summaryrefslogtreecommitdiff
path: root/src/compiler/nir/nir_opt_algebraic.py
blob: acd5a2bab68748a0c92cc0e755574ccca48d626c (plain)
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
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
2049
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
2093
2094
2095
2096
2097
2098
2099
2100
2101
2102
2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
2150
2151
2152
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
2170
2171
2172
2173
2174
2175
2176
2177
2178
2179
2180
2181
2182
2183
2184
2185
2186
2187
2188
2189
2190
2191
2192
2193
2194
2195
2196
2197
2198
2199
2200
2201
2202
2203
2204
2205
2206
2207
2208
2209
2210
2211
2212
2213
2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
2228
2229
2230
2231
2232
2233
2234
2235
2236
2237
2238
2239
2240
2241
2242
2243
2244
2245
2246
2247
2248
2249
2250
2251
2252
2253
2254
2255
2256
2257
2258
2259
2260
2261
2262
2263
2264
2265
2266
2267
2268
2269
2270
2271
2272
2273
2274
2275
2276
2277
2278
2279
2280
2281
2282
2283
2284
2285
2286
2287
2288
2289
#
# Copyright (C) 2014 Intel Corporation
#
# Permission is hereby granted, free of charge, to any person obtaining a
# copy of this software and associated documentation files (the "Software"),
# to deal in the Software without restriction, including without limitation
# the rights to use, copy, modify, merge, publish, distribute, sublicense,
# and/or sell copies of the Software, and to permit persons to whom the
# Software is furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice (including the next
# paragraph) shall be included in all copies or substantial portions of the
# Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
# THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
# FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
# IN THE SOFTWARE.
#
# Authors:
#    Jason Ekstrand (jason@jlekstrand.net)

from __future__ import print_function

from collections import OrderedDict
import nir_algebraic
from nir_opcodes import type_sizes
import itertools
import struct
from math import pi

# Convenience variables
a = 'a'
b = 'b'
c = 'c'
d = 'd'
e = 'e'

# Written in the form (<search>, <replace>) where <search> is an expression
# and <replace> is either an expression or a value.  An expression is
# defined as a tuple of the form ([~]<op>, <src0>, <src1>, <src2>, <src3>)
# where each source is either an expression or a value.  A value can be
# either a numeric constant or a string representing a variable name.
#
# If the opcode in a search expression is prefixed by a '~' character, this
# indicates that the operation is inexact.  Such operations will only get
# applied to SSA values that do not have the exact bit set.  This should be
# used by by any optimizations that are not bit-for-bit exact.  It should not,
# however, be used for backend-requested lowering operations as those need to
# happen regardless of precision.
#
# Variable names are specified as "[#]name[@type][(cond)][.swiz]" where:
# "#" indicates that the given variable will only match constants,
# type indicates that the given variable will only match values from ALU
#    instructions with the given output type,
# (cond) specifies an additional condition function (see nir_search_helpers.h),
# swiz is a swizzle applied to the variable (only in the <replace> expression)
#
# For constants, you have to be careful to make sure that it is the right
# type because python is unaware of the source and destination types of the
# opcodes.
#
# All expression types can have a bit-size specified.  For opcodes, this
# looks like "op@32", for variables it is "a@32" or "a@uint32" to specify a
# type and size.  In the search half of the expression this indicates that it
# should only match that particular bit-size.  In the replace half of the
# expression this indicates that the constructed value should have that
# bit-size.
#
# If the opcode in a replacement expression is prefixed by a '!' character,
# this indicated that the new expression will be marked exact.
#
# A special condition "many-comm-expr" can be used with expressions to note
# that the expression and its subexpressions have more commutative expressions
# than nir_replace_instr can handle.  If this special condition is needed with
# another condition, the two can be separated by a comma (e.g.,
# "(many-comm-expr,is_used_once)").

# based on https://web.archive.org/web/20180105155939/http://forum.devmaster.net/t/fast-and-accurate-sine-cosine/9648
def lowered_sincos(c):
    x = ('fsub', ('fmul', 2.0, ('ffract', ('fadd', ('fmul', 0.5 / pi, a), c))), 1.0)
    x = ('fmul', ('fsub', x, ('fmul', x, ('fabs', x))), 4.0)
    return ('ffma', ('ffma', x, ('fabs', x), ('fneg', x)), 0.225, x)

def intBitsToFloat(i):
    return struct.unpack('!f', struct.pack('!I', i))[0]

optimizations = [

   (('imul', a, '#b(is_pos_power_of_two)'), ('ishl', a, ('find_lsb', b)), '!options->lower_bitops'),
   (('imul', a, '#b(is_neg_power_of_two)'), ('ineg', ('ishl', a, ('find_lsb', ('iabs', b)))), '!options->lower_bitops'),
   (('ishl', a, '#b'), ('imul', a, ('ishl', 1, b)), 'options->lower_bitops'),

   (('unpack_64_2x32_split_x', ('imul_2x32_64(is_used_once)', a, b)), ('imul', a, b)),
   (('unpack_64_2x32_split_x', ('umul_2x32_64(is_used_once)', a, b)), ('imul', a, b)),
   (('imul_2x32_64', a, b), ('pack_64_2x32_split', ('imul', a, b), ('imul_high', a, b)), 'options->lower_mul_2x32_64'),
   (('umul_2x32_64', a, b), ('pack_64_2x32_split', ('imul', a, b), ('umul_high', a, b)), 'options->lower_mul_2x32_64'),
   (('udiv', a, 1), a),
   (('idiv', a, 1), a),
   (('umod', a, 1), 0),
   (('imod', a, 1), 0),
   (('udiv', a, '#b(is_pos_power_of_two)'), ('ushr', a, ('find_lsb', b)), '!options->lower_bitops'),
   (('idiv', a, '#b(is_pos_power_of_two)'), ('imul', ('isign', a), ('ushr', ('iabs', a), ('find_lsb', b))), '!options->lower_bitops'),
   (('idiv', a, '#b(is_neg_power_of_two)'), ('ineg', ('imul', ('isign', a), ('ushr', ('iabs', a), ('find_lsb', ('iabs', b))))), '!options->lower_bitops'),
   (('umod', a, '#b(is_pos_power_of_two)'),    ('iand', a, ('isub', b, 1))),

   (('~fneg', ('fneg', a)), a),
   (('ineg', ('ineg', a)), a),
   (('fabs', ('fneg', a)), ('fabs', a)),
   (('fabs', ('u2f', a)), ('u2f', a)),
   (('iabs', ('iabs', a)), ('iabs', a)),
   (('iabs', ('ineg', a)), ('iabs', a)),
   (('f2b', ('fneg', a)), ('f2b', a)),
   (('i2b', ('ineg', a)), ('i2b', a)),
   (('~fadd', a, 0.0), a),
   (('iadd', a, 0), a),
   (('usadd_4x8', a, 0), a),
   (('usadd_4x8', a, ~0), ~0),
   (('~fadd', ('fmul', a, b), ('fmul', a, c)), ('fmul', a, ('fadd', b, c))),
   (('iadd', ('imul', a, b), ('imul', a, c)), ('imul', a, ('iadd', b, c))),
   (('iand', ('ior', a, b), ('ior', a, c)), ('ior', a, ('iand', b, c))),
   (('ior', ('iand', a, b), ('iand', a, c)), ('iand', a, ('ior', b, c))),
   (('~fadd', ('fneg', a), a), 0.0),
   (('iadd', ('ineg', a), a), 0),
   (('iadd', ('ineg', a), ('iadd', a, b)), b),
   (('iadd', a, ('iadd', ('ineg', a), b)), b),
   (('~fadd', ('fneg', a), ('fadd', a, b)), b),
   (('~fadd', a, ('fadd', ('fneg', a), b)), b),
   (('fadd', ('fsat', a), ('fsat', ('fneg', a))), ('fsat', ('fabs', a))),
   (('~fmul', a, 0.0), 0.0),
   (('imul', a, 0), 0),
   (('umul_unorm_4x8', a, 0), 0),
   (('umul_unorm_4x8', a, ~0), a),
   (('~fmul', a, 1.0), a),
   (('imul', a, 1), a),
   (('fmul', a, -1.0), ('fneg', a)),
   (('imul', a, -1), ('ineg', a)),
   # If a < 0: fsign(a)*a*a => -1*a*a => -a*a => abs(a)*a
   # If a > 0: fsign(a)*a*a => 1*a*a => a*a => abs(a)*a
   # If a == 0: fsign(a)*a*a => 0*0*0 => abs(0)*0
   (('fmul', ('fsign', a), ('fmul', a, a)), ('fmul', ('fabs', a), a)),
   (('fmul', ('fmul', ('fsign', a), a), a), ('fmul', ('fabs', a), a)),
   (('~ffma', 0.0, a, b), b),
   (('~ffma', a, b, 0.0), ('fmul', a, b)),
   (('ffma', 1.0, a, b), ('fadd', a, b)),
   (('ffma', -1.0, a, b), ('fadd', ('fneg', a), b)),
   (('~flrp', a, b, 0.0), a),
   (('~flrp', a, b, 1.0), b),
   (('~flrp', a, a, b), a),
   (('~flrp', 0.0, a, b), ('fmul', a, b)),

   # flrp(a, a + b, c) => a + flrp(0, b, c) => a + (b * c)
   (('~flrp', a, ('fadd(is_used_once)', a, b), c), ('fadd', ('fmul', b, c), a)),
]

# Float sizes
for s in [16, 32, 64]:
    optimizations.extend([
       (('~flrp@{}'.format(s), a, ('fadd', a, b), c), ('fadd', ('fmul', b, c), a), 'options->lower_flrp{}'.format(s)),
       (('~flrp@{}'.format(s), ('fadd', a, b), ('fadd', a, c), d), ('fadd', ('flrp', b, c, d), a), 'options->lower_flrp{}'.format(s)),
       (('~flrp@{}'.format(s), a, ('fmul(is_used_once)', a, b), c), ('fmul', ('flrp', 1.0, b, c), a), 'options->lower_flrp{}'.format(s)),

       (('~flrp@{}'.format(s), a, b, ('b2f', 'c@1')), ('bcsel', c, b, a), 'options->lower_flrp{}'.format(s)),

       (('~fadd@{}'.format(s), ('fmul', a, ('fadd', 1.0, ('fneg', c))), ('fmul', b, c)), ('flrp', a, b, c), '!options->lower_flrp{}'.format(s)),
       # These are the same as the previous three rules, but it depends on
       # 1-fsat(x) <=> fsat(1-x).  See below.
       (('~fadd@{}'.format(s), ('fmul', a, ('fsat', ('fadd', 1.0, ('fneg', c)))), ('fmul', b, ('fsat', c))), ('flrp', a, b, ('fsat', c)), '!options->lower_flrp{}'.format(s)),
       (('~fadd@{}'.format(s), a, ('fmul', c, ('fadd', b, ('fneg', a)))), ('flrp', a, b, c), '!options->lower_flrp{}'.format(s)),

       (('~fadd@{}'.format(s),    ('fmul', a, ('fadd', 1.0, ('fneg', ('b2f', 'c@1')))), ('fmul', b, ('b2f',  c))), ('bcsel', c, b, a), 'options->lower_flrp{}'.format(s)),
       (('~fadd@{}'.format(s), a, ('fmul', ('b2f', 'c@1'), ('fadd', b, ('fneg', a)))), ('bcsel', c, b, a), 'options->lower_flrp{}'.format(s)),

       # 1 - ((1 - a) * (1 - b))
       # 1 - (1 - a - b + a*b)
       # 1 - 1 + a + b - a*b
       # a + b - a*b
       # a + b*(1 - a)
       # b*(1 - a) + 1*a
       # flrp(b, 1, a)
       (('~fadd@{}'.format(s), 1.0, ('fneg', ('fmul', ('fadd', 1.0, ('fneg', a)), ('fadd', 1.0, ('fneg', b))))), ('flrp', b, 1.0, a), '!options->lower_flrp{}'.format(s)),
    ])

optimizations.extend([
   (('~flrp', ('fmul(is_used_once)', a, b), ('fmul(is_used_once)', a, c), d), ('fmul', ('flrp', b, c, d), a)),

   (('~flrp', a, 0.0, c), ('fadd', ('fmul', ('fneg', a), c), a)),
   (('ftrunc', a), ('bcsel', ('flt', a, 0.0), ('fneg', ('ffloor', ('fabs', a))), ('ffloor', ('fabs', a))), 'options->lower_ftrunc'),
   (('ffloor', a), ('fsub', a, ('ffract', a)), 'options->lower_ffloor'),
   (('fadd', a, ('fneg', ('ffract', a))), ('ffloor', a), '!options->lower_ffloor'),
   (('ffract', a), ('fsub', a, ('ffloor', a)), 'options->lower_ffract'),
   (('fceil', a), ('fneg', ('ffloor', ('fneg', a))), 'options->lower_fceil'),
   (('ffma@16', a, b, c), ('fadd', ('fmul', a, b), c), 'options->lower_ffma16'),
   (('ffma@32', a, b, c), ('fadd', ('fmul', a, b), c), 'options->lower_ffma32'),
   (('ffma@64', a, b, c), ('fadd', ('fmul', a, b), c), 'options->lower_ffma64'),
   # Always lower inexact ffma, because it will be fused back by late optimizations (nir_opt_algebraic_late).
   (('~ffma@16', a, b, c), ('fadd', ('fmul', a, b), c), 'options->fuse_ffma16'),
   (('~ffma@32', a, b, c), ('fadd', ('fmul', a, b), c), 'options->fuse_ffma32'),
   (('~ffma@64', a, b, c), ('fadd', ('fmul', a, b), c), 'options->fuse_ffma64'),

   (('~fmul', ('fadd', ('iand', ('ineg', ('b2i', 'a@bool')), ('fmul', b, c)), '#d'), '#e'),
    ('bcsel', a, ('fmul', ('fadd', ('fmul', b, c), d), e), ('fmul', d, e))),

   (('fdph', a, b), ('fdot4', ('vec4', 'a.x', 'a.y', 'a.z', 1.0), b), 'options->lower_fdph'),

   (('fdot4', ('vec4', a, b,   c,   1.0), d), ('fdph',  ('vec3', a, b, c), d), '!options->lower_fdph'),
   (('fdot4', ('vec4', a, 0.0, 0.0, 0.0), b), ('fmul', a, b)),
   (('fdot4', ('vec4', a, b,   0.0, 0.0), c), ('fdot2', ('vec2', a, b), c)),
   (('fdot4', ('vec4', a, b,   c,   0.0), d), ('fdot3', ('vec3', a, b, c), d)),

   (('fdot3', ('vec3', a, 0.0, 0.0), b), ('fmul', a, b)),
   (('fdot3', ('vec3', a, b,   0.0), c), ('fdot2', ('vec2', a, b), c)),

   (('fdot2', ('vec2', a, 0.0), b), ('fmul', a, b)),
   (('fdot2', a, 1.0), ('fadd', 'a.x', 'a.y')),

   # Lower fdot to fsum when it is available
   (('fdot2', a, b), ('fsum2', ('fmul', a, b)), 'options->lower_fdot'),
   (('fdot3', a, b), ('fsum3', ('fmul', a, b)), 'options->lower_fdot'),
   (('fdot4', a, b), ('fsum4', ('fmul', a, b)), 'options->lower_fdot'),
   (('fsum2', a), ('fadd', 'a.x', 'a.y'), 'options->lower_fdot'),

   # If x >= 0 and x <= 1: fsat(1 - x) == 1 - fsat(x) trivially
   # If x < 0: 1 - fsat(x) => 1 - 0 => 1 and fsat(1 - x) => fsat(> 1) => 1
   # If x > 1: 1 - fsat(x) => 1 - 1 => 0 and fsat(1 - x) => fsat(< 0) => 0
   (('~fadd', ('fneg(is_used_once)', ('fsat(is_used_once)', 'a(is_not_fmul)')), 1.0), ('fsat', ('fadd', 1.0, ('fneg', a)))),

   # (a * #b + #c) << #d
   # ((a * #b) << #d) + (#c << #d)
   # (a * (#b << #d)) + (#c << #d)
   (('ishl', ('iadd', ('imul', a, '#b'), '#c'), '#d'),
    ('iadd', ('imul', a, ('ishl', b, d)), ('ishl', c, d))),

   # (a * #b) << #c
   # a * (#b << #c)
   (('ishl', ('imul', a, '#b'), '#c'), ('imul', a, ('ishl', b, c))),
])

# Care must be taken here.  Shifts in NIR uses only the lower log2(bitsize)
# bits of the second source.  These replacements must correctly handle the
# case where (b % bitsize) + (c % bitsize) >= bitsize.
for s in [8, 16, 32, 64]:
   mask = (1 << s) - 1

   ishl = "ishl@{}".format(s)
   ishr = "ishr@{}".format(s)
   ushr = "ushr@{}".format(s)

   in_bounds = ('ult', ('iadd', ('iand', b, mask), ('iand', c, mask)), s)

   optimizations.extend([
       ((ishl, (ishl, a, '#b'), '#c'), ('bcsel', in_bounds, (ishl, a, ('iadd', b, c)), 0)),
       ((ushr, (ushr, a, '#b'), '#c'), ('bcsel', in_bounds, (ushr, a, ('iadd', b, c)), 0)),

       # To get get -1 for large shifts of negative values, ishr must instead
       # clamp the shift count to the maximum value.
       ((ishr, (ishr, a, '#b'), '#c'),
        (ishr, a, ('imin', ('iadd', ('iand', b, mask), ('iand', c, mask)), s - 1))),
   ])

# Optimize a pattern of address calculation created by DXVK where the offset is
# divided by 4 and then multipled by 4. This can be turned into an iand and the
# additions before can be reassociated to CSE the iand instruction.
for log2 in range(1, 7): # powers of two from 2 to 64
   v = 1 << log2
   mask = 0xffffffff & ~(v - 1)
   b_is_multiple = '#b(is_unsigned_multiple_of_{})'.format(v)

   optimizations.extend([
       # 'a >> #b << #b' -> 'a & ~((1 << #b) - 1)'
       (('ishl', ('ushr', a, log2), log2), ('iand', a, mask)),

       # Reassociate for improved CSE
       (('iand', ('iadd', a, b_is_multiple), mask), ('iadd', ('iand', a, mask), b)),
   ])

# To save space in the state tables, reduce to the set that is known to help.
# Previously, this was range(1, 32).  In addition, a couple rules inside the
# loop are commented out.  Revisit someday, probably after mesa/#2635 has some
# resolution.
for i in [1, 2, 16, 24]:
    lo_mask = 0xffffffff >> i
    hi_mask = (0xffffffff << i) & 0xffffffff

    optimizations.extend([
        # This pattern seems to only help in the soft-fp64 code.
        (('ishl@32', ('iand', 'a@32', lo_mask), i), ('ishl', a, i)),
#        (('ushr@32', ('iand', 'a@32', hi_mask), i), ('ushr', a, i)),
#        (('ishr@32', ('iand', 'a@32', hi_mask), i), ('ishr', a, i)),

