aboutsummaryrefslogtreecommitdiff
path: root/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
blob: bf520a56040439a64c9173e3472984db378a304d (plain) (blame)
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
2290
2291
2292
2293
2294
2295
2296
2297
2298
2299
2300
2301
2302
2303
2304
2305
2306
2307
2308
2309
2310
2311
2312
2313
2314
2315
2316
2317
2318
2319
2320
2321
2322
2323
2324
2325
2326
2327
2328
2329
2330
2331
2332
2333
2334
2335
2336
2337
2338
2339
2340
2341
2342
2343
2344
2345
2346
2347
2348
2349
2350
2351
2352
2353
2354
2355
2356
2357
2358
2359
2360
2361
2362
2363
2364
2365
2366
2367
2368
2369
2370
2371
2372
2373
2374
2375
2376
2377
2378
2379
2380
2381
2382
2383
2384
2385
2386
2387
2388
2389
2390
2391
2392
2393
2394
2395
2396
2397
2398
2399
2400
2401
2402
2403
2404
2405
2406
2407
2408
2409
2410
2411
2412
2413
2414
2415
2416
2417
2418
2419
2420
2421
2422
2423
2424
2425
2426
2427
2428
2429
2430
2431
2432
2433
2434
2435
2436
2437
2438
2439
2440
2441
2442
2443
2444
2445
2446
2447
2448
2449
2450
2451
2452
2453
2454
2455
2456
2457
2458
2459
2460
2461
2462
2463
2464
2465
2466
2467
2468
2469
2470
2471
2472
2473
2474
2475
2476
2477
2478
2479
2480
2481
2482
2483
2484
2485
2486
2487
2488
2489
2490
2491
2492
2493
2494
2495
2496
2497
2498
2499
2500
2501
2502
2503
2504
2505
2506
2507
2508
2509
2510
2511
2512
2513
2514
2515
2516
2517
2518
2519
2520
2521
2522
2523
2524
2525
2526
2527
2528
2529
2530
2531
2532
2533
2534
2535
2536
2537
2538
2539
2540
2541
2542
2543
2544
2545
2546
2547
2548
2549
2550
2551
2552
2553
2554
2555
2556
2557
2558
2559
2560
2561
2562
2563
2564
2565
2566
2567
2568
2569
2570
2571
2572
2573
2574
2575
2576
2577
2578
2579
2580
2581
2582
2583
2584
2585
2586
2587
2588
2589
2590
2591
2592
2593
2594
2595
2596
2597
2598
2599
2600
2601
2602
2603
2604
2605
2606
2607
2608
2609
2610
2611
2612
2613
2614
2615
2616
2617
2618
2619
2620
2621
2622
2623
2624
2625
2626
2627
2628
2629
2630
2631
2632
2633
2634
2635
2636
2637
2638
2639
2640
2641
2642
2643
2644
2645
2646
2647
2648
2649
2650
2651
2652
2653
2654
2655
2656
2657
2658
2659
2660
2661
2662
2663
2664
2665
2666
2667
2668
2669
2670
2671
2672
2673
2674
2675
2676
2677
2678
2679
2680
2681
2682
2683
2684
2685
2686
2687
2688
2689
2690
2691
2692
2693
2694
2695
2696
2697
2698
2699
2700
2701
2702
2703
2704
2705
2706
2707
2708
2709
2710
2711
2712
2713
2714
2715
2716
2717
2718
2719
2720
2721
2722
2723
2724
2725
2726
2727
2728
2729
2730
2731
2732
2733
2734
2735
2736
2737
2738
2739
2740
2741
2742
2743
2744
2745
2746
2747
2748
2749
2750
2751
2752
2753
2754
2755
2756
2757
2758
2759
2760
2761
2762
2763
2764
2765
2766
2767
2768
2769
2770
2771
2772
2773
2774
2775
2776
2777
2778
2779
2780
2781
2782
2783
2784
2785
2786
2787
2788
2789
2790
2791
2792
2793
2794
2795
2796
2797
2798
2799
2800
2801
2802
2803
2804
2805
2806
2807
2808
2809
2810
2811
2812
2813
2814
2815
2816
2817
2818
2819
2820
2821
2822
2823
2824
2825
2826
2827
2828
2829
2830
2831
2832
2833
2834
2835
2836
2837
2838
2839
2840
2841
2842
2843
2844
2845
2846
2847
2848
2849
2850
2851
2852
2853
2854
2855
2856
2857
2858
2859
2860
2861
2862
2863
2864
2865
2866
2867
2868
2869
2870
2871
2872
2873
2874
2875
2876
2877
2878
2879
2880
2881
2882
2883
2884
2885
2886
2887
2888
2889
2890
2891
2892
2893
2894
2895
2896
2897
2898
2899
2900
2901
2902
2903
2904
2905
2906
2907
2908
2909
2910
2911
2912
2913
2914
2915
2916
2917
2918
2919
2920
2921
2922
2923
2924
2925
2926
2927
2928
2929
2930
2931
2932
2933
2934
2935
2936
2937
2938
2939
2940
2941
2942
2943
2944
2945
2946
2947
2948
2949
2950
2951
2952
2953
2954
2955
2956
2957
2958
2959
2960
2961
2962
2963
2964
2965
2966
2967
2968
2969
2970
2971
2972
2973
2974
2975
2976
2977
2978
2979
2980
2981
2982
2983
2984
2985
2986
2987
2988
2989
2990
2991
2992
2993
2994
2995
2996
2997
2998
2999
3000
3001
3002
3003
3004
3005
3006
3007
3008
3009
3010
3011
3012
3013
3014
3015
3016
3017
3018
3019
3020
3021
3022
3023
3024
3025
3026
3027
3028
3029
3030
3031
3032
3033
3034
3035
3036
3037
3038
3039
3040
3041
3042
3043
3044
3045
3046
3047
3048
3049
3050
3051
3052
3053
3054
3055
3056
3057
3058
3059
3060
3061
3062
3063
3064
3065
3066
3067
3068
3069
3070
3071
3072
3073
3074
3075
3076
3077
3078
3079
3080
3081
3082
3083
3084
3085
3086
3087
3088
3089
3090
3091
3092
3093
3094
3095
3096
3097
3098
3099
3100
3101
3102
3103
3104
3105
3106
3107
3108
3109
3110
3111
3112
3113
3114
3115
3116
3117
3118
3119
3120
3121
3122
3123
3124
3125
3126
3127
3128
3129
3130
3131
3132
3133
3134
3135
3136
3137
3138
3139
3140
3141
3142
3143
3144
3145
3146
3147
3148
3149
3150
3151
3152
3153
3154
3155
3156
3157
3158
3159
3160
3161
3162
3163
3164
3165
3166
3167
3168
3169
3170
3171
3172
3173
3174
3175
3176
3177
3178
3179
3180
3181
3182
3183
3184
3185
3186
3187
3188
3189
3190
3191
3192
3193
3194
3195
3196
3197
3198
3199
3200
3201
3202
3203
3204
3205
3206
3207
3208
3209
3210
3211
3212
3213
3214
3215
3216
3217
3218
3219
3220
3221
3222
3223
3224
3225
3226
3227
3228
3229
3230
3231
3232
3233
3234
3235
3236
3237
3238
3239
3240
3241
3242
3243
3244
3245
3246
3247
3248
3249
3250
3251
3252
3253
3254
3255
3256
3257
3258
3259
3260
3261
3262
3263
3264
3265
3266
3267
3268
3269
3270
3271
3272
3273
3274
3275
3276
3277
3278
3279
3280
3281
3282
3283
3284
3285
3286
3287
3288
3289
3290
3291
3292
3293
3294
3295
3296
3297
3298
3299
3300
3301
3302
3303
3304
3305
3306
3307
3308
3309
3310
3311
3312
3313
3314
3315
3316
3317
3318
3319
3320
3321
3322
3323
3324
3325
3326
3327
3328
3329
3330
3331
3332
3333
3334
3335
3336
3337
3338
3339
3340
3341
3342
3343
3344
3345
3346
3347
3348
3349
3350
3351
3352
3353
3354
3355
3356
3357
3358
3359
3360
3361
3362
3363
3364
3365
3366
3367
3368
3369
3370
3371
3372
3373
3374
3375
3376
3377
3378
3379
3380
3381
3382
3383
3384
3385
3386
3387
3388
3389
3390
3391
3392
3393
3394
3395
3396
3397
3398
3399
3400
3401
3402
3403
3404
3405
3406
3407
3408
3409
3410
3411
3412
3413
3414
3415
3416
3417
3418
3419
3420
3421
3422
3423
3424
3425
3426
3427
3428
3429
3430
3431
3432
3433
3434
3435
3436
3437
3438
3439
3440
3441
3442
3443
3444
3445
3446
3447
3448
3449
3450
3451
3452
3453
3454
3455
3456
3457
3458
3459
3460
3461
3462
3463
3464
3465
3466
3467
3468
3469
3470
3471
3472
3473
3474
3475
3476
3477
3478
3479
3480
3481
3482
3483
3484
3485
3486
3487
3488
3489
3490
3491
3492
3493
3494
3495
3496
3497
3498
3499
3500
3501
3502
3503
3504
3505
3506
3507
3508
3509
3510
3511
3512
3513
3514
3515
3516
3517
3518
3519
3520
3521
3522
3523
3524
3525
3526
3527
3528
3529
3530
3531
3532
3533
3534
3535
3536
3537
3538
3539
3540
3541
3542
3543
3544
3545
3546
3547
3548
3549
3550
3551
3552
3553
3554
3555
3556
3557
3558
3559
3560
3561
3562
3563
3564
3565
3566
3567
3568
3569
3570
3571
3572
3573
3574
3575
3576
3577
3578
3579
3580
3581
3582
3583
3584
3585
3586
3587
3588
3589
3590
3591
3592
3593
3594
3595
3596
3597
3598
3599
3600
3601
3602
3603
3604
3605
3606
3607
3608
3609
3610
3611
3612
3613
3614
3615
3616
3617
3618
3619
3620
3621
3622
3623
3624
3625
3626
3627
3628
3629
3630
3631
3632
3633
3634
3635
3636
3637
3638
3639
3640
3641
3642
3643
3644
3645
3646
3647
3648
3649
3650
3651
3652
3653
3654
3655
3656
3657
3658
3659
3660
3661
3662
3663
3664
3665
3666
3667
3668
3669
3670
3671
3672
3673
3674
3675
3676
3677
3678
3679
3680
3681
3682
3683
3684
3685
3686
3687
3688
3689
3690
3691
3692
3693
3694
3695
3696
3697
3698
3699
3700
3701
3702
3703
3704
3705
3706
3707
3708
3709
3710
3711
3712
3713
3714
3715
3716
3717
3718
3719
3720
3721
3722
3723
3724
3725
3726
3727
3728
3729
3730
3731
3732
3733
3734
3735
3736
3737
3738
3739
3740
3741
3742
3743
3744
3745
3746
3747
3748
3749
3750
3751
3752
3753
3754
3755
3756
3757
3758
3759
3760
3761
3762
3763
3764
3765
3766
3767
3768
3769
3770
3771
3772
3773
3774
3775
3776
3777
3778
3779
3780
3781
3782
3783
3784
3785
3786
3787
3788
3789
3790
3791
3792
3793
3794
3795
3796
3797
3798
3799
3800
3801
3802
3803
3804
3805
3806
3807
3808
3809
3810
3811
3812
3813
3814
3815
3816
3817
3818
3819
3820
3821
3822
3823
3824
3825
3826
3827
3828
3829
3830
3831
3832
3833
3834
3835
3836
3837
3838
3839
3840
3841
3842
3843
3844
3845
3846
3847
3848
3849
3850
3851
3852
3853
3854
3855
3856
3857
3858
3859
3860
3861
3862
3863
3864
3865
3866
3867
3868
3869
3870
3871
3872
3873
3874
3875
3876
3877
3878
3879
3880
3881
3882
3883
3884
3885
3886
3887
3888
3889
3890
3891
3892
3893
3894
3895
3896
3897
3898
3899
3900
3901
3902
3903
3904
3905
3906
3907
3908
3909
3910
3911
3912
3913
3914
3915
3916
3917
3918
3919
3920
3921
3922
3923
3924
3925
3926
3927
3928
3929
3930
3931
3932
3933
3934
3935
3936
3937
3938
3939
3940
3941
3942
3943
3944
3945
3946
3947
3948
3949
3950
3951
3952
3953
3954
3955
3956
3957
3958
3959
3960
3961
3962
3963
3964
3965
3966
3967
3968
3969
3970
3971
3972
3973
3974
3975
3976
3977
3978
3979
3980
3981
3982
3983
3984
3985
3986
3987
3988
3989
3990
3991
3992
3993
3994
3995
3996
3997
3998
3999
4000
4001
4002
4003
4004
4005
4006
4007
4008
4009
4010
4011
4012
4013
4014
4015
4016
4017
4018
4019
4020
4021
4022
4023
4024
4025
4026
4027
4028
4029
4030
4031
4032
4033
4034
4035
4036
4037
4038
4039
4040
4041
4042
4043
4044
4045
4046
4047
4048
4049
4050
4051
4052
4053
4054
4055
4056
4057
4058
4059
4060
4061
4062
4063
4064
4065
4066
4067
4068
4069
4070
4071
4072
4073
4074
4075
4076
4077
4078
4079
4080
4081
4082
4083
4084
4085
4086
4087
4088
4089
4090
4091
4092
4093
4094
4095
4096
4097
4098
4099
4100
4101
4102
4103
4104
4105
4106
4107
4108
4109
4110
4111
4112
4113
4114
4115
4116
4117
4118
4119
4120
4121
4122
4123
4124
4125
4126
4127
4128
4129
4130
4131
4132
4133
4134
4135
4136
4137
4138
4139
4140
4141
4142
4143
4144
4145
4146
4147
4148
4149
4150
4151
4152
4153
4154
4155
4156
4157
4158
4159
4160
4161
4162
4163
4164
4165
4166
4167
4168
4169
4170
4171
4172
4173
4174
4175
4176
4177
4178
4179
4180
4181
4182
4183
4184
4185
4186
4187
4188
4189
4190
4191
4192
4193
4194
4195
4196
4197
4198
4199
4200
4201
4202
4203
4204
4205
4206
4207
4208
4209
4210
4211
4212
4213
4214
4215
4216
4217
4218
4219
4220
4221
4222
4223
4224
4225
4226
4227
4228
4229
4230
4231
4232
4233
4234
4235
4236
4237
4238
4239
4240
4241
4242
4243
4244
4245
4246
4247
4248
4249
4250
4251
4252
4253
4254
4255
4256
4257
4258
4259
4260
4261
4262
4263
4264
4265
4266
4267
4268
4269
4270
4271
4272
4273
4274
4275
4276
4277
4278
4279
4280
4281
4282
4283
4284
4285
4286
4287
4288
4289
4290
4291
4292
4293
4294
4295
4296
4297
4298
4299
4300
4301
4302
4303
4304
4305
4306
4307
4308
4309
4310
4311
4312
4313
4314
4315
4316
4317
4318
4319
4320
4321
4322
4323
4324
4325
4326
4327
4328
4329
4330
4331
4332
4333
4334
4335
4336
4337
4338
4339
4340
4341
4342
4343
4344
4345
4346
4347
4348
4349
4350
4351
4352
4353
4354
4355
4356
4357
4358
4359
4360
4361
4362
4363
4364
4365
4366
4367
4368
4369
4370
4371
4372
4373
4374
4375
4376
4377
4378
4379
4380
4381
4382
4383
4384
4385
4386
4387
4388
4389
4390
4391
4392
4393
4394
4395
4396
4397
4398
4399
4400
4401
4402
4403
4404
4405
4406
4407
4408
4409
4410
4411
4412
4413
4414
4415
4416
4417
4418
4419
4420
4421
4422
4423
4424
4425
4426
4427
4428
4429
4430
4431
4432
4433
4434
4435
4436
4437
4438
4439
4440
4441
4442
4443
4444
4445
4446
4447
4448
4449
4450
4451
4452
4453
4454
4455
4456
4457
4458
4459
4460
4461
4462
4463
4464
4465
4466
4467
4468
4469
4470
4471
4472
4473
4474
4475
4476
4477
4478
4479
4480
4481
4482
4483
4484
4485
4486
4487
4488
4489
4490
4491
4492
4493
4494
4495
4496
4497
4498
4499
4500
4501
4502
4503
4504
4505
4506
4507
4508
4509
4510
4511
4512
4513
4514
4515
4516
4517
4518
4519
4520
4521
4522
4523
4524
4525
4526
4527
4528
4529
4530
4531
4532
4533
4534
4535
4536
4537
4538
4539
4540
4541
4542
4543
4544
4545
4546
4547
4548
4549
4550
4551
4552
4553
4554
4555
4556
4557
4558
4559
4560
4561
4562
4563
4564
4565
4566
4567
4568
4569
4570
4571
4572
4573
4574
4575
4576
4577
4578
4579
4580
4581
4582
4583
4584
4585
4586
4587
4588
4589
4590
4591
4592
4593
4594
4595
4596
4597
4598
4599
4600
4601
4602
4603
4604
4605
4606
4607
4608
4609
4610
4611
4612
4613
4614
4615
4616
4617
4618
4619
4620
4621
4622
4623
4624
4625
4626
4627
4628
4629
4630
4631
4632
4633
4634
4635
4636
4637
4638
4639
4640
4641
4642
4643
4644
4645
4646
4647
4648
4649
4650
4651
4652
4653
4654
4655
4656
4657
4658
4659
4660
4661
4662
4663
4664
4665
4666
4667
4668
4669
4670
4671
4672
4673
4674
4675
4676
4677
4678
4679
4680
4681
4682
4683
4684
4685
4686
4687
4688
4689
4690
4691
4692
4693
4694
4695
4696
4697
4698
4699
4700
4701
4702
4703
4704
4705
4706
4707
4708
4709
4710
4711
4712
4713
4714
4715
4716
4717
4718
4719
4720
4721
4722
4723
4724
4725
4726
4727
4728
4729
4730
4731
4732
4733
4734
4735
4736
4737
4738
4739
4740
4741
4742
4743
4744
4745
4746
4747
4748
4749
4750
4751
4752
4753
4754
4755
4756
4757
4758
4759
4760
4761
4762
4763
4764
4765
4766
4767
4768
4769
4770
4771
4772
4773
4774
4775
4776
4777
4778
4779
4780
4781
4782
4783
4784
4785
4786
4787
4788
4789
4790
4791
4792
4793
4794
4795
4796
4797
4798
4799
4800
4801
4802
4803
4804
4805
4806
4807
4808
4809
4810
4811
4812
4813
4814
4815
4816
4817
4818
//===-- AMDGPUISelLowering.cpp - AMDGPU Common DAG lowering functions -----===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
/// \file
/// This is the parent TargetLowering class for hardware code gen
/// targets.
//
//===----------------------------------------------------------------------===//

#include "AMDGPUISelLowering.h"
#include "AMDGPU.h"
#include "AMDGPUInstrInfo.h"
#include "AMDGPUMachineFunction.h"
#include "GCNSubtarget.h"
#include "SIMachineFunctionInfo.h"
#include "llvm/CodeGen/Analysis.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/IR/DiagnosticInfo.h"
#include "llvm/IR/IntrinsicsAMDGPU.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/KnownBits.h"
#include "llvm/Target/TargetMachine.h"

using namespace llvm;

#include "AMDGPUGenCallingConv.inc"

static cl::opt<bool> AMDGPUBypassSlowDiv(
  "amdgpu-bypass-slow-div",
  cl::desc("Skip 64-bit divide for dynamic 32-bit values"),
  cl::init(true));

// Find a larger type to do a load / store of a vector with.
EVT AMDGPUTargetLowering::getEquivalentMemType(LLVMContext &Ctx, EVT VT) {
  unsigned StoreSize = VT.getStoreSizeInBits();
  if (StoreSize <= 32)
    return EVT::getIntegerVT(Ctx, StoreSize);

  assert(StoreSize % 32 == 0 && "Store size not a multiple of 32");
  return EVT::getVectorVT(Ctx, MVT::i32, StoreSize / 32);
}

unsigned AMDGPUTargetLowering::numBitsUnsigned(SDValue Op, SelectionDAG &DAG) {
  return DAG.computeKnownBits(Op).countMaxActiveBits();
}

unsigned AMDGPUTargetLowering::numBitsSigned(SDValue Op, SelectionDAG &DAG) {
  // In order for this to be a signed 24-bit value, bit 23, must
  // be a sign bit.
  return DAG.ComputeMaxSignificantBits(Op);
}

AMDGPUTargetLowering::AMDGPUTargetLowering(const TargetMachine &TM,
                                           const AMDGPUSubtarget &STI)
    : TargetLowering(TM), Subtarget(&STI) {
  // Lower floating point store/load to integer store/load to reduce the number
  // of patterns in tablegen.
  setOperationAction(ISD::LOAD, MVT::f32, Promote);
  AddPromotedToType(ISD::LOAD, MVT::f32, MVT::i32);

  setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
  AddPromotedToType(ISD::LOAD, MVT::v2f32, MVT::v2i32);

  setOperationAction(ISD::LOAD, MVT::v3f32, Promote);
  AddPromotedToType(ISD::LOAD, MVT::v3f32, MVT::v3i32);

  setOperationAction(ISD::LOAD, MVT::v4f32, Promote);
  AddPromotedToType(ISD::LOAD, MVT::v4f32, MVT::v4i32);

  setOperationAction(ISD::LOAD, MVT::v5f32, Promote);
  AddPromotedToType(ISD::LOAD, MVT::v5f32, MVT::v5i32);

  setOperationAction(ISD::LOAD, MVT::v6f32, Promote);
  AddPromotedToType(ISD::LOAD, MVT::v6f32, MVT::v6i32);

  setOperationAction(ISD::LOAD, MVT::v7f32, Promote);
  AddPromotedToType(ISD::LOAD, MVT::v7f32, MVT::v7i32);

  setOperationAction(ISD::LOAD, MVT::v8f32, Promote);
  AddPromotedToType(ISD::LOAD, MVT::v8f32, MVT::v8i32);

  setOperationAction(ISD::LOAD, MVT::v16f32, Promote);
  AddPromotedToType(ISD::LOAD, MVT::v16f32, MVT::v16i32);

  setOperationAction(ISD::LOAD, MVT::v32f32, Promote);
  AddPromotedToType(ISD::LOAD, MVT::v32f32, MVT::v32i32);

  setOperationAction(ISD::LOAD, MVT::i64, Promote);
  AddPromotedToType(ISD::LOAD, MVT::i64, MVT::v2i32);

  setOperationAction(ISD::LOAD, MVT::v2i64, Promote);
  AddPromotedToType(ISD::LOAD, MVT::v2i64, MVT::v4i32);

  setOperationAction(ISD::LOAD, MVT::f64, Promote);
  AddPromotedToType(ISD::LOAD, MVT::f64, MVT::v2i32);

  setOperationAction(ISD::LOAD, MVT::v2f64, Promote);
  AddPromotedToType(ISD::LOAD, MVT::v2f64, MVT::v4i32);

  setOperationAction(ISD::LOAD, MVT::v3i64, Promote);
  AddPromotedToType(ISD::LOAD, MVT::v3i64, MVT::v6i32);

  setOperationAction(ISD::LOAD, MVT::v4i64, Promote);
  AddPromotedToType(ISD::LOAD, MVT::v4i64, MVT::v8i32);

  setOperationAction(ISD::LOAD, MVT::v3f64, Promote);
  AddPromotedToType(ISD::LOAD, MVT::v3f64, MVT::v6i32);

  setOperationAction(ISD::LOAD, MVT::v4f64, Promote);
  AddPromotedToType(ISD::LOAD, MVT::v4f64, MVT::v8i32);

  setOperationAction(ISD::LOAD, MVT::v8i64, Promote);
  AddPromotedToType(ISD::LOAD, MVT::v8i64, MVT::v16i32);

  setOperationAction(ISD::LOAD, MVT::v8f64, Promote);
  AddPromotedToType(ISD::LOAD, MVT::v8f64, MVT::v16i32);

  setOperationAction(ISD::LOAD, MVT::v16i64, Promote);
  AddPromotedToType(ISD::LOAD, MVT::v16i64, MVT::v32i32);

  setOperationAction(ISD::LOAD, MVT::v16f64, Promote);
  AddPromotedToType(ISD::LOAD, MVT::v16f64, MVT::v32i32);

