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
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
|
<HTML>
<HEAD>
<TITLE>Berkeley SoftFloat Library Interface</TITLE>
</HEAD>
<BODY>
<H1>Berkeley SoftFloat Release 3e: Library Interface</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. Types and Functions</TD></TR>
<TR><TD></TD><TD>4.1. Boolean and Integer Types</TD></TR>
<TR><TD></TD><TD>4.2. Floating-Point Types</TD></TR>
<TR><TD></TD><TD>4.3. Supported Floating-Point Functions</TD></TR>
<TR>
<TD></TD>
<TD>4.4. Non-canonical Representations in <CODE>extFloat80_t</CODE></TD>
</TR>
<TR><TD></TD><TD>4.5. Conventions for Passing Arguments and Results</TD></TR>
<TR><TD COLSPAN=2>5. Reserved Names</TD></TR>
<TR><TD COLSPAN=2>6. Mode Variables</TD></TR>
<TR><TD></TD><TD>6.1. Rounding Mode</TD></TR>
<TR><TD></TD><TD>6.2. Underflow Detection</TD></TR>
<TR>
<TD></TD>
<TD>6.3. Rounding Precision for the <NOBR>80-Bit</NOBR> Extended Format</TD>
</TR>
<TR><TD COLSPAN=2>7. Exceptions and Exception Flags</TD></TR>
<TR><TD COLSPAN=2>8. Function Details</TD></TR>
<TR><TD></TD><TD>8.1. Conversions from Integer to Floating-Point</TD></TR>
<TR><TD></TD><TD>8.2. Conversions from Floating-Point to Integer</TD></TR>
<TR><TD></TD><TD>8.3. Conversions Among Floating-Point Types</TD></TR>
<TR><TD></TD><TD>8.4. Basic Arithmetic Functions</TD></TR>
<TR><TD></TD><TD>8.5. Fused Multiply-Add Functions</TD></TR>
<TR><TD></TD><TD>8.6. Remainder Functions</TD></TR>
<TR><TD></TD><TD>8.7. Round-to-Integer Functions</TD></TR>
<TR><TD></TD><TD>8.8. Comparison Functions</TD></TR>
<TR><TD></TD><TD>8.9. Signaling NaN Test Functions</TD></TR>
<TR><TD></TD><TD>8.10. Raise-Exception Function</TD></TR>
<TR><TD COLSPAN=2>9. Changes from SoftFloat <NOBR>Release 2</NOBR></TD></TR>
<TR><TD></TD><TD>9.1. Name Changes</TD></TR>
<TR><TD></TD><TD>9.2. Changes to Function Arguments</TD></TR>
<TR><TD></TD><TD>9.3. Added Capabilities</TD></TR>
<TR><TD></TD><TD>9.4. Better Compatibility with the C Language</TD></TR>
<TR><TD></TD><TD>9.5. New Organization as a Library</TD></TR>
<TR><TD></TD><TD>9.6. Optimization Gains (and Losses)</TD></TR>
<TR><TD COLSPAN=2>10. Future Directions</TD></TR>
<TR><TD COLSPAN=2>11. Contact Information</TD></TR>
</TABLE>
</BLOCKQUOTE>
<H2>1. Introduction</H2>
<P>
Berkeley SoftFloat is a software implementation of binary floating-point that
conforms to the IEEE Standard for Floating-Point Arithmetic.
The current release supports five binary formats: <NOBR>16-bit</NOBR>
half-precision, <NOBR>32-bit</NOBR> single-precision, <NOBR>64-bit</NOBR>
double-precision, <NOBR>80-bit</NOBR> double-extended-precision, and
<NOBR>128-bit</NOBR> quadruple-precision.
The following functions are supported for each format:
<UL>
<LI>
addition, subtraction, multiplication, division, and square root;
<LI>
fused multiply-add as defined by the IEEE Standard, except for
<NOBR>80-bit</NOBR> double-extended-precision;
<LI>
remainder as defined by the IEEE Standard;
<LI>
round to integral value;
<LI>
comparisons;
<LI>
conversions to/from other supported formats; and
<LI>
conversions to/from <NOBR>32-bit</NOBR> and <NOBR>64-bit</NOBR> integers,
signed and unsigned.
</UL>
All operations required by the original 1985 version of the IEEE Floating-Point
Standard are implemented, except for conversions to and from decimal.
</P>
<P>
This document gives information about the types defined and the routines
implemented by SoftFloat.
It does not attempt to define or explain the IEEE Floating-Point Standard.
Information about the standard is available elsewhere.
</P>
<P>
The current version of SoftFloat is <NOBR>Release 3e</NOBR>.
This release modifies the behavior of the rarely used <I>odd</I> rounding mode
(<I>round to odd</I>, also known as <I>jamming</I>), and also adds some new
specialization and optimization examples for those compiling SoftFloat.
</P>
<P>
The previous <NOBR>Release 3d</NOBR> fixed bugs that were found in the square
root functions for the <NOBR>64-bit</NOBR>, <NOBR>80-bit</NOBR>, and
<NOBR>128-bit</NOBR> floating-point formats.
(Thanks to Alexei Sibidanov at the University of Victoria for reporting an
incorrect result.)
The bugs affected all prior <NOBR>Release-3</NOBR> versions of SoftFloat
<NOBR>through 3c</NOBR>.
The flaw in the <NOBR>64-bit</NOBR> floating-point square root function was of
very minor impact, causing a <NOBR>1-ulp</NOBR> error (<NOBR>1 unit</NOBR> in
the last place) a few times out of a billion.
The bugs in the <NOBR>80-bit</NOBR> and <NOBR>128-bit</NOBR> square root
functions were more serious.
Although incorrect results again occurred only a few times out of a billion,
when they did occur a large portion of the less-significant bits could be
wrong.
</P>
<P>
Among earlier releases, 3b was notable for adding support for the
<NOBR>16-bit</NOBR> half-precision format.
For more about the evolution of SoftFloat releases, see
<A HREF="SoftFloat-history.html"><NOBR><CODE>SoftFloat-history.html</CODE></NOBR></A>.
</P>
<P>
The functional interface of SoftFloat <NOBR>Release 3</NOBR> and later differs
in many details from the releases that came before.
For specifics of these differences, see <NOBR>section 9</NOBR> below,
<I>Changes from SoftFloat <NOBR>Release 2</NOBR></I>.
</P>
<H2>2. Limitations</H2>
<P>
SoftFloat assumes the computer has an addressable byte size of 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>
Most operations not required by the original 1985 version of the IEEE
Floating-Point Standard but added in the 2008 version are not yet supported in
SoftFloat <NOBR>Release 3e</NOBR>.
</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. Types and Functions</H2>
<P>
The types and functions of SoftFloat are declared in header file
<CODE>softfloat.h</CODE>.
</P>
<H3>4.1. Boolean and Integer Types</H3>
<P>
Header file <CODE>softfloat.h</CODE> depends on standard headers
<CODE><stdbool.h></CODE> and <CODE><stdint.h></CODE> to define type
<CODE>bool</CODE> and several integer types.
These standard headers have been part of the ISO C Standard Library since 1999.