        (('iand', ('ishl', 'a@32', i), hi_mask), ('ishl', a, i)),
        (('iand', ('ushr', 'a@32', i), lo_mask), ('ushr', a, i)),
#        (('iand', ('ishr', 'a@32', i), lo_mask), ('ushr', a, i)), # Yes, ushr is correct
    ])

optimizations.extend([
   # This is common for address calculations.  Reassociating may enable the
   # 'a<<c' to be CSE'd.  It also helps architectures that have an ISHLADD
   # instruction or a constant offset field for in load / store instructions.
   (('ishl', ('iadd', a, '#b'), '#c'), ('iadd', ('ishl', a, c), ('ishl', b, c))),

   # Comparison simplifications
   (('~inot', ('flt', a, b)), ('fge', a, b)),
   (('~inot', ('fge', a, b)), ('flt', a, b)),
   (('inot', ('feq', a, b)), ('fneu', a, b)),
   (('inot', ('fneu', a, b)), ('feq', a, b)),
   (('inot', ('ilt', a, b)), ('ige', a, b)),
   (('inot', ('ult', a, b)), ('uge', a, b)),
   (('inot', ('ige', a, b)), ('ilt', a, b)),
   (('inot', ('uge', a, b)), ('ult', a, b)),
   (('inot', ('ieq', a, b)), ('ine', a, b)),
   (('inot', ('ine', a, b)), ('ieq', a, b)),

   (('iand', ('feq', a, b), ('fneu', a, b)), False),
   (('iand', ('flt', a, b), ('flt', b, a)), False),
   (('iand', ('ieq', a, b), ('ine', a, b)), False),
   (('iand', ('ilt', a, b), ('ilt', b, a)), False),
   (('iand', ('ult', a, b), ('ult', b, a)), False),

   # This helps some shaders because, after some optimizations, they end up
   # with patterns like (-a < -b) || (b < a).  In an ideal world, this sort of
   # matching would be handled by CSE.
   (('flt', ('fneg', a), ('fneg', b)), ('flt', b, a)),
   (('fge', ('fneg', a), ('fneg', b)), ('fge', b, a)),
   (('feq', ('fneg', a), ('fneg', b)), ('feq', b, a)),
   (('fneu', ('fneg', a), ('fneg', b)), ('fneu', b, a)),
   (('flt', ('fneg', a), -1.0), ('flt', 1.0, a)),
   (('flt', -1.0, ('fneg', a)), ('flt', a, 1.0)),
   (('fge', ('fneg', a), -1.0), ('fge', 1.0, a)),
   (('fge', -1.0, ('fneg', a)), ('fge', a, 1.0)),
   (('fneu', ('fneg', a), -1.0), ('fneu', 1.0, a)),
   (('feq', -1.0, ('fneg', a)), ('feq', a, 1.0)),

   # flt(fsat(a), b > 0 && b < 1) is inexact if a is NaN (fsat(NaN) is 0)
   # because it returns True while flt(a, b) always returns False.
   (('~flt', ('fsat(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('flt', a, b)),
   (('flt', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a)), ('flt', b, a)),
   (('fge', ('fsat(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('fge', a, b)),
   # fge(b > 0 && b < 1, fsat(a)) is inexact if a is NaN (fsat(NaN) is 0)
   # because it returns True while fge(b, a) always returns False.
   (('~fge', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a)), ('fge', b, a)),
   (('feq', ('fsat(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('feq', a, b)),
   (('fneu', ('fsat(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('fneu', a, b)),

   (('fge', ('fsat(is_used_once)', a), 1.0), ('fge', a, 1.0)),
   (('flt', ('fsat(is_used_once)', a), 1.0), ('flt', a, 1.0)),
   (('fge', 0.0, ('fsat(is_used_once)', a)), ('fge', 0.0, a)),
   (('flt', 0.0, ('fsat(is_used_once)', a)), ('flt', 0.0, a)),

   # 0.0 >= b2f(a)
   # b2f(a) <= 0.0
   # b2f(a) == 0.0 because b2f(a) can only be 0 or 1
   # inot(a)
   (('fge', 0.0, ('b2f', 'a@1')), ('inot', a)),

   (('fge', ('fneg', ('b2f', 'a@1')), 0.0), ('inot', a)),

   (('fneu', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a, b)),
   (('fneu', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a, b)),
   (('fneu', ('bcsel', a, 1.0, ('b2f', 'b@1'))   , 0.0), ('ior', a, b)),
   (('fneu', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))),      ('ior', a, b)),
   (('fneu', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a, b)),
   (('fneu', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a, b)),
   (('fneu', ('bcsel', a, ('b2f', 'b@1'), 0.0)   , 0.0), ('iand', a, b)),
   (('fneu', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ixor', a, b)),
   (('fneu',          ('b2f', 'a@1') ,          ('b2f', 'b@1') ),      ('ixor', a, b)),
   (('fneu', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))),      ('ixor', a, b)),
   (('feq', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a, b))),
   (('feq', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a, b))),
   (('feq', ('bcsel', a, 1.0, ('b2f', 'b@1'))   , 0.0), ('inot', ('ior', a, b))),
   (('feq', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))),      ('inot', ('ior', a, b))),
   (('feq', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a, b))),
   (('feq', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a, b))),
   (('feq', ('bcsel', a, ('b2f', 'b@1'), 0.0)   , 0.0), ('inot', ('iand', a, b))),
   (('feq', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ieq', a, b)),
   (('feq',          ('b2f', 'a@1') ,          ('b2f', 'b@1') ),      ('ieq', a, b)),
   (('feq', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))),      ('ieq', a, b)),

   # -(b2f(a) + b2f(b)) < 0
   # 0 < b2f(a) + b2f(b)
   # 0 != b2f(a) + b2f(b)       b2f must be 0 or 1, so the sum is non-negative
   # a || b
   (('flt', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('ior', a, b)),
   (('flt', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('ior', a, b)),

   # -(b2f(a) + b2f(b)) >= 0
   # 0 >= b2f(a) + b2f(b)
   # 0 == b2f(a) + b2f(b)       b2f must be 0 or 1, so the sum is non-negative
   # !(a || b)
   (('fge', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('inot', ('ior', a, b))),
   (('fge', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('inot', ('ior', a, b))),

   (('flt', a, ('fneg', a)), ('flt', a, 0.0)),
   (('fge', a, ('fneg', a)), ('fge', a, 0.0)),

   # Some optimizations (below) convert things like (a < b || c < b) into
   # (min(a, c) < b).  However, this interfers with the previous optimizations
   # that try to remove comparisons with negated sums of b2f.  This just
   # breaks that apart.
   (('flt', ('fmin', c, ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')))), 0.0),
    ('ior', ('flt', c, 0.0), ('ior', a, b))),

   (('~flt', ('fadd', a, b), a), ('flt', b, 0.0)),
   (('~fge', ('fadd', a, b), a), ('fge', b, 0.0)),
   (('~feq', ('fadd', a, b), a), ('feq', b, 0.0)),
   (('~fneu', ('fadd', a, b), a), ('fneu', b, 0.0)),
   (('~flt',                        ('fadd(is_used_once)', a, '#b'),  '#c'), ('flt', a, ('fadd', c, ('fneg', b)))),
   (('~flt', ('fneg(is_used_once)', ('fadd(is_used_once)', a, '#b')), '#c'), ('flt', ('fneg', ('fadd', c, b)), a)),
   (('~fge',                        ('fadd(is_used_once)', a, '#b'),  '#c'), ('fge', a, ('fadd', c, ('fneg', b)))),
   (('~fge', ('fneg(is_used_once)', ('fadd(is_used_once)', a, '#b')), '#c'), ('fge', ('fneg', ('fadd', c, b)), a)),
   (('~feq',                        ('fadd(is_used_once)', a, '#b'),  '#c'), ('feq', a, ('fadd', c, ('fneg', b)))),
   (('~feq', ('fneg(is_used_once)', ('fadd(is_used_once)', a, '#b')), '#c'), ('feq', ('fneg', ('fadd', c, b)), a)),
   (('~fneu',                        ('fadd(is_used_once)', a, '#b'),  '#c'), ('fneu', a, ('fadd', c, ('fneg', b)))),
   (('~fneu', ('fneg(is_used_once)', ('fadd(is_used_once)', a, '#b')), '#c'), ('fneu', ('fneg', ('fadd', c, b)), a)),

   # Cannot remove the addition from ilt or ige due to overflow.
   (('ieq', ('iadd', a, b), a), ('ieq', b, 0)),
   (('ine', ('iadd', a, b), a), ('ine', b, 0)),

   # fmin(-b2f(a), b) >= 0.0
   # -b2f(a) >= 0.0 && b >= 0.0
   # -b2f(a) == 0.0 && b >= 0.0    -b2f can only be 0 or -1, never >0
   # b2f(a) == 0.0 && b >= 0.0
   # a == False && b >= 0.0
   # !a && b >= 0.0
   #
   # The fge in the second replacement is not a typo.  I leave the proof that
   # "fmin(-b2f(a), b) >= 0 <=> fmin(-b2f(a), b) == 0" as an exercise for the
   # reader.
   (('fge', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a), ('fge', b, 0.0))),
   (('feq', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a), ('fge', b, 0.0))),

   (('feq', ('b2f', 'a@1'), 0.0), ('inot', a)),
   (('~fneu', ('b2f', 'a@1'), 0.0), a),
   (('ieq', ('b2i', 'a@1'), 0),   ('inot', a)),
   (('ine', ('b2i', 'a@1'), 0),   a),

   (('fneu', ('u2f', a), 0.0), ('ine', a, 0)),
   (('feq', ('u2f', a), 0.0), ('ieq', a, 0)),
   (('fge', ('u2f', a), 0.0), True),
   (('fge', 0.0, ('u2f', a)), ('uge', 0, a)),    # ieq instead?
   (('flt', ('u2f', a), 0.0), False),
   (('flt', 0.0, ('u2f', a)), ('ult', 0, a)),    # ine instead?
   (('fneu', ('i2f', a), 0.0), ('ine', a, 0)),
   (('feq', ('i2f', a), 0.0), ('ieq', a, 0)),
   (('fge', ('i2f', a), 0.0), ('ige', a, 0)),
   (('fge', 0.0, ('i2f', a)), ('ige', 0, a)),
   (('flt', ('i2f', a), 0.0), ('ilt', a, 0)),
   (('flt', 0.0, ('i2f', a)), ('ilt', 0, a)),

   # 0.0 < fabs(a)
   # fabs(a) > 0.0
   # fabs(a) != 0.0 because fabs(a) must be >= 0
   # a != 0.0
   (('~flt', 0.0, ('fabs', a)), ('fneu', a, 0.0)),

   # -fabs(a) < 0.0
   # fabs(a) > 0.0
   (('~flt', ('fneg', ('fabs', a)), 0.0), ('fneu', a, 0.0)),

   # 0.0 >= fabs(a)
   # 0.0 == fabs(a)   because fabs(a) must be >= 0
   # 0.0 == a
   (('fge', 0.0, ('fabs', a)), ('feq', a, 0.0)),

   # -fabs(a) >= 0.0
   # 0.0 >= fabs(a)
   (('fge', ('fneg', ('fabs', a)), 0.0), ('feq', a, 0.0)),

   # (a >= 0.0) && (a <= 1.0) -> fsat(a) == a
   (('iand', ('fge', a, 0.0), ('fge', 1.0, a)), ('feq', a, ('fsat', a)), '!options->lower_fsat'),

   # (a < 0.0) || (a > 1.0)
   # !(!(a < 0.0) && !(a > 1.0))
   # !((a >= 0.0) && (a <= 1.0))
   # !(a == fsat(a))
   # a != fsat(a)
   (('ior', ('flt', a, 0.0), ('flt', 1.0, a)), ('fneu', a, ('fsat', a)), '!options->lower_fsat'),

   (('fmax',                        ('b2f(is_used_once)', 'a@1'),           ('b2f', 'b@1')),           ('b2f', ('ior', a, b))),
   (('fmax', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('ior', a, b)))),
   (('fmin',                        ('b2f(is_used_once)', 'a@1'),           ('b2f', 'b@1')),           ('b2f', ('iand', a, b))),
   (('fmin', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('iand', a, b)))),

   # fmin(b2f(a), b)
   # bcsel(a, fmin(b2f(a), b), fmin(b2f(a), b))
   # bcsel(a, fmin(b2f(True), b), fmin(b2f(False), b))
   # bcsel(a, fmin(1.0, b), fmin(0.0, b))
   #
   # Since b is a constant, constant folding will eliminate the fmin and the
   # fmax.  If b is > 1.0, the bcsel will be replaced with a b2f.
   (('fmin', ('b2f', 'a@1'), '#b'), ('bcsel', a, ('fmin', b, 1.0), ('fmin', b, 0.0))),

   (('flt', ('fadd(is_used_once)', a, ('fneg', b)), 0.0), ('flt', a, b)),

   (('fge', ('fneg', ('fabs', a)), 0.0), ('feq', a, 0.0)),
   (('~bcsel', ('flt', b, a), b, a), ('fmin', a, b)),
   (('~bcsel', ('flt', a, b), b, a), ('fmax', a, b)),
   (('~bcsel', ('fge', a, b), b, a), ('fmin', a, b)),
   (('~bcsel', ('fge', b, a), b, a), ('fmax', a, b)),
   (('bcsel', ('i2b', a), b, c), ('bcsel', ('ine', a, 0), b, c)),
   (('bcsel', ('inot', a), b, c), ('bcsel', a, c, b)),
   (('bcsel', a, ('bcsel', a, b, c), d), ('bcsel', a, b, d)),
   (('bcsel', a, b, ('bcsel', a, c, d)), ('bcsel', a, b, d)),
   (('bcsel', a, ('bcsel', b, c, d), ('bcsel(is_used_once)', b, c, 'e')), ('bcsel', b, c, ('bcsel', a, d, 'e'))),
   (('bcsel', a, ('bcsel(is_used_once)', b, c, d), ('bcsel', b, c, 'e')), ('bcsel', b, c, ('bcsel', a, d, 'e'))),
   (('bcsel', a, ('bcsel', b, c, d), ('bcsel(is_used_once)', b, 'e', d)), ('bcsel', b, ('bcsel', a, c, 'e'), d)),
   (('bcsel', a, ('bcsel(is_used_once)', b, c, d), ('bcsel', b, 'e', d)), ('bcsel', b, ('bcsel', a, c, 'e'), d)),
   (('bcsel', a, True, b), ('ior', a, b)),
   (('bcsel', a, a, b), ('ior', a, b)),
   (('bcsel', a, b, False), ('iand', a, b)),
   (('bcsel', a, b, a), ('iand', a, b)),
   (('~fmin', a, a), a),
   (('~fmax', a, a), a),
   (('imin', a, a), a),
   (('imax', a, a), a),
   (('umin', a, a), a),
   (('umax', a, a), a),
   (('fmax', ('fmax', a, b), b), ('fmax', a, b)),
   (('umax', ('umax', a, b), b), ('umax', a, b)),
   (('imax', ('imax', a, b), b), ('imax', a, b)),
   (('fmin', ('fmin', a, b), b), ('fmin', a, b)),
   (('umin', ('umin', a, b), b), ('umin', a, b)),
   (('imin', ('imin', a, b), b), ('imin', a, b)),
])

# Integer sizes
for s in [8, 16, 32, 64]:
    optimizations.extend([
       (('iand@{}'.format(s), a, ('inot', ('ishr', a, s - 1))), ('imax', a, 0)),

       # Simplify logic to detect sign of an integer.
       (('ieq', ('iand', 'a@{}'.format(s), 1 << (s - 1)), 0),            ('ige', a, 0)),
       (('ine', ('iand', 'a@{}'.format(s), 1 << (s - 1)), 1 << (s - 1)), ('ige', a, 0)),
       (('ine', ('iand', 'a@{}'.format(s), 1 << (s - 1)), 0),            ('ilt', a, 0)),
       (('ieq', ('iand', 'a@{}'.format(s), 1 << (s - 1)), 1 << (s - 1)), ('ilt', a, 0)),
       (('ine', ('ushr', 'a@{}'.format(s), s - 1), 0), ('ilt', a, 0)),
       (('ieq', ('ushr', 'a@{}'.format(s), s - 1), 0), ('ige', a, 0)),
       (('ieq', ('ushr', 'a@{}'.format(s), s - 1), 1), ('ilt', a, 0)),
       (('ine', ('ushr', 'a@{}'.format(s), s - 1), 1), ('ige', a, 0)),
       (('ine', ('ishr', 'a@{}'.format(s), s - 1), 0), ('ilt', a, 0)),
       (('ieq', ('ishr', 'a@{}'.format(s), s - 1), 0), ('ige', a, 0)),
       (('ieq', ('ishr', 'a@{}'.format(s), s - 1), -1), ('ilt', a, 0)),
       (('ine', ('ishr', 'a@{}'.format(s), s - 1), -1), ('ige', a, 0)),
    ])

optimizations.extend([
   (('fmin', a, ('fneg', a)), ('fneg', ('fabs', a))),
   (('imin', a, ('ineg', a)), ('ineg', ('iabs', a))),
   (('fmin', a, ('fneg', ('fabs', a))), ('fneg', ('fabs', a))),
   (('imin', a, ('ineg', ('iabs', a))), ('ineg', ('iabs', a))),
   (('~fmin', a, ('fabs', a)), a),
   (('imin', a, ('iabs', a)), a),
   (('~fmax', a, ('fneg', ('fabs', a))), a),
   (('imax', a, ('ineg', ('iabs', a))), a),
   (('fmax', a, ('fabs', a)), ('fabs', a)),
   (('imax', a, ('iabs', a)), ('iabs', a)),
   (('fmax', a, ('fneg', a)), ('fabs', a)),
   (('imax', a, ('ineg', a)), ('iabs', a), '!options->lower_iabs'),
   (('~fmax', ('fabs', a), 0.0), ('fabs', a)),
   (('fmin', ('fmax', a, 0.0), 1.0), ('fsat', a), '!options->lower_fsat'),
   # fmax(fmin(a, 1.0), 0.0) is inexact because it returns 1.0 on NaN, while
   # fsat(a) returns 0.0.
   (('~fmax', ('fmin', a, 1.0), 0.0), ('fsat', a), '!options->lower_fsat'),
   # fmin(fmax(a, -1.0), 0.0) is inexact because it returns -1.0 on NaN, while
   # fneg(fsat(fneg(a))) returns -0.0 on NaN.
   (('~fmin', ('fmax', a, -1.0),  0.0), ('fneg', ('fsat', ('fneg', a))), '!options->lower_fsat'),
   # fmax(fmin(a, 0.0), -1.0) is inexact because it returns 0.0 on NaN, while
   # fneg(fsat(fneg(a))) returns -0.0 on NaN. This only matters if
   # SignedZeroInfNanPreserve is set, but we don't currently have any way of
   # representing this in the optimizations other than the usual ~.
   (('~fmax', ('fmin', a,  0.0), -1.0), ('fneg', ('fsat', ('fneg', a))), '!options->lower_fsat'),
   (('fsat', ('fsign', a)), ('b2f', ('flt', 0.0, a))),
   (('fsat', ('b2f', a)), ('b2f', a)),
   (('fsat', a), ('fmin', ('fmax', a, 0.0), 1.0), 'options->lower_fsat'),
   (('fsat', ('fsat', a)), ('fsat', a)),
   (('fsat', ('fneg(is_used_once)', ('fadd(is_used_once)', a, b))), ('fsat', ('fadd', ('fneg', a), ('fneg', b))), '!options->lower_fsat'),
   (('fsat', ('fneg(is_used_once)', ('fmul(is_used_once)', a, b))), ('fsat', ('fmul', ('fneg', a), b)), '!options->lower_fsat'),
   (('fsat', ('fabs(is_used_once)', ('fmul(is_used_once)', a, b))), ('fsat', ('fmul', ('fabs', a), ('fabs', b))), '!options->lower_fsat'),
   (('fmin', ('fmax', ('fmin', ('fmax', a, b), c), b), c), ('fmin', ('fmax', a, b), c)),
   (('imin', ('imax', ('imin', ('imax', a, b), c), b), c), ('imin', ('imax', a, b), c)),
   (('umin', ('umax', ('umin', ('umax', a, b), c), b), c), ('umin', ('umax', a, b), c)),
   # Both the left and right patterns are "b" when isnan(a), so this is exact.
   (('fmax', ('fsat', a), '#b(is_zero_to_one)'), ('fsat', ('fmax', a, b))),
   # The left pattern is 0.0 when isnan(a) (because fmin(fsat(NaN), b) ->
   # fmin(0.0, b)) while the right one is "b", so this optimization is inexact.
   (('~fmin', ('fsat', a), '#b(is_zero_to_one)'), ('fsat', ('fmin', a, b))),

