aboutsummaryrefslogtreecommitdiffstats
path: root/contrib/tools/python3/src/Python/dtoa.c
blob: 733e70bc7916984628d0e4601b60aeeb319c47d3 (plain) (blame)
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
/****************************************************************
 *
 * The author of this software is David M. Gay.
 *
 * Copyright (c) 1991, 2000, 2001 by Lucent Technologies.
 *
 * Permission to use, copy, modify, and distribute this software for any
 * purpose without fee is hereby granted, provided that this entire notice
 * is included in all copies of any software which is or includes a copy
 * or modification of this software and in all copies of the supporting
 * documentation for such software.
 *
 * THIS SOFTWARE IS BEING PROVIDED "AS IS", WITHOUT ANY EXPRESS OR IMPLIED
 * WARRANTY.  IN PARTICULAR, NEITHER THE AUTHOR NOR LUCENT MAKES ANY
 * REPRESENTATION OR WARRANTY OF ANY KIND CONCERNING THE MERCHANTABILITY
 * OF THIS SOFTWARE OR ITS FITNESS FOR ANY PARTICULAR PURPOSE.
 *
 ***************************************************************/

/****************************************************************
 * This is dtoa.c by David M. Gay, downloaded from
 * http://www.netlib.org/fp/dtoa.c on April 15, 2009 and modified for
 * inclusion into the Python core by Mark E. T. Dickinson and Eric V. Smith.
 *
 * Please remember to check http://www.netlib.org/fp regularly (and especially
 * before any Python release) for bugfixes and updates.
 *
 * The major modifications from Gay's original code are as follows:
 *
 *  0. The original code has been specialized to Python's needs by removing
 *     many of the #ifdef'd sections.  In particular, code to support VAX and
 *     IBM floating-point formats, hex NaNs, hex floats, locale-aware
 *     treatment of the decimal point, and setting of the inexact flag have
 *     been removed.
 *
 *  1. We use PyMem_Malloc and PyMem_Free in place of malloc and free.
 *
 *  2. The public functions strtod, dtoa and freedtoa all now have
 *     a _Py_dg_ prefix.
 *
 *  3. Instead of assuming that PyMem_Malloc always succeeds, we thread
 *     PyMem_Malloc failures through the code.  The functions
 *
 *       Balloc, multadd, s2b, i2b, mult, pow5mult, lshift, diff, d2b
 *
 *     of return type *Bigint all return NULL to indicate a malloc failure.
 *     Similarly, rv_alloc and nrv_alloc (return type char *) return NULL on
 *     failure.  bigcomp now has return type int (it used to be void) and
 *     returns -1 on failure and 0 otherwise.  _Py_dg_dtoa returns NULL
 *     on failure.  _Py_dg_strtod indicates failure due to malloc failure
 *     by returning -1.0, setting errno=ENOMEM and *se to s00.
 *
 *  4. The static variable dtoa_result has been removed.  Callers of
 *     _Py_dg_dtoa are expected to call _Py_dg_freedtoa to free
 *     the memory allocated by _Py_dg_dtoa.
 *
 *  5. The code has been reformatted to better fit with Python's
 *     C style guide (PEP 7).
 *
 *  6. A bug in the memory allocation has been fixed: to avoid FREEing memory
 *     that hasn't been MALLOC'ed, private_mem should only be used when k <=
 *     Kmax.
 *
 *  7. _Py_dg_strtod has been modified so that it doesn't accept strings with
 *     leading whitespace.
 *
 *  8. A corner case where _Py_dg_dtoa didn't strip trailing zeros has been
 *     fixed. (bugs.python.org/issue40780)
 *
 ***************************************************************/

/* Please send bug reports for the original dtoa.c code to David M. Gay (dmg
 * at acm dot org, with " at " changed at "@" and " dot " changed to ".").
 * Please report bugs for this modified version using the Python issue tracker
 * (http://bugs.python.org). */

/* On a machine with IEEE extended-precision registers, it is
 * necessary to specify double-precision (53-bit) rounding precision
 * before invoking strtod or dtoa.  If the machine uses (the equivalent
 * of) Intel 80x87 arithmetic, the call
 *      _control87(PC_53, MCW_PC);
 * does this with many compilers.  Whether this or another call is
 * appropriate depends on the compiler; for this to work, it may be
 * necessary to #include "float.h" or another system-dependent header
 * file.
 */

/* strtod for IEEE-, VAX-, and IBM-arithmetic machines.
 *
 * This strtod returns a nearest machine number to the input decimal
 * string (or sets errno to ERANGE).  With IEEE arithmetic, ties are
 * broken by the IEEE round-even rule.  Otherwise ties are broken by
 * biased rounding (add half and chop).
 *
 * Inspired loosely by William D. Clinger's paper "How to Read Floating
 * Point Numbers Accurately" [Proc. ACM SIGPLAN '90, pp. 92-101].
 *
 * Modifications:
 *
 *      1. We only require IEEE, IBM, or VAX double-precision
 *              arithmetic (not IEEE double-extended).
 *      2. We get by with floating-point arithmetic in a case that
 *              Clinger missed -- when we're computing d * 10^n
 *              for a small integer d and the integer n is not too
 *              much larger than 22 (the maximum integer k for which
 *              we can represent 10^k exactly), we may be able to
 *              compute (d*10^k) * 10^(e-k) with just one roundoff.
 *      3. Rather than a bit-at-a-time adjustment of the binary
 *              result in the hard case, we use floating-point
 *              arithmetic to determine the adjustment to within
 *              one bit; only in really hard cases do we need to
 *              compute a second residual.
 *      4. Because of 3., we don't need a large table of powers of 10
 *              for ten-to-e (just some small tables, e.g. of 10^k
 *              for 0 <= k <= 22).
 */

/* Linking of Python's #defines to Gay's #defines starts here. */

#include "Python.h"
#include "pycore_dtoa.h"          // _PY_SHORT_FLOAT_REPR
#include <stdlib.h>               // exit()

/* if _PY_SHORT_FLOAT_REPR == 0, then don't even try to compile
   the following code */
#if _PY_SHORT_FLOAT_REPR == 1

#include "float.h"

#define MALLOC PyMem_Malloc
#define FREE PyMem_Free

/* This code should also work for ARM mixed-endian format on little-endian
   machines, where doubles have byte order 45670123 (in increasing address
   order, 0 being the least significant byte). */
#ifdef DOUBLE_IS_LITTLE_ENDIAN_IEEE754
#  define IEEE_8087
#endif
#if defined(DOUBLE_IS_BIG_ENDIAN_IEEE754) ||  \
  defined(DOUBLE_IS_ARM_MIXED_ENDIAN_IEEE754)
#  define IEEE_MC68k
#endif
#if defined(IEEE_8087) + defined(IEEE_MC68k) != 1
#error "Exactly one of IEEE_8087 or IEEE_MC68k should be defined."
#endif

/* The code below assumes that the endianness of integers matches the
   endianness of the two 32-bit words of a double.  Check this. */
#if defined(WORDS_BIGENDIAN) && (defined(DOUBLE_IS_LITTLE_ENDIAN_IEEE754) || \
                                 defined(DOUBLE_IS_ARM_MIXED_ENDIAN_IEEE754))
#error "doubles and ints have incompatible endianness"
#endif

#if !defined(WORDS_BIGENDIAN) && defined(DOUBLE_IS_BIG_ENDIAN_IEEE754)
#error "doubles and ints have incompatible endianness"
#endif


typedef uint32_t ULong;
typedef int32_t Long;
typedef uint64_t ULLong;

#undef DEBUG
#ifdef Py_DEBUG
#define DEBUG
#endif

/* End Python #define linking */

#ifdef DEBUG
#define Bug(x) {fprintf(stderr, "%s\n", x); exit(1);}
#endif

#ifndef PRIVATE_MEM
#define PRIVATE_MEM 2304
#endif
#define PRIVATE_mem ((PRIVATE_MEM+sizeof(double)-1)/sizeof(double))
static double private_mem[PRIVATE_mem], *pmem_next = private_mem;

#ifdef __cplusplus
extern "C" {
#endif

typedef union { double d; ULong L[2]; } U;

#ifdef IEEE_8087
#define word0(x) (x)->L[1]
#define word1(x) (x)->L[0]
#else
#define word0(x) (x)->L[0]
#define word1(x) (x)->L[1]
#endif
#define dval(x) (x)->d

#ifndef STRTOD_DIGLIM
#define STRTOD_DIGLIM 40
#endif

/* maximum permitted exponent value for strtod; exponents larger than
   MAX_ABS_EXP in absolute value get truncated to +-MAX_ABS_EXP.  MAX_ABS_EXP
   should fit into an int. */
#ifndef MAX_ABS_EXP
#define MAX_ABS_EXP 1100000000U
#endif
/* Bound on length of pieces of input strings in _Py_dg_strtod; specifically,
   this is used to bound the total number of digits ignoring leading zeros and
   the number of digits that follow the decimal point.  Ideally, MAX_DIGITS
   should satisfy MAX_DIGITS + 400 < MAX_ABS_EXP; that ensures that the
   exponent clipping in _Py_dg_strtod can't affect the value of the output. */
#ifndef MAX_DIGITS
#define MAX_DIGITS 1000000000U
#endif

/* Guard against trying to use the above values on unusual platforms with ints
 * of width less than 32 bits. */
#if MAX_ABS_EXP > INT_MAX
#error "MAX_ABS_EXP should fit in an int"
#endif
#if MAX_DIGITS > INT_MAX
#error "MAX_DIGITS should fit in an int"
#endif

/* The following definition of Storeinc is appropriate for MIPS processors.
 * An alternative that might be better on some machines is
 * #define Storeinc(a,b,c) (*a++ = b << 16 | c & 0xffff)
 */
#if defined(IEEE_8087)
#define Storeinc(a,b,c) (((unsigned short *)a)[1] = (unsigned short)b,  \
                         ((unsigned short *)a)[0] = (unsigned short)c, a++)
#else
#define Storeinc(a,b,c) (((unsigned short *)a)[0] = (unsigned short)b,  \
                         ((unsigned short *)a)[1] = (unsigned short)c, a++)
#endif

/* #define P DBL_MANT_DIG */
/* Ten_pmax = floor(P*log(2)/log(5)) */
/* Bletch = (highest power of 2 < DBL_MAX_10_EXP) / 16 */
/* Quick_max = floor((P-1)*log(FLT_RADIX)/log(10) - 1) */
/* Int_max = floor(P*log(FLT_RADIX)/log(10) - 1) */