With any recent compiler, they 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, a port of SoftFloat may have substitutes
for these headers.
Header <CODE>softfloat.h</CODE> depends only on the name <CODE>bool</CODE> from
<CODE><stdbool.h></CODE> and on these type names from
<CODE><stdint.h></CODE>:
<BLOCKQUOTE>
<PRE>
uint16_t
uint32_t
uint64_t
int32_t
int64_t
uint_fast8_t
uint_fast32_t
uint_fast64_t
int_fast32_t
int_fast64_t
</PRE>
</BLOCKQUOTE>
</P>
<H3>4.2. Floating-Point Types</H3>
<P>
The <CODE>softfloat.h</CODE> header defines five floating-point types:
<BLOCKQUOTE>
<TABLE CELLSPACING=0 CELLPADDING=0>
<TR>
<TD><CODE>float16_t</CODE></TD>
<TD><NOBR>16-bit</NOBR> half-precision binary format</TD>
</TR>
<TR>
<TD><CODE>float32_t</CODE></TD>
<TD><NOBR>32-bit</NOBR> single-precision binary format</TD>
</TR>
<TR>
<TD><CODE>float64_t</CODE></TD>
<TD><NOBR>64-bit</NOBR> double-precision binary format</TD>
</TR>
<TR>
<TD><CODE>extFloat80_t </CODE></TD>
<TD><NOBR>80-bit</NOBR> double-extended-precision binary format (old Intel or
Motorola format)</TD>
</TR>
<TR>
<TD><CODE>float128_t</CODE></TD>
<TD><NOBR>128-bit</NOBR> quadruple-precision binary format</TD>
</TR>
</TABLE>
</BLOCKQUOTE>
The non-extended types are each exactly the size specified:
<NOBR>16 bits</NOBR> for <CODE>float16_t</CODE>, <NOBR>32 bits</NOBR> for
<CODE>float32_t</CODE>, <NOBR>64 bits</NOBR> for <CODE>float64_t</CODE>, and
<NOBR>128 bits</NOBR> for <CODE>float128_t</CODE>.
Aside from these size requirements, the definitions of all these types may
differ for different ports of SoftFloat to specific systems.
A given port of SoftFloat may or may not define some of the floating-point
types as aliases for the C standard types <CODE>float</CODE>,
<CODE>double</CODE>, and <CODE>long</CODE> <CODE>double</CODE>.
</P>
<P>
Header file <CODE>softfloat.h</CODE> also defines a structure,
<CODE>struct</CODE> <CODE>extFloat80M</CODE>, for the representation of
<NOBR>80-bit</NOBR> double-extended-precision floating-point values in memory.
This structure is the same size as type <CODE>extFloat80_t</CODE> and contains
at least these two fields (not necessarily in this order):
<BLOCKQUOTE>
<PRE>
uint16_t signExp;
uint64_t signif;
</PRE>
</BLOCKQUOTE>
Field <CODE>signExp</CODE> contains the sign and exponent of the floating-point
value, with the sign in the most significant bit (<NOBR>bit 15</NOBR>) and the
encoded exponent in the other <NOBR>15 bits</NOBR>.
Field <CODE>signif</CODE> is the complete <NOBR>64-bit</NOBR> significand of
the floating-point value.
(In the usual encoding for <NOBR>80-bit</NOBR> extended floating-point, the
leading <NOBR>1 bit</NOBR> of normalized numbers is not implicit but is stored
in the most significant bit of the significand.)
</P>
<H3>4.3. Supported Floating-Point Functions</H3>
<P>
SoftFloat implements these arithmetic operations for its floating-point types:
<UL>
<LI>
conversions between any two floating-point formats;
<LI>
for each floating-point format, conversions to and from signed and unsigned
<NOBR>32-bit</NOBR> and <NOBR>64-bit</NOBR> integers;
<LI>
for each format, the usual addition, subtraction, multiplication, division, and
square root operations;
<LI>
for each format except <CODE>extFloat80_t</CODE>, the fused multiply-add
operation defined by the IEEE Standard;
<LI>
for each format, the floating-point remainder operation defined by the IEEE
Standard;
<LI>
for each format, a “round to integer” operation that rounds to the
nearest integer value in the same format; and
<LI>
comparisons between two values in the same floating-point format.
</UL>
</P>
<P>
The following operations required by the 2008 IEEE Floating-Point Standard are
not supported in SoftFloat <NOBR>Release 3e</NOBR>:
<UL>
<LI>
<B>nextUp</B>, <B>nextDown</B>, <B>minNum</B>, <B>maxNum</B>, <B>minNumMag</B>,
<B>maxNumMag</B>, <B>scaleB</B>, and <B>logB</B>;
<LI>
conversions between floating-point formats and decimal or hexadecimal character
sequences;
<LI>
all “quiet-computation” operations (<B>copy</B>, <B>negate</B>,
<B>abs</B>, and <B>copySign</B>, which all involve only simple copying and/or
manipulation of the floating-point sign bit); and
<LI>
all “non-computational” operations other than <B>isSignaling</B>
(which is supported).
</UL>
</P>
<H3>4.4. Non-canonical Representations in <CODE>extFloat80_t</CODE></H3>
<P>
Because the <NOBR>80-bit</NOBR> double-extended-precision format,
<CODE>extFloat80_t</CODE>, stores an explicit leading significand bit, many
finite floating-point numbers are encodable in this type in multiple equivalent
forms.
Of these multiple encodings, there is always a unique one with the least
encoded exponent value, and this encoding is considered the <I>canonical</I>
representation of the floating-point number.
Any other equivalent representations (having a higher encoded exponent value)
are <I>non-canonical</I>.
For a value in the subnormal range (including zero), the canonical
representation always has an encoded exponent of zero and a leading significand
bit <NOBR>of 0</NOBR>.
For finite values outside the subnormal range, the canonical representation
always has an encoded exponent that is nonzero and a leading significand bit
<NOBR>of 1</NOBR>.
</P>
<P>
For an infinity or NaN, the leading significand bit is similarly expected to
<NOBR>be 1</NOBR>.
An infinity or NaN with a leading significand bit <NOBR>of 0</NOBR> is again
considered non-canonical.
Hence, altogether, to be canonical, a value of type <CODE>extFloat80_t</CODE>
must have a leading significand bit <NOBR>of 1</NOBR>, unless the value is
subnormal or zero, in which case the leading significand bit and the encoded
exponent must both be zero.
</P>
<P>
SoftFloat’s functions are not guaranteed to operate as expected when
inputs of type <CODE>extFloat80_t</CODE> are non-canonical.
Assuming all of a function’s <CODE>extFloat80_t</CODE> inputs (if any)
are canonical, function outputs of type <CODE>extFloat80_t</CODE> will always
be canonical.
</P>
<H3>4.5. Conventions for Passing Arguments and Results</H3>
<P>
Values that are at most <NOBR>64 bits</NOBR> in size (i.e., not the
<NOBR>80-bit</NOBR> or <NOBR>128-bit</NOBR> floating-point formats) are in all
cases passed as function arguments by value.
