-
Notifications
You must be signed in to change notification settings - Fork 6
/
memmove64.asm
1073 lines (961 loc) · 36.9 KB
/
memmove64.asm
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
;************************* memmove64.asm ***********************************
; Author: Agner Fog
; Date created: 2008-07-18
; Last modified: 2013-09-11
; Description:
; Faster version of the standard memmove function:
; void * A_memmove(void *dest, const void *src, size_t count);
; Moves 'count' bytes from 'src' to 'dest'. src and dest may overlap.
;
; Overriding standard function memmove:
; The alias ?OVR_memmove is changed to _memmove in the object file if
; it is desired to override the standard library function memmove.
;
; CPU dispatching included for different CPUs
;
; Copyright (c) 2008-2013 GNU General Public License www.gnu.org/licenses
;******************************************************************************
default rel
global A_memmove: function ; Function A_memmove
global ?OVR_memmove: function ; ?OVR removed if standard function memmove overridden
global memmoveSSE2: function ; Version for processors with only SSE2
global memmoveSSSE3: function ; Version for processors with SSSE3
global memmoveU: function ; Version for processors with fast unaligned read
global memmoveU256: function ; Version for processors with fast 256-bit read/write
global SetMemcpyCacheLimit ; Change limit for bypassing cache
; Imported from memcpy64.asm:
extern A_memcpy ; function entry
extern memcpySSE2 ; CPU specific function entry
extern memcpySSSE3 ; CPU specific function entry
extern memcpyU ; CPU specific function entry
extern memcpyU256 ; CPU specific function entry
; Imported from instrset64.asm
extern InstructionSet ; Instruction set for CPU dispatcher
; Imported from unalignedisfaster64.asm:
extern UnalignedIsFaster ; Tells if unaligned read is faster than PALIGNR
extern Store256BitIsFaster ; Tells if a 256 bit store is faster than two 128 bit stores
; Imported from memcpy64.asm
extern GetMemcpyCacheLimit ; Get the size limit for bypassing cache when copying with memcpy and memmove
extern SetMemcpyCacheLimit1 ; Set the size limit for bypassing cache when copying with memcpy
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;
; Prolog macro. Determine if we should move forwards or backwards
;
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
; Define prolog for this function
; Parameter 1 is forward function label
%MACRO PROLOGM 1
%IFDEF WINDOWS
; Check if dest overlaps src
mov rax, rcx
sub rax, rdx
cmp rax, r8
; We can avoid testing for dest < src by using unsigned compare:
; (Assume that the memory block cannot span across address 0)
; Must move backwards if unsigned(dest-src) < count
jae %1 ; Jump to memcpy if we can move forwards
push rsi
push rdi
mov rdi, rcx ; dest
mov r9, rcx ; dest
mov rsi, rdx ; src
mov rcx, r8 ; count
%ELSE ; Unix
; Check if dest overlaps src
mov rax, rdi
sub rax, rsi
cmp rax, rdx
; Must move backwards if unsigned(dest-src) < count
jae %1 ; Jump to memcpy if we can move forwards
mov rcx, rdx ; count
mov r9, rdi ; dest
%ENDIF
%ENDM
; Define return from this function
%MACRO RETURNM 0
%IFDEF WINDOWS
pop rdi
pop rsi
%ENDIF
mov rax, r9 ; Return value = dest
ret
