aboutsummaryrefslogtreecommitdiff
path: root/gcc/local-alloc.c
blob: 2d7e32a5d1fc0f2866e83a4623ef81b87ee2b21e (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
2049
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
2093
2094
2095
2096
2097
2098
2099
2100
2101
2102
2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
2150
2151
2152
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
2170
2171
2172
2173
2174
2175
2176
2177
2178
2179
2180
2181
2182
2183
2184
2185
2186
2187
2188
2189
2190
2191
2192
2193
2194
2195
2196
2197
2198
2199
2200
2201
2202
2203
2204
2205
2206
2207
2208
2209
2210
2211
2212
2213
2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
2228
2229
2230
2231
2232
2233
2234
2235
2236
2237
2238
2239
2240
2241
2242
2243
2244
2245
2246
2247
2248
2249
2250
2251
2252
2253
2254
2255
2256
2257
2258
2259
2260
2261
2262
2263
2264
2265
2266
2267
2268
2269
2270
2271
2272
2273
2274
2275
2276
2277
2278
2279
2280
2281
2282
2283
2284
2285
2286
2287
2288
2289
2290
2291
2292
2293
2294
2295
2296
2297
2298
2299
2300
2301
2302
2303
2304
2305
2306
2307
2308
2309
2310
2311
2312
2313
2314
2315
2316
2317
2318
2319
2320
2321
2322
2323
2324
2325
2326
2327
2328
2329
2330
2331
2332
2333
2334
2335
2336
2337
2338
2339
2340
2341
2342
2343
2344
2345
2346
2347
2348
2349
2350
2351
2352
2353
2354
2355
2356
2357
2358
2359
2360
2361
2362
2363
2364
2365
2366
2367
2368
2369
2370
2371
2372
2373
2374
2375
2376
2377
2378
2379
2380
2381
2382
2383
2384
2385
2386
2387
2388
2389
2390
2391
2392
2393
2394
2395
2396
2397
2398
2399
2400
2401
2402
2403
2404
2405
2406
2407
2408
2409
2410
2411
2412
2413
2414
2415
2416
2417
2418
2419
2420
2421
2422
2423
2424
2425
2426
2427
2428
2429
2430
2431
2432
2433
2434
2435
2436
/* Allocate registers within a basic block, for GNU compiler.
   Copyright (C) 1987, 1988, 1991, 1993, 1994, 1995, 1996, 1997, 1998,
   1999, 2000, 2001, 2002, 2003, 2004 Free Software Foundation, Inc.

This file is part of GCC.

GCC is free software; you can redistribute it and/or modify it under
the terms of the GNU General Public License as published by the Free
Software Foundation; either version 2, or (at your option) any later
version.

GCC is distributed in the hope that it will be useful, but WITHOUT ANY
WARRANTY; without even the implied warranty of MERCHANTABILITY or
FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
for more details.

You should have received a copy of the GNU General Public License
along with GCC; see the file COPYING.  If not, write to the Free
Software Foundation, 59 Temple Place - Suite 330, Boston, MA
02111-1307, USA.  */

/* Allocation of hard register numbers to pseudo registers is done in
   two passes.  In this pass we consider only regs that are born and
   die once within one basic block.  We do this one basic block at a
   time.  Then the next pass allocates the registers that remain.
   Two passes are used because this pass uses methods that work only
   on linear code, but that do a better job than the general methods
   used in global_alloc, and more quickly too.

   The assignments made are recorded in the vector reg_renumber
   whose space is allocated here.  The rtl code itself is not altered.

   We assign each instruction in the basic block a number
   which is its order from the beginning of the block.
   Then we can represent the lifetime of a pseudo register with
   a pair of numbers, and check for conflicts easily.
   We can record the availability of hard registers with a
   HARD_REG_SET for each instruction.  The HARD_REG_SET
   contains 0 or 1 for each hard reg.

   To avoid register shuffling, we tie registers together when one
   dies by being copied into another, or dies in an instruction that
   does arithmetic to produce another.  The tied registers are
   allocated as one.  Registers with different reg class preferences
   can never be tied unless the class preferred by one is a subclass
   of the one preferred by the other.

   Tying is represented with "quantity numbers".
   A non-tied register is given a new quantity number.
   Tied registers have the same quantity number.

   We have provision to exempt registers, even when they are contained
   within the block, that can be tied to others that are not contained in it.
   This is so that global_alloc could process them both and tie them then.
   But this is currently disabled since tying in global_alloc is not
   yet implemented.  */

/* Pseudos allocated here can be reallocated by global.c if the hard register
   is used as a spill register.  Currently we don't allocate such pseudos
   here if their preferred class is likely to be used by spills.  */

#include "config.h"
#include "system.h"
#include "coretypes.h"
#include "tm.h"
#include "hard-reg-set.h"
#include "rtl.h"
#include "tm_p.h"
#include "flags.h"
#include "basic-block.h"
#include "regs.h"
#include "function.h"
#include "insn-config.h"
#include "insn-attr.h"
#include "recog.h"
#include "output.h"
#include "toplev.h"
#include "except.h"
#include "integrate.h"

/* Next quantity number available for allocation.  */

static int next_qty;

/* Information we maintain about each quantity.  */
struct qty
{
  /* The number of refs to quantity Q.  */

  int n_refs;

  /* The frequency of uses of quantity Q.  */

  int freq;

  /* Insn number (counting from head of basic block)
     where quantity Q was born.  -1 if birth has not been recorded.  */

  int birth;

  /* Insn number (counting from head of basic block)
     where given quantity died.  Due to the way tying is done,
     and the fact that we consider in this pass only regs that die but once,
     a quantity can die only once.  Each quantity's life span
     is a set of consecutive insns.  -1 if death has not been recorded.  */

  int death;

  /* Number of words needed to hold the data in given quantity.
     This depends on its machine mode.  It is used for these purposes:
     1. It is used in computing the relative importance of qtys,
	which determines the order in which we look for regs for them.
     2. It is used in rules that prevent tying several registers of
	different sizes in a way that is geometrically impossible
	(see combine_regs).  */

  int size;

  /* Number of times a reg tied to given qty lives across a CALL_INSN.  */

  int n_calls_crossed;

  /* The register number of one pseudo register whose reg_qty value is Q.
     This register should be the head of the chain
     maintained in reg_next_in_qty.  */

  int first_reg;

  /* Reg class contained in (smaller than) the preferred classes of all
     the pseudo regs that are tied in given quantity.
     This is the preferred class for allocating that quantity.  */

  enum reg_class min_class;

  /* Register class within which we allocate given qty if we can't get
     its preferred class.  */

  enum reg_class alternate_class;

  /* This holds the mode of the registers that are tied to given qty,
     or VOIDmode if registers with differing modes are tied together.  */

  enum machine_mode mode;

  /* the hard reg number chosen for given quantity,
     or -1 if none was found.  */

  short phys_reg;
};

static struct qty *qty;

/* These fields are kept separately to speedup their clearing.  */

/* We maintain two hard register sets that indicate suggested hard registers
   for each quantity.  The first, phys_copy_sugg, contains hard registers
   that are tied to the quantity by a simple copy.  The second contains all
   hard registers that are tied to the quantity via an arithmetic operation.

   The former register set is given priority for allocation.  This tends to
   eliminate copy insns.  */

/* Element Q is a set of hard registers that are suggested for quantity Q by
   copy insns.  */

static HARD_REG_SET *qty_phys_copy_sugg;

/* Element Q is a set of hard registers that are suggested for quantity Q by
   arithmetic insns.  */

static HARD_REG_SET *qty_phys_sugg;

/* Element Q is the number of suggested registers in qty_phys_copy_sugg.  */

static short *qty_phys_num_copy_sugg;

/* Element Q is the number of suggested registers in qty_phys_sugg.  */

static short *qty_phys_num_sugg;

/* If (REG N) has been assigned a quantity number, is a register number
   of another register assigned the same quantity number, or -1 for the
   end of the chain.  qty->first_reg point to the head of this chain.  */

static int *reg_next_in_qty;

/* reg_qty[N] (where N is a pseudo reg number) is the qty number of that reg
   if it is >= 0,
   of -1 if this register cannot be allocated by local-alloc,
   or -2 if not known yet.

   Note that if we see a use or death of pseudo register N with
   reg_qty[N] == -2, register N must be local to the current block.  If
   it were used in more than one block, we would have reg_qty[N] == -1.
   This relies on the fact that if reg_basic_block[N] is >= 0, register N
   will not appear in any other block.  We save a considerable number of
   tests by exploiting this.

   If N is < FIRST_PSEUDO_REGISTER, reg_qty[N] is undefined and should not
   be referenced.  */

static int *reg_qty;

/* The offset (in words) of register N within its quantity.
   This can be nonzero if register N is SImode, and has been tied
   to a subreg of a DImode register.  */

static char *reg_offset;

/* Vector of substitutions of register numbers,
   used to map pseudo regs into hardware regs.
   This is set up as a result of register allocation.
   Element N is the hard reg assigned to pseudo reg N,
   or is -1 if no hard reg was assigned.
   If N is a hard reg number, element N is N.  */

short *reg_renumber;

/* Set of hard registers live at the current point in the scan
   of the instructions in a basic block.  */

static HARD_REG_SET regs_live;

/* Each set of hard registers indicates registers live at a particular
   point in the basic block.  For N even, regs_live_at[N] says which
   hard registers are needed *after* insn N/2 (i.e., they may not
   conflict with the outputs of insn N/2 or the inputs of insn N/2 + 1.

   If an object is to conflict with the inputs of insn J but not the
   outputs of insn J + 1, we say it is born at index J*2 - 1.  Similarly,
   if it is to conflict with the outputs of insn J but not the inputs of
   insn J + 1, it is said to die at index J*2 + 1.  */

static HARD_REG_SET *regs_live_at;

/* Communicate local vars `insn_number' and `insn'
   from `block_alloc' to `reg_is_set', `wipe_dead_reg', and `alloc_qty'.  */
static int this_insn_number;
static rtx this_insn;

struct equivalence
{
  /* Set when an attempt should be made to replace a register
     with the associated src_p entry.  */

  char replace;

  /* Set when a REG_EQUIV note is found or created.  Use to
     keep track of what memory accesses might be created later,
     e.g. by reload.  */

  rtx replacement;

  rtx *src_p;

  /* Loop depth is used to recognize equivalences which appear
     to be present within the same loop (or in an inner loop).  */

  int loop_depth;

  /* The list of each instruction which initializes this register.  */

  rtx init_insns;
};

/* reg_equiv[N] (where N is a pseudo reg number) is the equivalence
   structure for that register.  */

static struct equivalence *reg_equiv;

/* Nonzero if we recorded an equivalence for a LABEL_REF.  */
static int recorded_label_ref;

static void alloc_qty (int, enum machine_mode, int, int);
static void validate_equiv_mem_from_store (rtx, rtx, void *);
static int validate_equiv_mem (rtx, rtx, rtx);
static int equiv_init_varies_p (rtx);
static int equiv_init_movable_p (rtx, int);
static int contains_replace_regs (rtx);
static int memref_referenced_p (rtx, rtx);
static int memref_used_between_p (rtx, rtx, rtx);
static void update_equiv_regs (void);
static void no_equiv (rtx, rtx, void *);
static void block_alloc (int);
static int qty_sugg_compare (int, int);
static int qty_sugg_compare_1 (const void *, const void *);
static int qty_compare (int, int);
static int qty_compare_1 (const void *, const void *);
static int combine_regs (rtx, rtx, int, int, rtx, int);
static int reg_meets_class_p (int, enum reg_class);
static void update_qty_class (int, int);
static void reg_is_set (rtx, rtx, void *);
static void reg_is_born (rtx, int);
static void wipe_dead_reg (rtx, int);
static int find_free_reg (enum reg_class, enum machine_mode, int, int, int,
			  int, int);
static void mark_life (int, enum machine_mode, int);
static void post_mark_life (int, enum machine_mode, int, int, int);
static int no_conflict_p (rtx, rtx, rtx);
static int requires_inout (const char *);

/* Allocate a new quantity (new within current basic block)
   for register number REGNO which is born at index BIRTH
   within the block.  MODE and SIZE are info on reg REGNO.  */

static void
alloc_qty (int regno, enum machine_mode mode, int size, int birth)
{
  int qtyno = next_qty++;

  reg_qty[regno] = qtyno;
  reg_offset[regno] = 0;
  reg_next_in_qty[regno] = -1;

  qty[qtyno].first_reg = regno;
  qty[qtyno].size = size;
  qty[qtyno].mode = mode;
  qty[qtyno].birth = birth;
  qty[qtyno].n_calls_crossed = REG_N_CALLS_CROSSED (regno);
  qty[qtyno].min_class = reg_preferred_class (regno);
  qty[qtyno].alternate_class = reg_alternate_class (regno);
  qty[qtyno].n_refs = REG_N_REFS (regno);
  qty[qtyno].freq = REG_FREQ (regno);
}

/* Main entry point of this file.  */

int
local_alloc (void)
{
  int i;
  int max_qty;
  basic_block b;

  /* We need to keep track of whether or not we recorded a LABEL_REF so
     that we know if the jump optimizer needs to be rerun.  */
  recorded_label_ref = 0;

  /* Leaf functions and non-leaf functions have different needs.
     If defined, let the machine say what kind of ordering we
     should use.  */
#ifdef ORDER_REGS_FOR_LOCAL_ALLOC
  ORDER_REGS_FOR_LOCAL_ALLOC;
#endif

  /* Promote REG_EQUAL notes to REG_EQUIV notes and adjust status of affected
     registers.  */
  if (optimize)
    update_equiv_regs ();

  /* This sets the maximum number of quantities we can have.  Quantity
     numbers start at zero and we can have one for each pseudo.  */
  max_qty = (max_regno - FIRST_PSEUDO_REGISTER);

  /* Allocate vectors of temporary data.
     See the declarations of these variables, above,
     for what they mean.  */

  qty = xmalloc (max_qty * sizeof (struct qty));
  qty_phys_copy_sugg = xmalloc (max_qty * sizeof (HARD_REG_SET));
  qty_phys_num_copy_sugg = xmalloc (max_qty * sizeof (short));
  qty_phys_sugg = xmalloc (max_qty * sizeof (HARD_REG_SET));
  qty_phys_num_sugg = xmalloc (max_qty * sizeof (short));

  reg_qty = xmalloc (max_regno * sizeof (int));
  reg_offset = xmalloc (max_regno * sizeof (char));
  reg_next_in_qty = xmalloc (max_regno * sizeof (int));

  /* Determine which pseudo-registers can be allocated by local-alloc.
     In general, these are the registers used only in a single block and
     which only die once.

     We need not be concerned with which block actually uses the register
     since we will never see it outside that block.  */

  for (i = FIRST_PSEUDO_REGISTER; i < max_regno; i++)
    {
      if (REG_BASIC_BLOCK (i) >= 0 && REG_N_DEATHS (i) == 1)
	reg_qty[i] = -2;
      else
	reg_qty[i] = -1;
    }

  /* Force loop below to initialize entire quantity array.  */
  next_qty = max_qty;

  /* Allocate each block's local registers, block by block.  */

  FOR_EACH_BB (b)
    {
      /* NEXT_QTY indicates which elements of the `qty_...'
	 vectors might need to be initialized because they were used
	 for the previous block; it is set to the entire array before
	 block 0.  Initialize those, with explicit loop if there are few,
	 else with bzero and bcopy.  Do not initialize vectors that are
	 explicit set by `alloc_qty'.  */

      if (next_qty < 6)
	{
	  for (i = 0; i < next_qty; i++)
	    {
	      CLEAR_HARD_REG_SET (qty_phys_copy_sugg[i]);
	      qty_phys_num_copy_sugg[i] = 0;
	      CLEAR_HARD_REG_SET (qty_phys_sugg[i]);
	      qty_phys_num_sugg[i] = 0;
	    }
	}
      else
	{
#define CLEAR(vector)  \
	  memset ((vector), 0, (sizeof (*(vector))) * next_qty);

	  CLEAR (qty_phys_copy_sugg);
	  CLEAR (qty_phys_num_copy_sugg);
	  CLEAR (qty_phys_sugg);
	  CLEAR (qty_phys_num_sugg);
	}

      next_qty = 0;

      block_alloc (b->index);
    }

  free (qty);
  free (qty_phys_copy_sugg);
  free (qty_phys_num_copy_sugg);
  free (qty_phys_sugg);
  free (qty_phys_num_sugg);

  free (reg_qty);
  free (reg_offset);
  free (reg_next_in_qty);

  return recorded_label_ref;
}

/* Used for communication between the following two functions: contains
   a MEM that we wish to ensure remains unchanged.  */
static rtx equiv_mem;

/* Set nonzero if EQUIV_MEM is modified.  */
static int equiv_mem_modified;

/* If EQUIV_MEM is modified by modifying DEST, indicate that it is modified.
   Called via note_stores.  */

static void
validate_equiv_mem_from_store (rtx dest, rtx set ATTRIBUTE_UNUSED,
			       void *data ATTRIBUTE_UNUSED)
{
  if ((GET_CODE (dest) == REG
       && reg_overlap_mentioned_p (dest, equiv_mem))
      || (GET_CODE (dest) == MEM
	  && true_dependence (dest, VOIDmode, equiv_mem, rtx_varies_p)))
    equiv_mem_modified = 1;
}

/* Verify that no store between START and the death of REG invalidates
   MEMREF.  MEMREF is invalidated by modifying a register used in MEMREF,
   by storing into an overlapping memory location, or with a non-const
   CALL_INSN.

   Return 1 if MEMREF remains valid.  */

static int
validate_equiv_mem (rtx start, rtx reg, rtx memref)
{
  rtx insn;
  rtx note;

  equiv_mem = memref;
  equiv_mem_modified = 0;

  /* If the memory reference has side effects or is volatile, it isn't a
     valid equivalence.  */
  if (side_effects_p (memref))
    return 0;

  for (insn = start; insn && ! equiv_mem_modified; insn = NEXT_INSN (insn))
    {
      if (! INSN_P (insn))
	continue;

      if (find_reg_note (insn, REG_DEAD, reg))
	return 1;

      if (GET_CODE (insn) == CALL_INSN && ! RTX_UNCHANGING_P (memref)
	  && ! CONST_OR_PURE_CALL_P (insn))
	return 0;

      note_stores (PATTERN (insn), validate_equiv_mem_from_store, NULL);

      /* If a register mentioned in MEMREF is modified via an
	 auto-increment, we lose the equivalence.  Do the same if one
	 dies; although we could extend the life, it doesn't seem worth
	 the trouble.  */

      for (note = REG_NOTES (insn); note; note = XEXP (note, 1))
	if ((REG_NOTE_KIND (note) == REG_INC
	     || REG_NOTE_KIND (note) == REG_DEAD)
	    && GET_CODE (XEXP (note, 0)) == REG
	    && reg_overlap_mentioned_p (XEXP (note, 0), memref))
	  return 0;
    }

  return 0;
}

/* Returns zero if X is known to be invariant.  */

static int
equiv_init_varies_p (rtx x)
{
  RTX_CODE code = GET_CODE (x);
  int i;
  const char *fmt;

  switch (code)
    {
    case MEM:
      return ! RTX_UNCHANGING_P (x) || equiv_init_varies_p (XEXP (x, 0));

    case QUEUED:
      return 1;

    case CONST:
    case CONST_INT:
    case CONST_DOUBLE:
    case CONST_VECTOR:
    case SYMBOL_REF:
    case LABEL_REF:
      return 0;

    case REG:
      return reg_equiv[REGNO (x)].replace == 0 && rtx_varies_p (x, 0);

    case ASM_OPERANDS:
      if (MEM_VOLATILE_P (x))
	return 1;

      /* Fall through.  */

    default:
      break;
    }

  fmt = GET_RTX_FORMAT (code);
  for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
    if (fmt[i] == 'e')
      {
	if (equiv_init_varies_p (XEXP (x, i)))
	  return 1;
      }
    else if (fmt[i] == 'E')
      {
	int j;
	for (j = 0; j < XVECLEN (x, i); j++)
	  if (equiv_init_varies_p (XVECEXP (x, i, j)))
	    return 1;
      }

  return 0;
}

/* Returns nonzero if X (used to initialize register REGNO) is movable.
   X is only movable if the registers it uses have equivalent initializations
   which appear to be within the same loop (or in an inner loop) and movable
   or if they are not candidates for local_alloc and don't vary.  */

static int
equiv_init_movable_p (rtx x, int regno)
{
  int i, j;
  const char *fmt;
  enum rtx_code code = GET_CODE (x);

  switch (code)
    {
    case SET:
      return equiv_init_movable_p (SET_SRC (x), regno);

    case CC0:
    case CLOBBER:
      return 0;

    case PRE_INC:
    case PRE_DEC:
    case POST_INC:
    case POST_DEC:
    case PRE_MODIFY:
    case POST_MODIFY:
      return 0;

    case REG:
      return (reg_equiv[REGNO (x)].loop_depth >= reg_equiv[regno].loop_depth
	      && reg_equiv[REGNO (x)].replace)
	     || (REG_BASIC_BLOCK (REGNO (x)) < 0 && ! rtx_varies_p (x, 0));

    case UNSPEC_VOLATILE:
      return 0;

    case ASM_OPERANDS:
      if (MEM_VOLATILE_P (x))
	return 0;

      /* Fall through.  */

    default:
      break;
    }

  fmt = GET_RTX_FORMAT (code);
  for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
    switch (fmt[i])
      {
      case 'e':
	if (! equiv_init_movable_p (XEXP (x, i), regno))
	  return 0;
	break;
      case 'E':
	for (j = XVECLEN (x, i) - 1; j >= 0; j--)
	  if (! equiv_init_movable_p (XVECEXP (x, i, j), regno))
	    return 0;
	break;
      }

  return 1;
}

/* TRUE if X uses any registers for which reg_equiv[REGNO].replace is true.  */

static int
contains_replace_regs (rtx x)
{
  int i, j;
  const char *fmt;
  enum rtx_code code = GET_CODE (x);

  switch (code)
    {
    case CONST_INT:
    case CONST:
    case LABEL_REF:
    case SYMBOL_REF:
    case CONST_DOUBLE:
    case CONST_VECTOR:
    case PC:
    case CC0:
    case HIGH:
      return 0;

    case REG:
      return reg_equiv[REGNO (x)].replace;

    default:
      break;
    }

  fmt = GET_RTX_FORMAT (code);
  for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
    switch (fmt[i])
      {
      case 'e':
	if (contains_replace_regs (XEXP (x, i)))
	  return 1;
	break;
      case 'E':
	for (j = XVECLEN (x, i) - 1; j >= 0; j--)
	  if (contains_replace_regs (XVECEXP (x, i, j)))
	    return 1;
	break;
      }

  return 0;
}

/* TRUE if X references a memory location that would be affected by a store
   to MEMREF.  */

static int
memref_referenced_p (rtx memref, rtx x)
{
  int i, j;
  const char *fmt;
  enum rtx_code code = GET_CODE (x);

  switch (code)
    {
    case CONST_INT:
    case CONST:
    case LABEL_REF:
    case SYMBOL_REF:
    case CONST_DOUBLE:
    case CONST_VECTOR:
    case PC:
    case CC0:
    case HIGH:
    case LO_SUM:
      return 0;

    case REG:
      return (reg_equiv[REGNO (x)].replacement
	      && memref_referenced_p (memref,
				      reg_equiv[REGNO (x)].replacement));

    case MEM:
      if (true_dependence (memref, VOIDmode, x, rtx_varies_p))
	return 1;
      break;

    case SET:
      /* If we are setting a MEM, it doesn't count (its address does), but any
	 other SET_DEST that has a MEM in it is referencing the MEM.  */
      if (GET_CODE (SET_DEST (x)) == MEM)
	{
	  if (memref_referenced_p (memref, XEXP (SET_DEST (x), 0)))
	    return 1;
	}
      else if (memref_referenced_p (memref, SET_DEST (x)))
	return 1;

      return memref_referenced_p (memref, SET_SRC (x));

    default:
      break;
    }

  fmt = GET_RTX_FORMAT (code);
  for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
    switch (fmt[i])
      {
      case 'e':
	if (memref_referenced_p (memref, XEXP (x, i)))
	  return 1;
	break;
      case 'E':
	for (j = XVECLEN (x, i) - 1; j >= 0; j--)
	  if (memref_referenced_p (memref, XVECEXP (x, i, j)))
	    return 1;
	break;
      }

  return 0;
}

/* TRUE if some insn in the range (START, END] references a memory location
   that would be affected by a store to MEMREF.  */

static int
memref_used_between_p (rtx memref, rtx start, rtx end)
{
  rtx insn;

  for (insn = NEXT_INSN (start); insn != NEXT_INSN (end);
       insn = NEXT_INSN (insn))
    if (INSN_P (insn) && memref_referenced_p (memref, PATTERN (insn)))
      return 1;

  return 0;
}

/* Return nonzero if the rtx X is invariant over the current function.  */
/* ??? Actually, the places this is used in reload expect exactly what
   is tested here, and not everything that is function invariant.  In
   particular, the frame pointer and arg pointer are special cased;
   pic_offset_table_rtx is not, and this will cause aborts when we
   go to spill these things to memory.  */

int
function_invariant_p (rtx x)
{
  if (CONSTANT_P (x))
    return 1;
  if (x == frame_pointer_rtx || x == arg_pointer_rtx)
    return 1;
  if (GET_CODE (x) == PLUS
      && (XEXP (x, 0) == frame_pointer_rtx || XEXP (x, 0) == arg_pointer_rtx)
      && CONSTANT_P (XEXP (x, 1)))
    return 1;
  return 0;
}

/* Find registers that are equivalent to a single value throughout the
   compilation (either because they can be referenced in memory or are set once
   from a single constant).  Lower their priority for a register.