Likewise, when an output of a function is no more than <NOBR>64 bits</NOBR>, it
is always returned directly as the function result.
Thus, for example, the SoftFloat function for adding two <NOBR>64-bit</NOBR>
floating-point values has this simple signature:
<BLOCKQUOTE>
<CODE>float64_t f64_add( float64_t, float64_t );</CODE>
</BLOCKQUOTE>
</P>
<P>
The story is more complex when function inputs and outputs are
<NOBR>80-bit</NOBR> and <NOBR>128-bit</NOBR> floating-point.
For these types, SoftFloat always provides a function that passes these larger
values into or out of the function indirectly, via pointers.
For example, for adding two <NOBR>128-bit</NOBR> floating-point values,
SoftFloat supplies this function:
<BLOCKQUOTE>
<CODE>void f128M_add( const float128_t *, const float128_t *, float128_t * );</CODE>
</BLOCKQUOTE>
The first two arguments point to the values to be added, and the last argument
points to the location where the sum will be stored.
The <CODE>M</CODE> in the name <CODE>f128M_add</CODE> is mnemonic for the fact
that the <NOBR>128-bit</NOBR> inputs and outputs are “in memory”,
pointed to by pointer arguments.
</P>
<P>
All ports of SoftFloat implement these <I>pass-by-pointer</I> functions for
types <CODE>extFloat80_t</CODE> and <CODE>float128_t</CODE>.
At the same time, SoftFloat ports may also implement alternate versions of
these same functions that pass <CODE>extFloat80_t</CODE> and
<CODE>float128_t</CODE> by value, like the smaller formats.
Thus, besides the function with name <CODE>f128M_add</CODE> shown above, a
SoftFloat port may also supply an equivalent function with this signature:
<BLOCKQUOTE>
<CODE>float128_t f128_add( float128_t, float128_t );</CODE>
</BLOCKQUOTE>
</P>
<P>
As a general rule, on computers where the machine word size is
<NOBR>32 bits</NOBR> or smaller, only the pass-by-pointer versions of functions
(e.g., <CODE>f128M_add</CODE>) are provided for types <CODE>extFloat80_t</CODE>
and <CODE>float128_t</CODE>, because passing such large types directly can have
significant extra cost.
On computers where the word size is <NOBR>64 bits</NOBR> or larger, both
function versions (<CODE>f128M_add</CODE> and <CODE>f128_add</CODE>) are
provided, because the cost of passing by value is then more reasonable.
Applications that must be portable accross both classes of computers must use
the pointer-based functions, as these are always implemented.
However, if it is known that SoftFloat includes the by-value functions for all
platforms of interest, programmers can use whichever version they prefer.
</P>
<H2>5. Reserved Names</H2>
<P>
In addition to the variables and functions documented here, SoftFloat defines
some symbol names for its own private use.
These private names always begin with the prefix
‘<CODE>softfloat_</CODE>’.
When a program includes header <CODE>softfloat.h</CODE> or links with the
SoftFloat library, all names with prefix ‘<CODE>softfloat_</CODE>’
are reserved for possible use by SoftFloat.
Applications that use SoftFloat should not define their own names with this
prefix, and should reference only such names as are documented.
</P>
<H2>6. Mode Variables</H2>
<P>
The following global variables control rounding mode, underflow detection, and
the <NOBR>80-bit</NOBR> extended format’s rounding precision:
<BLOCKQUOTE>
<CODE>softfloat_roundingMode</CODE><BR>
<CODE>softfloat_detectTininess</CODE><BR>
<CODE>extF80_roundingPrecision</CODE>
</BLOCKQUOTE>
These mode variables are covered in the next several subsections.
For some SoftFloat ports, these variables may be <I>per-thread</I> (declared
<CODE>thread_local</CODE>), meaning that different execution threads have their
own separate copies of the variables.
</P>
<H3>6.1. Rounding Mode</H3>
<P>
All five rounding modes defined by the 2008 IEEE Floating-Point Standard are
implemented for all operations that require rounding.
Some ports of SoftFloat may also implement the <I>round-to-odd</I> mode.
</P>
<P>
The rounding mode is selected by the global variable
<BLOCKQUOTE>
<CODE>uint_fast8_t softfloat_roundingMode;</CODE>
</BLOCKQUOTE>
This variable may be set to one of the values
<BLOCKQUOTE>
<TABLE CELLSPACING=0 CELLPADDING=0>
<TR>
<TD><CODE>softfloat_round_near_even</CODE></TD>
<TD>round to nearest, with ties to even</TD>
</TR>
<TR>
<TD><CODE>softfloat_round_near_maxMag </CODE></TD>
<TD>round to nearest, with ties to maximum magnitude (away from zero)</TD>
</TR>
<TR>
<TD><CODE>softfloat_round_minMag</CODE></TD>
<TD>round to minimum magnitude (toward zero)</TD>
</TR>
<TR>
<TD><CODE>softfloat_round_min</CODE></TD>
<TD>round to minimum (down)</TD>
</TR>
<TR>
<TD><CODE>softfloat_round_max</CODE></TD>
<TD>round to maximum (up)</TD>
</TR>
<TR>
<TD><CODE>softfloat_round_odd</CODE></TD>
<TD>round to odd (jamming), if supported by the SoftFloat port</TD>
</TR>
</TABLE>
</BLOCKQUOTE>
Variable <CODE>softfloat_roundingMode</CODE> is initialized to
<CODE>softfloat_round_near_even</CODE>.
</P>
<P>
When <CODE>softfloat_round_odd</CODE> is the rounding mode for a function that
rounds to an integer value (either conversion to an integer format or a
‘<CODE>roundToInt</CODE>’ function), if the input is not already an
integer, the rounded result is the closest <EM>odd</EM> integer.
For other operations, this rounding mode acts as though the floating-point
result is first rounded to minimum magnitude, the same as
<CODE>softfloat_round_minMag</CODE>, and then, if the result is inexact, the
least-significant bit of the result is set <NOBR>to 1</NOBR>.
Rounding to odd is also known as <EM>jamming</EM>.
</P>
<H3>6.2. Underflow Detection</H3>
<P>
In the terminology of the IEEE Standard, SoftFloat can detect tininess for
underflow either before or after rounding.
The choice is made by the global variable
<BLOCKQUOTE>
<CODE>uint_fast8_t softfloat_detectTininess;</CODE>
</BLOCKQUOTE>
which can be set to either
<BLOCKQUOTE>
<CODE>softfloat_tininess_beforeRounding</CODE><BR>
<CODE>softfloat_tininess_afterRounding</CODE>
</BLOCKQUOTE>
Detecting tininess after rounding is usually better because it results in fewer
spurious underflow signals.
The other option is provided for compatibility with some systems.
Like most systems (and as required by the newer 2008 IEEE Standard), SoftFloat
always detects loss of accuracy for underflow as an inexact result.
