summaryrefslogtreecommitdiff
path: root/src/panfrost/midgard/midgard_compile.c
blob: 5704f8475763119408b1849e28ee16abf9a51428 (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
/*
 * Copyright (C) 2018-2019 Alyssa Rosenzweig <alyssa@rosenzweig.io>
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
 * copy of this software and associated documentation files (the "Software"),
 * to deal in the Software without restriction, including without limitation
 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
 * and/or sell copies of the Software, and to permit persons to whom the
 * Software is furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice (including the next
 * paragraph) shall be included in all copies or substantial portions of the
 * Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 * SOFTWARE.
 */

#include <sys/types.h>
#include <sys/stat.h>
#include <sys/mman.h>
#include <fcntl.h>
#include <stdint.h>
#include <stdlib.h>
#include <stdio.h>
#include <err.h>

#include "main/mtypes.h"
#include "compiler/glsl/glsl_to_nir.h"
#include "compiler/nir_types.h"
#include "main/imports.h"
#include "compiler/nir/nir_builder.h"
#include "util/half_float.h"
#include "util/u_math.h"
#include "util/u_debug.h"
#include "util/u_dynarray.h"
#include "util/list.h"
#include "main/mtypes.h"

#include "midgard.h"
#include "midgard_nir.h"
#include "midgard_compile.h"
#include "midgard_ops.h"
#include "helpers.h"
#include "compiler.h"

#include "disassemble.h"

static const struct debug_named_value debug_options[] = {
        {"msgs",      MIDGARD_DBG_MSGS,		"Print debug messages"},
        {"shaders",   MIDGARD_DBG_SHADERS,	"Dump shaders in NIR and MIR"},
        {"shaderdb",  MIDGARD_DBG_SHADERDB,     "Prints shader-db statistics"},
        DEBUG_NAMED_VALUE_END
};

DEBUG_GET_ONCE_FLAGS_OPTION(midgard_debug, "MIDGARD_MESA_DEBUG", debug_options, 0)

unsigned SHADER_DB_COUNT = 0;

int midgard_debug = 0;

#define DBG(fmt, ...) \
		do { if (midgard_debug & MIDGARD_DBG_MSGS) \
			fprintf(stderr, "%s:%d: "fmt, \
				__FUNCTION__, __LINE__, ##__VA_ARGS__); } while (0)

static bool
midgard_is_branch_unit(unsigned unit)
{
        return (unit == ALU_ENAB_BRANCH) || (unit == ALU_ENAB_BR_COMPACT);
}

static midgard_block *
create_empty_block(compiler_context *ctx)
{
        midgard_block *blk = rzalloc(ctx, midgard_block);

        blk->predecessors = _mesa_set_create(blk,
                        _mesa_hash_pointer,
                        _mesa_key_pointer_equal);

        blk->source_id = ctx->block_source_count++;

        return blk;
}

static void
midgard_block_add_successor(midgard_block *block, midgard_block *successor)
{
        assert(block);
        assert(successor);

        /* Deduplicate */
        for (unsigned i = 0; i < block->nr_successors; ++i) {
                if (block->successors[i] == successor)
                        return;
        }

        block->successors[block->nr_successors++] = successor;
        assert(block->nr_successors <= ARRAY_SIZE(block->successors));

        /* Note the predecessor in the other direction */
        _mesa_set_add(successor->predecessors, block);
}

static void
schedule_barrier(compiler_context *ctx)
{
        midgard_block *temp = ctx->after_block;
        ctx->after_block = create_empty_block(ctx);
        ctx->block_count++;
        list_addtail(&ctx->after_block->link, &ctx->blocks);
        list_inithead(&ctx->after_block->instructions);
        midgard_block_add_successor(ctx->current_block, ctx->after_block);
        ctx->current_block = ctx->after_block;
        ctx->after_block = temp;
}

/* Helpers to generate midgard_instruction's using macro magic, since every
 * driver seems to do it that way */

#define EMIT(op, ...) emit_mir_instruction(ctx, v_##op(__VA_ARGS__));

#define M_LOAD_STORE(name, store) \
	static midgard_instruction m_##name(unsigned ssa, unsigned address) { \
		midgard_instruction i = { \
			.type = TAG_LOAD_STORE_4, \
                        .mask = 0xF, \
                        .dest = ~0, \
                        .src = { ~0, ~0, ~0 }, \
                        .swizzle = SWIZZLE_IDENTITY_4, \
			.load_store = { \
				.op = midgard_op_##name, \
				.address = address \
			} \
		}; \
                \
                if (store) \
                        i.src[0] = ssa; \
                else \
                        i.dest = ssa; \
		\
		return i; \
	}

#define M_LOAD(name) M_LOAD_STORE(name, false)
#define M_STORE(name) M_LOAD_STORE(name, true)

/* Inputs a NIR ALU source, with modifiers attached if necessary, and outputs
 * the corresponding Midgard source */

static midgard_vector_alu_src
vector_alu_modifiers(nir_alu_src *src, bool is_int, unsigned broadcast_count,
                     bool half, bool sext)
{
        /* Figure out how many components there are so we can adjust.
         * Specifically we want to broadcast the last channel so things like
         * ball2/3 work.
         */

        if (broadcast_count && src) {
                uint8_t last_component = src->swizzle[broadcast_count - 1];

                for (unsigned c = broadcast_count; c < NIR_MAX_VEC_COMPONENTS; ++c) {
                        src->swizzle[c] = last_component;
                }
        }

        midgard_vector_alu_src alu_src = {
                .rep_low = 0,
                .rep_high = 0,
                .half = half
        };

        if (is_int) {
                alu_src.mod = midgard_int_normal;

                /* Sign/zero-extend if needed */

                if (half) {
                        alu_src.mod = sext ?
                                      midgard_int_sign_extend
                                      : midgard_int_zero_extend;
                }

                /* These should have been lowered away */
                if (src)
                        assert(!(src->abs || src->negate));
        } else {
                if (src)
                        alu_src.mod = (src->abs << 0) | (src->negate << 1);
        }

        return alu_src;
}

/* load/store instructions have both 32-bit and 16-bit variants, depending on
 * whether we are using vectors composed of highp or mediump. At the moment, we
 * don't support half-floats -- this requires changes in other parts of the
 * compiler -- therefore the 16-bit versions are commented out. */

//M_LOAD(ld_attr_16);
M_LOAD(ld_attr_32);
//M_LOAD(ld_vary_16);
M_LOAD(ld_vary_32);
M_LOAD(ld_ubo_int4);
M_LOAD(ld_int4);
M_STORE(st_int4);
M_LOAD(ld_color_buffer_8);
//M_STORE(st_vary_16);
M_STORE(st_vary_32);
M_LOAD(ld_cubemap_coords);
M_LOAD(ld_compute_id);

static midgard_instruction
v_alu_br_compact_cond(midgard_jmp_writeout_op op, unsigned tag, signed offset, unsigned cond)
{
        midgard_branch_cond branch = {
                .op = op,
                .dest_tag = tag,
                .offset = offset,
                .cond = cond
        };

        uint16_t compact;
        memcpy(&compact, &branch, sizeof(branch));

        midgard_instruction ins = {
                .type = TAG_ALU_4,
                .unit = ALU_ENAB_BR_COMPACT,
                .prepacked_branch = true,
                .compact_branch = true,
                .br_compact = compact,
                .dest = ~0,
                .src = { ~0, ~0, ~0 },
        };

        if (op == midgard_jmp_writeout_op_writeout)
                ins.writeout = true;

        return ins;
}

static midgard_instruction
v_branch(bool conditional, bool invert)
{
        midgard_instruction ins = {
                .type = TAG_ALU_4,
                .unit = ALU_ENAB_BRANCH,
                .compact_branch = true,
                .branch = {
                        .conditional = conditional,
                        .invert_conditional = invert
                },
                .dest = ~0,
                .src = { ~0, ~0, ~0 },
        };

        return ins;
}

static midgard_branch_extended
midgard_create_branch_extended( midgard_condition cond,
                                midgard_jmp_writeout_op op,
                                unsigned dest_tag,
                                signed quadword_offset)
{
        /* The condition code is actually a LUT describing a function to
         * combine multiple condition codes. However, we only support a single
         * condition code at the moment, so we just duplicate over a bunch of
         * times. */

        uint16_t duplicated_cond =
                (cond << 14) |
                (cond << 12) |
                (cond << 10) |
                (cond << 8) |
                (cond << 6) |
                (cond << 4) |
                (cond << 2) |
                (cond << 0);

        midgard_branch_extended branch = {
                .op = op,
                .dest_tag = dest_tag,
                .offset = quadword_offset,
                .cond = duplicated_cond
        };

        return branch;
}

static void
attach_constants(compiler_context *ctx, midgard_instruction *ins, void *constants, int name)
{
        ins->has_constants = true;
        memcpy(&ins->constants, constants, 16);
}

static int
glsl_type_size(const struct glsl_type *type, bool bindless)
{
        return glsl_count_attribute_slots(type, false);
}

/* Lower fdot2 to a vector multiplication followed by channel addition  */
static void
midgard_nir_lower_fdot2_body(nir_builder *b, nir_alu_instr *alu)
{
        if (alu->op != nir_op_fdot2)
                return;

        b->cursor = nir_before_instr(&alu->instr);

        nir_ssa_def *src0 = nir_ssa_for_alu_src(b, alu, 0);
        nir_ssa_def *src1 = nir_ssa_for_alu_src(b, alu, 1);

        nir_ssa_def *product = nir_fmul(b, src0, src1);

        nir_ssa_def *sum = nir_fadd(b,
                                    nir_channel(b, product, 0),
                                    nir_channel(b, product, 1));

        /* Replace the fdot2 with this sum */
        nir_ssa_def_rewrite_uses(&alu->dest.dest.ssa, nir_src_for_ssa(sum));
}

static int
midgard_sysval_for_ssbo(nir_intrinsic_instr *instr)
{
        /* This is way too meta */
        bool is_store = instr->intrinsic == nir_intrinsic_store_ssbo;
        unsigned idx_idx = is_store ? 1 : 0;

        nir_src index = instr->src[idx_idx];
        assert(nir_src_is_const(index));
        uint32_t uindex = nir_src_as_uint(index);

        return PAN_SYSVAL(SSBO, uindex);
}

static int
midgard_nir_sysval_for_intrinsic(nir_intrinsic_instr *instr)
{
        switch (instr->intrinsic) {
        case nir_intrinsic_load_viewport_scale:
                return PAN_SYSVAL_VIEWPORT_SCALE;
        case nir_intrinsic_load_viewport_offset:
                return PAN_SYSVAL_VIEWPORT_OFFSET;
        case nir_intrinsic_load_num_work_groups:
                return PAN_SYSVAL_NUM_WORK_GROUPS;
        case nir_intrinsic_load_ssbo: 
        case nir_intrinsic_store_ssbo: 
                return midgard_sysval_for_ssbo(instr);
        default:
                return ~0;
        }
}

static int sysval_for_instr(compiler_context *ctx, nir_instr *instr,
                            unsigned *dest)
{
        nir_intrinsic_instr *intr;
        nir_dest *dst = NULL;
        nir_tex_instr *tex;
        int sysval = -1;

        bool is_store = false;

        switch (instr->type) {
        case nir_instr_type_intrinsic:
                intr = nir_instr_as_intrinsic(instr);
                sysval = midgard_nir_sysval_for_intrinsic(intr);
                dst = &intr->dest;
                is_store |= intr->intrinsic == nir_intrinsic_store_ssbo;
                break;
        case nir_instr_type_tex:
                tex = nir_instr_as_tex(instr);
                if (tex->op != nir_texop_txs)
                        break;

                sysval = PAN_SYSVAL(TEXTURE_SIZE,
                                    PAN_TXS_SYSVAL_ID(tex->texture_index,
                                                      nir_tex_instr_dest_size(tex) -
                                                      (tex->is_array ? 1 : 0),
                                                      tex->is_array));
                dst  = &tex->dest;
                break;
        default:
                break;
        }

        if (dest && dst && !is_store)
                *dest = nir_dest_index(ctx, dst);

        return sysval;
}

static void
midgard_nir_assign_sysval_body(compiler_context *ctx, nir_instr *instr)
{
        int sysval;

        sysval = sysval_for_instr(ctx, instr, NULL);
        if (sysval < 0)
                return;

