2
* Copyright (c) 2003, 2006 Matteo Frigo
3
* Copyright (c) 2003, 2006 Massachusetts Institute of Technology
5
* This program is free software; you can redistribute it and/or modify
6
* it under the terms of the GNU General Public License as published by
7
* the Free Software Foundation; either version 2 of the License, or
8
* (at your option) any later version.
10
* This program is distributed in the hope that it will be useful,
11
* but WITHOUT ANY WARRANTY; without even the implied warranty of
12
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13
* GNU General Public License for more details.
15
* You should have received a copy of the GNU General Public License
16
* along with this program; if not, write to the Free Software
17
* Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
21
/* This file was automatically generated --- DO NOT EDIT */
22
/* Generated on Sat Jul 1 22:45:32 EDT 2006 */
28
/* cheap-mode: VECTGRADE_FULL succeeded. (8036 steps) */
29
/* Generated by: ../../../genfft-k7/gen_notw -no-randomized-cse -n 128 -sign 1 -name n1k7i_128 */
33
* $Id: algsimp.ml,v 1.4 2006-01-05 03:04:27 stevenj Exp $
34
* $Id: fft.ml,v 1.4 2006-01-05 03:04:27 stevenj Exp $
35
* $Id: gen_notw.ml,v 1.12 2006-01-05 03:04:27 stevenj Exp $
38
/* The following asm code is Copyright (c) 2000-2001 Stefan Kral */
41
chs_lo: .long 0x80000000, 0x00000000
42
chs_hi: .long 0x00000000, 0x80000000
43
KP923879532KP382683432: .float +0.923879532511286756128183189396788286822416626, +0.382683432365089771728459984030398866761344562
44
KP382683432KP923879532: .float +0.382683432365089771728459984030398866761344562, +0.923879532511286756128183189396788286822416626
45
KP707106781KP707106781: .float +0.707106781186547524400844362104849039284835938, +0.707106781186547524400844362104849039284835938
46
KP555570233KP831469612: .float +0.555570233019602224742830813948532874374937191, +0.831469612302545237078788377617905756738560812
47
KP831469612KP555570233: .float +0.831469612302545237078788377617905756738560812, +0.555570233019602224742830813948532874374937191
48
KP980785280KP195090322: .float +0.980785280403230449126182236134239036973933731, +0.195090322016128267848284868477022240927691618
49
KP195090322KP980785280: .float +0.195090322016128267848284868477022240927691618, +0.980785280403230449126182236134239036973933731
50
KP773010453KP634393284: .float +0.773010453362736960810906609758469800971041293, +0.634393284163645498215171613225493370675687095
51
KP634393284KP773010453: .float +0.634393284163645498215171613225493370675687095, +0.773010453362736960810906609758469800971041293
52
KP995184726KP098017140: .float +0.995184726672196886244836953109479921575474869, +0.098017140329560601994195563888641845861136673
53
KP098017140KP995184726: .float +0.098017140329560601994195563888641845861136673, +0.995184726672196886244836953109479921575474869
54
KP956940335KP290284677: .float +0.956940335732208864935797886980269969482849206, +0.290284677254462367636192375817395274691476278
55
KP290284677KP956940335: .float +0.290284677254462367636192375817395274691476278, +0.956940335732208864935797886980269969482849206
56
KP881921264KP471396736: .float +0.881921264348355029712756863660388349508442621, +0.471396736825997648556387625905254377657460319
57
KP471396736KP881921264: .float +0.471396736825997648556387625905254377657460319, +0.881921264348355029712756863660388349508442621
58
KP146730474KP989176509: .float +0.146730474455361751658850129646717819706215317, +0.989176509964780973451673738016243063983689533
59
KP514102744KP857728610: .float +0.514102744193221726593693838968815772608049120, +0.857728610000272069902269984284770137042490799
60
KP857728610KP514102744: .float +0.857728610000272069902269984284770137042490799, +0.514102744193221726593693838968815772608049120
61
KP803207531KP595699304: .float +0.803207531480644909806676512963141923879569427, +0.595699304492433343467036528829969889511926338
62
KP242980179KP970031253: .float +0.242980179903263889948274162077471118320990783, +0.970031253194543992603984207286100251456865962
63
KP970031253KP242980179: .float +0.970031253194543992603984207286100251456865962, +0.242980179903263889948274162077471118320990783
64
KP595699304KP803207531: .float +0.595699304492433343467036528829969889511926338, +0.803207531480644909806676512963141923879569427
65
KP989176509KP146730474: .float +0.989176509964780973451673738016243063983689533, +0.146730474455361751658850129646717819706215317
66
KP427555093KP903989293: .float +0.427555093430282094320966856888798534304578629, +0.903989293123443331586200297230537048710132025
67
KP049067674KP998795456: .float +0.049067674327418014254954976942682658314745363, +0.998795456205172392714771604759100694443203615
68
KP998795456KP049067674: .float +0.998795456205172392714771604759100694443203615, +0.049067674327418014254954976942682658314745363
69
KP941544065KP336889853: .float +0.941544065183020778412509402599502357185589796, +0.336889853392220050689253212619147570477766780
70
KP671558954KP740951125: .float +0.671558954847018400625376850427421803228750632, +0.740951125354959091175616897495162729728955309
71
KP740951125KP671558954: .float +0.740951125354959091175616897495162729728955309, +0.671558954847018400625376850427421803228750632
72
KP336889853KP941544065: .float +0.336889853392220050689253212619147570477766780, +0.941544065183020778412509402599502357185589796
73
KP903989293KP427555093: .float +0.903989293123443331586200297230537048710132025, +0.427555093430282094320966856888798534304578629
100
/* promise simd cell size = 8 */
104
movq (%edx,%ebp,8), %mm3
105
movq (%eax,%ebp,8), %mm4
106
leal (%ecx,%ecx), %ebp
108
leal (%ebp,%ebp,2), %ebp
112
movq (%eax,%ebp,8), %mm7
114
movq (%edx,%ebp,8), %mm0
115
leal (%ecx,%ecx), %ebp
120
movq (%edx,%ebp,8), %mm2
121
movq (%eax,%ebp,8), %mm4
124
leal (%ecx,%ecx,2), %ebp
132
movq (%edx,%ebp,8), %mm2