</P>
<H3>6.3. Rounding Precision for the <NOBR>80-Bit</NOBR> Extended Format</H3>
<P>
For <CODE>extFloat80_t</CODE> only, the rounding precision of the basic
arithmetic operations is controlled by the global variable
<BLOCKQUOTE>
<CODE>uint_fast8_t extF80_roundingPrecision;</CODE>
</BLOCKQUOTE>
The operations affected are:
<BLOCKQUOTE>
<CODE>extF80_add</CODE><BR>
<CODE>extF80_sub</CODE><BR>
<CODE>extF80_mul</CODE><BR>
<CODE>extF80_div</CODE><BR>
<CODE>extF80_sqrt</CODE>
</BLOCKQUOTE>
When <CODE>extF80_roundingPrecision</CODE> is set to its default value of 80,
these operations are rounded to the full precision of the <NOBR>80-bit</NOBR>
double-extended-precision format, like occurs for other formats.
Setting <CODE>extF80_roundingPrecision</CODE> to 32 or to 64 causes the
operations listed to be rounded to <NOBR>32-bit</NOBR> precision (equivalent to
<CODE>float32_t</CODE>) or to <NOBR>64-bit</NOBR> precision (equivalent to
<CODE>float64_t</CODE>), respectively.
When rounding to reduced precision, additional bits in the result significand
beyond the rounding point are set to zero.
The consequences of setting <CODE>extF80_roundingPrecision</CODE> to a value
other than 32, 64, or 80 is not specified.
Operations other than the ones listed above are not affected by
<CODE>extF80_roundingPrecision</CODE>.
</P>
<H2>7. Exceptions and Exception Flags</H2>
<P>
All five exception flags required by the IEEE Floating-Point Standard are
implemented.
Each flag is stored as a separate bit in the global variable
<BLOCKQUOTE>
<CODE>uint_fast8_t softfloat_exceptionFlags;</CODE>
</BLOCKQUOTE>
The positions of the exception flag bits within this variable are determined by
the bit masks
<BLOCKQUOTE>
<CODE>softfloat_flag_inexact</CODE><BR>
<CODE>softfloat_flag_underflow</CODE><BR>
<CODE>softfloat_flag_overflow</CODE><BR>
<CODE>softfloat_flag_infinite</CODE><BR>
<CODE>softfloat_flag_invalid</CODE>
</BLOCKQUOTE>
Variable <CODE>softfloat_exceptionFlags</CODE> is initialized to all zeros,
meaning no exceptions.
</P>
<P>
For some SoftFloat ports, <CODE>softfloat_exceptionFlags</CODE> may be
<I>per-thread</I> (declared <CODE>thread_local</CODE>), meaning that different
execution threads have their own separate instances of it.
</P>
<P>
An individual exception flag can be cleared with the statement
<BLOCKQUOTE>
<CODE>softfloat_exceptionFlags &= ~softfloat_flag_<<I>exception</I>>;</CODE>
</BLOCKQUOTE>
where <CODE><<I>exception</I>></CODE> is the appropriate name.
To raise a floating-point exception, function <CODE>softfloat_raiseFlags</CODE>
should normally be used.
</P>
<P>
When SoftFloat detects an exception other than <I>inexact</I>, it calls
<CODE>softfloat_raiseFlags</CODE>.
The default version of this function simply raises the corresponding exception
flags.
Particular ports of SoftFloat may support alternate behavior, such as exception
traps, by modifying the default <CODE>softfloat_raiseFlags</CODE>.
A program may also supply its own <CODE>softfloat_raiseFlags</CODE> function to
override the one from the SoftFloat library.
</P>
<P>
Because inexact results occur frequently under most circumstances (and thus are
hardly exceptional), SoftFloat does not ordinarily call
<CODE>softfloat_raiseFlags</CODE> for <I>inexact</I> exceptions.
It does always raise the <I>inexact</I> exception flag as required.
</P>
<H2>8. Function Details</H2>
<P>
In this section, <CODE><<I>float</I>></CODE> appears in function names as
a substitute for one of these abbreviations:
<BLOCKQUOTE>
<TABLE CELLSPACING=0 CELLPADDING=0>
<TR>
<TD><CODE>f16</CODE></TD>
<TD>indicates <CODE>float16_t</CODE>, passed by value</TD>
</TR>
<TR>
<TD><CODE>f32</CODE></TD>
<TD>indicates <CODE>float32_t</CODE>, passed by value</TD>
</TR>
<TR>
<TD><CODE>f64</CODE></TD>
<TD>indicates <CODE>float64_t</CODE>, passed by value</TD>
</TR>
<TR>
<TD><CODE>extF80M </CODE></TD>
<TD>indicates <CODE>extFloat80_t</CODE>, passed indirectly via pointers</TD>
</TR>
<TR>
<TD><CODE>extF80</CODE></TD>
<TD>indicates <CODE>extFloat80_t</CODE>, passed by value</TD>
</TR>
<TR>
<TD><CODE>f128M</CODE></TD>
<TD>indicates <CODE>float128_t</CODE>, passed indirectly via pointers</TD>
</TR>
<TR>
<TD><CODE>f128</CODE></TD>
<TD>indicates <CODE>float128_t</CODE>, passed by value</TD>
</TR>
</TABLE>
</BLOCKQUOTE>
The circumstances under which values of floating-point types
<CODE>extFloat80_t</CODE> and <CODE>float128_t</CODE> may be passed either by
value or indirectly via pointers was discussed earlier in
<NOBR>section 4.5</NOBR>, <I>Conventions for Passing Arguments and Results</I>.
</P>
<H3>8.1. Conversions from Integer to Floating-Point</H3>
<P>
All conversions from a <NOBR>32-bit</NOBR> or <NOBR>64-bit</NOBR> integer,
signed or unsigned, to a floating-point format are supported.
Functions performing these conversions have these names:
<BLOCKQUOTE>
<CODE>ui32_to_<<I>float</I>></CODE><BR>
<CODE>ui64_to_<<I>float</I>></CODE><BR>
<CODE>i32_to_<<I>float</I>></CODE><BR>
<CODE>i64_to_<<I>float</I>></CODE>
</BLOCKQUOTE>
Conversions from <NOBR>32-bit</NOBR> integers to <NOBR>64-bit</NOBR>
double-precision and larger formats are always exact, and likewise conversions
from <NOBR>64-bit</NOBR> integers to <NOBR>80-bit</NOBR>
double-extended-precision and <NOBR>128-bit</NOBR> quadruple-precision are also
always exact.
</P>
<P>
Each conversion function takes one input of the appropriate type and generates
one output.
The following illustrates the signatures of these functions in cases when the
floating-point result is passed either by value or via pointers:
<BLOCKQUOTE>
<PRE>
float64_t i32_to_f64( int32_t <I>a</I> );
</PRE>
<PRE>
void i32_to_f128M( int32_t <I>a</I>, float128_t *<I>destPtr</I> );
</PRE>
</BLOCKQUOTE>
</P>
<H3>8.2. Conversions from Floating-Point to Integer</H3>
<P>
Conversions from a floating-point format to a <NOBR>32-bit</NOBR> or
<NOBR>64-bit</NOBR> integer, signed or unsigned, are supported with these
functions:
<BLOCKQUOTE>
<CODE><<I>float</I>>_to_ui32</CODE><BR>
<CODE><<I>float</I>>_to_ui64</CODE><BR>
<CODE><<I>float</I>>_to_i32</CODE><BR>
<CODE><<I>float</I>>_to_i64</CODE>
</BLOCKQUOTE>
The functions have signatures as follows, depending on whether the
floating-point input is passed by value or via pointers:
<BLOCKQUOTE>
<PRE>
int_fast32_t f64_to_i32( float64_t <I>a</I>, uint_fast8_t <I>roundingMode</I>, bool <I>exact</I> );
</PRE>
<PRE>
int_fast32_t
f128M_to_i32( const float128_t *<I>aPtr</I>, uint_fast8_t <I>roundingMode</I>, bool <I>exact</I> );
</PRE>
</BLOCKQUOTE>
</P>
<P>
The <CODE><I>roundingMode</I></CODE> argument specifies the rounding mode for
the conversion.
