blob: 0ebb03b60e797946ff5aae6952b7e94fae5d5d25 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/* Written 2003 by Andi Kleen, based on a kernel by Evandro Menezes */
2
Jan Beulich8d379da2006-09-26 10:52:32 +02003#include <linux/config.h>
4#include <linux/linkage.h>
5#include <asm/dwarf2.h>
6
7 ALIGN
8copy_page_c:
9 CFI_STARTPROC
10 movl $4096/8,%ecx
11 rep movsq
12 ret
13 CFI_ENDPROC
14ENDPROC(copy_page_c)
15
Linus Torvalds1da177e2005-04-16 15:20:36 -070016/* Don't use streaming store because it's better when the target
17 ends up in cache. */
18
19/* Could vary the prefetch distance based on SMP/UP */
20
Jan Beulich8d379da2006-09-26 10:52:32 +020021ENTRY(copy_page)
22 CFI_STARTPROC
Andi Kleen7bcd3f32006-02-03 21:51:02 +010023 subq $3*8,%rsp
Jan Beulich8d379da2006-09-26 10:52:32 +020024 CFI_ADJUST_CFA_OFFSET 3*8
Andi Kleen7bcd3f32006-02-03 21:51:02 +010025 movq %rbx,(%rsp)
Jan Beulich8d379da2006-09-26 10:52:32 +020026 CFI_REL_OFFSET rbx, 0
Andi Kleen7bcd3f32006-02-03 21:51:02 +010027 movq %r12,1*8(%rsp)
Jan Beulich8d379da2006-09-26 10:52:32 +020028 CFI_REL_OFFSET r12, 1*8
Andi Kleen7bcd3f32006-02-03 21:51:02 +010029 movq %r13,2*8(%rsp)
Jan Beulich8d379da2006-09-26 10:52:32 +020030 CFI_REL_OFFSET r13, 2*8
Andi Kleen7bcd3f32006-02-03 21:51:02 +010031
32 movl $(4096/64)-5,%ecx
33 .p2align 4
34.Loop64:
35 dec %rcx
36
37 movq (%rsi), %rax
38 movq 8 (%rsi), %rbx
39 movq 16 (%rsi), %rdx
40 movq 24 (%rsi), %r8
41 movq 32 (%rsi), %r9
42 movq 40 (%rsi), %r10
43 movq 48 (%rsi), %r11
44 movq 56 (%rsi), %r12
45
46 prefetcht0 5*64(%rsi)
47
48 movq %rax, (%rdi)
49 movq %rbx, 8 (%rdi)
50 movq %rdx, 16 (%rdi)
51 movq %r8, 24 (%rdi)
52 movq %r9, 32 (%rdi)
53 movq %r10, 40 (%rdi)
54 movq %r11, 48 (%rdi)
55 movq %r12, 56 (%rdi)
56
57 leaq 64 (%rsi), %rsi
58 leaq 64 (%rdi), %rdi
59
60 jnz .Loop64
61
62 movl $5,%ecx
63 .p2align 4
64.Loop2:
65 decl %ecx
66
67 movq (%rsi), %rax
68 movq 8 (%rsi), %rbx
69 movq 16 (%rsi), %rdx
70 movq 24 (%rsi), %r8
71 movq 32 (%rsi), %r9
72 movq 40 (%rsi), %r10
73 movq 48 (%rsi), %r11
74 movq 56 (%rsi), %r12
75
76 movq %rax, (%rdi)
77 movq %rbx, 8 (%rdi)
78 movq %rdx, 16 (%rdi)
79 movq %r8, 24 (%rdi)
80 movq %r9, 32 (%rdi)
81 movq %r10, 40 (%rdi)
82 movq %r11, 48 (%rdi)
83 movq %r12, 56 (%rdi)
84
85 leaq 64(%rdi),%rdi
86 leaq 64(%rsi),%rsi
87
88 jnz .Loop2
89
90 movq (%rsp),%rbx
Jan Beulich8d379da2006-09-26 10:52:32 +020091 CFI_RESTORE rbx
Andi Kleen7bcd3f32006-02-03 21:51:02 +010092 movq 1*8(%rsp),%r12
Jan Beulich8d379da2006-09-26 10:52:32 +020093 CFI_RESTORE r12
Andi Kleen7bcd3f32006-02-03 21:51:02 +010094 movq 2*8(%rsp),%r13
Jan Beulich8d379da2006-09-26 10:52:32 +020095 CFI_RESTORE r13
Andi Kleen7bcd3f32006-02-03 21:51:02 +010096 addq $3*8,%rsp
Jan Beulich8d379da2006-09-26 10:52:32 +020097 CFI_ADJUST_CFA_OFFSET -3*8
Andi Kleen7bcd3f32006-02-03 21:51:02 +010098 ret
Jan Beulich8d379da2006-09-26 10:52:32 +020099.Lcopy_page_end:
100 CFI_ENDPROC
101ENDPROC(copy_page)
Andi Kleen7bcd3f32006-02-03 21:51:02 +0100102
103 /* Some CPUs run faster using the string copy instructions.
104 It is also a lot simpler. Use this when possible */
105
106#include <asm/cpufeature.h>
107
Jan Beulich8d379da2006-09-26 10:52:32 +0200108 .section .altinstr_replacement,"ax"
1091: .byte 0xeb /* jmp <disp8> */
110 .byte (copy_page_c - copy_page) - (2f - 1b) /* offset */
1112:
112 .previous
Andi Kleen7bcd3f32006-02-03 21:51:02 +0100113 .section .altinstructions,"a"
114 .align 8
Jan Beulich8d379da2006-09-26 10:52:32 +0200115 .quad copy_page
116 .quad 1b
117 .byte X86_FEATURE_REP_GOOD
118 .byte .Lcopy_page_end - copy_page
119 .byte 2b - 1b
Andi Kleen7bcd3f32006-02-03 21:51:02 +0100120 .previous