  // There are no 64-bit extloads. These should be done as a 32-bit extload and
  // an extension to 64-bit.
  for (MVT VT : MVT::integer_valuetypes())
    setLoadExtAction({ISD::EXTLOAD, ISD::SEXTLOAD, ISD::ZEXTLOAD}, MVT::i64, VT,
                     Expand);

  for (MVT VT : MVT::integer_valuetypes()) {
    if (VT == MVT::i64)
      continue;

    for (auto Op : {ISD::SEXTLOAD, ISD::ZEXTLOAD, ISD::EXTLOAD}) {
      setLoadExtAction(Op, VT, MVT::i1, Promote);
      setLoadExtAction(Op, VT, MVT::i8, Legal);
      setLoadExtAction(Op, VT, MVT::i16, Legal);
      setLoadExtAction(Op, VT, MVT::i32, Expand);
    }
  }

  for (MVT VT : MVT::integer_fixedlen_vector_valuetypes())
    for (auto MemVT :
         {MVT::v2i8, MVT::v4i8, MVT::v2i16, MVT::v3i16, MVT::v4i16})
      setLoadExtAction({ISD::SEXTLOAD, ISD::ZEXTLOAD, ISD::EXTLOAD}, VT, MemVT,
                       Expand);

  setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand);
  setLoadExtAction(ISD::EXTLOAD, MVT::v2f32, MVT::v2f16, Expand);
  setLoadExtAction(ISD::EXTLOAD, MVT::v3f32, MVT::v3f16, Expand);
  setLoadExtAction(ISD::EXTLOAD, MVT::v4f32, MVT::v4f16, Expand);
  setLoadExtAction(ISD::EXTLOAD, MVT::v8f32, MVT::v8f16, Expand);
  setLoadExtAction(ISD::EXTLOAD, MVT::v16f32, MVT::v16f16, Expand);
  setLoadExtAction(ISD::EXTLOAD, MVT::v32f32, MVT::v32f16, Expand);

  setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f32, Expand);
  setLoadExtAction(ISD::EXTLOAD, MVT::v2f64, MVT::v2f32, Expand);
  setLoadExtAction(ISD::EXTLOAD, MVT::v3f64, MVT::v3f32, Expand);
  setLoadExtAction(ISD::EXTLOAD, MVT::v4f64, MVT::v4f32, Expand);
  setLoadExtAction(ISD::EXTLOAD, MVT::v8f64, MVT::v8f32, Expand);
  setLoadExtAction(ISD::EXTLOAD, MVT::v16f64, MVT::v16f32, Expand);

  setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand);
  setLoadExtAction(ISD::EXTLOAD, MVT::v2f64, MVT::v2f16, Expand);
  setLoadExtAction(ISD::EXTLOAD, MVT::v3f64, MVT::v3f16, Expand);
  setLoadExtAction(ISD::EXTLOAD, MVT::v4f64, MVT::v4f16, Expand);
  setLoadExtAction(ISD::EXTLOAD, MVT::v8f64, MVT::v8f16, Expand);
  setLoadExtAction(ISD::EXTLOAD, MVT::v16f64, MVT::v16f16, Expand);

  setOperationAction(ISD::STORE, MVT::f32, Promote);
  AddPromotedToType(ISD::STORE, MVT::f32, MVT::i32);

  setOperationAction(ISD::STORE, MVT::v2f32, Promote);
  AddPromotedToType(ISD::STORE, MVT::v2f32, MVT::v2i32);

  setOperationAction(ISD::STORE, MVT::v3f32, Promote);
  AddPromotedToType(ISD::STORE, MVT::v3f32, MVT::v3i32);

  setOperationAction(ISD::STORE, MVT::v4f32, Promote);
  AddPromotedToType(ISD::STORE, MVT::v4f32, MVT::v4i32);

  setOperationAction(ISD::STORE, MVT::v5f32, Promote);
  AddPromotedToType(ISD::STORE, MVT::v5f32, MVT::v5i32);

  setOperationAction(ISD::STORE, MVT::v6f32, Promote);
  AddPromotedToType(ISD::STORE, MVT::v6f32, MVT::v6i32);

  setOperationAction(ISD::STORE, MVT::v7f32, Promote);
  AddPromotedToType(ISD::STORE, MVT::v7f32, MVT::v7i32);

  setOperationAction(ISD::STORE, MVT::v8f32, Promote);
  AddPromotedToType(ISD::STORE, MVT::v8f32, MVT::v8i32);

  setOperationAction(ISD::STORE, MVT::v16f32, Promote);
  AddPromotedToType(ISD::STORE, MVT::v16f32, MVT::v16i32);

  setOperationAction(ISD::STORE, MVT::v32f32, Promote);
  AddPromotedToType(ISD::STORE, MVT::v32f32, MVT::v32i32);

  setOperationAction(ISD::STORE, MVT::i64, Promote);
  AddPromotedToType(ISD::STORE, MVT::i64, MVT::v2i32);

  setOperationAction(ISD::STORE, MVT::v2i64, Promote);
  AddPromotedToType(ISD::STORE, MVT::v2i64, MVT::v4i32);

  setOperationAction(ISD::STORE, MVT::f64, Promote);
  AddPromotedToType(ISD::STORE, MVT::f64, MVT::v2i32);

  setOperationAction(ISD::STORE, MVT::v2f64, Promote);
  AddPromotedToType(ISD::STORE, MVT::v2f64, MVT::v4i32);

  setOperationAction(ISD::STORE, MVT::v3i64, Promote);
  AddPromotedToType(ISD::STORE, MVT::v3i64, MVT::v6i32);

  setOperationAction(ISD::STORE, MVT::v3f64, Promote);
  AddPromotedToType(ISD::STORE, MVT::v3f64, MVT::v6i32);

  setOperationAction(ISD::STORE, MVT::v4i64, Promote);
  AddPromotedToType(ISD::STORE, MVT::v4i64, MVT::v8i32);

  setOperationAction(ISD::STORE, MVT::v4f64, Promote);
  AddPromotedToType(ISD::STORE, MVT::v4f64, MVT::v8i32);

  setOperationAction(ISD::STORE, MVT::v8i64, Promote);
  AddPromotedToType(ISD::STORE, MVT::v8i64, MVT::v16i32);

  setOperationAction(ISD::STORE, MVT::v8f64, Promote);
  AddPromotedToType(ISD::STORE, MVT::v8f64, MVT::v16i32);

  setOperationAction(ISD::STORE, MVT::v16i64, Promote);
  AddPromotedToType(ISD::STORE, MVT::v16i64, MVT::v32i32);

  setOperationAction(ISD::STORE, MVT::v16f64, Promote);
  AddPromotedToType(ISD::STORE, MVT::v16f64, MVT::v32i32);

  setTruncStoreAction(MVT::i64, MVT::i1, Expand);
  setTruncStoreAction(MVT::i64, MVT::i8, Expand);
  setTruncStoreAction(MVT::i64, MVT::i16, Expand);
  setTruncStoreAction(MVT::i64, MVT::i32, Expand);

  setTruncStoreAction(MVT::v2i64, MVT::v2i1, Expand);
  setTruncStoreAction(MVT::v2i64, MVT::v2i8, Expand);
  setTruncStoreAction(MVT::v2i64, MVT::v2i16, Expand);
  setTruncStoreAction(MVT::v2i64, MVT::v2i32, Expand);

  setTruncStoreAction(MVT::f32, MVT::f16, Expand);
  setTruncStoreAction(MVT::v2f32, MVT::v2f16, Expand);
  setTruncStoreAction(MVT::v3f32, MVT::v3f16, Expand);
  setTruncStoreAction(MVT::v4f32, MVT::v4f16, Expand);
  setTruncStoreAction(MVT::v8f32, MVT::v8f16, Expand);
  setTruncStoreAction(MVT::v16f32, MVT::v16f16, Expand);
  setTruncStoreAction(MVT::v32f32, MVT::v32f16, Expand);

  setTruncStoreAction(MVT::f64, MVT::f16, Expand);
  setTruncStoreAction(MVT::f64, MVT::f32, Expand);

  setTruncStoreAction(MVT::v2f64, MVT::v2f32, Expand);
  setTruncStoreAction(MVT::v2f64, MVT::v2f16, Expand);

  setTruncStoreAction(MVT::v3i64, MVT::v3i32, Expand);
  setTruncStoreAction(MVT::v3i64, MVT::v3i16, Expand);
  setTruncStoreAction(MVT::v3f64, MVT::v3f32, Expand);
  setTruncStoreAction(MVT::v3f64, MVT::v3f16, Expand);

  setTruncStoreAction(MVT::v4i64, MVT::v4i32, Expand);
  setTruncStoreAction(MVT::v4i64, MVT::v4i16, Expand);
  setTruncStoreAction(MVT::v4f64, MVT::v4f32, Expand);
  setTruncStoreAction(MVT::v4f64, MVT::v4f16, Expand);

  setTruncStoreAction(MVT::v8f64, MVT::v8f32, Expand);
  setTruncStoreAction(MVT::v8f64, MVT::v8f16, Expand);

  setTruncStoreAction(MVT::v16f64, MVT::v16f32, Expand);
  setTruncStoreAction(MVT::v16f64, MVT::v16f16, Expand);
  setTruncStoreAction(MVT::v16i64, MVT::v16i16, Expand);
  setTruncStoreAction(MVT::v16i64, MVT::v16i16, Expand);
  setTruncStoreAction(MVT::v16i64, MVT::v16i8, Expand);
  setTruncStoreAction(MVT::v16i64, MVT::v16i8, Expand);
  setTruncStoreAction(MVT::v16i64, MVT::v16i1, Expand);

  setOperationAction(ISD::Constant, {MVT::i32, MVT::i64}, Legal);
  setOperationAction(ISD::ConstantFP, {MVT::f32, MVT::f64}, Legal);

  setOperationAction({ISD::BR_JT, ISD::BRIND}, MVT::Other, Expand);

  // This is totally unsupported, just custom lower to produce an error.
  setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);

  // Library functions.  These default to Expand, but we have instructions
  // for them.
  setOperationAction({ISD::FCEIL, ISD::FEXP2, ISD::FPOW, ISD::FLOG2, ISD::FABS,
                      ISD::FFLOOR, ISD::FRINT, ISD::FTRUNC, ISD::FMINNUM,
                      ISD::FMAXNUM},
                     MVT::f32, Legal);

  setOperationAction(ISD::FROUND, {MVT::f32, MVT::f64}, Custom);

  setOperationAction({ISD::FLOG, ISD::FLOG10, ISD::FEXP}, MVT::f32, Custom);

  setOperationAction(ISD::FNEARBYINT, {MVT::f32, MVT::f64}, Custom);

  setOperationAction(ISD::FREM, {MVT::f16, MVT::f32, MVT::f64}, Custom);

  // Expand to fneg + fadd.
  setOperationAction(ISD::FSUB, MVT::f64, Expand);

  setOperationAction(ISD::CONCAT_VECTORS,
                     {MVT::v3i32, MVT::v3f32, MVT::v4i32, MVT::v4f32,
                      MVT::v5i32, MVT::v5f32, MVT::v6i32, MVT::v6f32,
                      MVT::v7i32, MVT::v7f32, MVT::v8i32, MVT::v8f32},
                     Custom);
  setOperationAction(
      ISD::EXTRACT_SUBVECTOR,
      {MVT::v2f16,  MVT::v2i16,  MVT::v4f16,  MVT::v4i16,  MVT::v2f32,
       MVT::v2i32,  MVT::v3f32,  MVT::v3i32,  MVT::v4f32,  MVT::v4i32,
       MVT::v5f32,  MVT::v5i32,  MVT::v6f32,  MVT::v6i32,  MVT::v7f32,
       MVT::v7i32,  MVT::v8f32,  MVT::v8i32,  MVT::v16f16, MVT::v16i16,
       MVT::v16f32, MVT::v16i32, MVT::v32f32, MVT::v32i32, MVT::v2f64,
       MVT::v2i64,  MVT::v3f64,  MVT::v3i64,  MVT::v4f64,  MVT::v4i64,
       MVT::v8f64,  MVT::v8i64,  MVT::v16f64, MVT::v16i64},
      Custom);

  setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
  setOperationAction(ISD::FP_TO_FP16, {MVT::f64, MVT::f32}, Custom);

  const MVT ScalarIntVTs[] = { MVT::i32, MVT::i64 };
  for (MVT VT : ScalarIntVTs) {
    // These should use [SU]DIVREM, so set them to expand
    setOperationAction({ISD::SDIV, ISD::UDIV, ISD::SREM, ISD::UREM}, VT,
                       Expand);

    // GPU does not have divrem function for signed or unsigned.
    setOperationAction({ISD::SDIVREM, ISD::UDIVREM}, VT, Custom);

    // GPU does not have [S|U]MUL_LOHI functions as a single instruction.
    setOperationAction({ISD::SMUL_LOHI, ISD::UMUL_LOHI}, VT, Expand);

    setOperationAction({ISD::BSWAP, ISD::CTTZ, ISD::CTLZ}, VT, Expand);

    // AMDGPU uses ADDC/SUBC/ADDE/SUBE
    setOperationAction({ISD::ADDC, ISD::SUBC, ISD::ADDE, ISD::SUBE}, VT, Legal);
  }

  // The hardware supports 32-bit FSHR, but not FSHL.
  setOperationAction(ISD::FSHR, MVT::i32, Legal);

  // The hardware supports 32-bit ROTR, but not ROTL.
  setOperationAction(ISD::ROTL, {MVT::i32, MVT::i64}, Expand);
  setOperationAction(ISD::ROTR, MVT::i64, Expand);

  setOperationAction({ISD::MULHU, ISD::MULHS}, MVT::i16, Expand);

  setOperationAction({ISD::MUL, ISD::MULHU, ISD::MULHS}, MVT::i64, Expand);
  setOperationAction(
      {ISD::UINT_TO_FP, ISD::SINT_TO_FP, ISD::FP_TO_SINT, ISD::FP_TO_UINT},
      MVT::i64, Custom);
  setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);

  setOperationAction({ISD::SMIN, ISD::UMIN, ISD::SMAX, ISD::UMAX}, MVT::i32,
                     Legal);

  setOperationAction(
      {ISD::CTTZ, ISD::CTTZ_ZERO_UNDEF, ISD::CTLZ, ISD::CTLZ_ZERO_UNDEF},
      MVT::i64, Custom);

  static const MVT::SimpleValueType VectorIntTypes[] = {
      MVT::v2i32, MVT::v3i32, MVT::v4i32, MVT::v5i32, MVT::v6i32, MVT::v7i32};

  for (MVT VT : VectorIntTypes) {
    // Expand the following operations for the current type by default.
    setOperationAction({ISD::ADD,        ISD::AND,     ISD::FP_TO_SINT,
                        ISD::FP_TO_UINT, ISD::MUL,     ISD::MULHU,
                        ISD::MULHS,      ISD::OR,      ISD::SHL,
                        ISD::SRA,        ISD::SRL,     ISD::ROTL,
                        ISD::ROTR,       ISD::SUB,     ISD::SINT_TO_FP,
                        ISD::UINT_TO_FP, ISD::SDIV,    ISD::UDIV,
                        ISD::SREM,       ISD::UREM,    ISD::SMUL_LOHI,
                        ISD::UMUL_LOHI,  ISD::SDIVREM, ISD::UDIVREM,
                        ISD::SELECT,     ISD::VSELECT, ISD::SELECT_CC,
                        ISD::XOR,        ISD::BSWAP,   ISD::CTPOP,
                        ISD::CTTZ,       ISD::CTLZ,    ISD::VECTOR_SHUFFLE,
                        ISD::SETCC},
                       VT, Expand);
  }

  static const MVT::SimpleValueType FloatVectorTypes[] = {
      MVT::v2f32, MVT::v3f32, MVT::v4f32, MVT::v5f32, MVT::v6f32, MVT::v7f32};

  for (MVT VT : FloatVectorTypes) {
    setOperationAction(
        {ISD::FABS,    ISD::FMINNUM,      ISD::FMAXNUM,   ISD::FADD,
         ISD::FCEIL,   ISD::FCOS,         ISD::FDIV,      ISD::FEXP2,
         ISD::FEXP,    ISD::FLOG2,        ISD::FREM,      ISD::FLOG,
         ISD::FLOG10,  ISD::FPOW,         ISD::FFLOOR,    ISD::FTRUNC,
         ISD::FMUL,    ISD::FMA,          ISD::FRINT,     ISD::FNEARBYINT,
         ISD::FSQRT,   ISD::FSIN,         ISD::FSUB,      ISD::FNEG,
         ISD::VSELECT, ISD::SELECT_CC,    ISD::FCOPYSIGN, ISD::VECTOR_SHUFFLE,
         ISD::SETCC,   ISD::FCANONICALIZE},
        VT, Expand);
  }

  // This causes using an unrolled select operation rather than expansion with
  // bit operations. This is in general better, but the alternative using BFI
  // instructions may be better if the select sources are SGPRs.
  setOperationAction(ISD::SELECT, MVT::v2f32, Promote);
  AddPromotedToType(ISD::SELECT, MVT::v2f32, MVT::v2i32);

  setOperationAction(ISD::SELECT, MVT::v3f32, Promote);
  AddPromotedToType(ISD::SELECT, MVT::v3f32, MVT::v3i32);

  setOperationAction(ISD::SELECT, MVT::v4f32, Promote);
  AddPromotedToType(ISD::SELECT, MVT::v4f32, MVT::v4i32);

  setOperationAction(ISD::SELECT, MVT::v5f32, Promote);
  AddPromotedToType(ISD::SELECT, MVT::v5f32, MVT::v5i32);

  setOperationAction(ISD::SELECT, MVT::v6f32, Promote);
  AddPromotedToType(ISD::SELECT, MVT::v6f32, MVT::v6i32);

  setOperationAction(ISD::SELECT, MVT::v7f32, Promote);
  AddPromotedToType(ISD::SELECT, MVT::v7f32, MVT::v7i32);

  // There are no libcalls of any kind.
  for (int I = 0; I < RTLIB::UNKNOWN_LIBCALL; ++I)
    setLibcallName(static_cast<RTLIB::Libcall>(I), nullptr);

  setSchedulingPreference(Sched::RegPressure);
  setJumpIsExpensive(true);

  // FIXME: This is only partially true. If we have to do vector compares, any
  // SGPR pair can be a condition register. If we have a uniform condition, we
  // are better off doing SALU operations, where there is only one SCC. For now,
  // we don't have a way of knowing during instruction selection if a condition
  // will be uniform and we always use vector compares. Assume we are using
  // vector compares until that is fixed.
  setHasMultipleConditionRegisters(true);

  setMinCmpXchgSizeInBits(32);
  setSupportsUnalignedAtomics(false);

  PredictableSelectIsExpensive = false;

  // We want to find all load dependencies for long chains of stores to enable
  // merging into very wide vectors. The problem is with vectors with > 4
  // elements. MergeConsecutiveStores will attempt to merge these because x8/x16
  // vectors are a legal type, even though we have to split the loads
  // usually. When we can more precisely specify load legality per address
  // space, we should be able to make FindBetterChain/MergeConsecutiveStores
  // smarter so that they can figure out what to do in 2 iterations without all
  // N > 4 stores on the same chain.
  GatherAllAliasesMaxDepth = 16;

  // memcpy/memmove/memset are expanded in the IR, so we shouldn't need to worry
  // about these during lowering.
  MaxStoresPerMemcpy  = 0xffffffff;
  MaxStoresPerMemmove = 0xffffffff;
  MaxStoresPerMemset  = 0xffffffff;

  // The expansion for 64-bit division is enormous.
  if (AMDGPUBypassSlowDiv)
    addBypassSlowDiv(64, 32);

  setTargetDAGCombine({ISD::BITCAST,    ISD::SHL,
                       ISD::SRA,        ISD::SRL,
                       ISD::TRUNCATE,   ISD::MUL,
                       ISD::SMUL_LOHI,  ISD::UMUL_LOHI,
                       ISD::MULHU,      ISD::MULHS,
                       ISD::SELECT,     ISD::SELECT_CC,
                       ISD::STORE,      ISD::FADD,
                       ISD::FSUB,       ISD::FNEG,
                       ISD::FABS,       ISD::AssertZext,
                       ISD::AssertSext, ISD::INTRINSIC_WO_CHAIN});
}

bool AMDGPUTargetLowering::mayIgnoreSignedZero(SDValue Op) const {
  if (getTargetMachine().Options.NoSignedZerosFPMath)
    return true;

  const auto Flags = Op.getNode()->getFlags();
  if (Flags.hasNoSignedZeros())
    return true;

  return false;
}

//===----------------------------------------------------------------------===//
// Target Information
//===----------------------------------------------------------------------===//

LLVM_READNONE
static bool fnegFoldsIntoOp(unsigned Opc) {
  switch (Opc) {
  case ISD::FADD:
  case ISD::FSUB:
  case ISD::FMUL:
  case ISD::FMA:
  case ISD::FMAD:
  case ISD::FMINNUM:
  case ISD::FMAXNUM:
  case ISD::FMINNUM_IEEE:
  case ISD::FMAXNUM_IEEE:
  case ISD::FSIN:
  case ISD::FTRUNC:
  case ISD::FRINT:
  case ISD::FNEARBYINT:
  case ISD::FCANONICALIZE:
  case AMDGPUISD::RCP:
  case AMDGPUISD::RCP_LEGACY:
  case AMDGPUISD::RCP_IFLAG:
  case AMDGPUISD::SIN_HW:
  case AMDGPUISD::FMUL_LEGACY:
  case AMDGPUISD::FMIN_LEGACY:
  case AMDGPUISD::FMAX_LEGACY:
  case AMDGPUISD::FMED3:
    // TODO: handle llvm.amdgcn.fma.legacy
    return true;
  default:
    return false;
  }
}

/// \p returns true if the operation will definitely need to use a 64-bit
/// encoding, and thus will use a VOP3 encoding regardless of the source
/// modifiers.
LLVM_READONLY
static bool opMustUseVOP3Encoding(const SDNode *N, MVT VT) {
  return N->getNumOperands() > 2 || VT == MVT::f64;
}

// Most FP instructions support source modifiers, but this could be refined
// slightly.
LLVM_READONLY
static bool hasSourceMods(const SDNode *N) {
  if (isa<MemSDNode>(N))
    return false;

  switch (N->getOpcode()) {
  case ISD::CopyToReg:
  case ISD::SELECT:
  case ISD::FDIV:
  case ISD::FREM:
  case ISD::INLINEASM:
  case ISD::INLINEASM_BR:
  case AMDGPUISD::DIV_SCALE:
  case ISD::INTRINSIC_W_CHAIN:

  // TODO: Should really be looking at the users of the bitcast. These are
  // problematic because bitcasts are used to legalize all stores to integer
  // types.
  case ISD::BITCAST:
    return false;
  case ISD::INTRINSIC_WO_CHAIN: {
    switch (cast<ConstantSDNode>(N->getOperand(0))->getZExtValue()) {
    case Intrinsic::amdgcn_interp_p1:
    case Intrinsic::amdgcn_interp_p2:
    case Intrinsic::amdgcn_interp_mov:
    case Intrinsic::amdgcn_interp_p1_f16:
    case Intrinsic::amdgcn_interp_p2_f16:
      return false;
    default:
      return true;
    }
  }
  default:
    return true;
  }
}

bool AMDGPUTargetLowering::allUsesHaveSourceMods(const SDNode *N,
                                                 unsigned CostThreshold) {
  // Some users (such as 3-operand FMA/MAD) must use a VOP3 encoding, and thus
  // it is truly free to use a source modifier in all cases. If there are
  // multiple users but for each one will necessitate using VOP3, there will be
  // a code size increase. Try to avoid increasing code size unless we know it
  // will save on the instruction count.
  unsigned NumMayIncreaseSize = 0;
  MVT VT = N->getValueType(0).getScalarType().getSimpleVT();

  // XXX - Should this limit number of uses to check?
  for (const SDNode *U : N->uses()) {
    if (!hasSourceMods(U))
      return false;

    if (!opMustUseVOP3Encoding(U, VT)) {
      if (++NumMayIncreaseSize > CostThreshold)
        return false;
    }
  }

  return true;
}

EVT AMDGPUTargetLowering::getTypeForExtReturn(LLVMContext &Context, EVT VT,
                                              ISD::NodeType ExtendKind) const {
  assert(!VT.isVector() && "only scalar expected");

  // Round to the next multiple of 32-bits.
  unsigned Size = VT.getSizeInBits();
  if (Size <= 32)
    return MVT::i32;
  return EVT::getIntegerVT(Context, 32 * ((Size + 31) / 32));
}

MVT AMDGPUTargetLowering::getVectorIdxTy(const DataLayout &) const {
  return MVT::i32;
}

bool AMDGPUTargetLowering::isSelectSupported(SelectSupportKind SelType) const {
  return true;
}

// The backend supports 32 and 64 bit floating point immediates.
// FIXME: Why are we reporting vectors of FP immediates as legal?
bool AMDGPUTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT,
                                        bool ForCodeSize) const {
  EVT ScalarVT = VT.getScalarType();
  return (ScalarVT == MVT::f32 || ScalarVT == MVT::f64 ||
         (ScalarVT == MVT::f16 && Subtarget->has16BitInsts()));
}

// We don't want to shrink f64 / f32 constants.
bool AMDGPUTargetLowering::ShouldShrinkFPConstant(EVT VT) const {
  EVT ScalarVT = VT.getScalarType();
  return (ScalarVT != MVT::f32 && ScalarVT != MVT::f64);
}

bool AMDGPUTargetLowering::shouldReduceLoadWidth(SDNode *N,
                                                 ISD::LoadExtType ExtTy,
                                                 EVT NewVT) const {
  // TODO: This may be worth removing. Check regression tests for diffs.
  if (!TargetLoweringBase::shouldReduceLoadWidth(N, ExtTy, NewVT))
    return false;

  unsigned NewSize = NewVT.getStoreSizeInBits();

  // If we are reducing to a 32-bit load or a smaller multi-dword load,
  // this is always better.
  if (NewSize >= 32)
    return true;

  EVT OldVT = N->getValueType(0);
  unsigned OldSize = OldVT.getStoreSizeInBits();

  MemSDNode *MN = cast<MemSDNode>(N);
  unsigned AS = MN->getAddressSpace();
  // Do not shrink an aligned scalar load to sub-dword.
  // Scalar engine cannot do sub-dword loads.
  if (OldSize >= 32 && NewSize < 32 && MN->getAlign() >= Align(4) &&
      (AS == AMDGPUAS::CONSTANT_ADDRESS ||
       AS == AMDGPUAS::CONSTANT_ADDRESS_32BIT ||
       (isa<LoadSDNode>(N) && AS == AMDGPUAS::GLOBAL_ADDRESS &&
        MN->isInvariant())) &&
      AMDGPUInstrInfo::isUniformMMO(MN->getMemOperand()))
    return false;

  // Don't produce extloads from sub 32-bit types. SI doesn't have scalar
  // extloads, so doing one requires using a buffer_load. In cases where we
  // still couldn't use a scalar load, using the wider load shouldn't really
  // hurt anything.