        /* We have a sysval load; check if it's already been assigned */

        if (_mesa_hash_table_u64_search(ctx->sysval_to_id, sysval))
                return;

        /* It hasn't -- so assign it now! */

        unsigned id = ctx->sysval_count++;
        _mesa_hash_table_u64_insert(ctx->sysval_to_id, sysval, (void *) ((uintptr_t) id + 1));
        ctx->sysvals[id] = sysval;
}

static void
midgard_nir_assign_sysvals(compiler_context *ctx, nir_shader *shader)
{
        ctx->sysval_count = 0;

        nir_foreach_function(function, shader) {
                if (!function->impl) continue;

                nir_foreach_block(block, function->impl) {
                        nir_foreach_instr_safe(instr, block) {
                                midgard_nir_assign_sysval_body(ctx, instr);
                        }
                }
        }
}

static bool
midgard_nir_lower_fdot2(nir_shader *shader)
{
        bool progress = false;

        nir_foreach_function(function, shader) {
                if (!function->impl) continue;

                nir_builder _b;
                nir_builder *b = &_b;
                nir_builder_init(b, function->impl);

                nir_foreach_block(block, function->impl) {
                        nir_foreach_instr_safe(instr, block) {
                                if (instr->type != nir_instr_type_alu) continue;

                                nir_alu_instr *alu = nir_instr_as_alu(instr);
                                midgard_nir_lower_fdot2_body(b, alu);

                                progress |= true;
                        }
                }

                nir_metadata_preserve(function->impl, nir_metadata_block_index | nir_metadata_dominance);

        }

        return progress;
}

/* Flushes undefined values to zero */

static void
optimise_nir(nir_shader *nir)
{
        bool progress;
        unsigned lower_flrp =
                (nir->options->lower_flrp16 ? 16 : 0) |
                (nir->options->lower_flrp32 ? 32 : 0) |
                (nir->options->lower_flrp64 ? 64 : 0);

        NIR_PASS(progress, nir, nir_lower_regs_to_ssa);
        NIR_PASS(progress, nir, midgard_nir_lower_fdot2);
        NIR_PASS(progress, nir, nir_lower_idiv, nir_lower_idiv_fast);

        nir_lower_tex_options lower_tex_options = {
                .lower_txs_lod = true,
                .lower_txp = ~0
        };

        NIR_PASS(progress, nir, nir_lower_tex, &lower_tex_options);

        do {
                progress = false;

                NIR_PASS(progress, nir, nir_lower_var_copies);
                NIR_PASS(progress, nir, nir_lower_vars_to_ssa);

                NIR_PASS(progress, nir, nir_copy_prop);
                NIR_PASS(progress, nir, nir_opt_dce);
                NIR_PASS(progress, nir, nir_opt_dead_cf);
                NIR_PASS(progress, nir, nir_opt_cse);
                NIR_PASS(progress, nir, nir_opt_peephole_select, 64, false, true);
                NIR_PASS(progress, nir, nir_opt_algebraic);
                NIR_PASS(progress, nir, nir_opt_constant_folding);

                if (lower_flrp != 0) {
                        bool lower_flrp_progress = false;
                        NIR_PASS(lower_flrp_progress,
                                 nir,
                                 nir_lower_flrp,
                                 lower_flrp,
                                 false /* always_precise */,
                                 nir->options->lower_ffma);
                        if (lower_flrp_progress) {
                                NIR_PASS(progress, nir,
                                         nir_opt_constant_folding);
                                progress = true;
                        }

                        /* Nothing should rematerialize any flrps, so we only
                         * need to do this lowering once.
                         */
                        lower_flrp = 0;
                }

                NIR_PASS(progress, nir, nir_opt_undef);
                NIR_PASS(progress, nir, nir_undef_to_zero);

                NIR_PASS(progress, nir, nir_opt_loop_unroll,
                         nir_var_shader_in |
                         nir_var_shader_out |
                         nir_var_function_temp);

                NIR_PASS(progress, nir, nir_opt_vectorize);
        } while (progress);

        /* Must be run at the end to prevent creation of fsin/fcos ops */
        NIR_PASS(progress, nir, midgard_nir_scale_trig);

        do {
                progress = false;

                NIR_PASS(progress, nir, nir_opt_dce);
                NIR_PASS(progress, nir, nir_opt_algebraic);
                NIR_PASS(progress, nir, nir_opt_constant_folding);
                NIR_PASS(progress, nir, nir_copy_prop);
        } while (progress);

        NIR_PASS(progress, nir, nir_opt_algebraic_late);

        /* We implement booleans as 32-bit 0/~0 */
        NIR_PASS(progress, nir, nir_lower_bool_to_int32);

        /* Now that booleans are lowered, we can run out late opts */
        NIR_PASS(progress, nir, midgard_nir_lower_algebraic_late);

        /* Lower mods for float ops only. Integer ops don't support modifiers
         * (saturate doesn't make sense on integers, neg/abs require dedicated
         * instructions) */

        NIR_PASS(progress, nir, nir_lower_to_source_mods, nir_lower_float_source_mods);
        NIR_PASS(progress, nir, nir_copy_prop);
        NIR_PASS(progress, nir, nir_opt_dce);

        /* Take us out of SSA */
        NIR_PASS(progress, nir, nir_lower_locals_to_regs);
        NIR_PASS(progress, nir, nir_convert_from_ssa, true);

        /* We are a vector architecture; write combine where possible */
        NIR_PASS(progress, nir, nir_move_vec_src_uses_to_dest);
        NIR_PASS(progress, nir, nir_lower_vec_to_movs);

        NIR_PASS(progress, nir, nir_opt_dce);
}

/* Do not actually emit a load; instead, cache the constant for inlining */

static void
emit_load_const(compiler_context *ctx, nir_load_const_instr *instr)
{
        nir_ssa_def def = instr->def;

        float *v = rzalloc_array(NULL, float, 4);
        nir_const_value_to_array(v, instr->value, instr->def.num_components, f32);

        /* Shifted for SSA, +1 for off-by-one */
        _mesa_hash_table_u64_insert(ctx->ssa_constants, (def.index << 1) + 1, v);
}

/* Normally constants are embedded implicitly, but for I/O and such we have to
 * explicitly emit a move with the constant source */

static void
emit_explicit_constant(compiler_context *ctx, unsigned node, unsigned to)
{
        void *constant_value = _mesa_hash_table_u64_search(ctx->ssa_constants, node + 1);

        if (constant_value) {
                midgard_instruction ins = v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT), to);
                attach_constants(ctx, &ins, constant_value, node + 1);
                emit_mir_instruction(ctx, ins);
        }
}

static bool
nir_is_non_scalar_swizzle(nir_alu_src *src, unsigned nr_components)
{
        unsigned comp = src->swizzle[0];

        for (unsigned c = 1; c < nr_components; ++c) {
                if (src->swizzle[c] != comp)
                        return true;
        }

        return false;
}

#define ALU_CASE(nir, _op) \
	case nir_op_##nir: \
		op = midgard_alu_op_##_op; \
                assert(src_bitsize == dst_bitsize); \
		break;

#define ALU_CASE_BCAST(nir, _op, count) \
        case nir_op_##nir: \
                op = midgard_alu_op_##_op; \
                broadcast_swizzle = count; \
                assert(src_bitsize == dst_bitsize); \
                break;
static bool
nir_is_fzero_constant(nir_src src)
{
        if (!nir_src_is_const(src))
                return false;

        for (unsigned c = 0; c < nir_src_num_components(src); ++c) {
                if (nir_src_comp_as_float(src, c) != 0.0)
                        return false;
        }

        return true;
}

/* Analyze the sizes of the inputs to determine which reg mode. Ops needed
 * special treatment override this anyway. */

static midgard_reg_mode
reg_mode_for_nir(nir_alu_instr *instr)
{
        unsigned src_bitsize = nir_src_bit_size(instr->src[0].src);

        switch (src_bitsize) {
        case 8:
                return midgard_reg_mode_8;
        case 16:
                return midgard_reg_mode_16;
        case 32:
                return midgard_reg_mode_32;
        case 64:
                return midgard_reg_mode_64;
        default:
                unreachable("Invalid bit size");
        }
}

static void
emit_alu(compiler_context *ctx, nir_alu_instr *instr)
{
        /* Derivatives end up emitted on the texture pipe, not the ALUs. This
         * is handled elsewhere */

        if (instr->op == nir_op_fddx || instr->op == nir_op_fddy) {
                midgard_emit_derivatives(ctx, instr);
                return;
        }

        bool is_ssa = instr->dest.dest.is_ssa;

        unsigned dest = nir_dest_index(ctx, &instr->dest.dest);
        unsigned nr_components = nir_dest_num_components(instr->dest.dest);
        unsigned nr_inputs = nir_op_infos[instr->op].num_inputs;

        /* Most Midgard ALU ops have a 1:1 correspondance to NIR ops; these are
         * supported. A few do not and are commented for now. Also, there are a
         * number of NIR ops which Midgard does not support and need to be
         * lowered, also TODO. This switch block emits the opcode and calling
         * convention of the Midgard instruction; actual packing is done in
         * emit_alu below */

        unsigned op;

        /* Number of components valid to check for the instruction (the rest
         * will be forced to the last), or 0 to use as-is. Relevant as
         * ball-type instructions have a channel count in NIR but are all vec4
         * in Midgard */

        unsigned broadcast_swizzle = 0;

        /* What register mode should we operate in? */
        midgard_reg_mode reg_mode =
                reg_mode_for_nir(instr);

        /* Do we need a destination override? Used for inline
         * type conversion */

        midgard_dest_override dest_override =
                midgard_dest_override_none;

