llvm.org GIT mirror llvm / b4989f0 lib / Target / AMDGPU / SIISelLowering.cpp
b4989f0

Tree @b4989f0 (Download .tar.gz)

SIISelLowering.cpp @b4989f0raw · history · 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
//===-- SIISelLowering.cpp - SI DAG Lowering Implementation ---------------===//
//
//                     The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
/// \file
/// \brief Custom DAG lowering for SI
//
//===----------------------------------------------------------------------===//

#ifdef _MSC_VER
// Provide M_PI.
#define _USE_MATH_DEFINES
#include <cmath>
#endif

#include "SIISelLowering.h"
#include "AMDGPU.h"
#include "AMDGPUDiagnosticInfoUnsupported.h"
#include "AMDGPUIntrinsicInfo.h"
#include "AMDGPUSubtarget.h"
#include "SIInstrInfo.h"
#include "SIMachineFunctionInfo.h"
#include "SIRegisterInfo.h"
#include "llvm/ADT/BitVector.h"
#include "llvm/CodeGen/CallingConvLower.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/SelectionDAG.h"
#include "llvm/IR/Function.h"
#include "llvm/ADT/SmallString.h"

using namespace llvm;

SITargetLowering::SITargetLowering(TargetMachine &TM,
                                   const AMDGPUSubtarget &STI)
    : AMDGPUTargetLowering(TM, STI) {
  addRegisterClass(MVT::i1, &AMDGPU::VReg_1RegClass);
  addRegisterClass(MVT::i64, &AMDGPU::SReg_64RegClass);

  addRegisterClass(MVT::v32i8, &AMDGPU::SReg_256RegClass);
  addRegisterClass(MVT::v64i8, &AMDGPU::SReg_512RegClass);

  addRegisterClass(MVT::i32, &AMDGPU::SReg_32RegClass);
  addRegisterClass(MVT::f32, &AMDGPU::VGPR_32RegClass);

  addRegisterClass(MVT::f64, &AMDGPU::VReg_64RegClass);
  addRegisterClass(MVT::v2i32, &AMDGPU::SReg_64RegClass);
  addRegisterClass(MVT::v2f32, &AMDGPU::VReg_64RegClass);

  addRegisterClass(MVT::v2i64, &AMDGPU::SReg_128RegClass);
  addRegisterClass(MVT::v2f64, &AMDGPU::SReg_128RegClass);

  addRegisterClass(MVT::v4i32, &AMDGPU::SReg_128RegClass);
  addRegisterClass(MVT::v4f32, &AMDGPU::VReg_128RegClass);

  addRegisterClass(MVT::v8i32, &AMDGPU::SReg_256RegClass);
  addRegisterClass(MVT::v8f32, &AMDGPU::VReg_256RegClass);

  addRegisterClass(MVT::v16i32, &AMDGPU::SReg_512RegClass);
  addRegisterClass(MVT::v16f32, &AMDGPU::VReg_512RegClass);

  computeRegisterProperties(STI.getRegisterInfo());

  setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i32, Expand);
  setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Expand);
  setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i32, Expand);
  setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16f32, Expand);

  setOperationAction(ISD::ADD, MVT::i32, Legal);
  setOperationAction(ISD::ADDC, MVT::i32, Legal);
  setOperationAction(ISD::ADDE, MVT::i32, Legal);
  setOperationAction(ISD::SUBC, MVT::i32, Legal);
  setOperationAction(ISD::SUBE, MVT::i32, Legal);

  setOperationAction(ISD::FSIN, MVT::f32, Custom);
  setOperationAction(ISD::FCOS, MVT::f32, Custom);

  setOperationAction(ISD::FMINNUM, MVT::f64, Legal);
  setOperationAction(ISD::FMAXNUM, MVT::f64, Legal);

  // We need to custom lower vector stores from local memory
  setOperationAction(ISD::LOAD, MVT::v4i32, Custom);
  setOperationAction(ISD::LOAD, MVT::v8i32, Custom);
  setOperationAction(ISD::LOAD, MVT::v16i32, Custom);

  setOperationAction(ISD::STORE, MVT::v8i32, Custom);
  setOperationAction(ISD::STORE, MVT::v16i32, Custom);

  setOperationAction(ISD::STORE, MVT::i1, Custom);
  setOperationAction(ISD::STORE, MVT::v4i32, Custom);

  setOperationAction(ISD::SELECT, MVT::i64, Custom);
  setOperationAction(ISD::SELECT, MVT::f64, Promote);
  AddPromotedToType(ISD::SELECT, MVT::f64, MVT::i64);

  setOperationAction(ISD::SELECT_CC, MVT::f32, Expand);
  setOperationAction(ISD::SELECT_CC, MVT::i32, Expand);
  setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);
  setOperationAction(ISD::SELECT_CC, MVT::f64, Expand);

  setOperationAction(ISD::SETCC, MVT::v2i1, Expand);
  setOperationAction(ISD::SETCC, MVT::v4i1, Expand);

  setOperationAction(ISD::BSWAP, MVT::i32, Legal);
  setOperationAction(ISD::BITREVERSE, MVT::i32, Legal);

  setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Legal);
  setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i1, Custom);
  setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i1, Custom);

  setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Legal);
  setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i8, Custom);
  setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i8, Custom);

  setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Legal);
  setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i16, Custom);
  setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i16, Custom);

  setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
  setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::Other, Custom);

  setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
  setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::f32, Custom);
  setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::v16i8, Custom);
  setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::v4f32, Custom);

  setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
  setOperationAction(ISD::BRCOND, MVT::Other, Custom);

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

    setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
    setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i8, Legal);
    setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i16, Legal);
    setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i32, Expand);

    setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i1, Promote);
    setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i8, Legal);
    setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i16, Legal);
    setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i32, Expand);

    setLoadExtAction(ISD::EXTLOAD, VT, MVT::i1, Promote);
    setLoadExtAction(ISD::EXTLOAD, VT, MVT::i8, Legal);
    setLoadExtAction(ISD::EXTLOAD, VT, MVT::i16, Legal);
    setLoadExtAction(ISD::EXTLOAD, VT, MVT::i32, Expand);
  }

  for (MVT VT : MVT::integer_vector_valuetypes()) {
    setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v8i16, Expand);
    setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v16i16, Expand);
  }

  for (MVT VT : MVT::fp_valuetypes())
    setLoadExtAction(ISD::EXTLOAD, VT, MVT::f32, Expand);

  setLoadExtAction(ISD::EXTLOAD, MVT::v2f64, MVT::v2f16, Expand);
  setLoadExtAction(ISD::EXTLOAD, MVT::v2f64, MVT::v2f32, Expand);

  setTruncStoreAction(MVT::i64, MVT::i32, Expand);
  setTruncStoreAction(MVT::v8i32, MVT::v8i16, Expand);
  setTruncStoreAction(MVT::v16i32, MVT::v16i8, Expand);
  setTruncStoreAction(MVT::v16i32, MVT::v16i16, Expand);


  setTruncStoreAction(MVT::v2i64, MVT::v2i32, Expand);

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

  setOperationAction(ISD::LOAD, MVT::i1, Custom);

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

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

  setOperationAction(ISD::ConstantPool, MVT::v2i64, Expand);

  setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
  setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
  setOperationAction(ISD::FrameIndex, MVT::i32, Custom);

  // These should use UDIVREM, so set them to expand
  setOperationAction(ISD::UDIV, MVT::i64, Expand);
  setOperationAction(ISD::UREM, MVT::i64, Expand);

  setOperationAction(ISD::SELECT_CC, MVT::i1, Expand);
  setOperationAction(ISD::SELECT, MVT::i1, Promote);

  setOperationAction(ISD::TRUNCATE, MVT::v2i32, Expand);


  setOperationAction(ISD::FP_ROUND, MVT::v2f32, Expand);

  // We only support LOAD/STORE and vector manipulation ops for vectors
  // with > 4 elements.
  for (MVT VT : {MVT::v8i32, MVT::v8f32, MVT::v16i32, MVT::v16f32, MVT::v2i64, MVT::v2f64}) {
    for (unsigned Op = 0; Op < ISD::BUILTIN_OP_END; ++Op) {
      switch(Op) {
      case ISD::LOAD:
      case ISD::STORE:
      case ISD::BUILD_VECTOR:
      case ISD::BITCAST:
      case ISD::EXTRACT_VECTOR_ELT:
      case ISD::INSERT_VECTOR_ELT:
      case ISD::INSERT_SUBVECTOR:
      case ISD::EXTRACT_SUBVECTOR:
      case ISD::SCALAR_TO_VECTOR:
        break;
      case ISD::CONCAT_VECTORS:
        setOperationAction(Op, VT, Custom);
        break;
      default:
        setOperationAction(Op, VT, Expand);
        break;
      }
    }
  }

  // Most operations are naturally 32-bit vector operations. We only support
  // load and store of i64 vectors, so promote v2i64 vector operations to v4i32.
  for (MVT Vec64 : { MVT::v2i64, MVT::v2f64 }) {
    setOperationAction(ISD::BUILD_VECTOR, Vec64, Promote);
    AddPromotedToType(ISD::BUILD_VECTOR, Vec64, MVT::v4i32);

    setOperationAction(ISD::EXTRACT_VECTOR_ELT, Vec64, Promote);
    AddPromotedToType(ISD::EXTRACT_VECTOR_ELT, Vec64, MVT::v4i32);

    setOperationAction(ISD::INSERT_VECTOR_ELT, Vec64, Promote);
    AddPromotedToType(ISD::INSERT_VECTOR_ELT, Vec64, MVT::v4i32);

    setOperationAction(ISD::SCALAR_TO_VECTOR, Vec64, Promote);
    AddPromotedToType(ISD::SCALAR_TO_VECTOR, Vec64, MVT::v4i32);
  }

  if (Subtarget->getGeneration() >= AMDGPUSubtarget::SEA_ISLANDS) {
    setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
    setOperationAction(ISD::FCEIL, MVT::f64, Legal);
    setOperationAction(ISD::FRINT, MVT::f64, Legal);
  }

  setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
  setOperationAction(ISD::FDIV, MVT::f32, Custom);
  setOperationAction(ISD::FDIV, MVT::f64, Custom);

  setTargetDAGCombine(ISD::FADD);
  setTargetDAGCombine(ISD::FSUB);
  setTargetDAGCombine(ISD::FMINNUM);
  setTargetDAGCombine(ISD::FMAXNUM);
  setTargetDAGCombine(ISD::SMIN);
  setTargetDAGCombine(ISD::SMAX);
  setTargetDAGCombine(ISD::UMIN);
  setTargetDAGCombine(ISD::UMAX);
  setTargetDAGCombine(ISD::SETCC);
  setTargetDAGCombine(ISD::AND);
  setTargetDAGCombine(ISD::OR);
  setTargetDAGCombine(ISD::UINT_TO_FP);

  // All memory operations. Some folding on the pointer operand is done to help
  // matching the constant offsets in the addressing modes.
  setTargetDAGCombine(ISD::LOAD);
  setTargetDAGCombine(ISD::STORE);
  setTargetDAGCombine(ISD::ATOMIC_LOAD);
  setTargetDAGCombine(ISD::ATOMIC_STORE);
  setTargetDAGCombine(ISD::ATOMIC_CMP_SWAP);
  setTargetDAGCombine(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS);
  setTargetDAGCombine(ISD::ATOMIC_SWAP);
  setTargetDAGCombine(ISD::ATOMIC_LOAD_ADD);
  setTargetDAGCombine(ISD::ATOMIC_LOAD_SUB);
  setTargetDAGCombine(ISD::ATOMIC_LOAD_AND);
  setTargetDAGCombine(ISD::ATOMIC_LOAD_OR);
  setTargetDAGCombine(ISD::ATOMIC_LOAD_XOR);
  setTargetDAGCombine(ISD::ATOMIC_LOAD_NAND);
  setTargetDAGCombine(ISD::ATOMIC_LOAD_MIN);
  setTargetDAGCombine(ISD::ATOMIC_LOAD_MAX);
  setTargetDAGCombine(ISD::ATOMIC_LOAD_UMIN);
  setTargetDAGCombine(ISD::ATOMIC_LOAD_UMAX);

  setSchedulingPreference(Sched::RegPressure);
}

//===----------------------------------------------------------------------===//
// TargetLowering queries
//===----------------------------------------------------------------------===//

bool SITargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &,
                                          EVT) const {
  // SI has some legal vector types, but no legal vector operations. Say no
  // shuffles are legal in order to prefer scalarizing some vector operations.
  return false;
}

bool SITargetLowering::isLegalFlatAddressingMode(const AddrMode &AM) const {
  // Flat instructions do not have offsets, and only have the register
  // address.
  return AM.BaseOffs == 0 && (AM.Scale == 0 || AM.Scale == 1);
}

bool SITargetLowering::isLegalMUBUFAddressingMode(const AddrMode &AM) const {
  // MUBUF / MTBUF instructions have a 12-bit unsigned byte offset, and
  // additionally can do r + r + i with addr64. 32-bit has more addressing
  // mode options. Depending on the resource constant, it can also do
  // (i64 r0) + (i32 r1) * (i14 i).
  //
  // Private arrays end up using a scratch buffer most of the time, so also
  // assume those use MUBUF instructions. Scratch loads / stores are currently
  // implemented as mubuf instructions with offen bit set, so slightly
  // different than the normal addr64.
  if (!isUInt<12>(AM.BaseOffs))
    return false;

