summaryrefslogtreecommitdiff
path: root/compiler/GHC/CmmToLlvm/CodeGen.hs
blob: bfeb39171de37f4b183ffc8ea92a237f2b2c6ee2 (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
{-# LANGUAGE CPP #-}
{-# LANGUAGE GADTs, MultiWayIf #-}
{-# OPTIONS_GHC -fno-warn-type-defaults #-}
{-# OPTIONS_GHC -Wno-incomplete-uni-patterns #-}

-- | Handle conversion of CmmProc to LLVM code.
module GHC.CmmToLlvm.CodeGen ( genLlvmProc ) where

import GHC.Prelude

import GHC.Driver.Session
import GHC.Driver.Ppr

import GHC.Platform
import GHC.Platform.Regs ( activeStgRegs )

import GHC.Llvm
import GHC.CmmToLlvm.Base
import GHC.CmmToLlvm.Regs

import GHC.Cmm.BlockId
import GHC.Cmm.CLabel
import GHC.Cmm
import GHC.Cmm.Ppr as PprCmm
import GHC.Cmm.Utils
import GHC.Cmm.Switch
import GHC.Cmm.Dataflow.Block
import GHC.Cmm.Dataflow.Graph
import GHC.Cmm.Dataflow.Collections

import GHC.Data.FastString
import GHC.Data.OrdList

import GHC.Types.ForeignCall
import GHC.Types.Unique.Supply
import GHC.Types.Unique

import GHC.Utils.Outputable
import GHC.Utils.Panic.Plain (massert)
import qualified GHC.Utils.Panic as Panic
import GHC.Utils.Misc

import Control.Monad.Trans.Class
import Control.Monad.Trans.Writer
import Control.Monad

import qualified Data.Semigroup as Semigroup
import Data.List ( nub )
import Data.Maybe ( catMaybes )

type Atomic = Bool
type LlvmStatements = OrdList LlvmStatement

data Signage = Signed | Unsigned deriving (Eq, Show)

-- -----------------------------------------------------------------------------
-- | Top-level of the LLVM proc Code generator
--
genLlvmProc :: RawCmmDecl -> LlvmM [LlvmCmmDecl]
genLlvmProc (CmmProc infos lbl live graph) = do
    let blocks = toBlockListEntryFirstFalseFallthrough graph
    (lmblocks, lmdata) <- basicBlocksCodeGen live blocks
    let info = mapLookup (g_entry graph) infos
        proc = CmmProc info lbl live (ListGraph lmblocks)
    return (proc:lmdata)

genLlvmProc _ = panic "genLlvmProc: case that shouldn't reach here!"

-- -----------------------------------------------------------------------------
-- * Block code generation
--

-- | Generate code for a list of blocks that make up a complete
-- procedure. The first block in the list is expected to be the entry
-- point.
basicBlocksCodeGen :: LiveGlobalRegs -> [CmmBlock]
                      -> LlvmM ([LlvmBasicBlock], [LlvmCmmDecl])
basicBlocksCodeGen _    []                     = panic "no entry block!"
basicBlocksCodeGen live cmmBlocks
  = do -- Emit the prologue
       -- N.B. this must be its own block to ensure that the entry block of the
       -- procedure has no predecessors, as required by the LLVM IR. See #17589
       -- and #11649.
       bid <- newBlockId
       (prologue, prologueTops) <- funPrologue live cmmBlocks
       let entryBlock = BasicBlock bid (fromOL prologue)

       -- Generate code
       (blocks, topss) <- fmap unzip $ mapM basicBlockCodeGen cmmBlocks

       -- Compose
       return (entryBlock : blocks, prologueTops ++ concat topss)


-- | Generate code for one block
basicBlockCodeGen :: CmmBlock -> LlvmM ( LlvmBasicBlock, [LlvmCmmDecl] )
basicBlockCodeGen block
  = do let (_, nodes, tail)  = blockSplit block
           id = entryLabel block
       (mid_instrs, top) <- stmtsToInstrs $ blockToList nodes
       (tail_instrs, top')  <- stmtToInstrs tail
       let instrs = fromOL (mid_instrs `appOL` tail_instrs)
       return (BasicBlock id instrs, top' ++ top)

-- -----------------------------------------------------------------------------
-- * CmmNode code generation
--

-- A statement conversion return data.
--   * LlvmStatements: The compiled LLVM statements.
--   * LlvmCmmDecl: Any global data needed.
type StmtData = (LlvmStatements, [LlvmCmmDecl])


-- | Convert a list of CmmNode's to LlvmStatement's
stmtsToInstrs :: [CmmNode e x] -> LlvmM StmtData
stmtsToInstrs stmts
   = do (instrss, topss) <- fmap unzip $ mapM stmtToInstrs stmts
        return (concatOL instrss, concat topss)


-- | Convert a CmmStmt to a list of LlvmStatement's
stmtToInstrs :: CmmNode e x -> LlvmM StmtData
stmtToInstrs stmt = case stmt of

    CmmComment _         -> return (nilOL, []) -- nuke comments
    CmmTick    _         -> return (nilOL, [])
    CmmUnwind  {}        -> return (nilOL, [])

    CmmAssign reg src    -> genAssign reg src
    CmmStore addr src    -> genStore addr src

    CmmBranch id         -> genBranch id
    CmmCondBranch arg true false likely
                         -> genCondBranch arg true false likely
    CmmSwitch arg ids    -> genSwitch arg ids

    -- Foreign Call
    CmmUnsafeForeignCall target res args
        -> genCall target res args

    -- Tail call
    CmmCall { cml_target = arg,
              cml_args_regs = live } -> genJump arg live

    _ -> panic "Llvm.CodeGen.stmtToInstrs"

-- | Wrapper function to declare an instrinct function by function type
getInstrinct2 :: LMString -> LlvmType -> LlvmM ExprData
getInstrinct2 fname fty@(LMFunction funSig) = do

    let fv   = LMGlobalVar fname fty (funcLinkage funSig) Nothing Nothing Constant

    fn <- funLookup fname
    tops <- case fn of
      Just _  ->
        return []
      Nothing -> do
        funInsert fname fty
        un <- getUniqueM
        let lbl = mkAsmTempLabel un
        return [CmmData (Section Data lbl) [([],[fty])]]

    return (fv, nilOL, tops)

getInstrinct2 _ _ = error "getInstrinct2: Non-function type!"

-- | Declares an instrinct function by return and parameter types
getInstrinct :: LMString -> LlvmType -> [LlvmType] -> LlvmM ExprData
getInstrinct fname retTy parTys =
    let funSig = LlvmFunctionDecl fname ExternallyVisible CC_Ccc retTy
                    FixedArgs (tysToParams parTys) Nothing
        fty = LMFunction funSig
    in getInstrinct2 fname fty

-- | Memory barrier instruction for LLVM >= 3.0
barrier :: LlvmM StmtData
barrier = do
    let s = Fence False SyncSeqCst
    return (unitOL s, [])

-- | Insert a 'barrier', unless the target platform is in the provided list of
--   exceptions (where no code will be emitted instead).
barrierUnless :: [Arch] -> LlvmM StmtData
barrierUnless exs = do
    platform <- getPlatform
    if platformArch platform `elem` exs
        then return (nilOL, [])
        else barrier

-- | Foreign Calls
genCall :: ForeignTarget -> [CmmFormal] -> [CmmActual] -> LlvmM StmtData

-- Barriers need to be handled specially as they are implemented as LLVM
-- intrinsic functions.
genCall (PrimTarget MO_ReadBarrier) _ _ =
    barrierUnless [ArchX86, ArchX86_64, ArchSPARC]

genCall (PrimTarget MO_WriteBarrier) _ _ =
    barrierUnless [ArchX86, ArchX86_64, ArchSPARC]

genCall (PrimTarget MO_Touch) _ _ =
    return (nilOL, [])

genCall (PrimTarget (MO_UF_Conv w)) [dst] [e] = runStmtsDecls $ do
    dstV <- getCmmRegW (CmmLocal dst)
    let ty = cmmToLlvmType $ localRegType dst
        width = widthToLlvmFloat w
    castV <- lift $ mkLocalVar ty
    ve <- exprToVarW e
    statement $ Assignment castV $ Cast LM_Uitofp ve width
    statement $ Store castV dstV

genCall (PrimTarget (MO_UF_Conv _)) [_] args =
    panic $ "genCall: Too many arguments to MO_UF_Conv. " ++
    "Can only handle 1, given" ++ show (length args) ++ "."

-- Handle prefetching data
genCall t@(PrimTarget (MO_Prefetch_Data localityInt)) [] args
  | 0 <= localityInt && localityInt <= 3 = runStmtsDecls $ do
    let argTy = [i8Ptr, i32, i32, i32]
        funTy = \name -> LMFunction $ LlvmFunctionDecl name ExternallyVisible
                             CC_Ccc LMVoid FixedArgs (tysToParams argTy) Nothing

    let (_, arg_hints) = foreignTargetHints t
    let args_hints' = zip args arg_hints
    argVars <- arg_varsW args_hints' ([], nilOL, [])
    fptr    <- liftExprData $ getFunPtr funTy t
    argVars' <- castVarsW Signed $ zip argVars argTy

    let argSuffix = [mkIntLit i32 0, mkIntLit i32 localityInt, mkIntLit i32 1]
    statement $ Expr $ Call StdCall fptr (argVars' ++ argSuffix) []
  | otherwise = panic $ "prefetch locality level integer must be between 0 and 3, given: " ++ (show localityInt)

-- Handle PopCnt, Clz, Ctz, and BSwap that need to only convert arg
-- and return types
genCall t@(PrimTarget (MO_PopCnt w)) dsts args =
    genCallSimpleCast w t dsts args

genCall t@(PrimTarget (MO_Pdep w)) dsts args =
    genCallSimpleCast2 w t dsts args
genCall t@(PrimTarget (MO_Pext w)) dsts args =
    genCallSimpleCast2 w t dsts args
genCall t@(PrimTarget (MO_Clz w)) dsts args =
    genCallSimpleCast w t dsts args
genCall t@(PrimTarget (MO_Ctz w)) dsts args =
    genCallSimpleCast w t dsts args
genCall t@(PrimTarget (MO_BSwap w)) dsts args =
    genCallSimpleCast w t dsts args
genCall t@(PrimTarget (MO_BRev w)) dsts args =
    genCallSimpleCast w t dsts args

genCall (PrimTarget (MO_AtomicRMW width amop)) [dst] [addr, n] = runStmtsDecls $ do
    addrVar <- exprToVarW addr
    nVar <- exprToVarW n
    let targetTy = widthToLlvmInt width
        ptrExpr = Cast LM_Inttoptr addrVar (pLift targetTy)
    ptrVar <- doExprW (pLift targetTy) ptrExpr
    dstVar <- getCmmRegW (CmmLocal dst)
    let op = case amop of
               AMO_Add  -> LAO_Add
               AMO_Sub  -> LAO_Sub
               AMO_And  -> LAO_And
               AMO_Nand -> LAO_Nand
               AMO_Or   -> LAO_Or
               AMO_Xor  -> LAO_Xor
    retVar <- doExprW targetTy $ AtomicRMW op ptrVar nVar SyncSeqCst
    statement $ Store retVar dstVar

genCall (PrimTarget (MO_AtomicRead _)) [dst] [addr] = runStmtsDecls $ do
    dstV <- getCmmRegW (CmmLocal dst)
    v1 <- genLoadW True addr (localRegType dst)
    statement $ Store v1 dstV

genCall (PrimTarget (MO_Cmpxchg _width))
        [dst] [addr, old, new] = runStmtsDecls $ do
    addrVar <- exprToVarW addr
    oldVar <- exprToVarW old
    newVar <- exprToVarW new
    let targetTy = getVarType oldVar
        ptrExpr = Cast LM_Inttoptr addrVar (pLift targetTy)
    ptrVar <- doExprW (pLift targetTy) ptrExpr
    dstVar <- getCmmRegW (CmmLocal dst)
    retVar <- doExprW (LMStructU [targetTy,i1])
              $ CmpXChg ptrVar oldVar newVar SyncSeqCst SyncSeqCst
    retVar' <- doExprW targetTy $ ExtractV retVar 0
    statement $ Store retVar' dstVar

genCall (PrimTarget (MO_Xchg _width)) [dst] [addr, val] = runStmtsDecls $ do
    dstV <- getCmmRegW (CmmLocal dst) :: WriterT LlvmAccum LlvmM LlvmVar
    addrVar <- exprToVarW addr
    valVar <- exprToVarW val
    let ptrTy = pLift $ getVarType valVar
        ptrExpr = Cast LM_Inttoptr addrVar ptrTy
    ptrVar <- doExprW ptrTy ptrExpr
    resVar <- doExprW (getVarType valVar) (AtomicRMW LAO_Xchg ptrVar valVar SyncSeqCst)
    statement $ Store resVar dstV

genCall (PrimTarget (MO_AtomicWrite _width)) [] [addr, val] = runStmtsDecls $ do
    addrVar <- exprToVarW addr
    valVar <- exprToVarW val
    let ptrTy = pLift $ getVarType valVar
        ptrExpr = Cast LM_Inttoptr addrVar ptrTy
    ptrVar <- doExprW ptrTy ptrExpr
    statement $ Expr $ AtomicRMW LAO_Xchg ptrVar valVar SyncSeqCst

