summaryrefslogtreecommitdiff
path: root/src/gallium/drivers/radeonsi/si_shader.c
blob: 7260a6b8df69297131d68e19b44827df2c9d526a (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
2049
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
2093
2094
2095
2096
2097
2098
2099
2100
2101
2102
2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
2150
2151
2152
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
2170
2171
2172
2173
2174
2175
2176
2177
2178
2179
2180
2181
2182
2183
2184
2185
2186
2187
2188
2189
2190
2191
2192
2193
2194
2195
2196
2197
2198
2199
2200
2201
2202
2203
2204
2205
2206
2207
2208
2209
2210
2211
2212
2213
2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
2228
2229
2230
2231
2232
2233
2234
2235
2236
2237
2238
2239
2240
2241
2242
2243
2244
2245
2246
2247
2248
2249
2250
2251
2252
2253
2254
2255
2256
2257
2258
2259
2260
2261
2262
2263
2264
2265
2266
2267
2268
2269
2270
2271
2272
2273
2274
2275
2276
2277
2278
2279
2280
2281
2282
2283
2284
2285
2286
2287
2288
2289
2290
2291
2292
2293
2294
2295
2296
2297
2298
2299
2300
2301
2302
2303
2304
2305
2306
2307
2308
2309
2310
2311
2312
2313
2314
2315
2316
2317
2318
2319
2320
2321
2322
2323
2324
2325
2326
2327
2328
2329
2330
2331
2332
2333
2334
2335
2336
2337
2338
2339
2340
2341
2342
2343
2344
2345
2346
2347
2348
2349
2350
2351
2352
2353
2354
2355
2356
2357
2358
2359
2360
2361
2362
2363
2364
2365
2366
2367
2368
2369
2370
2371
2372
2373
2374
2375
2376
2377
2378
2379
2380
2381
2382
2383
2384
2385
2386
2387
2388
2389
2390
2391
2392
2393
2394
2395
2396
2397
2398
2399
2400
2401
2402
2403
2404
2405
2406
2407
2408
2409
2410
2411
2412
2413
2414
2415
2416
2417
2418
2419
2420
2421
2422
2423
2424
2425
2426
2427
2428
2429
2430
2431
2432
2433
2434
2435
2436
2437
2438
2439
2440
2441
2442
2443
2444
2445
2446
2447
2448
2449
2450
2451
2452
2453
2454
2455
2456
2457
2458
2459
2460
2461
2462
2463
2464
2465
2466
2467
2468
2469
2470
2471
2472
2473
2474
2475
2476
2477
2478
2479
2480
2481
2482
2483
2484
2485
2486
2487
2488
2489
2490
2491
2492
2493
2494
2495
2496
2497
2498
2499
2500
2501
2502
2503
2504
2505
2506
2507
2508
2509
2510
2511
2512
2513
2514
2515
2516
2517
2518
2519
2520
2521
2522
2523
2524
2525
2526
2527
2528
2529
2530
2531
2532
2533
2534
2535
2536
2537
2538
2539
2540
2541
2542
2543
2544
2545
2546
2547
2548
2549
2550
2551
2552
2553
2554
2555
2556
2557
2558
2559
2560
2561
2562
2563
2564
2565
2566
2567
2568
2569
2570
2571
2572
2573
2574
2575
2576
2577
2578
2579
2580
2581
2582
2583
2584
2585
2586
2587
2588
2589
2590
2591
2592
2593
2594
2595
2596
2597
2598
2599
2600
2601
2602
2603
2604
2605
2606
2607
2608
2609
2610
2611
2612
2613
2614
2615
2616
2617
2618
2619
2620
2621
2622
2623
2624
2625
2626
2627
2628
2629
2630
2631
2632
2633
2634
2635
2636
2637
2638
2639
2640
2641
2642
2643
2644
2645
2646
2647
2648
2649
2650
2651
2652
2653
2654
2655
2656
2657
2658
2659
2660
2661
2662
2663
2664
2665
2666
2667
2668
2669
2670
2671
2672
2673
2674
2675
2676
2677
2678
2679
2680
2681
2682
2683
2684
2685
2686
2687
2688
2689
2690
2691
2692
2693
2694
2695
2696
2697
2698
2699
2700
2701
2702
2703
2704
2705
2706
2707
2708
2709
2710
2711
2712
2713
2714
2715
2716
2717
2718
2719
2720
2721
2722
2723
2724
2725
2726
2727
2728
2729
2730
2731
2732
2733
2734
2735
2736
2737
2738
2739
2740
2741
2742
2743
2744
2745
2746
2747
2748
2749
2750
2751
2752
2753
2754
2755
2756
2757
2758
2759
2760
2761
2762
2763
2764
2765
2766
2767
2768
2769
2770
2771
2772
2773
2774
2775
2776
2777
2778
2779
2780
2781
2782
2783
2784
2785
2786
2787
2788
2789
2790
2791
2792
2793
2794
2795
2796
2797
2798
2799
2800
2801
2802
2803
2804
2805
2806
2807
2808
2809
2810
2811
2812
2813
2814
2815
2816
2817
2818
2819
2820
2821
2822
2823
2824
2825
2826
2827
2828
2829
2830
2831
2832
2833
2834
2835
2836
2837
2838
2839
2840
2841
2842
2843
2844
2845
2846
2847
2848
2849
2850
2851
2852
2853
2854
2855
2856
2857
2858
2859
2860
2861
2862
2863
2864
2865
2866
2867
2868
2869
2870
2871
2872
2873
2874
2875
2876
2877
2878
2879
2880
2881
2882
2883
2884
2885
2886
2887
2888
2889
2890
2891
2892
2893
2894
2895
2896
2897
2898
2899
2900
2901
2902
2903
2904
2905
2906
2907
2908
2909
2910
2911
2912
2913
2914
2915
2916
2917
2918
2919
2920
2921
2922
2923
2924
2925
2926
2927
2928
2929
2930
2931
2932
2933
2934
2935
2936
2937
2938
2939
2940
2941
2942
2943
2944
2945
2946
2947
2948
2949
2950
2951
2952
2953
2954
2955
2956
2957
2958
2959
2960
2961
2962
2963
2964
2965
2966
2967
2968
2969
2970
2971
2972
2973
2974
2975
2976
2977
2978
2979
2980
2981
2982
2983
2984
2985
2986
2987
2988
2989
2990
2991
2992
2993
2994
2995
2996
2997
2998
2999
3000
3001
3002
3003
3004
3005
3006
3007
3008
3009
3010
3011
3012
3013
3014
3015
3016
3017
3018
3019
3020
3021
3022
3023
3024
3025
3026
3027
3028
3029
3030
3031
3032
3033
3034
3035
3036
3037
3038
3039
3040
3041
3042
3043
3044
3045
3046
3047
3048
3049
3050
3051
3052
3053
3054
3055
3056
3057
3058
3059
3060
3061
3062
3063
3064
3065
3066
3067
3068
3069
3070
3071
3072
3073
3074
3075
3076
3077
3078
3079
3080
3081
3082
3083
3084
3085
3086
3087
3088
3089
3090
3091
3092
3093
3094
3095
3096
3097
3098
3099
3100
3101
3102
3103
3104
3105
3106
3107
3108
3109
3110
3111
3112
3113
3114
3115
3116
3117
3118
3119
3120
3121
3122
3123
3124
3125
3126
3127
3128
3129
3130
3131
3132
3133
3134
3135
3136
3137
3138
3139
3140
3141
3142
3143
3144
3145
3146
3147
3148
3149
3150
3151
3152
3153
3154
3155
3156
3157
3158
3159
3160
3161
3162
3163
3164
3165
3166
3167
3168
3169
3170
3171
3172
3173
3174
3175
3176
3177
3178
3179
3180
3181
3182
3183
3184
3185
3186
3187
3188
3189
3190
3191
3192
3193
3194
3195
3196
3197
3198
3199
3200
3201
3202
3203
3204
3205
3206
3207
3208
3209
3210
3211
3212
3213
3214
3215
3216
3217
3218
3219
3220
3221
3222
3223
3224
3225
3226
3227
3228
3229
3230
3231
3232
3233
3234
3235
3236
3237
3238
3239
3240
3241
3242
3243
3244
3245
3246
3247
3248
3249
3250
3251
3252
3253
3254
3255
3256
3257
3258
3259
3260
3261
3262
3263
3264
3265
3266
3267
3268
3269
3270
3271
3272
3273
3274
3275
3276
3277
3278
3279
3280
3281
3282
3283
3284
3285
3286
3287
3288
3289
3290
3291
3292
3293
3294
3295
3296
3297
3298
3299
3300
3301
3302
3303
3304
3305
3306
3307
3308
3309
3310
3311
3312
3313
3314
3315
3316
3317
3318
3319
3320
3321
3322
3323
3324
3325
3326
3327
3328
3329
3330
3331
3332
3333
3334
3335
3336
3337
3338
3339
3340
3341
3342
3343
3344
3345
3346
3347
3348
3349
3350
3351
3352
3353
3354
3355
3356
3357
3358
3359
3360
3361
3362
3363
3364
3365
3366
3367
3368
3369
3370
3371
3372
3373
3374
3375
3376
3377
3378
3379
3380
3381
3382
3383
3384
3385
3386
3387
3388
3389
3390
3391
3392
3393
3394
3395
3396
3397
3398
3399
3400
3401
3402
3403
3404
3405
3406
3407
3408
3409
3410
3411
3412
3413
3414
3415
3416
3417
3418
3419
3420
3421
3422
3423
3424
3425
3426
3427
3428
3429
3430
3431
3432
3433
3434
3435
3436
3437
3438
3439
3440
3441
3442
3443
3444
3445
3446
3447
3448
3449
3450
3451
3452
3453
3454
3455
3456
3457
3458
3459
3460
3461
3462
3463
3464
3465
3466
3467
3468
3469
3470
3471
3472
3473
3474
3475
3476
3477
3478
3479
3480
3481
3482
3483
3484
3485
3486
3487
3488
3489
3490
3491
3492
3493
3494
3495
3496
3497
3498
3499
3500
3501
3502
3503
3504
3505
3506
3507
3508
3509
3510
3511
3512
3513
3514
3515
3516
3517
3518
3519
3520
3521
3522
3523
3524
3525
3526
3527
3528
3529
3530
3531
3532
3533
3534
3535
3536
3537
3538
3539
3540
3541
3542
3543
3544
3545
3546
3547
3548
3549
3550
3551
3552
3553
3554
3555
3556
3557
3558
3559
3560
3561
3562
3563
3564
3565
3566
3567
3568
3569
3570
3571
3572
3573
3574
3575
3576
3577
3578
3579
3580
3581
3582
3583
3584
3585
3586
3587
3588
3589
3590
3591
3592
3593
3594
3595
3596
3597
3598
3599
3600
3601
3602
3603
3604
3605
3606
3607
3608
3609
3610
3611
3612
3613
3614
3615
3616
3617
3618
3619
3620
3621
3622
3623
3624
3625
3626
3627
3628
3629
3630
3631
3632
3633
3634
3635
3636
3637
3638
3639
3640
3641
3642
3643
3644
3645
3646
3647
3648
3649
3650
3651
3652
3653
3654
3655
3656
3657
3658
3659
3660
3661
3662
3663
3664
3665
3666
3667
3668
3669
3670
3671
3672
3673
3674
3675
3676
3677
3678
3679
3680
3681
3682
3683
3684
3685
3686
3687
3688
3689
3690
3691
3692
3693
3694
3695
3696
3697
3698
3699
3700
3701
3702
3703
3704
3705
3706
3707
3708
3709
3710
3711
3712
3713
3714
3715
3716
3717
3718
3719
3720
3721
3722
3723
3724
3725
3726
3727
3728
3729
3730
3731
3732
3733
3734
3735
3736
3737
3738
3739
3740
3741
3742
3743
3744
3745
3746
3747
3748
3749
3750
3751
3752
3753
3754
3755
3756
3757
3758
3759
3760
3761
3762
3763
3764
3765
3766
3767
3768
3769
3770
3771
3772
3773
3774
3775
3776
3777
3778
3779
3780
3781
3782
3783
3784
3785
3786
3787
3788
3789
3790
3791
3792
3793
3794
3795
3796
3797
3798
3799
3800
3801
3802
3803
3804
3805
3806
3807
3808
3809
3810
3811
3812
3813
3814
3815
3816
3817
3818
3819
3820
3821
3822
3823
3824
3825
3826
3827
3828
3829
3830
3831
3832
3833
3834
3835
3836
3837
3838
3839
3840
3841
3842
3843
3844
3845
3846
3847
3848
3849
3850
3851
3852
3853
3854
3855
3856
3857
3858
3859
3860
3861
3862
3863
3864
3865
3866
3867
3868
3869
3870
3871
3872
3873
3874
3875
3876
3877
3878
3879
3880
3881
3882
3883
3884
3885
3886
3887
3888
3889
3890
3891
3892
3893
3894
3895
3896
3897
3898
3899
3900
3901
3902
3903
3904
3905
3906
3907
3908
3909
3910
3911
3912
3913
3914
3915
3916
3917
3918
3919
3920
3921
3922
3923
3924
3925
3926
3927
3928
3929
3930
3931
3932
3933
3934
3935
3936
3937
3938
3939
3940
3941
3942
3943
3944
3945
3946
3947
3948
3949
3950
3951
3952
3953
3954
3955
3956
3957
3958
3959
3960
3961
3962
3963
3964
3965
3966
3967
3968
3969
3970
3971
3972
3973
3974
3975
3976
3977
3978
3979
3980
3981
3982
3983
3984
3985
3986
3987
3988
3989
3990
3991
3992
3993
3994
3995
3996
3997
3998
3999
4000
4001
4002
4003
4004
4005
4006
4007
4008
4009
4010
4011
4012
4013
4014
4015
4016
4017
4018
4019
4020
4021
4022
4023
4024
4025
4026
4027
4028
4029
4030
4031
4032
4033
4034
4035
4036
4037
4038
4039
4040
4041
4042
4043
4044
4045
4046
4047
4048
4049
4050
4051
4052
4053
4054
4055
4056
4057
4058
4059
4060
4061
4062
4063
4064
4065
4066
4067
4068
4069
4070
4071
4072
4073
4074
4075
4076
4077
4078
4079
4080
4081
4082
4083
4084
4085
4086
4087
4088
4089
4090
4091
4092
4093
4094
4095
4096
4097
4098
4099
4100
4101
4102
4103
4104
4105
4106
4107
4108
4109
4110
4111
4112
4113
4114
4115
4116
4117
4118
4119
4120
4121
4122
4123
4124
4125
4126
4127
4128
4129
4130
4131
4132
4133
4134
4135
4136
4137
4138
4139
4140
4141
4142
4143
4144
4145
4146
4147
4148
4149
4150
4151
4152
4153
4154
4155
4156
4157
4158
4159
4160
4161
4162
4163
4164
4165
4166
4167
4168
4169
4170
4171
4172
4173
4174
4175
4176
4177
4178
4179
4180
4181
4182
4183
4184
4185
4186
4187
4188
4189
4190
4191
4192
4193
4194
4195
4196
4197
4198
4199
4200
4201
4202
4203
4204
4205
4206
4207
4208
4209
4210
4211
4212
4213
4214
4215
4216
4217
4218
4219
4220
4221
4222
4223
4224
4225
4226
4227
4228
4229
4230
4231
4232
4233
4234
4235
4236
4237
4238
4239
4240
4241
4242
4243
4244
4245
4246
4247
4248
4249
4250
4251
4252
4253
4254
4255
4256
4257
4258
4259
4260
4261
4262
4263
4264
4265
4266
4267
4268
4269
4270
4271
4272
4273
4274
4275
4276
4277
4278
4279
4280
4281
4282
4283
4284
4285
4286
4287
4288
4289
4290
4291
4292
4293
4294
4295
4296
4297
4298
4299
4300
4301
4302
4303
4304
4305
4306
4307
4308
4309
4310
4311
4312
4313
4314
4315
4316
4317
4318
4319
4320
4321
4322
4323
4324
4325
4326
4327
4328
4329
4330
4331
4332
4333
4334
4335
4336
4337
4338
4339
4340
4341
4342
4343
4344
4345
4346
4347
4348
4349
4350
4351
4352
4353
4354
4355
4356
4357
4358
4359
4360
4361
4362
4363
4364
4365
4366
4367
4368
4369
4370
4371
4372
4373
4374
4375
4376
4377
4378
4379
4380
4381
4382
4383
4384
4385
4386
4387
4388
4389
4390
4391
4392
4393
4394
4395
4396
4397
4398
4399
4400
4401
4402
4403
4404
4405
4406
4407
4408
4409
4410
4411
4412
4413
4414
4415
4416
4417
4418
4419
4420
4421
4422
4423
4424
4425
4426
4427
4428
4429
4430
4431
4432
4433
4434
4435
4436
4437
4438
4439
4440
4441
4442
4443
4444
4445
4446
4447
4448
4449
4450
4451
4452
4453
4454
4455
4456
4457
4458
4459
4460
4461
4462
4463
4464
4465
4466
4467
4468
4469
4470
4471
4472
4473
4474
4475
4476
4477
4478
4479
4480
4481
4482
4483
4484
4485
4486
4487
4488
4489
4490
4491
4492
4493
4494
4495
4496
4497
4498
4499
4500
4501
4502
4503
4504
4505
4506
4507
4508
4509
4510
4511
4512
4513
4514
4515
4516
4517
4518
4519
4520
4521
4522
4523
4524
4525
4526
4527
4528
4529
4530
4531
4532
4533
4534
4535
4536
4537
4538
4539
4540
4541
4542
4543
4544
4545
4546
4547
4548
4549
4550
4551
4552
4553
4554
4555
4556
4557
4558
4559
4560
4561
4562
4563
4564
4565
4566
4567
4568
4569
4570
4571
4572
4573
4574
4575
4576
4577
4578
4579
4580
4581
4582
4583
4584
4585
4586
4587
4588
4589
4590
4591
4592
4593
4594
4595
4596
4597
4598
4599
4600
4601
4602
4603
4604
4605
4606
4607
4608
4609
4610
4611
4612
4613
4614
4615
4616
4617
4618
4619
4620
4621
4622
4623
4624
4625
4626
4627
4628
4629
4630
4631
4632
4633
4634
4635
4636
4637
4638
4639
4640
4641
4642
4643
4644
4645
4646
4647
4648
4649
4650
4651
4652
4653
4654
4655
4656
4657
4658
4659
4660
4661
4662
4663
4664
4665
4666
4667
4668
4669
4670
4671
4672
4673
4674
4675
4676
4677
4678
4679
4680
4681
4682
4683
4684
4685
4686
4687
4688
4689
4690
4691
4692
4693
4694
4695
4696
4697
4698
4699
4700
4701
4702
4703
4704
4705
4706
4707
4708
4709
4710
4711
4712
4713
4714
4715
4716
4717
4718
4719
4720
4721
4722
4723
4724
4725
4726
4727
4728
4729
4730
4731
4732
4733
4734
4735
4736
4737
4738
4739
4740
4741
4742
4743
4744
4745
4746
4747
4748
4749
4750
4751
4752
4753
4754
4755
4756
4757
4758
4759
4760
4761
4762
4763
4764
4765
4766
4767
4768
4769
4770
4771
4772
4773
4774
4775
4776
4777
4778
4779
4780
4781
4782
4783
4784
4785
4786
4787
4788
4789
4790
4791
4792
4793
4794
4795
4796
4797
4798
4799
4800
4801
4802
4803
4804
4805
4806
4807
4808
4809
4810
4811
4812
4813
4814
4815
4816
4817
4818
4819
4820
4821
4822
4823
4824
4825
4826
4827
4828
4829
4830
4831
4832
4833
4834
4835
4836
4837
4838
4839
4840
4841
4842
4843
4844
4845
4846
4847
4848
4849
4850
4851
4852
4853
4854
4855
4856
4857
4858
4859
4860
4861
4862
4863
4864
4865
4866
4867
4868
4869
4870
4871
4872
4873
4874
4875
4876
4877
4878
4879
4880
4881
4882
4883
4884
4885
4886
4887
4888
4889
4890
4891
4892
4893
4894
4895
4896
4897
4898
4899
4900
4901
4902
4903
4904
4905
4906
4907
4908
4909
4910
4911
4912
4913
4914
4915
4916
4917
4918
4919
4920
4921
4922
4923
4924
4925
4926
4927
4928
4929
4930
4931
4932
4933
4934
4935
4936
4937
4938
4939
4940
4941
4942
4943
4944
4945
4946
4947
4948
4949
4950
4951
4952
4953
4954
4955
4956
4957
4958
4959
4960
4961
4962
4963
4964
4965
4966
4967
4968
4969
4970
4971
4972
4973
4974
4975
4976
4977
4978
4979
4980
4981
4982
4983
4984
4985
4986
4987
4988
4989
4990
4991
4992
4993
4994
4995
4996
4997
4998
4999
5000
5001
5002
5003
5004
5005
5006
5007
5008
5009
5010
5011
5012
5013
5014
5015
5016
5017
5018
5019
5020
5021
5022
5023
5024
5025
5026
5027
5028
5029
5030
5031
5032
5033
5034
5035
5036
5037
5038
5039
5040
5041
5042
5043
5044
5045
5046
5047
5048
5049
5050
5051
5052
5053
5054
5055
5056
5057
5058
5059
5060
5061
5062
5063
5064
5065
5066
5067
5068
5069
5070
5071
5072
5073
5074
5075
5076
5077
5078
5079
5080
5081
5082
5083
5084
5085
5086
5087
5088
5089
5090
5091
5092
5093
5094
5095
5096
5097
5098
5099
5100
5101
5102
5103
5104
5105
5106
5107
5108
5109
5110
5111
5112
5113
5114
5115
5116
5117
5118
5119
5120
5121
5122
5123
5124
5125
5126
5127
5128
5129
5130
5131
5132
5133
5134
5135
5136
5137
5138
5139
5140
5141
5142
5143
5144
5145
5146
5147
5148
5149
5150
5151
5152
5153
5154
5155
5156
5157
5158
5159
5160
5161
5162
5163
5164
5165
5166
5167
5168
5169
5170
5171
5172
5173
5174
5175
5176
5177
5178
5179
5180
5181
5182
5183
5184
5185
5186
5187
5188
5189
5190
5191
5192
5193
5194
5195
5196
5197
5198
5199
5200
5201
5202
5203
5204
5205
5206
5207
5208
5209
5210
5211
5212
5213
5214
5215
5216
5217
5218
5219
5220
5221
5222
5223
5224
5225
5226
5227
5228
5229
5230
5231
5232
5233
5234
5235
5236
5237
5238
5239
5240
5241
5242
5243
5244
5245
5246
5247
5248
5249
5250
5251
5252
5253
5254
5255
5256
5257
5258
5259
5260
5261
5262
5263
5264
5265
5266
5267
5268
5269
5270
5271
5272
5273
5274
5275
5276
5277
5278
5279
5280
5281
5282
5283
5284
5285
5286
5287
5288
5289
5290
5291
5292
5293
5294
5295
5296
5297
5298
5299
5300
5301
5302
5303
5304
5305
5306
5307
5308
5309
5310
5311
5312
5313
5314
5315
5316
5317
5318
5319
5320
5321
5322
5323
5324
5325
5326
5327
5328
5329
5330
5331
5332
5333
5334
5335
5336
5337
5338
5339
5340
5341
5342
5343
5344
5345
5346
5347
5348
5349
5350
5351
5352
5353
5354
5355
5356
5357
5358
5359
5360
5361
5362
5363
5364
5365
5366
5367
5368
5369
5370
5371
5372
5373
5374
5375
5376
5377
5378
5379
5380
5381
5382
5383
5384
5385
5386
5387
5388
5389
5390
5391
5392
5393
5394
5395
5396
5397
5398
5399
5400
5401
5402
5403
5404
5405
5406
5407
5408
5409
5410
5411
5412
5413
5414
5415
5416
5417
5418
5419
5420
5421
5422
5423
5424
5425
5426
5427
5428
5429
5430
5431
5432
5433
5434
5435
5436
5437
5438
5439
5440
5441
5442
5443
5444
5445
5446
5447
5448
5449
5450
5451
5452
5453
5454
5455
5456
5457
5458
5459
5460
5461
5462
5463
5464
5465
5466
5467
5468
5469
5470
5471
5472
5473
5474
5475
5476
5477
5478
5479
5480
5481
5482
5483
5484
5485
5486
5487
5488
5489
5490
5491
5492
5493
5494
5495
5496
5497
5498
5499
5500
5501
5502
5503
5504
5505
5506
5507
5508
5509
5510
5511
5512
5513
5514
5515
5516
5517
5518
5519
5520
5521
5522
5523
5524
5525
5526
5527
5528
5529
5530
5531
5532
5533
5534
5535
5536
5537
5538
5539
5540
5541
5542
5543
5544
5545
5546
5547
5548
5549
5550
5551
5552
5553
5554
5555
5556
5557
5558
5559
5560
5561
5562
5563
5564
5565
5566
5567
5568
5569
5570
5571
5572
5573
5574
5575
5576
5577
5578
5579
5580
5581
5582
5583
5584
5585
5586
5587
5588
5589
5590
5591
5592
5593
5594
5595
5596
5597
5598
5599
5600
5601
5602
5603
5604
5605
5606
5607
5608
5609
5610
5611
5612
5613
5614
5615
5616
5617
5618
5619
5620
5621
5622
5623
5624
5625
5626
5627
5628
5629
5630
5631
5632
5633
5634
5635
5636
5637
5638
5639
5640
5641
5642
5643
5644
5645
5646
5647
5648
5649
5650
5651
5652
5653
5654
5655
5656
5657
5658
5659
5660
5661
5662
5663
5664
5665
5666
5667
5668
5669
5670
5671
5672
5673
5674
5675
5676
5677
5678
5679
5680
5681
5682
5683
5684
5685
5686
5687
5688
5689
5690
5691
5692
5693
5694
5695
5696
5697
5698
5699
5700
5701
5702
5703
5704
5705
5706
5707
5708
5709
5710
5711
5712
5713
5714
5715
5716
5717
5718
5719
5720
5721
5722
5723
5724
5725
5726
5727
5728
5729
5730
5731
5732
5733
5734
5735
5736
5737
5738
5739
5740
5741
5742
5743
5744
5745
5746
5747
5748
5749
5750
5751
5752
5753
5754
5755
5756
5757
5758
5759
5760
5761
5762
5763
5764
5765
5766
5767
5768
5769
5770
5771
5772
5773
5774
5775
5776
5777
5778
5779
5780
5781
5782
5783
5784
5785
5786
5787
5788
5789
5790
5791
5792
5793
5794
5795
5796
5797
5798
5799
5800
5801
5802
5803
5804
5805
5806
5807
5808
5809
5810
5811
5812
5813
5814
5815
5816
5817
5818
5819
5820
5821
5822
5823
5824
5825
5826
5827
5828
5829
5830
5831
5832
5833
5834
5835
5836
5837
5838
5839
5840
5841
5842
5843
5844
5845
5846
5847
5848
5849
5850
5851
5852
5853
5854
5855
5856
5857
5858
5859
5860
5861
5862
5863
5864
5865
5866
5867
5868
5869
5870
5871
5872
5873
5874
5875
5876
5877
5878
5879
5880
5881
5882
5883
5884
5885
5886
5887
5888
5889
5890
5891
5892
5893
5894
5895
5896
5897
5898
5899
5900
5901
5902
5903
5904
5905
5906
5907
5908
5909
5910
5911
5912
5913
5914
5915
5916
5917
5918
5919
5920
5921
5922
5923
5924
5925
5926
5927
5928
5929
5930
5931
5932
5933
5934
5935
5936
5937
5938
5939
5940
5941
5942
5943
5944
5945
5946
5947
5948
5949
5950
5951
5952
5953
5954
5955
5956
5957
5958
5959
5960
5961
5962
5963
5964
5965
5966
5967
5968
5969
5970
5971
5972
5973
5974
5975
5976
5977
5978
5979
5980
5981
5982
5983
5984
5985
5986
5987
5988
5989
5990
5991
5992
5993
5994
5995
5996
5997
5998
5999
6000
6001
6002
6003
6004
6005
6006
6007
6008
6009
6010
6011
6012
6013
6014
6015
6016
6017
6018
6019
6020
6021
6022
6023
6024
6025
6026
6027
6028
6029
6030
6031
6032
6033
6034
6035
6036
6037
6038
6039
6040
6041
6042
6043
6044
6045
6046
6047
6048
6049
6050
6051
6052
6053
6054
6055
6056
6057
6058
6059
6060
6061
6062
6063
6064
6065
6066
6067
6068
6069
6070
6071
6072
6073
6074
6075
6076
6077
6078
6079
6080
6081
6082
6083
6084
6085
6086
6087
6088
6089
6090
6091
6092
6093
6094
6095
6096
6097
6098
6099
6100
6101
6102
6103
6104
6105
6106
6107
6108
6109
6110
6111
6112
6113
6114
6115
6116
6117
6118
6119
6120
6121
6122
6123
6124
6125
6126
6127
6128
6129
6130
6131
6132
6133
6134
6135
6136
6137
6138
6139
6140
6141
6142
6143
6144
6145
6146
6147
6148
6149
6150
6151
6152
6153
6154
6155
6156
6157
6158
6159
6160
6161
6162
6163
6164
6165
6166
6167
6168
6169
6170
6171
6172
6173
6174
6175
6176
6177
6178
6179
6180
6181
6182
6183
6184
6185
6186
6187
6188
6189
6190
6191
6192
6193
6194
6195
6196
6197
6198
6199
6200
6201
6202
6203
6204
6205
6206
6207
6208
6209
6210
6211
6212
6213
6214
6215
6216
6217
6218
6219
6220
6221
6222
6223
6224
6225
6226
6227
6228
6229
6230
6231
6232
6233
6234
6235
6236
6237
6238
6239
6240
6241
6242
6243
6244
6245
6246
6247
6248
6249
6250
6251
6252
6253
6254
6255
6256
6257
6258
6259
6260
6261
6262
6263
6264
6265
6266
6267
6268
6269
6270
6271
6272
6273
6274
6275
6276
6277
6278
6279
6280
6281
6282
6283
6284
6285
6286
6287
6288
6289
6290
6291
6292
6293
6294
6295
6296
6297
6298
6299
6300
6301
6302
6303
6304
6305
6306
6307
6308
6309
6310
6311
6312
6313
6314
6315
6316
6317
6318
6319
6320
6321
6322
6323
6324
6325
6326
6327
6328
6329
6330
6331
6332
6333
6334
6335
6336
6337
6338
6339
6340
6341
6342
6343
6344
6345
6346
6347
6348
6349
6350
6351
6352
6353
6354
6355
6356
6357
6358
6359
6360
6361
6362
6363
6364
6365
6366
6367
6368
6369
6370
6371
6372
6373
6374
6375
6376
6377
6378
6379
6380
6381
6382
6383
6384
6385
6386
6387
6388
6389
6390
6391
6392
6393
6394
6395
6396
6397
6398
6399
6400
6401
6402
6403
6404
6405
6406
6407
6408
6409
6410
6411
6412
6413
6414
6415
6416
6417
6418
6419
6420
6421
6422
6423
6424
6425
6426
6427
6428
6429
6430
6431
6432
6433
6434
6435
6436
6437
6438
6439
6440
6441
6442
6443
6444
6445
6446
6447
6448
6449
6450
6451
6452
6453
6454
6455
6456
6457
6458
6459
6460
6461
6462
6463
6464
6465
6466
6467
6468
6469
6470
6471
6472
6473
6474
6475
6476
6477
6478
6479
6480
6481
6482
6483
6484
6485
6486
6487
6488
6489
6490
6491
6492
6493
6494
6495
6496
6497
6498
6499
6500
6501
6502
6503
6504
6505
6506
6507
6508
6509
6510
6511
6512
6513
6514
6515
6516
6517
6518
6519
6520
6521
6522
6523
6524
6525
6526
6527
6528
6529
6530
6531
6532
6533
6534
6535
6536
6537
6538
6539
6540
6541
6542
6543
6544
6545
6546
6547
6548
6549
6550
6551
6552
6553
6554
6555
6556
6557
6558
6559
6560
6561
6562
6563
6564
6565
6566
6567
6568
6569
6570
6571
6572
6573
6574
6575
6576
6577
6578
6579
6580
6581
6582
6583
6584
6585
6586
6587
6588
6589
6590
6591
6592
6593
6594
6595
6596
6597
6598
6599
6600
6601
6602
6603
6604
6605
6606
6607
6608
6609
6610
6611
6612
6613
6614
6615
6616
6617
6618
6619
6620
6621
6622
6623
6624
6625
6626
6627
6628
6629
6630
6631
6632
6633
6634
6635
6636
6637
6638
6639
6640
6641
6642
6643
6644
6645
6646
6647
6648
6649
6650
6651
6652
6653
6654
6655
6656
6657
6658
6659
6660
6661
6662
6663
6664
6665
6666
6667
6668
6669
6670
6671
6672
6673
6674
6675
6676
6677
6678
6679
6680
6681
6682
6683
6684
6685
6686
6687
6688
6689
6690
6691
6692
6693
6694
6695
6696
6697
6698
6699
6700
6701
6702
6703
6704
6705
6706
6707
6708
6709
6710
6711
6712
6713
6714
6715
6716
6717
6718
6719
6720
6721
6722
6723
6724
6725
6726
6727
6728
6729
6730
6731
6732
6733
6734
6735
6736
6737
6738
6739
6740
6741
6742
6743
6744
6745
6746
6747
6748
6749
6750
6751
6752
6753
6754
6755
6756
6757
6758
6759
6760
6761
6762
6763
6764
6765
6766
6767
6768
6769
6770
6771
6772
6773
6774
6775
6776
6777
6778
6779
6780
6781
6782
6783
6784
6785
6786
6787
6788
6789
6790
6791
6792
6793
6794
6795
6796
6797
6798
6799
6800
6801
6802
6803
6804
6805
6806
6807
6808
6809
6810
6811
6812
6813
6814
6815
6816
6817
6818
6819
6820
6821
6822
6823
6824
6825
6826
6827
6828
6829
6830
6831
6832
6833
6834
6835
6836
6837
6838
6839
6840
6841
6842
6843
6844
6845
6846
6847
6848
6849
6850
6851
6852
6853
6854
6855
6856
6857
6858
6859
6860
6861
6862
6863
6864
6865
6866
6867
6868
6869
6870
6871
6872
6873
6874
6875
6876
6877
6878
6879
6880
6881
6882
6883
6884
6885
6886
6887
6888
6889
6890
6891
6892
6893
6894
6895
6896
6897
6898
6899
6900
6901
6902
6903
6904
6905
6906
6907
6908
6909
6910
6911
6912
6913
6914
6915
6916
6917
6918
6919
6920
6921
6922
6923
6924
6925
6926
6927
6928
6929
6930
6931
6932
6933
6934
6935
6936
6937
6938
6939
6940
6941
6942
6943
6944
6945
6946
6947
6948
6949
6950
6951
6952
6953
6954
6955
6956
6957
6958
6959
6960
6961
6962
6963
6964
6965
6966
6967
6968
6969
6970
6971
6972
6973
6974
6975
6976
6977
6978
6979
6980
6981
6982
6983
6984
6985
6986
6987
6988
6989
6990
6991
6992
6993
6994
6995
6996
6997
6998
6999
7000
7001
7002
7003
7004
7005
7006
7007
7008
7009
7010
7011
7012
7013
7014
7015
7016
7017
7018
7019
7020
7021
7022
7023
7024
7025
7026
7027
7028
7029
7030
7031
7032
7033
7034
7035
7036
7037
7038
7039
7040
7041
7042
7043
7044
7045
7046
7047
7048
7049
7050
7051
7052
7053
7054
7055
7056
7057
7058
7059
7060
7061
7062
7063
7064
7065
7066
7067
7068
7069
7070
7071
7072
7073
7074
7075
7076
7077
7078
7079
7080
7081
7082
7083
7084
7085
7086
7087
7088
7089
7090
7091
7092
7093
7094
7095
7096
7097
7098
7099
7100
7101
7102
7103
7104
7105
7106
7107
7108
7109
7110
7111
7112
7113
7114
7115
7116
7117
7118
7119
7120
7121
7122
7123
7124
7125
7126
7127
7128
7129
7130
7131
7132
7133
7134
7135
7136
7137
7138
7139
7140
7141
7142
7143
7144
7145
7146
7147
7148
7149
7150
7151
7152
7153
7154
7155
7156
7157
7158
7159
7160
7161
7162
7163
7164
7165
7166
7167
7168
7169
7170
7171
7172
7173
7174
7175
7176
7177
7178
7179
7180
7181
7182
7183
7184
7185
7186
7187
7188
7189
7190
7191
7192
7193
7194
7195
7196
7197
7198
7199
7200
7201
7202
7203
7204
7205
7206
7207
7208
7209
7210
7211
7212
7213
7214
7215
7216
7217
7218
7219
7220
7221
7222
7223
7224
7225
7226
7227
7228
7229
7230
7231
7232
7233
7234
7235
7236
7237
7238
7239
7240
7241
7242
7243
7244
7245
7246
7247
7248
7249
7250
7251
7252
7253
7254
7255
7256
7257
7258
7259
7260
7261
7262
7263
7264
7265
7266
7267
7268
7269
7270
7271
7272
7273
7274
7275
7276
7277
7278
7279
7280
7281
7282
7283
7284
7285
7286
7287
7288
7289
7290
7291
7292
7293
7294
7295
7296
7297
7298
7299
7300
7301
7302
7303
7304
7305
7306
7307
7308
7309
7310
7311
7312
7313
7314
7315
7316
7317
7318
7319
7320
7321
7322
7323
7324
7325
7326
7327
7328
7329
7330
7331
7332
7333
7334
7335
7336
7337
7338
7339
7340
7341
7342
7343
7344
7345
7346
7347
7348
7349
7350
7351
7352
7353
7354
7355
7356
7357
7358
7359
7360
7361
7362
7363
7364
7365
7366
7367
7368
7369
7370
7371
7372
7373
7374
7375
7376
7377
7378
7379
7380
7381
7382
7383
7384
7385
7386
7387
7388
7389
7390
7391
7392
7393
7394
7395
7396
7397
7398
7399
7400
7401
7402
7403
7404
7405
7406
7407
7408
7409
7410
7411
7412
7413
7414
7415
7416
7417
7418
7419
7420
7421
7422
7423
7424
7425
7426
7427
7428
7429
7430
7431
7432
7433
7434
7435
7436
7437
7438
7439
7440
7441
7442
7443
7444
7445
7446
7447
7448
7449
7450
7451
7452
7453
7454
7455
7456
7457
7458
7459
7460
7461
7462
7463
7464
7465
7466
7467
7468
7469
7470
7471
7472
7473
7474
7475
7476
7477
7478
7479
7480
7481
7482
7483
7484
7485
7486
7487
7488
7489
7490
7491
7492
7493
7494
7495
7496
7497
7498
7499
7500
7501
7502
7503
7504
7505
7506
7507
7508
7509
7510
7511
7512
7513
7514
7515
7516
7517
7518
7519
7520
7521
7522
7523
7524
7525
7526
7527
7528
7529
7530
7531
7532
7533
7534
7535
7536
7537
7538
7539
7540
7541
7542
7543
7544
7545
7546
7547
7548
7549
7550
7551
7552
7553
7554
7555
7556
7557
7558
7559
7560
7561
7562
7563
7564
7565
7566
7567
7568
7569
7570
7571
7572
7573
7574
7575
7576
7577
7578
7579
7580
7581
7582
7583
7584
7585
7586
7587
7588
7589
7590
7591
7592
7593
7594
7595
7596
7597
7598
7599
7600
7601
7602
7603
7604
7605
7606
7607
7608
7609
7610
7611
7612
7613
7614
7615
7616
7617
7618
7619
7620
7621
7622
7623
7624
7625
7626
7627
7628
7629
7630
7631
7632
7633
7634
7635
7636
7637
7638
7639
7640
7641
7642
7643
7644
7645
7646
7647
7648
7649
7650
7651
7652
7653
7654
7655
7656
7657
7658
7659
7660
7661
7662
7663
7664
7665
7666
7667
7668
7669
7670
7671
7672
7673
7674
7675
7676
7677
7678
7679
7680
7681
7682
7683
7684
7685
7686
7687
7688
7689
7690
7691
7692
7693
7694
7695
7696
7697
7698
7699
7700
7701
7702
7703
7704
7705
7706
7707
7708
7709
7710
7711
7712
7713
7714
7715
7716
7717
7718
7719
7720
7721
7722
7723
7724
7725
7726
7727
7728
7729
7730
7731
7732
7733
7734
7735
7736
7737
7738
7739
7740
7741
7742
7743
7744
7745
7746
7747
7748
7749
7750
7751
7752
7753
7754
7755
7756
7757
7758
7759
7760
7761
7762
7763
7764
7765
7766
7767
7768
7769
7770
7771
7772
7773
7774
7775
7776
7777
7778
7779
7780
7781
7782
7783
7784
7785
7786
7787
7788
7789
7790
7791
7792
7793
7794
7795
7796
7797
7798
7799
7800
7801
7802
7803
7804
7805
7806
7807
7808
7809
7810
7811
7812
7813
7814
7815
7816
7817
7818
7819
7820
7821
7822
7823
7824
7825
7826
7827
7828
7829
7830
7831
7832
7833
7834
7835
7836
7837
7838
7839
7840
7841
7842
7843
7844
7845
7846
7847
7848
7849
7850
7851
7852
7853
7854
7855
7856
7857
7858
7859
7860
7861
7862
7863
7864
7865
7866
7867
7868
7869
7870
7871
7872
7873
7874
7875
7876
7877
7878
7879
7880
7881
7882
7883
7884
7885
7886
7887
7888
7889
7890
7891
7892
7893
7894
7895
7896
7897
7898
7899
7900
7901
7902
7903
7904
7905
7906
7907
7908
7909
7910
7911
7912
7913
7914
7915
7916
7917
7918
7919
7920
7921
7922
7923
7924
7925
7926
7927
7928
7929
7930
7931
7932
7933
7934
7935
7936
7937
7938
7939
7940
7941
7942
7943
7944
7945
7946
7947
7948
7949
7950
7951
7952
7953
7954
7955
7956
7957
7958
7959
7960
7961
7962
7963
7964
7965
7966
7967
7968
7969
7970
7971
7972
7973
7974
7975
7976
7977
7978
7979
7980
7981
7982
7983
7984
7985
7986
7987
7988
7989
7990
7991
7992
7993
7994
7995
7996
7997
7998
7999
8000
8001
8002
8003
8004
8005
8006
8007
8008
8009
8010
8011
8012
8013
8014
8015
8016
8017
8018
8019
8020
8021
8022
8023
8024
8025
8026
8027
8028
8029
8030
8031
8032
8033
8034
8035
8036
8037
8038
8039
8040
8041
8042
8043
8044
8045
8046
8047
8048
8049
8050
8051
8052
8053
8054
8055
8056
8057
8058
8059
8060
8061
8062
8063
8064
8065
8066
8067
8068
8069
8070
8071
8072
8073
8074
8075
8076
8077
8078
8079
8080
8081
8082
8083
8084
8085
8086
8087
8088
8089
8090
8091
8092
8093
8094
8095
8096
8097
8098
8099
8100
8101
8102
8103
8104
/*
 * Copyright 2012 Advanced Micro Devices, Inc.
 * All Rights Reserved.
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
 * copy of this software and associated documentation files (the "Software"),
 * to deal in the Software without restriction, including without limitation
 * on the rights to use, copy, modify, merge, publish, distribute, sub
 * license, and/or sell copies of the Software, and to permit persons to whom
 * the Software is furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice (including the next
 * paragraph) shall be included in all copies or substantial portions of the
 * Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
 * USE OR OTHER DEALINGS IN THE SOFTWARE.
 */

#include "util/u_memory.h"
#include "util/u_string.h"
#include "tgsi/tgsi_build.h"
#include "tgsi/tgsi_strings.h"
#include "tgsi/tgsi_util.h"
#include "tgsi/tgsi_dump.h"

#include "ac_exp_param.h"
#include "ac_shader_util.h"
#include "ac_llvm_util.h"
#include "si_shader_internal.h"
#include "si_pipe.h"
#include "sid.h"

#include "compiler/nir/nir.h"

static const char *scratch_rsrc_dword0_symbol =
	"SCRATCH_RSRC_DWORD0";

static const char *scratch_rsrc_dword1_symbol =
	"SCRATCH_RSRC_DWORD1";

struct si_shader_output_values
{
	LLVMValueRef values[4];
	unsigned semantic_name;
	unsigned semantic_index;
	ubyte vertex_stream[4];
};

static void si_init_shader_ctx(struct si_shader_context *ctx,
			       struct si_screen *sscreen,
			       struct ac_llvm_compiler *compiler);

static void si_llvm_emit_barrier(const struct lp_build_tgsi_action *action,
				 struct lp_build_tgsi_context *bld_base,
				 struct lp_build_emit_data *emit_data);

static void si_dump_shader_key(unsigned processor, const struct si_shader *shader,
			       FILE *f);

static void si_build_vs_prolog_function(struct si_shader_context *ctx,
					union si_shader_part_key *key);
static void si_build_tcs_epilog_function(struct si_shader_context *ctx,
					 union si_shader_part_key *key);
static void si_build_ps_prolog_function(struct si_shader_context *ctx,
					union si_shader_part_key *key);
static void si_build_ps_epilog_function(struct si_shader_context *ctx,
					union si_shader_part_key *key);
static void si_fix_resource_usage(struct si_screen *sscreen,
				  struct si_shader *shader);

/* Ideally pass the sample mask input to the PS epilog as v14, which
 * is its usual location, so that the shader doesn't have to add v_mov.
 */
#define PS_EPILOG_SAMPLEMASK_MIN_LOC 14

static bool llvm_type_is_64bit(struct si_shader_context *ctx,
			       LLVMTypeRef type)
{
	if (type == ctx->ac.i64 || type == ctx->ac.f64)
		return true;

	return false;
}

static bool is_merged_shader(struct si_shader_context *ctx)
{
	if (ctx->screen->info.chip_class <= GFX8)
		return false;

	return ctx->shader->key.as_ls ||
	       ctx->shader->key.as_es ||
	       ctx->type == PIPE_SHADER_TESS_CTRL ||
	       ctx->type == PIPE_SHADER_GEOMETRY;
}

void si_init_function_info(struct si_function_info *fninfo)
{
	fninfo->num_params = 0;
	fninfo->num_sgpr_params = 0;
}

unsigned add_arg_assign(struct si_function_info *fninfo,
			enum si_arg_regfile regfile, LLVMTypeRef type,
			LLVMValueRef *assign)
{
	assert(regfile != ARG_SGPR || fninfo->num_sgpr_params == fninfo->num_params);

	unsigned idx = fninfo->num_params++;
	assert(idx < ARRAY_SIZE(fninfo->types));

	if (regfile == ARG_SGPR)
		fninfo->num_sgpr_params = fninfo->num_params;

	fninfo->types[idx] = type;
	fninfo->assign[idx] = assign;
	return idx;
}

static unsigned add_arg(struct si_function_info *fninfo,
			enum si_arg_regfile regfile, LLVMTypeRef type)
{
	return add_arg_assign(fninfo, regfile, type, NULL);
}

static void add_arg_assign_checked(struct si_function_info *fninfo,
				   enum si_arg_regfile regfile, LLVMTypeRef type,
				   LLVMValueRef *assign, unsigned idx)
{
	MAYBE_UNUSED unsigned actual = add_arg_assign(fninfo, regfile, type, assign);
	assert(actual == idx);
}

static void add_arg_checked(struct si_function_info *fninfo,
			    enum si_arg_regfile regfile, LLVMTypeRef type,
			    unsigned idx)
{
	add_arg_assign_checked(fninfo, regfile, type, NULL, idx);
}