%ENDMACRO
SECTION .text align=16
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;
; Common entry for dispatch
;
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
; extern "C" void * A_memmove(void * dest, const void * src, size_t count);
; Function entry:
A_memmove:
?OVR_memmove:
jmp qword [memmoveDispatch] ; Go to appropriate version, depending on instruction set
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;
; AVX Version for processors with fast unaligned read and fast 32 bytes write
;
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
align 16
memmoveU256: ; Version for processors with fast 256-bit read/write
memmoveU256@: ; local label
PROLOGM memcpyU256
cmp rcx, 40H
jb A1000 ; Use simpler code if count < 64
; count >= 64
; Note: this part will not always work if count < 64
; Calculate size of last block after last regular boundary of dest
lea edx, [rdi+rcx] ; end of dext
and edx, 1FH
jz B4300 ; Skip if end of dest aligned by 32
; edx = size of last partial block, 1 - 31 bytes
test dl, 3
jz B4210
test dl, 1
jz B4201 ; B4200 if we haven't tested edx,3
; move 1 byte
dec rcx
movzx eax, byte [rsi+rcx]
mov [rdi+rcx], al
B4200: test dl, 2
jz B4210
B4201: ; move 2 bytes
sub rcx, 2
movzx eax, word [rsi+rcx]
mov [rdi+rcx], ax
B4210: test dl, 4
jz B4220
; move 4 bytes
sub rcx, 4
mov eax, [rsi+rcx]
mov [rdi+rcx], eax
B4220: test dl, 8
jz B4230
; move 8 bytes
sub rcx, 8
mov rax, [rsi+rcx]
mov [rdi+rcx], rax
B4230: test dl, 16
jz B4300
; move 16 bytes
sub rcx, 16
movups xmm0, [rsi+rcx]
movaps [rdi+rcx], xmm0
B4300: ; Now end of dest is aligned by 32. Any partial block has been moved
mov rdx, rcx
and ecx, 1FH ; remaining size after 32 bytes blocks moved
and rdx, -20H ; number of 32 bytes blocks
jz H4100
add rsi, rcx
add rdi, rcx
; Check if count very big
cmp rdx, [CacheBypassLimit]
ja H4800 ; Use non-temporal store if count > _CacheBypassLimit
align 16
H4000: ; 32 bytes move loop
vmovups ymm0, [rsi+rdx-20H]
vmovaps [rdi+rdx-20H], ymm0
sub rdx, 20H
jnz H4000
vzeroupper
H4090: sub rsi, rcx
sub rdi, rcx
H4100: ; remaining 0-31 bytes
test ecx, ecx
jz H4600
test cl, 10H
jz H4200
; move 16 bytes
sub ecx, 10H
movups xmm0, [rsi+rcx]
movaps [rdi+rcx], xmm0
jz H4600 ; early out if count divisible by 16
H4200: test cl, 8
jz H4300
; move 8 bytes
sub ecx, 8
mov rax, [rsi+rcx]
mov [rdi+rcx], rax
H4300: test cl, 4
jz H4400
; move 4 bytes
sub ecx, 4
mov eax, [rsi+rcx]
mov [rdi+rcx], eax
jz H4600 ; early out if count divisible by 4
H4400: test cl, 2
jz H4500
; move 2 bytes
sub ecx, 2
movzx eax, word [rsi+rcx]
mov [rdi+rcx], ax
H4500: test cl, 1
jz H4600
; move 1 byte
movzx eax, byte [rsi] ; rcx-1 = 0
mov [rdi], al
H4600: ; finished
RETURNM
align 16
H4800: ; 32 bytes move loop, bypass cache
vmovups ymm0, [rsi+rdx-20H]
vmovntps [rdi+rdx-20H], ymm0
sub rdx, 20H
jnz H4800
vzeroupper
jmp H4090
A1000: ; count < 64. Move 32-16-8-4-2-1 bytes
test cl, 20H
jz A1100
; move 32 bytes
; movups is faster on processors with SSSE3
sub ecx, 20H
movups xmm0, [rsi+rcx+10H]
movups xmm1, [rsi+rcx]
movups [rdi+rcx+10H], xmm0
movups [rdi+rcx], xmm1
A1100: test cl, 10H
jz A1200
; move 16 bytes
sub ecx, 10H
movups xmm0, [rsi+rcx]
movups [rdi+rcx], xmm0
A1200: test cl, 8
jz A1300
; move 8 bytes
sub ecx, 8
mov rax, [rsi+rcx]
