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
|
{-
(c) The GRASP/AQUA Project, Glasgow University, 1993-1998
\section[Specialise]{Stamping out overloading, and (optionally) polymorphism}
-}
{-# LANGUAGE CPP #-}
{-# LANGUAGE DeriveFunctor #-}
{-# LANGUAGE ViewPatterns #-}
module Specialise ( specProgram, specUnfolding ) where
#include "HsVersions.h"
import GhcPrelude
import Id
import TcType hiding( substTy )
import Type hiding( substTy, extendTvSubstList )
import Predicate
import Module( Module, HasModule(..) )
import Coercion( Coercion )
import CoreMonad
import qualified CoreSubst
import CoreUnfold
import Var ( isLocalVar )
import VarSet
import VarEnv
import CoreSyn
import Rules
import CoreOpt ( collectBindersPushingCo )
import CoreUtils ( exprIsTrivial, mkCast, exprType )
import CoreFVs
import CoreArity ( etaExpandToJoinPointRule )
import UniqSupply
import Name
import MkId ( voidArgId, voidPrimId )
import Maybes ( mapMaybe, isJust )
import MonadUtils ( foldlM )
import BasicTypes
import HscTypes
import Bag
import DynFlags
import Util
import Outputable
import FastString
import State
import UniqDFM
import TyCoRep (TyCoBinder (..))
import Control.Monad
import qualified Control.Monad.Fail as MonadFail
{-
************************************************************************
* *
\subsection[notes-Specialise]{Implementation notes [SLPJ, Aug 18 1993]}
* *
************************************************************************
These notes describe how we implement specialisation to eliminate
overloading.
The specialisation pass works on Core
syntax, complete with all the explicit dictionary application,
abstraction and construction as added by the type checker. The
existing type checker remains largely as it is.
One important thought: the {\em types} passed to an overloaded
function, and the {\em dictionaries} passed are mutually redundant.
If the same function is applied to the same type(s) then it is sure to
be applied to the same dictionary(s)---or rather to the same {\em
values}. (The arguments might look different but they will evaluate
to the same value.)
Second important thought: we know that we can make progress by
treating dictionary arguments as static and worth specialising on. So
we can do without binding-time analysis, and instead specialise on
dictionary arguments and no others.
The basic idea
~~~~~~~~~~~~~~
Suppose we have
let f = <f_rhs>
in <body>
and suppose f is overloaded.
STEP 1: CALL-INSTANCE COLLECTION
We traverse <body>, accumulating all applications of f to types and
dictionaries.
(Might there be partial applications, to just some of its types and
dictionaries? In principle yes, but in practice the type checker only
builds applications of f to all its types and dictionaries, so partial
applications could only arise as a result of transformation, and even
then I think it's unlikely. In any case, we simply don't accumulate such
partial applications.)
STEP 2: EQUIVALENCES
So now we have a collection of calls to f:
f t1 t2 d1 d2
f t3 t4 d3 d4
...
Notice that f may take several type arguments. To avoid ambiguity, we
say that f is called at type t1/t2 and t3/t4.
We take equivalence classes using equality of the *types* (ignoring
the dictionary args, which as mentioned previously are redundant).
STEP 3: SPECIALISATION
For each equivalence class, choose a representative (f t1 t2 d1 d2),
and create a local instance of f, defined thus:
f@t1/t2 = <f_rhs> t1 t2 d1 d2
f_rhs presumably has some big lambdas and dictionary lambdas, so lots
of simplification will now result. However we don't actually *do* that
simplification. Rather, we leave it for the simplifier to do. If we
*did* do it, though, we'd get more call instances from the specialised
RHS. We can work out what they are by instantiating the call-instance
set from f's RHS with the types t1, t2.
Add this new id to f's IdInfo, to record that f has a specialised version.
Before doing any of this, check that f's IdInfo doesn't already
tell us about an existing instance of f at the required type/s.
(This might happen if specialisation was applied more than once, or
it might arise from user SPECIALIZE pragmas.)
Recursion
~~~~~~~~~
Wait a minute! What if f is recursive? Then we can't just plug in
its right-hand side, can we?
But it's ok. The type checker *always* creates non-recursive definitions
for overloaded recursive functions. For example:
f x = f (x+x) -- Yes I know its silly
becomes
f a (d::Num a) = let p = +.sel a d
in
letrec fl (y::a) = fl (p y y)
in
fl
We still have recursion for non-overloaded functions which we
specialise, but the recursive call should get specialised to the
same recursive version.
Polymorphism 1
~~~~~~~~~~~~~~
All this is crystal clear when the function is applied to *constant
types*; that is, types which have no type variables inside. But what if
it is applied to non-constant types? Suppose we find a call of f at type
t1/t2. There are two possibilities:
(a) The free type variables of t1, t2 are in scope at the definition point
of f. In this case there's no problem, we proceed just as before. A common
example is as follows. Here's the Haskell:
g y = let f x = x+x
in f y + f y
After typechecking we have
g a (d::Num a) (y::a) = let f b (d'::Num b) (x::b) = +.sel b d' x x
in +.sel a d (f a d y) (f a d y)
Notice that the call to f is at type type "a"; a non-constant type.
Both calls to f are at the same type, so we can specialise to give:
g a (d::Num a) (y::a) = let f@a (x::a) = +.sel a d x x
in +.sel a d (f@a y) (f@a y)
(b) The other case is when the type variables in the instance types
are *not* in scope at the definition point of f. The example we are
working with above is a good case. There are two instances of (+.sel a d),
but "a" is not in scope at the definition of +.sel. Can we do anything?
Yes, we can "common them up", a sort of limited common sub-expression deal.
This would give:
g a (d::Num a) (y::a) = let +.sel@a = +.sel a d
f@a (x::a) = +.sel@a x x
in +.sel@a (f@a y) (f@a y)
This can save work, and can't be spotted by the type checker, because
the two instances of +.sel weren't originally at the same type.
Further notes on (b)
* There are quite a few variations here. For example, the defn of
+.sel could be floated ouside the \y, to attempt to gain laziness.
It certainly mustn't be floated outside the \d because the d has to
be in scope too.
* We don't want to inline f_rhs in this case, because
that will duplicate code. Just commoning up the call is the point.
* Nothing gets added to +.sel's IdInfo.
* Don't bother unless the equivalence class has more than one item!
Not clear whether this is all worth it. It is of course OK to
simply discard call-instances when passing a big lambda.
Polymorphism 2 -- Overloading
~~~~~~~~~~~~~~
Consider a function whose most general type is
f :: forall a b. Ord a => [a] -> b -> b
There is really no point in making a version of g at Int/Int and another
at Int/Bool, because it's only instantiating the type variable "a" which
buys us any efficiency. Since g is completely polymorphic in b there
ain't much point in making separate versions of g for the different
b types.
That suggests that we should identify which of g's type variables
are constrained (like "a") and which are unconstrained (like "b").
Then when taking equivalence classes in STEP 2, we ignore the type args
corresponding to unconstrained type variable. In STEP 3 we make
polymorphic versions. Thus:
f@t1/ = /\b -> <f_rhs> t1 b d1 d2
We do this.
Dictionary floating
~~~~~~~~~~~~~~~~~~~
Consider this
f a (d::Num a) = let g = ...
in
...(let d1::Ord a = Num.Ord.sel a d in g a d1)...
Here, g is only called at one type, but the dictionary isn't in scope at the
definition point for g. Usually the type checker would build a
definition for d1 which enclosed g, but the transformation system
might have moved d1's defn inward. Solution: float dictionary bindings
outwards along with call instances.
Consider
f x = let g p q = p==q
h r s = (r+s, g r s)
in
h x x
Before specialisation, leaving out type abstractions we have
f df x = let g :: Eq a => a -> a -> Bool
g dg p q = == dg p q
h :: Num a => a -> a -> (a, Bool)
h dh r s = let deq = eqFromNum dh
in (+ dh r s, g deq r s)
in
h df x x
After specialising h we get a specialised version of h, like this:
h' r s = let deq = eqFromNum df
in (+ df r s, g deq r s)
But we can't naively make an instance for g from this, because deq is not in scope
at the defn of g. Instead, we have to float out the (new) defn of deq
to widen its scope. Notice that this floating can't be done in advance -- it only
shows up when specialisation is done.
User SPECIALIZE pragmas
~~~~~~~~~~~~~~~~~~~~~~~
Specialisation pragmas can be digested by the type checker, and implemented
by adding extra definitions along with that of f, in the same way as before
f@t1/t2 = <f_rhs> t1 t2 d1 d2
Indeed the pragmas *have* to be dealt with by the type checker, because
only it knows how to build the dictionaries d1 and d2! For example
g :: Ord a => [a] -> [a]
{-# SPECIALIZE f :: [Tree Int] -> [Tree Int] #-}
Here, the specialised version of g is an application of g's rhs to the
Ord dictionary for (Tree Int), which only the type checker can conjure
up. There might not even *be* one, if (Tree Int) is not an instance of
Ord! (All the other specialision has suitable dictionaries to hand
from actual calls.)
Problem. The type checker doesn't have to hand a convenient <f_rhs>, because
it is buried in a complex (as-yet-un-desugared) binding group.
Maybe we should say
f@t1/t2 = f* t1 t2 d1 d2
where f* is the Id f with an IdInfo which says "inline me regardless!".
Indeed all the specialisation could be done in this way.
That in turn means that the simplifier has to be prepared to inline absolutely
any in-scope let-bound thing.
Again, the pragma should permit polymorphism in unconstrained variables:
h :: Ord a => [a] -> b -> b
{-# SPECIALIZE h :: [Int] -> b -> b #-}
We *insist* that all overloaded type variables are specialised to ground types,
(and hence there can be no context inside a SPECIALIZE pragma).
We *permit* unconstrained type variables to be specialised to
- a ground type
- or left as a polymorphic type variable
but nothing in between. So
{-# SPECIALIZE h :: [Int] -> [c] -> [c] #-}
is *illegal*. (It can be handled, but it adds complication, and gains the
programmer nothing.)
SPECIALISING INSTANCE DECLARATIONS
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Consider
instance Foo a => Foo [a] where
...
{-# SPECIALIZE instance Foo [Int] #-}
The original instance decl creates a dictionary-function
definition:
dfun.Foo.List :: forall a. Foo a -> Foo [a]
The SPECIALIZE pragma just makes a specialised copy, just as for
ordinary function definitions:
dfun.Foo.List@Int :: Foo [Int]
dfun.Foo.List@Int = dfun.Foo.List Int dFooInt
The information about what instance of the dfun exist gets added to
the dfun's IdInfo in the same way as a user-defined function too.
Automatic instance decl specialisation?
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Can instance decls be specialised automatically? It's tricky.
We could collect call-instance information for each dfun, but
then when we specialised their bodies we'd get new call-instances
for ordinary functions; and when we specialised their bodies, we might get
new call-instances of the dfuns, and so on. This all arises because of
the unrestricted mutual recursion between instance decls and value decls.
Still, there's no actual problem; it just means that we may not do all
the specialisation we could theoretically do.
Furthermore, instance decls are usually exported and used non-locally,
so we'll want to compile enough to get those specialisations done.
Lastly, there's no such thing as a local instance decl, so we can
survive solely by spitting out *usage* information, and then reading that
back in as a pragma when next compiling the file. So for now,
we only specialise instance decls in response to pragmas.
SPITTING OUT USAGE INFORMATION
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
To spit out usage information we need to traverse the code collecting
call-instance information for all imported (non-prelude?) functions
and data types. Then we equivalence-class it and spit it out.
This is done at the top-level when all the call instances which escape
must be for imported functions and data types.
*** Not currently done ***
Partial specialisation by pragmas
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
What about partial specialisation:
k :: (Ord a, Eq b) => [a] -> b -> b -> [a]
{-# SPECIALIZE k :: Eq b => [Int] -> b -> b -> [a] #-}
or even
{-# SPECIALIZE k :: Eq b => [Int] -> [b] -> [b] -> [a] #-}
Seems quite reasonable. Similar things could be done with instance decls:
instance (Foo a, Foo b) => Foo (a,b) where
...
{-# SPECIALIZE instance Foo a => Foo (a,Int) #-}
{-# SPECIALIZE instance Foo b => Foo (Int,b) #-}
Ho hum. Things are complex enough without this. I pass.
Requirements for the simplifier
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
The simplifier has to be able to take advantage of the specialisation.
* When the simplifier finds an application of a polymorphic f, it looks in
f's IdInfo in case there is a suitable instance to call instead. This converts
f t1 t2 d1 d2 ===> f_t1_t2
Note that the dictionaries get eaten up too!
