1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
|
<HTML>
<HEAD>
<TITLE>Berkeley SoftFloat Source Documentation</TITLE>
</HEAD>
<BODY>
<H1>Berkeley SoftFloat Release 3e: Source Documentation</H1>
<P>
John R. Hauser<BR>
2018 January 20<BR>
</P>
<H2>Contents</H2>
<BLOCKQUOTE>
<TABLE BORDER=0 CELLSPACING=0 CELLPADDING=0>
<COL WIDTH=25>
<COL WIDTH=*>
<TR><TD COLSPAN=2>1. Introduction</TD></TR>
<TR><TD COLSPAN=2>2. Limitations</TD></TR>
<TR><TD COLSPAN=2>3. Acknowledgments and License</TD></TR>
<TR><TD COLSPAN=2>4. SoftFloat Package Directory Structure</TD></TR>
<TR><TD COLSPAN=2>5. Issues for Porting SoftFloat to a New Target</TD></TR>
<TR>
<TD></TD>
<TD>5.1. Standard Headers <CODE><stdbool.h></CODE> and
<CODE><stdint.h></CODE></TD>
</TR>
<TR><TD></TD><TD>5.2. Specializing Floating-Point Behavior</TD></TR>
<TR><TD></TD><TD>5.3. Macros for Build Options</TD></TR>
<TR><TD></TD><TD>5.4. Adapting a Template Target Directory</TD></TR>
<TR>
<TD></TD><TD>5.5. Target-Specific Optimization of Primitive Functions</TD>
</TR>
<TR><TD COLSPAN=2>6. Testing SoftFloat</TD></TR>
<TR>
<TD COLSPAN=2>7. Providing SoftFloat as a Common Library for Applications</TD>
</TR>
<TR><TD COLSPAN=2>8. Contact Information</TD></TR>
</TABLE>
</BLOCKQUOTE>
<H2>1. Introduction</H2>
<P>
This document gives information needed for compiling and/or porting Berkeley
SoftFloat, a library of C functions implementing binary floating-point
conforming to the IEEE Standard for Floating-Point Arithmetic.
For basic documentation about SoftFloat refer to
<A HREF="SoftFloat.html"><NOBR><CODE>SoftFloat.html</CODE></NOBR></A>.
</P>
<P>
The source code for SoftFloat is intended to be relatively machine-independent
and should be compilable with any ISO-Standard C compiler that also supports
<NOBR>64-bit</NOBR> integers.
SoftFloat has been successfully compiled with the GNU C Compiler
(<CODE>gcc</CODE>) for several platforms.
</P>
<P>
<NOBR>Release 3</NOBR> of SoftFloat was a complete rewrite relative to
<NOBR>Release 2</NOBR> or earlier.
Changes to the interface of SoftFloat functions are documented in
<A HREF="SoftFloat.html"><NOBR><CODE>SoftFloat.html</CODE></NOBR></A>.
The current version of SoftFloat is <NOBR>Release 3e</NOBR>.
</P>
<H2>2. Limitations</H2>
<P>
SoftFloat assumes the computer has an addressable byte size of either 8 or
<NOBR>16 bits</NOBR>.
(Nearly all computers in use today have <NOBR>8-bit</NOBR> bytes.)
</P>
<P>
SoftFloat is written in C and is designed to work with other C code.
The C compiler used must conform at a minimum to the 1989 ANSI standard for the
C language (same as the 1990 ISO standard) and must in addition support basic
arithmetic on <NOBR>64-bit</NOBR> integers.
Earlier releases of SoftFloat included implementations of <NOBR>32-bit</NOBR>
single-precision and <NOBR>64-bit</NOBR> double-precision floating-point that
did not require <NOBR>64-bit</NOBR> integers, but this option is not supported
starting with <NOBR>Release 3</NOBR>.
Since 1999, ISO standards for C have mandated compiler support for
<NOBR>64-bit</NOBR> integers.
A compiler conforming to the 1999 C Standard or later is recommended but not
strictly required.
</P>
<P>
<NOBR>C Standard</NOBR> header files <CODE><stdbool.h></CODE> and
<CODE><stdint.h></CODE> are required for defining standard Boolean and
integer types.
If these headers are not supplied with the C compiler, minimal substitutes must
be provided.
SoftFloat’s dependence on these headers is detailed later in
<NOBR>section 5.1</NOBR>, <I>Standard Headers <CODE><stdbool.h></CODE>
and <CODE><stdint.h></CODE></I>.