-- Handle memcpy function specifically since llvm's intrinsic version takes
-- some extra parameters.
genCall t@(PrimTarget op) [] args
 | Just align <- machOpMemcpyishAlign op
 = do
   platform <- getPlatform
   runStmtsDecls $ do
    let isVolTy = [i1]
        isVolVal = [mkIntLit i1 0]
        argTy | MO_Memset _ <- op = [i8Ptr, i8,    llvmWord platform, i32] ++ isVolTy
              | otherwise         = [i8Ptr, i8Ptr, llvmWord platform, i32] ++ isVolTy
        funTy = \name -> LMFunction $ LlvmFunctionDecl name ExternallyVisible
                             CC_Ccc LMVoid FixedArgs (tysToParams argTy) Nothing

    let (_, arg_hints) = foreignTargetHints t
    let args_hints = zip args arg_hints
    argVars       <- arg_varsW args_hints ([], nilOL, [])
    fptr          <- getFunPtrW funTy t
    argVars' <- castVarsW Signed $ zip argVars argTy

    let alignVal = mkIntLit i32 align
        arguments = argVars' ++ (alignVal:isVolVal)
    statement $ Expr $ Call StdCall fptr arguments []

-- We handle MO_U_Mul2 by simply using a 'mul' instruction, but with operands
-- twice the width (we first zero-extend them), e.g., on 64-bit arch we will
-- generate 'mul' on 128-bit operands. Then we only need some plumbing to
-- extract the two 64-bit values out of 128-bit result.
genCall (PrimTarget (MO_U_Mul2 w)) [dstH, dstL] [lhs, rhs] = runStmtsDecls $ do
    let width = widthToLlvmInt w
        bitWidth = widthInBits w
        width2x = LMInt (bitWidth * 2)
    -- First zero-extend the operands ('mul' instruction requires the operands
    -- and the result to be of the same type). Note that we don't use 'castVars'
    -- because it tries to do LM_Sext.
    lhsVar <- exprToVarW lhs
    rhsVar <- exprToVarW rhs
    lhsExt <- doExprW width2x $ Cast LM_Zext lhsVar width2x
    rhsExt <- doExprW width2x $ Cast LM_Zext rhsVar width2x
    -- Do the actual multiplication (note that the result is also 2x width).
    retV <- doExprW width2x $ LlvmOp LM_MO_Mul lhsExt rhsExt
    -- Extract the lower bits of the result into retL.
    retL <- doExprW width $ Cast LM_Trunc retV width
    -- Now we unsigned right-shift the higher bits by width.
    let widthLlvmLit = LMLitVar $ LMIntLit (fromIntegral bitWidth) width
    retShifted <- doExprW width2x $ LlvmOp LM_MO_LShr retV widthLlvmLit
    -- And extract them into retH.
    retH <- doExprW width $ Cast LM_Trunc retShifted width
    dstRegL <- getCmmRegW (CmmLocal dstL)
    dstRegH <- getCmmRegW (CmmLocal dstH)
    statement $ Store retL dstRegL
    statement $ Store retH dstRegH

genCall (PrimTarget (MO_S_Mul2 w)) [dstC, dstH, dstL] [lhs, rhs] = runStmtsDecls $ do
    let width = widthToLlvmInt w
        bitWidth = widthInBits w
        width2x = LMInt (bitWidth * 2)
    -- First sign-extend the operands ('mul' instruction requires the operands
    -- and the result to be of the same type). Note that we don't use 'castVars'
    -- because it tries to do LM_Sext.
    lhsVar <- exprToVarW lhs
    rhsVar <- exprToVarW rhs
    lhsExt <- doExprW width2x $ Cast LM_Sext lhsVar width2x
    rhsExt <- doExprW width2x $ Cast LM_Sext rhsVar width2x
    -- Do the actual multiplication (note that the result is also 2x width).
    retV <- doExprW width2x $ LlvmOp LM_MO_Mul lhsExt rhsExt
    -- Extract the lower bits of the result into retL.
    retL <- doExprW width $ Cast LM_Trunc retV width
    -- Now we signed right-shift the higher bits by width.
    let widthLlvmLit = LMLitVar $ LMIntLit (fromIntegral bitWidth) width
    retShifted <- doExprW width2x $ LlvmOp LM_MO_AShr retV widthLlvmLit
    -- And extract them into retH.
    retH <- doExprW width $ Cast LM_Trunc retShifted width
    -- Check if the carry is useful by doing a full arithmetic right shift on
    -- retL and comparing the result with retH
    let widthLlvmLitm1 = LMLitVar $ LMIntLit (fromIntegral bitWidth - 1) width
    retH' <- doExprW width $ LlvmOp LM_MO_AShr retL widthLlvmLitm1
    retC1  <- doExprW i1 $ Compare LM_CMP_Ne retH retH' -- Compare op returns a 1-bit value (i1)
    retC   <- doExprW width $ Cast LM_Zext retC1 width  -- so we zero-extend it
    dstRegL <- getCmmRegW (CmmLocal dstL)
    dstRegH <- getCmmRegW (CmmLocal dstH)
    dstRegC <- getCmmRegW (CmmLocal dstC)
    statement $ Store retL dstRegL
    statement $ Store retH dstRegH
    statement $ Store retC dstRegC

-- MO_U_QuotRem2 is another case we handle by widening the registers to double
-- the width and use normal LLVM instructions (similarly to the MO_U_Mul2). The
-- main difference here is that we need to combine two words into one register
-- and then use both 'udiv' and 'urem' instructions to compute the result.
genCall (PrimTarget (MO_U_QuotRem2 w))
        [dstQ, dstR] [lhsH, lhsL, rhs] = runStmtsDecls $ do
    let width = widthToLlvmInt w
        bitWidth = widthInBits w
        width2x = LMInt (bitWidth * 2)
    -- First zero-extend all parameters to double width.
    let zeroExtend expr = do
            var <- exprToVarW expr
            doExprW width2x $ Cast LM_Zext var width2x
    lhsExtH <- zeroExtend lhsH
    lhsExtL <- zeroExtend lhsL
    rhsExt <- zeroExtend rhs
    -- Now we combine the first two parameters (that represent the high and low
    -- bits of the value). So first left-shift the high bits to their position
    -- and then bit-or them with the low bits.
    let widthLlvmLit = LMLitVar $ LMIntLit (fromIntegral bitWidth) width
    lhsExtHShifted <- doExprW width2x $ LlvmOp LM_MO_Shl lhsExtH widthLlvmLit
    lhsExt <- doExprW width2x $ LlvmOp LM_MO_Or lhsExtHShifted lhsExtL
    -- Finally, we can call 'udiv' and 'urem' to compute the results.
    retExtDiv <- doExprW width2x $ LlvmOp LM_MO_UDiv lhsExt rhsExt
    retExtRem <- doExprW width2x $ LlvmOp LM_MO_URem lhsExt rhsExt
    -- And since everything is in 2x width, we need to truncate the results and
    -- then return them.
    let narrow var = doExprW width $ Cast LM_Trunc var width
    retDiv <- narrow retExtDiv
    retRem <- narrow retExtRem
    dstRegQ <- lift $ getCmmReg (CmmLocal dstQ)
    dstRegR <- lift $ getCmmReg (CmmLocal dstR)
    statement $ Store retDiv dstRegQ
    statement $ Store retRem dstRegR

-- Handle the MO_{Add,Sub}IntC separately. LLVM versions return a record from
-- which we need to extract the actual values.
genCall t@(PrimTarget (MO_AddIntC w)) [dstV, dstO] [lhs, rhs] =
    genCallWithOverflow t w [dstV, dstO] [lhs, rhs]
genCall t@(PrimTarget (MO_SubIntC w)) [dstV, dstO] [lhs, rhs] =
    genCallWithOverflow t w [dstV, dstO] [lhs, rhs]

-- Similar to MO_{Add,Sub}IntC, but MO_Add2 expects the first element of the
-- return tuple to be the overflow bit and the second element to contain the
-- actual result of the addition. So we still use genCallWithOverflow but swap
-- the return registers.
genCall t@(PrimTarget (MO_Add2 w)) [dstO, dstV] [lhs, rhs] =
    genCallWithOverflow t w [dstV, dstO] [lhs, rhs]

genCall t@(PrimTarget (MO_AddWordC w)) [dstV, dstO] [lhs, rhs] =
    genCallWithOverflow t w [dstV, dstO] [lhs, rhs]

genCall t@(PrimTarget (MO_SubWordC w)) [dstV, dstO] [lhs, rhs] =
    genCallWithOverflow t w [dstV, dstO] [lhs, rhs]

-- Handle all other foreign calls and prim ops.
genCall target res args = do
  platform <- getPlatform
  runStmtsDecls $ do

    -- extract Cmm call convention, and translate to LLVM call convention
    let lmconv = case target of
            ForeignTarget _ (ForeignConvention conv _ _ _) ->
              case conv of
                 StdCallConv  -> case platformArch platform of
                                 ArchX86    -> CC_X86_Stdcc
                                 ArchX86_64 -> CC_X86_Stdcc
                                 _          -> CC_Ccc
                 CCallConv    -> CC_Ccc
                 CApiConv     -> CC_Ccc
                 PrimCallConv       -> panic "GHC.CmmToLlvm.CodeGen.genCall: PrimCallConv"
                 JavaScriptCallConv -> panic "GHC.CmmToLlvm.CodeGen.genCall: JavaScriptCallConv"

            PrimTarget   _ -> CC_Ccc

    {-
        CC_Ccc of the possibilities here are a worry with the use of a custom
        calling convention for passing STG args. In practice the more
        dangerous combinations (e.g StdCall + llvmGhcCC) don't occur.

        The native code generator only handles StdCall and CCallConv.
    -}

    -- parameter types
    let arg_type (_, AddrHint) = (i8Ptr, [])
        -- cast pointers to i8*. Llvm equivalent of void*
        arg_type (expr, hint) =
            case cmmToLlvmType $ cmmExprType platform expr of
              ty@(LMInt n) | n < 64 && lmconv == CC_Ccc && platformCConvNeedsExtension platform
                 -> (ty, if hint == SignedHint then [SignExt] else [ZeroExt])
              ty -> (ty, [])

    -- ret type
    let ret_type [] = LMVoid
        ret_type [(_, AddrHint)] = i8Ptr
        ret_type [(reg, _)]      = cmmToLlvmType $ localRegType reg
        ret_type t = panic $ "genCall: Too many return values! Can only handle"
                        ++ " 0 or 1, given " ++ show (length t) ++ "."

