Mercurial > pmdwin

annotate memcpy.S @ 6:da588a3fb3cc

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
Make pmdwin a dynamically linked executable.
author Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
date Mon, 08 Sep 2014 17:15:41 +0200
parents c55ea9478c80
children
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
0
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
1 .text
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
2 .global memcpy
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
3 .type memcpy,%function
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
4 memcpy:
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
5 /* Cutoff for the big loop is a size of 64 bytes since otherwise
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
6 the loop will never be entered. */
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
7 cmpq $64, %rdx
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
8 movq %rdx, %rcx
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
9
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
10 /* We need this in any case. */
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
11 cld
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
12 jbe 1f
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
13
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
14 /* Align destination. */
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
15 movq %rdi, %rax
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
16 negq %rax
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
17 andq $15, %rax
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
18 subq %rax, %rcx
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
19 xchgq %rax, %rcx
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
20
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
21 rep; movsb
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
22
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
23 movq %rax, %rcx
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
24 subq $64, %rcx
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
25 js 2f
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
26
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
27 /* Next 3 insns are 11 bytes total, make sure we decode them in one go */
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
28 .p2align 4,,11
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
29 3:
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
30 /* Now correct the loop counter. Please note that in the following
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
31 code the flags are not changed anymore. */
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
32 subq $64, %rcx
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
33
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
34 movups (%rsi), %xmm0
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
35 movups 16(%rsi), %xmm1
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
36 movups 32(%rsi), %xmm2
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
37 movups 48(%rsi), %xmm3
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
38 movaps %xmm0, (%rdi)
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
39 movaps %xmm1, 16(%rdi)
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
40 movaps %xmm2, 32(%rdi)
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
41 movaps %xmm3, 48(%rdi)
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
42
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
43 leaq 64(%rsi), %rsi
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
44 leaq 64(%rdi), %rdi
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
45
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
46 jns 3b
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
47
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
48 /* Correct extra loop counter modification. */
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
49 2: addq $64, %rcx
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
50 1:
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
51 movq %rcx, %r10
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
52 shr $2, %rcx
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
53 rep; movsd
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
54
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
55 movq %r10, %rcx
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
56 andq $3, %rcx
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
57 rep; movsb
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
58
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
59 movq %rdi, %rax /* Set return value. */
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
60 ret
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
61 .size memcpy,.-memcpy
c55ea9478c80 Hello Gensokyo!
Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
parents:
diff changeset
62