#define Exp_shift  20
#define Exp_shift1 20
#define Exp_msk1    0x100000
#define Exp_msk11   0x100000
#define Exp_mask  0x7ff00000
#define P 53
#define Nbits 53
#define Bias 1023
#define Emax 1023
#define Emin (-1022)
#define Etiny (-1074)  /* smallest denormal is 2**Etiny */
#define Exp_1  0x3ff00000
#define Exp_11 0x3ff00000
#define Ebits 11
#define Frac_mask  0xfffff
#define Frac_mask1 0xfffff
#define Ten_pmax 22
#define Bletch 0x10
#define Bndry_mask  0xfffff
#define Bndry_mask1 0xfffff
#define Sign_bit 0x80000000
#define Log2P 1
#define Tiny0 0
#define Tiny1 1
#define Quick_max 14
#define Int_max 14

#ifndef Flt_Rounds
#ifdef FLT_ROUNDS
#define Flt_Rounds FLT_ROUNDS
#else
#define Flt_Rounds 1
#endif
#endif /*Flt_Rounds*/

#define Rounding Flt_Rounds

#define Big0 (Frac_mask1 | Exp_msk1*(DBL_MAX_EXP+Bias-1))
#define Big1 0xffffffff

/* Standard NaN used by _Py_dg_stdnan. */

#define NAN_WORD0 0x7ff80000
#define NAN_WORD1 0

/* Bits of the representation of positive infinity. */

#define POSINF_WORD0 0x7ff00000
#define POSINF_WORD1 0

/* struct BCinfo is used to pass information from _Py_dg_strtod to bigcomp */

typedef struct BCinfo BCinfo;
struct
BCinfo {
    int e0, nd, nd0, scale;
};

#define FFFFFFFF 0xffffffffUL

#define Kmax 7

/* struct Bigint is used to represent arbitrary-precision integers.  These
   integers are stored in sign-magnitude format, with the magnitude stored as
   an array of base 2**32 digits.  Bigints are always normalized: if x is a
   Bigint then x->wds >= 1, and either x->wds == 1 or x[wds-1] is nonzero.

   The Bigint fields are as follows:

     - next is a header used by Balloc and Bfree to keep track of lists
         of freed Bigints;  it's also used for the linked list of
         powers of 5 of the form 5**2**i used by pow5mult.
     - k indicates which pool this Bigint was allocated from
     - maxwds is the maximum number of words space was allocated for
       (usually maxwds == 2**k)
     - sign is 1 for negative Bigints, 0 for positive.  The sign is unused
       (ignored on inputs, set to 0 on outputs) in almost all operations
       involving Bigints: a notable exception is the diff function, which
       ignores signs on inputs but sets the sign of the output correctly.
     - wds is the actual number of significant words
     - x contains the vector of words (digits) for this Bigint, from least
       significant (x[0]) to most significant (x[wds-1]).
*/

struct
Bigint {
    struct Bigint *next;
    int k, maxwds, sign, wds;
    ULong x[1];
};

typedef struct Bigint Bigint;

#ifndef Py_USING_MEMORY_DEBUGGER

/* Memory management: memory is allocated from, and returned to, Kmax+1 pools
   of memory, where pool k (0 <= k <= Kmax) is for Bigints b with b->maxwds ==
   1 << k.  These pools are maintained as linked lists, with freelist[k]
   pointing to the head of the list for pool k.

   On allocation, if there's no free slot in the appropriate pool, MALLOC is
   called to get more memory.  This memory is not returned to the system until
   Python quits.  There's also a private memory pool that's allocated from
   in preference to using MALLOC.

   For Bigints with more than (1 << Kmax) digits (which implies at least 1233
   decimal digits), memory is directly allocated using MALLOC, and freed using
   FREE.

   XXX: it would be easy to bypass this memory-management system and
   translate each call to Balloc into a call to PyMem_Malloc, and each
   Bfree to PyMem_Free.  Investigate whether this has any significant
   performance on impact. */

static Bigint *freelist[Kmax+1];

/* Allocate space for a Bigint with up to 1<<k digits */

static Bigint *
Balloc(int k)
{
    int x;
    Bigint *rv;
    unsigned int len;

    if (k <= Kmax && (rv = freelist[k]))
        freelist[k] = rv->next;
    else {
        x = 1 << k;
        len = (sizeof(Bigint) + (x-1)*sizeof(ULong) + sizeof(double) - 1)
            /sizeof(double);
        if (k <= Kmax && pmem_next - private_mem + len <= (Py_ssize_t)PRIVATE_mem) {
            rv = (Bigint*)pmem_next;
            pmem_next += len;
        }
        else {
            rv = (Bigint*)MALLOC(len*sizeof(double));
            if (rv == NULL)
                return NULL;
        }
        rv->k = k;
        rv->maxwds = x;
    }
    rv->sign = rv->wds = 0;
    return rv;
}

/* Free a Bigint allocated with Balloc */

static void
Bfree(Bigint *v)
{
    if (v) {
        if (v->k > Kmax)
            FREE((void*)v);
        else {
            v->next = freelist[v->k];
            freelist[v->k] = v;
        }
    }
}

#else

/* Alternative versions of Balloc and Bfree that use PyMem_Malloc and
   PyMem_Free directly in place of the custom memory allocation scheme above.
   These are provided for the benefit of memory debugging tools like
   Valgrind. */

/* Allocate space for a Bigint with up to 1<<k digits */

static Bigint *
Balloc(int k)
{
    int x;
    Bigint *rv;
    unsigned int len;

    x = 1 << k;
    len = (sizeof(Bigint) + (x-1)*sizeof(ULong) + sizeof(double) - 1)
        /sizeof(double);

    rv = (Bigint*)MALLOC(len*sizeof(double));
    if (rv == NULL)
        return NULL;

    rv->k = k;
    rv->maxwds = x;
    rv->sign = rv->wds = 0;
    return rv;
}

/* Free a Bigint allocated with Balloc */

static void
Bfree(Bigint *v)
{
    if (v) {
        FREE((void*)v);
    }
}

#endif /* Py_USING_MEMORY_DEBUGGER */

#define Bcopy(x,y) memcpy((char *)&x->sign, (char *)&y->sign,   \
                          y->wds*sizeof(Long) + 2*sizeof(int))

/* Multiply a Bigint b by m and add a.  Either modifies b in place and returns
   a pointer to the modified b, or Bfrees b and returns a pointer to a copy.
   On failure, return NULL.  In this case, b will have been already freed. */

static Bigint *
multadd(Bigint *b, int m, int a)       /* multiply by m and add a */
{
    int i, wds;
    ULong *x;
    ULLong carry, y;
    Bigint *b1;

    wds = b->wds;
    x = b->x;
    i = 0;
    carry = a;
    do {
        y = *x * (ULLong)m + carry;
        carry = y >> 32;
        *x++ = (ULong)(y & FFFFFFFF);
    }
    while(++i < wds);
    if (carry) {
        if (wds >= b->maxwds) {
            b1 = Balloc(b->k+1);
            if (b1 == NULL){
                Bfree(b);
                return NULL;
            }
            Bcopy(b1, b);
            Bfree(b);
            b = b1;
        }
        b->x[wds++] = (ULong)carry;
        b->wds = wds;
    }
    return b;
}

/* convert a string s containing nd decimal digits (possibly containing a
   decimal separator at position nd0, which is ignored) to a Bigint.  This
   function carries on where the parsing code in _Py_dg_strtod leaves off: on
   entry, y9 contains the result of converting the first 9 digits.  Returns
   NULL on failure. */

static Bigint *
s2b(const char *s, int nd0, int nd, ULong y9)
{
    Bigint *b;
    int i, k;
    Long x, y;

    x = (nd + 8) / 9;
    for(k = 0, y = 1; x > y; y <<= 1, k++) ;
    b = Balloc(k);
    if (b == NULL)
        return NULL;
    b->x[0] = y9;
    b->wds = 1;

    if (nd <= 9)
      return b;

    s += 9;
    for (i = 9; i < nd0; i++) {
        b = multadd(b, 10, *s++ - '0');
        if (b == NULL)
            return NULL;
    }
    s++;
    for(; i < nd; i++) {
        b = multadd(b, 10, *s++ - '0');
        if (b == NULL)
            return NULL;
    }
    return b;
}

/* count leading 0 bits in the 32-bit integer x. */

static int
hi0bits(ULong x)
{
    int k = 0;

    if (!(x & 0xffff0000)) {
        k = 16;
        x <<= 16;
    }
    if (!(x & 0xff000000)) {
        k += 8;
        x <<= 8;
    }
    if (!(x & 0xf0000000)) {
        k += 4;
        x <<= 4;
    }
    if (!(x & 0xc0000000)) {
        k += 2;
        x <<= 2;
    }
    if (!(x & 0x80000000)) {
        k++;
        if (!(x & 0x40000000))
            return 32;
    }
    return k;
}

/* count trailing 0 bits in the 32-bit integer y, and shift y right by that
   number of bits. */

static int
lo0bits(ULong *y)
{
    int k;
    ULong x = *y;

    if (x & 7) {
        if (x & 1)
            return 0;
        if (x & 2) {
            *y = x >> 1;
            return 1;
        }
        *y = x >> 2;
        return 2;
    }
    k = 0;
    if (!(x & 0xffff)) {
        k = 16;
        x >>= 16;
    }
    if (!(x & 0xff)) {
        k += 8;
        x >>= 8;
    }
    if (!(x & 0xf)) {
        k += 4;
        x >>= 4;
    }
    if (!(x & 0x3)) {
        k += 2;
        x >>= 2;
    }
    if (!(x & 1)) {
        k++;
        x >>= 1;
        if (!x)
            return 32;
    }
    *y = x;
    return k;
}

/* convert a small nonnegative integer to a Bigint */

static Bigint *
i2b(int i)
{
    Bigint *b;

    b = Balloc(1);
    if (b == NULL)
        return NULL;
    b->x[0] = i;
    b->wds = 1;
    return b;
}

/* multiply two Bigints.  Returns a new Bigint, or NULL on failure.  Ignores
   the signs of a and b. */

static Bigint *
mult(Bigint *a, Bigint *b)
{
    Bigint *c;
    int k, wa, wb, wc;
    ULong *x, *xa, *xae, *xb, *xbe, *xc, *xc0;
    ULong y;
    ULLong carry, z;

    if ((!a->x[0] && a->wds == 1) || (!b->x[0] && b->wds == 1)) {
        c = Balloc(0);
        if (c == NULL)
            return NULL;
        c->wds = 1;
        c->x[0] = 0;
        return c;
    }

    if (a->wds < b->wds) {
        c = a;
        a = b;
        b = c;
    }
    k = a->k;
    wa = a->wds;
    wb = b->wds;
    wc = wa + wb;
    if (wc > a->maxwds)
        k++;
    c = Balloc(k);
    if (c == NULL)
        return NULL;
    for(x = c->x, xa = x + wc; x < xa; x++)
        *x = 0;
    xa = a->x;
    xae = xa + wa;
    xb = b->x;
    xbe = xb + wb;
    xc0 = c->x;
    for(; xb < xbe; xc0++) {
        if ((y = *xb++)) {
            x = xa;
            xc = xc0;
            carry = 0;
            do {
                z = *x++ * (ULLong)y + *xc + carry;
                carry = z >> 32;
                *xc++ = (ULong)(z & FFFFFFFF);
            }
            while(x < xae);
            *xc = (ULong)carry;
        }
    }
    for(xc0 = c->x, xc = xc0 + wc; wc > 0 && !*--xc; --wc) ;
    c->wds = wc;
    return c;
}

