SPARCV9 Manual

User Manual:

Open the PDF directly: View PDF .
Page Count: 399

Download
Open PDF In Browser	View PDF

The SPARC Architecture Manual
Version 9

SPARC International, Inc.
San Jose, California

David L. Weaver / Tom Germond
Editors

SAV09R1459912

PT R Prentice Hall, Englewood Cliffs, New Jersey 07632

SPARC® is a registered trademark of SPARC International, Inc.
The SPARC logo is a registered trademark of SPARC International, Inc.
UNIX® is a registered trademark of UNIX System Laboratories, Inc.

Copyright © 1994 SPARC International, Inc.
Published by PTR Prentice Hall
Prentice-Hall, Inc.
A Paramount Communications Company
Englewood Cliffs, New Jersey 07632
The publisher offers discounts on this book when ordered in bulk quantities. For more
information, contact:
Corporate Sales Department
PT R Prentice Hall
113 Sylvan Avenue
Englewood Cliffs, NJ 07632
Phone: (201) 592-2863
Fax:
(201) 592-2249
All rights reserved.
No part of this publication may be reproduced, stored in a retrieval system, or transmitted in any
form or by any means, electronic, mechanical, photocopying, recording or otherwise, without the
prior permission of the copyright owners.
Restricted rights legend: use, duplication, or disclosure by the U. S. Government is subject to
restrictions set forth in subparagraph (c)(1)(ii) of the Rights in Technical Data and Computer
Software clause at DFARS 52.227-7013 and in similar clauses in the FAR and NASA FAR Supplement.
Printed in the United States of America
10

ISBN

0-13-825001-4

PRENTICE-HALL INTERNATIONAL (UK) LIMITED, London
PRENTICE-HALL OF AUSTRALIA PTY. LIMITED, Sydney
PRENTICE-HALL CANADA INC., Toronto
PRENTICE-HALL HISPANOAMERICANA, S.A., Mexico
PRENTICE-HALL OF INDIA PRIVATE LIMITED, New Delhi
PRENTICE-HALL OF JAPAN, INC., Tokyo

SIMON & SCHUSTER ASIA PTE. LTD., Singapore
EDITORA PRENTICE-HALL DO BRASIL, LTDA., Rio de Janeiro

Contents
Introduction .............................................................................................................
0.1
SPARC ....................................................................................................
0.2
Processor Needs for the 90s and Beyond ................................................
0.3
SPARC-V9: A Robust RISC for the Next Century ................................
0.3.1
64-bit Data and Addresses .......................................................
0.3.2
Improved System Performance ................................................
0.3.3
Advanced Optimizing Compilers ............................................
0.3.4
Advanced Superscalar Processors ............................................
0.3.5
Advanced Operating Systems ..................................................
0.3.6
Fault Tolerance ........................................................................
0.3.7
Fast Traps and Context Switching ...........................................
0.3.8
Big- and Little-Endian Byte Orders .........................................
0.4
Summary .................................................................................................

xiii
xiii
xiv
xiv
xiv
xv
xvi
xvii
xvii
xviii
xviii
xix
xix

Editors’ Notes ..........................................................................................................
Acknowledgments ...............................................................................................
Personal Notes ....................................................................................................

xxi
xxi
xxi

Overview ............................................................................................................
1.1
Notes About this Book ............................................................................
1.1.1
Audience ..................................................................................
1.1.2
Where to Start ..........................................................................
1.1.3
Contents ...................................................................................
1.1.4
Editorial Conventions ..............................................................
1.2
The SPARC-V9 Architecture .................................................................
1.2.1
Features ....................................................................................
1.2.2
Attributes ..................................................................................
1.2.3
System Components ................................................................
1.2.4
Binary Compatibility ...............................................................
1.2.5
Architectural Definition ...........................................................
1.2.6
SPARC-V9 Compliance ..........................................................

1
1
1
1
1
3
4
4
5
6
6
7
7

Definitions ..........................................................................................................

Architectural Overview ....................................................................................
3.1
SPARC-V9 Processor .............................................................................
3.1.1
Integer Unit (IU) ......................................................................
3.1.2
Floating-Point Unit (FPU) ......................................................
3.2
Instructions ..............................................................................................
3.2.1
Memory Access .......................................................................

15
15
15
16
16
17

3.2.2
Arithmetic/Logical/Shift Instructions ......................................
3.2.3
Control Transfer .......................................................................
3.2.4
State Register Access ...............................................................
3.2.5
Floating-Point Operate .............................................................
3.2.6
Conditional Move ....................................................................
3.2.7
Register Window Management ................................................
Traps .......................................................................................................

19
19
20
20
20
20
21

Data Formats .....................................................................................................
4.1
Signed Integer Byte .................................................................................
4.2
Signed Integer Halfword .........................................................................
4.3
Signed Integer Word ...............................................................................
4.4
Signed Integer Double ............................................................................
4.5
Signed Extended Integer .........................................................................
4.6
Unsigned Integer Byte ............................................................................
4.7
Unsigned Integer Halfword .....................................................................
4.8
Unsigned Integer Word ...........................................................................
4.9
Unsigned Integer Double ........................................................................
4.10 Unsigned Extended Integer .....................................................................
4.11 Tagged Word ..........................................................................................
4.12 Floating-Point Single Precision ..............................................................
4.13 Floating-Point Double Precision .............................................................
4.14 Floating-Point Quad Precision ................................................................

23
23
24
24
24
24
24
24
25
25
25
25
25
26
26

Registers .............................................................................................................
5.1
Nonprivileged Registers ..........................................................................
5.1.1
General Purpose r Registers .....................................................
5.1.2
Special r Registers ...................................................................
5.1.3
IU Control/Status Registers .....................................................
5.1.4
Floating-Point Registers ..........................................................
5.1.5
Condition Codes Register (CCR) ............................................
5.1.6
Floating-Point Registers State (FPRS) Register ......................
5.1.7
Floating-Point State Register (FSR) ........................................
5.1.8
Address Space Identifier Register (ASI) .................................
5.1.9
TICK Register (TICK) .............................................................
5.2
Privileged Registers ................................................................................
5.2.1
Processor State Register (PSTATE) ........................................
5.2.2
Trap Level Register (TL) .........................................................
5.2.3
Processor Interrupt Level (PIL) ...............................................
5.2.4
Trap Program Counter (TPC) ..................................................
5.2.5
Trap Next Program Counter (TNPC) .......................................

29
30
30
34
35
36
40
42
43
50
50
51
51
54
54
55
55

3.3

5.2.6
5.2.7
5.2.8
5.2.9
5.2.10
5.2.11
5.2.12
5.2.13

Trap State (TSTATE) ..............................................................
Trap Type Register (TT) ..........................................................
Trap Base Address (TBA) .......................................................
Version Register (VER) ...........................................................
Register-Window State Registers ............................................
Ancillary State Registers (ASRs) .............................................
Floating-Point Deferred-Trap Queue (FQ) ..............................
IU Deferred-Trap Queue ..........................................................

56
56
57
57
58
60
61
61

Instructions ........................................................................................................
6.1
Instruction Execution ..............................................................................
6.2
Instruction Formats .................................................................................
6.2.1
Instruction Fields .....................................................................
6.3
Instruction Categories .............................................................................
6.3.1
Memory Access Instructions ....................................................
6.3.2
Memory Synchronization Instructions .....................................
6.3.3
Integer Arithmetic Instructions ................................................
6.3.4
Control-Transfer Instructions (CTIs) .......................................
6.3.5
Conditional Move Instructions ................................................
6.3.6
Register Window Management Instructions ............................
6.3.7
State Register Access ...............................................................
6.3.8
Privileged Register Access ......................................................
6.3.9
Floating-Point Operate (FPop) Instructions .............................
6.3.10 Implementation-Dependent Instructions ..................................
6.3.11 Reserved Opcodes and Instruction Fields ................................
6.4
Register Window Management ...............................................................
6.4.1
Register Window State Definition ...........................................
6.4.2
Register Window Traps ...........................................................

63
63
63
66
68
69
76
76
77
80
82
84
84
84
85
85
85
85
86

Traps ..................................................................................................................
7.1
Overview .................................................................................................
7.2
Processor States, Normal and Special Traps ...........................................
7.2.1
RED_state ................................................................................
7.2.2
Error_state ................................................................................
7.3
Trap Categories .......................................................................................
7.3.1
Precise Traps ............................................................................
7.3.2
Deferred Traps .........................................................................
7.3.3
Disrupting Traps ......................................................................
7.3.4
Reset Traps ...............................................................................
7.3.5
Uses of the Trap Categories .....................................................
7.4
Trap Control ............................................................................................

89
89
90
90
94
94
95
95
96
97
97
99

7.4.1
PIL Control ..............................................................................
7.4.2
TEM Control ............................................................................
Trap-Table Entry Addresses ...................................................................
7.5.1
Trap Table Organization ..........................................................
7.5.2
Trap Type (TT) ........................................................................
7.5.3
Trap Priorities ..........................................................................
Trap Processing .......................................................................................
7.6.1
Normal Trap Processing .........................................................
7.6.2
Special Trap Processing ...........................................................
Exception and Interrupt Descriptions .....................................................

99
100
100
101
101
104
105
106
108
113

Memory Models ................................................................................................
8.1
Introduction .............................................................................................
8.2
Memory, Real Memory, and I/O Locations ............................................
8.3
Addressing and Alternate Address Spaces .............................................
8.4
The SPARC-V9 Memory Model ............................................................
8.4.1
The SPARC-V9 Program Execution Model ............................
8.4.2
The Processor/Memory Interface Model .................................
8.4.3
The MEMBAR Instruction ......................................................
8.4.4
Memory Models .......................................................................
8.4.5
Mode Control ...........................................................................
8.4.6
Hardware Primitives for Mutual Exclusion .............................
8.4.7
Synchronizing Instruction and Data Memory ..........................

119
119
120
121
123
123
125
126
128
129
130
131

A Instruction Definitions (Normative) ................................................................
A.1
Overview .................................................................................................
A.2
Add ..........................................................................................................
A.3
Branch on Integer Register with Prediction (BPr) .................................
A.4
Branch on Floating-Point Condition Codes (FBfcc) ..............................
A.5
Branch on Floating-Point Condition Codes with Prediction (FBPfcc) ...
A.6
Branch on Integer Condition Codes (Bicc) .............................................
A.7
Branch on Integer Condition Codes with Prediction (BPcc) ...................
A.8
Call and Link ...........................................................................................
A.9
Compare and Swap .................................................................................
A.10 Divide (64-bit / 32-bit) ............................................................................
A.11 DONE and RETRY .................................................................................
A.12 Floating-Point Add and Subtract ............................................................
A.13 Floating-Point Compare ..........................................................................
A.14 Convert Floating-Point to Integer ...........................................................
A.15 Convert Between Floating-Point Formats ..............................................
A.16 Convert Integer to Floating-Point ...........................................................

133
133
137
138
140
143
146
148
151
152
154
157
158
159
161
162
163

7.5

7.6

7.7
8

A.17
A.18
A.19
A.20
A.21
A.22
A.23
A.24
A.25
A.26
A.27
A.28
A.29
A.30
A.31
A.32
A.33
A.34
A.35
A.36
A.37
A.38
A.39
A.40
A.41
A.42

A.43
A.44
A.45
A.46
A.47
A.48
A.49
A.50
A.51
A.52
A.53
A.54

Floating-Point Move ...............................................................................
Floating-Point Multiply and Divide ........................................................
Floating-Point Square Root
.........................................................
Flush Instruction Memory ......................................................................
Flush Register Windows ........................................................................
Illegal Instruction Trap ...........................................................................
Implementation-Dependent Instructions .................................................
Jump and Link .........................................................................................
Load Floating-Point ................................................................................
Load Floating-Point from Alternate Space .............................................
Load Integer ............................................................................................
Load Integer from Alternate Space .........................................................
Load-Store Unsigned Byte ......................................................................
Load-Store Unsigned Byte to Alternate Space .......................................
Logical Operations ..................................................................................
Memory Barrier ......................................................................................
Move Floating-Point Register on Condition (FMOVcc) ........................
Move F-P Register on Integer Register Condition (FMOVr) .................
Move Integer Register on Condition (MOVcc) ......................................
Move Integer Register on Register Condition (MOVR) .........................
Multiply and Divide (64-bit) ...................................................................
Multiply (32-bit) .....................................................................................
Multiply Step ..........................................................................................
No Operation ...........................................................................................
Population Count ....................................................................................
Prefetch Data ...........................................................................................
A.42.1 Prefetch Variants ......................................................................
A.42.2 General Comments ...................................................................
Read Privileged Register .........................................................................
Read State Register .................................................................................
RETURN .................................................................................................
SAVE and RESTORE .............................................................................
SAVED and RESTORED .......................................................................
SETHI .....................................................................................................
Shift .........................................................................................................
Software-Initiated Reset ..........................................................................
Store Barrier ............................................................................................
Store Floating-Point ................................................................................
Store Floating-Point into Alternate Space ..............................................
Store Integer ............................................................................................

164
165
166
167
169
170
171
172
173
176
178
180
182
183
184
186
188
192
194
198
199
200
202
204
205
206
207
209
211
214
216
217
219
220
221
223
224
225
227
229

A.55
A.56
A.57
A.58
A.59
A.60
A.61
A.62
A.63

Store Integer into Alternate Space ..........................................................
Subtract ...................................................................................................
Swap Register with Memory ..................................................................
Swap Register with Alternate Space Memory ........................................
Tagged Add ............................................................................................
Tagged Subtract ......................................................................................
Trap on Integer Condition Codes (Tcc) ..................................................
Write Privileged Register ........................................................................
Write State Register
....................................................................

231
233
234
235
237
238
240
242
244

B IEEE Std 754-1985 Requirements for SPARC-V9 (Normative) ..................
B.1
Traps Inhibit Results ...............................................................................
B.2
NaN Operand and Result Definitions .....................................................
B.2.1
Untrapped Result in Different Format from Operands ............
B.2.2
Untrapped Result in Same Format as Operands ......................
B.3
Trapped Underflow Definition (UFM = 1) .............................................
B.4
Untrapped Underflow Definition (UFM = 0) .........................................
B.5
Integer Overflow Definition ...................................................................
B.6
Floating-Point Nonstandard Mode ..........................................................

247
247
248
248
248
249
249
250
250

C SPARC-V9 Implementation Dependencies (Normative) ..............................
C.1
Definition of an Implementation Dependency ........................................
C.2
Hardware Characteristics ........................................................................
C.3
Implementation Dependency Categories ................................................
C.4
List of Implementation Dependencies ....................................................

251
251
251
252
252

D Formal Specification of the Memory Models (Normative) ...........................
D.1
Processors and Memory ..........................................................................
D.2
An Overview of the Memory Model Specification ................................
D.3
Memory Transactions .............................................................................
D.3.1
Memory Transactions ..............................................................
D.3.2
Program Order .........................................................................
D.3.3
Dependence Order ...................................................................
D.3.4
Memory Order .........................................................................
D.4
Specification of Relaxed Memory Order (RMO) ...................................
D.4.1
Value Atomicity .......................................................................
D.4.2
Store Atomicity ........................................................................
D.4.3
Atomic Memory Transactions .................................................
D.4.4
Memory Order Constraints ......................................................
D.4.5
Value of Memory Transactions ...............................................
D.4.6
Termination of Memory Transactions .....................................

261
261
262
263
263
263
264
265
265
265
266
266
266
266
267

D.4.7
Flush Memory Transaction ......................................................
Specification of Partial Store Order (PSO) .............................................
Specification of Total Store Order (TSO) ...............................................
Examples Of Program Executions ..........................................................
D.7.1
Observation of Store Atomicity ...............................................
D.7.2
Dekker’s Algorithm .................................................................
D.7.3
Indirection Through Processors ...............................................
D.7.4
PSO Behavior ...........................................................................
D.7.5
Application to Compilers .........................................................
D.7.6
Verifying Memory Models ......................................................

267
267
267
267
267
269
269
270
271
272

E Opcode Maps (Normative) ...............................................................................
E.1
Overview .................................................................................................
E.2
Tables ......................................................................................................

273
273
273

F SPARC-V9 MMU Requirements (Informative) ...........................................
F.1
Introduction .............................................................................................
F.1.1
Definitions ................................................................................
F.2
Overview .................................................................................................
F.3
The Processor-MMU Interface ...............................................................
F.3.1
Information the MMU Expects from the Processor .................
F.3.2
Attributes the MMU Associates with Each Mapping ..............
F.3.3
Information the MMU Sends to the Processor ........................
F.4
Components of the SPARC-V9 MMU Architecture ..............................
F.4.1
Virtual-to-Physical Address Translation ..................................
F.4.2
Memory Protection ..................................................................
F.4.3
Prefetch and Non-Faulting Load Violation ..............................
F.4.4
Contexts ...................................................................................
F.4.5
Fault Status and Fault Address ................................................
F.4.6
Referenced and Modified Statistics .........................................
F.5
RED_state Processing .............................................................................
F.6
Virtual Address Aliasing .........................................................................
F.7
MMU Demap Operation .........................................................................
F.8
SPARC-V9 Systems without an MMU ..................................................

281
281
281
281
282
283
284
284
285
285
286
286
286
287
288
288
288
288
289

G Suggested Assembly Language Syntax (Informative) ..................................
G.1
Notation Used .........................................................................................
G.1.1
Register Names ........................................................................
G.1.2
Special Symbol Names ............................................................
G.1.3
Values ......................................................................................
G.1.4
Labels .......................................................................................

291
291
291
292
294
295

D.5
D.6
D.7

G.1.5
Other Operand Syntax ..............................................................
G.1.6
Comments ................................................................................
Syntax Design .........................................................................................
Synthetic Instructions .............................................................................

295
296
296
297

H Software Considerations (Informative) .........................................................
H.1
Nonprivileged Software ..........................................................................
H.1.1
Registers ...................................................................................
H.1.2
Leaf-Procedure Optimization ..................................................
H.1.3
Example Code for a Procedure Call .........................................
H.1.4
Register Allocation within a Window .....................................
H.1.5
Other Register-Window-Usage Models ..................................
H.1.6
Self-Modifying Code ..............................................................
H.1.7
Thread Management ................................................................
H.1.8
Minimizing Branch Latency ....................................................
H.1.9
Prefetch ....................................................................................
H.1.10 Nonfaulting Load .....................................................................
H.2
Supervisor Software ................................................................................
H.2.1
Trap Handling ..........................................................................
H.2.2
Example Code for Spill Handler ..............................................
H.2.3
Client-Server Model .................................................................
H.2.4
User Trap Handlers ..................................................................

301
301
301
304
306
307
308
308
309
309
310
313
315
315
316
316
317

Extending the SPARC-V9 Architecture (Informative) ................................
I.1
Addition of SPARC-V9 Extensions .......................................................
I.1.1
Read/Write Ancillary State Registers (ASRs) .........................
I.1.2
Implementation-Dependent and Reserved Opcodes ................

321
321
321
321

Programming With the Memory Models (Informative) ..............................
J.1
Memory Operations ................................................................................
J.2
Memory Model Selection .......................................................................
J.3
Processors and Processes ........................................................................
J.4
Higher-Level Programming Languages and Memory Models ...............
J.5
Portability And Recommended Programming Style ...............................
J.6
Spin Locks ..............................................................................................
J.7
Producer-Consumer Relationship ...........................................................
J.8
Process Switch Sequence ........................................................................
J.9
Dekker’s Algorithm ................................................................................
J.10 Code Patching .........................................................................................
J.11 Fetch_and_Add .......................................................................................
J.12 Barrier Synchronization ..........................................................................

323
323
324
324
325
325
327
327
329
330
330
333
333

G.2
G.3

J.13
J.14

Linked List Insertion and Deletion .........................................................
Communicating With I/O Devices ..........................................................
J.14.1 I/O Registers With Side Effects ...............................................
J.14.2 The Control and Status Register (CSR) ...................................
J.14.3 The Descriptor .........................................................................
J.14.4 Lock-Controlled Access to a Device Register .........................

335
335
337
337
338
338

K Changes From SPARC-V8 to SPARC-V9 (Informative) .............................
K.1
Trap Model ..............................................................................................
K.2
Data Formats ...........................................................................................
K.3
Little-Endian Support ..............................................................................
K.4
Registers ..................................................................................................
K.5
Alternate Space Access ...........................................................................
K.6
Little-Endian Byte Order ........................................................................
K.7
Instruction Set .........................................................................................
K.8
Memory Model .......................................................................................

339
339
340
340
340
341
341
341
344

Bibliography ............................................................................................................
General References .............................................................................................
Memory Model References .................................................................................
Prefetching ..........................................................................................................

345
345
346
347

Index .........................................................................................................................

349

Introduction
Welcome to SPARC-V9, the most significant change to the SPARC architecture since it was
announced in 1987. SPARC-V9 extends the addresses of SPARC to 64 bits and adds a number of
new instructions and other enhancements to the architecture.1
SPARC-V9, like its predecessor SPARC-V8, is a microprocessor specification created by the
SPARC Architecture Committee of SPARC International. SPARC-V9 is not a specific chip; it is
an architectural specification that can be implemented as a microprocessor by anyone securing a
license from SPARC International.
SPARC International is a consortium of computer makers, with membership open to any company
in the world. Executive member companies each designate one voting member to participate on
the SPARC Architecture Committee. Over the past several years, the architecture committee has
been hard at work designing the next generation of the SPARC architecture.
Typically, microprocessors are designed and implemented in secret by a single company. Then the
company spends succeeding years defending its proprietary rights in court against its competitors.
With SPARC, it is our intention to make it easy for anyone to design and implement to this architectural specification. Several SPARC-V9 implementations are already underway, and we expect
many more companies to design systems around this microprocessor standard in the coming
years.

0.1 SPARC
SPARC stands for a Scalable Processor ARChitecture. SPARC has been implemented in processors used in a range of computers from laptops to supercomputers. SPARC International member
companies have implemented over a dozen different compatible microprocessors since SPARC
was first announced—more than any other microprocessor family with this level of binary compatibility. As a result, SPARC today boasts over 8000 compatible software application programs.
SPARC-V9 maintains upwards binary compatibility for application software, which is a very
important feature.
Throughout the past six years, the SPARC architecture has served our needs well. But at the same
time, VLSI technology, compiler techniques and users’ needs have changed. The time is right to
upgrade SPARC for the coming decade.

0.2 Processor Needs for the 90s and Beyond
The design of Reduced Instruction Set Processors (RISC) began in earnest in the early 1980s.
Early RISC processors typically were characterized by a load-store architecture, single instruction-per-cycle execution, and 32-bit addressing. The instruction set architecture of these early

1.For a complete list of changes between SPARC-V8 and SPARC-V9, see Appendix K.

RISC chips was well matched to the level of computer optimization available in the early 1980s,
and provided a minimal interface for the UNIX™ operating system.
The computer industry has grown significantly in the last decade. Computer users need more for
the 1990s than these early RISCs provided; they demand more powerful systems today, and yet
they continue to want their systems to have good performance growth and compatibility into the
future.The applications of the future—highly interactive and distributed across multiple platforms—will require larger address spaces and more sophisticated operating system interfaces.
Tomorrow’s architectures must provide better support for multiprocessors, lightweight threads,
and object oriented programming. Modern computer systems must also perform more reliably
than in the past.
It is interesting to observe the evolution of RISC architectures. Without sufficient instruction
encoding, some microprocessors have been unable to provide for either larger address spaces or
new instruction functionality. Others have provided 64-bit addressing, but still have not changed
much from the RISCs of the 1980s. Fortunately, SPARC’s designers had sufficient foresight to
allow for all of the changes we felt were needed to keep SPARC a viable architecture for the long
term.

0.3 SPARC-V9: A Robust RISC for the Next Century
SPARC-V9 is a robust RISC architecture that will remain competitive well into the next century.
The SPARC-V9 architecture delivers on this promise by enhancing SPARC-V8 to provide explicit
support for:
— 64-bit virtual addresses and 64-bit integer data
— Improved system performance
— Advanced optimizing compilers
— Superscalar implementations
— Advanced operating systems
— Fault tolerance
— Extremely fast trap handling and context switching
— Big- and little-endian byte orders
0.3.1 64-bit Data and Addresses
SPARC-V9 directly supports 64-bit virtual addresses and integer data sizes up to 64 bits. All
SPARC-V8 integer registers have been extended from 32 to 64 bits. There are also several new
instructions that explicitly manipulate 64-bit values. For example, 64-bit integer values can be
loaded and stored directly with the LDX and STX instructions.
Despite these changes, 64-bit SPARC-V9 microprocessors will be able to execute programs compiled for 32-bit SPARC-V8 processors. The principles of two’s complement arithmetic made

upward compatibility straightforward to accomplish. Arithmetic operations, for example, specified arithmetic on registers, independent of the length of the register. The low order 32-bits of
arithmetic operations will continue to generate the same values they did on SPARC-V8 processors. Since SPARC-V8 programs paid attention to only the low order 32-bits, these programs will
execute compatibly. Compatibility for SPARC-V9 was accomplished by making sure that all previously existing instructions continued to generate exactly the same result in the low order 32-bits
of registers. In some cases this meant adding new instructions to operate on 64-bit values. For
example, shift instructions now have an additional 64-bit form.
In order to take advantage of SPARC-V9’s extended addressing and advanced capabilities,
SPARC-V8 programs must be recompiled. SPARC-V9 compilers will take full advantage of the
new features of the architecture, extending the addressing range and providing access to all of the
added functionality.
0.3.2 Improved System Performance
Performance is one of the biggest concerns for both computer users and manufacturers. We’ve
changed some basic things in the architecture to allow SPARC-V9 systems to achieve higher performance. The new architecture contains 16 additional double-precision floating-point registers,
bringing the total to 32. These additional registers reduce memory traffic, allowing programs to
run faster. The new floating-point registers are also addressable as eight quad-precision registers.
SPARC-V9’s support for a 128-bit quad floating-point format is unique for microprocessors.
SPARC-V9 supports four floating-point condition code registers, where SPARC-V8 supported
only one. SPARC-V9 processors can provide more parallelism for a Superscalar machine by
launching several instructions at a time. With only one condition code register, instructions would
have a serial dependence waiting for the single condition code register to be updated. The new
floating-point condition code registers allow SPARC-V9 processors to initiate up to four floatingpoint compares simultaneously.
We’ve also extended the instruction set to increase performance by adding:
— 64-bit integer multiply and divide instructions.
— Load and store floating-point quadword instructions.
— Software settable branch prediction, which gives the hardware a greater probability of
keeping the processor pipeline full.
— Branches on register value, which eliminate the need to execute a compare instruction.
This provides the appearance of multiple integer condition codes, eliminating a potential
bottleneck and creating similar possibilities for parallelism in integer calculations that we
obtained from multiple floating-point condition codes.
— Conditional move instructions, which allow many branches to be eliminated.

0.3.3 Advanced Optimizing Compilers
We expect to see many new optimizing compilers in the coming decade, and we have included
features in SPARC-V9 that these compilers will be able to use to provide higher performance.
SPARC-V9 software can explicitly prefetch data and instructions, thus reducing the memory
latency, so a program need not wait as long for its code or data. If compilers generate code to
prefetch code and data far enough in advance, the data can be available as soon as the program
needs to use it, reducing cache miss penalties and pipeline stalls.
SPARC-V9 has support for loading data not aligned on “natural” boundaries. Because of the way
the FORTRAN language is specified, compilers often cannot determine whether double-precision
floating-point data is aligned on doubleword boundaries in memory. In many RISC architectures,
FORTRAN compilers generate two single-precision loads instead of one double-precision load.
This can be a severe performance bottleneck. SPARC-V9 allows the compiler to always use the
most efficient load and store instructions. On those rare occasions when the data is not aligned,
the underlying architecture provides for a fast trap to return the requested data, without the
encumbrances of providing unaligned accesses directly in the memory system hardware. This net
effect is higher performance on many FORTRAN programs.
SPARC-V9 also supports non-faulting loads, which allow compilers to move load instructions
ahead of conditional control structures that guard their use. The semantics of non-faulting loads
are the same as for other loads, except when a nonrecoverable fault such as an address-out-ofrange error occurs. These faults are ignored, and hardware and system software cooperate to make
the load appear to complete normally, returning a zero result. This optimization is particularly
useful when optimizing for superscalar processors. Consider this C program fragment:
if (p != NULL) x

= *p + y;

With non-faulting loads, the load of *p can be moved up by the compiler to before the check for
p != NULL, allowing overlapped execution. A normal load on many processors would cause the
program to be aborted if this optimization was performed and p was NULL. The effect is equivalent to this transformation:
temp_register

= *p;

if ( p != NULL ) x

= temp_register + y;

Imagine a superscalar processor that could execute four instructions per cycle, but only one of
which could be a load or store. In a loop of eight instructions containing two loads, it might turn
out that without this transformation it would not be possible to schedule either of the loads in the
first group of four instructions. In this case a third or possibly fourth clock cycle might be necessary for each loop iteration instead of the minimal two cycles. Improving opportunities for better
instruction scheduling could have made a factor of two difference in performance for this example. Good instruction scheduling is critical.
Alias detection is a particularly difficult problem for compilers. If a compiler cannot tell whether
two pointers might point to the same value in memory, then it is not at liberty to move loads up
past previous store instructions. This can create a difficult instruction scheduling bottleneck.
SPARC-V9 contains specific instructions to enable the hardware to detect pointer aliases, and
offers the compiler a simple solution to this difficult problem. Two pointers can be compared and

the results of this comparison stored in an integer register. The FMOVRZ instruction, for example, will conditionally move a floating-point register based on the result of this prior test. This
instruction can be used to correct aliasing problems and allow load instructions to be moved up
past stores. As with the previous example, this can make a significant difference in overall program performance.
Finally, we’ve added a TICK register, which is incremented once per machine cycle. This register
can be read by a user program to make simple and accurate measurements of program performance.
0.3.4 Advanced Superscalar Processors
SPARC-V9 includes support for advanced Superscalar processor designs. CPU designers are
learning to execute more instructions per cycle every year with new pipelines. Two to three
instructions at a time is becoming commonplace. We eventually expect to be able to execute eight
to sixteen instructions at a time with the SPARC architecture. To accomplish this, we’ve made
enhancements to provide better support for Superscalar execution.
Many of these changes were driven by the experience gained from implementing Texas Instruments’ SuperSPARC and Ross Technologies’ HyperSPARC, both Superscalar chips. SPARC’s
simple-to-decode, fixed-length instructions, and separate integer and floating-point units lend
themselves to Superscalar technology.
In addition, SPARC-V9 provides more floating-point registers, support for non-faulting loads,
multiple condition codes, branch prediction, and branches on integer register contents. All of
these features allow for more parallelism within the processor. For the memory system, we’ve
added a sophisticated memory barrier instruction, which allows system programmers to specify
the minimum synchronization needed to ensure correct operation.
0.3.5 Advanced Operating Systems
The operating system interface has been completely redesigned in SPARC-V9 to better support
operating systems of the 1990s. There are new privileged registers and a new structure to those
registers, which makes it much simpler to access important control information in the machine.
Remember, the change in the operating system interface has no effect on application software;
user-level programs do not see these changes, and thus, are binary compatible without recompilation.
Several changes were made to support the new microkernel style of operating system design.
Nested trap levels allow more modular structuring of code, and are more efficient as well.
SPARC-V9 provides improved support for lightweight threads and faster context switching than
was possible in previous SPARC architectures. We’ve accomplished this by making register windows more flexible than they were in earlier SPARC processors, allowing the kernel to provide a
separate register bank to each running process. Thus, the processor can perform a context switch
with essentially no overhead. The new register window implementation also provides better support for object-oriented operating systems by speeding up interprocess communication across different domains. There is a mechanism to provide efficient server access to client address spaces

using user address space identifiers. The definition of a nucleus address space allows the operating
system to exist in a different address space than that of the user program.
Earlier SPARC implementations supported multiprocessors; now we’ve added support for very
large-scale multiprocessors, including a memory barrier instruction and a new memory model we
call relaxed memory order (RMO). These features allow SPARC-V9 CPUs to schedule memory
operations to achieve high performance, while still doing the synchronization and locking operations needed for shared-memory multiprocessing.
Finally we’ve added architectural support that helps the operating system provide “clean” register
windows to its processes. A clean window is guaranteed to contain zeroes initially, and only data
or addresses generated by the process during its lifetime. This makes it easier to implement a
secure operating system, which must provide absolute isolation between its processes.
0.3.6 Fault Tolerance
Most existing microprocessor architectures do not provide explicit support for reliability and
fault-tolerance. You might build a reliable and fault-tolerant machine without explicit support, but
providing it saves a lot of work, and the machine will cost less in the long run.
We’ve incorporated a number of features in SPARC-V9 to address these shortcomings. First,
we’ve added a compare-and-swap instruction. This instruction has well-known fault-tolerant features and is also an efficient way to do multiprocessor synchronization.
We’ve also added support for multiple levels of nested traps, which allow systems to recover
gracefully from various kinds of faults, and to contain more efficient trap handlers. Nested traps
are described in the next section.
Finally, we’ve added a special new processor state called RED_state, short for Reset, Error and
Debug state. It fully defines the expected behavior when the system is faced with catastrophic
errors, and during reset processing when it is returning to service. This level of robustness is
required to build fault-tolerant systems.
0.3.7 Fast Traps and Context Switching
We have also worked hard to provide very fast traps and context switching in SPARC-V9. We
have re-architected the trap entry mechanism to transfer control into the trap handlers very
quickly. We’ve also added eight new registers called “alternate globals,” so the trap handler has a
fresh register set to use immediately upon entry; the software need not store registers before it can
begin to do its work. This allows very fast instruction emulation and very short interrupt response
times.
We have also added support for multiple levels of nested traps. It is very useful for the machine to
allow a trap handler to generate a trap. SPARC-V8 trap handlers were not allowed to cause
another trap. With support for nested traps, we have seen some trap handlers reduced from one
hundred instructions to less than twenty. Obviously, this creates a big performance improvement,
but it also allows a much simpler operating system design.

We’ve also found a way to reduce the number of registers saved and restored between process
executions, which provides faster context switching. The architecture provides separate dirty bits
for the original (lower) and the new (upper) floating-point registers. If a program has not modified
any register in one of the sets, there is no need to save that set during a context switch.
0.3.8 Big- and Little-Endian Byte Orders
Finally, we have provided support for data created on little-endian processors such as the 80x86
family. The architecture allows both user and supervisor code to explicitly access data in littleendian byte order. It is also possible to change the default byte order to little-endian in user mode
only, in supervisor mode only, or in both. This allows SPARC-V9 to support mixed byte order
systems.

0.4 Summary
As you can see, SPARC-V9 is a significant advance over its predecessors. We have provided 64bit data and addressing, support for fault tolerance, fast context switching, support for advanced
compiler optimizations, efficient design for Superscalar processors, and a clean structure for modern operating systems. And we’ve done it all with 100% upwards binary compatibility for application programs. We believe that this is a significant achievement.
In the future, we envision superior SPARC-V9 implementations providing high performance, stellar reliability, and excellent cost efficiency—just what computer users are asking for. SPARC has
been the RISC leader for the last five years. With the changes we have made in SPARC-V9, we
expect it to remain the RISC leader well into the next century.
Speaking for the Committee members, we sincerely hope that you profit from our work.
— David R. Ditzel
Chairman, SPARC Architecture Committee

Editors’ Notes
Acknowledgments
The members of SPARC International’s Architecture Committee devoted a great deal of time over
a period of three years designing the SPARC-V9 architecture. As of Summer 1993, the committee
membership was: Dennis Allison, Hisashige Ando, Jack Benkual, Joel Boney (vice-chair), David
Ditzel (chair), Hisakazu Edamatsu, Kees Mage, Steve Krueger, Craig Nelson, Chris Thomson,
David Weaver, and Winfried Wilcke.
Joel Boney wrote the original “V9 Delta Documents” that supplied much of the new material for
this specification.
Others who have made significant contributions to SPARC-V9 include Greg Blanck, Jeff Broughton (former vice-chair), David Chase, Steve Chessin, Bob Cmelik, David Dill, Kourosh Gharachorloo, David Hough, Bill Joy, Ed Kelly, Steve Kleiman, Jaspal Kohli, Les Kohn, Shing Kong,
Paul Loewenstein, Guillermo “Matute” Maturana, Mike McCammon, Bob Montoye, Chuck
Narad, Andreas Nowatzyk, Seungjoon Park, David Patterson, Mike Powell, John Platko, Steve
Richardson, Robert Setzer, Pradeep Sindhu, George Taylor, Marc Tremblay, Rudolf Usselmann, J.
J. Whelan, Malcolm Wing, and Robert Yung.
Joel Boney, Dennis Allison, Steve Chessin, and Steve Muchnick deserve distinction as “Ace”
reviewers. They performed meticulous reviews, eliminating countless bugs in the specification.
Our thanks to all of the above people for their support, critiques, and contributions to this book
over the last three years!
Personal Notes
Three years — that’s a long time to be in labor! It is with a great deal of pride (and frankly, relief!)
that I see this book go to print.
The SPARC Architecture Committee comprised roughly a dozen people, all top computer architects in the industry, from diverse companies. Yet — and this was the most incredible part of the
whole process — this group was able to set aside personal egos and individual company interests,
and work not just as a committee, but as a real Team. This kind of cooperation and synergy
doesn’t happen every day. Years from now, I’ll look back at this work and still be proud to have
been a part of this group, and of what we created. . . . “Way to go, gang — we done good!”
Special kudos are due Tom Germond, whose expertise and sharp eye for detail were instrumental
in preparing this book. He fearlessly performed a complex but accurate conversion of this specification from one document-preparation system to a wildly different one. Tom made countless
improvements to the specification’s substance and style, and tenaciously followed numerous open
technical issues through to resolution. This book would simply not have been the same without
him. Thanks for being there, Tom.
— David Weaver, Editor

Well, it’s three o’clock in the morning and I’m in the middle of yet another SPARC-V9 allnighter. I haven’t lost this much sleep since my firstborn was first born. But I must say, it’s been
great fun bringing this baby to life.
My deepest gratitude to every member of our team, and a tiny extra measure of thanks to a special
few. To Joel Boney for his generous and unwavering support. To Dennis Allison for his constant
striving for excellence and clarity. To Steve Muchnick for his astonishing mastery of the details.
To Steve Chessin for always going to the heart of the issues. And to Jane Bonnell, our editor at
Prentice-Hall, for helping us turn a technical specification into a real book.
And finally,warm thanks to Dave Weaver, a good friend and an easy person to work for. You created the opportunity for me to join the team, and you got me through the rough times with all
those great movie-and-hot-tub parties. Until next time....
— Tom Germond, Co-editor

1 Overview
This specification defines a 64-bit architecture called SPARC-V9, which is upward-compatible
with the existing 32-bit SPARC-V8 microprocessor architecture. This specification includes, but
is not limited to, the definition of the instruction set, register model, data types, instruction
opcodes, trap model, and memory model.

1.1 Notes About this Book
1.1.1 Audience
Audiences for this specification include implementors of the architecture, students of computer
architecture, and developers of SPARC-V9 system software (simulators, compilers, debuggers,
and operating systems, for example). Software developers who need to write SPARC-V9 software
in assembly language will also find this information useful.
1.1.2 Where to Start
If you are new to the SPARC architecture, read Chapter 2 and Chapter 3 for an overview, then
look into the subsequent chapters and appendixes for more details in areas of interest to you.
If you are already familiar with SPARC-V8, you will want to review the list of changes in Appendix K, “Changes From SPARC-V8 to SPARC-V9.” For additional detail, review the following
chapters:
— Chapter 5, “Registers,” for a description of the register set.
— Chapter 6, “Instructions,” for a description of the new instructions.
— Chapter 7, “Traps,” for a description of the trap model.
— Chapter 8, “Memory Models,” for a description of the memory models.
— Appendix A, “Instruction Definitions,” for descriptions of new or changed instructions.
1.1.3 Contents
The manual contains these chapters:
— Chapter 1, “Overview,” describes the background, design philosophy, and high-level features of the architecture.
— Chapter 2, “Definitions,” defines some of the terms used in the specification.
— Chapter 3, “Architectural Overview,” is an overview of the architecture: its organization,
instruction set, and trap model.
— Chapter 4, “Data Formats,” describes the supported data types.
— Chapter 5, “Registers,” describes the register set.

— Chapter 6, “Instructions,” describes the instruction set.
— Chapter 7, “Traps,” describes the trap model.
— Chapter 8, “Memory Models,” describes the memory models.
These appendixes follow the chapters:
— Appendix A, “Instruction Definitions,” contains definitions of all SPARC-V9 instructions,
including tables showing the recommended assembly language syntax for each instruction.
— Appendix B, “IEEE Std 754-1985 Requirements for SPARC-V9,” contains information
about the SPARC-V9 implementation of the IEEE 754 floating-point standard.
— Appendix C, “SPARC-V9 Implementation Dependencies,” contains information about
features that may differ among conforming implementations.
— Appendix D, “Formal Specification of the Memory Models,” contains a formal description
of the memory models.
— Appendix E, “Opcode Maps,” contains tables detailing the encoding of all opcodes.
— Appendix F, “SPARC-V9 MMU Requirements,” describes the requirements that SPARCV9 imposes on Memory Management Units.
— Appendix G, “Suggested Assembly Language Syntax,” defines the syntactic conventions
used in the appendixes for the suggested SPARC-V9 assembly language. It also lists synthetic instructions that may be supported by SPARC-V9 assemblers for the convenience of
assembly language programmers.
— Appendix H, “Software Considerations,” contains general SPARC-V9 software considerations.
— Appendix I, “Extending the SPARC-V9 Architecture,” contains information on how an
implementation can extend the instruction set or register set.
— Appendix J, “Programming With the Memory Models,” contains information on programming with the SPARC-V9 memory models.
— Appendix K, “Changes From SPARC-V8 to SPARC-V9,” describes the differences
between SPARC-V8 and SPARC-V9.
A bibliography and an index complete the book.

1.1.4 Editorial Conventions
1.1.4.1 Fonts and Notational Conventions
Fonts are used as follows:
— Italic font is used for register names, instruction fields, and read-only register fields. For
example: “The rs1 field contains....”
— Typewriter font is used for literals and for software examples.
— Bold font is used for emphasis and the first time a word is defined. For example: “A precise trap is induced....”
— UPPER CASE items are acronyms, instruction names, or writable register fields. Some
common acronyms appear in the glossary in Chapter 2. Note that names of some instructions contain both upper- and lower-case letters.
— Italic sans serif font is used for exception and trap names. For example, “The
privileged_action exception....”
— Underbar characters join words in register, register field, exception, and trap names. Note
that such words can be split across lines at the underbar without an intervening hyphen.
For example: “This is true whenever the integer_condition_code field....”
— Reduced-size font is used in informational notes. See 1.1.4.4, “Informational Notes.”
The following notational conventions are used:
— Square brackets ‘[ ]’ indicate a numbered register in a register file. For example: “r[0]
contains....”
— Angle brackets ‘< >’ indicate a bit number or colon-separated range of bit numbers within
a field. For example: “Bits FSR<29:28> and FSR<12> are....”
— Curly braces ‘{ }’ are used to indicate textual substitution. For example, the string
“ASI_PRIMARY{_LITTLE}”
expands
to
“ASI_PRIMARY”
and
“ASI_PRIMARY_LITTLE”.
— The symbol designates concatenation of bit vectors. A comma ‘,’ on the left side of an
assignment separates quantities that are concatenated for the purpose of assignment. For
example, if X, Y, and Z are 1-bit vectors, and the 2-bit vector T equals 112, then
(X, Y, Z) ← 0 T
results in X = 0, Y = 1, and Z = 1.
1.1.4.2 Implementation Dependencies
Definitions of SPARC-V9 architecture implementation dependencies are indicated by the notation
“IMPL. DEP. #nn: Some descriptive text.” The number nn is used to enumerate the dependencies in
Appendix C, “SPARC-V9 Implementation Dependencies.” References to SPARC-V9 implemen-

tation dependencies are indicated by the notation “(impl. dep. #nn).” Appendix C lists the page
number on which each definition and reference occurs.
1.1.4.3 Notation for Numbers
Numbers throughout this specification are decimal (base-10) unless otherwise indicated. Numbers
in other bases are followed by a numeric subscript indicating their base (for example, 10012,
FFFF 000016). Long binary and hex numbers within the text have spaces inserted every four characters to improve readability. Within C or assembly language examples, numbers may be preceded by “0x” to indicate base-16 (hexadecimal) notation (for example, 0xffff0000).
1.1.4.4 Informational Notes
This manual provides several different types of information in notes; the information appears in a
reduced-size font. The following are illustrations of the various note types:
Programming Note:
These contain incidental information about programming using the SPARC-V9 architecture.
Implementation Note:
These contain information that may be specific to an implementation or may differ in different implementations.
Compatibility Note:
These contain information about features of SPARC-V9 that may not be compatible with SPARC-V8 implementations.

1.2 The SPARC-V9 Architecture
1.2.1 Features
SPARC-V9 includes the following principal features:
— A linear address space with 64-bit addressing.
— Few and simple instruction formats: All instructions are 32 bits wide, and are aligned on
32-bit boundaries in memory. Only load and store instructions access memory and perform I/O.
— Few addressing modes: A memory address is given as either “register + register” or “register + immediate.”
— Triadic register addresses: Most computational instructions operate on two register operands or one register and a constant, and place the result in a third register.
— A large windowed register file: At any one instant, a program sees 8 global integer registers plus a 24-register window of a larger register file. The windowed registers can be used
as a cache of procedure arguments, local values, and return addresses.

— Floating-point: The architecture provides an IEEE 754-compatible floating-point instruction set, operating on a separate register file that provides 32 single-precision (32-bit), 32
double-precision (64-bit), 16 quad-precision (128-bit) registers, or a mixture thereof.
— Fast trap handlers: Traps are vectored through a table.
— Multiprocessor synchronization instructions: One instruction performs an atomic readthen-set-memory operation; another performs an atomic exchange-register-with-memory
operation; another compares the contents of a register with a value in memory and
exchanges memory with the contents of another register if the comparison was equal; two
others are used to synchronize the order of shared memory operations as observed by processors.
— Predicted branches: The branch with prediction instructions allow the compiler or assembly language programmer to give the hardware a hint about whether a branch will be
taken.
— Branch elimination instructions: Several instructions can be used to eliminate branches
altogether (e.g., move on condition). Eliminating branches increases performance in
superscalar and superpipelined implementations.
— Hardware trap stack: A hardware trap stack is provided to allow nested traps. It contains
all of the machine state necessary to return to the previous trap level. The trap stack makes
the handling of faults and error conditions simpler, faster, and safer.
— Relaxed memory order (RMO) model: This weak memory model allows the hardware to
schedule memory accesses in almost any order, as long as the program computes the correct result.
1.2.2 Attributes
SPARC-V9 is a CPU instruction set architecture (ISA) derived from SPARC-V8; both architectures come from a reduced instruction set computer (RISC) lineage. As architectures, SPARC-V9
and SPARC-V8 allow for a spectrum of chip and system implementations at a variety of price/
performance points for a range of applications, including scientific/engineering, programming,
real-time, and commercial.
1.2.2.1 Design Goals
SPARC-V9 is designed to be a target for optimizing compilers and high-performance hardware
implementations. SPARC-V9 implementations provide exceptionally high execution rates and
short time-to-market development schedules.
1.2.2.2 Register Windows
SPARC-V9 is derived from SPARC, which was formulated at Sun Microsystems in 1985. SPARC
is based on the RISC I and II designs engineered at the University of California at Berkeley from
1980 through 1982. SPARC’s “register window” architecture, pioneered in the UC Berkeley

designs, allows for straightforward, high-performance compilers and a significant reduction in
memory load/store instructions over other RISCs, particularly for large application programs. For
languages such as C++, where object-oriented programming is dominant, register windows result
in an even greater reduction in instructions executed.
Note that supervisor software, not user programs, manages the register windows. The supervisor
can save a minimum number of registers (approximately 24) during a context switch, thereby
optimizing context-switch latency.
One major difference between SPARC-V9 and the Berkeley RISC I and II is that SPARC-V9 provides greater flexibility to a compiler in its assignment of registers to program variables. SPARCV9 is more flexible because register window management is not tied to procedure call and return
instructions, as it is on the Berkeley machines. Instead, separate instructions (SAVE and
RESTORE) provide register window management. The management of register windows by privileged software is very different too, as discussed in Appendix H, “Software Considerations.”
1.2.3 System Components
The architecture allows for a spectrum of input/output (I/O), memory-management unit (MMU),
and cache system subarchitectures. SPARC-V9 assumes that these elements are best defined by
the specific requirements of particular systems. Note that they are invisible to nearly all user programs, and the interfaces to them can be limited to localized modules in an associated operating
system.
1.2.3.1 Reference MMU
The SPARC-V9 ISA does not mandate a single MMU design for all system implementations.
Rather, designers are free to use the MMU that is most appropriate for their application, or no
MMU at all, if they wish. Appendix F, “SPARC-V9 MMU Requirements,” discusses the boundary conditions that a SPARC-V9 MMU is expected to satisfy.
1.2.3.2 Privileged Software
SPARC-V9 does not assume that all implementations must execute identical privileged software.
Thus, certain traits of an implementation that are visible to privileged software can be tailored to
the requirements of the system. For example, SPARC-V9 allows for implementations with different instruction concurrency and different trap hardware.
1.2.4 Binary Compatibility
The most important SPARC-V9 architectural mandate is binary compatibility of nonprivileged
programs across implementations. Binaries executed in nonprivileged mode should behave identically on all SPARC-V9 systems when those systems are running an operating system known to
provide a standard execution environment. One example of such a standard environment is the
SPARC-V9 Application Binary Interface (ABI).

Although different SPARC-V9 systems may execute nonprivileged programs at different rates,
they will generate the same results, as long as they are run under the same memory model. See
Chapter 8, “Memory Models,” for more information.
Additionally, SPARC-V9 is designed to be binary upward-compatible from SPARC-V8 for applications running in nonprivileged mode that conform to the SPARC-V8 ABI.
1.2.5 Architectural Definition
The SPARC Version 9 Architecture is defined by the chapters and normative appendixes of this
document. A correct implementation of the architecture interprets a program strictly according to
the rules and algorithms specified in the chapters and normative appendixes. Only two classes of
deviations are permitted:
(1) Certain elements of the architecture are defined to be implementation-dependent. These
elements include registers and operations that may vary from implementation to implementation, and are explicitly identified in this document using the notation “IMPL. DEP.
#NN: Some descriptive text.” Appendix C, “SPARC-V9 Implementation Dependencies,”
describes each of these references.
(2) Functional extensions are permitted, insofar as they do not change the behavior of any
defined operation or register. Such extensions are discouraged, since they limit the portability of applications from one implementation to another. Appendix I, “Extending the
SPARC-V9 Architecture,” provides guidelines for incorporating enhancements in an
implementation.
This document defines a nonprivileged subset, designated SPARC-V9-NP. This includes only
those elements that may be executed or accessed while the processor is executing in nonprivileged
mode.
The informative appendixes provide supplementary information such as programming tips,
expected usage, and assembly language syntax. These appendixes are not binding on an implementation or user of a SPARC-V9 system.
The Architecture Committee of SPARC International has sole responsibility for clarification of
the definitions in this document.
1.2.6 SPARC-V9 Compliance
SPARC International is responsible for certifying that implementations comply with the SPARCV9 Architecture. Two levels of compliance are distinguished; an implementation may be certified
at either level.
Level 1:
The implementation correctly interprets all of the nonprivileged instructions by any
method, including direct execution, simulation, or emulation. This level supports user
applications and is the architecture component of the SPARC-V9 ABI.

Level 2:
The implementation correctly interprets both nonprivileged and privileged instructions by
any method, including direct execution, simulation, or emulation. A Level 2 implementation includes all hardware, supporting software, and firmware necessary to provide a complete and correct implementation.
Note that a Level-2-compliant implementation is also Level-1-compliant.
IMPL. DEP. #1: Whether an instruction is implemented directly by hardware, simulated by software, or
emulated by firmware is implementation-dependent.

SPARC International publishes a document, Implementation Characteristics of Current
SPARC-V9-based Products, Revision 9.x, listing which instructions are simulated or emulated in
existing SPARC-V9 implementations.
Compliant implementations shall not add to or deviate from this standard except in aspects
described as implementation-dependent. See Appendix C, “SPARC-V9 Implementation Dependencies.”
An implementation may be claimed to be compliant only if it has been
(1) Submitted to SPARC International for testing, and
(2) Issued a Certificate of Compliance by S. I.
A system incorporating a certified implementation may also claim compliance. A claim of compliance must designate the level of compliance.
Prior to testing, a statement must be submitted for each implementation; this statement must:
— Resolve the implementation dependencies listed in Appendix C
— Identify the presence (but not necessarily the function) of any extensions
— Designate any instructions that require emulation
These statements become the property of SPARC International, and may be released publicly.

2 Definitions
The following subsections define some of the most important words and acronyms used in this
manual
2.1

address space identifier: An eight-bit value that identifies an address space. For each
instruction or data access, the integer unit appends an ASI to the address. See also:
implicit ASI.

2.2

ASI: Abbreviation for address space identifier.

2.3

application program: A program executed with the processor in nonprivileged mode.
Note that statements made in this document regarding application programs may not be
applicable to programs (for example, debuggers) that have access to privileged processor
state (for example, as stored in a memory-image dump).

2.4

big-endian: An addressing convention. Within a multiple-byte integer, the byte with the
smallest address is the most significant; a byte’s significance decreases as its address
increases.

2.5

byte: Eight consecutive bits of data.

2.6

clean window: A register window in which all of the registers contain either zero, a valid
address from the current address space, or valid data from the current address space.

2.7

completed: A memory transaction is said to be completed when an idealized memory has
executed the transaction with respect to all processors. A load is considered completed
when no subsequent memory transaction can affect the value returned by the load. A store
is considered completed when no subsequent load can return the value that was overwritten by the store.

2.8

current window: The block of 24 r registers that is currently in use. The Current Window
Pointer (CWP) register points to the current window.

2.9

dispatch: Issue a fetched instruction to one or more functional units for execution.

2.10

doublet: Two bytes (16 bits) of data.

2.11

doubleword: An aligned octlet. Note that the definition of this term is architecture-dependent and may differ from that used in other processor architectures.

2.12

exception: A condition that makes it impossible for the processor to continue executing
the current instruction stream without software intervention.

2.13

extended word: An aligned octlet, nominally containing integer data. Note that the definition of this term is architecture-dependent and may differ from that used in other processor
architectures.

2.14

f register: A floating-point register. SPARC-V9 includes single- double- and quad- precision f registers.

2.15

fccn: One of the floating-point condition code fields: fcc0, fcc1, fcc2, or fcc3.

2.16

floating-point exception: An exception that occurs during the execution of a floatingpoint operate (FPop) instruction. The exceptions are: unfinished_FPop,
unimplemented_FPop,
sequence_error,
hardware_error,
invalid_fp_register,
and
IEEE_754_exception.

2.17

floating-point IEEE-754 exception: A floating-point exception, as specified by IEEE Std
754-1985. Listed within this manual as IEEE_754_exception.

2.18

floating-point trap type: The specific type of floating-point exception, encoded in the
FSR.ftt field.

2.19

floating-point operate (FPop) instructions: Instructions that perform floating-point calculations, as defined by the FPop1 and FPop2 opcodes. FPop instructions do not include
FBfcc instructions, or loads and stores between memory and the floating-point unit.

2.20

floating-point unit: A processing unit that contains the floating-point registers and performs floating-point operations, as defined by this specification.

2.21

FPU: Abbreviation for floating-point unit.

2.22

halfword: An aligned doublet. Note that the definition of this term is architecture-dependent and may differ from that used in other processor architectures.

2.23

hexlet: Sixteen bytes (128 bits) of data.

2.24

implementation: Hardware and/or software that conforms to all of the specifications of an
ISA.

2.25

implementation-dependent: An aspect of the architecture that may legitimately vary
among implementations. In many cases, the permitted range of variation is specified in the
standard. When a range is specified, compliant implementations shall not deviate from
that range.

2.26

implicit ASI: The address space identifier that is supplied by the hardware on all instruction accesses, and on data accesses that do not contain an explicit ASI or a reference to the
contents of the ASI register.

2.27

informative appendix: An appendix containing information that is useful but not required
to create an implementation that conforms to the SPARC-V9 specification. See also: normative appendix.

2.28

initiated. See issued.

2.29

instruction field: A bit field within an instruction word.

2.30

instruction set architecture (ISA): An ISA defines instructions, registers, instruction and
data memory, the effect of executed instructions on the registers and memory, and an algorithm for controlling instruction execution. An ISA does not define clock cycle times,
cycles per instruction, data paths, etc. This specification defines an ISA.

2.31

integer unit: A processing unit that performs integer and control-flow operations and contains general-purpose integer registers and processor state registers, as defined by this
specification.

2.32

interrupt request: A request for service presented to the processor by an external device.

2.33

IU: Abbreviation for integer unit.

2.34

ISA: Abbreviation for instruction set architecture.

2.35

issued: In reference to memory transaction, a load, store, or atomic load-store is said to be
issued when a processor has sent the transaction to the memory subsystem and the completion of the request is out of the processor’s control. Synonym: initiated.

2.36

leaf procedure: A procedure that is a leaf in the program’s call graph; that is, one that
does not call (using CALL or JMPL) any other procedures.

2.37

little-endian: An addressing convention. Within a multiple-byte integer, the byte with the
smallest address is the least significant; a byte’s significance increases as its address
increases.

2.38

may: A key word indicating flexibility of choice with no implied preference. Note: “may”
indicates that an action or operation is allowed, “can” indicates that it is possible.

2.39

must: Synonym: shall.

2.40

next program counter (nPC): A register that contains the address of the instruction to be
executed next, if a trap does not occur.

2.41

non-faulting load: A load operation that will either complete correctly (in the absence of
any faults) or will return a value (nominally zero) if a fault occurs. See speculative load.

2.42

nonprivileged: An adjective that describes (1) the state of the processor when
PSTATE.PRIV = 0, that is, nonprivileged mode; (2) processor state information that is
accessible to software while the processor is in either privileged mode or nonprivileged
mode, for example, nonprivileged registers, nonprivileged ASRs, or, in general, nonprivileged state; (3) an instruction that can be executed when the processor is in either privileged mode or nonprivileged mode.

2.43

nonprivileged mode: The processor mode when PSTATE.PRIV = 0. See also: nonprivileged.

2.44

normative appendix: An appendix containing specifications that must be met by an
implementation conforming to the SPARC-V9 specification. See also: informative
appendix.

2.45

NWINDOWS: The number of register windows present in an implementation.

2.46

octlet: Eight bytes (64 bits) of data. Not to be confused with an “octet,” which has been
commonly used to describe eight bits of data. In this document, the term byte, rather than
octet, is used to describe eight bits of data.

2.47

opcode: A bit pattern that identifies a particular instruction.

2.48

prefetchable: An attribute of a memory location which indicates to an MMU that
PREFETCH operations to that location may be applied. Normal memory is prefetchable.
Nonprefetchable locations include those that, when read, change state or cause external
events to occur. See also: side effect.

2.49

privileged: An adjective that describes (1) the state of the processor when PSTATE.PRIV
= 1, that is , privileged mode; (2) processor state information that is accessible to software only while the processor is in privileged mode, for example, privileged registers,
privileged ASRs, or, in general, privileged state; (3) an instruction that can be executed
only when the processor is in privileged mode.

2.50

privileged mode: The processor mode when PSTATE.PRIV = 1. See also: nonprivileged.

2.51

processor: The combination of the integer unit and the floating-point unit.

2.52

program counter (PC): A register that contains the address of the instruction currently
being executed by the IU.

2.53

quadlet: Four bytes (32 bits) of data.

2.54

quadword: An aligned hexlet. Note that the definition of this term is architecture-dependent and may be different from that used in other processor architectures.

2.55

r register: An integer register. Also called a general purpose register or working register.

2.56

RED_state: Reset, Error, and Debug state. The processor state when PSTATE.RED = 1.
A restricted execution environment used to process resets and traps that occur when
TL = MAXTL – 1.

2.57

reserved: Used to describe an instruction field, certain bit combinations within an instruction field, or a register field that is reserved for definition by future versions of the architecture. Reserved instruction fields shall read as zero, unless the implementation
supports extended instructions within the field. The behavior of SPARC-V9-compliant
processors when they encounter non-zero values in reserved instruction fields is undefined. Reserved bit combinations within instruction fields are defined in Appendix A;
in all cases, SPARC-V9-compliant processors shall decode and trap on these reserved

combinations. Reserved register fields should always be written by software with values
of those fields previously read from that register, or with zeroes; they should read as zero
in hardware. Software intended to run on future versions of SPARC-V9 should not assume
that these field will read as zero or any other particular value. Throughout this manual, figures and tables illustrating registers and instruction encodings indicate reserved fields and
combinations with an em dash ‘—’.
2.58

reset trap: A vectored transfer of control to privileged software through a fixed-address
reset trap table. Reset traps cause entry into RED_state.

2.59

restricted: An adjective used to describe an address space identifier (ASI) that may be
accessed only while the processor is operating in privileged mode.

2.60

rs1, rs2, rd: The integer register operands of an instruction, where rs1 and rs2 are the
source registers and rd is the destination register.

2.61

shall: A key word indicating a mandatory requirement. Designers shall implement all such
mandatory requirements to ensure interoperability with other SPARC-V9-conformant
products. Synonym: must.

2.62

should: A key word indicating flexibility of choice with a strongly preferred implementation. Synonym: it is recommended.

2.63

side effect: An operation has a side effect if it induces a secondary effect as well as its primary effect. For example, access to an I/O location may cause a register value in an I/O
device to change state or initiate an I/O operation. A memory location is deemed to have
side effects if additional actions beyond the reading or writing of data may occur when a
memory operation on that location is allowed to succeed. See also: prefetchable.

2.64

speculative load: A load operation that is issued by the processor speculatively, that is,
before it is known whether the load will be executed in the flow of the program. Speculative accesses are used by hardware to speed program execution and are transparent to
code. Contrast with non-faulting load, which is an explict load that always completes,
even in the presence of faults. Warning: some authors confuse speculative loads with nonfaulting loads.

2.65

supervisor software: Software that executes when the processor is in privileged mode.

2.66

trap: The action taken by the processor when it changes the instruction flow in response to
the presence of an exception, a Tcc instruction, or an interrupt. The action is a vectored
transfer of control to supervisor software through a table, the address of which is specified by the privileged Trap Base Address (TBA) register.

2.67

unassigned: A value (for example, an address space identifier), the semantics of which
are not architecturally mandated and may be determined independently by each implementation within any guidelines given.

2.68

undefined: An aspect of the architecture that has deliberately been left unspecified. Software should have no expectation of, nor make any assumptions about, an undefined feature or behavior. Use of such a feature may deliver random results, may or may not cause
a trap, may vary among implementations, and may vary with time on a given implementation. Notwithstanding any of the above, undefined aspects of the architecture shall not
cause security holes such as allowing user software to access privileged state, put the processor into supervisor mode, or put the processor into an unrecoverable state.

2.69

unrestricted: An adjective used to describe an address space identifier that may be used
regardless of the processor mode, that is, regardless of the value of PSTATE.PRIV.

2.70

user application program: Synonym: application program.

2.71

word: An aligned quadlet. Note that the definition of this term is architecture-dependent
and may differ from that used in other processor architectures.

3 Architectural Overview
SPARC-V9 is an instruction set architecture (ISA) with 32- and 64-bit integer and 32-, 64- and
128-bit floating-point as its principal data types. The 32- and 64- bit floating point types conforms
to IEEE Std 754-1985. The 128-bit floating-point type conforms to IEEE Std 1596.5-1992.
SPARC-V9 defines general-purpose integer, floating-point, and special state/status register
instructions, all encoded in 32-bit-wide instruction formats. The load/store instructions address a
linear, 264-byte address space.

3.1 SPARC-V9 Processor
A SPARC-V9 processor logically consists of an integer unit (IU) and a floating-point unit (FPU),
each with its own registers. This organization allows for implementations with concurrency
between integer and floating-point instruction execution. Integer registers are 64 bits wide; floating-point registers are 32, 64, or 128 bits wide. Instruction operands are single registers, register
pairs, register quadruples, or immediate constants.
The processor can be in either of two modes: privileged or nonprivileged. In privileged mode,
the processor can execute any instruction, including privileged instructions. In nonprivileged
mode, an attempt to execute a privileged instruction causes a trap to privileged software.
3.1.1 Integer Unit (IU)
The integer unit contains the general-purpose registers and controls the overall operation of the
processor. The IU executes the integer arithmetic instructions and computes memory addresses
for loads and stores. It also maintains the program counters and controls instruction execution for
the FPU.

IMPL. DEP. #2: An implementation of the IU may contain from 64 to 528 general-purpose 64-bit r registers.
this corresponds to a grouping of the registers into 8 global r registers, 8 alternate global r registers, plus a
circular stack of from 3 to 32 sets of 16 registers each, known as register windows. Since the number of
register windows present (NWINDOWS) is implementation-dependent, the total number of registers is
implementation-dependent.

At a given time, an instruction can access the 8 globals (or the 8 alternate globals) and a register
window into the r registers. The 24-register window consists of a 16-register set — divided into 8
in and 8 local registers — together with the 8 in registers of an adjacent register set, addressable
from the current window as its out registers. See figure 2 on page 32.
The current window is specified by the current window pointer (CWP) register. The processor
detects window spill and fill exceptions via the CANSAVE and CANRESTORE registers, respectively, which are controlled by hardware and supervisor software. The actual number of windows
in a SPARC-V9 implementation is invisible to a user application program.
Whenever the IU accesses an instruction or datum in memory, it appends an address space identifier (ASI), to the address. All instruction accesses and most data accesses append an implict
ASI, but some instructions allow the inclusion of an explict ASI, either as an immediate field
within the instruction, or from the ASI register. The ASI determines the byte order of the access.
All instructions are accessed in big-endian byte order; data can be referenced in either big- or little-endian order. See 5.2.1, “Processor State Register (PSTATE),” for information about changing
the default byte order.

3.1.2 Floating-Point Unit (FPU)
The FPU has 32 32-bit (single-precision) floating-point registers, 32 64-bit (double-precision)
floating-point registers, and 16 128-bit (quad-precision) floating-point registers, some of which
overlap. Double-precision values occupy an even-odd pair of single-precision registers, and quadprecision values occupy a quad-aligned group of four single-precision registers. The 32 singleprecision registers, the lower half of the double-precision registers, and the lower half of the quadprecision registers overlay each other. The upper half of the double-precision registers and the
upper half of the quad-precision registers overlay each other, but do not overlay any of the singleprecision registers. Thus, the floating-point registers can hold a maximum of 32 single-precision,
32 double-precision, or 16 quad-precision values. This is described in more detail in 5.1.4, “Floating-Point Registers.”
Floating-point load/store instructions are used to move data between the FPU and memory. The
memory address is calculated by the IU. Floating-Point operate (FPop) instructions perform the
floating-point arithmetic operations and comparisons.
The floating-point instruction set and 32- and 64-bit data formats conform to the IEEE Standard
for Binary Floating-Point Arithmetic, IEEE Std 754-1985. The 128-bit floating-point data type
conforms to the IEEE Standard for Shared Data Formats, IEEE Std 1596.5-1992.
If an FPU is not present or is not enabled, an attempt to execute a floating-point instruction generates an fp_disabled trap. In either case, privileged-mode software must:

— Enable the FPU and reexecute the trapping instruction, or
— Emulate the trapping instruction.

3.2 Instructions
Instructions fall into the following basic categories:
— Memory access
— Integer operate
— Control transfer
— State register access
— Floating-point operate
— Conditional move
— Register window management
These classes are discussed in the following subsections.
3.2.1 Memory Access
Load and store instructions and the atomic operations, CASX, SWAP, and LDSTUB, are the only
instructions that access memory. They use two r registers or an r register and a signed 13-bit
immediate value to calculate a 64-bit, byte-aligned memory address. The IU appends an ASI to
this address.
The destination field of the load/store instruction specifies either one or two r registers, or one,
two, or four f registers, that supply the data for a store or receive the data from a load.
Integer load and store instructions support byte, halfword (16-bit), word (32-bit), and doubleword
(64-bit) accesses. Some versions of integer load instructions perform sign extension on 8-, 16-,
and 32-bit values as they are loaded into a 64-bit destination register. Floating-point load and store
instructions support word, doubleword, and quadword memory accesses.
CAS, SWAP, and LDSTUB are special atomic memory access instructions that are used for synchonization and memory updates by concurrent processes.

3.2.1.1 Memory Alignment Restrictions
Halfword accesses shall be aligned on 2-byte boundaries; word accesses (which include instruction fetches) shall be aligned on 4-byte boundaries; extended-word and doubleword accesses shall
be aligned on 8-byte boundaries; and quadword quantities shall be aligned on 16-byte boundaries.
An improperly aligned address in a load, store, or load-store instruction causes a trap to occur,
with the possible exception of cases described in 6.3.1.1, “Memory Alignment Restrictions.”

3.2.1.2 Addressing Conventions
SPARC-V9 uses big-endian byte order by default: the address of a quadword, doubleword, word,
or halfword is the address of its most significant byte. Increasing the address means decreasing the
significance of the unit being accessed. All instruction accesses are performed using big-endian
byte order. SPARC-V9 also can support little-endian byte order for data accesses only: the address
of a quadword, doubleword, word, or halfword is the address of its least significant byte. Increasing the address means increasing the significance of the unit being accessed. See 5.2.1, Processor
State Register (PSTATE), for information about changing the implicit byte order to little-endian.
Addressing conventions are illustrated in figure 35 on page 71 and figure 36 on page 73.

3.2.1.3 Load/Store Alternate
Versions of load/store instructions, the load/store alternate instructions, can specify an arbitrary
8-bit address space identifier for the load/store data access. Access to alternate spaces 0016 ..7F16
is restricted, and access to alternate spaces 8016 ..FF16 is unrestricted. Some of the ASIs are available for implementation-dependent uses (impl. dep. #29). Supervisor software can use the implementation-dependent ASIs to access special protected registers, such as MMU, cache control, and
processor state registers, and other processor- or system-dependent values. See 6.3.1.3, “Address
Space Identifiers (ASIs),” for more information.
Alternate space addressing is also provided for the atomic memory access instructions, LDSTUB,
SWAP, and CASX.

3.2.1.4 Separate I and D Memories
Most of the specifications in this manual ignore the issues of memory mapping and caching. The
interpretation of addresses can be unified, in which case the same translations and caching are
applied to both instructions and data, or they can be split, in which case instruction references use
one translation mechanism and cache and data references another, although the same main memory is shared. In such split-memory systems, the coherency mechanism may be unified and
include both instructions and data, or it may be split. For this reason, programs that modify their
own code (self-modifying code) must issue FLUSH instructions, or a system call with a similar
effect, to bring the instruction and data memories into a consistent state.

3.2.1.5 Input/Output
SPARC-V9 assumes that input/output registers are accessed via load/store alternate instructions,
normal load/store instructions, or read/write Ancillary State Register instructions (RDASR,
WRASR).
IMPL. DEP. #123: The semantic effect of accessing input/output (I/O) locations is implementation-dependent.

IMPL. DEP. #6: Whether the I/O registers can be accessed by nonprovileged code is implementationdependent.
IMPL. DEP. #7: The addresses and contents of I/O registers are implementation-dependent.

3.2.1.6 Memory Synchronization
Two instructions are used for synchronization of memory operations: FLUSH and MEMBAR.
Their operation is explained in A.20, “Flush Instruction Memory,” and A.32, “Memory Barrier,”
respectively.

3.2.2 Arithmetic/Logical/Shift Instructions
The arithmetic/logical/shift instructions perform arithmetic, tagged arithmetic, logical, and shift
operations. With one exception, these instructions compute a result that is a function of two
source operands; the result is either written into a destination register or discarded. The exception,
SETHI, may be used in combination with another arithmentic or logical instruction to create a 32bit constant in an r register.
Shift instructions are used to shift the contents of an r register left or right by a given count. The
shift distance is specified by a constant in the instruction or by the contents of an r register.
The integer multiply instruction performs a 64 × 64 → 64-bit operation. The integer division
instructions perform 64 ÷ 64 → 64-bit operations. In addition, for compatibility with SPARC-V8,
32 × 32 → 64-bit multiply, 64 ÷ 32 → 32-bit divide, and multiply step instructions are included.
Division by zero causes a trap. Some versions of the 32-bit multiply and divide instructions set the
condition codes.
The tagged arithmetic instructions assume that the least-significant two bits of each operand are a
data-type tag. The nontrapping versions of these instructions set the integer condition code (icc)
and extended integer condition code (xcc) overflow bits on 32-bit (icc) or 64-bit (xcc) arithmetic
overflow. In addition, if any of the operands’ tag bits are nonzero, icc is set. The xcc overflow bit
is not affected by the tag bits.

3.2.3 Control Transfer
Control-transfer instructions (CTIs) include PC-relative branches and calls, register-indirect
jumps, and conditional traps. Most of the control-transfer instructions are delayed; that is, the
instruction immediately following a control-transfer instruction in logical sequence is dispatched
before the control transfer to the target address is completed. Note that the next instruction in logical sequence may not be the instruction following the control-transfer instruction in memory.
The instruction following a delayed control-transfer instruction is called a delay instruction. A bit
in a delayed control-transfer instruction (the annul bit) can cause the delay instruction to be
annulled (that is, to have no effect) if the branch is not taken (or in the “branch always” case, if the
branch is taken).

Compatibility Note:
SPARC-V8 specified that the delay instruction was always fetched, even if annulled, and that an annulled
instruction could not cause any traps. SPARC-V9 does not require the delay instruction to be fetched if it is
annulled.

Branch and CALL instructions use PC-relative displacements. The jump and link (JMPL) and
return (RETURN) instructions use a register-indirect target address. They compute their target
addresses as either the sum of two r registers, or the sum of an r register and a 13-bit signed
immediate value. The branch on condition codes without prediction instruction provides a displacement of ±8 Mbytes; the branch on condition codes with prediction instruction provides a displacement of ±1 Mbyte; the branch on register contents instruction provides a displacement of
±128 Kbytes, and the CALL instruction’s 30-bit word displacement allows a control transfer to
any address within ±2 gigabytes (±231 bytes). Note that when 32-bit address masking is enabled
(see 5.2.1.7, “PSTATE_address_mask (AM)”), the CALL instruction may transfer control to an
arbitrary 32-bit address. The return from privileged trap instructions (DONE and RETRY) get
their target address from the appropriate TPC or TNPC register.

3.2.4 State Register Access
The read and write state register instructions read and write the contents of state registers visible
to nonprivileged software (Y, CCR, ASI, PC, TICK, and FPRS). The read and write privileged
register instructions read and write the contents of state registers visible only to privileged software (TPC, TNPC, TSTATE, TT, TICK, TBA, PSTATE, TL, PIL, CWP, CANSAVE, CANRESTORE, CLEANWIN, OTHERWIN, WSTATE, FPQ, and VER).
IMPL. DEP. #8: Software can use read/write ancillary state register instructions to read/write implementation-dependent processor registers (ASRs 16..31).
IMPL. DEP. #9: Which if any of the implementation-dependent read/write ancillary state register instructions (for ASRS 16..31) is privileged is implementation-dependent.

3.2.5 Floating-Point Operate
Floating-point operate (FPop) instructions perform all floating-point calculations; they are register-to-register instructions that operate on the floating-point registers. Like arithmetic/logical/shift
instructions, FPops compute a result that is a function of one or two source operands. Specific
floating-point operations are selected by a subfield of the FPop1/FPop2 instruction formats.

3.2.6 Conditional Move
Conditional move instructions conditionally copy a value from a source register to a destination
register, depending on an integer or floating-point condition code or upon the contents of an integer register. These instructions increase performance by reducing the number of branches.

3.2.7 Register Window Management
These instructions are used to manage the register windows. SAVE and RESTORE are nonprivileged and cause a register window to be pushed or popped. FLUSHW is nonprivileged and causes
all of the windows except the current one to be flushed to memory. SAVED and RESTORED are
used by privileged software to end a window spill or fill trap handler.

3.3 Traps
A trap is a vectored transfer of control to privileged software through a trap table that may contain the first eight instructions (thirty-two for fill/spill traps) of each trap handler. The base address
of the table is established by software in a state register (the Trap Base Address register, TBA).
The displacement within the table is encoded in the type number of each trap and the level of the
trap. One half of the table is reserved for hardware traps; one quarter is reserved for software traps
generated by trap (Tcc) instructions; the final quarter is reserved for future expansion of the architecture.
A trap causes the current PC and nPC to be saved in the TPC and TNPC registers. It also causes
the CCR, ASI, PSTATE, and CWP registers to be saved in TSTATE. TPC, TNPC, and TSTATE
are entries in a hardware trap stack, where the number of entries in the trap stack is equal to the
number of trap levels supported (impl. dep. #101). A trap also sets bits in the PSTATE register,
one of which can enable an alternate set of global registers for use by the trap handler. Normally,
the CWP is not changed by a trap; on a window spill or fill trap, however, the CWP is changed to
point to the register window to be saved or restored.
A trap may be caused by a Tcc instruction, an asynchronous exception, an instruction-induced
exception, or an interrupt request not directly related to a particular instruction. Before executing each instruction, the processor behaves as though it determines if there are any pending exceptions or interrupt requests. If any are pending, the processor selects the highest-priority exception
or interrupt request and causes a trap.
See Chapter 7, “Traps,” for a complete description of traps.

4 Data Formats
The SPARC-V9 architecture recognizes these fundamental data types:
— Signed Integer: 8, 16, 32, and 64 bits
— Unsigned Integer: 8, 16, 32, and 64 bits
— Floating Point: 32, 64, and 128 bits
The widths of the data types are:
— Byte: 8 bits
— Halfword: 16 bits

— Word: 32 bits
— Extended Word: 64 bits
— Tagged Word: 32 bits (30-bit value plus 2-bit tag)
— Doubleword: 64 bits
— Quadword: 128 bits
The signed integer values are stored as two’s-complement numbers with a width commensurate
with their range. Unsigned integer values, bit strings, boolean values, strings, and other values
representable in binary form are stored as unsigned integers with a width commensurate with their
range. The floating-point formats conform to the IEEE Standard for Binary Floating-Point Arithmetic, IEEE Std 754-1985. In tagged words, the least significant two bits are treated as a tag; the
remaining 30 bits are treated as a signed integer.
Subsections 4.1 through 4.11 illustrate the signed integer, unsigned integer, and tagged formats.
Subsections 4.12 through 4.14 illustrate the floating-point formats. In 4.4, 4.9, 4.13, and 4.14, the
individual subwords of the multiword data formats are assigned names. The arrangement of the
subformats in memory and processor registers based on these names is shown in table 1. Tables 2
through 5 define the integer and floating-point formats.

4.1 Signed Integer Byte
S
7 6

4.2 Signed Integer Halfword
S
15 14

4.3 Signed Integer Word
S
31 30

4.4 Signed Integer Double
SD–0
signed_dbl_integer[62:32]

S
31 30

SD–1
signed_dbl_integer[31:0]
31

4.5 Signed Extended Integer
SX
signed_ext_integer

S
63 62

4.6 Unsigned Integer Byte

4.7 Unsigned Integer Halfword

4.8 Unsigned Integer Word

4.9 Unsigned Integer Double
UD–0
unsigned_dbl_integer[63:32]
31

UD–1
unsigned_dbl_integer[31:0]
31

4.10 Unsigned Extended Integer
UX
unsigned_ext_integer
63

4.11 Tagged Word
tag
2 1

4.12 Floating-Point Single Precision
S

exp[7:0]

31 30

fraction[22:0]
23 22

4.13 Floating-Point Double Precision
FD–0

S
31 30

FD–1

exp[10:0]

fraction[51:32]
20 19

fraction[31:0]
31

4.14 Floating-Point Quad Precision
FQ–0

exp[14:0]

31 30

fraction[111:96]
16 15

FQ–1

fraction[95:64]
31

FQ–2

fraction[63:32]
31

FQ–3

fraction[31:0]
31

Table 1—Double- and Quadwords in Memory & Registers
Subformat
Name
SD-0
SD-1
SX
UD-0
UD-1
UX
FD-0
FD-1

Subformat Field
signed_dbl_integer[63:32]
signed_dbl_integer[31:0]
signed_ext_integer[63:0]
unsigned_dbl_integer[63:32]
unsigned_dbl_integer[31:0]
unsigned_ext_integer[63:0]
s:exp[10:0]:fraction[51:32]
fraction[31:0]

Required
Address
Alignment
0 mod 8
4 mod 8
0 mod 8
0 mod 8
4 mod 8
0 mod 8
0 mod 4 †

Memory
Address

n
n+4
n
n
n+4
n

0 mod 2
1 mod 2
—
0 mod 2
1 mod 2
—

r
r+1
r
r
r+1
r

0 mod 2

0 mod 4

†

n+4

1 mod 2

f+1

‡

FQ-0

s:exp[14:0]:fraction[111:96]

0 mod 4

FQ-1

fraction[95:64]

0 mod 4 ‡

n+4

1 mod 4

f+1

fraction[63:32]

0 mod 4

‡

n+8

2 mod 4

f+2

0 mod 4

‡

n + 12

3 mod 4

f+3

FQ-2
FQ-3

fraction[31:0]

†

Although a floating-point doubleword is only required to be word-aligned in memory, it is recommended
that it be doubleword-aligned (i.e., the address of its FD-0 word should be 0 mod 8).