  // If the old size already had to be an extload, there's no harm in continuing
  // to reduce the width.
  return (OldSize < 32);
}

bool AMDGPUTargetLowering::isLoadBitCastBeneficial(EVT LoadTy, EVT CastTy,
                                                   const SelectionDAG &DAG,
                                                   const MachineMemOperand &MMO) const {

  assert(LoadTy.getSizeInBits() == CastTy.getSizeInBits());

  if (LoadTy.getScalarType() == MVT::i32)
    return false;

  unsigned LScalarSize = LoadTy.getScalarSizeInBits();
  unsigned CastScalarSize = CastTy.getScalarSizeInBits();

  if ((LScalarSize >= CastScalarSize) && (CastScalarSize < 32))
    return false;

  bool Fast = false;
  return allowsMemoryAccessForAlignment(*DAG.getContext(), DAG.getDataLayout(),
                                        CastTy, MMO, &Fast) &&
         Fast;
}

// SI+ has instructions for cttz / ctlz for 32-bit values. This is probably also
// profitable with the expansion for 64-bit since it's generally good to
// speculate things.
// FIXME: These should really have the size as a parameter.
bool AMDGPUTargetLowering::isCheapToSpeculateCttz() const {
  return true;
}

bool AMDGPUTargetLowering::isCheapToSpeculateCtlz() const {
  return true;
}

bool AMDGPUTargetLowering::isSDNodeAlwaysUniform(const SDNode *N) const {
  switch (N->getOpcode()) {
  case ISD::EntryToken:
  case ISD::TokenFactor:
    return true;
  case ISD::INTRINSIC_WO_CHAIN: {
    unsigned IntrID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
    switch (IntrID) {
    case Intrinsic::amdgcn_readfirstlane:
    case Intrinsic::amdgcn_readlane:
      return true;
    }
    return false;
  }
  case ISD::LOAD:
    if (cast<LoadSDNode>(N)->getMemOperand()->getAddrSpace() ==
        AMDGPUAS::CONSTANT_ADDRESS_32BIT)
      return true;
    return false;
  case AMDGPUISD::SETCC: // ballot-style instruction
    return true;
  }
  return false;
}

SDValue AMDGPUTargetLowering::getNegatedExpression(
    SDValue Op, SelectionDAG &DAG, bool LegalOperations, bool ForCodeSize,
    NegatibleCost &Cost, unsigned Depth) const {

  switch (Op.getOpcode()) {
  case ISD::FMA:
  case ISD::FMAD: {
    // Negating a fma is not free if it has users without source mods.
    if (!allUsesHaveSourceMods(Op.getNode()))
      return SDValue();
    break;
  }
  default:
    break;
  }

  return TargetLowering::getNegatedExpression(Op, DAG, LegalOperations,
                                              ForCodeSize, Cost, Depth);
}

//===---------------------------------------------------------------------===//
// Target Properties
//===---------------------------------------------------------------------===//

bool AMDGPUTargetLowering::isFAbsFree(EVT VT) const {
  assert(VT.isFloatingPoint());

  // Packed operations do not have a fabs modifier.
  return VT == MVT::f32 || VT == MVT::f64 ||
         (Subtarget->has16BitInsts() && VT == MVT::f16);
}

bool AMDGPUTargetLowering::isFNegFree(EVT VT) const {
  assert(VT.isFloatingPoint());
  // Report this based on the end legalized type.
  VT = VT.getScalarType();
  return VT == MVT::f32 || VT == MVT::f64 || VT == MVT::f16;
}

bool AMDGPUTargetLowering:: storeOfVectorConstantIsCheap(EVT MemVT,
                                                         unsigned NumElem,
                                                         unsigned AS) const {
  return true;
}

bool AMDGPUTargetLowering::aggressivelyPreferBuildVectorSources(EVT VecVT) const {
  // There are few operations which truly have vector input operands. Any vector
  // operation is going to involve operations on each component, and a
  // build_vector will be a copy per element, so it always makes sense to use a
  // build_vector input in place of the extracted element to avoid a copy into a
  // super register.
  //
  // We should probably only do this if all users are extracts only, but this
  // should be the common case.
  return true;
}

bool AMDGPUTargetLowering::isTruncateFree(EVT Source, EVT Dest) const {
  // Truncate is just accessing a subregister.

  unsigned SrcSize = Source.getSizeInBits();
  unsigned DestSize = Dest.getSizeInBits();

  return DestSize < SrcSize && DestSize % 32 == 0 ;
}

bool AMDGPUTargetLowering::isTruncateFree(Type *Source, Type *Dest) const {
  // Truncate is just accessing a subregister.

  unsigned SrcSize = Source->getScalarSizeInBits();
  unsigned DestSize = Dest->getScalarSizeInBits();

  if (DestSize== 16 && Subtarget->has16BitInsts())
    return SrcSize >= 32;

  return DestSize < SrcSize && DestSize % 32 == 0;
}

bool AMDGPUTargetLowering::isZExtFree(Type *Src, Type *Dest) const {
  unsigned SrcSize = Src->getScalarSizeInBits();
  unsigned DestSize = Dest->getScalarSizeInBits();

  if (SrcSize == 16 && Subtarget->has16BitInsts())
    return DestSize >= 32;

  return SrcSize == 32 && DestSize == 64;
}

bool AMDGPUTargetLowering::isZExtFree(EVT Src, EVT Dest) const {
  // Any register load of a 64-bit value really requires 2 32-bit moves. For all
  // practical purposes, the extra mov 0 to load a 64-bit is free.  As used,
  // this will enable reducing 64-bit operations the 32-bit, which is always
  // good.

  if (Src == MVT::i16)
    return Dest == MVT::i32 ||Dest == MVT::i64 ;

  return Src == MVT::i32 && Dest == MVT::i64;
}

bool AMDGPUTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
  return isZExtFree(Val.getValueType(), VT2);
}

bool AMDGPUTargetLowering::isNarrowingProfitable(EVT SrcVT, EVT DestVT) const {
  // There aren't really 64-bit registers, but pairs of 32-bit ones and only a
  // limited number of native 64-bit operations. Shrinking an operation to fit
  // in a single 32-bit register should always be helpful. As currently used,
  // this is much less general than the name suggests, and is only used in
  // places trying to reduce the sizes of loads. Shrinking loads to < 32-bits is
  // not profitable, and may actually be harmful.
  return SrcVT.getSizeInBits() > 32 && DestVT.getSizeInBits() == 32;
}

//===---------------------------------------------------------------------===//
// TargetLowering Callbacks
//===---------------------------------------------------------------------===//

CCAssignFn *AMDGPUCallLowering::CCAssignFnForCall(CallingConv::ID CC,
                                                  bool IsVarArg) {
  switch (CC) {
  case CallingConv::AMDGPU_VS:
  case CallingConv::AMDGPU_GS:
  case CallingConv::AMDGPU_PS:
  case CallingConv::AMDGPU_CS:
  case CallingConv::AMDGPU_HS:
  case CallingConv::AMDGPU_ES:
  case CallingConv::AMDGPU_LS:
    return CC_AMDGPU;
  case CallingConv::C:
  case CallingConv::Fast:
  case CallingConv::Cold:
    return CC_AMDGPU_Func;
  case CallingConv::AMDGPU_Gfx:
    return CC_SI_Gfx;
  case CallingConv::AMDGPU_KERNEL:
  case CallingConv::SPIR_KERNEL:
  default:
    report_fatal_error("Unsupported calling convention for call");
  }
}

CCAssignFn *AMDGPUCallLowering::CCAssignFnForReturn(CallingConv::ID CC,
                                                    bool IsVarArg) {
  switch (CC) {
  case CallingConv::AMDGPU_KERNEL:
  case CallingConv::SPIR_KERNEL:
    llvm_unreachable("kernels should not be handled here");
  case CallingConv::AMDGPU_VS:
  case CallingConv::AMDGPU_GS:
  case CallingConv::AMDGPU_PS:
  case CallingConv::AMDGPU_CS:
  case CallingConv::AMDGPU_HS:
  case CallingConv::AMDGPU_ES:
  case CallingConv::AMDGPU_LS:
    return RetCC_SI_Shader;
  case CallingConv::AMDGPU_Gfx:
    return RetCC_SI_Gfx;
  case CallingConv::C:
  case CallingConv::Fast:
  case CallingConv::Cold:
    return RetCC_AMDGPU_Func;
  default:
    report_fatal_error("Unsupported calling convention.");
  }
}

/// The SelectionDAGBuilder will automatically promote function arguments
/// with illegal types.  However, this does not work for the AMDGPU targets
/// since the function arguments are stored in memory as these illegal types.
/// In order to handle this properly we need to get the original types sizes
/// from the LLVM IR Function and fixup the ISD:InputArg values before
/// passing them to AnalyzeFormalArguments()

/// When the SelectionDAGBuilder computes the Ins, it takes care of splitting
/// input values across multiple registers.  Each item in the Ins array
/// represents a single value that will be stored in registers.  Ins[x].VT is
/// the value type of the value that will be stored in the register, so
/// whatever SDNode we lower the argument to needs to be this type.
///
/// In order to correctly lower the arguments we need to know the size of each
/// argument.  Since Ins[x].VT gives us the size of the register that will
/// hold the value, we need to look at Ins[x].ArgVT to see the 'real' type
/// for the original function argument so that we can deduce the correct memory
/// type to use for Ins[x].  In most cases the correct memory type will be
/// Ins[x].ArgVT.  However, this will not always be the case.  If, for example,
/// we have a kernel argument of type v8i8, this argument will be split into
/// 8 parts and each part will be represented by its own item in the Ins array.
/// For each part the Ins[x].ArgVT will be the v8i8, which is the full type of
/// the argument before it was split.  From this, we deduce that the memory type
/// for each individual part is i8.  We pass the memory type as LocVT to the
/// calling convention analysis function and the register type (Ins[x].VT) as
/// the ValVT.
void AMDGPUTargetLowering::analyzeFormalArgumentsCompute(
  CCState &State,
  const SmallVectorImpl<ISD::InputArg> &Ins) const {
  const MachineFunction &MF = State.getMachineFunction();
  const Function &Fn = MF.getFunction();
  LLVMContext &Ctx = Fn.getParent()->getContext();
  const AMDGPUSubtarget &ST = AMDGPUSubtarget::get(MF);
  const unsigned ExplicitOffset = ST.getExplicitKernelArgOffset(Fn);
  CallingConv::ID CC = Fn.getCallingConv();

  Align MaxAlign = Align(1);
  uint64_t ExplicitArgOffset = 0;
  const DataLayout &DL = Fn.getParent()->getDataLayout();

  unsigned InIndex = 0;

  for (const Argument &Arg : Fn.args()) {
    const bool IsByRef = Arg.hasByRefAttr();
    Type *BaseArgTy = Arg.getType();
    Type *MemArgTy = IsByRef ? Arg.getParamByRefType() : BaseArgTy;
    Align Alignment = DL.getValueOrABITypeAlignment(
        IsByRef ? Arg.getParamAlign() : None, MemArgTy);
    MaxAlign = std::max(Alignment, MaxAlign);
    uint64_t AllocSize = DL.getTypeAllocSize(MemArgTy);

    uint64_t ArgOffset = alignTo(ExplicitArgOffset, Alignment) + ExplicitOffset;
    ExplicitArgOffset = alignTo(ExplicitArgOffset, Alignment) + AllocSize;

    // We're basically throwing away everything passed into us and starting over
    // to get accurate in-memory offsets. The "PartOffset" is completely useless
    // to us as computed in Ins.
    //
    // We also need to figure out what type legalization is trying to do to get
    // the correct memory offsets.

    SmallVector<EVT, 16> ValueVTs;
    SmallVector<uint64_t, 16> Offsets;
    ComputeValueVTs(*this, DL, BaseArgTy, ValueVTs, &Offsets, ArgOffset);

    for (unsigned Value = 0, NumValues = ValueVTs.size();
         Value != NumValues; ++Value) {
      uint64_t BasePartOffset = Offsets[Value];

      EVT ArgVT = ValueVTs[Value];
      EVT MemVT = ArgVT;
      MVT RegisterVT = getRegisterTypeForCallingConv(Ctx, CC, ArgVT);
      unsigned NumRegs = getNumRegistersForCallingConv(Ctx, CC, ArgVT);

      if (NumRegs == 1) {
        // This argument is not split, so the IR type is the memory type.
        if (ArgVT.isExtended()) {
          // We have an extended type, like i24, so we should just use the
          // register type.
          MemVT = RegisterVT;
        } else {
          MemVT = ArgVT;
        }
      } else if (ArgVT.isVector() && RegisterVT.isVector() &&
                 ArgVT.getScalarType() == RegisterVT.getScalarType()) {
        assert(ArgVT.getVectorNumElements() > RegisterVT.getVectorNumElements());
        // We have a vector value which has been split into a vector with
        // the same scalar type, but fewer elements.  This should handle
        // all the floating-point vector types.
        MemVT = RegisterVT;
      } else if (ArgVT.isVector() &&
                 ArgVT.getVectorNumElements() == NumRegs) {
        // This arg has been split so that each element is stored in a separate
        // register.
        MemVT = ArgVT.getScalarType();
      } else if (ArgVT.isExtended()) {
        // We have an extended type, like i65.
        MemVT = RegisterVT;
      } else {
        unsigned MemoryBits = ArgVT.getStoreSizeInBits() / NumRegs;
        assert(ArgVT.getStoreSizeInBits() % NumRegs == 0);
        if (RegisterVT.isInteger()) {
          MemVT = EVT::getIntegerVT(State.getContext(), MemoryBits);
        } else if (RegisterVT.isVector()) {
          assert(!RegisterVT.getScalarType().isFloatingPoint());
          unsigned NumElements = RegisterVT.getVectorNumElements();
          assert(MemoryBits % NumElements == 0);
          // This vector type has been split into another vector type with
          // a different elements size.
          EVT ScalarVT = EVT::getIntegerVT(State.getContext(),
                                           MemoryBits / NumElements);
          MemVT = EVT::getVectorVT(State.getContext(), ScalarVT, NumElements);
        } else {
          llvm_unreachable("cannot deduce memory type.");
        }
      }

      // Convert one element vectors to scalar.
      if (MemVT.isVector() && MemVT.getVectorNumElements() == 1)
        MemVT = MemVT.getScalarType();

      // Round up vec3/vec5 argument.
      if (MemVT.isVector() && !MemVT.isPow2VectorType()) {
        assert(MemVT.getVectorNumElements() == 3 ||
               MemVT.getVectorNumElements() == 5);
        MemVT = MemVT.getPow2VectorType(State.getContext());
      } else if (!MemVT.isSimple() && !MemVT.isVector()) {
        MemVT = MemVT.getRoundIntegerType(State.getContext());
      }

      unsigned PartOffset = 0;
      for (unsigned i = 0; i != NumRegs; ++i) {
        State.addLoc(CCValAssign::getCustomMem(InIndex++, RegisterVT,
                                               BasePartOffset + PartOffset,
                                               MemVT.getSimpleVT(),
                                               CCValAssign::Full));
        PartOffset += MemVT.getStoreSize();
      }
    }
  }
}

SDValue AMDGPUTargetLowering::LowerReturn(
  SDValue Chain, CallingConv::ID CallConv,
  bool isVarArg,
  const SmallVectorImpl<ISD::OutputArg> &Outs,
  const SmallVectorImpl<SDValue> &OutVals,
  const SDLoc &DL, SelectionDAG &DAG) const {
  // FIXME: Fails for r600 tests
  //assert(!isVarArg && Outs.empty() && OutVals.empty() &&
  // "wave terminate should not have return values");
  return DAG.getNode(AMDGPUISD::ENDPGM, DL, MVT::Other, Chain);
}

//===---------------------------------------------------------------------===//
// Target specific lowering
//===---------------------------------------------------------------------===//

/// Selects the correct CCAssignFn for a given CallingConvention value.
CCAssignFn *AMDGPUTargetLowering::CCAssignFnForCall(CallingConv::ID CC,
                                                    bool IsVarArg) {
  return AMDGPUCallLowering::CCAssignFnForCall(CC, IsVarArg);
}

CCAssignFn *AMDGPUTargetLowering::CCAssignFnForReturn(CallingConv::ID CC,
                                                      bool IsVarArg) {
  return AMDGPUCallLowering::CCAssignFnForReturn(CC, IsVarArg);
}

SDValue AMDGPUTargetLowering::addTokenForArgument(SDValue Chain,
                                                  SelectionDAG &DAG,
                                                  MachineFrameInfo &MFI,
                                                  int ClobberedFI) const {
  SmallVector<SDValue, 8> ArgChains;
  int64_t FirstByte = MFI.getObjectOffset(ClobberedFI);
  int64_t LastByte = FirstByte + MFI.getObjectSize(ClobberedFI) - 1;

  // Include the original chain at the beginning of the list. When this is
  // used by target LowerCall hooks, this helps legalize find the
  // CALLSEQ_BEGIN node.
  ArgChains.push_back(Chain);

  // Add a chain value for each stack argument corresponding
  for (SDNode *U : DAG.getEntryNode().getNode()->uses()) {
    if (LoadSDNode *L = dyn_cast<LoadSDNode>(U)) {
      if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(L->getBasePtr())) {
        if (FI->getIndex() < 0) {
          int64_t InFirstByte = MFI.getObjectOffset(FI->getIndex());
          int64_t InLastByte = InFirstByte;
          InLastByte += MFI.getObjectSize(FI->getIndex()) - 1;

          if ((InFirstByte <= FirstByte && FirstByte <= InLastByte) ||
              (FirstByte <= InFirstByte && InFirstByte <= LastByte))
            ArgChains.push_back(SDValue(L, 1));
        }
      }
    }
  }

  // Build a tokenfactor for all the chains.
  return DAG.getNode(ISD::TokenFactor, SDLoc(Chain), MVT::Other, ArgChains);
}

SDValue AMDGPUTargetLowering::lowerUnhandledCall(CallLoweringInfo &CLI,
                                                 SmallVectorImpl<SDValue> &InVals,
                                                 StringRef Reason) const {
  SDValue Callee = CLI.Callee;
  SelectionDAG &DAG = CLI.DAG;

  const Function &Fn = DAG.getMachineFunction().getFunction();

  StringRef FuncName("<unknown>");

  if (const ExternalSymbolSDNode *G = dyn_cast<ExternalSymbolSDNode>(Callee))
    FuncName = G->getSymbol();
  else if (const GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
    FuncName = G->getGlobal()->getName();

  DiagnosticInfoUnsupported NoCalls(
    Fn, Reason + FuncName, CLI.DL.getDebugLoc());
  DAG.getContext()->diagnose(NoCalls);

  if (!CLI.IsTailCall) {
    for (unsigned I = 0, E = CLI.Ins.size(); I != E; ++I)
      InVals.push_back(DAG.getUNDEF(CLI.Ins[I].VT));
  }

  return DAG.getEntryNode();
}

SDValue AMDGPUTargetLowering::LowerCall(CallLoweringInfo &CLI,
                                        SmallVectorImpl<SDValue> &InVals) const {
  return lowerUnhandledCall(CLI, InVals, "unsupported call to function ");
}

SDValue AMDGPUTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
                                                      SelectionDAG &DAG) const {
  const Function &Fn = DAG.getMachineFunction().getFunction();

  DiagnosticInfoUnsupported NoDynamicAlloca(Fn, "unsupported dynamic alloca",
                                            SDLoc(Op).getDebugLoc());
  DAG.getContext()->diagnose(NoDynamicAlloca);
  auto Ops = {DAG.getConstant(0, SDLoc(), Op.getValueType()), Op.getOperand(0)};
  return DAG.getMergeValues(Ops, SDLoc());
}

SDValue AMDGPUTargetLowering::LowerOperation(SDValue Op,
                                             SelectionDAG &DAG) const {
  switch (Op.getOpcode()) {
  default:
    Op->print(errs(), &DAG);
    llvm_unreachable("Custom lowering code for this "
                     "instruction is not implemented yet!");
    break;
  case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op, DAG);
  case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
  case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG);
  case ISD::UDIVREM: return LowerUDIVREM(Op, DAG);
  case ISD::SDIVREM: return LowerSDIVREM(Op, DAG);
  case ISD::FREM: return LowerFREM(Op, DAG);
  case ISD::FCEIL: return LowerFCEIL(Op, DAG);
  case ISD::FTRUNC: return LowerFTRUNC(Op, DAG);
  case ISD::FRINT: return LowerFRINT(Op, DAG);
  case ISD::FNEARBYINT: return LowerFNEARBYINT(Op, DAG);
  case ISD::FROUND: return LowerFROUND(Op, DAG);
  case ISD::FFLOOR: return LowerFFLOOR(Op, DAG);
  case ISD::FLOG:
    return LowerFLOG(Op, DAG, numbers::ln2f);
  case ISD::FLOG10:
    return LowerFLOG(Op, DAG, numbers::ln2f / numbers::ln10f);
  case ISD::FEXP:
    return lowerFEXP(Op, DAG);
  case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
  case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
  case ISD::FP_TO_FP16: return LowerFP_TO_FP16(Op, DAG);
  case ISD::FP_TO_SINT:
  case ISD::FP_TO_UINT:
    return LowerFP_TO_INT(Op, DAG);
  case ISD::CTTZ:
  case ISD::CTTZ_ZERO_UNDEF:
  case ISD::CTLZ:
  case ISD::CTLZ_ZERO_UNDEF:
    return LowerCTLZ_CTTZ(Op, DAG);
  case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
  }
  return Op;
}

void AMDGPUTargetLowering::ReplaceNodeResults(SDNode *N,
                                              SmallVectorImpl<SDValue> &Results,
                                              SelectionDAG &DAG) const {
  switch (N->getOpcode()) {
  case ISD::SIGN_EXTEND_INREG:
    // Different parts of legalization seem to interpret which type of
    // sign_extend_inreg is the one to check for custom lowering. The extended
    // from type is what really matters, but some places check for custom
    // lowering of the result type. This results in trying to use
    // ReplaceNodeResults to sext_in_reg to an illegal type, so we'll just do
    // nothing here and let the illegal result integer be handled normally.
    return;
  default:
    return;
  }
}

SDValue AMDGPUTargetLowering::LowerGlobalAddress(AMDGPUMachineFunction* MFI,
                                                 SDValue Op,
                                                 SelectionDAG &DAG) const {

  const DataLayout &DL = DAG.getDataLayout();
  GlobalAddressSDNode *G = cast<GlobalAddressSDNode>(Op);
  const GlobalValue *GV = G->getGlobal();

  if (G->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
      G->getAddressSpace() == AMDGPUAS::REGION_ADDRESS) {
    if (!MFI->isModuleEntryFunction() &&
        !GV->getName().equals("llvm.amdgcn.module.lds")) {
      SDLoc DL(Op);
      const Function &Fn = DAG.getMachineFunction().getFunction();
      DiagnosticInfoUnsupported BadLDSDecl(
        Fn, "local memory global used by non-kernel function",
        DL.getDebugLoc(), DS_Warning);
      DAG.getContext()->diagnose(BadLDSDecl);

      // We currently don't have a way to correctly allocate LDS objects that
      // aren't directly associated with a kernel. We do force inlining of
      // functions that use local objects. However, if these dead functions are
      // not eliminated, we don't want a compile time error. Just emit a warning
      // and a trap, since there should be no callable path here.
      SDValue Trap = DAG.getNode(ISD::TRAP, DL, MVT::Other, DAG.getEntryNode());
      SDValue OutputChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
                                        Trap, DAG.getRoot());
      DAG.setRoot(OutputChain);
      return DAG.getUNDEF(Op.getValueType());
    }

    // XXX: What does the value of G->getOffset() mean?
    assert(G->getOffset() == 0 &&
         "Do not know what to do with an non-zero offset");

    // TODO: We could emit code to handle the initialization somewhere.
    // We ignore the initializer for now and legalize it to allow selection.
    // The initializer will anyway get errored out during assembly emission.
    unsigned Offset = MFI->allocateLDSGlobal(DL, *cast<GlobalVariable>(GV));
    return DAG.getConstant(Offset, SDLoc(Op), Op.getValueType());
  }
  return SDValue();
}

SDValue AMDGPUTargetLowering::LowerCONCAT_VECTORS(SDValue Op,
                                                  SelectionDAG &DAG) const {
  SmallVector<SDValue, 8> Args;

  EVT VT = Op.getValueType();
  if (VT == MVT::v4i16 || VT == MVT::v4f16) {
    SDLoc SL(Op);
    SDValue Lo = DAG.getNode(ISD::BITCAST, SL, MVT::i32, Op.getOperand(0));
    SDValue Hi = DAG.getNode(ISD::BITCAST, SL, MVT::i32, Op.getOperand(1));

    SDValue BV = DAG.getBuildVector(MVT::v2i32, SL, { Lo, Hi });
    return DAG.getNode(ISD::BITCAST, SL, VT, BV);
  }

  for (const SDUse &U : Op->ops())
    DAG.ExtractVectorElements(U.get(), Args);

  return DAG.getBuildVector(Op.getValueType(), SDLoc(Op), Args);
}

SDValue AMDGPUTargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
                                                     SelectionDAG &DAG) const {

  SmallVector<SDValue, 8> Args;
  unsigned Start = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
  EVT VT = Op.getValueType();
  EVT SrcVT = Op.getOperand(0).getValueType();

  // For these types, we have some TableGen patterns except if the index is 1
  if (((SrcVT == MVT::v4f16 && VT == MVT::v2f16) ||
       (SrcVT == MVT::v4i16 && VT == MVT::v2i16)) &&
      Start != 1)
    return Op;

  if (((SrcVT == MVT::v8f16 && VT == MVT::v4f16) ||
       (SrcVT == MVT::v8i16 && VT == MVT::v4i16)) &&
      (Start == 0 || Start == 4))
    return Op;

  if (((SrcVT == MVT::v16f16 && VT == MVT::v8f16) ||
       (SrcVT == MVT::v16i16 && VT == MVT::v8i16)) &&
      (Start == 0 || Start == 8))
    return Op;

  DAG.ExtractVectorElements(Op.getOperand(0), Args, Start,
                            VT.getVectorNumElements());

  return DAG.getBuildVector(Op.getValueType(), SDLoc(Op), Args);
}

/// Generate Min/Max node
SDValue AMDGPUTargetLowering::combineFMinMaxLegacy(const SDLoc &DL, EVT VT,
                                                   SDValue LHS, SDValue RHS,
                                                   SDValue True, SDValue False,
                                                   SDValue CC,
                                                   DAGCombinerInfo &DCI) const {
  if (!(LHS == True && RHS == False) && !(LHS == False && RHS == True))
    return SDValue();

  SelectionDAG &DAG = DCI.DAG;
  ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
  switch (CCOpcode) {
  case ISD::SETOEQ:
  case ISD::SETONE:
  case ISD::SETUNE:
  case ISD::SETNE:
  case ISD::SETUEQ:
  case ISD::SETEQ:
  case ISD::SETFALSE:
  case ISD::SETFALSE2:
  case ISD::SETTRUE:
  case ISD::SETTRUE2:
  case ISD::SETUO:
  case ISD::SETO:
    break;
  case ISD::SETULE:
  case ISD::SETULT: {
    if (LHS == True)
      return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, RHS, LHS);
    return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, LHS, RHS);
  }
  case ISD::SETOLE:
  case ISD::SETOLT:
  case ISD::SETLE:
  case ISD::SETLT: {
    // Ordered. Assume ordered for undefined.

