Attachment #5264 for bug #12518




#ifdef USE_AS_MEMPCPY
	add	%rdx, %rax
#endif
	/*
	 * The small cases we can do without checking for any
	 * overlap at all, since we do them as all loads followed
	 * by all stores.
	 *
	 * So just jump through the less-than-80bytes table.
	 */
	cmp	$79,%rdx
	lea	L(table_less_80bytes)(%rip), %r11
#endif
	cmp	$79, %rdx
	lea     L(table_less_80bytes)(%rip), %r11
	ja	L(80bytesormore)
	movslq	(%r11, %rdx, 4), %r9
	add	%rdx, %rsi

	jmp	*%r9
	ud2

	/*
	 * For the 80+ byte cases we need to check overlap
	 */
L(80bytesormore):
	lea	(%rsi,%rdx),%r9
	lea	(%rdi,%rdx),%r11
	cmp	%rdi,%r9		/* dest start >= source end */
	jae	L(nonoverlap)		/*  -> nonoverlapping */
	cmp	%rsi,%r11		/* source start >= destination end */
	jae	L(nonoverlap)		/*  -> nonoverlapping */
	cmp	%rsi, %rdi		/* overlap: */
	jb	L(copy_forward)		/* source < dest: forward copy */
	je	L(write_0bytes)		/* source == dest: no copy */
	jmp	L(copy_backward)	/* source > dest: backward copy */
L(nonoverlap):
	cmp	%dil, %sil
	jle	L(copy_backward)
L(copy_forward):

	movdqu	(%rsi), %xmm0
	mov	%rdi, %rcx
	and	$-16, %rdi

	movntdq	%xmm1, (%rdi)
	lea	16(%rdi), %rdi
	lea	-0x90(%rdx), %rdx
#ifdef USE_AS_MEMMOVE
	mov	%rsi, %r9
	sub	%rdi, %r9
	cmp	%rdx, %r9

	cmp	%rcx, %rdx
	jb	L(ll_cache_copy_fwd_start)
L(memmove_is_memcpy_fwd):
#endif
L(large_page_loop):
	movdqu	(%rsi), %xmm0
	movdqu	0x10(%rsi), %xmm1

	sfence
	BRANCH_TO_JMPTBL_ENTRY (L(table_less_80bytes), %rdx, 4)

#ifdef USE_AS_MEMMOVE
	ALIGN (4)
L(ll_cache_copy_fwd_start):
	prefetcht0 0x1c0(%rsi)

	add	%rdx, %rdi
	BRANCH_TO_JMPTBL_ENTRY (L(table_less_80bytes), %rdx, 4)

#endif
	ALIGN (4)
L(large_page_bwd):
	movdqu	-0x10(%rsi), %xmm1

	movdqa	%xmm1, -0x10(%rdi)
	lea	-16(%rdi), %rdi
	lea	-0x90(%rdx), %rdx
#ifdef USE_AS_MEMMOVE
	mov	%rdi, %r9
	sub	%rsi, %r9
	cmp	%rdx, %r9

	cmp	%rcx, %r9
	jb	L(ll_cache_copy_bwd_start)
L(memmove_is_memcpy_bwd):
#endif
L(large_page_bwd_loop):
	movdqu	-0x10(%rsi), %xmm0
	movdqu	-0x20(%rsi), %xmm1

	sfence
	BRANCH_TO_JMPTBL_ENTRY (L(table_less_80bytes), %rdx, 4)

#ifdef USE_AS_MEMMOVE
	ALIGN (4)
L(ll_cache_copy_bwd_start):
	prefetcht0 -0x1c0(%rsi)

	sub	$0x40, %rdx
L(large_page_ll_less_bwd_64bytes):
	BRANCH_TO_JMPTBL_ENTRY (L(table_less_80bytes), %rdx, 4)
#endif

END (MEMCPY)


(-)a/sysdeps/x86_64/multiarch/memmove-ssse3.S (-4 / +4 lines)

Lines 1-4	Link Here

#define USE_AS_MEMMOVE

#include <sysdep.h>

#define MEMCPY		__memmove_ssse3

#define MEMCPY_CHK	__memmove_chk_ssse3

strong_alias(__memmove_ssse3, __memcpy_ssse3);

#include "memcpy-ssse3.S"

strong_alias(__memmove_chk_ssse3, __memcpy_chk_ssse3);

Return to bug 12518