</P>
<H2>3. Acknowledgments and License</H2>
<P>
The SoftFloat package was written by me, <NOBR>John R.</NOBR> Hauser.
<NOBR>Release 3</NOBR> of SoftFloat was a completely new implementation
supplanting earlier releases.
The project to create <NOBR>Release 3</NOBR> (now <NOBR>through 3e</NOBR>) was
done in the employ of the University of California, Berkeley, within the
Department of Electrical Engineering and Computer Sciences, first for the
Parallel Computing Laboratory (Par Lab) and then for the ASPIRE Lab.
The work was officially overseen by Prof. Krste Asanovic, with funding provided
by these sources:
<BLOCKQUOTE>
<TABLE>
<COL>
<COL WIDTH=10>
<COL>
<TR>
<TD VALIGN=TOP><NOBR>Par Lab:</NOBR></TD>
<TD></TD>
<TD>
Microsoft (Award #024263), Intel (Award #024894), and U.C. Discovery
(Award #DIG07-10227), with additional support from Par Lab affiliates Nokia,
NVIDIA, Oracle, and Samsung.
</TD>
</TR>
<TR>
<TD VALIGN=TOP><NOBR>ASPIRE Lab:</NOBR></TD>
<TD></TD>
<TD>
DARPA PERFECT program (Award #HR0011-12-2-0016), with additional support from
ASPIRE industrial sponsor Intel and ASPIRE affiliates Google, Nokia, NVIDIA,
Oracle, and Samsung.
</TD>
</TR>
</TABLE>
</BLOCKQUOTE>
</P>
<P>
The following applies to the whole of SoftFloat <NOBR>Release 3e</NOBR> as well
as to each source file individually.
</P>
<P>
Copyright 2011, 2012, 2013, 2014, 2015, 2016, 2017, 2018 The Regents of the
University of California.
All rights reserved.
</P>
<P>
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are met:
<OL>
<LI>
<P>
Redistributions of source code must retain the above copyright notice, this
list of conditions, and the following disclaimer.
</P>
<LI>
<P>
Redistributions in binary form must reproduce the above copyright notice, this
list of conditions, and the following disclaimer in the documentation and/or
other materials provided with the distribution.
</P>
<LI>
<P>
Neither the name of the University nor the names of its contributors may be
used to endorse or promote products derived from this software without specific
prior written permission.
</P>
</OL>
</P>
<P>
THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS “AS IS”,
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE, ARE
DISCLAIMED.
IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT,
INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE
OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF
ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
</P>
<H2>4. SoftFloat Package Directory Structure</H2>
<P>
Because SoftFloat is targeted to multiple platforms, its source code is
slightly scattered between target-specific and target-independent directories
and files.
The supplied directory structure is as follows:
<BLOCKQUOTE>
<PRE>
doc
source
include
8086
8086-SSE
ARM-VFPv2
ARM-VFPv2-defaultNaN
build
template-FAST_INT64
template-not-FAST_INT64
Linux-386-GCC
Linux-386-SSE2-GCC
Linux-x86_64-GCC
Linux-ARM-VFPv2-GCC
Win32-MinGW
Win32-SSE2-MinGW
Win64-MinGW-w64
</PRE>
</BLOCKQUOTE>
The majority of the SoftFloat sources are provided in the <CODE>source</CODE>
directory.
The <CODE>include</CODE> subdirectory contains several header files
(unsurprisingly), while the other subdirectories of <CODE>source</CODE> contain
source files that specialize the floating-point behavior to match particular
processor families:
<BLOCKQUOTE>
<DL>
<DT><CODE>8086</CODE></DT>
<DD>
Intel’s older, 8087-derived floating-point, extended to all supported
floating-point types
</DD>
<DT><CODE>8086-SSE</CODE></DT>
<DD>
Intel’s x86 processors with Streaming SIMD Extensions (SSE) and later
compatible extensions, having 8087 behavior for <NOBR>80-bit</NOBR>
double-extended-precision (<CODE>extFloat80_t</CODE>) and SSE behavior for
other floating-point types
</DD>
<DT><CODE>ARM-VFPv2</CODE></DT>
<DD>
ARM’s VFPv2 or later floating-point, with NaN payload propagation
</DD>
<DT><CODE>ARM-VFPv2-defaultNaN</CODE></DT>
<DD>
ARM’s VFPv2 or later floating-point, with the “default NaN”
option
</DD>
</DL>
</BLOCKQUOTE>
If other specializations are attempted, these would be expected to be other
subdirectories of <CODE>source</CODE> alongside the ones listed above.