    // Only do this after legalization to avoid interfering with other combines
    // which might occur.
    if (DCI.getDAGCombineLevel() < AfterLegalizeDAG &&
        !DCI.isCalledByLegalizer())
      return SDValue();

    // We need to permute the operands to get the correct NaN behavior. The
    // selected operand is the second one based on the failing compare with NaN,
    // so permute it based on the compare type the hardware uses.
    if (LHS == True)
      return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, LHS, RHS);
    return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, RHS, LHS);
  }
  case ISD::SETUGE:
  case ISD::SETUGT: {
    if (LHS == True)
      return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, RHS, LHS);
    return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, LHS, RHS);
  }
  case ISD::SETGT:
  case ISD::SETGE:
  case ISD::SETOGE:
  case ISD::SETOGT: {
    if (DCI.getDAGCombineLevel() < AfterLegalizeDAG &&
        !DCI.isCalledByLegalizer())
      return SDValue();

    if (LHS == True)
      return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, LHS, RHS);
    return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, RHS, LHS);
  }
  case ISD::SETCC_INVALID:
    llvm_unreachable("Invalid setcc condcode!");
  }
  return SDValue();
}

std::pair<SDValue, SDValue>
AMDGPUTargetLowering::split64BitValue(SDValue Op, SelectionDAG &DAG) const {
  SDLoc SL(Op);

  SDValue Vec = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Op);

  const SDValue Zero = DAG.getConstant(0, SL, MVT::i32);
  const SDValue One = DAG.getConstant(1, SL, MVT::i32);

  SDValue Lo = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Vec, Zero);
  SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Vec, One);

  return std::make_pair(Lo, Hi);
}

SDValue AMDGPUTargetLowering::getLoHalf64(SDValue Op, SelectionDAG &DAG) const {
  SDLoc SL(Op);

  SDValue Vec = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Op);
  const SDValue Zero = DAG.getConstant(0, SL, MVT::i32);
  return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Vec, Zero);
}

SDValue AMDGPUTargetLowering::getHiHalf64(SDValue Op, SelectionDAG &DAG) const {
  SDLoc SL(Op);

  SDValue Vec = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Op);
  const SDValue One = DAG.getConstant(1, SL, MVT::i32);
  return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Vec, One);
}

// Split a vector type into two parts. The first part is a power of two vector.
// The second part is whatever is left over, and is a scalar if it would
// otherwise be a 1-vector.
std::pair<EVT, EVT>
AMDGPUTargetLowering::getSplitDestVTs(const EVT &VT, SelectionDAG &DAG) const {
  EVT LoVT, HiVT;
  EVT EltVT = VT.getVectorElementType();
  unsigned NumElts = VT.getVectorNumElements();
  unsigned LoNumElts = PowerOf2Ceil((NumElts + 1) / 2);
  LoVT = EVT::getVectorVT(*DAG.getContext(), EltVT, LoNumElts);
  HiVT = NumElts - LoNumElts == 1
             ? EltVT
             : EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts - LoNumElts);
  return std::make_pair(LoVT, HiVT);
}

// Split a vector value into two parts of types LoVT and HiVT. HiVT could be
// scalar.
std::pair<SDValue, SDValue>
AMDGPUTargetLowering::splitVector(const SDValue &N, const SDLoc &DL,
                                  const EVT &LoVT, const EVT &HiVT,
                                  SelectionDAG &DAG) const {
  assert(LoVT.getVectorNumElements() +
                 (HiVT.isVector() ? HiVT.getVectorNumElements() : 1) <=
             N.getValueType().getVectorNumElements() &&
         "More vector elements requested than available!");
  SDValue Lo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, LoVT, N,
                           DAG.getVectorIdxConstant(0, DL));
  SDValue Hi = DAG.getNode(
      HiVT.isVector() ? ISD::EXTRACT_SUBVECTOR : ISD::EXTRACT_VECTOR_ELT, DL,
      HiVT, N, DAG.getVectorIdxConstant(LoVT.getVectorNumElements(), DL));
  return std::make_pair(Lo, Hi);
}

SDValue AMDGPUTargetLowering::SplitVectorLoad(const SDValue Op,
                                              SelectionDAG &DAG) const {
  LoadSDNode *Load = cast<LoadSDNode>(Op);
  EVT VT = Op.getValueType();
  SDLoc SL(Op);


  // If this is a 2 element vector, we really want to scalarize and not create
  // weird 1 element vectors.
  if (VT.getVectorNumElements() == 2) {
    SDValue Ops[2];
    std::tie(Ops[0], Ops[1]) = scalarizeVectorLoad(Load, DAG);
    return DAG.getMergeValues(Ops, SL);
  }

  SDValue BasePtr = Load->getBasePtr();
  EVT MemVT = Load->getMemoryVT();

  const MachinePointerInfo &SrcValue = Load->getMemOperand()->getPointerInfo();

  EVT LoVT, HiVT;
  EVT LoMemVT, HiMemVT;
  SDValue Lo, Hi;

  std::tie(LoVT, HiVT) = getSplitDestVTs(VT, DAG);
  std::tie(LoMemVT, HiMemVT) = getSplitDestVTs(MemVT, DAG);
  std::tie(Lo, Hi) = splitVector(Op, SL, LoVT, HiVT, DAG);

  unsigned Size = LoMemVT.getStoreSize();
  Align BaseAlign = Load->getAlign();
  Align HiAlign = commonAlignment(BaseAlign, Size);

  SDValue LoLoad = DAG.getExtLoad(Load->getExtensionType(), SL, LoVT,
                                  Load->getChain(), BasePtr, SrcValue, LoMemVT,
                                  BaseAlign, Load->getMemOperand()->getFlags());
  SDValue HiPtr = DAG.getObjectPtrOffset(SL, BasePtr, TypeSize::Fixed(Size));
  SDValue HiLoad =
      DAG.getExtLoad(Load->getExtensionType(), SL, HiVT, Load->getChain(),
                     HiPtr, SrcValue.getWithOffset(LoMemVT.getStoreSize()),
                     HiMemVT, HiAlign, Load->getMemOperand()->getFlags());

  SDValue Join;
  if (LoVT == HiVT) {
    // This is the case that the vector is power of two so was evenly split.
    Join = DAG.getNode(ISD::CONCAT_VECTORS, SL, VT, LoLoad, HiLoad);
  } else {
    Join = DAG.getNode(ISD::INSERT_SUBVECTOR, SL, VT, DAG.getUNDEF(VT), LoLoad,
                       DAG.getVectorIdxConstant(0, SL));
    Join = DAG.getNode(
        HiVT.isVector() ? ISD::INSERT_SUBVECTOR : ISD::INSERT_VECTOR_ELT, SL,
        VT, Join, HiLoad,
        DAG.getVectorIdxConstant(LoVT.getVectorNumElements(), SL));
  }

  SDValue Ops[] = {Join, DAG.getNode(ISD::TokenFactor, SL, MVT::Other,
                                     LoLoad.getValue(1), HiLoad.getValue(1))};

  return DAG.getMergeValues(Ops, SL);
}

SDValue AMDGPUTargetLowering::WidenOrSplitVectorLoad(SDValue Op,
                                                     SelectionDAG &DAG) const {
  LoadSDNode *Load = cast<LoadSDNode>(Op);
  EVT VT = Op.getValueType();
  SDValue BasePtr = Load->getBasePtr();
  EVT MemVT = Load->getMemoryVT();
  SDLoc SL(Op);
  const MachinePointerInfo &SrcValue = Load->getMemOperand()->getPointerInfo();
  Align BaseAlign = Load->getAlign();
  unsigned NumElements = MemVT.getVectorNumElements();

  // Widen from vec3 to vec4 when the load is at least 8-byte aligned
  // or 16-byte fully dereferenceable. Otherwise, split the vector load.
  if (NumElements != 3 ||
      (BaseAlign < Align(8) &&
       !SrcValue.isDereferenceable(16, *DAG.getContext(), DAG.getDataLayout())))
    return SplitVectorLoad(Op, DAG);

  assert(NumElements == 3);

  EVT WideVT =
      EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(), 4);
  EVT WideMemVT =
      EVT::getVectorVT(*DAG.getContext(), MemVT.getVectorElementType(), 4);
  SDValue WideLoad = DAG.getExtLoad(
      Load->getExtensionType(), SL, WideVT, Load->getChain(), BasePtr, SrcValue,
      WideMemVT, BaseAlign, Load->getMemOperand()->getFlags());
  return DAG.getMergeValues(
      {DAG.getNode(ISD::EXTRACT_SUBVECTOR, SL, VT, WideLoad,
                   DAG.getVectorIdxConstant(0, SL)),
       WideLoad.getValue(1)},
      SL);
}

SDValue AMDGPUTargetLowering::SplitVectorStore(SDValue Op,
                                               SelectionDAG &DAG) const {
  StoreSDNode *Store = cast<StoreSDNode>(Op);
  SDValue Val = Store->getValue();
  EVT VT = Val.getValueType();

  // If this is a 2 element vector, we really want to scalarize and not create
  // weird 1 element vectors.
  if (VT.getVectorNumElements() == 2)
    return scalarizeVectorStore(Store, DAG);

  EVT MemVT = Store->getMemoryVT();
  SDValue Chain = Store->getChain();
  SDValue BasePtr = Store->getBasePtr();
  SDLoc SL(Op);

  EVT LoVT, HiVT;
  EVT LoMemVT, HiMemVT;
  SDValue Lo, Hi;

  std::tie(LoVT, HiVT) = getSplitDestVTs(VT, DAG);
  std::tie(LoMemVT, HiMemVT) = getSplitDestVTs(MemVT, DAG);
  std::tie(Lo, Hi) = splitVector(Val, SL, LoVT, HiVT, DAG);

  SDValue HiPtr = DAG.getObjectPtrOffset(SL, BasePtr, LoMemVT.getStoreSize());

  const MachinePointerInfo &SrcValue = Store->getMemOperand()->getPointerInfo();
  Align BaseAlign = Store->getAlign();
  unsigned Size = LoMemVT.getStoreSize();
  Align HiAlign = commonAlignment(BaseAlign, Size);

  SDValue LoStore =
      DAG.getTruncStore(Chain, SL, Lo, BasePtr, SrcValue, LoMemVT, BaseAlign,
                        Store->getMemOperand()->getFlags());
  SDValue HiStore =
      DAG.getTruncStore(Chain, SL, Hi, HiPtr, SrcValue.getWithOffset(Size),
                        HiMemVT, HiAlign, Store->getMemOperand()->getFlags());

  return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, LoStore, HiStore);
}

// This is a shortcut for integer division because we have fast i32<->f32
// conversions, and fast f32 reciprocal instructions. The fractional part of a
// float is enough to accurately represent up to a 24-bit signed integer.
SDValue AMDGPUTargetLowering::LowerDIVREM24(SDValue Op, SelectionDAG &DAG,
                                            bool Sign) const {
  SDLoc DL(Op);
  EVT VT = Op.getValueType();
  SDValue LHS = Op.getOperand(0);
  SDValue RHS = Op.getOperand(1);
  MVT IntVT = MVT::i32;
  MVT FltVT = MVT::f32;

  unsigned LHSSignBits = DAG.ComputeNumSignBits(LHS);
  if (LHSSignBits < 9)
    return SDValue();

  unsigned RHSSignBits = DAG.ComputeNumSignBits(RHS);
  if (RHSSignBits < 9)
    return SDValue();

  unsigned BitSize = VT.getSizeInBits();
  unsigned SignBits = std::min(LHSSignBits, RHSSignBits);
  unsigned DivBits = BitSize - SignBits;
  if (Sign)
    ++DivBits;

  ISD::NodeType ToFp = Sign ? ISD::SINT_TO_FP : ISD::UINT_TO_FP;
  ISD::NodeType ToInt = Sign ? ISD::FP_TO_SINT : ISD::FP_TO_UINT;

  SDValue jq = DAG.getConstant(1, DL, IntVT);

  if (Sign) {
    // char|short jq = ia ^ ib;
    jq = DAG.getNode(ISD::XOR, DL, VT, LHS, RHS);

    // jq = jq >> (bitsize - 2)
    jq = DAG.getNode(ISD::SRA, DL, VT, jq,
                     DAG.getConstant(BitSize - 2, DL, VT));

    // jq = jq | 0x1
    jq = DAG.getNode(ISD::OR, DL, VT, jq, DAG.getConstant(1, DL, VT));
  }

  // int ia = (int)LHS;
  SDValue ia = LHS;

  // int ib, (int)RHS;
  SDValue ib = RHS;

  // float fa = (float)ia;
  SDValue fa = DAG.getNode(ToFp, DL, FltVT, ia);

  // float fb = (float)ib;
  SDValue fb = DAG.getNode(ToFp, DL, FltVT, ib);

  SDValue fq = DAG.getNode(ISD::FMUL, DL, FltVT,
                           fa, DAG.getNode(AMDGPUISD::RCP, DL, FltVT, fb));

  // fq = trunc(fq);
  fq = DAG.getNode(ISD::FTRUNC, DL, FltVT, fq);

  // float fqneg = -fq;
  SDValue fqneg = DAG.getNode(ISD::FNEG, DL, FltVT, fq);

  MachineFunction &MF = DAG.getMachineFunction();
  const AMDGPUMachineFunction *MFI = MF.getInfo<AMDGPUMachineFunction>();

  // float fr = mad(fqneg, fb, fa);
  unsigned OpCode = !Subtarget->hasMadMacF32Insts() ?
                    (unsigned)ISD::FMA :
                    !MFI->getMode().allFP32Denormals() ?
                    (unsigned)ISD::FMAD :
                    (unsigned)AMDGPUISD::FMAD_FTZ;
  SDValue fr = DAG.getNode(OpCode, DL, FltVT, fqneg, fb, fa);

  // int iq = (int)fq;
  SDValue iq = DAG.getNode(ToInt, DL, IntVT, fq);

  // fr = fabs(fr);
  fr = DAG.getNode(ISD::FABS, DL, FltVT, fr);

  // fb = fabs(fb);
  fb = DAG.getNode(ISD::FABS, DL, FltVT, fb);

  EVT SetCCVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);

  // int cv = fr >= fb;
  SDValue cv = DAG.getSetCC(DL, SetCCVT, fr, fb, ISD::SETOGE);

  // jq = (cv ? jq : 0);
  jq = DAG.getNode(ISD::SELECT, DL, VT, cv, jq, DAG.getConstant(0, DL, VT));

  // dst = iq + jq;
  SDValue Div = DAG.getNode(ISD::ADD, DL, VT, iq, jq);

  // Rem needs compensation, it's easier to recompute it
  SDValue Rem = DAG.getNode(ISD::MUL, DL, VT, Div, RHS);
  Rem = DAG.getNode(ISD::SUB, DL, VT, LHS, Rem);

  // Truncate to number of bits this divide really is.
  if (Sign) {
    SDValue InRegSize
      = DAG.getValueType(EVT::getIntegerVT(*DAG.getContext(), DivBits));
    Div = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT, Div, InRegSize);
    Rem = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT, Rem, InRegSize);
  } else {
    SDValue TruncMask = DAG.getConstant((UINT64_C(1) << DivBits) - 1, DL, VT);
    Div = DAG.getNode(ISD::AND, DL, VT, Div, TruncMask);
    Rem = DAG.getNode(ISD::AND, DL, VT, Rem, TruncMask);
  }

  return DAG.getMergeValues({ Div, Rem }, DL);
}

void AMDGPUTargetLowering::LowerUDIVREM64(SDValue Op,
                                      SelectionDAG &DAG,
                                      SmallVectorImpl<SDValue> &Results) const {
  SDLoc DL(Op);
  EVT VT = Op.getValueType();

  assert(VT == MVT::i64 && "LowerUDIVREM64 expects an i64");

  EVT HalfVT = VT.getHalfSizedIntegerVT(*DAG.getContext());

  SDValue One = DAG.getConstant(1, DL, HalfVT);
  SDValue Zero = DAG.getConstant(0, DL, HalfVT);

  //HiLo split
  SDValue LHS = Op.getOperand(0);
  SDValue LHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, Zero);
  SDValue LHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, One);

  SDValue RHS = Op.getOperand(1);
  SDValue RHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, Zero);
  SDValue RHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, One);

  if (DAG.MaskedValueIsZero(RHS, APInt::getHighBitsSet(64, 32)) &&
      DAG.MaskedValueIsZero(LHS, APInt::getHighBitsSet(64, 32))) {

    SDValue Res = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(HalfVT, HalfVT),
                              LHS_Lo, RHS_Lo);

    SDValue DIV = DAG.getBuildVector(MVT::v2i32, DL, {Res.getValue(0), Zero});
    SDValue REM = DAG.getBuildVector(MVT::v2i32, DL, {Res.getValue(1), Zero});

    Results.push_back(DAG.getNode(ISD::BITCAST, DL, MVT::i64, DIV));
    Results.push_back(DAG.getNode(ISD::BITCAST, DL, MVT::i64, REM));
    return;
  }

  if (isTypeLegal(MVT::i64)) {
    // The algorithm here is based on ideas from "Software Integer Division",
    // Tom Rodeheffer, August 2008.

    MachineFunction &MF = DAG.getMachineFunction();
    const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();

    // Compute denominator reciprocal.
    unsigned FMAD = !Subtarget->hasMadMacF32Insts() ?
                    (unsigned)ISD::FMA :
                    !MFI->getMode().allFP32Denormals() ?
                    (unsigned)ISD::FMAD :
                    (unsigned)AMDGPUISD::FMAD_FTZ;

    SDValue Cvt_Lo = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, RHS_Lo);
    SDValue Cvt_Hi = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, RHS_Hi);
    SDValue Mad1 = DAG.getNode(FMAD, DL, MVT::f32, Cvt_Hi,
      DAG.getConstantFP(APInt(32, 0x4f800000).bitsToFloat(), DL, MVT::f32),
      Cvt_Lo);
    SDValue Rcp = DAG.getNode(AMDGPUISD::RCP, DL, MVT::f32, Mad1);
    SDValue Mul1 = DAG.getNode(ISD::FMUL, DL, MVT::f32, Rcp,
      DAG.getConstantFP(APInt(32, 0x5f7ffffc).bitsToFloat(), DL, MVT::f32));
    SDValue Mul2 = DAG.getNode(ISD::FMUL, DL, MVT::f32, Mul1,
      DAG.getConstantFP(APInt(32, 0x2f800000).bitsToFloat(), DL, MVT::f32));
    SDValue Trunc = DAG.getNode(ISD::FTRUNC, DL, MVT::f32, Mul2);
    SDValue Mad2 = DAG.getNode(FMAD, DL, MVT::f32, Trunc,
      DAG.getConstantFP(APInt(32, 0xcf800000).bitsToFloat(), DL, MVT::f32),
      Mul1);
    SDValue Rcp_Lo = DAG.getNode(ISD::FP_TO_UINT, DL, HalfVT, Mad2);
    SDValue Rcp_Hi = DAG.getNode(ISD::FP_TO_UINT, DL, HalfVT, Trunc);
    SDValue Rcp64 = DAG.getBitcast(VT,
                        DAG.getBuildVector(MVT::v2i32, DL, {Rcp_Lo, Rcp_Hi}));

    SDValue Zero64 = DAG.getConstant(0, DL, VT);
    SDValue One64  = DAG.getConstant(1, DL, VT);
    SDValue Zero1 = DAG.getConstant(0, DL, MVT::i1);
    SDVTList HalfCarryVT = DAG.getVTList(HalfVT, MVT::i1);

    // First round of UNR (Unsigned integer Newton-Raphson).
    SDValue Neg_RHS = DAG.getNode(ISD::SUB, DL, VT, Zero64, RHS);
    SDValue Mullo1 = DAG.getNode(ISD::MUL, DL, VT, Neg_RHS, Rcp64);
    SDValue Mulhi1 = DAG.getNode(ISD::MULHU, DL, VT, Rcp64, Mullo1);
    SDValue Mulhi1_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, Mulhi1,
                                    Zero);
    SDValue Mulhi1_Hi =
        DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, Mulhi1, One);
    SDValue Add1_Lo = DAG.getNode(ISD::ADDCARRY, DL, HalfCarryVT, Rcp_Lo,
                                  Mulhi1_Lo, Zero1);
    SDValue Add1_Hi = DAG.getNode(ISD::ADDCARRY, DL, HalfCarryVT, Rcp_Hi,
                                  Mulhi1_Hi, Add1_Lo.getValue(1));
    SDValue Add1 = DAG.getBitcast(VT,
                        DAG.getBuildVector(MVT::v2i32, DL, {Add1_Lo, Add1_Hi}));

    // Second round of UNR.
    SDValue Mullo2 = DAG.getNode(ISD::MUL, DL, VT, Neg_RHS, Add1);
    SDValue Mulhi2 = DAG.getNode(ISD::MULHU, DL, VT, Add1, Mullo2);
    SDValue Mulhi2_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, Mulhi2,
                                    Zero);
    SDValue Mulhi2_Hi =
        DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, Mulhi2, One);
    SDValue Add2_Lo = DAG.getNode(ISD::ADDCARRY, DL, HalfCarryVT, Add1_Lo,
                                  Mulhi2_Lo, Zero1);
    SDValue Add2_Hi = DAG.getNode(ISD::ADDCARRY, DL, HalfCarryVT, Add1_Hi,
                                  Mulhi2_Hi, Add2_Lo.getValue(1));
    SDValue Add2 = DAG.getBitcast(VT,
                        DAG.getBuildVector(MVT::v2i32, DL, {Add2_Lo, Add2_Hi}));

    SDValue Mulhi3 = DAG.getNode(ISD::MULHU, DL, VT, LHS, Add2);

    SDValue Mul3 = DAG.getNode(ISD::MUL, DL, VT, RHS, Mulhi3);

    SDValue Mul3_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, Mul3, Zero);
    SDValue Mul3_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, Mul3, One);
    SDValue Sub1_Lo = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, LHS_Lo,
                                  Mul3_Lo, Zero1);
    SDValue Sub1_Hi = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, LHS_Hi,
                                  Mul3_Hi, Sub1_Lo.getValue(1));
    SDValue Sub1_Mi = DAG.getNode(ISD::SUB, DL, HalfVT, LHS_Hi, Mul3_Hi);
    SDValue Sub1 = DAG.getBitcast(VT,
                        DAG.getBuildVector(MVT::v2i32, DL, {Sub1_Lo, Sub1_Hi}));

    SDValue MinusOne = DAG.getConstant(0xffffffffu, DL, HalfVT);
    SDValue C1 = DAG.getSelectCC(DL, Sub1_Hi, RHS_Hi, MinusOne, Zero,
                                 ISD::SETUGE);
    SDValue C2 = DAG.getSelectCC(DL, Sub1_Lo, RHS_Lo, MinusOne, Zero,
                                 ISD::SETUGE);
    SDValue C3 = DAG.getSelectCC(DL, Sub1_Hi, RHS_Hi, C2, C1, ISD::SETEQ);

    // TODO: Here and below portions of the code can be enclosed into if/endif.
    // Currently control flow is unconditional and we have 4 selects after
    // potential endif to substitute PHIs.

    // if C3 != 0 ...
    SDValue Sub2_Lo = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, Sub1_Lo,
                                  RHS_Lo, Zero1);
    SDValue Sub2_Mi = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, Sub1_Mi,
                                  RHS_Hi, Sub1_Lo.getValue(1));
    SDValue Sub2_Hi = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, Sub2_Mi,
                                  Zero, Sub2_Lo.getValue(1));
    SDValue Sub2 = DAG.getBitcast(VT,
                        DAG.getBuildVector(MVT::v2i32, DL, {Sub2_Lo, Sub2_Hi}));

    SDValue Add3 = DAG.getNode(ISD::ADD, DL, VT, Mulhi3, One64);

    SDValue C4 = DAG.getSelectCC(DL, Sub2_Hi, RHS_Hi, MinusOne, Zero,
                                 ISD::SETUGE);
    SDValue C5 = DAG.getSelectCC(DL, Sub2_Lo, RHS_Lo, MinusOne, Zero,
                                 ISD::SETUGE);
    SDValue C6 = DAG.getSelectCC(DL, Sub2_Hi, RHS_Hi, C5, C4, ISD::SETEQ);

    // if (C6 != 0)
    SDValue Add4 = DAG.getNode(ISD::ADD, DL, VT, Add3, One64);

    SDValue Sub3_Lo = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, Sub2_Lo,
                                  RHS_Lo, Zero1);
    SDValue Sub3_Mi = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, Sub2_Mi,
                                  RHS_Hi, Sub2_Lo.getValue(1));
    SDValue Sub3_Hi = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, Sub3_Mi,
                                  Zero, Sub3_Lo.getValue(1));
    SDValue Sub3 = DAG.getBitcast(VT,
                        DAG.getBuildVector(MVT::v2i32, DL, {Sub3_Lo, Sub3_Hi}));

    // endif C6
    // endif C3

    SDValue Sel1 = DAG.getSelectCC(DL, C6, Zero, Add4, Add3, ISD::SETNE);
    SDValue Div  = DAG.getSelectCC(DL, C3, Zero, Sel1, Mulhi3, ISD::SETNE);

    SDValue Sel2 = DAG.getSelectCC(DL, C6, Zero, Sub3, Sub2, ISD::SETNE);
    SDValue Rem  = DAG.getSelectCC(DL, C3, Zero, Sel2, Sub1, ISD::SETNE);

    Results.push_back(Div);
    Results.push_back(Rem);

    return;
  }

  // r600 expandion.
  // Get Speculative values
  SDValue DIV_Part = DAG.getNode(ISD::UDIV, DL, HalfVT, LHS_Hi, RHS_Lo);
  SDValue REM_Part = DAG.getNode(ISD::UREM, DL, HalfVT, LHS_Hi, RHS_Lo);

  SDValue REM_Lo = DAG.getSelectCC(DL, RHS_Hi, Zero, REM_Part, LHS_Hi, ISD::SETEQ);
  SDValue REM = DAG.getBuildVector(MVT::v2i32, DL, {REM_Lo, Zero});
  REM = DAG.getNode(ISD::BITCAST, DL, MVT::i64, REM);

  SDValue DIV_Hi = DAG.getSelectCC(DL, RHS_Hi, Zero, DIV_Part, Zero, ISD::SETEQ);
  SDValue DIV_Lo = Zero;

  const unsigned halfBitWidth = HalfVT.getSizeInBits();

  for (unsigned i = 0; i < halfBitWidth; ++i) {
    const unsigned bitPos = halfBitWidth - i - 1;
    SDValue POS = DAG.getConstant(bitPos, DL, HalfVT);
    // Get value of high bit
    SDValue HBit = DAG.getNode(ISD::SRL, DL, HalfVT, LHS_Lo, POS);
    HBit = DAG.getNode(ISD::AND, DL, HalfVT, HBit, One);
    HBit = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, HBit);

    // Shift
    REM = DAG.getNode(ISD::SHL, DL, VT, REM, DAG.getConstant(1, DL, VT));
    // Add LHS high bit
    REM = DAG.getNode(ISD::OR, DL, VT, REM, HBit);

    SDValue BIT = DAG.getConstant(1ULL << bitPos, DL, HalfVT);
    SDValue realBIT = DAG.getSelectCC(DL, REM, RHS, BIT, Zero, ISD::SETUGE);

    DIV_Lo = DAG.getNode(ISD::OR, DL, HalfVT, DIV_Lo, realBIT);

    // Update REM
    SDValue REM_sub = DAG.getNode(ISD::SUB, DL, VT, REM, RHS);
    REM = DAG.getSelectCC(DL, REM, RHS, REM_sub, REM, ISD::SETUGE);
  }

  SDValue DIV = DAG.getBuildVector(MVT::v2i32, DL, {DIV_Lo, DIV_Hi});
  DIV = DAG.getNode(ISD::BITCAST, DL, MVT::i64, DIV);
  Results.push_back(DIV);
  Results.push_back(REM);
}

SDValue AMDGPUTargetLowering::LowerUDIVREM(SDValue Op,
                                           SelectionDAG &DAG) const {
  SDLoc DL(Op);
  EVT VT = Op.getValueType();

  if (VT == MVT::i64) {
    SmallVector<SDValue, 2> Results;
    LowerUDIVREM64(Op, DAG, Results);
    return DAG.getMergeValues(Results, DL);
  }

  if (VT == MVT::i32) {
    if (SDValue Res = LowerDIVREM24(Op, DAG, false))
      return Res;
  }

  SDValue X = Op.getOperand(0);
  SDValue Y = Op.getOperand(1);

  // See AMDGPUCodeGenPrepare::expandDivRem32 for a description of the
  // algorithm used here.