        /* Should we use a smaller respective source and sign-extend?  */

        bool half_1 = false, sext_1 = false;
        bool half_2 = false, sext_2 = false;

        unsigned src_bitsize = nir_src_bit_size(instr->src[0].src);
        unsigned dst_bitsize = nir_dest_bit_size(instr->dest.dest);

        switch (instr->op) {
                ALU_CASE(fadd, fadd);
                ALU_CASE(fmul, fmul);
                ALU_CASE(fmin, fmin);
                ALU_CASE(fmax, fmax);
                ALU_CASE(imin, imin);
                ALU_CASE(imax, imax);
                ALU_CASE(umin, umin);
                ALU_CASE(umax, umax);
                ALU_CASE(ffloor, ffloor);
                ALU_CASE(fround_even, froundeven);
                ALU_CASE(ftrunc, ftrunc);
                ALU_CASE(fceil, fceil);
                ALU_CASE(fdot3, fdot3);
                ALU_CASE(fdot4, fdot4);
                ALU_CASE(iadd, iadd);
                ALU_CASE(isub, isub);
                ALU_CASE(imul, imul);

                /* Zero shoved as second-arg */
                ALU_CASE(iabs, iabsdiff);

                ALU_CASE(mov, imov);

                ALU_CASE(feq32, feq);
                ALU_CASE(fne32, fne);
                ALU_CASE(flt32, flt);
                ALU_CASE(ieq32, ieq);
                ALU_CASE(ine32, ine);
                ALU_CASE(ilt32, ilt);
                ALU_CASE(ult32, ult);

                /* We don't have a native b2f32 instruction. Instead, like many
                 * GPUs, we exploit booleans as 0/~0 for false/true, and
                 * correspondingly AND
                 * by 1.0 to do the type conversion. For the moment, prime us
                 * to emit:
                 *
                 * iand [whatever], #0
                 *
                 * At the end of emit_alu (as MIR), we'll fix-up the constant
                 */

                ALU_CASE(b2f32, iand);
                ALU_CASE(b2i32, iand);

                /* Likewise, we don't have a dedicated f2b32 instruction, but
                 * we can do a "not equal to 0.0" test. */

                ALU_CASE(f2b32, fne);
                ALU_CASE(i2b32, ine);

                ALU_CASE(frcp, frcp);
                ALU_CASE(frsq, frsqrt);
                ALU_CASE(fsqrt, fsqrt);
                ALU_CASE(fexp2, fexp2);
                ALU_CASE(flog2, flog2);

                ALU_CASE(f2i32, f2i_rtz);
                ALU_CASE(f2u32, f2u_rtz);
                ALU_CASE(i2f32, i2f_rtz);
                ALU_CASE(u2f32, u2f_rtz);

                ALU_CASE(f2i16, f2i_rtz);
                ALU_CASE(f2u16, f2u_rtz);
                ALU_CASE(i2f16, i2f_rtz);
                ALU_CASE(u2f16, u2f_rtz);

                ALU_CASE(fsin, fsin);
                ALU_CASE(fcos, fcos);

                /* We'll set invert */
                ALU_CASE(inot, imov);
                ALU_CASE(iand, iand);
                ALU_CASE(ior, ior);
                ALU_CASE(ixor, ixor);
                ALU_CASE(ishl, ishl);
                ALU_CASE(ishr, iasr);
                ALU_CASE(ushr, ilsr);

                ALU_CASE_BCAST(b32all_fequal2, fball_eq, 2);
                ALU_CASE_BCAST(b32all_fequal3, fball_eq, 3);
                ALU_CASE(b32all_fequal4, fball_eq);

                ALU_CASE_BCAST(b32any_fnequal2, fbany_neq, 2);
                ALU_CASE_BCAST(b32any_fnequal3, fbany_neq, 3);
                ALU_CASE(b32any_fnequal4, fbany_neq);

                ALU_CASE_BCAST(b32all_iequal2, iball_eq, 2);
                ALU_CASE_BCAST(b32all_iequal3, iball_eq, 3);
                ALU_CASE(b32all_iequal4, iball_eq);

                ALU_CASE_BCAST(b32any_inequal2, ibany_neq, 2);
                ALU_CASE_BCAST(b32any_inequal3, ibany_neq, 3);
                ALU_CASE(b32any_inequal4, ibany_neq);

                /* Source mods will be shoved in later */
                ALU_CASE(fabs, fmov);
                ALU_CASE(fneg, fmov);
                ALU_CASE(fsat, fmov);

        /* For size conversion, we use a move. Ideally though we would squash
         * these ops together; maybe that has to happen after in NIR as part of
         * propagation...? An earlier algebraic pass ensured we step down by
         * only / exactly one size. If stepping down, we use a dest override to
         * reduce the size; if stepping up, we use a larger-sized move with a
         * half source and a sign/zero-extension modifier */

        case nir_op_i2i8:
        case nir_op_i2i16:
        case nir_op_i2i32:
                /* If we end up upscale, we'll need a sign-extend on the
                 * operand (the second argument) */

                sext_2 = true;
                /* fallthrough */
        case nir_op_u2u8:
        case nir_op_u2u16:
        case nir_op_u2u32: {
                op = midgard_alu_op_imov;

                if (dst_bitsize == (src_bitsize * 2)) {
                        /* Converting up */
                        half_2 = true;

                        /* Use a greater register mode */
                        reg_mode++;
                } else if (src_bitsize == (dst_bitsize * 2)) {
                        /* Converting down */
                        dest_override = midgard_dest_override_lower;
                }

                break;
        }

        case nir_op_f2f16: {
                assert(src_bitsize == 32);

                op = midgard_alu_op_fmov;
                dest_override = midgard_dest_override_lower;
                break;
        }

        case nir_op_f2f32: {
                assert(src_bitsize == 16);

                op = midgard_alu_op_fmov;
                half_2 = true;
                reg_mode++;
                break;
        }


        /* For greater-or-equal, we lower to less-or-equal and flip the
         * arguments */

        case nir_op_fge:
        case nir_op_fge32:
        case nir_op_ige32:
        case nir_op_uge32: {
                op =
                        instr->op == nir_op_fge   ? midgard_alu_op_fle :
                        instr->op == nir_op_fge32 ? midgard_alu_op_fle :
                        instr->op == nir_op_ige32 ? midgard_alu_op_ile :
                        instr->op == nir_op_uge32 ? midgard_alu_op_ule :
                        0;

                /* Swap via temporary */
                nir_alu_src temp = instr->src[1];
                instr->src[1] = instr->src[0];
                instr->src[0] = temp;

                break;
        }

        case nir_op_b32csel: {
                /* Midgard features both fcsel and icsel, depending on
                 * the type of the arguments/output. However, as long
                 * as we're careful we can _always_ use icsel and
                 * _never_ need fcsel, since the latter does additional
                 * floating-point-specific processing whereas the
                 * former just moves bits on the wire. It's not obvious
                 * why these are separate opcodes, save for the ability
                 * to do things like sat/pos/abs/neg for free */

                bool mixed = nir_is_non_scalar_swizzle(&instr->src[0], nr_components);
                op = mixed ? midgard_alu_op_icsel_v : midgard_alu_op_icsel;

                /* The condition is the first argument; move the other
                 * arguments up one to be a binary instruction for
                 * Midgard with the condition last */

                nir_alu_src temp = instr->src[2];

                instr->src[2] = instr->src[0];
                instr->src[0] = instr->src[1];
                instr->src[1] = temp;

                break;
        }

        default:
                DBG("Unhandled ALU op %s\n", nir_op_infos[instr->op].name);
                assert(0);
                return;
        }

        /* Midgard can perform certain modifiers on output of an ALU op */
        unsigned outmod;

        if (midgard_is_integer_out_op(op)) {
                outmod = midgard_outmod_int_wrap;
        } else {
                bool sat = instr->dest.saturate || instr->op == nir_op_fsat;
                outmod = sat ? midgard_outmod_sat : midgard_outmod_none;
        }

        /* fmax(a, 0.0) can turn into a .pos modifier as an optimization */

        if (instr->op == nir_op_fmax) {
                if (nir_is_fzero_constant(instr->src[0].src)) {
                        op = midgard_alu_op_fmov;
                        nr_inputs = 1;
                        outmod = midgard_outmod_pos;
                        instr->src[0] = instr->src[1];
                } else if (nir_is_fzero_constant(instr->src[1].src)) {
                        op = midgard_alu_op_fmov;
                        nr_inputs = 1;
                        outmod = midgard_outmod_pos;
                }
        }

        /* Fetch unit, quirks, etc information */
        unsigned opcode_props = alu_opcode_props[op].props;
        bool quirk_flipped_r24 = opcode_props & QUIRK_FLIPPED_R24;

        /* src0 will always exist afaik, but src1 will not for 1-argument
         * instructions. The latter can only be fetched if the instruction
         * needs it, or else we may segfault. */

        unsigned src0 = nir_alu_src_index(ctx, &instr->src[0]);
        unsigned src1 = nr_inputs >= 2 ? nir_alu_src_index(ctx, &instr->src[1]) : ~0;
        unsigned src2 = nr_inputs == 3 ? nir_alu_src_index(ctx, &instr->src[2]) : ~0;
        assert(nr_inputs <= 3);

        /* Rather than use the instruction generation helpers, we do it
         * ourselves here to avoid the mess */

        midgard_instruction ins = {
                .type = TAG_ALU_4,
                .src = {
                        quirk_flipped_r24 ? ~0 : src0,
                        quirk_flipped_r24 ? src0       : src1,
                        src2,
                },
                .dest = dest,
        };

        nir_alu_src *nirmods[3] = { NULL };

        if (nr_inputs >= 2) {
                nirmods[0] = &instr->src[0];
                nirmods[1] = &instr->src[1];
        } else if (nr_inputs == 1) {
                nirmods[quirk_flipped_r24] = &instr->src[0];
        } else {
                assert(0);
        }

        if (nr_inputs == 3)
                nirmods[2] = &instr->src[2];

        /* These were lowered to a move, so apply the corresponding mod */

        if (instr->op == nir_op_fneg || instr->op == nir_op_fabs) {
                nir_alu_src *s = nirmods[quirk_flipped_r24];

                if (instr->op == nir_op_fneg)
                        s->negate = !s->negate;

                if (instr->op == nir_op_fabs)
                        s->abs = !s->abs;
        }

        bool is_int = midgard_is_integer_op(op);

        ins.mask = mask_of(nr_components);

        midgard_vector_alu alu = {
                .op = op,
                .reg_mode = reg_mode,
                .dest_override = dest_override,
                .outmod = outmod,

                .src1 = vector_alu_srco_unsigned(vector_alu_modifiers(nirmods[0], is_int, broadcast_swizzle, half_1, sext_1)),
                .src2 = vector_alu_srco_unsigned(vector_alu_modifiers(nirmods[1], is_int, broadcast_swizzle, half_2, sext_2)),
        };

        /* Apply writemask if non-SSA, keeping in mind that we can't write to components that don't exist */

        if (!is_ssa)
                ins.mask &= instr->dest.write_mask;

        for (unsigned m = 0; m < 3; ++m) {
                if (!nirmods[m])
                        continue;

                for (unsigned c = 0; c < NIR_MAX_VEC_COMPONENTS; ++c)
                        ins.swizzle[m][c] = nirmods[m]->swizzle[c];