#ifndef Py_USING_MEMORY_DEBUGGER

/* p5s is a linked list of powers of 5 of the form 5**(2**i), i >= 2 */

static Bigint *p5s;

/* multiply the Bigint b by 5**k.  Returns a pointer to the result, or NULL on
   failure; if the returned pointer is distinct from b then the original
   Bigint b will have been Bfree'd.   Ignores the sign of b. */

static Bigint *
pow5mult(Bigint *b, int k)
{
    Bigint *b1, *p5, *p51;
    int i;
    static const int p05[3] = { 5, 25, 125 };

    if ((i = k & 3)) {
        b = multadd(b, p05[i-1], 0);
        if (b == NULL)
            return NULL;
    }

    if (!(k >>= 2))
        return b;
    p5 = p5s;
    if (!p5) {
        /* first time */
        p5 = i2b(625);
        if (p5 == NULL) {
            Bfree(b);
            return NULL;
        }
        p5s = p5;
        p5->next = 0;
    }
    for(;;) {
        if (k & 1) {
            b1 = mult(b, p5);
            Bfree(b);
            b = b1;
            if (b == NULL)
                return NULL;
        }
        if (!(k >>= 1))
            break;
        p51 = p5->next;
        if (!p51) {
            p51 = mult(p5,p5);
            if (p51 == NULL) {
                Bfree(b);
                return NULL;
            }
            p51->next = 0;
            p5->next = p51;
        }
        p5 = p51;
    }
    return b;
}

#else

/* Version of pow5mult that doesn't cache powers of 5. Provided for
   the benefit of memory debugging tools like Valgrind. */

static Bigint *
pow5mult(Bigint *b, int k)
{
    Bigint *b1, *p5, *p51;
    int i;
    static const int p05[3] = { 5, 25, 125 };

    if ((i = k & 3)) {
        b = multadd(b, p05[i-1], 0);
        if (b == NULL)
            return NULL;
    }

    if (!(k >>= 2))
        return b;
    p5 = i2b(625);
    if (p5 == NULL) {
        Bfree(b);
        return NULL;
    }

    for(;;) {
        if (k & 1) {
            b1 = mult(b, p5);
            Bfree(b);
            b = b1;
            if (b == NULL) {
                Bfree(p5);
                return NULL;
            }
        }
        if (!(k >>= 1))
            break;
        p51 = mult(p5, p5);
        Bfree(p5);
        p5 = p51;
        if (p5 == NULL) {
            Bfree(b);
            return NULL;
        }
    }
    Bfree(p5);
    return b;
}

#endif /* Py_USING_MEMORY_DEBUGGER */

/* shift a Bigint b left by k bits.  Return a pointer to the shifted result,
   or NULL on failure.  If the returned pointer is distinct from b then the
   original b will have been Bfree'd.   Ignores the sign of b. */

static Bigint *
lshift(Bigint *b, int k)
{
    int i, k1, n, n1;
    Bigint *b1;
    ULong *x, *x1, *xe, z;

    if (!k || (!b->x[0] && b->wds == 1))
        return b;

    n = k >> 5;
    k1 = b->k;
    n1 = n + b->wds + 1;
    for(i = b->maxwds; n1 > i; i <<= 1)
        k1++;
    b1 = Balloc(k1);
    if (b1 == NULL) {
        Bfree(b);
        return NULL;
    }
    x1 = b1->x;
    for(i = 0; i < n; i++)
        *x1++ = 0;
    x = b->x;
    xe = x + b->wds;
    if (k &= 0x1f) {
        k1 = 32 - k;
        z = 0;
        do {
            *x1++ = *x << k | z;
            z = *x++ >> k1;
        }
        while(x < xe);
        if ((*x1 = z))
            ++n1;
    }
    else do
             *x1++ = *x++;
        while(x < xe);
    b1->wds = n1 - 1;
    Bfree(b);
    return b1;
}

/* Do a three-way compare of a and b, returning -1 if a < b, 0 if a == b and
   1 if a > b.  Ignores signs of a and b. */

static int
cmp(Bigint *a, Bigint *b)
{
    ULong *xa, *xa0, *xb, *xb0;
    int i, j;

    i = a->wds;
    j = b->wds;
#ifdef DEBUG
    if (i > 1 && !a->x[i-1])
        Bug("cmp called with a->x[a->wds-1] == 0");
    if (j > 1 && !b->x[j-1])
        Bug("cmp called with b->x[b->wds-1] == 0");
#endif
    if (i -= j)
        return i;
    xa0 = a->x;
    xa = xa0 + j;
    xb0 = b->x;
    xb = xb0 + j;
    for(;;) {
        if (*--xa != *--xb)
            return *xa < *xb ? -1 : 1;
        if (xa <= xa0)
            break;
    }
    return 0;
}

/* Take the difference of Bigints a and b, returning a new Bigint.  Returns
   NULL on failure.  The signs of a and b are ignored, but the sign of the
   result is set appropriately. */

static Bigint *
diff(Bigint *a, Bigint *b)
{
    Bigint *c;
    int i, wa, wb;
    ULong *xa, *xae, *xb, *xbe, *xc;
    ULLong borrow, y;

    i = cmp(a,b);
    if (!i) {
        c = Balloc(0);
        if (c == NULL)
            return NULL;
        c->wds = 1;
        c->x[0] = 0;
        return c;
    }
    if (i < 0) {
        c = a;
        a = b;
        b = c;
        i = 1;
    }
    else
        i = 0;
    c = Balloc(a->k);
    if (c == NULL)
        return NULL;
    c->sign = i;
    wa = a->wds;
    xa = a->x;
    xae = xa + wa;
    wb = b->wds;
    xb = b->x;
    xbe = xb + wb;
    xc = c->x;
    borrow = 0;
    do {
        y = (ULLong)*xa++ - *xb++ - borrow;
        borrow = y >> 32 & (ULong)1;
        *xc++ = (ULong)(y & FFFFFFFF);
    }
    while(xb < xbe);
    while(xa < xae) {
        y = *xa++ - borrow;
        borrow = y >> 32 & (ULong)1;
        *xc++ = (ULong)(y & FFFFFFFF);
    }
    while(!*--xc)
        wa--;
    c->wds = wa;
    return c;
}

/* Given a positive normal double x, return the difference between x and the
   next double up.  Doesn't give correct results for subnormals. */

static double
ulp(U *x)
{
    Long L;
    U u;

    L = (word0(x) & Exp_mask) - (P-1)*Exp_msk1;
    word0(&u) = L;
    word1(&u) = 0;
    return dval(&u);
}

/* Convert a Bigint to a double plus an exponent */

static double
b2d(Bigint *a, int *e)
{
    ULong *xa, *xa0, w, y, z;
    int k;
    U d;

    xa0 = a->x;
    xa = xa0 + a->wds;
    y = *--xa;
#ifdef DEBUG
    if (!y) Bug("zero y in b2d");
#endif
    k = hi0bits(y);
    *e = 32 - k;
    if (k < Ebits) {
        word0(&d) = Exp_1 | y >> (Ebits - k);
        w = xa > xa0 ? *--xa : 0;
        word1(&d) = y << ((32-Ebits) + k) | w >> (Ebits - k);
        goto ret_d;
    }
    z = xa > xa0 ? *--xa : 0;
    if (k -= Ebits) {
        word0(&d) = Exp_1 | y << k | z >> (32 - k);
        y = xa > xa0 ? *--xa : 0;
        word1(&d) = z << k | y >> (32 - k);
    }
    else {
        word0(&d) = Exp_1 | y;
        word1(&d) = z;
    }
  ret_d:
    return dval(&d);
}

/* Convert a scaled double to a Bigint plus an exponent.  Similar to d2b,
   except that it accepts the scale parameter used in _Py_dg_strtod (which
   should be either 0 or 2*P), and the normalization for the return value is
   different (see below).  On input, d should be finite and nonnegative, and d
   / 2**scale should be exactly representable as an IEEE 754 double.

   Returns a Bigint b and an integer e such that

     dval(d) / 2**scale = b * 2**e.

   Unlike d2b, b is not necessarily odd: b and e are normalized so
   that either 2**(P-1) <= b < 2**P and e >= Etiny, or b < 2**P
   and e == Etiny.  This applies equally to an input of 0.0: in that
   case the return values are b = 0 and e = Etiny.

   The above normalization ensures that for all possible inputs d,
   2**e gives ulp(d/2**scale).

   Returns NULL on failure.
*/

static Bigint *
sd2b(U *d, int scale, int *e)
{
    Bigint *b;

    b = Balloc(1);
    if (b == NULL)
        return NULL;

    /* First construct b and e assuming that scale == 0. */
    b->wds = 2;
    b->x[0] = word1(d);
    b->x[1] = word0(d) & Frac_mask;
    *e = Etiny - 1 + (int)((word0(d) & Exp_mask) >> Exp_shift);
    if (*e < Etiny)
        *e = Etiny;
    else
        b->x[1] |= Exp_msk1;

    /* Now adjust for scale, provided that b != 0. */
    if (scale && (b->x[0] || b->x[1])) {
        *e -= scale;
        if (*e < Etiny) {
            scale = Etiny - *e;
            *e = Etiny;
            /* We can't shift more than P-1 bits without shifting out a 1. */
            assert(0 < scale && scale <= P - 1);
            if (scale >= 32) {
                /* The bits shifted out should all be zero. */
                assert(b->x[0] == 0);
                b->x[0] = b->x[1];
                b->x[1] = 0;
                scale -= 32;
            }
            if (scale) {
                /* The bits shifted out should all be zero. */
                assert(b->x[0] << (32 - scale) == 0);
                b->x[0] = (b->x[0] >> scale) | (b->x[1] << (32 - scale));
                b->x[1] >>= scale;
            }
        }
    }
    /* Ensure b is normalized. */
    if (!b->x[1])
        b->wds = 1;

    return b;
}

/* Convert a double to a Bigint plus an exponent.  Return NULL on failure.

   Given a finite nonzero double d, return an odd Bigint b and exponent *e
   such that fabs(d) = b * 2**e.  On return, *bbits gives the number of
   significant bits of b; that is, 2**(*bbits-1) <= b < 2**(*bbits).