  // FIXME: Since we can split immediate into soffset and immediate offset,
  // would it make sense to allow any immediate?

  switch (AM.Scale) {
  case 0: // r + i or just i, depending on HasBaseReg.
    return true;
  case 1:
    return true; // We have r + r or r + i.
  case 2:
    if (AM.HasBaseReg) {
      // Reject 2 * r + r.
      return false;
    }

    // Allow 2 * r as r + r
    // Or  2 * r + i is allowed as r + r + i.
    return true;
  default: // Don't allow n * r
    return false;
  }
}

bool SITargetLowering::isLegalAddressingMode(const DataLayout &DL,
                                             const AddrMode &AM, Type *Ty,
                                             unsigned AS) const {
  // No global is ever allowed as a base.
  if (AM.BaseGV)
    return false;

  switch (AS) {
  case AMDGPUAS::GLOBAL_ADDRESS: {
    if (Subtarget->getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) {
      // Assume the we will use FLAT for all global memory accesses
      // on VI.
      // FIXME: This assumption is currently wrong.  On VI we still use
      // MUBUF instructions for the r + i addressing mode.  As currently
      // implemented, the MUBUF instructions only work on buffer < 4GB.
      // It may be possible to support > 4GB buffers with MUBUF instructions,
      // by setting the stride value in the resource descriptor which would
      // increase the size limit to (stride * 4GB).  However, this is risky,
      // because it has never been validated.
      return isLegalFlatAddressingMode(AM);
    }

    return isLegalMUBUFAddressingMode(AM);
  }
  case AMDGPUAS::CONSTANT_ADDRESS: {
    // If the offset isn't a multiple of 4, it probably isn't going to be
    // correctly aligned.
    if (AM.BaseOffs % 4 != 0)
      return isLegalMUBUFAddressingMode(AM);

    // There are no SMRD extloads, so if we have to do a small type access we
    // will use a MUBUF load.
    // FIXME?: We also need to do this if unaligned, but we don't know the
    // alignment here.
    if (DL.getTypeStoreSize(Ty) < 4)
      return isLegalMUBUFAddressingMode(AM);

    if (Subtarget->getGeneration() == AMDGPUSubtarget::SOUTHERN_ISLANDS) {
      // SMRD instructions have an 8-bit, dword offset on SI.
      if (!isUInt<8>(AM.BaseOffs / 4))
        return false;
    } else if (Subtarget->getGeneration() == AMDGPUSubtarget::SEA_ISLANDS) {
      // On CI+, this can also be a 32-bit literal constant offset. If it fits
      // in 8-bits, it can use a smaller encoding.
      if (!isUInt<32>(AM.BaseOffs / 4))
        return false;
    } else if (Subtarget->getGeneration() == AMDGPUSubtarget::VOLCANIC_ISLANDS) {
      // On VI, these use the SMEM format and the offset is 20-bit in bytes.
      if (!isUInt<20>(AM.BaseOffs))
        return false;
    } else
      llvm_unreachable("unhandled generation");

    if (AM.Scale == 0) // r + i or just i, depending on HasBaseReg.
      return true;

    if (AM.Scale == 1 && AM.HasBaseReg)
      return true;

    return false;
  }

  case AMDGPUAS::PRIVATE_ADDRESS:
  case AMDGPUAS::UNKNOWN_ADDRESS_SPACE:
    return isLegalMUBUFAddressingMode(AM);

  case AMDGPUAS::LOCAL_ADDRESS:
  case AMDGPUAS::REGION_ADDRESS: {
    // Basic, single offset DS instructions allow a 16-bit unsigned immediate
    // field.
    // XXX - If doing a 4-byte aligned 8-byte type access, we effectively have
    // an 8-bit dword offset but we don't know the alignment here.
    if (!isUInt<16>(AM.BaseOffs))
      return false;

    if (AM.Scale == 0) // r + i or just i, depending on HasBaseReg.
      return true;

    if (AM.Scale == 1 && AM.HasBaseReg)
      return true;

    return false;
  }
  case AMDGPUAS::FLAT_ADDRESS:
    return isLegalFlatAddressingMode(AM);

  default:
    llvm_unreachable("unhandled address space");
  }
}

bool SITargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
                                                      unsigned AddrSpace,
                                                      unsigned Align,
                                                      bool *IsFast) const {
  if (IsFast)
    *IsFast = false;

  // TODO: I think v3i32 should allow unaligned accesses on CI with DS_READ_B96,
  // which isn't a simple VT.
  if (!VT.isSimple() || VT == MVT::Other)
    return false;

  // TODO - CI+ supports unaligned memory accesses, but this requires driver
  // support.

  // XXX - The only mention I see of this in the ISA manual is for LDS direct
  // reads the "byte address and must be dword aligned". Is it also true for the
  // normal loads and stores?
  if (AddrSpace == AMDGPUAS::LOCAL_ADDRESS) {
    // ds_read/write_b64 require 8-byte alignment, but we can do a 4 byte
    // aligned, 8 byte access in a single operation using ds_read2/write2_b32
    // with adjacent offsets.
    bool AlignedBy4 = (Align % 4 == 0);
    if (IsFast)
      *IsFast = AlignedBy4;
    return AlignedBy4;
  }

  // Smaller than dword value must be aligned.
  // FIXME: This should be allowed on CI+
  if (VT.bitsLT(MVT::i32))
    return false;

  // 8.1.6 - For Dword or larger reads or writes, the two LSBs of the
  // byte-address are ignored, thus forcing Dword alignment.
  // This applies to private, global, and constant memory.
  if (IsFast)
    *IsFast = true;

  return VT.bitsGT(MVT::i32) && Align % 4 == 0;
}

EVT SITargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign,
                                          unsigned SrcAlign, bool IsMemset,
                                          bool ZeroMemset,
                                          bool MemcpyStrSrc,
                                          MachineFunction &MF) const {
  // FIXME: Should account for address space here.

  // The default fallback uses the private pointer size as a guess for a type to
  // use. Make sure we switch these to 64-bit accesses.

  if (Size >= 16 && DstAlign >= 4) // XXX: Should only do for global
    return MVT::v4i32;

  if (Size >= 8 && DstAlign >= 4)
    return MVT::v2i32;

  // Use the default.
  return MVT::Other;
}

static bool isFlatGlobalAddrSpace(unsigned AS) {
  return AS == AMDGPUAS::GLOBAL_ADDRESS ||
    AS == AMDGPUAS::FLAT_ADDRESS ||
    AS == AMDGPUAS::CONSTANT_ADDRESS;
}

bool SITargetLowering::isNoopAddrSpaceCast(unsigned SrcAS,
                                           unsigned DestAS) const {
  return isFlatGlobalAddrSpace(SrcAS) &&  isFlatGlobalAddrSpace(DestAS);
}


bool SITargetLowering::isMemOpUniform(const SDNode *N) const {
  const MemSDNode *MemNode = cast<MemSDNode>(N);
  const Value *Ptr = MemNode->getMemOperand()->getValue();

  // UndefValue means this is a load of a kernel input.  These are uniform.
  // Sometimes LDS instructions have constant pointers
  if (isa<UndefValue>(Ptr) || isa<Argument>(Ptr) || isa<Constant>(Ptr) ||
      isa<GlobalValue>(Ptr))
    return true;

  const Instruction *I = dyn_cast_or_null<Instruction>(Ptr);
  return I && I->getMetadata("amdgpu.uniform");
}

TargetLoweringBase::LegalizeTypeAction
SITargetLowering::getPreferredVectorAction(EVT VT) const {
  if (VT.getVectorNumElements() != 1 && VT.getScalarType().bitsLE(MVT::i16))
    return TypeSplitVector;

  return TargetLoweringBase::getPreferredVectorAction(VT);
}

bool SITargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
                                                         Type *Ty) const {
  const SIInstrInfo *TII =
      static_cast<const SIInstrInfo *>(Subtarget->getInstrInfo());
  return TII->isInlineConstant(Imm);
}

SDValue SITargetLowering::LowerParameter(SelectionDAG &DAG, EVT VT, EVT MemVT,
                                         SDLoc SL, SDValue Chain,
                                         unsigned Offset, bool Signed) const {
  const DataLayout &DL = DAG.getDataLayout();
  MachineFunction &MF = DAG.getMachineFunction();
  const SIRegisterInfo *TRI =
      static_cast<const SIRegisterInfo*>(Subtarget->getRegisterInfo());
  unsigned InputPtrReg = TRI->getPreloadedValue(MF, SIRegisterInfo::KERNARG_SEGMENT_PTR);

  Type *Ty = VT.getTypeForEVT(*DAG.getContext());

  MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
  MVT PtrVT = getPointerTy(DL, AMDGPUAS::CONSTANT_ADDRESS);
  PointerType *PtrTy = PointerType::get(Ty, AMDGPUAS::CONSTANT_ADDRESS);
  SDValue BasePtr = DAG.getCopyFromReg(Chain, SL,
                                       MRI.getLiveInVirtReg(InputPtrReg), PtrVT);
  SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT, BasePtr,
                            DAG.getConstant(Offset, SL, PtrVT));
  SDValue PtrOffset = DAG.getUNDEF(PtrVT);
  MachinePointerInfo PtrInfo(UndefValue::get(PtrTy));

  unsigned Align = DL.getABITypeAlignment(Ty);

  ISD::LoadExtType ExtTy = Signed ? ISD::SEXTLOAD : ISD::ZEXTLOAD;
  if (MemVT.isFloatingPoint())
    ExtTy = ISD::EXTLOAD;

  return DAG.getLoad(ISD::UNINDEXED, ExtTy,
                     VT, SL, Chain, Ptr, PtrOffset, PtrInfo, MemVT,
                     false, // isVolatile
                     true, // isNonTemporal
                     true, // isInvariant
                     Align); // Alignment
}

SDValue SITargetLowering::LowerFormalArguments(
    SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
    const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
    SmallVectorImpl<SDValue> &InVals) const {
  const SIRegisterInfo *TRI =
      static_cast<const SIRegisterInfo *>(Subtarget->getRegisterInfo());

  MachineFunction &MF = DAG.getMachineFunction();
  FunctionType *FType = MF.getFunction()->getFunctionType();
  SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
  const AMDGPUSubtarget &ST = MF.getSubtarget<AMDGPUSubtarget>();

  if (Subtarget->isAmdHsaOS() && Info->getShaderType() != ShaderType::COMPUTE) {
    const Function *Fn = MF.getFunction();
    DiagnosticInfoUnsupported NoGraphicsHSA(*Fn, "non-compute shaders with HSA");
    DAG.getContext()->diagnose(NoGraphicsHSA);
    return SDValue();
  }

  // FIXME: We currently assume all calling conventions are kernels.

  SmallVector<ISD::InputArg, 16> Splits;
  BitVector Skipped(Ins.size());

  for (unsigned i = 0, e = Ins.size(), PSInputNum = 0; i != e; ++i) {
    const ISD::InputArg &Arg = Ins[i];

    // First check if it's a PS input addr
    if (Info->getShaderType() == ShaderType::PIXEL && !Arg.Flags.isInReg() &&
        !Arg.Flags.isByVal() && PSInputNum <= 15) {

      if (!Arg.Used && !Info->isPSInputAllocated(PSInputNum)) {
        // We can safely skip PS inputs
        Skipped.set(i);
        ++PSInputNum;
        continue;
      }

      Info->markPSInputAllocated(PSInputNum);
      if (Arg.Used)
        Info->PSInputEna |= 1 << PSInputNum;

      ++PSInputNum;
    }

    // Second split vertices into their elements
    if (Info->getShaderType() != ShaderType::COMPUTE && Arg.VT.isVector()) {
      ISD::InputArg NewArg = Arg;
      NewArg.Flags.setSplit();
      NewArg.VT = Arg.VT.getVectorElementType();

      // We REALLY want the ORIGINAL number of vertex elements here, e.g. a
      // three or five element vertex only needs three or five registers,
      // NOT four or eight.
      Type *ParamType = FType->getParamType(Arg.getOrigArgIndex());
      unsigned NumElements = ParamType->getVectorNumElements();

      for (unsigned j = 0; j != NumElements; ++j) {
        Splits.push_back(NewArg);
        NewArg.PartOffset += NewArg.VT.getStoreSize();
      }

    } else if (Info->getShaderType() != ShaderType::COMPUTE) {
      Splits.push_back(Arg);
    }
  }

  SmallVector<CCValAssign, 16> ArgLocs;
  CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
                 *DAG.getContext());

  // At least one interpolation mode must be enabled or else the GPU will hang.
  //
  // Check PSInputAddr instead of PSInputEna. The idea is that if the user set
  // PSInputAddr, the user wants to enable some bits after the compilation
  // based on run-time states. Since we can't know what the final PSInputEna
  // will look like, so we shouldn't do anything here and the user should take
  // responsibility for the correct programming.
  //
  // Otherwise, the following restrictions apply:
  // - At least one of PERSP_* (0xF) or LINEAR_* (0x70) must be enabled.
  // - If POS_W_FLOAT (11) is enabled, at least one of PERSP_* must be
  //   enabled too.
  if (Info->getShaderType() == ShaderType::PIXEL &&
      ((Info->getPSInputAddr() & 0x7F) == 0 ||
       ((Info->getPSInputAddr() & 0xF) == 0 &&
	Info->isPSInputAllocated(11)))) {
    CCInfo.AllocateReg(AMDGPU::VGPR0);
    CCInfo.AllocateReg(AMDGPU::VGPR1);
    Info->markPSInputAllocated(0);
    Info->PSInputEna |= 1;
  }

  if (Info->getShaderType() == ShaderType::COMPUTE) {
    getOriginalFunctionArgs(DAG, DAG.getMachineFunction().getFunction(), Ins,
                            Splits);
  }

