aboutsummaryrefslogtreecommitdiffstats
path: root/gcc-4.9/gcc/config/ia64/ia64.c
blob: 4ec3e3abe6e3e10e927991cf0c01fbdf85604354 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
2049
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
2093
2094
2095
2096
2097
2098
2099
2100
2101
2102
2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
2150
2151
2152
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
2170
2171
2172
2173
2174
2175
2176
2177
2178
2179
2180
2181
2182
2183
2184
2185
2186
2187
2188
2189
2190
2191
2192
2193
2194
2195
2196
2197
2198
2199
2200
2201
2202
2203
2204
2205
2206
2207
2208
2209
2210
2211
2212
2213
2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
2228
2229
2230
2231
2232
2233
2234
2235
2236
2237
2238
2239
2240
2241
2242
2243
2244
2245
2246
2247
2248
2249
2250
2251
2252
2253
2254
2255
2256
2257
2258
2259
2260
2261
2262
2263
2264
2265
2266
2267
2268
2269
2270
2271
2272
2273
2274
2275
2276
2277
2278
2279
2280
2281
2282
2283
2284
2285
2286
2287
2288
2289
2290
2291
2292
2293
2294
2295
2296
2297
2298
2299
2300
2301
2302
2303
2304
2305
2306
2307
2308
2309
2310
2311
2312
2313
2314
2315
2316
2317
2318
2319
2320
2321
2322
2323
2324
2325
2326
2327
2328
2329
2330
2331
2332
2333
2334
2335
2336
2337
2338
2339
2340
2341
2342
2343
2344
2345
2346
2347
2348
2349
2350
2351
2352
2353
2354
2355
2356
2357
2358
2359
2360
2361
2362
2363
2364
2365
2366
2367
2368
2369
2370
2371
2372
2373
2374
2375
2376
2377
2378
2379
2380
2381
2382
2383
2384
2385
2386
2387
2388
2389
2390
2391
2392
2393
2394
2395
2396
2397
2398
2399
2400
2401
2402
2403
2404
2405
2406
2407
2408
2409
2410
2411
2412
2413
2414
2415
2416
2417
2418
2419
2420
2421
2422
2423
2424
2425
2426
2427
2428
2429
2430
2431
2432
2433
2434
2435
2436
2437
2438
2439
2440
2441
2442
2443
2444
2445
2446
2447
2448
2449
2450
2451
2452
2453
2454
2455
2456
2457
2458
2459
2460
2461
2462
2463
2464
2465
2466
2467
2468
2469
2470
2471
2472
2473
2474
2475
2476
2477
2478
2479
2480
2481
2482
2483
2484
2485
2486
2487
2488
2489
2490
2491
2492
2493
2494
2495
2496
2497
2498
2499
2500
2501
2502
2503
2504
2505
2506
2507
2508
2509
2510
2511
2512
2513
2514
2515
2516
2517
2518
2519
2520
2521
2522
2523
2524
2525
2526
2527
2528
2529
2530
2531
2532
2533
2534
2535
2536
2537
2538
2539
2540
2541
2542
2543
2544
2545
2546
2547
2548
2549
2550
2551
2552
2553
2554
2555
2556
2557
2558
2559
2560
2561
2562
2563
2564
2565
2566
2567
2568
2569
2570
2571
2572
2573
2574
2575
2576
2577
2578
2579
2580
2581
2582
2583
2584
2585
2586
2587
2588
2589
2590
2591
2592
2593
2594
2595
2596
2597
2598
2599
2600
2601
2602
2603
2604
2605
2606
2607
2608
2609
2610
2611
2612
2613
2614
2615
2616
2617
2618
2619
2620
2621
2622
2623
2624
2625
2626
2627
2628
2629
2630
2631
2632
2633
2634
2635
2636
2637
2638
2639
2640
2641
2642
2643
2644
2645
2646
2647
2648
2649
2650
2651
2652
2653
2654
2655
2656
2657
2658
2659
2660
2661
2662
2663
2664
2665
2666
2667
2668
2669
2670
2671
2672
2673
2674
2675
2676
2677
2678
2679
2680
2681
2682
2683
2684
2685
2686
2687
2688
2689
2690
2691
2692
2693
2694
2695
2696
2697
2698
2699
2700
2701
2702
2703
2704
2705
2706
2707
2708
2709
2710
2711
2712
2713
2714
2715
2716
2717
2718
2719
2720
2721
2722
2723
2724
2725
2726
2727
2728
2729
2730
2731
2732
2733
2734
2735
2736
2737
2738
2739
2740
2741
2742
2743
2744
2745
2746
2747
2748
2749
2750
2751
2752
2753
2754
2755
2756
2757
2758
2759
2760
2761
2762
2763
2764
2765
2766
2767
2768
2769
2770
2771
2772
2773
2774
2775
2776
2777
2778
2779
2780
2781
2782
2783
2784
2785
2786
2787
2788
2789
2790
2791
2792
2793
2794
2795
2796
2797
2798
2799
2800
2801
2802
2803
2804
2805
2806
2807
2808
2809
2810
2811
2812
2813
2814
2815
2816
2817
2818
2819
2820
2821
2822
2823
2824
2825
2826
2827
2828
2829
2830
2831
2832
2833
2834
2835
2836
2837
2838
2839
2840
2841
2842
2843
2844
2845
2846
2847
2848
2849
2850
2851
2852
2853
2854
2855
2856
2857
2858
2859
2860
2861
2862
2863
2864
2865
2866
2867
2868
2869
2870
2871
2872
2873
2874
2875
2876
2877
2878
2879
2880
2881
2882
2883
2884
2885
2886
2887
2888
2889
2890
2891
2892
2893
2894
2895
2896
2897
2898
2899
2900
2901
2902
2903
2904
2905
2906
2907
2908
2909
2910
2911
2912
2913
2914
2915
2916
2917
2918
2919
2920
2921
2922
2923
2924
2925
2926
2927
2928
2929
2930
2931
2932
2933
2934
2935
2936
2937
2938
2939
2940
2941
2942
2943
2944
2945
2946
2947
2948
2949
2950
2951
2952
2953
2954
2955
2956
2957
2958
2959
2960
2961
2962
2963
2964
2965
2966
2967
2968
2969
2970
2971
2972
2973
2974
2975
2976
2977
2978
2979
2980
2981
2982
2983
2984
2985
2986
2987
2988
2989
2990
2991
2992
2993
2994
2995
2996
2997
2998
2999
3000
3001
3002
3003
3004
3005
3006
3007
3008
3009
3010
3011
3012
3013
3014
3015
3016
3017
3018
3019
3020
3021
3022
3023
3024
3025
3026
3027
3028
3029
3030
3031
3032
3033
3034
3035
3036
3037
3038
3039
3040
3041
3042
3043
3044
3045
3046
3047
3048
3049
3050
3051
3052
3053
3054
3055
3056
3057
3058
3059
3060
3061
3062
3063
3064
3065
3066
3067
3068
3069
3070
3071
3072
3073
3074
3075
3076
3077
3078
3079
3080
3081
3082
3083
3084
3085
3086
3087
3088
3089
3090
3091
3092
3093
3094
3095
3096
3097
3098
3099
3100
3101
3102
3103
3104
3105
3106
3107
3108
3109
3110
3111
3112
3113
3114
3115
3116
3117
3118
3119
3120
3121
3122
3123
3124
3125
3126
3127
3128
3129
3130
3131
3132
3133
3134
3135
3136
3137
3138
3139
3140
3141
3142
3143
3144
3145
3146
3147
3148
3149
3150
3151
3152
3153
3154
3155
3156
3157
3158
3159
3160
3161
3162
3163
3164
3165
3166
3167
3168
3169
3170
3171
3172
3173
3174
3175
3176
3177
3178
3179
3180
3181
3182
3183
3184
3185
3186
3187
3188
3189
3190
3191
3192
3193
3194
3195
3196
3197
3198
3199
3200
3201
3202
3203
3204
3205
3206
3207
3208
3209
3210
3211
3212
3213
3214
3215
3216
3217
3218
3219
3220
3221
3222
3223
3224
3225
3226
3227
3228
3229
3230
3231
3232
3233
3234
3235
3236
3237
3238
3239
3240
3241
3242
3243
3244
3245
3246
3247
3248
3249
3250
3251
3252
3253
3254
3255
3256
3257
3258
3259
3260
3261
3262
3263
3264
3265
3266
3267
3268
3269
3270
3271
3272
3273
3274
3275
3276
3277
3278
3279
3280
3281
3282
3283
3284
3285
3286
3287
3288
3289
3290
3291
3292
3293
3294
3295
3296
3297
3298
3299
3300
3301
3302
3303
3304
3305
3306
3307
3308
3309
3310
3311
3312
3313
3314
3315
3316
3317
3318
3319
3320
3321
3322
3323
3324
3325
3326
3327
3328
3329
3330
3331
3332
3333
3334
3335
3336
3337
3338
3339
3340
3341
3342
3343
3344
3345
3346
3347
3348
3349
3350
3351
3352
3353
3354
3355
3356
3357
3358
3359
3360
3361
3362
3363
3364
3365
3366
3367
3368
3369
3370
3371
3372
3373
3374
3375
3376
3377
3378
3379
3380
3381
3382
3383
3384
3385
3386
3387
3388
3389
3390
3391
3392
3393
3394
3395
3396
3397
3398
3399
3400
3401
3402
3403
3404
3405
3406
3407
3408
3409
3410
3411
3412
3413
3414
3415
3416
3417
3418
3419
3420
3421
3422
3423
3424
3425
3426
3427
3428
3429
3430
3431
3432
3433
3434
3435
3436
3437
3438
3439
3440
3441
3442
3443
3444
3445
3446
3447
3448
3449
3450
3451
3452
3453
3454
3455
3456
3457
3458
3459
3460
3461
3462
3463
3464
3465
3466
3467
3468
3469
3470
3471
3472
3473
3474
3475
3476
3477
3478
3479
3480
3481
3482
3483
3484
3485
3486
3487
3488
3489
3490
3491
3492
3493
3494
3495
3496
3497
3498
3499
3500
3501
3502
3503
3504
3505
3506
3507
3508
3509
3510
3511
3512
3513
3514
3515
3516
3517
3518
3519
3520
3521
3522
3523
3524
3525
3526
3527
3528
3529
3530
3531
3532
3533
3534
3535
3536
3537
3538
3539
3540
3541
3542
3543
3544
3545
3546
3547
3548
3549
3550
3551
3552
3553
3554
3555
3556
3557
3558
3559
3560
3561
3562
3563
3564
3565
3566
3567
3568
3569
3570
3571
3572
3573
3574
3575
3576
3577
3578
3579
3580
3581
3582
3583
3584
3585
3586
3587
3588
3589
3590
3591
3592
3593
3594
3595
3596
3597
3598
3599
3600
3601
3602
3603
3604
3605
3606
3607
3608
3609
3610
3611
3612
3613
3614
3615
3616
3617
3618
3619
3620
3621
3622
3623
3624
3625
3626
3627
3628
3629
3630
3631
3632
3633
3634
3635
3636
3637
3638
3639
3640
3641
3642
3643
3644
3645
3646
3647
3648
3649
3650
3651
3652
3653
3654
3655
3656
3657
3658
3659
3660
3661
3662
3663
3664
3665
3666
3667
3668
3669
3670
3671
3672
3673
3674
3675
3676
3677
3678
3679
3680
3681
3682
3683
3684
3685
3686
3687
3688
3689
3690
3691
3692
3693
3694
3695
3696
3697
3698
3699
3700
3701
3702
3703
3704
3705
3706
3707
3708
3709
3710
3711
3712
3713
3714
3715
3716
3717
3718
3719
3720
3721
3722
3723
3724
3725
3726
3727
3728
3729
3730
3731
3732
3733
3734
3735
3736
3737
3738
3739
3740
3741
3742
3743
3744
3745
3746
3747
3748
3749
3750
3751
3752
3753
3754
3755
3756
3757
3758
3759
3760
3761
3762
3763
3764
3765
3766
3767
3768
3769
3770
3771
3772
3773
3774
3775
3776
3777
3778
3779
3780
3781
3782
3783
3784
3785
3786
3787
3788
3789
3790
3791
3792
3793
3794
3795
3796
3797
3798
3799
3800
3801
3802
3803
3804
3805
3806
3807
3808
3809
3810
3811
3812
3813
3814
3815
3816
3817
3818
3819
3820
3821
3822
3823
3824
3825
3826
3827
3828
3829
3830
3831
3832
3833
3834
3835
3836
3837
3838
3839
3840
3841
3842
3843
3844
3845
3846
3847
3848
3849
3850
3851
3852
3853
3854
3855
3856
3857
3858
3859
3860
3861
3862
3863
3864
3865
3866
3867
3868
3869
3870
3871
3872
3873
3874
3875
3876
3877
3878
3879
3880
3881
3882
3883
3884
3885
3886
3887
3888
3889
3890
3891
3892
3893
3894
3895
3896
3897
3898
3899
3900
3901
3902
3903
3904
3905
3906
3907
3908
3909
3910
3911
3912
3913
3914
3915
3916
3917
3918
3919
3920
3921
3922
3923
3924
3925
3926
3927
3928
3929
3930
3931
3932
3933
3934
3935
3936
3937
3938
3939
3940
3941
3942
3943
3944
3945
3946
3947
3948
3949
3950
3951
3952
3953
3954
3955
3956
3957
3958
3959
3960
3961
3962
3963
3964
3965
3966
3967
3968
3969
3970
3971
3972
3973
3974
3975
3976
3977
3978
3979
3980
3981
3982
3983
3984
3985
3986
3987
3988
3989
3990
3991
3992
3993
3994
3995
3996
3997
3998
3999
4000
4001
4002
4003
4004
4005
4006
4007
4008
4009
4010
4011
4012
4013
4014
4015
4016
4017
4018
4019
4020
4021
4022
4023
4024
4025
4026
4027
4028
4029
4030
4031
4032
4033
4034
4035
4036
4037
4038
4039
4040
4041
4042
4043
4044
4045
4046
4047
4048
4049
4050
4051
4052
4053
4054
4055
4056
4057
4058
4059
4060
4061
4062
4063
4064
4065
4066
4067
4068
4069
4070
4071
4072
4073
4074
4075
4076
4077
4078
4079
4080
4081
4082
4083
4084
4085
4086
4087
4088
4089
4090
4091
4092
4093
4094
4095
4096
4097
4098
4099
4100
4101
4102
4103
4104
4105
4106
4107
4108
4109
4110
4111
4112
4113
4114
4115
4116
4117
4118
4119
4120
4121
4122
4123
4124
4125
4126
4127
4128
4129
4130
4131
4132
4133
4134
4135
4136
4137
4138
4139
4140
4141
4142
4143
4144
4145
4146
4147
4148
4149
4150
4151
4152
4153
4154
4155
4156
4157
4158
4159
4160
4161
4162
4163
4164
4165
4166
4167
4168
4169
4170
4171
4172
4173
4174
4175
4176
4177
4178
4179
4180
4181
4182
4183
4184
4185
4186
4187
4188
4189
4190
4191
4192
4193
4194
4195
4196
4197
4198
4199
4200
4201
4202
4203
4204
4205
4206
4207
4208
4209
4210
4211
4212
4213
4214
4215
4216
4217
4218
4219
4220
4221
4222
4223
4224
4225
4226
4227
4228
4229
4230
4231
4232
4233
4234
4235
4236
4237
4238
4239
4240
4241
4242
4243
4244
4245
4246
4247
4248
4249
4250
4251
4252
4253
4254
4255
4256
4257
4258
4259
4260
4261
4262
4263
4264
4265
4266
4267
4268
4269
4270
4271
4272
4273
4274
4275
4276
4277
4278
4279
4280
4281
4282
4283
4284
4285
4286
4287
4288
4289
4290
4291
4292
4293
4294
4295
4296
4297
4298
4299
4300
4301
4302
4303
4304
4305
4306
4307
4308
4309
4310
4311
4312
4313
4314
4315
4316
4317
4318
4319
4320
4321
4322
4323
4324
4325
4326
4327
4328
4329
4330
4331
4332
4333
4334
4335
4336
4337
4338
4339
4340
4341
4342
4343
4344
4345
4346
4347
4348
4349
4350
4351
4352
4353
4354
4355
4356
4357
4358
4359
4360
4361
4362
4363
4364
4365
4366
4367
4368
4369
4370
4371
4372
4373
4374
4375
4376
4377
4378
4379
4380
4381
4382
4383
4384
4385
4386
4387
4388
4389
4390
4391
4392
4393
4394
4395
4396
4397
4398
4399
4400
4401
4402
4403
4404
4405
4406
4407
4408
4409
4410
4411
4412
4413
4414
4415
4416
4417
4418
4419
4420
4421
4422
4423
4424
4425
4426
4427
4428
4429
4430
4431
4432
4433
4434
4435
4436
4437
4438
4439
4440
4441
4442
4443
4444
4445
4446
4447
4448
4449
4450
4451
4452
4453
4454
4455
4456
4457
4458
4459
4460
4461
4462
4463
4464
4465
4466
4467
4468
4469
4470
4471
4472
4473
4474
4475
4476
4477
4478
4479
4480
4481
4482
4483
4484
4485
4486
4487
4488
4489
4490
4491
4492
4493
4494
4495
4496
4497
4498
4499
4500
4501
4502
4503
4504
4505
4506
4507
4508
4509
4510
4511
4512
4513
4514
4515
4516
4517
4518
4519
4520
4521
4522
4523
4524
4525
4526
4527
4528
4529
4530
4531
4532
4533
4534
4535
4536
4537
4538
4539
4540
4541
4542
4543
4544
4545
4546
4547
4548
4549
4550
4551
4552
4553
4554
4555
4556
4557
4558
4559
4560
4561
4562
4563
4564
4565
4566
4567
4568
4569
4570
4571
4572
4573
4574
4575
4576
4577
4578
4579
4580
4581
4582
4583
4584
4585
4586
4587
4588
4589
4590
4591
4592
4593
4594
4595
4596
4597
4598
4599
4600
4601
4602
4603
4604
4605
4606
4607
4608
4609
4610
4611
4612
4613
4614
4615
4616
4617
4618
4619
4620
4621
4622
4623
4624
4625
4626
4627
4628
4629
4630
4631
4632
4633
4634
4635
4636
4637
4638
4639
4640
4641
4642
4643
4644
4645
4646
4647
4648
4649
4650
4651
4652
4653
4654
4655
4656
4657
4658
4659
4660
4661
4662
4663
4664
4665
4666
4667
4668
4669
4670
4671
4672
4673
4674
4675
4676
4677
4678
4679
4680
4681
4682
4683
4684
4685
4686
4687
4688
4689
4690
4691
4692
4693
4694
4695
4696
4697
4698
4699
4700
4701
4702
4703
4704
4705
4706
4707
4708
4709
4710
4711
4712
4713
4714
4715
4716
4717
4718
4719
4720
4721
4722
4723
4724
4725
4726
4727
4728
4729
4730
4731
4732
4733
4734
4735
4736
4737
4738
4739
4740
4741
4742
4743
4744
4745
4746
4747
4748
4749
4750
4751
4752
4753
4754
4755
4756
4757
4758
4759
4760
4761
4762
4763
4764
4765
4766
4767
4768
4769
4770
4771
4772
4773
4774
4775
4776
4777
4778
4779
4780
4781
4782
4783
4784
4785
4786
4787
4788
4789
4790
4791
4792
4793
4794
4795
4796
4797
4798
4799
4800
4801
4802
4803
4804
4805
4806
4807
4808
4809
4810
4811
4812
4813
4814
4815
4816
4817
4818
4819
4820
4821
4822
4823
4824
4825
4826
4827
4828
4829
4830
4831
4832
4833
4834
4835
4836
4837
4838
4839
4840
4841
4842
4843
4844
4845
4846
4847
4848
4849
4850
4851
4852
4853
4854
4855
4856
4857
4858
4859
4860
4861
4862
4863
4864
4865
4866
4867
4868
4869
4870
4871
4872
4873
4874
4875
4876
4877
4878
4879
4880
4881
4882
4883
4884
4885
4886
4887
4888
4889
4890
4891
4892
4893
4894
4895
4896
4897
4898
4899
4900
4901
4902
4903
4904
4905
4906
4907
4908
4909
4910
4911
4912
4913
4914
4915
4916
4917
4918
4919
4920
4921
4922
4923
4924
4925
4926
4927
4928
4929
4930
4931
4932
4933
4934
4935
4936
4937
4938
4939
4940
4941
4942
4943
4944
4945
4946
4947
4948
4949
4950
4951
4952
4953
4954
4955
4956
4957
4958
4959
4960
4961
4962
4963
4964
4965
4966
4967
4968
4969
4970
4971
4972
4973
4974
4975
4976
4977
4978
4979
4980
4981
4982
4983
4984
4985
4986
4987
4988
4989
4990
4991
4992
4993
4994
4995
4996
4997
4998
4999
5000
5001
5002
5003
5004
5005
5006
5007
5008
5009
5010
5011
5012
5013
5014
5015
5016
5017
5018
5019
5020
5021
5022
5023
5024
5025
5026
5027
5028
5029
5030
5031
5032
5033
5034
5035
5036
5037
5038
5039
5040
5041
5042
5043
5044
5045
5046
5047
5048
5049
5050
5051
5052
5053
5054
5055
5056
5057
5058
5059
5060
5061
5062
5063
5064
5065
5066
5067
5068
5069
5070
5071
5072
5073
5074
5075
5076
5077
5078
5079
5080
5081
5082
5083
5084
5085
5086
5087
5088
5089
5090
5091
5092
5093
5094
5095
5096
5097
5098
5099
5100
5101
5102
5103
5104
5105
5106
5107
5108
5109
5110
5111
5112
5113
5114
5115
5116
5117
5118
5119
5120
5121
5122
5123
5124
5125
5126
5127
5128
5129
5130
5131
5132
5133
5134
5135
5136
5137
5138
5139
5140
5141
5142
5143
5144
5145
5146
5147
5148
5149
5150
5151
5152
5153
5154
5155
5156
5157
5158
5159
5160
5161
5162
5163
5164
5165
5166
5167
5168
5169
5170
5171
5172
5173
5174
5175
5176
5177
5178
5179
5180
5181
5182
5183
5184
5185
5186
5187
5188
5189
5190
5191
5192
5193
5194
5195
5196
5197
5198
5199
5200
5201
5202
5203
5204
5205
5206
5207
5208
5209
5210
5211
5212
5213
5214
5215
5216
5217
5218
5219
5220
5221
5222
5223
5224
5225
5226
5227
5228
5229
5230
5231
5232
5233
5234
5235
5236
5237
5238
5239
5240
5241
5242
5243
5244
5245
5246
5247
5248
5249
5250
5251
5252
5253
5254
5255
5256
5257
5258
5259
5260
5261
5262
5263
5264
5265
5266
5267
5268
5269
5270
5271
5272
5273
5274
5275
5276
5277
5278
5279
5280
5281
5282
5283
5284
5285
5286
5287
5288
5289
5290
5291
5292
5293
5294
5295
5296
5297
5298
5299
5300
5301
5302
5303
5304
5305
5306
5307
5308
5309
5310
5311
5312
5313
5314
5315
5316
5317
5318
5319
5320
5321
5322
5323
5324
5325
5326
5327
5328
5329
5330
5331
5332
5333
5334
5335
5336
5337
5338
5339
5340
5341
5342
5343
5344
5345
5346
5347
5348
5349
5350
5351
5352
5353
5354
5355
5356
5357
5358
5359
5360
5361
5362
5363
5364
5365
5366
5367
5368
5369
5370
5371
5372
5373
5374
5375
5376
5377
5378
5379
5380
5381
5382
5383
5384
5385
5386
5387
5388
5389
5390
5391
5392
5393
5394
5395
5396
5397
5398
5399
5400
5401
5402
5403
5404
5405
5406
5407
5408
5409
5410
5411
5412
5413
5414
5415
5416
5417
5418
5419
5420
5421
5422
5423
5424
5425
5426
5427
5428
5429
5430
5431
5432
5433
5434
5435
5436
5437
5438
5439
5440
5441
5442
5443
5444
5445
5446
5447
5448
5449
5450
5451
5452
5453
5454
5455
5456
5457
5458
5459
5460
5461
5462
5463
5464
5465
5466
5467
5468
5469
5470
5471
5472
5473
5474
5475
5476
5477
5478
5479
5480
5481
5482
5483
5484
5485
5486
5487
5488
5489
5490
5491
5492
5493
5494
5495
5496
5497
5498
5499
5500
5501
5502
5503
5504
5505
5506
5507
5508
5509
5510
5511
5512
5513
5514
5515
5516
5517
5518
5519
5520
5521
5522
5523
5524
5525
5526
5527
5528
5529
5530
5531
5532
5533
5534
5535
5536
5537
5538
5539
5540
5541
5542
5543
5544
5545
5546
5547
5548
5549
5550
5551
5552
5553
5554
5555
5556
5557
5558
5559
5560
5561
5562
5563
5564
5565
5566
5567
5568
5569
5570
5571
5572
5573
5574
5575
5576
5577
5578
5579
5580
5581
5582
5583
5584
5585
5586
5587
5588
5589
5590
5591
5592
5593
5594
5595
5596
5597
5598
5599
5600
5601
5602
5603
5604
5605
5606
5607
5608
5609
5610
5611
5612
5613
5614
5615
5616
5617
5618
5619
5620
5621
5622
5623
5624
5625
5626
5627
5628
5629
5630
5631
5632
5633
5634
5635
5636
5637
5638
5639
5640
5641
5642
5643
5644
5645
5646
5647
5648
5649
5650
5651
5652
5653
5654
5655
5656
5657
5658
5659
5660
5661
5662
5663
5664
5665
5666
5667
5668
5669
5670
5671
5672
5673
5674
5675
5676
5677
5678
5679
5680
5681
5682
5683
5684
5685
5686
5687
5688
5689
5690
5691
5692
5693
5694
5695
5696
5697
5698
5699
5700
5701
5702
5703
5704
5705
5706
5707
5708
5709
5710
5711
5712
5713
5714
5715
5716
5717
5718
5719
5720
5721
5722
5723
5724
5725
5726
5727
5728
5729
5730
5731
5732
5733
5734
5735
5736
5737
5738
5739
5740
5741
5742
5743
5744
5745
5746
5747
5748
5749
5750
5751
5752
5753
5754
5755
5756
5757
5758
5759
5760
5761
5762
5763
5764
5765
5766
5767
5768
5769
5770
5771
5772
5773
5774
5775
5776
5777
5778
5779
5780
5781
5782
5783
5784
5785
5786
5787
5788
5789
5790
5791
5792
5793
5794
5795
5796
5797
5798
5799
5800
5801
5802
5803
5804
5805
5806
5807
5808
5809
5810
5811
5812
5813
5814
5815
5816
5817
5818
5819
5820
5821
5822
5823
5824
5825
5826
5827
5828
5829
5830
5831
5832
5833
5834
5835
5836
5837
5838
5839
5840
5841
5842
5843
5844
5845
5846
5847
5848
5849
5850
5851
5852
5853
5854
5855
5856
5857
5858
5859
5860
5861
5862
5863
5864
5865
5866
5867
5868
5869
5870
5871
5872
5873
5874
5875
5876
5877
5878
5879
5880
5881
5882
5883
5884
5885
5886
5887
5888
5889
5890
5891
5892
5893
5894
5895
5896
5897
5898
5899
5900
5901
5902
5903
5904
5905
5906
5907
5908
5909
5910
5911
5912
5913
5914
5915
5916
5917
5918
5919
5920
5921
5922
5923
5924
5925
5926
5927
5928
5929
5930
5931
5932
5933
5934
5935
5936
5937
5938
5939
5940
5941
5942
5943
5944
5945
5946
5947
5948
5949
5950
5951
5952
5953
5954
5955
5956
5957
5958
5959
5960
5961
5962
5963
5964
5965
5966
5967
5968
5969
5970
5971
5972
5973
5974
5975
5976
5977
5978
5979
5980
5981
5982
5983
5984
5985
5986
5987
5988
5989
5990
5991
5992
5993
5994
5995
5996
5997
5998
5999
6000
6001
6002
6003
6004
6005
6006
6007
6008
6009
6010
6011
6012
6013
6014
6015
6016
6017
6018
6019
6020
6021
6022
6023
6024
6025
6026
6027
6028
6029
6030
6031
6032
6033
6034
6035
6036
6037
6038
6039
6040
6041
6042
6043
6044
6045
6046
6047
6048
6049
6050
6051
6052
6053
6054
6055
6056
6057
6058
6059
6060
6061
6062
6063
6064
6065
6066
6067
6068
6069
6070
6071
6072
6073
6074
6075
6076
6077
6078
6079
6080
6081
6082
6083
6084
6085
6086
6087
6088
6089
6090
6091
6092
6093
6094
6095
6096
6097
6098
6099
6100
6101
6102
6103
6104
6105
6106
6107
6108
6109
6110
6111
6112
6113
6114
6115
6116
6117
6118
6119
6120
6121
6122
6123
6124
6125
6126
6127
6128
6129
6130
6131
6132
6133
6134
6135
6136
6137
6138
6139
6140
6141
6142
6143
6144
6145
6146
6147
6148
6149
6150
6151
6152
6153
6154
6155
6156
6157
6158
6159
6160
6161
6162
6163
6164
6165
6166
6167
6168
6169
6170
6171
6172
6173
6174
6175
6176
6177
6178
6179
6180
6181
6182
6183
6184
6185
6186
6187
6188
6189
6190
6191
6192
6193
6194
6195
6196
6197
6198
6199
6200
6201
6202
6203
6204
6205
6206
6207
6208
6209
6210
6211
6212
6213
6214
6215
6216
6217
6218
6219
6220
6221
6222
6223
6224
6225
6226
6227
6228
6229
6230
6231
6232
6233
6234
6235
6236
6237
6238
6239
6240
6241
6242
6243
6244
6245
6246
6247
6248
6249
6250
6251
6252
6253
6254
6255
6256
6257
6258
6259
6260
6261
6262
6263
6264
6265
6266
6267
6268
6269
6270
6271
6272
6273
6274
6275
6276
6277
6278
6279
6280
6281
6282
6283
6284
6285
6286
6287
6288
6289
6290
6291
6292
6293
6294
6295
6296
6297
6298
6299
6300
6301
6302
6303
6304
6305
6306
6307
6308
6309
6310
6311
6312
6313
6314
6315
6316
6317
6318
6319
6320
6321
6322
6323
6324
6325
6326
6327
6328
6329
6330
6331
6332
6333
6334
6335
6336
6337
6338
6339
6340
6341
6342
6343
6344
6345
6346
6347
6348
6349
6350
6351
6352
6353
6354
6355
6356
6357
6358
6359
6360
6361
6362
6363
6364
6365
6366
6367
6368
6369
6370
6371
6372
6373
6374
6375
6376
6377
6378
6379
6380
6381
6382
6383
6384
6385
6386
6387
6388
6389
6390
6391
6392
6393
6394
6395
6396
6397
6398
6399
6400
6401
6402
6403
6404
6405
6406
6407
6408
6409
6410
6411
6412
6413
6414
6415
6416
6417
6418
6419
6420
6421
6422
6423
6424
6425
6426
6427
6428
6429
6430
6431
6432
6433
6434
6435
6436
6437
6438
6439
6440
6441
6442
6443
6444
6445
6446
6447
6448
6449
6450
6451
6452
6453
6454
6455
6456
6457
6458
6459
6460
6461
6462
6463
6464
6465
6466
6467
6468
6469
6470
6471
6472
6473
6474
6475
6476
6477
6478
6479
6480
6481
6482
6483
6484
6485
6486
6487
6488
6489
6490
6491
6492
6493
6494
6495
6496
6497
6498
6499
6500
6501
6502
6503
6504
6505
6506
6507
6508
6509
6510
6511
6512
6513
6514
6515
6516
6517
6518
6519
6520
6521
6522
6523
6524
6525
6526
6527
6528
6529
6530
6531
6532
6533
6534
6535
6536
6537
6538
6539
6540
6541
6542
6543
6544
6545
6546
6547
6548
6549
6550
6551
6552
6553
6554
6555
6556
6557
6558
6559
6560
6561
6562
6563
6564
6565
6566
6567
6568
6569
6570
6571
6572
6573
6574
6575
6576
6577
6578
6579
6580
6581
6582
6583
6584
6585
6586
6587
6588
6589
6590
6591
6592
6593
6594
6595
6596
6597
6598
6599
6600
6601
6602
6603
6604
6605
6606
6607
6608
6609
6610
6611
6612
6613
6614
6615
6616
6617
6618
6619
6620
6621
6622
6623
6624
6625
6626
6627
6628
6629
6630
6631
6632
6633
6634
6635
6636
6637
6638
6639
6640
6641
6642
6643
6644
6645
6646
6647
6648
6649
6650
6651
6652
6653
6654
6655
6656
6657
6658
6659
6660
6661
6662
6663
6664
6665
6666
6667
6668
6669
6670
6671
6672
6673
6674
6675
6676
6677
6678
6679
6680
6681
6682
6683
6684
6685
6686
6687
6688
6689
6690
6691
6692
6693
6694
6695
6696
6697
6698
6699
6700
6701
6702
6703
6704
6705
6706
6707
6708
6709
6710
6711
6712
6713
6714
6715
6716
6717
6718
6719
6720
6721
6722
6723
6724
6725
6726
6727
6728
6729
6730
6731
6732
6733
6734
6735
6736
6737
6738
6739
6740
6741
6742
6743
6744
6745
6746
6747
6748
6749
6750
6751
6752
6753
6754
6755
6756
6757
6758
6759
6760
6761
6762
6763
6764
6765
6766
6767
6768
6769
6770
6771
6772
6773
6774
6775
6776
6777
6778
6779
6780
6781
6782
6783
6784
6785
6786
6787
6788
6789
6790
6791
6792
6793
6794
6795
6796
6797
6798
6799
6800
6801
6802
6803
6804
6805
6806
6807
6808
6809
6810
6811
6812
6813
6814
6815
6816
6817
6818
6819
6820
6821
6822
6823
6824
6825
6826
6827
6828
6829
6830
6831
6832
6833
6834
6835
6836
6837
6838
6839
6840
6841
6842
6843
6844
6845
6846
6847
6848
6849
6850
6851
6852
6853
6854
6855
6856
6857
6858
6859
6860
6861
6862
6863
6864
6865
6866
6867
6868
6869
6870
6871
6872
6873
6874
6875
6876
6877
6878
6879
6880
6881
6882
6883
6884
6885
6886
6887
6888
6889
6890
6891
6892
6893
6894
6895
6896
6897
6898
6899
6900
6901
6902
6903
6904
6905
6906
6907
6908
6909
6910
6911
6912
6913
6914
6915
6916
6917
6918
6919
6920
6921
6922
6923
6924
6925
6926
6927
6928
6929
6930
6931
6932
6933
6934
6935
6936
6937
6938
6939
6940
6941
6942
6943
6944
6945
6946
6947
6948
6949
6950
6951
6952
6953
6954
6955
6956
6957
6958
6959
6960
6961
6962
6963
6964
6965
6966
6967
6968
6969
6970
6971
6972
6973
6974
6975
6976
6977
6978
6979
6980
6981
6982
6983
6984
6985
6986
6987
6988
6989
6990
6991
6992
6993
6994
6995
6996
6997
6998
6999
7000
7001
7002
7003
7004
7005
7006
7007
7008
7009
7010
7011
7012
7013
7014
7015
7016
7017
7018
7019
7020
7021
7022
7023
7024
7025
7026
7027
7028
7029
7030
7031
7032
7033
7034
7035
7036
7037
7038
7039
7040
7041
7042
7043
7044
7045
7046
7047
7048
7049
7050
7051
7052
7053
7054
7055
7056
7057
7058
7059
7060
7061
7062
7063
7064
7065
7066
7067
7068
7069
7070
7071
7072
7073
7074
7075
7076
7077
7078
7079
7080
7081
7082
7083
7084
7085
7086
7087
7088
7089
7090
7091
7092
7093
7094
7095
7096
7097
7098
7099
7100
7101
7102
7103
7104
7105
7106
7107
7108
7109
7110
7111
7112
7113
7114
7115
7116
7117
7118
7119
7120
7121
7122
7123
7124
7125
7126
7127
7128
7129
7130
7131
7132
7133
7134
7135
7136
7137
7138
7139
7140
7141
7142
7143
7144
7145
7146
7147
7148
7149
7150
7151
7152
7153
7154
7155
7156
7157
7158
7159
7160
7161
7162
7163
7164
7165
7166
7167
7168
7169
7170
7171
7172
7173
7174
7175
7176
7177
7178
7179
7180
7181
7182
7183
7184
7185
7186
7187
7188
7189
7190
7191
7192
7193
7194
7195
7196
7197
7198
7199
7200
7201
7202
7203
7204
7205
7206
7207
7208
7209
7210
7211
7212
7213
7214
7215
7216
7217
7218
7219
7220
7221
7222
7223
7224
7225
7226
7227
7228
7229
7230
7231
7232
7233
7234
7235
7236
7237
7238
7239
7240
7241
7242
7243
7244
7245
7246
7247
7248
7249
7250
7251
7252
7253
7254
7255
7256
7257
7258
7259
7260
7261
7262
7263
7264
7265
7266
7267
7268
7269
7270
7271
7272
7273
7274
7275
7276
7277
7278
7279
7280
7281
7282
7283
7284
7285
7286
7287
7288
7289
7290
7291
7292
7293
7294
7295
7296
7297
7298
7299
7300
7301
7302
7303
7304
7305
7306
7307
7308
7309
7310
7311
7312
7313
7314
7315
7316
7317
7318
7319
7320
7321
7322
7323
7324
7325
7326
7327
7328
7329
7330
7331
7332
7333
7334
7335
7336
7337
7338
7339
7340
7341
7342
7343
7344
7345
7346
7347
7348
7349
7350
7351
7352
7353
7354
7355
7356
7357
7358
7359
7360
7361
7362
7363
7364
7365
7366
7367
7368
7369
7370
7371
7372
7373
7374
7375
7376
7377
7378
7379
7380
7381
7382
7383
7384
7385
7386
7387
7388
7389
7390
7391
7392
7393
7394
7395
7396
7397
7398
7399
7400
7401
7402
7403
7404
7405
7406
7407
7408
7409
7410
7411
7412
7413
7414
7415
7416
7417
7418
7419
7420
7421
7422
7423
7424
7425
7426
7427
7428
7429
7430
7431
7432
7433
7434
7435
7436
7437
7438
7439
7440
7441
7442
7443
7444
7445
7446
7447
7448
7449
7450
7451
7452
7453
7454
7455
7456
7457
7458
7459
7460
7461
7462
7463
7464
7465
7466
7467
7468
7469
7470
7471
7472
7473
7474
7475
7476
7477
7478
7479
7480
7481
7482
7483
7484
7485
7486
7487
7488
7489
7490
7491
7492
7493
7494
7495
7496
7497
7498
7499
7500
7501
7502
7503
7504
7505
7506
7507
7508
7509
7510
7511
7512
7513
7514
7515
7516
7517
7518
7519
7520
7521
7522
7523
7524
7525
7526
7527
7528
7529
7530
7531
7532
7533
7534
7535
7536
7537
7538
7539
7540
7541
7542
7543
7544
7545
7546
7547
7548
7549
7550
7551
7552
7553
7554
7555
7556
7557
7558
7559
7560
7561
7562
7563
7564
7565
7566
7567
7568
7569
7570
7571
7572
7573
7574
7575
7576
7577
7578
7579
7580
7581
7582
7583
7584
7585
7586
7587
7588
7589
7590
7591
7592
7593
7594
7595
7596
7597
7598
7599
7600
7601
7602
7603
7604
7605
7606
7607
7608
7609
7610
7611
7612
7613
7614
7615
7616
7617
7618
7619
7620
7621
7622
7623
7624
7625
7626
7627
7628
7629
7630
7631
7632
7633
7634
7635
7636
7637
7638
7639
7640
7641
7642
7643
7644
7645
7646
7647
7648
7649
7650
7651
7652
7653
7654
7655
7656
7657
7658
7659
7660
7661
7662
7663
7664
7665
7666
7667
7668
7669
7670
7671
7672
7673
7674
7675
7676
7677
7678
7679
7680
7681
7682
7683
7684
7685
7686
7687
7688
7689
7690
7691
7692
7693
7694
7695
7696
7697
7698
7699
7700
7701
7702
7703
7704
7705
7706
7707
7708
7709
7710
7711
7712
7713
7714
7715
7716
7717
7718
7719
7720
7721
7722
7723
7724
7725
7726
7727
7728
7729
7730
7731
7732
7733
7734
7735
7736
7737
7738
7739
7740
7741
7742
7743
7744
7745
7746
7747
7748
7749
7750
7751
7752
7753
7754
7755
7756
7757
7758
7759
7760
7761
7762
7763
7764
7765
7766
7767
7768
7769
7770
7771
7772
7773
7774
7775
7776
7777
7778
7779
7780
7781
7782
7783
7784
7785
7786
7787
7788
7789
7790
7791
7792
7793
7794
7795
7796
7797
7798
7799
7800
7801
7802
7803
7804
7805
7806
7807
7808
7809
7810
7811
7812
7813
7814
7815
7816
7817
7818
7819
7820
7821
7822
7823
7824
7825
7826
7827
7828
7829
7830
7831
7832
7833
7834
7835
7836
7837
7838
7839
7840
7841
7842
7843
7844
7845
7846
7847
7848
7849
7850
7851
7852
7853
7854
7855
7856
7857
7858
7859
7860
7861
7862
7863
7864
7865
7866
7867
7868
7869
7870
7871
7872
7873
7874
7875
7876
7877
7878
7879
7880
7881
7882
7883
7884
7885
7886
7887
7888
7889
7890
7891
7892
7893
7894
7895
7896
7897
7898
7899
7900
7901
7902
7903
7904
7905
7906
7907
7908
7909
7910
7911
7912
7913
7914
7915
7916
7917
7918
7919
7920
7921
7922
7923
7924
7925
7926
7927
7928
7929
7930
7931
7932
7933
7934
7935
7936
7937
7938
7939
7940
7941
7942
7943
7944
7945
7946
7947
7948
7949
7950
7951
7952
7953
7954
7955
7956
7957
7958
7959
7960
7961
7962
7963
7964
7965
7966
7967
7968
7969
7970
7971
7972
7973
7974
7975
7976
7977
7978
7979
7980
7981
7982
7983
7984
7985
7986
7987
7988
7989
7990
7991
7992
7993
7994
7995
7996
7997
7998
7999
8000
8001
8002
8003
8004
8005
8006
8007
8008
8009
8010
8011
8012
8013
8014
8015
8016
8017
8018
8019
8020
8021
8022
8023
8024
8025
8026
8027
8028
8029
8030
8031
8032
8033
8034
8035
8036
8037
8038
8039
8040
8041
8042
8043
8044
8045
8046
8047
8048
8049
8050
8051
8052
8053
8054
8055
8056
8057
8058
8059
8060
8061
8062
8063
8064
8065
8066
8067
8068
8069
8070
8071
8072
8073
8074
8075
8076
8077
8078
8079
8080
8081
8082
8083
8084
8085
8086
8087
8088
8089
8090
8091
8092
8093
8094
8095
8096
8097
8098
8099
8100
8101
8102
8103
8104
8105
8106
8107
8108
8109
8110
8111
8112
8113
8114
8115
8116
8117
8118
8119
8120
8121
8122
8123
8124
8125
8126
8127
8128
8129
8130
8131
8132
8133
8134
8135
8136
8137
8138
8139
8140
8141
8142
8143
8144
8145
8146
8147
8148
8149
8150
8151
8152
8153
8154
8155
8156
8157
8158
8159
8160
8161
8162
8163
8164
8165
8166
8167
8168
8169
8170
8171
8172
8173
8174
8175
8176
8177
8178
8179
8180
8181
8182
8183
8184
8185
8186
8187
8188
8189
8190
8191
8192
8193
8194
8195
8196
8197
8198
8199
8200
8201
8202
8203
8204
8205
8206
8207
8208
8209
8210
8211
8212
8213
8214
8215
8216
8217
8218
8219
8220
8221
8222
8223
8224
8225
8226
8227
8228
8229
8230
8231
8232
8233
8234
8235
8236
8237
8238
8239
8240
8241
8242
8243
8244
8245
8246
8247
8248
8249
8250
8251
8252
8253
8254
8255
8256
8257
8258
8259
8260
8261
8262
8263
8264
8265
8266
8267
8268
8269
8270
8271
8272
8273
8274
8275
8276
8277
8278
8279
8280
8281
8282
8283
8284
8285
8286
8287
8288
8289
8290
8291
8292
8293
8294
8295
8296
8297
8298
8299
8300
8301
8302
8303
8304
8305
8306
8307
8308
8309
8310
8311
8312
8313
8314
8315
8316
8317
8318
8319
8320
8321
8322
8323
8324
8325
8326
8327
8328
8329
8330
8331
8332
8333
8334
8335
8336
8337
8338
8339
8340
8341
8342
8343
8344
8345
8346
8347
8348
8349
8350
8351
8352
8353
8354
8355
8356
8357
8358
8359
8360
8361
8362
8363
8364
8365
8366
8367
8368
8369
8370
8371
8372
8373
8374
8375
8376
8377
8378
8379
8380
8381
8382
8383
8384
8385
8386
8387
8388
8389
8390
8391
8392
8393
8394
8395
8396
8397
8398
8399
8400
8401
8402
8403
8404
8405
8406
8407
8408
8409
8410
8411
8412
8413
8414
8415
8416
8417
8418
8419
8420
8421
8422
8423
8424
8425
8426
8427
8428
8429
8430
8431
8432
8433
8434
8435
8436
8437
8438
8439
8440
8441
8442
8443
8444
8445
8446
8447
8448
8449
8450
8451
8452
8453
8454
8455
8456
8457
8458
8459
8460
8461
8462
8463
8464
8465
8466
8467
8468
8469
8470
8471
8472
8473
8474
8475
8476
8477
8478
8479
8480
8481
8482
8483
8484
8485
8486
8487
8488
8489
8490
8491
8492
8493
8494
8495
8496
8497
8498
8499
8500
8501
8502
8503
8504
8505
8506
8507
8508
8509
8510
8511
8512
8513
8514
8515
8516
8517
8518
8519
8520
8521
8522
8523
8524
8525
8526
8527
8528
8529
8530
8531
8532
8533
8534
8535
8536
8537
8538
8539
8540
8541
8542
8543
8544
8545
8546
8547
8548
8549
8550
8551
8552
8553
8554
8555
8556
8557
8558
8559
8560
8561
8562
8563
8564
8565
8566
8567
8568
8569
8570
8571
8572
8573
8574
8575
8576
8577
8578
8579
8580
8581
8582
8583
8584
8585
8586
8587
8588
8589
8590
8591
8592
8593
8594
8595
8596
8597
8598
8599
8600
8601
8602
8603
8604
8605
8606
8607
8608
8609
8610
8611
8612
8613
8614
8615
8616
8617
8618
8619
8620
8621
8622
8623
8624
8625
8626
8627
8628
8629
8630
8631
8632
8633
8634
8635
8636
8637
8638
8639
8640
8641
8642
8643
8644
8645
8646
8647
8648
8649
8650
8651
8652
8653
8654
8655
8656
8657
8658
8659
8660
8661
8662
8663
8664
8665
8666
8667
8668
8669
8670
8671
8672
8673
8674
8675
8676
8677
8678
8679
8680
8681
8682
8683
8684
8685
8686
8687
8688
8689
8690
8691
8692
8693
8694
8695
8696
8697
8698
8699
8700
8701
8702
8703
8704
8705
8706
8707
8708
8709
8710
8711
8712
8713
8714
8715
8716
8717
8718
8719
8720
8721
8722
8723
8724
8725
8726
8727
8728
8729
8730
8731
8732
8733
8734
8735
8736
8737
8738
8739
8740
8741
8742
8743
8744
8745
8746
8747
8748
8749
8750
8751
8752
8753
8754
8755
8756
8757
8758
8759
8760
8761
8762
8763
8764
8765
8766
8767
8768
8769
8770
8771
8772
8773
8774
8775
8776
8777
8778
8779
8780
8781
8782
8783
8784
8785
8786
8787
8788
8789
8790
8791
8792
8793
8794
8795
8796
8797
8798
8799
8800
8801
8802
8803
8804
8805
8806
8807
8808
8809
8810
8811
8812
8813
8814
8815
8816
8817
8818
8819
8820
8821
8822
8823
8824
8825
8826
8827
8828
8829
8830
8831
8832
8833
8834
8835
8836
8837
8838
8839
8840
8841
8842
8843
8844
8845
8846
8847
8848
8849
8850
8851
8852
8853
8854
8855
8856
8857
8858
8859
8860
8861
8862
8863
8864
8865
8866
8867
8868
8869
8870
8871
8872
8873
8874
8875
8876
8877
8878
8879
8880
8881
8882
8883
8884
8885
8886
8887
8888
8889
8890
8891
8892
8893
8894
8895
8896
8897
8898
8899
8900
8901
8902
8903
8904
8905
8906
8907
8908
8909
8910
8911
8912
8913
8914
8915
8916
8917
8918
8919
8920
8921
8922
8923
8924
8925
8926
8927
8928
8929
8930
8931
8932
8933
8934
8935
8936
8937
8938
8939
8940
8941
8942
8943
8944
8945
8946
8947
8948
8949
8950
8951
8952
8953
8954
8955
8956
8957
8958
8959
8960
8961
8962
8963
8964
8965
8966
8967
8968
8969
8970
8971
8972
8973
8974
8975
8976
8977
8978
8979
8980
8981
8982
8983
8984
8985
8986
8987
8988
8989
8990
8991
8992
8993
8994
8995
8996
8997
8998
8999
9000
9001
9002
9003
9004
9005
9006
9007
9008
9009
9010
9011
9012
9013
9014
9015
9016
9017
9018
9019
9020
9021
9022
9023
9024
9025
9026
9027
9028
9029
9030
9031
9032
9033
9034
9035
9036
9037
9038
9039
9040
9041
9042
9043
9044
9045
9046
9047
9048
9049
9050
9051
9052
9053
9054
9055
9056
9057
9058
9059
9060
9061
9062
9063
9064
9065
9066
9067
9068
9069
9070
9071
9072
9073
9074
9075
9076
9077
9078
9079
9080
9081
9082
9083
9084
9085
9086
9087
9088
9089
9090
9091
9092
9093
9094
9095
9096
9097
9098
9099
9100
9101
9102
9103
9104
9105
9106
9107
9108
9109
9110
9111
9112
9113
9114
9115
9116
9117
9118
9119
9120
9121
9122
9123
9124
9125
9126
9127
9128
9129
9130
9131
9132
9133
9134
9135
9136
9137
9138
9139
9140
9141
9142
9143
9144
9145
9146
9147
9148
9149
9150
9151
9152
9153
9154
9155
9156
9157
9158
9159
9160
9161
9162
9163
9164
9165
9166
9167
9168
9169
9170
9171
9172
9173
9174
9175
9176
9177
9178
9179
9180
9181
9182
9183
9184
9185
9186
9187
9188
9189
9190
9191
9192
9193
9194
9195
9196
9197
9198
9199
9200
9201
9202
9203
9204
9205
9206
9207
9208
9209
9210
9211
9212
9213
9214
9215
9216
9217
9218
9219
9220
9221
9222
9223
9224
9225
9226
9227
9228
9229
9230
9231
9232
9233
9234
9235
9236
9237
9238
9239
9240
9241
9242
9243
9244
9245
9246
9247
9248
9249
9250
9251
9252
9253
9254
9255
9256
9257
9258
9259
9260
9261
9262
9263
9264
9265
9266
9267
9268
9269
9270
9271
9272
9273
9274
9275
9276
9277
9278
9279
9280
9281
9282
9283
9284
9285
9286
9287
9288
9289
9290
9291
9292
9293
9294
9295
9296
9297
9298
9299
9300
9301
9302
9303
9304
9305
9306
9307
9308
9309
9310
9311
9312
9313
9314
9315
9316
9317
9318
9319
9320
9321
9322
9323
9324
9325
9326
9327
9328
9329
9330
9331
9332
9333
9334
9335
9336
9337
9338
9339
9340
9341
9342
9343
9344
9345
9346
9347
9348
9349
9350
9351
9352
9353
9354
9355
9356
9357
9358
9359
9360
9361
9362
9363
9364
9365
9366
9367
9368
9369
9370
9371
9372
9373
9374
9375
9376
9377
9378
9379
9380
9381
9382
9383
9384
9385
9386
9387
9388
9389
9390
9391
9392
9393
9394
9395
9396
9397
9398
9399
9400
9401
9402
9403
9404
9405
9406
9407
9408
9409
9410
9411
9412
9413
9414
9415
9416
9417
9418
9419
9420
9421
9422
9423
9424
9425
9426
9427
9428
9429
9430
9431
9432
9433
9434
9435
9436
9437
9438
9439
9440
9441
9442
9443
9444
9445
9446
9447
9448
9449
9450
9451
9452
9453
9454
9455
9456
9457
9458
9459
9460
9461
9462
9463
9464
9465
9466
9467
9468
9469
9470
9471
9472
9473
9474
9475
9476
9477
9478
9479
9480
9481
9482
9483
9484
9485
9486
9487
9488
9489
9490
9491
9492
9493
9494
9495
9496
9497
9498
9499
9500
9501
9502
9503
9504
9505
9506
9507
9508
9509
9510
9511
9512
9513
9514
9515
9516
9517
9518
9519
9520
9521
9522
9523
9524
9525
9526
9527
9528
9529
9530
9531
9532
9533
9534
9535
9536
9537
9538
9539
9540
9541
9542
9543
9544
9545
9546
9547
9548
9549
9550
9551
9552
9553
9554
9555
9556
9557
9558
9559
9560
9561
9562
9563
9564
9565
9566
9567
9568
9569
9570
9571
9572
9573
9574
9575
9576
9577
9578
9579
9580
9581
9582
9583
9584
9585
9586
9587
9588
9589
9590
9591
9592
9593
9594
9595
9596
9597
9598
9599
9600
9601
9602
9603
9604
9605
9606
9607
9608
9609
9610
9611
9612
9613
9614
9615
9616
9617
9618
9619
9620
9621
9622
9623
9624
9625
9626
9627
9628
9629
9630
9631
9632
9633
9634
9635
9636
9637
9638
9639
9640
9641
9642
9643
9644
9645
9646
9647
9648
9649
9650
9651
9652
9653
9654
9655
9656
9657
9658
9659
9660
9661
9662
9663
9664
9665
9666
9667
9668
9669
9670
9671
9672
9673
9674
9675
9676
9677
9678
9679
9680
9681
9682
9683
9684
9685
9686
9687
9688
9689
9690
9691
9692
9693
9694
9695
9696
9697
9698
9699
9700
9701
9702
9703
9704
9705
9706
9707
9708
9709
9710
9711
9712
9713
9714
9715
9716
9717
9718
9719
9720
9721
9722
9723
9724
9725
9726
9727
9728
9729
9730
9731
9732
9733
9734
9735
9736
9737
9738
9739
9740
9741
9742
9743
9744
9745
9746
9747
9748
9749
9750
9751
9752
9753
9754
9755
9756
9757
9758
9759
9760
9761
9762
9763
9764
9765
9766
9767
9768
9769
9770
9771
9772
9773
9774
9775
9776
9777
9778
9779
9780
9781
9782
9783
9784
9785
9786
9787
9788
9789
9790
9791
9792
9793
9794
9795
9796
9797
9798
9799
9800
9801
9802
9803
9804
9805
9806
9807
9808
9809
9810
9811
9812
9813
9814
9815
9816
9817
9818
9819
9820
9821
9822
9823
9824
9825
9826
9827
9828
9829
9830
9831
9832
9833
9834
9835
9836
9837
9838
9839
9840
9841
9842
9843
9844
9845
9846
9847
9848
9849
9850
9851
9852
9853
9854
9855
9856
9857
9858
9859
9860
9861
9862
9863
9864
9865
9866
9867
9868
9869
9870
9871
9872
9873
9874
9875
9876
9877
9878
9879
9880
9881
9882
9883
9884
9885
9886
9887
9888
9889
9890
9891
9892
9893
9894
9895
9896
9897
9898
9899
9900
9901
9902
9903
9904
9905
9906
9907
9908
9909
9910
9911
9912
9913
9914
9915
9916
9917
9918
9919
9920
9921
9922
9923
9924
9925
9926
9927
9928
9929
9930
9931
9932
9933
9934
9935
9936
9937
9938
9939
9940
9941
9942
9943
9944
9945
9946
9947
9948
9949
9950
9951
9952
9953
9954
9955
9956
9957
9958
9959
9960
9961
9962
9963
9964
9965
9966
9967
9968
9969
9970
9971
9972
9973
9974
9975
9976
9977
9978
9979
9980
9981
9982
9983
9984
9985
9986
9987
9988
9989
9990
9991
9992
9993
9994
9995
9996
9997
9998
9999
10000
10001
10002
10003
10004
10005
10006
10007
10008
10009
10010
10011
10012
10013
10014
10015
10016
10017
10018
10019
10020
10021
10022
10023
10024
10025
10026
10027
10028
10029
10030
10031
10032
10033
10034
10035
10036
10037
10038
10039
10040
10041
10042
10043
10044
10045
10046
10047
10048
10049
10050
10051
10052
10053
10054
10055
10056
10057
10058
10059
10060
10061
10062
10063
10064
10065
10066
10067
10068
10069
10070
10071
10072
10073
10074
10075
10076
10077
10078
10079
10080
10081
10082
10083
10084
10085
10086
10087
10088
10089
10090
10091
10092
10093
10094
10095
10096
10097
10098
10099
10100
10101
10102
10103
10104
10105
10106
10107
10108
10109
10110
10111
10112
10113
10114
10115
10116
10117
10118
10119
10120
10121
10122
10123
10124
10125
10126
10127
10128
10129
10130
10131
10132
10133
10134
10135
10136
10137
10138
10139
10140
10141
10142
10143
10144
10145
10146
10147
10148
10149
10150
10151
10152
10153
10154
10155
10156
10157
10158
10159
10160
10161
10162
10163
10164
10165
10166
10167
10168
10169
10170
10171
10172
10173
10174
10175
10176
10177
10178
10179
10180
10181
10182
10183
10184
10185
10186
10187
10188
10189
10190
10191
10192
10193
10194
10195
10196
10197
10198
10199
10200
10201
10202
10203
10204
10205
10206
10207
10208
10209
10210
10211
10212
10213
10214
10215
10216
10217
10218
10219
10220
10221
10222
10223
10224
10225
10226
10227
10228
10229
10230
10231
10232
10233
10234
10235
10236
10237
10238
10239
10240
10241
10242
10243
10244
10245
10246
10247
10248
10249
10250
10251
10252
10253
10254
10255
10256
10257
10258
10259
10260
10261
10262
10263
10264
10265
10266
10267
10268
10269
10270
10271
10272
10273
10274
10275
10276
10277
10278
10279
10280
10281
10282
10283
10284
10285
10286
10287
10288
10289
10290
10291
10292
10293
10294
10295
10296
10297
10298
10299
10300
10301
10302
10303
10304
10305
10306
10307
10308
10309
10310
10311
10312
10313
10314
10315
10316
10317
10318
10319
10320
10321
10322
10323
10324
10325
10326
10327
10328
10329
10330
10331
10332
10333
10334
10335
10336
10337
10338
10339
10340
10341
10342
10343
10344
10345
10346
10347
10348
10349
10350
10351
10352
10353
10354
10355
10356
10357
10358
10359
10360
10361
10362
10363
10364
10365
10366
10367
10368
10369
10370
10371
10372
10373
10374
10375
10376
10377
10378
10379
10380
10381
10382
10383
10384
10385
10386
10387
10388
10389
10390
10391
10392
10393
10394
10395
10396
10397
10398
10399
10400
10401
10402
10403
10404
10405
10406
10407
10408
10409
10410
10411
10412
10413
10414
10415
10416
10417
10418
10419
10420
10421
10422
10423
10424
10425
10426
10427
10428
10429
10430
10431
10432
10433
10434
10435
10436
10437
10438
10439
10440
10441
10442
10443
10444
10445
10446
10447
10448
10449
10450
10451
10452
10453
10454
10455
10456
10457
10458
10459
10460
10461
10462
10463
10464
10465
10466
10467
10468
10469
10470
10471
10472
10473
10474
10475
10476
10477
10478
10479
10480
10481
10482
10483
10484
10485
10486
10487
10488
10489
10490
10491
10492
10493
10494
10495
10496
10497
10498
10499
10500
10501
10502
10503
10504
10505
10506
10507
10508
10509
10510
10511
10512
10513
10514
10515
10516
10517
10518
10519
10520
10521
10522
10523
10524
10525
10526
10527
10528
10529
10530
10531
10532
10533
10534
10535
10536
10537
10538
10539
10540
10541
10542
10543
10544
10545
10546
10547
10548
10549
10550
10551
10552
10553
10554
10555
10556
10557
10558
10559
10560
10561
10562
10563
10564
10565
10566
10567
10568
10569
10570
10571
10572
10573
10574
10575
10576
10577
10578
10579
10580
10581
10582
10583
10584
10585
10586
10587
10588
10589
10590
10591
10592
10593
10594
10595
10596
10597
10598
10599
10600
10601
10602
10603
10604
10605
10606
10607
10608
10609
10610
10611
10612
10613
10614
10615
10616
10617
10618
10619
10620
10621
10622
10623
10624
10625
10626
10627
10628
10629
10630
10631
10632
10633
10634
10635
10636
10637
10638
10639
10640
10641
10642
10643
10644
10645
10646
10647
10648
10649
10650
10651
10652
10653
10654
10655
10656
10657
10658
10659
10660
10661
10662
10663
10664
10665
10666
10667
10668
10669
10670
10671
10672
10673
10674
10675
10676
10677
10678
10679
10680
10681
10682
10683
10684
10685
10686
10687
10688
10689
10690
10691
10692
10693
10694
10695
10696
10697
10698
10699
10700
10701
10702
10703
10704
10705
10706
10707
10708
10709
10710
10711
10712
10713
10714
10715
10716
10717
10718
10719
10720
10721
10722
10723
10724
10725
10726
10727
10728
10729
10730
10731
10732
10733
10734
10735
10736
10737
10738
10739
10740
10741
10742
10743
10744
10745
10746
10747
10748
10749
10750
10751
10752
10753
10754
10755
10756
10757
10758
10759
10760
10761
10762
10763
10764
10765
10766
10767
10768
10769
10770
10771
10772
10773
10774
10775
10776
10777
10778
10779
10780
10781
10782
10783
10784
10785
10786
10787
10788
10789
10790
10791
10792
10793
10794
10795
10796
10797
10798
10799
10800
10801
10802
10803
10804
10805
10806
10807
10808
10809
10810
10811
10812
10813
10814
10815
10816
10817
10818
10819
10820
10821
10822
10823
10824
10825
10826
10827
10828
10829
10830
10831
10832
10833
10834
10835
10836
10837
10838
10839
10840
10841
10842
10843
10844
10845
10846
10847
10848
10849
10850
10851
10852
10853
10854
10855
10856
10857
10858
10859
10860
10861
10862
10863
10864
10865
10866
10867
10868
10869
10870
10871
10872
10873
10874
10875
10876
10877
10878
10879
10880
10881
10882
10883
10884
10885
10886
10887
10888
10889
10890
10891
10892
10893
10894
10895
10896
10897
10898
10899
10900
10901
10902
10903
10904
10905
10906
10907
10908
10909
10910
10911
10912
10913
10914
10915
10916
10917
10918
10919
10920
10921
10922
10923
10924
10925
10926
10927
10928
10929
10930
10931
10932
10933
10934
10935
10936
10937
10938
10939
10940
10941
10942
10943
10944
10945
10946
10947
10948
10949
10950
10951
10952
10953
10954
10955
10956
10957
10958
10959
10960
10961
10962
10963
10964
10965
10966
10967
10968
10969
10970
10971
10972
10973
10974
10975
10976
10977
10978
10979
10980
10981
10982
10983
10984
10985
10986
10987
10988
10989
10990
10991
10992
10993
10994
10995
10996
10997
10998
10999
11000
11001
11002
11003
11004
11005
11006
11007
11008
11009
11010
11011
11012
11013
11014
11015
11016
11017
11018
11019
11020
11021
11022
11023
11024
11025
11026
11027
11028
11029
11030
11031
11032
11033
11034
11035
11036
11037
11038
11039
11040
11041
11042
11043
11044
11045
11046
11047
11048
11049
11050
11051
11052
11053
11054
11055
11056
11057
11058
11059
11060
11061
11062
11063
11064
11065
11066
11067
11068
11069
11070
11071
11072
11073
11074
11075
11076
11077
11078
11079
11080
11081
11082
11083
11084
11085
11086
11087
11088
11089
11090
11091
11092
11093
11094
11095
11096
11097
11098
11099
11100
11101
11102
11103
11104
11105
11106
11107
11108
11109
11110
11111
11112
11113
11114
11115
11116
11117
11118
11119
11120
11121
11122
11123
11124
11125
11126
11127
11128
11129
11130
11131
11132
11133
11134
11135
11136
11137
11138
11139
11140
11141
11142
11143
11144
11145
11146
11147
11148
11149
11150
11151
11152
11153
11154
11155
11156
11157
11158
11159
11160
11161
11162
11163
11164
11165
11166
11167
11168
11169
11170
11171
11172
11173
11174
11175
11176
11177
11178
11179
11180
11181
11182
11183
11184
11185
11186
11187
11188
11189
11190
11191
11192
11193
11194
11195
11196
11197
11198
11199
11200
11201
11202
11203
11204
11205
11206
11207
11208
11209
11210
11211
11212
11213
11214
11215
11216
11217
11218
11219
11220
11221
11222
11223
11224
11225
11226
11227
11228
11229
11230
11231
11232
11233
11234
11235
11236
11237
11238
11239
11240
11241
11242
11243
11244
11245
11246
11247
11248
11249
11250
11251
11252
11253
11254
11255
11256
11257
11258
11259
11260
11261
11262
11263
11264
11265
11266
11267
11268
11269
11270
11271
11272
11273
11274
11275
11276
11277
11278
11279
11280
11281
11282
11283
11284
11285
11286
11287
11288
11289
11290
11291
11292
11293
11294
11295
11296
11297
11298
11299
11300
11301
11302
11303
11304
11305
11306
11307
11308
11309
11310
11311
11312
11313
11314
11315
11316
11317
11318
11319
11320
11321
11322
11323
11324
11325
11326
11327
11328
11329
11330
11331
11332
11333
11334
11335
11336
11337
11338
11339
11340
11341
11342
11343
11344
11345
11346
11347
11348
11349
11350
11351
11352
11353
11354
11355
11356
11357
11358
11359
11360
11361
11362
11363
11364
11365
11366
11367
11368
11369
11370
11371
11372
11373
11374
11375
11376
11377
11378
11379
11380
11381
11382
11383
11384
11385
11386
11387
11388
11389
11390
11391
11392
11393
11394
11395
11396
11397
11398
11399
11400
11401
11402
11403
11404
11405
11406
11407
11408
11409
11410
11411
11412
11413
11414
11415
11416
11417
11418
11419
11420
11421
11422
11423
11424
11425
11426
11427
11428
11429
11430
11431
11432
11433
11434
11435
11436
11437
11438
11439
11440
11441
11442
11443
11444
11445
11446
11447
11448
11449
11450
11451
11452
11453
11454
11455
11456
11457
11458
11459
11460
11461
11462
11463
11464
11465
11466
11467
11468
11469
11470
11471
11472
11473
11474
11475
11476
11477
11478
11479
11480
11481
11482
11483
11484
11485
11486
11487
11488
11489
11490
11491
11492
11493
11494
11495
11496
11497
11498
11499
11500
11501
11502
11503
11504
11505
11506
11507
11508
11509
11510
11511
11512
11513
11514
11515
11516
11517
11518
11519
11520
11521
11522
11523
11524
11525
11526
11527
11528
11529
11530
11531
11532
11533
11534
11535
11536
11537
11538
11539
11540
11541
11542
11543
11544
11545
11546
11547
11548
11549
11550
11551
11552
11553
11554
11555
11556
11557
11558
11559
11560
11561
11562
11563
11564
11565
11566
11567
11568
11569
11570
11571
11572
11573
11574
11575
11576
11577
11578
11579
11580
11581
11582
11583
11584
11585
11586
11587
11588
11589
11590
11591
11592
11593
11594
11595
11596
11597
11598
11599
11600
11601
11602
11603
11604
11605
11606
11607
11608
11609
11610
11611
11612
11613
11614
11615
11616
11617
11618
11619
11620
11621
11622
11623
11624
11625
11626
11627
11628
11629
11630
11631
11632
11633
11634
11635
11636
11637
11638
11639
11640
11641
11642
11643
11644
11645
11646
11647
11648
11649
11650
11651
11652
11653
11654
11655
11656
11657
11658
11659
11660
11661
11662
11663
11664
11665
11666
11667
11668
11669
11670
11671
11672
11673
11674
11675
11676
11677
11678
11679
11680
11681
11682
11683
11684
11685
11686
11687
11688
11689
11690
11691
11692
11693
11694
11695
11696
11697
11698
11699
11700
11701
11702
11703
11704
11705
11706
11707
11708
11709
11710
11711
11712
11713
11714
11715
11716
11717
11718
11719
11720
11721
11722
11723
11724
11725
11726
11727
11728
11729
11730
11731
11732
11733
11734
11735
11736
11737
11738
11739
11740
11741
11742
11743
11744
11745
11746
11747
11748
11749
11750
11751
11752
11753
11754
11755
11756
11757
/* Definitions of target machine for GNU compiler.
   Copyright (C) 1999-2014 Free Software Foundation, Inc.
   Contributed by James E. Wilson <wilson@cygnus.com> and
		  David Mosberger <davidm@hpl.hp.com>.