   If d is zero, then b == 0, *e == -1010, *bbits = 0.
 */

static Bigint *
d2b(U *d, int *e, int *bits)
{
    Bigint *b;
    int de, k;
    ULong *x, y, z;
    int i;

    b = Balloc(1);
    if (b == NULL)
        return NULL;
    x = b->x;

    z = word0(d) & Frac_mask;
    word0(d) &= 0x7fffffff;   /* clear sign bit, which we ignore */
    if ((de = (int)(word0(d) >> Exp_shift)))
        z |= Exp_msk1;
    if ((y = word1(d))) {
        if ((k = lo0bits(&y))) {
            x[0] = y | z << (32 - k);
            z >>= k;
        }
        else
            x[0] = y;
        i =
            b->wds = (x[1] = z) ? 2 : 1;
    }
    else {
        k = lo0bits(&z);
        x[0] = z;
        i =
            b->wds = 1;
        k += 32;
    }
    if (de) {
        *e = de - Bias - (P-1) + k;
        *bits = P - k;
    }
    else {
        *e = de - Bias - (P-1) + 1 + k;
        *bits = 32*i - hi0bits(x[i-1]);
    }
    return b;
}

/* Compute the ratio of two Bigints, as a double.  The result may have an
   error of up to 2.5 ulps. */

static double
ratio(Bigint *a, Bigint *b)
{
    U da, db;
    int k, ka, kb;

    dval(&da) = b2d(a, &ka);
    dval(&db) = b2d(b, &kb);
    k = ka - kb + 32*(a->wds - b->wds);
    if (k > 0)
        word0(&da) += k*Exp_msk1;
    else {
        k = -k;
        word0(&db) += k*Exp_msk1;
    }
    return dval(&da) / dval(&db);
}

static const double
tens[] = {
    1e0, 1e1, 1e2, 1e3, 1e4, 1e5, 1e6, 1e7, 1e8, 1e9,
    1e10, 1e11, 1e12, 1e13, 1e14, 1e15, 1e16, 1e17, 1e18, 1e19,
    1e20, 1e21, 1e22
};

static const double
bigtens[] = { 1e16, 1e32, 1e64, 1e128, 1e256 };
static const double tinytens[] = { 1e-16, 1e-32, 1e-64, 1e-128,
                                   9007199254740992.*9007199254740992.e-256
                                   /* = 2^106 * 1e-256 */
};
/* The factor of 2^53 in tinytens[4] helps us avoid setting the underflow */
/* flag unnecessarily.  It leads to a song and dance at the end of strtod. */
#define Scale_Bit 0x10
#define n_bigtens 5

#define ULbits 32
#define kshift 5
#define kmask 31


static int
dshift(Bigint *b, int p2)
{
    int rv = hi0bits(b->x[b->wds-1]) - 4;
    if (p2 > 0)
        rv -= p2;
    return rv & kmask;
}

/* special case of Bigint division.  The quotient is always in the range 0 <=
   quotient < 10, and on entry the divisor S is normalized so that its top 4
   bits (28--31) are zero and bit 27 is set. */

static int
quorem(Bigint *b, Bigint *S)
{
    int n;
    ULong *bx, *bxe, q, *sx, *sxe;
    ULLong borrow, carry, y, ys;

    n = S->wds;
#ifdef DEBUG
    /*debug*/ if (b->wds > n)
        /*debug*/       Bug("oversize b in quorem");
#endif
    if (b->wds < n)
        return 0;
    sx = S->x;
    sxe = sx + --n;
    bx = b->x;
    bxe = bx + n;
    q = *bxe / (*sxe + 1);      /* ensure q <= true quotient */
#ifdef DEBUG
    /*debug*/ if (q > 9)
        /*debug*/       Bug("oversized quotient in quorem");
#endif
    if (q) {
        borrow = 0;
        carry = 0;
        do {
            ys = *sx++ * (ULLong)q + carry;
            carry = ys >> 32;
            y = *bx - (ys & FFFFFFFF) - borrow;
            borrow = y >> 32 & (ULong)1;
            *bx++ = (ULong)(y & FFFFFFFF);
        }
        while(sx <= sxe);
        if (!*bxe) {
            bx = b->x;
            while(--bxe > bx && !*bxe)
                --n;
            b->wds = n;
        }
    }
    if (cmp(b, S) >= 0) {
        q++;
        borrow = 0;
        carry = 0;
        bx = b->x;
        sx = S->x;
        do {
            ys = *sx++ + carry;
            carry = ys >> 32;
            y = *bx - (ys & FFFFFFFF) - borrow;
            borrow = y >> 32 & (ULong)1;
            *bx++ = (ULong)(y & FFFFFFFF);
        }
        while(sx <= sxe);
        bx = b->x;
        bxe = bx + n;
        if (!*bxe) {
            while(--bxe > bx && !*bxe)
                --n;
            b->wds = n;
        }
    }
    return q;
}

/* sulp(x) is a version of ulp(x) that takes bc.scale into account.

   Assuming that x is finite and nonnegative (positive zero is fine
   here) and x / 2^bc.scale is exactly representable as a double,
   sulp(x) is equivalent to 2^bc.scale * ulp(x / 2^bc.scale). */

static double
sulp(U *x, BCinfo *bc)
{
    U u;

    if (bc->scale && 2*P + 1 > (int)((word0(x) & Exp_mask) >> Exp_shift)) {
        /* rv/2^bc->scale is subnormal */
        word0(&u) = (P+2)*Exp_msk1;
        word1(&u) = 0;
        return u.d;
    }
    else {
        assert(word0(x) || word1(x)); /* x != 0.0 */
        return ulp(x);
    }
}

/* The bigcomp function handles some hard cases for strtod, for inputs
   with more than STRTOD_DIGLIM digits.  It's called once an initial
   estimate for the double corresponding to the input string has
   already been obtained by the code in _Py_dg_strtod.

   The bigcomp function is only called after _Py_dg_strtod has found a
   double value rv such that either rv or rv + 1ulp represents the
   correctly rounded value corresponding to the original string.  It
   determines which of these two values is the correct one by
   computing the decimal digits of rv + 0.5ulp and comparing them with
   the corresponding digits of s0.

   In the following, write dv for the absolute value of the number represented
   by the input string.

   Inputs:

     s0 points to the first significant digit of the input string.

     rv is a (possibly scaled) estimate for the closest double value to the
        value represented by the original input to _Py_dg_strtod.  If
        bc->scale is nonzero, then rv/2^(bc->scale) is the approximation to
        the input value.

     bc is a struct containing information gathered during the parsing and
        estimation steps of _Py_dg_strtod.  Description of fields follows:

        bc->e0 gives the exponent of the input value, such that dv = (integer
           given by the bd->nd digits of s0) * 10**e0

        bc->nd gives the total number of significant digits of s0.  It will
           be at least 1.

        bc->nd0 gives the number of significant digits of s0 before the
           decimal separator.  If there's no decimal separator, bc->nd0 ==
           bc->nd.

        bc->scale is the value used to scale rv to avoid doing arithmetic with
           subnormal values.  It's either 0 or 2*P (=106).

   Outputs:

     On successful exit, rv/2^(bc->scale) is the closest double to dv.

     Returns 0 on success, -1 on failure (e.g., due to a failed malloc call). */

static int
bigcomp(U *rv, const char *s0, BCinfo *bc)
{
    Bigint *b, *d;
    int b2, d2, dd, i, nd, nd0, odd, p2, p5;

    nd = bc->nd;
    nd0 = bc->nd0;
    p5 = nd + bc->e0;
    b = sd2b(rv, bc->scale, &p2);
    if (b == NULL)
        return -1;

    /* record whether the lsb of rv/2^(bc->scale) is odd:  in the exact halfway
       case, this is used for round to even. */
    odd = b->x[0] & 1;

    /* left shift b by 1 bit and or a 1 into the least significant bit;
       this gives us b * 2**p2 = rv/2^(bc->scale) + 0.5 ulp. */
    b = lshift(b, 1);
    if (b == NULL)
        return -1;
    b->x[0] |= 1;
    p2--;

    p2 -= p5;
    d = i2b(1);
    if (d == NULL) {
        Bfree(b);
        return -1;
    }
    /* Arrange for convenient computation of quotients:
     * shift left if necessary so divisor has 4 leading 0 bits.
     */
    if (p5 > 0) {
        d = pow5mult(d, p5);
        if (d == NULL) {
            Bfree(b);
            return -1;
        }
    }
    else if (p5 < 0) {
        b = pow5mult(b, -p5);
        if (b == NULL) {
            Bfree(d);
            return -1;
        }
    }
    if (p2 > 0) {
        b2 = p2;
        d2 = 0;
    }
    else {
        b2 = 0;
        d2 = -p2;
    }
    i = dshift(d, d2);
    if ((b2 += i) > 0) {
        b = lshift(b, b2);
        if (b == NULL) {
            Bfree(d);
            return -1;
        }
    }
    if ((d2 += i) > 0) {
        d = lshift(d, d2);
        if (d == NULL) {
            Bfree(b);
            return -1;
        }
    }

    /* Compare s0 with b/d: set dd to -1, 0, or 1 according as s0 < b/d, s0 ==
     * b/d, or s0 > b/d.  Here the digits of s0 are thought of as representing
     * a number in the range [0.1, 1). */
    if (cmp(b, d) >= 0)
        /* b/d >= 1 */
        dd = -1;
    else {
        i = 0;
        for(;;) {
            b = multadd(b, 10, 0);
            if (b == NULL) {
                Bfree(d);
                return -1;
            }
            dd = s0[i < nd0 ? i : i+1] - '0' - quorem(b, d);
            i++;

            if (dd)
                break;
            if (!b->x[0] && b->wds == 1) {
                /* b/d == 0 */
                dd = i < nd;
                break;
            }
            if (!(i < nd)) {
                /* b/d != 0, but digits of s0 exhausted */
                dd = -1;
                break;
            }
        }
    }
    Bfree(b);
    Bfree(d);
    if (dd > 0 || (dd == 0 && odd))
        dval(rv) += sulp(rv, bc);
    return 0;
}

/* Return a 'standard' NaN value.