‡

Although a floating-point quadword is only required to be word-aligned in memory, it is recommended that
it be quadword-aligned (i.e., the address of its FQ-0 word should be 0 mod 16).

Table 2—Signed Integer, Unsigned Integer, and Tagged Format Ranges
Data type

Width (bits)

Signed integer byte
Signed integer halfword
Signed integer word
Signed integer tagged word
Signed integer double
Signed extended integer

8
16
32
32
64
64

Unsigned integer byte

Unsigned integer halfword
Unsigned integer word
Unsigned integer tagged word
Unsigned integer double
Unsigned extended integer

16
32
32
64
64

Range

−27

to 27 − 1
−215 to 215 − 1
−231 to 231 − 1
−229 to 229 − 1
−263 to 263 − 1

−263 to 263 − 1
0 to 28 − 1
0 to 216 − 1
0 to 232 − 1
0 to 230 − 1
0 to 264 − 1
0 to 264 − 1

Table 3—Floating-Point Single-Precision Format Definition
s
e
f
u

= sign (1 bit)
= biased exponent (8 bits)
= fraction (23 bits)
= undefined

Normalized value (0 < e < 255):
Subnormal value (e = 0):
Zero (e = 0)
Signalling NaN

(−1)s × 2e−127 × 1.f

(−1)s × 2−126 × 0.f
(−1)s × 0
s = u; e = 255 (max); f = .0uu--uu
(At least one bit of the fraction must be nonzero)

Quiet NaN
− ∞ (negative infinity)
+ ∞ (positive infinity)

s = u; e = 255 (max); f = .1uu--uu
s = 1; e = 255 (max); f = .000--00
s = 0; e = 255 (max); f = .000--00

Table 4—Floating-Point Double-Precision Format Definition
s
e
f
u

= sign (1 bit)
= biased exponent (11 bits)
= fraction (52 bits)
= undefined

Normalized value (0 < e < 2047):
Subnormal value (e = 0):
Zero (e = 0)
Signalling NaN

(−1)s × 2e−1023 × 1.f
(−1)s × 2−1022 × 0.f
(−1)s × 0
s = u; e = 2047 (max); f = .0uu--uu
(At least one bit of the fraction must be nonzero)

Quiet NaN
− ∞ (negative infinity)
+ ∞ (positive infinity)

s = u; e = 2047 (max); f = .1uu--uu
s = 1; e = 2047 (max); f = .000--00
s = 0; e = 2047 (max); f = .000--00

Table 5—Floating-Point Quad-Precision Format Definition
s
e
f
u

= sign (1 bit)
= biased exponent (15 bits)
= fraction (112 bits)
= undefined

Normalized value (0 < e < 32767):
Subnormal value (e = 0):
Zero (e = 0)
Signalling NaN

(-1)s × 2e−16383 × 1.f
(-1)s × 2−16382 × 0.f
(-1)s × 0
s = u; e = 32767 (max); f = .0uu--uu
(At least one bit of the fraction must be nonzero)

Quiet NaN
− ∞ (negative infinity)
+ ∞ (positive infinity)

s = u; e = 32767 (max); f = .1uu--uu
s = 1; e = 32767 (max); f = .000--00
s = 0; e = 32767 (max); f = .000--00

5 Registers
A SPARC-V9 processor includes two types of registers: general-purpose, or working data registers, and control/status registers.
Working registers include:
— Integer working registers (r registers)
— Floating-point working registers (f registers)
Control/status registers include:
— Program Counter register (PC)
— Next Program Counter register (nPC)

— Processor State register (PSTATE)
— Trap Base Address register (TBA)
— Y register (Y)
— Processor Interrupt Level register (PIL)
— Current Window Pointer register (CWP)
— Trap Type register (TT)
— Condition Codes Register (CCR)
— Address Space Identifier register (ASI)
— Trap Level register (TL)
— Trap Program Counter register (TPC)
— Trap Next Program Counter register (TNPC)
— Trap State register (TSTATE)
— Hardware clock-tick counter register (TICK)
— Savable windows register (CANSAVE)
— Restorable windows register (CANRESTORE)
— Other windows register (OTHERWIN)
— Clean windows register (CLEANWIN)
— Window State register (WSTATE)
— Version register (VER)
— Implementation-dependent Ancillary State Registers (ASRs) (impl. dep. #8)
— Implementation-dependent IU Deferred-Trap Queue (impl. dep. #16)
— Floating-Point State Register (FSR)
— Floating-Point Registers State register (FPRS)
— Implementation-dependent Floating-Point Deferred-Trap Queue (FQ) (impl. dep. #24)
For convenience, some registers in this chapter are illustrated as fewer than 64 bits wide. Any bits
not shown are reserved for future extensions to the architecture. Such reserved bits read as zeroes
and, when written by software, should always be written with the values of those bits previously
read from that register, or with zeroes.

5.1 Nonprivileged Registers
The registers described in this subsection are visible to nonprivileged (application, or “usermode”) software.
5.1.1 General Purpose r Registers
At any moment, general-purpose registers appear to nonprivileged software as shown in figure 1.
An implementation of the IU may contain from 64 to 528 general-purpose 64-bit r registers. They
are partitioned into 8 global registers, 8 alternate global registers, plus an implementation-dependent number of 16-register sets (impl. dep. #2). A register window consists of the current 8 in registers, 8 local registers, and 8 out registers. See table 6.
5.1.1.1 Global r Registers
Registers r[0]..r[7] refer to a set of eight registers called the global registers (g0..g7). At any
time, one of two sets of eight registers is enabled and can be accessed as the global registers.
Which set of globals is currently enabled is selected by the AG (alternate global) field in the
PSTATE register. See 5.2.1, “Processor State Register (PSTATE),” for a description of the AG
field.
Global register zero (g0) always reads as zero; writes to it have no program-visible effect.
Compatibility Note:
Since the PSTATE register is only writable by privileged software, existing nonprivileged SPARC-V8 software will operate correctly on a SPARC-V9 implementation if supervisor software ensures that nonprivileged software sees a consistent set of global registers.

r [31]

r [30]

r [29]

r [28]

r [27]

r [26]

r [25]

r [24]

r [23]

r [22]

r [21]

r [20]

r [19]

r [18]

r [17]

r [16]

r [15]

r [14]

r [13]

r [12]

r [11]

r [10]

r [9]

r [8]

r [7]

r [6]

r [5]

r [4]

r [3]

r [2]

r [1]

r [0]

Figure 1—General-Purpose Registers (Nonprivileged View)

Programming Note:
The alternate global registers are present to give trap handlers a set of scratch registers that are independent
of nonprivileged software’s registers. The AG bit in PSTATE allows supervisor software to access the normal global registers if required (for example, during instruction emulation).

5.1.1.2 Windowed r Registers
At any time, an instruction can access the 8 globals and a 24-register window into the r registers.
A register window comprises the 8 in and 8 local registers of a particular register set, together

with the 8 in registers of an adjacent register set, which are addressable from the current window
as out registers. See figure 2 and table 6.
Window (CWP – 1)
r [31]

.
.

ins

r [24]
r [23]

.
.

locals

r [16]

Window (CWP )
r [31]

r [15]

.
.
r [ 8]

.
.

outs

ins

r [24]
r [23]

.
.

locals

r [16]

Window (CWP + 1)
r [31]

r [15]

.
.

outs

ins

r [24]

r [ 8]

r [23]

.
.

locals

r [16]
r [15]

.
.

outs

r [ 8]

r[ 7]
.
.

globals

r[ 1]
r[ 0]
63

0
0

Figure 2—Three Overlapping Windows and the Eight Global Registers

The number of windows or register sets, NWINDOWS, is implementation-dependent and ranges
from 3 to 32 (impl. dep. #2). The total number of r registers in a given implementation is 8 (for the

globals), plus 8 (for the alternate globals), plus the number of sets times 16 registers/set. Thus, the
minimum number of r registers is 64 (3 sets plus the 16 globals and alternate globals) and the
maximum number is 528 (32 sets plus the 16 globals and alternate globals).
Table 6—Window Addressing
Windowed Register Address
in[0] – in[7]
local[0] – local[7]
out[0] – out[7]
global[0] – global[7]

r Register Address
r[24] – r[31]
r[16] – r[23]
r[ 8] – r[15]
r[ 0] – r[ 7]

The current window into the r registers is given by the current window pointer (CWP) register.
The CWP is decremented by the RESTORE instruction and incremented by the SAVE instruction.
Window overflow is detected via the CANSAVE register and window underflow is detected via
the CANRESTORE register, both of which are controlled by privileged software. A window overflow (underflow) condition causes a window spill (fill) trap.
5.1.1.3 Overlapping Windows
Each window shares its ins with one adjacent window and its outs with another. The outs of the
CWP–1 (modulo NWINDOWS) window are addressable as the ins of the current window, and the
outs in the current window are the ins of the CWP+1 (modulo NWINDOWS) window. The locals
are unique to each window.
An r register with address o, where 8 ≤ o ≤ 15, refers to exactly the same register as (o+16) does
after the CWP is incremented by 1 (modulo NWINDOWS). Likewise, a register with address i,
where 24 ≤ i ≤ 31, refers to exactly the same register as address (i−16) does after the CWP is decremented by 1 (modulo NWINDOWS). See figures 2 and 3.
Since CWP arithmetic is performed modulo NWINDOWS, the highest numbered implemented
window overlaps with window 0. The outs of window NWINDOWS−1 are the ins of window 0.
Implemented windows must be numbered contiguously from 0 through NWINDOWS−1.
Programming Note:
Since the procedure call instructions (CALL and JMPL) do not change the CWP, a procedure can be called
without changing the window. See H.1.2, “Leaf-Procedure Optimization.”
Because the windows overlap, the number of windows available to software is one less than the number of
implemented windows, or NWINDOWS−1. When the register file is full, the outs of the newest window are
the ins of the oldest window, which still contains valid data.
The local and out registers of a register window are guaranteed to contain either zeroes or an old value that
belongs to the current context upon reentering the window through a SAVE instruction. If a program executes a RESTORE followed by a SAVE, the resulting window’s locals and outs may not be valid after the
SAVE, since a trap may have occurred between the RESTORE and the SAVE. However, if the clean_window
protocol is being used, system software must guarantee that registers in the current window after a SAVE
will always contain only zeroes or valid data from that context. See 5.2.10.6, “Clean Windows (CLEANWIN) Register.”

Subsection 6.4, “Register Window Management,” describes how the windowed integer registers
are managed.
CWP = 0
(current window pointer)

w1 ins
w7 outs

w1 locals
w0 locals

CANRESTORE = 1

w0 outs

w0 ins
w1 outs

w7 locals
w2 ins

CANSAVE = 3

w7 ins
w2 locals

w6 outs
RESTORE

SAVE

w6 locals

w2 outs

w3 ins

w6 ins
w5 outs

w3 locals
w4 ins
w4 outs
w5 locals

w4 locals
w3 outs

OTHERWIN = 2

w5 ins
(Overlap)

CANSAVE + CANRESTORE + OTHERWIN = NWINDOWS – 2
The current window (window 0) and the overlap window (window 4) account for the two windows
in the right-hand side of the equation. The “overlap window” is the window that must remain
unused because its ins and outs overlap two other valid windows.

Figure 3—The Windowed r Registers for NWINDOWS = 8

5.1.2 Special r Registers
The usage of two of the r registers is fixed, in whole or in part, by the architecture:
— The value of r[0] is always zero; writes to it have no program-visible effect.

— The CALL instruction writes its own address into register r[15] (out register 7).
5.1.2.1 Register-Pair Operands
LDD, LDDA, STD, and STDA instructions access a pair of words in adjacent r registers and
require even-odd register alignment. The least-significant bit of an r register number in these
instructions is reserved, and should be supplied as zero by software.
When the r[0] – r[1] register pair is used as a destination in LDD or LDDA, only r[1] is modified.
When the r[0] – r[1] register pair is used as a source in STD or STDA, a zero is written to the 32bit word at the lowest address and the least significant 32 bits of r[1] are written to the 32-bit word
at the highest address (in big-endian mode).
An attempt to execute an LDD, LDDA, STD, or STDA instruction that refers to a misaligned
(odd) destination register number causes an illegal_instruction trap.
5.1.2.2 Register Usage
See H.1.1, “Registers,” for information about the conventional usage of the r registers.
In figure 3, NWINDOWS = 8. The 8 globals are not illustrated. CWP = 0, CANSAVE = 3,
OTHERWIN = 2, and CANRESTORE = 1. If the procedure using window w0 executes a
RESTORE, window w7 becomes the current window. If the procedure using window w0 executes
a SAVE, window w1 becomes the current window.
5.1.3 IU Control/Status Registers
The nonprivileged IU control/status registers include the program counters (PC and nPC), the 32bit multiply/divide (Y) register (and possibly optional) implementation-dependent Ancillary State
Registers (ASRs) (impl. dep. #8).
5.1.3.1 Program Counters (PC, nPC)
The PC contains the address of the instruction currently being executed by the IU. The nPC holds
the address of the next instruction to be executed, if a trap does not occur. The low-order two bits
of PC and nPC always contain zero.
For a delayed control transfer, the instruction that immediately follows the transfer instruction is
known as the delay instruction. This delay instruction is executed (unless the control transfer
instruction annuls it) before control is transferred to the target. During execution of the delay
instruction, the nPC points to the target of the control transfer instruction, while the PC points to
the delay instruction. See Chapter 6, “Instructions.”
The PC is used implicitly as a destination register by CALL, Bicc, BPcc, BPr, FBfcc, FBPfcc,
JMPL, and RETURN instructions. It can be read directly by an RDPC instruction.

5.1.3.2 32-Bit Multiply/Divide Register (Y)
The Y register is deprecated; it is provided only for compatibility with previous versions of the architecture. It should not be used in new SPARC-V9 software. It is
recommended that all instructions that reference the Y register (i.e., SMUL,
SMULcc, UMUL, UMULcc, MULScc, SDIV, SDIVcc, UDIV, UDIVcc, RDY, and
WRY) be avoided. See the appropriate pages in Appendix A, “Instruction Definitions,” for suitable substitute instructions.

—
63

product<63:32> or dividend<63:32>
32 31

Figure 4—Y Register

The low-order 32 bits of the Y register, illustrated in figure 4, contain the more significant word of
the 64-bit product of an integer multiplication, as a result of either a 32-bit integer multiply
(SMUL, SMULcc, UMUL, UMULcc) instruction or an integer multiply step (MULScc) instruction. The Y register also holds the more significant word of the 64-bit dividend for a 32-bit integer
divide (SDIV, SDIVcc, UDIV, UDIVcc) instruction.
Although Y is a 64-bit register, its high-order 32 bits are reserved and always read as 0.
The Y register is read and written with the RDY and WRY instructions, respectively.
5.1.3.3 Ancillary State Registers (ASRs)
SPARC-V9 provides for optional ancillary state registers (ASRs). Access to a particular ASR may
be privileged or nonprivileged (impl. dep. #9); see 5.2.11, “Ancillary State Registers (ASRs),” for
a more complete description of ASRs.
5.1.4 Floating-Point Registers
The FPU contains:
— 32 single-precision (32-bit) floating-point registers, numbered f[0], f[1], .. f[31].
— 32 double-precision (64-bit) floating-point registers, numbered f[0], f[2], .. f[62].
— 16 quad-precision (128-bit) floating-point registers, numbered f[0], f[4], .. f[60].
The floating-point registers are arranged so that some of them overlap, that is, are aliased. The
layout and numbering of the floating-point registers are shown in figures 5, 6, and 7. Unlike the
windowed r registers, all of the floating-point registers are accessible at any time. The floatingpoint registers can be read and written by FPop (FPop1/FPop2 format) instructions, and by load/
store single/double/quad floating-point instructions.

Figure 5—Single-Precision Floating-Point Registers, with Aliasing
Operand
register ID
f31
f30
f29
f28
f27
f26
f25
f24
f23
f22
f21
f20
f19
f18
f17
f16
f15
f14
f13
f12
f11
f10
f9
f8
f7
f6
f5
f4
f3
f2
f1
f0

Operand
from
f31<31:0>
f30<31:0>
f29<31:0>
f28<31:0>
f27<31:0>
f26<31:0>
f25<31:0>
f24<31:0>
f23<31:0>
f22<31:0>
f21<31:0>
f20<31:0>
f19<31:0>
f18<31:0>
f17<31:0>
f16<31:0>
f15<31:0>
f14<31:0>
f13<31:0>
f12<31:0>
f11<31:0>
f10<31:0>
f9<31:0>
f8<31:0>
f7<31:0>
f6<31:0>
f5<31:0>
f4<31:0>
f3<31:0>
f2<31:0>
f1<31:0>
f0<31:0>

Figure 6—Double-Precision Floating-Point Registers, with Aliasing
Operand
register ID
f62
f60
f58
f56
f54
f52
f50
f48
f46
f44
f42
f40
f38
f36
f34
f32
f30
f28
f26
f24
f22
f20
f18
f16
f14
f12
f10
f8
f6
f4
f2
f0

Operand
field
<63:0>
<63:0>
<63:0>
<63:0>
<63:0>
<63:0>
<63:0>
<63:0>
<63:0>
<63:0>
<63:0>
<63:0>
<63:0>
<63:0>
<63:0>
<63:0>
<31:0>
<63:32>
<31:0>
<63:32>
<31:0>
<63:32>
<31:0>
<63:32>
<31:0>
<63:32>
<31:0>
<63:32>
<31:0>
<63:32>
<31:0>
<63:32>
<31:0>
<63:32>
<31:0>
<63:32>
<31:0>
<63:32>
<31:0>
<63:32>
<31:0>
<63:32>
<31:0>
<63:32>
<31:0>
<63:32>
<31:0>
<63:32>

From
register
f62<63:0>
f60<63:0>
f58<63:0>
f56<63:0>
f54<63:0>
f52<63:0>
f50<63:0>
f48<63:0>
f46<63:0>
f44<63:0>
f42<63:0>
f40<63:0>
f38<63:0>
f36<63:0>
f34<63:0>
f32<63:0>
f31<31:0>
f30<31:0>
f29<31:0>
f28<31:0>
f27<31:0>
f26<31:0>
f25<31:0>
f24<31:0>
f23<31:0>
f22<31:0>
f21<31:0>
f20<31:0>
f19<31:0>
f18<31:0>
f17<31:0>
f16<31:0>
f15<31:0>
f14<31:0>
f13<31:0>
f12<31:0>
f11<31:0>
f10<31:0>
f9<31:0>
f8<31:0>
f7<31:0>
f6<31:0>
f5<31:0>
f4<31:0>
f3<31:0>
f2<31:0>
f1<31:0>
f0<31:0>

Figure 7—Quad-Precision Floating-Point Registers, with Aliasing
Operand
register ID
f60
f56
f52
f48
f44
f40
f36
f32

f28

f24

f20

f16

f12

Operand
field
<63:0>
<127:64>
<63:0>
<127:64>
<63:0>
<127:64>
<63:0>
<127:64>
<63:0>
<127:64>
<63:0>
<127:64>
<63:0>
<127:64>
<63:0>
<127:64>
<31:0>
<63:32>
<95:64>
<127:96>
<31:0>
<63:32>
<95:64>
<127:96>
<31:0>
<63:32>
<95:64>
<127:96>
<31:0>
<63:32>
<95:64>
<127:96>
<31:0>
<63:32>
<95:64>
<127:96>
<31:0>
<63:32>
<95:64>
<127:96>
<31:0>
<63:32>
<95:64>
<127:96>
<31:0>
<63:32>
<95:64>
<127:96>

5.1.4.1 Floating-Point Register Number Encoding
Register numbers for single, double, and quad registers are encoded differently in the 5-bit register number field in a floating-point instruction. If the bits in a register number field are labeled:
b<4>..b<0> (where b<4> is the most-significant bit of the register number), the encoding of floating-point register numbers into 5-bit instruction fields is as given in table 7.
Table 7—Floating-Point Register Number Encoding
Register
operand
type
Single f.p. or
32-bit integer
Double f.p. or
64-bit integer
Quad f.p.

Encoding in a
5-bit register field
in an instruction

6-bit register number
0

b<4>

b<3>

b<2>

b<1>

b<0>

b<4>

b<3>

b<2>

b<1>

b<0>

b<5>

b<4>

b<3>

b<2>

b<1>

b<4>

b<3>

b<2>

b<1>

b<5>

b<4>

b<3>

b<2>

b<4>

b<3>

b<2>

b<5>

Compatibility Note:
In SPARC-V8, bit 0 of double and quad register numbers encoded in instruction fields was required to be
zero. Therefore, all SPARC-V8 floating-point instructions can run unchanged on a SPARC-V9 implementation using the encoding in table 7.

5.1.4.2 Double and Quad Floating-Point Operands
A single f register can hold one single-precision operand, a double-precision operand requires an
aligned pair of f registers, and a quad-precision operand requires an aligned quadruple of f registers. At a given time, the floating-point registers can hold a maximum of 32 single-precision, 16
double-precision, or 8 quad-precision values in the lower half of the floating-point register file,
plus an additional 16 double-precision or 8 quad-precision values in the upper half, or mixtures of
the three sizes.
Programming Note:
Data to be loaded into a floating-point double or quad register that is not doubleword-aligned in memory
must be loaded into the lower 16 double registers (8 quad registers) using single-precision LDF instructions.
If desired, it can then be copied into the upper 16 double registers (8 quad registers).

An attempt to execute an instruction that refers to a misaligned floating-point register operand
(that is, a quad-precision operand in a register whose 6-bit register number is not 0 mod 4) shall
cause an fp_exception_other trap, with FSR.ftt = 6 (invalid_fp_register).
Programming Note:
Given the encoding in table 7, it is impossible to specify a misaligned double-precision register.

5.1.5 Condition Codes Register (CCR)

xcc

CCR
7

icc
4

Figure 8—Condition Codes Register

The Condition Codes Register (CCR) holds the integer condition codes.
5.1.5.1 CCR Condition Code Fields (xcc and icc)
All instructions that set integer condition codes set both the xcc and icc fields. The xcc condition
codes indicate the result of an operation when viewed as a 64-bit operation. The icc condition
codes indicate the result of an operation when viewed as a 32-bit operation. For example, if an
operation results in the 64-bit value 0000 0000 FFFF FFFF16, the 32-bit result is negative (icc.N is
set to 1) but the 64-bit result is nonnegative (xcc.N is set to 0).
Each of the 4-bit condition-code fields is composed of four 1-bit subfields, as shown in figure 9.
n z v c
xcc:
icc:

Figure 9—Integer Condition Codes (CCR_icc and CCR_xcc)

The n bits indicate whether the 2’s-complement ALU result was negative for the last instruction
that modified the integer condition codes. 1 = negative, 0 = not negative.
The z bits indicate whether the ALU result was zero for the last instruction that modified the integer condition codes. 1 = zero, 0 = nonzero.
The v bits indicate whether the ALU result was within the range of (was representable in) 64-bit
(xcc) or 32-bit (icc) 2’s complement notation for the last instruction that modified the integer condition codes. 1 = overflow, 0 = no overflow.
The c bits indicate whether a 2’s complement carry (or borrow) occurred during the last instruction that modified the integer condition codes. Carry is set on addition if there is a carry out of bit
63 (xcc) or bit 31 (icc). Carry is set on subtraction if there is a borrow into bit 63 (xcc) or bit 31
(icc). 1 = carry, 0 = no carry.

5.1.5.1.1 CCR_extended_integer_cond_codes (xcc)
Bits 7 through 4 are the IU condition codes that indicate the results of an integer operation with
both of the operands considered to be 64 bits long. These bits are modified by the arithmetic and
logical instructions the names of which end with the letters “cc” (e.g., ANDcc) and by the
WRCCR instruction. They can be modified by a DONE or RETRY instruction, which replaces
these bits with the CCR field of the TSTATE register. The BPcc and Tcc instructions may cause a
transfer of control based on the values of these bits. The MOVcc instruction can conditionally
move the contents of an integer register based on the state of these bits. The FMOVcc instruction
can conditionally move the contents of a floating-point register based on the state of these bits.

5.1.5.1.2 CCR_integer_cond_codes (icc)
Bits 3 through 0 are the IU condition codes, which indicate the results of an integer operation with
both of the operands considered to be 32 bits. These bits are modified by the arithmetic and logical instructions the names of which end with the letters “cc” (e.g., ANDcc) and by the WRCCR
instruction. They can be modified by a DONE or RETRY instruction, which replaces these bits
with the CCR field of the TSTATE register. The BPcc, Bicc, and Tcc instructions may cause a
transfer of control based on the values of these bits. The MOVcc instruction can conditionally
move the contents of an integer register based on the state of these bits. The FMOVcc instruction
can conditionally move the contents of a floating-point register based on the state of these bits.
5.1.6 Floating-Point Registers State (FPRS) Register
FPRS

FEF DU DL
2

Figure 10—Floating-Point Registers State Register

The Floating-Point Registers State (FPRS) register holds control information for the floatingpoint register file; this information is readable and writable by nonprivileged software.
5.1.6.1 FPRS_enable_fp (FEF)
Bit 2, FEF, determines whether the FPU is enabled. If it is disabled, executing a floating-point
instruction causes an fp_disabled trap. If this bit is set but the PSTATE.PEF bit is not set, then executing a floating-point instruction causes an fp_disabled trap; that is, both FPRS.FEF and
PSTATE.PEF must be set to enable floating-point operations.
5.1.6.2 FPRS_dirty_upper (DU)
Bit 1 is the “dirty” bit for the upper half of the floating-point registers; that is, f32..f62. It is set
whenever any of the upper floating-point registers is modified. Its setting may be pessimistic; that
is, it may be set in some cases even though no register was actually modified. It is cleared only by
software.
5.1.6.3 FPRS_dirty_lower (DL)
Bit 0 is the “dirty” bit for the lower 32 floating-point registers; that is, f0..f31. It is set whenever
any of the lower floating-point registers is modified. Its setting may be pessimistic; that is, it may
be set in some cases even though no register was actually modified. It is cleared only by software.
Implementation Note:
The pessimistic setting of FPRS.DL and FPRS.DU allows hardware to set these bits even though the modification of a floating-point register might be cancelled before data is written.

5.1.7 Floating-Point State Register (FSR)
The FSR register fields, illustrated in figure 11, contain FPU mode and status information. The
lower 32 bits of the FSR are read and written by the STFSR and LDFSR instructions; all 64 bits of
the FSR are read and written by the STXFSR and LDXFSR instructions, respectively. The ver, ftt,
and reserved fields are not modified by LDFSR or LDXFSR.

—

fcc3

38 37 36

—

31 30 29 28 27

TEM

—

ver

23 22 21 20 19

ftt
17 16

qne — fcc0
14 13 12 11 10 9

aexc

fcc2

fcc1

35 34 33 32

cexc
5

Figure 11—FSR Fields

Bits 63..38, 29..28, 21..20, and 12 are reserved. When read by an STXFSR instruction, these bits
shall read as zero. Software should only issue LDXFSR instructions with zero values in these bits,
unless the values of these bits are exactly those derived from a previous STFSR.
Subsections 5.1.7.1 through 5.1.7.10.5 describe the remaining fields in the FSR.
5.1.7.1 FSR_fp_condition_codes (fcc0, fcc1, fcc2, fcc3)
There are four sets of floating-point condition code fields, labeled fcc0, fcc1, fcc2, and fcc3.
Compatibility Note:
SPARC-V9’s fcc0 is the same as SPARC-V8’s fcc.

The fcc0 field consists of bits 11 and 10 of the FSR, fcc1 consists of bits 33 and 32, fcc2 consists
of bits 35 and 34, and fcc3 consists of bits 37 and 36. Execution of a floating-point compare
instruction (FCMP or FCMPE) updates one of the fccn fields in the FSR, as selected by the
instruction. The fccn fields are read and written by STXFSR and LDXFSR instructions, respectively. The fcc0 field may also be read and written by STFSR and LDFSR, respectively. FBfcc and
FBPfcc instructions base their control transfers on these fields. The MOVcc and FMOVcc instructions can conditionally copy a register based on the state of these fields.
In table 8, frs1 and frs2 correspond to the single, double, or quad values in the floating-point registers specified by a floating-point compare instruction’s rs1 and rs2 fields. The question mark (‘?’)
indicates an unordered relation, which is true if either frs1 or frs2 is a signalling NaN or quiet NaN.
If FCMP or FCMPE generates an fp_exception_ieee_754 exception, then fccn is unchanged.

Table 8—Floating-Point Condition Codes (fccn) Fields of FSR
Content of
fccn
0
1
2
3

Indicated relation
frs1 = frs2
frs1 < frs2
frs1 > frs2
frs1 ? frs2 (unordered)

5.1.7.2 FSR_rounding_direction (RD)
Bits 31 and 30 select the rounding direction for floating-point results according to IEEE Std 7541985. Table 9 shows the encodings.
Table 9—Rounding Direction (RD) Field of FSR
RD
0
1
2
3

Round toward
Nearest (even if tie)
0
+∞
−∞

5.1.7.3 FSR_trap_enable_mask (TEM)
Bits 27 through 23 are enable bits for each of the five IEEE-754 floating-point exceptions that can
be indicated in the current_exception field (cexc). See figure 12 on page 67. If a floating-point
operate instruction generates one or more exceptions and the TEM bit corresponding to any of the
exceptions is 1, an fp_exception_ieee_754 trap is caused. A TEM bit value of 0 prevents the corresponding exception type from generating a trap.
5.1.7.4 FSR_nonstandard_fp (NS)
IMPL. DEP. #18: When set to 1, bit 22 causes the FPU to produce implementation-defined results that
may not correspond to IEEE Std 754-1985.

For instance, to obtain higher performance, implementations may convert a subnormal floatingpoint operand or result to zero when FSR.NS is set. SPARC-V9 implementations are permitted
but not encouraged to deviate from IEEE 754 requirements when the nonstandard mode bit of the
FSR is 1. For implementations in which no nonstandard floating-point mode exists, the NS bit of
the FSR should always read as 0, and writes to it should be ignored.
See Implementation Characteristics of Current SPARC-V9-based Products, Revision 9.x,
a document available from SPARC International, for a description of how this field is used in
existing implementations.

5.1.7.5 FSR_version (ver)
IMPL. DEP. #19: Bits 19 through 17 identify one or more particular implementations of the FPU architecture.

For each SPARC-V9 IU implementation (as identified by its VER.impl field), there may be one or
more FPU implementations, or none. This field identifies the particular FPU implementation
present. Version number 7 is reserved to indicate that no hardware floating-point controller is
present. See Implementation Characteristics of Current SPARC-V9-based Products, Revision 9.x, a document available from SPARC International, for a description of the values of this
field in existing implementations.
The ver field is read-only; it cannot be modified by the LDFSR and LDXFSR instructions.
5.1.7.6 FSR_floating-point_trap_type (ftt)
Several conditions can cause a floating-point exception trap. When a floating-point exception trap
occurs, ftt (bits 16 through 14 of the FSR) identifies the cause of the exception, the “floating-point
trap type.” After a floating-point exception occurs, the ftt field encodes the type of the floatingpoint exception until an STFSR or an FPop is executed.
The ftt field can be read by the STFSR and STXFSR instructions. The LDFSR and LDXFSR
instructions do not affect ftt.
Privileged software that handles floating-point traps must execute an STFSR (or STXFSR) to
determine the floating-point trap type. STFSR and STXFSR shall zero ftt after the store completes
without error. If the store generates an error and does not complete, ftt shall remain unchanged.
Programming Note:
Neither LDFSR nor LDXFSR can be used for this purpose, since both leave ftt unchanged. However, executing a nontrapping FPop such as “fmovs %f0,%f0” prior to returning to nonprivileged mode will zero ftt.
The ftt remains valid until the next FPop instruction completes execution.

The ftt field encodes the floating-point trap type according to table 10. Note that the value “7” is
reserved for future expansion.
Table 10—Floating-Point Trap Type (ftt) Field of FSR
ftt
0
1
2
3
4
5
6
7

Trap type
None
IEEE_754_exception
unfinished_FPop
unimplemented_FPop
sequence_error
hardware_error
invalid_fp_register
—

The sequence_error and hardware_error trap types are unlikely to arise in the normal course of
computation. They are essentially unrecoverable from the point of view of user applications. In
contrast, IEEE_754_exception, unfinished_FPop, and unimplemented_FPop will likely arise occasionally in the normal course of computation and must be recoverable by system software.
When a floating-point trap occurs, the following results are observed by user software:
(1) The value of aexc is unchanged.
(2) The value of cexc is unchanged, except that for an IEEE_754_exception a bit corresponding
to the trapping exception is set. The unfinished_FPop, unimplemented_FPop, sequence_error,
and invalid_fp_register floating-point trap types do not affect cexc.
(3) The source registers are unchanged (usually implemented by leaving the destination registers unchanged).
(4) The value of fccn is unchanged.
The foregoing describes the result seen by a user trap handler if an IEEE exception is signalled,
either immediately from an IEEE_754_exception or after recovery from an unfinished_FPop or
unimplemented_FPop. In either case, cexc as seen by the trap handler reflects the exception causing
the trap.
In the cases of unfinished_FPop and unimplemented_FPop exceptions that do not subsequently generate IEEE traps, the recovery software should define cexc, aexc, and the destination registers or
fccs, as appropriate.
5.1.7.6.1 ftt = IEEE_754_exception
The IEEE_754_exception floating-point trap type indicates that a floating-point exception conforming to IEEE Std 754-1985 has occurred. The exception type is encoded in the cexc field. Note that
aexc, the fccs, and the destination f register are not affected by an IEEE_754_exception trap.
5.1.7.6.2 ftt = unfinished_FPop
The unfinished_FPop floating-point trap type indicates that an implementation’s FPU was unable to
generate correct results, or that exceptions as defined by IEEE Std 754-1985 have occurred. In the
latter case, the cexc field is unchanged.
5.1.7.6.3 ftt = unimplemented_FPop
The unimplemented_FPop floating-point trap type indicates that an implementation’s FPU decoded
an FPop that it does not implement. In this case, the cexc field is unchanged.
Programming Note:
For the unfinished_FPop and unimplemented_FPop floating-point traps, software should emulate or reexecute the exception-causing instruction and update the FSR, destination f register(s), and fccs.

5.1.7.6.4 ftt = sequence_error
The sequence_error floating-point trap type indicates one of three abnormal error conditions in the
FPU, all caused by erroneous supervisor software:
— An attempt was made to read the floating-point deferred-trap queue (FQ) on an implementation without an FQ.
Implementation Note:
IMPL. DEP #25: On implementations without a floating-point queue, an attempt to read the fq with
an RDPR instruction shall cause either an illegal_instruction exception or an fp_exception_other
exception with FSR.ftt set to 4 (sequence_error).

— An attempt was made to execute a floating-point instruction when the FPU was unable to
accept one. This type of sequence_error arises from a logic error in supervisor software
that has caused a previous floating-point trap to be incompletely serviced (for example, the
floating-point queue was not emptied after a previous floating-point exception).
— An attempt was made to read the floating-point deferred-trap queue (FQ) with a RDPR
instruction when the FQ was empty; that is, when FSR.qne = 0. Note that generation of
sequence_error is recommended but not required in this case.
Programming Note:
If a sequence_error floating-point exception occurs while executing user code due to any of the above conditions, it may not be possible to recover sufficient state to continue execution of the user application.

5.1.7.6.5 ftt = hardware_error
The hardware_error floating-point trap type indicates that the FPU detected a catastrophic internal
error, such as an illegal state or a parity error on an f register access.
Programming Note:
If a hardware_error occurs while executing user code, it may not be possible to recover sufficient state to
continue execution of the user application.

5.1.7.6.6 ftt = invalid_fp_register
The invalid_fp_register trap indicates that one (or more) operands of an FPop are misaligned; that
is, a quad-precision register number is not 0 mod 4. An implementation shall generate an
fp_exception_other trap with FSR.ftt = invalid_fp_register in this case.
5.1.7.7 FSR_FQ_not_empty (qne)
Bit 13 indicates whether the optional floating-point deferred-trap queue (FQ) is empty after a
deferred floating-point exception trap or after a read privileged register (RDPR) instruction that
reads the queue has been executed. If qne = 0, the queue is empty; if qne = 1, the queue is not
empty.

The qne bit can be read by the STFSR and STXFSR instructions. The LDFSR and LDXFSR
instructions do not affect qne. However, executing successive “RDPR %fpq” instructions will
(eventually) cause the FQ to become empty (qne = 0). If an implementation does not provide an
FQ, this bit shall read as zero. Supervisor software must arrange for this bit to always read as zero
to user-mode software.
5.1.7.8 FSR_accrued_exception (aexc)
Bits 9 through 5 accumulate IEEE_754 floating-point exceptions while floating-point exception
traps are disabled using the TEM field. See figure 13 on page 68. After an FPop completes, the
TEM and cexc fields are logically ANDed together. If the result is nonzero, aexc is left unchanged
and an fp_exception_ieee_754 trap is generated; otherwise, the new cexc field is ORed into the aexc
field and no trap is generated. Thus, while (and only while) traps are masked, exceptions are accumulated in the aexc field.
5.1.7.9 FSR_current_exception (cexc)
Bits 4 through 0 indicate that one or more IEEE_754 floating-point exceptions were generated by
the most recently executed FPop instruction. The absence of an exception causes the corresponding bit to be cleared. See figure 14 on page 68.
The cexc bits are set as described in 5.1.7.10, “Floating-Point Exception Fields,” by the execution
of an FPop that either does not cause a trap or causes an fp_exception_ieee_754 trap with
FSR.ftt = IEEE_754_exception. An IEEE_754_exception that traps shall cause exactly one bit in
FSR.cexc to be set, corresponding to the detected IEEE Std 754-1985 exception.
In the case of an overflow (underflow) IEEE_754_exception that does not trap (because neither
OFM (UFM) nor NXM is set), more than one bit in cexc is set: such an overflow (underflow) sets
both ofc (ufc) and nxc. An overflow (underflow) IEEE_754_exception that does trap (because OFM
(UFM) or NXM or both are set) shall set ofc (ufc), but not nxc.
If the execution of an FPop causes a trap other than an fp_exception_ieee_754 due to an IEEE Std
754-1985 exception, FSR.cexc is left unchanged.
5.1.7.10 Floating-Point Exception Fields
The current and accrued exception fields and the trap enable mask assume the following definitions of the floating-point exception conditions (per IEEE Std 754-1985):
NVM OFM UFM DZM NXM
27

Figure 12—Trap Enable Mask (TEM) Fields of FSR

nva

ofa

ufa

dza

nxa

Figure 13—Accrued Exception Bits (aexc) Fields of FSR

nvc

ofc

ufc

dzc

nxc

Figure 14—Current Exception Bits (cexc) Fields of FSR

5.1.7.10.1 FSR_invalid (nvc, nva)
An operand is improper for the operation to be performed. For example, 0.0 ÷ 0.0 and ∞ – ∞ are
invalid. 1 = invalid operand(s), 0 = valid operand(s).
5.1.7.10.2 FSR_overflow (ofc, ofa)
The result, rounded as if the exponent range were unbounded, would be larger in magnitude than
the destination format’s largest finite number. 1 = overflow, 0 = no overflow.
5.1.7.10.3 FSR_underflow (ufc, ufa)
The rounded result is inexact and would be smaller in magnitude than the smallest normalized
number in the indicated format. 1 = underflow, 0 = no underflow.
Underflow is never indicated when the correct unrounded result is zero. Otherwise:
If UFM = 0: Underflow occurs if a nonzero result is tiny and a loss of accuracy occurs. Tininess may be detected before or after rounding (impl. dep. #55). Loss of accuracy may be either a denormalization loss or an inexact result.
If UFM = 1: Underflow occurs if a nonzero result is tiny. Tininess may be detected before or
after rounding (impl. dep. #55).
5.1.7.10.4 FSR_division-by-zero (dzc, dza)
X ÷ 0.0, where X is subnormal or normalized. Note that 0.0 ÷ 0.0 does not set the dzc or dza bits.
1 = division by zero, 0 = no division by zero.
5.1.7.10.5 FSR_inexact (nxc, nxa)
The rounded result of an operation differs from the infinitely precise unrounded result. 1 = inexact
result, 0 = exact result.

5.1.7.11 FSR Conformance
IMPL. DEP. #22: An implementation may choose to implement the TEM, cexc, and aexc fields in hardware
in either of two ways (both of which comply with IEEE Std 754-1985):

(1) Implement all three fields conformant to IEEE Std 754-1985.
(2) Implement the NXM, nxa, and nxc bits of these fields conformant to IEEE Std 754-1985.
Implement each of the remaining bits in the three fields either
(a) Conformant to IEEE Std 754-1985, or
(b) As a state bit that may be set by software that calculates the IEEE Std 754-1985 value
of the bit. For any bit implemented as a state bit:
[1] The IEEE exception corresponding to the state bit must always cause an exception
(specifically, an unfinished_FPop exception). During exception processing in the
trap handler, the bit in the state field can be written to the appropriate value by an
LDFSR or LDXFSR instruction.
[2] The state bit must be implemented in such a way that if it is written to a particular
value by an LDFSR or LDXFSR instruction, it will be read back as the same value
by a subsequent STFSR or STXFSR.
Programming Note:
Software must be capable of simulating the operation of the FPU in order to handle the
unimplemented_FPop, unfinished_FPop, and IEEE_754_exception floating-point trap types properly. Thus, a
user application program always sees an FSR that is fully compliant with IEEE Std 754-1985.

5.1.8 Address Space Identifier Register (ASI)
ASI
7

Figure 15—ASI Register

The ASI register specifies the address space identifier to be used for load and store alternate
instructions that use the “rs1 + simm13” addressing form. Nonprivileged (user-mode) software
may write any value into the ASI register; however, values with bit 7 = 0 indicate restricted ASIs.
When a nonprivileged instruction makes an access that uses an ASI with bit 7 = 0, a
privileged_action exception is generated. See 6.3.1.3, “Address Space Identifiers (ASIs),” for
details.
5.1.9 TICK Register (TICK)
TICK

counter

NPT
63

Figure 16—TICK Register

The counter field of the TICK register is a 63-bit counter that counts CPU clock cycles. Bit 63 of
the TICK register is the Nonprivileged Trap (NPT) bit, which controls access to the TICK register
by nonprivileged software. Privileged software can always read the TICK register with either the
RDPR or RDTICK instruction. Privileged software can always write the TICK register with the
WRPR instruction; there is no WRTICK instruction.
Nonprivileged software can read the TICK register using the RDTICK instruction; TICK.NPT
must be 0. When TICK.NPT = 1, an attempt by nonprivileged software to read the TICK register
causes a privileged_action exception. Nonprivileged software cannot write the TICK register.
TICK.NPT is set to 1 by a power-on reset trap. The value of TICK.counter is undefined after a
power-on reset trap.
After the TICK register is written, reading the TICK register returns a value incremented (by one
or more) from the last value written, rather than from some previous value of the counter. The
number of counts between a write and a subsequent read need not accurately reflect the number of
processor cycles between the write and the read. Software may only rely on read-to-read counts of
the TICK register for accurate timing, not on write-to-read counts.
IMPL. DEP. #105: The difference between the values read from the TICK register on two reads should
reflect the number of processor cycles executed between the reads. If an accurate count cannot always be
returned, any inaccuracy should be small, bounded, and documented. An implementation may implement
fewer than 63 bits in TICK.counter ; however, the counter as implemented must be able to count for at least
10 years without overflowing. Any upper bits not implemented must read as zero.
Programming Note:
TICK.NPT may be used by a secure operating system to control access by user software to high-accuracy
timing information. The operation of the timer might be emulated by the trap handler, which could read
TICK.counter and “fuzz” the value to lower accuracy.

5.2 Privileged Registers
The registers described in this subsection are visible only to software running in privileged mode;
that is, when PSTATE.PRIV = 1. Privileged registers are written using the WRPR instruction and
read using the RDPR instruction.
5.2.1 Processor State Register (PSTATE)
PSTATE PID1 PID0
11

CLE

TLE

MM
7

RED PEF
6

PRIV

Figure 17—PSTATE Fields

The PSTATE register holds the current state of the processor. There is only one instance of the
PSTATE register. See Chapter 7, “Traps,” for more details.

Writing PSTATE is nondelayed; that is, new machine state written to PSTATE is visible to the
next instruction executed. The privileged RDPR and WRPR instructions are used to read and
write PSTATE, respectively.
Implementation Note:
To ensure the nondelayed semantics, a write to PSTATE may take multiple cycles to complete on some
implementations.

5.2.1.2 through 5.2.1.10 describe the fields contained in the PSTATE register.
5.2.1.1 PSTATE_impldep (PID1, PID0)
IMPL. DEP. #127: The presence and semantics of PSTATE.PID1 and PSTATE.PID0 are implementationdependent. Software intended to run on multiple implementations should only write these bits to values
previously read from PSTATE, or to zeroes.

See also TSTATE bits 19..18.
5.2.1.2 PSTATE_current_little_endian (CLE)
When PSTATE.CLE = 1, all data reads and writes using an implicit ASI are performed in littleendian byte order with an ASI of ASI_PRIMARY_LITTLE. When PSTATE.CLE = 0, all data
reads and writes using an implicit ASI are performed in big-endian byte order with an ASI of
ASI_PRIMARY. Instruction accesses are always big-endian.
5.2.1.3 PSTATE_trap_little_endian (TLE)
When a trap is taken, the current PSTATE register is pushed onto the trap stack and the
PSTATE.TLE bit is copied into PSTATE.CLE in the new PSTATE register. This allows system
software to have a different implicit byte ordering than the current process. Thus, if PSTATE.TLE
is set to 1, data accesses using an implicit ASI in the trap handler are little-endian. The original
state of PSTATE.CLE is restored when the original PSTATE register is restored from the trap
stack.
5.2.1.4 PSTATE_mem_model (MM)
This 2-bit field determines the memory model in use by the processor. Its values are:
Value
00
01
10
11

Memory model
Total Store Order (TSO)
Partial Store Order (PSO)
Relaxed Memory Order (RMO)
—

An implementation must provide a memory model that allows programs conforming to the TSO
model to run correctly; that is, TSO or a stronger model. Whether the Partial Store Order (PSO)

model or the Relaxed Memory Ordering (RMO) model is supported is implementation-dependent
(impl. dep. #113).
The current memory model is determined by the value of PSTATE.MM. The effect of setting
PSTATE.MM to an unsupported value is implementation-dependent (impl. dep. #119).
5.2.1.5 PSTATE_RED_state (RED)
When PSTATE.RED is set to 1, the processor is operating in RED (Reset, Error, and Debug) state.
See 7.2.1, “RED_state.” The IU sets PSTATE.RED when any hardware reset occurs. It also sets
PSTATE.RED when a trap is taken while TL = (MAXTL – 1). Software can exit RED_state by
one of two methods:
(1) Execute a DONE or RETRY instruction, which restores the stacked copy of PSTATE and
clears PSTATE.RED if it was 0 in the stacked copy.
(2) Write a 0 to PSTATE.RED with a WRPR instruction.
Programming Note:
Changing PSTATE.RED may cause a change in address mapping on some systems. It is recommended that
writes of PSTATE.RED be placed in the delay slot of a JMPL; the target of this JMPL should be in the new
address mapping. The JMPL sets the nPC, which becomes the PC for the instruction that folows the WPR in
its delay slot. The effect of the WPR instruction is immediate.

5.2.1.6 PSTATE_enable_floating-point (PEF)
When set to 1, this bit enables the floating-point unit, which allows privileged software to manage
the FPU. For the floating-point unit to be usable, both PSTATE.PEF and FPRS.FEF must be set.
Otherwise, a floating-point instruction that tries to reference the FPU will cause an fp_disabled
trap.
5.2.1.7 PSTATE_address_mask (AM)
When PSTATE.AM = 1, both instruction and data addresses are interpreted as if the high-order 32
bits were masked to zero before being presented to the MMU or memory system. Thirty-two-bit
application software must run with this bit set.
Branch target addresses (sent to the nPC) and addresses sent to registers by CALL, JMPL, and
RDPC instructions are always 64-bit values, but the value of the high-order 32-bits are implementation-dependent. Similarly, the value of the high-order 32-bits of TPC and TNPC after a trap
taken while PSTATE.AM = 1 is implementation-dependent.
IMPL. DEP. #125: When PSTATE.AM = 1, the value of the high-order 32-bits of the PC transmitted to the
specified destination register(s) by CALL, JMLP, RDPC, and on a trap is implementation-dependent.

5.2.1.8 PSTATE_privileged_mode (PRIV)
When PSTATE.PRIV = 1, the processor is in privileged mode.

5.2.1.9 PSTATE_interrupt_enable (IE)
When PSTATE.IE = 1, the processor can accept interrupts.
5.2.1.10 PSTATE_alternate_globals (AG)
When PSTATE.AG = 0, the processor interprets integer register numbers in the range 0..7 as
referring to the normal global register set. When PSTATE.AG = 1, the processor interprets integer
register numbers in the range 0..7 as referring to the alternate global register set.
5.2.2 Trap Level Register (TL)
TL

TL
2

Figure 18—Trap Level Register

The trap level register specifies the current trap level. TL = 0 is the normal (nontrap) level of operation. TL > 0 implies that one or more traps are being processed. The maximum valid value that
the TL register may contain is “MAXTL.” This is always equal to the number of supported trap
levels beyond level 0. See Chapter 7, “Traps,” for more details about the TL register. An implementation shall support at least four levels of traps beyond level 0; that is, MAXTL shall be ≥ 4.
IMPL. DEP. #101: How many additional trap levels, if any, past level 4 are supported is implementationdependent.

The remainder of this subsection assumes that there are four trap levels beyond level 0.
Programming Note:
Writing the TL register with a wrpr %tl instruction does not alter any other machine state; that is, it is not
equivalent to taking or returning from a trap.

5.2.3 Processor Interrupt Level (PIL)
PIL

PIL
3

Figure 19—Processor Interrupt Level Register

The processor interrupt level (PIL) is the interrupt level above which the processor will accept an
interrupt. Interrupt priorities are mapped such that interrupt level 2 has greater priority than interrupt level 1, and so on. See table 15 on page 103 for a list of exception and interrupt priorities.
Compatibility Note:
On SPARC-V8 processors, the level 15 interrupt is considered to be nonmaskable, so it has different semantics from other interrupt levels. SPARC-V9 processors do not treat level 15 interrupts differently from other

interrupt levels. See 7.6.2.4, “Externally Initiated Reset (XIR) Traps,” for a facility in SPARC-V9 that is
similar to a nonmaskable interrupt.

5.2.4 Trap Program Counter (TPC)
TPC1

PC from trap while TL = 0

TPC2

PC from trap while TL = 1

TPC3

PC from trap while TL = 2

TPC4

PC from trap while TL = 3

1 0

Figure 20—Trap Program Counter Register

The TPC register contains the program counter (PC) from the previous trap level. There are
MAXTL instances of the TPC (impl. dep. #101), but only one is accessible at any time. The current value in the TL register determines which instance of the TPC register is accessible. An
attempt to read or write the TPC register when TL = 0 shall cause an illegal_instruction exception.
5.2.5 Trap Next Program Counter (TNPC)
TNPC1

nPC from trap while TL = 0

TNPC2

nPC from trap while TL = 1

TNPC3

nPC from trap while TL = 2

TNPC4

nPC from trap while TL = 3

1 0

Figure 21—Trap Next Program Counter Register

The TNPC register is the next program counter (nPC) from the previous trap level. There are
MAXTL instances of the TNPC (impl. dep. #101), but only one is accessible at any time. The current value in the TL register determines which instance of the TNPC register is accessible. An
attempt to read or write the TNPC register when TL = 0 shall cause an illegal_instruction exception.

5.2.6 Trap State (TSTATE)
TSTATE1

CCR from TL = 0

ASI from TL = 0

—

PSTATE from TL = 0

—

CWP from TL = 0

TSTATE2

CCR from TL = 1

ASI from TL = 1

—

PSTATE from TL = 1

—

CWP from TL = 1

TSTATE3

CCR from TL = 2

ASI from TL = 2

—

PSTATE from TL = 2

—

CWP from TL = 2

TSTATE4

CCR from TL = 3

ASI from TL = 3

—

PSTATE from TL = 3

—

CWP from TL = 3

32 31

24 23

20 19

8 7

5 4

Figure 22—Trap State Register

TSTATE contains the state from the previous trap level, comprising the contents of the CCR, ASI,
CWP, and PSTATE registers from the previous trap level. There are MAXTL instances of the
TSTATE register, but only one is accessible at a time. The current value in the TL register determines which instance of TSTATE is accessible. An attempt to read or write the TSTATE register
when TL = 0 causes an illegal_instruction exception.
TSTATE bits 19 and 18 are implementation-dependent. IMPL.DEP. #127: If PSTATE bit 11 (10) is
implemented, TSTATE bit 19 (18) shall be implemented and contain the state of PSTATE bit 11 (10) from
the previous trap level. If PSTATE bit 11 (10) is not implemented, TSTATE bit 19 (18) shall read as zero.
Software intended to run on multiple implementations should only write these bits to values previously read
from PSTATE, or to zeroes.

5.2.7 Trap Type Register (TT)
TT1 Trap type from trap while TL = 0
TT2 Trap type from trap while TL = 1
TT3 Trap type from trap while TL = 2
TT4 Trap type from trap while TL = 3
8

Figure 23—Trap Type Register

The TT register normally contains the trap type of the trap that caused entry to the current trap
level. On a reset trap the TT field contains the trap type of the reset (see 7.2.1.1, “RED_state Trap
Table”), except when a watchdog (WDR) or externally initiated (XIR) reset occurs while the processor is in error_state. When this occurs, the TT register will contain the trap type of the exception that caused entry into error_state.
There are MAXTL instances of the TT register (impl. dep. #101), but only one is accessible at a
time. The current value in the TL register determines which instance of the TT register is accessible. An attempt to read or write the TT register when TL = 0 shall cause an illegal_instruction
exception.

5.2.8 Trap Base Address (TBA)
Trap Base Address

000000000000000

15 14

Figure 24—Trap Base Address Register

The TBA register provides the upper 49 bits of the address used to select the trap vector for a trap.
The lower 15 bits of the TBA always read as zero, and writes to them are ignored.
The full address for a trap vector is specified by the TBA, TL, TT[TL], and five zeroes:
TBA<63:15>

TTTL

TL>0

00000
5

Figure 25—Trap Vector Address

Note that the “(TL>0)” bit is 0 if TL = 0 when the trap was taken, and 1 if TL > 0 when the trap
was taken. This implies that there are two trap tables: one for traps from TL = 0 and one for traps
from TL > 0. See Chapter 7, “Traps,” for more details on trap vectors.
5.2.9 Version Register (VER)
manuf
63

impl
48 47

mask
32 31

maxtl

—
24 23

16 15

— maxwin
8 7 5 4

Figure 26—Version Register

The version register specifies the fixed parameters pertaining to a particular CPU implementation
and mask set. The VER register is read-only.
IMPL. DEP. #104: VER.manuf contains a 16-bit manufacturer code. This field is optional and, if not
present, shall read as 0. VER.manuf may indicate the original supplier of a second-sourced chip. It is
intended that the contents of VER.manuf track the JEDEC semiconductor manufacturer code as closely as
possible. If the manufacturer does not have a JEDEC semiconductor manufacturer code, SPARC International will assign a value for VER.manuf.
IMPL. DEP. #13: VER.impl uniquely identifies an implementation or class of software-compatible implementations of the architecture. Values FFF016 ..FFFF16 are reserved and are not available for assignment.

The value of VER.impl is assigned as described in C.3, “Implementation Dependency Categories.”
VER.mask specifies the current mask set revision, and is chosen by the implementor. It generally
increases numerically with successive releases of the processor, but does not necessarily increase
by one for consecutive releases.
VER.maxtl contains the maximum number of trap levels supported by an implementation (impl.
dep. #101), that is, MAXTL, the maximum value of the contents of the TL register.

VER.maxwin contains the maximum index number available for use as a valid CWP value in an
implementation; that is, VER.maxwin contains the value “NWINDOWS – 1” (impl. dep. #2).
5.2.10 Register-Window State Registers
The state of the register windows is determined by a set of privileged registers. They can be read/
written by privileged software using the RDPR/WRPR instructions. In addition, these registers
are modified by instructions related to register windows and are used to generate traps that allow
supervisor software to spill, fill, and clean register windows.
IMPL. DEP. #126: Privileged registers CWP, CANSAVE, CANRESTORE, OTHERWIN, and CLEANWIN
contain values in the range 0..NWINDOWS-1. The effect of writing a value greater than NWINDOWS-1 to
any of these registers is undefined. Although the width of each of these five registers is nominally 5 bits,
the width is implementation-dependent and shall be between log2(NWINDOWS) and 5 bits, inclusive. If
fewer than 5 bits are implemented, the unimplemented upper bits shall read as 0, and writes to them shall
have no effect. All five registers should have the same width.

The details of how the window-management registers are used by hardware are presented in 6.3.6,
“Register Window Management Instructions.”
5.2.10.1 Current Window Pointer (CWP)
CWP

Current Window #
4

Figure 27—Current Window Pointer Register

The CWP register is a counter that identifies the current window into the set of integer registers.
See 6.3.6, “Register Window Management Instructions,” and Chapter 7, “Traps,” for information
on how hardware manipulates the CWP register.
The effect of writing a value greater than NWINDOWS-1 to this register is undefined (impl. dep.
#126).
Compatibility Note:
The following differences between SPARC-V8 and SPARC-V9 are visible only to privileged software; they
are invisible to nonprivileged software:
1) In SPARC-V9, SAVE increments CWP and RESTORE decrements CWP. In SPARC-V8, the opposite is
true: SAVE decrements PSR.CWP and RESTORE increments PSR.CWP.
2) PSR.CWP in SPARC-V8 is changed on each trap. In SPARC-V9, CWP is affected only by a trap caused
by a window fill or spill exception.
3) In SPARC-V8, writing a value into PSR.CWP that is greater than or equal to the number of implemented windows causes an illegal_instruction exception. In SPARC-V9, the effect of writing an out-ofrange value to CWP is undefined.

5.2.10.2 Savable Windows (CANSAVE) Register
CANSAVE
4

Figure 28—CANSAVE Register

The CANSAVE register contains the number of register windows following CWP that are not in
use and are, hence, available to be allocated by a SAVE instruction without generating a window
spill exception
The effect of writing a value greater than NWINDOWS-1 to this register is undefined (impl. dep.
#126).
5.2.10.3 Restorable Windows (CANRESTORE) Register
CANRESTORE
4

Figure 29—CANRESTORE Register

The CANRESTORE register contains the number of register windows preceding CWP that are in
use by the current program and can be restored (via the RESTORE instruction) without generating a window fill exception.
The effect of writing a value greater than NWINDOWS-1 to this register is undefined (impl. dep.
#126).
5.2.10.4 Other Windows (OTHERWIN) Register
OTHERWIN
4

Figure 30—OTHERWIN Register

The OTHERWIN register contains the count of register windows that will be spilled/filled using a
separate set of trap vectors based on the contents of WSTATE_OTHER. If OTHERWIN is zero,
register windows are spilled/filled using trap vectors based on the contents of
WSTATE_NORMAL.
The OTHERWIN register can be used to split the register windows among different address
spaces and handle spill/fill traps efficiently by using separate spill/fill vectors.
The effect of writing a value greater than NWINDOWS-1 to this register is undefined (impl. dep.
#126).

5.2.10.5 Window State (WSTATE) Register
OTHER

WSTATE
5

NORMAL
3

Figure 31—WSTATE Register

The WSTATE register specifies bits that are inserted into TTTL<4:2> on traps caused by window
spill and fill exceptions. These bits are used to select one of eight different window spill and fill
handlers. If OTHERWIN = 0 at the time a trap is taken due to a window spill or window fill
exception, then the WSTATE.NORMAL bits are inserted into TT[TL]. Otherwise, the
WSTATE.OTHER bits are inserted into TT[TL]. See 6.4, “Register Window Management,” for
details of the semantics of OTHERWIN.
5.2.10.6 Clean Windows (CLEANWIN) Register
CLEANWIN
4

Figure 32—CLEANWIN Register

The CLEANWIN register contains the number of windows that can be used by the SAVE instruction without causing a clean_window exception.
The CLEANWIN register counts the number of register windows that are “clean” with respect to
the current program; that is, register windows that contain only zeros, valid addresses, or valid
data from that program. Registers in these windows need not be cleaned before they can be used.
The count includes the register windows that can be restored (the value in the CANRESTORE
register) and the register windows following CWP that can be used without cleaning. When a
clean window is requested (via a dSAVE instruction) and none is available, a clean_window exception occurs to cause the next window to be cleaned.
The effect of writing a value greater than NWINDOWS-1 to this register is undefined (impl. dep.
#126).
5.2.11 Ancillary State Registers (ASRs)
SPARC-V9 provides for up to 25 ancillary state registers (ASRs), numbered from 7 through 31.
ASRs numbered 7..15 are reserved for future use by the architecture and should not be referenced
by software.
ASRs numbered 16..31 are available for implementation-dependent uses (impl. dep. #8), such as
timers, counters, diagnostic registers, self-test registers, and trap-control registers. An IU may
choose to implement from zero to sixteen of these ASRs. The semantics of accessing any of these
ASRs is implementation-dependent. Whether access to a particular ancillary state register is privileged is implementation-dependent (impl. dep. #9).

An ASR is read and written with the RDASR and WRASR instructions, respectively. An RDASR
or WRASR instruction is privileged if the accessed register is privileged.
5.2.12 Floating-Point Deferred-Trap Queue (FQ)
If present in an implementation, the FQ contains sufficient state information to implement resumable, deferred floating-point traps.
IMPL. DEP. #23: Floating-point traps may be precise or deferred. If deferred, a floating-point deferred-trap
queue (FQ) shall be present.

The FQ can be read with the read privileged register (RDPR) floating-point queue instruction. In a
given implementation, it may also be readable or writable via privileged load/store double alternate instructions (LDDA, STDA), or by read/write ancillary state register instructions (RDASR,
WRASR).
IMPL. DEP. #24: The presence, contents of, and operations upon the FQ are implementation-dependent.

If an FQ is present, however, supervisor software must be able to deduce the exception-causing
instruction’s opcode (opf), operands, and address from its FQ entry. This also must be true of any
other pending floating-point operations in the queue. See Implementation Characteristics
of Current SPARC-V9-based Products, Revision 9.x, a document available from SPARC International, for a discussion of the formats and operation of implemented floating-point queues in
existing SPARC-V9 implementations.
In implementations with a floating-point queue, an attempt to read the FQ with a RDPR instruction when the FQ is empty (FSR.qne = 0) shall cause an fp_exception_other trap with FSR.ftt set to
4 (sequence_error).In implementations without an FQ, the qne bit in the FSR is always 0.
IMPL. DEP. #25: In implementations without a floating-point queue, an attempt to read the FQ with an
RDPR instruction shall cause either an illegal_instruction trap or an fp_exception_other trap with FSR.ftt
SET TO 4 (sequence_error).

5.2.13 IU Deferred-Trap Queue
An implementation may contain zero or more IU deferred-trap queues. Such a queue contains sufficient state to implement resumable deferred traps caused by the IU. See 7.3.2, “Deferred Traps,”
for more information. Note that deferred floating-point traps are handled by the floating-point
deferred-trap queue. See Implementation Characteristics of Current SPARC-V9-based
Products, Revision 9.x, a document available from SPARC International, for a discussion of such
queues in existing implementations.
IMPL. DEP. #16: The existence, contents, and operation of an IU deferred-trap queue are implementationdependent; it is not visible to user application programs under normal conditions.

6 Instructions
Instructions are accessed by the processor from memory and are executed, annulled, or trapped.
Instructions are encoded in four major formats and partitioned into eleven general categories.

6.1 Instruction Execution
The instruction at the memory location specified by the program counter is fetched and then executed. Instruction execution may change program-visible processor and/or memory state. As a
side-effect of its execution, new values are assigned to the program counter (PC) and the next program counter (nPC).
An instruction may generate an exception if it encounters some condition that makes it impossible
to complete normal execution. Such an exception may in turn generate a precise trap. Other events
may also cause traps: an exception caused by a previous instruction (a deferred trap), an interrupt
or asynchronous error (a disrupting trap), or a reset request (a reset trap). If a trap occurs, control
is vectored into a trap table. See Chapter 7, “Traps,” for a detailed description of exception and
trap processing.
If a trap does not occur and the instruction is not a control transfer, the next program counter
(nPC) is copied into the PC and the nPC is incremented by 4 (ignoring overflow, if any). If the
instruction is a control-transfer instruction, the next program counter (nPC) is copied into the PC
and the target address is written to nPC. Thus, the two program counters provide for a delayedbranch execution model.
For each instruction access and each normal data access, the IU appends an 8-bit address space
identifier, or ASI, to the 64-bit memory address. Load/store alternate instructions (see 6.3.1.3,
“Address Space Identifiers (ASIs),”) can provide an arbitrary ASI with their data addresses, or use
the ASI value currently contained in the ASI register.
Implementation Note:
The time required to execute an instruction is implementation-dependent, as is the degree of execution concurrency. In the absence of traps, an implementation should cause the same program-visible register and
memory state changes as if a program had executed according to the sequential model implied in this document. See Chapter 7, “Traps,” for a definition of architectural compliance in the presence of traps.

6.2 Instruction Formats
Instructions are encoded in four major 32-bit formats and several minor formats, as shown in figures 33 and 34.

Format 1 (op = 1): CALL
op

disp30

31 30 29

Format 2 (op = 0): SETHI & Branches (Bicc, BPcc, BPr, FBfcc, FBPfcc)
op

op2

imm22

disp22

cond

op2

cond

op2

cc1cc0 p

op2

d16hi p

rcond

30 29 28

25 24

disp19

rs1

22 21 20 19 18

d16lo
14 13

Format 3 (op = 2 or 3): Arithmetic, Logical, MOVr, MEMBAR, Load, and Store
op

op3

rs1

i=0

op3

rs1

i=1

—

op3

rs1

i=0

—

op3

rs1

i=1

op3

rs1

i=0

rcond

op3

rs1

i=1

rcond

op3

rs1

i=0

op3

rs1

i=1

—

op3

rs1

i=0

imm_asi

impl-dep

op3

31 30 29

27 26 25 24

—

rs2

simm13

—

rs2

simm13

—

rs2

simm10

—

rs2

cmask

mmask

rs2

impl-dep
19 18

14 13 12

Figure 33—Summary of Instruction Formats: Formats 1, 2, and 3

Format 3 (op = 2 or 3): Continued
op

op3

rs1

i=0 x

—

rs2

op3

rs1

i=1 x=0

—

shcnt32

op3

rs1

i=1 x=1

000

cc1 cc0

shcnt64

op3

—

opf

rs2

op3

rs1

opf

rs2

opf

rs2

op3

rs1

op3

rs1

fcn

op3

31 30 29

—

25 24

—

—
19 18

14 13 12 11

5 4

Format 4 (op = 2): MOVcc, FMOVr, FMOVcc, and Tcc
op

op3

rs1

i=0 cc1cc0

op3

rs1

i=1 cc1cc0

op3

cc2

cond

i=0 cc1 cc0

op3

cc2

cond

i=1 cc1 cc0

op3

rs1

i=1 cc1cc0

op3

rs1

op3

31 30 29

25 24

0
19 18 17

cond

rcond

opf_cc
14 13 12 11 10 9

rs2

—

simm11

—

rs2

simm11

—

sw_trap#

opf_low

rs2

opf_low
7

rs2
6 5

Figure 34—Summary of Instruction Formats: Formats 3 and 4

6.2.1 Instruction Fields
The instruction fields are interpreted as follows:
a:
The a bit annuls the execution of the following instruction if the branch is conditional and
untaken, or if it is unconditional and taken.
cc0, cc1, and cc2:
cc2:cc1:cc0 specify the condition codes (icc, xcc, fcc0, fcc1, fcc2, fcc3) to be used in the
instruction. Individual bits of the same logical field are present in several other instructions: Branch on Floating-Point Condition Codes with Prediction Instructions (FBPfcc),
Branch on Integer Condition Codes with Prediction (BPcc), Floating-Point Compare
Instructions, Move Integer Register if Condition is Satisfied (MOVcc), Move FloatingPoint Register if Condition is Satisfied (FMOVcc), and Trap on Integer Condition Codes
(Tcc). In instructions such as Tcc that do not contain the cc2 bit, the missing cc2 bit takes
on a default value. See table 38 on page 279 for a description of these fields’ values.
cmask:
This 3-bit field specifies sequencing constraints on the order of memory references and the
processing of instructions before and after a MEMBAR instruction.
cond:
This 4-bit field selects the condition tested by a branch instruction. See Appendix E,
“Opcode Maps,” for descriptions of its values.
d16hi and d16lo:
These 2-bit and 14-bit fields together comprise a word-aligned, sign-extended, PC-relative
displacement for a branch-on-register-contents with prediction (BPr) instruction.
disp19:
This 19-bit field is a word-aligned, sign-extended, PC-relative displacement for an integer
branch-with-prediction (BPcc) instruction or a floating-point branch-with-prediction
(FBPfcc) instruction.
disp22 and disp30:
These 22-bit and 30-bit fields are word-aligned, sign-extended, PC-relative displacements
for a branch or call, respectively.
fcn:
This 5-bit field provides additional opcode bits to encode the DONE and RETRY instructions.
i:
The i bit selects the second operand for integer arithmetic and load/store instructions. If
i = 0, the operand is r[rs2]. If i = 1, the operand is simm10, simm11, or simm13, depending
on the instruction, sign-extended to 64 bits.

imm22:
This 22-bit field is a constant that SETHI places in bits 31..10 of a destination register.
imm_asi:
This 8-bit field is the address space identifier in instructions that access alternate space.
impl-dep:
The meaning of these fields is completely implementation-dependent for IMPDEP1 and
IMPDEP2 instructions.
mmask:
This 4-bit field imposes order constraints on memory references appearing before and
after a MEMBAR instruction.
op and op2:
These 2- and 3-bit fields encode the three major formats and the Format 2 instructions. See
Appendix E, “Opcode Maps,” for descriptions of their values.
op3:
This 6-bit field (together with one bit from op) encodes the Format 3 instructions. See
Appendix E, “Opcode Maps,” for descriptions of its values.
opf:
This 9-bit field encodes the operation for a floating-point operate (FPop) instruction. See
Appendix E, “Opcode Maps,” for possible values and their meanings.
opf_cc:
Specifies the condition codes to be used in FMOVcc instructions. See cc0, cc1, and cc2
above for details.
opf_low:
This 6-bit field encodes the specific operation for a Move Floating-Point Register if Condition is satisfied (FMOVcc) or Move Floating-Point register if contents of integer register
match condition (FMOVr) instruction.
p:
This 1-bit field encodes static prediction for BPcc and FBPfcc instructions, as follows:
p
0
1

Branch prediction
Predict branch will not be taken
Predict branch will be taken

rcond:
This 3-bit field selects the register-contents condition to test for a move based on register
contents (MOVr or FMOVr) instruction or a branch on register contents with prediction
(BPr) instruction. See Appendix E, “Opcode Maps,” for descriptions of its values.
rd:
This 5-bit field is the address of the destination (or source) r or f register(s) for a load,
arithmetic, or store instruction.

rs1:
This 5-bit field is the address of the first r or f register(s) source operand.
rs2:
This 5-bit field is the address of the second r or f register(s) source operand with i = 0.
shcnt32:
This 5-bit field provides the shift count for 32-bit shift instructions.
shcnt64:
This 6-bit field provides the shift count for 64-bit shift instructions.
simm10:
This 10-bit field is an immediate value that is sign-extended to 64 bits and used as the second ALU operand for a MOVr instruction when i = 1.
simm11:
This 11-bit field is an immediate value that is sign-extended to 64 bits and used as the second ALU operand for a MOVcc instruction when i = 1.
simm13:
This 13-bit field is an immediate value that is sign-extended to 64 bits and used as the second ALU operand for an integer arithmetic instruction or for a load/store instruction when
i = 1.
sw_trap#:
This 7-bit field is an immediate value that is used as the second ALU operand for a Trap
on Condition Code instruction.
x:
The x bit selects whether a 32- or 64-bit shift will be performed..

6.3 Instruction Categories
SPARC-V9 instructions can be grouped into the following categories:
— Memory access
— Memory synchronization
— Integer arithmetic
— Control transfer (CTI)
— Conditional moves
— Register window management
— State register access
— Privileged register access
— Floating-point operate

— Implementation-dependent
— Reserved
Each of these categories is further described in the following subsections.
6.3.1 Memory Access Instructions
Load, Store, Prefetch, Load Store Unsigned Byte, Swap, and Compare and Swap are the only
instructions that access memory. All of the instructions except Compare and Swap use either two
r registers or an r register and simm13 to calculate a 64-bit byte memory address. Compare and
Swap uses a single r register to specify a 64-bit byte memory address. To this 64-bit address, the
IU appends an ASI that encodes address space information.
The destination field of a memory reference instruction specifies the r or f register(s) that supply
the data for a store or receive the data from a load or LDSTUB. For SWAP, the destination register identifies the r register to be exchanged atomically with the calculated memory location. For
Compare and Swap, an r register is specified whose value is compared with the value in memory
at the computed address. If the values are equal, the destination field specifies the r register that is
to be exchanged atomically with the addressed memory location. If the values are unequal, the
destination field specifies the r register that is to receive the value at the addressed memory location; in this case, the addressed memory location remains unchanged.
The destination field of a PREFETCH instruction is used to encode the type of the prefetch.
Integer load and store instructions support byte (8-bit), halfword (16-bit), word (32-bit), and doubleword (64-bit) accesses. Floating-point load and store instructions support word, doubleword,
and quadword memory accesses. LDSTUB accesses bytes, SWAP accesses words, and CAS
accesses words or doublewords. PREFETCH accesses at least 64 bytes.
Programming Note:
By setting i = 1 and rs1 = 0, any location in the lowest or highest 4K bytes of an address space can be
accessed without using a register to hold part of the address.

6.3.1.1 Memory Alignment Restrictions
Halfword accesses shall be aligned on 2-byte boundaries, word accesses (which include instruction fetches) shall be aligned on 4-byte boundaries, extended word and doubleword accesses shall
be aligned on 8-byte boundaries, and quadword accesses shall be aligned on 16-byte boundaries,
with the following exceptions.
An improperly aligned address in a load, store, or load-store instruction causes a
mem_address_not_aligned exception to occur, except:
— An LDDF or LDDFA instruction accessing an address that is word-aligned but not doubleword-aligned may cause an LDDF_mem_address_not_aligned exception, or may complete
the operation in hardware (impl. dep. #109).

— An STDF or STDFA instruction accessing an address that is word-aligned but not doubleword-aligned may cause an STDF_mem_address_not_aligned exception or may complete
the operation in hardware (impl. dep. #110).
— An LDQF or LDQFA instruction accessing an address that is word-aligned but not quadword-aligned may cause an LDQF_mem_address_not_aligned exception or may complete
the operation in hardware (impl. dep. #111).
— An STQF or STQFA instruction accessing an address that is word-aligned but not quadword aligned may cause an STQF_mem_address_not_aligned exception or may complete
the operation in hardware (impl. dep. #112).
6.3.1.2 Addressing Conventions
SPARC-V9 uses big-endian byte order for all instruction accesses and, by default, for data
accesses. It is possible to access data in little-endian format by using selected ASIs. It is also possible to change the default byte order for implicit data accesses. See 5.2.1, “Processor State Register (PSTATE),” for more information.1
6.3.1.2.1 Big-Endian Addressing Convention
Within a multiple-byte integer, the byte with the smallest address is the most significant; a byte’s
significance decreases as its address increases. The big-endian addressing conventions are illustrated in figure 35 and defined as follows:
byte:
A load/store byte instruction accesses the addressed byte in both big- and little-endian
modes.
halfword:
For a load/store halfword instruction, two bytes are accessed. The most significant byte
(bits 15..8) is accessed at the address specified in the instruction; the least significant byte
(bits 7..0) is accessed at the address + 1.
word:
For a load/store word instruction, four bytes are accessed. The most significant byte (bits
31..24) is accessed at the address specified in the instruction; the least significant byte
(bits 7..0) is accessed at the address + 3.
doubleword or extended word:
For a load/store extended or floating-point load/store double instruction, eight bytes are
accessed. The most significant byte (bits 63..56) is accessed at the address specified in the
instruction; the least significant byte (bits 7..0) is accessed at the address + 7.
For the deprecated integer load/store double instructions (LDD/STD), two big-endian
words are accessed. The word at the address specified in the instruction corresponds to the

1.See Cohen, D., “On Holy Wars and a Plea for Peace,” Computer 14:10 (October 1981), pp. 48-54.

even register specified in the instruction; the word at address + 4 corresponds to the following odd-numbered register.
quadword:
For a load/store quadword instruction, sixteen bytes are accessed. The most significant
byte (bits 127..120) is accessed at the address specified in the instruction; the least significant byte (bits 7..0) is accessed at the address + 15.
Byte

Address
7

Halfword

Address<0> =

0
15

Word

Address<1:0> =

8 7

00
31

Doubleword / Address<2:0> =
Extended word

000

001

Address<3:0> =

0100

Address<3:0> =

1000

Address<3:0> =

1100
31

0111

1010

1101

72 71

48 47

24 23

0011

0110

1001

104 103

80 79

56 55

111

0010

0101

8 7

112 111

88 87

011

110

0001

40 39

16 15

120 119

010

101

0000
127

11
8 7

48 47

24 23

10
16 15

56 55

100

Address<3:0> =

24 23

Address<2:0> =

Quadword

1011
40 39

1110
16 15

1111
8 7

Figure 35—Big-Endian Addressing Conventions

6.3.1.2.2 Little-Endian Addressing Convention
Within a multiple-byte integer, the byte with the smallest address is the least significant; a byte’s
significance increases as its address increases. The little-endian addressing conventions are illustrated in figure 36 and defined as follows:
byte:
A load/store byte instruction accesses the addressed byte in both big- and little-endian
modes.

halfword:
For a load/store halfword instruction, two bytes are accessed. The least significant byte
(bits 7..0) is accessed at the address specified in the instruction; the most significant byte
(bits 15..8) is accessed at the address + 1.
word:
For a load/store word instruction, four bytes are accessed. The least significant byte (bits
7..0) is accessed at the address specified in the instruction; the most significant byte (bits
31..24) is accessed at the address + 3.
doubleword or extended word:
For a load/store extended or floating-point load/store double instruction, eight bytes are
accessed. The least significant byte (bits 7..0) is accessed at the address specified in the
instruction; the most significant byte (bits 63..56) is accessed at the address + 7.
For the deprecated integer load/store double instructions (LDD/STD), two little-endian
words are accessed. The word at the address specified in the instruction + 4 corresponds to
the even register specified in the instruction; the word at the address specified in the
instruction corresponds to the following odd-numbered register.

quadword:
For a load/store quadword instruction, sixteen bytes are accessed. The least significant
byte (bits 7..0) is accessed at the address specified in the instruction; the most significant
byte (bits 127..120) is accessed at the address + 15
Byte

Address
7

Halfword

Address<0> =

0
7

Word

Address<1:0> =

000

001

100

Address<3:0> =

0100

Address<3:0> =

1000
71

Address<3:0> =

1100
103

0111
48 63

1010
72 87

1101
96 111

0011

0110

1001

16 31

40 55

64 79

111

0010

0101

48 63

8 23

32 47

011

110

0001

16 31

40 55

0 15

010

101

0000

11
16 31

8 23

32 47

10
8 23

0 15

Address<3:0> =

0 15

Address<2:0> =

Quadword

00
7

Doubleword / Address<2:0> =
Extended word

1
0

1011
80 95

1110
104 119

1111
112 127

120

Figure 36—Little-Endian Addressing Conventions

6.3.1.3 Address Space Identifiers (ASIs)
Load and store instructions provide an implicit ASI value of ASI_PRIMARY or
ASI_PRIMARY_LITTLE. Load and store alternate instructions provide an explicit ASI, specified
by the imm_asi instruction field when i = 0, or the contents of the ASI register when i = 1.
ASIs 0016 through 7F16 are restricted; only privileged software is allowed to access them. An
attempt to access a restricted ASI by nonprivileged software results in a privileged_action excep-

tion. ASIs 8016 through FF16 are unrestricted; software is allowed to access them whether the processor is operating in privileged or nonprivileged mode. This is illustrated in table 11.
Table 11—Allowed Accesses to ASIs
Value

Access Type

0016 ..7F16

Restricted

8016 ..FF16

Unrestricted

Processor state
(PSTATE.PRIV)
Nonprivileged (0)
Privileged (1)
Nonprivileged (0)
Privileged (1)

Result of ASI access
privileged_action exception

Valid access
Valid access
Valid access

The required ASI assignments are shown in table 12. In the table, “R” indicates a restricted ASI,
and “U” indicates an unrestricted ASI.
IMPL. DEP. #29: These ASI assignments are implementation-dependent: restricted ASIs 0016 ..0316,
0516 ..0B16, 0D16 ..0F16, 1216 ..1716, AND 1A16 ..7F16; and unrestricted ASIs C016 .. FF16.
IMPL. DEP. #30: An implementation may choose to decode only a subset of the 8-bit ASI specifier; however, it shall decode at least enough of the ASI to distinguish ASI_PRIMARY, ASI_PRIMARY_LITTLE,
ASI_AS_IF_USER_PRIMARY,
ASI_AS_IF_USER_PRIMARY_LITTLE,
ASI_PRIMARY_NOFAULT,
ASI_PRIMARY_NOFAULT_LITTLE,
ASI_SECONDARY,
ASI_SECONDARY_LITTLE,
ASI_AS_IF_USER_SECONDARY,
ASI_AS_IF_USER_SECONDARY_LITTLE,
ASI_SECONDARY_NOFAULT, and ASI_SECONDARY_NOFAULT_LITTLE. If the nucleus context is supported, then ASI_NUCLEUS and ASI_NUCLEUS_LITTLE must also be decoded (impl. dep. #124).
Finally, an implementation must always decode ASI bit<7> while PSTATE.PRIV = 0, so that an attempt by
nonprivileged software to access a restricted ASI will always cause a privileged_action exception.