   # If a in [0,b] then b-a is also in [0,b].  Since b in [0,1], max(b-a, 0) =
   # fsat(b-a).
   #
   # If a > b, then b-a < 0 and max(b-a, 0) = fsat(b-a) = 0
   #
   # This should be NaN safe since max(NaN, 0) = fsat(NaN) = 0.
   (('fmax', ('fadd(is_used_once)', ('fneg', 'a(is_not_negative)'), '#b(is_zero_to_one)'), 0.0),
    ('fsat', ('fadd', ('fneg',  a), b)), '!options->lower_fsat'),

   (('extract_u8', ('imin', ('imax', a, 0), 0xff), 0), ('imin', ('imax', a, 0), 0xff)),
   (('~ior', ('flt(is_used_once)', a, b), ('flt', a, c)), ('flt', a, ('fmax', b, c))),
   (('~ior', ('flt(is_used_once)', a, c), ('flt', b, c)), ('flt', ('fmin', a, b), c)),
   (('~ior', ('fge(is_used_once)', a, b), ('fge', a, c)), ('fge', a, ('fmin', b, c))),
   (('~ior', ('fge(is_used_once)', a, c), ('fge', b, c)), ('fge', ('fmax', a, b), c)),
   (('~ior', ('flt', a, '#b'), ('flt', a, '#c')), ('flt', a, ('fmax', b, c))),
   (('~ior', ('flt', '#a', c), ('flt', '#b', c)), ('flt', ('fmin', a, b), c)),
   (('~ior', ('fge', a, '#b'), ('fge', a, '#c')), ('fge', a, ('fmin', b, c))),
   (('~ior', ('fge', '#a', c), ('fge', '#b', c)), ('fge', ('fmax', a, b), c)),
   (('~iand', ('flt(is_used_once)', a, b), ('flt', a, c)), ('flt', a, ('fmin', b, c))),
   (('~iand', ('flt(is_used_once)', a, c), ('flt', b, c)), ('flt', ('fmax', a, b), c)),
   (('~iand', ('fge(is_used_once)', a, b), ('fge', a, c)), ('fge', a, ('fmax', b, c))),
   (('~iand', ('fge(is_used_once)', a, c), ('fge', b, c)), ('fge', ('fmin', a, b), c)),
   (('~iand', ('flt', a, '#b'), ('flt', a, '#c')), ('flt', a, ('fmin', b, c))),
   (('~iand', ('flt', '#a', c), ('flt', '#b', c)), ('flt', ('fmax', a, b), c)),
   (('~iand', ('fge', a, '#b'), ('fge', a, '#c')), ('fge', a, ('fmax', b, c))),
   (('~iand', ('fge', '#a', c), ('fge', '#b', c)), ('fge', ('fmin', a, b), c)),

   (('ior', ('ilt(is_used_once)', a, b), ('ilt', a, c)), ('ilt', a, ('imax', b, c))),
   (('ior', ('ilt(is_used_once)', a, c), ('ilt', b, c)), ('ilt', ('imin', a, b), c)),
   (('ior', ('ige(is_used_once)', a, b), ('ige', a, c)), ('ige', a, ('imin', b, c))),
   (('ior', ('ige(is_used_once)', a, c), ('ige', b, c)), ('ige', ('imax', a, b), c)),
   (('ior', ('ult(is_used_once)', a, b), ('ult', a, c)), ('ult', a, ('umax', b, c))),
   (('ior', ('ult(is_used_once)', a, c), ('ult', b, c)), ('ult', ('umin', a, b), c)),
   (('ior', ('uge(is_used_once)', a, b), ('uge', a, c)), ('uge', a, ('umin', b, c))),
   (('ior', ('uge(is_used_once)', a, c), ('uge', b, c)), ('uge', ('umax', a, b), c)),
   (('iand', ('ilt(is_used_once)', a, b), ('ilt', a, c)), ('ilt', a, ('imin', b, c))),
   (('iand', ('ilt(is_used_once)', a, c), ('ilt', b, c)), ('ilt', ('imax', a, b), c)),
   (('iand', ('ige(is_used_once)', a, b), ('ige', a, c)), ('ige', a, ('imax', b, c))),
   (('iand', ('ige(is_used_once)', a, c), ('ige', b, c)), ('ige', ('imin', a, b), c)),
   (('iand', ('ult(is_used_once)', a, b), ('ult', a, c)), ('ult', a, ('umin', b, c))),
   (('iand', ('ult(is_used_once)', a, c), ('ult', b, c)), ('ult', ('umax', a, b), c)),
   (('iand', ('uge(is_used_once)', a, b), ('uge', a, c)), ('uge', a, ('umax', b, c))),
   (('iand', ('uge(is_used_once)', a, c), ('uge', b, c)), ('uge', ('umin', a, b), c)),
])

# Float sizes
for s in [16, 32, 64]:
    fp_one = {16: 0x3c00, 32: 0x3f800000, 64: 0x3ff0000000000000}[s]

    optimizations.extend([
       # These derive from the previous patterns with the application of b < 0 <=>
       # 0 < -b.  The transformation should be applied if either comparison is
       # used once as this ensures that the number of comparisons will not
       # increase.  The sources to the ior and iand are not symmetric, so the
       # rules have to be duplicated to get this behavior.
       (('~ior', ('flt(is_used_once)', 0.0, 'a@{}'.format(s)), ('flt', 'b@{}'.format(s), 0.0)), ('flt', 0.0, ('fmax', a, ('fneg', b)))),
       (('~ior', ('flt', 0.0, 'a@{}'.format(s)), ('flt(is_used_once)', 'b@{}'.format(s), 0.0)), ('flt', 0.0, ('fmax', a, ('fneg', b)))),
       (('~ior', ('fge(is_used_once)', 0.0, 'a@{}'.format(s)), ('fge', 'b@{}'.format(s), 0.0)), ('fge', 0.0, ('fmin', a, ('fneg', b)))),
       (('~ior', ('fge', 0.0, 'a@{}'.format(s)), ('fge(is_used_once)', 'b@{}'.format(s), 0.0)), ('fge', 0.0, ('fmin', a, ('fneg', b)))),
       (('~iand', ('flt(is_used_once)', 0.0, 'a@{}'.format(s)), ('flt', 'b@{}'.format(s), 0.0)), ('flt', 0.0, ('fmin', a, ('fneg', b)))),
       (('~iand', ('flt', 0.0, 'a@{}'.format(s)), ('flt(is_used_once)', 'b@{}'.format(s), 0.0)), ('flt', 0.0, ('fmin', a, ('fneg', b)))),
       (('~iand', ('fge(is_used_once)', 0.0, 'a@{}'.format(s)), ('fge', 'b@{}'.format(s), 0.0)), ('fge', 0.0, ('fmax', a, ('fneg', b)))),
       (('~iand', ('fge', 0.0, 'a@{}'.format(s)), ('fge(is_used_once)', 'b@{}'.format(s), 0.0)), ('fge', 0.0, ('fmax', a, ('fneg', b)))),

       # The (i2f32, ...) part is an open-coded fsign.  When that is combined with
       # the bcsel, it's basically copysign(1.0, a).  There is no copysign in NIR,
       # so emit an open-coded version of that.
       (('bcsel@{}'.format(s), ('feq', a, 0.0), 1.0, ('i2f{}'.format(s), ('iadd', ('b2i{}'.format(s), ('flt', 0.0, 'a@{}'.format(s))), ('ineg', ('b2i{}'.format(s), ('flt', 'a@{}'.format(s), 0.0)))))),
        ('ior', fp_one, ('iand', a, 1 << (s - 1)))),

       (('bcsel', a, ('b2f(is_used_once)', 'b@{}'.format(s)), ('b2f', 'c@{}'.format(s))), ('b2f', ('bcsel', a, b, c))),

       # The C spec says, "If the value of the integral part cannot be represented
       # by the integer type, the behavior is undefined."  "Undefined" can mean
       # "the conversion doesn't happen at all."
       (('~i2f{}'.format(s), ('f2i', 'a@{}'.format(s))), ('ftrunc', a)),

       # Ironically, mark these as imprecise because removing the conversions may
       # preserve more precision than doing the conversions (e.g.,
       # uint(float(0x81818181u)) == 0x81818200).
       (('~f2i{}'.format(s), ('i2f', 'a@{}'.format(s))), a),
       (('~f2i{}'.format(s), ('u2f', 'a@{}'.format(s))), a),
       (('~f2u{}'.format(s), ('i2f', 'a@{}'.format(s))), a),
       (('~f2u{}'.format(s), ('u2f', 'a@{}'.format(s))), a),

       (('fadd', ('b2f{}'.format(s), ('flt', 0.0, 'a@{}'.format(s))), ('fneg', ('b2f{}'.format(s), ('flt', 'a@{}'.format(s), 0.0)))), ('fsign', a), '!options->lower_fsign'),
       (('iadd', ('b2i{}'.format(s), ('flt', 0, 'a@{}'.format(s))), ('ineg', ('b2i{}'.format(s), ('flt', 'a@{}'.format(s), 0)))), ('f2i{}'.format(s), ('fsign', a)), '!options->lower_fsign'),
    ])

    # float? -> float? -> floatS ==> float? -> floatS
    (('~f2f{}'.format(s), ('f2f', a)), ('f2f{}'.format(s), a)),

    # int? -> float? -> floatS ==> int? -> floatS
    (('~f2f{}'.format(s), ('u2f', a)), ('u2f{}'.format(s), a)),
    (('~f2f{}'.format(s), ('i2f', a)), ('i2f{}'.format(s), a)),

    # float? -> float? -> intS ==> float? -> intS
    (('~f2u{}'.format(s), ('f2f', a)), ('f2u{}'.format(s), a)),
    (('~f2i{}'.format(s), ('f2f', a)), ('f2i{}'.format(s), a)),

    for B in [32, 64]:
        if s < B:
            optimizations.extend([
               # S = smaller, B = bigger
               # typeS -> typeB -> typeS ==> identity
               (('f2f{}'.format(s), ('f2f{}'.format(B), 'a@{}'.format(s))), a),
               (('i2i{}'.format(s), ('i2i{}'.format(B), 'a@{}'.format(s))), a),
               (('u2u{}'.format(s), ('u2u{}'.format(B), 'a@{}'.format(s))), a),

               # bool1 -> typeB -> typeS ==> bool1 -> typeS
               (('f2f{}'.format(s), ('b2f{}'.format(B), 'a@1')), ('b2f{}'.format(s), a)),
               (('i2i{}'.format(s), ('b2i{}'.format(B), 'a@1')), ('b2i{}'.format(s), a)),
               (('u2u{}'.format(s), ('b2i{}'.format(B), 'a@1')), ('b2i{}'.format(s), a)),

               # floatS -> floatB -> intB ==> floatS -> intB
               (('f2u{}'.format(B), ('f2f{}'.format(B), 'a@{}'.format(s))), ('f2u{}'.format(B), a)),
               (('f2i{}'.format(B), ('f2f{}'.format(B), 'a@{}'.format(s))), ('f2i{}'.format(B), a)),

               # int? -> floatB -> floatS ==> int? -> floatS
               (('f2f{}'.format(s), ('u2f{}'.format(B), a)), ('u2f{}'.format(s), a)),
               (('f2f{}'.format(s), ('i2f{}'.format(B), a)), ('i2f{}'.format(s), a)),

               # intS -> intB -> floatB ==> intS -> floatB
               (('u2f{}'.format(B), ('u2u{}'.format(B), 'a@{}'.format(s))), ('u2f{}'.format(B), a)),
               (('i2f{}'.format(B), ('i2i{}'.format(B), 'a@{}'.format(s))), ('i2f{}'.format(B), a)),
            ])

# mediump variants of the above
optimizations.extend([
    # int32 -> float32 -> float16 ==> int32 -> float16
    (('f2fmp', ('u2f32', 'a@32')), ('u2fmp', a)),
    (('f2fmp', ('i2f32', 'a@32')), ('i2fmp', a)),

    # float32 -> float16 -> int16 ==> float32 -> int16
    (('f2u16', ('f2fmp', 'a@32')), ('f2u16', a)),
    (('f2i16', ('f2fmp', 'a@32')), ('f2i16', a)),

    # float32 -> int32 -> int16 ==> float32 -> int16
    (('i2imp', ('f2u32', 'a@32')), ('f2ump', a)),
    (('i2imp', ('f2i32', 'a@32')), ('f2imp', a)),

    # int32 -> int16 -> float16 ==> int32 -> float16
    (('u2f16', ('i2imp', 'a@32')), ('u2f16', a)),
    (('i2f16', ('i2imp', 'a@32')), ('i2f16', a)),
])

# Integer sizes
for s in [8, 16, 32, 64]:
    optimizations.extend([
       (('iand', ('ieq', 'a@{}'.format(s), 0), ('ieq', 'b@{}'.format(s), 0)), ('ieq', ('ior', a, b), 0), 'options->lower_umax'),
       (('ior',  ('ine', 'a@{}'.format(s), 0), ('ine', 'b@{}'.format(s), 0)), ('ine', ('ior', a, b), 0), 'options->lower_umin'),
       (('iand', ('ieq', 'a@{}'.format(s), 0), ('ieq', 'b@{}'.format(s), 0)), ('ieq', ('umax', a, b), 0), '!options->lower_umax'),
       (('ior',  ('ieq', 'a@{}'.format(s), 0), ('ieq', 'b@{}'.format(s), 0)), ('ieq', ('umin', a, b), 0), '!options->lower_umin'),
       (('iand', ('ine', 'a@{}'.format(s), 0), ('ine', 'b@{}'.format(s), 0)), ('ine', ('umin', a, b), 0), '!options->lower_umin'),
       (('ior',  ('ine', 'a@{}'.format(s), 0), ('ine', 'b@{}'.format(s), 0)), ('ine', ('umax', a, b), 0), '!options->lower_umax'),

       # True/False are ~0 and 0 in NIR.  b2i of True is 1, and -1 is ~0 (True).
       (('ineg', ('b2i{}'.format(s), 'a@{}'.format(s))), a),

       # SM5 32-bit shifts are defined to use the 5 least significant bits (or 4 bits for 16 bits)
       (('ishl', 'a@{}'.format(s), ('iand', s - 1, b)), ('ishl', a, b)),
       (('ishr', 'a@{}'.format(s), ('iand', s - 1, b)), ('ishr', a, b)),
       (('ushr', 'a@{}'.format(s), ('iand', s - 1, b)), ('ushr', a, b)),
    ])

optimizations.extend([
   # Common pattern like 'if (i == 0 || i == 1 || ...)'
   (('ior', ('ieq', a, 0), ('ieq', a, 1)), ('uge', 1, a)),
   (('ior', ('uge', 1, a), ('ieq', a, 2)), ('uge', 2, a)),
   (('ior', ('uge', 2, a), ('ieq', a, 3)), ('uge', 3, a)),

   (('ior', a, ('ieq', a, False)), True),
   (('ior', a, ('inot', a)), -1),

   (('ine', ('ineg', ('b2i', 'a@1')), ('ineg', ('b2i', 'b@1'))), ('ine', a, b)),
   (('b2i', ('ine', 'a@1', 'b@1')), ('b2i', ('ixor', a, b))),

   # This pattern occurs coutresy of __flt64_nonnan in the soft-fp64 code.
   # The first part of the iand comes from the !__feq64_nonnan.
   #
   # The second pattern is a reformulation of the first based on the relation
   # (a == 0 || y == 0) <=> umin(a, y) == 0, where b in the first equation
   # happens to be y == 0.
   (('iand', ('inot', ('iand', ('ior', ('ieq', a, 0),  b), c)), ('ilt', a, 0)),
    ('iand', ('inot', ('iand',                         b , c)), ('ilt', a, 0))),
   (('iand', ('inot', ('iand', ('ieq', ('umin', a, b), 0), c)), ('ilt', a, 0)),
    ('iand', ('inot', ('iand', ('ieq',             b , 0), c)), ('ilt', a, 0))),

   # These patterns can result when (a < b || a < c) => (a < min(b, c))
   # transformations occur before constant propagation and loop-unrolling.
   (('~flt', a, ('fmax', b, a)), ('flt', a, b)),
   (('~flt', ('fmin', a, b), a), ('flt', b, a)),
   (('~fge', a, ('fmin', b, a)), True),
   (('~fge', ('fmax', a, b), a), True),
   (('~flt', a, ('fmin', b, a)), False),
   (('~flt', ('fmax', a, b), a), False),
   (('~fge', a, ('fmax', b, a)), ('fge', a, b)),
   (('~fge', ('fmin', a, b), a), ('fge', b, a)),

   (('ilt', a, ('imax', b, a)), ('ilt', a, b)),
   (('ilt', ('imin', a, b), a), ('ilt', b, a)),
   (('ige', a, ('imin', b, a)), True),
   (('ige', ('imax', a, b), a), True),
   (('ult', a, ('umax', b, a)), ('ult', a, b)),
   (('ult', ('umin', a, b), a), ('ult', b, a)),
   (('uge', a, ('umin', b, a)), True),
   (('uge', ('umax', a, b), a), True),
   (('ilt', a, ('imin', b, a)), False),
   (('ilt', ('imax', a, b), a), False),
   (('ige', a, ('imax', b, a)), ('ige', a, b)),
   (('ige', ('imin', a, b), a), ('ige', b, a)),
   (('ult', a, ('umin', b, a)), False),
   (('ult', ('umax', a, b), a), False),
   (('uge', a, ('umax', b, a)), ('uge', a, b)),
   (('uge', ('umin', a, b), a), ('uge', b, a)),
   (('ult', a, ('iand', b, a)), False),
   (('ult', ('ior', a, b), a), False),
   (('uge', a, ('iand', b, a)), True),
   (('uge', ('ior', a, b), a), True),