  // FIXME: How should these inputs interact with inreg / custom SGPR inputs?
  if (Info->hasPrivateSegmentBuffer()) {
    unsigned PrivateSegmentBufferReg = Info->addPrivateSegmentBuffer(*TRI);
    MF.addLiveIn(PrivateSegmentBufferReg, &AMDGPU::SReg_128RegClass);
    CCInfo.AllocateReg(PrivateSegmentBufferReg);
  }

  if (Info->hasDispatchPtr()) {
    unsigned DispatchPtrReg = Info->addDispatchPtr(*TRI);
    MF.addLiveIn(DispatchPtrReg, &AMDGPU::SReg_64RegClass);
    CCInfo.AllocateReg(DispatchPtrReg);
  }

  if (Info->hasKernargSegmentPtr()) {
    unsigned InputPtrReg = Info->addKernargSegmentPtr(*TRI);
    MF.addLiveIn(InputPtrReg, &AMDGPU::SReg_64RegClass);
    CCInfo.AllocateReg(InputPtrReg);
  }

  if (Info->hasFlatScratchInit()) {
    unsigned FlatScratchInitReg = Info->addFlatScratchInit(*TRI);
    MF.addLiveIn(FlatScratchInitReg, &AMDGPU::SReg_64RegClass);
    CCInfo.AllocateReg(FlatScratchInitReg);
  }

  AnalyzeFormalArguments(CCInfo, Splits);

  SmallVector<SDValue, 16> Chains;

  for (unsigned i = 0, e = Ins.size(), ArgIdx = 0; i != e; ++i) {

    const ISD::InputArg &Arg = Ins[i];
    if (Skipped[i]) {
      InVals.push_back(DAG.getUNDEF(Arg.VT));
      continue;
    }

    CCValAssign &VA = ArgLocs[ArgIdx++];
    MVT VT = VA.getLocVT();

    if (VA.isMemLoc()) {
      VT = Ins[i].VT;
      EVT MemVT = Splits[i].VT;
      const unsigned Offset = Subtarget->getExplicitKernelArgOffset() +
                              VA.getLocMemOffset();
      // The first 36 bytes of the input buffer contains information about
      // thread group and global sizes.
      SDValue Arg = LowerParameter(DAG, VT, MemVT,  DL, Chain,
                                   Offset, Ins[i].Flags.isSExt());
      Chains.push_back(Arg.getValue(1));

      auto *ParamTy =
        dyn_cast<PointerType>(FType->getParamType(Ins[i].getOrigArgIndex()));
      if (Subtarget->getGeneration() == AMDGPUSubtarget::SOUTHERN_ISLANDS &&
          ParamTy && ParamTy->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS) {
        // On SI local pointers are just offsets into LDS, so they are always
        // less than 16-bits.  On CI and newer they could potentially be
        // real pointers, so we can't guarantee their size.
        Arg = DAG.getNode(ISD::AssertZext, DL, Arg.getValueType(), Arg,
                          DAG.getValueType(MVT::i16));
      }

      InVals.push_back(Arg);
      Info->ABIArgOffset = Offset + MemVT.getStoreSize();
      continue;
    }
    assert(VA.isRegLoc() && "Parameter must be in a register!");

    unsigned Reg = VA.getLocReg();

    if (VT == MVT::i64) {
      // For now assume it is a pointer
      Reg = TRI->getMatchingSuperReg(Reg, AMDGPU::sub0,
                                     &AMDGPU::SReg_64RegClass);
      Reg = MF.addLiveIn(Reg, &AMDGPU::SReg_64RegClass);
      SDValue Copy = DAG.getCopyFromReg(Chain, DL, Reg, VT);
      InVals.push_back(Copy);
      continue;
    }

    const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg, VT);

    Reg = MF.addLiveIn(Reg, RC);
    SDValue Val = DAG.getCopyFromReg(Chain, DL, Reg, VT);

    if (Arg.VT.isVector()) {

      // Build a vector from the registers
      Type *ParamType = FType->getParamType(Arg.getOrigArgIndex());
      unsigned NumElements = ParamType->getVectorNumElements();

      SmallVector<SDValue, 4> Regs;
      Regs.push_back(Val);
      for (unsigned j = 1; j != NumElements; ++j) {
        Reg = ArgLocs[ArgIdx++].getLocReg();
        Reg = MF.addLiveIn(Reg, RC);

        SDValue Copy = DAG.getCopyFromReg(Chain, DL, Reg, VT);
        Regs.push_back(Copy);
      }

      // Fill up the missing vector elements
      NumElements = Arg.VT.getVectorNumElements() - NumElements;
      Regs.append(NumElements, DAG.getUNDEF(VT));

      InVals.push_back(DAG.getNode(ISD::BUILD_VECTOR, DL, Arg.VT, Regs));
      continue;
    }

    InVals.push_back(Val);
  }

  // TODO: Add GridWorkGroupCount user SGPRs when used. For now with HSA we read
  // these from the dispatch pointer.

  // Start adding system SGPRs.
  if (Info->hasWorkGroupIDX()) {
    unsigned Reg = Info->addWorkGroupIDX();
    MF.addLiveIn(Reg, &AMDGPU::SReg_32RegClass);
    CCInfo.AllocateReg(Reg);
  } else
    llvm_unreachable("work group id x is always enabled");

  if (Info->hasWorkGroupIDY()) {
    unsigned Reg = Info->addWorkGroupIDY();
    MF.addLiveIn(Reg, &AMDGPU::SReg_32RegClass);
    CCInfo.AllocateReg(Reg);
  }

  if (Info->hasWorkGroupIDZ()) {
    unsigned Reg = Info->addWorkGroupIDZ();
    MF.addLiveIn(Reg, &AMDGPU::SReg_32RegClass);
    CCInfo.AllocateReg(Reg);
  }

  if (Info->hasWorkGroupInfo()) {
    unsigned Reg = Info->addWorkGroupInfo();
    MF.addLiveIn(Reg, &AMDGPU::SReg_32RegClass);
    CCInfo.AllocateReg(Reg);
  }

  if (Info->hasPrivateSegmentWaveByteOffset()) {
    // Scratch wave offset passed in system SGPR.
    unsigned PrivateSegmentWaveByteOffsetReg
      = Info->addPrivateSegmentWaveByteOffset();

    MF.addLiveIn(PrivateSegmentWaveByteOffsetReg, &AMDGPU::SGPR_32RegClass);
    CCInfo.AllocateReg(PrivateSegmentWaveByteOffsetReg);
  }

  // Now that we've figured out where the scratch register inputs are, see if
  // should reserve the arguments and use them directly.
  bool HasStackObjects = MF.getFrameInfo()->hasStackObjects();
  // Record that we know we have non-spill stack objects so we don't need to
  // check all stack objects later.
  if (HasStackObjects)
    Info->setHasNonSpillStackObjects(true);

  if (ST.isAmdHsaOS()) {
    // TODO: Assume we will spill without optimizations.
    if (HasStackObjects) {
      // If we have stack objects, we unquestionably need the private buffer
      // resource. For the HSA ABI, this will be the first 4 user SGPR
      // inputs. We can reserve those and use them directly.

      unsigned PrivateSegmentBufferReg = TRI->getPreloadedValue(
        MF, SIRegisterInfo::PRIVATE_SEGMENT_BUFFER);
      Info->setScratchRSrcReg(PrivateSegmentBufferReg);

      unsigned PrivateSegmentWaveByteOffsetReg = TRI->getPreloadedValue(
        MF, SIRegisterInfo::PRIVATE_SEGMENT_WAVE_BYTE_OFFSET);
      Info->setScratchWaveOffsetReg(PrivateSegmentWaveByteOffsetReg);
    } else {
      unsigned ReservedBufferReg
        = TRI->reservedPrivateSegmentBufferReg(MF);
      unsigned ReservedOffsetReg
        = TRI->reservedPrivateSegmentWaveByteOffsetReg(MF);

      // We tentatively reserve the last registers (skipping the last two
      // which may contain VCC). After register allocation, we'll replace
      // these with the ones immediately after those which were really
      // allocated. In the prologue copies will be inserted from the argument
      // to these reserved registers.
      Info->setScratchRSrcReg(ReservedBufferReg);
      Info->setScratchWaveOffsetReg(ReservedOffsetReg);
    }
  } else {
    unsigned ReservedBufferReg = TRI->reservedPrivateSegmentBufferReg(MF);

    // Without HSA, relocations are used for the scratch pointer and the
    // buffer resource setup is always inserted in the prologue. Scratch wave
    // offset is still in an input SGPR.
    Info->setScratchRSrcReg(ReservedBufferReg);

    if (HasStackObjects) {
      unsigned ScratchWaveOffsetReg = TRI->getPreloadedValue(
        MF, SIRegisterInfo::PRIVATE_SEGMENT_WAVE_BYTE_OFFSET);
      Info->setScratchWaveOffsetReg(ScratchWaveOffsetReg);
    } else {
      unsigned ReservedOffsetReg
        = TRI->reservedPrivateSegmentWaveByteOffsetReg(MF);
      Info->setScratchWaveOffsetReg(ReservedOffsetReg);
    }
  }

  if (Info->hasWorkItemIDX()) {
    unsigned Reg = TRI->getPreloadedValue(MF, SIRegisterInfo::WORKITEM_ID_X);
    MF.addLiveIn(Reg, &AMDGPU::VGPR_32RegClass);
    CCInfo.AllocateReg(Reg);
  } else
    llvm_unreachable("workitem id x should always be enabled");

  if (Info->hasWorkItemIDY()) {
    unsigned Reg = TRI->getPreloadedValue(MF, SIRegisterInfo::WORKITEM_ID_Y);
    MF.addLiveIn(Reg, &AMDGPU::VGPR_32RegClass);
    CCInfo.AllocateReg(Reg);
  }

  if (Info->hasWorkItemIDZ()) {
    unsigned Reg = TRI->getPreloadedValue(MF, SIRegisterInfo::WORKITEM_ID_Z);
    MF.addLiveIn(Reg, &AMDGPU::VGPR_32RegClass);
    CCInfo.AllocateReg(Reg);
  }

  if (Chains.empty())
    return Chain;

  return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
}

SDValue SITargetLowering::LowerReturn(SDValue Chain,
                                      CallingConv::ID CallConv,
                                      bool isVarArg,
                                      const SmallVectorImpl<ISD::OutputArg> &Outs,
                                      const SmallVectorImpl<SDValue> &OutVals,
                                      SDLoc DL, SelectionDAG &DAG) const {
  MachineFunction &MF = DAG.getMachineFunction();
  SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();

  if (Info->getShaderType() == ShaderType::COMPUTE)
    return AMDGPUTargetLowering::LowerReturn(Chain, CallConv, isVarArg, Outs,
                                             OutVals, DL, DAG);

  Info->setIfReturnsVoid(Outs.size() == 0);

  SmallVector<ISD::OutputArg, 48> Splits;
  SmallVector<SDValue, 48> SplitVals;

  // Split vectors into their elements.
  for (unsigned i = 0, e = Outs.size(); i != e; ++i) {
    const ISD::OutputArg &Out = Outs[i];

    if (Out.VT.isVector()) {
      MVT VT = Out.VT.getVectorElementType();
      ISD::OutputArg NewOut = Out;
      NewOut.Flags.setSplit();
      NewOut.VT = VT;

      // We want the original number of vector elements here, e.g.
      // three or five, not four or eight.
      unsigned NumElements = Out.ArgVT.getVectorNumElements();

      for (unsigned j = 0; j != NumElements; ++j) {
        SDValue Elem = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, OutVals[i],
                                   DAG.getConstant(j, DL, MVT::i32));
        SplitVals.push_back(Elem);
        Splits.push_back(NewOut);
        NewOut.PartOffset += NewOut.VT.getStoreSize();
      }
    } else {
      SplitVals.push_back(OutVals[i]);
      Splits.push_back(Out);
    }
  }

  // CCValAssign - represent the assignment of the return value to a location.
  SmallVector<CCValAssign, 48> RVLocs;

  // CCState - Info about the registers and stack slots.
  CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
                 *DAG.getContext());

  // Analyze outgoing return values.
  AnalyzeReturn(CCInfo, Splits);

  SDValue Flag;
  SmallVector<SDValue, 48> RetOps;
  RetOps.push_back(Chain); // Operand #0 = Chain (updated below)

  // Copy the result values into the output registers.
  for (unsigned i = 0, realRVLocIdx = 0;
       i != RVLocs.size();
       ++i, ++realRVLocIdx) {
    CCValAssign &VA = RVLocs[i];
    assert(VA.isRegLoc() && "Can only return in registers!");

    SDValue Arg = SplitVals[realRVLocIdx];

    // Copied from other backends.
    switch (VA.getLocInfo()) {
    default: llvm_unreachable("Unknown loc info!");
    case CCValAssign::Full:
      break;
    case CCValAssign::BCvt:
      Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
      break;
    }

    Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Arg, Flag);
    Flag = Chain.getValue(1);
    RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
  }

  // Update chain and glue.
  RetOps[0] = Chain;
  if (Flag.getNode())
    RetOps.push_back(Flag);

  return DAG.getNode(AMDGPUISD::RET_FLAG, DL, MVT::Other, RetOps);
}

MachineBasicBlock * SITargetLowering::EmitInstrWithCustomInserter(
    MachineInstr * MI, MachineBasicBlock * BB) const {

  switch (MI->getOpcode()) {
  default:
    return AMDGPUTargetLowering::EmitInstrWithCustomInserter(MI, BB);
  case AMDGPU::BRANCH:
    return BB;
  }
  return BB;
}

bool SITargetLowering::enableAggressiveFMAFusion(EVT VT) const {
  // This currently forces unfolding various combinations of fsub into fma with
  // free fneg'd operands. As long as we have fast FMA (controlled by
  // isFMAFasterThanFMulAndFAdd), we should perform these.