   If such a register is only referenced once, try substituting its value
   into the using insn.  If it succeeds, we can eliminate the register
   completely.  */

static void
update_equiv_regs (void)
{
  rtx insn;
  basic_block bb;
  int loop_depth;
  regset_head cleared_regs;
  int clear_regnos = 0;

  reg_equiv = xcalloc (max_regno, sizeof *reg_equiv);
  INIT_REG_SET (&cleared_regs);

  init_alias_analysis ();

  /* Scan the insns and find which registers have equivalences.  Do this
     in a separate scan of the insns because (due to -fcse-follow-jumps)
     a register can be set below its use.  */
  FOR_EACH_BB (bb)
    {
      loop_depth = bb->loop_depth;

      for (insn = BB_HEAD (bb);
	   insn != NEXT_INSN (BB_END (bb));
	   insn = NEXT_INSN (insn))
	{
	  rtx note;
	  rtx set;
	  rtx dest, src;
	  int regno;

	  if (! INSN_P (insn))
	    continue;

	  for (note = REG_NOTES (insn); note; note = XEXP (note, 1))
	    if (REG_NOTE_KIND (note) == REG_INC)
	      no_equiv (XEXP (note, 0), note, NULL);

	  set = single_set (insn);

	  /* If this insn contains more (or less) than a single SET,
	     only mark all destinations as having no known equivalence.  */
	  if (set == 0)
	    {
	      note_stores (PATTERN (insn), no_equiv, NULL);
	      continue;
	    }
	  else if (GET_CODE (PATTERN (insn)) == PARALLEL)
	    {
	      int i;

	      for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
		{
		  rtx part = XVECEXP (PATTERN (insn), 0, i);
		  if (part != set)
		    note_stores (part, no_equiv, NULL);
		}
	    }

	  dest = SET_DEST (set);
	  src = SET_SRC (set);

	  /* If this sets a MEM to the contents of a REG that is only used
	     in a single basic block, see if the register is always equivalent
	     to that memory location and if moving the store from INSN to the
	     insn that set REG is safe.  If so, put a REG_EQUIV note on the
	     initializing insn.

	     Don't add a REG_EQUIV note if the insn already has one.  The existing
	     REG_EQUIV is likely more useful than the one we are adding.

	     If one of the regs in the address has reg_equiv[REGNO].replace set,
	     then we can't add this REG_EQUIV note.  The reg_equiv[REGNO].replace
	     optimization may move the set of this register immediately before
	     insn, which puts it after reg_equiv[REGNO].init_insns, and hence
	     the mention in the REG_EQUIV note would be to an uninitialized
	     pseudo.  */
	  /* ????? This test isn't good enough; we might see a MEM with a use of
	     a pseudo register before we see its setting insn that will cause
	     reg_equiv[].replace for that pseudo to be set.
	     Equivalences to MEMs should be made in another pass, after the
	     reg_equiv[].replace information has been gathered.  */

	  if (GET_CODE (dest) == MEM && GET_CODE (src) == REG
	      && (regno = REGNO (src)) >= FIRST_PSEUDO_REGISTER
	      && REG_BASIC_BLOCK (regno) >= 0
	      && REG_N_SETS (regno) == 1
	      && reg_equiv[regno].init_insns != 0
	      && reg_equiv[regno].init_insns != const0_rtx
	      && ! find_reg_note (XEXP (reg_equiv[regno].init_insns, 0),
				  REG_EQUIV, NULL_RTX)
	      && ! contains_replace_regs (XEXP (dest, 0)))
	    {
	      rtx init_insn = XEXP (reg_equiv[regno].init_insns, 0);
	      if (validate_equiv_mem (init_insn, src, dest)
		  && ! memref_used_between_p (dest, init_insn, insn))
		REG_NOTES (init_insn)
		  = gen_rtx_EXPR_LIST (REG_EQUIV, dest, REG_NOTES (init_insn));
	    }

	  /* We only handle the case of a pseudo register being set
	     once, or always to the same value.  */
	  /* ??? The mn10200 port breaks if we add equivalences for
	     values that need an ADDRESS_REGS register and set them equivalent
	     to a MEM of a pseudo.  The actual problem is in the over-conservative
	     handling of INPADDR_ADDRESS / INPUT_ADDRESS / INPUT triples in
	     calculate_needs, but we traditionally work around this problem
	     here by rejecting equivalences when the destination is in a register
	     that's likely spilled.  This is fragile, of course, since the
	     preferred class of a pseudo depends on all instructions that set
	     or use it.  */

	  if (GET_CODE (dest) != REG
	      || (regno = REGNO (dest)) < FIRST_PSEUDO_REGISTER
	      || reg_equiv[regno].init_insns == const0_rtx
	      || (CLASS_LIKELY_SPILLED_P (reg_preferred_class (regno))
		  && GET_CODE (src) == MEM))
	    {
	      /* This might be setting a SUBREG of a pseudo, a pseudo that is
		 also set somewhere else to a constant.  */
	      note_stores (set, no_equiv, NULL);
	      continue;
	    }

	  note = find_reg_note (insn, REG_EQUAL, NULL_RTX);

	  /* cse sometimes generates function invariants, but doesn't put a
	     REG_EQUAL note on the insn.  Since this note would be redundant,
	     there's no point creating it earlier than here.  */
	  if (! note && ! rtx_varies_p (src, 0))
	    note = set_unique_reg_note (insn, REG_EQUAL, src);

	  /* Don't bother considering a REG_EQUAL note containing an EXPR_LIST
	     since it represents a function call */
	  if (note && GET_CODE (XEXP (note, 0)) == EXPR_LIST)
	    note = NULL_RTX;

	  if (REG_N_SETS (regno) != 1
	      && (! note
		  || rtx_varies_p (XEXP (note, 0), 0)
		  || (reg_equiv[regno].replacement
		      && ! rtx_equal_p (XEXP (note, 0),
					reg_equiv[regno].replacement))))
	    {
	      no_equiv (dest, set, NULL);
	      continue;
	    }
	  /* Record this insn as initializing this register.  */
	  reg_equiv[regno].init_insns
	    = gen_rtx_INSN_LIST (VOIDmode, insn, reg_equiv[regno].init_insns);

	  /* If this register is known to be equal to a constant, record that
	     it is always equivalent to the constant.  */
	  if (note && ! rtx_varies_p (XEXP (note, 0), 0))
	    PUT_MODE (note, (enum machine_mode) REG_EQUIV);

	  /* If this insn introduces a "constant" register, decrease the priority
	     of that register.  Record this insn if the register is only used once
	     more and the equivalence value is the same as our source.

	     The latter condition is checked for two reasons:  First, it is an
	     indication that it may be more efficient to actually emit the insn
	     as written (if no registers are available, reload will substitute
	     the equivalence).  Secondly, it avoids problems with any registers
	     dying in this insn whose death notes would be missed.

	     If we don't have a REG_EQUIV note, see if this insn is loading
	     a register used only in one basic block from a MEM.  If so, and the
	     MEM remains unchanged for the life of the register, add a REG_EQUIV
	     note.  */

	  note = find_reg_note (insn, REG_EQUIV, NULL_RTX);

	  if (note == 0 && REG_BASIC_BLOCK (regno) >= 0
	      && GET_CODE (SET_SRC (set)) == MEM
	      && validate_equiv_mem (insn, dest, SET_SRC (set)))
	    REG_NOTES (insn) = note = gen_rtx_EXPR_LIST (REG_EQUIV, SET_SRC (set),
							 REG_NOTES (insn));

	  if (note)
	    {
	      int regno = REGNO (dest);

	      /* Record whether or not we created a REG_EQUIV note for a LABEL_REF.
		 We might end up substituting the LABEL_REF for uses of the
		 pseudo here or later.  That kind of transformation may turn an
		 indirect jump into a direct jump, in which case we must rerun the
		 jump optimizer to ensure that the JUMP_LABEL fields are valid.  */
	      if (GET_CODE (XEXP (note, 0)) == LABEL_REF
		  || (GET_CODE (XEXP (note, 0)) == CONST
		      && GET_CODE (XEXP (XEXP (note, 0), 0)) == PLUS
		      && (GET_CODE (XEXP (XEXP (XEXP (note, 0), 0), 0))
			  == LABEL_REF)))
		recorded_label_ref = 1;

	      reg_equiv[regno].replacement = XEXP (note, 0);
	      reg_equiv[regno].src_p = &SET_SRC (set);
	      reg_equiv[regno].loop_depth = loop_depth;

	      /* Don't mess with things live during setjmp.  */
	      if (REG_LIVE_LENGTH (regno) >= 0 && optimize)
		{
		  /* Note that the statement below does not affect the priority
		     in local-alloc!  */
		  REG_LIVE_LENGTH (regno) *= 2;


		  /* If the register is referenced exactly twice, meaning it is
		     set once and used once, indicate that the reference may be
		     replaced by the equivalence we computed above.  Do this
		     even if the register is only used in one block so that
		     dependencies can be handled where the last register is
		     used in a different block (i.e. HIGH / LO_SUM sequences)
		     and to reduce the number of registers alive across
		     calls.  */

		    if (REG_N_REFS (regno) == 2
			&& (rtx_equal_p (XEXP (note, 0), src)
			    || ! equiv_init_varies_p (src))
			&& GET_CODE (insn) == INSN
			&& equiv_init_movable_p (PATTERN (insn), regno))
		      reg_equiv[regno].replace = 1;
		}
	    }
	}
    }

  /* Now scan all regs killed in an insn to see if any of them are
     registers only used that once.  If so, see if we can replace the
     reference with the equivalent from.  If we can, delete the
     initializing reference and this register will go away.  If we
     can't replace the reference, and the initializing reference is
     within the same loop (or in an inner loop), then move the register
     initialization just before the use, so that they are in the same
     basic block.  */
  FOR_EACH_BB_REVERSE (bb)
    {
      loop_depth = bb->loop_depth;
      for (insn = BB_END (bb);
	   insn != PREV_INSN (BB_HEAD (bb));
	   insn = PREV_INSN (insn))
	{
	  rtx link;

	  if (! INSN_P (insn))
	    continue;

	  for (link = REG_NOTES (insn); link; link = XEXP (link, 1))
	    {
	      if (REG_NOTE_KIND (link) == REG_DEAD
		  /* Make sure this insn still refers to the register.  */
		  && reg_mentioned_p (XEXP (link, 0), PATTERN (insn)))
		{
		  int regno = REGNO (XEXP (link, 0));
		  rtx equiv_insn;

		  if (! reg_equiv[regno].replace
		      || reg_equiv[regno].loop_depth < loop_depth)
		    continue;

		  /* reg_equiv[REGNO].replace gets set only when
		     REG_N_REFS[REGNO] is 2, i.e. the register is set
		     once and used once.  (If it were only set, but not used,
		     flow would have deleted the setting insns.)  Hence
		     there can only be one insn in reg_equiv[REGNO].init_insns.  */
		  if (reg_equiv[regno].init_insns == NULL_RTX
		      || XEXP (reg_equiv[regno].init_insns, 1) != NULL_RTX)
		    abort ();
		  equiv_insn = XEXP (reg_equiv[regno].init_insns, 0);

		  /* We may not move instructions that can throw, since
		     that changes basic block boundaries and we are not
		     prepared to adjust the CFG to match.  */
		  if (can_throw_internal (equiv_insn))
		    continue;

		  if (asm_noperands (PATTERN (equiv_insn)) < 0
		      && validate_replace_rtx (regno_reg_rtx[regno],
					       *(reg_equiv[regno].src_p), insn))
		    {
		      rtx equiv_link;
		      rtx last_link;
		      rtx note;

		      /* Find the last note.  */
		      for (last_link = link; XEXP (last_link, 1);
			   last_link = XEXP (last_link, 1))
			;

		      /* Append the REG_DEAD notes from equiv_insn.  */
		      equiv_link = REG_NOTES (equiv_insn);
		      while (equiv_link)
			{
			  note = equiv_link;
			  equiv_link = XEXP (equiv_link, 1);
			  if (REG_NOTE_KIND (note) == REG_DEAD)
			    {
			      remove_note (equiv_insn, note);
			      XEXP (last_link, 1) = note;
			      XEXP (note, 1) = NULL_RTX;
			      last_link = note;
			    }
			}

		      remove_death (regno, insn);
		      REG_N_REFS (regno) = 0;
		      REG_FREQ (regno) = 0;
		      delete_insn (equiv_insn);

		      reg_equiv[regno].init_insns
			= XEXP (reg_equiv[regno].init_insns, 1);
		    }
		  /* Move the initialization of the register to just before
		     INSN.  Update the flow information.  */
		  else if (PREV_INSN (insn) != equiv_insn)
		    {
		      rtx new_insn;

		      new_insn = emit_insn_before (PATTERN (equiv_insn), insn);
		      REG_NOTES (new_insn) = REG_NOTES (equiv_insn);
		      REG_NOTES (equiv_insn) = 0;

		      /* Make sure this insn is recognized before reload begins,
			 otherwise eliminate_regs_in_insn will abort.  */
		      INSN_CODE (new_insn) = INSN_CODE (equiv_insn);

		      delete_insn (equiv_insn);

		      XEXP (reg_equiv[regno].init_insns, 0) = new_insn;

		      REG_BASIC_BLOCK (regno) = bb->index;
		      REG_N_CALLS_CROSSED (regno) = 0;
		      REG_LIVE_LENGTH (regno) = 2;

		      if (insn == BB_HEAD (bb))
			BB_HEAD (bb) = PREV_INSN (insn);

		      /* Remember to clear REGNO from all basic block's live
			 info.  */
		      SET_REGNO_REG_SET (&cleared_regs, regno);
		      clear_regnos++;
		    }
		}
	    }
	}
    }

  /* Clear all dead REGNOs from all basic block's live info.  */
  if (clear_regnos)
    {
      int j;
      if (clear_regnos > 8)
	{
	  FOR_EACH_BB (bb)
	    {
	      AND_COMPL_REG_SET (bb->global_live_at_start, &cleared_regs);
	      AND_COMPL_REG_SET (bb->global_live_at_end, &cleared_regs);
	    }
	}
      else
	EXECUTE_IF_SET_IN_REG_SET (&cleared_regs, 0, j,
	  {
	    FOR_EACH_BB (bb)
	      {
	        CLEAR_REGNO_REG_SET (bb->global_live_at_start, j);
	        CLEAR_REGNO_REG_SET (bb->global_live_at_end, j);
	      }
	  });
    }

  /* Clean up.  */
  end_alias_analysis ();
  CLEAR_REG_SET (&cleared_regs);
  free (reg_equiv);
}

/* Mark REG as having no known equivalence.
   Some instructions might have been processed before and furnished
   with REG_EQUIV notes for this register; these notes will have to be
   removed.
   STORE is the piece of RTL that does the non-constant / conflicting
   assignment - a SET, CLOBBER or REG_INC note.  It is currently not used,
   but needs to be there because this function is called from note_stores.  */
static void
no_equiv (rtx reg, rtx store ATTRIBUTE_UNUSED, void *data ATTRIBUTE_UNUSED)
{
  int regno;
  rtx list;

  if (GET_CODE (reg) != REG)
    return;
  regno = REGNO (reg);
  list = reg_equiv[regno].init_insns;
  if (list == const0_rtx)
    return;
  for (; list; list =  XEXP (list, 1))
    {
      rtx insn = XEXP (list, 0);
      remove_note (insn, find_reg_note (insn, REG_EQUIV, NULL_RTX));
    }
  reg_equiv[regno].init_insns = const0_rtx;
  reg_equiv[regno].replacement = NULL_RTX;
}

/* Allocate hard regs to the pseudo regs used only within block number B.
   Only the pseudos that die but once can be handled.  */

static void
block_alloc (int b)
{
  int i, q;
  rtx insn;
  rtx note, hard_reg;
  int insn_number = 0;
  int insn_count = 0;
  int max_uid = get_max_uid ();
  int *qty_order;
  int no_conflict_combined_regno = -1;

  /* Count the instructions in the basic block.  */

  insn = BB_END (BASIC_BLOCK (b));
  while (1)
    {
      if (GET_CODE (insn) != NOTE)
	if (++insn_count > max_uid)
	  abort ();
      if (insn == BB_HEAD (BASIC_BLOCK (b)))
	break;
      insn = PREV_INSN (insn);
    }

  /* +2 to leave room for a post_mark_life at the last insn and for
     the birth of a CLOBBER in the first insn.  */
  regs_live_at = xcalloc ((2 * insn_count + 2), sizeof (HARD_REG_SET));

  /* Initialize table of hardware registers currently live.  */

  REG_SET_TO_HARD_REG_SET (regs_live, BASIC_BLOCK (b)->global_live_at_start);

  /* This loop scans the instructions of the basic block
     and assigns quantities to registers.
     It computes which registers to tie.  */

  insn = BB_HEAD (BASIC_BLOCK (b));
  while (1)
    {
      if (GET_CODE (insn) != NOTE)
	insn_number++;

      if (INSN_P (insn))
	{
	  rtx link, set;
	  int win = 0;
	  rtx r0, r1 = NULL_RTX;
	  int combined_regno = -1;
	  int i;

	  this_insn_number = insn_number;
	  this_insn = insn;

	  extract_insn (insn);
	  which_alternative = -1;

	  /* Is this insn suitable for tying two registers?
	     If so, try doing that.
	     Suitable insns are those with at least two operands and where
	     operand 0 is an output that is a register that is not
	     earlyclobber.

	     We can tie operand 0 with some operand that dies in this insn.
	     First look for operands that are required to be in the same
	     register as operand 0.  If we find such, only try tying that
	     operand or one that can be put into that operand if the
	     operation is commutative.  If we don't find an operand
	     that is required to be in the same register as operand 0,
	     we can tie with any operand.

	     Subregs in place of regs are also ok.

	     If tying is done, WIN is set nonzero.  */

	  if (optimize
	      && recog_data.n_operands > 1
	      && recog_data.constraints[0][0] == '='
	      && recog_data.constraints[0][1] != '&')
	    {
	      /* If non-negative, is an operand that must match operand 0.  */
	      int must_match_0 = -1;
	      /* Counts number of alternatives that require a match with
		 operand 0.  */
	      int n_matching_alts = 0;

	      for (i = 1; i < recog_data.n_operands; i++)
		{
		  const char *p = recog_data.constraints[i];
		  int this_match = requires_inout (p);

		  n_matching_alts += this_match;
		  if (this_match == recog_data.n_alternatives)
		    must_match_0 = i;
		}

	      r0 = recog_data.operand[0];
	      for (i = 1; i < recog_data.n_operands; i++)
		{
		  /* Skip this operand if we found an operand that
		     must match operand 0 and this operand isn't it
		     and can't be made to be it by commutativity.  */

		  if (must_match_0 >= 0 && i != must_match_0
		      && ! (i == must_match_0 + 1
			    && recog_data.constraints[i-1][0] == '%')
		      && ! (i == must_match_0 - 1
			    && recog_data.constraints[i][0] == '%'))
		    continue;

		  /* Likewise if each alternative has some operand that
		     must match operand zero.  In that case, skip any
		     operand that doesn't list operand 0 since we know that
		     the operand always conflicts with operand 0.  We
		     ignore commutativity in this case to keep things simple.  */
		  if (n_matching_alts == recog_data.n_alternatives
		      && 0 == requires_inout (recog_data.constraints[i]))
		    continue;

		  r1 = recog_data.operand[i];

		  /* If the operand is an address, find a register in it.
		     There may be more than one register, but we only try one
		     of them.  */
		  if (recog_data.constraints[i][0] == 'p'
		      || EXTRA_ADDRESS_CONSTRAINT (recog_data.constraints[i][0],
						   recog_data.constraints[i]))
		    while (GET_CODE (r1) == PLUS || GET_CODE (r1) == MULT)
		      r1 = XEXP (r1, 0);

		  /* Avoid making a call-saved register unnecessarily
                     clobbered.  */
		  hard_reg = get_hard_reg_initial_reg (cfun, r1);
		  if (hard_reg != NULL_RTX)
		    {
		      if (GET_CODE (hard_reg) == REG
			  && IN_RANGE (REGNO (hard_reg),
				       0, FIRST_PSEUDO_REGISTER - 1)
			  && ! call_used_regs[REGNO (hard_reg)])
			continue;
		    }

		  if (GET_CODE (r0) == REG || GET_CODE (r0) == SUBREG)
		    {
		      /* We have two priorities for hard register preferences.
			 If we have a move insn or an insn whose first input
			 can only be in the same register as the output, give
			 priority to an equivalence found from that insn.  */
		      int may_save_copy
			= (r1 == recog_data.operand[i] && must_match_0 >= 0);

		      if (GET_CODE (r1) == REG || GET_CODE (r1) == SUBREG)
			win = combine_regs (r1, r0, may_save_copy,
					    insn_number, insn, 0);
		    }
		  if (win)
		    break;
		}
	    }

	  /* Recognize an insn sequence with an ultimate result
	     which can safely overlap one of the inputs.
	     The sequence begins with a CLOBBER of its result,
	     and ends with an insn that copies the result to itself
	     and has a REG_EQUAL note for an equivalent formula.
	     That note indicates what the inputs are.
	     The result and the input can overlap if each insn in
	     the sequence either doesn't mention the input
	     or has a REG_NO_CONFLICT note to inhibit the conflict.

	     We do the combining test at the CLOBBER so that the
	     destination register won't have had a quantity number
	     assigned, since that would prevent combining.  */

	  if (optimize
	      && GET_CODE (PATTERN (insn)) == CLOBBER
	      && (r0 = XEXP (PATTERN (insn), 0),
		  GET_CODE (r0) == REG)
	      && (link = find_reg_note (insn, REG_LIBCALL, NULL_RTX)) != 0
	      && XEXP (link, 0) != 0
	      && GET_CODE (XEXP (link, 0)) == INSN
	      && (set = single_set (XEXP (link, 0))) != 0
	      && SET_DEST (set) == r0 && SET_SRC (set) == r0
	      && (note = find_reg_note (XEXP (link, 0), REG_EQUAL,
					NULL_RTX)) != 0)
	    {
	      if (r1 = XEXP (note, 0), GET_CODE (r1) == REG
		  /* Check that we have such a sequence.  */
		  && no_conflict_p (insn, r0, r1))
		win = combine_regs (r1, r0, 1, insn_number, insn, 1);
	      else if (GET_RTX_FORMAT (GET_CODE (XEXP (note, 0)))[0] == 'e'
		       && (r1 = XEXP (XEXP (note, 0), 0),
			   GET_CODE (r1) == REG || GET_CODE (r1) == SUBREG)
		       && no_conflict_p (insn, r0, r1))
		win = combine_regs (r1, r0, 0, insn_number, insn, 1);

	      /* Here we care if the operation to be computed is
		 commutative.  */
	      else if ((GET_CODE (XEXP (note, 0)) == EQ
			|| GET_CODE (XEXP (note, 0)) == NE
			|| GET_RTX_CLASS (GET_CODE (XEXP (note, 0))) == 'c')
		       && (r1 = XEXP (XEXP (note, 0), 1),
			   (GET_CODE (r1) == REG || GET_CODE (r1) == SUBREG))
		       && no_conflict_p (insn, r0, r1))
		win = combine_regs (r1, r0, 0, insn_number, insn, 1);

	      /* If we did combine something, show the register number
		 in question so that we know to ignore its death.  */
	      if (win)
		no_conflict_combined_regno = REGNO (r1);
	    }

	  /* If registers were just tied, set COMBINED_REGNO
	     to the number of the register used in this insn
	     that was tied to the register set in this insn.
	     This register's qty should not be "killed".  */

	  if (win)
	    {
	      while (GET_CODE (r1) == SUBREG)
		r1 = SUBREG_REG (r1);
	      combined_regno = REGNO (r1);
	    }

	  /* Mark the death of everything that dies in this instruction,
	     except for anything that was just combined.  */

	  for (link = REG_NOTES (insn); link; link = XEXP (link, 1))
	    if (REG_NOTE_KIND (link) == REG_DEAD
		&& GET_CODE (XEXP (link, 0)) == REG
		&& combined_regno != (int) REGNO (XEXP (link, 0))
		&& (no_conflict_combined_regno != (int) REGNO (XEXP (link, 0))
		    || ! find_reg_note (insn, REG_NO_CONFLICT,
					XEXP (link, 0))))
	      wipe_dead_reg (XEXP (link, 0), 0);

	  /* Allocate qty numbers for all registers local to this block
	     that are born (set) in this instruction.
	     A pseudo that already has a qty is not changed.  */

	  note_stores (PATTERN (insn), reg_is_set, NULL);

	  /* If anything is set in this insn and then unused, mark it as dying
	     after this insn, so it will conflict with our outputs.  This
	     can't match with something that combined, and it doesn't matter
	     if it did.  Do this after the calls to reg_is_set since these
	     die after, not during, the current insn.  */

	  for (link = REG_NOTES (insn); link; link = XEXP (link, 1))
	    if (REG_NOTE_KIND (link) == REG_UNUSED
		&& GET_CODE (XEXP (link, 0)) == REG)
	      wipe_dead_reg (XEXP (link, 0), 1);

	  /* If this is an insn that has a REG_RETVAL note pointing at a
	     CLOBBER insn, we have reached the end of a REG_NO_CONFLICT
	     block, so clear any register number that combined within it.  */
	  if ((note = find_reg_note (insn, REG_RETVAL, NULL_RTX)) != 0
	      && GET_CODE (XEXP (note, 0)) == INSN
	      && GET_CODE (PATTERN (XEXP (note, 0))) == CLOBBER)
	    no_conflict_combined_regno = -1;
	}

      /* Set the registers live after INSN_NUMBER.  Note that we never
	 record the registers live before the block's first insn, since no
	 pseudos we care about are live before that insn.  */

      IOR_HARD_REG_SET (regs_live_at[2 * insn_number], regs_live);
      IOR_HARD_REG_SET (regs_live_at[2 * insn_number + 1], regs_live);

      if (insn == BB_END (BASIC_BLOCK (b)))
	break;

      insn = NEXT_INSN (insn);
    }

  /* Now every register that is local to this basic block
     should have been given a quantity, or else -1 meaning ignore it.
     Every quantity should have a known birth and death.