mov [rdi+rcx], rax
A1300: test cl, 4
jz A1400
; move 4 bytes
sub ecx, 4
mov eax, [rsi+rcx]
mov [rdi+rcx], eax
jz A1900 ; early out if count divisible by 4
A1400: test cl, 2
jz A1500
; move 2 bytes
sub ecx, 2
movzx eax, word [rsi+rcx]
mov [rdi+rcx], ax
A1500: test cl, 1
jz A1900
; move 1 byte
movzx eax, byte [rsi] ; rcx-1 = 0
mov [rdi], al
A1900: ; finished
RETURNM
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;
; Version for processors with fast unaligned read and fast 16 bytes write
;
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
align 16
memmoveU: ; Version for processors with fast unaligned read
memmoveU@: ; local label
PROLOGM memcpyU
cmp rcx, 40H
jb A1000 ; Use simpler code if count < 64
; count >= 64
; Note: this part will not always work if count < 64
; Calculate size of last block after last regular boundary of dest
lea edx, [rdi+rcx] ; end of dext
and edx, 0FH
jz B3300 ; Skip if end of dest aligned by 16
; edx = size of last partial block, 1 - 15 bytes
test dl, 3
jz B3210
test dl, 1
jz B3201 ; B3200 if we haven't tested edx,3
; move 1 byte
dec rcx
movzx eax, byte [rsi+rcx]
mov [rdi+rcx], al
B3200: test dl, 2
jz B3210
B3201: ; move 2 bytes
sub rcx, 2
movzx eax, word [rsi+rcx]
mov [rdi+rcx], ax
B3210: test dl, 4
jz B3220
; move 4 bytes
sub rcx, 4
mov eax, [rsi+rcx]
mov [rdi+rcx], eax
B3220: test dl, 8
jz B3300
; move 8 bytes
sub rcx, 8
mov rax, [rsi+rcx]
mov [rdi+rcx], rax
B3300: ; Now end of dest is aligned by 16. Any partial block has been moved
mov rdx, rcx
and ecx, 1FH ; remaining size after 32 bytes blocks moved
and rdx, -20H ; number of 32 bytes blocks
jz H1100
add rsi, rcx
add rdi, rcx
; Check if count very big
cmp rdx, [CacheBypassLimit]
ja H1800 ; Use non-temporal store if count > _CacheBypassLimit
align 16 ; minimize 16-bytes boundaries in H1000 loop
H1000: ; 32 bytes move loop
movups xmm1, [rsi+rdx-20H]
movups xmm0, [rsi+rdx-10H]
movaps [rdi+rdx-20H], xmm1
movaps [rdi+rdx-10H], xmm0
sub rdx, 20H
jnz H1000
H1090: sub rsi, rcx
sub rdi, rcx
H1100: ; remaining 0-31 bytes
test ecx, ecx
jz H1600
test cl, 10H
jz H1200
; move 16 bytes
sub ecx, 10H
movups xmm0, [rsi+rcx]
movaps [rdi+rcx], xmm0
jz H1600 ; early out if count divisible by 16
H1200: test cl, 8
jz H1300
; move 8 bytes
sub ecx, 8
mov rax, [rsi+rcx]
mov [rdi+rcx], rax
H1300: test cl, 4
jz H1400
; move 4 bytes
sub ecx, 4
mov eax, [rsi+rcx]
mov [rdi+rcx], eax
jz H1600 ; early out if count divisible by 4
H1400: test cl, 2
jz H1500
; move 2 bytes
sub ecx, 2
movzx eax, word [rsi+rcx]
mov [rdi+rcx], ax
H1500: test cl, 1
jz H1600
; move 1 byte
movzx eax, byte [rsi] ; rcx-1 = 0
mov [rdi], al
H1600: ; finished
RETURNM
align 16
H1800: ; 32 bytes move loop, bypass cache
movups xmm1, [rsi+rdx-20H]
movups xmm0, [rsi+rdx-10H]
movntps [rdi+rdx-20H], xmm1
movntps [rdi+rdx-10H], xmm0
sub rdx, 20H
jnz H1800
jmp H1090
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;
; Version for processors with SSSE3. Aligned read + shift + aligned write
;
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
align 16
memmoveSSSE3: ; SSSE3 version begins here
memmoveSSSE3@: ; local label
PROLOGM memcpySSSE3
; Cannot use memcpy. Must move backwards because of overlap between src and dest
cmp rcx, 40H
jb A1000 ; Use simpler code if count < 64