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 3, 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 COPYING3.  If not see
<http://www.gnu.org/licenses/>.  */

#include "config.h"
#include "system.h"
#include "coretypes.h"
#include "tm.h"
#include "rtl.h"
#include "tree.h"
#include "stringpool.h"
#include "stor-layout.h"
#include "calls.h"
#include "varasm.h"
#include "regs.h"
#include "hard-reg-set.h"
#include "insn-config.h"
#include "conditions.h"
#include "output.h"
#include "insn-attr.h"
#include "flags.h"
#include "recog.h"
#include "expr.h"
#include "optabs.h"
#include "except.h"
#include "function.h"
#include "ggc.h"
#include "basic-block.h"
#include "libfuncs.h"
#include "diagnostic-core.h"
#include "sched-int.h"
#include "timevar.h"
#include "target.h"
#include "target-def.h"
#include "common/common-target.h"
#include "tm_p.h"
#include "hash-table.h"
#include "langhooks.h"
#include "pointer-set.h"
#include "vec.h"
#include "basic-block.h"
#include "tree-ssa-alias.h"
#include "internal-fn.h"
#include "gimple-fold.h"
#include "tree-eh.h"
#include "gimple-expr.h"
#include "is-a.h"
#include "gimple.h"
#include "gimplify.h"
#include "intl.h"
#include "df.h"
#include "debug.h"
#include "params.h"
#include "dbgcnt.h"
#include "tm-constrs.h"
#include "sel-sched.h"
#include "reload.h"
#include "opts.h"
#include "dumpfile.h"

/* This is used for communication between ASM_OUTPUT_LABEL and
   ASM_OUTPUT_LABELREF.  */
int ia64_asm_output_label = 0;

/* Register names for ia64_expand_prologue.  */
static const char * const ia64_reg_numbers[96] =
{ "r32", "r33", "r34", "r35", "r36", "r37", "r38", "r39",
  "r40", "r41", "r42", "r43", "r44", "r45", "r46", "r47",
  "r48", "r49", "r50", "r51", "r52", "r53", "r54", "r55",
  "r56", "r57", "r58", "r59", "r60", "r61", "r62", "r63",
  "r64", "r65", "r66", "r67", "r68", "r69", "r70", "r71",
  "r72", "r73", "r74", "r75", "r76", "r77", "r78", "r79",
  "r80", "r81", "r82", "r83", "r84", "r85", "r86", "r87",
  "r88", "r89", "r90", "r91", "r92", "r93", "r94", "r95",
  "r96", "r97", "r98", "r99", "r100","r101","r102","r103",
  "r104","r105","r106","r107","r108","r109","r110","r111",
  "r112","r113","r114","r115","r116","r117","r118","r119",
  "r120","r121","r122","r123","r124","r125","r126","r127"};

/* ??? These strings could be shared with REGISTER_NAMES.  */
static const char * const ia64_input_reg_names[8] =
{ "in0",  "in1",  "in2",  "in3",  "in4",  "in5",  "in6",  "in7" };

/* ??? These strings could be shared with REGISTER_NAMES.  */
static const char * const ia64_local_reg_names[80] =
{ "loc0", "loc1", "loc2", "loc3", "loc4", "loc5", "loc6", "loc7",
  "loc8", "loc9", "loc10","loc11","loc12","loc13","loc14","loc15",
  "loc16","loc17","loc18","loc19","loc20","loc21","loc22","loc23",
  "loc24","loc25","loc26","loc27","loc28","loc29","loc30","loc31",
  "loc32","loc33","loc34","loc35","loc36","loc37","loc38","loc39",
  "loc40","loc41","loc42","loc43","loc44","loc45","loc46","loc47",
  "loc48","loc49","loc50","loc51","loc52","loc53","loc54","loc55",
  "loc56","loc57","loc58","loc59","loc60","loc61","loc62","loc63",
  "loc64","loc65","loc66","loc67","loc68","loc69","loc70","loc71",
  "loc72","loc73","loc74","loc75","loc76","loc77","loc78","loc79" };

/* ??? These strings could be shared with REGISTER_NAMES.  */
static const char * const ia64_output_reg_names[8] =
{ "out0", "out1", "out2", "out3", "out4", "out5", "out6", "out7" };

/* Variables which are this size or smaller are put in the sdata/sbss
   sections.  */

unsigned int ia64_section_threshold;

/* The following variable is used by the DFA insn scheduler.  The value is
   TRUE if we do insn bundling instead of insn scheduling.  */
int bundling_p = 0;

enum ia64_frame_regs
{
   reg_fp,
   reg_save_b0,
   reg_save_pr,
   reg_save_ar_pfs,
   reg_save_ar_unat,
   reg_save_ar_lc,
   reg_save_gp,
   number_of_ia64_frame_regs
};

/* Structure to be filled in by ia64_compute_frame_size with register
   save masks and offsets for the current function.  */

struct ia64_frame_info
{
  HOST_WIDE_INT total_size;	/* size of the stack frame, not including
				   the caller's scratch area.  */
  HOST_WIDE_INT spill_cfa_off;	/* top of the reg spill area from the cfa.  */
  HOST_WIDE_INT spill_size;	/* size of the gr/br/fr spill area.  */
  HOST_WIDE_INT extra_spill_size;  /* size of spill area for others.  */
  HARD_REG_SET mask;		/* mask of saved registers.  */
  unsigned int gr_used_mask;	/* mask of registers in use as gr spill
				   registers or long-term scratches.  */
  int n_spilled;		/* number of spilled registers.  */
  int r[number_of_ia64_frame_regs];  /* Frame related registers.  */
  int n_input_regs;		/* number of input registers used.  */
  int n_local_regs;		/* number of local registers used.  */
  int n_output_regs;		/* number of output registers used.  */
  int n_rotate_regs;		/* number of rotating registers used.  */

  char need_regstk;		/* true if a .regstk directive needed.  */
  char initialized;		/* true if the data is finalized.  */
};

/* Current frame information calculated by ia64_compute_frame_size.  */
static struct ia64_frame_info current_frame_info;
/* The actual registers that are emitted.  */
static int emitted_frame_related_regs[number_of_ia64_frame_regs];

static int ia64_first_cycle_multipass_dfa_lookahead (void);
static void ia64_dependencies_evaluation_hook (rtx, rtx);
static void ia64_init_dfa_pre_cycle_insn (void);
static rtx ia64_dfa_pre_cycle_insn (void);
static int ia64_first_cycle_multipass_dfa_lookahead_guard (rtx);
static bool ia64_first_cycle_multipass_dfa_lookahead_guard_spec (const_rtx);
static int ia64_dfa_new_cycle (FILE *, int, rtx, int, int, int *);
static void ia64_h_i_d_extended (void);
static void * ia64_alloc_sched_context (void);
static void ia64_init_sched_context (void *, bool);
static void ia64_set_sched_context (void *);
static void ia64_clear_sched_context (void *);
static void ia64_free_sched_context (void *);
static int ia64_mode_to_int (enum machine_mode);
static void ia64_set_sched_flags (spec_info_t);
static ds_t ia64_get_insn_spec_ds (rtx);
static ds_t ia64_get_insn_checked_ds (rtx);
static bool ia64_skip_rtx_p (const_rtx);
static int ia64_speculate_insn (rtx, ds_t, rtx *);
static bool ia64_needs_block_p (ds_t);
static rtx ia64_gen_spec_check (rtx, rtx, ds_t);
static int ia64_spec_check_p (rtx);
static int ia64_spec_check_src_p (rtx);
static rtx gen_tls_get_addr (void);
static rtx gen_thread_pointer (void);
static int find_gr_spill (enum ia64_frame_regs, int);
static int next_scratch_gr_reg (void);
static void mark_reg_gr_used_mask (rtx, void *);
static void ia64_compute_frame_size (HOST_WIDE_INT);
static void setup_spill_pointers (int, rtx, HOST_WIDE_INT);
static void finish_spill_pointers (void);
static rtx spill_restore_mem (rtx, HOST_WIDE_INT);
static void do_spill (rtx (*)(rtx, rtx, rtx), rtx, HOST_WIDE_INT, rtx);
static void do_restore (rtx (*)(rtx, rtx, rtx), rtx, HOST_WIDE_INT);
static rtx gen_movdi_x (rtx, rtx, rtx);
static rtx gen_fr_spill_x (rtx, rtx, rtx);
static rtx gen_fr_restore_x (rtx, rtx, rtx);

static void ia64_option_override (void);
static bool ia64_can_eliminate (const int, const int);
static enum machine_mode hfa_element_mode (const_tree, bool);
static void ia64_setup_incoming_varargs (cumulative_args_t, enum machine_mode,
					 tree, int *, int);
static int ia64_arg_partial_bytes (cumulative_args_t, enum machine_mode,
				   tree, bool);
static rtx ia64_function_arg_1 (cumulative_args_t, enum machine_mode,
				const_tree, bool, bool);
static rtx ia64_function_arg (cumulative_args_t, enum machine_mode,
			      const_tree, bool);
static rtx ia64_function_incoming_arg (cumulative_args_t,
				       enum machine_mode, const_tree, bool);
static void ia64_function_arg_advance (cumulative_args_t, enum machine_mode,
				       const_tree, bool);
static unsigned int ia64_function_arg_boundary (enum machine_mode,
						const_tree);
static bool ia64_function_ok_for_sibcall (tree, tree);
static bool ia64_return_in_memory (const_tree, const_tree);
static rtx ia64_function_value (const_tree, const_tree, bool);
static rtx ia64_libcall_value (enum machine_mode, const_rtx);
static bool ia64_function_value_regno_p (const unsigned int);
static int ia64_register_move_cost (enum machine_mode, reg_class_t,
                                    reg_class_t);
static int ia64_memory_move_cost (enum machine_mode mode, reg_class_t,
				  bool);
static bool ia64_rtx_costs (rtx, int, int, int, int *, bool);
static int ia64_unspec_may_trap_p (const_rtx, unsigned);
static void fix_range (const char *);
static struct machine_function * ia64_init_machine_status (void);
static void emit_insn_group_barriers (FILE *);
static void emit_all_insn_group_barriers (FILE *);
static void final_emit_insn_group_barriers (FILE *);
static void emit_predicate_relation_info (void);
static void ia64_reorg (void);
static bool ia64_in_small_data_p (const_tree);
static void process_epilogue (FILE *, rtx, bool, bool);

static bool ia64_assemble_integer (rtx, unsigned int, int);
static void ia64_output_function_prologue (FILE *, HOST_WIDE_INT);
static void ia64_output_function_epilogue (FILE *, HOST_WIDE_INT);
static void ia64_output_function_end_prologue (FILE *);

static void ia64_print_operand (FILE *, rtx, int);
static void ia64_print_operand_address (FILE *, rtx);
static bool ia64_print_operand_punct_valid_p (unsigned char code);

static int ia64_issue_rate (void);
static int ia64_adjust_cost_2 (rtx, int, rtx, int, dw_t);
static void ia64_sched_init (FILE *, int, int);
static void ia64_sched_init_global (FILE *, int, int);
static void ia64_sched_finish_global (FILE *, int);
static void ia64_sched_finish (FILE *, int);
static int ia64_dfa_sched_reorder (FILE *, int, rtx *, int *, int, int);
static int ia64_sched_reorder (FILE *, int, rtx *, int *, int);
static int ia64_sched_reorder2 (FILE *, int, rtx *, int *, int);
static int ia64_variable_issue (FILE *, int, rtx, int);

static void ia64_asm_unwind_emit (FILE *, rtx);
static void ia64_asm_emit_except_personality (rtx);
static void ia64_asm_init_sections (void);

static enum unwind_info_type ia64_debug_unwind_info (void);

static struct bundle_state *get_free_bundle_state (void);
static void free_bundle_state (struct bundle_state *);
static void initiate_bundle_states (void);
static void finish_bundle_states (void);
static int insert_bundle_state (struct bundle_state *);
static void initiate_bundle_state_table (void);
static void finish_bundle_state_table (void);
static int try_issue_nops (struct bundle_state *, int);
static int try_issue_insn (struct bundle_state *, rtx);
static void issue_nops_and_insn (struct bundle_state *, int, rtx, int, int);
static int get_max_pos (state_t);
static int get_template (state_t, int);

static rtx get_next_important_insn (rtx, rtx);
static bool important_for_bundling_p (rtx);
static bool unknown_for_bundling_p (rtx);
static void bundling (FILE *, int, rtx, rtx);

static void ia64_output_mi_thunk (FILE *, tree, HOST_WIDE_INT,
				  HOST_WIDE_INT, tree);
static void ia64_file_start (void);
static void ia64_globalize_decl_name (FILE *, tree);

static int ia64_hpux_reloc_rw_mask (void) ATTRIBUTE_UNUSED;
static int ia64_reloc_rw_mask (void) ATTRIBUTE_UNUSED;
static section *ia64_select_rtx_section (enum machine_mode, rtx,
					 unsigned HOST_WIDE_INT);
static void ia64_output_dwarf_dtprel (FILE *, int, rtx)
     ATTRIBUTE_UNUSED;
static unsigned int ia64_section_type_flags (tree, const char *, int);
static void ia64_init_libfuncs (void)
     ATTRIBUTE_UNUSED;
static void ia64_hpux_init_libfuncs (void)
     ATTRIBUTE_UNUSED;
static void ia64_sysv4_init_libfuncs (void)
     ATTRIBUTE_UNUSED;
static void ia64_vms_init_libfuncs (void)
     ATTRIBUTE_UNUSED;
static void ia64_soft_fp_init_libfuncs (void)
     ATTRIBUTE_UNUSED;
static bool ia64_vms_valid_pointer_mode (enum machine_mode mode)
     ATTRIBUTE_UNUSED;
static tree ia64_vms_common_object_attribute (tree *, tree, tree, int, bool *)
     ATTRIBUTE_UNUSED;

static tree ia64_handle_model_attribute (tree *, tree, tree, int, bool *);
static tree ia64_handle_version_id_attribute (tree *, tree, tree, int, bool *);
static void ia64_encode_section_info (tree, rtx, int);
static rtx ia64_struct_value_rtx (tree, int);
static tree ia64_gimplify_va_arg (tree, tree, gimple_seq *, gimple_seq *);
static bool ia64_scalar_mode_supported_p (enum machine_mode mode);
static bool ia64_vector_mode_supported_p (enum machine_mode mode);
static bool ia64_legitimate_constant_p (enum machine_mode, rtx);
static bool ia64_legitimate_address_p (enum machine_mode, rtx, bool);
static bool ia64_cannot_force_const_mem (enum machine_mode, rtx);
static const char *ia64_mangle_type (const_tree);
static const char *ia64_invalid_conversion (const_tree, const_tree);
static const char *ia64_invalid_unary_op (int, const_tree);
static const char *ia64_invalid_binary_op (int, const_tree, const_tree);
static enum machine_mode ia64_c_mode_for_suffix (char);
static void ia64_trampoline_init (rtx, tree, rtx);
static void ia64_override_options_after_change (void);
static bool ia64_member_type_forces_blk (const_tree, enum machine_mode);

static tree ia64_builtin_decl (unsigned, bool);

static reg_class_t ia64_preferred_reload_class (rtx, reg_class_t);
static enum machine_mode ia64_get_reg_raw_mode (int regno);
static section * ia64_hpux_function_section (tree, enum node_frequency,
					     bool, bool);

static bool ia64_vectorize_vec_perm_const_ok (enum machine_mode vmode,
					      const unsigned char *sel);

#define MAX_VECT_LEN	8

struct expand_vec_perm_d
{
  rtx target, op0, op1;
  unsigned char perm[MAX_VECT_LEN];
  enum machine_mode vmode;
  unsigned char nelt;
  bool one_operand_p;
  bool testing_p; 
};

static bool ia64_expand_vec_perm_const_1 (struct expand_vec_perm_d *d);


/* Table of valid machine attributes.  */
static const struct attribute_spec ia64_attribute_table[] =
{
  /* { name, min_len, max_len, decl_req, type_req, fn_type_req, handler,
       affects_type_identity } */
  { "syscall_linkage", 0, 0, false, true,  true,  NULL, false },
  { "model",	       1, 1, true, false, false, ia64_handle_model_attribute,
    false },
#if TARGET_ABI_OPEN_VMS
  { "common_object",   1, 1, true, false, false,
    ia64_vms_common_object_attribute, false },
#endif
  { "version_id",      1, 1, true, false, false,
    ia64_handle_version_id_attribute, false },
  { NULL,	       0, 0, false, false, false, NULL, false }
};

/* Initialize the GCC target structure.  */
#undef TARGET_ATTRIBUTE_TABLE
#define TARGET_ATTRIBUTE_TABLE ia64_attribute_table

#undef TARGET_INIT_BUILTINS
#define TARGET_INIT_BUILTINS ia64_init_builtins

#undef TARGET_EXPAND_BUILTIN
#define TARGET_EXPAND_BUILTIN ia64_expand_builtin

#undef TARGET_BUILTIN_DECL
#define TARGET_BUILTIN_DECL ia64_builtin_decl

#undef TARGET_ASM_BYTE_OP
#define TARGET_ASM_BYTE_OP "\tdata1\t"
#undef TARGET_ASM_ALIGNED_HI_OP
#define TARGET_ASM_ALIGNED_HI_OP "\tdata2\t"
#undef TARGET_ASM_ALIGNED_SI_OP
#define TARGET_ASM_ALIGNED_SI_OP "\tdata4\t"
#undef TARGET_ASM_ALIGNED_DI_OP
#define TARGET_ASM_ALIGNED_DI_OP "\tdata8\t"
#undef TARGET_ASM_UNALIGNED_HI_OP
#define TARGET_ASM_UNALIGNED_HI_OP "\tdata2.ua\t"
#undef TARGET_ASM_UNALIGNED_SI_OP
#define TARGET_ASM_UNALIGNED_SI_OP "\tdata4.ua\t"
#undef TARGET_ASM_UNALIGNED_DI_OP
#define TARGET_ASM_UNALIGNED_DI_OP "\tdata8.ua\t"
#undef TARGET_ASM_INTEGER
#define TARGET_ASM_INTEGER ia64_assemble_integer

#undef TARGET_OPTION_OVERRIDE
#define TARGET_OPTION_OVERRIDE ia64_option_override

#undef TARGET_ASM_FUNCTION_PROLOGUE
#define TARGET_ASM_FUNCTION_PROLOGUE ia64_output_function_prologue
#undef TARGET_ASM_FUNCTION_END_PROLOGUE
#define TARGET_ASM_FUNCTION_END_PROLOGUE ia64_output_function_end_prologue
#undef TARGET_ASM_FUNCTION_EPILOGUE
#define TARGET_ASM_FUNCTION_EPILOGUE ia64_output_function_epilogue

#undef TARGET_PRINT_OPERAND
#define TARGET_PRINT_OPERAND ia64_print_operand
#undef TARGET_PRINT_OPERAND_ADDRESS
#define TARGET_PRINT_OPERAND_ADDRESS ia64_print_operand_address
#undef TARGET_PRINT_OPERAND_PUNCT_VALID_P
#define TARGET_PRINT_OPERAND_PUNCT_VALID_P ia64_print_operand_punct_valid_p

#undef TARGET_IN_SMALL_DATA_P
#define TARGET_IN_SMALL_DATA_P  ia64_in_small_data_p

#undef TARGET_SCHED_ADJUST_COST_2
#define TARGET_SCHED_ADJUST_COST_2 ia64_adjust_cost_2
#undef TARGET_SCHED_ISSUE_RATE
#define TARGET_SCHED_ISSUE_RATE ia64_issue_rate
#undef TARGET_SCHED_VARIABLE_ISSUE
#define TARGET_SCHED_VARIABLE_ISSUE ia64_variable_issue
#undef TARGET_SCHED_INIT
#define TARGET_SCHED_INIT ia64_sched_init
#undef TARGET_SCHED_FINISH
#define TARGET_SCHED_FINISH ia64_sched_finish
#undef TARGET_SCHED_INIT_GLOBAL
#define TARGET_SCHED_INIT_GLOBAL ia64_sched_init_global
#undef TARGET_SCHED_FINISH_GLOBAL
#define TARGET_SCHED_FINISH_GLOBAL ia64_sched_finish_global
#undef TARGET_SCHED_REORDER
#define TARGET_SCHED_REORDER ia64_sched_reorder
#undef TARGET_SCHED_REORDER2
#define TARGET_SCHED_REORDER2 ia64_sched_reorder2

#undef TARGET_SCHED_DEPENDENCIES_EVALUATION_HOOK
#define TARGET_SCHED_DEPENDENCIES_EVALUATION_HOOK ia64_dependencies_evaluation_hook

#undef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD
#define TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD ia64_first_cycle_multipass_dfa_lookahead

#undef TARGET_SCHED_INIT_DFA_PRE_CYCLE_INSN
#define TARGET_SCHED_INIT_DFA_PRE_CYCLE_INSN ia64_init_dfa_pre_cycle_insn
#undef TARGET_SCHED_DFA_PRE_CYCLE_INSN
#define TARGET_SCHED_DFA_PRE_CYCLE_INSN ia64_dfa_pre_cycle_insn

#undef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD_GUARD
#define TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD_GUARD\
  ia64_first_cycle_multipass_dfa_lookahead_guard

#undef TARGET_SCHED_DFA_NEW_CYCLE
#define TARGET_SCHED_DFA_NEW_CYCLE ia64_dfa_new_cycle

#undef TARGET_SCHED_H_I_D_EXTENDED
#define TARGET_SCHED_H_I_D_EXTENDED ia64_h_i_d_extended

#undef TARGET_SCHED_ALLOC_SCHED_CONTEXT
#define TARGET_SCHED_ALLOC_SCHED_CONTEXT ia64_alloc_sched_context

#undef TARGET_SCHED_INIT_SCHED_CONTEXT
#define TARGET_SCHED_INIT_SCHED_CONTEXT ia64_init_sched_context

#undef TARGET_SCHED_SET_SCHED_CONTEXT
#define TARGET_SCHED_SET_SCHED_CONTEXT ia64_set_sched_context

#undef TARGET_SCHED_CLEAR_SCHED_CONTEXT
#define TARGET_SCHED_CLEAR_SCHED_CONTEXT ia64_clear_sched_context

#undef TARGET_SCHED_FREE_SCHED_CONTEXT
#define TARGET_SCHED_FREE_SCHED_CONTEXT ia64_free_sched_context

#undef TARGET_SCHED_SET_SCHED_FLAGS
#define TARGET_SCHED_SET_SCHED_FLAGS ia64_set_sched_flags

#undef TARGET_SCHED_GET_INSN_SPEC_DS
#define TARGET_SCHED_GET_INSN_SPEC_DS ia64_get_insn_spec_ds

#undef TARGET_SCHED_GET_INSN_CHECKED_DS
#define TARGET_SCHED_GET_INSN_CHECKED_DS ia64_get_insn_checked_ds

#undef TARGET_SCHED_SPECULATE_INSN
#define TARGET_SCHED_SPECULATE_INSN ia64_speculate_insn

#undef TARGET_SCHED_NEEDS_BLOCK_P
#define TARGET_SCHED_NEEDS_BLOCK_P ia64_needs_block_p

#undef TARGET_SCHED_GEN_SPEC_CHECK
#define TARGET_SCHED_GEN_SPEC_CHECK ia64_gen_spec_check

#undef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD_GUARD_SPEC
#define TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD_GUARD_SPEC\
  ia64_first_cycle_multipass_dfa_lookahead_guard_spec

#undef TARGET_SCHED_SKIP_RTX_P
#define TARGET_SCHED_SKIP_RTX_P ia64_skip_rtx_p

#undef TARGET_FUNCTION_OK_FOR_SIBCALL
#define TARGET_FUNCTION_OK_FOR_SIBCALL ia64_function_ok_for_sibcall
#undef TARGET_ARG_PARTIAL_BYTES
#define TARGET_ARG_PARTIAL_BYTES ia64_arg_partial_bytes
#undef TARGET_FUNCTION_ARG
#define TARGET_FUNCTION_ARG ia64_function_arg
#undef TARGET_FUNCTION_INCOMING_ARG
#define TARGET_FUNCTION_INCOMING_ARG ia64_function_incoming_arg
#undef TARGET_FUNCTION_ARG_ADVANCE
#define TARGET_FUNCTION_ARG_ADVANCE ia64_function_arg_advance
#undef TARGET_FUNCTION_ARG_BOUNDARY
#define TARGET_FUNCTION_ARG_BOUNDARY ia64_function_arg_boundary

#undef TARGET_ASM_OUTPUT_MI_THUNK
#define TARGET_ASM_OUTPUT_MI_THUNK ia64_output_mi_thunk
#undef TARGET_ASM_CAN_OUTPUT_MI_THUNK
#define TARGET_ASM_CAN_OUTPUT_MI_THUNK hook_bool_const_tree_hwi_hwi_const_tree_true

#undef TARGET_ASM_FILE_START
#define TARGET_ASM_FILE_START ia64_file_start

#undef TARGET_ASM_GLOBALIZE_DECL_NAME
#define TARGET_ASM_GLOBALIZE_DECL_NAME ia64_globalize_decl_name

#undef TARGET_REGISTER_MOVE_COST
#define TARGET_REGISTER_MOVE_COST ia64_register_move_cost
#undef TARGET_MEMORY_MOVE_COST
#define TARGET_MEMORY_MOVE_COST ia64_memory_move_cost
#undef TARGET_RTX_COSTS
#define TARGET_RTX_COSTS ia64_rtx_costs
#undef TARGET_ADDRESS_COST
#define TARGET_ADDRESS_COST hook_int_rtx_mode_as_bool_0

#undef TARGET_UNSPEC_MAY_TRAP_P
#define TARGET_UNSPEC_MAY_TRAP_P ia64_unspec_may_trap_p

#undef TARGET_MACHINE_DEPENDENT_REORG
#define TARGET_MACHINE_DEPENDENT_REORG ia64_reorg

#undef TARGET_ENCODE_SECTION_INFO
#define TARGET_ENCODE_SECTION_INFO ia64_encode_section_info

#undef  TARGET_SECTION_TYPE_FLAGS
#define TARGET_SECTION_TYPE_FLAGS  ia64_section_type_flags

#ifdef HAVE_AS_TLS
#undef TARGET_ASM_OUTPUT_DWARF_DTPREL
#define TARGET_ASM_OUTPUT_DWARF_DTPREL ia64_output_dwarf_dtprel
#endif

/* ??? Investigate.  */
#if 0
#undef TARGET_PROMOTE_PROTOTYPES
#define TARGET_PROMOTE_PROTOTYPES hook_bool_tree_true
#endif

#undef TARGET_FUNCTION_VALUE
#define TARGET_FUNCTION_VALUE ia64_function_value
#undef TARGET_LIBCALL_VALUE
#define TARGET_LIBCALL_VALUE ia64_libcall_value
#undef TARGET_FUNCTION_VALUE_REGNO_P
#define TARGET_FUNCTION_VALUE_REGNO_P ia64_function_value_regno_p

#undef TARGET_STRUCT_VALUE_RTX
#define TARGET_STRUCT_VALUE_RTX ia64_struct_value_rtx
#undef TARGET_RETURN_IN_MEMORY
#define TARGET_RETURN_IN_MEMORY ia64_return_in_memory
#undef TARGET_SETUP_INCOMING_VARARGS
#define TARGET_SETUP_INCOMING_VARARGS ia64_setup_incoming_varargs
#undef TARGET_STRICT_ARGUMENT_NAMING
#define TARGET_STRICT_ARGUMENT_NAMING hook_bool_CUMULATIVE_ARGS_true
#undef TARGET_MUST_PASS_IN_STACK
#define TARGET_MUST_PASS_IN_STACK must_pass_in_stack_var_size
#undef TARGET_GET_RAW_RESULT_MODE
#define TARGET_GET_RAW_RESULT_MODE ia64_get_reg_raw_mode
#undef TARGET_GET_RAW_ARG_MODE
#define TARGET_GET_RAW_ARG_MODE ia64_get_reg_raw_mode

#undef TARGET_MEMBER_TYPE_FORCES_BLK
#define TARGET_MEMBER_TYPE_FORCES_BLK ia64_member_type_forces_blk

#undef TARGET_GIMPLIFY_VA_ARG_EXPR
#define TARGET_GIMPLIFY_VA_ARG_EXPR ia64_gimplify_va_arg

#undef TARGET_ASM_UNWIND_EMIT
#define TARGET_ASM_UNWIND_EMIT ia64_asm_unwind_emit
#undef TARGET_ASM_EMIT_EXCEPT_PERSONALITY
#define TARGET_ASM_EMIT_EXCEPT_PERSONALITY  ia64_asm_emit_except_personality
#undef TARGET_ASM_INIT_SECTIONS
#define TARGET_ASM_INIT_SECTIONS  ia64_asm_init_sections

#undef TARGET_DEBUG_UNWIND_INFO
#define TARGET_DEBUG_UNWIND_INFO  ia64_debug_unwind_info

#undef TARGET_SCALAR_MODE_SUPPORTED_P
#define TARGET_SCALAR_MODE_SUPPORTED_P ia64_scalar_mode_supported_p
#undef TARGET_VECTOR_MODE_SUPPORTED_P
#define TARGET_VECTOR_MODE_SUPPORTED_P ia64_vector_mode_supported_p

#undef TARGET_LEGITIMATE_CONSTANT_P
#define TARGET_LEGITIMATE_CONSTANT_P ia64_legitimate_constant_p
#undef TARGET_LEGITIMATE_ADDRESS_P
#define TARGET_LEGITIMATE_ADDRESS_P ia64_legitimate_address_p

#undef TARGET_CANNOT_FORCE_CONST_MEM
#define TARGET_CANNOT_FORCE_CONST_MEM ia64_cannot_force_const_mem

#undef TARGET_MANGLE_TYPE
#define TARGET_MANGLE_TYPE ia64_mangle_type

#undef TARGET_INVALID_CONVERSION
#define TARGET_INVALID_CONVERSION ia64_invalid_conversion
#undef TARGET_INVALID_UNARY_OP
#define TARGET_INVALID_UNARY_OP ia64_invalid_unary_op
#undef TARGET_INVALID_BINARY_OP
#define TARGET_INVALID_BINARY_OP ia64_invalid_binary_op

#undef TARGET_C_MODE_FOR_SUFFIX
#define TARGET_C_MODE_FOR_SUFFIX ia64_c_mode_for_suffix

#undef TARGET_CAN_ELIMINATE
#define TARGET_CAN_ELIMINATE ia64_can_eliminate

#undef TARGET_TRAMPOLINE_INIT
#define TARGET_TRAMPOLINE_INIT ia64_trampoline_init

#undef TARGET_CAN_USE_DOLOOP_P
#define TARGET_CAN_USE_DOLOOP_P can_use_doloop_if_innermost
#undef TARGET_INVALID_WITHIN_DOLOOP
#define TARGET_INVALID_WITHIN_DOLOOP hook_constcharptr_const_rtx_null

#undef TARGET_OVERRIDE_OPTIONS_AFTER_CHANGE
#define TARGET_OVERRIDE_OPTIONS_AFTER_CHANGE ia64_override_options_after_change

#undef TARGET_PREFERRED_RELOAD_CLASS
#define TARGET_PREFERRED_RELOAD_CLASS ia64_preferred_reload_class

#undef TARGET_DELAY_SCHED2
#define TARGET_DELAY_SCHED2 true

/* Variable tracking should be run after all optimizations which
   change order of insns.  It also needs a valid CFG.  */
#undef TARGET_DELAY_VARTRACK
#define TARGET_DELAY_VARTRACK true

#undef TARGET_VECTORIZE_VEC_PERM_CONST_OK
#define TARGET_VECTORIZE_VEC_PERM_CONST_OK ia64_vectorize_vec_perm_const_ok

struct gcc_target targetm = TARGET_INITIALIZER;

typedef enum
  {
    ADDR_AREA_NORMAL,	/* normal address area */
    ADDR_AREA_SMALL	/* addressable by "addl" (-2MB < addr < 2MB) */
  }
ia64_addr_area;

static GTY(()) tree small_ident1;
static GTY(()) tree small_ident2;

static void
init_idents (void)
{
  if (small_ident1 == 0)
    {
      small_ident1 = get_identifier ("small");
      small_ident2 = get_identifier ("__small__");
    }
}

/* Retrieve the address area that has been chosen for the given decl.  */

static ia64_addr_area
ia64_get_addr_area (tree decl)
{
  tree model_attr;

  model_attr = lookup_attribute ("model", DECL_ATTRIBUTES (decl));
  if (model_attr)
    {
      tree id;

      init_idents ();
      id = TREE_VALUE (TREE_VALUE (model_attr));
      if (id == small_ident1 || id == small_ident2)
	return ADDR_AREA_SMALL;
    }
  return ADDR_AREA_NORMAL;
}

static tree
ia64_handle_model_attribute (tree *node, tree name, tree args,
			     int flags ATTRIBUTE_UNUSED, bool *no_add_attrs)
{
  ia64_addr_area addr_area = ADDR_AREA_NORMAL;
  ia64_addr_area area;
  tree arg, decl = *node;

  init_idents ();
  arg = TREE_VALUE (args);
  if (arg == small_ident1 || arg == small_ident2)
    {
      addr_area = ADDR_AREA_SMALL;
    }
  else
    {
      warning (OPT_Wattributes, "invalid argument of %qE attribute",
	       name);
      *no_add_attrs = true;
    }

  switch (TREE_CODE (decl))
    {
    case VAR_DECL:
      if ((DECL_CONTEXT (decl) && TREE_CODE (DECL_CONTEXT (decl))
	   == FUNCTION_DECL)
	  && !TREE_STATIC (decl))
	{
	  error_at (DECL_SOURCE_LOCATION (decl),
		    "an address area attribute cannot be specified for "
		    "local variables");
	  *no_add_attrs = true;
	}
      area = ia64_get_addr_area (decl);
      if (area != ADDR_AREA_NORMAL && addr_area != area)
	{
	  error ("address area of %q+D conflicts with previous "
		 "declaration", decl);
	  *no_add_attrs = true;
	}
      break;

    case FUNCTION_DECL:
      error_at (DECL_SOURCE_LOCATION (decl),
		"address area attribute cannot be specified for "
		"functions");
      *no_add_attrs = true;
      break;

    default:
      warning (OPT_Wattributes, "%qE attribute ignored",
	       name);
      *no_add_attrs = true;
      break;
    }

  return NULL_TREE;
}

/* Part of the low level implementation of DEC Ada pragma Common_Object which
   enables the shared use of variables stored in overlaid linker areas
   corresponding to the use of Fortran COMMON.  */

static tree
ia64_vms_common_object_attribute (tree *node, tree name, tree args,
				  int flags ATTRIBUTE_UNUSED,
				  bool *no_add_attrs)
{
    tree decl = *node;
    tree id;

    gcc_assert (DECL_P (decl));
  
    DECL_COMMON (decl) = 1;
    id = TREE_VALUE (args);
    if (TREE_CODE (id) != IDENTIFIER_NODE && TREE_CODE (id) != STRING_CST)
      {
	error ("%qE attribute requires a string constant argument", name);
	*no_add_attrs = true;
	return NULL_TREE;
      }
    return NULL_TREE;
}

/* Part of the low level implementation of DEC Ada pragma Common_Object.  */

void
ia64_vms_output_aligned_decl_common (FILE *file, tree decl, const char *name,
				     unsigned HOST_WIDE_INT size,
				     unsigned int align)
{
  tree attr = DECL_ATTRIBUTES (decl);

  if (attr)
    attr = lookup_attribute ("common_object", attr);
  if (attr)
    {
      tree id = TREE_VALUE (TREE_VALUE (attr));
      const char *name;

      if (TREE_CODE (id) == IDENTIFIER_NODE)
        name = IDENTIFIER_POINTER (id);
      else if (TREE_CODE (id) == STRING_CST)
        name = TREE_STRING_POINTER (id);
      else
        abort ();

      fprintf (file, "\t.vms_common\t\"%s\",", name);
    }
  else
    fprintf (file, "%s", COMMON_ASM_OP);

  /*  Code from elfos.h.  */
  assemble_name (file, name);
  fprintf (file, ","HOST_WIDE_INT_PRINT_UNSIGNED",%u",
           size, align / BITS_PER_UNIT);

  fputc ('\n', file);
}

static void
ia64_encode_addr_area (tree decl, rtx symbol)
{
  int flags;

  flags = SYMBOL_REF_FLAGS (symbol);
  switch (ia64_get_addr_area (decl))
    {
    case ADDR_AREA_NORMAL: break;
    case ADDR_AREA_SMALL: flags |= SYMBOL_FLAG_SMALL_ADDR; break;
    default: gcc_unreachable ();
    }
  SYMBOL_REF_FLAGS (symbol) = flags;
}

static void
ia64_encode_section_info (tree decl, rtx rtl, int first)
{
  default_encode_section_info (decl, rtl, first);

  /* Careful not to prod global register variables.  */
  if (TREE_CODE (decl) == VAR_DECL
      && GET_CODE (DECL_RTL (decl)) == MEM
      && GET_CODE (XEXP (DECL_RTL (decl), 0)) == SYMBOL_REF
      && (TREE_STATIC (decl) || DECL_EXTERNAL (decl)))
    ia64_encode_addr_area (decl, XEXP (rtl, 0));
}

/* Return 1 if the operands of a move are ok.  */

int
ia64_move_ok (rtx dst, rtx src)
{
  /* If we're under init_recog_no_volatile, we'll not be able to use
     memory_operand.  So check the code directly and don't worry about
     the validity of the underlying address, which should have been
     checked elsewhere anyway.  */
  if (GET_CODE (dst) != MEM)
    return 1;
  if (GET_CODE (src) == MEM)
    return 0;
  if (register_operand (src, VOIDmode))
    return 1;

  /* Otherwise, this must be a constant, and that either 0 or 0.0 or 1.0.  */
  if (INTEGRAL_MODE_P (GET_MODE (dst)))
    return src == const0_rtx;
  else
    return satisfies_constraint_G (src);
}

/* Return 1 if the operands are ok for a floating point load pair.  */

int
ia64_load_pair_ok (rtx dst, rtx src)
{
  /* ??? There is a thinko in the implementation of the "x" constraint and the
     FP_REGS class.  The constraint will also reject (reg f30:TI) so we must
     also return false for it.  */
  if (GET_CODE (dst) != REG
      || !(FP_REGNO_P (REGNO (dst)) && FP_REGNO_P (REGNO (dst) + 1)))
    return 0;
  if (GET_CODE (src) != MEM || MEM_VOLATILE_P (src))
    return 0;
  switch (GET_CODE (XEXP (src, 0)))
    {
    case REG:
    case POST_INC:
      break;
    case POST_DEC:
      return 0;
    case POST_MODIFY:
      {
	rtx adjust = XEXP (XEXP (XEXP (src, 0), 1), 1);

	if (GET_CODE (adjust) != CONST_INT
	    || INTVAL (adjust) != GET_MODE_SIZE (GET_MODE (src)))
	  return 0;
      }
      break;
    default:
      abort ();
    }
  return 1;
}

int
addp4_optimize_ok (rtx op1, rtx op2)
{
  return (basereg_operand (op1, GET_MODE(op1)) !=
	  basereg_operand (op2, GET_MODE(op2)));
}

/* Check if OP is a mask suitable for use with SHIFT in a dep.z instruction.
   Return the length of the field, or <= 0 on failure.  */

int
ia64_depz_field_mask (rtx rop, rtx rshift)
{
  unsigned HOST_WIDE_INT op = INTVAL (rop);
  unsigned HOST_WIDE_INT shift = INTVAL (rshift);

  /* Get rid of the zero bits we're shifting in.  */
  op >>= shift;

  /* We must now have a solid block of 1's at bit 0.  */
  return exact_log2 (op + 1);
}

/* Return the TLS model to use for ADDR.  */

static enum tls_model
tls_symbolic_operand_type (rtx addr)
{
  enum tls_model tls_kind = TLS_MODEL_NONE;

  if (GET_CODE (addr) == CONST)
    {
      if (GET_CODE (XEXP (addr, 0)) == PLUS
	  && GET_CODE (XEXP (XEXP (addr, 0), 0)) == SYMBOL_REF)
        tls_kind = SYMBOL_REF_TLS_MODEL (XEXP (XEXP (addr, 0), 0));
    }
  else if (GET_CODE (addr) == SYMBOL_REF)
    tls_kind = SYMBOL_REF_TLS_MODEL (addr);

  return tls_kind;
}

/* Returns true if REG (assumed to be a `reg' RTX) is valid for use
   as a base register.  */

static inline bool
ia64_reg_ok_for_base_p (const_rtx reg, bool strict)
{
  if (strict
      && REGNO_OK_FOR_BASE_P (REGNO (reg)))
    return true;
  else if (!strict
	   && (GENERAL_REGNO_P (REGNO (reg))
	       || !HARD_REGISTER_P (reg)))
    return true;
  else
    return false;
}

static bool
ia64_legitimate_address_reg (const_rtx reg, bool strict)
{
  if ((REG_P (reg) && ia64_reg_ok_for_base_p (reg, strict))
      || (GET_CODE (reg) == SUBREG && REG_P (XEXP (reg, 0))
	  && ia64_reg_ok_for_base_p (XEXP (reg, 0), strict)))
    return true;

  return false;
}

static bool
ia64_legitimate_address_disp (const_rtx reg, const_rtx disp, bool strict)
{
  if (GET_CODE (disp) == PLUS
      && rtx_equal_p (reg, XEXP (disp, 0))
      && (ia64_legitimate_address_reg (XEXP (disp, 1), strict)
	  || (CONST_INT_P (XEXP (disp, 1))
	      && IN_RANGE (INTVAL (XEXP (disp, 1)), -256, 255))))
    return true;

  return false;
}

/* Implement TARGET_LEGITIMATE_ADDRESS_P.  */

static bool
ia64_legitimate_address_p (enum machine_mode mode ATTRIBUTE_UNUSED,
			   rtx x, bool strict)
{
  if (ia64_legitimate_address_reg (x, strict))
    return true;
  else if ((GET_CODE (x) == POST_INC || GET_CODE (x) == POST_DEC)
	   && ia64_legitimate_address_reg (XEXP (x, 0), strict)
	   && XEXP (x, 0) != arg_pointer_rtx) 
    return true;
  else if (GET_CODE (x) == POST_MODIFY
	   && ia64_legitimate_address_reg (XEXP (x, 0), strict)
	   && XEXP (x, 0) != arg_pointer_rtx
	   && ia64_legitimate_address_disp (XEXP (x, 0), XEXP (x, 1), strict))
    return true;
  else
    return false;
}

/* Return true if X is a constant that is valid for some immediate
   field in an instruction.  */

static bool
ia64_legitimate_constant_p (enum machine_mode mode, rtx x)
{
  switch (GET_CODE (x))
    {
    case CONST_INT:
    case LABEL_REF:
      return true;

    case CONST_DOUBLE:
      if (GET_MODE (x) == VOIDmode || mode == SFmode || mode == DFmode)
	return true;
      return satisfies_constraint_G (x);

    case CONST:
    case SYMBOL_REF:
      /* ??? Short term workaround for PR 28490.  We must make the code here
	 match the code in ia64_expand_move and move_operand, even though they
	 are both technically wrong.  */
      if (tls_symbolic_operand_type (x) == 0)
	{
	  HOST_WIDE_INT addend = 0;
	  rtx op = x;

	  if (GET_CODE (op) == CONST
	      && GET_CODE (XEXP (op, 0)) == PLUS
	      && GET_CODE (XEXP (XEXP (op, 0), 1)) == CONST_INT)
	    {
	      addend = INTVAL (XEXP (XEXP (op, 0), 1));
	      op = XEXP (XEXP (op, 0), 0);
	    }

          if (any_offset_symbol_operand (op, mode)
              || function_operand (op, mode))
            return true;
	  if (aligned_offset_symbol_operand (op, mode))
	    return (addend & 0x3fff) == 0;
	  return false;
	}
      return false;

    case CONST_VECTOR:
      if (mode == V2SFmode)
	return satisfies_constraint_Y (x);

      return (GET_MODE_CLASS (mode) == MODE_VECTOR_INT
	      && GET_MODE_SIZE (mode) <= 8);

    default:
      return false;
    }
}

/* Don't allow TLS addresses to get spilled to memory.  */

static bool
ia64_cannot_force_const_mem (enum machine_mode mode, rtx x)
{
  if (mode == RFmode)
    return true;
  return tls_symbolic_operand_type (x) != 0;
}

/* Expand a symbolic constant load.  */

bool
ia64_expand_load_address (rtx dest, rtx src)
{
  gcc_assert (GET_CODE (dest) == REG);

  /* ILP32 mode still loads 64-bits of data from the GOT.  This avoids
     having to pointer-extend the value afterward.  Other forms of address
     computation below are also more natural to compute as 64-bit quantities.
     If we've been given an SImode destination register, change it.  */
  if (GET_MODE (dest) != Pmode)
    dest = gen_rtx_REG_offset (dest, Pmode, REGNO (dest),
			       byte_lowpart_offset (Pmode, GET_MODE (dest)));

  if (TARGET_NO_PIC)
    return false;
  if (small_addr_symbolic_operand (src, VOIDmode))
    return false;

  if (TARGET_AUTO_PIC)
    emit_insn (gen_load_gprel64 (dest, src));
  else if (GET_CODE (src) == SYMBOL_REF && SYMBOL_REF_FUNCTION_P (src))
    emit_insn (gen_load_fptr (dest, src));
  else if (sdata_symbolic_operand (src, VOIDmode))
    emit_insn (gen_load_gprel (dest, src));
  else
    {
      HOST_WIDE_INT addend = 0;
      rtx tmp;

      /* We did split constant offsets in ia64_expand_move, and we did try
	 to keep them split in move_operand, but we also allowed reload to
	 rematerialize arbitrary constants rather than spill the value to
	 the stack and reload it.  So we have to be prepared here to split
	 them apart again.  */
      if (GET_CODE (src) == CONST)
	{
	  HOST_WIDE_INT hi, lo;

	  hi = INTVAL (XEXP (XEXP (src, 0), 1));
	  lo = ((hi & 0x3fff) ^ 0x2000) - 0x2000;
	  hi = hi - lo;

	  if (lo != 0)
	    {
	      addend = lo;
	      src = plus_constant (Pmode, XEXP (XEXP (src, 0), 0), hi);
	    }
	}

      tmp = gen_rtx_HIGH (Pmode, src);
      tmp = gen_rtx_PLUS (Pmode, tmp, pic_offset_table_rtx);
      emit_insn (gen_rtx_SET (VOIDmode, dest, tmp));

      tmp = gen_rtx_LO_SUM (Pmode, gen_const_mem (Pmode, dest), src);
      emit_insn (gen_rtx_SET (VOIDmode, dest, tmp));

      if (addend)
	{
	  tmp = gen_rtx_PLUS (Pmode, dest, GEN_INT (addend));
	  emit_insn (gen_rtx_SET (VOIDmode, dest, tmp));
	}
    }

  return true;
}

static GTY(()) rtx gen_tls_tga;
static rtx
gen_tls_get_addr (void)
{
  if (!gen_tls_tga)
    gen_tls_tga = init_one_libfunc ("__tls_get_addr");
  return gen_tls_tga;
}

static GTY(()) rtx thread_pointer_rtx;
static rtx
gen_thread_pointer (void)
{
  if (!thread_pointer_rtx)
    thread_pointer_rtx = gen_rtx_REG (Pmode, 13);
  return thread_pointer_rtx;
}

static rtx
ia64_expand_tls_address (enum tls_model tls_kind, rtx op0, rtx op1,
			 rtx orig_op1, HOST_WIDE_INT addend)
{
  rtx tga_op1, tga_op2, tga_ret, tga_eqv, tmp, insns;
  rtx orig_op0 = op0;
  HOST_WIDE_INT addend_lo, addend_hi;

  switch (tls_kind)
    {
    case TLS_MODEL_GLOBAL_DYNAMIC:
      start_sequence ();

      tga_op1 = gen_reg_rtx (Pmode);
      emit_insn (gen_load_dtpmod (tga_op1, op1));

      tga_op2 = gen_reg_rtx (Pmode);
      emit_insn (gen_load_dtprel (tga_op2, op1));

      tga_ret = emit_library_call_value (gen_tls_get_addr (), NULL_RTX,
					 LCT_CONST, Pmode, 2, tga_op1,
					 Pmode, tga_op2, Pmode);

      insns = get_insns ();
      end_sequence ();

      if (GET_MODE (op0) != Pmode)
	op0 = tga_ret;
      emit_libcall_block (insns, op0, tga_ret, op1);
      break;

    case TLS_MODEL_LOCAL_DYNAMIC:
      /* ??? This isn't the completely proper way to do local-dynamic
	 If the call to __tls_get_addr is used only by a single symbol,
	 then we should (somehow) move the dtprel to the second arg
	 to avoid the extra add.  */
      start_sequence ();

      tga_op1 = gen_reg_rtx (Pmode);
      emit_insn (gen_load_dtpmod (tga_op1, op1));

      tga_op2 = const0_rtx;

      tga_ret = emit_library_call_value (gen_tls_get_addr (), NULL_RTX,
					 LCT_CONST, Pmode, 2, tga_op1,
					 Pmode, tga_op2, Pmode);

      insns = get_insns ();
      end_sequence ();

      tga_eqv = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, const0_rtx),
				UNSPEC_LD_BASE);
      tmp = gen_reg_rtx (Pmode);
      emit_libcall_block (insns, tmp, tga_ret, tga_eqv);

      if (!register_operand (op0, Pmode))
	op0 = gen_reg_rtx (Pmode);
      if (TARGET_TLS64)
	{
	  emit_insn (gen_load_dtprel (op0, op1));
	  emit_insn (gen_adddi3 (op0, tmp, op0));
	}
      else
	emit_insn (gen_add_dtprel (op0, op1, tmp));
      break;

    case TLS_MODEL_INITIAL_EXEC:
      addend_lo = ((addend & 0x3fff) ^ 0x2000) - 0x2000;
      addend_hi = addend - addend_lo;

      op1 = plus_constant (Pmode, op1, addend_hi);
      addend = addend_lo;

      tmp = gen_reg_rtx (Pmode);
      emit_insn (gen_load_tprel (tmp, op1));

      if (!register_operand (op0, Pmode))
	op0 = gen_reg_rtx (Pmode);
      emit_insn (gen_adddi3 (op0, tmp, gen_thread_pointer ()));
      break;

    case TLS_MODEL_LOCAL_EXEC:
      if (!register_operand (op0, Pmode))
	op0 = gen_reg_rtx (Pmode);

      op1 = orig_op1;
      addend = 0;
      if (TARGET_TLS64)
	{
	  emit_insn (gen_load_tprel (op0, op1));
	  emit_insn (gen_adddi3 (op0, op0, gen_thread_pointer ()));
	}
      else
	emit_insn (gen_add_tprel (op0, op1, gen_thread_pointer ()));
      break;

    default:
      gcc_unreachable ();
    }

  if (addend)
    op0 = expand_simple_binop (Pmode, PLUS, op0, GEN_INT (addend),
			       orig_op0, 1, OPTAB_DIRECT);
  if (orig_op0 == op0)
    return NULL_RTX;
  if (GET_MODE (orig_op0) == Pmode)
    return op0;
  return gen_lowpart (GET_MODE (orig_op0), op0);
}

rtx
ia64_expand_move (rtx op0, rtx op1)
{
  enum machine_mode mode = GET_MODE (op0);

  if (!reload_in_progress && !reload_completed && !ia64_move_ok (op0, op1))
    op1 = force_reg (mode, op1);

  if ((mode == Pmode || mode == ptr_mode) && symbolic_operand (op1, VOIDmode))
    {
      HOST_WIDE_INT addend = 0;
      enum tls_model tls_kind;
      rtx sym = op1;

      if (GET_CODE (op1) == CONST
	  && GET_CODE (XEXP (op1, 0)) == PLUS
	  && GET_CODE (XEXP (XEXP (op1, 0), 1)) == CONST_INT)
	{
	  addend = INTVAL (XEXP (XEXP (op1, 0), 1));
	  sym = XEXP (XEXP (op1, 0), 0);
	}

      tls_kind = tls_symbolic_operand_type (sym);
      if (tls_kind)
	return ia64_expand_tls_address (tls_kind, op0, sym, op1, addend);

      if (any_offset_symbol_operand (sym, mode))
	addend = 0;
      else if (aligned_offset_symbol_operand (sym, mode))
	{
	  HOST_WIDE_INT addend_lo, addend_hi;
	      
	  addend_lo = ((addend & 0x3fff) ^ 0x2000) - 0x2000;
	  addend_hi = addend - addend_lo;

	  if (addend_lo != 0)
	    {
	      op1 = plus_constant (mode, sym, addend_hi);
	      addend = addend_lo;
	    }
	  else
	    addend = 0;
	}
      else
	op1 = sym;

      if (reload_completed)
	{
	  /* We really should have taken care of this offset earlier.  */
	  gcc_assert (addend == 0);
	  if (ia64_expand_load_address (op0, op1))
	    return NULL_RTX;
	}

      if (addend)
	{
	  rtx subtarget = !can_create_pseudo_p () ? op0 : gen_reg_rtx (mode);

	  emit_insn (gen_rtx_SET (VOIDmode, subtarget, op1));

	  op1 = expand_simple_binop (mode, PLUS, subtarget,
				     GEN_INT (addend), op0, 1, OPTAB_DIRECT);
	  if (op0 == op1)
	    return NULL_RTX;
	}
    }

  return op1;
}

/* Split a move from OP1 to OP0 conditional on COND.  */

void
ia64_emit_cond_move (rtx op0, rtx op1, rtx cond)
{
  rtx insn, first = get_last_insn ();

  emit_move_insn (op0, op1);

  for (insn = get_last_insn (); insn != first; insn = PREV_INSN (insn))
    if (INSN_P (insn))
      PATTERN (insn) = gen_rtx_COND_EXEC (VOIDmode, copy_rtx (cond),
					  PATTERN (insn));
}

/* Split a post-reload TImode or TFmode reference into two DImode
   components.  This is made extra difficult by the fact that we do
   not get any scratch registers to work with, because reload cannot
   be prevented from giving us a scratch that overlaps the register
   pair involved.  So instead, when addressing memory, we tweak the
   pointer register up and back down with POST_INCs.  Or up and not
   back down when we can get away with it.