   (('ilt', '#a', ('imax', '#b', c)), ('ior', ('ilt', a, b), ('ilt', a, c))),
   (('ilt', ('imin', '#a', b), '#c'), ('ior', ('ilt', a, c), ('ilt', b, c))),
   (('ige', '#a', ('imin', '#b', c)), ('ior', ('ige', a, b), ('ige', a, c))),
   (('ige', ('imax', '#a', b), '#c'), ('ior', ('ige', a, c), ('ige', b, c))),
   (('ult', '#a', ('umax', '#b', c)), ('ior', ('ult', a, b), ('ult', a, c))),
   (('ult', ('umin', '#a', b), '#c'), ('ior', ('ult', a, c), ('ult', b, c))),
   (('uge', '#a', ('umin', '#b', c)), ('ior', ('uge', a, b), ('uge', a, c))),
   (('uge', ('umax', '#a', b), '#c'), ('ior', ('uge', a, c), ('uge', b, c))),
   (('ilt', '#a', ('imin', '#b', c)), ('iand', ('ilt', a, b), ('ilt', a, c))),
   (('ilt', ('imax', '#a', b), '#c'), ('iand', ('ilt', a, c), ('ilt', b, c))),
   (('ige', '#a', ('imax', '#b', c)), ('iand', ('ige', a, b), ('ige', a, c))),
   (('ige', ('imin', '#a', b), '#c'), ('iand', ('ige', a, c), ('ige', b, c))),
   (('ult', '#a', ('umin', '#b', c)), ('iand', ('ult', a, b), ('ult', a, c))),
   (('ult', ('umax', '#a', b), '#c'), ('iand', ('ult', a, c), ('ult', b, c))),
   (('uge', '#a', ('umax', '#b', c)), ('iand', ('uge', a, b), ('uge', a, c))),
   (('uge', ('umin', '#a', b), '#c'), ('iand', ('uge', a, c), ('uge', b, c))),

   # Thanks to sign extension, the ishr(a, b) is negative if and only if a is
   # negative.
   (('bcsel', ('ilt', a, 0), ('ineg', ('ishr', a, b)), ('ishr', a, b)),
    ('iabs', ('ishr', a, b))),
   (('iabs', ('ishr', ('iabs', a), b)), ('ishr', ('iabs', a), b)),

   (('fabs', ('slt', a, b)), ('slt', a, b)),
   (('fabs', ('sge', a, b)), ('sge', a, b)),
   (('fabs', ('seq', a, b)), ('seq', a, b)),
   (('fabs', ('sne', a, b)), ('sne', a, b)),
   (('slt', a, b), ('b2f', ('flt', a, b)), 'options->lower_scmp'),
   (('sge', a, b), ('b2f', ('fge', a, b)), 'options->lower_scmp'),
   (('seq', a, b), ('b2f', ('feq', a, b)), 'options->lower_scmp'),
   (('sne', a, b), ('b2f', ('fneu', a, b)), 'options->lower_scmp'),
   (('seq', ('seq', a, b), 1.0), ('seq', a, b)),
   (('seq', ('sne', a, b), 1.0), ('sne', a, b)),
   (('seq', ('slt', a, b), 1.0), ('slt', a, b)),
   (('seq', ('sge', a, b), 1.0), ('sge', a, b)),
   (('sne', ('seq', a, b), 0.0), ('seq', a, b)),
   (('sne', ('sne', a, b), 0.0), ('sne', a, b)),
   (('sne', ('slt', a, b), 0.0), ('slt', a, b)),
   (('sne', ('sge', a, b), 0.0), ('sge', a, b)),
   (('seq', ('seq', a, b), 0.0), ('sne', a, b)),
   (('seq', ('sne', a, b), 0.0), ('seq', a, b)),
   (('seq', ('slt', a, b), 0.0), ('sge', a, b)),
   (('seq', ('sge', a, b), 0.0), ('slt', a, b)),
   (('sne', ('seq', a, b), 1.0), ('sne', a, b)),
   (('sne', ('sne', a, b), 1.0), ('seq', a, b)),
   (('sne', ('slt', a, b), 1.0), ('sge', a, b)),
   (('sne', ('sge', a, b), 1.0), ('slt', a, b)),
   (('fall_equal2', a, b), ('fmin', ('seq', 'a.x', 'b.x'), ('seq', 'a.y', 'b.y')), 'options->lower_vector_cmp'),
   (('fall_equal3', a, b), ('seq', ('fany_nequal3', a, b), 0.0), 'options->lower_vector_cmp'),
   (('fall_equal4', a, b), ('seq', ('fany_nequal4', a, b), 0.0), 'options->lower_vector_cmp'),
   (('fany_nequal2', a, b), ('fmax', ('sne', 'a.x', 'b.x'), ('sne', 'a.y', 'b.y')), 'options->lower_vector_cmp'),
   (('fany_nequal3', a, b), ('fsat', ('fdot3', ('sne', a, b), ('sne', a, b))), 'options->lower_vector_cmp'),
   (('fany_nequal4', a, b), ('fsat', ('fdot4', ('sne', a, b), ('sne', a, b))), 'options->lower_vector_cmp'),

   (('ball_iequal2', a, b), ('iand', ('ieq', 'a.x', 'b.x'), ('ieq', 'a.y', 'b.y')), 'options->lower_vector_cmp'),
   (('ball_iequal3', a, b), ('iand', ('iand', ('ieq', 'a.x', 'b.x'), ('ieq', 'a.y', 'b.y')), ('ieq', 'a.z', 'b.z')), 'options->lower_vector_cmp'),
   (('ball_iequal4', a, b), ('iand', ('iand', ('ieq', 'a.x', 'b.x'), ('ieq', 'a.y', 'b.y')), ('iand', ('ieq', 'a.z', 'b.z'), ('ieq', 'a.w', 'b.w'))), 'options->lower_vector_cmp'),

   (('bany_inequal2', a, b), ('ior', ('ine', 'a.x', 'b.x'), ('ine', 'a.y', 'b.y')), 'options->lower_vector_cmp'),
   (('bany_inequal3', a, b), ('ior', ('ior', ('ine', 'a.x', 'b.x'), ('ine', 'a.y', 'b.y')), ('ine', 'a.z', 'b.z')), 'options->lower_vector_cmp'),
   (('bany_inequal4', a, b), ('ior', ('ior', ('ine', 'a.x', 'b.x'), ('ine', 'a.y', 'b.y')), ('ior', ('ine', 'a.z', 'b.z'), ('ine', 'a.w', 'b.w'))), 'options->lower_vector_cmp'),

   (('ball_fequal2', a, b), ('iand', ('feq', 'a.x', 'b.x'), ('feq', 'a.y', 'b.y')), 'options->lower_vector_cmp'),
   (('ball_fequal3', a, b), ('iand', ('iand', ('feq', 'a.x', 'b.x'), ('feq', 'a.y', 'b.y')), ('feq', 'a.z', 'b.z')), 'options->lower_vector_cmp'),
   (('ball_fequal4', a, b), ('iand', ('iand', ('feq', 'a.x', 'b.x'), ('feq', 'a.y', 'b.y')), ('iand', ('feq', 'a.z', 'b.z'), ('feq', 'a.w', 'b.w'))), 'options->lower_vector_cmp'),

   (('bany_fnequal2', a, b), ('ior', ('fneu', 'a.x', 'b.x'), ('fneu', 'a.y', 'b.y')), 'options->lower_vector_cmp'),
   (('bany_fnequal3', a, b), ('ior', ('ior', ('fneu', 'a.x', 'b.x'), ('fneu', 'a.y', 'b.y')), ('fneu', 'a.z', 'b.z')), 'options->lower_vector_cmp'),
   (('bany_fnequal4', a, b), ('ior', ('ior', ('fneu', 'a.x', 'b.x'), ('fneu', 'a.y', 'b.y')), ('ior', ('fneu', 'a.z', 'b.z'), ('fneu', 'a.w', 'b.w'))), 'options->lower_vector_cmp'),

   (('fneu', ('fneg', a), a), ('fneu', a, 0.0)),
   (('feq', ('fneg', a), a), ('feq', a, 0.0)),
   # Emulating booleans
   (('imul', ('b2i', 'a@1'), ('b2i', 'b@1')), ('b2i', ('iand', a, b))),
   (('iand', ('b2i', 'a@1'), ('b2i', 'b@1')), ('b2i', ('iand', a, b))),
   (('ior', ('b2i', 'a@1'), ('b2i', 'b@1')), ('b2i', ('ior', a, b))),
   (('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a, b))),
   (('fsat', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('b2f', ('ior', a, b))),
   (('iand', 'a@bool16', 1.0), ('b2f', a)),
   (('iand', 'a@bool32', 1.0), ('b2f', a)),
   (('flt', ('fneg', ('b2f', 'a@1')), 0), a), # Generated by TGSI KILL_IF.
   # Comparison with the same args.  Note that these are not done for
   # the float versions because NaN always returns false on float
   # inequalities.
   (('ilt', a, a), False),
   (('ige', a, a), True),
   (('ieq', a, a), True),
   (('ine', a, a), False),
   (('ult', a, a), False),
   (('uge', a, a), True),
   # Logical and bit operations
   (('iand', a, a), a),
   (('iand', a, ~0), a),
   (('iand', a, 0), 0),
   (('ior', a, a), a),
   (('ior', a, 0), a),
   (('ior', a, True), True),
   (('ixor', a, a), 0),
   (('ixor', a, 0), a),
   (('inot', ('inot', a)), a),
   (('ior', ('iand', a, b), b), b),
   (('ior', ('ior', a, b), b), ('ior', a, b)),
   (('iand', ('ior', a, b), b), b),
   (('iand', ('iand', a, b), b), ('iand', a, b)),
   # DeMorgan's Laws
   (('iand', ('inot', a), ('inot', b)), ('inot', ('ior',  a, b))),
   (('ior',  ('inot', a), ('inot', b)), ('inot', ('iand', a, b))),
   # Shift optimizations
   (('ishl', 0, a), 0),
   (('ishl', a, 0), a),
   (('ishr', 0, a), 0),
   (('ishr', a, 0), a),
   (('ushr', 0, a), 0),
   (('ushr', a, 0), a),
   (('ior', ('ishl@16', a, b), ('ushr@16', a, ('iadd', 16, ('ineg', b)))), ('urol', a, b), '!options->lower_rotate'),
   (('ior', ('ishl@16', a, b), ('ushr@16', a, ('isub', 16, b))), ('urol', a, b), '!options->lower_rotate'),
   (('ior', ('ishl@32', a, b), ('ushr@32', a, ('iadd', 32, ('ineg', b)))), ('urol', a, b), '!options->lower_rotate'),
   (('ior', ('ishl@32', a, b), ('ushr@32', a, ('isub', 32, b))), ('urol', a, b), '!options->lower_rotate'),
   (('ior', ('ushr@16', a, b), ('ishl@16', a, ('iadd', 16, ('ineg', b)))), ('uror', a, b), '!options->lower_rotate'),
   (('ior', ('ushr@16', a, b), ('ishl@16', a, ('isub', 16, b))), ('uror', a, b), '!options->lower_rotate'),
   (('ior', ('ushr@32', a, b), ('ishl@32', a, ('iadd', 32, ('ineg', b)))), ('uror', a, b), '!options->lower_rotate'),
   (('ior', ('ushr@32', a, b), ('ishl@32', a, ('isub', 32, b))), ('uror', a, b), '!options->lower_rotate'),
   (('urol@16', a, b), ('ior', ('ishl', a, b), ('ushr', a, ('isub', 16, b))), 'options->lower_rotate'),
   (('urol@32', a, b), ('ior', ('ishl', a, b), ('ushr', a, ('isub', 32, b))), 'options->lower_rotate'),
   (('uror@16', a, b), ('ior', ('ushr', a, b), ('ishl', a, ('isub', 16, b))), 'options->lower_rotate'),
   (('uror@32', a, b), ('ior', ('ushr', a, b), ('ishl', a, ('isub', 32, b))), 'options->lower_rotate'),
   # Exponential/logarithmic identities
   (('~fexp2', ('flog2', a)), a), # 2^lg2(a) = a
   (('~flog2', ('fexp2', a)), a), # lg2(2^a) = a
   (('fpow', a, b), ('fexp2', ('fmul', ('flog2', a), b)), 'options->lower_fpow'), # a^b = 2^(lg2(a)*b)
   (('~fexp2', ('fmul', ('flog2', a), b)), ('fpow', a, b), '!options->lower_fpow'), # 2^(lg2(a)*b) = a^b
   (('~fexp2', ('fadd', ('fmul', ('flog2', a), b), ('fmul', ('flog2', c), d))),
    ('~fmul', ('fpow', a, b), ('fpow', c, d)), '!options->lower_fpow'), # 2^(lg2(a) * b + lg2(c) + d) = a^b * c^d
   (('~fexp2', ('fmul', ('flog2', a), 0.5)), ('fsqrt', a)),
   (('~fexp2', ('fmul', ('flog2', a), 2.0)), ('fmul', a, a)),
   (('~fexp2', ('fmul', ('flog2', a), 4.0)), ('fmul', ('fmul', a, a), ('fmul', a, a))),
   (('~fpow', a, 1.0), a),
   (('~fpow', a, 2.0), ('fmul', a, a)),
   (('~fpow', a, 4.0), ('fmul', ('fmul', a, a), ('fmul', a, a))),
   (('~fpow', 2.0, a), ('fexp2', a)),
   (('~fpow', ('fpow', a, 2.2), 0.454545), a),
   (('~fpow', ('fabs', ('fpow', a, 2.2)), 0.454545), ('fabs', a)),
   (('~fsqrt', ('fexp2', a)), ('fexp2', ('fmul', 0.5, a))),
   (('~frcp', ('fexp2', a)), ('fexp2', ('fneg', a))),
   (('~frsq', ('fexp2', a)), ('fexp2', ('fmul', -0.5, a))),
   (('~flog2', ('fsqrt', a)), ('fmul', 0.5, ('flog2', a))),
   (('~flog2', ('frcp', a)), ('fneg', ('flog2', a))),
   (('~flog2', ('frsq', a)), ('fmul', -0.5, ('flog2', a))),
   (('~flog2', ('fpow', a, b)), ('fmul', b, ('flog2', a))),
   (('~fmul', ('fexp2(is_used_once)', a), ('fexp2(is_used_once)', b)), ('fexp2', ('fadd', a, b))),
   (('bcsel', ('flt', a, 0.0), 0.0, ('fsqrt', a)), ('fsqrt', ('fmax', a, 0.0))),
   (('~fmul', ('fsqrt', a), ('fsqrt', a)), ('fabs',a)),
   # Division and reciprocal
   (('~fdiv', 1.0, a), ('frcp', a)),
   (('fdiv', a, b), ('fmul', a, ('frcp', b)), 'options->lower_fdiv'),
   (('~frcp', ('frcp', a)), a),
   (('~frcp', ('fsqrt', a)), ('frsq', a)),
   (('fsqrt', a), ('frcp', ('frsq', a)), 'options->lower_fsqrt'),
   (('~frcp', ('frsq', a)), ('fsqrt', a), '!options->lower_fsqrt'),
   # Trig
   (('fsin', a), lowered_sincos(0.5), 'options->lower_sincos'),
   (('fcos', a), lowered_sincos(0.75), 'options->lower_sincos'),
   # Boolean simplifications
   (('i2b16(is_used_by_if)', a), ('ine16', a, 0)),
   (('i2b32(is_used_by_if)', a), ('ine32', a, 0)),
   (('i2b1(is_used_by_if)', a), ('ine', a, 0)),
   (('ieq', a, True), a),
   (('ine(is_not_used_by_if)', a, True), ('inot', a)),
   (('ine', a, False), a),
   (('ieq(is_not_used_by_if)', a, False), ('inot', 'a')),
   (('bcsel', a, True, False), a),
   (('bcsel', a, False, True), ('inot', a)),
   (('bcsel', a, 1.0, 0.0), ('b2f', a)),
   (('bcsel', a, 0.0, 1.0), ('b2f', ('inot', a))),
   (('bcsel', a, -1.0, -0.0), ('fneg', ('b2f', a))),
   (('bcsel', a, -0.0, -1.0), ('fneg', ('b2f', ('inot', a)))),
   (('bcsel', True, b, c), b),
   (('bcsel', False, b, c), c),

   (('bcsel', a, b, b), b),
   (('~fcsel', a, b, b), b),

   # D3D Boolean emulation
   (('bcsel', a, -1, 0), ('ineg', ('b2i', 'a@1'))),
   (('bcsel', a, 0, -1), ('ineg', ('b2i', ('inot', a)))),
   (('bcsel', a, 1, 0), ('b2i', 'a@1')),
   (('bcsel', a, 0, 1), ('b2i', ('inot', a))),
   (('iand', ('ineg', ('b2i', 'a@1')), ('ineg', ('b2i', 'b@1'))),
    ('ineg', ('b2i', ('iand', a, b)))),
   (('ior', ('ineg', ('b2i','a@1')), ('ineg', ('b2i', 'b@1'))),
    ('ineg', ('b2i', ('ior', a, b)))),
   (('ieq', ('ineg', ('b2i', 'a@1')), 0), ('inot', a)),
   (('ieq', ('ineg', ('b2i', 'a@1')), -1), a),
   (('ine', ('ineg', ('b2i', 'a@1')), 0), a),
   (('ine', ('ineg', ('b2i', 'a@1')), -1), ('inot', a)),
   (('iand', ('ineg', ('b2i', a)), 1.0), ('b2f', a)),
   (('iand', ('ineg', ('b2i', a)), 1),   ('b2i', a)),

   # Conversions
   (('i2b16', ('b2i', 'a@16')), a),
   (('i2b32', ('b2i', 'a@32')), a),
   (('f2i', ('ftrunc', a)), ('f2i', a)),
   (('f2u', ('ftrunc', a)), ('f2u', a)),
   (('i2b', ('ineg', a)), ('i2b', a)),
   (('i2b', ('iabs', a)), ('i2b', a)),
   (('inot', ('f2b1', a)), ('feq', a, 0.0)),

   # Conversions from 16 bits to 32 bits and back can always be removed
   (('f2fmp', ('f2f32', 'a@16')), a),
   (('i2imp', ('i2i32', 'a@16')), a),
   (('i2imp', ('u2u32', 'a@16')), a),

   (('f2imp', ('f2f32', 'a@16')), ('f2i16', a)),
   (('f2ump', ('f2f32', 'a@16')), ('f2u16', a)),
   (('i2fmp', ('i2i32', 'a@16')), ('i2f16', a)),
   (('u2fmp', ('u2u32', 'a@16')), ('u2f16', a)),

   (('f2fmp', ('b2f32', 'a@1')), ('b2f16', a)),
   (('i2imp', ('b2i32', 'a@1')), ('b2i16', a)),
   (('i2imp', ('b2i32', 'a@1')), ('b2i16', a)),

   (('f2imp', ('b2f32', 'a@1')), ('b2i16', a)),
   (('f2ump', ('b2f32', 'a@1')), ('b2i16', a)),
   (('i2fmp', ('b2i32', 'a@1')), ('b2f16', a)),
   (('u2fmp', ('b2i32', 'a@1')), ('b2f16', a)),

   # Conversions to 16 bits would be lossy so they should only be removed if
   # the instruction was generated by the precision lowering pass.
   (('f2f32', ('f2fmp', 'a@32')), a),
   (('i2i32', ('i2imp', 'a@32')), a),
   (('u2u32', ('i2imp', 'a@32')), a),