; count >= 64
; Note: this part will not always work if count < 64
; Calculate size of last block after last regular boundary of dest
lea edx, [rdi+rcx] ; end of dext
and edx, 0FH
jz B1300 ; Skip if end of dest aligned by 16
; edx = size of last partial block, 1 - 15 bytes
test dl, 3
jz B1210
test dl, 1
jz B1201 ; B1200 if we haven't tested edx,3
; move 1 byte
dec rcx
movzx eax, byte [rsi+rcx]
mov [rdi+rcx], al
B1200: test dl, 2
jz B1210
B1201: ; move 2 bytes
sub rcx, 2
movzx eax, word [rsi+rcx]
mov [rdi+rcx], ax
B1210: test dl, 4
jz B1220
; move 4 bytes
sub rcx, 4
mov eax, [rsi+rcx]
mov [rdi+rcx], eax
B1220: test dl, 8
jz B1300
; move 8 bytes
sub rcx, 8
mov rax, [rsi+rcx]
mov [rdi+rcx], rax
B1300: ; Now end of dest is aligned by 16. Any partial block has been moved
; Find alignment of end of src modulo 16 at this point:
lea eax, [rsi+rcx]
and eax, 0FH
; Set up for loop moving 32 bytes per iteration:
mov edx, ecx ; Save count
and rcx, -20H ; Round down to nearest multiple of 32
sub edx, ecx ; Remaining data after loop
sub rsi, rax ; Nearest preceding aligned block of src
; Add the same to rsi and rdi as we have subtracted from rcx
add rsi, rdx
add rdi, rdx
; Check if count very big
cmp rcx, [CacheBypassLimit]
ja B1400 ; Use non-temporal store if count > CacheBypassLimit
; Dispatch to different codes depending on src alignment
lea r8, [MAlignmentDispatchSSSE3]
jmp near [r8+rax*8]
B1400: ; Dispatch to different codes depending on src alignment
lea r8, [MAlignmentDispatchNT]
jmp near [r8+rax*8]
align 16
C100: ; Code for aligned src. SSE2 and later CPUs
; The nice case, src and dest have same alignment.
; Loop. rcx has positive index from the beginning, counting down to zero
movaps xmm0, [rsi+rcx-10H]
movaps xmm1, [rsi+rcx-20H]
movaps [rdi+rcx-10H], xmm0
movaps [rdi+rcx-20H], xmm1
sub rcx, 20H
jnz C100
; Move the remaining edx bytes (0 - 31):
; move 16-8-4-2-1 bytes, aligned
test edx, edx
jz C500 ; Early out if no more data
test dl, 10H
jz C200
; move 16 bytes
sub rcx, 10H
movaps xmm0, [rsi+rcx]
movaps [rdi+rcx], xmm0
C200: ; Other branches come in here, rcx may contain arbitrary offset
test edx, edx
jz C500 ; Early out if no more data
test dl, 8
jz C210
; move 8 bytes
sub rcx, 8
mov rax, [rsi+rcx]
mov [rdi+rcx], rax
C210: test dl, 4
jz C220
; move 4 bytes
sub rcx, 4
mov eax, [rsi+rcx]
mov [rdi+rcx], eax
jz C500 ; Early out if count divisible by 4
C220: test dl, 2
jz C230
; move 2 bytes
sub rcx, 2
movzx eax, word [rsi+rcx]
mov [rdi+rcx], ax
C230: test dl, 1
jz C500
; move 1 byte
movzx eax, byte [rsi+rcx-1] ; rcx-1 is not always 0 here
mov [rdi+rcx-1], al
C500: ; finished
RETURNM
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;
; Version for processors with SSE2. Aligned read + shift + aligned write
;
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
memmoveSSE2: ; SSE2 version begins here
memmoveSSE2@: ; local label
PROLOGM memcpySSE2
; Cannot use memcpy. Must move backwards because of overlap between src and dest
cmp rcx, 40H
jae B0100 ; Use simpler code if count < 64
; count < 64. Move 32-16-8-4-2-1 bytes
test cl, 20H
jz A100
; move 32 bytes
; mov is faster than movdqu on SSE2 processors,
; movdqu is faster on later processors
sub ecx, 20H
mov rax, [rsi+rcx+18H]