* Dictionary selection operations on constant dictionaries must be
short-circuited:
+.sel Int d ===> +Int
The obvious way to do this is in the same way as other specialised
calls: +.sel has inside it some IdInfo which tells that if it's applied
to the type Int then it should eat a dictionary and transform to +Int.
In short, dictionary selectors need IdInfo inside them for constant
methods.
* Exactly the same applies if a superclass dictionary is being
extracted:
Eq.sel Int d ===> dEqInt
* Something similar applies to dictionary construction too. Suppose
dfun.Eq.List is the function taking a dictionary for (Eq a) to
one for (Eq [a]). Then we want
dfun.Eq.List Int d ===> dEq.List_Int
Where does the Eq [Int] dictionary come from? It is built in
response to a SPECIALIZE pragma on the Eq [a] instance decl.
In short, dfun Ids need IdInfo with a specialisation for each
constant instance of their instance declaration.
All this uses a single mechanism: the SpecEnv inside an Id
What does the specialisation IdInfo look like?
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
The SpecEnv of an Id maps a list of types (the template) to an expression
[Type] |-> Expr
For example, if f has this RuleInfo:
[Int, a] -> \d:Ord Int. f' a
it means that we can replace the call
f Int t ===> (\d. f' t)
This chucks one dictionary away and proceeds with the
specialised version of f, namely f'.
What can't be done this way?
~~~~~~~~~~~~~~~~~~~~~~~~~~~~
There is no way, post-typechecker, to get a dictionary for (say)
Eq a from a dictionary for Eq [a]. So if we find
==.sel [t] d
we can't transform to
eqList (==.sel t d')
where
eqList :: (a->a->Bool) -> [a] -> [a] -> Bool
Of course, we currently have no way to automatically derive
eqList, nor to connect it to the Eq [a] instance decl, but you
can imagine that it might somehow be possible. Taking advantage
of this is permanently ruled out.
Still, this is no great hardship, because we intend to eliminate
overloading altogether anyway!
A note about non-tyvar dictionaries
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Some Ids have types like
forall a,b,c. Eq a -> Ord [a] -> tau
This seems curious at first, because we usually only have dictionary
args whose types are of the form (C a) where a is a type variable.
But this doesn't hold for the functions arising from instance decls,
which sometimes get arguments with types of form (C (T a)) for some
type constructor T.
Should we specialise wrt this compound-type dictionary? We used to say
"no", saying:
"This is a heuristic judgement, as indeed is the fact that we
specialise wrt only dictionaries. We choose *not* to specialise
wrt compound dictionaries because at the moment the only place
they show up is in instance decls, where they are simply plugged
into a returned dictionary. So nothing is gained by specialising
wrt them."
But it is simpler and more uniform to specialise wrt these dicts too;
and in future GHC is likely to support full fledged type signatures
like
f :: Eq [(a,b)] => ...
************************************************************************
* *
\subsubsection{The new specialiser}
* *
************************************************************************
Our basic game plan is this. For let(rec) bound function
f :: (C a, D c) => (a,b,c,d) -> Bool
* Find any specialised calls of f, (f ts ds), where
ts are the type arguments t1 .. t4, and
ds are the dictionary arguments d1 .. d2.
* Add a new definition for f1 (say):
f1 = /\ b d -> (..body of f..) t1 b t3 d d1 d2
Note that we abstract over the unconstrained type arguments.
* Add the mapping
[t1,b,t3,d] |-> \d1 d2 -> f1 b d
to the specialisations of f. This will be used by the
simplifier to replace calls
(f t1 t2 t3 t4) da db
by
(\d1 d1 -> f1 t2 t4) da db
All the stuff about how many dictionaries to discard, and what types
to apply the specialised function to, are handled by the fact that the
SpecEnv contains a template for the result of the specialisation.
We don't build *partial* specialisations for f. For example:
f :: Eq a => a -> a -> Bool
{-# SPECIALISE f :: (Eq b, Eq c) => (b,c) -> (b,c) -> Bool #-}
Here, little is gained by making a specialised copy of f.
There's a distinct danger that the specialised version would
first build a dictionary for (Eq b, Eq c), and then select the (==)
method from it! Even if it didn't, not a great deal is saved.
We do, however, generate polymorphic, but not overloaded, specialisations:
f :: Eq a => [a] -> b -> b -> b
... SPECIALISE f :: [Int] -> b -> b -> b ...
Hence, the invariant is this:
*** no specialised version is overloaded ***
************************************************************************
* *
\subsubsection{The exported function}
* *
************************************************************************
-}
-- | Specialise calls to type-class overloaded functions occurring in a program.
specProgram :: ModGuts -> CoreM ModGuts
specProgram guts@(ModGuts { mg_module = this_mod
, mg_rules = local_rules
, mg_binds = binds })
= do { dflags <- getDynFlags
-- Specialise the bindings of this module
; (binds', uds) <- runSpecM dflags this_mod (go binds)
-- Specialise imported functions
; hpt_rules <- getRuleBase
; let rule_base = extendRuleBaseList hpt_rules local_rules
; (new_rules, spec_binds) <- specImports dflags this_mod top_env emptyVarSet
[] rule_base uds
; let final_binds
| null spec_binds = binds'
| otherwise = Rec (flattenBinds spec_binds) : binds'
-- Note [Glom the bindings if imported functions are specialised]
; return (guts { mg_binds = final_binds
, mg_rules = new_rules ++ local_rules }) }
where
-- We need to start with a Subst that knows all the things
-- that are in scope, so that the substitution engine doesn't
-- accidentally re-use a unique that's already in use
-- Easiest thing is to do it all at once, as if all the top-level
-- decls were mutually recursive
top_env = SE { se_subst = CoreSubst.mkEmptySubst $ mkInScopeSet $ mkVarSet $
bindersOfBinds binds
, se_interesting = emptyVarSet }
go [] = return ([], emptyUDs)
go (bind:binds) = do (binds', uds) <- go binds
(bind', uds') <- specBind top_env bind uds
return (bind' ++ binds', uds')
{-
Note [Wrap bindings returned by specImports]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
'specImports' returns a set of specialized bindings. However, these are lacking
necessary floated dictionary bindings, which are returned by
UsageDetails(ud_binds). These dictionaries need to be brought into scope with
'wrapDictBinds' before the bindings returned by 'specImports' can be used. See,
for instance, the 'specImports' call in 'specProgram'.
Note [Disabling cross-module specialisation]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Since GHC 7.10 we have performed specialisation of INLINABLE bindings living
in modules outside of the current module. This can sometimes uncover user code
which explodes in size when aggressively optimized. The
-fno-cross-module-specialise option was introduced to allow users to being
bitten by such instances to revert to the pre-7.10 behavior.
See #10491
-}
-- | An argument that we might want to specialise.
-- See Note [Specialising Calls] for the nitty gritty details.
data SpecArg
=
-- | Type arguments that should be specialised, due to appearing
-- free in the type of a 'SpecDict'.
SpecType Type
-- | Type arguments that should remain polymorphic.
| UnspecType
-- | Dictionaries that should be specialised.
| SpecDict DictExpr
-- | Value arguments that should not be specialised.
| UnspecArg
instance Outputable SpecArg where
ppr (SpecType t) = text "SpecType" <+> ppr t
ppr UnspecType = text "UnspecType"
ppr (SpecDict d) = text "SpecDict" <+> ppr d
ppr UnspecArg = text "UnspecArg"
getSpecDicts :: [SpecArg] -> [DictExpr]
getSpecDicts = mapMaybe go
where
go (SpecDict d) = Just d
go _ = Nothing
getSpecTypes :: [SpecArg] -> [Type]
getSpecTypes = mapMaybe go
where
go (SpecType t) = Just t
go _ = Nothing
isUnspecArg :: SpecArg -> Bool
isUnspecArg UnspecArg = True
isUnspecArg UnspecType = True
isUnspecArg _ = False
isValueArg :: SpecArg -> Bool
isValueArg UnspecArg = True
isValueArg (SpecDict _) = True
isValueArg _ = False
-- | Given binders from an original function 'f', and the 'SpecArg's
-- corresponding to its usage, compute everything necessary to build
-- a specialisation.
--
-- We will use a running example. Consider the function
--
-- foo :: forall a b. Eq a => Int -> blah
-- foo @a @b dEqA i = blah
--
-- which is called with the 'CallInfo'
--
-- [SpecType T1, UnspecType, SpecDict dEqT1, UnspecArg]
--
-- We'd eventually like to build the RULE
--
-- RULE "SPEC foo @T1 _"
-- forall @a @b (dEqA' :: Eq a).
-- foo @T1 @b dEqA' = $sfoo @b
--
-- and the specialisation '$sfoo'
--
-- $sfoo :: forall b. Int -> blah
-- $sfoo @b = \i -> SUBST[a->T1, dEqA->dEqA'] blah
--
-- The cases for 'specHeader' below are presented in the same order as this
-- running example. The result of 'specHeader' for this example is as follows:
--
-- ( -- Returned arguments
-- env + [a -> T1, deqA -> dEqA']
-- , []
--
-- -- RULE helpers
-- , [b, dx', i]
-- , [T1, b, dx', i]
--
-- -- Specialised function helpers
-- , [b, i]
-- , [dx]
-- , [T1, b, dx_spec, i]
-- )
specHeader
:: SpecEnv
-> [CoreBndr] -- The binders from the original function 'f'
-> [SpecArg] -- From the CallInfo
-> SpecM ( -- Returned arguments
SpecEnv -- Substitution to apply to the body of 'f'
, [CoreBndr] -- All the remaining unspecialised args from the original function 'f'
-- RULE helpers
, [CoreBndr] -- Binders for the RULE
, [CoreArg] -- Args for the LHS of the rule
-- Specialised function helpers
, [CoreBndr] -- Binders for $sf
, [DictBind] -- Auxiliary dictionary bindings
, [CoreExpr] -- Specialised arguments for unfolding
)
-- We want to specialise on type 'T1', and so we must construct a substitution
-- 'a->T1', as well as a LHS argument for the resulting RULE and unfolding
-- details.
specHeader env (bndr : bndrs) (SpecType t : args)
= do { let env' = extendTvSubstList env [(bndr, t)]
; (env'', unused_bndrs, rule_bs, rule_es, bs', dx, spec_args)
<- specHeader env' bndrs args
; pure ( env''
, unused_bndrs
, rule_bs
, Type t : rule_es
, bs'
, dx
, Type t : spec_args
)
}
-- Next we have a type that we don't want to specialise. We need to perform
-- a substitution on it (in case the type refers to 'a'). Additionally, we need
-- to produce a binder, LHS argument and RHS argument for the resulting rule,
-- /and/ a binder for the specialised body.
specHeader env (bndr : bndrs) (UnspecType : args)
= do { let (env', bndr') = substBndr env bndr
; (env'', unused_bndrs, rule_bs, rule_es, bs', dx, spec_args)
<- specHeader env' bndrs args
; pure ( env''
, unused_bndrs
, bndr' : rule_bs
, varToCoreExpr bndr' : rule_es
, bndr' : bs'
, dx
, varToCoreExpr bndr' : spec_args
)
}
-- Next we want to specialise the 'Eq a' dict away. We need to construct
-- a wildcard binder to match the dictionary (See Note [Specialising Calls] for
-- the nitty-gritty), as a LHS rule and unfolding details.
specHeader env (bndr : bndrs) (SpecDict d : args)
= do { inst_dict_id <- newDictBndr env bndr
; let (rhs_env2, dx_binds, spec_dict_args')
= bindAuxiliaryDicts env [bndr] [d] [inst_dict_id]
; (env', unused_bndrs, rule_bs, rule_es, bs', dx, spec_args)
<- specHeader rhs_env2 bndrs args
; pure ( env'
, unused_bndrs
-- See Note [Evidence foralls]
, exprFreeIdsList (varToCoreExpr inst_dict_id) ++ rule_bs
, varToCoreExpr inst_dict_id : rule_es
, bs'
, dx_binds ++ dx
, spec_dict_args' ++ spec_args
)
}
-- Finally, we have the unspecialised argument 'i'. We need to produce
-- a binder, LHS and RHS argument for the RULE, and a binder for the
-- specialised body.