139
movq (%eax,%ebp,8), %mm0
140
leal (%ecx,%ecx,2), %ebp
142
leal (%ecx,%ebp,2), %ebp
145
movq (%eax,%ebp,8), %mm7
148
movq (%edx,%ebp,8), %mm3
149
leal (%ecx,%ecx,4), %ebp
158
movq (%edx,%ebp,8), %mm0
162
movq (%eax,%ebp,8), %mm2
164
leal (%ecx,%ecx,8), %ebp
166
movq (%edx,%ecx,8), %mm3
171
movq (%eax,%ecx,8), %mm0
214
pfmul KP923879532KP382683432, %mm1
215
pfmul KP382683432KP923879532, %mm7
223
pfmul KP382683432KP923879532, %mm5
224
pfmul KP923879532KP382683432, %mm6
231
pfmul KP923879532KP382683432, %mm4
235
pfmul KP382683432KP923879532, %mm7
244
pfmul KP707106781KP707106781, %mm5
246
pfmul KP707106781KP707106781, %mm2
247
pfmul KP382683432KP923879532, %mm0
248
pfmul KP707106781KP707106781, %mm1
249
pfmul KP707106781KP707106781, %mm6
250
pfmul KP923879532KP382683432, %mm3
253
movq (%edx,%ebp,4), %mm5
254
movq (%eax,%ebp,4), %mm2
255
leal (%ecx,%ecx,2), %ebp
258
leal (%ecx,%ebp,4), %ebp
260
movq (%edx,%ecx,4), %mm1
261
movq (%eax,%ecx,4), %mm6
266
movq (%eax,%ebp,4), %mm0
268
movq (%edx,%ebp,4), %mm5
269
leal (%ecx,%ecx,4), %ebp
282
movq (%edx,%ebp,4), %mm0
283
movq (%eax,%ebp,4), %mm3
285
leal (%ecx,%ecx,2), %ebp
287
leal (%ebp,%ecx,8), %ebp
297
movq (%edx,%ebp,4), %mm3
299
movq (%eax,%ebp,4), %mm5
300
leal (%ecx,%ecx,2), %ebp
304
movq (%edx,%ebp,4), %mm0
305
movq (%eax,%ebp,4), %mm1
308
leal (%ecx,%ecx,2), %ebp
310
leal (%ecx,%ebp,2), %ebp
322
movq (%edx,%ebp,4), %mm0
323
movq (%eax,%ebp,4), %mm3
324
leal (%ecx,%ecx,2), %ebp
326
leal (%ebp,%ebp,4), %ebp
332
movq (%eax,%ebp,4), %mm0
333
movq (%edx,%ebp,4), %mm1
334
leal (%ecx,%ecx), %ebp
336
leal (%ecx,%ebp,8), %ebp
375
pfmul KP707106781KP707106781, %mm2
376
pfmul KP707106781KP707106781, %mm5
398
pfmul KP707106781KP707106781, %mm5
399
pfmul KP707106781KP707106781, %mm6
415
pfmul KP382683432KP923879532, %mm1
418
pfmul KP923879532KP382683432, %mm6
419
pfmul KP923879532KP382683432, %mm0
421
pfmul KP382683432KP923879532, %mm2
425
pfmul KP923879532KP382683432, %mm5
427
pfmul KP382683432KP923879532, %mm3
428
pfmul KP382683432KP923879532, %mm7
429
pfmul KP923879532KP382683432, %mm4
431
movq (%edx,%ebp,2), %mm1
433
movq (%eax,%ebp,2), %mm6
434
leal (%ecx,%ecx,2), %ebp
437
movq (%edx,%ecx,2), %mm4
438
leal (%ecx,%ebp,8), %ebp
441
movq (%eax,%ecx,2), %mm5
445
movq (%eax,%ebp,2), %mm0
447
movq (%edx,%ebp,2), %mm1
448
leal (%ecx,%ecx,8), %ebp
453
movq (%edx,%ebp,2), %mm6
458
movq (%eax,%ebp,2), %mm5
459
leal (%ecx,%ecx,2), %ebp
463
leal (%ebp,%ecx,8), %ebp
470
movq (%edx,%ebp,2), %mm3
471
movq (%eax,%ebp,2), %mm4
472
leal (%ecx,%ecx,2), %ebp
482
movq (%edx,%ebp,2), %mm5
484
movq (%eax,%ebp,2), %mm0
485
leal (%ecx,%ecx,2), %ebp
488
leal (%ebp,%ebp,8), %ebp
491
movq (%eax,%ebp,2), %mm7
492
movq (%edx,%ebp,2), %mm6
493
leal (%ecx,%ecx,8), %ebp
497
leal (%ecx,%ebp,2), %ebp
500
movq (%edx,%ebp,2), %mm5
501
movq (%eax,%ebp,2), %mm4
504
leal (%ecx,%ecx,2), %ebp
506
leal (%ebp,%ebp,4), %ebp
528
movq (%edx,%ebp,2), %mm0
532
movq (%eax,%ebp,2), %mm4
533
leal (%ecx,%ecx,2), %ebp
536
leal (%ebp,%ebp,4), %ebp
540
leal (%ecx,%ebp,2), %ebp
544
movq (%eax,%ebp,2), %mm1
545
movq (%edx,%ebp,2), %mm5
546
leal (%ecx,%ecx,2), %ebp
549
leal (%ebp,%ebp,4), %ebp
551
leal (%ebp,%ecx,8), %ebp
555
movq (%eax,%ebp,2), %mm0
556
movq (%edx,%ebp,2), %mm7
557
leal (%ecx,%ecx,2), %ebp
560
leal (%ecx,%ebp,2), %ebp
565
movq (%edx,%ebp,2), %mm3
566
movq (%eax,%ebp,2), %mm6
567
leal (%ecx,%ecx,2), %ebp
570
leal (%ecx,%ebp,4), %ebp
575
movq (%eax,%ebp,2), %mm2
576
movq (%edx,%ebp,2), %mm5
577
leal (%ecx,%ecx,4), %ebp
586
movq (%eax,%ebp,2), %mm5
590
movq (%edx,%ebp,2), %mm6
591
leal (%ecx,%ecx,2), %ebp
594
leal (%ebp,%ebp,8), %ebp
598
leal (%ebp,%ecx,2), %ebp
602
movq (%eax,%ebp,2), %mm7
603
movq (%edx,%ebp,2), %mm1
604
leal (%ecx,%ecx,4), %ebp
606
leal (%ecx,%ebp,4), %ebp
608
movq (%edx,%ebp,2), %mm6
609
movq (%eax,%ebp,2), %mm5
612
leal (%ecx,%ecx,4), %ebp
614
leal (%ebp,%ebp,8), %ebp
617
leal (%ebp,%ecx,2), %ebp
667
pfmul KP707106781KP707106781, %mm7
668
pfmul KP707106781KP707106781, %mm1
686
pfmul KP707106781KP707106781, %mm1
688
pfmul KP707106781KP707106781, %mm4
712
pfmul KP707106781KP707106781, %mm0
714
pfmul KP707106781KP707106781, %mm5
719
pfmul KP707106781KP707106781, %mm7
724
pfmul KP707106781KP707106781, %mm6
773
pfmul KP382683432KP923879532, %mm4
778
pfmul KP923879532KP382683432, %mm5
782
pfmul KP382683432KP923879532, %mm1
783
pfmul KP382683432KP923879532, %mm2
788
pfmul KP382683432KP923879532, %mm6
791
pfmul KP923879532KP382683432, %mm0
792
pfmul KP382683432KP923879532, %mm7
795
pfmul KP382683432KP923879532, %mm3
802
pfmul KP382683432KP923879532, %mm1
805
pfmul KP923879532KP382683432, %mm7
807
pfmul KP923879532KP382683432, %mm5
811
pfmul KP923879532KP382683432, %mm4
819
pfmul KP923879532KP382683432, %mm0
821
pfmul KP382683432KP923879532, %mm2
823
pfmul KP923879532KP382683432, %mm6
825
pfmul KP923879532KP382683432, %mm3
828
movq (%eax,%ebp), %mm1
830
movq (%edx,%ebp), %mm6
831
leal (%ecx,%ecx,2), %ebp
834
leal (%ebp,%ebp,4), %ebp
836
movq (%edx,%ebp), %mm5
837
movq (%eax,%ebp), %mm7
838
leal (%ecx,%ecx,2), %ebp
843
leal (%ebp,%ebp,4), %ebp
846
leal (%ecx,%ebp,2), %ebp
851
movq (%edx,%ebp), %mm0
852
movq (%eax,%ebp), %mm5
853
leal (%ecx,%ecx,2), %ebp
857
leal (%ecx,%ebp,2), %ebp
861
leal (%ebp,%ebp,8), %ebp
867
movq (%eax,%ebp), %mm7
868
movq (%edx,%ebp), %mm0
869
leal (%ecx,%ecx,2), %ebp
872
leal (%ebp,%ebp,4), %ebp
877
leal (%ebp,%ecx,8), %ebp
881
movq (%edx,%ebp), %mm2
882
movq (%eax,%ebp), %mm1
883
leal (%ecx,%ecx,2), %ebp
887
leal (%ecx,%ebp,2), %ebp
890
movq (%edx,%ebp), %mm3
891
movq (%eax,%ebp), %mm4
892
leal (%ecx,%ecx,2), %ebp
896
leal (%ebp,%ebp,8), %ebp
899
leal (%ecx,%ebp,2), %ebp
903
movq (%eax,%ebp), %mm0
904
movq (%edx,%ebp), %mm2
905
leal (%ecx,%ecx,2), %ebp
907
leal (%ecx,%ebp,4), %ebp
909
leal (%ebp,%ebp,2), %ebp
912
movq (%edx,%ebp), %mm3
913
movq (%eax,%ebp), %mm1
915
leal (%ecx,%ecx), %ebp
916
leal (%ecx,%ebp,8), %ebp
920