     Order the qtys so we assign them registers in order of the
     number of suggested registers they need so we allocate those with
     the most restrictive needs first.  */

  qty_order = xmalloc (next_qty * sizeof (int));
  for (i = 0; i < next_qty; i++)
    qty_order[i] = i;

#define EXCHANGE(I1, I2)  \
  { i = qty_order[I1]; qty_order[I1] = qty_order[I2]; qty_order[I2] = i; }

  switch (next_qty)
    {
    case 3:
      /* Make qty_order[2] be the one to allocate last.  */
      if (qty_sugg_compare (0, 1) > 0)
	EXCHANGE (0, 1);
      if (qty_sugg_compare (1, 2) > 0)
	EXCHANGE (2, 1);

      /* ... Fall through ...  */
    case 2:
      /* Put the best one to allocate in qty_order[0].  */
      if (qty_sugg_compare (0, 1) > 0)
	EXCHANGE (0, 1);

      /* ... Fall through ...  */

    case 1:
    case 0:
      /* Nothing to do here.  */
      break;

    default:
      qsort (qty_order, next_qty, sizeof (int), qty_sugg_compare_1);
    }

  /* Try to put each quantity in a suggested physical register, if it has one.
     This may cause registers to be allocated that otherwise wouldn't be, but
     this seems acceptable in local allocation (unlike global allocation).  */
  for (i = 0; i < next_qty; i++)
    {
      q = qty_order[i];
      if (qty_phys_num_sugg[q] != 0 || qty_phys_num_copy_sugg[q] != 0)
	qty[q].phys_reg = find_free_reg (qty[q].min_class, qty[q].mode, q,
					 0, 1, qty[q].birth, qty[q].death);
      else
	qty[q].phys_reg = -1;
    }

  /* Order the qtys so we assign them registers in order of
     decreasing length of life.  Normally call qsort, but if we
     have only a very small number of quantities, sort them ourselves.  */

  for (i = 0; i < next_qty; i++)
    qty_order[i] = i;

#define EXCHANGE(I1, I2)  \
  { i = qty_order[I1]; qty_order[I1] = qty_order[I2]; qty_order[I2] = i; }

  switch (next_qty)
    {
    case 3:
      /* Make qty_order[2] be the one to allocate last.  */
      if (qty_compare (0, 1) > 0)
	EXCHANGE (0, 1);
      if (qty_compare (1, 2) > 0)
	EXCHANGE (2, 1);

      /* ... Fall through ...  */
    case 2:
      /* Put the best one to allocate in qty_order[0].  */
      if (qty_compare (0, 1) > 0)
	EXCHANGE (0, 1);

      /* ... Fall through ...  */

    case 1:
    case 0:
      /* Nothing to do here.  */
      break;

    default:
      qsort (qty_order, next_qty, sizeof (int), qty_compare_1);
    }

  /* Now for each qty that is not a hardware register,
     look for a hardware register to put it in.
     First try the register class that is cheapest for this qty,
     if there is more than one class.  */

  for (i = 0; i < next_qty; i++)
    {
      q = qty_order[i];
      if (qty[q].phys_reg < 0)
	{
#ifdef INSN_SCHEDULING
	  /* These values represent the adjusted lifetime of a qty so
	     that it conflicts with qtys which appear near the start/end
	     of this qty's lifetime.

	     The purpose behind extending the lifetime of this qty is to
	     discourage the register allocator from creating false
	     dependencies.

	     The adjustment value is chosen to indicate that this qty
	     conflicts with all the qtys in the instructions immediately
	     before and after the lifetime of this qty.

	     Experiments have shown that higher values tend to hurt
	     overall code performance.

	     If allocation using the extended lifetime fails we will try
	     again with the qty's unadjusted lifetime.  */
	  int fake_birth = MAX (0, qty[q].birth - 2 + qty[q].birth % 2);
	  int fake_death = MIN (insn_number * 2 + 1,
				qty[q].death + 2 - qty[q].death % 2);
#endif

	  if (N_REG_CLASSES > 1)
	    {
#ifdef INSN_SCHEDULING
	      /* We try to avoid using hard registers allocated to qtys which
		 are born immediately after this qty or die immediately before
		 this qty.

		 This optimization is only appropriate when we will run
		 a scheduling pass after reload and we are not optimizing
		 for code size.  */
	      if (flag_schedule_insns_after_reload
		  && !optimize_size
		  && !SMALL_REGISTER_CLASSES)
		{
		  qty[q].phys_reg = find_free_reg (qty[q].min_class,
						   qty[q].mode, q, 0, 0,
						   fake_birth, fake_death);
		  if (qty[q].phys_reg >= 0)
		    continue;
		}
#endif
	      qty[q].phys_reg = find_free_reg (qty[q].min_class,
					       qty[q].mode, q, 0, 0,
					       qty[q].birth, qty[q].death);
	      if (qty[q].phys_reg >= 0)
		continue;
	    }

#ifdef INSN_SCHEDULING
	  /* Similarly, avoid false dependencies.  */
	  if (flag_schedule_insns_after_reload
	      && !optimize_size
	      && !SMALL_REGISTER_CLASSES
	      && qty[q].alternate_class != NO_REGS)
	    qty[q].phys_reg = find_free_reg (qty[q].alternate_class,
					     qty[q].mode, q, 0, 0,
					     fake_birth, fake_death);
#endif
	  if (qty[q].alternate_class != NO_REGS)
	    qty[q].phys_reg = find_free_reg (qty[q].alternate_class,
					     qty[q].mode, q, 0, 0,
					     qty[q].birth, qty[q].death);
	}
    }

  /* Now propagate the register assignments
     to the pseudo regs belonging to the qtys.  */

  for (q = 0; q < next_qty; q++)
    if (qty[q].phys_reg >= 0)
      {
	for (i = qty[q].first_reg; i >= 0; i = reg_next_in_qty[i])
	  reg_renumber[i] = qty[q].phys_reg + reg_offset[i];
      }

  /* Clean up.  */
  free (regs_live_at);
  free (qty_order);
}

/* Compare two quantities' priority for getting real registers.
   We give shorter-lived quantities higher priority.
   Quantities with more references are also preferred, as are quantities that
   require multiple registers.  This is the identical prioritization as
   done by global-alloc.

   We used to give preference to registers with *longer* lives, but using
   the same algorithm in both local- and global-alloc can speed up execution
   of some programs by as much as a factor of three!  */

/* Note that the quotient will never be bigger than
   the value of floor_log2 times the maximum number of
   times a register can occur in one insn (surely less than 100)
   weighted by frequency (max REG_FREQ_MAX).
   Multiplying this by 10000/REG_FREQ_MAX can't overflow.
   QTY_CMP_PRI is also used by qty_sugg_compare.  */

#define QTY_CMP_PRI(q)		\
  ((int) (((double) (floor_log2 (qty[q].n_refs) * qty[q].freq * qty[q].size) \
	  / (qty[q].death - qty[q].birth)) * (10000 / REG_FREQ_MAX)))

static int
qty_compare (int q1, int q2)
{
  return QTY_CMP_PRI (q2) - QTY_CMP_PRI (q1);
}

static int
qty_compare_1 (const void *q1p, const void *q2p)
{
  int q1 = *(const int *) q1p, q2 = *(const int *) q2p;
  int tem = QTY_CMP_PRI (q2) - QTY_CMP_PRI (q1);

  if (tem != 0)
    return tem;

  /* If qtys are equally good, sort by qty number,
     so that the results of qsort leave nothing to chance.  */
  return q1 - q2;
}

/* Compare two quantities' priority for getting real registers.  This version
   is called for quantities that have suggested hard registers.  First priority
   goes to quantities that have copy preferences, then to those that have
   normal preferences.  Within those groups, quantities with the lower
   number of preferences have the highest priority.  Of those, we use the same
   algorithm as above.  */

#define QTY_CMP_SUGG(q)		\
  (qty_phys_num_copy_sugg[q]		\
    ? qty_phys_num_copy_sugg[q]	\
    : qty_phys_num_sugg[q] * FIRST_PSEUDO_REGISTER)

static int
qty_sugg_compare (int q1, int q2)
{
  int tem = QTY_CMP_SUGG (q1) - QTY_CMP_SUGG (q2);

  if (tem != 0)
    return tem;

  return QTY_CMP_PRI (q2) - QTY_CMP_PRI (q1);
}

static int
qty_sugg_compare_1 (const void *q1p, const void *q2p)
{
  int q1 = *(const int *) q1p, q2 = *(const int *) q2p;
  int tem = QTY_CMP_SUGG (q1) - QTY_CMP_SUGG (q2);

  if (tem != 0)
    return tem;

  tem = QTY_CMP_PRI (q2) - QTY_CMP_PRI (q1);
  if (tem != 0)
    return tem;

  /* If qtys are equally good, sort by qty number,
     so that the results of qsort leave nothing to chance.  */
  return q1 - q2;
}

#undef QTY_CMP_SUGG
#undef QTY_CMP_PRI

/* Attempt to combine the two registers (rtx's) USEDREG and SETREG.
   Returns 1 if have done so, or 0 if cannot.

   Combining registers means marking them as having the same quantity
   and adjusting the offsets within the quantity if either of
   them is a SUBREG.

   We don't actually combine a hard reg with a pseudo; instead
   we just record the hard reg as the suggestion for the pseudo's quantity.
   If we really combined them, we could lose if the pseudo lives
   across an insn that clobbers the hard reg (eg, movstr).

   ALREADY_DEAD is nonzero if USEDREG is known to be dead even though
   there is no REG_DEAD note on INSN.  This occurs during the processing
   of REG_NO_CONFLICT blocks.

   MAY_SAVE_COPY is nonzero if this insn is simply copying USEDREG to
   SETREG or if the input and output must share a register.
   In that case, we record a hard reg suggestion in QTY_PHYS_COPY_SUGG.

   There are elaborate checks for the validity of combining.  */

static int
combine_regs (rtx usedreg, rtx setreg, int may_save_copy, int insn_number,
	      rtx insn, int already_dead)
{
  int ureg, sreg;
  int offset = 0;
  int usize, ssize;
  int sqty;

  /* Determine the numbers and sizes of registers being used.  If a subreg
     is present that does not change the entire register, don't consider
     this a copy insn.  */

  while (GET_CODE (usedreg) == SUBREG)
    {
      rtx subreg = SUBREG_REG (usedreg);

      if (GET_CODE (subreg) == REG)
	{
	  if (GET_MODE_SIZE (GET_MODE (subreg)) > UNITS_PER_WORD)
	    may_save_copy = 0;

	  if (REGNO (subreg) < FIRST_PSEUDO_REGISTER)
	    offset += subreg_regno_offset (REGNO (subreg),
					   GET_MODE (subreg),
					   SUBREG_BYTE (usedreg),
					   GET_MODE (usedreg));
	  else
	    offset += (SUBREG_BYTE (usedreg)
		      / REGMODE_NATURAL_SIZE (GET_MODE (usedreg)));
	}

      usedreg = subreg;
    }

  if (GET_CODE (usedreg) != REG)
    return 0;

  ureg = REGNO (usedreg);
  if (ureg < FIRST_PSEUDO_REGISTER)
    usize = hard_regno_nregs[ureg][GET_MODE (usedreg)];
  else
    usize = ((GET_MODE_SIZE (GET_MODE (usedreg))
	      + (REGMODE_NATURAL_SIZE (GET_MODE (usedreg)) - 1))
	     / REGMODE_NATURAL_SIZE (GET_MODE (usedreg)));

  while (GET_CODE (setreg) == SUBREG)
    {
      rtx subreg = SUBREG_REG (setreg);

      if (GET_CODE (subreg) == REG)
	{
	  if (GET_MODE_SIZE (GET_MODE (subreg)) > UNITS_PER_WORD)
	    may_save_copy = 0;

	  if (REGNO (subreg) < FIRST_PSEUDO_REGISTER)
	    offset -= subreg_regno_offset (REGNO (subreg),
					   GET_MODE (subreg),
					   SUBREG_BYTE (setreg),
					   GET_MODE (setreg));
	  else
	    offset -= (SUBREG_BYTE (setreg)
		      / REGMODE_NATURAL_SIZE (GET_MODE (setreg)));
	}

      setreg = subreg;
    }

  if (GET_CODE (setreg) != REG)
    return 0;

  sreg = REGNO (setreg);
  if (sreg < FIRST_PSEUDO_REGISTER)
    ssize = hard_regno_nregs[sreg][GET_MODE (setreg)];
  else
    ssize = ((GET_MODE_SIZE (GET_MODE (setreg))
	      + (REGMODE_NATURAL_SIZE (GET_MODE (setreg)) - 1))
	     / REGMODE_NATURAL_SIZE (GET_MODE (setreg)));

  /* If UREG is a pseudo-register that hasn't already been assigned a
     quantity number, it means that it is not local to this block or dies
     more than once.  In either event, we can't do anything with it.  */
  if ((ureg >= FIRST_PSEUDO_REGISTER && reg_qty[ureg] < 0)
      /* Do not combine registers unless one fits within the other.  */
      || (offset > 0 && usize + offset > ssize)
      || (offset < 0 && usize + offset < ssize)
      /* Do not combine with a smaller already-assigned object
	 if that smaller object is already combined with something bigger.  */
      || (ssize > usize && ureg >= FIRST_PSEUDO_REGISTER
	  && usize < qty[reg_qty[ureg]].size)
      /* Can't combine if SREG is not a register we can allocate.  */
      || (sreg >= FIRST_PSEUDO_REGISTER && reg_qty[sreg] == -1)
      /* Don't combine with a pseudo mentioned in a REG_NO_CONFLICT note.
	 These have already been taken care of.  This probably wouldn't
	 combine anyway, but don't take any chances.  */
      || (ureg >= FIRST_PSEUDO_REGISTER
	  && find_reg_note (insn, REG_NO_CONFLICT, usedreg))
      /* Don't tie something to itself.  In most cases it would make no
	 difference, but it would screw up if the reg being tied to itself
	 also dies in this insn.  */
      || ureg == sreg
      /* Don't try to connect two different hardware registers.  */
      || (ureg < FIRST_PSEUDO_REGISTER && sreg < FIRST_PSEUDO_REGISTER)
      /* Don't connect two different machine modes if they have different
	 implications as to which registers may be used.  */
      || !MODES_TIEABLE_P (GET_MODE (usedreg), GET_MODE (setreg)))
    return 0;

  /* Now, if UREG is a hard reg and SREG is a pseudo, record the hard reg in
     qty_phys_sugg for the pseudo instead of tying them.

     Return "failure" so that the lifespan of UREG is terminated here;
     that way the two lifespans will be disjoint and nothing will prevent
     the pseudo reg from being given this hard reg.  */

  if (ureg < FIRST_PSEUDO_REGISTER)
    {
      /* Allocate a quantity number so we have a place to put our
	 suggestions.  */
      if (reg_qty[sreg] == -2)
	reg_is_born (setreg, 2 * insn_number);

      if (reg_qty[sreg] >= 0)
	{
	  if (may_save_copy
	      && ! TEST_HARD_REG_BIT (qty_phys_copy_sugg[reg_qty[sreg]], ureg))
	    {
	      SET_HARD_REG_BIT (qty_phys_copy_sugg[reg_qty[sreg]], ureg);
	      qty_phys_num_copy_sugg[reg_qty[sreg]]++;
	    }
	  else if (! TEST_HARD_REG_BIT (qty_phys_sugg[reg_qty[sreg]], ureg))
	    {
	      SET_HARD_REG_BIT (qty_phys_sugg[reg_qty[sreg]], ureg);
	      qty_phys_num_sugg[reg_qty[sreg]]++;
	    }
	}
      return 0;
    }

  /* Similarly for SREG a hard register and UREG a pseudo register.  */

  if (sreg < FIRST_PSEUDO_REGISTER)
    {
      if (may_save_copy
	  && ! TEST_HARD_REG_BIT (qty_phys_copy_sugg[reg_qty[ureg]], sreg))
	{
	  SET_HARD_REG_BIT (qty_phys_copy_sugg[reg_qty[ureg]], sreg);
	  qty_phys_num_copy_sugg[reg_qty[ureg]]++;
	}
      else if (! TEST_HARD_REG_BIT (qty_phys_sugg[reg_qty[ureg]], sreg))
	{
	  SET_HARD_REG_BIT (qty_phys_sugg[reg_qty[ureg]], sreg);
	  qty_phys_num_sugg[reg_qty[ureg]]++;
	}
      return 0;
    }

  /* At this point we know that SREG and UREG are both pseudos.
     Do nothing if SREG already has a quantity or is a register that we
     don't allocate.  */
  if (reg_qty[sreg] >= -1
      /* If we are not going to let any regs live across calls,
	 don't tie a call-crossing reg to a non-call-crossing reg.  */
      || (current_function_has_nonlocal_label
	  && ((REG_N_CALLS_CROSSED (ureg) > 0)
	      != (REG_N_CALLS_CROSSED (sreg) > 0))))
    return 0;

  /* We don't already know about SREG, so tie it to UREG
     if this is the last use of UREG, provided the classes they want
     are compatible.  */

  if ((already_dead || find_regno_note (insn, REG_DEAD, ureg))
      && reg_meets_class_p (sreg, qty[reg_qty[ureg]].min_class))
    {
      /* Add SREG to UREG's quantity.  */
      sqty = reg_qty[ureg];
      reg_qty[sreg] = sqty;
      reg_offset[sreg] = reg_offset[ureg] + offset;
      reg_next_in_qty[sreg] = qty[sqty].first_reg;
      qty[sqty].first_reg = sreg;

      /* If SREG's reg class is smaller, set qty[SQTY].min_class.  */
      update_qty_class (sqty, sreg);

      /* Update info about quantity SQTY.  */
      qty[sqty].n_calls_crossed += REG_N_CALLS_CROSSED (sreg);
      qty[sqty].n_refs += REG_N_REFS (sreg);
      qty[sqty].freq += REG_FREQ (sreg);
      if (usize < ssize)
	{
	  int i;

	  for (i = qty[sqty].first_reg; i >= 0; i = reg_next_in_qty[i])
	    reg_offset[i] -= offset;

	  qty[sqty].size = ssize;
	  qty[sqty].mode = GET_MODE (setreg);
	}
    }
  else
    return 0;

  return 1;
}

/* Return 1 if the preferred class of REG allows it to be tied
   to a quantity or register whose class is CLASS.
   True if REG's reg class either contains or is contained in CLASS.  */

static int
reg_meets_class_p (int reg, enum reg_class class)
{
  enum reg_class rclass = reg_preferred_class (reg);
  return (reg_class_subset_p (rclass, class)
	  || reg_class_subset_p (class, rclass));
}

/* Update the class of QTYNO assuming that REG is being tied to it.  */

static void
update_qty_class (int qtyno, int reg)
{
  enum reg_class rclass = reg_preferred_class (reg);
  if (reg_class_subset_p (rclass, qty[qtyno].min_class))
    qty[qtyno].min_class = rclass;

  rclass = reg_alternate_class (reg);
  if (reg_class_subset_p (rclass, qty[qtyno].alternate_class))
    qty[qtyno].alternate_class = rclass;
}

/* Handle something which alters the value of an rtx REG.

   REG is whatever is set or clobbered.  SETTER is the rtx that
   is modifying the register.

   If it is not really a register, we do nothing.
   The file-global variables `this_insn' and `this_insn_number'
   carry info from `block_alloc'.  */

static void
reg_is_set (rtx reg, rtx setter, void *data ATTRIBUTE_UNUSED)
{
  /* Note that note_stores will only pass us a SUBREG if it is a SUBREG of
     a hard register.  These may actually not exist any more.  */

  if (GET_CODE (reg) != SUBREG
      && GET_CODE (reg) != REG)
    return;

  /* Mark this register as being born.  If it is used in a CLOBBER, mark
     it as being born halfway between the previous insn and this insn so that
     it conflicts with our inputs but not the outputs of the previous insn.  */

  reg_is_born (reg, 2 * this_insn_number - (GET_CODE (setter) == CLOBBER));
}

/* Handle beginning of the life of register REG.
   BIRTH is the index at which this is happening.  */

static void
reg_is_born (rtx reg, int birth)
{
  int regno;

  if (GET_CODE (reg) == SUBREG)
    {
      regno = REGNO (SUBREG_REG (reg));
      if (regno < FIRST_PSEUDO_REGISTER)
	regno = subreg_hard_regno (reg, 1);
    }
  else
    regno = REGNO (reg);

  if (regno < FIRST_PSEUDO_REGISTER)
    {
      mark_life (regno, GET_MODE (reg), 1);

      /* If the register was to have been born earlier that the present
	 insn, mark it as live where it is actually born.  */
      if (birth < 2 * this_insn_number)
	post_mark_life (regno, GET_MODE (reg), 1, birth, 2 * this_insn_number);
    }
  else
    {
      if (reg_qty[regno] == -2)
	alloc_qty (regno, GET_MODE (reg), PSEUDO_REGNO_SIZE (regno), birth);

      /* If this register has a quantity number, show that it isn't dead.  */
      if (reg_qty[regno] >= 0)
	qty[reg_qty[regno]].death = -1;
    }
}

/* Record the death of REG in the current insn.  If OUTPUT_P is nonzero,
   REG is an output that is dying (i.e., it is never used), otherwise it
   is an input (the normal case).
   If OUTPUT_P is 1, then we extend the life past the end of this insn.  */

static void
wipe_dead_reg (rtx reg, int output_p)
{
  int regno = REGNO (reg);

  /* If this insn has multiple results,
     and the dead reg is used in one of the results,
     extend its life to after this insn,
     so it won't get allocated together with any other result of this insn.

     It is unsafe to use !single_set here since it will ignore an unused
     output.  Just because an output is unused does not mean the compiler
     can assume the side effect will not occur.   Consider if REG appears
     in the address of an output and we reload the output.  If we allocate
     REG to the same hard register as an unused output we could set the hard
     register before the output reload insn.  */
  if (GET_CODE (PATTERN (this_insn)) == PARALLEL
      && multiple_sets (this_insn))
    {
      int i;
      for (i = XVECLEN (PATTERN (this_insn), 0) - 1; i >= 0; i--)
	{
	  rtx set = XVECEXP (PATTERN (this_insn), 0, i);
	  if (GET_CODE (set) == SET
	      && GET_CODE (SET_DEST (set)) != REG
	      && !rtx_equal_p (reg, SET_DEST (set))
	      && reg_overlap_mentioned_p (reg, SET_DEST (set)))
	    output_p = 1;
	}
    }

  /* If this register is used in an auto-increment address, then extend its
     life to after this insn, so that it won't get allocated together with
     the result of this insn.  */
  if (! output_p && find_regno_note (this_insn, REG_INC, regno))
    output_p = 1;

  if (regno < FIRST_PSEUDO_REGISTER)
    {
      mark_life (regno, GET_MODE (reg), 0);

      /* If a hard register is dying as an output, mark it as in use at
	 the beginning of this insn (the above statement would cause this
	 not to happen).  */
      if (output_p)
	post_mark_life (regno, GET_MODE (reg), 1,
			2 * this_insn_number, 2 * this_insn_number + 1);
    }

  else if (reg_qty[regno] >= 0)
    qty[reg_qty[regno]].death = 2 * this_insn_number + output_p;
}

/* Find a block of SIZE words of hard regs in reg_class CLASS
   that can hold something of machine-mode MODE
     (but actually we test only the first of the block for holding MODE)
   and still free between insn BORN_INDEX and insn DEAD_INDEX,
   and return the number of the first of them.
   Return -1 if such a block cannot be found.
   If QTYNO crosses calls, insist on a register preserved by calls,
   unless ACCEPT_CALL_CLOBBERED is nonzero.

   If JUST_TRY_SUGGESTED is nonzero, only try to see if the suggested
   register is available.  If not, return -1.  */

static int
find_free_reg (enum reg_class class, enum machine_mode mode, int qtyno,
	       int accept_call_clobbered, int just_try_suggested,
	       int born_index, int dead_index)
{
  int i, ins;
  HARD_REG_SET first_used, used;
#ifdef ELIMINABLE_REGS
  static const struct {const int from, to; } eliminables[] = ELIMINABLE_REGS;
#endif

  /* Validate our parameters.  */
  if (born_index < 0 || born_index > dead_index)
    abort ();

  /* Don't let a pseudo live in a reg across a function call
     if we might get a nonlocal goto.  */
  if (current_function_has_nonlocal_label
      && qty[qtyno].n_calls_crossed > 0)
    return -1;

  if (accept_call_clobbered)
    COPY_HARD_REG_SET (used, call_fixed_reg_set);
  else if (qty[qtyno].n_calls_crossed == 0)
    COPY_HARD_REG_SET (used, fixed_reg_set);
  else
    COPY_HARD_REG_SET (used, call_used_reg_set);

  if (accept_call_clobbered)
    IOR_HARD_REG_SET (used, losing_caller_save_reg_set);

  for (ins = born_index; ins < dead_index; ins++)
    IOR_HARD_REG_SET (used, regs_live_at[ins]);

  IOR_COMPL_HARD_REG_SET (used, reg_class_contents[(int) class]);

  /* Don't use the frame pointer reg in local-alloc even if
     we may omit the frame pointer, because if we do that and then we
     need a frame pointer, reload won't know how to move the pseudo
     to another hard reg.  It can move only regs made by global-alloc.

