blob: 176cca67212b7072687f42450ef56018c03ebaaa [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/* Written 2003 by Andi Kleen, based on a kernel by Evandro Menezes */
Dave Jones038b0a62006-10-04 03:38:54 -04002
Jan Beulich8d379da2006-09-26 10:52:32 +02003#include <linux/linkage.h>
4#include <asm/dwarf2.h>
Andy Lutomirski59e97e42011-07-13 09:24:10 -04005#include <asm/alternative-asm.h>
Jan Beulich8d379da2006-09-26 10:52:32 +02006
7 ALIGN
Ma Ling269833b2012-10-18 03:52:45 +08008copy_page_rep:
Jan Beulich8d379da2006-09-26 10:52:32 +02009 CFI_STARTPROC
Ma Ling269833b2012-10-18 03:52:45 +080010 movl $4096/8, %ecx
11 rep movsq
Jan Beulich8d379da2006-09-26 10:52:32 +020012 ret
13 CFI_ENDPROC
Ma Ling269833b2012-10-18 03:52:45 +080014ENDPROC(copy_page_rep)
Jan Beulich8d379da2006-09-26 10:52:32 +020015
Ma Ling269833b2012-10-18 03:52:45 +080016/*
17 * Don't use streaming copy unless the CPU indicates X86_FEATURE_REP_GOOD.
18 * Could vary the prefetch distance based on SMP/UP.
19*/
Linus Torvalds1da177e2005-04-16 15:20:36 -070020
Jan Beulich8d379da2006-09-26 10:52:32 +020021ENTRY(copy_page)
22 CFI_STARTPROC
Ma Ling269833b2012-10-18 03:52:45 +080023 subq $2*8, %rsp
Jan Beulich42693292012-01-05 16:12:25 +000024 CFI_ADJUST_CFA_OFFSET 2*8
Ma Ling269833b2012-10-18 03:52:45 +080025 movq %rbx, (%rsp)
Jan Beulich8d379da2006-09-26 10:52:32 +020026 CFI_REL_OFFSET rbx, 0
Ma Ling269833b2012-10-18 03:52:45 +080027 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
Ma Ling269833b2012-10-18 03:52:45 +080030 movl $(4096/64)-5, %ecx
Andi Kleen7bcd3f32006-02-03 21:51:02 +010031 .p2align 4
32.Loop64:
Ma Ling269833b2012-10-18 03:52:45 +080033 dec %rcx
34 movq 0x8*0(%rsi), %rax
35 movq 0x8*1(%rsi), %rbx
36 movq 0x8*2(%rsi), %rdx
37 movq 0x8*3(%rsi), %r8
38 movq 0x8*4(%rsi), %r9
39 movq 0x8*5(%rsi), %r10
40 movq 0x8*6(%rsi), %r11
41 movq 0x8*7(%rsi), %r12
Andi Kleen7bcd3f32006-02-03 21:51:02 +010042
43 prefetcht0 5*64(%rsi)
44
Ma Ling269833b2012-10-18 03:52:45 +080045 movq %rax, 0x8*0(%rdi)
46 movq %rbx, 0x8*1(%rdi)
47 movq %rdx, 0x8*2(%rdi)
48 movq %r8, 0x8*3(%rdi)
49 movq %r9, 0x8*4(%rdi)
50 movq %r10, 0x8*5(%rdi)
51 movq %r11, 0x8*6(%rdi)
52 movq %r12, 0x8*7(%rdi)
Andi Kleen7bcd3f32006-02-03 21:51:02 +010053
Ma Ling269833b2012-10-18 03:52:45 +080054 leaq 64 (%rsi), %rsi
55 leaq 64 (%rdi), %rdi
Andi Kleen7bcd3f32006-02-03 21:51:02 +010056
Ma Ling269833b2012-10-18 03:52:45 +080057 jnz .Loop64
Andi Kleen7bcd3f32006-02-03 21:51:02 +010058
Ma Ling269833b2012-10-18 03:52:45 +080059 movl $5, %ecx
Andi Kleen7bcd3f32006-02-03 21:51:02 +010060 .p2align 4
61.Loop2:
Ma Ling269833b2012-10-18 03:52:45 +080062 decl %ecx
Andi Kleen7bcd3f32006-02-03 21:51:02 +010063
Ma Ling269833b2012-10-18 03:52:45 +080064 movq 0x8*0(%rsi), %rax
65 movq 0x8*1(%rsi), %rbx
66 movq 0x8*2(%rsi), %rdx
67 movq 0x8*3(%rsi), %r8
68 movq 0x8*4(%rsi), %r9
69 movq 0x8*5(%rsi), %r10
70 movq 0x8*6(%rsi), %r11
71 movq 0x8*7(%rsi), %r12
Andi Kleen7bcd3f32006-02-03 21:51:02 +010072
Ma Ling269833b2012-10-18 03:52:45 +080073 movq %rax, 0x8*0(%rdi)
74 movq %rbx, 0x8*1(%rdi)
75 movq %rdx, 0x8*2(%rdi)
76 movq %r8, 0x8*3(%rdi)
77 movq %r9, 0x8*4(%rdi)
78 movq %r10, 0x8*5(%rdi)
79 movq %r11, 0x8*6(%rdi)
80 movq %r12, 0x8*7(%rdi)
Andi Kleen7bcd3f32006-02-03 21:51:02 +010081
Ma Ling269833b2012-10-18 03:52:45 +080082 leaq 64(%rdi), %rdi
83 leaq 64(%rsi), %rsi
Andi Kleen7bcd3f32006-02-03 21:51:02 +010084 jnz .Loop2
85
Ma Ling269833b2012-10-18 03:52:45 +080086 movq (%rsp), %rbx
Jan Beulich8d379da2006-09-26 10:52:32 +020087 CFI_RESTORE rbx
Ma Ling269833b2012-10-18 03:52:45 +080088 movq 1*8(%rsp), %r12
Jan Beulich8d379da2006-09-26 10:52:32 +020089 CFI_RESTORE r12
Ma Ling269833b2012-10-18 03:52:45 +080090 addq $2*8, %rsp
Jan Beulich42693292012-01-05 16:12:25 +000091 CFI_ADJUST_CFA_OFFSET -2*8
Andi Kleen7bcd3f32006-02-03 21:51:02 +010092 ret
Jan Beulich8d379da2006-09-26 10:52:32 +020093.Lcopy_page_end:
94 CFI_ENDPROC
95ENDPROC(copy_page)
Andi Kleen7bcd3f32006-02-03 21:51:02 +010096
97 /* Some CPUs run faster using the string copy instructions.
98 It is also a lot simpler. Use this when possible */
99
100#include <asm/cpufeature.h>
101
Jan Beulich8d379da2006-09-26 10:52:32 +0200102 .section .altinstr_replacement,"ax"
1031: .byte 0xeb /* jmp <disp8> */
Ma Ling269833b2012-10-18 03:52:45 +0800104 .byte (copy_page_rep - copy_page) - (2f - 1b) /* offset */
Jan Beulich8d379da2006-09-26 10:52:32 +02001052:
106 .previous
Andi Kleen7bcd3f32006-02-03 21:51:02 +0100107 .section .altinstructions,"a"
Andy Lutomirski59e97e42011-07-13 09:24:10 -0400108 altinstruction_entry copy_page, 1b, X86_FEATURE_REP_GOOD, \
109 .Lcopy_page_end-copy_page, 2b-1b
Andi Kleen7bcd3f32006-02-03 21:51:02 +0100110 .previous