Specialization is covered later, in <NOBR>section 5.2</NOBR>, <I>Specializing
Floating-Point Behavior</I>.
</P>
<P>
The <CODE>build</CODE> directory is intended to contain a subdirectory for each
target platform for which a build of the SoftFloat library may be created.
For each build target, the target’s subdirectory is where all derived
object files and the completed SoftFloat library (typically
<CODE>softfloat.a</CODE> or <CODE>libsoftfloat.a</CODE>) are created.
The two <CODE>template</CODE> subdirectories are not actual build targets but
contain sample files for creating new target directories.
(The meaning of <CODE>FAST_INT64</CODE> will be explained later.)
</P>
<P>
Ignoring the <CODE>template</CODE> directories, the supplied target directories
are intended to follow a naming system of
<NOBR><CODE><<I>execution-environment</I>>-<<I>compiler</I>></CODE></NOBR>.
For the example targets,
<NOBR><CODE><<I>execution-environment</I>></CODE></NOBR> is
<NOBR><CODE>Linux-386</CODE></NOBR>, <NOBR><CODE>Linux-386-SSE2</CODE></NOBR>,
<NOBR><CODE>Linux-x86_64</CODE></NOBR>,
<NOBR><CODE>Linux-ARM-VFPv2</CODE></NOBR>, <CODE>Win32</CODE>,
<NOBR><CODE>Win32-SSE2</CODE></NOBR>, or <CODE>Win64</CODE>, and
<NOBR><CODE><<I>compiler</I>></CODE></NOBR> is <CODE>GCC</CODE>,
<CODE>MinGW</CODE>, or <NOBR><CODE>MinGW-w64</CODE></NOBR>.
</P>
<P>
All of the supplied target directories are merely examples that may or may not
be correct for compiling on any particular system.
Despite requests, there are currently no plans to include and maintain in the
SoftFloat package the build files needed for a great many users’
compilation environments, which can span a huge range of operating systems,
compilers, and other tools.
</P>
<P>
As supplied, each target directory contains two files:
<BLOCKQUOTE>
<PRE>
Makefile
platform.h
</PRE>
</BLOCKQUOTE>
The provided <CODE>Makefile</CODE> is written for GNU <CODE>make</CODE>.
A build of SoftFloat for the specific target is begun by executing the
<CODE>make</CODE> command with the target directory as the current directory.
A completely different build tool can be used if an appropriate
<CODE>Makefile</CODE> equivalent is created.
</P>
<P>
The <CODE>platform.h</CODE> header file exists to provide a location for
additional C declarations specific to the build target.
Every C source file of SoftFloat contains a <CODE>#include</CODE> for
<CODE>platform.h</CODE>.
In many cases, the contents of <CODE>platform.h</CODE> can be as simple as one
or two lines of code.
At the other extreme, to get maximal performance from SoftFloat, it may be
desirable to include in header <CODE>platform.h</CODE> (directly or via
<CODE>#include</CODE>) declarations for numerous target-specific optimizations.
Such possibilities are discussed in the next section, <I>Issues for Porting
SoftFloat to a New Target</I>.
If the target’s compiler or library has bugs or other shortcomings,
workarounds for these issues may also be possible with target-specific
declarations in <CODE>platform.h</CODE>, avoiding the need to modify the main
SoftFloat sources.
</P>
<H2>5. Issues for Porting SoftFloat to a New Target</H2>
<H3>5.1. Standard Headers <CODE><stdbool.h></CODE> and <CODE><stdint.h></CODE></H3>
<P>
The SoftFloat sources make use of standard headers
<CODE><stdbool.h></CODE> and <CODE><stdint.h></CODE>, which have
been part of the ISO C Standard Library since 1999.
With any recent compiler, these standard headers are likely to be supported,
even if the compiler does not claim complete conformance to the latest ISO C
Standard.
For older or nonstandard compilers, substitutes for
<CODE><stdbool.h></CODE> and <CODE><stdint.h></CODE> may need to be
created.
SoftFloat depends on these names from <CODE><stdbool.h></CODE>:
<BLOCKQUOTE>
<PRE>
bool
true
false
</PRE>
</BLOCKQUOTE>
and on these names from <CODE><stdint.h></CODE>:
<BLOCKQUOTE>
<PRE>
uint16_t
uint32_t
uint64_t
int32_t
int64_t
UINT64_C
INT64_C
uint_least8_t
uint_fast8_t
uint_fast16_t
uint_fast32_t
uint_fast64_t
int_fast8_t
int_fast16_t
int_fast32_t
int_fast64_t
</PRE>
</BLOCKQUOTE>
</P>
<H3>5.2. Specializing Floating-Point Behavior</H3>
<P>
The IEEE Floating-Point Standard allows for some flexibility in a conforming
implementation, particularly concerning NaNs.