/**
 * Returns a unique index for a per-patch semantic name and index. The index
 * must be less than 32, so that a 32-bit bitmask of used inputs or outputs
 * can be calculated.
 */
unsigned si_shader_io_get_unique_index_patch(unsigned semantic_name, unsigned index)
{
	switch (semantic_name) {
	case TGSI_SEMANTIC_TESSOUTER:
		return 0;
	case TGSI_SEMANTIC_TESSINNER:
		return 1;
	case TGSI_SEMANTIC_PATCH:
		assert(index < 30);
		return 2 + index;

	default:
		assert(!"invalid semantic name");
		return 0;
	}
}

/**
 * Returns a unique index for a semantic name and index. The index must be
 * less than 64, so that a 64-bit bitmask of used inputs or outputs can be
 * calculated.
 */
unsigned si_shader_io_get_unique_index(unsigned semantic_name, unsigned index,
				       unsigned is_varying)
{
	switch (semantic_name) {
	case TGSI_SEMANTIC_POSITION:
		return 0;
	case TGSI_SEMANTIC_GENERIC:
		/* Since some shader stages use the the highest used IO index
		 * to determine the size to allocate for inputs/outputs
		 * (in LDS, tess and GS rings). GENERIC should be placed right
		 * after POSITION to make that size as small as possible.
		 */
		if (index < SI_MAX_IO_GENERIC)
			return 1 + index;

		assert(!"invalid generic index");
		return 0;
	case TGSI_SEMANTIC_PSIZE:
		return SI_MAX_IO_GENERIC + 1;
	case TGSI_SEMANTIC_CLIPDIST:
		assert(index <= 1);
		return SI_MAX_IO_GENERIC + 2 + index;
	case TGSI_SEMANTIC_FOG:
		return SI_MAX_IO_GENERIC + 4;
	case TGSI_SEMANTIC_LAYER:
		return SI_MAX_IO_GENERIC + 5;
	case TGSI_SEMANTIC_VIEWPORT_INDEX:
		return SI_MAX_IO_GENERIC + 6;
	case TGSI_SEMANTIC_PRIMID:
		return SI_MAX_IO_GENERIC + 7;
	case TGSI_SEMANTIC_COLOR:
		assert(index < 2);
		return SI_MAX_IO_GENERIC + 8 + index;
	case TGSI_SEMANTIC_BCOLOR:
		assert(index < 2);
		/* If it's a varying, COLOR and BCOLOR alias. */
		if (is_varying)
			return SI_MAX_IO_GENERIC + 8 + index;
		else
			return SI_MAX_IO_GENERIC + 10 + index;
	case TGSI_SEMANTIC_TEXCOORD:
		assert(index < 8);
		STATIC_ASSERT(SI_MAX_IO_GENERIC + 12 + 8 <= 63);
		return SI_MAX_IO_GENERIC + 12 + index;
	case TGSI_SEMANTIC_CLIPVERTEX:
		return 63;
	default:
		fprintf(stderr, "invalid semantic name = %u\n", semantic_name);
		assert(!"invalid semantic name");
		return 0;
	}
}

/**
 * Get the value of a shader input parameter and extract a bitfield.
 */
static LLVMValueRef unpack_llvm_param(struct si_shader_context *ctx,
				      LLVMValueRef value, unsigned rshift,
				      unsigned bitwidth)
{
	if (LLVMGetTypeKind(LLVMTypeOf(value)) == LLVMFloatTypeKind)
		value = ac_to_integer(&ctx->ac, value);

	if (rshift)
		value = LLVMBuildLShr(ctx->ac.builder, value,
				      LLVMConstInt(ctx->i32, rshift, 0), "");

	if (rshift + bitwidth < 32) {
		unsigned mask = (1 << bitwidth) - 1;
		value = LLVMBuildAnd(ctx->ac.builder, value,
				     LLVMConstInt(ctx->i32, mask, 0), "");
	}

	return value;
}

LLVMValueRef si_unpack_param(struct si_shader_context *ctx,
			     unsigned param, unsigned rshift,
			     unsigned bitwidth)
{
	LLVMValueRef value = LLVMGetParam(ctx->main_fn, param);

	return unpack_llvm_param(ctx, value, rshift, bitwidth);
}

static LLVMValueRef get_rel_patch_id(struct si_shader_context *ctx)
{
	switch (ctx->type) {
	case PIPE_SHADER_TESS_CTRL:
		return unpack_llvm_param(ctx, ctx->abi.tcs_rel_ids, 0, 8);

	case PIPE_SHADER_TESS_EVAL:
		return LLVMGetParam(ctx->main_fn,
				    ctx->param_tes_rel_patch_id);

	default:
		assert(0);
		return NULL;
	}
}

/* Tessellation shaders pass outputs to the next shader using LDS.
 *
 * LS outputs = TCS inputs
 * TCS outputs = TES inputs
 *
 * The LDS layout is:
 * - TCS inputs for patch 0
 * - TCS inputs for patch 1
 * - TCS inputs for patch 2		= get_tcs_in_current_patch_offset (if RelPatchID==2)
 * - ...
 * - TCS outputs for patch 0            = get_tcs_out_patch0_offset
 * - Per-patch TCS outputs for patch 0  = get_tcs_out_patch0_patch_data_offset
 * - TCS outputs for patch 1
 * - Per-patch TCS outputs for patch 1
 * - TCS outputs for patch 2            = get_tcs_out_current_patch_offset (if RelPatchID==2)
 * - Per-patch TCS outputs for patch 2  = get_tcs_out_current_patch_data_offset (if RelPatchID==2)
 * - ...
 *
 * All three shaders VS(LS), TCS, TES share the same LDS space.
 */

static LLVMValueRef
get_tcs_in_patch_stride(struct si_shader_context *ctx)
{
	return si_unpack_param(ctx, ctx->param_vs_state_bits, 8, 13);
}

static unsigned get_tcs_out_vertex_dw_stride_constant(struct si_shader_context *ctx)
{
	assert(ctx->type == PIPE_SHADER_TESS_CTRL);

	if (ctx->shader->key.mono.u.ff_tcs_inputs_to_copy)
		return util_last_bit64(ctx->shader->key.mono.u.ff_tcs_inputs_to_copy) * 4;

	return util_last_bit64(ctx->shader->selector->outputs_written) * 4;
}

static LLVMValueRef get_tcs_out_vertex_dw_stride(struct si_shader_context *ctx)
{
	unsigned stride = get_tcs_out_vertex_dw_stride_constant(ctx);

	return LLVMConstInt(ctx->i32, stride, 0);
}

static LLVMValueRef get_tcs_out_patch_stride(struct si_shader_context *ctx)
{
	if (ctx->shader->key.mono.u.ff_tcs_inputs_to_copy)
		return si_unpack_param(ctx, ctx->param_tcs_out_lds_layout, 0, 13);

	const struct tgsi_shader_info *info = &ctx->shader->selector->info;
	unsigned tcs_out_vertices = info->properties[TGSI_PROPERTY_TCS_VERTICES_OUT];
	unsigned vertex_dw_stride = get_tcs_out_vertex_dw_stride_constant(ctx);
	unsigned num_patch_outputs = util_last_bit64(ctx->shader->selector->patch_outputs_written);
	unsigned patch_dw_stride = tcs_out_vertices * vertex_dw_stride +
				   num_patch_outputs * 4;
	return LLVMConstInt(ctx->i32, patch_dw_stride, 0);
}

static LLVMValueRef
get_tcs_out_patch0_offset(struct si_shader_context *ctx)
{
	return LLVMBuildMul(ctx->ac.builder,
			    si_unpack_param(ctx,
					    ctx->param_tcs_out_lds_offsets,
					    0, 16),
			    LLVMConstInt(ctx->i32, 4, 0), "");
}

static LLVMValueRef
get_tcs_out_patch0_patch_data_offset(struct si_shader_context *ctx)
{
	return LLVMBuildMul(ctx->ac.builder,
			    si_unpack_param(ctx,
					    ctx->param_tcs_out_lds_offsets,
					    16, 16),
			    LLVMConstInt(ctx->i32, 4, 0), "");
}

static LLVMValueRef
get_tcs_in_current_patch_offset(struct si_shader_context *ctx)
{
	LLVMValueRef patch_stride = get_tcs_in_patch_stride(ctx);
	LLVMValueRef rel_patch_id = get_rel_patch_id(ctx);

	return LLVMBuildMul(ctx->ac.builder, patch_stride, rel_patch_id, "");
}

static LLVMValueRef
get_tcs_out_current_patch_offset(struct si_shader_context *ctx)
{
	LLVMValueRef patch0_offset = get_tcs_out_patch0_offset(ctx);
	LLVMValueRef patch_stride = get_tcs_out_patch_stride(ctx);
	LLVMValueRef rel_patch_id = get_rel_patch_id(ctx);

	return ac_build_imad(&ctx->ac, patch_stride, rel_patch_id, patch0_offset);
}

static LLVMValueRef
get_tcs_out_current_patch_data_offset(struct si_shader_context *ctx)
{
	LLVMValueRef patch0_patch_data_offset =
		get_tcs_out_patch0_patch_data_offset(ctx);
	LLVMValueRef patch_stride = get_tcs_out_patch_stride(ctx);
	LLVMValueRef rel_patch_id = get_rel_patch_id(ctx);

	return ac_build_imad(&ctx->ac, patch_stride, rel_patch_id, patch0_patch_data_offset);
}

static LLVMValueRef get_num_tcs_out_vertices(struct si_shader_context *ctx)
{
	unsigned tcs_out_vertices =
		ctx->shader->selector ?
		ctx->shader->selector->info.properties[TGSI_PROPERTY_TCS_VERTICES_OUT] : 0;

	/* If !tcs_out_vertices, it's either the fixed-func TCS or the TCS epilog. */
	if (ctx->type == PIPE_SHADER_TESS_CTRL && tcs_out_vertices)
		return LLVMConstInt(ctx->i32, tcs_out_vertices, 0);

	return si_unpack_param(ctx, ctx->param_tcs_offchip_layout, 6, 6);
}

static LLVMValueRef get_tcs_in_vertex_dw_stride(struct si_shader_context *ctx)
{
	unsigned stride;

	switch (ctx->type) {
	case PIPE_SHADER_VERTEX:
		stride = ctx->shader->selector->lshs_vertex_stride / 4;
		return LLVMConstInt(ctx->i32, stride, 0);

	case PIPE_SHADER_TESS_CTRL:
		if (ctx->screen->info.chip_class >= GFX9 &&
		    ctx->shader->is_monolithic) {
			stride = ctx->shader->key.part.tcs.ls->lshs_vertex_stride / 4;
			return LLVMConstInt(ctx->i32, stride, 0);
		}
		return si_unpack_param(ctx, ctx->param_vs_state_bits, 24, 8);

	default:
		assert(0);
		return NULL;
	}
}

static LLVMValueRef unpack_sint16(struct si_shader_context *ctx,
				 LLVMValueRef i32, unsigned index)
{
	assert(index <= 1);

	if (index == 1)
		return LLVMBuildAShr(ctx->ac.builder, i32,
				     LLVMConstInt(ctx->i32, 16, 0), "");

	return LLVMBuildSExt(ctx->ac.builder,
			     LLVMBuildTrunc(ctx->ac.builder, i32,
					    ctx->ac.i16, ""),
			     ctx->i32, "");
}

void si_llvm_load_input_vs(
	struct si_shader_context *ctx,
	unsigned input_index,
	LLVMValueRef out[4])
{
	const struct tgsi_shader_info *info = &ctx->shader->selector->info;
	unsigned vs_blit_property = info->properties[TGSI_PROPERTY_VS_BLIT_SGPRS];

	if (vs_blit_property) {
		LLVMValueRef vertex_id = ctx->abi.vertex_id;
		LLVMValueRef sel_x1 = LLVMBuildICmp(ctx->ac.builder,
						    LLVMIntULE, vertex_id,
						    ctx->i32_1, "");
		/* Use LLVMIntNE, because we have 3 vertices and only
		 * the middle one should use y2.
		 */
		LLVMValueRef sel_y1 = LLVMBuildICmp(ctx->ac.builder,
						    LLVMIntNE, vertex_id,
						    ctx->i32_1, "");

		if (input_index == 0) {
			/* Position: */
			LLVMValueRef x1y1 = LLVMGetParam(ctx->main_fn,
							 ctx->param_vs_blit_inputs);
			LLVMValueRef x2y2 = LLVMGetParam(ctx->main_fn,
							 ctx->param_vs_blit_inputs + 1);

			LLVMValueRef x1 = unpack_sint16(ctx, x1y1, 0);
			LLVMValueRef y1 = unpack_sint16(ctx, x1y1, 1);
			LLVMValueRef x2 = unpack_sint16(ctx, x2y2, 0);
			LLVMValueRef y2 = unpack_sint16(ctx, x2y2, 1);

			LLVMValueRef x = LLVMBuildSelect(ctx->ac.builder, sel_x1,
							 x1, x2, "");
			LLVMValueRef y = LLVMBuildSelect(ctx->ac.builder, sel_y1,
							 y1, y2, "");

			out[0] = LLVMBuildSIToFP(ctx->ac.builder, x, ctx->f32, "");
			out[1] = LLVMBuildSIToFP(ctx->ac.builder, y, ctx->f32, "");
			out[2] = LLVMGetParam(ctx->main_fn,
					      ctx->param_vs_blit_inputs + 2);
			out[3] = ctx->ac.f32_1;
			return;
		}

		/* Color or texture coordinates: */
		assert(input_index == 1);

		if (vs_blit_property == SI_VS_BLIT_SGPRS_POS_COLOR) {
			for (int i = 0; i < 4; i++) {
				out[i] = LLVMGetParam(ctx->main_fn,
						      ctx->param_vs_blit_inputs + 3 + i);
			}
		} else {
			assert(vs_blit_property == SI_VS_BLIT_SGPRS_POS_TEXCOORD);
			LLVMValueRef x1 = LLVMGetParam(ctx->main_fn,
						       ctx->param_vs_blit_inputs + 3);
			LLVMValueRef y1 = LLVMGetParam(ctx->main_fn,
						       ctx->param_vs_blit_inputs + 4);
			LLVMValueRef x2 = LLVMGetParam(ctx->main_fn,
						       ctx->param_vs_blit_inputs + 5);
			LLVMValueRef y2 = LLVMGetParam(ctx->main_fn,
						       ctx->param_vs_blit_inputs + 6);

			out[0] = LLVMBuildSelect(ctx->ac.builder, sel_x1,
						 x1, x2, "");
			out[1] = LLVMBuildSelect(ctx->ac.builder, sel_y1,
						 y1, y2, "");
			out[2] = LLVMGetParam(ctx->main_fn,
					      ctx->param_vs_blit_inputs + 7);
			out[3] = LLVMGetParam(ctx->main_fn,
					      ctx->param_vs_blit_inputs + 8);
		}
		return;
	}

	union si_vs_fix_fetch fix_fetch;
	LLVMValueRef t_list_ptr;
	LLVMValueRef t_offset;
	LLVMValueRef t_list;
	LLVMValueRef vertex_index;
	LLVMValueRef tmp;

	/* Load the T list */
	t_list_ptr = LLVMGetParam(ctx->main_fn, ctx->param_vertex_buffers);

	t_offset = LLVMConstInt(ctx->i32, input_index, 0);

	t_list = ac_build_load_to_sgpr(&ctx->ac, t_list_ptr, t_offset);

	vertex_index = LLVMGetParam(ctx->main_fn,
				    ctx->param_vertex_index0 +
				    input_index);

	/* Use the open-coded implementation for all loads of doubles and
	 * of dword-sized data that needs fixups. We need to insert conversion
	 * code anyway, and the amd/common code does it for us.
	 *
	 * Note: On LLVM <= 8, we can only open-code formats with
	 * channel size >= 4 bytes.
	 */
	bool opencode = ctx->shader->key.mono.vs_fetch_opencode & (1 << input_index);
	fix_fetch.bits = ctx->shader->key.mono.vs_fix_fetch[input_index].bits;
	if (opencode ||
	    (fix_fetch.u.log_size == 3 && fix_fetch.u.format == AC_FETCH_FORMAT_FLOAT) ||
	    (fix_fetch.u.log_size == 2)) {
		tmp = ac_build_opencoded_load_format(
				&ctx->ac, fix_fetch.u.log_size, fix_fetch.u.num_channels_m1 + 1,
				fix_fetch.u.format, fix_fetch.u.reverse, !opencode,
				t_list, vertex_index, ctx->ac.i32_0, ctx->ac.i32_0,
				false, false, true);
		for (unsigned i = 0; i < 4; ++i)
			out[i] = LLVMBuildExtractElement(ctx->ac.builder, tmp, LLVMConstInt(ctx->i32, i, false), "");
		return;
	}

	/* Do multiple loads for special formats. */
	unsigned required_channels = util_last_bit(info->input_usage_mask[input_index]);
	LLVMValueRef fetches[4];
	unsigned num_fetches;
	unsigned fetch_stride;
	unsigned channels_per_fetch;

	if (fix_fetch.u.log_size <= 1 && fix_fetch.u.num_channels_m1 == 2) {
		num_fetches = MIN2(required_channels, 3);
		fetch_stride = 1 << fix_fetch.u.log_size;
		channels_per_fetch = 1;
	} else {
		num_fetches = 1;
		fetch_stride = 0;
		channels_per_fetch = required_channels;
	}

	for (unsigned i = 0; i < num_fetches; ++i) {
		LLVMValueRef voffset = LLVMConstInt(ctx->i32, fetch_stride * i, 0);
		fetches[i] = ac_build_buffer_load_format(&ctx->ac, t_list, vertex_index, voffset,
							 channels_per_fetch, false, true);
	}

	if (num_fetches == 1 && channels_per_fetch > 1) {
		LLVMValueRef fetch = fetches[0];
		for (unsigned i = 0; i < channels_per_fetch; ++i) {
			tmp = LLVMConstInt(ctx->i32, i, false);
			fetches[i] = LLVMBuildExtractElement(
				ctx->ac.builder, fetch, tmp, "");
		}
		num_fetches = channels_per_fetch;
		channels_per_fetch = 1;
	}

	for (unsigned i = num_fetches; i < 4; ++i)
		fetches[i] = LLVMGetUndef(ctx->f32);

	if (fix_fetch.u.log_size <= 1 && fix_fetch.u.num_channels_m1 == 2 &&
	    required_channels == 4) {
		if (fix_fetch.u.format == AC_FETCH_FORMAT_UINT || fix_fetch.u.format == AC_FETCH_FORMAT_SINT)
			fetches[3] = ctx->ac.i32_1;
		else
			fetches[3] = ctx->ac.f32_1;
	} else if (fix_fetch.u.log_size == 3 &&
		   (fix_fetch.u.format == AC_FETCH_FORMAT_SNORM ||
		    fix_fetch.u.format == AC_FETCH_FORMAT_SSCALED ||
		    fix_fetch.u.format == AC_FETCH_FORMAT_SINT) &&
		   required_channels == 4) {
		/* For 2_10_10_10, the hardware returns an unsigned value;
		 * convert it to a signed one.
		 */
		LLVMValueRef tmp = fetches[3];
		LLVMValueRef c30 = LLVMConstInt(ctx->i32, 30, 0);

		/* First, recover the sign-extended signed integer value. */
		if (fix_fetch.u.format == AC_FETCH_FORMAT_SSCALED)
			tmp = LLVMBuildFPToUI(ctx->ac.builder, tmp, ctx->i32, "");
		else
			tmp = ac_to_integer(&ctx->ac, tmp);

		/* For the integer-like cases, do a natural sign extension.
		 *
		 * For the SNORM case, the values are 0.0, 0.333, 0.666, 1.0
		 * and happen to contain 0, 1, 2, 3 as the two LSBs of the
		 * exponent.
		 */
		tmp = LLVMBuildShl(ctx->ac.builder, tmp,
				   fix_fetch.u.format == AC_FETCH_FORMAT_SNORM ?
				   LLVMConstInt(ctx->i32, 7, 0) : c30, "");
		tmp = LLVMBuildAShr(ctx->ac.builder, tmp, c30, "");

		/* Convert back to the right type. */
		if (fix_fetch.u.format == AC_FETCH_FORMAT_SNORM) {
			LLVMValueRef clamp;
			LLVMValueRef neg_one = LLVMConstReal(ctx->f32, -1.0);
			tmp = LLVMBuildSIToFP(ctx->ac.builder, tmp, ctx->f32, "");
			clamp = LLVMBuildFCmp(ctx->ac.builder, LLVMRealULT, tmp, neg_one, "");
			tmp = LLVMBuildSelect(ctx->ac.builder, clamp, neg_one, tmp, "");
		} else if (fix_fetch.u.format == AC_FETCH_FORMAT_SSCALED) {
			tmp = LLVMBuildSIToFP(ctx->ac.builder, tmp, ctx->f32, "");
		}

		fetches[3] = tmp;
	}

	for (unsigned i = 0; i < 4; ++i)
		out[i] = ac_to_float(&ctx->ac, fetches[i]);
}

static void declare_input_vs(
	struct si_shader_context *ctx,
	unsigned input_index,
	const struct tgsi_full_declaration *decl,
	LLVMValueRef out[4])
{
	si_llvm_load_input_vs(ctx, input_index, out);
}

static LLVMValueRef get_primitive_id(struct si_shader_context *ctx,
				     unsigned swizzle)
{
	if (swizzle > 0)
		return ctx->i32_0;

	switch (ctx->type) {
	case PIPE_SHADER_VERTEX:
		return LLVMGetParam(ctx->main_fn,
				    ctx->param_vs_prim_id);
	case PIPE_SHADER_TESS_CTRL:
		return ctx->abi.tcs_patch_id;
	case PIPE_SHADER_TESS_EVAL:
		return ctx->abi.tes_patch_id;
	case PIPE_SHADER_GEOMETRY:
		return ctx->abi.gs_prim_id;
	default:
		assert(0);
		return ctx->i32_0;
	}
}

/**
 * Return the value of tgsi_ind_register for indexing.
 * This is the indirect index with the constant offset added to it.
 */
LLVMValueRef si_get_indirect_index(struct si_shader_context *ctx,
				   const struct tgsi_ind_register *ind,
				   unsigned addr_mul,
				   int rel_index)
{
	LLVMValueRef result;

	if (ind->File == TGSI_FILE_ADDRESS) {
		result = ctx->addrs[ind->Index][ind->Swizzle];
		result = LLVMBuildLoad(ctx->ac.builder, result, "");
	} else {
		struct tgsi_full_src_register src = {};

		src.Register.File = ind->File;
		src.Register.Index = ind->Index;

		/* Set the second index to 0 for constants. */
		if (ind->File == TGSI_FILE_CONSTANT)
			src.Register.Dimension = 1;

		result = ctx->bld_base.emit_fetch_funcs[ind->File](&ctx->bld_base, &src,
								   TGSI_TYPE_SIGNED,
								   ind->Swizzle);
		result = ac_to_integer(&ctx->ac, result);
	}

	return ac_build_imad(&ctx->ac, result, LLVMConstInt(ctx->i32, addr_mul, 0),
			     LLVMConstInt(ctx->i32, rel_index, 0));
}

/**
 * Like si_get_indirect_index, but restricts the return value to a (possibly
 * undefined) value inside [0..num).
 */
LLVMValueRef si_get_bounded_indirect_index(struct si_shader_context *ctx,
					   const struct tgsi_ind_register *ind,
					   int rel_index, unsigned num)
{
	LLVMValueRef result = si_get_indirect_index(ctx, ind, 1, rel_index);

	return si_llvm_bound_index(ctx, result, num);
}

static LLVMValueRef get_dw_address_from_generic_indices(struct si_shader_context *ctx,
							LLVMValueRef vertex_dw_stride,
							LLVMValueRef base_addr,
							LLVMValueRef vertex_index,
							LLVMValueRef param_index,
							unsigned input_index,
							ubyte *name,
							ubyte *index,
							bool is_patch)
{
	if (vertex_dw_stride) {
		base_addr = ac_build_imad(&ctx->ac, vertex_index,
					  vertex_dw_stride, base_addr);
	}

	if (param_index) {
		base_addr = ac_build_imad(&ctx->ac, param_index,
					  LLVMConstInt(ctx->i32, 4, 0), base_addr);
	}

	int param = is_patch ?
		si_shader_io_get_unique_index_patch(name[input_index],
						    index[input_index]) :
		si_shader_io_get_unique_index(name[input_index],
					      index[input_index], false);

	/* Add the base address of the element. */
	return LLVMBuildAdd(ctx->ac.builder, base_addr,
			    LLVMConstInt(ctx->i32, param * 4, 0), "");
}

/**
 * Calculate a dword address given an input or output register and a stride.
 */
static LLVMValueRef get_dw_address(struct si_shader_context *ctx,
				   const struct tgsi_full_dst_register *dst,
				   const struct tgsi_full_src_register *src,
				   LLVMValueRef vertex_dw_stride,
				   LLVMValueRef base_addr)
{
	struct tgsi_shader_info *info = &ctx->shader->selector->info;
	ubyte *name, *index, *array_first;
	int input_index;
	struct tgsi_full_dst_register reg;
	LLVMValueRef vertex_index = NULL;
	LLVMValueRef ind_index = NULL;

	/* Set the register description. The address computation is the same
	 * for sources and destinations. */
	if (src) {
		reg.Register.File = src->Register.File;
		reg.Register.Index = src->Register.Index;
		reg.Register.Indirect = src->Register.Indirect;
		reg.Register.Dimension = src->Register.Dimension;
		reg.Indirect = src->Indirect;
		reg.Dimension = src->Dimension;
		reg.DimIndirect = src->DimIndirect;
	} else
		reg = *dst;

	/* If the register is 2-dimensional (e.g. an array of vertices
	 * in a primitive), calculate the base address of the vertex. */
	if (reg.Register.Dimension) {
		if (reg.Dimension.Indirect)
			vertex_index = si_get_indirect_index(ctx, &reg.DimIndirect,
						      1, reg.Dimension.Index);
		else
			vertex_index = LLVMConstInt(ctx->i32, reg.Dimension.Index, 0);
	}

	/* Get information about the register. */
	if (reg.Register.File == TGSI_FILE_INPUT) {
		name = info->input_semantic_name;
		index = info->input_semantic_index;
		array_first = info->input_array_first;
	} else if (reg.Register.File == TGSI_FILE_OUTPUT) {
		name = info->output_semantic_name;
		index = info->output_semantic_index;
		array_first = info->output_array_first;
	} else {
		assert(0);
		return NULL;
	}

	if (reg.Register.Indirect) {
		/* Add the relative address of the element. */
		if (reg.Indirect.ArrayID)
			input_index = array_first[reg.Indirect.ArrayID];
		else
			input_index = reg.Register.Index;

		ind_index = si_get_indirect_index(ctx, &reg.Indirect,
						  1, reg.Register.Index - input_index);
	} else {
		input_index = reg.Register.Index;
	}

	return get_dw_address_from_generic_indices(ctx, vertex_dw_stride,
						   base_addr, vertex_index,
						   ind_index, input_index,
						   name, index,
						   !reg.Register.Dimension);
}

/* The offchip buffer layout for TCS->TES is
 *
 * - attribute 0 of patch 0 vertex 0
 * - attribute 0 of patch 0 vertex 1
 * - attribute 0 of patch 0 vertex 2
 *   ...
 * - attribute 0 of patch 1 vertex 0
 * - attribute 0 of patch 1 vertex 1
 *   ...
 * - attribute 1 of patch 0 vertex 0
 * - attribute 1 of patch 0 vertex 1
 *   ...
 * - per patch attribute 0 of patch 0
 * - per patch attribute 0 of patch 1
 *   ...
 *
 * Note that every attribute has 4 components.
 */
static LLVMValueRef get_tcs_tes_buffer_address(struct si_shader_context *ctx,
					       LLVMValueRef rel_patch_id,
                                               LLVMValueRef vertex_index,
                                               LLVMValueRef param_index)
{
	LLVMValueRef base_addr, vertices_per_patch, num_patches, total_vertices;
	LLVMValueRef param_stride, constant16;

	vertices_per_patch = get_num_tcs_out_vertices(ctx);
	num_patches = si_unpack_param(ctx, ctx->param_tcs_offchip_layout, 0, 6);
	total_vertices = LLVMBuildMul(ctx->ac.builder, vertices_per_patch,
	                              num_patches, "");

	constant16 = LLVMConstInt(ctx->i32, 16, 0);
	if (vertex_index) {
		base_addr = ac_build_imad(&ctx->ac, rel_patch_id,
					  vertices_per_patch, vertex_index);
		param_stride = total_vertices;
	} else {
		base_addr = rel_patch_id;
		param_stride = num_patches;
	}

	base_addr = ac_build_imad(&ctx->ac, param_index, param_stride, base_addr);
	base_addr = LLVMBuildMul(ctx->ac.builder, base_addr, constant16, "");

	if (!vertex_index) {
		LLVMValueRef patch_data_offset =
		           si_unpack_param(ctx, ctx->param_tcs_offchip_layout, 12, 20);

		base_addr = LLVMBuildAdd(ctx->ac.builder, base_addr,
		                         patch_data_offset, "");
	}
	return base_addr;
}

/* This is a generic helper that can be shared by the NIR and TGSI backends */
static LLVMValueRef get_tcs_tes_buffer_address_from_generic_indices(
					struct si_shader_context *ctx,
					LLVMValueRef vertex_index,
					LLVMValueRef param_index,
					unsigned param_base,
					ubyte *name,
					ubyte *index,
					bool is_patch)
{
	unsigned param_index_base;

	param_index_base = is_patch ?
		si_shader_io_get_unique_index_patch(name[param_base], index[param_base]) :
		si_shader_io_get_unique_index(name[param_base], index[param_base], false);

	if (param_index) {
		param_index = LLVMBuildAdd(ctx->ac.builder, param_index,
					   LLVMConstInt(ctx->i32, param_index_base, 0),
					   "");
	} else {
		param_index = LLVMConstInt(ctx->i32, param_index_base, 0);
	}

	return get_tcs_tes_buffer_address(ctx, get_rel_patch_id(ctx),
					  vertex_index, param_index);
}

static LLVMValueRef get_tcs_tes_buffer_address_from_reg(
                                       struct si_shader_context *ctx,
                                       const struct tgsi_full_dst_register *dst,
                                       const struct tgsi_full_src_register *src)
{
	struct tgsi_shader_info *info = &ctx->shader->selector->info;
	ubyte *name, *index, *array_first;
	struct tgsi_full_src_register reg;
	LLVMValueRef vertex_index = NULL;
	LLVMValueRef param_index = NULL;
	unsigned param_base;

	reg = src ? *src : tgsi_full_src_register_from_dst(dst);

	if (reg.Register.Dimension) {

		if (reg.Dimension.Indirect)
			vertex_index = si_get_indirect_index(ctx, &reg.DimIndirect,
							     1, reg.Dimension.Index);
		else
			vertex_index = LLVMConstInt(ctx->i32, reg.Dimension.Index, 0);
	}

	/* Get information about the register. */
	if (reg.Register.File == TGSI_FILE_INPUT) {
		name = info->input_semantic_name;
		index = info->input_semantic_index;
		array_first = info->input_array_first;
	} else if (reg.Register.File == TGSI_FILE_OUTPUT) {
		name = info->output_semantic_name;
		index = info->output_semantic_index;
		array_first = info->output_array_first;
	} else {
		assert(0);
		return NULL;
	}

	if (reg.Register.Indirect) {
		if (reg.Indirect.ArrayID)
			param_base = array_first[reg.Indirect.ArrayID];
		else
			param_base = reg.Register.Index;

		param_index = si_get_indirect_index(ctx, &reg.Indirect,
						    1, reg.Register.Index - param_base);

	} else {
		param_base = reg.Register.Index;
	}

	return get_tcs_tes_buffer_address_from_generic_indices(ctx, vertex_index,
							       param_index, param_base,
							       name, index, !reg.Register.Dimension);
}

static LLVMValueRef buffer_load(struct lp_build_tgsi_context *bld_base,
                                LLVMTypeRef type, unsigned swizzle,
                                LLVMValueRef buffer, LLVMValueRef offset,
                                LLVMValueRef base, bool can_speculate)
{
	struct si_shader_context *ctx = si_shader_context(bld_base);
	LLVMValueRef value, value2;
	LLVMTypeRef vec_type = LLVMVectorType(type, 4);

	if (swizzle == ~0) {
		value = ac_build_buffer_load(&ctx->ac, buffer, 4, NULL, base, offset,
					     0, 1, 0, can_speculate, false);

		return LLVMBuildBitCast(ctx->ac.builder, value, vec_type, "");
	}

	if (!llvm_type_is_64bit(ctx, type)) {
		value = ac_build_buffer_load(&ctx->ac, buffer, 4, NULL, base, offset,
					     0, 1, 0, can_speculate, false);

		value = LLVMBuildBitCast(ctx->ac.builder, value, vec_type, "");
		return LLVMBuildExtractElement(ctx->ac.builder, value,
		                    LLVMConstInt(ctx->i32, swizzle, 0), "");
	}

	value = ac_build_buffer_load(&ctx->ac, buffer, 1, NULL, base, offset,
	                          swizzle * 4, 1, 0, can_speculate, false);

	value2 = ac_build_buffer_load(&ctx->ac, buffer, 1, NULL, base, offset,
	                           swizzle * 4 + 4, 1, 0, can_speculate, false);

	return si_llvm_emit_fetch_64bit(bld_base, type, value, value2);
}

/**
 * Load from LDS.
 *
 * \param type		output value type
 * \param swizzle	offset (typically 0..3); it can be ~0, which loads a vec4
 * \param dw_addr	address in dwords
 */
static LLVMValueRef lds_load(struct lp_build_tgsi_context *bld_base,
			     LLVMTypeRef type, unsigned swizzle,
			     LLVMValueRef dw_addr)
{
	struct si_shader_context *ctx = si_shader_context(bld_base);
	LLVMValueRef value;

	if (swizzle == ~0) {
		LLVMValueRef values[TGSI_NUM_CHANNELS];

		for (unsigned chan = 0; chan < TGSI_NUM_CHANNELS; chan++)
			values[chan] = lds_load(bld_base, type, chan, dw_addr);

		return ac_build_gather_values(&ctx->ac, values,
					      TGSI_NUM_CHANNELS);
	}

	/* Split 64-bit loads. */
	if (llvm_type_is_64bit(ctx, type)) {
		LLVMValueRef lo, hi;

		lo = lds_load(bld_base, ctx->i32, swizzle, dw_addr);
		hi = lds_load(bld_base, ctx->i32, swizzle + 1, dw_addr);
		return si_llvm_emit_fetch_64bit(bld_base, type, lo, hi);
	}

	dw_addr = LLVMBuildAdd(ctx->ac.builder, dw_addr,
			       LLVMConstInt(ctx->i32, swizzle, 0), "");

	value = ac_lds_load(&ctx->ac, dw_addr);

	return LLVMBuildBitCast(ctx->ac.builder, value, type, "");
}

/**
 * Store to LDS.
 *
 * \param swizzle	offset (typically 0..3)
 * \param dw_addr	address in dwords
 * \param value		value to store
 */
static void lds_store(struct si_shader_context *ctx,
		      unsigned dw_offset_imm, LLVMValueRef dw_addr,
		      LLVMValueRef value)
{
	dw_addr = LLVMBuildAdd(ctx->ac.builder, dw_addr,
			       LLVMConstInt(ctx->i32, dw_offset_imm, 0), "");

	ac_lds_store(&ctx->ac, dw_addr, value);
}

enum si_tess_ring {
	TCS_FACTOR_RING,
	TESS_OFFCHIP_RING_TCS,
	TESS_OFFCHIP_RING_TES,
};

static LLVMValueRef get_tess_ring_descriptor(struct si_shader_context *ctx,
					     enum si_tess_ring ring)
{
	LLVMBuilderRef builder = ctx->ac.builder;
	unsigned param = ring == TESS_OFFCHIP_RING_TES ? ctx->param_tes_offchip_addr :
							 ctx->param_tcs_out_lds_layout;
	LLVMValueRef addr = LLVMGetParam(ctx->main_fn, param);

	/* TCS only receives high 13 bits of the address. */
	if (ring == TESS_OFFCHIP_RING_TCS || ring == TCS_FACTOR_RING) {
		addr = LLVMBuildAnd(builder, addr,
				    LLVMConstInt(ctx->i32, 0xfff80000, 0), "");
	}

	if (ring == TCS_FACTOR_RING) {
		unsigned tf_offset = ctx->screen->tess_offchip_ring_size;
		addr = LLVMBuildAdd(builder, addr,
				    LLVMConstInt(ctx->i32, tf_offset, 0), "");
	}

	LLVMValueRef desc[4];
	desc[0] = addr;
	desc[1] = LLVMConstInt(ctx->i32,
			       S_008F04_BASE_ADDRESS_HI(ctx->screen->info.address32_hi), 0);
	desc[2] = LLVMConstInt(ctx->i32, 0xffffffff, 0);
	desc[3] = LLVMConstInt(ctx->i32,
			       S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) |
			       S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
			       S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) |
			       S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W) |
			       S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) |
			       S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32), 0);

	return ac_build_gather_values(&ctx->ac, desc, 4);
}

static LLVMValueRef fetch_input_tcs(
	struct lp_build_tgsi_context *bld_base,
	const struct tgsi_full_src_register *reg,
	enum tgsi_opcode_type type, unsigned swizzle_in)
{
	struct si_shader_context *ctx = si_shader_context(bld_base);
	LLVMValueRef dw_addr, stride;
	unsigned swizzle = swizzle_in & 0xffff;
	stride = get_tcs_in_vertex_dw_stride(ctx);
	dw_addr = get_tcs_in_current_patch_offset(ctx);
	dw_addr = get_dw_address(ctx, NULL, reg, stride, dw_addr);

	return lds_load(bld_base, tgsi2llvmtype(bld_base, type), swizzle, dw_addr);
}

static LLVMValueRef si_nir_load_tcs_varyings(struct ac_shader_abi *abi,
					     LLVMTypeRef type,
					     LLVMValueRef vertex_index,
					     LLVMValueRef param_index,
					     unsigned const_index,
					     unsigned location,
					     unsigned driver_location,
					     unsigned component,
					     unsigned num_components,
					     bool is_patch,
					     bool is_compact,
					     bool load_input)
{
	struct si_shader_context *ctx = si_shader_context_from_abi(abi);
	struct tgsi_shader_info *info = &ctx->shader->selector->info;
	struct lp_build_tgsi_context *bld_base = &ctx->bld_base;
	LLVMValueRef dw_addr, stride;

	driver_location = driver_location / 4;

	if (load_input) {
		stride = get_tcs_in_vertex_dw_stride(ctx);
		dw_addr = get_tcs_in_current_patch_offset(ctx);
	} else {
		if (is_patch) {
			stride = NULL;
			dw_addr = get_tcs_out_current_patch_data_offset(ctx);
		} else {
			stride = get_tcs_out_vertex_dw_stride(ctx);
			dw_addr = get_tcs_out_current_patch_offset(ctx);
		}
	}

	if (param_index) {
		/* Add the constant index to the indirect index */
		param_index = LLVMBuildAdd(ctx->ac.builder, param_index,
					   LLVMConstInt(ctx->i32, const_index, 0), "");
	} else {
		param_index = LLVMConstInt(ctx->i32, const_index, 0);
	}

	ubyte *names;
	ubyte *indices;
	if (load_input) {
		names = info->input_semantic_name;
		indices = info->input_semantic_index;
	} else {
		names = info->output_semantic_name;
		indices = info->output_semantic_index;
	}

	dw_addr = get_dw_address_from_generic_indices(ctx, stride, dw_addr,
						      vertex_index, param_index,
						      driver_location,
						      names, indices,
						      is_patch);

	LLVMValueRef value[4];
	for (unsigned i = 0; i < num_components; i++) {
		unsigned offset = i;
		if (llvm_type_is_64bit(ctx, type))
			offset *= 2;

		offset += component;
		value[i + component] = lds_load(bld_base, type, offset, dw_addr);
	}

	return ac_build_varying_gather_values(&ctx->ac, value, num_components, component);
}

static LLVMValueRef fetch_output_tcs(
		struct lp_build_tgsi_context *bld_base,
		const struct tgsi_full_src_register *reg,
		enum tgsi_opcode_type type, unsigned swizzle_in)
{
	struct si_shader_context *ctx = si_shader_context(bld_base);
	LLVMValueRef dw_addr, stride;
	unsigned swizzle = (swizzle_in & 0xffff);

	if (reg->Register.Dimension) {
		stride = get_tcs_out_vertex_dw_stride(ctx);
		dw_addr = get_tcs_out_current_patch_offset(ctx);
		dw_addr = get_dw_address(ctx, NULL, reg, stride, dw_addr);
	} else {
		dw_addr = get_tcs_out_current_patch_data_offset(ctx);
		dw_addr = get_dw_address(ctx, NULL, reg, NULL, dw_addr);
	}

	return lds_load(bld_base, tgsi2llvmtype(bld_base, type), swizzle, dw_addr);
}

static LLVMValueRef fetch_input_tes(
	struct lp_build_tgsi_context *bld_base,
	const struct tgsi_full_src_register *reg,
	enum tgsi_opcode_type type, unsigned swizzle_in)
{
	struct si_shader_context *ctx = si_shader_context(bld_base);
	LLVMValueRef base, addr;
	unsigned swizzle = (swizzle_in & 0xffff);

	base = LLVMGetParam(ctx->main_fn, ctx->param_tcs_offchip_offset);
	addr = get_tcs_tes_buffer_address_from_reg(ctx, NULL, reg);

	return buffer_load(bld_base, tgsi2llvmtype(bld_base, type), swizzle,
			   ctx->tess_offchip_ring, base, addr, true);
}

LLVMValueRef si_nir_load_input_tes(struct ac_shader_abi *abi,
				   LLVMTypeRef type,
				   LLVMValueRef vertex_index,
				   LLVMValueRef param_index,
				   unsigned const_index,
				   unsigned location,
				   unsigned driver_location,
				   unsigned component,
				   unsigned num_components,
				   bool is_patch,
				   bool is_compact,
				   bool load_input)
{
	struct si_shader_context *ctx = si_shader_context_from_abi(abi);
	struct tgsi_shader_info *info = &ctx->shader->selector->info;
	LLVMValueRef base, addr;

	driver_location = driver_location / 4;

	base = LLVMGetParam(ctx->main_fn, ctx->param_tcs_offchip_offset);

	if (param_index) {
		/* Add the constant index to the indirect index */
		param_index = LLVMBuildAdd(ctx->ac.builder, param_index,
					   LLVMConstInt(ctx->i32, const_index, 0), "");
	} else {
		param_index = LLVMConstInt(ctx->i32, const_index, 0);
	}

	addr = get_tcs_tes_buffer_address_from_generic_indices(ctx, vertex_index,
							       param_index, driver_location,
							       info->input_semantic_name,
							       info->input_semantic_index,
							       is_patch);

	/* TODO: This will generate rather ordinary llvm code, although it
	 * should be easy for the optimiser to fix up. In future we might want
	 * to refactor buffer_load(), but for now this maximises code sharing
	 * between the NIR and TGSI backends.
	 */
	LLVMValueRef value[4];
	for (unsigned i = 0; i < num_components; i++) {
		unsigned offset = i;
		if (llvm_type_is_64bit(ctx, type))
			offset *= 2;

		offset += component;
		value[i + component] = buffer_load(&ctx->bld_base, type, offset,
						   ctx->tess_offchip_ring, base, addr, true);
	}

	return ac_build_varying_gather_values(&ctx->ac, value, num_components, component);
}

static void store_output_tcs(struct lp_build_tgsi_context *bld_base,
			     const struct tgsi_full_instruction *inst,
			     const struct tgsi_opcode_info *info,
			     unsigned index,
			     LLVMValueRef dst[4])
{
	struct si_shader_context *ctx = si_shader_context(bld_base);
	const struct tgsi_full_dst_register *reg = &inst->Dst[index];
	const struct tgsi_shader_info *sh_info = &ctx->shader->selector->info;
	unsigned chan_index;
	LLVMValueRef dw_addr, stride;
	LLVMValueRef buffer, base, buf_addr;
	LLVMValueRef values[4];
	bool skip_lds_store;
	bool is_tess_factor = false, is_tess_inner = false;

	/* Only handle per-patch and per-vertex outputs here.
	 * Vectors will be lowered to scalars and this function will be called again.
	 */
	if (reg->Register.File != TGSI_FILE_OUTPUT ||
	    (dst[0] && LLVMGetTypeKind(LLVMTypeOf(dst[0])) == LLVMVectorTypeKind)) {
		si_llvm_emit_store(bld_base, inst, info, index, dst);
		return;
	}

	if (reg->Register.Dimension) {
		stride = get_tcs_out_vertex_dw_stride(ctx);
		dw_addr = get_tcs_out_current_patch_offset(ctx);
		dw_addr = get_dw_address(ctx, reg, NULL, stride, dw_addr);
		skip_lds_store = !sh_info->reads_pervertex_outputs;
	} else {
		dw_addr = get_tcs_out_current_patch_data_offset(ctx);
		dw_addr = get_dw_address(ctx, reg, NULL, NULL, dw_addr);
		skip_lds_store = !sh_info->reads_perpatch_outputs;

		if (!reg->Register.Indirect) {
			int name = sh_info->output_semantic_name[reg->Register.Index];

			/* Always write tess factors into LDS for the TCS epilog. */
			if (name == TGSI_SEMANTIC_TESSINNER ||
			    name == TGSI_SEMANTIC_TESSOUTER) {
				/* The epilog doesn't read LDS if invocation 0 defines tess factors. */
				skip_lds_store = !sh_info->reads_tessfactor_outputs &&
						 ctx->shader->selector->tcs_info.tessfactors_are_def_in_all_invocs;
				is_tess_factor = true;
				is_tess_inner = name == TGSI_SEMANTIC_TESSINNER;
			}
		}
	}

	buffer = get_tess_ring_descriptor(ctx, TESS_OFFCHIP_RING_TCS);

	base = LLVMGetParam(ctx->main_fn, ctx->param_tcs_offchip_offset);
	buf_addr = get_tcs_tes_buffer_address_from_reg(ctx, reg, NULL);

	uint32_t writemask = reg->Register.WriteMask;
	while (writemask) {
		chan_index = u_bit_scan(&writemask);
		LLVMValueRef value = dst[chan_index];

		if (inst->Instruction.Saturate)
			value = ac_build_clamp(&ctx->ac, value);

		/* Skip LDS stores if there is no LDS read of this output. */
		if (!skip_lds_store)
			lds_store(ctx, chan_index, dw_addr, value);

		value = ac_to_integer(&ctx->ac, value);
		values[chan_index] = value;

		if (reg->Register.WriteMask != 0xF && !is_tess_factor) {
			ac_build_buffer_store_dword(&ctx->ac, buffer, value, 1,
						    buf_addr, base,
						    4 * chan_index, 1, 0, true, false);
		}