                /* Replicate. TODO: remove when vec16 lands */
                for (unsigned c = NIR_MAX_VEC_COMPONENTS; c < MIR_VEC_COMPONENTS; ++c)
                        ins.swizzle[m][c] = nirmods[m]->swizzle[NIR_MAX_VEC_COMPONENTS - 1];
        }

        if (nr_inputs == 3) {
                /* Conditions can't have mods */
                assert(!nirmods[2]->abs);
                assert(!nirmods[2]->negate);
        }

        ins.alu = alu;

        /* Late fixup for emulated instructions */

        if (instr->op == nir_op_b2f32 || instr->op == nir_op_b2i32) {
                /* Presently, our second argument is an inline #0 constant.
                 * Switch over to an embedded 1.0 constant (that can't fit
                 * inline, since we're 32-bit, not 16-bit like the inline
                 * constants) */

                ins.has_inline_constant = false;
                ins.src[1] = SSA_FIXED_REGISTER(REGISTER_CONSTANT);
                ins.has_constants = true;

                if (instr->op == nir_op_b2f32) {
                        float f = 1.0f;
                        memcpy(&ins.constants, &f, sizeof(float));
                } else {
                        ins.constants[0] = 1;
                }


                for (unsigned c = 0; c < 16; ++c)
                        ins.swizzle[1][c] = 0;
        } else if (nr_inputs == 1 && !quirk_flipped_r24) {
                /* Lots of instructions need a 0 plonked in */
                ins.has_inline_constant = false;
                ins.src[1] = SSA_FIXED_REGISTER(REGISTER_CONSTANT);
                ins.has_constants = true;
                ins.constants[0] = 0;

                for (unsigned c = 0; c < 16; ++c)
                        ins.swizzle[1][c] = 0;
        } else if (instr->op == nir_op_inot) {
                ins.invert = true;
        }

        if ((opcode_props & UNITS_ALL) == UNIT_VLUT) {
                /* To avoid duplicating the lookup tables (probably), true LUT
                 * instructions can only operate as if they were scalars. Lower
                 * them here by changing the component. */

                unsigned orig_mask = ins.mask;

                for (int i = 0; i < nr_components; ++i) {
                        /* Mask the associated component, dropping the
                         * instruction if needed */

                        ins.mask = 1 << i;
                        ins.mask &= orig_mask;

                        if (!ins.mask)
                                continue;

                        for (unsigned j = 0; j < MIR_VEC_COMPONENTS; ++j)
                                ins.swizzle[0][j] = nirmods[0]->swizzle[i]; /* Pull from the correct component */

                        emit_mir_instruction(ctx, ins);
                }
        } else {
                emit_mir_instruction(ctx, ins);
        }
}

#undef ALU_CASE

static unsigned
mir_mask_for_intr(nir_instr *instr, bool is_read)
{
        nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);

        if (is_read)
                return mask_of(nir_intrinsic_dest_components(intr));
        else
                return nir_intrinsic_write_mask(intr);
}

/* Uniforms and UBOs use a shared code path, as uniforms are just (slightly
 * optimized) versions of UBO #0 */

midgard_instruction *
emit_ubo_read(
        compiler_context *ctx,
        nir_instr *instr,
        unsigned dest,
        unsigned offset,
        nir_src *indirect_offset,
        unsigned index)
{
        /* TODO: half-floats */

        midgard_instruction ins = m_ld_ubo_int4(dest, offset);

        assert((offset & 0xF) == 0);
        offset /= 16;

        /* TODO: Don't split */
        ins.load_store.varying_parameters = (offset & 7) << 7;
        ins.load_store.address = offset >> 3;
        ins.mask = mir_mask_for_intr(instr, true);

        if (indirect_offset) {
                ins.src[2] = nir_src_index(ctx, indirect_offset);
                ins.load_store.arg_2 = 0x80;
        } else {
                ins.load_store.arg_2 = 0x1E;
        }

        ins.load_store.arg_1 = index;

        return emit_mir_instruction(ctx, ins);
}

/* SSBO reads are like UBO reads if you squint */

static void
emit_ssbo_access(
        compiler_context *ctx,
        nir_instr *instr,
        bool is_read,
        unsigned srcdest,
        unsigned offset,
        nir_src *indirect_offset,
        unsigned index)
{
        /* TODO: types */

        midgard_instruction ins; 

        if (is_read)
                ins = m_ld_int4(srcdest, offset);
        else
                ins = m_st_int4(srcdest, offset);

        /* SSBO reads use a generic memory read interface, so we need the
         * address of the SSBO as the first argument. This is a sysval. */

        unsigned addr = make_compiler_temp(ctx);
        emit_sysval_read(ctx, instr, addr, 2);

        /* The source array:
         *
         *  src[0] = store ? value : unused
         *  src[1] = arg_1
         *  src[2] = arg_2
         *
         * We would like arg_1 = the address and
         * arg_2 = the offset.
         */

        ins.src[1] = addr;

        /* TODO: What is this? It looks superficially like a shift << 5, but
         * arg_1 doesn't take a shift Should it be E0 or A0? We also need the
         * indirect offset. */

        if (indirect_offset) {
                ins.load_store.arg_1 |= 0xE0;
                ins.src[2] = nir_src_index(ctx, indirect_offset);
        } else {
                ins.load_store.arg_2 = 0x7E;
        }

        /* TODO: Bounds check */

        /* Finally, we emit the direct offset */

        ins.load_store.varying_parameters = (offset & 0x1FF) << 1;
        ins.load_store.address = (offset >> 9);
        ins.mask = mir_mask_for_intr(instr, is_read);

        emit_mir_instruction(ctx, ins);
}

static void
emit_varying_read(
        compiler_context *ctx,
        unsigned dest, unsigned offset,
        unsigned nr_comp, unsigned component,
        nir_src *indirect_offset, nir_alu_type type)
{
        /* XXX: Half-floats? */
        /* TODO: swizzle, mask */

        midgard_instruction ins = m_ld_vary_32(dest, offset);
        ins.mask = mask_of(nr_comp);

        for (unsigned i = 0; i < ARRAY_SIZE(ins.swizzle[0]); ++i)
                ins.swizzle[0][i] = MIN2(i + component, COMPONENT_W);

        midgard_varying_parameter p = {
                .is_varying = 1,
                .interpolation = midgard_interp_default,
                .flat = /*var->data.interpolation == INTERP_MODE_FLAT*/ 0
        };

        unsigned u;
        memcpy(&u, &p, sizeof(p));
        ins.load_store.varying_parameters = u;

        if (indirect_offset)
                ins.src[2] = nir_src_index(ctx, indirect_offset);
        else
                ins.load_store.arg_2 = 0x1E;

        ins.load_store.arg_1 = 0x9E;

        /* Use the type appropriate load */
        switch (type) {
        case nir_type_uint:
        case nir_type_bool:
                ins.load_store.op = midgard_op_ld_vary_32u;
                break;
        case nir_type_int:
                ins.load_store.op = midgard_op_ld_vary_32i;
                break;
        case nir_type_float:
                ins.load_store.op = midgard_op_ld_vary_32;
                break;
        default:
                unreachable("Attempted to load unknown type");
                break;
        }

        emit_mir_instruction(ctx, ins);
}

void
emit_sysval_read(compiler_context *ctx, nir_instr *instr, signed dest_override,
                unsigned nr_components)
{
        unsigned dest = 0;

        /* Figure out which uniform this is */
        int sysval = sysval_for_instr(ctx, instr, &dest);
        void *val = _mesa_hash_table_u64_search(ctx->sysval_to_id, sysval);

        if (dest_override >= 0)
                dest = dest_override;

        /* Sysvals are prefix uniforms */
        unsigned uniform = ((uintptr_t) val) - 1;

        /* Emit the read itself -- this is never indirect */
        midgard_instruction *ins =
                emit_ubo_read(ctx, instr, dest, uniform * 16, NULL, 0);

        ins->mask = mask_of(nr_components);
}

static unsigned
compute_builtin_arg(nir_op op)
{
        switch (op) {
        case nir_intrinsic_load_work_group_id:
                return 0x14;
        case nir_intrinsic_load_local_invocation_id:
                return 0x10;
        default:
                unreachable("Invalid compute paramater loaded");
        }
}

/* Emit store for a fragment shader, which is encoded via a fancy branch. TODO:
 * Handle MRT here */

static void
emit_fragment_store(compiler_context *ctx, unsigned src, unsigned rt)
{
        emit_explicit_constant(ctx, src, src);

        /* If we're doing MRT, we need to specify the render target */

        midgard_instruction rt_move = {
                .dest = ~0
        };

        if (rt != 0) {
                /* We'll write to r1.z */
                rt_move = v_mov(~0, SSA_FIXED_REGISTER(1));
                rt_move.mask = 1 << COMPONENT_Z;
                rt_move.unit = UNIT_SADD;

                /* r1.z = (rt * 0x100) */
                rt_move.has_inline_constant = true;
                rt_move.inline_constant = (rt * 0x100);

                /* r1 */
                ctx->work_registers = MAX2(ctx->work_registers, 1);

                /* Do the write */
                emit_mir_instruction(ctx, rt_move);
        }

        /* Next, generate the branch. For R render targets in the writeout, the
         * i'th render target jumps to pseudo-offset [2(R-1) + i] */

        unsigned outputs = ctx->is_blend ? 1 : ctx->nir->num_outputs;
        unsigned offset = (2 * (outputs - 1)) + rt;

        struct midgard_instruction ins =
                v_alu_br_compact_cond(midgard_jmp_writeout_op_writeout, TAG_ALU_4, offset, midgard_condition_always);

        /* Add dependencies */
        ins.src[0] = src;
        ins.src[1] = rt_move.dest;

        /* Emit the branch */
        emit_mir_instruction(ctx, ins);
}

static void
emit_compute_builtin(compiler_context *ctx, nir_intrinsic_instr *instr)
{
        unsigned reg = nir_dest_index(ctx, &instr->dest);
        midgard_instruction ins = m_ld_compute_id(reg, 0);
        ins.mask = mask_of(3);
        ins.load_store.arg_1 = compute_builtin_arg(instr->intrinsic);
        emit_mir_instruction(ctx, ins);
}
static void
emit_intrinsic(compiler_context *ctx, nir_intrinsic_instr *instr)
{
        unsigned offset = 0, reg;

        switch (instr->intrinsic) {
        case nir_intrinsic_discard_if:
        case nir_intrinsic_discard: {
                bool conditional = instr->intrinsic == nir_intrinsic_discard_if;
                struct midgard_instruction discard = v_branch(conditional, false);
                discard.branch.target_type = TARGET_DISCARD;

                if (conditional)
                        discard.src[0] = nir_src_index(ctx, &instr->src[0]);

                emit_mir_instruction(ctx, discard);
                schedule_barrier(ctx);

                break;
        }

        case nir_intrinsic_load_uniform:
        case nir_intrinsic_load_ubo:
        case nir_intrinsic_load_ssbo:
        case nir_intrinsic_load_input: {
                bool is_uniform = instr->intrinsic == nir_intrinsic_load_uniform;
                bool is_ubo = instr->intrinsic == nir_intrinsic_load_ubo;
                bool is_ssbo = instr->intrinsic == nir_intrinsic_load_ssbo;