mov rdx, [rsi+rcx+10H]
mov [rdi+rcx+18H], rax
mov [rdi+rcx+10H], rdx
mov rax, [rsi+rcx+8]
mov rdx, [rsi+rcx]
mov [rdi+rcx+8], rax
mov [rdi+rcx], rdx
A100: test cl, 10H
jz A200
; move 16 bytes
sub ecx, 10H
mov rax, [rsi+rcx+8]
mov rdx, [rsi+rcx]
mov [rdi+rcx+8], rax
mov [rdi+rcx], rdx
A200: test cl, 8
jz A300
; move 8 bytes
sub ecx, 8
mov rax, [rsi+rcx]
mov [rdi+rcx], rax
A300: test cl, 4
jz A400
; move 4 bytes
sub ecx, 4
mov eax, [rsi+rcx]
mov [rdi+rcx], eax
jz A900 ; early out if count divisible by 4
A400: test cl, 2
jz A500
; move 2 bytes
sub ecx, 2
movzx eax, word [rsi+rcx]
mov [rdi+rcx], ax
A500: test cl, 1
jz A900
; move 1 byte
movzx eax, byte [rsi] ; rcx-1 = 0
mov [rdi], al
A900: ; finished
RETURNM
B0100: ; count >= 64
; Note: this part will not always work if count < 64
; Calculate size of last block after last regular boundary of dest
lea edx, [rdi+rcx] ; end of dext
and edx, 0FH
jz B0300 ; Skip if end of dest aligned by 16
; edx = size of last partial block, 1 - 15 bytes
test dl, 3
jz B0210
test dl, 1
jz B0201 ; B0200 if we haven't tested edx,3
; move 1 byte
dec rcx
movzx eax, byte [rsi+rcx]
mov [rdi+rcx], al
B0200: test dl, 2
jz B0210
B0201: ; move 2 bytes
sub rcx, 2
movzx eax, word [rsi+rcx]
mov [rdi+rcx], ax
B0210: test dl, 4
jz B0220
; move 4 bytes
sub rcx, 4
mov eax, [rsi+rcx]
mov [rdi+rcx], eax
B0220: test dl, 8
jz B0300
; move 8 bytes
sub rcx, 8
mov rax, [rsi+rcx]
mov [rdi+rcx], rax
B0300: ; Now end of dest is aligned by 16. Any partial block has been moved
; Find alignment of end of src modulo 16 at this point:
lea eax, [rsi+rcx]
and eax, 0FH
; Set up for loop moving 32 bytes per iteration:
mov edx, ecx ; Save count
and rcx, -20H ; Round down to nearest multiple of 32
sub edx, ecx ; Remaining data after loop
sub rsi, rax ; Nearest preceding aligned block of src
; Add the same to rsi and rdi as we have subtracted from rcx
add rsi, rdx
add rdi, rdx
; Check if count very big
cmp rcx, [CacheBypassLimit]
ja B0400 ; Use non-temporal store if count > CacheBypassLimit
; Dispatch to different codes depending on src alignment
lea r8, [MAlignmentDispatchSSE2]
jmp near [r8+rax*8]
B0400: ; Dispatch to different codes depending on src alignment
lea r8, [MAlignmentDispatchNT]
jmp near [r8+rax*8]
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;
; Macros and alignment jump tables
;
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
; Macros for each src alignment, SSE2 instruction set:
; Make separate code for each alignment u because the shift instructions
; have the shift count as a constant:
%MACRO MOVE_REVERSE_UNALIGNED_SSE2 2 ; u, nt
; Move rcx + rdx bytes of data
; Source is misaligned. (src-dest) modulo 16 = %1
; %2 = 1 if non-temporal store desired
; eax = %1
; rsi = src - %1 = nearest preceding 16-bytes boundary
; rdi = dest (aligned)
; rcx = count rounded down to nearest divisible by 32
; edx = remaining bytes to move after loop
movdqa xmm0, [rsi+rcx] ; Read from nearest following 16B boundary
%%L1: ; Loop. rcx has positive index from the beginning, counting down to zero
sub rcx, 20H
movdqa xmm1, [rsi+rcx+10H] ; Read next two blocks aligned
movdqa xmm2, [rsi+rcx]
movdqa xmm3, xmm1 ; Copy because used twice
pslldq xmm0, 16-%1 ; shift left
psrldq xmm1, %1 ; shift right