     This is true of any register that can be eliminated.  */
#ifdef ELIMINABLE_REGS
  for (i = 0; i < (int) ARRAY_SIZE (eliminables); i++)
    SET_HARD_REG_BIT (used, eliminables[i].from);
#if FRAME_POINTER_REGNUM != HARD_FRAME_POINTER_REGNUM
  /* If FRAME_POINTER_REGNUM is not a real register, then protect the one
     that it might be eliminated into.  */
  SET_HARD_REG_BIT (used, HARD_FRAME_POINTER_REGNUM);
#endif
#else
  SET_HARD_REG_BIT (used, FRAME_POINTER_REGNUM);
#endif

#ifdef CANNOT_CHANGE_MODE_CLASS
  cannot_change_mode_set_regs (&used, mode, qty[qtyno].first_reg);
#endif

  /* Normally, the registers that can be used for the first register in
     a multi-register quantity are the same as those that can be used for
     subsequent registers.  However, if just trying suggested registers,
     restrict our consideration to them.  If there are copy-suggested
     register, try them.  Otherwise, try the arithmetic-suggested
     registers.  */
  COPY_HARD_REG_SET (first_used, used);

  if (just_try_suggested)
    {
      if (qty_phys_num_copy_sugg[qtyno] != 0)
	IOR_COMPL_HARD_REG_SET (first_used, qty_phys_copy_sugg[qtyno]);
      else
	IOR_COMPL_HARD_REG_SET (first_used, qty_phys_sugg[qtyno]);
    }

  /* If all registers are excluded, we can't do anything.  */
  GO_IF_HARD_REG_SUBSET (reg_class_contents[(int) ALL_REGS], first_used, fail);

  /* If at least one would be suitable, test each hard reg.  */

  for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
    {
#ifdef REG_ALLOC_ORDER
      int regno = reg_alloc_order[i];
#else
      int regno = i;
#endif
      if (! TEST_HARD_REG_BIT (first_used, regno)
	  && HARD_REGNO_MODE_OK (regno, mode)
	  && (qty[qtyno].n_calls_crossed == 0
	      || accept_call_clobbered
	      || ! HARD_REGNO_CALL_PART_CLOBBERED (regno, mode)))
	{
	  int j;
	  int size1 = hard_regno_nregs[regno][mode];
	  for (j = 1; j < size1 && ! TEST_HARD_REG_BIT (used, regno + j); j++);
	  if (j == size1)
	    {
	      /* Mark that this register is in use between its birth and death
		 insns.  */
	      post_mark_life (regno, mode, 1, born_index, dead_index);
	      return regno;
	    }
#ifndef REG_ALLOC_ORDER
	  /* Skip starting points we know will lose.  */
	  i += j;
#endif
	}
    }

 fail:
  /* If we are just trying suggested register, we have just tried copy-
     suggested registers, and there are arithmetic-suggested registers,
     try them.  */

  /* If it would be profitable to allocate a call-clobbered register
     and save and restore it around calls, do that.  */
  if (just_try_suggested && qty_phys_num_copy_sugg[qtyno] != 0
      && qty_phys_num_sugg[qtyno] != 0)
    {
      /* Don't try the copy-suggested regs again.  */
      qty_phys_num_copy_sugg[qtyno] = 0;
      return find_free_reg (class, mode, qtyno, accept_call_clobbered, 1,
			    born_index, dead_index);
    }

  /* We need not check to see if the current function has nonlocal
     labels because we don't put any pseudos that are live over calls in
     registers in that case.  */

  if (! accept_call_clobbered
      && flag_caller_saves
      && ! just_try_suggested
      && qty[qtyno].n_calls_crossed != 0
      && CALLER_SAVE_PROFITABLE (qty[qtyno].n_refs,
				 qty[qtyno].n_calls_crossed))
    {
      i = find_free_reg (class, mode, qtyno, 1, 0, born_index, dead_index);
      if (i >= 0)
	caller_save_needed = 1;
      return i;
    }
  return -1;
}

/* Mark that REGNO with machine-mode MODE is live starting from the current
   insn (if LIFE is nonzero) or dead starting at the current insn (if LIFE
   is zero).  */

static void
mark_life (int regno, enum machine_mode mode, int life)
{
  int j = hard_regno_nregs[regno][mode];
  if (life)
    while (--j >= 0)
      SET_HARD_REG_BIT (regs_live, regno + j);
  else
    while (--j >= 0)
      CLEAR_HARD_REG_BIT (regs_live, regno + j);
}

/* Mark register number REGNO (with machine-mode MODE) as live (if LIFE
   is nonzero) or dead (if LIFE is zero) from insn number BIRTH (inclusive)
   to insn number DEATH (exclusive).  */

static void
post_mark_life (int regno, enum machine_mode mode, int life, int birth,
		int death)
{
  int j = hard_regno_nregs[regno][mode];
  HARD_REG_SET this_reg;

  CLEAR_HARD_REG_SET (this_reg);
  while (--j >= 0)
    SET_HARD_REG_BIT (this_reg, regno + j);

  if (life)
    while (birth < death)
      {
	IOR_HARD_REG_SET (regs_live_at[birth], this_reg);
	birth++;
      }
  else
    while (birth < death)
      {
	AND_COMPL_HARD_REG_SET (regs_live_at[birth], this_reg);
	birth++;
      }
}

/* INSN is the CLOBBER insn that starts a REG_NO_NOCONFLICT block, R0
   is the register being clobbered, and R1 is a register being used in
   the equivalent expression.

   If R1 dies in the block and has a REG_NO_CONFLICT note on every insn
   in which it is used, return 1.

   Otherwise, return 0.  */

static int
no_conflict_p (rtx insn, rtx r0 ATTRIBUTE_UNUSED, rtx r1)
{
  int ok = 0;
  rtx note = find_reg_note (insn, REG_LIBCALL, NULL_RTX);
  rtx p, last;

  /* If R1 is a hard register, return 0 since we handle this case
     when we scan the insns that actually use it.  */

  if (note == 0
      || (GET_CODE (r1) == REG && REGNO (r1) < FIRST_PSEUDO_REGISTER)
      || (GET_CODE (r1) == SUBREG && GET_CODE (SUBREG_REG (r1)) == REG
	  && REGNO (SUBREG_REG (r1)) < FIRST_PSEUDO_REGISTER))
    return 0;

  last = XEXP (note, 0);

  for (p = NEXT_INSN (insn); p && p != last; p = NEXT_INSN (p))
    if (INSN_P (p))
      {
	if (find_reg_note (p, REG_DEAD, r1))
	  ok = 1;

	/* There must be a REG_NO_CONFLICT note on every insn, otherwise
	   some earlier optimization pass has inserted instructions into
	   the sequence, and it is not safe to perform this optimization.
	   Note that emit_no_conflict_block always ensures that this is
	   true when these sequences are created.  */
	if (! find_reg_note (p, REG_NO_CONFLICT, r1))
	  return 0;
      }

  return ok;
}

/* Return the number of alternatives for which the constraint string P
   indicates that the operand must be equal to operand 0 and that no register
   is acceptable.  */

static int
requires_inout (const char *p)
{
  char c;
  int found_zero = 0;
  int reg_allowed = 0;
  int num_matching_alts = 0;
  int len;

  for ( ; (c = *p); p += len)
    {
      len = CONSTRAINT_LEN (c, p);
      switch (c)
	{
	case '=':  case '+':  case '?':
	case '#':  case '&':  case '!':
	case '*':  case '%':
	case 'm':  case '<':  case '>':  case 'V':  case 'o':
	case 'E':  case 'F':  case 'G':  case 'H':
	case 's':  case 'i':  case 'n':
	case 'I':  case 'J':  case 'K':  case 'L':
	case 'M':  case 'N':  case 'O':  case 'P':
	case 'X':
	  /* These don't say anything we care about.  */
	  break;

	case ',':
	  if (found_zero && ! reg_allowed)
	    num_matching_alts++;

	  found_zero = reg_allowed = 0;
	  break;

	case '0':
	  found_zero = 1;
	  break;

	case '1':  case '2':  case '3':  case '4': case '5':
	case '6':  case '7':  case '8':  case '9':
	  /* Skip the balance of the matching constraint.  */
	  do
	    p++;
	  while (ISDIGIT (*p));
	  len = 0;
	  break;

	default:
	  if (REG_CLASS_FROM_CONSTRAINT (c, p) == NO_REGS
	      && !EXTRA_ADDRESS_CONSTRAINT (c, p))
	    break;
	  /* Fall through.  */
	case 'p':
	case 'g': case 'r':
	  reg_allowed = 1;
	  break;
	}
    }

  if (found_zero && ! reg_allowed)
    num_matching_alts++;