   REVERSED is true when the loads must be done in reversed order
   (high word first) for correctness.  DEAD is true when the pointer
   dies with the second insn we generate and therefore the second
   address must not carry a postmodify.

   May return an insn which is to be emitted after the moves.  */

static rtx
ia64_split_tmode (rtx out[2], rtx in, bool reversed, bool dead)
{
  rtx fixup = 0;

  switch (GET_CODE (in))
    {
    case REG:
      out[reversed] = gen_rtx_REG (DImode, REGNO (in));
      out[!reversed] = gen_rtx_REG (DImode, REGNO (in) + 1);
      break;

    case CONST_INT:
    case CONST_DOUBLE:
      /* Cannot occur reversed.  */
      gcc_assert (!reversed);
      
      if (GET_MODE (in) != TFmode)
	split_double (in, &out[0], &out[1]);
      else
	/* split_double does not understand how to split a TFmode
	   quantity into a pair of DImode constants.  */
	{
	  REAL_VALUE_TYPE r;
	  unsigned HOST_WIDE_INT p[2];
	  long l[4];  /* TFmode is 128 bits */

	  REAL_VALUE_FROM_CONST_DOUBLE (r, in);
	  real_to_target (l, &r, TFmode);

	  if (FLOAT_WORDS_BIG_ENDIAN)
	    {
	      p[0] = (((unsigned HOST_WIDE_INT) l[0]) << 32) + l[1];
	      p[1] = (((unsigned HOST_WIDE_INT) l[2]) << 32) + l[3];
	    }
	  else
	    {
	      p[0] = (((unsigned HOST_WIDE_INT) l[1]) << 32) + l[0];
	      p[1] = (((unsigned HOST_WIDE_INT) l[3]) << 32) + l[2];
	    }
	  out[0] = GEN_INT (p[0]);
	  out[1] = GEN_INT (p[1]);
	}
      break;

    case MEM:
      {
	rtx base = XEXP (in, 0);
	rtx offset;

	switch (GET_CODE (base))
	  {
	  case REG:
	    if (!reversed)
	      {
		out[0] = adjust_automodify_address
		  (in, DImode, gen_rtx_POST_INC (Pmode, base), 0);
		out[1] = adjust_automodify_address
		  (in, DImode, dead ? 0 : gen_rtx_POST_DEC (Pmode, base), 8);
	      }
	    else
	      {
		/* Reversal requires a pre-increment, which can only
		   be done as a separate insn.  */
		emit_insn (gen_adddi3 (base, base, GEN_INT (8)));
		out[0] = adjust_automodify_address
		  (in, DImode, gen_rtx_POST_DEC (Pmode, base), 8);
		out[1] = adjust_address (in, DImode, 0);
	      }
	    break;

	  case POST_INC:
	    gcc_assert (!reversed && !dead);
	    
	    /* Just do the increment in two steps.  */
	    out[0] = adjust_automodify_address (in, DImode, 0, 0);
	    out[1] = adjust_automodify_address (in, DImode, 0, 8);
	    break;

	  case POST_DEC:
	    gcc_assert (!reversed && !dead);
	    
	    /* Add 8, subtract 24.  */
	    base = XEXP (base, 0);
	    out[0] = adjust_automodify_address
	      (in, DImode, gen_rtx_POST_INC (Pmode, base), 0);
	    out[1] = adjust_automodify_address
	      (in, DImode,
	       gen_rtx_POST_MODIFY (Pmode, base,
				    plus_constant (Pmode, base, -24)),
	       8);
	    break;

	  case POST_MODIFY:
	    gcc_assert (!reversed && !dead);

	    /* Extract and adjust the modification.  This case is
	       trickier than the others, because we might have an
	       index register, or we might have a combined offset that
	       doesn't fit a signed 9-bit displacement field.  We can
	       assume the incoming expression is already legitimate.  */
	    offset = XEXP (base, 1);
	    base = XEXP (base, 0);

	    out[0] = adjust_automodify_address
	      (in, DImode, gen_rtx_POST_INC (Pmode, base), 0);

	    if (GET_CODE (XEXP (offset, 1)) == REG)
	      {
		/* Can't adjust the postmodify to match.  Emit the
		   original, then a separate addition insn.  */
		out[1] = adjust_automodify_address (in, DImode, 0, 8);
		fixup = gen_adddi3 (base, base, GEN_INT (-8));
	      }
	    else
	      {
		gcc_assert (GET_CODE (XEXP (offset, 1)) == CONST_INT);
		if (INTVAL (XEXP (offset, 1)) < -256 + 8)
		  {
		    /* Again the postmodify cannot be made to match,
		       but in this case it's more efficient to get rid
		       of the postmodify entirely and fix up with an
		       add insn.  */
		    out[1] = adjust_automodify_address (in, DImode, base, 8);
		    fixup = gen_adddi3
		      (base, base, GEN_INT (INTVAL (XEXP (offset, 1)) - 8));
		  }
		else
		  {
		    /* Combined offset still fits in the displacement field.
		       (We cannot overflow it at the high end.)  */
		    out[1] = adjust_automodify_address
		      (in, DImode, gen_rtx_POST_MODIFY
		       (Pmode, base, gen_rtx_PLUS
			(Pmode, base,
			 GEN_INT (INTVAL (XEXP (offset, 1)) - 8))),
		       8);
		  }
	      }
	    break;

	  default:
	    gcc_unreachable ();
	  }
	break;
      }

    default:
      gcc_unreachable ();
    }

  return fixup;
}

/* Split a TImode or TFmode move instruction after reload.
   This is used by *movtf_internal and *movti_internal.  */
void
ia64_split_tmode_move (rtx operands[])
{
  rtx in[2], out[2], insn;
  rtx fixup[2];
  bool dead = false;
  bool reversed = false;

  /* It is possible for reload to decide to overwrite a pointer with
     the value it points to.  In that case we have to do the loads in
     the appropriate order so that the pointer is not destroyed too
     early.  Also we must not generate a postmodify for that second
     load, or rws_access_regno will die.  And we must not generate a
     postmodify for the second load if the destination register 
     overlaps with the base register.  */
  if (GET_CODE (operands[1]) == MEM
      && reg_overlap_mentioned_p (operands[0], operands[1]))
    {
      rtx base = XEXP (operands[1], 0);
      while (GET_CODE (base) != REG)
	base = XEXP (base, 0);

      if (REGNO (base) == REGNO (operands[0]))
	reversed = true;

      if (refers_to_regno_p (REGNO (operands[0]),
			     REGNO (operands[0])+2,
			     base, 0))
	dead = true;
    }
  /* Another reason to do the moves in reversed order is if the first
     element of the target register pair is also the second element of
     the source register pair.  */
  if (GET_CODE (operands[0]) == REG && GET_CODE (operands[1]) == REG
      && REGNO (operands[0]) == REGNO (operands[1]) + 1)
    reversed = true;

  fixup[0] = ia64_split_tmode (in, operands[1], reversed, dead);
  fixup[1] = ia64_split_tmode (out, operands[0], reversed, dead);

#define MAYBE_ADD_REG_INC_NOTE(INSN, EXP)				\
  if (GET_CODE (EXP) == MEM						\
      && (GET_CODE (XEXP (EXP, 0)) == POST_MODIFY			\
	  || GET_CODE (XEXP (EXP, 0)) == POST_INC			\
	  || GET_CODE (XEXP (EXP, 0)) == POST_DEC))			\
    add_reg_note (insn, REG_INC, XEXP (XEXP (EXP, 0), 0))

  insn = emit_insn (gen_rtx_SET (VOIDmode, out[0], in[0]));
  MAYBE_ADD_REG_INC_NOTE (insn, in[0]);
  MAYBE_ADD_REG_INC_NOTE (insn, out[0]);

  insn = emit_insn (gen_rtx_SET (VOIDmode, out[1], in[1]));
  MAYBE_ADD_REG_INC_NOTE (insn, in[1]);
  MAYBE_ADD_REG_INC_NOTE (insn, out[1]);

  if (fixup[0])
    emit_insn (fixup[0]);
  if (fixup[1])
    emit_insn (fixup[1]);

#undef MAYBE_ADD_REG_INC_NOTE
}

/* ??? Fixing GR->FR XFmode moves during reload is hard.  You need to go
   through memory plus an extra GR scratch register.  Except that you can
   either get the first from SECONDARY_MEMORY_NEEDED or the second from
   SECONDARY_RELOAD_CLASS, but not both.

   We got into problems in the first place by allowing a construct like
   (subreg:XF (reg:TI)), which we got from a union containing a long double.
   This solution attempts to prevent this situation from occurring.  When
   we see something like the above, we spill the inner register to memory.  */

static rtx
spill_xfmode_rfmode_operand (rtx in, int force, enum machine_mode mode)
{
  if (GET_CODE (in) == SUBREG
      && GET_MODE (SUBREG_REG (in)) == TImode
      && GET_CODE (SUBREG_REG (in)) == REG)
    {
      rtx memt = assign_stack_temp (TImode, 16);
      emit_move_insn (memt, SUBREG_REG (in));
      return adjust_address (memt, mode, 0);
    }
  else if (force && GET_CODE (in) == REG)
    {
      rtx memx = assign_stack_temp (mode, 16);
      emit_move_insn (memx, in);
      return memx;
    }
  else
    return in;
}

/* Expand the movxf or movrf pattern (MODE says which) with the given
   OPERANDS, returning true if the pattern should then invoke
   DONE.  */

bool
ia64_expand_movxf_movrf (enum machine_mode mode, rtx operands[])
{
  rtx op0 = operands[0];

  if (GET_CODE (op0) == SUBREG)
    op0 = SUBREG_REG (op0);

  /* We must support XFmode loads into general registers for stdarg/vararg,
     unprototyped calls, and a rare case where a long double is passed as
     an argument after a float HFA fills the FP registers.  We split them into
     DImode loads for convenience.  We also need to support XFmode stores
     for the last case.  This case does not happen for stdarg/vararg routines,
     because we do a block store to memory of unnamed arguments.  */

  if (GET_CODE (op0) == REG && GR_REGNO_P (REGNO (op0)))
    {
      rtx out[2];

      /* We're hoping to transform everything that deals with XFmode
	 quantities and GR registers early in the compiler.  */
      gcc_assert (can_create_pseudo_p ());

      /* Struct to register can just use TImode instead.  */
      if ((GET_CODE (operands[1]) == SUBREG
	   && GET_MODE (SUBREG_REG (operands[1])) == TImode)
	  || (GET_CODE (operands[1]) == REG
	      && GR_REGNO_P (REGNO (operands[1]))))
	{
	  rtx op1 = operands[1];

	  if (GET_CODE (op1) == SUBREG)
	    op1 = SUBREG_REG (op1);
	  else
	    op1 = gen_rtx_REG (TImode, REGNO (op1));

	  emit_move_insn (gen_rtx_REG (TImode, REGNO (op0)), op1);
	  return true;
	}

      if (GET_CODE (operands[1]) == CONST_DOUBLE)
	{
	  /* Don't word-swap when reading in the constant.  */
	  emit_move_insn (gen_rtx_REG (DImode, REGNO (op0)),
			  operand_subword (operands[1], WORDS_BIG_ENDIAN,
					   0, mode));
	  emit_move_insn (gen_rtx_REG (DImode, REGNO (op0) + 1),
			  operand_subword (operands[1], !WORDS_BIG_ENDIAN,
					   0, mode));
	  return true;
	}

      /* If the quantity is in a register not known to be GR, spill it.  */
      if (register_operand (operands[1], mode))
	operands[1] = spill_xfmode_rfmode_operand (operands[1], 1, mode);

      gcc_assert (GET_CODE (operands[1]) == MEM);

      /* Don't word-swap when reading in the value.  */
      out[0] = gen_rtx_REG (DImode, REGNO (op0));
      out[1] = gen_rtx_REG (DImode, REGNO (op0) + 1);

      emit_move_insn (out[0], adjust_address (operands[1], DImode, 0));
      emit_move_insn (out[1], adjust_address (operands[1], DImode, 8));
      return true;
    }

  if (GET_CODE (operands[1]) == REG && GR_REGNO_P (REGNO (operands[1])))
    {
      /* We're hoping to transform everything that deals with XFmode
	 quantities and GR registers early in the compiler.  */
      gcc_assert (can_create_pseudo_p ());

      /* Op0 can't be a GR_REG here, as that case is handled above.
	 If op0 is a register, then we spill op1, so that we now have a
	 MEM operand.  This requires creating an XFmode subreg of a TImode reg
	 to force the spill.  */
      if (register_operand (operands[0], mode))
	{
	  rtx op1 = gen_rtx_REG (TImode, REGNO (operands[1]));
	  op1 = gen_rtx_SUBREG (mode, op1, 0);
	  operands[1] = spill_xfmode_rfmode_operand (op1, 0, mode);
	}

      else
	{
	  rtx in[2];

	  gcc_assert (GET_CODE (operands[0]) == MEM);

	  /* Don't word-swap when writing out the value.  */
	  in[0] = gen_rtx_REG (DImode, REGNO (operands[1]));
	  in[1] = gen_rtx_REG (DImode, REGNO (operands[1]) + 1);

	  emit_move_insn (adjust_address (operands[0], DImode, 0), in[0]);
	  emit_move_insn (adjust_address (operands[0], DImode, 8), in[1]);
	  return true;
	}
    }

  if (!reload_in_progress && !reload_completed)
    {
      operands[1] = spill_xfmode_rfmode_operand (operands[1], 0, mode);

      if (GET_MODE (op0) == TImode && GET_CODE (op0) == REG)
	{
	  rtx memt, memx, in = operands[1];
	  if (CONSTANT_P (in))
	    in = validize_mem (force_const_mem (mode, in));
	  if (GET_CODE (in) == MEM)
	    memt = adjust_address (in, TImode, 0);
	  else
	    {
	      memt = assign_stack_temp (TImode, 16);
	      memx = adjust_address (memt, mode, 0);
	      emit_move_insn (memx, in);
	    }
	  emit_move_insn (op0, memt);
	  return true;
	}

      if (!ia64_move_ok (operands[0], operands[1]))
	operands[1] = force_reg (mode, operands[1]);
    }

  return false;
}

/* Emit comparison instruction if necessary, replacing *EXPR, *OP0, *OP1
   with the expression that holds the compare result (in VOIDmode).  */

static GTY(()) rtx cmptf_libfunc;

void
ia64_expand_compare (rtx *expr, rtx *op0, rtx *op1)
{
  enum rtx_code code = GET_CODE (*expr);
  rtx cmp;

  /* If we have a BImode input, then we already have a compare result, and
     do not need to emit another comparison.  */
  if (GET_MODE (*op0) == BImode)
    {
      gcc_assert ((code == NE || code == EQ) && *op1 == const0_rtx);
      cmp = *op0;
    }
  /* HPUX TFmode compare requires a library call to _U_Qfcmp, which takes a
     magic number as its third argument, that indicates what to do.
     The return value is an integer to be compared against zero.  */
  else if (TARGET_HPUX && GET_MODE (*op0) == TFmode)
    {
      enum qfcmp_magic {
	QCMP_INV = 1,	/* Raise FP_INVALID on NaNs as a side effect.  */
	QCMP_UNORD = 2,
	QCMP_EQ = 4,
	QCMP_LT = 8,
	QCMP_GT = 16
      };
      int magic;
      enum rtx_code ncode;
      rtx ret, insns;
      
      gcc_assert (cmptf_libfunc && GET_MODE (*op1) == TFmode);
      switch (code)
	{
	  /* 1 = equal, 0 = not equal.  Equality operators do
	     not raise FP_INVALID when given a NaN operand.  */
	case EQ:        magic = QCMP_EQ;                  ncode = NE; break;
	case NE:        magic = QCMP_EQ;                  ncode = EQ; break;
	  /* isunordered() from C99.  */
	case UNORDERED: magic = QCMP_UNORD;               ncode = NE; break;
	case ORDERED:   magic = QCMP_UNORD;               ncode = EQ; break;
	  /* Relational operators raise FP_INVALID when given
	     a NaN operand.  */
	case LT:        magic = QCMP_LT        |QCMP_INV; ncode = NE; break;
	case LE:        magic = QCMP_LT|QCMP_EQ|QCMP_INV; ncode = NE; break;
	case GT:        magic = QCMP_GT        |QCMP_INV; ncode = NE; break;
	case GE:        magic = QCMP_GT|QCMP_EQ|QCMP_INV; ncode = NE; break;
          /* Unordered relational operators do not raise FP_INVALID
	     when given a NaN operand.  */
	case UNLT:    magic = QCMP_LT        |QCMP_UNORD; ncode = NE; break;
	case UNLE:    magic = QCMP_LT|QCMP_EQ|QCMP_UNORD; ncode = NE; break;
	case UNGT:    magic = QCMP_GT        |QCMP_UNORD; ncode = NE; break;
	case UNGE:    magic = QCMP_GT|QCMP_EQ|QCMP_UNORD; ncode = NE; break;
	  /* Not supported.  */
	case UNEQ:
	case LTGT:
	default: gcc_unreachable ();
	}

      start_sequence ();

      ret = emit_library_call_value (cmptf_libfunc, 0, LCT_CONST, DImode, 3,
				     *op0, TFmode, *op1, TFmode,
				     GEN_INT (magic), DImode);
      cmp = gen_reg_rtx (BImode);
      emit_insn (gen_rtx_SET (VOIDmode, cmp,
			      gen_rtx_fmt_ee (ncode, BImode,
					      ret, const0_rtx)));

      insns = get_insns ();
      end_sequence ();

      emit_libcall_block (insns, cmp, cmp,
			  gen_rtx_fmt_ee (code, BImode, *op0, *op1));
      code = NE;
    }
  else
    {
      cmp = gen_reg_rtx (BImode);
      emit_insn (gen_rtx_SET (VOIDmode, cmp,
			      gen_rtx_fmt_ee (code, BImode, *op0, *op1)));
      code = NE;
    }

  *expr = gen_rtx_fmt_ee (code, VOIDmode, cmp, const0_rtx);
  *op0 = cmp;
  *op1 = const0_rtx;
}

/* Generate an integral vector comparison.  Return true if the condition has
   been reversed, and so the sense of the comparison should be inverted.  */

static bool
ia64_expand_vecint_compare (enum rtx_code code, enum machine_mode mode,
			    rtx dest, rtx op0, rtx op1)
{
  bool negate = false;
  rtx x;

  /* Canonicalize the comparison to EQ, GT, GTU.  */
  switch (code)
    {
    case EQ:
    case GT:
    case GTU:
      break;

    case NE:
    case LE:
    case LEU:
      code = reverse_condition (code);
      negate = true;
      break;

    case GE:
    case GEU:
      code = reverse_condition (code);
      negate = true;
      /* FALLTHRU */

    case LT:
    case LTU:
      code = swap_condition (code);
      x = op0, op0 = op1, op1 = x;
      break;

    default:
      gcc_unreachable ();
    }

  /* Unsigned parallel compare is not supported by the hardware.  Play some
     tricks to turn this into a signed comparison against 0.  */
  if (code == GTU)
    {
      switch (mode)
	{
	case V2SImode:
	  {
	    rtx t1, t2, mask;

	    /* Subtract (-(INT MAX) - 1) from both operands to make
	       them signed.  */
	    mask = GEN_INT (0x80000000);
	    mask = gen_rtx_CONST_VECTOR (V2SImode, gen_rtvec (2, mask, mask));
	    mask = force_reg (mode, mask);
	    t1 = gen_reg_rtx (mode);
	    emit_insn (gen_subv2si3 (t1, op0, mask));
	    t2 = gen_reg_rtx (mode);
	    emit_insn (gen_subv2si3 (t2, op1, mask));
	    op0 = t1;
	    op1 = t2;
	    code = GT;
	  }
	  break;

	case V8QImode:
	case V4HImode:
	  /* Perform a parallel unsigned saturating subtraction.  */
	  x = gen_reg_rtx (mode);
	  emit_insn (gen_rtx_SET (VOIDmode, x,
				  gen_rtx_US_MINUS (mode, op0, op1)));

	  code = EQ;
	  op0 = x;
	  op1 = CONST0_RTX (mode);
	  negate = !negate;
	  break;

	default:
	  gcc_unreachable ();
	}
    }

  x = gen_rtx_fmt_ee (code, mode, op0, op1);
  emit_insn (gen_rtx_SET (VOIDmode, dest, x));

  return negate;
}

/* Emit an integral vector conditional move.  */

void
ia64_expand_vecint_cmov (rtx operands[])
{
  enum machine_mode mode = GET_MODE (operands[0]);
  enum rtx_code code = GET_CODE (operands[3]);
  bool negate;
  rtx cmp, x, ot, of;

  cmp = gen_reg_rtx (mode);
  negate = ia64_expand_vecint_compare (code, mode, cmp,
				       operands[4], operands[5]);

  ot = operands[1+negate];
  of = operands[2-negate];

  if (ot == CONST0_RTX (mode))
    {
      if (of == CONST0_RTX (mode))
	{
	  emit_move_insn (operands[0], ot);
	  return;
	}

      x = gen_rtx_NOT (mode, cmp);
      x = gen_rtx_AND (mode, x, of);
      emit_insn (gen_rtx_SET (VOIDmode, operands[0], x));
    }
  else if (of == CONST0_RTX (mode))
    {
      x = gen_rtx_AND (mode, cmp, ot);
      emit_insn (gen_rtx_SET (VOIDmode, operands[0], x));
    }
  else
    {
      rtx t, f;

      t = gen_reg_rtx (mode);
      x = gen_rtx_AND (mode, cmp, operands[1+negate]);
      emit_insn (gen_rtx_SET (VOIDmode, t, x));

      f = gen_reg_rtx (mode);
      x = gen_rtx_NOT (mode, cmp);
      x = gen_rtx_AND (mode, x, operands[2-negate]);
      emit_insn (gen_rtx_SET (VOIDmode, f, x));

      x = gen_rtx_IOR (mode, t, f);
      emit_insn (gen_rtx_SET (VOIDmode, operands[0], x));
    }
}

/* Emit an integral vector min or max operation.  Return true if all done.  */

bool
ia64_expand_vecint_minmax (enum rtx_code code, enum machine_mode mode,
			   rtx operands[])
{
  rtx xops[6];

  /* These four combinations are supported directly.  */
  if (mode == V8QImode && (code == UMIN || code == UMAX))
    return false;
  if (mode == V4HImode && (code == SMIN || code == SMAX))
    return false;

  /* This combination can be implemented with only saturating subtraction.  */
  if (mode == V4HImode && code == UMAX)
    {
      rtx x, tmp = gen_reg_rtx (mode);

      x = gen_rtx_US_MINUS (mode, operands[1], operands[2]);
      emit_insn (gen_rtx_SET (VOIDmode, tmp, x));

      emit_insn (gen_addv4hi3 (operands[0], tmp, operands[2]));
      return true;
    }

  /* Everything else implemented via vector comparisons.  */
  xops[0] = operands[0];
  xops[4] = xops[1] = operands[1];
  xops[5] = xops[2] = operands[2];

  switch (code)
    {
    case UMIN:
      code = LTU;
      break;
    case UMAX:
      code = GTU;
      break;
    case SMIN:
      code = LT;
      break;
    case SMAX:
      code = GT;
      break;
    default:
      gcc_unreachable ();
    }
  xops[3] = gen_rtx_fmt_ee (code, VOIDmode, operands[1], operands[2]);

  ia64_expand_vecint_cmov (xops);
  return true;
}

/* The vectors LO and HI each contain N halves of a double-wide vector.
   Reassemble either the first N/2 or the second N/2 elements.  */

void
ia64_unpack_assemble (rtx out, rtx lo, rtx hi, bool highp)
{
  enum machine_mode vmode = GET_MODE (lo);
  unsigned int i, high, nelt = GET_MODE_NUNITS (vmode);
  struct expand_vec_perm_d d;
  bool ok;

  d.target = gen_lowpart (vmode, out);
  d.op0 = (TARGET_BIG_ENDIAN ? hi : lo);
  d.op1 = (TARGET_BIG_ENDIAN ? lo : hi);
  d.vmode = vmode;
  d.nelt = nelt;
  d.one_operand_p = false;
  d.testing_p = false;

  high = (highp ? nelt / 2 : 0);
  for (i = 0; i < nelt / 2; ++i)
    {
      d.perm[i * 2] = i + high;
      d.perm[i * 2 + 1] = i + high + nelt;
    }

  ok = ia64_expand_vec_perm_const_1 (&d);
  gcc_assert (ok);
}

/* Return a vector of the sign-extension of VEC.  */

static rtx
ia64_unpack_sign (rtx vec, bool unsignedp)
{
  enum machine_mode mode = GET_MODE (vec);
  rtx zero = CONST0_RTX (mode);

  if (unsignedp)
    return zero;
  else
    {
      rtx sign = gen_reg_rtx (mode);
      bool neg;

      neg = ia64_expand_vecint_compare (LT, mode, sign, vec, zero);
      gcc_assert (!neg);

      return sign;
    }
}

/* Emit an integral vector unpack operation.  */

void
ia64_expand_unpack (rtx operands[3], bool unsignedp, bool highp)
{
  rtx sign = ia64_unpack_sign (operands[1], unsignedp);
  ia64_unpack_assemble (operands[0], operands[1], sign, highp);
}

/* Emit an integral vector widening sum operations.  */

void
ia64_expand_widen_sum (rtx operands[3], bool unsignedp)
{
  enum machine_mode wmode;
  rtx l, h, t, sign;

  sign = ia64_unpack_sign (operands[1], unsignedp);

  wmode = GET_MODE (operands[0]);
  l = gen_reg_rtx (wmode);
  h = gen_reg_rtx (wmode);

  ia64_unpack_assemble (l, operands[1], sign, false);
  ia64_unpack_assemble (h, operands[1], sign, true);

  t = expand_binop (wmode, add_optab, l, operands[2], NULL, 0, OPTAB_DIRECT);
  t = expand_binop (wmode, add_optab, h, t, operands[0], 0, OPTAB_DIRECT);
  if (t != operands[0])
    emit_move_insn (operands[0], t);
}

/* Emit the appropriate sequence for a call.  */

void
ia64_expand_call (rtx retval, rtx addr, rtx nextarg ATTRIBUTE_UNUSED,
		  int sibcall_p)
{
  rtx insn, b0;

  addr = XEXP (addr, 0);
  addr = convert_memory_address (DImode, addr);
  b0 = gen_rtx_REG (DImode, R_BR (0));

  /* ??? Should do this for functions known to bind local too.  */
  if (TARGET_NO_PIC || TARGET_AUTO_PIC)
    {
      if (sibcall_p)
	insn = gen_sibcall_nogp (addr);
      else if (! retval)
	insn = gen_call_nogp (addr, b0);
      else
	insn = gen_call_value_nogp (retval, addr, b0);
      insn = emit_call_insn (insn);
    }
  else
    {
      if (sibcall_p)
	insn = gen_sibcall_gp (addr);
      else if (! retval)
	insn = gen_call_gp (addr, b0);
      else
	insn = gen_call_value_gp (retval, addr, b0);
      insn = emit_call_insn (insn);

      use_reg (&CALL_INSN_FUNCTION_USAGE (insn), pic_offset_table_rtx);
    }

  if (sibcall_p)
    use_reg (&CALL_INSN_FUNCTION_USAGE (insn), b0);

  if (TARGET_ABI_OPEN_VMS)
    use_reg (&CALL_INSN_FUNCTION_USAGE (insn),
	     gen_rtx_REG (DImode, GR_REG (25)));
}

static void
reg_emitted (enum ia64_frame_regs r)
{
  if (emitted_frame_related_regs[r] == 0)
    emitted_frame_related_regs[r] = current_frame_info.r[r];
  else
    gcc_assert (emitted_frame_related_regs[r] == current_frame_info.r[r]);
}

static int
get_reg (enum ia64_frame_regs r)
{
  reg_emitted (r);
  return current_frame_info.r[r];
}

static bool
is_emitted (int regno)
{
  unsigned int r;

  for (r = reg_fp; r < number_of_ia64_frame_regs; r++)
    if (emitted_frame_related_regs[r] == regno)
      return true;
  return false;
}

void
ia64_reload_gp (void)
{
  rtx tmp;

  if (current_frame_info.r[reg_save_gp])
    {
      tmp = gen_rtx_REG (DImode, get_reg (reg_save_gp));
    }
  else
    {
      HOST_WIDE_INT offset;
      rtx offset_r;

      offset = (current_frame_info.spill_cfa_off
	        + current_frame_info.spill_size);
      if (frame_pointer_needed)
        {
          tmp = hard_frame_pointer_rtx;
          offset = -offset;
        }
      else
        {
          tmp = stack_pointer_rtx;
          offset = current_frame_info.total_size - offset;
        }

      offset_r = GEN_INT (offset);
      if (satisfies_constraint_I (offset_r))
        emit_insn (gen_adddi3 (pic_offset_table_rtx, tmp, offset_r));
      else
        {
          emit_move_insn (pic_offset_table_rtx, offset_r);
          emit_insn (gen_adddi3 (pic_offset_table_rtx,
			         pic_offset_table_rtx, tmp));
        }

      tmp = gen_rtx_MEM (DImode, pic_offset_table_rtx);
    }

  emit_move_insn (pic_offset_table_rtx, tmp);
}

void
ia64_split_call (rtx retval, rtx addr, rtx retaddr, rtx scratch_r,
		 rtx scratch_b, int noreturn_p, int sibcall_p)
{
  rtx insn;
  bool is_desc = false;

  /* If we find we're calling through a register, then we're actually
     calling through a descriptor, so load up the values.  */
  if (REG_P (addr) && GR_REGNO_P (REGNO (addr)))
    {
      rtx tmp;
      bool addr_dead_p;

      /* ??? We are currently constrained to *not* use peep2, because
	 we can legitimately change the global lifetime of the GP
	 (in the form of killing where previously live).  This is
	 because a call through a descriptor doesn't use the previous
	 value of the GP, while a direct call does, and we do not
	 commit to either form until the split here.

	 That said, this means that we lack precise life info for
	 whether ADDR is dead after this call.  This is not terribly
	 important, since we can fix things up essentially for free
	 with the POST_DEC below, but it's nice to not use it when we
	 can immediately tell it's not necessary.  */
      addr_dead_p = ((noreturn_p || sibcall_p
		      || TEST_HARD_REG_BIT (regs_invalidated_by_call,
					    REGNO (addr)))
		     && !FUNCTION_ARG_REGNO_P (REGNO (addr)));

      /* Load the code address into scratch_b.  */
      tmp = gen_rtx_POST_INC (Pmode, addr);
      tmp = gen_rtx_MEM (Pmode, tmp);
      emit_move_insn (scratch_r, tmp);
      emit_move_insn (scratch_b, scratch_r);

      /* Load the GP address.  If ADDR is not dead here, then we must
	 revert the change made above via the POST_INCREMENT.  */
      if (!addr_dead_p)
	tmp = gen_rtx_POST_DEC (Pmode, addr);
      else
	tmp = addr;
      tmp = gen_rtx_MEM (Pmode, tmp);
      emit_move_insn (pic_offset_table_rtx, tmp);

      is_desc = true;
      addr = scratch_b;
    }

  if (sibcall_p)
    insn = gen_sibcall_nogp (addr);
  else if (retval)
    insn = gen_call_value_nogp (retval, addr, retaddr);
  else
    insn = gen_call_nogp (addr, retaddr);
  emit_call_insn (insn);

  if ((!TARGET_CONST_GP || is_desc) && !noreturn_p && !sibcall_p)
    ia64_reload_gp ();
}

/* Expand an atomic operation.  We want to perform MEM <CODE>= VAL atomically.

   This differs from the generic code in that we know about the zero-extending
   properties of cmpxchg, and the zero-extending requirements of ar.ccv.  We
   also know that ld.acq+cmpxchg.rel equals a full barrier.

   The loop we want to generate looks like

	cmp_reg = mem;
      label:
        old_reg = cmp_reg;
	new_reg = cmp_reg op val;
	cmp_reg = compare-and-swap(mem, old_reg, new_reg)
	if (cmp_reg != old_reg)
	  goto label;

   Note that we only do the plain load from memory once.  Subsequent
   iterations use the value loaded by the compare-and-swap pattern.  */

void
ia64_expand_atomic_op (enum rtx_code code, rtx mem, rtx val,
		       rtx old_dst, rtx new_dst, enum memmodel model)
{
  enum machine_mode mode = GET_MODE (mem);
  rtx old_reg, new_reg, cmp_reg, ar_ccv, label;
  enum insn_code icode;

  /* Special case for using fetchadd.  */
  if ((mode == SImode || mode == DImode)
      && (code == PLUS || code == MINUS)
      && fetchadd_operand (val, mode))
    {
      if (code == MINUS)
	val = GEN_INT (-INTVAL (val));

      if (!old_dst)
        old_dst = gen_reg_rtx (mode);

      switch (model)
	{
	case MEMMODEL_ACQ_REL:
	case MEMMODEL_SEQ_CST:
	  emit_insn (gen_memory_barrier ());
	  /* FALLTHRU */
	case MEMMODEL_RELAXED:
	case MEMMODEL_ACQUIRE:
	case MEMMODEL_CONSUME:
	  if (mode == SImode)
	    icode = CODE_FOR_fetchadd_acq_si;
	  else
	    icode = CODE_FOR_fetchadd_acq_di;
	  break;
	case MEMMODEL_RELEASE:
	  if (mode == SImode)
	    icode = CODE_FOR_fetchadd_rel_si;
	  else
	    icode = CODE_FOR_fetchadd_rel_di;
	  break;

	default:
	  gcc_unreachable ();
	}

      emit_insn (GEN_FCN (icode) (old_dst, mem, val));

      if (new_dst)
	{
	  new_reg = expand_simple_binop (mode, PLUS, old_dst, val, new_dst,
					 true, OPTAB_WIDEN);
	  if (new_reg != new_dst)
	    emit_move_insn (new_dst, new_reg);
	}
      return;
    }

  /* Because of the volatile mem read, we get an ld.acq, which is the
     front half of the full barrier.  The end half is the cmpxchg.rel.
     For relaxed and release memory models, we don't need this.  But we
     also don't bother trying to prevent it either.  */
  gcc_assert (model == MEMMODEL_RELAXED
	      || model == MEMMODEL_RELEASE
	      || MEM_VOLATILE_P (mem));

  old_reg = gen_reg_rtx (DImode);
  cmp_reg = gen_reg_rtx (DImode);
  label = gen_label_rtx ();

  if (mode != DImode)
    {
      val = simplify_gen_subreg (DImode, val, mode, 0);
      emit_insn (gen_extend_insn (cmp_reg, mem, DImode, mode, 1));
    }
  else
    emit_move_insn (cmp_reg, mem);

  emit_label (label);

  ar_ccv = gen_rtx_REG (DImode, AR_CCV_REGNUM);
  emit_move_insn (old_reg, cmp_reg);
  emit_move_insn (ar_ccv, cmp_reg);

  if (old_dst)
    emit_move_insn (old_dst, gen_lowpart (mode, cmp_reg));

  new_reg = cmp_reg;
  if (code == NOT)
    {
      new_reg = expand_simple_binop (DImode, AND, new_reg, val, NULL_RTX,
				     true, OPTAB_DIRECT);
      new_reg = expand_simple_unop (DImode, code, new_reg, NULL_RTX, true);
    }
  else
    new_reg = expand_simple_binop (DImode, code, new_reg, val, NULL_RTX,
				   true, OPTAB_DIRECT);

  if (mode != DImode)
    new_reg = gen_lowpart (mode, new_reg);
  if (new_dst)
    emit_move_insn (new_dst, new_reg);

  switch (model)
    {
    case MEMMODEL_RELAXED:
    case MEMMODEL_ACQUIRE:
    case MEMMODEL_CONSUME:
      switch (mode)
	{
	case QImode: icode = CODE_FOR_cmpxchg_acq_qi;  break;
	case HImode: icode = CODE_FOR_cmpxchg_acq_hi;  break;
	case SImode: icode = CODE_FOR_cmpxchg_acq_si;  break;
	case DImode: icode = CODE_FOR_cmpxchg_acq_di;  break;
	default:
	  gcc_unreachable ();
	}
      break;

    case MEMMODEL_RELEASE:
    case MEMMODEL_ACQ_REL:
    case MEMMODEL_SEQ_CST:
      switch (mode)
	{
	case QImode: icode = CODE_FOR_cmpxchg_rel_qi;  break;
	case HImode: icode = CODE_FOR_cmpxchg_rel_hi;  break;
	case SImode: icode = CODE_FOR_cmpxchg_rel_si;  break;
	case DImode: icode = CODE_FOR_cmpxchg_rel_di;  break;
	default:
	  gcc_unreachable ();
	}
      break;

    default:
      gcc_unreachable ();
    }

  emit_insn (GEN_FCN (icode) (cmp_reg, mem, ar_ccv, new_reg));

  emit_cmp_and_jump_insns (cmp_reg, old_reg, NE, NULL, DImode, true, label);
}

/* Begin the assembly file.  */

static void
ia64_file_start (void)
{
  default_file_start ();
  emit_safe_across_calls ();
}

void
emit_safe_across_calls (void)
{
  unsigned int rs, re;
  int out_state;

  rs = 1;
  out_state = 0;
  while (1)
    {
      while (rs < 64 && call_used_regs[PR_REG (rs)])
	rs++;
      if (rs >= 64)
	break;
      for (re = rs + 1; re < 64 && ! call_used_regs[PR_REG (re)]; re++)
	continue;
      if (out_state == 0)
	{
	  fputs ("\t.pred.safe_across_calls ", asm_out_file);
	  out_state = 1;
	}
      else
	fputc (',', asm_out_file);
      if (re == rs + 1)
	fprintf (asm_out_file, "p%u", rs);
      else
	fprintf (asm_out_file, "p%u-p%u", rs, re - 1);
      rs = re + 1;
    }
  if (out_state)
    fputc ('\n', asm_out_file);
}

/* Globalize a declaration.  */

static void
ia64_globalize_decl_name (FILE * stream, tree decl)
{
  const char *name = XSTR (XEXP (DECL_RTL (decl), 0), 0);
  tree version_attr = lookup_attribute ("version_id", DECL_ATTRIBUTES (decl));
  if (version_attr)
    {
      tree v = TREE_VALUE (TREE_VALUE (version_attr));
      const char *p = TREE_STRING_POINTER (v);
      fprintf (stream, "\t.alias %s#, \"%s{%s}\"\n", name, name, p);
    }
  targetm.asm_out.globalize_label (stream, name);
  if (TREE_CODE (decl) == FUNCTION_DECL)
    ASM_OUTPUT_TYPE_DIRECTIVE (stream, name, "function");
}

/* Helper function for ia64_compute_frame_size: find an appropriate general
   register to spill some special register to.  SPECIAL_SPILL_MASK contains
   bits in GR0 to GR31 that have already been allocated by this routine.
   TRY_LOCALS is true if we should attempt to locate a local regnum.  */

static int
find_gr_spill (enum ia64_frame_regs r, int try_locals)
{
  int regno;

  if (emitted_frame_related_regs[r] != 0)
    {
      regno = emitted_frame_related_regs[r];
      if (regno >= LOC_REG (0) && regno < LOC_REG (80 - frame_pointer_needed)
	  && current_frame_info.n_local_regs < regno - LOC_REG (0) + 1)
        current_frame_info.n_local_regs = regno - LOC_REG (0) + 1;
      else if (crtl->is_leaf
               && regno >= GR_REG (1) && regno <= GR_REG (31))
        current_frame_info.gr_used_mask |= 1 << regno;

      return regno;
    }

  /* If this is a leaf function, first try an otherwise unused
     call-clobbered register.  */
  if (crtl->is_leaf)
    {
      for (regno = GR_REG (1); regno <= GR_REG (31); regno++)
	if (! df_regs_ever_live_p (regno)
	    && call_used_regs[regno]
	    && ! fixed_regs[regno]
	    && ! global_regs[regno]
	    && ((current_frame_info.gr_used_mask >> regno) & 1) == 0
            && ! is_emitted (regno))
	  {
	    current_frame_info.gr_used_mask |= 1 << regno;
	    return regno;
	  }
    }

  if (try_locals)
    {
      regno = current_frame_info.n_local_regs;
      /* If there is a frame pointer, then we can't use loc79, because
	 that is HARD_FRAME_POINTER_REGNUM.  In particular, see the
	 reg_name switching code in ia64_expand_prologue.  */
      while (regno < (80 - frame_pointer_needed))
	if (! is_emitted (LOC_REG (regno++)))
	  {
	    current_frame_info.n_local_regs = regno;
	    return LOC_REG (regno - 1);
	  }
    }

  /* Failed to find a general register to spill to.  Must use stack.  */
  return 0;
}

/* In order to make for nice schedules, we try to allocate every temporary
   to a different register.  We must of course stay away from call-saved,
   fixed, and global registers.  We must also stay away from registers
   allocated in current_frame_info.gr_used_mask, since those include regs
   used all through the prologue.

   Any register allocated here must be used immediately.  The idea is to
   aid scheduling, not to solve data flow problems.  */

static int last_scratch_gr_reg;

static int
next_scratch_gr_reg (void)
{
  int i, regno;

  for (i = 0; i < 32; ++i)
    {
      regno = (last_scratch_gr_reg + i + 1) & 31;
      if (call_used_regs[regno]
	  && ! fixed_regs[regno]
	  && ! global_regs[regno]
	  && ((current_frame_info.gr_used_mask >> regno) & 1) == 0)
	{
	  last_scratch_gr_reg = regno;
	  return regno;
	}
    }

  /* There must be _something_ available.  */
  gcc_unreachable ();
}

/* Helper function for ia64_compute_frame_size, called through
   diddle_return_value.  Mark REG in current_frame_info.gr_used_mask.  */

static void
mark_reg_gr_used_mask (rtx reg, void *data ATTRIBUTE_UNUSED)
{
  unsigned int regno = REGNO (reg);
  if (regno < 32)
    {
      unsigned int i, n = hard_regno_nregs[regno][GET_MODE (reg)];
      for (i = 0; i < n; ++i)
	current_frame_info.gr_used_mask |= 1 << (regno + i);
    }
}


/* Returns the number of bytes offset between the frame pointer and the stack
   pointer for the current function.  SIZE is the number of bytes of space
   needed for local variables.  */

static void
ia64_compute_frame_size (HOST_WIDE_INT size)
{
  HOST_WIDE_INT total_size;
  HOST_WIDE_INT spill_size = 0;
  HOST_WIDE_INT extra_spill_size = 0;
  HOST_WIDE_INT pretend_args_size;
  HARD_REG_SET mask;
  int n_spilled = 0;
  int spilled_gr_p = 0;
  int spilled_fr_p = 0;
  unsigned int regno;
  int min_regno;
  int max_regno;
  int i;

  if (current_frame_info.initialized)
    return;

  memset (&current_frame_info, 0, sizeof current_frame_info);
  CLEAR_HARD_REG_SET (mask);

  /* Don't allocate scratches to the return register.  */
  diddle_return_value (mark_reg_gr_used_mask, NULL);

  /* Don't allocate scratches to the EH scratch registers.  */
  if (cfun->machine->ia64_eh_epilogue_sp)
    mark_reg_gr_used_mask (cfun->machine->ia64_eh_epilogue_sp, NULL);
  if (cfun->machine->ia64_eh_epilogue_bsp)
    mark_reg_gr_used_mask (cfun->machine->ia64_eh_epilogue_bsp, NULL);

  /* Static stack checking uses r2 and r3.  */
  if (flag_stack_check == STATIC_BUILTIN_STACK_CHECK)
    current_frame_info.gr_used_mask |= 0xc;

  /* Find the size of the register stack frame.  We have only 80 local
     registers, because we reserve 8 for the inputs and 8 for the
     outputs.  */

  /* Skip HARD_FRAME_POINTER_REGNUM (loc79) when frame_pointer_needed,
     since we'll be adjusting that down later.  */
  regno = LOC_REG (78) + ! frame_pointer_needed;
  for (; regno >= LOC_REG (0); regno--)
    if (df_regs_ever_live_p (regno) && !is_emitted (regno))
      break;
  current_frame_info.n_local_regs = regno - LOC_REG (0) + 1;

  /* For functions marked with the syscall_linkage attribute, we must mark
     all eight input registers as in use, so that locals aren't visible to
     the caller.  */

  if (cfun->machine->n_varargs > 0
      || lookup_attribute ("syscall_linkage",
			   TYPE_ATTRIBUTES (TREE_TYPE (current_function_decl))))
    current_frame_info.n_input_regs = 8;
  else
    {
      for (regno = IN_REG (7); regno >= IN_REG (0); regno--)
	if (df_regs_ever_live_p (regno))
	  break;
      current_frame_info.n_input_regs = regno - IN_REG (0) + 1;
    }

  for (regno = OUT_REG (7); regno >= OUT_REG (0); regno--)
    if (df_regs_ever_live_p (regno))
      break;
  i = regno - OUT_REG (0) + 1;

#ifndef PROFILE_HOOK
  /* When -p profiling, we need one output register for the mcount argument.
     Likewise for -a profiling for the bb_init_func argument.  For -ax
     profiling, we need two output registers for the two bb_init_trace_func
     arguments.  */
  if (crtl->profile)
    i = MAX (i, 1);
#endif
  current_frame_info.n_output_regs = i;

  /* ??? No rotating register support yet.  */
  current_frame_info.n_rotate_regs = 0;

  /* Discover which registers need spilling, and how much room that
     will take.  Begin with floating point and general registers,
     which will always wind up on the stack.  */

  for (regno = FR_REG (2); regno <= FR_REG (127); regno++)
    if (df_regs_ever_live_p (regno) && ! call_used_regs[regno])
      {
	SET_HARD_REG_BIT (mask, regno);
	spill_size += 16;
	n_spilled += 1;
	spilled_fr_p = 1;
      }

  for (regno = GR_REG (1); regno <= GR_REG (31); regno++)
    if (df_regs_ever_live_p (regno) && ! call_used_regs[regno])
      {
	SET_HARD_REG_BIT (mask, regno);
	spill_size += 8;
	n_spilled += 1;
	spilled_gr_p = 1;
      }

  for (regno = BR_REG (1); regno <= BR_REG (7); regno++)
    if (df_regs_ever_live_p (regno) && ! call_used_regs[regno])
      {
	SET_HARD_REG_BIT (mask, regno);
	spill_size += 8;
	n_spilled += 1;
      }

  /* Now come all special registers that might get saved in other
     general registers.  */

  if (frame_pointer_needed)
    {
      current_frame_info.r[reg_fp] = find_gr_spill (reg_fp, 1);
      /* If we did not get a register, then we take LOC79.  This is guaranteed
	 to be free, even if regs_ever_live is already set, because this is
	 HARD_FRAME_POINTER_REGNUM.  This requires incrementing n_local_regs,
	 as we don't count loc79 above.  */
      if (current_frame_info.r[reg_fp] == 0)
	{
	  current_frame_info.r[reg_fp] = LOC_REG (79);
	  current_frame_info.n_local_regs = LOC_REG (79) - LOC_REG (0) + 1;
	}
    }

  if (! crtl->is_leaf)
    {
      /* Emit a save of BR0 if we call other functions.  Do this even
	 if this function doesn't return, as EH depends on this to be
	 able to unwind the stack.  */
      SET_HARD_REG_BIT (mask, BR_REG (0));

      current_frame_info.r[reg_save_b0] = find_gr_spill (reg_save_b0, 1);
      if (current_frame_info.r[reg_save_b0] == 0)
	{
	  extra_spill_size += 8;
	  n_spilled += 1;
	}

      /* Similarly for ar.pfs.  */
      SET_HARD_REG_BIT (mask, AR_PFS_REGNUM);
      current_frame_info.r[reg_save_ar_pfs] = find_gr_spill (reg_save_ar_pfs, 1);
      if (current_frame_info.r[reg_save_ar_pfs] == 0)
	{
	  extra_spill_size += 8;
	  n_spilled += 1;
	}

      /* Similarly for gp.  Note that if we're calling setjmp, the stacked
	 registers are clobbered, so we fall back to the stack.  */
      current_frame_info.r[reg_save_gp]
	= (cfun->calls_setjmp ? 0 : find_gr_spill (reg_save_gp, 1));
      if (current_frame_info.r[reg_save_gp] == 0)
	{
	  SET_HARD_REG_BIT (mask, GR_REG (1));
	  spill_size += 8;
	  n_spilled += 1;
	}
    }
  else
    {
      if (df_regs_ever_live_p (BR_REG (0)) && ! call_used_regs[BR_REG (0)])
	{
	  SET_HARD_REG_BIT (mask, BR_REG (0));
	  extra_spill_size += 8;
	  n_spilled += 1;
	}

      if (df_regs_ever_live_p (AR_PFS_REGNUM))
	{
	  SET_HARD_REG_BIT (mask, AR_PFS_REGNUM);
 	  current_frame_info.r[reg_save_ar_pfs] 
            = find_gr_spill (reg_save_ar_pfs, 1);
	  if (current_frame_info.r[reg_save_ar_pfs] == 0)
	    {
	      extra_spill_size += 8;
	      n_spilled += 1;
	    }
	}
    }

  /* Unwind descriptor hackery: things are most efficient if we allocate
     consecutive GR save registers for RP, PFS, FP in that order. However,
     it is absolutely critical that FP get the only hard register that's
     guaranteed to be free, so we allocated it first.  If all three did
     happen to be allocated hard regs, and are consecutive, rearrange them
     into the preferred order now.  
     
     If we have already emitted code for any of those registers,
     then it's already too late to change.  */
  min_regno = MIN (current_frame_info.r[reg_fp],
		   MIN (current_frame_info.r[reg_save_b0],
			current_frame_info.r[reg_save_ar_pfs]));
  max_regno = MAX (current_frame_info.r[reg_fp],
		   MAX (current_frame_info.r[reg_save_b0],
			current_frame_info.r[reg_save_ar_pfs]));
  if (min_regno > 0
      && min_regno + 2 == max_regno
      && (current_frame_info.r[reg_fp] == min_regno + 1
	  || current_frame_info.r[reg_save_b0] == min_regno + 1
	  || current_frame_info.r[reg_save_ar_pfs] == min_regno + 1)
      && (emitted_frame_related_regs[reg_save_b0] == 0
	  || emitted_frame_related_regs[reg_save_b0] == min_regno)
      && (emitted_frame_related_regs[reg_save_ar_pfs] == 0
	  || emitted_frame_related_regs[reg_save_ar_pfs] == min_regno + 1)
      && (emitted_frame_related_regs[reg_fp] == 0
	  || emitted_frame_related_regs[reg_fp] == min_regno + 2))
    {
      current_frame_info.r[reg_save_b0] = min_regno;
      current_frame_info.r[reg_save_ar_pfs] = min_regno + 1;
      current_frame_info.r[reg_fp] = min_regno + 2;
    }

  /* See if we need to store the predicate register block.  */
  for (regno = PR_REG (0); regno <= PR_REG (63); regno++)
    if (df_regs_ever_live_p (regno) && ! call_used_regs[regno])
      break;
  if (regno <= PR_REG (63))
    {
      SET_HARD_REG_BIT (mask, PR_REG (0));
      current_frame_info.r[reg_save_pr] = find_gr_spill (reg_save_pr, 1);
      if (current_frame_info.r[reg_save_pr] == 0)
	{
	  extra_spill_size += 8;
	  n_spilled += 1;
	}

      /* ??? Mark them all as used so that register renaming and such
	 are free to use them.  */
      for (regno = PR_REG (0); regno <= PR_REG (63); regno++)
	df_set_regs_ever_live (regno, true);
    }

  /* If we're forced to use st8.spill, we're forced to save and restore
     ar.unat as well.  The check for existing liveness allows inline asm
     to touch ar.unat.  */
  if (spilled_gr_p || cfun->machine->n_varargs
      || df_regs_ever_live_p (AR_UNAT_REGNUM))
    {
      df_set_regs_ever_live (AR_UNAT_REGNUM, true);
      SET_HARD_REG_BIT (mask, AR_UNAT_REGNUM);
      current_frame_info.r[reg_save_ar_unat] 
        = find_gr_spill (reg_save_ar_unat, spill_size == 0);
      if (current_frame_info.r[reg_save_ar_unat] == 0)
	{
	  extra_spill_size += 8;
	  n_spilled += 1;
	}
    }

  if (df_regs_ever_live_p (AR_LC_REGNUM))
    {
      SET_HARD_REG_BIT (mask, AR_LC_REGNUM);
      current_frame_info.r[reg_save_ar_lc] 
        = find_gr_spill (reg_save_ar_lc, spill_size == 0);
      if (current_frame_info.r[reg_save_ar_lc] == 0)
	{
	  extra_spill_size += 8;
	  n_spilled += 1;
	}
    }

  /* If we have an odd number of words of pretend arguments written to
     the stack, then the FR save area will be unaligned.  We round the
     size of this area up to keep things 16 byte aligned.  */
  if (spilled_fr_p)
    pretend_args_size = IA64_STACK_ALIGN (crtl->args.pretend_args_size);
  else
    pretend_args_size = crtl->args.pretend_args_size;

  total_size = (spill_size + extra_spill_size + size + pretend_args_size
		+ crtl->outgoing_args_size);
  total_size = IA64_STACK_ALIGN (total_size);

  /* We always use the 16-byte scratch area provided by the caller, but
     if we are a leaf function, there's no one to which we need to provide
     a scratch area.  However, if the function allocates dynamic stack space,
     the dynamic offset is computed early and contains STACK_POINTER_OFFSET,
     so we need to cope.  */
  if (crtl->is_leaf && !cfun->calls_alloca)
    total_size = MAX (0, total_size - 16);

  current_frame_info.total_size = total_size;
  current_frame_info.spill_cfa_off = pretend_args_size - 16;
  current_frame_info.spill_size = spill_size;
  current_frame_info.extra_spill_size = extra_spill_size;
  COPY_HARD_REG_SET (current_frame_info.mask, mask);
  current_frame_info.n_spilled = n_spilled;
  current_frame_info.initialized = reload_completed;
}

/* Worker function for TARGET_CAN_ELIMINATE.  */

bool
ia64_can_eliminate (const int from ATTRIBUTE_UNUSED, const int to)
{
  return (to == BR_REG (0) ? crtl->is_leaf : true);
}

/* Compute the initial difference between the specified pair of registers.  */

HOST_WIDE_INT
ia64_initial_elimination_offset (int from, int to)
{
  HOST_WIDE_INT offset;

  ia64_compute_frame_size (get_frame_size ());
  switch (from)
    {
    case FRAME_POINTER_REGNUM:
      switch (to)
	{
	case HARD_FRAME_POINTER_REGNUM:
	  offset = -current_frame_info.total_size;
	  if (!crtl->is_leaf || cfun->calls_alloca)
	    offset += 16 + crtl->outgoing_args_size;
	  break;

	case STACK_POINTER_REGNUM:
	  offset = 0;
	  if (!crtl->is_leaf || cfun->calls_alloca)
	    offset += 16 + crtl->outgoing_args_size;
	  break;

	default:
	  gcc_unreachable ();
	}
      break;

    case ARG_POINTER_REGNUM:
      /* Arguments start above the 16 byte save area, unless stdarg
	 in which case we store through the 16 byte save area.  */
      switch (to)
	{
	case HARD_FRAME_POINTER_REGNUM:
	  offset = 16 - crtl->args.pretend_args_size;
	  break;

	case STACK_POINTER_REGNUM:
	  offset = (current_frame_info.total_size
		    + 16 - crtl->args.pretend_args_size);
	  break;

	default:
	  gcc_unreachable ();
	}
      break;

    default:
      gcc_unreachable ();
    }

  return offset;
}

/* If there are more than a trivial number of register spills, we use
   two interleaved iterators so that we can get two memory references
   per insn group.