leal (%ecx,%ebp,2), %ebp
933
movq %mm7, 1024(%esp)
941
movq %mm1, 1048(%esp)
943
movq (%edx,%ebp), %mm1
944
movq (%eax,%ebp), %mm7
945
leal (%ecx,%ecx,2), %ebp
947
movq %mm2, 1096(%esp)
949
movq (%edx,%ebp), %mm2
950
movq (%eax,%ebp), %mm5
951
leal (%ecx,%ecx), %ebp
952
movq %mm6, 1016(%esp)
953
leal (%ecx,%ebp,8), %ebp
954
movq %mm4, 1000(%esp)
956
movq %mm3, 1008(%esp)
958
leal (%ebp,%ebp,2), %ebp
962
movq (%eax,%ebp), %mm3
963
movq (%edx,%ebp), %mm6
964
leal (%ecx,%ecx,8), %ebp
966
movq %mm7, 1040(%esp)
967
leal (%ecx,%ebp,2), %ebp
968
movq %mm5, 1032(%esp)
969
movq (%eax,%ebp), %mm5
971
movq (%edx,%ebp), %mm2
974
leal (%ecx,%ecx,2), %ebp
977
leal (%ebp,%ecx,8), %ebp
985
movq %mm1, 1064(%esp)
990
movq (%edx,%ebp), %mm7
993
movq %mm3, 1072(%esp)
994
movq (%eax,%ebp), %mm3
995
imull $59, %ecx, %ebp
998
pfmul KP707106781KP707106781, %mm5
1002
pfmul KP707106781KP707106781, %mm0
1003
movq %mm4, 1104(%esp)
1004
movq (%eax,%ebp), %mm2
1006
movq (%edx,%ebp), %mm4
1007
leal (%ecx,%ecx,2), %ebp
1008
movq %mm1, 1056(%esp)
1009
leal (%ebp,%ebp,8), %ebp
1010
movq %mm5, 1080(%esp)
1011
movq %mm6, 1112(%esp)
1012
movq (%edx,%ebp), %mm1
1013
movq (%eax,%ebp), %mm5
1014
leal (%ecx,%ecx,4), %ebp
1016
movq %mm0, 1088(%esp)
1018
leal (%ecx,%ebp,4), %ebp
1020
leal (%ecx,%ebp,2), %ebp
1023
movq (%eax,%ebp), %mm0
1024
movq (%edx,%ebp), %mm2
1027
leal (,%ecx,4), %ebp
1028
leal (%ecx,%ebp,8), %ebp
1041
movq %mm0, 1152(%esp)
1045
movq 1112(%esp), %mm6
1048
movq %mm3, 1128(%esp)
1049
movq 1104(%esp), %mm3
1051
movq %mm4, 1120(%esp)
1063
movq 1096(%esp), %mm6
1066
movq 1048(%esp), %mm3
1067
pfmul KP707106781KP707106781, %mm5
1068
pfmul KP707106781KP707106781, %mm1
1069
pfmul KP707106781KP707106781, %mm4
1070
pfmul KP707106781KP707106781, %mm0
1071
movq %mm1, 1160(%esp)
1074
movq %mm4, 1144(%esp)
1075
movq 968(%esp), %mm4
1077
movq %mm0, 1136(%esp)
1078
movq 960(%esp), %mm0
1079
pfmul KP707106781KP707106781, %mm6
1081
pfmul KP707106781KP707106781, %mm3
1090
movq %mm4, 1168(%esp)
1095
movq %mm0, 1176(%esp)
1096
movq %mm1, 1184(%esp)
1097
movq %mm6, 1192(%esp)
1098
movq 1040(%esp), %mm5
1099
movq 1032(%esp), %mm0
1100
movq 1088(%esp), %mm6
1103
pfmul KP555570233KP831469612, %mm4
1106
pfmul KP831469612KP555570233, %mm3
1113
pfmul KP831469612KP555570233, %mm0
1114
pfmul KP831469612KP555570233, %mm6
1117
pfmul KP980785280KP195090322, %mm7
1120
pfmul KP195090322KP980785280, %mm5
1121
pfmul KP195090322KP980785280, %mm4
1122
movq %mm0, 1240(%esp)
1123
pfmul KP195090322KP980785280, %mm6
1124
movq %mm3, 1216(%esp)
1125
movq 952(%esp), %mm0
1126
movq 944(%esp), %mm3
1130
movq 936(%esp), %mm6
1133
movq %mm5, 1208(%esp)
1134
movq 928(%esp), %mm5
1137
movq %mm4, 1200(%esp)
1142
pfmul KP707106781KP707106781, %mm6
1143
pfmul KP707106781KP707106781, %mm5
1146
movq 1160(%esp), %mm6
1150
movq %mm0, 1232(%esp)
1151
movq %mm4, 1264(%esp)
1152
movq 1080(%esp), %mm0
1156
movq %mm7, 1224(%esp)
1157
movq %mm3, 1288(%esp)
1162
pfmul KP555570233KP831469612, %mm5
1163
pfmul KP980785280KP195090322, %mm7
1165
pfmul KP831469612KP555570233, %mm4
1167
pfmul KP980785280KP195090322, %mm1
1168
pfmul KP555570233KP831469612, %mm6
1169
pfmul KP555570233KP831469612, %mm3
1170
pfmul KP980785280KP195090322, %mm2
1171
pfmul KP195090322KP980785280, %mm0
1173
movq 1024(%esp), %mm4
1175
movq 984(%esp), %mm7
1178
movq %mm5, 1280(%esp)
1179
movq 920(%esp), %mm5
1181
movq %mm1, 1248(%esp)
1182
movq 912(%esp), %mm0
1185
movq %mm3, 1272(%esp)
1186
movq 1152(%esp), %mm7
1189
movq %mm2, 1256(%esp)
1190
movq 1128(%esp), %mm2
1199
movq 1072(%esp), %mm2
1203
pfmul KP923879532KP382683432, %mm7
1205
pfmul KP382683432KP923879532, %mm5
1206
movq %mm4, 1296(%esp)
1207
movq 1064(%esp), %mm4
1209
movq %mm3, 1304(%esp)
1214
pfmul KP923879532KP382683432, %mm4
1215
pfmul KP923879532KP382683432, %mm2
1217
movq 1056(%esp), %mm5
1219
movq %mm4, 1320(%esp)
1220
movq 1120(%esp), %mm2
1222
movq %mm7, 1312(%esp)
1228
movq 1008(%esp), %mm6
1233
movq %mm7, 1336(%esp)
1234
movq 1016(%esp), %mm4
1236
pfmul KP382683432KP923879532, %mm6
1237
movq %mm5, 1328(%esp)
1238
pfmul KP382683432KP923879532, %mm7
1239
movq %mm1, 1392(%esp)
1241
pfmul KP923879532KP382683432, %mm4
1242
movq %mm2, 1400(%esp)
1243
movq 1000(%esp), %mm1
1244
pfmul KP382683432KP923879532, %mm5
1246
pfmul KP382683432KP923879532, %mm1
1248
pfmul KP923879532KP382683432, %mm2
1250
movq 992(%esp), %mm7
1255
pfmul KP923879532KP382683432, %mm7
1256
pfmul KP923879532KP382683432, %mm4
1257
movq %mm6, 1344(%esp)
1259
pfmul KP382683432KP923879532, %mm3
1260
movq %mm5, 1352(%esp)
1262
pfmul KP923879532KP382683432, %mm0
1263
pfmul KP382683432KP923879532, %mm6
1264
pfmul KP382683432KP923879532, %mm5
1267
movq (%edx,%ebp), %mm4
1269
movq (%eax,%ebp), %mm0
1270
leal (%ecx,%ecx), %ebp
1273
leal (%ebp,%ebp,2), %ebp
1276
movq %mm3, 1384(%esp)
1277
movq (%edx,%ecx), %mm1
1278
movq (%eax,%ecx), %mm3
1280
leal (%ecx,%ebp,8), %ebp
1282
movq %mm7, 1368(%esp)
1283
movq %mm5, 1376(%esp)
1285
movq (%eax,%ebp), %mm5
1286
movq (%edx,%ebp), %mm4
1287
leal (%ecx,%ecx), %ebp
1288
movq %mm2, 1360(%esp)
1291
leal (%ecx,%ebp,8), %ebp
1293
movq %mm6, 1448(%esp)
1295
movq (%edx,%ebp), %mm1
1297
movq %mm3, 1440(%esp)
1298
movq (%eax,%ebp), %mm3
1299
leal (%ecx,%ecx,2), %ebp
1303
leal (%ecx,%ebp,8), %ebp
1308
movq %mm7, 1408(%esp)
1309
movq (%edx,%ebp), %mm7
1311
movq %mm6, 1464(%esp)
1312
movq (%eax,%ebp), %mm6
1313
leal (%ecx,%ecx,8), %ebp
1320
movq (%edx,%ebp), %mm3
1322
movq %mm5, 1416(%esp)
1323
movq (%eax,%ebp), %mm5
1324
leal (%ecx,%ecx,2), %ebp
1326
movq %mm1, 1424(%esp)
1327
leal (%ecx,%ebp,2), %ebp
1329
movq %mm0, 1456(%esp)
1330
movq %mm4, 1432(%esp)
1331
leal (%ecx,%ebp,8), %ebp
1334
movq (%eax,%ebp), %mm7
1335
movq (%edx,%ebp), %mm0
1336
leal (%ecx,%ecx,4), %ebp
1339
leal (%ecx,%ebp,8), %ebp
1340
movq %mm5, 1472(%esp)
1342
movq (%eax,%ebp), %mm5
1344
movq (%edx,%ebp), %mm3
1346
leal (%ecx,%ecx,4), %ebp
1348
leal (%ecx,%ebp,4), %ebp