   There are exactly two quiet NaNs that don't arise by 'quieting' signaling
   NaNs (see IEEE 754-2008, section 6.2.1).  If sign == 0, return the one whose
   sign bit is cleared.  Otherwise, return the one whose sign bit is set.
*/

double
_Py_dg_stdnan(int sign)
{
    U rv;
    word0(&rv) = NAN_WORD0;
    word1(&rv) = NAN_WORD1;
    if (sign)
        word0(&rv) |= Sign_bit;
    return dval(&rv);
}

/* Return positive or negative infinity, according to the given sign (0 for
 * positive infinity, 1 for negative infinity). */

double
_Py_dg_infinity(int sign)
{
    U rv;
    word0(&rv) = POSINF_WORD0;
    word1(&rv) = POSINF_WORD1;
    return sign ? -dval(&rv) : dval(&rv);
}

double
_Py_dg_strtod(const char *s00, char **se)
{
    int bb2, bb5, bbe, bd2, bd5, bs2, c, dsign, e, e1, error;
    int esign, i, j, k, lz, nd, nd0, odd, sign;
    const char *s, *s0, *s1;
    double aadj, aadj1;
    U aadj2, adj, rv, rv0;
    ULong y, z, abs_exp;
    Long L;
    BCinfo bc;
    Bigint *bb = NULL, *bd = NULL, *bd0 = NULL, *bs = NULL, *delta = NULL;
    size_t ndigits, fraclen;
    double result;

    dval(&rv) = 0.;

    /* Start parsing. */
    c = *(s = s00);

    /* Parse optional sign, if present. */
    sign = 0;
    switch (c) {
    case '-':
        sign = 1;
        /* fall through */
    case '+':
        c = *++s;
    }

    /* Skip leading zeros: lz is true iff there were leading zeros. */
    s1 = s;
    while (c == '0')
        c = *++s;
    lz = s != s1;

    /* Point s0 at the first nonzero digit (if any).  fraclen will be the
       number of digits between the decimal point and the end of the
       digit string.  ndigits will be the total number of digits ignoring
       leading zeros. */
    s0 = s1 = s;
    while ('0' <= c && c <= '9')
        c = *++s;
    ndigits = s - s1;
    fraclen = 0;

    /* Parse decimal point and following digits. */
    if (c == '.') {
        c = *++s;
        if (!ndigits) {
            s1 = s;
            while (c == '0')
                c = *++s;
            lz = lz || s != s1;
            fraclen += (s - s1);
            s0 = s;
        }
        s1 = s;
        while ('0' <= c && c <= '9')
            c = *++s;
        ndigits += s - s1;
        fraclen += s - s1;
    }

    /* Now lz is true if and only if there were leading zero digits, and
       ndigits gives the total number of digits ignoring leading zeros.  A
       valid input must have at least one digit. */
    if (!ndigits && !lz) {
        if (se)
            *se = (char *)s00;
        goto parse_error;
    }

    /* Range check ndigits and fraclen to make sure that they, and values
       computed with them, can safely fit in an int. */
    if (ndigits > MAX_DIGITS || fraclen > MAX_DIGITS) {
        if (se)
            *se = (char *)s00;
        goto parse_error;
    }
    nd = (int)ndigits;
    nd0 = (int)ndigits - (int)fraclen;

    /* Parse exponent. */
    e = 0;
    if (c == 'e' || c == 'E') {
        s00 = s;
        c = *++s;

        /* Exponent sign. */
        esign = 0;
        switch (c) {
        case '-':
            esign = 1;
            /* fall through */
        case '+':
            c = *++s;
        }

        /* Skip zeros.  lz is true iff there are leading zeros. */
        s1 = s;
        while (c == '0')
            c = *++s;
        lz = s != s1;

        /* Get absolute value of the exponent. */
        s1 = s;
        abs_exp = 0;
        while ('0' <= c && c <= '9') {
            abs_exp = 10*abs_exp + (c - '0');
            c = *++s;
        }

        /* abs_exp will be correct modulo 2**32.  But 10**9 < 2**32, so if
           there are at most 9 significant exponent digits then overflow is
           impossible. */
        if (s - s1 > 9 || abs_exp > MAX_ABS_EXP)
            e = (int)MAX_ABS_EXP;
        else
            e = (int)abs_exp;
        if (esign)
            e = -e;

        /* A valid exponent must have at least one digit. */
        if (s == s1 && !lz)
            s = s00;
    }

    /* Adjust exponent to take into account position of the point. */
    e -= nd - nd0;
    if (nd0 <= 0)
        nd0 = nd;

    /* Finished parsing.  Set se to indicate how far we parsed */
    if (se)
        *se = (char *)s;

    /* If all digits were zero, exit with return value +-0.0.  Otherwise,
       strip trailing zeros: scan back until we hit a nonzero digit. */
    if (!nd)
        goto ret;
    for (i = nd; i > 0; ) {
        --i;
        if (s0[i < nd0 ? i : i+1] != '0') {
            ++i;
            break;
        }
    }
    e += nd - i;
    nd = i;
    if (nd0 > nd)
        nd0 = nd;

    /* Summary of parsing results.  After parsing, and dealing with zero
     * inputs, we have values s0, nd0, nd, e, sign, where:
     *
     *  - s0 points to the first significant digit of the input string
     *
     *  - nd is the total number of significant digits (here, and
     *    below, 'significant digits' means the set of digits of the
     *    significand of the input that remain after ignoring leading
     *    and trailing zeros).
     *
     *  - nd0 indicates the position of the decimal point, if present; it
     *    satisfies 1 <= nd0 <= nd.  The nd significant digits are in
     *    s0[0:nd0] and s0[nd0+1:nd+1] using the usual Python half-open slice
     *    notation.  (If nd0 < nd, then s0[nd0] contains a '.'  character; if
     *    nd0 == nd, then s0[nd0] could be any non-digit character.)
     *
     *  - e is the adjusted exponent: the absolute value of the number
     *    represented by the original input string is n * 10**e, where
     *    n is the integer represented by the concatenation of
     *    s0[0:nd0] and s0[nd0+1:nd+1]
     *
     *  - sign gives the sign of the input:  1 for negative, 0 for positive
     *
     *  - the first and last significant digits are nonzero
     */

    /* put first DBL_DIG+1 digits into integer y and z.
     *
     *  - y contains the value represented by the first min(9, nd)
     *    significant digits
     *
     *  - if nd > 9, z contains the value represented by significant digits
     *    with indices in [9, min(16, nd)).  So y * 10**(min(16, nd) - 9) + z
     *    gives the value represented by the first min(16, nd) sig. digits.
     */

    bc.e0 = e1 = e;
    y = z = 0;
    for (i = 0; i < nd; i++) {
        if (i < 9)
            y = 10*y + s0[i < nd0 ? i : i+1] - '0';
        else if (i < DBL_DIG+1)
            z = 10*z + s0[i < nd0 ? i : i+1] - '0';
        else
            break;
    }

    k = nd < DBL_DIG + 1 ? nd : DBL_DIG + 1;
    dval(&rv) = y;
    if (k > 9) {
        dval(&rv) = tens[k - 9] * dval(&rv) + z;
    }
    if (nd <= DBL_DIG
        && Flt_Rounds == 1
        ) {
        if (!e)
            goto ret;
        if (e > 0) {
            if (e <= Ten_pmax) {
                dval(&rv) *= tens[e];
                goto ret;
            }
            i = DBL_DIG - nd;
            if (e <= Ten_pmax + i) {
                /* A fancier test would sometimes let us do
                 * this for larger i values.
                 */
                e -= i;
                dval(&rv) *= tens[i];
                dval(&rv) *= tens[e];
                goto ret;
            }
        }
        else if (e >= -Ten_pmax) {
            dval(&rv) /= tens[-e];
            goto ret;
        }
    }
    e1 += nd - k;

    bc.scale = 0;

    /* Get starting approximation = rv * 10**e1 */

    if (e1 > 0) {
        if ((i = e1 & 15))
            dval(&rv) *= tens[i];
        if (e1 &= ~15) {
            if (e1 > DBL_MAX_10_EXP)
                goto ovfl;
            e1 >>= 4;
            for(j = 0; e1 > 1; j++, e1 >>= 1)
                if (e1 & 1)
                    dval(&rv) *= bigtens[j];
            /* The last multiplication could overflow. */
            word0(&rv) -= P*Exp_msk1;
            dval(&rv) *= bigtens[j];
            if ((z = word0(&rv) & Exp_mask)
                > Exp_msk1*(DBL_MAX_EXP+Bias-P))
                goto ovfl;
            if (z > Exp_msk1*(DBL_MAX_EXP+Bias-1-P)) {
                /* set to largest number */
                /* (Can't trust DBL_MAX) */
                word0(&rv) = Big0;
                word1(&rv) = Big1;
            }
            else
                word0(&rv) += P*Exp_msk1;
        }
    }
    else if (e1 < 0) {
        /* The input decimal value lies in [10**e1, 10**(e1+16)).

           If e1 <= -512, underflow immediately.
           If e1 <= -256, set bc.scale to 2*P.

           So for input value < 1e-256, bc.scale is always set;
           for input value >= 1e-240, bc.scale is never set.
           For input values in [1e-256, 1e-240), bc.scale may or may
           not be set. */

        e1 = -e1;
        if ((i = e1 & 15))
            dval(&rv) /= tens[i];
        if (e1 >>= 4) {
            if (e1 >= 1 << n_bigtens)
                goto undfl;
            if (e1 & Scale_Bit)
                bc.scale = 2*P;
            for(j = 0; e1 > 0; j++, e1 >>= 1)
                if (e1 & 1)
                    dval(&rv) *= tinytens[j];
            if (bc.scale && (j = 2*P + 1 - ((word0(&rv) & Exp_mask)
                                            >> Exp_shift)) > 0) {
                /* scaled rv is denormal; clear j low bits */
                if (j >= 32) {
                    word1(&rv) = 0;
                    if (j >= 53)
                        word0(&rv) = (P+2)*Exp_msk1;
                    else
                        word0(&rv) &= 0xffffffff << (j-32);
                }
                else
                    word1(&rv) &= 0xffffffff << j;
            }
            if (!dval(&rv))
                goto undfl;
        }
    }

    /* Now the hard part -- adjusting rv to the correct value.*/

    /* Put digits into bd: true value = bd * 10^e */

    bc.nd = nd;
    bc.nd0 = nd0;       /* Only needed if nd > STRTOD_DIGLIM, but done here */
                        /* to silence an erroneous warning about bc.nd0 */
                        /* possibly not being initialized. */
    if (nd > STRTOD_DIGLIM) {
        /* ASSERT(STRTOD_DIGLIM >= 18); 18 == one more than the */
        /* minimum number of decimal digits to distinguish double values */
        /* in IEEE arithmetic. */

        /* Truncate input to 18 significant digits, then discard any trailing
           zeros on the result by updating nd, nd0, e and y suitably. (There's
           no need to update z; it's not reused beyond this point.) */
        for (i = 18; i > 0; ) {
            /* scan back until we hit a nonzero digit.  significant digit 'i'
            is s0[i] if i < nd0, s0[i+1] if i >= nd0. */
            --i;
            if (s0[i < nd0 ? i : i+1] != '0') {
                ++i;
                break;
            }
        }
        e += nd - i;
        nd = i;
        if (nd0 > nd)
            nd0 = nd;
        if (nd < 9) { /* must recompute y */
            y = 0;
            for(i = 0; i < nd0; ++i)
                y = 10*y + s0[i] - '0';
            for(; i < nd; ++i)
                y = 10*y + s0[i+1] - '0';
        }
    }
    bd0 = s2b(s0, nd0, nd, y);
    if (bd0 == NULL)
        goto failed_malloc;

    /* Notation for the comments below.  Write:

         - dv for the absolute value of the number represented by the original
           decimal input string.