--
-- NB: Calls to 'specHeader' will trim off any trailing 'UnspecArg's, which is
-- why 'i' doesn't appear in our RULE above. But we have no guarantee that
-- there aren't 'UnspecArg's which come /before/ all of the dictionaries, so
-- this case must be here.
specHeader env (bndr : bndrs) (UnspecArg : args)
= do { let (env', bndr') = substBndr env bndr
; (env'', unused_bndrs, rule_bs, rule_es, bs', dx, spec_args)
<- specHeader env' bndrs args
; pure ( env''
, unused_bndrs
, bndr' : rule_bs
, varToCoreExpr bndr' : rule_es
, bndr' : bs'
, dx
, varToCoreExpr bndr' : spec_args
)
}
-- Return all remaining binders from the original function. These have the
-- invariant that they should all correspond to unspecialised arguments, so
-- it's safe to stop processing at this point.
specHeader env bndrs [] = pure (env, bndrs, [], [], [], [], [])
specHeader env [] _ = pure (env, [], [], [], [], [], [])
-- | Specialise a set of calls to imported bindings
specImports :: DynFlags
-> Module
-> SpecEnv -- Passed in so that all top-level Ids are in scope
-> VarSet -- Don't specialise these ones
-- See Note [Avoiding recursive specialisation]
-> [Id] -- Stack of imported functions being specialised
-> RuleBase -- Rules from this module and the home package
-- (but not external packages, which can change)
-> UsageDetails -- Calls for imported things, and floating bindings
-> CoreM ( [CoreRule] -- New rules
, [CoreBind] ) -- Specialised bindings
-- See Note [Wrapping bindings returned by specImports]
specImports dflags this_mod top_env done callers rule_base
(MkUD { ud_binds = dict_binds, ud_calls = calls })
-- See Note [Disabling cross-module specialisation]
| not $ gopt Opt_CrossModuleSpecialise dflags
= return ([], [])
| otherwise
= do { let import_calls = dVarEnvElts calls
; (rules, spec_binds) <- go rule_base import_calls
-- Don't forget to wrap the specialized bindings with
-- bindings for the needed dictionaries.
-- See Note [Wrap bindings returned by specImports]
; let spec_binds' = wrapDictBinds dict_binds spec_binds
; return (rules, spec_binds') }
where
go :: RuleBase -> [CallInfoSet] -> CoreM ([CoreRule], [CoreBind])
go _ [] = return ([], [])
go rb (cis@(CIS fn _) : other_calls)
= do { let ok_calls = filterCalls cis dict_binds
-- Drop calls that (directly or indirectly) refer to fn
-- See Note [Avoiding loops]
-- ; debugTraceMsg (text "specImport" <+> vcat [ ppr fn
-- , text "calls" <+> ppr cis
-- , text "ud_binds =" <+> ppr dict_binds
-- , text "dump set =" <+> ppr dump_set
-- , text "filtered calls =" <+> ppr ok_calls ])
; (rules1, spec_binds1) <- specImport dflags this_mod top_env
done callers rb fn ok_calls
; (rules2, spec_binds2) <- go (extendRuleBaseList rb rules1) other_calls
; return (rules1 ++ rules2, spec_binds1 ++ spec_binds2) }
specImport :: DynFlags
-> Module
-> SpecEnv -- Passed in so that all top-level Ids are in scope
-> VarSet -- Don't specialise these
-- See Note [Avoiding recursive specialisation]
-> [Id] -- Stack of imported functions being specialised
-> RuleBase -- Rules from this module
-> Id -> [CallInfo] -- Imported function and calls for it
-> CoreM ( [CoreRule] -- New rules
, [CoreBind] ) -- Specialised bindings
specImport dflags this_mod top_env done callers rb fn calls_for_fn
| fn `elemVarSet` done
= return ([], []) -- No warning. This actually happens all the time
-- when specialising a recursive function, because
-- the RHS of the specialised function contains a recursive
-- call to the original function
| null calls_for_fn -- We filtered out all the calls in deleteCallsMentioning
= return ([], [])
| wantSpecImport dflags unfolding
, Just rhs <- maybeUnfoldingTemplate unfolding
= do { -- Get rules from the external package state
-- We keep doing this in case we "page-fault in"
-- more rules as we go along
; hsc_env <- getHscEnv
; eps <- liftIO $ hscEPS hsc_env
; vis_orphs <- getVisibleOrphanMods
; let full_rb = unionRuleBase rb (eps_rule_base eps)
rules_for_fn = getRules (RuleEnv full_rb vis_orphs) fn
; (rules1, spec_pairs, uds)
<- -- pprTrace "specImport1" (vcat [ppr fn, ppr calls_for_fn, ppr rhs]) $
runSpecM dflags this_mod $
specCalls (Just this_mod) top_env rules_for_fn calls_for_fn fn rhs
; let spec_binds1 = [NonRec b r | (b,r) <- spec_pairs]
-- After the rules kick in we may get recursion, but
-- we rely on a global GlomBinds to sort that out later
-- See Note [Glom the bindings if imported functions are specialised]
-- Now specialise any cascaded calls
; (rules2, spec_binds2) <- -- pprTrace "specImport 2" (ppr fn $$ ppr rules1 $$ ppr spec_binds1) $
specImports dflags this_mod top_env
(extendVarSet done fn)
(fn:callers)
(extendRuleBaseList rb rules1)
uds
; let final_binds = spec_binds2 ++ spec_binds1
; return (rules2 ++ rules1, final_binds) }
| otherwise = do { tryWarnMissingSpecs dflags callers fn calls_for_fn
; return ([], [])}
where
unfolding = realIdUnfolding fn -- We want to see the unfolding even for loop breakers
-- | Returns whether or not to show a missed-spec warning.
-- If -Wall-missed-specializations is on, show the warning.
-- Otherwise, if -Wmissed-specializations is on, only show a warning
-- if there is at least one imported function being specialized,
-- and if all imported functions are marked with an inline pragma
-- Use the most specific warning as the reason.
tryWarnMissingSpecs :: DynFlags -> [Id] -> Id -> [CallInfo] -> CoreM ()
-- See Note [Warning about missed specialisations]
tryWarnMissingSpecs dflags callers fn calls_for_fn
| wopt Opt_WarnMissedSpecs dflags
&& not (null callers)
&& allCallersInlined = doWarn $ Reason Opt_WarnMissedSpecs
| wopt Opt_WarnAllMissedSpecs dflags = doWarn $ Reason Opt_WarnAllMissedSpecs
| otherwise = return ()
where
allCallersInlined = all (isAnyInlinePragma . idInlinePragma) callers
doWarn reason =
warnMsg reason
(vcat [ hang (text ("Could not specialise imported function") <+> quotes (ppr fn))
2 (vcat [ text "when specialising" <+> quotes (ppr caller)
| caller <- callers])
, whenPprDebug (text "calls:" <+> vcat (map (pprCallInfo fn) calls_for_fn))
, text "Probable fix: add INLINABLE pragma on" <+> quotes (ppr fn) ])
wantSpecImport :: DynFlags -> Unfolding -> Bool
-- See Note [Specialise imported INLINABLE things]
wantSpecImport dflags unf
= case unf of
NoUnfolding -> False
BootUnfolding -> False
OtherCon {} -> False
DFunUnfolding {} -> True
CoreUnfolding { uf_src = src, uf_guidance = _guidance }
| gopt Opt_SpecialiseAggressively dflags -> True
| isStableSource src -> True
-- Specialise even INLINE things; it hasn't inlined yet,
-- so perhaps it never will. Moreover it may have calls
-- inside it that we want to specialise
| otherwise -> False -- Stable, not INLINE, hence INLINABLE
{- Note [Warning about missed specialisations]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Suppose
* In module Lib, you carefully mark a function 'foo' INLINABLE
* Import Lib(foo) into another module M
* Call 'foo' at some specialised type in M
Then you jolly well expect it to be specialised in M. But what if
'foo' calls another function 'Lib.bar'. Then you'd like 'bar' to be
specialised too. But if 'bar' is not marked INLINABLE it may well
not be specialised. The warning Opt_WarnMissedSpecs warns about this.
It's more noisy to warning about a missed specialisation opportunity
for /every/ overloaded imported function, but sometimes useful. That
is what Opt_WarnAllMissedSpecs does.
ToDo: warn about missed opportunities for local functions.
Note [Specialise imported INLINABLE things]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
What imported functions do we specialise? The basic set is
* DFuns and things with INLINABLE pragmas.
but with -fspecialise-aggressively we add
* Anything with an unfolding template
#8874 has a good example of why we want to auto-specialise DFuns.
We have the -fspecialise-aggressively flag (usually off), because we
risk lots of orphan modules from over-vigorous specialisation.
However it's not a big deal: anything non-recursive with an
unfolding-template will probably have been inlined already.
Note [Glom the bindings if imported functions are specialised]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Suppose we have an imported, *recursive*, INLINABLE function
f :: Eq a => a -> a
f = /\a \d x. ...(f a d)...
In the module being compiled we have
g x = f (x::Int)
Now we'll make a specialised function
f_spec :: Int -> Int
f_spec = \x -> ...(f Int dInt)...
{-# RULE f Int _ = f_spec #-}
g = \x. f Int dInt x
Note that f_spec doesn't look recursive
After rewriting with the RULE, we get
f_spec = \x -> ...(f_spec)...
BUT since f_spec was non-recursive before it'll *stay* non-recursive.
The occurrence analyser never turns a NonRec into a Rec. So we must
make sure that f_spec is recursive. Easiest thing is to make all
the specialisations for imported bindings recursive.
Note [Avoiding recursive specialisation]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
When we specialise 'f' we may find new overloaded calls to 'g', 'h' in
'f's RHS. So we want to specialise g,h. But we don't want to
specialise f any more! It's possible that f's RHS might have a
recursive yet-more-specialised call, so we'd diverge in that case.
And if the call is to the same type, one specialisation is enough.
Avoiding this recursive specialisation loop is the reason for the
'done' VarSet passed to specImports and specImport.
************************************************************************
* *
\subsubsection{@specExpr@: the main function}
* *
************************************************************************
-}
data SpecEnv
= SE { se_subst :: CoreSubst.Subst
-- We carry a substitution down:
-- a) we must clone any binding that might float outwards,
-- to avoid name clashes
-- b) we carry a type substitution to use when analysing
-- the RHS of specialised bindings (no type-let!)
, se_interesting :: VarSet
-- Dict Ids that we know something about
-- and hence may be worth specialising against
-- See Note [Interesting dictionary arguments]
}
specVar :: SpecEnv -> Id -> CoreExpr
specVar env v = CoreSubst.lookupIdSubst (text "specVar") (se_subst env) v
specExpr :: SpecEnv -> CoreExpr -> SpecM (CoreExpr, UsageDetails)
---------------- First the easy cases --------------------
specExpr env (Type ty) = return (Type (substTy env ty), emptyUDs)
specExpr env (Coercion co) = return (Coercion (substCo env co), emptyUDs)
specExpr env (Var v) = return (specVar env v, emptyUDs)
specExpr _ (Lit lit) = return (Lit lit, emptyUDs)
specExpr env (Cast e co)
= do { (e', uds) <- specExpr env e
; return ((mkCast e' (substCo env co)), uds) }
specExpr env (Tick tickish body)
= do { (body', uds) <- specExpr env body
; return (Tick (specTickish env tickish) body', uds) }
---------------- Applications might generate a call instance --------------------
specExpr env expr@(App {})
= go expr []
where
go (App fun arg) args = do (arg', uds_arg) <- specExpr env arg
(fun', uds_app) <- go fun (arg':args)
return (App fun' arg', uds_arg `plusUDs` uds_app)
go (Var f) args = case specVar env f of
Var f' -> return (Var f', mkCallUDs env f' args)
e' -> return (e', emptyUDs) -- I don't expect this!
go other _ = specExpr env other
---------------- Lambda/case require dumping of usage details --------------------
specExpr env e@(Lam _ _) = do
(body', uds) <- specExpr env' body
let (free_uds, dumped_dbs) = dumpUDs bndrs' uds
return (mkLams bndrs' (wrapDictBindsE dumped_dbs body'), free_uds)
where
(bndrs, body) = collectBinders e
(env', bndrs') = substBndrs env bndrs
-- More efficient to collect a group of binders together all at once
-- and we don't want to split a lambda group with dumped bindings
specExpr env (Case scrut case_bndr ty alts)
= do { (scrut', scrut_uds) <- specExpr env scrut
; (scrut'', case_bndr', alts', alts_uds)
<- specCase env scrut' case_bndr alts
; return (Case scrut'' case_bndr' (substTy env ty) alts'
, scrut_uds `plusUDs` alts_uds) }
---------------- Finally, let is the interesting case --------------------
specExpr env (Let bind body)
= do { -- Clone binders
(rhs_env, body_env, bind') <- cloneBindSM env bind
-- Deal with the body
; (body', body_uds) <- specExpr body_env body
-- Deal with the bindings
; (binds', uds) <- specBind rhs_env bind' body_uds
-- All done
; return (foldr Let body' binds', uds) }
specTickish :: SpecEnv -> Tickish Id -> Tickish Id
specTickish env (Breakpoint ix ids)
= Breakpoint ix [ id' | id <- ids, Var id' <- [specVar env id]]
-- drop vars from the list if they have a non-variable substitution.