The variable that usually indicates rounding mode,
<CODE>softfloat_roundingMode</CODE>, is ignored.
Argument <CODE><I>exact</I></CODE> determines whether the <I>inexact</I>
exception flag is raised if the conversion is not exact.
If <CODE><I>exact</I></CODE> is <CODE>true</CODE>, the <I>inexact</I> flag may
be raised;
otherwise, it will not be, even if the conversion is inexact.
</P>
<P>
A conversion from floating-point to integer format raises the <I>invalid</I>
exception if the source value cannot be rounded to a representable integer of
the desired size (32 or 64 bits).
In such circumstances, the integer result returned is determined by the
particular port of SoftFloat, although typically this value will be either the
maximum or minimum value of the integer format.
The functions that convert to integer types never raise the floating-point
<I>overflow</I> exception.
</P>
<P>
Because languages such <NOBR>as C</NOBR> require that conversions to integers
be rounded toward zero, the following functions are provided for improved speed
and convenience:
<BLOCKQUOTE>
<CODE><<I>float</I>>_to_ui32_r_minMag</CODE><BR>
<CODE><<I>float</I>>_to_ui64_r_minMag</CODE><BR>
<CODE><<I>float</I>>_to_i32_r_minMag</CODE><BR>
<CODE><<I>float</I>>_to_i64_r_minMag</CODE>
</BLOCKQUOTE>
These functions round only toward zero (to minimum magnitude).
The signatures for these functions are the same as above without the redundant
<CODE><I>roundingMode</I></CODE> argument:
<BLOCKQUOTE>
<PRE>
int_fast32_t f64_to_i32_r_minMag( float64_t <I>a</I>, bool <I>exact</I> );
</PRE>
<PRE>
int_fast32_t f128M_to_i32_r_minMag( const float128_t *<I>aPtr</I>, bool <I>exact</I> );
</PRE>
</BLOCKQUOTE>
</P>
<H3>8.3. Conversions Among Floating-Point Types</H3>
<P>
Conversions between floating-point formats are done by functions with these
names:
<BLOCKQUOTE>
<CODE><<I>float</I>>_to_<<I>float</I>></CODE>
</BLOCKQUOTE>
All combinations of source and result type are supported where the source and
result are different formats.
There are four different styles of signature for these functions, depending on
whether the input and the output floating-point values are passed by value or
via pointers:
<BLOCKQUOTE>
<PRE>
float32_t f64_to_f32( float64_t <I>a</I> );
</PRE>
<PRE>
float32_t f128M_to_f32( const float128_t *<I>aPtr</I> );
</PRE>
<PRE>
void f32_to_f128M( float32_t <I>a</I>, float128_t *<I>destPtr</I> );
</PRE>
<PRE>
void extF80M_to_f128M( const extFloat80_t *<I>aPtr</I>, float128_t *<I>destPtr</I> );
</PRE>
</BLOCKQUOTE>
</P>
<P>
Conversions from a smaller to a larger floating-point format are always exact
and so require no rounding.
</P>
<H3>8.4. Basic Arithmetic Functions</H3>
<P>
The following basic arithmetic functions are provided:
<BLOCKQUOTE>
<CODE><<I>float</I>>_add</CODE><BR>
<CODE><<I>float</I>>_sub</CODE><BR>
<CODE><<I>float</I>>_mul</CODE><BR>
<CODE><<I>float</I>>_div</CODE><BR>
<CODE><<I>float</I>>_sqrt</CODE>
</BLOCKQUOTE>
Each floating-point operation takes two operands, except for <CODE>sqrt</CODE>
(square root) which takes only one.
The operands and result are all of the same floating-point format.
Signatures for these functions take the following forms:
<BLOCKQUOTE>
<PRE>
float64_t f64_add( float64_t <I>a</I>, float64_t <I>b</I> );
</PRE>
<PRE>
void
f128M_add(
const float128_t *<I>aPtr</I>, const float128_t *<I>bPtr</I>, float128_t *<I>destPtr</I> );
</PRE>
<PRE>
float64_t f64_sqrt( float64_t <I>a</I> );
</PRE>
<PRE>
void f128M_sqrt( const float128_t *<I>aPtr</I>, float128_t *<I>destPtr</I> );
</PRE>
</BLOCKQUOTE>
When floating-point values are passed indirectly through pointers, arguments
<CODE><I>aPtr</I></CODE> and <CODE><I>bPtr</I></CODE> point to the input
operands, and the last argument, <CODE><I>destPtr</I></CODE>, points to the
location where the result is stored.
</P>
<P>
Rounding of the <NOBR>80-bit</NOBR> double-extended-precision
(<CODE>extFloat80_t</CODE>) functions is affected by variable
<CODE>extF80_roundingPrecision</CODE>, as explained earlier in
<NOBR>section 6.3</NOBR>,
<I>Rounding Precision for the <NOBR>80-Bit</NOBR> Extended Format</I>.
</P>
<H3>8.5. Fused Multiply-Add Functions</H3>
<P>
The 2008 version of the IEEE Floating-Point Standard defines a <I>fused
multiply-add</I> operation that does a combined multiplication and addition
with only a single rounding.
SoftFloat implements fused multiply-add with functions
<BLOCKQUOTE>
<CODE><<I>float</I>>_mulAdd</CODE>
</BLOCKQUOTE>
Unlike other operations, fused multiple-add is not supported for the
<NOBR>80-bit</NOBR> double-extended-precision format,
<CODE>extFloat80_t</CODE>.
</P>
<P>
Depending on whether floating-point values are passed by value or via pointers,
the fused multiply-add functions have signatures of these forms:
<BLOCKQUOTE>
<PRE>
float64_t f64_mulAdd( float64_t <I>a</I>, float64_t <I>b</I>, float64_t <I>c</I> );
</PRE>
<PRE>
void
f128M_mulAdd(
const float128_t *<I>aPtr</I>,
const float128_t *<I>bPtr</I>,
const float128_t *<I>cPtr</I>,
float128_t *<I>destPtr</I>
);
</PRE>
</BLOCKQUOTE>
The functions compute
<NOBR>(<CODE><I>a</I></CODE> × <CODE><I>b</I></CODE>)
+ <CODE><I>c</I></CODE></NOBR>
with a single rounding.