   (('i2i32', ('f2imp', 'a@32')), ('f2i32', a)),
   (('u2u32', ('f2ump', 'a@32')), ('f2u32', a)),
   (('f2f32', ('i2fmp', 'a@32')), ('i2f32', a)),
   (('f2f32', ('u2fmp', 'a@32')), ('u2f32', a)),

   (('ffloor', 'a(is_integral)'), a),
   (('fceil', 'a(is_integral)'), a),
   (('ftrunc', 'a(is_integral)'), a),
   # fract(x) = x - floor(x), so fract(NaN) = NaN
   (('~ffract', 'a(is_integral)'), 0.0),
   (('fabs', 'a(is_not_negative)'), a),
   (('iabs', 'a(is_not_negative)'), a),
   (('fsat', 'a(is_not_positive)'), 0.0),

   # Section 5.4.1 (Conversion and Scalar Constructors) of the GLSL 4.60 spec
   # says:
   #
   #    It is undefined to convert a negative floating-point value to an
   #    uint.
   #
   # Assuming that (uint)some_float behaves like (uint)(int)some_float allows
   # some optimizations in the i965 backend to proceed.
   (('ige', ('f2u', a), b), ('ige', ('f2i', a), b)),
   (('ige', b, ('f2u', a)), ('ige', b, ('f2i', a))),
   (('ilt', ('f2u', a), b), ('ilt', ('f2i', a), b)),
   (('ilt', b, ('f2u', a)), ('ilt', b, ('f2i', a))),

   (('~fmin', 'a(is_not_negative)', 1.0), ('fsat', a), '!options->lower_fsat'),

   # The result of the multiply must be in [-1, 0], so the result of the ffma
   # must be in [0, 1].
   (('flt', ('fadd', ('fmul', ('fsat', a), ('fneg', ('fsat', a))), 1.0), 0.0), False),
   (('flt', ('fadd', ('fneg', ('fmul', ('fsat', a), ('fsat', a))), 1.0), 0.0), False),
   (('fmax', ('fadd', ('fmul', ('fsat', a), ('fneg', ('fsat', a))), 1.0), 0.0), ('fadd', ('fmul', ('fsat', a), ('fneg', ('fsat', a))), 1.0)),
   (('fmax', ('fadd', ('fneg', ('fmul', ('fsat', a), ('fsat', a))), 1.0), 0.0), ('fadd', ('fneg', ('fmul', ('fsat', a), ('fsat', a))), 1.0)),

   (('fneu', 'a(is_not_zero)', 0.0), True),
   (('feq', 'a(is_not_zero)', 0.0), False),

   # In this chart, + means value > 0 and - means value < 0.
   #
   # + >= + -> unknown  0 >= + -> false    - >= + -> false
   # + >= 0 -> true     0 >= 0 -> true     - >= 0 -> false
   # + >= - -> true     0 >= - -> true     - >= - -> unknown
   #
   # Using grouping conceptually similar to a Karnaugh map...
   #
   # (+ >= 0, + >= -, 0 >= 0, 0 >= -) == (is_not_negative >= is_not_positive) -> true
   # (0 >= +, - >= +) == (is_not_positive >= gt_zero) -> false
   # (- >= +, - >= 0) == (lt_zero >= is_not_negative) -> false
   #
   # The flt / ilt cases just invert the expected result.
   #
   # The results expecting true, must be marked imprecise.  The results
   # expecting false are fine because NaN compared >= or < anything is false.

   (('~fge', 'a(is_not_negative)', 'b(is_not_positive)'), True),
   (('fge',  'a(is_not_positive)', 'b(is_gt_zero)'),      False),
   (('fge',  'a(is_lt_zero)',      'b(is_not_negative)'), False),

   (('flt',  'a(is_not_negative)', 'b(is_not_positive)'), False),
   (('~flt', 'a(is_not_positive)', 'b(is_gt_zero)'),      True),
   (('~flt', 'a(is_lt_zero)',      'b(is_not_negative)'), True),

   (('ine', 'a(is_not_zero)', 0), True),
   (('ieq', 'a(is_not_zero)', 0), False),

   (('ige', 'a(is_not_negative)', 'b(is_not_positive)'), True),
   (('ige', 'a(is_not_positive)', 'b(is_gt_zero)'),      False),
   (('ige', 'a(is_lt_zero)',      'b(is_not_negative)'), False),

   (('ilt', 'a(is_not_negative)', 'b(is_not_positive)'), False),
   (('ilt', 'a(is_not_positive)', 'b(is_gt_zero)'),      True),
   (('ilt', 'a(is_lt_zero)',      'b(is_not_negative)'), True),

   (('ult', 0, 'a(is_gt_zero)'), True),
   (('ult', a, 0), False),

   # Packing and then unpacking does nothing
   (('unpack_64_2x32_split_x', ('pack_64_2x32_split', a, b)), a),
   (('unpack_64_2x32_split_y', ('pack_64_2x32_split', a, b)), b),
   (('unpack_64_2x32', ('pack_64_2x32_split', a, b)), ('vec2', a, b)),
   (('unpack_64_2x32', ('pack_64_2x32', a)), a),
   (('pack_64_2x32_split', ('unpack_64_2x32_split_x', a),
                           ('unpack_64_2x32_split_y', a)), a),
   (('pack_64_2x32', ('vec2', ('unpack_64_2x32_split_x', a),
                              ('unpack_64_2x32_split_y', a))), a),
   (('pack_64_2x32', ('unpack_64_2x32', a)), a),

   # Comparing two halves of an unpack separately.  While this optimization
   # should be correct for non-constant values, it's less obvious that it's
   # useful in that case.  For constant values, the pack will fold and we're
   # guaranteed to reduce the whole tree to one instruction.
   (('iand', ('ieq', ('unpack_32_2x16_split_x', a), '#b'),
             ('ieq', ('unpack_32_2x16_split_y', a), '#c')),
    ('ieq', a, ('pack_32_2x16_split', b, c))),

   # Byte extraction
   (('ushr', 'a@16',  8), ('extract_u8', a, 1), '!options->lower_extract_byte'),
   (('ushr', 'a@32', 24), ('extract_u8', a, 3), '!options->lower_extract_byte'),
   (('ushr', 'a@64', 56), ('extract_u8', a, 7), '!options->lower_extract_byte'),
   (('ishr', 'a@16',  8), ('extract_i8', a, 1), '!options->lower_extract_byte'),
   (('ishr', 'a@32', 24), ('extract_i8', a, 3), '!options->lower_extract_byte'),
   (('ishr', 'a@64', 56), ('extract_i8', a, 7), '!options->lower_extract_byte'),
   (('iand', 0xff, a), ('extract_u8', a, 0), '!options->lower_extract_byte'),

   (('ubfe', a,  0, 8), ('extract_u8', a, 0), '!options->lower_extract_byte'),
   (('ubfe', a,  8, 8), ('extract_u8', a, 1), '!options->lower_extract_byte'),
   (('ubfe', a, 16, 8), ('extract_u8', a, 2), '!options->lower_extract_byte'),
   (('ubfe', a, 24, 8), ('extract_u8', a, 3), '!options->lower_extract_byte'),
   (('ibfe', a,  0, 8), ('extract_i8', a, 0), '!options->lower_extract_byte'),
   (('ibfe', a,  8, 8), ('extract_i8', a, 1), '!options->lower_extract_byte'),
   (('ibfe', a, 16, 8), ('extract_i8', a, 2), '!options->lower_extract_byte'),
   (('ibfe', a, 24, 8), ('extract_i8', a, 3), '!options->lower_extract_byte'),

    # Word extraction
   (('ushr', ('ishl', 'a@32', 16), 16), ('extract_u16', a, 0), '!options->lower_extract_word'),
   (('ushr', 'a@32', 16), ('extract_u16', a, 1), '!options->lower_extract_word'),
   (('ishr', ('ishl', 'a@32', 16), 16), ('extract_i16', a, 0), '!options->lower_extract_word'),
   (('ishr', 'a@32', 16), ('extract_i16', a, 1), '!options->lower_extract_word'),
   (('iand', 0xffff, a), ('extract_u16', a, 0), '!options->lower_extract_word'),

   (('ubfe', a,  0, 16), ('extract_u16', a, 0), '!options->lower_extract_word'),
   (('ubfe', a, 16, 16), ('extract_u16', a, 1), '!options->lower_extract_word'),
   (('ibfe', a,  0, 16), ('extract_i16', a, 0), '!options->lower_extract_word'),
   (('ibfe', a, 16, 16), ('extract_i16', a, 1), '!options->lower_extract_word'),

   # Lower pack/unpack
   (('pack_64_2x32_split', a, b), ('ior', ('u2u64', a), ('ishl', ('u2u64', b), 32)), 'options->lower_pack_64_2x32_split'),
   (('pack_32_2x16_split', a, b), ('ior', ('u2u32', a), ('ishl', ('u2u32', b), 16)), 'options->lower_pack_32_2x16_split'),
   (('unpack_64_2x32_split_x', a), ('u2u32', a), 'options->lower_unpack_64_2x32_split'),
   (('unpack_64_2x32_split_y', a), ('u2u32', ('ushr', a, 32)), 'options->lower_unpack_64_2x32_split'),
   (('unpack_32_2x16_split_x', a), ('u2u16', a), 'options->lower_unpack_32_2x16_split'),
   (('unpack_32_2x16_split_y', a), ('u2u16', ('ushr', a, 16)), 'options->lower_unpack_32_2x16_split'),

   # Useless masking before unpacking
   (('unpack_half_2x16_split_x', ('iand', a, 0xffff)), ('unpack_half_2x16_split_x', a)),
   (('unpack_32_2x16_split_x', ('iand', a, 0xffff)), ('unpack_32_2x16_split_x', a)),
   (('unpack_64_2x32_split_x', ('iand', a, 0xffffffff)), ('unpack_64_2x32_split_x', a)),
   (('unpack_half_2x16_split_y', ('iand', a, 0xffff0000)), ('unpack_half_2x16_split_y', a)),
   (('unpack_32_2x16_split_y', ('iand', a, 0xffff0000)), ('unpack_32_2x16_split_y', a)),
   (('unpack_64_2x32_split_y', ('iand', a, 0xffffffff00000000)), ('unpack_64_2x32_split_y', a)),

   (('unpack_half_2x16_split_x', ('extract_u16', a, 0)), ('unpack_half_2x16_split_x', a)),
   (('unpack_half_2x16_split_x', ('extract_u16', a, 1)), ('unpack_half_2x16_split_y', a)),
   (('unpack_32_2x16_split_x', ('extract_u16', a, 0)), ('unpack_32_2x16_split_x', a)),
   (('unpack_32_2x16_split_x', ('extract_u16', a, 1)), ('unpack_32_2x16_split_y', a)),

   # Optimize half packing
   (('ishl', ('pack_half_2x16', ('vec2', a, 0)), 16), ('pack_half_2x16', ('vec2', 0, a))),
   (('ushr', ('pack_half_2x16', ('vec2', 0, a)), 16), ('pack_half_2x16', ('vec2', a, 0))),

   (('iadd', ('pack_half_2x16', ('vec2', a, 0)), ('pack_half_2x16', ('vec2', 0, b))),
    ('pack_half_2x16', ('vec2', a, b))),
   (('ior', ('pack_half_2x16', ('vec2', a, 0)), ('pack_half_2x16', ('vec2', 0, b))),
    ('pack_half_2x16', ('vec2', a, b))),

   (('ishl', ('pack_half_2x16_split', a, 0), 16), ('pack_half_2x16_split', 0, a)),
   (('ushr', ('pack_half_2x16_split', 0, a), 16), ('pack_half_2x16_split', a, 0)),
   (('extract_u16', ('pack_half_2x16_split', 0, a), 1), ('pack_half_2x16_split', a, 0)),

   (('iadd', ('pack_half_2x16_split', a, 0), ('pack_half_2x16_split', 0, b)), ('pack_half_2x16_split', a, b)),
   (('ior',  ('pack_half_2x16_split', a, 0), ('pack_half_2x16_split', 0, b)), ('pack_half_2x16_split', a, b)),
])

# After the ('extract_u8', a, 0) pattern, above, triggers, there will be
# patterns like those below.
for op in ('ushr', 'ishr'):
   optimizations.extend([(('extract_u8', (op, 'a@16',  8),     0), ('extract_u8', a, 1))])
   optimizations.extend([(('extract_u8', (op, 'a@32',  8 * i), 0), ('extract_u8', a, i)) for i in range(1, 4)])
   optimizations.extend([(('extract_u8', (op, 'a@64',  8 * i), 0), ('extract_u8', a, i)) for i in range(1, 8)])

optimizations.extend([(('extract_u8', ('extract_u16', a, 1), 0), ('extract_u8', a, 2))])

# After the ('extract_[iu]8', a, 3) patterns, above, trigger, there will be
# patterns like those below.
for op in ('extract_u8', 'extract_i8'):
   optimizations.extend([((op, ('ishl', 'a@16',      8),     1), (op, a, 0))])
   optimizations.extend([((op, ('ishl', 'a@32', 24 - 8 * i), 3), (op, a, i)) for i in range(2, -1, -1)])
   optimizations.extend([((op, ('ishl', 'a@64', 56 - 8 * i), 7), (op, a, i)) for i in range(6, -1, -1)])

optimizations.extend([
   # Subtracts
   (('ussub_4x8', a, 0), a),
   (('ussub_4x8', a, ~0), 0),
   # Lower all Subtractions first - they can get recombined later
   (('fsub', a, b), ('fadd', a, ('fneg', b))),
   (('isub', a, b), ('iadd', a, ('ineg', b))),
   (('uabs_usub', a, b), ('bcsel', ('ult', a, b), ('ineg', ('isub', a, b)), ('isub', a, b))),
   # This is correct.  We don't need isub_sat because the result type is unsigned, so it cannot overflow.
   (('uabs_isub', a, b), ('bcsel', ('ilt', a, b), ('ineg', ('isub', a, b)), ('isub', a, b))),

   # Propagate negation up multiplication chains
   (('fmul(is_used_by_non_fsat)', ('fneg', a), b), ('fneg', ('fmul', a, b))),
   (('imul', ('ineg', a), b), ('ineg', ('imul', a, b))),

   # Propagate constants up multiplication chains
   (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fmul', ('fmul', a, c), b)),
   (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('imul', ('imul', a, c), b)),
   (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fadd', ('fadd', a, c), b)),
   (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('iadd', ('iadd', a, c), b)),

   # Reassociate constants in add/mul chains so they can be folded together.
   # For now, we mostly only handle cases where the constants are separated by
   # a single non-constant.  We could do better eventually.
   (('~fmul', '#a', ('fmul', 'b(is_not_const)', '#c')), ('fmul', ('fmul', a, c), b)),
   (('imul', '#a', ('imul', 'b(is_not_const)', '#c')), ('imul', ('imul', a, c), b)),
   (('~fadd', '#a',          ('fadd', 'b(is_not_const)', '#c')),  ('fadd', ('fadd', a,          c),           b)),
   (('~fadd', '#a', ('fneg', ('fadd', 'b(is_not_const)', '#c'))), ('fadd', ('fadd', a, ('fneg', c)), ('fneg', b))),
   (('iadd', '#a', ('iadd', 'b(is_not_const)', '#c')), ('iadd', ('iadd', a, c), b)),
   (('iand', '#a', ('iand', 'b(is_not_const)', '#c')), ('iand', ('iand', a, c), b)),
   (('ior',  '#a', ('ior',  'b(is_not_const)', '#c')), ('ior',  ('ior',  a, c), b)),
   (('ixor', '#a', ('ixor', 'b(is_not_const)', '#c')), ('ixor', ('ixor', a, c), b)),

   # Drop mul-div by the same value when there's no wrapping.
   (('idiv', ('imul(no_signed_wrap)', a, b), b), a),

   # By definition...
   (('bcsel', ('ige', ('find_lsb', a), 0), ('find_lsb', a), -1), ('find_lsb', a)),
   (('bcsel', ('ige', ('ifind_msb', a), 0), ('ifind_msb', a), -1), ('ifind_msb', a)),
   (('bcsel', ('ige', ('ufind_msb', a), 0), ('ufind_msb', a), -1), ('ufind_msb', a)),

   (('bcsel', ('ine', a, 0), ('find_lsb', a), -1), ('find_lsb', a)),
   (('bcsel', ('ine', a, 0), ('ifind_msb', a), -1), ('ifind_msb', a)),
   (('bcsel', ('ine', a, 0), ('ufind_msb', a), -1), ('ufind_msb', a)),

   (('bcsel', ('ine', a, -1), ('ifind_msb', a), -1), ('ifind_msb', a)),

   (('~fmul', ('bcsel(is_used_once)', c, -1.0, 1.0), b), ('bcsel', c, ('fneg', b), b)),
   (('~fmul', ('bcsel(is_used_once)', c, 1.0, -1.0), b), ('bcsel', c, b, ('fneg', b))),
   (('~bcsel', ('flt', a, 0.0), ('fneg', a), a), ('fabs', a)),

   (('bcsel', a, ('bcsel', b, c, d), d), ('bcsel', ('iand', a, b), c, d)),
   (('bcsel', a, b, ('bcsel', c, b, d)), ('bcsel', ('ior', a, c), b, d)),

   # Misc. lowering
   (('fmod', a, b), ('fsub', a, ('fmul', b, ('ffloor', ('fdiv', a, b)))), 'options->lower_fmod'),
   (('frem', a, b), ('fsub', a, ('fmul', b, ('ftrunc', ('fdiv', a, b)))), 'options->lower_fmod'),
   (('uadd_carry', a, b), ('b2i', ('ult', ('iadd', a, b), a)), 'options->lower_uadd_carry'),
   (('usub_borrow@32', a, b), ('b2i', ('ult', a, b)), 'options->lower_usub_borrow'),

   (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
    ('bcsel', ('ult', 31, 'bits'), 'insert',
              ('bfi', ('bfm', 'bits', 'offset'), 'insert', 'base')),
    'options->lower_bitfield_insert'),
   (('ihadd', a, b), ('iadd', ('iand', a, b), ('ishr', ('ixor', a, b), 1)), 'options->lower_hadd'),
   (('uhadd', a, b), ('iadd', ('iand', a, b), ('ushr', ('ixor', a, b), 1)), 'options->lower_hadd'),
   (('irhadd', a, b), ('isub', ('ior', a, b), ('ishr', ('ixor', a, b), 1)), 'options->lower_hadd'),
   (('urhadd', a, b), ('isub', ('ior', a, b), ('ushr', ('ixor', a, b), 1)), 'options->lower_hadd'),
   (('ihadd@64', a, b), ('iadd', ('iand', a, b), ('ishr', ('ixor', a, b), 1)), 'options->lower_hadd64 || (options->lower_int64_options & nir_lower_iadd64) != 0'),
   (('uhadd@64', a, b), ('iadd', ('iand', a, b), ('ushr', ('ixor', a, b), 1)), 'options->lower_hadd64 || (options->lower_int64_options & nir_lower_iadd64) != 0'),
   (('irhadd@64', a, b), ('isub', ('ior', a, b), ('ishr', ('ixor', a, b), 1)), 'options->lower_hadd64 || (options->lower_int64_options & nir_lower_iadd64) != 0'),
   (('urhadd@64', a, b), ('isub', ('ior', a, b), ('ushr', ('ixor', a, b), 1)), 'options->lower_hadd64 || (options->lower_int64_options & nir_lower_iadd64) != 0'),