   In order to simplify things in the prologue and epilogue expanders,
   we use helper functions to fix up the memory references after the
   fact with the appropriate offsets to a POST_MODIFY memory mode.
   The following data structure tracks the state of the two iterators
   while insns are being emitted.  */

struct spill_fill_data
{
  rtx init_after;		/* point at which to emit initializations */
  rtx init_reg[2];		/* initial base register */
  rtx iter_reg[2];		/* the iterator registers */
  rtx *prev_addr[2];		/* address of last memory use */
  rtx prev_insn[2];		/* the insn corresponding to prev_addr */
  HOST_WIDE_INT prev_off[2];	/* last offset */
  int n_iter;			/* number of iterators in use */
  int next_iter;		/* next iterator to use */
  unsigned int save_gr_used_mask;
};

static struct spill_fill_data spill_fill_data;

static void
setup_spill_pointers (int n_spills, rtx init_reg, HOST_WIDE_INT cfa_off)
{
  int i;

  spill_fill_data.init_after = get_last_insn ();
  spill_fill_data.init_reg[0] = init_reg;
  spill_fill_data.init_reg[1] = init_reg;
  spill_fill_data.prev_addr[0] = NULL;
  spill_fill_data.prev_addr[1] = NULL;
  spill_fill_data.prev_insn[0] = NULL;
  spill_fill_data.prev_insn[1] = NULL;
  spill_fill_data.prev_off[0] = cfa_off;
  spill_fill_data.prev_off[1] = cfa_off;
  spill_fill_data.next_iter = 0;
  spill_fill_data.save_gr_used_mask = current_frame_info.gr_used_mask;

  spill_fill_data.n_iter = 1 + (n_spills > 2);
  for (i = 0; i < spill_fill_data.n_iter; ++i)
    {
      int regno = next_scratch_gr_reg ();
      spill_fill_data.iter_reg[i] = gen_rtx_REG (DImode, regno);
      current_frame_info.gr_used_mask |= 1 << regno;
    }
}

static void
finish_spill_pointers (void)
{
  current_frame_info.gr_used_mask = spill_fill_data.save_gr_used_mask;
}

static rtx
spill_restore_mem (rtx reg, HOST_WIDE_INT cfa_off)
{
  int iter = spill_fill_data.next_iter;
  HOST_WIDE_INT disp = spill_fill_data.prev_off[iter] - cfa_off;
  rtx disp_rtx = GEN_INT (disp);
  rtx mem;

  if (spill_fill_data.prev_addr[iter])
    {
      if (satisfies_constraint_N (disp_rtx))
	{
	  *spill_fill_data.prev_addr[iter]
	    = gen_rtx_POST_MODIFY (DImode, spill_fill_data.iter_reg[iter],
				   gen_rtx_PLUS (DImode,
						 spill_fill_data.iter_reg[iter],
						 disp_rtx));
	  add_reg_note (spill_fill_data.prev_insn[iter],
			REG_INC, spill_fill_data.iter_reg[iter]);
	}
      else
	{
	  /* ??? Could use register post_modify for loads.  */
	  if (!satisfies_constraint_I (disp_rtx))
	    {
	      rtx tmp = gen_rtx_REG (DImode, next_scratch_gr_reg ());
	      emit_move_insn (tmp, disp_rtx);
	      disp_rtx = tmp;
	    }
	  emit_insn (gen_adddi3 (spill_fill_data.iter_reg[iter],
				 spill_fill_data.iter_reg[iter], disp_rtx));
	}
    }
  /* Micro-optimization: if we've created a frame pointer, it's at
     CFA 0, which may allow the real iterator to be initialized lower,
     slightly increasing parallelism.  Also, if there are few saves
     it may eliminate the iterator entirely.  */
  else if (disp == 0
	   && spill_fill_data.init_reg[iter] == stack_pointer_rtx
	   && frame_pointer_needed)
    {
      mem = gen_rtx_MEM (GET_MODE (reg), hard_frame_pointer_rtx);
      set_mem_alias_set (mem, get_varargs_alias_set ());
      return mem;
    }
  else
    {
      rtx seq, insn;

      if (disp == 0)
	seq = gen_movdi (spill_fill_data.iter_reg[iter],
			 spill_fill_data.init_reg[iter]);
      else
	{
	  start_sequence ();

	  if (!satisfies_constraint_I (disp_rtx))
	    {
	      rtx tmp = gen_rtx_REG (DImode, next_scratch_gr_reg ());
	      emit_move_insn (tmp, disp_rtx);
	      disp_rtx = tmp;
	    }

	  emit_insn (gen_adddi3 (spill_fill_data.iter_reg[iter],
				 spill_fill_data.init_reg[iter],
				 disp_rtx));

	  seq = get_insns ();
	  end_sequence ();
	}

      /* Careful for being the first insn in a sequence.  */
      if (spill_fill_data.init_after)
	insn = emit_insn_after (seq, spill_fill_data.init_after);
      else
	{
	  rtx first = get_insns ();
	  if (first)
	    insn = emit_insn_before (seq, first);
	  else
	    insn = emit_insn (seq);
	}
      spill_fill_data.init_after = insn;
    }

  mem = gen_rtx_MEM (GET_MODE (reg), spill_fill_data.iter_reg[iter]);

  /* ??? Not all of the spills are for varargs, but some of them are.
     The rest of the spills belong in an alias set of their own.  But
     it doesn't actually hurt to include them here.  */
  set_mem_alias_set (mem, get_varargs_alias_set ());

  spill_fill_data.prev_addr[iter] = &XEXP (mem, 0);
  spill_fill_data.prev_off[iter] = cfa_off;

  if (++iter >= spill_fill_data.n_iter)
    iter = 0;
  spill_fill_data.next_iter = iter;

  return mem;
}

static void
do_spill (rtx (*move_fn) (rtx, rtx, rtx), rtx reg, HOST_WIDE_INT cfa_off,
	  rtx frame_reg)
{
  int iter = spill_fill_data.next_iter;
  rtx mem, insn;

  mem = spill_restore_mem (reg, cfa_off);
  insn = emit_insn ((*move_fn) (mem, reg, GEN_INT (cfa_off)));
  spill_fill_data.prev_insn[iter] = insn;

  if (frame_reg)
    {
      rtx base;
      HOST_WIDE_INT off;

      RTX_FRAME_RELATED_P (insn) = 1;

      /* Don't even pretend that the unwind code can intuit its way
	 through a pair of interleaved post_modify iterators.  Just
	 provide the correct answer.  */

      if (frame_pointer_needed)
	{
	  base = hard_frame_pointer_rtx;
	  off = - cfa_off;
	}
      else
	{
	  base = stack_pointer_rtx;
	  off = current_frame_info.total_size - cfa_off;
	}

      add_reg_note (insn, REG_CFA_OFFSET,
		    gen_rtx_SET (VOIDmode,
				 gen_rtx_MEM (GET_MODE (reg),
					      plus_constant (Pmode,
							     base, off)),
				 frame_reg));
    }
}

static void
do_restore (rtx (*move_fn) (rtx, rtx, rtx), rtx reg, HOST_WIDE_INT cfa_off)
{
  int iter = spill_fill_data.next_iter;
  rtx insn;

  insn = emit_insn ((*move_fn) (reg, spill_restore_mem (reg, cfa_off),
				GEN_INT (cfa_off)));
  spill_fill_data.prev_insn[iter] = insn;
}

/* Wrapper functions that discards the CONST_INT spill offset.  These
   exist so that we can give gr_spill/gr_fill the offset they need and
   use a consistent function interface.  */

static rtx
gen_movdi_x (rtx dest, rtx src, rtx offset ATTRIBUTE_UNUSED)
{
  return gen_movdi (dest, src);
}

static rtx
gen_fr_spill_x (rtx dest, rtx src, rtx offset ATTRIBUTE_UNUSED)
{
  return gen_fr_spill (dest, src);
}

static rtx
gen_fr_restore_x (rtx dest, rtx src, rtx offset ATTRIBUTE_UNUSED)
{
  return gen_fr_restore (dest, src);
}

#define PROBE_INTERVAL (1 << STACK_CHECK_PROBE_INTERVAL_EXP)

/* See Table 6.2 of the IA-64 Software Developer Manual, Volume 2.  */
#define BACKING_STORE_SIZE(N) ((N) > 0 ? ((N) + (N)/63 + 1) * 8 : 0)

/* Emit code to probe a range of stack addresses from FIRST to FIRST+SIZE,
   inclusive.  These are offsets from the current stack pointer.  BS_SIZE
   is the size of the backing store.  ??? This clobbers r2 and r3.  */

static void
ia64_emit_probe_stack_range (HOST_WIDE_INT first, HOST_WIDE_INT size,
			     int bs_size)
{
  rtx r2 = gen_rtx_REG (Pmode, GR_REG (2));
  rtx r3 = gen_rtx_REG (Pmode, GR_REG (3));
  rtx p6 = gen_rtx_REG (BImode, PR_REG (6));

  /* On the IA-64 there is a second stack in memory, namely the Backing Store
     of the Register Stack Engine.  We also need to probe it after checking
     that the 2 stacks don't overlap.  */
  emit_insn (gen_bsp_value (r3));
  emit_move_insn (r2, GEN_INT (-(first + size)));

  /* Compare current value of BSP and SP registers.  */
  emit_insn (gen_rtx_SET (VOIDmode, p6,
			  gen_rtx_fmt_ee (LTU, BImode,
					  r3, stack_pointer_rtx)));

  /* Compute the address of the probe for the Backing Store (which grows
     towards higher addresses).  We probe only at the first offset of
     the next page because some OS (eg Linux/ia64) only extend the
     backing store when this specific address is hit (but generate a SEGV
     on other address).  Page size is the worst case (4KB).  The reserve
     size is at least 4096 - (96 + 2) * 8 = 3312 bytes, which is enough.
     Also compute the address of the last probe for the memory stack
     (which grows towards lower addresses).  */
  emit_insn (gen_rtx_SET (VOIDmode, r3, plus_constant (Pmode, r3, 4095)));
  emit_insn (gen_rtx_SET (VOIDmode, r2,
			  gen_rtx_PLUS (Pmode, stack_pointer_rtx, r2)));

  /* Compare them and raise SEGV if the former has topped the latter.  */
  emit_insn (gen_rtx_COND_EXEC (VOIDmode,
				gen_rtx_fmt_ee (NE, VOIDmode, p6, const0_rtx),
				gen_rtx_SET (VOIDmode, p6,
					     gen_rtx_fmt_ee (GEU, BImode,
							     r3, r2))));
  emit_insn (gen_rtx_SET (VOIDmode,
			  gen_rtx_ZERO_EXTRACT (DImode, r3, GEN_INT (12),
						const0_rtx),
			  const0_rtx));
  emit_insn (gen_rtx_COND_EXEC (VOIDmode,
				gen_rtx_fmt_ee (NE, VOIDmode, p6, const0_rtx),
				gen_rtx_TRAP_IF (VOIDmode, const1_rtx,
						 GEN_INT (11))));

  /* Probe the Backing Store if necessary.  */
  if (bs_size > 0)
    emit_stack_probe (r3);

  /* Probe the memory stack if necessary.  */
  if (size == 0)
    ;

  /* See if we have a constant small number of probes to generate.  If so,
     that's the easy case.  */
  else if (size <= PROBE_INTERVAL)
    emit_stack_probe (r2);

  /* The run-time loop is made up of 8 insns in the generic case while this
     compile-time loop is made up of 5+2*(n-2) insns for n # of intervals.  */
  else if (size <= 4 * PROBE_INTERVAL)
    {
      HOST_WIDE_INT i;

      emit_move_insn (r2, GEN_INT (-(first + PROBE_INTERVAL)));
      emit_insn (gen_rtx_SET (VOIDmode, r2,
			      gen_rtx_PLUS (Pmode, stack_pointer_rtx, r2)));
      emit_stack_probe (r2);

      /* Probe at FIRST + N * PROBE_INTERVAL for values of N from 2 until
	 it exceeds SIZE.  If only two probes are needed, this will not
	 generate any code.  Then probe at FIRST + SIZE.  */
      for (i = 2 * PROBE_INTERVAL; i < size; i += PROBE_INTERVAL)
	{
	  emit_insn (gen_rtx_SET (VOIDmode, r2,
				  plus_constant (Pmode, r2, -PROBE_INTERVAL)));
	  emit_stack_probe (r2);
	}

      emit_insn (gen_rtx_SET (VOIDmode, r2,
			      plus_constant (Pmode, r2,
					     (i - PROBE_INTERVAL) - size)));
      emit_stack_probe (r2);
    }

  /* Otherwise, do the same as above, but in a loop.  Note that we must be
     extra careful with variables wrapping around because we might be at
     the very top (or the very bottom) of the address space and we have
     to be able to handle this case properly; in particular, we use an
     equality test for the loop condition.  */
  else
    {
      HOST_WIDE_INT rounded_size;

      emit_move_insn (r2, GEN_INT (-first));


      /* Step 1: round SIZE to the previous multiple of the interval.  */

      rounded_size = size & -PROBE_INTERVAL;


      /* Step 2: compute initial and final value of the loop counter.  */

      /* TEST_ADDR = SP + FIRST.  */
      emit_insn (gen_rtx_SET (VOIDmode, r2,
			      gen_rtx_PLUS (Pmode, stack_pointer_rtx, r2)));

      /* LAST_ADDR = SP + FIRST + ROUNDED_SIZE.  */
      if (rounded_size > (1 << 21))
	{
	  emit_move_insn (r3, GEN_INT (-rounded_size));
	  emit_insn (gen_rtx_SET (VOIDmode, r3, gen_rtx_PLUS (Pmode, r2, r3)));
	}
      else
        emit_insn (gen_rtx_SET (VOIDmode, r3,
				gen_rtx_PLUS (Pmode, r2,
					      GEN_INT (-rounded_size))));


      /* Step 3: the loop

	 while (TEST_ADDR != LAST_ADDR)
	   {
	     TEST_ADDR = TEST_ADDR + PROBE_INTERVAL
	     probe at TEST_ADDR
	   }

	 probes at FIRST + N * PROBE_INTERVAL for values of N from 1
	 until it is equal to ROUNDED_SIZE.  */

      emit_insn (gen_probe_stack_range (r2, r2, r3));


      /* Step 4: probe at FIRST + SIZE if we cannot assert at compile-time
	 that SIZE is equal to ROUNDED_SIZE.  */

      /* TEMP = SIZE - ROUNDED_SIZE.  */
      if (size != rounded_size)
	{
	  emit_insn (gen_rtx_SET (VOIDmode, r2,
				  plus_constant (Pmode, r2,
						 rounded_size - size)));
	  emit_stack_probe (r2);
	}
    }

  /* Make sure nothing is scheduled before we are done.  */
  emit_insn (gen_blockage ());
}

/* Probe a range of stack addresses from REG1 to REG2 inclusive.  These are
   absolute addresses.  */

const char *
output_probe_stack_range (rtx reg1, rtx reg2)
{
  static int labelno = 0;
  char loop_lab[32], end_lab[32];
  rtx xops[3];

  ASM_GENERATE_INTERNAL_LABEL (loop_lab, "LPSRL", labelno);
  ASM_GENERATE_INTERNAL_LABEL (end_lab, "LPSRE", labelno++);

  ASM_OUTPUT_INTERNAL_LABEL (asm_out_file, loop_lab);

  /* Jump to END_LAB if TEST_ADDR == LAST_ADDR.  */
  xops[0] = reg1;
  xops[1] = reg2;
  xops[2] = gen_rtx_REG (BImode, PR_REG (6));
  output_asm_insn ("cmp.eq %2, %I2 = %0, %1", xops);
  fprintf (asm_out_file, "\t(%s) br.cond.dpnt ", reg_names [REGNO (xops[2])]);
  assemble_name_raw (asm_out_file, end_lab);
  fputc ('\n', asm_out_file);

  /* TEST_ADDR = TEST_ADDR + PROBE_INTERVAL.  */
  xops[1] = GEN_INT (-PROBE_INTERVAL);
  output_asm_insn ("addl %0 = %1, %0", xops);
  fputs ("\t;;\n", asm_out_file);

  /* Probe at TEST_ADDR and branch.  */
  output_asm_insn ("probe.w.fault %0, 0", xops);
  fprintf (asm_out_file, "\tbr ");
  assemble_name_raw (asm_out_file, loop_lab);
  fputc ('\n', asm_out_file);

  ASM_OUTPUT_INTERNAL_LABEL (asm_out_file, end_lab);

  return "";
}

/* Called after register allocation to add any instructions needed for the
   prologue.  Using a prologue insn is favored compared to putting all of the
   instructions in output_function_prologue(), since it allows the scheduler
   to intermix instructions with the saves of the caller saved registers.  In
   some cases, it might be necessary to emit a barrier instruction as the last
   insn to prevent such scheduling.

   Also any insns generated here should have RTX_FRAME_RELATED_P(insn) = 1
   so that the debug info generation code can handle them properly.

   The register save area is laid out like so:
   cfa+16
	[ varargs spill area ]
	[ fr register spill area ]
	[ br register spill area ]
	[ ar register spill area ]
	[ pr register spill area ]
	[ gr register spill area ] */

/* ??? Get inefficient code when the frame size is larger than can fit in an
   adds instruction.  */

void
ia64_expand_prologue (void)
{
  rtx insn, ar_pfs_save_reg, ar_unat_save_reg;
  int i, epilogue_p, regno, alt_regno, cfa_off, n_varargs;
  rtx reg, alt_reg;

  ia64_compute_frame_size (get_frame_size ());
  last_scratch_gr_reg = 15;

  if (flag_stack_usage_info)
    current_function_static_stack_size = current_frame_info.total_size;

  if (flag_stack_check == STATIC_BUILTIN_STACK_CHECK)
    {
      HOST_WIDE_INT size = current_frame_info.total_size;
      int bs_size = BACKING_STORE_SIZE (current_frame_info.n_input_regs
					  + current_frame_info.n_local_regs);

      if (crtl->is_leaf && !cfun->calls_alloca)
	{
	  if (size > PROBE_INTERVAL && size > STACK_CHECK_PROTECT)
	    ia64_emit_probe_stack_range (STACK_CHECK_PROTECT,
					 size - STACK_CHECK_PROTECT,
					 bs_size);
	  else if (size + bs_size > STACK_CHECK_PROTECT)
	    ia64_emit_probe_stack_range (STACK_CHECK_PROTECT, 0, bs_size);
	}
      else if (size + bs_size > 0)
	ia64_emit_probe_stack_range (STACK_CHECK_PROTECT, size, bs_size);
    }

  if (dump_file) 
    {
      fprintf (dump_file, "ia64 frame related registers "
               "recorded in current_frame_info.r[]:\n");
#define PRINTREG(a) if (current_frame_info.r[a]) \
        fprintf(dump_file, "%s = %d\n", #a, current_frame_info.r[a])
      PRINTREG(reg_fp);
      PRINTREG(reg_save_b0);
      PRINTREG(reg_save_pr);
      PRINTREG(reg_save_ar_pfs);
      PRINTREG(reg_save_ar_unat);
      PRINTREG(reg_save_ar_lc);
      PRINTREG(reg_save_gp);
#undef PRINTREG
    }

  /* If there is no epilogue, then we don't need some prologue insns.
     We need to avoid emitting the dead prologue insns, because flow
     will complain about them.  */
  if (optimize)
    {
      edge e;
      edge_iterator ei;

      FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR_FOR_FN (cfun)->preds)
	if ((e->flags & EDGE_FAKE) == 0
	    && (e->flags & EDGE_FALLTHRU) != 0)
	  break;
      epilogue_p = (e != NULL);
    }
  else
    epilogue_p = 1;

  /* Set the local, input, and output register names.  We need to do this
     for GNU libc, which creates crti.S/crtn.S by splitting initfini.c in
     half.  If we use in/loc/out register names, then we get assembler errors
     in crtn.S because there is no alloc insn or regstk directive in there.  */
  if (! TARGET_REG_NAMES)
    {
      int inputs = current_frame_info.n_input_regs;
      int locals = current_frame_info.n_local_regs;
      int outputs = current_frame_info.n_output_regs;

      for (i = 0; i < inputs; i++)
	reg_names[IN_REG (i)] = ia64_reg_numbers[i];
      for (i = 0; i < locals; i++)
	reg_names[LOC_REG (i)] = ia64_reg_numbers[inputs + i];
      for (i = 0; i < outputs; i++)
	reg_names[OUT_REG (i)] = ia64_reg_numbers[inputs + locals + i];
    }

  /* Set the frame pointer register name.  The regnum is logically loc79,
     but of course we'll not have allocated that many locals.  Rather than
     worrying about renumbering the existing rtxs, we adjust the name.  */
  /* ??? This code means that we can never use one local register when
     there is a frame pointer.  loc79 gets wasted in this case, as it is
     renamed to a register that will never be used.  See also the try_locals
     code in find_gr_spill.  */
  if (current_frame_info.r[reg_fp])
    {
      const char *tmp = reg_names[HARD_FRAME_POINTER_REGNUM];
      reg_names[HARD_FRAME_POINTER_REGNUM]
	= reg_names[current_frame_info.r[reg_fp]];
      reg_names[current_frame_info.r[reg_fp]] = tmp;
    }

  /* We don't need an alloc instruction if we've used no outputs or locals.  */
  if (current_frame_info.n_local_regs == 0
      && current_frame_info.n_output_regs == 0
      && current_frame_info.n_input_regs <= crtl->args.info.int_regs
      && !TEST_HARD_REG_BIT (current_frame_info.mask, AR_PFS_REGNUM))
    {
      /* If there is no alloc, but there are input registers used, then we
	 need a .regstk directive.  */
      current_frame_info.need_regstk = (TARGET_REG_NAMES != 0);
      ar_pfs_save_reg = NULL_RTX;
    }
  else
    {
      current_frame_info.need_regstk = 0;

      if (current_frame_info.r[reg_save_ar_pfs])
        {
	  regno = current_frame_info.r[reg_save_ar_pfs];
	  reg_emitted (reg_save_ar_pfs);
	}
      else
	regno = next_scratch_gr_reg ();
      ar_pfs_save_reg = gen_rtx_REG (DImode, regno);

      insn = emit_insn (gen_alloc (ar_pfs_save_reg,
				   GEN_INT (current_frame_info.n_input_regs),
				   GEN_INT (current_frame_info.n_local_regs),
				   GEN_INT (current_frame_info.n_output_regs),
				   GEN_INT (current_frame_info.n_rotate_regs)));
      if (current_frame_info.r[reg_save_ar_pfs])
	{
	  RTX_FRAME_RELATED_P (insn) = 1;
	  add_reg_note (insn, REG_CFA_REGISTER,
			gen_rtx_SET (VOIDmode,
				     ar_pfs_save_reg,
				     gen_rtx_REG (DImode, AR_PFS_REGNUM)));
	}
    }

  /* Set up frame pointer, stack pointer, and spill iterators.  */

  n_varargs = cfun->machine->n_varargs;
  setup_spill_pointers (current_frame_info.n_spilled + n_varargs,
			stack_pointer_rtx, 0);

  if (frame_pointer_needed)
    {
      insn = emit_move_insn (hard_frame_pointer_rtx, stack_pointer_rtx);
      RTX_FRAME_RELATED_P (insn) = 1;

      /* Force the unwind info to recognize this as defining a new CFA,
	 rather than some temp register setup.  */
      add_reg_note (insn, REG_CFA_ADJUST_CFA, NULL_RTX);
    }

  if (current_frame_info.total_size != 0)
    {
      rtx frame_size_rtx = GEN_INT (- current_frame_info.total_size);
      rtx offset;

      if (satisfies_constraint_I (frame_size_rtx))
	offset = frame_size_rtx;
      else
	{
	  regno = next_scratch_gr_reg ();
	  offset = gen_rtx_REG (DImode, regno);
	  emit_move_insn (offset, frame_size_rtx);
	}

      insn = emit_insn (gen_adddi3 (stack_pointer_rtx,
				    stack_pointer_rtx, offset));

      if (! frame_pointer_needed)
	{
	  RTX_FRAME_RELATED_P (insn) = 1;
	  add_reg_note (insn, REG_CFA_ADJUST_CFA,
			gen_rtx_SET (VOIDmode,
				     stack_pointer_rtx,
				     gen_rtx_PLUS (DImode,
						   stack_pointer_rtx,
						   frame_size_rtx)));
	}

      /* ??? At this point we must generate a magic insn that appears to
	 modify the stack pointer, the frame pointer, and all spill
	 iterators.  This would allow the most scheduling freedom.  For
	 now, just hard stop.  */
      emit_insn (gen_blockage ());
    }

  /* Must copy out ar.unat before doing any integer spills.  */
  if (TEST_HARD_REG_BIT (current_frame_info.mask, AR_UNAT_REGNUM))
    {
      if (current_frame_info.r[reg_save_ar_unat])
        {
	  ar_unat_save_reg
	    = gen_rtx_REG (DImode, current_frame_info.r[reg_save_ar_unat]);
	  reg_emitted (reg_save_ar_unat);
	}
      else
	{
	  alt_regno = next_scratch_gr_reg ();
	  ar_unat_save_reg = gen_rtx_REG (DImode, alt_regno);
	  current_frame_info.gr_used_mask |= 1 << alt_regno;
	}

      reg = gen_rtx_REG (DImode, AR_UNAT_REGNUM);
      insn = emit_move_insn (ar_unat_save_reg, reg);
      if (current_frame_info.r[reg_save_ar_unat])
	{
	  RTX_FRAME_RELATED_P (insn) = 1;
	  add_reg_note (insn, REG_CFA_REGISTER, NULL_RTX);
	}

      /* Even if we're not going to generate an epilogue, we still
	 need to save the register so that EH works.  */
      if (! epilogue_p && current_frame_info.r[reg_save_ar_unat])
	emit_insn (gen_prologue_use (ar_unat_save_reg));
    }
  else
    ar_unat_save_reg = NULL_RTX;

  /* Spill all varargs registers.  Do this before spilling any GR registers,
     since we want the UNAT bits for the GR registers to override the UNAT
     bits from varargs, which we don't care about.  */

  cfa_off = -16;
  for (regno = GR_ARG_FIRST + 7; n_varargs > 0; --n_varargs, --regno)
    {
      reg = gen_rtx_REG (DImode, regno);
      do_spill (gen_gr_spill, reg, cfa_off += 8, NULL_RTX);
    }

  /* Locate the bottom of the register save area.  */
  cfa_off = (current_frame_info.spill_cfa_off
	     + current_frame_info.spill_size
	     + current_frame_info.extra_spill_size);

  /* Save the predicate register block either in a register or in memory.  */
  if (TEST_HARD_REG_BIT (current_frame_info.mask, PR_REG (0)))
    {
      reg = gen_rtx_REG (DImode, PR_REG (0));
      if (current_frame_info.r[reg_save_pr] != 0)
	{
	  alt_reg = gen_rtx_REG (DImode, current_frame_info.r[reg_save_pr]);
	  reg_emitted (reg_save_pr);
	  insn = emit_move_insn (alt_reg, reg);

	  /* ??? Denote pr spill/fill by a DImode move that modifies all
	     64 hard registers.  */
	  RTX_FRAME_RELATED_P (insn) = 1;
	  add_reg_note (insn, REG_CFA_REGISTER, NULL_RTX);

	  /* Even if we're not going to generate an epilogue, we still
	     need to save the register so that EH works.  */
	  if (! epilogue_p)
	    emit_insn (gen_prologue_use (alt_reg));
	}
      else
	{
	  alt_regno = next_scratch_gr_reg ();
	  alt_reg = gen_rtx_REG (DImode, alt_regno);
	  insn = emit_move_insn (alt_reg, reg);
	  do_spill (gen_movdi_x, alt_reg, cfa_off, reg);
	  cfa_off -= 8;
	}
    }

  /* Handle AR regs in numerical order.  All of them get special handling.  */
  if (TEST_HARD_REG_BIT (current_frame_info.mask, AR_UNAT_REGNUM)
      && current_frame_info.r[reg_save_ar_unat] == 0)
    {
      reg = gen_rtx_REG (DImode, AR_UNAT_REGNUM);
      do_spill (gen_movdi_x, ar_unat_save_reg, cfa_off, reg);
      cfa_off -= 8;
    }

  /* The alloc insn already copied ar.pfs into a general register.  The
     only thing we have to do now is copy that register to a stack slot
     if we'd not allocated a local register for the job.  */
  if (TEST_HARD_REG_BIT (current_frame_info.mask, AR_PFS_REGNUM)
      && current_frame_info.r[reg_save_ar_pfs] == 0)
    {
      reg = gen_rtx_REG (DImode, AR_PFS_REGNUM);
      do_spill (gen_movdi_x, ar_pfs_save_reg, cfa_off, reg);
      cfa_off -= 8;
    }

  if (TEST_HARD_REG_BIT (current_frame_info.mask, AR_LC_REGNUM))
    {
      reg = gen_rtx_REG (DImode, AR_LC_REGNUM);
      if (current_frame_info.r[reg_save_ar_lc] != 0)
	{
	  alt_reg = gen_rtx_REG (DImode, current_frame_info.r[reg_save_ar_lc]);
	  reg_emitted (reg_save_ar_lc);
	  insn = emit_move_insn (alt_reg, reg);
	  RTX_FRAME_RELATED_P (insn) = 1;
	  add_reg_note (insn, REG_CFA_REGISTER, NULL_RTX);

	  /* Even if we're not going to generate an epilogue, we still
	     need to save the register so that EH works.  */
	  if (! epilogue_p)
	    emit_insn (gen_prologue_use (alt_reg));
	}
      else
	{
	  alt_regno = next_scratch_gr_reg ();
	  alt_reg = gen_rtx_REG (DImode, alt_regno);
	  emit_move_insn (alt_reg, reg);
	  do_spill (gen_movdi_x, alt_reg, cfa_off, reg);
	  cfa_off -= 8;
	}
    }

  /* Save the return pointer.  */
  if (TEST_HARD_REG_BIT (current_frame_info.mask, BR_REG (0)))
    {
      reg = gen_rtx_REG (DImode, BR_REG (0));
      if (current_frame_info.r[reg_save_b0] != 0)
	{
          alt_reg = gen_rtx_REG (DImode, current_frame_info.r[reg_save_b0]);
          reg_emitted (reg_save_b0);
	  insn = emit_move_insn (alt_reg, reg);
	  RTX_FRAME_RELATED_P (insn) = 1;
	  add_reg_note (insn, REG_CFA_REGISTER,
			gen_rtx_SET (VOIDmode, alt_reg, pc_rtx));

	  /* Even if we're not going to generate an epilogue, we still
	     need to save the register so that EH works.  */
	  if (! epilogue_p)
	    emit_insn (gen_prologue_use (alt_reg));
	}
      else
	{
	  alt_regno = next_scratch_gr_reg ();
	  alt_reg = gen_rtx_REG (DImode, alt_regno);
	  emit_move_insn (alt_reg, reg);
	  do_spill (gen_movdi_x, alt_reg, cfa_off, reg);
	  cfa_off -= 8;
	}
    }

  if (current_frame_info.r[reg_save_gp])
    {
      reg_emitted (reg_save_gp);
      insn = emit_move_insn (gen_rtx_REG (DImode,
					  current_frame_info.r[reg_save_gp]),
			     pic_offset_table_rtx);
    }

  /* We should now be at the base of the gr/br/fr spill area.  */
  gcc_assert (cfa_off == (current_frame_info.spill_cfa_off
			  + current_frame_info.spill_size));

  /* Spill all general registers.  */
  for (regno = GR_REG (1); regno <= GR_REG (31); ++regno)
    if (TEST_HARD_REG_BIT (current_frame_info.mask, regno))
      {
	reg = gen_rtx_REG (DImode, regno);
	do_spill (gen_gr_spill, reg, cfa_off, reg);
	cfa_off -= 8;
      }

  /* Spill the rest of the BR registers.  */
  for (regno = BR_REG (1); regno <= BR_REG (7); ++regno)
    if (TEST_HARD_REG_BIT (current_frame_info.mask, regno))
      {
	alt_regno = next_scratch_gr_reg ();
	alt_reg = gen_rtx_REG (DImode, alt_regno);
	reg = gen_rtx_REG (DImode, regno);
	emit_move_insn (alt_reg, reg);
	do_spill (gen_movdi_x, alt_reg, cfa_off, reg);
	cfa_off -= 8;
      }

  /* Align the frame and spill all FR registers.  */
  for (regno = FR_REG (2); regno <= FR_REG (127); ++regno)
    if (TEST_HARD_REG_BIT (current_frame_info.mask, regno))
      {
        gcc_assert (!(cfa_off & 15));
	reg = gen_rtx_REG (XFmode, regno);
	do_spill (gen_fr_spill_x, reg, cfa_off, reg);
	cfa_off -= 16;
      }

  gcc_assert (cfa_off == current_frame_info.spill_cfa_off);

  finish_spill_pointers ();
}

/* Output the textual info surrounding the prologue.  */

void
ia64_start_function (FILE *file, const char *fnname,
		     tree decl ATTRIBUTE_UNUSED)
{
#if TARGET_ABI_OPEN_VMS
  vms_start_function (fnname);
#endif

  fputs ("\t.proc ", file);
  assemble_name (file, fnname);
  fputc ('\n', file);
  ASM_OUTPUT_LABEL (file, fnname);
}

/* Called after register allocation to add any instructions needed for the
   epilogue.  Using an epilogue insn is favored compared to putting all of the
   instructions in output_function_prologue(), since it allows the scheduler
   to intermix instructions with the saves of the caller saved registers.  In
   some cases, it might be necessary to emit a barrier instruction as the last
   insn to prevent such scheduling.  */

void
ia64_expand_epilogue (int sibcall_p)
{
  rtx insn, reg, alt_reg, ar_unat_save_reg;
  int regno, alt_regno, cfa_off;

  ia64_compute_frame_size (get_frame_size ());

  /* If there is a frame pointer, then we use it instead of the stack
     pointer, so that the stack pointer does not need to be valid when
     the epilogue starts.  See EXIT_IGNORE_STACK.  */
  if (frame_pointer_needed)
    setup_spill_pointers (current_frame_info.n_spilled,
			  hard_frame_pointer_rtx, 0);
  else
    setup_spill_pointers (current_frame_info.n_spilled, stack_pointer_rtx,
			  current_frame_info.total_size);

  if (current_frame_info.total_size != 0)
    {
      /* ??? At this point we must generate a magic insn that appears to
         modify the spill iterators and the frame pointer.  This would
	 allow the most scheduling freedom.  For now, just hard stop.  */
      emit_insn (gen_blockage ());
    }

  /* Locate the bottom of the register save area.  */
  cfa_off = (current_frame_info.spill_cfa_off
	     + current_frame_info.spill_size
	     + current_frame_info.extra_spill_size);

  /* Restore the predicate registers.  */
  if (TEST_HARD_REG_BIT (current_frame_info.mask, PR_REG (0)))
    {
      if (current_frame_info.r[reg_save_pr] != 0)
        {
	  alt_reg = gen_rtx_REG (DImode, current_frame_info.r[reg_save_pr]);
	  reg_emitted (reg_save_pr);
	}
      else
	{
	  alt_regno = next_scratch_gr_reg ();
	  alt_reg = gen_rtx_REG (DImode, alt_regno);
	  do_restore (gen_movdi_x, alt_reg, cfa_off);
	  cfa_off -= 8;
	}
      reg = gen_rtx_REG (DImode, PR_REG (0));
      emit_move_insn (reg, alt_reg);
    }

  /* Restore the application registers.  */

  /* Load the saved unat from the stack, but do not restore it until
     after the GRs have been restored.  */
  if (TEST_HARD_REG_BIT (current_frame_info.mask, AR_UNAT_REGNUM))
    {
      if (current_frame_info.r[reg_save_ar_unat] != 0)
        {
          ar_unat_save_reg
	    = gen_rtx_REG (DImode, current_frame_info.r[reg_save_ar_unat]);
	  reg_emitted (reg_save_ar_unat);
	}
      else
	{
	  alt_regno = next_scratch_gr_reg ();
	  ar_unat_save_reg = gen_rtx_REG (DImode, alt_regno);
	  current_frame_info.gr_used_mask |= 1 << alt_regno;
	  do_restore (gen_movdi_x, ar_unat_save_reg, cfa_off);
	  cfa_off -= 8;
	}
    }
  else
    ar_unat_save_reg = NULL_RTX;

  if (current_frame_info.r[reg_save_ar_pfs] != 0)
    {
      reg_emitted (reg_save_ar_pfs);
      alt_reg = gen_rtx_REG (DImode, current_frame_info.r[reg_save_ar_pfs]);
      reg = gen_rtx_REG (DImode, AR_PFS_REGNUM);
      emit_move_insn (reg, alt_reg);
    }
  else if (TEST_HARD_REG_BIT (current_frame_info.mask, AR_PFS_REGNUM))
    {
      alt_regno = next_scratch_gr_reg ();
      alt_reg = gen_rtx_REG (DImode, alt_regno);
      do_restore (gen_movdi_x, alt_reg, cfa_off);
      cfa_off -= 8;
      reg = gen_rtx_REG (DImode, AR_PFS_REGNUM);
      emit_move_insn (reg, alt_reg);
    }

  if (TEST_HARD_REG_BIT (current_frame_info.mask, AR_LC_REGNUM))
    {
      if (current_frame_info.r[reg_save_ar_lc] != 0)
        {
	  alt_reg = gen_rtx_REG (DImode, current_frame_info.r[reg_save_ar_lc]);
          reg_emitted (reg_save_ar_lc);
	}
      else
	{
	  alt_regno = next_scratch_gr_reg ();
	  alt_reg = gen_rtx_REG (DImode, alt_regno);
	  do_restore (gen_movdi_x, alt_reg, cfa_off);
	  cfa_off -= 8;
	}
      reg = gen_rtx_REG (DImode, AR_LC_REGNUM);
      emit_move_insn (reg, alt_reg);
    }

  /* Restore the return pointer.  */
  if (TEST_HARD_REG_BIT (current_frame_info.mask, BR_REG (0)))
    {
      if (current_frame_info.r[reg_save_b0] != 0)
        {
         alt_reg = gen_rtx_REG (DImode, current_frame_info.r[reg_save_b0]);
         reg_emitted (reg_save_b0);
        }
      else
	{
	  alt_regno = next_scratch_gr_reg ();
	  alt_reg = gen_rtx_REG (DImode, alt_regno);
	  do_restore (gen_movdi_x, alt_reg, cfa_off);
	  cfa_off -= 8;
	}
      reg = gen_rtx_REG (DImode, BR_REG (0));
      emit_move_insn (reg, alt_reg);
    }

  /* We should now be at the base of the gr/br/fr spill area.  */
  gcc_assert (cfa_off == (current_frame_info.spill_cfa_off
			  + current_frame_info.spill_size));

  /* The GP may be stored on the stack in the prologue, but it's
     never restored in the epilogue.  Skip the stack slot.  */
  if (TEST_HARD_REG_BIT (current_frame_info.mask, GR_REG (1)))
    cfa_off -= 8;

  /* Restore all general registers.  */
  for (regno = GR_REG (2); regno <= GR_REG (31); ++regno)
    if (TEST_HARD_REG_BIT (current_frame_info.mask, regno))
      {
	reg = gen_rtx_REG (DImode, regno);
	do_restore (gen_gr_restore, reg, cfa_off);
	cfa_off -= 8;
      }

  /* Restore the branch registers.  */
  for (regno = BR_REG (1); regno <= BR_REG (7); ++regno)
    if (TEST_HARD_REG_BIT (current_frame_info.mask, regno))
      {
	alt_regno = next_scratch_gr_reg ();
	alt_reg = gen_rtx_REG (DImode, alt_regno);
	do_restore (gen_movdi_x, alt_reg, cfa_off);
	cfa_off -= 8;
	reg = gen_rtx_REG (DImode, regno);
	emit_move_insn (reg, alt_reg);
      }

  /* Restore floating point registers.  */
  for (regno = FR_REG (2); regno <= FR_REG (127); ++regno)
    if (TEST_HARD_REG_BIT (current_frame_info.mask, regno))
      {
        gcc_assert (!(cfa_off & 15));
	reg = gen_rtx_REG (XFmode, regno);
	do_restore (gen_fr_restore_x, reg, cfa_off);
	cfa_off -= 16;
      }

  /* Restore ar.unat for real.  */
  if (TEST_HARD_REG_BIT (current_frame_info.mask, AR_UNAT_REGNUM))
    {
      reg = gen_rtx_REG (DImode, AR_UNAT_REGNUM);
      emit_move_insn (reg, ar_unat_save_reg);
    }

  gcc_assert (cfa_off == current_frame_info.spill_cfa_off);

  finish_spill_pointers ();

  if (current_frame_info.total_size
      || cfun->machine->ia64_eh_epilogue_sp
      || frame_pointer_needed)
    {
      /* ??? At this point we must generate a magic insn that appears to
         modify the spill iterators, the stack pointer, and the frame
	 pointer.  This would allow the most scheduling freedom.  For now,
	 just hard stop.  */
      emit_insn (gen_blockage ());
    }

  if (cfun->machine->ia64_eh_epilogue_sp)
    emit_move_insn (stack_pointer_rtx, cfun->machine->ia64_eh_epilogue_sp);
  else if (frame_pointer_needed)
    {
      insn = emit_move_insn (stack_pointer_rtx, hard_frame_pointer_rtx);
      RTX_FRAME_RELATED_P (insn) = 1;
      add_reg_note (insn, REG_CFA_ADJUST_CFA, NULL);
    }
  else if (current_frame_info.total_size)
    {
      rtx offset, frame_size_rtx;

      frame_size_rtx = GEN_INT (current_frame_info.total_size);
      if (satisfies_constraint_I (frame_size_rtx))
	offset = frame_size_rtx;
      else
	{
	  regno = next_scratch_gr_reg ();
	  offset = gen_rtx_REG (DImode, regno);
	  emit_move_insn (offset, frame_size_rtx);
	}

      insn = emit_insn (gen_adddi3 (stack_pointer_rtx, stack_pointer_rtx,
				    offset));

      RTX_FRAME_RELATED_P (insn) = 1;
      add_reg_note (insn, REG_CFA_ADJUST_CFA,
		    gen_rtx_SET (VOIDmode,
				 stack_pointer_rtx,
				 gen_rtx_PLUS (DImode,
					       stack_pointer_rtx,
					       frame_size_rtx)));
    }

  if (cfun->machine->ia64_eh_epilogue_bsp)
    emit_insn (gen_set_bsp (cfun->machine->ia64_eh_epilogue_bsp));

  if (! sibcall_p)
    emit_jump_insn (gen_return_internal (gen_rtx_REG (DImode, BR_REG (0))));
  else
    {
      int fp = GR_REG (2);
      /* We need a throw away register here, r0 and r1 are reserved,
	 so r2 is the first available call clobbered register.  If
	 there was a frame_pointer register, we may have swapped the
	 names of r2 and HARD_FRAME_POINTER_REGNUM, so we have to make
	 sure we're using the string "r2" when emitting the register
	 name for the assembler.  */
      if (current_frame_info.r[reg_fp] 
          && current_frame_info.r[reg_fp] == GR_REG (2))
	fp = HARD_FRAME_POINTER_REGNUM;

      /* We must emit an alloc to force the input registers to become output
	 registers.  Otherwise, if the callee tries to pass its parameters
	 through to another call without an intervening alloc, then these
	 values get lost.  */
      /* ??? We don't need to preserve all input registers.  We only need to
	 preserve those input registers used as arguments to the sibling call.
	 It is unclear how to compute that number here.  */
      if (current_frame_info.n_input_regs != 0)
	{
	  rtx n_inputs = GEN_INT (current_frame_info.n_input_regs);

	  insn = emit_insn (gen_alloc (gen_rtx_REG (DImode, fp),
				const0_rtx, const0_rtx,
				n_inputs, const0_rtx));
	  RTX_FRAME_RELATED_P (insn) = 1;

	  /* ??? We need to mark the alloc as frame-related so that it gets
	     passed into ia64_asm_unwind_emit for ia64-specific unwinding.
	     But there's nothing dwarf2 related to be done wrt the register
	     windows.  If we do nothing, dwarf2out will abort on the UNSPEC;
	     the empty parallel means dwarf2out will not see anything.  */
	  add_reg_note (insn, REG_FRAME_RELATED_EXPR,
			gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (0)));
	}
    }
}

/* Return 1 if br.ret can do all the work required to return from a
   function.  */

int
ia64_direct_return (void)
{
  if (reload_completed && ! frame_pointer_needed)
    {
      ia64_compute_frame_size (get_frame_size ());

      return (current_frame_info.total_size == 0
	      && current_frame_info.n_spilled == 0
	      && current_frame_info.r[reg_save_b0] == 0
	      && current_frame_info.r[reg_save_pr] == 0
	      && current_frame_info.r[reg_save_ar_pfs] == 0
	      && current_frame_info.r[reg_save_ar_unat] == 0
	      && current_frame_info.r[reg_save_ar_lc] == 0);
    }
  return 0;
}

/* Return the magic cookie that we use to hold the return address
   during early compilation.  */

rtx
ia64_return_addr_rtx (HOST_WIDE_INT count, rtx frame ATTRIBUTE_UNUSED)
{
  if (count != 0)
    return NULL;
  return gen_rtx_UNSPEC (Pmode, gen_rtvec (1, const0_rtx), UNSPEC_RET_ADDR);
}

/* Split this value after reload, now that we know where the return
   address is saved.  */

void
ia64_split_return_addr_rtx (rtx dest)
{
  rtx src;

  if (TEST_HARD_REG_BIT (current_frame_info.mask, BR_REG (0)))
    {
      if (current_frame_info.r[reg_save_b0] != 0)
        {
	  src = gen_rtx_REG (DImode, current_frame_info.r[reg_save_b0]);
	  reg_emitted (reg_save_b0);
	}
      else
	{
	  HOST_WIDE_INT off;
	  unsigned int regno;
	  rtx off_r;

	  /* Compute offset from CFA for BR0.  */
	  /* ??? Must be kept in sync with ia64_expand_prologue.  */
	  off = (current_frame_info.spill_cfa_off
		 + current_frame_info.spill_size);
	  for (regno = GR_REG (1); regno <= GR_REG (31); ++regno)
	    if (TEST_HARD_REG_BIT (current_frame_info.mask, regno))
	      off -= 8;

	  /* Convert CFA offset to a register based offset.  */
	  if (frame_pointer_needed)
	    src = hard_frame_pointer_rtx;
	  else
	    {
	      src = stack_pointer_rtx;
	      off += current_frame_info.total_size;
	    }

	  /* Load address into scratch register.  */
	  off_r = GEN_INT (off);
	  if (satisfies_constraint_I (off_r))
	    emit_insn (gen_adddi3 (dest, src, off_r));
	  else
	    {
	      emit_move_insn (dest, off_r);
	      emit_insn (gen_adddi3 (dest, src, dest));
	    }

	  src = gen_rtx_MEM (Pmode, dest);
	}
    }
  else
    src = gen_rtx_REG (DImode, BR_REG (0));

  emit_move_insn (dest, src);
}

int
ia64_hard_regno_rename_ok (int from, int to)
{
  /* Don't clobber any of the registers we reserved for the prologue.  */
  unsigned int r;

  for (r = reg_fp; r <= reg_save_ar_lc; r++)
    if (to == current_frame_info.r[r] 
        || from == current_frame_info.r[r]
        || to == emitted_frame_related_regs[r]
        || from == emitted_frame_related_regs[r])
      return 0;

  /* Don't use output registers outside the register frame.  */
  if (OUT_REGNO_P (to) && to >= OUT_REG (current_frame_info.n_output_regs))
    return 0;

  /* Retain even/oddness on predicate register pairs.  */
  if (PR_REGNO_P (from) && PR_REGNO_P (to))
    return (from & 1) == (to & 1);

  return 1;
}

/* Target hook for assembling integer objects.  Handle word-sized
   aligned objects and detect the cases when @fptr is needed.  */

static bool
ia64_assemble_integer (rtx x, unsigned int size, int aligned_p)
{
  if (size == POINTER_SIZE / BITS_PER_UNIT
      && !(TARGET_NO_PIC || TARGET_AUTO_PIC)
      && GET_CODE (x) == SYMBOL_REF
      && SYMBOL_REF_FUNCTION_P (x))
    {
      static const char * const directive[2][2] = {
	  /* 64-bit pointer */  /* 32-bit pointer */
	{ "\tdata8.ua\t@fptr(", "\tdata4.ua\t@fptr("},	/* unaligned */
	{ "\tdata8\t@fptr(",    "\tdata4\t@fptr("}	/* aligned */
      };
      fputs (directive[(aligned_p != 0)][POINTER_SIZE == 32], asm_out_file);
      output_addr_const (asm_out_file, x);
      fputs (")\n", asm_out_file);
      return true;
    }
  return default_assemble_integer (x, size, aligned_p);
}

/* Emit the function prologue.  */

static void
ia64_output_function_prologue (FILE *file, HOST_WIDE_INT size ATTRIBUTE_UNUSED)
{
  int mask, grsave, grsave_prev;

  if (current_frame_info.need_regstk)
    fprintf (file, "\t.regstk %d, %d, %d, %d\n",
	     current_frame_info.n_input_regs,
	     current_frame_info.n_local_regs,
	     current_frame_info.n_output_regs,
	     current_frame_info.n_rotate_regs);

  if (ia64_except_unwind_info (&global_options) != UI_TARGET)
    return;

  /* Emit the .prologue directive.  */

  mask = 0;
  grsave = grsave_prev = 0;
  if (current_frame_info.r[reg_save_b0] != 0)
    {
      mask |= 8;
      grsave = grsave_prev = current_frame_info.r[reg_save_b0];
    }
  if (current_frame_info.r[reg_save_ar_pfs] != 0
      && (grsave_prev == 0
	  || current_frame_info.r[reg_save_ar_pfs] == grsave_prev + 1))
    {
      mask |= 4;
      if (grsave_prev == 0)
	grsave = current_frame_info.r[reg_save_ar_pfs];
      grsave_prev = current_frame_info.r[reg_save_ar_pfs];
    }
  if (current_frame_info.r[reg_fp] != 0
      && (grsave_prev == 0
	  || current_frame_info.r[reg_fp] == grsave_prev + 1))
    {
      mask |= 2;
      if (grsave_prev == 0)
	grsave = HARD_FRAME_POINTER_REGNUM;
      grsave_prev = current_frame_info.r[reg_fp];
    }
  if (current_frame_info.r[reg_save_pr] != 0
      && (grsave_prev == 0
	  || current_frame_info.r[reg_save_pr] == grsave_prev + 1))
    {
      mask |= 1;
      if (grsave_prev == 0)
	grsave = current_frame_info.r[reg_save_pr];
    }

  if (mask && TARGET_GNU_AS)
    fprintf (file, "\t.prologue %d, %d\n", mask,
	     ia64_dbx_register_number (grsave));
  else
    fputs ("\t.prologue\n", file);

  /* Emit a .spill directive, if necessary, to relocate the base of
     the register spill area.  */
  if (current_frame_info.spill_cfa_off != -16)
    fprintf (file, "\t.spill %ld\n",
	     (long) (current_frame_info.spill_cfa_off
		     + current_frame_info.spill_size));
}

/* Emit the .body directive at the scheduled end of the prologue.  */

static void
ia64_output_function_end_prologue (FILE *file)
{
  if (ia64_except_unwind_info (&global_options) != UI_TARGET)
    return;

  fputs ("\t.body\n", file);
}

/* Emit the function epilogue.  */

static void
ia64_output_function_epilogue (FILE *file ATTRIBUTE_UNUSED,
			       HOST_WIDE_INT size ATTRIBUTE_UNUSED)
{
  int i;

  if (current_frame_info.r[reg_fp])
    {
      const char *tmp = reg_names[HARD_FRAME_POINTER_REGNUM];
      reg_names[HARD_FRAME_POINTER_REGNUM]
	= reg_names[current_frame_info.r[reg_fp]];
      reg_names[current_frame_info.r[reg_fp]] = tmp;
      reg_emitted (reg_fp);
    }
  if (! TARGET_REG_NAMES)
    {
      for (i = 0; i < current_frame_info.n_input_regs; i++)
	reg_names[IN_REG (i)] = ia64_input_reg_names[i];
      for (i = 0; i < current_frame_info.n_local_regs; i++)
	reg_names[LOC_REG (i)] = ia64_local_reg_names[i];
      for (i = 0; i < current_frame_info.n_output_regs; i++)
	reg_names[OUT_REG (i)] = ia64_output_reg_names[i];
    }

  current_frame_info.initialized = 0;
}

int
ia64_dbx_register_number (int regno)
{
  /* In ia64_expand_prologue we quite literally renamed the frame pointer
     from its home at loc79 to something inside the register frame.  We
     must perform the same renumbering here for the debug info.  */
  if (current_frame_info.r[reg_fp])
    {
      if (regno == HARD_FRAME_POINTER_REGNUM)
	regno = current_frame_info.r[reg_fp];
      else if (regno == current_frame_info.r[reg_fp])
	regno = HARD_FRAME_POINTER_REGNUM;
    }

  if (IN_REGNO_P (regno))
    return 32 + regno - IN_REG (0);
  else if (LOC_REGNO_P (regno))
    return 32 + current_frame_info.n_input_regs + regno - LOC_REG (0);
  else if (OUT_REGNO_P (regno))
    return (32 + current_frame_info.n_input_regs
	    + current_frame_info.n_local_regs + regno - OUT_REG (0));
  else
    return regno;
}

/* Implement TARGET_TRAMPOLINE_INIT.

   The trampoline should set the static chain pointer to value placed
   into the trampoline and should branch to the specified routine.
   To make the normal indirect-subroutine calling convention work,
   the trampoline must look like a function descriptor; the first
   word being the target address and the second being the target's
   global pointer.

   We abuse the concept of a global pointer by arranging for it
   to point to the data we need to load.  The complete trampoline
   has the following form:

		+-------------------+ \
	TRAMP:	| __ia64_trampoline | |
		+-------------------+  > fake function descriptor
		| TRAMP+16          | |
		+-------------------+ /
		| target descriptor |
		+-------------------+
		| static link	    |
		+-------------------+
*/

static void
ia64_trampoline_init (rtx m_tramp, tree fndecl, rtx static_chain)
{
  rtx fnaddr = XEXP (DECL_RTL (fndecl), 0);
  rtx addr, addr_reg, tramp, eight = GEN_INT (8);

  /* The Intel assembler requires that the global __ia64_trampoline symbol
     be declared explicitly */
  if (!TARGET_GNU_AS)
    {
      static bool declared_ia64_trampoline = false;

      if (!declared_ia64_trampoline)
	{
	  declared_ia64_trampoline = true;
	  (*targetm.asm_out.globalize_label) (asm_out_file,
					      "__ia64_trampoline");
	}
    }

  /* Make sure addresses are Pmode even if we are in ILP32 mode. */
  addr = convert_memory_address (Pmode, XEXP (m_tramp, 0));
  fnaddr = convert_memory_address (Pmode, fnaddr);
  static_chain = convert_memory_address (Pmode, static_chain);

  /* Load up our iterator.  */
  addr_reg = copy_to_reg (addr);
  m_tramp = adjust_automodify_address (m_tramp, Pmode, addr_reg, 0);

  /* The first two words are the fake descriptor:
     __ia64_trampoline, ADDR+16.  */
  tramp = gen_rtx_SYMBOL_REF (Pmode, "__ia64_trampoline");
  if (TARGET_ABI_OPEN_VMS)
    {
      /* HP decided to break the ELF ABI on VMS (to deal with an ambiguity
	 in the Macro-32 compiler) and changed the semantics of the LTOFF22
	 relocation against function symbols to make it identical to the
	 LTOFF_FPTR22 relocation.  Emit the latter directly to stay within
	 strict ELF and dereference to get the bare code address.  */
      rtx reg = gen_reg_rtx (Pmode);
      SYMBOL_REF_FLAGS (tramp) |= SYMBOL_FLAG_FUNCTION;
      emit_move_insn (reg, tramp);
      emit_move_insn (reg, gen_rtx_MEM (Pmode, reg));
      tramp = reg;
   }
  emit_move_insn (m_tramp, tramp);
  emit_insn (gen_adddi3 (addr_reg, addr_reg, eight));
  m_tramp = adjust_automodify_address (m_tramp, VOIDmode, NULL, 8);

  emit_move_insn (m_tramp, force_reg (Pmode, plus_constant (Pmode, addr, 16)));
  emit_insn (gen_adddi3 (addr_reg, addr_reg, eight));
  m_tramp = adjust_automodify_address (m_tramp, VOIDmode, NULL, 8);

  /* The third word is the target descriptor.  */
  emit_move_insn (m_tramp, force_reg (Pmode, fnaddr));
  emit_insn (gen_adddi3 (addr_reg, addr_reg, eight));
  m_tramp = adjust_automodify_address (m_tramp, VOIDmode, NULL, 8);

  /* The fourth word is the static chain.  */
  emit_move_insn (m_tramp, static_chain);
}

/* Do any needed setup for a variadic function.  CUM has not been updated
   for the last named argument which has type TYPE and mode MODE.