  return num_matching_alts;
}

void
dump_local_alloc (FILE *file)
{
  int i;
  for (i = FIRST_PSEUDO_REGISTER; i < max_regno; i++)
    if (reg_renumber[i] != -1)
      fprintf (file, ";; Register %d in %d.\n", i, reg_renumber[i]);
}
ppc">#include "hash-set.h" #include "machmode.h" #include "vec.h" #include "double-int.h" #include "input.h" #include "alias.h" #include "symtab.h" #include "options.h" #include "wide-int.h" #include "inchash.h" #include "tree.h" #include "fold-const.h" #include "gimple-expr.h" #include "diagnostic-core.h" /* For internal_error/fatal_error. */ #include "flags.h" #include "constructor.h" #include "trans.h" #include "trans-stmt.h" #include "trans-types.h" #include "trans-array.h" #include "trans-const.h" #include "dependency.h" #include "wide-int.h" static bool gfc_get_array_constructor_size (mpz_t *, gfc_constructor_base); /* The contents of this structure aren't actually used, just the address. */ static gfc_ss gfc_ss_terminator_var; gfc_ss * const gfc_ss_terminator = &gfc_ss_terminator_var; static tree gfc_array_dataptr_type (tree desc) { return (GFC_TYPE_ARRAY_DATAPTR_TYPE (TREE_TYPE (desc))); } /* Build expressions to access the members of an array descriptor. It's surprisingly easy to mess up here, so never access an array descriptor by "brute force", always use these functions. This also avoids problems if we change the format of an array descriptor. To understand these magic numbers, look at the comments before gfc_build_array_type() in trans-types.c. The code within these defines should be the only code which knows the format of an array descriptor. Any code just needing to read obtain the bounds of an array should use gfc_conv_array_* rather than the following functions as these will return know constant values, and work with arrays which do not have descriptors. Don't forget to #undef these! */ #define DATA_FIELD 0 #define OFFSET_FIELD 1 #define DTYPE_FIELD 2 #define DIMENSION_FIELD 3 #define CAF_TOKEN_FIELD 4 #define STRIDE_SUBFIELD 0 #define LBOUND_SUBFIELD 1 #define UBOUND_SUBFIELD 2 /* This provides READ-ONLY access to the data field. The field itself doesn't have the proper type. */ tree gfc_conv_descriptor_data_get (tree desc) { tree field, type, t; type = TREE_TYPE (desc); gcc_assert (GFC_DESCRIPTOR_TYPE_P (type)); field = TYPE_FIELDS (type); gcc_assert (DATA_FIELD == 0); t = fold_build3_loc (input_location, COMPONENT_REF, TREE_TYPE (field), desc, field, NULL_TREE); t = fold_convert (GFC_TYPE_ARRAY_DATAPTR_TYPE (type), t); return t; } /* This provides WRITE access to the data field. TUPLES_P is true if we are generating tuples. This function gets called through the following macros: gfc_conv_descriptor_data_set gfc_conv_descriptor_data_set. */ void gfc_conv_descriptor_data_set (stmtblock_t *block, tree desc, tree value) { tree field, type, t; type = TREE_TYPE (desc); gcc_assert (GFC_DESCRIPTOR_TYPE_P (type)); field = TYPE_FIELDS (type); gcc_assert (DATA_FIELD == 0); t = fold_build3_loc (input_location, COMPONENT_REF, TREE_TYPE (field), desc, field, NULL_TREE); gfc_add_modify (block, t, fold_convert (TREE_TYPE (field), value)); } /* This provides address access to the data field. This should only be used by array allocation, passing this on to the runtime. */ tree gfc_conv_descriptor_data_addr (tree desc) { tree field, type, t; type = TREE_TYPE (desc); gcc_assert (GFC_DESCRIPTOR_TYPE_P (type)); field = TYPE_FIELDS (type); gcc_assert (DATA_FIELD == 0); t = fold_build3_loc (input_location, COMPONENT_REF, TREE_TYPE (field), desc, field, NULL_TREE); return gfc_build_addr_expr (NULL_TREE, t); } static tree gfc_conv_descriptor_offset (tree desc) { tree type; tree field; type = TREE_TYPE (desc); gcc_assert (GFC_DESCRIPTOR_TYPE_P (type)); field = gfc_advance_chain (TYPE_FIELDS (type), OFFSET_FIELD); gcc_assert (field != NULL_TREE && TREE_TYPE (field) == gfc_array_index_type); return fold_build3_loc (input_location, COMPONENT_REF, TREE_TYPE (field), desc, field, NULL_TREE); } tree gfc_conv_descriptor_offset_get (tree desc) { return gfc_conv_descriptor_offset (desc); } void gfc_conv_descriptor_offset_set (stmtblock_t *block, tree desc, tree value) { tree t = gfc_conv_descriptor_offset (desc); gfc_add_modify (block, t, fold_convert (TREE_TYPE (t), value)); } tree gfc_conv_descriptor_dtype (tree desc) { tree field; tree type; type = TREE_TYPE (desc); gcc_assert (GFC_DESCRIPTOR_TYPE_P (type)); field = gfc_advance_chain (TYPE_FIELDS (type), DTYPE_FIELD); gcc_assert (field != NULL_TREE && TREE_TYPE (field) == gfc_array_index_type); return fold_build3_loc (input_location, COMPONENT_REF, TREE_TYPE (field), desc, field, NULL_TREE); } tree gfc_conv_descriptor_rank (tree desc) { tree tmp; tree dtype; dtype = gfc_conv_descriptor_dtype (desc); tmp = build_int_cst (TREE_TYPE (dtype), GFC_DTYPE_RANK_MASK); tmp = fold_build2_loc (input_location, BIT_AND_EXPR, TREE_TYPE (dtype), dtype, tmp); return fold_convert (gfc_get_int_type (gfc_default_integer_kind), tmp); } tree gfc_get_descriptor_dimension (tree desc) { tree type, field; type = TREE_TYPE (desc); gcc_assert (GFC_DESCRIPTOR_TYPE_P (type)); field = gfc_advance_chain (TYPE_FIELDS (type), DIMENSION_FIELD); gcc_assert (field != NULL_TREE && TREE_CODE (TREE_TYPE (field)) == ARRAY_TYPE && TREE_CODE (TREE_TYPE (TREE_TYPE (field))) == RECORD_TYPE); return fold_build3_loc (input_location, COMPONENT_REF, TREE_TYPE (field), desc, field, NULL_TREE); } static tree gfc_conv_descriptor_dimension (tree desc, tree dim) { tree tmp; tmp = gfc_get_descriptor_dimension (desc); return gfc_build_array_ref (tmp, dim, NULL); } tree gfc_conv_descriptor_token (tree desc) { tree type; tree field; type = TREE_TYPE (desc); gcc_assert (GFC_DESCRIPTOR_TYPE_P (type)); gcc_assert (flag_coarray == GFC_FCOARRAY_LIB); field = gfc_advance_chain (TYPE_FIELDS (type), CAF_TOKEN_FIELD); /* Should be a restricted pointer - except in the finalization wrapper. */ gcc_assert (field != NULL_TREE && (TREE_TYPE (field) == prvoid_type_node || TREE_TYPE (field) == pvoid_type_node)); return fold_build3_loc (input_location, COMPONENT_REF, TREE_TYPE (field), desc, field, NULL_TREE); } static tree gfc_conv_descriptor_stride (tree desc, tree dim) { tree tmp; tree field; tmp = gfc_conv_descriptor_dimension (desc, dim); field = TYPE_FIELDS (TREE_TYPE (tmp)); field = gfc_advance_chain (field, STRIDE_SUBFIELD); gcc_assert (field != NULL_TREE && TREE_TYPE (field) == gfc_array_index_type); tmp = fold_build3_loc (input_location, COMPONENT_REF, TREE_TYPE (field), tmp, field, NULL_TREE); return tmp; } tree gfc_conv_descriptor_stride_get (tree desc, tree dim) { tree type = TREE_TYPE (desc); gcc_assert (GFC_DESCRIPTOR_TYPE_P (type)); if (integer_zerop (dim) && (GFC_TYPE_ARRAY_AKIND (type) == GFC_ARRAY_ALLOCATABLE ||GFC_TYPE_ARRAY_AKIND (type) == GFC_ARRAY_ASSUMED_SHAPE_CONT ||GFC_TYPE_ARRAY_AKIND (type) == GFC_ARRAY_ASSUMED_RANK_CONT ||GFC_TYPE_ARRAY_AKIND (type) == GFC_ARRAY_POINTER_CONT)) return gfc_index_one_node; return gfc_conv_descriptor_stride (desc, dim); } void gfc_conv_descriptor_stride_set (stmtblock_t *block, tree desc, tree dim, tree value) { tree t = gfc_conv_descriptor_stride (desc, dim); gfc_add_modify (block, t, fold_convert (TREE_TYPE (t), value)); } static tree gfc_conv_descriptor_lbound (tree desc, tree dim) { tree tmp; tree field; tmp = gfc_conv_descriptor_dimension (desc, dim); field = TYPE_FIELDS (TREE_TYPE (tmp)); field = gfc_advance_chain (field, LBOUND_SUBFIELD); gcc_assert (field != NULL_TREE && TREE_TYPE (field) == gfc_array_index_type); tmp = fold_build3_loc (input_location, COMPONENT_REF, TREE_TYPE (field), tmp, field, NULL_TREE); return tmp; } tree gfc_conv_descriptor_lbound_get (tree desc, tree dim) { return gfc_conv_descriptor_lbound (desc, dim); } void gfc_conv_descriptor_lbound_set (stmtblock_t *block, tree desc, tree dim, tree value) { tree t = gfc_conv_descriptor_lbound (desc, dim); gfc_add_modify (block, t, fold_convert (TREE_TYPE (t), value)); } static tree gfc_conv_descriptor_ubound (tree desc, tree dim) { tree tmp; tree field; tmp = gfc_conv_descriptor_dimension (desc, dim); field = TYPE_FIELDS (TREE_TYPE (tmp)); field = gfc_advance_chain (field, UBOUND_SUBFIELD); gcc_assert (field != NULL_TREE && TREE_TYPE (field) == gfc_array_index_type); tmp = fold_build3_loc (input_location, COMPONENT_REF, TREE_TYPE (field), tmp, field, NULL_TREE); return tmp; } tree gfc_conv_descriptor_ubound_get (tree desc, tree dim) { return gfc_conv_descriptor_ubound (desc, dim); } void gfc_conv_descriptor_ubound_set (stmtblock_t *block, tree desc, tree dim, tree value) { tree t = gfc_conv_descriptor_ubound (desc, dim); gfc_add_modify (block, t, fold_convert (TREE_TYPE (t), value)); } /* Build a null array descriptor constructor. */ tree gfc_build_null_descriptor (tree type) { tree field; tree tmp; gcc_assert (GFC_DESCRIPTOR_TYPE_P (type)); gcc_assert (DATA_FIELD == 0); field = TYPE_FIELDS (type); /* Set a NULL data pointer. */ tmp = build_constructor_single (type, field, null_pointer_node); TREE_CONSTANT (tmp) = 1; /* All other fields are ignored. */ return tmp; } /* Modify a descriptor such that the lbound of a given dimension is the value specified. This also updates ubound and offset accordingly. */ void gfc_conv_shift_descriptor_lbound (stmtblock_t* block, tree desc, int dim, tree new_lbound) { tree offs, ubound, lbound, stride; tree diff, offs_diff; new_lbound = fold_convert (gfc_array_index_type, new_lbound); offs = gfc_conv_descriptor_offset_get (desc); lbound = gfc_conv_descriptor_lbound_get (desc, gfc_rank_cst[dim]); ubound = gfc_conv_descriptor_ubound_get (desc, gfc_rank_cst[dim]); stride = gfc_conv_descriptor_stride_get (desc, gfc_rank_cst[dim]); /* Get difference (new - old) by which to shift stuff. */ diff = fold_build2_loc (input_location, MINUS_EXPR, gfc_array_index_type, new_lbound, lbound); /* Shift ubound and offset accordingly. This has to be done before updating the lbound, as they depend on the lbound expression! */ ubound = fold_build2_loc (input_location, PLUS_EXPR, gfc_array_index_type, ubound, diff); gfc_conv_descriptor_ubound_set (block, desc, gfc_rank_cst[dim], ubound); offs_diff = fold_build2_loc (input_location, MULT_EXPR, gfc_array_index_type, diff, stride); offs = fold_build2_loc (input_location, MINUS_EXPR, gfc_array_index_type, offs, offs_diff); gfc_conv_descriptor_offset_set (block, desc, offs); /* Finally set lbound to value we want. */ gfc_conv_descriptor_lbound_set (block, desc, gfc_rank_cst[dim], new_lbound); } /* Cleanup those #defines. */ #undef DATA_FIELD #undef OFFSET_FIELD #undef DTYPE_FIELD #undef DIMENSION_FIELD #undef CAF_TOKEN_FIELD #undef STRIDE_SUBFIELD #undef LBOUND_SUBFIELD #undef UBOUND_SUBFIELD /* Mark a SS chain as used. Flags specifies in which loops the SS is used. flags & 1 = Main loop body. flags & 2 = temp copy loop. */ void gfc_mark_ss_chain_used (gfc_ss * ss, unsigned flags) { for (; ss != gfc_ss_terminator; ss = ss->next) ss->info->useflags = flags; } /* Free a gfc_ss chain. */ void gfc_free_ss_chain (gfc_ss * ss) { gfc_ss *next; while (ss != gfc_ss_terminator) { gcc_assert (ss != NULL); next = ss->next; gfc_free_ss (ss); ss = next; } } static void free_ss_info (gfc_ss_info *ss_info) { int n; ss_info->refcount--; if (ss_info->refcount > 0) return; gcc_assert (ss_info->refcount == 0); switch (ss_info->type) { case GFC_SS_SECTION: for (n = 0; n < GFC_MAX_DIMENSIONS; n++) if (ss_info->data.array.subscript[n]) gfc_free_ss_chain (ss_info->data.array.subscript[n]); break; default: break; } free (ss_info); } /* Free a SS. */ void gfc_free_ss (gfc_ss * ss) { free_ss_info (ss->info); free (ss); } /* Creates and initializes an array type gfc_ss struct. */ gfc_ss * gfc_get_array_ss (gfc_ss *next, gfc_expr *expr, int dimen, gfc_ss_type type) { gfc_ss *ss; gfc_ss_info *ss_info; int i; ss_info = gfc_get_ss_info (); ss_info->refcount++; ss_info->type = type; ss_info->expr = expr; ss = gfc_get_ss (); ss->info = ss_info; ss->next = next; ss->dimen = dimen; for (i = 0; i < ss->dimen; i++) ss->dim[i] = i; return ss; } /* Creates and initializes a temporary type gfc_ss struct. */ gfc_ss * gfc_get_temp_ss (tree type, tree string_length, int dimen) { gfc_ss *ss; gfc_ss_info *ss_info; int i; ss_info = gfc_get_ss_info (); ss_info->refcount++; ss_info->type = GFC_SS_TEMP; ss_info->string_length = string_length; ss_info->data.temp.type = type; ss = gfc_get_ss (); ss->info = ss_info; ss->next = gfc_ss_terminator; ss->dimen = dimen; for (i = 0; i < ss->dimen; i++) ss->dim[i] = i; return ss; } /* Creates and initializes a scalar type gfc_ss struct. */ gfc_ss * gfc_get_scalar_ss (gfc_ss *next, gfc_expr *expr) { gfc_ss *ss; gfc_ss_info *ss_info; ss_info = gfc_get_ss_info (); ss_info->refcount++; ss_info->type = GFC_SS_SCALAR; ss_info->expr = expr; ss = gfc_get_ss (); ss->info = ss_info; ss->next = next; return ss; } /* Free all the SS associated with a loop. */ void gfc_cleanup_loop (gfc_loopinfo * loop) { gfc_loopinfo *loop_next, **ploop; gfc_ss *ss; gfc_ss *next; ss = loop->ss; while (ss != gfc_ss_terminator) { gcc_assert (ss != NULL); next = ss->loop_chain; gfc_free_ss (ss); ss = next; } /* Remove reference to self in the parent loop. */ if (loop->parent) for (ploop = &loop->parent->nested; *ploop; ploop = &(*ploop)->next) if (*ploop == loop) { *ploop = loop->next; break; } /* Free non-freed nested loops. */ for (loop = loop->nested; loop; loop = loop_next) { loop_next = loop->next; gfc_cleanup_loop (loop); free (loop); } } static void set_ss_loop (gfc_ss *ss, gfc_loopinfo *loop) { int n; for (; ss != gfc_ss_terminator; ss = ss->next) { ss->loop = loop; if (ss->info->type == GFC_SS_SCALAR || ss->info->type == GFC_SS_REFERENCE || ss->info->type == GFC_SS_TEMP) continue; for (n = 0; n < GFC_MAX_DIMENSIONS; n++) if (ss->info->data.array.subscript[n] != NULL) set_ss_loop (ss->info->data.array.subscript[n], loop); } } /* Associate a SS chain with a loop. */ void gfc_add_ss_to_loop (gfc_loopinfo * loop, gfc_ss * head) { gfc_ss *ss; gfc_loopinfo *nested_loop; if (head == gfc_ss_terminator) return; set_ss_loop (head, loop); ss = head; for (; ss && ss != gfc_ss_terminator; ss = ss->next) { if (ss->nested_ss) { nested_loop = ss->nested_ss->loop; /* More than one ss can belong to the same loop. Hence, we add the loop to the chain only if it is different from the previously added one, to avoid duplicate nested loops. */ if (nested_loop != loop->nested) { gcc_assert (nested_loop->parent == NULL); nested_loop->parent = loop; gcc_assert (nested_loop->next == NULL); nested_loop->next = loop->nested; loop->nested = nested_loop; } else gcc_assert (nested_loop->parent == loop); } if (ss->next == gfc_ss_terminator) ss->loop_chain = loop->ss; else ss->loop_chain = ss->next; } gcc_assert (ss == gfc_ss_terminator); loop->ss = head; } /* Generate an initializer for a static pointer or allocatable array. */ void gfc_trans_static_array_pointer (gfc_symbol * sym) { tree type; gcc_assert (TREE_STATIC (sym->backend_decl)); /* Just zero the data member. */ type = TREE_TYPE (sym->backend_decl); DECL_INITIAL (sym->backend_decl) = gfc_build_null_descriptor (type); } /* If the bounds of SE's loop have not yet been set, see if they can be determined from array spec AS, which is the array spec of a called function. MAPPING maps the callee's dummy arguments to the values that the caller is passing. Add any initialization and finalization code to SE. */ void gfc_set_loop_bounds_from_array_spec (gfc_interface_mapping * mapping, gfc_se * se, gfc_array_spec * as) { int n, dim, total_dim; gfc_se tmpse; gfc_ss *ss; tree lower; tree upper; tree tmp; total_dim = 0; if (!as || as->type != AS_EXPLICIT) return; for (ss = se->ss; ss; ss = ss->parent) { total_dim += ss->loop->dimen; for (n = 0; n < ss->loop->dimen; n++) { /* The bound is known, nothing to do. */ if (ss->loop->to[n] != NULL_TREE) continue; dim = ss->dim[n]; gcc_assert (dim < as->rank); gcc_assert (ss->loop->dimen <= as->rank); /* Evaluate the lower bound. */ gfc_init_se (&tmpse, NULL); gfc_apply_interface_mapping (mapping, &tmpse, as->lower[dim]); gfc_add_block_to_block (&se->pre, &tmpse.pre); gfc_add_block_to_block (&se->post, &tmpse.post); lower = fold_convert (gfc_array_index_type, tmpse.expr); /* ...and the upper bound. */ gfc_init_se (&tmpse, NULL); gfc_apply_interface_mapping (mapping, &tmpse, as->upper[dim]); gfc_add_block_to_block (&se->pre, &tmpse.pre); gfc_add_block_to_block (&se->post, &tmpse.post); upper = fold_convert (gfc_array_index_type, tmpse.expr); /* Set the upper bound of the loop to UPPER - LOWER. */ tmp = fold_build2_loc (input_location, MINUS_EXPR, gfc_array_index_type, upper, lower); tmp = gfc_evaluate_now (tmp, &se->pre); ss->loop->to[n] = tmp; } } gcc_assert (total_dim == as->rank); } /* Generate code to allocate an array temporary, or create a variable to hold the data. If size is NULL, zero the descriptor so that the callee will allocate the array. If DEALLOC is true, also generate code to free the array afterwards. If INITIAL is not NULL, it is packed using internal_pack and the result used as data instead of allocating a fresh, unitialized area of memory. Initialization code is added to PRE and finalization code to POST. DYNAMIC is true if the caller may want to extend the array later using realloc. This prevents us from putting the array on the stack. */ static void gfc_trans_allocate_array_storage (stmtblock_t * pre, stmtblock_t * post, gfc_array_info * info, tree size, tree nelem, tree initial, bool dynamic, bool dealloc) { tree tmp; tree desc; bool onstack; desc = info->descriptor; info->offset = gfc_index_zero_node; if (size == NULL_TREE || integer_zerop (size)) { /* A callee allocated array. */ gfc_conv_descriptor_data_set (pre, desc, null_pointer_node); onstack = FALSE; } else { /* Allocate the temporary. */ onstack = !dynamic && initial == NULL_TREE && (flag_stack_arrays || gfc_can_put_var_on_stack (size)); if (onstack) { /* Make a temporary variable to hold the data. */ tmp = fold_build2_loc (input_location, MINUS_EXPR, TREE_TYPE (nelem), nelem, gfc_index_one_node); tmp = gfc_evaluate_now (tmp, pre); tmp = build_range_type (gfc_array_index_type, gfc_index_zero_node, tmp); tmp = build_array_type (gfc_get_element_type (TREE_TYPE (desc)), tmp); tmp = gfc_create_var (tmp, "A"); /* If we're here only because of -fstack-arrays we have to emit a DECL_EXPR to make the gimplifier emit alloca calls. */ if (!gfc_can_put_var_on_stack (size)) gfc_add_expr_to_block (pre, fold_build1_loc (input_location, DECL_EXPR, TREE_TYPE (tmp), tmp)); tmp = gfc_build_addr_expr (NULL_TREE, tmp); gfc_conv_descriptor_data_set (pre, desc, tmp); } else { /* Allocate memory to hold the data or call internal_pack. */ if (initial == NULL_TREE) { tmp = gfc_call_malloc (pre, NULL, size); tmp = gfc_evaluate_now (tmp, pre); } else { tree packed; tree source_data; tree was_packed; stmtblock_t do_copying; tmp = TREE_TYPE (initial); /* Pointer to descriptor. */ gcc_assert (TREE_CODE (tmp) == POINTER_TYPE); tmp = TREE_TYPE (tmp); /* The descriptor itself. */ tmp = gfc_get_element_type (tmp); gcc_assert (tmp == gfc_get_element_type (TREE_TYPE (desc))); packed = gfc_create_var (build_pointer_type (tmp), "data"); tmp = build_call_expr_loc (input_location, gfor_fndecl_in_pack, 1, initial); tmp = fold_convert (TREE_TYPE (packed), tmp); gfc_add_modify (pre, packed, tmp); tmp = build_fold_indirect_ref_loc (input_location, initial); source_data = gfc_conv_descriptor_data_get (tmp); /* internal_pack may return source->data without any allocation or copying if it is already packed. If that's the case, we need to allocate and copy manually. */ gfc_start_block (&do_copying); tmp = gfc_call_malloc (&do_copying, NULL, size); tmp = fold_convert (TREE_TYPE (packed), tmp); gfc_add_modify (&do_copying, packed, tmp); tmp = gfc_build_memcpy_call (packed, source_data, size); gfc_add_expr_to_block (&do_copying, tmp); was_packed = fold_build2_loc (input_location, EQ_EXPR, boolean_type_node, packed, source_data); tmp = gfc_finish_block (&do_copying); tmp = build3_v (COND_EXPR, was_packed, tmp, build_empty_stmt (input_location)); gfc_add_expr_to_block (pre, tmp); tmp = fold_convert (pvoid_type_node, packed); } gfc_conv_descriptor_data_set (pre, desc, tmp); } } info->data = gfc_conv_descriptor_data_get (desc); /* The offset is zero because we create temporaries with a zero lower bound. */ gfc_conv_descriptor_offset_set (pre, desc, gfc_index_zero_node); if (dealloc && !onstack) { /* Free the temporary. */ tmp = gfc_conv_descriptor_data_get (desc); tmp = gfc_call_free (fold_convert (pvoid_type_node, tmp)); gfc_add_expr_to_block (post, tmp); } } /* Get the scalarizer array dimension corresponding to actual array dimension given by ARRAY_DIM. For example, if SS represents the array ref a(1,:,:,1), it is a bidimensional scalarizer array, and the result would be 0 for ARRAY_DIM=1, and 1 for ARRAY_DIM=2. If SS represents transpose(a(:,1,1,:)), it is again a bidimensional scalarizer array, and the result would be 1 for ARRAY_DIM=0 and 0 for ARRAY_DIM=3. If SS represents sum(a(:,:,:,1), dim=1), it is a 2+1-dimensional scalarizer array. If called on the inner ss, the result would be respectively 0,1,2 for ARRAY_DIM=0,1,2. If called on the outer ss, the result would be 0,1 for ARRAY_DIM=1,2. */ static int get_scalarizer_dim_for_array_dim (gfc_ss *ss, int array_dim) { int array_ref_dim; int n; array_ref_dim = 0; for (; ss; ss = ss->parent) for (n = 0; n < ss->dimen; n++) if (ss->dim[n] < array_dim) array_ref_dim++; return array_ref_dim; } static gfc_ss * innermost_ss (gfc_ss *ss) { while (ss->nested_ss != NULL) ss = ss->nested_ss; return ss; } /* Get the array reference dimension corresponding to the given loop dimension. It is different from the true array dimension given by the dim array in the case of a partial array reference (i.e. a(:,:,1,:) for example) It is different from the loop dimension in the case of a transposed array. */ static int get_array_ref_dim_for_loop_dim (gfc_ss *ss, int loop_dim) { return get_scalarizer_dim_for_array_dim (innermost_ss (ss), ss->dim[loop_dim]); } /* Generate code to create and initialize the descriptor for a temporary array. This is used for both temporaries needed by the scalarizer, and functions returning arrays. Adjusts the loop variables to be zero-based, and calculates the loop bounds for callee allocated arrays. Allocate the array unless it's callee allocated (we have a callee allocated array if 'callee_alloc' is true, or if loop->to[n] is NULL_TREE for any n). Also fills in the descriptor, data and offset fields of info if known. Returns the size of the array, or NULL for a callee allocated array. 'eltype' == NULL signals that the temporary should be a class object. The 'initial' expression is used to obtain the size of the dynamic type; otherwise the allocation and initialization proceeds as for any other expression PRE, POST, INITIAL, DYNAMIC and DEALLOC are as for gfc_trans_allocate_array_storage. */ tree gfc_trans_create_temp_array (stmtblock_t * pre, stmtblock_t * post, gfc_ss * ss, tree eltype, tree initial, bool dynamic, bool dealloc, bool callee_alloc, locus * where) { gfc_loopinfo *loop; gfc_ss *s; gfc_array_info *info; tree from[GFC_MAX_DIMENSIONS], to[GFC_MAX_DIMENSIONS]; tree type; tree desc; tree tmp; tree size; tree nelem; tree cond; tree or_expr; tree class_expr = NULL_TREE; int n, dim, tmp_dim; int total_dim = 0; /* This signals a class array for which we need the size of the dynamic type. Generate an eltype and then the class expression. */ if (eltype == NULL_TREE && initial) { gcc_assert (POINTER_TYPE_P (TREE_TYPE (initial))); class_expr = build_fold_indirect_ref_loc (input_location, initial); eltype = TREE_TYPE (class_expr); eltype = gfc_get_element_type (eltype); /* Obtain the structure (class) expression. */ class_expr = TREE_OPERAND (class_expr, 0); gcc_assert (class_expr); } memset (from, 0, sizeof (from)); memset (to, 0, sizeof (to)); info = &ss->info->data.array; gcc_assert (ss->dimen > 0); gcc_assert (ss->loop->dimen == ss->dimen); if (warn_array_temporaries && where) gfc_warning (OPT_Warray_temporaries, "Creating array temporary at %L", where); /* Set the lower bound to zero. */ for (s = ss; s; s = s->parent) { loop = s->loop; total_dim += loop->dimen; for (n = 0; n < loop->dimen; n++) { dim = s->dim[n]; /* Callee allocated arrays may not have a known bound yet. */ if (loop->to[n]) loop->to[n] = gfc_evaluate_now ( fold_build2_loc (input_location, MINUS_EXPR, gfc_array_index_type, loop->to[n], loop->from[n]), pre); loop->from[n] = gfc_index_zero_node; /* We have just changed the loop bounds, we must clear the corresponding specloop, so that delta calculation is not skipped later in gfc_set_delta. */ loop->specloop[n] = NULL; /* We are constructing the temporary's descriptor based on the loop dimensions. As the dimensions may be accessed in arbitrary order (think of transpose) the size taken from the n'th loop may not map to the n'th dimension of the array. We need to reconstruct loop infos in the right order before using it to set the descriptor bounds. */ tmp_dim = get_scalarizer_dim_for_array_dim (ss, dim); from[tmp_dim] = loop->from[n]; to[tmp_dim] = loop->to[n]; info->delta[dim] = gfc_index_zero_node; info->start[dim] = gfc_index_zero_node; info->end[dim] = gfc_index_zero_node; info->stride[dim] = gfc_index_one_node; } } /* Initialize the descriptor. */ type = gfc_get_array_type_bounds (eltype, total_dim, 0, from, to, 1, GFC_ARRAY_UNKNOWN, true); desc = gfc_create_var (type, "atmp"); GFC_DECL_PACKED_ARRAY (desc) = 1; info->descriptor = desc; size = gfc_index_one_node; /* Fill in the array dtype. */ tmp = gfc_conv_descriptor_dtype (desc); gfc_add_modify (pre, tmp, gfc_get_dtype (TREE_TYPE (desc))); /* Fill in the bounds and stride. This is a packed array, so: size = 1; for (n = 0; n < rank; n++) { stride[n] = size delta = ubound[n] + 1 - lbound[n]; size = size * delta; } size = size * sizeof(element); */ or_expr = NULL_TREE; /* If there is at least one null loop->to[n], it is a callee allocated array. */ for (n = 0; n < total_dim; n++) if (to[n] == NULL_TREE) { size = NULL_TREE; break; } if (size == NULL_TREE) for (s = ss; s; s = s->parent) for (n = 0; n < s->loop->dimen; n++) { dim = get_scalarizer_dim_for_array_dim (ss, s->dim[n]); /* For a callee allocated array express the loop bounds in terms of the descriptor fields. */ tmp = fold_build2_loc (input_location, MINUS_EXPR, gfc_array_index_type, gfc_conv_descriptor_ubound_get (desc, gfc_rank_cst[dim]), gfc_conv_descriptor_lbound_get (desc, gfc_rank_cst[dim])); s->loop->to[n] = tmp; } else { for (n = 0; n < total_dim; n++) { /* Store the stride