-- should never happen, but it's harmless to drop them anyway.
specTickish _ other_tickish = other_tickish
specCase :: SpecEnv
-> CoreExpr -- Scrutinee, already done
-> Id -> [CoreAlt]
-> SpecM ( CoreExpr -- New scrutinee
, Id
, [CoreAlt]
, UsageDetails)
specCase env scrut' case_bndr [(con, args, rhs)]
| isDictId case_bndr -- See Note [Floating dictionaries out of cases]
, interestingDict env scrut'
, not (isDeadBinder case_bndr && null sc_args')
= do { (case_bndr_flt : sc_args_flt) <- mapM clone_me (case_bndr' : sc_args')
; let sc_rhss = [ Case (Var case_bndr_flt) case_bndr' (idType sc_arg')
[(con, args', Var sc_arg')]
| sc_arg' <- sc_args' ]
-- Extend the substitution for RHS to map the *original* binders
-- to their floated versions.
mb_sc_flts :: [Maybe DictId]
mb_sc_flts = map (lookupVarEnv clone_env) args'
clone_env = zipVarEnv sc_args' sc_args_flt
subst_prs = (case_bndr, Var case_bndr_flt)
: [ (arg, Var sc_flt)
| (arg, Just sc_flt) <- args `zip` mb_sc_flts ]
env_rhs' = env_rhs { se_subst = CoreSubst.extendIdSubstList (se_subst env_rhs) subst_prs
, se_interesting = se_interesting env_rhs `extendVarSetList`
(case_bndr_flt : sc_args_flt) }
; (rhs', rhs_uds) <- specExpr env_rhs' rhs
; let scrut_bind = mkDB (NonRec case_bndr_flt scrut')
case_bndr_set = unitVarSet case_bndr_flt
sc_binds = [(NonRec sc_arg_flt sc_rhs, case_bndr_set)
| (sc_arg_flt, sc_rhs) <- sc_args_flt `zip` sc_rhss ]
flt_binds = scrut_bind : sc_binds
(free_uds, dumped_dbs) = dumpUDs (case_bndr':args') rhs_uds
all_uds = flt_binds `addDictBinds` free_uds
alt' = (con, args', wrapDictBindsE dumped_dbs rhs')
; return (Var case_bndr_flt, case_bndr', [alt'], all_uds) }
where
(env_rhs, (case_bndr':args')) = substBndrs env (case_bndr:args)
sc_args' = filter is_flt_sc_arg args'
clone_me bndr = do { uniq <- getUniqueM
; return (mkUserLocalOrCoVar occ uniq ty loc) }
where
name = idName bndr
ty = idType bndr
occ = nameOccName name
loc = getSrcSpan name
arg_set = mkVarSet args'
is_flt_sc_arg var = isId var
&& not (isDeadBinder var)
&& isDictTy var_ty
&& not (tyCoVarsOfType var_ty `intersectsVarSet` arg_set)
where
var_ty = idType var
specCase env scrut case_bndr alts
= do { (alts', uds_alts) <- mapAndCombineSM spec_alt alts
; return (scrut, case_bndr', alts', uds_alts) }
where
(env_alt, case_bndr') = substBndr env case_bndr
spec_alt (con, args, rhs) = do
(rhs', uds) <- specExpr env_rhs rhs
let (free_uds, dumped_dbs) = dumpUDs (case_bndr' : args') uds
return ((con, args', wrapDictBindsE dumped_dbs rhs'), free_uds)
where
(env_rhs, args') = substBndrs env_alt args
{-
Note [Floating dictionaries out of cases]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Consider
g = \d. case d of { MkD sc ... -> ...(f sc)... }
Naively we can't float d2's binding out of the case expression,
because 'sc' is bound by the case, and that in turn means we can't
specialise f, which seems a pity.
So we invert the case, by floating out a binding
for 'sc_flt' thus:
sc_flt = case d of { MkD sc ... -> sc }
Now we can float the call instance for 'f'. Indeed this is just
what'll happen if 'sc' was originally bound with a let binding,
but case is more efficient, and necessary with equalities. So it's
good to work with both.
You might think that this won't make any difference, because the
call instance will only get nuked by the \d. BUT if 'g' itself is
specialised, then transitively we should be able to specialise f.
In general, given
case e of cb { MkD sc ... -> ...(f sc)... }
we transform to
let cb_flt = e
sc_flt = case cb_flt of { MkD sc ... -> sc }
in
case cb_flt of bg { MkD sc ... -> ....(f sc_flt)... }
The "_flt" things are the floated binds; we use the current substitution
to substitute sc -> sc_flt in the RHS
************************************************************************
* *
Dealing with a binding
* *
************************************************************************
-}
specBind :: SpecEnv -- Use this for RHSs
-> CoreBind -- Binders are already cloned by cloneBindSM,
-- but RHSs are un-processed
-> UsageDetails -- Info on how the scope of the binding
-> SpecM ([CoreBind], -- New bindings
UsageDetails) -- And info to pass upstream
-- Returned UsageDetails:
-- No calls for binders of this bind
specBind rhs_env (NonRec fn rhs) body_uds
= do { (rhs', rhs_uds) <- specExpr rhs_env rhs
; (fn', spec_defns, body_uds1) <- specDefn rhs_env body_uds fn rhs
; let pairs = spec_defns ++ [(fn', rhs')]
-- fn' mentions the spec_defns in its rules,
-- so put the latter first
combined_uds = body_uds1 `plusUDs` rhs_uds
(free_uds, dump_dbs, float_all) = dumpBindUDs [fn] combined_uds
final_binds :: [DictBind]
-- See Note [From non-recursive to recursive]
final_binds
| not (isEmptyBag dump_dbs)
, not (null spec_defns)
= [recWithDumpedDicts pairs dump_dbs]
| otherwise
= [mkDB $ NonRec b r | (b,r) <- pairs]
++ bagToList dump_dbs
; if float_all then
-- Rather than discard the calls mentioning the bound variables
-- we float this (dictionary) binding along with the others
return ([], free_uds `snocDictBinds` final_binds)
else
-- No call in final_uds mentions bound variables,
-- so we can just leave the binding here
return (map fst final_binds, free_uds) }
specBind rhs_env (Rec pairs) body_uds
-- Note [Specialising a recursive group]
= do { let (bndrs,rhss) = unzip pairs
; (rhss', rhs_uds) <- mapAndCombineSM (specExpr rhs_env) rhss
; let scope_uds = body_uds `plusUDs` rhs_uds
-- Includes binds and calls arising from rhss
; (bndrs1, spec_defns1, uds1) <- specDefns rhs_env scope_uds pairs
; (bndrs3, spec_defns3, uds3)
<- if null spec_defns1 -- Common case: no specialisation
then return (bndrs1, [], uds1)
else do { -- Specialisation occurred; do it again
(bndrs2, spec_defns2, uds2)
<- specDefns rhs_env uds1 (bndrs1 `zip` rhss)
; return (bndrs2, spec_defns2 ++ spec_defns1, uds2) }
; let (final_uds, dumped_dbs, float_all) = dumpBindUDs bndrs uds3
final_bind = recWithDumpedDicts (spec_defns3 ++ zip bndrs3 rhss')
dumped_dbs
; if float_all then
return ([], final_uds `snocDictBind` final_bind)
else
return ([fst final_bind], final_uds) }
---------------------------
specDefns :: SpecEnv
-> UsageDetails -- Info on how it is used in its scope
-> [(OutId,InExpr)] -- The things being bound and their un-processed RHS
-> SpecM ([OutId], -- Original Ids with RULES added
[(OutId,OutExpr)], -- Extra, specialised bindings
UsageDetails) -- Stuff to fling upwards from the specialised versions
-- Specialise a list of bindings (the contents of a Rec), but flowing usages
-- upwards binding by binding. Example: { f = ...g ...; g = ...f .... }
-- Then if the input CallDetails has a specialised call for 'g', whose specialisation
-- in turn generates a specialised call for 'f', we catch that in this one sweep.
-- But not vice versa (it's a fixpoint problem).
specDefns _env uds []
= return ([], [], uds)
specDefns env uds ((bndr,rhs):pairs)
= do { (bndrs1, spec_defns1, uds1) <- specDefns env uds pairs
; (bndr1, spec_defns2, uds2) <- specDefn env uds1 bndr rhs
; return (bndr1 : bndrs1, spec_defns1 ++ spec_defns2, uds2) }
---------------------------
specDefn :: SpecEnv
-> UsageDetails -- Info on how it is used in its scope
-> OutId -> InExpr -- The thing being bound and its un-processed RHS
-> SpecM (Id, -- Original Id with added RULES
[(Id,CoreExpr)], -- Extra, specialised bindings
UsageDetails) -- Stuff to fling upwards from the specialised versions
specDefn env body_uds fn rhs
= do { let (body_uds_without_me, calls_for_me) = callsForMe fn body_uds
rules_for_me = idCoreRules fn
; (rules, spec_defns, spec_uds) <- specCalls Nothing env rules_for_me
calls_for_me fn rhs
; return ( fn `addIdSpecialisations` rules
, spec_defns
, body_uds_without_me `plusUDs` spec_uds) }
-- It's important that the `plusUDs` is this way
-- round, because body_uds_without_me may bind
-- dictionaries that are used in calls_for_me passed
-- to specDefn. So the dictionary bindings in
-- spec_uds may mention dictionaries bound in
-- body_uds_without_me
---------------------------
specCalls :: Maybe Module -- Just this_mod => specialising imported fn
-- Nothing => specialising local fn
-> SpecEnv
-> [CoreRule] -- Existing RULES for the fn
-> [CallInfo]
-> OutId -> InExpr
-> SpecM SpecInfo -- New rules, specialised bindings, and usage details
-- This function checks existing rules, and does not create
-- duplicate ones. So the caller does not need to do this filtering.
-- See 'already_covered'
type SpecInfo = ( [CoreRule] -- Specialisation rules
, [(Id,CoreExpr)] -- Specialised definition
, UsageDetails ) -- Usage details from specialised RHSs
specCalls mb_mod env existing_rules calls_for_me fn rhs
-- The first case is the interesting one
| callSpecArity pis <= fn_arity -- See Note [Specialisation Must Preserve Sharing]
&& notNull calls_for_me -- And there are some calls to specialise
&& not (isNeverActive (idInlineActivation fn))
-- Don't specialise NOINLINE things
-- See Note [Auto-specialisation and RULES]
-- && not (certainlyWillInline (idUnfolding fn)) -- And it's not small
-- See Note [Inline specialisation] for why we do not
-- switch off specialisation for inline functions
= -- pprTrace "specDefn: some" (ppr fn $$ ppr calls_for_me $$ ppr existing_rules) $
foldlM spec_call ([], [], emptyUDs) calls_for_me
| otherwise -- No calls or RHS doesn't fit our preconceptions
= WARN( not (exprIsTrivial rhs) && notNull calls_for_me,
text "Missed specialisation opportunity for"
<+> ppr fn $$ _trace_doc )
-- Note [Specialisation shape]
-- pprTrace "specDefn: none" (ppr fn <+> ppr calls_for_me) $
return ([], [], emptyUDs)
where
_trace_doc = sep [ ppr rhs_tyvars, ppr rhs_bndrs
, ppr (idInlineActivation fn) ]
fn_type = idType fn
fn_arity = idArity fn
fn_unf = realIdUnfolding fn -- Ignore loop-breaker-ness here
pis = fst $ splitPiTys fn_type
theta = getTheta pis
n_dicts = length theta
inl_prag = idInlinePragma fn
inl_act = inlinePragmaActivation inl_prag
is_local = isLocalId fn
-- Figure out whether the function has an INLINE pragma
-- See Note [Inline specialisations]
(rhs_bndrs, rhs_body) = collectBindersPushingCo rhs
-- See Note [Account for casts in binding]
rhs_tyvars = filter isTyVar rhs_bndrs
in_scope = CoreSubst.substInScope (se_subst env)
already_covered :: DynFlags -> [CoreRule] -> [CoreExpr] -> Bool
already_covered dflags new_rules args -- Note [Specialisations already covered]
= isJust (lookupRule dflags (in_scope, realIdUnfolding)
(const True) fn args
(new_rules ++ existing_rules))
-- NB: we look both in the new_rules (generated by this invocation
-- of specCalls), and in existing_rules (passed in to specCalls)
----------------------------------------------------------
-- Specialise to one particular call pattern
spec_call :: SpecInfo -- Accumulating parameter
-> CallInfo -- Call instance
-> SpecM SpecInfo
spec_call spec_acc@(rules_acc, pairs_acc, uds_acc)
(CI { ci_key = call_args, ci_arity = call_arity })
= ASSERT(call_arity <= fn_arity)
-- See Note [Specialising Calls]
do { (rhs_env2, unused_bndrs, rule_bndrs, rule_args, unspec_bndrs, dx_binds, spec_args)
<- specHeader env rhs_bndrs $ dropWhileEndLE isUnspecArg call_args
; let rhs_body' = mkLams unused_bndrs rhs_body
; dflags <- getDynFlags
; if already_covered dflags rules_acc rule_args
then return spec_acc
else -- pprTrace "spec_call" (vcat [ ppr _call_info, ppr fn, ppr rhs_dict_ids
-- , text "rhs_env2" <+> ppr (se_subst rhs_env2)
-- , ppr dx_binds ]) $
do
{ -- Figure out the type of the specialised function
let body = mkLams unspec_bndrs rhs_body'
body_ty = substTy rhs_env2 $ exprType body
(lam_extra_args, app_args) -- See Note [Specialisations Must Be Lifted]
| isUnliftedType body_ty -- C.f. WwLib.mkWorkerArgs
, not (isJoinId fn)
= ([voidArgId], unspec_bndrs ++ [voidPrimId])
| otherwise = ([], unspec_bndrs)
join_arity_change = length app_args - length rule_args
spec_join_arity | Just orig_join_arity <- isJoinId_maybe fn
= Just (orig_join_arity + join_arity_change)
| otherwise
= Nothing
; (spec_rhs, rhs_uds) <- specExpr rhs_env2 (mkLams lam_extra_args body)
; let spec_id_ty = exprType spec_rhs
; spec_f <- newSpecIdSM fn spec_id_ty spec_join_arity
; this_mod <- getModule
; let
-- The rule to put in the function's specialisation is:
-- forall x @b d1' d2'.