Table 12—Address Space Identifiers (ASIs)

Value

Name

Access

Address space

0016 ..0316

—

Implementation-dependent1

0416

ASI_NUCLEUS

Implementation-dependent2

0516 ..0B16

—

Implementation-dependent1

0C16

ASI_NUCLEUS_LITTLE

Implementation-dependent2

0D16 ..0F16

—

Implementation-dependent1

1016

ASI_AS_IF_USER_PRIMARY

Primary address space, user privilege3

1116

ASI_AS_IF_USER_SECONDARY

Secondary address space, user privilege3

1216 ..1716
1816

—

Implementation-dependent1

ASI_AS_IF_USER_PRIMARY_LITTLE

Primary address space, user privilege, little-endian3

1916

ASI_AS_IF_USER_SECONDARY_LITTLE

Secondary address space, user priv., little-endian3

1A16 ..7F16

—

Implementation-dependent1

8016

ASI_PRIMARY

Primary address space

8116

ASI_SECONDARY

Secondary address space

8216

ASI_PRIMARY_NOFAULT

Primary address space, no fault4

8316

ASI_SECONDARY_NOFAULT

Secondary address space, no fault4

8416 ..8716

—

Reserved

8816

ASI_PRIMARY_LITTLE

Primary address space, little-endian

8916

ASI_SECONDARY_LITTLE

Secondary address space, little-endian

8A16

ASI_PRIMARY_NOFAULT_LITTLE

Primary address space, no fault, little-endian4

8B16

ASI_SECONDARY_NOFAULT_LITTLE

Secondary address space, no fault, little-endian4

8C16 ..BF16

—

Reserved

C016 ..FF16

—

Implementation-dependent1

These ASI assignments are implementation-dependent (impl. dep. #29) and available for use by implementors. Code that references any of these ASIs may not be portable.

2 ASI_NUCLEUS{_LITTLE} are implementation-dependent (impl. dep. #124); they may not be supported in all

implementations. See F.4.4, “Contexts,” for more information.
3

Use of these ASIs causes access checks to be performed as if the memory access instruction were issued
while PSTATE.PRIV = 0 (that is, in nonprivileged mode) and directed towards the corresponding address
space.

4 ASI_PRIMARY_NOFAULT{_LITTLE}

and ASI_SECONDARY_NOFAULT{_LITTLE} refer to the same address
spaces as ASI_PRIMARY{_LITTLE} and ASI_SECONDARY{_LITTLE}, respectively, with additional semantics as described in 8.3, “Addressing and Alternate Address Spaces.”

6.3.1.4 Separate Instruction Memory
A SPARC-V9 implementation may choose to place instruction and data in the same shared
address space and use hardware to keep the data and instruction memory consistent at all times. It
may also choose to overload independent address spaces for data and instructions and allow them
to become inconsistent when data writes are made to addresses shared with the instruction space.
A program containing such self-modifying code must issue a FLUSH instruction or appropriate
calls to system software to bring the address spaces to a consistent state. See H.1.6, “Self-Modifying Code,” for more information.

6.3.2 Memory Synchronization Instructions
Two forms of memory barrier (MEMBAR) instructions allow programs to manage the order and
completion of memory references. Ordering MEMBARs induce a partial ordering between sets of
loads and stores and future loads and stores. Sequencing MEMBARs exert explicit control over
completion of loads and stores. Both barrier forms are encoded in a single instruction, with subfunctions bit-encoded in an immediate field.
Compatibility Note:
The deprecated STBAR instruction is a subcase of the MEMBAR instruction; it is identical in operation to
the STBAR instruction of SPARC-V8, and is included only for compatibility.

6.3.3 Integer Arithmetic Instructions
The integer arithmetic instructions are generally triadic-register-address instructions that compute
a result which is a function of two source operands. They either write the result into the destination register r[rd] or discard it. One of the source operands is always r[rs1]. The other source
operand depends on the i bit in the instruction; if i = 0, the operand is r[rs2]; if i = 1, the operand
is the constant simm10, simm11, or simm13 sign-extended to 64 bits.
Note that the value of r[0] always reads as zero, and writes to it are ignored.
6.3.3.1 Setting Condition Codes
Most integer arithmetic instructions have two versions; one sets the integer condition codes (icc
and xcc) as a side effect; the other does not affect the condition codes. A special comparison
instruction for integer values is not needed, since it is easily synthesized using the “subtract and
set condition codes” (SUBcc) instruction. See G.3, “Synthetic Instructions,” for details.
6.3.3.2 Shift Instructions
Shift instructions shift an r register left or right by a constant or variable amount. None of the shift
instructions changes the condition codes.
6.3.3.3 Set High 22 Bits of Low Word
The “set high 22 bits of low word of an r register” instruction (SETHI) writes a 22-bit constant
from the instruction into bits 31 through 10 of the destination register. It clears the low-order 10
bits and high-order 32 bits, and does not affect the condition codes. Its primary use is to construct
constants in registers.
6.3.3.4 Integer Multiply/Divide
The integer multiply instruction performs a 64 × 64 → 64-bit operation; the integer divide
instructions perform 64 ÷ 64 → 64-bit operations. For compatibility with SPARC-V8,

32 × 32 → 64-bit multiply instructions, 64 ÷ 32 → 32-bit divide instructions, and the multiply
step instruction are provided. Division by zero causes a division_by_zero exception.
6.3.3.5 Tagged Add/Subtract
The tagged add/subtract instructions assume tagged-format data, in which the tag is the two loworder bits of each operand. If either of the two operands has a nonzero tag, or if 32-bit arithmetic
overflow occurs, tag overflow is detected. TADDcc and TSUBcc set the CCR.icc.V bit if tag overflow occurs; they set the CCR.xcc.V bit if 64-bit arithmetic overflow occurs. The trapping versions (TADDccTV, TSUBccTV) of these instructions cause a tag_overflow trap if tag overflow
occurs. If 64-bit arithmetic overflow occurs but tag overflow does not, TADDccTV and TSUBccTV set the CCR.xcc.V bit but do not trap.
6.3.4 Control-Transfer Instructions (CTIs)
These are the basic control-transfer instruction types:
— Conditional branch (Bicc, BPcc, BPr, FBfcc, FBPfcc)
— Unconditional Branch
— Call and Link (CALL)
— Jump and Link (JMPL, RETURN)
— Return from trap (DONE, RETRY)
— Trap (Tcc)
A control-transfer instruction functions by changing the value of the next program counter (nPC)
or by changing the value of both the program counter (PC) and the next program counter (nPC).
When only the next program counter, nPC, is changed, the effect of the transfer of control is
delayed by one instruction. Most control transfers in SPARC-V9 are of the delayed variety. The
instruction following a delayed control transfer instruction is said to be in the delay slot of the
control transfer instruction. Some control transfer instructions (branches) can optionally annul,
that is, not execute, the instruction in the delay slot, depending upon whether the transfer is taken
or not-taken. Annulled instructions have no effect upon the program-visible state nor can they
cause a trap.
Programming Note:
The annul bit increases the likelihood that a compiler can find a useful instruction to fill the delay slot after a
branch, thereby reducing the number of instructions executed by a program. For example, the annul bit can
be used to move an instruction from within a loop to fill the delay slot of the branch that closes the loop.
Likewise, the annul bit can be used to move an instruction from either the “else” or “then” branch of an “ifthen-else” program block to the delay slot of the branch that selects between them. Since a full set of conditions are provided, a compiler can arrange the code (possibly reversing the sense of the condition) so that an
instruction from either the “else” branch or the “then” branch can be moved to the delay slot.

Table 13 below defines the value of the program counter and the value of the next program
counter after execution of each instruction. Conditional branches have two forms: branches that

test a condition, represented in the table by “Bcc,” and branches that are unconditional, that is,
always or never taken, represented in the table by “B.” The effect of an annulled branch is shown
in the table through explicit transfers of control, rather than by fetching and annulling the instruction.
The effective address, EA in table 13, specifies the target of the control transfer instruction. The
effective address is computed in different ways, depending on the particular instruction:
PC-relative Effective Address:
A PC-relative effective address is computed by sign extending the instruction’s immediate
field to 64-bits, left-shifting the word displacement by two bits to create a byte displacement, and adding the result to the contents of the PC.
Register-Indirect Effective Address:
A register-indirect effective address computes its target address as either r[rs1]+r[rs2] if i
= 0, or r[rs1]+sign_ext(simm13) if i = 1.
Trap Vector Effective Address:
A trap vector effective address first computes the software trap number as the least significant seven bits of r[rs1]+r[rs2] if i = 0, or as the least significant seven bits of
r[rs1]+sw_trap# if i = 1. The trap level, TL, is incremented. The hardware trap type is
computed as 256 + sw_trap# and stored in TT[TL]. The effective address is generated by
concatenating the contents of the TBA register, the “TL>0” bit, and the contents of
TT[TL]. See 5.2.8, “Trap Base Address (TBA),” for details.
Trap State Effective Address:
A trap state effective address is not computed, but is taken directly from either TPC[TL] or
TNPC[TL].
Compatibility Note:
SPARC-V8 specified that the delay instruction was always fetched, even if annulled, and that an annulled
instruction could not cause any traps. SPARC-V9 does not require the delay instruction to be fetched if it is
annulled.
Compatibility Note:
SPARC-V8 left as undefined the result of executing a delayed conditional branch that had a delayed control
transfer in its delay slot. For this reason, programmers should avoid such constructs when backwards compatibility is an issue.

Table 13—Control Transfer Characteristics
Instruction group
Non-CTIs
Bcc
Bcc
Bcc
Bcc
B
B
B
B
CALL
JMPL, RETURN
DONE
RETRY
Tcc
Tcc

Address
form
—
PC-relative
PC-relative
PC-relative
PC-relative
PC-relative
PC-relative
PC-relative
PC-relative
PC-relative
Register-ind.
Trap state
Trap state
Trap vector
Trap vector

Delayed

Taken

—
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
No
No
No
No

—
Yes
No
Yes
No
Yes
No
Yes
No
—
—
—
—
Yes
No

Annul
bit
—
0
0
1
1
0
0
1
1
—
—
—
—
—
—

New PC

New nPC

nPC
nPC
nPC
nPC
nPC + 4
nPC
nPC
EA
nPC + 4
nPC
nPC
TNPC[TL]
TPC[TL]
EA
nPC

nPC + 4
EA
nPC + 4
EA
nPC + 8
EA
nPC + 4
EA + 4
nPC + 8
EA
EA
TNPC[TL] + 4
TNPC[TL]
EA + 4
nPC + 4

6.3.4.1 Conditional Branches
A conditional branch transfers control if the specified condition is true. If the annul bit is 0, the
instruction in the delay slot is always executed. If the annul bit is 1, the instruction in the delay
slot is not executed unless the conditional branch is taken. Note that the annul behavior of a taken
conditional branch is different from that of an unconditional branch.
6.3.4.2 Unconditional Branches
An unconditional branch transfers control unconditionally if its specified condition is “always”; it
never transfers control if its specified condition is “never.” If the annul bit is 0, the instruction in
the delay slot is always executed. If the annul bit is 1, the instruction in the delay slot is never executed. Note that the annul behavior of an unconditional branch is different from that of a taken
conditional branch.
6.3.4.3 CALL and JMPL instructions
The CALL instruction writes the contents of the PC, which points to the CALL instruction itself,
into r[15] (out register 7) and then causes a delayed transfer of control to a PC-relative effective
address. The value written into r[15] is visible to the instruction in the delay slot.
The JMPL instruction writes the contents of the PC, which points to the JMPL instruction itself,
into r[rd] and then causes a delayed transfer of control to a PC-relative effective address. The
value written into r[rd] is visible to the instruction in the delay slot.

When PSTATE.AM = 1, the value of the high order 32-bits transmitted to r[15] by the CALL
instruction or to r[rd] by the JMPL instruction is implementation-dependent. (impl. dep #125).
6.3.4.4 RETURN Instruction
The RETURN instruction is used to return from a trap handler executing in nonpriviliged mode.
RETURN combines the control-transfer characteristics of a JMPL instruction with r[0] specified
as the destination register and the register-window semantics of a RESTORE instruction.
6.3.4.5 DONE and RETRY Instructions
The DONE and RETRY instructions are used by privileged software to return from a trap. These
instructions restore the machine state to values saved in the TSTATE register.
RETRY returns to the instruction that caused the trap in order to reexecute it. DONE returns to the
instruction pointed to by the value of nPC associated with the instruction that caused the trap, that
is, the next logical instruction in the program. DONE presumes that the trap handler did whatever
was requested by the program and that execution should continue.
6.3.4.6 Trap Instruction (Tcc)
The Tcc instruction initiates a trap if the condition specified by its cond field matches the current
state of the condition code register specified by its cc field, otherwise it executes as a NOP. If the
trap is taken, it increments the TL register, computes a trap type which is stored in TT[TL], and
transfers to a computed address in the trap table pointed to by TBA. See 5.2.8, “Trap Base
Address (TBA).”
A Tcc instruction can specify one of 128 software trap types. When a Tcc is taken, 256 plus the
seven least significant bits of the sum of the Tcc’s source operands is written to TT[TL]. The only
visible difference between a software trap generated by a Tcc instruction and a hardware trap is
the trap number in the TT register. See Chapter 7, “Traps,” for more information.
Programming Note:
Tcc can be used to implement breakpointing, tracing, and calls to supervisor software. Tcc can also be used
for run-time checks, such as out-of-range array index checks or integer overflow checks.

6.3.5 Conditional Move Instructions
6.3.5.1 MOVcc and FMOVcc Instructions
The MOVcc and FMOVcc instructions copy the contents of any integer or floating-point register
to a destination integer or floating-point register if a condition is satisfied. The condition to test is
specified in the instruction and may be any of the conditions allowed in conditional delayed control-transfer instructions. This condition is tested against one of the six condition codes (icc, xcc,
fcc0, fcc1, fcc2, and fcc3) as specified by the instruction. For example:
fmovdg

%fcc2, %f20, %f22

moves the contents of the double-precision floating-point register %f20 to register %f22 if floating-point condition code number 2 (fcc2) indicates a greater-than relation (FSR.fcc2 = 2). If fcc2
does not indicate a greater-than relation (FSR.fcc2 ≠ 2), then the move is not performed.
The MOVcc and FMOVcc instructions can be used to eliminate some branches in programs. In
most implementations, branches will be more expensive than the MOVcc or FMOVcc instructions. For example, the following C statement:
if (A > B) X

= 1; else X

= 0;

can be coded as:
cmp
or
movg

%i0, %i2
%g0, 0, %i3
%xcc, %g0,1, %i3

! (A > B)
! set X = 0
! overwrite X with 1 if A > B

which eliminates the need for a branch.
6.3.5.2 MOVr and FMOVr Instructions
The MOVr and FMOVr instructions allow the contents of any integer or floating-point register to
be moved to a destination integer or floating-point register if a condition specified by the instruction is satisfied. The condition to test may be any of the following:
Condition
NZ
Z
GEZ
LZ
LEZ
GZ

Description
Nonzero
Zero
Greater than or equal to zero
Less than zero
Less than or equal to zero
Greater than zero

Any of the integer registers may be tested for one of the conditions, and the result used to control
the move. For example,
movrnz

%i2, %l4, %l6

moves integer register %l4 to integer register %l6 if integer register %i2 contains a nonzero
value.
MOVr and FMOVr can be used to eliminate some branches in programs, or to emulate multiple
unsigned condition codes by using an integer register to hold the result of a comparison.
6.3.6 Register Window Management Instructions
This subsection describes the instructions used to manage register windows in SPARC-V9. The
privileged registers affected by these instructions are described in 5.2.10, “Register-Window State
Registers.”

6.3.6.1 SAVE Instruction
The SAVE instruction allocates a new register window and saves the caller’s register window by
incrementing the CWP register.
If CANSAVE = 0, execution of a SAVE instruction causes a window_spill exception.
If CANSAVE ≠ 0, but the number of clean windows is zero, that is:
(CLEANWIN – CANRESTORE) = 0
then SAVE causes a clean_window exception.
If SAVE does not cause an exception, it performs an ADD operation, decrements CANSAVE, and
increments CANRESTORE. The source registers for the ADD are from the old window (the one
to which CWP pointed before the SAVE), while the result is written into a register in the new window (the one to which the incremented CWP points).
6.3.6.2 RESTORE Instruction
The RESTORE instruction restores the previous register window by decrementing the CWP register.
If CANRESTORE = 0, execution of a RESTORE instruction causes a window_fill exception.
If RESTORE does not cause an exception, it performs an ADD operation, decrements CANRESTORE, and increments CANSAVE. The source registers for the ADD are from the “old” window (the one to which CWP pointed before the RESTORE), while the result is written into a
register in the “new” window (the one to which the decremented CWP points).
Programming Note:
The following describes a common convention for use of register windows, SAVE, RESTORE, CALL, and
JMPL instructions.
A procedure is invoked by executing a CALL (or a JMPL) instruction. If the procedure requires a register
window, it executes a SAVE instruction. A routine that does not allocate a register window of its own (possibly a leaf procedure) should not modify any windowed registers except out registers 0 through 6. See H.1.2,
“Leaf-Procedure Optimization.”
A procedure that uses a register window returns by executing both a RESTORE and a JMPL instruction. A
procedure that has not allocated a register window returns by executing a JMPL only. The target address for
the JMPL instruction is normally eight plus the address saved by the calling instruction, that is, to the
instruction after the instruction in the delay slot of the calling instruction.
The SAVE and RESTORE instructions can be used to atomically establish a new memory stack pointer in an
r register and switch to a new or previous register window. See H.1.4, “Register Allocation within a Window.”

6.3.6.3 SAVED Instruction
The SAVED instruction should be used by a spill trap handler to indicate that a window spill has
completed successfully. It increments CANSAVE:

CANSAVE ← (CANSAVE + 1)
If the saved window belongs to a different address space (OTHERWIN ≠ 0), it decrements OTHERWIN:
OTHERWIN ← (OTHERWIN – 1)
Otherwise, the saved window belongs to the current address space (OTHERWIN = 0), so SAVED
decrements CANRESTORE:
CANRESTORE ← (CANRESTORE – 1)
6.3.6.4 RESTORED Instruction
The RESTORED instruction should be used by a fill trap handler to indicate that a window has
been filled successfully. It increments CANRESTORE:
CANRESTORE ← (CANRESTORE + 1)
If the restored window replaces a window that belongs to a different address space
(OTHERWIN ≠ 0), it decrements OTHERWIN:
OTHERWIN ← (OTHERWIN – 1)
Otherwise, the restored window belongs to the current address space (OTHERWIN = 0), so
RESTORED decrements CANSAVE:
CANSAVE ← (CANSAVE – 1)
If CLEANWIN is less than NWINDOWS-1, the RESTORED instruction increments CLEANWIN:
if (CLEANWIN < (NWINDOWS-1)) then CLEANWIN ← (CLEANWIN + 1)
6.3.6.5 Flush Windows Instruction
The FLUSHW instruction flushes all of the register windows except the current window, by performing repetitive spill traps. The FLUSHW instruction is implemented by causing a spill trap if
any register window (other than the current window) has valid contents. The number of windows
with valid contents is computed as
NWINDOWS – 2 – CANSAVE
If this number is nonzero, the FLUSHW instruction causes a spill trap. Otherwise, FLUSHW has
no effect. If the spill trap handler exits with a RETRY instruction, the FLUSHW instruction will
continue causing spill traps until all the register windows except the current window have been
flushed.

6.3.7 State Register Access
The read/write state register instructions access program-visible state and status registers. These
instructions read/write the state registers into/from r registers. A read/write Ancillary State Register instruction is privileged only if the accessed register is privileged.
6.3.8 Privileged Register Access
The read/write privileged register instructions access state and status registers that are visible only
to privileged software. These instructions read/write privileged registers into/from r registers. The
read/write privileged register instructions are privileged.
6.3.9 Floating-Point Operate (FPop) Instructions
Floating-point operate instructions (FPops) are generally triadic-register-address instructions.
They compute a result that is a function of one or two source operands and place the result in one
or more destination f registers. The exceptions are:
— Floating-point convert operations, which use one source and one destination operand
— Floating-point compare operations, which do not write to an f register, but update one of
the fccn fields of the FSR instead
The term “FPop” refers to those instructions encoded by the FPop1 and FPop2 opcodes and does
not include branches based on the floating-point condition codes (FBfcc and FBPfcc) or the load/
store floating-point instructions.
The FMOVcc instructions function for the floating-point registers as the MOVcc instructions do
for the integer registers. See 6.3.5.1, “MOVcc and FMOVcc Instructions.”
The FMOVr instructions function for the floating-point registers as the MOVr instructions do for
the integer registers. See 6.3.5.2, “MOVr and FMOVr Instructions.”
If there is no floating-point unit present or if PSTATE.PEF = 0 or FPRS.FEF = 0, any instruction
that attempts to access an FPU register, including an FPop instruction, generates an fp_disabled
exception.
All FPop instructions clear the ftt field and set the cexc field, unless they generate an exception.
Floating-point compare instructions also write one of the fccn fields. All FPop instructions that
can generate IEEE exceptions set the cexc and aexc fields, unless they generate an exception.
FABS(s,d,q), FMOV(s,d,q), FMOVcc(s,d,q), FMOVr(s,d,q), and FNEG(s,d,q) cannot generate
IEEE exceptions, so they clear cexc and leave aexc unchanged. FMOVcc and FMOVr instructions
clear these FSR fields regardless of the value of the conditional predicate.
IMPL. DEP. #3: An implementation may indicate that a floating-point instruction did not produce a correct
IEEE STD 754-1985 result by generating a special unfinished_FPop or unimplemented_FPop exception.
Privileged-mode software must emulate any functionality not present in the hardware.

6.3.10 Implementation-Dependent Instructions
SPARC-V9 provides two instructions that are entirely implementation-dependent, IMPDEP1 and
IMPDEP2 (impl. dep. #106).
Compatibility Note:
The IMPDEPn instructions replace the CPopn instructions in SPARC-V8.

See A.23, “Implementation-Dependent Instructions,” for more information.
6.3.11 Reserved Opcodes and Instruction Fields
An attempt to execute an opcode to which no instruction is assigned shall cause a trap. Specifically, attempting to execute a reserved FPop causes an fp_exception_other trap (with
FSR.ftt = unimplemented_FPop); attempting to execute any other reserved opcode shall cause an
illegal_instruction trap. See Appendix E, “Opcode Maps,” for a complete enumeration of the
reserved opcodes.

6.4 Register Window Management
The state of the register windows is determined by the contents of the set of privileged registers
described in 5.2.10, “Register-Window State Registers.” Those registers are affected by the
instructions described in 6.3.6, “Register Window Management Instructions.” Privileged software
can read/write these state registers directly by using RDPR/WRPR instructions.
6.4.1 Register Window State Definition
In order for the state of the register windows to be consistent, the following must always be true:
CANSAVE + CANRESTORE + OTHERWIN = NWINDOWS – 2
Figure 3 on page 34 shows how the register windows are partitioned to obtain the above equation.
In figure 3, the partitions are as follows:
— The current window and the window that overlaps two other valid windows and so must
not be used (in the figure, windows 0 and 4, respectively) are always present and account
for the 2 subtracted from NWINDOWS in the right-hand side of the equation.
— Windows that do not have valid contents and can be used (via a SAVE instruction) without
causing a spill trap. These windows (windows 1, 2 and 3 in the figure) are counted in
CANSAVE.
— Windows that have valid contents for the current address space and can be used (via the
RESTORE instruction) without causing a fill trap. These windows (window 7 in the figure) are counted in CANRESTORE.
— Windows that have valid contents for an address space other than the current address
space. An attempt to use these windows via a SAVE (RESTORE) instruction results in a

spill (fill) trap to a separate set of trap vectors, as discussed in the following subsection.
These windows (windows 5 and 6 in the figure) are counted in OTHERWIN.
In addition,
CLEANWIN ≥ CANRESTORE
since CLEANWIN is the sum of CANRESTORE and the number of clean windows following
CWP.
In order to use the window-management features of the architecture as described here, the state of
the register windows must be kept consistent at all times, except in trap handlers for window spilling, filling, and cleaning. While handling window traps the state may be inconsistent. Window
spill/fill strap handlers should be written such that a nested trap can be taken without destroying
state.
6.4.2 Register Window Traps
Window traps are used to manage overflow and underflow conditions in the register windows, to
support clean windows, and to implement the FLUSHW instruction.
6.4.2.1 Window Spill and Fill Traps
A window overflow occurs when a SAVE instruction is executed and the next register window is
occupied (CANSAVE = 0). An overflow causes a spill trap that allows privileged software to save
the occupied register window in memory, thereby making it available for use.
A window underflow occurs when a RESTORE instruction is executed and the previous register
window is not valid (CANRESTORE = 0). An underflow causes a fill trap that allows privileged
software to load the registers from memory.
6.4.2.2 Clean-Window Trap
SPARC-V9 provides the clean_window trap so that software can create a secure environment in
which it is guaranteed that register windows contain only data from the same address space.
A clean register window is one in which all of the registers, including uninitialized registers, contain either zero or data assigned by software executing in the address space to which the window
belongs. A clean window cannot contain register values from another process, that is, software
operating in a different address space.
Supervisor software specifies the number of windows that are clean with respect to the current
address space in the CLEANWIN register. This includes register windows that can be restored
(the value in the CANRESTORE register) and the register windows following CWP that can be
used without cleaning. Therefore, the number of clean windows that are available to be used by
the SAVE instruction is
CLEANWIN – CANRESTORE

The SAVE instruction causes a clean_window trap if this value is zero. This allows supervisor software to clean a register window before it is accessed by a user.
6.4.2.3 Vectoring of Fill/Spill Traps
In order to make handling of fill and spill traps efficient, SPARC-V9 provides multiple trap vectors for the fill and spill traps. These trap vectors are determined as follows:
— Supervisor software can mark a set of contiguous register windows as belonging to an
address space different from the current one. The count of these register windows is kept
in the OTHERWIN register. A separate set of trap vectors (fill_n_other and spill_n_other) is
provided for spill and fill traps for these register windows (as opposed to register windows
that belong to the current address space).
— Supervisor software can specify the trap vectors for fill and spill traps by presetting the
fields in the WSTATE register. This register contains two subfields, each three bits wide.
The WSTATE.NORMAL field is used to determine one of eight spill (fill) vectors to be
used when the register window to be spilled (filled) belongs to the current address space
(OTHERWIN = 0). If the OTHERWIN register is nonzero, the WSTATE.OTHER field
selects one of eight fill_n_other (spill_n_other) trap vectors.
See Chapter 7, “Traps,” for more details on how the trap address is determined.
6.4.2.4 CWP on Window Traps
On a window trap the CWP is set to point to the window that must be accessed by the trap handler,
as follows (note that all arithmetic on CWP is done modulo NWINDOWS):
— If the spill trap occurs due to a SAVE instruction (when CANSAVE = 0), there is an overlap window between the CWP and the next register window to be spilled
CWP ← (CWP + 2) mod NWINDOWS
If the spill trap occurs due to a FLUSHW instruction, there can be unused windows (CANSAVE) in addition to the overlap window, between the CWP and the window to be spilled
CWP ← (CWP + CANSAVE + 2) mod NWINDOWS
Implementation Note:
All spill traps can use:
CWP ← (CWP + CANSAVE + 2) mod NWINDOWS
since CANSAVE is zero whenever a trap occurs due to a SAVE instruction.

— On a fill trap, the window preceding CWP must be filled
CWP ← (CWP – 1) mod NWINDOWS
— On a clean_window trap, the window following CWP must be cleaned. Then
CWP ← (CWP + 1) mod NWINDOWS

6.4.2.5 Window Trap Handlers
The trap handlers for fill, spill and clean_window traps must handle the trap appropriately and
return using the RETRY instruction, to reexecute the trapped instruction. The state of the register
windows must be updated by the trap handler, and the relationship among CLEANWIN, CANSAVE, CANRESTORE, and OTHERWIN must remain consistent. The following recommendations should be followed:
— A spill trap handler should execute the SAVED instruction for each window that it spills.
— A fill trap handler should execute the RESTORED instruction for each window that it fills.
— A clean_window trap handler should increment CLEANWIN for each window that it
cleans:
CLEANWIN ← (CLEANWIN + 1)
Window trap handlers in SPARC-V9 can be very efficient. See H.2.2, “Example Code for Spill
Handler,” for details and sample code.

7 Traps
7.1 Overview
A trap is a vectored transfer of control to supervisor software through a trap table that contains the
first eight (thirty-two for fill/spill traps) instructions of each trap handler. The base address of the
table is established by supervisor software, by writing the Trap Base Address (TBA) register. The
displacement within the table is determined by the trap type and the current trap level (TL). Onehalf of the table is reserved for hardware traps; one-quarter is reserved for software traps generated by Tcc instructions; the remaining quarter is reserved for future use.
A trap behaves like an unexpected procedure call. It causes the hardware to
(1) Save certain processor state (program counters, CWP, ASI, CCR, PSTATE, and the trap
type) on a hardware register stack
(2) Enter privileged execution mode with a predefined PSTATE
(3) Begin executing trap handler code in the trap vector
When the trap handler has finished, it uses either a DONE or RETRY instruction to return.
A trap may be caused by a Tcc instruction, an SIR instruction, an instruction-induced exception, a
reset, an asynchronous exception, or an interrupt request not directly related to a particular
instruction. The processor must appear to behave as though, before executing each instruction, it
determines if there are any pending exceptions or interrupt requests. If there are pending exceptions or interrupt requests, the processor selects the highest-priority exception or interrupt request
and causes a trap.
Thus, an exception is a condition that makes it impossible for the processor to continue executing
the current instruction stream without software intervention. A trap is the action taken by the pro-

cessor when it changes the instruction flow in response to the presence of an exception, interrupt,
or Tcc instruction.
A catastrophic error exception is due to the detection of a hardware malfunction from which, due
to the nature of the error, the state of the machine at the time of the exception cannot be restored.
Since the machine state cannot be restored, execution after such an exception may not be resumable. An example of such an error is an uncorrectable bus parity error.
IMPL. DEP. #31: The causes and effects of catastrophic errors are implementation-dependent. They may
cause precise, deferred, or disrupting traps.

7.2 Processor States, Normal and Special Traps
The processor is always in one of three discrete states:
— execute_state, which is the normal execution state of the processor
— RED_state (Reset, Error, and Debug state), which is a restricted execution state reserved
for processing traps that occur when TL = MAXTL – 1, and for processing hardware- and
software-initiated resets
— error_state, which is a halted state that is entered as a result of a trap when TL = MAXTL,
or due to an unrecoverable error
Traps processed in execute_state are called normal traps. Traps processed in RED_state are
called special traps. Exceptions that cause the processor to enter error_state are recorded by the
hardware and are made available in the TT field after the processor is reset.
Figure 37 shows the processor state diagram.
Trap @
TL = MAXTL

Trap or SIR @
TL = MAXTL

Trap @
TL = MAXTL–1,

execute_state

Trap or SIR @
TL< MAXTL,
RED = 1

DONE,
RETRY,
RED = 0

RED_state

POR,
WDR,

Trap @
TL < MAXTL–1

Trap or SIR @
TL < MAXTL

XIR
Any State
Including Power Off

Figure 37—Processor State Diagram

error_state

7.2.1 RED_state
RED_state is an acronym for Reset, Error, and Debug state. The processor enters RED_state
under any one of the following conditions:
— A trap is taken when TL = MAXTL–1.
— Any of the four reset requests occurs (POR, WDR, XIR, SIR).
— An implementation-dependent trap, internal_processor_error exception, or catastrophic_error
exception occurs.
— System software sets PSTATE.RED = 1.
RED_state serves two mutually exclusive purposes:
— During trap processing, it indicates that there are no more available trap levels; that is, if
another nested trap is taken, the processor will enter error_state and halt. RED_state provides system software with a restricted execution environment.
— It provides the execution environment for all reset processing.
RED_state is indicated by PSTATE.RED. When this bit is set, the processor is in RED_state;
when this bit is clear, the processor is not in RED_state, independent of the value of TL. Executing a DONE or RETRY instruction in RED_state restores the stacked copy of the PSTATE register, which clears the PSTATE.RED flag if the stacked copy had it cleared. System software can
also set or clear the PSTATE.RED flag with a WRPR instruction, which also forces the processor
to enter or exit RED_state, respectively. In this case, the WRPR instruction should be placed in
the delay slot of a jump, so that the PC can be changed in concert with the state change.
Programming Note:
Setting TL = MAXTL with a WRPR instruction does not also set PSTATE.RED = 1; nor does it alter any
other machine state. The values of PSTATE.RED and TL are independent.

7.2.1.1 RED_state Trap Table
Traps occurring in RED_state or traps that cause the processor to enter RED_state use an abbreviated trap vector. The RED_state trap vector is constructed so that it can overlay the normal trap
vector if necessary. Figure 38 illustrates the RED_state trap vector.
Offset

Reason

0016
2016

0
1

Reserved (SPARC-V8 reset)
Power-on reset (POR)

4016

2†

Watchdog reset (WDR)

6016
8016
A016

3‡
4
*

Externally initiated reset (XIR)
Software-initiated reset (SIR)
All other exceptions in RED_state

† TT

= 2 if a watchdog reset occurs while the processor is not in error_state; TT = trap type of the exception
that caused entry into error_state if a watchdog reset (WDR) occurs in error_state.

‡ TT

= 3 if an externally initiated reset (XIR) occurs while the processor is not in error_state; TT = trap type
of the exception that caused entry into error_state if the externally initiated reset occurs in error_state.

* TT

= trap type of the exception. See table 14 on page 124.
Figure 38—RED_state Trap Vector Layout

IMPL. DEP. #114: The RED_state trap vector is located at an implementation-dependent address referred
to as RSTVaddr.
Implementation Note:
The RED_state trap handlers should be located in trusted memory, for example, in ROM. The value of RSTVaddr may be hard-wired in an implementation, but it is suggested that it be externally settable, for instance
by scan, or read from pins at power-on reset.

7.2.1.2 RED_state Execution Environment
In RED_state the processor is forced to execute in a restricted environment by overriding the values of some processor controls and state registers.
Programming Note:
The values are overridden, not set. This is to allow them to be switched atomically.
IMPL. DEP. #115: A processor’s behavior in RED_state is implementation-dependent.

The following are recommended:
(1) Instruction address translation is a straight-through physical map; that is, the MMU is
always suppressed for instruction access in RED_state.
(2) Data address translation is handled normally; that is, the MMU is used if it is enabled.
However, any event that causes the processor to enter RED_state also disables the MMU.
The handler executing in RED_state can reenable the MMU.
(3) All references are uncached.

(4) Cache coherence in RED_state is the problem of the system designer and system programmer. Normally, cache enables are left unchanged by RED_state; thus, if a cache is enabled,
it will continue to participate in cache coherence until explicitly disabled by recovery
code. A cache may be disabled automatically if an error is detected in the cache.
(5) Unessential functional units (for example, the floating-point unit) and capabilities (for
example, superscalar execution) should be disabled.
(6) If a store buffer is present, it should be emptied, if possible, before entering RED_state.
(7) PSTATE.MM is set to TSO.
Programming Note:
When RED_state is entered due to component failures, the handler should attempt to recover from potentially catastrophic error conditions or to disable the failing components. When RED_state is entered after a
reset, the software should create the environment necessary to restore the system to a running state.

7.2.1.3 RED_state Entry Traps
The following traps are processed in RED_state in all cases
— POR (Power-on reset)
— WDR (Watchdog reset)
— XIR (Externally initiated reset)
In addition, the following trap is processed in RED_state if TL < MAXTL when the trap is taken.
Otherwise it is processed in error_state.
— SIR (Software-initiated Reset)
An implementation also may elect to set PSTATE.RED = 1 after an internal_processor_error trap
(impl. dep. #31), or any of the implementation-dependent traps (impl. dep. #35).
Implementation-dependent traps may force additional state changes, such as disabling failing
components.
Traps that occur when TL = MAXTL – 1 also set PSTATE.RED = 1; that is, any trap handler
entered with TL = MAXTL runs in RED_state.
Any nonreset trap that sets PSTATE.RED = 1 or that occurs when PSTATE.RED = 1, branches to
a special entry in the RED_state trap vector at RSTVaddr + A016.
In systems in which it is desired that traps not enter RED_state, the RED_state handler may transfer to the normal trap vector by executing the following code:
! Assumptions:
!
-- In RED_state handler, therefore we know that
!
PSTATE.RED = 1, so a WRPR can directly toggle it to 0
!
and, we don’t have to worry about intervening traps.
!
-- Registers %g1 and %g2 are available as scratch registers.
...
#define PSTATE_RED
0x0020
! PSTATE.RED is bit 5

...
rdpr
rdpr
add
rdpr
jmpl
wrpr

%tt,%g1
%tba,%g2
%g1,%g2,%g2
%pstate,%g1
%g2
%g1,PSTATE_RED,%pstate

! Get the normal trap vector
! address in %g2.
! Read PSTATE into %g1.
! Jump to normal trap vector,
! toggling PSTATE.RED to 0.

7.2.1.4 RED_state Software Considerations
In effect, RED_state reserves one level of the trap stack for recovery and reset processing. Software should be designed to require only MAXTL – 1 trap levels for normal processing. That is,
any trap that causes TL = MAXTL is an exceptional condition that should cause entry to
RED_state.
Since the minimum value for MAXTL is 4, typical usage of the trap levels is as follows:
TL

Usage

0
1
2
3
4

Normal execution
System calls; interrupt handlers; instruction emulation
Window spill / fill
Page-fault handler
RED_state handler

Programming Note:
In order to log the state of the processor, RED_state-handler software needs either a spare register or a preloaded pointer to a save area. To support recovery, the operating system might reserve one of the alternate
global registers, (for example, %a7) for use in RED_state.

7.2.2 Error_state
The processor enters error_state when a trap occurs while the processor is already at its maximum
supported trap level, that is, when TL = MAXTL.
IMPL. DEP. #39: The processor may enter error_state when an implementation-dependent error condition
occurs.
IMPL. DEP. #40: Effects when error_state is entered are implementation-dependent, but it is recommended that as much processor state as possible be preserved upon entry to error_state.

In particular:
(1) The processor should present an external indication that it has entered error_state.
(2) The processor should halt, that is, make no further changes to system state.
(3) The processor should be restarted by a watchdog reset (WDR). Alternatively, it may be
restarted by an externally initiated reset (XIR) or a power-on reset (POR).

After a reset that brings the processor out of error_state, the processor enters RED_state with TL
set as defined in table 18 on page 130; the trap state describes the state at the time of entry into
error_state. In particular, for WDR and XIR, TT is set to the value of the original trap that caused
entry to error_state, not the normal TT value for the WDR or XIR.

7.3 Trap Categories
An exception or interrupt request can cause any of the following trap types:
— A precise trap
— A deferred trap
— A disrupting trap
— A reset trap
7.3.1 Precise Traps
A precise trap is induced by a particular instruction and occurs before any program-visible state
has been changed by the trap-inducing instruction. When a precise trap occurs, several conditions
must be true.
— The PC saved in TPC[TL] points to the instruction that induced the trap, and the nPC
saved in NTPC[TL] points to the instruction that was to be executed next.
— All instructions issued before the one that induced the trap have completed execution.
— Any instructions issued after the one that induced the trap remain unexecuted.
Programming Note:
Among the actions the trap handler software might take after a precise trap are:

— Return to the instruction that caused the trap and reexecute it, by executing a RETRY instruction (PC
← old PC, nPC ← old nPC)
— Emulate the instruction that caused the trap and return to the succeeding instruction by executing a
DONE instruction (PC ← old nPC, nPC ← old nPC + 4)
— Terminate the program or process associated with the trap

7.3.2 Deferred Traps
A deferred trap is also induced by a particular instruction, but unlike a precise trap, a deferred
trap may occur after program-visible state has been changed. Such state may have been changed
by the execution of either the trap-inducing instruction itself or by one or more other instructions.
If an instruction induces a deferred trap and a precise trap occurs simultaneously, the deferred trap
may not be deferred past the precise trap, except that a floating-point exception may be deferred
past a precise trap.
Associated with a particular deferred-trap implementation, there must exist:
— An instruction that causes a potentially outstanding deferred-trap exception to be taken as
a trap.
— Privileged instructions that access the deferred-trap queues. This queue contains the state
information needed by supervisor software to emulate the deferred-trap-inducing instruction, and to resume execution of the trapped instruction stream. See 5.2.13, “IU DeferredTrap Queue.”)
Note that resuming execution may require the emulation of instructions that had not completed
execution at the time of the deferred trap, that is, those instructions in the deferred-trap queue.
IMPL. DEP. #32: Whether any deferred traps (and associated deferred-trap queues) are present is implementation-dependent.

Note that to avoid deferred traps entirely, an implementation would need to execute all implemented floating-point instructions synchronously with the execution of integer instructions, causing all generated exceptions to be precise. A deferred-trap queue (e.g., FQ) would be superfluous
in such an implementation.
Programming Note:
Among the actions software can take after a deferred trap are:
— Emulate the instruction that caused the exception, emulate or cause to execute any other executiondeferred instructions that were in an associated deferred-trap state queue, and use RETRY to return control to the instruction at which the deferred trap was invoked, or
— Terminate the program or process associated with the trap.

7.3.3 Disrupting Traps
A disrupting trap is neither a precise trap nor a deferred trap. A disrupting trap is caused by a
condition (e.g., an interrupt), rather than directly by a particular instruction; this distinguishes it
from precise and deferred traps. When a disrupting trap has been serviced, program execution
resumes where it left off. This differentiates disrupting traps from reset traps, which resume execution at the unique reset address.
Disrupting traps are controlled by a combination of the Processor Interrupt Level (PIL) register
and the Interrupt Enable (IE) field of PSTATE. A disrupting trap condition is ignored when inter-

rupts are disabled (PSTATE.IE = 0) or when the condition’s interrupt level is lower than that specified in PIL.
A disrupting trap may be due to either an interrupt request not directly related to a previously executed instruction, or to an exception related to a previously executed instruction. Interrupt
requests may be either internal or external. An interrupt request can be induced by the assertion of
a signal not directly related to any particular processor or memory state. Examples of this are the
assertion of an “I/O done” signal or setting external interrupt request lines.
A disrupting trap related to an earlier instruction causing an exception is similar to a deferred trap
in that it occurs after instructions following the trap-inducing instruction have modified the processor or memory state. The difference is that the condition which caused the instruction to induce
the trap may lead to unrecoverable errors, since the implementation may not preserve the necessary state. An example of this is an ECC data-access error reported after the corresponding load
instruction has completed.
Disrupting trap conditions should persist until the corresponding trap is taken.
Programming Note:
Among the actions that trap-handler software might take after a disrupting trap are:
— Use RETRY to return to the instruction at which the trap was invoked
(PC ← old PC, nPC ← old nPC), or
— Terminate the program or process associated with the trap.

7.3.4 Reset Traps
A reset trap occurs when supervisor software or the implementation’s hardware determines that
the machine must be reset to a known state. Reset traps differ from disrupting traps, since they do
not resume execution of the program that was running when the reset trap occurred.
IMPL. DEP. #37: Some of a processor’s behavior during a reset trap is implementation-dependent.

See

7.6.2, “Special Trap Processing,” for details.
The following reset traps are defined for SPARC-V9:
Software-initiated reset (SIR):
Initiated by software by executing the SIR instruction.
Power-on reset (POR):
Initiated when power is applied (or reapplied) to the processor.
Watchdog reset (WDR):
Initiated in response to error_state or expiration of a watchdog timer.
Externally initiated reset (XIR):
Initiated in response to an external signal. This reset trap is normally used for critical system events, such as power failure.

7.3.5 Uses of the Trap Categories
The SPARC-V9 trap model stipulates that:
(1) Reset traps, except software_initiated_reset traps, occur asynchronously to program execution.
(2) When recovery from an exception can affect the interpretation of subsequent instructions,
such exceptions shall be precise. These exceptions are:
— software_initiated_reset
— instruction_access_exception
— privileged_action
— privileged_opcode
— trap_instruction
— instruction_access_error
— clean_window
— fp_disabled
— LDDF_mem_address_not_aligned
— STDF_mem_address_not_aligned
— tag_overflow
— unimplemented_LDD
— unimplemented_STD
— spill_n_normal
— spill_n_other
— fill_n_normal

— fill_n_other
(3) IMPL. DEP. #33: Exceptions that occur as the result of program execution may be precise or
deferred, although it is recommended that such exceptions be precise. Examples:
mem_address_not_aligned, division_by_zero.
(4) An exception caused after the initial access of a multiple-access load or store instruction
(load-store doubleword, LDSTUB, CASA, CASXA, or SWAP) that causes a catastrophic
exception may be precise, deferred, or disrupting. Thus, a trap due to the second memory
access can occur after the processor or memory state has been modified by the first access.
(5) Implementation-dependent catastrophic exceptions may cause precise, deferred, or disrupting traps (impl. dep. #31).
(6) Exceptions caused by external events unrelated to the instruction stream, such as interrupts, are disrupting.

For the purposes of this subsection, we must distinguish between the dispatch of an instruction
and its execution by some functional unit. An instruction is deemed to have been dispatched
when the software-visible PC advances beyond that instruction in the instruction stream. An
instruction is deemed to have been executed when the results of that instruction are available to
subsequent instructions.
For most instructions, dispatch and execution appear to occur simultaneously; when the PC has
advanced beyond the instruction, its results are immediately available to subsequent instructions.
For floating-point instructions, however, the PC may advance beyond the instruction as soon as
the IU places the instruction into a floating-point queue; the instruction itself may not have completed (or even begun) execution, and results may not be available to subsequent instructions for
some time. In particular, the fact that a floating-point instruction will generate an exception may
not be noticed by the hardware until additional floating-point instructions have been placed into
the queue by the IU. This creates the condition for a deferred trap.
A deferred trap may occur one or more instructions after the trap-inducing instruction is dispatched. However, a deferred trap must occur before the execution (but not necessarily the dispatch) of any instruction that depends on the trap-inducing instruction. That is, a deferred trap
may not be deferred past the execution of an instruction that specifies source registers, destination
registers, condition codes, or any software-visible machine state that could be modified by the
trap-inducing instruction.
In the case of floating-point instructions, if a floating-point exception is currently deferred, an
attempt to dispatch a floating-point instruction (FPop, FBfcc, FBPfcc, or floating-point load/store)
invokes or causes the outstanding fp_exception_ieee_754 trap.
Implementation Note:
To provide the capability to terminate a user process on the occurrence of a catastrophic exception that can
cause a deferred or disrupting trap, an implementation should provide one or more instructions that provoke
an outstanding exception to be taken as a trap. For example, an outstanding floating-point exception might
cause an fp_exception_ieee_754 trap when any of an FPop, load or store floating-point register (including
the FSR), FBfcc, or FBPfcc instruction is executed.

7.4 Trap Control
Several registers control how any given trap is processed:
— The interrupt enable (IE) field in PSTATE and the processor interrupt level (PIL) register
control interrupt processing.
— The enable floating-point unit (FEF) field in FPRS, the floating-point unit enable (PEF)
field in PSTATE, and the trap enable mask (TEM) in the FSR control floating-point traps.
— The TL register, which contains the current level of trap nesting, controls whether a trap
causes entry to execute_state, RED_state, or error_state.
— PSTATE.TLE determines whether implicit data accesses in the trap routine will be performed using the big- or little-endian byte order.
7.4.1 PIL Control
Between the execution of instructions, the IU prioritizes the outstanding exceptions and interrupt
requests according to table 15. At any given time, only the highest priority exception or interrupt
request is taken as a trap.1 When there are multiple outstanding exceptions or interrupt requests,

1. The highest priority exception or interrupt is the one with the lowest priority value in table 15. For example, a
priority 2 exception is processed before a priority 3 exception.

SPARC-V9 assumes that lower-priority interrupt requests will persist and lower-priority exceptions will recur if an exception-causing instruction is reexecuted.
For interrupt requests, the IU compares the interrupt request level against the processor interrupt
level (PIL) register. If the interrupt request level is greater than PIL, the processor takes the interrupt request trap, assuming there are no higher-priority exceptions outstanding
IMPL. DEP. #34: How quickly a processor responds to an interrupt request and the method by which an
interrupt request is removed are implementation-dependent.

7.4.2 TEM Control
The occurrence of floating-point traps of type IEEE_754_exception can be controlled with the useraccessible trap enable mask (TEM) field of the FSR. If a particular bit of TEM is 1, the associated
IEEE_754_exception can cause an fp_exception_ieee_754 trap.
If a particular bit of TEM is 0, the associated IEEE_754_exception does not cause an
fp_exception_ieee_754 trap. Instead, the occurrence of the exception is recorded in the FSR’s
accrued exception field (aexc).
If an IEEE_754_exception results in an fp_exception_ieee_754 trap, then the destination f register,
fccn, and aexc fields remain unchanged. However, if an IEEE_754_exception does not result in a
trap, then the f register, fccn, and aexc fields are updated to their new values.

7.5 Trap-Table Entry Addresses
Privileged software initializes the trap base address (TBA) register to the upper 49 bits of the traptable base address. Bit 14 of the vector address (the “TL>0” field) is set based on the value of TL
at the time the trap is taken; that is, to 0 if TL = 0 and to 1 if TL > 0. Bits 13..5 of the trap vector
address are the contents of the TT register. The lowest five bits of the trap address, bits 4..0, are
always 0 (hence each trap-table entry is at least 25 or 32 bytes long). Figure 39 illustrates this.
TBA<63:15>
63

TL>0
15

Figure 39—Trap Vector Address

TTTL
13

00000
5

7.5.1 Trap Table Organization
The trap table layout is as illustrated in figure 40.
Trap Table Contents
Value of TL
Before the Trap
TL = 0

TL > 0

Trap Type

Hardware traps

00016 ..07F16

Spill/fill traps

08016 ..0FF16

Software traps

10016 ..17F16

Reserved

18016 ..1FF16

Hardware traps

20016 ..27F16

Spill/fill traps

28016 ..2FF16

Software traps

30016 ..37F16

Reserved

38016 ..3FF16

Figure 40—Trap Table Layout

The trap table for TL = 0 comprises 512 32-byte entries; the trap table for TL > 0 comprises 512
more 32-byte entries. Therefore, the total size of a full trap table is 512 × 32 × 2, or 32K bytes.
However, if privileged software does not use software traps (Tcc instructions) at TL > 0, the table
can be made 24K bytes long.
7.5.2 Trap Type (TT)
When a normal trap occurs, a value that uniquely identifies the trap is written into the current 9-bit
TT register (TT[TL]) by hardware. Control is then transferred into the trap table to an address
formed by the TBA register (“TL>0”) and TT[TL] (see 5.2.8, “Trap Base Address (TBA)”).
Since the lowest five bits of the address are always zero, each entry in the trap table may contain
the first eight instructions of the corresponding trap handler.
Programming Note:
The spill/fill and clean_window trap types are spaced such that their trap table entries are 128 bytes (32
instructions) long. This allows the complete code for one spill/fill or clean_window routine to reside in one
trap table entry.

When a special trap occurs, the TT register is set as described in 7.2.1, “RED_state.” Control is
then transferred into the RED_state trap table to an address formed by the RSTVaddr and an offset
depending on the condition.
TT values 00016 ..0FF16 are reserved for hardware traps. TT values 10016 ..17F16 are reserved for
software traps (traps caused by execution of a Tcc instruction). TT values 18016 ..1FF16 are
reserved for future uses. The assignment of TT values to traps is shown in table 14; table 15 lists
the traps in priority order. Traps marked with an open bullet ‘❍’ are optional and possibly implementation-dependent. Traps marked with a closed bullet ‘●’ are mandatory; that is, hardware
must detect and trap these exceptions and interrupts and must set the defined TT values.
The trap type for the clean_window exception is 02416. Three subsequent trap vectors
(02516 ..02716) are reserved to allow for an inline (branchless) trap handler. Window spill/fill traps
are described in 7.5.2.1. Three subsequent trap vectors are reserved for each spill/fill vector, to
allow for an inline (branchless) trap handler.

Table 14—Exception and Interrupt Requests, Sorted by TT Value
M/O
●
●
❍
❍
●
●
●
●
❍
❍
●
●
●
❍
❍
●
●
❍
❍
●
❍
●
❍
●
●
❍
❍
❍
●
❍
❍
●
❍
❍
●
❍
●
●
❍
●
●
●
●
●

Exception or interrupt request
Reserved
power_on_reset
watchdog_reset
externally_initiated_reset
software_initiated_reset
RED_state_exception

Reserved
instruction_access_exception
instruction_access_MMU_miss
instruction_access_error

Reserved
illegal_instruction
privileged_opcode
unimplemented_LDD
unimplemented_STD

Reserved
fp_disabled
fp_exception_ieee_754
fp_exception_other
tag_overflow
clean_window
division_by_zero
internal_processor_error

Reserved
data_access_exception
data_access_MMU_miss
data_access_error
data_access_protection
mem_address_not_aligned
LDDF_mem_address_not_aligned (impl. dep. #109)
STDF_mem_address_not_aligned (impl. dep. #110)
privileged_action
LDQF_mem_address_not_aligned (impl. dep. #111)
STQF_mem_address_not_aligned (impl. dep. #112)

Reserved
async_data_error
interrupt_level_n (n = 1..15)

Reserved
implementation_dependent_exception_n (impl. dep. #35)
spill_n_normal (n = 0..7)
spill_n_other (n = 0..7)
fill_n_normal (n = 0..7)
fill_n_other (n = 0..7)
trap_instruction

Priority

00016
00116
00216
00316
00416
00516
00616 ..00716
00816
00916
00A16
00B16 ..00F16
01016
01116
01216
01316
01416 ..01F16
02016
02116
02216
02316
02416 ..02716
02816
02916
02A16 ..02F16
03016
03116
03216
03316
03416
03516
03616
03716
03816
03916
03A16 ..03F16
04016
04116 ..04F16
05016 ..05F16
06016 ..07F16
08016 ..09F16
0A016 ..0BF16
0C016 ..0DF16
0E016 ..0FF16
10016 ..17F16

n/a
0
1
1
1
1
n/a
5
2
3
n/a
7
6
6
6
n/a
8
11
11
14
10
15
4
n/a
12
12
12
12
10
10
10
11
10
10
n/a
2
32–n
n/a
impl.-dep.
9
9
9
9
16

Table 14—Exception and Interrupt Requests, Sorted by TT Value
M/O
●

Exception or interrupt request
Reserved

Priority

18016 ..1FF16

n/a

Table 15—Exception and Interrupt Requests, Sorted by Priority (0 = Highest; 31 = Lowest)
M/O

Exception or Interrupt Request

Priority

●

power_on_reset

00116

❍

watchdog_reset

00216

❍

externally_initiated_reset

00316

●

software_initiated_reset

00416

●

RED_state_exception

00516

❍

instruction_access_MMU_miss

00916

❍

async_data_error

04016

❍

instruction_access_error

00A16

❍

internal_processor_error

02916

●

instruction_access_exception

00816

●

privileged_opcode

01116

❍

unimplemented_LDD

01216

❍

unimplemented_STD

01316

●

illegal_instruction

01016

●

fp_disabled

02016

●

spill_n_normal (n = 0..7)

08016 ..09F16

●

spill_n_other (n = 0..7)

0A016 ..0BF16

●

fill_n_normal (n = 0..7)

0C016 ..0DF16

●

fill_n_other (n = 0..7)

0E016 ..0FF16

❍

clean_window

02416 ..02716

●

mem_address_not_aligned

03416

❍

LDDF_mem_address_not_aligned (impl. dep. #109)

03516

❍

STDF_mem_address_not_aligned (impl. dep. #110)

03616

❍

LDQF_mem_address_not_aligned (impl. dep. #111)

03816

❍

STQF_mem_address_not_aligned (impl. dep. #112)

03916

❍

fp_exception_ieee_754

02116

❍

fp_exception_other

02216

●

privileged_action

03716

●

data_access_exception

03016

❍

data_access_MMU_miss

03116

❍

data_access_error

03216

❍

data_access_protection

03316

●

tag_overflow

02316

●

division_by_zero

02816

●

trap_instruction

10016 ..17F16

●

interrupt_level_n (n = 1..15)

04116 ..04F16

32–n

❍

implementation_dependent_exception_n (impl. dep. #35)

06016 ..07F16

impl.-dep.

Compatibility Note:
Support for some trap types is optional because they are associated with specific instruction(s), which, in a
given implementation, might be implemented purely in software. In such a case, hardware would never generate that type of trap; therefore, support for it would be superfluous. Examples of trap types to which this
applies are fp_exception_ieee_754 and fp_exception_other.

Since the assignment of exceptions and interrupt requests to particular trap vector addresses and
the priority levels are not visible to a user program, an implementation is allowed to define additional hardware traps.
IMPL. DEP. #35: TT values 06016 TO 07F16 are reserved for implementation-dependent exceptions. The
existence of implementation_dependent_n traps and whether any that do exist are precise, deferred, or
disrupting is implementation-dependent. See Appendix C, “SPARC-V9 Implementation Dependen-

cies.”
Trap Type values marked “Reserved” in table 14 are reserved for future versions of the architecture.
7.5.2.1 Trap Type for Spill/Fill Traps
The trap type for window spill/fill traps is determined based on the contents of the OTHERWIN
and WSTATE registers as follows:
Trap Type

SPILL_OR_FILL
8

OTHER

WTYPE
4

The fields have the following values:
SPILL_OR_FILL:
0102 for spill traps; 0112 for fill traps
OTHER:
(OTHERWIN≠0)
WTYPE:
If (OTHER) then WSTATE.OTHER else WSTATE.NORMAL
7.5.3 Trap Priorities
Table 14 shows the assignment of traps to TT values and the relative priority of traps and interrupt
requests. Priority 0 is highest, priority 31 is lowest; that is, if X < Y, a pending exception or interrupt request with priority X is taken instead of a pending exception or interrupt request with priority Y.
IMPL. DEP. #36: The priorities of particular traps are relative and are implementation-dependent, because
a future version of the architecture may define new traps, and an implementation may define implementation-dependent traps that establish new relative priorities.

However, the TT values for the exceptions and interrupt requests shown in table 14 must remain
the same for every implementation.

7.6 Trap Processing
The processor’s action during trap processing depends on the trap type, the current level of trap
nesting (given in the TL register), and the processor state. All traps use normal trap processing,
except those due to reset requests, catastrophic errors, traps taken when TL = MAXTL – 1, and
traps taken when the processor is in RED_state. These traps use special RED_state trap processing.
During normal operation, the processor is in execute_state. It processes traps in execute_state and
continues.
When a normal trap or software-initiated reset (SIR) occurs with TL = MAXTL, there are no
more levels on the trap stack, so the processor enters error_state and halts. In order to avoid this
catastrophic failure, SPARC-V9 provides the RED_state processor state. Traps processed in
RED_state use a special trap vector and a special trap-vectoring algorithm. RED_state vectoring
and the setting of the TT value for RED_state traps are described in 7.2.1, “RED_state.”
Traps that occur with TL = MAXTL – 1 are processed in RED_state. In addition, reset traps are
also processed in RED_state. Reset trap processing is described in 7.6.2, “Special Trap Processing.” Finally, supervisor software can force the processor into RED_state by setting the
PSTATE.RED flag to one.
Once the processor has entered RED_state, no matter how it got there, all subsequent traps are
processed in RED_state until software returns the processor to execute_state or a normal or SIR
trap is taken when TL = MAXTL, which puts the processor in error_state. Tables 16, 17, and 18
describe the processor mode and trap level transitions involved in handling traps:
Table 16—Trap Received While in execute_state
New State, after receiving trap type
Original state

Normal trap
or interrupt

POR

WDR, XIR,
Impl. Dep.

SIR

execute_state

RED_state

TL < MAXTL – 1
execute_state

TL + 1
RED_state

MAXTL
RED_state

TL + 1
RED_state

TL = MAXTL – 1

MAXTL

execute_state†

error_state

RED_state

error_state

TL = MAXTL

MAXTL

†This

state occurs when software changes TL to MAXTL and does not set PSTATE.RED, or if it clears
PSTATE.RED while at MAXTL.

Table 17—Trap Received While in RED_state
New State, after receiving trap type
Original state

Normal trap
or interrupt

POR

WDR, XIR,
Impl. Dep.

SIR

RED_state

TL < MAXTL – 1
RED_state

TL + 1
RED_state

MAXTL
RED_state

TL + 1
RED_state

TL = MAXTL – 1
RED_state

MAXTL
error_state

MAXTL
RED_state

MAXTL
error_state

TL = MAXTL

MAXTL

Table 18—Reset Received While in error_state
New State, after receiving trap type
Original state
error_state
TL < MAXTL – 1
error_state
TL = MAXTL – 1
error_state
TL = MAXTL

Normal trap
or interrupt
—
—
—

POR

WDR, XIR,
Impl. Dep.

RED_state

MAXTL
RED_state

TL + 1
RED_state

MAXTL
RED_state

MAXTL

SIR
—
—
—

Implementation Note:
The processor shall not recognize interrupts while it is in error_state.

7.6.1 Normal Trap Processing
A normal trap causes the following state changes to occur:
— If the processor is already in RED_state, the new trap is processed in RED_state unless
TL = MAXTL. See 7.6.2.6, “Normal Traps When the Processor is in RED_state.”
— If the processor is in execute_state and the trap level is one less than its maximum value,
that is, TL = MAXTL–1, the processor enters RED_state. See 7.2.1, “RED_state,” and
7.6.2.1, “Normal Traps with TL = MAXTL – 1.”
— If the processor is in either execute_state or RED_state, and the trap level is already at its
maximum value, that is, TL = MAXTL, the processor enters error_state. See 7.2.2,
“Error_state.”
Otherwise, the trap uses normal trap processing, and the following state changes occur:
— The trap level is set. This provides access to a fresh set of privileged trap-state registers
used to save the current state, in effect, pushing a frame on the trap stack.
TL ← TL + 1

— Existing state is preserved
TSTATE[TL].CCR
TSTATE[TL].ASI
TSTATE[TL].PSTATE
TSTATE[TL].CWP
TPC[TL]
TNPC[TL]

←
←
←
←
←
←

CCR
ASI
PSTATE
CWP
PC
nPC

— The trap type is preserved.
TT[TL] ← the trap type
— The PSTATE register is updated to a predefined state
PSTATE.MM
is unchanged
PSTATE.RED
← 0
PSTATE.PEF
← 1 if FPU is present, 0 otherwise
PSTATE.AM
← 0 (address masking is turned off)
PSTATE.PRIV
← 1 (the processor enters privileged mode)
PSTATE.IE
← 0 (interrupts are disabled)
PSTATE.AG
← 1 (global regs are replaced with alternate globals)
PSTATE.CLE
← PSTATE.TLE (set endian mode for traps)
— For a register-window trap only, CWP is set to point to the register window that must be
accessed by the trap-handler software, that is:
• If TT[TL] = 02416 (a clean_window trap), then CWP ← CWP + 1.
• If (08016 ≤ TT[TL] ≤ 0BF16) (window spill trap), then CWP ←
CWP + CANSAVE + 2.
• If (0C016 ≤ TT[TL] ≤ 0FF16) (window fill trap), then CWP ← CWP–1.
For non-register-window traps, CWP is not changed.
— Control is transferred into the trap table:
PC ← TBA<63:15> (TL>0) TT[TL]
nPC ← TBA<63:15> (TL>0) TT[TL]
where “(TL>0)” is 0 if TL = 0, and 1 if TL > 0.

0 0000
0 0100

Interrupts are ignored as long as PSTATE.IE = 0.
Programming Note:
State in TPC[n], TNPC[n], TSTATE[n], and TT[n] is only changed autonomously by the processor when a
trap is taken while TL = n–1, however, software can change any of these values with a WRPR instruction
when TL = n.

7.6.2 Special Trap Processing
The following conditions invoke special trap processing:
— Traps taken with TL = MAXTL – 1
— Power-on reset traps
— Watchdog reset traps
— Externally initiated reset traps
— Software-initiated reset traps
— Traps taken when the processor is already in RED_state
— Implementation-dependent traps
IMPL. DEP. #38: Implementation-dependent registers may or may not be affected by the various reset
traps.

7.6.2.1 Normal Traps with TL = MAXTL – 1
Normal traps that occur when TL = MAXTL – 1 are processed in RED_state. The following state
changes occur:
— The trap level is advanced.
TL ← MAXTL
— Existing state is preserved
TSTATE[TL].CCR
TSTATE[TL].ASI
TSTATE[TL].PSTATE
TSTATE[TL].CWP
TPC[TL]
TNPC[TL]

←
←
←
←
←
←

CCR
ASI
PSTATE
CWP
PC
nPC

— The trap type is preserved.
TT[TL] ← the trap type
— The PSTATE register is set as follows:
PSTATE.MΜ
← 002 (TSO)
PSTATE.RED
← 1 (enter RED_state)
PSTATE.PEF
← 1 if FPU is present, 0 otherwise
PSTATE.AM
← 0 (address masking is turned off)
PSTATE.PRIV
← 1 (the processor enters privileged mode)
PSTATE.IE
← 0 (interrupts are disabled)
PSTATE.AG
← 1 (global regs are replaced with alternate globals)
PSTATE.CLE
← PSTATE.TLE (set endian mode for traps)

— For a register-window trap only, CWP is set to point to the register window that must be
accessed by the trap-handler software, that is:
• If TT[TL] = 02416 (a clean_window trap), then CWP ← CWP + 1.
• If (08016 ≤ TT[TL] ≤ 0BF16) (window spill trap), then CWP ←
CWP + CANSAVE + 2.
• If (0C016 ≤ TT[TL] ≤ 0FF16)(window fill trap), then CWP ← CWP–1.
For non-register-window traps, CWP is not changed.
— Implementation-specific state changes; for example, disabling an MMU
— Control is transferred into the RED_state trap table
PC ← RSTVaddr<63:8> 1010 00002
nPC ← RSTVaddr<63:8>

1010 01002

7.6.2.2 Power-On Reset (POR) Traps
Initiated when power is applied to the processor. If the processor is in error_state, a power-on
reset (POR) brings the processor out of error_state and places it in RED_state. Processor state is
undefined after POR, except for the following:
— The trap level is set.
TL ← MAXTL
— The trap type is set.
TT[TL] ← 00116
— The PSTATE register is set as follows:
PSTATE.MM
← 002 (TSO)
PSTATE.RED
← 1 (enter RED_state)
PSTATE.PEF
← 1 if FPU is present, 0 otherwise
PSTATE.AM
← 0 (address masking is turned off)
PSTATE.PRIV
← 1 (the processor enters privileged mode)
PSTATE.IE
← 0 (interrupts are disabled)
PSTATE.AG
← 1 (global regs are replaced with alternate globals)
PSTATE.TLE
← 0 (big-endian mode for traps)
PSTATE.CLE
← 0 (big-endian mode for non-traps)
— The TICK register is protected.
TICK.NPT ← 1 (TICK unreadable by nonprivileged software)
— Implementation-specific state changes; for example, disabling an MMU
— Control is transferred into the RED_state trap table
PC ← RSTVaddr<63:8> 0010 00002
nPC ← RSTVaddr<63:8> 0010 01002

For any reset when TL = MAXTL, for all n 0100 00002
nPC ← RSTVaddr<63:8> 0100 01002

If a watchdog reset occurs when the processor is in error_state, the TT field gives the type of the
trap that caused entry into error_state. If a watchdog reset occurs with the processor in
execute_state, TT is set to 2 (WDR).
For any reset when TL = MAXTL, for all n 0110 00002
nPC ← RSTVaddr<63:8> 0110 01002

TT is set in the same manner as for watchdog reset. If the processor is in execute_state when the
externally initiated reset (XIR) occurs, TT = 3. If the processor is in error_state when the XIR
occurs, TT identifies the exception that caused entry into error_state.
For any reset when TL = MAXTL, for all n 1000 00002
nPC ← RSTVaddr<63:8> 1000 01002

For any reset when TL = MAXTL, for all n < MAXTL, the values in TPC[n], TNPC[n], and
TSTATE[n] are undefined.
7.6.2.6 Normal Traps When the Processor is in RED_state
Normal traps taken when the processor is already in RED_state are also processed in RED_state,
unless TL = MAXTL, in which case the processor enters error_state.
The processor state shall be set as follows:
— The trap level is set.
TL ← TL + 1
— Existing state is preserved.
TSTATE[TL].CCR
TSTATE[TL].ASI
TSTATE[TL].PSTATE
TSTATE[TL].CWP
TPC[TL]
TNPC[TL]

←
←
←
←
←
←

CCR
ASI
PSTATE
CWP
PC
nPC

— The trap type is preserved.
TT[TL] ← trap type
— The PSTATE register is set as follows:
PSTATE.MM
← 002 (TSO)
PSTATE.RED
← 1 (enter RED_state)
PSTATE.PEF
← 1 if FPU is present, 0 otherwise
PSTATE.AM
← 0 (address masking is turned off)
PSTATE.PRIV
← 1 (the processor enters privileged mode)
PSTATE.IE
← 0 (interrupts are disabled)
PSTATE.AG
← 1 (global regs are replaced with alternate globals)
PSTATE.CLE
← PSTATE.TLE (set endian mode for traps)
— For a register-window trap only, CWP is set to point to the register window that must be
accessed by the trap-handler software, that is:
• If TT[TL] = 02416 (a clean_window trap), then CWP ← CWP + 1.
• If (08016 ≤ TT[TL] ≤ 0BF16) (window spill trap), then
CWP ← CWP + CANSAVE + 2.
• If (0C016 ≤ TT[TL] ≤ 0FF16) (window fill trap), then CWP ← CWP – 1.
For non-register-window traps, CWP is not changed.
— Implementation-specific state changes; for example, disabling an MMU
— Control is transferred into the RED_state trap table
PC ← RSTVaddr<63:8> 1010 00002
nPC ← RSTVaddr<63:8>

1010 01002

7.6.2.7 Implementation-Dependent Traps
The operation of the processor for implementation_dependent_exception_n traps is implementationdependent (impl. dep. #35).

7.7 Exception and Interrupt Descriptions
The following paragraphs describe the various exceptions and interrupt requests and the conditions that cause them. Each exception and interrupt request describes the corresponding trap type
as defined by the trap model. An open bullet ‘❍’ identifies optional and possibly implementationdependent traps; traps marked with a closed bullet ‘●’ are mandatory. Each trap is marked as precise, deferred, disrupting, or reset. Example exception conditions are included for each exception
type. Appendix A, “Instruction Definitions,” enumerates which traps can be generated by each
instruction.
❍ async_data_error [tt = 04016] (Disrupting)
An asynchronous data error occurred on a data access. Examples: an ECC error occurred
while writing data from a cache store buffer to memory, or an ECC error occurred on an
MMU hardware table walk. When an async_data_error occurs, the TPC and TNPC stacked
by the trap are not necessarily related to the instruction or data access that caused the
error; that is, async_data_error causes a disrupting trap.
Compatibility Note:
The SPARC-V9 async_data_error exception supersedes the less general SPARC-V8 data_store_error
exception.

❍ clean_window [tt = 02416 ..02716] (Precise)
A SAVE instruction discovered that the window about to be used contains data from
another address space; the window must be cleaned before it can be used.
IMPL. DEP. #102: An implementation may choose either to implement automatic cleaning of register windows in hardware, or to generate a clean_window trap, when needed, so that window(s) can
be cleaned by software. If an implementation chooses the latter option, then support for this

trap type is mandatory.
❍ data_access_error [tt = 03216] (Precise, Deferred, or Disrupting)
A catastrophic error exception occurred on a data access from/to memory (for example, a
parity error on a data cache access, or an uncorrectable ECC memory error) (impl. dep.
#31).
● data_access_exception [tt = 03016] (Precise or Deferred)
An exception occurred on a data access. For example, an MMU indicated that a page was
invalid or protected (impl. dep. #33).
❍ data_access_MMU_miss [tt = 03116] (Precise or Deferred)
A miss in an MMU occurred on a data access from/to memory. For example, a page
descriptor cache or translation lookaside buffer did not contain a translation for the virtual
address. (impl. dep. #33)

❍ data_access_protection [tt = 03316] (Precise or Deferred)
A protection fault occurred on a data access; for example, an MMU indicated that the page
was write-protected (impl. dep. #33).
● division_by_zero [tt = 02816] (Precise or Deferred)
An integer divide instruction attempted to divide by zero (impl. dep. #33).
❍ externally_initiated_reset [tt = 00316] (Reset)
An external signal was asserted. This trap is used for catastrophic events such as power
failure, reset button pressed, and system-wide reset in multiprocessor systems.
● fill_n_normal [tt = 0C016 ..0DF16] (Precise)
● fill_n_other [tt = 0E016 ..0FF16] (Precise)
A RESTORE or RETURN instruction has determined that the contents of a register window must be restored from memory.
Compatibility Note:
The SPARC-V9 fill_n_* exceptions supersede the SPARC-V8 window_underflow exception.

● fp_disabled [tt = 02016] (Precise)
An attempt was made to execute an FPop, a floating-point branch, or a floating-point load/
store instruction while an FPU was not present, PSTATE.PEF = 0, or FPRS.FEF = 0.
❍ fp_exception_ieee_754 [tt = 02116] (Precise or Deferred (impl. dep. #23))
An FPop instruction generated an IEEE_754_exception and its corresponding trap enable
mask (TEM) bit was 1. The floating-point exception type, IEEE_754_exception, is encoded
in the FSR.ftt, and specific IEEE_754_exception information is encoded in FSR.cexc.
❍ fp_exception_other [tt = 02216] (Precise or Deferred (impl. dep. #23))
An FPop instruction generated an exception other than an IEEE_754_exception. For example, the FPop is unimplemented, or the FPop did not complete, or there was a sequence or
hardware error in the FPU. The floating-point exception type is encoded in the FSR’s ftt
field.
● illegal_instruction [tt = 01016] (Precise or Deferred)
An attempt was made to execute an instruction with an unimplemented opcode, an ILLTRAP instruction, an instruction with invalid field usage, or an instruction that would
result in illegal processor state. Note that unimplemented FPop instructions generate
fp_exception_other traps.
❍ implementation_dependent_exception_n [tt = 06016 ..07F16] (Pre, Def, or Dis)
These exceptions are implementation-dependent (impl. dep. #35).
❍ instruction_access_error [tt = 00A16] (Precise, Deferred, or Disrupting)
A catastrophic error exception occurred on an instruction access. For example, a parity
error on an instruction cache access (impl. dep. #31).

● instruction_access_exception [tt = 00816] (Precise)
An exception occurred on an instruction access. For example, an MMU indicated that the
page was invalid or not executable.
❍ instruction_access_MMU_miss [tt = 00916] (Precise, Deferred, or Disrupting)
A miss in an MMU occurred on an instruction access from memory. For example, a PDC
or TLB did not contain a translation for the virtual address. (impl. dep. #33)
❍ internal_processor_error [tt = 02916] (Precise, Deferred, or Disrupting)
A catastrophic error exception occurred in the main processor. For example, a parity or
uncorrectable ECC error on an internal register or bus (impl. dep. #31).
Compatibility Note:
The SPARC-V9

internal_processor_error
r_register_access_error exception.

exception

supersedes

the

less

general

SPARC-V8

● interrupt_level_n [tt = 04116 ..04F16] (Disrupting)
An interrupt request level of n was presented to the IU, while PSTATE.IE = 1 and (interrupt request level > PIL).
❍ LDDF_mem_address_not_aligned [tt = 03516] (Precise)
An attempt was made to execute an LDDF instruction and the effective address was wordaligned but not doubleword-aligned. Use of this exception is implementation-dependent
(impl. dep. #109). A separate trap entry for this exception supports fast software emulation
of the LDDF instruction when the effective address is word-aligned but not doublewordaligned. See A.25, “Load Floating-Point.”
❍ LDQF_mem_address_not_aligned [tt = 03816] (Precise)
An attempt was made to execute an LDQF instruction and the effective address was wordaligned but not quadword-aligned. Use of this exception is implementation-dependent
(impl. dep. #111). A separate trap entry for this exception supports fast software emulation
of the LDQF instruction when the effective address is word-aligned but not quadwordaligned. See A.25, “Load Floating-Point.”
● mem_address_not_aligned [tt = 03416] (Precise or Deferred)
A load/store instruction generated a memory address that was not properly aligned according to the instruction, or a JMPL or RETURN instruction generated a non-word-aligned
address (impl. dep. #33).
● power_on_reset [tt = 00116] (Reset)
An external signal was asserted. This trap isused to bring a system reliably from the
power-off to the power-on state.
● privileged_action [tt = 03716] (Precise)
An action defined to be privileged has been attempted while PSTATE.PRIV = 0. Examples: a data access by nonprivileged software using an ASI value with its most significant
bit = 0 (a restricted ASI), or an attempt to read the TICK register by nonprivileged software when TICK.NPT = 1.

● privileged_opcode [tt = 01116] (Precise)
An attempt was made to execute a privileged instruction while PSTATE.PRIV = 0.
Compatibility Note:
This trap type is identical to the SPARC-V8 privileged_instruction trap. The name was changed to distinguish it from the new privileged_action trap type.

