linux/arch/x86/lib/memcpy_64.S
Ma Ling 59daa706fb x86, mem: Optimize memcpy by avoiding memory false dependece
All read operations after allocation stage can run speculatively,
all write operation will run in program order, and if addresses are
different read may run before older write operation, otherwise wait
until write commit. However CPU don't check each address bit,
so read could fail to recognize different address even they
are in different page.For example if rsi is 0xf004, rdi is 0xe008,
in following operation there will generate big performance latency.
1. movq (%rsi),	%rax
2. movq %rax,	(%rdi)
3. movq 8(%rsi), %rax
4. movq %rax,	8(%rdi)

If %rsi and rdi were in really the same meory page, there are TRUE
read-after-write dependence because instruction 2 write 0x008 and
instruction 3 read 0x00c, the two address are overlap partially.
Actually there are in different page and no any issues,
but without checking each address bit CPU could think they are
in the same page, and instruction 3 have to wait for instruction 2
to write data into cache from write buffer, then load data from cache,
the cost time read spent is equal to mfence instruction. We may avoid it by
tuning operation sequence as follow.

1. movq 8(%rsi), %rax
2. movq %rax,	8(%rdi)
3. movq (%rsi),	%rax
4. movq %rax,	(%rdi)

Instruction 3 read 0x004, instruction 2 write address 0x010, no any
dependence.  At last on Core2 we gain 1.83x speedup compared with
original instruction sequence.  In this patch we first handle small
size(less 20bytes), then jump to different copy mode. Based on our
micro-benchmark small bytes from 1 to 127 bytes, we got up to 2X
improvement, and up to 1.5X improvement for 1024 bytes on Corei7.  (We
use our micro-benchmark, and will do further test according to your
requirment)

Signed-off-by: Ma Ling <ling.ma@intel.com>
LKML-Reference: <1277753065-18610-1-git-send-email-ling.ma@intel.com>
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
2010-08-23 14:56:41 -07:00

192 lines
3.3 KiB
ArmAsm

/* Copyright 2002 Andi Kleen */
#include <linux/linkage.h>
#include <asm/cpufeature.h>
#include <asm/dwarf2.h>
/*
* memcpy - Copy a memory block.
*
* Input:
* rdi destination
* rsi source
* rdx count
*
* Output:
* rax original destination
*/
/*
* memcpy_c() - fast string ops (REP MOVSQ) based variant.
*
* This gets patched over the unrolled variant (below) via the
* alternative instructions framework:
*/
.section .altinstr_replacement, "ax", @progbits
.Lmemcpy_c:
movq %rdi, %rax
movl %edx, %ecx
shrl $3, %ecx
andl $7, %edx
rep movsq
movl %edx, %ecx
rep movsb
ret
.Lmemcpy_e:
.previous
ENTRY(__memcpy)
ENTRY(memcpy)
CFI_STARTPROC
movq %rdi, %rax
/*
* Use 32bit CMP here to avoid long NOP padding.
*/
cmp $0x20, %edx
jb .Lhandle_tail
/*
* We check whether memory false dependece could occur,
* then jump to corresponding copy mode.
*/
cmp %dil, %sil
jl .Lcopy_backward
subl $0x20, %edx
.Lcopy_forward_loop:
subq $0x20, %rdx
/*
* Move in blocks of 4x8 bytes:
*/
movq 0*8(%rsi), %r8
movq 1*8(%rsi), %r9
movq 2*8(%rsi), %r10
movq 3*8(%rsi), %r11
leaq 4*8(%rsi), %rsi
movq %r8, 0*8(%rdi)
movq %r9, 1*8(%rdi)
movq %r10, 2*8(%rdi)
movq %r11, 3*8(%rdi)
leaq 4*8(%rdi), %rdi
jae .Lcopy_forward_loop
addq $0x20, %rdx
jmp .Lhandle_tail
.Lcopy_backward:
/*
* Calculate copy position to tail.
*/
addq %rdx, %rsi
addq %rdx, %rdi
subq $0x20, %rdx
/*
* At most 3 ALU operations in one cycle,
* so append NOPS in the same 16bytes trunk.
*/
.p2align 4
.Lcopy_backward_loop:
subq $0x20, %rdx
movq -1*8(%rsi), %r8
movq -2*8(%rsi), %r9
movq -3*8(%rsi), %r10
movq -4*8(%rsi), %r11
leaq -4*8(%rsi), %rsi
movq %r8, -1*8(%rdi)
movq %r9, -2*8(%rdi)
movq %r10, -3*8(%rdi)
movq %r11, -4*8(%rdi)
leaq -4*8(%rdi), %rdi
jae .Lcopy_backward_loop
/*
* Calculate copy position to head.
*/
addq $0x20, %rdx
subq %rdx, %rsi
subq %rdx, %rdi
.Lhandle_tail:
cmpq $16, %rdx
jb .Lless_16bytes
/*
* Move data from 16 bytes to 31 bytes.
*/
movq 0*8(%rsi), %r8
movq 1*8(%rsi), %r9
movq -2*8(%rsi, %rdx), %r10
movq -1*8(%rsi, %rdx), %r11
movq %r8, 0*8(%rdi)
movq %r9, 1*8(%rdi)
movq %r10, -2*8(%rdi, %rdx)
movq %r11, -1*8(%rdi, %rdx)
retq
.p2align 4
.Lless_16bytes:
cmpq $8, %rdx
jb .Lless_8bytes
/*
* Move data from 8 bytes to 15 bytes.
*/
movq 0*8(%rsi), %r8
movq -1*8(%rsi, %rdx), %r9
movq %r8, 0*8(%rdi)
movq %r9, -1*8(%rdi, %rdx)
retq
.p2align 4
.Lless_8bytes:
cmpq $4, %rdx
jb .Lless_3bytes
/*
* Move data from 4 bytes to 7 bytes.
*/
movl (%rsi), %ecx
movl -4(%rsi, %rdx), %r8d
movl %ecx, (%rdi)
movl %r8d, -4(%rdi, %rdx)
retq
.p2align 4
.Lless_3bytes:
cmpl $0, %edx
je .Lend
/*
* Move data from 1 bytes to 3 bytes.
*/
.Lloop_1:
movb (%rsi), %r8b
movb %r8b, (%rdi)
incq %rdi
incq %rsi
decl %edx
jnz .Lloop_1
.Lend:
retq
CFI_ENDPROC
ENDPROC(memcpy)
ENDPROC(__memcpy)
/*
* Some CPUs run faster using the string copy instructions.
* It is also a lot simpler. Use this when possible:
*/
.section .altinstructions, "a"
.align 8
.quad memcpy
.quad .Lmemcpy_c
.word X86_FEATURE_REP_GOOD
/*
* Replace only beginning, memcpy is used to apply alternatives,
* so it is silly to overwrite itself with nops - reboot is the
* only outcome...
*/
.byte .Lmemcpy_e - .Lmemcpy_c
.byte .Lmemcpy_e - .Lmemcpy_c
.previous