and bound components in the descriptor. */ gfc_conv_descriptor_stride_set (pre, desc, gfc_rank_cst[n], size); gfc_conv_descriptor_lbound_set (pre, desc, gfc_rank_cst[n], gfc_index_zero_node); gfc_conv_descriptor_ubound_set (pre, desc, gfc_rank_cst[n], to[n]); tmp = fold_build2_loc (input_location, PLUS_EXPR, gfc_array_index_type, to[n], gfc_index_one_node); /* Check whether the size for this dimension is negative. */ cond = fold_build2_loc (input_location, LE_EXPR, boolean_type_node, tmp, gfc_index_zero_node); cond = gfc_evaluate_now (cond, pre); if (n == 0) or_expr = cond; else or_expr = fold_build2_loc (input_location, TRUTH_OR_EXPR, boolean_type_node, or_expr, cond); size = fold_build2_loc (input_location, MULT_EXPR, gfc_array_index_type, size, tmp); size = gfc_evaluate_now (size, pre); } } /* Get the size of the array. */ if (size && !callee_alloc) { tree elemsize; /* If or_expr is true, then the extent in at least one dimension is zero and the size is set to zero. */ size = fold_build3_loc (input_location, COND_EXPR, gfc_array_index_type, or_expr, gfc_index_zero_node, size); nelem = size; if (class_expr == NULL_TREE) elemsize = fold_convert (gfc_array_index_type, TYPE_SIZE_UNIT (gfc_get_element_type (type))); else elemsize = gfc_vtable_size_get (class_expr); size = fold_build2_loc (input_location, MULT_EXPR, gfc_array_index_type, size, elemsize); } else { nelem = size; size = NULL_TREE; } gfc_trans_allocate_array_storage (pre, post, info, size, nelem, initial, dynamic, dealloc); while (ss->parent) ss = ss->parent; if (ss->dimen > ss->loop->temp_dim) ss->loop->temp_dim = ss->dimen; return size; } /* Return the number of iterations in a loop that starts at START, ends at END, and has step STEP. */ static tree gfc_get_iteration_count (tree start, tree end, tree step) { tree tmp; tree type; type = TREE_TYPE (step); tmp = fold_build2_loc (input_location, MINUS_EXPR, type, end, start); tmp = fold_build2_loc (input_location, FLOOR_DIV_EXPR, type, tmp, step); tmp = fold_build2_loc (input_location, PLUS_EXPR, type, tmp, build_int_cst (type, 1)); tmp = fold_build2_loc (input_location, MAX_EXPR, type, tmp, build_int_cst (type, 0)); return fold_convert (gfc_array_index_type, tmp); } /* Extend the data in array DESC by EXTRA elements. */ static void gfc_grow_array (stmtblock_t * pblock, tree desc, tree extra) { tree arg0, arg1; tree tmp; tree size; tree ubound; if (integer_zerop (extra)) return; ubound = gfc_conv_descriptor_ubound_get (desc, gfc_rank_cst[0]); /* Add EXTRA to the upper bound. */ tmp = fold_build2_loc (input_location, PLUS_EXPR, gfc_array_index_type, ubound, extra); gfc_conv_descriptor_ubound_set (pblock, desc, gfc_rank_cst[0], tmp); /* Get the value of the current data pointer. */ arg0 = gfc_conv_descriptor_data_get (desc); /* Calculate the new array size. */ size = TYPE_SIZE_UNIT (gfc_get_element_type (TREE_TYPE (desc))); tmp = fold_build2_loc (input_location, PLUS_EXPR, gfc_array_index_type, ubound, gfc_index_one_node); arg1 = fold_build2_loc (input_location, MULT_EXPR, size_type_node, fold_convert (size_type_node, tmp), fold_convert (size_type_node, size)); /* Call the realloc() function. */ tmp = gfc_call_realloc (pblock, arg0, arg1); gfc_conv_descriptor_data_set (pblock, desc, tmp); } /* Return true if the bounds of iterator I can only be determined at run time. */ static inline bool gfc_iterator_has_dynamic_bounds (gfc_iterator * i) { return (i->start->expr_type != EXPR_CONSTANT || i->end->expr_type != EXPR_CONSTANT || i->step->expr_type != EXPR_CONSTANT); } /* Split the size of constructor element EXPR into the sum of two terms, one of which can be determined at compile time and one of which must be calculated at run time. Set *SIZE to the former and return true if the latter might be nonzero. */ static bool gfc_get_array_constructor_element_size (mpz_t * size, gfc_expr * expr) { if (expr->expr_type == EXPR_ARRAY) return gfc_get_array_constructor_size (size, expr->value.constructor); else if (expr->rank > 0) { /* Calculate everything at run time. */ mpz_set_ui (*size, 0); return true; } else { /* A single element. */ mpz_set_ui (*size, 1); return false; } } /* Like gfc_get_array_constructor_element_size, but applied to the whole of array constructor C. */ static bool gfc_get_array_constructor_size (mpz_t * size, gfc_constructor_base base) { gfc_constructor *c; gfc_iterator *i; mpz_t val; mpz_t len; bool dynamic; mpz_set_ui (*size, 0); mpz_init (len); mpz_init (val); dynamic = false; for (c = gfc_constructor_first (base); c; c = gfc_constructor_next (c)) { i = c->iterator; if (i && gfc_iterator_has_dynamic_bounds (i)) dynamic = true; else { dynamic |= gfc_get_array_constructor_element_size (&len, c->expr); if (i) { /* Multiply the static part of the element size by the number of iterations. */ mpz_sub (val, i->end->value.integer, i->start->value.integer); mpz_fdiv_q (val, val, i->step->value.integer); mpz_add_ui (val, val, 1); if (mpz_sgn (val) > 0) mpz_mul (len, len, val); else mpz_set_ui (len, 0); } mpz_add (*size, *size, len); } } mpz_clear (len); mpz_clear (val); return dynamic; } /* Make sure offset is a variable. */ static void gfc_put_offset_into_var (stmtblock_t * pblock, tree * poffset, tree * offsetvar) { /* We should have already created the offset variable. We cannot create it here because we may be in an inner scope. */ gcc_assert (*offsetvar != NULL_TREE); gfc_add_modify (pblock, *offsetvar, *poffset); *poffset = *offsetvar; TREE_USED (*offsetvar) = 1; } /* Variables needed for bounds-checking. */ static bool first_len; static tree first_len_val; static bool typespec_chararray_ctor; static void gfc_trans_array_ctor_element (stmtblock_t * pblock, tree desc, tree offset, gfc_se * se, gfc_expr * expr) { tree tmp; gfc_conv_expr (se, expr); /* Store the value. */ tmp = build_fold_indirect_ref_loc (input_location, gfc_conv_descriptor_data_get (desc)); tmp = gfc_build_array_ref (tmp, offset, NULL); if (expr->ts.type == BT_CHARACTER) { int i = gfc_validate_kind (BT_CHARACTER, expr->ts.kind, false); tree esize; esize = size_in_bytes (gfc_get_element_type (TREE_TYPE (desc))); esize = fold_convert (gfc_charlen_type_node, esize); esize = fold_build2_loc (input_location, TRUNC_DIV_EXPR, gfc_charlen_type_node, esize, build_int_cst (gfc_charlen_type_node, gfc_character_kinds[i].bit_size / 8)); gfc_conv_string_parameter (se); if (POINTER_TYPE_P (TREE_TYPE (tmp))) { /* The temporary is an array of pointers. */ se->expr = fold_convert (TREE_TYPE (tmp), se->expr); gfc_add_modify (&se->pre, tmp, se->expr); } else { /* The temporary is an array of string values. */ tmp = gfc_build_addr_expr (gfc_get_pchar_type (expr->ts.kind), tmp); /* We know the temporary and the value will be the same length, so can use memcpy. */ gfc_trans_string_copy (&se->pre, esize, tmp, expr->ts.kind, se->string_length, se->expr, expr->ts.kind); } if ((gfc_option.rtcheck & GFC_RTCHECK_BOUNDS) && !typespec_chararray_ctor) { if (first_len) { gfc_add_modify (&se->pre, first_len_val, se->string_length); first_len = false; } else { /* Verify that all constructor elements are of the same length. */ tree cond = fold_build2_loc (input_location, NE_EXPR, boolean_type_node, first_len_val, se->string_length); gfc_trans_runtime_check (true, false, cond, &se->pre, &expr->where, "Different CHARACTER lengths (%ld/%ld) in array constructor", fold_convert (long_integer_type_node, first_len_val), fold_convert (long_integer_type_node, se->string_length)); } } } else { /* TODO: Should the frontend already have done this conversion? */ se->expr = fold_convert (TREE_TYPE (tmp), se->expr); gfc_add_modify (&se->pre, tmp, se->expr); } gfc_add_block_to_block (pblock, &se->pre); gfc_add_block_to_block (pblock, &se->post); } /* Add the contents of an array to the constructor. DYNAMIC is as for gfc_trans_array_constructor_value. */ static void gfc_trans_array_constructor_subarray (stmtblock_t * pblock, tree type ATTRIBUTE_UNUSED, tree desc, gfc_expr * expr, tree * poffset, tree * offsetvar, bool dynamic) { gfc_se se; gfc_ss *ss; gfc_loopinfo loop; stmtblock_t body; tree tmp; tree size; int n; /* We need this to be a variable so we can increment it. */ gfc_put_offset_into_var (pblock, poffset, offsetvar); gfc_init_se (&se, NULL); /* Walk the array expression. */ ss = gfc_walk_expr (expr); gcc_assert (ss != gfc_ss_terminator); /* Initialize the scalarizer. */ gfc_init_loopinfo (&loop); gfc_add_ss_to_loop (&loop, ss); /* Initialize the loop. */ gfc_conv_ss_startstride (&loop); gfc_conv_loop_setup (&loop, &expr->where); /* Make sure the constructed array has room for the new data. */ if (dynamic) { /* Set SIZE to the total number of elements in the subarray. */ size = gfc_index_one_node; for (n = 0; n < loop.dimen; n++) { tmp = gfc_get_iteration_count (loop.from[n], loop.to[n], gfc_index_one_node); size = fold_build2_loc (input_location, MULT_EXPR, gfc_array_index_type, size, tmp); } /* Grow the constructed array by SIZE elements. */ gfc_grow_array (&loop.pre, desc, size); } /* Make the loop body. */ gfc_mark_ss_chain_used (ss, 1); gfc_start_scalarized_body (&loop, &body); gfc_copy_loopinfo_to_se (&se, &loop); se.ss = ss; gfc_trans_array_ctor_element (&body, desc, *poffset, &se, expr); gcc_assert (se.ss == gfc_ss_terminator); /* Increment the offset. */ tmp = fold_build2_loc (input_location, PLUS_EXPR, gfc_array_index_type, *poffset, gfc_index_one_node); gfc_add_modify (&body, *poffset, tmp); /* Finish the loop. */ gfc_trans_scalarizing_loops (&loop, &body); gfc_add_block_to_block (&loop.pre, &loop.post); tmp = gfc_finish_block (&loop.pre); gfc_add_expr_to_block (pblock, tmp); gfc_cleanup_loop (&loop); } /* Assign the values to the elements of an array constructor. DYNAMIC is true if descriptor DESC only contains enough data for the static size calculated by gfc_get_array_constructor_size. When true, memory for the dynamic parts must be allocated using realloc. */ static void gfc_trans_array_constructor_value (stmtblock_t * pblock, tree type, tree desc, gfc_constructor_base base, tree * poffset, tree * offsetvar, bool dynamic) { tree tmp; tree start = NULL_TREE; tree end = NULL_TREE; tree step = NULL_TREE; stmtblock_t body; gfc_se se; mpz_t size; gfc_constructor *c; tree shadow_loopvar = NULL_TREE; gfc_saved_var saved_loopvar; mpz_init (size); for (c = gfc_constructor_first (base); c; c = gfc_constructor_next (c)) { /* If this is an iterator or an array, the offset must be a variable. */ if ((c->iterator || c->expr->rank > 0) && INTEGER_CST_P (*poffset)) gfc_put_offset_into_var (pblock, poffset, offsetvar); /* Shadowing the iterator avoids changing its value and saves us from keeping track of it. Further, it makes sure that there's always a backend-decl for the symbol, even if there wasn't one before, e.g. in the case of an iterator that appears in a specification expression in an interface mapping. */ if (c->iterator) { gfc_symbol *sym; tree type; /* Evaluate loop bounds before substituting the loop variable in case they depend on it. Such a case is invalid, but it is not more expensive to do the right thing here. See PR 44354. */ gfc_init_se (&se, NULL); gfc_conv_expr_val (&se, c->iterator->start); gfc_add_block_to_block (pblock, &se.pre); start = gfc_evaluate_now (se.expr, pblock); gfc_init_se (&se, NULL); gfc_conv_expr_val (&se, c->iterator->end); gfc_add_block_to_block (pblock, &se.pre); end = gfc_evaluate_now (se.expr, pblock); gfc_init_se (&se, NULL); gfc_conv_expr_val (&se, c->iterator->step); gfc_add_block_to_block (pblock, &se.pre); step = gfc_evaluate_now (se.expr, pblock); sym = c->iterator->var->symtree->n.sym; type = gfc_typenode_for_spec (&sym->ts); shadow_loopvar = gfc_create_var (type, "shadow_loopvar"); gfc_shadow_sym (sym, shadow_loopvar, &saved_loopvar); } gfc_start_block (&body); if (c->expr->expr_type == EXPR_ARRAY) { /* Array constructors can be nested. */ gfc_trans_array_constructor_value (&body, type, desc, c->expr->value.constructor, poffset, offsetvar, dynamic); } else if (c->expr->rank > 0) { gfc_trans_array_constructor_subarray (&body, type, desc, c->expr, poffset, offsetvar, dynamic); } else { /* This code really upsets the gimplifier so don't bother for now. */ gfc_constructor *p; HOST_WIDE_INT n; HOST_WIDE_INT size; p = c; n = 0; while (p && !(p->iterator || p->expr->expr_type != EXPR_CONSTANT)) { p = gfc_constructor_next (p); n++; } if (n < 4) { /* Scalar values. */ gfc_init_se (&se, NULL); gfc_trans_array_ctor_element (&body, desc, *poffset, &se, c->expr); *poffset = fold_build2_loc (input_location, PLUS_EXPR, gfc_array_index_type, *poffset, gfc_index_one_node); } else { /* Collect multiple scalar constants into a constructor. */ vec<constructor_elt, va_gc> *v = NULL; tree init; tree bound; tree tmptype; HOST_WIDE_INT idx = 0; p = c; /* Count the number of consecutive scalar constants. */ while (p && !(p->iterator || p->expr->expr_type != EXPR_CONSTANT)) { gfc_init_se (&se, NULL); gfc_conv_constant (&se, p->expr); if (c->expr->ts.type != BT_CHARACTER) se.expr = fold_convert (type, se.expr); /* For constant character array constructors we build an array of pointers. */ else if (POINTER_TYPE_P (type)) se.expr = gfc_build_addr_expr (gfc_get_pchar_type (p->expr->ts.kind), se.expr); CONSTRUCTOR_APPEND_ELT (v, build_int_cst (gfc_array_index_type, idx++), se.expr); c = p; p = gfc_constructor_next (p); } bound = size_int (n - 1); /* Create an array type to hold them. */ tmptype = build_range_type (gfc_array_index_type, gfc_index_zero_node, bound); tmptype = build_array_type (type, tmptype); init = build_constructor (tmptype, v); TREE_CONSTANT (init) = 1; TREE_STATIC (init) = 1; /* Create a static variable to hold the data. */ tmp = gfc_create_var (tmptype, "data"); TREE_STATIC (tmp) = 1; TREE_CONSTANT (tmp) = 1; TREE_READONLY (tmp) = 1; DECL_INITIAL (tmp) = init; init = tmp; /* Use BUILTIN_MEMCPY to assign the values. */ tmp = gfc_conv_descriptor_data_get (desc); tmp = build_fold_indirect_ref_loc (input_location, tmp); tmp = gfc_build_array_ref (tmp, *poffset, NULL); tmp = gfc_build_addr_expr (NULL_TREE, tmp); init = gfc_build_addr_expr (NULL_TREE, init); size = TREE_INT_CST_LOW (TYPE_SIZE_UNIT (type)); bound = build_int_cst (size_type_node, n * size); tmp = build_call_expr_loc (input_location, builtin_decl_explicit (BUILT_IN_MEMCPY), 3, tmp, init, bound); gfc_add_expr_to_block (&body, tmp); *poffset = fold_build2_loc (input_location, PLUS_EXPR, gfc_array_index_type, *poffset, build_int_cst (gfc_array_index_type, n)); } if (!INTEGER_CST_P (*poffset)) { gfc_add_modify (&body, *offsetvar, *poffset); *poffset = *offsetvar; } } /* The frontend should already have done any expansions at compile-time. */ if (!c->iterator) { /* Pass the code as is. */ tmp = gfc_finish_block (&body); gfc_add_expr_to_block (pblock, tmp); } else { /* Build the implied do-loop. */ stmtblock_t implied_do_block; tree cond; tree exit_label; tree loopbody; tree tmp2; loopbody = gfc_finish_block (&body); /* Create a new block that holds the implied-do loop. A temporary loop-variable is used. */ gfc_start_block(&implied_do_block); /* Initialize the loop. */ gfc_add_modify (&implied_do_block, shadow_loopvar, start); /* If this array expands dynamically, and the number of iterations is not constant, we won't have allocated space for the static part of C->EXPR's size. Do that now. */ if (dynamic && gfc_iterator_has_dynamic_bounds (c->iterator)) { /* Get the number of iterations. */ tmp = gfc_get_iteration_count (shadow_loopvar, end, step); /* Get the static part of C->EXPR's size. */ gfc_get_array_constructor_element_size (&size, c->expr); tmp2 = gfc_conv_mpz_to_tree (size, gfc_index_integer_kind); /* Grow the array by TMP * TMP2 elements. */ tmp = fold_build2_loc (input_location, MULT_EXPR, gfc_array_index_type, tmp, tmp2); gfc_grow_array (&implied_do_block, desc, tmp); } /* Generate the loop body. */ exit_label = gfc_build_label_decl (NULL_TREE); gfc_start_block (&body); /* Generate the exit condition. Depending on the sign of the step variable we have to generate the correct comparison. */ tmp = fold_build2_loc (input_location, GT_EXPR, boolean_type_node, step, build_int_cst (TREE_TYPE (step), 0)); cond = fold_build3_loc (input_location, COND_EXPR, boolean_type_node, tmp, fold_build2_loc (input_location, GT_EXPR, boolean_type_node, shadow_loopvar, end), fold_build2_loc (input_location, LT_EXPR, boolean_type_node, shadow_loopvar, end)); tmp = build1_v (GOTO_EXPR, exit_label); TREE_USED (exit_label) = 1; tmp = build3_v (COND_EXPR, cond, tmp, build_empty_stmt (input_location)); gfc_add_expr_to_block (&body, tmp); /* The main loop body. */ gfc_add_expr_to_block (&body, loopbody); /* Increase loop variable by step. */ tmp = fold_build2_loc (input_location, PLUS_EXPR, TREE_TYPE (shadow_loopvar), shadow_loopvar, step); gfc_add_modify (&body, shadow_loopvar, tmp); /* Finish the loop. */ tmp = gfc_finish_block (&body); tmp = build1_v (LOOP_EXPR, tmp); gfc_add_expr_to_block (&implied_do_block, tmp); /* Add the exit label. */ tmp = build1_v (LABEL_EXPR, exit_label); gfc_add_expr_to_block (&implied_do_block, tmp); /* Finish the implied-do loop. */ tmp = gfc_finish_block(&implied_do_block); gfc_add_expr_to_block(pblock, tmp); gfc_restore_sym (c->iterator->var->symtree->n.sym, &saved_loopvar); } } mpz_clear (size); } /* A catch-all to obtain the string length for anything that is not a substring of non-constant length, a constant, array or variable. */ static void get_array_ctor_all_strlen (stmtblock_t *block, gfc_expr *e, tree *len) { gfc_se se; /* Don't bother if we already know the length is a constant. */ if (*len && INTEGER_CST_P (*len)) return; if (!e->ref && e->ts.u.cl && e->ts.u.cl->length && e->ts.u.cl->length->expr_type == EXPR_CONSTANT) { /* This is easy. */ gfc_conv_const_charlen (e->ts.u.cl); *len = e->ts.u.cl->backend_decl; } else { /* Otherwise, be brutal even if inefficient. */ gfc_init_se (&se, NULL); /* No function call, in case of side effects. */ se.no_function_call = 1; if (e->rank == 0) gfc_conv_expr (&se, e); else gfc_conv_expr_descriptor (&se, e); /* Fix the value. */ *len = gfc_evaluate_now (se.string_length, &se.pre); gfc_add_block_to_block (block, &se.pre); gfc_add_block_to_block (block, &se.post); e->ts.u.cl->backend_decl = *len; } } /* Figure out the string length of a variable reference expression. Used by get_array_ctor_strlen. */ static void get_array_ctor_var_strlen (stmtblock_t *block, gfc_expr * expr, tree * len) { gfc_ref *ref; gfc_typespec *ts; mpz_t char_len; /* Don't bother if we already know the length is a constant. */ if (*len && INTEGER_CST_P (*len)) return; ts = &expr->symtree->n.sym->ts; for (ref = expr->ref; ref; ref = ref->next) { switch (ref->type) { case REF_ARRAY: /* Array references don't change the string length. */ break; case REF_COMPONENT: /* Use the length of the component. */ ts = &ref->u.c.component->ts; break; case REF_SUBSTRING: if (ref->u.ss.start->expr_type != EXPR_CONSTANT || ref->u.ss.end->expr_type != EXPR_CONSTANT) { /* Note that this might evaluate expr. */ get_array_ctor_all_strlen (block, expr, len); return; } mpz_init_set_ui (char_len, 1); mpz_add (char_len, char_len, ref->u.ss.end->value.integer); mpz_sub (char_len, char_len, ref->u.ss.start->value.integer); *len = gfc_conv_mpz_to_tree (char_len, gfc_default_integer_kind); *len = convert (gfc_charlen_type_node, *len); mpz_clear (char_len); return; default: gcc_unreachable (); } } *len = ts->u.cl->backend_decl; } /* Figure out the string length of a character array constructor. If len is NULL, don't calculate the length; this happens for recursive calls when a sub-array-constructor is an element but not at the first position, so when we're not interested in the length. Returns TRUE if all elements are character constants. */ bool get_array_ctor_strlen (stmtblock_t *block, gfc_constructor_base base, tree * len) { gfc_constructor *c; bool is_const; is_const = TRUE; if (gfc_constructor_first (base) == NULL) { if (len) *len = build_int_cstu (gfc_charlen_type_node, 0); return is_const; } /* Loop over all constructor elements to find out is_const, but in len we want to store the length of the first, not the last, element. We can of course exit the loop as soon as is_const is found to be false. */ for (c = gfc_constructor_first (base); c && is_const; c = gfc_constructor_next (c)) { switch (c->expr->expr_type) { case EXPR_CONSTANT: if (len && !(*len && INTEGER_CST_P (*len))) *len = build_int_cstu (gfc_charlen_type_node, c->expr->value.character.length); break; case EXPR_ARRAY: if (!get_array_ctor_strlen (block, c->expr->value.constructor, len)) is_const = false; break; case EXPR_VARIABLE: is_const = false; if (len) get_array_ctor_var_strlen (block, c->expr, len); break; default: is_const = false; if (len) get_array_ctor_all_strlen (block, c->expr, len); break; } /* After the first iteration, we don't want the length modified. */ len = NULL; } return is_const; } /* Check whether the array constructor C consists entirely of constant elements, and if so returns the number of those elements, otherwise return zero. Note, an empty or NULL array constructor returns zero. */ unsigned HOST_WIDE_INT gfc_constant_array_constructor_p (gfc_constructor_base base) { unsigned HOST_WIDE_INT nelem = 0; gfc_constructor *c = gfc_constructor_first (base); while (c) { if (c->iterator || c->expr->rank > 0 || c->expr->expr_type != EXPR_CONSTANT) return 0; c = gfc_constructor_next (c); nelem++; } return nelem; } /* Given EXPR, the constant array constructor specified by an EXPR_ARRAY, and the tree type of it's elements, TYPE, return a static constant variable that is compile-time initialized. */ tree gfc_build_constant_array_constructor (gfc_expr * expr, tree type) { tree tmptype, init, tmp; HOST_WIDE_INT nelem; gfc_constructor *c; gfc_array_spec as; gfc_se se; int i; vec<constructor_elt, va_gc> *v = NULL; /* First traverse the constructor list, converting the constants to tree to build an initializer. */ nelem = 0; c = gfc_constructor_first (expr->value.constructor); while (c) { gfc_init_se (&se, NULL); gfc_conv_constant (&se, c->expr); if (c->expr->ts.type != BT_CHARACTER) se.expr = fold_convert (type, se.expr); else if (POINTER_TYPE_P (type)) se.expr = gfc_build_addr_expr (gfc_get_pchar_type (c->expr->ts.kind), se.expr); CONSTRUCTOR_APPEND_ELT (v, build_int_cst (gfc_array_index_type, nelem), se.expr); c = gfc_constructor_next (c); nelem++; } /* Next determine the tree type for the array. We use the gfortran front-end's gfc_get_nodesc_array_type in order to create a suitable GFC_ARRAY_TYPE_P that may be used by the scalarizer. */ memset (&as, 0, sizeof (gfc_array_spec)); as.rank = expr->rank; as.type = AS_EXPLICIT; if (!expr->shape) { as.lower[0] = gfc_get_int_expr (gfc_default_integer_kind, NULL, 0); as.upper[0] = gfc_get_int_expr (gfc_default_integer_kind, NULL, nelem - 1); } else for (i = 0; i < expr->rank; i++) { int tmp = (int) mpz_get_si (expr->shape[i]); as.lower[i] = gfc_get_int_expr (gfc_default_integer_kind, NULL, 0); as.upper[i] = gfc_get_int_expr (gfc_default_integer_kind, NULL, tmp - 1); } tmptype = gfc_get_nodesc_array_type (type, &as, PACKED_STATIC, true); /* as is not needed anymore. */ for (i = 0; i < as.rank + as.corank; i++) { gfc_free_expr (as.lower[i]); gfc_free_expr (as.upper[i]); } init = build_constructor (tmptype, v); TREE_CONSTANT (init) = 1; TREE_STATIC (init) = 1; tmp = build_decl (input_location, VAR_DECL, create_tmp_var_name ("A"), tmptype); DECL_ARTIFICIAL (tmp) = 1; DECL_IGNORED_P (tmp) = 1; TREE_STATIC (tmp) = 1; TREE_CONSTANT (tmp) = 1; TREE_READONLY (tmp) = 1; DECL_INITIAL (tmp) = init; pushdecl (tmp); return tmp; } /* Translate a constant EXPR_ARRAY array constructor for the scalarizer. This mostly initializes the scalarizer state info structure with the appropriate values to directly use the array created by the function gfc_build_constant_array_constructor. */ static void trans_constant_array_constructor (gfc_ss * ss, tree type) { gfc_array_info *info; tree tmp; int i; tmp = gfc_build_constant_array_constructor (ss->info->expr, type); info = &ss->info->data.array; info->descriptor = tmp; info->data = gfc_build_addr_expr (NULL_TREE, tmp); info->offset = gfc_index_zero_node; for (i = 0; i < ss->dimen; i++) { info->delta[i] = gfc_index_zero_node; info->start[i] = gfc_index_zero_node; info->end[i] = gfc_index_zero_node; info->stride[i] = gfc_index_one_node; } } static int get_rank (gfc_loopinfo *loop) { int rank; rank = 0; for (; loop; loop = loop->parent) rank += loop->dimen; return rank; } /* Helper routine of gfc_trans_array_constructor to determine if the bounds of the loop specified by LOOP are constant and simple enough to use with trans_constant_array_constructor. Returns the iteration count of the loop if suitable, and NULL_TREE otherwise. */ static tree constant_array_constructor_loop_size (gfc_loopinfo * l) { gfc_loopinfo *loop; tree size = gfc_index_one_node; tree tmp; int i, total_dim; total_dim = get_rank (l); for (loop = l; loop; loop = loop->parent) { for (i = 0; i < loop->dimen; i++) { /* If the bounds aren't constant, return NULL_TREE. */ if (!INTEGER_CST_P (loop->from[i]) || !INTEGER_CST_P (loop->to[i])) return NULL_TREE; if (!integer_zerop (loop->from[i])) { /* Only allow nonzero "from" in one-dimensional arrays. */ if (total_dim != 1) return NULL_TREE; tmp = fold_build2_loc (input_location, MINUS_EXPR, gfc_array_index_type, loop->to[i], loop->from[i]); } else tmp = loop->to[i]; tmp = fold_build2_loc (input_location, PLUS_EXPR, gfc_array_index_type, tmp, gfc_index_one_node); size = fold_build2_loc (input_location, MULT_EXPR, gfc_array_index_type, size, tmp); } } return size; } static tree * get_loop_upper_bound_for_array (gfc_ss *array, int array_dim) { gfc_ss *ss; int n; gcc_assert (array->nested_ss == NULL); for (ss = array; ss; ss = ss->parent) for (n = 0; n < ss->loop->dimen; n++) if (array_dim == get_array_ref_dim_for_loop_dim (ss, n)) return &(ss->loop->to[n]); gcc_unreachable (); } static gfc_loopinfo * outermost_loop (gfc_loopinfo * loop) { while (loop->parent != NULL) loop = loop->parent; return loop; } /* Array constructors are handled by constructing a temporary, then using that within the scalarization loop. This is not optimal, but seems by far the simplest method. */ static void trans_array_constructor (gfc_ss * ss, locus * where) { gfc_constructor_base c; tree offset; tree offsetvar; tree desc; tree type; tree tmp; tree *loop_ubound0; bool dynamic; bool old_first_len, old_typespec_chararray_ctor; tree old_first_len_val; gfc_loopinfo *loop, *outer_loop; gfc_ss_info *ss_info; gfc_expr *expr; gfc_ss *s; /* Save the old