   We generate the actual spill instructions during prologue generation.  */

static void
ia64_setup_incoming_varargs (cumulative_args_t cum, enum machine_mode mode,
			     tree type, int * pretend_size,
			     int second_time ATTRIBUTE_UNUSED)
{
  CUMULATIVE_ARGS next_cum = *get_cumulative_args (cum);

  /* Skip the current argument.  */
  ia64_function_arg_advance (pack_cumulative_args (&next_cum), mode, type, 1);

  if (next_cum.words < MAX_ARGUMENT_SLOTS)
    {
      int n = MAX_ARGUMENT_SLOTS - next_cum.words;
      *pretend_size = n * UNITS_PER_WORD;
      cfun->machine->n_varargs = n;
    }
}

/* Check whether TYPE is a homogeneous floating point aggregate.  If
   it is, return the mode of the floating point type that appears
   in all leafs.  If it is not, return VOIDmode.

   An aggregate is a homogeneous floating point aggregate is if all
   fields/elements in it have the same floating point type (e.g,
   SFmode).  128-bit quad-precision floats are excluded.

   Variable sized aggregates should never arrive here, since we should
   have already decided to pass them by reference.  Top-level zero-sized
   aggregates are excluded because our parallels crash the middle-end.  */

static enum machine_mode
hfa_element_mode (const_tree type, bool nested)
{
  enum machine_mode element_mode = VOIDmode;
  enum machine_mode mode;
  enum tree_code code = TREE_CODE (type);
  int know_element_mode = 0;
  tree t;

  if (!nested && (!TYPE_SIZE (type) || integer_zerop (TYPE_SIZE (type))))
    return VOIDmode;

  switch (code)
    {
    case VOID_TYPE:	case INTEGER_TYPE:	case ENUMERAL_TYPE:
    case BOOLEAN_TYPE:	case POINTER_TYPE:
    case OFFSET_TYPE:	case REFERENCE_TYPE:	case METHOD_TYPE:
    case LANG_TYPE:		case FUNCTION_TYPE:
      return VOIDmode;

      /* Fortran complex types are supposed to be HFAs, so we need to handle
	 gcc's COMPLEX_TYPEs as HFAs.  We need to exclude the integral complex
	 types though.  */
    case COMPLEX_TYPE:
      if (GET_MODE_CLASS (TYPE_MODE (type)) == MODE_COMPLEX_FLOAT
	  && TYPE_MODE (type) != TCmode)
	return GET_MODE_INNER (TYPE_MODE (type));
      else
	return VOIDmode;

    case REAL_TYPE:
      /* We want to return VOIDmode for raw REAL_TYPEs, but the actual
	 mode if this is contained within an aggregate.  */
      if (nested && TYPE_MODE (type) != TFmode)
	return TYPE_MODE (type);
      else
	return VOIDmode;

    case ARRAY_TYPE:
      return hfa_element_mode (TREE_TYPE (type), 1);

    case RECORD_TYPE:
    case UNION_TYPE:
    case QUAL_UNION_TYPE:
      for (t = TYPE_FIELDS (type); t; t = DECL_CHAIN (t))
	{
	  if (TREE_CODE (t) != FIELD_DECL)
	    continue;

	  mode = hfa_element_mode (TREE_TYPE (t), 1);
	  if (know_element_mode)
	    {
	      if (mode != element_mode)
		return VOIDmode;
	    }
	  else if (GET_MODE_CLASS (mode) != MODE_FLOAT)
	    return VOIDmode;
	  else
	    {
	      know_element_mode = 1;
	      element_mode = mode;
	    }
	}
      return element_mode;

    default:
      /* If we reach here, we probably have some front-end specific type
	 that the backend doesn't know about.  This can happen via the
	 aggregate_value_p call in init_function_start.  All we can do is
	 ignore unknown tree types.  */
      return VOIDmode;
    }

  return VOIDmode;
}

/* Return the number of words required to hold a quantity of TYPE and MODE
   when passed as an argument.  */
static int
ia64_function_arg_words (const_tree type, enum machine_mode mode)
{
  int words;

  if (mode == BLKmode)
    words = int_size_in_bytes (type);
  else
    words = GET_MODE_SIZE (mode);

  return (words + UNITS_PER_WORD - 1) / UNITS_PER_WORD;  /* round up */
}

/* Return the number of registers that should be skipped so the current
   argument (described by TYPE and WORDS) will be properly aligned.

   Integer and float arguments larger than 8 bytes start at the next
   even boundary.  Aggregates larger than 8 bytes start at the next
   even boundary if the aggregate has 16 byte alignment.  Note that
   in the 32-bit ABI, TImode and TFmode have only 8-byte alignment
   but are still to be aligned in registers.

   ??? The ABI does not specify how to handle aggregates with
   alignment from 9 to 15 bytes, or greater than 16.  We handle them
   all as if they had 16 byte alignment.  Such aggregates can occur
   only if gcc extensions are used.  */
static int
ia64_function_arg_offset (const CUMULATIVE_ARGS *cum,
			  const_tree type, int words)
{
  /* No registers are skipped on VMS.  */
  if (TARGET_ABI_OPEN_VMS || (cum->words & 1) == 0)
    return 0;

  if (type
      && TREE_CODE (type) != INTEGER_TYPE
      && TREE_CODE (type) != REAL_TYPE)
    return TYPE_ALIGN (type) > 8 * BITS_PER_UNIT;
  else
    return words > 1;
}

/* Return rtx for register where argument is passed, or zero if it is passed
   on the stack.  */
/* ??? 128-bit quad-precision floats are always passed in general
   registers.  */

static rtx
ia64_function_arg_1 (cumulative_args_t cum_v, enum machine_mode mode,
		     const_tree type, bool named, bool incoming)
{
  const CUMULATIVE_ARGS *cum = get_cumulative_args (cum_v);

  int basereg = (incoming ? GR_ARG_FIRST : AR_ARG_FIRST);
  int words = ia64_function_arg_words (type, mode);
  int offset = ia64_function_arg_offset (cum, type, words);
  enum machine_mode hfa_mode = VOIDmode;

  /* For OPEN VMS, emit the instruction setting up the argument register here,
     when we know this will be together with the other arguments setup related
     insns.  This is not the conceptually best place to do this, but this is
     the easiest as we have convenient access to cumulative args info.  */

  if (TARGET_ABI_OPEN_VMS && mode == VOIDmode && type == void_type_node
      && named == 1)
    {
      unsigned HOST_WIDE_INT regval = cum->words;
      int i;

      for (i = 0; i < 8; i++)
	regval |= ((int) cum->atypes[i]) << (i * 3 + 8);

      emit_move_insn (gen_rtx_REG (DImode, GR_REG (25)),
		      GEN_INT (regval));
    }

  /* If all argument slots are used, then it must go on the stack.  */
  if (cum->words + offset >= MAX_ARGUMENT_SLOTS)
    return 0;

  /* On OpenVMS argument is either in Rn or Fn.  */
  if (TARGET_ABI_OPEN_VMS)
    {
      if (FLOAT_MODE_P (mode))
	return gen_rtx_REG (mode, FR_ARG_FIRST + cum->words);
      else
	return gen_rtx_REG (mode, basereg + cum->words);
    }

  /* Check for and handle homogeneous FP aggregates.  */
  if (type)
    hfa_mode = hfa_element_mode (type, 0);

  /* Unnamed prototyped hfas are passed as usual.  Named prototyped hfas
     and unprototyped hfas are passed specially.  */
  if (hfa_mode != VOIDmode && (! cum->prototype || named))
    {
      rtx loc[16];
      int i = 0;
      int fp_regs = cum->fp_regs;
      int int_regs = cum->words + offset;
      int hfa_size = GET_MODE_SIZE (hfa_mode);
      int byte_size;
      int args_byte_size;

      /* If prototyped, pass it in FR regs then GR regs.
	 If not prototyped, pass it in both FR and GR regs.

	 If this is an SFmode aggregate, then it is possible to run out of
	 FR regs while GR regs are still left.  In that case, we pass the
	 remaining part in the GR regs.  */

      /* Fill the FP regs.  We do this always.  We stop if we reach the end
	 of the argument, the last FP register, or the last argument slot.  */

      byte_size = ((mode == BLKmode)
		   ? int_size_in_bytes (type) : GET_MODE_SIZE (mode));
      args_byte_size = int_regs * UNITS_PER_WORD;
      offset = 0;
      for (; (offset < byte_size && fp_regs < MAX_ARGUMENT_SLOTS
	      && args_byte_size < (MAX_ARGUMENT_SLOTS * UNITS_PER_WORD)); i++)
	{
	  loc[i] = gen_rtx_EXPR_LIST (VOIDmode,
				      gen_rtx_REG (hfa_mode, (FR_ARG_FIRST
							      + fp_regs)),
				      GEN_INT (offset));
	  offset += hfa_size;
	  args_byte_size += hfa_size;
	  fp_regs++;
	}

      /* If no prototype, then the whole thing must go in GR regs.  */
      if (! cum->prototype)
	offset = 0;
      /* If this is an SFmode aggregate, then we might have some left over
	 that needs to go in GR regs.  */
      else if (byte_size != offset)
	int_regs += offset / UNITS_PER_WORD;

      /* Fill in the GR regs.  We must use DImode here, not the hfa mode.  */

      for (; offset < byte_size && int_regs < MAX_ARGUMENT_SLOTS; i++)
	{
	  enum machine_mode gr_mode = DImode;
	  unsigned int gr_size;

	  /* If we have an odd 4 byte hunk because we ran out of FR regs,
	     then this goes in a GR reg left adjusted/little endian, right
	     adjusted/big endian.  */
	  /* ??? Currently this is handled wrong, because 4-byte hunks are
	     always right adjusted/little endian.  */
	  if (offset & 0x4)
	    gr_mode = SImode;
	  /* If we have an even 4 byte hunk because the aggregate is a
	     multiple of 4 bytes in size, then this goes in a GR reg right
	     adjusted/little endian.  */
	  else if (byte_size - offset == 4)
	    gr_mode = SImode;

	  loc[i] = gen_rtx_EXPR_LIST (VOIDmode,
				      gen_rtx_REG (gr_mode, (basereg
							     + int_regs)),
				      GEN_INT (offset));

	  gr_size = GET_MODE_SIZE (gr_mode);
	  offset += gr_size;
	  if (gr_size == UNITS_PER_WORD
	      || (gr_size < UNITS_PER_WORD && offset % UNITS_PER_WORD == 0))
	    int_regs++;
	  else if (gr_size > UNITS_PER_WORD)
	    int_regs += gr_size / UNITS_PER_WORD;
	}
      return gen_rtx_PARALLEL (mode, gen_rtvec_v (i, loc));
    }
  
  /* Integral and aggregates go in general registers.  If we have run out of
     FR registers, then FP values must also go in general registers.  This can
     happen when we have a SFmode HFA.  */
  else if (mode == TFmode || mode == TCmode
	   || (! FLOAT_MODE_P (mode) || cum->fp_regs == MAX_ARGUMENT_SLOTS))
    {
      int byte_size = ((mode == BLKmode)
                       ? int_size_in_bytes (type) : GET_MODE_SIZE (mode));
      if (BYTES_BIG_ENDIAN
	&& (mode == BLKmode || (type && AGGREGATE_TYPE_P (type)))
	&& byte_size < UNITS_PER_WORD
	&& byte_size > 0)
	{
	  rtx gr_reg = gen_rtx_EXPR_LIST (VOIDmode,
					  gen_rtx_REG (DImode,
						       (basereg + cum->words
							+ offset)),
					  const0_rtx);
	  return gen_rtx_PARALLEL (mode, gen_rtvec (1, gr_reg));
	}
      else
	return gen_rtx_REG (mode, basereg + cum->words + offset);

    }

  /* If there is a prototype, then FP values go in a FR register when
     named, and in a GR register when unnamed.  */
  else if (cum->prototype)
    {
      if (named)
	return gen_rtx_REG (mode, FR_ARG_FIRST + cum->fp_regs);
      /* In big-endian mode, an anonymous SFmode value must be represented
         as (parallel:SF [(expr_list (reg:DI n) (const_int 0))]) to force
	 the value into the high half of the general register.  */
      else if (BYTES_BIG_ENDIAN && mode == SFmode)
	return gen_rtx_PARALLEL (mode,
		 gen_rtvec (1,
                   gen_rtx_EXPR_LIST (VOIDmode,
		     gen_rtx_REG (DImode, basereg + cum->words + offset),
				      const0_rtx)));
      else
	return gen_rtx_REG (mode, basereg + cum->words + offset);
    }
  /* If there is no prototype, then FP values go in both FR and GR
     registers.  */
  else
    {
      /* See comment above.  */
      enum machine_mode inner_mode =
	(BYTES_BIG_ENDIAN && mode == SFmode) ? DImode : mode;

      rtx fp_reg = gen_rtx_EXPR_LIST (VOIDmode,
				      gen_rtx_REG (mode, (FR_ARG_FIRST
							  + cum->fp_regs)),
				      const0_rtx);
      rtx gr_reg = gen_rtx_EXPR_LIST (VOIDmode,
				      gen_rtx_REG (inner_mode,
						   (basereg + cum->words
						    + offset)),
				      const0_rtx);

      return gen_rtx_PARALLEL (mode, gen_rtvec (2, fp_reg, gr_reg));
    }
}

/* Implement TARGET_FUNCION_ARG target hook.  */

static rtx
ia64_function_arg (cumulative_args_t cum, enum machine_mode mode,
		   const_tree type, bool named)
{
  return ia64_function_arg_1 (cum, mode, type, named, false);
}

/* Implement TARGET_FUNCION_INCOMING_ARG target hook.  */

static rtx
ia64_function_incoming_arg (cumulative_args_t cum,
			    enum machine_mode mode,
			    const_tree type, bool named)
{
  return ia64_function_arg_1 (cum, mode, type, named, true);
}

/* Return number of bytes, at the beginning of the argument, that must be
   put in registers.  0 is the argument is entirely in registers or entirely
   in memory.  */

static int
ia64_arg_partial_bytes (cumulative_args_t cum_v, enum machine_mode mode,
			tree type, bool named ATTRIBUTE_UNUSED)
{
  CUMULATIVE_ARGS *cum = get_cumulative_args (cum_v);

  int words = ia64_function_arg_words (type, mode);
  int offset = ia64_function_arg_offset (cum, type, words);

  /* If all argument slots are used, then it must go on the stack.  */
  if (cum->words + offset >= MAX_ARGUMENT_SLOTS)
    return 0;

  /* It doesn't matter whether the argument goes in FR or GR regs.  If
     it fits within the 8 argument slots, then it goes entirely in
     registers.  If it extends past the last argument slot, then the rest
     goes on the stack.  */

  if (words + cum->words + offset <= MAX_ARGUMENT_SLOTS)
    return 0;

  return (MAX_ARGUMENT_SLOTS - cum->words - offset) * UNITS_PER_WORD;
}

/* Return ivms_arg_type based on machine_mode.  */

static enum ivms_arg_type
ia64_arg_type (enum machine_mode mode)
{
  switch (mode)
    {
    case SFmode:
      return FS;
    case DFmode:
      return FT;
    default:
      return I64;
    }
}

/* Update CUM to point after this argument.  This is patterned after
   ia64_function_arg.  */

static void
ia64_function_arg_advance (cumulative_args_t cum_v, enum machine_mode mode,
			   const_tree type, bool named)
{
  CUMULATIVE_ARGS *cum = get_cumulative_args (cum_v);
  int words = ia64_function_arg_words (type, mode);
  int offset = ia64_function_arg_offset (cum, type, words);
  enum machine_mode hfa_mode = VOIDmode;

  /* If all arg slots are already full, then there is nothing to do.  */
  if (cum->words >= MAX_ARGUMENT_SLOTS)
    {
      cum->words += words + offset;
      return;
    }

  cum->atypes[cum->words] = ia64_arg_type (mode);
  cum->words += words + offset;

  /* On OpenVMS argument is either in Rn or Fn.  */
  if (TARGET_ABI_OPEN_VMS)
    {
      cum->int_regs = cum->words;
      cum->fp_regs = cum->words;
      return;
    }

  /* Check for and handle homogeneous FP aggregates.  */
  if (type)
    hfa_mode = hfa_element_mode (type, 0);

  /* Unnamed prototyped hfas are passed as usual.  Named prototyped hfas
     and unprototyped hfas are passed specially.  */
  if (hfa_mode != VOIDmode && (! cum->prototype || named))
    {
      int fp_regs = cum->fp_regs;
      /* This is the original value of cum->words + offset.  */
      int int_regs = cum->words - words;
      int hfa_size = GET_MODE_SIZE (hfa_mode);
      int byte_size;
      int args_byte_size;

      /* If prototyped, pass it in FR regs then GR regs.
	 If not prototyped, pass it in both FR and GR regs.

	 If this is an SFmode aggregate, then it is possible to run out of
	 FR regs while GR regs are still left.  In that case, we pass the
	 remaining part in the GR regs.  */

      /* Fill the FP regs.  We do this always.  We stop if we reach the end
	 of the argument, the last FP register, or the last argument slot.  */

      byte_size = ((mode == BLKmode)
		   ? int_size_in_bytes (type) : GET_MODE_SIZE (mode));
      args_byte_size = int_regs * UNITS_PER_WORD;
      offset = 0;
      for (; (offset < byte_size && fp_regs < MAX_ARGUMENT_SLOTS
	      && args_byte_size < (MAX_ARGUMENT_SLOTS * UNITS_PER_WORD));)
	{
	  offset += hfa_size;
	  args_byte_size += hfa_size;
	  fp_regs++;
	}

      cum->fp_regs = fp_regs;
    }

  /* Integral and aggregates go in general registers.  So do TFmode FP values.
     If we have run out of FR registers, then other FP values must also go in
     general registers.  This can happen when we have a SFmode HFA.  */
  else if (mode == TFmode || mode == TCmode
           || (! FLOAT_MODE_P (mode) || cum->fp_regs == MAX_ARGUMENT_SLOTS))
    cum->int_regs = cum->words;

  /* If there is a prototype, then FP values go in a FR register when
     named, and in a GR register when unnamed.  */
  else if (cum->prototype)
    {
      if (! named)
	cum->int_regs = cum->words;
      else
	/* ??? Complex types should not reach here.  */
	cum->fp_regs += (GET_MODE_CLASS (mode) == MODE_COMPLEX_FLOAT ? 2 : 1);
    }
  /* If there is no prototype, then FP values go in both FR and GR
     registers.  */
  else
    {
      /* ??? Complex types should not reach here.  */
      cum->fp_regs += (GET_MODE_CLASS (mode) == MODE_COMPLEX_FLOAT ? 2 : 1);
      cum->int_regs = cum->words;
    }
}

/* Arguments with alignment larger than 8 bytes start at the next even
   boundary.  On ILP32 HPUX, TFmode arguments start on next even boundary
   even though their normal alignment is 8 bytes.  See ia64_function_arg.  */

static unsigned int
ia64_function_arg_boundary (enum machine_mode mode, const_tree type)
{
  if (mode == TFmode && TARGET_HPUX && TARGET_ILP32)
    return PARM_BOUNDARY * 2;

  if (type)
    {
      if (TYPE_ALIGN (type) > PARM_BOUNDARY)
        return PARM_BOUNDARY * 2;
      else
        return PARM_BOUNDARY;
    }

  if (GET_MODE_BITSIZE (mode) > PARM_BOUNDARY)
    return PARM_BOUNDARY * 2;
  else
    return PARM_BOUNDARY;
}

/* True if it is OK to do sibling call optimization for the specified
   call expression EXP.  DECL will be the called function, or NULL if
   this is an indirect call.  */
static bool
ia64_function_ok_for_sibcall (tree decl, tree exp ATTRIBUTE_UNUSED)
{
  /* We can't perform a sibcall if the current function has the syscall_linkage
     attribute.  */
  if (lookup_attribute ("syscall_linkage",
			TYPE_ATTRIBUTES (TREE_TYPE (current_function_decl))))
    return false;

  /* We must always return with our current GP.  This means we can
     only sibcall to functions defined in the current module unless
     TARGET_CONST_GP is set to true.  */
  return (decl && (*targetm.binds_local_p) (decl)) || TARGET_CONST_GP;
}


/* Implement va_arg.  */

static tree
ia64_gimplify_va_arg (tree valist, tree type, gimple_seq *pre_p,
		      gimple_seq *post_p)
{
  /* Variable sized types are passed by reference.  */
  if (pass_by_reference (NULL, TYPE_MODE (type), type, false))
    {
      tree ptrtype = build_pointer_type (type);
      tree addr = std_gimplify_va_arg_expr (valist, ptrtype, pre_p, post_p);
      return build_va_arg_indirect_ref (addr);
    }

  /* Aggregate arguments with alignment larger than 8 bytes start at
     the next even boundary.  Integer and floating point arguments
     do so if they are larger than 8 bytes, whether or not they are
     also aligned larger than 8 bytes.  */
  if ((TREE_CODE (type) == REAL_TYPE || TREE_CODE (type) == INTEGER_TYPE)
      ? int_size_in_bytes (type) > 8 : TYPE_ALIGN (type) > 8 * BITS_PER_UNIT)
    {
      tree t = fold_build_pointer_plus_hwi (valist, 2 * UNITS_PER_WORD - 1);
      t = build2 (BIT_AND_EXPR, TREE_TYPE (t), t,
		  build_int_cst (TREE_TYPE (t), -2 * UNITS_PER_WORD));
      gimplify_assign (unshare_expr (valist), t, pre_p);
    }

  return std_gimplify_va_arg_expr (valist, type, pre_p, post_p);
}

/* Return 1 if function return value returned in memory.  Return 0 if it is
   in a register.  */

static bool
ia64_return_in_memory (const_tree valtype, const_tree fntype ATTRIBUTE_UNUSED)
{
  enum machine_mode mode;
  enum machine_mode hfa_mode;
  HOST_WIDE_INT byte_size;

  mode = TYPE_MODE (valtype);
  byte_size = GET_MODE_SIZE (mode);
  if (mode == BLKmode)
    {
      byte_size = int_size_in_bytes (valtype);
      if (byte_size < 0)
	return true;
    }

  /* Hfa's with up to 8 elements are returned in the FP argument registers.  */

  hfa_mode = hfa_element_mode (valtype, 0);
  if (hfa_mode != VOIDmode)
    {
      int hfa_size = GET_MODE_SIZE (hfa_mode);

      if (byte_size / hfa_size > MAX_ARGUMENT_SLOTS)
	return true;
      else
	return false;
    }
  else if (byte_size > UNITS_PER_WORD * MAX_INT_RETURN_SLOTS)
    return true;
  else
    return false;
}

/* Return rtx for register that holds the function return value.  */

static rtx
ia64_function_value (const_tree valtype,
		     const_tree fn_decl_or_type,
		     bool outgoing ATTRIBUTE_UNUSED)
{
  enum machine_mode mode;
  enum machine_mode hfa_mode;
  int unsignedp;
  const_tree func = fn_decl_or_type;

  if (fn_decl_or_type
      && !DECL_P (fn_decl_or_type))
    func = NULL;
  
  mode = TYPE_MODE (valtype);
  hfa_mode = hfa_element_mode (valtype, 0);

  if (hfa_mode != VOIDmode)
    {
      rtx loc[8];
      int i;
      int hfa_size;
      int byte_size;
      int offset;

      hfa_size = GET_MODE_SIZE (hfa_mode);
      byte_size = ((mode == BLKmode)
		   ? int_size_in_bytes (valtype) : GET_MODE_SIZE (mode));
      offset = 0;
      for (i = 0; offset < byte_size; i++)
	{
	  loc[i] = gen_rtx_EXPR_LIST (VOIDmode,
				      gen_rtx_REG (hfa_mode, FR_ARG_FIRST + i),
				      GEN_INT (offset));
	  offset += hfa_size;
	}
      return gen_rtx_PARALLEL (mode, gen_rtvec_v (i, loc));
    }
  else if (FLOAT_TYPE_P (valtype) && mode != TFmode && mode != TCmode)
    return gen_rtx_REG (mode, FR_ARG_FIRST);
  else
    {
      bool need_parallel = false;

      /* In big-endian mode, we need to manage the layout of aggregates
	 in the registers so that we get the bits properly aligned in
	 the highpart of the registers.  */
      if (BYTES_BIG_ENDIAN
	  && (mode == BLKmode || (valtype && AGGREGATE_TYPE_P (valtype))))
	need_parallel = true;

      /* Something like struct S { long double x; char a[0] } is not an
	 HFA structure, and therefore doesn't go in fp registers.  But
	 the middle-end will give it XFmode anyway, and XFmode values
	 don't normally fit in integer registers.  So we need to smuggle
	 the value inside a parallel.  */
      else if (mode == XFmode || mode == XCmode || mode == RFmode)
	need_parallel = true;

      if (need_parallel)
	{
	  rtx loc[8];
	  int offset;
	  int bytesize;
	  int i;

	  offset = 0;
	  bytesize = int_size_in_bytes (valtype);
	  /* An empty PARALLEL is invalid here, but the return value
	     doesn't matter for empty structs.  */
	  if (bytesize == 0)
	    return gen_rtx_REG (mode, GR_RET_FIRST);
	  for (i = 0; offset < bytesize; i++)
	    {
	      loc[i] = gen_rtx_EXPR_LIST (VOIDmode,
					  gen_rtx_REG (DImode,
						       GR_RET_FIRST + i),
					  GEN_INT (offset));
	      offset += UNITS_PER_WORD;
	    }
	  return gen_rtx_PARALLEL (mode, gen_rtvec_v (i, loc));
	}

      mode = promote_function_mode (valtype, mode, &unsignedp,
                                    func ? TREE_TYPE (func) : NULL_TREE,
                                    true);

      return gen_rtx_REG (mode, GR_RET_FIRST);
    }
}

/* Worker function for TARGET_LIBCALL_VALUE.  */

static rtx
ia64_libcall_value (enum machine_mode mode,
		    const_rtx fun ATTRIBUTE_UNUSED)
{
  return gen_rtx_REG (mode,
		      (((GET_MODE_CLASS (mode) == MODE_FLOAT
			 || GET_MODE_CLASS (mode) == MODE_COMPLEX_FLOAT)
			&& (mode) != TFmode)
		       ? FR_RET_FIRST : GR_RET_FIRST));
}

/* Worker function for FUNCTION_VALUE_REGNO_P.  */

static bool
ia64_function_value_regno_p (const unsigned int regno)
{
  return ((regno >= GR_RET_FIRST && regno <= GR_RET_LAST)
          || (regno >= FR_RET_FIRST && regno <= FR_RET_LAST));
}

/* This is called from dwarf2out.c via TARGET_ASM_OUTPUT_DWARF_DTPREL.
   We need to emit DTP-relative relocations.  */

static void
ia64_output_dwarf_dtprel (FILE *file, int size, rtx x)
{
  gcc_assert (size == 4 || size == 8);
  if (size == 4)
    fputs ("\tdata4.ua\t@dtprel(", file);
  else
    fputs ("\tdata8.ua\t@dtprel(", file);
  output_addr_const (file, x);
  fputs (")", file);
}

/* Print a memory address as an operand to reference that memory location.  */

/* ??? Do we need this?  It gets used only for 'a' operands.  We could perhaps
   also call this from ia64_print_operand for memory addresses.  */

static void
ia64_print_operand_address (FILE * stream ATTRIBUTE_UNUSED,
			    rtx address ATTRIBUTE_UNUSED)
{
}

/* Print an operand to an assembler instruction.
   C	Swap and print a comparison operator.
   D	Print an FP comparison operator.
   E    Print 32 - constant, for SImode shifts as extract.
   e    Print 64 - constant, for DImode rotates.
   F	A floating point constant 0.0 emitted as f0, or 1.0 emitted as f1, or
        a floating point register emitted normally.
   G	A floating point constant.
   I	Invert a predicate register by adding 1.
   J    Select the proper predicate register for a condition.
   j    Select the inverse predicate register for a condition.
   O	Append .acq for volatile load.
   P	Postincrement of a MEM.
   Q	Append .rel for volatile store.
   R	Print .s .d or nothing for a single, double or no truncation.
   S	Shift amount for shladd instruction.
   T	Print an 8-bit sign extended number (K) as a 32-bit unsigned number
	for Intel assembler.
   U	Print an 8-bit sign extended number (K) as a 64-bit unsigned number
	for Intel assembler.
   X	A pair of floating point registers.
   r	Print register name, or constant 0 as r0.  HP compatibility for
	Linux kernel.
   v    Print vector constant value as an 8-byte integer value.  */

static void
ia64_print_operand (FILE * file, rtx x, int code)
{
  const char *str;

  switch (code)
    {
    case 0:
      /* Handled below.  */
      break;

    case 'C':
      {
	enum rtx_code c = swap_condition (GET_CODE (x));
	fputs (GET_RTX_NAME (c), file);
	return;
      }

    case 'D':
      switch (GET_CODE (x))
	{
	case NE:
	  str = "neq";
	  break;
	case UNORDERED:
	  str = "unord";
	  break;
	case ORDERED:
	  str = "ord";
	  break;
	case UNLT:
	  str = "nge";
	  break;
	case UNLE:
	  str = "ngt";
	  break;
	case UNGT:
	  str = "nle";
	  break;
	case UNGE:
	  str = "nlt";
	  break;
	case UNEQ:
	case LTGT:
	  gcc_unreachable ();
	default:
	  str = GET_RTX_NAME (GET_CODE (x));
	  break;
	}
      fputs (str, file);
      return;

    case 'E':
      fprintf (file, HOST_WIDE_INT_PRINT_DEC, 32 - INTVAL (x));
      return;

    case 'e':
      fprintf (file, HOST_WIDE_INT_PRINT_DEC, 64 - INTVAL (x));
      return;

    case 'F':
      if (x == CONST0_RTX (GET_MODE (x)))
	str = reg_names [FR_REG (0)];
      else if (x == CONST1_RTX (GET_MODE (x)))
	str = reg_names [FR_REG (1)];
      else
	{
	  gcc_assert (GET_CODE (x) == REG);
	  str = reg_names [REGNO (x)];
	}
      fputs (str, file);
      return;

    case 'G':
      {
	long val[4];
	REAL_VALUE_TYPE rv;
	REAL_VALUE_FROM_CONST_DOUBLE (rv, x);
	real_to_target (val, &rv, GET_MODE (x));
	if (GET_MODE (x) == SFmode)
	  fprintf (file, "0x%08lx", val[0] & 0xffffffff);
	else if (GET_MODE (x) == DFmode)
	  fprintf (file, "0x%08lx%08lx", (WORDS_BIG_ENDIAN ? val[0] : val[1])
					  & 0xffffffff,
					 (WORDS_BIG_ENDIAN ? val[1] : val[0])
					  & 0xffffffff);
	else
	  output_operand_lossage ("invalid %%G mode");
      }
      return;

    case 'I':
      fputs (reg_names [REGNO (x) + 1], file);
      return;

    case 'J':
    case 'j':
      {
	unsigned int regno = REGNO (XEXP (x, 0));
	if (GET_CODE (x) == EQ)
	  regno += 1;
	if (code == 'j')
	  regno ^= 1;
        fputs (reg_names [regno], file);
      }
      return;

    case 'O':
      if (MEM_VOLATILE_P (x))
	fputs(".acq", file);
      return;

    case 'P':
      {
	HOST_WIDE_INT value;

	switch (GET_CODE (XEXP (x, 0)))
	  {
	  default:
	    return;

	  case POST_MODIFY:
	    x = XEXP (XEXP (XEXP (x, 0), 1), 1);
	    if (GET_CODE (x) == CONST_INT)
	      value = INTVAL (x);
	    else
	      {
		gcc_assert (GET_CODE (x) == REG);
		fprintf (file, ", %s", reg_names[REGNO (x)]);
		return;
	      }
	    break;

	  case POST_INC:
	    value = GET_MODE_SIZE (GET_MODE (x));
	    break;

	  case POST_DEC:
	    value = - (HOST_WIDE_INT) GET_MODE_SIZE (GET_MODE (x));
	    break;
	  }

	fprintf (file, ", " HOST_WIDE_INT_PRINT_DEC, value);
	return;
      }

    case 'Q':
      if (MEM_VOLATILE_P (x))
	fputs(".rel", file);
      return;

    case 'R':
      if (x == CONST0_RTX (GET_MODE (x)))
	fputs(".s", file);
      else if (x == CONST1_RTX (GET_MODE (x)))
	fputs(".d", file);
      else if (x == CONST2_RTX (GET_MODE (x)))
	;
      else
	output_operand_lossage ("invalid %%R value");
      return;

    case 'S':
      fprintf (file, "%d", exact_log2 (INTVAL (x)));
      return;

    case 'T':
      if (! TARGET_GNU_AS && GET_CODE (x) == CONST_INT)
	{
	  fprintf (file, "0x%x", (int) INTVAL (x) & 0xffffffff);
	  return;
	}
      break;

    case 'U':
      if (! TARGET_GNU_AS && GET_CODE (x) == CONST_INT)
	{
	  const char *prefix = "0x";
	  if (INTVAL (x) & 0x80000000)
	    {
	      fprintf (file, "0xffffffff");
	      prefix = "";
	    }
	  fprintf (file, "%s%x", prefix, (int) INTVAL (x) & 0xffffffff);
	  return;
	}
      break;

    case 'X':
      {
	unsigned int regno = REGNO (x);
	fprintf (file, "%s, %s", reg_names [regno], reg_names [regno + 1]);
      }
      return;

    case 'r':
      /* If this operand is the constant zero, write it as register zero.
	 Any register, zero, or CONST_INT value is OK here.  */
      if (GET_CODE (x) == REG)
	fputs (reg_names[REGNO (x)], file);
      else if (x == CONST0_RTX (GET_MODE (x)))
	fputs ("r0", file);
      else if (GET_CODE (x) == CONST_INT)
	output_addr_const (file, x);
      else
	output_operand_lossage ("invalid %%r value");
      return;

    case 'v':
      gcc_assert (GET_CODE (x) == CONST_VECTOR);
      x = simplify_subreg (DImode, x, GET_MODE (x), 0);
      break;

    case '+':
      {
	const char *which;

	/* For conditional branches, returns or calls, substitute
	   sptk, dptk, dpnt, or spnt for %s.  */
	x = find_reg_note (current_output_insn, REG_BR_PROB, 0);
	if (x)
	  {
	    int pred_val = XINT (x, 0);

	    /* Guess top and bottom 10% statically predicted.  */
	    if (pred_val < REG_BR_PROB_BASE / 50
		&& br_prob_note_reliable_p (x))
	      which = ".spnt";
	    else if (pred_val < REG_BR_PROB_BASE / 2)
	      which = ".dpnt";
	    else if (pred_val < REG_BR_PROB_BASE / 100 * 98
		     || !br_prob_note_reliable_p (x))
	      which = ".dptk";
	    else
	      which = ".sptk";
	  }
	else if (CALL_P (current_output_insn))
	  which = ".sptk";
	else
	  which = ".dptk";

	fputs (which, file);
	return;
      }

    case ',':
      x = current_insn_predicate;
      if (x)
	{
	  unsigned int regno = REGNO (XEXP (x, 0));
	  if (GET_CODE (x) == EQ)
	    regno += 1;
          fprintf (file, "(%s) ", reg_names [regno]);
	}
      return;

    default:
      output_operand_lossage ("ia64_print_operand: unknown code");
      return;
    }

  switch (GET_CODE (x))
    {
      /* This happens for the spill/restore instructions.  */
    case POST_INC:
    case POST_DEC:
    case POST_MODIFY:
      x = XEXP (x, 0);
      /* ... fall through ...  */

    case REG:
      fputs (reg_names [REGNO (x)], file);
      break;

    case MEM:
      {
	rtx addr = XEXP (x, 0);
	if (GET_RTX_CLASS (GET_CODE (addr)) == RTX_AUTOINC)
	  addr = XEXP (addr, 0);
	fprintf (file, "[%s]", reg_names [REGNO (addr)]);
	break;
      }

    default:
      output_addr_const (file, x);
      break;
    }

  return;
}

/* Worker function for TARGET_PRINT_OPERAND_PUNCT_VALID_P.  */

static bool
ia64_print_operand_punct_valid_p (unsigned char code)
{
  return (code == '+' || code == ',');
}

/* Compute a (partial) cost for rtx X.  Return true if the complete
   cost has been computed, and false if subexpressions should be
   scanned.  In either case, *TOTAL contains the cost result.  */
/* ??? This is incomplete.  */

static bool
ia64_rtx_costs (rtx x, int code, int outer_code, int opno ATTRIBUTE_UNUSED,
		int *total, bool speed ATTRIBUTE_UNUSED)
{
  switch (code)
    {
    case CONST_INT:
      switch (outer_code)
        {
        case SET:
	  *total = satisfies_constraint_J (x) ? 0 : COSTS_N_INSNS (1);
	  return true;
        case PLUS:
	  if (satisfies_constraint_I (x))
	    *total = 0;
	  else if (satisfies_constraint_J (x))
	    *total = 1;
	  else
	    *total = COSTS_N_INSNS (1);
	  return true;
        default:
	  if (satisfies_constraint_K (x) || satisfies_constraint_L (x))
	    *total = 0;
	  else
	    *total = COSTS_N_INSNS (1);
	  return true;
	}

    case CONST_DOUBLE:
      *total = COSTS_N_INSNS (1);
      return true;

    case CONST:
    case SYMBOL_REF:
    case LABEL_REF:
      *total = COSTS_N_INSNS (3);
      return true;

    case FMA:
      *total = COSTS_N_INSNS (4);
      return true;

    case MULT:
      /* For multiplies wider than HImode, we have to go to the FPU,
         which normally involves copies.  Plus there's the latency
         of the multiply itself, and the latency of the instructions to
         transfer integer regs to FP regs.  */
      if (FLOAT_MODE_P (GET_MODE (x)))
	*total = COSTS_N_INSNS (4);
      else if (GET_MODE_SIZE (GET_MODE (x)) > 2)
        *total = COSTS_N_INSNS (10);
      else
	*total = COSTS_N_INSNS (2);
      return true;

    case PLUS:
    case MINUS:
      if (FLOAT_MODE_P (GET_MODE (x)))
	{
	  *total = COSTS_N_INSNS (4);
	  return true;
	}
      /* FALLTHRU */

    case ASHIFT:
    case ASHIFTRT:
    case LSHIFTRT:
      *total = COSTS_N_INSNS (1);
      return true;

    case DIV:
    case UDIV:
    case MOD:
    case UMOD:
      /* We make divide expensive, so that divide-by-constant will be
         optimized to a multiply.  */
      *total = COSTS_N_INSNS (60);
      return true;

    default:
      return false;
    }
}

/* Calculate the cost of moving data from a register in class FROM to
   one in class TO, using MODE.  */

static int
ia64_register_move_cost (enum machine_mode mode, reg_class_t from,
			 reg_class_t to)
{
  /* ADDL_REGS is the same as GR_REGS for movement purposes.  */
  if (to == ADDL_REGS)
    to = GR_REGS;
  if (from == ADDL_REGS)
    from = GR_REGS;

  /* All costs are symmetric, so reduce cases by putting the
     lower number class as the destination.  */
  if (from < to)
    {
      reg_class_t tmp = to;
      to = from, from = tmp;
    }

  /* Moving from FR<->GR in XFmode must be more expensive than 2,
     so that we get secondary memory reloads.  Between FR_REGS,
     we have to make this at least as expensive as memory_move_cost
     to avoid spectacularly poor register class preferencing.  */
  if (mode == XFmode || mode == RFmode)
    {
      if (to != GR_REGS || from != GR_REGS)
        return memory_move_cost (mode, to, false);
      else
	return 3;
    }

  switch (to)
    {
    case PR_REGS:
      /* Moving between PR registers takes two insns.  */
      if (from == PR_REGS)
	return 3;
      /* Moving between PR and anything but GR is impossible.  */
      if (from != GR_REGS)
	return memory_move_cost (mode, to, false);
      break;

    case BR_REGS:
      /* Moving between BR and anything but GR is impossible.  */
      if (from != GR_REGS && from != GR_AND_BR_REGS)
	return memory_move_cost (mode, to, false);
      break;

    case AR_I_REGS:
    case AR_M_REGS:
      /* Moving between AR and anything but GR is impossible.  */
      if (from != GR_REGS)
	return memory_move_cost (mode, to, false);
      break;

    case GR_REGS:
    case FR_REGS:
    case FP_REGS:
    case GR_AND_FR_REGS:
    case GR_AND_BR_REGS:
    case ALL_REGS:
      break;

    default:
      gcc_unreachable ();
    }

  return 2;
}

/* Calculate the cost of moving data of MODE from a register to or from
   memory.  */

static int
ia64_memory_move_cost (enum machine_mode mode ATTRIBUTE_UNUSED,
		       reg_class_t rclass,
		       bool in ATTRIBUTE_UNUSED)
{
  if (rclass == GENERAL_REGS
      || rclass == FR_REGS
      || rclass == FP_REGS
      || rclass == GR_AND_FR_REGS)
    return 4;
  else
    return 10;
}

/* Implement TARGET_PREFERRED_RELOAD_CLASS.  Place additional restrictions
   on RCLASS to use when copying X into that class.  */

static reg_class_t
ia64_preferred_reload_class (rtx x, reg_class_t rclass)
{
  switch (rclass)
    {
    case FR_REGS:
    case FP_REGS:
      /* Don't allow volatile mem reloads into floating point registers.
	 This is defined to force reload to choose the r/m case instead
	 of the f/f case when reloading (set (reg fX) (mem/v)).  */
      if (MEM_P (x) && MEM_VOLATILE_P (x))
	return NO_REGS;
      
      /* Force all unrecognized constants into the constant pool.  */
      if (CONSTANT_P (x))
	return NO_REGS;
      break;

    case AR_M_REGS:
    case AR_I_REGS:
      if (!OBJECT_P (x))
	return NO_REGS;
      break;

    default:
      break;
    }

  return rclass;
}

/* This function returns the register class required for a secondary
   register when copying between one of the registers in RCLASS, and X,
   using MODE.  A return value of NO_REGS means that no secondary register
   is required.  */

enum reg_class
ia64_secondary_reload_class (enum reg_class rclass,
			     enum machine_mode mode ATTRIBUTE_UNUSED, rtx x)
{
  int regno = -1;

  if (GET_CODE (x) == REG || GET_CODE (x) == SUBREG)
    regno = true_regnum (x);

  switch (rclass)
    {
    case BR_REGS:
    case AR_M_REGS:
    case AR_I_REGS:
      /* ??? BR<->BR register copies can happen due to a bad gcse/cse/global
	 interaction.  We end up with two pseudos with overlapping lifetimes
	 both of which are equiv to the same constant, and both which need
	 to be in BR_REGS.  This seems to be a cse bug.  cse_basic_block_end
	 changes depending on the path length, which means the qty_first_reg
	 check in make_regs_eqv can give different answers at different times.
	 At some point I'll probably need a reload_indi pattern to handle
	 this.

	 We can also get GR_AND_FR_REGS to BR_REGS/AR_REGS copies, where we
	 wound up with a FP register from GR_AND_FR_REGS.  Extend that to all
	 non-general registers for good measure.  */
      if (regno >= 0 && ! GENERAL_REGNO_P (regno))
	return GR_REGS;

      /* This is needed if a pseudo used as a call_operand gets spilled to a
	 stack slot.  */
      if (GET_CODE (x) == MEM)
	return GR_REGS;
      break;

    case FR_REGS:
    case FP_REGS:
      /* Need to go through general registers to get to other class regs.  */
      if (regno >= 0 && ! (FR_REGNO_P (regno) || GENERAL_REGNO_P (regno)))
	return GR_REGS;

      /* This can happen when a paradoxical subreg is an operand to the
	 muldi3 pattern.  */
      /* ??? This shouldn't be necessary after instruction scheduling is
	 enabled, because paradoxical subregs are not accepted by
	 register_operand when INSN_SCHEDULING is defined.  Or alternatively,
	 stop the paradoxical subreg stupidity in the *_operand functions
	 in recog.c.  */
      if (GET_CODE (x) == MEM
	  && (GET_MODE (x) == SImode || GET_MODE (x) == HImode
	      || GET_MODE (x) == QImode))
	return GR_REGS;

      /* This can happen because of the ior/and/etc patterns that accept FP
	 registers as operands.  If the third operand is a constant, then it
	 needs to be reloaded into a FP register.  */
      if (GET_CODE (x) == CONST_INT)
	return GR_REGS;

      /* This can happen because of register elimination in a muldi3 insn.
	 E.g. `26107 * (unsigned long)&u'.  */
      if (GET_CODE (x) == PLUS)
	return GR_REGS;
      break;

    case PR_REGS:
      /* ??? This happens if we cse/gcse a BImode value across a call,
	 and the function has a nonlocal goto.  This is because global
	 does not allocate call crossing pseudos to hard registers when
	 crtl->has_nonlocal_goto is true.  This is relatively
	 common for C++ programs that use exceptions.  To reproduce,
	 return NO_REGS and compile libstdc++.  */
      if (GET_CODE (x) == MEM)
	return GR_REGS;

      /* This can happen when we take a BImode subreg of a DImode value,
	 and that DImode value winds up in some non-GR register.  */
      if (regno >= 0 && ! GENERAL_REGNO_P (regno) && ! PR_REGNO_P (regno))
	return GR_REGS;
      break;

    default:
      break;
    }

  return NO_REGS;
}


/* Implement targetm.unspec_may_trap_p hook.  */
static int
ia64_unspec_may_trap_p (const_rtx x, unsigned flags)
{
  switch (XINT (x, 1))
    {
    case UNSPEC_LDA:
    case UNSPEC_LDS:
    case UNSPEC_LDSA:
    case UNSPEC_LDCCLR:
    case UNSPEC_CHKACLR:
    case UNSPEC_CHKS:
      /* These unspecs are just wrappers.  */
      return may_trap_p_1 (XVECEXP (x, 0, 0), flags);
    }

  return default_unspec_may_trap_p (x, flags);
}


/* Parse the -mfixed-range= option string.  */

static void
fix_range (const char *const_str)
{
  int i, first, last;
  char *str, *dash, *comma;

  /* str must be of the form REG1'-'REG2{,REG1'-'REG} where REG1 and
     REG2 are either register names or register numbers.  The effect
     of this option is to mark the registers in the range from REG1 to
     REG2 as ``fixed'' so they won't be used by the compiler.  This is
     used, e.g., to ensure that kernel mode code doesn't use f32-f127.  */

  i = strlen (const_str);
  str = (char *) alloca (i + 1);
  memcpy (str, const_str, i + 1);

  while (1)
    {
      dash = strchr (str, '-');
      if (!dash)
	{
	  warning (0, "value of -mfixed-range must have form REG1-REG2");
	  return;
	}
      *dash = '\0';

      comma = strchr (dash + 1, ',');
      if (comma)
	*comma = '\0';

      first = decode_reg_name (str);
      if (first < 0)
	{
	  warning (0, "unknown register name: %s", str);
	  return;
	}

      last = decode_reg_name (dash + 1);
      if (last < 0)
	{
	  warning (0, "unknown register name: %s", dash + 1);
	  return;
	}

      *dash = '-';

      if (first > last)
	{
	  warning (0, "%s-%s is an empty range", str, dash + 1);
	  return;
	}

      for (i = first; i <= last; ++i)
	fixed_regs[i] = call_used_regs[i] = 1;

      if (!comma)
	break;

      *comma = ',';
      str = comma + 1;
    }
}

/* Implement TARGET_OPTION_OVERRIDE.  */

static void
ia64_option_override (void)
{
  unsigned int i;
  cl_deferred_option *opt;
  vec<cl_deferred_option> *v
    = (vec<cl_deferred_option> *) ia64_deferred_options;

  if (v)
    FOR_EACH_VEC_ELT (*v, i, opt)
      {
	switch (opt->opt_index)
	  {
	  case OPT_mfixed_range_:
	    fix_range (opt->arg);
	    break;

	  default:
	    gcc_unreachable ();
	  }
      }

  if (TARGET_AUTO_PIC)
    target_flags |= MASK_CONST_GP;

  /* Numerous experiment shows that IRA based loop pressure
     calculation works better for RTL loop invariant motion on targets
     with enough (>= 32) registers.  It is an expensive optimization.
     So it is on only for peak performance.  */
  if (optimize >= 3)
    flag_ira_loop_pressure = 1;


  ia64_section_threshold = (global_options_set.x_g_switch_value
			    ? g_switch_value
			    : IA64_DEFAULT_GVALUE);

  init_machine_status = ia64_init_machine_status;

  if (align_functions <= 0)
    align_functions = 64;
  if (align_loops <= 0)
    align_loops = 32;
  if (TARGET_ABI_OPEN_VMS)
    flag_no_common = 1;

  ia64_override_options_after_change();
}

/* Implement targetm.override_options_after_change.  */

static void
ia64_override_options_after_change (void)
{
  if (optimize >= 3
      && !global_options_set.x_flag_selective_scheduling
      && !global_options_set.x_flag_selective_scheduling2)
    {
      flag_selective_scheduling2 = 1;
      flag_sel_sched_pipelining = 1;
    }
  if (mflag_sched_control_spec == 2)
    {
      /* Control speculation is on by default for the selective scheduler,
         but not for the Haifa scheduler.  */
      mflag_sched_control_spec = flag_selective_scheduling2 ? 1 : 0;
    }
  if (flag_sel_sched_pipelining && flag_auto_inc_dec)
    {
      /* FIXME: remove this when we'd implement breaking autoinsns as
         a transformation.  */
      flag_auto_inc_dec = 0;
    }
}

/* Initialize the record of emitted frame related registers.  */

void ia64_init_expanders (void)
{
  memset (&emitted_frame_related_regs, 0, sizeof (emitted_frame_related_regs));
}

static struct machine_function *
ia64_init_machine_status (void)
{
  return ggc_alloc_cleared_machine_function ();
}

static enum attr_itanium_class ia64_safe_itanium_class (rtx);
static enum attr_type ia64_safe_type (rtx);

static enum attr_itanium_class
ia64_safe_itanium_class (rtx insn)
{
  if (recog_memoized (insn) >= 0)
    return get_attr_itanium_class (insn);
  else if (DEBUG_INSN_P (insn))
    return ITANIUM_CLASS_IGNORE;
  else
    return ITANIUM_CLASS_UNKNOWN;
}

static enum attr_type
ia64_safe_type (rtx insn)
{
  if (recog_memoized (insn) >= 0)
    return get_attr_type (insn);
  else
    return TYPE_UNKNOWN;
}

/* The following collection of routines emit instruction group stop bits as
   necessary to avoid dependencies.  */

/* Need to track some additional registers as far as serialization is
   concerned so we can properly handle br.call and br.ret.  We could
   make these registers visible to gcc, but since these registers are
   never explicitly used in gcc generated code, it seems wasteful to
   do so (plus it would make the call and return patterns needlessly
   complex).  */
#define REG_RP		(BR_REG (0))
#define REG_AR_CFM	(FIRST_PSEUDO_REGISTER + 1)
/* This is used for volatile asms which may require a stop bit immediately
   before and after them.  */
#define REG_VOLATILE	(FIRST_PSEUDO_REGISTER + 2)
#define AR_UNAT_BIT_0	(FIRST_PSEUDO_REGISTER + 3)
#define NUM_REGS	(AR_UNAT_BIT_0 + 64)

/* For each register, we keep track of how it has been written in the
   current instruction group.

   If a register is written unconditionally (no qualifying predicate),
   WRITE_COUNT is set to 2 and FIRST_PRED is ignored.

   If a register is written if its qualifying predicate P is true, we
   set WRITE_COUNT to 1 and FIRST_PRED to P.  Later on, the same register
   may be written again by the complement of P (P^1) and when this happens,
   WRITE_COUNT gets set to 2.

   The result of this is that whenever an insn attempts to write a register
   whose WRITE_COUNT is two, we need to issue an insn group barrier first.

   If a predicate register is written by a floating-point insn, we set
   WRITTEN_BY_FP to true.