-- f x @T1 @b @T2 d1' d2' = f1 x @b
-- See Note [Specialising Calls]
herald = case mb_mod of
Nothing -- Specialising local fn
-> text "SPEC"
Just this_mod -- Specialising imported fn
-> text "SPEC/" <> ppr this_mod
rule_name = mkFastString $ showSDoc dflags $
herald <+> ftext (occNameFS (getOccName fn))
<+> hsep (mapMaybe ppr_call_key_ty call_args)
-- This name ends up in interface files, so use occNameString.
-- Otherwise uniques end up there, making builds
-- less deterministic (See #4012 comment:61 ff)
rule_wout_eta = mkRule
this_mod
True {- Auto generated -}
is_local
rule_name
inl_act -- Note [Auto-specialisation and RULES]
(idName fn)
rule_bndrs
rule_args
(mkVarApps (Var spec_f) app_args)
spec_rule
= case isJoinId_maybe fn of
Just join_arity -> etaExpandToJoinPointRule join_arity
rule_wout_eta
Nothing -> rule_wout_eta
-- Add the { d1' = dx1; d2' = dx2 } usage stuff
-- See Note [Specialising Calls]
spec_uds = foldr consDictBind rhs_uds dx_binds
--------------------------------------
-- Add a suitable unfolding if the spec_inl_prag says so
-- See Note [Inline specialisations]
(spec_inl_prag, spec_unf)
| not is_local && isStrongLoopBreaker (idOccInfo fn)
= (neverInlinePragma, noUnfolding)
-- See Note [Specialising imported functions] in OccurAnal
| InlinePragma { inl_inline = Inlinable } <- inl_prag
= (inl_prag { inl_inline = NoUserInline }, noUnfolding)
| otherwise
= (inl_prag, specUnfolding dflags unspec_bndrs spec_app n_dicts fn_unf)
spec_app e = e `mkApps` spec_args
--------------------------------------
-- Adding arity information just propagates it a bit faster
-- See Note [Arity decrease] in Simplify
-- Copy InlinePragma information from the parent Id.
-- So if f has INLINE[1] so does spec_f
spec_f_w_arity = spec_f `setIdArity` max 0 (fn_arity - n_dicts)
`setInlinePragma` spec_inl_prag
`setIdUnfolding` spec_unf
`asJoinId_maybe` spec_join_arity
_rule_trace_doc = vcat [ ppr spec_f, ppr fn_type, ppr spec_id_ty
, ppr rhs_bndrs, ppr call_args
, ppr spec_rule
]
; -- pprTrace "spec_call: rule" _rule_trace_doc
return ( spec_rule : rules_acc
, (spec_f_w_arity, spec_rhs) : pairs_acc
, spec_uds `plusUDs` uds_acc
) } }
{- Note [Specialisation Must Preserve Sharing]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Consider a function:
f :: forall a. Eq a => a -> blah
f =
if expensive
then f1
else f2
As written, all calls to 'f' will share 'expensive'. But if we specialise 'f'
at 'Int', eg:
$sfInt = SUBST[a->Int,dict->dEqInt] (if expensive then f1 else f2)
RULE "SPEC f"
forall (d :: Eq Int).
f Int _ = $sfIntf
We've now lost sharing between 'f' and '$sfInt' for 'expensive'. Yikes!
To avoid this, we only generate specialisations for functions whose arity is
enough to bind all of the arguments we need to specialise. This ensures our
specialised functions don't do any work before receiving all of their dicts,
and thus avoids the 'f' case above.
Note [Specialisations Must Be Lifted]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Consider a function 'f':
f = forall a. Eq a => Array# a
used like
case x of
True -> ...f @Int dEqInt...
False -> 0
Naively, we might generate an (expensive) specialisation
$sfInt :: Array# Int
even in the case that @x = False@! Instead, we add a dummy 'Void#' argument to
the specialisation '$sfInt' ($sfInt :: Void# -> Array# Int) in order to
preserve laziness.
Note [Specialising Calls]
~~~~~~~~~~~~~~~~~~~~~~~~~
Suppose we have a function:
f :: Int -> forall a b c. (Foo a, Foo c) => Bar -> Qux
f = \x -> /\ a b c -> \d1 d2 bar -> rhs
and suppose it is called at:
f 7 @T1 @T2 @T3 dFooT1 dFooT3 bar
This call is described as a 'CallInfo' whose 'ci_key' is
[ UnspecArg, SpecType T1, UnspecType, SpecType T3, SpecDict dFooT1
, SpecDict dFooT3, UnspecArg ]
Why are 'a' and 'c' identified as 'SpecType', while 'b' is 'UnspecType'?
Because we must specialise the function on type variables that appear
free in its *dictionary* arguments; but not on type variables that do not
appear in any dictionaries, i.e. are fully polymorphic.
Because this call has dictionaries applied, we'd like to specialise
the call on any type argument that appears free in those dictionaries.
In this case, those are (a ~ T1, c ~ T3).
As a result, we'd like to generate a function:
$sf :: Int -> forall b. Bar -> Qux
$sf = SUBST[a->T1, c->T3, d1->d1', d2->d2'] (\x -> /\ b -> \bar -> rhs)
Note that the substitution is applied to the whole thing. This is
convenient, but just slightly fragile. Notably:
* There had better be no name clashes in a/b/c
We must construct a rewrite rule:
RULE "SPEC f @T1 _ @T3"
forall (x :: Int) (@b :: Type) (d1' :: Foo T1) (d2' :: Foo T3).
f x @T1 @b @T3 d1' d2' = $sf x @b
In the rule, d1' and d2' are just wildcards, not used in the RHS. Note
additionally that 'bar' isn't captured by this rule --- we bind only
enough etas in order to capture all of the *specialised* arguments.
Finally, we must also construct the usage-details
{ d1' = dx1; d2' = dx2 }
where d1', d2' are cloned versions of d1,d2, with the type substitution
applied. These auxiliary bindings just avoid duplication of dx1, dx2.
Note [Account for casts in binding]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Consider
f :: Eq a => a -> IO ()
{-# INLINABLE f
StableUnf = (/\a \(d:Eq a) (x:a). blah) |> g
#-}
f = ...
In f's stable unfolding we have done some modest simplification which
has pushed the cast to the outside. (I wonder if this is the Right
Thing, but it's what happens now; see SimplUtils Note [Casts and
lambdas].) Now that stable unfolding must be specialised, so we want
to push the cast back inside. It would be terrible if the cast
defeated specialisation! Hence the use of collectBindersPushingCo.
Note [Evidence foralls]
~~~~~~~~~~~~~~~~~~~~~~~~~~
Suppose (#12212) that we are specialising
f :: forall a b. (Num a, F a ~ F b) => blah
with a=b=Int. Then the RULE will be something like
RULE forall (d:Num Int) (g :: F Int ~ F Int).
f Int Int d g = f_spec
But both varToCoreExpr (when constructing the LHS args), and the
simplifier (when simplifying the LHS args), will transform to
RULE forall (d:Num Int) (g :: F Int ~ F Int).
f Int Int d <F Int> = f_spec
by replacing g with Refl. So now 'g' is unbound, which results in a later
crash. So we use Refl right off the bat, and do not forall-quantify 'g':
* varToCoreExpr generates a Refl
* exprsFreeIdsList returns the Ids bound by the args,
which won't include g
You might wonder if this will match as often, but the simplifier replaces
complicated Refl coercions with Refl pretty aggressively.
Note [Orphans and auto-generated rules]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
When we specialise an INLINABLE function, or when we have
-fspecialise-aggressively, we auto-generate RULES that are orphans.
We don't want to warn about these, or we'd generate a lot of warnings.
Thus, we only warn about user-specified orphan rules.
Indeed, we don't even treat the module as an orphan module if it has
auto-generated *rule* orphans. Orphan modules are read every time we
compile, so they are pretty obtrusive and slow down every compilation,
even non-optimised ones. (Reason: for type class instances it's a
type correctness issue.) But specialisation rules are strictly for
*optimisation* only so it's fine not to read the interface.
What this means is that a SPEC rules from auto-specialisation in
module M will be used in other modules only if M.hi has been read for
some other reason, which is actually pretty likely.
-}
bindAuxiliaryDicts
:: SpecEnv
-> [DictId] -> [CoreExpr] -- Original dict bndrs, and the witnessing expressions
-> [DictId] -- A cloned dict-id for each dict arg
-> (SpecEnv, -- Substitute for all orig_dicts
[DictBind], -- Auxiliary dict bindings
[CoreExpr]) -- Witnessing expressions (all trivial)
-- Bind any dictionary arguments to fresh names, to preserve sharing
bindAuxiliaryDicts env@(SE { se_subst = subst, se_interesting = interesting })
orig_dict_ids call_ds inst_dict_ids
= (env', dx_binds, spec_dict_args)
where
(dx_binds, spec_dict_args) = go call_ds inst_dict_ids
env' = env { se_subst = subst `CoreSubst.extendSubstList`
(orig_dict_ids `zip` spec_dict_args)
`CoreSubst.extendInScopeList` dx_ids
, se_interesting = interesting `unionVarSet` interesting_dicts }
dx_ids = [dx_id | (NonRec dx_id _, _) <- dx_binds]
interesting_dicts = mkVarSet [ dx_id | (NonRec dx_id dx, _) <- dx_binds
, interestingDict env dx ]
-- See Note [Make the new dictionaries interesting]
go :: [CoreExpr] -> [CoreBndr] -> ([DictBind], [CoreExpr])
go [] _ = ([], [])
go (dx:dxs) (dx_id:dx_ids)
| exprIsTrivial dx = (dx_binds, dx : args)
| otherwise = (mkDB (NonRec dx_id dx) : dx_binds, Var dx_id : args)
where
(dx_binds, args) = go dxs dx_ids
-- In the first case extend the substitution but not bindings;
-- in the latter extend the bindings but not the substitution.
-- For the former, note that we bind the *original* dict in the substitution,
-- overriding any d->dx_id binding put there by substBndrs
go _ _ = pprPanic "bindAuxiliaryDicts" (ppr orig_dict_ids $$ ppr call_ds $$ ppr inst_dict_ids)
{-
Note [Make the new dictionaries interesting]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Important! We're going to substitute dx_id1 for d
and we want it to look "interesting", else we won't gather *any*
consequential calls. E.g.
f d = ...g d....