  // When fma is quarter rate, for f64 where add / sub are at best half rate,
  // most of these combines appear to be cycle neutral but save on instruction
  // count / code size.
  return true;
}

EVT SITargetLowering::getSetCCResultType(const DataLayout &DL, LLVMContext &Ctx,
                                         EVT VT) const {
  if (!VT.isVector()) {
    return MVT::i1;
  }
  return EVT::getVectorVT(Ctx, MVT::i1, VT.getVectorNumElements());
}

MVT SITargetLowering::getScalarShiftAmountTy(const DataLayout &, EVT) const {
  return MVT::i32;
}

// Answering this is somewhat tricky and depends on the specific device which
// have different rates for fma or all f64 operations.
//
// v_fma_f64 and v_mul_f64 always take the same number of cycles as each other
// regardless of which device (although the number of cycles differs between
// devices), so it is always profitable for f64.
//
// v_fma_f32 takes 4 or 16 cycles depending on the device, so it is profitable
// only on full rate devices. Normally, we should prefer selecting v_mad_f32
// which we can always do even without fused FP ops since it returns the same
// result as the separate operations and since it is always full
// rate. Therefore, we lie and report that it is not faster for f32. v_mad_f32
// however does not support denormals, so we do report fma as faster if we have
// a fast fma device and require denormals.
//
bool SITargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
  VT = VT.getScalarType();

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

  switch (VT.getSimpleVT().SimpleTy) {
  case MVT::f32:
    // This is as fast on some subtargets. However, we always have full rate f32
    // mad available which returns the same result as the separate operations
    // which we should prefer over fma. We can't use this if we want to support
    // denormals, so only report this in these cases.
    return Subtarget->hasFP32Denormals() && Subtarget->hasFastFMAF32();
  case MVT::f64:
    return true;
  default:
    break;
  }

  return false;
}

//===----------------------------------------------------------------------===//
// Custom DAG Lowering Operations
//===----------------------------------------------------------------------===//

SDValue SITargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
  switch (Op.getOpcode()) {
  default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
  case ISD::FrameIndex: return LowerFrameIndex(Op, DAG);
  case ISD::BRCOND: return LowerBRCOND(Op, DAG);
  case ISD::LOAD: {
    SDValue Result = LowerLOAD(Op, DAG);
    assert((!Result.getNode() ||
            Result.getNode()->getNumValues() == 2) &&
           "Load should return a value and a chain");
    return Result;
  }

  case ISD::FSIN:
  case ISD::FCOS:
    return LowerTrig(Op, DAG);
  case ISD::SELECT: return LowerSELECT(Op, DAG);
  case ISD::FDIV: return LowerFDIV(Op, DAG);
  case ISD::STORE: return LowerSTORE(Op, DAG);
  case ISD::GlobalAddress: {
    MachineFunction &MF = DAG.getMachineFunction();
    SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
    return LowerGlobalAddress(MFI, Op, DAG);
  }
  case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
  case ISD::INTRINSIC_VOID: return LowerINTRINSIC_VOID(Op, DAG);
  }
  return SDValue();
}

/// \brief Helper function for LowerBRCOND
static SDNode *findUser(SDValue Value, unsigned Opcode) {

  SDNode *Parent = Value.getNode();
  for (SDNode::use_iterator I = Parent->use_begin(), E = Parent->use_end();
       I != E; ++I) {

    if (I.getUse().get() != Value)
      continue;

    if (I->getOpcode() == Opcode)
      return *I;
  }
  return nullptr;
}

SDValue SITargetLowering::LowerFrameIndex(SDValue Op, SelectionDAG &DAG) const {

  SDLoc SL(Op);
  FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Op);
  unsigned FrameIndex = FINode->getIndex();

  // A FrameIndex node represents a 32-bit offset into scratch memory.  If
  // the high bit of a frame index offset were to be set, this would mean
  // that it represented an offset of ~2GB * 64 = ~128GB from the start of the
  // scratch buffer, with 64 being the number of threads per wave.
  //
  // If we know the machine uses less than 128GB of scratch, then we can
  // amrk the high bit of the FrameIndex node as known zero,
  // which is important, because it means in most situations we can
  // prove that values derived from FrameIndex nodes are non-negative.
  // This enables us to take advantage of more addressing modes when
  // accessing scratch buffers, since for scratch reads/writes, the register
  // offset must always be positive.

  SDValue TFI = DAG.getTargetFrameIndex(FrameIndex, MVT::i32);
  if (Subtarget->enableHugeScratchBuffer())
    return TFI;

  return DAG.getNode(ISD::AssertZext, SL, MVT::i32, TFI,
                    DAG.getValueType(EVT::getIntegerVT(*DAG.getContext(), 31)));
}

/// This transforms the control flow intrinsics to get the branch destination as
/// last parameter, also switches branch target with BR if the need arise
SDValue SITargetLowering::LowerBRCOND(SDValue BRCOND,
                                      SelectionDAG &DAG) const {

  SDLoc DL(BRCOND);

  SDNode *Intr = BRCOND.getOperand(1).getNode();
  SDValue Target = BRCOND.getOperand(2);
  SDNode *BR = nullptr;

  if (Intr->getOpcode() == ISD::SETCC) {
    // As long as we negate the condition everything is fine
    SDNode *SetCC = Intr;
    assert(SetCC->getConstantOperandVal(1) == 1);
    assert(cast<CondCodeSDNode>(SetCC->getOperand(2).getNode())->get() ==
           ISD::SETNE);
    Intr = SetCC->getOperand(0).getNode();

  } else {
    // Get the target from BR if we don't negate the condition
    BR = findUser(BRCOND, ISD::BR);
    Target = BR->getOperand(1);
  }

  assert(Intr->getOpcode() == ISD::INTRINSIC_W_CHAIN);

  // Build the result and
  ArrayRef<EVT> Res(Intr->value_begin() + 1, Intr->value_end());

  // operands of the new intrinsic call
  SmallVector<SDValue, 4> Ops;
  Ops.push_back(BRCOND.getOperand(0));
  Ops.append(Intr->op_begin() + 1, Intr->op_end());
  Ops.push_back(Target);

  // build the new intrinsic call
  SDNode *Result = DAG.getNode(
    Res.size() > 1 ? ISD::INTRINSIC_W_CHAIN : ISD::INTRINSIC_VOID, DL,
    DAG.getVTList(Res), Ops).getNode();

  if (BR) {
    // Give the branch instruction our target
    SDValue Ops[] = {
      BR->getOperand(0),
      BRCOND.getOperand(2)
    };
    SDValue NewBR = DAG.getNode(ISD::BR, DL, BR->getVTList(), Ops);
    DAG.ReplaceAllUsesWith(BR, NewBR.getNode());
    BR = NewBR.getNode();
  }

  SDValue Chain = SDValue(Result, Result->getNumValues() - 1);

  // Copy the intrinsic results to registers
  for (unsigned i = 1, e = Intr->getNumValues() - 1; i != e; ++i) {
    SDNode *CopyToReg = findUser(SDValue(Intr, i), ISD::CopyToReg);
    if (!CopyToReg)
      continue;

    Chain = DAG.getCopyToReg(
      Chain, DL,
      CopyToReg->getOperand(1),
      SDValue(Result, i - 1),
      SDValue());

    DAG.ReplaceAllUsesWith(SDValue(CopyToReg, 0), CopyToReg->getOperand(0));
  }

  // Remove the old intrinsic from the chain
  DAG.ReplaceAllUsesOfValueWith(
    SDValue(Intr, Intr->getNumValues() - 1),
    Intr->getOperand(0));

  return Chain;
}

SDValue SITargetLowering::LowerGlobalAddress(AMDGPUMachineFunction *MFI,
                                             SDValue Op,
                                             SelectionDAG &DAG) const {
  GlobalAddressSDNode *GSD = cast<GlobalAddressSDNode>(Op);

  if (GSD->getAddressSpace() != AMDGPUAS::CONSTANT_ADDRESS)
    return AMDGPUTargetLowering::LowerGlobalAddress(MFI, Op, DAG);

  SDLoc DL(GSD);
  const GlobalValue *GV = GSD->getGlobal();
  MVT PtrVT = getPointerTy(DAG.getDataLayout(), GSD->getAddressSpace());

  SDValue GA = DAG.getTargetGlobalAddress(GV, DL, MVT::i32);
  return DAG.getNode(AMDGPUISD::CONST_DATA_PTR, DL, PtrVT, GA);
}

SDValue SITargetLowering::copyToM0(SelectionDAG &DAG, SDValue Chain, SDLoc DL,
                                   SDValue V) const {
  // We can't use CopyToReg, because MachineCSE won't combine COPY instructions,
  // so we will end up with redundant moves to m0.
  //
  // We can't use S_MOV_B32, because there is no way to specify m0 as the
  // destination register.
  //
  // We have to use them both.  Machine cse will combine all the S_MOV_B32
  // instructions and the register coalescer eliminate the extra copies.
  SDNode *M0 = DAG.getMachineNode(AMDGPU::S_MOV_B32, DL, V.getValueType(), V);
  return DAG.getCopyToReg(Chain, DL, DAG.getRegister(AMDGPU::M0, MVT::i32),
                          SDValue(M0, 0), SDValue()); // Glue
                                                      // A Null SDValue creates
                                                      // a glue result.
}

SDValue SITargetLowering::lowerImplicitZextParam(SelectionDAG &DAG,
                                                 SDValue Op,
                                                 MVT VT,
                                                 unsigned Offset) const {
  SDLoc SL(Op);
  SDValue Param = LowerParameter(DAG, MVT::i32, MVT::i32, SL,
                                 DAG.getEntryNode(), Offset, false);
  // The local size values will have the hi 16-bits as zero.
  return DAG.getNode(ISD::AssertZext, SL, MVT::i32, Param,
                     DAG.getValueType(VT));
}

SDValue SITargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
                                                  SelectionDAG &DAG) const {
  MachineFunction &MF = DAG.getMachineFunction();
  auto MFI = MF.getInfo<SIMachineFunctionInfo>();
  const SIRegisterInfo *TRI =
      static_cast<const SIRegisterInfo *>(Subtarget->getRegisterInfo());

  EVT VT = Op.getValueType();
  SDLoc DL(Op);
  unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();

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

  switch (IntrinsicID) {
  case Intrinsic::amdgcn_dispatch_ptr:
    if (!Subtarget->isAmdHsaOS()) {
      DiagnosticInfoUnsupported BadIntrin(*MF.getFunction(),
                                          "hsa intrinsic without hsa target");
      DAG.getContext()->diagnose(BadIntrin);
      return DAG.getUNDEF(VT);
    }

    return CreateLiveInRegister(DAG, &AMDGPU::SReg_64RegClass,
      TRI->getPreloadedValue(MF, SIRegisterInfo::DISPATCH_PTR), VT);

  case Intrinsic::r600_read_ngroups_x:
    return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
                          SI::KernelInputOffsets::NGROUPS_X, false);
  case Intrinsic::r600_read_ngroups_y:
    return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
                          SI::KernelInputOffsets::NGROUPS_Y, false);
  case Intrinsic::r600_read_ngroups_z:
    return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
                          SI::KernelInputOffsets::NGROUPS_Z, false);
  case Intrinsic::r600_read_global_size_x:
    return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
                          SI::KernelInputOffsets::GLOBAL_SIZE_X, false);
  case Intrinsic::r600_read_global_size_y:
    return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
                          SI::KernelInputOffsets::GLOBAL_SIZE_Y, false);
  case Intrinsic::r600_read_global_size_z:
    return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
                          SI::KernelInputOffsets::GLOBAL_SIZE_Z, false);
  case Intrinsic::r600_read_local_size_x:
    return lowerImplicitZextParam(DAG, Op, MVT::i16,
                                  SI::KernelInputOffsets::LOCAL_SIZE_X);
  case Intrinsic::r600_read_local_size_y:
    return lowerImplicitZextParam(DAG, Op, MVT::i16,
                                  SI::KernelInputOffsets::LOCAL_SIZE_Y);
  case Intrinsic::r600_read_local_size_z:
    return lowerImplicitZextParam(DAG, Op, MVT::i16,
                                  SI::KernelInputOffsets::LOCAL_SIZE_Z);
  case Intrinsic::AMDGPU_read_workdim:
    // Really only 2 bits.
    return lowerImplicitZextParam(DAG, Op, MVT::i8,
                                  getImplicitParameterOffset(MFI, GRID_DIM));
  case Intrinsic::r600_read_tgid_x:
    return CreateLiveInRegister(DAG, &AMDGPU::SReg_32RegClass,
      TRI->getPreloadedValue(MF, SIRegisterInfo::WORKGROUP_ID_X), VT);
  case Intrinsic::r600_read_tgid_y:
    return CreateLiveInRegister(DAG, &AMDGPU::SReg_32RegClass,
      TRI->getPreloadedValue(MF, SIRegisterInfo::WORKGROUP_ID_Y), VT);
  case Intrinsic::r600_read_tgid_z:
    return CreateLiveInRegister(DAG, &AMDGPU::SReg_32RegClass,
      TRI->getPreloadedValue(MF, SIRegisterInfo::WORKGROUP_ID_Z), VT);
  case Intrinsic::r600_read_tidig_x:
    return CreateLiveInRegister(DAG, &AMDGPU::VGPR_32RegClass,
      TRI->getPreloadedValue(MF, SIRegisterInfo::WORKITEM_ID_X), VT);
  case Intrinsic::r600_read_tidig_y:
    return CreateLiveInRegister(DAG, &AMDGPU::VGPR_32RegClass,
      TRI->getPreloadedValue(MF, SIRegisterInfo::WORKITEM_ID_Y), VT);
  case Intrinsic::r600_read_tidig_z:
    return CreateLiveInRegister(DAG, &AMDGPU::VGPR_32RegClass,
      TRI->getPreloadedValue(MF, SIRegisterInfo::WORKITEM_ID_Z), VT);
  case AMDGPUIntrinsic::SI_load_const: {
    SDValue Ops[] = {
      Op.getOperand(1),
      Op.getOperand(2)
    };