por xmm0, xmm1 ; combine blocks
%IF %2 == 0
movdqa [rdi+rcx+10H], xmm0 ; Save aligned
%ELSE
movntdq [rdi+rcx+10H], xmm0 ; Save aligned
%ENDIF
movdqa xmm0, xmm2 ; Save for next iteration
pslldq xmm3, 16-%1 ; shift left
psrldq xmm2, %1 ; shift right
por xmm3, xmm2 ; combine blocks
%IF %2 == 0
movdqa [rdi+rcx], xmm3 ; Save aligned
%ELSE
movntdq [rdi+rcx], xmm3 ; Save aligned
%ENDIF
jnz %%L1
; Move edx remaining bytes
test dl, 10H
jz %%L2
; One more 16-bytes block to move
sub rcx, 10H
movdqa xmm1, [rsi+rcx]
pslldq xmm0, 16-%1 ; shift left
psrldq xmm1, %1 ; shift right
por xmm0, xmm1 ; combine blocks
%IF %2 == 0
movdqa [rdi+rcx], xmm0 ; Save aligned
%ELSE
movntdq [rdi+rcx], xmm0 ; Save aligned
%ENDIF
%%L2: ; Get src pointer back to misaligned state
add rsi, rax
; Move remaining 0 - 15 bytes, unaligned
jmp C200
%ENDMACRO
%MACRO MOVE_REVERSE_UNALIGNED_SSE2_4 1 ; nt
; Special case: u = 4
; %1 = 1 if non-temporal store desired
movaps xmm0, [rsi+rcx] ; Read from nearest following 16B boundary
%%L1: ; Loop. rcx has positive index from the beginning, counting down to zero
sub rcx, 20H
movaps xmm1, [rsi+rcx+10H] ; Read next two blocks aligned
movaps xmm2, [rsi+rcx]
movaps xmm3, xmm0
movaps xmm0, xmm2
movss xmm2, xmm1
shufps xmm2, xmm2, 00111001B ; Rotate right
movss xmm1, xmm3
shufps xmm1, xmm1, 00111001B ; Rotate right
%IF %1 == 0
movaps [rdi+rcx+10H], xmm1 ; Save aligned
movaps [rdi+rcx], xmm2 ; Save aligned
%ELSE
movntps [rdi+rcx+10H], xmm1 ; Non-temporal save
movntps [rdi+rcx], xmm2 ; Non-temporal save
%ENDIF
jnz %%L1
; Move edx remaining bytes
test dl, 10H
jz %%L2
; One more 16-bytes block to move
sub rcx, 10H
movaps xmm1, [rsi+rcx]
movss xmm1, xmm0
shufps xmm1, xmm1, 00111001B ; Rotate right
%IF %1 == 0
movaps [rdi+rcx], xmm1 ; Save aligned
%ELSE
movntps [rdi+rcx], xmm1 ; Non-temporal save
%ENDIF
%%L2: ; Get src pointer back to misaligned state
add rsi, rax
; Move remaining 0 - 15 bytes, unaligned
jmp C200
%ENDMACRO
%MACRO MOVE_REVERSE_UNALIGNED_SSE2_8 1 ; nt
; Special case: u = 8
; %1 = 1 if non-temporal store desired
movaps xmm0, [rsi+rcx] ; Read from nearest following 16B boundary
shufps xmm0, xmm0, 01001110B ; Rotate
%%L1: ; Loop. rcx has positive index from the beginning, counting down to zero
sub rcx, 20H
movaps xmm1, [rsi+rcx+10H] ; Read next two blocks aligned
shufps xmm1, xmm1, 01001110B ; Rotate
movsd xmm0, xmm1
%IF %1 == 0
movaps [rdi+rcx+10H], xmm0 ; Save aligned
%ELSE
movntps [rdi+rcx+10H], xmm0 ; Non-temporal save
%ENDIF
movaps xmm0, [rsi+rcx]
shufps xmm0, xmm0, 01001110B ; Rotate
movsd xmm1, xmm0
%IF %1 == 0
movaps [rdi+rcx], xmm1 ; Save aligned
%ELSE
movntps [rdi+rcx], xmm1 ; Non-temporal save
%ENDIF
jnz %%L1
; Move edx remaining bytes
test dl, 10H
jz %%L2
; One more 16-bytes block to move
sub rcx, 10H
movaps xmm1, [rsi+rcx]
shufps xmm1, xmm1, 01001110B ; Rotate
movsd xmm0, xmm1
%IF %1 == 0
movaps [rdi+rcx], xmm0 ; Save aligned
%ELSE
movntps [rdi+rcx], xmm0 ; Non-temporal save
%ENDIF
%%L2: ; Get src pointer back to misaligned state
add rsi, rax
; Move remaining 0 - 15 bytes, unaligned
jmp C200
%ENDMACRO
%MACRO MOVE_REVERSE_UNALIGNED_SSE2_12 1 ; nt
; Special case: u = 12
; %1 = 1 if non-temporal store desired
movaps xmm0, [rsi+rcx] ; Read from nearest following 16B boundary