1360
movq %mm0, 1488(%esp)
1361
movq %mm3, 1504(%esp)
1362
movq 1472(%esp), %mm0
1369
movq %mm3, 1480(%esp)
1370
movq (%edx,%ebp), %mm3
1372
movq %mm1, 1496(%esp)
1373
movq (%eax,%ebp), %mm1
1374
leal (%ecx,%ecx,4), %ebp
1381
movq (%edx,%ebp), %mm0
1383
movq %mm6, 1528(%esp)
1384
movq (%eax,%ebp), %mm6
1385
leal (%ecx,%ecx,8), %ebp
1387
movq %mm5, 1520(%esp)
1388
leal (%ecx,%ebp,4), %ebp
1390
movq %mm7, 1536(%esp)
1391
movq %mm2, 1512(%esp)
1392
movq (%edx,%ebp), %mm3
1393
movq (%eax,%ebp), %mm7
1394
leal (%ecx,%ecx,2), %ebp
1397
leal (%ecx,%ebp,4), %ebp
1398
movq %mm4, 1544(%esp)
1400
leal (%ecx,%ebp,4), %ebp
1403
movq (%edx,%ebp), %mm0
1404
movq (%eax,%ebp), %mm4
1406
leal (%ecx,%ecx,2), %ebp
1407
leal (%ecx,%ebp,4), %ebp
1417
movq %mm4, 1552(%esp)
1422
movq 1544(%esp), %mm7
1426
movq %mm4, 1568(%esp)
1428
movq %mm5, 1600(%esp)
1429
movq (%edx,%ebp), %mm5
1434
movq (%eax,%ebp), %mm1
1435
leal (%ecx,%ecx,2), %ebp
1437
leal (%ebp,%ebp,4), %ebp
1438
movq %mm0, 1584(%esp)
1439
movq %mm2, 1576(%esp)
1440
leal (%ecx,%ebp,4), %ebp
1441
movq %mm6, 1608(%esp)
1445
movq (%eax,%ebp), %mm2
1446
movq (%edx,%ebp), %mm6
1447
leal (%ecx,%ecx,2), %ebp
1450
leal (%ebp,%ebp,8), %ebp
1452
movq %mm1, 1616(%esp)
1453
leal (%ebp,%ecx,2), %ebp
1456
pfmul KP707106781KP707106781, %mm3
1457
pfmul KP707106781KP707106781, %mm7
1458
movq (%edx,%ebp), %mm5
1460
movq (%eax,%ebp), %mm1
1461
leal (%ecx,%ecx,4), %ebp
1462
leal (%ebp,%ebp,8), %ebp
1463
movq %mm3, 1592(%esp)
1464
movq %mm7, 1560(%esp)
1465
movq (%edx,%ebp), %mm3
1466
movq (%eax,%ebp), %mm2
1469
/* simd data load/store barrier */
1470
leal (,%esi,4), %ebp
1472
addl 2796(%esp), %eax
1473
addl 2796(%esp), %edx
1483
movq %mm2, 1624(%esp)
1488
movq 1616(%esp), %mm1
1492
movq %mm2, 1648(%esp)
1499
movq 1608(%esp), %mm0
1501
movq %mm7, 1656(%esp)
1505
movq 1464(%esp), %mm2
1508
movq 1456(%esp), %mm6
1509
pfmul KP707106781KP707106781, %mm1
1510
pfmul KP707106781KP707106781, %mm5
1511
movq %mm0, 1640(%esp)
1515
movq %mm1, 1632(%esp)
1516
movq 1536(%esp), %mm6
1522
movq %mm2, 1664(%esp)
1523
movq 1528(%esp), %mm6
1526
movq %mm1, 1672(%esp)
1528
movq 1448(%esp), %mm6
1531
movq %mm2, 1680(%esp)
1533
movq 1440(%esp), %mm2
1536
movq %mm0, 1688(%esp)
1537
movq 1432(%esp), %mm0
1539
movq %mm1, 1696(%esp)
1540
movq 1424(%esp), %mm1
1547
pfmul KP707106781KP707106781, %mm0
1548
pfmul KP707106781KP707106781, %mm1
1554
movq %mm6, 1720(%esp)
1555
movq %mm2, 1712(%esp)
1556
movq %mm7, 1728(%esp)
1557
movq 1416(%esp), %mm1
1558
movq 1408(%esp), %mm6
1559
movq 1520(%esp), %mm7
1561
movq %mm0, 1704(%esp)
1562
movq 1512(%esp), %mm0
1571
pfmul KP707106781KP707106781, %mm7
1572
pfmul KP707106781KP707106781, %mm0
1580
movq %mm1, 1736(%esp)
1582
movq %mm2, 1744(%esp)
1583
movq %mm6, 1792(%esp)
1584
movq %mm7, 1752(%esp)
1585
movq 1600(%esp), %mm5
1586
movq 1592(%esp), %mm2
1589
pfmul KP980785280KP195090322, %mm0
1592
pfmul KP980785280KP195090322, %mm1
1594
pfmul KP980785280KP195090322, %mm5
1596
pfmul KP555570233KP831469612, %mm6
1597
pfmul KP195090322KP980785280, %mm7
1598
pfmul KP831469612KP555570233, %mm2
1601
pfmul KP555570233KP831469612, %mm4
1602
pfmul KP555570233KP831469612, %mm0
1604
movq %mm5, 1760(%esp)
1605
movq %mm1, 1768(%esp)
1606
movq 1584(%esp), %mm7
1607
movq 1624(%esp), %mm1
1610
movq 1656(%esp), %mm4
1615
movq %mm6, 1776(%esp)
1616
movq 1552(%esp), %mm6
1618
movq %mm0, 1784(%esp)
1619
movq 1576(%esp), %mm0
1630
pfmul KP923879532KP382683432, %mm2
1632
pfmul KP707106781KP707106781, %mm5
1633
pfmul KP707106781KP707106781, %mm3
1635
pfmul KP923879532KP382683432, %mm6
1636
pfmul KP382683432KP923879532, %mm7
1637
pfmul KP923879532KP382683432, %mm0
1638
movq %mm5, 1816(%esp)
1639
movq %mm3, 1824(%esp)
1640
movq 1568(%esp), %mm5
1641
movq 1560(%esp), %mm3
1644
movq 1648(%esp), %mm7
1647
movq %mm2, 1808(%esp)
1648
movq 1632(%esp), %mm2
1651
movq %mm0, 1800(%esp)
1655
pfmul KP831469612KP555570233, %mm6
1656
pfmul KP831469612KP555570233, %mm0
1658
pfmul KP831469612KP555570233, %mm7
1659
pfmul KP555570233KP831469612, %mm2
1662
pfmul KP195090322KP980785280, %mm3
1665
pfmul KP195090322KP980785280, %mm5
1666
pfmul KP980785280KP195090322, %mm6
1667
movq %mm7, 1856(%esp)
1668
pfmul KP195090322KP980785280, %mm2
1669
movq %mm0, 1872(%esp)
1670
movq 1504(%esp), %mm7
1672
movq 1496(%esp), %mm5
1674
pfmul KP382683432KP923879532, %mm7
1676
pfmul KP923879532KP382683432, %mm0
1678
pfmul KP382683432KP923879532, %mm5
1679
movq %mm3, 1832(%esp)
1680
movq 1488(%esp), %mm3
1681
pfmul KP382683432KP923879532, %mm6
1682
movq %mm2, 1840(%esp)
1684
pfmul KP923879532KP382683432, %mm3
1686
pfmul KP382683432KP923879532, %mm2
1688
movq 1480(%esp), %mm0
1693
pfmul KP923879532KP382683432, %mm0
1694
pfmul KP923879532KP382683432, %mm5
1695
movq %mm7, 1848(%esp)
1697
pfmul KP382683432KP923879532, %mm4
1698
movq %mm6, 1880(%esp)
1700
pfmul KP923879532KP382683432, %mm1
1701
pfmul KP382683432KP923879532, %mm7
1702
pfmul KP382683432KP923879532, %mm6
1705
movq 1400(%esp), %mm2
1708
movq 1696(%esp), %mm1
1712
movq %mm7, 1896(%esp)
1713
movq 192(%esp), %mm0
1714
movq 400(%esp), %mm7
1717
movq %mm3, 1888(%esp)
1718
movq %mm6, 1904(%esp)
1719
movq 904(%esp), %mm6
1722
movq %mm5, 1864(%esp)
1730
movq 1392(%esp), %mm6
1734
movq 1688(%esp), %mm1
1738
movq %mm0, (%ebx,%ebp,8)
1740
movq %mm2, (%edi,%ebp,8)
1743
movq 896(%esp), %mm5
1744
leal (%esi,%esi), %ebp
1754
movq 120(%esp), %mm0
1756
pfmul KP707106781KP707106781, %mm4
1758
movq 360(%esp), %mm5
1759
pfmul KP707106781KP707106781, %mm6
1769
movq 888(%esp), %mm6
1770
movq %mm7, (%ebx,%ebp,8)
1771
movq 880(%esp), %mm7
1773
movq %mm3, (%edi,%ebp,8)
1774
leal (%esi,%esi), %ebp
1776
movq 1336(%esp), %mm5
1777
leal (%ebp,%ebp,2), %ebp
1780
movq %mm0, 1920(%esp)
1781
movq 1328(%esp), %mm0
1782
movq %mm1, (%ebx,%ebp,8)
1783
movq 1640(%esp), %mm1
1785
movq 1680(%esp), %mm3
1787
movq %mm2, (%edi,%ebp,8)
1788
leal (%esi,%esi,2), %ebp
1789
pfmul KP707106781KP707106781, %mm6