    MachineMemOperand *MMO = MF.getMachineMemOperand(
      MachinePointerInfo(),
      MachineMemOperand::MOLoad | MachineMemOperand::MOInvariant,
      VT.getStoreSize(), 4);
    return DAG.getMemIntrinsicNode(AMDGPUISD::LOAD_CONSTANT, DL,
                                   Op->getVTList(), Ops, VT, MMO);
  }
  case AMDGPUIntrinsic::SI_sample:
    return LowerSampleIntrinsic(AMDGPUISD::SAMPLE, Op, DAG);
  case AMDGPUIntrinsic::SI_sampleb:
    return LowerSampleIntrinsic(AMDGPUISD::SAMPLEB, Op, DAG);
  case AMDGPUIntrinsic::SI_sampled:
    return LowerSampleIntrinsic(AMDGPUISD::SAMPLED, Op, DAG);
  case AMDGPUIntrinsic::SI_samplel:
    return LowerSampleIntrinsic(AMDGPUISD::SAMPLEL, Op, DAG);
  case AMDGPUIntrinsic::SI_vs_load_input:
    return DAG.getNode(AMDGPUISD::LOAD_INPUT, DL, VT,
                       Op.getOperand(1),
                       Op.getOperand(2),
                       Op.getOperand(3));

  case AMDGPUIntrinsic::AMDGPU_fract:
  case AMDGPUIntrinsic::AMDIL_fraction: // Legacy name.
    return DAG.getNode(ISD::FSUB, DL, VT, Op.getOperand(1),
                       DAG.getNode(ISD::FFLOOR, DL, VT, Op.getOperand(1)));
  case AMDGPUIntrinsic::SI_fs_constant: {
    SDValue M0 = copyToM0(DAG, DAG.getEntryNode(), DL, Op.getOperand(3));
    SDValue Glue = M0.getValue(1);
    return DAG.getNode(AMDGPUISD::INTERP_MOV, DL, MVT::f32,
                       DAG.getConstant(2, DL, MVT::i32), // P0
                       Op.getOperand(1), Op.getOperand(2), Glue);
  }
  case AMDGPUIntrinsic::SI_packf16:
    if (Op.getOperand(1).isUndef() && Op.getOperand(2).isUndef())
      return DAG.getUNDEF(MVT::i32);
    return Op;
  case AMDGPUIntrinsic::SI_fs_interp: {
    SDValue IJ = Op.getOperand(4);
    SDValue I = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, IJ,
                            DAG.getConstant(0, DL, MVT::i32));
    SDValue J = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, IJ,
                            DAG.getConstant(1, DL, MVT::i32));
    SDValue M0 = copyToM0(DAG, DAG.getEntryNode(), DL, Op.getOperand(3));
    SDValue Glue = M0.getValue(1);
    SDValue P1 = DAG.getNode(AMDGPUISD::INTERP_P1, DL,
                             DAG.getVTList(MVT::f32, MVT::Glue),
                             I, Op.getOperand(1), Op.getOperand(2), Glue);
    Glue = SDValue(P1.getNode(), 1);
    return DAG.getNode(AMDGPUISD::INTERP_P2, DL, MVT::f32, P1, J,
                             Op.getOperand(1), Op.getOperand(2), Glue);
  }
  case Intrinsic::amdgcn_interp_p1: {
    SDValue M0 = copyToM0(DAG, DAG.getEntryNode(), DL, Op.getOperand(4));
    SDValue Glue = M0.getValue(1);
    return DAG.getNode(AMDGPUISD::INTERP_P1, DL, MVT::f32, Op.getOperand(1),
                       Op.getOperand(2), Op.getOperand(3), Glue);
  }
  case Intrinsic::amdgcn_interp_p2: {
    SDValue M0 = copyToM0(DAG, DAG.getEntryNode(), DL, Op.getOperand(5));
    SDValue Glue = SDValue(M0.getNode(), 1);
    return DAG.getNode(AMDGPUISD::INTERP_P2, DL, MVT::f32, Op.getOperand(1),
                       Op.getOperand(2), Op.getOperand(3), Op.getOperand(4),
                       Glue);
  }
  default:
    return AMDGPUTargetLowering::LowerOperation(Op, DAG);
  }
}

SDValue SITargetLowering::LowerINTRINSIC_VOID(SDValue Op,
                                              SelectionDAG &DAG) const {
  MachineFunction &MF = DAG.getMachineFunction();
  SDLoc DL(Op);
  SDValue Chain = Op.getOperand(0);
  unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();

  switch (IntrinsicID) {
  case AMDGPUIntrinsic::SI_sendmsg: {
    Chain = copyToM0(DAG, Chain, DL, Op.getOperand(3));
    SDValue Glue = Chain.getValue(1);
    return DAG.getNode(AMDGPUISD::SENDMSG, DL, MVT::Other, Chain,
                       Op.getOperand(2), Glue);
  }
  case AMDGPUIntrinsic::SI_tbuffer_store: {
    SDValue Ops[] = {
      Chain,
      Op.getOperand(2),
      Op.getOperand(3),
      Op.getOperand(4),
      Op.getOperand(5),
      Op.getOperand(6),
      Op.getOperand(7),
      Op.getOperand(8),
      Op.getOperand(9),
      Op.getOperand(10),
      Op.getOperand(11),
      Op.getOperand(12),
      Op.getOperand(13),
      Op.getOperand(14)
    };

    EVT VT = Op.getOperand(3).getValueType();

    MachineMemOperand *MMO = MF.getMachineMemOperand(
      MachinePointerInfo(),
      MachineMemOperand::MOStore,
      VT.getStoreSize(), 4);
    return DAG.getMemIntrinsicNode(AMDGPUISD::TBUFFER_STORE_FORMAT, DL,
                                   Op->getVTList(), Ops, VT, MMO);
  }
  default:
    return SDValue();
  }
}

SDValue SITargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
  SDLoc DL(Op);
  LoadSDNode *Load = cast<LoadSDNode>(Op);

  if (Op.getValueType().isVector()) {
    assert(Op.getValueType().getVectorElementType() == MVT::i32 &&
           "Custom lowering for non-i32 vectors hasn't been implemented.");
    unsigned NumElements = Op.getValueType().getVectorNumElements();
    assert(NumElements != 2 && "v2 loads are supported for all address spaces.");

    switch (Load->getAddressSpace()) {
      default: break;
      case AMDGPUAS::CONSTANT_ADDRESS:
      if (isMemOpUniform(Load))
        break;
        // Non-uniform loads will be selected to MUBUF instructions, so they
        // have the same legalization requires ments as global and private
        // loads.
        //
        // Fall-through
      case AMDGPUAS::GLOBAL_ADDRESS:
      case AMDGPUAS::PRIVATE_ADDRESS:
        if (NumElements >= 8)
          return SplitVectorLoad(Op, DAG);

        // v4 loads are supported for private and global memory.
        if (NumElements <= 4)
          break;
        // fall-through
      case AMDGPUAS::LOCAL_ADDRESS:
        // If properly aligned, if we split we might be able to use ds_read_b64.
        return SplitVectorLoad(Op, DAG);
    }
  }

  return AMDGPUTargetLowering::LowerLOAD(Op, DAG);
}

SDValue SITargetLowering::LowerSampleIntrinsic(unsigned Opcode,
                                               const SDValue &Op,
                                               SelectionDAG &DAG) const {
  return DAG.getNode(Opcode, SDLoc(Op), Op.getValueType(), Op.getOperand(1),
                     Op.getOperand(2),
                     Op.getOperand(3),
                     Op.getOperand(4));
}

SDValue SITargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
  if (Op.getValueType() != MVT::i64)
    return SDValue();

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

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

  SDValue LHS = DAG.getNode(ISD::BITCAST, DL, MVT::v2i32, Op.getOperand(1));
  SDValue RHS = DAG.getNode(ISD::BITCAST, DL, MVT::v2i32, Op.getOperand(2));

  SDValue Lo0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, LHS, Zero);
  SDValue Lo1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, RHS, Zero);

  SDValue Lo = DAG.getSelect(DL, MVT::i32, Cond, Lo0, Lo1);

  SDValue Hi0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, LHS, One);
  SDValue Hi1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, RHS, One);

  SDValue Hi = DAG.getSelect(DL, MVT::i32, Cond, Hi0, Hi1);

  SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v2i32, Lo, Hi);
  return DAG.getNode(ISD::BITCAST, DL, MVT::i64, Res);
}

// Catch division cases where we can use shortcuts with rcp and rsq
// instructions.
SDValue SITargetLowering::LowerFastFDIV(SDValue Op, SelectionDAG &DAG) const {
  SDLoc SL(Op);
  SDValue LHS = Op.getOperand(0);
  SDValue RHS = Op.getOperand(1);
  EVT VT = Op.getValueType();
  bool Unsafe = DAG.getTarget().Options.UnsafeFPMath;

  if (const ConstantFPSDNode *CLHS = dyn_cast<ConstantFPSDNode>(LHS)) {
    if ((Unsafe || (VT == MVT::f32 && !Subtarget->hasFP32Denormals())) &&
        CLHS->isExactlyValue(1.0)) {
      // v_rcp_f32 and v_rsq_f32 do not support denormals, and according to
      // the CI documentation has a worst case error of 1 ulp.
      // OpenCL requires <= 2.5 ulp for 1.0 / x, so it should always be OK to
      // use it as long as we aren't trying to use denormals.

      // 1.0 / sqrt(x) -> rsq(x)
      //
      // XXX - Is UnsafeFPMath sufficient to do this for f64? The maximum ULP
      // error seems really high at 2^29 ULP.
      if (RHS.getOpcode() == ISD::FSQRT)
        return DAG.getNode(AMDGPUISD::RSQ, SL, VT, RHS.getOperand(0));

      // 1.0 / x -> rcp(x)
      return DAG.getNode(AMDGPUISD::RCP, SL, VT, RHS);
    }
  }

  if (Unsafe) {
    // Turn into multiply by the reciprocal.
    // x / y -> x * (1.0 / y)
    SDNodeFlags Flags;
    Flags.setUnsafeAlgebra(true);
    SDValue Recip = DAG.getNode(AMDGPUISD::RCP, SL, VT, RHS);
    return DAG.getNode(ISD::FMUL, SL, VT, LHS, Recip, &Flags);
  }

  return SDValue();
}

SDValue SITargetLowering::LowerFDIV32(SDValue Op, SelectionDAG &DAG) const {
  SDValue FastLowered = LowerFastFDIV(Op, DAG);
  if (FastLowered.getNode())
    return FastLowered;

  // This uses v_rcp_f32 which does not handle denormals. Let this hit a
  // selection error for now rather than do something incorrect.
  if (Subtarget->hasFP32Denormals())
    return SDValue();

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

  SDValue r1 = DAG.getNode(ISD::FABS, SL, MVT::f32, RHS);

  const APFloat K0Val(BitsToFloat(0x6f800000));
  const SDValue K0 = DAG.getConstantFP(K0Val, SL, MVT::f32);

  const APFloat K1Val(BitsToFloat(0x2f800000));
  const SDValue K1 = DAG.getConstantFP(K1Val, SL, MVT::f32);

  const SDValue One = DAG.getConstantFP(1.0, SL, MVT::f32);

  EVT SetCCVT =
      getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::f32);

  SDValue r2 = DAG.getSetCC(SL, SetCCVT, r1, K0, ISD::SETOGT);

  SDValue r3 = DAG.getNode(ISD::SELECT, SL, MVT::f32, r2, K1, One);

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

  r1 = DAG.getNode(ISD::FMUL, SL, MVT::f32, RHS, r3);

  SDValue r0 = DAG.getNode(AMDGPUISD::RCP, SL, MVT::f32, r1);

  SDValue Mul = DAG.getNode(ISD::FMUL, SL, MVT::f32, LHS, r0);

  return DAG.getNode(ISD::FMUL, SL, MVT::f32, r3, Mul);
}

SDValue SITargetLowering::LowerFDIV64(SDValue Op, SelectionDAG &DAG) const {
  if (DAG.getTarget().Options.UnsafeFPMath)
    return LowerFastFDIV(Op, DAG);

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

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

  SDVTList ScaleVT = DAG.getVTList(MVT::f64, MVT::i1);