   (('uadd_sat@64', a, b), ('bcsel', ('ult', ('iadd', a, b), a), -1, ('iadd', a, b)), 'options->lower_add_sat || (options->lower_int64_options & nir_lower_iadd64) != 0'),
   (('uadd_sat', a, b), ('bcsel', ('ult', ('iadd', a, b), a), -1, ('iadd', a, b)), 'options->lower_add_sat'),
   (('usub_sat', a, b), ('bcsel', ('ult', a, b), 0, ('isub', a, b)), 'options->lower_add_sat'),
   (('usub_sat@64', a, b), ('bcsel', ('ult', a, b), 0, ('isub', a, b)), 'options->lower_usub_sat64 || (options->lower_int64_options & nir_lower_iadd64) != 0'),

   # int64_t sum = a + b;
   #
   # if (a < 0 && b < 0 && a < sum)
   #    sum = INT64_MIN;
   # } else if (a >= 0 && b >= 0 && sum < a)
   #    sum = INT64_MAX;
   # }
   #
   # A couple optimizations are applied.
   #
   # 1. a < sum => sum >= 0.  This replacement works because it is known that
   #    a < 0 and b < 0, so sum should also be < 0 unless there was
   #    underflow.
   #
   # 2. sum < a => sum < 0.  This replacement works because it is known that
   #    a >= 0 and b >= 0, so sum should also be >= 0 unless there was
   #    overflow.
   #
   # 3. Invert the second if-condition and swap the order of parameters for
   #    the bcsel. !(a >= 0 && b >= 0 && sum < 0) becomes !(a >= 0) || !(b >=
   #    0) || !(sum < 0), and that becomes (a < 0) || (b < 0) || (sum >= 0)
   #
   # On Intel Gen11, this saves ~11 instructions.
   (('iadd_sat@64', a, b), ('bcsel',
                            ('iand', ('iand', ('ilt', a, 0), ('ilt', b, 0)), ('ige', ('iadd', a, b), 0)),
                            0x8000000000000000,
                            ('bcsel',
                             ('ior', ('ior', ('ilt', a, 0), ('ilt', b, 0)), ('ige', ('iadd', a, b), 0)),
                             ('iadd', a, b),
                             0x7fffffffffffffff)),
    '(options->lower_int64_options & nir_lower_iadd64) != 0'),

   # int64_t sum = a - b;
   #
   # if (a < 0 && b >= 0 && a < sum)
   #    sum = INT64_MIN;
   # } else if (a >= 0 && b < 0 && a >= sum)
   #    sum = INT64_MAX;
   # }
   #
   # Optimizations similar to the iadd_sat case are applied here.
   (('isub_sat@64', a, b), ('bcsel',
                            ('iand', ('iand', ('ilt', a, 0), ('ige', b, 0)), ('ige', ('isub', a, b), 0)),
                            0x8000000000000000,
                            ('bcsel',
                             ('ior', ('ior', ('ilt', a, 0), ('ige', b, 0)), ('ige', ('isub', a, b), 0)),
                             ('isub', a, b),
                             0x7fffffffffffffff)),
    '(options->lower_int64_options & nir_lower_iadd64) != 0'),

   # These are done here instead of in the backend because the int64 lowering
   # pass will make a mess of the patterns.  The first patterns are
   # conditioned on nir_lower_minmax64 because it was not clear that it was
   # always an improvement on platforms that have real int64 support.  No
   # shaders in shader-db hit this, so it was hard to say one way or the
   # other.
   (('ilt', ('imax(is_used_once)', 'a@64', 'b@64'), 0), ('ilt', ('imax', ('unpack_64_2x32_split_y', a), ('unpack_64_2x32_split_y', b)), 0), '(options->lower_int64_options & nir_lower_minmax64) != 0'),
   (('ilt', ('imin(is_used_once)', 'a@64', 'b@64'), 0), ('ilt', ('imin', ('unpack_64_2x32_split_y', a), ('unpack_64_2x32_split_y', b)), 0), '(options->lower_int64_options & nir_lower_minmax64) != 0'),
   (('ige', ('imax(is_used_once)', 'a@64', 'b@64'), 0), ('ige', ('imax', ('unpack_64_2x32_split_y', a), ('unpack_64_2x32_split_y', b)), 0), '(options->lower_int64_options & nir_lower_minmax64) != 0'),
   (('ige', ('imin(is_used_once)', 'a@64', 'b@64'), 0), ('ige', ('imin', ('unpack_64_2x32_split_y', a), ('unpack_64_2x32_split_y', b)), 0), '(options->lower_int64_options & nir_lower_minmax64) != 0'),
   (('ilt', 'a@64', 0), ('ilt', ('unpack_64_2x32_split_y', a), 0), '(options->lower_int64_options & nir_lower_icmp64) != 0'),
   (('ige', 'a@64', 0), ('ige', ('unpack_64_2x32_split_y', a), 0), '(options->lower_int64_options & nir_lower_icmp64) != 0'),

   (('ine', 'a@64', 0), ('ine', ('ior', ('unpack_64_2x32_split_x', a), ('unpack_64_2x32_split_y', a)), 0), '(options->lower_int64_options & nir_lower_icmp64) != 0'),
   (('ieq', 'a@64', 0), ('ieq', ('ior', ('unpack_64_2x32_split_x', a), ('unpack_64_2x32_split_y', a)), 0), '(options->lower_int64_options & nir_lower_icmp64) != 0'),
   # 0u < uint(a) <=> uint(a) != 0u
   (('ult', 0, 'a@64'), ('ine', ('ior', ('unpack_64_2x32_split_x', a), ('unpack_64_2x32_split_y', a)), 0), '(options->lower_int64_options & nir_lower_icmp64) != 0'),

   # Alternative lowering that doesn't rely on bfi.
   (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
    ('bcsel', ('ult', 31, 'bits'),
     'insert',
    (('ior',
     ('iand', 'base', ('inot', ('ishl', ('isub', ('ishl', 1, 'bits'), 1), 'offset'))),
     ('iand', ('ishl', 'insert', 'offset'), ('ishl', ('isub', ('ishl', 1, 'bits'), 1), 'offset'))))),
    'options->lower_bitfield_insert_to_shifts'),

   # Alternative lowering that uses bitfield_select.
   (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
    ('bcsel', ('ult', 31, 'bits'), 'insert',
              ('bitfield_select', ('bfm', 'bits', 'offset'), ('ishl', 'insert', 'offset'), 'base')),
    'options->lower_bitfield_insert_to_bitfield_select'),

   (('ibitfield_extract', 'value', 'offset', 'bits'),
    ('bcsel', ('ult', 31, 'bits'), 'value',
              ('ibfe', 'value', 'offset', 'bits')),
    'options->lower_bitfield_extract'),

   (('ubitfield_extract', 'value', 'offset', 'bits'),
    ('bcsel', ('ult', 31, 'bits'), 'value',
              ('ubfe', 'value', 'offset', 'bits')),
    'options->lower_bitfield_extract'),

   # Note that these opcodes are defined to only use the five least significant bits of 'offset' and 'bits'
   (('ubfe', 'value', 'offset', ('iand', 31, 'bits')), ('ubfe', 'value', 'offset', 'bits')),
   (('ubfe', 'value', ('iand', 31, 'offset'), 'bits'), ('ubfe', 'value', 'offset', 'bits')),
   (('ibfe', 'value', 'offset', ('iand', 31, 'bits')), ('ibfe', 'value', 'offset', 'bits')),
   (('ibfe', 'value', ('iand', 31, 'offset'), 'bits'), ('ibfe', 'value', 'offset', 'bits')),
   (('bfm', 'bits', ('iand', 31, 'offset')), ('bfm', 'bits', 'offset')),
   (('bfm', ('iand', 31, 'bits'), 'offset'), ('bfm', 'bits', 'offset')),

   # Section 8.8 (Integer Functions) of the GLSL 4.60 spec says:
   #
   #    If bits is zero, the result will be zero.
   #
   # These patterns prevent other patterns from generating invalid results
   # when count is zero.
   (('ubfe', a, b, 0), 0),
   (('ibfe', a, b, 0), 0),

   (('ubfe', a, 0, '#b'), ('iand', a, ('ushr', 0xffffffff, ('ineg', b)))),

   (('b2i32', ('i2b', ('ubfe', a, b, 1))), ('ubfe', a, b, 1)),
   (('b2i32', ('i2b', ('ibfe', a, b, 1))), ('ubfe', a, b, 1)), # ubfe in the replacement is correct
   (('ine', ('ibfe(is_used_once)', a, '#b', '#c'), 0), ('ine', ('iand', a, ('ishl', ('ushr', 0xffffffff, ('ineg', c)), b)), 0)),
   (('ieq', ('ibfe(is_used_once)', a, '#b', '#c'), 0), ('ieq', ('iand', a, ('ishl', ('ushr', 0xffffffff, ('ineg', c)), b)), 0)),
   (('ine', ('ubfe(is_used_once)', a, '#b', '#c'), 0), ('ine', ('iand', a, ('ishl', ('ushr', 0xffffffff, ('ineg', c)), b)), 0)),
   (('ieq', ('ubfe(is_used_once)', a, '#b', '#c'), 0), ('ieq', ('iand', a, ('ishl', ('ushr', 0xffffffff, ('ineg', c)), b)), 0)),

   (('ibitfield_extract', 'value', 'offset', 'bits'),
    ('bcsel', ('ieq', 0, 'bits'),
     0,
     ('ishr',
       ('ishl', 'value', ('isub', ('isub', 32, 'bits'), 'offset')),
       ('isub', 32, 'bits'))),
    'options->lower_bitfield_extract_to_shifts'),

   (('ubitfield_extract', 'value', 'offset', 'bits'),
    ('iand',
     ('ushr', 'value', 'offset'),
     ('bcsel', ('ieq', 'bits', 32),
      0xffffffff,
      ('isub', ('ishl', 1, 'bits'), 1))),
    'options->lower_bitfield_extract_to_shifts'),

   (('ifind_msb', 'value'),
    ('ufind_msb', ('bcsel', ('ilt', 'value', 0), ('inot', 'value'), 'value')),
    'options->lower_ifind_msb'),

   (('find_lsb', 'value'),
    ('ufind_msb', ('iand', 'value', ('ineg', 'value'))),
    'options->lower_find_lsb'),

   (('extract_i8', a, 'b@32'),
    ('ishr', ('ishl', a, ('imul', ('isub', 3, b), 8)), 24),
    'options->lower_extract_byte'),

   (('extract_u8', a, 'b@32'),
    ('iand', ('ushr', a, ('imul', b, 8)), 0xff),
    'options->lower_extract_byte'),

   (('extract_i16', a, 'b@32'),
    ('ishr', ('ishl', a, ('imul', ('isub', 1, b), 16)), 16),
    'options->lower_extract_word'),

   (('extract_u16', a, 'b@32'),
    ('iand', ('ushr', a, ('imul', b, 16)), 0xffff),
    'options->lower_extract_word'),

    (('pack_unorm_2x16', 'v'),
     ('pack_uvec2_to_uint',
        ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 65535.0)))),
     'options->lower_pack_unorm_2x16'),

    (('pack_unorm_4x8', 'v'),
     ('pack_uvec4_to_uint',
        ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 255.0)))),
     'options->lower_pack_unorm_4x8'),

    (('pack_snorm_2x16', 'v'),
     ('pack_uvec2_to_uint',
        ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 32767.0)))),
     'options->lower_pack_snorm_2x16'),

    (('pack_snorm_4x8', 'v'),
     ('pack_uvec4_to_uint',
        ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 127.0)))),
     'options->lower_pack_snorm_4x8'),

    (('unpack_unorm_2x16', 'v'),
     ('fdiv', ('u2f32', ('vec2', ('extract_u16', 'v', 0),
                                  ('extract_u16', 'v', 1))),
              65535.0),
     'options->lower_unpack_unorm_2x16'),

    (('unpack_unorm_4x8', 'v'),
     ('fdiv', ('u2f32', ('vec4', ('extract_u8', 'v', 0),
                                  ('extract_u8', 'v', 1),
                                  ('extract_u8', 'v', 2),
                                  ('extract_u8', 'v', 3))),
              255.0),
     'options->lower_unpack_unorm_4x8'),

    (('unpack_snorm_2x16', 'v'),
     ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec2', ('extract_i16', 'v', 0),
                                                            ('extract_i16', 'v', 1))),
                                           32767.0))),
     'options->lower_unpack_snorm_2x16'),

    (('unpack_snorm_4x8', 'v'),
     ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec4', ('extract_i8', 'v', 0),
                                                            ('extract_i8', 'v', 1),
                                                            ('extract_i8', 'v', 2),
                                                            ('extract_i8', 'v', 3))),
                                           127.0))),
     'options->lower_unpack_snorm_4x8'),

   (('pack_half_2x16_split', 'a@32', 'b@32'),
    ('ior', ('ishl', ('u2u32', ('f2f16', b)), 16), ('u2u32', ('f2f16', a))),
    'options->lower_pack_split'),

   (('unpack_half_2x16_split_x', 'a@32'),
    ('f2f32', ('u2u16', a)),
    'options->lower_pack_split'),

   (('unpack_half_2x16_split_y', 'a@32'),
    ('f2f32', ('u2u16', ('ushr', a, 16))),
    'options->lower_pack_split'),

   (('pack_32_2x16_split', 'a@16', 'b@16'),
    ('ior', ('ishl', ('u2u32', b), 16), ('u2u32', a)),
    'options->lower_pack_split'),

   (('unpack_32_2x16_split_x', 'a@32'),
    ('u2u16', a),
    'options->lower_pack_split'),

   (('unpack_32_2x16_split_y', 'a@32'),
    ('u2u16', ('ushr', 'a', 16)),
    'options->lower_pack_split'),

   (('isign', a), ('imin', ('imax', a, -1), 1), 'options->lower_isign'),
   (('imin', ('imax', a, -1), 1), ('isign', a), '!options->lower_isign'),
   (('imax', ('imin', a, 1), -1), ('isign', a), '!options->lower_isign'),
   (('fsign', a), ('fsub', ('b2f', ('flt', 0.0, a)), ('b2f', ('flt', a, 0.0))), 'options->lower_fsign'),

   # Address/offset calculations:
   # Drivers supporting imul24 should use the nir_lower_amul() pass, this
   # rule converts everyone else to imul:
   (('amul', a, b), ('imul', a, b), '!options->has_imul24'),

   (('umul24', a, b),
    ('imul', ('iand', a, 0xffffff), ('iand', b, 0xffffff)),
    '!options->has_umul24'),
   (('umad24', a, b, c),
    ('iadd', ('imul', ('iand', a, 0xffffff), ('iand', b, 0xffffff)), c),
    '!options->has_umad24'),

   (('imad24_ir3', a, b, 0), ('imul24', a, b)),
   (('imad24_ir3', a, 0, c), (c)),
   (('imad24_ir3', a, 1, c), ('iadd', a, c)),

   # if first two srcs are const, crack apart the imad so constant folding
   # can clean up the imul:
   # TODO ffma should probably get a similar rule:
   (('imad24_ir3', '#a', '#b', c), ('iadd', ('imul', a, b), c)),

   # These will turn 24b address/offset calc back into 32b shifts, but
   # it should be safe to get back some of the bits of precision that we
   # already decided were no necessary:
   (('imul24', a, '#b@32(is_pos_power_of_two)'), ('ishl', a, ('find_lsb', b)), '!options->lower_bitops'),
   (('imul24', a, '#b@32(is_neg_power_of_two)'), ('ineg', ('ishl', a, ('find_lsb', ('iabs', b)))), '!options->lower_bitops'),
   (('imul24', a, 0), (0)),
])

# bit_size dependent lowerings
for bit_size in [8, 16, 32, 64]:
   # convenience constants
   intmax = (1 << (bit_size - 1)) - 1
   intmin = 1 << (bit_size - 1)

   optimizations += [
      (('iadd_sat@' + str(bit_size), a, b),
       ('bcsel', ('ige', b, 1), ('bcsel', ('ilt', ('iadd', a, b), a), intmax, ('iadd', a, b)),
                                ('bcsel', ('ilt', a, ('iadd', a, b)), intmin, ('iadd', a, b))), 'options->lower_add_sat'),
      (('isub_sat@' + str(bit_size), a, b),
       ('bcsel', ('ilt', b, 0), ('bcsel', ('ilt', ('isub', a, b), a), intmax, ('isub', a, b)),
                                ('bcsel', ('ilt', a, ('isub', a, b)), intmin, ('isub', a, b))), 'options->lower_add_sat'),
   ]

invert = OrderedDict([('feq', 'fneu'), ('fneu', 'feq')])

for left, right in itertools.combinations_with_replacement(invert.keys(), 2):
   optimizations.append((('inot', ('ior(is_used_once)', (left, a, b), (right, c, d))),
                         ('iand', (invert[left], a, b), (invert[right], c, d))))
   optimizations.append((('inot', ('iand(is_used_once)', (left, a, b), (right, c, d))),
                         ('ior', (invert[left], a, b), (invert[right], c, d))))

# Optimize x2bN(b2x(x)) -> x
for size in type_sizes('bool'):
    aN = 'a@' + str(size)
    f2bN = 'f2b' + str(size)
    i2bN = 'i2b' + str(size)
    optimizations.append(((f2bN, ('b2f', aN)), a))
    optimizations.append(((i2bN, ('b2i', aN)), a))

# Optimize x2yN(b2x(x)) -> b2y
for x, y in itertools.product(['f', 'u', 'i'], ['f', 'u', 'i']):
   if x != 'f' and y != 'f' and x != y:
      continue

   b2x = 'b2f' if x == 'f' else 'b2i'
   b2y = 'b2f' if y == 'f' else 'b2i'
   x2yN = '{}2{}'.format(x, y)
   optimizations.append(((x2yN, (b2x, a)), (b2y, a)))

# Optimize away x2xN(a@N)
for t in ['int', 'uint', 'float', 'bool']:
   for N in type_sizes(t):
      x2xN = '{0}2{0}{1}'.format(t[0], N)
      aN = 'a@{0}'.format(N)
      optimizations.append(((x2xN, aN), a))

# Optimize x2xN(y2yM(a@P)) -> y2yN(a) for integers
# In particular, we can optimize away everything except upcast of downcast and
# upcasts where the type differs from the other cast
for N, M in itertools.product(type_sizes('uint'), type_sizes('uint')):
   if N < M:
      # The outer cast is a down-cast.  It doesn't matter what the size of the
      # argument of the inner cast is because we'll never been in the upcast
      # of downcast case.  Regardless of types, we'll always end up with y2yN
      # in the end.
      for x, y in itertools.product(['i', 'u'], ['i', 'u']):
         x2xN = '{0}2{0}{1}'.format(x, N)
         y2yM = '{0}2{0}{1}'.format(y, M)
         y2yN = '{0}2{0}{1}'.format(y, N)
         optimizations.append(((x2xN, (y2yM, a)), (y2yN, a)))
   elif N > M:
      # If the outer cast is an up-cast, we have to be more careful about the
      # size of the argument of the inner cast and with types.  In this case,
      # the type is always the type of type up-cast which is given by the
      # outer cast.
      for P in type_sizes('uint'):
         # We can't optimize away up-cast of down-cast.
         if M < P:
            continue