values for nested checking. */ old_first_len = first_len; old_first_len_val = first_len_val; old_typespec_chararray_ctor = typespec_chararray_ctor; loop = ss->loop; outer_loop = outermost_loop (loop); ss_info = ss->info; expr = ss_info->expr; /* Do bounds-checking here and in gfc_trans_array_ctor_element only if no typespec was given for the array constructor. */ typespec_chararray_ctor = (expr->ts.u.cl && expr->ts.u.cl->length_from_typespec); if ((gfc_option.rtcheck & GFC_RTCHECK_BOUNDS) && expr->ts.type == BT_CHARACTER && !typespec_chararray_ctor) { first_len_val = gfc_create_var (gfc_charlen_type_node, "len"); first_len = true; } gcc_assert (ss->dimen == ss->loop->dimen); c = expr->value.constructor; if (expr->ts.type == BT_CHARACTER) { bool const_string; /* get_array_ctor_strlen walks the elements of the constructor, if a typespec was given, we already know the string length and want the one specified there. */ if (typespec_chararray_ctor && expr->ts.u.cl->length && expr->ts.u.cl->length->expr_type != EXPR_CONSTANT) { gfc_se length_se; const_string = false; gfc_init_se (&length_se, NULL); gfc_conv_expr_type (&length_se, expr->ts.u.cl->length, gfc_charlen_type_node); ss_info->string_length = length_se.expr; gfc_add_block_to_block (&outer_loop->pre, &length_se.pre); gfc_add_block_to_block (&outer_loop->post, &length_se.post); } else const_string = get_array_ctor_strlen (&outer_loop->pre, c, &ss_info->string_length); /* Complex character array constructors should have been taken care of and not end up here. */ gcc_assert (ss_info->string_length); expr->ts.u.cl->backend_decl = ss_info->string_length; type = gfc_get_character_type_len (expr->ts.kind, ss_info->string_length); if (const_string) type = build_pointer_type (type); } else type = gfc_typenode_for_spec (&expr->ts); /* See if the constructor determines the loop bounds. */ dynamic = false; loop_ubound0 = get_loop_upper_bound_for_array (ss, 0); if (expr->shape && get_rank (loop) > 1 && *loop_ubound0 == NULL_TREE) { /* We have a multidimensional parameter. */ for (s = ss; s; s = s->parent) { int n; for (n = 0; n < s->loop->dimen; n++) { s->loop->from[n] = gfc_index_zero_node; s->loop->to[n] = gfc_conv_mpz_to_tree (expr->shape[s->dim[n]], gfc_index_integer_kind); s->loop->to[n] = fold_build2_loc (input_location, MINUS_EXPR, gfc_array_index_type, s->loop->to[n], gfc_index_one_node); } } } if (*loop_ubound0 == NULL_TREE) { mpz_t size; /* We should have a 1-dimensional, zero-based loop. */ gcc_assert (loop->parent == NULL && loop->nested == NULL); gcc_assert (loop->dimen == 1); gcc_assert (integer_zerop (loop->from[0])); /* Split the constructor size into a static part and a dynamic part. Allocate the static size up-front and record whether the dynamic size might be nonzero. */ mpz_init (size); dynamic = gfc_get_array_constructor_size (&size, c); mpz_sub_ui (size, size, 1); loop->to[0] = gfc_conv_mpz_to_tree (size, gfc_index_integer_kind); mpz_clear (size); } /* Special case constant array constructors. */ if (!dynamic) { unsigned HOST_WIDE_INT nelem = gfc_constant_array_constructor_p (c); if (nelem > 0) { tree size = constant_array_constructor_loop_size (loop); if (size && compare_tree_int (size, nelem) == 0) { trans_constant_array_constructor (ss, type); goto finish; } } } gfc_trans_create_temp_array (&outer_loop->pre, &outer_loop->post, ss, type, NULL_TREE, dynamic, true, false, where); desc = ss_info->data.array.descriptor; offset = gfc_index_zero_node; offsetvar = gfc_create_var_np (gfc_array_index_type, "offset"); TREE_NO_WARNING (offsetvar) = 1; TREE_USED (offsetvar) = 0; gfc_trans_array_constructor_value (&outer_loop->pre, type, desc, c, &offset, &offsetvar, dynamic); /* If the array grows dynamically, the upper bound of the loop variable is determined by the array's final upper bound. */ if (dynamic) { tmp = fold_build2_loc (input_location, MINUS_EXPR, gfc_array_index_type, offsetvar, gfc_index_one_node); tmp = gfc_evaluate_now (tmp, &outer_loop->pre); gfc_conv_descriptor_ubound_set (&loop->pre, desc, gfc_rank_cst[0], tmp); if (*loop_ubound0 && TREE_CODE (*loop_ubound0) == VAR_DECL) gfc_add_modify (&outer_loop->pre, *loop_ubound0, tmp); else *loop_ubound0 = tmp; } if (TREE_USED (offsetvar)) pushdecl (offsetvar); else gcc_assert (INTEGER_CST_P (offset)); #if 0 /* Disable bound checking for now because it's probably broken. */ if (gfc_option.rtcheck & GFC_RTCHECK_BOUNDS) { gcc_unreachable (); } #endif finish: /* Restore old values of globals. */ first_len = old_first_len; first_len_val = old_first_len_val; typespec_chararray_ctor = old_typespec_chararray_ctor; } /* INFO describes a GFC_SS_SECTION in loop LOOP, and this function is called after evaluating all of INFO's vector dimensions. Go through each such vector dimension and see if we can now fill in any missing loop bounds. */ static void set_vector_loop_bounds (gfc_ss * ss) { gfc_loopinfo *loop, *outer_loop; gfc_array_info *info; gfc_se se; tree tmp; tree desc; tree zero; int n; int dim; outer_loop = outermost_loop (ss->loop); info = &ss->info->data.array; for (; ss; ss = ss->parent) { loop = ss->loop; for (n = 0; n < loop->dimen; n++) { dim = ss->dim[n]; if (info->ref->u.ar.dimen_type[dim] != DIMEN_VECTOR || loop->to[n] != NULL) continue; /* Loop variable N indexes vector dimension DIM, and we don't yet know the upper bound of loop variable N. Set it to the difference between the vector's upper and lower bounds. */ gcc_assert (loop->from[n] == gfc_index_zero_node); gcc_assert (info->subscript[dim] && info->subscript[dim]->info->type == GFC_SS_VECTOR); gfc_init_se (&se, NULL); desc = info->subscript[dim]->info->data.array.descriptor; zero = gfc_rank_cst[0]; tmp = fold_build2_loc (input_location, MINUS_EXPR, gfc_array_index_type, gfc_conv_descriptor_ubound_get (desc, zero), gfc_conv_descriptor_lbound_get (desc, zero)); tmp = gfc_evaluate_now (tmp, &outer_loop->pre); loop->to[n] = tmp; } } } /* Add the pre and post chains for all the scalar expressions in a SS chain to loop. This is called after the loop parameters have been calculated, but before the actual scalarizing loops. */ static void gfc_add_loop_ss_code (gfc_loopinfo * loop, gfc_ss * ss, bool subscript, locus * where) { gfc_loopinfo *nested_loop, *outer_loop; gfc_se se; gfc_ss_info *ss_info; gfc_array_info *info; gfc_expr *expr; int n; /* Don't evaluate the arguments for realloc_lhs_loop_for_fcn_call; otherwise, arguments could get evaluated multiple times. */ if (ss->is_alloc_lhs) return; outer_loop = outermost_loop (loop); /* TODO: This can generate bad code if there are ordering dependencies, e.g., a callee allocated function and an unknown size constructor. */ gcc_assert (ss != NULL); for (; ss != gfc_ss_terminator; ss = ss->loop_chain) { gcc_assert (ss); /* Cross loop arrays are handled from within the most nested loop. */ if (ss->nested_ss != NULL) continue; ss_info = ss->info; expr = ss_info->expr; info = &ss_info->data.array; switch (ss_info->type) { case GFC_SS_SCALAR: /* Scalar expression. Evaluate this now. This includes elemental dimension indices, but not array section bounds. */ gfc_init_se (&se, NULL); gfc_conv_expr (&se, expr); gfc_add_block_to_block (&outer_loop->pre, &se.pre); if (expr->ts.type != BT_CHARACTER && !gfc_is_alloc_class_scalar_function (expr)) { /* Move the evaluation of scalar expressions outside the scalarization loop, except for WHERE assignments. */ if (subscript) se.expr = convert(gfc_array_index_type, se.expr); if (!ss_info->where) se.expr = gfc_evaluate_now (se.expr, &outer_loop->pre); gfc_add_block_to_block (&outer_loop->pre, &se.post); } else gfc_add_block_to_block (&outer_loop->post, &se.post); ss_info->data.scalar.value = se.expr; ss_info->string_length = se.string_length; break; case GFC_SS_REFERENCE: /* Scalar argument to elemental procedure. */ gfc_init_se (&se, NULL); if (ss_info->can_be_null_ref) { /* If the actual argument can be absent (in other words, it can be a NULL reference), don't try to evaluate it; pass instead the reference directly. */ gfc_conv_expr_reference (&se, expr); } else { /* Otherwise, evaluate the argument outside the loop and pass a reference to the value. */ gfc_conv_expr (&se, expr); } /* Ensure that a pointer to the string is stored. */ if (expr->ts.type == BT_CHARACTER) gfc_conv_string_parameter (&se); gfc_add_block_to_block (&outer_loop->pre, &se.pre); gfc_add_block_to_block (&outer_loop->post, &se.post); if (gfc_is_class_scalar_expr (expr)) /* This is necessary because the dynamic type will always be large than the declared type. In consequence, assigning the value to a temporary could segfault. OOP-TODO: see if this is generally correct or is the value has to be written to an allocated temporary, whose address is passed via ss_info. */ ss_info->data.scalar.value = se.expr; else ss_info->data.scalar.value = gfc_evaluate_now (se.expr, &outer_loop->pre); ss_info->string_length = se.string_length; break; case GFC_SS_SECTION: /* Add the expressions for scalar and vector subscripts. */ for (n = 0; n < GFC_MAX_DIMENSIONS; n++) if (info->subscript[n]) gfc_add_loop_ss_code (loop, info->subscript[n], true, where); set_vector_loop_bounds (ss); break; case GFC_SS_VECTOR: /* Get the vector's descriptor and store it in SS. */ gfc_init_se (&se, NULL); gfc_conv_expr_descriptor (&se, expr); gfc_add_block_to_block (&outer_loop->pre, &se.pre); gfc_add_block_to_block (&outer_loop->post, &se.post); info->descriptor = se.expr; break; case GFC_SS_INTRINSIC: gfc_add_intrinsic_ss_code (loop, ss); break; case GFC_SS_FUNCTION: /* Array function return value. We call the function and save its result in a temporary for use inside the loop. */ gfc_init_se (&se, NULL); se.loop = loop; se.ss = ss; gfc_conv_expr (&se, expr); gfc_add_block_to_block (&outer_loop->pre, &se.pre); gfc_add_block_to_block (&outer_loop->post, &se.post); ss_info->string_length = se.string_length; break; case GFC_SS_CONSTRUCTOR: if (expr->ts.type == BT_CHARACTER && ss_info->string_length == NULL && expr->ts.u.cl && expr->ts.u.cl->length) { gfc_init_se (&se, NULL); gfc_conv_expr_type (&se, expr->ts.u.cl->length, gfc_charlen_type_node); ss_info->string_length = se.expr; gfc_add_block_to_block (&outer_loop->pre, &se.pre); gfc_add_block_to_block (&outer_loop->post, &se.post); } trans_array_constructor (ss, where); break; case GFC_SS_TEMP: case GFC_SS_COMPONENT: /* Do nothing. These are handled elsewhere. */ break; default: gcc_unreachable (); } } if (!subscript) for (nested_loop = loop->nested; nested_loop; nested_loop = nested_loop->next) gfc_add_loop_ss_code (nested_loop, nested_loop->ss, subscript, where); } /* Translate expressions for the descriptor and data pointer of a SS. */ /*GCC ARRAYS*/ static void gfc_conv_ss_descriptor (stmtblock_t * block, gfc_ss * ss, int base) { gfc_se se; gfc_ss_info *ss_info; gfc_array_info *info; tree tmp; ss_info = ss->info; info = &ss_info->data.array; /* Get the descriptor for the array to be scalarized. */ gcc_assert (ss_info->expr->expr_type == EXPR_VARIABLE); gfc_init_se (&se, NULL); se.descriptor_only = 1; gfc_conv_expr_lhs (&se, ss_info->expr); gfc_add_block_to_block (block, &se.pre); info->descriptor = se.expr; ss_info->string_length = se.string_length; if (base) { /* Also the data pointer. */ tmp = gfc_conv_array_data (se.expr); /* If this is a variable or address of a variable we use it directly. Otherwise we must evaluate it now to avoid breaking dependency analysis by pulling the expressions for elemental array indices inside the loop. */ if (!(DECL_P (tmp) || (TREE_CODE (tmp) == ADDR_EXPR && DECL_P (TREE_OPERAND (tmp, 0))))) tmp = gfc_evaluate_now (tmp, block); info->data = tmp; tmp = gfc_conv_array_offset (se.expr); info->offset = gfc_evaluate_now (tmp, block); /* Make absolutely sure that the saved_offset is indeed saved so that the variable is still accessible after the loops are translated. */ info->saved_offset = info->offset; } } /* Initialize a gfc_loopinfo structure. */ void gfc_init_loopinfo (gfc_loopinfo * loop) { int n; memset (loop, 0, sizeof (gfc_loopinfo)); gfc_init_block (&loop->pre); gfc_init_block (&loop->post); /* Initially scalarize in order and default to no loop reversal. */ for (n = 0; n < GFC_MAX_DIMENSIONS; n++) { loop->order[n] = n; loop->reverse[n] = GFC_INHIBIT_REVERSE; } loop->ss = gfc_ss_terminator; } /* Copies the loop variable info to a gfc_se structure. Does not copy the SS chain. */ void gfc_copy_loopinfo_to_se (gfc_se * se, gfc_loopinfo * loop) { se->loop = loop; } /* Return an expression for the data pointer of an array. */ tree gfc_conv_array_data (tree descriptor) { tree type; type = TREE_TYPE (descriptor); if (GFC_ARRAY_TYPE_P (type)) { if (TREE_CODE (type) == POINTER_TYPE) return descriptor; else { /* Descriptorless arrays. */ return gfc_build_addr_expr (NULL_TREE, descriptor); } } else return gfc_conv_descriptor_data_get (descriptor); } /* Return an expression for the base offset of an array. */ tree gfc_conv_array_offset (tree descriptor) { tree type; type = TREE_TYPE (descriptor); if (GFC_ARRAY_TYPE_P (type)) return GFC_TYPE_ARRAY_OFFSET (type); else return gfc_conv_descriptor_offset_get (descriptor); } /* Get an expression for the array stride. */ tree gfc_conv_array_stride (tree descriptor, int dim) { tree tmp; tree type; type = TREE_TYPE (descriptor); /* For descriptorless arrays use the array size. */ tmp = GFC_TYPE_ARRAY_STRIDE (type, dim); if (tmp != NULL_TREE) return tmp; tmp = gfc_conv_descriptor_stride_get (descriptor, gfc_rank_cst[dim]); return tmp; } /* Like gfc_conv_array_stride, but for the lower bound. */ tree gfc_conv_array_lbound (tree descriptor, int dim) { tree tmp; tree type; type = TREE_TYPE (descriptor); tmp = GFC_TYPE_ARRAY_LBOUND (type, dim); if (tmp != NULL_TREE) return tmp; tmp = gfc_conv_descriptor_lbound_get (descriptor, gfc_rank_cst[dim]); return tmp; } /* Like gfc_conv_array_stride, but for the upper bound. */ tree gfc_conv_array_ubound (tree descriptor, int dim) { tree tmp; tree type; type = TREE_TYPE (descriptor); tmp = GFC_TYPE_ARRAY_UBOUND (type, dim); if (tmp != NULL_TREE) return tmp; /* This should only ever happen when passing an assumed shape array as an actual parameter. The value will never be used. */ if (GFC_ARRAY_TYPE_P (TREE_TYPE (descriptor))) return gfc_index_zero_node; tmp = gfc_conv_descriptor_ubound_get (descriptor, gfc_rank_cst[dim]); return tmp; } /* Generate code to perform an array index bound check. */ static tree trans_array_bound_check (gfc_se * se, gfc_ss *ss, tree index, int n, locus * where, bool check_upper) { tree fault; tree tmp_lo, tmp_up; tree descriptor; char *msg; const char * name = NULL; if (!(gfc_option.rtcheck & GFC_RTCHECK_BOUNDS)) return index; descriptor = ss->info->data.array.descriptor; index = gfc_evaluate_now (index, &se->pre); /* We find a name for the error message. */ name = ss->info->expr->symtree->n.sym->name; gcc_assert (name != NULL); if (TREE_CODE (descriptor) == VAR_DECL) name = IDENTIFIER_POINTER (DECL_NAME (descriptor)); /* If upper bound is present, include both bounds in the error message. */ if (check_upper) { tmp_lo = gfc_conv_array_lbound (descriptor, n); tmp_up = gfc_conv_array_ubound (descriptor, n); if (name) msg = xasprintf ("Index '%%ld' of dimension %d of array '%s' " "outside of expected range (%%ld:%%ld)", n+1, name); else msg = xasprintf ("Index '%%ld' of dimension %d " "outside of expected range (%%ld:%%ld)", n+1); fault = fold_build2_loc (input_location, LT_EXPR, boolean_type_node, index, tmp_lo); gfc_trans_runtime_check (true, false, fault, &se->pre, where, msg, fold_convert (long_integer_type_node, index), fold_convert (long_integer_type_node, tmp_lo), fold_convert (long_integer_type_node, tmp_up)); fault = fold_build2_loc (input_location, GT_EXPR, boolean_type_node, index, tmp_up); gfc_trans_runtime_check (true, false, fault, &se->pre, where, msg, fold_convert (long_integer_type_node, index), fold_convert (long_integer_type_node, tmp_lo), fold_convert (long_integer_type_node, tmp_up)); free (msg); } else { tmp_lo = gfc_conv_array_lbound (descriptor, n); if (name) msg = xasprintf ("Index '%%ld' of dimension %d of array '%s' " "below lower bound of %%ld", n+1, name); else msg = xasprintf ("Index '%%ld' of dimension %d " "below lower bound of %%ld", n+1); fault = fold_build2_loc (input_location, LT_EXPR, boolean_type_node, index, tmp_lo); gfc_trans_runtime_check (true, false, fault, &se->pre, where, msg, fold_convert (long_integer_type_node, index), fold_convert (long_integer_type_node, tmp_lo)); free (msg); } return index; } /* Return the offset for an index. Performs bound checking for elemental dimensions. Single element references are processed separately. DIM is the array dimension, I is the loop dimension. */ static tree conv_array_index_offset (gfc_se * se, gfc_ss * ss, int dim, int i, gfc_array_ref * ar, tree stride) { gfc_array_info *info; tree index; tree desc; tree data; info = &ss->info->data.array; /* Get the index into the array for this dimension. */ if (ar) { gcc_assert (ar->type != AR_ELEMENT); switch (ar->dimen_type[dim]) { case DIMEN_THIS_IMAGE: gcc_unreachable (); break; case DIMEN_ELEMENT: /* Elemental dimension. */ gcc_assert (info->subscript[dim] && info->subscript[dim]->info->type == GFC_SS_SCALAR); /* We've already translated this value outside the loop. */ index = info->subscript[dim]->info->data.scalar.value; index = trans_array_bound_check (se, ss, index, dim, &ar->where, ar->as->type != AS_ASSUMED_SIZE || dim < ar->dimen - 1); break; case DIMEN_VECTOR: gcc_assert (info && se->loop); gcc_assert (info->subscript[dim] && info->subscript[dim]->info->type == GFC_SS_VECTOR); desc = info->subscript[dim]->info->data.array.descriptor; /* Get a zero-based index into the vector. */ index = fold_build2_loc (input_location, MINUS_EXPR, gfc_array_index_type, se->loop->loopvar[i], se->loop->from[i]); /* Multiply the index by the stride. */ index = fold_build2_loc (input_location, MULT_EXPR, gfc_array_index_type, index, gfc_conv_array_stride (desc, 0)); /* Read the vector to get an index into info->descriptor. */ data = build_fold_indirect_ref_loc (input_location, gfc_conv_array_data (desc)); index = gfc_build_array_ref (data, index, NULL); index = gfc_evaluate_now (index, &se->pre); index = fold_convert (gfc_array_index_type, index); /* Do any bounds checking on the final info->descriptor index. */ index = trans_array_bound_check (se, ss, index, dim, &ar->where, ar->as->type != AS_ASSUMED_SIZE || dim < ar->dimen - 1); break; case DIMEN_RANGE: /* Scalarized dimension. */ gcc_assert (info && se->loop); /* Multiply the loop variable by the stride and delta. */ index = se->loop->loopvar[i]; if (!integer_onep (info->stride[dim])) index = fold_build2_loc (input_location, MULT_EXPR, gfc_array_index_type, index, info->stride[dim]); if (!integer_zerop (info->delta[dim])) index = fold_build2_loc (input_location, PLUS_EXPR, gfc_array_index_type, index, info->delta[dim]); break; default: gcc_unreachable (); } } else { /* Temporary array or derived type component. */ gcc_assert (se->loop); index = se->loop->loopvar[se->loop->order[i]]; /* Pointer functions can have stride[0] different from unity. Use the stride returned by the function call and stored in the descriptor for the temporary. */ if (se->ss && se->ss->info->type == GFC_SS_FUNCTION && se->ss->info->expr && se->ss->info->expr->symtree && se->ss->info->expr->symtree->n.sym->result && se->ss->info->expr->symtree->n.sym->result->attr.pointer) stride = gfc_conv_descriptor_stride_get (info->descriptor, gfc_rank_cst[dim]); if (info->delta[dim] && !integer_zerop (info->delta[dim])) index = fold_build2_loc (input_location, PLUS_EXPR, gfc_array_index_type, index, info->delta[dim]); } /* Multiply by the stride. */ if (!integer_onep (stride)) index = fold_build2_loc (input_location, MULT_EXPR, gfc_array_index_type, index, stride); return index; } /* Build a scalarized array reference using the vptr 'size'. */ static bool build_class_array_ref (gfc_se *se, tree base, tree index) { tree type; tree size; tree offset; tree decl; tree tmp; gfc_expr *expr = se->ss->info->expr; gfc_ref *ref; gfc_ref *class_ref; gfc_typespec *ts; if (expr == NULL || (expr->ts.type != BT_CLASS && !gfc_is_alloc_class_array_function (expr))) return false; if (expr->symtree && expr->symtree->n.sym->ts.type == BT_CLASS) ts = &expr->symtree->n.sym->ts; else ts = NULL; class_ref = NULL; for (ref = expr->ref; ref; ref = ref->next) { if (ref->type == REF_COMPONENT && ref->u.c.component->ts.type == BT_CLASS && ref->next && ref->next->type == REF_COMPONENT && strcmp (ref->next->u.c.component->name, "_data") == 0 && ref->next->next && ref->next->next->type == REF_ARRAY && ref->next->next->u.ar.type != AR_ELEMENT) { ts = &ref->u.c.component->ts; class_ref = ref; break; } } if (ts == NULL) return false; if (class_ref == NULL && expr->symtree->n.sym->attr.function && expr->symtree->n.sym == expr->symtree->n.sym->result) { gcc_assert (expr->symtree->n.sym->backend_decl == current_function_decl); decl = gfc_get_fake_result_decl (expr->symtree->n.sym, 0); } else if (gfc_is_alloc_class_array_function (expr)) { size = NULL_TREE; decl = NULL_TREE; for (tmp = base; tmp; tmp = TREE_OPERAND (tmp, 0)) { tree type; type = TREE_TYPE (tmp); while (type) { if (GFC_CLASS_TYPE_P (type)) decl = tmp; if (type != TYPE_CANONICAL (type)) type = TYPE_CANONICAL (type); else type = NULL_TREE; } if (TREE_CODE (tmp) == VAR_DECL) break; } if (decl == NULL_TREE) return false; } else if (class_ref == NULL) decl = expr->symtree->n.sym->backend_decl; else { /* Remove everything after the last class reference, convert the expression and then recover its tailend once more. */ gfc_se tmpse; ref = class_ref->next; class_ref->next = NULL; gfc_init_se (&tmpse, NULL); gfc_conv_expr (&tmpse, expr); decl = tmpse.expr; class_ref->next = ref; } if (POINTER_TYPE_P (TREE_TYPE (decl))) decl = build_fold_indirect_ref_loc (input_location, decl); if (!GFC_CLASS_TYPE_P (TREE_TYPE (decl))) return false; size = gfc_vtable_size_get (decl); /* Build the address of the element. */ type = TREE_TYPE (TREE_TYPE (base)); size = fold_convert (TREE_TYPE (index), size); offset = fold_build2_loc (input_location, MULT_EXPR, gfc_array_index_type, index, size); tmp = gfc_build_addr_expr (pvoid_type_node, base); tmp = fold_build_pointer_plus_loc (input_location, tmp, offset); tmp = fold_convert (build_pointer_type (type), tmp); /* Return the element in the se expression. */ se->expr = build_fold_indirect_ref_loc (input_location, tmp); return true; } /* Build a scalarized reference to an array. */ static void gfc_conv_scalarized_array_ref (gfc_se * se, gfc_array_ref * ar) { gfc_array_info *info; tree decl = NULL_TREE; tree index; tree tmp; gfc_ss *ss; gfc_expr *expr; int n; ss = se->ss; expr = ss->info->expr; info = &ss->info->data.array; if (ar) n = se->loop->order[0]; else n = 0; index = conv_array_index_offset (se, ss, ss->dim[n], n, ar, info->stride0); /* Add the offset for this dimension to the stored offset for all other dimensions. */ if (info->offset && !integer_zerop (info->offset)) index = fold_build2_loc (input_location, PLUS_EXPR, gfc_array_index_type, index, info->offset); if (expr && is_subref_array (expr)) decl = expr->symtree->n.sym->backend_decl; tmp = build_fold_indirect_ref_loc (input_location, info->data); /* Use the vptr 'size' field to access a class the element of a class array. */ if (build_class_array_ref (se, tmp, index)) return; se->expr = gfc_build_array_ref (tmp, index, decl); } /* Translate access of temporary array. */ void gfc_conv_tmp_array_ref (gfc_se * se) { se->string_length = se->ss->info->string_length; gfc_conv_scalarized_array_ref (se, NULL); gfc_advance_se_ss_chain (se); } /* Add T to the offset pair *OFFSET, *CST_OFFSET. */ static void add_to_offset (tree *cst_offset, tree *offset, tree t) { if (TREE_CODE (t) == INTEGER_CST) *cst_offset = int_const_binop (PLUS_EXPR, *cst_offset, t); else { if (!integer_zerop (*offset)) *offset = fold_build2_loc (input_location, PLUS_EXPR, gfc_array_index_type, *offset, t); else *offset = t; } } static tree build_array_ref (tree desc, tree offset, tree decl) { tree tmp; tree type; /* Class container types do not always have the GFC_CLASS_TYPE_P but the canonical type does. */ if (GFC_DESCRIPTOR_TYPE_P (TREE_TYPE (desc)) && TREE_CODE (desc) == COMPONENT_REF) { type = TREE_TYPE (TREE_OPERAND (desc, 0)); if (TYPE_CANONICAL (type) && GFC_CLASS_TYPE_P (TYPE_CANONICAL (type))) type = TYPE_CANONICAL (type); } else type = NULL; /* Class array references need special treatment because the assigned type size needs to be used to point to the element. */ if (type && GFC_CLASS_TYPE_P (type)) { type = gfc_get_element_type (TREE_TYPE (desc)); tmp = TREE_OPERAND (desc, 0); tmp = gfc_get_class_array_ref (offset, tmp); tmp = fold_convert (build_pointer_type (type), tmp); tmp = build_fold_indirect_ref_loc (input_location, tmp); return tmp; } tmp = gfc_conv_array_data (desc); tmp = build_fold_indirect_ref_loc (input_location, tmp); tmp = gfc_build_array_ref (tmp, offset, decl); return tmp; } /* Build an array reference. se->expr already holds the array descriptor. This should be either a variable, indirect variable reference or component reference. For arrays which do not have a descriptor, se->expr will be the data pointer. a(i, j, k) = base[offset + i * stride[0] + j * stride[1] + k * stride[2]]*/ void gfc_conv_array_ref (gfc_se * se, gfc_array_ref * ar, gfc_expr *expr, locus * where) { int n; tree offset, cst_offset; tree tmp; tree stride; gfc_se indexse; gfc_se tmpse; gfc_symbol * sym = expr->symtree->n.sym; char *var_name = NULL; if (ar->dimen == 0) { gcc_assert (ar->codimen); if (GFC_DESCRIPTOR_TYPE_P (TREE_TYPE (se->expr))) se->expr = build_fold_indirect_ref (gfc_conv_array_data (se->expr)); else { if (GFC_ARRAY_TYPE_P (TREE_TYPE (se->expr)) && TREE_CODE (TREE_TYPE (se->expr)) == POINTER_TYPE) se->expr = build_fold_indirect_ref_loc (input_location, se->expr); /* Use the actual tree type and not the wrapped coarray. */ if (!se->want_pointer) se->expr = fold_convert (TYPE_MAIN_VARIANT (TREE_TYPE (se->expr)), se->expr); } return; } /* Handle scalarized references separately. */ if (ar->type != AR_ELEMENT) { gfc_conv_scalarized_array_ref (se, ar); gfc_advance_se_ss_chain (se); return; } if (gfc_option.rtcheck & GFC_RTCHECK_BOUNDS) { size_t len; gfc_ref *ref; len = strlen (sym->name) + 1; for (ref = expr->ref; ref; ref = ref->next) { if (ref->type == REF_ARRAY && &ref->u.ar == ar) break; if (ref->type == REF_COMPONENT) len += 1 + strlen (ref->u.c.component->name); } var_name = XALLOCAVEC (char, len); strcpy (var_name, sym->name); for (ref = expr->ref; ref; ref = ref->next) { if (ref->type == REF_ARRAY && &ref->u.ar == ar) break; if (ref->type == REF_COMPONENT) { strcat (var_name, "%%"); strcat (var_name, ref->u.c.component->name); } } } cst_offset = offset = gfc_index_zero_node; add_to_offset (&cst_offset, &offset, gfc_conv_array_offset (se->expr)); /* Calculate the offsets from all the dimensions. Make sure to associate the final offset so that we form a chain of loop invariant summands. */ for (n = ar->dimen - 1; n >= 0; n--) { /* Calculate the index for this dimension. */ gfc_init_se (&indexse, se); gfc_conv_expr_type (&indexse, ar->start[n], gfc_array_index_type); gfc_add_block_to_block (&se->pre, &indexse.pre); if (gfc_option.rtcheck & GFC_RTCHECK_BOUNDS) { /* Check array bounds. */ tree cond; char *msg; /* Evaluate the indexse.expr only once. */ indexse.expr = save_expr (indexse.expr); /* Lower bound. */ tmp = gfc_conv_array_lbound (se->expr, n); if (sym->attr.temporary) { gfc_init_se (&tmpse, se); gfc_conv_expr_type (&tmpse, ar->as->lower[n], gfc_array_index_type); gfc_add_block_to_block (&se->pre, &tmpse.pre); tmp = tmpse.expr; } cond = fold_build2_loc (input_location, LT_EXPR, boolean_type_node, indexse.expr, tmp); msg = xasprintf ("Index '%%ld' of dimension %d of array '%s' " "below lower bound of %%ld", n+1, var_name); gfc_trans_runtime_check (true, false, cond, &se->pre, where, msg, fold_convert (long_integer_type_node, indexse.expr), fold_convert (long_integer_type_node, tmp)); free (msg); /* Upper bound, but not for the last dimension of assumed-size arrays. */ if (n < ar->dimen - 1 || ar->as->type != AS_ASSUMED_SIZE) { tmp = gfc_conv_array_ubound (se->expr, n); if (sym->attr.temporary) { gfc_init_se (&tmpse, se); gfc_conv_expr_type (&tmpse, ar->as->upper[n], gfc_array_index_type); gfc_add_block_to_block (&se->pre, &tmpse.pre); tmp = tmpse.expr; } cond = fold_build2_loc (input_location, GT_EXPR, boolean_type_node, indexse.expr, tmp); msg = xasprintf ("Index '%%ld' of dimension %d of array '%s' " "above upper bound of %%ld", n+1, var_name); gfc_trans_runtime_check (true, false, cond, &se->pre, where, msg, fold_convert (long_integer_type_node, indexse.expr), fold_convert (long_integer_type_node, tmp)); free (msg); } } /* Multiply the index by the stride. */ stride = gfc_conv_array_stride (se->expr, n); tmp = fold_build2_loc (input_location, MULT_EXPR, gfc_array_index_type, indexse.expr, stride); /* And add it to the total. */ add_to_offset (&cst_offset, &offset, tmp); } if (!integer_zerop (cst_offset)) offset = fold_build2_loc (input_location, PLUS_EXPR, gfc_array_index_type, offset, cst_offset); se->expr = build_array_ref (se->expr, offset, sym->backend_decl); } /* Add the offset corresponding to array's ARRAY_DIM dimension and loop's LOOP_DIM dimension (if any) to array's offset. */ static void add_array_offset (stmtblock_t *pblock, gfc_loopinfo *loop, gfc_ss *ss, gfc_array_ref *ar, int array_dim, int loop_dim) { gfc_se se; gfc_array_info *info; tree stride, index; info = &ss->info->data.array; gfc_init_se (&se, NULL); se.loop = loop; se.expr = info->descriptor; stride = gfc_conv_array_stride (info->descriptor, array_dim); index = conv_array_index_offset (&se, ss, array_dim, loop_dim, ar, stride); gfc_add_block_to_block (pblock, &se.pre); info->offset = fold_build2_loc (input_location, PLUS_EXPR, gfc_array_index_type, info->offset, index); info->offset = gfc_evaluate_now (info->offset, pblock); } /* Generate the code to be executed immediately before entering a scalarization loop. */ static void gfc_trans_preloop_setup (gfc_loopinfo * loop, int dim, int flag, stmtblock_t * pblock) { tree stride; gfc_ss_info *ss_info; gfc_array_info *info; gfc_ss_type ss_type; gfc_ss *ss, *pss; gfc_loopinfo *ploop; gfc_array_ref *ar; int i; /* This code will be executed before entering the scalarization loop for this dimension. */ for (ss = loop->ss; ss != gfc_ss_terminator; ss = ss->loop_chain) { ss_info = ss->info; if ((ss_info->useflags & flag) == 0) continue; ss_type = ss_info->type; if (ss_type != GFC_SS_SECTION && ss_type != GFC_SS_FUNCTION && ss_type != GFC_SS_CONSTRUCTOR && ss_type != GFC_SS_COMPONENT) continue; info = &ss_info->data.array; gcc_assert (dim < ss->dimen); gcc_assert (ss->dimen == loop->dimen); if (info->ref) ar = &info->ref->u.ar; else ar = NULL; if (dim == loop->dimen - 1 && loop->parent != NULL) { /* If we are in the outermost dimension of this loop, the previous dimension shall be in the parent loop. */ gcc_assert (ss->parent != NULL); pss = ss->parent; ploop = loop->parent; /* ss and ss->parent are about the same array. */ gcc_assert (ss_info == pss->info); } else { ploop = loop; pss = ss; } if (dim == loop->dimen - 1) i = 0; else i = dim + 1; /* For the time being, there is no loop reordering. */ gcc_assert (i == ploop->order[i]); i = ploop->order[i]; if (dim == loop->dimen - 1 && loop->parent == NULL) { stride = gfc_conv_array_stride (info->descriptor, innermost_ss (ss)->dim[i]); /* Calculate the stride of the innermost loop. Hopefully this will allow the backend optimizers to do their stuff more effectively. */ info->stride0 = gfc_evaluate_now (stride, pblock); /* For the outermost loop calculate the offset due to any elemental dimensions. It will have been initialized with the base offset of the array. */ if (info->ref) { for (i = 0; i < ar->dimen; i++) { if (ar->dimen_type[i] != DIMEN_ELEMENT) continue; add_array_offset (pblock, loop, ss, ar, i, /* unused */ -1); } } } else /* Add the offset for the previous loop dimension. */ add_array_offset (pblock, ploop, ss, ar, pss->dim[i], i); /* Remember this offset for the second loop. */ if (dim == loop->temp_dim - 1 && loop->parent == NULL) info->saved_offset = info->offset; } } /* Start a scalarized expression. Creates a scope and declares loop variables. */ void gfc_start_scalarized_body (gfc_loopinfo * loop, stmtblock_t * pbody) { int dim; int n; int flags; gcc_assert (!loop->array_parameter); for (dim = loop->dimen - 1; dim >= 0; dim--) { n = loop->order[dim]; gfc_start_block (&loop->code[n]); /* Create the loop variable. */ loop->loopvar[n] = gfc_create_var (gfc_array_index_type, "S"); if (dim < loop->temp_dim) flags = 3; else flags = 1; /* Calculate values that will be constant within this loop. */ gfc_trans_preloop_setup (loop, dim, flags, &loop->code[n]); } gfc_start_block (pbody); } /* Generates the actual loop code for a scalarization loop. */ void gfc_trans_scalarized_loop_end (gfc_loopinfo * loop, int n, stmtblock_t * pbody) { stmtblock_t block; tree cond; tree tmp; tree loopbody; tree exit_label; tree stmt; tree init; tree incr; if ((ompws_flags & (OMPWS_WORKSHARE_FLAG | OMPWS_SCALARIZER_WS)) == (OMPWS_WORKSHARE_FLAG | OMPWS_SCALARIZER_WS) && n == loop->dimen - 1) { /* We create an OMP_FOR construct for the outermost scalarized loop. */ init = make_tree_vec (1); cond = make_tree_vec (1); incr = make_tree_vec (1); /* Cycle statement is implemented with a goto. Exit statement must not be present for this loop. */ exit_label = gfc_build_label_decl (NULL_TREE); TREE_USED (exit_label) = 1; /* Label for cycle statements (if needed). */ tmp = build1_v (LABEL_EXPR, exit_label); gfc_add_expr_to_block (pbody, tmp); stmt = make_node (OMP_FOR); TREE_TYPE (stmt) = void_type_node; OMP_FOR_BODY (stmt) = loopbody = gfc_finish_block (pbody); OMP_FOR_CLAUSES (stmt) = build_omp_clause (input_location, OMP_CLAUSE_SCHEDULE); OMP_CLAUSE_SCHEDULE_KIND (OMP_FOR_CLAUSES (stmt)) = OMP_CLAUSE_SCHEDULE_STATIC; if (ompws_flags & OMPWS_NOWAIT) OMP_CLAUSE_CHAIN (OMP_FOR_CLAUSES (stmt)) = build_omp_clause (input_location, OMP_CLAUSE_NOWAIT); /* Initialize the loopvar. */ TREE_VEC_ELT (init, 0) = build2_v (MODIFY_EXPR, loop->loopvar[n], loop->from[n]); OMP_FOR_INIT (stmt) = init; /* The exit condition. */ TREE_VEC_ELT (cond, 0) = build2_loc (input_location, LE_EXPR, boolean_type_node, loop->loopvar[n], loop->to[n]); SET_EXPR_LOCATION (TREE_VEC_ELT (cond, 0), input_location); OMP_FOR_COND (stmt) = cond; /* Increment the loopvar. */ tmp = build2_loc (input_location, PLUS_EXPR, gfc_array_index_type, loop->loopvar[n], gfc_index_one_node); TREE_VEC_ELT (incr, 0) = fold_build2_loc (input_location, MODIFY_EXPR, void_type_node, loop->loopvar[n], tmp); OMP_FOR_INCR (stmt) = incr; ompws_flags &= ~OMPWS_CURR_SINGLEUNIT; gfc_add_expr_to_block (&loop->code[n], stmt); } else { bool reverse_loop = (loop->reverse[n] == GFC_REVERSE_SET) && (loop->temp_ss == NULL); loopbody = gfc_finish_block (pbody); if (reverse_loop) { tmp = loop->from[n]; loop->from[n] = loop->to[n]; loop->to[n] = tmp; } /* Initialize the loopvar. */ if (loop->loopvar[n] != loop->from[n]) gfc_add_modify (&loop->code[n], loop->loopvar[n], loop->from[n]); exit_label = gfc_build_label_decl (NULL_TREE); /* Generate the loop body. */ gfc_init_block (&block); /* The exit condition. */ cond = fold_build2_loc (input_location, reverse_loop ? LT_EXPR : GT_EXPR, boolean_type_node, loop->loopvar[n], loop->to[n]); tmp = build1_v (GOTO_EXPR, exit_label); TREE_USED (exit_label) = 1; tmp = build3_v (COND_EXPR, cond, tmp, build_empty_stmt (input_location)); gfc_add_expr_to_block (&block, tmp); /* The main body. */ gfc_add_expr_to_block (&block, loopbody); /* Increment the loopvar. */ tmp = fold_build2_loc (input_location, reverse_loop ? MINUS_EXPR : PLUS_EXPR, gfc_array_index_type, loop->loopvar[n], gfc_index_one_node); gfc_add_modify (&block, loop->loopvar[n], tmp); /* Build the loop. */ tmp = gfc_finish_block (&block); tmp = build1_v (LOOP_EXPR, tmp); gfc_add_expr_to_block (&loop->code[n], tmp); /* Add the exit label. */ tmp = build1_v (LABEL_EXPR, exit_label); gfc_add_expr_to_block (&loop->code[n], tmp); } } /* Finishes and generates the loops for a scalarized expression. */ void gfc_trans_scalarizing_loops (gfc_loopinfo * loop, stmtblock_t * body) { int dim; int n; gfc_ss *ss; stmtblock_t *pblock; tree tmp; pblock = body; /* Generate the loops. */ for (dim = 0; dim < loop->dimen; dim++) { n = loop->order[dim]; gfc_trans_scalarized_loop_end (loop, n, pblock); loop->loopvar[n] = NULL_TREE; pblock = &loop->code[n]; } tmp = gfc_finish_block (pblock); gfc_add_expr_to_block (&loop->pre, tmp); /* Clear all the used flags. */ for (ss = loop->ss; ss != gfc_ss_terminator; ss = ss->loop_chain) if (ss->parent == NULL) ss->info->useflags = 0; } /* Finish the main body of a scalarized expression, and start the secondary copying body. */ void gfc_trans_scalarized_loop_boundary (gfc_loopinfo * loop, stmtblock_t * body) { int dim; int n; stmtblock_t *pblock; gfc_ss *ss; pblock = body; /* We finish as many loops as are used by the temporary. */ for (dim = 0; dim < loop->temp_dim - 1; dim++) { n = loop->order[dim]; gfc_trans_scalarized_loop_end (loop, n, pblock); loop->loopvar[n] = NULL_TREE; pblock = &loop->code[n]; } /* We don't want to finish the outermost loop entirely. */ n = loop->order[loop->temp_dim - 1]; gfc_trans_scalarized_loop_end (loop, n, pblock); /* Restore the initial offsets. */ for (ss = loop->ss; ss != gfc_ss_terminator; ss = ss->loop_chain) { gfc_ss_type ss_type; gfc_ss_info *ss_info; ss_info = ss->info; if ((ss_info->useflags & 2) == 0) continue; ss_type = ss_info->type; if (ss_type != GFC_SS_SECTION && ss_type != GFC_SS_FUNCTION && ss_type != GFC_SS_CONSTRUCTOR && ss_type != GFC_SS_COMPONENT) continue; ss_info->data.array.offset = ss_info->data.array.saved_offset; } /* Restart all the inner loops we just finished. */ for (dim = loop->temp_dim - 2; dim >= 0; dim--) { n = loop->order[dim]; gfc_start_block (&loop->code[n]); loop->loopvar[n] = gfc_create_var (gfc_array_index_type, "Q"); gfc_trans_preloop_setup (loop, dim, 2, &loop->code[n]); } /* Start a block for the secondary copying code. */ gfc_start_block (body); } /* Precalculate (either lower or upper) bound of an array section. BLOCK: Block in which the (pre)calculation code will go. BOUNDS[DIM]: Where the bound value will be stored once evaluated. VALUES[DIM]: Specified bound (NULL <=> unspecified). DESC: Array descriptor from which the bound will be picked if unspecified (either lower or upper bound according to LBOUND). */ static void evaluate_bound (stmtblock_t *block, tree *bounds, gfc_expr ** values, tree desc, int dim, bool lbound) { gfc_se se; gfc_expr * input_val = values[dim]; tree *output = &bounds[dim]; if (input_val) { /* Specified section bound. */ gfc_init_se (&se, NULL); gfc_conv_expr_type (&se, input_val, gfc_array_index_type); gfc_add_block_to_block (block, &se.pre); *output = se.expr; } else { /* No specific bound specified so use the bound of the array. */ *output = lbound ? gfc_conv_array_lbound (desc, dim) : gfc_conv_array_ubound (desc, dim); } *output = gfc_evaluate_now (*output, block); } /* Calculate the lower bound of an array section. */ static void gfc_conv_section_startstride (stmtblock_t * block, gfc_ss * ss, int dim) { gfc_expr *stride = NULL; tree desc; gfc_se se; gfc_array_info *info; gfc_array_ref *ar; gcc_assert (ss->info->type == GFC_SS_SECTION); info = &ss->info->data.array; ar = &info->ref->u.ar; if (ar->dimen_type[dim] == DIMEN_VECTOR) { /* We use a zero-based index to access the vector. */ info->start[dim] = gfc_index_zero_node; info->end[dim] = NULL; info->stride[dim] = gfc_index_one_node; return; } gcc_assert (ar->dimen_type[dim] == DIMEN_RANGE || ar->dimen_type[dim] == DIMEN_THIS_IMAGE); desc = info->descriptor; stride = ar->stride[dim]; /* Calculate the start of the range. For vector subscripts this will be the range of the vector. */ evaluate_bound (block, info->start, ar->start, desc, dim, true); /* Similarly calculate the end. Although this is not used in the scalarizer, it is needed when checking bounds and where the end is an expression with side-effects. */ evaluate_bound (block, info->end, ar->end, desc, dim, false); /* Calculate the stride. */ if (stride == NULL) info->stride[dim] = gfc_index_one_node; else { gfc_init_se (&se, NULL); gfc_conv_expr_type (&se, stride, gfc_array_index_type); gfc_add_block_to_block (block, &se.pre); info->stride[dim] = gfc_evaluate_now (se.expr, block); } } /* Calculates the range start and stride for a SS chain. Also gets the descriptor and data pointer. The range of vector subscripts is the size of the vector. Array bounds are also checked. */ void gfc_conv_ss_startstride (gfc_loopinfo * loop) { int n; tree tmp; gfc_ss *ss; tree desc; gfc_loopinfo * const outer_loop = outermost_loop (loop); loop->dimen = 0; /* Determine the rank of the loop. */ for (ss = loop->ss; ss != gfc_ss_terminator; ss = ss->loop_chain) { switch (ss->info->type) { case GFC_SS_SECTION: case GFC_SS_CONSTRUCTOR: case GFC_SS_FUNCTION: case GFC_SS_COMPONENT: loop->dimen = ss->dimen; goto done; /* As usual, lbound and ubound are exceptions!. */ case GFC_SS_INTRINSIC: switch (ss->info->expr->value.function.isym->id) { case GFC_ISYM_LBOUND: case GFC_ISYM_UBOUND: case GFC_ISYM_LCOBOUND: case GFC_ISYM_UCOBOUND: case GFC_ISYM_THIS_IMAGE: loop->dimen = ss->dimen; goto done; default: break; } default: break; } } /* We should have determined the rank of the expression by now. If not, that's bad news. */ gcc_unreachable (); done: /* Loop over all the SS in the chain. */ for (ss = loop->ss; ss != gfc_ss_terminator; ss = ss->loop_chain) { gfc_ss_info *ss_info; gfc_array_info *info; gfc_expr *expr; ss_info = ss->info; expr = ss_info->expr; info = &ss_info->data.array; if (expr && expr->shape && !info->shape) info->shape = expr->shape; switch (ss_info->type) { case GFC_SS_SECTION: /* Get the descriptor for the array. If it is a cross loops array, we got the descriptor already in the outermost loop. */ if (ss->parent == NULL) gfc_conv_ss_descriptor (&outer_loop->pre, ss, !loop->array_parameter); for (n = 0; n < ss->dimen; n++) gfc_conv_section_startstride (&outer_loop->pre, ss, ss->dim[n]); break; case GFC_SS_INTRINSIC: switch (expr->value.function.isym->id) { /* Fall through to supply start and stride. */ case GFC_ISYM_LBOUND: case GFC_ISYM_UBOUND: { gfc_expr *arg; /* This is the variant without DIM=... */ gcc_assert (expr->value.function.actual->next->expr == NULL); arg = expr->value.function.actual->expr; if (arg->rank == -1) { gfc_se se; tree rank, tmp; /* The rank (hence the return value's shape) is unknown, we have to retrieve it. */ gfc_init_se (&se, NULL); se.descriptor_only = 1; gfc_conv_expr (&se, arg); /* This is a bare variable, so there is no preliminary or cleanup code. */ gcc_assert (se.pre.head == NULL_TREE && se.post.head == NULL_TREE); rank = gfc_conv_descriptor_rank (se.expr); tmp = fold_build2_loc (input_location, MINUS_EXPR, gfc_array_index_type, fold_convert (gfc_array_index_type, rank), gfc_index_one_node); info->end[0] = gfc_evaluate_now (tmp, &outer_loop->pre); info->start[0] = gfc_index_zero_node; info->stride[0] = gfc_index_one_node; continue; } /* Otherwise fall through GFC_SS_FUNCTION. */ } case GFC_ISYM_LCOBOUND: case GFC_ISYM_UCOBOUND: case GFC_ISYM_THIS_IMAGE: break; default: continue; } case GFC_SS_CONSTRUCTOR: case GFC_SS_FUNCTION: for (n = 0; n < ss->dimen; n++) { int dim = ss->dim[n]; info->start[dim] = gfc_index_zero_node; info->end[dim] = gfc_index_zero_node; info->stride[dim] = gfc_index_one_node; } break; default: break; } } /* The rest is just runtime bound checking. */ if (gfc_option.rtcheck & GFC_RTCHECK_BOUNDS) { stmtblock_t block; tree lbound, ubound; tree end; tree size[GFC_MAX_DIMENSIONS]; tree stride_pos, stride_neg, non_zerosized, tmp2, tmp3; gfc_array_info *info; char *msg; int dim; gfc_start_block (&block); for (n = 0; n < loop->dimen; n++) size[n] = NULL_TREE; for (ss = loop->ss; ss != gfc_ss_terminator; ss = ss->loop_chain) { stmtblock_t inner; gfc_ss_info *ss_info; gfc_expr *expr; locus *expr_loc; const char *expr_name; ss_info = ss->info; if (ss_info->type != GFC_SS_SECTION) continue; /* Catch allocatable lhs in f2003. */ if (flag_realloc_lhs && ss->is_alloc_lhs) continue; expr = ss_info->expr; expr_loc = &expr->where; expr_name = expr->symtree->name; gfc_start_block (&inner); /* TODO: range checking for mapped dimensions. */ info = &ss_info->data.array; /* This code only checks ranges. Elemental and vector dimensions are checked later. */ for (n = 0; n < loop->dimen; n++) { bool check_upper; dim = ss->dim[n]; if (info->ref->u.ar.dimen_type[dim] != DIMEN_RANGE) continue; if (dim == info->ref->u.ar.dimen - 1 && info->ref->u.ar.as->type == AS_ASSUMED_SIZE) check_upper = false; else check_upper = true; /* Zero stride is not allowed. */ tmp = fold_build2_loc (input_location, EQ_EXPR, boolean_type_node, info->stride[dim], gfc_index_zero_node); msg = xasprintf ("Zero stride is not allowed, for dimension %d " "of array '%s'", dim + 1, expr_name); gfc_trans_runtime_check (true, false, tmp, &inner, expr_loc, msg); free (msg); desc = info->descriptor; /* This is the run-time equivalent of resolve.c's check_dimension(). The logical is more readable there than it is here, with all the trees. */ lbound = gfc_conv_array_lbound (desc, dim); end = info->end[dim]; if (check_upper) ubound = gfc_conv_array_ubound (desc, dim); else ubound = NULL; /* non_zerosized is true when the selected range is not empty. */ stride_pos = fold_build2_loc (input_location, GT_EXPR, boolean_type_node, info->stride[dim], gfc_index_zero_node); tmp = fold_build2_loc (input_location, LE_EXPR, boolean_type_node, info->start[dim], end); stride_pos = fold_build2_loc (input_location, TRUTH_AND_EXPR, boolean_type_node, stride_pos, tmp); stride_neg = fold_build2_loc (input_location, LT_EXPR, boolean_type_node, info->stride[dim], gfc_index_zero_node); tmp = fold_build2_loc (input_location, GE_EXPR, boolean_type_node, info->start[dim], end); stride_neg = fold_build2_loc (input_location, TRUTH_AND_EXPR, boolean_type_node, stride_neg, tmp); non_zerosized = fold_build2_loc (input_location, TRUTH_OR_EXPR, boolean_type_node, stride_pos, stride_neg); /* Check the start of the range against the lower and upper bounds of the array, if the range is not empty. If upper bound is present, include both bounds in the error message. */ if (check_upper) { tmp = fold_build2_loc (input_location, LT_EXPR, boolean_type_node, info->start[dim], lbound); tmp = fold_build2_loc (input_location, TRUTH_AND_EXPR, boolean_type_node, non_zerosized, tmp); tmp2 = fold_build2_loc (input_location, GT_EXPR, boolean_type_node, info->start[dim], ubound); tmp2 = fold_build2_loc (input_location, TRUTH_AND_EXPR, boolean_type_node, non_zerosized, tmp2); msg = xasprintf ("Index '%%ld' of dimension %d of array '%s' " "outside of expected range (%%ld:%%ld)", dim + 1, expr_name); gfc_trans_runtime_check (true, false, tmp, &inner, expr_loc, msg, fold_convert (long_integer_type_node, info->start[dim]), fold_convert (long_integer_type_node, lbound), fold_convert (long_integer_type_node, ubound)); gfc_trans_runtime_check (true, false, tmp2, &inner, expr_loc, msg, fold_convert (long_integer_type_node, info->start[dim]), fold_convert (long_integer_type_node, lbound), fold_convert (long_integer_type_node, ubound)); free (msg); } else { tmp = fold_build2_loc (input_location, LT_EXPR, boolean_type_node, info->start[dim], lbound); tmp = fold_build2_loc (input_location, TRUTH_AND_EXPR, boolean_type_node, non_zerosized, tmp); msg = xasprintf ("Index '%%ld' of dimension %d of array '%s' " "below lower bound of %%ld", dim + 1, expr_name); gfc_trans_runtime_check (true, false, tmp, &inner, expr_loc, msg, fold_convert (long_integer_type_node, info->start[dim]), fold_convert (long_integer_type_node, lbound)); free (msg); } /* Compute the last element of the range, which is not necessarily "end" (think 0:5:3, which doesn't contain 5) and check it against both lower and upper bounds. */ tmp = fold_build2_loc (input_location, MINUS_EXPR, gfc_array_index_type, end, info->start[dim]); tmp = fold_build2_loc (input_location, TRUNC_MOD_EXPR, gfc_array_index_type, tmp, info->stride[dim]); tmp = fold_build2_loc (input_location, MINUS_EXPR, gfc_array_index_type, end, tmp); tmp2 = fold_build2_loc (input_location, LT_EXPR, boolean_type_node, tmp, lbound); tmp2 = fold_build2_loc (input_location, TRUTH_AND_EXPR, boolean_type_node, non_zerosized, tmp2); if (check_upper) { tmp3 = fold_build2_loc (input_location, GT_EXPR, boolean_type_node, tmp, ubound); tmp3 = fold_build2_loc (input_location, TRUTH_AND_EXPR, boolean_type_node, non_zerosized, tmp3); msg = xasprintf ("Index '%%ld' of dimension %d of array '%s' " "outside of expected range (%%ld:%%ld)", dim + 1, expr_name); gfc_trans_runtime_check (true, false, tmp2, &inner, expr_loc, msg, fold_convert (long_integer_type_node, tmp), fold_convert (long_integer_type_node, ubound), fold_convert (long_integer_type_node, lbound)); gfc_trans_runtime_check (true, false, tmp3, &inner, expr_loc, msg, fold_convert (long_integer_type_node, tmp), fold_convert (long_integer_type_node, ubound), fold_convert (long_integer_type_node, lbound)); free (msg); } else { msg = xasprintf ("Index '%%ld' of dimension %d of array '%s' " "below lower bound of %%ld", dim + 1, expr_name); gfc_trans_runtime_check (true, false, tmp2, &inner, expr_loc, msg, fold_convert (long_integer_type_node, tmp), fold_convert (long_integer_type_node, lbound)); free (msg); } /* Check the section sizes match. */ tmp = fold_build2_loc (input_location, MINUS_EXPR, gfc_array_index_type, end, info->start[dim]); tmp = fold_build2_loc (input_location, FLOOR_DIV_EXPR, gfc_array_index_type, tmp, info->stride[dim]); tmp = fold_build2_loc (input_location, PLUS_EXPR, gfc_array_index_type, gfc_index_one_node, tmp); tmp = fold_build2_loc (input_location, MAX_EXPR, gfc_array_index_type, tmp, build_int_cst (gfc_array_index_type, 0)); /* We remember the size of the first section, and check all the others against this. */ if (size[n]) { tmp3 = fold_build2_loc (input_location, NE_EXPR, boolean_type_node, tmp, size[n]); msg = xasprintf ("Array bound mismatch for dimension %d " "of array '%s' (%%ld/%%ld)", dim + 1, expr_name); gfc_trans_runtime_check (true, false, tmp3, &inner, expr_loc, msg, fold_convert (long_integer_type_node, tmp), fold_convert (long_integer_type_node, size[n])); free (msg); } else size[n] = gfc_evaluate_now (tmp, &inner); } tmp = gfc_finish_block (&inner); /* For optional arguments, only check bounds if the argument is present. */ if (expr->symtree->n.sym->attr.optional || expr->symtree->n.sym->attr.not_always_present) tmp = build3_v (COND_EXPR, gfc_conv_expr_present (expr->symtree->n.sym), tmp, build_empty_stmt (input_location)); gfc_add_expr_to_block (&block, tmp); } tmp = gfc_finish_block (&block); gfc_add_expr_to_block (&outer_loop->pre, tmp); } for (loop = loop->nested; loop; loop = loop->next) gfc_conv_ss_startstride (loop); } /* Return true if both symbols could refer to the same data object. Does not take account of aliasing due to equivalence statements. */ static int symbols_could_alias (gfc_symbol *lsym, gfc_symbol *rsym, bool lsym_pointer, bool lsym_target, bool rsym_pointer, bool rsym_target) { /* Aliasing isn't possible if the symbols have different base types. */ if (gfc_compare_types (&lsym->ts, &rsym->ts) == 0) return 0; /* Pointers can point to other pointers and target objects. */ if ((lsym_pointer && (rsym_pointer || rsym_target)) || (rsym_pointer && (lsym_pointer || lsym_target))) return 1; /* Special case: Argument association, cf. F90 12.4.1.6, F2003 12.4.1.7 and F2008 12.5.2.13 items 3b and 4b. The pointer case (a) is already checked above. */ if (lsym_target && rsym_target && ((lsym->attr.dummy && !lsym->attr.contiguous && (!lsym->attr.dimension || lsym->as->type == AS_ASSUMED_SHAPE)) || (rsym->attr.dummy && !rsym->attr.contiguous && (!rsym->attr.dimension || rsym->as->type == AS_ASSUMED_SHAPE)))) return 1; return 0; } /* Return true if the two SS could be aliased, i.e. both point to the same data object. */ /* TODO: resolve aliases based on frontend expressions. */ static int gfc_could_be_alias (gfc_ss * lss, gfc_ss * rss) { gfc_ref *lref; gfc_ref *rref; gfc_expr *lexpr, *rexpr; gfc_symbol *lsym; gfc_symbol *rsym; bool lsym_pointer, lsym_target, rsym_pointer, rsym_target; lexpr = lss->info->expr; rexpr = rss->info->expr; lsym = lexpr->symtree->n.sym; rsym = rexpr->symtree->n.sym; lsym_pointer = lsym->attr.pointer; lsym_target = lsym->attr.target; rsym_pointer = rsym->attr.pointer; rsym_target = rsym->attr.target; if (symbols_could_alias (lsym, rsym, lsym_pointer, lsym_target, rsym_pointer, rsym_target)) return 1; if (rsym->ts.type != BT_DERIVED && rsym->ts.type != BT_CLASS && lsym->ts.type != BT_DERIVED && lsym->ts.type != BT_CLASS) return 0; /* For derived types we must check all the component types. We can ignore array references as these will have the same base type as the previous