● software_initiated_reset [tt = 00416] (Reset)
Caused by the execution of the SIR, Software-Initiated Reset, instruction. It allows system
software to reset the processor.
● spill_n_normal [tt = 08016 ..09F16] (Precise)
● spill_n_other [tt = 0A016 ..0BF16] (Precise)
A SAVE or FLUSHW instruction has determined that the contents of a register window
must be saved to memory.
Compatibility Note:
The SPARC-V9 spill_n_* exceptions supersede the SPARC-V8 window_overflow exception.

❍ STDF_mem_address_not_aligned [tt = 03616] (Precise)
An attempt was made to execute an STDF instruction and the effective address was wordaligned but not doubleword-aligned. Use of this exception is implementation-dependent
(impl. dep. #110). A separate trap entry for this exception supports fast software emulation
of the STDF instruction when the effective address is word-aligned but not doublewordaligned. See A.52, “Store Floating-Point.”
❍ STQF_mem_address_not_aligned [tt = 03916] (Precise)
An attempt was made to execute an STQF instruction and the effective address was wordaligned but not quadword-aligned. Use of this exception is implementation-dependent
(impl. dep. #112). A separate trap entry for this exception supports fast software emulation
of the STQF instruction when the effective address is word-aligned but not quadwordaligned. See A.52, “Store Floating-Point.”
● tag_overflow [tt = 02316] (Precise)
A TADDccTV or TSUBccTV instruction was executed, and either 32-bit arithmetic overflow occurred or at least one of the tag bits of the operands was nonzero.
● trap_instruction [tt = 10016 ..17F16] (Precise)
A Tcc instruction was executed and the trap condition evaluated to TRUE.
❍ unimplemented_LDD [tt = 01216] (Precise)
An attempt was made to execute an LDD instruction, which is not implemented in hardware on this implementation (impl. dep. #107).
❍ unimplemented_STD [tt = 01316] (Precise)
An attempt was made to execute an STD instruction which is not implemented in hardware on this implementation (impl. dep. #108).

● watchdog_reset [tt = 00216] (Precise)
An external signal was asserted. This trap exists to break a system deadlock created when
an expected external event does not happen within the expected time. In simple systems it
is also used to bring a system out of error_state, through RED_state, and ultimately back
to execute_state.
All other trap types are reserved.

8 Memory Models
8.1 Introduction
The SPARC-V9 memory models define the semantics of memory operations. The instruction set
semantics require that loads and stores seem to be performed in the order in which they appear in
the dynamic control flow of the program. The actual order in which they are processed by the
memory may be different. The purpose of the memory models is to specify what constraints, if
any, are placed on the order of memory operations.
The memory models apply both to uniprocessor and to shared-memory multiprocessors. Formal
memory models are necessary in order to precisely define the interactions between multiple processors and input/output devices in a shared-memory configuration. Programming shared-memory multiprocessors requires a detailed understanding of the operative memory model and the
ability to specify memory operations at a low level in order to build programs that can safely and
reliably coordinate their activities. See Appendix J, “Programming With the Memory Models,”
for additional information on the use of the models in programming real systems.
The SPARC-V9 architecture is a model that specifies the behavior observable by software on
SPARC-V9 systems. Therefore, access to memory can be implemented in any manner, as long as
the behavior observed by software conforms to that of the models described here and formally
defined in Appendix D, “Formal Specification of the Memory Models.”
The SPARC-V9 architecture defines three different memory models: Total Store Order (TSO),
Partial Store Order (PSO), and Relaxed Memory Order (RMO). All SPARC-V9 processors
must provide Total Store Order (or a more strongly ordered model, for example, Sequential Consistency) to ensure SPARC-V8 compatibility.
IMPL. DEP. 113: Whether the PSO or RMO models are supported is implementation-dependent.

Figure 41 shows the relationship of the various SPARC-V9 memory models, from the least
restrictive to the most restrictive. Programs written assuming one model will function correctly on
any included model.

RMO

PSO

TSO

Figure 41—Memory Models from Least Restrictive (RMO) to Most Restrictive (TSO)

SPARC-V9 provides multiple memory models so that:
— Implementations can schedule memory operations for high performance.
— Programmers can create synchronization primitives using shared memory.
These models are described informally in this subsection and formally in Appendix D, “Formal
Specification of the Memory Models.” If there is a conflict in interpretation between the informal
description provided here and the formal models, the formal models supersede the informal
description.
There is no preferred memory model for SPARC-V9. Programs written for Relaxed Memory
Order will work in Partial Store Order and Total Store Order as well. Programs written for Partial
Store Order will work in Total Store Order. Programs written for a weak model, such as RMO,
may execute more quickly, since the model exposes more scheduling opportunities, but may also
require extra instructions to ensure synchronization. Multiprocessor programs written for a stronger model will behave unpredictably if run in a weaker model.
Machines that implement sequential consistency (also called strong ordering or strong consistency) automatically support programs written for TSO, PSO, and RMO. Sequential consistency
is not a SPARC-V9 memory model. In sequential consistency, the loads, stores, and atomic loadstores of all processors are performed by memory in a serial order that conforms to the order in
which these instructions are issued by individual processors. A machine that implements sequential consistency may deliver lower performance than an equivalent machine that implements a
weaker model. Although particular SPARC-V9 implementations may support sequential consistency, portable software must not rely on having this model available.

8.2 Memory, Real Memory, and I/O Locations
Memory is the collection of locations accessed by the load and store instructions (described in
Appendix A, “Instruction Definitions”). Each location is identified by an address consisting of
two elements: an address space identifier (ASI), which identifies an address space, and a 64-bit
address, which is a byte offset into that address space. Memory addresses may be interpreted by
the memory subsystem to be either physical addresses or virtual addresses; addresses may be

remapped and values cached, provided that memory properties are preserved transparently and
coherency is maintained.
When two or more data addresses refer to the same datum, the address is said to be aliased. In this
case, the processor and memory system must cooperate to maintain consistency; that is, a store to
an aliased address must change all values aliased to that address.
Memory addresses identify either real memory or I/O locations.
Real memory stores information without side effects. A load operation returns the value most
recently stored. Operations are side-effect-free in the sense that a load, store, or atomic load-store
to a location in real memory has no program-observable effect, except upon that location.
I/O locations may not behave like memory and may have side effects. Load, store, and atomic
load-store operations performed on I/O locations may have observable side effects and loads may
not return the value most recently stored. The value semantics of operations on I/O locations are
not defined by the memory models, but the constraints on the order in which operations are performed is the same as it would be if the I/O locations were real memory. The storage properties,
contents, semantics, ASI assignments, and addresses of I/O registers are implementation-dependent (impl. dep. #6) (impl. dep. #7) (impl. dep. #123).
IMPL. DEP. #118: The manner in which I/O locations are identified is implementation-dependent.

See F.3.2, “Attributes the MMU Associates with Each Mapping,” for example.
IMPL. DEP #120: The coherence and atomicity of memory operations between processors and I/O DMA
memory accesses are implementation-dependent.
Compatibility Note:
Operations to I/O locations are not guaranteed to be sequentially consistent between themselves, as they are
in SPARC-V8.
SPARC-V9 does not distinguish real memory from I/O locations in terms of ordering. All references, both to
I/O locations and real memory, conform to the memory model’s order constraints. References to I/O locations may need to be interspersed with MEMBAR instructions to guarantee the desired ordering. Loads following stores to locations with side effects may return unexpected results due to lookaside into the
processor’s store buffer, which may subsume the memory transaction. This can be avoided by using a MEMBAR #LookAside.
Systems supporting SPARC-V8 applications that use memory mapped I/O locations must ensure that
SPARC-V8 sequential consistency of I/O locations can be maintained when those locations are referenced
by a SPARC-V8 application. The MMU either must enforce such consistency or cooperate with system software and/or the processor to provide it.
IMPL. DEP #121: An implementation may choose to identify certain addresses and use an implementation-dependent memory model for references to them.

For example, an implementation might choose to process addresses tagged with a flag bit in the
memory management unit (see Appendix F, “SPARC-V9 MMU Requirements”), or to treat those
that utilize a particular ASI (see 8.3, “Addressing and Alternate Address Spaces,” below) as using
a sequentially consistent model.

8.3 Addressing and Alternate Address Spaces
An address in SPARC-V9 is a tuple consisting of an 8-bit address space identifier (ASI) and a 64bit byte-address offset in the specified address space. Memory is byte-addressed, with halfword
accesses aligned on 2-byte boundaries, word accesses (which include instruction fetches) aligned
on 4-byte boundaries, extended-word and doubleword accesses aligned on 8-byte boundaries, and
quadword quantities aligned on 16-byte boundaries. With the possible exception of the cases
described in 6.3.1.1, “Memory Alignment Restrictions,” an improperly aligned address in a load,
store, or load-store instruction always causes a trap to occur. The largest datum that is guaranteed
to be atomically read or written is an aligned doubleword. Also, memory references to different
bytes, halfwords, and words in a given doubleword are treated for ordering purposes as references
to the same location. Thus, the unit of ordering for memory is a doubleword.
Programming Note:
While the doubleword is the coherency unit for update, programmers should not assume that doubleword
floating-point values are updated as a unit unless they are doubleword-aligned and always updated using
double-precision loads and stores. Some programs use pairs of single-precision operations to load and store
double-precision floating-point values when the compiler cannot determine that they are doublewordaligned. Also, while quad-precision operations are defined in the SPARC-V9 architecture, the granularity of
loads and stores for quad-precision floating-point values may be word or doubleword.

The processor provides an address space identifier with every address. This ASI may serve several
purposes:
— To identify which of several distinguished address spaces the 64-bit address offset is to be
interpreted as addressing
— To provide additional access control and attribute information, for example, the processing
which is to be taken if an access fault occurs or to specify the endian-ness of the reference
— To specify the address of an internal control register in the processor, cache, or memory
management hardware
The memory management hardware can associate an independent 264-byte memory address space
with each ASI. If this is done, it becomes possible to allow system software easy access to the
address space of the faulting program when processing exceptions, or to implement access to a
client program’s memory space by a server program.
The architecturally specified ASIs are listed in table 12 on page 75. ASIs need not be fully
decoded by the hardware (impl. dep. #30). In particular, specifying an architecturally undefined
ASI value in a memory reference instruction or in the ASI register may produce unexpected
implementation-dependent results.
When TL = 0, normal accesses by the processor to memory when fetching instructions and performing loads and stores implicitly specify ASI_PRIMARY or ASI_PRIMARY_LITTLE,
depending on the setting of the PSTATE.CLE bit.
IMPL. DEP. #124: When TL > 0, the implicit ASI for instruction fetches, loads, and stores is implementation-dependent.

Implementation Note:
Implementations that support the nucleus context should use ASI_NUCLEUS{_LITTLE}; those that do not
should use ASI_PRIMARY{_LITTLE}. See F.4.4, “Contexts,” for more information about the nucleus context.

Accesses to other address spaces use the load/store alternate instructions. For these accesses, the
ASI is either contained in the instruction (for the register-register addressing mode) or taken from
the ASI register (for register-immediate addressing).
ASIs are either unrestricted or restricted. An unrestricted ASI is one that may be used independent
of the privilege level (PSTATE.PRIV) at which the processor is running. Restricted ASIs require
that the processor be in privileged mode for a legal access to occur. Restricted ASIs have their
high-order bit equal to zero. The relationship between processor state and ASI restriction is shown
in table 11 on page 74.
Several restricted ASIs must be provided: ASI_AS_IF_USER_PRIMARY{_LITTLE} and
ASI_AS_IF_USER_SECONDARY{_LITTLE}. The intent of these ASIs is to give system software efficient access to the memory space of a program.
The normal address space is primary address space, which is accessed by the unrestricted
ASI_PRIMARY{_LITTLE}. The secondary address space, which is accessed by the unrestricted
ASI_SECONDARY{_LITTLE}, is provided to allow a server program to access a client program’s address space.
ASI_PRIMARY_NOFAULT{_LITTLE} and ASI_SECONDARY_NOFAULT{_LITTLE} support nonfaulting loads. These ASIs are aliased to ASI_PRIMARY{_LITTLE} and
ASI_SECONDARY{_LITTLE}, respectively, and have exactly the same action. They may be
used to color (that is, distinguish into classes) loads in the instruction stream so that, in combination with a judicious mapping of low memory and a specialized trap handler, an optimizing compiler can move loads outside of conditional control structures.
Programming Note:
Nonfaulting loads allow optimizations that move loads ahead of conditional control structures which guard
their use; thus, they can minimize the effects of load latency by improving instruction scheduling. The
semantics of nonfaulting load are the same as for any other load, except when non-recoverable catastrophic
faults occur (for example, address-out-of-range errors). When such a fault occurs, it is ignored and the hardware and system software cooperate to make the load appear to complete normally, returning a zero result.
The compiler’s optimizer generates load-alternate instructions with the ASI field or register set to
ASI_PRIMARY_NOFAULT{_LITTLE} or ASI_SECONDARY_NOFAULT{_LITTLE} for those loads it
determines should be nonfaulting. To minimize unnecessary processing if a fault does occur, it is desirable to
map low addresses (especially address zero) to a page of all zeros, so that references through a NULL
pointer do not cause unnecessary traps.
Implementation Note:
An implementation, through a combination of hardware and system software, must prevent nonfaulting
loads on memory locations that have side effects; otherwise, such accesses produce undefined results.

8.4 The SPARC-V9 Memory Model
The SPARC-V9 processor architecture specifies the organization and structure of a SPARC-V9
central processing unit, but does not specify a memory system architecture. Appendix F,

“SPARC-V9 MMU Requirements,” summarizes the MMU support required by a SPARC-V9 central processing unit.
The memory models specify the possible order relationships between memory-reference instructions issued by a processor and the order and visibility of those instructions as seen by other processors. The memory model is intimately intertwined with the program execution model for
instructions.
8.4.1 The SPARC-V9 Program Execution Model
The SPARC-V9 processor model consists of three units: an issue unit, a reorder unit, and an execute unit, as shown in figure 42.
The issue unit reads instructions over the instruction path from memory and issues them in program order. Program order is precisely the order determined by the control flow of the program
and the instruction semantics, under the assumption that each instruction is performed independently and sequentially.
Issued instructions are collected, reordered, and then dispatched to the execute unit. Instruction
reordering allows an implementation to perform some operations in parallel and to better allocate
resources. The reordering of instructions is constrained to ensure that the results of program execution are the same as they would be if the instructions were performed in program order. This
property is called processor self-consistency.
Processor
Data Path
Issue

Reorder

Execute

Instruction Path

Memory

Figure 42—Processor Model: Uniprocessor System

Processor self-consistency requires that the result of execution, in the absence of any shared memory interaction with another processor, be identical to the result that would be observed if the
instructions were performed in program order. In the model in figure 42, instructions are issued in
program order and placed in the reorder buffer. The processor is allowed to reorder instructions,
provided it does not violate any of the data-flow constraints for registers or for memory.
The data-flow order constraints for register reference instructions are:
— An instruction cannot be performed until all earlier instructions that set a register it uses
have been performed (read-after-write hazard; write-after-write hazard).
— An instruction cannot be performed until all earlier instructions that use a register it sets
have been performed (write-after-read hazard).

An implementation can avoid blocking instruction execution in the second case by using a renaming mechanism which provides the old value of the register to earlier instructions and the new
value to later uses.
The data-flow order constraints for memory-reference instructions are those for register reference
instructions, plus the following additional constraints:
(1) A memory-reference instruction that sets (stores to) a location cannot be performed until
all previous instructions that use (load from) the location have been performed (writeafter-read hazard).
(2) A memory-reference instruction that uses (loads) the value at a location cannot be performed until all earlier memory-reference instructions that set (store to) the location have
been performed (read-after-write hazard).
As with the case for registers, implementations can avoid blocking instructions in case (2) by providing an additional mechanism, in this case, a write buffer which guarantees that the value
returned by a load is that which would be returned by the most recent store, even though the store
has not completed. As a result, the value associated with an address may appear to be different
when observed from a processor that has written the location and is holding the value in its write
buffer than it would be when observed from a processor that references memory (or its own write
buffer). Moreover, the load that was satisfied by the write buffer never appears at the memory.
Memory-barrier instructions (MEMBAR and STBAR) and the active memory model specified by
PSTATE.MM also constrain the issue of memory-reference instructions. See 8.4.3, “The MEMBAR Instruction,” and 8.4.4, “Memory Models,” for a detailed description.
The constraints on instruction execution assert a partial ordering on the instructions in the reorder
buffer. Every one of the several possible orderings is a legal execution ordering for the program.
See Appendix D, “Formal Specification of the Memory Models,” for more information.

8.4.2 The Processor/Memory Interface Model
Each processor in a multiprocessor system is modelled as shown in figure 43; that is, having two
independent paths to memory: one for instructions and one for data. Caches and mappings are
considered to be part of the memory. Data caches are maintained by hardware to be consistent
(coherent). Instruction caches need not be kept consistent with data caches and, therefore, require
explicit program action to ensure consistency when a program modifies an executing instruction
stream. Memory is shared in terms of address space, but may be inhomogeneous and distributed
in an implementation. Mapping and caches are ignored in the model, since their functions are
transparent to the memory model.1
In real systems addresses may have attributes that the processor must respect. The processor executes loads, stores, and atomic load-stores in whatever order it chooses, as constrained by program order and the current memory model. The ASI address-couples it generates are translated by
1. The model described here is only a model. Implementations of SPARC-V9 systems are unsonstrained so long
as their observable behaviors match those of the model.

a memory management unit (MMU), which associates attributes with the address and may, in
some instances, abort the memory transaction and signal an exception to the CPU. For example, a
region of memory may be marked as non-prefetchable, non-cacheable, read-only, or restricted. It
is the MMU’s responsibility, working in conjunction with system software, to ensure that memory
attribute constraints are not violated. See Appendix F, “SPARC-V9 MMU Requirements,” for
more information.
Instructions are performed in an order constrained by local dependencies. Using this dependency
ordering, an execution unit submits one or more pending memory transactions to the memory. The
memory performs transactions in memory order. The memory unit may perform transactions
submitted to it out of order; hence, the execution unit must not submit two or more transactions
concurrently that are required to be ordered.
Memory Transactions
In Memory Order

Processors

Instructions
Data

Memory

Figure 43—Data Memory Paths: Multiprocessor System

The memory accepts transactions, performs them, and then acknowledges their completion. Multiple memory operations may be in progress at any time and may be initiated in a nondeterministic
fashion in any order, provided that all transactions to a location preserve the per-processor partial
orders. Memory transactions may complete in any order. Once initiated, all memory operations
are performed atomically: loads from one location all see the same value, and the result of stores
are visible to all potential requestors at the same instant.
The order of memory operations observed at a single location is a total order that preserves the
partial orderings of each processor’s transactions to this address. There may be many legal total
orders for a given program’s execution.

8.4.3 The MEMBAR Instruction
MEMBAR serves two distinct functions in SPARC-V9. One variant of the MEMBAR, the ordering MEMBAR, provides a way for the programmer to control the order of loads and stores issued
by a processor. The other variant of MEMBAR, the sequencing MEMBAR, allows the programmer to explicitly control order and completion for memory operations. Sequencing MEMBARs
are needed only when a program requires that the effect of an operation become globally visible,

rather than simply being scheduled.1 As both forms are bit-encoded into the instruction, a single
MEMBAR can function both as an ordering MEMBAR and as a sequencing MEMBAR.

8.4.3.1 Ordering MEMBAR Instructions
Ordering MEMBAR instructions induce an ordering in the instruction stream of a single processor. Sets of loads and stores that appear before the MEMBAR in program order are ordered with
respect to sets of loads and stores that follow the MEMBAR in program order. Atomic operations
(LDSTUB(A), SWAP(A), CASA, and CASXA) are ordered by MEMBAR as if they were both a
load and a store, since they share the semantics of both. An STBAR instruction, with semantics
that are a subset of MEMBAR, is provided for SPARC-V8 compatibility. MEMBAR and STBAR
operate on all pending memory operations in the reorder buffer, independent of their address or
ASI, ordering them with respect to all future memory operations. This ordering applies only to
memory-reference instructions issued by the processor issuing the MEMBAR. Memory-reference
instructions issued by other processors are unaffected.
The ordering relationships are bit-encoded as shown in table 19. For example, MEMBAR 0116,
written as “membar #LoadLoad” in assembly language, requires that all load operations
appearing before the MEMBAR in program order complete before any of the load operations following the MEMBAR in program order complete. Store operations are unconstrained in this case.
MEMBAR 0816 (#StoreStore) is equivalent to the STBAR instruction; it requires that the values stored by store instructions appearing in program order prior to the STBAR instruction be visible to other processors prior to issuing any store operations that appear in program order
following the STBAR.
In table 19 these ordering relationships are specified by the ‘ 0
r[rs1] ≥ 0

Operation
Reserved
Branch on Register Zero
Branch on Register Less Than or Equal to Zero
Branch on Register Less Than Zero
Reserved
Branch on Register Not Zero
Branch on Register Greater Than Zero
Branch on Register Greater Than or Equal to Zero

Format (2):
00

rcond

31 30 29 28 27

25 24

011

d16hi

22 21 20 19 18

rs1

d16lo
14 13

Suggested Assembly Language Syntax
brz{,a}{,pt|,pn}
regrs1, label
brlez{,a}{,pt|,pn}
regrs1, label
brlz{,a}{,pt|,pn}
regrs1, label
brnz{,a}{,pt|,pn}
regrs1, label
brgz{,a}{,pt|,pn}
regrs1, label
brgez{,a}{,pt|,pn}
regrs1, label
Programming Note:
To set the annul bit for BPr instructions, append “,a” to the opcode mnemonic. For example, use “brz,a
%i3,label.” The preceding table indicates that the “,a” is optional by enclosing it in braces. To set the
branch prediction bit “p,” append either “,pt” for predict taken or “,pn” for predict not taken to the
opcode mnemonic. If neither “,pt” nor “,pn” is specified, the assembler shall default to “,pt”.

Description:
These instructions branch based on the contents of r[rs1]. They treat the register contents as a
signed integer value.
A BPr instruction examines all 64 bits of r[rs1] according to the rcond field of the instruction,
producing either a TRUE or FALSE result. If TRUE, the branch is taken; that is, the instruction
causes a PC-relative, delayed control transfer to the address “PC + (4 * sign_ext(d16hi
d16lo)).” If FALSE, the branch is not taken.
If the branch is taken, the delay instruction is always executed, regardless of the value of the annul
bit. If the branch is not taken and the annul bit (a) is 1, the delay instruction is annulled (not executed).

The predict bit (p) is used to give the hardware a hint about whether the branch is expected to be
taken. A 1 in the p bit indicates that the branch is expected to be taken; a 0 indicates that the
branch is expected not to be taken.
Annulment, delay instructions, prediction, and delayed control transfers are described further in
Chapter 6, “Instructions.”
Implementation Note:
If this instruction is implemented by tagging each register value with an N (negative) bit and Z (zero) bit, the
following table can be used to determine if rcond is TRUE:
Branch
BRNZ
BRZ
BRGEZ
BRLZ
BRLEZ
BRGZ

Exceptions:
illegal_instruction (if rcond = 0002 or 1002)

Test
not Z
Z
not N
N
N or Z
not (N or Z)

A.4 Branch on Floating-Point Condition Codes (FBfcc)
The FBfcc instructions are deprecated; they are provided only for compatibility
with previous versions of the architecture. They should not be used in new SPARCV9 software. It is recommended that the FBPfcc instructions be used in their place.

Opcode
FBAD
FBND
FBUD
FBGD
FBUGD
FBLD
FBULD
FBLGD
FBNED
FBED
FBUED
FBGED
FBUGED
FBLED
FBULED
FBOD

cond
1000
0000
0111
0110
0101
0100
0011
0010
0001
1001
1010
1011
1100
1101
1110
1111

Operation

fcc test

Branch Always
Branch Never
Branch on Unordered
Branch on Greater
Branch on Unordered or Greater
Branch on Less
Branch on Unordered or Less
Branch on Less or Greater
Branch on Not Equal
Branch on Equal
Branch on Unordered or Equal
Branch on Greater or Equal
Branch on Unordered or Greater or Equal
Branch on Less or Equal
Branch on Unordered or Less or Equal
Branch on Ordered

1
0
U
G
G or U
L
L or U
L or G
L or G or U
E
E or U
E or G
E or G or U
E or L
E or L or U
E or L or G

Format (2):
00

31 30 29 28

cond

110
25 24

disp22
22 21

Suggested Assembly Language Syntax
fba{,a}
label
fbn{,a}
label
fbu{,a}
label
fbg{,a}
label
fbug{,a}
label
fbl{,a}
label
fbul{,a}
label
fblg{,a}
label
fbne{,a}
label
(synonym: fbnz)
fbe{,a}
label
(synonym: fbz)
fbue{,a}
label
fbge{,a}
label
fbuge{,a}
label
fble{,a}
label
fbule{,a}
label
fbo{,a}
label
Programming Note:
To set the annul bit for FBfcc instructions, append “,a” to the opcode mnemonic. For example, use “fbl,a
label.” The preceding table indicates that the “,a” is optional by enclosing it in braces .

Description:
Unconditional Branches (FBA, FBN):
If its annul field is 0, an FBN (Branch Never) instruction acts like a NOP. If its annul field
is 1, the following (delay) instruction is annulled (not executed) when the FBN is executed. In neither case does a transfer of control take place.
FBA (Branch Always) causes a PC-relative, delayed control transfer to the address
“PC + (4 × sign_ext(disp22)),” regardless of the value of the floating-point condition code
bits. If the annul field of the branch instruction is 1, the delay instruction is annulled (not
executed). If the annul field is 0, the delay instruction is executed.
Fcc-Conditional Branches:
Conditional FBfcc instructions (except FBA and FBN) evaluate floating-point condition
code zero (fcc0) according to the cond field of the instruction. Such evaluation produces
either a TRUE or FALSE result. If TRUE, the branch is taken, that is, the instruction
causes
a
PC-relative,
delayed
control
transfer
to
the
address
“PC + (4 × sign_ext(disp22)).” If FALSE, the branch is not taken.
If a conditional branch is taken, the delay instruction is always executed, regardless of the
value of the annul field. If a conditional branch is not taken and the a (annul) field is 1, the
delay instruction is annulled (not executed). Note that the annul bit has a different effect
on conditional branches than it does on unconditional branches.

Annulment, delay instructions, and delayed control transfers are described further in
Chapter 6, “Instructions.”
Compatibility Note:
Unlike SPARC-V8, SPARC-V9 does not require an instruction between a floating-point compare operation
and a floating-point branch (FBfcc, FBPfcc).

If FPRS.FEF = 0 or PSTATE.PEF = 0, or if an FPU is not present, the FBfcc instruction is not
executed and instead, generates an fp_disabled exception.
Exceptions:
fp_disabled

A.5 Branch on Floating-Point Condition Codes with Prediction
(FBPfcc)
Opcode
FBPA
FBPN
FBPU
FBPG
FBPUG
FBPL
FBPUL
FBPLG
FBPNE
FBPE
FBPUE
FBPGE
FBPUGE
FBPLE
FBPULE
FBPO

cond
1000
0000
0111
0110
0101
0100
0011
0010
0001
1001
1010
1011
1100
1101
1110
1111

Operation
Branch Always
Branch Never
Branch on Unordered
Branch on Greater
Branch on Unordered or Greater
Branch on Less
Branch on Unordered or Less
Branch on Less or Greater
Branch on Not Equal
Branch on Equal
Branch on Unordered or Equal
Branch on Greater or Equal
Branch on Unordered or Greater or Equal
Branch on Less or Equal
Branch on Unordered or Less or Equal
Branch on Ordered

fcc test
1
0
U
G
G or U
L
L or U
L or G
L or G or U
E
E or U
E or G
E or G or U
E or L
E or L or U
E or L or G

Format (2):
00

31 30 29 28

cond

101
25 24

cc1 cc0 p

disp19

22 21 20 19 18

cc1

cc0
00
01
10
11

Condition
code
fcc0
fcc1
fcc2
fcc3

Suggested Assembly Language Syntax
fba{,a}{,pt|,pn}
%fccn, label
fbn{,a}{,pt|,pn}
%fccn, label
fbu{,a}{,pt|,pn}
%fccn, label
fbg{,a}{,pt|,pn}
%fccn, label
fbug{,a}{,pt|,pn}
%fccn, label
fbl{,a}{,pt|,pn}
%fccn, label
fbul{,a}{,pt|,pn}
%fccn, label
fblg{,a}{,pt|,pn}
%fccn, label
fbne{,a}{,pt|,pn}
%fccn, label
(synonym: fbnz)
fbe{,a}{,pt|,pn}
%fccn, label
(synonym: fbz)
fbue{,a}{,pt|,pn}
%fccn, label
fbge{,a}{,pt|,pn}
%fccn, label
fbuge{,a}{,pt|,pn}
%fccn, label
fble{,a}{,pt|,pn}
%fccn, label
fbule{,a}{,pt|,pn}
%fccn, label
fbo{,a}{,pt|,pn}
%fccn, label
Programming Note:
To set the annul bit for FBPfcc instructions, append “,a” to the opcode mnemonic. For example, use
“fbl,a %fcc3,label.” The preceding table indicates that the “,a” is optional by enclosing it in braces.
To set the branch prediction bit, append either “,pt” (for predict taken) or “pn” (for predict not taken) to
the opcode mnemonic. If neither “,pt” nor “,pn” is specified, the assembler shall default to “,pt”. To
select the appropriate floating-point condition code, include "%fcc0", "%fcc1", "%fcc2", or "%fcc3"
before the label.

Description:
Unconditional Branches (FBPA, FBPN):
If its annul field is 0, an FBPN (Floating-Point Branch Never with Prediction) instruction
acts like a NOP. If the Branch Never’s annul field is 0, the following (delay) instruction is
executed; if the annul field is 1, the following instruction is annulled (not executed). In no
case does an FBPN cause a transfer of control to take place.
FBPA (Floating-Point Branch Always with Prediction) causes an unconditional PC-relative, delayed control transfer to the address “PC + (4 × sign_ext(disp19)).” If the annul
field of the branch instruction is 1, the delay instruction is annulled (not executed). If the
annul field is 0, the delay instruction is executed.
Fcc-Conditional Branches:
Conditional FBPfcc instructions (except FBPA and FBPN) evaluate one of the four floating-point condition codes (fcc0, fcc1, fcc2, fcc3) as selected by cc0 and cc1, according to
the cond field of the instruction, producing either a TRUE or FALSE result. If TRUE, the
branch is taken, that is, the instruction causes a PC-relative, delayed control transfer to the
address “PC + (4 × sign_ext(disp19)).” If FALSE, the branch is not taken.

If a conditional branch is taken, the delay instruction is always executed, regardless of the
value of the annul field. If a conditional branch is not taken and the a (annul) field is 1, the
delay instruction is annulled (not executed). Note that the annul bit has a different effect
on conditional branches than it does on unconditional branches.
The predict bit (p) is used to give the hardware a hint about whether the branch is expected
to be taken. A 1 in the p bit indicates that the branch is expected to be taken. A 0 indicates
that the branch is expected not to be taken.
Annulment, delay instructions, and delayed control transfers are described further in
Chapter 6, “Instructions.”
If FPRS.FEF = 0 or PSTATE.PEF = 0, or if an FPU is not present, an FBPfcc instruction is not
executed and instead, generates an fp_disabled exception.
Compatibility Note:
Unlike SPARC-V8, SPARC-V9 does not require an instruction between a floating-point compare operation
and a floating-point branch (FBfcc, FBPfcc).

Exceptions:
fp_disabled

A.6 Branch on Integer Condition Codes (Bicc)
The Bicc instructions are deprecated; they are provided only for compatibility with
previous versions of the architecture. They should not be used in new SPARC-V9
software. It is recommended that the BPcc instructions be used in their place.

Opcode
BAD
BND
BNED
BED
BGD
BLED
BGED
BLD
BGUD
BLEUD
BCCD
BCSD
BPOSD
BNEGD
BVCD
BVSD

cond
1000
0000
1001
0001
1010
0010
1011
0011
1100
0100
1101
0101
1110
0110
1111
0111

Operation

icc test

Branch Always
Branch Never
Branch on Not Equal
Branch on Equal
Branch on Greater
Branch on Less or Equal
Branch on Greater or Equal
Branch on Less
Branch on Greater Unsigned
Branch on Less or Equal Unsigned
Branch on Carry Clear (Greater than or Equal, Unsigned)
Branch on Carry Set (Less than, Unsigned)
Branch on Positive
Branch on Negative
Branch on Overflow Clear
Branch on Overflow Set

1
0
not Z
Z
not (Z or (N xor V))
Z or (N xor V)
not (N xor V)
N xor V
not (C or Z)
C or Z
not C
C
not N
N
not V
V

Format (2):
00

31 30 29 28

cond

010
25 24

disp22
22 21

Suggested Assembly Language Syntax
ba{,a}
label
bn{,a}
label
bne{,a}
label
(synonym: bnz)
be{,a}
label
(synonym: bz)
bg{,a}
label
ble{,a}
label
bge{,a}
label
bl{,a}
label
bgu{,a}
label
bleu{,a}
label
bcc{,a}
label
(synonym: bgeu)
bcs{,a}
label
(synonym: blu)
bpos{,a}
label
bneg{,a}
label
bvc{,a}
label
bvs{,a}
label
Programming Note:
To set the annul bit for Bicc instructions, append “,a” to the opcode mnemonic. For example, use “bgu,a
label.” The preceding table indicates that the “,a” is optional by enclosing it in braces.

Description:
Unconditional Branches (BA, BN):
If its annul field is 0, a BN (Branch Never) instruction acts like a NOP. If its annul field is
1, the following (delay) instruction is annulled (not executed). In neither case does a transfer of control take place.
BA (Branch Always) causes an unconditional PC-relative, delayed control transfer to the
address “PC + (4 × sign_ext(disp22)).” If the annul field of the branch instruction is 1, the
delay instruction is annulled (not executed). If the annul field is 0, the delay instruction is
executed.
Icc-Conditional Branches:
Conditional Bicc instructions (all except BA and BN) evaluate the 32-bit integer condition
codes (icc), according to the cond field of the instruction, producing either a TRUE or
FALSE result. If TRUE, the branch is taken, that is, the instruction causes a PC-relative,
delayed control transfer to the address “PC + (4 × sign_ext(disp22)).” If FALSE, the
branch is not taken.
If a conditional branch is taken, the delay instruction is always executed regardless of the
value of the annul field. If a conditional branch is not taken and the a (annul) field is 1, the
delay instruction is annulled (not executed). Note that the annul bit has a different effect
on conditional branches than it does on unconditional branches.
Annulment, delay instructions, and delayed control transfers are described further in
Chapter 6, “Instructions.”
Exceptions:

(none)

A.7 Branch on Integer Condition Codes with Prediction (BPcc)
Opcode
BPA
BPN
BPNE
BPE
BPG
BPLE
BPGE
BPL
BPGU
BPLEU
BPCC
BPCS
BPPOS
BPNEG
BPVC
BPVS

cond
1000
0000
1001
0001
1010
0010
1011
0011
1100
0100
1101
0101
1110
0110
1111
0111

Operation
Branch Always
Branch Never
Branch on Not Equal
Branch on Equal
Branch on Greater
Branch on Less or Equal
Branch on Greater or Equal
Branch on Less
Branch on Greater Unsigned
Branch on Less or Equal Unsigned
Branch on Carry Clear (Greater Than or Equal, Unsigned)
Branch on Carry Set (Less than, Unsigned)
Branch on Positive
Branch on Negative
Branch on Overflow Clear
Branch on Overflow Set

icc test
1
0
not Z
Z
not (Z or (N xor V))
Z or (N xor V)
not (N xor V)
N xor V
not (C or Z)
C or Z
not C
C
not N
N
not V
V

Format (2):
00

31 30 29 28

cond

001
25 24

cc1 cc0 p

disp19

22 21 20 19 18

cc1

cc0
00
01
10
11

Condition
code
icc
—
xcc
—

Suggested Assembly Language Syntax
ba{,a}{,pt|,pn}
i_or_x_cc, label
bn{,a}{,pt|,pn}
i_or_x_cc, label
(or: iprefetch label)
bne{,a}{,pt|,pn}
i_or_x_cc, label
(synonym: bnz)
be{,a}{,pt|,pn}
i_or_x_cc, label
(synonym: bz)
bg{,a}{,pt|,pn}
i_or_x_cc, label
ble{,a}{,pt|,pn}
i_or_x_cc, label
bge{,a}{,pt|,pn}
i_or_x_cc, label
bl{,a}{,pt|,pn}
i_or_x_cc, label
bgu{,a}{,pt|,pn}
i_or_x_cc, label
bleu{,a}{,pt|,pn}
i_or_x_cc, label
bcc{,a}{,pt|,pn}
i_or_x_cc, label
(synonym: bgeu)
(synonym: blu)
bcs{,a}{,pt|,pn}
i_or_x_cc, label
bpos{,a}{,pt|,pn}
i_or_x_cc, label
bneg{,a}{,pt|,pn}
i_or_x_cc, label
bvc{,a}{,pt|,pn}
i_or_x_cc, label
bvs{,a}{,pt|,pn}
i_or_x_cc, label
Programming Note:
To set the annul bit for BPcc instructions, append “,a” to the opcode mnemonic. For example, use “bgu,a
%icc,label.” The preceding table indicates that the “,a” is optional by enclosing it in braces. To set the
branch prediction bit, append to an opcode mnemonic either “,pt” for predict taken or “,pn” for predict
not taken. If neither “,pt” nor “,pn” is specified, the assembler shall default to “,pt”. To select the appropriate integer condition code, include “%icc” or “%xcc” before the label.

Description:
Unconditional Branches (BPA, BPN):
A BPN (Branch Never with Prediction) instruction for this branch type (op2 = 1) is used
in SPARC-V9 as an instruction prefetch; that is, the effective address (PC + (4 ×
sign_ext(disp19))) specifies an address of an instruction that is expected to be executed
soon. The processor may use this information to begin prefetching instructions from that
address. Like the PREFETCH instruction, this instruction may be treated as a NOP by an
implementation. If the Branch Never’s annul field is 1, the following (delay) instruction is
annulled (not executed). If the annul field is 0, the following instruction is executed. In no
case does a Branch Never cause a transfer of control to take place.
BPA (Branch Always with Prediction) causes an unconditional PC-relative, delayed control transfer to the address “PC + (4 × sign_ext(disp19)).” If the annul field of the branch
instruction is 1, the delay instruction is annulled (not executed). If the annul field is 0, the
delay instruction is executed.
Conditional Branches:
Conditional BPcc instructions (except BPA and BPN) evaluate one of the two integer condition codes (icc or xcc), as selected by cc0 and cc1, according to the cond field of the
instruction, producing either a TRUE or FALSE result. If TRUE, the branch is taken; that

is, the instruction causes a PC-relative, delayed control transfer to the address “PC + (4 ×
sign_ext(disp19)).” If FALSE, the branch is not taken.
If a conditional branch is taken, the delay instruction is always executed regardless of the
value of the annul field. If a conditional branch is not taken and the a (annul) field is 1, the
delay instruction is annulled (not executed). Note that the annul bit has a different effect
for conditional branches than it does for unconditional branches.
The predict bit (p) is used to give the hardware a hint about whether the branch is expected
to be taken. A 1 in the p bit indicates that the branch is expected to be taken; a 0 indicates
that the branch is expected not to be taken.
Annulment, delay instructions, prediction, and delayed control transfers are described further in Chapter 6, “Instructions.”
Exceptions:
illegal_instruction (cc1

cc0 = 012 or 112)

A.8 Call and Link
Opcode
CALL

op
01

Operation
Call and Link

Format (1):
01

disp30

31 30 29

Suggested Assembly Language Syntax
call
label

Description:
The CALL instruction causes an unconditional, delayed, PC-relative control transfer to address
PC + (4 × sign_ext(disp30)). Since the word displacement (disp30) field is 30 bits wide, the target
address lies within a range of –231 to +231 – 4 bytes. The PC-relative displacement is formed by
sign-extending the 30-bit word displacement field to 62 bits and appending two low-order zeros to
obtain a 64-bit byte displacement.
The CALL instruction also writes the value of PC, which contains the address of the CALL, into
r[15] (out register 7). The high-order 32-bits of the PC value stored in r[15] are implementationdependent when PSTATE.AM = 1 (impl. dep. #125). The value written into r[15] is visible to the
instruction in the delay slot.
Exceptions:
(none)

A.9 Compare and Swap
Opcode
CASAPASI
CASXAPASI

op3
11 1100
11 1110

Operation
Compare and Swap Word from Alternate space
Compare and Swap Extended from Alternate space

Format (3):
11

op3

rs1

i=0

imm_asi

rs2

op3

rs1

i=1

—

rs2

31 30 29

25 24

19 18

14 13 12

Suggested Assembly Language Syntax
casa
[regrs1] imm_asi, regrs2, regrd
casa
[regrs1] %asi, regrs2, regrd
casxa
[regrs1] imm_asi, regrs2, regrd
casxa
[regrs1] %asi, regrs2, regrd

Description:
These instructions are used for synchronization and memory updates by concurrent processes.
Uses of compare-and-swap include spin-lock operations, updates of shared counters, and updates
of linked-list pointers. The latter two can use wait-free (nonlocking) protocols.
The CASXA instruction compares the value in register r[rs2] with the doubleword in memory
pointed to by the doubleword address in r[rs1]. If the values are equal, the value in r[rd] is
swapped with the doubleword pointed to by the doubleword address in r[rs1]. If the values are not
equal, the contents of the doubleword pointed to by r[rs1] replaces the value in r[rd], but the
memory location remains unchanged.
The CASA instruction compares the low-order 32 bits of register r[rs2] with a word in memory
pointed to by the word address in r[rs1]. If the values are equal, the low-order 32 bits of register
r[rd] are swapped with the contents of the memory word pointed to by the address in r[rs1] and
the high-order 32 bits of register r[rd] are set to zero. If the values are not equal, the memory location remains unchanged, but the zero-extended contents of the memory word pointed to by r[rs1]
replace the low-order 32 bits of r[rd] and the high-order 32 bits of register r[rd] are set to zero.
A compare-and-swap instruction comprises three operations: a load, a compare, and a swap. The
overall instruction is atomic; that is, no intervening interrupts or deferred traps are recognized by
the processor, and no intervening update resulting from a compare-and-swap, swap, load, loadstore unsigned byte, or store instruction to the doubleword containing the addressed location, or
any portion of it, is performed by the memory system.
A compare-and-swap operation does not imply any memory barrier semantics. When compareand-swap is used for synchronization, the same consideration should be given to memory barriers
as if a load, store, or swap instruction were used.

A compare-and-swap operation behaves as if it performs a store, either of a new value from r[rd]
or of the previous value in memory. The addressed location must be writable, even if the values in
memory and r[rs2] are not equal.
If i = 0, the address space of the memory location is specified in the imm_asi field; if i = 1, the
address space is specified in the ASI register.
A mem_address_not_aligned exception is generated if the address in r[rs1] is not properly aligned.
CASXA and CASA cause a privileged_action exception if PSTATE.PRIV = 0 and bit 7 of the ASI
is zero.
The coherence and atomicity of memory operations between processors and I/O DMA memory
accesses are implementation-dependent (impl. dep #120).
Implementation Note:
An implementation might cause an exception due to an error during the store memory access, even though
there was no error during the load memory access.
Programming Note:
Compare and Swap (CAS) and Compare and Swap Extended (CASX) synthetic instructions are available for
“big endian” memory accesses. Compare and Swap Little (CASL) and Compare and Swap Extended Little
(CASXL) synthetic instructions are available for “little endian” memory accesses. See G.3, “Synthetic
Instructions,” for these synthetic instructions’ syntax.

The compare-and-swap instructions do not affect the condition codes.
Exceptions:
privileged_action
mem_address_not_aligned
data_access_exception
data_access_MMU_miss
data_access_protection
data_access_error
async_data_error

A.10 Divide (64-bit / 32-bit)
The UDIV, UDIVcc, SDIV, and SDIVcc instructions are deprecated; they are provided only for compatibility with previous versions of the architecture. They should
not be used in new SPARC-V9 software. It is recommended that the UDIVX and
SDIVX instructions be used in their place.

Opcode
UDIVD
SDIVD
UDIVccD
SDIVccD

op3
00 1110
00 1111
01 1110
01 1111

Operation
Unsigned Integer Divide
Signed Integer Divide
Unsigned Integer Divide and modify cc’s
Signed Integer Divide and modify cc’s

Format (3):
10

op3

rs1

i=0

op3

rs1

i=1

31 30 29

25 24

19 18

14 13 12

—

rs2

simm13
5

Suggested Assembly Language Syntax
udiv
regrs1, reg_or_imm, regrd
sdiv
regrs1, reg_or_imm, regrd
udivcc
regrs1, reg_or_imm, regrd
sdivcc
regrs1, reg_or_imm, regrd

Description:
The divide instructions perform 64-bit by 32-bit division, producing a 32-bit result. If i = 0, they
compute “(Y lower 32 bits of r[rs1]) ÷ lower 32 bits of r[rs2].” Otherwise (i.e., if i = 1), the
divide instructions compute “(Y lower 32 bits of r[rs1]) ÷ lower 32 bits of sign_ext(simm13).”
In either case, if overflow does not occur, the less significant 32 bits of the integer quotient are
sign-or zero-extended to 64 bits and are written into r[rd].
The contents of the Y register are undefined after any 64-bit by 32-bit integer divide operation.
Unsigned Divide:
Unsigned divide (UDIV, UDIVcc) assumes an unsigned integer doubleword dividend (Y lower
32 bits of r[rs1]) and an unsigned integer word divisor (lower 32 bits of r[rs2] or lower 32 bits of
sign_ext(simm13)) and computes an unsigned integer word quotient (r[rd]). Immediate values in
simm13 are in the ranges 0..212 – 1 and 232 – 212 .. 232 – 1 for unsigned divide instructions.
Unsigned division rounds an inexact rational quotient toward zero .

Programming Note:
The rational quotient is the infinitely precise result quotient. It includes both the integer part and the fractional part of the result. For example, the rational quotient of 11/4 = 2.75 (Integer part = 2, fractional part
= .75).

The result of an unsigned divide instruction can overflow the low-order 32 bits of the destination
register r[rd] under certain conditions. When overflow occurs the largest appropriate unsigned
integer is returned as the quotient in r[rd]. The condition under which overflow occurs and the
value returned in r[rd] under this condition is specified in the following table.
Table 23—UDIV / UDIVcc Overflow Detection and Value Returned
Condition under which overflow occurs
Rational quotient ≥ 232

Value returned in r[rd]
232 −1
(0000 0000 FFFF FFFF16)

When no overflow occurs, the 32-bit result is zero-extended to 64 bits and written into register
r[rd].
UDIV does not affect the condition code bits. UDIVcc writes the integer condition code bits as
shown in the following table. Note that negative (N) and zero (Z) are set according to the value of
r[rd] after it has been set to reflect overflow, if any.
Bit
icc.N
icc.Z
icc.V
icc.C
xcc.N
xcc.Z
xcc.V
xcc.C

UDIVcc
Set if r[rd]<31> = 1
Set if r[rd]<31:0> = 0
Set if overflow (per table 23)
Zero
Set if r[rd]<63> = 1
Set if r[rd]<63:0> = 0
Zero
Zero

Signed Divide:
Signed divide (SDIV, SDIVcc) assumes a signed integer doubleword dividend (Y lower 32 bits
of r[rs1]) and a signed integer word divisor (lower 32 bits of r[rs2] or lower 32 bits of
sign_ext(simm13)) and computes a signed integer word quotient (r[rd]).
Signed division rounds an inexact quotient toward zero. For example, –7 ÷ 4 equals the rational
quotient of –1.75, which rounds to –1 (not –2) when rounding toward zero.
The result of a signed divide can overflow the low-order 32 bits of the destination register r[rd]
under certain conditions. When overflow occurs the largest appropriate signed integer is returned
as the quotient in r[rd]. The conditions under which overflow occurs and the value returned in
r[rd] under those conditions are specified in the following table.

Table 24—SDIV / SDIVcc Overflow Detection and Value Returned
Condition under which overflow occurs
Rational quotient ≥ 231
Rational quotient ≤ -231−1

Value returned in r[rd]
231 −1
(0000 0000 7FFF FFFF16)
−231
(FFFF FFFF 8000 000016)

When no overflow occurs, the 32-bit result is sign-extended to 64 bits and written into register
r[rd].
SDIV does not affect the condition code bits. SDIVcc writes the integer condition code bits as
shown in the following table. Note that negative (N) and zero (Z) are set according to the value of
r[rd] after it has been set to reflect overflow, if any.
Bit
icc.N
icc.Z
icc.V
icc.C
xcc.N
xcc.Z
xcc.V
xcc.C

Exceptions:
division_by_zero

SDIVcc
Set if r[rd]<31> = 1
Set if r[rd]<31:0> = 0
Set if overflow (per table 24)
Zero
Set if r[rd]<63]> = 1
Set if r[rd]<63:0> = 0
Zero
Zero

A.11 DONE and RETRY
Opcode
DONEP
RETRYP
—

op3
11 1110
11 1110
11 1110

fcn
0
1
2..31

Operation
Return from Trap (skip trapped instruction)
Return from Trap (retry trapped instruction)
Reserved

Format (3):
10
31 30 29

fcn

op3
25 24

—
19 18

Suggested Assembly Language Syntax
done
retry

Description:
The DONE and RETRY instructions restore the saved state from TSTATE (CWP, ASI, CCR, and
PSTATE), set PC and nPC, and decrement TL.
The RETRY instruction resumes execution with the trapped instruction by setting PC←TPC[TL]
(the saved value of PC on trap) and nPC←TNPC[TL] (the saved value of nPC on trap).
The DONE instruction skips the trapped instruction by setting PC←TNPC[TL] and
nPC←TNPC[TL]+4.
Execution of a DONE or RETRY instruction in the delay slot of a control-transfer instruction produces undefined results.
Programming Note:
The DONE and RETRY instructions should be used to return from privileged trap handlers.

Exceptions:
privileged_opcode
illegal_instruction (if TL = 0 or fcn = 2..31)

A.12 Floating-Point Add and Subtract
Opcode
FADDs
FADDd
FADDq
FSUBs
FSUBd
FSUBq

op3
11 0100
11 0100
11 0100
11 0100
11 0100
11 0100

opf
0 0100 0001
0 0100 0010
0 0100 0011
0 0100 0101
0 0100 0110
0 0100 0111

Operation
Add Single
Add Double
Add Quad
Subtract Single
Subtract Double
Subtract Quad

Format (3):
10
31 30 29

op3
25 24

rs1
19 18

opf
14 13

rs2
5 4

Suggested Assembly Language Syntax
fadds
fregrs1, fregrs2, fregrd
faddd
fregrs1, fregrs2, fregrd
faddq
fregrs1, fregrs2, fregrd
fsubs
fregrs1, fregrs2, fregrd
fsubd
fregrs1, fregrs2, fregrd
fsubq
fregrs1, fregrs2, fregrd

Description:
The floating-point add instructions add the floating-point register(s) specified by the rs1 field and
the floating-point register(s) specified by the rs2 field, and write the sum into the floating-point
register(s) specified by the rd field.
The floating-point subtract instructions subtract the floating-point register(s) specified by the rs2
field from the floating-point register(s) specified by the rs1 field, and write the difference into the
floating-point register(s) specified by the rd field.
Rounding is performed as specified by the FSR.RD field.
Exceptions:
fp_disabled
fp_exception_ieee_754 (OF, UF, NX, NV)
fp_exception_other (invalid_fp_register (only FADDQ and FSUBQ))

A.13 Floating-Point Compare
Opcode
FCMPs
FCMPd
FCMPq
FCMPEs
FCMPEd
FCMPEq

op3
11 0101
11 0101
11 0101
11 0101
11 0101
11 0101

opf
0 0101 0001
0 0101 0010
0 0101 0011
0 0101 0101
0 0101 0110
0 0101 0111

Operation
Compare Single
Compare Double
Compare Quad
Compare Single and Exception if Unordered
Compare Double and Exception if Unordered
Compare Quad and Exception if Unordered

Format (3):
10

000

31 30 29

cc1 cc0

op3

27 26 25 24

rs1
19 18

opf
14 13

rs2
5 4

Suggested Assembly Language Syntax
fcmps
%fccn, fregrs1, fregrs2
fcmpd
%fccn, fregrs1, fregrs2
fcmpq
%fccn, fregrs1, fregrs2
fcmpes
%fccn, fregrs1, fregrs2
fcmped
%fccn, fregrs1, fregrs2
fcmpeq
%fccn, fregrs1, fregrs2

cc1

cc0
00
01
10
11

Condition
code
fcc0
fcc1
fcc2
fcc3

Description:
These instructions compare the floating-point register(s) specified by the rs1 field with the floating-point register(s) specified by the rs2 field, and set the selected floating-point condition code
(fccn) according to the following table:
fcc value
0
1
2
3

Relation
fregrs1 = fregrs2
fregrs1 < fregrs2
fregrs1 > fregrs2
fregrs1 ? fregrs2 (unordered)

The “?” in the above table indicates that the comparison is unordered. The unordered condition
occurs when one or both of the operands to the compare is a signaling or quiet NaN.
The “compare and cause exception if unordered” (FCMPEs, FCMPEd, and FCMPEq) instructions cause an invalid (NV) exception if either operand is a NaN.
FCMP causes an invalid (NV) exception if either operand is a signaling NaN.
Compatibility Note:
Unlike SPARC-V8, SPARC-V9 does not require an instruction between a floating-point compare operation
and a floating-point branch (FBfcc, FBPfcc).
Compatibility Note:
SPARC-V8 floating-point compare instructions are required to have a zero in the r[rd] field. In SPARC-V9,
bits 26 and 25 of the r[rd] field are used to specify the floating-point condition code to be set. Legal SPARCV8 code will work on SPARC-V9 because the zeroes in the r[rd] field are interpreted as fcc0, and the
FBfcc instruction branches based on fcc0.

Exceptions:
fp_disabled
fp_exception_ieee_754 (NV)
fp_exception_other (invalid_fp_register (FCMPq, FCMPEq only))

A.14 Convert Floating-Point to Integer
Opcode
FsTOx
FdTOx
FqTOx
FsTOi
FdTOi
FqTOi

op3
11 0100
11 0100
11 0100
11 0100
11 0100
11 0100

opf
0 1000 0001
0 1000 0010
0 1000 0011
0 1101 0001
0 1101 0010
0 1101 0011

Operation
Convert Single to 64-bit Integer
Convert Double to 64-bit Integer
Convert Quad to 64-bit Integer
Convert Single to 32-bit Integer
Convert Double to 32-bit Integer
Convert Quad to 32-bit Integer

Format (3):
10
31 30 29

op3
25 24

—
19 18

opf
14 13

rs2
5 4

Suggested Assembly Language Syntax
fstox
fregrs2, fregrd
fdtox
fregrs2, fregrd
fqtox
fregrs2, fregrd
fstoi
fregrs2, fregrd
fdtoi
fregrs2, fregrd
fqtoi
fregrs2, fregrd

Description:
FsTOx, FdTOx, and FqTOx convert the floating-point operand in the floating-point register(s)
specified by rs2 to a 64-bit integer in the floating-point register(s) specified by rd.
FsTOi, FdTOi, and FqTOi convert the floating-point operand in the floating-point register(s) specified by rs2 to a 32-bit integer in the floating-point register specified by rd.
The result is always rounded toward zero; that is, the rounding direction (RD) field of the FSR
register is ignored.
If the floating-point operand’s value is too large to be converted to an integer of the specified size,
or is a NaN or infinity, an invalid (NV) exception occurs. The value written into the floating-point
register(s) specified by rd in these cases is defined in B.5, “Integer Overflow Definition.”
Exceptions:
fp_disabled
fp_exception_ieee_754 (NV, NX)
fp_exception_other (invalid_fp_register (FqTOi, FqTOx only))

A.15 Convert Between Floating-Point Formats
Opcode
FsTOd
FsTOq
FdTOs
FdTOq
FqTOs
FqTOd

op3
11 0100
11 0100
11 0100
11 0100
11 0100
11 0100

opf
0 1100 1001
0 1100 1101
0 1100 0110
0 1100 1110
0 1100 0111
0 1100 1011

Operation
Convert Single to Double
Convert Single to Quad
Convert Double to Single
Convert Double to Quad
Convert Quad to Single
Convert Quad to Double

Format (3):
10
31 30 29

op3
25 24

—
19 18

opf
14 13

rs2
5 4

Suggested Assembly Language Syntax
fstod
fregrs2, fregrd
fstoq
fregrs2, fregrd
fdtos
fregrs2, fregrd
fdtoq
fregrs2, fregrd
fqtos
fregrs2, fregrd
fqtod
fregrs2, fregrd

Description:
These instructions convert the floating-point operand in the floating-point register(s) specified by
rs2 to a floating-point number in the destination format. They write the result into the floatingpoint register(s) specified by rd.
Rounding is performed as specified by the FSR.RD field.
FqTOd, FqTOs, and FdTOs (the “narrowing” conversion instructions) can raise OF, UF, and NX
exceptions. FdTOq, FsTOq, and FsTOd (the “widening” conversion instructions) cannot.
Any of these six instructions can trigger an NV exception if the source operand is a signaling
NaN.
B.2.1, “Untrapped Result in Different Format from Operands,” defines the rules for converting
NaNs from one floating-point format to another.
Exceptions:
fp_disabled
fp_exception_ieee_754 (OF, UF, NV, NX)
fp_exception_other (invalid_fp_register) (FsTOq, FdTOq, FqTOs, FqTOd)

A.16 Convert Integer to Floating-Point
Opcode
FxTOs
FxTOd
FxTOq
FiTOs
FiTOd
FiTOq

op3
11 0100
11 0100
11 0100
11 0100
11 0100
11 0100

opf
0 1000 0100
0 1000 1000
0 1000 1100
0 1100 0100
0 1100 1000
0 1100 1100

Operation
Convert 64-bit Integer to Single
Convert 64-bit Integer to Double
Convert 64-bit Integer to Quad
Convert 32-bit Integer to Single
Convert 32-bit Integer to Double
Convert 32-bit Integer to Quad

Format (3):
10
31 30 29

op3
25 24

—
19 18

opf
14 13

rs2
5 4

Suggested Assembly Language Syntax
fxtos
fregrs2, fregrd
fxtod
fregrs2, fregrd
fxtoq
fregrs2, fregrd
fitos
fregrs2, fregrd
fitod
fregrs2, fregrd
fitoq
fregrs2, fregrd

Description:
FxTOs, FxTOd, and FxTOq convert the 64-bit signed integer operand in the floating-point register(s) specified by rs2 into a floating-point number in the destination format. The source register,
floating-point register(s) specified by rs2, must be an even-numbered (that is, double-precision)
floating-point register.
FiTOs, FiTOd, and FiTOq convert the 32-bit signed integer operand in floating-point register(s)
specified by rs2 into a floating-point number in the destination format. All write their result into
the floating-point register(s) specified by rd.
FiTOs, FxTOs, and FxTOd round as specified by the FSR.RD field.
Exceptions:
fp_disabled
fp_exception_ieee_754 (NX (FiTOs, FxTOs, FxTOd only))
fp_exception_other (invalid_fp_register (FiTOq, FxTOq only))

A.17 Floating-Point Move
Opcode
FMOVs
FMOVd
FMOVq
FNEGs
FNEGd
FNEGq
FABSs
FABSd
FABSq

op3
11 0100
11 0100
11 0100
11 0100
11 0100
11 0100
11 0100
11 0100
11 0100

opf
0 0000 0001
0 0000 0010
0 0000 0011
0 0000 0101
0 0000 0110
0 0000 0111
0 0000 1001
0 0000 1010
0 0000 1011

Operation
Move Single
Move Double
Move Quad
Negate Single
Negate Double
Negate Quad
Absolute Value Single
Absolute Value Double
Absolute Value Quad

Format (3):
10
31 30 29

op3
25 24

—
19 18

opf
14 13

rs2
5 4

Suggested Assembly Language Syntax
fmovs
fregrs2, fregrd
fmovd
fregrs2, fregrd
fmovq
fregrs2, fregrd
fnegs
fregrs2, fregrd
fnegd
fregrs2, fregrd
fnegq
fregrs2, fregrd
fabss
fregrs2, fregrd
fabsd
fregrs2, fregrd
fabsq
fregrs2, fregrd

Description:
The single-precision versions of these instructions copy the contents of a single-precision floating-point register to the destination. The double-precision forms copy the contents of a doubleprecision floating-point register to the destination. The quad-precision versions copy a quad-precision value in floating-point registers to the destination.
FMOV copies the source to the destination unaltered.
FNEG copies the source to the destination with the sign bit complemented.
FABS copies the source to the destination with the sign bit cleared.
These instructions do not round.
Exceptions:
fp_disabled
fp_exception_other (invalid_fp_register (FMOVq, FNEGq, FABSq only))

A.18 Floating-Point Multiply and Divide
Opcode
FMULs
FMULd
FMULq
FsMULd
FdMULq
FDIVs
FDIVd
FDIVq

op3
11 0100
11 0100
11 0100
11 0100
11 0100
11 0100
11 0100
11 0100

opf
0 0100 1001
0 0100 1010
0 0100 1011
0 0110 1001
0 0110 1110
0 0100 1101
0 0100 1110
0 0100 1111

Operation
Multiply Single
Multiply Double
Multiply Quad
Multiply Single to Double
Multiply Double to Quad
Divide Single
Divide Double
Divide Quad

Format (3):
10
31 30 29

op3
25 24

rs1
19 18

opf
14 13

rs2
5 4

Suggested Assembly Language Syntax
fmuls
fregrs1, fregrs2, fregrd
fmuld
fregrs1, fregrs2, fregrd
fmulq
fregrs1, fregrs2, fregrd
fsmuld
fregrs1, fregrs2, fregrd
fdmulq
fregrs1, fregrs2, fregrd
fdivs
fregrs1, fregrs2, fregrd
fdivd
fregrs1, fregrs2, fregrd
fdivq
fregrs1, fregrs2, fregrd

Description:
The floating-point multiply instructions multiply the contents of the floating-point register(s)
specified by the rs1 field by the contents of the floating-point register(s) specified by the rs2 field,
and write the product into the floating-point register(s) specified by the rd field.
The FsMULd instruction provides the exact double-precision product of two single-precision
operands, without underflow, overflow, or rounding error. Similarly, FdMULq provides the exact
quad-precision product of two double-precision operands.
The floating-point divide instructions divide the contents of the floating-point register(s) specified
by the rs1 field by the contents of the floating-point register(s) specified by the rs2 field, and write
the quotient into the floating-point register(s) specified by the rd field.
Rounding is performed as specified by the FSR.RD field.
Exceptions:
fp_disabled
fp_exception_ieee_754 (OF, UF, DZ (FDIV only), NV, NX)
fp_exception_other (invalid_fp_register (FMULq, FdMULq, and FDIVq only))

A.19 Floating-Point Square Root
Opcode
FSQRTs
FSQRTd
FSQRTq

op3
11 0100
11 0100
11 0100

opf
0 0010 1001
0 0010 1010
0 0010 1011

Operation
Square Root Single
Square Root Double
Square Root Quad

Format (3):
10
31 30 29

op3
25 24

—
19 18

opf
14 13

rs2
5 4

Suggested Assembly Language Syntax
fsqrts fregrs2, fregrd
fsqrtd fregrs2, fregrd
fsqrtq fregrs2, fregrd

Description:
These instructions generate the square root of the floating-point operand in the floating-point register(s) specified by the rs2 field, and place the result in the destination floating-point register(s)
specified by the rd field.
Rounding is performed as specified by the FSR.RD field.
Implementation Note:
See Implementation Characteristics of Current SPARC-V9-based Products, Revision 9.x, a document available from SPARC International, for information on whether the FSQRT instructions are implemented in hardware or software in the various SPARC-V9 implementations.

Exceptions:
fp_disabled
fp_exception_ieee_754 (IEEE_754_exception (NV, NX))
fp_exception_other (invalid_fp_register (FSQRTq))

A.20 Flush Instruction Memory
Opcode
FLUSH

op3
11 1011

Operation
Flush Instruction Memory

Format (3):
10

—

op3

rs1

i=0

—

op3

rs1

i=1

31 30 29

25 24

19 18

14 13 12

—

rs2

simm13
5

Suggested Assembly Language Syntax
flush
address

Description:
FLUSH ensures that the doubleword specified as the effective address is consistent across any
local caches and, in a multiprocessor system, will eventually become consistent everywhere.
In the following discussion PFLUSH refers to the processor that executed the FLUSH instruction.
FLUSH ensures that instruction fetches from the specified effective address by PFLUSH appear to
execute after any loads, stores, and atomic load-stores to that address issued by PFLUSH prior to the
FLUSH. In a multiprocessor system, FLUSH also ensures that these values will eventually
become visible to the instruction fetches of all other processors. FLUSH behaves as if it were a
store with respect to MEMBAR-induced orderings. See A.32, “Memory Barrier.”
FLUSH operates on at least the doubleword containing the addressed location.
The effective address operand for the FLUSH instruction is “r[rs1] + r[rs2]” if i = 0, or
“r[rs1] + sign_ext(simm13)” if i = 1. The least significant two address bits of the effective address
are unused and should be supplied as zeros by software. Bit 2 of the address is ignored, because
FLUSH operates on at least a doubleword.
Programming Notes:
(1) Typically, FLUSH is used in self-modifying code. See H.1.6, “Self-Modifying Code,” for information about
use of the FLUSH instruction in portable self-modifying code. The use of self-modifying code is discouraged.
(2) The order in which memory is modified can be controlled by using FLUSH and MEMBAR instructions
interspersed appropriately between stores and atomic load-stores. FLUSH is needed only between a store
and a subsequent instruction fetch from the modified location. When multiple processes may concurrently
modify live (that is, potentially executing) code, care must be taken to ensure that the order of update maintains the program in a semantically correct form at all times.
(3) The memory model guarantees in a uniprocessor that data loads observe the results of the most recent store,
even if there is no intervening FLUSH.

(4) FLUSH may be time-consuming. Some implementations may trap rather than implement FLUSH in hardware. In a multiprocessor configuration, FLUSH requires all processors that may be referencing the
addressed doubleword to flush their instruction caches, a potentially disruptive activity.
(5) In a multiprocessor system, the time it takes for a FLUSH to take effect is implementation-dependent (impl.
dep. #122). No mechanism is provided to ensure or test completion.
(6) Because FLUSH is designed to act on a doubleword, and because, on some implementations, FLUSH may
trap to system software, it is recommended that system software provide a user-callable service routine for
flushing arbitrarily sized regions of memory. On some implementations, this routine would issue a series of
FLUSH instructions; on others, it might issue a single trap to system software that would then flush the
entire region.
Implementation Notes:
(1) IMPL. DEP. #42: If FLUSH is not implemented in hardware, it causes an illegal_instruction exception and the function of FLUSH is performed by system software. Whether FLUSH traps is implementation-dependent.
(2) The effect of a FLUSH instruction as observed from PFLUSH is immediate. Other processors in a multiprocessor system eventually will see the effect of the FLUSH, but the latency is implementation-dependent
(impl. dep. #122).

Exceptions:
(none)

A.21 Flush Register Windows
Opcode
FLUSHW

op3
10 1011

Operation
Flush Register Windows

Format (3):
10

—

31 30 29

op3
25 24

—
19 18

i=0
14 13 12

—
0

Suggested Assembly Language Syntax
flushw

Description:
FLUSHW causes all active register windows except the current window to be flushed to memory
at locations determined by privileged software. FLUSHW behaves as a NOP if there are no active
windows other than the current window. At the completion of the FLUSHW instruction, the only
active register window is the current one.
Programming Note:
The FLUSHW instruction can be used by application software to switch memory stacks or examine register
contents for previous stack frames.

FLUSHW acts as a NOP if CANSAVE = NWINDOWS – 2. Otherwise, there is more than one
active window, so FLUSHW causes a spill exception. The trap vector for the spill exception is
based on the contents of OTHERWIN and WSTATE. The spill trap handler is invoked with the
CWP set to the window to be spilled (that is, (CWP + CANSAVE + 2) mod NWINDOWS). See
6.3.6, “Register Window Management Instructions.”
Programming Note:
Typically, the spill handler will save a window on a memory stack and return to reexecute the FLUSHW
instruction. Thus, FLUSHW will trap and reexecute until all active windows other than the current window
have been spilled.

Exceptions:
spill_n_normal
spill_n_other

A.22 Illegal Instruction Trap
Opcode
ILLTRAP

op
00

op2
000

Operation
illegal_instruction trap

Format (2):
00
31 30 29

—

000
25 24

const22
22 21

Suggested Assembly Language Syntax
illtrap

const22

Description:
The ILLTRAP instruction causes an illegal_instruction exception. The const22 value is ignored by
the hardware; specifically, this field is not reserved by the architecture for any future use.
Compatibility Note:
Except for its name, this instruction is identical to the SPARC-V8 UNIMP instruction.

Exceptions:
illegal_instruction

A.23 Implementation-Dependent Instructions
Opcode
IMPDEP1
IMPDEP2

op3
11 0110
11 0111

Operation
Implementation-Dependent Instruction 1
Implementation-Dependent Instruction 2

Format (3):
10
31 30 29

impl-dep

op3
25 24

19 18

Description:
IMPL. DEP. #106: The IMPDEP1 and IMPDEP2 instructions are completely implementation-dependent.
Implementation-dependent aspects include their operation, the interpretation of bits 29..25 and 18..0 in
their encodings, and which (if any) exceptions they may cause.

See I.1.2, “Implementation-Dependent and Reserved Opcodes,” for information about extending
the SPARC-V9 instruction set using the implementation-dependent instructions.
Compatibility Note:
These instructions replace the CPopn instructions in SPARC-V8.

Exceptions:
illegal_instruction (if the implementation does not define the instructions)
implementation-dependent (if the implementation defines the instructions)

A.24 Jump and Link
Opcode
JMPL

op3
11 1000

Operation
Jump and Link

Format (3):
10

op3

rs1

i=0

op3

rs1

i=1

31 30 29

25 24

19 18

14 13 12

—

rs2

simm13
5

Suggested Assembly Language Syntax
jmpl
address, reg rd

Description:
The JMPL instruction causes a register-indirect delayed control transfer to the address given by
“r[rs1] + r[rs2]” if i field = 0, or “r[rs1] + sign_ext(simm13)” if i = 1.
The JMPL instruction copies the PC, which contains the address of the JMPL instruction, into
register r[rd]. The high-order 32-bits of the PC value stored in r[rd] are implementation-dependent when PSTATE.AM = 1 (impl. dep. #125). The value written into r[rd] is visible to the
instruction in the delay slot.
If either of the low-order two bits of the jump address is nonzero, a mem_address_not_aligned
exception occurs.
Programming Note:
A JMPL instruction with rd = 15 functions as a register-indirect call using the standard link register.
JMPL with rd = 0 can be used to return from a subroutine. The typical return address is “r[31] + 8,” if a
nonleaf routine (one that uses the SAVE instruction) is entered by a CALL instruction, or “r[15] + 8” if a
leaf routine (one that does not use the SAVE instruction) is entered by a CALL instruction or by a JMPL
instruction with rd = 15.

Exceptions:
mem_address_not_aligned

A.25 Load Floating-Point
The LDFSR instruction is deprecated; it is provided only for compatibility with
previous versions of the architecture. It should not be used in new SPARC-V9 software. It is recommended that the LDXFSR instruction be used in its place.

Opcode
LDF
LDDF
LDQF
LDFSRD
LDXFSR
—
†

op3
10 000
0
10 001
1
10 001
0
10 000
1
10 000
1
10 000
1

rd
0..31

Operation
Load Floating-Point Register

†

Load Double Floating-Point Register

†

Load Quad Floating-Point Register

Load Floating-Point State Register Lower

Load Floating-Point State Register

2..31

Reserved

Encoded floating-point register value, as described in 5.1.4.1

Format (3):
11

op3

rs1

i=0

op3

rs1

i=1

31 30 29

25 24

19 18

14 13 12

rs2

—

simm13
5

Suggested Assembly Language Syntax
ld
[address], fregrd
ldd
[address], fregrd
ldq
[address], fregrd
ld
[address], %fsr
ldx
[address], %fsr

Description:
The load single floating-point instruction (LDF) copies a word from memory into f[rd].
The load doubleword floating-point instruction (LDDF) copies a word-aligned doubleword from
memory into a double-precision floating-point register.
The load quad floating-point instruction (LDQF) copies a word-aligned quadword from memory
into a quad-precision floating-point register.

The load floating-point state register lower instruction (LDFSR) waits for all FPop instructions
that have not finished execution to complete, and then loads a word from memory into the lower
32 bits of the FSR. The upper 32 bits of FSR are unaffected by LDFSR.
The load floating-point state register instruction (LDXFSR) waits for all FPop instructions that
have not finished execution to complete, and then loads a doubleword from memory into the FSR.
Compatibility Note:
SPARC-V9 supports two different instructions to load the FSR; the SPARC-V8 LDFSR instruction is
defined to load only the lower 32 bits into the FSR, whereas LDXFSR allows SPARC-V9 programs to load
all 64 bits of the FSR.

Load floating-point instructions access the primary address space (ASI = 8016). The effective
address for these instructions is “r[rs1] + r[rs2]” if i = 0, or “r[rs1] + sign_ext(simm13)” if i = 1.
LDF, LDFSR, LDDF, and LDQF cause a mem_address_not_aligned exception if the effective
memory address is not word-aligned; LDXFSR causes a mem_address_not_aligned exception if the
address is not doubleword-aligned. If the floating-point unit is not enabled (per FPRS.FEF and
PSTATE.PEF), or if no FPU is present, a load floating-point instruction causes an fp_disabled
exception.
IMPL. DEP. #109(1): LDDF requires only word alignment. However, if the effective address is word-aligned
but not doubleword-aligned, LDDF may cause an LDDF_mem_address_not_aligned exception. In this
case the trap handler software shall emulate the LDDF instruction and return.
IMPL. DEP. #111(1): LDQF requires only word alignment. However, if the effective address is word-aligned
but not quadword-aligned, LDQF may cause an LDQF_mem_address_not_aligned exception. In this case
the trap handler software shall emulate the LDQF instruction and return.
Programming Note:
In SPARC-V8, some compilers issued sequences of single-precision loads when they could not determine
that double- or quadword operands were properly aligned. For SPARC-V9, since emulation of misaligned
loads is expected to be fast, it is recommended that compilers issue sets of single-precision loads only when
they can determine that double- or quadword operands are not properly aligned.
Implementation Note:
IMPL. DEP. #44: If a load floating-point instruction traps with any type of access error, the contents
of the destination floating-point register(s) remain unchanged or are undefined.

Exceptions:
async_data_error
illegal_instruction (op3=2116 and rd = 2..31)
fp_disabled
LDDF_mem_address_not_aligned (LDDF only) (impl. dep. #109)
LDQF_mem_address_not_aligned (LDQF only) (impl. dep. #111)
fp_exception_other (invalid_fp_register (LDQF only))
mem_address_not_aligned
data_access_MMU_miss
data_access_exception
data_access_error
data_access_protection

A.26 Load Floating-Point from Alternate Space

†

Opcode
LDFAPASI

op3
11 000
0

rd
0..31

Operation
Load Floating-Point Register from Alternate space

LDDFAPASI

11 001
1

†

Load Double Floating-Point Register from Alternate space

LDQFAPASI

11 001
0

†

Load QuadFloating-Point Register from Alternate space

Encoded floating-point register value, as described in 5.1.4.1

Format (3):
11

op3

rs1

i=0

op3

rs1

i=1

31 30 29

25 24

19 18

imm_asi

14 13 12

rs2

simm13
5

Suggested Assembly Language Syntax
lda
[regaddr] imm_asi, fregrd
lda
[reg_plus_imm] %asi, fregrd
ldda
[regaddr] imm_asi, fregrd
ldda
[reg_plus_imm] %asi, fregrd
ldqa
[regaddr] imm_asi, fregrd
ldqa
[reg_plus_imm] %asi, fregrd

Description:
The load single floating-point from alternate space instruction (LDFA) copies a word from memory into f[rd].
The load doubleword floating-point from alternate space instruction (LDDFA) copies a wordaligned doubleword from memory into a double-precision floating-point register.
The load quad floating-point from alternate space instruction (LDQFA) copies a word-aligned
quadword from memory into a quad-precision floating-point register.
Load floating-point from alternate space instructions contain the address space identifier (ASI) to
be used for the load in the imm_asi field if i = 0, or in the ASI register if i = 1. The access is privileged if bit seven of the ASI is zero; otherwise, it is not privileged. The effective address for these
instructions is “r[rs1] + r[rs2]” if i = 0, or “r[rs1] + sign_ext(simm13)” if i = 1.
LDFA, LDDFA, and LDQFA cause a mem_address_not_aligned exception if the effective memory
address is not word-aligned; If the floating-point unit is not enabled (per FPRS.FEF and
PSTATE.PEF), or if no FPU is present, load floating-point from alternate space instructions cause
an fp_disabled exception. LDFA, LDDFA and LDQFA cause a privileged_action exception if
PSTATE.PRIV = 0 and bit 7 of the ASI is zero.

IMPL. DEP. #109(2): LDDFA requires only word alignment. However, if the effective address is wordaligned but not doubleword-aligned, LDDFA may cause an LDDF_mem_address_not_aligned exception.
In this case the trap handler software shall emulate the LDDF instruction and return.
IMPL. DEP. #111(2): LDQFA requires only word alignment. however, if the effective address is wordaligned but not quadword-aligned, LDQFA may cause an ldqf_mem_address_not_aligned exception. In
this case the trap handler software shall emulate the LDQF instruction and return.
Programming Note:
In SPARC-V8, some compilers issued sequences of single-precision loads when they could not determine
that double- or quadword operands were properly aligned. For SPARC-V9, since emulation of mis-aligned
loads is expected to be fast, it is recommended that compilers issue sets of single-precision loads only when
they can determine that double- or quadword operands are not properly aligned.
Implementation Note:
If a load floating-point instruction traps with any type of access error, the destination floating-point register(s) either remain unchanged or are undefined. (impl. dep. #44)

Exceptions:
async_data_error
fp_disabled
LDDF_mem_address_not_aligned (LDDFA only) (impl. dep. #109)
LDQF_mem_address_not_aligned (LDQFA only) (impl. dep. #111)
fp_exception_other (invalid_fp_register (LDQFA only))
mem_address_not_aligned
privileged_action
data_access_MMU_miss
data_access_exception
data_access_error
data_access_protection

A.27 Load Integer
The LDD instruction is deprecated; it is provided only for compatibility with previous versions of the architecture. It should not be used in new SPARC-V9 software.
It is recommended that the LDX instruction be used in its place.

Opcode
LDSB
LDSH
LDSW
LDUB
LDUH
LDUW
LDX
LDDD

op3
00 1001
00 1010
00 1000
00 0001
00 0010
00 0000
00 1011
00 0011

Operation
Load Signed Byte
Load Signed Halfword
Load Signed Word
Load Unsigned Byte
Load Unsigned Halfword
Load Unsigned Word
Load Extended Word
Load Doubleword

Format (3):
11

op3

rs1

i=0

op3

rs1

i=1

31 30 29

25 24

19 18

ldsb
ldsh
ldsw
ldub
lduh
lduw
ldx
ldd

—

14 13 12

rs2

simm13
5

Suggested Assembly Language Syntax
[address], regrd
[address], regrd
[address], regrd
[address], regrd
[address], regrd
[address], regrd
[address], regrd
[address], regrd

(synonym: ld)

Description:
The load integer instructions copy a byte, a halfword, a word, an extended word, or a doubleword
from memory. All except LDD copy the fetched value into r[rd]. A fetched byte, halfword, or
word is right-justified in the destination register r[rd]; it is either sign-extended or zero-filled on
the left, depending on whether the opcode specifies a signed or unsigned operation, respectively.
The load doubleword integer instructions (LDD) copy a doubleword from memory into an r-register pair. The word at the effective memory address is copied into the even r register. The word at

the effective memory address + 4 is copied into the following odd-numbered r register. The upper
32 bits of both the even-numbered and odd-numbered r registers are zero-filled. Note that a load
doubleword with rd = 0 modifies only r[1]. The least significant bit of the rd field in an LDD
instruction is unused and should be set to zero by software. An attempt to execute a load doubleword instruction that refers to a misaligned (odd-numbered) destination register causes an
illegal_instruction exception.
IMPL. DEP. #107(1): It is implementation-dependent whether LDD is implemented in hardware. If not, an
attempt to execute it will cause an unimplemented_ldd exception.