The SoftFloat <CODE>source</CODE> directory is supplied with some
<I>specialization</I> subdirectories containing possible definitions for this
implementation-specific behavior.
For example, the <CODE>8086</CODE> and <NOBR><CODE>8086-SSE</CODE></NOBR>
subdirectories have source files that specialize SoftFloat’s behavior to
match that of Intel’s x86 line of processors.
The files in a specialization subdirectory must determine:
<UL>
<LI>
whether tininess for underflow is detected before or after rounding by default;
<LI>
how signaling NaNs are distinguished from quiet NaNs;
<LI>
what (if anything) special happens when exceptions are raised;
<LI>
the default generated quiet NaNs;
<LI>
how NaNs are propagated from function inputs to output; and
<LI>
the integer results returned when conversions to integer type raise the
<I>invalid</I> exception.
</UL>
</P>
<P>
As provided, the build process for a target expects to involve exactly
<EM>one</EM> specialization directory that defines <EM>all</EM> of these
implementation-specific details for the target.
A specialization directory such as <CODE>8086</CODE> is expected to contain a
header file called <CODE>specialize.h</CODE>, together with whatever other
source files are needed to complete the specialization.
</P>
<P>
A new build target may use an existing specialization, such as the ones
provided by the <CODE>8086</CODE> and <NOBR><CODE>8086-SSE</CODE></NOBR>
subdirectories.
If a build target needs a new specialization, different from any existing ones,
it is recommended that a new specialization directory be created for this
purpose.
The <CODE>specialize.h</CODE> header file from any of the provided
specialization subdirectories can be used as a model for what definitions are
needed.
</P>
<H3>5.3. Macros for Build Options</H3>
<P>
The SoftFloat source files adapt the floating-point implementation according to
several C preprocessor macros:
<BLOCKQUOTE>
<DL>
<DT><CODE>LITTLEENDIAN</CODE>
<DD>
Must be defined for little-endian machines; must not be defined for big-endian
machines.
<DT><CODE>INLINE</CODE>
<DD>
Specifies the sequence of tokens used to indicate that a C function should be
inlined.
If macro <CODE>INLINE_LEVEL</CODE> is defined with a value of 1 or higher, this
macro must be defined; otherwise, this macro is ignored and need not be
defined.
For compilers that conform to the C Standard’s rules for inline
functions, this macro can be defined as the single keyword <CODE>inline</CODE>.
For other compilers that follow a convention pre-dating the standardization of
<CODE>inline</CODE>, this macro may need to be defined to <CODE>extern</CODE>
<CODE>inline</CODE>.
<DT><CODE>THREAD_LOCAL</CODE>
<DD>
Can be defined to a sequence of tokens that, when appearing at the start of a
variable declaration, indicates to the C compiler that the variable is
<I>per-thread</I>, meaning that each execution thread gets its own separate
instance of the variable.
This macro is used in header <CODE>softfloat.h</CODE> in the declarations of
variables <CODE>softfloat_roundingMode</CODE>,
<CODE>softfloat_detectTininess</CODE>, <CODE>extF80_roundingPrecision</CODE>,
and <CODE>softfloat_exceptionFlags</CODE>.
If macro <CODE>THREAD_LOCAL</CODE> is left undefined, these variables will
default to being ordinary global variables.
Depending on the compiler, possible valid definitions of this macro include
<CODE>_Thread_local</CODE> and <CODE>__thread</CODE>.
</DL>
<DL>
<DT><CODE>SOFTFLOAT_ROUND_ODD</CODE>
<DD>
Can be defined to enable support for optional rounding mode
<CODE>softfloat_round_odd</CODE>.
</DL>
<DL>
<DT><CODE>INLINE_LEVEL</CODE>
<DD>
Can be defined to an integer to determine the degree of inlining requested of
the compiler.
Larger numbers request that more inlining be done.
If this macro is not defined or is defined to a value less <NOBR>than 1</NOBR>
(zero or negative), no inlining is requested.
The maximum effective value is no higher <NOBR>than 5</NOBR>.
Defining this macro to a value greater than 5 is the same as defining it
<NOBR>to 5</NOBR>.