		/* Write tess factors into VGPRs for the epilog. */
		if (is_tess_factor &&
		    ctx->shader->selector->tcs_info.tessfactors_are_def_in_all_invocs) {
			if (!is_tess_inner) {
				LLVMBuildStore(ctx->ac.builder, value, /* outer */
					       ctx->invoc0_tess_factors[chan_index]);
			} else if (chan_index < 2) {
				LLVMBuildStore(ctx->ac.builder, value, /* inner */
					       ctx->invoc0_tess_factors[4 + chan_index]);
			}
		}
	}

	if (reg->Register.WriteMask == 0xF && !is_tess_factor) {
		LLVMValueRef value = ac_build_gather_values(&ctx->ac,
		                                            values, 4);
		ac_build_buffer_store_dword(&ctx->ac, buffer, value, 4, buf_addr,
					    base, 0, 1, 0, true, false);
	}
}

static void si_nir_store_output_tcs(struct ac_shader_abi *abi,
				    const struct nir_variable *var,
				    LLVMValueRef vertex_index,
				    LLVMValueRef param_index,
				    unsigned const_index,
				    LLVMValueRef src,
				    unsigned writemask)
{
	struct si_shader_context *ctx = si_shader_context_from_abi(abi);
	struct tgsi_shader_info *info = &ctx->shader->selector->info;
	const unsigned component = var->data.location_frac;
	const bool is_patch = var->data.patch;
	unsigned driver_location = var->data.driver_location;
	LLVMValueRef dw_addr, stride;
	LLVMValueRef buffer, base, addr;
	LLVMValueRef values[4];
	bool skip_lds_store;
	bool is_tess_factor = false, is_tess_inner = false;

	driver_location = driver_location / 4;

	if (param_index) {
		/* Add the constant index to the indirect index */
		param_index = LLVMBuildAdd(ctx->ac.builder, param_index,
					   LLVMConstInt(ctx->i32, const_index, 0), "");
	} else {
		if (const_index != 0)
			param_index = LLVMConstInt(ctx->i32, const_index, 0);
	}

	if (!is_patch) {
		stride = get_tcs_out_vertex_dw_stride(ctx);
		dw_addr = get_tcs_out_current_patch_offset(ctx);
		dw_addr = get_dw_address_from_generic_indices(ctx, stride, dw_addr,
							      vertex_index, param_index,
							      driver_location,
							      info->output_semantic_name,
							      info->output_semantic_index,
							      is_patch);

		skip_lds_store = !info->reads_pervertex_outputs;
	} else {
		dw_addr = get_tcs_out_current_patch_data_offset(ctx);
		dw_addr = get_dw_address_from_generic_indices(ctx, NULL, dw_addr,
							      vertex_index, param_index,
							      driver_location,
							      info->output_semantic_name,
							      info->output_semantic_index,
							      is_patch);

		skip_lds_store = !info->reads_perpatch_outputs;

		if (!param_index) {
			int name = info->output_semantic_name[driver_location];

			/* Always write tess factors into LDS for the TCS epilog. */
			if (name == TGSI_SEMANTIC_TESSINNER ||
			    name == TGSI_SEMANTIC_TESSOUTER) {
				/* The epilog doesn't read LDS if invocation 0 defines tess factors. */
				skip_lds_store = !info->reads_tessfactor_outputs &&
						 ctx->shader->selector->tcs_info.tessfactors_are_def_in_all_invocs;
				is_tess_factor = true;
				is_tess_inner = name == TGSI_SEMANTIC_TESSINNER;
			}
		}
	}

	buffer = get_tess_ring_descriptor(ctx, TESS_OFFCHIP_RING_TCS);

	base = LLVMGetParam(ctx->main_fn, ctx->param_tcs_offchip_offset);

	addr = get_tcs_tes_buffer_address_from_generic_indices(ctx, vertex_index,
							       param_index, driver_location,
							       info->output_semantic_name,
							       info->output_semantic_index,
							       is_patch);

	for (unsigned chan = 0; chan < 4; chan++) {
		if (!(writemask & (1 << chan)))
			continue;
		LLVMValueRef value = ac_llvm_extract_elem(&ctx->ac, src, chan - component);

		/* Skip LDS stores if there is no LDS read of this output. */
		if (!skip_lds_store)
			lds_store(ctx, chan, dw_addr, value);

		value = ac_to_integer(&ctx->ac, value);
		values[chan] = value;

		if (writemask != 0xF && !is_tess_factor) {
			ac_build_buffer_store_dword(&ctx->ac, buffer, value, 1,
						    addr, base,
						    4 * chan, 1, 0, true, false);
		}

		/* Write tess factors into VGPRs for the epilog. */
		if (is_tess_factor &&
		    ctx->shader->selector->tcs_info.tessfactors_are_def_in_all_invocs) {
			if (!is_tess_inner) {
				LLVMBuildStore(ctx->ac.builder, value, /* outer */
					       ctx->invoc0_tess_factors[chan]);
			} else if (chan < 2) {
				LLVMBuildStore(ctx->ac.builder, value, /* inner */
					       ctx->invoc0_tess_factors[4 + chan]);
			}
		}
	}

	if (writemask == 0xF && !is_tess_factor) {
		LLVMValueRef value = ac_build_gather_values(&ctx->ac,
		                                            values, 4);
		ac_build_buffer_store_dword(&ctx->ac, buffer, value, 4, addr,
					    base, 0, 1, 0, true, false);
	}
}

LLVMValueRef si_llvm_load_input_gs(struct ac_shader_abi *abi,
				   unsigned input_index,
				   unsigned vtx_offset_param,
				   LLVMTypeRef type,
				   unsigned swizzle)
{
	struct si_shader_context *ctx = si_shader_context_from_abi(abi);
	struct lp_build_tgsi_context *bld_base = &ctx->bld_base;
	struct si_shader *shader = ctx->shader;
	LLVMValueRef vtx_offset, soffset;
	struct tgsi_shader_info *info = &shader->selector->info;
	unsigned semantic_name = info->input_semantic_name[input_index];
	unsigned semantic_index = info->input_semantic_index[input_index];
	unsigned param;
	LLVMValueRef value;

	param = si_shader_io_get_unique_index(semantic_name, semantic_index, false);

	/* GFX9 has the ESGS ring in LDS. */
	if (ctx->screen->info.chip_class >= GFX9) {
		unsigned index = vtx_offset_param;

		switch (index / 2) {
		case 0:
			vtx_offset = si_unpack_param(ctx, ctx->param_gs_vtx01_offset,
						  index % 2 ? 16 : 0, 16);
			break;
		case 1:
			vtx_offset = si_unpack_param(ctx, ctx->param_gs_vtx23_offset,
						  index % 2 ? 16 : 0, 16);
			break;
		case 2:
			vtx_offset = si_unpack_param(ctx, ctx->param_gs_vtx45_offset,
						  index % 2 ? 16 : 0, 16);
			break;
		default:
			assert(0);
			return NULL;
		}

		vtx_offset = LLVMBuildAdd(ctx->ac.builder, vtx_offset,
					  LLVMConstInt(ctx->i32, param * 4, 0), "");
		return lds_load(bld_base, type, swizzle, vtx_offset);
	}

	/* GFX6: input load from the ESGS ring in memory. */
	if (swizzle == ~0) {
		LLVMValueRef values[TGSI_NUM_CHANNELS];
		unsigned chan;
		for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
			values[chan] = si_llvm_load_input_gs(abi, input_index, vtx_offset_param,
							     type, chan);
		}
		return ac_build_gather_values(&ctx->ac, values,
					      TGSI_NUM_CHANNELS);
	}

	/* Get the vertex offset parameter on GFX6. */
	LLVMValueRef gs_vtx_offset = ctx->gs_vtx_offset[vtx_offset_param];

	vtx_offset = LLVMBuildMul(ctx->ac.builder, gs_vtx_offset,
				  LLVMConstInt(ctx->i32, 4, 0), "");

	soffset = LLVMConstInt(ctx->i32, (param * 4 + swizzle) * 256, 0);

	value = ac_build_buffer_load(&ctx->ac, ctx->esgs_ring, 1, ctx->i32_0,
				     vtx_offset, soffset, 0, 1, 0, true, false);
	if (llvm_type_is_64bit(ctx, type)) {
		LLVMValueRef value2;
		soffset = LLVMConstInt(ctx->i32, (param * 4 + swizzle + 1) * 256, 0);

		value2 = ac_build_buffer_load(&ctx->ac, ctx->esgs_ring, 1,
					      ctx->i32_0, vtx_offset, soffset,
					      0, 1, 0, true, false);
		return si_llvm_emit_fetch_64bit(bld_base, type, value, value2);
	}
	return LLVMBuildBitCast(ctx->ac.builder, value, type, "");
}

static LLVMValueRef si_nir_load_input_gs(struct ac_shader_abi *abi,
					 unsigned location,
					 unsigned driver_location,
					 unsigned component,
					 unsigned num_components,
					 unsigned vertex_index,
					 unsigned const_index,
					 LLVMTypeRef type)
{
	struct si_shader_context *ctx = si_shader_context_from_abi(abi);

	LLVMValueRef value[4];
	for (unsigned i = 0; i < num_components; i++) {
		unsigned offset = i;
		if (llvm_type_is_64bit(ctx, type))
			offset *= 2;

		offset += component;
		value[i + component] = si_llvm_load_input_gs(&ctx->abi, driver_location  / 4,
							     vertex_index, type, offset);
	}

	return ac_build_varying_gather_values(&ctx->ac, value, num_components, component);
}

static LLVMValueRef fetch_input_gs(
	struct lp_build_tgsi_context *bld_base,
	const struct tgsi_full_src_register *reg,
	enum tgsi_opcode_type type,
	unsigned swizzle_in)
{
	struct si_shader_context *ctx = si_shader_context(bld_base);
	struct tgsi_shader_info *info = &ctx->shader->selector->info;
	unsigned swizzle = swizzle_in & 0xffff;

	unsigned semantic_name = info->input_semantic_name[reg->Register.Index];
	if (swizzle != ~0 && semantic_name == TGSI_SEMANTIC_PRIMID)
		return get_primitive_id(ctx, swizzle);

	if (!reg->Register.Dimension)
		return NULL;

	return si_llvm_load_input_gs(&ctx->abi, reg->Register.Index,
				     reg->Dimension.Index,
				     tgsi2llvmtype(bld_base, type),
				     swizzle);
}

static int lookup_interp_param_index(unsigned interpolate, unsigned location)
{
	switch (interpolate) {
	case TGSI_INTERPOLATE_CONSTANT:
		return 0;

	case TGSI_INTERPOLATE_LINEAR:
		if (location == TGSI_INTERPOLATE_LOC_SAMPLE)
			return SI_PARAM_LINEAR_SAMPLE;
		else if (location == TGSI_INTERPOLATE_LOC_CENTROID)
			return SI_PARAM_LINEAR_CENTROID;
		else
			return SI_PARAM_LINEAR_CENTER;
		break;
	case TGSI_INTERPOLATE_COLOR:
	case TGSI_INTERPOLATE_PERSPECTIVE:
		if (location == TGSI_INTERPOLATE_LOC_SAMPLE)
			return SI_PARAM_PERSP_SAMPLE;
		else if (location == TGSI_INTERPOLATE_LOC_CENTROID)
			return SI_PARAM_PERSP_CENTROID;
		else
			return SI_PARAM_PERSP_CENTER;
		break;
	default:
		fprintf(stderr, "Warning: Unhandled interpolation mode.\n");
		return -1;
	}
}

static LLVMValueRef si_build_fs_interp(struct si_shader_context *ctx,
				       unsigned attr_index, unsigned chan,
				       LLVMValueRef prim_mask,
				       LLVMValueRef i, LLVMValueRef j)
{
	if (i || j) {
		return ac_build_fs_interp(&ctx->ac,
					  LLVMConstInt(ctx->i32, chan, 0),
					  LLVMConstInt(ctx->i32, attr_index, 0),
					  prim_mask, i, j);
	}
	return ac_build_fs_interp_mov(&ctx->ac,
				      LLVMConstInt(ctx->i32, 2, 0), /* P0 */
				      LLVMConstInt(ctx->i32, chan, 0),
				      LLVMConstInt(ctx->i32, attr_index, 0),
				      prim_mask);
}

/**
 * Interpolate a fragment shader input.
 *
 * @param ctx		context
 * @param input_index		index of the input in hardware
 * @param semantic_name		TGSI_SEMANTIC_*
 * @param semantic_index	semantic index
 * @param num_interp_inputs	number of all interpolated inputs (= BCOLOR offset)
 * @param colors_read_mask	color components read (4 bits for each color, 8 bits in total)
 * @param interp_param		interpolation weights (i,j)
 * @param prim_mask		SI_PARAM_PRIM_MASK
 * @param face			SI_PARAM_FRONT_FACE
 * @param result		the return value (4 components)
 */
static void interp_fs_input(struct si_shader_context *ctx,
			    unsigned input_index,
			    unsigned semantic_name,
			    unsigned semantic_index,
			    unsigned num_interp_inputs,
			    unsigned colors_read_mask,
			    LLVMValueRef interp_param,
			    LLVMValueRef prim_mask,
			    LLVMValueRef face,
			    LLVMValueRef result[4])
{
	LLVMValueRef i = NULL, j = NULL;
	unsigned chan;

	/* fs.constant returns the param from the middle vertex, so it's not
	 * really useful for flat shading. It's meant to be used for custom
	 * interpolation (but the intrinsic can't fetch from the other two
	 * vertices).
	 *
	 * Luckily, it doesn't matter, because we rely on the FLAT_SHADE state
	 * to do the right thing. The only reason we use fs.constant is that
	 * fs.interp cannot be used on integers, because they can be equal
	 * to NaN.
	 *
	 * When interp is false we will use fs.constant or for newer llvm,
         * amdgcn.interp.mov.
	 */
	bool interp = interp_param != NULL;

	if (interp) {
		interp_param = LLVMBuildBitCast(ctx->ac.builder, interp_param,
						LLVMVectorType(ctx->f32, 2), "");

		i = LLVMBuildExtractElement(ctx->ac.builder, interp_param,
						ctx->i32_0, "");
		j = LLVMBuildExtractElement(ctx->ac.builder, interp_param,
						ctx->i32_1, "");
	}

	if (semantic_name == TGSI_SEMANTIC_COLOR &&
	    ctx->shader->key.part.ps.prolog.color_two_side) {
		LLVMValueRef is_face_positive;

		/* If BCOLOR0 is used, BCOLOR1 is at offset "num_inputs + 1",
		 * otherwise it's at offset "num_inputs".
		 */
		unsigned back_attr_offset = num_interp_inputs;
		if (semantic_index == 1 && colors_read_mask & 0xf)
			back_attr_offset += 1;

		is_face_positive = LLVMBuildICmp(ctx->ac.builder, LLVMIntNE,
						 face, ctx->i32_0, "");

		for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
			LLVMValueRef front, back;

			front = si_build_fs_interp(ctx,
						   input_index, chan,
						   prim_mask, i, j);
			back = si_build_fs_interp(ctx,
						  back_attr_offset, chan,
						  prim_mask, i, j);

			result[chan] = LLVMBuildSelect(ctx->ac.builder,
						is_face_positive,
						front,
						back,
						"");
		}
	} else if (semantic_name == TGSI_SEMANTIC_FOG) {
		result[0] = si_build_fs_interp(ctx, input_index,
					       0, prim_mask, i, j);
		result[1] =
		result[2] = LLVMConstReal(ctx->f32, 0.0f);
		result[3] = LLVMConstReal(ctx->f32, 1.0f);
	} else {
		for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
			result[chan] = si_build_fs_interp(ctx,
							  input_index, chan,
							  prim_mask, i, j);
		}
	}
}

void si_llvm_load_input_fs(
	struct si_shader_context *ctx,
	unsigned input_index,
	LLVMValueRef out[4])
{
	struct si_shader *shader = ctx->shader;
	struct tgsi_shader_info *info = &shader->selector->info;
	LLVMValueRef main_fn = ctx->main_fn;
	LLVMValueRef interp_param = NULL;
	int interp_param_idx;
	enum tgsi_semantic semantic_name = info->input_semantic_name[input_index];
	unsigned semantic_index = info->input_semantic_index[input_index];
	enum tgsi_interpolate_mode interp_mode = info->input_interpolate[input_index];
	enum tgsi_interpolate_loc interp_loc = info->input_interpolate_loc[input_index];

	/* Get colors from input VGPRs (set by the prolog). */
	if (semantic_name == TGSI_SEMANTIC_COLOR) {
		unsigned colors_read = shader->selector->info.colors_read;
		unsigned mask = colors_read >> (semantic_index * 4);
		unsigned offset = SI_PARAM_POS_FIXED_PT + 1 +
				  (semantic_index ? util_bitcount(colors_read & 0xf) : 0);
		LLVMValueRef undef = LLVMGetUndef(ctx->f32);

		out[0] = mask & 0x1 ? LLVMGetParam(main_fn, offset++) : undef;
		out[1] = mask & 0x2 ? LLVMGetParam(main_fn, offset++) : undef;
		out[2] = mask & 0x4 ? LLVMGetParam(main_fn, offset++) : undef;
		out[3] = mask & 0x8 ? LLVMGetParam(main_fn, offset++) : undef;
		return;
	}

	interp_param_idx = lookup_interp_param_index(interp_mode, interp_loc);
	if (interp_param_idx == -1)
		return;
	else if (interp_param_idx) {
		interp_param = LLVMGetParam(ctx->main_fn, interp_param_idx);
	}

	interp_fs_input(ctx, input_index, semantic_name,
			semantic_index, 0, /* this param is unused */
			shader->selector->info.colors_read, interp_param,
			ctx->abi.prim_mask,
			LLVMGetParam(main_fn, SI_PARAM_FRONT_FACE),
			&out[0]);
}

static void declare_input_fs(
	struct si_shader_context *ctx,
	unsigned input_index,
	const struct tgsi_full_declaration *decl,
	LLVMValueRef out[4])
{
	si_llvm_load_input_fs(ctx, input_index, out);
}

LLVMValueRef si_get_sample_id(struct si_shader_context *ctx)
{
	return si_unpack_param(ctx, SI_PARAM_ANCILLARY, 8, 4);
}

static LLVMValueRef get_base_vertex(struct ac_shader_abi *abi)
{
	struct si_shader_context *ctx = si_shader_context_from_abi(abi);

	/* For non-indexed draws, the base vertex set by the driver
	 * (for direct draws) or the CP (for indirect draws) is the
	 * first vertex ID, but GLSL expects 0 to be returned.
	 */
	LLVMValueRef vs_state = LLVMGetParam(ctx->main_fn,
					     ctx->param_vs_state_bits);
	LLVMValueRef indexed;

	indexed = LLVMBuildLShr(ctx->ac.builder, vs_state, ctx->i32_1, "");
	indexed = LLVMBuildTrunc(ctx->ac.builder, indexed, ctx->i1, "");

	return LLVMBuildSelect(ctx->ac.builder, indexed, ctx->abi.base_vertex,
			       ctx->i32_0, "");
}

static LLVMValueRef get_block_size(struct ac_shader_abi *abi)
{
	struct si_shader_context *ctx = si_shader_context_from_abi(abi);

	LLVMValueRef values[3];
	LLVMValueRef result;
	unsigned i;
	unsigned *properties = ctx->shader->selector->info.properties;

	if (properties[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH] != 0) {
		unsigned sizes[3] = {
			properties[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH],
			properties[TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT],
			properties[TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH]
		};

		for (i = 0; i < 3; ++i)
			values[i] = LLVMConstInt(ctx->i32, sizes[i], 0);

		result = ac_build_gather_values(&ctx->ac, values, 3);
	} else {
		result = LLVMGetParam(ctx->main_fn, ctx->param_block_size);
	}

	return result;
}

/**
 * Load a dword from a constant buffer.
 */
static LLVMValueRef buffer_load_const(struct si_shader_context *ctx,
				      LLVMValueRef resource,
				      LLVMValueRef offset)
{
	return ac_build_buffer_load(&ctx->ac, resource, 1, NULL, offset, NULL,
				    0, 0, 0, true, true);
}

static LLVMValueRef load_sample_position(struct ac_shader_abi *abi, LLVMValueRef sample_id)
{
	struct si_shader_context *ctx = si_shader_context_from_abi(abi);
	LLVMValueRef desc = LLVMGetParam(ctx->main_fn, ctx->param_rw_buffers);
	LLVMValueRef buf_index = LLVMConstInt(ctx->i32, SI_PS_CONST_SAMPLE_POSITIONS, 0);
	LLVMValueRef resource = ac_build_load_to_sgpr(&ctx->ac, desc, buf_index);

	/* offset = sample_id * 8  (8 = 2 floats containing samplepos.xy) */
	LLVMValueRef offset0 = LLVMBuildMul(ctx->ac.builder, sample_id, LLVMConstInt(ctx->i32, 8, 0), "");
	LLVMValueRef offset1 = LLVMBuildAdd(ctx->ac.builder, offset0, LLVMConstInt(ctx->i32, 4, 0), "");

	LLVMValueRef pos[4] = {
		buffer_load_const(ctx, resource, offset0),
		buffer_load_const(ctx, resource, offset1),
		LLVMConstReal(ctx->f32, 0),
		LLVMConstReal(ctx->f32, 0)
	};

	return ac_build_gather_values(&ctx->ac, pos, 4);
}

static LLVMValueRef load_sample_mask_in(struct ac_shader_abi *abi)
{
	struct si_shader_context *ctx = si_shader_context_from_abi(abi);
	return ac_to_integer(&ctx->ac, abi->sample_coverage);
}

static LLVMValueRef si_load_tess_coord(struct ac_shader_abi *abi)
{
	struct si_shader_context *ctx = si_shader_context_from_abi(abi);
	LLVMValueRef coord[4] = {
		LLVMGetParam(ctx->main_fn, ctx->param_tes_u),
		LLVMGetParam(ctx->main_fn, ctx->param_tes_v),
		ctx->ac.f32_0,
		ctx->ac.f32_0
	};

	/* For triangles, the vector should be (u, v, 1-u-v). */
	if (ctx->shader->selector->info.properties[TGSI_PROPERTY_TES_PRIM_MODE] ==
	    PIPE_PRIM_TRIANGLES) {
		coord[2] = LLVMBuildFSub(ctx->ac.builder, ctx->ac.f32_1,
					 LLVMBuildFAdd(ctx->ac.builder,
						       coord[0], coord[1], ""), "");
	}
	return ac_build_gather_values(&ctx->ac, coord, 4);
}

static LLVMValueRef load_tess_level(struct si_shader_context *ctx,
				    unsigned semantic_name)
{
	LLVMValueRef base, addr;

	int param = si_shader_io_get_unique_index_patch(semantic_name, 0);

	base = LLVMGetParam(ctx->main_fn, ctx->param_tcs_offchip_offset);
	addr = get_tcs_tes_buffer_address(ctx, get_rel_patch_id(ctx), NULL,
					  LLVMConstInt(ctx->i32, param, 0));

	return buffer_load(&ctx->bld_base, ctx->f32,
			   ~0, ctx->tess_offchip_ring, base, addr, true);

}

static LLVMValueRef si_load_tess_level(struct ac_shader_abi *abi,
				       unsigned varying_id)
{
	struct si_shader_context *ctx = si_shader_context_from_abi(abi);
	unsigned semantic_name;

	switch (varying_id) {
	case VARYING_SLOT_TESS_LEVEL_INNER:
		semantic_name = TGSI_SEMANTIC_TESSINNER;
		break;
	case VARYING_SLOT_TESS_LEVEL_OUTER:
		semantic_name = TGSI_SEMANTIC_TESSOUTER;
		break;
	default:
		unreachable("unknown tess level");
	}

	return load_tess_level(ctx, semantic_name);

}

static LLVMValueRef si_load_patch_vertices_in(struct ac_shader_abi *abi)
{
	struct si_shader_context *ctx = si_shader_context_from_abi(abi);
	if (ctx->type == PIPE_SHADER_TESS_CTRL)
		return si_unpack_param(ctx, ctx->param_tcs_out_lds_layout, 13, 6);
	else if (ctx->type == PIPE_SHADER_TESS_EVAL)
		return get_num_tcs_out_vertices(ctx);
	else
		unreachable("invalid shader stage for TGSI_SEMANTIC_VERTICESIN");
}

void si_load_system_value(struct si_shader_context *ctx,
			  unsigned index,
			  const struct tgsi_full_declaration *decl)
{
	LLVMValueRef value = 0;

	assert(index < RADEON_LLVM_MAX_SYSTEM_VALUES);

	switch (decl->Semantic.Name) {
	case TGSI_SEMANTIC_INSTANCEID:
		value = ctx->abi.instance_id;
		break;

	case TGSI_SEMANTIC_VERTEXID:
		value = LLVMBuildAdd(ctx->ac.builder,
				     ctx->abi.vertex_id,
				     ctx->abi.base_vertex, "");
		break;

	case TGSI_SEMANTIC_VERTEXID_NOBASE:
		/* Unused. Clarify the meaning in indexed vs. non-indexed
		 * draws if this is ever used again. */
		assert(false);
		break;

	case TGSI_SEMANTIC_BASEVERTEX:
		value = get_base_vertex(&ctx->abi);
		break;

	case TGSI_SEMANTIC_BASEINSTANCE:
		value = ctx->abi.start_instance;
		break;

	case TGSI_SEMANTIC_DRAWID:
		value = ctx->abi.draw_id;
		break;

	case TGSI_SEMANTIC_INVOCATIONID:
		if (ctx->type == PIPE_SHADER_TESS_CTRL)
			value = unpack_llvm_param(ctx, ctx->abi.tcs_rel_ids, 8, 5);
		else if (ctx->type == PIPE_SHADER_GEOMETRY)
			value = ctx->abi.gs_invocation_id;
		else
			assert(!"INVOCATIONID not implemented");
		break;

	case TGSI_SEMANTIC_POSITION:
	{
		LLVMValueRef pos[4] = {
			LLVMGetParam(ctx->main_fn, SI_PARAM_POS_X_FLOAT),
			LLVMGetParam(ctx->main_fn, SI_PARAM_POS_Y_FLOAT),
			LLVMGetParam(ctx->main_fn, SI_PARAM_POS_Z_FLOAT),
			ac_build_fdiv(&ctx->ac, ctx->ac.f32_1,
				      LLVMGetParam(ctx->main_fn, SI_PARAM_POS_W_FLOAT)),
		};
		value = ac_build_gather_values(&ctx->ac, pos, 4);
		break;
	}

	case TGSI_SEMANTIC_FACE:
		value = ctx->abi.front_face;
		break;

	case TGSI_SEMANTIC_SAMPLEID:
		value = si_get_sample_id(ctx);
		break;

	case TGSI_SEMANTIC_SAMPLEPOS: {
		LLVMValueRef pos[4] = {
			LLVMGetParam(ctx->main_fn, SI_PARAM_POS_X_FLOAT),
			LLVMGetParam(ctx->main_fn, SI_PARAM_POS_Y_FLOAT),
			LLVMConstReal(ctx->f32, 0),
			LLVMConstReal(ctx->f32, 0)
		};
		pos[0] = ac_build_fract(&ctx->ac, pos[0], 32);
		pos[1] = ac_build_fract(&ctx->ac, pos[1], 32);
		value = ac_build_gather_values(&ctx->ac, pos, 4);
		break;
	}

	case TGSI_SEMANTIC_SAMPLEMASK:
		/* This can only occur with the OpenGL Core profile, which
		 * doesn't support smoothing.
		 */
		value = LLVMGetParam(ctx->main_fn, SI_PARAM_SAMPLE_COVERAGE);
		break;

	case TGSI_SEMANTIC_TESSCOORD:
		value = si_load_tess_coord(&ctx->abi);
		break;

	case TGSI_SEMANTIC_VERTICESIN:
		value = si_load_patch_vertices_in(&ctx->abi);
		break;

	case TGSI_SEMANTIC_TESSINNER:
	case TGSI_SEMANTIC_TESSOUTER:
		value = load_tess_level(ctx, decl->Semantic.Name);
		break;

	case TGSI_SEMANTIC_DEFAULT_TESSOUTER_SI:
	case TGSI_SEMANTIC_DEFAULT_TESSINNER_SI:
	{
		LLVMValueRef buf, slot, val[4];
		int i, offset;

		slot = LLVMConstInt(ctx->i32, SI_HS_CONST_DEFAULT_TESS_LEVELS, 0);
		buf = LLVMGetParam(ctx->main_fn, ctx->param_rw_buffers);
		buf = ac_build_load_to_sgpr(&ctx->ac, buf, slot);
		offset = decl->Semantic.Name == TGSI_SEMANTIC_DEFAULT_TESSINNER_SI ? 4 : 0;

		for (i = 0; i < 4; i++)
			val[i] = buffer_load_const(ctx, buf,
						   LLVMConstInt(ctx->i32, (offset + i) * 4, 0));
		value = ac_build_gather_values(&ctx->ac, val, 4);
		break;
	}

	case TGSI_SEMANTIC_PRIMID:
		value = get_primitive_id(ctx, 0);
		break;

	case TGSI_SEMANTIC_GRID_SIZE:
		value = ctx->abi.num_work_groups;
		break;

	case TGSI_SEMANTIC_BLOCK_SIZE:
		value = get_block_size(&ctx->abi);
		break;

	case TGSI_SEMANTIC_BLOCK_ID:
	{
		LLVMValueRef values[3];

		for (int i = 0; i < 3; i++) {
			values[i] = ctx->i32_0;
			if (ctx->abi.workgroup_ids[i]) {
				values[i] = ctx->abi.workgroup_ids[i];
			}
		}
		value = ac_build_gather_values(&ctx->ac, values, 3);
		break;
	}

	case TGSI_SEMANTIC_THREAD_ID:
		value = ctx->abi.local_invocation_ids;
		break;

	case TGSI_SEMANTIC_HELPER_INVOCATION:
		value = ac_build_load_helper_invocation(&ctx->ac);
		break;

	case TGSI_SEMANTIC_SUBGROUP_SIZE:
		value = LLVMConstInt(ctx->i32, 64, 0);
		break;

	case TGSI_SEMANTIC_SUBGROUP_INVOCATION:
		value = ac_get_thread_id(&ctx->ac);
		break;

	case TGSI_SEMANTIC_SUBGROUP_EQ_MASK:
	{
		LLVMValueRef id = ac_get_thread_id(&ctx->ac);
		id = LLVMBuildZExt(ctx->ac.builder, id, ctx->i64, "");
		value = LLVMBuildShl(ctx->ac.builder, LLVMConstInt(ctx->i64, 1, 0), id, "");
		value = LLVMBuildBitCast(ctx->ac.builder, value, ctx->v2i32, "");
		break;
	}

	case TGSI_SEMANTIC_SUBGROUP_GE_MASK:
	case TGSI_SEMANTIC_SUBGROUP_GT_MASK:
	case TGSI_SEMANTIC_SUBGROUP_LE_MASK:
	case TGSI_SEMANTIC_SUBGROUP_LT_MASK:
	{
		LLVMValueRef id = ac_get_thread_id(&ctx->ac);
		if (decl->Semantic.Name == TGSI_SEMANTIC_SUBGROUP_GT_MASK ||
		    decl->Semantic.Name == TGSI_SEMANTIC_SUBGROUP_LE_MASK) {
			/* All bits set except LSB */
			value = LLVMConstInt(ctx->i64, -2, 0);
		} else {
			/* All bits set */
			value = LLVMConstInt(ctx->i64, -1, 0);
		}
		id = LLVMBuildZExt(ctx->ac.builder, id, ctx->i64, "");
		value = LLVMBuildShl(ctx->ac.builder, value, id, "");
		if (decl->Semantic.Name == TGSI_SEMANTIC_SUBGROUP_LE_MASK ||
		    decl->Semantic.Name == TGSI_SEMANTIC_SUBGROUP_LT_MASK)
			value = LLVMBuildNot(ctx->ac.builder, value, "");
		value = LLVMBuildBitCast(ctx->ac.builder, value, ctx->v2i32, "");
		break;
	}

	case TGSI_SEMANTIC_CS_USER_DATA:
		value = LLVMGetParam(ctx->main_fn, ctx->param_cs_user_data);
		break;

	default:
		assert(!"unknown system value");
		return;
	}

	ctx->system_values[index] = value;
}

void si_declare_compute_memory(struct si_shader_context *ctx)
{
	struct si_shader_selector *sel = ctx->shader->selector;
	unsigned lds_size = sel->info.properties[TGSI_PROPERTY_CS_LOCAL_SIZE];

	LLVMTypeRef i8p = LLVMPointerType(ctx->i8, AC_ADDR_SPACE_LDS);
	LLVMValueRef var;

	assert(!ctx->ac.lds);

	var = LLVMAddGlobalInAddressSpace(ctx->ac.module,
	                                  LLVMArrayType(ctx->i8, lds_size),
	                                  "compute_lds",
	                                  AC_ADDR_SPACE_LDS);
	LLVMSetAlignment(var, 4);

	ctx->ac.lds = LLVMBuildBitCast(ctx->ac.builder, var, i8p, "");
}

void si_tgsi_declare_compute_memory(struct si_shader_context *ctx,
				    const struct tgsi_full_declaration *decl)
{
	assert(decl->Declaration.MemType == TGSI_MEMORY_TYPE_SHARED);
	assert(decl->Range.First == decl->Range.Last);

	si_declare_compute_memory(ctx);
}

static LLVMValueRef load_const_buffer_desc_fast_path(struct si_shader_context *ctx)
{
	LLVMValueRef ptr =
		LLVMGetParam(ctx->main_fn, ctx->param_const_and_shader_buffers);
	struct si_shader_selector *sel = ctx->shader->selector;

	/* Do the bounds checking with a descriptor, because
	 * doing computation and manual bounds checking of 64-bit
	 * addresses generates horrible VALU code with very high
	 * VGPR usage and very low SIMD occupancy.
	 */
	ptr = LLVMBuildPtrToInt(ctx->ac.builder, ptr, ctx->ac.intptr, "");

	LLVMValueRef desc0, desc1;
	desc0 = ptr;
	desc1 = LLVMConstInt(ctx->i32,
			     S_008F04_BASE_ADDRESS_HI(ctx->screen->info.address32_hi), 0);

	LLVMValueRef desc_elems[] = {
		desc0,
		desc1,
		LLVMConstInt(ctx->i32, (sel->info.const_file_max[0] + 1) * 16, 0),
		LLVMConstInt(ctx->i32,
			S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) |
			S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
			S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) |
			S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W) |
			S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) |
			S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32), 0)
	};

	return ac_build_gather_values(&ctx->ac, desc_elems, 4);
}

static LLVMValueRef load_const_buffer_desc(struct si_shader_context *ctx, int i)
{
	LLVMValueRef list_ptr = LLVMGetParam(ctx->main_fn,
					     ctx->param_const_and_shader_buffers);

	return ac_build_load_to_sgpr(&ctx->ac, list_ptr,
				     LLVMConstInt(ctx->i32, si_get_constbuf_slot(i), 0));
}

static LLVMValueRef load_ubo(struct ac_shader_abi *abi, LLVMValueRef index)
{
	struct si_shader_context *ctx = si_shader_context_from_abi(abi);
	struct si_shader_selector *sel = ctx->shader->selector;

	LLVMValueRef ptr = LLVMGetParam(ctx->main_fn, ctx->param_const_and_shader_buffers);

	if (sel->info.const_buffers_declared == 1 &&
	    sel->info.shader_buffers_declared == 0) {
		return load_const_buffer_desc_fast_path(ctx);
	}

	index = si_llvm_bound_index(ctx, index, ctx->num_const_buffers);
	index = LLVMBuildAdd(ctx->ac.builder, index,
			     LLVMConstInt(ctx->i32, SI_NUM_SHADER_BUFFERS, 0), "");

	return ac_build_load_to_sgpr(&ctx->ac, ptr, index);
}

static LLVMValueRef
load_ssbo(struct ac_shader_abi *abi, LLVMValueRef index, bool write)
{
	struct si_shader_context *ctx = si_shader_context_from_abi(abi);
	LLVMValueRef rsrc_ptr = LLVMGetParam(ctx->main_fn,
					     ctx->param_const_and_shader_buffers);

	index = si_llvm_bound_index(ctx, index, ctx->num_shader_buffers);
	index = LLVMBuildSub(ctx->ac.builder,
			     LLVMConstInt(ctx->i32, SI_NUM_SHADER_BUFFERS - 1, 0),
			     index, "");

	return ac_build_load_to_sgpr(&ctx->ac, rsrc_ptr, index);
}

static LLVMValueRef fetch_constant(
	struct lp_build_tgsi_context *bld_base,
	const struct tgsi_full_src_register *reg,
	enum tgsi_opcode_type type,
	unsigned swizzle_in)
{
	struct si_shader_context *ctx = si_shader_context(bld_base);
	struct si_shader_selector *sel = ctx->shader->selector;
	const struct tgsi_ind_register *ireg = &reg->Indirect;
	unsigned buf, idx;
	unsigned swizzle = swizzle_in & 0xffff;

	LLVMValueRef addr, bufp;

	if (swizzle_in == LP_CHAN_ALL) {
		unsigned chan;
		LLVMValueRef values[4];
		for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan)
			values[chan] = fetch_constant(bld_base, reg, type, chan);

		return ac_build_gather_values(&ctx->ac, values, 4);
	}

	/* Split 64-bit loads. */
	if (tgsi_type_is_64bit(type)) {
		LLVMValueRef lo, hi;

		lo = fetch_constant(bld_base, reg, TGSI_TYPE_UNSIGNED, swizzle);
		hi = fetch_constant(bld_base, reg, TGSI_TYPE_UNSIGNED, (swizzle_in >> 16));
		return si_llvm_emit_fetch_64bit(bld_base, tgsi2llvmtype(bld_base, type),
						lo, hi);
	}

	idx = reg->Register.Index * 4 + swizzle;
	if (reg->Register.Indirect) {
		addr = si_get_indirect_index(ctx, ireg, 16, idx * 4);
	} else {
		addr = LLVMConstInt(ctx->i32, idx * 4, 0);
	}

	/* Fast path when user data SGPRs point to constant buffer 0 directly. */
	if (sel->info.const_buffers_declared == 1 &&
	    sel->info.shader_buffers_declared == 0) {
		LLVMValueRef desc = load_const_buffer_desc_fast_path(ctx);
		LLVMValueRef result = buffer_load_const(ctx, desc, addr);
		return bitcast(bld_base, type, result);
	}

	assert(reg->Register.Dimension);
	buf = reg->Dimension.Index;

	if (reg->Dimension.Indirect) {
		LLVMValueRef ptr = LLVMGetParam(ctx->main_fn, ctx->param_const_and_shader_buffers);
		LLVMValueRef index;
		index = si_get_bounded_indirect_index(ctx, &reg->DimIndirect,
						      reg->Dimension.Index,
						      ctx->num_const_buffers);
		index = LLVMBuildAdd(ctx->ac.builder, index,
				     LLVMConstInt(ctx->i32, SI_NUM_SHADER_BUFFERS, 0), "");
		bufp = ac_build_load_to_sgpr(&ctx->ac, ptr, index);
	} else
		bufp = load_const_buffer_desc(ctx, buf);

	return bitcast(bld_base, type, buffer_load_const(ctx, bufp, addr));
}

/* Initialize arguments for the shader export intrinsic */
static void si_llvm_init_export_args(struct si_shader_context *ctx,
				     LLVMValueRef *values,
				     unsigned target,
				     struct ac_export_args *args)
{
	LLVMValueRef f32undef = LLVMGetUndef(ctx->ac.f32);
	unsigned spi_shader_col_format = V_028714_SPI_SHADER_32_ABGR;
	unsigned chan;
	bool is_int8, is_int10;

	/* Default is 0xf. Adjusted below depending on the format. */
	args->enabled_channels = 0xf; /* writemask */

	/* Specify whether the EXEC mask represents the valid mask */
	args->valid_mask = 0;

	/* Specify whether this is the last export */
	args->done = 0;

	/* Specify the target we are exporting */
	args->target = target;

	if (ctx->type == PIPE_SHADER_FRAGMENT) {
		const struct si_shader_key *key = &ctx->shader->key;
		unsigned col_formats = key->part.ps.epilog.spi_shader_col_format;
		int cbuf = target - V_008DFC_SQ_EXP_MRT;

		assert(cbuf >= 0 && cbuf < 8);
		spi_shader_col_format = (col_formats >> (cbuf * 4)) & 0xf;
		is_int8 = (key->part.ps.epilog.color_is_int8 >> cbuf) & 0x1;
		is_int10 = (key->part.ps.epilog.color_is_int10 >> cbuf) & 0x1;
	}

	args->compr = false;
	args->out[0] = f32undef;
	args->out[1] = f32undef;
	args->out[2] = f32undef;
	args->out[3] = f32undef;

	LLVMValueRef (*packf)(struct ac_llvm_context *ctx, LLVMValueRef args[2]) = NULL;
	LLVMValueRef (*packi)(struct ac_llvm_context *ctx, LLVMValueRef args[2],
			      unsigned bits, bool hi) = NULL;

	switch (spi_shader_col_format) {
	case V_028714_SPI_SHADER_ZERO:
		args->enabled_channels = 0; /* writemask */
		args->target = V_008DFC_SQ_EXP_NULL;
		break;

	case V_028714_SPI_SHADER_32_R:
		args->enabled_channels = 1; /* writemask */
		args->out[0] = values[0];
		break;

	case V_028714_SPI_SHADER_32_GR:
		args->enabled_channels = 0x3; /* writemask */
		args->out[0] = values[0];
		args->out[1] = values[1];
		break;

	case V_028714_SPI_SHADER_32_AR:
		args->enabled_channels = 0x9; /* writemask */
		args->out[0] = values[0];
		args->out[3] = values[3];
		break;

	case V_028714_SPI_SHADER_FP16_ABGR:
		packf = ac_build_cvt_pkrtz_f16;
		break;

	case V_028714_SPI_SHADER_UNORM16_ABGR:
		packf = ac_build_cvt_pknorm_u16;
		break;

	case V_028714_SPI_SHADER_SNORM16_ABGR:
		packf = ac_build_cvt_pknorm_i16;
		break;

	case V_028714_SPI_SHADER_UINT16_ABGR:
		packi = ac_build_cvt_pk_u16;
		break;

	case V_028714_SPI_SHADER_SINT16_ABGR:
		packi = ac_build_cvt_pk_i16;
		break;

	case V_028714_SPI_SHADER_32_ABGR:
		memcpy(&args->out[0], values, sizeof(values[0]) * 4);
		break;
	}

	/* Pack f16 or norm_i16/u16. */
	if (packf) {
		for (chan = 0; chan < 2; chan++) {
			LLVMValueRef pack_args[2] = {
				values[2 * chan],
				values[2 * chan + 1]
			};
			LLVMValueRef packed;

			packed = packf(&ctx->ac, pack_args);
			args->out[chan] = ac_to_float(&ctx->ac, packed);
		}
		args->compr = 1; /* COMPR flag */
	}
	/* Pack i16/u16. */
	if (packi) {
		for (chan = 0; chan < 2; chan++) {
			LLVMValueRef pack_args[2] = {
				ac_to_integer(&ctx->ac, values[2 * chan]),
				ac_to_integer(&ctx->ac, values[2 * chan + 1])
			};
			LLVMValueRef packed;

			packed = packi(&ctx->ac, pack_args,
				       is_int8 ? 8 : is_int10 ? 10 : 16,
				       chan == 1);
			args->out[chan] = ac_to_float(&ctx->ac, packed);
		}
		args->compr = 1; /* COMPR flag */
	}
}

static void si_alpha_test(struct lp_build_tgsi_context *bld_base,
			  LLVMValueRef alpha)
{
	struct si_shader_context *ctx = si_shader_context(bld_base);

	if (ctx->shader->key.part.ps.epilog.alpha_func != PIPE_FUNC_NEVER) {
		static LLVMRealPredicate cond_map[PIPE_FUNC_ALWAYS + 1] = {
			[PIPE_FUNC_LESS] = LLVMRealOLT,
			[PIPE_FUNC_EQUAL] = LLVMRealOEQ,
			[PIPE_FUNC_LEQUAL] = LLVMRealOLE,
			[PIPE_FUNC_GREATER] = LLVMRealOGT,
			[PIPE_FUNC_NOTEQUAL] = LLVMRealONE,
			[PIPE_FUNC_GEQUAL] = LLVMRealOGE,
		};
		LLVMRealPredicate cond = cond_map[ctx->shader->key.part.ps.epilog.alpha_func];
		assert(cond);

		LLVMValueRef alpha_ref = LLVMGetParam(ctx->main_fn,
				SI_PARAM_ALPHA_REF);
		LLVMValueRef alpha_pass =
			LLVMBuildFCmp(ctx->ac.builder, cond, alpha, alpha_ref, "");
		ac_build_kill_if_false(&ctx->ac, alpha_pass);
	} else {
		ac_build_kill_if_false(&ctx->ac, ctx->i1false);
	}
}

static LLVMValueRef si_scale_alpha_by_sample_mask(struct lp_build_tgsi_context *bld_base,
						  LLVMValueRef alpha,
						  unsigned samplemask_param)
{
	struct si_shader_context *ctx = si_shader_context(bld_base);
	LLVMValueRef coverage;

	/* alpha = alpha * popcount(coverage) / SI_NUM_SMOOTH_AA_SAMPLES */
	coverage = LLVMGetParam(ctx->main_fn,
				samplemask_param);
	coverage = ac_to_integer(&ctx->ac, coverage);

	coverage = ac_build_intrinsic(&ctx->ac, "llvm.ctpop.i32",
				   ctx->i32,
				   &coverage, 1, AC_FUNC_ATTR_READNONE);

	coverage = LLVMBuildUIToFP(ctx->ac.builder, coverage,
				   ctx->f32, "");

	coverage = LLVMBuildFMul(ctx->ac.builder, coverage,
				 LLVMConstReal(ctx->f32,
					1.0 / SI_NUM_SMOOTH_AA_SAMPLES), "");

	return LLVMBuildFMul(ctx->ac.builder, alpha, coverage, "");
}

static void si_llvm_emit_clipvertex(struct si_shader_context *ctx,
				    struct ac_export_args *pos, LLVMValueRef *out_elts)
{
	unsigned reg_index;
	unsigned chan;
	unsigned const_chan;
	LLVMValueRef base_elt;
	LLVMValueRef ptr = LLVMGetParam(ctx->main_fn, ctx->param_rw_buffers);
	LLVMValueRef constbuf_index = LLVMConstInt(ctx->i32,
						   SI_VS_CONST_CLIP_PLANES, 0);
	LLVMValueRef const_resource = ac_build_load_to_sgpr(&ctx->ac, ptr, constbuf_index);

	for (reg_index = 0; reg_index < 2; reg_index ++) {
		struct ac_export_args *args = &pos[2 + reg_index];

		args->out[0] =
		args->out[1] =
		args->out[2] =
		args->out[3] = LLVMConstReal(ctx->f32, 0.0f);