    -- call attributes
    let fnAttrs | never_returns = NoReturn : llvmStdFunAttrs
                | otherwise     = llvmStdFunAttrs

        never_returns = case target of
             ForeignTarget _ (ForeignConvention _ _ _ CmmNeverReturns) -> True
             _ -> False

    -- fun type
    let (res_hints, arg_hints) = foreignTargetHints target
    let args_hints = zip args arg_hints
    let ress_hints = zip res  res_hints
    let ccTy  = StdCall -- tail calls should be done through CmmJump
    let retTy = ret_type ress_hints
    let argTy = map arg_type args_hints
    let funTy = \name -> LMFunction $ LlvmFunctionDecl name ExternallyVisible
                             lmconv retTy FixedArgs argTy (llvmFunAlign platform)


    argVars <- arg_varsW args_hints ([], nilOL, [])
    fptr    <- getFunPtrW funTy target

    let doReturn | ccTy == TailCall  = statement $ Return Nothing
                 | never_returns     = statement $ Unreachable
                 | otherwise         = return ()


    -- make the actual call
    case retTy of
        LMVoid ->
            statement $ Expr $ Call ccTy fptr argVars fnAttrs
        _ -> do
            v1 <- doExprW retTy $ Call ccTy fptr argVars fnAttrs
            -- get the return register
            let ret_reg [reg] = reg
                ret_reg t = panic $ "genCall: Bad number of registers! Can only handle"
                                ++ " 1, given " ++ show (length t) ++ "."
            let creg = ret_reg res
            vreg <- getCmmRegW (CmmLocal creg)
            if retTy == pLower (getVarType vreg)
                then do
                    statement $ Store v1 vreg
                    doReturn
                else do
                    let ty = pLower $ getVarType vreg
                    let op = case ty of
                            vt | isPointer vt -> LM_Bitcast
                               | isInt     vt -> LM_Ptrtoint
                               | otherwise    ->
                                   panic $ "genCall: CmmReg bad match for"
                                        ++ " returned type!"

                    v2 <- doExprW ty $ Cast op v1 ty
                    statement $ Store v2 vreg
                    doReturn

-- | Generate a call to an LLVM intrinsic that performs arithmetic operation
-- with overflow bit (i.e., returns a struct containing the actual result of the
-- operation and an overflow bit). This function will also extract the overflow
-- bit and zero-extend it (all the corresponding Cmm PrimOps represent the
-- overflow "bit" as a usual Int# or Word#).
genCallWithOverflow
  :: ForeignTarget -> Width -> [CmmFormal] -> [CmmActual] -> LlvmM StmtData
genCallWithOverflow t@(PrimTarget op) w [dstV, dstO] [lhs, rhs] = do
    -- So far this was only tested for the following four CallishMachOps.
    let valid = op `elem`   [ MO_Add2 w
                            , MO_AddIntC w
                            , MO_SubIntC w
                            , MO_AddWordC w
                            , MO_SubWordC w
                            ]
    massert valid
    let width = widthToLlvmInt w
    -- This will do most of the work of generating the call to the intrinsic and
    -- extracting the values from the struct.
    (value, overflowBit, (stmts, top)) <-
      genCallExtract t w (lhs, rhs) (width, i1)
    -- value is i<width>, but overflowBit is i1, so we need to cast (Cmm expects
    -- both to be i<width>)
    (overflow, zext) <- doExpr width $ Cast LM_Zext overflowBit width
    dstRegV <- getCmmReg (CmmLocal dstV)
    dstRegO <- getCmmReg (CmmLocal dstO)
    let storeV = Store value dstRegV
        storeO = Store overflow dstRegO
    return (stmts `snocOL` zext `snocOL` storeV `snocOL` storeO, top)
genCallWithOverflow _ _ _ _ =
    panic "genCallExtract: wrong ForeignTarget or number of arguments"

-- | A helper function for genCallWithOverflow that handles generating the call
-- to the LLVM intrinsic and extracting the result from the struct to LlvmVars.
genCallExtract
    :: ForeignTarget           -- ^ PrimOp
    -> Width                   -- ^ Width of the operands.
    -> (CmmActual, CmmActual)  -- ^ Actual arguments.
    -> (LlvmType, LlvmType)    -- ^ LLVM types of the returned struct.
    -> LlvmM (LlvmVar, LlvmVar, StmtData)
genCallExtract target@(PrimTarget op) w (argA, argB) (llvmTypeA, llvmTypeB) = do
    let width = widthToLlvmInt w
        argTy = [width, width]
        retTy = LMStructU [llvmTypeA, llvmTypeB]

    -- Process the arguments.
    let args_hints = zip [argA, argB] (snd $ foreignTargetHints target)
    (argsV1, args1, top1) <- arg_vars args_hints ([], nilOL, [])
    (argsV2, args2) <- castVars Signed $ zip argsV1 argTy

    -- Get the function and make the call.
    fname <- cmmPrimOpFunctions op
    (fptr, _, top2) <- getInstrinct fname retTy argTy
    -- We use StdCall for primops. See also the last case of genCall.
    (retV, call) <- doExpr retTy $ Call StdCall fptr argsV2 []

    -- This will result in a two element struct, we need to use "extractvalue"
    -- to get them out of it.
    (res1, ext1) <- doExpr llvmTypeA (ExtractV retV 0)
    (res2, ext2) <- doExpr llvmTypeB (ExtractV retV 1)

    let stmts = args1 `appOL` args2 `snocOL` call `snocOL` ext1 `snocOL` ext2
        tops = top1 ++ top2
    return (res1, res2, (stmts, tops))

genCallExtract _ _ _ _ =
    panic "genCallExtract: unsupported ForeignTarget"

-- Handle simple function call that only need simple type casting, of the form:
--   truncate arg >>= \a -> call(a) >>= zext
--
-- since GHC only really has i32 and i64 types and things like Word8 are backed
-- by an i32 and just present a logical i8 range. So we must handle conversions
-- from i32 to i8 explicitly as LLVM is strict about types.
genCallSimpleCast :: Width -> ForeignTarget -> [CmmFormal] -> [CmmActual]
              -> LlvmM StmtData
genCallSimpleCast w t@(PrimTarget op) [dst] args = do
    let width = widthToLlvmInt w
        dstTy = cmmToLlvmType $ localRegType dst

    fname                       <- cmmPrimOpFunctions op
    (fptr, _, top3)             <- getInstrinct fname width [width]

    dstV                        <- getCmmReg (CmmLocal dst)

    let (_, arg_hints) = foreignTargetHints t
    let args_hints = zip args arg_hints
    (argsV, stmts2, top2)       <- arg_vars args_hints ([], nilOL, [])
    (argsV', stmts4)            <- castVars Signed $ zip argsV [width]
    (retV, s1)                  <- doExpr width $ Call StdCall fptr argsV' []
    (retVs', stmts5)            <- castVars (cmmPrimOpRetValSignage op) [(retV,dstTy)]
    let retV'                    = singletonPanic "genCallSimpleCast" retVs'
    let s2                       = Store retV' dstV

    let stmts = stmts2 `appOL` stmts4 `snocOL`
                s1 `appOL` stmts5 `snocOL` s2
    return (stmts, top2 ++ top3)
genCallSimpleCast _ _ dsts _ =
    panic ("genCallSimpleCast: " ++ show (length dsts) ++ " dsts")

-- Handle simple function call that only need simple type casting, of the form:
--   truncate arg >>= \a -> call(a) >>= zext
--
-- since GHC only really has i32 and i64 types and things like Word8 are backed
-- by an i32 and just present a logical i8 range. So we must handle conversions
-- from i32 to i8 explicitly as LLVM is strict about types.
genCallSimpleCast2 :: Width -> ForeignTarget -> [CmmFormal] -> [CmmActual]
              -> LlvmM StmtData
genCallSimpleCast2 w t@(PrimTarget op) [dst] args = do
    let width = widthToLlvmInt w
        dstTy = cmmToLlvmType $ localRegType dst

    fname                       <- cmmPrimOpFunctions op
    (fptr, _, top3)             <- getInstrinct fname width (const width <$> args)

    dstV                        <- getCmmReg (CmmLocal dst)

    let (_, arg_hints) = foreignTargetHints t
    let args_hints = zip args arg_hints
    (argsV, stmts2, top2)       <- arg_vars args_hints ([], nilOL, [])
    (argsV', stmts4)            <- castVars Signed $ zip argsV (const width <$> argsV)
    (retV, s1)                  <- doExpr width $ Call StdCall fptr argsV' []
    (retVs', stmts5)             <- castVars (cmmPrimOpRetValSignage op) [(retV,dstTy)]
    let retV'                    = singletonPanic "genCallSimpleCast2" retVs'
    let s2                       = Store retV' dstV

    let stmts = stmts2 `appOL` stmts4 `snocOL`
                s1 `appOL` stmts5 `snocOL` s2
    return (stmts, top2 ++ top3)
genCallSimpleCast2 _ _ dsts _ =
    panic ("genCallSimpleCast2: " ++ show (length dsts) ++ " dsts")

-- | Create a function pointer from a target.
getFunPtrW :: (LMString -> LlvmType) -> ForeignTarget
           -> WriterT LlvmAccum LlvmM LlvmVar
getFunPtrW funTy targ = liftExprData $ getFunPtr funTy targ

-- | Create a function pointer from a target.
getFunPtr :: (LMString -> LlvmType) -> ForeignTarget
          -> LlvmM ExprData
getFunPtr funTy targ = case targ of
    ForeignTarget (CmmLit (CmmLabel lbl)) _ -> do
        name <- strCLabel_llvm lbl
        getHsFunc' name (funTy name)

    ForeignTarget expr _ -> do
        (v1, stmts, top) <- exprToVar expr
        dflags <- getDynFlags
        let fty = funTy $ fsLit "dynamic"
            cast = case getVarType v1 of
                ty | isPointer ty -> LM_Bitcast
                ty | isInt ty     -> LM_Inttoptr

                ty -> panic $ "genCall: Expr is of bad type for function"
                              ++ " call! (" ++ showSDoc dflags (ppr ty) ++ ")"

        (v2,s1) <- doExpr (pLift fty) $ Cast cast v1 (pLift fty)
        return (v2, stmts `snocOL` s1, top)

    PrimTarget mop -> do
        name <- cmmPrimOpFunctions mop
        let fty = funTy name
        getInstrinct2 name fty

-- | Conversion of call arguments.
arg_varsW :: [(CmmActual, ForeignHint)]
          -> ([LlvmVar], LlvmStatements, [LlvmCmmDecl])
          -> WriterT LlvmAccum LlvmM [LlvmVar]
arg_varsW xs ys = do
    (vars, stmts, decls) <- lift $ arg_vars xs ys
    tell $ LlvmAccum stmts decls
    return vars

-- | Conversion of call arguments.
arg_vars :: [(CmmActual, ForeignHint)]
         -> ([LlvmVar], LlvmStatements, [LlvmCmmDecl])
         -> LlvmM ([LlvmVar], LlvmStatements, [LlvmCmmDecl])

arg_vars [] (vars, stmts, tops)
  = return (vars, stmts, tops)

arg_vars ((e, AddrHint):rest) (vars, stmts, tops)
  = do (v1, stmts', top') <- exprToVar e
       dflags <- getDynFlags
       let op = case getVarType v1 of
               ty | isPointer ty -> LM_Bitcast
               ty | isInt ty     -> LM_Inttoptr

               a  -> panic $ "genCall: Can't cast llvmType to i8*! ("
                           ++ showSDoc dflags (ppr a) ++ ")"