When floating-point values are passed indirectly through pointers, arguments
<CODE><I>aPtr</I></CODE>, <CODE><I>bPtr</I></CODE>, and
<CODE><I>cPtr</I></CODE> point to operands <CODE><I>a</I></CODE>,
<CODE><I>b</I></CODE>, and <CODE><I>c</I></CODE> respectively, and
<CODE><I>destPtr</I></CODE> points to the location where the result is stored.
</P>
<P>
If one of the multiplication operands <CODE><I>a</I></CODE> and
<CODE><I>b</I></CODE> is infinite and the other is zero, these functions raise
the invalid exception even if operand <CODE><I>c</I></CODE> is a quiet NaN.
</P>
<H3>8.6. Remainder Functions</H3>
<P>
For each format, SoftFloat implements the remainder operation defined by the
IEEE Floating-Point Standard.
The remainder functions have names
<BLOCKQUOTE>
<CODE><<I>float</I>>_rem</CODE>
</BLOCKQUOTE>
Each remainder operation takes two floating-point operands of the same format
and returns a result in the same format.
Depending on whether floating-point values are passed by value or via pointers,
the remainder functions have signatures of these forms:
<BLOCKQUOTE>
<PRE>
float64_t f64_rem( float64_t <I>a</I>, float64_t <I>b</I> );
</PRE>
<PRE>
void
f128M_rem(
const float128_t *<I>aPtr</I>, const float128_t *<I>bPtr</I>, float128_t *<I>destPtr</I> );
</PRE>
</BLOCKQUOTE>
When floating-point values are passed indirectly through pointers, arguments
<CODE><I>aPtr</I></CODE> and <CODE><I>bPtr</I></CODE> point to operands
<CODE><I>a</I></CODE> and <CODE><I>b</I></CODE> respectively, and
<CODE><I>destPtr</I></CODE> points to the location where the result is stored.
</P>
<P>
The IEEE Standard remainder operation computes the value
<NOBR><CODE><I>a</I></CODE>
− <I>n</I> × <CODE><I>b</I></CODE></NOBR>,
where <I>n</I> is the integer closest to
<NOBR><CODE><I>a</I></CODE> ÷ <CODE><I>b</I></CODE></NOBR>.
If <NOBR><CODE><I>a</I></CODE> ÷ <CODE><I>b</I></CODE></NOBR> is exactly
halfway between two integers, <I>n</I> is the <EM>even</EM> integer closest to
<NOBR><CODE><I>a</I></CODE> ÷ <CODE><I>b</I></CODE></NOBR>.
The IEEE Standard’s remainder operation is always exact and so requires
no rounding.
</P>
<P>
Depending on the relative magnitudes of the operands, the remainder
functions can take considerably longer to execute than the other SoftFloat
functions.
This is an inherent characteristic of the remainder operation itself and is not
a flaw in the SoftFloat implementation.
</P>
<H3>8.7. Round-to-Integer Functions</H3>
<P>
For each format, SoftFloat implements the round-to-integer operation specified
by the IEEE Floating-Point Standard.
These functions are named
<BLOCKQUOTE>
<CODE><<I>float</I>>_roundToInt</CODE>
</BLOCKQUOTE>
Each round-to-integer operation takes a single floating-point operand.
This operand is rounded to an integer according to a specified rounding mode,
and the resulting integer value is returned in the same floating-point format.
(Note that the result is not an integer type.)
</P>
<P>
The signatures of the round-to-integer functions are similar to those for
conversions to an integer type:
<BLOCKQUOTE>
<PRE>
float64_t f64_roundToInt( float64_t <I>a</I>, uint_fast8_t <I>roundingMode</I>, bool <I>exact</I> );
</PRE>
<PRE>
void
f128M_roundToInt(
const float128_t *<I>aPtr</I>,
uint_fast8_t <I>roundingMode</I>,
bool <I>exact</I>,
float128_t *<I>destPtr</I>
);
</PRE>
</BLOCKQUOTE>
When floating-point values are passed indirectly through pointers,
<CODE><I>aPtr</I></CODE> points to the input operand and
<CODE><I>destPtr</I></CODE> points to the location where the result is stored.
</P>
<P>
The <CODE><I>roundingMode</I></CODE> argument specifies the rounding mode to
apply.
The variable that usually indicates rounding mode,
<CODE>softfloat_roundingMode</CODE>, is ignored.
Argument <CODE><I>exact</I></CODE> determines whether the <I>inexact</I>
exception flag is raised if the conversion is not exact.
If <CODE><I>exact</I></CODE> is <CODE>true</CODE>, the <I>inexact</I> flag may
be raised;
otherwise, it will not be, even if the conversion is inexact.
</P>
<H3>8.8. Comparison Functions</H3>
<P>
For each format, the following floating-point comparison functions are
provided:
<BLOCKQUOTE>
<CODE><<I>float</I>>_eq</CODE><BR>
<CODE><<I>float</I>>_le</CODE><BR>
<CODE><<I>float</I>>_lt</CODE>
</BLOCKQUOTE>
Each comparison takes two operands of the same type and returns a Boolean.
The abbreviation <CODE>eq</CODE> stands for “equal” (=);
<CODE>le</CODE> stands for “less than or equal” (≤);
and <CODE>lt</CODE> stands for “less than” (<).
Depending on whether the floating-point operands are passed by value or via
pointers, the comparison functions have signatures of these forms:
<BLOCKQUOTE>
<PRE>
bool f64_eq( float64_t <I>a</I>, float64_t <I>b</I> );
</PRE>
<PRE>
bool f128M_eq( const float128_t *<I>aPtr</I>, const float128_t *<I>bPtr</I> );
</PRE>
</BLOCKQUOTE>
</P>
<P>
The usual greater-than (>), greater-than-or-equal (≥), and not-equal
(≠) comparisons are easily obtained from the functions provided.
The not-equal function is just the logical complement of the equal function.
The greater-than-or-equal function is identical to the less-than-or-equal
function with the arguments in reverse order, and likewise the greater-than
function is identical to the less-than function with the arguments reversed.
</P>
<P>
The IEEE Floating-Point Standard specifies that the less-than-or-equal and
less-than comparisons by default raise the <I>invalid</I> exception if either
operand is any kind of NaN.
Equality comparisons, on the other hand, are defined by default to raise the
<I>invalid</I> exception only for signaling NaNs, not quiet NaNs.
For completeness, SoftFloat provides these complementary functions:
<BLOCKQUOTE>
<CODE><<I>float</I>>_eq_signaling</CODE><BR>
<CODE><<I>float</I>>_le_quiet</CODE><BR>
<CODE><<I>float</I>>_lt_quiet</CODE>
</BLOCKQUOTE>
The <CODE>signaling</CODE> equality comparisons are identical to the default
equality comparisons except that the <I>invalid</I> exception is raised for any
NaN input, not just for signaling NaNs.
Similarly, the <CODE>quiet</CODE> comparison functions are identical to their
default counterparts except that the <I>invalid</I> exception is not raised for
quiet NaNs.
