Last change
on this file since 690 was
444,
checked in by satin@…, 6 years ago
|
add newlib,libalmos-mkh, restructure shared_syscalls.h and mini-libc
|
File size:
1.7 KB
|
Line | |
---|
1 | /* |
---|
2 | * ==================================================== |
---|
3 | * Copyright (C) 2007 by Ellips BV. All rights reserved. |
---|
4 | * |
---|
5 | * Permission to use, copy, modify, and distribute this |
---|
6 | * software is freely granted, provided that this notice |
---|
7 | * is preserved. |
---|
8 | * ==================================================== |
---|
9 | */ |
---|
10 | |
---|
11 | #include "x86_64mach.h" |
---|
12 | |
---|
13 | .global SYM (memset) |
---|
14 | SOTYPE_FUNCTION(memset) |
---|
15 | |
---|
16 | SYM (memset): |
---|
17 | movq rdi, r9 /* Save return value */ |
---|
18 | movq rsi, rax |
---|
19 | movq rdx, rcx |
---|
20 | cmpq $16, rdx |
---|
21 | jb byte_set |
---|
22 | |
---|
23 | movq rdi, r8 /* Align on quad word boundary */ |
---|
24 | andq $7, r8 |
---|
25 | jz quadword_aligned |
---|
26 | movq $8, rcx |
---|
27 | subq r8, rcx |
---|
28 | subq rcx, rdx |
---|
29 | rep stosb |
---|
30 | movq rdx, rcx |
---|
31 | |
---|
32 | quadword_aligned: |
---|
33 | movabs $0x0101010101010101, r8 |
---|
34 | movzbl sil, eax |
---|
35 | imul r8, rax |
---|
36 | cmpq $256, rdx |
---|
37 | jb quadword_set |
---|
38 | |
---|
39 | shrq $7, rcx /* Store 128 bytes at a time with minimum cache polution */ |
---|
40 | |
---|
41 | .p2align 4 |
---|
42 | loop: |
---|
43 | movntiq rax, (rdi) |
---|
44 | movntiq rax, 8 (rdi) |
---|
45 | movntiq rax, 16 (rdi) |
---|
46 | movntiq rax, 24 (rdi) |
---|
47 | movntiq rax, 32 (rdi) |
---|
48 | movntiq rax, 40 (rdi) |
---|
49 | movntiq rax, 48 (rdi) |
---|
50 | movntiq rax, 56 (rdi) |
---|
51 | movntiq rax, 64 (rdi) |
---|
52 | movntiq rax, 72 (rdi) |
---|
53 | movntiq rax, 80 (rdi) |
---|
54 | movntiq rax, 88 (rdi) |
---|
55 | movntiq rax, 96 (rdi) |
---|
56 | movntiq rax, 104 (rdi) |
---|
57 | movntiq rax, 112 (rdi) |
---|
58 | movntiq rax, 120 (rdi) |
---|
59 | |
---|
60 | leaq 128 (rdi), rdi |
---|
61 | |
---|
62 | dec rcx |
---|
63 | jnz loop |
---|
64 | |
---|
65 | sfence |
---|
66 | movq rdx, rcx |
---|
67 | andq $127, rcx |
---|
68 | rep stosb |
---|
69 | movq r9, rax |
---|
70 | ret |
---|
71 | |
---|
72 | |
---|
73 | byte_set: |
---|
74 | rep stosb |
---|
75 | movq r9, rax |
---|
76 | ret |
---|
77 | |
---|
78 | |
---|
79 | quadword_set: |
---|
80 | shrq $3, rcx |
---|
81 | .p2align 4 |
---|
82 | rep stosq |
---|
83 | movq rdx, rcx |
---|
84 | andq $7, rcx |
---|
85 | rep stosb /* Store the remaining bytes */ |
---|
86 | movq r9, rax |
---|
87 | ret |
---|
88 | |
---|
Note: See
TracBrowser
for help on using the repository browser.