Load integer instructions access the primary address space (ASI = 8016). The effective address is
“r[rs1] + r[rs2]” if i = 0, or “r[rs1] + sign_ext(simm13)” if i = 1.
A successful load (notably, load extended and load doubleword) instruction operates atomically.
LDUH and LDSH cause a mem_address_not_aligned exception if the address is not halfwordaligned. LDUW and LDSW cause a mem_address_not_aligned exception if the effective address is
not word-aligned. LDX and LDD cause a mem_address_not_aligned exception if the address is not
doubleword-aligned.
Programming Note:
LDD is provided for compatibility with SPARC-V8. It may execute slowly on SPARC-V9 machines because
of data path and register-access difficulties. In some systems it may trap to emulation code. It is suggested
that programmers and compilers avoid using these instructions.
If LDD is emulated in software, an LDX instruction should be used for the memory access in order to preserve atomicity.
Compatibility Note:
The SPARC-V8 LD instruction has been renamed LDUW in SPARC-V9. The LDSW instruction is new in
SPARC-V9.

Exceptions:
async_data_error
unimplemented_LDD (LDD only (impl. dep. #107))
illegal_instruction (LDD with odd rd)
mem_address_not_aligned (all except LDSB, LDUB)
data_access_exception
data_access_protection
data_access_MMU_miss
data_access_error

A.28 Load Integer from Alternate Space
The LDDA instruction is deprecated; it is provided only for compatibility with previous versions of the architecture. It should not be used in new SPARC-V9 software. It is recommended that the LDXA instruction be used in its place.

Opcode
LDSBAPASI
LDSHAPASI
LDSWAPASI
LDUBAPASI
LDUHAPASI
LDUWAPASI
LDXAPASI
LDDAD, PASI

op3
01 1001
01 1010
01 1000
01 0001
01 0010
01 0000
01 1011
01 0011

Operation
Load Signed Byte from Alternate space
Load Signed Halfword from Alternate space
Load Signed Word from Alternate space
Load Unsigned Byte from Alternate space
Load Unsigned Halfword from Alternate space
Load Unsigned Word from Alternate space
Load Extended Word from Alternate space
Load Doubleword from Alternate space

op3

rs1

i=0

op3

rs1

i=1

Format (3):

31 30 29

Description:

25 24

19 18

imm_asi

rs2

simm13

14 13 12

ldsba
ldsha
ldswa
lduba
lduha
lduwa
ldxa

Suggested Assembly Language Syntax
[regaddr] imm_asi, reg rd
[regaddr] imm_asi, reg rd
[regaddr] imm_asi, reg rd
[regaddr] imm_asi, reg rd
[regaddr] imm_asi, reg rd
[regaddr] imm_asi, reg rd
(synonym: lda)
[regaddr] imm_asi, reg rd

ldda
ldsba
ldsha
ldswa
lduba
lduha
lduwa
ldxa
ldda

[regaddr] imm_asi, reg rd
[reg_plus_imm] %asi, reg rd
[reg_plus_imm] %asi, reg rd
[reg_plus_imm] %asi, reg rd
[reg_plus_imm] %asi, reg rd
[reg_plus_imm] %asi, reg rd
[reg_plus_imm] %asi, reg rd
[reg_plus_imm] %asi, reg rd
[reg_plus_imm] %asi, reg rd

(synonym: lda)

The load integer from alternate space instructions copy a byte, a halfword, a word, an extended
word, or a doubleword from memory. All except LDDA copy the fetched value into r[rd]. A
fetched byte, halfword, or word is right-justified in the destination register r[rd]; it is either signextended or zero-filled on the left, depending on whether the opcode specifies a signed or
unsigned operation, respectively.
The load doubleword integer from alternate space instruction (LDDA) copies a doubleword from
memory into an r-register pair. The word at the effective memory address is copied into the even r
register. The word at the effective memory address + 4 is copied into the following odd-numbered
r register. The upper 32 bits of both the even-numbered and odd-numbered r registers are zerofilled. Note that a load doubleword with rd = 0 modifies only r[1]. The least significant bit of the
rd field in an LDDA instruction is unused and should be set to zero by software. An attempt to
execute a load doubleword instruction that refers to a misaligned (odd-numbered) destination register causes an illegal_instruction exception.
IMPL. DEP. #107(2): It is implementation-dependent whether LDDA is implemented in hardware. If not, an
attempt to execute it will cause an unimplemented_ldd exception.

The load integer from alternate space instructions contain the address space identifier (ASI) to be
used for the load in the imm_asi field if i = 0, or in the ASI register if i = 1. The access is privileged if bit seven of the ASI is zero; otherwise, it is not privileged. The effective address for these
instructions is “r[rs1] + r[rs2]” if i = 0, or “r[rs1] + sign_ext(simm13)” if i = 1.
A successful load (notably, load extended and load doubleword) instruction operates atomically.
LDUHA, and LDSHA cause a mem_address_not_aligned exception if the address is not halfwordaligned. LDUWA and LDSWA cause a mem_address_not_aligned exception if the effective
address is not word-aligned; LDXA and LDDA cause a mem_address_not_aligned exception if the
address is not doubleword-aligned.
These instructions cause a privileged_action exception if PSTATE.PRIV = 0 and bit 7 of the ASI is
zero.
Programming Note:
LDDA is provided for compatibility with SPARC-V8. It may execute slowly on SPARC-V9 machines
because of data path and register-access difficulties. In some systems it may trap to emulation code. It is suggested that programmers and compilers avoid using this instruction.
If LDDA is emulated in software, an LDXA instruction should be used for the memory access in order to
preserve atomicity.
Compatibility Note:
The SPARC-V8 instruction LDA has been renamed LDUWA in SPARC-V9. The LDSWA instruction is new
in SPARC-V9.

Exceptions:
async_data_error
privileged_action
unimplemented_LDD (LDDA only (impl. dep. #107))
illegal_instruction (LDDA with odd rd)
mem_address_not_aligned (all except LDSBA and LDUBA)
data_access_exception
data_access_protection

data_access_MMU_miss
data_access_error

A.29 Load-Store Unsigned Byte
Opcode
LDSTUB

op3
00 1101

Operation
Load-Store Unsigned Byte

Format (3):
11

op3

rs1

i=0

op3

rs1

i=1

31 30 29

25 24

19 18

rs2

—

simm13

14 13 12

Suggested Assembly Language Syntax
ldstub [address], regrd

Description:
The load-store unsigned byte instruction copies a byte from memory into r[rd], and then rewrites
the addressed byte in memory to all ones. The fetched byte is right-justified in the destination register r[rd] and zero-filled on the left.
The operation is performed atomically, that is, without allowing intervening interrupts or deferred
traps. In a multiprocessor system, two or more processors executing LDSTUB, LDSTUBA,
CASA, CASXA, SWAP, or SWAPA instructions addressing all or parts of the same doubleword
simultaneously are guaranteed to execute them in an undefined but serial order.
The effective address for these
“r[rs1] + sign_ext(simm13)” if i = 1.

instructions

“r[rs1] + r[rs2]”

i = 0,

The coherence and atomicity of memory operations between processors and I/O DMA memory
accesses are implementation-dependent (impl. dep #120).
Exceptions:
async_data_error
data_access_exception
data_access_error
data_access_protection
data_access_MMU_miss

A.30 Load-Store Unsigned Byte to Alternate Space
Opcode
LDSTUBAPASI

op3
01 1101

Operation
Load-Store Unsigned Byte into Alternate space

Format (3):
11

op3

rs1

i=0

op3

rs1

i=1

31 30 29

25 24

19 18

imm_asi

14 13 12

rs2

simm13
5

Suggested Assembly Language Syntax
ldstuba [regaddr] imm_asi, regrd
ldstuba [reg_plus_imm] %asi, regrd

Description:
The load-store unsigned byte into alternate space instruction copies a byte from memory into
r[rd], then rewrites the addressed byte in memory to all ones. The fetched byte is right-justified in
the destination register r[rd] and zero-filled on the left.
The operation is performed atomically, that is, without allowing intervening interrupts or deferred
traps. In a multiprocessor system, two or more processors executing LDSTUB, LDSTUBA,
CASA, CASXA, SWAP, or SWAPA instructions addressing all or parts of the same doubleword
simultaneously are guaranteed to execute them in an undefined, but serial order.
LDSTUBA contains the address space identifier (ASI) to be used for the load in the imm_asi field
if i = 0, or in the ASI register if i = 1. The access is privileged if bit seven of the ASI is zero; otherwise, it is not privileged. The effective address is “r[rs1] + r[rs2]” if i = 0, or
“r[rs1] + sign_ext(simm13)” if i = 1.
LDSTUBA causes a privileged_action exception if PSTATE.PRIV = 0 and bit 7 of the ASI is zero.
The coherence and atomicity of memory operations between processors and I/O DMA memory
accesses are implementation-dependent (impl. dep #120).
Exceptions:
async_data_error
privileged_action
data_access_exception
data_access_error
data_access_protection
data_access_MMU_miss

A.31 Logical Operations
Opcode
AND
ANDcc
ANDN
ANDNcc
OR
ORcc
ORN
ORNcc
XOR
XORcc
XNOR
XNORcc

op3
00 0001
01 0001
00 0101
01 0101
00 0010
01 0010
00 0110
01 0110
00 0011
01 0011
00 0111
01 0111

Operation
And
And and modify cc’s
And Not
And Not and modify cc’s
Inclusive Or
Inclusive Or and modify cc’s
Inclusive Or Not
Inclusive Or Not and modify cc’s
Exclusive Or
Exclusive Or and modify cc’s
Exclusive Nor
Exclusive Nor and modify cc’s

Format (3):
10

op3

rs1

i=0

op3

rs1

i=1

31 30 29

25 24

19 18

14 13 12

—

rs2

simm13
5

Suggested Assembly Language Syntax
and
regrs1, reg_or_imm, regrd
andcc
regrs1, reg_or_imm, regrd
andn
regrs1, reg_or_imm, regrd
andncc
regrs1, reg_or_imm, regrd
or
regrs1, reg_or_imm, regrd
orcc
regrs1, reg_or_imm, regrd
orn
regrs1, reg_or_imm, regrd
orncc
regrs1, reg_or_imm, regrd
xor
regrs1, reg_or_imm, regrd
xorcc
regrs1, reg_or_imm, regrd
xnor
regrs1, reg_or_imm, regrd
xnorcc
regrs1, reg_or_imm, regrd

Description:
These instructions implement bitwise logical operations. They compute “r[rs1] op r[rs2]” if i = 0,
or “r[rs1] op sign_ext(simm13)” if i = 1, and write the result into r[rd].
ANDcc, ANDNcc, ORcc, ORNcc, XORcc, and XNORcc modify the integer condition codes (icc
and xcc). They set icc.v, icc.c, xcc.v, and xcc.c to zero, icc.n to bit 31 of the result, xcc.n to bit 63

of the result, icc.z to 1 if bits 31:0 of the result are zero (otherwise to 0), and xcc.z to 1 if all 64 bits
of the result are zero (otherwise to 0).
ANDN, ANDNcc, ORN, and ORNcc logically negate their second operand before applying the
main (AND or OR) operation.
Programming Note:
XNOR and XNORcc are identical to the XOR-Not and XOR-Not-cc logical operations, respectively.

Exceptions:
(none)

A.32 Memory Barrier
Opcode
MEMBAR

op3
10 1000

Operation
Memory Barrier

Format (3):
10
31 30 29

op3
25 24

0 1111
19 18

i=1
14 13 12

mmask

cmask

—

4 3

7 6

Suggested Assembly Language Syntax
membar
membar_mask

Description:
The memory barrier instruction, MEMBAR, has two complementary functions: to express order
constraints between memory references and to provide explicit control of memory-reference completion. The membar_mask field in the suggested assembly language is the bitwise OR of the
cmask and mmask instruction fields.
MEMBAR introduces an order constraint between classes of memory references appearing before
the MEMBAR and memory references following it in a program. The particular classes of memory references are specified by the mmask field. Memory references are classified as loads (including load instructions, LDSTUB(A), SWAP(A), CASA, and CASXA) and stores (including store
instructions, LDSTUB(A), SWAP(A), CASA, CASXA, and FLUSH). The mmask field specifies
the classes of memory references subject to ordering, as described below. MEMBAR applies to all
memory operations in all address spaces referenced by the issuing processor, but has no effect on
memory references by other processors. When the cmask field is nonzero, completion as well as
order constraints are imposed, and the order imposed can be more stringent than that specifiable
by the mmask field alone.
A load has been performed when the value loaded has been transmitted from memory and cannot
be modified by another processor. A store has been performed when the value stored has become
visible, that is, when the previous value can no longer be read by any processor. In specifying the
effect of MEMBAR, instructions are considered to be executed as if they were processed in a
strictly sequential fashion, with each instruction completed before the next has begun.
The mmask field is encoded in bits 3 through 0 of the instruction. Table 25 specifies the order constraint that each bit of mmask (selected when set to 1) imposes on memory references appearing
before and after the MEMBAR. From zero to four mask bits may be selected in the mmask field.

Table 25—MEMBAR mmask Encodings
Mask bit
mmask<3>

Name
#StoreStore

Description
The effects of all stores appearing prior to the MEMBAR instruction
must be visible to all processors before the effect of any stores following
the MEMBAR. Equivalent to the deprecated STBAR instruction

mmask<2>

#LoadStore

mmask<1>

#StoreLoad

mmask<0>

#LoadLoad

All loads appearing prior to the MEMBAR instruction must have been
performed before the effect of any stores following the MEMBAR is visible to any other processor.
The effects of all stores appearing prior to the MEMBAR instruction
must be visible to all processors before loads following the MEMBAR
may be performed.
All loads appearing prior to the MEMBAR instruction must have been
performed before any loads following the MEMBAR may be performed.

The cmask field is encoded in bits 6 through 4 of the instruction. Bits in the cmask field, illustrated
in table 26, specify additional constraints on the order of memory references and the processing of
instructions. If cmask is zero, then MEMBAR enforces the partial ordering specified by the
mmask field; if cmask is nonzero, then completion as well as partial order constraints are applied.
Table 26—MEMBAR cmask Encodings
Mask bit
Function
Name
cmask<2> Synchronization #Sync
barrier

cmask<1>

cmask<0>

Description
All operations (including nonmemory reference operations)
appearing prior to the MEMBAR must have been performed
and the effects of any exceptions become visible before any
instruction after the MEMBAR may be initiated.
Memory issue #MemIssue All memory reference operations appearing prior to the
barrier
MEMBAR must have been performed before any memory
operation after the MEMBAR may be initiated.
Lookaside
#Lookaside A store appearing prior to the MEMBAR must complete
barrier
before any load following the MEMBAR referencing the
same address can be initiated.

For information on the use of MEMBAR, see 8.4.3, “The MEMBAR Instruction,” and Appendix
J, “Programming With the Memory Models.” Chapter 8, “Memory Models,” and Appendix F,
“SPARC-V9 MMU Requirements,” contain additional information about the memory models
themselves.
The encoding of MEMBAR is identical to that of the RDASR instruction, except that rs1 = 15,
rd = 0, and i = 1.
The coherence and atomicity of memory operations between processors and I/O DMA memory
accesses are implementation-dependent (impl. dep #120).
Compatibility Note:
MEMBAR with mmask = 816 and cmask = 016 (“membar #StoreStore”) is identical in function to the
SPARC-V8 STBAR instruction, which is deprecated.

Exceptions:
(none)

A.33 Move Floating-Point Register on Condition (FMOVcc)
For Integer Condition Codes:
Opcode
FMOVA
FMOVN
FMOVNE
FMOVE
FMOVG
FMOVLE
FMOVGE
FMOVL
FMOVGU
FMOVLEU
FMOVCC
FMOVCS
FMOVPOS
FMOVNEG
FMOVVC
FMOVVS

op3
11 0101
11 0101
11 0101
11 0101
11 0101
11 0101
11 0101
11 0101
11 0101
11 0101
11 0101
11 0101
11 0101
11 0101
11 0101
11 0101

cond
1000
0000
1001
0001
1010
0010
1011
0011
1100
0100
1101
0101
1110
0110
1111
0111

Operation

icc/xcc test

Move Always
Move Never
Move if Not Equal
Move if Equal
Move if Greater
Move if Less or Equal
Move if Greater or Equal
Move if Less
Move if Greater Unsigned
Move if Less or Equal Unsigned
Move if Carry Clear (Greater or Equal, Unsigned)

1
0
not Z
Z
not (Z or (N xor V))
Z or (N xor V)
not (N xor V)
N xor V
not (C or Z)
(C or Z)
not C
C
not N
N
not V
V

Move if Carry Set (Less than, Unsigned)
Move if Positive
Move if Negative
Move if Overflow Clear
Move if Overflow Set

For Floating-Point Condition Codes:
Opcode
FMOVFA
FMOVFN
FMOVFU
FMOVFG
FMOVFUG
FMOVFL
FMOVFUL
FMOVFLG
FMOVFNE
FMOVFE
FMOVFUE
FMOVFGE
FMOVFUGE
FMOVFLE
FMOVFULE
FMOVFO

op3
11 0101
11 0101
11 0101
11 0101
11 0101
11 0101
11 0101
11 0101
11 0101
11 0101
11 0101
11 0101
11 0101
11 0101
11 0101
11 0101

cond
1000
0000
0111
0110
0101
0100
0011
0010
0001
1001
1010
1011
1100
1101
1110
1111

Operation
Move Always
Move Never
Move if Unordered
Move if Greater
Move if Unordered or Greater
Move if Less
Move if Unordered or Less
Move if Less or Greater
Move if Not Equal
Move if Equal
Move if Unordered or Equal
Move if Greater or Equal
Move if Unordered or Greater or Equal
Move if Less or Equal
Move if Unordered or Less or Equal
Move if Ordered

fcc test
1
0
U
G
G or U
L
L or U
L or G
L or G or U
E
E or U
E or G
E or G or U
E or L
E or L or U
E or L or G

Format (4):
10
31 30 29

op3
25 24

cond

0
19 18 17

opf_cc
14 13

11 10

opf_low

rs2
5

Encoding of the opf_cc field (also see table 38 on page 273):
opf_cc
000
001
010
011
100
101
110
111

Encoding of opf field (opf_cc

Condition code
fcc0
fcc1
fcc2
fcc3
icc
—
xcc
—

opf_low):

Instruction variation
FMOVScc
%fccn,rs2,rd
FMOVDcc
%fccn,rs2,rd
FMOVQcc
%fccn,rs2,rd
FMOVScc
%icc, rs2,rd
FMOVDcc
%icc, rs2,rd
FMOVQcc
%icc, rs2,rd
FMOVScc
%xcc, rs2,rd
FMOVDcc
%xcc, rs2,rd
FMOVQcc
%xcc, rs2,rd

opf_cc
0nn
0nn
0nn
100
100
100
110
110
110

opf_low
00 0001
00 0010
00 0011
00 0001
00 0010
00 0011
00 0001
00 0010
00 0011

opf
0 nn00 0001
0 nn00 0010
0 nn00 0011
1 0000 0001
1 0000 0010
1 0000 0011
1 1000 0001
1 1000 0010
1 1000 0011

For Integer Condition Codes:

fmov{s,d,q}a
fmov{s,d,q}n
fmov{s,d,q}ne
fmov{s,d,q}e
fmov{s,d,q}g
fmov{s,d,q}le
fmov{s,d,q}ge
fmov{s,d,q}l
fmov{s,d,q}gu
fmov{s,d,q}leu
fmov{s,d,q}cc
fmov{s,d,q}cs
fmov{s,d,q}pos
fmov{s,d,q}neg
fmov{s,d,q}vc
fmov{s,d,q}vs

Suggested Assembly Language Syntax
i_or_x_cc, fregrs2, fregrd
i_or_x_cc, fregrs2, fregrd
i_or_x_cc, fregrs2, fregrd
(synonyms: fmov{s,d,q}nz)
i_or_x_cc, fregrs2, fregrd
(synonyms: fmov{s,d,q}z)
i_or_x_cc, fregrs2, fregrd
i_or_x_cc, fregrs2, fregrd
i_or_x_cc, fregrs2, fregrd
i_or_x_cc, fregrs2, fregrd
i_or_x_cc, fregrs2, fregrd
i_or_x_cc, fregrs2, fregrd
i_or_x_cc, fregrs2, fregrd
(synonyms: fmov{s,d,q}geu)
(synonyms: fmov{s,d,q}lu)
i_or_x_cc, fregrs2, fregrd
i_or_x_cc, fregrs2, fregrd
i_or_x_cc, fregrs2, fregrd
i_or_x_cc, fregrs2, fregrd
i_or_x_cc, fregrs2, fregrd

Programming Note:
To select the appropriate condition code, include “%icc” or “%xcc” before the registers.

For Floating-Point Condition Codes:

fmov{s,d,q}a
fmov{s,d,q}n
fmov{s,d,q}u
fmov{s,d,q}g
fmov{s,d,q}ug
fmov{s,d,q}l
fmov{s,d,q}ul
fmov{s,d,q}lg
fmov{s,d,q}ne
fmov{s,d,q}e
fmov{s,d,q}ue
fmov{s,d,q}ge
fmov{s,d,q}uge
fmov{s,d,q}le
fmov{s,d,q}ule
fmov{s,d,q}o

Suggested Assembly Language Syntax
%fccn, fregrs2, fregrd
%fccn, fregrs2, fregrd
%fccn, fregrs2, fregrd
%fccn, fregrs2, fregrd
%fccn, fregrs2, fregrd
%fccn, fregrs2, fregrd
%fccn, fregrs2, fregrd
%fccn, fregrs2, fregrd
%fccn, fregrs2, fregrd
(synonyms: fmov{s,d,q}nz)
%fccn, fregrs2, fregrd
(synonyms: fmov{s,d,q}z)
%fccn, fregrs2, fregrd
%fccn, fregrs2, fregrd
%fccn, fregrs2, fregrd
%fccn, fregrs2, fregrd
%fccn, fregrs2, fregrd
%fccn, fregrs2, fregrd

Description:
These instructions copy the floating-point register(s) specified by rs2 to the floating-point register(s) specified by rd if the condition indicated by the cond field is satisfied by the selected condition code. The condition code used is specified by the opf_cc field of the instruction. If the
condition is FALSE, then the destination register(s) are not changed.
These instructions do not modify any condition codes.
Programming Note:
Branches cause most implementations’ performance to degrade significantly. Frrequently, the MOVcc and
FMOVcc instructions can be used to avoid branches. For example, the following C language segment:
double A, B, X;
if (A > B) then X

= 1.03; else X

= 0.0;

can be coded as
! assume A is in %f0; B is in %f2; %xx points to constant area
ldd
[%xx+C_1.03],%f4 ! X = 1.03
fcmpd
%fcc3,%f0,%f2
! A > B
fble ,a
%fcc3,label
! following only executed if the branch is taken
fsubd
%f4,%f4,%f4
! X = 0.0
label:...
This takes four instructions including a branch.
Using FMOVcc, this could be coded as
ldd
fsubd
fcmpd
fmovdle

[%xx+C_1.03],%f4
%f4,%f4,%f6
%fcc3,%f0,%f2
%fcc3,%f6,%f4

!
!
!
!

X = 1.03
X’ = 0.0
A > B
X = 0.0

This also takes four instructions, but requires no branches and may boost performance significantly. It is suggested that MOVcc and FMOVcc be used instead of branches wherever they would improve performance.

Exceptions:
fp_disabled
fp_exception_other (invalid_fp_register (quad forms only))
fp_exception_other (ftt = unimplemented_FPop (opf_cc = 1012 or 1112)

A.34 Move F-P Register on Integer Register Condition (FMOVr)
Opcode
—
FMOVRZ
FMOVRLEZ
FMOVRLZ
—
FMOVRNZ
FMOVRGZ
FMOVRGEZ

op3
11 0101
11 0101
11 0101
11 0101
11 0101
11 0101
11 0101
11 0101

rcond
000
001
010
011
100
101
110
111

Operation
Reserved
Move if Register Zero
Move if Register Less Than or Equal to Zero
Move if Register Less Than Zero
Reserved
Move if Register Not Zero
Move if Register Greater Than Zero
Move if Register Greater Than or Equal to Zero

Test
—
r[rs1] = 0
r[rs1] ≤ 0
r[rs1] < 0
—
r[rs1] ≠ 0
r[rs1] > 0
r[rs1] ≥ 0

Format (4):
10

31 30 29

op3
25 24

rs1
19 18

rcond

14 13 12

10 9

opf_low

rs2
5

Encoding of opf_low field:
Instruction variation
FMOVSrcond rs1, rs2, rd
FMOVDrcond rs1, rs2, rd
FMOVQrcond rs1, rs2, rd

opf_low
0 0101
0 0110
0 0111

Suggested Assembly Language Syntax
(synonym: fmovr{s,d,q}z)
fmovr{s,d,q}e
regrs1, fregrs2, fregrd
fmovr{s,d,q}lez regrs1, fregrs2, fregrd
fmovr{s,d,q}lz
regrs1, fregrs2, fregrd
fmovr{s,d,q}ne
regrs1, fregrs2, fregrd
(synonym: fmovr{s,d,q}nz)
fmovr{s,d,q}gz
regrs1, fregrs2, fregrd
fmovr{s,d,q}gez regrs1, fregrs2, fregrd

Description:
If the contents of integer register r[rs1] satisfy the condition specified in the rcond field, these
instructions copy the contents of the floating-point register(s) specified by the rs2 field to the
floating-point register(s) specified by the rd field. If the contents of r[rs1] do not satisfy the condition, the floating-point register(s) specified by the rd field are not modified.
These instructions treat the integer register contents as a signed integer value; they do not modify
any condition codes.

Implementation Note:
If this instruction is implemented by tagging each register value with an N (negative) and a Z (zero) bit, use
the following table to determine whether rcond is TRUE:
Branch
FMOVRNZ
FMOVRZ
FMOVGEZ
FMOVRLZ
FMOVRLEZ
FMOVRGZ

Test
not Z
Z
not N
N
N or Z
N nor Z

Exceptions:
fp_disabled
fp_exception_other (invalid_fp_register (quad forms only))
fp_exception_other (unimplemented_FPop (rcond = 0002 or 1002))

A.35 Move Integer Register on Condition (MOVcc)
For Integer Condition Codes:
Opcode
MOVA
MOVN
MOVNE
MOVE
MOVG
MOVLE
MOVGE
MOVL
MOVGU
MOVLEU
MOVCC
MOVCS
MOVPOS
MOVNEG
MOVVC
MOVVS

op3
10 1100
10 1100
10 1100
10 1100
10 1100
10 1100
10 1100
10 1100
10 1100
10 1100
10 1100
10 1100
10 1100
10 1100
10 1100
10 1100

cond
1000
0000
1001
0001
1010
0010
1011
0011
1100
0100
1101
0101
1110
0110
1111
0111

Operation
Move Always
Move Never
Move if Not Equal
Move if Equal
Move if Greater
Move if Less or Equal
Move if Greater or Equal
Move if Less
Move if Greater Unsigned
Move if Less or Equal Unsigned
Move if Carry Clear (Greater or Equal, Unsigned)
Move if Carry Set (Less than, Unsigned)
Move if Positive
Move if Negative
Move if Overflow Clear
Move if Overflow Set

icc/xcc test
1
0
not Z
Z
not (Z or (N xorV))
Z or (N xorV)
not (N xorV)
N xorV
not (C orZ)
(C orZ)
not C
C
not N
N
not V
V

For Floating-Point Condition Codes:
Opcode
MOVFA
MOVFN
MOVFU
MOVFG
MOVFUG
MOVFL
MOVFUL
MOVFLG
MOVFNE
MOVFE
MOVFUE
MOVFGE
MOVFUGE
MOVFLE
MOVFULE
MOVFO

op3
10 1100
10 1100
10 1100
10 1100
10 1100
10 1100
10 1100
10 1100
10 1100
10 1100
10 1100
10 1100
10 1100
10 1100
10 1100
10 1100

cond
1000
0000
0111
0110
0101
0100
0011
0010
0001
1001
1010
1011
1100
1101
1110
1111

fcc test
1
0
U
G
G or U
L
L or U
L or G
L or G or U
E
E or U
E or G
E or G or U
E or L
E or L or U
E or L or G

Format (4):
10

op3

cc2

cond

i=0 cc1 cc0

op3

cc2

cond

i=1 cc1 cc0

simm11

14 13 12 11 10

31 30 29

25 24

19 18 17

cc2

cc1
000
001
010
011
100
101
110
111

cc0

—

rs2

Condition code
fcc0
fcc1
fcc2
fcc3
icc
Reserved
xcc
Reserved

For Integer Condition Codes:

mova
movn
movne
move
movg
movle
movge
movl
movgu
movleu
movcc
movcs
movpos
movneg
movvc
movvs

Suggested Assembly Language Syntax
i_or_x_cc, reg_or_imm11, regrd
i_or_x_cc, reg_or_imm11, regrd
i_or_x_cc, reg_or_imm11, regrd
(synonym: movnz)
i_or_x_cc, reg_or_imm11, regrd
(synonym: movz)
i_or_x_cc, reg_or_imm11, regrd
i_or_x_cc, reg_or_imm11, regrd
i_or_x_cc, reg_or_imm11, regrd
i_or_x_cc, reg_or_imm11, regrd
i_or_x_cc, reg_or_imm11, regrd
i_or_x_cc, reg_or_imm11, regrd
i_or_x_cc, reg_or_imm11, regrd
(synonym: movgeu)
i_or_x_cc, reg_or_imm11, regrd
(synonym: movlu)
i_or_x_cc, reg_or_imm11, regrd
i_or_x_cc, reg_or_imm11, regrd
i_or_x_cc, reg_or_imm11, regrd
i_or_x_cc, reg_or_imm11, regrd

Programming Note:
To select the appropriate condition code, include “%icc” or “%xcc” before the register or immediate field.

For Floating-Point Condition Codes:

mova

Suggested Assembly Language Syntax
%fccn, reg_or_imm11, regrd

movn
movu
movg
movug
movl
movul
movlg
movne
move
movue
movge
movuge
movle
movule
movo

%fccn,
%fccn,
%fccn,
%fccn,
%fccn,
%fccn,
%fccn,
%fccn,
%fccn,
%fccn,
%fccn,
%fccn,
%fccn,
%fccn,
%fccn,

reg_or_imm11, regrd
reg_or_imm11, regrd
reg_or_imm11, regrd
reg_or_imm11, regrd
reg_or_imm11, regrd
reg_or_imm11, regrd
reg_or_imm11, regrd
reg_or_imm11, regrd
reg_or_imm11, regrd
reg_or_imm11, regrd
reg_or_imm11, regrd
reg_or_imm11, regrd
reg_or_imm11, regrd
reg_or_imm11, regrd
reg_or_imm11, regrd

(synonym: movnz)
(synonym: movz)

Programming Note:
To select the appropriate condition code, include “%fcc0,” “%fcc1,” “%fcc2,” or “%fcc3” before the
register or immediate field.

Description:
These instructions test to see if cond is TRUE for the selected condition codes. If so, they copy the
value in r[rs2] if i field = 0, or “sign_ext(simm11)” if i = 1 into r[rd]. The condition code used is
specified by the cc2, cc1, and cc0 fields of the instruction. If the condition is FALSE, then r[rd] is
not changed.
These instructions copy an integer register to another integer register if the condition is TRUE.
The condition code that is used to determine whether the move will occur can be either integer
condition code (icc or xcc) or any floating-point condition code (fcc0, fcc1, fcc2, or fcc3).
These instructions do not modify any condition codes.
Programming Note:
Branches cause many implementations’ performance to degrade significantly. Frequently, the MOVcc and
FMOVcc instructions can be used to avoid branches. For example, the C language if-then-else statement
if (A > B) then X = 1; else X = 0;
can be coded as
cmp
bg,a
or
or
label:...

%i0,%i2
%xcc,label
%g0,1,%i3
%g0,0,%i3

! X
! X

= 1
= 0

This takes four instructions including a branch. Using MOVcc this could be coded as
cmp
or
movle

%i0,%i2
%g0,1,%i3
%xcc,0,%i3

! assume X = 1
! overwrite with X = 0

This takes only three instructions and no branches and may boost performance significantly. It is suggested
that MOVcc and FMOVcc be used instead of branches wherever they would increase performance.

Exceptions:
illegal_instruction (cc2
cc1 cc0 = 1012 or 1112)
fp_disabled (cc2
cc1 cc0 = 0002 , 0012 , 0102 , or 0112 and the FPU is disabled)

A.36 Move Integer Register on Register Condition (MOVR)
Opcode
—
MOVRZ
MOVRLEZ
MOVRLZ
—
MOVRNZ
MOVRGZ
MOVRGEZ

op3
10 1111
10 1111
10 1111
10 1111
10 1111
10 1111
10 1111
10 1111

rcond
000
001
010
011
100
101
110
111

Test
—
r[rs1] = 0
r[rs1] ≤ 0
r[rs1] < 0
—
r[rs1] ≠ 0
r[rs1] > 0
r[rs1] ≥ 0

Format (3):
10

op3

rs1

i=0 rcond

op3

rs1

i=1 rcond

31 30 29

25 24

movrz
movrlez
movrlz
movrnz
movrgz
movrgez

19 18

14 13 12

—

10 9

rs2

simm10
5

Suggested Assembly Language Syntax
regrs1, reg_or_imm10, regrd
(synonym: movre)
regrs1, reg_or_imm10, regrd
regrs1, reg_or_imm10, regrd
regrs1, reg_or_imm10, regrd
(synonym: movrne)
regrs1, reg_or_imm10, regrd
regrs1, reg_or_imm10, regrd

Description:
If the contents of integer register r[rs1] satisfies the condition specified in the rcond field, these
instructions copy r[rs2] (if i = 0) or sign_ext(simm10) (if i = 1) into r[rd]. If the contents of r[rs1]
does not satisfy the condition then r[rd] is not modified. These instructions treat the register contents as a signed integer value; they do not modify any condition codes.
Implementation Note:
If this instruction is implemented by tagging each register value with an N (negative) and a Z (zero) bit, use
the following table to determine if rcond is TRUE:
Branch
MOVRNZ
MOVRZ
MOVRGEZ
MOVRLZ
MOVRLEZ
MOVRGZ

Test
not Z
Z
not N
N
N or Z
N nor Z

Exceptions:
illegal_instruction (rcond = 0002 or 1002)

A.37 Multiply and Divide (64-bit)
Opcode
MULX
SDIVX
UDIVX

op3
00 1001
10 1101
00 1101

Operation
Multiply (signed or unsigned)
Signed Divide
Unsigned Divide

Format (3):
10

op3

rs1

i=0

op3

rs1

i=1

31 30 29

25 24

19 18

14 13 12

—

rs2

simm13
5

Suggested Assembly Language Syntax
mulx
regrs1, reg_or_imm, regrd
sdivx regrs1, reg_or_imm, regrd
udivx regrs1, reg_or_imm, regrd

Description:
MULX computes “r[rs1] × r[rs2]” if i = 0 or “r[rs1] × sign_ext(simm13)” if i = 1, and writes the
64-bit product into r[rd]. MULX can be used to calculate the 64-bit product for signed or
unsigned operands (the product is the same).
SDIVX and UDIVX compute “r[rs1] ÷ r[rs2]” if i = 0 or “r[rs1] ÷ sign_ext(simm13)” if i = 1,
and write the 64-bit result into r[rd]. SDIVX operates on the operands as signed integers and produces a corresponding signed result. UDIVX operates on the operands as unsigned integers and
produces a corresponding unsigned result.
For SDIVX, if the largest negative number is divided by –1, the result should be the largest negative number. That is:
8000 0000 0000 000016 ÷ FFFF FFFF FFFF FFFF16 = 8000 0000 0000 000016.
These instructions do not modify any condition codes.
Exceptions:
division_by_zero

A.38 Multiply (32-bit)
The UMUL, UMULcc, SMUL, and SMULcc instructions are deprecated; they are
provided only for compatibility with previous versions of the architecture. They
should not be used in new SPARC-V9 software. It is recommended that the MULX
instruction be used in their place.

Opcode
UMULD
SMULD
UMULccD
SMULccD

op3
00 1010
00 1011
01 1010
01 1011

Operation
Unsigned Integer Multiply
Signed Integer Multiply
Unsigned Integer Multiply and modify cc’s
Signed Integer Multiply and modify cc’s

Format (3):
10

op3

rs1

i=0

op3

rs1

i=1

31 30 29

25 24

19 18

14 13 12

—

rs2

simm13
5

Suggested Assembly Language Syntax
umul
regrs1, reg_or_imm, regrd
smul
regrs1, reg_or_imm, regrd
umulcc
regrs1, reg_or_imm, regrd
smulcc
regrs1, reg_or_imm, regrd

Description:
The multiply instructions perform 32-bit by 32-bit multiplications, producing 64-bit results. They
compute “r[rs1]<31:0> × r[rs2]<31:0>” if i = 0, or “r[rs1]<31:0> × sign_ext(simm13)<31:0>” if
i = 1. They write the 32 most significant bits of the product into the Y register and all 64 bits of
the product into r[rd].
Unsigned multiply (UMUL, UMULcc) operates on unsigned integer word operands and computes an unsigned integer doubleword product. Signed multiply (SMUL, SMULcc) operates on
signed integer word operands and computes a signed integer doubleword product.
UMUL and SMUL do not affect the condition code bits. UMULcc and SMULcc write the integer
condition code bits, icc and xcc, as follows. Note that 32-bit negative (icc.N) and zero (icc.Z) condition codes are set according to the less significant word of the product, and not according to the
full 64-bit result.

Bit
icc.N
icc.Z
icc.V
icc.C
xcc.N
xcc.Z
xcc.V
xcc.C

UMULcc / SMULcc
Set if product[31] = 1
Set if product[31:0] = 0
Zero
Zero
Set if product[63] = 1
Set if product[63:0] = 0
Zero
Zero

Programming Note:
32-bit overflow after UMUL / UMULcc is indicated by Y ≠ 0.
32-bit overflow after SMUL / SMULcc is indicated by Y ≠ (r[rd] >> 31), where “>>” indicates 32-bit arithmetic right shift.
Implementation Note:
An implementation may assume that the smaller operand typically will be r[rs2] or simm13.
Implementation Note:
See Implementation Characteristics of Current SPARC-V9-based Products, Revision 9.x, a document available from SPARC International, for information on whether these instructions are implemented by
hardware or software in the various SPARC-V9 implementations.

Exceptions:
(none)

A.39 Multiply Step
The MULScc instruction is deprecated; it is provided only for compatibility with
previous versions of the architecture. It should not be used in new SPARC-V9 software. It is recommended that the MULX instruction be used in its place.

Opcode
MULSccD

op3
10 0100

Operation
Multiply Step and modify cc’s

Format (3):
10

op3

rs1

i=0

op3

rs1

i=1

31 30 29

25 24

19 18

—

14 13 12

rs2

simm13
5

Suggested Assembly Language Syntax
mulscc
regrs1, reg_or_imm, regrd

Description:
MULScc treats the lower 32 bits of both r[rs1] and the Y register as a single 64-bit, right-shiftable
doubleword register. The least significant bit of r[rs1] is treated as if it were adjacent to bit 31 of
the Y register. The MULScc instruction adds, based on the least significant bit of Y.
Multiplication assumes that the Y register initially contains the multiplier, r[rs1] contains the
most significant bits of the product, and r[rs2] contains the multiplicand. Upon completion of the
multiplication, the Y register contains the least significant bits of the product.
Note that a standard MULScc instruction has rs1 = rd.
MULScc operates as follows:
(1) The multiplicand is r[rs2] if i = 0, or sign_ext(simm13) if i = 1.
(2) A 32-bit value is computed by shifting r[rs1] right by one bit with
“CCR.icc.n xor CCR.icc.v” replacing bit 31 of r[rs1]. (This is the proper sign for the previous partial product.)
(3) If the least significant bit of Y = 1, the shifted value from step (2) and the multiplicand are
added. If the least significant bit of the Y = 0, then 0 is added to the shifted value from step
(2).

(4) The sum from step (3) is written into r[rd]. The upper 32-bits of r[rd] are undefined. The
integer condition codes are updated according to the addition performed in step (3). The
values of the extended condition codes are undefined.
(5) The Y register is shifted right by one bit, with the least significant bit of the unshifted
r[rs1] replacing bit 31of Y.
Exceptions:
(none)

A.40 No Operation
Opcode
NOP

op
00

op2
100

Operation
No Operation

Format (2):
00
31 30 29

op2
25 24

0000000000000000000000
22 21

Suggested Assembly Language Syntax
nop

Description:
The NOP instruction changes no program-visible state (except the PC and nPC).
Note that NOP is a special case of the SETHI instruction, with imm22 = 0 and rd = 0.
Exceptions:
(none)

A.41 Population Count
Opcode
POPC

op3
10 1110

Operation
Population Count

Format (3):
10

op3

0 0000

i=0

op3

0 0000

i=1

31 30 29

25 24

19 18

14 13 12

—

rs2

simm13
5

Suggested Assembly Language Syntax
popc
reg_or_imm, regrd

Description:
POPC counts the number of one bits in r[rs2] if i = 0, or the number of one bits in
sign_ext(simm13) if i = 1, and stores the count in r[rd]. This instruction does not modify the condition codes.
Implementation Note:
Instruction bits 18 through 14 must be zero for POPC. Other encodings of this field (rs1) may be used in
future versions of the SPARC architecture for other instructions.
Programming Note:
POPC can be used to “find first bit set” in a register. A C program illustrating how POPC can be used for this
purpose follows:
int ffs(zz)
/* finds first 1 bit, counting from the LSB */
unsigned zz;
{
return popc ( zz ^ (∼ (–zz))); /* for nonzero zz */
}
Inline assembly language code for ffs( ) is
neg
xnor
popc
movrz

%IN, %M_IN
%IN, %M_IN, %TEMP
%TEMP,%RESULT
%IN,%g0,%RESULT

!
!
!
!

–zz( 2’s complement)
^ ∼ –zz (exclusive nor)
result = popc( zz ^ ∼ –zz)
%RESULT should be 0 for %IN=0

where IN, M_IN, TEMP, and RESULT are integer registers.

Example:
IN
–IN
∼ –IN
IN ^ ∼ –IN
popc(IN ^ ∼ –IN)

Exceptions:

=
=
=
=
=

...00101000! 1st 1 bit from rt is 4th bit
...11011000
...00100111
...00001111
4

illegal_instruction (instruction<18:14> ≠ 0)

A.42 Prefetch Data
Opcode
PREFETCH
PREFETCHAPASI

op3
10 1101
11 1101

Operation
Prefetch Data
Prefetch Data from Alternate Space

Format (3) PREFETCH:
11

fcn

op3

rs1

i=0

fcn

op3

rs1

i=1

31 30 29

25 24

19 18

—

rs2

simm13

14 13 12

Format (3) PREFETCHA:
11

fcn

op3

rs1

i=0

fcn

op3

rs1

i=1

31 30 29

25 24

19 18

fcn
0
1
2
3
4
5–15
16–31

imm_asi

14 13 12

rs2

simm13
5

Prefetch function
Prefetch for several reads
Prefetch for one read
Prefetch for several writes
Prefetch for one write
Prefetch page
Reserved
Implementation-dependent

Suggested Assembly Language Syntax
prefetch
[address], prefetch_fcn
prefetcha
[regaddr] imm_asi, prefetch_fcn
prefetcha
[reg_plus_imm] %asi, prefetch_fcn

Description:
In nonprivileged code, a prefetch instruction has the same observable effect as a NOP; its execution is nonblocking and cannot cause an observable trap. In particular, a prefetch instruction shall
not trap if it is applied to an illegal or nonexistent memory address.
IMPL. DEP. #103(1): Whether the execution of a PREFETCH instruction has observable effects in privileged code is implementation-dependent.

IMPL. DEP. #103(2): Whether the execution of a PREFETCH
data_access_mmu_miss exception is implementation-dependent.

instruction

can

cause

Whether prefetch always succeeds when the MMU is disabled is implementation-dependent (impl.
dep. # 117).
Implementation Note:
Any effects of prefetch in privileged code should be reasonable (e.g., handling ECC errors, no page prefetching allowed within code that handles page faults). The benefits of prefetching should be available to most
privileged code.

Execution of a prefetch instruction initiates data movement (or preparation for future data movement or address mapping) to reduce the latency of subsequent loads and stores to the specified
address range.
A successful prefetch initiates movement of a block of data containing the addressed byte from
memory toward the processor.
IMPL. DEP. #103(3): The size and alignment in memory of the data block is implementation-dependent;
the minimum size is 64 bytes and the minimum alignment is a 64-byte boundary.
Programming Note:
Software may prefetch 64 bytes beginning at an arbitrary address address by issuing the instructions
prefetch
prefetch

[address], prefetch_fcn
[address + 63], prefetch_fcn

Implementation Note:
Prefetching may be used to help manage memory cache(s). A prefetch from a nonprefetchable location has
no effect. It is up to memory management hardware to determine how locations are identified as not
prefetchable.

Prefetch instructions that do not load from an alternate address space access the primary address
space (ASI_PRIMARY{_LITTLE}). Prefetch instructions that do load from an alternate address
space contain the address space identifier (ASI) to be used for the load in the imm_asi field if
i = 0, or in the ASI register if i = 1. The access is privileged if bit seven of the ASI is zero; otherwise, it is not privileged. The effective address for these instructions is “r[rs1] + r[rs2]” if i = 0, or
“r[rs1] + sign_ext(simm13)” if i = 1.
Variants of the prefetch instruction can be used to prepare the memory system for different types
of accesses.
IMPL. DEP. #103(4): An implementation may implement none, some, or all of these variants. A variant not
implemented shall execute as a nop. An implemented variant may support its full semantics, or may support just the simple common-case prefetching semantics.

A.42.1 Prefetch Variants
The prefetch variant is selected by the fcn field of the instruction. fcn values 5..15 are reserved for
future extensions of the architecture.

IMPL. DEP. #103(5): PREFETCH fcn values of 16..31 are implementation-dependent.

Each prefetch variant reflects an intent on the part of the compiler or programmer. This is different
from other instructions in SPARC-V9 (except BPN), all of which specify specific actions. An
implementation may implement a prefetch variant by any technique, as long as the intent of the
variant is achieved.
The prefetch instruction is designed to treat the common cases well. The variants are intended to
provide scalability for future improvements in both hardware and compilers. If a variant is implemented, then it should have the effects described below. In case some of the variants listed below
are implemented and some are not, there is a recommended overloading of the unimplemented
variants (see the Implementation Note labeled “Recommended Overloadings” in A.42.2).
A.42.1.1 Prefetch for Several Reads (fcn = 0)
The intent of this variant is to cause movement of data into the data cache nearest the processor,
with “reasonable” efforts made to obtain the data.
Implementation Note:
If, for example, some TLB misses are handled in hardware, then they should be handled. On the other hand,
a multiple ECC error is reasonable cause for cancellation of a prefetch.

This is the most important case of prefetching.
If the addressed data is already present (and owned, if necessary) in the cache, then this variant
has no effect.
A.42.1.2 Prefetch for One Read (fcn = 1)
This variant indicates that, if possible, the data cache should be minimally disturbed by the data
read from the given address, because that data is expected to be read once and not reused (read or
written) soon after that.
If the data is already present in the cache, then this variant has no effect.
Programming Note:
The intended use of this variant is in streaming large amounts of data into the processor without overwriting
data in cache memory.

A.42.1.3 Prefetch for Several Writes (and Possibly Reads) (fcn = 2)
The intent of this variant is to cause movement of data in preparation for writing.
If the addressed data is already present in the data cache, then this variant has no effect.
Programming Note:
An example use of this variant is to write a dirty cache line back to memory, or to initialize a cache line in
preparation for a partial write.

Implementation Note:
On a multiprocessor, this variant indicates that exclusive ownership of the addressed data is needed, so it
may have the additional effect of obtaining exclusive ownership of the addressed cache line.
Implementation Note:
On a uniprocessor, there is no distinction between Prefetch for Several Reads and this variant.

A.42.1.4 Prefetch for One Write (fcn = 3)
This variant indicates that, if possible, the data cache should be minimally disturbed by the data
written to this address, because that data is not expected to be reused (read or written) soon after it
has been written once.
If the data is already present in the cache, then this variant has no effect.
A.42.1.5 Prefetch Page (fcn = 4)
In a virtual-memory system, the intended action of this variant is for the supervisor software or
hardware to initiate asynchronous mapping of the referenced virtual address, assuming that it is
legal to do so.
Programming Note:
The desire is to avoid a later page fault for the given address, or at least to shorten the latency of a page fault.

In a nonvirtual-memory system, or if the addressed page is already mapped, this variant has no
effect.
The referenced page need not be mapped when the instruction completes. Loads and stores issued
before the page is mapped should block just as they would if the prefetch had never been issued.
When the activity associated with the mapping has completed, the loads and stores may proceed.
Implementation Note:
An example of mapping activity is DMA from secondary storage.
Implementation Note:
Use of this variant may be disabled or restricted in privileged code that is not permitted to cause page faults.

A.42.1.6 Implementation-Dependent Prefetch (fcn = 16..31)
These values are available for implementations to use. An implementation shall treat any unimplemented prefetch fcn values as NOPs (impl. dep. #103).
A.42.2 General Comments
There is no variant of PREFETCH for instruction prefetching. Instruction prefetching should be
encoded using the Branch Never (BPN) form of the BPcc instruction (see A.7, “Branch on Integer
Condition Codes with Prediction (BPcc)”).

One error to avoid in thinking about prefetch instructions is that they should have “no cost to execute.” As long as the cost of executing a prefetch instruction is well less than one-third the cost of
a cache miss, use of prefetching is a net win. It does not appear that prefetching causes a significant number of useless fetches from memory, though it may increase the rate of useful fetches
(and hence the bandwidth), because it more efficiently overlaps computing with fetching.
Implementation Note:
Recommended Overloadings. There are four recommended sets of overloadings for the prefetch variants,
based on a simplistic classification of SPARC-V9 systems into cost (low-cost vs. high-cost) and processor
multiplicity (uniprocessor vs. multiprocessor) categories. These overloadings are chosen to help ensure efficient portability of software across a range of implementations.
In a uniprocessor, there is no need to support multiprocessor cache protocols; hence, Prefetch for Several
Reads and Prefetch for Several Writes may behave identically. In a low-cost implementation, Prefetch for
One Read and Prefetch for One Write may be identical to Prefetch for Several Reads and Prefetch for Several Writes, respectively.

Multiplicity

Cost

Uniprocessor

Low

Uniprocessor

High

Multiprocessor

Low

Multiprocessor

High

Prefetch for ..
One read
Several reads
One write
Several writes
One read
Several reads
One write
Several writes
One read
Several reads
One write
Several writes
One read
Several reads
One write
Several writes

Could be overloaded
to mean the same as
Prefetch for ..
Several writes
Several writes
Several writes
—
—
Several writes
—
—
Several reads
—
Several writes
—
—
—
—
—

Programming Note:
A SPARC-V9 compiler that generates PREFETCH instructions should generate each of the four variants
where it is most appropriate. The overloadings suggested in the previous Implementation NoteNOTE ensure
that such code will be portable and reasonably efficient across a range of hardware configurations.
Implementation Note:
The Prefetch for One Read and Prefetch for One Write variants assume the existence of a “bypass cache,” so
that the bulk of the “real cache” remains undisturbed. If such a bypass cache is used, it should be large
enough to properly shield the processor from memory latency. Such a cache should probably be small,
highly associative, and use a FIFO replacement policy.

Exceptions:

data_access_MMU_miss (implementation-dependent (impl. dep. #103))
illegal_instruction (fcn=5..15)

A.43 Read Privileged Register
Opcode
RDPRP

op3
10 1010

Operation
Read Privileged Register

Format (3):
10
31 30 29

op3
25 24

rs1
19 18

rs1
0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16..30
31

—
14 13

Privileged register
TPC
TNPC
TSTATE
TT
TICK
TBA
PSTATE
TL
PIL
CWP
CANSAVE
CANRESTORE
CLEANWIN
OTHERWIN
WSTATE
FQ
—
VER

Suggested Assembly Language Syntax
rdpr
%tpc, regrd
rdpr
%tnpc, regrd
rdpr
%tstate, regrd
rdpr
%tt, regrd
rdpr
%tick, regrd
rdpr
%tba, regrd
rdpr
%pstate, regrd
rdpr
%tl, regrd
rdpr
%pil, regrd
rdpr
%cwp, regrd
rdpr
%cansave, regrd
rdpr
%canrestore, regrd
rdpr
%cleanwin, regrd
rdpr
%otherwin, regrd
rdpr
%wstate, regrd
rdpr
%fq, regrd
rdpr
%ver, regrd

Description:
The rs1 field in the instruction determines the privileged register that is read. There are MAXTL
copies of the TPC, TNPC, TT, and TSTATE registers. A read from one of these registers returns
the value in the register indexed by the current value in the trap level register (TL). A read of TPC,
TNPC, TT, or TSTATE when the trap level is zero (TL = 0) causes an illegal_instruction exception.
RDPR instructions with rs1 in the range 16..30 are reserved; executing a RDPR instruction with
rs1 in that range causes an illegal_instruction exception.
A read from the FQ (Floating-Point Deferred-Trap Queue) register copies the front doubleword of
the queue into r[rd]. The semantics of reading the FQ and the data returned are implementationdependent (impl. dep. #24). However, the address of a trapping floating-point instruction must be
available to the privileged trap handler. On an implementation with a floating-point queue, an
attempt to execute RDPR of FQ when the queue is empty (FSR.qne = 0) shall cause an
fp_exception exception with FSR.ftt set to 4 (sequence_error). In an implementation without a floating-point queue, an attempt to execute RDPR of FQ shall cause either an illegal_instruction exception or an fp_exception_other exception with FSR.ftt set to 3 (unimplemented_FPop) (impl. dep.
#25).
Programming Note:
On an implementation with precise floating-point traps, the address of a trapping instruction will be in the
TPC[TL] register when the trap code begins execution. On an implementation with deferred floating-point
traps, the address of the trapping instruction might be a value obtained from the FQ.

Exceptions:
privileged_opcode
illegal_instruction ((rs1 = 16..30) or ((rs1≤3) and (TL = 0)))

fp_exception_other (sequence_error) (RDPR of FQ when FSR.qne = 0 in a system with an

FQ; (impl. dep. #25)
illegal_instruction (RDPR of FQ in a system without an FQ; (impl. dep. #25)

A.44 Read State Register
The RDY instruction is deprecated; it is provided only for compatibility with previous versions of the architecture. It should not be used in new SPARC-V9 software.
It is recommended that all instructions which reference the Y register be avoided.

Opcode
RDYD
—
RDCCR
RDASI
RDTICKPNPT
RDPC
RDFPRS
RDASRPASR
See text
RDASRPASR

op3
10 1000
10 1000
10 1000
10 1000
10 1000
10 1000
10 1000
10 1000
10 1000
10 1000

rs1
0
1
2
3
4
5
6
7−14
15
16−31

Operation
Read Y Register
reserved
Read Condition Codes Register
Read ASI Register
Read Tick Register
Read Program Counter
Read Floating-Point Registers Status Register
Read Ancillary State Register (reserved)
See text
Implementation-dependent (impl. dep. #47)

Format (3):
10
31 30 29

op3
25 24

rs1
19 18

i=0
14 13 12

—
0

Suggested Assembly Language Syntax
rd
%y, regrd
rd
%ccr, regrd
rd
%asi, regrd
rd
%tick, regrd
rd
%pc, regrd
rd
%fprs, regrd
rd
asr_regrs1, regrd

Description:
These instructions read the specified state register into r[rd].
Note that RDY, RDCCR, RDASI, RDPC, RDTICK, RDFPRS, and RDASR are distinguished
only by the value in the rs1 field.
If rs1 ≥ 7, an ancillary state register is read. Values of rs1 in the range 7..14 are reserved for future
versions of the architecture; values in the range 16..31 are available for implementations to use
(impl. dep. #8). A RDASR instruction with rs1 = 15, rd = 0, and i = 0 is defined to be an STBAR
instruction (see A.51). An RDASR instruction with rs1 = 15, rd = 0, and i = 1 is defined to be a

MEMBAR instruction (see A.32). RDASR with rs1 = 15 and rd≠0 is reserved for future versions
of the architecture; it causes an illegal_instruction exception.
RDTICK causes a privileged_action exception if PSTATE.PRIV = 0 and TICK.NPT = 1.
For RDPC, the high-order 32-bits of the PC value stored in r[rd] are implementation-dependent
when PSTATE.AM = 1 (impl. dep. #125).
RDFPRS waits for all pending FPops and loads of floating-point registers to complete before
reading the FPRS register.
IMPL. DEP. #47: RDASR instructions with rd in the range 16..31 are available for implementation-dependent uses (impl. dep. #8). For a RDASR instruction with rs1 in the range 16 .. 31, the following are implementation-dependent: the interpretation of bits 13:0 and 29:25 in the instruction, whether the instruction is
privileged (impl. dep. #9), and whether the instruction causes an illegal_instruction exception.

See I.1.1, “Read/Write Ancillary State Registers (ASRs),” for a discussion of extending the
SPARC-V9 instruction set using read/write ASR instructions.
Implementation Note:
Ancillary state registers may include (for example) timer, counter, diagnostic, self-test, and trap-control registers. See Implementation Characteristics of Current SPARC-V9-based Products, Revision 9.x, a
document available from SPARC International, for information on implemented ancillary state registers.
Compatibility Note:
The SPARC-V8 RDPSR, RDWIM, and RDTBR instructions do not exist in SPARC-V9 since the PSR,
WIM, and TBR registers do not exist in SPARC-V9.

Exceptions:
privileged_opcode (RDASR only; implementation-dependent (impl. dep. #47))
illegal_instruction (RDASR with rs1 = 1 or 7..14; RDASR with rs1 = 15 and rd≠0; RDASR

with rs1 = 16..31 and the implementation does not define the instruction as an extension; implementation-dependent (impl. dep. #47))
privileged_action (RDTICK only)

A.45 RETURN
Opcode
RETURN

op3
11 1001

Operation
RETURN

Format (3):
10

—

op3

rs1

i=0

—

op3

rs1

i=1

31 30 29

25 24

19 18

14 13 12

—

rs2

simm13
5

Suggested Assembly Language Syntax
return
address

Description:
The RETURN instruction causes a delayed transfer of control to the target address and has the
window semantics of a RESTORE instruction; that is, it restores the register window prior to the
last SAVE instruction. The target address is “r[rs1] + r[rs2]” if i = 0, or
“r[rs1] + sign_ext(simm13)” if i = 1. Registers r[rs1] and r[rs2] come from the old window.
The RETURN instruction may cause an exception. It may cause a window_fill exception as part of
its RESTORE semantics or it may cause a mem_address_not_aligned exception if either of the two
low-order bits of the target address are nonzero.
Programming Note:
To reexecute the trapped instruction when returning from a user trap handler, use the RETURN instruction in
the delay slot of a JMPL instruction, for example:
jmpl
return

%l6,%g0
%l7

! Trapped PC supplied to user trap handler
! Trapped nPC supplied to user trap handler

Programming Note:
A routine that uses a register window may be structured either as
save
. . .
ret
restore

%sp,-framesize, %sp
! Same as jmpl %i7 + 8, %g0
! Something useful like “restore %o2,%l2,%o0”

or as
save
. . .
return
nop

%sp,-framesize, %sp
%i7

Exceptions:
mem_address_not_aligned
fill_n_normal (n = 0..7)

! Could do some useful work in the caller’s
! window e.g. “or %o1, %o2,%o0”

fill_n_other (n = 0..7)

A.46 SAVE and RESTORE
Opcode
SAVE
RESTORE

op3
11 1100
11 1101

Operation
Save caller’s window
Restore caller’s window

Format (3):
10

op3

rs1

i=0

op3

rs1

i=1

31 30 29

25 24

19 18

14 13 12

—

rs2

simm13
5

Suggested Assembly Language Syntax
save
regrs1, reg_or_imm, regrd
restore

regrs1, reg_or_imm, regrd

Description (Effect on Nonprivileged State):
The SAVE instruction provides the routine executing it with a new register window. The out registers from the old window become the in registers of the new window. The contents of the out and
the local registers in the new window are zero or contain values from the executing process; that
is, the process sees a clean window.
The RESTORE instruction restores the register window saved by the last SAVE instruction executed by the current process. The in registers of the old window become the out registers of the
new window. The in and local registers in the new window contain the previous values.
Furthermore, if and only if a spill or fill trap is not generated, SAVE and RESTORE behave like
normal ADD instructions, except that the source operands r[rs1] and/or r[rs2] are read from the
old window (that is, the window addressed by the original CWP) and the sum is written into r[rd]
of the new window (that is, the window addressed by the new CWP).
Note that CWP arithmetic is performed modulo the number of implemented windows, NWINDOWS.
Programming Note:
Typically, if a SAVE (RESTORE) instruction traps, the spill (fill) trap handler returns to the trapped instruction to reexecute it. So, although the ADD operation is not performed the first time (when the instruction
traps), it is performed the second time the instruction executes. The same applies to changing the CWP.
Programming Note:
The SAVE instruction can be used to atomically allocate a new window in the register file and a new software stack frame in memory. See H.1.2, “Leaf-Procedure Optimization,” for details.
Programming Note:
There is a performance tradeoff to consider between using SAVE/RESTORE and saving and restoring
selected registers explicitly.

Description (effect on privileged state):
If the SAVE instruction does not trap, it increments the CWP (mod NWINDOWS) to provide a
new register window and updates the state of the register windows by decrementing CANSAVE
and incrementing CANRESTORE.
If the new register window is occupied (that is, CANSAVE = 0), a spill trap is generated. The trap
vector for the spill trap is based on the value of OTHERWIN and WSTATE. The spill trap handler
is invoked with the CWP set to point to the window to be spilled (that is, old CWP + 2).
If CANSAVE ≠ 0, the SAVE instruction checks whether the new window needs to be cleaned. It
causes a clean_window trap if the number of unused clean windows is zero, that is, (CLEANWIN –
CANRESTORE) = 0. The clean_window trap handler is invoked with the CWP set to point to the
window to be cleaned (that is, old CWP + 1).
If the RESTORE instruction does not trap, it decrements the CWP (mod NWINDOWS) to restore
the register window that was in use prior to the last SAVE instruction executed by the current process. It also updates the state of the register windows by decrementing CANRESTORE and incrementing CANSAVE.
If the register window to be restored has been spilled (CANRESTORE = 0), a fill trap is generated. The trap vector for the fill trap is based on the values of OTHERWIN and WSTATE, as
described in 7.5.2.1, “Trap Type for Spill/Fill Traps.” The fill trap handler is invoked with CWP
set to point to the window to be filled, that is, old CWP – 1.
Programming Note:
The vectoring of spill and fill traps can be controlled by setting the value of the OTHERWIN and WSTATE
registers appropriately. For details, see the unnumbered subsection titled “Splitting the Register Windows”
in H.2.3, “Client-Server Model.”
Programming Note:
The spill (fill) handler normally will end with a SAVED (RESTORED) instruction followed by a RETRY
instruction.

Exceptions:
clean_window (SAVE only)
fill_n_normal (RESTORE only, n =0..7)
fill_n_other (RESTORE only, n = 0..7)
spill_n_normal (SAVE only, n = 0..7)
spill_n_other (SAVE only, n = 0..7)

A.47 SAVED and RESTORED
Opcode
SAVEDP
RESTOREDP
—

op3
11 0001
11 0001
11 0001

fcn
0
1
2..31

Operation
Window has been Saved
Window has been Restored
Reserved

Format (3):
10
31 30 29

fcn

op3
25 24

—
19 18

Suggested Assembly Language Syntax
saved
restored

Description:
SAVED and RESTORED adjust the state of the register-windows control registers.
SAVED increments CANSAVE. If OTHERWIN = 0, it decrements CANRESTORE.
If OTHERWIN≠0, it decrements OTHERWIN.
RESTORED increments CANRESTORE. If CLEANWIN < (NWINDOWS-1), RESTORED
increments CLEANWIN. If OTHERWIN = 0, it decrements CANSAVE.
If OTHERWIN ≠ 0, it decrements OTHERWIN.
Programming Note:
The spill (fill) handlers use the SAVED (RESTORED) instruction to indicate that a window has been spilled
(filled) successfully. See H.2.2, “Example Code for Spill Handler,” for details.
Programming Note:
Normal privileged software would probably not do a SAVED or RESTORED from trap level zero (TL = 0).
However, it is not illegal to do so, and does not cause a trap.
Programming Note:
Executing a SAVED (RESTORED) instruction outside of a window spill (fill) trap handler is likely to create
an inconsistent window state. Hardware will not signal an exception, however, since maintaining a consistent window state is the responsibility of privileged software.

Exceptions:
privileged_opcode
illegal_instruction (fcn=2..31)

A.48 SETHI
Opcode
SETHI

op
00

op2
100

Operation
Set High 22 Bits of Low Word

Format (2):
00

31 30 29

100
25 24

imm22

22 21

Suggested Assembly Language Syntax
sethi
const22, regrd
sethi
%hi (value), regrd

Description:
SETHI zeroes the least significant 10 bits and the most significant 32 bits of r[rd], and replaces
bits 31 through 10 of r[rd] with the value from its imm22 field.
SETHI does not affect the condition codes.
A SETHI instruction with rd = 0 and imm22 = 0 is defined to be a NOP instruction, which is
defined in A.40.
Programming Note:
The most common form of 64-bit constant generation is creating stack offsets whose magnitude is less than
232. The code below can be used to create the constant 0000 0000 ABCD 123416:
sethi
or

%hi(0xabcd1234),%o0
%o0, 0x234, %o0

The following code shows how to create a negative constant. Note that the immediate field of the xor
instruction is sign extended and can be used to get 1s in all of the upper 32 bits. For example, to set the negative constant FFFF FFFF ABCD 123416:
sethi
xor

Exceptions:
(none)

%hi(0x5432edcb),%o0! note 0x5432EDCB, not 0xABCD1234
%o0, 0x1e34, %o0
! part of imm. overlaps upper bits

A.49 Shift
Opcode
SLL
SRL
SRA
SLLX
SRLX
SRAX

op3
10 0101
10 0110
10 0111
10 0101
10 0110
10 0111

x
0
0
0
1
1
1

Operation
Shift Left Logical - 32 Bits
Shift Right Logical - 32 Bits
Shift Right Arithmetic - 32 Bits
Shift Left Logical - 64 Bits
Shift Right Logical - 64 Bits
Shift Right Arithmetic - 64 Bits

Format (3):
10

op3

rs1

i=0 x

—

rs2

op3

rs1

i=1 x=0

—

shcnt32

op3

rs1

i=1 x=1

31 30 29

25 24

19 18

14 13 12

—

shcnt64
6 5 4

Suggested Assembly Language Syntax
sll
regrs1, reg_or_shcnt, regrd
srl
regrs1, reg_or_shcnt, regrd
sra
regrs1, reg_or_shcnt, regrd
sllx
regrs1, reg_or_shcnt, regrd
srlx
regrs1, reg_or_shcnt, regrd
srax
regrs1, reg_or_shcnt, regrd

Description:
When i = 0 and x = 0, the shift count is the least significant five bits of r[rs2]. When i = 0 and
x = 1, the shift count is the least significant six bits of r[rs2]. When i = 1 and x = 0, the shift count
is the immediate value specified in bits 0 through 4 of the instruction. When i = 1 and x = 1, the
shift count is the immediate value specified in bits 0 through 5 of the instruction.
i
0
0
1
1

x
0
1
0
1

Shift count
bits 4 .. 0 of r[rs2]
bits 5 .. 0 of r[rs2]
bits 4..0 of instruction
bits 5..0 of instruction

SLL and SLLX shift all 64 bits of the value in r[rs1] left by the number of bits specified by the
shift count, replacing the vacated positions with zeroes, and write the shifted result to r[rd].

SRL shifts the low 32 bits of the value in r[rs1] right by the number of bits specified by the shift
count. Zeroes are shifted into bit 31. The upper 32 bits are set to zero, and the result is written to
r[rd].
SRLX shifts all 64 bits of the value in r[rs1] right by the number of bits specified by the shift
count. Zeroes are shifted into the vacated high-order bit positions, and the shifted result is written
to r[rd].
SRA shifts the low 32 bits of the value in r[rs1] right by the number of bits specified by the shift
count, and replaces the vacated positions with bit 31 of r[rs1]. The high order 32 bits of the result
are all set with bit 31 of r[rs1], and the result is written to r[rd].
SRAX shifts all 64 bits of the value in r[rs1] right by the number of bits specified by the shift
count, and replaces the vacated positions with bit 63 of r[rs1]. The shifted result is written to
r[rd].
No shift occurs when the shift count is zero, but the high-order bits are affected by the 32-bit
shifts as noted above.
These instructions do not modify the condition codes.
Programming Note:
“Arithmetic left shift by 1 (and calculate overflow)” can be effected with the ADDcc instruction.
Programming Note:
The instruction “sra rs1,0,rd” can be used to convert a 32-bit value to 64 bits, with sign extension into
the upper word. “srl rs1,0,rd” can be used to clear the upper 32 bits of r[rd].

Exceptions:
(none)

A.50 Software-Initiated Reset
Opcode
SIR

op3
11 0000

rd
15

Operation
Software-initiated reset

Format (3):
10
31 30 29

0 1111

op3
25 24

0 0000
19 18

i=1
14 13 12

simm13
0

Suggested Assembly Language Syntax
sir
simm13

Description:
SIR is used to generate a software-initiated reset (SIR). It may be executed in either privileged or
nonprivileged mode, with slightly different effect. As with other traps, a software-initiated reset
performs different actions when TL = MAXTL than it does when TL < MAXTL.
When executed in user mode, the action of SIR is conditional on the SIR_enable control flag.
IMPL. DEP. #116: The location of the SIR_enable control flag and the means of accessing the SIR_enable
control flag are implementation-dependent. In some implementations it may be permanently zero.

When SIR_enable is 0, SIR executes without effect (as a NOP) in user mode. When SIR is executed in privileged mode or in user mode with SIR_enable = 1, the processor performs a softwareinitiated reset. See 7.6.2.5, “Software-Initiated Reset (SIR) Traps,” for more information about
software-initiated resets.
Programming Note:
This instruction is never illegal. It is not a privileged instruction, even though its action in privileged mode is
different than in user mode.

Exceptions:
software_initiated_reset

A.51 Store Barrier
The STBAR instruction is deprecated; it is provided only for compatibility with
previous versions of the architecture. It should not be used in new SPARC-V9 software. It is recommended that the MEMBAR instruction be used in its place.

Opcode
STBARD

op3
10 1000

Operation
Store Barrier

Format (3):
10
31 30 29

op3
25 24

0 1111
19 18

0
14 13 12

—
0

Suggested Assembly Language Syntax
stbar

Description:
The store barrier instruction (STBAR) forces all store and atomic load-store operations issued by
a processor prior to the STBAR to complete their effects on memory before any store or atomic
load-store operations issued by that processor subsequent to the STBAR are executed by memory.
Note that the encoding of STBAR is identical to that of the RDASR instruction except that
rs1 = 15 and rd = 0, and is identical to that of the MEMBAR instruction except that bit 13 (i) = 0.
The coherence and atomicity of memory operations between processors and I/O DMA memory
accesses are implementation-dependent (impl. dep #120).
Compatibility Note:
STBAR is identical in function to a MEMBAR instruction with mmask = 816. STBAR is retained for compatibility with SPARC-V8.
Implementation Note:
For correctness, it is sufficient for a processor to stop issuing new store and atomic load-store operations
when an STBAR is encountered and resume after all stores have completed and are observed in memory by
all processors. More efficient implementations may take advantage of the fact that the processor is allowed
to issue store and load-store operations after the STBAR, as long as those operations are guaranteed not to
become visible before all the earlier stores and atomic load-stores have become visible to all processors.

Exceptions:
(none)

A.52 Store Floating-Point
The STFSR instruction is deprecated; it is provided only for compatibility with previous versions of the architecture. It should not be used in new SPARC-V9 software. It is recommended that the STXFSR instruction be used in its place.

†

Opcode
STF

op3
10 0100

rd
0..31

STDF

10 0111

†

Store Double Floating-Point Register

STQF
STFSRD
STXFSR
—

10 0110
10 0101
10 0101
10 0101

†

Store Quad Floating-Point Register
Store Floating-Point State Register Lower
Store Floating-Point State Register
Reserved

0
1
2..31

Operation
Store Floating-Point Register

Encoded floating-point register value, as described in 5.1.4.1

Format (3):
11

op3

rs1

i=0

op3

rs1

i=1

31 30 29

25 24

19 18

14 13 12

—

rs2

simm13
5

Suggested Assembly Language Syntax
fregrd, [address]
st
std
fregrd, [address]
stq
fregrd, [address]
st
%fsr, [address]
stx
%fsr, [address]

Description:
The store single floating-point instruction (STF) copies f [rd] into memory.
The store double floating-point instruction (STDF) copies a doubleword from a double floatingpoint register into a word-aligned doubleword in memory.
The store quad floating-point instruction (STQF) copies the contents of a quad floating-point register into a word-aligned quadword in memory.
The store floating-point state register lower instruction (STFSR) waits for any currently executing
FPop instructions to complete, and then writes the lower 32 bits of the FSR into memory.
The store floating-point state register instruction (STXFSR) waits for any currently executing
FPop instructions to complete, and then writes all 64 bits of the FSR into memory.

Compatibility Note:
SPARC-V9 needs two store-FSR instructions, since the SPARC-V8 STFSR instruction is defined to store
only 32 bits of the FSR into memory. STXFSR allows SPARC-V9 programs to store all 64 bits of the FSR.

STFSR and STXFSR zero FSR.ftt after writing the FSR to memory.
Implementation Note:
FSR.ftt should not be zeroed until it is known that the store will not cause a precise trap.

The effective address for these
“r[rs1] + sign_ext(simm13)” if i = 1.

instructions

“r[rs1] + r[rs2]”

i = 0,

STF, STFSR, STDF, and STQF cause a mem_address_not_aligned exception if the effective memory address is not word-aligned; STXFSR causes a mem_address_not_aligned exception if the
address is not doubleword-aligned. If the floating-point unit is not enabled for the source register
rd (per FPRS.FEF and PSTATE.PEF), or if the FPU is not present, a store floating-point instruction causes an fp_disabled exception.
IMPL. DEP. #110(1): STDF requires only word alignment in memory. If the effective address is wordaligned but not doubleword-aligned, it may cause an STDF_mem_address_not_aligned exception. In this
case the trap handler software shall emulate the STDF instruction and return.
IMPL. DEP. #112(1): STQF requires only word alignment in memory. If the effective address is wordaligned but not quadword-aligned, it may cause an STQF_mem_address_not_aligned exception. In this
case the trap handler software shall emulate the STQF instruction and return.
Programming Note:
In SPARC-V8, some compilers issued sets of single-precision stores when they could not determine that
double- or quadword operands were properly aligned. For SPARC-V9, since emulation of misaligned stores
is expected to be fast, it is recommended that compilers issue sets of single-precision stores only when they
can determine that double- or quadword operands are not properly aligned.

Exceptions:
async_data_error
fp_disabled
mem_address_not_aligned
STDF_mem_address_not_aligned (STDF only) (impl. dep. #110)
STQF_mem_address_not_aligned (STQF only) (impl. dep. #112)
data_access_exception
data_access_protection
data_access_MMU_miss
data_access_error
illegal_instruction (op3 = 2516 and rd = 2..31)
fp_exception_other (invalid_fp_register (STQF only))

A.53 Store Floating-Point into Alternate Space

†

Opcode
STFAPASI

op3
11 0100

rd
0..31

Operation
Store Floating-Point Register to Alternate Space

STDFAPASI

11 0111

†

Store Double Floating-Point Register to Alternate Space

STQFAPASI

11 0110

†

Store Quad Floating-Point Register to Alternate Space

Encoded floating-point register value, as described in 5.1.4.1

Format (3):
11

op3

rs1

i=0

op3

rs1

i=1

31 30 29

25 24

19 18

rs2

imm_asi

14 13 12

simm13
5

Suggested Assembly Language Syntax
fregrd, [regaddr] imm_asi
sta
sta
fregrd, [reg_plus_imm] %asi
stda
fregrd, [regaddr] imm_asi
stda
fregrd, [reg_plus_imm] %asi
stqa
fregrd, [regaddr] imm_asi
stqa
fregrd, [reg_plus_imm] %asi

Description:
The store single floating-point into alternate space instruction (STFA) copies f[rd] into memory.
The store double floating-point into alternate space instruction (STDFA) copies a doubleword
from a double floating-point register into a word-aligned doubleword in memory.
The store quad floating-point into alternate space instruction (STQFA) copies the contents of a
quad floating-point register into a word-aligned quadword in memory.
Store floating-point into alternate space instructions contain the address space identifier (ASI) to
be used for the load in the imm_asi field if i = 0, or in the ASI register if i = 1. The access is privileged if bit seven of the ASI is zero; otherwise, it is not privileged. The effective address for these
instructions is “r[rs1] + r[rs2]” if i = 0, or “r[rs1] + sign_ext(simm13)” if i = 1.
STFA, STDFA, and STQFA cause a mem_address_not_aligned exception if the effective memory
address is not word-aligned. If the floating-point unit is not enabled for the source register rd (per
FPRS.FEF and PSTATE.PEF), or if the FPU is not present, store floating-point into alternate
space instructions cause an fp_disabled exception.
STFA, STDFA, and STQFA cause a privileged_action exception if PSTATE.PRIV = 0 and bit 7 of
the ASI is zero.

IMPL. DEP. #110(2): STDFA requires only word alignment in memory. If the effective address is wordaligned but not doubleword-aligned, it may cause an STDF_mem_address_not_aligned exception. In this
case the trap handler software shall emulate the STDFA instruction and return.
IMPL. DEP. #112(2): STQFA requires only word alignment in memory. If the effective address is wordaligned but not quadword-aligned, it may cause an STQF_mem_address_not_aligned exception. In this
case the trap handler software shall emulate the STQFA instruction and return.
Programming Note:
In SPARC-V8, some compilers issued sets of single-precision stores when they could not determine that
double- or quadword operands were properly aligned. For SPARC-V9, since emulation of misaligned stores
is expected to be fast, it is recommended that compilers issue sets of single-precision stores only when they
can determine that double- or quadword operands are not properly aligned.

Exceptions:
async_data_error
fp_disabled
mem_address_not_aligned
STDF_mem_address_not_aligned (STDFA only) (impl. dep. #110)
STQF_mem_address_not_aligned (STQFA only) (impl. dep. #112)
privileged_action
data_access_exception
data_access_protection
data_access_MMU_miss
data_access_error
fp_exception_other (invalid_fp_register (STQFA only))

A.54 Store Integer
The STD instruction isdeprecated; it is provided only for compatibility with previous versions of the architecture. It should not be used in new SPARC-V9 software.
It is recommended that the STX instruction be used in its place.

Opcode
STB
STH
STW
STX
STDD

op3
00 0101
00 0110
00 0100
00 1110

Operation
Store Byte
Store Halfword
Store Word
Store Extended Word

00 0111

Store Doubleword

Format (3):
11

op3

rs1

i=0

op3

rs1

i=1

31 30 29

25 24

stb
sth
stw
stx
std

19 18

—

rs2

simm13

14 13 12

Suggested Assembly Language Syntax
(synonyms: stub, stsb)
regrd, [address]
regrd, [address]
(synonyms: stuh, stsh)
regrd, [address]
(synonyms: st, stuw, stsw)
regrd, [address]
regrd, [address]

Description:
The store integer instructions (except store doubleword) copy the whole extended (64-bit) integer,
the less-significant word, the least significant halfword, or the least significant byte of r[rd] into
memory.
The store doubleword integer instruction (STD) copies two words from an r register pair into
memory. The least significant 32 bits of the even-numbered r register are written into memory at
the effective address, and the least significant 32 bits of the following odd-numbered r register are
written into memory at the “effective address + 4.” The least significant bit of the rd field of a
store doubleword instruction is unused and should always be set to zero by software. An attempt
to execute a store doubleword instruction that refers to a misaligned (odd-numbered) rd causes an
illegal_instruction exception.
IMPL. DEP. #108(1): IT is implementation-dependent whether STD is implemented in hardware. if not, an
attempt to execute it will cause an unimplemented_STD exception.