</P>
<H3>8.9. Signaling NaN Test Functions</H3>
<P>
Functions for testing whether a floating-point value is a signaling NaN are
provided with these names:
<BLOCKQUOTE>
<CODE><<I>float</I>>_isSignalingNaN</CODE>
</BLOCKQUOTE>
The functions take one floating-point operand and return a Boolean indicating
whether the operand is a signaling NaN.
Accordingly, the functions have the forms
<BLOCKQUOTE>
<PRE>
bool f64_isSignalingNaN( float64_t <I>a</I> );
</PRE>
<PRE>
bool f128M_isSignalingNaN( const float128_t *<I>aPtr</I> );
</PRE>
</BLOCKQUOTE>
</P>
<H3>8.10. Raise-Exception Function</H3>
<P>
SoftFloat provides a single function for raising floating-point exceptions:
<BLOCKQUOTE>
<PRE>
void softfloat_raiseFlags( uint_fast8_t <I>exceptions</I> );
</PRE>
</BLOCKQUOTE>
The <CODE><I>exceptions</I></CODE> argument is a mask indicating the set of
exceptions to raise.
(See earlier section 7, <I>Exceptions and Exception Flags</I>.)
In addition to setting the specified exception flags in variable
<CODE>softfloat_exceptionFlags</CODE>, the <CODE>softfloat_raiseFlags</CODE>
function may cause a trap or abort appropriate for the current system.
</P>
<H2>9. Changes from SoftFloat <NOBR>Release 2</NOBR></H2>
<P>
Apart from a change in the legal use license, <NOBR>Release 3</NOBR> of
SoftFloat introduced numerous technical differences compared to earlier
releases.
</P>
<H3>9.1. Name Changes</H3>
<P>
The most obvious and pervasive difference compared to <NOBR>Release 2</NOBR>
is that the names of most functions and variables have changed, even when the
behavior has not.
First, the floating-point types, the mode variables, the exception flags
variable, the function to raise exceptions, and various associated constants
have been renamed as follows:
<BLOCKQUOTE>
<TABLE>
<TR>
<TD>old name, Release 2:</TD>
<TD>new name, Release 3:</TD>
</TR>
<TR>
<TD><CODE>float32</CODE></TD>
<TD><CODE>float32_t</CODE></TD>
</TR>
<TR>
<TD><CODE>float64</CODE></TD>
<TD><CODE>float64_t</CODE></TD>
</TR>
<TR>
<TD><CODE>floatx80</CODE></TD>
<TD><CODE>extFloat80_t</CODE></TD>
</TR>
<TR>
<TD><CODE>float128</CODE></TD>
<TD><CODE>float128_t</CODE></TD>
</TR>
<TR>
<TD><CODE>float_rounding_mode</CODE></TD>
<TD><CODE>softfloat_roundingMode</CODE></TD>
</TR>
<TR>
<TD><CODE>float_round_nearest_even</CODE></TD>
<TD><CODE>softfloat_round_near_even</CODE></TD>
</TR>
<TR>
<TD><CODE>float_round_to_zero</CODE></TD>
<TD><CODE>softfloat_round_minMag</CODE></TD>
</TR>
<TR>
<TD><CODE>float_round_down</CODE></TD>
<TD><CODE>softfloat_round_min</CODE></TD>
</TR>
<TR>
<TD><CODE>float_round_up</CODE></TD>
<TD><CODE>softfloat_round_max</CODE></TD>
</TR>
<TR>
<TD><CODE>float_detect_tininess</CODE></TD>
<TD><CODE>softfloat_detectTininess</CODE></TD>
</TR>
<TR>
<TD><CODE>float_tininess_before_rounding </CODE></TD>
<TD><CODE>softfloat_tininess_beforeRounding</CODE></TD>
</TR>
<TR>
<TD><CODE>float_tininess_after_rounding</CODE></TD>
<TD><CODE>softfloat_tininess_afterRounding</CODE></TD>
</TR>
<TR>
<TD><CODE>floatx80_rounding_precision</CODE></TD>
<TD><CODE>extF80_roundingPrecision</CODE></TD>
</TR>
<TR>
<TD><CODE>float_exception_flags</CODE></TD>
<TD><CODE>softfloat_exceptionFlags</CODE></TD>
</TR>
<TR>
<TD><CODE>float_flag_inexact</CODE></TD>
<TD><CODE>softfloat_flag_inexact</CODE></TD>
</TR>
<TR>
<TD><CODE>float_flag_underflow</CODE></TD>
<TD><CODE>softfloat_flag_underflow</CODE></TD>
</TR>
<TR>
<TD><CODE>float_flag_overflow</CODE></TD>
<TD><CODE>softfloat_flag_overflow</CODE></TD>
</TR>
<TR>
<TD><CODE>float_flag_divbyzero</CODE></TD>
<TD><CODE>softfloat_flag_infinite</CODE></TD>
</TR>
<TR>
<TD><CODE>float_flag_invalid</CODE></TD>
<TD><CODE>softfloat_flag_invalid</CODE></TD>
</TR>
<TR>
<TD><CODE>float_raise</CODE></TD>
<TD><CODE>softfloat_raiseFlags</CODE></TD>
</TR>
</TABLE>
</BLOCKQUOTE>
</P>
<P>
Furthermore, <NOBR>Release 3</NOBR> adopted the following new abbreviations for
function names:
<BLOCKQUOTE>
<TABLE>
<TR>
<TD>used in names in Release 2:<CODE> </CODE></TD>
<TD>used in names in Release 3:</TD>
</TR>
<TR> <TD><CODE>int32</CODE></TD> <TD><CODE>i32</CODE></TD> </TR>
<TR> <TD><CODE>int64</CODE></TD> <TD><CODE>i64</CODE></TD> </TR>
<TR> <TD><CODE>float32</CODE></TD> <TD><CODE>f32</CODE></TD> </TR>
<TR> <TD><CODE>float64</CODE></TD> <TD><CODE>f64</CODE></TD> </TR>
<TR> <TD><CODE>floatx80</CODE></TD> <TD><CODE>extF80</CODE></TD> </TR>
<TR> <TD><CODE>float128</CODE></TD> <TD><CODE>f128</CODE></TD> </TR>
</TABLE>
</BLOCKQUOTE>
Thus, for example, the function to add two <NOBR>32-bit</NOBR> floating-point
numbers, previously called <CODE>float32_add</CODE> in <NOBR>Release 2</NOBR>,
is now <CODE>f32_add</CODE>.