		/* Compute dot products of position and user clip plane vectors */
		for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
			for (const_chan = 0; const_chan < TGSI_NUM_CHANNELS; const_chan++) {
				LLVMValueRef addr =
					LLVMConstInt(ctx->i32, ((reg_index * 4 + chan) * 4 +
								const_chan) * 4, 0);
				base_elt = buffer_load_const(ctx, const_resource,
							     addr);
				args->out[chan] = ac_build_fmad(&ctx->ac, base_elt,
								out_elts[const_chan], args->out[chan]);
			}
		}

		args->enabled_channels = 0xf;
		args->valid_mask = 0;
		args->done = 0;
		args->target = V_008DFC_SQ_EXP_POS + 2 + reg_index;
		args->compr = 0;
	}
}

static void si_dump_streamout(struct pipe_stream_output_info *so)
{
	unsigned i;

	if (so->num_outputs)
		fprintf(stderr, "STREAMOUT\n");

	for (i = 0; i < so->num_outputs; i++) {
		unsigned mask = ((1 << so->output[i].num_components) - 1) <<
				so->output[i].start_component;
		fprintf(stderr, "  %i: BUF%i[%i..%i] <- OUT[%i].%s%s%s%s\n",
			i, so->output[i].output_buffer,
			so->output[i].dst_offset, so->output[i].dst_offset + so->output[i].num_components - 1,
			so->output[i].register_index,
			mask & 1 ? "x" : "",
		        mask & 2 ? "y" : "",
		        mask & 4 ? "z" : "",
		        mask & 8 ? "w" : "");
	}
}

static void emit_streamout_output(struct si_shader_context *ctx,
				  LLVMValueRef const *so_buffers,
				  LLVMValueRef const *so_write_offsets,
				  struct pipe_stream_output *stream_out,
				  struct si_shader_output_values *shader_out)
{
	unsigned buf_idx = stream_out->output_buffer;
	unsigned start = stream_out->start_component;
	unsigned num_comps = stream_out->num_components;
	LLVMValueRef out[4];

	assert(num_comps && num_comps <= 4);
	if (!num_comps || num_comps > 4)
		return;

	/* Load the output as int. */
	for (int j = 0; j < num_comps; j++) {
		assert(stream_out->stream == shader_out->vertex_stream[start + j]);

		out[j] = ac_to_integer(&ctx->ac, shader_out->values[start + j]);
	}

	/* Pack the output. */
	LLVMValueRef vdata = NULL;

	switch (num_comps) {
	case 1: /* as i32 */
		vdata = out[0];
		break;
	case 2: /* as v2i32 */
	case 3: /* as v4i32 (aligned to 4) */
		out[3] = LLVMGetUndef(ctx->i32);
		/* fall through */
	case 4: /* as v4i32 */
		vdata = ac_build_gather_values(&ctx->ac, out, util_next_power_of_two(num_comps));
		break;
	}

	ac_build_buffer_store_dword(&ctx->ac, so_buffers[buf_idx],
				    vdata, num_comps,
				    so_write_offsets[buf_idx],
				    ctx->i32_0,
				    stream_out->dst_offset * 4, 1, 1, true, false);
}

/**
 * Write streamout data to buffers for vertex stream @p stream (different
 * vertex streams can occur for GS copy shaders).
 */
static void si_llvm_emit_streamout(struct si_shader_context *ctx,
				   struct si_shader_output_values *outputs,
				   unsigned noutput, unsigned stream)
{
	struct si_shader_selector *sel = ctx->shader->selector;
	struct pipe_stream_output_info *so = &sel->so;
	LLVMBuilderRef builder = ctx->ac.builder;
	int i;
	struct lp_build_if_state if_ctx;

	/* Get bits [22:16], i.e. (so_param >> 16) & 127; */
	LLVMValueRef so_vtx_count =
		si_unpack_param(ctx, ctx->param_streamout_config, 16, 7);

	LLVMValueRef tid = ac_get_thread_id(&ctx->ac);

	/* can_emit = tid < so_vtx_count; */
	LLVMValueRef can_emit =
		LLVMBuildICmp(builder, LLVMIntULT, tid, so_vtx_count, "");

	/* Emit the streamout code conditionally. This actually avoids
	 * out-of-bounds buffer access. The hw tells us via the SGPR
	 * (so_vtx_count) which threads are allowed to emit streamout data. */
	lp_build_if(&if_ctx, &ctx->gallivm, can_emit);
	{
		/* The buffer offset is computed as follows:
		 *   ByteOffset = streamout_offset[buffer_id]*4 +
		 *                (streamout_write_index + thread_id)*stride[buffer_id] +
		 *                attrib_offset
                 */

		LLVMValueRef so_write_index =
			LLVMGetParam(ctx->main_fn,
				     ctx->param_streamout_write_index);

		/* Compute (streamout_write_index + thread_id). */
		so_write_index = LLVMBuildAdd(builder, so_write_index, tid, "");

		/* Load the descriptor and compute the write offset for each
		 * enabled buffer. */
		LLVMValueRef so_write_offset[4] = {};
		LLVMValueRef so_buffers[4];
		LLVMValueRef buf_ptr = LLVMGetParam(ctx->main_fn,
						    ctx->param_rw_buffers);

		for (i = 0; i < 4; i++) {
			if (!so->stride[i])
				continue;

			LLVMValueRef offset = LLVMConstInt(ctx->i32,
							   SI_VS_STREAMOUT_BUF0 + i, 0);

			so_buffers[i] = ac_build_load_to_sgpr(&ctx->ac, buf_ptr, offset);

			LLVMValueRef so_offset = LLVMGetParam(ctx->main_fn,
							      ctx->param_streamout_offset[i]);
			so_offset = LLVMBuildMul(builder, so_offset, LLVMConstInt(ctx->i32, 4, 0), "");

			so_write_offset[i] = ac_build_imad(&ctx->ac, so_write_index,
							   LLVMConstInt(ctx->i32, so->stride[i]*4, 0),
							   so_offset);
		}

		/* Write streamout data. */
		for (i = 0; i < so->num_outputs; i++) {
			unsigned reg = so->output[i].register_index;

			if (reg >= noutput)
				continue;

			if (stream != so->output[i].stream)
				continue;

			emit_streamout_output(ctx, so_buffers, so_write_offset,
					      &so->output[i], &outputs[reg]);
		}
	}
	lp_build_endif(&if_ctx);
}

static void si_export_param(struct si_shader_context *ctx, unsigned index,
			    LLVMValueRef *values)
{
	struct ac_export_args args;

	si_llvm_init_export_args(ctx, values,
				 V_008DFC_SQ_EXP_PARAM + index, &args);
	ac_build_export(&ctx->ac, &args);
}

static void si_build_param_exports(struct si_shader_context *ctx,
				   struct si_shader_output_values *outputs,
			           unsigned noutput)
{
	struct si_shader *shader = ctx->shader;
	unsigned param_count = 0;

	for (unsigned i = 0; i < noutput; i++) {
		unsigned semantic_name = outputs[i].semantic_name;
		unsigned semantic_index = outputs[i].semantic_index;

		if (outputs[i].vertex_stream[0] != 0 &&
		    outputs[i].vertex_stream[1] != 0 &&
		    outputs[i].vertex_stream[2] != 0 &&
		    outputs[i].vertex_stream[3] != 0)
			continue;

		switch (semantic_name) {
		case TGSI_SEMANTIC_LAYER:
		case TGSI_SEMANTIC_VIEWPORT_INDEX:
		case TGSI_SEMANTIC_CLIPDIST:
		case TGSI_SEMANTIC_COLOR:
		case TGSI_SEMANTIC_BCOLOR:
		case TGSI_SEMANTIC_PRIMID:
		case TGSI_SEMANTIC_FOG:
		case TGSI_SEMANTIC_TEXCOORD:
		case TGSI_SEMANTIC_GENERIC:
			break;
		default:
			continue;
		}

		if ((semantic_name != TGSI_SEMANTIC_GENERIC ||
		     semantic_index < SI_MAX_IO_GENERIC) &&
		    shader->key.opt.kill_outputs &
		    (1ull << si_shader_io_get_unique_index(semantic_name,
							   semantic_index, true)))
			continue;

		si_export_param(ctx, param_count, outputs[i].values);

		assert(i < ARRAY_SIZE(shader->info.vs_output_param_offset));
		shader->info.vs_output_param_offset[i] = param_count++;
	}

	shader->info.nr_param_exports = param_count;
}

/* Generate export instructions for hardware VS shader stage */
static void si_llvm_export_vs(struct si_shader_context *ctx,
			      struct si_shader_output_values *outputs,
			      unsigned noutput)
{
	struct si_shader *shader = ctx->shader;
	struct ac_export_args pos_args[4] = {};
	LLVMValueRef psize_value = NULL, edgeflag_value = NULL, layer_value = NULL, viewport_index_value = NULL;
	unsigned pos_idx;
	int i;

	/* Build position exports. */
	for (i = 0; i < noutput; i++) {
		switch (outputs[i].semantic_name) {
		case TGSI_SEMANTIC_POSITION:
			si_llvm_init_export_args(ctx, outputs[i].values,
						 V_008DFC_SQ_EXP_POS, &pos_args[0]);
			break;
		case TGSI_SEMANTIC_PSIZE:
			psize_value = outputs[i].values[0];
			break;
		case TGSI_SEMANTIC_LAYER:
			layer_value = outputs[i].values[0];
			break;
		case TGSI_SEMANTIC_VIEWPORT_INDEX:
			viewport_index_value = outputs[i].values[0];
			break;
		case TGSI_SEMANTIC_EDGEFLAG:
			edgeflag_value = outputs[i].values[0];
			break;
		case TGSI_SEMANTIC_CLIPDIST:
			if (!shader->key.opt.clip_disable) {
				unsigned index = 2 + outputs[i].semantic_index;
				si_llvm_init_export_args(ctx, outputs[i].values,
							 V_008DFC_SQ_EXP_POS + index,
							 &pos_args[index]);
			}
			break;
		case TGSI_SEMANTIC_CLIPVERTEX:
			if (!shader->key.opt.clip_disable) {
				si_llvm_emit_clipvertex(ctx, pos_args,
							outputs[i].values);
			}
			break;
		}
	}

	/* We need to add the position output manually if it's missing. */
	if (!pos_args[0].out[0]) {
		pos_args[0].enabled_channels = 0xf; /* writemask */
		pos_args[0].valid_mask = 0; /* EXEC mask */
		pos_args[0].done = 0; /* last export? */
		pos_args[0].target = V_008DFC_SQ_EXP_POS;
		pos_args[0].compr = 0; /* COMPR flag */
		pos_args[0].out[0] = ctx->ac.f32_0; /* X */
		pos_args[0].out[1] = ctx->ac.f32_0; /* Y */
		pos_args[0].out[2] = ctx->ac.f32_0; /* Z */
		pos_args[0].out[3] = ctx->ac.f32_1;  /* W */
	}

	/* Write the misc vector (point size, edgeflag, layer, viewport). */
	if (shader->selector->info.writes_psize ||
	    shader->selector->info.writes_edgeflag ||
	    shader->selector->info.writes_viewport_index ||
	    shader->selector->info.writes_layer) {
		pos_args[1].enabled_channels = shader->selector->info.writes_psize |
					       (shader->selector->info.writes_edgeflag << 1) |
					       (shader->selector->info.writes_layer << 2);

		pos_args[1].valid_mask = 0; /* EXEC mask */
		pos_args[1].done = 0; /* last export? */
		pos_args[1].target = V_008DFC_SQ_EXP_POS + 1;
		pos_args[1].compr = 0; /* COMPR flag */
		pos_args[1].out[0] = ctx->ac.f32_0; /* X */
		pos_args[1].out[1] = ctx->ac.f32_0; /* Y */
		pos_args[1].out[2] = ctx->ac.f32_0; /* Z */
		pos_args[1].out[3] = ctx->ac.f32_0; /* W */

		if (shader->selector->info.writes_psize)
			pos_args[1].out[0] = psize_value;

		if (shader->selector->info.writes_edgeflag) {
			/* The output is a float, but the hw expects an integer
			 * with the first bit containing the edge flag. */
			edgeflag_value = LLVMBuildFPToUI(ctx->ac.builder,
							 edgeflag_value,
							 ctx->i32, "");
			edgeflag_value = ac_build_umin(&ctx->ac,
						      edgeflag_value,
						      ctx->i32_1);

			/* The LLVM intrinsic expects a float. */
			pos_args[1].out[1] = ac_to_float(&ctx->ac, edgeflag_value);
		}

		if (ctx->screen->info.chip_class >= GFX9) {
			/* GFX9 has the layer in out.z[10:0] and the viewport
			 * index in out.z[19:16].
			 */
			if (shader->selector->info.writes_layer)
				pos_args[1].out[2] = layer_value;

			if (shader->selector->info.writes_viewport_index) {
				LLVMValueRef v = viewport_index_value;

				v = ac_to_integer(&ctx->ac, v);
				v = LLVMBuildShl(ctx->ac.builder, v,
						 LLVMConstInt(ctx->i32, 16, 0), "");
				v = LLVMBuildOr(ctx->ac.builder, v,
						ac_to_integer(&ctx->ac,  pos_args[1].out[2]), "");
				pos_args[1].out[2] = ac_to_float(&ctx->ac, v);
				pos_args[1].enabled_channels |= 1 << 2;
			}
		} else {
			if (shader->selector->info.writes_layer)
				pos_args[1].out[2] = layer_value;

			if (shader->selector->info.writes_viewport_index) {
				pos_args[1].out[3] = viewport_index_value;
				pos_args[1].enabled_channels |= 1 << 3;
			}
		}
	}

	for (i = 0; i < 4; i++)
		if (pos_args[i].out[0])
			shader->info.nr_pos_exports++;

	pos_idx = 0;
	for (i = 0; i < 4; i++) {
		if (!pos_args[i].out[0])
			continue;

		/* Specify the target we are exporting */
		pos_args[i].target = V_008DFC_SQ_EXP_POS + pos_idx++;

		if (pos_idx == shader->info.nr_pos_exports)
			/* Specify that this is the last export */
			pos_args[i].done = 1;

		ac_build_export(&ctx->ac, &pos_args[i]);
	}

	/* Build parameter exports. */
	si_build_param_exports(ctx, outputs, noutput);
}

/**
 * Forward all outputs from the vertex shader to the TES. This is only used
 * for the fixed function TCS.
 */
static void si_copy_tcs_inputs(struct lp_build_tgsi_context *bld_base)
{
	struct si_shader_context *ctx = si_shader_context(bld_base);
	LLVMValueRef invocation_id, buffer, buffer_offset;
	LLVMValueRef lds_vertex_stride, lds_base;
	uint64_t inputs;

	invocation_id = unpack_llvm_param(ctx, ctx->abi.tcs_rel_ids, 8, 5);
	buffer = get_tess_ring_descriptor(ctx, TESS_OFFCHIP_RING_TCS);
	buffer_offset = LLVMGetParam(ctx->main_fn, ctx->param_tcs_offchip_offset);

	lds_vertex_stride = get_tcs_in_vertex_dw_stride(ctx);
	lds_base = get_tcs_in_current_patch_offset(ctx);
	lds_base = ac_build_imad(&ctx->ac, invocation_id, lds_vertex_stride,
				 lds_base);

	inputs = ctx->shader->key.mono.u.ff_tcs_inputs_to_copy;
	while (inputs) {
		unsigned i = u_bit_scan64(&inputs);

		LLVMValueRef lds_ptr = LLVMBuildAdd(ctx->ac.builder, lds_base,
		                            LLVMConstInt(ctx->i32, 4 * i, 0),
		                             "");

		LLVMValueRef buffer_addr = get_tcs_tes_buffer_address(ctx,
					      get_rel_patch_id(ctx),
		                              invocation_id,
		                              LLVMConstInt(ctx->i32, i, 0));

		LLVMValueRef value = lds_load(bld_base, ctx->ac.i32, ~0,
		                              lds_ptr);

		ac_build_buffer_store_dword(&ctx->ac, buffer, value, 4, buffer_addr,
					    buffer_offset, 0, 1, 0, true, false);
	}
}

static void si_write_tess_factors(struct lp_build_tgsi_context *bld_base,
				  LLVMValueRef rel_patch_id,
				  LLVMValueRef invocation_id,
				  LLVMValueRef tcs_out_current_patch_data_offset,
				  LLVMValueRef invoc0_tf_outer[4],
				  LLVMValueRef invoc0_tf_inner[2])
{
	struct si_shader_context *ctx = si_shader_context(bld_base);
	struct si_shader *shader = ctx->shader;
	unsigned tess_inner_index, tess_outer_index;
	LLVMValueRef lds_base, lds_inner, lds_outer, byteoffset, buffer;
	LLVMValueRef out[6], vec0, vec1, tf_base, inner[4], outer[4];
	unsigned stride, outer_comps, inner_comps, i, offset;
	struct lp_build_if_state if_ctx, inner_if_ctx;

	/* Add a barrier before loading tess factors from LDS. */
	if (!shader->key.part.tcs.epilog.invoc0_tess_factors_are_def)
		si_llvm_emit_barrier(NULL, bld_base, NULL);

	/* Do this only for invocation 0, because the tess levels are per-patch,
	 * not per-vertex.
	 *
	 * This can't jump, because invocation 0 executes this. It should
	 * at least mask out the loads and stores for other invocations.
	 */
	lp_build_if(&if_ctx, &ctx->gallivm,
		    LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ,
				  invocation_id, ctx->i32_0, ""));

	/* Determine the layout of one tess factor element in the buffer. */
	switch (shader->key.part.tcs.epilog.prim_mode) {
	case PIPE_PRIM_LINES:
		stride = 2; /* 2 dwords, 1 vec2 store */
		outer_comps = 2;
		inner_comps = 0;
		break;
	case PIPE_PRIM_TRIANGLES:
		stride = 4; /* 4 dwords, 1 vec4 store */
		outer_comps = 3;
		inner_comps = 1;
		break;
	case PIPE_PRIM_QUADS:
		stride = 6; /* 6 dwords, 2 stores (vec4 + vec2) */
		outer_comps = 4;
		inner_comps = 2;
		break;
	default:
		assert(0);
		return;
	}

	for (i = 0; i < 4; i++) {
		inner[i] = LLVMGetUndef(ctx->i32);
		outer[i] = LLVMGetUndef(ctx->i32);
	}

	if (shader->key.part.tcs.epilog.invoc0_tess_factors_are_def) {
		/* Tess factors are in VGPRs. */
		for (i = 0; i < outer_comps; i++)
			outer[i] = out[i] = invoc0_tf_outer[i];
		for (i = 0; i < inner_comps; i++)
			inner[i] = out[outer_comps+i] = invoc0_tf_inner[i];
	} else {
		/* Load tess_inner and tess_outer from LDS.
		 * Any invocation can write them, so we can't get them from a temporary.
		 */
		tess_inner_index = si_shader_io_get_unique_index_patch(TGSI_SEMANTIC_TESSINNER, 0);
		tess_outer_index = si_shader_io_get_unique_index_patch(TGSI_SEMANTIC_TESSOUTER, 0);

		lds_base = tcs_out_current_patch_data_offset;
		lds_inner = LLVMBuildAdd(ctx->ac.builder, lds_base,
					 LLVMConstInt(ctx->i32,
						      tess_inner_index * 4, 0), "");
		lds_outer = LLVMBuildAdd(ctx->ac.builder, lds_base,
					 LLVMConstInt(ctx->i32,
						      tess_outer_index * 4, 0), "");

		for (i = 0; i < outer_comps; i++) {
			outer[i] = out[i] =
				lds_load(bld_base, ctx->ac.i32, i, lds_outer);
		}
		for (i = 0; i < inner_comps; i++) {
			inner[i] = out[outer_comps+i] =
				lds_load(bld_base, ctx->ac.i32, i, lds_inner);
		}
	}

	if (shader->key.part.tcs.epilog.prim_mode == PIPE_PRIM_LINES) {
		/* For isolines, the hardware expects tess factors in the
		 * reverse order from what GLSL / TGSI specify.
		 */
		LLVMValueRef tmp = out[0];
		out[0] = out[1];
		out[1] = tmp;
	}

	/* Convert the outputs to vectors for stores. */
	vec0 = ac_build_gather_values(&ctx->ac, out, MIN2(stride, 4));
	vec1 = NULL;

	if (stride > 4)
		vec1 = ac_build_gather_values(&ctx->ac, out+4, stride - 4);

	/* Get the buffer. */
	buffer = get_tess_ring_descriptor(ctx, TCS_FACTOR_RING);

	/* Get the offset. */
	tf_base = LLVMGetParam(ctx->main_fn,
			       ctx->param_tcs_factor_offset);
	byteoffset = LLVMBuildMul(ctx->ac.builder, rel_patch_id,
				  LLVMConstInt(ctx->i32, 4 * stride, 0), "");

	lp_build_if(&inner_if_ctx, &ctx->gallivm,
		    LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ,
				  rel_patch_id, ctx->i32_0, ""));

	/* Store the dynamic HS control word. */
	offset = 0;
	if (ctx->screen->info.chip_class <= GFX8) {
		ac_build_buffer_store_dword(&ctx->ac, buffer,
					    LLVMConstInt(ctx->i32, 0x80000000, 0),
					    1, ctx->i32_0, tf_base,
					    offset, 1, 0, true, false);
		offset += 4;
	}

	lp_build_endif(&inner_if_ctx);

	/* Store the tessellation factors. */
	ac_build_buffer_store_dword(&ctx->ac, buffer, vec0,
				    MIN2(stride, 4), byteoffset, tf_base,
				    offset, 1, 0, true, false);
	offset += 16;
	if (vec1)
		ac_build_buffer_store_dword(&ctx->ac, buffer, vec1,
					    stride - 4, byteoffset, tf_base,
					    offset, 1, 0, true, false);

	/* Store the tess factors into the offchip buffer if TES reads them. */
	if (shader->key.part.tcs.epilog.tes_reads_tess_factors) {
		LLVMValueRef buf, base, inner_vec, outer_vec, tf_outer_offset;
		LLVMValueRef tf_inner_offset;
		unsigned param_outer, param_inner;

		buf = get_tess_ring_descriptor(ctx, TESS_OFFCHIP_RING_TCS);
		base = LLVMGetParam(ctx->main_fn, ctx->param_tcs_offchip_offset);

		param_outer = si_shader_io_get_unique_index_patch(
				      TGSI_SEMANTIC_TESSOUTER, 0);
		tf_outer_offset = get_tcs_tes_buffer_address(ctx, rel_patch_id, NULL,
					LLVMConstInt(ctx->i32, param_outer, 0));

		outer_vec = ac_build_gather_values(&ctx->ac, outer,
						   util_next_power_of_two(outer_comps));

		ac_build_buffer_store_dword(&ctx->ac, buf, outer_vec,
					    outer_comps, tf_outer_offset,
					    base, 0, 1, 0, true, false);
		if (inner_comps) {
			param_inner = si_shader_io_get_unique_index_patch(
					      TGSI_SEMANTIC_TESSINNER, 0);
			tf_inner_offset = get_tcs_tes_buffer_address(ctx, rel_patch_id, NULL,
					LLVMConstInt(ctx->i32, param_inner, 0));

			inner_vec = inner_comps == 1 ? inner[0] :
				    ac_build_gather_values(&ctx->ac, inner, inner_comps);
			ac_build_buffer_store_dword(&ctx->ac, buf, inner_vec,
						    inner_comps, tf_inner_offset,
						    base, 0, 1, 0, true, false);
		}
	}

	lp_build_endif(&if_ctx);
}

static LLVMValueRef
si_insert_input_ret(struct si_shader_context *ctx, LLVMValueRef ret,
		    unsigned param, unsigned return_index)
{
	return LLVMBuildInsertValue(ctx->ac.builder, ret,
				    LLVMGetParam(ctx->main_fn, param),
				    return_index, "");
}

static LLVMValueRef
si_insert_input_ret_float(struct si_shader_context *ctx, LLVMValueRef ret,
			  unsigned param, unsigned return_index)
{
	LLVMBuilderRef builder = ctx->ac.builder;
	LLVMValueRef p = LLVMGetParam(ctx->main_fn, param);

	return LLVMBuildInsertValue(builder, ret,
				    ac_to_float(&ctx->ac, p),
				    return_index, "");
}

static LLVMValueRef
si_insert_input_ptr(struct si_shader_context *ctx, LLVMValueRef ret,
		    unsigned param, unsigned return_index)
{
	LLVMBuilderRef builder = ctx->ac.builder;
	LLVMValueRef ptr = LLVMGetParam(ctx->main_fn, param);
	ptr = LLVMBuildPtrToInt(builder, ptr, ctx->i32, "");
	return LLVMBuildInsertValue(builder, ret, ptr, return_index, "");
}

/* This only writes the tessellation factor levels. */
static void si_llvm_emit_tcs_epilogue(struct ac_shader_abi *abi,
				      unsigned max_outputs,
				      LLVMValueRef *addrs)
{
	struct si_shader_context *ctx = si_shader_context_from_abi(abi);
	struct lp_build_tgsi_context *bld_base = &ctx->bld_base;
	LLVMBuilderRef builder = ctx->ac.builder;
	LLVMValueRef rel_patch_id, invocation_id, tf_lds_offset;

	si_copy_tcs_inputs(bld_base);

	rel_patch_id = get_rel_patch_id(ctx);
	invocation_id = unpack_llvm_param(ctx, ctx->abi.tcs_rel_ids, 8, 5);
	tf_lds_offset = get_tcs_out_current_patch_data_offset(ctx);

	if (ctx->screen->info.chip_class >= GFX9) {
		LLVMBasicBlockRef blocks[2] = {
			LLVMGetInsertBlock(builder),
			ctx->merged_wrap_if_state.entry_block
		};
		LLVMValueRef values[2];

		lp_build_endif(&ctx->merged_wrap_if_state);

		values[0] = rel_patch_id;
		values[1] = LLVMGetUndef(ctx->i32);
		rel_patch_id = ac_build_phi(&ctx->ac, ctx->i32, 2, values, blocks);

		values[0] = tf_lds_offset;
		values[1] = LLVMGetUndef(ctx->i32);
		tf_lds_offset = ac_build_phi(&ctx->ac, ctx->i32, 2, values, blocks);

		values[0] = invocation_id;
		values[1] = ctx->i32_1; /* cause the epilog to skip threads */
		invocation_id = ac_build_phi(&ctx->ac, ctx->i32, 2, values, blocks);
	}

	/* Return epilog parameters from this function. */
	LLVMValueRef ret = ctx->return_value;
	unsigned vgpr;

	if (ctx->screen->info.chip_class >= GFX9) {
		ret = si_insert_input_ret(ctx, ret, ctx->param_tcs_offchip_layout,
					  8 + GFX9_SGPR_TCS_OFFCHIP_LAYOUT);
		ret = si_insert_input_ret(ctx, ret, ctx->param_tcs_out_lds_layout,
					  8 + GFX9_SGPR_TCS_OUT_LAYOUT);
		/* Tess offchip and tess factor offsets are at the beginning. */
		ret = si_insert_input_ret(ctx, ret, ctx->param_tcs_offchip_offset, 2);
		ret = si_insert_input_ret(ctx, ret, ctx->param_tcs_factor_offset, 4);
		vgpr = 8 + GFX9_SGPR_TCS_OUT_LAYOUT + 1;
	} else {
		ret = si_insert_input_ret(ctx, ret, ctx->param_tcs_offchip_layout,
					  GFX6_SGPR_TCS_OFFCHIP_LAYOUT);
		ret = si_insert_input_ret(ctx, ret, ctx->param_tcs_out_lds_layout,
					  GFX6_SGPR_TCS_OUT_LAYOUT);
		/* Tess offchip and tess factor offsets are after user SGPRs. */
		ret = si_insert_input_ret(ctx, ret, ctx->param_tcs_offchip_offset,
					  GFX6_TCS_NUM_USER_SGPR);
		ret = si_insert_input_ret(ctx, ret, ctx->param_tcs_factor_offset,
					  GFX6_TCS_NUM_USER_SGPR + 1);
		vgpr = GFX6_TCS_NUM_USER_SGPR + 2;
	}

	/* VGPRs */
	rel_patch_id = ac_to_float(&ctx->ac, rel_patch_id);
	invocation_id = ac_to_float(&ctx->ac, invocation_id);
	tf_lds_offset = ac_to_float(&ctx->ac, tf_lds_offset);

	/* Leave a hole corresponding to the two input VGPRs. This ensures that
	 * the invocation_id output does not alias the tcs_rel_ids input,
	 * which saves a V_MOV on gfx9.
	 */
	vgpr += 2;

	ret = LLVMBuildInsertValue(builder, ret, rel_patch_id, vgpr++, "");
	ret = LLVMBuildInsertValue(builder, ret, invocation_id, vgpr++, "");

	if (ctx->shader->selector->tcs_info.tessfactors_are_def_in_all_invocs) {
		vgpr++; /* skip the tess factor LDS offset */
		for (unsigned i = 0; i < 6; i++) {
			LLVMValueRef value =
				LLVMBuildLoad(builder, ctx->invoc0_tess_factors[i], "");
			value = ac_to_float(&ctx->ac, value);
			ret = LLVMBuildInsertValue(builder, ret, value, vgpr++, "");
		}
	} else {
		ret = LLVMBuildInsertValue(builder, ret, tf_lds_offset, vgpr++, "");
	}
	ctx->return_value = ret;
}

/* Pass TCS inputs from LS to TCS on GFX9. */
static void si_set_ls_return_value_for_tcs(struct si_shader_context *ctx)
{
	LLVMValueRef ret = ctx->return_value;

	ret = si_insert_input_ptr(ctx, ret, 0, 0);
	ret = si_insert_input_ptr(ctx, ret, 1, 1);
	ret = si_insert_input_ret(ctx, ret, ctx->param_tcs_offchip_offset, 2);
	ret = si_insert_input_ret(ctx, ret, ctx->param_merged_wave_info, 3);
	ret = si_insert_input_ret(ctx, ret, ctx->param_tcs_factor_offset, 4);
	ret = si_insert_input_ret(ctx, ret, ctx->param_merged_scratch_offset, 5);

	ret = si_insert_input_ptr(ctx, ret, ctx->param_rw_buffers,
				  8 + SI_SGPR_RW_BUFFERS);
	ret = si_insert_input_ptr(ctx, ret,
				  ctx->param_bindless_samplers_and_images,
				  8 + SI_SGPR_BINDLESS_SAMPLERS_AND_IMAGES);

	ret = si_insert_input_ret(ctx, ret, ctx->param_vs_state_bits,
				  8 + SI_SGPR_VS_STATE_BITS);

	ret = si_insert_input_ret(ctx, ret, ctx->param_tcs_offchip_layout,
				  8 + GFX9_SGPR_TCS_OFFCHIP_LAYOUT);
	ret = si_insert_input_ret(ctx, ret, ctx->param_tcs_out_lds_offsets,
				  8 + GFX9_SGPR_TCS_OUT_OFFSETS);
	ret = si_insert_input_ret(ctx, ret, ctx->param_tcs_out_lds_layout,
				  8 + GFX9_SGPR_TCS_OUT_LAYOUT);

	unsigned vgpr = 8 + GFX9_TCS_NUM_USER_SGPR;
	ret = LLVMBuildInsertValue(ctx->ac.builder, ret,
				   ac_to_float(&ctx->ac, ctx->abi.tcs_patch_id),
				   vgpr++, "");
	ret = LLVMBuildInsertValue(ctx->ac.builder, ret,
				   ac_to_float(&ctx->ac, ctx->abi.tcs_rel_ids),
				   vgpr++, "");
	ctx->return_value = ret;
}

/* Pass GS inputs from ES to GS on GFX9. */
static void si_set_es_return_value_for_gs(struct si_shader_context *ctx)
{
	LLVMValueRef ret = ctx->return_value;

	ret = si_insert_input_ptr(ctx, ret, 0, 0);
	ret = si_insert_input_ptr(ctx, ret, 1, 1);
	ret = si_insert_input_ret(ctx, ret, ctx->param_gs2vs_offset, 2);
	ret = si_insert_input_ret(ctx, ret, ctx->param_merged_wave_info, 3);
	ret = si_insert_input_ret(ctx, ret, ctx->param_merged_scratch_offset, 5);

	ret = si_insert_input_ptr(ctx, ret, ctx->param_rw_buffers,
				  8 + SI_SGPR_RW_BUFFERS);
	ret = si_insert_input_ptr(ctx, ret,
				  ctx->param_bindless_samplers_and_images,
				  8 + SI_SGPR_BINDLESS_SAMPLERS_AND_IMAGES);

	unsigned vgpr;
	if (ctx->type == PIPE_SHADER_VERTEX)
		vgpr = 8 + GFX9_VSGS_NUM_USER_SGPR;
	else
		vgpr = 8 + GFX9_TESGS_NUM_USER_SGPR;

	for (unsigned i = 0; i < 5; i++) {
		unsigned param = ctx->param_gs_vtx01_offset + i;
		ret = si_insert_input_ret_float(ctx, ret, param, vgpr++);
	}
	ctx->return_value = ret;
}

static void si_llvm_emit_ls_epilogue(struct ac_shader_abi *abi,
				     unsigned max_outputs,
				     LLVMValueRef *addrs)
{
	struct si_shader_context *ctx = si_shader_context_from_abi(abi);
	struct si_shader *shader = ctx->shader;
	struct tgsi_shader_info *info = &shader->selector->info;
	unsigned i, chan;
	LLVMValueRef vertex_id = LLVMGetParam(ctx->main_fn,
					      ctx->param_rel_auto_id);
	LLVMValueRef vertex_dw_stride = get_tcs_in_vertex_dw_stride(ctx);
	LLVMValueRef base_dw_addr = LLVMBuildMul(ctx->ac.builder, vertex_id,
						 vertex_dw_stride, "");

	/* Write outputs to LDS. The next shader (TCS aka HS) will read
	 * its inputs from it. */
	for (i = 0; i < info->num_outputs; i++) {
		unsigned name = info->output_semantic_name[i];
		unsigned index = info->output_semantic_index[i];

		/* The ARB_shader_viewport_layer_array spec contains the
		 * following issue:
		 *
		 *    2) What happens if gl_ViewportIndex or gl_Layer is
		 *    written in the vertex shader and a geometry shader is
		 *    present?
		 *
		 *    RESOLVED: The value written by the last vertex processing
		 *    stage is used. If the last vertex processing stage
		 *    (vertex, tessellation evaluation or geometry) does not
		 *    statically assign to gl_ViewportIndex or gl_Layer, index
		 *    or layer zero is assumed.
		 *
		 * So writes to those outputs in VS-as-LS are simply ignored.
		 */
		if (name == TGSI_SEMANTIC_LAYER ||
		    name == TGSI_SEMANTIC_VIEWPORT_INDEX)
			continue;

		int param = si_shader_io_get_unique_index(name, index, false);
		LLVMValueRef dw_addr = LLVMBuildAdd(ctx->ac.builder, base_dw_addr,
					LLVMConstInt(ctx->i32, param * 4, 0), "");

		for (chan = 0; chan < 4; chan++) {
			if (!(info->output_usagemask[i] & (1 << chan)))
				continue;

			lds_store(ctx, chan, dw_addr,
				  LLVMBuildLoad(ctx->ac.builder, addrs[4 * i + chan], ""));
		}
	}

	if (ctx->screen->info.chip_class >= GFX9)
		si_set_ls_return_value_for_tcs(ctx);
}

static void si_llvm_emit_es_epilogue(struct ac_shader_abi *abi,
				     unsigned max_outputs,
				     LLVMValueRef *addrs)
{
	struct si_shader_context *ctx = si_shader_context_from_abi(abi);
	struct si_shader *es = ctx->shader;
	struct tgsi_shader_info *info = &es->selector->info;
	LLVMValueRef soffset = LLVMGetParam(ctx->main_fn,
					    ctx->param_es2gs_offset);
	LLVMValueRef lds_base = NULL;
	unsigned chan;
	int i;

	if (ctx->screen->info.chip_class >= GFX9 && info->num_outputs) {
		unsigned itemsize_dw = es->selector->esgs_itemsize / 4;
		LLVMValueRef vertex_idx = ac_get_thread_id(&ctx->ac);
		LLVMValueRef wave_idx = si_unpack_param(ctx, ctx->param_merged_wave_info, 24, 4);
		vertex_idx = LLVMBuildOr(ctx->ac.builder, vertex_idx,
					 LLVMBuildMul(ctx->ac.builder, wave_idx,
						      LLVMConstInt(ctx->i32, 64, false), ""), "");
		lds_base = LLVMBuildMul(ctx->ac.builder, vertex_idx,
					LLVMConstInt(ctx->i32, itemsize_dw, 0), "");
	}

	for (i = 0; i < info->num_outputs; i++) {
		int param;

		if (info->output_semantic_name[i] == TGSI_SEMANTIC_VIEWPORT_INDEX ||
		    info->output_semantic_name[i] == TGSI_SEMANTIC_LAYER)
			continue;

		param = si_shader_io_get_unique_index(info->output_semantic_name[i],
						      info->output_semantic_index[i], false);

		for (chan = 0; chan < 4; chan++) {
			if (!(info->output_usagemask[i] & (1 << chan)))
				continue;

			LLVMValueRef out_val = LLVMBuildLoad(ctx->ac.builder, addrs[4 * i + chan], "");
			out_val = ac_to_integer(&ctx->ac, out_val);

			/* GFX9 has the ESGS ring in LDS. */
			if (ctx->screen->info.chip_class >= GFX9) {
				lds_store(ctx, param * 4 + chan, lds_base, out_val);
				continue;
			}

			ac_build_buffer_store_dword(&ctx->ac,
						    ctx->esgs_ring,
						    out_val, 1, NULL, soffset,
						    (4 * param + chan) * 4,
						    1, 1, true, true);
		}
	}

	if (ctx->screen->info.chip_class >= GFX9)
		si_set_es_return_value_for_gs(ctx);
}

static LLVMValueRef si_get_gs_wave_id(struct si_shader_context *ctx)
{
	if (ctx->screen->info.chip_class >= GFX9)
		return si_unpack_param(ctx, ctx->param_merged_wave_info, 16, 8);
	else
		return LLVMGetParam(ctx->main_fn, ctx->param_gs_wave_id);
}

static void emit_gs_epilogue(struct si_shader_context *ctx)
{
	ac_build_sendmsg(&ctx->ac, AC_SENDMSG_GS_OP_NOP | AC_SENDMSG_GS_DONE,
			 si_get_gs_wave_id(ctx));

	if (ctx->screen->info.chip_class >= GFX9)
		lp_build_endif(&ctx->merged_wrap_if_state);
}

static void si_llvm_emit_gs_epilogue(struct ac_shader_abi *abi,
				     unsigned max_outputs,
				     LLVMValueRef *addrs)
{
	struct si_shader_context *ctx = si_shader_context_from_abi(abi);
	struct tgsi_shader_info UNUSED *info = &ctx->shader->selector->info;

	assert(info->num_outputs <= max_outputs);

	emit_gs_epilogue(ctx);
}

static void si_tgsi_emit_gs_epilogue(struct lp_build_tgsi_context *bld_base)
{
	struct si_shader_context *ctx = si_shader_context(bld_base);
	emit_gs_epilogue(ctx);
}

static void si_llvm_emit_vs_epilogue(struct ac_shader_abi *abi,
				     unsigned max_outputs,
				     LLVMValueRef *addrs)
{
	struct si_shader_context *ctx = si_shader_context_from_abi(abi);
	struct tgsi_shader_info *info = &ctx->shader->selector->info;
	struct si_shader_output_values *outputs = NULL;
	int i,j;

	assert(!ctx->shader->is_gs_copy_shader);
	assert(info->num_outputs <= max_outputs);

	outputs = MALLOC((info->num_outputs + 1) * sizeof(outputs[0]));

	/* Vertex color clamping.
	 *
	 * This uses a state constant loaded in a user data SGPR and
	 * an IF statement is added that clamps all colors if the constant
	 * is true.
	 */
	struct lp_build_if_state if_ctx;
	LLVMValueRef cond = NULL;
	LLVMValueRef addr, val;

	for (i = 0; i < info->num_outputs; i++) {
		if (info->output_semantic_name[i] != TGSI_SEMANTIC_COLOR &&
		    info->output_semantic_name[i] != TGSI_SEMANTIC_BCOLOR)
			continue;

		/* We've found a color. */
		if (!cond) {
			/* The state is in the first bit of the user SGPR. */
			cond = LLVMGetParam(ctx->main_fn,
					    ctx->param_vs_state_bits);
			cond = LLVMBuildTrunc(ctx->ac.builder, cond,
					      ctx->i1, "");
			lp_build_if(&if_ctx, &ctx->gallivm, cond);
		}

		for (j = 0; j < 4; j++) {
			addr = addrs[4 * i + j];
			val = LLVMBuildLoad(ctx->ac.builder, addr, "");
			val = ac_build_clamp(&ctx->ac, val);
			LLVMBuildStore(ctx->ac.builder, val, addr);
		}
	}

	if (cond)
		lp_build_endif(&if_ctx);

	for (i = 0; i < info->num_outputs; i++) {
		outputs[i].semantic_name = info->output_semantic_name[i];
		outputs[i].semantic_index = info->output_semantic_index[i];

		for (j = 0; j < 4; j++) {
			outputs[i].values[j] =
				LLVMBuildLoad(ctx->ac.builder,
					      addrs[4 * i + j],
					      "");
			outputs[i].vertex_stream[j] =
				(info->output_streams[i] >> (2 * j)) & 3;
		}
	}

	if (ctx->shader->selector->so.num_outputs)
		si_llvm_emit_streamout(ctx, outputs, i, 0);

	/* Export PrimitiveID. */
	if (ctx->shader->key.mono.u.vs_export_prim_id) {
		outputs[i].semantic_name = TGSI_SEMANTIC_PRIMID;
		outputs[i].semantic_index = 0;
		outputs[i].values[0] = ac_to_float(&ctx->ac, get_primitive_id(ctx, 0));
		for (j = 1; j < 4; j++)
			outputs[i].values[j] = LLVMConstReal(ctx->f32, 0);

		memset(outputs[i].vertex_stream, 0,
		       sizeof(outputs[i].vertex_stream));
		i++;
	}

	si_llvm_export_vs(ctx, outputs, i);
	FREE(outputs);
}

static void si_llvm_emit_prim_discard_cs_epilogue(struct ac_shader_abi *abi,
						  unsigned max_outputs,
						  LLVMValueRef *addrs)
{
	struct si_shader_context *ctx = si_shader_context_from_abi(abi);
	struct tgsi_shader_info *info = &ctx->shader->selector->info;
	LLVMValueRef pos[4] = {};

	assert(info->num_outputs <= max_outputs);

	for (unsigned i = 0; i < info->num_outputs; i++) {
		if (info->output_semantic_name[i] != TGSI_SEMANTIC_POSITION)
			continue;

		for (unsigned chan = 0; chan < 4; chan++)
			pos[chan] = LLVMBuildLoad(ctx->ac.builder, addrs[4 * i + chan], "");
		break;
	}
	assert(pos[0] != NULL);

	/* Return the position output. */
	LLVMValueRef ret = ctx->return_value;
	for (unsigned chan = 0; chan < 4; chan++)
		ret = LLVMBuildInsertValue(ctx->ac.builder, ret, pos[chan], chan, "");
	ctx->return_value = ret;
}

static void si_tgsi_emit_epilogue(struct lp_build_tgsi_context *bld_base)
{
	struct si_shader_context *ctx = si_shader_context(bld_base);

	ctx->abi.emit_outputs(&ctx->abi, RADEON_LLVM_MAX_OUTPUTS,
			      &ctx->outputs[0][0]);
}

struct si_ps_exports {
	unsigned num;
	struct ac_export_args args[10];
};

static void si_export_mrt_z(struct lp_build_tgsi_context *bld_base,
			    LLVMValueRef depth, LLVMValueRef stencil,
			    LLVMValueRef samplemask, struct si_ps_exports *exp)
{
	struct si_shader_context *ctx = si_shader_context(bld_base);
	struct ac_export_args args;

	ac_export_mrt_z(&ctx->ac, depth, stencil, samplemask, &args);

	memcpy(&exp->args[exp->num++], &args, sizeof(args));
}

static void si_export_mrt_color(struct lp_build_tgsi_context *bld_base,
				LLVMValueRef *color, unsigned index,
				unsigned samplemask_param,
				bool is_last, struct si_ps_exports *exp)
{
	struct si_shader_context *ctx = si_shader_context(bld_base);
	int i;

	/* Clamp color */
	if (ctx->shader->key.part.ps.epilog.clamp_color)
		for (i = 0; i < 4; i++)
			color[i] = ac_build_clamp(&ctx->ac, color[i]);

	/* Alpha to one */
	if (ctx->shader->key.part.ps.epilog.alpha_to_one)
		color[3] = ctx->ac.f32_1;

	/* Alpha test */
	if (index == 0 &&
	    ctx->shader->key.part.ps.epilog.alpha_func != PIPE_FUNC_ALWAYS)
		si_alpha_test(bld_base, color[3]);

	/* Line & polygon smoothing */
	if (ctx->shader->key.part.ps.epilog.poly_line_smoothing)
		color[3] = si_scale_alpha_by_sample_mask(bld_base, color[3],
							 samplemask_param);

	/* If last_cbuf > 0, FS_COLOR0_WRITES_ALL_CBUFS is true. */
	if (ctx->shader->key.part.ps.epilog.last_cbuf > 0) {
		struct ac_export_args args[8];
		int c, last = -1;

		/* Get the export arguments, also find out what the last one is. */
		for (c = 0; c <= ctx->shader->key.part.ps.epilog.last_cbuf; c++) {
			si_llvm_init_export_args(ctx, color,
						 V_008DFC_SQ_EXP_MRT + c, &args[c]);
			if (args[c].enabled_channels)
				last = c;
		}

		/* Emit all exports. */
		for (c = 0; c <= ctx->shader->key.part.ps.epilog.last_cbuf; c++) {
			if (is_last && last == c) {
				args[c].valid_mask = 1; /* whether the EXEC mask is valid */
				args[c].done = 1; /* DONE bit */
			} else if (!args[c].enabled_channels)
				continue; /* unnecessary NULL export */

			memcpy(&exp->args[exp->num++], &args[c], sizeof(args[c]));
		}
	} else {
		struct ac_export_args args;

		/* Export */
		si_llvm_init_export_args(ctx, color, V_008DFC_SQ_EXP_MRT + index,
					 &args);
		if (is_last) {
			args.valid_mask = 1; /* whether the EXEC mask is valid */
			args.done = 1; /* DONE bit */
		} else if (!args.enabled_channels)
			return; /* unnecessary NULL export */

		memcpy(&exp->args[exp->num++], &args, sizeof(args));
	}
}

static void si_emit_ps_exports(struct si_shader_context *ctx,
			       struct si_ps_exports *exp)
{
	for (unsigned i = 0; i < exp->num; i++)
		ac_build_export(&ctx->ac, &exp->args[i]);
}

/**
 * Return PS outputs in this order:
 *
 * v[0:3] = color0.xyzw
 * v[4:7] = color1.xyzw
 * ...
 * vN+0 = Depth
 * vN+1 = Stencil
 * vN+2 = SampleMask
 * vN+3 = SampleMaskIn (used for OpenGL smoothing)
 *
 * The alpha-ref SGPR is returned via its original location.
 */
static void si_llvm_return_fs_outputs(struct ac_shader_abi *abi,
				      unsigned max_outputs,
				      LLVMValueRef *addrs)
{
	struct si_shader_context *ctx = si_shader_context_from_abi(abi);
	struct si_shader *shader = ctx->shader;
	struct tgsi_shader_info *info = &shader->selector->info;
	LLVMBuilderRef builder = ctx->ac.builder;
	unsigned i, j, first_vgpr, vgpr;