       (v2, s1) <- doExpr i8Ptr $ Cast op v1 i8Ptr
       arg_vars rest (vars ++ [v2], stmts `appOL` stmts' `snocOL` s1,
                               tops ++ top')

arg_vars ((e, _):rest) (vars, stmts, tops)
  = do (v1, stmts', top') <- exprToVar e
       arg_vars rest (vars ++ [v1], stmts `appOL` stmts', tops ++ top')


-- | Cast a collection of LLVM variables to specific types.
castVarsW :: Signage
          -> [(LlvmVar, LlvmType)]
          -> WriterT LlvmAccum LlvmM [LlvmVar]
castVarsW signage vars = do
    (vars, stmts) <- lift $ castVars signage vars
    tell $ LlvmAccum stmts mempty
    return vars

-- | Cast a collection of LLVM variables to specific types.
castVars :: Signage -> [(LlvmVar, LlvmType)]
         -> LlvmM ([LlvmVar], LlvmStatements)
castVars signage vars = do
                done <- mapM (uncurry (castVar signage)) vars
                let (vars', stmts) = unzip done
                return (vars', toOL stmts)

-- | Cast an LLVM variable to a specific type, panicing if it can't be done.
castVar :: Signage -> LlvmVar -> LlvmType -> LlvmM (LlvmVar, LlvmStatement)
castVar signage v t | getVarType v == t
            = return (v, Nop)

            | otherwise
            = do dflags <- getDynFlags
                 platform <- getPlatform
                 let op = case (getVarType v, t) of
                      (LMInt n, LMInt m)
                          -> if n < m then extend else LM_Trunc
                      (vt, _) | isFloat vt && isFloat t
                          -> if llvmWidthInBits platform vt < llvmWidthInBits platform t
                                then LM_Fpext else LM_Fptrunc
                      (vt, _) | isInt vt && isFloat t       -> LM_Sitofp
                      (vt, _) | isFloat vt && isInt t       -> LM_Fptosi
                      (vt, _) | isInt vt && isPointer t     -> LM_Inttoptr
                      (vt, _) | isPointer vt && isInt t     -> LM_Ptrtoint
                      (vt, _) | isPointer vt && isPointer t -> LM_Bitcast
                      (vt, _) | isVector vt && isVector t   -> LM_Bitcast

                      (vt, _) -> panic $ "castVars: Can't cast this type ("
                                  ++ showSDoc dflags (ppr vt) ++ ") to (" ++ showSDoc dflags (ppr t) ++ ")"
                 doExpr t $ Cast op v t
    where extend = case signage of
            Signed      -> LM_Sext
            Unsigned    -> LM_Zext


cmmPrimOpRetValSignage :: CallishMachOp -> Signage
cmmPrimOpRetValSignage mop = case mop of
    MO_Pdep _   -> Unsigned
    MO_Pext _   -> Unsigned
    _           -> Signed

-- | Decide what C function to use to implement a CallishMachOp
cmmPrimOpFunctions :: CallishMachOp -> LlvmM LMString
cmmPrimOpFunctions mop = do

  dflags <- getDynFlags
  platform <- getPlatform
  let intrinTy1 = "p0i8.p0i8." ++ showSDoc dflags (ppr $ llvmWord platform)
      intrinTy2 = "p0i8." ++ showSDoc dflags (ppr $ llvmWord platform)
      unsupported = panic ("cmmPrimOpFunctions: " ++ show mop
                        ++ " not supported here")

  return $ case mop of
    MO_F32_Exp    -> fsLit "expf"
    MO_F32_ExpM1  -> fsLit "expm1f"
    MO_F32_Log    -> fsLit "logf"
    MO_F32_Log1P  -> fsLit "log1pf"
    MO_F32_Sqrt   -> fsLit "llvm.sqrt.f32"
    MO_F32_Fabs   -> fsLit "llvm.fabs.f32"
    MO_F32_Pwr    -> fsLit "llvm.pow.f32"

    MO_F32_Sin    -> fsLit "llvm.sin.f32"
    MO_F32_Cos    -> fsLit "llvm.cos.f32"
    MO_F32_Tan    -> fsLit "tanf"

    MO_F32_Asin   -> fsLit "asinf"
    MO_F32_Acos   -> fsLit "acosf"
    MO_F32_Atan   -> fsLit "atanf"

    MO_F32_Sinh   -> fsLit "sinhf"
    MO_F32_Cosh   -> fsLit "coshf"
    MO_F32_Tanh   -> fsLit "tanhf"

    MO_F32_Asinh  -> fsLit "asinhf"
    MO_F32_Acosh  -> fsLit "acoshf"
    MO_F32_Atanh  -> fsLit "atanhf"

    MO_F64_Exp    -> fsLit "exp"
    MO_F64_ExpM1  -> fsLit "expm1"
    MO_F64_Log    -> fsLit "log"
    MO_F64_Log1P  -> fsLit "log1p"
    MO_F64_Sqrt   -> fsLit "llvm.sqrt.f64"
    MO_F64_Fabs   -> fsLit "llvm.fabs.f64"
    MO_F64_Pwr    -> fsLit "llvm.pow.f64"

    MO_F64_Sin    -> fsLit "llvm.sin.f64"
    MO_F64_Cos    -> fsLit "llvm.cos.f64"
    MO_F64_Tan    -> fsLit "tan"

    MO_F64_Asin   -> fsLit "asin"
    MO_F64_Acos   -> fsLit "acos"
    MO_F64_Atan   -> fsLit "atan"

    MO_F64_Sinh   -> fsLit "sinh"
    MO_F64_Cosh   -> fsLit "cosh"
    MO_F64_Tanh   -> fsLit "tanh"

    MO_F64_Asinh  -> fsLit "asinh"
    MO_F64_Acosh  -> fsLit "acosh"
    MO_F64_Atanh  -> fsLit "atanh"

    MO_Memcpy _   -> fsLit $ "llvm.memcpy."  ++ intrinTy1
    MO_Memmove _  -> fsLit $ "llvm.memmove." ++ intrinTy1
    MO_Memset _   -> fsLit $ "llvm.memset."  ++ intrinTy2
    MO_Memcmp _   -> fsLit $ "memcmp"

    MO_SuspendThread -> fsLit $ "suspendThread"
    MO_ResumeThread  -> fsLit $ "resumeThread"

    (MO_PopCnt w) -> fsLit $ "llvm.ctpop."      ++ showSDoc dflags (ppr $ widthToLlvmInt w)
    (MO_BSwap w)  -> fsLit $ "llvm.bswap."      ++ showSDoc dflags (ppr $ widthToLlvmInt w)
    (MO_BRev w)   -> fsLit $ "llvm.bitreverse." ++ showSDoc dflags (ppr $ widthToLlvmInt w)
    (MO_Clz w)    -> fsLit $ "llvm.ctlz."       ++ showSDoc dflags (ppr $ widthToLlvmInt w)
    (MO_Ctz w)    -> fsLit $ "llvm.cttz."       ++ showSDoc dflags (ppr $ widthToLlvmInt w)

    (MO_Pdep w)   ->  let w' = showSDoc dflags (ppr $ widthInBits w)
                      in  if isBmi2Enabled dflags
                            then fsLit $ "llvm.x86.bmi.pdep."   ++ w'
                            else fsLit $ "hs_pdep"              ++ w'
    (MO_Pext w)   ->  let w' = showSDoc dflags (ppr $ widthInBits w)
                      in  if isBmi2Enabled dflags
                            then fsLit $ "llvm.x86.bmi.pext."   ++ w'
                            else fsLit $ "hs_pext"              ++ w'

    (MO_Prefetch_Data _ )-> fsLit "llvm.prefetch"

    MO_AddIntC w    -> fsLit $ "llvm.sadd.with.overflow."
                             ++ showSDoc dflags (ppr $ widthToLlvmInt w)
    MO_SubIntC w    -> fsLit $ "llvm.ssub.with.overflow."
                             ++ showSDoc dflags (ppr $ widthToLlvmInt w)
    MO_Add2 w       -> fsLit $ "llvm.uadd.with.overflow."
                             ++ showSDoc dflags (ppr $ widthToLlvmInt w)
    MO_AddWordC w   -> fsLit $ "llvm.uadd.with.overflow."
                             ++ showSDoc dflags (ppr $ widthToLlvmInt w)
    MO_SubWordC w   -> fsLit $ "llvm.usub.with.overflow."
                             ++ showSDoc dflags (ppr $ widthToLlvmInt w)

    MO_S_Mul2    {}  -> unsupported
    MO_S_QuotRem {}  -> unsupported
    MO_U_QuotRem {}  -> unsupported
    MO_U_QuotRem2 {} -> unsupported
    -- We support MO_U_Mul2 through ordinary LLVM mul instruction, see the
    -- appropriate case of genCall.
    MO_U_Mul2 {}     -> unsupported
    MO_ReadBarrier   -> unsupported
    MO_WriteBarrier  -> unsupported
    MO_Touch         -> unsupported
    MO_UF_Conv _     -> unsupported

    MO_AtomicRead _  -> unsupported
    MO_AtomicRMW _ _ -> unsupported
    MO_AtomicWrite _ -> unsupported
    MO_Cmpxchg _     -> unsupported
    MO_Xchg _        -> unsupported

-- | Tail function calls
genJump :: CmmExpr -> [GlobalReg] -> LlvmM StmtData

-- Call to known function
genJump (CmmLit (CmmLabel lbl)) live = do
    (vf, stmts, top) <- getHsFunc live lbl
    (stgRegs, stgStmts) <- funEpilogue live
    let s1  = Expr $ Call TailCall vf stgRegs llvmStdFunAttrs
    let s2  = Return Nothing
    return (stmts `appOL` stgStmts `snocOL` s1 `snocOL` s2, top)


-- Call to unknown function / address
genJump expr live = do
    fty <- llvmFunTy live
    (vf, stmts, top) <- exprToVar expr
    dflags <- getDynFlags

    let cast = case getVarType vf of
         ty | isPointer ty -> LM_Bitcast
         ty | isInt ty     -> LM_Inttoptr

         ty -> panic $ "genJump: Expr is of bad type for function call! ("
                     ++ showSDoc dflags (ppr ty) ++ ")"

    (v1, s1) <- doExpr (pLift fty) $ Cast cast vf (pLift fty)
    (stgRegs, stgStmts) <- funEpilogue live
    let s2 = Expr $ Call TailCall v1 stgRegs llvmStdFunAttrs
    let s3 = Return Nothing
    return (stmts `snocOL` s1 `appOL` stgStmts `snocOL` s2 `snocOL` s3,
            top)


-- | CmmAssign operation
--
-- We use stack allocated variables for CmmReg. The optimiser will replace
-- these with registers when possible.
genAssign :: CmmReg -> CmmExpr -> LlvmM StmtData
genAssign reg val = do
    vreg <- getCmmReg reg
    (vval, stmts2, top2) <- exprToVar val
    let stmts = stmts2

    let ty = (pLower . getVarType) vreg
    platform <- getPlatform
    case ty of
      -- Some registers are pointer types, so need to cast value to pointer
      LMPointer _ | getVarType vval == llvmWord platform -> do
          (v, s1) <- doExpr ty $ Cast LM_Inttoptr vval ty
          let s2 = Store v vreg
          return (stmts `snocOL` s1 `snocOL` s2, top2)

      LMVector _ _ -> do
          (v, s1) <- doExpr ty $ Cast LM_Bitcast vval ty
          let s2 = Store v vreg
          return (stmts `snocOL` s1 `snocOL` s2, top2)

      _ -> do
          let s1 = Store vval vreg
          return (stmts `snocOL` s1, top2)