If we specialise f for a call (f (dfun dNumInt)), we'll get
a consequent call (g d') with an auxiliary definition
d' = df dNumInt
We want that consequent call to look interesting
Note [From non-recursive to recursive]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Even in the non-recursive case, if any dict-binds depend on 'fn' we might
have built a recursive knot
f a d x = <blah>
MkUD { ud_binds = NonRec d7 (MkD ..f..)
, ud_calls = ...(f T d7)... }
The we generate
Rec { fs x = <blah>[T/a, d7/d]
f a d x = <blah>
RULE f T _ = fs
d7 = ...f... }
Here the recursion is only through the RULE.
However we definitely should /not/ make the Rec in this wildly common
case:
d = ...
MkUD { ud_binds = NonRec d7 (...d...)
, ud_calls = ...(f T d7)... }
Here we want simply to add d to the floats, giving
MkUD { ud_binds = NonRec d (...)
NonRec d7 (...d...)
, ud_calls = ...(f T d7)... }
In general, we need only make this Rec if
- there are some specialisations (spec_binds non-empty)
- there are some dict_binds that depend on f (dump_dbs non-empty)
Note [Avoiding loops]
~~~~~~~~~~~~~~~~~~~~~
When specialising /dictionary functions/ we must be very careful to
avoid building loops. Here is an example that bit us badly: #3591
class Eq a => C a
instance Eq [a] => C [a]
This translates to
dfun :: Eq [a] -> C [a]
dfun a d = MkD a d (meth d)
d4 :: Eq [T] = <blah>
d2 :: C [T] = dfun T d4
d1 :: Eq [T] = $p1 d2
d3 :: C [T] = dfun T d1
None of these definitions is recursive. What happened was that we
generated a specialisation:
RULE forall d. dfun T d = dT :: C [T]
dT = (MkD a d (meth d)) [T/a, d1/d]
= MkD T d1 (meth d1)
But now we use the RULE on the RHS of d2, to get
d2 = dT = MkD d1 (meth d1)
d1 = $p1 d2
and now d1 is bottom! The problem is that when specialising 'dfun' we
should first dump "below" the binding all floated dictionary bindings
that mention 'dfun' itself. So d2 and d3 (and hence d1) must be
placed below 'dfun', and thus unavailable to it when specialising
'dfun'. That in turn means that the call (dfun T d1) must be
discarded. On the other hand, the call (dfun T d4) is fine, assuming
d4 doesn't mention dfun.
Solution:
Discard all calls that mention dictionaries that depend
(directly or indirectly) on the dfun we are specialising.
This is done by 'filterCalls'
--------------
Here's another example, this time for an imported dfun, so the call
to filterCalls is in specImports (#13429). Suppose we have
class Monoid v => C v a where ...
We start with a call
f @ [Integer] @ Integer $fC[]Integer
Specialising call to 'f' gives dict bindings
$dMonoid_1 :: Monoid [Integer]
$dMonoid_1 = M.$p1C @ [Integer] $fC[]Integer
$dC_1 :: C [Integer] (Node [Integer] Integer)
$dC_1 = M.$fCvNode @ [Integer] $dMonoid_1
...plus a recursive call to
f @ [Integer] @ (Node [Integer] Integer) $dC_1
Specialising that call gives
$dMonoid_2 :: Monoid [Integer]
$dMonoid_2 = M.$p1C @ [Integer] $dC_1
$dC_2 :: C [Integer] (Node [Integer] Integer)
$dC_2 = M.$fCvNode @ [Integer] $dMonoid_2
Now we have two calls to the imported function
M.$fCvNode :: Monoid v => C v a
M.$fCvNode @v @a m = C m some_fun
But we must /not/ use the call (M.$fCvNode @ [Integer] $dMonoid_2)
for specialisation, else we get:
$dC_1 = M.$fCvNode @ [Integer] $dMonoid_1
$dMonoid_2 = M.$p1C @ [Integer] $dC_1
$s$fCvNode = C $dMonoid_2 ...
RULE M.$fCvNode [Integer] _ _ = $s$fCvNode
Now use the rule to rewrite the call in the RHS of $dC_1
and we get a loop!
--------------
Here's yet another example
class C a where { foo,bar :: [a] -> [a] }
instance C Int where
foo x = r_bar x
bar xs = reverse xs
r_bar :: C a => [a] -> [a]
r_bar xs = bar (xs ++ xs)
That translates to:
r_bar a (c::C a) (xs::[a]) = bar a d (xs ++ xs)
Rec { $fCInt :: C Int = MkC foo_help reverse
foo_help (xs::[Int]) = r_bar Int $fCInt xs }
The call (r_bar $fCInt) mentions $fCInt,
which mentions foo_help,
which mentions r_bar
But we DO want to specialise r_bar at Int:
Rec { $fCInt :: C Int = MkC foo_help reverse
foo_help (xs::[Int]) = r_bar Int $fCInt xs
r_bar a (c::C a) (xs::[a]) = bar a d (xs ++ xs)
RULE r_bar Int _ = r_bar_Int
r_bar_Int xs = bar Int $fCInt (xs ++ xs)
}
Note that, because of its RULE, r_bar joins the recursive
group. (In this case it'll unravel a short moment later.)
Note [Specialising a recursive group]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Consider
let rec { f x = ...g x'...
; g y = ...f y'.... }
in f 'a'
Here we specialise 'f' at Char; but that is very likely to lead to
a specialisation of 'g' at Char. We must do the latter, else the
whole point of specialisation is lost.
But we do not want to keep iterating to a fixpoint, because in the
presence of polymorphic recursion we might generate an infinite number
of specialisations.
So we use the following heuristic:
* Arrange the rec block in dependency order, so far as possible
(the occurrence analyser already does this)
* Specialise it much like a sequence of lets
* Then go through the block a second time, feeding call-info from
the RHSs back in the bottom, as it were
In effect, the ordering maxmimises the effectiveness of each sweep,
and we do just two sweeps. This should catch almost every case of
monomorphic recursion -- the exception could be a very knotted-up
recursion with multiple cycles tied up together.
This plan is implemented in the Rec case of specBindItself.
Note [Specialisations already covered]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
We obviously don't want to generate two specialisations for the same
argument pattern. There are two wrinkles
1. We do the already-covered test in specDefn, not when we generate
the CallInfo in mkCallUDs. We used to test in the latter place, but
we now iterate the specialiser somewhat, and the Id at the call site
might therefore not have all the RULES that we can see in specDefn
2. What about two specialisations where the second is an *instance*
of the first? If the more specific one shows up first, we'll generate
specialisations for both. If the *less* specific one shows up first,
we *don't* currently generate a specialisation for the more specific
one. (See the call to lookupRule in already_covered.) Reasons:
(a) lookupRule doesn't say which matches are exact (bad reason)
(b) if the earlier specialisation is user-provided, it's
far from clear that we should auto-specialise further
Note [Auto-specialisation and RULES]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Consider:
g :: Num a => a -> a
g = ...
f :: (Int -> Int) -> Int
f w = ...
{-# RULE f g = 0 #-}
Suppose that auto-specialisation makes a specialised version of
g::Int->Int That version won't appear in the LHS of the RULE for f.
So if the specialisation rule fires too early, the rule for f may
never fire.
It might be possible to add new rules, to "complete" the rewrite system.
Thus when adding
RULE forall d. g Int d = g_spec
also add
RULE f g_spec = 0
But that's a bit complicated. For now we ask the programmer's help,
by *copying the INLINE activation pragma* to the auto-specialised
rule. So if g says {-# NOINLINE[2] g #-}, then the auto-spec rule
will also not be active until phase 2. And that's what programmers
should jolly well do anyway, even aside from specialisation, to ensure
that g doesn't inline too early.
This in turn means that the RULE would never fire for a NOINLINE
thing so not much point in generating a specialisation at all.
Note [Specialisation shape]
~~~~~~~~~~~~~~~~~~~~~~~~~~~
We only specialise a function if it has visible top-level lambdas
corresponding to its overloading. E.g. if
f :: forall a. Eq a => ....
then its body must look like
f = /\a. \d. ...
Reason: when specialising the body for a call (f ty dexp), we want to
substitute dexp for d, and pick up specialised calls in the body of f.
This doesn't always work. One example I came across was this:
newtype Gen a = MkGen{ unGen :: Int -> a }
choose :: Eq a => a -> Gen a
choose n = MkGen (\r -> n)
oneof = choose (1::Int)
It's a silly example, but we get
choose = /\a. g `cast` co
where choose doesn't have any dict arguments. Thus far I have not
tried to fix this (wait till there's a real example).
Mind you, then 'choose' will be inlined (since RHS is trivial) so
it doesn't matter. This comes up with single-method classes
class C a where { op :: a -> a }
instance C a => C [a] where ....
==>
$fCList :: C a => C [a]
$fCList = $copList |> (...coercion>...)
....(uses of $fCList at particular types)...
So we suppress the WARN if the rhs is trivial.
Note [Inline specialisations]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Here is what we do with the InlinePragma of the original function
* Activation/RuleMatchInfo: both transferred to the
specialised function
* InlineSpec:
(a) An INLINE pragma is transferred
(b) An INLINABLE pragma is *not* transferred
Why (a): transfer INLINE pragmas? The point of INLINE was precisely to
specialise the function at its call site, and arguably that's not so
important for the specialised copies. BUT *pragma-directed*
specialisation now takes place in the typechecker/desugarer, with
manually specified INLINEs. The specialisation here is automatic.
It'd be very odd if a function marked INLINE was specialised (because
of some local use), and then forever after (including importing
modules) the specialised version wasn't INLINEd. After all, the
programmer said INLINE!
You might wonder why we specialise INLINE functions at all. After
all they should be inlined, right? Two reasons:
* Even INLINE functions are sometimes not inlined, when they aren't
applied to interesting arguments. But perhaps the type arguments
alone are enough to specialise (even though the args are too boring
to trigger inlining), and it's certainly better to call the
specialised version.
* The RHS of an INLINE function might call another overloaded function,
and we'd like to generate a specialised version of that function too.
This actually happens a lot. Consider
replicateM_ :: (Monad m) => Int -> m a -> m ()
{-# INLINABLE replicateM_ #-}
replicateM_ d x ma = ...
The strictness analyser may transform to
replicateM_ :: (Monad m) => Int -> m a -> m ()
{-# INLINE replicateM_ #-}
replicateM_ d x ma = case x of I# x' -> $wreplicateM_ d x' ma
$wreplicateM_ :: (Monad m) => Int# -> m a -> m ()
{-# INLINABLE $wreplicateM_ #-}
$wreplicateM_ = ...
Now an importing module has a specialised call to replicateM_, say
(replicateM_ dMonadIO). We certainly want to specialise $wreplicateM_!
This particular example had a huge effect on the call to replicateM_
in nofib/shootout/n-body.
Why (b): discard INLINABLE pragmas? See #4874 for persuasive examples.
Suppose we have
{-# INLINABLE f #-}
f :: Ord a => [a] -> Int
f xs = letrec f' = ...f'... in f'
Then, when f is specialised and optimised we might get
wgo :: [Int] -> Int#
wgo = ...wgo...
f_spec :: [Int] -> Int
f_spec xs = case wgo xs of { r -> I# r }
and we clearly want to inline f_spec at call sites. But if we still
have the big, un-optimised of f (albeit specialised) captured in an
INLINABLE pragma for f_spec, we won't get that optimisation.
So we simply drop INLINABLE pragmas when specialising. It's not really
a complete solution; ignoring specialisation for now, INLINABLE functions
don't get properly strictness analysed, for example. But it works well
for examples involving specialisation, which is the dominant use of
INLINABLE. See #4874.
************************************************************************
* *
\subsubsection{UsageDetails and suchlike}
* *
************************************************************************
-}
data UsageDetails
= MkUD {
ud_binds :: !(Bag DictBind),
-- See Note [Floated dictionary bindings]
-- The order is important;
-- in ds1 `union` ds2, bindings in ds2 can depend on those in ds1
-- (Remember, Bags preserve order in GHC.)
ud_calls :: !CallDetails
-- INVARIANT: suppose bs = bindersOf ud_binds
-- Then 'calls' may *mention* 'bs',
-- but there should be no calls *for* bs
}
-- | A 'DictBind' is a binding along with a cached set containing its free
-- variables (both type variables and dictionaries)
type DictBind = (CoreBind, VarSet)
{- Note [Floated dictionary bindings]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
We float out dictionary bindings for the reasons described under
"Dictionary floating" above. But not /just/ dictionary bindings.
Consider
f :: Eq a => blah
f a d = rhs
$c== :: T -> T -> Bool
$c== x y = ...