	LLVMValueRef color[8][4] = {};
	LLVMValueRef depth = NULL, stencil = NULL, samplemask = NULL;
	LLVMValueRef ret;

	if (ctx->postponed_kill)
		ac_build_kill_if_false(&ctx->ac, LLVMBuildLoad(builder, ctx->postponed_kill, ""));

	/* Read the output values. */
	for (i = 0; i < info->num_outputs; i++) {
		unsigned semantic_name = info->output_semantic_name[i];
		unsigned semantic_index = info->output_semantic_index[i];

		switch (semantic_name) {
		case TGSI_SEMANTIC_COLOR:
			assert(semantic_index < 8);
			for (j = 0; j < 4; j++) {
				LLVMValueRef ptr = addrs[4 * i + j];
				LLVMValueRef result = LLVMBuildLoad(builder, ptr, "");
				color[semantic_index][j] = result;
			}
			break;
		case TGSI_SEMANTIC_POSITION:
			depth = LLVMBuildLoad(builder,
					      addrs[4 * i + 2], "");
			break;
		case TGSI_SEMANTIC_STENCIL:
			stencil = LLVMBuildLoad(builder,
						addrs[4 * i + 1], "");
			break;
		case TGSI_SEMANTIC_SAMPLEMASK:
			samplemask = LLVMBuildLoad(builder,
						   addrs[4 * i + 0], "");
			break;
		default:
			fprintf(stderr, "Warning: GFX6 unhandled fs output type:%d\n",
				semantic_name);
		}
	}

	/* Fill the return structure. */
	ret = ctx->return_value;

	/* Set SGPRs. */
	ret = LLVMBuildInsertValue(builder, ret,
				   ac_to_integer(&ctx->ac,
                                                 LLVMGetParam(ctx->main_fn,
                                                              SI_PARAM_ALPHA_REF)),
				   SI_SGPR_ALPHA_REF, "");

	/* Set VGPRs */
	first_vgpr = vgpr = SI_SGPR_ALPHA_REF + 1;
	for (i = 0; i < ARRAY_SIZE(color); i++) {
		if (!color[i][0])
			continue;

		for (j = 0; j < 4; j++)
			ret = LLVMBuildInsertValue(builder, ret, color[i][j], vgpr++, "");
	}
	if (depth)
		ret = LLVMBuildInsertValue(builder, ret, depth, vgpr++, "");
	if (stencil)
		ret = LLVMBuildInsertValue(builder, ret, stencil, vgpr++, "");
	if (samplemask)
		ret = LLVMBuildInsertValue(builder, ret, samplemask, vgpr++, "");

	/* Add the input sample mask for smoothing at the end. */
	if (vgpr < first_vgpr + PS_EPILOG_SAMPLEMASK_MIN_LOC)
		vgpr = first_vgpr + PS_EPILOG_SAMPLEMASK_MIN_LOC;
	ret = LLVMBuildInsertValue(builder, ret,
				   LLVMGetParam(ctx->main_fn,
						SI_PARAM_SAMPLE_COVERAGE), vgpr++, "");

	ctx->return_value = ret;
}

static void membar_emit(
		const struct lp_build_tgsi_action *action,
		struct lp_build_tgsi_context *bld_base,
		struct lp_build_emit_data *emit_data)
{
	struct si_shader_context *ctx = si_shader_context(bld_base);
	LLVMValueRef src0 = lp_build_emit_fetch(bld_base, emit_data->inst, 0, 0);
	unsigned flags = LLVMConstIntGetZExtValue(src0);
	unsigned waitcnt = NOOP_WAITCNT;

	if (flags & TGSI_MEMBAR_THREAD_GROUP)
		waitcnt &= VM_CNT & LGKM_CNT;

	if (flags & (TGSI_MEMBAR_ATOMIC_BUFFER |
		     TGSI_MEMBAR_SHADER_BUFFER |
		     TGSI_MEMBAR_SHADER_IMAGE))
		waitcnt &= VM_CNT;

	if (flags & TGSI_MEMBAR_SHARED)
		waitcnt &= LGKM_CNT;

	if (waitcnt != NOOP_WAITCNT)
		ac_build_waitcnt(&ctx->ac, waitcnt);
}

static void clock_emit(
		const struct lp_build_tgsi_action *action,
		struct lp_build_tgsi_context *bld_base,
		struct lp_build_emit_data *emit_data)
{
	struct si_shader_context *ctx = si_shader_context(bld_base);
	LLVMValueRef tmp = ac_build_shader_clock(&ctx->ac);

	emit_data->output[0] =
		LLVMBuildExtractElement(ctx->ac.builder, tmp, ctx->i32_0, "");
	emit_data->output[1] =
		LLVMBuildExtractElement(ctx->ac.builder, tmp, ctx->i32_1, "");
}

static void si_llvm_emit_ddxy(
	const struct lp_build_tgsi_action *action,
	struct lp_build_tgsi_context *bld_base,
	struct lp_build_emit_data *emit_data)
{
	struct si_shader_context *ctx = si_shader_context(bld_base);
	unsigned opcode = emit_data->info->opcode;
	LLVMValueRef val;
	int idx;
	unsigned mask;

	if (opcode == TGSI_OPCODE_DDX_FINE)
		mask = AC_TID_MASK_LEFT;
	else if (opcode == TGSI_OPCODE_DDY_FINE)
		mask = AC_TID_MASK_TOP;
	else
		mask = AC_TID_MASK_TOP_LEFT;

	/* for DDX we want to next X pixel, DDY next Y pixel. */
	idx = (opcode == TGSI_OPCODE_DDX || opcode == TGSI_OPCODE_DDX_FINE) ? 1 : 2;

	val = ac_to_integer(&ctx->ac, emit_data->args[0]);
	val = ac_build_ddxy(&ctx->ac, mask, idx, val);
	emit_data->output[emit_data->chan] = val;
}

static void build_interp_intrinsic(const struct lp_build_tgsi_action *action,
				struct lp_build_tgsi_context *bld_base,
				struct lp_build_emit_data *emit_data)
{
	struct si_shader_context *ctx = si_shader_context(bld_base);
	struct si_shader *shader = ctx->shader;
	const struct tgsi_shader_info *info = &shader->selector->info;
	LLVMValueRef interp_param;
	const struct tgsi_full_instruction *inst = emit_data->inst;
	const struct tgsi_full_src_register *input = &inst->Src[0];
	int input_base, input_array_size;
	int chan;
	int i;
	LLVMValueRef prim_mask = ctx->abi.prim_mask;
	LLVMValueRef array_idx, offset_x = NULL, offset_y = NULL;
	int interp_param_idx;
	unsigned interp;
	unsigned location;

	if (inst->Instruction.Opcode == TGSI_OPCODE_INTERP_OFFSET) {
		/* offset is in second src, first two channels */
		offset_x = lp_build_emit_fetch(bld_base, emit_data->inst, 1,
					       TGSI_CHAN_X);
		offset_y = lp_build_emit_fetch(bld_base, emit_data->inst, 1,
					       TGSI_CHAN_Y);
	} else if (inst->Instruction.Opcode == TGSI_OPCODE_INTERP_SAMPLE) {
		LLVMValueRef sample_position;
		LLVMValueRef sample_id;
		LLVMValueRef halfval = LLVMConstReal(ctx->f32, 0.5f);

		/* fetch sample ID, then fetch its sample position,
		 * and place into first two channels.
		 */
		sample_id = lp_build_emit_fetch(bld_base,
						emit_data->inst, 1, TGSI_CHAN_X);
		sample_id = ac_to_integer(&ctx->ac, sample_id);

		/* Section 8.13.2 (Interpolation Functions) of the OpenGL Shading
		 * Language 4.50 spec says about interpolateAtSample:
		 *
		 *    "Returns the value of the input interpolant variable at
		 *     the location of sample number sample. If multisample
		 *     buffers are not available, the input variable will be
		 *     evaluated at the center of the pixel. If sample sample
		 *     does not exist, the position used to interpolate the
		 *     input variable is undefined."
		 *
		 * This means that sample_id values outside of the valid are
		 * in fact valid input, and the usual mechanism for loading the
		 * sample position doesn't work.
		 */
		if (ctx->shader->key.mono.u.ps.interpolate_at_sample_force_center) {
			LLVMValueRef center[4] = {
				LLVMConstReal(ctx->f32, 0.5),
				LLVMConstReal(ctx->f32, 0.5),
				ctx->ac.f32_0,
				ctx->ac.f32_0,
			};

			sample_position = ac_build_gather_values(&ctx->ac, center, 4);
		} else {
			sample_position = load_sample_position(&ctx->abi, sample_id);
		}

		offset_x = LLVMBuildExtractElement(ctx->ac.builder, sample_position,
						   ctx->i32_0, "");

		offset_x = LLVMBuildFSub(ctx->ac.builder, offset_x, halfval, "");
		offset_y = LLVMBuildExtractElement(ctx->ac.builder, sample_position,
						   ctx->i32_1, "");
		offset_y = LLVMBuildFSub(ctx->ac.builder, offset_y, halfval, "");
	}

	assert(input->Register.File == TGSI_FILE_INPUT);

	if (input->Register.Indirect) {
		unsigned array_id = input->Indirect.ArrayID;

		if (array_id) {
			input_base = info->input_array_first[array_id];
			input_array_size = info->input_array_last[array_id] - input_base + 1;
		} else {
			input_base = inst->Src[0].Register.Index;
			input_array_size = info->num_inputs - input_base;
		}

		array_idx = si_get_indirect_index(ctx, &input->Indirect,
						  1, input->Register.Index - input_base);
	} else {
		input_base = inst->Src[0].Register.Index;
		input_array_size = 1;
		array_idx = ctx->i32_0;
	}

	interp = shader->selector->info.input_interpolate[input_base];

	if (inst->Instruction.Opcode == TGSI_OPCODE_INTERP_OFFSET ||
	    inst->Instruction.Opcode == TGSI_OPCODE_INTERP_SAMPLE)
		location = TGSI_INTERPOLATE_LOC_CENTER;
	else
		location = TGSI_INTERPOLATE_LOC_CENTROID;

	interp_param_idx = lookup_interp_param_index(interp, location);
	if (interp_param_idx == -1)
		return;
	else if (interp_param_idx)
		interp_param = LLVMGetParam(ctx->main_fn, interp_param_idx);
	else
		interp_param = NULL;

	if (inst->Instruction.Opcode == TGSI_OPCODE_INTERP_OFFSET ||
	    inst->Instruction.Opcode == TGSI_OPCODE_INTERP_SAMPLE) {
		LLVMValueRef ij_out[2];
		LLVMValueRef ddxy_out = ac_build_ddxy_interp(&ctx->ac, interp_param);

		/*
		 * take the I then J parameters, and the DDX/Y for it, and
		 * calculate the IJ inputs for the interpolator.
		 * temp1 = ddx * offset/sample.x + I;
		 * interp_param.I = ddy * offset/sample.y + temp1;
		 * temp1 = ddx * offset/sample.x + J;
		 * interp_param.J = ddy * offset/sample.y + temp1;
		 */
		for (i = 0; i < 2; i++) {
			LLVMValueRef ix_ll = LLVMConstInt(ctx->i32, i, 0);
			LLVMValueRef iy_ll = LLVMConstInt(ctx->i32, i + 2, 0);
			LLVMValueRef ddx_el = LLVMBuildExtractElement(ctx->ac.builder,
								      ddxy_out, ix_ll, "");
			LLVMValueRef ddy_el = LLVMBuildExtractElement(ctx->ac.builder,
								      ddxy_out, iy_ll, "");
			LLVMValueRef interp_el = LLVMBuildExtractElement(ctx->ac.builder,
									 interp_param, ix_ll, "");
			LLVMValueRef temp;

			interp_el = ac_to_float(&ctx->ac, interp_el);

			temp = ac_build_fmad(&ctx->ac, ddx_el, offset_x, interp_el);
			ij_out[i] = ac_build_fmad(&ctx->ac, ddy_el, offset_y, temp);
		}
		interp_param = ac_build_gather_values(&ctx->ac, ij_out, 2);
	}

	if (interp_param)
		interp_param = ac_to_float(&ctx->ac, interp_param);

	for (chan = 0; chan < 4; chan++) {
		LLVMValueRef gather = LLVMGetUndef(LLVMVectorType(ctx->f32, input_array_size));
		unsigned schan = tgsi_util_get_full_src_register_swizzle(&inst->Src[0], chan);

		for (unsigned idx = 0; idx < input_array_size; ++idx) {
			LLVMValueRef v, i = NULL, j = NULL;

			if (interp_param) {
				i = LLVMBuildExtractElement(
					ctx->ac.builder, interp_param, ctx->i32_0, "");
				j = LLVMBuildExtractElement(
					ctx->ac.builder, interp_param, ctx->i32_1, "");
			}
			v = si_build_fs_interp(ctx, input_base + idx, schan,
					       prim_mask, i, j);

			gather = LLVMBuildInsertElement(ctx->ac.builder,
				gather, v, LLVMConstInt(ctx->i32, idx, false), "");
		}

		emit_data->output[chan] = LLVMBuildExtractElement(
			ctx->ac.builder, gather, array_idx, "");
	}
}

static void vote_all_emit(
	const struct lp_build_tgsi_action *action,
	struct lp_build_tgsi_context *bld_base,
	struct lp_build_emit_data *emit_data)
{
	struct si_shader_context *ctx = si_shader_context(bld_base);

        LLVMValueRef tmp = ac_build_vote_all(&ctx->ac, emit_data->args[0]);
	emit_data->output[emit_data->chan] =
		LLVMBuildSExt(ctx->ac.builder, tmp, ctx->i32, "");
}

static void vote_any_emit(
	const struct lp_build_tgsi_action *action,
	struct lp_build_tgsi_context *bld_base,
	struct lp_build_emit_data *emit_data)
{
	struct si_shader_context *ctx = si_shader_context(bld_base);

        LLVMValueRef tmp = ac_build_vote_any(&ctx->ac, emit_data->args[0]);
	emit_data->output[emit_data->chan] =
		LLVMBuildSExt(ctx->ac.builder, tmp, ctx->i32, "");
}

static void vote_eq_emit(
	const struct lp_build_tgsi_action *action,
	struct lp_build_tgsi_context *bld_base,
	struct lp_build_emit_data *emit_data)
{
	struct si_shader_context *ctx = si_shader_context(bld_base);

        LLVMValueRef tmp = ac_build_vote_eq(&ctx->ac, emit_data->args[0]);
	emit_data->output[emit_data->chan] =
		LLVMBuildSExt(ctx->ac.builder, tmp, ctx->i32, "");
}

static void ballot_emit(
	const struct lp_build_tgsi_action *action,
	struct lp_build_tgsi_context *bld_base,
	struct lp_build_emit_data *emit_data)
{
	struct si_shader_context *ctx = si_shader_context(bld_base);
	LLVMBuilderRef builder = ctx->ac.builder;
	LLVMValueRef tmp;

	tmp = lp_build_emit_fetch(bld_base, emit_data->inst, 0, TGSI_CHAN_X);
	tmp = ac_build_ballot(&ctx->ac, tmp);
	tmp = LLVMBuildBitCast(builder, tmp, ctx->v2i32, "");

	emit_data->output[0] = LLVMBuildExtractElement(builder, tmp, ctx->i32_0, "");
	emit_data->output[1] = LLVMBuildExtractElement(builder, tmp, ctx->i32_1, "");
}

static void read_lane_emit(
	const struct lp_build_tgsi_action *action,
	struct lp_build_tgsi_context *bld_base,
	struct lp_build_emit_data *emit_data)
{
	struct si_shader_context *ctx = si_shader_context(bld_base);

	if (emit_data->inst->Instruction.Opcode == TGSI_OPCODE_READ_INVOC) {
		emit_data->args[0] = lp_build_emit_fetch(bld_base, emit_data->inst,
							 0, emit_data->src_chan);

		/* Always read the source invocation (= lane) from the X channel. */
		emit_data->args[1] = lp_build_emit_fetch(bld_base, emit_data->inst,
							 1, TGSI_CHAN_X);
		emit_data->arg_count = 2;
	}

	/* We currently have no other way to prevent LLVM from lifting the icmp
	 * calls to a dominating basic block.
	 */
	ac_build_optimization_barrier(&ctx->ac, &emit_data->args[0]);

	for (unsigned i = 0; i < emit_data->arg_count; ++i)
		emit_data->args[i] = ac_to_integer(&ctx->ac, emit_data->args[i]);

	emit_data->output[emit_data->chan] =
		ac_build_intrinsic(&ctx->ac, action->intr_name,
				   ctx->i32, emit_data->args, emit_data->arg_count,
				   AC_FUNC_ATTR_READNONE |
				   AC_FUNC_ATTR_CONVERGENT);
}

static unsigned si_llvm_get_stream(struct lp_build_tgsi_context *bld_base,
				       struct lp_build_emit_data *emit_data)
{
	struct si_shader_context *ctx = si_shader_context(bld_base);
	struct tgsi_src_register src0 = emit_data->inst->Src[0].Register;
	LLVMValueRef imm;
	unsigned stream;

	assert(src0.File == TGSI_FILE_IMMEDIATE);

	imm = ctx->imms[src0.Index * TGSI_NUM_CHANNELS + src0.SwizzleX];
	stream = LLVMConstIntGetZExtValue(imm) & 0x3;
	return stream;
}

/* Emit one vertex from the geometry shader */
static void si_llvm_emit_vertex(struct ac_shader_abi *abi,
				unsigned stream,
				LLVMValueRef *addrs)
{
	struct si_shader_context *ctx = si_shader_context_from_abi(abi);
	struct tgsi_shader_info *info = &ctx->shader->selector->info;
	struct si_shader *shader = ctx->shader;
	struct lp_build_if_state if_state;
	LLVMValueRef soffset = LLVMGetParam(ctx->main_fn,
					    ctx->param_gs2vs_offset);
	LLVMValueRef gs_next_vertex;
	LLVMValueRef can_emit;
	unsigned chan, offset;
	int i;

	/* Write vertex attribute values to GSVS ring */
	gs_next_vertex = LLVMBuildLoad(ctx->ac.builder,
				       ctx->gs_next_vertex[stream],
				       "");

	/* If this thread has already emitted the declared maximum number of
	 * vertices, skip the write: excessive vertex emissions are not
	 * supposed to have any effect.
	 *
	 * If the shader has no writes to memory, kill it instead. This skips
	 * further memory loads and may allow LLVM to skip to the end
	 * altogether.
	 */
	can_emit = LLVMBuildICmp(ctx->ac.builder, LLVMIntULT, gs_next_vertex,
				 LLVMConstInt(ctx->i32,
					      shader->selector->gs_max_out_vertices, 0), "");

	bool use_kill = !info->writes_memory;
	if (use_kill) {
		ac_build_kill_if_false(&ctx->ac, can_emit);
	} else {
		lp_build_if(&if_state, &ctx->gallivm, can_emit);
	}

	offset = 0;
	for (i = 0; i < info->num_outputs; i++) {
		for (chan = 0; chan < 4; chan++) {
			if (!(info->output_usagemask[i] & (1 << chan)) ||
			    ((info->output_streams[i] >> (2 * chan)) & 3) != stream)
				continue;

			LLVMValueRef out_val = LLVMBuildLoad(ctx->ac.builder, addrs[4 * i + chan], "");
			LLVMValueRef voffset =
				LLVMConstInt(ctx->i32, offset *
					     shader->selector->gs_max_out_vertices, 0);
			offset++;

			voffset = LLVMBuildAdd(ctx->ac.builder, voffset, gs_next_vertex, "");
			voffset = LLVMBuildMul(ctx->ac.builder, voffset,
					       LLVMConstInt(ctx->i32, 4, 0), "");

			out_val = ac_to_integer(&ctx->ac, out_val);

			ac_build_buffer_store_dword(&ctx->ac,
						    ctx->gsvs_ring[stream],
						    out_val, 1,
						    voffset, soffset, 0,
						    1, 1, true, true);
		}
	}

	gs_next_vertex = LLVMBuildAdd(ctx->ac.builder, gs_next_vertex, ctx->i32_1, "");
	LLVMBuildStore(ctx->ac.builder, gs_next_vertex, ctx->gs_next_vertex[stream]);

	/* Signal vertex emission if vertex data was written. */
	if (offset) {
		ac_build_sendmsg(&ctx->ac, AC_SENDMSG_GS_OP_EMIT | AC_SENDMSG_GS | (stream << 8),
				 si_get_gs_wave_id(ctx));
	}

	if (!use_kill)
		lp_build_endif(&if_state);
}

/* Emit one vertex from the geometry shader */
static void si_tgsi_emit_vertex(
	const struct lp_build_tgsi_action *action,
	struct lp_build_tgsi_context *bld_base,
	struct lp_build_emit_data *emit_data)
{
	struct si_shader_context *ctx = si_shader_context(bld_base);
	unsigned stream = si_llvm_get_stream(bld_base, emit_data);

	si_llvm_emit_vertex(&ctx->abi, stream, ctx->outputs[0]);
}

/* Cut one primitive from the geometry shader */
static void si_llvm_emit_primitive(struct ac_shader_abi *abi,
				   unsigned stream)
{
	struct si_shader_context *ctx = si_shader_context_from_abi(abi);

	/* Signal primitive cut */
	ac_build_sendmsg(&ctx->ac, AC_SENDMSG_GS_OP_CUT | AC_SENDMSG_GS | (stream << 8),
			 si_get_gs_wave_id(ctx));
}

/* Cut one primitive from the geometry shader */
static void si_tgsi_emit_primitive(
	const struct lp_build_tgsi_action *action,
	struct lp_build_tgsi_context *bld_base,
	struct lp_build_emit_data *emit_data)
{
	struct si_shader_context *ctx = si_shader_context(bld_base);

	si_llvm_emit_primitive(&ctx->abi, si_llvm_get_stream(bld_base, emit_data));
}

static void si_llvm_emit_barrier(const struct lp_build_tgsi_action *action,
				 struct lp_build_tgsi_context *bld_base,
				 struct lp_build_emit_data *emit_data)
{
	struct si_shader_context *ctx = si_shader_context(bld_base);

	/* GFX6 only (thanks to a hw bug workaround):
	 * The real barrier instruction isn’t needed, because an entire patch
	 * always fits into a single wave.
	 */
	if (ctx->screen->info.chip_class == GFX6 &&
	    ctx->type == PIPE_SHADER_TESS_CTRL) {
		ac_build_waitcnt(&ctx->ac, LGKM_CNT & VM_CNT);
		return;
	}

	ac_build_s_barrier(&ctx->ac);
}

void si_create_function(struct si_shader_context *ctx,
			const char *name,
			LLVMTypeRef *returns, unsigned num_returns,
			struct si_function_info *fninfo,
			unsigned max_workgroup_size)
{
	int i;

	si_llvm_create_func(ctx, name, returns, num_returns,
			    fninfo->types, fninfo->num_params);
	ctx->return_value = LLVMGetUndef(ctx->return_type);

	for (i = 0; i < fninfo->num_sgpr_params; ++i) {
		LLVMValueRef P = LLVMGetParam(ctx->main_fn, i);

		/* The combination of:
		 * - noalias
		 * - dereferenceable
		 * - invariant.load
		 * allows the optimization passes to move loads and reduces
		 * SGPR spilling significantly.
		 */
		ac_add_function_attr(ctx->ac.context, ctx->main_fn, i + 1,
				     AC_FUNC_ATTR_INREG);

		if (LLVMGetTypeKind(LLVMTypeOf(P)) == LLVMPointerTypeKind) {
			ac_add_function_attr(ctx->ac.context, ctx->main_fn, i + 1,
					     AC_FUNC_ATTR_NOALIAS);
			ac_add_attr_dereferenceable(P, UINT64_MAX);
		}
	}

	for (i = 0; i < fninfo->num_params; ++i) {
		if (fninfo->assign[i])
			*fninfo->assign[i] = LLVMGetParam(ctx->main_fn, i);
	}

	if (ctx->screen->info.address32_hi) {
		ac_llvm_add_target_dep_function_attr(ctx->main_fn,
						     "amdgpu-32bit-address-high-bits",
						     ctx->screen->info.address32_hi);
	}

	if (max_workgroup_size) {
		ac_llvm_add_target_dep_function_attr(ctx->main_fn,
						     "amdgpu-max-work-group-size",
						     max_workgroup_size);
	}
	LLVMAddTargetDependentFunctionAttr(ctx->main_fn,
					   "no-signed-zeros-fp-math",
					   "true");

	if (ctx->screen->debug_flags & DBG(UNSAFE_MATH)) {
		/* These were copied from some LLVM test. */
		LLVMAddTargetDependentFunctionAttr(ctx->main_fn,
						   "less-precise-fpmad",
						   "true");
		LLVMAddTargetDependentFunctionAttr(ctx->main_fn,
						   "no-infs-fp-math",
						   "true");
		LLVMAddTargetDependentFunctionAttr(ctx->main_fn,
						   "no-nans-fp-math",
						   "true");
		LLVMAddTargetDependentFunctionAttr(ctx->main_fn,
						   "unsafe-fp-math",
						   "true");
	}
}

static void declare_streamout_params(struct si_shader_context *ctx,
				     struct pipe_stream_output_info *so,
				     struct si_function_info *fninfo)
{
	int i;

	/* Streamout SGPRs. */
	if (so->num_outputs) {
		if (ctx->type != PIPE_SHADER_TESS_EVAL)
			ctx->param_streamout_config = add_arg(fninfo, ARG_SGPR, ctx->ac.i32);
		else
			ctx->param_streamout_config = fninfo->num_params - 1;

		ctx->param_streamout_write_index = add_arg(fninfo, ARG_SGPR, ctx->ac.i32);
	}
	/* A streamout buffer offset is loaded if the stride is non-zero. */
	for (i = 0; i < 4; i++) {
		if (!so->stride[i])
			continue;

		ctx->param_streamout_offset[i] = add_arg(fninfo, ARG_SGPR, ctx->ac.i32);
	}
}

static unsigned si_get_max_workgroup_size(const struct si_shader *shader)
{
	switch (shader->selector->type) {
	case PIPE_SHADER_TESS_CTRL:
		/* Return this so that LLVM doesn't remove s_barrier
		 * instructions on chips where we use s_barrier. */
		return shader->selector->screen->info.chip_class >= GFX7 ? 128 : 64;

	case PIPE_SHADER_GEOMETRY:
		return shader->selector->screen->info.chip_class >= GFX9 ? 128 : 64;

	case PIPE_SHADER_COMPUTE:
		break; /* see below */

	default:
		return 0;
	}

	const unsigned *properties = shader->selector->info.properties;
	unsigned max_work_group_size =
	               properties[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH] *
	               properties[TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT] *
	               properties[TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH];

	if (!max_work_group_size) {
		/* This is a variable group size compute shader,
		 * compile it for the maximum possible group size.
		 */
		max_work_group_size = SI_MAX_VARIABLE_THREADS_PER_BLOCK;
	}
	return max_work_group_size;
}

static void declare_const_and_shader_buffers(struct si_shader_context *ctx,
					     struct si_function_info *fninfo,
					     bool assign_params)
{
	LLVMTypeRef const_shader_buf_type;

	if (ctx->shader->selector->info.const_buffers_declared == 1 &&
	    ctx->shader->selector->info.shader_buffers_declared == 0)
		const_shader_buf_type = ctx->f32;
	else
		const_shader_buf_type = ctx->v4i32;

	unsigned const_and_shader_buffers =
		add_arg(fninfo, ARG_SGPR,
			ac_array_in_const32_addr_space(const_shader_buf_type));

	if (assign_params)
		ctx->param_const_and_shader_buffers = const_and_shader_buffers;
}

static void declare_samplers_and_images(struct si_shader_context *ctx,
					struct si_function_info *fninfo,
					bool assign_params)
{
	unsigned samplers_and_images =
		add_arg(fninfo, ARG_SGPR,
			ac_array_in_const32_addr_space(ctx->v8i32));

	if (assign_params)
		ctx->param_samplers_and_images = samplers_and_images;
}

static void declare_per_stage_desc_pointers(struct si_shader_context *ctx,
					    struct si_function_info *fninfo,
					    bool assign_params)
{
	declare_const_and_shader_buffers(ctx, fninfo, assign_params);
	declare_samplers_and_images(ctx, fninfo, assign_params);
}

static void declare_global_desc_pointers(struct si_shader_context *ctx,
					 struct si_function_info *fninfo)
{
	ctx->param_rw_buffers = add_arg(fninfo, ARG_SGPR,
		ac_array_in_const32_addr_space(ctx->v4i32));
	ctx->param_bindless_samplers_and_images = add_arg(fninfo, ARG_SGPR,
		ac_array_in_const32_addr_space(ctx->v8i32));
}

static void declare_vs_specific_input_sgprs(struct si_shader_context *ctx,
					    struct si_function_info *fninfo)
{
	ctx->param_vs_state_bits = add_arg(fninfo, ARG_SGPR, ctx->i32);
	add_arg_assign(fninfo, ARG_SGPR, ctx->i32, &ctx->abi.base_vertex);
	add_arg_assign(fninfo, ARG_SGPR, ctx->i32, &ctx->abi.start_instance);
	add_arg_assign(fninfo, ARG_SGPR, ctx->i32, &ctx->abi.draw_id);
}

static void declare_vs_input_vgprs(struct si_shader_context *ctx,
				   struct si_function_info *fninfo,
				   unsigned *num_prolog_vgprs)
{
	struct si_shader *shader = ctx->shader;

	add_arg_assign(fninfo, ARG_VGPR, ctx->i32, &ctx->abi.vertex_id);
	if (shader->key.as_ls) {
		ctx->param_rel_auto_id = add_arg(fninfo, ARG_VGPR, ctx->i32);
		add_arg_assign(fninfo, ARG_VGPR, ctx->i32, &ctx->abi.instance_id);
	} else {
		add_arg_assign(fninfo, ARG_VGPR, ctx->i32, &ctx->abi.instance_id);
		ctx->param_vs_prim_id = add_arg(fninfo, ARG_VGPR, ctx->i32);
	}
	add_arg(fninfo, ARG_VGPR, ctx->i32); /* unused */

	if (!shader->is_gs_copy_shader) {
		/* Vertex load indices. */
		ctx->param_vertex_index0 = fninfo->num_params;
		for (unsigned i = 0; i < shader->selector->info.num_inputs; i++)
			add_arg(fninfo, ARG_VGPR, ctx->i32);
		*num_prolog_vgprs += shader->selector->info.num_inputs;
	}
}

static void declare_vs_blit_inputs(struct si_shader_context *ctx,
				   struct si_function_info *fninfo,
				   unsigned vs_blit_property)
{
	ctx->param_vs_blit_inputs = fninfo->num_params;
	add_arg(fninfo, ARG_SGPR, ctx->i32); /* i16 x1, y1 */
	add_arg(fninfo, ARG_SGPR, ctx->i32); /* i16 x2, y2 */
	add_arg(fninfo, ARG_SGPR, ctx->f32); /* depth */

	if (vs_blit_property == SI_VS_BLIT_SGPRS_POS_COLOR) {
		add_arg(fninfo, ARG_SGPR, ctx->f32); /* color0 */
		add_arg(fninfo, ARG_SGPR, ctx->f32); /* color1 */
		add_arg(fninfo, ARG_SGPR, ctx->f32); /* color2 */
		add_arg(fninfo, ARG_SGPR, ctx->f32); /* color3 */
	} else if (vs_blit_property == SI_VS_BLIT_SGPRS_POS_TEXCOORD) {
		add_arg(fninfo, ARG_SGPR, ctx->f32); /* texcoord.x1 */
		add_arg(fninfo, ARG_SGPR, ctx->f32); /* texcoord.y1 */
		add_arg(fninfo, ARG_SGPR, ctx->f32); /* texcoord.x2 */
		add_arg(fninfo, ARG_SGPR, ctx->f32); /* texcoord.y2 */
		add_arg(fninfo, ARG_SGPR, ctx->f32); /* texcoord.z */
		add_arg(fninfo, ARG_SGPR, ctx->f32); /* texcoord.w */
	}
}

static void declare_tes_input_vgprs(struct si_shader_context *ctx,
				    struct si_function_info *fninfo)
{
	ctx->param_tes_u = add_arg(fninfo, ARG_VGPR, ctx->f32);
	ctx->param_tes_v = add_arg(fninfo, ARG_VGPR, ctx->f32);
	ctx->param_tes_rel_patch_id = add_arg(fninfo, ARG_VGPR, ctx->i32);
	add_arg_assign(fninfo, ARG_VGPR, ctx->i32, &ctx->abi.tes_patch_id);
}

enum {
	/* Convenient merged shader definitions. */
	SI_SHADER_MERGED_VERTEX_TESSCTRL = PIPE_SHADER_TYPES,
	SI_SHADER_MERGED_VERTEX_OR_TESSEVAL_GEOMETRY,
};

static void create_function(struct si_shader_context *ctx)
{
	struct si_shader *shader = ctx->shader;
	struct si_function_info fninfo;
	LLVMTypeRef returns[16+32*4];
	unsigned i, num_return_sgprs;
	unsigned num_returns = 0;
	unsigned num_prolog_vgprs = 0;
	unsigned type = ctx->type;
	unsigned vs_blit_property =
		shader->selector->info.properties[TGSI_PROPERTY_VS_BLIT_SGPRS];

	si_init_function_info(&fninfo);

	/* Set MERGED shaders. */
	if (ctx->screen->info.chip_class >= GFX9) {
		if (shader->key.as_ls || type == PIPE_SHADER_TESS_CTRL)
			type = SI_SHADER_MERGED_VERTEX_TESSCTRL; /* LS or HS */
		else if (shader->key.as_es || type == PIPE_SHADER_GEOMETRY)
			type = SI_SHADER_MERGED_VERTEX_OR_TESSEVAL_GEOMETRY;
	}

	LLVMTypeRef v3i32 = LLVMVectorType(ctx->i32, 3);

	switch (type) {
	case PIPE_SHADER_VERTEX:
		declare_global_desc_pointers(ctx, &fninfo);

		if (vs_blit_property) {
			declare_vs_blit_inputs(ctx, &fninfo, vs_blit_property);

			/* VGPRs */
			declare_vs_input_vgprs(ctx, &fninfo, &num_prolog_vgprs);
			break;
		}

		declare_per_stage_desc_pointers(ctx, &fninfo, true);
		declare_vs_specific_input_sgprs(ctx, &fninfo);
		ctx->param_vertex_buffers = add_arg(&fninfo, ARG_SGPR,
			ac_array_in_const32_addr_space(ctx->v4i32));

		if (shader->key.as_es) {
			ctx->param_es2gs_offset = add_arg(&fninfo, ARG_SGPR, ctx->i32);
		} else if (shader->key.as_ls) {
			/* no extra parameters */
		} else {
			if (shader->is_gs_copy_shader) {
				fninfo.num_params = ctx->param_vs_state_bits + 1;
				fninfo.num_sgpr_params = fninfo.num_params;
			}

			/* The locations of the other parameters are assigned dynamically. */
			declare_streamout_params(ctx, &shader->selector->so,
						 &fninfo);
		}

		/* VGPRs */
		declare_vs_input_vgprs(ctx, &fninfo, &num_prolog_vgprs);

		/* Return values */
		if (shader->key.opt.vs_as_prim_discard_cs) {
			for (i = 0; i < 4; i++)
				returns[num_returns++] = ctx->f32; /* VGPRs */
		}
		break;

	case PIPE_SHADER_TESS_CTRL: /* GFX6-GFX8 */
		declare_global_desc_pointers(ctx, &fninfo);
		declare_per_stage_desc_pointers(ctx, &fninfo, true);
		ctx->param_tcs_offchip_layout = add_arg(&fninfo, ARG_SGPR, ctx->i32);
		ctx->param_tcs_out_lds_offsets = add_arg(&fninfo, ARG_SGPR, ctx->i32);
		ctx->param_tcs_out_lds_layout = add_arg(&fninfo, ARG_SGPR, ctx->i32);
		ctx->param_vs_state_bits = add_arg(&fninfo, ARG_SGPR, ctx->i32);
		ctx->param_tcs_offchip_offset = add_arg(&fninfo, ARG_SGPR, ctx->i32);
		ctx->param_tcs_factor_offset = add_arg(&fninfo, ARG_SGPR, ctx->i32);

		/* VGPRs */
		add_arg_assign(&fninfo, ARG_VGPR, ctx->i32, &ctx->abi.tcs_patch_id);
		add_arg_assign(&fninfo, ARG_VGPR, ctx->i32, &ctx->abi.tcs_rel_ids);

		/* param_tcs_offchip_offset and param_tcs_factor_offset are
		 * placed after the user SGPRs.
		 */
		for (i = 0; i < GFX6_TCS_NUM_USER_SGPR + 2; i++)
			returns[num_returns++] = ctx->i32; /* SGPRs */
		for (i = 0; i < 11; i++)
			returns[num_returns++] = ctx->f32; /* VGPRs */
		break;

	case SI_SHADER_MERGED_VERTEX_TESSCTRL:
		/* Merged stages have 8 system SGPRs at the beginning. */
		/* SPI_SHADER_USER_DATA_ADDR_LO/HI_HS */
		declare_per_stage_desc_pointers(ctx, &fninfo,
						ctx->type == PIPE_SHADER_TESS_CTRL);
		ctx->param_tcs_offchip_offset = add_arg(&fninfo, ARG_SGPR, ctx->i32);
		ctx->param_merged_wave_info = add_arg(&fninfo, ARG_SGPR, ctx->i32);
		ctx->param_tcs_factor_offset = add_arg(&fninfo, ARG_SGPR, ctx->i32);
		ctx->param_merged_scratch_offset = add_arg(&fninfo, ARG_SGPR, ctx->i32);
		add_arg(&fninfo, ARG_SGPR, ctx->i32); /* unused */
		add_arg(&fninfo, ARG_SGPR, ctx->i32); /* unused */

		declare_global_desc_pointers(ctx, &fninfo);
		declare_per_stage_desc_pointers(ctx, &fninfo,
						ctx->type == PIPE_SHADER_VERTEX);
		declare_vs_specific_input_sgprs(ctx, &fninfo);

		ctx->param_tcs_offchip_layout = add_arg(&fninfo, ARG_SGPR, ctx->i32);
		ctx->param_tcs_out_lds_offsets = add_arg(&fninfo, ARG_SGPR, ctx->i32);
		ctx->param_tcs_out_lds_layout = add_arg(&fninfo, ARG_SGPR, ctx->i32);
		ctx->param_vertex_buffers = add_arg(&fninfo, ARG_SGPR,
			ac_array_in_const32_addr_space(ctx->v4i32));

		/* VGPRs (first TCS, then VS) */
		add_arg_assign(&fninfo, ARG_VGPR, ctx->i32, &ctx->abi.tcs_patch_id);
		add_arg_assign(&fninfo, ARG_VGPR, ctx->i32, &ctx->abi.tcs_rel_ids);

		if (ctx->type == PIPE_SHADER_VERTEX) {
			declare_vs_input_vgprs(ctx, &fninfo,
					       &num_prolog_vgprs);

			/* LS return values are inputs to the TCS main shader part. */
			for (i = 0; i < 8 + GFX9_TCS_NUM_USER_SGPR; i++)
				returns[num_returns++] = ctx->i32; /* SGPRs */
			for (i = 0; i < 2; i++)
				returns[num_returns++] = ctx->f32; /* VGPRs */
		} else {
			/* TCS return values are inputs to the TCS epilog.
			 *
			 * param_tcs_offchip_offset, param_tcs_factor_offset,
			 * param_tcs_offchip_layout, and param_rw_buffers
			 * should be passed to the epilog.
			 */
			for (i = 0; i <= 8 + GFX9_SGPR_TCS_OUT_LAYOUT; i++)
				returns[num_returns++] = ctx->i32; /* SGPRs */
			for (i = 0; i < 11; i++)
				returns[num_returns++] = ctx->f32; /* VGPRs */
		}
		break;

	case SI_SHADER_MERGED_VERTEX_OR_TESSEVAL_GEOMETRY:
		/* Merged stages have 8 system SGPRs at the beginning. */
		/* SPI_SHADER_USER_DATA_ADDR_LO/HI_GS */
		declare_per_stage_desc_pointers(ctx, &fninfo,
						ctx->type == PIPE_SHADER_GEOMETRY);
		ctx->param_gs2vs_offset = add_arg(&fninfo, ARG_SGPR, ctx->i32);
		ctx->param_merged_wave_info = add_arg(&fninfo, ARG_SGPR, ctx->i32);
		ctx->param_tcs_offchip_offset = add_arg(&fninfo, ARG_SGPR, ctx->i32);
		ctx->param_merged_scratch_offset = add_arg(&fninfo, ARG_SGPR, ctx->i32);
		add_arg(&fninfo, ARG_SGPR, ctx->i32); /* unused (SPI_SHADER_PGM_LO/HI_GS << 8) */
		add_arg(&fninfo, ARG_SGPR, ctx->i32); /* unused (SPI_SHADER_PGM_LO/HI_GS >> 24) */

		declare_global_desc_pointers(ctx, &fninfo);
		declare_per_stage_desc_pointers(ctx, &fninfo,
						(ctx->type == PIPE_SHADER_VERTEX ||
						 ctx->type == PIPE_SHADER_TESS_EVAL));
		if (ctx->type == PIPE_SHADER_VERTEX) {
			declare_vs_specific_input_sgprs(ctx, &fninfo);
		} else {
			ctx->param_vs_state_bits = add_arg(&fninfo, ARG_SGPR, ctx->i32);
			ctx->param_tcs_offchip_layout = add_arg(&fninfo, ARG_SGPR, ctx->i32);
			ctx->param_tes_offchip_addr = add_arg(&fninfo, ARG_SGPR, ctx->i32);
			/* Declare as many input SGPRs as the VS has. */
		}

		if (ctx->type == PIPE_SHADER_VERTEX) {
			ctx->param_vertex_buffers = add_arg(&fninfo, ARG_SGPR,
				ac_array_in_const32_addr_space(ctx->v4i32));
		}

		/* VGPRs (first GS, then VS/TES) */
		ctx->param_gs_vtx01_offset = add_arg(&fninfo, ARG_VGPR, ctx->i32);
		ctx->param_gs_vtx23_offset = add_arg(&fninfo, ARG_VGPR, ctx->i32);
		add_arg_assign(&fninfo, ARG_VGPR, ctx->i32, &ctx->abi.gs_prim_id);
		add_arg_assign(&fninfo, ARG_VGPR, ctx->i32, &ctx->abi.gs_invocation_id);
		ctx->param_gs_vtx45_offset = add_arg(&fninfo, ARG_VGPR, ctx->i32);

		if (ctx->type == PIPE_SHADER_VERTEX) {
			declare_vs_input_vgprs(ctx, &fninfo,
					       &num_prolog_vgprs);
		} else if (ctx->type == PIPE_SHADER_TESS_EVAL) {
			declare_tes_input_vgprs(ctx, &fninfo);
		}

		if (ctx->type == PIPE_SHADER_VERTEX ||
		    ctx->type == PIPE_SHADER_TESS_EVAL) {
			unsigned num_user_sgprs;

			if (ctx->type == PIPE_SHADER_VERTEX)
				num_user_sgprs = GFX9_VSGS_NUM_USER_SGPR;
			else
				num_user_sgprs = GFX9_TESGS_NUM_USER_SGPR;

			/* ES return values are inputs to GS. */
			for (i = 0; i < 8 + num_user_sgprs; i++)
				returns[num_returns++] = ctx->i32; /* SGPRs */
			for (i = 0; i < 5; i++)
				returns[num_returns++] = ctx->f32; /* VGPRs */
		}
		break;

	case PIPE_SHADER_TESS_EVAL:
		declare_global_desc_pointers(ctx, &fninfo);
		declare_per_stage_desc_pointers(ctx, &fninfo, true);
		ctx->param_vs_state_bits = add_arg(&fninfo, ARG_SGPR, ctx->i32);
		ctx->param_tcs_offchip_layout = add_arg(&fninfo, ARG_SGPR, ctx->i32);
		ctx->param_tes_offchip_addr = add_arg(&fninfo, ARG_SGPR, ctx->i32);

		if (shader->key.as_es) {
			ctx->param_tcs_offchip_offset = add_arg(&fninfo, ARG_SGPR, ctx->i32);
			add_arg(&fninfo, ARG_SGPR, ctx->i32);
			ctx->param_es2gs_offset = add_arg(&fninfo, ARG_SGPR, ctx->i32);
		} else {
			add_arg(&fninfo, ARG_SGPR, ctx->i32);
			declare_streamout_params(ctx, &shader->selector->so,
						 &fninfo);
			ctx->param_tcs_offchip_offset = add_arg(&fninfo, ARG_SGPR, ctx->i32);
		}

		/* VGPRs */
		declare_tes_input_vgprs(ctx, &fninfo);
		break;

	case PIPE_SHADER_GEOMETRY:
		declare_global_desc_pointers(ctx, &fninfo);
		declare_per_stage_desc_pointers(ctx, &fninfo, true);
		ctx->param_gs2vs_offset = add_arg(&fninfo, ARG_SGPR, ctx->i32);
		ctx->param_gs_wave_id = add_arg(&fninfo, ARG_SGPR, ctx->i32);