  // Initial estimate of inv(y).
  SDValue Z = DAG.getNode(AMDGPUISD::URECIP, DL, VT, Y);

  // One round of UNR.
  SDValue NegY = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), Y);
  SDValue NegYZ = DAG.getNode(ISD::MUL, DL, VT, NegY, Z);
  Z = DAG.getNode(ISD::ADD, DL, VT, Z,
                  DAG.getNode(ISD::MULHU, DL, VT, Z, NegYZ));

  // Quotient/remainder estimate.
  SDValue Q = DAG.getNode(ISD::MULHU, DL, VT, X, Z);
  SDValue R =
      DAG.getNode(ISD::SUB, DL, VT, X, DAG.getNode(ISD::MUL, DL, VT, Q, Y));

  // First quotient/remainder refinement.
  EVT CCVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);
  SDValue One = DAG.getConstant(1, DL, VT);
  SDValue Cond = DAG.getSetCC(DL, CCVT, R, Y, ISD::SETUGE);
  Q = DAG.getNode(ISD::SELECT, DL, VT, Cond,
                  DAG.getNode(ISD::ADD, DL, VT, Q, One), Q);
  R = DAG.getNode(ISD::SELECT, DL, VT, Cond,
                  DAG.getNode(ISD::SUB, DL, VT, R, Y), R);

  // Second quotient/remainder refinement.
  Cond = DAG.getSetCC(DL, CCVT, R, Y, ISD::SETUGE);
  Q = DAG.getNode(ISD::SELECT, DL, VT, Cond,
                  DAG.getNode(ISD::ADD, DL, VT, Q, One), Q);
  R = DAG.getNode(ISD::SELECT, DL, VT, Cond,
                  DAG.getNode(ISD::SUB, DL, VT, R, Y), R);

  return DAG.getMergeValues({Q, R}, DL);
}

SDValue AMDGPUTargetLowering::LowerSDIVREM(SDValue Op,
                                           SelectionDAG &DAG) const {
  SDLoc DL(Op);
  EVT VT = Op.getValueType();

  SDValue LHS = Op.getOperand(0);
  SDValue RHS = Op.getOperand(1);

  SDValue Zero = DAG.getConstant(0, DL, VT);
  SDValue NegOne = DAG.getConstant(-1, DL, VT);

  if (VT == MVT::i32) {
    if (SDValue Res = LowerDIVREM24(Op, DAG, true))
      return Res;
  }

  if (VT == MVT::i64 &&
      DAG.ComputeNumSignBits(LHS) > 32 &&
      DAG.ComputeNumSignBits(RHS) > 32) {
    EVT HalfVT = VT.getHalfSizedIntegerVT(*DAG.getContext());

    //HiLo split
    SDValue LHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, Zero);
    SDValue RHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, Zero);
    SDValue DIVREM = DAG.getNode(ISD::SDIVREM, DL, DAG.getVTList(HalfVT, HalfVT),
                                 LHS_Lo, RHS_Lo);
    SDValue Res[2] = {
      DAG.getNode(ISD::SIGN_EXTEND, DL, VT, DIVREM.getValue(0)),
      DAG.getNode(ISD::SIGN_EXTEND, DL, VT, DIVREM.getValue(1))
    };
    return DAG.getMergeValues(Res, DL);
  }

  SDValue LHSign = DAG.getSelectCC(DL, LHS, Zero, NegOne, Zero, ISD::SETLT);
  SDValue RHSign = DAG.getSelectCC(DL, RHS, Zero, NegOne, Zero, ISD::SETLT);
  SDValue DSign = DAG.getNode(ISD::XOR, DL, VT, LHSign, RHSign);
  SDValue RSign = LHSign; // Remainder sign is the same as LHS

  LHS = DAG.getNode(ISD::ADD, DL, VT, LHS, LHSign);
  RHS = DAG.getNode(ISD::ADD, DL, VT, RHS, RHSign);

  LHS = DAG.getNode(ISD::XOR, DL, VT, LHS, LHSign);
  RHS = DAG.getNode(ISD::XOR, DL, VT, RHS, RHSign);

  SDValue Div = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(VT, VT), LHS, RHS);
  SDValue Rem = Div.getValue(1);

  Div = DAG.getNode(ISD::XOR, DL, VT, Div, DSign);
  Rem = DAG.getNode(ISD::XOR, DL, VT, Rem, RSign);

  Div = DAG.getNode(ISD::SUB, DL, VT, Div, DSign);
  Rem = DAG.getNode(ISD::SUB, DL, VT, Rem, RSign);

  SDValue Res[2] = {
    Div,
    Rem
  };
  return DAG.getMergeValues(Res, DL);
}

// (frem x, y) -> (fma (fneg (ftrunc (fdiv x, y))), y, x)
SDValue AMDGPUTargetLowering::LowerFREM(SDValue Op, SelectionDAG &DAG) const {
  SDLoc SL(Op);
  EVT VT = Op.getValueType();
  auto Flags = Op->getFlags();
  SDValue X = Op.getOperand(0);
  SDValue Y = Op.getOperand(1);

  SDValue Div = DAG.getNode(ISD::FDIV, SL, VT, X, Y, Flags);
  SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, VT, Div, Flags);
  SDValue Neg = DAG.getNode(ISD::FNEG, SL, VT, Trunc, Flags);
  // TODO: For f32 use FMAD instead if !hasFastFMA32?
  return DAG.getNode(ISD::FMA, SL, VT, Neg, Y, X, Flags);
}

SDValue AMDGPUTargetLowering::LowerFCEIL(SDValue Op, SelectionDAG &DAG) const {
  SDLoc SL(Op);
  SDValue Src = Op.getOperand(0);

  // result = trunc(src)
  // if (src > 0.0 && src != result)
  //   result += 1.0

  SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);

  const SDValue Zero = DAG.getConstantFP(0.0, SL, MVT::f64);
  const SDValue One = DAG.getConstantFP(1.0, SL, MVT::f64);

  EVT SetCCVT =
      getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::f64);

  SDValue Lt0 = DAG.getSetCC(SL, SetCCVT, Src, Zero, ISD::SETOGT);
  SDValue NeTrunc = DAG.getSetCC(SL, SetCCVT, Src, Trunc, ISD::SETONE);
  SDValue And = DAG.getNode(ISD::AND, SL, SetCCVT, Lt0, NeTrunc);

  SDValue Add = DAG.getNode(ISD::SELECT, SL, MVT::f64, And, One, Zero);
  // TODO: Should this propagate fast-math-flags?
  return DAG.getNode(ISD::FADD, SL, MVT::f64, Trunc, Add);
}

static SDValue extractF64Exponent(SDValue Hi, const SDLoc &SL,
                                  SelectionDAG &DAG) {
  const unsigned FractBits = 52;
  const unsigned ExpBits = 11;

  SDValue ExpPart = DAG.getNode(AMDGPUISD::BFE_U32, SL, MVT::i32,
                                Hi,
                                DAG.getConstant(FractBits - 32, SL, MVT::i32),
                                DAG.getConstant(ExpBits, SL, MVT::i32));
  SDValue Exp = DAG.getNode(ISD::SUB, SL, MVT::i32, ExpPart,
                            DAG.getConstant(1023, SL, MVT::i32));

  return Exp;
}

SDValue AMDGPUTargetLowering::LowerFTRUNC(SDValue Op, SelectionDAG &DAG) const {
  SDLoc SL(Op);
  SDValue Src = Op.getOperand(0);

  assert(Op.getValueType() == MVT::f64);

  const SDValue Zero = DAG.getConstant(0, SL, MVT::i32);

  // Extract the upper half, since this is where we will find the sign and
  // exponent.
  SDValue Hi = getHiHalf64(Src, DAG);

  SDValue Exp = extractF64Exponent(Hi, SL, DAG);

  const unsigned FractBits = 52;

  // Extract the sign bit.
  const SDValue SignBitMask = DAG.getConstant(UINT32_C(1) << 31, SL, MVT::i32);
  SDValue SignBit = DAG.getNode(ISD::AND, SL, MVT::i32, Hi, SignBitMask);

  // Extend back to 64-bits.
  SDValue SignBit64 = DAG.getBuildVector(MVT::v2i32, SL, {Zero, SignBit});
  SignBit64 = DAG.getNode(ISD::BITCAST, SL, MVT::i64, SignBit64);

  SDValue BcInt = DAG.getNode(ISD::BITCAST, SL, MVT::i64, Src);
  const SDValue FractMask
    = DAG.getConstant((UINT64_C(1) << FractBits) - 1, SL, MVT::i64);

  SDValue Shr = DAG.getNode(ISD::SRA, SL, MVT::i64, FractMask, Exp);
  SDValue Not = DAG.getNOT(SL, Shr, MVT::i64);
  SDValue Tmp0 = DAG.getNode(ISD::AND, SL, MVT::i64, BcInt, Not);

  EVT SetCCVT =
      getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::i32);

  const SDValue FiftyOne = DAG.getConstant(FractBits - 1, SL, MVT::i32);

  SDValue ExpLt0 = DAG.getSetCC(SL, SetCCVT, Exp, Zero, ISD::SETLT);
  SDValue ExpGt51 = DAG.getSetCC(SL, SetCCVT, Exp, FiftyOne, ISD::SETGT);

  SDValue Tmp1 = DAG.getNode(ISD::SELECT, SL, MVT::i64, ExpLt0, SignBit64, Tmp0);
  SDValue Tmp2 = DAG.getNode(ISD::SELECT, SL, MVT::i64, ExpGt51, BcInt, Tmp1);

  return DAG.getNode(ISD::BITCAST, SL, MVT::f64, Tmp2);
}

SDValue AMDGPUTargetLowering::LowerFRINT(SDValue Op, SelectionDAG &DAG) const {
  SDLoc SL(Op);
  SDValue Src = Op.getOperand(0);

  assert(Op.getValueType() == MVT::f64);

  APFloat C1Val(APFloat::IEEEdouble(), "0x1.0p+52");
  SDValue C1 = DAG.getConstantFP(C1Val, SL, MVT::f64);
  SDValue CopySign = DAG.getNode(ISD::FCOPYSIGN, SL, MVT::f64, C1, Src);

  // TODO: Should this propagate fast-math-flags?

  SDValue Tmp1 = DAG.getNode(ISD::FADD, SL, MVT::f64, Src, CopySign);
  SDValue Tmp2 = DAG.getNode(ISD::FSUB, SL, MVT::f64, Tmp1, CopySign);

  SDValue Fabs = DAG.getNode(ISD::FABS, SL, MVT::f64, Src);

  APFloat C2Val(APFloat::IEEEdouble(), "0x1.fffffffffffffp+51");
  SDValue C2 = DAG.getConstantFP(C2Val, SL, MVT::f64);

  EVT SetCCVT =
      getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::f64);
  SDValue Cond = DAG.getSetCC(SL, SetCCVT, Fabs, C2, ISD::SETOGT);

  return DAG.getSelect(SL, MVT::f64, Cond, Src, Tmp2);
}

SDValue AMDGPUTargetLowering::LowerFNEARBYINT(SDValue Op, SelectionDAG &DAG) const {
  // FNEARBYINT and FRINT are the same, except in their handling of FP
  // exceptions. Those aren't really meaningful for us, and OpenCL only has
  // rint, so just treat them as equivalent.
  return DAG.getNode(ISD::FRINT, SDLoc(Op), Op.getValueType(), Op.getOperand(0));
}

// XXX - May require not supporting f32 denormals?

// Don't handle v2f16. The extra instructions to scalarize and repack around the
// compare and vselect end up producing worse code than scalarizing the whole
// operation.
SDValue AMDGPUTargetLowering::LowerFROUND(SDValue Op, SelectionDAG &DAG) const {
  SDLoc SL(Op);
  SDValue X = Op.getOperand(0);
  EVT VT = Op.getValueType();

  SDValue T = DAG.getNode(ISD::FTRUNC, SL, VT, X);

  // TODO: Should this propagate fast-math-flags?

  SDValue Diff = DAG.getNode(ISD::FSUB, SL, VT, X, T);

  SDValue AbsDiff = DAG.getNode(ISD::FABS, SL, VT, Diff);

  const SDValue Zero = DAG.getConstantFP(0.0, SL, VT);
  const SDValue One = DAG.getConstantFP(1.0, SL, VT);
  const SDValue Half = DAG.getConstantFP(0.5, SL, VT);

  SDValue SignOne = DAG.getNode(ISD::FCOPYSIGN, SL, VT, One, X);

  EVT SetCCVT =
      getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);

  SDValue Cmp = DAG.getSetCC(SL, SetCCVT, AbsDiff, Half, ISD::SETOGE);

  SDValue Sel = DAG.getNode(ISD::SELECT, SL, VT, Cmp, SignOne, Zero);

  return DAG.getNode(ISD::FADD, SL, VT, T, Sel);
}

SDValue AMDGPUTargetLowering::LowerFFLOOR(SDValue Op, SelectionDAG &DAG) const {
  SDLoc SL(Op);
  SDValue Src = Op.getOperand(0);

  // result = trunc(src);
  // if (src < 0.0 && src != result)
  //   result += -1.0.

  SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);

  const SDValue Zero = DAG.getConstantFP(0.0, SL, MVT::f64);
  const SDValue NegOne = DAG.getConstantFP(-1.0, SL, MVT::f64);

  EVT SetCCVT =
      getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::f64);

  SDValue Lt0 = DAG.getSetCC(SL, SetCCVT, Src, Zero, ISD::SETOLT);
  SDValue NeTrunc = DAG.getSetCC(SL, SetCCVT, Src, Trunc, ISD::SETONE);
  SDValue And = DAG.getNode(ISD::AND, SL, SetCCVT, Lt0, NeTrunc);

  SDValue Add = DAG.getNode(ISD::SELECT, SL, MVT::f64, And, NegOne, Zero);
  // TODO: Should this propagate fast-math-flags?
  return DAG.getNode(ISD::FADD, SL, MVT::f64, Trunc, Add);
}

SDValue AMDGPUTargetLowering::LowerFLOG(SDValue Op, SelectionDAG &DAG,
                                        double Log2BaseInverted) const {
  EVT VT = Op.getValueType();

  SDLoc SL(Op);
  SDValue Operand = Op.getOperand(0);
  SDValue Log2Operand = DAG.getNode(ISD::FLOG2, SL, VT, Operand);
  SDValue Log2BaseInvertedOperand = DAG.getConstantFP(Log2BaseInverted, SL, VT);

  return DAG.getNode(ISD::FMUL, SL, VT, Log2Operand, Log2BaseInvertedOperand);
}

// exp2(M_LOG2E_F * f);
SDValue AMDGPUTargetLowering::lowerFEXP(SDValue Op, SelectionDAG &DAG) const {
  EVT VT = Op.getValueType();
  SDLoc SL(Op);
  SDValue Src = Op.getOperand(0);

  const SDValue K = DAG.getConstantFP(numbers::log2e, SL, VT);
  SDValue Mul = DAG.getNode(ISD::FMUL, SL, VT, Src, K, Op->getFlags());
  return DAG.getNode(ISD::FEXP2, SL, VT, Mul, Op->getFlags());
}

static bool isCtlzOpc(unsigned Opc) {
  return Opc == ISD::CTLZ || Opc == ISD::CTLZ_ZERO_UNDEF;
}

static bool isCttzOpc(unsigned Opc) {
  return Opc == ISD::CTTZ || Opc == ISD::CTTZ_ZERO_UNDEF;
}

SDValue AMDGPUTargetLowering::LowerCTLZ_CTTZ(SDValue Op, SelectionDAG &DAG) const {
  SDLoc SL(Op);
  SDValue Src = Op.getOperand(0);

  assert(isCtlzOpc(Op.getOpcode()) || isCttzOpc(Op.getOpcode()));
  bool Ctlz = isCtlzOpc(Op.getOpcode());
  unsigned NewOpc = Ctlz ? AMDGPUISD::FFBH_U32 : AMDGPUISD::FFBL_B32;

  bool ZeroUndef = Op.getOpcode() == ISD::CTLZ_ZERO_UNDEF ||
                   Op.getOpcode() == ISD::CTTZ_ZERO_UNDEF;

  if (Src.getValueType() == MVT::i32) {
    // (ctlz hi:lo) -> (umin (ffbh src), 32)
    // (cttz hi:lo) -> (umin (ffbl src), 32)
    // (ctlz_zero_undef src) -> (ffbh src)
    // (cttz_zero_undef src) -> (ffbl src)
    SDValue NewOpr = DAG.getNode(NewOpc, SL, MVT::i32, Src);
    if (!ZeroUndef) {
      const SDValue Const32 = DAG.getConstant(32, SL, MVT::i32);
      NewOpr = DAG.getNode(ISD::UMIN, SL, MVT::i32, NewOpr, Const32);
    }
    return NewOpr;
  }

  SDValue Lo, Hi;
  std::tie(Lo, Hi) = split64BitValue(Src, DAG);

  SDValue OprLo = DAG.getNode(NewOpc, SL, MVT::i32, Lo);
  SDValue OprHi = DAG.getNode(NewOpc, SL, MVT::i32, Hi);

  // (ctlz hi:lo) -> (umin3 (ffbh hi), (uaddsat (ffbh lo), 32), 64)
  // (cttz hi:lo) -> (umin3 (uaddsat (ffbl hi), 32), (ffbl lo), 64)
  // (ctlz_zero_undef hi:lo) -> (umin (ffbh hi), (add (ffbh lo), 32))
  // (cttz_zero_undef hi:lo) -> (umin (add (ffbl hi), 32), (ffbl lo))

  unsigned AddOpc = ZeroUndef ? ISD::ADD : ISD::UADDSAT;
  const SDValue Const32 = DAG.getConstant(32, SL, MVT::i32);
  if (Ctlz)
    OprLo = DAG.getNode(AddOpc, SL, MVT::i32, OprLo, Const32);
  else
    OprHi = DAG.getNode(AddOpc, SL, MVT::i32, OprHi, Const32);

  SDValue NewOpr;
  NewOpr = DAG.getNode(ISD::UMIN, SL, MVT::i32, OprLo, OprHi);
  if (!ZeroUndef) {
    const SDValue Const64 = DAG.getConstant(64, SL, MVT::i32);
    NewOpr = DAG.getNode(ISD::UMIN, SL, MVT::i32, NewOpr, Const64);
  }

  return DAG.getNode(ISD::ZERO_EXTEND, SL, MVT::i64, NewOpr);
}

SDValue AMDGPUTargetLowering::LowerINT_TO_FP32(SDValue Op, SelectionDAG &DAG,
                                               bool Signed) const {
  // The regular method converting a 64-bit integer to float roughly consists of
  // 2 steps: normalization and rounding. In fact, after normalization, the
  // conversion from a 64-bit integer to a float is essentially the same as the
  // one from a 32-bit integer. The only difference is that it has more
  // trailing bits to be rounded. To leverage the native 32-bit conversion, a
  // 64-bit integer could be preprocessed and fit into a 32-bit integer then
  // converted into the correct float number. The basic steps for the unsigned
  // conversion are illustrated in the following pseudo code:
  //
  // f32 uitofp(i64 u) {
  //   i32 hi, lo = split(u);
  //   // Only count the leading zeros in hi as we have native support of the
  //   // conversion from i32 to f32. If hi is all 0s, the conversion is
  //   // reduced to a 32-bit one automatically.
  //   i32 shamt = clz(hi); // Return 32 if hi is all 0s.
  //   u <<= shamt;
  //   hi, lo = split(u);
  //   hi |= (lo != 0) ? 1 : 0; // Adjust rounding bit in hi based on lo.
  //   // convert it as a 32-bit integer and scale the result back.
  //   return uitofp(hi) * 2^(32 - shamt);
  // }
  //
  // The signed one follows the same principle but uses 'ffbh_i32' to count its
  // sign bits instead. If 'ffbh_i32' is not available, its absolute value is
  // converted instead followed by negation based its sign bit.

  SDLoc SL(Op);
  SDValue Src = Op.getOperand(0);

  SDValue Lo, Hi;
  std::tie(Lo, Hi) = split64BitValue(Src, DAG);
  SDValue Sign;
  SDValue ShAmt;
  if (Signed && Subtarget->isGCN()) {
    // We also need to consider the sign bit in Lo if Hi has just sign bits,
    // i.e. Hi is 0 or -1. However, that only needs to take the MSB into
    // account. That is, the maximal shift is
    // - 32 if Lo and Hi have opposite signs;
    // - 33 if Lo and Hi have the same sign.
    //
    // Or, MaxShAmt = 33 + OppositeSign, where
    //
    // OppositeSign is defined as ((Lo ^ Hi) >> 31), which is
    // - -1 if Lo and Hi have opposite signs; and
    // -  0 otherwise.
    //
    // All in all, ShAmt is calculated as
    //
    //  umin(sffbh(Hi), 33 + (Lo^Hi)>>31) - 1.
    //
    // or
    //
    //  umin(sffbh(Hi) - 1, 32 + (Lo^Hi)>>31).
    //
    // to reduce the critical path.
    SDValue OppositeSign = DAG.getNode(
        ISD::SRA, SL, MVT::i32, DAG.getNode(ISD::XOR, SL, MVT::i32, Lo, Hi),
        DAG.getConstant(31, SL, MVT::i32));
    SDValue MaxShAmt =
        DAG.getNode(ISD::ADD, SL, MVT::i32, DAG.getConstant(32, SL, MVT::i32),
                    OppositeSign);
    // Count the leading sign bits.
    ShAmt = DAG.getNode(AMDGPUISD::FFBH_I32, SL, MVT::i32, Hi);
    // Different from unsigned conversion, the shift should be one bit less to
    // preserve the sign bit.
    ShAmt = DAG.getNode(ISD::SUB, SL, MVT::i32, ShAmt,
                        DAG.getConstant(1, SL, MVT::i32));
    ShAmt = DAG.getNode(ISD::UMIN, SL, MVT::i32, ShAmt, MaxShAmt);
  } else {
    if (Signed) {
      // Without 'ffbh_i32', only leading zeros could be counted. Take the
      // absolute value first.
      Sign = DAG.getNode(ISD::SRA, SL, MVT::i64, Src,
                         DAG.getConstant(63, SL, MVT::i64));
      SDValue Abs =
          DAG.getNode(ISD::XOR, SL, MVT::i64,
                      DAG.getNode(ISD::ADD, SL, MVT::i64, Src, Sign), Sign);
      std::tie(Lo, Hi) = split64BitValue(Abs, DAG);
    }
    // Count the leading zeros.
    ShAmt = DAG.getNode(ISD::CTLZ, SL, MVT::i32, Hi);
    // The shift amount for signed integers is [0, 32].
  }
  // Normalize the given 64-bit integer.
  SDValue Norm = DAG.getNode(ISD::SHL, SL, MVT::i64, Src, ShAmt);
  // Split it again.
  std::tie(Lo, Hi) = split64BitValue(Norm, DAG);
  // Calculate the adjust bit for rounding.
  // (lo != 0) ? 1 : 0 => (lo >= 1) ? 1 : 0 => umin(1, lo)
  SDValue Adjust = DAG.getNode(ISD::UMIN, SL, MVT::i32,
                               DAG.getConstant(1, SL, MVT::i32), Lo);
  // Get the 32-bit normalized integer.
  Norm = DAG.getNode(ISD::OR, SL, MVT::i32, Hi, Adjust);
  // Convert the normalized 32-bit integer into f32.
  unsigned Opc =
      (Signed && Subtarget->isGCN()) ? ISD::SINT_TO_FP : ISD::UINT_TO_FP;
  SDValue FVal = DAG.getNode(Opc, SL, MVT::f32, Norm);

  // Finally, need to scale back the converted floating number as the original
  // 64-bit integer is converted as a 32-bit one.
  ShAmt = DAG.getNode(ISD::SUB, SL, MVT::i32, DAG.getConstant(32, SL, MVT::i32),
                      ShAmt);
  // On GCN, use LDEXP directly.
  if (Subtarget->isGCN())
    return DAG.getNode(AMDGPUISD::LDEXP, SL, MVT::f32, FVal, ShAmt);

  // Otherwise, align 'ShAmt' to the exponent part and add it into the exponent
  // part directly to emulate the multiplication of 2^ShAmt. That 8-bit
  // exponent is enough to avoid overflowing into the sign bit.
  SDValue Exp = DAG.getNode(ISD::SHL, SL, MVT::i32, ShAmt,
                            DAG.getConstant(23, SL, MVT::i32));
  SDValue IVal =
      DAG.getNode(ISD::ADD, SL, MVT::i32,
                  DAG.getNode(ISD::BITCAST, SL, MVT::i32, FVal), Exp);
  if (Signed) {
    // Set the sign bit.
    Sign = DAG.getNode(ISD::SHL, SL, MVT::i32,
                       DAG.getNode(ISD::TRUNCATE, SL, MVT::i32, Sign),
                       DAG.getConstant(31, SL, MVT::i32));
    IVal = DAG.getNode(ISD::OR, SL, MVT::i32, IVal, Sign);
  }
  return DAG.getNode(ISD::BITCAST, SL, MVT::f32, IVal);
}