                /* Get the base type of the intrinsic */
                /* TODO: Infer type? Does it matter? */
                nir_alu_type t =
                        (is_ubo || is_ssbo) ? nir_type_uint : nir_intrinsic_type(instr);
                t = nir_alu_type_get_base_type(t);

                if (!(is_ubo || is_ssbo)) {
                        offset = nir_intrinsic_base(instr);
                }

                unsigned nr_comp = nir_intrinsic_dest_components(instr);

                nir_src *src_offset = nir_get_io_offset_src(instr);

                bool direct = nir_src_is_const(*src_offset);
                nir_src *indirect_offset = direct ? NULL : src_offset;

                if (direct)
                        offset += nir_src_as_uint(*src_offset);

                /* We may need to apply a fractional offset */
                int component = instr->intrinsic == nir_intrinsic_load_input ?
                                nir_intrinsic_component(instr) : 0;
                reg = nir_dest_index(ctx, &instr->dest);

                if (is_uniform && !ctx->is_blend) {
                        emit_ubo_read(ctx, &instr->instr, reg, (ctx->sysval_count + offset) * 16, indirect_offset, 0);
                } else if (is_ubo) {
                        nir_src index = instr->src[0];

                        /* We don't yet support indirect UBOs. For indirect
                         * block numbers (if that's possible), we don't know
                         * enough about the hardware yet. For indirect sources,
                         * we know what we need but we need to add some NIR
                         * support for lowering correctly with respect to
                         * 128-bit reads */

                        assert(nir_src_is_const(index));
                        assert(nir_src_is_const(*src_offset));

                        uint32_t uindex = nir_src_as_uint(index) + 1;
                        emit_ubo_read(ctx, &instr->instr, reg, offset, NULL, uindex);
                } else if (is_ssbo) {
                        nir_src index = instr->src[0];
                        assert(nir_src_is_const(index));
                        uint32_t uindex = nir_src_as_uint(index);

                        emit_ssbo_access(ctx, &instr->instr, true, reg, offset, indirect_offset, uindex);
                } else if (ctx->stage == MESA_SHADER_FRAGMENT && !ctx->is_blend) {
                        emit_varying_read(ctx, reg, offset, nr_comp, component, !direct ? &instr->src[0] : NULL, t);
                } else if (ctx->is_blend) {
                        /* For blend shaders, load the input color, which is
                         * preloaded to r0 */

                        midgard_instruction move = v_mov(SSA_FIXED_REGISTER(0), reg);
                        emit_mir_instruction(ctx, move);
                        schedule_barrier(ctx);
                }  else if (ctx->stage == MESA_SHADER_VERTEX) {
                        midgard_instruction ins = m_ld_attr_32(reg, offset);
                        ins.load_store.arg_1 = 0x1E;
                        ins.load_store.arg_2 = 0x1E;
                        ins.mask = mask_of(nr_comp);

                        /* Use the type appropriate load */
                        switch (t) {
                        case nir_type_uint:
                        case nir_type_bool:
                                ins.load_store.op = midgard_op_ld_attr_32u;
                                break;
                        case nir_type_int:
                                ins.load_store.op = midgard_op_ld_attr_32i;
                                break;
                        case nir_type_float:
                                ins.load_store.op = midgard_op_ld_attr_32;
                                break;
                        default:
                                unreachable("Attempted to load unknown type");
                                break;
                        }

                        emit_mir_instruction(ctx, ins);
                } else {
                        DBG("Unknown load\n");
                        assert(0);
                }

                break;
        }

        /* Reads 128-bit value raw off the tilebuffer during blending, tasty */

        case nir_intrinsic_load_raw_output_pan:
        case nir_intrinsic_load_output_u8_as_fp16_pan:
                reg = nir_dest_index(ctx, &instr->dest);
                assert(ctx->is_blend);

                /* T720 and below use different blend opcodes with slightly
                 * different semantics than T760 and up */

                midgard_instruction ld = m_ld_color_buffer_8(reg, 0);
                bool old_blend = ctx->gpu_id < 0x750;

                if (instr->intrinsic == nir_intrinsic_load_output_u8_as_fp16_pan) {
                        ld.load_store.op = old_blend ?
                                midgard_op_ld_color_buffer_u8_as_fp16_old :
                                midgard_op_ld_color_buffer_u8_as_fp16;

                        if (old_blend) {
                                ld.load_store.address = 1;
                                ld.load_store.arg_2 = 0x1E;
                        }

                        for (unsigned c = 2; c < 16; ++c)
                                ld.swizzle[0][c] = 0;
                }

                emit_mir_instruction(ctx, ld);
                break;

        case nir_intrinsic_load_blend_const_color_rgba: {
                assert(ctx->is_blend);
                reg = nir_dest_index(ctx, &instr->dest);

                /* Blend constants are embedded directly in the shader and
                 * patched in, so we use some magic routing */

                midgard_instruction ins = v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT), reg);
                ins.has_constants = true;
                ins.has_blend_constant = true;
                emit_mir_instruction(ctx, ins);
                break;
        }

        case nir_intrinsic_store_output:
                assert(nir_src_is_const(instr->src[1]) && "no indirect outputs");

                offset = nir_intrinsic_base(instr) + nir_src_as_uint(instr->src[1]);

                reg = nir_src_index(ctx, &instr->src[0]);

                if (ctx->stage == MESA_SHADER_FRAGMENT) {
                        /* Determine number of render targets */
                        emit_fragment_store(ctx, reg, offset);
                } else if (ctx->stage == MESA_SHADER_VERTEX) {
                        /* We should have been vectorized, though we don't
                         * currently check that st_vary is emitted only once
                         * per slot (this is relevant, since there's not a mask
                         * parameter available on the store [set to 0 by the
                         * blob]). We do respect the component by adjusting the
                         * swizzle. If this is a constant source, we'll need to
                         * emit that explicitly. */

                        emit_explicit_constant(ctx, reg, reg);

                        unsigned component = nir_intrinsic_component(instr);
                        unsigned nr_comp = nir_src_num_components(instr->src[0]);

                        midgard_instruction st = m_st_vary_32(reg, offset);
                        st.load_store.arg_1 = 0x9E;
                        st.load_store.arg_2 = 0x1E;

                        for (unsigned i = 0; i < ARRAY_SIZE(st.swizzle[0]); ++i)
                                st.swizzle[0][i] = MIN2(i + component, nr_comp);

                        emit_mir_instruction(ctx, st);
                } else {
                        DBG("Unknown store\n");
                        assert(0);
                }

                break;

        /* Special case of store_output for lowered blend shaders */
        case nir_intrinsic_store_raw_output_pan:
                assert (ctx->stage == MESA_SHADER_FRAGMENT);
                reg = nir_src_index(ctx, &instr->src[0]);
                emit_fragment_store(ctx, reg, 0);

                break;

        case nir_intrinsic_store_ssbo:
                assert(nir_src_is_const(instr->src[1]));

                bool direct_offset = nir_src_is_const(instr->src[2]);
                offset = direct_offset ? nir_src_as_uint(instr->src[2]) : 0;
                nir_src *indirect_offset = direct_offset ? NULL : &instr->src[2];
                reg = nir_src_index(ctx, &instr->src[0]);

                uint32_t uindex = nir_src_as_uint(instr->src[1]);

                emit_explicit_constant(ctx, reg, reg);
                emit_ssbo_access(ctx, &instr->instr, false, reg, offset, indirect_offset, uindex);
                break;

        case nir_intrinsic_load_viewport_scale:
        case nir_intrinsic_load_viewport_offset:
        case nir_intrinsic_load_num_work_groups:
                emit_sysval_read(ctx, &instr->instr, ~0, 3);
                break;

        case nir_intrinsic_load_work_group_id:
        case nir_intrinsic_load_local_invocation_id:
                emit_compute_builtin(ctx, instr);
                break;

        default:
                printf ("Unhandled intrinsic\n");
                assert(0);
                break;
        }
}

static unsigned
midgard_tex_format(enum glsl_sampler_dim dim)
{
        switch (dim) {
        case GLSL_SAMPLER_DIM_1D:
        case GLSL_SAMPLER_DIM_BUF:
                return MALI_TEX_1D;

        case GLSL_SAMPLER_DIM_2D:
        case GLSL_SAMPLER_DIM_EXTERNAL:
        case GLSL_SAMPLER_DIM_RECT:
                return MALI_TEX_2D;

        case GLSL_SAMPLER_DIM_3D:
                return MALI_TEX_3D;

        case GLSL_SAMPLER_DIM_CUBE:
                return MALI_TEX_CUBE;

        default:
                DBG("Unknown sampler dim type\n");
                assert(0);
                return 0;
        }
}

/* Tries to attach an explicit LOD / bias as a constant. Returns whether this
 * was successful */

static bool
pan_attach_constant_bias(
        compiler_context *ctx,
        nir_src lod,
        midgard_texture_word *word)
{
        /* To attach as constant, it has to *be* constant */

        if (!nir_src_is_const(lod))
                return false;

        float f = nir_src_as_float(lod);

        /* Break into fixed-point */
        signed lod_int = f;
        float lod_frac = f - lod_int;

        /* Carry over negative fractions */
        if (lod_frac < 0.0) {
                lod_int--;
                lod_frac += 1.0;
        }

        /* Encode */
        word->bias = float_to_ubyte(lod_frac);
        word->bias_int = lod_int;

        return true;
}

static enum mali_sampler_type
midgard_sampler_type(nir_alu_type t) {
        switch (nir_alu_type_get_base_type(t))
        {
        case nir_type_float:
                                return MALI_SAMPLER_FLOAT;
        case nir_type_int:
                return MALI_SAMPLER_SIGNED;
        case nir_type_uint:
                return MALI_SAMPLER_UNSIGNED;
        default:
                unreachable("Unknown sampler type");
        }
}

static void
emit_texop_native(compiler_context *ctx, nir_tex_instr *instr,
                  unsigned midgard_texop)
{
        /* TODO */
        //assert (!instr->sampler);
        //assert (!instr->texture_array_size);

        int texture_index = instr->texture_index;
        int sampler_index = texture_index;

        /* No helper to build texture words -- we do it all here */
        midgard_instruction ins = {
                .type = TAG_TEXTURE_4,
                .mask = 0xF,
                .dest = nir_dest_index(ctx, &instr->dest),
                .src = { ~0, ~0, ~0 },
                .swizzle = SWIZZLE_IDENTITY_4,
                .texture = {
                        .op = midgard_texop,
                        .format = midgard_tex_format(instr->sampler_dim),
                        .texture_handle = texture_index,
                        .sampler_handle = sampler_index,

                        /* TODO: half */
                        .in_reg_full = 1,
                        .out_full = 1,

                        .sampler_type = midgard_sampler_type(instr->dest_type),
                }
        };

        for (unsigned i = 0; i < instr->num_srcs; ++i) {
                int index = nir_src_index(ctx, &instr->src[i].src);
                unsigned nr_components = nir_src_num_components(instr->src[i].src);

                switch (instr->src[i].src_type) {
                case nir_tex_src_coord: {
                        emit_explicit_constant(ctx, index, index);