		/* VGPRs */
		add_arg_assign(&fninfo, ARG_VGPR, ctx->i32, &ctx->gs_vtx_offset[0]);
		add_arg_assign(&fninfo, ARG_VGPR, ctx->i32, &ctx->gs_vtx_offset[1]);
		add_arg_assign(&fninfo, ARG_VGPR, ctx->i32, &ctx->abi.gs_prim_id);
		add_arg_assign(&fninfo, ARG_VGPR, ctx->i32, &ctx->gs_vtx_offset[2]);
		add_arg_assign(&fninfo, ARG_VGPR, ctx->i32, &ctx->gs_vtx_offset[3]);
		add_arg_assign(&fninfo, ARG_VGPR, ctx->i32, &ctx->gs_vtx_offset[4]);
		add_arg_assign(&fninfo, ARG_VGPR, ctx->i32, &ctx->gs_vtx_offset[5]);
		add_arg_assign(&fninfo, ARG_VGPR, ctx->i32, &ctx->abi.gs_invocation_id);
		break;

	case PIPE_SHADER_FRAGMENT:
		declare_global_desc_pointers(ctx, &fninfo);
		declare_per_stage_desc_pointers(ctx, &fninfo, true);
		add_arg_checked(&fninfo, ARG_SGPR, ctx->f32, SI_PARAM_ALPHA_REF);
		add_arg_assign_checked(&fninfo, ARG_SGPR, ctx->i32,
				       &ctx->abi.prim_mask, SI_PARAM_PRIM_MASK);

		add_arg_checked(&fninfo, ARG_VGPR, ctx->v2i32, SI_PARAM_PERSP_SAMPLE);
		add_arg_checked(&fninfo, ARG_VGPR, ctx->v2i32, SI_PARAM_PERSP_CENTER);
		add_arg_checked(&fninfo, ARG_VGPR, ctx->v2i32, SI_PARAM_PERSP_CENTROID);
		add_arg_checked(&fninfo, ARG_VGPR, v3i32, SI_PARAM_PERSP_PULL_MODEL);
		add_arg_checked(&fninfo, ARG_VGPR, ctx->v2i32, SI_PARAM_LINEAR_SAMPLE);
		add_arg_checked(&fninfo, ARG_VGPR, ctx->v2i32, SI_PARAM_LINEAR_CENTER);
		add_arg_checked(&fninfo, ARG_VGPR, ctx->v2i32, SI_PARAM_LINEAR_CENTROID);
		add_arg_checked(&fninfo, ARG_VGPR, ctx->f32, SI_PARAM_LINE_STIPPLE_TEX);
		add_arg_assign_checked(&fninfo, ARG_VGPR, ctx->f32,
				       &ctx->abi.frag_pos[0], SI_PARAM_POS_X_FLOAT);
		add_arg_assign_checked(&fninfo, ARG_VGPR, ctx->f32,
				       &ctx->abi.frag_pos[1], SI_PARAM_POS_Y_FLOAT);
		add_arg_assign_checked(&fninfo, ARG_VGPR, ctx->f32,
				       &ctx->abi.frag_pos[2], SI_PARAM_POS_Z_FLOAT);
		add_arg_assign_checked(&fninfo, ARG_VGPR, ctx->f32,
				       &ctx->abi.frag_pos[3], SI_PARAM_POS_W_FLOAT);
		add_arg_assign_checked(&fninfo, ARG_VGPR, ctx->i32,
				       &ctx->abi.front_face, SI_PARAM_FRONT_FACE);
		shader->info.face_vgpr_index = 20;
		add_arg_assign_checked(&fninfo, ARG_VGPR, ctx->i32,
				       &ctx->abi.ancillary, SI_PARAM_ANCILLARY);
		shader->info.ancillary_vgpr_index = 21;
		add_arg_assign_checked(&fninfo, ARG_VGPR, ctx->f32,
				       &ctx->abi.sample_coverage, SI_PARAM_SAMPLE_COVERAGE);
		add_arg_checked(&fninfo, ARG_VGPR, ctx->i32, SI_PARAM_POS_FIXED_PT);

		/* Color inputs from the prolog. */
		if (shader->selector->info.colors_read) {
			unsigned num_color_elements =
				util_bitcount(shader->selector->info.colors_read);

			assert(fninfo.num_params + num_color_elements <= ARRAY_SIZE(fninfo.types));
			for (i = 0; i < num_color_elements; i++)
				add_arg(&fninfo, ARG_VGPR, ctx->f32);

			num_prolog_vgprs += num_color_elements;
		}

		/* Outputs for the epilog. */
		num_return_sgprs = SI_SGPR_ALPHA_REF + 1;
		num_returns =
			num_return_sgprs +
			util_bitcount(shader->selector->info.colors_written) * 4 +
			shader->selector->info.writes_z +
			shader->selector->info.writes_stencil +
			shader->selector->info.writes_samplemask +
			1 /* SampleMaskIn */;

		num_returns = MAX2(num_returns,
				   num_return_sgprs +
				   PS_EPILOG_SAMPLEMASK_MIN_LOC + 1);

		for (i = 0; i < num_return_sgprs; i++)
			returns[i] = ctx->i32;
		for (; i < num_returns; i++)
			returns[i] = ctx->f32;
		break;

	case PIPE_SHADER_COMPUTE:
		declare_global_desc_pointers(ctx, &fninfo);
		declare_per_stage_desc_pointers(ctx, &fninfo, true);
		if (shader->selector->info.uses_grid_size)
			add_arg_assign(&fninfo, ARG_SGPR, v3i32, &ctx->abi.num_work_groups);
		if (shader->selector->info.uses_block_size &&
		    shader->selector->info.properties[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH] == 0)
			ctx->param_block_size = add_arg(&fninfo, ARG_SGPR, v3i32);

		unsigned cs_user_data_dwords =
			shader->selector->info.properties[TGSI_PROPERTY_CS_USER_DATA_DWORDS];
		if (cs_user_data_dwords) {
			ctx->param_cs_user_data = add_arg(&fninfo, ARG_SGPR,
							  LLVMVectorType(ctx->i32, cs_user_data_dwords));
		}

		for (i = 0; i < 3; i++) {
			ctx->abi.workgroup_ids[i] = NULL;
			if (shader->selector->info.uses_block_id[i])
				add_arg_assign(&fninfo, ARG_SGPR, ctx->i32, &ctx->abi.workgroup_ids[i]);
		}

		add_arg_assign(&fninfo, ARG_VGPR, v3i32, &ctx->abi.local_invocation_ids);
		break;
	default:
		assert(0 && "unimplemented shader");
		return;
	}

	si_create_function(ctx, "main", returns, num_returns, &fninfo,
			   si_get_max_workgroup_size(shader));

	/* Reserve register locations for VGPR inputs the PS prolog may need. */
	if (ctx->type == PIPE_SHADER_FRAGMENT && !ctx->shader->is_monolithic) {
		ac_llvm_add_target_dep_function_attr(ctx->main_fn,
						     "InitialPSInputAddr",
						     S_0286D0_PERSP_SAMPLE_ENA(1) |
						     S_0286D0_PERSP_CENTER_ENA(1) |
						     S_0286D0_PERSP_CENTROID_ENA(1) |
						     S_0286D0_LINEAR_SAMPLE_ENA(1) |
						     S_0286D0_LINEAR_CENTER_ENA(1) |
						     S_0286D0_LINEAR_CENTROID_ENA(1) |
						     S_0286D0_FRONT_FACE_ENA(1) |
						     S_0286D0_ANCILLARY_ENA(1) |
						     S_0286D0_POS_FIXED_PT_ENA(1));
	}

	shader->info.num_input_sgprs = 0;
	shader->info.num_input_vgprs = 0;

	for (i = 0; i < fninfo.num_sgpr_params; ++i)
		shader->info.num_input_sgprs += ac_get_type_size(fninfo.types[i]) / 4;

	for (; i < fninfo.num_params; ++i)
		shader->info.num_input_vgprs += ac_get_type_size(fninfo.types[i]) / 4;

	assert(shader->info.num_input_vgprs >= num_prolog_vgprs);
	shader->info.num_input_vgprs -= num_prolog_vgprs;

	if (shader->key.as_ls ||
	    ctx->type == PIPE_SHADER_TESS_CTRL ||
	    /* GFX9 has the ESGS ring buffer in LDS. */
	    type == SI_SHADER_MERGED_VERTEX_OR_TESSEVAL_GEOMETRY)
		ac_declare_lds_as_pointer(&ctx->ac);
}

/**
 * Load ESGS and GSVS ring buffer resource descriptors and save the variables
 * for later use.
 */
static void preload_ring_buffers(struct si_shader_context *ctx)
{
	LLVMBuilderRef builder = ctx->ac.builder;

	LLVMValueRef buf_ptr = LLVMGetParam(ctx->main_fn,
					    ctx->param_rw_buffers);

	if (ctx->screen->info.chip_class <= GFX8 &&
	    (ctx->shader->key.as_es || ctx->type == PIPE_SHADER_GEOMETRY)) {
		unsigned ring =
			ctx->type == PIPE_SHADER_GEOMETRY ? SI_GS_RING_ESGS
							     : SI_ES_RING_ESGS;
		LLVMValueRef offset = LLVMConstInt(ctx->i32, ring, 0);

		ctx->esgs_ring =
			ac_build_load_to_sgpr(&ctx->ac, buf_ptr, offset);
	}

	if (ctx->shader->is_gs_copy_shader) {
		LLVMValueRef offset = LLVMConstInt(ctx->i32, SI_RING_GSVS, 0);

		ctx->gsvs_ring[0] =
			ac_build_load_to_sgpr(&ctx->ac, buf_ptr, offset);
	} else if (ctx->type == PIPE_SHADER_GEOMETRY) {
		const struct si_shader_selector *sel = ctx->shader->selector;
		LLVMValueRef offset = LLVMConstInt(ctx->i32, SI_RING_GSVS, 0);
		LLVMValueRef base_ring;

		base_ring = ac_build_load_to_sgpr(&ctx->ac, buf_ptr, offset);

		/* The conceptual layout of the GSVS ring is
		 *   v0c0 .. vLv0 v0c1 .. vLc1 ..
		 * but the real memory layout is swizzled across
		 * threads:
		 *   t0v0c0 .. t15v0c0 t0v1c0 .. t15v1c0 ... t15vLcL
		 *   t16v0c0 ..
		 * Override the buffer descriptor accordingly.
		 */
		LLVMTypeRef v2i64 = LLVMVectorType(ctx->i64, 2);
		uint64_t stream_offset = 0;

		for (unsigned stream = 0; stream < 4; ++stream) {
			unsigned num_components;
			unsigned stride;
			unsigned num_records;
			LLVMValueRef ring, tmp;

			num_components = sel->info.num_stream_output_components[stream];
			if (!num_components)
				continue;

			stride = 4 * num_components * sel->gs_max_out_vertices;

			/* Limit on the stride field for <= GFX7. */
			assert(stride < (1 << 14));

			num_records = 64;

			ring = LLVMBuildBitCast(builder, base_ring, v2i64, "");
			tmp = LLVMBuildExtractElement(builder, ring, ctx->i32_0, "");
			tmp = LLVMBuildAdd(builder, tmp,
					   LLVMConstInt(ctx->i64,
							stream_offset, 0), "");
			stream_offset += stride * 64;

			ring = LLVMBuildInsertElement(builder, ring, tmp, ctx->i32_0, "");
			ring = LLVMBuildBitCast(builder, ring, ctx->v4i32, "");
			tmp = LLVMBuildExtractElement(builder, ring, ctx->i32_1, "");
			tmp = LLVMBuildOr(builder, tmp,
				LLVMConstInt(ctx->i32,
					     S_008F04_STRIDE(stride) |
					     S_008F04_SWIZZLE_ENABLE(1), 0), "");
			ring = LLVMBuildInsertElement(builder, ring, tmp, ctx->i32_1, "");
			ring = LLVMBuildInsertElement(builder, ring,
					LLVMConstInt(ctx->i32, num_records, 0),
					LLVMConstInt(ctx->i32, 2, 0), "");
			ring = LLVMBuildInsertElement(builder, ring,
				LLVMConstInt(ctx->i32,
					     S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) |
					     S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
					     S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) |
					     S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W) |
					     S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) |
					     S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32) |
					     S_008F0C_ELEMENT_SIZE(1) | /* element_size = 4 (bytes) */
					     S_008F0C_INDEX_STRIDE(1) | /* index_stride = 16 (elements) */
					     S_008F0C_ADD_TID_ENABLE(1),
					     0),
				LLVMConstInt(ctx->i32, 3, 0), "");

			ctx->gsvs_ring[stream] = ring;
		}
	} else if (ctx->type == PIPE_SHADER_TESS_EVAL) {
		ctx->tess_offchip_ring = get_tess_ring_descriptor(ctx, TESS_OFFCHIP_RING_TES);
	}
}

static void si_llvm_emit_polygon_stipple(struct si_shader_context *ctx,
					 LLVMValueRef param_rw_buffers,
					 unsigned param_pos_fixed_pt)
{
	LLVMBuilderRef builder = ctx->ac.builder;
	LLVMValueRef slot, desc, offset, row, bit, address[2];

	/* Use the fixed-point gl_FragCoord input.
	 * Since the stipple pattern is 32x32 and it repeats, just get 5 bits
	 * per coordinate to get the repeating effect.
	 */
	address[0] = si_unpack_param(ctx, param_pos_fixed_pt, 0, 5);
	address[1] = si_unpack_param(ctx, param_pos_fixed_pt, 16, 5);

	/* Load the buffer descriptor. */
	slot = LLVMConstInt(ctx->i32, SI_PS_CONST_POLY_STIPPLE, 0);
	desc = ac_build_load_to_sgpr(&ctx->ac, param_rw_buffers, slot);

	/* The stipple pattern is 32x32, each row has 32 bits. */
	offset = LLVMBuildMul(builder, address[1],
			      LLVMConstInt(ctx->i32, 4, 0), "");
	row = buffer_load_const(ctx, desc, offset);
	row = ac_to_integer(&ctx->ac, row);
	bit = LLVMBuildLShr(builder, row, address[0], "");
	bit = LLVMBuildTrunc(builder, bit, ctx->i1, "");
	ac_build_kill_if_false(&ctx->ac, bit);
}

void si_shader_binary_read_config(struct ac_shader_binary *binary,
				  struct si_shader_config *conf,
				  unsigned symbol_offset)
{
	unsigned i;
	const unsigned char *config =
		ac_shader_binary_config_start(binary, symbol_offset);
	bool really_needs_scratch = false;

	/* LLVM adds SGPR spills to the scratch size.
	 * Find out if we really need the scratch buffer.
	 */
	for (i = 0; i < binary->reloc_count; i++) {
		const struct ac_shader_reloc *reloc = &binary->relocs[i];

		if (!strcmp(scratch_rsrc_dword0_symbol, reloc->name) ||
		    !strcmp(scratch_rsrc_dword1_symbol, reloc->name)) {
			really_needs_scratch = true;
			break;
		}
	}

	/* XXX: We may be able to emit some of these values directly rather than
	 * extracting fields to be emitted later.
	 */

	for (i = 0; i < binary->config_size_per_symbol; i+= 8) {
		unsigned reg = util_le32_to_cpu(*(uint32_t*)(config + i));
		unsigned value = util_le32_to_cpu(*(uint32_t*)(config + i + 4));
		switch (reg) {
		case R_00B028_SPI_SHADER_PGM_RSRC1_PS:
		case R_00B128_SPI_SHADER_PGM_RSRC1_VS:
		case R_00B228_SPI_SHADER_PGM_RSRC1_GS:
		case R_00B428_SPI_SHADER_PGM_RSRC1_HS:
		case R_00B848_COMPUTE_PGM_RSRC1:
			conf->num_sgprs = MAX2(conf->num_sgprs, (G_00B028_SGPRS(value) + 1) * 8);
			conf->num_vgprs = MAX2(conf->num_vgprs, (G_00B028_VGPRS(value) + 1) * 4);
			conf->float_mode =  G_00B028_FLOAT_MODE(value);
			conf->rsrc1 = value;
			break;
		case R_00B02C_SPI_SHADER_PGM_RSRC2_PS:
			conf->lds_size = MAX2(conf->lds_size, G_00B02C_EXTRA_LDS_SIZE(value));
			break;
		case R_00B84C_COMPUTE_PGM_RSRC2:
			conf->lds_size = MAX2(conf->lds_size, G_00B84C_LDS_SIZE(value));
			conf->rsrc2 = value;
			break;
		case R_0286CC_SPI_PS_INPUT_ENA:
			conf->spi_ps_input_ena = value;
			break;
		case R_0286D0_SPI_PS_INPUT_ADDR:
			conf->spi_ps_input_addr = value;
			break;
		case R_0286E8_SPI_TMPRING_SIZE:
		case R_00B860_COMPUTE_TMPRING_SIZE:
			/* WAVESIZE is in units of 256 dwords. */
			if (really_needs_scratch)
				conf->scratch_bytes_per_wave =
					G_00B860_WAVESIZE(value) * 256 * 4;
			break;
		case 0x4: /* SPILLED_SGPRS */
			conf->spilled_sgprs = value;
			break;
		case 0x8: /* SPILLED_VGPRS */
			conf->spilled_vgprs = value;
			break;
		default:
			{
				static bool printed;

				if (!printed) {
					fprintf(stderr, "Warning: LLVM emitted unknown "
						"config register: 0x%x\n", reg);
					printed = true;
				}
			}
			break;
		}
	}

	if (!conf->spi_ps_input_addr)
		conf->spi_ps_input_addr = conf->spi_ps_input_ena;
}

void si_shader_apply_scratch_relocs(struct si_shader *shader,
				    uint64_t scratch_va)
{
	unsigned i;
	uint32_t scratch_rsrc_dword0 = scratch_va;
	uint32_t scratch_rsrc_dword1 =
		S_008F04_BASE_ADDRESS_HI(scratch_va >> 32);

	/* Enable scratch coalescing. */
	scratch_rsrc_dword1 |= S_008F04_SWIZZLE_ENABLE(1);

	for (i = 0 ; i < shader->binary.reloc_count; i++) {
		const struct ac_shader_reloc *reloc =
					&shader->binary.relocs[i];
		if (!strcmp(scratch_rsrc_dword0_symbol, reloc->name)) {
			util_memcpy_cpu_to_le32(shader->binary.code + reloc->offset,
			&scratch_rsrc_dword0, 4);
		} else if (!strcmp(scratch_rsrc_dword1_symbol, reloc->name)) {
			util_memcpy_cpu_to_le32(shader->binary.code + reloc->offset,
			&scratch_rsrc_dword1, 4);
		}
	}
}

/* For the UMR disassembler. */
#define DEBUGGER_END_OF_CODE_MARKER	0xbf9f0000 /* invalid instruction */
#define DEBUGGER_NUM_MARKERS		5

static unsigned si_get_shader_binary_size(const struct si_shader *shader)
{
	unsigned size = shader->binary.code_size;

	if (shader->prolog)
		size += shader->prolog->binary.code_size;
	if (shader->previous_stage)
		size += shader->previous_stage->binary.code_size;
	if (shader->prolog2)
		size += shader->prolog2->binary.code_size;
	if (shader->epilog)
		size += shader->epilog->binary.code_size;
	return size + DEBUGGER_NUM_MARKERS * 4;
}

int si_shader_binary_upload(struct si_screen *sscreen, struct si_shader *shader)
{
	const struct ac_shader_binary *prolog =
		shader->prolog ? &shader->prolog->binary : NULL;
	const struct ac_shader_binary *previous_stage =
		shader->previous_stage ? &shader->previous_stage->binary : NULL;
	const struct ac_shader_binary *prolog2 =
		shader->prolog2 ? &shader->prolog2->binary : NULL;
	const struct ac_shader_binary *epilog =
		shader->epilog ? &shader->epilog->binary : NULL;
	const struct ac_shader_binary *mainb = &shader->binary;
	unsigned bo_size = si_get_shader_binary_size(shader) +
			   (!epilog ? mainb->rodata_size : 0);
	unsigned char *ptr;

	assert(!prolog || !prolog->rodata_size);
	assert(!previous_stage || !previous_stage->rodata_size);
	assert(!prolog2 || !prolog2->rodata_size);
	assert((!prolog && !previous_stage && !prolog2 && !epilog) ||
	       !mainb->rodata_size);
	assert(!epilog || !epilog->rodata_size);

	si_resource_reference(&shader->bo, NULL);
	shader->bo = si_aligned_buffer_create(&sscreen->b,
					      sscreen->cpdma_prefetch_writes_memory ?
						0 : SI_RESOURCE_FLAG_READ_ONLY,
                                              PIPE_USAGE_IMMUTABLE,
                                              align(bo_size, SI_CPDMA_ALIGNMENT),
                                              256);
	if (!shader->bo)
		return -ENOMEM;

	/* Upload. */
	ptr = sscreen->ws->buffer_map(shader->bo->buf, NULL,
					PIPE_TRANSFER_READ_WRITE |
					PIPE_TRANSFER_UNSYNCHRONIZED |
					RADEON_TRANSFER_TEMPORARY);

	/* Don't use util_memcpy_cpu_to_le32. LLVM binaries are
	 * endian-independent. */
	if (prolog) {
		memcpy(ptr, prolog->code, prolog->code_size);
		ptr += prolog->code_size;
	}
	if (previous_stage) {
		memcpy(ptr, previous_stage->code, previous_stage->code_size);
		ptr += previous_stage->code_size;
	}
	if (prolog2) {
		memcpy(ptr, prolog2->code, prolog2->code_size);
		ptr += prolog2->code_size;
	}

	memcpy(ptr, mainb->code, mainb->code_size);
	ptr += mainb->code_size;

	if (epilog) {
		memcpy(ptr, epilog->code, epilog->code_size);
		ptr += epilog->code_size;
	} else if (mainb->rodata_size > 0) {
		memcpy(ptr, mainb->rodata, mainb->rodata_size);
		ptr += mainb->rodata_size;
	}

	/* Add end-of-code markers for the UMR disassembler. */
	uint32_t *ptr32 = (uint32_t*)ptr;
	for (unsigned i = 0; i < DEBUGGER_NUM_MARKERS; i++)
		ptr32[i] = DEBUGGER_END_OF_CODE_MARKER;

	sscreen->ws->buffer_unmap(shader->bo->buf);
	return 0;
}

static void si_shader_dump_disassembly(const struct ac_shader_binary *binary,
				       struct pipe_debug_callback *debug,
				       const char *name, FILE *file)
{
	char *line, *p;
	unsigned i, count;

	if (binary->disasm_string) {
		fprintf(file, "Shader %s disassembly:\n", name);
		fprintf(file, "%s", binary->disasm_string);

		if (debug && debug->debug_message) {
			/* Very long debug messages are cut off, so send the
			 * disassembly one line at a time. This causes more
			 * overhead, but on the plus side it simplifies
			 * parsing of resulting logs.
			 */
			pipe_debug_message(debug, SHADER_INFO,
					   "Shader Disassembly Begin");

			line = binary->disasm_string;
			while (*line) {
				p = util_strchrnul(line, '\n');
				count = p - line;

				if (count) {
					pipe_debug_message(debug, SHADER_INFO,
							   "%.*s", count, line);
				}

				if (!*p)
					break;
				line = p + 1;
			}

			pipe_debug_message(debug, SHADER_INFO,
					   "Shader Disassembly End");
		}
	} else {
		fprintf(file, "Shader %s binary:\n", name);
		for (i = 0; i < binary->code_size; i += 4) {
			fprintf(file, "@0x%x: %02x%02x%02x%02x\n", i,
				binary->code[i + 3], binary->code[i + 2],
				binary->code[i + 1], binary->code[i]);
		}
	}
}

static void si_calculate_max_simd_waves(struct si_shader *shader)
{
	struct si_screen *sscreen = shader->selector->screen;
	struct si_shader_config *conf = &shader->config;
	unsigned num_inputs = shader->selector->info.num_inputs;
	unsigned lds_increment = sscreen->info.chip_class >= GFX7 ? 512 : 256;
	unsigned lds_per_wave = 0;
	unsigned max_simd_waves;

	max_simd_waves = ac_get_max_simd_waves(sscreen->info.family);

	/* Compute LDS usage for PS. */
	switch (shader->selector->type) {
	case PIPE_SHADER_FRAGMENT:
		/* The minimum usage per wave is (num_inputs * 48). The maximum
		 * usage is (num_inputs * 48 * 16).
		 * We can get anything in between and it varies between waves.
		 *
		 * The 48 bytes per input for a single primitive is equal to
		 * 4 bytes/component * 4 components/input * 3 points.
		 *
		 * Other stages don't know the size at compile time or don't
		 * allocate LDS per wave, but instead they do it per thread group.
		 */
		lds_per_wave = conf->lds_size * lds_increment +
			       align(num_inputs * 48, lds_increment);
		break;
	case PIPE_SHADER_COMPUTE:
		if (shader->selector) {
			unsigned max_workgroup_size =
				si_get_max_workgroup_size(shader);
			lds_per_wave = (conf->lds_size * lds_increment) /
				       DIV_ROUND_UP(max_workgroup_size, 64);
		}
		break;
	}

	/* Compute the per-SIMD wave counts. */
	if (conf->num_sgprs) {
		max_simd_waves =
			MIN2(max_simd_waves,
			     ac_get_num_physical_sgprs(sscreen->info.chip_class) / conf->num_sgprs);
	}

	if (conf->num_vgprs)
		max_simd_waves = MIN2(max_simd_waves, 256 / conf->num_vgprs);

	/* LDS is 64KB per CU (4 SIMDs), which is 16KB per SIMD (usage above
	 * 16KB makes some SIMDs unoccupied). */
	if (lds_per_wave)
		max_simd_waves = MIN2(max_simd_waves, 16384 / lds_per_wave);

	conf->max_simd_waves = max_simd_waves;
}

void si_shader_dump_stats_for_shader_db(const struct si_shader *shader,
					struct pipe_debug_callback *debug)
{
	const struct si_shader_config *conf = &shader->config;

	pipe_debug_message(debug, SHADER_INFO,
			   "Shader Stats: SGPRS: %d VGPRS: %d Code Size: %d "
			   "LDS: %d Scratch: %d Max Waves: %d Spilled SGPRs: %d "
			   "Spilled VGPRs: %d PrivMem VGPRs: %d",
			   conf->num_sgprs, conf->num_vgprs,
			   si_get_shader_binary_size(shader),
			   conf->lds_size, conf->scratch_bytes_per_wave,
			   conf->max_simd_waves, conf->spilled_sgprs,
			   conf->spilled_vgprs, conf->private_mem_vgprs);
}

static void si_shader_dump_stats(struct si_screen *sscreen,
				 const struct si_shader *shader,
			         unsigned processor,
				 FILE *file,
				 bool check_debug_option)
{
	const struct si_shader_config *conf = &shader->config;

	if (!check_debug_option ||
	    si_can_dump_shader(sscreen, processor)) {
		if (processor == PIPE_SHADER_FRAGMENT) {
			fprintf(file, "*** SHADER CONFIG ***\n"
				"SPI_PS_INPUT_ADDR = 0x%04x\n"
				"SPI_PS_INPUT_ENA  = 0x%04x\n",
				conf->spi_ps_input_addr, conf->spi_ps_input_ena);
		}

		fprintf(file, "*** SHADER STATS ***\n"
			"SGPRS: %d\n"
			"VGPRS: %d\n"
		        "Spilled SGPRs: %d\n"
			"Spilled VGPRs: %d\n"
			"Private memory VGPRs: %d\n"
			"Code Size: %d bytes\n"
			"LDS: %d blocks\n"
			"Scratch: %d bytes per wave\n"
			"Max Waves: %d\n"
			"********************\n\n\n",
			conf->num_sgprs, conf->num_vgprs,
			conf->spilled_sgprs, conf->spilled_vgprs,
			conf->private_mem_vgprs,
			si_get_shader_binary_size(shader),
			conf->lds_size, conf->scratch_bytes_per_wave,
			conf->max_simd_waves);
	}
}

const char *si_get_shader_name(const struct si_shader *shader, unsigned processor)
{
	switch (processor) {
	case PIPE_SHADER_VERTEX:
		if (shader->key.as_es)
			return "Vertex Shader as ES";
		else if (shader->key.as_ls)
			return "Vertex Shader as LS";
		else if (shader->key.opt.vs_as_prim_discard_cs)
			return "Vertex Shader as Primitive Discard CS";
		else
			return "Vertex Shader as VS";
	case PIPE_SHADER_TESS_CTRL:
		return "Tessellation Control Shader";
	case PIPE_SHADER_TESS_EVAL:
		if (shader->key.as_es)
			return "Tessellation Evaluation Shader as ES";
		else
			return "Tessellation Evaluation Shader as VS";
	case PIPE_SHADER_GEOMETRY:
		if (shader->is_gs_copy_shader)
			return "GS Copy Shader as VS";
		else
			return "Geometry Shader";
	case PIPE_SHADER_FRAGMENT:
		return "Pixel Shader";
	case PIPE_SHADER_COMPUTE:
		return "Compute Shader";
	default:
		return "Unknown Shader";
	}
}

void si_shader_dump(struct si_screen *sscreen, const struct si_shader *shader,
		    struct pipe_debug_callback *debug, unsigned processor,
		    FILE *file, bool check_debug_option)
{
	if (!check_debug_option ||
	    si_can_dump_shader(sscreen, processor))
		si_dump_shader_key(processor, shader, file);

	if (!check_debug_option && shader->binary.llvm_ir_string) {
		if (shader->previous_stage &&
		    shader->previous_stage->binary.llvm_ir_string) {
			fprintf(file, "\n%s - previous stage - LLVM IR:\n\n",
				si_get_shader_name(shader, processor));
			fprintf(file, "%s\n", shader->previous_stage->binary.llvm_ir_string);
		}

		fprintf(file, "\n%s - main shader part - LLVM IR:\n\n",
			si_get_shader_name(shader, processor));
		fprintf(file, "%s\n", shader->binary.llvm_ir_string);
	}

	if (!check_debug_option ||
	    (si_can_dump_shader(sscreen, processor) &&
	     !(sscreen->debug_flags & DBG(NO_ASM)))) {
		fprintf(file, "\n%s:\n", si_get_shader_name(shader, processor));

		if (shader->prolog)
			si_shader_dump_disassembly(&shader->prolog->binary,
						   debug, "prolog", file);
		if (shader->previous_stage)
			si_shader_dump_disassembly(&shader->previous_stage->binary,
						   debug, "previous stage", file);
		if (shader->prolog2)
			si_shader_dump_disassembly(&shader->prolog2->binary,
						   debug, "prolog2", file);

		si_shader_dump_disassembly(&shader->binary, debug, "main", file);

		if (shader->epilog)
			si_shader_dump_disassembly(&shader->epilog->binary,
						   debug, "epilog", file);
		fprintf(file, "\n");
	}

	si_shader_dump_stats(sscreen, shader, processor, file,
			     check_debug_option);
}

static int si_compile_llvm(struct si_screen *sscreen,
			   struct ac_shader_binary *binary,
			   struct si_shader_config *conf,
			   struct ac_llvm_compiler *compiler,
			   LLVMModuleRef mod,
			   struct pipe_debug_callback *debug,
			   unsigned processor,
			   const char *name,
			   bool less_optimized)
{
	int r = 0;
	unsigned count = p_atomic_inc_return(&sscreen->num_compilations);

	if (si_can_dump_shader(sscreen, processor)) {
		fprintf(stderr, "radeonsi: Compiling shader %d\n", count);

		if (!(sscreen->debug_flags & (DBG(NO_IR) | DBG(PREOPT_IR)))) {
			fprintf(stderr, "%s LLVM IR:\n\n", name);
			ac_dump_module(mod);
			fprintf(stderr, "\n");
		}
	}

	if (sscreen->record_llvm_ir) {
		char *ir = LLVMPrintModuleToString(mod);
		binary->llvm_ir_string = strdup(ir);
		LLVMDisposeMessage(ir);
	}

	if (!si_replace_shader(count, binary)) {
		r = si_llvm_compile(mod, binary, compiler, debug,
				    less_optimized);
		if (r)
			return r;
	}

	si_shader_binary_read_config(binary, conf, 0);

	/* Enable 64-bit and 16-bit denormals, because there is no performance
	 * cost.
	 *
	 * If denormals are enabled, all floating-point output modifiers are
	 * ignored.
	 *
	 * Don't enable denormals for 32-bit floats, because:
	 * - Floating-point output modifiers would be ignored by the hw.
	 * - Some opcodes don't support denormals, such as v_mad_f32. We would
	 *   have to stop using those.
	 * - GFX6 & GFX7 would be very slow.
	 */
	conf->float_mode |= V_00B028_FP_64_DENORMS;

	FREE(binary->config);
	FREE(binary->global_symbol_offsets);
	binary->config = NULL;
	binary->global_symbol_offsets = NULL;

	/* Some shaders can't have rodata because their binaries can be
	 * concatenated.
	 */
	if (binary->rodata_size &&
	    (processor == PIPE_SHADER_VERTEX ||
	     processor == PIPE_SHADER_TESS_CTRL ||
	     processor == PIPE_SHADER_TESS_EVAL ||
	     processor == PIPE_SHADER_FRAGMENT)) {
		fprintf(stderr, "radeonsi: The shader can't have rodata.");
		return -EINVAL;
	}

	return r;
}

static void si_llvm_build_ret(struct si_shader_context *ctx, LLVMValueRef ret)
{
	if (LLVMGetTypeKind(LLVMTypeOf(ret)) == LLVMVoidTypeKind)
		LLVMBuildRetVoid(ctx->ac.builder);
	else
		LLVMBuildRet(ctx->ac.builder, ret);
}

/* Generate code for the hardware VS shader stage to go with a geometry shader */
struct si_shader *
si_generate_gs_copy_shader(struct si_screen *sscreen,
			   struct ac_llvm_compiler *compiler,
			   struct si_shader_selector *gs_selector,
			   struct pipe_debug_callback *debug)
{
	struct si_shader_context ctx;
	struct si_shader *shader;
	LLVMBuilderRef builder;
	struct si_shader_output_values outputs[SI_MAX_VS_OUTPUTS];
	struct tgsi_shader_info *gsinfo = &gs_selector->info;
	int i, r;


	shader = CALLOC_STRUCT(si_shader);
	if (!shader)
		return NULL;

	/* We can leave the fence as permanently signaled because the GS copy
	 * shader only becomes visible globally after it has been compiled. */
	util_queue_fence_init(&shader->ready);

	shader->selector = gs_selector;
	shader->is_gs_copy_shader = true;

	si_init_shader_ctx(&ctx, sscreen, compiler);
	ctx.shader = shader;
	ctx.type = PIPE_SHADER_VERTEX;

	builder = ctx.ac.builder;

	create_function(&ctx);
	preload_ring_buffers(&ctx);

	LLVMValueRef voffset =
		LLVMBuildMul(ctx.ac.builder, ctx.abi.vertex_id,
			     LLVMConstInt(ctx.i32, 4, 0), "");

	/* Fetch the vertex stream ID.*/
	LLVMValueRef stream_id;

	if (gs_selector->so.num_outputs)
		stream_id = si_unpack_param(&ctx, ctx.param_streamout_config, 24, 2);
	else
		stream_id = ctx.i32_0;

	/* Fill in output information. */
	for (i = 0; i < gsinfo->num_outputs; ++i) {
		outputs[i].semantic_name = gsinfo->output_semantic_name[i];
		outputs[i].semantic_index = gsinfo->output_semantic_index[i];

		for (int chan = 0; chan < 4; chan++) {
			outputs[i].vertex_stream[chan] =
				(gsinfo->output_streams[i] >> (2 * chan)) & 3;
		}
	}

	LLVMBasicBlockRef end_bb;
	LLVMValueRef switch_inst;

	end_bb = LLVMAppendBasicBlockInContext(ctx.ac.context, ctx.main_fn, "end");
	switch_inst = LLVMBuildSwitch(builder, stream_id, end_bb, 4);

	for (int stream = 0; stream < 4; stream++) {
		LLVMBasicBlockRef bb;
		unsigned offset;

		if (!gsinfo->num_stream_output_components[stream])
			continue;

		if (stream > 0 && !gs_selector->so.num_outputs)
			continue;

		bb = LLVMInsertBasicBlockInContext(ctx.ac.context, end_bb, "out");
		LLVMAddCase(switch_inst, LLVMConstInt(ctx.i32, stream, 0), bb);
		LLVMPositionBuilderAtEnd(builder, bb);

		/* Fetch vertex data from GSVS ring */
		offset = 0;
		for (i = 0; i < gsinfo->num_outputs; ++i) {
			for (unsigned chan = 0; chan < 4; chan++) {
				if (!(gsinfo->output_usagemask[i] & (1 << chan)) ||
				    outputs[i].vertex_stream[chan] != stream) {
					outputs[i].values[chan] = LLVMGetUndef(ctx.f32);
					continue;
				}

				LLVMValueRef soffset = LLVMConstInt(ctx.i32,
					offset * gs_selector->gs_max_out_vertices * 16 * 4, 0);
				offset++;

				outputs[i].values[chan] =
					ac_build_buffer_load(&ctx.ac,
							     ctx.gsvs_ring[0], 1,
							     ctx.i32_0, voffset,
							     soffset, 0, 1, 1,
							     true, false);
			}
		}

		/* Streamout and exports. */
		if (gs_selector->so.num_outputs) {
			si_llvm_emit_streamout(&ctx, outputs,
					       gsinfo->num_outputs,
					       stream);
		}

		if (stream == 0) {
			/* Vertex color clamping.
			 *
			 * This uses a state constant loaded in a user data SGPR and
			 * an IF statement is added that clamps all colors if the constant
			 * is true.
			 */
			struct lp_build_if_state if_ctx;
			LLVMValueRef v[2], cond = NULL;
			LLVMBasicBlockRef blocks[2];

			for (unsigned i = 0; i < gsinfo->num_outputs; i++) {
				if (gsinfo->output_semantic_name[i] != TGSI_SEMANTIC_COLOR &&
				    gsinfo->output_semantic_name[i] != TGSI_SEMANTIC_BCOLOR)
					continue;

				/* We've found a color. */
				if (!cond) {
					/* The state is in the first bit of the user SGPR. */
					cond = LLVMGetParam(ctx.main_fn,
							    ctx.param_vs_state_bits);
					cond = LLVMBuildTrunc(ctx.ac.builder, cond,
							      ctx.i1, "");
					lp_build_if(&if_ctx, &ctx.gallivm, cond);
					/* Remember blocks for Phi. */
					blocks[0] = if_ctx.true_block;
					blocks[1] = if_ctx.entry_block;
				}

				for (unsigned j = 0; j < 4; j++) {
					/* Insert clamp into the true block. */
					v[0] = ac_build_clamp(&ctx.ac, outputs[i].values[j]);
					v[1] = outputs[i].values[j];

					/* Insert Phi into the endif block. */
					LLVMPositionBuilderAtEnd(ctx.ac.builder, if_ctx.merge_block);
					outputs[i].values[j] = ac_build_phi(&ctx.ac, ctx.f32, 2, v, blocks);
					LLVMPositionBuilderAtEnd(ctx.ac.builder, if_ctx.true_block);
				}
			}
			if (cond)
				lp_build_endif(&if_ctx);

			si_llvm_export_vs(&ctx, outputs, gsinfo->num_outputs);
		}

		LLVMBuildBr(builder, end_bb);
	}

	LLVMPositionBuilderAtEnd(builder, end_bb);

	LLVMBuildRetVoid(ctx.ac.builder);

	ctx.type = PIPE_SHADER_GEOMETRY; /* override for shader dumping */
	si_llvm_optimize_module(&ctx);

	r = si_compile_llvm(sscreen, &ctx.shader->binary,
			    &ctx.shader->config, ctx.compiler,
			    ctx.ac.module,
			    debug, PIPE_SHADER_GEOMETRY,
			    "GS Copy Shader", false);
	if (!r) {
		if (si_can_dump_shader(sscreen, PIPE_SHADER_GEOMETRY))
			fprintf(stderr, "GS Copy Shader:\n");
		si_shader_dump(sscreen, ctx.shader, debug,
			       PIPE_SHADER_GEOMETRY, stderr, true);
		r = si_shader_binary_upload(sscreen, ctx.shader);
	}

	si_llvm_dispose(&ctx);

	if (r != 0) {
		FREE(shader);
		shader = NULL;
	} else {
		si_fix_resource_usage(sscreen, shader);
	}
	return shader;
}

static void si_dump_shader_key_vs(const struct si_shader_key *key,
				  const struct si_vs_prolog_bits *prolog,
				  const char *prefix, FILE *f)
{
	fprintf(f, "  %s.instance_divisor_is_one = %u\n",
		prefix, prolog->instance_divisor_is_one);
	fprintf(f, "  %s.instance_divisor_is_fetched = %u\n",
		prefix, prolog->instance_divisor_is_fetched);
	fprintf(f, "  %s.unpack_instance_id_from_vertex_id = %u\n",
		prefix, prolog->unpack_instance_id_from_vertex_id);
	fprintf(f, "  %s.ls_vgpr_fix = %u\n",
		prefix, prolog->ls_vgpr_fix);

	fprintf(f, "  mono.vs.fetch_opencode = %x\n", key->mono.vs_fetch_opencode);
	fprintf(f, "  mono.vs.fix_fetch = {");
	for (int i = 0; i < SI_MAX_ATTRIBS; i++) {
		union si_vs_fix_fetch fix = key->mono.vs_fix_fetch[i];
		if (i)
			fprintf(f, ", ");
		if (!fix.bits)
			fprintf(f, "0");
		else
			fprintf(f, "%u.%u.%u.%u", fix.u.reverse, fix.u.log_size,
				fix.u.num_channels_m1, fix.u.format);
	}
	fprintf(f, "}\n");
}

static void si_dump_shader_key(unsigned processor, const struct si_shader *shader,
			       FILE *f)
{
	const struct si_shader_key *key = &shader->key;

	fprintf(f, "SHADER KEY\n");

	switch (processor) {
	case PIPE_SHADER_VERTEX:
		si_dump_shader_key_vs(key, &key->part.vs.prolog,
				      "part.vs.prolog", f);
		fprintf(f, "  as_es = %u\n", key->as_es);
		fprintf(f, "  as_ls = %u\n", key->as_ls);
		fprintf(f, "  mono.u.vs_export_prim_id = %u\n",
			key->mono.u.vs_export_prim_id);
		fprintf(f, "  opt.vs_as_prim_discard_cs = %u\n",
			key->opt.vs_as_prim_discard_cs);
		fprintf(f, "  opt.cs_prim_type = %s\n",
			tgsi_primitive_names[key->opt.cs_prim_type]);
		fprintf(f, "  opt.cs_indexed = %u\n",
			key->opt.cs_indexed);
		fprintf(f, "  opt.cs_instancing = %u\n",
			key->opt.cs_instancing);
		fprintf(f, "  opt.cs_primitive_restart = %u\n",
			key->opt.cs_primitive_restart);
		fprintf(f, "  opt.cs_provoking_vertex_first = %u\n",
			key->opt.cs_provoking_vertex_first);
		fprintf(f, "  opt.cs_need_correct_orientation = %u\n",
			key->opt.cs_need_correct_orientation);
		fprintf(f, "  opt.cs_cull_front = %u\n",
			key->opt.cs_cull_front);
		fprintf(f, "  opt.cs_cull_back = %u\n",
			key->opt.cs_cull_back);
		fprintf(f, "  opt.cs_cull_z = %u\n",
			key->opt.cs_cull_z);
		fprintf(f, "  opt.cs_halfz_clip_space = %u\n",
			key->opt.cs_halfz_clip_space);
		break;

	case PIPE_SHADER_TESS_CTRL:
		if (shader->selector->screen->info.chip_class >= GFX9) {
			si_dump_shader_key_vs(key, &key->part.tcs.ls_prolog,
					      "part.tcs.ls_prolog", f);
		}
		fprintf(f, "  part.tcs.epilog.prim_mode = %u\n", key->part.tcs.epilog.prim_mode);
		fprintf(f, "  mono.u.ff_tcs_inputs_to_copy = 0x%"PRIx64"\n", key->mono.u.ff_tcs_inputs_to_copy);
		break;

	case PIPE_SHADER_TESS_EVAL:
		fprintf(f, "  as_es = %u\n", key->as_es);
		fprintf(f, "  mono.u.vs_export_prim_id = %u\n",
			key->mono.u.vs_export_prim_id);
		break;

	case PIPE_SHADER_GEOMETRY:
		if (shader->is_gs_copy_shader)
			break;

		if (shader->selector->screen->info.chip_class >= GFX9 &&
		    key->part.gs.es->type == PIPE_SHADER_VERTEX) {
			si_dump_shader_key_vs(key, &key->part.gs.vs_prolog,
					      "part.gs.vs_prolog", f);
		}
		fprintf(f, "  part.gs.prolog.tri_strip_adj_fix = %u\n", key->part.gs.prolog.tri_strip_adj_fix);
		break;

	case PIPE_SHADER_COMPUTE:
		break;

	case PIPE_SHADER_FRAGMENT:
		fprintf(f, "  part.ps.prolog.color_two_side = %u\n", key->part.ps.prolog.color_two_side);
		fprintf(f, "  part.ps.prolog.flatshade_colors = %u\n", key->part.ps.prolog.flatshade_colors);
		fprintf(f, "  part.ps.prolog.poly_stipple = %u\n", key->part.ps.prolog.poly_stipple);
		fprintf(f, "  part.ps.prolog.force_persp_sample_interp = %u\n", key->part.ps.prolog.force_persp_sample_interp);
		fprintf(f, "  part.ps.prolog.force_linear_sample_interp = %u\n", key->part.ps.prolog.force_linear_sample_interp);
		fprintf(f, "  part.ps.prolog.force_persp_center_interp = %u\n", key->part.ps.prolog.force_persp_center_interp);
		fprintf(f, "  part.ps.prolog.force_linear_center_interp = %u\n", key->part.ps.prolog.force_linear_center_interp);
		fprintf(f, "  part.ps.prolog.bc_optimize_for_persp = %u\n", key->part.ps.prolog.bc_optimize_for_persp);
		fprintf(f, "  part.ps.prolog.bc_optimize_for_linear = %u\n", key->part.ps.prolog.bc_optimize_for_linear);
		fprintf(f, "  part.ps.epilog.spi_shader_col_format = 0x%x\n", key->part.ps.epilog.spi_shader_col_format);
		fprintf(f, "  part.ps.epilog.color_is_int8 = 0x%X\n", key->part.ps.epilog.color_is_int8);
		fprintf(f, "  part.ps.epilog.color_is_int10 = 0x%X\n", key->part.ps.epilog.color_is_int10);
		fprintf(f, "  part.ps.epilog.last_cbuf = %u\n", key->part.ps.epilog.last_cbuf);
		fprintf(f, "  part.ps.epilog.alpha_func = %u\n", key->part.ps.epilog.alpha_func);
		fprintf(f, "  part.ps.epilog.alpha_to_one = %u\n", key->part.ps.epilog.alpha_to_one);
		fprintf(f, "  part.ps.epilog.poly_line_smoothing = %u\n", key->part.ps.epilog.poly_line_smoothing);
		fprintf(f, "  part.ps.epilog.clamp_color = %u\n", key->part.ps.epilog.clamp_color);
		break;

	default:
		assert(0);
	}

	if ((processor == PIPE_SHADER_GEOMETRY ||
	     processor == PIPE_SHADER_TESS_EVAL ||
	     processor == PIPE_SHADER_VERTEX) &&
	    !key->as_es && !key->as_ls) {
		fprintf(f, "  opt.kill_outputs = 0x%"PRIx64"\n", key->opt.kill_outputs);
		fprintf(f, "  opt.clip_disable = %u\n", key->opt.clip_disable);
	}
}

static void si_init_shader_ctx(struct si_shader_context *ctx,
			       struct si_screen *sscreen,
			       struct ac_llvm_compiler *compiler)
{
	struct lp_build_tgsi_context *bld_base;

	si_llvm_context_init(ctx, sscreen, compiler);

	bld_base = &ctx->bld_base;
	bld_base->emit_fetch_funcs[TGSI_FILE_CONSTANT] = fetch_constant;

	bld_base->op_actions[TGSI_OPCODE_INTERP_CENTROID].emit = build_interp_intrinsic;
	bld_base->op_actions[TGSI_OPCODE_INTERP_SAMPLE].emit = build_interp_intrinsic;
	bld_base->op_actions[TGSI_OPCODE_INTERP_OFFSET].emit = build_interp_intrinsic;

	bld_base->op_actions[TGSI_OPCODE_MEMBAR].emit = membar_emit;

	bld_base->op_actions[TGSI_OPCODE_CLOCK].emit = clock_emit;

	bld_base->op_actions[TGSI_OPCODE_DDX].emit = si_llvm_emit_ddxy;
	bld_base->op_actions[TGSI_OPCODE_DDY].emit = si_llvm_emit_ddxy;
	bld_base->op_actions[TGSI_OPCODE_DDX_FINE].emit = si_llvm_emit_ddxy;
	bld_base->op_actions[TGSI_OPCODE_DDY_FINE].emit = si_llvm_emit_ddxy;

	bld_base->op_actions[TGSI_OPCODE_VOTE_ALL].emit = vote_all_emit;
	bld_base->op_actions[TGSI_OPCODE_VOTE_ANY].emit = vote_any_emit;
	bld_base->op_actions[TGSI_OPCODE_VOTE_EQ].emit = vote_eq_emit;
	bld_base->op_actions[TGSI_OPCODE_BALLOT].emit = ballot_emit;
	bld_base->op_actions[TGSI_OPCODE_READ_FIRST].intr_name = "llvm.amdgcn.readfirstlane";
	bld_base->op_actions[TGSI_OPCODE_READ_FIRST].emit = read_lane_emit;
	bld_base->op_actions[TGSI_OPCODE_READ_INVOC].intr_name = "llvm.amdgcn.readlane";
	bld_base->op_actions[TGSI_OPCODE_READ_INVOC].emit = read_lane_emit;

	bld_base->op_actions[TGSI_OPCODE_EMIT].emit = si_tgsi_emit_vertex;
	bld_base->op_actions[TGSI_OPCODE_ENDPRIM].emit = si_tgsi_emit_primitive;
	bld_base->op_actions[TGSI_OPCODE_BARRIER].emit = si_llvm_emit_barrier;
}

static void si_optimize_vs_outputs(struct si_shader_context *ctx)
{
	struct si_shader *shader = ctx->shader;
	struct tgsi_shader_info *info = &shader->selector->info;

	if ((ctx->type != PIPE_SHADER_VERTEX &&
	     ctx->type != PIPE_SHADER_TESS_EVAL) ||
	    shader->key.as_ls ||
	    shader->key.as_es)
		return;

	ac_optimize_vs_outputs(&ctx->ac,
			       ctx->main_fn,
			       shader->info.vs_output_param_offset,
			       info->num_outputs,
			       &shader->info.nr_param_exports);
}

static void si_init_exec_from_input(struct si_shader_context *ctx,
				    unsigned param, unsigned bitoffset)
{
	LLVMValueRef args[] = {
		LLVMGetParam(ctx->main_fn, param),
		LLVMConstInt(ctx->i32, bitoffset, 0),
	};
	ac_build_intrinsic(&ctx->ac,
			   "llvm.amdgcn.init.exec.from.input",
			   ctx->voidt, args, 2, AC_FUNC_ATTR_CONVERGENT);
}

static bool si_vs_needs_prolog(const struct si_shader_selector *sel,
			       const struct si_vs_prolog_bits *key)
{
	/* VGPR initialization fixup for Vega10 and Raven is always done in the
	 * VS prolog. */
	return sel->vs_needs_prolog || key->ls_vgpr_fix;
}

static bool si_compile_tgsi_main(struct si_shader_context *ctx)
{
	struct si_shader *shader = ctx->shader;
	struct si_shader_selector *sel = shader->selector;
	struct lp_build_tgsi_context *bld_base = &ctx->bld_base;