-- | CmmStore operation
genStore :: CmmExpr -> CmmExpr -> LlvmM StmtData

-- First we try to detect a few common cases and produce better code for
-- these then the default case. We are mostly trying to detect Cmm code
-- like I32[Sp + n] and use 'getelementptr' operations instead of the
-- generic case that uses casts and pointer arithmetic
genStore addr@(CmmReg (CmmGlobal r)) val
    = genStore_fast addr r 0 val

genStore addr@(CmmRegOff (CmmGlobal r) n) val
    = genStore_fast addr r n val

genStore addr@(CmmMachOp (MO_Add _) [
                            (CmmReg (CmmGlobal r)),
                            (CmmLit (CmmInt n _))])
                val
    = genStore_fast addr r (fromInteger n) val

genStore addr@(CmmMachOp (MO_Sub _) [
                            (CmmReg (CmmGlobal r)),
                            (CmmLit (CmmInt n _))])
                val
    = genStore_fast addr r (negate $ fromInteger n) val

-- generic case
genStore addr val
    = getTBAAMeta topN >>= genStore_slow addr val

-- | CmmStore operation
-- This is a special case for storing to a global register pointer
-- offset such as I32[Sp+8].
genStore_fast :: CmmExpr -> GlobalReg -> Int -> CmmExpr
              -> LlvmM StmtData
genStore_fast addr r n val
  = do platform <- getPlatform
       (gv, grt, s1) <- getCmmRegVal (CmmGlobal r)
       meta          <- getTBAARegMeta r
       let (ix,rem) = n `divMod` ((llvmWidthInBits platform . pLower) grt  `div` 8)
       case isPointer grt && rem == 0 of
            True -> do
                (vval,  stmts, top) <- exprToVar val
                (ptr, s2) <- doExpr grt $ GetElemPtr True gv [toI32 ix]
                -- We might need a different pointer type, so check
                case pLower grt == getVarType vval of
                     -- were fine
                     True  -> do
                         let s3 = MetaStmt meta $ Store vval ptr
                         return (stmts `appOL` s1 `snocOL` s2
                                 `snocOL` s3, top)

                     -- cast to pointer type needed
                     False -> do
                         let ty = (pLift . getVarType) vval
                         (ptr', s3) <- doExpr ty $ Cast LM_Bitcast ptr ty
                         let s4 = MetaStmt meta $ Store vval ptr'
                         return (stmts `appOL` s1 `snocOL` s2
                                 `snocOL` s3 `snocOL` s4, top)

            -- If its a bit type then we use the slow method since
            -- we can't avoid casting anyway.
            False -> genStore_slow addr val meta


-- | CmmStore operation
-- Generic case. Uses casts and pointer arithmetic if needed.
genStore_slow :: CmmExpr -> CmmExpr -> [MetaAnnot] -> LlvmM StmtData
genStore_slow addr val meta = do
    (vaddr, stmts1, top1) <- exprToVar addr
    (vval,  stmts2, top2) <- exprToVar val

    let stmts = stmts1 `appOL` stmts2
    dflags <- getDynFlags
    platform <- getPlatform
    opts <- getLlvmOpts
    case getVarType vaddr of
        -- sometimes we need to cast an int to a pointer before storing
        LMPointer ty@(LMPointer _) | getVarType vval == llvmWord platform -> do
            (v, s1) <- doExpr ty $ Cast LM_Inttoptr vval ty
            let s2 = MetaStmt meta $ Store v vaddr
            return (stmts `snocOL` s1 `snocOL` s2, top1 ++ top2)

        LMPointer _ -> do
            let s1 = MetaStmt meta $ Store vval vaddr
            return (stmts `snocOL` s1, top1 ++ top2)

        i@(LMInt _) | i == llvmWord platform -> do
            let vty = pLift $ getVarType vval
            (vptr, s1) <- doExpr vty $ Cast LM_Inttoptr vaddr vty
            let s2 = MetaStmt meta $ Store vval vptr
            return (stmts `snocOL` s1 `snocOL` s2, top1 ++ top2)

        other ->
            pprPanic "genStore: ptr not right type!"
                    (PprCmm.pprExpr platform addr <+> text (
                        "Size of Ptr: " ++ show (llvmPtrBits platform) ++
                        ", Size of var: " ++ show (llvmWidthInBits platform other) ++
                        ", Var: " ++ showSDoc dflags (ppVar opts vaddr)))


-- | Unconditional branch
genBranch :: BlockId -> LlvmM StmtData
genBranch id =
    let label = blockIdToLlvm id
    in return (unitOL $ Branch label, [])


-- | Conditional branch
genCondBranch :: CmmExpr -> BlockId -> BlockId -> Maybe Bool -> LlvmM StmtData
genCondBranch cond idT idF likely = do
    let labelT = blockIdToLlvm idT
    let labelF = blockIdToLlvm idF
    -- See Note [Literals and branch conditions].
    (vc, stmts1, top1) <- exprToVarOpt i1Option cond
    if getVarType vc == i1
        then do
            (vc', (stmts2, top2)) <- case likely of
              Just b -> genExpectLit (if b then 1 else 0) i1  vc
              _      -> pure (vc, (nilOL, []))
            let s1 = BranchIf vc' labelT labelF
            return (stmts1 `appOL` stmts2 `snocOL` s1, top1 ++ top2)
        else do
            dflags <- getDynFlags
            opts <- getLlvmOpts
            panic $ "genCondBranch: Cond expr not bool! (" ++ showSDoc dflags (ppVar opts vc) ++ ")"


-- | Generate call to llvm.expect.x intrinsic. Assigning result to a new var.
genExpectLit :: Integer -> LlvmType -> LlvmVar -> LlvmM (LlvmVar, StmtData)
genExpectLit expLit expTy var = do
  dflags <- getDynFlags

  let
    lit = LMLitVar $ LMIntLit expLit expTy

    llvmExpectName
      | isInt expTy = fsLit $ "llvm.expect." ++ showSDoc dflags (ppr expTy)
      | otherwise   = panic $ "genExpectedLit: Type not an int!"

  (llvmExpect, stmts, top) <-
    getInstrinct llvmExpectName expTy [expTy, expTy]
  (var', call) <- doExpr expTy $ Call StdCall llvmExpect [var, lit] []
  return (var', (stmts `snocOL` call, top))

{- Note [Literals and branch conditions]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

It is important that whenever we generate branch conditions for
literals like '1', they are properly narrowed to an LLVM expression of
type 'i1' (for bools.) Otherwise, nobody is happy. So when we convert
a CmmExpr to an LLVM expression for a branch conditional, exprToVarOpt
must be certain to return a properly narrowed type. genLit is
responsible for this, in the case of literal integers.

Often, we won't see direct statements like:

    if(1) {
      ...
    } else {
      ...
    }

at this point in the pipeline, because the Glorious Code Generator
will do trivial branch elimination in the sinking pass (among others,)
which will eliminate the expression entirely.

However, it's certainly possible and reasonable for this to occur in
hand-written C-- code. Consider something like:

    #if !defined(SOME_CONDITIONAL)
    #define CHECK_THING(x) 1
    #else
    #define CHECK_THING(x) some_operation((x))
    #endif

    f() {

      if (CHECK_THING(xyz)) {
        ...
      } else {
        ...
      }

    }

In such an instance, CHECK_THING might result in an *expression* in
one case, and a *literal* in the other, depending on what in
particular was #define'd. So we must be sure to properly narrow the
literal in this case to i1 as it won't be eliminated beforehand.

For a real example of this, see ./rts/StgStdThunks.cmm

-}



-- | Switch branch
genSwitch :: CmmExpr -> SwitchTargets -> LlvmM StmtData
genSwitch cond ids = do
    (vc, stmts, top) <- exprToVar cond
    let ty = getVarType vc

    let labels = [ (mkIntLit ty ix, blockIdToLlvm b)
                 | (ix, b) <- switchTargetsCases ids ]
    -- out of range is undefined, so let's just branch to first label
    let defLbl | Just l <- switchTargetsDefault ids = blockIdToLlvm l
               | otherwise                          = snd (head labels)

    let s1 = Switch vc defLbl labels
    return $ (stmts `snocOL` s1, top)


-- -----------------------------------------------------------------------------
-- * CmmExpr code generation
--

-- | An expression conversion return data:
--   * LlvmVar: The var holding the result of the expression
--   * LlvmStatements: Any statements needed to evaluate the expression
--   * LlvmCmmDecl: Any global data needed for this expression
type ExprData = (LlvmVar, LlvmStatements, [LlvmCmmDecl])

-- | Values which can be passed to 'exprToVar' to configure its
-- behaviour in certain circumstances.
--
-- Currently just used for determining if a comparison should return
-- a boolean (i1) or a word. See Note [Literals and branch conditions].
newtype EOption = EOption { i1Expected :: Bool }
-- XXX: EOption is an ugly and inefficient solution to this problem.

-- | i1 type expected (condition scrutinee).
i1Option :: EOption
i1Option = EOption True

-- | Word type expected (usual).
wordOption :: EOption
wordOption = EOption False

-- | Convert a CmmExpr to a list of LlvmStatements with the result of the
-- expression being stored in the returned LlvmVar.
exprToVar :: CmmExpr -> LlvmM ExprData
exprToVar = exprToVarOpt wordOption

exprToVarOpt :: EOption -> CmmExpr -> LlvmM ExprData
exprToVarOpt opt e = case e of

    CmmLit lit
        -> genLit opt lit

    CmmLoad e' ty
        -> genLoad False e' ty

    -- Cmmreg in expression is the value, so must load. If you want actual
    -- reg pointer, call getCmmReg directly.
    CmmReg r -> do
        (v1, ty, s1) <- getCmmRegVal r
        case isPointer ty of
             True  -> do
                 -- Cmm wants the value, so pointer types must be cast to ints
                 platform <- getPlatform
                 (v2, s2) <- doExpr (llvmWord platform) $ Cast LM_Ptrtoint v1 (llvmWord platform)
                 return (v2, s1 `snocOL` s2, [])

             False -> return (v1, s1, [])

    CmmMachOp op exprs
        -> genMachOp opt op exprs

    CmmRegOff r i
        -> do platform <- getPlatform
              exprToVar $ expandCmmReg platform (r, i)

    CmmStackSlot _ _
        -> panic "exprToVar: CmmStackSlot not supported!"