<DT><CODE>SOFTFLOAT_FAST_INT64</CODE>
<DD>
Can be defined to indicate that the build target’s implementation of
<NOBR>64-bit</NOBR> arithmetic is efficient.
For newer <NOBR>64-bit</NOBR> processors, this macro should usually be defined.
For very small microprocessors whose buses and registers are <NOBR>8-bit</NOBR>
or <NOBR>16-bit</NOBR> in size, this macro should usually not be defined.
Whether this macro should be defined for a <NOBR>32-bit</NOBR> processor may
depend on the target machine and the applications that will use SoftFloat.
<DT><CODE>SOFTFLOAT_FAST_DIV32TO16</CODE>
<DD>
Can be defined to indicate that the target’s division operator
<NOBR>in C</NOBR> (written as <CODE>/</CODE>) is reasonably efficient for
dividing a <NOBR>32-bit</NOBR> unsigned integer by a <NOBR>16-bit</NOBR>
unsigned integer.
Setting this macro may affect the performance of function <CODE>f16_div</CODE>.
<DT><CODE>SOFTFLOAT_FAST_DIV64TO32</CODE>
<DD>
Can be defined to indicate that the target’s division operator
<NOBR>in C</NOBR> (written as <CODE>/</CODE>) is reasonably efficient for
dividing a <NOBR>64-bit</NOBR> unsigned integer by a <NOBR>32-bit</NOBR>
unsigned integer.
Setting this macro may affect the performance of division, remainder, and
square root operations other than <CODE>f16_div</CODE>.
</DL>
</BLOCKQUOTE>
</P>
<P>
Following the usual custom <NOBR>for C</NOBR>, for most of these macros (all
except <CODE>INLINE</CODE>, <CODE>THREAD_LOCAL</CODE>, and
<CODE>INLINE_LEVEL</CODE>), the content of any definition is irrelevant;
what matters is a macro’s effect on <CODE>#ifdef</CODE> directives.
</P>
<P>
It is recommended that any definitions of macros <CODE>LITTLEENDIAN</CODE>,
<CODE>INLINE</CODE>, and <CODE>THREAD_LOCAL</CODE> be made in a build
target’s <CODE>platform.h</CODE> header file, because these macros are
expected to be determined inflexibly by the target machine and compiler.
The other five macros select options and control optimization, and thus might
be better located in the target’s Makefile (or its equivalent).
</P>
<H3>5.4. Adapting a Template Target Directory</H3>
<P>
In the <CODE>build</CODE> directory, two <CODE>template</CODE> subdirectories
provide models for new target directories.
Two different templates exist because different functions are needed in the
SoftFloat library depending on whether macro <CODE>SOFTFLOAT_FAST_INT64</CODE>
is defined.
If macro <CODE>SOFTFLOAT_FAST_INT64</CODE> will be defined,
<NOBR><CODE>template-FAST_INT64</CODE></NOBR> is the template to use;
otherwise, <NOBR><CODE>template-not-FAST_INT64</CODE></NOBR> is the appropriate
template.
A new target directory can be created by copying the correct template directory
and editing the files inside.
To avoid confusion, it would be wise to refrain from editing the files within a
template directory directly.
</P>
<H3>5.5. Target-Specific Optimization of Primitive Functions</H3>
<P>
Header file <CODE>primitives.h</CODE> (in directory
<CODE>source/include</CODE>) declares macros and functions for numerous
underlying arithmetic operations upon which many of SoftFloat’s
floating-point functions are ultimately built.
The SoftFloat sources include implementations of all of these functions/macros,
written as standard C code, so a complete and correct SoftFloat library can be
created using only the supplied code for all functions.
However, for many targets, SoftFloat’s performance can be improved by
substituting target-specific implementations of some of the functions/macros
declared in <CODE>primitives.h</CODE>.
</P>
<P>
For example, <CODE>primitives.h</CODE> declares a function called
<CODE>softfloat_countLeadingZeros32</CODE> that takes an unsigned
<NOBR>32-bit</NOBR> integer as an argument and returns the number of the
integer’s most-significant bits that are zeros.
While the SoftFloat sources include an implementation of this function written
in <NOBR>standard C</NOBR>, many processors can perform this same function
directly in only one or two machine instructions.
An alternative, target-specific implementation that maps to those instructions
is likely to be more efficient than the generic C code from the SoftFloat
package.