1793
pfmul KP707106781KP707106781, %mm7
1797
movq %mm6, 1912(%esp)
1799
pfmul KP382683432KP923879532, %mm5
1802
pfmul KP923879532KP382683432, %mm6
1804
pfmul KP923879532KP382683432, %mm3
1805
pfmul KP923879532KP382683432, %mm1
1808
pfmul KP923879532KP382683432, %mm2
1811
pfmul KP382683432KP923879532, %mm5
1812
pfmul KP382683432KP923879532, %mm1
1813
pfmul KP382683432KP923879532, %mm0
1829
movq 1920(%esp), %mm5
1831
movq %mm4, (%ebx,%ebp,8)
1832
movq 1912(%esp), %mm4
1834
movq %mm2, (%edi,%ebp,8)
1835
leal (%esi,%esi,2), %ebp
1837
leal (%esi,%ebp,2), %ebp
1840
movq %mm1, (%edi,%ebp,8)
1841
movq 280(%esp), %mm1
1842
movq %mm7, (%ebx,%ebp,8)
1843
movq 288(%esp), %mm7
1844
leal (%esi,%esi,4), %ebp
1853
movq 808(%esp), %mm3
1854
movq %mm5, (%edi,%ebp,8)
1856
movq 112(%esp), %mm6
1857
movq %mm0, (%ebx,%esi,8)
1859
movq %mm2, (%ebx,%ebp,8)
1860
leal (%esi,%esi,2), %ebp
1861
pfmul KP707106781KP707106781, %mm1
1862
movq %mm4, (%edi,%esi,8)
1863
movq 800(%esp), %mm4
1866
pfmul KP382683432KP923879532, %mm3
1867
leal (%esi,%ebp,2), %ebp
1868
pfmul KP923879532KP382683432, %mm2
1871
pfmul KP707106781KP707106781, %mm7
1872
movq %mm1, 1928(%esp)
1874
pfmul KP923879532KP382683432, %mm4
1875
pfmul KP382683432KP923879532, %mm1
1878
movq %mm0, 1936(%esp)
1879
movq 1672(%esp), %mm0
1885
movq 1824(%esp), %mm7
1890
movq %mm6, 1952(%esp)
1892
movq 1304(%esp), %mm7
1894
movq %mm1, 1944(%esp)
1895
movq 1144(%esp), %mm2
1898
pfmul KP195090322KP980785280, %mm4
1900
movq %mm3, 1960(%esp)
1903
pfmul KP831469612KP555570233, %mm0
1904
pfmul KP195090322KP980785280, %mm6
1905
pfmul KP831469612KP555570233, %mm3
1907
pfmul KP831469612KP555570233, %mm7
1908
pfmul KP555570233KP831469612, %mm2
1910
movq 1960(%esp), %mm3
1913
pfmul KP195090322KP980785280, %mm1
1914
pfmul KP980785280KP195090322, %mm2
1930
movq 1952(%esp), %mm1
1931
movq %mm5, (%ebx,%ebp,4)
1932
movq 1944(%esp), %mm5
1934
movq %mm3, (%edi,%ebp,4)
1936
leal (%esi,%esi,2), %ebp
1938
leal (%ebp,%ebp,4), %ebp
1941
movq %mm2, (%ebx,%ebp,4)
1942
movq 792(%esp), %mm2
1943
movq %mm6, (%edi,%ebp,4)
1944
leal (%esi,%esi,2), %ebp
1951
movq 1936(%esp), %mm0
1952
pfmul KP923879532KP382683432, %mm2
1954
movq 1928(%esp), %mm7
1955
pfmul KP382683432KP923879532, %mm6
1956
movq %mm1, (%edi,%ebp,4)
1957
movq %mm4, (%ebx,%ebp,4)
1958
movq 784(%esp), %mm4
1961
leal (%esi,%esi,2), %ebp
1963
leal (%ebp,%esi,8), %ebp
1965
pfmul KP382683432KP923879532, %mm4
1966
movq %mm3, (%edi,%ebp,4)
1967
movq 1296(%esp), %mm3
1968
pfmul KP923879532KP382683432, %mm7
1969
movq %mm5, (%ebx,%ebp,4)
1970
leal (%esi,%esi,4), %ebp
1971
movq %mm1, 1976(%esp)
1972
movq 1136(%esp), %mm5
1978
movq 1664(%esp), %mm5
1982
pfmul KP555570233KP831469612, %mm1
1984
movq 1816(%esp), %mm4
1985
pfmul KP555570233KP831469612, %mm6
1986
movq %mm7, 1968(%esp)
1989
movq %mm2, 1984(%esp)
1992
pfmul KP980785280KP195090322, %mm3
1993
pfmul KP980785280KP195090322, %mm2
1995
pfmul KP980785280KP195090322, %mm5
1996
pfmul KP195090322KP980785280, %mm4
1998
movq 1984(%esp), %mm2
2001
pfmul KP555570233KP831469612, %mm7
2002
pfmul KP831469612KP555570233, %mm4
2019
movq 1968(%esp), %mm7
2021
movq 1976(%esp), %mm1
2022
movq %mm0, (%ebx,%ebp,4)
2023
movq %mm4, (%edi,%ebp,4)
2024
leal (%esi,%esi,2), %ebp
2027
leal (%esi,%ebp,4), %ebp
2029
movq %mm6, (%edi,%ebp,4)
2030
movq 648(%esp), %mm6
2033
movq %mm2, (%ebx,%ebp,4)
2034
movq 776(%esp), %mm2
2035
leal (%esi,%esi,8), %ebp
2041
movq %mm1, (%edi,%ebp,4)
2042
movq 768(%esp), %mm1
2045
movq 728(%esp), %mm2
2046
movq %mm4, (%ebx,%ebp,4)
2047
movq %mm0, (%edi,%esi,4)
2048
leal (%esi,%esi,8), %ebp
2049
movq %mm7, (%ebx,%esi,4)
2052
pfmul KP555570233KP831469612, %mm6
2053
leal (%esi,%ebp,2), %ebp
2056
pfmul KP980785280KP195090322, %mm3
2058
pfmul KP831469612KP555570233, %mm5
2059
pfmul KP195090322KP980785280, %mm4
2061
pfmul KP831469612KP555570233, %mm1
2063
pfmul KP980785280KP195090322, %mm7
2064
pfmul KP555570233KP831469612, %mm0
2065
pfmul KP195090322KP980785280, %mm2
2069
movq 392(%esp), %mm0
2071
movq 384(%esp), %mm3
2079
movq %mm5, 2040(%esp)
2081
movq 184(%esp), %mm5
2084
movq %mm2, 2024(%esp)
2085
movq %mm1, 2008(%esp)
2087
movq %mm4, 1992(%esp)
2088
movq 1384(%esp), %mm4
2091
movq 1376(%esp), %mm5
2099
movq 1192(%esp), %mm0
2101
movq %mm7, 2032(%esp)
2103
movq %mm6, 2016(%esp)
2104
movq %mm2, 2000(%esp)
2105
movq 1904(%esp), %mm3
2106
movq 1896(%esp), %mm6
2108
movq %mm1, 2048(%esp)
2112
movq 1792(%esp), %mm4
2117
pfmul KP773010453KP634393284, %mm7
2120
pfmul KP773010453KP634393284, %mm1
2122
pfmul KP634393284KP773010453, %mm4
2123
pfmul KP773010453KP634393284, %mm3
2126
pfmul KP995184726KP098017140, %mm2
2129
pfmul KP098017140KP995184726, %mm0
2130
movq %mm7, 2056(%esp)
2131
movq 1752(%esp), %mm7
2132
pfmul KP098017140KP995184726, %mm4
2133
pfmul KP098017140KP995184726, %mm1
2134
movq %mm3, 2064(%esp)
2138
movq 1184(%esp), %mm2
2144
pfmul KP956940335KP290284677, %mm7
2145
pfmul KP956940335KP290284677, %mm1
2148
pfmul KP290284677KP956940335, %mm2
2149
pfmul KP956940335KP290284677, %mm5
2152
pfmul KP881921264KP471396736, %mm6
2155
pfmul KP471396736KP881921264, %mm2
2156
pfmul KP471396736KP881921264, %mm1
2157
pfmul KP471396736KP881921264, %mm3
2159
movq 2048(%esp), %mm1
2161
movq %mm2, 2072(%esp)
2162
movq 2040(%esp), %mm6
2176
movq %mm6, (%ebx,%ebp,2)
2177
movq 2032(%esp), %mm6
2179
movq %mm2, (%edi,%ebp,2)
2180
movq 2024(%esp), %mm2
2182
leal (%esi,%esi,2), %ebp
2184
movq %mm1, (%ebx,%ebp,2)
2186
movq %mm5, (%edi,%ebp,2)
2187
leal (%esi,%esi,2), %ebp
2189
movq 2016(%esp), %mm2
2190
leal (%ebp,%ebp,4), %ebp
2196
movq 2008(%esp), %mm4
2199
movq %mm6, (%edi,%ebp,2)
2200
movq 2072(%esp), %mm6
2201
movq %mm1, (%ebx,%ebp,2)
2203
leal (%esi,%esi,2), %ebp
2206
leal (%ebp,%ebp,4), %ebp
2209
leal (%esi,%ebp,2), %ebp
2211
movq %mm5, (%ebx,%ebp,2)
2212
movq 2000(%esp), %mm5
2213
movq %mm7, (%edi,%ebp,2)
2215
leal (%esi,%esi,2), %ebp
2217
movq 1992(%esp), %mm7
2219