         - if we've truncated dv, write tdv for the truncated value.
           Otherwise, set tdv == dv.

         - srv for the quantity rv/2^bc.scale; so srv is the current binary
           approximation to tdv (and dv).  It should be exactly representable
           in an IEEE 754 double.
    */

    for(;;) {

        /* This is the main correction loop for _Py_dg_strtod.

           We've got a decimal value tdv, and a floating-point approximation
           srv=rv/2^bc.scale to tdv.  The aim is to determine whether srv is
           close enough (i.e., within 0.5 ulps) to tdv, and to compute a new
           approximation if not.

           To determine whether srv is close enough to tdv, compute integers
           bd, bb and bs proportional to tdv, srv and 0.5 ulp(srv)
           respectively, and then use integer arithmetic to determine whether
           |tdv - srv| is less than, equal to, or greater than 0.5 ulp(srv).
        */

        bd = Balloc(bd0->k);
        if (bd == NULL) {
            goto failed_malloc;
        }
        Bcopy(bd, bd0);
        bb = sd2b(&rv, bc.scale, &bbe);   /* srv = bb * 2^bbe */
        if (bb == NULL) {
            goto failed_malloc;
        }
        /* Record whether lsb of bb is odd, in case we need this
           for the round-to-even step later. */
        odd = bb->x[0] & 1;

        /* tdv = bd * 10**e;  srv = bb * 2**bbe */
        bs = i2b(1);
        if (bs == NULL) {
            goto failed_malloc;
        }

        if (e >= 0) {
            bb2 = bb5 = 0;
            bd2 = bd5 = e;
        }
        else {
            bb2 = bb5 = -e;
            bd2 = bd5 = 0;
        }
        if (bbe >= 0)
            bb2 += bbe;
        else
            bd2 -= bbe;
        bs2 = bb2;
        bb2++;
        bd2++;

        /* At this stage bd5 - bb5 == e == bd2 - bb2 + bbe, bb2 - bs2 == 1,
           and bs == 1, so:

              tdv == bd * 10**e = bd * 2**(bbe - bb2 + bd2) * 5**(bd5 - bb5)
              srv == bb * 2**bbe = bb * 2**(bbe - bb2 + bb2)
              0.5 ulp(srv) == 2**(bbe-1) = bs * 2**(bbe - bb2 + bs2)

           It follows that:

              M * tdv = bd * 2**bd2 * 5**bd5
              M * srv = bb * 2**bb2 * 5**bb5
              M * 0.5 ulp(srv) = bs * 2**bs2 * 5**bb5

           for some constant M.  (Actually, M == 2**(bb2 - bbe) * 5**bb5, but
           this fact is not needed below.)
        */

        /* Remove factor of 2**i, where i = min(bb2, bd2, bs2). */
        i = bb2 < bd2 ? bb2 : bd2;
        if (i > bs2)
            i = bs2;
        if (i > 0) {
            bb2 -= i;
            bd2 -= i;
            bs2 -= i;
        }

        /* Scale bb, bd, bs by the appropriate powers of 2 and 5. */
        if (bb5 > 0) {
            bs = pow5mult(bs, bb5);
            if (bs == NULL) {
                goto failed_malloc;
            }
            Bigint *bb1 = mult(bs, bb);
            Bfree(bb);
            bb = bb1;
            if (bb == NULL) {
                goto failed_malloc;
            }
        }
        if (bb2 > 0) {
            bb = lshift(bb, bb2);
            if (bb == NULL) {
                goto failed_malloc;
            }
        }
        if (bd5 > 0) {
            bd = pow5mult(bd, bd5);
            if (bd == NULL) {
                goto failed_malloc;
            }
        }
        if (bd2 > 0) {
            bd = lshift(bd, bd2);
            if (bd == NULL) {
                goto failed_malloc;
            }
        }
        if (bs2 > 0) {
            bs = lshift(bs, bs2);
            if (bs == NULL) {
                goto failed_malloc;
            }
        }

        /* Now bd, bb and bs are scaled versions of tdv, srv and 0.5 ulp(srv),
           respectively.  Compute the difference |tdv - srv|, and compare
           with 0.5 ulp(srv). */

        delta = diff(bb, bd);
        if (delta == NULL) {
            goto failed_malloc;
        }
        dsign = delta->sign;
        delta->sign = 0;
        i = cmp(delta, bs);
        if (bc.nd > nd && i <= 0) {
            if (dsign)
                break;  /* Must use bigcomp(). */

            /* Here rv overestimates the truncated decimal value by at most
               0.5 ulp(rv).  Hence rv either overestimates the true decimal
               value by <= 0.5 ulp(rv), or underestimates it by some small
               amount (< 0.1 ulp(rv)); either way, rv is within 0.5 ulps of
               the true decimal value, so it's possible to exit.

               Exception: if scaled rv is a normal exact power of 2, but not
               DBL_MIN, then rv - 0.5 ulp(rv) takes us all the way down to the
               next double, so the correctly rounded result is either rv - 0.5
               ulp(rv) or rv; in this case, use bigcomp to distinguish. */

            if (!word1(&rv) && !(word0(&rv) & Bndry_mask)) {
                /* rv can't be 0, since it's an overestimate for some
                   nonzero value.  So rv is a normal power of 2. */
                j = (int)(word0(&rv) & Exp_mask) >> Exp_shift;
                /* rv / 2^bc.scale = 2^(j - 1023 - bc.scale); use bigcomp if
                   rv / 2^bc.scale >= 2^-1021. */
                if (j - bc.scale >= 2) {
                    dval(&rv) -= 0.5 * sulp(&rv, &bc);
                    break; /* Use bigcomp. */
                }
            }

            {
                bc.nd = nd;
                i = -1; /* Discarded digits make delta smaller. */
            }
        }

        if (i < 0) {
            /* Error is less than half an ulp -- check for
             * special case of mantissa a power of two.
             */
            if (dsign || word1(&rv) || word0(&rv) & Bndry_mask
                || (word0(&rv) & Exp_mask) <= (2*P+1)*Exp_msk1
                ) {
                break;
            }
            if (!delta->x[0] && delta->wds <= 1) {
                /* exact result */
                break;
            }
            delta = lshift(delta,Log2P);
            if (delta == NULL) {
                goto failed_malloc;
            }
            if (cmp(delta, bs) > 0)
                goto drop_down;
            break;
        }
        if (i == 0) {
            /* exactly half-way between */
            if (dsign) {
                if ((word0(&rv) & Bndry_mask1) == Bndry_mask1
                    &&  word1(&rv) == (
                        (bc.scale &&
                         (y = word0(&rv) & Exp_mask) <= 2*P*Exp_msk1) ?
                        (0xffffffff & (0xffffffff << (2*P+1-(y>>Exp_shift)))) :
                        0xffffffff)) {
                    /*boundary case -- increment exponent*/
                    word0(&rv) = (word0(&rv) & Exp_mask)
                        + Exp_msk1
                        ;
                    word1(&rv) = 0;
                    /* dsign = 0; */
                    break;
                }
            }
            else if (!(word0(&rv) & Bndry_mask) && !word1(&rv)) {
              drop_down:
                /* boundary case -- decrement exponent */
                if (bc.scale) {
                    L = word0(&rv) & Exp_mask;
                    if (L <= (2*P+1)*Exp_msk1) {
                        if (L > (P+2)*Exp_msk1)
                            /* round even ==> */
                            /* accept rv */
                            break;
                        /* rv = smallest denormal */
                        if (bc.nd > nd)
                            break;
                        goto undfl;
                    }
                }
                L = (word0(&rv) & Exp_mask) - Exp_msk1;
                word0(&rv) = L | Bndry_mask1;
                word1(&rv) = 0xffffffff;
                break;
            }
            if (!odd)
                break;
            if (dsign)
                dval(&rv) += sulp(&rv, &bc);
            else {
                dval(&rv) -= sulp(&rv, &bc);
                if (!dval(&rv)) {
                    if (bc.nd >nd)
                        break;
                    goto undfl;
                }
            }
            /* dsign = 1 - dsign; */
            break;
        }
        if ((aadj = ratio(delta, bs)) <= 2.) {
            if (dsign)
                aadj = aadj1 = 1.;
            else if (word1(&rv) || word0(&rv) & Bndry_mask) {
                if (word1(&rv) == Tiny1 && !word0(&rv)) {
                    if (bc.nd >nd)
                        break;
                    goto undfl;
                }
                aadj = 1.;
                aadj1 = -1.;
            }
            else {
                /* special case -- power of FLT_RADIX to be */
                /* rounded down... */

                if (aadj < 2./FLT_RADIX)
                    aadj = 1./FLT_RADIX;
                else
                    aadj *= 0.5;
                aadj1 = -aadj;
            }
        }
        else {
            aadj *= 0.5;
            aadj1 = dsign ? aadj : -aadj;
            if (Flt_Rounds == 0)
                aadj1 += 0.5;
        }
        y = word0(&rv) & Exp_mask;

        /* Check for overflow */

        if (y == Exp_msk1*(DBL_MAX_EXP+Bias-1)) {
            dval(&rv0) = dval(&rv);
            word0(&rv) -= P*Exp_msk1;
            adj.d = aadj1 * ulp(&rv);
            dval(&rv) += adj.d;
            if ((word0(&rv) & Exp_mask) >=
                Exp_msk1*(DBL_MAX_EXP+Bias-P)) {
                if (word0(&rv0) == Big0 && word1(&rv0) == Big1) {
                    goto ovfl;
                }
                word0(&rv) = Big0;
                word1(&rv) = Big1;
                goto cont;
            }
            else
                word0(&rv) += P*Exp_msk1;
        }
        else {
            if (bc.scale && y <= 2*P*Exp_msk1) {
                if (aadj <= 0x7fffffff) {
                    if ((z = (ULong)aadj) <= 0)
                        z = 1;
                    aadj = z;
                    aadj1 = dsign ? aadj : -aadj;
                }
                dval(&aadj2) = aadj1;
                word0(&aadj2) += (2*P+1)*Exp_msk1 - y;
                aadj1 = dval(&aadj2);
            }
            adj.d = aadj1 * ulp(&rv);
            dval(&rv) += adj.d;
        }
        z = word0(&rv) & Exp_mask;
        if (bc.nd == nd) {
            if (!bc.scale)
                if (y == z) {
                    /* Can we stop now? */
                    L = (Long)aadj;
                    aadj -= L;
                    /* The tolerances below are conservative. */
                    if (dsign || word1(&rv) || word0(&rv) & Bndry_mask) {
                        if (aadj < .4999999 || aadj > .5000001)
                            break;
                    }
                    else if (aadj < .4999999/FLT_RADIX)
                        break;
                }
        }
      cont:
        Bfree(bb); bb = NULL;
        Bfree(bd); bd = NULL;
        Bfree(bs); bs = NULL;
        Bfree(delta); delta = NULL;
    }
    if (bc.nd > nd) {
        error = bigcomp(&rv, s0, &bc);
        if (error)
            goto failed_malloc;
    }

    if (bc.scale) {
        word0(&rv0) = Exp_1 - 2*P*Exp_msk1;
        word1(&rv0) = 0;
        dval(&rv) *= dval(&rv0);
    }

  ret:
    result = sign ? -dval(&rv) : dval(&rv);
    goto done;

  parse_error:
    result = 0.0;
    goto done;

  failed_malloc:
    errno = ENOMEM;
    result = -1.0;
    goto done;

  undfl:
    result = sign ? -0.0 : 0.0;
    goto done;

  ovfl:
    errno = ERANGE;
    /* Can't trust HUGE_VAL */
    word0(&rv) = Exp_mask;
    word1(&rv) = 0;
    result = sign ? -dval(&rv) : dval(&rv);
    goto done;

  done:
    Bfree(bb);
    Bfree(bd);
    Bfree(bs);
    Bfree(bd0);
    Bfree(delta);
    return result;