-- | Handle CmmMachOp expressions
genMachOp :: EOption -> MachOp -> [CmmExpr] -> LlvmM ExprData

-- Unary Machop
genMachOp _ op [x] = case op of

    MO_Not w ->
        let all1 = mkIntLit (widthToLlvmInt w) (-1)
        in negate (widthToLlvmInt w) all1 LM_MO_Xor

    MO_S_Neg w ->
        let all0 = mkIntLit (widthToLlvmInt w) 0
        in negate (widthToLlvmInt w) all0 LM_MO_Sub

    MO_F_Neg w ->
        let all0 = LMLitVar $ LMFloatLit (-0) (widthToLlvmFloat w)
        in negate (widthToLlvmFloat w) all0 LM_MO_FSub

    MO_SF_Conv _ w -> fiConv (widthToLlvmFloat w) LM_Sitofp
    MO_FS_Conv _ w -> fiConv (widthToLlvmInt w) LM_Fptosi

    MO_SS_Conv from to
        -> sameConv from (widthToLlvmInt to) LM_Trunc LM_Sext

    MO_UU_Conv from to
        -> sameConv from (widthToLlvmInt to) LM_Trunc LM_Zext

    MO_XX_Conv from to
        -> sameConv from (widthToLlvmInt to) LM_Trunc LM_Zext

    MO_FF_Conv from to
        -> sameConv from (widthToLlvmFloat to) LM_Fptrunc LM_Fpext

    MO_VS_Neg len w ->
        let ty    = widthToLlvmInt w
            vecty = LMVector len ty
            all0  = LMIntLit (-0) ty
            all0s = LMLitVar $ LMVectorLit (replicate len all0)
        in negateVec vecty all0s LM_MO_Sub

    MO_VF_Neg len w ->
        let ty    = widthToLlvmFloat w
            vecty = LMVector len ty
            all0  = LMFloatLit (-0) ty
            all0s = LMLitVar $ LMVectorLit (replicate len all0)
        in negateVec vecty all0s LM_MO_FSub

    MO_AlignmentCheck _ _ -> panic "-falignment-sanitisation is not supported by -fllvm"

    -- Handle unsupported cases explicitly so we get a warning
    -- of missing case when new MachOps added
    MO_Add _          -> panicOp
    MO_Mul _          -> panicOp
    MO_Sub _          -> panicOp
    MO_S_MulMayOflo _ -> panicOp
    MO_S_Quot _       -> panicOp
    MO_S_Rem _        -> panicOp
    MO_U_MulMayOflo _ -> panicOp
    MO_U_Quot _       -> panicOp
    MO_U_Rem _        -> panicOp

    MO_Eq  _          -> panicOp
    MO_Ne  _          -> panicOp
    MO_S_Ge _         -> panicOp
    MO_S_Gt _         -> panicOp
    MO_S_Le _         -> panicOp
    MO_S_Lt _         -> panicOp
    MO_U_Ge _         -> panicOp
    MO_U_Gt _         -> panicOp
    MO_U_Le _         -> panicOp
    MO_U_Lt _         -> panicOp

    MO_F_Add        _ -> panicOp
    MO_F_Sub        _ -> panicOp
    MO_F_Mul        _ -> panicOp
    MO_F_Quot       _ -> panicOp
    MO_F_Eq         _ -> panicOp
    MO_F_Ne         _ -> panicOp
    MO_F_Ge         _ -> panicOp
    MO_F_Gt         _ -> panicOp
    MO_F_Le         _ -> panicOp
    MO_F_Lt         _ -> panicOp

    MO_And          _ -> panicOp
    MO_Or           _ -> panicOp
    MO_Xor          _ -> panicOp
    MO_Shl          _ -> panicOp
    MO_U_Shr        _ -> panicOp
    MO_S_Shr        _ -> panicOp

    MO_V_Insert   _ _ -> panicOp
    MO_V_Extract  _ _ -> panicOp

    MO_V_Add      _ _ -> panicOp
    MO_V_Sub      _ _ -> panicOp
    MO_V_Mul      _ _ -> panicOp

    MO_VS_Quot    _ _ -> panicOp
    MO_VS_Rem     _ _ -> panicOp

    MO_VU_Quot    _ _ -> panicOp
    MO_VU_Rem     _ _ -> panicOp

    MO_VF_Insert  _ _ -> panicOp
    MO_VF_Extract _ _ -> panicOp

    MO_VF_Add     _ _ -> panicOp
    MO_VF_Sub     _ _ -> panicOp
    MO_VF_Mul     _ _ -> panicOp
    MO_VF_Quot    _ _ -> panicOp

    where
        negate ty v2 negOp = do
            (vx, stmts, top) <- exprToVar x
            (v1, s1) <- doExpr ty $ LlvmOp negOp v2 vx
            return (v1, stmts `snocOL` s1, top)

        negateVec ty v2 negOp = do
            (vx, stmts1, top) <- exprToVar x
            (vxs', stmts2) <- castVars Signed [(vx, ty)]
            let vx' = singletonPanic "genMachOp: negateVec" vxs'
            (v1, s1) <- doExpr ty $ LlvmOp negOp v2 vx'
            return (v1, stmts1 `appOL` stmts2 `snocOL` s1, top)

        fiConv ty convOp = do
            (vx, stmts, top) <- exprToVar x
            (v1, s1) <- doExpr ty $ Cast convOp vx ty
            return (v1, stmts `snocOL` s1, top)

        sameConv from ty reduce expand = do
            x'@(vx, stmts, top) <- exprToVar x
            let sameConv' op = do
                    (v1, s1) <- doExpr ty $ Cast op vx ty
                    return (v1, stmts `snocOL` s1, top)
            platform <- getPlatform
            let toWidth = llvmWidthInBits platform ty
            -- LLVM doesn't like trying to convert to same width, so
            -- need to check for that as we do get Cmm code doing it.
            case widthInBits from  of
                 w | w < toWidth -> sameConv' expand
                 w | w > toWidth -> sameConv' reduce
                 _w              -> return x'

        panicOp = panic $ "LLVM.CodeGen.genMachOp: non unary op encountered"
                       ++ "with one argument! (" ++ show op ++ ")"

-- Handle GlobalRegs pointers
genMachOp opt o@(MO_Add _) e@[(CmmReg (CmmGlobal r)), (CmmLit (CmmInt n _))]
    = genMachOp_fast opt o r (fromInteger n) e

genMachOp opt o@(MO_Sub _) e@[(CmmReg (CmmGlobal r)), (CmmLit (CmmInt n _))]
    = genMachOp_fast opt o r (negate . fromInteger $ n) e

-- Generic case
genMachOp opt op e = genMachOp_slow opt op e


-- | Handle CmmMachOp expressions
-- This is a specialised method that handles Global register manipulations like
-- 'Sp - 16', using the getelementptr instruction.
genMachOp_fast :: EOption -> MachOp -> GlobalReg -> Int -> [CmmExpr]
               -> LlvmM ExprData
genMachOp_fast opt op r n e
  = do (gv, grt, s1) <- getCmmRegVal (CmmGlobal r)
       platform <- getPlatform
       let (ix,rem) = n `divMod` ((llvmWidthInBits platform . pLower) grt  `div` 8)
       case isPointer grt && rem == 0 of
            True -> do
                (ptr, s2) <- doExpr grt $ GetElemPtr True gv [toI32 ix]
                (var, s3) <- doExpr (llvmWord platform) $ Cast LM_Ptrtoint ptr (llvmWord platform)
                return (var, s1 `snocOL` s2 `snocOL` s3, [])

            False -> genMachOp_slow opt op e


-- | Handle CmmMachOp expressions
-- This handles all the cases not handle by the specialised genMachOp_fast.
genMachOp_slow :: EOption -> MachOp -> [CmmExpr] -> LlvmM ExprData

-- Element extraction
genMachOp_slow _ (MO_V_Extract l w) [val, idx] = runExprData $ do
    vval <- exprToVarW val
    vidx <- exprToVarW idx
    vval' <- singletonPanic "genMachOp_slow" <$>
             castVarsW Signed [(vval, LMVector l ty)]
    doExprW ty $ Extract vval' vidx
  where
    ty = widthToLlvmInt w

genMachOp_slow _ (MO_VF_Extract l w) [val, idx] = runExprData $ do
    vval <- exprToVarW val
    vidx <- exprToVarW idx
    vval' <- singletonPanic "genMachOp_slow" <$>
             castVarsW Signed [(vval, LMVector l ty)]
    doExprW ty $ Extract vval' vidx
  where
    ty = widthToLlvmFloat w

-- Element insertion
genMachOp_slow _ (MO_V_Insert l w) [val, elt, idx] = runExprData $ do
    vval <- exprToVarW val
    velt <- exprToVarW elt
    vidx <- exprToVarW idx
    vval' <- singletonPanic "genMachOp_slow" <$>
             castVarsW Signed [(vval, ty)]
    doExprW ty $ Insert vval' velt vidx
  where
    ty = LMVector l (widthToLlvmInt w)

genMachOp_slow _ (MO_VF_Insert l w) [val, elt, idx] = runExprData $ do
    vval <- exprToVarW val
    velt <- exprToVarW elt
    vidx <- exprToVarW idx
    vval' <- singletonPanic "genMachOp_slow" <$>
             castVarsW Signed [(vval, ty)]
    doExprW ty $ Insert vval' velt vidx
  where
    ty = LMVector l (widthToLlvmFloat w)

-- Binary MachOp
genMachOp_slow opt op [x, y] = case op of

    MO_Eq _   -> genBinComp opt LM_CMP_Eq
    MO_Ne _   -> genBinComp opt LM_CMP_Ne

    MO_S_Gt _ -> genBinComp opt LM_CMP_Sgt
    MO_S_Ge _ -> genBinComp opt LM_CMP_Sge
    MO_S_Lt _ -> genBinComp opt LM_CMP_Slt
    MO_S_Le _ -> genBinComp opt LM_CMP_Sle

    MO_U_Gt _ -> genBinComp opt LM_CMP_Ugt
    MO_U_Ge _ -> genBinComp opt LM_CMP_Uge
    MO_U_Lt _ -> genBinComp opt LM_CMP_Ult
    MO_U_Le _ -> genBinComp opt LM_CMP_Ule

    MO_Add _ -> genBinMach LM_MO_Add
    MO_Sub _ -> genBinMach LM_MO_Sub
    MO_Mul _ -> genBinMach LM_MO_Mul

    MO_U_MulMayOflo _ -> panic "genMachOp: MO_U_MulMayOflo unsupported!"

    MO_S_MulMayOflo w -> isSMulOK w x y

    MO_S_Quot _ -> genBinMach LM_MO_SDiv
    MO_S_Rem  _ -> genBinMach LM_MO_SRem

    MO_U_Quot _ -> genBinMach LM_MO_UDiv
    MO_U_Rem  _ -> genBinMach LM_MO_URem

    MO_F_Eq _ -> genBinComp opt LM_CMP_Feq
    MO_F_Ne _ -> genBinComp opt LM_CMP_Fne
    MO_F_Gt _ -> genBinComp opt LM_CMP_Fgt
    MO_F_Ge _ -> genBinComp opt LM_CMP_Fge
    MO_F_Lt _ -> genBinComp opt LM_CMP_Flt
    MO_F_Le _ -> genBinComp opt LM_CMP_Fle

    MO_F_Add  _ -> genBinMach LM_MO_FAdd
    MO_F_Sub  _ -> genBinMach LM_MO_FSub
    MO_F_Mul  _ -> genBinMach LM_MO_FMul
    MO_F_Quot _ -> genBinMach LM_MO_FDiv

    MO_And _   -> genBinMach LM_MO_And
    MO_Or  _   -> genBinMach LM_MO_Or
    MO_Xor _   -> genBinMach LM_MO_Xor
    MO_Shl _   -> genBinCastYMach LM_MO_Shl
    MO_U_Shr _ -> genBinCastYMach LM_MO_LShr
    MO_S_Shr _ -> genBinCastYMach LM_MO_AShr

    MO_V_Add l w   -> genCastBinMach (LMVector l (widthToLlvmInt w)) LM_MO_Add
    MO_V_Sub l w   -> genCastBinMach (LMVector l (widthToLlvmInt w)) LM_MO_Sub
    MO_V_Mul l w   -> genCastBinMach (LMVector l (widthToLlvmInt w)) LM_MO_Mul

    MO_VS_Quot l w -> genCastBinMach (LMVector l (widthToLlvmInt w)) LM_MO_SDiv
    MO_VS_Rem  l w -> genCastBinMach (LMVector l (widthToLlvmInt w)) LM_MO_SRem

    MO_VU_Quot l w -> genCastBinMach (LMVector l (widthToLlvmInt w)) LM_MO_UDiv
    MO_VU_Rem  l w -> genCastBinMach (LMVector l (widthToLlvmInt w)) LM_MO_URem

    MO_VF_Add  l w -> genCastBinMach (LMVector l (widthToLlvmFloat w)) LM_MO_FAdd
    MO_VF_Sub  l w -> genCastBinMach (LMVector l (widthToLlvmFloat w)) LM_MO_FSub
    MO_VF_Mul  l w -> genCastBinMach (LMVector l (widthToLlvmFloat w)) LM_MO_FMul
    MO_VF_Quot l w -> genCastBinMach (LMVector l (widthToLlvmFloat w)) LM_MO_FDiv

    MO_Not _       -> panicOp
    MO_S_Neg _     -> panicOp
    MO_F_Neg _     -> panicOp