SDValue AMDGPUTargetLowering::LowerINT_TO_FP64(SDValue Op, SelectionDAG &DAG,
                                               bool Signed) const {
  SDLoc SL(Op);
  SDValue Src = Op.getOperand(0);

  SDValue Lo, Hi;
  std::tie(Lo, Hi) = split64BitValue(Src, DAG);

  SDValue CvtHi = DAG.getNode(Signed ? ISD::SINT_TO_FP : ISD::UINT_TO_FP,
                              SL, MVT::f64, Hi);

  SDValue CvtLo = DAG.getNode(ISD::UINT_TO_FP, SL, MVT::f64, Lo);

  SDValue LdExp = DAG.getNode(AMDGPUISD::LDEXP, SL, MVT::f64, CvtHi,
                              DAG.getConstant(32, SL, MVT::i32));
  // TODO: Should this propagate fast-math-flags?
  return DAG.getNode(ISD::FADD, SL, MVT::f64, LdExp, CvtLo);
}

SDValue AMDGPUTargetLowering::LowerUINT_TO_FP(SDValue Op,
                                               SelectionDAG &DAG) const {
  // TODO: Factor out code common with LowerSINT_TO_FP.
  EVT DestVT = Op.getValueType();
  SDValue Src = Op.getOperand(0);
  EVT SrcVT = Src.getValueType();

  if (SrcVT == MVT::i16) {
    if (DestVT == MVT::f16)
      return Op;
    SDLoc DL(Op);

    // Promote src to i32
    SDValue Ext = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, Src);
    return DAG.getNode(ISD::UINT_TO_FP, DL, DestVT, Ext);
  }

  assert(SrcVT == MVT::i64 && "operation should be legal");

  if (Subtarget->has16BitInsts() && DestVT == MVT::f16) {
    SDLoc DL(Op);

    SDValue IntToFp32 = DAG.getNode(Op.getOpcode(), DL, MVT::f32, Src);
    SDValue FPRoundFlag = DAG.getIntPtrConstant(0, SDLoc(Op));
    SDValue FPRound =
        DAG.getNode(ISD::FP_ROUND, DL, MVT::f16, IntToFp32, FPRoundFlag);

    return FPRound;
  }

  if (DestVT == MVT::f32)
    return LowerINT_TO_FP32(Op, DAG, false);

  assert(DestVT == MVT::f64);
  return LowerINT_TO_FP64(Op, DAG, false);
}

SDValue AMDGPUTargetLowering::LowerSINT_TO_FP(SDValue Op,
                                              SelectionDAG &DAG) const {
  EVT DestVT = Op.getValueType();

  SDValue Src = Op.getOperand(0);
  EVT SrcVT = Src.getValueType();

  if (SrcVT == MVT::i16) {
    if (DestVT == MVT::f16)
      return Op;

    SDLoc DL(Op);
    // Promote src to i32
    SDValue Ext = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i32, Src);
    return DAG.getNode(ISD::SINT_TO_FP, DL, DestVT, Ext);
  }

  assert(SrcVT == MVT::i64 && "operation should be legal");

  // TODO: Factor out code common with LowerUINT_TO_FP.

  if (Subtarget->has16BitInsts() && DestVT == MVT::f16) {
    SDLoc DL(Op);
    SDValue Src = Op.getOperand(0);

    SDValue IntToFp32 = DAG.getNode(Op.getOpcode(), DL, MVT::f32, Src);
    SDValue FPRoundFlag = DAG.getIntPtrConstant(0, SDLoc(Op));
    SDValue FPRound =
        DAG.getNode(ISD::FP_ROUND, DL, MVT::f16, IntToFp32, FPRoundFlag);

    return FPRound;
  }

  if (DestVT == MVT::f32)
    return LowerINT_TO_FP32(Op, DAG, true);

  assert(DestVT == MVT::f64);
  return LowerINT_TO_FP64(Op, DAG, true);
}

SDValue AMDGPUTargetLowering::LowerFP_TO_INT64(SDValue Op, SelectionDAG &DAG,
                                               bool Signed) const {
  SDLoc SL(Op);

  SDValue Src = Op.getOperand(0);
  EVT SrcVT = Src.getValueType();

  assert(SrcVT == MVT::f32 || SrcVT == MVT::f64);

  // The basic idea of converting a floating point number into a pair of 32-bit
  // integers is illustrated as follows:
  //
  //     tf := trunc(val);
  //    hif := floor(tf * 2^-32);
  //    lof := tf - hif * 2^32; // lof is always positive due to floor.
  //     hi := fptoi(hif);
  //     lo := fptoi(lof);
  //
  SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, SrcVT, Src);
  SDValue Sign;
  if (Signed && SrcVT == MVT::f32) {
    // However, a 32-bit floating point number has only 23 bits mantissa and
    // it's not enough to hold all the significant bits of `lof` if val is
    // negative. To avoid the loss of precision, We need to take the absolute
    // value after truncating and flip the result back based on the original
    // signedness.
    Sign = DAG.getNode(ISD::SRA, SL, MVT::i32,
                       DAG.getNode(ISD::BITCAST, SL, MVT::i32, Trunc),
                       DAG.getConstant(31, SL, MVT::i32));
    Trunc = DAG.getNode(ISD::FABS, SL, SrcVT, Trunc);
  }

  SDValue K0, K1;
  if (SrcVT == MVT::f64) {
    K0 = DAG.getConstantFP(BitsToDouble(UINT64_C(/*2^-32*/ 0x3df0000000000000)),
                           SL, SrcVT);
    K1 = DAG.getConstantFP(BitsToDouble(UINT64_C(/*-2^32*/ 0xc1f0000000000000)),
                           SL, SrcVT);
  } else {
    K0 = DAG.getConstantFP(BitsToFloat(UINT32_C(/*2^-32*/ 0x2f800000)), SL,
                           SrcVT);
    K1 = DAG.getConstantFP(BitsToFloat(UINT32_C(/*-2^32*/ 0xcf800000)), SL,
                           SrcVT);
  }
  // TODO: Should this propagate fast-math-flags?
  SDValue Mul = DAG.getNode(ISD::FMUL, SL, SrcVT, Trunc, K0);

  SDValue FloorMul = DAG.getNode(ISD::FFLOOR, SL, SrcVT, Mul);

  SDValue Fma = DAG.getNode(ISD::FMA, SL, SrcVT, FloorMul, K1, Trunc);

  SDValue Hi = DAG.getNode((Signed && SrcVT == MVT::f64) ? ISD::FP_TO_SINT
                                                         : ISD::FP_TO_UINT,
                           SL, MVT::i32, FloorMul);
  SDValue Lo = DAG.getNode(ISD::FP_TO_UINT, SL, MVT::i32, Fma);

  SDValue Result = DAG.getNode(ISD::BITCAST, SL, MVT::i64,
                               DAG.getBuildVector(MVT::v2i32, SL, {Lo, Hi}));

  if (Signed && SrcVT == MVT::f32) {
    assert(Sign);
    // Flip the result based on the signedness, which is either all 0s or 1s.
    Sign = DAG.getNode(ISD::BITCAST, SL, MVT::i64,
                       DAG.getBuildVector(MVT::v2i32, SL, {Sign, Sign}));
    // r := xor(r, sign) - sign;
    Result =
        DAG.getNode(ISD::SUB, SL, MVT::i64,
                    DAG.getNode(ISD::XOR, SL, MVT::i64, Result, Sign), Sign);
  }

  return Result;
}

SDValue AMDGPUTargetLowering::LowerFP_TO_FP16(SDValue Op, SelectionDAG &DAG) const {
  SDLoc DL(Op);
  SDValue N0 = Op.getOperand(0);

  // Convert to target node to get known bits
  if (N0.getValueType() == MVT::f32)
    return DAG.getNode(AMDGPUISD::FP_TO_FP16, DL, Op.getValueType(), N0);

  if (getTargetMachine().Options.UnsafeFPMath) {
    // There is a generic expand for FP_TO_FP16 with unsafe fast math.
    return SDValue();
  }

  assert(N0.getSimpleValueType() == MVT::f64);

  // f64 -> f16 conversion using round-to-nearest-even rounding mode.
  const unsigned ExpMask = 0x7ff;
  const unsigned ExpBiasf64 = 1023;
  const unsigned ExpBiasf16 = 15;
  SDValue Zero = DAG.getConstant(0, DL, MVT::i32);
  SDValue One = DAG.getConstant(1, DL, MVT::i32);
  SDValue U = DAG.getNode(ISD::BITCAST, DL, MVT::i64, N0);
  SDValue UH = DAG.getNode(ISD::SRL, DL, MVT::i64, U,
                           DAG.getConstant(32, DL, MVT::i64));
  UH = DAG.getZExtOrTrunc(UH, DL, MVT::i32);
  U = DAG.getZExtOrTrunc(U, DL, MVT::i32);
  SDValue E = DAG.getNode(ISD::SRL, DL, MVT::i32, UH,
                          DAG.getConstant(20, DL, MVT::i64));
  E = DAG.getNode(ISD::AND, DL, MVT::i32, E,
                  DAG.getConstant(ExpMask, DL, MVT::i32));
  // Subtract the fp64 exponent bias (1023) to get the real exponent and
  // add the f16 bias (15) to get the biased exponent for the f16 format.
  E = DAG.getNode(ISD::ADD, DL, MVT::i32, E,
                  DAG.getConstant(-ExpBiasf64 + ExpBiasf16, DL, MVT::i32));

  SDValue M = DAG.getNode(ISD::SRL, DL, MVT::i32, UH,
                          DAG.getConstant(8, DL, MVT::i32));
  M = DAG.getNode(ISD::AND, DL, MVT::i32, M,
                  DAG.getConstant(0xffe, DL, MVT::i32));

  SDValue MaskedSig = DAG.getNode(ISD::AND, DL, MVT::i32, UH,
                                  DAG.getConstant(0x1ff, DL, MVT::i32));
  MaskedSig = DAG.getNode(ISD::OR, DL, MVT::i32, MaskedSig, U);

  SDValue Lo40Set = DAG.getSelectCC(DL, MaskedSig, Zero, Zero, One, ISD::SETEQ);
  M = DAG.getNode(ISD::OR, DL, MVT::i32, M, Lo40Set);

  // (M != 0 ? 0x0200 : 0) | 0x7c00;
  SDValue I = DAG.getNode(ISD::OR, DL, MVT::i32,
      DAG.getSelectCC(DL, M, Zero, DAG.getConstant(0x0200, DL, MVT::i32),
                      Zero, ISD::SETNE), DAG.getConstant(0x7c00, DL, MVT::i32));

  // N = M | (E << 12);
  SDValue N = DAG.getNode(ISD::OR, DL, MVT::i32, M,
      DAG.getNode(ISD::SHL, DL, MVT::i32, E,
                  DAG.getConstant(12, DL, MVT::i32)));

  // B = clamp(1-E, 0, 13);
  SDValue OneSubExp = DAG.getNode(ISD::SUB, DL, MVT::i32,
                                  One, E);
  SDValue B = DAG.getNode(ISD::SMAX, DL, MVT::i32, OneSubExp, Zero);
  B = DAG.getNode(ISD::SMIN, DL, MVT::i32, B,
                  DAG.getConstant(13, DL, MVT::i32));

  SDValue SigSetHigh = DAG.getNode(ISD::OR, DL, MVT::i32, M,
                                   DAG.getConstant(0x1000, DL, MVT::i32));

  SDValue D = DAG.getNode(ISD::SRL, DL, MVT::i32, SigSetHigh, B);
  SDValue D0 = DAG.getNode(ISD::SHL, DL, MVT::i32, D, B);
  SDValue D1 = DAG.getSelectCC(DL, D0, SigSetHigh, One, Zero, ISD::SETNE);
  D = DAG.getNode(ISD::OR, DL, MVT::i32, D, D1);

  SDValue V = DAG.getSelectCC(DL, E, One, D, N, ISD::SETLT);
  SDValue VLow3 = DAG.getNode(ISD::AND, DL, MVT::i32, V,
                              DAG.getConstant(0x7, DL, MVT::i32));
  V = DAG.getNode(ISD::SRL, DL, MVT::i32, V,
                  DAG.getConstant(2, DL, MVT::i32));
  SDValue V0 = DAG.getSelectCC(DL, VLow3, DAG.getConstant(3, DL, MVT::i32),
                               One, Zero, ISD::SETEQ);
  SDValue V1 = DAG.getSelectCC(DL, VLow3, DAG.getConstant(5, DL, MVT::i32),
                               One, Zero, ISD::SETGT);
  V1 = DAG.getNode(ISD::OR, DL, MVT::i32, V0, V1);
  V = DAG.getNode(ISD::ADD, DL, MVT::i32, V, V1);

  V = DAG.getSelectCC(DL, E, DAG.getConstant(30, DL, MVT::i32),
                      DAG.getConstant(0x7c00, DL, MVT::i32), V, ISD::SETGT);
  V = DAG.getSelectCC(DL, E, DAG.getConstant(1039, DL, MVT::i32),
                      I, V, ISD::SETEQ);

  // Extract the sign bit.
  SDValue Sign = DAG.getNode(ISD::SRL, DL, MVT::i32, UH,
                            DAG.getConstant(16, DL, MVT::i32));
  Sign = DAG.getNode(ISD::AND, DL, MVT::i32, Sign,
                     DAG.getConstant(0x8000, DL, MVT::i32));

  V = DAG.getNode(ISD::OR, DL, MVT::i32, Sign, V);
  return DAG.getZExtOrTrunc(V, DL, Op.getValueType());
}

SDValue AMDGPUTargetLowering::LowerFP_TO_INT(SDValue Op,
                                             SelectionDAG &DAG) const {
  SDValue Src = Op.getOperand(0);
  unsigned OpOpcode = Op.getOpcode();
  EVT SrcVT = Src.getValueType();
  EVT DestVT = Op.getValueType();

  // Will be selected natively
  if (SrcVT == MVT::f16 && DestVT == MVT::i16)
    return Op;

  // Promote i16 to i32
  if (DestVT == MVT::i16 && (SrcVT == MVT::f32 || SrcVT == MVT::f64)) {
    SDLoc DL(Op);

    SDValue FpToInt32 = DAG.getNode(OpOpcode, DL, MVT::i32, Src);
    return DAG.getNode(ISD::TRUNCATE, DL, MVT::i16, FpToInt32);
  }

  if (SrcVT == MVT::f16 ||
      (SrcVT == MVT::f32 && Src.getOpcode() == ISD::FP16_TO_FP)) {
    SDLoc DL(Op);

    SDValue FpToInt32 = DAG.getNode(OpOpcode, DL, MVT::i32, Src);
    unsigned Ext =
        OpOpcode == ISD::FP_TO_SINT ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
    return DAG.getNode(Ext, DL, MVT::i64, FpToInt32);
  }

  if (DestVT == MVT::i64 && (SrcVT == MVT::f32 || SrcVT == MVT::f64))
    return LowerFP_TO_INT64(Op, DAG, OpOpcode == ISD::FP_TO_SINT);

  return SDValue();
}

SDValue AMDGPUTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
                                                     SelectionDAG &DAG) const {
  EVT ExtraVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
  MVT VT = Op.getSimpleValueType();
  MVT ScalarVT = VT.getScalarType();

  assert(VT.isVector());

  SDValue Src = Op.getOperand(0);
  SDLoc DL(Op);

  // TODO: Don't scalarize on Evergreen?
  unsigned NElts = VT.getVectorNumElements();
  SmallVector<SDValue, 8> Args;
  DAG.ExtractVectorElements(Src, Args, 0, NElts);

  SDValue VTOp = DAG.getValueType(ExtraVT.getScalarType());
  for (unsigned I = 0; I < NElts; ++I)
    Args[I] = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, ScalarVT, Args[I], VTOp);

  return DAG.getBuildVector(VT, DL, Args);
}

//===----------------------------------------------------------------------===//
// Custom DAG optimizations
//===----------------------------------------------------------------------===//

static bool isU24(SDValue Op, SelectionDAG &DAG) {
  return AMDGPUTargetLowering::numBitsUnsigned(Op, DAG) <= 24;
}

static bool isI24(SDValue Op, SelectionDAG &DAG) {
  EVT VT = Op.getValueType();
  return VT.getSizeInBits() >= 24 && // Types less than 24-bit should be treated
                                     // as unsigned 24-bit values.
         AMDGPUTargetLowering::numBitsSigned(Op, DAG) <= 24;
}

static SDValue simplifyMul24(SDNode *Node24,
                             TargetLowering::DAGCombinerInfo &DCI) {
  SelectionDAG &DAG = DCI.DAG;
  const TargetLowering &TLI = DAG.getTargetLoweringInfo();
  bool IsIntrin = Node24->getOpcode() == ISD::INTRINSIC_WO_CHAIN;

  SDValue LHS = IsIntrin ? Node24->getOperand(1) : Node24->getOperand(0);
  SDValue RHS = IsIntrin ? Node24->getOperand(2) : Node24->getOperand(1);
  unsigned NewOpcode = Node24->getOpcode();
  if (IsIntrin) {
    unsigned IID = cast<ConstantSDNode>(Node24->getOperand(0))->getZExtValue();
    switch (IID) {
    case Intrinsic::amdgcn_mul_i24:
      NewOpcode = AMDGPUISD::MUL_I24;
      break;
    case Intrinsic::amdgcn_mul_u24:
      NewOpcode = AMDGPUISD::MUL_U24;
      break;
    case Intrinsic::amdgcn_mulhi_i24:
      NewOpcode = AMDGPUISD::MULHI_I24;
      break;
    case Intrinsic::amdgcn_mulhi_u24:
      NewOpcode = AMDGPUISD::MULHI_U24;
      break;
    default:
      llvm_unreachable("Expected 24-bit mul intrinsic");
    }
  }

  APInt Demanded = APInt::getLowBitsSet(LHS.getValueSizeInBits(), 24);

  // First try to simplify using SimplifyMultipleUseDemandedBits which allows
  // the operands to have other uses, but will only perform simplifications that
  // involve bypassing some nodes for this user.
  SDValue DemandedLHS = TLI.SimplifyMultipleUseDemandedBits(LHS, Demanded, DAG);
  SDValue DemandedRHS = TLI.SimplifyMultipleUseDemandedBits(RHS, Demanded, DAG);
  if (DemandedLHS || DemandedRHS)
    return DAG.getNode(NewOpcode, SDLoc(Node24), Node24->getVTList(),
                       DemandedLHS ? DemandedLHS : LHS,
                       DemandedRHS ? DemandedRHS : RHS);

  // Now try SimplifyDemandedBits which can simplify the nodes used by our
  // operands if this node is the only user.
  if (TLI.SimplifyDemandedBits(LHS, Demanded, DCI))
    return SDValue(Node24, 0);
  if (TLI.SimplifyDemandedBits(RHS, Demanded, DCI))
    return SDValue(Node24, 0);

  return SDValue();
}

template <typename IntTy>
static SDValue constantFoldBFE(SelectionDAG &DAG, IntTy Src0, uint32_t Offset,
                               uint32_t Width, const SDLoc &DL) {
  if (Width + Offset < 32) {
    uint32_t Shl = static_cast<uint32_t>(Src0) << (32 - Offset - Width);
    IntTy Result = static_cast<IntTy>(Shl) >> (32 - Width);
    return DAG.getConstant(Result, DL, MVT::i32);
  }

  return DAG.getConstant(Src0 >> Offset, DL, MVT::i32);
}

static bool hasVolatileUser(SDNode *Val) {
  for (SDNode *U : Val->uses()) {
    if (MemSDNode *M = dyn_cast<MemSDNode>(U)) {
      if (M->isVolatile())
        return true;
    }
  }

  return false;
}

bool AMDGPUTargetLowering::shouldCombineMemoryType(EVT VT) const {
  // i32 vectors are the canonical memory type.
  if (VT.getScalarType() == MVT::i32 || isTypeLegal(VT))
    return false;

  if (!VT.isByteSized())
    return false;

  unsigned Size = VT.getStoreSize();

  if ((Size == 1 || Size == 2 || Size == 4) && !VT.isVector())
    return false;

  if (Size == 3 || (Size > 4 && (Size % 4 != 0)))
    return false;

  return true;
}

// Replace load of an illegal type with a store of a bitcast to a friendlier
// type.
SDValue AMDGPUTargetLowering::performLoadCombine(SDNode *N,
                                                 DAGCombinerInfo &DCI) const {
  if (!DCI.isBeforeLegalize())
    return SDValue();

  LoadSDNode *LN = cast<LoadSDNode>(N);
  if (!LN->isSimple() || !ISD::isNormalLoad(LN) || hasVolatileUser(LN))
    return SDValue();

  SDLoc SL(N);
  SelectionDAG &DAG = DCI.DAG;
  EVT VT = LN->getMemoryVT();

  unsigned Size = VT.getStoreSize();
  Align Alignment = LN->getAlign();
  if (Alignment < Size && isTypeLegal(VT)) {
    bool IsFast;
    unsigned AS = LN->getAddressSpace();

    // Expand unaligned loads earlier than legalization. Due to visitation order
    // problems during legalization, the emitted instructions to pack and unpack
    // the bytes again are not eliminated in the case of an unaligned copy.
    if (!allowsMisalignedMemoryAccesses(
            VT, AS, Alignment, LN->getMemOperand()->getFlags(), &IsFast)) {
      if (VT.isVector())
        return SplitVectorLoad(SDValue(LN, 0), DAG);

      SDValue Ops[2];
      std::tie(Ops[0], Ops[1]) = expandUnalignedLoad(LN, DAG);

      return DAG.getMergeValues(Ops, SDLoc(N));
    }

    if (!IsFast)
      return SDValue();
  }

  if (!shouldCombineMemoryType(VT))
    return SDValue();

  EVT NewVT = getEquivalentMemType(*DAG.getContext(), VT);

  SDValue NewLoad
    = DAG.getLoad(NewVT, SL, LN->getChain(),
                  LN->getBasePtr(), LN->getMemOperand());

  SDValue BC = DAG.getNode(ISD::BITCAST, SL, VT, NewLoad);
  DCI.CombineTo(N, BC, NewLoad.getValue(1));
  return SDValue(N, 0);
}

// Replace store of an illegal type with a store of a bitcast to a friendlier
// type.
SDValue AMDGPUTargetLowering::performStoreCombine(SDNode *N,
                                                  DAGCombinerInfo &DCI) const {
  if (!DCI.isBeforeLegalize())
    return SDValue();

  StoreSDNode *SN = cast<StoreSDNode>(N);
  if (!SN->isSimple() || !ISD::isNormalStore(SN))
    return SDValue();

  EVT VT = SN->getMemoryVT();
  unsigned Size = VT.getStoreSize();

  SDLoc SL(N);
  SelectionDAG &DAG = DCI.DAG;
  Align Alignment = SN->getAlign();
  if (Alignment < Size && isTypeLegal(VT)) {
    bool IsFast;
    unsigned AS = SN->getAddressSpace();

    // Expand unaligned stores earlier than legalization. Due to visitation
    // order problems during legalization, the emitted instructions to pack and
    // unpack the bytes again are not eliminated in the case of an unaligned
    // copy.
    if (!allowsMisalignedMemoryAccesses(
            VT, AS, Alignment, SN->getMemOperand()->getFlags(), &IsFast)) {
      if (VT.isVector())
        return SplitVectorStore(SDValue(SN, 0), DAG);

      return expandUnalignedStore(SN, DAG);
    }

    if (!IsFast)
      return SDValue();
  }

  if (!shouldCombineMemoryType(VT))
    return SDValue();

  EVT NewVT = getEquivalentMemType(*DAG.getContext(), VT);
  SDValue Val = SN->getValue();

  //DCI.AddToWorklist(Val.getNode());

  bool OtherUses = !Val.hasOneUse();
  SDValue CastVal = DAG.getNode(ISD::BITCAST, SL, NewVT, Val);
  if (OtherUses) {
    SDValue CastBack = DAG.getNode(ISD::BITCAST, SL, VT, CastVal);
    DAG.ReplaceAllUsesOfValueWith(Val, CastBack);
  }

  return DAG.getStore(SN->getChain(), SL, CastVal,
                      SN->getBasePtr(), SN->getMemOperand());
}

// FIXME: This should go in generic DAG combiner with an isTruncateFree check,
// but isTruncateFree is inaccurate for i16 now because of SALU vs. VALU
// issues.
SDValue AMDGPUTargetLowering::performAssertSZExtCombine(SDNode *N,
                                                        DAGCombinerInfo &DCI) const {
  SelectionDAG &DAG = DCI.DAG;
  SDValue N0 = N->getOperand(0);

  // (vt2 (assertzext (truncate vt0:x), vt1)) ->
  //     (vt2 (truncate (assertzext vt0:x, vt1)))
  if (N0.getOpcode() == ISD::TRUNCATE) {
    SDValue N1 = N->getOperand(1);
    EVT ExtVT = cast<VTSDNode>(N1)->getVT();
    SDLoc SL(N);

    SDValue Src = N0.getOperand(0);
    EVT SrcVT = Src.getValueType();
    if (SrcVT.bitsGE(ExtVT)) {
      SDValue NewInReg = DAG.getNode(N->getOpcode(), SL, SrcVT, Src, N1);
      return DAG.getNode(ISD::TRUNCATE, SL, N->getValueType(0), NewInReg);
    }
  }

  return SDValue();
}

SDValue AMDGPUTargetLowering::performIntrinsicWOChainCombine(
  SDNode *N, DAGCombinerInfo &DCI) const {
  unsigned IID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
  switch (IID) {
  case Intrinsic::amdgcn_mul_i24:
  case Intrinsic::amdgcn_mul_u24:
  case Intrinsic::amdgcn_mulhi_i24:
  case Intrinsic::amdgcn_mulhi_u24:
    return simplifyMul24(N, DCI);
  case Intrinsic::amdgcn_fract:
  case Intrinsic::amdgcn_rsq:
  case Intrinsic::amdgcn_rcp_legacy:
  case Intrinsic::amdgcn_rsq_legacy:
  case Intrinsic::amdgcn_rsq_clamp:
  case Intrinsic::amdgcn_ldexp: {
    // FIXME: This is probably wrong. If src is an sNaN, it won't be quieted
    SDValue Src = N->getOperand(1);
    return Src.isUndef() ? Src : SDValue();
  }
  default:
    return SDValue();
  }
}