                        /* Texelfetch coordinates uses all four elements
                         * (xyz/index) regardless of texture dimensionality,
                         * which means it's necessary to zero the unused
                         * components to keep everything happy */

                        if (midgard_texop == TEXTURE_OP_TEXEL_FETCH) {
                                unsigned old_index = index;

                                index = make_compiler_temp(ctx);

                                /* mov index, old_index */
                                midgard_instruction mov = v_mov(old_index, index);
                                mov.mask = 0x3;
                                emit_mir_instruction(ctx, mov);

                                /* mov index.zw, #0 */
                                mov = v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT), index);
                                mov.has_constants = true;
                                mov.mask = (1 << COMPONENT_Z) | (1 << COMPONENT_W);
                                emit_mir_instruction(ctx, mov);
                        }

                        if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE) {
                                /* texelFetch is undefined on samplerCube */
                                assert(midgard_texop != TEXTURE_OP_TEXEL_FETCH);

                                /* For cubemaps, we use a special ld/st op to
                                 * select the face and copy the xy into the
                                 * texture register */

                                unsigned temp = make_compiler_temp(ctx);
                                midgard_instruction ld = m_ld_cubemap_coords(temp, 0);
                                ld.src[1] = index;
                                ld.mask = 0x3; /* xy */
                                ld.load_store.arg_1 = 0x20;
                                ld.swizzle[1][3] = COMPONENT_X;
                                emit_mir_instruction(ctx, ld);

                                ins.src[1] = temp;
                                /* xyzw -> xyxx */
                                ins.swizzle[1][2] = COMPONENT_X;
                                ins.swizzle[1][3] = COMPONENT_X;
                        } else {
                                ins.src[1] = index;
                        }

                        if (instr->sampler_dim == GLSL_SAMPLER_DIM_2D) {
                                /* Array component in w but NIR wants it in z */
                                if (nr_components == 3) {
                                        ins.swizzle[1][2] = COMPONENT_Z;
                                        ins.swizzle[1][3] = COMPONENT_Z;
                                } else if (nr_components == 2) {
                                        ins.swizzle[1][2] = COMPONENT_X;
                                        ins.swizzle[1][3] = COMPONENT_X;
                                } else
                                        unreachable("Invalid texture 2D components");
                        }

                        break;
                }

                case nir_tex_src_bias:
                case nir_tex_src_lod: {
                        /* Try as a constant if we can */

                        bool is_txf = midgard_texop == TEXTURE_OP_TEXEL_FETCH;
                        if (!is_txf && pan_attach_constant_bias(ctx, instr->src[i].src, &ins.texture))
                                break;

                        ins.texture.lod_register = true;
                        ins.src[2] = index;
                        emit_explicit_constant(ctx, index, index);

                        break;
                };

                default:
                        unreachable("Unknown texture source type\n");
                }
        }

        emit_mir_instruction(ctx, ins);

        /* Used for .cont and .last hinting */
        ctx->texture_op_count++;
}

static void
emit_tex(compiler_context *ctx, nir_tex_instr *instr)
{
        /* Fixup op, since only textureLod is permitted in VS but NIR can give
         * generic tex in some cases (which confuses the hardware) */

        bool is_vertex = ctx->stage == MESA_SHADER_VERTEX;

        if (is_vertex && instr->op == nir_texop_tex)
                instr->op = nir_texop_txl;

        switch (instr->op) {
        case nir_texop_tex:
        case nir_texop_txb:
                emit_texop_native(ctx, instr, TEXTURE_OP_NORMAL);
                break;
        case nir_texop_txl:
                emit_texop_native(ctx, instr, TEXTURE_OP_LOD);
                break;
        case nir_texop_txf:
                emit_texop_native(ctx, instr, TEXTURE_OP_TEXEL_FETCH);
                break;
        case nir_texop_txs:
                emit_sysval_read(ctx, &instr->instr, ~0, 4);
                break;
        default:
                unreachable("Unhanlded texture op");
        }
}

static void
emit_jump(compiler_context *ctx, nir_jump_instr *instr)
{
        switch (instr->type) {
        case nir_jump_break: {
                /* Emit a branch out of the loop */
                struct midgard_instruction br = v_branch(false, false);
                br.branch.target_type = TARGET_BREAK;
                br.branch.target_break = ctx->current_loop_depth;
                emit_mir_instruction(ctx, br);
                break;
        }

        default:
                DBG("Unknown jump type %d\n", instr->type);
                break;
        }
}

static void
emit_instr(compiler_context *ctx, struct nir_instr *instr)
{
        switch (instr->type) {
        case nir_instr_type_load_const:
                emit_load_const(ctx, nir_instr_as_load_const(instr));
                break;

        case nir_instr_type_intrinsic:
                emit_intrinsic(ctx, nir_instr_as_intrinsic(instr));
                break;

        case nir_instr_type_alu:
                emit_alu(ctx, nir_instr_as_alu(instr));
                break;

        case nir_instr_type_tex:
                emit_tex(ctx, nir_instr_as_tex(instr));
                break;

        case nir_instr_type_jump:
                emit_jump(ctx, nir_instr_as_jump(instr));
                break;

        case nir_instr_type_ssa_undef:
                /* Spurious */
                break;

        default:
                DBG("Unhandled instruction type\n");
                break;
        }
}


/* ALU instructions can inline or embed constants, which decreases register
 * pressure and saves space. */

#define CONDITIONAL_ATTACH(idx) { \
	void *entry = _mesa_hash_table_u64_search(ctx->ssa_constants, alu->src[idx] + 1); \
\
	if (entry) { \
		attach_constants(ctx, alu, entry, alu->src[idx] + 1); \
		alu->src[idx] = SSA_FIXED_REGISTER(REGISTER_CONSTANT); \
	} \
}

static void
inline_alu_constants(compiler_context *ctx, midgard_block *block)
{
        mir_foreach_instr_in_block(block, alu) {
                /* Other instructions cannot inline constants */
                if (alu->type != TAG_ALU_4) continue;
                if (alu->compact_branch) continue;

                /* If there is already a constant here, we can do nothing */
                if (alu->has_constants) continue;

                CONDITIONAL_ATTACH(0);

                if (!alu->has_constants) {
                        CONDITIONAL_ATTACH(1)
                } else if (!alu->inline_constant) {
                        /* Corner case: _two_ vec4 constants, for instance with a
                         * csel. For this case, we can only use a constant
                         * register for one, we'll have to emit a move for the
                         * other. Note, if both arguments are constants, then
                         * necessarily neither argument depends on the value of
                         * any particular register. As the destination register
                         * will be wiped, that means we can spill the constant
                         * to the destination register.
                         */

                        void *entry = _mesa_hash_table_u64_search(ctx->ssa_constants, alu->src[1] + 1);
                        unsigned scratch = alu->dest;

                        if (entry) {
                                midgard_instruction ins = v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT), scratch);
                                attach_constants(ctx, &ins, entry, alu->src[1] + 1);

                                /* Set the source */
                                alu->src[1] = scratch;

                                /* Inject us -before- the last instruction which set r31 */
                                mir_insert_instruction_before(ctx, mir_prev_op(alu), ins);
                        }
                }
        }
}

/* Being a little silly with the names, but returns the op that is the bitwise
 * inverse of the op with the argument switched. I.e. (f and g are
 * contrapositives):
 *
 * f(a, b) = ~g(b, a)
 *
 * Corollary: if g is the contrapositve of f, f is the contrapositive of g:
 *
 *      f(a, b) = ~g(b, a)
 *      ~f(a, b) = g(b, a)
 *      ~f(a, b) = ~h(a, b) where h is the contrapositive of g
 *      f(a, b) = h(a, b)
 *
 * Thus we define this function in pairs.
 */

static inline midgard_alu_op
mir_contrapositive(midgard_alu_op op)
{
        switch (op) {
        case midgard_alu_op_flt:
                return midgard_alu_op_fle;
        case midgard_alu_op_fle:
                return midgard_alu_op_flt;

        case midgard_alu_op_ilt:
                return midgard_alu_op_ile;
        case midgard_alu_op_ile:
                return midgard_alu_op_ilt;

        default:
                unreachable("No known contrapositive");
        }
}

/* Midgard supports two types of constants, embedded constants (128-bit) and
 * inline constants (16-bit). Sometimes, especially with scalar ops, embedded
 * constants can be demoted to inline constants, for space savings and
 * sometimes a performance boost */

static void
embedded_to_inline_constant(compiler_context *ctx, midgard_block *block)
{
        mir_foreach_instr_in_block(block, ins) {
                if (!ins->has_constants) continue;
                if (ins->has_inline_constant) continue;

                /* Blend constants must not be inlined by definition */
                if (ins->has_blend_constant) continue;

                /* We can inline 32-bit (sometimes) or 16-bit (usually) */
                bool is_16 = ins->alu.reg_mode == midgard_reg_mode_16;
                bool is_32 = ins->alu.reg_mode == midgard_reg_mode_32;

                if (!(is_16 || is_32))
                        continue;

                /* src1 cannot be an inline constant due to encoding
                 * restrictions. So, if possible we try to flip the arguments
                 * in that case */

                int op = ins->alu.op;

                if (ins->src[0] == SSA_FIXED_REGISTER(REGISTER_CONSTANT)) {
                        bool flip = alu_opcode_props[op].props & OP_COMMUTES;

                        switch (op) {
                        /* Conditionals can be inverted */
                        case midgard_alu_op_flt:
                        case midgard_alu_op_ilt:
                        case midgard_alu_op_fle:
                        case midgard_alu_op_ile:
                                ins->alu.op = mir_contrapositive(ins->alu.op);
                                ins->invert = true;
                                flip = true;
                                break;

                        case midgard_alu_op_fcsel:
                        case midgard_alu_op_icsel:
                                DBG("Missed non-commutative flip (%s)\n", alu_opcode_props[op].name);
                        default:
                                break;
                        }

                        if (flip)
                                mir_flip(ins);
                }

                if (ins->src[1] == SSA_FIXED_REGISTER(REGISTER_CONSTANT)) {
                        /* Extract the source information */

                        midgard_vector_alu_src *src;
                        int q = ins->alu.src2;
                        midgard_vector_alu_src *m = (midgard_vector_alu_src *) &q;
                        src = m;

                        /* Component is from the swizzle. Take a nonzero component */
                        assert(ins->mask);
                        unsigned first_comp = ffs(ins->mask) - 1;
                        unsigned component = ins->swizzle[1][first_comp];

                        /* Scale constant appropriately, if we can legally */
                        uint16_t scaled_constant = 0;

                        if (midgard_is_integer_op(op) || is_16) {
                                unsigned int *iconstants = (unsigned int *) ins->constants;
                                scaled_constant = (uint16_t) iconstants[component];

                                /* Constant overflow after resize */
                                if (scaled_constant != iconstants[component])
                                        continue;
                        } else {
                                float *f = (float *) ins->constants;
                                float original = f[component];
                                scaled_constant = _mesa_float_to_half(original);

                                /* Check for loss of precision. If this is
                                 * mediump, we don't care, but for a highp
                                 * shader, we need to pay attention. NIR
                                 * doesn't yet tell us which mode we're in!
                                 * Practically this prevents most constants
                                 * from being inlined, sadly. */