   If a predicate register is written by an AND.ORCM we set WRITTEN_BY_AND
   to true; if it was written by an OR.ANDCM we set WRITTEN_BY_OR to true.  */

#if GCC_VERSION >= 4000
#define RWS_FIELD_TYPE __extension__ unsigned short
#else
#define RWS_FIELD_TYPE unsigned int
#endif
struct reg_write_state
{
  RWS_FIELD_TYPE write_count : 2;
  RWS_FIELD_TYPE first_pred : 10;
  RWS_FIELD_TYPE written_by_fp : 1;
  RWS_FIELD_TYPE written_by_and : 1;
  RWS_FIELD_TYPE written_by_or : 1;
};

/* Cumulative info for the current instruction group.  */
struct reg_write_state rws_sum[NUM_REGS];
#ifdef ENABLE_CHECKING
/* Bitmap whether a register has been written in the current insn.  */
HARD_REG_ELT_TYPE rws_insn[(NUM_REGS + HOST_BITS_PER_WIDEST_FAST_INT - 1)
			   / HOST_BITS_PER_WIDEST_FAST_INT];

static inline void
rws_insn_set (int regno)
{
  gcc_assert (!TEST_HARD_REG_BIT (rws_insn, regno));
  SET_HARD_REG_BIT (rws_insn, regno);
}

static inline int
rws_insn_test (int regno)
{
  return TEST_HARD_REG_BIT (rws_insn, regno);
}
#else
/* When not checking, track just REG_AR_CFM and REG_VOLATILE.  */
unsigned char rws_insn[2];

static inline void
rws_insn_set (int regno)
{
  if (regno == REG_AR_CFM)
    rws_insn[0] = 1;
  else if (regno == REG_VOLATILE)
    rws_insn[1] = 1;
}

static inline int
rws_insn_test (int regno)
{
  if (regno == REG_AR_CFM)
    return rws_insn[0];
  if (regno == REG_VOLATILE)
    return rws_insn[1];
  return 0;
}
#endif

/* Indicates whether this is the first instruction after a stop bit,
   in which case we don't need another stop bit.  Without this,
   ia64_variable_issue will die when scheduling an alloc.  */
static int first_instruction;

/* Misc flags needed to compute RAW/WAW dependencies while we are traversing
   RTL for one instruction.  */
struct reg_flags
{
  unsigned int is_write : 1;	/* Is register being written?  */
  unsigned int is_fp : 1;	/* Is register used as part of an fp op?  */
  unsigned int is_branch : 1;	/* Is register used as part of a branch?  */
  unsigned int is_and : 1;	/* Is register used as part of and.orcm?  */
  unsigned int is_or : 1;	/* Is register used as part of or.andcm?  */
  unsigned int is_sibcall : 1;	/* Is this a sibling or normal call?  */
};

static void rws_update (int, struct reg_flags, int);
static int rws_access_regno (int, struct reg_flags, int);
static int rws_access_reg (rtx, struct reg_flags, int);
static void update_set_flags (rtx, struct reg_flags *);
static int set_src_needs_barrier (rtx, struct reg_flags, int);
static int rtx_needs_barrier (rtx, struct reg_flags, int);
static void init_insn_group_barriers (void);
static int group_barrier_needed (rtx);
static int safe_group_barrier_needed (rtx);
static int in_safe_group_barrier;

/* Update *RWS for REGNO, which is being written by the current instruction,
   with predicate PRED, and associated register flags in FLAGS.  */

static void
rws_update (int regno, struct reg_flags flags, int pred)
{
  if (pred)
    rws_sum[regno].write_count++;
  else
    rws_sum[regno].write_count = 2;
  rws_sum[regno].written_by_fp |= flags.is_fp;
  /* ??? Not tracking and/or across differing predicates.  */
  rws_sum[regno].written_by_and = flags.is_and;
  rws_sum[regno].written_by_or = flags.is_or;
  rws_sum[regno].first_pred = pred;
}

/* Handle an access to register REGNO of type FLAGS using predicate register
   PRED.  Update rws_sum array.  Return 1 if this access creates
   a dependency with an earlier instruction in the same group.  */

static int
rws_access_regno (int regno, struct reg_flags flags, int pred)
{
  int need_barrier = 0;

  gcc_assert (regno < NUM_REGS);

  if (! PR_REGNO_P (regno))
    flags.is_and = flags.is_or = 0;

  if (flags.is_write)
    {
      int write_count;

      rws_insn_set (regno);
      write_count = rws_sum[regno].write_count;

      switch (write_count)
	{
	case 0:
	  /* The register has not been written yet.  */
	  if (!in_safe_group_barrier)
	    rws_update (regno, flags, pred);
	  break;

	case 1:
	  /* The register has been written via a predicate.  Treat
	     it like a unconditional write and do not try to check
	     for complementary pred reg in earlier write.  */
	  if (flags.is_and && rws_sum[regno].written_by_and)
	    ;
	  else if (flags.is_or && rws_sum[regno].written_by_or)
	    ;
	  else
	    need_barrier = 1;
	  if (!in_safe_group_barrier)
	    rws_update (regno, flags, pred);
	  break;

	case 2:
	  /* The register has been unconditionally written already.  We
	     need a barrier.  */
	  if (flags.is_and && rws_sum[regno].written_by_and)
	    ;
	  else if (flags.is_or && rws_sum[regno].written_by_or)
	    ;
	  else
	    need_barrier = 1;
	  if (!in_safe_group_barrier)
	    {
	      rws_sum[regno].written_by_and = flags.is_and;
	      rws_sum[regno].written_by_or = flags.is_or;
	    }
	  break;

	default:
	  gcc_unreachable ();
	}
    }
  else
    {
      if (flags.is_branch)
	{
	  /* Branches have several RAW exceptions that allow to avoid
	     barriers.  */

	  if (REGNO_REG_CLASS (regno) == BR_REGS || regno == AR_PFS_REGNUM)
	    /* RAW dependencies on branch regs are permissible as long
	       as the writer is a non-branch instruction.  Since we
	       never generate code that uses a branch register written
	       by a branch instruction, handling this case is
	       easy.  */
	    return 0;

	  if (REGNO_REG_CLASS (regno) == PR_REGS
	      && ! rws_sum[regno].written_by_fp)
	    /* The predicates of a branch are available within the
	       same insn group as long as the predicate was written by
	       something other than a floating-point instruction.  */
	    return 0;
	}

      if (flags.is_and && rws_sum[regno].written_by_and)
	return 0;
      if (flags.is_or && rws_sum[regno].written_by_or)
	return 0;

      switch (rws_sum[regno].write_count)
	{
	case 0:
	  /* The register has not been written yet.  */
	  break;

	case 1:
	  /* The register has been written via a predicate, assume we
	     need a barrier (don't check for complementary regs).  */
	  need_barrier = 1;
	  break;

	case 2:
	  /* The register has been unconditionally written already.  We
	     need a barrier.  */
	  need_barrier = 1;
	  break;

	default:
	  gcc_unreachable ();
	}
    }

  return need_barrier;
}

static int
rws_access_reg (rtx reg, struct reg_flags flags, int pred)
{
  int regno = REGNO (reg);
  int n = HARD_REGNO_NREGS (REGNO (reg), GET_MODE (reg));

  if (n == 1)
    return rws_access_regno (regno, flags, pred);
  else
    {
      int need_barrier = 0;
      while (--n >= 0)
	need_barrier |= rws_access_regno (regno + n, flags, pred);
      return need_barrier;
    }
}

/* Examine X, which is a SET rtx, and update the flags, the predicate, and
   the condition, stored in *PFLAGS, *PPRED and *PCOND.  */

static void
update_set_flags (rtx x, struct reg_flags *pflags)
{
  rtx src = SET_SRC (x);

  switch (GET_CODE (src))
    {
    case CALL:
      return;

    case IF_THEN_ELSE:
      /* There are four cases here:
	 (1) The destination is (pc), in which case this is a branch,
	 nothing here applies.
	 (2) The destination is ar.lc, in which case this is a
	 doloop_end_internal,
	 (3) The destination is an fp register, in which case this is
	 an fselect instruction.
	 (4) The condition has (unspec [(reg)] UNSPEC_LDC), in which case 
	 this is a check load.
	 In all cases, nothing we do in this function applies.  */
      return;

    default:
      if (COMPARISON_P (src)
	  && SCALAR_FLOAT_MODE_P (GET_MODE (XEXP (src, 0))))
	/* Set pflags->is_fp to 1 so that we know we're dealing
	   with a floating point comparison when processing the
	   destination of the SET.  */
	pflags->is_fp = 1;

      /* Discover if this is a parallel comparison.  We only handle
	 and.orcm and or.andcm at present, since we must retain a
	 strict inverse on the predicate pair.  */
      else if (GET_CODE (src) == AND)
	pflags->is_and = 1;
      else if (GET_CODE (src) == IOR)
	pflags->is_or = 1;

      break;
    }
}

/* Subroutine of rtx_needs_barrier; this function determines whether the
   source of a given SET rtx found in X needs a barrier.  FLAGS and PRED
   are as in rtx_needs_barrier.  COND is an rtx that holds the condition
   for this insn.  */

static int
set_src_needs_barrier (rtx x, struct reg_flags flags, int pred)
{
  int need_barrier = 0;
  rtx dst;
  rtx src = SET_SRC (x);

  if (GET_CODE (src) == CALL)
    /* We don't need to worry about the result registers that
       get written by subroutine call.  */
    return rtx_needs_barrier (src, flags, pred);
  else if (SET_DEST (x) == pc_rtx)
    {
      /* X is a conditional branch.  */
      /* ??? This seems redundant, as the caller sets this bit for
	 all JUMP_INSNs.  */
      if (!ia64_spec_check_src_p (src))
	flags.is_branch = 1;
      return rtx_needs_barrier (src, flags, pred);
    }

  if (ia64_spec_check_src_p (src))
    /* Avoid checking one register twice (in condition 
       and in 'then' section) for ldc pattern.  */
    {
      gcc_assert (REG_P (XEXP (src, 2)));
      need_barrier = rtx_needs_barrier (XEXP (src, 2), flags, pred);
		  
      /* We process MEM below.  */
      src = XEXP (src, 1);
    }

  need_barrier |= rtx_needs_barrier (src, flags, pred);

  dst = SET_DEST (x);
  if (GET_CODE (dst) == ZERO_EXTRACT)
    {
      need_barrier |= rtx_needs_barrier (XEXP (dst, 1), flags, pred);
      need_barrier |= rtx_needs_barrier (XEXP (dst, 2), flags, pred);
    }
  return need_barrier;
}

/* Handle an access to rtx X of type FLAGS using predicate register
   PRED.  Return 1 if this access creates a dependency with an earlier
   instruction in the same group.  */

static int
rtx_needs_barrier (rtx x, struct reg_flags flags, int pred)
{
  int i, j;
  int is_complemented = 0;
  int need_barrier = 0;
  const char *format_ptr;
  struct reg_flags new_flags;
  rtx cond;

  if (! x)
    return 0;

  new_flags = flags;

  switch (GET_CODE (x))
    {
    case SET:
      update_set_flags (x, &new_flags);
      need_barrier = set_src_needs_barrier (x, new_flags, pred);
      if (GET_CODE (SET_SRC (x)) != CALL)
	{
	  new_flags.is_write = 1;
	  need_barrier |= rtx_needs_barrier (SET_DEST (x), new_flags, pred);
	}
      break;

    case CALL:
      new_flags.is_write = 0;
      need_barrier |= rws_access_regno (AR_EC_REGNUM, new_flags, pred);

      /* Avoid multiple register writes, in case this is a pattern with
	 multiple CALL rtx.  This avoids a failure in rws_access_reg.  */
      if (! flags.is_sibcall && ! rws_insn_test (REG_AR_CFM))
	{
	  new_flags.is_write = 1;
	  need_barrier |= rws_access_regno (REG_RP, new_flags, pred);
	  need_barrier |= rws_access_regno (AR_PFS_REGNUM, new_flags, pred);
	  need_barrier |= rws_access_regno (REG_AR_CFM, new_flags, pred);
	}
      break;

    case COND_EXEC:
      /* X is a predicated instruction.  */

      cond = COND_EXEC_TEST (x);
      gcc_assert (!pred);
      need_barrier = rtx_needs_barrier (cond, flags, 0);

      if (GET_CODE (cond) == EQ)
	is_complemented = 1;
      cond = XEXP (cond, 0);
      gcc_assert (GET_CODE (cond) == REG
		  && REGNO_REG_CLASS (REGNO (cond)) == PR_REGS);
      pred = REGNO (cond);
      if (is_complemented)
	++pred;

      need_barrier |= rtx_needs_barrier (COND_EXEC_CODE (x), flags, pred);
      return need_barrier;

    case CLOBBER:
    case USE:
      /* Clobber & use are for earlier compiler-phases only.  */
      break;

    case ASM_OPERANDS:
    case ASM_INPUT:
      /* We always emit stop bits for traditional asms.  We emit stop bits
	 for volatile extended asms if TARGET_VOL_ASM_STOP is true.  */
      if (GET_CODE (x) != ASM_OPERANDS
	  || (MEM_VOLATILE_P (x) && TARGET_VOL_ASM_STOP))
	{
	  /* Avoid writing the register multiple times if we have multiple
	     asm outputs.  This avoids a failure in rws_access_reg.  */
	  if (! rws_insn_test (REG_VOLATILE))
	    {
	      new_flags.is_write = 1;
	      rws_access_regno (REG_VOLATILE, new_flags, pred);
	    }
	  return 1;
	}

      /* For all ASM_OPERANDS, we must traverse the vector of input operands.
	 We cannot just fall through here since then we would be confused
	 by the ASM_INPUT rtx inside ASM_OPERANDS, which do not indicate
	 traditional asms unlike their normal usage.  */

      for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; --i)
	if (rtx_needs_barrier (ASM_OPERANDS_INPUT (x, i), flags, pred))
	  need_barrier = 1;
      break;

    case PARALLEL:
      for (i = XVECLEN (x, 0) - 1; i >= 0; --i)
	{
	  rtx pat = XVECEXP (x, 0, i);
	  switch (GET_CODE (pat))
	    {
	    case SET:
	      update_set_flags (pat, &new_flags);
	      need_barrier |= set_src_needs_barrier (pat, new_flags, pred);
	      break;

	    case USE:
	    case CALL:
	    case ASM_OPERANDS:
	      need_barrier |= rtx_needs_barrier (pat, flags, pred);
	      break;

	    case CLOBBER:
	      if (REG_P (XEXP (pat, 0))
		  && extract_asm_operands (x) != NULL_RTX
		  && REGNO (XEXP (pat, 0)) != AR_UNAT_REGNUM)
		{
		  new_flags.is_write = 1;
		  need_barrier |= rtx_needs_barrier (XEXP (pat, 0),
						     new_flags, pred);
		  new_flags = flags;
		}
	      break;

	    case RETURN:
	      break;

	    default:
	      gcc_unreachable ();
	    }
	}
      for (i = XVECLEN (x, 0) - 1; i >= 0; --i)
	{
	  rtx pat = XVECEXP (x, 0, i);
	  if (GET_CODE (pat) == SET)
	    {
	      if (GET_CODE (SET_SRC (pat)) != CALL)
		{
		  new_flags.is_write = 1;
		  need_barrier |= rtx_needs_barrier (SET_DEST (pat), new_flags,
						     pred);
		}
	    }
	  else if (GET_CODE (pat) == CLOBBER || GET_CODE (pat) == RETURN)
	    need_barrier |= rtx_needs_barrier (pat, flags, pred);
	}
      break;

    case SUBREG:
      need_barrier |= rtx_needs_barrier (SUBREG_REG (x), flags, pred);
      break;
    case REG:
      if (REGNO (x) == AR_UNAT_REGNUM)
	{
	  for (i = 0; i < 64; ++i)
	    need_barrier |= rws_access_regno (AR_UNAT_BIT_0 + i, flags, pred);
	}
      else
	need_barrier = rws_access_reg (x, flags, pred);
      break;

    case MEM:
      /* Find the regs used in memory address computation.  */
      new_flags.is_write = 0;
      need_barrier = rtx_needs_barrier (XEXP (x, 0), new_flags, pred);
      break;

    case CONST_INT:   case CONST_DOUBLE:  case CONST_VECTOR:
    case SYMBOL_REF:  case LABEL_REF:     case CONST:
      break;

      /* Operators with side-effects.  */
    case POST_INC:    case POST_DEC:
      gcc_assert (GET_CODE (XEXP (x, 0)) == REG);

      new_flags.is_write = 0;
      need_barrier  = rws_access_reg (XEXP (x, 0), new_flags, pred);
      new_flags.is_write = 1;
      need_barrier |= rws_access_reg (XEXP (x, 0), new_flags, pred);
      break;

    case POST_MODIFY:
      gcc_assert (GET_CODE (XEXP (x, 0)) == REG);

      new_flags.is_write = 0;
      need_barrier  = rws_access_reg (XEXP (x, 0), new_flags, pred);
      need_barrier |= rtx_needs_barrier (XEXP (x, 1), new_flags, pred);
      new_flags.is_write = 1;
      need_barrier |= rws_access_reg (XEXP (x, 0), new_flags, pred);
      break;

      /* Handle common unary and binary ops for efficiency.  */
    case COMPARE:  case PLUS:    case MINUS:   case MULT:      case DIV:
    case MOD:      case UDIV:    case UMOD:    case AND:       case IOR:
    case XOR:      case ASHIFT:  case ROTATE:  case ASHIFTRT:  case LSHIFTRT:
    case ROTATERT: case SMIN:    case SMAX:    case UMIN:      case UMAX:
    case NE:       case EQ:      case GE:      case GT:        case LE:
    case LT:       case GEU:     case GTU:     case LEU:       case LTU:
      need_barrier = rtx_needs_barrier (XEXP (x, 0), new_flags, pred);
      need_barrier |= rtx_needs_barrier (XEXP (x, 1), new_flags, pred);
      break;

    case NEG:      case NOT:	        case SIGN_EXTEND:     case ZERO_EXTEND:
    case TRUNCATE: case FLOAT_EXTEND:   case FLOAT_TRUNCATE:  case FLOAT:
    case FIX:      case UNSIGNED_FLOAT: case UNSIGNED_FIX:    case ABS:
    case SQRT:     case FFS:		case POPCOUNT:
      need_barrier = rtx_needs_barrier (XEXP (x, 0), flags, pred);
      break;

    case VEC_SELECT:
      /* VEC_SELECT's second argument is a PARALLEL with integers that
	 describe the elements selected.  On ia64, those integers are
	 always constants.  Avoid walking the PARALLEL so that we don't
	 get confused with "normal" parallels and then die.  */
      need_barrier = rtx_needs_barrier (XEXP (x, 0), flags, pred);
      break;

    case UNSPEC:
      switch (XINT (x, 1))
	{
	case UNSPEC_LTOFF_DTPMOD:
	case UNSPEC_LTOFF_DTPREL:
	case UNSPEC_DTPREL:
	case UNSPEC_LTOFF_TPREL:
	case UNSPEC_TPREL:
	case UNSPEC_PRED_REL_MUTEX:
	case UNSPEC_PIC_CALL:
        case UNSPEC_MF:
        case UNSPEC_FETCHADD_ACQ:
        case UNSPEC_FETCHADD_REL:
	case UNSPEC_BSP_VALUE:
	case UNSPEC_FLUSHRS:
	case UNSPEC_BUNDLE_SELECTOR:
          break;

	case UNSPEC_GR_SPILL:
	case UNSPEC_GR_RESTORE:
	  {
	    HOST_WIDE_INT offset = INTVAL (XVECEXP (x, 0, 1));
	    HOST_WIDE_INT bit = (offset >> 3) & 63;

	    need_barrier = rtx_needs_barrier (XVECEXP (x, 0, 0), flags, pred);
	    new_flags.is_write = (XINT (x, 1) == UNSPEC_GR_SPILL);
	    need_barrier |= rws_access_regno (AR_UNAT_BIT_0 + bit,
					      new_flags, pred);
	    break;
	  }

	case UNSPEC_FR_SPILL:
	case UNSPEC_FR_RESTORE:
	case UNSPEC_GETF_EXP:
	case UNSPEC_SETF_EXP:
        case UNSPEC_ADDP4:
	case UNSPEC_FR_SQRT_RECIP_APPROX:
	case UNSPEC_FR_SQRT_RECIP_APPROX_RES:
	case UNSPEC_LDA:
	case UNSPEC_LDS:
	case UNSPEC_LDS_A:
	case UNSPEC_LDSA:
	case UNSPEC_CHKACLR:
        case UNSPEC_CHKS:
	  need_barrier = rtx_needs_barrier (XVECEXP (x, 0, 0), flags, pred);
	  break;

	case UNSPEC_FR_RECIP_APPROX:
	case UNSPEC_SHRP:
	case UNSPEC_COPYSIGN:
	case UNSPEC_FR_RECIP_APPROX_RES:
	  need_barrier = rtx_needs_barrier (XVECEXP (x, 0, 0), flags, pred);
	  need_barrier |= rtx_needs_barrier (XVECEXP (x, 0, 1), flags, pred);
	  break;

        case UNSPEC_CMPXCHG_ACQ:
        case UNSPEC_CMPXCHG_REL:
	  need_barrier = rtx_needs_barrier (XVECEXP (x, 0, 1), flags, pred);
	  need_barrier |= rtx_needs_barrier (XVECEXP (x, 0, 2), flags, pred);
	  break;

	default:
	  gcc_unreachable ();
	}
      break;

    case UNSPEC_VOLATILE:
      switch (XINT (x, 1))
	{
	case UNSPECV_ALLOC:
	  /* Alloc must always be the first instruction of a group.
	     We force this by always returning true.  */
	  /* ??? We might get better scheduling if we explicitly check for
	     input/local/output register dependencies, and modify the
	     scheduler so that alloc is always reordered to the start of
	     the current group.  We could then eliminate all of the
	     first_instruction code.  */
	  rws_access_regno (AR_PFS_REGNUM, flags, pred);

	  new_flags.is_write = 1;
	  rws_access_regno (REG_AR_CFM, new_flags, pred);
	  return 1;

	case UNSPECV_SET_BSP:
	case UNSPECV_PROBE_STACK_RANGE:
	  need_barrier = 1;
          break;

	case UNSPECV_BLOCKAGE:
	case UNSPECV_INSN_GROUP_BARRIER:
	case UNSPECV_BREAK:
	case UNSPECV_PSAC_ALL:
	case UNSPECV_PSAC_NORMAL:
	  return 0;

	case UNSPECV_PROBE_STACK_ADDRESS:
	  need_barrier = rtx_needs_barrier (XVECEXP (x, 0, 0), flags, pred);
	  break;

	default:
	  gcc_unreachable ();
	}
      break;

    case RETURN:
      new_flags.is_write = 0;
      need_barrier  = rws_access_regno (REG_RP, flags, pred);
      need_barrier |= rws_access_regno (AR_PFS_REGNUM, flags, pred);

      new_flags.is_write = 1;
      need_barrier |= rws_access_regno (AR_EC_REGNUM, new_flags, pred);
      need_barrier |= rws_access_regno (REG_AR_CFM, new_flags, pred);
      break;

    default:
      format_ptr = GET_RTX_FORMAT (GET_CODE (x));
      for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--)
	switch (format_ptr[i])
	  {
	  case '0':	/* unused field */
	  case 'i':	/* integer */
	  case 'n':	/* note */
	  case 'w':	/* wide integer */
	  case 's':	/* pointer to string */
	  case 'S':	/* optional pointer to string */
	    break;

	  case 'e':
	    if (rtx_needs_barrier (XEXP (x, i), flags, pred))
	      need_barrier = 1;
	    break;

	  case 'E':
	    for (j = XVECLEN (x, i) - 1; j >= 0; --j)
	      if (rtx_needs_barrier (XVECEXP (x, i, j), flags, pred))
		need_barrier = 1;
	    break;

	  default:
	    gcc_unreachable ();
	  }
      break;
    }
  return need_barrier;
}

/* Clear out the state for group_barrier_needed at the start of a
   sequence of insns.  */

static void
init_insn_group_barriers (void)
{
  memset (rws_sum, 0, sizeof (rws_sum));
  first_instruction = 1;
}

/* Given the current state, determine whether a group barrier (a stop bit) is
   necessary before INSN.  Return nonzero if so.  This modifies the state to
   include the effects of INSN as a side-effect.  */

static int
group_barrier_needed (rtx insn)
{
  rtx pat;
  int need_barrier = 0;
  struct reg_flags flags;

  memset (&flags, 0, sizeof (flags));
  switch (GET_CODE (insn))
    {
    case NOTE:
    case DEBUG_INSN:
      break;

    case BARRIER:
      /* A barrier doesn't imply an instruction group boundary.  */
      break;

    case CODE_LABEL:
      memset (rws_insn, 0, sizeof (rws_insn));
      return 1;

    case CALL_INSN:
      flags.is_branch = 1;
      flags.is_sibcall = SIBLING_CALL_P (insn);
      memset (rws_insn, 0, sizeof (rws_insn));

      /* Don't bundle a call following another call.  */
      if ((pat = prev_active_insn (insn)) && CALL_P (pat))
	{
	  need_barrier = 1;
	  break;
	}

      need_barrier = rtx_needs_barrier (PATTERN (insn), flags, 0);
      break;

    case JUMP_INSN:
      if (!ia64_spec_check_p (insn))
	flags.is_branch = 1;

      /* Don't bundle a jump following a call.  */
      if ((pat = prev_active_insn (insn)) && CALL_P (pat))
	{
	  need_barrier = 1;
	  break;
	}
      /* FALLTHRU */

    case INSN:
      if (GET_CODE (PATTERN (insn)) == USE
	  || GET_CODE (PATTERN (insn)) == CLOBBER)
	/* Don't care about USE and CLOBBER "insns"---those are used to
	   indicate to the optimizer that it shouldn't get rid of
	   certain operations.  */
	break;

      pat = PATTERN (insn);

      /* Ug.  Hack hacks hacked elsewhere.  */
      switch (recog_memoized (insn))
	{
	  /* We play dependency tricks with the epilogue in order
	     to get proper schedules.  Undo this for dv analysis.  */
	case CODE_FOR_epilogue_deallocate_stack:
	case CODE_FOR_prologue_allocate_stack:
	  pat = XVECEXP (pat, 0, 0);
	  break;

	  /* The pattern we use for br.cloop confuses the code above.
	     The second element of the vector is representative.  */
	case CODE_FOR_doloop_end_internal:
	  pat = XVECEXP (pat, 0, 1);
	  break;

	  /* Doesn't generate code.  */
	case CODE_FOR_pred_rel_mutex:
	case CODE_FOR_prologue_use:
	  return 0;

	default:
	  break;
	}

      memset (rws_insn, 0, sizeof (rws_insn));
      need_barrier = rtx_needs_barrier (pat, flags, 0);

      /* Check to see if the previous instruction was a volatile
	 asm.  */
      if (! need_barrier)
	need_barrier = rws_access_regno (REG_VOLATILE, flags, 0);

      break;

    default:
      gcc_unreachable ();
    }

  if (first_instruction && important_for_bundling_p (insn))
    {
      need_barrier = 0;
      first_instruction = 0;
    }

  return need_barrier;
}

/* Like group_barrier_needed, but do not clobber the current state.  */

static int
safe_group_barrier_needed (rtx insn)
{
  int saved_first_instruction;
  int t;

  saved_first_instruction = first_instruction;
  in_safe_group_barrier = 1;

  t = group_barrier_needed (insn);

  first_instruction = saved_first_instruction;
  in_safe_group_barrier = 0;

  return t;
}

/* Scan the current function and insert stop bits as necessary to
   eliminate dependencies.  This function assumes that a final
   instruction scheduling pass has been run which has already
   inserted most of the necessary stop bits.  This function only
   inserts new ones at basic block boundaries, since these are
   invisible to the scheduler.  */

static void
emit_insn_group_barriers (FILE *dump)
{
  rtx insn;
  rtx last_label = 0;
  int insns_since_last_label = 0;

  init_insn_group_barriers ();

  for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
    {
      if (LABEL_P (insn))
	{
	  if (insns_since_last_label)
	    last_label = insn;
	  insns_since_last_label = 0;
	}
      else if (NOTE_P (insn)
	       && NOTE_KIND (insn) == NOTE_INSN_BASIC_BLOCK)
	{
	  if (insns_since_last_label)
	    last_label = insn;
	  insns_since_last_label = 0;
	}
      else if (NONJUMP_INSN_P (insn)
	       && GET_CODE (PATTERN (insn)) == UNSPEC_VOLATILE
	       && XINT (PATTERN (insn), 1) == UNSPECV_INSN_GROUP_BARRIER)
	{
	  init_insn_group_barriers ();
	  last_label = 0;
	}
      else if (NONDEBUG_INSN_P (insn))
	{
	  insns_since_last_label = 1;

	  if (group_barrier_needed (insn))
	    {
	      if (last_label)
		{
		  if (dump)
		    fprintf (dump, "Emitting stop before label %d\n",
			     INSN_UID (last_label));
		  emit_insn_before (gen_insn_group_barrier (GEN_INT (3)), last_label);
		  insn = last_label;

		  init_insn_group_barriers ();
		  last_label = 0;
		}
	    }
	}
    }
}

/* Like emit_insn_group_barriers, but run if no final scheduling pass was run.
   This function has to emit all necessary group barriers.  */

static void
emit_all_insn_group_barriers (FILE *dump ATTRIBUTE_UNUSED)
{
  rtx insn;

  init_insn_group_barriers ();

  for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
    {
      if (BARRIER_P (insn))
	{
	  rtx last = prev_active_insn (insn);

	  if (! last)
	    continue;
	  if (JUMP_TABLE_DATA_P (last))
	    last = prev_active_insn (last);
	  if (recog_memoized (last) != CODE_FOR_insn_group_barrier)
	    emit_insn_after (gen_insn_group_barrier (GEN_INT (3)), last);

	  init_insn_group_barriers ();
	}
      else if (NONDEBUG_INSN_P (insn))
	{
	  if (recog_memoized (insn) == CODE_FOR_insn_group_barrier)
	    init_insn_group_barriers ();
	  else if (group_barrier_needed (insn))
	    {
	      emit_insn_before (gen_insn_group_barrier (GEN_INT (3)), insn);
	      init_insn_group_barriers ();
	      group_barrier_needed (insn);
	    }
	}
    }
}



/* Instruction scheduling support.  */

#define NR_BUNDLES 10

/* A list of names of all available bundles.  */

static const char *bundle_name [NR_BUNDLES] =
{
  ".mii",
  ".mmi",
  ".mfi",
  ".mmf",
#if NR_BUNDLES == 10
  ".bbb",
  ".mbb",
#endif
  ".mib",
  ".mmb",
  ".mfb",
  ".mlx"
};

/* Nonzero if we should insert stop bits into the schedule.  */

int ia64_final_schedule = 0;

/* Codes of the corresponding queried units: */

static int _0mii_, _0mmi_, _0mfi_, _0mmf_;
static int _0bbb_, _0mbb_, _0mib_, _0mmb_, _0mfb_, _0mlx_;

static int _1mii_, _1mmi_, _1mfi_, _1mmf_;
static int _1bbb_, _1mbb_, _1mib_, _1mmb_, _1mfb_, _1mlx_;

static int pos_1, pos_2, pos_3, pos_4, pos_5, pos_6;

/* The following variable value is an insn group barrier.  */

static rtx dfa_stop_insn;

/* The following variable value is the last issued insn.  */

static rtx last_scheduled_insn;

/* The following variable value is pointer to a DFA state used as
   temporary variable.  */

static state_t temp_dfa_state = NULL;

/* The following variable value is DFA state after issuing the last
   insn.  */

static state_t prev_cycle_state = NULL;

/* The following array element values are TRUE if the corresponding
   insn requires to add stop bits before it.  */

static char *stops_p = NULL;

/* The following variable is used to set up the mentioned above array.  */

static int stop_before_p = 0;

/* The following variable value is length of the arrays `clocks' and
   `add_cycles'. */

static int clocks_length;

/* The following variable value is number of data speculations in progress.  */
static int pending_data_specs = 0;

/* Number of memory references on current and three future processor cycles.  */
static char mem_ops_in_group[4];

/* Number of current processor cycle (from scheduler's point of view).  */
static int current_cycle;

static rtx ia64_single_set (rtx);
static void ia64_emit_insn_before (rtx, rtx);

/* Map a bundle number to its pseudo-op.  */

const char *
get_bundle_name (int b)
{
  return bundle_name[b];
}


/* Return the maximum number of instructions a cpu can issue.  */

static int
ia64_issue_rate (void)
{
  return 6;
}

/* Helper function - like single_set, but look inside COND_EXEC.  */

static rtx
ia64_single_set (rtx insn)
{
  rtx x = PATTERN (insn), ret;
  if (GET_CODE (x) == COND_EXEC)
    x = COND_EXEC_CODE (x);
  if (GET_CODE (x) == SET)
    return x;

  /* Special case here prologue_allocate_stack and epilogue_deallocate_stack.
     Although they are not classical single set, the second set is there just
     to protect it from moving past FP-relative stack accesses.  */
  switch (recog_memoized (insn))
    {
    case CODE_FOR_prologue_allocate_stack:
    case CODE_FOR_prologue_allocate_stack_pr:
    case CODE_FOR_epilogue_deallocate_stack:
    case CODE_FOR_epilogue_deallocate_stack_pr:
      ret = XVECEXP (x, 0, 0);
      break;

    default:
      ret = single_set_2 (insn, x);
      break;
    }

  return ret;
}

/* Adjust the cost of a scheduling dependency.
   Return the new cost of a dependency of type DEP_TYPE or INSN on DEP_INSN.
   COST is the current cost, DW is dependency weakness.  */
static int
ia64_adjust_cost_2 (rtx insn, int dep_type1, rtx dep_insn, int cost, dw_t dw)
{
  enum reg_note dep_type = (enum reg_note) dep_type1;
  enum attr_itanium_class dep_class;
  enum attr_itanium_class insn_class;

  insn_class = ia64_safe_itanium_class (insn);
  dep_class = ia64_safe_itanium_class (dep_insn);

  /* Treat true memory dependencies separately.  Ignore apparent true
     dependence between store and call (call has a MEM inside a SYMBOL_REF).  */
  if (dep_type == REG_DEP_TRUE
      && (dep_class == ITANIUM_CLASS_ST || dep_class == ITANIUM_CLASS_STF)
      && (insn_class == ITANIUM_CLASS_BR || insn_class == ITANIUM_CLASS_SCALL))
    return 0;

  if (dw == MIN_DEP_WEAK)
    /* Store and load are likely to alias, use higher cost to avoid stall.  */
    return PARAM_VALUE (PARAM_SCHED_MEM_TRUE_DEP_COST);
  else if (dw > MIN_DEP_WEAK)
    {
      /* Store and load are less likely to alias.  */
      if (mflag_sched_fp_mem_deps_zero_cost && dep_class == ITANIUM_CLASS_STF)
	/* Assume there will be no cache conflict for floating-point data.
	   For integer data, L1 conflict penalty is huge (17 cycles), so we
	   never assume it will not cause a conflict.  */
	return 0;
      else
	return cost;
    }

  if (dep_type != REG_DEP_OUTPUT)
    return cost;

  if (dep_class == ITANIUM_CLASS_ST || dep_class == ITANIUM_CLASS_STF
      || insn_class == ITANIUM_CLASS_ST || insn_class == ITANIUM_CLASS_STF)
    return 0;

  return cost;
}

/* Like emit_insn_before, but skip cycle_display notes.
   ??? When cycle display notes are implemented, update this.  */

static void
ia64_emit_insn_before (rtx insn, rtx before)
{
  emit_insn_before (insn, before);
}

/* The following function marks insns who produce addresses for load
   and store insns.  Such insns will be placed into M slots because it
   decrease latency time for Itanium1 (see function
   `ia64_produce_address_p' and the DFA descriptions).  */

static void
ia64_dependencies_evaluation_hook (rtx head, rtx tail)
{
  rtx insn, next, next_tail;

  /* Before reload, which_alternative is not set, which means that
     ia64_safe_itanium_class will produce wrong results for (at least)
     move instructions.  */
  if (!reload_completed)
    return;

  next_tail = NEXT_INSN (tail);
  for (insn = head; insn != next_tail; insn = NEXT_INSN (insn))
    if (INSN_P (insn))
      insn->call = 0;
  for (insn = head; insn != next_tail; insn = NEXT_INSN (insn))
    if (INSN_P (insn)
	&& ia64_safe_itanium_class (insn) == ITANIUM_CLASS_IALU)
      {
	sd_iterator_def sd_it;
	dep_t dep;
	bool has_mem_op_consumer_p = false;

	FOR_EACH_DEP (insn, SD_LIST_FORW, sd_it, dep)
	  {
	    enum attr_itanium_class c;

	    if (DEP_TYPE (dep) != REG_DEP_TRUE)
	      continue;

	    next = DEP_CON (dep);
	    c = ia64_safe_itanium_class (next);
	    if ((c == ITANIUM_CLASS_ST
		 || c == ITANIUM_CLASS_STF)
		&& ia64_st_address_bypass_p (insn, next))
	      {
		has_mem_op_consumer_p = true;
		break;
	      }
	    else if ((c == ITANIUM_CLASS_LD
		      || c == ITANIUM_CLASS_FLD
		      || c == ITANIUM_CLASS_FLDP)
		     && ia64_ld_address_bypass_p (insn, next))
	      {
		has_mem_op_consumer_p = true;
		break;
	      }
	  }

	insn->call = has_mem_op_consumer_p;
      }
}

/* We're beginning a new block.  Initialize data structures as necessary.  */

static void
ia64_sched_init (FILE *dump ATTRIBUTE_UNUSED,
		 int sched_verbose ATTRIBUTE_UNUSED,
		 int max_ready ATTRIBUTE_UNUSED)
{
#ifdef ENABLE_CHECKING
  rtx insn;

  if (!sel_sched_p () && reload_completed)
    for (insn = NEXT_INSN (current_sched_info->prev_head);
	 insn != current_sched_info->next_tail;
	 insn = NEXT_INSN (insn))
      gcc_assert (!SCHED_GROUP_P (insn));
#endif
  last_scheduled_insn = NULL_RTX;
  init_insn_group_barriers ();

  current_cycle = 0;
  memset (mem_ops_in_group, 0, sizeof (mem_ops_in_group));
}

/* We're beginning a scheduling pass.  Check assertion.  */

static void
ia64_sched_init_global (FILE *dump ATTRIBUTE_UNUSED,
                        int sched_verbose ATTRIBUTE_UNUSED,
                        int max_ready ATTRIBUTE_UNUSED)
{  
  gcc_assert (pending_data_specs == 0);
}

/* Scheduling pass is now finished.  Free/reset static variable.  */
static void
ia64_sched_finish_global (FILE *dump ATTRIBUTE_UNUSED,
			  int sched_verbose ATTRIBUTE_UNUSED)
{
  gcc_assert (pending_data_specs == 0);
}

/* Return TRUE if INSN is a load (either normal or speculative, but not a
   speculation check), FALSE otherwise.  */
static bool
is_load_p (rtx insn)
{
  enum attr_itanium_class insn_class = ia64_safe_itanium_class (insn);

  return
   ((insn_class == ITANIUM_CLASS_LD || insn_class == ITANIUM_CLASS_FLD)
    && get_attr_check_load (insn) == CHECK_LOAD_NO);
}

/* If INSN is a memory reference, memoize it in MEM_OPS_IN_GROUP global array
   (taking account for 3-cycle cache reference postponing for stores: Intel
   Itanium 2 Reference Manual for Software Development and Optimization,
   6.7.3.1).  */
static void
record_memory_reference (rtx insn)
{
  enum attr_itanium_class insn_class = ia64_safe_itanium_class (insn);

  switch (insn_class) {
    case ITANIUM_CLASS_FLD:
    case ITANIUM_CLASS_LD:
      mem_ops_in_group[current_cycle % 4]++;
      break;
    case ITANIUM_CLASS_STF:
    case ITANIUM_CLASS_ST:
      mem_ops_in_group[(current_cycle + 3) % 4]++;
      break;
    default:;
  }
}

/* We are about to being issuing insns for this clock cycle.
   Override the default sort algorithm to better slot instructions.  */

static int
ia64_dfa_sched_reorder (FILE *dump, int sched_verbose, rtx *ready,
			int *pn_ready, int clock_var,
			int reorder_type)
{
  int n_asms;
  int n_ready = *pn_ready;
  rtx *e_ready = ready + n_ready;
  rtx *insnp;

  if (sched_verbose)
    fprintf (dump, "// ia64_dfa_sched_reorder (type %d):\n", reorder_type);

  if (reorder_type == 0)
    {
      /* First, move all USEs, CLOBBERs and other crud out of the way.  */
      n_asms = 0;
      for (insnp = ready; insnp < e_ready; insnp++)
	if (insnp < e_ready)
	  {
	    rtx insn = *insnp;
	    enum attr_type t = ia64_safe_type (insn);
	    if (t == TYPE_UNKNOWN)
	      {
		if (GET_CODE (PATTERN (insn)) == ASM_INPUT
		    || asm_noperands (PATTERN (insn)) >= 0)
		  {
		    rtx lowest = ready[n_asms];
		    ready[n_asms] = insn;
		    *insnp = lowest;
		    n_asms++;
		  }
		else
		  {
		    rtx highest = ready[n_ready - 1];
		    ready[n_ready - 1] = insn;
		    *insnp = highest;
		    return 1;
		  }
	      }
	  }

      if (n_asms < n_ready)
	{
	  /* Some normal insns to process.  Skip the asms.  */
	  ready += n_asms;
	  n_ready -= n_asms;
	}
      else if (n_ready > 0)
	return 1;
    }

  if (ia64_final_schedule)
    {
      int deleted = 0;
      int nr_need_stop = 0;

      for (insnp = ready; insnp < e_ready; insnp++)
	if (safe_group_barrier_needed (*insnp))
	  nr_need_stop++;

      if (reorder_type == 1 && n_ready == nr_need_stop)
	return 0;
      if (reorder_type == 0)
	return 1;
      insnp = e_ready;
      /* Move down everything that needs a stop bit, preserving
	 relative order.  */
      while (insnp-- > ready + deleted)
	while (insnp >= ready + deleted)
	  {
	    rtx insn = *insnp;
	    if (! safe_group_barrier_needed (insn))
	      break;
	    memmove (ready + 1, ready, (insnp - ready) * sizeof (rtx));
	    *ready = insn;
	    deleted++;
	  }
      n_ready -= deleted;
      ready += deleted;
    }

  current_cycle = clock_var;
  if (reload_completed && mem_ops_in_group[clock_var % 4] >= ia64_max_memory_insns)
    {
      int moved = 0;

      insnp = e_ready;
      /* Move down loads/stores, preserving relative order.  */
      while (insnp-- > ready + moved)
	while (insnp >= ready + moved)
	  {
	    rtx insn = *insnp;
	    if (! is_load_p (insn))
	      break;
	    memmove (ready + 1, ready, (insnp - ready) * sizeof (rtx));
	    *ready = insn;
	    moved++;
	  }
      n_ready -= moved;
      ready += moved;
    }

  return 1;
}

/* We are about to being issuing insns for this clock cycle.  Override
   the default sort algorithm to better slot instructions.  */

static int
ia64_sched_reorder (FILE *dump, int sched_verbose, rtx *ready, int *pn_ready,
		    int clock_var)
{
  return ia64_dfa_sched_reorder (dump, sched_verbose, ready,
				 pn_ready, clock_var, 0);
}

/* Like ia64_sched_reorder, but called after issuing each insn.
   Override the default sort algorithm to better slot instructions.  */

static int
ia64_sched_reorder2 (FILE *dump ATTRIBUTE_UNUSED,
		     int sched_verbose ATTRIBUTE_UNUSED, rtx *ready,
		     int *pn_ready, int clock_var)
{
  return ia64_dfa_sched_reorder (dump, sched_verbose, ready, pn_ready,
				 clock_var, 1);
}

/* We are about to issue INSN.  Return the number of insns left on the
   ready queue that can be issued this cycle.  */

static int
ia64_variable_issue (FILE *dump ATTRIBUTE_UNUSED,
		     int sched_verbose ATTRIBUTE_UNUSED,
		     rtx insn ATTRIBUTE_UNUSED,
		     int can_issue_more ATTRIBUTE_UNUSED)
{
  if (sched_deps_info->generate_spec_deps && !sel_sched_p ())
    /* Modulo scheduling does not extend h_i_d when emitting
       new instructions.  Don't use h_i_d, if we don't have to.  */
    {
      if (DONE_SPEC (insn) & BEGIN_DATA)
	pending_data_specs++;
      if (CHECK_SPEC (insn) & BEGIN_DATA)
	pending_data_specs--;
    }

  if (DEBUG_INSN_P (insn))
    return 1;

  last_scheduled_insn = insn;
  memcpy (prev_cycle_state, curr_state, dfa_state_size);
  if (reload_completed)
    {
      int needed = group_barrier_needed (insn);
      
      gcc_assert (!needed);
      if (CALL_P (insn))
	init_insn_group_barriers ();
      stops_p [INSN_UID (insn)] = stop_before_p;
      stop_before_p = 0;

      record_memory_reference (insn);
    }
  return 1;
}

/* We are choosing insn from the ready queue.  Return nonzero if INSN
   can be chosen.  */

static int
ia64_first_cycle_multipass_dfa_lookahead_guard (rtx insn)
{
  gcc_assert (insn && INSN_P (insn));
  return ((!reload_completed
	   || !safe_group_barrier_needed (insn))
	  && ia64_first_cycle_multipass_dfa_lookahead_guard_spec (insn)
	  && (!mflag_sched_mem_insns_hard_limit
	      || !is_load_p (insn)
	      || mem_ops_in_group[current_cycle % 4] < ia64_max_memory_insns));
}

/* We are choosing insn from the ready queue.  Return nonzero if INSN
   can be chosen.  */

static bool
ia64_first_cycle_multipass_dfa_lookahead_guard_spec (const_rtx insn)
{
  gcc_assert (insn  && INSN_P (insn));
  /* Size of ALAT is 32.  As far as we perform conservative data speculation,
     we keep ALAT half-empty.  */
  return (pending_data_specs < 16
	  || !(TODO_SPEC (insn) & BEGIN_DATA));
}

/* The following variable value is pseudo-insn used by the DFA insn
   scheduler to change the DFA state when the simulated clock is
   increased.  */

static rtx dfa_pre_cycle_insn;

/* Returns 1 when a meaningful insn was scheduled between the last group
   barrier and LAST.  */
static int
scheduled_good_insn (rtx last)
{
  if (last && recog_memoized (last) >= 0)
    return 1;

  for ( ;
       last != NULL && !NOTE_INSN_BASIC_BLOCK_P (last)
       && !stops_p[INSN_UID (last)];
       last = PREV_INSN (last))
    /* We could hit a NOTE_INSN_DELETED here which is actually outside
       the ebb we're scheduling.  */
    if (INSN_P (last) && recog_memoized (last) >= 0)
      return 1;

  return 0;
}

/* We are about to being issuing INSN.  Return nonzero if we cannot
   issue it on given cycle CLOCK and return zero if we should not sort
   the ready queue on the next clock start.  */

static int
ia64_dfa_new_cycle (FILE *dump, int verbose, rtx insn, int last_clock,
		    int clock, int *sort_p)
{
  gcc_assert (insn && INSN_P (insn));

  if (DEBUG_INSN_P (insn))
    return 0;

  /* When a group barrier is needed for insn, last_scheduled_insn
     should be set.  */
  gcc_assert (!(reload_completed && safe_group_barrier_needed (insn))
              || last_scheduled_insn);

  if ((reload_completed
       && (safe_group_barrier_needed (insn)
	   || (mflag_sched_stop_bits_after_every_cycle
	       && last_clock != clock
	       && last_scheduled_insn
	       && scheduled_good_insn (last_scheduled_insn))))
      || (last_scheduled_insn
	  && (CALL_P (last_scheduled_insn)
	      || unknown_for_bundling_p (last_scheduled_insn))))
    {
      init_insn_group_barriers ();

      if (verbose && dump)
	fprintf (dump, "//    Stop should be before %d%s\n", INSN_UID (insn),
		 last_clock == clock ? " + cycle advance" : "");

      stop_before_p = 1;
      current_cycle = clock;
      mem_ops_in_group[current_cycle % 4] = 0;

      if (last_clock == clock)
	{
	  state_transition (curr_state, dfa_stop_insn);
	  if (TARGET_EARLY_STOP_BITS)
	    *sort_p = (last_scheduled_insn == NULL_RTX
		       || ! CALL_P (last_scheduled_insn));
	  else
	    *sort_p = 0;
	  return 1;
	}

      if (last_scheduled_insn)
	{
	  if (unknown_for_bundling_p (last_scheduled_insn))
	    state_reset (curr_state);
	  else
	    {
	      memcpy (curr_state, prev_cycle_state, dfa_state_size);
	      state_transition (curr_state, dfa_stop_insn);
	      state_transition (curr_state, dfa_pre_cycle_insn);
	      state_transition (curr_state, NULL);
	    }
	}
    }
  return 0;
}

/* Implement targetm.sched.h_i_d_extended hook.
   Extend internal data structures.  */
static void
ia64_h_i_d_extended (void)
{
  if (stops_p != NULL) 
    {
      int new_clocks_length = get_max_uid () * 3 / 2;
      stops_p = (char *) xrecalloc (stops_p, new_clocks_length, clocks_length, 1);
      clocks_length = new_clocks_length;
    }
}


/* This structure describes the data used by the backend to guide scheduling.
   When the current scheduling point is switched, this data should be saved
   and restored later, if the scheduler returns to this point.  */
struct _ia64_sched_context
{
  state_t prev_cycle_state;
  rtx last_scheduled_insn;
  struct reg_write_state rws_sum[NUM_REGS];
  struct reg_write_state rws_insn[NUM_REGS];
  int first_instruction;
  int pending_data_specs;
  int current_cycle;
  char mem_ops_in_group[4];
};
typedef struct _ia64_sched_context *ia64_sched_context_t;

/* Allocates a scheduling context.  */
static void *
ia64_alloc_sched_context (void)
{
  return xmalloc (sizeof (struct _ia64_sched_context));
}

/* Initializes the _SC context with clean data, if CLEAN_P, and from
   the global context otherwise.  */
static void
ia64_init_sched_context (void *_sc, bool clean_p)
{
  ia64_sched_context_t sc = (ia64_sched_context_t) _sc;

  sc->prev_cycle_state = xmalloc (dfa_state_size);
  if (clean_p)
    {
      state_reset (sc->prev_cycle_state);
      sc->last_scheduled_insn = NULL_RTX;
      memset (sc->rws_sum, 0, sizeof (rws_sum));
      memset (sc->rws_insn, 0, sizeof (rws_insn));
      sc->first_instruction = 1;
      sc->pending_data_specs = 0;
      sc->current_cycle = 0;
      memset (sc->mem_ops_in_group, 0, sizeof (mem_ops_in_group));
    }
  else
    {
      memcpy (sc->prev_cycle_state, prev_cycle_state, dfa_state_size);
      sc->last_scheduled_insn = last_scheduled_insn;
      memcpy (sc->rws_sum, rws_sum, sizeof (rws_sum));
      memcpy (sc->rws_insn, rws_insn, sizeof (rws_insn));
      sc->first_instruction = first_instruction;
      sc->pending_data_specs = pending_data_specs;
      sc->current_cycle = current_cycle;
      memcpy (sc->mem_ops_in_group, mem_ops_in_group, sizeof (mem_ops_in_group));
    }
}

/* Sets the global scheduling context to the one pointed to by _SC.  */
static void
ia64_set_sched_context (void *_sc)
{
  ia64_sched_context_t sc = (ia64_sched_context_t) _sc;

  gcc_assert (sc != NULL);

  memcpy (prev_cycle_state, sc->prev_cycle_state, dfa_state_size);
  last_scheduled_insn = sc->last_scheduled_insn;
  memcpy (rws_sum, sc->rws_sum, sizeof (rws_sum));
  memcpy (rws_insn, sc->rws_insn, sizeof (rws_insn));
  first_instruction = sc->first_instruction;
  pending_data_specs = sc->pending_data_specs;
  current_cycle = sc->current_cycle;
  memcpy (mem_ops_in_group, sc->mem_ops_in_group, sizeof (mem_ops_in_group));
}

/* Clears the data in the _SC scheduling context.  */
static void
ia64_clear_sched_context (void *_sc)
{
  ia64_sched_context_t sc = (ia64_sched_context_t) _sc;
  
  free (sc->prev_cycle_state);
  sc->prev_cycle_state = NULL;
}

/* Frees the _SC scheduling context.  */
static void
ia64_free_sched_context (void *_sc)
{
  gcc_assert (_sc != NULL);

  free (_sc);
}

typedef rtx (* gen_func_t) (rtx, rtx);

/* Return a function that will generate a load of mode MODE_NO
   with speculation types TS.  */
static gen_func_t
get_spec_load_gen_function (ds_t ts, int mode_no)
{
  static gen_func_t gen_ld_[] = {
    gen_movbi,
    gen_movqi_internal,
    gen_movhi_internal,
    gen_movsi_internal,
    gen_movdi_internal,
    gen_movsf_internal,
    gen_movdf_internal,
    gen_movxf_internal,
    gen_movti_internal,
    gen_zero_extendqidi2,
    gen_zero_extendhidi2,
    gen_zero_extendsidi2,
  };

  static gen_func_t gen_ld_a[] = {
    gen_movbi_advanced,
    gen_movqi_advanced,
    gen_movhi_advanced,
    gen_movsi_advanced,
    gen_movdi_advanced,
    gen_movsf_advanced,
    gen_movdf_advanced,
    gen_movxf_advanced,
    gen_movti_advanced,
    gen_zero_extendqidi2_advanced,
    gen_zero_extendhidi2_advanced,
    gen_zero_extendsidi2_advanced,
  };
  static gen_func_t gen_ld_s[] = {
    gen_movbi_speculative,
    gen_movqi_speculative,
    gen_movhi_speculative,
    gen_movsi_speculative,
    gen_movdi_speculative,
    gen_movsf_speculative,
    gen_movdf_speculative,
    gen_movxf_speculative,
    gen_movti_speculative,
    gen_zero_extendqidi2_speculative,
    gen_zero_extendhidi2_speculative,
    gen_zero_extendsidi2_speculative,
  };
  static gen_func_t gen_ld_sa[] = {
    gen_movbi_speculative_advanced,
    gen_movqi_speculative_advanced,
    gen_movhi_speculative_advanced,
    gen_movsi_speculative_advanced,
    gen_movdi_speculative_advanced,
    gen_movsf_speculative_advanced,
    gen_movdf_speculative_advanced,
    gen_movxf_speculative_advanced,
    gen_movti_speculative_advanced,
    gen_zero_extendqidi2_speculative_advanced,
    gen_zero_extendhidi2_speculative_advanced,
    gen_zero_extendsidi2_speculative_advanced,
  };
  static gen_func_t gen_ld_s_a[] = {
    gen_movbi_speculative_a,
    gen_movqi_speculative_a,
    gen_movhi_speculative_a,
    gen_movsi_speculative_a,
    gen_movdi_speculative_a,
    gen_movsf_speculative_a,
    gen_movdf_speculative_a,
    gen_movxf_speculative_a,
    gen_movti_speculative_a,
    gen_zero_extendqidi2_speculative_a,
    gen_zero_extendhidi2_speculative_a,
    gen_zero_extendsidi2_speculative_a,
  };

  gen_func_t *gen_ld;

  if (ts & BEGIN_DATA)
    {
      if (ts & BEGIN_CONTROL)
	gen_ld = gen_ld_sa;
      else
	gen_ld = gen_ld_a;
    }
  else if (ts & BEGIN_CONTROL)
    {
      if ((spec_info->flags & SEL_SCHED_SPEC_DONT_CHECK_CONTROL)
	  || ia64_needs_block_p (ts))
	gen_ld = gen_ld_s;
      else
	gen_ld = gen_ld_s_a;
    }
  else if (ts == 0)
    gen_ld = gen_ld_;
  else
    gcc_unreachable ();

  return gen_ld[mode_no];
}

/* Constants that help mapping 'enum machine_mode' to int.  */
enum SPEC_MODES
  {
    SPEC_MODE_INVALID = -1,
    SPEC_MODE_FIRST = 0,
    SPEC_MODE_FOR_EXTEND_FIRST = 1,
    SPEC_MODE_FOR_EXTEND_LAST = 3,
    SPEC_MODE_LAST = 8
  };

enum
  {
    /* Offset to reach ZERO_EXTEND patterns.  */
    SPEC_GEN_EXTEND_OFFSET = SPEC_MODE_LAST - SPEC_MODE_FOR_EXTEND_FIRST + 1
  };

/* Return index of the MODE.  */
static int
ia64_mode_to_int (enum machine_mode mode)
{
  switch (mode)
    {
    case BImode: return 0; /* SPEC_MODE_FIRST  */
    case QImode: return 1; /* SPEC_MODE_FOR_EXTEND_FIRST  */
    case HImode: return 2;
    case SImode: return 3; /* SPEC_MODE_FOR_EXTEND_LAST  */
    case DImode: return 4;
    case SFmode: return 5;
    case DFmode: return 6;
    case XFmode: return 7;
    case TImode:
      /* ??? This mode needs testing.  Bypasses for ldfp8 instruction are not
	 mentioned in itanium[12].md.  Predicate fp_register_operand also
	 needs to be defined.  Bottom line: better disable for now.  */
      return SPEC_MODE_INVALID;
    default:     return SPEC_MODE_INVALID;
    }
}

/* Provide information about speculation capabilities.  */
static void
ia64_set_sched_flags (spec_info_t spec_info)
{
  unsigned int *flags = &(current_sched_info->flags);

  if (*flags & SCHED_RGN
      || *flags & SCHED_EBB
      || *flags & SEL_SCHED)
    {
      int mask = 0;

      if ((mflag_sched_br_data_spec && !reload_completed && optimize > 0)
          || (mflag_sched_ar_data_spec && reload_completed))
	{
	  mask |= BEGIN_DATA;

	  if (!sel_sched_p ()
	      && ((mflag_sched_br_in_data_spec && !reload_completed)
		  || (mflag_sched_ar_in_data_spec && reload_completed)))
	    mask |= BE_IN_DATA;
	}
      
      if (mflag_sched_control_spec
          && (!sel_sched_p ()
	      || reload_completed))
	{
	  mask |= BEGIN_CONTROL;
	  
	  if (!sel_sched_p () && mflag_sched_in_control_spec)
	    mask |= BE_IN_CONTROL;
	}

      spec_info->mask = mask;

      if (mask)
	{
	  *flags |= USE_DEPS_LIST | DO_SPECULATION;

	  if (mask & BE_IN_SPEC)
	    *flags |= NEW_BBS;
	  
	  spec_info->flags = 0;
      
	  if ((mask & DATA_SPEC) && mflag_sched_prefer_non_data_spec_insns)
	    spec_info->flags |= PREFER_NON_DATA_SPEC;

	  if (mask & CONTROL_SPEC)
	    {
	      if (mflag_sched_prefer_non_control_spec_insns)
		spec_info->flags |= PREFER_NON_CONTROL_SPEC;

	      if (sel_sched_p () && mflag_sel_sched_dont_check_control_spec)
		spec_info->flags |= SEL_SCHED_SPEC_DONT_CHECK_CONTROL;
	    }

	  if (sched_verbose >= 1)
	    spec_info->dump = sched_dump;
	  else
	    spec_info->dump = 0;
	  
	  if (mflag_sched_count_spec_in_critical_path)
	    spec_info->flags |= COUNT_SPEC_IN_CRITICAL_PATH;
	}
    }
  else
    spec_info->mask = 0;
}

/* If INSN is an appropriate load return its mode.
   Return -1 otherwise.  */
static int
get_mode_no_for_insn (rtx insn)
{
  rtx reg, mem, mode_rtx;
  int mode_no;
  bool extend_p;

  extract_insn_cached (insn);

  /* We use WHICH_ALTERNATIVE only after reload.  This will
     guarantee that reload won't touch a speculative insn.  */

  if (recog_data.n_operands != 2)
    return -1;

  reg = recog_data.operand[0];
  mem = recog_data.operand[1];

  /* We should use MEM's mode since REG's mode in presence of
     ZERO_EXTEND will always be DImode.  */
  if (get_attr_speculable1 (insn) == SPECULABLE1_YES)
    /* Process non-speculative ld.  */
    {
      if (!reload_completed)
	{
	  /* Do not speculate into regs like ar.lc.  */
	  if (!REG_P (reg) || AR_REGNO_P (REGNO (reg)))
	    return -1;

	  if (!MEM_P (mem))
	    return -1;

	  {
	    rtx mem_reg = XEXP (mem, 0);

	    if (!REG_P (mem_reg))
	      return -1;
	  }

	  mode_rtx = mem;
	}
      else if (get_attr_speculable2 (insn) == SPECULABLE2_YES)
	{
	  gcc_assert (REG_P (reg) && MEM_P (mem));
	  mode_rtx = mem;
	}
      else
	return -1;
    }
  else if (get_attr_data_speculative (insn) == DATA_SPECULATIVE_YES
	   || get_attr_control_speculative (insn) == CONTROL_SPECULATIVE_YES
	   || get_attr_check_load (insn) == CHECK_LOAD_YES)
    /* Process speculative ld or ld.c.  */
    {
      gcc_assert (REG_P (reg) && MEM_P (mem));
      mode_rtx = mem;
    }
  else
    {
      enum attr_itanium_class attr_class = get_attr_itanium_class (insn);

      if (attr_class == ITANIUM_CLASS_CHK_A
	  || attr_class == ITANIUM_CLASS_CHK_S_I
	  || attr_class == ITANIUM_CLASS_CHK_S_F)
	/* Process chk.  */
	mode_rtx = reg;
      else
	return -1;
    }

  mode_no = ia64_mode_to_int (GET_MODE (mode_rtx));

  if (mode_no == SPEC_MODE_INVALID)
    return -1;

  extend_p = (GET_MODE (reg) != GET_MODE (mode_rtx));

  if (extend_p)
    {
      if (!(SPEC_MODE_FOR_EXTEND_FIRST <= mode_no
	    && mode_no <= SPEC_MODE_FOR_EXTEND_LAST))
	return -1;

      mode_no += SPEC_GEN_EXTEND_OFFSET;
    }

  return mode_no;
}

/* If X is an unspec part of a speculative load, return its code.
   Return -1 otherwise.  */
static int
get_spec_unspec_code (const_rtx x)
{
  if (GET_CODE (x) != UNSPEC)
    return -1;

  {
    int code;

    code = XINT (x, 1);

    switch (code)
      {
      case UNSPEC_LDA:
      case UNSPEC_LDS:
      case UNSPEC_LDS_A:
      case UNSPEC_LDSA:
	return code;

      default:
	return -1;
      }
  }
}

/* Implement skip_rtx_p hook.  */
static bool
ia64_skip_rtx_p (const_rtx x)
{
  return get_spec_unspec_code (x) != -1;
}

/* If INSN is a speculative load, return its UNSPEC code.
   Return -1 otherwise.  */
static int
get_insn_spec_code (const_rtx insn)
{
  rtx pat, reg, mem;

  pat = PATTERN (insn);

  if (GET_CODE (pat) == COND_EXEC)
    pat = COND_EXEC_CODE (pat);

  if (GET_CODE (pat) != SET)
    return -1;

  reg = SET_DEST (pat);
  if (!REG_P (reg))
    return -1;

  mem = SET_SRC (pat);
  if (GET_CODE (mem) == ZERO_EXTEND)
    mem = XEXP (mem, 0);

  return get_spec_unspec_code (mem);
}

/* If INSN is a speculative load, return a ds with the speculation types.
   Otherwise [if INSN is a normal instruction] return 0.  */
static ds_t
ia64_get_insn_spec_ds (rtx insn)
{
  int code = get_insn_spec_code (insn);

  switch (code)
    {
    case UNSPEC_LDA:
      return BEGIN_DATA;

    case UNSPEC_LDS:
    case UNSPEC_LDS_A:
      return BEGIN_CONTROL;

    case UNSPEC_LDSA:
      return BEGIN_DATA | BEGIN_CONTROL;

    default:
      return 0;
    }
}

/* If INSN is a speculative load return a ds with the speculation types that
   will be checked.
   Otherwise [if INSN is a normal instruction] return 0.  */
static ds_t
ia64_get_insn_checked_ds (rtx insn)
{
  int code = get_insn_spec_code (insn);

  switch (code)
    {
    case UNSPEC_LDA:
      return BEGIN_DATA | BEGIN_CONTROL;

    case UNSPEC_LDS:
      return BEGIN_CONTROL;

    case UNSPEC_LDS_A:
    case UNSPEC_LDSA:
      return BEGIN_DATA | BEGIN_CONTROL;

    default:
      return 0;
    }
}

/* If GEN_P is true, calculate the index of needed speculation check and return
   speculative pattern for INSN with speculative mode TS, machine mode
   MODE_NO and with ZERO_EXTEND (if EXTEND_P is true).
   If GEN_P is false, just calculate the index of needed speculation check.  */
static rtx
ia64_gen_spec_load (rtx insn, ds_t ts, int mode_no)
{
  rtx pat, new_pat;
  gen_func_t gen_load;

  gen_load = get_spec_load_gen_function (ts, mode_no);

  new_pat = gen_load (copy_rtx (recog_data.operand[0]),
		      copy_rtx (recog_data.operand[1]));

  pat = PATTERN (insn);
  if (GET_CODE (pat) == COND_EXEC)
    new_pat = gen_rtx_COND_EXEC (VOIDmode, copy_rtx (COND_EXEC_TEST (pat)),
				 new_pat);

  return new_pat;
}

static bool
insn_can_be_in_speculative_p (rtx insn ATTRIBUTE_UNUSED,
			      ds_t ds ATTRIBUTE_UNUSED)
{
  return false;
}