    MO_SF_Conv _ _ -> panicOp
    MO_FS_Conv _ _ -> panicOp
    MO_SS_Conv _ _ -> panicOp
    MO_UU_Conv _ _ -> panicOp
    MO_XX_Conv _ _ -> panicOp
    MO_FF_Conv _ _ -> panicOp

    MO_V_Insert  {} -> panicOp

    MO_VS_Neg {} -> panicOp

    MO_VF_Insert  {} -> panicOp

    MO_VF_Neg {} -> panicOp

    MO_AlignmentCheck {} -> panicOp

#if __GLASGOW_HASKELL__ < 811
    MO_VF_Extract {} -> panicOp
    MO_V_Extract {} -> panicOp
#endif

    where
        binLlvmOp ty binOp allow_y_cast = do
          platform <- getPlatform
          runExprData $ do
            vx <- exprToVarW x
            vy <- exprToVarW y

            if | getVarType vx == getVarType vy
               -> doExprW (ty vx) $ binOp vx vy

               | allow_y_cast
               -> do
                    vy' <- singletonPanic "binLlvmOp cast"<$>
                            castVarsW Signed [(vy, (ty vx))]
                    doExprW (ty vx) $ binOp vx vy'

               | otherwise
               -> do
                    -- Error. Continue anyway so we can debug the generated ll file.
                    dflags <- getDynFlags
                    let style = PprCode CStyle
                        toString doc = renderWithContext (initSDocContext dflags style) doc
                        cmmToStr = (lines . toString . PprCmm.pprExpr platform)
                    statement $ Comment $ map fsLit $ cmmToStr x
                    statement $ Comment $ map fsLit $ cmmToStr y
                    doExprW (ty vx) $ binOp vx vy

        binCastLlvmOp ty binOp = runExprData $ do
            vx <- exprToVarW x
            vy <- exprToVarW y
            vxy' <- castVarsW Signed [(vx, ty), (vy, ty)]
            case vxy' of
              [vx',vy'] -> doExprW ty $ binOp vx' vy'
              _         -> panic "genMachOp_slow: binCastLlvmOp"

        -- | Need to use EOption here as Cmm expects word size results from
        -- comparisons while LLVM return i1. Need to extend to llvmWord type
        -- if expected. See Note [Literals and branch conditions].
        genBinComp opt cmp = do
            ed@(v1, stmts, top) <- binLlvmOp (\_ -> i1) (Compare cmp) False
            dflags <- getDynFlags
            platform <- getPlatform
            if getVarType v1 == i1
                then case i1Expected opt of
                    True  -> return ed
                    False -> do
                        let w_ = llvmWord platform
                        (v2, s1) <- doExpr w_ $ Cast LM_Zext v1 w_
                        return (v2, stmts `snocOL` s1, top)
                else
                    panic $ "genBinComp: Compare returned type other then i1! "
                        ++ (showSDoc dflags $ ppr $ getVarType v1)

        genBinMach op = binLlvmOp getVarType (LlvmOp op) False

        genBinCastYMach op = binLlvmOp getVarType (LlvmOp op) True

        genCastBinMach ty op = binCastLlvmOp ty (LlvmOp op)

        -- | Detect if overflow will occur in signed multiply of the two
        -- CmmExpr's. This is the LLVM assembly equivalent of the NCG
        -- implementation. Its much longer due to type information/safety.
        -- This should actually compile to only about 3 asm instructions.
        isSMulOK :: Width -> CmmExpr -> CmmExpr -> LlvmM ExprData
        isSMulOK _ x y = do
          platform <- getPlatform
          dflags <- getDynFlags
          runExprData $ do
            vx <- exprToVarW x
            vy <- exprToVarW y

            let word  = getVarType vx
            let word2 = LMInt $ 2 * (llvmWidthInBits platform $ getVarType vx)
            let shift = llvmWidthInBits platform word
            let shift1 = toIWord platform (shift - 1)
            let shift2 = toIWord platform shift

            if isInt word
                then do
                    x1     <- doExprW word2 $ Cast LM_Sext vx word2
                    y1     <- doExprW word2 $ Cast LM_Sext vy word2
                    r1     <- doExprW word2 $ LlvmOp LM_MO_Mul x1 y1
                    rlow1  <- doExprW word $ Cast LM_Trunc r1 word
                    rlow2  <- doExprW word $ LlvmOp LM_MO_AShr rlow1 shift1
                    rhigh1 <- doExprW word2 $ LlvmOp LM_MO_AShr r1 shift2
                    rhigh2 <- doExprW word $ Cast LM_Trunc rhigh1 word
                    doExprW word $ LlvmOp LM_MO_Sub rlow2 rhigh2

                else
                    panic $ "isSMulOK: Not bit type! (" ++ showSDoc dflags (ppr word) ++ ")"

        panicOp = panic $ "LLVM.CodeGen.genMachOp_slow: unary op encountered"
                       ++ "with two arguments! (" ++ show op ++ ")"

-- More than two expression, invalid!
genMachOp_slow _ _ _ = panic "genMachOp: More than 2 expressions in MachOp!"


-- | Handle CmmLoad expression.
genLoad :: Atomic -> CmmExpr -> CmmType -> LlvmM ExprData

-- First we try to detect a few common cases and produce better code for
-- these then the default case. We are mostly trying to detect Cmm code
-- like I32[Sp + n] and use 'getelementptr' operations instead of the
-- generic case that uses casts and pointer arithmetic
genLoad atomic e@(CmmReg (CmmGlobal r)) ty
    = genLoad_fast atomic e r 0 ty

genLoad atomic e@(CmmRegOff (CmmGlobal r) n) ty
    = genLoad_fast atomic e r n ty

genLoad atomic e@(CmmMachOp (MO_Add _) [
                            (CmmReg (CmmGlobal r)),
                            (CmmLit (CmmInt n _))])
                ty
    = genLoad_fast atomic e r (fromInteger n) ty

genLoad atomic e@(CmmMachOp (MO_Sub _) [
                            (CmmReg (CmmGlobal r)),
                            (CmmLit (CmmInt n _))])
                ty
    = genLoad_fast atomic e r (negate $ fromInteger n) ty

-- generic case
genLoad atomic e ty
    = getTBAAMeta topN >>= genLoad_slow atomic e ty

-- | Handle CmmLoad expression.
-- This is a special case for loading from a global register pointer
-- offset such as I32[Sp+8].
genLoad_fast :: Atomic -> CmmExpr -> GlobalReg -> Int -> CmmType
             -> LlvmM ExprData
genLoad_fast atomic e r n ty = do
    platform <- getPlatform
    (gv, grt, s1) <- getCmmRegVal (CmmGlobal r)
    meta          <- getTBAARegMeta r
    let ty'      = cmmToLlvmType ty
        (ix,rem) = n `divMod` ((llvmWidthInBits platform . pLower) grt  `div` 8)
    case isPointer grt && rem == 0 of
            True  -> do
                (ptr, s2) <- doExpr grt $ GetElemPtr True gv [toI32 ix]
                -- We might need a different pointer type, so check
                case grt == ty' of
                     -- were fine
                     True -> do
                         (var, s3) <- doExpr ty' (MExpr meta $ loadInstr ptr)
                         return (var, s1 `snocOL` s2 `snocOL` s3,
                                     [])

                     -- cast to pointer type needed
                     False -> do
                         let pty = pLift ty'
                         (ptr', s3) <- doExpr pty $ Cast LM_Bitcast ptr pty
                         (var, s4) <- doExpr ty' (MExpr meta $ loadInstr ptr')
                         return (var, s1 `snocOL` s2 `snocOL` s3
                                    `snocOL` s4, [])

            -- If its a bit type then we use the slow method since
            -- we can't avoid casting anyway.
            False -> genLoad_slow atomic  e ty meta
  where
    loadInstr ptr | atomic    = ALoad SyncSeqCst False ptr
                  | otherwise = Load ptr

-- | Handle Cmm load expression.
-- Generic case. Uses casts and pointer arithmetic if needed.
genLoad_slow :: Atomic -> CmmExpr -> CmmType -> [MetaAnnot] -> LlvmM ExprData
genLoad_slow atomic e ty meta = do
  platform <- getPlatform
  dflags <- getDynFlags
  opts <- getLlvmOpts
  runExprData $ do
    iptr <- exprToVarW e
    case getVarType iptr of
        LMPointer _ ->
                    doExprW (cmmToLlvmType ty) (MExpr meta $ loadInstr iptr)

        i@(LMInt _) | i == llvmWord platform -> do
                    let pty = LMPointer $ cmmToLlvmType ty
                    ptr <- doExprW pty $ Cast LM_Inttoptr iptr pty
                    doExprW (cmmToLlvmType ty) (MExpr meta $ loadInstr ptr)

        other -> pprPanic "exprToVar: CmmLoad expression is not right type!"
                     (PprCmm.pprExpr platform e <+> text (
                         "Size of Ptr: " ++ show (llvmPtrBits platform) ++
                         ", Size of var: " ++ show (llvmWidthInBits platform other) ++
                         ", Var: " ++ showSDoc dflags (ppVar opts iptr)))
  where
    loadInstr ptr | atomic    = ALoad SyncSeqCst False ptr
                  | otherwise = Load ptr


-- | Handle CmmReg expression. This will return a pointer to the stack
-- location of the register. Throws an error if it isn't allocated on
-- the stack.
getCmmReg :: CmmReg -> LlvmM LlvmVar
getCmmReg (CmmLocal (LocalReg un _))
  = do exists <- varLookup un
       dflags <- getDynFlags
       case exists of
         Just ety -> return (LMLocalVar un $ pLift ety)
         Nothing  -> panic $ "getCmmReg: Cmm register " ++ showSDoc dflags (ppr un) ++ " was not allocated!"
           -- This should never happen, as every local variable should
           -- have been assigned a value at some point, triggering
           -- "funPrologue" to allocate it on the stack.

getCmmReg (CmmGlobal g)
  = do onStack <- checkStackReg g
       dflags <- getDynFlags
       platform <- getPlatform
       if onStack
         then return (lmGlobalRegVar platform g)
         else panic $ "getCmmReg: Cmm register " ++ showSDoc dflags (ppr g) ++ " not stack-allocated!"

-- | Return the value of a given register, as well as its type. Might
-- need to be load from stack.
getCmmRegVal :: CmmReg -> LlvmM (LlvmVar, LlvmType, LlvmStatements)
getCmmRegVal reg =
  case reg of
    CmmGlobal g -> do
      onStack <- checkStackReg g
      platform <- getPlatform
      if onStack then loadFromStack else do
        let r = lmGlobalRegArg platform g
        return (r, getVarType r, nilOL)
    _ -> loadFromStack
 where loadFromStack = do
         ptr <- getCmmReg reg
         let ty = pLower $ getVarType ptr
         (v, s) <- doExpr ty (Load ptr)
         return (v, ty, unitOL s)

-- | Allocate a local CmmReg on the stack
allocReg :: CmmReg -> (LlvmVar, LlvmStatements)
allocReg (CmmLocal (LocalReg un ty))
  = let ty' = cmmToLlvmType ty
        var = LMLocalVar un (LMPointer ty')
        alc = Alloca ty' 1
    in (var, unitOL $ Assignment var alc)

allocReg _ = panic $ "allocReg: Global reg encountered! Global registers should"
                    ++ " have been handled elsewhere!"