/// Split the 64-bit value \p LHS into two 32-bit components, and perform the
/// binary operation \p Opc to it with the corresponding constant operands.
SDValue AMDGPUTargetLowering::splitBinaryBitConstantOpImpl(
  DAGCombinerInfo &DCI, const SDLoc &SL,
  unsigned Opc, SDValue LHS,
  uint32_t ValLo, uint32_t ValHi) const {
  SelectionDAG &DAG = DCI.DAG;
  SDValue Lo, Hi;
  std::tie(Lo, Hi) = split64BitValue(LHS, DAG);

  SDValue LoRHS = DAG.getConstant(ValLo, SL, MVT::i32);
  SDValue HiRHS = DAG.getConstant(ValHi, SL, MVT::i32);

  SDValue LoAnd = DAG.getNode(Opc, SL, MVT::i32, Lo, LoRHS);
  SDValue HiAnd = DAG.getNode(Opc, SL, MVT::i32, Hi, HiRHS);

  // Re-visit the ands. It's possible we eliminated one of them and it could
  // simplify the vector.
  DCI.AddToWorklist(Lo.getNode());
  DCI.AddToWorklist(Hi.getNode());

  SDValue Vec = DAG.getBuildVector(MVT::v2i32, SL, {LoAnd, HiAnd});
  return DAG.getNode(ISD::BITCAST, SL, MVT::i64, Vec);
}

SDValue AMDGPUTargetLowering::performShlCombine(SDNode *N,
                                                DAGCombinerInfo &DCI) const {
  EVT VT = N->getValueType(0);

  ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N->getOperand(1));
  if (!RHS)
    return SDValue();

  SDValue LHS = N->getOperand(0);
  unsigned RHSVal = RHS->getZExtValue();
  if (!RHSVal)
    return LHS;

  SDLoc SL(N);
  SelectionDAG &DAG = DCI.DAG;

  switch (LHS->getOpcode()) {
  default:
    break;
  case ISD::ZERO_EXTEND:
  case ISD::SIGN_EXTEND:
  case ISD::ANY_EXTEND: {
    SDValue X = LHS->getOperand(0);

    if (VT == MVT::i32 && RHSVal == 16 && X.getValueType() == MVT::i16 &&
        isOperationLegal(ISD::BUILD_VECTOR, MVT::v2i16)) {
      // Prefer build_vector as the canonical form if packed types are legal.
      // (shl ([asz]ext i16:x), 16 -> build_vector 0, x
      SDValue Vec = DAG.getBuildVector(MVT::v2i16, SL,
       { DAG.getConstant(0, SL, MVT::i16), LHS->getOperand(0) });
      return DAG.getNode(ISD::BITCAST, SL, MVT::i32, Vec);
    }

    // shl (ext x) => zext (shl x), if shift does not overflow int
    if (VT != MVT::i64)
      break;
    KnownBits Known = DAG.computeKnownBits(X);
    unsigned LZ = Known.countMinLeadingZeros();
    if (LZ < RHSVal)
      break;
    EVT XVT = X.getValueType();
    SDValue Shl = DAG.getNode(ISD::SHL, SL, XVT, X, SDValue(RHS, 0));
    return DAG.getZExtOrTrunc(Shl, SL, VT);
  }
  }

  if (VT != MVT::i64)
    return SDValue();

  // i64 (shl x, C) -> (build_pair 0, (shl x, C -32))

  // On some subtargets, 64-bit shift is a quarter rate instruction. In the
  // common case, splitting this into a move and a 32-bit shift is faster and
  // the same code size.
  if (RHSVal < 32)
    return SDValue();

  SDValue ShiftAmt = DAG.getConstant(RHSVal - 32, SL, MVT::i32);

  SDValue Lo = DAG.getNode(ISD::TRUNCATE, SL, MVT::i32, LHS);
  SDValue NewShift = DAG.getNode(ISD::SHL, SL, MVT::i32, Lo, ShiftAmt);

  const SDValue Zero = DAG.getConstant(0, SL, MVT::i32);

  SDValue Vec = DAG.getBuildVector(MVT::v2i32, SL, {Zero, NewShift});
  return DAG.getNode(ISD::BITCAST, SL, MVT::i64, Vec);
}

SDValue AMDGPUTargetLowering::performSraCombine(SDNode *N,
                                                DAGCombinerInfo &DCI) const {
  if (N->getValueType(0) != MVT::i64)
    return SDValue();

  const ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N->getOperand(1));
  if (!RHS)
    return SDValue();

  SelectionDAG &DAG = DCI.DAG;
  SDLoc SL(N);
  unsigned RHSVal = RHS->getZExtValue();

  // (sra i64:x, 32) -> build_pair x, (sra hi_32(x), 31)
  if (RHSVal == 32) {
    SDValue Hi = getHiHalf64(N->getOperand(0), DAG);
    SDValue NewShift = DAG.getNode(ISD::SRA, SL, MVT::i32, Hi,
                                   DAG.getConstant(31, SL, MVT::i32));

    SDValue BuildVec = DAG.getBuildVector(MVT::v2i32, SL, {Hi, NewShift});
    return DAG.getNode(ISD::BITCAST, SL, MVT::i64, BuildVec);
  }

  // (sra i64:x, 63) -> build_pair (sra hi_32(x), 31), (sra hi_32(x), 31)
  if (RHSVal == 63) {
    SDValue Hi = getHiHalf64(N->getOperand(0), DAG);
    SDValue NewShift = DAG.getNode(ISD::SRA, SL, MVT::i32, Hi,
                                   DAG.getConstant(31, SL, MVT::i32));
    SDValue BuildVec = DAG.getBuildVector(MVT::v2i32, SL, {NewShift, NewShift});
    return DAG.getNode(ISD::BITCAST, SL, MVT::i64, BuildVec);
  }

  return SDValue();
}

SDValue AMDGPUTargetLowering::performSrlCombine(SDNode *N,
                                                DAGCombinerInfo &DCI) const {
  auto *RHS = dyn_cast<ConstantSDNode>(N->getOperand(1));
  if (!RHS)
    return SDValue();

  EVT VT = N->getValueType(0);
  SDValue LHS = N->getOperand(0);
  unsigned ShiftAmt = RHS->getZExtValue();
  SelectionDAG &DAG = DCI.DAG;
  SDLoc SL(N);

  // fold (srl (and x, c1 << c2), c2) -> (and (srl(x, c2), c1)
  // this improves the ability to match BFE patterns in isel.
  if (LHS.getOpcode() == ISD::AND) {
    if (auto *Mask = dyn_cast<ConstantSDNode>(LHS.getOperand(1))) {
      unsigned MaskIdx, MaskLen;
      if (Mask->getAPIntValue().isShiftedMask(MaskIdx, MaskLen) &&
          MaskIdx == ShiftAmt) {
        return DAG.getNode(
            ISD::AND, SL, VT,
            DAG.getNode(ISD::SRL, SL, VT, LHS.getOperand(0), N->getOperand(1)),
            DAG.getNode(ISD::SRL, SL, VT, LHS.getOperand(1), N->getOperand(1)));
      }
    }
  }

  if (VT != MVT::i64)
    return SDValue();

  if (ShiftAmt < 32)
    return SDValue();

  // srl i64:x, C for C >= 32
  // =>
  //   build_pair (srl hi_32(x), C - 32), 0
  SDValue Zero = DAG.getConstant(0, SL, MVT::i32);

  SDValue Hi = getHiHalf64(LHS, DAG);

  SDValue NewConst = DAG.getConstant(ShiftAmt - 32, SL, MVT::i32);
  SDValue NewShift = DAG.getNode(ISD::SRL, SL, MVT::i32, Hi, NewConst);

  SDValue BuildPair = DAG.getBuildVector(MVT::v2i32, SL, {NewShift, Zero});

  return DAG.getNode(ISD::BITCAST, SL, MVT::i64, BuildPair);
}

SDValue AMDGPUTargetLowering::performTruncateCombine(
  SDNode *N, DAGCombinerInfo &DCI) const {
  SDLoc SL(N);
  SelectionDAG &DAG = DCI.DAG;
  EVT VT = N->getValueType(0);
  SDValue Src = N->getOperand(0);

  // vt1 (truncate (bitcast (build_vector vt0:x, ...))) -> vt1 (bitcast vt0:x)
  if (Src.getOpcode() == ISD::BITCAST && !VT.isVector()) {
    SDValue Vec = Src.getOperand(0);
    if (Vec.getOpcode() == ISD::BUILD_VECTOR) {
      SDValue Elt0 = Vec.getOperand(0);
      EVT EltVT = Elt0.getValueType();
      if (VT.getFixedSizeInBits() <= EltVT.getFixedSizeInBits()) {
        if (EltVT.isFloatingPoint()) {
          Elt0 = DAG.getNode(ISD::BITCAST, SL,
                             EltVT.changeTypeToInteger(), Elt0);
        }

        return DAG.getNode(ISD::TRUNCATE, SL, VT, Elt0);
      }
    }
  }

  // Equivalent of above for accessing the high element of a vector as an
  // integer operation.
  // trunc (srl (bitcast (build_vector x, y))), 16 -> trunc (bitcast y)
  if (Src.getOpcode() == ISD::SRL && !VT.isVector()) {
    if (auto K = isConstOrConstSplat(Src.getOperand(1))) {
      if (2 * K->getZExtValue() == Src.getValueType().getScalarSizeInBits()) {
        SDValue BV = stripBitcast(Src.getOperand(0));
        if (BV.getOpcode() == ISD::BUILD_VECTOR &&
            BV.getValueType().getVectorNumElements() == 2) {
          SDValue SrcElt = BV.getOperand(1);
          EVT SrcEltVT = SrcElt.getValueType();
          if (SrcEltVT.isFloatingPoint()) {
            SrcElt = DAG.getNode(ISD::BITCAST, SL,
                                 SrcEltVT.changeTypeToInteger(), SrcElt);
          }

          return DAG.getNode(ISD::TRUNCATE, SL, VT, SrcElt);
        }
      }
    }
  }

  // Partially shrink 64-bit shifts to 32-bit if reduced to 16-bit.
  //
  // i16 (trunc (srl i64:x, K)), K <= 16 ->
  //     i16 (trunc (srl (i32 (trunc x), K)))
  if (VT.getScalarSizeInBits() < 32) {
    EVT SrcVT = Src.getValueType();
    if (SrcVT.getScalarSizeInBits() > 32 &&
        (Src.getOpcode() == ISD::SRL ||
         Src.getOpcode() == ISD::SRA ||
         Src.getOpcode() == ISD::SHL)) {
      SDValue Amt = Src.getOperand(1);
      KnownBits Known = DAG.computeKnownBits(Amt);
      unsigned Size = VT.getScalarSizeInBits();
      if ((Known.isConstant() && Known.getConstant().ule(Size)) ||
          (Known.countMaxActiveBits() <= Log2_32(Size))) {
        EVT MidVT = VT.isVector() ?
          EVT::getVectorVT(*DAG.getContext(), MVT::i32,
                           VT.getVectorNumElements()) : MVT::i32;

        EVT NewShiftVT = getShiftAmountTy(MidVT, DAG.getDataLayout());
        SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SL, MidVT,
                                    Src.getOperand(0));
        DCI.AddToWorklist(Trunc.getNode());

        if (Amt.getValueType() != NewShiftVT) {
          Amt = DAG.getZExtOrTrunc(Amt, SL, NewShiftVT);
          DCI.AddToWorklist(Amt.getNode());
        }

        SDValue ShrunkShift = DAG.getNode(Src.getOpcode(), SL, MidVT,
                                          Trunc, Amt);
        return DAG.getNode(ISD::TRUNCATE, SL, VT, ShrunkShift);
      }
    }
  }

  return SDValue();
}

// We need to specifically handle i64 mul here to avoid unnecessary conversion
// instructions. If we only match on the legalized i64 mul expansion,
// SimplifyDemandedBits will be unable to remove them because there will be
// multiple uses due to the separate mul + mulh[su].
static SDValue getMul24(SelectionDAG &DAG, const SDLoc &SL,
                        SDValue N0, SDValue N1, unsigned Size, bool Signed) {
  if (Size <= 32) {
    unsigned MulOpc = Signed ? AMDGPUISD::MUL_I24 : AMDGPUISD::MUL_U24;
    return DAG.getNode(MulOpc, SL, MVT::i32, N0, N1);
  }

  unsigned MulLoOpc = Signed ? AMDGPUISD::MUL_I24 : AMDGPUISD::MUL_U24;
  unsigned MulHiOpc = Signed ? AMDGPUISD::MULHI_I24 : AMDGPUISD::MULHI_U24;

  SDValue MulLo = DAG.getNode(MulLoOpc, SL, MVT::i32, N0, N1);
  SDValue MulHi = DAG.getNode(MulHiOpc, SL, MVT::i32, N0, N1);

  return DAG.getNode(ISD::BUILD_PAIR, SL, MVT::i64, MulLo, MulHi);
}

SDValue AMDGPUTargetLowering::performMulCombine(SDNode *N,
                                                DAGCombinerInfo &DCI) const {
  EVT VT = N->getValueType(0);

  // Don't generate 24-bit multiplies on values that are in SGPRs, since
  // we only have a 32-bit scalar multiply (avoid values being moved to VGPRs
  // unnecessarily). isDivergent() is used as an approximation of whether the
  // value is in an SGPR.
  if (!N->isDivergent())
    return SDValue();

  unsigned Size = VT.getSizeInBits();
  if (VT.isVector() || Size > 64)
    return SDValue();

  // There are i16 integer mul/mad.
  if (Subtarget->has16BitInsts() && VT.getScalarType().bitsLE(MVT::i16))
    return SDValue();

  SelectionDAG &DAG = DCI.DAG;
  SDLoc DL(N);

  SDValue N0 = N->getOperand(0);
  SDValue N1 = N->getOperand(1);

  // SimplifyDemandedBits has the annoying habit of turning useful zero_extends
  // in the source into any_extends if the result of the mul is truncated. Since
  // we can assume the high bits are whatever we want, use the underlying value
  // to avoid the unknown high bits from interfering.
  if (N0.getOpcode() == ISD::ANY_EXTEND)
    N0 = N0.getOperand(0);

  if (N1.getOpcode() == ISD::ANY_EXTEND)
    N1 = N1.getOperand(0);

  SDValue Mul;

  if (Subtarget->hasMulU24() && isU24(N0, DAG) && isU24(N1, DAG)) {
    N0 = DAG.getZExtOrTrunc(N0, DL, MVT::i32);
    N1 = DAG.getZExtOrTrunc(N1, DL, MVT::i32);
    Mul = getMul24(DAG, DL, N0, N1, Size, false);
  } else if (Subtarget->hasMulI24() && isI24(N0, DAG) && isI24(N1, DAG)) {
    N0 = DAG.getSExtOrTrunc(N0, DL, MVT::i32);
    N1 = DAG.getSExtOrTrunc(N1, DL, MVT::i32);
    Mul = getMul24(DAG, DL, N0, N1, Size, true);
  } else {
    return SDValue();
  }

  // We need to use sext even for MUL_U24, because MUL_U24 is used
  // for signed multiply of 8 and 16-bit types.
  return DAG.getSExtOrTrunc(Mul, DL, VT);
}

SDValue
AMDGPUTargetLowering::performMulLoHiCombine(SDNode *N,
                                            DAGCombinerInfo &DCI) const {
  if (N->getValueType(0) != MVT::i32)
    return SDValue();

  SelectionDAG &DAG = DCI.DAG;
  SDLoc DL(N);

  SDValue N0 = N->getOperand(0);
  SDValue N1 = N->getOperand(1);

  // SimplifyDemandedBits has the annoying habit of turning useful zero_extends
  // in the source into any_extends if the result of the mul is truncated. Since
  // we can assume the high bits are whatever we want, use the underlying value
  // to avoid the unknown high bits from interfering.
  if (N0.getOpcode() == ISD::ANY_EXTEND)
    N0 = N0.getOperand(0);
  if (N1.getOpcode() == ISD::ANY_EXTEND)
    N1 = N1.getOperand(0);

  // Try to use two fast 24-bit multiplies (one for each half of the result)
  // instead of one slow extending multiply.
  unsigned LoOpcode, HiOpcode;
  if (Subtarget->hasMulU24() && isU24(N0, DAG) && isU24(N1, DAG)) {
    N0 = DAG.getZExtOrTrunc(N0, DL, MVT::i32);
    N1 = DAG.getZExtOrTrunc(N1, DL, MVT::i32);
    LoOpcode = AMDGPUISD::MUL_U24;
    HiOpcode = AMDGPUISD::MULHI_U24;
  } else if (Subtarget->hasMulI24() && isI24(N0, DAG) && isI24(N1, DAG)) {
    N0 = DAG.getSExtOrTrunc(N0, DL, MVT::i32);
    N1 = DAG.getSExtOrTrunc(N1, DL, MVT::i32);
    LoOpcode = AMDGPUISD::MUL_I24;
    HiOpcode = AMDGPUISD::MULHI_I24;
  } else {
    return SDValue();
  }

  SDValue Lo = DAG.getNode(LoOpcode, DL, MVT::i32, N0, N1);
  SDValue Hi = DAG.getNode(HiOpcode, DL, MVT::i32, N0, N1);
  DCI.CombineTo(N, Lo, Hi);
  return SDValue(N, 0);
}

SDValue AMDGPUTargetLowering::performMulhsCombine(SDNode *N,
                                                  DAGCombinerInfo &DCI) const {
  EVT VT = N->getValueType(0);

  if (!Subtarget->hasMulI24() || VT.isVector())
    return SDValue();

  // Don't generate 24-bit multiplies on values that are in SGPRs, since
  // we only have a 32-bit scalar multiply (avoid values being moved to VGPRs
  // unnecessarily). isDivergent() is used as an approximation of whether the
  // value is in an SGPR.
  // This doesn't apply if no s_mul_hi is available (since we'll end up with a
  // valu op anyway)
  if (Subtarget->hasSMulHi() && !N->isDivergent())
    return SDValue();

  SelectionDAG &DAG = DCI.DAG;
  SDLoc DL(N);

  SDValue N0 = N->getOperand(0);
  SDValue N1 = N->getOperand(1);

  if (!isI24(N0, DAG) || !isI24(N1, DAG))
    return SDValue();

  N0 = DAG.getSExtOrTrunc(N0, DL, MVT::i32);
  N1 = DAG.getSExtOrTrunc(N1, DL, MVT::i32);

  SDValue Mulhi = DAG.getNode(AMDGPUISD::MULHI_I24, DL, MVT::i32, N0, N1);
  DCI.AddToWorklist(Mulhi.getNode());
  return DAG.getSExtOrTrunc(Mulhi, DL, VT);
}

SDValue AMDGPUTargetLowering::performMulhuCombine(SDNode *N,
                                                  DAGCombinerInfo &DCI) const {
  EVT VT = N->getValueType(0);

  if (!Subtarget->hasMulU24() || VT.isVector() || VT.getSizeInBits() > 32)
    return SDValue();

  // Don't generate 24-bit multiplies on values that are in SGPRs, since
  // we only have a 32-bit scalar multiply (avoid values being moved to VGPRs
  // unnecessarily). isDivergent() is used as an approximation of whether the
  // value is in an SGPR.
  // This doesn't apply if no s_mul_hi is available (since we'll end up with a
  // valu op anyway)
  if (Subtarget->hasSMulHi() && !N->isDivergent())
    return SDValue();

  SelectionDAG &DAG = DCI.DAG;
  SDLoc DL(N);

  SDValue N0 = N->getOperand(0);
  SDValue N1 = N->getOperand(1);

  if (!isU24(N0, DAG) || !isU24(N1, DAG))
    return SDValue();

  N0 = DAG.getZExtOrTrunc(N0, DL, MVT::i32);
  N1 = DAG.getZExtOrTrunc(N1, DL, MVT::i32);

  SDValue Mulhi = DAG.getNode(AMDGPUISD::MULHI_U24, DL, MVT::i32, N0, N1);
  DCI.AddToWorklist(Mulhi.getNode());
  return DAG.getZExtOrTrunc(Mulhi, DL, VT);
}

static bool isNegativeOne(SDValue Val) {
  if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val))
    return C->isAllOnes();
  return false;
}

SDValue AMDGPUTargetLowering::getFFBX_U32(SelectionDAG &DAG,
                                          SDValue Op,
                                          const SDLoc &DL,
                                          unsigned Opc) const {
  EVT VT = Op.getValueType();
  EVT LegalVT = getTypeToTransformTo(*DAG.getContext(), VT);
  if (LegalVT != MVT::i32 && (Subtarget->has16BitInsts() &&
                              LegalVT != MVT::i16))
    return SDValue();

  if (VT != MVT::i32)
    Op = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, Op);

  SDValue FFBX = DAG.getNode(Opc, DL, MVT::i32, Op);
  if (VT != MVT::i32)
    FFBX = DAG.getNode(ISD::TRUNCATE, DL, VT, FFBX);

  return FFBX;
}

// The native instructions return -1 on 0 input. Optimize out a select that
// produces -1 on 0.
//
// TODO: If zero is not undef, we could also do this if the output is compared
// against the bitwidth.
//
// TODO: Should probably combine against FFBH_U32 instead of ctlz directly.
SDValue AMDGPUTargetLowering::performCtlz_CttzCombine(const SDLoc &SL, SDValue Cond,
                                                 SDValue LHS, SDValue RHS,
                                                 DAGCombinerInfo &DCI) const {
  ConstantSDNode *CmpRhs = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
  if (!CmpRhs || !CmpRhs->isZero())
    return SDValue();

  SelectionDAG &DAG = DCI.DAG;
  ISD::CondCode CCOpcode = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
  SDValue CmpLHS = Cond.getOperand(0);

  // select (setcc x, 0, eq), -1, (ctlz_zero_undef x) -> ffbh_u32 x
  // select (setcc x, 0, eq), -1, (cttz_zero_undef x) -> ffbl_u32 x
  if (CCOpcode == ISD::SETEQ &&
      (isCtlzOpc(RHS.getOpcode()) || isCttzOpc(RHS.getOpcode())) &&
      RHS.getOperand(0) == CmpLHS && isNegativeOne(LHS)) {
    unsigned Opc =
        isCttzOpc(RHS.getOpcode()) ? AMDGPUISD::FFBL_B32 : AMDGPUISD::FFBH_U32;
    return getFFBX_U32(DAG, CmpLHS, SL, Opc);
  }

  // select (setcc x, 0, ne), (ctlz_zero_undef x), -1 -> ffbh_u32 x
  // select (setcc x, 0, ne), (cttz_zero_undef x), -1 -> ffbl_u32 x
  if (CCOpcode == ISD::SETNE &&
      (isCtlzOpc(LHS.getOpcode()) || isCttzOpc(LHS.getOpcode())) &&
      LHS.getOperand(0) == CmpLHS && isNegativeOne(RHS)) {
    unsigned Opc =
        isCttzOpc(LHS.getOpcode()) ? AMDGPUISD::FFBL_B32 : AMDGPUISD::FFBH_U32;

    return getFFBX_U32(DAG, CmpLHS, SL, Opc);
  }

  return SDValue();
}

static SDValue distributeOpThroughSelect(TargetLowering::DAGCombinerInfo &DCI,
                                         unsigned Op,
                                         const SDLoc &SL,
                                         SDValue Cond,
                                         SDValue N1,
                                         SDValue N2) {
  SelectionDAG &DAG = DCI.DAG;
  EVT VT = N1.getValueType();

  SDValue NewSelect = DAG.getNode(ISD::SELECT, SL, VT, Cond,
                                  N1.getOperand(0), N2.getOperand(0));
  DCI.AddToWorklist(NewSelect.getNode());
  return DAG.getNode(Op, SL, VT, NewSelect);
}

// Pull a free FP operation out of a select so it may fold into uses.
//
// select c, (fneg x), (fneg y) -> fneg (select c, x, y)
// select c, (fneg x), k -> fneg (select c, x, (fneg k))
//
// select c, (fabs x), (fabs y) -> fabs (select c, x, y)
// select c, (fabs x), +k -> fabs (select c, x, k)
static SDValue foldFreeOpFromSelect(TargetLowering::DAGCombinerInfo &DCI,
                                    SDValue N) {
  SelectionDAG &DAG = DCI.DAG;
  SDValue Cond = N.getOperand(0);
  SDValue LHS = N.getOperand(1);
  SDValue RHS = N.getOperand(2);

  EVT VT = N.getValueType();
  if ((LHS.getOpcode() == ISD::FABS && RHS.getOpcode() == ISD::FABS) ||
      (LHS.getOpcode() == ISD::FNEG && RHS.getOpcode() == ISD::FNEG)) {
    return distributeOpThroughSelect(DCI, LHS.getOpcode(),
                                     SDLoc(N), Cond, LHS, RHS);
  }

  bool Inv = false;
  if (RHS.getOpcode() == ISD::FABS || RHS.getOpcode() == ISD::FNEG) {
    std::swap(LHS, RHS);
    Inv = true;
  }

  // TODO: Support vector constants.
  ConstantFPSDNode *CRHS = dyn_cast<ConstantFPSDNode>(RHS);
  if ((LHS.getOpcode() == ISD::FNEG || LHS.getOpcode() == ISD::FABS) && CRHS) {
    SDLoc SL(N);
    // If one side is an fneg/fabs and the other is a constant, we can push the
    // fneg/fabs down. If it's an fabs, the constant needs to be non-negative.
    SDValue NewLHS = LHS.getOperand(0);
    SDValue NewRHS = RHS;

    // Careful: if the neg can be folded up, don't try to pull it back down.
    bool ShouldFoldNeg = true;

    if (NewLHS.hasOneUse()) {
      unsigned Opc = NewLHS.getOpcode();
      if (LHS.getOpcode() == ISD::FNEG && fnegFoldsIntoOp(Opc))
        ShouldFoldNeg = false;
      if (LHS.getOpcode() == ISD::FABS && Opc == ISD::FMUL)
        ShouldFoldNeg = false;
    }

    if (ShouldFoldNeg) {
      if (LHS.getOpcode() == ISD::FNEG)
        NewRHS = DAG.getNode(ISD::FNEG, SL, VT, RHS);
      else if (CRHS->isNegative())
        return SDValue();

      if (Inv)
        std::swap(NewLHS, NewRHS);