movq 2064(%esp), %mm6
2220
leal (%ebp,%ebp,8), %ebp
2224
movq %mm2, (%ebx,%ebp,2)
2225
movq 2056(%esp), %mm2
2227
movq %mm1, (%edi,%ebp,2)
2228
leal (%esi,%esi,2), %ebp
2231
leal (%ebp,%esi,8), %ebp
2235
movq %mm4, (%ebx,%ebp,2)
2237
movq %mm0, (%edi,%ebp,2)
2238
leal (%esi,%esi,2), %ebp
2239
movq 720(%esp), %mm4
2241
leal (%esi,%ebp,2), %ebp
2243
movq 640(%esp), %mm6
2246
movq 712(%esp), %mm2
2248
pfmul KP980785280KP195090322, %mm4
2249
movq %mm1, (%ebx,%ebp,2)
2250
movq %mm3, (%edi,%ebp,2)
2252
leal (%esi,%esi,2), %ebp
2253
pfmul KP980785280KP195090322, %mm0
2255
leal (%ebp,%ebp,4), %ebp
2257
leal (%ebp,%esi,8), %ebp
2259
pfmul KP195090322KP980785280, %mm6
2260
movq %mm7, (%ebx,%ebp,2)
2261
movq 704(%esp), %mm7
2262
movq %mm5, (%edi,%ebp,2)
2264
pfmul KP980785280KP195090322, %mm3
2265
leal (%esi,%esi), %ebp
2266
pfmul KP555570233KP831469612, %mm2
2267
pfmul KP831469612KP555570233, %mm1
2268
leal (%esi,%ebp,8), %ebp
2270
pfmul KP555570233KP831469612, %mm7
2272
pfmul KP555570233KP831469612, %mm5
2277
movq 176(%esp), %mm1
2283
movq %mm4, 2096(%esp)
2284
movq 376(%esp), %mm4
2286
movq %mm3, 2128(%esp)
2287
movq 368(%esp), %mm3
2295
movq %mm2, 2080(%esp)
2297
movq %mm7, 2112(%esp)
2298
movq 1320(%esp), %mm2
2299
movq 1312(%esp), %mm7
2305
movq 1176(%esp), %mm3
2307
movq %mm6, 2104(%esp)
2309
movq %mm5, 2088(%esp)
2310
movq %mm1, 2136(%esp)
2311
movq 1808(%esp), %mm4
2312
movq 1800(%esp), %mm5
2314
movq %mm0, 2120(%esp)
2318
movq 1744(%esp), %mm2
2323
pfmul KP881921264KP471396736, %mm6
2326
pfmul KP881921264KP471396736, %mm0
2328
pfmul KP471396736KP881921264, %mm2
2329
pfmul KP881921264KP471396736, %mm4
2332
pfmul KP956940335KP290284677, %mm1
2335
pfmul KP290284677KP956940335, %mm3
2336
movq %mm6, 2144(%esp)
2337
movq 1736(%esp), %mm6
2338
pfmul KP290284677KP956940335, %mm2
2339
pfmul KP290284677KP956940335, %mm0
2340
movq %mm4, 2152(%esp)
2344
movq 1168(%esp), %mm1
2350
pfmul KP995184726KP098017140, %mm6
2351
pfmul KP995184726KP098017140, %mm0
2354
pfmul KP098017140KP995184726, %mm1
2355
pfmul KP995184726KP098017140, %mm7
2358
pfmul KP773010453KP634393284, %mm5
2361
pfmul KP634393284KP773010453, %mm1
2362
pfmul KP634393284KP773010453, %mm0
2363
pfmul KP634393284KP773010453, %mm4
2365
movq 2136(%esp), %mm0
2367
movq %mm1, 2160(%esp)
2368
movq 2128(%esp), %mm5
2382
movq %mm5, (%ebx,%ebp,2)
2383
movq 2120(%esp), %mm5
2385
movq %mm1, (%edi,%ebp,2)
2386
movq 2112(%esp), %mm1
2388
movq %mm0, (%edi,%esi,2)
2389
leal (%esi,%esi,2), %ebp
2390
movq %mm7, (%ebx,%esi,2)
2392
leal (%esi,%ebp,4), %ebp
2395
movq 2104(%esp), %mm1
2401
movq 2096(%esp), %mm2
2404
movq %mm5, (%edi,%ebp,2)
2405
movq 2160(%esp), %mm5
2406
movq %mm6, (%ebx,%ebp,2)
2408
leal (%esi,%esi,2), %ebp
2411
leal (%ebp,%ebp,8), %ebp
2414
leal (%ebp,%esi,2), %ebp
2416
movq %mm7, (%ebx,%ebp,2)
2417
movq 2088(%esp), %mm7
2418
movq %mm0, (%edi,%ebp,2)
2420
leal (%esi,%esi,2), %ebp
2422
movq 2080(%esp), %mm0
2424
leal (%esi,%ebp,8), %ebp
2427
movq 2152(%esp), %mm4
2429
movq %mm1, (%ebx,%ebp,2)
2430
movq 2144(%esp), %mm1
2433
movq %mm6, (%edi,%ebp,2)
2434
leal (%esi,%esi,8), %ebp
2437
movq %mm2, (%ebx,%ebp,2)
2439
movq 760(%esp), %mm0
2440
movq %mm3, (%edi,%ebp,2)
2443
leal (%esi,%esi,4), %ebp
2444
movq 856(%esp), %mm3
2447
movq 696(%esp), %mm4
2451
movq %mm7, (%edi,%ebp,2)
2452
movq 872(%esp), %mm7
2456
movq %mm6, (%ebx,%ebp,2)
2457
leal (%esi,%esi,4), %ebp
2460
leal (%esi,%ebp,4), %ebp
2463
pfmul KP290284677KP956940335, %mm6
2464
pfmul KP290284677KP956940335, %mm0
2465
movq %mm5, (%edi,%ebp,2)
2466
movq %mm2, (%ebx,%ebp,2)
2467
leal (%esi,%esi,4), %ebp
2469
pfmul KP881921264KP471396736, %mm1
2471
pfmul KP290284677KP956940335, %mm4
2472
leal (%ebp,%ebp,8), %ebp
2473
pfmul KP881921264KP471396736, %mm7
2474
pfmul KP471396736KP881921264, %mm3
2475
pfmul KP881921264KP471396736, %mm5
2476
pfmul KP956940335KP290284677, %mm2
2479
movq 352(%esp), %mm3
2481
movq 136(%esp), %mm7
2483
movq %mm4, 2240(%esp)
2484
movq %mm1, 2168(%esp)
2485
movq 344(%esp), %mm2
2486
movq 128(%esp), %mm4
2488
pfmul KP831469612KP555570233, %mm3
2489
movq %mm5, 2184(%esp)
2490
movq %mm0, 2232(%esp)
2491
movq 104(%esp), %mm5
2494
pfmul KP555570233KP831469612, %mm6
2496
pfmul KP555570233KP831469612, %mm2
2498
movq 168(%esp), %mm0
2499
pfmul KP831469612KP555570233, %mm1
2500
movq %mm4, 2176(%esp)
2512
movq %mm4, 2192(%esp)
2513
movq 1728(%esp), %mm4
2515
movq 1888(%esp), %mm0
2522
movq %mm5, 2200(%esp)
2523
movq 1784(%esp), %mm6
2524
movq 1776(%esp), %mm5
2526
movq %mm1, 2208(%esp)
2528
movq %mm3, 2216(%esp)
2530
movq %mm7, 2224(%esp)
2540
pfmul KP146730474KP989176509, %mm4
2542
pfmul KP514102744KP857728610, %mm2
2543
pfmul KP146730474KP989176509, %mm7
2545
pfmul KP857728610KP514102744, %mm1
2547
pfmul KP803207531KP595699304, %mm0
2549
pfmul KP242980179KP970031253, %mm3
2550
movq %mm4, 2248(%esp)
2551
movq %mm2, 2264(%esp)
2552
movq %mm7, 2256(%esp)
2553
movq %mm1, 2272(%esp)
2554
movq %mm0, 2280(%esp)
2555
movq 1288(%esp), %mm4
2556
movq 1368(%esp), %mm2
2557
movq 1272(%esp), %mm7
2558
movq 1280(%esp), %mm1
2560
pfmul KP970031253KP242980179, %mm5
2562
pfmul KP803207531KP595699304, %mm6
2571
pfmul KP970031253KP242980179, %mm4
2573
pfmul KP970031253KP242980179, %mm1
2580
pfmul KP514102744KP857728610, %mm7
2581
movq %mm4, 2304(%esp)
2583
pfmul KP514102744KP857728610, %mm2
2584
movq 2264(%esp), %mm4
2586
pfmul KP595699304KP803207531, %mm3
2587
pfmul KP803207531KP595699304, %mm1
2588
movq %mm5, 2312(%esp)
2590
pfmul KP989176509KP146730474, %mm0
2591
pfpnacc 2272(%esp), %mm7
2592
pfmul KP146730474KP989176509, %mm5
2594
movq 2280(%esp), %mm2
2596
movq 2240(%esp), %mm6
2598
movq %mm4, 2296(%esp)
2599
movq %mm7, 2288(%esp)
2600
movq 2248(%esp), %mm3
2601
movq 2232(%esp), %mm7
2602
pfpnacc 2256(%esp), %mm5
2605
movq 2224(%esp), %mm0
2621
movq %mm4, (%ebx,%ebp)
2622
movq 2216(%esp), %mm4
2624
movq %mm6, (%edi,%ebp)
2625
leal (%esi,%esi,2), %ebp
2627
leal (%esi,%ebp,4), %ebp