-- | Generate code for a literal
genLit :: EOption -> CmmLit -> LlvmM ExprData
genLit opt (CmmInt i w)
  -- See Note [Literals and branch conditions].
  = let width | i1Expected opt = i1
              | otherwise      = LMInt (widthInBits w)
        -- comm  = Comment [ fsLit $ "EOption: " ++ show opt
        --                 , fsLit $ "Width  : " ++ show w
        --                 , fsLit $ "Width' : " ++ show (widthInBits w)
        --                 ]
    in return (mkIntLit width i, nilOL, [])

genLit _ (CmmFloat r w)
  = return (LMLitVar $ LMFloatLit (fromRational r) (widthToLlvmFloat w),
              nilOL, [])

genLit opt (CmmVec ls)
  = do llvmLits <- mapM toLlvmLit ls
       return (LMLitVar $ LMVectorLit llvmLits, nilOL, [])
  where
    toLlvmLit :: CmmLit -> LlvmM LlvmLit
    toLlvmLit lit = do
        (llvmLitVar, _, _) <- genLit opt lit
        case llvmLitVar of
          LMLitVar llvmLit -> return llvmLit
          _ -> panic "genLit"

genLit _ cmm@(CmmLabel l)
  = do var <- getGlobalPtr =<< strCLabel_llvm l
       platform <- getPlatform
       let lmty = cmmToLlvmType $ cmmLitType platform cmm
       (v1, s1) <- doExpr lmty $ Cast LM_Ptrtoint var (llvmWord platform)
       return (v1, unitOL s1, [])

genLit opt (CmmLabelOff label off) = do
    platform <- getPlatform
    (vlbl, stmts, stat) <- genLit opt (CmmLabel label)
    let voff = toIWord platform off
    (v1, s1) <- doExpr (getVarType vlbl) $ LlvmOp LM_MO_Add vlbl voff
    return (v1, stmts `snocOL` s1, stat)

genLit opt (CmmLabelDiffOff l1 l2 off w) = do
    platform <- getPlatform
    (vl1, stmts1, stat1) <- genLit opt (CmmLabel l1)
    (vl2, stmts2, stat2) <- genLit opt (CmmLabel l2)
    let voff = toIWord platform off
    let ty1 = getVarType vl1
    let ty2 = getVarType vl2
    if (isInt ty1) && (isInt ty2)
       && (llvmWidthInBits platform ty1 == llvmWidthInBits platform ty2)
       then do
            (v1, s1) <- doExpr (getVarType vl1) $ LlvmOp LM_MO_Sub vl1 vl2
            (v2, s2) <- doExpr (getVarType v1 ) $ LlvmOp LM_MO_Add v1 voff
            let ty = widthToLlvmInt w
            let stmts = stmts1 `appOL` stmts2 `snocOL` s1 `snocOL` s2
            if w /= wordWidth platform
              then do
                (v3, s3) <- doExpr ty $ Cast LM_Trunc v2 ty
                return (v3, stmts `snocOL` s3, stat1 ++ stat2)
              else
                return (v2, stmts, stat1 ++ stat2)
        else
            panic "genLit: CmmLabelDiffOff encountered with different label ty!"

genLit opt (CmmBlock b)
  = genLit opt (CmmLabel $ infoTblLbl b)

genLit _ CmmHighStackMark
  = panic "genStaticLit - CmmHighStackMark unsupported!"


-- -----------------------------------------------------------------------------
-- * Misc
--

-- | Find CmmRegs that get assigned and allocate them on the stack
--
-- Any register that gets written needs to be allocated on the
-- stack. This avoids having to map a CmmReg to an equivalent SSA form
-- and avoids having to deal with Phi node insertion.  This is also
-- the approach recommended by LLVM developers.
--
-- On the other hand, this is unnecessarily verbose if the register in
-- question is never written. Therefore we skip it where we can to
-- save a few lines in the output and hopefully speed compilation up a
-- bit.
funPrologue :: LiveGlobalRegs -> [CmmBlock] -> LlvmM StmtData
funPrologue live cmmBlocks = do

  let getAssignedRegs :: CmmNode O O -> [CmmReg]
      getAssignedRegs (CmmAssign reg _)  = [reg]
      getAssignedRegs (CmmUnsafeForeignCall _ rs _) = map CmmLocal rs
      getAssignedRegs _                  = []
      getRegsBlock (_, body, _)          = concatMap getAssignedRegs $ blockToList body
      assignedRegs = nub $ concatMap (getRegsBlock . blockSplit) cmmBlocks
      isLive r     = r `elem` alwaysLive || r `elem` live

  platform <- getPlatform
  stmtss <- forM assignedRegs $ \reg ->
    case reg of
      CmmLocal (LocalReg un _) -> do
        let (newv, stmts) = allocReg reg
        varInsert un (pLower $ getVarType newv)
        return stmts
      CmmGlobal r -> do
        let reg   = lmGlobalRegVar platform r
            arg   = lmGlobalRegArg platform r
            ty    = (pLower . getVarType) reg
            trash = LMLitVar $ LMUndefLit ty
            rval  = if isLive r then arg else trash
            alloc = Assignment reg $ Alloca (pLower $ getVarType reg) 1
        markStackReg r
        return $ toOL [alloc, Store rval reg]

  return (concatOL stmtss `snocOL` jumpToEntry, [])
  where
    entryBlk : _ = cmmBlocks
    jumpToEntry = Branch $ blockIdToLlvm (entryLabel entryBlk)

-- | Function epilogue. Load STG variables to use as argument for call.
-- STG Liveness optimisation done here.
funEpilogue :: LiveGlobalRegs -> LlvmM ([LlvmVar], LlvmStatements)
funEpilogue live = do
    platform <- getPlatform

    let paddingRegs = padLiveArgs platform live

    -- Set to value or "undef" depending on whether the register is
    -- actually live
    let loadExpr r = do
          (v, _, s) <- getCmmRegVal (CmmGlobal r)
          return (Just $ v, s)
        loadUndef r = do
          let ty = (pLower . getVarType $ lmGlobalRegVar platform r)
          return (Just $ LMLitVar $ LMUndefLit ty, nilOL)

    -- Note that floating-point registers in `activeStgRegs` must be sorted
    -- according to the calling convention.
    --  E.g. for X86:
    --     GOOD: F1,D1,XMM1,F2,D2,XMM2,...
    --     BAD : F1,F2,F3,D1,D2,D3,XMM1,XMM2,XMM3,...
    --  As Fn, Dn and XMMn use the same register (XMMn) to be passed, we don't
    --  want to pass F2 before D1 for example, otherwise we could get F2 -> XMM1
    --  and D1 -> XMM2.
    let allRegs = activeStgRegs platform
    loads <- forM allRegs $ \r -> if
      -- load live registers
      | r `elem` alwaysLive  -> loadExpr r
      | r `elem` live        -> loadExpr r
      -- load all non Floating-Point Registers
      | not (isFPR r)        -> loadUndef r
      -- load padding Floating-Point Registers
      | r `elem` paddingRegs -> loadUndef r
      | otherwise            -> return (Nothing, nilOL)

    let (vars, stmts) = unzip loads
    return (catMaybes vars, concatOL stmts)

-- | Get a function pointer to the CLabel specified.
--
-- This is for Haskell functions, function type is assumed, so doesn't work
-- with foreign functions.
getHsFunc :: LiveGlobalRegs -> CLabel -> LlvmM ExprData
getHsFunc live lbl
  = do fty <- llvmFunTy live
       name <- strCLabel_llvm lbl
       getHsFunc' name fty

getHsFunc' :: LMString -> LlvmType -> LlvmM ExprData
getHsFunc' name fty
  = do fun <- getGlobalPtr name
       if getVarType fun == fty
         then return (fun, nilOL, [])
         else do (v1, s1) <- doExpr (pLift fty)
                               $ Cast LM_Bitcast fun (pLift fty)
                 return  (v1, unitOL s1, [])

-- | Create a new local var
mkLocalVar :: LlvmType -> LlvmM LlvmVar
mkLocalVar ty = do
    un <- getUniqueM
    return $ LMLocalVar un ty


-- | Execute an expression, assigning result to a var
doExpr :: LlvmType -> LlvmExpression -> LlvmM (LlvmVar, LlvmStatement)
doExpr ty expr = do
    v <- mkLocalVar ty
    return (v, Assignment v expr)


-- | Expand CmmRegOff
expandCmmReg :: Platform -> (CmmReg, Int) -> CmmExpr
expandCmmReg platform (reg, off)
  = let width = typeWidth (cmmRegType platform reg)
        voff  = CmmLit $ CmmInt (fromIntegral off) width
    in CmmMachOp (MO_Add width) [CmmReg reg, voff]


-- | Convert a block id into a appropriate Llvm label
blockIdToLlvm :: BlockId -> LlvmVar
blockIdToLlvm bid = LMLocalVar (getUnique bid) LMLabel

-- | Create Llvm int Literal
mkIntLit :: Integral a => LlvmType -> a -> LlvmVar
mkIntLit ty i = LMLitVar $ LMIntLit (toInteger i) ty

-- | Convert int type to a LLvmVar of word or i32 size
toI32 :: Integral a => a -> LlvmVar
toI32 = mkIntLit i32

toIWord :: Integral a => Platform -> a -> LlvmVar
toIWord platform = mkIntLit (llvmWord platform)


-- | Error functions
panic :: HasCallStack => String -> a
panic s = Panic.panic $ "GHC.CmmToLlvm.CodeGen." ++ s

pprPanic :: HasCallStack => String -> SDoc -> a
pprPanic s d = Panic.pprPanic ("GHC.CmmToLlvm.CodeGen." ++ s) d


-- | Returns TBAA meta data by unique
getTBAAMeta :: Unique -> LlvmM [MetaAnnot]
getTBAAMeta u = do
    mi <- getUniqMeta u
    return [MetaAnnot tbaa (MetaNode i) | let Just i = mi]

-- | Returns TBAA meta data for given register
getTBAARegMeta :: GlobalReg -> LlvmM [MetaAnnot]
getTBAARegMeta = getTBAAMeta . getTBAA


-- | A more convenient way of accumulating LLVM statements and declarations.
data LlvmAccum = LlvmAccum LlvmStatements [LlvmCmmDecl]

instance Semigroup LlvmAccum where
  LlvmAccum stmtsA declsA <> LlvmAccum stmtsB declsB =
        LlvmAccum (stmtsA Semigroup.<> stmtsB) (declsA Semigroup.<> declsB)

instance Monoid LlvmAccum where
    mempty = LlvmAccum nilOL []
    mappend = (Semigroup.<>)

liftExprData :: LlvmM ExprData -> WriterT LlvmAccum LlvmM LlvmVar
liftExprData action = do
    (var, stmts, decls) <- lift action
    tell $ LlvmAccum stmts decls
    return var

statement :: LlvmStatement -> WriterT LlvmAccum LlvmM ()
statement stmt = tell $ LlvmAccum (unitOL stmt) []

doExprW :: LlvmType -> LlvmExpression -> WriterT LlvmAccum LlvmM LlvmVar
doExprW a b = do
    (var, stmt) <- lift $ doExpr a b
    statement stmt
    return var

exprToVarW :: CmmExpr -> WriterT LlvmAccum LlvmM LlvmVar
exprToVarW = liftExprData . exprToVar

runExprData :: WriterT LlvmAccum LlvmM LlvmVar -> LlvmM ExprData
runExprData action = do
    (var, LlvmAccum stmts decls) <- runWriterT action
    return (var, stmts, decls)

runStmtsDecls :: WriterT LlvmAccum LlvmM () -> LlvmM (LlvmStatements, [LlvmCmmDecl])
runStmtsDecls action = do
    LlvmAccum stmts decls <- execWriterT action
    return (stmts, decls)

getCmmRegW :: CmmReg -> WriterT LlvmAccum LlvmM LlvmVar
getCmmRegW = lift . getCmmReg

genLoadW :: Atomic -> CmmExpr -> CmmType -> WriterT LlvmAccum LlvmM LlvmVar
genLoadW atomic e ty = liftExprData $ genLoad atomic e ty

-- | Return element of single-element list; 'panic' if list is not a single-element list
singletonPanic :: String -> [a] -> a
singletonPanic _ [x] = x
singletonPanic s _ = panic s