Lastly, there have been a few other changes to function names:
<BLOCKQUOTE>
<TABLE>
<TR>
<TD>used in names in Release 2:<CODE> </CODE></TD>
<TD>used in names in Release 3:<CODE> </CODE></TD>
<TD>relevant functions:</TD>
</TR>
<TR>
<TD><CODE>_round_to_zero</CODE></TD>
<TD><CODE>_r_minMag</CODE></TD>
<TD>conversions from floating-point to integer (<NOBR>section 8.2</NOBR>)</TD>
</TR>
<TR>
<TD><CODE>round_to_int</CODE></TD>
<TD><CODE>roundToInt</CODE></TD>
<TD>round-to-integer functions (<NOBR>section 8.7</NOBR>)</TD>
</TR>
<TR>
<TD><CODE>is_signaling_nan </CODE></TD>
<TD><CODE>isSignalingNaN</CODE></TD>
<TD>signaling NaN test functions (<NOBR>section 8.9</NOBR>)</TD>
</TR>
</TABLE>
</BLOCKQUOTE>
</P>
<H3>9.2. Changes to Function Arguments</H3>
<P>
Besides simple name changes, some operations were given a different interface
in <NOBR>Release 3</NOBR> than they had in <NOBR>Release 2</NOBR>:
<UL>
<LI>
<P>
Since <NOBR>Release 3</NOBR>, integer arguments and results of functions have
standard types from header <CODE><stdint.h></CODE>, such as
<CODE>uint32_t</CODE>, whereas previously their types could be defined
differently for each port of SoftFloat, usually using traditional C types such
as <CODE>unsigned</CODE> <CODE>int</CODE>.
Likewise, functions in <NOBR>Release 3</NOBR> and later pass Booleans as
standard type <CODE>bool</CODE> from <CODE><stdbool.h></CODE>, whereas
previously these were again passed as a port-specific type (usually
<CODE>int</CODE>).
</P>
<LI>
<P>
As explained earlier in <NOBR>section 4.5</NOBR>, <I>Conventions for Passing
Arguments and Results</I>, SoftFloat functions in <NOBR>Release 3</NOBR> and
later may pass <NOBR>80-bit</NOBR> and <NOBR>128-bit</NOBR> floating-point
values through pointers, meaning that functions take pointer arguments and then
read or write floating-point values at the locations indicated by the pointers.
In <NOBR>Release 2</NOBR>, floating-point arguments and results were always
passed by value, regardless of their size.
</P>
<LI>
<P>
Functions that round to an integer have additional
<CODE><I>roundingMode</I></CODE> and <CODE><I>exact</I></CODE> arguments that
they did not have in <NOBR>Release 2</NOBR>.
Refer to sections 8.2 <NOBR>and 8.7</NOBR> for descriptions of these functions
since <NOBR>Release 3</NOBR>.
For <NOBR>Release 2</NOBR>, the rounding mode, when needed, was taken from the
same global variable that affects the basic arithmetic operations (now called
<CODE>softfloat_roundingMode</CODE> but previously known as
<CODE>float_rounding_mode</CODE>).
Also, for <NOBR>Release 2</NOBR>, if the original floating-point input was not
an exact integer value, and if the <I>invalid</I> exception was not raised by
the function, the <I>inexact</I> exception was always raised.
<NOBR>Release 2</NOBR> had no option to suppress raising <I>inexact</I> in this
case.
Applications using SoftFloat <NOBR>Release 3</NOBR> or later can get the same
effect as <NOBR>Release 2</NOBR> by passing variable
<CODE>softfloat_roundingMode</CODE> for argument
<CODE><I>roundingMode</I></CODE> and <CODE>true</CODE> for argument
<CODE><I>exact</I></CODE>.
</P>
</UL>
</P>
<H3>9.3. Added Capabilities</H3>
<P>
With <NOBR>Release 3</NOBR>, some new features have been added that were not
present in <NOBR>Release 2</NOBR>:
<UL>
<LI>
<P>
A port of SoftFloat can now define any of the floating-point types
<CODE>float32_t</CODE>, <CODE>float64_t</CODE>, <CODE>extFloat80_t</CODE>, and
<CODE>float128_t</CODE> as aliases for C’s standard floating-point types
<CODE>float</CODE>, <CODE>double</CODE>, and <CODE>long</CODE>
<CODE>double</CODE>, using either <CODE>#define</CODE> or <CODE>typedef</CODE>.
This potential convenience was not supported under <NOBR>Release 2</NOBR>.
</P>
<P>
(Note, however, that there may be a performance cost to defining
SoftFloat’s floating-point types this way, depending on the platform and
the applications using SoftFloat.
Ports of SoftFloat may choose to forgo the convenience in favor of better
speed.)
</P>
<P>
<LI>
As of <NOBR>Release 3b</NOBR>, <NOBR>16-bit</NOBR> half-precision,
<CODE>float16_t</CODE>, is supported.
</P>
<P>
<LI>
Functions have been added for converting between the floating-point types and
unsigned integers.
<NOBR>Release 2</NOBR> supported only signed integers, not unsigned.
</P>
<P>
<LI>
Fused multiply-add functions have been added for all floating-point formats
except <NOBR>80-bit</NOBR> double-extended-precision,
<CODE>extFloat80_t</CODE>.
</P>
<P>
<LI>
New rounding modes are supported:
<CODE>softfloat_round_near_maxMag</CODE> (round to nearest, with ties to
maximum magnitude, away from zero), and, as of <NOBR>Release 3c</NOBR>,
optional <CODE>softfloat_round_odd</CODE> (round to odd, also known as
jamming).
</P>
</UL>
</P>
<H3>9.4. Better Compatibility with the C Language</H3>
<P>
<NOBR>Release 3</NOBR> of SoftFloat was written to conform better to the ISO C
Standard’s rules for portability.
For example, older releases of SoftFloat employed type conversions in ways
that, while commonly practiced, are not fully defined by the C Standard.
Such problematic type conversions have generally been replaced by the use of
unions, the behavior around which is more strictly regulated these days.
</P>
<H3>9.5. New Organization as a Library</H3>
<P>
Starting with <NOBR>Release 3</NOBR>, SoftFloat now builds as a library.
Previously, SoftFloat compiled into a single, monolithic object file containing
all the SoftFloat functions, with the consequence that a program linking with
SoftFloat would get every SoftFloat function in its binary file even if only a
few functions were actually used.
With SoftFloat in the form of a library, a program that is linked by a standard
linker will include only those functions of SoftFloat that it needs and no
others.
</P>
<H3>9.6. Optimization Gains (and Losses)</H3>
<P>
Individual SoftFloat functions have been variously improved in
<NOBR>Release 3</NOBR> compared to earlier releases.
In particular, better, faster algorithms have been deployed for the operations
of division, square root, and remainder.
For functions operating on the larger <NOBR>80-bit</NOBR> and
<NOBR>128-bit</NOBR> formats, <CODE>extFloat80_t</CODE> and
<CODE>float128_t</CODE>, code size has also generally been reduced.
</P>
<P>
However, because <NOBR>Release 2</NOBR> compiled all of SoftFloat together as a
single object file, compilers could make optimizations across function calls
when one SoftFloat function calls another.
Now that the functions of SoftFloat are compiled separately and only afterward
linked together into a program, there is not usually the same opportunity to
optimize across function calls.
Some loss of speed has been observed due to this change.
</P>
<H2>10. Future Directions</H2>
<P>
The following improvements are anticipated for future releases of SoftFloat:
<UL>
<LI>
more functions from the 2008 version of the IEEE Floating-Point Standard;
<LI>
consistent, defined behavior for non-canonical representations of extended
format <CODE>extFloat80_t</CODE> (discussed in <NOBR>section 4.4</NOBR>,
<I>Non-canonical Representations in <CODE>extFloat80_t</CODE></I>).
</UL>
</P>
<H2>11. 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>
|