  SDValue DivScale0 = DAG.getNode(AMDGPUISD::DIV_SCALE, SL, ScaleVT, Y, Y, X);

  SDValue NegDivScale0 = DAG.getNode(ISD::FNEG, SL, MVT::f64, DivScale0);

  SDValue Rcp = DAG.getNode(AMDGPUISD::RCP, SL, MVT::f64, DivScale0);

  SDValue Fma0 = DAG.getNode(ISD::FMA, SL, MVT::f64, NegDivScale0, Rcp, One);

  SDValue Fma1 = DAG.getNode(ISD::FMA, SL, MVT::f64, Rcp, Fma0, Rcp);

  SDValue Fma2 = DAG.getNode(ISD::FMA, SL, MVT::f64, NegDivScale0, Fma1, One);

  SDValue DivScale1 = DAG.getNode(AMDGPUISD::DIV_SCALE, SL, ScaleVT, X, Y, X);

  SDValue Fma3 = DAG.getNode(ISD::FMA, SL, MVT::f64, Fma1, Fma2, Fma1);
  SDValue Mul = DAG.getNode(ISD::FMUL, SL, MVT::f64, DivScale1, Fma3);

  SDValue Fma4 = DAG.getNode(ISD::FMA, SL, MVT::f64,
                             NegDivScale0, Mul, DivScale1);

  SDValue Scale;

  if (Subtarget->getGeneration() == AMDGPUSubtarget::SOUTHERN_ISLANDS) {
    // Workaround a hardware bug on SI where the condition output from div_scale
    // is not usable.

    const SDValue Hi = DAG.getConstant(1, SL, MVT::i32);

    // Figure out if the scale to use for div_fmas.
    SDValue NumBC = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, X);
    SDValue DenBC = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Y);
    SDValue Scale0BC = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, DivScale0);
    SDValue Scale1BC = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, DivScale1);

    SDValue NumHi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, NumBC, Hi);
    SDValue DenHi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, DenBC, Hi);

    SDValue Scale0Hi
      = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Scale0BC, Hi);
    SDValue Scale1Hi
      = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Scale1BC, Hi);

    SDValue CmpDen = DAG.getSetCC(SL, MVT::i1, DenHi, Scale0Hi, ISD::SETEQ);
    SDValue CmpNum = DAG.getSetCC(SL, MVT::i1, NumHi, Scale1Hi, ISD::SETEQ);
    Scale = DAG.getNode(ISD::XOR, SL, MVT::i1, CmpNum, CmpDen);
  } else {
    Scale = DivScale1.getValue(1);
  }

  SDValue Fmas = DAG.getNode(AMDGPUISD::DIV_FMAS, SL, MVT::f64,
                             Fma4, Fma3, Mul, Scale);

  return DAG.getNode(AMDGPUISD::DIV_FIXUP, SL, MVT::f64, Fmas, Y, X);
}

SDValue SITargetLowering::LowerFDIV(SDValue Op, SelectionDAG &DAG) const {
  EVT VT = Op.getValueType();

  if (VT == MVT::f32)
    return LowerFDIV32(Op, DAG);

  if (VT == MVT::f64)
    return LowerFDIV64(Op, DAG);

  llvm_unreachable("Unexpected type for fdiv");
}

SDValue SITargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
  SDLoc DL(Op);
  StoreSDNode *Store = cast<StoreSDNode>(Op);
  EVT VT = Store->getMemoryVT();

  // These stores are legal.
  if (Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS) {
    if (VT.isVector() && VT.getVectorNumElements() > 4)
      return ScalarizeVectorStore(Op, DAG);
    return SDValue();
  }

  SDValue Ret = AMDGPUTargetLowering::LowerSTORE(Op, DAG);
  if (Ret.getNode())
    return Ret;

  if (VT.isVector() && VT.getVectorNumElements() >= 8)
      return SplitVectorStore(Op, DAG);

  if (VT == MVT::i1)
    return DAG.getTruncStore(Store->getChain(), DL,
                        DAG.getSExtOrTrunc(Store->getValue(), DL, MVT::i32),
                        Store->getBasePtr(), MVT::i1, Store->getMemOperand());

  return SDValue();
}

SDValue SITargetLowering::LowerTrig(SDValue Op, SelectionDAG &DAG) const {
  SDLoc DL(Op);
  EVT VT = Op.getValueType();
  SDValue Arg = Op.getOperand(0);
  // TODO: Should this propagate fast-math-flags?
  SDValue FractPart = DAG.getNode(AMDGPUISD::FRACT, DL, VT,
                                  DAG.getNode(ISD::FMUL, DL, VT, Arg,
                                              DAG.getConstantFP(0.5/M_PI, DL,
                                                                VT)));

  switch (Op.getOpcode()) {
  case ISD::FCOS:
    return DAG.getNode(AMDGPUISD::COS_HW, SDLoc(Op), VT, FractPart);
  case ISD::FSIN:
    return DAG.getNode(AMDGPUISD::SIN_HW, SDLoc(Op), VT, FractPart);
  default:
    llvm_unreachable("Wrong trig opcode");
  }
}

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

SDValue SITargetLowering::performUCharToFloatCombine(SDNode *N,
                                                     DAGCombinerInfo &DCI) const {
  EVT VT = N->getValueType(0);
  EVT ScalarVT = VT.getScalarType();
  if (ScalarVT != MVT::f32)
    return SDValue();

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

  SDValue Src = N->getOperand(0);
  EVT SrcVT = Src.getValueType();

  // TODO: We could try to match extracting the higher bytes, which would be
  // easier if i8 vectors weren't promoted to i32 vectors, particularly after
  // types are legalized. v4i8 -> v4f32 is probably the only case to worry
  // about in practice.
  if (DCI.isAfterLegalizeVectorOps() && SrcVT == MVT::i32) {
    if (DAG.MaskedValueIsZero(Src, APInt::getHighBitsSet(32, 24))) {
      SDValue Cvt = DAG.getNode(AMDGPUISD::CVT_F32_UBYTE0, DL, VT, Src);
      DCI.AddToWorklist(Cvt.getNode());
      return Cvt;
    }
  }

  // We are primarily trying to catch operations on illegal vector types
  // before they are expanded.
  // For scalars, we can use the more flexible method of checking masked bits
  // after legalization.
  if (!DCI.isBeforeLegalize() ||
      !SrcVT.isVector() ||
      SrcVT.getVectorElementType() != MVT::i8) {
    return SDValue();
  }

  assert(DCI.isBeforeLegalize() && "Unexpected legal type");

  // Weird sized vectors are a pain to handle, but we know 3 is really the same
  // size as 4.
  unsigned NElts = SrcVT.getVectorNumElements();
  if (!SrcVT.isSimple() && NElts != 3)
    return SDValue();

  // Handle v4i8 -> v4f32 extload. Replace the v4i8 with a legal i32 load to
  // prevent a mess from expanding to v4i32 and repacking.
  if (ISD::isNormalLoad(Src.getNode()) && Src.hasOneUse()) {
    EVT LoadVT = getEquivalentMemType(*DAG.getContext(), SrcVT);
    EVT RegVT = getEquivalentLoadRegType(*DAG.getContext(), SrcVT);
    EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f32, NElts);
    LoadSDNode *Load = cast<LoadSDNode>(Src);

    unsigned AS = Load->getAddressSpace();
    unsigned Align = Load->getAlignment();
    Type *Ty = LoadVT.getTypeForEVT(*DAG.getContext());
    unsigned ABIAlignment = DAG.getDataLayout().getABITypeAlignment(Ty);

    // Don't try to replace the load if we have to expand it due to alignment
    // problems. Otherwise we will end up scalarizing the load, and trying to
    // repack into the vector for no real reason.
    if (Align < ABIAlignment &&
        !allowsMisalignedMemoryAccesses(LoadVT, AS, Align, nullptr)) {
      return SDValue();
    }

    SDValue NewLoad = DAG.getExtLoad(ISD::ZEXTLOAD, DL, RegVT,
                                     Load->getChain(),
                                     Load->getBasePtr(),
                                     LoadVT,
                                     Load->getMemOperand());

    // Make sure successors of the original load stay after it by updating
    // them to use the new Chain.
    DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 1), NewLoad.getValue(1));

    SmallVector<SDValue, 4> Elts;
    if (RegVT.isVector())
      DAG.ExtractVectorElements(NewLoad, Elts);
    else
      Elts.push_back(NewLoad);

    SmallVector<SDValue, 4> Ops;

    unsigned EltIdx = 0;
    for (SDValue Elt : Elts) {
      unsigned ComponentsInElt = std::min(4u, NElts - 4 * EltIdx);
      for (unsigned I = 0; I < ComponentsInElt; ++I) {
        unsigned Opc = AMDGPUISD::CVT_F32_UBYTE0 + I;
        SDValue Cvt = DAG.getNode(Opc, DL, MVT::f32, Elt);
        DCI.AddToWorklist(Cvt.getNode());
        Ops.push_back(Cvt);
      }

      ++EltIdx;
    }

    assert(Ops.size() == NElts);

    return DAG.getNode(ISD::BUILD_VECTOR, DL, FloatVT, Ops);
  }

  return SDValue();
}

/// \brief Return true if the given offset Size in bytes can be folded into
/// the immediate offsets of a memory instruction for the given address space.
static bool canFoldOffset(unsigned OffsetSize, unsigned AS,
                          const AMDGPUSubtarget &STI) {
  switch (AS) {
  case AMDGPUAS::GLOBAL_ADDRESS: {
    // MUBUF instructions a 12-bit offset in bytes.
    return isUInt<12>(OffsetSize);
  }
  case AMDGPUAS::CONSTANT_ADDRESS: {
    // SMRD instructions have an 8-bit offset in dwords on SI and
    // a 20-bit offset in bytes on VI.
    if (STI.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS)
      return isUInt<20>(OffsetSize);
    else
      return (OffsetSize % 4 == 0) && isUInt<8>(OffsetSize / 4);
  }
  case AMDGPUAS::LOCAL_ADDRESS:
  case AMDGPUAS::REGION_ADDRESS: {
    // The single offset versions have a 16-bit offset in bytes.
    return isUInt<16>(OffsetSize);
  }
  case AMDGPUAS::PRIVATE_ADDRESS:
  // Indirect register addressing does not use any offsets.
  default:
    return 0;
  }
}

// (shl (add x, c1), c2) -> add (shl x, c2), (shl c1, c2)

// This is a variant of
// (mul (add x, c1), c2) -> add (mul x, c2), (mul c1, c2),
//
// The normal DAG combiner will do this, but only if the add has one use since
// that would increase the number of instructions.
//
// This prevents us from seeing a constant offset that can be folded into a
// memory instruction's addressing mode. If we know the resulting add offset of
// a pointer can be folded into an addressing offset, we can replace the pointer
// operand with the add of new constant offset. This eliminates one of the uses,
// and may allow the remaining use to also be simplified.
//
SDValue SITargetLowering::performSHLPtrCombine(SDNode *N,
                                               unsigned AddrSpace,
                                               DAGCombinerInfo &DCI) const {
  SDValue N0 = N->getOperand(0);
  SDValue N1 = N->getOperand(1);

  if (N0.getOpcode() != ISD::ADD)
    return SDValue();

  const ConstantSDNode *CN1 = dyn_cast<ConstantSDNode>(N1);
  if (!CN1)
    return SDValue();

  const ConstantSDNode *CAdd = dyn_cast<ConstantSDNode>(N0.getOperand(1));
  if (!CAdd)
    return SDValue();

  // If the resulting offset is too large, we can't fold it into the addressing
  // mode offset.
  APInt Offset = CAdd->getAPIntValue() << CN1->getAPIntValue();
  if (!canFoldOffset(Offset.getZExtValue(), AddrSpace, *Subtarget))
    return SDValue();

  SelectionDAG &DAG = DCI.DAG;
  SDLoc SL(N);
  EVT VT = N->getValueType(0);

  SDValue ShlX = DAG.getNode(ISD::SHL, SL, VT, N0.getOperand(0), N1);
  SDValue COffset = DAG.getConstant(Offset, SL, MVT::i32);

  return DAG.getNode(ISD::ADD, SL, VT, ShlX, COffset);
}

SDValue SITargetLowering::performAndCombine(SDNode *N,
                                            DAGCombinerInfo &DCI) const {
  if (DCI.isBeforeLegalize())
    return SDValue();

  SelectionDAG &DAG = DCI.DAG;

  // (and (fcmp ord x, x), (fcmp une (fabs x), inf)) ->
  // fp_class x, ~(s_nan | q_nan | n_infinity | p_infinity)
  SDValue LHS = N->getOperand(0);
  SDValue RHS = N->getOperand(1);

  if (LHS.getOpcode() == ISD::SETCC &&
      RHS.getOpcode() == ISD::SETCC) {
    ISD::CondCode LCC = cast<CondCodeSDNode>(LHS.getOperand(2))->get();
    ISD::CondCode RCC = cast<CondCodeSDNode>(RHS.getOperand(2))->get();