/* Implement targetm.sched.speculate_insn hook.
   Check if the INSN can be TS speculative.
   If 'no' - return -1.
   If 'yes' - generate speculative pattern in the NEW_PAT and return 1.
   If current pattern of the INSN already provides TS speculation,
   return 0.  */
static int
ia64_speculate_insn (rtx insn, ds_t ts, rtx *new_pat)
{  
  int mode_no;
  int res;
  
  gcc_assert (!(ts & ~SPECULATIVE));

  if (ia64_spec_check_p (insn))
    return -1;

  if ((ts & BE_IN_SPEC)
      && !insn_can_be_in_speculative_p (insn, ts))
    return -1;

  mode_no = get_mode_no_for_insn (insn);

  if (mode_no != SPEC_MODE_INVALID)
    {
      if (ia64_get_insn_spec_ds (insn) == ds_get_speculation_types (ts))
	res = 0;
      else
	{
	  res = 1;
	  *new_pat = ia64_gen_spec_load (insn, ts, mode_no);
	}
    }
  else
    res = -1;

  return res;
}

/* Return a function that will generate a check for speculation TS with mode
   MODE_NO.
   If simple check is needed, pass true for SIMPLE_CHECK_P.
   If clearing check is needed, pass true for CLEARING_CHECK_P.  */
static gen_func_t
get_spec_check_gen_function (ds_t ts, int mode_no,
			     bool simple_check_p, bool clearing_check_p)
{
  static gen_func_t gen_ld_c_clr[] = {
    gen_movbi_clr,
    gen_movqi_clr,
    gen_movhi_clr,
    gen_movsi_clr,
    gen_movdi_clr,
    gen_movsf_clr,
    gen_movdf_clr,
    gen_movxf_clr,
    gen_movti_clr,
    gen_zero_extendqidi2_clr,
    gen_zero_extendhidi2_clr,
    gen_zero_extendsidi2_clr,
  };
  static gen_func_t gen_ld_c_nc[] = {
    gen_movbi_nc,
    gen_movqi_nc,
    gen_movhi_nc,
    gen_movsi_nc,
    gen_movdi_nc,
    gen_movsf_nc,
    gen_movdf_nc,
    gen_movxf_nc,
    gen_movti_nc,
    gen_zero_extendqidi2_nc,
    gen_zero_extendhidi2_nc,
    gen_zero_extendsidi2_nc,
  };
  static gen_func_t gen_chk_a_clr[] = {
    gen_advanced_load_check_clr_bi,
    gen_advanced_load_check_clr_qi,
    gen_advanced_load_check_clr_hi,
    gen_advanced_load_check_clr_si,
    gen_advanced_load_check_clr_di,
    gen_advanced_load_check_clr_sf,
    gen_advanced_load_check_clr_df,
    gen_advanced_load_check_clr_xf,
    gen_advanced_load_check_clr_ti,
    gen_advanced_load_check_clr_di,
    gen_advanced_load_check_clr_di,
    gen_advanced_load_check_clr_di,
  };
  static gen_func_t gen_chk_a_nc[] = {
    gen_advanced_load_check_nc_bi,
    gen_advanced_load_check_nc_qi,
    gen_advanced_load_check_nc_hi,
    gen_advanced_load_check_nc_si,
    gen_advanced_load_check_nc_di,
    gen_advanced_load_check_nc_sf,
    gen_advanced_load_check_nc_df,
    gen_advanced_load_check_nc_xf,
    gen_advanced_load_check_nc_ti,
    gen_advanced_load_check_nc_di,
    gen_advanced_load_check_nc_di,
    gen_advanced_load_check_nc_di,
  };
  static gen_func_t gen_chk_s[] = {
    gen_speculation_check_bi,
    gen_speculation_check_qi,
    gen_speculation_check_hi,
    gen_speculation_check_si,
    gen_speculation_check_di,
    gen_speculation_check_sf,
    gen_speculation_check_df,
    gen_speculation_check_xf,
    gen_speculation_check_ti,
    gen_speculation_check_di,
    gen_speculation_check_di,
    gen_speculation_check_di,
  };

  gen_func_t *gen_check;

  if (ts & BEGIN_DATA)
    {
      /* We don't need recovery because even if this is ld.sa
	 ALAT entry will be allocated only if NAT bit is set to zero.
	 So it is enough to use ld.c here.  */

      if (simple_check_p)
	{
	  gcc_assert (mflag_sched_spec_ldc);

	  if (clearing_check_p)
	    gen_check = gen_ld_c_clr;
	  else
	    gen_check = gen_ld_c_nc;
	}
      else
	{
	  if (clearing_check_p)
	    gen_check = gen_chk_a_clr;
	  else
	    gen_check = gen_chk_a_nc;
	}
    }
  else if (ts & BEGIN_CONTROL)
    {
      if (simple_check_p)
	/* We might want to use ld.sa -> ld.c instead of
	   ld.s -> chk.s.  */
	{
	  gcc_assert (!ia64_needs_block_p (ts));

	  if (clearing_check_p)
	    gen_check = gen_ld_c_clr;
	  else
	    gen_check = gen_ld_c_nc;
	}
      else
	{
	  gen_check = gen_chk_s;
	}
    }
  else
    gcc_unreachable ();

  gcc_assert (mode_no >= 0);
  return gen_check[mode_no];
}

/* Return nonzero, if INSN needs branchy recovery check.  */
static bool
ia64_needs_block_p (ds_t ts)
{
  if (ts & BEGIN_DATA)
    return !mflag_sched_spec_ldc;

  gcc_assert ((ts & BEGIN_CONTROL) != 0);

  return !(mflag_sched_spec_control_ldc && mflag_sched_spec_ldc);
}

/* Generate (or regenerate) a recovery check for INSN.  */
static rtx
ia64_gen_spec_check (rtx insn, rtx label, ds_t ds)
{
  rtx op1, pat, check_pat;
  gen_func_t gen_check;
  int mode_no;

  mode_no = get_mode_no_for_insn (insn);
  gcc_assert (mode_no >= 0);

  if (label)
    op1 = label;
  else
    {
      gcc_assert (!ia64_needs_block_p (ds));
      op1 = copy_rtx (recog_data.operand[1]);
    }
      
  gen_check = get_spec_check_gen_function (ds, mode_no, label == NULL_RTX,
					   true);

  check_pat = gen_check (copy_rtx (recog_data.operand[0]), op1);
    
  pat = PATTERN (insn);
  if (GET_CODE (pat) == COND_EXEC)
    check_pat = gen_rtx_COND_EXEC (VOIDmode, copy_rtx (COND_EXEC_TEST (pat)),
				   check_pat);

  return check_pat;
}

/* Return nonzero, if X is branchy recovery check.  */
static int
ia64_spec_check_p (rtx x)
{
  x = PATTERN (x);
  if (GET_CODE (x) == COND_EXEC)
    x = COND_EXEC_CODE (x);
  if (GET_CODE (x) == SET)
    return ia64_spec_check_src_p (SET_SRC (x));
  return 0;
}

/* Return nonzero, if SRC belongs to recovery check.  */
static int
ia64_spec_check_src_p (rtx src)
{
  if (GET_CODE (src) == IF_THEN_ELSE)
    {
      rtx t;

      t = XEXP (src, 0);
      if (GET_CODE (t) == NE)
	{
	  t = XEXP (t, 0);	    

	  if (GET_CODE (t) == UNSPEC)
	    {
	      int code;
	      
	      code = XINT (t, 1);
	     
	      if (code == UNSPEC_LDCCLR
		  || code == UNSPEC_LDCNC
		  || code == UNSPEC_CHKACLR
		  || code == UNSPEC_CHKANC
		  || code == UNSPEC_CHKS)
		{
		  gcc_assert (code != 0);
		  return code;
		}
	    }
	}
    }
  return 0;
}


/* The following page contains abstract data `bundle states' which are
   used for bundling insns (inserting nops and template generation).  */

/* The following describes state of insn bundling.  */

struct bundle_state
{
  /* Unique bundle state number to identify them in the debugging
     output  */
  int unique_num;
  rtx insn;     /* corresponding insn, NULL for the 1st and the last state  */
  /* number nops before and after the insn  */
  short before_nops_num, after_nops_num;
  int insn_num; /* insn number (0 - for initial state, 1 - for the 1st
                   insn */
  int cost;     /* cost of the state in cycles */
  int accumulated_insns_num; /* number of all previous insns including
				nops.  L is considered as 2 insns */
  int branch_deviation; /* deviation of previous branches from 3rd slots  */
  int middle_bundle_stops; /* number of stop bits in the middle of bundles */
  struct bundle_state *next;  /* next state with the same insn_num  */
  struct bundle_state *originator; /* originator (previous insn state)  */
  /* All bundle states are in the following chain.  */
  struct bundle_state *allocated_states_chain;
  /* The DFA State after issuing the insn and the nops.  */
  state_t dfa_state;
};

/* The following is map insn number to the corresponding bundle state.  */

static struct bundle_state **index_to_bundle_states;

/* The unique number of next bundle state.  */

static int bundle_states_num;

/* All allocated bundle states are in the following chain.  */

static struct bundle_state *allocated_bundle_states_chain;

/* All allocated but not used bundle states are in the following
   chain.  */

static struct bundle_state *free_bundle_state_chain;


/* The following function returns a free bundle state.  */

static struct bundle_state *
get_free_bundle_state (void)
{
  struct bundle_state *result;

  if (free_bundle_state_chain != NULL)
    {
      result = free_bundle_state_chain;
      free_bundle_state_chain = result->next;
    }
  else
    {
      result = XNEW (struct bundle_state);
      result->dfa_state = xmalloc (dfa_state_size);
      result->allocated_states_chain = allocated_bundle_states_chain;
      allocated_bundle_states_chain = result;
    }
  result->unique_num = bundle_states_num++;
  return result;

}

/* The following function frees given bundle state.  */

static void
free_bundle_state (struct bundle_state *state)
{
  state->next = free_bundle_state_chain;
  free_bundle_state_chain = state;
}

/* Start work with abstract data `bundle states'.  */

static void
initiate_bundle_states (void)
{
  bundle_states_num = 0;
  free_bundle_state_chain = NULL;
  allocated_bundle_states_chain = NULL;
}

/* Finish work with abstract data `bundle states'.  */

static void
finish_bundle_states (void)
{
  struct bundle_state *curr_state, *next_state;

  for (curr_state = allocated_bundle_states_chain;
       curr_state != NULL;
       curr_state = next_state)
    {
      next_state = curr_state->allocated_states_chain;
      free (curr_state->dfa_state);
      free (curr_state);
    }
}

/* Hashtable helpers.  */

struct bundle_state_hasher : typed_noop_remove <bundle_state>
{
  typedef bundle_state value_type;
  typedef bundle_state compare_type;
  static inline hashval_t hash (const value_type *);
  static inline bool equal (const value_type *, const compare_type *);
};

/* The function returns hash of BUNDLE_STATE.  */

inline hashval_t
bundle_state_hasher::hash (const value_type *state)
{
  unsigned result, i;

  for (result = i = 0; i < dfa_state_size; i++)
    result += (((unsigned char *) state->dfa_state) [i]
	       << ((i % CHAR_BIT) * 3 + CHAR_BIT));
  return result + state->insn_num;
}

/* The function returns nonzero if the bundle state keys are equal.  */

inline bool
bundle_state_hasher::equal (const value_type *state1,
			    const compare_type *state2)
{
  return (state1->insn_num == state2->insn_num
	  && memcmp (state1->dfa_state, state2->dfa_state,
		     dfa_state_size) == 0);
}

/* Hash table of the bundle states.  The key is dfa_state and insn_num
   of the bundle states.  */

static hash_table <bundle_state_hasher> bundle_state_table;

/* The function inserts the BUNDLE_STATE into the hash table.  The
   function returns nonzero if the bundle has been inserted into the
   table.  The table contains the best bundle state with given key.  */

static int
insert_bundle_state (struct bundle_state *bundle_state)
{
  struct bundle_state **entry_ptr;

  entry_ptr = bundle_state_table.find_slot (bundle_state, INSERT);
  if (*entry_ptr == NULL)
    {
      bundle_state->next = index_to_bundle_states [bundle_state->insn_num];
      index_to_bundle_states [bundle_state->insn_num] = bundle_state;
      *entry_ptr = bundle_state;
      return TRUE;
    }
  else if (bundle_state->cost < (*entry_ptr)->cost
	   || (bundle_state->cost == (*entry_ptr)->cost
	       && ((*entry_ptr)->accumulated_insns_num
		   > bundle_state->accumulated_insns_num
		   || ((*entry_ptr)->accumulated_insns_num
		       == bundle_state->accumulated_insns_num
		       && ((*entry_ptr)->branch_deviation
			   > bundle_state->branch_deviation
			   || ((*entry_ptr)->branch_deviation
			       == bundle_state->branch_deviation
			       && (*entry_ptr)->middle_bundle_stops
			       > bundle_state->middle_bundle_stops))))))

    {
      struct bundle_state temp;

      temp = **entry_ptr;
      **entry_ptr = *bundle_state;
      (*entry_ptr)->next = temp.next;
      *bundle_state = temp;
    }
  return FALSE;
}

/* Start work with the hash table.  */

static void
initiate_bundle_state_table (void)
{
  bundle_state_table.create (50);
}

/* Finish work with the hash table.  */

static void
finish_bundle_state_table (void)
{
  bundle_state_table.dispose ();
}



/* The following variable is a insn `nop' used to check bundle states
   with different number of inserted nops.  */

static rtx ia64_nop;

/* The following function tries to issue NOPS_NUM nops for the current
   state without advancing processor cycle.  If it failed, the
   function returns FALSE and frees the current state.  */

static int
try_issue_nops (struct bundle_state *curr_state, int nops_num)
{
  int i;

  for (i = 0; i < nops_num; i++)
    if (state_transition (curr_state->dfa_state, ia64_nop) >= 0)
      {
	free_bundle_state (curr_state);
	return FALSE;
      }
  return TRUE;
}

/* The following function tries to issue INSN for the current
   state without advancing processor cycle.  If it failed, the
   function returns FALSE and frees the current state.  */

static int
try_issue_insn (struct bundle_state *curr_state, rtx insn)
{
  if (insn && state_transition (curr_state->dfa_state, insn) >= 0)
    {
      free_bundle_state (curr_state);
      return FALSE;
    }
  return TRUE;
}

/* The following function tries to issue BEFORE_NOPS_NUM nops and INSN
   starting with ORIGINATOR without advancing processor cycle.  If
   TRY_BUNDLE_END_P is TRUE, the function also/only (if
   ONLY_BUNDLE_END_P is TRUE) tries to issue nops to fill all bundle.
   If it was successful, the function creates new bundle state and
   insert into the hash table and into `index_to_bundle_states'.  */

static void
issue_nops_and_insn (struct bundle_state *originator, int before_nops_num,
		     rtx insn, int try_bundle_end_p, int only_bundle_end_p)
{
  struct bundle_state *curr_state;

  curr_state = get_free_bundle_state ();
  memcpy (curr_state->dfa_state, originator->dfa_state, dfa_state_size);
  curr_state->insn = insn;
  curr_state->insn_num = originator->insn_num + 1;
  curr_state->cost = originator->cost;
  curr_state->originator = originator;
  curr_state->before_nops_num = before_nops_num;
  curr_state->after_nops_num = 0;
  curr_state->accumulated_insns_num
    = originator->accumulated_insns_num + before_nops_num;
  curr_state->branch_deviation = originator->branch_deviation;
  curr_state->middle_bundle_stops = originator->middle_bundle_stops;
  gcc_assert (insn);
  if (INSN_CODE (insn) == CODE_FOR_insn_group_barrier)
    {
      gcc_assert (GET_MODE (insn) != TImode);
      if (!try_issue_nops (curr_state, before_nops_num))
	return;
      if (!try_issue_insn (curr_state, insn))
	return;
      memcpy (temp_dfa_state, curr_state->dfa_state, dfa_state_size);
      if (curr_state->accumulated_insns_num % 3 != 0)
	curr_state->middle_bundle_stops++;
      if (state_transition (temp_dfa_state, dfa_pre_cycle_insn) >= 0
	  && curr_state->accumulated_insns_num % 3 != 0)
	{
	  free_bundle_state (curr_state);
	  return;
	}
    }
  else if (GET_MODE (insn) != TImode)
    {
      if (!try_issue_nops (curr_state, before_nops_num))
	return;
      if (!try_issue_insn (curr_state, insn))
	return;
      curr_state->accumulated_insns_num++;
      gcc_assert (!unknown_for_bundling_p (insn));

      if (ia64_safe_type (insn) == TYPE_L)
	curr_state->accumulated_insns_num++;
    }
  else
    {
      /* If this is an insn that must be first in a group, then don't allow
	 nops to be emitted before it.  Currently, alloc is the only such
	 supported instruction.  */
      /* ??? The bundling automatons should handle this for us, but they do
	 not yet have support for the first_insn attribute.  */
      if (before_nops_num > 0 && get_attr_first_insn (insn) == FIRST_INSN_YES)
	{
	  free_bundle_state (curr_state);
	  return;
	}

      state_transition (curr_state->dfa_state, dfa_pre_cycle_insn);
      state_transition (curr_state->dfa_state, NULL);
      curr_state->cost++;
      if (!try_issue_nops (curr_state, before_nops_num))
	return;
      if (!try_issue_insn (curr_state, insn))
	return;
      curr_state->accumulated_insns_num++;
      if (unknown_for_bundling_p (insn))
	{
	  /* Finish bundle containing asm insn.  */
	  curr_state->after_nops_num
	    = 3 - curr_state->accumulated_insns_num % 3;
	  curr_state->accumulated_insns_num
	    += 3 - curr_state->accumulated_insns_num % 3;
	}
      else if (ia64_safe_type (insn) == TYPE_L)
	curr_state->accumulated_insns_num++;
    }
  if (ia64_safe_type (insn) == TYPE_B)
    curr_state->branch_deviation
      += 2 - (curr_state->accumulated_insns_num - 1) % 3;
  if (try_bundle_end_p && curr_state->accumulated_insns_num % 3 != 0)
    {
      if (!only_bundle_end_p && insert_bundle_state (curr_state))
	{
	  state_t dfa_state;
	  struct bundle_state *curr_state1;
	  struct bundle_state *allocated_states_chain;

	  curr_state1 = get_free_bundle_state ();
	  dfa_state = curr_state1->dfa_state;
	  allocated_states_chain = curr_state1->allocated_states_chain;
	  *curr_state1 = *curr_state;
	  curr_state1->dfa_state = dfa_state;
	  curr_state1->allocated_states_chain = allocated_states_chain;
	  memcpy (curr_state1->dfa_state, curr_state->dfa_state,
		  dfa_state_size);
	  curr_state = curr_state1;
	}
      if (!try_issue_nops (curr_state,
			   3 - curr_state->accumulated_insns_num % 3))
	return;
      curr_state->after_nops_num
	= 3 - curr_state->accumulated_insns_num % 3;
      curr_state->accumulated_insns_num
	+= 3 - curr_state->accumulated_insns_num % 3;
    }
  if (!insert_bundle_state (curr_state))
    free_bundle_state (curr_state);
  return;
}

/* The following function returns position in the two window bundle
   for given STATE.  */

static int
get_max_pos (state_t state)
{
  if (cpu_unit_reservation_p (state, pos_6))
    return 6;
  else if (cpu_unit_reservation_p (state, pos_5))
    return 5;
  else if (cpu_unit_reservation_p (state, pos_4))
    return 4;
  else if (cpu_unit_reservation_p (state, pos_3))
    return 3;
  else if (cpu_unit_reservation_p (state, pos_2))
    return 2;
  else if (cpu_unit_reservation_p (state, pos_1))
    return 1;
  else
    return 0;
}

/* The function returns code of a possible template for given position
   and state.  The function should be called only with 2 values of
   position equal to 3 or 6.  We avoid generating F NOPs by putting
   templates containing F insns at the end of the template search
   because undocumented anomaly in McKinley derived cores which can
   cause stalls if an F-unit insn (including a NOP) is issued within a
   six-cycle window after reading certain application registers (such
   as ar.bsp).  Furthermore, power-considerations also argue against
   the use of F-unit instructions unless they're really needed.  */

static int
get_template (state_t state, int pos)
{
  switch (pos)
    {
    case 3:
      if (cpu_unit_reservation_p (state, _0mmi_))
	return 1;
      else if (cpu_unit_reservation_p (state, _0mii_))
	return 0;
      else if (cpu_unit_reservation_p (state, _0mmb_))
	return 7;
      else if (cpu_unit_reservation_p (state, _0mib_))
	return 6;
      else if (cpu_unit_reservation_p (state, _0mbb_))
	return 5;
      else if (cpu_unit_reservation_p (state, _0bbb_))
	return 4;
      else if (cpu_unit_reservation_p (state, _0mmf_))
	return 3;
      else if (cpu_unit_reservation_p (state, _0mfi_))
	return 2;
      else if (cpu_unit_reservation_p (state, _0mfb_))
	return 8;
      else if (cpu_unit_reservation_p (state, _0mlx_))
	return 9;
      else
	gcc_unreachable ();
    case 6:
      if (cpu_unit_reservation_p (state, _1mmi_))
	return 1;
      else if (cpu_unit_reservation_p (state, _1mii_))
	return 0;
      else if (cpu_unit_reservation_p (state, _1mmb_))
	return 7;
      else if (cpu_unit_reservation_p (state, _1mib_))
	return 6;
      else if (cpu_unit_reservation_p (state, _1mbb_))
	return 5;
      else if (cpu_unit_reservation_p (state, _1bbb_))
	return 4;
      else if (_1mmf_ >= 0 && cpu_unit_reservation_p (state, _1mmf_))
	return 3;
      else if (cpu_unit_reservation_p (state, _1mfi_))
	return 2;
      else if (cpu_unit_reservation_p (state, _1mfb_))
	return 8;
      else if (cpu_unit_reservation_p (state, _1mlx_))
	return 9;
      else
	gcc_unreachable ();
    default:
      gcc_unreachable ();
    }
}

/* True when INSN is important for bundling.  */

static bool
important_for_bundling_p (rtx insn)
{
  return (INSN_P (insn)
	  && ia64_safe_itanium_class (insn) != ITANIUM_CLASS_IGNORE
	  && GET_CODE (PATTERN (insn)) != USE
	  && GET_CODE (PATTERN (insn)) != CLOBBER);
}

/* The following function returns an insn important for insn bundling
   followed by INSN and before TAIL.  */

static rtx
get_next_important_insn (rtx insn, rtx tail)
{
  for (; insn && insn != tail; insn = NEXT_INSN (insn))
    if (important_for_bundling_p (insn))
      return insn;
  return NULL_RTX;
}

/* True when INSN is unknown, but important, for bundling.  */

static bool
unknown_for_bundling_p (rtx insn)
{
  return (INSN_P (insn)
	  && ia64_safe_itanium_class (insn) == ITANIUM_CLASS_UNKNOWN
	  && GET_CODE (PATTERN (insn)) != USE
	  && GET_CODE (PATTERN (insn)) != CLOBBER);
}

/* Add a bundle selector TEMPLATE0 before INSN.  */

static void
ia64_add_bundle_selector_before (int template0, rtx insn)
{
  rtx b = gen_bundle_selector (GEN_INT (template0));

  ia64_emit_insn_before (b, insn);
#if NR_BUNDLES == 10
  if ((template0 == 4 || template0 == 5)
      && ia64_except_unwind_info (&global_options) == UI_TARGET)
    {
      int i;
      rtx note = NULL_RTX;

      /* In .mbb and .bbb bundles, check if CALL_INSN isn't in the
	 first or second slot.  If it is and has REG_EH_NOTE set, copy it
	 to following nops, as br.call sets rp to the address of following
	 bundle and therefore an EH region end must be on a bundle
	 boundary.  */
      insn = PREV_INSN (insn);
      for (i = 0; i < 3; i++)
	{
	  do
	    insn = next_active_insn (insn);
	  while (NONJUMP_INSN_P (insn)
		 && get_attr_empty (insn) == EMPTY_YES);
	  if (CALL_P (insn))
	    note = find_reg_note (insn, REG_EH_REGION, NULL_RTX);
	  else if (note)
	    {
	      int code;

	      gcc_assert ((code = recog_memoized (insn)) == CODE_FOR_nop
			  || code == CODE_FOR_nop_b);
	      if (find_reg_note (insn, REG_EH_REGION, NULL_RTX))
		note = NULL_RTX;
	      else
		add_reg_note (insn, REG_EH_REGION, XEXP (note, 0));
	    }
	}
    }
#endif
}

/* The following function does insn bundling.  Bundling means
   inserting templates and nop insns to fit insn groups into permitted
   templates.  Instruction scheduling uses NDFA (non-deterministic
   finite automata) encoding informations about the templates and the
   inserted nops.  Nondeterminism of the automata permits follows
   all possible insn sequences very fast.

   Unfortunately it is not possible to get information about inserting
   nop insns and used templates from the automata states.  The
   automata only says that we can issue an insn possibly inserting
   some nops before it and using some template.  Therefore insn
   bundling in this function is implemented by using DFA
   (deterministic finite automata).  We follow all possible insn
   sequences by inserting 0-2 nops (that is what the NDFA describe for
   insn scheduling) before/after each insn being bundled.  We know the
   start of simulated processor cycle from insn scheduling (insn
   starting a new cycle has TImode).

   Simple implementation of insn bundling would create enormous
   number of possible insn sequences satisfying information about new
   cycle ticks taken from the insn scheduling.  To make the algorithm
   practical we use dynamic programming.  Each decision (about
   inserting nops and implicitly about previous decisions) is described
   by structure bundle_state (see above).  If we generate the same
   bundle state (key is automaton state after issuing the insns and
   nops for it), we reuse already generated one.  As consequence we
   reject some decisions which cannot improve the solution and
   reduce memory for the algorithm.

   When we reach the end of EBB (extended basic block), we choose the
   best sequence and then, moving back in EBB, insert templates for
   the best alternative.  The templates are taken from querying
   automaton state for each insn in chosen bundle states.

   So the algorithm makes two (forward and backward) passes through
   EBB.  */

static void
bundling (FILE *dump, int verbose, rtx prev_head_insn, rtx tail)
{
  struct bundle_state *curr_state, *next_state, *best_state;
  rtx insn, next_insn;
  int insn_num;
  int i, bundle_end_p, only_bundle_end_p, asm_p;
  int pos = 0, max_pos, template0, template1;
  rtx b;
  rtx nop;
  enum attr_type type;

  insn_num = 0;
  /* Count insns in the EBB.  */
  for (insn = NEXT_INSN (prev_head_insn);
       insn && insn != tail;
       insn = NEXT_INSN (insn))
    if (INSN_P (insn))
      insn_num++;
  if (insn_num == 0)
    return;
  bundling_p = 1;
  dfa_clean_insn_cache ();
  initiate_bundle_state_table ();
  index_to_bundle_states = XNEWVEC (struct bundle_state *, insn_num + 2);
  /* First (forward) pass -- generation of bundle states.  */
  curr_state = get_free_bundle_state ();
  curr_state->insn = NULL;
  curr_state->before_nops_num = 0;
  curr_state->after_nops_num = 0;
  curr_state->insn_num = 0;
  curr_state->cost = 0;
  curr_state->accumulated_insns_num = 0;
  curr_state->branch_deviation = 0;
  curr_state->middle_bundle_stops = 0;
  curr_state->next = NULL;
  curr_state->originator = NULL;
  state_reset (curr_state->dfa_state);
  index_to_bundle_states [0] = curr_state;
  insn_num = 0;
  /* Shift cycle mark if it is put on insn which could be ignored.  */
  for (insn = NEXT_INSN (prev_head_insn);
       insn != tail;
       insn = NEXT_INSN (insn))
    if (INSN_P (insn)
	&& !important_for_bundling_p (insn)
	&& GET_MODE (insn) == TImode)
      {
	PUT_MODE (insn, VOIDmode);
	for (next_insn = NEXT_INSN (insn);
	     next_insn != tail;
	     next_insn = NEXT_INSN (next_insn))
	  if (important_for_bundling_p (next_insn)
	      && INSN_CODE (next_insn) != CODE_FOR_insn_group_barrier)
	    {
	      PUT_MODE (next_insn, TImode);
	      break;
	    }
      }
  /* Forward pass: generation of bundle states.  */
  for (insn = get_next_important_insn (NEXT_INSN (prev_head_insn), tail);
       insn != NULL_RTX;
       insn = next_insn)
    {
      gcc_assert (important_for_bundling_p (insn));
      type = ia64_safe_type (insn);
      next_insn = get_next_important_insn (NEXT_INSN (insn), tail);
      insn_num++;
      index_to_bundle_states [insn_num] = NULL;
      for (curr_state = index_to_bundle_states [insn_num - 1];
	   curr_state != NULL;
	   curr_state = next_state)
	{
	  pos = curr_state->accumulated_insns_num % 3;
	  next_state = curr_state->next;
	  /* We must fill up the current bundle in order to start a
	     subsequent asm insn in a new bundle.  Asm insn is always
	     placed in a separate bundle.  */
	  only_bundle_end_p
	    = (next_insn != NULL_RTX
	       && INSN_CODE (insn) == CODE_FOR_insn_group_barrier
	       && unknown_for_bundling_p (next_insn));
	  /* We may fill up the current bundle if it is the cycle end
	     without a group barrier.  */
	  bundle_end_p
	    = (only_bundle_end_p || next_insn == NULL_RTX
	       || (GET_MODE (next_insn) == TImode
		   && INSN_CODE (insn) != CODE_FOR_insn_group_barrier));
	  if (type == TYPE_F || type == TYPE_B || type == TYPE_L
	      || type == TYPE_S)
	    issue_nops_and_insn (curr_state, 2, insn, bundle_end_p,
				 only_bundle_end_p);
	  issue_nops_and_insn (curr_state, 1, insn, bundle_end_p,
			       only_bundle_end_p);
	  issue_nops_and_insn (curr_state, 0, insn, bundle_end_p,
			       only_bundle_end_p);
	}
      gcc_assert (index_to_bundle_states [insn_num]);
      for (curr_state = index_to_bundle_states [insn_num];
	   curr_state != NULL;
	   curr_state = curr_state->next)
	if (verbose >= 2 && dump)
	  {
	    /* This structure is taken from generated code of the
	       pipeline hazard recognizer (see file insn-attrtab.c).
	       Please don't forget to change the structure if a new
	       automaton is added to .md file.  */
	    struct DFA_chip
	    {
	      unsigned short one_automaton_state;
	      unsigned short oneb_automaton_state;
	      unsigned short two_automaton_state;
	      unsigned short twob_automaton_state;
	    };

	    fprintf
	      (dump,
	       "//    Bundle state %d (orig %d, cost %d, nops %d/%d, insns %d, branch %d, mid.stops %d state %d) for %d\n",
	       curr_state->unique_num,
	       (curr_state->originator == NULL
		? -1 : curr_state->originator->unique_num),
	       curr_state->cost,
	       curr_state->before_nops_num, curr_state->after_nops_num,
	       curr_state->accumulated_insns_num, curr_state->branch_deviation,
	       curr_state->middle_bundle_stops,
	       ((struct DFA_chip *) curr_state->dfa_state)->twob_automaton_state,
	       INSN_UID (insn));
	  }
    }
  
  /* We should find a solution because the 2nd insn scheduling has
     found one.  */
  gcc_assert (index_to_bundle_states [insn_num]);
  /* Find a state corresponding to the best insn sequence.  */
  best_state = NULL;
  for (curr_state = index_to_bundle_states [insn_num];
       curr_state != NULL;
       curr_state = curr_state->next)
    /* We are just looking at the states with fully filled up last
       bundle.  The first we prefer insn sequences with minimal cost
       then with minimal inserted nops and finally with branch insns
       placed in the 3rd slots.  */
    if (curr_state->accumulated_insns_num % 3 == 0
	&& (best_state == NULL || best_state->cost > curr_state->cost
	    || (best_state->cost == curr_state->cost
		&& (curr_state->accumulated_insns_num
		    < best_state->accumulated_insns_num
		    || (curr_state->accumulated_insns_num
			== best_state->accumulated_insns_num
			&& (curr_state->branch_deviation
			    < best_state->branch_deviation
			    || (curr_state->branch_deviation
				== best_state->branch_deviation
				&& curr_state->middle_bundle_stops
				< best_state->middle_bundle_stops)))))))
      best_state = curr_state;
  /* Second (backward) pass: adding nops and templates.  */
  gcc_assert (best_state);
  insn_num = best_state->before_nops_num;
  template0 = template1 = -1;
  for (curr_state = best_state;
       curr_state->originator != NULL;
       curr_state = curr_state->originator)
    {
      insn = curr_state->insn;
      asm_p = unknown_for_bundling_p (insn);
      insn_num++;
      if (verbose >= 2 && dump)
	{
	  struct DFA_chip
	  {
	    unsigned short one_automaton_state;
	    unsigned short oneb_automaton_state;
	    unsigned short two_automaton_state;
	    unsigned short twob_automaton_state;
	  };

	  fprintf
	    (dump,
	     "//    Best %d (orig %d, cost %d, nops %d/%d, insns %d, branch %d, mid.stops %d, state %d) for %d\n",
	     curr_state->unique_num,
	     (curr_state->originator == NULL
	      ? -1 : curr_state->originator->unique_num),
	     curr_state->cost,
	     curr_state->before_nops_num, curr_state->after_nops_num,
	     curr_state->accumulated_insns_num, curr_state->branch_deviation,
	     curr_state->middle_bundle_stops,
	     ((struct DFA_chip *) curr_state->dfa_state)->twob_automaton_state,
	     INSN_UID (insn));
	}
      /* Find the position in the current bundle window.  The window can
	 contain at most two bundles.  Two bundle window means that
	 the processor will make two bundle rotation.  */
      max_pos = get_max_pos (curr_state->dfa_state);
      if (max_pos == 6
	  /* The following (negative template number) means that the
	     processor did one bundle rotation.  */
	  || (max_pos == 3 && template0 < 0))
	{
	  /* We are at the end of the window -- find template(s) for
	     its bundle(s).  */
	  pos = max_pos;
	  if (max_pos == 3)
	    template0 = get_template (curr_state->dfa_state, 3);
	  else
	    {
	      template1 = get_template (curr_state->dfa_state, 3);
	      template0 = get_template (curr_state->dfa_state, 6);
	    }
	}
      if (max_pos > 3 && template1 < 0)
	/* It may happen when we have the stop inside a bundle.  */
	{
	  gcc_assert (pos <= 3);
	  template1 = get_template (curr_state->dfa_state, 3);
	  pos += 3;
	}
      if (!asm_p)
	/* Emit nops after the current insn.  */
	for (i = 0; i < curr_state->after_nops_num; i++)
	  {
	    nop = gen_nop ();
	    emit_insn_after (nop, insn);
	    pos--;
	    gcc_assert (pos >= 0);
	    if (pos % 3 == 0)
	      {
		/* We are at the start of a bundle: emit the template
		   (it should be defined).  */
		gcc_assert (template0 >= 0);
		ia64_add_bundle_selector_before (template0, nop);
		/* If we have two bundle window, we make one bundle
		   rotation.  Otherwise template0 will be undefined
		   (negative value).  */
		template0 = template1;
		template1 = -1;
	      }
	  }
      /* Move the position backward in the window.  Group barrier has
	 no slot.  Asm insn takes all bundle.  */
      if (INSN_CODE (insn) != CODE_FOR_insn_group_barrier
	  && !unknown_for_bundling_p (insn))
	pos--;
      /* Long insn takes 2 slots.  */
      if (ia64_safe_type (insn) == TYPE_L)
	pos--;
      gcc_assert (pos >= 0);
      if (pos % 3 == 0
	  && INSN_CODE (insn) != CODE_FOR_insn_group_barrier
	  && !unknown_for_bundling_p (insn))
	{
	  /* The current insn is at the bundle start: emit the
	     template.  */
	  gcc_assert (template0 >= 0);
	  ia64_add_bundle_selector_before (template0, insn);
	  b = PREV_INSN (insn);
	  insn = b;
	  /* See comment above in analogous place for emitting nops
	     after the insn.  */
	  template0 = template1;
	  template1 = -1;
	}
      /* Emit nops after the current insn.  */
      for (i = 0; i < curr_state->before_nops_num; i++)
	{
	  nop = gen_nop ();
	  ia64_emit_insn_before (nop, insn);
	  nop = PREV_INSN (insn);
	  insn = nop;
	  pos--;
	  gcc_assert (pos >= 0);
	  if (pos % 3 == 0)
	    {
	      /* See comment above in analogous place for emitting nops
		 after the insn.  */
	      gcc_assert (template0 >= 0);
	      ia64_add_bundle_selector_before (template0, insn);
	      b = PREV_INSN (insn);
	      insn = b;
	      template0 = template1;
	      template1 = -1;
	    }
	}
    }

#ifdef ENABLE_CHECKING
  {
    /* Assert right calculation of middle_bundle_stops.  */
    int num = best_state->middle_bundle_stops;
    bool start_bundle = true, end_bundle = false;

    for (insn = NEXT_INSN (prev_head_insn);
	 insn && insn != tail;
	 insn = NEXT_INSN (insn))
      {
	if (!INSN_P (insn))
	  continue;
	if (recog_memoized (insn) == CODE_FOR_bundle_selector)
	  start_bundle = true;
	else
	  {
	    rtx next_insn;

	    for (next_insn = NEXT_INSN (insn);
		 next_insn && next_insn != tail;
		 next_insn = NEXT_INSN (next_insn))
	      if (INSN_P (next_insn)
		  && (ia64_safe_itanium_class (next_insn)
		      != ITANIUM_CLASS_IGNORE
		      || recog_memoized (next_insn)
		      == CODE_FOR_bundle_selector)
		  && GET_CODE (PATTERN (next_insn)) != USE
		  && GET_CODE (PATTERN (next_insn)) != CLOBBER)
		break;

	    end_bundle = next_insn == NULL_RTX
	     || next_insn == tail
	     || (INSN_P (next_insn)
		 && recog_memoized (next_insn)
		 == CODE_FOR_bundle_selector);
	    if (recog_memoized (insn) == CODE_FOR_insn_group_barrier
		&& !start_bundle && !end_bundle
		&& next_insn
		&& !unknown_for_bundling_p (next_insn))
	      num--;

	    start_bundle = false;
	  }
      }

    gcc_assert (num == 0);
  }
#endif

  free (index_to_bundle_states);
  finish_bundle_state_table ();
  bundling_p = 0;
  dfa_clean_insn_cache ();
}

/* The following function is called at the end of scheduling BB or
   EBB.  After reload, it inserts stop bits and does insn bundling.  */

static void
ia64_sched_finish (FILE *dump, int sched_verbose)
{
  if (sched_verbose)
    fprintf (dump, "// Finishing schedule.\n");
  if (!reload_completed)
    return;
  if (reload_completed)
    {
      final_emit_insn_group_barriers (dump);
      bundling (dump, sched_verbose, current_sched_info->prev_head,
		current_sched_info->next_tail);
      if (sched_verbose && dump)
	fprintf (dump, "//    finishing %d-%d\n",
		 INSN_UID (NEXT_INSN (current_sched_info->prev_head)),
		 INSN_UID (PREV_INSN (current_sched_info->next_tail)));

      return;
    }
}

/* The following function inserts stop bits in scheduled BB or EBB.  */

static void
final_emit_insn_group_barriers (FILE *dump ATTRIBUTE_UNUSED)
{
  rtx insn;
  int need_barrier_p = 0;
  int seen_good_insn = 0;

  init_insn_group_barriers ();

  for (insn = NEXT_INSN (current_sched_info->prev_head);
       insn != current_sched_info->next_tail;
       insn = NEXT_INSN (insn))
    {
      if (BARRIER_P (insn))
	{
	  rtx last = prev_active_insn (insn);

	  if (! last)
	    continue;
	  if (JUMP_TABLE_DATA_P (last))
	    last = prev_active_insn (last);
	  if (recog_memoized (last) != CODE_FOR_insn_group_barrier)
	    emit_insn_after (gen_insn_group_barrier (GEN_INT (3)), last);

	  init_insn_group_barriers ();
	  seen_good_insn = 0;
	  need_barrier_p = 0;
	}
      else if (NONDEBUG_INSN_P (insn))
	{
	  if (recog_memoized (insn) == CODE_FOR_insn_group_barrier)
	    {
	      init_insn_group_barriers ();
	      seen_good_insn = 0;
	      need_barrier_p = 0;
	    }
	  else if (need_barrier_p || group_barrier_needed (insn)
		   || (mflag_sched_stop_bits_after_every_cycle
		       && GET_MODE (insn) == TImode
		       && seen_good_insn))
	    {
	      if (TARGET_EARLY_STOP_BITS)
		{
		  rtx last;

		  for (last = insn;
		       last != current_sched_info->prev_head;
		       last = PREV_INSN (last))
		    if (INSN_P (last) && GET_MODE (last) == TImode
			&& stops_p [INSN_UID (last)])
		      break;
		  if (last == current_sched_info->prev_head)
		    last = insn;
		  last = prev_active_insn (last);
		  if (last
		      && recog_memoized (last) != CODE_FOR_insn_group_barrier)
		    emit_insn_after (gen_insn_group_barrier (GEN_INT (3)),
				     last);
		  init_insn_group_barriers ();
		  for (last = NEXT_INSN (last);
		       last != insn;
		       last = NEXT_INSN (last))
		    if (INSN_P (last))
		      {
			group_barrier_needed (last);
			if (recog_memoized (last) >= 0
			    && important_for_bundling_p (last))
			  seen_good_insn = 1;
		      }
		}
	      else
		{
		  emit_insn_before (gen_insn_group_barrier (GEN_INT (3)),
				    insn);
		  init_insn_group_barriers ();
		  seen_good_insn = 0;
		}
	      group_barrier_needed (insn);
	      if (recog_memoized (insn) >= 0
		  && important_for_bundling_p (insn))
		seen_good_insn = 1;
	    }
	  else if (recog_memoized (insn) >= 0
		   && important_for_bundling_p (insn))
	    seen_good_insn = 1;
	  need_barrier_p = (CALL_P (insn) || unknown_for_bundling_p (insn));
	}
    }
}



/* If the following function returns TRUE, we will use the DFA
   insn scheduler.  */

static int
ia64_first_cycle_multipass_dfa_lookahead (void)
{
  return (reload_completed ? 6 : 4);
}

/* The following function initiates variable `dfa_pre_cycle_insn'.  */

static void
ia64_init_dfa_pre_cycle_insn (void)
{
  if (temp_dfa_state == NULL)
    {
      dfa_state_size = state_size ();
      temp_dfa_state = xmalloc (dfa_state_size);
      prev_cycle_state = xmalloc (dfa_state_size);
    }
  dfa_pre_cycle_insn = make_insn_raw (gen_pre_cycle ());
  PREV_INSN (dfa_pre_cycle_insn) = NEXT_INSN (dfa_pre_cycle_insn) = NULL_RTX;
  recog_memoized (dfa_pre_cycle_insn);
  dfa_stop_insn = make_insn_raw (gen_insn_group_barrier (GEN_INT (3)));
  PREV_INSN (dfa_stop_insn) = NEXT_INSN (dfa_stop_insn) = NULL_RTX;
  recog_memoized (dfa_stop_insn);
}

/* The following function returns the pseudo insn DFA_PRE_CYCLE_INSN
   used by the DFA insn scheduler.  */

static rtx
ia64_dfa_pre_cycle_insn (void)
{
  return dfa_pre_cycle_insn;
}

/* The following function returns TRUE if PRODUCER (of type ilog or
   ld) produces address for CONSUMER (of type st or stf). */

int
ia64_st_address_bypass_p (rtx producer, rtx consumer)
{
  rtx dest, reg, mem;

  gcc_assert (producer && consumer);
  dest = ia64_single_set (producer);
  gcc_assert (dest);
  reg = SET_DEST (dest);
  gcc_assert (reg);
  if (GET_CODE (reg) == SUBREG)
    reg = SUBREG_REG (reg);
  gcc_assert (GET_CODE (reg) == REG);
  
  dest = ia64_single_set (consumer);
  gcc_assert (dest);
  mem = SET_DEST (dest);
  gcc_assert (mem && GET_CODE (mem) == MEM);
  return reg_mentioned_p (reg, mem);
}

/* The following function returns TRUE if PRODUCER (of type ilog or
   ld) produces address for CONSUMER (of type ld or fld). */

int
ia64_ld_address_bypass_p (rtx producer, rtx consumer)
{
  rtx dest, src, reg, mem;

  gcc_assert (producer && consumer);
  dest = ia64_single_set (producer);
  gcc_assert (dest);
  reg = SET_DEST (dest);
  gcc_assert (reg);
  if (GET_CODE (reg) == SUBREG)
    reg = SUBREG_REG (reg);
  gcc_assert (GET_CODE (reg) == REG);
  
  src = ia64_single_set (consumer);
  gcc_assert (src);
  mem = SET_SRC (src);
  gcc_assert (mem);
 
  if (GET_CODE (mem) == UNSPEC && XVECLEN (mem, 0) > 0)
    mem = XVECEXP (mem, 0, 0);
  else if (GET_CODE (mem) == IF_THEN_ELSE)
    /* ??? Is this bypass necessary for ld.c?  */
    {
      gcc_assert (XINT (XEXP (XEXP (mem, 0), 0), 1) == UNSPEC_LDCCLR);
      mem = XEXP (mem, 1);
    }
     
  while (GET_CODE (mem) == SUBREG || GET_CODE (mem) == ZERO_EXTEND)
    mem = XEXP (mem, 0);

  if (GET_CODE (mem) == UNSPEC)
    {
      int c = XINT (mem, 1);

      gcc_assert (c == UNSPEC_LDA || c == UNSPEC_LDS || c == UNSPEC_LDS_A
		  || c == UNSPEC_LDSA);
      mem = XVECEXP (mem, 0, 0);
    }

  /* Note that LO_SUM is used for GOT loads.  */
  gcc_assert (GET_CODE (mem) == LO_SUM || GET_CODE (mem) == MEM);

  return reg_mentioned_p (reg, mem);
}

/* The following function returns TRUE if INSN produces address for a
   load/store insn.  We will place such insns into M slot because it
   decreases its latency time.  */

int
ia64_produce_address_p (rtx insn)
{
  return insn->call;
}


/* Emit pseudo-ops for the assembler to describe predicate relations.
   At present this assumes that we only consider predicate pairs to
   be mutex, and that the assembler can deduce proper values from
   straight-line code.  */

static void
emit_predicate_relation_info (void)
{
  basic_block bb;

  FOR_EACH_BB_REVERSE_FN (bb, cfun)
    {
      int r;
      rtx head = BB_HEAD (bb);

      /* We only need such notes at code labels.  */
      if (! LABEL_P (head))
	continue;
      if (NOTE_INSN_BASIC_BLOCK_P (NEXT_INSN (head)))
	head = NEXT_INSN (head);

      /* Skip p0, which may be thought to be live due to (reg:DI p0)
	 grabbing the entire block of predicate registers.  */
      for (r = PR_REG (2); r < PR_REG (64); r += 2)
	if (REGNO_REG_SET_P (df_get_live_in (bb), r))
	  {
	    rtx p = gen_rtx_REG (BImode, r);
	    rtx n = emit_insn_after (gen_pred_rel_mutex (p), head);
	    if (head == BB_END (bb))
	      BB_END (bb) = n;
	    head = n;
	  }
    }

  /* Look for conditional calls that do not return, and protect predicate
     relations around them.  Otherwise the assembler will assume the call
     returns, and complain about uses of call-clobbered predicates after
     the call.  */
  FOR_EACH_BB_REVERSE_FN (bb, cfun)
    {
      rtx insn = BB_HEAD (bb);

      while (1)
	{
	  if (CALL_P (insn)
	      && GET_CODE (PATTERN (insn)) == COND_EXEC
	      && find_reg_note (insn, REG_NORETURN, NULL_RTX))
	    {
	      rtx b = emit_insn_before (gen_safe_across_calls_all (), insn);
	      rtx a = emit_insn_after (gen_safe_across_calls_normal (), insn);
	      if (BB_HEAD (bb) == insn)
		BB_HEAD (bb) = b;
	      if (BB_END (bb) == insn)
		BB_END (bb) = a;
	    }

	  if (insn == BB_END (bb))
	    break;
	  insn = NEXT_INSN (insn);
	}
    }
}

/* Perform machine dependent operations on the rtl chain INSNS.  */

static void
ia64_reorg (void)
{
  /* We are freeing block_for_insn in the toplev to keep compatibility
     with old MDEP_REORGS that are not CFG based.  Recompute it now.  */
  compute_bb_for_insn ();

  /* If optimizing, we'll have split before scheduling.  */
  if (optimize == 0)
    split_all_insns ();

  if (optimize && flag_schedule_insns_after_reload
      && dbg_cnt (ia64_sched2))
    {
      basic_block bb;
      timevar_push (TV_SCHED2);
      ia64_final_schedule = 1;

      /* We can't let modulo-sched prevent us from scheduling any bbs,
	 since we need the final schedule to produce bundle information.  */
      FOR_EACH_BB_FN (bb, cfun)
	bb->flags &= ~BB_DISABLE_SCHEDULE;

      initiate_bundle_states ();
      ia64_nop = make_insn_raw (gen_nop ());
      PREV_INSN (ia64_nop) = NEXT_INSN (ia64_nop) = NULL_RTX;
      recog_memoized (ia64_nop);
      clocks_length = get_max_uid () + 1;
      stops_p = XCNEWVEC (char, clocks_length);

      if (ia64_tune == PROCESSOR_ITANIUM2)
	{
	  pos_1 = get_cpu_unit_code ("2_1");
	  pos_2 = get_cpu_unit_code ("2_2");
	  pos_3 = get_cpu_unit_code ("2_3");
	  pos_4 = get_cpu_unit_code ("2_4");
	  pos_5 = get_cpu_unit_code ("2_5");
	  pos_6 = get_cpu_unit_code ("2_6");
	  _0mii_ = get_cpu_unit_code ("2b_0mii.");
	  _0mmi_ = get_cpu_unit_code ("2b_0mmi.");
	  _0mfi_ = get_cpu_unit_code ("2b_0mfi.");
	  _0mmf_ = get_cpu_unit_code ("2b_0mmf.");
	  _0bbb_ = get_cpu_unit_code ("2b_0bbb.");
	  _0mbb_ = get_cpu_unit_code ("2b_0mbb.");
	  _0mib_ = get_cpu_unit_code ("2b_0mib.");
	  _0mmb_ = get_cpu_unit_code ("2b_0mmb.");
	  _0mfb_ = get_cpu_unit_code ("2b_0mfb.");
	  _0mlx_ = get_cpu_unit_code ("2b_0mlx.");
	  _1mii_ = get_cpu_unit_code ("2b_1mii.");
	  _1mmi_ = get_cpu_unit_code ("2b_1mmi.");
	  _1mfi_ = get_cpu_unit_code ("2b_1mfi.");
	  _1mmf_ = get_cpu_unit_code ("2b_1mmf.");
	  _1bbb_ = get_cpu_unit_code ("2b_1bbb.");
	  _1mbb_ = get_cpu_unit_code ("2b_1mbb.");
	  _1mib_ = get_cpu_unit_code ("2b_1mib.");
	  _1mmb_ = get_cpu_unit_code ("2b_1mmb.");
	  _1mfb_ = get_cpu_unit_code ("2b_1mfb.");
	  _1mlx_ = get_cpu_unit_code ("2b_1mlx.");
	}
      else
	{
	  pos_1 = get_cpu_unit_code ("1_1");
	  pos_2 = get_cpu_unit_code ("1_2");
	  pos_3 = get_cpu_unit_code ("1_3");
	  pos_4 = get_cpu_unit_code ("1_4");
	  pos_5 = get_cpu_unit_code ("1_5");
	  pos_6 = get_cpu_unit_code ("1_6");
	  _0mii_ = get_cpu_unit_code ("1b_0mii.");
	  _0mmi_ = get_cpu_unit_code ("1b_0mmi.");
	  _0mfi_ = get_cpu_unit_code ("1b_0mfi.");
	  _0mmf_ = get_cpu_unit_code ("1b_0mmf.");
	  _0bbb_ = get_cpu_unit_code ("1b_0bbb.");
	  _0mbb_ = get_cpu_unit_code ("1b_0mbb.");
	  _0mib_ = get_cpu_unit_code ("1b_0mib.");
	  _0mmb_ = get_cpu_unit_code ("1b_0mmb.");
	  _0mfb_ = get_cpu_unit_code ("1b_0mfb.");
	  _0mlx_ = get_cpu_unit_code ("1b_0mlx.");
	  _1mii_ = get_cpu_unit_code ("1b_1mii.");
	  _1mmi_ = get_cpu_unit_code ("1b_1mmi.");
	  _1mfi_ = get_cpu_unit_code ("1b_1mfi.");
	  _1mmf_ = get_cpu_unit_code ("1b_1mmf.");
	  _1bbb_ = get_cpu_unit_code ("1b_1bbb.");
	  _1mbb_ = get_cpu_unit_code ("1b_1mbb.");
	  _1mib_ = get_cpu_unit_code ("1b_1mib.");
	  _1mmb_ = get_cpu_unit_code ("1b_1mmb.");
	  _1mfb_ = get_cpu_unit_code ("1b_1mfb.");
	  _1mlx_ = get_cpu_unit_code ("1b_1mlx.");
	}

      if (flag_selective_scheduling2
	  && !maybe_skip_selective_scheduling ())
        run_selective_scheduling ();
      else
	schedule_ebbs ();

      /* Redo alignment computation, as it might gone wrong.  */
      compute_alignments ();

      /* We cannot reuse this one because it has been corrupted by the
	 evil glat.  */
      finish_bundle_states ();
      free (stops_p);
      stops_p = NULL;
      emit_insn_group_barriers (dump_file);

      ia64_final_schedule = 0;
      timevar_pop (TV_SCHED2);
    }
  else
    emit_all_insn_group_barriers (dump_file);

  df_analyze ();
 
  /* A call must not be the last instruction in a function, so that the
     return address is still within the function, so that unwinding works
     properly.  Note that IA-64 differs from dwarf2 on this point.  */
  if (ia64_except_unwind_info (&global_options) == UI_TARGET)
    {
      rtx insn;
      int saw_stop = 0;

      insn = get_last_insn ();
      if (! INSN_P (insn))
        insn = prev_active_insn (insn);
      if (insn)
	{
	  /* Skip over insns that expand to nothing.  */
	  while (NONJUMP_INSN_P (insn)
		 && get_attr_empty (insn) == EMPTY_YES)
	    {
	      if (GET_CODE (PATTERN (insn)) == UNSPEC_VOLATILE
		  && XINT (PATTERN (insn), 1) == UNSPECV_INSN_GROUP_BARRIER)
		saw_stop = 1;
	      insn = prev_active_insn (insn);
	    }
	  if (CALL_P (insn))
	    {
	      if (! saw_stop)
		emit_insn (gen_insn_group_barrier (GEN_INT (3)));
	      emit_insn (gen_break_f ());
	      emit_insn (gen_insn_group_barrier (GEN_INT (3)));
	    }
	}
    }

  emit_predicate_relation_info ();

  if (flag_var_tracking)
    {
      timevar_push (TV_VAR_TRACKING);
      variable_tracking_main ();
      timevar_pop (TV_VAR_TRACKING);
    }
  df_finish_pass (false);
}

/* Return true if REGNO is used by the epilogue.  */

int
ia64_epilogue_uses (int regno)
{
  switch (regno)
    {
    case R_GR (1):
      /* With a call to a function in another module, we will write a new
	 value to "gp".  After returning from such a call, we need to make
	 sure the function restores the original gp-value, even if the
	 function itself does not use the gp anymore.  */
      return !(TARGET_AUTO_PIC || TARGET_NO_PIC);

    case IN_REG (0): case IN_REG (1): case IN_REG (2): case IN_REG (3):
    case IN_REG (4): case IN_REG (5): case IN_REG (6): case IN_REG (7):
      /* For functions defined with the syscall_linkage attribute, all
	 input registers are marked as live at all function exits.  This
	 prevents the register allocator from using the input registers,
	 which in turn makes it possible to restart a system call after
	 an interrupt without having to save/restore the input registers.
	 This also prevents kernel data from leaking to application code.  */
      return lookup_attribute ("syscall_linkage",
	   TYPE_ATTRIBUTES (TREE_TYPE (current_function_decl))) != NULL;

    case R_BR (0):
      /* Conditional return patterns can't represent the use of `b0' as
         the return address, so we force the value live this way.  */
      return 1;

    case AR_PFS_REGNUM:
      /* Likewise for ar.pfs, which is used by br.ret.  */
      return 1;

    default:
      return 0;
    }
}

/* Return true if REGNO is used by the frame unwinder.  */

int
ia64_eh_uses (int regno)
{
  unsigned int r;

  if (! reload_completed)
    return 0;

  if (regno == 0)
    return 0;

  for (r = reg_save_b0; r <= reg_save_ar_lc; r++)
    if (regno == current_frame_info.r[r]
       || regno == emitted_frame_related_regs[r])
      return 1;

  return 0;
}

/* Return true if this goes in small data/bss.  */

/* ??? We could also support own long data here.  Generating movl/add/ld8
   instead of addl,ld8/ld8.  This makes the code bigger, but should make the
   code faster because there is one less load.  This also includes incomplete
   types which can't go in sdata/sbss.  */

static bool
ia64_in_small_data_p (const_tree exp)
{
  if (TARGET_NO_SDATA)
    return false;

  /* We want to merge strings, so we never consider them small data.  */
  if (TREE_CODE (exp) == STRING_CST)
    return false;

  /* Functions are never small data.  */
  if (TREE_CODE (exp) == FUNCTION_DECL)
    return false;

  if (TREE_CODE (exp) == VAR_DECL && DECL_SECTION_NAME (exp))
    {
      const char *section = TREE_STRING_POINTER (DECL_SECTION_NAME (exp));

      if (strcmp (section, ".sdata") == 0
	  || strncmp (section, ".sdata.", 7) == 0
	  || strncmp (section, ".gnu.linkonce.s.", 16) == 0
	  || strcmp (section, ".sbss") == 0
	  || strncmp (section, ".sbss.", 6) == 0
	  || strncmp (section, ".gnu.linkonce.sb.", 17) == 0)
	return true;
    }
  else
    {
      HOST_WIDE_INT size = int_size_in_bytes (TREE_TYPE (exp));

      /* If this is an incomplete type with size 0, then we can't put it
	 in sdata because it might be too big when completed.  */
      if (size > 0 && size <= ia64_section_threshold)
	return true;
    }

  return false;
}

/* Output assembly directives for prologue regions.  */

/* The current basic block number.  */

static bool last_block;

/* True if we need a copy_state command at the start of the next block.  */

static bool need_copy_state;

#ifndef MAX_ARTIFICIAL_LABEL_BYTES
# define MAX_ARTIFICIAL_LABEL_BYTES 30
#endif

/* The function emits unwind directives for the start of an epilogue.  */

static void
process_epilogue (FILE *asm_out_file, rtx insn ATTRIBUTE_UNUSED,
		  bool unwind, bool frame ATTRIBUTE_UNUSED)
{
  /* If this isn't the last block of the function, then we need to label the
     current state, and copy it back in at the start of the next block.  */

  if (!last_block)
    {
      if (unwind)
	fprintf (asm_out_file, "\t.label_state %d\n",
		 ++cfun->machine->state_num);
      need_copy_state = true;
    }

  if (unwind)
    fprintf (asm_out_file, "\t.restore sp\n");
}

/* This function processes a SET pattern for REG_CFA_ADJUST_CFA.  */

static void
process_cfa_adjust_cfa (FILE *asm_out_file, rtx pat, rtx insn,
			bool unwind, bool frame)
{
  rtx dest = SET_DEST (pat);
  rtx src = SET_SRC (pat);

  if (dest == stack_pointer_rtx)
    {
      if (GET_CODE (src) == PLUS)
	{
	  rtx op0 = XEXP (src, 0);
	  rtx op1 = XEXP (src, 1);
	  
	  gcc_assert (op0 == dest && GET_CODE (op1) == CONST_INT);
	  
	  if (INTVAL (op1) < 0)
	    {
	      gcc_assert (!frame_pointer_needed);
	      if (unwind)
		fprintf (asm_out_file,
			 "\t.fframe "HOST_WIDE_INT_PRINT_DEC"\n",
			 -INTVAL (op1));
	    }
	  else
	    process_epilogue (asm_out_file, insn, unwind, frame);
	}
      else
	{
	  gcc_assert (src == hard_frame_pointer_rtx);
	  process_epilogue (asm_out_file, insn, unwind, frame);
	}
    }
  else if (dest == hard_frame_pointer_rtx)
    {
      gcc_assert (src == stack_pointer_rtx);
      gcc_assert (frame_pointer_needed);

      if (unwind)
	fprintf (asm_out_file, "\t.vframe r%d\n",
		 ia64_dbx_register_number (REGNO (dest)));
    }
  else
    gcc_unreachable ();
}