}

static char *
rv_alloc(int i)
{
    int j, k, *r;

    j = sizeof(ULong);
    for(k = 0;
        sizeof(Bigint) - sizeof(ULong) - sizeof(int) + j <= (unsigned)i;
        j <<= 1)
        k++;
    r = (int*)Balloc(k);
    if (r == NULL)
        return NULL;
    *r = k;
    return (char *)(r+1);
}

static char *
nrv_alloc(const char *s, char **rve, int n)
{
    char *rv, *t;

    rv = rv_alloc(n);
    if (rv == NULL)
        return NULL;
    t = rv;
    while((*t = *s++)) t++;
    if (rve)
        *rve = t;
    return rv;
}

/* freedtoa(s) must be used to free values s returned by dtoa
 * when MULTIPLE_THREADS is #defined.  It should be used in all cases,
 * but for consistency with earlier versions of dtoa, it is optional
 * when MULTIPLE_THREADS is not defined.
 */

void
_Py_dg_freedtoa(char *s)
{
    Bigint *b = (Bigint *)((int *)s - 1);
    b->maxwds = 1 << (b->k = *(int*)b);
    Bfree(b);
}

/* dtoa for IEEE arithmetic (dmg): convert double to ASCII string.
 *
 * Inspired by "How to Print Floating-Point Numbers Accurately" by
 * Guy L. Steele, Jr. and Jon L. White [Proc. ACM SIGPLAN '90, pp. 112-126].
 *
 * Modifications:
 *      1. Rather than iterating, we use a simple numeric overestimate
 *         to determine k = floor(log10(d)).  We scale relevant
 *         quantities using O(log2(k)) rather than O(k) multiplications.
 *      2. For some modes > 2 (corresponding to ecvt and fcvt), we don't
 *         try to generate digits strictly left to right.  Instead, we
 *         compute with fewer bits and propagate the carry if necessary
 *         when rounding the final digit up.  This is often faster.
 *      3. Under the assumption that input will be rounded nearest,
 *         mode 0 renders 1e23 as 1e23 rather than 9.999999999999999e22.
 *         That is, we allow equality in stopping tests when the
 *         round-nearest rule will give the same floating-point value
 *         as would satisfaction of the stopping test with strict
 *         inequality.
 *      4. We remove common factors of powers of 2 from relevant
 *         quantities.
 *      5. When converting floating-point integers less than 1e16,
 *         we use floating-point arithmetic rather than resorting
 *         to multiple-precision integers.
 *      6. When asked to produce fewer than 15 digits, we first try
 *         to get by with floating-point arithmetic; we resort to
 *         multiple-precision integer arithmetic only if we cannot
 *         guarantee that the floating-point calculation has given
 *         the correctly rounded result.  For k requested digits and
 *         "uniformly" distributed input, the probability is
 *         something like 10^(k-15) that we must resort to the Long
 *         calculation.
 */

/* Additional notes (METD): (1) returns NULL on failure.  (2) to avoid memory
   leakage, a successful call to _Py_dg_dtoa should always be matched by a
   call to _Py_dg_freedtoa. */

char *
_Py_dg_dtoa(double dd, int mode, int ndigits,
            int *decpt, int *sign, char **rve)
{
    /*  Arguments ndigits, decpt, sign are similar to those
        of ecvt and fcvt; trailing zeros are suppressed from
        the returned string.  If not null, *rve is set to point
        to the end of the return value.  If d is +-Infinity or NaN,
        then *decpt is set to 9999.

        mode:
        0 ==> shortest string that yields d when read in
        and rounded to nearest.
        1 ==> like 0, but with Steele & White stopping rule;
        e.g. with IEEE P754 arithmetic , mode 0 gives
        1e23 whereas mode 1 gives 9.999999999999999e22.
        2 ==> max(1,ndigits) significant digits.  This gives a
        return value similar to that of ecvt, except
        that trailing zeros are suppressed.
        3 ==> through ndigits past the decimal point.  This
        gives a return value similar to that from fcvt,
        except that trailing zeros are suppressed, and
        ndigits can be negative.
        4,5 ==> similar to 2 and 3, respectively, but (in
        round-nearest mode) with the tests of mode 0 to
        possibly return a shorter string that rounds to d.
        With IEEE arithmetic and compilation with
        -DHonor_FLT_ROUNDS, modes 4 and 5 behave the same
        as modes 2 and 3 when FLT_ROUNDS != 1.
        6-9 ==> Debugging modes similar to mode - 4:  don't try
        fast floating-point estimate (if applicable).

        Values of mode other than 0-9 are treated as mode 0.

        Sufficient space is allocated to the return value
        to hold the suppressed trailing zeros.
    */

    int bbits, b2, b5, be, dig, i, ieps, ilim, ilim0, ilim1,
        j, j1, k, k0, k_check, leftright, m2, m5, s2, s5,
        spec_case, try_quick;
    Long L;
    int denorm;
    ULong x;
    Bigint *b, *b1, *delta, *mlo, *mhi, *S;
    U d2, eps, u;
    double ds;
    char *s, *s0;

    /* set pointers to NULL, to silence gcc compiler warnings and make
       cleanup easier on error */
    mlo = mhi = S = 0;
    s0 = 0;

    u.d = dd;
    if (word0(&u) & Sign_bit) {
        /* set sign for everything, including 0's and NaNs */
        *sign = 1;
        word0(&u) &= ~Sign_bit; /* clear sign bit */
    }
    else
        *sign = 0;

    /* quick return for Infinities, NaNs and zeros */
    if ((word0(&u) & Exp_mask) == Exp_mask)
    {
        /* Infinity or NaN */
        *decpt = 9999;
        if (!word1(&u) && !(word0(&u) & 0xfffff))
            return nrv_alloc("Infinity", rve, 8);
        return nrv_alloc("NaN", rve, 3);
    }
    if (!dval(&u)) {
        *decpt = 1;
        return nrv_alloc("0", rve, 1);
    }

    /* compute k = floor(log10(d)).  The computation may leave k
       one too large, but should never leave k too small. */
    b = d2b(&u, &be, &bbits);
    if (b == NULL)
        goto failed_malloc;
    if ((i = (int)(word0(&u) >> Exp_shift1 & (Exp_mask>>Exp_shift1)))) {
        dval(&d2) = dval(&u);
        word0(&d2) &= Frac_mask1;
        word0(&d2) |= Exp_11;

        /* log(x)       ~=~ log(1.5) + (x-1.5)/1.5
         * log10(x)      =  log(x) / log(10)
         *              ~=~ log(1.5)/log(10) + (x-1.5)/(1.5*log(10))
         * log10(d) = (i-Bias)*log(2)/log(10) + log10(d2)
         *
         * This suggests computing an approximation k to log10(d) by
         *
         * k = (i - Bias)*0.301029995663981
         *      + ( (d2-1.5)*0.289529654602168 + 0.176091259055681 );
         *
         * We want k to be too large rather than too small.
         * The error in the first-order Taylor series approximation
         * is in our favor, so we just round up the constant enough
         * to compensate for any error in the multiplication of
         * (i - Bias) by 0.301029995663981; since |i - Bias| <= 1077,
         * and 1077 * 0.30103 * 2^-52 ~=~ 7.2e-14,
         * adding 1e-13 to the constant term more than suffices.
         * Hence we adjust the constant term to 0.1760912590558.
         * (We could get a more accurate k by invoking log10,
         *  but this is probably not worthwhile.)
         */

        i -= Bias;
        denorm = 0;
    }
    else {
        /* d is denormalized */

        i = bbits + be + (Bias + (P-1) - 1);
        x = i > 32  ? word0(&u) << (64 - i) | word1(&u) >> (i - 32)
            : word1(&u) << (32 - i);
        dval(&d2) = x;
        word0(&d2) -= 31*Exp_msk1; /* adjust exponent */
        i -= (Bias + (P-1) - 1) + 1;
        denorm = 1;
    }
    ds = (dval(&d2)-1.5)*0.289529654602168 + 0.1760912590558 +
        i*0.301029995663981;
    k = (int)ds;
    if (ds < 0. && ds != k)
        k--;    /* want k = floor(ds) */
    k_check = 1;
    if (k >= 0 && k <= Ten_pmax) {
        if (dval(&u) < tens[k])
            k--;
        k_check = 0;
    }
    j = bbits - i - 1;
    if (j >= 0) {
        b2 = 0;
        s2 = j;
    }
    else {
        b2 = -j;
        s2 = 0;
    }
    if (k >= 0) {
        b5 = 0;
        s5 = k;
        s2 += k;
    }
    else {
        b2 -= k;
        b5 = -k;
        s5 = 0;
    }
    if (mode < 0 || mode > 9)
        mode = 0;

    try_quick = 1;

    if (mode > 5) {
        mode -= 4;
        try_quick = 0;
    }
    leftright = 1;
    ilim = ilim1 = -1;  /* Values for cases 0 and 1; done here to */
    /* silence erroneous "gcc -Wall" warning. */
    switch(mode) {
    case 0:
    case 1:
        i = 18;
        ndigits = 0;
        break;
    case 2:
        leftright = 0;
        /* fall through */
    case 4:
        if (ndigits <= 0)
            ndigits = 1;
        ilim = ilim1 = i = ndigits;
        break;
    case 3:
        leftright = 0;
        /* fall through */
    case 5:
        i = ndigits + k + 1;
        ilim = i;
        ilim1 = i - 1;
        if (i <= 0)
            i = 1;
    }
    s0 = rv_alloc(i);
    if (s0 == NULL)
        goto failed_malloc;
    s = s0;