The effective address for these instructions is “r[rs1] + r[rs2]” if i = 0, or
“r[rs1] + sign_ext(simm13)” if i = 1.
A successful store (notably, store extended and store doubleword) instruction operates atomically.
STH causes a mem_address_not_aligned exception if the effective address is not halfword-aligned.
STW causes a mem_address_not_aligned exception if the effective address is not word-aligned.
STX and STD causes a mem_address_not_aligned exception if the effective address is not doubleword-aligned.
Programming Note:
STD is provided for compatibility with SPARC-V8. It may execute slowly on SPARC-V9 machines because
of data path and register-access difficulties. In some SPARC-V9 systems it may cause a trap to emulation
code; therefore, STD should be avoided.
If STD is emulated in software, STX should be used in order to preserve atomicity.
Compatibility Note:
The SPARC-V8 ST instruction has been renamed STW in SPARC-V9.

Exceptions:
async_data_error
unimplemented_STD (STD only) (impl. dep. #108)
illegal_instruction (STD with odd rd)
mem_address_not_aligned (all except STB)
data_access_exception
data_access_error
data_access_protection
data_access_MMU_miss

A.55 Store Integer into Alternate Space
The STDA instruction is deprecated; it is provided only for compatibility with previous versions of the architecture. It should not be used in new SPARC-V9 software. It is recommended that the STXA instruction be used in its place.

Opcode
STBAPASI
STHAPASI
STWAPASI
STXAPASI
STDAD, PASI

op3
01 0101
01 0110
01 0100
01 1110
01 0111

Operation
Store Byte into Alternate space
Store Halfword into Alternate space
Store Word into Alternate space
Store Extended Word into Alternate space
Store Doubleword into Alternate space

Format (3):
11

op3

rs1

i=0

op3

rs1

i=1

31 30 29

25 24

19 18

14 13 12

imm_asi

rs2

simm13
5

stba
stha
stwa
stxa
stda
stba
stha
stwa
stxa

Suggested Assembly Language Syntax
(synonyms: stuba, stsba)
regrd, [regaddr] imm_asi
regrd, [regaddr] imm_asi
(synonyms: stuha, stsha)
regrd, [regaddr] imm_asi
(synonyms: sta, stuwa, stswa)
regrd, [regaddr] imm_asi
regrd, [regaddr] imm_asi
regrd, [reg_plus_imm] %asi
(synonyms: stuba, stsba)
regrd, [reg_plus_imm] %asi
(synonyms: stuha, stsha)
regrd, [reg_plus_imm] %asi
(synonyms: sta, stuwa, stswa)
regrd, [reg_plus_imm] %asi

stda

regrd, [reg_plus_imm] %asi

Description:
The store integer into alternate space instructions (except store doubleword) copy the whole
extended (64-bit) integer, the less-significant word, the least-significant halfword, or the least-significant byte of r[rd] into memory.
The store doubleword integer instruction (STDA) copies two words from an r register pair into
memory. The least-significant 32 bits of the even-numbered r register are written into memory at
the effective address, and the least-significant 32 bits of the following odd-numbered r register are
written into memory at the “effective address + 4.” The least significant bit of the rd field of a
store doubleword instruction is unused and should always be set to zero by software. An attempt

to execute a store doubleword instruction that refers to a misaligned (odd-numbered) rd causes an
illegal_instruction exception.
IMPL. DEP. #108(2): It is implementation-dependent whether STDA is implemented in hardware. If not, an
attempt to execute it will cause an unimplemented_STD exception.

Store integer to alternate space instructions contain the address space identifier (ASI) to be used
for the store in the imm_asi field if i = 0, or in the ASI register if i = 1. The access is privileged if
bit seven of the ASI is zero; otherwise, it is not privileged. The effective address for these instructions is “r[rs1] + r[rs2]” if i = 0, or “r[rs1]+sign_ext(simm13)” if i = 1.
A successful store (notably, store extended and store doubleword) instruction operates atomically.
STHA causes a mem_address_not_aligned exception if the effective address is not halfwordaligned. STWA causes a mem_address_not_aligned exception if the effective address is not wordaligned. STXA and STDA cause a mem_address_not_aligned exception if the effective address is
not doubleword-aligned.
A store integer into alternate space instruction causes a privileged_action exception if
PSTATE.PRIV = 0 and bit 7 of the ASI is zero.
Programming Note:
STDA is provided for compatibility with SPARC-V8. It may execute slowly on SPARC-V9 machines
because of data path and register-access difficulties. In some SPARC-V9 systems it may cause a trap to emulation code; therefore, STDA should be avoided.
If STDA is emulated in software, STXA should be used in order to preserve atomicity.
Compatibility Note:
The SPARC-V8 STA instruction is renamed STWA in SPARC-V9.

Exceptions:
async_data_error
unimplemented_STD (STDA only) (impl. dep. #108)
illegal_instruction (STDA with odd rd)
privileged_action
mem_address_not_aligned (all except STBA)
data_access_exception
data_access_error
data_access_protection
data_access_MMU_miss

A.56 Subtract
Opcode
SUB
SUBcc
SUBC
SUBCcc

op3
00 0100
01 0100
00 1100
01 1100

Operation
Subtract
Subtract and modify cc’s
Subtract with Carry
Subtract with Carry and modify cc’s

Format (3):
10

op3

rs1

i=0

op3

rs1

i=1

31 30 29

25 24

19 18

—

rs2

simm13

14 13 12

Suggested Assembly Language Syntax
sub
regrs1, reg_or_imm, regrd
subcc
regrs1, reg_or_imm, regrd
subc
regrs1, reg_or_imm, regrd
subccc
regrs1, reg_or_imm, regrd

Description:
These instructions compute “r[rs1] – r[rs2]” if i = 0, or “r[rs1] – sign_ext(simm13)” if i = 1, and
write the difference into r[rd].
SUBC and SUBCcc (“SUBtract with carry”) also subtract the CCR register’s 32-bit carry (icc.c)
bit; that is, they compute “r[rs1] – r[rs2] – icc.c” or “r[rs1] – sign_ext(simm13) –icc.c,” and write
the difference into r[rd].
SUBcc and SUBCcc modify the integer condition codes (CCR.icc and CCR.xcc). 32-bit overflow
(CCR.icc.v) occurs on subtraction if bit 31 (the sign) of the operands differ and bit 31 (the sign) of
the difference differs from r[rs1]<31>. 64-bit overflow (CCR.xcc.v) occurs on subtraction if bit 63
(the sign) of the operands differ and bit 63 (the sign) of the difference differs from r[rs1]<63>.
Programming Note:
A SUBcc with rd = 0 can be used to effect a signed or unsigned integer comparison. See the CMP synthetic
instruction in Appendix G.
Programming Note:
SUBC and SUBCcc read the 32-bit condition codes’ carry bit (CCR.icc.c), not the 64-bit condition codes’
carry bit (CCR.xcc.c).
Compatibility Note:
SUBC and SUBCcc were named SUBX and SUBXcc, respectively, in SPARC-V8.

Exceptions:
(none)

A.57 Swap Register with Memory
The SWAP instruction is deprecated; it is provided only for compatibility with previous versions of the architecture. It should not be used in new SPARC-V9 software. It is recommended that the CASA (or CASXA) instruction be used in its
place.

Opcode
SWAPD

op3
00 1111

Operation
SWAP register with memory

Format (3):
11

op3

rs1

i=0

op3

rs1

i=1

31 30 29

25 24

19 18

14 13 12

rs2

—

simm13
5

Suggested Assembly Language Syntax
swap

[address], regrd

Description:
SWAP exchanges the lower 32 bits of r[rd] with the contents of the word at the addressed memory
location. The upper 32 bits of r[rd] are set to zero. The operation is performed atomically, that is,
without allowing intervening interrupts or deferred traps. In a multiprocessor system, two or more
processors executing CASA, CASXA, SWAP, SWAPA, LDSTUB, or LDSTUBA instructions
addressing any or all of the same doubleword simultaneously are guaranteed to execute them in an
undefined but serial order.
The effective address for these instructions is “r[rs1] + r[rs2]” if i = 0, or
“r[rs1] + sign_ext(simm13)” if i = 1. This instruction causes a mem_address_not_aligned exception
if the effective address is not word-aligned.
The coherence and atomicity of memory operations between processors and I/O DMA memory
accesses are implementation-dependent (impl. dep #120).
Implementation Note:
See Implementation Characteristics of Current SPARC-V9-based Products, Revision 9.x, a document available from SPARC International, for information on the presence of hardware support for these
instructions in the various SPARC-V9 implementations.

Exceptions:
mem_address_not_aligned
data_access_exception
data_access_error
data_access_protection
data_access_MMU_miss

async_data_error

A.58 Swap Register with Alternate Space Memory
The SWAPA instruction is deprecated; it is provided only for compatibility with
previous versions of the architecture. It should not be used in new SPARC-V9 software. It is recommended that the CASXA instruction be used in its place.

Opcode
SWAPAD, PASI

op3
01 1111

Operation
SWAP register with Alternate space memory

Format (3):
11

op3

rs1

i=0

op3

rs1

i=1

31 30 29

25 24

19 18

rs2

imm_asi

14 13 12

simm13
5

Suggested Assembly Language Syntax
swapa
[regaddr] imm_asi, regrd
swapa
[reg_plus_imm] %asi, regrd

Description:
SWAPA exchanges the lower 32 bits of r[rd] with the contents of the word at the addressed memory location. The upper 32 bits of r[rd] are set to zero. The operation is performed atomically, that
is, without allowing intervening interrupts or deferred traps. In a multiprocessor system, two or
more processors executing CASA, CASXA, SWAP, SWAPA, LDSTUB, or LDSTUBA instructions addressing any or all of the same doubleword simultaneously are guaranteed to execute them
in an undefined, but serial order.
The SWAPA instruction contains the address space identifier (ASI) to be used for the load in the
imm_asi field if i = 0, or in the ASI register if i = 1. The access is privileged if bit seven of the ASI
is zero; otherwise, it is not privileged. The effective address for this instruction is “r[rs1] + r[rs2]”
if i = 0, or “r[rs1] + sign_ext(simm13)” if i = 1.
This instruction causes a mem_address_not_aligned exception if the effective address is not wordaligned. It causes a privileged_action exception if PSTATE.PRIV = 0 and bit 7 of the ASI is zero.
The coherence and atomicity of memory operations between processors and I/O DMA memory
accesses are implementation-dependent (impl. dep #120).
Implementation Note:
See Implementation Characteristics of Current SPARC-V9-based Products, Revision 9.x, a document available from SPARC International, for information on the presence of hardware support for this
instruction in the various SPARC-V9 implementations.

Exceptions:
mem_address_not_aligned
privileged_action
data_access_exception
data_access_error
data_access_protection
data_access_MMU_miss
async_data_error

A.59 Tagged Add
The TADDccTV instruction is deprecated; it is provided only for compatibility
with previous versions of the architecture. It should not be used in new SPARC-V9
software. It is recommended that TADDcc followed by BPVS be used in its place
(with instructions to save the pre-TADDcc integer condition codes, if necessary).

Opcode
TADDcc
TADDccTVD

op3
10 0000
10 0010

Operation
Tagged Add and modify cc’s
Tagged Add and modify cc’s, or Trap on Overflow

Format (3):
10

op3

rs1

i=0

op3

rs1

i=1

31 30 29

25 24

19 18

14 13 12

—

rs2

simm13
5

Suggested Assembly Language Syntax
taddcc
regrs1, reg_or_imm, regrd
taddcctv
regrs1, reg_or_imm, regrd

Description:
These instructions compute a sum that is “r[rs1] + r[rs2]” if i = 0, or “r[rs1] + sign_ext(simm13)”
if i = 1.
TADDcc modifies the integer condition codes (icc and xcc), and TADDccTV does so also, if it
does not trap.
A tag_overflow exception occurs if bit 1 or bit 0 of either operand is nonzero, or if the addition generates 32-bit arithmetic overflow (i.e., both operands have the same value in bit 31, and bit 31 of
the sum is different).
If TADDccTV causes a tag overflow, a tag_overflow exception is generated, and r[rd] and the integer condition codes remain unchanged. If a TADDccTV does not cause a tag overflow, the sum is
written into r[rd], and the integer condition codes are updated. CCR.icc.v is set to 0 to indicate no
32-bit overflow. If a TADDcc causes a tag overflow, the 32-bit overflow bit (CCR.icc.v) is set to 1;
if it does not cause a tag overflow, CCR.icc.v is cleared.
In either case, the remaining integer condition codes (both the other CCR.icc bits and all the
CCR.xcc bits) are also updated as they would be for a normal ADD instruction. In particular, the
setting of the CCR.xcc.v bit is not determined by the tag overflow condition (tag overflow is used
only to set the 32-bit overflow bit). CCR.xcc.v is set only based on the normal 64-bit arithemetic
overflow condition, like a normal 64-bit add.
Compatibility Note:
TADDccTV traps based on the 32-bit overflow condition, just as in SPARC-V8. Although the tagged-add
instructions set the 64-bit condition codes CCR.xcc, there is no form of the instruction that traps the 64-bit
overflow condition.

Exceptions:
tag_overflow (TADDccTV only)

A.60 Tagged Subtract
The TSUBccTV instruction is deprecated; it is provided only for compatibility with
previous versions of the architecture. It should not be used in new SPARC-V9 software. It is recommended that TSUBcc followed by BPVS be used in its place (with
instructions to save the pre-TSUBcc integer condition codes, if necessary).

Opcode
TSUBcc
TSUBccTVD

op3
10 0001
10 0011

Operation
Tagged Subtract and modify cc’s
Tagged Subtract and modify cc’s, or Trap on Overflow

Format (3):
10

op3

rs1

i=0

op3

rs1

i=1

31 30 29

25 24

19 18

14 13 12

—

rs2

simm13
5

Suggested Assembly Language Syntax
tsubcc
regrs1, reg_or_imm, regrd
tsubcctv regrs1, reg_or_imm, regrd

Description:
These instructions compute “r[rs1] – r[rs2]” if i = 0, or “r[rs1] – sign_ext(simm13)” if i = 1.
TSUBcc modifies the integer condition codes (icc and xcc); TSUBccTV also modifies the integer
condition codes, if it does not trap.
A tag overflow occurs if bit 1 or bit 0 of either operand is nonzero, or if the subtraction generates
32-bit arithmetic overflow; that is, the operands have different values in bit 31 (the 32-bit sign bit)
and the sign of the 32-bit difference in bit 31 differs from bit 31 of r[rs1].
If TSUBccTV causes a tag overflow, a tag_overflow exception is generated and r[rd] and the integer condition codes remain unchanged. If a TSUBccTV does not cause a tag overflow condition,
the difference is written into r[rd], and the integer condition codes are updated. CCR.icc.v is set to
0 to indicate no 32-bit overflow. If a TSUBcc causes a tag overflow, the 32-bit overflow bit
(CCR.icc.v) is set to 1; if it does not cause a tag overflow, CCR.icc.v is cleared.
In either case, the remaining integer condition codes (both the other CCR.icc bits and all the
CCR.xcc bits) are also updated as they would be for a normal subtract instruction. In particular,
the setting of the CCR.xcc.v bit is not determined by the tag overflow condition (tag overflow is
used only to set the 32-bit overflow bit). CCR.xcc.v is set only based on the normal 64-bit
arithemetic overflow condition, like a normal 64-bit subtract.

Compatibility Note:
TSUBccTV traps are based on the 32-bit overflow condition, just as in SPARC-V8. Although the taggedsubtract instructions set the 64-bit condition codes CCR.xcc, there is no form of the instruction that traps on
64-bit overflow.

Exceptions:
tag_overflow (TSUBccTV only)

A.61 Trap on Integer Condition Codes (Tcc)
Opcode
TA
TN
TNE
TE
TG
TLE
TGE
TL
TGU
TLEU
TCC
TCS
TPOS
TNEG
TVC
TVS

op3
11 1010
11 1010
11 1010
11 1010
11 1010
11 1010
11 1010
11 1010
11 1010
11 1010
11 1010
11 1010
11 1010
11 1010
11 1010
11 1010

cond
1000
0000
1001
0001
1010
0010
1011
0011
1100
0100
1101
0101
1110
0110
1111
0111

Operation
icc test
Trap Always
1
Trap Never
0
Trap on Not Equal
not Z
Trap on Equal
Z
Trap on Greater
not (Z or (N xor V))
Trap on Less or Equal
Z (N xor V)
Trap on Greater or Equal
not (N xor V)
Trap on Less
N xor V
Trap on Greater Unsigned
not (C or Z)
Trap on Less or Equal Unsigned
(C or Z)
Trap on Carry Clear (Greater than or Equal, Unsigned)
not C
Trap on Carry Set (Less Than, Unsigned)
C
Trap on Positive or zero
not N
Trap on Negative
N
Trap on Overflow Clear
not V
Trap on Overflow Set
V

Format (4):
10

—

cond

op3

rs1

i=0 cc1 cc0

—

cond

op3

rs1

i=1 cc1 cc0

31 30 29 28

25 24

19 18

cc1

cc0
00
01
10
11

14 13 12 11 10

Condition codes
icc
—
xcc
—

—

rs2

—

sw_trap_#
7

6 5

ta
tn
tne
te
tg
tle
tge
tl
tgu
tleu
tcc
tcs
tpos
tneg
tvc
tvs

Suggested Assembly Language Syntax
i_or_x_cc, software_trap_number
i_or_x_cc, software_trap_number
i_or_x_cc, software_trap_number
(synonym: tnz)
i_or_x_cc, software_trap_number
(synonym: tz)
i_or_x_cc, software_trap_number
i_or_x_cc, software_trap_number
i_or_x_cc, software_trap_number
i_or_x_cc, software_trap_number
i_or_x_cc, software_trap_number
i_or_x_cc, software_trap_number
i_or_x_cc, software_trap_number
(synonym: tgeu)
i_or_x_cc, software_trap_number
(synonym: tlu)
i_or_x_cc, software_trap_number
i_or_x_cc, software_trap_number
i_or_x_cc, software_trap_number
i_or_x_cc, software_trap_number

Description:
The Tcc instruction evaluates the selected integer condition codes (icc or xcc) according to the
cond field of the instruction, producing either a TRUE or FALSE result. If TRUE and no higherpriority exceptions or interrupt requests are pending, then a trap_instruction exception is generated.
If FALSE, a trap_instruction exception does not occur and the instruction behaves like a NOP.
The software trap number is specified by the least significant seven bits of “r[rs1] + r[rs2]” if
i = 0, or the least significant seven bits of “r[rs1] + sw_trap_#” if i = 1.
When i = 1, bits 7 through 10 are reserved and should be supplied as zeros by software. When
i = 0, bits 5 through 10 are reserved, and the most significant 57 bits of “r[rs1] + r[rs2]” are
unused, and both should be supplied as zeros by software.
Description (Effect on Privileged State):
If a trap_instruction traps, 256 plus the software trap number is written into TT[TL]. Then the trap
is taken, and the processor performs the normal trap entry procedure, as described in Chapter 7,
“Traps.”
Programming Note:
Tcc can be used to implement breakpointing, tracing, and calls to supervisor software. It can also be used for
run-time checks, such as out-of-range array indexes, integer overflow, etc.
Compatibility Note:
Tcc is upward compatible with the SPARC-V8 Ticc instruction, with one qualification: a Ticc with i = 1 and
simm13 < 0 may execute differently on a SPARC-V9 processor. Use of the i = 1 form of Ticc is believed to
be rare in SPARC-V8 software, and simm13 < 0 is probably not used at all, so it is believed that, in practice,
full software compatibillity will be achieved.

Exceptions:
trap_instruction

illegal_instruction (cc1

cc0 = 012 or 112)

A.62 Write Privileged Register
Opcode
WRPRP

op3
11 0010

Operation
Write Privileged Register

Format (3):
10

op3

rs1

i=0

op3

rs1

i=1

31 30 29

25 24

19 18

rd
0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15..31

14 13 12

Privileged register
TPC
TNPC
TSTATE
TT
TICK
TBA
PSTATE
TL
PIL
CWP
CANSAVE
CANRESTORE
CLEANWIN
OTHERWIN
WSTATE
Reserved

—

rs2

simm13
5

Suggested Assembly Language Syntax
wrpr
regrs1, reg_or_imm, %tpc
wrpr
regrs1, reg_or_imm, %tnpc
wrpr
regrs1, reg_or_imm, %tstate
wrpr
regrs1, reg_or_imm, %tt
wrpr
regrs1, reg_or_imm, %tick
wrpr
regrs1, reg_or_imm, %tba
wrpr
regrs1, reg_or_imm, %pstate
wrpr
regrs1, reg_or_imm, %tl
wrpr
regrs1, reg_or_imm, %pil
wrpr
regrs1, reg_or_imm, %cwp
wrpr
regrs1, reg_or_imm, %cansave
wrpr
regrs1, reg_or_imm, %canrestore
wrpr
regrs1, reg_or_imm, %cleanwin
wrpr
regrs1, reg_or_imm, %otherwin
wrpr
regrs1, reg_or_imm, %wstate

Description:
This instruction stores the value “r[rs1] xor r[rs2]” if i = 0, or “r[rs1] xor sign_ext(simm13)” if
i = 1 to the writable fields of the specified privileged state register. Note the exclusive-or operation.
The rd field in the instruction determines the privileged register that is written. There are at least
four copies of the TPC, TNPC, TT, and TSTATE registers, one for each trap level. A write to one
of these registers sets the register indexed by the current value in the trap level register (TL). A
write to TPC, TNPC, TT, or TSTATE when the trap level is zero (TL = 0) causes an
illegal_instruction exception.
A WRPR of TL does not cause a trap or return from trap; it does not alter any other machine state.
Programming Note:
A WRPR of TL can be used to read the values of TPC, TNPC, and TSTATE for any trap level, however, care
must be taken that traps do not occur while the TL register is modified.

The WRPR instruction is a nondelayed-write instruction. The instruction immediately following
the WRPR observes any changes made to processor state made by the WRPR.
WRPR instructions with rd in the range 15..31 are reserved for future versions of the architecture;
executing a WRPR instruction with rd in that range causes an illegal_instruction exception.
Programming Note:
On an implementation that provides a floating-point queue, supervisor software should be aware of the state
of the FQ before disabling the floating-point unit (changing PSTATE.PEF from 1 to 0 with a WRPR instruction) (impl. dep. #24). Typically, supervisor software ensures that the FQ is empty (FSR.qne = 0) before disabling the floating-point unit.

Exceptions:
privileged_opcode
illegal_instruction ((rd = 15..31) or ((rd ≤ 3) and (TL = 0)))

A.63 Write State Register
The WRY instruction is deprecated; it is provided only for compatibility with previous versions of the architecture. It should not be used in new SPARC-V9 software.
It is recommended that all instructions which reference the Y register be avoided.

Opcode
WRYD
—
WRCCR
WRASI
WRASRPASR
WRFPRS
WRASRPASR
See text
WRASRPASR

op3
11 0000
11 0000
11 0000
11 0000
11 0000
11 0000
11 0000
11 0000
11 0000

rd
0
1
2
3
4, 5
6
7..14
15
16 .. 31

Operation
Write Y register
Reserved
Write Condition Codes Register
Write ASI register
Write Ancillary State Register (reserved)
Write Floating-Point Registers Status register
Write Ancillary State Register (reserved)
See text
Implementation-dependent (impl. dep. #48)

Format (3):
10

op3

rs1

i=0

op3

rs1

i=1

31 30 29

25 24

19 18

wr
wr
wr
wr
wr
†

—

rs2

simm13

14 13 12

Suggested Assembly Language Syntax
regrs1, reg_or_imm, %y
regrs1, reg_or_imm, %ccr
regrs1, reg_or_imm, %asi
regrs1, reg_or_imm, %fprs
regrs1, reg_or_imm, asr_regrd †

Syntax for WRASR with rd=16..31 may vary (impl. dep. #48)

Description:
WRY, WRCCR, WRFPRS, and WRASI stores the value “r[rs1] xor r[rs2]” if i = 0, or “r[rs1] xor
sign_ext(simm13)” if i = 1, to the writable fields of the specified state register. Note the exclusiveor operation.
Note that WRY, WRCCR, WRASI, WRFPRS, and WRASR are distinguished only by the rd field.

WRASR writes a value to the ancillary state register (ASR) indicated by rd. The operation performed to generate the value written may be rd-dependent or implementation-dependent (see
below). A WRASR instruction is indicated by op = 216, rd = 4, 5, or ≥ 7 and op3 = 3016.
An instruction with op = 216, op3 = 3016, rd = 15, rs1 = 0, and i = 1 is defined as a SIR instruction.
See A.50, “Software-Initiated Reset.” When op = 216, op3 = 3016, and rd = 15, if either rs1≠0 or
i≠1, then an illegal_instruction exception shall be generated.
IMPL. DEP. #48: WRASR instructions with rd in the range 16..31 are available for implementation-dependent uses (impl. dep. #8). For a WRASR instruction with rd in the range 16..31, the following are implementation-dependent: the interpretation of bits 18:0 in the instruction, the operation(s) performed (for
example, XOR) to generate the value written to the ASR, whether the instruction is privileged (impl. dep.
#9), and whether the instruction causes an illegal_instruction exception.

See I.1.1, “Read/Write Ancillary State Registers (ASRs),” for a discussion of extending the
SPARC-V9 instruction set using read/write ASR instructions.
The WRY, WRCCR, WRFPRS, and WRASI instructions are not delayed-write instructions. The
instruction immediately following a WRY, WRCCR, WRFPRS, or WRASI observes the new
value of the Y, CCR, FPRS, or ASI register.
WRFPRS waits for any pending floating-point operations to complete before writing the FPRS
register.
Implementation Note:
Ancillary state registers may include (for example) timer, counter, diagnostic, self-test, and trap-control registers. See Implementation Characteristics of Current SPARC-V9-based Products, Revision 9.x, a
document available from SPARC International, for information on ancillary state registers provided by specific implementations.
Compatibility Note:
The SPARC-V8 WRIER, WRPSR, WRWIM, and WRTBR instructions do not exist in SPARC-V9, since the
IER, PSR, TBR, and WIM registers do not exist in SPARC-V9.

Exceptions:
privileged_opcode (WRASR only; implementation-dependent (impl. dep. #48))
illegal_instruction (WRASR with rd = 16..31 and the implementation does not define the

instruction as an extension; implementation-dependent (impl. dep. #48), or WRASR
with rd equal to 1, 4, 5, or in the range 7..14), WRASR with rd equal to 15 and rs1≠0
or i≠1

B IEEE Std 754-1985 Requirements for SPARC-V9
The IEEE Std 754-1985 floating-point standard contains a number of implementation-dependencies. This appendix specifies choices for these implementation-dependencies, to ensure that
SPARC-V9 implementations are as consistent as possible.

B.1 Traps Inhibit Results
As described in 5.1.7, “Floating-Point State Register (FSR),” and elsewhere, when a floatingpoint trap occurs:
— The destination floating-point register(s) (the f registers) are unchanged.
— The floating-point condition codes (fcc0, fcc1, fcc2, and fcc3) are unchanged.
— The FSR.aexc (accrued exceptions) field is unchanged.
— The FSR.cexc (current exceptions) field is unchanged except for IEEE_754_exceptions; in
that case, cexc contains a bit set to “1” corresponding to the exception that caused the trap.
Only one bit shall be set in cexc.
Instructions causing an fp_exception_other trap due to unfinished or unimplemented FPops execute
as if by hardware; that is, a trap is undetectable by user software, except that timing may be
affected. A user-mode trap handler invoked for an IEEE_754_exception, whether as a direct result
of a hardware fp_exception_ieee_754 trap or as an indirect result of supervisor handling of an
unfinished_FPop or unimplemented_FPop, can rely on the following:
— The address of the instruction that caused the exception will be available to it.
— The destination floating-point register(s) are unchanged from their state prior to that
instruction’s execution.
— The floating-point condition codes (fcc0, fcc1, fcc2, and fcc3) are unchanged.
— The FSR aexc field is unchanged.
— The FSR cexc field contains exactly one bit set to 1, corresponding to the exception that
caused the trap.
— The FSR ftt, qne, and reserved fields are zero.
Supervisor software is responsible for enforcing these requirements if the hardware trap mechanism does not.

B.2 NaN Operand and Result Definitions
An untrapped floating-point result can be in a format that is either the same as, or different from,
the format of the source operands. These two cases are described separately below.
B.2.1 Untrapped Result in Different Format from Operands
F[sdq]TO[sdq] with a quiet NaN operand:
No exception caused; result is a quiet NaN. The operand is transformed as follows:
NaN transformation: The most significant bits of the operand fraction are copied to the
most significant bits of the result fraction. When converting to a narrower format, excess
low-order bits of the operand fraction are discarded. When converting to a wider format,
excess low-order bits of the result fraction are set to 0. The quiet bit (the most significant

bit of the result fraction) is always set to 1, so the NaN transformation always produces a
quiet NaN. The sign bit is copied from the operand to the result without modification.
F[sdq]TO[sdq] with a signaling NaN operand:
Invalid exception; result is the signaling NaN operand processed by the NaN transformation above to produce a quiet NaN.
FCMPE[sdq] with any NaN operand:
Invalid exception; the selected floating-point condition code is set to unordered.
FCMP[sdq] with any signaling NaN operand:
Invalid exception; the selected floating-point condition code is set to unordered.
FCMP[sdq] with any quiet NaN operand but no signaling NaN operand:
No exception; the selected floating-point condition code is set to unordered.
B.2.2 Untrapped Result in Same Format as Operands
No NaN operand:
For an invalid operation such as sqrt(–1.0) or 0.0 ÷ 0.0, the result is the quiet NaN with
sign = zero, exponent = all ones, and fraction = all ones. The sign is zero to distinguish
such results from storage initialized to all ones.
One operand, a quiet NaN:
No exception; result is the quiet NaN operand.
One operand, a signaling NaN:
Invalid exception; result is the signaling NaN with its quiet bit (most significant bit of fraction field) set to 1.
Two operands, both quiet NaNs:
No exception; result is the rs2 (second source) operand.
Two operands, both signaling NaNs:
Invalid exception; result is the rs2 operand with the quiet bit set to 1.
Two operands, only one a signaling NaN:
Invalid exception; result is the signaling NaN operand with the quiet bit set to 1.
Two operands, neither a signaling NaN, only one a quiet NaN:
No exception; result is the quiet NaN operand.

In table 27 NaNn means that the NaN is in rsn, Q means quiet, S signaling.
Table 27—Untrapped Floating-Point Results

rs1
operand

None
Number
QNaN1
SNaN1

Number

rs2 operand
QNaN2

SNaN2

IEEE 754
IEEE 754
QNaN1
QSNaN1

QNaN2
QNaN2
QNaN2
QSNaN1

QSNaN2
QSNaN2
QSNaN2
QSNaN2

QSNaNn means a quiet NaN produced by the NaN transformation on a signaling NaN from rsn;
the invalid exception is always indicated. The QNaNn results in the table never generate an exception, but IEEE 754 specifies several cases of invalid exceptions, and QNaN results from operands
that are both numbers.

B.3 Trapped Underflow Definition (UFM = 1)
Underflow occurs if the exact unrounded result has magnitude between zero and the smallest normalized number in the destination format.
IMPL. DEP. #55: Whether tininess (in IEEE 754 terms) is detected before or after rounding is implementation-dependent. It is recommended that tininess be detected before rounding.

Note that the wrapped exponent results intended to be delivered on trapped underflows and overflows in IEEE 754 are irrelevant to SPARC-V9 at the hardware and supervisor software levels; if
they are created at all, it would be by user software in a user-mode trap handler.

B.4 Untrapped Underflow Definition (UFM = 0)
Underflow occurs if the exact unrounded result has magnitude between zero and the smallest normalized number in the destination format, and the correctly rounded result in the destination format is inexact.
Table 28 summarizes what happens when an exact unrounded value u satisfying
0 ≤ |u| ≤ smallest normalized number
would round, if no trap intervened, to a rounded value r which might be zero, subnormal, or the
smallest normalized value. “UF” means underflow trap (with ufc set in cexc), “NX” means inexact
trap (with nxc set in cexc), “uf” means untrapped underflow exception (with ufc set in cexc and
ufa in aexc), and “nx” means untrapped inexact exception (with nxc set in cexc and nxa in aexc).

Table 28—Untrapped Floating-Point Underflow

u=r

u≠r

Underflow trap:
Inexact trap:
r is minimum normal
r is subnormal
r is zero

UFM = 1
NXM = ?
None
UF
None

UFM = 0
NXM = 1
None
None
None

UFM = 0
NXM = 0
None
None
None

r is minimum normal
r is subnormal
r is zero

UF †
UF
UF

NX
NX
NX

uf nx
uf nx
uf nx

† If tininess is detected after rounding and NXM

= 1, then NX, otherwise “None” (impl. dep.

#55).

B.5 Integer Overflow Definition
F[sdq]TOi:
When a NaN, infinity, large positive argument ≥ 2147483648.0, or large negative argument ≤ –2147483649.0 is converted to an integer, the invalid_current (nvc) bit of FSR.cexc
should be set and fp_exception_IEEE_754 should be raised. If the floating-point invalid trap
is disabled (FSR.TEM.NVM = 0), no trap occurs and a numerical result is generated: if
the sign bit of the operand is 0, the result is 2147483647; if the sign bit of the operand is 1,
the result is –2147483648.
F[sdq]TOx:
When a NaN, infinity, large positive argument ≥ 263, or large negative argument ≤ –
(263 + 1), is converted to an extended integer, the invalid_current (nvc) bit of FSR.cexc
should be set and fp_exception_IEEE_754 should be raised. If the floating-point invalid trap
is disabled (FSR.TEM.NVM = 0), no trap occurs and a numerical result is generated: if
the sign bit of the operand is 0, the result is 263 – 1; if the sign bit of the operand is 1, the
result is –263.

B.6 Floating-Point Nonstandard Mode
SPARC-V9 implementations are permitted but not encouraged to deviate from IEEE Std 7541985 requirements when the nonstandard mode (NS) bit of the FSR is set (impl. dep. #18).

C SPARC-V9 Implementation Dependencies
This appendixchapter provides a summary of all implementation dependencies in the SPARC-V9
standard. The notation “IMPL. DEP. #nn:” is used to identify the definition of an implementation
dependency; the notation “(impl. dep. #nn)” is used to identify a reference to an implementation
dependency. The number nn provides an index into table 29 on page 282.

SPARC International maintains a document, Implementation Characteristics of Current
SPARC-V9-based Products, Revision 9.x, which describes the implementation-dependent design
features of SPARC-V9-compliant implementations. Contact SPARC International for this document at
SPARC International
535 Middlefield Rd, Suite 210
Menlo Park, CA 94025
(415) 321-8692

C.1 Definition of an Implementation Dependency
The SPARC-V9 architecture is a model that specifies unambiguously the behavior observed by
software on SPARC-V9 systems. Therefore, it does not necessarily describe the operation of the
hardware of any actual implementation.
An implementation is not required to execute every instruction in hardware. An attempt to execute a SPARC-V9 instruction that is not implemented in hardware generates a trap. Whether an
instruction is implemented directly by hardware, simulated by software, or emulated by firmware
is implementation-dependent (impl. dep. #1).
The two levels of SPARC-V9 compliance are described in 1.2.6, “SPARC-V9 Compliance.”
Some elements of the architecture are defined to be implementation-dependent. These elements
include certain registers and operations that may vary from implementation to implementation,
and are explicitly identified as such in this appendixchapter.
Implementation elements (such as instructions or registers) that appear in an implementation but
are not defined in this document (or its updates) are not considered to be SPARC-V9 elements of
that implementation.

C.2 Hardware Characteristics
Hardware characteristics that do not affect the behavior observed by software on SPARC-V9 systems are not considered architectural implementation dependencies. A hardware characteristic
may be relevant to the user system design (for example, the speed of execution of an instruction)
or may be transparent to the user (for example, the method used for achieving cache consistency).
The SPARC International document, Implementation Characteristics of Current SPARCV9-based Products, Revision 9.x, provides a useful list of these hardware characteristics, along
with the list of implementation-dependent design features of SPARC-V9-compliant implementations.
In general, hardware characteristics deal with
— Instruction execution speed
— Whether instructions are implemented in hardware

— The nature and degree of concurrency of the various hardware units comprising a SPARCV9 implementation.

C.3 Implementation Dependency Categories
Many of the implementation dependencies can be grouped into four categories, abbreviated by
their first letters throughout this appendixchapter:
Value (v):
The semantics of an architectural feature are well-defined, except that a value associated
with it may differ across implementations. A typical example is the number of implemented register windows (Implementation dependency #2).
Assigned Value (a):
The semantics of an architectural feature are well-defined, except that a value associated
with it may differ across implementations and the actual value is assigned by SPARC
International. Typical examples are the impl field of Version register (VER) (Implemententation dependency #13) and the FSR.ver field (Implementation dependency #19).
Functional Choice (f):
The SPARC-V9 architecture allows implementors to choose among several possible
semantics related to an architectural function. A typical example is the treatment of a catastrophic error exception, which may cause either a deferred or a disrupting trap (Implementation dependency #31).
Total Unit (t):
The existence of the architectural unit or function is recognized, but details are left to each
implementation. Examples include the handling of I/O registers (Implementation dependency #7) and some alternate address spaces (Implementation dependency #29).

C.4 List of Implementation Dependencies
Table 29 provides a complete list of the implementation dependencies in the architecture, the definition of each, and references to the page numbers in the standard where each is defined or referenced. Most implementation dependencies occur because of the address spaces, I/O registers,
registers (including ASRs), the type of trapping used for an exception, the handling of errors, or
miscellaneous non-SPARC-V9-architectural units such as the MMU or caches (which affect the
FLUSH instruction).
Table 29—Implementation Dependencies
Number Category
1

Def / Ref
page #
8, 281

Description
Software emulation of instructions
Whether an instruction is implemented directly by hardware, simulated by software, or emulated by firmware is implementationdependent.

Table 29—Implementation Dependencies (Continued)
Number Category

Def / Ref
page #

Description

15, 30, 32, 58

Number of IU registers
An implementation of the IU may contain from 64 to 528 generalpurpose 64-bit r registers. This corresponds to a grouping of the
registers into two sets of eight global r registers, plus a circular
stack of from three to 32 sets of 16 registers each, known as register windows. Since the number of register windows present
(NWINDOWS) is implementation-dependent, the total number of
registers is also implementation-dependent.

Incorrect IEEE Std 754-1985 results
An implementation may indicate that a floating-point instruction
did not produce a correct IEEE Std 754-1985 result by generating
a special floating-point unfinished or unimplemented exception. In
this case, privileged mode software shall emulate any functionality
not present in the hardware.

4-5

—

Reserved

18, 121

I/O registers privileged status
Whether I/O registers can be accessed by nonprivileged code is
implementation-dependent.

18, 121

I/O register definitions
The contents and addresses of I/O registers are implementationdependent.

20, 30, 35, 60, RDASR/WRASR target registers
214, 215, 245, Software can use read/write ancillary state register instructions to
286, 54
read/write implementation-dependent processor registers (ASRs
16-31).

20, 36, 60, 245, RDASR/WRASR privileged status
79, 242
Whether each of the implementation-dependent read/write ancillary state register instructions (for ASRs 16-31) is privileged is
implementation-dependent.

10-12

—

Reserved

VER.impl
VER.impl uniquely identifies an implementation or class of software-compatible implementations of the architecture. Values
FFF016 ..FFFF16 are reserved and are not available for assignment.

14-15

—

Reserved

IU deferred-trap queue
The existence, contents, and operation of an IU deferred-trap
queue are implementation-dependent; it is not visible to user application programs under normal operating conditions.

—

Reserved

44, 250

Nonstandard IEEE 754-1985 results
Bit 22 of the FSR, FSR_nonstandard_fp (NS), when set to 1,
causes the FPU to produce implementation-defined results that
may not correspond to IEEE Standard 754-1985.

Table 29—Implementation Dependencies (Continued)
Number Category

Def / Ref
page #

Description

FPU version, FSR.ver
Bits 19:17 of the FSR, FSR.ver, identify one or more implementations of the FPU architecture.

20-21

—

Reserved

FPU TEM, cexc, and aexc
An implementation may choose to implement the TEM, cexc, and
aexc fields in hardware in either of two ways (see 5.1.7.11 for
details).

61, 115, 115

Floating-point traps
Floating-point traps may be precise or deferred. If deferred, a
floating-point deferred-trap queue (FQ) must be present.

30, 212

FPU deferred-trap queue (FQ)
The presence, contents of, and operations on the floating-point
deferred-trap queue (FQ) are implementation-dependent.

47, 212, 213,
213

RDPR of FQ with nonexistent FQ
On implementations without a floating-point queue, an attempt to
read the FQ with an RDPR instruction shall cause either an illegal_
instruction exception or an fp_exception_other exception with
FSR.ftt set to 4 (sequence_error).

26-28

—

18, 74, 75

Address space identifier (ASI) definitions
The following ASI assignments are implementation-dependent:
restricted ASIs 0016 ..0316, 0516 ..0B16, 0D16 ..0F16, 1216 ..1716,
and 1A16 ..7F16; and unrestricted ASIs C016 .. FF16.

ASI address decoding
An implementation may choose to decode only a subset of the 8bit ASI specifier; however, it shall decode at least enough of the
ASI to distinguish ASI_PRIMARY, ASI_PRIMARY_LITTLE, ASI_
AS_IF_USER_PRIMARY,
ASI_AS_IF_USER_PRIMARY_LITTLE,
ASI_PRIMARY_NOFAULT,
ASI_PRIMARY_NOFAULT_LITTLE,
ASI_SECONDARY, ASI_SECONDARY_LITTLE, ASI_AS_IF_USER_
SECONDARY, ASI_AS_IF_USER_SECONDARY_LITTLE, ASI_
SECONDARY_NOFAULT, and ASI_SECONDARY_NOFAULT_LITTLE. If ASI_NUCLEUS and ASI_NUCLEUS_LITTLE are supported
(impl. dep. #124), they must be decoded also. Finally, an implementation must always decode ASI bit<7> while
PSTATE.PRIV = 0, so that an attempt by nonprivileged software
to access a restricted ASI will always cause a privileged_action
exception.

Reserved

90, 93, 114, 115, Catastrophic error exceptions
115
The causes and effects of catastrophic error exceptions are implementation-dependent. They may cause precise, deferred, or disrupting traps.
96

Deferred traps
Whether any deferred traps (and associated deferred-trap queues)
are present is implementation-dependent.

Table 29—Implementation Dependencies (Continued)
Number Category

Def / Ref
page #

Description

98, 114, 114, Trap precision
114, 114, 115, Exceptions that occur as the result of program execution may be
116
precise or deferred, although it is recommended that such exceptions be precise. Examples include mem_address_not_aligned and
division_by_zero.

100

Interrupt clearing
How quickly a processor responds to an interrupt request and the
method by which an interrupt request is removed are implementation-dependent.

93, 102, 103,
104, 113, 115

Implementation-dependent traps
Trap type (TT) values 06016 ..07F16 are reserved for implementation-dependent exceptions. The existence of implementation_
dependent_n traps and whether any that do exist are precise,
deferred, or disrupting is implementation-dependent.

104

Trap priorities
The priorities of particular traps are relative and are implementation-dependent, because a future version of the architecture may
define new traps, and implementations may define implementation-dependent traps that establish new relative priorities.

Reset trap
Some of a processor’s behavior during a reset trap is implementation-dependent.

108

Effect of reset trap on implementation-dependent registers
Implementation-dependent registers may or may not be affected
by the various reset traps.

Entering error_state on implementation-dependent errors
The processor may enter error_state when an implementationdependent error condition occurs.

Error_state processor state
What occurs after error_state is entered is implementation-dependent, but it is recommended that as much processor state as possible be preserved upon entry to error_state.

—

Reserved

t,f,v

168

FLUSH instruction
If FLUSH is not implemented in hardware, it causes an illegal_
instruction exception and its function is performed by system software. Whether FLUSH traps is implementation-dependent.

—

Reserved

174, 177

45 - 46

—

Data access FPU trap
If a load floating-point instruction traps with any type of access
error exception, the contents of the destination floating-point register(s) either remain unchanged or are undefined.
Reserved

Table 29—Implementation Dependencies (Continued)
Number Category

Def / Ref
page #

Description

214, 215, 215, RDASR
215
RDASR instructions with rd in the range 16..31 are available for
implementation-dependent uses (impl. dep. #8). For an RDASR
instruction with rs1 in the range 16..31, the following are implementation-dependent: the interpretation of bits 13:0 and 29:25 in
the instruction, whether the instruction is privileged (impl. dep.
#9), and whether it causes an illegal_instruction trap.

244, 244, 245, WRASR
245, 245
WRASR instructions with rd in the range 16..31 are available for
implementation-dependent uses (impl. dep. #8). For a WRASR
instruction with rd in the range 16..31, the following are implementation-dependent: the interpretation of bits 18:0 in the instruction, the operation(s) performed (for example, xor) to generate the
value written to the ASR, whether the instruction is privileged
(impl. dep. #9), and whether it causes an illegal_instruction trap.

49-54

—

56-100

—

101

102

103

104

—

Reserved

49, 49, 249, 250 Floating-point underflow detection
Whether "tininess" (in IEEE 754 terms) is detected before or after
rounding is implementation-dependent. It is recommended that
tininess be detected before rounding.
—

Reserved

21, 54, 55, 55, Maximum trap level
56, 57
It is implementation-dependent how many additional levels, if any,
past level 4 are supported.
114

Clean windows trap
An implementation may choose either to implement automatic
“cleaning” of register windows in hardware, or generate a clean_
window trap, when needed, for window(s) to be cleaned by software.

206, 206, 207, Prefetch instructions
207, 207, 209, The following aspects of the PREFETCH and PREFETCHA
210
instructions are implementation-dependent: (1) whether they have
an observable effect in privileged code; (2) whether they can cause
a data_access_MMU_miss exception; (3) the attributes of the
block of memory prefetched: its size (minimum = 64 bytes) and its
alignment (minimum = 64-byte alignment); (4) whether each variant is implemented as a NOP, with its full semantics, or with common-case prefetching semantics; (5) whether and how variants
16..31 are implemented.
57

VER.manuf
VER.manuf contains a 16-bit semiconductor manufacturer code.
This field is optional, and if not present reads as zero. VER.manuf
may indicate the original supplier of a second-sourced chip in
cases involving mask-level second-sourcing. It is intended that the
contents of VER.manuf track the JEDEC semiconductor manufacturer code as closely as possible. If the manufacturer does not have
a JEDEC semiconductor manufacturer code, SPARC International
will assign a VER.manuf value.

Table 29—Implementation Dependencies (Continued)
Number Category

Def / Ref
page #

Description

105

TICK register
The difference between the values read from the TICK register on
two reads should reflect the number of processor cycles executed
between the reads. If an accurate count cannot always be returned,
any inaccuracy should be small, bounded, and documented. An
implementation may implement fewer than 63 bits in
TICK.counter; however, the counter as implemented must be able
to count for at least 10 years without overflowing. Any upper bits
not implemented must read as zero.

106

85, 171

IMPDEPn instructions
The IMPDEP1 and IMPDEP2 instructions are completely implementation-dependent. Implementation-dependent aspects include
their operation, the interpretation of bits 29:25 and 18:0 in their
encodings, and which (if any) exceptions they may cause.

107

179, 179, 181, Unimplemented LDD trap
181
It is implementation-dependent whether LDD and LDDA are
implemented in hardware. If not, an attempt to execute either will
cause an unimplemented_LDD trap.

108

117, 229, 230, Unimplemented STD trap
232, 232
It is implementation-dependent whether STD and STDA are
implemented in hardware. If not, an attempt to execute either will
cause an unimplemented_STD trap.

109

115, 174, 174, LDDF_mem_address_not_aligned
177
LDDF and LDDFA require only word alignment. However, if the
effective address is word-aligned but not doubleword-aligned,
either may cause an LDDF_mem_address_not_aligned trap, in
which case the trap handler software shall emulate the LDDF (or
LDDFA) instruction and return.

110

116, 226, 226, STDF_mem_address_not_aligned
228, 228
STDF and STDFA require only word alignment in memory. However, if the effective address is word-aligned but not doublewordaligned, either may cause an STDF_mem_address_not_aligned
trap, in which case the trap handler software shall emulate the
STDF or STDFA instruction and return.

111

116, 174, 174, LDQF_mem_address_not_aligned
177
LDQF and LDQFA require only word alignment. However, if the
effective address is word-aligned but not quadword-aligned, either
may cause an LDQF_mem_address_not_aligned trap, in which
case the trap handler software shall emulate the LDQF (or
LDQFA) instruction and return.

112

117, 226, 226, STQF_mem_address_not_aligned
228, 228
STQF and STQFA require only word alignment in memory. However, if the effective address is word-aligned but not quadwordaligned, either may cause an STQF_mem_address_not_aligned
trap, in which case the trap handler software shall emulate the
STQF or STQFA instruction and return.

Table 29—Implementation Dependencies (Continued)
Number Category

Def / Ref
page #

Description

113

52, 119
Whether the Partial Store Order (PSO) or Relaxed Memory Order
(RMO) models are supported is implementation-dependent.

114

RED_state trap vector address (RSTVaddr)
The RED_state trap vector is located at an implementation-dependent address referred to as RSTVaddr.

115

RED_state processor state
What occurs after the processor enters RED_state is implementation-dependent.

116

223

SIR_enable control flag
The location of the SIR_enable control flag and the means of
accessing the SIR_enable control flag are implementation-dependent. In some implementations, it may be permanently zero.

117

118

119

53,129
The effect of writing an unimplemented memory-mode designation into PSTATE.MM is implementation-dependent.

120

121, 130,
153, 182, 187,
224, 234, 235

Coherence and atomicity of memory operations
The coherence and atomicity of memory operations between processors and I/O DMA memory accesses are implementationdependent.

121

Implementation-dependent memory model
An implementation may choose to identify certain addresses and
use an implementation-dependent memory model for references to
them.

122

123

18, 121, 130

Input/output (I/O) semantics
The semantic effect of accessing input/output (I/O) registers is
implementation-dependent.

124

75, 74,
122, 96

Implicit ASI when TL > 0
When TL > 0, the implicit ASI for instruction fetches, loads, and
stores is implementation-dependent. See F.4.4, “Contexts,” for
more information.

125

207, 282
Whether Prefetch and Non-faulting Load always succeed when the
MMU is disabled is implementation-dependent.
121

Identifying I/O locations
The manner in which I/O locations are identified is implementation-dependent.

131, 168, The latency between the execution of FLUSH on one processor
168
and the point at which the modified instructions have replaced outdated instructions in a multiprocessor is implementation-dependent.

53, 80, 151, 172, Address masking
215
When PSTATE.AM = 1, the value of the high-order 32-bits of the
PC transmitted to the specified destination register(s) by CALL,
JMPL, RDPC, and on a trap is implementation-dependent.

Table 29—Implementation Dependencies (Continued)
Number Category
126

127

Def / Ref
page #

Description

58, 58, 59, 59, Register Windows State Registers Width
59, 60
Privileged registers CWP, CANSAVE, CANRESTORE, OTHERWIN, and CLEANWIN contain values in the range
0..NWINDOWS-1. The effect of writing a value greater than
NWINDOWS-1 to any of these registers is undefined.Although the
width of each of these five registers is nominally 5 bits, the width
is
implementation-dependent
and
shall
be
between
log2(NWINDOWS) and 5 bits, inclusive. If fewer than 5 bits are
implemented, the unimplemented upper bits shall read as 0 and
writes to them shall have no effect. All five registers should have
the same width.
52, 56

PSTATE PID bits
The presence and semantics of PSTATE.PID1 and PSTATE.PID0
are implementation-dependent. The presence of TSTATE bits 19
and 18 is implementation-dependent. If PSTATE bit 11 (10) is
implemented, TSTATE bit 19 (18) shall be implemented and contain the state of PSTATE bit 11 (10) from the previous trap level..
If PSTATE bit 11 (10) is not implemented, TSTATE bit 19 (18)
shall read as zero. Software intended to run on multiple implementations should only write these bits to values previously read from
PSTATE, or to zeroes.

D Formal Specification of the Memory Models
This appendix provides a formal description of the SPARC-V9 processor’s interaction with memory. The formal description is more complete and more precise than the description of Chapter 8,
“Memory Models,” and therefore represents the definitive specification. Implementations must
conform to this model, and programmers must use this description to resolve any ambiguity.
This formal specification is not intended to be a description of an actual implementation, only to
describe in a precise and rigorous fashion the behavior that any conforming implementation must
provide.

D.1 Processors and Memory
The system model consists of a collection of processors, P0, P1,..Pn-1. Each processor executes its
own instruction stream.1 Processors may share address space and access to real memory and I/O
locations.
To improve performance, processors may interpose a cache or caches in the path between the processor and memory. For data and I/O references, caches are required to be transparent. The memory model specifies the functional behavior of the entire memory subsystem, which includes any

1. Processors are equivalent to their software abstraction, processes, provided that context switching is properly performed. See Appendix J, “Programming With the Memory Models,” for an example of context switch code.

form of caching. Implementations must use appropriate cache coherency mechanisms to achieve
this transparency.1
The SPARC-V9 memory model requires that all data references be consistent but does not require
that instruction references or input/output references be maintained consistent. The FLUSH
instruction or an appropriate operating system call may be used to ensure that instruction and data
spaces are consistent. Likewise, system software is needed to manage the consistency of I/O operations.
The memory model is a local property of a processor that determines the order properties of memory references. The ordering properties have global implications when memory is shared, since
the memory model determines what data is visible to observing processors and in what order.
Moreover, the operative memory model of the observing processor affects the apparent order of
shared data reads and writes that it observes.

D.2 An Overview of the Memory Model Specification
The underlying goal of the memory model is to place the weakest possible constraints on the processor implementations and to provide a precise specification of the possible orderings of memory
operations so that shared-memory multiprocessors can be constructed.
An execution trace is a sequence of instructions with a specified initial instruction. An execution
trace constitutes one possible execution of a program and may involve arbitrary reorderings and
parallel execution of instructions. A self-consistent execution trace is one that generates precisely
the same results as those produced by a program order execution trace.
A program order execution trace is an execution trace that begins with a specified initial
instruction and executes one instruction at a time in such a fashion that all the semantic effects of
each instruction take effect before the next instruction is begun. The execution trace this process
generates is defined to be program order.
A program is defined by the collection of all possible program order execution traces.
Dependence order is a partial order on the instructions in an execution trace that is adequate to
ensure that the execution trace is self-consistent. Dependence order can be constructed using conventional data dependence analysis techniques. Dependence order holds only between instructions in the instruction trace of a single processor; instructions that are part of execution traces on
different processors are never dependence-ordered.
Memory order is a total order on the memory reference instructions (loads, stores, and atomic
load/stores) which satisfies the dependence order and, possibly, other order constraints such as
those introduced implicitly by the choice of memory model or explicitly by the appearance of
memory barrier (MEMBAR) instructions in the execution trace. The existence of a global memory order on the performance of all stores implies that memory access is write-atomic.2
1. Philip Bitar and Alvin M. Despain, “Multiprocessor Cache Synchronization: Issues, Innovations, Evolution,”
Proc. 13th Annual International Symposium on Computer Architecture, Computer Architecture News 14:2, June
1986, pp.424-433.

A memory model is a set of rules that constrain the order of memory references. The SPARC-V9
architecture supports three memory models: total store order (TSO), partial store order (PSO), and
relaxed memory order (RMO). The memory models are defined only for memory and not for I/O
locations. See 8.2, “Memory, Real Memory, and I/O Locations,” for more information.
The formal definition used in the SPARC-V8 specification1 remains valid for the definition of
PSO and TSO, except for the FLUSH instruction, which has been modified slightly.2 The
SPARC-V9 architecture introduces a new memory model, RMO, which differs from TSO and
PSO in that it allows load operations to be reordered as long as single thread programs remain
self-consistent.

D.3 Memory Transactions
D.3.1 Memory Transactions
A memory transaction is one of the following:
Store:
A request by a processor to replace the value of a specified memory location. The address
and new value are bound to the store transaction when the processor initiates the store
transaction. A store is complete when the new value is visible to all processors in the system.
Load:
A request by a processor to retrieve the value of the specified memory location. The
address is bound to the load transaction when the processor initiates the load transaction.
A load is complete when the value being returned cannot be modified by a store made by
another processor.
Atomic:
A load/store pair with the guarantee that no other memory transaction will alter the state
of the memory between the load and the store. The SPARC-V9 instruction set includes
three atomic instructions: LDSTUB, SWAP and CAS.3 An atomic transaction is considered to be both a load and a store.4

2. W.W. Collier, “Reasoning About Parallel Architectures”, Prentice-Hall, 1992 includes an excellent discussion of
write-atomicity and related memory model topics.
1. Pradeep Sindhu, Jean-Marc Frailong, and Michel Ceklov. “Formal Specification of Memory Models,” Xerox Palo
Alto Research Center Report CSL-91-11, December 1991.
2. In SPARC-V8, a FLUSH instruction needs at least five instruction execution cycles before it is guaranteed to have
local effects; in SPARC-V9 this five-cycle requirement has been removed.
3. There are three generic forms. CASA and CASXA reference 32-bit and 64-bit objects respectively. Both normal
and alternate ASI forms exist for LDSTUB and SWAP. CASA and CASXA only have alternate forms, however, a
CASA (CASXA) with ASI = ASI_PRIMARY{_LITTLE} is equivalent to CAS (CASX). Synthetic instructions
for CAS and CASX are suggested in G.3, “Synthetic Instructions.”

Flush:
A request by a processor to force changes in the data space aliased to the instruction space
to become consistent. Flush transactions are considered to be store operations for memory
model purposes.
Memory transactions are referred to by capital letters: Xna, which denotes a specific memory
transaction X by processor n to memory address a. The processor index and the address are specified only if needed. The predicate S(X) is true if and only if X has store semantics. The predicate
L(X) is true if and only if X has load semantics.
MEMBAR instructions are not memory transactions; rather they convey order information above
and beyond the implicit ordering implied by the memory model in use. MEMBAR instructions
are applied in program order.
D.3.2 Program Order
The program order is a per-processor total order that denotes the sequence in which processor n
logically executes instructions. The program order relation is denoted by 0)
MOVr
See table 43

—
FPop1
See table 34
FPop2
See table 41
IMPDEP1
IMPDEP2

JMPL

DONEP (fcn = 0)
RETRYP (fcn = 1)

—

Table 33—op3[5:0] (op = 3)
op3 [5:4]
0

0
LDUW

1
LDUWAPASI

LDUB

LDUBAPASI

LDUH
LDDD
STW

PASI

STB

2
3
4
5

op3
[3:0]

6
7
8

2
LDF
LDFSRD, LDXFSR

3
LDFAPASI

LDUHA
LDDAD, PASI
STWAPASI

LDQF
LDDF
STF

LDQFAPASI
LDDFAPASI
STFAPASI

STBAPASI

STFSRD, STXFSR

—

STQF
STDF

STQFAPASI
STDFAPASI

—
—
—
—
—

—
—
—
—

PASI

STH
STDD
LDSW

STHA
STDAPASI
LDSWAPASI

LDSB

LDSBAPASI

LDSH

LDSHAPASI

LDX

LDXAPASI

—

D
E

LDSTUB
STX

LDSTUBAPASI
STXAPASI

SWAPD

SWAPAD, PASI

PREFETCH

—
—

—

CASAPASI
PREFETCHAPASI
CASXAPASI

—

This appendix is informative only.
It is not part of the SPARC-V9 specification.

Table 34—opf[8:0] (op = 2,op3 = 3416 = FPop1)
opf[3:0]
opf
[8:4]
00

—
01
—
02
—
03
—
04
—
05
—
06
—
07
—
08
—
09
—
0A
—
0B
—
0C
—
0D
—
0E..1F —

—
—
—

—
—
—
—
—
—
—
—

FsTOx

FdTOx

FqTOx

FxTOs

—
—
—
—

—
—
—

FsTOi

FdTOi

FqTOi

—

FMOVs FMOVd FMOVq

—
—
—

FADDs FADDd FADDq

FiTOs

—
—

FNEGs FNEGd FNEGq

—
—
—
FSUBs

—
—
—
—
—
—
—
—
—
—

—
—
—

FSUBd FSUBq

—
—
—
—
—
—
—
—

FABSs

FABSd

FABSq

—

—
—
—
—
—
—
—
—

—
—
—
—

FDIVs

FDIVd

FDIVq

—

—
—
—
—
—
—
—
—
—
—

FSQRTs FSQRTd FSQRTq

—

FMULs

FMULd

FMULq

—

—
—
—
—
—
—
—
—
—
—

—
—
—
—
—
—
—

FxTOq

—
—
—

—
—
—
—
—
—
—

FqTOd

FiTOq

FsTOq

FdTOq

—
—

—
—
—
—
—
—
—

FxTOd

—
—
—

—
—
—
—
—

FdTOs

FqTOs

FiTOd

FsTOd

—
—

FsMULd

(This Annex is not a part of SPARC-V9/R1.4.5,
The SPARC Architecture Manual;
it is included for information only.)

F SPARC-V9 MMU Requirements
F.1 Introduction
This appendix describes the boundary conditions that all SPARC-V9 MMUs must satisfy. The
appendix does not define the architecture of any specific memory management unit. It is possible
to build a SPARC-V9-compliant system without an MMU.

FdMULq

—
—
—
—
—

F.1.1 Definitions
address space:
A range of locations accessible with a 64-bit virtual address. Different address spaces may
use the same virtual address to refer to different physical locations.
aliases:
Two virtual addresses are aliases of each other if they refer to the same physical address.
context:
A set of translations used to support a particular address space.
page:
The range of virtual addresses translated by a single translation element. The size of a
page is the size of the range translated by a single translation element. Different pages may
have different sizes. Associated with a page or with a translation element are attributes
(e.g., restricted, permission, etc.) and statistics (e.g., referenced, modified, etc.)
translation element:
Used to translate a range of virtual addresses to a range of physical addresses.

F.2 Overview
All SPARC-V9 MMUs must provide the following basic functions:
— Translate 64-bit virtual addresses to physical addresses. This translation may be implemented with one or more page sizes.
— Provide the RED_state operation, as defined in 7.2.1, “RED_state.”
— Provide a method for disabling the MMU. When the MMU is disabled, no translation
occurs: Physical Address = Virtual Address, where N is implementationdependent. Furthermore, the disabled MMU will not perform any memory protection (see
F.4.2, “Memory Protection”) or prefetch and non-faulting load violation (see F.4.3,
“Prefetch and Non-Faulting Load Violation”) checks.
IMPL. DEP. #117: Whether PREFETCH and non-faulting load always succeed when the MMU is
disabled is implementation-dependent.

— Provide page-level protections. Conventional protections (Read, Write, Execute) for both
privileged and nonprivileged accesses may be provided.
— Provide page-level enabling and disabling of prefetch and non-faulting load operation.
The MMU, however, need not provide separate protection mechanisms for prefetch and
non-faulting load.
— Support multiple address spaces (ASIs). The MMU must support the address spaces as
defined in F.3.1, “Information the MMU Expects from the Processor.”
— Provide page-level statistics such as referenced and modified.

The above requirements apply only to those systems that include SPARC-V9 MMUs. See F.8,
“SPARC-V9 Systems without an MMU.”

F.3 The Processor-MMU Interface
A SPARC-V9 MMU must support at least two types of addresses:
(1) Virtual Addresses, which map all system-wide, program-visible memory. A SPARC-V9
MMU may choose not to support translation for the entire 64-bit virtual address space, as
long as addresses outside the supported virtual address range are treated either as
No_translation or Translation_not_valid (see F.3.3, “Information the MMU Sends to the
Processor”).
(2) Physical Addresses, which map real physical memory and I/O device space. There is no
minimum requirement for how many physical address bits a SPARC-V9 MMU must support.
A SPARC-V9 MMU translates virtual addresses from the processor into physical addresses, as
illustrated in figure 48.
Data
Physical
Address
Space

Processor
Virtual
Address

MMU

Physical
Address

Real

I/O

Memory

Locations

Figure 48—Logical Diagram of a SPARC-V9 System with an MMU

Figure 48 shows only the address and data paths between the processor and the MMU. The control interface between the processor and the MMU is discussed in F.3.1, “Information the MMU
Expects from the Processor,” and F.3.3, “Information the MMU Sends to the Processor.”
F.3.1 Information the MMU Expects from the Processor
A SPARC-V9 MMU expects the following information to accompany each virtual address from
the processor:
RED_state:
Indicates whether the MMU should operate in RED_state, as defined in 7.2.1,
“RED_state.”
Data / Instruction:
Indicates whether the access is an instruction fetch or data access (load or store).

Prefetch:
Indicates whether the data (Data / Instruction = Data) access was initiated by one of the
SPARC-V9 prefetch instructions.
Privileged:
Indicates whether this access is privileged.
Read / Write:
Indicates whether this access is a read (instruction fetch or data load) or a write (data
store) operation.
Atomic:
Indicates whether this is an atomic load-store operation. Whenever atomic is asserted, the
value of “Read/Write” is treated by the MMU as “don’t care.”
ASI:
An 8-bit address space identifier. See 6.3.1.3, “Address Space Identifiers (ASIs),” for the
list of ASIs that the MMU must support.
F.3.2 Attributes the MMU Associates with Each Mapping
In addition to translating virtual addresses to physical addresses, a SPARC-V9 MMU also stores
associated attributes, either with each mapping or with each page, depending upon the implementation. Some of these attributes may be associated implicitly, as opposed to explicitly, with the
mapping. This information includes
Restricted:
Only privileged accesses are allowed (see F.3.1, “Information the MMU Expects from the
Processor”); nonprivileged accesses are disallowed.
Read, Write, and Execute Permissions:
An MMU may allow zero or more of read, write, and execute permissions, on a per-mapping basis. Read permission is necessary for data read accesses and atomic accesses. Write
permission is necessary for data write accesses and atomic accesses. Execute permission is
necessary for instruction accesses. At a minimum, an MMU must allow for “all permissions,” “no permissions,” and “no write permission”; optionally, it can provide “execute
only” and “write only,” or any combination of “read/write/execute” permissions.
Prefetchable:
The presence of this attribute indicates that accesses made with the prefetch indication
from the processor are allowed; otherwise, they are disallowed. See F.3.1, “Information
the MMU Expects from the Processor.”
Non-faultable:
The presence

this

attribute indicates that accesses made with
and ASI_SECONDARY_NOFAULT{_LITTLE} are allowed;
otherwise, they are disallowed. An implementation may choose to combine the prefetchable and non-faultable attributes into a single “No Side Effects” attribute; that is, “reads
from this address do not cause side effects, such as clear on read.”
ASI_PRIMARY_NOFAULT{_LITTLE}

F.3.3 Information the MMU Sends to the Processor
The processor can expect one and only one of the following signals coming from any SPARC-V9
MMU for each translation requested:
Translation_error:
The MMU has detected an error (for example, parity error) in the translation process. Can
cause a data_access_error or instruction_access_error exception.
No_translation:
The MMU is unable to translate the virtual address, since no translation exists for it. Some
implementations may not provide this information and provide only
Translation_not_valid.
Can cause either a data_access_exception or an
instruction_access_exception exception.
Translation_not_valid:
The MMU is unable to translate the virtual address, since it cannot find a valid translation.
Some implementations may not provide this information and provide only No_translation.
Can cause either a data_access_MMU_miss or an instruction_access_MMU_miss exception.
Privilege_violation:
The MMU has detected a privilege violation, i.e., an access to a restricted page when the
access does not have the required privilege (see F.3.1, “Information the MMU Expects
from the Processor”). Can cause either a data_access_protection or an
instruction_access_protection exception.
Protection_violation:
The MMU has detected a protection violation, which is defined to be a read, write, or
instruction fetch attempt to a page that does not have read, write, or execute permission,
respectively. Can cause either a data_access_protection or an instruction_access_protection
exception.
Prefetch_violation:
The MMU has detected an attempt to prefetch from a page for which prefetching is disabled.
NF-Load_violation:
The MMU has detected an attempt to perform a non-faulting load from a page for which
non-faulting loads are disabled.
Translation_successful:
The MMU has successfully translated the virtual address to a physical address; none of the
conditions described above has been detected.

F.4 Components of the SPARC-V9 MMU Architecture
A SPARC-V9 MMU should contain the following:
— Logic that implements virtual-to-physical address translation

— Logic that provides memory protection
— Logic that supports prefetching as noted in A.42, “Prefetch Data”
— Logic that supports non-faulting loading, as noted in 8.3, “Addressing and Alternate
Address Spaces”
— A method for specifying the primary, secondary and, optionally, nucleus address spaces
— A method for supplying information related to failed translations
— A method for collecting “referenced” and “modified” statistics
F.4.1 Virtual-to-Physical Address Translation
A SPARC-V9 MMU tries to translate every virtual address it receives into a physical address as
long as:
— The MMU is enabled.
— The processor indicates that this is a non-RED_state instruction fetch (see the Data/
Instruction description in F.3.1, “Information the MMU Expects from the Processor”) or a
data access with an ASI that indicates a translatable address space.
Although the MMU will attempt to translate every virtual address that meets the above two conditions, it need not guarantee that it can provide a translation every time. When the MMU encounters a virtual address that it cannot translate, it asserts either Translation_error, No_translation, or
Translation_not_valid, as discussed in F.3.3, “Information the MMU Sends to the Processor.”
F.4.2 Memory Protection
For each virtual address for which a SPARC-V9 MMU can provide a translation, the MMU
checks whether memory protection would be violated. More specifically, the MMU
— Indicates Privilege_violation (see F.3.3) if the translation information indicates a restricted
page but the access was not privileged (see F.3.1)
— Indicates Protection_violation (see F.3.3) if a read, write, or instruction fetch uses a translation that does not grant read, write, or execute permission, respectively
— Indicates Protection_violation (see F.3.3) if an atomic load-store uses a translation that
does not grant both read and write permission
F.4.3 Prefetch and Non-Faulting Load Violation
For each virtual address, the MMU checks for prefetch or non-faulting load violation as long as
— The MMU can provide a translation, and
— The MMU does not detect any memory protection violation, as discussed in F.4.2, “Memory Protection.”

More specifically, the MMU performs the following before sending the physical address to the
rest of the memory system:
— Asserts Prefetch_violation (see F.3.3) if an access with the prefetch indication (see F.3.1)
uses a translation that lacks the prefetchable attribute (see F.3.2)
— Asserts NF-Load_violation (see F.3.3) if the ASI (see F.3.1) indicates this access is a nonfaulting load, but the translation it uses lacks the non-faultable attribute (see F.3.2)
F.4.4 Contexts
The MMU must support two contexts:
(1) Primary Context
(2) Secondary Context
In addition, it is also recommended that the MMU support a third context:
(3) Nucleus Context
On data accesss, the MMU decides which of these three contexts to use based on the ASI field, as
illustrated in table 35. Because the SPARC-V9 MMU cannot determine the instruction opcode, it
treats all data accesses with ASI_PRIMARY{_LITTLE} as normal loads or stores, even though
the processor may issue them with load/store alternate instructions.
Table 35—Context Used for Data Access
MMU Inputs
ASI

†

Output
Context

ASI_PRIMARY
ASI_PRIMARY_LITTLE
ASI_PRIMARY_NOFAULT
ASI_PRIMARY_NOFAULT_LITTLE
ASI_AS_IF_USER_PRIMARY
ASI_AS_IF_USER_PRIMARY_LITTLE
ASI_SECONDARY
ASI_SECONDARY_LITTLE
ASI_SECONDARY_NOFAULT
ASI_SECONDARY_NOFAULT_LITTLE
ASI_AS_IF_USER_SECONDARY
ASI_AS_IF_USER_SECONDARY_LITTLE

Mode
Either
Either
Either
Either
Privileged
Privileged
Either
Either
Either
Either
Privileged
Privileged

Primary
Primary
Primary
Primary
Primary
Primary
Secondary
Secondary
Secondary
Secondary
Secondary
Secondary

ASI_NUCLEUS †

Privileged

Nucleus

ASI_NUCLEUS_LITTLE †

Privileged

Nucleus

Support for the nucleus context is only a recommendation; if an implementation does not support the
nucleus context it may ignore this row.

On instruction fetch, the MMU decides which context to use based on the ASI field, as illustrated
in table 36. Note that the secondary context is never used for instruction fetch.
Table 36—Context Used for Instruction Access
ASI
ASI_PRIMARY
ASI_NUCLEUS †

Mode
Either

Context
Primary

Privileged ‡

Nucleus

†

Support for the Nucleus Context is only a recommendation; if an implementation does not support the
Nucleus Context it may ignore this row.

‡

It is implementation-dependent whether instruction fetch using ASI_NUCLEUS in nonprivileged
mode is allowed.

F.4.5 Fault Status and Fault Address
A SPARC-V9 MMU must provide the following:
— Fault status information that specifies which condition listed in F.3.3, “Information the
MMU Sends to the Processor,” has resulted in a translation-related processor trap, and any
other information necessary for privileged software to determine the cause of the trap; for
example, ASI, Read/Write, Data/Instruction, etc.
— The Fault address associated with the failed translation. Since the address from an instruction translation failure is available in the processor as the trap PC, the MMU is not
required to save the address of an instruction translation failure.
F.4.6 Referenced and Modified Statistics
A SPARC-V9 MMU shall allow, either through hardware, software, or some combination thereof,
for the collection of “referenced” and “modified” statistics associated with translations and/or
physical pages. That is, there must be a method to determine if a page has been referenced, a
method to determine if a page has been modified, and a method for clearing the indications that a
page has been referenced and/or modified. These statistics may be kept on either a per-translation
basis or a per-physical-page basis.
It is implementation-dependent whether the referenced and/or modified statistics are updated
when an access is performed or when the translation for that access is performed.

F.5 RED_state Processing
It is recommended that the MMU perform as follows when the processor is in RED_state:
— Instruction address translation is a straight-through physical map; that is, the MMU is
always suppressed for instruction access in RED_state.
— Data address translation is handled normally; that is, the MMU is used if it is enabled.
Note that any event which causes the processor to enter RED_state also disables the
MMU, however, the handler executing in RED_state may reenable the MMU.

This appendix is informative only.
It is not part of the SPARC-V9 specification.

F.6 Virtual Address Aliasing
Hardware and privileged software must cooperate so that multiple virtual addresses aliased to the
same physical address appear to be consistent as defined by the memory models described in
Chapter 8, “Memory Models.” Depending upon the implementation, this may require allowing
multiple translations to coexist only if they meet some implementation-dependent alignment constraint, or it may require that software ensure that only one translation is in effect at any given
time.

F.7 MMU Demap Operation
The SPARC-V9 MMU must provide a mechanism for privileged software to invalidate some or
all of the virtual-to-physical address translations.

F.8 SPARC-V9 Systems without an MMU
It is possible to build a SPARC-V9 system that does not have an MMU. Such a system should
behave as if contains an MMU that is disabled.

(This Annex is not a part of SPARC-V9/R1.4.5,
The SPARC Architecture Manual;
it is included for information only.)

G Suggested Assembly Language Syntax
This appendix supports Appendix A, “Instruction Definitions.” Each instruction description in
Appendix A includes a table that describes the suggested assembly language format for that
instruction. This appendix describes the notation used in those assembly language syntax descriptions and lists some synthetic instructions that may be provided by a SPARC-V9 assembler for the
convenience of assembly language programmers.

G.1 Notation Used
The notations defined here are also used in the syntax descriptions in Appendix A.
Items in typewriter font are literals to be written exactly as they appear. Items in italic font
are metasymbols that are to be replaced by numeric or symbolic values in actual SPARC-V9

assembly language code. For example, “imm_asi” would be replaced by a number in the range 0
to 255 (the value of the imm_asi bits in the binary instruction), or by a symbol bound to such a
number.
Subscripts on metasymbols further identify the placement of the operand in the generated binary
instruction. For example, regrs2 is a reg (register name) whose binary value will be placed in the
rs2 field of the resulting instruction.
G.1.1 Register Names
reg:
A reg is an integer register name. It may have any of the following values:1
%r0..%r31
%g0..%g7
(global registers; same as %r0..%r7)
%o0..%o7
(out registers; same as %r8 ..%r15)
%l0..%l7
(local registers; same as %r16..%r23)
%i0..%i7
(in
registers; same as %r24..%r31)
%fp
(frame pointer; conventionally same as %i6)
%sp
(stack pointer; conventionally same as %o6)
Subscripts identify the placement of the operand in the binary instruction as one of the following:
regrs1
(rs1 field)
regrs2
(rs2 field)
regrd
(rd field)
freg:
An freg is a floating-point register name. It may have the following values:
%f0, %f1, %f2 .. %f63
See 5.1.4, “Floating-Point Registers”
Subscripts further identify the placement of the operand in the binary instruction as one of
the following:
fregrs1
(rs1 field)
fregrs2
(rs2 field)
fregrd
(rd field)
asr_reg:
An asr_reg is an Ancillary State Register name. It may have one of the following values:
%asr16..%asr31
Subscripts further identify the placement of the operand in the binary instruction as one of
the following:
asr_regrs1
(rs1 field)
asr_regrd
(rd field)
1.In actual usage, the %sp, %fp, %gn, %on, %ln, and %in forms are preferred over %rn.

i_or_x_cc:
An i_or_x_cc specifies a set of integer condition codes, those based on either the 32-bit
result of an operation (icc) or on the full 64-bit result (xcc). It may have either of the
following values:
%icc
%xcc
fccn:
An fccn specifies a set of floating-point condition codes. It may have any of the following
values:
%fcc0
%fcc1
%fcc2
%fcc3
G.1.2 Special Symbol Names
Certain special symbols appear in the syntax table in typewriter font. They must be written
exactly as they are shown, including the leading percent sign (%).
The symbol names and the registers or operators to which they refer are as follows:
%asi
Address Space Identifier register
%canrestore Restorable Windows register
%cansave
Savable Windows register
%cleanwin
Clean Windows register
%cwp
Current Window Pointer register
%fq
Floating-Point Queue
%fsr
Floating-Point State Register
%otherwin
Other Windows register
%pc
Program Counter register
%pil
Processor Interrupt Level register
%pstate
Processor State register
%tba
Trap Base Address register
%tick
Tick (cycle count) register
%tl
Trap Level register
%tnpc
Trap Next Program Counter register
%tpc
Trap Program Counter register
%tstate
Trap State register
%tt
Trap Type register
%ccr
Condition Codes Register
%fprs
Floating-Point Registers State register
%ver
Version register
%wstate
Window State register

Y register

The following special symbol names are unary operators that perform the functions described:
%uhi
Extracts bits 63..42 (high 22 bits of upper word) of its operand
%ulo
Extracts bits 41..32 (low-order 10 bits of upper word) of its operand
%hi
Extracts bits 31..10 (high-order 22 bits of low-order word) of its operand
%lo
Extracts bits 9..0 (low-order 10 bits) of its operand
Certain predefined value names appear in the syntax table in typewriter font. They must be
written exactly as they are shown, including the leading sharp sign (#).
The value names and the values to which they refer are as follows:
#n_reads
0
(for PREFETCH instruction)
#one_read
1
(for PREFETCH instruction)
#n_writes
2
(for PREFETCH instruction)
#one_write
3
(for PREFETCH instruction)
#page
4
(for PREFETCH instruction)
#Sync
4016
(for MEMBAR instruction cmask field)
#MemIssue
2016
(for MEMBAR instruction cmask field)
#Lookaside
1016
(for MEMBAR instruction cmask field)
#StoreStore 0816
(for MEMBAR instruction mmask field)
#LoadStore
0416
(for MEMBAR instruction mmask field)
#StoreLoad
0216
(for MEMBAR instruction mmask field)
#LoadLoad
0116
(for MEMBAR instruction mmask field)
#ASI_AIUP
1016
ASI_AS_IF_USER_PRIMARY
#ASI_AIUS
1116
ASI_AS_IF_USER_SECONDARY
#ASI_AIUP_L 1816
ASI_AS_IF_USER_PRIMARY_LITTLE
#ASI_AIUS_L 1916
ASI_AS_IF_USER_SECONDARY_LITTLE
#ASI_P
8016
ASI_PRIMARY
#ASI_S
8116
ASI_SECONDARY
#ASI_PNF
8216
ASI_PRIMARY_NOFAULT
#ASI_SNF
8316
ASI_SECONDARY_NOFAULT
#ASI_P_L
8816
ASI_PRIMARY_LITTLE
#ASI_S_L
8916
ASI_SECONDARY_LITTLE
#ASI_PNF_L
8A16
ASI_PRIMARY_NOFAULT_LITTLE
#ASI_SNF_L
8B16
ASI_SECONDARY_NOFAULT_LITTLE
The full names of the ASIs may also be defined:
#ASI_AS_IF_USER_PRIMARY
1016
#ASI_AS_IF_USER_SECONDARY
1116
#ASI_AS_IF_USER_PRIMARY_LITTLE
1816
#ASI_AS_IF_USER_SECONDARY_LITTLE 1916

#ASI_PRIMARY
#ASI_SECONDARY
#ASI_PRIMARY_NOFAULT
#ASI_SECONDARY_NOFAULT
#ASI_PRIMARY_LITTLE
#ASI_SECONDARY_LITTLE
#ASI_PRIMARY_NOFAULT_LITTLE
#ASI_SECONDARY_NOFAULT_LITTLE

8016
8116
8216
8316
8816
8916
8A16
8B16

G.1.3 Values
Some instructions use operand values as follows:
const4
A constant that can be represented in 4 bits
const22
A constant that can be represented in 22 bits
imm_asi
An alternate address space identifier (0..255)
simm7
A signed immediate constant that can be represented in 7 bits
simm10
A signed immediate constant that can be represented in 10 bits
simm11
A signed immediate constant that can be represented in 11 bits
simm13
A signed immediate constant that can be represented in 13 bits
value
Any 64-bit value
shcnt32
A shift count from 0..31
shcnt64
A shift count from 0..63
G.1.4 Labels
A label is a sequence of characters that comprises alphabetic letters (a–z, A–Z [with upper and
lower case distinct]), underscores (_), dollar signs ($), periods (.), and decimal digits (0-9). A
label may contain decimal digits, but may not begin with one. A local label contains digits only.
G.1.5 Other Operand Syntax
Some instructions allow several operand syntaxes, as follows:
reg_plus_imm may be any of the following:
regrs1
(equivalent to regrs1 + %g0)
regrs1 + simm13
regrs1 – simm13
simm13
(equivalent to %g0 + simm13)
simm13 + regrs1 (equivalent to regrs1 + simm13)
address may be any of the following:
regrs1
(equivalent to regrs1 + %g0)
regrs1 + simm13

regrs1 – simm13
simm13
simm13 + regrs1
regrs1 + regrs2

(equivalent to %g0 + simm13)
(equivalent to regrs1 + simm13)

membar_mask is the following:
const7
A constant that can be represented in 7 bits. Typically, this is an expression involving the logical or of some combination of #Lookaside,
#MemIssue, #Sync, #StoreStore, #LoadStore, #StoreLoad, and #LoadLoad.
prefetch_fcn (prefetch function) may be any of the following:
#n_reads
#one_read
#n_writes
#one_write
#page
0..31
regaddr (register-only address) may be any of the following:
regrs1
(equivalent to regrs1 + %g0)
regrs1 + regrs2
reg_or_imm (register or immediate value) may be either of:
regrs2
simm13
reg_or_imm10 (register or immediate value) may be either of:
regrs2
simm10
reg_or_imm11 (register or immediate value) may be either of:
regrs2
simm11
reg_or_shcnt (register or shift count value) may be any of:
regrs2
shcnt32
shcnt64
software_trap_number may be any of the following:
regrs1
(equivalent to regrs1 + %g0)
regrs1 + simm7
regrs1 – simm7
simm7
(equivalent to %g0 + simm7)
simm7 + regrs1
(equivalent to regrs1 + simm7)

regrs1 + regrs2
The resulting operand value (software trap number) must be in the range 0..127, inclusive.
G.1.6 Comments
It is suggested that two types of comments be accepted by SPARC-V9 assemblers: C-style “/
*...*/” comments, which may span multiple lines, and “!...” comments, which extend from
the “!” to the end of the line.