$df :: Eq T
$df = Eq $c== ...
gurgle = ...(f @T $df)...
We gather the call info for (f @T $df), and we don't want to drop it
when we come across the binding for $df. So we add $df to the floats
and continue. But then we have to add $c== to the floats, and so on.
These all float above the binding for 'f', and now we can
successfully specialise 'f'.
So the DictBinds in (ud_binds :: Bag DictBind) may contain
non-dictionary bindings too.
-}
instance Outputable UsageDetails where
ppr (MkUD { ud_binds = dbs, ud_calls = calls })
= text "MkUD" <+> braces (sep (punctuate comma
[text "binds" <+> equals <+> ppr dbs,
text "calls" <+> equals <+> ppr calls]))
emptyUDs :: UsageDetails
emptyUDs = MkUD { ud_binds = emptyBag, ud_calls = emptyDVarEnv }
------------------------------------------------------------
type CallDetails = DIdEnv CallInfoSet
-- The order of specialized binds and rules depends on how we linearize
-- CallDetails, so to get determinism we must use a deterministic set here.
-- See Note [Deterministic UniqFM] in UniqDFM
data CallInfoSet = CIS Id (Bag CallInfo)
-- The list of types and dictionaries is guaranteed to
-- match the type of f
-- The Bag may contain duplicate calls (i.e. f @T and another f @T)
-- These dups are eliminated by already_covered in specCalls
data CallInfo
= CI { ci_key :: [SpecArg] -- All arguments
, ci_arity :: Int -- The number of variables necessary to bind
-- all of the specialised arguments
, ci_fvs :: VarSet -- Free vars of the ci_key
-- call (including tyvars)
-- [*not* include the main id itself, of course]
}
type DictExpr = CoreExpr
ciSetFilter :: (CallInfo -> Bool) -> CallInfoSet -> CallInfoSet
ciSetFilter p (CIS id a) = CIS id (filterBag p a)
instance Outputable CallInfoSet where
ppr (CIS fn map) = hang (text "CIS" <+> ppr fn)
2 (ppr map)
pprCallInfo :: Id -> CallInfo -> SDoc
pprCallInfo fn (CI { ci_key = key })
= ppr fn <+> ppr key
ppr_call_key_ty :: SpecArg -> Maybe SDoc
ppr_call_key_ty (SpecType ty) = Just $ char '@' <+> pprParendType ty
ppr_call_key_ty UnspecType = Just $ char '_'
ppr_call_key_ty (SpecDict _) = Nothing
ppr_call_key_ty UnspecArg = Nothing
instance Outputable CallInfo where
ppr (CI { ci_key = key, ci_fvs = fvs })
= text "CI" <> braces (hsep [ fsep (mapMaybe ppr_call_key_ty key), ppr fvs ])
unionCalls :: CallDetails -> CallDetails -> CallDetails
unionCalls c1 c2 = plusDVarEnv_C unionCallInfoSet c1 c2
unionCallInfoSet :: CallInfoSet -> CallInfoSet -> CallInfoSet
unionCallInfoSet (CIS f calls1) (CIS _ calls2) =
CIS f (calls1 `unionBags` calls2)
callDetailsFVs :: CallDetails -> VarSet
callDetailsFVs calls =
nonDetFoldUDFM (unionVarSet . callInfoFVs) emptyVarSet calls
-- It's OK to use nonDetFoldUDFM here because we forget the ordering
-- immediately by converting to a nondeterministic set.
callInfoFVs :: CallInfoSet -> VarSet
callInfoFVs (CIS _ call_info) =
foldr (\(CI { ci_fvs = fv }) vs -> unionVarSet fv vs) emptyVarSet call_info
computeArity :: [SpecArg] -> Int
computeArity = length . filter isValueArg . dropWhileEndLE isUnspecArg
callSpecArity :: [TyCoBinder] -> Int
callSpecArity = length . filter (not . isNamedBinder) . dropWhileEndLE isVisibleBinder
getTheta :: [TyCoBinder] -> [PredType]
getTheta = fmap tyBinderType . filter isInvisibleBinder . filter (not . isNamedBinder)
------------------------------------------------------------
singleCall :: Id -> [SpecArg] -> UsageDetails
singleCall id args
= MkUD {ud_binds = emptyBag,
ud_calls = unitDVarEnv id $ CIS id $
unitBag (CI { ci_key = args -- used to be tys
, ci_arity = computeArity args
, ci_fvs = call_fvs }) }
where
tys = getSpecTypes args
dicts = getSpecDicts args
call_fvs = exprsFreeVars dicts `unionVarSet` tys_fvs
tys_fvs = tyCoVarsOfTypes tys
-- The type args (tys) are guaranteed to be part of the dictionary
-- types, because they are just the constrained types,
-- and the dictionary is therefore sure to be bound
-- inside the binding for any type variables free in the type;
-- hence it's safe to neglect tyvars free in tys when making
-- the free-var set for this call
-- BUT I don't trust this reasoning; play safe and include tys_fvs
--
-- We don't include the 'id' itself.
mkCallUDs, mkCallUDs' :: SpecEnv -> Id -> [CoreExpr] -> UsageDetails
mkCallUDs env f args
= -- pprTrace "mkCallUDs" (vcat [ ppr f, ppr args, ppr res ])
res
where
res = mkCallUDs' env f args
mkCallUDs' env f args
| not (want_calls_for f) -- Imported from elsewhere
|| null theta -- Not overloaded
= emptyUDs
| not (all type_determines_value theta)
|| not (computeArity ci_key <= idArity f)
|| not (length dicts == length theta)
|| not (any (interestingDict env) dicts) -- Note [Interesting dictionary arguments]
-- See also Note [Specialisations already covered]
= -- pprTrace "mkCallUDs: discarding" _trace_doc
emptyUDs -- Not overloaded, or no specialisation wanted
| otherwise
= -- pprTrace "mkCallUDs: keeping" _trace_doc
singleCall f ci_key
where
_trace_doc = vcat [ppr f, ppr args, ppr (map (interestingDict env) dicts)]
pis = fst $ splitPiTys $ idType f
theta = getTheta pis
constrained_tyvars = tyCoVarsOfTypes theta
ci_key :: [SpecArg]
ci_key = fmap (\(t, a) ->
case t of
Named (binderVar -> tyVar)
| tyVar `elemVarSet` constrained_tyvars
-> case a of
Type ty -> SpecType ty
_ -> pprPanic "ci_key" $ ppr a
| otherwise
-> UnspecType
Anon InvisArg _ -> SpecDict a
Anon VisArg _ -> UnspecArg
) $ zip pis args
dicts = getSpecDicts ci_key
want_calls_for f = isLocalId f || isJust (maybeUnfoldingTemplate (realIdUnfolding f))
-- For imported things, we gather call instances if
-- there is an unfolding that we could in principle specialise
-- We might still decide not to use it (consulting dflags)
-- in specImports
-- Use 'realIdUnfolding' to ignore the loop-breaker flag!
type_determines_value pred -- See Note [Type determines value]
= case classifyPredType pred of
ClassPred cls _ -> not (isIPClass cls) -- Superclasses can't be IPs
EqPred {} -> True
IrredPred {} -> True -- Things like (D []) where D is a
-- Constraint-ranged family; #7785
ForAllPred {} -> True
{-
Note [Type determines value]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Only specialise if all overloading is on non-IP *class* params,
because these are the ones whose *type* determines their *value*. In
parrticular, with implicit params, the type args *don't* say what the
value of the implicit param is! See #7101
However, consider
type family D (v::*->*) :: Constraint
type instance D [] = ()
f :: D v => v Char -> Int
If we see a call (f "foo"), we'll pass a "dictionary"
() |> (g :: () ~ D [])
and it's good to specialise f at this dictionary.
So the question is: can an implicit parameter "hide inside" a
type-family constraint like (D a). Well, no. We don't allow
type instance D Maybe = ?x:Int
Hence the IrredPred case in type_determines_value.
See #7785.
Note [Interesting dictionary arguments]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Consider this
\a.\d:Eq a. let f = ... in ...(f d)...
There really is not much point in specialising f wrt the dictionary d,
because the code for the specialised f is not improved at all, because
d is lambda-bound. We simply get junk specialisations.
What is "interesting"? Just that it has *some* structure. But what about
variables?
* A variable might be imported, in which case its unfolding
will tell us whether it has useful structure
* Local variables are cloned on the way down (to avoid clashes when
we float dictionaries), and cloning drops the unfolding
(cloneIdBndr). Moreover, we make up some new bindings, and it's a
nuisance to give them unfoldings. So we keep track of the
"interesting" dictionaries as a VarSet in SpecEnv.
We have to take care to put any new interesting dictionary
bindings in the set.
We accidentally lost accurate tracking of local variables for a long
time, because cloned variables don't have unfoldings. But makes a
massive difference in a few cases, eg #5113. For nofib as a
whole it's only a small win: 2.2% improvement in allocation for ansi,
1.2% for bspt, but mostly 0.0! Average 0.1% increase in binary size.
-}
interestingDict :: SpecEnv -> CoreExpr -> Bool
-- A dictionary argument is interesting if it has *some* structure
-- NB: "dictionary" arguments include constraints of all sorts,
-- including equality constraints; hence the Coercion case
interestingDict env (Var v) = hasSomeUnfolding (idUnfolding v)
|| isDataConWorkId v
|| v `elemVarSet` se_interesting env
interestingDict _ (Type _) = False
interestingDict _ (Coercion _) = False
interestingDict env (App fn (Type _)) = interestingDict env fn
interestingDict env (App fn (Coercion _)) = interestingDict env fn
interestingDict env (Tick _ a) = interestingDict env a
interestingDict env (Cast e _) = interestingDict env e
interestingDict _ _ = True
plusUDs :: UsageDetails -> UsageDetails -> UsageDetails
plusUDs (MkUD {ud_binds = db1, ud_calls = calls1})
(MkUD {ud_binds = db2, ud_calls = calls2})
= MkUD { ud_binds = db1 `unionBags` db2
, ud_calls = calls1 `unionCalls` calls2 }
-----------------------------
_dictBindBndrs :: Bag DictBind -> [Id]
_dictBindBndrs dbs = foldr ((++) . bindersOf . fst) [] dbs
-- | Construct a 'DictBind' from a 'CoreBind'
mkDB :: CoreBind -> DictBind
mkDB bind = (bind, bind_fvs bind)
-- | Identify the free variables of a 'CoreBind'
bind_fvs :: CoreBind -> VarSet
bind_fvs (NonRec bndr rhs) = pair_fvs (bndr,rhs)
bind_fvs (Rec prs) = foldl' delVarSet rhs_fvs bndrs
where
bndrs = map fst prs
rhs_fvs = unionVarSets (map pair_fvs prs)
pair_fvs :: (Id, CoreExpr) -> VarSet
pair_fvs (bndr, rhs) = exprSomeFreeVars interesting rhs
`unionVarSet` idFreeVars bndr
-- idFreeVars: don't forget variables mentioned in
-- the rules of the bndr. C.f. OccAnal.addRuleUsage
-- Also tyvars mentioned in its type; they may not appear
-- in the RHS
-- type T a = Int
-- x :: T a = 3
where
interesting :: InterestingVarFun
interesting v = isLocalVar v || (isId v && isDFunId v)
-- Very important: include DFunIds /even/ if it is imported
-- Reason: See Note [Avoiding loops], the second exmaple
-- involving an imported dfun. We must know whether
-- a dictionary binding depends on an imported dfun,
-- in case we try to specialise that imported dfun
-- #13429 illustrates
-- | Flatten a set of "dumped" 'DictBind's, and some other binding
-- pairs, into a single recursive binding.