	// TODO clean all this up!
	switch (ctx->type) {
	case PIPE_SHADER_VERTEX:
		ctx->load_input = declare_input_vs;
		if (shader->key.as_ls)
			ctx->abi.emit_outputs = si_llvm_emit_ls_epilogue;
		else if (shader->key.as_es)
			ctx->abi.emit_outputs = si_llvm_emit_es_epilogue;
		else if (shader->key.opt.vs_as_prim_discard_cs)
			ctx->abi.emit_outputs = si_llvm_emit_prim_discard_cs_epilogue;
		else
			ctx->abi.emit_outputs = si_llvm_emit_vs_epilogue;
		bld_base->emit_epilogue = si_tgsi_emit_epilogue;
		ctx->abi.load_base_vertex = get_base_vertex;
		break;
	case PIPE_SHADER_TESS_CTRL:
		bld_base->emit_fetch_funcs[TGSI_FILE_INPUT] = fetch_input_tcs;
		ctx->abi.load_tess_varyings = si_nir_load_tcs_varyings;
		bld_base->emit_fetch_funcs[TGSI_FILE_OUTPUT] = fetch_output_tcs;
		bld_base->emit_store = store_output_tcs;
		ctx->abi.store_tcs_outputs = si_nir_store_output_tcs;
		ctx->abi.emit_outputs = si_llvm_emit_tcs_epilogue;
		ctx->abi.load_patch_vertices_in = si_load_patch_vertices_in;
		bld_base->emit_epilogue = si_tgsi_emit_epilogue;
		break;
	case PIPE_SHADER_TESS_EVAL:
		bld_base->emit_fetch_funcs[TGSI_FILE_INPUT] = fetch_input_tes;
		ctx->abi.load_tess_varyings = si_nir_load_input_tes;
		ctx->abi.load_tess_coord = si_load_tess_coord;
		ctx->abi.load_tess_level = si_load_tess_level;
		ctx->abi.load_patch_vertices_in = si_load_patch_vertices_in;
		if (shader->key.as_es)
			ctx->abi.emit_outputs = si_llvm_emit_es_epilogue;
		else
			ctx->abi.emit_outputs = si_llvm_emit_vs_epilogue;
		bld_base->emit_epilogue = si_tgsi_emit_epilogue;
		break;
	case PIPE_SHADER_GEOMETRY:
		bld_base->emit_fetch_funcs[TGSI_FILE_INPUT] = fetch_input_gs;
		ctx->abi.load_inputs = si_nir_load_input_gs;
		ctx->abi.emit_vertex = si_llvm_emit_vertex;
		ctx->abi.emit_primitive = si_llvm_emit_primitive;
		ctx->abi.emit_outputs = si_llvm_emit_gs_epilogue;
		bld_base->emit_epilogue = si_tgsi_emit_gs_epilogue;
		break;
	case PIPE_SHADER_FRAGMENT:
		ctx->load_input = declare_input_fs;
		ctx->abi.emit_outputs = si_llvm_return_fs_outputs;
		bld_base->emit_epilogue = si_tgsi_emit_epilogue;
		ctx->abi.lookup_interp_param = si_nir_lookup_interp_param;
		ctx->abi.load_sample_position = load_sample_position;
		ctx->abi.load_sample_mask_in = load_sample_mask_in;
		ctx->abi.emit_kill = si_llvm_emit_kill;
		break;
	case PIPE_SHADER_COMPUTE:
		ctx->abi.load_local_group_size = get_block_size;
		break;
	default:
		assert(!"Unsupported shader type");
		return false;
	}

	ctx->abi.load_ubo = load_ubo;
	ctx->abi.load_ssbo = load_ssbo;

	create_function(ctx);
	preload_ring_buffers(ctx);

	/* For GFX9 merged shaders:
	 * - Set EXEC for the first shader. If the prolog is present, set
	 *   EXEC there instead.
	 * - Add a barrier before the second shader.
	 * - In the second shader, reset EXEC to ~0 and wrap the main part in
	 *   an if-statement. This is required for correctness in geometry
	 *   shaders, to ensure that empty GS waves do not send GS_EMIT and
	 *   GS_CUT messages.
	 *
	 * For monolithic merged shaders, the first shader is wrapped in an
	 * if-block together with its prolog in si_build_wrapper_function.
	 */
	if (ctx->screen->info.chip_class >= GFX9) {
		if (!shader->is_monolithic &&
		    sel->info.num_instructions > 1 && /* not empty shader */
		    (shader->key.as_es || shader->key.as_ls) &&
		    (ctx->type == PIPE_SHADER_TESS_EVAL ||
		     (ctx->type == PIPE_SHADER_VERTEX &&
		      !si_vs_needs_prolog(sel, &shader->key.part.vs.prolog)))) {
			si_init_exec_from_input(ctx,
						ctx->param_merged_wave_info, 0);
		} else if (ctx->type == PIPE_SHADER_TESS_CTRL ||
			   ctx->type == PIPE_SHADER_GEOMETRY) {
			if (!shader->is_monolithic)
				ac_init_exec_full_mask(&ctx->ac);

			LLVMValueRef num_threads = si_unpack_param(ctx, ctx->param_merged_wave_info, 8, 8);
			LLVMValueRef ena =
				LLVMBuildICmp(ctx->ac.builder, LLVMIntULT,
					    ac_get_thread_id(&ctx->ac), num_threads, "");
			lp_build_if(&ctx->merged_wrap_if_state, &ctx->gallivm, ena);

			/* The barrier must execute for all shaders in a
			 * threadgroup.
			 *
			 * Execute the barrier inside the conditional block,
			 * so that empty waves can jump directly to s_endpgm,
			 * which will also signal the barrier.
			 *
			 * If the shader is TCS and the TCS epilog is present
			 * and contains a barrier, it will wait there and then
			 * reach s_endpgm.
			 */
			si_llvm_emit_barrier(NULL, bld_base, NULL);
		}
	}

	if (ctx->type == PIPE_SHADER_TESS_CTRL &&
	    sel->tcs_info.tessfactors_are_def_in_all_invocs) {
		for (unsigned i = 0; i < 6; i++) {
			ctx->invoc0_tess_factors[i] =
				ac_build_alloca_undef(&ctx->ac, ctx->i32, "");
		}
	}

	if (ctx->type == PIPE_SHADER_GEOMETRY) {
		int i;
		for (i = 0; i < 4; i++) {
			ctx->gs_next_vertex[i] =
				ac_build_alloca(&ctx->ac, ctx->i32, "");
		}
	}

	if (sel->force_correct_derivs_after_kill) {
		ctx->postponed_kill = ac_build_alloca_undef(&ctx->ac, ctx->i1, "");
		/* true = don't kill. */
		LLVMBuildStore(ctx->ac.builder, ctx->i1true,
			       ctx->postponed_kill);
	}

	if (sel->tokens) {
		if (!lp_build_tgsi_llvm(bld_base, sel->tokens)) {
			fprintf(stderr, "Failed to translate shader from TGSI to LLVM\n");
			return false;
		}
	} else {
		if (!si_nir_build_llvm(ctx, sel->nir)) {
			fprintf(stderr, "Failed to translate shader from NIR to LLVM\n");
			return false;
		}
	}

	si_llvm_build_ret(ctx, ctx->return_value);
	return true;
}

/**
 * Compute the VS prolog key, which contains all the information needed to
 * build the VS prolog function, and set shader->info bits where needed.
 *
 * \param info             Shader info of the vertex shader.
 * \param num_input_sgprs  Number of input SGPRs for the vertex shader.
 * \param prolog_key       Key of the VS prolog
 * \param shader_out       The vertex shader, or the next shader if merging LS+HS or ES+GS.
 * \param key              Output shader part key.
 */
static void si_get_vs_prolog_key(const struct tgsi_shader_info *info,
				 unsigned num_input_sgprs,
				 const struct si_vs_prolog_bits *prolog_key,
				 struct si_shader *shader_out,
				 union si_shader_part_key *key)
{
	memset(key, 0, sizeof(*key));
	key->vs_prolog.states = *prolog_key;
	key->vs_prolog.num_input_sgprs = num_input_sgprs;
	key->vs_prolog.last_input = MAX2(1, info->num_inputs) - 1;
	key->vs_prolog.as_ls = shader_out->key.as_ls;
	key->vs_prolog.as_es = shader_out->key.as_es;

	if (shader_out->selector->type == PIPE_SHADER_TESS_CTRL) {
		key->vs_prolog.as_ls = 1;
		key->vs_prolog.num_merged_next_stage_vgprs = 2;
	} else if (shader_out->selector->type == PIPE_SHADER_GEOMETRY) {
		key->vs_prolog.as_es = 1;
		key->vs_prolog.num_merged_next_stage_vgprs = 5;
	}

	/* Enable loading the InstanceID VGPR. */
	uint16_t input_mask = u_bit_consecutive(0, info->num_inputs);

	if ((key->vs_prolog.states.instance_divisor_is_one |
	     key->vs_prolog.states.instance_divisor_is_fetched) & input_mask)
		shader_out->info.uses_instanceid = true;
}

/**
 * Compute the PS prolog key, which contains all the information needed to
 * build the PS prolog function, and set related bits in shader->config.
 */
static void si_get_ps_prolog_key(struct si_shader *shader,
				 union si_shader_part_key *key,
				 bool separate_prolog)
{
	struct tgsi_shader_info *info = &shader->selector->info;

	memset(key, 0, sizeof(*key));
	key->ps_prolog.states = shader->key.part.ps.prolog;
	key->ps_prolog.colors_read = info->colors_read;
	key->ps_prolog.num_input_sgprs = shader->info.num_input_sgprs;
	key->ps_prolog.num_input_vgprs = shader->info.num_input_vgprs;
	key->ps_prolog.wqm = info->uses_derivatives &&
		(key->ps_prolog.colors_read ||
		 key->ps_prolog.states.force_persp_sample_interp ||
		 key->ps_prolog.states.force_linear_sample_interp ||
		 key->ps_prolog.states.force_persp_center_interp ||
		 key->ps_prolog.states.force_linear_center_interp ||
		 key->ps_prolog.states.bc_optimize_for_persp ||
		 key->ps_prolog.states.bc_optimize_for_linear);
	key->ps_prolog.ancillary_vgpr_index = shader->info.ancillary_vgpr_index;

	if (info->colors_read) {
		unsigned *color = shader->selector->color_attr_index;

		if (shader->key.part.ps.prolog.color_two_side) {
			/* BCOLORs are stored after the last input. */
			key->ps_prolog.num_interp_inputs = info->num_inputs;
			key->ps_prolog.face_vgpr_index = shader->info.face_vgpr_index;
			shader->config.spi_ps_input_ena |= S_0286CC_FRONT_FACE_ENA(1);
		}

		for (unsigned i = 0; i < 2; i++) {
			unsigned interp = info->input_interpolate[color[i]];
			unsigned location = info->input_interpolate_loc[color[i]];

			if (!(info->colors_read & (0xf << i*4)))
				continue;

			key->ps_prolog.color_attr_index[i] = color[i];

			if (shader->key.part.ps.prolog.flatshade_colors &&
			    interp == TGSI_INTERPOLATE_COLOR)
				interp = TGSI_INTERPOLATE_CONSTANT;

			switch (interp) {
			case TGSI_INTERPOLATE_CONSTANT:
				key->ps_prolog.color_interp_vgpr_index[i] = -1;
				break;
			case TGSI_INTERPOLATE_PERSPECTIVE:
			case TGSI_INTERPOLATE_COLOR:
				/* Force the interpolation location for colors here. */
				if (shader->key.part.ps.prolog.force_persp_sample_interp)
					location = TGSI_INTERPOLATE_LOC_SAMPLE;
				if (shader->key.part.ps.prolog.force_persp_center_interp)
					location = TGSI_INTERPOLATE_LOC_CENTER;

				switch (location) {
				case TGSI_INTERPOLATE_LOC_SAMPLE:
					key->ps_prolog.color_interp_vgpr_index[i] = 0;
					shader->config.spi_ps_input_ena |=
						S_0286CC_PERSP_SAMPLE_ENA(1);
					break;
				case TGSI_INTERPOLATE_LOC_CENTER:
					key->ps_prolog.color_interp_vgpr_index[i] = 2;
					shader->config.spi_ps_input_ena |=
						S_0286CC_PERSP_CENTER_ENA(1);
					break;
				case TGSI_INTERPOLATE_LOC_CENTROID:
					key->ps_prolog.color_interp_vgpr_index[i] = 4;
					shader->config.spi_ps_input_ena |=
						S_0286CC_PERSP_CENTROID_ENA(1);
					break;
				default:
					assert(0);
				}
				break;
			case TGSI_INTERPOLATE_LINEAR:
				/* Force the interpolation location for colors here. */
				if (shader->key.part.ps.prolog.force_linear_sample_interp)
					location = TGSI_INTERPOLATE_LOC_SAMPLE;
				if (shader->key.part.ps.prolog.force_linear_center_interp)
					location = TGSI_INTERPOLATE_LOC_CENTER;

				/* The VGPR assignment for non-monolithic shaders
				 * works because InitialPSInputAddr is set on the
				 * main shader and PERSP_PULL_MODEL is never used.
				 */
				switch (location) {
				case TGSI_INTERPOLATE_LOC_SAMPLE:
					key->ps_prolog.color_interp_vgpr_index[i] =
						separate_prolog ? 6 : 9;
					shader->config.spi_ps_input_ena |=
						S_0286CC_LINEAR_SAMPLE_ENA(1);
					break;
				case TGSI_INTERPOLATE_LOC_CENTER:
					key->ps_prolog.color_interp_vgpr_index[i] =
						separate_prolog ? 8 : 11;
					shader->config.spi_ps_input_ena |=
						S_0286CC_LINEAR_CENTER_ENA(1);
					break;
				case TGSI_INTERPOLATE_LOC_CENTROID:
					key->ps_prolog.color_interp_vgpr_index[i] =
						separate_prolog ? 10 : 13;
					shader->config.spi_ps_input_ena |=
						S_0286CC_LINEAR_CENTROID_ENA(1);
					break;
				default:
					assert(0);
				}
				break;
			default:
				assert(0);
			}
		}
	}
}

/**
 * Check whether a PS prolog is required based on the key.
 */
static bool si_need_ps_prolog(const union si_shader_part_key *key)
{
	return key->ps_prolog.colors_read ||
	       key->ps_prolog.states.force_persp_sample_interp ||
	       key->ps_prolog.states.force_linear_sample_interp ||
	       key->ps_prolog.states.force_persp_center_interp ||
	       key->ps_prolog.states.force_linear_center_interp ||
	       key->ps_prolog.states.bc_optimize_for_persp ||
	       key->ps_prolog.states.bc_optimize_for_linear ||
	       key->ps_prolog.states.poly_stipple ||
	       key->ps_prolog.states.samplemask_log_ps_iter;
}

/**
 * Compute the PS epilog key, which contains all the information needed to
 * build the PS epilog function.
 */
static void si_get_ps_epilog_key(struct si_shader *shader,
				 union si_shader_part_key *key)
{
	struct tgsi_shader_info *info = &shader->selector->info;
	memset(key, 0, sizeof(*key));
	key->ps_epilog.colors_written = info->colors_written;
	key->ps_epilog.writes_z = info->writes_z;
	key->ps_epilog.writes_stencil = info->writes_stencil;
	key->ps_epilog.writes_samplemask = info->writes_samplemask;
	key->ps_epilog.states = shader->key.part.ps.epilog;
}

/**
 * Build the GS prolog function. Rotate the input vertices for triangle strips
 * with adjacency.
 */
static void si_build_gs_prolog_function(struct si_shader_context *ctx,
					union si_shader_part_key *key)
{
	unsigned num_sgprs, num_vgprs;
	struct si_function_info fninfo;
	LLVMBuilderRef builder = ctx->ac.builder;
	LLVMTypeRef returns[48];
	LLVMValueRef func, ret;

	si_init_function_info(&fninfo);

	if (ctx->screen->info.chip_class >= GFX9) {
		if (key->gs_prolog.states.gfx9_prev_is_vs)
			num_sgprs = 8 + GFX9_VSGS_NUM_USER_SGPR;
		else
			num_sgprs = 8 + GFX9_TESGS_NUM_USER_SGPR;
		num_vgprs = 5; /* ES inputs are not needed by GS */
	} else {
		num_sgprs = GFX6_GS_NUM_USER_SGPR + 2;
		num_vgprs = 8;
	}

	for (unsigned i = 0; i < num_sgprs; ++i) {
		add_arg(&fninfo, ARG_SGPR, ctx->i32);
		returns[i] = ctx->i32;
	}

	for (unsigned i = 0; i < num_vgprs; ++i) {
		add_arg(&fninfo, ARG_VGPR, ctx->i32);
		returns[num_sgprs + i] = ctx->f32;
	}

	/* Create the function. */
	si_create_function(ctx, "gs_prolog", returns, num_sgprs + num_vgprs,
			   &fninfo, 0);
	func = ctx->main_fn;

	/* Set the full EXEC mask for the prolog, because we are only fiddling
	 * with registers here. The main shader part will set the correct EXEC
	 * mask.
	 */
	if (ctx->screen->info.chip_class >= GFX9 && !key->gs_prolog.is_monolithic)
		ac_init_exec_full_mask(&ctx->ac);

	/* Copy inputs to outputs. This should be no-op, as the registers match,
	 * but it will prevent the compiler from overwriting them unintentionally.
	 */
	ret = ctx->return_value;
	for (unsigned i = 0; i < num_sgprs; i++) {
		LLVMValueRef p = LLVMGetParam(func, i);
		ret = LLVMBuildInsertValue(builder, ret, p, i, "");
	}
	for (unsigned i = 0; i < num_vgprs; i++) {
		LLVMValueRef p = LLVMGetParam(func, num_sgprs + i);
		p = ac_to_float(&ctx->ac, p);
		ret = LLVMBuildInsertValue(builder, ret, p, num_sgprs + i, "");
	}

	if (key->gs_prolog.states.tri_strip_adj_fix) {
		/* Remap the input vertices for every other primitive. */
		const unsigned gfx6_vtx_params[6] = {
			num_sgprs,
			num_sgprs + 1,
			num_sgprs + 3,
			num_sgprs + 4,
			num_sgprs + 5,
			num_sgprs + 6
		};
		const unsigned gfx9_vtx_params[3] = {
			num_sgprs,
			num_sgprs + 1,
			num_sgprs + 4,
		};
		LLVMValueRef vtx_in[6], vtx_out[6];
		LLVMValueRef prim_id, rotate;

		if (ctx->screen->info.chip_class >= GFX9) {
			for (unsigned i = 0; i < 3; i++) {
				vtx_in[i*2] = si_unpack_param(ctx, gfx9_vtx_params[i], 0, 16);
				vtx_in[i*2+1] = si_unpack_param(ctx, gfx9_vtx_params[i], 16, 16);
			}
		} else {
			for (unsigned i = 0; i < 6; i++)
				vtx_in[i] = LLVMGetParam(func, gfx6_vtx_params[i]);
		}

		prim_id = LLVMGetParam(func, num_sgprs + 2);
		rotate = LLVMBuildTrunc(builder, prim_id, ctx->i1, "");

		for (unsigned i = 0; i < 6; ++i) {
			LLVMValueRef base, rotated;
			base = vtx_in[i];
			rotated = vtx_in[(i + 4) % 6];
			vtx_out[i] = LLVMBuildSelect(builder, rotate, rotated, base, "");
		}

		if (ctx->screen->info.chip_class >= GFX9) {
			for (unsigned i = 0; i < 3; i++) {
				LLVMValueRef hi, out;

				hi = LLVMBuildShl(builder, vtx_out[i*2+1],
						  LLVMConstInt(ctx->i32, 16, 0), "");
				out = LLVMBuildOr(builder, vtx_out[i*2], hi, "");
				out = ac_to_float(&ctx->ac, out);
				ret = LLVMBuildInsertValue(builder, ret, out,
							   gfx9_vtx_params[i], "");
			}
		} else {
			for (unsigned i = 0; i < 6; i++) {
				LLVMValueRef out;

				out = ac_to_float(&ctx->ac, vtx_out[i]);
				ret = LLVMBuildInsertValue(builder, ret, out,
							   gfx6_vtx_params[i], "");
			}
		}
	}

	LLVMBuildRet(builder, ret);
}

/**
 * Given a list of shader part functions, build a wrapper function that
 * runs them in sequence to form a monolithic shader.
 */
static void si_build_wrapper_function(struct si_shader_context *ctx,
				      LLVMValueRef *parts,
				      unsigned num_parts,
				      unsigned main_part,
				      unsigned next_shader_first_part)
{
	LLVMBuilderRef builder = ctx->ac.builder;
	/* PS epilog has one arg per color component; gfx9 merged shader
	 * prologs need to forward 32 user SGPRs.
	 */
	struct si_function_info fninfo;
	LLVMValueRef initial[64], out[64];
	LLVMTypeRef function_type;
	unsigned num_first_params;
	unsigned num_out, initial_num_out;
	MAYBE_UNUSED unsigned num_out_sgpr; /* used in debug checks */
	MAYBE_UNUSED unsigned initial_num_out_sgpr; /* used in debug checks */
	unsigned num_sgprs, num_vgprs;
	unsigned gprs;
	struct lp_build_if_state if_state;

	si_init_function_info(&fninfo);

	for (unsigned i = 0; i < num_parts; ++i) {
		ac_add_function_attr(ctx->ac.context, parts[i], -1,
				     AC_FUNC_ATTR_ALWAYSINLINE);
		LLVMSetLinkage(parts[i], LLVMPrivateLinkage);
	}

	/* The parameters of the wrapper function correspond to those of the
	 * first part in terms of SGPRs and VGPRs, but we use the types of the
	 * main part to get the right types. This is relevant for the
	 * dereferenceable attribute on descriptor table pointers.
	 */
	num_sgprs = 0;
	num_vgprs = 0;

	function_type = LLVMGetElementType(LLVMTypeOf(parts[0]));
	num_first_params = LLVMCountParamTypes(function_type);

	for (unsigned i = 0; i < num_first_params; ++i) {
		LLVMValueRef param = LLVMGetParam(parts[0], i);

		if (ac_is_sgpr_param(param)) {
			assert(num_vgprs == 0);
			num_sgprs += ac_get_type_size(LLVMTypeOf(param)) / 4;
		} else {
			num_vgprs += ac_get_type_size(LLVMTypeOf(param)) / 4;
		}
	}

	gprs = 0;
	while (gprs < num_sgprs + num_vgprs) {
		LLVMValueRef param = LLVMGetParam(parts[main_part], fninfo.num_params);
		LLVMTypeRef type = LLVMTypeOf(param);
		unsigned size = ac_get_type_size(type) / 4;

		add_arg(&fninfo, gprs < num_sgprs ? ARG_SGPR : ARG_VGPR, type);

		assert(ac_is_sgpr_param(param) == (gprs < num_sgprs));
		assert(gprs + size <= num_sgprs + num_vgprs &&
		       (gprs >= num_sgprs || gprs + size <= num_sgprs));

		gprs += size;
	}

	/* Prepare the return type. */
	unsigned num_returns = 0;
	LLVMTypeRef returns[32], last_func_type, return_type;

	last_func_type = LLVMGetElementType(LLVMTypeOf(parts[num_parts - 1]));
	return_type = LLVMGetReturnType(last_func_type);

	switch (LLVMGetTypeKind(return_type)) {
	case LLVMStructTypeKind:
		num_returns = LLVMCountStructElementTypes(return_type);
		assert(num_returns <= ARRAY_SIZE(returns));
		LLVMGetStructElementTypes(return_type, returns);
		break;
	case LLVMVoidTypeKind:
		break;
	default:
		unreachable("unexpected type");
	}

	si_create_function(ctx, "wrapper", returns, num_returns, &fninfo,
			   si_get_max_workgroup_size(ctx->shader));

	if (is_merged_shader(ctx))
		ac_init_exec_full_mask(&ctx->ac);

	/* Record the arguments of the function as if they were an output of
	 * a previous part.
	 */
	num_out = 0;
	num_out_sgpr = 0;

	for (unsigned i = 0; i < fninfo.num_params; ++i) {
		LLVMValueRef param = LLVMGetParam(ctx->main_fn, i);
		LLVMTypeRef param_type = LLVMTypeOf(param);
		LLVMTypeRef out_type = i < fninfo.num_sgpr_params ? ctx->i32 : ctx->f32;
		unsigned size = ac_get_type_size(param_type) / 4;

		if (size == 1) {
			if (LLVMGetTypeKind(param_type) == LLVMPointerTypeKind) {
				param = LLVMBuildPtrToInt(builder, param, ctx->i32, "");
				param_type = ctx->i32;
			}

			if (param_type != out_type)
				param = LLVMBuildBitCast(builder, param, out_type, "");
			out[num_out++] = param;
		} else {
			LLVMTypeRef vector_type = LLVMVectorType(out_type, size);

			if (LLVMGetTypeKind(param_type) == LLVMPointerTypeKind) {
				param = LLVMBuildPtrToInt(builder, param, ctx->i64, "");
				param_type = ctx->i64;
			}

			if (param_type != vector_type)
				param = LLVMBuildBitCast(builder, param, vector_type, "");

			for (unsigned j = 0; j < size; ++j)
				out[num_out++] = LLVMBuildExtractElement(
					builder, param, LLVMConstInt(ctx->i32, j, 0), "");
		}

		if (i < fninfo.num_sgpr_params)
			num_out_sgpr = num_out;
	}

	memcpy(initial, out, sizeof(out));
	initial_num_out = num_out;
	initial_num_out_sgpr = num_out_sgpr;

	/* Now chain the parts. */
	LLVMValueRef ret;
	for (unsigned part = 0; part < num_parts; ++part) {
		LLVMValueRef in[48];
		LLVMTypeRef ret_type;
		unsigned out_idx = 0;
		unsigned num_params = LLVMCountParams(parts[part]);

		/* Merged shaders are executed conditionally depending
		 * on the number of enabled threads passed in the input SGPRs. */
		if (is_merged_shader(ctx) && part == 0) {
			LLVMValueRef ena, count = initial[3];

			count = LLVMBuildAnd(builder, count,
					     LLVMConstInt(ctx->i32, 0x7f, 0), "");
			ena = LLVMBuildICmp(builder, LLVMIntULT,
					    ac_get_thread_id(&ctx->ac), count, "");
			lp_build_if(&if_state, &ctx->gallivm, ena);
		}

		/* Derive arguments for the next part from outputs of the
		 * previous one.
		 */
		for (unsigned param_idx = 0; param_idx < num_params; ++param_idx) {
			LLVMValueRef param;
			LLVMTypeRef param_type;
			bool is_sgpr;
			unsigned param_size;
			LLVMValueRef arg = NULL;

			param = LLVMGetParam(parts[part], param_idx);
			param_type = LLVMTypeOf(param);
			param_size = ac_get_type_size(param_type) / 4;
			is_sgpr = ac_is_sgpr_param(param);

			if (is_sgpr) {
				ac_add_function_attr(ctx->ac.context, parts[part],
						     param_idx + 1, AC_FUNC_ATTR_INREG);
			} else if (out_idx < num_out_sgpr) {
				/* Skip returned SGPRs the current part doesn't
				 * declare on the input. */
				out_idx = num_out_sgpr;
			}

			assert(out_idx + param_size <= (is_sgpr ? num_out_sgpr : num_out));

			if (param_size == 1)
				arg = out[out_idx];
			else
				arg = ac_build_gather_values(&ctx->ac, &out[out_idx], param_size);

			if (LLVMTypeOf(arg) != param_type) {
				if (LLVMGetTypeKind(param_type) == LLVMPointerTypeKind) {
					if (LLVMGetPointerAddressSpace(param_type) ==
					    AC_ADDR_SPACE_CONST_32BIT) {
						arg = LLVMBuildBitCast(builder, arg, ctx->i32, "");
						arg = LLVMBuildIntToPtr(builder, arg, param_type, "");
					} else {
						arg = LLVMBuildBitCast(builder, arg, ctx->i64, "");
						arg = LLVMBuildIntToPtr(builder, arg, param_type, "");
					}
				} else {
					arg = LLVMBuildBitCast(builder, arg, param_type, "");
				}
			}

			in[param_idx] = arg;
			out_idx += param_size;
		}

		ret = LLVMBuildCall(builder, parts[part], in, num_params, "");

		if (is_merged_shader(ctx) &&
		    part + 1 == next_shader_first_part) {
			lp_build_endif(&if_state);

			/* The second half of the merged shader should use
			 * the inputs from the toplevel (wrapper) function,
			 * not the return value from the last call.
			 *
			 * That's because the last call was executed condi-
			 * tionally, so we can't consume it in the main
			 * block.
			 */
			memcpy(out, initial, sizeof(initial));
			num_out = initial_num_out;
			num_out_sgpr = initial_num_out_sgpr;
			continue;
		}

		/* Extract the returned GPRs. */
		ret_type = LLVMTypeOf(ret);
		num_out = 0;
		num_out_sgpr = 0;

		if (LLVMGetTypeKind(ret_type) != LLVMVoidTypeKind) {
			assert(LLVMGetTypeKind(ret_type) == LLVMStructTypeKind);

			unsigned ret_size = LLVMCountStructElementTypes(ret_type);

			for (unsigned i = 0; i < ret_size; ++i) {
				LLVMValueRef val =
					LLVMBuildExtractValue(builder, ret, i, "");

				assert(num_out < ARRAY_SIZE(out));
				out[num_out++] = val;

				if (LLVMTypeOf(val) == ctx->i32) {
					assert(num_out_sgpr + 1 == num_out);
					num_out_sgpr = num_out;
				}
			}
		}
	}

	/* Return the value from the last part. */
	if (LLVMGetTypeKind(LLVMTypeOf(ret)) == LLVMVoidTypeKind)
		LLVMBuildRetVoid(builder);
	else
		LLVMBuildRet(builder, ret);
}

static bool si_should_optimize_less(struct ac_llvm_compiler *compiler,
				    struct si_shader_selector *sel)
{
	if (!compiler->low_opt_passes)
		return false;

	/* Assume a slow CPU. */
	assert(!sel->screen->info.has_dedicated_vram &&
	       sel->screen->info.chip_class <= GFX8);

	/* For a crazy dEQP test containing 2597 memory opcodes, mostly
	 * buffer stores. */
	return sel->type == PIPE_SHADER_COMPUTE &&
	       sel->info.num_memory_instructions > 1000;
}

int si_compile_tgsi_shader(struct si_screen *sscreen,
			   struct ac_llvm_compiler *compiler,
			   struct si_shader *shader,
			   struct pipe_debug_callback *debug)
{
	struct si_shader_selector *sel = shader->selector;
	struct si_shader_context ctx;
	int r = -1;

	/* Dump TGSI code before doing TGSI->LLVM conversion in case the
	 * conversion fails. */
	if (si_can_dump_shader(sscreen, sel->info.processor) &&
	    !(sscreen->debug_flags & DBG(NO_TGSI))) {
		if (sel->tokens)
			tgsi_dump(sel->tokens, 0);
		else
			nir_print_shader(sel->nir, stderr);
		si_dump_streamout(&sel->so);
	}

	si_init_shader_ctx(&ctx, sscreen, compiler);
	si_llvm_context_set_tgsi(&ctx, shader);

	memset(shader->info.vs_output_param_offset, AC_EXP_PARAM_UNDEFINED,
	       sizeof(shader->info.vs_output_param_offset));

	shader->info.uses_instanceid = sel->info.uses_instanceid;

	if (!si_compile_tgsi_main(&ctx)) {
		si_llvm_dispose(&ctx);
		return -1;
	}

	if (shader->is_monolithic && ctx.type == PIPE_SHADER_VERTEX) {
		LLVMValueRef parts[2];
		bool need_prolog = sel->vs_needs_prolog;

		parts[1] = ctx.main_fn;

		if (need_prolog) {
			union si_shader_part_key prolog_key;
			si_get_vs_prolog_key(&sel->info,
					     shader->info.num_input_sgprs,
					     &shader->key.part.vs.prolog,
					     shader, &prolog_key);
			si_build_vs_prolog_function(&ctx, &prolog_key);
			parts[0] = ctx.main_fn;
		}

		si_build_wrapper_function(&ctx, parts + !need_prolog,
					  1 + need_prolog, need_prolog, 0);

		if (ctx.shader->key.opt.vs_as_prim_discard_cs)
			si_build_prim_discard_compute_shader(&ctx);
	} else if (shader->is_monolithic && ctx.type == PIPE_SHADER_TESS_CTRL) {
		if (sscreen->info.chip_class >= GFX9) {
			struct si_shader_selector *ls = shader->key.part.tcs.ls;
			LLVMValueRef parts[4];
			bool vs_needs_prolog =
				si_vs_needs_prolog(ls, &shader->key.part.tcs.ls_prolog);

			/* TCS main part */
			parts[2] = ctx.main_fn;

			/* TCS epilog */
			union si_shader_part_key tcs_epilog_key;
			memset(&tcs_epilog_key, 0, sizeof(tcs_epilog_key));
			tcs_epilog_key.tcs_epilog.states = shader->key.part.tcs.epilog;
			si_build_tcs_epilog_function(&ctx, &tcs_epilog_key);
			parts[3] = ctx.main_fn;

			/* VS as LS main part */
			struct si_shader shader_ls = {};
			shader_ls.selector = ls;
			shader_ls.key.as_ls = 1;
			shader_ls.key.mono = shader->key.mono;
			shader_ls.key.opt = shader->key.opt;
			shader_ls.is_monolithic = true;
			si_llvm_context_set_tgsi(&ctx, &shader_ls);

			if (!si_compile_tgsi_main(&ctx)) {
				si_llvm_dispose(&ctx);
				return -1;
			}
			shader->info.uses_instanceid |= ls->info.uses_instanceid;
			parts[1] = ctx.main_fn;

			/* LS prolog */
			if (vs_needs_prolog) {
				union si_shader_part_key vs_prolog_key;
				si_get_vs_prolog_key(&ls->info,
						     shader_ls.info.num_input_sgprs,
						     &shader->key.part.tcs.ls_prolog,
						     shader, &vs_prolog_key);
				vs_prolog_key.vs_prolog.is_monolithic = true;
				si_build_vs_prolog_function(&ctx, &vs_prolog_key);
				parts[0] = ctx.main_fn;
			}

			/* Reset the shader context. */
			ctx.shader = shader;
			ctx.type = PIPE_SHADER_TESS_CTRL;

			si_build_wrapper_function(&ctx,
						  parts + !vs_needs_prolog,
						  4 - !vs_needs_prolog, vs_needs_prolog,
						  vs_needs_prolog ? 2 : 1);
		} else {
			LLVMValueRef parts[2];
			union si_shader_part_key epilog_key;

			parts[0] = ctx.main_fn;

			memset(&epilog_key, 0, sizeof(epilog_key));
			epilog_key.tcs_epilog.states = shader->key.part.tcs.epilog;
			si_build_tcs_epilog_function(&ctx, &epilog_key);
			parts[1] = ctx.main_fn;

			si_build_wrapper_function(&ctx, parts, 2, 0, 0);
		}
	} else if (shader->is_monolithic && ctx.type == PIPE_SHADER_GEOMETRY) {
		if (ctx.screen->info.chip_class >= GFX9) {
			struct si_shader_selector *es = shader->key.part.gs.es;
			LLVMValueRef es_prolog = NULL;
			LLVMValueRef es_main = NULL;
			LLVMValueRef gs_prolog = NULL;
			LLVMValueRef gs_main = ctx.main_fn;

			/* GS prolog */
			union si_shader_part_key gs_prolog_key;
			memset(&gs_prolog_key, 0, sizeof(gs_prolog_key));
			gs_prolog_key.gs_prolog.states = shader->key.part.gs.prolog;
			gs_prolog_key.gs_prolog.is_monolithic = true;
			si_build_gs_prolog_function(&ctx, &gs_prolog_key);
			gs_prolog = ctx.main_fn;

			/* ES main part */
			struct si_shader shader_es = {};
			shader_es.selector = es;
			shader_es.key.as_es = 1;
			shader_es.key.mono = shader->key.mono;
			shader_es.key.opt = shader->key.opt;
			shader_es.is_monolithic = true;
			si_llvm_context_set_tgsi(&ctx, &shader_es);

			if (!si_compile_tgsi_main(&ctx)) {
				si_llvm_dispose(&ctx);
				return -1;
			}
			shader->info.uses_instanceid |= es->info.uses_instanceid;
			es_main = ctx.main_fn;

			/* ES prolog */
			if (es->vs_needs_prolog) {
				union si_shader_part_key vs_prolog_key;
				si_get_vs_prolog_key(&es->info,
						     shader_es.info.num_input_sgprs,
						     &shader->key.part.gs.vs_prolog,
						     shader, &vs_prolog_key);
				vs_prolog_key.vs_prolog.is_monolithic = true;
				si_build_vs_prolog_function(&ctx, &vs_prolog_key);
				es_prolog = ctx.main_fn;
			}

			/* Reset the shader context. */
			ctx.shader = shader;
			ctx.type = PIPE_SHADER_GEOMETRY;

			/* Prepare the array of shader parts. */
			LLVMValueRef parts[4];
			unsigned num_parts = 0, main_part, next_first_part;

			if (es_prolog)
				parts[num_parts++] = es_prolog;

			parts[main_part = num_parts++] = es_main;
			parts[next_first_part = num_parts++] = gs_prolog;
			parts[num_parts++] = gs_main;

			si_build_wrapper_function(&ctx, parts, num_parts,
						  main_part, next_first_part);
		} else {
			LLVMValueRef parts[2];
			union si_shader_part_key prolog_key;

			parts[1] = ctx.main_fn;

			memset(&prolog_key, 0, sizeof(prolog_key));
			prolog_key.gs_prolog.states = shader->key.part.gs.prolog;
			si_build_gs_prolog_function(&ctx, &prolog_key);
			parts[0] = ctx.main_fn;

			si_build_wrapper_function(&ctx, parts, 2, 1, 0);
		}
	} else if (shader->is_monolithic && ctx.type == PIPE_SHADER_FRAGMENT) {
		LLVMValueRef parts[3];
		union si_shader_part_key prolog_key;
		union si_shader_part_key epilog_key;
		bool need_prolog;

		si_get_ps_prolog_key(shader, &prolog_key, false);
		need_prolog = si_need_ps_prolog(&prolog_key);

		parts[need_prolog ? 1 : 0] = ctx.main_fn;

		if (need_prolog) {
			si_build_ps_prolog_function(&ctx, &prolog_key);
			parts[0] = ctx.main_fn;
		}

		si_get_ps_epilog_key(shader, &epilog_key);
		si_build_ps_epilog_function(&ctx, &epilog_key);
		parts[need_prolog ? 2 : 1] = ctx.main_fn;

		si_build_wrapper_function(&ctx, parts, need_prolog ? 3 : 2,
					  need_prolog ? 1 : 0, 0);
	}

	si_llvm_optimize_module(&ctx);

	/* Post-optimization transformations and analysis. */
	si_optimize_vs_outputs(&ctx);

	if ((debug && debug->debug_message) ||
	    si_can_dump_shader(sscreen, ctx.type)) {
		ctx.shader->config.private_mem_vgprs =
			ac_count_scratch_private_memory(ctx.main_fn);
	}

	/* Make sure the input is a pointer and not integer followed by inttoptr. */
	assert(LLVMGetTypeKind(LLVMTypeOf(LLVMGetParam(ctx.main_fn, 0))) ==
	       LLVMPointerTypeKind);

	/* Compile to bytecode. */
	r = si_compile_llvm(sscreen, &shader->binary, &shader->config, compiler,
			    ctx.ac.module, debug, ctx.type,
			    si_get_shader_name(shader, ctx.type),
			    si_should_optimize_less(compiler, shader->selector));
	si_llvm_dispose(&ctx);
	if (r) {
		fprintf(stderr, "LLVM failed to compile shader\n");
		return r;
	}

	/* Validate SGPR and VGPR usage for compute to detect compiler bugs.
	 * LLVM 3.9svn has this bug.
	 */
	if (sel->type == PIPE_SHADER_COMPUTE) {
		unsigned wave_size = 64;
		unsigned max_vgprs = 256;
		unsigned max_sgprs = sscreen->info.chip_class >= GFX8 ? 800 : 512;
		unsigned max_sgprs_per_wave = 128;
		unsigned max_block_threads = si_get_max_workgroup_size(shader);
		unsigned min_waves_per_cu = DIV_ROUND_UP(max_block_threads, wave_size);
		unsigned min_waves_per_simd = DIV_ROUND_UP(min_waves_per_cu, 4);

		max_vgprs = max_vgprs / min_waves_per_simd;
		max_sgprs = MIN2(max_sgprs / min_waves_per_simd, max_sgprs_per_wave);

		if (shader->config.num_sgprs > max_sgprs ||
		    shader->config.num_vgprs > max_vgprs) {
			fprintf(stderr, "LLVM failed to compile a shader correctly: "
				"SGPR:VGPR usage is %u:%u, but the hw limit is %u:%u\n",
				shader->config.num_sgprs, shader->config.num_vgprs,
				max_sgprs, max_vgprs);

			/* Just terminate the process, because dependent
			 * shaders can hang due to bad input data, but use
			 * the env var to allow shader-db to work.
			 */
			if (!debug_get_bool_option("SI_PASS_BAD_SHADERS", false))
				abort();
		}
	}

	/* Add the scratch offset to input SGPRs. */
	if (shader->config.scratch_bytes_per_wave && !is_merged_shader(&ctx))
		shader->info.num_input_sgprs += 1; /* scratch byte offset */

	/* Calculate the number of fragment input VGPRs. */
	if (ctx.type == PIPE_SHADER_FRAGMENT) {
		shader->info.num_input_vgprs = 0;
		shader->info.face_vgpr_index = -1;
		shader->info.ancillary_vgpr_index = -1;

		if (G_0286CC_PERSP_SAMPLE_ENA(shader->config.spi_ps_input_addr))
			shader->info.num_input_vgprs += 2;
		if (G_0286CC_PERSP_CENTER_ENA(shader->config.spi_ps_input_addr))
			shader->info.num_input_vgprs += 2;
		if (G_0286CC_PERSP_CENTROID_ENA(shader->config.spi_ps_input_addr))
			shader->info.num_input_vgprs += 2;
		if (G_0286CC_PERSP_PULL_MODEL_ENA(shader->config.spi_ps_input_addr))
			shader->info.num_input_vgprs += 3;
		if (G_0286CC_LINEAR_SAMPLE_ENA(shader->config.spi_ps_input_addr))
			shader->info.num_input_vgprs += 2;
		if (G_0286CC_LINEAR_CENTER_ENA(shader->config.spi_ps_input_addr))
			shader->info.num_input_vgprs += 2;
		if (G_0286CC_LINEAR_CENTROID_ENA(shader->config.spi_ps_input_addr))
			shader->info.num_input_vgprs += 2;
		if (G_0286CC_LINE_STIPPLE_TEX_ENA(shader->config.spi_ps_input_addr))
			shader->info.num_input_vgprs += 1;
		if (G_0286CC_POS_X_FLOAT_ENA(shader->config.spi_ps_input_addr))
			shader->info.num_input_vgprs += 1;
		if (G_0286CC_POS_Y_FLOAT_ENA(shader->config.spi_ps_input_addr))
			shader->info.num_input_vgprs += 1;
		if (G_0286CC_POS_Z_FLOAT_ENA(shader->config.spi_ps_input_addr))
			shader->info.num_input_vgprs += 1;
		if (G_0286CC_POS_W_FLOAT_ENA(shader->config.spi_ps_input_addr))
			shader->info.num_input_vgprs += 1;
		if (G_0286CC_FRONT_FACE_ENA(shader->config.spi_ps_input_addr)) {
			shader->info.face_vgpr_index = shader->info.num_input_vgprs;
			shader->info.num_input_vgprs += 1;
		}
		if (G_0286CC_ANCILLARY_ENA(shader->config.spi_ps_input_addr)) {
			shader->info.ancillary_vgpr_index = shader->info.num_input_vgprs;
			shader->info.num_input_vgprs += 1;
		}
		if (G_0286CC_SAMPLE_COVERAGE_ENA(shader->config.spi_ps_input_addr))
			shader->info.num_input_vgprs += 1;
		if (G_0286CC_POS_FIXED_PT_ENA(shader->config.spi_ps_input_addr))
			shader->info.num_input_vgprs += 1;
	}

	si_calculate_max_simd_waves(shader);
	si_shader_dump_stats_for_shader_db(shader, debug);
	return 0;
}