    SDValue X = LHS.getOperand(0);
    SDValue Y = RHS.getOperand(0);
    if (Y.getOpcode() != ISD::FABS || Y.getOperand(0) != X)
      return SDValue();

    if (LCC == ISD::SETO) {
      if (X != LHS.getOperand(1))
        return SDValue();

      if (RCC == ISD::SETUNE) {
        const ConstantFPSDNode *C1 = dyn_cast<ConstantFPSDNode>(RHS.getOperand(1));
        if (!C1 || !C1->isInfinity() || C1->isNegative())
          return SDValue();

        const uint32_t Mask = SIInstrFlags::N_NORMAL |
                              SIInstrFlags::N_SUBNORMAL |
                              SIInstrFlags::N_ZERO |
                              SIInstrFlags::P_ZERO |
                              SIInstrFlags::P_SUBNORMAL |
                              SIInstrFlags::P_NORMAL;

        static_assert(((~(SIInstrFlags::S_NAN |
                          SIInstrFlags::Q_NAN |
                          SIInstrFlags::N_INFINITY |
                          SIInstrFlags::P_INFINITY)) & 0x3ff) == Mask,
                      "mask not equal");

        SDLoc DL(N);
        return DAG.getNode(AMDGPUISD::FP_CLASS, DL, MVT::i1,
                           X, DAG.getConstant(Mask, DL, MVT::i32));
      }
    }
  }

  return SDValue();
}

SDValue SITargetLowering::performOrCombine(SDNode *N,
                                           DAGCombinerInfo &DCI) const {
  SelectionDAG &DAG = DCI.DAG;
  SDValue LHS = N->getOperand(0);
  SDValue RHS = N->getOperand(1);

  // or (fp_class x, c1), (fp_class x, c2) -> fp_class x, (c1 | c2)
  if (LHS.getOpcode() == AMDGPUISD::FP_CLASS &&
      RHS.getOpcode() == AMDGPUISD::FP_CLASS) {
    SDValue Src = LHS.getOperand(0);
    if (Src != RHS.getOperand(0))
      return SDValue();

    const ConstantSDNode *CLHS = dyn_cast<ConstantSDNode>(LHS.getOperand(1));
    const ConstantSDNode *CRHS = dyn_cast<ConstantSDNode>(RHS.getOperand(1));
    if (!CLHS || !CRHS)
      return SDValue();

    // Only 10 bits are used.
    static const uint32_t MaxMask = 0x3ff;

    uint32_t NewMask = (CLHS->getZExtValue() | CRHS->getZExtValue()) & MaxMask;
    SDLoc DL(N);
    return DAG.getNode(AMDGPUISD::FP_CLASS, DL, MVT::i1,
                       Src, DAG.getConstant(NewMask, DL, MVT::i32));
  }

  return SDValue();
}

SDValue SITargetLowering::performClassCombine(SDNode *N,
                                              DAGCombinerInfo &DCI) const {
  SelectionDAG &DAG = DCI.DAG;
  SDValue Mask = N->getOperand(1);

  // fp_class x, 0 -> false
  if (const ConstantSDNode *CMask = dyn_cast<ConstantSDNode>(Mask)) {
    if (CMask->isNullValue())
      return DAG.getConstant(0, SDLoc(N), MVT::i1);
  }

  return SDValue();
}

static unsigned minMaxOpcToMin3Max3Opc(unsigned Opc) {
  switch (Opc) {
  case ISD::FMAXNUM:
    return AMDGPUISD::FMAX3;
  case ISD::SMAX:
    return AMDGPUISD::SMAX3;
  case ISD::UMAX:
    return AMDGPUISD::UMAX3;
  case ISD::FMINNUM:
    return AMDGPUISD::FMIN3;
  case ISD::SMIN:
    return AMDGPUISD::SMIN3;
  case ISD::UMIN:
    return AMDGPUISD::UMIN3;
  default:
    llvm_unreachable("Not a min/max opcode");
  }
}

SDValue SITargetLowering::performMin3Max3Combine(SDNode *N,
                                                 DAGCombinerInfo &DCI) const {
  SelectionDAG &DAG = DCI.DAG;

  unsigned Opc = N->getOpcode();
  SDValue Op0 = N->getOperand(0);
  SDValue Op1 = N->getOperand(1);

  // Only do this if the inner op has one use since this will just increases
  // register pressure for no benefit.

  // max(max(a, b), c)
  if (Op0.getOpcode() == Opc && Op0.hasOneUse()) {
    SDLoc DL(N);
    return DAG.getNode(minMaxOpcToMin3Max3Opc(Opc),
                       DL,
                       N->getValueType(0),
                       Op0.getOperand(0),
                       Op0.getOperand(1),
                       Op1);
  }

  // max(a, max(b, c))
  if (Op1.getOpcode() == Opc && Op1.hasOneUse()) {
    SDLoc DL(N);
    return DAG.getNode(minMaxOpcToMin3Max3Opc(Opc),
                       DL,
                       N->getValueType(0),
                       Op0,
                       Op1.getOperand(0),
                       Op1.getOperand(1));
  }

  return SDValue();
}

SDValue SITargetLowering::performSetCCCombine(SDNode *N,
                                              DAGCombinerInfo &DCI) const {
  SelectionDAG &DAG = DCI.DAG;
  SDLoc SL(N);

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

  if (VT != MVT::f32 && VT != MVT::f64)
    return SDValue();

  // Match isinf pattern
  // (fcmp oeq (fabs x), inf) -> (fp_class x, (p_infinity | n_infinity))
  ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
  if (CC == ISD::SETOEQ && LHS.getOpcode() == ISD::FABS) {
    const ConstantFPSDNode *CRHS = dyn_cast<ConstantFPSDNode>(RHS);
    if (!CRHS)
      return SDValue();

    const APFloat &APF = CRHS->getValueAPF();
    if (APF.isInfinity() && !APF.isNegative()) {
      unsigned Mask = SIInstrFlags::P_INFINITY | SIInstrFlags::N_INFINITY;
      return DAG.getNode(AMDGPUISD::FP_CLASS, SL, MVT::i1, LHS.getOperand(0),
                         DAG.getConstant(Mask, SL, MVT::i32));
    }
  }

  return SDValue();
}

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

  switch (N->getOpcode()) {
  default:
    return AMDGPUTargetLowering::PerformDAGCombine(N, DCI);
  case ISD::SETCC:
    return performSetCCCombine(N, DCI);
  case ISD::FMAXNUM: // TODO: What about fmax_legacy?
  case ISD::FMINNUM:
  case ISD::SMAX:
  case ISD::SMIN:
  case ISD::UMAX:
  case ISD::UMIN: {
    if (DCI.getDAGCombineLevel() >= AfterLegalizeDAG &&
        N->getValueType(0) != MVT::f64 &&
        getTargetMachine().getOptLevel() > CodeGenOpt::None)
      return performMin3Max3Combine(N, DCI);
    break;
  }

  case AMDGPUISD::CVT_F32_UBYTE0:
  case AMDGPUISD::CVT_F32_UBYTE1:
  case AMDGPUISD::CVT_F32_UBYTE2:
  case AMDGPUISD::CVT_F32_UBYTE3: {
    unsigned Offset = N->getOpcode() - AMDGPUISD::CVT_F32_UBYTE0;

    SDValue Src = N->getOperand(0);
    APInt Demanded = APInt::getBitsSet(32, 8 * Offset, 8 * Offset + 8);

    APInt KnownZero, KnownOne;
    TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
                                          !DCI.isBeforeLegalizeOps());
    const TargetLowering &TLI = DAG.getTargetLoweringInfo();
    if (TLO.ShrinkDemandedConstant(Src, Demanded) ||
        TLI.SimplifyDemandedBits(Src, Demanded, KnownZero, KnownOne, TLO)) {
      DCI.CommitTargetLoweringOpt(TLO);
    }

    break;
  }

  case ISD::UINT_TO_FP: {
    return performUCharToFloatCombine(N, DCI);
  }
  case ISD::FADD: {
    if (DCI.getDAGCombineLevel() < AfterLegalizeDAG)
      break;

    EVT VT = N->getValueType(0);
    if (VT != MVT::f32)
      break;

    // Only do this if we are not trying to support denormals. v_mad_f32 does
    // not support denormals ever.
    if (Subtarget->hasFP32Denormals())
      break;

    SDValue LHS = N->getOperand(0);
    SDValue RHS = N->getOperand(1);

    // These should really be instruction patterns, but writing patterns with
    // source modiifiers is a pain.

    // fadd (fadd (a, a), b) -> mad 2.0, a, b
    if (LHS.getOpcode() == ISD::FADD) {
      SDValue A = LHS.getOperand(0);
      if (A == LHS.getOperand(1)) {
        const SDValue Two = DAG.getConstantFP(2.0, DL, MVT::f32);
        return DAG.getNode(ISD::FMAD, DL, VT, Two, A, RHS);
      }
    }

    // fadd (b, fadd (a, a)) -> mad 2.0, a, b
    if (RHS.getOpcode() == ISD::FADD) {
      SDValue A = RHS.getOperand(0);
      if (A == RHS.getOperand(1)) {
        const SDValue Two = DAG.getConstantFP(2.0, DL, MVT::f32);
        return DAG.getNode(ISD::FMAD, DL, VT, Two, A, LHS);
      }
    }

    return SDValue();
  }
  case ISD::FSUB: {
    if (DCI.getDAGCombineLevel() < AfterLegalizeDAG)
      break;

    EVT VT = N->getValueType(0);

    // Try to get the fneg to fold into the source modifier. This undoes generic
    // DAG combines and folds them into the mad.
    //
    // Only do this if we are not trying to support denormals. v_mad_f32 does
    // not support denormals ever.
    if (VT == MVT::f32 &&
        !Subtarget->hasFP32Denormals()) {
      SDValue LHS = N->getOperand(0);
      SDValue RHS = N->getOperand(1);
      if (LHS.getOpcode() == ISD::FADD) {
        // (fsub (fadd a, a), c) -> mad 2.0, a, (fneg c)

        SDValue A = LHS.getOperand(0);
        if (A == LHS.getOperand(1)) {
          const SDValue Two = DAG.getConstantFP(2.0, DL, MVT::f32);
          SDValue NegRHS = DAG.getNode(ISD::FNEG, DL, VT, RHS);

          return DAG.getNode(ISD::FMAD, DL, VT, Two, A, NegRHS);
        }
      }

      if (RHS.getOpcode() == ISD::FADD) {
        // (fsub c, (fadd a, a)) -> mad -2.0, a, c

        SDValue A = RHS.getOperand(0);
        if (A == RHS.getOperand(1)) {
          const SDValue NegTwo = DAG.getConstantFP(-2.0, DL, MVT::f32);
          return DAG.getNode(ISD::FMAD, DL, VT, NegTwo, A, LHS);
        }
      }

      return SDValue();
    }

    break;
  }
  case ISD::LOAD:
  case ISD::STORE:
  case ISD::ATOMIC_LOAD:
  case ISD::ATOMIC_STORE:
  case ISD::ATOMIC_CMP_SWAP:
  case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
  case ISD::ATOMIC_SWAP:
  case ISD::ATOMIC_LOAD_ADD:
  case ISD::ATOMIC_LOAD_SUB:
  case ISD::ATOMIC_LOAD_AND:
  case ISD::ATOMIC_LOAD_OR:
  case ISD::ATOMIC_LOAD_XOR:
  case ISD::ATOMIC_LOAD_NAND:
  case ISD::ATOMIC_LOAD_MIN:
  case ISD::ATOMIC_LOAD_MAX:
  case ISD::ATOMIC_LOAD_UMIN:
  case ISD::ATOMIC_LOAD_UMAX: { // TODO: Target mem intrinsics.
    if (DCI.isBeforeLegalize())
      break;

    MemSDNode *MemNode = cast<MemSDNode>(N);
    SDValue Ptr = MemNode->getBasePtr();

    // TODO: We could also do this for multiplies.
    unsigned AS = MemNode->getAddressSpace();
    if (Ptr.getOpcode() == ISD::SHL && AS != AMDGPUAS::PRIVATE_ADDRESS) {
      SDValue NewPtr = performSHLPtrCombine(Ptr.getNode(), AS, DCI);
      if (NewPtr) {
        SmallVector<SDValue, 8> NewOps(MemNode->op_begin(), MemNode->op_end());

        NewOps[N->getOpcode() == ISD::STORE ? 2 : 1] = NewPtr;
        return SDValue(DAG.UpdateNodeOperands(MemNode, NewOps), 0);
      }
    }
    break;
  }
  case ISD::AND:
    return performAndCombine(N, DCI);
  case ISD::OR:
    return performOrCombine(N, DCI);
  case AMDGPUISD::FP_CLASS:
    return performClassCombine(N, DCI);
  }
  return AMDGPUTargetLowering::PerformDAGCombine(N, DCI);
}

/// \brief Analyze the possible immediate value Op
///
/// Returns -1 if it isn't an immediate, 0 if it's and inline immediate
/// and the immediate value if it's a literal immediate
int32_t SITargetLowering::analyzeImmediate(const SDNode *N) const {

  const SIInstrInfo *TII =
      static_cast<const SIInstrInfo *>(Subtarget->getInstrInfo());

  if (const ConstantSDNode *Node = dyn_cast<ConstantSDNode>(N)) {
    if (TII->isInlineConstant(Node->getAPIntValue()))
      return 0;

    uint64_t Val = Node->getZExtValue();
    return isUInt<32>(Val) ? Val : -1;
  }

  if (const ConstantFPSDNode *Node = dyn_cast<ConstantFPSDNode>(N)) {
    if (TII->isInlineConstant(Node->getValueAPF().bitcastToAPInt()))
      return 0;

    if (Node->getValueType(0) == MVT::f32)
      return FloatToBits(Node->getValueAPF().convertToFloat());

    return -1;
  }

  return -1;
}

/// \brief Helper function for adjustWritemask
static unsigned SubIdx2Lane(unsigned Idx) {
  switch (Idx) {
  default: return 0;
  case AMDGPU::sub0: return 0;
  case AMDGPU::sub1: return 1;
  case AMDGPU::sub2: return 2;
  case AMDGPU::sub3: return 3;
  }
}