2630
movq %mm1, (%ebx,%ebp)
2632
movq 2184(%esp), %mm7
2633
movq %mm0, (%edi,%ebp)
2634
movq 2168(%esp), %mm0
2635
leal (%esi,%esi,2), %ebp
2638
leal (%ebp,%ebp,4), %ebp
2643
leal (%esi,%ebp,4), %ebp
2645
movq 2208(%esp), %mm3
2647
movq %mm4, (%edi,%ebp)
2649
movq 2304(%esp), %mm5
2650
movq %mm2, (%ebx,%ebp)
2651
movq 2312(%esp), %mm2
2652
leal (%esi,%esi,2), %ebp
2655
leal (%ebp,%ebp,8), %ebp
2659
leal (%ebp,%esi,2), %ebp
2661
movq %mm6, (%ebx,%ebp)
2663
movq %mm1, (%edi,%ebp)
2665
movq 2200(%esp), %mm1
2666
leal (%esi,%esi,4), %ebp
2670
movq 2296(%esp), %mm5
2674
movq %mm3, (%ebx,%ebp)
2675
movq 2288(%esp), %mm3
2678
movq %mm6, (%edi,%ebp)
2679
leal (%esi,%esi,8), %ebp
2680
movq 336(%esp), %mm6
2682
leal (%esi,%ebp,4), %ebp
2685
movq %mm7, (%ebx,%ebp)
2687
pfmul KP831469612KP555570233, %mm6
2688
movq %mm4, (%edi,%ebp)
2690
leal (%esi,%esi,4), %ebp
2692
pfmul KP555570233KP831469612, %mm7
2694
movq 312(%esp), %mm5
2695
leal (%esi,%ebp,4), %ebp
2698
movq %mm0, (%edi,%ebp)
2699
movq 160(%esp), %mm0
2701
movq %mm1, (%ebx,%ebp)
2702
leal (%esi,%esi,2), %ebp
2703
pfmul KP555570233KP831469612, %mm5
2704
movq 152(%esp), %mm1
2705
pfmul KP831469612KP555570233, %mm3
2706
leal (%esi,%ebp,4), %ebp
2707
leal (%esi,%ebp,4), %ebp
2708
movq %mm2, (%edi,%ebp)
2709
movq 144(%esp), %mm2
2711
movq %mm4, (%ebx,%ebp)
2715
leal (%esi,%esi), %ebp
2718
leal (%esi,%ebp,8), %ebp
2723
leal (%esi,%ebp,2), %ebp
2725
movq %mm1, 2320(%esp)
2730
movq %mm5, 2328(%esp)
2731
movq 752(%esp), %mm0
2732
movq 824(%esp), %mm5
2740
movq %mm7, 2352(%esp)
2741
movq 688(%esp), %mm7
2744
movq %mm1, 2376(%esp)
2745
movq %mm2, 2336(%esp)
2746
movq %mm4, 2344(%esp)
2747
movq 840(%esp), %mm4
2750
pfmul KP290284677KP956940335, %mm0
2752
pfmul KP471396736KP881921264, %mm3
2753
pfmul KP956940335KP290284677, %mm6
2755
pfmul KP881921264KP471396736, %mm1
2758
pfmul KP956940335KP290284677, %mm7
2760
pfmul KP881921264KP471396736, %mm5
2761
pfmul KP290284677KP956940335, %mm2
2762
pfmul KP471396736KP881921264, %mm4
2764
movq 1264(%esp), %mm0
2766
movq 1216(%esp), %mm5
2768
movq 1352(%esp), %mm2
2770
movq %mm7, 2392(%esp)
2771
movq 1240(%esp), %mm7
2773
movq %mm3, 2360(%esp)
2774
movq %mm6, 2384(%esp)
2776
movq %mm4, 2368(%esp)
2788
pfmul KP857728610KP514102744, %mm0
2790
pfmul KP242980179KP970031253, %mm4
2791
pfmul KP857728610KP514102744, %mm3
2794
pfmul KP146730474KP989176509, %mm1
2796
pfmul KP989176509KP146730474, %mm7
2797
pfmul KP242980179KP970031253, %mm6
2798
movq %mm0, 2400(%esp)
2799
movq %mm4, 2416(%esp)
2800
movq %mm3, 2408(%esp)
2801
movq 1720(%esp), %mm0
2802
movq 1880(%esp), %mm4
2803
pfmul KP803207531KP595699304, %mm2
2804
pfmul KP595699304KP803207531, %mm5
2806
movq %mm6, 2424(%esp)
2808
movq 1872(%esp), %mm6
2810
movq %mm3, 2432(%esp)
2811
movq 1856(%esp), %mm4
2819
pfmul KP989176509KP146730474, %mm0
2820
pfmul KP989176509KP146730474, %mm6
2823
pfmul KP595699304KP803207531, %mm3
2825
movq 2432(%esp), %mm6
2827
pfmul KP595699304KP803207531, %mm1
2833
movq 2416(%esp), %mm1
2835
pfmul KP970031253KP242980179, %mm6
2836
pfmul KP242980179KP970031253, %mm4
2839
pfmul KP857728610KP514102744, %mm2
2840
pfmul KP514102744KP857728610, %mm6
2841
pfpnacc 2424(%esp), %mm4
2842
movq %mm1, 2440(%esp)
2843
movq 2400(%esp), %mm1
2844
pfpnacc 2408(%esp), %mm2
2846
movq %mm4, 2448(%esp)
2847
movq %mm1, 2456(%esp)
2848
movq %mm2, 2464(%esp)
2849
movq 2384(%esp), %mm6
2850
movq 2392(%esp), %mm4
2851
movq 2376(%esp), %mm1
2868
movq %mm2, (%edi,%ebp)
2869
movq 2352(%esp), %mm2
2871
movq %mm4, (%ebx,%ebp)
2873
leal (%esi,%esi,2), %ebp
2874
movq %mm7, (%ebx,%ebp)
2877
movq %mm1, (%edi,%ebp)
2878
leal (%esi,%esi), %ebp
2879
movq 2368(%esp), %mm1
2881
movq 2360(%esp), %mm6
2882
leal (%esi,%ebp,8), %ebp
2885
leal (%ebp,%ebp,2), %ebp
2891
movq 2344(%esp), %mm3
2892
movq %mm2, (%edi,%ebp)
2895
movq 2464(%esp), %mm5
2896
movq %mm0, (%ebx,%ebp)
2897
movq 2456(%esp), %mm0
2898
leal (%esi,%esi,8), %ebp
2901
leal (%esi,%ebp,2), %ebp
2905
movq %mm7, (%edi,%ebp)
2906
movq %mm4, (%ebx,%ebp)
2908
leal (%esi,%esi,2), %ebp
2910
movq 2336(%esp), %mm1
2912
leal (%ebp,%esi,8), %ebp
2918
movq 2448(%esp), %mm0
2919
movq %mm3, (%ebx,%ebp)
2920
movq 2440(%esp), %mm3
2922
movq %mm7, (%edi,%ebp)
2923
leal (%esi,%esi,4), %ebp
2926
leal (%esi,%ebp,4), %ebp
2929
leal (%esi,%ebp,2), %ebp
2932
movq %mm4, (%ebx,%ebp)
2933
movq 744(%esp), %mm4
2934
movq %mm2, (%edi,%ebp)
2936
leal (%esi,%esi,2), %ebp
2938
movq 816(%esp), %mm2
2939
leal (%ebp,%ebp,8), %ebp
2942
movq 680(%esp), %mm3
2944
movq %mm6, (%ebx,%ebp)
2945
movq 832(%esp), %mm6
2946
movq %mm1, (%edi,%ebp)
2948
imull $59, %esi, %ebp
2955
pfmul KP773010453KP634393284, %mm0
2956
pfmul KP098017140KP995184726, %mm6
2957
movq %mm5, (%edi,%ebp)
2959
pfmul KP098017140KP995184726, %mm4
2960
movq %mm7, (%ebx,%ebp)
2962
pfmul KP098017140KP995184726, %mm1
2963
leal (%esi,%esi,2), %ebp
2964
pfmul KP634393284KP773010453, %mm5
2965
pfmul KP773010453KP634393284, %mm2
2966
leal (%esi,%ebp,4), %ebp
2967
pfmul KP773010453KP634393284, %mm3
2968
pfmul KP995184726KP098017140, %mm7
2969
leal (%ebp,%ebp,2), %ebp
2971
movq 328(%esp), %mm6
2973
movq 304(%esp), %mm5
2977
pfmul KP195090322KP980785280, %mm6
2979
pfmul KP980785280KP195090322, %mm5
2980
movq %mm1, 2496(%esp)
2981
pfmul KP980785280KP195090322, %mm2
2982
movq 2320(%esp), %mm1
2983
pfmul KP195090322KP980785280, %mm7
2984
movq %mm3, 2528(%esp)
2985
movq %mm0, 2520(%esp)
2986
movq %mm4, 2472(%esp)
2987
movq 2328(%esp), %mm3
2992
movq 1712(%esp), %mm2
2999
movq 1848(%esp), %mm4
3005
movq %mm1, 2512(%esp)
3006
movq 1840(%esp), %mm6
3007
movq 1832(%esp), %mm1
3009
movq %mm0, 2504(%esp)
3011
movq %mm5, 2488(%esp)
3013
movq %mm3, 2480(%esp)
3023
pfmul KP427555093KP903989293, %mm2
3025
pfmul KP049067674KP998795456, %mm7
3026
pfmul KP427555093KP903989293, %mm3
3028
pfmul KP998795456KP049067674, %mm0
3030
pfmul KP941544065KP336889853, %mm4
3032
pfmul KP671558954KP740951125, %mm5
3033