recWithDumpedDicts :: [(Id,CoreExpr)] -> Bag DictBind ->DictBind
recWithDumpedDicts pairs dbs
= (Rec bindings, fvs)
where
(bindings, fvs) = foldr add
([], emptyVarSet)
(dbs `snocBag` mkDB (Rec pairs))
add (NonRec b r, fvs') (pairs, fvs) =
((b,r) : pairs, fvs `unionVarSet` fvs')
add (Rec prs1, fvs') (pairs, fvs) =
(prs1 ++ pairs, fvs `unionVarSet` fvs')
snocDictBinds :: UsageDetails -> [DictBind] -> UsageDetails
-- Add ud_binds to the tail end of the bindings in uds
snocDictBinds uds dbs
= uds { ud_binds = ud_binds uds `unionBags` listToBag dbs }
consDictBind :: DictBind -> UsageDetails -> UsageDetails
consDictBind bind uds = uds { ud_binds = bind `consBag` ud_binds uds }
addDictBinds :: [DictBind] -> UsageDetails -> UsageDetails
addDictBinds binds uds = uds { ud_binds = listToBag binds `unionBags` ud_binds uds }
snocDictBind :: UsageDetails -> DictBind -> UsageDetails
snocDictBind uds bind = uds { ud_binds = ud_binds uds `snocBag` bind }
wrapDictBinds :: Bag DictBind -> [CoreBind] -> [CoreBind]
wrapDictBinds dbs binds
= foldr add binds dbs
where
add (bind,_) binds = bind : binds
wrapDictBindsE :: Bag DictBind -> CoreExpr -> CoreExpr
wrapDictBindsE dbs expr
= foldr add expr dbs
where
add (bind,_) expr = Let bind expr
----------------------
dumpUDs :: [CoreBndr] -> UsageDetails -> (UsageDetails, Bag DictBind)
-- Used at a lambda or case binder; just dump anything mentioning the binder
dumpUDs bndrs uds@(MkUD { ud_binds = orig_dbs, ud_calls = orig_calls })
| null bndrs = (uds, emptyBag) -- Common in case alternatives
| otherwise = -- pprTrace "dumpUDs" (ppr bndrs $$ ppr free_uds $$ ppr dump_dbs) $
(free_uds, dump_dbs)
where
free_uds = MkUD { ud_binds = free_dbs, ud_calls = free_calls }
bndr_set = mkVarSet bndrs
(free_dbs, dump_dbs, dump_set) = splitDictBinds orig_dbs bndr_set
free_calls = deleteCallsMentioning dump_set $ -- Drop calls mentioning bndr_set on the floor
deleteCallsFor bndrs orig_calls -- Discard calls for bndr_set; there should be
-- no calls for any of the dicts in dump_dbs
dumpBindUDs :: [CoreBndr] -> UsageDetails -> (UsageDetails, Bag DictBind, Bool)
-- Used at a let(rec) binding.
-- We return a boolean indicating whether the binding itself is mentioned,
-- directly or indirectly, by any of the ud_calls; in that case we want to
-- float the binding itself;
-- See Note [Floated dictionary bindings]
dumpBindUDs bndrs (MkUD { ud_binds = orig_dbs, ud_calls = orig_calls })
= -- pprTrace "dumpBindUDs" (ppr bndrs $$ ppr free_uds $$ ppr dump_dbs) $
(free_uds, dump_dbs, float_all)
where
free_uds = MkUD { ud_binds = free_dbs, ud_calls = free_calls }
bndr_set = mkVarSet bndrs
(free_dbs, dump_dbs, dump_set) = splitDictBinds orig_dbs bndr_set
free_calls = deleteCallsFor bndrs orig_calls
float_all = dump_set `intersectsVarSet` callDetailsFVs free_calls
callsForMe :: Id -> UsageDetails -> (UsageDetails, [CallInfo])
callsForMe fn (MkUD { ud_binds = orig_dbs, ud_calls = orig_calls })
= -- pprTrace ("callsForMe")
-- (vcat [ppr fn,
-- text "Orig dbs =" <+> ppr (_dictBindBndrs orig_dbs),
-- text "Orig calls =" <+> ppr orig_calls,
-- text "Dep set =" <+> ppr dep_set,
-- text "Calls for me =" <+> ppr calls_for_me]) $
(uds_without_me, calls_for_me)
where
uds_without_me = MkUD { ud_binds = orig_dbs
, ud_calls = delDVarEnv orig_calls fn }
calls_for_me = case lookupDVarEnv orig_calls fn of
Nothing -> []
Just cis -> filterCalls cis orig_dbs
-- filterCalls: drop calls that (directly or indirectly)
-- refer to fn. See Note [Avoiding loops]
----------------------
filterCalls :: CallInfoSet -> Bag DictBind -> [CallInfo]
-- See Note [Avoiding loops]
filterCalls (CIS fn call_bag) dbs
= filter ok_call (bagToList call_bag)
where
dump_set = foldl' go (unitVarSet fn) dbs
-- This dump-set could also be computed by splitDictBinds
-- (_,_,dump_set) = splitDictBinds dbs {fn}
-- But this variant is shorter
go so_far (db,fvs) | fvs `intersectsVarSet` so_far
= extendVarSetList so_far (bindersOf db)
| otherwise = so_far
ok_call (CI { ci_fvs = fvs }) = not (fvs `intersectsVarSet` dump_set)
----------------------
splitDictBinds :: Bag DictBind -> IdSet -> (Bag DictBind, Bag DictBind, IdSet)
-- splitDictBinds dbs bndrs returns
-- (free_dbs, dump_dbs, dump_set)
-- where
-- * dump_dbs depends, transitively on bndrs
-- * free_dbs does not depend on bndrs
-- * dump_set = bndrs `union` bndrs(dump_dbs)
splitDictBinds dbs bndr_set
= foldl' split_db (emptyBag, emptyBag, bndr_set) dbs
-- Important that it's foldl' not foldr;
-- we're accumulating the set of dumped ids in dump_set
where
split_db (free_dbs, dump_dbs, dump_idset) db@(bind, fvs)
| dump_idset `intersectsVarSet` fvs -- Dump it
= (free_dbs, dump_dbs `snocBag` db,
extendVarSetList dump_idset (bindersOf bind))
| otherwise -- Don't dump it
= (free_dbs `snocBag` db, dump_dbs, dump_idset)
----------------------
deleteCallsMentioning :: VarSet -> CallDetails -> CallDetails
-- Remove calls *mentioning* bs in any way
deleteCallsMentioning bs calls
= mapDVarEnv (ciSetFilter keep_call) calls
where
keep_call (CI { ci_fvs = fvs }) = not (fvs `intersectsVarSet` bs)
deleteCallsFor :: [Id] -> CallDetails -> CallDetails
-- Remove calls *for* bs
deleteCallsFor bs calls = delDVarEnvList calls bs
{-
************************************************************************
* *
\subsubsection{Boring helper functions}
* *
************************************************************************
-}
newtype SpecM a = SpecM (State SpecState a) deriving (Functor)
data SpecState = SpecState {
spec_uniq_supply :: UniqSupply,
spec_module :: Module,
spec_dflags :: DynFlags
}
instance Applicative SpecM where
pure x = SpecM $ return x
(<*>) = ap
instance Monad SpecM where
SpecM x >>= f = SpecM $ do y <- x
case f y of
SpecM z ->
z
#if !MIN_VERSION_base(4,13,0)
fail = MonadFail.fail
#endif
instance MonadFail.MonadFail SpecM where
fail str = SpecM $ error str
instance MonadUnique SpecM where
getUniqueSupplyM
= SpecM $ do st <- get
let (us1, us2) = splitUniqSupply $ spec_uniq_supply st
put $ st { spec_uniq_supply = us2 }
return us1
getUniqueM
= SpecM $ do st <- get
let (u,us') = takeUniqFromSupply $ spec_uniq_supply st
put $ st { spec_uniq_supply = us' }
return u
instance HasDynFlags SpecM where
getDynFlags = SpecM $ liftM spec_dflags get
instance HasModule SpecM where
getModule = SpecM $ liftM spec_module get
runSpecM :: DynFlags -> Module -> SpecM a -> CoreM a
runSpecM dflags this_mod (SpecM spec)
= do us <- getUniqueSupplyM
let initialState = SpecState {
spec_uniq_supply = us,
spec_module = this_mod,
spec_dflags = dflags
}
return $ evalState spec initialState
mapAndCombineSM :: (a -> SpecM (b, UsageDetails)) -> [a] -> SpecM ([b], UsageDetails)
mapAndCombineSM _ [] = return ([], emptyUDs)
mapAndCombineSM f (x:xs) = do (y, uds1) <- f x
(ys, uds2) <- mapAndCombineSM f xs
return (y:ys, uds1 `plusUDs` uds2)
extendTvSubstList :: SpecEnv -> [(TyVar,Type)] -> SpecEnv
extendTvSubstList env tv_binds
= env { se_subst = CoreSubst.extendTvSubstList (se_subst env) tv_binds }
substTy :: SpecEnv -> Type -> Type
substTy env ty = CoreSubst.substTy (se_subst env) ty
substCo :: SpecEnv -> Coercion -> Coercion
substCo env co = CoreSubst.substCo (se_subst env) co
substBndr :: SpecEnv -> CoreBndr -> (SpecEnv, CoreBndr)
substBndr env bs = case CoreSubst.substBndr (se_subst env) bs of
(subst', bs') -> (env { se_subst = subst' }, bs')
substBndrs :: SpecEnv -> [CoreBndr] -> (SpecEnv, [CoreBndr])
substBndrs env bs = case CoreSubst.substBndrs (se_subst env) bs of
(subst', bs') -> (env { se_subst = subst' }, bs')
cloneBindSM :: SpecEnv -> CoreBind -> SpecM (SpecEnv, SpecEnv, CoreBind)
-- Clone the binders of the bind; return new bind with the cloned binders
-- Return the substitution to use for RHSs, and the one to use for the body
cloneBindSM env@(SE { se_subst = subst, se_interesting = interesting }) (NonRec bndr rhs)
= do { us <- getUniqueSupplyM
; let (subst', bndr') = CoreSubst.cloneIdBndr subst us bndr
interesting' | interestingDict env rhs
= interesting `extendVarSet` bndr'
| otherwise = interesting
; return (env, env { se_subst = subst', se_interesting = interesting' }
, NonRec bndr' rhs) }
cloneBindSM env@(SE { se_subst = subst, se_interesting = interesting }) (Rec pairs)
= do { us <- getUniqueSupplyM
; let (subst', bndrs') = CoreSubst.cloneRecIdBndrs subst us (map fst pairs)
env' = env { se_subst = subst'
, se_interesting = interesting `extendVarSetList`
[ v | (v,r) <- pairs, interestingDict env r ] }
; return (env', env', Rec (bndrs' `zip` map snd pairs)) }
newDictBndr :: SpecEnv -> CoreBndr -> SpecM CoreBndr
-- Make up completely fresh binders for the dictionaries
-- Their bindings are going to float outwards
newDictBndr env b = do { uniq <- getUniqueM
; let n = idName b
ty' = substTy env (idType b)
; return (mkUserLocalOrCoVar (nameOccName n) uniq ty' (getSrcSpan n)) }
newSpecIdSM :: Id -> Type -> Maybe JoinArity -> SpecM Id
-- Give the new Id a similar occurrence name to the old one
newSpecIdSM old_id new_ty join_arity_maybe
= do { uniq <- getUniqueM
; let name = idName old_id
new_occ = mkSpecOcc (nameOccName name)
new_id = mkUserLocalOrCoVar new_occ uniq new_ty (getSrcSpan name)
`asJoinId_maybe` join_arity_maybe
; return new_id }
{-
Old (but interesting) stuff about unboxed bindings
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
What should we do when a value is specialised to a *strict* unboxed value?
map_*_* f (x:xs) = let h = f x
t = map f xs
in h:t
Could convert let to case:
map_*_Int# f (x:xs) = case f x of h# ->
let t = map f xs
in h#:t
This may be undesirable since it forces evaluation here, but the value
may not be used in all branches of the body. In the general case this
transformation is impossible since the mutual recursion in a letrec
cannot be expressed as a case.
There is also a problem with top-level unboxed values, since our
implementation cannot handle unboxed values at the top level.
Solution: Lift the binding of the unboxed value and extract it when it
is used:
map_*_Int# f (x:xs) = let h = case (f x) of h# -> _Lift h#
t = map f xs
in case h of
_Lift h# -> h#:t
Now give it to the simplifier and the _Lifting will be optimised away.
The benefit is that we have given the specialised "unboxed" values a
very simple lifted semantics and then leave it up to the simplifier to
optimise it --- knowing that the overheads will be removed in nearly
all cases.
In particular, the value will only be evaluated in the branches of the
program which use it, rather than being forced at the point where the
value is bound. For example:
filtermap_*_* p f (x:xs)
= let h = f x
t = ...
in case p x of
True -> h:t
False -> t
==>
filtermap_*_Int# p f (x:xs)
= let h = case (f x) of h# -> _Lift h#
t = ...
in case p x of
True -> case h of _Lift h#
-> h#:t
False -> t
The binding for h can still be inlined in the one branch and the
_Lifting eliminated.
Question: When won't the _Lifting be eliminated?
Answer: When they at the top-level (where it is necessary) or when
inlining would duplicate work (or possibly code depending on
options). However, the _Lifting will still be eliminated if the
strictness analyser deems the lifted binding strict.
-}
|