                                float fp32 = _mesa_half_to_float(scaled_constant);

                                if (fp32 != original)
                                        continue;
                        }

                        /* We don't know how to handle these with a constant */

                        if (mir_nontrivial_source2_mod_simple(ins) || src->rep_low || src->rep_high) {
                                DBG("Bailing inline constant...\n");
                                continue;
                        }

                        /* Make sure that the constant is not itself a vector
                         * by checking if all accessed values are the same. */

                        uint32_t *cons = ins->constants;
                        uint32_t value = cons[component];

                        bool is_vector = false;
                        unsigned mask = effective_writemask(&ins->alu, ins->mask);

                        for (unsigned c = 0; c < MIR_VEC_COMPONENTS; ++c) {
                                /* We only care if this component is actually used */
                                if (!(mask & (1 << c)))
                                        continue;

                                uint32_t test = cons[ins->swizzle[1][c]];

                                if (test != value) {
                                        is_vector = true;
                                        break;
                                }
                        }

                        if (is_vector)
                                continue;

                        /* Get rid of the embedded constant */
                        ins->has_constants = false;
                        ins->src[1] = ~0;
                        ins->has_inline_constant = true;
                        ins->inline_constant = scaled_constant;
                }
        }
}

/* Dead code elimination for branches at the end of a block - only one branch
 * per block is legal semantically */

static void
midgard_opt_cull_dead_branch(compiler_context *ctx, midgard_block *block)
{
        bool branched = false;

        mir_foreach_instr_in_block_safe(block, ins) {
                if (!midgard_is_branch_unit(ins->unit)) continue;

                if (branched)
                        mir_remove_instruction(ins);

                branched = true;
        }
}

/* fmov.pos is an idiom for fpos. Propoagate the .pos up to the source, so then
 * the move can be propagated away entirely */

static bool
mir_compose_float_outmod(midgard_outmod_float *outmod, midgard_outmod_float comp)
{
        /* Nothing to do */
        if (comp == midgard_outmod_none)
                return true;

        if (*outmod == midgard_outmod_none) {
                *outmod = comp;
                return true;
        }

        /* TODO: Compose rules */
        return false;
}

static bool
midgard_opt_pos_propagate(compiler_context *ctx, midgard_block *block)
{
        bool progress = false;

        mir_foreach_instr_in_block_safe(block, ins) {
                if (ins->type != TAG_ALU_4) continue;
                if (ins->alu.op != midgard_alu_op_fmov) continue;
                if (ins->alu.outmod != midgard_outmod_pos) continue;

                /* TODO: Registers? */
                unsigned src = ins->src[1];
                if (src & IS_REG) continue;

                /* There might be a source modifier, too */
                if (mir_nontrivial_source2_mod(ins)) continue;

                /* Backpropagate the modifier */
                mir_foreach_instr_in_block_from_rev(block, v, mir_prev_op(ins)) {
                        if (v->type != TAG_ALU_4) continue;
                        if (v->dest != src) continue;

                        /* Can we even take a float outmod? */
                        if (midgard_is_integer_out_op(v->alu.op)) continue;

                        midgard_outmod_float temp = v->alu.outmod;
                        progress |= mir_compose_float_outmod(&temp, ins->alu.outmod);

                        /* Throw in the towel.. */
                        if (!progress) break;

                        /* Otherwise, transfer the modifier */
                        v->alu.outmod = temp;
                        ins->alu.outmod = midgard_outmod_none;

                        break;
                }
        }

        return progress;
}

static void
emit_fragment_epilogue(compiler_context *ctx)
{
        /* Just emit the last chunk with the branch */
        EMIT(alu_br_compact_cond, midgard_jmp_writeout_op_writeout, TAG_ALU_4, ~0, midgard_condition_always);
}

static midgard_block *
emit_block(compiler_context *ctx, nir_block *block)
{
        midgard_block *this_block = ctx->after_block;
        ctx->after_block = NULL;

        if (!this_block)
                this_block = create_empty_block(ctx);

        list_addtail(&this_block->link, &ctx->blocks);

        this_block->is_scheduled = false;
        ++ctx->block_count;

        /* Set up current block */
        list_inithead(&this_block->instructions);
        ctx->current_block = this_block;

        nir_foreach_instr(instr, block) {
                emit_instr(ctx, instr);
                ++ctx->instruction_count;
        }

        return this_block;
}

static midgard_block *emit_cf_list(struct compiler_context *ctx, struct exec_list *list);

static void
emit_if(struct compiler_context *ctx, nir_if *nif)
{
        midgard_block *before_block = ctx->current_block;

        /* Speculatively emit the branch, but we can't fill it in until later */
        EMIT(branch, true, true);
        midgard_instruction *then_branch = mir_last_in_block(ctx->current_block);
        then_branch->src[0] = nir_src_index(ctx, &nif->condition);

        /* Emit the two subblocks. */
        midgard_block *then_block = emit_cf_list(ctx, &nif->then_list);
        midgard_block *end_then_block = ctx->current_block;

        /* Emit a jump from the end of the then block to the end of the else */
        EMIT(branch, false, false);
        midgard_instruction *then_exit = mir_last_in_block(ctx->current_block);

        /* Emit second block, and check if it's empty */

        int else_idx = ctx->block_count;
        int count_in = ctx->instruction_count;
        midgard_block *else_block = emit_cf_list(ctx, &nif->else_list);
        midgard_block *end_else_block = ctx->current_block;
        int after_else_idx = ctx->block_count;

        /* Now that we have the subblocks emitted, fix up the branches */

        assert(then_block);
        assert(else_block);

        if (ctx->instruction_count == count_in) {
                /* The else block is empty, so don't emit an exit jump */
                mir_remove_instruction(then_exit);
                then_branch->branch.target_block = after_else_idx;
        } else {
                then_branch->branch.target_block = else_idx;
                then_exit->branch.target_block = after_else_idx;
        }

        /* Wire up the successors */

        ctx->after_block = create_empty_block(ctx);

        midgard_block_add_successor(before_block, then_block);
        midgard_block_add_successor(before_block, else_block);

        midgard_block_add_successor(end_then_block, ctx->after_block);
        midgard_block_add_successor(end_else_block, ctx->after_block);
}

static void
emit_loop(struct compiler_context *ctx, nir_loop *nloop)
{
        /* Remember where we are */
        midgard_block *start_block = ctx->current_block;

        /* Allocate a loop number, growing the current inner loop depth */
        int loop_idx = ++ctx->current_loop_depth;

        /* Get index from before the body so we can loop back later */
        int start_idx = ctx->block_count;

        /* Emit the body itself */
        midgard_block *loop_block = emit_cf_list(ctx, &nloop->body);

        /* Branch back to loop back */
        struct midgard_instruction br_back = v_branch(false, false);
        br_back.branch.target_block = start_idx;
        emit_mir_instruction(ctx, br_back);

        /* Mark down that branch in the graph. */
        midgard_block_add_successor(start_block, loop_block);
        midgard_block_add_successor(ctx->current_block, loop_block);

        /* Find the index of the block about to follow us (note: we don't add
         * one; blocks are 0-indexed so we get a fencepost problem) */
        int break_block_idx = ctx->block_count;

        /* Fix up the break statements we emitted to point to the right place,
         * now that we can allocate a block number for them */
        ctx->after_block = create_empty_block(ctx);

        list_for_each_entry_from(struct midgard_block, block, start_block, &ctx->blocks, link) {
                mir_foreach_instr_in_block(block, ins) {
                        if (ins->type != TAG_ALU_4) continue;
                        if (!ins->compact_branch) continue;
                        if (ins->prepacked_branch) continue;

                        /* We found a branch -- check the type to see if we need to do anything */
                        if (ins->branch.target_type != TARGET_BREAK) continue;

                        /* It's a break! Check if it's our break */
                        if (ins->branch.target_break != loop_idx) continue;

                        /* Okay, cool, we're breaking out of this loop.
                         * Rewrite from a break to a goto */

                        ins->branch.target_type = TARGET_GOTO;
                        ins->branch.target_block = break_block_idx;

                        midgard_block_add_successor(block, ctx->after_block);
                }
        }

        /* Now that we've finished emitting the loop, free up the depth again
         * so we play nice with recursion amid nested loops */
        --ctx->current_loop_depth;

        /* Dump loop stats */
        ++ctx->loop_count;
}

static midgard_block *
emit_cf_list(struct compiler_context *ctx, struct exec_list *list)
{
        midgard_block *start_block = NULL;

        foreach_list_typed(nir_cf_node, node, node, list) {
                switch (node->type) {
                case nir_cf_node_block: {
                        midgard_block *block = emit_block(ctx, nir_cf_node_as_block(node));

                        if (!start_block)
                                start_block = block;

                        break;
                }

                case nir_cf_node_if:
                        emit_if(ctx, nir_cf_node_as_if(node));
                        break;

                case nir_cf_node_loop:
                        emit_loop(ctx, nir_cf_node_as_loop(node));
                        break;

                case nir_cf_node_function:
                        assert(0);
                        break;
                }
        }

        return start_block;
}

/* Due to lookahead, we need to report the first tag executed in the command
 * stream and in branch targets. An initial block might be empty, so iterate
 * until we find one that 'works' */

static unsigned
midgard_get_first_tag_from_block(compiler_context *ctx, unsigned block_idx)
{
        midgard_block *initial_block = mir_get_block(ctx, block_idx);

        unsigned first_tag = 0;

        mir_foreach_block_from(ctx, initial_block, v) {
                if (v->quadword_count) {
                        midgard_bundle *initial_bundle =
                                util_dynarray_element(&v->bundles, midgard_bundle, 0);

                        first_tag = initial_bundle->tag;
                        break;
                }
        }

        return first_tag;
}

int
midgard_compile_shader_nir(struct midgard_screen *screen, nir_shader *nir, midgard_program *program, bool is_blend, unsigned gpu_id)
{
        struct util_dynarray *compiled = &program->compiled;

        midgard_debug = debug_get_option_midgard_debug();

        /* TODO: Bound against what? */
        compiler_context *ctx = rzalloc(NULL, compiler_context);

        ctx->nir = nir;
        ctx->screen = screen;
        ctx->stage = nir->info.stage;
        ctx->is_blend = is_blend;
        ctx->alpha_ref = program->alpha_ref;
        ctx->gpu_id = gpu_id;

        /* Start off with a safe cutoff, allowing usage of all 16 work
         * registers. Later, we'll promote uniform reads to uniform registers
         * if we determine it is beneficial to do so */
        ctx->uniform_cutoff = 8;

        /* Initialize at a global (not block) level hash tables */

        ctx->ssa_constants = _mesa_hash_table_u64_create(NULL);
        ctx->hash_to_temp = _mesa_hash_table_u64_create(NULL);
        ctx->sysval_to_id = _mesa_hash_table_u64_create(NULL);

        /* Record the varying mapping for the command stream's bookkeeping */

        struct exec_list *varyings =
                        ctx->stage == MESA_SHADER_VERTEX ? &nir->outputs : &nir->inputs;

        unsigned max_varying = 0;
        nir_foreach_variable(var, varyings) {
                unsigned loc = var->data.driver_location;
                unsigned sz = glsl_type_size(var->type, FALSE);

                for (int c = 0; c < sz; ++c) {
                        program->varyings[loc + c] = var->data.location + c;
                        max_varying = MAX2(max_varying, loc + c);
                }
        }