         # Because we're doing down-cast of down-cast, the types always have
         # to match between the two casts
         for x in ['i', 'u']:
            x2xN = '{0}2{0}{1}'.format(x, N)
            x2xM = '{0}2{0}{1}'.format(x, M)
            aP = 'a@{0}'.format(P)
            optimizations.append(((x2xN, (x2xM, aP)), (x2xN, a)))
   else:
      # The N == M case is handled by other optimizations
      pass

# Downcast operations should be able to see through pack
for t in ['i', 'u']:
    for N in [8, 16, 32]:
        x2xN = '{0}2{0}{1}'.format(t, N)
        optimizations += [
            ((x2xN, ('pack_64_2x32_split', a, b)), (x2xN, a)),
            ((x2xN, ('pack_64_2x32_split', a, b)), (x2xN, a)),
        ]

# Optimize comparisons with up-casts
for t in ['int', 'uint', 'float']:
    for N, M in itertools.product(type_sizes(t), repeat=2):
        if N == 1 or N >= M:
            continue

        cond = 'true'
        if N == 8:
            cond = 'options->support_8bit_alu'
        elif N == 16:
            cond = 'options->support_16bit_alu'
        x2xM = '{0}2{0}{1}'.format(t[0], M)
        x2xN = '{0}2{0}{1}'.format(t[0], N)
        aN = 'a@' + str(N)
        bN = 'b@' + str(N)
        xeq = 'feq' if t == 'float' else 'ieq'
        xne = 'fneu' if t == 'float' else 'ine'
        xge = '{0}ge'.format(t[0])
        xlt = '{0}lt'.format(t[0])

        # Up-casts are lossless so for correctly signed comparisons of
        # up-casted values we can do the comparison at the largest of the two
        # original sizes and drop one or both of the casts.  (We have
        # optimizations to drop the no-op casts which this may generate.)
        for P in type_sizes(t):
            if P == 1 or P > N:
                continue

            bP = 'b@' + str(P)
            optimizations += [
                ((xeq, (x2xM, aN), (x2xM, bP)), (xeq, a, (x2xN, b)), cond),
                ((xne, (x2xM, aN), (x2xM, bP)), (xne, a, (x2xN, b)), cond),
                ((xge, (x2xM, aN), (x2xM, bP)), (xge, a, (x2xN, b)), cond),
                ((xlt, (x2xM, aN), (x2xM, bP)), (xlt, a, (x2xN, b)), cond),
                ((xge, (x2xM, bP), (x2xM, aN)), (xge, (x2xN, b), a), cond),
                ((xlt, (x2xM, bP), (x2xM, aN)), (xlt, (x2xN, b), a), cond),
            ]

        # The next bit doesn't work on floats because the range checks would
        # get way too complicated.
        if t in ['int', 'uint']:
            if t == 'int':
                xN_min = -(1 << (N - 1))
                xN_max = (1 << (N - 1)) - 1
            elif t == 'uint':
                xN_min = 0
                xN_max = (1 << N) - 1
            else:
                assert False

            # If we're up-casting and comparing to a constant, we can unfold
            # the comparison into a comparison with the shrunk down constant
            # and a check that the constant fits in the smaller bit size.
            optimizations += [
                ((xeq, (x2xM, aN), '#b'),
                 ('iand', (xeq, a, (x2xN, b)), (xeq, (x2xM, (x2xN, b)), b)), cond),
                ((xne, (x2xM, aN), '#b'),
                 ('ior', (xne, a, (x2xN, b)), (xne, (x2xM, (x2xN, b)), b)), cond),
                ((xlt, (x2xM, aN), '#b'),
                 ('iand', (xlt, xN_min, b),
                          ('ior', (xlt, xN_max, b), (xlt, a, (x2xN, b)))), cond),
                ((xlt, '#a', (x2xM, bN)),
                 ('iand', (xlt, a, xN_max),
                          ('ior', (xlt, a, xN_min), (xlt, (x2xN, a), b))), cond),
                ((xge, (x2xM, aN), '#b'),
                 ('iand', (xge, xN_max, b),
                          ('ior', (xge, xN_min, b), (xge, a, (x2xN, b)))), cond),
                ((xge, '#a', (x2xM, bN)),
                 ('iand', (xge, a, xN_min),
                          ('ior', (xge, a, xN_max), (xge, (x2xN, a), b))), cond),
            ]

# Convert masking followed by signed downcast to just unsigned downcast
optimizations += [
    (('i2i32', ('iand', 'a@64', 0xffffffff)), ('u2u32', a)),
    (('i2i16', ('iand', 'a@32', 0xffff)), ('u2u16', a)),
    (('i2i16', ('iand', 'a@64', 0xffff)), ('u2u16', a)),
    (('i2i8', ('iand', 'a@16', 0xff)), ('u2u8', a)),
    (('i2i8', ('iand', 'a@32', 0xff)), ('u2u8', a)),
    (('i2i8', ('iand', 'a@64', 0xff)), ('u2u8', a)),
]

def fexp2i(exp, bits):
   # Generate an expression which constructs value 2.0^exp or 0.0.
   #
   # We assume that exp is already in a valid range:
   #
   #   * [-15, 15] for 16-bit float
   #   * [-127, 127] for 32-bit float
   #   * [-1023, 1023] for 16-bit float
   #
   # If exp is the lowest value in the valid range, a value of 0.0 is
   # constructed.  Otherwise, the value 2.0^exp is constructed.
   if bits == 16:
      return ('i2i16', ('ishl', ('iadd', exp, 15), 10))
   elif bits == 32:
      return ('ishl', ('iadd', exp, 127), 23)
   elif bits == 64:
      return ('pack_64_2x32_split', 0, ('ishl', ('iadd', exp, 1023), 20))
   else:
      assert False

def ldexp(f, exp, bits):
   # The maximum possible range for a normal exponent is [-126, 127] and,
   # throwing in denormals, you get a maximum range of [-149, 127].  This
   # means that we can potentially have a swing of +-276.  If you start with
   # FLT_MAX, you actually have to do ldexp(FLT_MAX, -278) to get it to flush
   # all the way to zero.  The GLSL spec only requires that we handle a subset
   # of this range.  From version 4.60 of the spec:
   #
   #    "If exp is greater than +128 (single-precision) or +1024
   #    (double-precision), the value returned is undefined. If exp is less
   #    than -126 (single-precision) or -1022 (double-precision), the value
   #    returned may be flushed to zero. Additionally, splitting the value
   #    into a significand and exponent using frexp() and then reconstructing
   #    a floating-point value using ldexp() should yield the original input
   #    for zero and all finite non-denormalized values."
   #
   # The SPIR-V spec has similar language.
   #
   # In order to handle the maximum value +128 using the fexp2i() helper
   # above, we have to split the exponent in half and do two multiply
   # operations.
   #
   # First, we clamp exp to a reasonable range.  Specifically, we clamp to
   # twice the full range that is valid for the fexp2i() function above.  If
   # exp/2 is the bottom value of that range, the fexp2i() expression will
   # yield 0.0f which, when multiplied by f, will flush it to zero which is
   # allowed by the GLSL and SPIR-V specs for low exponent values.  If the
   # value is clamped from above, then it must have been above the supported
   # range of the GLSL built-in and therefore any return value is acceptable.
   if bits == 16:
      exp = ('imin', ('imax', exp, -30), 30)
   elif bits == 32:
      exp = ('imin', ('imax', exp, -254), 254)
   elif bits == 64:
      exp = ('imin', ('imax', exp, -2046), 2046)
   else:
      assert False

   # Now we compute two powers of 2, one for exp/2 and one for exp-exp/2.
   # (We use ishr which isn't the same for -1, but the -1 case still works
   # since we use exp-exp/2 as the second exponent.)  While the spec
   # technically defines ldexp as f * 2.0^exp, simply multiplying once doesn't
   # work with denormals and doesn't allow for the full swing in exponents
   # that you can get with normalized values.  Instead, we create two powers
   # of two and multiply by them each in turn.  That way the effective range
   # of our exponent is doubled.
   pow2_1 = fexp2i(('ishr', exp, 1), bits)
   pow2_2 = fexp2i(('isub', exp, ('ishr', exp, 1)), bits)
   return ('fmul', ('fmul', f, pow2_1), pow2_2)

optimizations += [
   (('ldexp@16', 'x', 'exp'), ldexp('x', 'exp', 16), 'options->lower_ldexp'),
   (('ldexp@32', 'x', 'exp'), ldexp('x', 'exp', 32), 'options->lower_ldexp'),
   (('ldexp@64', 'x', 'exp'), ldexp('x', 'exp', 64), 'options->lower_ldexp'),
]

# Unreal Engine 4 demo applications open-codes bitfieldReverse()
def bitfield_reverse(u):
    step1 = ('ior', ('ishl', u, 16), ('ushr', u, 16))
    step2 = ('ior', ('ishl', ('iand', step1, 0x00ff00ff), 8), ('ushr', ('iand', step1, 0xff00ff00), 8))
    step3 = ('ior', ('ishl', ('iand', step2, 0x0f0f0f0f), 4), ('ushr', ('iand', step2, 0xf0f0f0f0), 4))
    step4 = ('ior', ('ishl', ('iand', step3, 0x33333333), 2), ('ushr', ('iand', step3, 0xcccccccc), 2))
    step5 = ('ior(many-comm-expr)', ('ishl', ('iand', step4, 0x55555555), 1), ('ushr', ('iand', step4, 0xaaaaaaaa), 1))

    return step5

optimizations += [(bitfield_reverse('x@32'), ('bitfield_reverse', 'x'), '!options->lower_bitfield_reverse')]

# "all_equal(eq(a, b), vec(~0))" is the same as "all_equal(a, b)"
# "any_nequal(neq(a, b), vec(0))" is the same as "any_nequal(a, b)"
for ncomp in [2, 3, 4, 8, 16]:
   optimizations += [
      (('ball_iequal' + str(ncomp), ('ieq', a, b), ~0), ('ball_iequal' + str(ncomp), a, b)),
      (('ball_iequal' + str(ncomp), ('feq', a, b), ~0), ('ball_fequal' + str(ncomp), a, b)),
      (('bany_inequal' + str(ncomp), ('ine', a, b), 0), ('bany_inequal' + str(ncomp), a, b)),
      (('bany_inequal' + str(ncomp), ('fneu', a, b), 0), ('bany_fnequal' + str(ncomp), a, b)),
   ]

# For any float comparison operation, "cmp", if you have "a == a && a cmp b"
# then the "a == a" is redundant because it's equivalent to "a is not NaN"
# and, if a is a NaN then the second comparison will fail anyway.
for op in ['flt', 'fge', 'feq']:
   optimizations += [
      (('iand', ('feq', a, a), (op, a, b)), ('!' + op, a, b)),
      (('iand', ('feq', a, a), (op, b, a)), ('!' + op, b, a)),
   ]

# Add optimizations to handle the case where the result of a ternary is
# compared to a constant.  This way we can take things like
#
# (a ? 0 : 1) > 0
#
# and turn it into
#
# a ? (0 > 0) : (1 > 0)
#
# which constant folding will eat for lunch.  The resulting ternary will
# further get cleaned up by the boolean reductions above and we will be
# left with just the original variable "a".
for op in ['flt', 'fge', 'feq', 'fneu',
           'ilt', 'ige', 'ieq', 'ine', 'ult', 'uge']:
   optimizations += [
      ((op, ('bcsel', 'a', '#b', '#c'), '#d'),
       ('bcsel', 'a', (op, 'b', 'd'), (op, 'c', 'd'))),
      ((op, '#d', ('bcsel', a, '#b', '#c')),
       ('bcsel', 'a', (op, 'd', 'b'), (op, 'd', 'c'))),
   ]


# For example, this converts things like
#
#    1 + mix(0, a - 1, condition)
#
# into
#
#    mix(1, (a-1)+1, condition)
#
# Other optimizations will rearrange the constants.
for op in ['fadd', 'fmul', 'iadd', 'imul']:
   optimizations += [
      ((op, ('bcsel(is_used_once)', a, '#b', c), '#d'), ('bcsel', a, (op, b, d), (op, c, d)))
   ]

# For derivatives in compute shaders, GLSL_NV_compute_shader_derivatives
# states:
#
#     If neither layout qualifier is specified, derivatives in compute shaders
#     return zero, which is consistent with the handling of built-in texture
#     functions like texture() in GLSL 4.50 compute shaders.
for op in ['fddx', 'fddx_fine', 'fddx_coarse',
           'fddy', 'fddy_fine', 'fddy_coarse']:
   optimizations += [
      ((op, 'a'), 0.0, 'info->stage == MESA_SHADER_COMPUTE && info->cs.derivative_group == DERIVATIVE_GROUP_NONE')
]

# Some optimizations for ir3-specific instructions.
optimizations += [
   # 'al * bl': If either 'al' or 'bl' is zero, return zero.
   (('umul_low', '#a(is_lower_half_zero)', 'b'), (0)),
   # '(ah * bl) << 16 + c': If either 'ah' or 'bl' is zero, return 'c'.
   (('imadsh_mix16', '#a@32(is_lower_half_zero)', 'b@32', 'c@32'), ('c')),
   (('imadsh_mix16', 'a@32', '#b@32(is_upper_half_zero)', 'c@32'), ('c')),
]

# These kinds of sequences can occur after nir_opt_peephole_select.
#
# NOTE: fadd is not handled here because that gets in the way of ffma
# generation in the i965 driver.  Instead, fadd and ffma are handled in
# late_optimizations.

for op in ['flrp']:
    optimizations += [
        (('bcsel', a, (op + '(is_used_once)', b, c, d), (op, b, c, e)), (op, b, c, ('bcsel', a, d, e))),
        (('bcsel', a, (op, b, c, d), (op + '(is_used_once)', b, c, e)), (op, b, c, ('bcsel', a, d, e))),
        (('bcsel', a, (op + '(is_used_once)', b, c, d), (op, b, e, d)), (op, b, ('bcsel', a, c, e), d)),
        (('bcsel', a, (op, b, c, d), (op + '(is_used_once)', b, e, d)), (op, b, ('bcsel', a, c, e), d)),
        (('bcsel', a, (op + '(is_used_once)', b, c, d), (op, e, c, d)), (op, ('bcsel', a, b, e), c, d)),
        (('bcsel', a, (op, b, c, d), (op + '(is_used_once)', e, c, d)), (op, ('bcsel', a, b, e), c, d)),
    ]

for op in ['fmul', 'iadd', 'imul', 'iand', 'ior', 'ixor', 'fmin', 'fmax', 'imin', 'imax', 'umin', 'umax']:
    optimizations += [
        (('bcsel', a, (op + '(is_used_once)', b, c), (op, b, 'd(is_not_const)')), (op, b, ('bcsel', a, c, d))),
        (('bcsel', a, (op + '(is_used_once)', b, 'c(is_not_const)'), (op, b, d)), (op, b, ('bcsel', a, c, d))),
        (('bcsel', a, (op, b, 'c(is_not_const)'), (op + '(is_used_once)', b, d)), (op, b, ('bcsel', a, c, d))),
        (('bcsel', a, (op, b, c), (op + '(is_used_once)', b, 'd(is_not_const)')), (op, b, ('bcsel', a, c, d))),
    ]

for op in ['fpow']:
    optimizations += [
        (('bcsel', a, (op + '(is_used_once)', b, c), (op, b, d)), (op, b, ('bcsel', a, c, d))),
        (('bcsel', a, (op, b, c), (op + '(is_used_once)', b, d)), (op, b, ('bcsel', a, c, d))),
        (('bcsel', a, (op + '(is_used_once)', b, c), (op, d, c)), (op, ('bcsel', a, b, d), c)),
        (('bcsel', a, (op, b, c), (op + '(is_used_once)', d, c)), (op, ('bcsel', a, b, d), c)),
    ]

for op in ['frcp', 'frsq', 'fsqrt', 'fexp2', 'flog2', 'fsign', 'fsin', 'fcos', 'fneg', 'fabs', 'fsign']:
    optimizations += [
        (('bcsel', c, (op + '(is_used_once)', a), (op + '(is_used_once)', b)), (op, ('bcsel', c, a, b))),
    ]

for op in ['ineg', 'iabs', 'inot', 'isign']:
    optimizations += [
        ((op, ('bcsel', c, '#a', '#b')), ('bcsel', c, (op, a), (op, b))),
    ]

# This section contains optimizations to propagate downsizing conversions of
# constructed vectors into vectors of downsized components. Whether this is
# useful depends on the SIMD semantics of the backend. On a true SIMD machine,
# this reduces the register pressure of the vector itself and often enables the
# conversions to be eliminated via other algebraic rules or constant folding.
# In the worst case on a SIMD architecture, the propagated conversions may be
# revectorized via nir_opt_vectorize so instruction count is minimally
# impacted.
#
# On a machine with SIMD-within-a-register only, this actually
# counterintuitively hurts instruction count. These machines are the same that
# require vectorize_vec2_16bit, so we predicate the optimizations on that flag
# not being set.
#
# Finally for scalar architectures, there should be no difference in generated
# code since it all ends up scalarized at the end, but it might minimally help
# compile-times.

for i in range(2, 4 + 1):
   for T in ('f', 'u', 'i'):
      vec_inst = ('vec' + str(i),)

      indices = ['a', 'b', 'c', 'd']
      suffix_in = tuple((indices[j] + '@32') for j in range(i))

      to_16 = '{}2{}16'.format(T, T)
      to_mp = '{}2{}mp'.format(T, T)

      out_16 = tuple((to_16, indices[j]) for j in range(i))
      out_mp = tuple((to_mp, indices[j]) for j in range(i))

      optimizations  += [
         ((to_16, vec_inst + suffix_in), vec_inst + out_16, '!options->vectorize_vec2_16bit'),
      ]
      # u2ump doesn't exist, because it's equal to i2imp
      if T in ['f', 'i']:
          optimizations  += [
             ((to_mp, vec_inst + suffix_in), vec_inst + out_mp, '!options->vectorize_vec2_16bit')
          ]

# This section contains "late" optimizations that should be run before
# creating ffmas and calling regular optimizations for the final time.
# Optimizations should go here if they help code generation and conflict
# with the regular optimizations.
before_ffma_optimizations = [
   # Propagate constants down multiplication chains
   (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fmul', ('fmul', a, c), b)),
   (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('imul', ('imul', a, c), b)),
   (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fadd', ('fadd', a, c), b)),
   (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('iadd', ('iadd', a, c), b)),

   (('~fadd', ('fmul', a, b), ('fmul', a, c)), ('fmul', a, ('fadd', b, c))),
   (('iadd', ('imul', a, b), ('imul', a, c)), ('imul', a, ('iadd', b, c))),
   (('~fadd', ('fneg', a), a), 0.0),
   (('iadd', ('ineg', a), a), 0),
   (('iadd', ('ineg', a), ('iadd', a, b)), b),
   (('iadd', a, ('iadd', ('ineg', a), b)), b),
   (('~fadd', ('fneg', a), ('fadd', a, b)), b),
   (('~fadd', a, ('fadd', ('fneg', a), b)), b),