    if (ilim >= 0 && ilim <= Quick_max && try_quick) {

        /* Try to get by with floating-point arithmetic. */

        i = 0;
        dval(&d2) = dval(&u);
        k0 = k;
        ilim0 = ilim;
        ieps = 2; /* conservative */
        if (k > 0) {
            ds = tens[k&0xf];
            j = k >> 4;
            if (j & Bletch) {
                /* prevent overflows */
                j &= Bletch - 1;
                dval(&u) /= bigtens[n_bigtens-1];
                ieps++;
            }
            for(; j; j >>= 1, i++)
                if (j & 1) {
                    ieps++;
                    ds *= bigtens[i];
                }
            dval(&u) /= ds;
        }
        else if ((j1 = -k)) {
            dval(&u) *= tens[j1 & 0xf];
            for(j = j1 >> 4; j; j >>= 1, i++)
                if (j & 1) {
                    ieps++;
                    dval(&u) *= bigtens[i];
                }
        }
        if (k_check && dval(&u) < 1. && ilim > 0) {
            if (ilim1 <= 0)
                goto fast_failed;
            ilim = ilim1;
            k--;
            dval(&u) *= 10.;
            ieps++;
        }
        dval(&eps) = ieps*dval(&u) + 7.;
        word0(&eps) -= (P-1)*Exp_msk1;
        if (ilim == 0) {
            S = mhi = 0;
            dval(&u) -= 5.;
            if (dval(&u) > dval(&eps))
                goto one_digit;
            if (dval(&u) < -dval(&eps))
                goto no_digits;
            goto fast_failed;
        }
        if (leftright) {
            /* Use Steele & White method of only
             * generating digits needed.
             */
            dval(&eps) = 0.5/tens[ilim-1] - dval(&eps);
            for(i = 0;;) {
                L = (Long)dval(&u);
                dval(&u) -= L;
                *s++ = '0' + (int)L;
                if (dval(&u) < dval(&eps))
                    goto ret1;
                if (1. - dval(&u) < dval(&eps))
                    goto bump_up;
                if (++i >= ilim)
                    break;
                dval(&eps) *= 10.;
                dval(&u) *= 10.;
            }
        }
        else {
            /* Generate ilim digits, then fix them up. */
            dval(&eps) *= tens[ilim-1];
            for(i = 1;; i++, dval(&u) *= 10.) {
                L = (Long)(dval(&u));
                if (!(dval(&u) -= L))
                    ilim = i;
                *s++ = '0' + (int)L;
                if (i == ilim) {
                    if (dval(&u) > 0.5 + dval(&eps))
                        goto bump_up;
                    else if (dval(&u) < 0.5 - dval(&eps)) {
                        while(*--s == '0');
                        s++;
                        goto ret1;
                    }
                    break;
                }
            }
        }
      fast_failed:
        s = s0;
        dval(&u) = dval(&d2);
        k = k0;
        ilim = ilim0;
    }

    /* Do we have a "small" integer? */

    if (be >= 0 && k <= Int_max) {
        /* Yes. */
        ds = tens[k];
        if (ndigits < 0 && ilim <= 0) {
            S = mhi = 0;
            if (ilim < 0 || dval(&u) <= 5*ds)
                goto no_digits;
            goto one_digit;
        }
        for(i = 1;; i++, dval(&u) *= 10.) {
            L = (Long)(dval(&u) / ds);
            dval(&u) -= L*ds;
            *s++ = '0' + (int)L;
            if (!dval(&u)) {
                break;
            }
            if (i == ilim) {
                dval(&u) += dval(&u);
                if (dval(&u) > ds || (dval(&u) == ds && L & 1)) {
                  bump_up:
                    while(*--s == '9')
                        if (s == s0) {
                            k++;
                            *s = '0';
                            break;
                        }
                    ++*s++;
                }
                else {
                    /* Strip trailing zeros. This branch was missing from the
                       original dtoa.c, leading to surplus trailing zeros in
                       some cases. See bugs.python.org/issue40780. */
                    while (s > s0 && s[-1] == '0') {
                        --s;
                    }
                }
                break;
            }
        }
        goto ret1;
    }

    m2 = b2;
    m5 = b5;
    if (leftright) {
        i =
            denorm ? be + (Bias + (P-1) - 1 + 1) :
            1 + P - bbits;
        b2 += i;
        s2 += i;
        mhi = i2b(1);
        if (mhi == NULL)
            goto failed_malloc;
    }
    if (m2 > 0 && s2 > 0) {
        i = m2 < s2 ? m2 : s2;
        b2 -= i;
        m2 -= i;
        s2 -= i;
    }
    if (b5 > 0) {
        if (leftright) {
            if (m5 > 0) {
                mhi = pow5mult(mhi, m5);
                if (mhi == NULL)
                    goto failed_malloc;
                b1 = mult(mhi, b);
                Bfree(b);
                b = b1;
                if (b == NULL)
                    goto failed_malloc;
            }
            if ((j = b5 - m5)) {
                b = pow5mult(b, j);
                if (b == NULL)
                    goto failed_malloc;
            }
        }
        else {
            b = pow5mult(b, b5);
            if (b == NULL)
                goto failed_malloc;
        }
    }
    S = i2b(1);
    if (S == NULL)
        goto failed_malloc;
    if (s5 > 0) {
        S = pow5mult(S, s5);
        if (S == NULL)
            goto failed_malloc;
    }

    /* Check for special case that d is a normalized power of 2. */

    spec_case = 0;
    if ((mode < 2 || leftright)
        ) {
        if (!word1(&u) && !(word0(&u) & Bndry_mask)
            && word0(&u) & (Exp_mask & ~Exp_msk1)
            ) {
            /* The special case */
            b2 += Log2P;
            s2 += Log2P;
            spec_case = 1;
        }
    }

    /* Arrange for convenient computation of quotients:
     * shift left if necessary so divisor has 4 leading 0 bits.
     *
     * Perhaps we should just compute leading 28 bits of S once
     * and for all and pass them and a shift to quorem, so it
     * can do shifts and ors to compute the numerator for q.
     */
#define iInc 28
    i = dshift(S, s2);
    b2 += i;
    m2 += i;
    s2 += i;
    if (b2 > 0) {
        b = lshift(b, b2);
        if (b == NULL)
            goto failed_malloc;
    }
    if (s2 > 0) {
        S = lshift(S, s2);
        if (S == NULL)
            goto failed_malloc;
    }
    if (k_check) {
        if (cmp(b,S) < 0) {
            k--;
            b = multadd(b, 10, 0);      /* we botched the k estimate */
            if (b == NULL)
                goto failed_malloc;
            if (leftright) {
                mhi = multadd(mhi, 10, 0);
                if (mhi == NULL)
                    goto failed_malloc;
            }
            ilim = ilim1;
        }
    }
    if (ilim <= 0 && (mode == 3 || mode == 5)) {
        if (ilim < 0) {
            /* no digits, fcvt style */
          no_digits:
            k = -1 - ndigits;
            goto ret;
        }
        else {
            S = multadd(S, 5, 0);
            if (S == NULL)
                goto failed_malloc;
            if (cmp(b, S) <= 0)
                goto no_digits;
        }
      one_digit:
        *s++ = '1';
        k++;
        goto ret;
    }
    if (leftright) {
        if (m2 > 0) {
            mhi = lshift(mhi, m2);
            if (mhi == NULL)
                goto failed_malloc;
        }

        /* Compute mlo -- check for special case
         * that d is a normalized power of 2.
         */

        mlo = mhi;
        if (spec_case) {
            mhi = Balloc(mhi->k);
            if (mhi == NULL)
                goto failed_malloc;
            Bcopy(mhi, mlo);
            mhi = lshift(mhi, Log2P);
            if (mhi == NULL)
                goto failed_malloc;
        }

        for(i = 1;;i++) {
            dig = quorem(b,S) + '0';
            /* Do we yet have the shortest decimal string
             * that will round to d?
             */
            j = cmp(b, mlo);
            delta = diff(S, mhi);
            if (delta == NULL)
                goto failed_malloc;
            j1 = delta->sign ? 1 : cmp(b, delta);
            Bfree(delta);
            if (j1 == 0 && mode != 1 && !(word1(&u) & 1)
                ) {
                if (dig == '9')
                    goto round_9_up;
                if (j > 0)
                    dig++;
                *s++ = dig;
                goto ret;
            }
            if (j < 0 || (j == 0 && mode != 1
                          && !(word1(&u) & 1)
                    )) {
                if (!b->x[0] && b->wds <= 1) {
                    goto accept_dig;
                }
                if (j1 > 0) {
                    b = lshift(b, 1);
                    if (b == NULL)
                        goto failed_malloc;
                    j1 = cmp(b, S);
                    if ((j1 > 0 || (j1 == 0 && dig & 1))
                        && dig++ == '9')
                        goto round_9_up;
                }
              accept_dig:
                *s++ = dig;
                goto ret;
            }
            if (j1 > 0) {
                if (dig == '9') { /* possible if i == 1 */
                  round_9_up:
                    *s++ = '9';
                    goto roundoff;
                }
                *s++ = dig + 1;
                goto ret;
            }
            *s++ = dig;
            if (i == ilim)
                break;
            b = multadd(b, 10, 0);
            if (b == NULL)
                goto failed_malloc;
            if (mlo == mhi) {
                mlo = mhi = multadd(mhi, 10, 0);
                if (mlo == NULL)
                    goto failed_malloc;
            }
            else {
                mlo = multadd(mlo, 10, 0);
                if (mlo == NULL)
                    goto failed_malloc;
                mhi = multadd(mhi, 10, 0);
                if (mhi == NULL)
                    goto failed_malloc;
            }
        }
    }
    else
        for(i = 1;; i++) {
            *s++ = dig = quorem(b,S) + '0';
            if (!b->x[0] && b->wds <= 1) {
                goto ret;
            }
            if (i >= ilim)
                break;
            b = multadd(b, 10, 0);
            if (b == NULL)
                goto failed_malloc;
        }

    /* Round off last digit */

    b = lshift(b, 1);
    if (b == NULL)
        goto failed_malloc;
    j = cmp(b, S);
    if (j > 0 || (j == 0 && dig & 1)) {
      roundoff:
        while(*--s == '9')
            if (s == s0) {
                k++;
                *s++ = '1';
                goto ret;
            }
        ++*s++;
    }
    else {
        while(*--s == '0');
        s++;
    }
  ret:
    Bfree(S);
    if (mhi) {
        if (mlo && mlo != mhi)
            Bfree(mlo);
        Bfree(mhi);
    }
  ret1:
    Bfree(b);
    *s = 0;
    *decpt = k + 1;
    if (rve)
        *rve = s;
    return s0;
  failed_malloc:
    if (S)
        Bfree(S);
    if (mlo && mlo != mhi)
        Bfree(mlo);
    if (mhi)
        Bfree(mhi);
    if (b)
        Bfree(b);
    if (s0)
        _Py_dg_freedtoa(s0);
    return NULL;
}
#ifdef __cplusplus
}
#endif

#endif  // _PY_SHORT_FLOAT_REPR == 1