        /* Lower gl_Position pre-optimisation, but after lowering vars to ssa
         * (so we don't accidentally duplicate the epilogue since mesa/st has
         * messed with our I/O quite a bit already) */

        NIR_PASS_V(nir, nir_lower_vars_to_ssa);

        if (ctx->stage == MESA_SHADER_VERTEX) {
                NIR_PASS_V(nir, nir_lower_viewport_transform);
                NIR_PASS_V(nir, nir_lower_point_size, 1.0, 1024.0);
        }

        NIR_PASS_V(nir, nir_lower_var_copies);
        NIR_PASS_V(nir, nir_lower_vars_to_ssa);
        NIR_PASS_V(nir, nir_split_var_copies);
        NIR_PASS_V(nir, nir_lower_var_copies);
        NIR_PASS_V(nir, nir_lower_global_vars_to_local);
        NIR_PASS_V(nir, nir_lower_var_copies);
        NIR_PASS_V(nir, nir_lower_vars_to_ssa);

        NIR_PASS_V(nir, nir_lower_io, nir_var_all, glsl_type_size, 0);

        /* Optimisation passes */

        optimise_nir(nir);

        if (midgard_debug & MIDGARD_DBG_SHADERS) {
                nir_print_shader(nir, stdout);
        }

        /* Assign sysvals and counts, now that we're sure
         * (post-optimisation) */

        midgard_nir_assign_sysvals(ctx, nir);

        program->uniform_count = nir->num_uniforms;
        program->sysval_count = ctx->sysval_count;
        memcpy(program->sysvals, ctx->sysvals, sizeof(ctx->sysvals[0]) * ctx->sysval_count);

        nir_foreach_function(func, nir) {
                if (!func->impl)
                        continue;

                list_inithead(&ctx->blocks);
                ctx->block_count = 0;
                ctx->func = func;

                emit_cf_list(ctx, &func->impl->body);

                /* Emit empty exit block with successor */

                struct midgard_block *semi_end = ctx->current_block;

                struct midgard_block *end =
                        emit_block(ctx, func->impl->end_block);

                if (ctx->stage == MESA_SHADER_FRAGMENT)
                        emit_fragment_epilogue(ctx);

                midgard_block_add_successor(semi_end, end);

                break; /* TODO: Multi-function shaders */
        }

        util_dynarray_init(compiled, NULL);

        /* Per-block lowering before opts */

        mir_foreach_block(ctx, block) {
                inline_alu_constants(ctx, block);
                midgard_opt_promote_fmov(ctx, block);
                embedded_to_inline_constant(ctx, block);
        }
        /* MIR-level optimizations */

        bool progress = false;

        do {
                progress = false;

                mir_foreach_block(ctx, block) {
                        progress |= midgard_opt_pos_propagate(ctx, block);
                        progress |= midgard_opt_copy_prop(ctx, block);
                        progress |= midgard_opt_dead_code_eliminate(ctx, block);
                        progress |= midgard_opt_combine_projection(ctx, block);
                        progress |= midgard_opt_varying_projection(ctx, block);
                        progress |= midgard_opt_not_propagate(ctx, block);
                        progress |= midgard_opt_fuse_src_invert(ctx, block);
                        progress |= midgard_opt_fuse_dest_invert(ctx, block);
                        progress |= midgard_opt_csel_invert(ctx, block);
                }
        } while (progress);

        mir_foreach_block(ctx, block) {
                midgard_lower_invert(ctx, block);
                midgard_lower_derivatives(ctx, block);
        }

        /* Nested control-flow can result in dead branches at the end of the
         * block. This messes with our analysis and is just dead code, so cull
         * them */
        mir_foreach_block(ctx, block) {
                midgard_opt_cull_dead_branch(ctx, block);
        }

        /* Ensure we were lowered */
        mir_foreach_instr_global(ctx, ins) {
                assert(!ins->invert);
        }

        /* Schedule! */
        schedule_program(ctx);

        /* Now that all the bundles are scheduled and we can calculate block
         * sizes, emit actual branch instructions rather than placeholders */

        int br_block_idx = 0;

        mir_foreach_block(ctx, block) {
                util_dynarray_foreach(&block->bundles, midgard_bundle, bundle) {
                        for (int c = 0; c < bundle->instruction_count; ++c) {
                                midgard_instruction *ins = bundle->instructions[c];

                                if (!midgard_is_branch_unit(ins->unit)) continue;

                                if (ins->prepacked_branch) continue;

                                /* Parse some basic branch info */
                                bool is_compact = ins->unit == ALU_ENAB_BR_COMPACT;
                                bool is_conditional = ins->branch.conditional;
                                bool is_inverted = ins->branch.invert_conditional;
                                bool is_discard = ins->branch.target_type == TARGET_DISCARD;

                                /* Determine the block we're jumping to */
                                int target_number = ins->branch.target_block;

                                /* Report the destination tag */
                                int dest_tag = is_discard ? 0 : midgard_get_first_tag_from_block(ctx, target_number);

                                /* Count up the number of quadwords we're
                                 * jumping over = number of quadwords until
                                 * (br_block_idx, target_number) */

                                int quadword_offset = 0;

                                if (is_discard) {
                                        /* Ignored */
                                } else if (target_number > br_block_idx) {
                                        /* Jump forward */

                                        for (int idx = br_block_idx + 1; idx < target_number; ++idx) {
                                                midgard_block *blk = mir_get_block(ctx, idx);
                                                assert(blk);

                                                quadword_offset += blk->quadword_count;
                                        }
                                } else {
                                        /* Jump backwards */

                                        for (int idx = br_block_idx; idx >= target_number; --idx) {
                                                midgard_block *blk = mir_get_block(ctx, idx);
                                                assert(blk);

                                                quadword_offset -= blk->quadword_count;
                                        }
                                }

                                /* Unconditional extended branches (far jumps)
                                 * have issues, so we always use a conditional
                                 * branch, setting the condition to always for
                                 * unconditional. For compact unconditional
                                 * branches, cond isn't used so it doesn't
                                 * matter what we pick. */

                                midgard_condition cond =
                                        !is_conditional ? midgard_condition_always :
                                        is_inverted ? midgard_condition_false :
                                        midgard_condition_true;

                                midgard_jmp_writeout_op op =
                                        is_discard ? midgard_jmp_writeout_op_discard :
                                        (is_compact && !is_conditional) ? midgard_jmp_writeout_op_branch_uncond :
                                        midgard_jmp_writeout_op_branch_cond;

                                if (!is_compact) {
                                        midgard_branch_extended branch =
                                                midgard_create_branch_extended(
                                                        cond, op,
                                                        dest_tag,
                                                        quadword_offset);

                                        memcpy(&ins->branch_extended, &branch, sizeof(branch));
                                } else if (is_conditional || is_discard) {
                                        midgard_branch_cond branch = {
                                                .op = op,
                                                .dest_tag = dest_tag,
                                                .offset = quadword_offset,
                                                .cond = cond
                                        };

                                        assert(branch.offset == quadword_offset);

                                        memcpy(&ins->br_compact, &branch, sizeof(branch));
                                } else {
                                        assert(op == midgard_jmp_writeout_op_branch_uncond);

                                        midgard_branch_uncond branch = {
                                                .op = op,
                                                .dest_tag = dest_tag,
                                                .offset = quadword_offset,
                                                .unknown = 1
                                        };

                                        assert(branch.offset == quadword_offset);

                                        memcpy(&ins->br_compact, &branch, sizeof(branch));
                                }
                        }
                }

                ++br_block_idx;
        }

        /* Emit flat binary from the instruction arrays. Iterate each block in
         * sequence. Save instruction boundaries such that lookahead tags can
         * be assigned easily */

        /* Cache _all_ bundles in source order for lookahead across failed branches */

        int bundle_count = 0;
        mir_foreach_block(ctx, block) {
                bundle_count += block->bundles.size / sizeof(midgard_bundle);
        }
        midgard_bundle **source_order_bundles = malloc(sizeof(midgard_bundle *) * bundle_count);
        int bundle_idx = 0;
        mir_foreach_block(ctx, block) {
                util_dynarray_foreach(&block->bundles, midgard_bundle, bundle) {
                        source_order_bundles[bundle_idx++] = bundle;
                }
        }

        int current_bundle = 0;

        /* Midgard prefetches instruction types, so during emission we
         * need to lookahead. Unless this is the last instruction, in
         * which we return 1. Or if this is the second to last and the
         * last is an ALU, then it's also 1... */

        mir_foreach_block(ctx, block) {
                mir_foreach_bundle_in_block(block, bundle) {
                        int lookahead = 1;

                        if (current_bundle + 1 < bundle_count) {
                                uint8_t next = source_order_bundles[current_bundle + 1]->tag;

                                if (!(current_bundle + 2 < bundle_count) && IS_ALU(next)) {
                                        lookahead = 1;
                                } else {
                                        lookahead = next;
                                }
                        }

                        emit_binary_bundle(ctx, bundle, compiled, lookahead);
                        ++current_bundle;
                }

                /* TODO: Free deeper */
                //util_dynarray_fini(&block->instructions);
        }

        free(source_order_bundles);

        /* Report the very first tag executed */
        program->first_tag = midgard_get_first_tag_from_block(ctx, 0);

        /* Deal with off-by-one related to the fencepost problem */
        program->work_register_count = ctx->work_registers + 1;
        program->uniform_cutoff = ctx->uniform_cutoff;

        program->blend_patch_offset = ctx->blend_constant_offset;
        program->tls_size = ctx->tls_size;

        if (midgard_debug & MIDGARD_DBG_SHADERS)
                disassemble_midgard(program->compiled.data, program->compiled.size, gpu_id, ctx->stage);

        if (midgard_debug & MIDGARD_DBG_SHADERDB) {
                unsigned nr_bundles = 0, nr_ins = 0;

                /* Count instructions and bundles */

                mir_foreach_block(ctx, block) {
                        nr_bundles += util_dynarray_num_elements(
                                              &block->bundles, midgard_bundle);

                        mir_foreach_bundle_in_block(block, bun)
                                nr_ins += bun->instruction_count;
                }

                /* Calculate thread count. There are certain cutoffs by
                 * register count for thread count */

                unsigned nr_registers = program->work_register_count;

                unsigned nr_threads =
                        (nr_registers <= 4) ? 4 :
                        (nr_registers <= 8) ? 2 :
                        1;

                /* Dump stats */

                fprintf(stderr, "shader%d - %s shader: "
                        "%u inst, %u bundles, %u quadwords, "
                        "%u registers, %u threads, %u loops, "
                        "%u:%u spills:fills\n",
                        SHADER_DB_COUNT++,
                        gl_shader_stage_name(ctx->stage),
                        nr_ins, nr_bundles, ctx->quadword_count,
                        nr_registers, nr_threads,
                        ctx->loop_count,
                        ctx->spills, ctx->fills);
        }

        ralloc_free(ctx);

        return 0;
}