G.2 Syntax Design
The suggested SPARC-V9 assembly language syntax is designed so that
— The destination operand (if any) is consistently specified as the last (rightmost) operand in
an assembly language instruction.
— A reference to the contents of a memory location (in a Load, Store, CASA, CASXA,
LDSTUB(A), or SWAP(A) instruction) is always indicated by square brackets ([]); a reference to the address of a memory location (such as in a JMPL, CALL, or SETHI) is
specified directly, without square brackets.

G.3 Synthetic Instructions
Table 37 describes the mapping of a set of synthetic (or “pseudo”) instructions to actual instructions. These and other synthetic instructions may be provided in a SPARC-V9 assembler for the
convenience of assembly language programmers.
Note that synthetic instructions should not be confused with “pseudo-ops,” which typically provide information to the assembler but do not generate instructions. Synthetic instructions always
generate instructions; they provide more mnemonic syntax for standard SPARC-V9 instructions.
Table 37—Mapping Synthetic to SPARC-V9 Instructions
Synthetic instruction
cmp
regrs1, reg_or_imm
jmp
address
call
address
iprefetch label
tst
regrs1
ret
retl
restore
save

setuw

value,regrd

SPARC-V9 instruction(s)
subcc
jmpl
jmpl
bn,a,pt
orcc
jmpl
jmpl
restore
save

regrs1, reg_or_imm, %g0
address, %g0
address, %o7
%xcc,label
%g0, regrs1, %g0
%i7+8, %g0
%o7+8, %g0
%g0, %g0, %g0
%g0, %g0, %g0

sethi

%hi(value), regrd

Comment
compare

instruction prefetch
test
return from subroutine
return from leaf subroutine
trivial restore
trivial save
(Warning: trivial save
should only be used in kernel
code!)
(when ((value&3FF16) = = 0))

Table 37—Mapping Synthetic to SPARC-V9 Instructions (Continued)
Synthetic instruction

SPARC-V9 instruction(s)

sethi
or

— or —
%g0, value, regrd
— or —
%hi(value), regrd;
regrd, %lo(value), regrd

sethi

%hi(value), regrd

set
setsw

value,regrd
value,regrd

or
sethi

sethi
or
sra

regrd, %g0, regrd
— or —
%hi(value), regrd;
regrd, %lo(value), regrd
— or —
%hi(value), regrd;
regrd, %lo(value), regrd
regrd, %g0, regrd

sethi
or
sllx
sethi
or
or

%uhi(value), reg
reg, %ulo(value), reg
reg,32,reg
%hi(value), regrd
regrd, reg, regrd
regrd, %lo(value), regrd

sra
sethi
or

setx

value, reg, regrd

— or —
%g0, value, regrd
— or —
%hi(value), regrd

Comment
(when 0 ≤value≤ 4095)
(otherwise)
Warning: do not use setuw in
the delay slot of a DCTI.
synonym for setuw
(when (value> = 0) and
((value & 3FF16) = = 0))
(when -4096≤value≤ 4095)
(otherwise, if (value < 0) and
((value & 3FF16) = = 0))

(otherwise, if value> = 0)

(otherwise, if value<0)
Warning: do not use setsw in
the delay slot of a CTI.
create 64-bit constant
(“reg” is used as a temporary
register)
Note: setx optimizations are
possible, but not enumerated here. The worst-case is
shown.Warning: do not use
setx in the delay slot of a
CTI.

Table 37—Mapping Synthetic to SPARC-V9 Instructions (Continued)
Synthetic instruction

SPARC-V9 instruction(s)

signx
signx
not
not
neg
neg
cas
casl
casx
casxl

regrs1, regrd
regrd
regrs1, regrd
regrd
regrs2, regrd
regrd
[regrs1], regrs2, regrd
[regrs1], regrs2, regrd
[regrs1], regrs2, regrd
[regrs1], regrs2, regrd

sra
sra
xnor
xnor
sub
sub
casa
casa
casxa
casxa

regrs1, %g0, regrd
regrd, %g0, regrd
regrs1, %g0, regrd
regrd, %g0, regrd
%g0, regrs2, regrd
%g0, regrd, regrd
[regrs1]#ASI_P, regrs2, regrd
[regrs1]#ASI_P_L, regrs2, regrd
[regrs1]#ASI_P, regrs2, regrd
[regrs1]#ASI_P_L, regrs2, regrd

inc
inc
inccc
inccc
dec
dec
deccc
deccc
btst
bset
bclr
btog
clr
clrb
clrh
clr
clrx
clruw
clruw
mov
mov
mov
mov
mov

regrd
const13,regrd
regrd
const13,regrd
regrd
const13, regrd
regrd
const13, regrd
reg_or_imm, regrs1
reg_or_imm, regrd
reg_or_imm, regrd
reg_or_imm, regrd
regrd
[address]
[address]
[address]
[address]
regrs1, regrd
regrd
reg_or_imm, regrd
%y, regrd
%asrn, regrd
reg_or_imm, %y
reg_or_imm, %asrn

add
add
addcc
addcc
sub
sub
subcc
subcc
andcc
or
andn
xor
or
stb
sth
stw
stx
srl
srl
or
rd
rd
wr
wr

regrd, 1, regrd
regrd, const13, regrd
regrd, 1, regrd
regrd, const13, regrd
regrd, 1, regrd
regrd, const13, regrd
regrd, 1, regrd
regrd, const13, regrd
regrs1, reg_or_imm, %g0
regrd, reg_or_imm, regrd
regrd, reg_or_imm, regrd
regrd, reg_or_imm, regrd
%g0, %g0, regrd
%g0, [address]
%g0, [address]
%g0, [address]
%g0, [address]
regrs1, %g0, regrd
regrd, %g0, regrd
%g0, reg_or_imm, regrd
%y, regrd
%asrn, regrd
%g0, reg_or_imm, %y
%g0, reg_or_imm, %asrn

Comment
sign-extend 32-bit value to
64 bits
one’s complement
one’s complement
two’s complement
two’s complement
compare and swap
compare and swap, little-endian
compare and swap extended
compare and swap extended,
little-endian
increment by 1
increment by const13
incr by 1; set icc & xcc
incr by const13; set icc & xcc
decrement by 1
decrement by const13
decr by 1; set icc & xcc
decr by const13; set icc & xcc
bit test
bit set
bit clear
bit toggle
clear (zero) register
clear byte
clear halfword
clear word
clear extended word
copy and clear upper word
clear upper word

(This Annex is not a part of SPARC-V9/R1.4.5,
The SPARC Architecture Manual;
it is included for information only.)

This appendix is informative only.
It is not part of the SPARC-V9 specification.

H Software Considerations
This appendix describes how software can use the SPARC-V9 architecture effectively. Examples
do not necessarily conform to any specific Application Binary Interface (ABI).

H.1 Nonprivileged Software
This subsection describes software conventions that have proven or may prove useful, assumptions that compilers may make about the resources available, and how compilers can use those
resources. It does not discuss how supervisor software (an operating system) may use the architecture. Although a set of software conventions is described, software is free to use other conventions more appropriate for specific applications.
The following are the primary goals for many of the software conventions described in this subsection:
— Minimizing average procedure-call overhead
— Minimizing latency due to branches
— Minimizing latency due to memory access
H.1.1 Registers
Register usage is a critical resource allocation issue for compilers. The SPARC-V9 architecture
provides windowed integer registers (in, out, local), global integer registers, and floating-point
registers.
H.1.1.1 In and Out Registers
The in and out registers are used primarily for passing parameters to and receiving results from
subroutines, and for keeping track of the memory stack. When a procedure is called and executes
a SAVE instruction, the caller’s outs become the callee’s ins.
One of a procedure’s out registers (%o6) is used as its stack pointer, %sp. It points to an area in
which the system can store %r16..%r31 (%l0..%l7 and %i0..%i7) when the register file overflows (spill trap), and is used to address most values located on the stack. A trap can occur at any
time1, which may precipitate a subsequent spill trap. During this spill trap, the contents of the
user’s register window at the time of the original trap are spilled to the memory to which its %sp
points.

A procedure may store temporary values in its out registers (except %sp) with the understanding
that those values are volatile across procedure calls. %sp cannot be used for temporary values for
the reasons described in H.1.1.3, “Register Windows and %sp.”
Up to six parameters1 may be passed by placing them in out registers %o0..%o5; additional
parameters are passed in the memory stack. The stack pointer is implicitly passed in %o6, and a
CALL instruction places its own address in %o7.2 Floating-point parameters may also be passed
in floating-point registers.
After a callee is entered and its SAVE instruction has been executed, the caller’s out registers are
accessible as the callee’s in registers.
The caller’s stack pointer %sp (%o6) automatically becomes the current procedure’s frame
pointer %fp (%i6) when the SAVE instruction is executed.
The callee finds its first six integer parameters in %i0..%i5, and the remainder (if any) on the
stack.
A function returns a scalar integer value by writing it into its ins (which are the caller’s outs),
starting with %i0. A scalar floating-point value is returned in the floating-point registers, starting
with %f0.
A procedure’s return address, normally the address of the instruction just after the CALL’s delayslot instruction, is as %i7+8.3
H.1.1.2 Local Registers
The locals are used for automatic4 variables and for most temporary values. For access efficiency,
a compiler may also copy parameters (that is, those past the sixth) from the memory stack into the
locals and use them from there.
See H.1.4, “Register Allocation within a Window,” for methods of allocating more or fewer than
eight registers for local values.

For example, due to an error in executing an instruction (for example, a mem_address_not_aligned trap), or due
to any type of external interrupt.

Six is more than adequate, since the overwhelming majority of procedures in system code take fewer than six
parameters. According to studies cited by Weicker (Weicker, R. P., “Dhrystone: A Synthetic Systems Programming Benchmark,” CACM 27:10, October 1984), at least 97% (measured statically) take fewer than six parameters. The average number of parameters did not exceed 2.1, measured either statically or dynamically, in any of
these studies.

If a JMPL instruction is used in place of a CALL, it should place its address in %o7 for consistency.

For convenience, SPARC-V9 assemblers may provide a “ret” (return) synthetic instruction that generates a
“jmpl %i7+8, %g0” hardware instruction. See G.3, “Synthetic Instructions.”

In the C language, an automatic variable is a local variable whose lifetime is no longer than that of its containing
procedure.

H.1.1.3 Register Windows and %sp
Some caveats about the use of %sp and the SAVE and RESTORE instructions are appropriate. If
the operating system and user code use register windows, it is essential that
— %sp always contains a correct value, so that when (and if) a register window spill/fill trap
occurs, the register window can be correctly stored to or reloaded from memory.1
— Nonprivileged code uses SAVE and RESTORE instructions carefully. In particular, “walking” the call chain through the register windows using RESTOREs, expecting to be able to
return to where one started using SAVEs, does not work as one might suppose. Since user
code cannot disable traps, a trap (e.g., an interrupt) could write over the contents of a user
register window that has “temporarily” been RESTOREd2. The safe method is to flush the
register windows to user memory (the stack) by using the FLUSHW instruction. Then,
user code can safely “walk” the call chain through user memory, instead of through the
register windows.
To avoid such problems, consider all data memory at addresses just less than %sp to be volatile,
and the contents of all register windows “below” the current one to be volatile.
H.1.1.4 Global Registers
Unlike the ins, locals, and outs, the globals are not part of any register window. The globals are a
set of eight registers with global scope, like the register sets of more traditional processor architectures. An ABI may define conventions that the globals (except %g0) must obey. For example,
if the convention assumes that globals are volatile across procedure calls, either the caller or the
callee must take responsibility for saving and restoring their contents.
Global register %g0 has a hardwired value of zero; it always reads as zero, and writes to it have no
program-visible effect.
Typically, the global registers other than %g0 are used for temporaries, global variables, or global
pointers — either user variables, or values maintained as part of the program’s execution environment. For example, one could use globals in the execution environment by establishing a convention that global scalars are addressed via offsets from a global base register. In the most general
case, memory accessed at an arbitrary address requires six instructions; for example,
sethi
or
sllx
sethi
or
ld

%uhi(address),tmp
tmp, %ulo(address), tmp
tmp, 32, tmp
%hi(address), reg
reg, %lo(address), reg
[reg+tmp], reg

Typically, the SAVE instruction is used to generate a new %sp value while shifting to a new register window, all
in one atomic operation. When SAVE is used this way, synchronization of the two operations should not be a
problem.

Another reason this might fail is that user code has no way to determine how many register windows are implemented by the hardware.

Use of a global base register for frequently accessed global values would provide faster (singleinstruction) access to 213 bytes of those values; for example,
ld

[%gn+offset], reg

Additional global registers could be used to provide single-instruction access to correspondingly
more global values.
H.1.1.5 Floating-Point Registers
There are sixteen quad-precision floating-point registers. The registers can also be accessed as
thirty-two double-precision registers. In addition, the first eight quad registers can also be
accessed as thirty-two single-precision registers. Floating-point registers are accessed with different instructions than the integer registers; their contents can be moved among themselves, and to
or from memory. See 5.1.4, “Floating-Point Registers,” for more information about floating-point
register aliasing.
Like the global registers, the floating-point registers must be managed by software. Compilers use
the floating-point registers for user variables and compiler temporaries, pass floating-point parameters, and return floating-point results in them.
H.1.1.6 The Memory Stack
A stack is maintained to hold automatic variables, temporary variables, and return information for
each invocation of a procedure. When a procedure is called, a stack frame is allocated; it is
released when the procedure returns. The use of a stack for this purpose allows simple and efficient implementation of recursive procedures.
Under certain conditions, optimization can allow a leaf procedure to use its caller’s stack frame
instead of one of its own. In that case, the procedure allocates no space of its own for a stack
frame. See H.1.2, “Leaf-Procedure Optimization,” for more information.
The stack pointer %sp must always maintain the alignment required by the operating system’s
ABI. This is at least doubleword alignment, possibly with a constant offset to increase stack
addressability using constant offset addressing.
H.1.2 Leaf-Procedure Optimization
A leaf procedure is one that is a “leaf” in the program’s call graph; that is, one that does not call
(e.g., via CALL or JMPL) any other procedures.
Each procedure, including leaf procedures, normally uses a SAVE instruction to allocate a stack
frame and obtain a register window for itself, and a corresponding RESTORE instruction to deallocate it. The time costs associated with this are
— Possible generation of register-window spill/fill traps at runtime. This only happens occasionally,1 but when either a spill or fill trap does occur, it costs several machine cycles to
process.

— The cycles expended by the SAVE and RESTORE instructions themselves.
There are also space costs associated with this convention, the cumulative cache effects of which
may be nonnegligible. The space costs include
— The space occupied on the stack by the procedure’s stack frame
— The two words occupied by the SAVE and RESTORE instructions
Of the above costs, the trap-processing cycles typically are the most significant.
Some leaf procedures can be made to operate without their own register window or stack frame,
using their caller’s instead. This can be done when the candidate leaf procedure meets all of the
following conditions:1
— It contains no references to %sp, except in its SAVE instruction.
— It contains no references to %fp.
— It refers to (or can be made to refer to) no more than eight of the thirty-two integer registers, including %o7 (the return address).
If a procedure conforms to the above conditions, it can be made to operate using its caller’s stack
frame and registers, an optimization that saves both time and space. This optimization is called
leaf procedure optimization. The optimized procedure may safely use only registers that its
caller already assumes to be volatile across a procedure call.
The optimization can be performed at the assembly language level using the following steps:
(1) Change all references to registers in the procedure to registers that the caller assumes volatile across the call.
(a) Leave references to %o7 unchanged.
(b) Leave any references to %g0..%g7 unchanged.
(c) Change %i0..%i5 to %o0..%o5, respectively. If an in register is changed to an out
register that was already referenced in the original unoptimized version of the procedure, all original references to that out register must be changed to refer to an unused
out or global register.
(d) Change references to each local register into references to any unused register that is
assumed to be volatile across a procedure call.
(2) Delete the SAVE instruction. If it was in a delay slot, replace it with a NOP instruction. If
its destination register was not %g0 or %sp, convert the SAVE into the corresponding
ADD instruction instead of deleting it.

The frequency of overflow and underflow traps depends on the application and on the number of register windows (NWINDOWS) implemented in hardware.

Although slightly less restrictive conditions could be used, the optimization would become more complex to perform and the incremental gain would usually be small.

(3) If the RESTORE’s implicit addition operation is used for a productive purpose (such as
setting the procedure’s return value), convert the RESTORE to the corresponding ADD
instruction. Otherwise, the RESTORE is only used for stack and register-window deallocation; replace it with a NOP instruction (it is probably in the delay slot of the RET, and so
cannot be deleted).
(4) Change the RET (return) synthetic instruction to RETL (return-from-leaf-procedure synthetic instruction).
(5) Perform any optimizations newly made possible, such as combining instructions or filling
the delay slot of the RETL (or the delay slot occupied by the SAVE) with a productive
instruction.
After the above changes, there should be no SAVE or RESTORE instructions, and no references
to in or local registers in the procedure body. All original references to ins are now to outs. All
other register references are to registers that are assumed to be volatile across a procedure call.
Costs of optimizing leaf procedures in this way include
— Additional intelligence in a peephole optimizer to recognize and optimize candidate leaf
procedures
— Additional intelligence in debuggers to properly report the call chain and the stack traceback for optimized leaf procedures1
H.1.3 Example Code for a Procedure Call
This subsection illustrates common parameter-passing conventions and gives a simple example of
leaf-procedure optimization.
The code fragment in example 1 shows a simple procedure call with a value returned, and the procedure itself.
Since sum3 does not call any other procedures (i.e., it is a leaf procedure), it can be optimized to
become:
sum3:
add
retl
add

%o0, %o1, %o0
%o0, %o2, %o0

! (must use RETL, not RET,
! to return from leaf procedure)

H.1.4 Register Allocation within a Window
The usual SPARC-V9 software convention is to allocate eight registers (%l0 .. %l7) for local values. A compiler could allocate more registers for local values at the expense of having fewer outs
and ins available for argument passing. For example, if instead of assuming that the boundary
1.

A debugger can recognize an optimized leaf procedure by scanning it, noting the absence of a SAVE instruction.
Compilers often constrain the SAVE, if present, to appear within the first few instructions of a procedure; in such
a case, only those instruction positions need be examined.

! CALLER:
!
int i;
!
i = sum3( 1, 2, 3 );
...
mov
1, %o0
mov
2, %o1
call
sum3
mov
3, %o2
mov
%o0, %l7
...
#define SA(x) (((x)+15)&(~0x1F))
*/
#define MINFRAME ((16+1+6)*8)

/* compiler assigns "i" to register %l7 */

!
!
!
!
!

first arg to sum3 is 1
second arg to sum3 is 2
the call to sum3
last parameter to sum3 in delay slot
copy return value to %l7 (variable "i")

/* rounds "x" up to extended word boundary
/* minimum size stack frame, in bytes;
*
16 extended words for saving the

current
*
*
*
*
*/

! CALLEE:
!
int sum3( a, b, c )
!
int a, b, c;
/* args received in %i0, %i1, and %i2 */
!
{
!
return a+b+c;
!
}
sum3:
save
%sp,-SA(MINFRAME),%sp! set up new %sp; alloc min. stack frame
add
%i0, %i1, %l7
! compute sum in local %l7
add
%l7, %i2, %l7
!
(or %i0 could have been used directly)
ret
! return from sum3, and...
restore
%l7, 0, %o0
! move result into output reg & restore

Example 1—Simple Procedure Call with Value Returned

between local values and input arguments is between r[23] and r[24] (%l7 and %i0), software
could, by convention, assume that the boundary is between r[25] and r[26] (%i1 and %i2). This
would provide ten registers for local values and six in and out registers. This is shown in table 38.
Table 38—Register Allocation within a Window

Registers for local values
In / out registers
Reserved for %sp / %fp
Reserved for return address
Available for argument passing
Total ins / outs

Standard
register
model
8

10 local
register
model
10

Arbitrary
register
model
n

1
1
6
8

1
1
4
6

1
1
14 − n
16 − n

H.1.5 Other Register-Window-Usage Models
So far, this appendix has described SPARC-V9 software conventions that are appropriate for use
in a general-purpose multitasking computer system. However, SPARC-V9 is used in many other
applications, notably embedded and/or real-time systems. In such applications, other schemes for
allocation of SPARC-V9’s register windows might be more nearly optimal than the one described
above.
One possibility is to avoid using the normal register-window mechanism by not using SAVE and
RESTORE instructions. Software would see 32 general-purpose registers instead of SPARC-V9’s
usual windowed register file. In this mode, SPARC-V9 would operate like processors with more
traditional (flat) register architectures. Procedure call times would be more determinate (due to
lack of spill/fill traps), but for most types of software, average procedure call time would significantly increase, due to increased memory traffic for parameter passing and saving/restoring local
variables.
Effective use of this software convention would require compilers to generate different code
(direct register spills/fills to memory and no SAVE/RESTORE instructions) than for the software
conventions described above.
It would be awkward, at best, to attempt to mix (link) code that uses the SAVE/RESTORE convention with code that does not use it. If both conventions were used in the same system, two versions of each library would be required.
It would be possible to run user code with one register-usage convention and supervisor code with
another. With sufficient intelligence in supervisor software, user processes with different register
conventions could be run simultaneously.1
H.1.6 Self-Modifying Code
If a program includes self-modifying code, it must issue a FLUSH instruction for each modified
doubleword of instructions (or a call to supervisor software having an equivalent effect).
Note that self-modifying code intended to be portable must use FLUSH instruction(s) (or a call to
supervisor software having equivalent effect) after storing into the instruction stream.
All SPARC-V9 instruction accesses are big-endian. If a program is running in little-endian mode
and wishes to modify instructions, it must do one of the following:
— Use an explicit big-endian ASI to write the modified instruction to memory, or
— Reverse the byte ordering shown in the instruction formats in Appendix A, “Instruction
Definitions,” before doing a little-endian store, since the stored data will be reordered
before the bytes are written to memory.

Although technically possible, this is not to suggest that there would be significant utility in mixing user processes with differing register-usage conventions.

H.1.7 Thread Management
SPARC-V9 provides support for the efficient management of user-level threads. The cost of
thread switching can be reduced by using the following features:
User Management of FPU:
The FEF bit in the FPRS register allows nonprivileged code to manage the FPU. This is in
addition to the management done by the supervisor code via the PEF bit in the PSTATE
register. A thread-management library can implement efficient switching of the FPU
among threads by manipulating the FEF bit in the FPRS register and by providing a user
trap handler (with support from the supervisor software) for the fp_disabled exception. See
the description of User Traps in H.2.4, “User Trap Handlers.”
FLUSHW Instruction:
The FLUSHW instruction is an efficient way for a thread library to flush the register windows during a thread switch. The instruction executes as a NOP if there are no windows to
flush.
H.1.8 Minimizing Branch Latency
The SPARC-V9 architecture contains several instructions that can be used to minimize branch
latency. These are described below.
Conditional Moves:
The conditional move instructions for both integer and floating-point registers can be used
to eliminate branches from the code generated for simple expressions and/or assignments.
The following example illustrates this.
The C code segment
double
x,y;
int
i;
...
i = (x > y) ? 1 : 2;

can be compiled to use a conditional move as follows:
fcmp
mov
movfle

%fcc1, x, y
1, i
%fcc1, 2, i

! x and y are double regs
! i is int; assume x > y
! fix i if wrong

Branch or Move Based on Register Contents:
The use of register contents as conditions for branch and move instructions allows any
integer register (other than r0) to hold a boolean value or the results of a comparison. This
allows conditions to be used more efficiently in nested cases. It allows the generation of a
condition to be moved further from its use, thereby minimizing latency. In addition, it can
eliminate the need for additional arithmetic instructions to set the condition codes. This is
illustrated in the following example.
The test for finding the maximum of an array of integers,
if (A[i] > max)
max = A[i];

can be compiled as follows, allowing the condition for the loop to be set before the
sequence and checked after it:
ldx
sub
movrgz

[addr_of_Ai], Ai
Ai, max, tmp
tmp, Ai, max

H.1.9 Prefetch
The SPARC-V9 architecture includes a prefetch instruction intended to help hide the latency of
accessing memory.1
As a general rule, given a loop of the following form (using C for assembly language, and assuming a cache line size of 64 bytes and that A and B are arrays of 8-byte values)
for (i = 0; i < N; i++) {
load A[i]
load B[i]
...
}

which takes C cycles per iteration (assuming all loads hit in cache) and given L cycles of latency
to memory, prefetch instructions may be inserted for data that will be needed ceiling(L/C') iterations in the future, where C' is number of cycles per iteration of the modified loop. Thus, the loop
would be transformed into
K = ceiling(L/C');
for (i = 0; i < N; i++) {
load A[i]
load B[i]
prefetch A[i+K]
prefetch B[i+K]
...
}

This ensures that the loads will find their data in the cache, and will thus complete more quickly.
The first K iterations will not get any benefit from prefetching, so if the number of iterations is
small (see below), then prefetching will not help.

Two papers describing the use of prefetch instructions are Callahan, D., K. Kennedy, A. Porterfield, “Software
Prefetching,” Proceedings of the Fourth International Conference on Architectural Support for Programming
Languages and Operating Systems, April 1991, pp. 40-52, and Mowry, T., M. Lam, and A. Gupta, “Design and
Evaluation of a Compiler Algorithm for Prefetching,” Proceedings of the Fifth International Conference on
Architectural Support for Programming Languages and Operating Systems, October 1992, pp. 62-73.

Note that in cases of contiguous access (like this one), many of the prefetch instructions will in
fact be unnecessary and may slow the program down. To avoid this, note that the prefetch instruction always obtains at least 64 (cache-line-aligned) bytes.
/* Round up access to next cache line. */
K' = (ceiling(L/C') + 7) & ~7;
for (i

= 0; i < N; i++) {

load A[i]
load B[i]
if ( ((int)(A+i) & 63)

= 0) {

prefetch A[i+K']
prefetch B[i+K']
}
...
}

or (unrolled eight times, assuming A and B are arrays of 8-byte values)
/* Be sure that we access the next cache line. */
K'' = ceiling(L/C') + 7;
for (i

= 0; i < N; i++) {

load A[i]
load B[i]
prefetch A[i+K'']
prefetch B[i+K'']
...
load A[i+1]
load B[i+2]
... (no prefetching)
...
load A[i+7]
load B[i+7]
...
}

In the first case, the prefetching is performed exactly when needed, and thus the distance need not
be adjusted. However, the prefetching may not start on the first iteration, resulting in as many as
K' + 7 iterations without prefetching.
In the second case, the prefetching occurs somewhere within a cache line, and thus, it is not
known exactly how long it will be until the next cache line is needed. However, by prefetching
seven further ahead, we ensure that the next cache line will be prefetched soon enough. In the
worst case, as many as K'' (≤ K' + 7) iterations will execute without any benefit from prefetching.
Table 39 illustrates the cost tradeoffs between no prefetching, naive prefetching, and smart
prefetching (the second choice) for a small loop (two cycles) with varying uncovered latencies to

memory. Some of the latency may be overlapped with execution of surrounding instructions; that
which is not is uncovered.
Table 39—Prefetch Cost Tradeoffs

C
2
2
2

C'
4
4
4

L
8
16
32

K
4
8
16

K''
11
15
23

Limit cycles/iteration
No pf
Naive
Smart
C+L/8
C'
(7C+C')/8
3
4
2.25
4
4
2.25
6
4
2.25

Smart startup costs
Worst
Worst
Misses
Breakeven
2
N = 21
2
N = 18
3
N = 26

Here, we treat the arrays accessed as if one were not in the cache. Thus, every eight iterations, a
cache line must be fetched from memory in the no-prefetch case; and thus, the amortized cost of
an iteration is C + L/8. The cost estimate for the smart case ignores any benefits from unrolling,
since it is reasonable to expect that the loop would be unrolled or pipelined in this fashion, even if
prefetching were not used. The startup costs assume an alignment within the cache that maximizes the initial misses. The break-even cost was chosen by solving the following equation for N.
N ∗ (C + L/8) = WM ∗ L + N ∗ (7C + C')/8 {e.g., 3N = 16 + 2.25N ⇒ N = 21}
Of course, this is a simplified model.
Another possibility to consider is the worst-case cost of prefetching. If, in the example provided,
everything accessed is always cached, then the smart-prefetching loop takes 12.5% longer. For
each memory latency, there is a break-even point (in terms of how often one of the array operands
is cached) at which the prefetching loop begins to run faster. Table 40 illustrates this.
Table 40—Cache Break-Even Points

L
8
16
32
64

C-cached
2
2
2
2

C-missed
3
4
6
10

C-smart
2.25
2.25
2.25
2.25

Break-even
% cached
operands
75%
88%
94%
97%

Break-even
loop cache miss
rate
1.56%
0.75%
0.375%
0.188%

Note that one uncached operand corresponds to one load out of sixteen missing the cache; the
operand miss rate is sixteen times higher than the load miss rate. Note that this is the miss rate for
this loop alone; extrapolation from whole-program miss rates is not advised.
Binaries that run efficiently across different SPARC-V9 implementations can be created for cases
like this (where memory accesses are regular, though not necessarily contiguous) by parameterizing the prefetch distance by machine type. In privileged code the machine type is available in the
VER register; nonprivileged code should be able to obtain this from the operating system or ABI.
Based on information about known machines and estimated loop execution times, a compiler
could precalculate values for K'' (assuming smart prefetching) and store them in a table. At execution time, the proper value for K'' would be fetched from the table before entering the loop.

For regular but noncontiguous accesses, a prefetch would be issued for every load. If cache blocking is used, the prefetching strategy must be adjusted accordingly, since there is no point in
prefetching data that is expected to be in the cache already.
The prefetch variant should be chosen based on what is known about the local and global use of
the data prefetched. If the data is not being written locally, then variant 0 (several reads) should be
used. If it is being written (and possibly also read), then variant 2 (several writes) should be used.
If, in addition, it is known that this is likely to be the last use of the data for some time (for example, if the loop iteration count is one million and dependence analysis reveals no reuse of data),
then it is appropriate to use either variant 1 (one read) or 3 (one write). If reuse of data is expected
to occur soon, then use of variants 1 or 3 is not appropriate, because of the risk of increased bus
and memory traffic on a multiprocessor.
If the hardware does not implement all variants, it is expected to provide a sensible overloading of
the unimplemented variants. Thus, correct use of a specific variant need not be tied to a particular
SPARC-V9 implementation or multi/uniprocessor configuration.
H.1.10 Nonfaulting Load
The SPARC-V9 architecture includes a way to specify load instructions that do not generate visible faults, so that compilers can have more freedom in scheduling instructions. Note that these are
not speculative loads, which may fault if their results are later used; these are normal load instructions, but tagged to indicate to the kernel and/or hardware that a fault should not be delivered to
the code executing the instruction.
Five important rules govern the use of nonfaulting loads:
(1) Volatile memory references in the source language should not use nonfaulting load
instructions.
(2) Code compiled for debugging should not use nonfaulting loads, because they remove the
ability to detect common errors.
(3) If nonfaulting loads are used, page zero should be a page of zero values, mapped readonly. Compilers that routinely use negative offsets to register pointers should map page “–
1” similarly, if the operating software permits it.
(4) Any use of nonfaulting loads in privileged code must be aware of how they are treated by
the host SPARC-V9 implementation.
(5) Nonfaulting loads from unaligned addresses may be substantially more expensive than
nonfaulting loads from other addresses.
Nonfaulting loads can be used to solve three scheduling problems.
— On super-scalar machines, it is often desirable to obtain the right mix of instructions to
avoid conflicts for any given execution unit. A nonfaulting load can be moved (backwards)
past a basic block boundary to even out the instruction mix.

— On pipelined machines, there may be latency between loads and uses. A nonfaulting load
can be moved past a block boundary to place more instructions between a load into a register and the next use of that register.
— Software pipelining improves the scheduling of loops, but if a loop iteration begins with a
load instruction and contains an early exit, it may not be eligible for pipelining. If the load
is replaced with a nonfaulting load, then the loop can be pipelined.
In the branch-laden code shown in example 2, nonfaulting loads could be used to separate loads
from uses. The result also has a somewhat better mix of instructions and is somewhat pipelined.
The basic blocks are separated.
Source Code:
while ( x ! = 0 && x -> key ! = goal) x

= x -> next;

With Normal Loads:
entry:
brnz,a

x,loop

[x],t1

! (pre)load1 (key)

cmp

t1,goal

! use1

bpe

%xcc,out

ldx
loop:

nop

! no filling from loop.

ldx

[x+8],x

! load2 (next)

brnz,a

x,loop

! use2

ldx

[x],t1

! load1

out: ...

With Nonfaulting Loads:
entry:
mov
mov
ldxa
loop:
mov
brz,pn
ldxa
cmp
bpne
ldxa
out: ...

x,t2
#ASI_PNF, %asi
[t2]%asi,t1
! (pre)load1 (nf-load for key)
t2,x
t2,out
[t2+8]%asi,t2
t1,goal
%xcc,loop
[t2],%asi,t1

! begin loop body
! load2 (nf-load for next)
! use1
! use2, load1 ! nf-load for x

Example 2—Branch-Laden Code with Nonfaulting Loads

In the loop shown in example 3, nonfaulting loads allow pipelining. This loop might be improved
further using unrolling, prefetching, and multiple FCCs, but that is beyond the scope of this discussion.

Source Code:
d_ne_index (double * d1, double * d2) {
int i = 0;
while(d1[i]

= d2[i]) i++;

return i;
}

With Normal Loads:
mov
mov
loop:
lddf
lddf
add
fcmpd
fbe,a
add

0,t
0,i
[d1+t],a1
[d2+t],a2
t,8,t
a1,a2
loop
i,1,i

! load
! use
! fcc use

With Nonfaulting Loads:
lddf
lddf
mov
mov
loop:
fcmpd
lddfa
lddfa
add
fbe,a
add

[d1],a1
[d2],a2
8,t
0,i
a1,a2
! use, fcc def
[d1+t],%asi,a1
[d2+t],%asi,a2 ! load
t,8,t
loop
! fcc use
i,1,i

Example 3—Loop with Nonfaulting Loads

H.2 Supervisor Software
This subsection discusses how supervisor software can use the SPARC-V9 privileged architecture. It is intended to illustrate how the architecture can be used in an efficient manner. An implementation may choose to utilize different strategies based on its requirements and
implementation-specific aspects of the architecture.
H.2.1 Trap Handling
The SPARC-V9 privileged architecture provides support for efficient trap handling, especially for
window traps. The following features of the SPARC-V9 privileged architecture can be used to
write efficient trap handlers:
Multiple Trap Levels:
The trap handlers for trap levels less than MAXTL – 1 can be written to ignore exceptional conditions and execute the common case efficiently (without checks and branches).
For example, the fill/spill handlers can access pageable memory without first checking if it

is resident. If the memory is not resident, the access will cause a trap that will be handled
at the next trap level.
Vectoring of Fill/Spill Traps:
Supervisor software can set up the vectoring of fill/spill traps prior to executing code that
uses register windows and may cause spill/fill traps. This feature can be used to support
SPARC-V8 and SPARC-V7 binaries. These binaries create stack frames with save areas
for 32-bit registers. SPARC-V9 binaries create stack frames with save areas for 64-bit registers. By setting up the spill/fill trap vector based on the type of binary being executed, the
trap handlers can avoid checking and branching to use the appropriate load/store instructions.
Saved Trap State:
Trap handlers need not save (restore) processor state that is automatically saved (restored)
on a trap (return from trap). For example, the fill/spill trap handlers can load
ASI_AS_IF_USER_PRIMARY{_LITTLE} into the ASI register in order to access the
user’s address space without the overhead of having to save and restore the ASI register.
SAVED and RESTORED Instructions:
The SAVED (RESTORED) instruction provides an efficient way to update the state of the
register windows after successfully spilling (filling) a register window. They implement a
default policy of spilling (filling) one register window at a time. If desired, the supervisor
software can implement a different policy by directly updating the state registers.
Alternate Globals:
The alternate global registers can be used to avoid saving and restoring the normal global
registers. They can be used like the local registers of the trap window in SPARC-V8.
Large Trap Vectors for Spill/Fill:
The definition of the spill and fill trap vectors with reserved space between each pair of
vectors allows spill and fill trap handlers to be up to thirty-two instructions long, thus
avoiding a branch in the handler.
H.2.2 Example Code for Spill Handler
The code in example 4 shows a spill handler for a SPARC-V9 user binary. The handler is located
at the vector for trap type spill_0_normal (08016). It is assumed that supervisor software has set the
WSTATE register to 0 before executing the user binary. The handler is invoked when user code
executes a SAVE instruction that results in a window overflow.
H.2.3 Client-Server Model
SPARC-V9 provides mechanisms to support client-server computing efficiently. A call from a client to a server (where the client and server have separate address spaces) can be implemented efficiently using a software trap that switches the address space. This is often referred to as a crossdomain call. A system call in most operating systems can be viewed as a special case of a crossdomain call. The following features are useful in implementing a cross-domain call:

T_NORMAL_SPILL_0:
!Set ASI to access user addr space
wr
#ASI_AIUP, %asi
stxa
%l0, [%sp+(8* 0)]%asi
!Store window in memory stack
stxa
%l1, [%sp+(8* 1)]%asi
stxa
%l2, [%sp+(8* 2)]%asi
stxa
%l3, [%sp+(8* 3)]%asi
stxa
%l4, [%sp+(8* 4)]%asi
stxa
%l5, [%sp+(8* 5)]%asi
stxa
%l6, [%sp+(8* 6)]%asi
stxa
%l7, [%sp+(8* 7)]%asi
stxa
%i0, [%sp+(8* 8)]%asi
stxa
%i1, [%sp+(8* 9)]%asi
stxa
%i2, [%sp+(8*10)]%asi
stxa
%i3, [%sp+(8*11)]%asi
stxa
%i4, [%sp+(8*12)]%asi
stxa
%i5, [%sp+(8*13)]%asi
stxa
%i6, [%sp+(8*14)]%asi
stxa
%i7, [%sp+(8*15)]%asi
saved
! Update state
retry
! Retry trapped instruction
! Restores old %asi

Example 4—Spill Handler

Splitting the Register Windows
The register windows can be shared efficiently between multiple address spaces by using the
OTHERWIN register and providing additional trap handlers to handle spill/fill traps for the other
(not the current) address spaces. On a cross-domain call (a software trap), the supervisor can set
the OTHERWIN register to the number of register windows used by the client (equal to CANRESTORE) and CANRESTORE to zero. At the same time the WSTATE bit vectors can be set to
vector the spill/fill traps appropriately for each address space.
The sequence in example 5 shows a cross-domain call and return. The example assumes the simple case, where only a single client-server pair can occupy the register windows. More general
schemes can be developed along the same lines.
ASI_SECONDARY{_LITTLE}
Supervisor software can use these unrestricted ASIs to support cross-address-space access
between clients and nonprivileged servers. For example, some services that are currently provided
as part of a large monolithic supervisor can be separated out as nonprivileged servers (potentially
occupying a separate address space). This is often referred to as the microkernel approach.
H.2.4 User Trap Handlers
Supervisor software can provide efficient support for user (nonprivileged) trap handlers on
SPARC-V9. The RETURN instruction allows nonprivileged code to retry an instruction pointed
to by the previous stack frame. This provides the semantics required for returning from a user trap
handler without any change in processor state. Supervisor software can invoke the user trap han-

cross_domain_call:
save
! create a new register window for the server
..
! Switch to the execution environment for the server;
..
! Save trap state as necessary.
! Set CWP for caller in TSTATE
rdpr
%tstate, %g1
rdpr
%cwp, %g2
bclr
TSTATE_CWP, %g1
wrpr
%g1, %g2, %tstate
rdpr
%canrestore, %g1
wrpr
%g0, 0, %canrestore
wrpr
%g0, %g1, %otherwin
rdpr
%wstate, %g1
sll
%g1, 3, %g1
or

%g1, WSTATE_SERVER, %g1

wrpr
..
done

%g0, %g1, %wstate

! Move WSTATE_NORMAL (client’s
! vector)to WSTATE_OTHER
! Set WSTATE_NORMAL to the
!
vector for the server
! Load trap state for server
! Execute server code

cross_domain_return:
rdpr
%otherwin, %g1
wrpr
%g0, %g1, %canrestore
wrpr
%g0, 0, %otherwin
rdpr
%wstate, %g1
srl
%g1, 3, %g1
wrpr
%g0, %g1, %wstate
..
restore

! Reset WSTATE_NORMAL to
! client’s vector
! Restore saved trap state as necessary; this includes
! the return PC for the caller.
! Go back to the caller’s register window.

! Set CWP for caller in TSTATE
rdpr
%tstate, %g1
rdpr
%cwp, %g2
bclr
TSTATE_CWP, %g1
wrpr
%g1, %g2, %tstate
done

! return to the caller

Example 5—Cross-Domain Call and Return

dler by first creating a new register window (and stack frame) on its behalf and passing the necessary arguments (including the PC and nPC for the trapped instruction) in the local registers. The
code in example 6 shows how a user trap handler may be invoked and how it returns:

(This Annex is not a part of SPARC-V9/R1.4.5,
The SPARC Architecture Manual;
it is included for information only.)

T_EXAMPLE_TRAP:
save

! Supervisor trap handler for T_EXAMPLE_TRAP trap
! Create a window for the user trap handler

!Set CWP for new window in TSTATE
rdpr
%tstate, %l6
rdpr
%cwp, %l5
bclr
TSTATE_CWP, %l6
wrpr
%l6, %l5, %tstate
rdpr
rdpr
..

%tpc,%l6 !Put PC for trapped instruction in local register
%tnpc,%l7 !Put nPC for trapped instruction in local register
!Get the address of the user trap handler in %l5;
! for example, from a supervisor data structure.

wrpr
done

%l5, %tnpc

USER_EXAMPLE_TRAP:
..

! Put PC for user trap handler in %tnpc.
! Execute user trap handler.
!User trap handler for T_EXAMPLE_TRAP trap
!Execute trap handler logic. Local registers
! can be used as scratch.

jmpl
return

%l6
%l7

!Return to retry the trapped instruction.

Example 6—User Trap Handler

This appendix is informative only.
It is not part of the SPARC-V9 specification.

Extending the SPARC-V9 Architecture

This appendix describes how extensions can be effectively added to the SPARC-V9 architecture.
It describes how new instructions can be added through the use of read and write ancillary state
register (ASR) and implementation-dependent (IMPDEP1/IMPDEP2) instructions.
— WARNING —
Programs that make use of SPARC-V9 architectural extensions
may not be portable and likely will not conform to any current or
future SPARC-V9 binary standards.

I.1 Addition of SPARC-V9 Extensions
There are two approved methods of adding extensions to an implementation of the SPARC-V9
architecture. An implementor who wishes to define and implement a new SPARC-V9 instruction
should, if possible, use one of the following methods.

I.1.1 Read/Write Ancillary State Registers (ASRs)
The first method of adding instructions to SPARC-V9 is through the use of the implementationdependent Write Ancillary State Register (WRASR) and Read Ancillary State Register (RDASR)
instructions operating on ASRs 16..31. Through a read/write instruction pair, any instruction that
requires an rs1, reg_or_imm, and rd field can be implemented. A WRASR instruction can also
perform an arbitrary operation on two register sources, or on one register source and a signed
immediate value, and place the result in an ASR. A subsequent RDASR instruction can read the
result ASR and place its value in an integer destination register.

I.1.2 Implementation-Dependent and Reserved Opcodes
The meaning of “reserved” for SPARC-V9 opcodes differs from its meaning in SPARC-V8. The
SPARC-V9 definition of “reserved” allows implementations to use reserved opcodes for implementation-specific purposes. While a hardware implementation that uses reserved opcodes will be
SPARC-V9-compliant, SPARC-V9 ABI-compliant programs cannot use these reserved opcodes
and remain compliant. A SPARC-V9 platform that implements instructions using reserved
opcodes must provide software libraries that provide the interface between SPARC-V9 ABI-compliant programs and these instructions. Graphics libraries provide a good example of this. Hardware platforms have many diverse implementations of graphics acceleration hardware, but
graphics application programs are insulated from this diversity through libraries.
There is no guarantee that a reserved opcode will not be used for additional instructions in a future
version of the SPARC architecture. Implementors who use reserved opcodes should keep this in
mind.
In some cases forward compatibility may not be an issue; for example, in an embedded application, binary compatibility may not be an issue. These implementations can use any reserved
opcodes for extensions.
Even when forward compatibility is an issue, future SPARC revisions are likely to contain few
changes to opcode assignments, given that backward compatibility with previous versions must
be maintained. It is recommended that implementations wishing to remain forward-compatible
use the new IMPDEP1 and IMPDEP2 reserved opcodes with op3[5:0] = 11 01102 and 11 01112.
Compatibility Note:
IMPDEP1 and IMPDEP2 replace the SPARC-V8 CPop1 and CPop2 opcodes. SPARC-V9 includes neither
the SPARC-V8 coprocessor opcodes nor any other SPARC-V8 architectural support for coprocessors. The
coprocessor opcodes were eliminated because they have not been used in SPARC-V7 and SPARC-V8, as
witnessed by the lack of coprocessor implementations.

It is further recommended that SPARC International be notified of any use of IMPDEP1,
IMPDEP2, or other reserved opcodes. When and if future revisions to SPARC are contemplated,
and if any SPARC-V9 implementations have made use of reserved opcodes, SPARC International
will make every effort not to use those opcodes. By going through SPARC International, there can
be feedback and coordination in the choice of opcodes that maximizes the probability of forward
compatibility. Given the historically small number of implementation-specific changes, coordinating through SPARC International should be sufficient to ensure future compatibility.

This appendix is informative only.
It is not part of the SPARC-V9 specification.

(This Annex is not a part of SPARC-V9/R1.4.5,
The SPARC Architecture Manual;
it is included for information only.)

J Programming With the Memory Models
This appendix describes how to program with the SPARC-V9 memory models. An intuitive
description of the models is provided in Chapter 8, “Memory Models.” A complete formal specification appears in Appendix D, “Formal Specification of the Memory Models.” In this subsection, general programming guidelines are given first, followed by specific examples showing how
low-level synchronization can be implemented in TSO, PSO, and RMO.
Note that code written for a weaker memory model will execute correctly in any of the stronger
memory models. Furthermore, the only possible difference between code written for a weaker
memory model and the corresponding code for a stronger memory model is the presence of memory ordering instructions (MEMBARs) that are not needed for the stronger memory model.
Hence, transforming code from/to a stronger memory model to/from a weaker memory model
means adding/removing certain memory ordering instructions.1 The required memory ordering
directives are monotonically ordered with respect to the strength of the memory model, with the
weakest memory model requiring the strongest memory ordering instructions.
The code examples given below are written to run correctly using the RMO memory model. The
comments on the MEMBAR instructions indicate which ordering constraints (if any) are required
for the PSO and TSO memory models.

J.1 Memory Operations
Programs access memory via five types of operations, namely, load, store, LDSTUB, SWAP, and
compare-and-swap. Load copies a value from memory or an I/O location to a register. Store copies a value from a register into memory or an I/O location. LDSTUB, SWAP, and compare-andswap are atomic load-store instructions that store a value into memory or an I/O location and
return the old value in a register. The value written by the atomic instructions depends on the
instruction. LDSTUB stores all ones in the accessed byte, SWAP stores the supplied value, and

1. MEMBAR instructions specify seven independent ordering constraints; thus, there are cases where the transition
to a stronger memory model allows the use of a less restrictive MEMBAR instruction, but still requires a MEMBAR instruction. To demonstrate this property, the code examples given in this subsection use multiple MEMBAR
instructions if some of the ordering constraints are needed in some but not all memory models. Multiple, adjacent
MEMBAR instructions can always be replaced with a single MEMBAR instruction by ORing the arguments.

compare-and-swap stores the supplied value only if the old value equals the second supplied
value.
Memory order and consistency are controlled by MEMBAR instructions. For example, a MEMBAR #StoreStore (equivalent to a STBAR in SPARC-V8) ensures that all previous stores
have been performed before subsequent stores and atomic load-stores are executed by memory.
This particular memory order is guaranteed implicitly in TSO, but PSO and RMO require this
instruction if the correctness of a program depends on the order in which two store instructions
can be observed by another processor.1
FLUSH is not a memory operation, but it is relevant here in the context of synchronizing stores
with instruction execution. When a processor modifies an instruction at address A, it does a store
to A followed by a FLUSH A. The FLUSH ensures that the change made by the store will become
visible to the instruction fetch units of all processors in the system.

J.2 Memory Model Selection
Given that all SPARC-V9 systems are required to support TSO, programs written for any memory
model will be able to run on any SPARC-V9 system. However, a system running with the TSO
model generally will offer lower performance than PSO or RMO, because less concurrency is
exposed to the CPU and the memory system. The motivation for weakening the memory model is
to allow the CPU to issue multiple, concurrent memory references in order to hide memory
latency and increase access bandwidth. For example, PSO and RMO allow the CPU to initiate
new store operations before an outstanding store has completed.
Using a weaker memory model for an MP (multiprocessor) application that exhibits a high degree
of read-write memory sharing with fine granularity and a high frequency of synchronization operations may result in frequent MEMBAR instructions.
In general, it is expected that the weaker memory models offer a performance advantage for multiprocessor SPARC-V9 implementations.

J.3 Processors and Processes
In the SPARC-V9 memory models, the term “processor” may be replaced systematically by the
term “process” or “thread,” as long as the code for switching processes or threads is written properly. The correct process-switch sequence is given in J.8, “Process Switch Sequence.” If an operating system implements this process-switch sequence, application programmers may completely
ignore the difference between a process/thread and a processor.

1. Memory order is of concern only to programs containing multiple threads that share writable memory and that
may run on multiple processors, and to those programs which reference I/O locations. Note that from the processor’s point of view, I/O devices behave like other processors.

J.4 Higher-Level Programming Languages and Memory Models
The SPARC-V9 memory models are defined at the machine instruction level. Special attention is
required to write the critical parts of MP/MT (multi-threaded) applications in a higher-level language. Current higher-level languages do not support memory ordering instructions and atomic
operations. As a result, MP/MT applications that are written in a higher-level language generally
will rely on a library of MP/MT support functions, for example, the parmacs library from
Argonne National Laboratory.1 The details of constructing and using such libraries are beyond the
scope of this document.
Compiler optimizations such as code motion and instruction scheduling generally do not preserve
the order in which memory is accessed but they do preserve the data dependencies of a single
thread. Compilers do not, in general, deal with the additional dependency requirements to support
sharing read-write data among multiple concurrent threads. Hence, the memory semantics of a
SPARC-V9 system in general are not preserved by optimizing compilers. For this reason, and
because memory ordering directives are not available from higher-level languages, the examples
presented in this subsection use assembly language.
Future compilers may have the ability to present the programmer with a sequentially consistent
memory model despite the underlying hardware’s providing a weaker memory model.2

J.5 Portability And Recommended Programming Style
Whether a program is portable across various memory models depends on how it synchronizes
access to shared read-write data. Two aspects of a program’s style are relevant to portability:
— Good semantics refers to whether the synchronization primitives chosen and the way in
which they are used are such that changing the memory model does not involve making
any changes to the code that uses the primitives.
— Good structure refers to whether the code for synchronization is encapsulated through
the use of primitives such that when the memory model is changed, required changes to
the code are confined to the primitives.
Good semantics are a prerequisite for portability, while good structure makes porting easier.
Programs that use single-writer/multiple-reader locks to protect all access to shared read-write
data are portable across all memory models. The code that implements the lock primitives themselves is portable across all models only if it is written to run correctly on RMO. If the lock primitives are collected into a library, then, at worst, only the library routines must be changed. Note
that mutual exclusion (mutex) locks are a degenerate type of single-writer/multiple-readers lock.
Programs that use write locks to protect write accesses but read without locking are portable
across all memory models only if writes to shared data are separated by MEMBAR #Store1. Lusk, E. L., R.A. Overbeek, “Use of Monitors in Fortran: A Tutorial on the Barrier, Self-scheduling Do-Loop, and
Askfor Monitors,” TR# ANL-84-51, Argonne National Laboratory, June 1987.
2. See Gharachorloo, K., S.V. Adve, A. Gupta, J.L. Hennessy, and M.D. Hill, “Programming for Different Memory
Consistency Models,” Journal of Parallel and Distributed Systems, 15:4, August 1992.

Store instructions, and if reading the lock is followed by a MEMBAR #LoadLoad instruction. If the MEMBAR instructions are omitted, the code is portable only across TSO and Strong
Consistency,1 but generally it will not work with PSO and RMO. The code that implements the
lock primitives is portable across all models only if it is written to run correctly on RMO. If the
lock routines are collected into a library, the only possible changes not confined to the library routines are the MEMBAR instructions.
Programs that do synchronization without using single-writer/multiple-reader locks, write locks,
or their equivalent are, in general, not portable across different memory models. More precisely,
the memory models are ordered from RMO (which is the weakest, least constrained, and most
concurrent), PSO, TSO, to sequentially consistent (which is the strongest, most constrained, and
least concurrent). A program written to run correctly for any particular memory model will also
run correctly in any of the stronger memory models, but not vice versa. Thus, programs written
for RMO are the most portable, those written for TSO are less so, and those written for strong
consistency are the least portable. This general relationship between the memory models is shown
graphically in figure 49.
Strong Consistency
TSO
PSO

RMO

Figure 49—Portability Relations among Memory Models

The style recommendations may be summarized as follows: Programs should use single-writer/
multiple-reader locks, or their equivalent, when possible. Other lower-level forms of synchronization (such as Dekker’s algorithm for locking) should be avoided when possible. When use of such
low-level primitives is unavoidable, it is recommended that the code be written to work on the
RMO model to ensure portability. Additionally, lock primitives should be collected together into a
library and written for RMO to ensure portability.

1. Programs that assume a sequentially consistent memory are not guaranteed to run correctly on any SPARC-V9compliant system, since TSO is the strongest memory model required to be supported. However, sequential consistency is the most natural and intuitive programming model. This motivates the development of compiler techniques that allow programs written for sequential consistency to be translated into code that runs correctly (and
efficiently) on systems with weaker memory models.

Appendix D, “Formal Specification of the Memory Models,” describes a tool and method that
allows short code sequences to be formally verified for correctness.

J.6 Spin Locks
A spin lock is a lock for which the “lock held” condition is handled by busy waiting. The code in
example 7 shows how spin locks can be implemented using LDSTUB. A nonzero value for the
lock represents the locked condition, while a zero value means that the lock is free. Note that the
code busy waits by doing loads to avoid generating expensive stores to a potentially shared location. The MEMBAR #StoreStore in UnLockWithLDSTUB ensures that pending stores are
completed before the store that frees the lock.
LockWithLDSTUB(lock)
retry:
ldstub
tst
be
nop
loop:
ldub
tst
bne
nop
ba,a
out:
membar

[lock],%l0
%l0
out

[lock],%l0
%l0
loop
retry
#LoadLoad | #LoadStore

UnLockWithLDSTUB(lock)
membar
#StoreStore
membar
#LoadStore
stub
%g0,[lock]

!RMO and PSO only
!RMO only

Example 7—Lock and Unlock Using LDSTUB

The code in example 8 shows how spin locks can be implemented using CASA. Again, a nonzero
value for the lock represents the locked condition, while a zero value means the lock is free. The
nonzero lock value (ID) is supplied by the caller and may be used to identify the current owner of
the lock. This value is available while spinning and could be used to maintain a time-out or to verify that the thread holding the lock is still running. As in the previous case, the code busy-waits by
doing loads, not stores.

J.7 Producer-Consumer Relationship
In a producer-consumer relationship,the producer process generates data and puts it into a buffer,
while the consumer process takes data from the buffer and uses it. If the buffer is full, the producer
process stalls when trying to put data into the buffer. If the buffer is empty, the consumer process
stalls when trying to remove data.

LockWithCAS(lock, ID)
retry:
mov
[ID],%l0
cas
[lock],%g0,%l0
tst
%l0
be
out
nop
loop:
ld
[lock],%l0
tst
%l0
bne
loop
nop
ba,a
retry
out:
membar

#LoadLoad | #LoadStore

UnLockWithCAS(lock)
membar
#StoreStore
membar
#LoadStore
st
%g0,[lock]

!See example 7

!RMO and PSO only
!RMO only

Example 8—Lock and Unlock Using CAS

Figure 50 shows the buffer data structure and register usage. Example 9 shows the producer and
consumer code. The code assumes the existence of two procedures, IncrHead and IncrTail,

which increment the head and tail pointers of the buffer in a wraparound manner and return the
incremented value, but do not modify the pointers in the buffer.
Buffer Data Structure:
buffer
( = %i0)

bufhead

buffer+4

buftail

Buffer Empty Condition:
bufhead == buftail
Buffer Full Condition:
IncrTail(buffer) == bufhead

bufdata
l
l
l
Register Usage:
%i0 and %i1

parameters

%l0 and %l1

local values

%o0

result

Figure 50—Data Structures for Producer-Consumer Code

Produce(buffer, data)
call
IncrTail
full:
ld
[%i0],%l0
cmp
%l0,%o0
be
full
ld
[%i0+4],%l0
st
%i1,[%l0]
membar
#StoreStore
st
%o0,[%i0+4]
Consume(buffer)
ld
[%i0],%l0
empty:
ld
[%i0+4],%l1
cmp
%l0,%l1
be
empty
call
IncrHead
ld
[%l0],%l0
membar
#LoadStore
st
%o0,[%i0]
mov
%l0,%o0

!RMO and PSO only

!RMO only

Example 9—Producer and Consumer Code

J.8 Process Switch Sequence
This subsection provides code that must be used during process or thread switching to ensure that
the memory model seen by a process or thread is the one seen by a processor. The HeadSequence must be inserted at the beginning of a process or thread when it starts executing on a
processor. The TailSequence must be inserted at the end of a process or thread when it relinquishes a processor.
Example 10 shows the head and tail sequences. The two sequences refer to a per-process variable
tailDone. The value 0 for tailDone means that the process is running, while the value –1 (all ones)
means that the process has completed its tail sequence and may be migrated to another processor
if the process is runnable. When a new process is created, tailDone is initialized to –1.
HeadSequence(tailDone)
nrdy:
ld
[tailDone],%l0
cmp
%l0,-1
bne
nrdy
st
%g0, [tailDone]
membar
#StoreLoad
TailSequence(tailDone)
mov
-1,%l0
membar
#StoreStore
membar
#LoadStore
st
%l0,[tailDone]

!RMO and PSO only
!RMO only (combine with above)

Example 10—Process or Thread Switch Sequence

The MEMBAR in HeadSequence is required to be able to provide a switching sequence that
ensures that the state observed by a process in its source processor will also be seen by the process
in its destination processor. Since FLUSHes and stores are totally ordered, the head sequence
need not do anything special to ensure that FLUSHes performed prior to the switch are visible by
the new processor.
Programming Note:
A conservative implementation may simply use a MEMBAR with all barriers set.

J.9 Dekker’s Algorithm
Dekker’s algorithm is the classical sequence for synchronizing entry into a critical section using
loads and stores only. The reason for showing this example here is to illustrate how one may
ensure that a store followed by a load in issuing order will be executed by the memory system in
that order. Dekker’s algorithm is not a valid synchronization primitive for SPARC-V9, because it
requires a sequentially consistent (SC) memory model in order to work. Dekker’s algorithm (and
similar synchronization sequences) can be coded on RMO, PSO, and TSO by adding appropriate
MEMBAR instructions. This example also illustrates how future compilers can provide the equivalent of sequential consistency on systems with weaker memory models.

Example 11 shows the entry and exit sequences for Dekker’s algorithm. The locations A and B are
used for synchronization; A = 0 means that process P1 is outside its critical section, while any
other value means that P1 is inside it; similarly, B = 0 means that P2 is outside its critical section,
and any other value means that P2 is inside it.
P1Entry( )
mov
busy:
st
membar
ld
tst
bne,a
st
P1Exit( )
membar
membar
st
P2Entry( )
mov
busy:
st
membar
ld
tst
bne,a
st
P2Exit( )
membar
membar
st

-1,%l0
%l0,[A]
#StoreLoad
[B],%l1
%l1
busy
%g0,[A]

#StoreStore
#LoadStore
%g0,[A]

!RMO and PSO only
!RMO only

-1,%l0
%l0,[B]
#StoreLoad
[A],%l1
%l1
busy
%g0,[B]

#StoreStore
#LoadStore
%g0,[B]

!RMO and PSO only
!RMO only

Example 11—Dekker’s Algorithm

Dekker’s algorithm guarantees mutual exclusion, but it does not guarantee freedom from deadlock. In this case, it is possible that both processors end up trying to enter the critical region without success. The code above tries to address this problem by briefly releasing the lock in each
retry loop. However, both stores are likely to be combined in a store buffer, so the release has no
chance of success. A more realistic implementation would use a probabilistic back-off strategy
that increases the released period exponentially while waiting. If any randomization is used, such
an algorithm will avoid deadlock with arbitrarily high probability.

J.10 Code Patching
The code patching example illustrates how to modify code that is potentially being executed at the
time of modification. Two common uses of code patching are in debuggers and dynamic linking.

Code patching involves a modifying process, Pm, and one or more target processes Pt. For simplicity, assume that the sequence to be modified is four instructions long: the old sequence is
(Old1, Old2, Old3, Old4), and the new sequence is (New1, New2, New3, New4). There are two
examples: noncooperative modification, in which the changes are made without cooperation
from Pt; and cooperative modification, in which the changes require explicit cooperation from Pt.
In noncooperative modification, illustrated in example 12, changes are made in reverse execution
order. The three partially modified sequences (Old1, Old2, Old3, New4), (Old1, Old2, New3,
New4), and (Old1, New2, New3, New4) must be legal sequences for Pt, in that Pt must do the right
thing if it executes any of them. Additionally, none of the locations to be modified, except the first,
may be the target of a branch. The code assumes that %i0 contains the starting address of the area
to be patched and %i1, %i2, %i3, and %i4 contain New1, New2, New3, and New4.
NonCoopPatch(addr, instructions...)
st
%i4,[%i0+12]
flush
%i0+12
membar
#StoreStore
!RMO and PSO only
st
%i3,[%i0+8]
flush
%i0+8
membar
#StoreStore
!RMO and PSO only
st
%i2,[%i0+4]
flush
%i0+4
membar
#StoreStore
!RMO and PSO only
st
%i1,[%i0]
flush
%i0
Example 12—Nonxooperative Code Patching

The constraint that all partially modified sequences must be legal is quite restrictive. When this
constraint cannot be satisfied, noncooperative code patching may require the target processor to
execute FLUSH instructions. One method of triggering such a non-local FLUSH would be to send
an interrupt to the target processor.
In cooperative code patching, illustrated in example 13, changes to instructions can be made in
any order. When Pm is finished with the changes, it writes into the shared variable done to notify
Pt. Pt waits for done to change from 0 to some other value as a signal that the changes have been
completed. The code assumes that %i0 contains the starting address of the area to be patched,
%i1, %i2, %i3, and %i4 contain New1, New2, New3, and New4, and %g1 contains the address
of done. The FLUSH instructions in Pt ensure that the instruction buffer of Pt’s processor is
flushed so that the old instructions are not executed.

J.11 Fetch_and_Add
Fetch_and_Add performs the sequence a = a + b atomically with respect to other Fetch_and_Adds
to location a. Two versions of Fetch_and_Add are shown. The first (example 14) uses the routine
LockWithLDSTUB described above. This approach uses a lock to guard the value. Since the memory model dependency is embodied in the lock access routines, the code does not depend on the
memory model.1

CoopPatch(addr, instructions...)
st
%i1,[%i0]
st
%i2,[%i0+4]
st
%i3,[%i0+8]
st
%i4,[%i0+12]
mov
-1,%l0
membar
#StoreStore
st
%l0,[%g1]

!%i0 = addr, %i1..%i4 = instructions

!RMO and PSO only

TargetCode( )
wait:
ld
[%g1],%l0
cmp
%l0,0
be
wait
flush
A
flush
A+4
flush
A+8
flush
A+12
A:
Old1
Old2
Old3
Old4
Example 13—Cooperative Code Patching
/*Fetch and Add using LDSTUB*/
int Fetch_And_Add(Index, Increment, Lock)
int *Index;
int Increment;
int *Lock;
{
int old_value;
LockWithLDSTUB(Lock);
old_value = *Index;
*Index = old_value + Increment;
UnlockWithLDSTUB(Lock);
return(old_value);
}
Example 14—Fetch and Add Using LDSTUB

Fetch_and_Add originally was invented to avoid lock contention and to provide an efficient means
to maintain queues and buffers without cumbersome locks. Hence, using a lock is inefficient and
contrary to the intentions of the Fetch_and_Add. The CAS synthetic instruction allows a more
efficient version, as shown in example 15.

1. Inlining of the lock-access functions with subsequent optimization may break this code.

FetchAndAddCAS(address, increment)
retry:
ld
[%i0],%l0
add
%l0,%i1,%l1
cas
[%i0],%l0,%l1
cmp
%l0,%l1
bne
retry
mov
%l1,%o0

!%i0 = address, %i1 = increment

!return old value

Example 15—Fetch and Add Using CAS

J.12 Barrier Synchronization
Barrier synchronization ensures that each of N processes is blocked until all of them reach a given
state. The point in the flow of control at which this state is reached is called the barrier; hence the
name. The code uses the variable Count initialized to N. As each process reaches its desired state,
it decrements Count and waits for Count to reach zero before proceeding further.
Similar to the fetch and add operation, barrier synchronization is easily implemented using a lock
to guard the counter variable, as shown in example 16.
/*Barrier Synchronization using LDSTUB*/
Barrier(Count,Lock)
int *Count;
int *Lock;
{
LockWithLdstUB(Lock);
*Count = *Count - 1;
UnlockWithLdstUB(Lock);
while(*Count > 0) { ; /*busy-wait*/ }
}
Example 16—Barrier Synchronization Using LDSTUB

The CAS implementation of barrier synchronization, shown in example 17, avoids the extra lock
access.
A practical barrier synchronization must be reusable because it is typically used once per iteration
in applications that require many iterations. Barriers that are based on counters must have means
to reset the counter. One solution to this problem is to alternate between two complementary versions of the barrier: one that counts down to 0 and the other that counts up to N. In this case, passing one barrier also initializes the counter for the next barrier.
Passing a barrier can also signal that the results of one iteration are ready for processing by the
next iteration. In this case, RMO and PSO require a MEMBAR #StoreStore instruction prior to
the barrier code to ensure that all local results become globally visible prior to passing the barrier.
Barrier synchronization among a large number of processors will lead to access contention on the
counter variable, which may degrade performance. This problem can be solved by using multiple

BarrierCAS(Count)
!%i0 = address of counter variable
retry:
ld
[%i0],%l0
add
%l0,-1,%l1
cas
[%i0],%l0,%l1
cmp
%l0,%l1
bne
retry
nop
wait:
ld
[%i0],%l0
tst
%l0
bne
wait
nop
Example 17—Barrier Synchronization Using CAS

counters. The butterfly barrier uses a divide-and-conquer technique to avoid any contention and
can be implemented without atomic operations.1

J.13 Linked List Insertion and Deletion
Linked lists are dynamic data structures that might be used by a multi-threaded application. As in
the previous examples, a lock can be used to guard access to the entire data structure. However,
single locks guarding large and frequently shared data structures can be inefficient.
In example 18, the CAS synthetic instruction is used to operate on a linked list without explicit
locking. Each list element starts with an address field that contains either the address of the next
list element or zero. The head of the list is the address of a variable that holds the address of the
first list element. The head is zero for empty lists.
In the example, there is little difference in performance between the CAS and lock approaches,
however, more complex data structures can allow more concurrency. For example, a binary tree
allows the concurrent insertion and removal of nodes in different branches.

J.14 Communicating With I/O Devices
I/O accesses may be reordered just as other memory reference are reordered. Because of this, the
programmer must take special care to meet the constraint requirements of physical devices, in
both the uniprocessor and multiprocessor cases.
Accesses to I/O locations require sequencing MEMBARs under certain circumstances to properly
manage the order of accesses arriving at the device, and the order of device accesses with respect
to memory accesses. The following rules describe the use of MEMBARs in these situations.
Maintaining the order of accesses to multiple devices will require higher-level software constructs, which are beyond the scope of this discussion.
1. Brooks, E. D., “The Butterfly Barrier,” International Journal of Parallel Programming 15(4), pp. 295-307, 1986.

ListInsert(Head, Element) !%i0 = Head addr, %i1 = Element addr
retry:
ld
[%i0],%l0
st
%l0, [%i1]
mov
%i1, %l1
cas
[%i0],%l0,%l1
cmp
%l0,%l1
bne
retry
nop
ListRemove(Head)
!%i0
retry:
ld
[%i0],%o0
tst
%o0
be
empty
nop
ld
[%o0],%l0
cas
[%i0],%o0,%l0
cmp
%o0,%l0
bne
retry

= Head addr

empty:
nop
Example 18—List Insertion and Removal

(1) Accesses to the same I/O location address:
— A store followed by a store is ordered in all memory models.
— A load followed by a load requires a MEMBAR #LoadLoad in RMO only..
Compatibility Note:
This MEMBAR is not needed in implementations that provide SPARC-V8 compatibility for I/
O accesses in RMO.

— A load followed by a store is ordered in all memory models.
— A store followed by a load requires MEMBAR #Lookaside between the accesses
for all memory models; however, implementations that provide SPARC-V8 compatiblity for I/O accesses in any of TSO, PSO, and RMO do not need the MEMBAR in any
model that provides this compatibility.
(2) Accesses to different I/O location addresses:
— The appropriate ordering MEMBAR is required to guarantee order within a range of
addresses assigned to a device.
— Device-specific synchronization of completion, such as reading back from an address
after a store, may be required to coordinate accesses to multiple devices. This is
beyond the scope of this discussion.
(3) Accesses to an I/O location address and a memory address.

— A MEMBAR #MemIssue is required between an I/O access and a memory access if
it is required that the I/O access reaches global visibility before the memory access
reaches global visibility. For example, if the memory location is a lock that controls
access to an I/O address, then MEMBAR #MemIssue is required between the last
access to the I/O location and the store that clears the lock.
(4) Accesses to different I/O location addresses within an implementation-dependent range of
addresses are strongly ordered once they reach global visiblity. Beyond the point of global
visibility there is no guarantee of global order of accesses arriving at different devices having disjoint implementation-dependent address ranges defining the device. Programmers
can rely on this behavior from implementations.
(5) Accesses to I/O locations protected by a lock in shared memory that is subsequently
released, with attention to the above barrier rules, are strongly ordered with respect to any
subsequent accesses to those locations that respect the lock.
J.14.1 I/O Registers With Side Effects
I/O registers with side effects are commonly used in hardware devices such as UARTs. One register is used to address an internal register of the I/O device, and a second register is used to transfer
data to or from the selected internal register.
In examples 19 and 20, let X be the address of a device with two such registers; X.P is a port register, and X.D is a data register. The address of an internal register is stored into X.P; that internal
register can then be read or written by loading into or storing from X.D.
st
membar
st

%i1, [X+P]
#StoreStore
%i2, [X+D]

! PSO and RMO only

Example 19—I/O Registers With Side-Effects: Store Followed by Store

st
membar
ld

%i1, [X+P]
#StoreLoad |#MemIssue
[X+D], %i2

! RMO only

Example 20—I/O Registers With Side-Effects: Store Followed by Load

J.14.2 The Control and Status Register (CSR)
A control and status register is an I/O location which is updated by an I/O device independent of
access by the processor. For example, such a register might contain the current sector under the
head of a disk drive.
In example 21, let Y be the address of a control and status register that is read to obtain status and
written to assert control. Bits read differ from the last data that was stored to them.
ld
st
membar
ld

[Y], %i1
%i2, [Y]
#Lookaside
[Y], %i3

!
!
!
!

obtain status
write a command
make sure we really read the register
obtain new status

Example 21—Accessing a Control/Status Register

Access to these registers, of course, must be protected by a mutual-exclusion lock to ensure that
multiple threads accessing the registers do not interfere. The sequencing MEMBAR is needed to
ensure the value produced by the load comes from the register and not from the write buffer since
the write has side-effects. No MEMBAR is needed between the load and the store, because of the
anti-dependency on the memory address.
J.14.3 The Descriptor
In example 22, let A be the address of a descriptor in memory. After initializing the descriptor
with information, the address of the descriptor is stored into device register D or made available to
some other portion of the program that will make decisions based upon the value(s) in the descriptor. It is important to ensure that the stores of the data have completed before making the address
(and hence the data in the descriptor) visible to the device or program component.
st
st
...
membar
st

%i1, [A]
%i2, [A+4]
#StoreStore
A, [D]

! more stores
! PSO and RMO only

Example 22—Accessing a Memory Descriptor

Access must be protected by a mutual-exclusion lock to ensure that multiple threads accessing the
registers do not interfere. In addition, the agent reading the descriptor must use a load-barrier
MEMBAR after reading D to ensure that the most recent values are read.
J.14.4 Lock-Controlled Access to a Device Register
Let A be a lock in memory that is used to control access to a device register D. The code that
accesses the device might look like that show in example 23.
The sequencing MEMBAR is needed to ensure that another CPU which grabs the lock and loads
from the device register will actually see any changes in the device induced by the store. The

set A, %l1
set D, %l2
call lock
mov %l1, %o0
ld [%l2], %i1

! address of lock
! address of device register
! lock(A);

...

! do some computation

st %i2, [%l2]
membar #MemIssue
call unlock
mov %l1, %o0

! write the register
! all memory models
! unlock(A);

! read the register

Example 23—Accessing a Device Register

This appendix is informative only.
It is not part of the SPARC-V9 specification.
ordering MEMBARs in the lock and unlock code (see J.6, “Spin Locks”), while ensuring correctness when protecting ordinary memory, are insufficient for this purpose when accessing device
registers. Compare with J.14.1, “I/O Registers With Side Effects.”
(This Annex is not a part of SPARC-V9/R1.4.5,
The SPARC Architecture Manual;
it is included for information only.)

K Changes From SPARC-V8 to SPARC-V9
SPARC-V9 is complimentary to the SPARC-V8 architecture; it does not replace it. SPARC-V9
was designed to be a higher-performance peer to SPARC-V8.
Application software for the 32-bit SPARC-V8 (Version 8) microprocessor architecture can execute, unchanged, on SPARC-V9 systems. SPARC-V8 software executes natively on SPARC-V9conformant processors; no special compatibility mode is required.
Changes to the SPARC-V9 architecture since SPARC-V8 are in six main areas: the trap model,
data formats, the registers, alternate address space access, the instruction set, and the memory
model.

K.1 Trap Model
The trap model, visible only to privileged software, has changed substantially.
— Instead of one level of traps, four or more levels are now supported. This allows first-level
trap handlers, notably register window spill and fill (formerly called overflow and underflow) traps, to execute much faster. This is because such trap handlers can now execute
without costly run-time checks for lower-level trap conditions, such as page faults or a

misaligned stack pointer. Also, multiple trap levels support more robust fault-tolerance
mechanisms.
— Most traps no longer change the CWP. Instead, the trap state (including the CWP register)
is saved in register stacks called TSTATE, TT, TPC, and TNPC.
— New instructions (DONE and RETRY) are used to return from a trap handler, instead of
RETT.
— A new instruction (RETURN) is provided for returning from a trap handler running in
nonprivileged mode, providing support for user trap handlers.
— Terminology about privileged-mode execution has changed, from “supervisor/user” to
“privileged/nonprivileged.”
— A new processor state, RED_state, has been added to facilitate processing resets and
nested traps that would exceed MAXTL.

K.2 Data Formats
Data formats for extended (64-bit) integers have been added.

K.3 Little-Endian Support
Data accesses can be either big-endian or little-endian. Bits in the PSTATE register control the
implicit endianness of data accesses. Special ASI values are provided to allow specific data
accesses to be in a specific endianness.

K.4 Registers
These privileged SPARC-V8 registers have been deleted:
— PSR: Processor State Register
— TBR: Trap Base Register
— WIM: Window Invalid Mask
These registers have been widened from 32 to 64 bits:
— All integer registers
— All state registers: FSR, PC, nPC, Y
The contents of the following register has changed:
— FSR: Floating-Point State Register: fcc1, fcc2, and fcc3 (additional floating-point condition code) bits have been added and the register widened to 64-bits..