shufps xmm0, xmm0, 10010011B ; Rotate right
%%L1: ; Loop. rcx has positive index from the beginning, counting down to zero
sub rcx, 20H
movaps xmm1, [rsi+rcx+10H] ; Read next two blocks aligned
shufps xmm1, xmm1, 10010011B ; Rotate left
movss xmm0, xmm1
%IF %1 == 0
movaps [rdi+rcx+10H], xmm0 ; Save aligned
%ELSE
movntps [rdi+rcx+10H], xmm0 ; Non-temporal save
%ENDIF
movaps xmm0, [rsi+rcx]
shufps xmm0, xmm0, 10010011B ; Rotate left
movss xmm1, xmm0
%IF %1 == 0
movaps [rdi+rcx], xmm1 ; Save aligned
%ELSE
movntps [rdi+rcx], xmm1 ; Non-temporal save
%ENDIF
jnz %%L1
; Move edx remaining bytes
test dl, 10H
jz %%L2
; One more 16-bytes block to move
sub rcx, 10H
movaps xmm1, [rsi+rcx]
shufps xmm1, xmm1, 10010011B ; Rotate left
movss xmm0, xmm1
%IF %1 == 0
movaps [rdi+rcx], xmm0 ; Save aligned
%ELSE
movntps [rdi+rcx], xmm0 ; Non-temporal save
%ENDIF
%%L2: ; Get src pointer back to misaligned state
add rsi, rax
; Move remaining 0 - 15 bytes, unaligned
jmp C200
%ENDMACRO
; Macros for each src alignment, Suppl.SSE3 instruction set:
; Code for unaligned src, Suppl.SSE3 instruction set.
; Make separate code for each alignment u because the palignr instruction
; has the shift count as a constant:
%MACRO MOVE_REVERSE_UNALIGNED_SSSE3 1; u
; Move rcx + rdx bytes of data
; Source is misaligned. (src-dest) modulo 16 = %1
; eax = %1
; rsi = src - %1 = nearest preceding 16-bytes boundary
; rdi = dest (aligned)
; rcx = - (count rounded down to nearest divisible by 32)
; edx = remaining bytes to move after loop
movdqa xmm0, [rsi+rcx] ; Read from nearest following 16B boundary
%%L1: ; Loop. rcx has positive index from the beginning, counting down to zero
movdqa xmm1, [rsi+rcx-10H] ; Read next two blocks
palignr xmm0, xmm1, %1 ; Combine parts into aligned block
movdqa [rdi+rcx-10H], xmm0 ; Save aligned
movdqa xmm0, [rsi+rcx-20H]
palignr xmm1, xmm0, %1 ; Combine parts into aligned block
movdqa [rdi+rcx-20H], xmm1 ; Save aligned
sub rcx, 20H
jnz %%L1
; Set up for edx remaining bytes
test dl, 10H
jz %%L2
; One more 16-bytes block to move
sub rcx, 10H
movdqa xmm1, [rsi+rcx] ; Read next two blocks
palignr xmm0, xmm1, %1 ; Combine parts into aligned block
movdqa [rdi+rcx], xmm0 ; Save aligned
%%L2: ; Get src pointer back to misaligned state
add rsi, rax
; Move remaining 0 - 15 bytes
jmp C200
%ENDMACRO
; Make 15 instances of SSE2 macro for each value of the alignment u.
; These are pointed to by the jump table MAlignmentDispatchSSE2 below
; (aligns and fillers are inserted manually to minimize the
; number of 16-bytes boundaries inside loops)
align 16
D104: MOVE_REVERSE_UNALIGNED_SSE2_4 0
D108: MOVE_REVERSE_UNALIGNED_SSE2_8 0
D10C: MOVE_REVERSE_UNALIGNED_SSE2_12 0
D101: MOVE_REVERSE_UNALIGNED_SSE2 1, 0
D102: MOVE_REVERSE_UNALIGNED_SSE2 2, 0
D103: MOVE_REVERSE_UNALIGNED_SSE2 3, 0
D105: MOVE_REVERSE_UNALIGNED_SSE2 5, 0
D106: MOVE_REVERSE_UNALIGNED_SSE2 6, 0
D107: MOVE_REVERSE_UNALIGNED_SSE2 7, 0
D109: MOVE_REVERSE_UNALIGNED_SSE2 9, 0
D10A: MOVE_REVERSE_UNALIGNED_SSE2 0AH, 0
D10B: MOVE_REVERSE_UNALIGNED_SSE2 0BH, 0
D10D: MOVE_REVERSE_UNALIGNED_SSE2 0DH, 0
D10E: MOVE_REVERSE_UNALIGNED_SSE2 0EH, 0
D10F: MOVE_REVERSE_UNALIGNED_SSE2 0FH, 0
; Make 15 instances of Suppl-SSE3 macro for each value of the alignment u.
