kernel_optimize_test/arch/x86/lib/memcpy_64.S
Jan Beulich 9d8e22777e x86-64: Handle byte-wise tail copying in memcpy() without a loop
While hard to measure, reducing the number of possibly/likely
mis-predicted branches can generally be expected to be slightly
better.

Other than apparent at the first glance, this also doesn't grow
the function size (the alignment gap to the next function just
gets smaller).

Signed-off-by: Jan Beulich <jbeulich@suse.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Link: http://lkml.kernel.org/r/4F218584020000780006F422@nat28.tlf.novell.com
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2012-01-26 21:19:20 +01:00

207 lines
4.1 KiB
ArmAsm

/* Copyright 2002 Andi Kleen */
#include <linux/linkage.h>
#include <asm/cpufeature.h>
#include <asm/dwarf2.h>
#include <asm/alternative-asm.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
movq %rdx, %rcx
shrq $3, %rcx
andl $7, %edx
rep movsq
movl %edx, %ecx
rep movsb
ret
.Lmemcpy_e:
.previous
/*
* memcpy_c_e() - enhanced fast string memcpy. This is faster and simpler than
* memcpy_c. Use memcpy_c_e when possible.
*
* This gets patched over the unrolled variant (below) via the
* alternative instructions framework:
*/
.section .altinstr_replacement, "ax", @progbits
.Lmemcpy_c_e:
movq %rdi, %rax
movq %rdx, %rcx
rep movsb
ret
.Lmemcpy_e_e:
.previous
ENTRY(__memcpy)
ENTRY(memcpy)
CFI_STARTPROC
movq %rdi, %rax
cmpq $0x20, %rdx
jb .Lhandle_tail
/*
* We check whether memory false dependence could occur,
* then jump to corresponding copy mode.
*/
cmp %dil, %sil
jl .Lcopy_backward
subq $0x20, %rdx
.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
addl $0x20, %edx
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.
*/
addl $0x20, %edx
subq %rdx, %rsi
subq %rdx, %rdi
.Lhandle_tail:
cmpl $16, %edx
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:
cmpl $8, %edx
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:
cmpl $4, %edx
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:
subl $1, %edx
jb .Lend
/*
* Move data from 1 bytes to 3 bytes.
*/
movzbl (%rsi), %ecx
jz .Lstore_1byte
movzbq 1(%rsi), %r8
movzbq (%rsi, %rdx), %r9
movb %r8b, 1(%rdi)
movb %r9b, (%rdi, %rdx)
.Lstore_1byte:
movb %cl, (%rdi)
.Lend:
retq
CFI_ENDPROC
ENDPROC(memcpy)
ENDPROC(__memcpy)
/*
* Some CPUs are adding enhanced REP MOVSB/STOSB feature
* If the feature is supported, memcpy_c_e() is the first choice.
* If enhanced rep movsb copy is not available, use fast string copy
* memcpy_c() when possible. This is faster and code is simpler than
* original memcpy().
* Otherwise, original memcpy() is used.
* In .altinstructions section, ERMS feature is placed after REG_GOOD
* feature to implement the right patch order.
*
* Replace only beginning, memcpy is used to apply alternatives,
* so it is silly to overwrite itself with nops - reboot is the
* only outcome...
*/
.section .altinstructions, "a"
altinstruction_entry memcpy,.Lmemcpy_c,X86_FEATURE_REP_GOOD,\
.Lmemcpy_e-.Lmemcpy_c,.Lmemcpy_e-.Lmemcpy_c
altinstruction_entry memcpy,.Lmemcpy_c_e,X86_FEATURE_ERMS, \
.Lmemcpy_e_e-.Lmemcpy_c_e,.Lmemcpy_e_e-.Lmemcpy_c_e
.previous