/**
 * Create, compile and return a shader part (prolog or epilog).
 *
 * \param sscreen	screen
 * \param list		list of shader parts of the same category
 * \param type		shader type
 * \param key		shader part key
 * \param prolog	whether the part being requested is a prolog
 * \param tm		LLVM target machine
 * \param debug		debug callback
 * \param build		the callback responsible for building the main function
 * \return		non-NULL on success
 */
static struct si_shader_part *
si_get_shader_part(struct si_screen *sscreen,
		   struct si_shader_part **list,
		   enum pipe_shader_type type,
		   bool prolog,
		   union si_shader_part_key *key,
		   struct ac_llvm_compiler *compiler,
		   struct pipe_debug_callback *debug,
		   void (*build)(struct si_shader_context *,
				 union si_shader_part_key *),
		   const char *name)
{
	struct si_shader_part *result;

	mtx_lock(&sscreen->shader_parts_mutex);

	/* Find existing. */
	for (result = *list; result; result = result->next) {
		if (memcmp(&result->key, key, sizeof(*key)) == 0) {
			mtx_unlock(&sscreen->shader_parts_mutex);
			return result;
		}
	}

	/* Compile a new one. */
	result = CALLOC_STRUCT(si_shader_part);
	result->key = *key;

	struct si_shader shader = {};
	struct si_shader_context ctx;

	si_init_shader_ctx(&ctx, sscreen, compiler);
	ctx.shader = &shader;
	ctx.type = type;

	switch (type) {
	case PIPE_SHADER_VERTEX:
		shader.key.as_ls = key->vs_prolog.as_ls;
		shader.key.as_es = key->vs_prolog.as_es;
		break;
	case PIPE_SHADER_TESS_CTRL:
		assert(!prolog);
		shader.key.part.tcs.epilog = key->tcs_epilog.states;
		break;
	case PIPE_SHADER_GEOMETRY:
		assert(prolog);
		break;
	case PIPE_SHADER_FRAGMENT:
		if (prolog)
			shader.key.part.ps.prolog = key->ps_prolog.states;
		else
			shader.key.part.ps.epilog = key->ps_epilog.states;
		break;
	default:
		unreachable("bad shader part");
	}

	build(&ctx, key);

	/* Compile. */
	si_llvm_optimize_module(&ctx);

	if (si_compile_llvm(sscreen, &result->binary, &result->config, compiler,
			    ctx.ac.module, debug, ctx.type, name, false)) {
		FREE(result);
		result = NULL;
		goto out;
	}

	result->next = *list;
	*list = result;

out:
	si_llvm_dispose(&ctx);
	mtx_unlock(&sscreen->shader_parts_mutex);
	return result;
}

static LLVMValueRef si_prolog_get_rw_buffers(struct si_shader_context *ctx)
{
	LLVMValueRef ptr[2], list;
	bool merged_shader = is_merged_shader(ctx);

	ptr[0] = LLVMGetParam(ctx->main_fn, (merged_shader ? 8 : 0) + SI_SGPR_RW_BUFFERS);
	list = LLVMBuildIntToPtr(ctx->ac.builder, ptr[0],
				 ac_array_in_const32_addr_space(ctx->v4i32), "");
	return list;
}

/**
 * Build the vertex shader prolog function.
 *
 * The inputs are the same as VS (a lot of SGPRs and 4 VGPR system values).
 * All inputs are returned unmodified. The vertex load indices are
 * stored after them, which will be used by the API VS for fetching inputs.
 *
 * For example, the expected outputs for instance_divisors[] = {0, 1, 2} are:
 *   input_v0,
 *   input_v1,
 *   input_v2,
 *   input_v3,
 *   (VertexID + BaseVertex),
 *   (InstanceID + StartInstance),
 *   (InstanceID / 2 + StartInstance)
 */
static void si_build_vs_prolog_function(struct si_shader_context *ctx,
					union si_shader_part_key *key)
{
	struct si_function_info fninfo;
	LLVMTypeRef *returns;
	LLVMValueRef ret, func;
	int num_returns, i;
	unsigned first_vs_vgpr = key->vs_prolog.num_merged_next_stage_vgprs;
	unsigned num_input_vgprs = key->vs_prolog.num_merged_next_stage_vgprs + 4;
	LLVMValueRef input_vgprs[9];
	unsigned num_all_input_regs = key->vs_prolog.num_input_sgprs +
				      num_input_vgprs;
	unsigned user_sgpr_base = key->vs_prolog.num_merged_next_stage_vgprs ? 8 : 0;

	si_init_function_info(&fninfo);

	/* 4 preloaded VGPRs + vertex load indices as prolog outputs */
	returns = alloca((num_all_input_regs + key->vs_prolog.last_input + 1) *
			 sizeof(LLVMTypeRef));
	num_returns = 0;

	/* Declare input and output SGPRs. */
	for (i = 0; i < key->vs_prolog.num_input_sgprs; i++) {
		add_arg(&fninfo, ARG_SGPR, ctx->i32);
		returns[num_returns++] = ctx->i32;
	}

	/* Preloaded VGPRs (outputs must be floats) */
	for (i = 0; i < num_input_vgprs; i++) {
		add_arg_assign(&fninfo, ARG_VGPR, ctx->i32, &input_vgprs[i]);
		returns[num_returns++] = ctx->f32;
	}

	/* Vertex load indices. */
	for (i = 0; i <= key->vs_prolog.last_input; i++)
		returns[num_returns++] = ctx->f32;

	/* Create the function. */
	si_create_function(ctx, "vs_prolog", returns, num_returns, &fninfo, 0);
	func = ctx->main_fn;

	if (key->vs_prolog.num_merged_next_stage_vgprs) {
		if (!key->vs_prolog.is_monolithic)
			si_init_exec_from_input(ctx, 3, 0);

		if (key->vs_prolog.as_ls &&
		    ctx->screen->has_ls_vgpr_init_bug) {
			/* If there are no HS threads, SPI loads the LS VGPRs
			 * starting at VGPR 0. Shift them back to where they
			 * belong.
			 */
			LLVMValueRef has_hs_threads =
				LLVMBuildICmp(ctx->ac.builder, LLVMIntNE,
				    si_unpack_param(ctx, 3, 8, 8),
				    ctx->i32_0, "");

			for (i = 4; i > 0; --i) {
				input_vgprs[i + 1] =
					LLVMBuildSelect(ctx->ac.builder, has_hs_threads,
						        input_vgprs[i + 1],
						        input_vgprs[i - 1], "");
			}
		}
	}

	unsigned vertex_id_vgpr = first_vs_vgpr;
	unsigned instance_id_vgpr = first_vs_vgpr + (key->vs_prolog.as_ls ? 2 : 1);

	ctx->abi.vertex_id = input_vgprs[vertex_id_vgpr];
	ctx->abi.instance_id = input_vgprs[instance_id_vgpr];

	/* InstanceID = VertexID >> 16;
	 * VertexID   = VertexID & 0xffff;
	 */
	if (key->vs_prolog.states.unpack_instance_id_from_vertex_id) {
		ctx->abi.instance_id = LLVMBuildLShr(ctx->ac.builder, ctx->abi.vertex_id,
						     LLVMConstInt(ctx->i32, 16, 0), "");
		ctx->abi.vertex_id = LLVMBuildAnd(ctx->ac.builder, ctx->abi.vertex_id,
						  LLVMConstInt(ctx->i32, 0xffff, 0), "");
	}

	/* Copy inputs to outputs. This should be no-op, as the registers match,
	 * but it will prevent the compiler from overwriting them unintentionally.
	 */
	ret = ctx->return_value;
	for (i = 0; i < key->vs_prolog.num_input_sgprs; i++) {
		LLVMValueRef p = LLVMGetParam(func, i);
		ret = LLVMBuildInsertValue(ctx->ac.builder, ret, p, i, "");
	}
	for (i = 0; i < num_input_vgprs; i++) {
		LLVMValueRef p = input_vgprs[i];

		if (i == vertex_id_vgpr)
			p = ctx->abi.vertex_id;
		else if (i == instance_id_vgpr)
			p = ctx->abi.instance_id;

		p = ac_to_float(&ctx->ac, p);
		ret = LLVMBuildInsertValue(ctx->ac.builder, ret, p,
					   key->vs_prolog.num_input_sgprs + i, "");
	}

	/* Compute vertex load indices from instance divisors. */
	LLVMValueRef instance_divisor_constbuf = NULL;

	if (key->vs_prolog.states.instance_divisor_is_fetched) {
		LLVMValueRef list = si_prolog_get_rw_buffers(ctx);
		LLVMValueRef buf_index =
			LLVMConstInt(ctx->i32, SI_VS_CONST_INSTANCE_DIVISORS, 0);
		instance_divisor_constbuf =
			ac_build_load_to_sgpr(&ctx->ac, list, buf_index);
	}

	for (i = 0; i <= key->vs_prolog.last_input; i++) {
		bool divisor_is_one =
			key->vs_prolog.states.instance_divisor_is_one & (1u << i);
		bool divisor_is_fetched =
			key->vs_prolog.states.instance_divisor_is_fetched & (1u << i);
		LLVMValueRef index = NULL;

		if (divisor_is_one) {
			index = ctx->abi.instance_id;
		} else if (divisor_is_fetched) {
			LLVMValueRef udiv_factors[4];

			for (unsigned j = 0; j < 4; j++) {
				udiv_factors[j] =
					buffer_load_const(ctx, instance_divisor_constbuf,
							  LLVMConstInt(ctx->i32, i*16 + j*4, 0));
				udiv_factors[j] = ac_to_integer(&ctx->ac, udiv_factors[j]);
			}
			/* The faster NUW version doesn't work when InstanceID == UINT_MAX.
			 * Such InstanceID might not be achievable in a reasonable time though.
			 */
			index = ac_build_fast_udiv_nuw(&ctx->ac, ctx->abi.instance_id,
						       udiv_factors[0], udiv_factors[1],
						       udiv_factors[2], udiv_factors[3]);
		}

		if (divisor_is_one || divisor_is_fetched) {
			/* Add StartInstance. */
			index = LLVMBuildAdd(ctx->ac.builder, index,
					     LLVMGetParam(ctx->main_fn, user_sgpr_base +
							  SI_SGPR_START_INSTANCE), "");
		} else {
			/* VertexID + BaseVertex */
			index = LLVMBuildAdd(ctx->ac.builder,
					     ctx->abi.vertex_id,
					     LLVMGetParam(func, user_sgpr_base +
								SI_SGPR_BASE_VERTEX), "");
		}

		index = ac_to_float(&ctx->ac, index);
		ret = LLVMBuildInsertValue(ctx->ac.builder, ret, index,
					   fninfo.num_params + i, "");
	}

	si_llvm_build_ret(ctx, ret);
}

static bool si_get_vs_prolog(struct si_screen *sscreen,
			     struct ac_llvm_compiler *compiler,
			     struct si_shader *shader,
			     struct pipe_debug_callback *debug,
			     struct si_shader *main_part,
			     const struct si_vs_prolog_bits *key)
{
	struct si_shader_selector *vs = main_part->selector;

	if (!si_vs_needs_prolog(vs, key))
		return true;

	/* Get the prolog. */
	union si_shader_part_key prolog_key;
	si_get_vs_prolog_key(&vs->info, main_part->info.num_input_sgprs,
			     key, shader, &prolog_key);

	shader->prolog =
		si_get_shader_part(sscreen, &sscreen->vs_prologs,
				   PIPE_SHADER_VERTEX, true, &prolog_key, compiler,
				   debug, si_build_vs_prolog_function,
				   "Vertex Shader Prolog");
	return shader->prolog != NULL;
}

/**
 * Select and compile (or reuse) vertex shader parts (prolog & epilog).
 */
static bool si_shader_select_vs_parts(struct si_screen *sscreen,
				      struct ac_llvm_compiler *compiler,
				      struct si_shader *shader,
				      struct pipe_debug_callback *debug)
{
	return si_get_vs_prolog(sscreen, compiler, shader, debug, shader,
				&shader->key.part.vs.prolog);
}

/**
 * Compile the TCS epilog function. This writes tesselation factors to memory
 * based on the output primitive type of the tesselator (determined by TES).
 */
static void si_build_tcs_epilog_function(struct si_shader_context *ctx,
					 union si_shader_part_key *key)
{
	struct lp_build_tgsi_context *bld_base = &ctx->bld_base;
	struct si_function_info fninfo;
	LLVMValueRef func;

	si_init_function_info(&fninfo);

	if (ctx->screen->info.chip_class >= GFX9) {
		add_arg(&fninfo, ARG_SGPR, ctx->i32);
		add_arg(&fninfo, ARG_SGPR, ctx->i32);
		ctx->param_tcs_offchip_offset = add_arg(&fninfo, ARG_SGPR, ctx->i32);
		add_arg(&fninfo, ARG_SGPR, ctx->i32); /* wave info */
		ctx->param_tcs_factor_offset = add_arg(&fninfo, ARG_SGPR, ctx->i32);
		add_arg(&fninfo, ARG_SGPR, ctx->i32);
		add_arg(&fninfo, ARG_SGPR, ctx->i32);
		add_arg(&fninfo, ARG_SGPR, ctx->i32);
		add_arg(&fninfo, ARG_SGPR, ctx->ac.intptr);
		add_arg(&fninfo, ARG_SGPR, ctx->ac.intptr);
		add_arg(&fninfo, ARG_SGPR, ctx->ac.intptr);
		add_arg(&fninfo, ARG_SGPR, ctx->ac.intptr);
		add_arg(&fninfo, ARG_SGPR, ctx->i32);
		add_arg(&fninfo, ARG_SGPR, ctx->i32);
		add_arg(&fninfo, ARG_SGPR, ctx->i32);
		add_arg(&fninfo, ARG_SGPR, ctx->i32);
		ctx->param_tcs_offchip_layout = add_arg(&fninfo, ARG_SGPR, ctx->i32);
		add_arg(&fninfo, ARG_SGPR, ctx->i32);
		ctx->param_tcs_out_lds_layout = add_arg(&fninfo, ARG_SGPR, ctx->i32);
	} else {
		add_arg(&fninfo, ARG_SGPR, ctx->ac.intptr);
		add_arg(&fninfo, ARG_SGPR, ctx->ac.intptr);
		add_arg(&fninfo, ARG_SGPR, ctx->ac.intptr);
		add_arg(&fninfo, ARG_SGPR, ctx->ac.intptr);
		ctx->param_tcs_offchip_layout = add_arg(&fninfo, ARG_SGPR, ctx->i32);
		add_arg(&fninfo, ARG_SGPR, ctx->i32);
		ctx->param_tcs_out_lds_layout = add_arg(&fninfo, ARG_SGPR, ctx->i32);
		add_arg(&fninfo, ARG_SGPR, ctx->i32);
		ctx->param_tcs_offchip_offset = add_arg(&fninfo, ARG_SGPR, ctx->i32);
		ctx->param_tcs_factor_offset = add_arg(&fninfo, ARG_SGPR, ctx->i32);
	}

	add_arg(&fninfo, ARG_VGPR, ctx->i32); /* VGPR gap */
	add_arg(&fninfo, ARG_VGPR, ctx->i32); /* VGPR gap */
	unsigned tess_factors_idx =
		add_arg(&fninfo, ARG_VGPR, ctx->i32); /* patch index within the wave (REL_PATCH_ID) */
	add_arg(&fninfo, ARG_VGPR, ctx->i32); /* invocation ID within the patch */
	add_arg(&fninfo, ARG_VGPR, ctx->i32); /* LDS offset where tess factors should be loaded from */

	for (unsigned i = 0; i < 6; i++)
		add_arg(&fninfo, ARG_VGPR, ctx->i32); /* tess factors */

	/* Create the function. */
	si_create_function(ctx, "tcs_epilog", NULL, 0, &fninfo,
			   ctx->screen->info.chip_class >= GFX7 ? 128 : 64);
	ac_declare_lds_as_pointer(&ctx->ac);
	func = ctx->main_fn;

	LLVMValueRef invoc0_tess_factors[6];
	for (unsigned i = 0; i < 6; i++)
		invoc0_tess_factors[i] = LLVMGetParam(func, tess_factors_idx + 3 + i);

	si_write_tess_factors(bld_base,
			      LLVMGetParam(func, tess_factors_idx),
			      LLVMGetParam(func, tess_factors_idx + 1),
			      LLVMGetParam(func, tess_factors_idx + 2),
			      invoc0_tess_factors, invoc0_tess_factors + 4);

	LLVMBuildRetVoid(ctx->ac.builder);
}

/**
 * Select and compile (or reuse) TCS parts (epilog).
 */
static bool si_shader_select_tcs_parts(struct si_screen *sscreen,
				       struct ac_llvm_compiler *compiler,
				       struct si_shader *shader,
				       struct pipe_debug_callback *debug)
{
	if (sscreen->info.chip_class >= GFX9) {
		struct si_shader *ls_main_part =
			shader->key.part.tcs.ls->main_shader_part_ls;

		if (!si_get_vs_prolog(sscreen, compiler, shader, debug, ls_main_part,
				      &shader->key.part.tcs.ls_prolog))
			return false;

		shader->previous_stage = ls_main_part;
	}

	/* Get the epilog. */
	union si_shader_part_key epilog_key;
	memset(&epilog_key, 0, sizeof(epilog_key));
	epilog_key.tcs_epilog.states = shader->key.part.tcs.epilog;

	shader->epilog = si_get_shader_part(sscreen, &sscreen->tcs_epilogs,
					    PIPE_SHADER_TESS_CTRL, false,
					    &epilog_key, compiler, debug,
					    si_build_tcs_epilog_function,
					    "Tessellation Control Shader Epilog");
	return shader->epilog != NULL;
}

/**
 * Select and compile (or reuse) GS parts (prolog).
 */
static bool si_shader_select_gs_parts(struct si_screen *sscreen,
				      struct ac_llvm_compiler *compiler,
				      struct si_shader *shader,
				      struct pipe_debug_callback *debug)
{
	if (sscreen->info.chip_class >= GFX9) {
		struct si_shader *es_main_part =
			shader->key.part.gs.es->main_shader_part_es;

		if (shader->key.part.gs.es->type == PIPE_SHADER_VERTEX &&
		    !si_get_vs_prolog(sscreen, compiler, shader, debug, es_main_part,
				      &shader->key.part.gs.vs_prolog))
			return false;

		shader->previous_stage = es_main_part;
	}

	if (!shader->key.part.gs.prolog.tri_strip_adj_fix)
		return true;

	union si_shader_part_key prolog_key;
	memset(&prolog_key, 0, sizeof(prolog_key));
	prolog_key.gs_prolog.states = shader->key.part.gs.prolog;

	shader->prolog2 = si_get_shader_part(sscreen, &sscreen->gs_prologs,
					    PIPE_SHADER_GEOMETRY, true,
					    &prolog_key, compiler, debug,
					    si_build_gs_prolog_function,
					    "Geometry Shader Prolog");
	return shader->prolog2 != NULL;
}

/**
 * Build the pixel shader prolog function. This handles:
 * - two-side color selection and interpolation
 * - overriding interpolation parameters for the API PS
 * - polygon stippling
 *
 * All preloaded SGPRs and VGPRs are passed through unmodified unless they are
 * overriden by other states. (e.g. per-sample interpolation)
 * Interpolated colors are stored after the preloaded VGPRs.
 */
static void si_build_ps_prolog_function(struct si_shader_context *ctx,
					union si_shader_part_key *key)
{
	struct si_function_info fninfo;
	LLVMValueRef ret, func;
	int num_returns, i, num_color_channels;

	assert(si_need_ps_prolog(key));

	si_init_function_info(&fninfo);

	/* Declare inputs. */
	for (i = 0; i < key->ps_prolog.num_input_sgprs; i++)
		add_arg(&fninfo, ARG_SGPR, ctx->i32);

	for (i = 0; i < key->ps_prolog.num_input_vgprs; i++)
		add_arg(&fninfo, ARG_VGPR, ctx->f32);

	/* Declare outputs (same as inputs + add colors if needed) */
	num_returns = fninfo.num_params;
	num_color_channels = util_bitcount(key->ps_prolog.colors_read);
	for (i = 0; i < num_color_channels; i++)
		fninfo.types[num_returns++] = ctx->f32;

	/* Create the function. */
	si_create_function(ctx, "ps_prolog", fninfo.types, num_returns,
			   &fninfo, 0);
	func = ctx->main_fn;

	/* Copy inputs to outputs. This should be no-op, as the registers match,
	 * but it will prevent the compiler from overwriting them unintentionally.
	 */
	ret = ctx->return_value;
	for (i = 0; i < fninfo.num_params; i++) {
		LLVMValueRef p = LLVMGetParam(func, i);
		ret = LLVMBuildInsertValue(ctx->ac.builder, ret, p, i, "");
	}

	/* Polygon stippling. */
	if (key->ps_prolog.states.poly_stipple) {
		/* POS_FIXED_PT is always last. */
		unsigned pos = key->ps_prolog.num_input_sgprs +
			       key->ps_prolog.num_input_vgprs - 1;
		LLVMValueRef list = si_prolog_get_rw_buffers(ctx);

		si_llvm_emit_polygon_stipple(ctx, list, pos);
	}

	if (key->ps_prolog.states.bc_optimize_for_persp ||
	    key->ps_prolog.states.bc_optimize_for_linear) {
		unsigned i, base = key->ps_prolog.num_input_sgprs;
		LLVMValueRef center[2], centroid[2], tmp, bc_optimize;

		/* The shader should do: if (PRIM_MASK[31]) CENTROID = CENTER;
		 * The hw doesn't compute CENTROID if the whole wave only
		 * contains fully-covered quads.
		 *
		 * PRIM_MASK is after user SGPRs.
		 */
		bc_optimize = LLVMGetParam(func, SI_PS_NUM_USER_SGPR);
		bc_optimize = LLVMBuildLShr(ctx->ac.builder, bc_optimize,
					    LLVMConstInt(ctx->i32, 31, 0), "");
		bc_optimize = LLVMBuildTrunc(ctx->ac.builder, bc_optimize,
					     ctx->i1, "");

		if (key->ps_prolog.states.bc_optimize_for_persp) {
			/* Read PERSP_CENTER. */
			for (i = 0; i < 2; i++)
				center[i] = LLVMGetParam(func, base + 2 + i);
			/* Read PERSP_CENTROID. */
			for (i = 0; i < 2; i++)
				centroid[i] = LLVMGetParam(func, base + 4 + i);
			/* Select PERSP_CENTROID. */
			for (i = 0; i < 2; i++) {
				tmp = LLVMBuildSelect(ctx->ac.builder, bc_optimize,
						      center[i], centroid[i], "");
				ret = LLVMBuildInsertValue(ctx->ac.builder, ret,
							   tmp, base + 4 + i, "");
			}
		}
		if (key->ps_prolog.states.bc_optimize_for_linear) {
			/* Read LINEAR_CENTER. */
			for (i = 0; i < 2; i++)
				center[i] = LLVMGetParam(func, base + 8 + i);
			/* Read LINEAR_CENTROID. */
			for (i = 0; i < 2; i++)
				centroid[i] = LLVMGetParam(func, base + 10 + i);
			/* Select LINEAR_CENTROID. */
			for (i = 0; i < 2; i++) {
				tmp = LLVMBuildSelect(ctx->ac.builder, bc_optimize,
						      center[i], centroid[i], "");
				ret = LLVMBuildInsertValue(ctx->ac.builder, ret,
							   tmp, base + 10 + i, "");
			}
		}
	}

	/* Force per-sample interpolation. */
	if (key->ps_prolog.states.force_persp_sample_interp) {
		unsigned i, base = key->ps_prolog.num_input_sgprs;
		LLVMValueRef persp_sample[2];

		/* Read PERSP_SAMPLE. */
		for (i = 0; i < 2; i++)
			persp_sample[i] = LLVMGetParam(func, base + i);
		/* Overwrite PERSP_CENTER. */
		for (i = 0; i < 2; i++)
			ret = LLVMBuildInsertValue(ctx->ac.builder, ret,
						   persp_sample[i], base + 2 + i, "");
		/* Overwrite PERSP_CENTROID. */
		for (i = 0; i < 2; i++)
			ret = LLVMBuildInsertValue(ctx->ac.builder, ret,
						   persp_sample[i], base + 4 + i, "");
	}
	if (key->ps_prolog.states.force_linear_sample_interp) {
		unsigned i, base = key->ps_prolog.num_input_sgprs;
		LLVMValueRef linear_sample[2];

		/* Read LINEAR_SAMPLE. */
		for (i = 0; i < 2; i++)
			linear_sample[i] = LLVMGetParam(func, base + 6 + i);
		/* Overwrite LINEAR_CENTER. */
		for (i = 0; i < 2; i++)
			ret = LLVMBuildInsertValue(ctx->ac.builder, ret,
						   linear_sample[i], base + 8 + i, "");
		/* Overwrite LINEAR_CENTROID. */
		for (i = 0; i < 2; i++)
			ret = LLVMBuildInsertValue(ctx->ac.builder, ret,
						   linear_sample[i], base + 10 + i, "");
	}

	/* Force center interpolation. */
	if (key->ps_prolog.states.force_persp_center_interp) {
		unsigned i, base = key->ps_prolog.num_input_sgprs;
		LLVMValueRef persp_center[2];

		/* Read PERSP_CENTER. */
		for (i = 0; i < 2; i++)
			persp_center[i] = LLVMGetParam(func, base + 2 + i);
		/* Overwrite PERSP_SAMPLE. */
		for (i = 0; i < 2; i++)
			ret = LLVMBuildInsertValue(ctx->ac.builder, ret,
						   persp_center[i], base + i, "");
		/* Overwrite PERSP_CENTROID. */
		for (i = 0; i < 2; i++)
			ret = LLVMBuildInsertValue(ctx->ac.builder, ret,
						   persp_center[i], base + 4 + i, "");
	}
	if (key->ps_prolog.states.force_linear_center_interp) {
		unsigned i, base = key->ps_prolog.num_input_sgprs;
		LLVMValueRef linear_center[2];

		/* Read LINEAR_CENTER. */
		for (i = 0; i < 2; i++)
			linear_center[i] = LLVMGetParam(func, base + 8 + i);
		/* Overwrite LINEAR_SAMPLE. */
		for (i = 0; i < 2; i++)
			ret = LLVMBuildInsertValue(ctx->ac.builder, ret,
						   linear_center[i], base + 6 + i, "");
		/* Overwrite LINEAR_CENTROID. */
		for (i = 0; i < 2; i++)
			ret = LLVMBuildInsertValue(ctx->ac.builder, ret,
						   linear_center[i], base + 10 + i, "");
	}

	/* Interpolate colors. */
	unsigned color_out_idx = 0;
	for (i = 0; i < 2; i++) {
		unsigned writemask = (key->ps_prolog.colors_read >> (i * 4)) & 0xf;
		unsigned face_vgpr = key->ps_prolog.num_input_sgprs +
				     key->ps_prolog.face_vgpr_index;
		LLVMValueRef interp[2], color[4];
		LLVMValueRef interp_ij = NULL, prim_mask = NULL, face = NULL;

		if (!writemask)
			continue;

		/* If the interpolation qualifier is not CONSTANT (-1). */
		if (key->ps_prolog.color_interp_vgpr_index[i] != -1) {
			unsigned interp_vgpr = key->ps_prolog.num_input_sgprs +
					       key->ps_prolog.color_interp_vgpr_index[i];

			/* Get the (i,j) updated by bc_optimize handling. */
			interp[0] = LLVMBuildExtractValue(ctx->ac.builder, ret,
							  interp_vgpr, "");
			interp[1] = LLVMBuildExtractValue(ctx->ac.builder, ret,
							  interp_vgpr + 1, "");
			interp_ij = ac_build_gather_values(&ctx->ac, interp, 2);
		}

		/* Use the absolute location of the input. */
		prim_mask = LLVMGetParam(func, SI_PS_NUM_USER_SGPR);

		if (key->ps_prolog.states.color_two_side) {
			face = LLVMGetParam(func, face_vgpr);
			face = ac_to_integer(&ctx->ac, face);
		}

		interp_fs_input(ctx,
				key->ps_prolog.color_attr_index[i],
				TGSI_SEMANTIC_COLOR, i,
				key->ps_prolog.num_interp_inputs,
				key->ps_prolog.colors_read, interp_ij,
				prim_mask, face, color);

		while (writemask) {
			unsigned chan = u_bit_scan(&writemask);
			ret = LLVMBuildInsertValue(ctx->ac.builder, ret, color[chan],
						   fninfo.num_params + color_out_idx++, "");
		}
	}

	/* Section 15.2.2 (Shader Inputs) of the OpenGL 4.5 (Core Profile) spec
	 * says:
	 *
	 *    "When per-sample shading is active due to the use of a fragment
	 *     input qualified by sample or due to the use of the gl_SampleID
	 *     or gl_SamplePosition variables, only the bit for the current
	 *     sample is set in gl_SampleMaskIn. When state specifies multiple
	 *     fragment shader invocations for a given fragment, the sample
	 *     mask for any single fragment shader invocation may specify a
	 *     subset of the covered samples for the fragment. In this case,
	 *     the bit corresponding to each covered sample will be set in
	 *     exactly one fragment shader invocation."
	 *
	 * The samplemask loaded by hardware is always the coverage of the
	 * entire pixel/fragment, so mask bits out based on the sample ID.
	 */
	if (key->ps_prolog.states.samplemask_log_ps_iter) {
		/* The bit pattern matches that used by fixed function fragment
		 * processing. */
		static const uint16_t ps_iter_masks[] = {
			0xffff, /* not used */
			0x5555,
			0x1111,
			0x0101,
			0x0001,
		};
		assert(key->ps_prolog.states.samplemask_log_ps_iter < ARRAY_SIZE(ps_iter_masks));

		uint32_t ps_iter_mask = ps_iter_masks[key->ps_prolog.states.samplemask_log_ps_iter];
		unsigned ancillary_vgpr = key->ps_prolog.num_input_sgprs +
					  key->ps_prolog.ancillary_vgpr_index;
		LLVMValueRef sampleid = si_unpack_param(ctx, ancillary_vgpr, 8, 4);
		LLVMValueRef samplemask = LLVMGetParam(func, ancillary_vgpr + 1);

		samplemask = ac_to_integer(&ctx->ac, samplemask);
		samplemask = LLVMBuildAnd(
			ctx->ac.builder,
			samplemask,
			LLVMBuildShl(ctx->ac.builder,
				     LLVMConstInt(ctx->i32, ps_iter_mask, false),
				     sampleid, ""),
			"");
		samplemask = ac_to_float(&ctx->ac, samplemask);

		ret = LLVMBuildInsertValue(ctx->ac.builder, ret, samplemask,
					   ancillary_vgpr + 1, "");
	}

	/* Tell LLVM to insert WQM instruction sequence when needed. */
	if (key->ps_prolog.wqm) {
		LLVMAddTargetDependentFunctionAttr(func,
						   "amdgpu-ps-wqm-outputs", "");
	}

	si_llvm_build_ret(ctx, ret);
}

/**
 * Build the pixel shader epilog function. This handles everything that must be
 * emulated for pixel shader exports. (alpha-test, format conversions, etc)
 */
static void si_build_ps_epilog_function(struct si_shader_context *ctx,
					union si_shader_part_key *key)
{
	struct lp_build_tgsi_context *bld_base = &ctx->bld_base;
	struct si_function_info fninfo;
	LLVMValueRef depth = NULL, stencil = NULL, samplemask = NULL;
	int i;
	struct si_ps_exports exp = {};

	si_init_function_info(&fninfo);

	/* Declare input SGPRs. */
	ctx->param_rw_buffers = add_arg(&fninfo, ARG_SGPR, ctx->ac.intptr);
	ctx->param_bindless_samplers_and_images = add_arg(&fninfo, ARG_SGPR, ctx->ac.intptr);
	ctx->param_const_and_shader_buffers = add_arg(&fninfo, ARG_SGPR, ctx->ac.intptr);
	ctx->param_samplers_and_images = add_arg(&fninfo, ARG_SGPR, ctx->ac.intptr);
	add_arg_checked(&fninfo, ARG_SGPR, ctx->f32, SI_PARAM_ALPHA_REF);

	/* Declare input VGPRs. */
	unsigned required_num_params =
		     fninfo.num_sgpr_params +
		     util_bitcount(key->ps_epilog.colors_written) * 4 +
		     key->ps_epilog.writes_z +
		     key->ps_epilog.writes_stencil +
		     key->ps_epilog.writes_samplemask;

	required_num_params = MAX2(required_num_params,
				   fninfo.num_sgpr_params + PS_EPILOG_SAMPLEMASK_MIN_LOC + 1);

	while (fninfo.num_params < required_num_params)
		add_arg(&fninfo, ARG_VGPR, ctx->f32);

	/* Create the function. */
	si_create_function(ctx, "ps_epilog", NULL, 0, &fninfo, 0);
	/* Disable elimination of unused inputs. */
	ac_llvm_add_target_dep_function_attr(ctx->main_fn,
					     "InitialPSInputAddr", 0xffffff);

	/* Process colors. */
	unsigned vgpr = fninfo.num_sgpr_params;
	unsigned colors_written = key->ps_epilog.colors_written;
	int last_color_export = -1;

	/* Find the last color export. */
	if (!key->ps_epilog.writes_z &&
	    !key->ps_epilog.writes_stencil &&
	    !key->ps_epilog.writes_samplemask) {
		unsigned spi_format = key->ps_epilog.states.spi_shader_col_format;

		/* If last_cbuf > 0, FS_COLOR0_WRITES_ALL_CBUFS is true. */
		if (colors_written == 0x1 && key->ps_epilog.states.last_cbuf > 0) {
			/* Just set this if any of the colorbuffers are enabled. */
			if (spi_format &
			    ((1ull << (4 * (key->ps_epilog.states.last_cbuf + 1))) - 1))
				last_color_export = 0;
		} else {
			for (i = 0; i < 8; i++)
				if (colors_written & (1 << i) &&
				    (spi_format >> (i * 4)) & 0xf)
					last_color_export = i;
		}
	}

	while (colors_written) {
		LLVMValueRef color[4];
		int mrt = u_bit_scan(&colors_written);

		for (i = 0; i < 4; i++)
			color[i] = LLVMGetParam(ctx->main_fn, vgpr++);

		si_export_mrt_color(bld_base, color, mrt,
				    fninfo.num_params - 1,
				    mrt == last_color_export, &exp);
	}

	/* Process depth, stencil, samplemask. */
	if (key->ps_epilog.writes_z)
		depth = LLVMGetParam(ctx->main_fn, vgpr++);
	if (key->ps_epilog.writes_stencil)
		stencil = LLVMGetParam(ctx->main_fn, vgpr++);
	if (key->ps_epilog.writes_samplemask)
		samplemask = LLVMGetParam(ctx->main_fn, vgpr++);

	if (depth || stencil || samplemask)
		si_export_mrt_z(bld_base, depth, stencil, samplemask, &exp);
	else if (last_color_export == -1)
		ac_build_export_null(&ctx->ac);

	if (exp.num)
		si_emit_ps_exports(ctx, &exp);

	/* Compile. */
	LLVMBuildRetVoid(ctx->ac.builder);
}

/**
 * Select and compile (or reuse) pixel shader parts (prolog & epilog).
 */
static bool si_shader_select_ps_parts(struct si_screen *sscreen,
				      struct ac_llvm_compiler *compiler,
				      struct si_shader *shader,
				      struct pipe_debug_callback *debug)
{
	union si_shader_part_key prolog_key;
	union si_shader_part_key epilog_key;

	/* Get the prolog. */
	si_get_ps_prolog_key(shader, &prolog_key, true);

	/* The prolog is a no-op if these aren't set. */
	if (si_need_ps_prolog(&prolog_key)) {
		shader->prolog =
			si_get_shader_part(sscreen, &sscreen->ps_prologs,
					   PIPE_SHADER_FRAGMENT, true,
					   &prolog_key, compiler, debug,
					   si_build_ps_prolog_function,
					   "Fragment Shader Prolog");
		if (!shader->prolog)
			return false;
	}

	/* Get the epilog. */
	si_get_ps_epilog_key(shader, &epilog_key);

	shader->epilog =
		si_get_shader_part(sscreen, &sscreen->ps_epilogs,
				   PIPE_SHADER_FRAGMENT, false,
				   &epilog_key, compiler, debug,
				   si_build_ps_epilog_function,
				   "Fragment Shader Epilog");
	if (!shader->epilog)
		return false;

	/* Enable POS_FIXED_PT if polygon stippling is enabled. */
	if (shader->key.part.ps.prolog.poly_stipple) {
		shader->config.spi_ps_input_ena |= S_0286CC_POS_FIXED_PT_ENA(1);
		assert(G_0286CC_POS_FIXED_PT_ENA(shader->config.spi_ps_input_addr));
	}

	/* Set up the enable bits for per-sample shading if needed. */
	if (shader->key.part.ps.prolog.force_persp_sample_interp &&
	    (G_0286CC_PERSP_CENTER_ENA(shader->config.spi_ps_input_ena) ||
	     G_0286CC_PERSP_CENTROID_ENA(shader->config.spi_ps_input_ena))) {
		shader->config.spi_ps_input_ena &= C_0286CC_PERSP_CENTER_ENA;
		shader->config.spi_ps_input_ena &= C_0286CC_PERSP_CENTROID_ENA;
		shader->config.spi_ps_input_ena |= S_0286CC_PERSP_SAMPLE_ENA(1);
	}
	if (shader->key.part.ps.prolog.force_linear_sample_interp &&
	    (G_0286CC_LINEAR_CENTER_ENA(shader->config.spi_ps_input_ena) ||
	     G_0286CC_LINEAR_CENTROID_ENA(shader->config.spi_ps_input_ena))) {
		shader->config.spi_ps_input_ena &= C_0286CC_LINEAR_CENTER_ENA;
		shader->config.spi_ps_input_ena &= C_0286CC_LINEAR_CENTROID_ENA;
		shader->config.spi_ps_input_ena |= S_0286CC_LINEAR_SAMPLE_ENA(1);
	}
	if (shader->key.part.ps.prolog.force_persp_center_interp &&
	    (G_0286CC_PERSP_SAMPLE_ENA(shader->config.spi_ps_input_ena) ||
	     G_0286CC_PERSP_CENTROID_ENA(shader->config.spi_ps_input_ena))) {
		shader->config.spi_ps_input_ena &= C_0286CC_PERSP_SAMPLE_ENA;
		shader->config.spi_ps_input_ena &= C_0286CC_PERSP_CENTROID_ENA;
		shader->config.spi_ps_input_ena |= S_0286CC_PERSP_CENTER_ENA(1);
	}
	if (shader->key.part.ps.prolog.force_linear_center_interp &&
	    (G_0286CC_LINEAR_SAMPLE_ENA(shader->config.spi_ps_input_ena) ||
	     G_0286CC_LINEAR_CENTROID_ENA(shader->config.spi_ps_input_ena))) {
		shader->config.spi_ps_input_ena &= C_0286CC_LINEAR_SAMPLE_ENA;
		shader->config.spi_ps_input_ena &= C_0286CC_LINEAR_CENTROID_ENA;
		shader->config.spi_ps_input_ena |= S_0286CC_LINEAR_CENTER_ENA(1);
	}

	/* POW_W_FLOAT requires that one of the perspective weights is enabled. */
	if (G_0286CC_POS_W_FLOAT_ENA(shader->config.spi_ps_input_ena) &&
	    !(shader->config.spi_ps_input_ena & 0xf)) {
		shader->config.spi_ps_input_ena |= S_0286CC_PERSP_CENTER_ENA(1);
		assert(G_0286CC_PERSP_CENTER_ENA(shader->config.spi_ps_input_addr));
	}

	/* At least one pair of interpolation weights must be enabled. */
	if (!(shader->config.spi_ps_input_ena & 0x7f)) {
		shader->config.spi_ps_input_ena |= S_0286CC_LINEAR_CENTER_ENA(1);
		assert(G_0286CC_LINEAR_CENTER_ENA(shader->config.spi_ps_input_addr));
	}

	/* Samplemask fixup requires the sample ID. */
	if (shader->key.part.ps.prolog.samplemask_log_ps_iter) {
		shader->config.spi_ps_input_ena |= S_0286CC_ANCILLARY_ENA(1);
		assert(G_0286CC_ANCILLARY_ENA(shader->config.spi_ps_input_addr));
	}

	/* The sample mask input is always enabled, because the API shader always
	 * passes it through to the epilog. Disable it here if it's unused.
	 */
	if (!shader->key.part.ps.epilog.poly_line_smoothing &&
	    !shader->selector->info.reads_samplemask)
		shader->config.spi_ps_input_ena &= C_0286CC_SAMPLE_COVERAGE_ENA;

	return true;
}

void si_multiwave_lds_size_workaround(struct si_screen *sscreen,
				      unsigned *lds_size)
{
	/* If tessellation is all offchip and on-chip GS isn't used, this
	 * workaround is not needed.
	 */
	return;

	/* SPI barrier management bug:
	 *   Make sure we have at least 4k of LDS in use to avoid the bug.
	 *   It applies to workgroup sizes of more than one wavefront.
	 */
	if (sscreen->info.family == CHIP_BONAIRE ||
	    sscreen->info.family == CHIP_KABINI ||
	    sscreen->info.family == CHIP_MULLINS)
		*lds_size = MAX2(*lds_size, 8);
}

static void si_fix_resource_usage(struct si_screen *sscreen,
				  struct si_shader *shader)
{
	unsigned min_sgprs = shader->info.num_input_sgprs + 2; /* VCC */

	shader->config.num_sgprs = MAX2(shader->config.num_sgprs, min_sgprs);

	if (shader->selector->type == PIPE_SHADER_COMPUTE &&
	    si_get_max_workgroup_size(shader) > 64) {
		si_multiwave_lds_size_workaround(sscreen,
						 &shader->config.lds_size);
	}
}

int si_shader_create(struct si_screen *sscreen, struct ac_llvm_compiler *compiler,
		     struct si_shader *shader,
		     struct pipe_debug_callback *debug)
{
	struct si_shader_selector *sel = shader->selector;
	struct si_shader *mainp = *si_get_main_shader_part(sel, &shader->key);
	int r;

	/* LS, ES, VS are compiled on demand if the main part hasn't been
	 * compiled for that stage.
	 *
	 * Vertex shaders are compiled on demand when a vertex fetch
	 * workaround must be applied.
	 */
	if (shader->is_monolithic) {
		/* Monolithic shader (compiled as a whole, has many variants,
		 * may take a long time to compile).
		 */
		r = si_compile_tgsi_shader(sscreen, compiler, shader, debug);
		if (r)
			return r;
	} else {
		/* The shader consists of several parts:
		 *
		 * - the middle part is the user shader, it has 1 variant only
		 *   and it was compiled during the creation of the shader
		 *   selector
		 * - the prolog part is inserted at the beginning
		 * - the epilog part is inserted at the end
		 *
		 * The prolog and epilog have many (but simple) variants.
		 *
		 * Starting with gfx9, geometry and tessellation control
		 * shaders also contain the prolog and user shader parts of
		 * the previous shader stage.
		 */

		if (!mainp)
			return -1;

		/* Copy the compiled TGSI shader data over. */
		shader->is_binary_shared = true;
		shader->binary = mainp->binary;
		shader->config = mainp->config;
		shader->info.num_input_sgprs = mainp->info.num_input_sgprs;
		shader->info.num_input_vgprs = mainp->info.num_input_vgprs;
		shader->info.face_vgpr_index = mainp->info.face_vgpr_index;
		shader->info.ancillary_vgpr_index = mainp->info.ancillary_vgpr_index;
		memcpy(shader->info.vs_output_param_offset,
		       mainp->info.vs_output_param_offset,
		       sizeof(mainp->info.vs_output_param_offset));
		shader->info.uses_instanceid = mainp->info.uses_instanceid;
		shader->info.nr_pos_exports = mainp->info.nr_pos_exports;
		shader->info.nr_param_exports = mainp->info.nr_param_exports;

		/* Select prologs and/or epilogs. */
		switch (sel->type) {
		case PIPE_SHADER_VERTEX:
			if (!si_shader_select_vs_parts(sscreen, compiler, shader, debug))
				return -1;
			break;
		case PIPE_SHADER_TESS_CTRL:
			if (!si_shader_select_tcs_parts(sscreen, compiler, shader, debug))
				return -1;
			break;
		case PIPE_SHADER_TESS_EVAL:
			break;
		case PIPE_SHADER_GEOMETRY:
			if (!si_shader_select_gs_parts(sscreen, compiler, shader, debug))
				return -1;
			break;
		case PIPE_SHADER_FRAGMENT:
			if (!si_shader_select_ps_parts(sscreen, compiler, shader, debug))
				return -1;

			/* Make sure we have at least as many VGPRs as there
			 * are allocated inputs.
			 */
			shader->config.num_vgprs = MAX2(shader->config.num_vgprs,
							shader->info.num_input_vgprs);
			break;
		}

		/* Update SGPR and VGPR counts. */
		if (shader->prolog) {
			shader->config.num_sgprs = MAX2(shader->config.num_sgprs,
							shader->prolog->config.num_sgprs);
			shader->config.num_vgprs = MAX2(shader->config.num_vgprs,
							shader->prolog->config.num_vgprs);
		}
		if (shader->previous_stage) {
			shader->config.num_sgprs = MAX2(shader->config.num_sgprs,
							shader->previous_stage->config.num_sgprs);
			shader->config.num_vgprs = MAX2(shader->config.num_vgprs,
							shader->previous_stage->config.num_vgprs);
			shader->config.spilled_sgprs =
				MAX2(shader->config.spilled_sgprs,
				     shader->previous_stage->config.spilled_sgprs);
			shader->config.spilled_vgprs =
				MAX2(shader->config.spilled_vgprs,
				     shader->previous_stage->config.spilled_vgprs);
			shader->config.private_mem_vgprs =
				MAX2(shader->config.private_mem_vgprs,
				     shader->previous_stage->config.private_mem_vgprs);
			shader->config.scratch_bytes_per_wave =
				MAX2(shader->config.scratch_bytes_per_wave,
				     shader->previous_stage->config.scratch_bytes_per_wave);
			shader->info.uses_instanceid |=
				shader->previous_stage->info.uses_instanceid;
		}
		if (shader->prolog2) {
			shader->config.num_sgprs = MAX2(shader->config.num_sgprs,
							shader->prolog2->config.num_sgprs);
			shader->config.num_vgprs = MAX2(shader->config.num_vgprs,
							shader->prolog2->config.num_vgprs);
		}
		if (shader->epilog) {
			shader->config.num_sgprs = MAX2(shader->config.num_sgprs,
							shader->epilog->config.num_sgprs);
			shader->config.num_vgprs = MAX2(shader->config.num_vgprs,
							shader->epilog->config.num_vgprs);
		}
		si_calculate_max_simd_waves(shader);
	}

	si_fix_resource_usage(sscreen, shader);
	si_shader_dump(sscreen, shader, debug, sel->info.processor,
		       stderr, true);

	/* Upload. */
	r = si_shader_binary_upload(sscreen, shader);
	if (r) {
		fprintf(stderr, "LLVM failed to upload shader\n");
		return r;
	}

	return 0;
}

void si_shader_destroy(struct si_shader *shader)
{
	if (shader->scratch_bo)
		si_resource_reference(&shader->scratch_bo, NULL);

	si_resource_reference(&shader->bo, NULL);

	if (!shader->is_binary_shared)
		ac_shader_binary_clean(&shader->binary);

	free(shader->shader_log);
}