movq %mm2, 2536(%esp)
3034
movq %mm7, 2552(%esp)
3035
movq %mm3, 2544(%esp)
3036
movq %mm0, 2560(%esp)
3037
movq %mm4, 2568(%esp)
3038
movq 1232(%esp), %mm2
3039
movq 1344(%esp), %mm7
3040
movq 1200(%esp), %mm3
3041
movq 1208(%esp), %mm0
3043
pfmul KP740951125KP671558954, %mm1
3045
pfmul KP941544065KP336889853, %mm6
3054
pfmul KP740951125KP671558954, %mm2
3056
pfmul KP740951125KP671558954, %mm0
3063
pfmul KP049067674KP998795456, %mm3
3064
movq %mm2, 2592(%esp)
3066
pfmul KP049067674KP998795456, %mm7
3067
movq 2552(%esp), %mm2
3069
pfmul KP336889853KP941544065, %mm5
3070
pfmul KP941544065KP336889853, %mm0
3071
movq %mm1, 2600(%esp)
3073
pfmul KP903989293KP427555093, %mm4
3074
pfpnacc 2560(%esp), %mm3
3075
pfmul KP427555093KP903989293, %mm1
3077
movq 2568(%esp), %mm7
3079
movq 2528(%esp), %mm6
3081
movq %mm2, 2584(%esp)
3082
movq %mm3, 2576(%esp)
3083
movq 2536(%esp), %mm5
3084
movq 2520(%esp), %mm3
3085
pfpnacc 2544(%esp), %mm1
3088
movq 2512(%esp), %mm4
3104
movq %mm2, (%ebx,%ebp)
3105
movq 2504(%esp), %mm2
3107
movq %mm6, (%edi,%ebp)
3108
leal (%esi,%esi,2), %ebp
3110
leal (%esi,%ebp,2), %ebp
3113
movq %mm0, (%ebx,%ebp)
3115
movq 2496(%esp), %mm3
3116
movq %mm4, (%edi,%ebp)
3117
movq 2472(%esp), %mm4
3118
leal (%esi,%esi,2), %ebp
3121
leal (%ebp,%ebp,8), %ebp
3126
leal (%esi,%ebp,2), %ebp
3128
movq 2488(%esp), %mm5
3130
movq %mm2, (%edi,%ebp)
3132
movq 2592(%esp), %mm1
3133
movq %mm7, (%ebx,%ebp)
3134
movq 2600(%esp), %mm7
3135
leal (%esi,%esi,2), %ebp
3138
leal (%ebp,%ebp,4), %ebp
3142
leal (%ebp,%esi,8), %ebp
3144
movq %mm0, (%edi,%ebp)
3146
movq %mm6, (%ebx,%ebp)
3147
leal (%esi,%esi,2), %ebp
3149
movq 2480(%esp), %mm6
3152
leal (%ebp,%ebp,4), %ebp
3155
movq 2584(%esp), %mm1
3158
movq %mm5, (%ebx,%ebp)
3159
movq 2576(%esp), %mm5
3160
movq %mm0, (%edi,%ebp)
3161
leal (%esi,%esi,4), %ebp
3164
leal (%ebp,%ebp,8), %ebp
3167
leal (%ebp,%esi,2), %ebp
3170
movq %mm2, (%edi,%ebp)
3171
movq 320(%esp), %mm2
3173
movq %mm3, (%ebx,%ebp)
3174
leal (%esi,%esi,2), %ebp
3176
movq 296(%esp), %mm1
3178
leal (%ebp,%ebp,4), %ebp
3181
pfmul KP195090322KP980785280, %mm2
3182
leal (%esi,%ebp,2), %ebp
3184
pfmul KP980785280KP195090322, %mm3
3185
pfmul KP980785280KP195090322, %mm1
3186
pfmul KP195090322KP980785280, %mm5
3187
movq %mm0, (%ebx,%ebp)
3188
movq %mm6, (%edi,%ebp)
3189
leal (%esi,%esi,2), %ebp
3190
movq 2192(%esp), %mm0
3191
movq 2176(%esp), %mm6
3192
leal (%esi,%ebp,2), %ebp
3193
leal (%ebp,%ebp,8), %ebp
3197
movq %mm4, (%ebx,%ebp)
3199
movq %mm7, (%edi,%ebp)
3200
leal (%esi,%esi,4), %ebp
3202
movq 736(%esp), %mm6
3203
leal (%esi,%ebp,8), %ebp
3209
movq 848(%esp), %mm5
3215
movq %mm0, 2648(%esp)
3216
movq 672(%esp), %mm1
3217
movq 864(%esp), %mm0
3220
movq %mm4, 2624(%esp)
3221
movq %mm7, 2608(%esp)
3223
movq %mm2, 2616(%esp)
3226
pfmul KP773010453KP634393284, %mm6
3229
pfmul KP995184726KP098017140, %mm3
3230
pfmul KP634393284KP773010453, %mm4
3231
pfmul KP098017140KP995184726, %mm7
3233
pfmul KP634393284KP773010453, %mm1
3235
pfmul KP995184726KP098017140, %mm5
3236
pfmul KP773010453KP634393284, %mm2
3237
pfmul KP098017140KP995184726, %mm0
3239
movq 1224(%esp), %mm6
3241
movq 1360(%esp), %mm7
3244
movq %mm1, 2664(%esp)
3245
movq 1256(%esp), %mm0
3246
movq 1248(%esp), %mm1
3248
movq %mm5, 2640(%esp)
3250
movq %mm4, 2656(%esp)
3253
movq %mm3, 2632(%esp)
3263
pfmul KP998795456KP049067674, %mm6
3265
pfmul KP671558954KP740951125, %mm4
3266
pfmul KP998795456KP049067674, %mm3
3269
pfmul KP427555093KP903989293, %mm2
3271
pfmul KP903989293KP427555093, %mm0
3272
pfmul KP671558954KP740951125, %mm5
3273
movq %mm6, 2672(%esp)
3274
movq %mm4, 2688(%esp)
3275
movq %mm3, 2680(%esp)
3276
movq 1704(%esp), %mm6
3277
movq 1864(%esp), %mm4
3278
pfmul KP941544065KP336889853, %mm7
3279
pfmul KP336889853KP941544065, %mm1
3281
movq %mm5, 2696(%esp)
3283
movq 1768(%esp), %mm5
3285
movq %mm3, 2704(%esp)
3286
movq 1760(%esp), %mm4
3294
pfmul KP903989293KP427555093, %mm6
3295
pfmul KP903989293KP427555093, %mm5
3298
pfmul KP336889853KP941544065, %mm3
3300
movq 2704(%esp), %mm5
3302
pfmul KP336889853KP941544065, %mm2
3308
movq 2688(%esp), %mm2
3310
pfmul KP740951125KP671558954, %mm5
3311
pfmul KP671558954KP740951125, %mm4
3314
pfmul KP998795456KP049067674, %mm7
3315
pfmul KP049067674KP998795456, %mm5
3316
pfpnacc 2696(%esp), %mm4
3317
movq %mm2, 2712(%esp)
3318
movq 2672(%esp), %mm2
3319
pfpnacc 2680(%esp), %mm7
3321
movq %mm4, 2720(%esp)
3322
movq %mm2, 2728(%esp)
3323
movq %mm7, 2736(%esp)
3324
movq 2656(%esp), %mm5
3325
movq 2664(%esp), %mm4
3326
movq 2648(%esp), %mm2
3343
movq %mm7, (%edi,%ebp)
3344
movq 2624(%esp), %mm7
3346
movq %mm4, (%ebx,%ebp)
3348
leal (%esi,%esi,8), %ebp
3349
movq %mm0, (%ebx,%ebp)
3352
movq %mm2, (%edi,%ebp)
3353
leal (%esi,%esi,2), %ebp
3354
movq 2640(%esp), %mm2
3356
movq 2632(%esp), %mm5
3357
leal (%esi,%ebp,2), %ebp
3360
leal (%esi,%ebp,8), %ebp
3366
movq 2616(%esp), %mm3
3367
movq %mm7, (%edi,%ebp)
3370
movq 2736(%esp), %mm1
3371
movq %mm6, (%ebx,%ebp)
3372
movq 2728(%esp), %mm6
3373
leal (%esi,%esi,2), %ebp
3376
leal (%esi,%ebp,8), %ebp
3380
movq %mm0, (%edi,%ebp)
3382
movq 2608(%esp), %mm2
3383
movq %mm4, (%ebx,%ebp)
3385
leal (,%esi,4), %ebp
3388
leal (%esi,%ebp,8), %ebp
3393
movq 2720(%esp), %mm6
3394
movq %mm3, (%ebx,%esi)
3395
movq 2712(%esp), %mm3
3397
movq %mm0, (%edi,%esi)
3399
movq %mm4, (%ebx,%ebp)
3401
movq %mm7, (%edi,%ebp)
3402
leal (%esi,%esi), %ebp
3406
leal (%esi,%ebp,8), %ebp
3412
movq %mm5, (%ebx,%ebp)
3413
movq %mm2, (%edi,%ebp)
3414
leal (%esi,%esi), %ebp
3415
leal (%ebp,%ebp,2), %ebp
3416
leal (%esi,%ebp,8), %ebp
3417
movq %mm4, (%ebx,%ebp)
3418
addl 2800(%esp), %ebx
3419
movq %mm1, (%edi,%ebp)
3420
addl 2800(%esp), %edi
3424
movl 2760(%esp), %ebx
3425
movl 2756(%esp), %esi
3426
movl 2752(%esp), %edi
3427
movl 2748(%esp), %ebp
3442
.long fftwf_kdft_k7_pgenus
3450
.globl fftwf_codelet_n1k7i_128
3451
fftwf_codelet_n1k7i_128:
3457
call fftwf_kdft_register