/* This function processes a SET pattern for REG_CFA_REGISTER.  */

static void
process_cfa_register (FILE *asm_out_file, rtx pat, bool unwind)
{
  rtx dest = SET_DEST (pat);
  rtx src = SET_SRC (pat);
  int dest_regno = REGNO (dest);
  int src_regno;

  if (src == pc_rtx)
    {
      /* Saving return address pointer.  */
      if (unwind)
	fprintf (asm_out_file, "\t.save rp, r%d\n",
		 ia64_dbx_register_number (dest_regno));
      return;
    }

  src_regno = REGNO (src);

  switch (src_regno)
    {
    case PR_REG (0):
      gcc_assert (dest_regno == current_frame_info.r[reg_save_pr]);
      if (unwind)
	fprintf (asm_out_file, "\t.save pr, r%d\n",
		 ia64_dbx_register_number (dest_regno));
      break;

    case AR_UNAT_REGNUM:
      gcc_assert (dest_regno == current_frame_info.r[reg_save_ar_unat]);
      if (unwind)
	fprintf (asm_out_file, "\t.save ar.unat, r%d\n",
		 ia64_dbx_register_number (dest_regno));
      break;

    case AR_LC_REGNUM:
      gcc_assert (dest_regno == current_frame_info.r[reg_save_ar_lc]);
      if (unwind)
	fprintf (asm_out_file, "\t.save ar.lc, r%d\n",
		 ia64_dbx_register_number (dest_regno));
      break;

    default:
      /* Everything else should indicate being stored to memory.  */
      gcc_unreachable ();
    }
}

/* This function processes a SET pattern for REG_CFA_OFFSET.  */

static void
process_cfa_offset (FILE *asm_out_file, rtx pat, bool unwind)
{
  rtx dest = SET_DEST (pat);
  rtx src = SET_SRC (pat);
  int src_regno = REGNO (src);
  const char *saveop;
  HOST_WIDE_INT off;
  rtx base;

  gcc_assert (MEM_P (dest));
  if (GET_CODE (XEXP (dest, 0)) == REG)
    {
      base = XEXP (dest, 0);
      off = 0;
    }
  else
    {
      gcc_assert (GET_CODE (XEXP (dest, 0)) == PLUS
		  && GET_CODE (XEXP (XEXP (dest, 0), 1)) == CONST_INT);
      base = XEXP (XEXP (dest, 0), 0);
      off = INTVAL (XEXP (XEXP (dest, 0), 1));
    }

  if (base == hard_frame_pointer_rtx)
    {
      saveop = ".savepsp";
      off = - off;
    }
  else
    {
      gcc_assert (base == stack_pointer_rtx);
      saveop = ".savesp";
    }

  src_regno = REGNO (src);
  switch (src_regno)
    {
    case BR_REG (0):
      gcc_assert (!current_frame_info.r[reg_save_b0]);
      if (unwind)
	fprintf (asm_out_file, "\t%s rp, " HOST_WIDE_INT_PRINT_DEC "\n",
		 saveop, off);
      break;

    case PR_REG (0):
      gcc_assert (!current_frame_info.r[reg_save_pr]);
      if (unwind)
	fprintf (asm_out_file, "\t%s pr, " HOST_WIDE_INT_PRINT_DEC "\n",
		 saveop, off);
      break;

    case AR_LC_REGNUM:
      gcc_assert (!current_frame_info.r[reg_save_ar_lc]);
      if (unwind)
	fprintf (asm_out_file, "\t%s ar.lc, " HOST_WIDE_INT_PRINT_DEC "\n",
		 saveop, off);
      break;

    case AR_PFS_REGNUM:
      gcc_assert (!current_frame_info.r[reg_save_ar_pfs]);
      if (unwind)
	fprintf (asm_out_file, "\t%s ar.pfs, " HOST_WIDE_INT_PRINT_DEC "\n",
		 saveop, off);
      break;

    case AR_UNAT_REGNUM:
      gcc_assert (!current_frame_info.r[reg_save_ar_unat]);
      if (unwind)
	fprintf (asm_out_file, "\t%s ar.unat, " HOST_WIDE_INT_PRINT_DEC "\n",
		 saveop, off);
      break;

    case GR_REG (4):
    case GR_REG (5):
    case GR_REG (6):
    case GR_REG (7):
      if (unwind)
	fprintf (asm_out_file, "\t.save.g 0x%x\n",
		 1 << (src_regno - GR_REG (4)));
      break;

    case BR_REG (1):
    case BR_REG (2):
    case BR_REG (3):
    case BR_REG (4):
    case BR_REG (5):
      if (unwind)
	fprintf (asm_out_file, "\t.save.b 0x%x\n",
		 1 << (src_regno - BR_REG (1)));
      break;

    case FR_REG (2):
    case FR_REG (3):
    case FR_REG (4):
    case FR_REG (5):
      if (unwind)
	fprintf (asm_out_file, "\t.save.f 0x%x\n",
		 1 << (src_regno - FR_REG (2)));
      break;

    case FR_REG (16): case FR_REG (17): case FR_REG (18): case FR_REG (19):
    case FR_REG (20): case FR_REG (21): case FR_REG (22): case FR_REG (23):
    case FR_REG (24): case FR_REG (25): case FR_REG (26): case FR_REG (27):
    case FR_REG (28): case FR_REG (29): case FR_REG (30): case FR_REG (31):
      if (unwind)
	fprintf (asm_out_file, "\t.save.gf 0x0, 0x%x\n",
		 1 << (src_regno - FR_REG (12)));
      break;

    default:
      /* ??? For some reason we mark other general registers, even those
	 we can't represent in the unwind info.  Ignore them.  */
      break;
    }
}

/* This function looks at a single insn and emits any directives
   required to unwind this insn.  */

static void
ia64_asm_unwind_emit (FILE *asm_out_file, rtx insn)
{
  bool unwind = ia64_except_unwind_info (&global_options) == UI_TARGET;
  bool frame = dwarf2out_do_frame ();
  rtx note, pat;
  bool handled_one;

  if (!unwind && !frame)
    return;

  if (NOTE_INSN_BASIC_BLOCK_P (insn))
    {
      last_block = NOTE_BASIC_BLOCK (insn)->next_bb
     == EXIT_BLOCK_PTR_FOR_FN (cfun);

      /* Restore unwind state from immediately before the epilogue.  */
      if (need_copy_state)
	{
	  if (unwind)
	    {
	      fprintf (asm_out_file, "\t.body\n");
	      fprintf (asm_out_file, "\t.copy_state %d\n",
		       cfun->machine->state_num);
	    }
	  need_copy_state = false;
	}
    }

  if (NOTE_P (insn) || ! RTX_FRAME_RELATED_P (insn))
    return;

  /* Look for the ALLOC insn.  */
  if (INSN_CODE (insn) == CODE_FOR_alloc)
    {
      rtx dest = SET_DEST (XVECEXP (PATTERN (insn), 0, 0));
      int dest_regno = REGNO (dest);

      /* If this is the final destination for ar.pfs, then this must
	 be the alloc in the prologue.  */
      if (dest_regno == current_frame_info.r[reg_save_ar_pfs])
	{
	  if (unwind)
	    fprintf (asm_out_file, "\t.save ar.pfs, r%d\n",
		     ia64_dbx_register_number (dest_regno));
	}
      else
	{
	  /* This must be an alloc before a sibcall.  We must drop the
	     old frame info.  The easiest way to drop the old frame
	     info is to ensure we had a ".restore sp" directive
	     followed by a new prologue.  If the procedure doesn't
	     have a memory-stack frame, we'll issue a dummy ".restore
	     sp" now.  */
	  if (current_frame_info.total_size == 0 && !frame_pointer_needed)
	    /* if haven't done process_epilogue() yet, do it now */
	    process_epilogue (asm_out_file, insn, unwind, frame);
	  if (unwind)
	    fprintf (asm_out_file, "\t.prologue\n");
	}
      return;
    }

  handled_one = false;
  for (note = REG_NOTES (insn); note; note = XEXP (note, 1))
    switch (REG_NOTE_KIND (note))
      {
      case REG_CFA_ADJUST_CFA:
	pat = XEXP (note, 0);
	if (pat == NULL)
	  pat = PATTERN (insn);
	process_cfa_adjust_cfa (asm_out_file, pat, insn, unwind, frame);
	handled_one = true;
	break;

      case REG_CFA_OFFSET:
	pat = XEXP (note, 0);
	if (pat == NULL)
	  pat = PATTERN (insn);
	process_cfa_offset (asm_out_file, pat, unwind);
	handled_one = true;
	break;

      case REG_CFA_REGISTER:
	pat = XEXP (note, 0);
	if (pat == NULL)
	  pat = PATTERN (insn);
	process_cfa_register (asm_out_file, pat, unwind);
	handled_one = true;
	break;

      case REG_FRAME_RELATED_EXPR:
      case REG_CFA_DEF_CFA:
      case REG_CFA_EXPRESSION:
      case REG_CFA_RESTORE:
      case REG_CFA_SET_VDRAP:
	/* Not used in the ia64 port.  */
	gcc_unreachable ();

      default:
	/* Not a frame-related note.  */
	break;
      }

  /* All REG_FRAME_RELATED_P insns, besides ALLOC, are marked with the
     explicit action to take.  No guessing required.  */
  gcc_assert (handled_one);
}

/* Implement TARGET_ASM_EMIT_EXCEPT_PERSONALITY.  */

static void
ia64_asm_emit_except_personality (rtx personality)
{
  fputs ("\t.personality\t", asm_out_file);
  output_addr_const (asm_out_file, personality);
  fputc ('\n', asm_out_file);
}

/* Implement TARGET_ASM_INITIALIZE_SECTIONS.  */

static void
ia64_asm_init_sections (void)
{
  exception_section = get_unnamed_section (0, output_section_asm_op,
					   "\t.handlerdata");
}

/* Implement TARGET_DEBUG_UNWIND_INFO.  */

static enum unwind_info_type
ia64_debug_unwind_info (void)
{
  return UI_TARGET;
}

enum ia64_builtins
{
  IA64_BUILTIN_BSP,
  IA64_BUILTIN_COPYSIGNQ,
  IA64_BUILTIN_FABSQ,
  IA64_BUILTIN_FLUSHRS,
  IA64_BUILTIN_INFQ,
  IA64_BUILTIN_HUGE_VALQ,
  IA64_BUILTIN_max
};

static GTY(()) tree ia64_builtins[(int) IA64_BUILTIN_max];

void
ia64_init_builtins (void)
{
  tree fpreg_type;
  tree float80_type;
  tree decl;

  /* The __fpreg type.  */
  fpreg_type = make_node (REAL_TYPE);
  TYPE_PRECISION (fpreg_type) = 82;
  layout_type (fpreg_type);
  (*lang_hooks.types.register_builtin_type) (fpreg_type, "__fpreg");

  /* The __float80 type.  */
  float80_type = make_node (REAL_TYPE);
  TYPE_PRECISION (float80_type) = 80;
  layout_type (float80_type);
  (*lang_hooks.types.register_builtin_type) (float80_type, "__float80");

  /* The __float128 type.  */
  if (!TARGET_HPUX)
    {
      tree ftype;
      tree float128_type = make_node (REAL_TYPE);

      TYPE_PRECISION (float128_type) = 128;
      layout_type (float128_type);
      (*lang_hooks.types.register_builtin_type) (float128_type, "__float128");

      /* TFmode support builtins.  */
      ftype = build_function_type_list (float128_type, NULL_TREE);
      decl = add_builtin_function ("__builtin_infq", ftype,
				   IA64_BUILTIN_INFQ, BUILT_IN_MD,
				   NULL, NULL_TREE);
      ia64_builtins[IA64_BUILTIN_INFQ] = decl;

      decl = add_builtin_function ("__builtin_huge_valq", ftype,
				   IA64_BUILTIN_HUGE_VALQ, BUILT_IN_MD,
				   NULL, NULL_TREE);
      ia64_builtins[IA64_BUILTIN_HUGE_VALQ] = decl;

      ftype = build_function_type_list (float128_type,
					float128_type,
					NULL_TREE);
      decl = add_builtin_function ("__builtin_fabsq", ftype,
				   IA64_BUILTIN_FABSQ, BUILT_IN_MD,
				   "__fabstf2", NULL_TREE);
      TREE_READONLY (decl) = 1;
      ia64_builtins[IA64_BUILTIN_FABSQ] = decl;

      ftype = build_function_type_list (float128_type,
					float128_type,
					float128_type,
					NULL_TREE);
      decl = add_builtin_function ("__builtin_copysignq", ftype,
				   IA64_BUILTIN_COPYSIGNQ, BUILT_IN_MD,
				   "__copysigntf3", NULL_TREE);
      TREE_READONLY (decl) = 1;
      ia64_builtins[IA64_BUILTIN_COPYSIGNQ] = decl;
    }
  else
    /* Under HPUX, this is a synonym for "long double".  */
    (*lang_hooks.types.register_builtin_type) (long_double_type_node,
					       "__float128");

  /* Fwrite on VMS is non-standard.  */
#if TARGET_ABI_OPEN_VMS
  vms_patch_builtins ();
#endif

#define def_builtin(name, type, code)					\
  add_builtin_function ((name), (type), (code), BUILT_IN_MD,	\
		       NULL, NULL_TREE)

  decl = def_builtin ("__builtin_ia64_bsp",
		      build_function_type_list (ptr_type_node, NULL_TREE),
		      IA64_BUILTIN_BSP);
  ia64_builtins[IA64_BUILTIN_BSP] = decl;

  decl = def_builtin ("__builtin_ia64_flushrs",
		      build_function_type_list (void_type_node, NULL_TREE),
		      IA64_BUILTIN_FLUSHRS);
  ia64_builtins[IA64_BUILTIN_FLUSHRS] = decl;

#undef def_builtin

  if (TARGET_HPUX)
    {
      if ((decl = builtin_decl_explicit (BUILT_IN_FINITE)) != NULL_TREE)
	set_user_assembler_name (decl, "_Isfinite");
      if ((decl = builtin_decl_explicit (BUILT_IN_FINITEF)) != NULL_TREE)
	set_user_assembler_name (decl, "_Isfinitef");
      if ((decl = builtin_decl_explicit (BUILT_IN_FINITEL)) != NULL_TREE)
	set_user_assembler_name (decl, "_Isfinitef128");
    }
}

rtx
ia64_expand_builtin (tree exp, rtx target, rtx subtarget ATTRIBUTE_UNUSED,
		     enum machine_mode mode ATTRIBUTE_UNUSED,
		     int ignore ATTRIBUTE_UNUSED)
{
  tree fndecl = TREE_OPERAND (CALL_EXPR_FN (exp), 0);
  unsigned int fcode = DECL_FUNCTION_CODE (fndecl);

  switch (fcode)
    {
    case IA64_BUILTIN_BSP:
      if (! target || ! register_operand (target, DImode))
	target = gen_reg_rtx (DImode);
      emit_insn (gen_bsp_value (target));
#ifdef POINTERS_EXTEND_UNSIGNED
      target = convert_memory_address (ptr_mode, target);
#endif
      return target;

    case IA64_BUILTIN_FLUSHRS:
      emit_insn (gen_flushrs ());
      return const0_rtx;

    case IA64_BUILTIN_INFQ:
    case IA64_BUILTIN_HUGE_VALQ:
      {
        enum machine_mode target_mode = TYPE_MODE (TREE_TYPE (exp));
	REAL_VALUE_TYPE inf;
	rtx tmp;

	real_inf (&inf);
	tmp = CONST_DOUBLE_FROM_REAL_VALUE (inf, target_mode);

	tmp = validize_mem (force_const_mem (target_mode, tmp));

	if (target == 0)
	  target = gen_reg_rtx (target_mode);

	emit_move_insn (target, tmp);
	return target;
      }

    case IA64_BUILTIN_FABSQ:
    case IA64_BUILTIN_COPYSIGNQ:
      return expand_call (exp, target, ignore);

    default:
      gcc_unreachable ();
    }

  return NULL_RTX;
}

/* Return the ia64 builtin for CODE.  */

static tree
ia64_builtin_decl (unsigned code, bool initialize_p ATTRIBUTE_UNUSED)
{
  if (code >= IA64_BUILTIN_max)
    return error_mark_node;

  return ia64_builtins[code];
}

/* For the HP-UX IA64 aggregate parameters are passed stored in the
   most significant bits of the stack slot.  */

enum direction
ia64_hpux_function_arg_padding (enum machine_mode mode, const_tree type)
{
   /* Exception to normal case for structures/unions/etc.  */

   if (type && AGGREGATE_TYPE_P (type)
       && int_size_in_bytes (type) < UNITS_PER_WORD)
     return upward;

   /* Fall back to the default.  */
   return DEFAULT_FUNCTION_ARG_PADDING (mode, type);
}

/* Emit text to declare externally defined variables and functions, because
   the Intel assembler does not support undefined externals.  */

void
ia64_asm_output_external (FILE *file, tree decl, const char *name)
{
  /* We output the name if and only if TREE_SYMBOL_REFERENCED is
     set in order to avoid putting out names that are never really
     used. */
  if (TREE_SYMBOL_REFERENCED (DECL_ASSEMBLER_NAME (decl)))
    {
      /* maybe_assemble_visibility will return 1 if the assembler
	 visibility directive is output.  */
      int need_visibility = ((*targetm.binds_local_p) (decl)
			     && maybe_assemble_visibility (decl));

      /* GNU as does not need anything here, but the HP linker does
	 need something for external functions.  */
      if ((TARGET_HPUX_LD || !TARGET_GNU_AS)
	  && TREE_CODE (decl) == FUNCTION_DECL)
	  (*targetm.asm_out.globalize_decl_name) (file, decl);
      else if (need_visibility && !TARGET_GNU_AS)
	(*targetm.asm_out.globalize_label) (file, name);
    }
}

/* Set SImode div/mod functions, init_integral_libfuncs only initializes
   modes of word_mode and larger.  Rename the TFmode libfuncs using the
   HPUX conventions. __divtf3 is used for XFmode. We need to keep it for
   backward compatibility. */

static void
ia64_init_libfuncs (void)
{
  set_optab_libfunc (sdiv_optab, SImode, "__divsi3");
  set_optab_libfunc (udiv_optab, SImode, "__udivsi3");
  set_optab_libfunc (smod_optab, SImode, "__modsi3");
  set_optab_libfunc (umod_optab, SImode, "__umodsi3");

  set_optab_libfunc (add_optab, TFmode, "_U_Qfadd");
  set_optab_libfunc (sub_optab, TFmode, "_U_Qfsub");
  set_optab_libfunc (smul_optab, TFmode, "_U_Qfmpy");
  set_optab_libfunc (sdiv_optab, TFmode, "_U_Qfdiv");
  set_optab_libfunc (neg_optab, TFmode, "_U_Qfneg");

  set_conv_libfunc (sext_optab, TFmode, SFmode, "_U_Qfcnvff_sgl_to_quad");
  set_conv_libfunc (sext_optab, TFmode, DFmode, "_U_Qfcnvff_dbl_to_quad");
  set_conv_libfunc (sext_optab, TFmode, XFmode, "_U_Qfcnvff_f80_to_quad");
  set_conv_libfunc (trunc_optab, SFmode, TFmode, "_U_Qfcnvff_quad_to_sgl");
  set_conv_libfunc (trunc_optab, DFmode, TFmode, "_U_Qfcnvff_quad_to_dbl");
  set_conv_libfunc (trunc_optab, XFmode, TFmode, "_U_Qfcnvff_quad_to_f80");

  set_conv_libfunc (sfix_optab, SImode, TFmode, "_U_Qfcnvfxt_quad_to_sgl");
  set_conv_libfunc (sfix_optab, DImode, TFmode, "_U_Qfcnvfxt_quad_to_dbl");
  set_conv_libfunc (sfix_optab, TImode, TFmode, "_U_Qfcnvfxt_quad_to_quad");
  set_conv_libfunc (ufix_optab, SImode, TFmode, "_U_Qfcnvfxut_quad_to_sgl");
  set_conv_libfunc (ufix_optab, DImode, TFmode, "_U_Qfcnvfxut_quad_to_dbl");

  set_conv_libfunc (sfloat_optab, TFmode, SImode, "_U_Qfcnvxf_sgl_to_quad");
  set_conv_libfunc (sfloat_optab, TFmode, DImode, "_U_Qfcnvxf_dbl_to_quad");
  set_conv_libfunc (sfloat_optab, TFmode, TImode, "_U_Qfcnvxf_quad_to_quad");
  /* HP-UX 11.23 libc does not have a function for unsigned
     SImode-to-TFmode conversion.  */
  set_conv_libfunc (ufloat_optab, TFmode, DImode, "_U_Qfcnvxuf_dbl_to_quad");
}

/* Rename all the TFmode libfuncs using the HPUX conventions.  */

static void
ia64_hpux_init_libfuncs (void)
{
  ia64_init_libfuncs ();

  /* The HP SI millicode division and mod functions expect DI arguments.
     By turning them off completely we avoid using both libgcc and the
     non-standard millicode routines and use the HP DI millicode routines
     instead.  */

  set_optab_libfunc (sdiv_optab, SImode, 0);
  set_optab_libfunc (udiv_optab, SImode, 0);
  set_optab_libfunc (smod_optab, SImode, 0);
  set_optab_libfunc (umod_optab, SImode, 0);

  set_optab_libfunc (sdiv_optab, DImode, "__milli_divI");
  set_optab_libfunc (udiv_optab, DImode, "__milli_divU");
  set_optab_libfunc (smod_optab, DImode, "__milli_remI");
  set_optab_libfunc (umod_optab, DImode, "__milli_remU");

  /* HP-UX libc has TF min/max/abs routines in it.  */
  set_optab_libfunc (smin_optab, TFmode, "_U_Qfmin");
  set_optab_libfunc (smax_optab, TFmode, "_U_Qfmax");
  set_optab_libfunc (abs_optab, TFmode, "_U_Qfabs");

  /* ia64_expand_compare uses this.  */
  cmptf_libfunc = init_one_libfunc ("_U_Qfcmp");

  /* These should never be used.  */
  set_optab_libfunc (eq_optab, TFmode, 0);
  set_optab_libfunc (ne_optab, TFmode, 0);
  set_optab_libfunc (gt_optab, TFmode, 0);
  set_optab_libfunc (ge_optab, TFmode, 0);
  set_optab_libfunc (lt_optab, TFmode, 0);
  set_optab_libfunc (le_optab, TFmode, 0);
}

/* Rename the division and modulus functions in VMS.  */

static void
ia64_vms_init_libfuncs (void)
{
  set_optab_libfunc (sdiv_optab, SImode, "OTS$DIV_I");
  set_optab_libfunc (sdiv_optab, DImode, "OTS$DIV_L");
  set_optab_libfunc (udiv_optab, SImode, "OTS$DIV_UI");
  set_optab_libfunc (udiv_optab, DImode, "OTS$DIV_UL");
  set_optab_libfunc (smod_optab, SImode, "OTS$REM_I");
  set_optab_libfunc (smod_optab, DImode, "OTS$REM_L");
  set_optab_libfunc (umod_optab, SImode, "OTS$REM_UI");
  set_optab_libfunc (umod_optab, DImode, "OTS$REM_UL");
  abort_libfunc = init_one_libfunc ("decc$abort");
  memcmp_libfunc = init_one_libfunc ("decc$memcmp");
#ifdef MEM_LIBFUNCS_INIT
  MEM_LIBFUNCS_INIT;
#endif
}

/* Rename the TFmode libfuncs available from soft-fp in glibc using
   the HPUX conventions.  */

static void
ia64_sysv4_init_libfuncs (void)
{
  ia64_init_libfuncs ();

  /* These functions are not part of the HPUX TFmode interface.  We
     use them instead of _U_Qfcmp, which doesn't work the way we
     expect.  */
  set_optab_libfunc (eq_optab, TFmode, "_U_Qfeq");
  set_optab_libfunc (ne_optab, TFmode, "_U_Qfne");
  set_optab_libfunc (gt_optab, TFmode, "_U_Qfgt");
  set_optab_libfunc (ge_optab, TFmode, "_U_Qfge");
  set_optab_libfunc (lt_optab, TFmode, "_U_Qflt");
  set_optab_libfunc (le_optab, TFmode, "_U_Qfle");

  /* We leave out _U_Qfmin, _U_Qfmax and _U_Qfabs since soft-fp in
     glibc doesn't have them.  */
}

/* Use soft-fp.  */

static void
ia64_soft_fp_init_libfuncs (void)
{
}

static bool
ia64_vms_valid_pointer_mode (enum machine_mode mode)
{
  return (mode == SImode || mode == DImode);
}

/* For HPUX, it is illegal to have relocations in shared segments.  */

static int
ia64_hpux_reloc_rw_mask (void)
{
  return 3;
}

/* For others, relax this so that relocations to local data goes in
   read-only segments, but we still cannot allow global relocations
   in read-only segments.  */

static int
ia64_reloc_rw_mask (void)
{
  return flag_pic ? 3 : 2;
}

/* Return the section to use for X.  The only special thing we do here
   is to honor small data.  */

static section *
ia64_select_rtx_section (enum machine_mode mode, rtx x,
			 unsigned HOST_WIDE_INT align)
{
  if (GET_MODE_SIZE (mode) > 0
      && GET_MODE_SIZE (mode) <= ia64_section_threshold
      && !TARGET_NO_SDATA)
    return sdata_section;
  else
    return default_elf_select_rtx_section (mode, x, align);
}

static unsigned int
ia64_section_type_flags (tree decl, const char *name, int reloc)
{
  unsigned int flags = 0;

  if (strcmp (name, ".sdata") == 0
      || strncmp (name, ".sdata.", 7) == 0
      || strncmp (name, ".gnu.linkonce.s.", 16) == 0
      || strncmp (name, ".sdata2.", 8) == 0
      || strncmp (name, ".gnu.linkonce.s2.", 17) == 0
      || strcmp (name, ".sbss") == 0
      || strncmp (name, ".sbss.", 6) == 0
      || strncmp (name, ".gnu.linkonce.sb.", 17) == 0)
    flags = SECTION_SMALL;

  flags |= default_section_type_flags (decl, name, reloc);
  return flags;
}

/* Returns true if FNTYPE (a FUNCTION_TYPE or a METHOD_TYPE) returns a
   structure type and that the address of that type should be passed
   in out0, rather than in r8.  */

static bool
ia64_struct_retval_addr_is_first_parm_p (tree fntype)
{
  tree ret_type = TREE_TYPE (fntype);

  /* The Itanium C++ ABI requires that out0, rather than r8, be used
     as the structure return address parameter, if the return value
     type has a non-trivial copy constructor or destructor.  It is not
     clear if this same convention should be used for other
     programming languages.  Until G++ 3.4, we incorrectly used r8 for
     these return values.  */
  return (abi_version_at_least (2)
	  && ret_type
	  && TYPE_MODE (ret_type) == BLKmode 
	  && TREE_ADDRESSABLE (ret_type)
	  && strcmp (lang_hooks.name, "GNU C++") == 0);
}

/* Output the assembler code for a thunk function.  THUNK_DECL is the
   declaration for the thunk function itself, FUNCTION is the decl for
   the target function.  DELTA is an immediate constant offset to be
   added to THIS.  If VCALL_OFFSET is nonzero, the word at
   *(*this + vcall_offset) should be added to THIS.  */

static void
ia64_output_mi_thunk (FILE *file, tree thunk ATTRIBUTE_UNUSED,
		      HOST_WIDE_INT delta, HOST_WIDE_INT vcall_offset,
		      tree function)
{
  rtx this_rtx, insn, funexp;
  unsigned int this_parmno;
  unsigned int this_regno;
  rtx delta_rtx;

  reload_completed = 1;
  epilogue_completed = 1;

  /* Set things up as ia64_expand_prologue might.  */
  last_scratch_gr_reg = 15;

  memset (&current_frame_info, 0, sizeof (current_frame_info));
  current_frame_info.spill_cfa_off = -16;
  current_frame_info.n_input_regs = 1;
  current_frame_info.need_regstk = (TARGET_REG_NAMES != 0);

  /* Mark the end of the (empty) prologue.  */
  emit_note (NOTE_INSN_PROLOGUE_END);

  /* Figure out whether "this" will be the first parameter (the
     typical case) or the second parameter (as happens when the
     virtual function returns certain class objects).  */
  this_parmno
    = (ia64_struct_retval_addr_is_first_parm_p (TREE_TYPE (thunk))
       ? 1 : 0);
  this_regno = IN_REG (this_parmno);
  if (!TARGET_REG_NAMES)
    reg_names[this_regno] = ia64_reg_numbers[this_parmno];

  this_rtx = gen_rtx_REG (Pmode, this_regno);

  /* Apply the constant offset, if required.  */
  delta_rtx = GEN_INT (delta);
  if (TARGET_ILP32)
    {
      rtx tmp = gen_rtx_REG (ptr_mode, this_regno);
      REG_POINTER (tmp) = 1;
      if (delta && satisfies_constraint_I (delta_rtx))
	{
	  emit_insn (gen_ptr_extend_plus_imm (this_rtx, tmp, delta_rtx));
	  delta = 0;
	}
      else
	emit_insn (gen_ptr_extend (this_rtx, tmp));
    }
  if (delta)
    {
      if (!satisfies_constraint_I (delta_rtx))
	{
	  rtx tmp = gen_rtx_REG (Pmode, 2);
	  emit_move_insn (tmp, delta_rtx);
	  delta_rtx = tmp;
	}
      emit_insn (gen_adddi3 (this_rtx, this_rtx, delta_rtx));
    }

  /* Apply the offset from the vtable, if required.  */
  if (vcall_offset)
    {
      rtx vcall_offset_rtx = GEN_INT (vcall_offset);
      rtx tmp = gen_rtx_REG (Pmode, 2);

      if (TARGET_ILP32)
	{
	  rtx t = gen_rtx_REG (ptr_mode, 2);
	  REG_POINTER (t) = 1;
	  emit_move_insn (t, gen_rtx_MEM (ptr_mode, this_rtx));
	  if (satisfies_constraint_I (vcall_offset_rtx))
	    {
	      emit_insn (gen_ptr_extend_plus_imm (tmp, t, vcall_offset_rtx));
	      vcall_offset = 0;
	    }
	  else
	    emit_insn (gen_ptr_extend (tmp, t));
	}
      else
	emit_move_insn (tmp, gen_rtx_MEM (Pmode, this_rtx));

      if (vcall_offset)
	{
	  if (!satisfies_constraint_J (vcall_offset_rtx))
	    {
	      rtx tmp2 = gen_rtx_REG (Pmode, next_scratch_gr_reg ());
	      emit_move_insn (tmp2, vcall_offset_rtx);
	      vcall_offset_rtx = tmp2;
	    }
	  emit_insn (gen_adddi3 (tmp, tmp, vcall_offset_rtx));
	}

      if (TARGET_ILP32)
	emit_insn (gen_zero_extendsidi2 (tmp, gen_rtx_MEM (ptr_mode, tmp)));
      else
	emit_move_insn (tmp, gen_rtx_MEM (Pmode, tmp));

      emit_insn (gen_adddi3 (this_rtx, this_rtx, tmp));
    }

  /* Generate a tail call to the target function.  */
  if (! TREE_USED (function))
    {
      assemble_external (function);
      TREE_USED (function) = 1;
    }
  funexp = XEXP (DECL_RTL (function), 0);
  funexp = gen_rtx_MEM (FUNCTION_MODE, funexp);
  ia64_expand_call (NULL_RTX, funexp, NULL_RTX, 1);
  insn = get_last_insn ();
  SIBLING_CALL_P (insn) = 1;

  /* Code generation for calls relies on splitting.  */
  reload_completed = 1;
  epilogue_completed = 1;
  try_split (PATTERN (insn), insn, 0);

  emit_barrier ();

  /* Run just enough of rest_of_compilation to get the insns emitted.
     There's not really enough bulk here to make other passes such as
     instruction scheduling worth while.  Note that use_thunk calls
     assemble_start_function and assemble_end_function.  */

  emit_all_insn_group_barriers (NULL);
  insn = get_insns ();
  shorten_branches (insn);
  final_start_function (insn, file, 1);
  final (insn, file, 1);
  final_end_function ();

  reload_completed = 0;
  epilogue_completed = 0;
}

/* Worker function for TARGET_STRUCT_VALUE_RTX.  */

static rtx
ia64_struct_value_rtx (tree fntype,
		       int incoming ATTRIBUTE_UNUSED)
{
  if (TARGET_ABI_OPEN_VMS ||
      (fntype && ia64_struct_retval_addr_is_first_parm_p (fntype)))
    return NULL_RTX;
  return gen_rtx_REG (Pmode, GR_REG (8));
}

static bool
ia64_scalar_mode_supported_p (enum machine_mode mode)
{
  switch (mode)
    {
    case QImode:
    case HImode:
    case SImode:
    case DImode:
    case TImode:
      return true;

    case SFmode:
    case DFmode:
    case XFmode:
    case RFmode:
      return true;

    case TFmode:
      return true;

    default:
      return false;
    }
}

static bool
ia64_vector_mode_supported_p (enum machine_mode mode)
{
  switch (mode)
    {
    case V8QImode:
    case V4HImode:
    case V2SImode:
      return true;

    case V2SFmode:
      return true;

    default:
      return false;
    }
}

/* Implement the FUNCTION_PROFILER macro.  */

void
ia64_output_function_profiler (FILE *file, int labelno)
{
  bool indirect_call;

  /* If the function needs a static chain and the static chain
     register is r15, we use an indirect call so as to bypass
     the PLT stub in case the executable is dynamically linked,
     because the stub clobbers r15 as per 5.3.6 of the psABI.
     We don't need to do that in non canonical PIC mode.  */

  if (cfun->static_chain_decl && !TARGET_NO_PIC && !TARGET_AUTO_PIC)
    {
      gcc_assert (STATIC_CHAIN_REGNUM == 15);
      indirect_call = true;
    }
  else
    indirect_call = false;

  if (TARGET_GNU_AS)
    fputs ("\t.prologue 4, r40\n", file);
  else
    fputs ("\t.prologue\n\t.save ar.pfs, r40\n", file);
  fputs ("\talloc out0 = ar.pfs, 8, 0, 4, 0\n", file);

  if (NO_PROFILE_COUNTERS)
    fputs ("\tmov out3 = r0\n", file);
  else
    {
      char buf[20];
      ASM_GENERATE_INTERNAL_LABEL (buf, "LP", labelno);

      if (TARGET_AUTO_PIC)
	fputs ("\tmovl out3 = @gprel(", file);
      else
	fputs ("\taddl out3 = @ltoff(", file);
      assemble_name (file, buf);
      if (TARGET_AUTO_PIC)
	fputs (")\n", file);
      else
	fputs ("), r1\n", file);
    }

  if (indirect_call)
    fputs ("\taddl r14 = @ltoff(@fptr(_mcount)), r1\n", file);
  fputs ("\t;;\n", file);

  fputs ("\t.save rp, r42\n", file);
  fputs ("\tmov out2 = b0\n", file);
  if (indirect_call)
    fputs ("\tld8 r14 = [r14]\n\t;;\n", file);
  fputs ("\t.body\n", file);
  fputs ("\tmov out1 = r1\n", file);
  if (indirect_call)
    {
      fputs ("\tld8 r16 = [r14], 8\n\t;;\n", file);
      fputs ("\tmov b6 = r16\n", file);
      fputs ("\tld8 r1 = [r14]\n", file);
      fputs ("\tbr.call.sptk.many b0 = b6\n\t;;\n", file);
    }
  else
    fputs ("\tbr.call.sptk.many b0 = _mcount\n\t;;\n", file);
}

static GTY(()) rtx mcount_func_rtx;
static rtx
gen_mcount_func_rtx (void)
{
  if (!mcount_func_rtx)
    mcount_func_rtx = init_one_libfunc ("_mcount");
  return mcount_func_rtx;
}

void
ia64_profile_hook (int labelno)
{
  rtx label, ip;

  if (NO_PROFILE_COUNTERS)
    label = const0_rtx;
  else
    {
      char buf[30];
      const char *label_name;
      ASM_GENERATE_INTERNAL_LABEL (buf, "LP", labelno);
      label_name = ggc_strdup ((*targetm.strip_name_encoding) (buf));
      label = gen_rtx_SYMBOL_REF (Pmode, label_name);
      SYMBOL_REF_FLAGS (label) = SYMBOL_FLAG_LOCAL;
    }
  ip = gen_reg_rtx (Pmode);
  emit_insn (gen_ip_value (ip));
  emit_library_call (gen_mcount_func_rtx (), LCT_NORMAL,
                     VOIDmode, 3,
		     gen_rtx_REG (Pmode, BR_REG (0)), Pmode,
		     ip, Pmode,
		     label, Pmode);
}

/* Return the mangling of TYPE if it is an extended fundamental type.  */

static const char *
ia64_mangle_type (const_tree type)
{
  type = TYPE_MAIN_VARIANT (type);

  if (TREE_CODE (type) != VOID_TYPE && TREE_CODE (type) != BOOLEAN_TYPE
      && TREE_CODE (type) != INTEGER_TYPE && TREE_CODE (type) != REAL_TYPE)
    return NULL;

  /* On HP-UX, "long double" is mangled as "e" so __float128 is
     mangled as "e".  */
  if (!TARGET_HPUX && TYPE_MODE (type) == TFmode)
    return "g";
  /* On HP-UX, "e" is not available as a mangling of __float80 so use
     an extended mangling.  Elsewhere, "e" is available since long
     double is 80 bits.  */
  if (TYPE_MODE (type) == XFmode)
    return TARGET_HPUX ? "u9__float80" : "e";
  if (TYPE_MODE (type) == RFmode)
    return "u7__fpreg";
  return NULL;
}

/* Return the diagnostic message string if conversion from FROMTYPE to
   TOTYPE is not allowed, NULL otherwise.  */
static const char *
ia64_invalid_conversion (const_tree fromtype, const_tree totype)
{
  /* Reject nontrivial conversion to or from __fpreg.  */
  if (TYPE_MODE (fromtype) == RFmode
      && TYPE_MODE (totype) != RFmode
      && TYPE_MODE (totype) != VOIDmode)
    return N_("invalid conversion from %<__fpreg%>");
  if (TYPE_MODE (totype) == RFmode
      && TYPE_MODE (fromtype) != RFmode)
    return N_("invalid conversion to %<__fpreg%>");
  return NULL;
}

/* Return the diagnostic message string if the unary operation OP is
   not permitted on TYPE, NULL otherwise.  */
static const char *
ia64_invalid_unary_op (int op, const_tree type)
{
  /* Reject operations on __fpreg other than unary + or &.  */
  if (TYPE_MODE (type) == RFmode
      && op != CONVERT_EXPR
      && op != ADDR_EXPR)
    return N_("invalid operation on %<__fpreg%>");
  return NULL;
}

/* Return the diagnostic message string if the binary operation OP is
   not permitted on TYPE1 and TYPE2, NULL otherwise.  */
static const char *
ia64_invalid_binary_op (int op ATTRIBUTE_UNUSED, const_tree type1, const_tree type2)
{
  /* Reject operations on __fpreg.  */
  if (TYPE_MODE (type1) == RFmode || TYPE_MODE (type2) == RFmode)
    return N_("invalid operation on %<__fpreg%>");
  return NULL;
}

/* HP-UX version_id attribute.
   For object foo, if the version_id is set to 1234 put out an alias
   of '.alias foo "foo{1234}"  We can't use "foo{1234}" in anything
   other than an alias statement because it is an illegal symbol name.  */

static tree
ia64_handle_version_id_attribute (tree *node ATTRIBUTE_UNUSED,
                                 tree name ATTRIBUTE_UNUSED,
                                 tree args,
                                 int flags ATTRIBUTE_UNUSED,
                                 bool *no_add_attrs)
{
  tree arg = TREE_VALUE (args);

  if (TREE_CODE (arg) != STRING_CST)
    {
      error("version attribute is not a string");
      *no_add_attrs = true;
      return NULL_TREE;
    }
  return NULL_TREE;
}

/* Target hook for c_mode_for_suffix.  */

static enum machine_mode
ia64_c_mode_for_suffix (char suffix)
{
  if (suffix == 'q')
    return TFmode;
  if (suffix == 'w')
    return XFmode;

  return VOIDmode;
}

static GTY(()) rtx ia64_dconst_0_5_rtx;

rtx
ia64_dconst_0_5 (void)
{
  if (! ia64_dconst_0_5_rtx)
    {
      REAL_VALUE_TYPE rv;
      real_from_string (&rv, "0.5");
      ia64_dconst_0_5_rtx = const_double_from_real_value (rv, DFmode);
    }
  return ia64_dconst_0_5_rtx;
}

static GTY(()) rtx ia64_dconst_0_375_rtx;

rtx
ia64_dconst_0_375 (void)
{
  if (! ia64_dconst_0_375_rtx)
    {
      REAL_VALUE_TYPE rv;
      real_from_string (&rv, "0.375");
      ia64_dconst_0_375_rtx = const_double_from_real_value (rv, DFmode);
    }
  return ia64_dconst_0_375_rtx;
}

static enum machine_mode
ia64_get_reg_raw_mode (int regno)
{
  if (FR_REGNO_P (regno))
    return XFmode;
  return default_get_reg_raw_mode(regno);
}

/* Implement TARGET_MEMBER_TYPE_FORCES_BLK.  ??? Might not be needed
   anymore.  */

bool
ia64_member_type_forces_blk (const_tree, enum machine_mode mode)
{
  return TARGET_HPUX && mode == TFmode;
}

/* Always default to .text section until HP-UX linker is fixed.  */

ATTRIBUTE_UNUSED static section *
ia64_hpux_function_section (tree decl ATTRIBUTE_UNUSED,
			    enum node_frequency freq ATTRIBUTE_UNUSED,
			    bool startup ATTRIBUTE_UNUSED,
			    bool exit ATTRIBUTE_UNUSED)
{
  return NULL;
}

/* Construct (set target (vec_select op0 (parallel perm))) and
   return true if that's a valid instruction in the active ISA.  */

static bool
expand_vselect (rtx target, rtx op0, const unsigned char *perm, unsigned nelt)
{
  rtx rperm[MAX_VECT_LEN], x;
  unsigned i;

  for (i = 0; i < nelt; ++i)
    rperm[i] = GEN_INT (perm[i]);

  x = gen_rtx_PARALLEL (VOIDmode, gen_rtvec_v (nelt, rperm));
  x = gen_rtx_VEC_SELECT (GET_MODE (target), op0, x);
  x = gen_rtx_SET (VOIDmode, target, x);

  x = emit_insn (x);
  if (recog_memoized (x) < 0)
    {
      remove_insn (x);
      return false;
    }
  return true;
}

/* Similar, but generate a vec_concat from op0 and op1 as well.  */

static bool
expand_vselect_vconcat (rtx target, rtx op0, rtx op1,
			const unsigned char *perm, unsigned nelt)
{
  enum machine_mode v2mode;
  rtx x;

  v2mode = GET_MODE_2XWIDER_MODE (GET_MODE (op0));
  x = gen_rtx_VEC_CONCAT (v2mode, op0, op1);
  return expand_vselect (target, x, perm, nelt);
}

/* Try to expand a no-op permutation.  */

static bool
expand_vec_perm_identity (struct expand_vec_perm_d *d)
{
  unsigned i, nelt = d->nelt;

  for (i = 0; i < nelt; ++i)
    if (d->perm[i] != i)
      return false;

  if (!d->testing_p)
    emit_move_insn (d->target, d->op0);

  return true;
}

/* Try to expand D via a shrp instruction.  */

static bool
expand_vec_perm_shrp (struct expand_vec_perm_d *d)
{
  unsigned i, nelt = d->nelt, shift, mask;
  rtx tmp, hi, lo;

  /* ??? Don't force V2SFmode into the integer registers.  */
  if (d->vmode == V2SFmode)
    return false;

  mask = (d->one_operand_p ? nelt - 1 : 2 * nelt - 1);

  shift = d->perm[0];
  if (BYTES_BIG_ENDIAN && shift > nelt)
    return false;

  for (i = 1; i < nelt; ++i)
    if (d->perm[i] != ((shift + i) & mask))
      return false;

  if (d->testing_p)
    return true;

  hi = shift < nelt ? d->op1 : d->op0;
  lo = shift < nelt ? d->op0 : d->op1;

  shift %= nelt;

  shift *= GET_MODE_UNIT_SIZE (d->vmode) * BITS_PER_UNIT;

  /* We've eliminated the shift 0 case via expand_vec_perm_identity.  */
  gcc_assert (IN_RANGE (shift, 1, 63));

  /* Recall that big-endian elements are numbered starting at the top of
     the register.  Ideally we'd have a shift-left-pair.  But since we
     don't, convert to a shift the other direction.  */
  if (BYTES_BIG_ENDIAN)
    shift = 64 - shift;

  tmp = gen_reg_rtx (DImode);
  hi = gen_lowpart (DImode, hi);
  lo = gen_lowpart (DImode, lo);
  emit_insn (gen_shrp (tmp, hi, lo, GEN_INT (shift)));

  emit_move_insn (d->target, gen_lowpart (d->vmode, tmp));
  return true;
}

/* Try to instantiate D in a single instruction.  */

static bool
expand_vec_perm_1 (struct expand_vec_perm_d *d)
{     
  unsigned i, nelt = d->nelt;
  unsigned char perm2[MAX_VECT_LEN];

  /* Try single-operand selections.  */
  if (d->one_operand_p)
    {
      if (expand_vec_perm_identity (d))
	return true;
      if (expand_vselect (d->target, d->op0, d->perm, nelt))
	return true;
    }

  /* Try two operand selections.  */
  if (expand_vselect_vconcat (d->target, d->op0, d->op1, d->perm, nelt))
    return true;

  /* Recognize interleave style patterns with reversed operands.  */
  if (!d->one_operand_p)
    {
      for (i = 0; i < nelt; ++i)
	{
	  unsigned e = d->perm[i];
	  if (e >= nelt)
	    e -= nelt;
	  else
	    e += nelt;
	  perm2[i] = e;
	}

      if (expand_vselect_vconcat (d->target, d->op1, d->op0, perm2, nelt))
	return true;
    }

  if (expand_vec_perm_shrp (d))
    return true;

  /* ??? Look for deposit-like permutations where most of the result 
     comes from one vector unchanged and the rest comes from a 
     sequential hunk of the other vector.  */

  return false;
}

/* Pattern match broadcast permutations.  */

static bool
expand_vec_perm_broadcast (struct expand_vec_perm_d *d)
{
  unsigned i, elt, nelt = d->nelt;
  unsigned char perm2[2];
  rtx temp;
  bool ok;

  if (!d->one_operand_p)
    return false;

  elt = d->perm[0];
  for (i = 1; i < nelt; ++i)
    if (d->perm[i] != elt)
      return false;

  switch (d->vmode)
    {
    case V2SImode:
    case V2SFmode:
      /* Implementable by interleave.  */
      perm2[0] = elt;
      perm2[1] = elt + 2;
      ok = expand_vselect_vconcat (d->target, d->op0, d->op0, perm2, 2);
      gcc_assert (ok);
      break;

    case V8QImode:
      /* Implementable by extract + broadcast.  */
      if (BYTES_BIG_ENDIAN)
	elt = 7 - elt;
      elt *= BITS_PER_UNIT;
      temp = gen_reg_rtx (DImode);
      emit_insn (gen_extzv (temp, gen_lowpart (DImode, d->op0),
			    GEN_INT (8), GEN_INT (elt)));
      emit_insn (gen_mux1_brcst_qi (d->target, gen_lowpart (QImode, temp)));
      break;

    case V4HImode:
      /* Should have been matched directly by vec_select.  */
    default:
      gcc_unreachable ();
    }

  return true;
}

/* A subroutine of ia64_expand_vec_perm_const_1.  Try to simplify a
   two vector permutation into a single vector permutation by using
   an interleave operation to merge the vectors.  */

static bool
expand_vec_perm_interleave_2 (struct expand_vec_perm_d *d)
{
  struct expand_vec_perm_d dremap, dfinal;
  unsigned char remap[2 * MAX_VECT_LEN];
  unsigned contents, i, nelt, nelt2;
  unsigned h0, h1, h2, h3;
  rtx seq;
  bool ok;

  if (d->one_operand_p)
    return false;

  nelt = d->nelt;
  nelt2 = nelt / 2;

  /* Examine from whence the elements come.  */
  contents = 0;
  for (i = 0; i < nelt; ++i)
    contents |= 1u << d->perm[i];

  memset (remap, 0xff, sizeof (remap));
  dremap = *d;

  h0 = (1u << nelt2) - 1;
  h1 = h0 << nelt2;
  h2 = h0 << nelt;
  h3 = h0 << (nelt + nelt2);
  
  if ((contents & (h0 | h2)) == contents)	/* punpck even halves */
    {
      for (i = 0; i < nelt; ++i)
	{
	  unsigned which = i / 2 + (i & 1 ? nelt : 0);
	  remap[which] = i;
	  dremap.perm[i] = which;
	}
    }
  else if ((contents & (h1 | h3)) == contents)	/* punpck odd halves */
    {
      for (i = 0; i < nelt; ++i)
	{
	  unsigned which = i / 2 + nelt2 + (i & 1 ? nelt : 0);
	  remap[which] = i;
	  dremap.perm[i] = which;
	}
    }
  else if ((contents & 0x5555) == contents)	/* mix even elements */
    {
      for (i = 0; i < nelt; ++i)
	{
	  unsigned which = (i & ~1) + (i & 1 ? nelt : 0);
	  remap[which] = i;
	  dremap.perm[i] = which;
	}
    }
  else if ((contents & 0xaaaa) == contents)	/* mix odd elements */
    {
      for (i = 0; i < nelt; ++i)
	{
	  unsigned which = (i | 1) + (i & 1 ? nelt : 0);
	  remap[which] = i;
	  dremap.perm[i] = which;
	}
    }
  else if (floor_log2 (contents) - ctz_hwi (contents) < (int)nelt) /* shrp */
    {
      unsigned shift = ctz_hwi (contents);
      for (i = 0; i < nelt; ++i)
	{
	  unsigned which = (i + shift) & (2 * nelt - 1);
	  remap[which] = i;
	  dremap.perm[i] = which;
	}
    }
  else
    return false;

  /* Use the remapping array set up above to move the elements from their
     swizzled locations into their final destinations.  */
  dfinal = *d;
  for (i = 0; i < nelt; ++i)
    {
      unsigned e = remap[d->perm[i]];
      gcc_assert (e < nelt);
      dfinal.perm[i] = e;
    }
  dfinal.op0 = gen_reg_rtx (dfinal.vmode);
  dfinal.op1 = dfinal.op0;
  dfinal.one_operand_p = true;
  dremap.target = dfinal.op0;

  /* Test if the final remap can be done with a single insn.  For V4HImode
     this *will* succeed.  For V8QImode or V2SImode it may not.  */
  start_sequence ();
  ok = expand_vec_perm_1 (&dfinal);
  seq = get_insns ();
  end_sequence ();
  if (!ok)
    return false;
  if (d->testing_p)
    return true;

  ok = expand_vec_perm_1 (&dremap);
  gcc_assert (ok);

  emit_insn (seq);
  return true;
}

/* A subroutine of ia64_expand_vec_perm_const_1.  Emit a full V4HImode
   constant permutation via two mux2 and a merge.  */

static bool
expand_vec_perm_v4hi_5 (struct expand_vec_perm_d *d)
{
  unsigned char perm2[4];
  rtx rmask[4];
  unsigned i;
  rtx t0, t1, mask, x;
  bool ok;

  if (d->vmode != V4HImode || d->one_operand_p)
    return false;
  if (d->testing_p)
    return true;

  for (i = 0; i < 4; ++i)
    {
      perm2[i] = d->perm[i] & 3;
      rmask[i] = (d->perm[i] & 4 ? const0_rtx : constm1_rtx);
    }
  mask = gen_rtx_CONST_VECTOR (V4HImode, gen_rtvec_v (4, rmask));
  mask = force_reg (V4HImode, mask);

  t0 = gen_reg_rtx (V4HImode);
  t1 = gen_reg_rtx (V4HImode);

  ok = expand_vselect (t0, d->op0, perm2, 4);
  gcc_assert (ok);
  ok = expand_vselect (t1, d->op1, perm2, 4);
  gcc_assert (ok);

  x = gen_rtx_AND (V4HImode, mask, t0);
  emit_insn (gen_rtx_SET (VOIDmode, t0, x));

  x = gen_rtx_NOT (V4HImode, mask);
  x = gen_rtx_AND (V4HImode, x, t1);
  emit_insn (gen_rtx_SET (VOIDmode, t1, x));

  x = gen_rtx_IOR (V4HImode, t0, t1);
  emit_insn (gen_rtx_SET (VOIDmode, d->target, x));

  return true;
}

/* The guts of ia64_expand_vec_perm_const, also used by the ok hook.
   With all of the interface bits taken care of, perform the expansion
   in D and return true on success.  */

static bool
ia64_expand_vec_perm_const_1 (struct expand_vec_perm_d *d)
{
  if (expand_vec_perm_1 (d))
    return true;
  if (expand_vec_perm_broadcast (d))
    return true;
  if (expand_vec_perm_interleave_2 (d))
    return true;
  if (expand_vec_perm_v4hi_5 (d))
    return true;
  return false;
}

bool
ia64_expand_vec_perm_const (rtx operands[4])
{
  struct expand_vec_perm_d d;
  unsigned char perm[MAX_VECT_LEN];
  int i, nelt, which;
  rtx sel;

  d.target = operands[0];
  d.op0 = operands[1];
  d.op1 = operands[2];
  sel = operands[3];

  d.vmode = GET_MODE (d.target);
  gcc_assert (VECTOR_MODE_P (d.vmode));
  d.nelt = nelt = GET_MODE_NUNITS (d.vmode);
  d.testing_p = false;

  gcc_assert (GET_CODE (sel) == CONST_VECTOR);
  gcc_assert (XVECLEN (sel, 0) == nelt);
  gcc_checking_assert (sizeof (d.perm) == sizeof (perm));

  for (i = which = 0; i < nelt; ++i)
    {
      rtx e = XVECEXP (sel, 0, i);
      int ei = INTVAL (e) & (2 * nelt - 1);

      which |= (ei < nelt ? 1 : 2);
      d.perm[i] = ei;
      perm[i] = ei;
    }

  switch (which)
    {
    default:
      gcc_unreachable();

    case 3:
      if (!rtx_equal_p (d.op0, d.op1))
	{
	  d.one_operand_p = false;
	  break;
	}

      /* The elements of PERM do not suggest that only the first operand
	 is used, but both operands are identical.  Allow easier matching
	 of the permutation by folding the permutation into the single
	 input vector.  */
      for (i = 0; i < nelt; ++i)
	if (d.perm[i] >= nelt)
	  d.perm[i] -= nelt;
      /* FALLTHRU */

    case 1:
      d.op1 = d.op0;
      d.one_operand_p = true;
      break;

    case 2:
      for (i = 0; i < nelt; ++i)
        d.perm[i] -= nelt;
      d.op0 = d.op1;
      d.one_operand_p = true;
      break;
    }

  if (ia64_expand_vec_perm_const_1 (&d))
    return true;

  /* If the mask says both arguments are needed, but they are the same,
     the above tried to expand with one_operand_p true.  If that didn't
     work, retry with one_operand_p false, as that's what we used in _ok.  */
  if (which == 3 && d.one_operand_p)
    {
      memcpy (d.perm, perm, sizeof (perm));
      d.one_operand_p = false;
      return ia64_expand_vec_perm_const_1 (&d);
    }

  return false;
}

/* Implement targetm.vectorize.vec_perm_const_ok.  */

static bool
ia64_vectorize_vec_perm_const_ok (enum machine_mode vmode,
				  const unsigned char *sel)
{
  struct expand_vec_perm_d d;
  unsigned int i, nelt, which;
  bool ret;

  d.vmode = vmode;
  d.nelt = nelt = GET_MODE_NUNITS (d.vmode);
  d.testing_p = true;

  /* Extract the values from the vector CST into the permutation
     array in D.  */
  memcpy (d.perm, sel, nelt);
  for (i = which = 0; i < nelt; ++i)
    {
      unsigned char e = d.perm[i];
      gcc_assert (e < 2 * nelt);
      which |= (e < nelt ? 1 : 2);
    }

  /* For all elements from second vector, fold the elements to first.  */
  if (which == 2)
    for (i = 0; i < nelt; ++i)
      d.perm[i] -= nelt;

  /* Check whether the mask can be applied to the vector type.  */
  d.one_operand_p = (which != 3);

  /* Otherwise we have to go through the motions and see if we can
     figure out how to generate the requested permutation.  */
  d.target = gen_raw_REG (d.vmode, LAST_VIRTUAL_REGISTER + 1);
  d.op1 = d.op0 = gen_raw_REG (d.vmode, LAST_VIRTUAL_REGISTER + 2);
  if (!d.one_operand_p)
    d.op1 = gen_raw_REG (d.vmode, LAST_VIRTUAL_REGISTER + 3);

  start_sequence ();
  ret = ia64_expand_vec_perm_const_1 (&d);
  end_sequence ();

  return ret;
}

void
ia64_expand_vec_setv2sf (rtx operands[3])
{
  struct expand_vec_perm_d d;
  unsigned int which;
  bool ok;
  
  d.target = operands[0];
  d.op0 = operands[0];
  d.op1 = gen_reg_rtx (V2SFmode);
  d.vmode = V2SFmode;
  d.nelt = 2;
  d.one_operand_p = false;
  d.testing_p = false;

  which = INTVAL (operands[2]);
  gcc_assert (which <= 1);
  d.perm[0] = 1 - which;
  d.perm[1] = which + 2;

  emit_insn (gen_fpack (d.op1, operands[1], CONST0_RTX (SFmode)));

  ok = ia64_expand_vec_perm_const_1 (&d);
  gcc_assert (ok);
}

void
ia64_expand_vec_perm_even_odd (rtx target, rtx op0, rtx op1, int odd)
{
  struct expand_vec_perm_d d;
  enum machine_mode vmode = GET_MODE (target);
  unsigned int i, nelt = GET_MODE_NUNITS (vmode);
  bool ok;

  d.target = target;
  d.op0 = op0;
  d.op1 = op1;
  d.vmode = vmode;
  d.nelt = nelt;
  d.one_operand_p = false;
  d.testing_p = false;

  for (i = 0; i < nelt; ++i)
    d.perm[i] = i * 2 + odd;

  ok = ia64_expand_vec_perm_const_1 (&d);
  gcc_assert (ok);
}

#include "gt-ia64.h"