   (('~flrp', ('fadd(is_used_once)', a, -1.0), ('fadd(is_used_once)', a,  1.0), d), ('fadd', ('flrp', -1.0,  1.0, d), a)),
   (('~flrp', ('fadd(is_used_once)', a,  1.0), ('fadd(is_used_once)', a, -1.0), d), ('fadd', ('flrp',  1.0, -1.0, d), a)),
   (('~flrp', ('fadd(is_used_once)', a, '#b'), ('fadd(is_used_once)', a, '#c'), d), ('fadd', ('fmul', d, ('fadd', c, ('fneg', b))), ('fadd', a, b))),
]

# This section contains "late" optimizations that should be run after the
# regular optimizations have finished.  Optimizations should go here if
# they help code generation but do not necessarily produce code that is
# more easily optimizable.
late_optimizations = [
   # Most of these optimizations aren't quite safe when you get infinity or
   # Nan involved but the first one should be fine.
   (('flt',          ('fadd', a, b),  0.0), ('flt',          a, ('fneg', b))),
   (('flt', ('fneg', ('fadd', a, b)), 0.0), ('flt', ('fneg', a),         b)),
   (('~fge',          ('fadd', a, b),  0.0), ('fge',          a, ('fneg', b))),
   (('~fge', ('fneg', ('fadd', a, b)), 0.0), ('fge', ('fneg', a),         b)),
   (('~feq', ('fadd', a, b), 0.0), ('feq', a, ('fneg', b))),
   (('~fneu', ('fadd', a, b), 0.0), ('fneu', a, ('fneg', b))),

   # nir_lower_to_source_mods will collapse this, but its existence during the
   # optimization loop can prevent other optimizations.
   (('fneg', ('fneg', a)), a),

   # Subtractions get lowered during optimization, so we need to recombine them
   (('fadd', 'a', ('fneg', 'b')), ('fsub', 'a', 'b'), '!options->lower_sub'),
   (('iadd', 'a', ('ineg', 'b')), ('isub', 'a', 'b'), '!options->lower_sub'),
   (('fneg', a), ('fsub', 0.0, a), 'options->lower_negate'),
   (('ineg', a), ('isub', 0, a), 'options->lower_negate'),
   (('iabs', a), ('imax', a, ('ineg', a)), 'options->lower_iabs'),
   (('~fadd@16', ('fmul', a, b), c), ('ffma', a, b, c), 'options->fuse_ffma16'),
   (('~fadd@32', ('fmul', a, b), c), ('ffma', a, b, c), 'options->fuse_ffma32'),
   (('~fadd@64', ('fmul', a, b), c), ('ffma', a, b, c), 'options->fuse_ffma64'),

   # These are duplicated from the main optimizations table.  The late
   # patterns that rearrange expressions like x - .5 < 0 to x < .5 can create
   # new patterns like these.  The patterns that compare with zero are removed
   # because they are unlikely to be created in by anything in
   # late_optimizations.
   (('flt', ('fsat(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('flt', a, b)),
   (('flt', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a)), ('flt', b, a)),
   (('fge', ('fsat(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('fge', a, b)),
   (('fge', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a)), ('fge', b, a)),
   (('feq', ('fsat(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('feq', a, b)),
   (('fneu', ('fsat(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('fneu', a, b)),

   (('fge', ('fsat(is_used_once)', a), 1.0), ('fge', a, 1.0)),
   # flt(fsat(a), 1.0) is inexact because it returns True if a is NaN
   # (fsat(NaN) is 0), while flt(a, 1.0) always returns FALSE.
   (('~flt', ('fsat(is_used_once)', a), 1.0), ('flt', a, 1.0)),

   (('~fge', ('fmin(is_used_once)', ('fadd(is_used_once)', a, b), ('fadd', c, d)), 0.0), ('iand', ('fge', a, ('fneg', b)), ('fge', c, ('fneg', d)))),

   (('flt', ('fneg', a), ('fneg', b)), ('flt', b, a)),
   (('fge', ('fneg', a), ('fneg', b)), ('fge', b, a)),
   (('feq', ('fneg', a), ('fneg', b)), ('feq', b, a)),
   (('fneu', ('fneg', a), ('fneg', b)), ('fneu', b, a)),
   (('flt', ('fneg', a), -1.0), ('flt', 1.0, a)),
   (('flt', -1.0, ('fneg', a)), ('flt', a, 1.0)),
   (('fge', ('fneg', a), -1.0), ('fge', 1.0, a)),
   (('fge', -1.0, ('fneg', a)), ('fge', a, 1.0)),
   (('fneu', ('fneg', a), -1.0), ('fneu', 1.0, a)),
   (('feq', -1.0, ('fneg', a)), ('feq', a, 1.0)),

   (('ior', a, a), a),
   (('iand', a, a), a),

   (('~fadd', ('fneg(is_used_once)', ('fsat(is_used_once)', 'a(is_not_fmul)')), 1.0), ('fsat', ('fadd', 1.0, ('fneg', a)))),

   (('fdot2', a, b), ('fdot_replicated2', a, b), 'options->fdot_replicates'),
   (('fdot3', a, b), ('fdot_replicated3', a, b), 'options->fdot_replicates'),
   (('fdot4', a, b), ('fdot_replicated4', a, b), 'options->fdot_replicates'),
   (('fdph', a, b), ('fdph_replicated', a, b), 'options->fdot_replicates'),

   (('~flrp', ('fadd(is_used_once)', a, b), ('fadd(is_used_once)', a, c), d), ('fadd', ('flrp', b, c, d), a)),

   # A similar operation could apply to any ffma(#a, b, #(-a/2)), but this
   # particular operation is common for expanding values stored in a texture
   # from [0,1] to [-1,1].
   (('~ffma@32', a,  2.0, -1.0), ('flrp', -1.0,  1.0,          a ), '!options->lower_flrp32'),
   (('~ffma@32', a, -2.0, -1.0), ('flrp', -1.0,  1.0, ('fneg', a)), '!options->lower_flrp32'),
   (('~ffma@32', a, -2.0,  1.0), ('flrp',  1.0, -1.0,          a ), '!options->lower_flrp32'),
   (('~ffma@32', a,  2.0,  1.0), ('flrp',  1.0, -1.0, ('fneg', a)), '!options->lower_flrp32'),
   (('~fadd@32', ('fmul(is_used_once)',  2.0, a), -1.0), ('flrp', -1.0,  1.0,          a ), '!options->lower_flrp32'),
   (('~fadd@32', ('fmul(is_used_once)', -2.0, a), -1.0), ('flrp', -1.0,  1.0, ('fneg', a)), '!options->lower_flrp32'),
   (('~fadd@32', ('fmul(is_used_once)', -2.0, a),  1.0), ('flrp',  1.0, -1.0,          a ), '!options->lower_flrp32'),
   (('~fadd@32', ('fmul(is_used_once)',  2.0, a),  1.0), ('flrp',  1.0, -1.0, ('fneg', a)), '!options->lower_flrp32'),

    # flrp(a, b, a)
    # a*(1-a) + b*a
    # a + -a*a + a*b    (1)
    # a + a*(b - a)
    # Option 1: ffma(a, (b-a), a)
    #
    # Alternately, after (1):
    # a*(1+b) + -a*a
    # a*((1+b) + -a)
    #
    # Let b=1
    #
    # Option 2: ffma(a, 2, -(a*a))
    # Option 3: ffma(a, 2, (-a)*a)
    # Option 4: ffma(a, -a, (2*a)
    # Option 5: a * (2 - a)
    #
    # There are a lot of other possible combinations.
   (('~ffma@32', ('fadd', b, ('fneg', a)), a, a), ('flrp', a, b, a), '!options->lower_flrp32'),
   (('~ffma@32', a, 2.0, ('fneg', ('fmul', a, a))), ('flrp', a, 1.0, a), '!options->lower_flrp32'),
   (('~ffma@32', a, 2.0, ('fmul', ('fneg', a), a)), ('flrp', a, 1.0, a), '!options->lower_flrp32'),
   (('~ffma@32', a, ('fneg', a), ('fmul', 2.0, a)), ('flrp', a, 1.0, a), '!options->lower_flrp32'),
   (('~fmul@32', a, ('fadd', 2.0, ('fneg', a))),    ('flrp', a, 1.0, a), '!options->lower_flrp32'),

   # we do these late so that we don't get in the way of creating ffmas
   (('fmin', ('fadd(is_used_once)', '#c', a), ('fadd(is_used_once)', '#c', b)), ('fadd', c, ('fmin', a, b))),
   (('fmax', ('fadd(is_used_once)', '#c', a), ('fadd(is_used_once)', '#c', b)), ('fadd', c, ('fmax', a, b))),

   # Putting this in 'optimizations' interferes with the bcsel(a, op(b, c),
   # op(b, d)) => op(b, bcsel(a, c, d)) transformations.  I do not know why.
   (('bcsel', ('feq', ('fsqrt', 'a(is_not_negative)'), 0.0), intBitsToFloat(0x7f7fffff), ('frsq', a)),
    ('fmin', ('frsq', a), intBitsToFloat(0x7f7fffff))),

   # Things that look like DPH in the source shader may get expanded to
   # something that looks like dot(v1.xyz, v2.xyz) + v1.w by the time it gets
   # to NIR.  After FFMA is generated, this can look like:
   #
   #    fadd(ffma(v1.z, v2.z, ffma(v1.y, v2.y, fmul(v1.x, v2.x))), v1.w)
   #
   # Reassociate the last addition into the first multiplication.
   #
   # Some shaders do not use 'invariant' in vertex and (possibly) geometry
   # shader stages on some outputs that are intended to be invariant.  For
   # various reasons, this optimization may not be fully applied in all
   # shaders used for different rendering passes of the same geometry.  This
   # can result in Z-fighting artifacts (at best).  For now, disable this
   # optimization in these stages.  See bugzilla #111490.  In tessellation
   # stages applications seem to use 'precise' when necessary, so allow the
   # optimization in those stages.
   (('~fadd', ('ffma(is_used_once)', a, b, ('ffma', c, d, ('fmul', 'e(is_not_const_and_not_fsign)', 'f(is_not_const_and_not_fsign)'))), 'g(is_not_const)'),
    ('ffma', a, b, ('ffma', c, d, ('ffma', e, 'f', 'g'))), '(info->stage != MESA_SHADER_VERTEX && info->stage != MESA_SHADER_GEOMETRY) && !options->intel_vec4'),
   (('~fadd', ('ffma(is_used_once)', a, b, ('fmul', 'c(is_not_const_and_not_fsign)', 'd(is_not_const_and_not_fsign)') ), 'e(is_not_const)'),
    ('ffma', a, b, ('ffma', c, d, e)), '(info->stage != MESA_SHADER_VERTEX && info->stage != MESA_SHADER_GEOMETRY) && !options->intel_vec4'),

   # Section 8.8 (Integer Functions) of the GLSL 4.60 spec says:
   #
   #    If bits is zero, the result will be zero.
   #
   # These prevent the next two lowerings generating incorrect results when
   # count is zero.
   (('ubfe', a, b, 0), 0),
   (('ibfe', a, b, 0), 0),

   # On Intel GPUs, BFE is a 3-source instruction.  Like all 3-source
   # instructions on Intel GPUs, it cannot have an immediate values as
   # sources.  There are also limitations on source register strides.  As a
   # result, it is very easy for 3-source instruction combined with either
   # loads of immediate values or copies from weird register strides to be
   # more expensive than the primitive instructions it represents.
   (('ubfe', a, '#b', '#c'), ('iand', ('ushr', 0xffffffff, ('ineg', c)), ('ushr', a, b)), 'options->lower_bfe_with_two_constants'),

   # b is the lowest order bit to be extracted and c is the number of bits to
   # extract.  The inner shift removes the bits above b + c by shifting left
   # 32 - (b + c).  ishl only sees the low 5 bits of the shift count, which is
   # -(b + c).  The outer shift moves the bit that was at b to bit zero.
   # After the first shift, that bit is now at b + (32 - (b + c)) or 32 - c.
   # This means that it must be shifted right by 32 - c or -c bits.
   (('ibfe', a, '#b', '#c'), ('ishr', ('ishl', a, ('ineg', ('iadd', b, c))), ('ineg', c)), 'options->lower_bfe_with_two_constants'),

   # Clean up no-op shifts that may result from the bfe lowerings.
   (('ishl', a, 0), a),
   (('ishl', a, -32), a),
   (('ishr', a, 0), a),
   (('ishr', a, -32), a),
   (('ushr', a, 0), a),
]

# Integer sizes
for s in [8, 16, 32, 64]:
    late_optimizations.extend([
        (('iand', ('ine(is_used_once)', 'a@{}'.format(s), 0), ('ine', 'b@{}'.format(s), 0)), ('ine', ('umin', a, b), 0)),
        (('ior',  ('ieq(is_used_once)', 'a@{}'.format(s), 0), ('ieq', 'b@{}'.format(s), 0)), ('ieq', ('umin', a, b), 0)),
    ])

# Float sizes
for s in [16, 32, 64]:
    late_optimizations.extend([
       (('~fadd@{}'.format(s), 1.0, ('fmul(is_used_once)', c , ('fadd', b, -1.0 ))), ('fadd', ('fadd', 1.0, ('fneg', c)), ('fmul', b, c)), 'options->lower_flrp{}'.format(s)),
       (('bcsel', a, 0, ('b2f{}'.format(s), ('inot', 'b@bool'))), ('b2f{}'.format(s), ('inot', ('ior', a, b)))),
    ])

for op in ['fadd']:
    late_optimizations += [
        (('bcsel', a, (op + '(is_used_once)', b, c), (op, b, d)), (op, b, ('bcsel', a, c, d))),
        (('bcsel', a, (op, b, c), (op + '(is_used_once)', b, d)), (op, b, ('bcsel', a, c, d))),
    ]

for op in ['ffma']:
    late_optimizations += [
        (('bcsel', a, (op + '(is_used_once)', b, c, d), (op, b, c, e)), (op, b, c, ('bcsel', a, d, e))),
        (('bcsel', a, (op, b, c, d), (op + '(is_used_once)', b, c, e)), (op, b, c, ('bcsel', a, d, e))),

        (('bcsel', a, (op + '(is_used_once)', b, c, d), (op, b, e, d)), (op, b, ('bcsel', a, c, e), d)),
        (('bcsel', a, (op, b, c, d), (op + '(is_used_once)', b, e, d)), (op, b, ('bcsel', a, c, e), d)),
    ]

# mediump: If an opcode is surrounded by conversions, remove the conversions.
# The rationale is that type conversions + the low precision opcode are more
# expensive that the same arithmetic opcode at higher precision.
#
# This must be done in late optimizations, because we need normal optimizations to
# first eliminate temporary up-conversions such as in op1(f2fmp(f2f32(op2()))).
#
# Unary opcodes
for op in ['fabs', 'fceil', 'fcos', 'fddx', 'fddx_coarse', 'fddx_fine', 'fddy',
           'fddy_coarse', 'fddy_fine', 'fexp2', 'ffloor', 'ffract', 'flog2', 'fneg',
           'frcp', 'fround_even', 'frsq', 'fsat', 'fsign', 'fsin', 'fsqrt']:
    late_optimizations += [(('~f2f32', (op, ('f2fmp', a))), (op, a))]

# Binary opcodes
for op in ['fadd', 'fdiv', 'fmax', 'fmin', 'fmod', 'fmul', 'fpow', 'frem']:
    late_optimizations += [(('~f2f32', (op, ('f2fmp', a), ('f2fmp', b))), (op, a, b))]

# Ternary opcodes
for op in ['ffma', 'flrp']:
    late_optimizations += [(('~f2f32', (op, ('f2fmp', a), ('f2fmp', b), ('f2fmp', c))), (op, a, b, c))]

# Comparison opcodes
for op in ['feq', 'fge', 'flt', 'fneu']:
    late_optimizations += [(('~' + op, ('f2fmp', a), ('f2fmp', b)), (op, a, b))]

# Do this last, so that the f2fmp patterns above have effect.
late_optimizations += [
  # Convert *2*mp instructions to concrete *2*16 instructions. At this point
  # any conversions that could have been removed will have been removed in
  # nir_opt_algebraic so any remaining ones are required.
  (('f2fmp', a), ('f2f16', a)),
  (('f2imp', a), ('f2i16', a)),
  (('f2ump', a), ('f2u16', a)),
  (('i2imp', a), ('i2i16', a)),
  (('i2fmp', a), ('i2f16', a)),
  (('i2imp', a), ('u2u16', a)),
  (('u2fmp', a), ('u2f16', a)),
]

distribute_src_mods = [
   # Try to remove some spurious negations rather than pushing them down.
   (('fmul', ('fneg', a), ('fneg', b)), ('fmul', a, b)),
   (('ffma', ('fneg', a), ('fneg', b), c), ('ffma', a, b, c)),
   (('fdot_replicated2', ('fneg', a), ('fneg', b)), ('fdot_replicated2', a, b)),
   (('fdot_replicated3', ('fneg', a), ('fneg', b)), ('fdot_replicated3', a, b)),
   (('fdot_replicated4', ('fneg', a), ('fneg', b)), ('fdot_replicated4', a, b)),
   (('fneg', ('fneg', a)), a),

   (('fneg', ('fmul(is_used_once)', a, b)), ('fmul', ('fneg', a), b)),
   (('fabs', ('fmul(is_used_once)', a, b)), ('fmul', ('fabs', a), ('fabs', b))),

   (('fneg', ('ffma(is_used_once)', a, b, c)), ('ffma', ('fneg', a), b, ('fneg', c))),
   (('fneg', ('flrp(is_used_once)', a, b, c)), ('flrp', ('fneg', a), ('fneg', b), c)),
   (('fneg', ('fadd(is_used_once)', a, b)), ('fadd', ('fneg', a), ('fneg', b))),

   # Note that fmin <-> fmax.  I don't think there is a way to distribute
   # fabs() into fmin or fmax.
   (('fneg', ('fmin(is_used_once)', a, b)), ('fmax', ('fneg', a), ('fneg', b))),
   (('fneg', ('fmax(is_used_once)', a, b)), ('fmin', ('fneg', a), ('fneg', b))),

   (('fneg', ('fdot_replicated2(is_used_once)', a, b)), ('fdot_replicated2', ('fneg', a), b)),
   (('fneg', ('fdot_replicated3(is_used_once)', a, b)), ('fdot_replicated3', ('fneg', a), b)),
   (('fneg', ('fdot_replicated4(is_used_once)', a, b)), ('fdot_replicated4', ('fneg', a), b)),

   # fdph works mostly like fdot, but to get the correct result, the negation
   # must be applied to the second source.
   (('fneg', ('fdph_replicated(is_used_once)', a, b)), ('fdph_replicated', a, ('fneg', b))),

   (('fneg', ('fsign(is_used_once)', a)), ('fsign', ('fneg', a))),
   (('fabs', ('fsign(is_used_once)', a)), ('fsign', ('fabs', a))),
]

print(nir_algebraic.AlgebraicPass("nir_opt_algebraic", optimizations).render())
print(nir_algebraic.AlgebraicPass("nir_opt_algebraic_before_ffma",
                                  before_ffma_optimizations).render())
print(nir_algebraic.AlgebraicPass("nir_opt_algebraic_late",
                                  late_optimizations).render())
print(nir_algebraic.AlgebraicPass("nir_opt_algebraic_distribute_src_mods",
                                  distribute_src_mods).render())