</P>
<P>
A build target can replace the supplied version of any function or macro of
<CODE>primitives.h</CODE> by defining a macro with the same name in the
target’s <CODE>platform.h</CODE> header file.
For this purpose, it may be helpful for <CODE>platform.h</CODE> to
<CODE>#include</CODE> header file <CODE>primitiveTypes.h</CODE>, which defines
types used for arguments and results of functions declared in
<CODE>primitives.h</CODE>.
When a desired replacement implementation is a function, not a macro, it is
sufficient for <CODE>platform.h</CODE> to include the line
<BLOCKQUOTE>
<PRE>
#define <<I>function-name</I>> <<I>function-name</I>>
</PRE>
</BLOCKQUOTE>
where <NOBR><CODE><<I>function-name</I>></CODE></NOBR> is the name of the
function.
This technically defines <NOBR><CODE><<I>function-name</I>></CODE></NOBR>
as a macro, but one that resolves to the same name, which may then be a
function.
(A preprocessor that conforms to the C Standard is required to limit recursive
macro expansion from being applied more than once.)
</P>
<P>
The supplied header file <CODE>opts-GCC.h</CODE> (in directory
<CODE>source/include</CODE>) provides an example of target-specific
optimization for the GCC compiler.
Each GCC target example in the <CODE>build</CODE> directory has
<BLOCKQUOTE>
<CODE>#include "opts-GCC.h"</CODE>
</BLOCKQUOTE>
in its <CODE>platform.h</CODE> header file.
Before <CODE>opts-GCC.h</CODE> is included, the following macros must be
defined (or not) to control which features are invoked:
<BLOCKQUOTE>
<DL>
<DT><CODE>SOFTFLOAT_BUILTIN_CLZ</CODE></DT>
<DD>
If defined, SoftFloat’s internal
‘<CODE>countLeadingZeros</CODE>’ functions use intrinsics
<CODE>__builtin_clz</CODE> and <CODE>__builtin_clzll</CODE>.
</DD>
<DT><CODE>SOFTFLOAT_INTRINSIC_INT128</CODE></DT>
<DD>
If defined, SoftFloat makes use of GCC’s nonstandard <NOBR>128-bit</NOBR>
integer type <CODE>__int128</CODE>.
</DD>
</DL>
</BLOCKQUOTE>
On some machines, these improvements are observed to increase the speeds of
<CODE>f64_mul</CODE> and <CODE>f128_mul</CODE> by around 20 to 25%, although
other functions receive less dramatic boosts, or none at all.
Results can vary greatly across different platforms.
</P>
<H2>6. Testing SoftFloat</H2>
<P>
SoftFloat can be tested using the <CODE>testsoftfloat</CODE> program by the
same author.
This program is part of the Berkeley TestFloat package available at the Web
page
<A HREF="http://www.jhauser.us/arithmetic/TestFloat.html"><NOBR><CODE>http://www.jhauser.us/arithmetic/TestFloat.html</CODE></NOBR></A>.
The TestFloat package also has a program called <CODE>timesoftfloat</CODE> that
measures the speed of SoftFloat’s floating-point functions.
</P>
<H2>7. Providing SoftFloat as a Common Library for Applications</H2>
<P>
Header file <CODE>softfloat.h</CODE> defines the SoftFloat interface as seen by
clients.
If the SoftFloat library will be made a common library for programs on a
system, the supplied <CODE>softfloat.h</CODE> has a couple of deficiencies for
this purpose:
<UL>
<LI>
As supplied, <CODE>softfloat.h</CODE> depends on another header,
<CODE>softfloat_types.h</CODE>, that is not intended for public use but which
must also be visible to the programmer’s compiler.
<LI>
More troubling, at the time <CODE>softfloat.h</CODE> is included in a C source
file, macros <CODE>SOFTFLOAT_FAST_INT64</CODE> and <CODE>THREAD_LOCAL</CODE>
must be defined, or not defined, consistent with how these macro were defined
when the SoftFloat library was built.
</UL>
In the situation that new programs may regularly <CODE>#include</CODE> header
file <CODE>softfloat.h</CODE>, it is recommended that a custom, self-contained
version of this header file be created that eliminates these issues.
</P>
<H2>8. Contact Information</H2>
<P>
At the time of this writing, the most up-to-date information about SoftFloat
and the latest release can be found at the Web page
<A HREF="http://www.jhauser.us/arithmetic/SoftFloat.html"><NOBR><CODE>http://www.jhauser.us/arithmetic/SoftFloat.html</CODE></NOBR></A>.
</P>
</BODY>
|