/// \brief Adjust the writemask of MIMG instructions
void SITargetLowering::adjustWritemask(MachineSDNode *&Node,
                                       SelectionDAG &DAG) const {
  SDNode *Users[4] = { };
  unsigned Lane = 0;
  unsigned OldDmask = Node->getConstantOperandVal(0);
  unsigned NewDmask = 0;

  // Try to figure out the used register components
  for (SDNode::use_iterator I = Node->use_begin(), E = Node->use_end();
       I != E; ++I) {

    // Abort if we can't understand the usage
    if (!I->isMachineOpcode() ||
        I->getMachineOpcode() != TargetOpcode::EXTRACT_SUBREG)
      return;

    // Lane means which subreg of %VGPRa_VGPRb_VGPRc_VGPRd is used.
    // Note that subregs are packed, i.e. Lane==0 is the first bit set
    // in OldDmask, so it can be any of X,Y,Z,W; Lane==1 is the second bit
    // set, etc.
    Lane = SubIdx2Lane(I->getConstantOperandVal(1));

    // Set which texture component corresponds to the lane.
    unsigned Comp;
    for (unsigned i = 0, Dmask = OldDmask; i <= Lane; i++) {
      assert(Dmask);
      Comp = countTrailingZeros(Dmask);
      Dmask &= ~(1 << Comp);
    }

    // Abort if we have more than one user per component
    if (Users[Lane])
      return;

    Users[Lane] = *I;
    NewDmask |= 1 << Comp;
  }

  // Abort if there's no change
  if (NewDmask == OldDmask)
    return;

  // Adjust the writemask in the node
  std::vector<SDValue> Ops;
  Ops.push_back(DAG.getTargetConstant(NewDmask, SDLoc(Node), MVT::i32));
  Ops.insert(Ops.end(), Node->op_begin() + 1, Node->op_end());
  Node = (MachineSDNode*)DAG.UpdateNodeOperands(Node, Ops);

  // If we only got one lane, replace it with a copy
  // (if NewDmask has only one bit set...)
  if (NewDmask && (NewDmask & (NewDmask-1)) == 0) {
    SDValue RC = DAG.getTargetConstant(AMDGPU::VGPR_32RegClassID, SDLoc(),
                                       MVT::i32);
    SDNode *Copy = DAG.getMachineNode(TargetOpcode::COPY_TO_REGCLASS,
                                      SDLoc(), Users[Lane]->getValueType(0),
                                      SDValue(Node, 0), RC);
    DAG.ReplaceAllUsesWith(Users[Lane], Copy);
    return;
  }

  // Update the users of the node with the new indices
  for (unsigned i = 0, Idx = AMDGPU::sub0; i < 4; ++i) {

    SDNode *User = Users[i];
    if (!User)
      continue;

    SDValue Op = DAG.getTargetConstant(Idx, SDLoc(User), MVT::i32);
    DAG.UpdateNodeOperands(User, User->getOperand(0), Op);

    switch (Idx) {
    default: break;
    case AMDGPU::sub0: Idx = AMDGPU::sub1; break;
    case AMDGPU::sub1: Idx = AMDGPU::sub2; break;
    case AMDGPU::sub2: Idx = AMDGPU::sub3; break;
    }
  }
}

static bool isFrameIndexOp(SDValue Op) {
  if (Op.getOpcode() == ISD::AssertZext)
    Op = Op.getOperand(0);

  return isa<FrameIndexSDNode>(Op);
}

/// \brief Legalize target independent instructions (e.g. INSERT_SUBREG)
/// with frame index operands.
/// LLVM assumes that inputs are to these instructions are registers.
void SITargetLowering::legalizeTargetIndependentNode(SDNode *Node,
                                                     SelectionDAG &DAG) const {

  SmallVector<SDValue, 8> Ops;
  for (unsigned i = 0; i < Node->getNumOperands(); ++i) {
    if (!isFrameIndexOp(Node->getOperand(i))) {
      Ops.push_back(Node->getOperand(i));
      continue;
    }

    SDLoc DL(Node);
    Ops.push_back(SDValue(DAG.getMachineNode(AMDGPU::S_MOV_B32, DL,
                                     Node->getOperand(i).getValueType(),
                                     Node->getOperand(i)), 0));
  }

  DAG.UpdateNodeOperands(Node, Ops);
}

/// \brief Fold the instructions after selecting them.
SDNode *SITargetLowering::PostISelFolding(MachineSDNode *Node,
                                          SelectionDAG &DAG) const {
  const SIInstrInfo *TII =
      static_cast<const SIInstrInfo *>(Subtarget->getInstrInfo());

  if (TII->isMIMG(Node->getMachineOpcode()))
    adjustWritemask(Node, DAG);

  if (Node->getMachineOpcode() == AMDGPU::INSERT_SUBREG ||
      Node->getMachineOpcode() == AMDGPU::REG_SEQUENCE) {
    legalizeTargetIndependentNode(Node, DAG);
    return Node;
  }
  return Node;
}

/// \brief Assign the register class depending on the number of
/// bits set in the writemask
void SITargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
                                                     SDNode *Node) const {
  const SIInstrInfo *TII =
      static_cast<const SIInstrInfo *>(Subtarget->getInstrInfo());

  MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();

  if (TII->isVOP3(MI->getOpcode())) {
    // Make sure constant bus requirements are respected.
    TII->legalizeOperandsVOP3(MRI, MI);
    return;
  }

  if (TII->isMIMG(*MI)) {
    unsigned VReg = MI->getOperand(0).getReg();
    unsigned Writemask = MI->getOperand(1).getImm();
    unsigned BitsSet = 0;
    for (unsigned i = 0; i < 4; ++i)
      BitsSet += Writemask & (1 << i) ? 1 : 0;

    const TargetRegisterClass *RC;
    switch (BitsSet) {
    default: return;
    case 1:  RC = &AMDGPU::VGPR_32RegClass; break;
    case 2:  RC = &AMDGPU::VReg_64RegClass; break;
    case 3:  RC = &AMDGPU::VReg_96RegClass; break;
    }

    unsigned NewOpcode = TII->getMaskedMIMGOp(MI->getOpcode(), BitsSet);
    MI->setDesc(TII->get(NewOpcode));
    MRI.setRegClass(VReg, RC);
    return;
  }

  // Replace unused atomics with the no return version.
  int NoRetAtomicOp = AMDGPU::getAtomicNoRetOp(MI->getOpcode());
  if (NoRetAtomicOp != -1) {
    if (!Node->hasAnyUseOfValue(0)) {
      MI->setDesc(TII->get(NoRetAtomicOp));
      MI->RemoveOperand(0);
    }

    return;
  }
}

static SDValue buildSMovImm32(SelectionDAG &DAG, SDLoc DL, uint64_t Val) {
  SDValue K = DAG.getTargetConstant(Val, DL, MVT::i32);
  return SDValue(DAG.getMachineNode(AMDGPU::S_MOV_B32, DL, MVT::i32, K), 0);
}

MachineSDNode *SITargetLowering::wrapAddr64Rsrc(SelectionDAG &DAG,
                                                SDLoc DL,
                                                SDValue Ptr) const {
  const SIInstrInfo *TII =
    static_cast<const SIInstrInfo *>(Subtarget->getInstrInfo());

  // Build the half of the subregister with the constants before building the
  // full 128-bit register. If we are building multiple resource descriptors,
  // this will allow CSEing of the 2-component register.
  const SDValue Ops0[] = {
    DAG.getTargetConstant(AMDGPU::SGPR_64RegClassID, DL, MVT::i32),
    buildSMovImm32(DAG, DL, 0),
    DAG.getTargetConstant(AMDGPU::sub0, DL, MVT::i32),
    buildSMovImm32(DAG, DL, TII->getDefaultRsrcDataFormat() >> 32),
    DAG.getTargetConstant(AMDGPU::sub1, DL, MVT::i32)
  };

  SDValue SubRegHi = SDValue(DAG.getMachineNode(AMDGPU::REG_SEQUENCE, DL,
                                                MVT::v2i32, Ops0), 0);

  // Combine the constants and the pointer.
  const SDValue Ops1[] = {
    DAG.getTargetConstant(AMDGPU::SReg_128RegClassID, DL, MVT::i32),
    Ptr,
    DAG.getTargetConstant(AMDGPU::sub0_sub1, DL, MVT::i32),
    SubRegHi,
    DAG.getTargetConstant(AMDGPU::sub2_sub3, DL, MVT::i32)
  };

  return DAG.getMachineNode(AMDGPU::REG_SEQUENCE, DL, MVT::v4i32, Ops1);
}

/// \brief Return a resource descriptor with the 'Add TID' bit enabled
///        The TID (Thread ID) is multiplied by the stride value (bits [61:48]
///        of the resource descriptor) to create an offset, which is added to
///        the resource pointer.
MachineSDNode *SITargetLowering::buildRSRC(SelectionDAG &DAG,
                                           SDLoc DL,
                                           SDValue Ptr,
                                           uint32_t RsrcDword1,
                                           uint64_t RsrcDword2And3) const {
  SDValue PtrLo = DAG.getTargetExtractSubreg(AMDGPU::sub0, DL, MVT::i32, Ptr);
  SDValue PtrHi = DAG.getTargetExtractSubreg(AMDGPU::sub1, DL, MVT::i32, Ptr);
  if (RsrcDword1) {
    PtrHi = SDValue(DAG.getMachineNode(AMDGPU::S_OR_B32, DL, MVT::i32, PtrHi,
                                     DAG.getConstant(RsrcDword1, DL, MVT::i32)),
                    0);
  }

  SDValue DataLo = buildSMovImm32(DAG, DL,
                                  RsrcDword2And3 & UINT64_C(0xFFFFFFFF));
  SDValue DataHi = buildSMovImm32(DAG, DL, RsrcDword2And3 >> 32);

  const SDValue Ops[] = {
    DAG.getTargetConstant(AMDGPU::SReg_128RegClassID, DL, MVT::i32),
    PtrLo,
    DAG.getTargetConstant(AMDGPU::sub0, DL, MVT::i32),
    PtrHi,
    DAG.getTargetConstant(AMDGPU::sub1, DL, MVT::i32),
    DataLo,
    DAG.getTargetConstant(AMDGPU::sub2, DL, MVT::i32),
    DataHi,
    DAG.getTargetConstant(AMDGPU::sub3, DL, MVT::i32)
  };

  return DAG.getMachineNode(AMDGPU::REG_SEQUENCE, DL, MVT::v4i32, Ops);
}

SDValue SITargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
                                               const TargetRegisterClass *RC,
                                               unsigned Reg, EVT VT) const {
  SDValue VReg = AMDGPUTargetLowering::CreateLiveInRegister(DAG, RC, Reg, VT);

  return DAG.getCopyFromReg(DAG.getEntryNode(), SDLoc(DAG.getEntryNode()),
                            cast<RegisterSDNode>(VReg)->getReg(), VT);
}

//===----------------------------------------------------------------------===//
//                         SI Inline Assembly Support
//===----------------------------------------------------------------------===//

std::pair<unsigned, const TargetRegisterClass *>
SITargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
                                               StringRef Constraint,
                                               MVT VT) const {

  if (Constraint.size() == 1) {
    switch (Constraint[0]) {
    case 's':
    case 'r':
      switch (VT.getSizeInBits()) {
      default:
        return std::make_pair(0U, nullptr);
      case 32:
        return std::make_pair(0U, &AMDGPU::SGPR_32RegClass);
      case 64:
        return std::make_pair(0U, &AMDGPU::SGPR_64RegClass);
      case 128:
        return std::make_pair(0U, &AMDGPU::SReg_128RegClass);
      case 256:
        return std::make_pair(0U, &AMDGPU::SReg_256RegClass);
      }

    case 'v':
      switch (VT.getSizeInBits()) {
      default:
        return std::make_pair(0U, nullptr);
      case 32:
        return std::make_pair(0U, &AMDGPU::VGPR_32RegClass);
      case 64:
        return std::make_pair(0U, &AMDGPU::VReg_64RegClass);
      case 96:
        return std::make_pair(0U, &AMDGPU::VReg_96RegClass);
      case 128:
        return std::make_pair(0U, &AMDGPU::VReg_128RegClass);
      case 256:
        return std::make_pair(0U, &AMDGPU::VReg_256RegClass);
      case 512:
        return std::make_pair(0U, &AMDGPU::VReg_512RegClass);
      }
    }
  }

  if (Constraint.size() > 1) {
    const TargetRegisterClass *RC = nullptr;
    if (Constraint[1] == 'v') {
      RC = &AMDGPU::VGPR_32RegClass;
    } else if (Constraint[1] == 's') {
      RC = &AMDGPU::SGPR_32RegClass;
    }

    if (RC) {
      uint32_t Idx;
      bool Failed = Constraint.substr(2).getAsInteger(10, Idx);
      if (!Failed && Idx < RC->getNumRegs())
        return std::make_pair(RC->getRegister(Idx), RC);
    }
  }
  return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
}

SITargetLowering::ConstraintType
SITargetLowering::getConstraintType(StringRef Constraint) const {
  if (Constraint.size() == 1) {
    switch (Constraint[0]) {
    default: break;
    case 's':
    case 'v':
      return C_RegisterClass;
    }
  }
  return TargetLowering::getConstraintType(Constraint);
}