These SPARC-V9 registers are fields within a register in SPARC-V8:
— PIL: Processor Interrupt Level register
— CWP: Current Window Pointer register
— TT[MAXTL]: Trap Type register
— TBA: Trap Base Address register
— VER: Version register
— CCR: Condition Codes Register
These registers have been added:
— Sixteen additional double-precision floating-point registers, f[32]..f[62], which are
aliased with and overlap eight additional quad-precision floating-point registers,
f[32]..f[60]
— FPRS: Floating-Point Register State register
— ASI: ASI register
— PSTATE: Processor State register
— TL: Trap Level register
— TPC[MAXTL]: Trap Program Counter register
— TNPC[MAXTL]: Trap Next Program Counter register
— TSTATE[MAXTL]: Trap State register
— TICK: Hardware clock-tick counter
— CANSAVE: Savable windows register
— CANRESTORE: Restorable windows register
— OTHERWIN: Other windows register
— CLEANWIN: Clean windows register
— WSTATE: Window State register
The SPARC-V9 CWP register is incremented during a SAVE instruction and decremented during
a RESTORE instruction. Although this is the opposite of PSR.CWP’s behavior in SPARC-V8, the

only software it should affect is a few trap handlers that operate in privileged mode, and that must
be rewritten for SPARC-V9 anyway. This change will have no effect on nonprivileged software.

K.5 Alternate Space Access
In SPARC-V8, access to all alternate address spaces is privileged. In SPARC-V9, loads and stores
to ASIs 0016 ..7f16 are privileged; those to ASIs 8016 ..FF16 are nonprivileged. That is, load- and
store-alternate instructions to one-half of the alternate spaces can now be used in user code.

K.6 Little-Endian Byte Order
In SPARC-V8, all instruction and data accesses were performed in big-endian byte order.
SPARC-V9 supports both big- and little-endian byte orders for data accesses only; instruction
accesses in SPARC-V9 are always performed using big-endian order.

K.7 Instruction Set
All changes to the instruction set were made such that application software written for SPARC-V8
can run unchanged on a SPARC-V9 processor. Application software written for SPARC-V8
should not even be able to detect that its instructions now process 64 bit values.
The definitions of the following instructions were extended or modified to work with the 64-bit
model:
— FCMP, FCMPE: Floating-Point Compare—can set any of the four floating-point condition
codes
— LDUW, LDUWA(same as “LD, LDA” in SPARC-V8)
— LDFSR, STFSR: Load/Store FSR: only affect low-order 32 bits of FSR
— RDASR/WRASR: Read/Write State Registers: access additional registers
— SAVE/RESTORE
— SETHI
— SRA, SRL, SLL: Shifts: split into 32-bit and 64-bit versions
— Tcc: (was Ticc) operates with either the 32-bit integer condition codes (icc), or the 64-bit
integer condition codes (xcc)
— All other arithmetic operations now operate on 64-bit operands and produce 64-bit results.
Application software written for SPARC-V8 cannot detect that arithmetic operations are
now 64 bits wide. This is due to retention of the 32-bit integer condition codes (icc), addition of 64-bit integer condition codes (xcc), and the carry-propagation rules of 2’s-complement arithmetic.

The following instructions have been added to provide support for 64-bit operations and/or
addressing:
— F[sdq]TOx: Convert floating point to 64-bit word
— FxTO[sdq]: Convert 64-bit word to floating point
— FMOV[dq]: Floating-point Move, double and quad
— FNEG[dq]: Floating-point Negate, double and quad
— FABS[dq]: Floating-point Absolute Value, double and quad
— LDDFA, STDFA, LDFA, STFA: Alternate address space forms of LDDF, STDF, LDF, and
STF
— LDSW: Load a signed word
— LDSWA: Load a signed word from an alternate space
— LDX: Load an extended word
— LDXA: Load an extended word from an alternate space
— LDXFSR: Load all 64 bits of the FSR register
— STX: Store an extended word
— STXA: Store an extended word into an alternate space
— STXFSR: Store all 64 bits of the FSR register
The following instructions have been added to support the new trap model:
— DONE: Return from trap and skip instruction that trapped
— RDPR and WRPR: Read and Write privileged registers
— RESTORED: Adjust state of register windows after RESTORE
— RETRY: Return from trap and reexecute instruction that trapped
— RETURN: Return
— SAVED: Adjust state of register windows after SAVE
— SIR: Signal Monitor (generate Software Initiated Reset)
The following instructions have been added to support implementation of higher-performance
systems:
— BPcc: Branch on integer condition code with prediction
— BPr: Branch on integer register contents with prediction
— CASA, CASXA: Compare and Swap from an alternate space
— FBPfcc: Branch on floating-point condition code with prediction

— FLUSHW: Flush windows
— FMOVcc: Move floating-point register if condition code is satisfied
— FMOVr: Move floating-point register if integer register contents satisfy condition
— LDQF(A), STQF(A): Load/Store Quad Floating-point (in an alternate space)
— MOVcc: Move integer register if condition code is satisfied
— MOVr: Move integer register if register contents satisfy condition
— MULX: Generic 64-bit multiply
— POPC: Population Count
— PREFETCH, PREFETCHA: Prefetch Data
— SDIVX, UDIVX: Signed and Unsigned 64-bit divide
The definitions of the following instructions have changed:
— IMPDEPn: Implementation-Dependent instructions (replace SPARC-V8 CPop instructions)
The following instruction was added to support memory synchronization:
— MEMBAR: Memory barrier
The following instructions have been deleted:
— Coprocessor loads and stores
— RDTBR and WRTBR: TBR no longer exists. It has been replaced by TBA, which can be
read/written with RDPR/WRPR instructions.
— RDWIM and WRWIM: WIM no longer exists. WIM has been subsumed by several register-window state registers.
— RDPSR and WRPSR: PSR no longer exists. It has been replaced by several separate registers which are read/written with other instructions.
— RETT: Return from trap (replaced by DONE/RETRY).
— STDFQ: Store Double from Floating-point Queue (replaced by the RDPR FQ instruction).

K.8 Memory Model
SPARC-V9 defines a new memory model called Relaxed Memory Order (RMO). This very weak
model allows the CPU hardware to schedule memory accesses such as loads and stores in nearly
any order, as long as the program computes the correct answer. Hence, the hardware can instanta-

This bibliography is informative only.
It is not part of the SPARC-V9 specification.
neously adjust to resource contentions and schedule accesses in the most efficient order, leading to
much faster memory operations and better performance.
(This Bibliography is not a part of SPARC-V9/R1.4.5,
The SPARC Architecture Manual;
it is included for information only.)

Bibliography
General References
For general information, see the following:
-----. The SPARC Architecture Manual, Version 8, Prentice-Hall, Inc., 1992.
Boney, Joel [1992]. “SPARC Version 9 Points the Way to the Next Generation RISC,” SunWorld,
October 1992, pp. 100-105.
Catanzaro, Ben, ed. The SPARC Technical Papers, Springer-Verlag, 1991.
Cmelik, R. F., S. I. Kong, D. R. Ditzel, and E. J. Kelly, “An Analysis of MIPS and SPARC Instruction Set Utilization on the SPEC Benchmarks,” Proceedings of the Fourth International Symposium on Architectural Support for Programming Languages and Operating Systems, April 8-11,
1991.
Dewar, R. B. K. and M. Smosna. Microprocessors: A Programmer’s View, McGraw-Hill, Inc.,
1990.
Ditzel, David R. [1993]. “SPARC Version 9: Adding 64-Bit Addressing and Robustness to an
Existing RISC Architecture.” Videotape available from University Video Communications, P. O.
Box 5129, Stanford, CA, 94309.
Garner, R. B. [1988]. “SPARC: The Scalable Processor Architecture,” SunTechnology, vol. 1, no.
3, Summer, 1988; also appeared in M. Hall and J. Barry (eds.), The SunTechnology Papers,
Springer-Verlag, 1990, pp. 75-99.
Garner, R. B., A. Agrawal, F. Briggs, E. W. Brown, D. Hough, W. N. Joy, S. Kleiman, S. Muchnick, M. Namjoo, D. Patterson, J. Pendleton, K. G. Tan, and R. Tuck [1988]. “The Scalable Processor Architecture (SPARC),” 33rd Annual IEEE Computer Conference (COMPCON), February,
1988, San Francisco, CA.

Hennessy, J. and D. Patterson. Computer Architecture: A Quantitative Approach, Morgan Kaufman Publishers, Inc., San Mateo, CA. 1990.
IEEE Standard for Binary Floating-Point Arithmetic, IEEE Std 754-1985, IEEE, New York, NY,
1985.
Katevenis, M. [1983]. Reduced Instruction Set Computer Architectures for VLSI, Ph.D. dissertation, Computer Science Div., Univ. of California, Berkeley, 1983; also published by M.I.T. Press,
Cambridge, MA, 1985.
Kleiman, S. and D. Williams [1988]. “SunOS on SPARC,” 33rd Annual IEEE Comp. Conf. (COMPCON), February, 1988, San Francisco, CA; also appeared in M. Hall and J. Barry (eds.), The
SunTechnology Papers, Springer-Verlag, 1990, pp. 13-27.
Muchnick, S. [1988]. “Optimizing Compilers for SPARC,” SunTechnology, Summer 1988, pp. 6471; also appeared in W. Stallings (ed.), Reduced Instruction Set Computers (2nd edition), IEEE
Computer Society Press, 1990, pp. 160-173, and in M. Hall and J. Barry (eds.), The SunTechnology Papers, Springer-Verlag, 1990, pp. 41-68.
Patterson, D. [1985]. “Reduced Instruction Set Computers,” Communications of the ACM, vol. 28,
no. 1, January, 1985.
Patterson, D., and D. R. Ditzel, “The Case for the Reduced Instruction Set Computer,” Computer
Architecture News, vol 8, no. 7, 1980.

Memory Model References
The concept of a memory model has become a significant one as shared memory multiprocessors
are more widely used. The issues are complex and interesting, and have created an active and
extensive literature. A partial annotated list of references is as follows:
Collier, W. W. Reasoning About Parallel Architectures, Prentice Hall, 1992.
Provides a mathematical framework for the study of parallel processors and their interaction
with memory.
Dill, David, Seungjoon Park, and Andreas G. Nowatzyk, “Formal Specification of Abstract Memory Models” in Research on Integrated Systems: Proceedings of the 1993 Symposium, Ed. Gaetano Borriello and Carl Ebeling, MIT Press, 1993.
Describes an application of software tools to the verification of the TSO and PSO memory
models.
Gharachorloo, K., D. Lenoski, J. Laudon, P. Gibbon, A. Gupta, and J. Hennessy. “Memory Consistency and Event Ordering in Scalable Shared-Memory Multiprocessors,” Proceedings of the
17th Annual International Symposium on Computer Architecture, May 1990, pp. 15-29.
Provides an overview of contemporary research in memory models.

Gharachorloo, K., S. Adve, A. Gupta, J. Hennessy, and M. Hill. “Programming for Different
Memory Consistency Models,” Journal of Parallel and Distributed Processing, 15:4, August
1992.
This paper proposes a new programming model which allows programmers to reason about
programs that have not been reduced to sequential consistency.
Gharachorloo, K., A. Gupta, and J. Hennessy, “Performance Evaluation of Memory Consistency
Models for Shared Memory Multiprocessors,” Proceedings of the 4th International Conference
on Architectural Support for Programming Languages and Operating Systems, pp. 245-257,
ACM, New York, 1991.
This paper discusses the performance benefits that can be obtained when a relaxed memory
model is used in a shared-memory model processor.
Lamport, Leslie. “How to Make a Multiprocessor Computer That Correctly Executes Multiprocess Programs,” IEEE Transactions on Computers, C-28, 9, September 1979, pp. 690-691.
Defines sequential consistency and shows how it can be used in simple shared-memory systems.
Reynal, M. Algorithms for Mutual Exclusion, MIT Press, 1986.
Provides an overview of the mutual exclusion problem and the extensive literature associated
with it.
Scheurich, C., and M. Dubois. “Dependency and Hazard Resolution in Multiprocessors,” Proceedings of the 14th International Symposium on Computer Architecture, pp. 234-243, IEEE CS
Press, Los Alamitos, CA, 1987.
Sindhu, Predeep, Jean-Marc Frailong, and Michel Ceklov. “Formal Specification of Memory
Models,” Xerox Palo Alto Research Center Report CSL-91-11, December 1991.
Introduces the formal framework used to define the SPARC-V8 TSO and PSO memory models.
Treiber, R. Kent. “Systems Programming: Coping with Parallelism,” IBM Research Report
RJ5118 (53162), 1986.
Provides an overview of the operational issues for systems programming in a multiprocessing
environment.

Prefetching
Callahan, D., K. Kennedy, A. Porterfield. “Software Prefetching,” Proceedings of the Fourth
International Conference on Architectural Support for Programming Languages and Operating
Systems, April 1991, pp. 40-52.

Mowry, T., M. Lam, and A. Gupta. “Design and Evaluation of a Compiler Algorithm for Prefetching.” Proceedings of the Fifth International Conference on Architectural Support for Programming Languages and Operating Systems, October 1992, pp. 62-73.

A
a field of instructions, 66, 138, 141, 144, 147, 148, 152
ABI, see SPARC-V9 Application Binary Interface (ABI)
accrued exception (aexc) field of FSR register, 46, 48, 100, 247, 254
activation record, see stack frame
ADD instruction, 137, 299
ADDC instruction, 137
ADDcc instruction, 137, 222, 299
ADDCcc instruction, 137
address, 120
aliased, 120
physical, 120, 281
virtual, 120, 281
address, 295
address aliases, 281
address mask (AM) field of PSTATE register, 53, 151, 172, 215
address space, 4, 281, 282
address space identifier (ASI), 9, 16, 17, 50, 63, 67, 69, 73, 120, 121, 174, 179, 207, 227, 254, 283, 317, 341
architecturally specified, 122
restricted, 74, 122, 254
unrestricted, 74, 122, 254
address space identifier (ASI) register, 16, 21, 50, 56, 73, 89, 122, 157, 176, 181, 183, 207, 227, 232, 235, 245,
316
addressing conventions, 17, 70
addressing modes, 4
ADDX instruction (SPARC-V8), 137
ADDXcc instruction (SPARC-V8), 137
aexc, see accrued exception (aexc) field of FSR register
AG, see alternate globals enable (AG) field of PSTATE register
aggregate data values, see data aggregates
alias
address, 120
floating-point registers, 36
alignment, 304
data (load/store), 17, 69, 121
doubleword, 17, 69, 121
extended-word, 69
halfword, 17, 69, 121
instructions, 17, 69, 121
integer registers, 179, 181
memory, 121
quadword, 17, 69, 121
stack pointer, 304
word, 17, 69, 121
alternate address space, 207
alternate global registers, 15, 30, 30, 316

alternate globals enable (AG) field of PSTATE register, 30, 31, 54
alternate space instructions, 18, 50, 341
AM, see address mask (AM) field of PSTATE register
ancillary state registers (ASRs), 18, 35, 36, 60, 214, 215, 244, 245, 252, 253, 292, 321
AND instruction, 184
ANDcc instruction, 184, 299
ANDN instruction, 184, 299
ANDNcc instruction, 184
annul bit, 35, 138
in conditional branches, 141
annulled branches, 138
application program, 9, 14, 16, 30, 46, 47, 50, 61, 104, 341
architectural extensions, 7, 321
arguments to a subroutine, 302
arithmetic overflow, 41
ASI register, see address space identifier (ASI) register
ASI, see address space identifier (ASI)
ASI_AS_IF_USER_PRIMARY, 75, 123, 254, 287, 316
ASI_AS_IF_USER_PRIMARY_LITTLE, 75, 123, 254, 287, 316
ASI_AS_IF_USER_SECONDARY, 75, 123, 254, 287
ASI_AS_IF_USER_SECONDARY_LITTLE, 75, 123, 254, 287
ASI_NUCLEUS, 75, 75, 122, 254, 287
ASI_NUCLEUS_LITTLE, 75, 122, 254, 287
ASI_PRIMARY, 73, 75, 75, 122, 123, 254, 287
ASI_PRIMARY_LITTLE, 52, 75, 122, 254, 287
ASI_PRIMARY_NOFAULT, 75, 75, 123, 254, 284, 287
ASI_PRIMARY_NOFAULT_LITTLE, 75, 254
ASI_SECONDARY, 75, 75, 123, 254, 287, 317
ASI_SECONDARY_LITTLE, 75, 254, 287, 317
ASI_SECONDARY_NOFAULT, 75, 75, 123, 254, 284, 287
ASI_SECONDARY_NOFAULT_LITTLE, 75, 254, 287
asr_reg, 292
assembler
synthetic instructions, 297
assigned value
implementation-dependent, 252
async_data_error exception, 113, 133, 153, 174, 177, 179, 181, 182, 183, 226, 228, 230, 232, 234, 236
atomic, 130, 230, 232
memory operations, 127, 130
atomic load-store instructions, 69, 152
compare and swap, 98, 152
load-store unsigned byte, 182, 234, 235
load-store unsigned byte to alternate space, 183
swap r register with alternate space memory, 235
swap r register with memory, 152, 234
atomicity, 121, 224, 258
automatic variables, 302

B
BA instruction, 147, 278
BCC instruction, 146, 278
BCLR synthetic instruction, 299
BCS instruction, 146, 278

BE instruction, 146, 278
Berkeley RISCs, xiv, 5
BG instruction, 146, 278
BGE instruction, 146, 278
BGU instruction, 146, 278
bibliography, 345
Bicc instructions, 35, 42, 146, 273, 278
big-endian btye order, 9
big-endian byte order, 17, 52, 70
binary compatibility, 6
bit vector concatenation, 3
BL instruction, 278
BLE instruction, 146, 278
BLEU instruction, 146, 278
BN instruction, 146, 147, 209, 278, 297
BNE instruction, 146, 278
BNEG instruction, 146, 278
BPA instruction, 148, 278
BPCC instruction, 148, 278
BPcc instructions, 35, 41, 42, 66, 67, 148, 209
BPCS instruction, 148, 278
BPE instruction, 148, 278
BPG instruction, 148, 278
BPGE instruction, 148, 278
BPGU instruction, 148, 278
BPL instruction, 148, 278
BPLE instruction, 148, 278
BPLEU instruction, 148, 278
BPN instruction, 148, 278
BPNE instruction, 148, 278
BPNEG instruction, 148, 278
BPOS instruction, 146, 278
BPPOS instruction, 148, 278
BPr instructions, 35, 66, 67, 138, 273, 278
BPVC instruction, 148, 278
BPVS instruction, 148, 278
branch
annulled, 138
delayed, 63
elimination, 81
fcc-conditional, 141, 144
icc-conditional, 147
prediction bit, 138
unconditional, 141, 144, 147, 149
with prediction, 5
branch if contents of integer register match condition instructions, 138
branch on floating-point condition codes instructions, 140
branch on floating-point condition codes with prediction instructions, 143
branch on integer condition codes instructions, 146
branch on integer condition codes with prediction instructions, 148
BRGEZ instruction, 138
BRGZ instruction, 138
BRLEZ instruction, 138
BRLZ instruction, 138

BRNZ instruction, 138
BRZ instruction, 138
BSET synthetic instruction, 299
BTOG synthetic instruction, 299
BTST synthetic instruction, 299
BVC instruction, 146, 278
BVS instruction, 146, 278
byte, 9
addressing, 70, 71
data format, 23
order, 17, 70
order, big-endian, 17, 52
order, implicit, 52
order, little-endian, 17, 52

C
C condition code bit, see carry (C) bit of condition fields of CCR
cache
coherence in RED_state, 92
data, 125
in RED_state, 92
instruction, 125
memory, 253
miss, 209
non-consistent instruction cache, 125
system, 6
call chain
walking, 303
CALL instruction, 19, 33, 34, 35, 151, 172, 302, 304
CALL synthetic instruction, 297
CANRESTORE, see restorable windows (CANRESTORE) register
CANSAVE, see savable windows (CANSAVE) register
carry (C) bit of condition fields of CCR, 41
CAS synthetic instruction, 127, 299
CASA instruction, 98, 130, 152, 182, 183, 234, 235, 299
CASX synthetic instruction, 127, 130, 299
CASXA instruction, 98, 130, 152, 182, 183, 234, 235, 299
catastrophic_error exception, 89, 91, 98, 99, 113, 114, 115
cc0 field of instructions, 66, 144, 148, 159, 195
cc1 field of instructions, 66, 144, 148, 159, 195
cc2 field of instructions, 66, 195
CCR, see condition codes (CCR) register
certificate of compliance, 8
cexc, see current exception (cexc) field of FSR register
CLE, see current_little-endian (CLE) field of PSTATE register
clean register window, 9, 33, 58, 60, 82, 86, 88, 114, 217
clean windows (CLEANWIN) register, 58, 60, 82, 83, 86, 87, 88, 211, 242, 259
clean_window exception, 60, 82, 87, 98, 101, 114, 218, 256
clock cycle, 51
clock-tick register (TICK), 51, 116, 211, 242, 257
CLR synthetic instruction, 299
CMP synthetic instruction, 233, 297
coherence, 120, 258

memory, 121, 224
unit, memory, 122
compare and swap instructions, 98, 152
comparison instruction, 76, 233
compatibility note, 4
compatibility with SPARC-V8, 4, 19, 30, 40, 43, 54, 58, 76, 78, 85, 104, 114, 115, 116, 121, 142, 145, 160,
170, 171, 174, 179, 181, 187, 215, 224, 226, 230, 232, 233, 237, 239, 241, 245, 322, 336
compatibility with SPARC-V9, 137
compliance, 8
certificate of, 8
certification process, 8
Level I, 7
Level II, 8
compliant SPARC-V9 implementation, 7
concatenation of bit vectors, 3
concurrency, 15
cond field of instructions, 66, 141, 144, 147, 148, 189, 195
condition codes, 153
floating-point, 141
integer, 41
condition codes (CCR) register, 21, 89, 137, 157, 202, 245
conditional branches, 141, 144, 147
conditional move instructions, 20
conforming SPARC-V9 implementation, 7
const22 field of instructions, 170
constants
generating, 220
contexts
Nucleus, 122, 287
control and status registers, 35
control-transfer instructions (CTIs), 19, 157
conventions
software, 301
convert between floating-point formats instructions, 162, 248
convert floating-point to integer instructions, 161, 250
convert integer to floating-point instructions, 163
coprocessor, 322
counter field of TICK register, 51
CPopn instructions (SPARC-V8), 171, 322
cross-domain call, 316
CTI, see control-transfer instructions (CTIs)
current exception (cexc) field of FSR register, 44, 46, 48, 84, 115, 247, 254
current window, 9
current window pointer (CWP) register, 9, 15, 21, 33, 56, 58, 58, 60, 82, 87, 89, 157, 169, 211, 217, 218, 242,
259
current_little_endian (CLE) field of PSTATE register, 52, 52, 122
CWP, see current window pointer (CWP) register

D
d16hi field of instructions, 66, 138
d16lo field of instructions, 66, 138
data access
RED_state, 92

data aggregate
argument passed by value, 302
examples of, 302
data alignment, see alignment
data cache, 125
data flow order constraints
memory reference instructions, 124
register reference instructions, 124
data formats
byte, 23
doubleword, 23
extended word, 23
halfword, 23
quadword, 23
tagged word, 23
word, 23
data memory, 131
data types, 23
floating-point, 23
signed integer, 23
unsigned integer, 23
data_access_error exception, 114, 153, 174, 177, 179, 181, 182, 183, 226, 228, 230, 232, 234, 236
data_access_exception exception, 114, 153, 174, 177, 179, 181, 182, 183, 226, 228, 230, 232, 234, 236
data_access_MMU_miss exception, 114, 153, 174, 177, 179, 181, 182, 183, 210, 226, 228, 230, 232, 234, 236,
256
data_access_protection exception, 114, 153, 175, 177, 179, 181, 182, 183, 226, 228, 230, 232
data_protection exception, 234, 236
DEC synthetic instruction, 299
DECcc synthetic instruction, 299
deferred trap, 95, 95, 96, 99, 254
avoiding, 96
floating-point, 212
deferred-trap queue, 95
floating-point (FQ), 47, 61, 96, 211, 243
integer unit, 61
Dekker's algorithm, 326
delay instruction, 19, 35, 138, 141, 144, 150, 157, 216, 302, 306
delayed branch, 63
delayed control transfer, 35, 138
deprecated instructions
BCC, 146
BCS, 146
BE, 146
BG, 146
BGE, 146
BGU, 146
Bicc, 146
BLE, 146
BLEU, 146
BN, 146
BNE, 146
BNEG, 146
BPOS, 146
BVC, 146

BVS, 146
FBE, 140
FBfcc, 140
FBG, 140
FBGE, 140
FBL, 140
FBLE, 140
FBLG, 140
FBN, 140
FBNE, 140
FBO, 140
FBU, 140
FBUE, 140
FBUGE, 140
FBUL, 140
FBULE, 140
LDDA, 180
LDFSR, 173
MULScc, 202
SDIV, 154
SDIVcc, 154
SMULcc, 200
STFSR, 225
SWAP, 234
SWAPA, 235
TSUBccTV, 237, 238
UDIVcc, 154
UMULcc, 200
destination register, 13
dirty bits, see lower and upper registers dirty (DL and DU) fields of FPRS register
disp19 field of instructions, 66, 144, 148
disp22 field of instructions, 66, 141, 147
disp30 field of instructions, 66, 151
disrupting traps, 95, 96, 97, 98, 254
divide instructions, 19, 154, 199
divide-by-zero mask (DZM) bit of TEM field of FSR register, 48
division_by_zero exception, 77, 98, 104, 114, 156, 199
division-by-zero accrued (dza) bit of aexc field of FSR register, 49
division-by-zero current (dzc) bit of cexc field of FSR register, 49
DL, see lower registers dirty (DL) field of FPRS register
DONE instruction, 20, 41, 42, 89, 91, 95
doublet, 9
doubleword, 9, 17, 69, 121
addressing, 70, 72
in memory, 35
doubleword data format, 23
DU, see upper registers dirty (DU) field of FPRS register
dza, see division-by-zero accrued (dza) bit of aexc field of FSR register
dzc, see division-by-zero current (dzc) bit of cexc field of FSR register
DZM, see divide-by-zero mask (DZM) bit of TEM field of FSR register

E
emulating multiple unsigned condition codes, 81

enable floating-point (FEF) field of FPRS register, 42, 53, 84, 99, 114, 142, 145, 174, 176, 226, 227, 243
enable floating-point (PEF) field of PSTATE register, 42, 53, 84, 99, 114, 142, 145, 174, 176, 226, 227, 309
enable RED_state (RED) field of PSTATE register, 91
error_state processor state, 56, 90, 91, 94, 105, 106, 109, 110, 111, 112, 117, 255
exceptions, 21, 89
async_data_error, 113, 133, 153, 174, 177, 179, 181, 182, 183, 226, 228, 230, 232, 234, 236
catastrophic, 98
catastrophic_error, 89, 91, 99, 113, 114, 115
clean_window, 60, 82, 87, 98, 101, 114, 218, 256
data_access_error, 114, 153, 174, 177, 179, 181, 182, 183, 226, 228, 230, 232, 234, 236
data_access_exception, 114, 153, 174, 177, 179, 181, 182, 183, 226, 228, 230, 232, 234, 236
data_access_MMU_miss, 114, 153, 174, 177, 179, 181, 182, 183, 210, 226, 228, 230, 232, 234, 236, 256
data_access_protection, 114, 153, 175, 177, 179, 181, 226, 228, 230, 232
data_protection, 234, 236
division_by_zero, 77, 98, 104, 114, 156, 199
externally_initiated_reset (XIR), 56, 108, 110
fill_n_normal, 98, 114, 216, 218
fill_n_other, 98, 114, 216, 218
floating-point, 10, 99
fp_disabled, 16, 42, 84, 98, 114, 142, 145, 158, 160, 161, 162, 163, 164, 165, 166, 174, 176, 177, 191,
193, 197, 226, 227, 228, 309
fp_exception, 45
fp_exception_ieee_754, 44, 48, 99, 100, 104, 115, 158, 160, 161, 162, 163, 165, 166, 247
fp_exception_other, 40, 47, 61, 85, 104, 115, 158, 160, 161, 162, 163, 164, 165, 166, 174, 177, 191, 193,
213, 226, 228, 247
illegal_instruction, 35, 47, 58, 85, 115, 133, 139, 150, 157, 168, 170, 171, 174, 179, 181, 197, 198, 205,
212, 213, 215, 219, 226, 229, 230, 231, 232, 241, 243, 245, 254, 255, 256
implementation_dependent_n, 91, 104, 255
instruction_access, 97
instruction_access_error, 98, 115, 133
instruction_access_exception, 115, 133
instruction_access_MMU_miss, 115, 133
internal_processor_error, 91, 115, 133
invalid_exception, 161
LDDF_mem_address_not_aligned, 70, 98, 115, 174, 177, 226, 228, 257
LDQF_mem_address_not_aligned, 70, 116, 174, 177, 257
mem_address_not_aligned, 69, 98, 116, 153, 172, 174, 177, 179, 181, 216, 226, 228, 230, 232, 234, 236
persistence, 100
power_on_reset (POR), 108, 116
privileged_action, 51, 73, 97, 116, 153, 176, 177, 181, 183, 215, 227, 228, 232, 236
privileged_instruction (SPARC-V8), 116
privileged_opcode, 98, 116, 157, 212, 215, 219, 243, 245
r_register_access_error (SPARC-V8), 115
software_initiated reset (SIR), 105
software_initiated_reset, 97
software_initiated_reset (SIR), 97, 111, 116, 223
spill_n_normal, 98, 116, 169, 218
spill_n_other, 116, 169, 218
STDF_mem_address_not_aligned, 70, 98, 116, 226, 228, 257
STQF_mem_address_not_aligned, 70, 116, 226, 228, 257
tag_overflow, 77, 98, 104, 117, 237, 239
trap_instruction, 98, 117, 241
unimplemented_LDD, 98, 117, 179, 181, 257
unimplemented_STD, 98, 117, 230, 232, 257

watchdog_reset (WDR), 108
window_fill, 58, 59, 60, 82, 216, 305
window_overflow, 301
window_spill, 58, 60, 305

exceptions, also see trap types
execute protection, 282
execute unit, 123
execute_state, 90, 105, 106, 110, 111, 117
extended word, 10
extended word addressing, 70, 72
extended word data format, 23
extensions
architectural, 7, 321
externally_initiated_reset (XIR), 56, 91, 93, 97, 108, 110, 111

F
f registers, 16, 36, 100, 247, 255
FABSd instruction, 164, 275, 276, 277
FABSq instruction, 164, 275, 276, 277
FABSs instruction, 164, 275
FADDd instruction, 158, 275
FADDq instruction, 158, 275
FADDs instruction, 158, 275
FBA instruction, 141, 278
FBE instruction, 140, 278
FBfcc instructions, 35, 43, 84, 99, 114, 140, 142, 273, 278
FBG instruction, 140, 278
FBGE instruction, 140, 278
FBL instruction, 140, 278
FBLE instruction, 140, 278
FBLG instruction, 140, 278
FBN instruction, 140, 141, 278
FBNE instruction, 140, 278
FBO instruction, 140, 278
FBPA instruction, 143, 144, 278
FBPcc instructions, 66
FBPE instruction, 143, 278
FBPfcc instructions, 35, 43, 66, 67, 84, 99, 142, 143, 273, 278
FBPG instruction, 143, 278
FBPGE instruction, 143, 278
FBPL instruction, 143, 278
FBPLE instruction, 143, 278
FBPLG instruction, 143, 278
FBPN instruction, 143, 144, 278
FBPNE instruction, 143, 278
FBPO instruction, 143, 278
FBPU instruction, 143, 278
FBPUE instruction, 143, 278
FBPUG instruction, 143, 278
FBPUGE instruction, 143, 278
FBPUL instruction, 143, 278
FBPULE instruction, 143, 278
FBU instruction, 140, 278

FBUE instruction, 140, 278
FBUG instruction, 140, 278
FBUGE instruction, 140, 278
FBUL instruction, 140, 278
FBULE instruction, 140, 278
fcc, see floating-point condition codes (fcc) fields of FSR register
fcc-conditional branches, 141, 144
fccN, 10
FCMP* instructions, 43, 159
FCMPd instruction, 159, 248, 277
FCMPE* instructions, 43, 159
FCMPEd instruction, 159, 248, 277
FCMPEq instruction, 159, 248, 277
FCMPEs instruction, 159, 248, 277
FCMPq instruction, 159, 248, 277
FCMPs instruction, 159, 248, 277
fcn field of instructions, 157, 206
FDIVd instruction, 165, 275
FDIVq instruction, 165, 275
FDIVs instruction, 165, 275
FdMULq instruction, 165, 275
FdTOi instruction, 161, 250, 275
FdTOq instruction, 162, 248, 275
FdTOs instruction, 162, 248, 275
FdTOx instruction, 161, 275, 276, 277
FEF, see enable floating-point (FEF) field of FPRS register
fill register window, 33, 58, 59, 82, 83, 86, 87, 88, 114, 217, 218, 219, 316
fill_n_normal exception, 98, 114, 216, 218
fill_n_other exception, 98, 114, 216, 218
FiTOd instruction, 163, 275
FiTOq instruction, 163, 275
FiTOs instruction, 163, 275
floating-point add and subtract instructions, 158
floating-point compare instructions, 43, 159, 159, 248
floating-point condition code bits, 141
floating-point condition codes (fcc) fields of FSR register, 43, 46, 100, 141, 144, 159, 247, 292
floating-point data type, 23
floating-point deferred-trap queue (FQ), 47, 61, 96, 211, 212, 243, 254
floating-point enable (FEF) field of FPRS register, 309
floating-point exception, 10, 99
floating-point move instructions, 164
floating-point multiply and divide instructions, 165
floating-point operate (FPop) instructions, 10, 20, 36, 45, 48, 67, 84, 99, 114, 115, 174
floating-point queue, see floating-point deferred-trap queue (FQ)
floating-point registers, 40, 247, 255, 304
floating-point registers state (FPRS) register, 42, 215, 245
floating-point square root instructions, 166
floating-point state (FSR) register, 43, 48, 50, 174, 225, 226, 247, 250, 254
floating-point trap type (ftt) field of FSR register, 10, 43, 45, 48, 61, 84, 85, 115, 212, 226, 247
floating-point trap types
fp_disabled, 53
FPop_unfinished, 85
FPop_unimplemented, 85
hardware_error, 10, 45, 47

IEEE_754_exception, 10, 46, 46, 48, 50, 100, 115, 247
invalid_fp_register, 10, 40, 46, 158, 160, 161, 162, 163, 164, 165, 166, 174, 177, 191, 193, 226, 228

numeric values, 45
sequence_error, 45, 46, 47, 61, 212, 213
unfinished_FPop, 10, 46, 46, 50, 247, 253
unimplemented_FPop, 10, 46, 46, 50, 85, 191, 212, 247, 253
floating-point traps
deferred, 212
precise, 212
floating-point unit (FPU), 10, 16
FLUSH instruction, 131, 167, 253, 258, 308, 324
in multiprocess environment, 132
flush instruction memory instruction, 167, 324
FLUSH latency, 258
flush register windows instruction, 169
FLUSHW instruction, 20, 83, 86, 87, 116, 169, 303
FMOVA instruction, 188
FMOVCC instruction, 188
FMOVcc instructions, 41, 42, 43, 66, 67, 80, 84, 188, 191, 196, 197, 278
FMOVccd instruction, 277
FMOVccq instruction, 277
FMOVccs instruction, 277
FMOVCS instruction, 188
FMOVd instruction, 164, 275, 276, 277
FMOVE instruction, 188
FMOVFA instruction, 188
FMOVFE instruction, 188
FMOVFG instruction, 188
FMOVFGE instruction, 188
FMOVFL instruction, 188
FMOVFLE instruction, 188
FMOVFLG instruction, 188
FMOVFN instruction, 188
FMOVFNE instruction, 188
FMOVFO instruction, 188
FMOVFU instruction, 188
FMOVFUE instruction, 188
FMOVFUG instruction, 188
FMOVFUGE instruction, 188
FMOVFUL instruction, 188
FMOVFULE instruction, 188
FMOVG instruction, 188
FMOVGE instruction, 188
FMOVGU instruction, 188
FMOVL instruction, 188
FMOVLE instruction, 188
FMOVLEU instruction, 188
FMOVN instruction, 188
FMOVNE instruction, 188
FMOVNEG instruction, 188
FMOVPOS instruction, 188
FMOVq instruction, 164, 275, 276, 277
FMOVr instructions, 67, 84, 192
FMOVRGEZ instruction, 192

FMOVRGZ instruction, 192
FMOVRLEZ instruction, 192
FMOVRLZ instruction, 192
FMOVRNZ instruction, 192
FMOVRZ instruction, 192
FMOVs instruction, 164, 275
FMOVVC instruction, 188
FMOVVS instruction, 188
FMULd instruction, 165, 275
FMULq instruction, 165, 275
FMULs instruction, 165, 275
FNEGd instruction, 164, 275, 276, 277
FNEGq instruction, 164, 275, 276, 277
FNEGs instruction, 164, 275
formats
instruction, 63
fp_disabled floating-point trap type, 16, 42, 53, 84, 98, 114, 142, 145, 158, 160, 161, 162, 163, 164, 165, 166,
174, 176, 177, 191, 193, 197, 226, 227, 228, 309
fp_exception exception, 45, 48
fp_exception_ieee_754 exception, 44, 48, 99, 100, 104, 115, 158, 160, 161, 162, 163, 165, 166, 247
fp_exception_other exception, 40, 47, 61, 85, 104, 115, 158, 160, 161, 162, 163, 164, 165, 166, 174, 177, 191,
193, 212, 213, 226, 228, 247
FPop instructions, see floating-point operate (FPop) instructions
FPop_unimplemented floating-point trap type, 85
FPop1 instructions, 10
FPop2 instructions, 10
FPRS, see floating-point register state (FPRS) register
FPU, see floating-point unit
FQ, see floating-point deferred-trap queue (FQ)
FqTOd instruction, 162, 248, 275
FqTOi instruction, 161, 250, 275
FqTOs instruction, 162, 248, 275
FqTOx instruction, 161, 275, 276, 277
frame pointer register, 302
freg, 292
FsMULd instruction, 165, 275
FSQRTd instruction, 166, 275
FSQRTq instruction, 166, 275
FSQRTs instruction, 166, 275
FsTOd instruction, 162, 248, 275
FsTOi instruction, 161, 250, 275
FsTOq instruction, 162, 248, 275
FsTOx instruction, 161, 275, 276, 277
FSUBd instruction, 158, 275
FSUBq instruction, 158, 275
FSUBs instruction, 158, 275
ftt, see floating-point trap type (ftt) field of FSR register
function return value, 302
functional choice
implementation-dependent, 252
FxTOd instruction, 163, 275, 276, 277
FxTOq instruction, 163, 275, 276, 277
FxTOs instruction, 163, 275, 276, 277

G
generating constants, 220
global registers, 4, 15, 30, 30, 30, 303

H
halfword, 10, 17, 69, 121
addressing, 70, 72
data format, 23
halt, 105
hardware
dependency, 251
traps, 101
hardware_error floating-point trap type, 10, 45, 47
has, 7
hexlet, 10

I
i field of instructions, 66, 137, 154, 167, 169, 172, 173, 176, 178, 180, 182, 183, 184, 195, 198, 199, 200, 202,
205, 206, 214, 216
I/O, see input/output (I/O)
i_or_x_cc, 292
icc field of CCR register, 41, 42, 137, 147, 149, 155, 156, 184, 196, 200, 202, 203, 233, 237, 241
icc-conditional branches, 147
IE, see interrupt enable (IE) field of PSTATE register
IEEE Std 754-1985, 10, 15, 44, 46, 48, 50, 85, 247, 249, 250, 253, 254
IEEE_754_exception floating-point trap type, 10, 46, 46, 48, 50, 100, 115, 247
IER register (SPARC-V8), 245
illegal_instruction exception, 35, 47, 58, 85, 115, 133, 139, 150, 157, 168, 170, 171, 174, 179, 181, 197, 198,
205, 210, 212, 213, 215, 219, 226, 229, 230, 231, 232, 241, 243, 245, 254, 255, 256
ILLTRAP instruction, 115, 170, 273
imm_asi field of instructions, 67, 73, 152, 173, 176, 178, 180, 182, 183, 206
imm22 field of instructions, 67
IMPDEPn instructions, see implementation-dependent (IMPDEPn) instructions
IMPL, 171
impl field of VER register, 45
impl_dep (PID) fields of PSTATE register, 52
implementation, 10
implementation dependency, 7, 251
implementation note, 4
implementation number (impl) field of VER register, 57
implementation_dependent_n exception, 91, 104, 115, 255
implementation-dependent, 10
assigned value (a), 252
functional choice (c), 252
total unit (t), 252
trap, 108
value (v), 252
implementation-dependent (IMPDEPn) instructions, 85, 171, 257, 321
implicit
ASI, 73
byte order, 52

in registers, 15, 30, 33, 217, 301
INC synthetic instruction, 299
INCcc synthetic instruction, 299
inexact accrued (nxa) bit of aexc field of FSR register, 49, 250
inexact current (nxc) bit of cexc field of FSR register, 48, 49, 249, 250
inexact mask (NXM) bit of TEM field of FSR register, 48, 48
inexact quotient, 154, 155
infinity, 250
initiated, 11
input/output (I/O), 6, 18
input/output (I/O) locations, 120, 121, 130, 253, 258
order, 121
value semantics, 121
instruction
access in RED_state, 92
alignment, 17, 69, 121
cache, 125
dispatch, 98
execution, 98
fetch, 69
formats, 4, 63
memory, 131
reordering, 124
instruction fields, 11
a, 66, 138, 141, 147, 148, 152
cc0, 66, 144, 148, 159, 195
cc1, 66, 144, 148, 159, 195
cc2, 66, 195
cond, 66, 141, 144, 147, 148, 189, 195
const22, 170
d16hi, 66, 138
d16lo, 66, 138
disp19, 66, 144, 148
disp22, 66, 141, 147
disp30, 66, 151
fcn, 157, 206
i, 66, 137, 154, 167, 169, 172, 173, 176, 178, 180, 182, 183, 184, 195, 198, 199, 200, 202, 205, 206, 214,
216
imm_asi, 67, 73, 152, 173, 176, 178, 180, 206
imm22, 67
mmask, 67, 224
op3, 67, 137, 152, 154, 157, 167, 169, 172, 173, 176, 178, 180, 182, 183, 184, 199, 200, 202, 206, 211,
214, 216
opf, 67, 158, 159, 161, 162, 163, 164, 165, 166
opf_cc, 67, 189
opf_low, 67, 189, 192
p, 67, 138, 139, 144, 148
rcond, 67, 138, 192, 198
rd, 13, 68, 137, 152, 154, 158, 161, 162, 163, 164, 165, 166, 172, 173, 176, 178, 180, 182, 183, 184, 189,
192, 195, 198, 199, 200, 202, 205, 211, 214, 321
reg_or_imm, 321
reserved, 133
rs1, 13, 68, 137, 138, 152, 154, 158, 159, 165, 167, 172, 173, 176, 178, 180, 182, 183, 184, 192, 198,
199, 200, 202, 206, 211, 214, 216, 321

rs2, 13, 68, 137, 152, 154, 158, 159, 161, 162, 163, 164, 165, 166, 167, 172, 173, 176, 178, 180, 182,
183, 184, 189, 192, 195, 198, 199, 200, 202, 205, 206, 216
simm10, 68, 198
simm11, 68, 195
simm13, 68, 137, 154, 167, 172, 173, 176, 178, 180, 182, 183, 184, 199, 200, 202, 205, 206, 216
sw_trap#, 68
undefined, 171
instruction set architecture, 5, 10, 11
instruction_access exception, 97
instruction_access_error exception, 98, 115, 133
instruction_access_exception exception, 115, 133
instruction_access_MMU_miss exception, 115, 133
instructions
atomic, 152
atomic load-store, 69, 98, 152, 182, 183, 234, 235
branch if contents of integer register match condition, 138
branch on floating-point condition codes, 140
branch on floating-point condition codes with prediction, 143
branch on integer condition codes, 146
branch on integer condition codes with prediction, 148
compare and swap, 98, 152
comparison, 76, 233
conditional move, 20
control-transfer (CTIs), 19, 157
convert between floating-point formats, 162, 248
convert floating-point to integer, 161, 250
convert integer to floating-point, 163
divide, 19, 154, 199
floating-point add and subtract, 158
floating-point compare, 43, 159, 159, 248
floating-point move, 164
floating-point multiply and divide, 165
floating-point operate (FPop), 20, 45, 48, 99, 174
floating-point square root, 166
flush instruction memory, 167, 324
flush register windows, 169
implementation-dependent (IMPDEPn), 85, 171
jump and link, 19, 172
load, 323
load floating-point, 69, 173
load floating-point from alternate space, 176
load integer, 69, 178
load integer from alternate space, 180
load-store unsigned byte, 98, 152, 182, 234, 235
load-store unsigned byte to alternate space, 183
logical, 184
move floating-point register if condition is true, 188
move floating-point register if contents of integer register satisfy condition, 192
move integer register if condition is satisfied, 194
move integer register if contents of integer register satisfies condition, 198
move on condition, 5
multiply, 19, 199, 200, 200
multiply step, 19, 202
prefetch data, 206

read privileged register, 211
read state register, 20, 214
register window management, 20
reserved, 85
reserved fields, 133
shift, 19, 221
software-initiated reset, 223
store, 323
store floating point, 69
store floating-point, 225
store floating-point into alternate space, 227
store integer, 69, 229, 231
subtract, 233
swap r register with alternate space memory, 235
swap r register with memory, 234
synthetic, 297
tagged add, 237
tagged arithmetic, 19
test-and-set, 131
timing, 133
trap on integer condition codes, 240
write privileged register, 242
write state register, 244
integer condition codes, see icc field of CCR register
integer divide instructions, see divide instructions
integer multiply instructions, see multiply instructions
integer unit (IU), 11, 11, 15
integer unit deferred-trap queue, 61
internal_processor_error exception, 91, 115, 133
and RED_state, 93
interrupt enable (IE) field of PSTATE register, 54, 96, 99, 115
interrupt level, 54
interrupt request, 11, 21, 89, 133
interrupts, 54
invalid accrued (nva) bit of aexc field of FSR register, 49
invalid current (nvc) bit of cexc field of FSR register, 49, 250
invalid mask (NVM) bit of TEM field of FSR register, 48
invalid_exception exception, 161
invalid_fp_register floating-point trap type, 10, 40, 46, 158, 160, 161, 162, 163, 164, 165, 166, 174, 177, 191,
193, 226, 228
IPREFETCH synthetic instruction, 297
ISA, see instruction set architecture
issue unit, 123, 124
issued, 11
italic font
in assembly language syntax, 291
IU, see integer unit

J
JMP synthetic instruction, 297
JMPL instruction, 19, 33, 35, 116, 172, 216, 297, 304
jump and link instruction, 19, 172

L
LD instruction (SPARC-V8), 179
LDA instruction (SPARC-V8), 181
LDD instruction, 35, 98, 117, 178, 257
LDDA instruction, 35, 61, 98, 180, 257
LDDF instruction, 70, 98, 115, 173
LDDF_mem_address_not_aligned exception, 70, 98, 115, 174, 177, 228, 257
LDDFA instruction, 70, 98, 176
LDF instruction, 173
LDFSR instruction, 43, 45, 48, 50, 173
LDQF instruction, 70, 116, 173
LDQF_mem_address_not_aligned exception, 70, 116, 174, 177, 257
LDQFA instruction, 70, 176
LDSB instruction, 178
LDSBA instruction, 180
LDSH instruction, 178
LDSHA instruction, 180
LDSTUB insruction, 69
LDSTUB instruction, 98, 127, 131, 182, 327
LDSTUBA instruction, 98, 183
LDSW instruction, 178
LDSWA instruction, 180
LDUB instruction, 178
LDUBA instruction, 180
LDUH instruction, 178
LDUHA instruction, 180
LDUW instruction, 178
LDUWA instruction, 180
LDX instruction, 98, 178
LDXA instruction, 98, 180
LDXFSR instruction, 43, 45, 48, 50, 173
leaf procedure, 11, 82, 304, 304
optimization, 305, 306
Level I compliance, 7
Level II compliance, 8
little-endian byte order, 11, 17, 52
load floating-point from alternate space instructions, 176
load floating-point instructions, 173
load instructions, 69, 323
load integer from alternate space instructions, 180
load integer instructions, 178
LoadLoad MEMBAR relationship, 127, 187
loads
non-faulting, 123, 123
loads from alternate space, 18, 50, 73, 341
load-store alignment, 17, 69, 121
load-store instructions, 17, 98
compare and swap, 98, 152
load-store unsigned byte, 152, 182, 234, 235
load-store unsigned byte to alternate space, 183
swap r register with alternate space memory, 235
swap r register with memory, 152, 234
LoadStore MEMBAR relationship, 127, 128, 187

local registers, 15, 30, 33, 217, 302, 307
logical instructions, 184
Lookaside MEMBAR relationship, 187
lower registers dirty (DL) field of FPRS register, 42

M
manual
audience, 1
fonts, 3
where to start, 1
manufacturer (manuf) field of VER register, 57, 256
mask number (mask) field of VER register, 57
maximum trap levels (maxtl) field of VER register, 57
MAXTL, 54, 90, 106, 223
maxtl, see maximum trap levels (maxtl) field of VER register
may, 11
mem_address_not_aligned exception, 69, 98, 116, 153, 172, 174, 177, 179, 181, 216, 226, 228, 230, 232, 234,
236
MEMBAR instruction, 67, 76, 121, 125, 126–128, 129, 131, 167, 186, 214, 224, 324
membar_mask, 295
MemIssue MEMBAR relationship, 187
memory
alignment, 121
atomicity, 258
coherence, 120, 121, 258
coherency unit, 122
data, 131
instruction, 131
ordering unit, 121
page, 281
real, 120, 121
stack layout, 304
memory access instructions, 17
memory management unit (MMU), 6, 114, 115, 253, 291
address translation, 286
ASI input, 283
atomic input, 283
context, 281
Data / Instr input, 283
diagram, 283
disabled, 207
disabling, 282
fault address, 288
fault status, 288
in RED_state, 92
memory protection, 286
modified statistics, 282, 288
NF-Load_violation output, 285
No_translation output, 284
Non-faultable attribute, 284
Nucleus Context, 287
Prefetch input, 283
Prefetch_violation output, 285

Prefetchable attribute, 284
Primary Context, 286
Privilege input, 283
Privilege_violation output, 285, 286
Protection_violation output, 285, 286
Read / Write input, 283
Read, Write, and Execute attributes, 284
RED_state, 92, 288
RED_state input, 283
referenced statistics, 282, 288
Restricted attribute, 284
Secondary Context, 286
Translation_error output, 284
Translation_not_valid output, 284
Translation_successful output, 285
memory model, 119–132
barrier synchronization, 333, 334
Dekker's algorithm, 326
issuing order, 330
mode control, 129
mutex (mutual exclusion) locks, 326
operations, 323
overview, 119
partial store order (PSO), 119, 128, 130, 257, 323
portability and recommended programming style, 324
processors and processes, 324
programming with, 323–335
relaxed memory order (RMO), 119, 128, 130, 257, 323
sequential consistency, 120
SPARC-V9, 128
spin lock, 327
strong, 120
strong consistency, 120, 325, 330
total store order (TSO), 92, 119, 129, 130, 323
weak, 120
memory operations
atomic, 130
memory order, 125
program order, 124
memory reference instructions
data flow order constraints, 124
memory_model (MM) field of PSTATE register, 52, 92, 125, 129, 130, 258
microkernel, 317
MM, see memory_model (MM) field of PSTATE register
mmask field of instructions, 67, 224
MMU, see memory management unit (MMU)
mode
nonprivileged, 6, 15, 75
privileged, 15, 51, 85, 122
user, 30, 50, 303
MOV synthetic instruction, 299
MOVA instruction, 194
MOVCC instruction, 194
MOVcc instructions, 41, 42, 43, 66, 68, 80, 191, 194, 196, 197, 278

MOVCS instruction, 194
move floating-point register if condition is true, 188
move floating-point register if contents of integer register satisfy condition, 192
MOVE instruction, 194
move integer register if condition is satisfied instructions, 194
move integer register if contents of integer register satisfies condition instructions, 198
move on condition instructions, 5
MOVFA instruction, 194
MOVFE instruction, 194
MOVFG instruction, 194
MOVFGE instruction, 194
MOVFL instruction, 194
MOVFLE instruction, 194
MOVFLG instruction, 194
MOVFN instruction, 194
MOVFNE instruction, 194
MOVFO instruction, 194
MOVFU instruction, 194
MOVFUE instruction, 194
MOVFUG instruction, 194
MOVFUGE instruction, 194
MOVFUL instruction, 194
MOVFULE instruction, 194
MOVG instruction, 194
MOVGE instruction, 194
MOVGU instruction, 194
MOVL instruction, 194
MOVLE instruction, 194
MOVLEU instruction, 194
MOVN instruction, 194
MOVNE instruction, 194
MOVNEG instruction, 194
MOVPOS instruction, 194
MOVr instruction, 67
MOVr instructions, 68, 81, 198
MOVRGEZ instruction, 198
MOVRGZ instruction, 198
MOVRLEZ instruction, 198
MOVRLZ instruction, 198
MOVRNZ instruction, 198
MOVRZ instruction, 198
MOVVC instruction, 194
MOVVS instruction, 194
MULScc (multiply step) instruction, 19, 202
multiple unsigned condition codes
emulating, 81
multiply instructions, 19, 199, 200, 200
multiply step instruction, see MULScc (multiply step) instruction
multiply/divide register, see Y register
multiprocessor synchronization instructions, 5, 152, 234, 235
multiprocessor system, 5, 125, 167, 208, 210, 234, 235, 258
MULX instruction, 199
must, 11
mutex (mutual exclusion) locks, 326

N
N condition code bit, see negative (N) bit of condition fields of CCR
NaN (not-a-number), 161, 248, 250
quiet, 159, 160, 248
signaling, 43, 159, 160, 162, 248
NEG synthetic instruction, 299
negative (N) bit of condition fields of CCR, 41
negative infinity, 250
nested traps, 5
next program counter (nPC), 11, 21, 35, 35, 55, 63, 95, 97, 157, 204, 318
non-faulting load, 11, 123, 123, 123
non-leaf routine, 172
nonprivileged
mode, 6, 9, 12, 15, 45, 75
registers, 30
software, 42
nonprivileged trap (NPT) field of TICK register, 51, 215
nonstandard floating-point (NS) field of FSR register, 44, 44, 250, 254
nonstandard modes
in FPU, 44
non-virtual memory, 209
NOP instruction, 141, 144, 147, 204, 206, 220, 241
normal traps, 90, 101, 106, 106, 108
NOT synthetic instruction, 299
note
compatibility, 4
implementation, 4
programming, 4
nPC, see next program counter (nPC)
NPT, see nonprivileged trap (NPT) field of TICK register)
NS, see nonstandard floating-point (NS) field of FSR register
Nucleus Context, 122, 287
number of windows (maxwin) field of VER register, 58, 87
nva, see invalid accrued (nva) bit of aexc field of FSR register
nvc, see invalid current (nvc) bit of cexc field of FSR register
NVM, see invalid mask (NVM) bit of TEM field of FSR register
NWINDOWS, 12, 15, 32, 33, 58, 217, 218, 253, 259
nxa, see inexact accrued (nxa) bit of aexc field of FSR register
nxc, see inexact current (nxc) bit of cexc field of FSR register
NXM, see inexact mask (NXM) bit of TEM field of FSR register

O
object-oriented programming, 6
octlet, 12
ofa, see overflow accrued (ofa) bit of aexc field of FSR register
ofc, see overflow current (ofc) bit of cexc field of FSR register
OFM, see overflow mask (OFM) bit of TEM field of FSR register
op3 field of instructions, 67, 137, 152, 154, 157, 167, 169, 172, 173, 176, 178, 180, 182, 183, 184, 199, 200,
202, 206, 211, 214, 216
opcode, 12
reserved, 321
opf field of instructions, 67, 158, 159, 161, 162, 163, 164, 165, 166

opf_cc field of instructions, 67, 189
opf_low field of instructions, 67, 189, 192
optimized leaf procedure, see leaf procedure (optimized)
OR instruction, 184, 299
ORcc instruction, 184, 297
ordering unit
memory, 121
ORN instruction, 184
ORNcc instruction, 184
other windows (OTHERWIN) register, 58, 59, 60, 83, 86, 87, 169, 211, 218, 242, 259, 317
out register #7, 34, 151, 172, 215
out registers, 15, 30, 33, 217, 301
overflow, 86
window, 316
overflow (V) bit of condition fields of CCR, 41, 77
overflow accrued (ofa) bit of aexc field of FSR register, 49
overflow current (ofc) bit of cexc field of FSR register, 48, 49
overflow mask (OFM) bit of TEM field of FSR register, 48, 48

P
p field of instructions, 67, 138, 139, 144, 148
page attributes, 281
page descriptor cache (PDC), 114, 115
page fault, 209
page-level protections, 282
parameters to a subroutine, 302
parity error, 115
partial store order (PSO) memory model, 52, 119, 120, 128, 130, 257, 323
PC, see program counter (PC)
PDC, see page descriptor cache (PDC)
PEF, see enable floating-point (PEF) field of PSTATE register
physical address, 120, 281, 282
PID0, PID1 fields of PSTATE register, 52
PIL, see processor interrupt level (PIL) register
POPC instruction, 205
positive infinity, 250
power failure, 97, 110
power-on reset, 51, 92, 93, 97, 109
power-on reset (POR) trap, 108
power-on_reset, 91
precise floating-point traps, 212
precise trap, 94, 95, 95, 96, 254
predict bit, 139
prefetch
for one read, 208
for one write, 209
for several reads, 208
for several writes, 208
implementation dependent, 209
instruction, 209
page, 209
prefetch data instruction, 206
PREFETCH instruction, 69, 149, 206, 256

prefetch_fcn, 295
PREFETCHA instruction, 206, 256
prefetchable, 12
PRIV, see privileged (PRIV) field of PSTATE register
privileged, 11, 12
mode, 12, 15, 51, 85, 122, 223
registers, 51
software, 6, 33, 45, 53, 73, 101, 169, 256, 288
privileged (PRIV) field of PSTATE register, 14, 53, 116, 122, 153, 176, 183, 215, 227, 232, 235
privileged_action exception, 51, 73, 97, 116, 153, 176, 177, 181, 183, 215, 227, 228, 232, 236
privileged_instruction exception (SPARC-V8), 116
privileged_opcode exception, 98, 116, 157, 212, 215, 219, 243, 245
processor, 12, 15
execute unit, 123
halt, 94, 105
issue unit, 123, 124
model, 123
reorder unit, 123
self-consistency, 124
state diagram, 90
processor interrupt level (PIL) register, 54, 96, 99, 100, 115, 211, 242
processor state (PSTATE) register, 21, 30, 51, 52, 56, 89, 91, 157, 211, 242
processor states
error_state, 56, 91, 94, 105, 106, 109, 110, 111, 112, 117, 255
execute_state, 105, 106, 110, 111, 117
RED_state, 90, 91, 94, 101, 105, 106, 108, 109, 110, 111, 112, 117, 130, 258
program counter (PC), 12, 21, 35, 35, 55, 63, 89, 95, 97, 151, 157, 172, 204, 318
program order, 124, 124
programming note, 4
protection
execute, 282
read, 282
write, 282
PSO, see partial store ordering (PSO) memory model
PSR register (SPARC-V8), 245
PTD, see page table descriptor (PTD)
PTE, see page table entry (PTE)

Q
qne, see queue not empty (qne) field of FSR register
quadlet, 12
quadword, 12, 17, 69, 121
addressing, 71, 73
data format, 23
queue not empty (qne) field of FSR register, 47, 47, 61, 212, 213, 243, 247
quiet NaN (not-a-number), 43, 159, 160, 248

R
r register, 30
#15, 34, 151, 172
alignment, 179, 181

r registers, 253
r_register_access_error exception (SPARC-V8), 115

rcond field of instructions, 67, 138, 192, 198
rd field of instructions, 13, 68, 137, 152, 154, 158, 161, 162, 163, 164, 165, 166, 172, 173, 176, 178, 180, 182,
183, 184, 189, 192, 195, 198, 199, 200, 202, 205, 211, 214, 321
RD, see rounding direction (RD) field of FSR register
RDASI instruction, 214
RDASR instruction, 18, 61, 214, 224, 256, 299, 321
RDCCR instruction, 214
RDFPRS instruction, 214
RDPC instruction, 35, 214
RDPR instruction, 47, 51, 52, 58, 61, 85, 211, 215
RDTICK instruction, 214, 215
RDY instruction, 36, 214, 299
read privileged register instruction, 211
read protection, 282
read state register instructions, 20, 214
read-after-write memory hazard, 124
real memory, 120, 121
real-time software, 308
RED, see enable RED_state (RED) field of PSTATE register
RED_state, 13, 90, 91, 94, 101, 105, 106, 108, 109, 110, 111, 112, 117, 282
and internal_processor_error exception, 93
cache behavior, 92
cache coherence in, 92
data access, 92
instruction access, 92
memory management unit (MMU) in, 92
restricted environment, 92
RED_state (RED) field of PSTATE register, 53, 91, 93
RED_state processor state, 130, 258
RED_state trap table, 101
RED_state trap vector, 91, 92, 258
RED_state trap vector address (RSTVaddr), 258
reference MMU, 6, 291
references, 345
reg, 291
reg_or_imm, 296
reg_or_imm field of instructions, 296, 321
reg_plus_imm, 295
regaddr, 295
register
allocation within a window, 307
destination, 13
renaming mechanism, 124
sets, 29, 33
window usage models, 308
register reference instructions
data flow order constraints, 124
register window management instructions, 20
register windows, 4, 5, 15, 33, 301, 303
clean, 9, 58, 60, 82, 86, 88, 114
fill, 33, 58, 59, 82, 83, 86, 87, 88, 114, 218, 219
spill, 33, 58, 59, 82, 83, 85, 86, 87, 88, 116, 218, 219

registers
address space identifier (ASI), 89, 122, 157, 176, 181, 183, 207, 227, 232, 235, 245, 316
alternate global, 15, 30, 30, 316
ancillary state registers (ASRs), 18, 36, 60, 252, 321
ASI, 50, 56
clean windows (CLEANWIN), 58, 60, 82, 83, 86, 87, 88, 211, 242, 259
clock-tick (TICK), 116
condition codes register (CCR), 56, 89, 137, 157, 202, 245
control and status, 29, 35
current window pointer (CWP), 15, 33, 56, 58, 58, 60, 87, 89, 157, 169, 211, 217, 218, 242, 259
f, 36, 100, 247, 255
floating-point, 16, 40, 255, 304
floating-point deferred-trap queue (FQ), 212
floating-point registers state (FPRS), 42, 215, 245
floating-point state (FSR), 43, 48, 50, 174, 225, 247, 250, 254
frame pointer, 302
global, 4, 15, 30, 30, 30, 303
IER (SPARC-V8), 245
in, 15, 30, 33, 217, 301
input/output (I/O), 18, 252
local, 15, 30, 33, 217, 302, 307
nonprivileged, 30
other windows (OTHERWIN), 58, 59, 60, 83, 86, 87, 169, 211, 218, 242, 259, 317
out, 15, 30, 33, 217, 301
out #7, 34, 151, 172, 215
privileged, 51
processor interrupt level (PIL), 54, 211, 242
processor state (PSTATE), 30, 51, 52, 56, 89, 157, 211, 242
PSR (SPARC-V8), 245
PSTATE, 91
r, 30, 253

r register
#15, 34, 151, 172, 215
restorable windows (CANRESTORE), 16, 33, 58, 59, 60, 82, 83, 86, 87, 211, 218, 219, 242, 259, 317
savable windows (CANSAVE), 16, 33, 58, 59, 82, 83, 86, 87, 169, 211, 218, 219, 242, 259
stack pointer, 301, 303
TBR (SPARC-V8), 245
TICK, 51, 211, 242
trap base address (TBA), 14, 57, 89, 100, 211, 242
trap level (TL), 54, 54, 55, 56, 57, 60, 89, 94, 157, 211, 212, 219, 223, 242, 243
trap next program counter (TNPC), 55, 95, 113, 211, 242
trap program counter (TPC), 55, 95, 113, 211, 212, 242
trap state (TSTATE), 52, 56, 157, 211, 242
trap type (TT), 56, 57, 60, 101, 105, 110, 111, 211, 241, 242, 255
version register (VER), 57, 211
WIM (SPARC-V8), 245
window state (WSTATE), 58, 60, 87, 169, 211, 218, 242, 316, 317
working, 29
Y, 35, 36, 154, 200, 202, 245
relaxed memory order (RMO) memory model, 5, 52, 119, 128, 130, 257, 323
renaming mechanism
register, 124
reorder unit, 123
reordering

instruction, 124
reserved, 13
fields in instructions, 133
instructions, 85
opcodes, 321
reset
externally initiated (XIR), 91, 93, 97, 111
externally_initiated (XIR), 91
externally_initiated_reset (XIR), 56, 110
power_on_reset (POR) trap, 116
power-on, 51, 91, 92, 93, 97, 109
processing, 91
request, 91, 116
reset
trap, 51, 56, 96, 97
software_initiated (SIR), 91
software_initiated_reset (SIR), 97, 111, 116
software-initiated, 93, 97, 105
trap, 13, 51, 95, 97, 105, 255
trap table, 13
watchdog, 56, 91, 93, 94, 97, 109, 110, 111
Reset, Error, and Debug state, 90
restorable windows (CANRESTORE) register, 16, 33, 58, 59, 60, 82, 83, 86, 87, 211, 218, 219, 242, 259, 317
RESTORE instruction, 6, 20, 33, 35, 58, 59, 82, 86, 114, 217, 303, 305, 306, 308
RESTORE synthetic instruction, 297
RESTORED instruction, 20, 83, 88, 218, 219, 316
restricted, 13
restricted address space identifier, 73, 74, 254
RET synthetic instruction, 297, 306
RETL synthetic instruction, 297, 306
RETRY instruction, 20, 41, 42, 88, 89, 91, 95, 96, 97, 157, 218
return address, 302, 305
return from trap (DONE) instruction, see DONE instruction
return from trap (RETRY) instruction, see RETRY instruction
RETURN instruction, 19, 35, 114, 116, 216, 317
RMO, see relaxed memory ordering (RMO) memory model
rounding
in signed division, 155
rounding direction (RD) field of FSR register, 44, 158, 161, 162, 163, 165, 166
routine
non-leaf, 172
rs1 field of instructions, 13, 68, 137, 138, 152, 154, 158, 159, 165, 167, 172, 173, 176, 178, 180, 182, 183, 184,
192, 198, 199, 200, 202, 206, 211, 214, 216, 321
rs2 field of instructions, 13, 68, 137, 152, 154, 158, 159, 161, 162, 163, 164, 165, 166, 167, 172, 173, 176, 178,
180, 184, 189, 192, 195, 198, 199, 200, 202, 205, 206
RSTVaddr, 92, 101, 258

S
savable windows (CANSAVE) register, 16, 33, 58, 59, 82, 83, 86, 87, 169, 211, 218, 219, 242, 259
SAVE instruction, 6, 20, 33, 35, 58, 59, 60, 82, 85, 86, 87, 114, 116, 172, 216, 217, 302, 303, 305, 306, 308
SAVE synthetic instruction, 297
SAVED instruction, 20, 83, 88, 218, 219, 316

SDIV instruction, 36, 154
SDIVcc instruction, 36, 154
SDIVX instruction, 199
self-consistency
processor, 124
self-modifying code, 167, 308
sequence_error floating-point trap type, 10, 45, 46, 47, 61, 115, 212, 213
sequential consistency memory model, 120
SET synthetic instruction, 297
SETHI instruction, 19, 67, 76, 204, 220, 273, 297, 304
shall (special term), 13
shared memory, 119, 325, 326, 327, 332
shift instructions, 19, 76, 221
should (special term), 13
side effects, 13, 120, 121, 123
signal handler, see trap handler
signal monitor instruction, 223
signaling NaN (not-a-number), 43, 159, 160, 162, 248
signed integer data type, 23
sign-extended 64-bit constant, 68
sign-extension, 299
SIGNX synthetic instruction, 299
simm10 field of instructions, 68, 198
simm11 field of instructions, 68, 195
simm13 field of instructions, 68, 137, 154, 167, 172, 173, 176, 178, 180, 182, 183, 184, 199, 200, 202, 205, 206,
216
SIR instruction, 89, 97, 111, 116, 223
SIR, see software_initiated_reset (SIR)
SIR_enable control flag, 223, 258
SLL instruction, 221
SLLX instruction, 221, 297
SMUL instruction, 36, 200
SMULcc instruction, 36, 200
software conventions, 301
software trap, 101, 101, 241
software_initiated_reset (SIR), 91, 97, 105, 108, 111, 116, 223
software_trap_number, 296
software-initiated_reset, 93, 97
SPARC Architecture Committee, 7
SPARC-V8 compatibility, 4, 19, 30, 40, 43, 54, 58, 76, 78, 104, 114, 115, 116, 121, 137, 142, 145, 160, 170,
171, 174, 179, 181, 187, 215, 224, 226, 230, 232, 233, 237, 239, 241, 245, 322, 336
SPARC-V8 compatiblity, 85
SPARC-V9 Application Binary Interface (ABI), 6, 7, 75
SPARC-V9 features, 4
SPARC-V9 memory models, 128
SPARC-V9-NP, 7
special terms
shall, 13
should, 13
special traps, 90, 101
speculative load, 13
spill register window, 33, 58, 59, 82, 83, 85, 86, 87, 88, 116, 218, 219, 316
spill windows, 217
spill_n_normal exception, 98, 116, 169, 218

spill_n_other exception, 116, 169, 218
spin lock, 327
SRA instruction, 221, 299
SRAX instruction, 221
SRL instruction, 221
SRLX instruction, 221
ST instruction, 299
stack frame, 217
stack pointer alignment, 304
stack pointer register, 301, 303
STB instruction, 229, 231, 299
STBA instruction, 229, 231
STBAR instruction, 76, 125, 127, 187, 214, 224
STD instruction, 35, 98, 117, 229, 231, 257
STDA instruction, 35, 61, 98, 229, 231, 257
STDF instruction, 70, 116, 225
STDF_mem_address_not_aligned exception, 70, 98, 116, 226, 228, 257
STDFA instruction, 70, 98, 227
STF instruction, 225
STFSR instruction, 43, 45, 48, 50, 225
STH instruction, 229, 231, 299
STHA instruction, 229, 231
store floating-point instructions, 225
store floating-point into alternate space instructions, 227
store instructions, 69, 323
store integer instructions, 229, 231
StoreLoad MEMBAR relationship, 127, 187
stores to alternate space, 18, 50, 73, 341
StoreStore MEMBAR relationship, 127, 187
STQF instruction, 70, 116, 225
STQF_mem_address_not_aligned exception, 70, 116, 226, 228, 257
STQFA instruction, 70, 227
strong consistency memory model, 120, 325, 330
strong ordering, see strong consistency memory model
STW instruction, 229, 231
STWA instruction, 229, 231
STX instruction, 98, 229, 231
STXA instruction, 98, 229, 231
STXFSR instruction, 43, 45, 48, 50, 225
SUB instruction, 233, 299
SUBC instruction, 233
SUBcc instruction, 76, 233, 297
SUBCcc instruction, 233
subtract instructions, 233
SUBX instruction (SPARC-V8), 233
SUBXcc instruction (SPARC-V8), 233
supervisor software, 13, 18, 30, 31, 46, 47, 48, 61, 89, 95, 105, 111, 243, 249, 253, 301, 315, 316, 317
supervisor-mode trap handler, 101
sw_trap# field of instructions, 68
SWAP instruction, 69, 127, 131, 182, 183, 234, 327
swap r register with alternate space memory instructions, 235
swap r register with memory instructions, 152, 234
SWAPA instruction, 182, 183, 235
Sync MEMBAR relationship, 187

synthetic instructions, 2
BCLR, 299
BSET, 299
BTOG, 299
BTST, 299
CALL, 297
CAS, 299
CASX, 299
CLR, 299
CMP, 233, 297
DEC, 299
DECcc, 299
INC, 299
INCcc, 299
IPREFETCH, 297
JMP, 297
MOV, 299
NEG, 299
NOT, 299
RESTORE, 297
RET, 297, 306
RETL, 297, 306
SAVE, 297
SET, 297
SIGNX, 299
TST, 297
synthetic instructions in assembler, 2, 297
system call, 316
system software, 116, 122, 123, 132, 168, 255, 303, 304, 308, 309, 316, 317

T
TA instruction, 278
TADDcc instruction, 77, 237
TADDccTV instruction, 77, 117, 237
tag overflow, 77
tag_overflow exception, 77, 98, 104, 117, 237, 239
tagged add instructions, 237
tagged arithmetic, 77
tagged arithmetic instructions, 19
tagged word data format, 23
task switching, see context switching
TBR register (SPARC-V8), 245
Tcc instructions, 21, 41, 42, 66, 89, 101, 117, 240, 278
TCS instruction, 278
TE instruction, 278
TEM, see trap enable mask (TEM) field of FSR register
test-and-set instruction, 131
TG instruction, 278
TGE instruction, 278
TGU instruction, 278
threads, see multithreaded software
Ticc instruction (SPARC-V8), 241
TICK, see clock-tick register (TICK)

timing
instruction, 133
tininess (floating-point), 49, 249, 256
TL instruction, 278
TLB, see page descriptor cache (PDC)
TLE instruction, 278
TLE, see trap_little_endian (TLE) field of PSTATE register
TLEU instruction, 278
TN instruction, 278
TNE instruction, 278
TNEG instruction, 278
total order, 126
total store order (TSO) memory model, 52, 92, 119, 129, 130, 323
total unit
implementation-dependent, 252
TPOS instruction, 278
Translation Lookaside Buffer (TLB), see page descriptor cache (PDC)
trap, 14, 21, 21, 89, 302
trap base address (TBA) register, 14, 57, 89, 100, 211, 242
trap categories
deferred, 95, 96, 99
disrupting, 96, 97, 98
precise, 95, 95, 96
reset, 97
trap enable mask (TEM) field of FSR register, 44, 48, 99, 100, 115, 254
trap handler, 157
supervisor-mode, 101
user, 46, 249, 317
trap level, 54
trap level (TL) register, 54, 54, 55, 56, 57, 60, 89, 94, 157, 211, 212, 219, 223, 242, 243
trap model, 97
trap next program counter (TNPC) register, 55, 95, 113, 211, 242
trap on integer condition codes instructions, 240
trap priorities, 104
trap processing, 91, 105
trap program counter (TPC) register, 55, 95, 113, 211, 212, 242
trap stack, 5, 106
trap state (TSTATE) register, 52, 56, 157, 211, 242
trap type (TT) register, 56, 57, 60, 90, 101, 105, 110, 111, 211, 241, 242, 255
trap types, also see exceptions
trap vector
RED_state, 91
trap_instruction exception, 98, 117, 241
trap_little_endian (TLE) field of PSTATE register, 52, 52
traps
also see exceptions
causes, 21
deferred, 95, 254
disrupting, 95, 254
hardware, 101
implementation-dependent, 108
nested, 5
normal, 90, 101, 106, 106, 108
precise, 94, 95, 254

reset, 56, 95, 96, 97, 105, 255
software, 101, 241
software-initiated reset (SIR), 108
special, 90, 101
window fill, 101
window spill, 101
TSO, see total store ordering (TSO) memory model
TST synthetic instruction, 297
TSUBcc instruction, 77
TSUBccTV instruction, 77, 117
TVC instruction, 278
TVS instruction, 278
typewriter font
in assembly language syntax, 291

U
UDIV instruction, 36
UDIVcc instruction, 36, 154
UDIVX instruction, 199
ufa, see underflow accrued (ufa) bit of aexc field of FSR register
ufc, see underflow current (ufc) bit of cexc field of FSR register
UFM, see underflow mask (UFM) bit of TEM field of FSR register
UMUL instruction, 36, 200
UMULcc instruction, 36, 200
unassigned, 14
unconditional branches, 141, 144, 147, 149
undefined, 14
underflow, 86
underflow accrued (ufa) bit of aexc field of FSR register, 49, 249
underflow current (ufc) bit of cexc field of FSR register, 48, 49, 249
underflow mask (UFM) bit of TEM field of FSR register, 48, 48, 49, 249
unfinished_FPop floating-point trap type, 10, 46, 46, 50, 85, 247, 253
UNIMP instruction (SPARC-V8), 170
unimplemented_FPop floating-point trap type, 10, 46, 46, 50, 85, 191, 212, 247, 253
unimplemented_LDD exception, 98, 117, 179, 181, 257
unimplemented_STD exception, 98, 117, 230, 232, 257
unrestricted, 14
unrestricted address space identifier, 74, 254, 317
unsigned integer data type, 23
upper registers dirty (DU) field of FPRS register, 42
user
mode, 30, 48, 50, 223, 303
program, 253
software, 308
trap handler, 46, 249, 317
user application program, see application program

V
V condition code bit, see overflow (V) bit of condition fields of CCR
value
implementation-dependent, 252

value semantics of input/output (I/O) locations, 121
variables
automatic, 302
ver, see version (ver) field of FSR register
version (ver) field of FSR register, 45, 254
version register (VER), 57, 211
virtual address, 120, 281, 282
virtual address aliasing, 288
virtual memory, 209

W
walking the call chain, 303
watchdog reset, 56, 91, 93, 94, 97, 109, 110, 111
watchdog timer, 109
watchdog_reset, 91
watchdog_reset (WDR), 108
WIM register (SPARC-V8), 245
window
clean, 217
window fill exception, 58, 60
window fill trap, 101
window fill trap handler, 20
window overflow, 33, 86, 316
window spill trap, 101
window spill trap handler, 20
window state (WSTATE) register, 58, 60, 87, 169, 211, 218, 242, 316, 317
window underflow, 33, 86
window_fill exception, 59, 82, 216, 305
window_overflow exception, 301
window_spill exception, 58, 60
windows
register, 303
windows, see register windows
word, 14, 17, 69, 121
word data format, 23
WRASI instruction, 244
WRASR instruction, 18, 61, 244, 256, 299, 321
WRCCR instruction, 41, 42, 244
WRFPRS instruction, 243, 244
WRIER instruction (SPARC-V8), 245
write privileged register instruction, 242
write protection, 282
write state register instructions, 244
write-after-read memory hazard, 124
write-after-write memory hazard, 124
WRPR instruction, 51, 52, 58, 85, 91, 242
WRPSR instruction (SPARC-V8), 245
WRTBR instruction (SPARC-V8), 245
WRWIM instruction (SPARC-V8), 245
WRY instruction, 36, 244, 299
WTYPE subfield field of trap type field, 104

X
xcc field of CCR register, 41, 137, 149, 155, 156, 184, 196, 200, 203, 233, 237
XIR, see externally_initiated_reset (XIR)
XNOR instruction, 184, 299
XNORcc instruction, 184
XOR instruction, 184, 299
XORcc instruction, 184

Y
Y register, 35, 36, 154, 200, 202, 245

Z
Z condition code bit, see zero (Z) bit of condition fields of CCR
zero (Z) bit of condition fields of CCR, 41

Source Exif Data:

File Type                       : PDF
File Type Extension             : pdf
MIME Type                       : application/pdf
PDF Version                     : 1.2
Linearized                      : No
Page Count                      : 399
Page Mode                       : UseOutlines
Producer                        : Acrobat Distiller Command 3.01 for Solaris 2.3 and later (SPARC)
Title                           : SPARCV9
Creator                         : FrameMaker 5.5.6.
Modify Date                     : 2003:03:26 10:49:25
Create Date                     : 2003:03:20 14:37:18

EXIF Metadata provided by EXIF.tools

SPARCV9 Manual

Navigation menu

Versions of this User Manual:

Views

Navigation