; These are pointed to by the jump table MAlignmentDispatchSupSSE3 below
align 16
E104: MOVE_REVERSE_UNALIGNED_SSSE3 4
E108: MOVE_REVERSE_UNALIGNED_SSSE3 8
E10C: MOVE_REVERSE_UNALIGNED_SSSE3 0CH
E101: MOVE_REVERSE_UNALIGNED_SSSE3 1
E102: MOVE_REVERSE_UNALIGNED_SSSE3 2
E103: MOVE_REVERSE_UNALIGNED_SSSE3 3
E105: MOVE_REVERSE_UNALIGNED_SSSE3 5
E106: MOVE_REVERSE_UNALIGNED_SSSE3 6
E107: MOVE_REVERSE_UNALIGNED_SSSE3 7
E109: MOVE_REVERSE_UNALIGNED_SSSE3 9
E10A: MOVE_REVERSE_UNALIGNED_SSSE3 0AH
E10B: MOVE_REVERSE_UNALIGNED_SSSE3 0BH
E10D: MOVE_REVERSE_UNALIGNED_SSSE3 0DH
E10E: MOVE_REVERSE_UNALIGNED_SSSE3 0EH
E10F: MOVE_REVERSE_UNALIGNED_SSSE3 0FH
align 16
F100: ; Non-temporal move, src and dest have same alignment.
; Loop. rcx has positive index from the beginning, counting down to zero
sub rcx, 20H
movaps xmm0, [rsi+rcx+10H]
movaps xmm1, [rsi+rcx]
movntps [rdi+rcx+10H], xmm0
movntps [rdi+rcx], xmm1
jnz F100
; Move the remaining edx bytes (0 - 31):
; move 16-8-4-2-1 bytes, aligned
test dl, 10H
jz C200
; move 16 bytes
sub rcx, 10H
movaps xmm0, [rsi+rcx]
movntps [rdi+rcx], xmm0
; move the remaining 0 - 15 bytes
jmp C200
; Non-temporal move, src and dest have different alignment.
; Make 15 instances of SSE2 macro for each value of the alignment u.
; These are pointed to by the jump table MAlignmentDispatchNT below
align 16
F101: MOVE_REVERSE_UNALIGNED_SSE2 1, 1
F102: MOVE_REVERSE_UNALIGNED_SSE2 2, 1
F103: MOVE_REVERSE_UNALIGNED_SSE2 3, 1
F104: MOVE_REVERSE_UNALIGNED_SSE2_4 1
F105: MOVE_REVERSE_UNALIGNED_SSE2 5, 1
F106: MOVE_REVERSE_UNALIGNED_SSE2 6, 1
F107: MOVE_REVERSE_UNALIGNED_SSE2 7, 1
F108: MOVE_REVERSE_UNALIGNED_SSE2_8 1
F109: MOVE_REVERSE_UNALIGNED_SSE2 9, 1
F10A: MOVE_REVERSE_UNALIGNED_SSE2 0AH, 1
F10B: MOVE_REVERSE_UNALIGNED_SSE2 0BH, 1
F10C: MOVE_REVERSE_UNALIGNED_SSE2_12 1
F10D: MOVE_REVERSE_UNALIGNED_SSE2 0DH, 1
F10E: MOVE_REVERSE_UNALIGNED_SSE2 0EH, 1
F10F: MOVE_REVERSE_UNALIGNED_SSE2 0FH, 1
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;
; CPU dispatcher
;
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
memmoveCPUDispatch: ; CPU dispatcher, check for Suppl-SSE3 instruction set
; This part is executed only once
push rbx
push rcx
push rdx
push rsi
push rdi
push r8
; set CacheBypassLimit to half the size of the largest level cache
%ifdef WINDOWS
xor ecx, ecx ; 0 means default
%else
xor edi, edi