      SDValue NewSelect = DAG.getNode(ISD::SELECT, SL, VT,
                                      Cond, NewLHS, NewRHS);
      DCI.AddToWorklist(NewSelect.getNode());
      return DAG.getNode(LHS.getOpcode(), SL, VT, NewSelect);
    }
  }

  return SDValue();
}


SDValue AMDGPUTargetLowering::performSelectCombine(SDNode *N,
                                                   DAGCombinerInfo &DCI) const {
  if (SDValue Folded = foldFreeOpFromSelect(DCI, SDValue(N, 0)))
    return Folded;

  SDValue Cond = N->getOperand(0);
  if (Cond.getOpcode() != ISD::SETCC)
    return SDValue();

  EVT VT = N->getValueType(0);
  SDValue LHS = Cond.getOperand(0);
  SDValue RHS = Cond.getOperand(1);
  SDValue CC = Cond.getOperand(2);

  SDValue True = N->getOperand(1);
  SDValue False = N->getOperand(2);

  if (Cond.hasOneUse()) { // TODO: Look for multiple select uses.
    SelectionDAG &DAG = DCI.DAG;
    if (DAG.isConstantValueOfAnyType(True) &&
        !DAG.isConstantValueOfAnyType(False)) {
      // Swap cmp + select pair to move constant to false input.
      // This will allow using VOPC cndmasks more often.
      // select (setcc x, y), k, x -> select (setccinv x, y), x, k

      SDLoc SL(N);
      ISD::CondCode NewCC =
          getSetCCInverse(cast<CondCodeSDNode>(CC)->get(), LHS.getValueType());

      SDValue NewCond = DAG.getSetCC(SL, Cond.getValueType(), LHS, RHS, NewCC);
      return DAG.getNode(ISD::SELECT, SL, VT, NewCond, False, True);
    }

    if (VT == MVT::f32 && Subtarget->hasFminFmaxLegacy()) {
      SDValue MinMax
        = combineFMinMaxLegacy(SDLoc(N), VT, LHS, RHS, True, False, CC, DCI);
      // Revisit this node so we can catch min3/max3/med3 patterns.
      //DCI.AddToWorklist(MinMax.getNode());
      return MinMax;
    }
  }

  // There's no reason to not do this if the condition has other uses.
  return performCtlz_CttzCombine(SDLoc(N), Cond, True, False, DCI);
}

static bool isInv2Pi(const APFloat &APF) {
  static const APFloat KF16(APFloat::IEEEhalf(), APInt(16, 0x3118));
  static const APFloat KF32(APFloat::IEEEsingle(), APInt(32, 0x3e22f983));
  static const APFloat KF64(APFloat::IEEEdouble(), APInt(64, 0x3fc45f306dc9c882));

  return APF.bitwiseIsEqual(KF16) ||
         APF.bitwiseIsEqual(KF32) ||
         APF.bitwiseIsEqual(KF64);
}

// 0 and 1.0 / (0.5 * pi) do not have inline immmediates, so there is an
// additional cost to negate them.
bool AMDGPUTargetLowering::isConstantCostlierToNegate(SDValue N) const {
  if (const ConstantFPSDNode *C = isConstOrConstSplatFP(N)) {
    if (C->isZero() && !C->isNegative())
      return true;

    if (Subtarget->hasInv2PiInlineImm() && isInv2Pi(C->getValueAPF()))
      return true;
  }

  return false;
}

static unsigned inverseMinMax(unsigned Opc) {
  switch (Opc) {
  case ISD::FMAXNUM:
    return ISD::FMINNUM;
  case ISD::FMINNUM:
    return ISD::FMAXNUM;
  case ISD::FMAXNUM_IEEE:
    return ISD::FMINNUM_IEEE;
  case ISD::FMINNUM_IEEE:
    return ISD::FMAXNUM_IEEE;
  case AMDGPUISD::FMAX_LEGACY:
    return AMDGPUISD::FMIN_LEGACY;
  case AMDGPUISD::FMIN_LEGACY:
    return  AMDGPUISD::FMAX_LEGACY;
  default:
    llvm_unreachable("invalid min/max opcode");
  }
}

SDValue AMDGPUTargetLowering::performFNegCombine(SDNode *N,
                                                 DAGCombinerInfo &DCI) const {
  SelectionDAG &DAG = DCI.DAG;
  SDValue N0 = N->getOperand(0);
  EVT VT = N->getValueType(0);

  unsigned Opc = N0.getOpcode();

  // If the input has multiple uses and we can either fold the negate down, or
  // the other uses cannot, give up. This both prevents unprofitable
  // transformations and infinite loops: we won't repeatedly try to fold around
  // a negate that has no 'good' form.
  if (N0.hasOneUse()) {
    // This may be able to fold into the source, but at a code size cost. Don't
    // fold if the fold into the user is free.
    if (allUsesHaveSourceMods(N, 0))
      return SDValue();
  } else {
    if (fnegFoldsIntoOp(Opc) &&
        (allUsesHaveSourceMods(N) || !allUsesHaveSourceMods(N0.getNode())))
      return SDValue();
  }

  SDLoc SL(N);
  switch (Opc) {
  case ISD::FADD: {
    if (!mayIgnoreSignedZero(N0))
      return SDValue();

    // (fneg (fadd x, y)) -> (fadd (fneg x), (fneg y))
    SDValue LHS = N0.getOperand(0);
    SDValue RHS = N0.getOperand(1);

    if (LHS.getOpcode() != ISD::FNEG)
      LHS = DAG.getNode(ISD::FNEG, SL, VT, LHS);
    else
      LHS = LHS.getOperand(0);

    if (RHS.getOpcode() != ISD::FNEG)
      RHS = DAG.getNode(ISD::FNEG, SL, VT, RHS);
    else
      RHS = RHS.getOperand(0);

    SDValue Res = DAG.getNode(ISD::FADD, SL, VT, LHS, RHS, N0->getFlags());
    if (Res.getOpcode() != ISD::FADD)
      return SDValue(); // Op got folded away.
    if (!N0.hasOneUse())
      DAG.ReplaceAllUsesWith(N0, DAG.getNode(ISD::FNEG, SL, VT, Res));
    return Res;
  }
  case ISD::FMUL:
  case AMDGPUISD::FMUL_LEGACY: {
    // (fneg (fmul x, y)) -> (fmul x, (fneg y))
    // (fneg (fmul_legacy x, y)) -> (fmul_legacy x, (fneg y))
    SDValue LHS = N0.getOperand(0);
    SDValue RHS = N0.getOperand(1);

    if (LHS.getOpcode() == ISD::FNEG)
      LHS = LHS.getOperand(0);
    else if (RHS.getOpcode() == ISD::FNEG)
      RHS = RHS.getOperand(0);
    else
      RHS = DAG.getNode(ISD::FNEG, SL, VT, RHS);

    SDValue Res = DAG.getNode(Opc, SL, VT, LHS, RHS, N0->getFlags());
    if (Res.getOpcode() != Opc)
      return SDValue(); // Op got folded away.
    if (!N0.hasOneUse())
      DAG.ReplaceAllUsesWith(N0, DAG.getNode(ISD::FNEG, SL, VT, Res));
    return Res;
  }
  case ISD::FMA:
  case ISD::FMAD: {
    // TODO: handle llvm.amdgcn.fma.legacy
    if (!mayIgnoreSignedZero(N0))
      return SDValue();

    // (fneg (fma x, y, z)) -> (fma x, (fneg y), (fneg z))
    SDValue LHS = N0.getOperand(0);
    SDValue MHS = N0.getOperand(1);
    SDValue RHS = N0.getOperand(2);

    if (LHS.getOpcode() == ISD::FNEG)
      LHS = LHS.getOperand(0);
    else if (MHS.getOpcode() == ISD::FNEG)
      MHS = MHS.getOperand(0);
    else
      MHS = DAG.getNode(ISD::FNEG, SL, VT, MHS);

    if (RHS.getOpcode() != ISD::FNEG)
      RHS = DAG.getNode(ISD::FNEG, SL, VT, RHS);
    else
      RHS = RHS.getOperand(0);

    SDValue Res = DAG.getNode(Opc, SL, VT, LHS, MHS, RHS);
    if (Res.getOpcode() != Opc)
      return SDValue(); // Op got folded away.
    if (!N0.hasOneUse())
      DAG.ReplaceAllUsesWith(N0, DAG.getNode(ISD::FNEG, SL, VT, Res));
    return Res;
  }
  case ISD::FMAXNUM:
  case ISD::FMINNUM:
  case ISD::FMAXNUM_IEEE:
  case ISD::FMINNUM_IEEE:
  case AMDGPUISD::FMAX_LEGACY:
  case AMDGPUISD::FMIN_LEGACY: {
    // fneg (fmaxnum x, y) -> fminnum (fneg x), (fneg y)
    // fneg (fminnum x, y) -> fmaxnum (fneg x), (fneg y)
    // fneg (fmax_legacy x, y) -> fmin_legacy (fneg x), (fneg y)
    // fneg (fmin_legacy x, y) -> fmax_legacy (fneg x), (fneg y)

    SDValue LHS = N0.getOperand(0);
    SDValue RHS = N0.getOperand(1);

    // 0 doesn't have a negated inline immediate.
    // TODO: This constant check should be generalized to other operations.
    if (isConstantCostlierToNegate(RHS))
      return SDValue();

    SDValue NegLHS = DAG.getNode(ISD::FNEG, SL, VT, LHS);
    SDValue NegRHS = DAG.getNode(ISD::FNEG, SL, VT, RHS);
    unsigned Opposite = inverseMinMax(Opc);

    SDValue Res = DAG.getNode(Opposite, SL, VT, NegLHS, NegRHS, N0->getFlags());
    if (Res.getOpcode() != Opposite)
      return SDValue(); // Op got folded away.
    if (!N0.hasOneUse())
      DAG.ReplaceAllUsesWith(N0, DAG.getNode(ISD::FNEG, SL, VT, Res));
    return Res;
  }
  case AMDGPUISD::FMED3: {
    SDValue Ops[3];
    for (unsigned I = 0; I < 3; ++I)
      Ops[I] = DAG.getNode(ISD::FNEG, SL, VT, N0->getOperand(I), N0->getFlags());

    SDValue Res = DAG.getNode(AMDGPUISD::FMED3, SL, VT, Ops, N0->getFlags());
    if (Res.getOpcode() != AMDGPUISD::FMED3)
      return SDValue(); // Op got folded away.

    if (!N0.hasOneUse()) {
      SDValue Neg = DAG.getNode(ISD::FNEG, SL, VT, Res);
      DAG.ReplaceAllUsesWith(N0, Neg);

      for (SDNode *U : Neg->uses())
        DCI.AddToWorklist(U);
    }

    return Res;
  }
  case ISD::FP_EXTEND:
  case ISD::FTRUNC:
  case ISD::FRINT:
  case ISD::FNEARBYINT: // XXX - Should fround be handled?
  case ISD::FSIN:
  case ISD::FCANONICALIZE:
  case AMDGPUISD::RCP:
  case AMDGPUISD::RCP_LEGACY:
  case AMDGPUISD::RCP_IFLAG:
  case AMDGPUISD::SIN_HW: {
    SDValue CvtSrc = N0.getOperand(0);
    if (CvtSrc.getOpcode() == ISD::FNEG) {
      // (fneg (fp_extend (fneg x))) -> (fp_extend x)
      // (fneg (rcp (fneg x))) -> (rcp x)
      return DAG.getNode(Opc, SL, VT, CvtSrc.getOperand(0));
    }

    if (!N0.hasOneUse())
      return SDValue();

    // (fneg (fp_extend x)) -> (fp_extend (fneg x))
    // (fneg (rcp x)) -> (rcp (fneg x))
    SDValue Neg = DAG.getNode(ISD::FNEG, SL, CvtSrc.getValueType(), CvtSrc);
    return DAG.getNode(Opc, SL, VT, Neg, N0->getFlags());
  }
  case ISD::FP_ROUND: {
    SDValue CvtSrc = N0.getOperand(0);

    if (CvtSrc.getOpcode() == ISD::FNEG) {
      // (fneg (fp_round (fneg x))) -> (fp_round x)
      return DAG.getNode(ISD::FP_ROUND, SL, VT,
                         CvtSrc.getOperand(0), N0.getOperand(1));
    }

    if (!N0.hasOneUse())
      return SDValue();

    // (fneg (fp_round x)) -> (fp_round (fneg x))
    SDValue Neg = DAG.getNode(ISD::FNEG, SL, CvtSrc.getValueType(), CvtSrc);
    return DAG.getNode(ISD::FP_ROUND, SL, VT, Neg, N0.getOperand(1));
  }
  case ISD::FP16_TO_FP: {
    // v_cvt_f32_f16 supports source modifiers on pre-VI targets without legal
    // f16, but legalization of f16 fneg ends up pulling it out of the source.
    // Put the fneg back as a legal source operation that can be matched later.
    SDLoc SL(N);

    SDValue Src = N0.getOperand(0);
    EVT SrcVT = Src.getValueType();

    // fneg (fp16_to_fp x) -> fp16_to_fp (xor x, 0x8000)
    SDValue IntFNeg = DAG.getNode(ISD::XOR, SL, SrcVT, Src,
                                  DAG.getConstant(0x8000, SL, SrcVT));
    return DAG.getNode(ISD::FP16_TO_FP, SL, N->getValueType(0), IntFNeg);
  }
  default:
    return SDValue();
  }
}

SDValue AMDGPUTargetLowering::performFAbsCombine(SDNode *N,
                                                 DAGCombinerInfo &DCI) const {
  SelectionDAG &DAG = DCI.DAG;
  SDValue N0 = N->getOperand(0);

  if (!N0.hasOneUse())
    return SDValue();

  switch (N0.getOpcode()) {
  case ISD::FP16_TO_FP: {
    assert(!Subtarget->has16BitInsts() && "should only see if f16 is illegal");
    SDLoc SL(N);
    SDValue Src = N0.getOperand(0);
    EVT SrcVT = Src.getValueType();

    // fabs (fp16_to_fp x) -> fp16_to_fp (and x, 0x7fff)
    SDValue IntFAbs = DAG.getNode(ISD::AND, SL, SrcVT, Src,
                                  DAG.getConstant(0x7fff, SL, SrcVT));
    return DAG.getNode(ISD::FP16_TO_FP, SL, N->getValueType(0), IntFAbs);
  }
  default:
    return SDValue();
  }
}

SDValue AMDGPUTargetLowering::performRcpCombine(SDNode *N,
                                                DAGCombinerInfo &DCI) const {
  const auto *CFP = dyn_cast<ConstantFPSDNode>(N->getOperand(0));
  if (!CFP)
    return SDValue();

  // XXX - Should this flush denormals?
  const APFloat &Val = CFP->getValueAPF();
  APFloat One(Val.getSemantics(), "1.0");
  return DCI.DAG.getConstantFP(One / Val, SDLoc(N), N->getValueType(0));
}

SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
                                                DAGCombinerInfo &DCI) const {
  SelectionDAG &DAG = DCI.DAG;
  SDLoc DL(N);

  switch(N->getOpcode()) {
  default:
    break;
  case ISD::BITCAST: {
    EVT DestVT = N->getValueType(0);

    // Push casts through vector builds. This helps avoid emitting a large
    // number of copies when materializing floating point vector constants.
    //
    // vNt1 bitcast (vNt0 (build_vector t0:x, t0:y)) =>
    //   vnt1 = build_vector (t1 (bitcast t0:x)), (t1 (bitcast t0:y))
    if (DestVT.isVector()) {
      SDValue Src = N->getOperand(0);
      if (Src.getOpcode() == ISD::BUILD_VECTOR) {
        EVT SrcVT = Src.getValueType();
        unsigned NElts = DestVT.getVectorNumElements();

        if (SrcVT.getVectorNumElements() == NElts) {
          EVT DestEltVT = DestVT.getVectorElementType();

          SmallVector<SDValue, 8> CastedElts;
          SDLoc SL(N);
          for (unsigned I = 0, E = SrcVT.getVectorNumElements(); I != E; ++I) {
            SDValue Elt = Src.getOperand(I);
            CastedElts.push_back(DAG.getNode(ISD::BITCAST, DL, DestEltVT, Elt));
          }

          return DAG.getBuildVector(DestVT, SL, CastedElts);
        }
      }
    }

    if (DestVT.getSizeInBits() != 64 || !DestVT.isVector())
      break;

    // Fold bitcasts of constants.
    //
    // v2i32 (bitcast i64:k) -> build_vector lo_32(k), hi_32(k)
    // TODO: Generalize and move to DAGCombiner
    SDValue Src = N->getOperand(0);
    if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Src)) {
      SDLoc SL(N);
      uint64_t CVal = C->getZExtValue();
      SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, SL, MVT::v2i32,
                               DAG.getConstant(Lo_32(CVal), SL, MVT::i32),
                               DAG.getConstant(Hi_32(CVal), SL, MVT::i32));
      return DAG.getNode(ISD::BITCAST, SL, DestVT, BV);
    }

    if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Src)) {
      const APInt &Val = C->getValueAPF().bitcastToAPInt();
      SDLoc SL(N);
      uint64_t CVal = Val.getZExtValue();
      SDValue Vec = DAG.getNode(ISD::BUILD_VECTOR, SL, MVT::v2i32,
                                DAG.getConstant(Lo_32(CVal), SL, MVT::i32),
                                DAG.getConstant(Hi_32(CVal), SL, MVT::i32));

      return DAG.getNode(ISD::BITCAST, SL, DestVT, Vec);
    }

    break;
  }
  case ISD::SHL: {
    if (DCI.getDAGCombineLevel() < AfterLegalizeDAG)
      break;

    return performShlCombine(N, DCI);
  }
  case ISD::SRL: {
    if (DCI.getDAGCombineLevel() < AfterLegalizeDAG)
      break;

    return performSrlCombine(N, DCI);
  }
  case ISD::SRA: {
    if (DCI.getDAGCombineLevel() < AfterLegalizeDAG)
      break;

    return performSraCombine(N, DCI);
  }
  case ISD::TRUNCATE:
    return performTruncateCombine(N, DCI);
  case ISD::MUL:
    return performMulCombine(N, DCI);
  case ISD::SMUL_LOHI:
  case ISD::UMUL_LOHI:
    return performMulLoHiCombine(N, DCI);
  case ISD::MULHS:
    return performMulhsCombine(N, DCI);
  case ISD::MULHU:
    return performMulhuCombine(N, DCI);
  case AMDGPUISD::MUL_I24:
  case AMDGPUISD::MUL_U24:
  case AMDGPUISD::MULHI_I24:
  case AMDGPUISD::MULHI_U24:
    return simplifyMul24(N, DCI);
  case ISD::SELECT:
    return performSelectCombine(N, DCI);
  case ISD::FNEG:
    return performFNegCombine(N, DCI);
  case ISD::FABS:
    return performFAbsCombine(N, DCI);
  case AMDGPUISD::BFE_I32:
  case AMDGPUISD::BFE_U32: {
    assert(!N->getValueType(0).isVector() &&
           "Vector handling of BFE not implemented");
    ConstantSDNode *Width = dyn_cast<ConstantSDNode>(N->getOperand(2));
    if (!Width)
      break;

    uint32_t WidthVal = Width->getZExtValue() & 0x1f;
    if (WidthVal == 0)
      return DAG.getConstant(0, DL, MVT::i32);

    ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
    if (!Offset)
      break;

    SDValue BitsFrom = N->getOperand(0);
    uint32_t OffsetVal = Offset->getZExtValue() & 0x1f;

    bool Signed = N->getOpcode() == AMDGPUISD::BFE_I32;

    if (OffsetVal == 0) {
      // This is already sign / zero extended, so try to fold away extra BFEs.
      unsigned SignBits =  Signed ? (32 - WidthVal + 1) : (32 - WidthVal);

      unsigned OpSignBits = DAG.ComputeNumSignBits(BitsFrom);
      if (OpSignBits >= SignBits)
        return BitsFrom;

      EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), WidthVal);
      if (Signed) {
        // This is a sign_extend_inreg. Replace it to take advantage of existing
        // DAG Combines. If not eliminated, we will match back to BFE during
        // selection.

        // TODO: The sext_inreg of extended types ends, although we can could
        // handle them in a single BFE.
        return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, BitsFrom,
                           DAG.getValueType(SmallVT));
      }

      return DAG.getZeroExtendInReg(BitsFrom, DL, SmallVT);
    }

    if (ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(BitsFrom)) {
      if (Signed) {
        return constantFoldBFE<int32_t>(DAG,
                                        CVal->getSExtValue(),
                                        OffsetVal,
                                        WidthVal,
                                        DL);
      }

      return constantFoldBFE<uint32_t>(DAG,
                                       CVal->getZExtValue(),
                                       OffsetVal,
                                       WidthVal,
                                       DL);
    }

    if ((OffsetVal + WidthVal) >= 32 &&
        !(Subtarget->hasSDWA() && OffsetVal == 16 && WidthVal == 16)) {
      SDValue ShiftVal = DAG.getConstant(OffsetVal, DL, MVT::i32);
      return DAG.getNode(Signed ? ISD::SRA : ISD::SRL, DL, MVT::i32,
                         BitsFrom, ShiftVal);
    }

    if (BitsFrom.hasOneUse()) {
      APInt Demanded = APInt::getBitsSet(32,
                                         OffsetVal,
                                         OffsetVal + WidthVal);

      KnownBits Known;
      TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
                                            !DCI.isBeforeLegalizeOps());
      const TargetLowering &TLI = DAG.getTargetLoweringInfo();
      if (TLI.ShrinkDemandedConstant(BitsFrom, Demanded, TLO) ||
          TLI.SimplifyDemandedBits(BitsFrom, Demanded, Known, TLO)) {
        DCI.CommitTargetLoweringOpt(TLO);
      }
    }

    break;
  }
  case ISD::LOAD:
    return performLoadCombine(N, DCI);
  case ISD::STORE:
    return performStoreCombine(N, DCI);
  case AMDGPUISD::RCP:
  case AMDGPUISD::RCP_IFLAG:
    return performRcpCombine(N, DCI);
  case ISD::AssertZext:
  case ISD::AssertSext:
    return performAssertSZExtCombine(N, DCI);
  case ISD::INTRINSIC_WO_CHAIN:
    return performIntrinsicWOChainCombine(N, DCI);
  }
  return SDValue();
}

//===----------------------------------------------------------------------===//
// Helper functions
//===----------------------------------------------------------------------===//

SDValue AMDGPUTargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
                                                   const TargetRegisterClass *RC,
                                                   Register Reg, EVT VT,
                                                   const SDLoc &SL,
                                                   bool RawReg) const {
  MachineFunction &MF = DAG.getMachineFunction();
  MachineRegisterInfo &MRI = MF.getRegInfo();
  Register VReg;

  if (!MRI.isLiveIn(Reg)) {
    VReg = MRI.createVirtualRegister(RC);
    MRI.addLiveIn(Reg, VReg);
  } else {
    VReg = MRI.getLiveInVirtReg(Reg);
  }

  if (RawReg)
    return DAG.getRegister(VReg, VT);

  return DAG.getCopyFromReg(DAG.getEntryNode(), SL, VReg, VT);
}

// This may be called multiple times, and nothing prevents creating multiple
// objects at the same offset. See if we already defined this object.
static int getOrCreateFixedStackObject(MachineFrameInfo &MFI, unsigned Size,
                                       int64_t Offset) {
  for (int I = MFI.getObjectIndexBegin(); I < 0; ++I) {
    if (MFI.getObjectOffset(I) == Offset) {
      assert(MFI.getObjectSize(I) == Size);
      return I;
    }
  }

  return MFI.CreateFixedObject(Size, Offset, true);
}

SDValue AMDGPUTargetLowering::loadStackInputValue(SelectionDAG &DAG,
                                                  EVT VT,
                                                  const SDLoc &SL,
                                                  int64_t Offset) const {
  MachineFunction &MF = DAG.getMachineFunction();
  MachineFrameInfo &MFI = MF.getFrameInfo();
  int FI = getOrCreateFixedStackObject(MFI, VT.getStoreSize(), Offset);

  auto SrcPtrInfo = MachinePointerInfo::getStack(MF, Offset);
  SDValue Ptr = DAG.getFrameIndex(FI, MVT::i32);

  return DAG.getLoad(VT, SL, DAG.getEntryNode(), Ptr, SrcPtrInfo, Align(4),
                     MachineMemOperand::MODereferenceable |
                         MachineMemOperand::MOInvariant);
}

SDValue AMDGPUTargetLowering::storeStackInputValue(SelectionDAG &DAG,
                                                   const SDLoc &SL,
                                                   SDValue Chain,
                                                   SDValue ArgVal,
                                                   int64_t Offset) const {
  MachineFunction &MF = DAG.getMachineFunction();
  MachinePointerInfo DstInfo = MachinePointerInfo::getStack(MF, Offset);
  const SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();

  SDValue Ptr = DAG.getConstant(Offset, SL, MVT::i32);
  // Stores to the argument stack area are relative to the stack pointer.
  SDValue SP =
      DAG.getCopyFromReg(Chain, SL, Info->getStackPtrOffsetReg(), MVT::i32);
  Ptr = DAG.getNode(ISD::ADD, SL, MVT::i32, SP, Ptr);
  SDValue Store = DAG.getStore(Chain, SL, ArgVal, Ptr, DstInfo, Align(4),
                               MachineMemOperand::MODereferenceable);
  return Store;
}

SDValue AMDGPUTargetLowering::loadInputValue(SelectionDAG &DAG,
                                             const TargetRegisterClass *RC,
                                             EVT VT, const SDLoc &SL,
                                             const ArgDescriptor &Arg) const {
  assert(Arg && "Attempting to load missing argument");

  SDValue V = Arg.isRegister() ?
    CreateLiveInRegister(DAG, RC, Arg.getRegister(), VT, SL) :
    loadStackInputValue(DAG, VT, SL, Arg.getStackOffset());

  if (!Arg.isMasked())
    return V;

  unsigned Mask = Arg.getMask();
  unsigned Shift = countTrailingZeros<unsigned>(Mask);
  V = DAG.getNode(ISD::SRL, SL, VT, V,
                  DAG.getShiftAmountConstant(Shift, VT, SL