From fdf4289679fd41d76553ce224750e9737cd80eea Mon Sep 17 00:00:00 2001
From: "Ma, Ling" <ling.ma@intel.com>
Date: Mon, 23 Aug 2010 14:11:12 -0700
Subject: x86, mem: Don't implement forward memmove() as memcpy()

memmove() allow source and destination address to be overlap, but
there is no such limitation for memcpy().  Therefore, explicitly
implement memmove() in both the forwards and backward directions, to
give us the ability to optimize memcpy().

Signed-off-by: Ma Ling <ling.ma@intel.com>
LKML-Reference: <C10D3FB0CD45994C8A51FEC1227CE22F0E483AD86A@shsmsx502.ccr.corp.intel.com>
Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
---
 arch/x86/lib/memcpy_32.c  | 38 +++++++++++++++++++++++++++-----------
 arch/x86/lib/memmove_64.c | 46 +++++++++++++++++++++++++++++++++++++++++-----
 2 files changed, 68 insertions(+), 16 deletions(-)

(limited to 'arch')

diff --git a/arch/x86/lib/memcpy_32.c b/arch/x86/lib/memcpy_32.c
index 5415a9d06f5..be424dfcf36 100644
--- a/arch/x86/lib/memcpy_32.c
+++ b/arch/x86/lib/memcpy_32.c
@@ -25,19 +25,35 @@ void *memmove(void *dest, const void *src, size_t n)
 	int d0, d1, d2;
 
 	if (dest < src) {
-		memcpy(dest, src, n);
+		if ((dest + n) < src)
+			 return memcpy(dest, src, n);
+		else
+			__asm__ __volatile__(
+				"rep\n\t"
+				"movsb\n\t"
+				: "=&c" (d0), "=&S" (d1), "=&D" (d2)
+				:"0" (n),
+				 "1" (src),
+				 "2" (dest)
+				:"memory");
+
 	} else {
-		__asm__ __volatile__(
-			"std\n\t"
-			"rep\n\t"
-			"movsb\n\t"
-			"cld"
-			: "=&c" (d0), "=&S" (d1), "=&D" (d2)
-			:"0" (n),
-			 "1" (n-1+src),
-			 "2" (n-1+dest)
-			:"memory");
+
+		if((src + count) < dest)
+			return memcpy(dest, src, count);
+		else
+			__asm__ __volatile__(
+				"std\n\t"
+				"rep\n\t"
+				"movsb\n\t"
+				"cld"
+				: "=&c" (d0), "=&S" (d1), "=&D" (d2)
+				:"0" (n),
+				 "1" (n-1+src),
+				 "2" (n-1+dest)
+				:"memory");
 	}
+
 	return dest;
 }
 EXPORT_SYMBOL(memmove);
diff --git a/arch/x86/lib/memmove_64.c b/arch/x86/lib/memmove_64.c
index 0a33909bf12..ecacc4b3d9e 100644
--- a/arch/x86/lib/memmove_64.c
+++ b/arch/x86/lib/memmove_64.c
@@ -8,13 +8,49 @@
 #undef memmove
 void *memmove(void *dest, const void *src, size_t count)
 {
+	unsigned long d0, d1, d2, d3;
 	if (dest < src) {
-		return memcpy(dest, src, count);
+		if ((dest + count) < src)
+			 return memcpy(dest, src, count);
+		else
+			__asm__ __volatile__(
+				"movq %0, %3\n\t"
+				"shr $3, %0\n\t"
+				"andq $7, %3\n\t"
+				"rep\n\t"
+				"movsq\n\t"
+				"movq %3, %0\n\t"
+				"rep\n\t"
+				"movsb"
+				: "=&c" (d0), "=&S" (d1), "=&D" (d2), "=r" (d3)
+				:"0" (count),
+				 "1" (src),
+				 "2" (dest)
+				:"memory");
 	} else {
-		char *p = dest + count;
-		const char *s = src + count;
-		while (count--)
-			*--p = *--s;
+		if((src + count) < dest)
+			return memcpy(dest, src, count);
+		else
+			__asm__ __volatile__(
+				"movq %0, %3\n\t"
+				"lea -8(%1, %0), %1\n\t"
+				"lea -8(%2, %0), %2\n\t"
+				"shr $3, %0\n\t"
+				"andq $7, %3\n\t"
+				"std\n\t"
+				"rep\n\t"
+				"movsq\n\t"
+				"lea 7(%1), %1\n\t"
+				"lea 7(%2), %2\n\t"
+				"movq %3, %0\n\t"
+				"rep\n\t"
+				"movsb\n\t"
+				"cld"
+				: "=&c" (d0), "=&S" (d1), "=&D" (d2), "=r" (d3)
+				:"0" (count),
+				 "1" (src),
+				 "2" (dest)
+				:"memory");
 	}
 	return dest;
 }
-- 
cgit v1.2.3


From 59daa706fbec745684702741b9f5373142dd9fdc Mon Sep 17 00:00:00 2001
From: Ma Ling <ling.ma@intel.com>
Date: Tue, 29 Jun 2010 03:24:25 +0800
Subject: 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>
---
 arch/x86/lib/memcpy_32.c |   6 +-
 arch/x86/lib/memcpy_64.S | 158 ++++++++++++++++++++++++++++++-----------------
 2 files changed, 105 insertions(+), 59 deletions(-)

(limited to 'arch')

diff --git a/arch/x86/lib/memcpy_32.c b/arch/x86/lib/memcpy_32.c
index be424dfcf36..81130d477ee 100644
--- a/arch/x86/lib/memcpy_32.c
+++ b/arch/x86/lib/memcpy_32.c
@@ -36,11 +36,9 @@ void *memmove(void *dest, const void *src, size_t n)
 				 "1" (src),
 				 "2" (dest)
 				:"memory");
-
 	} else {
-
-		if((src + count) < dest)
-			return memcpy(dest, src, count);
+		if((src + n) < dest)
+			return memcpy(dest, src, n);
 		else
 			__asm__ __volatile__(
 				"std\n\t"
diff --git a/arch/x86/lib/memcpy_64.S b/arch/x86/lib/memcpy_64.S
index bcbcd1e0f7d..75ef61e35e3 100644
--- a/arch/x86/lib/memcpy_64.S
+++ b/arch/x86/lib/memcpy_64.S
@@ -40,84 +40,132 @@
 ENTRY(__memcpy)
 ENTRY(memcpy)
 	CFI_STARTPROC
+	movq %rdi, %rax
 
 	/*
-	 * Put the number of full 64-byte blocks into %ecx.
-	 * Tail portion is handled at the end:
+	 * Use 32bit CMP here to avoid long NOP padding.
 	 */
-	movq %rdi, %rax
-	movl %edx, %ecx
-	shrl   $6, %ecx
-	jz .Lhandle_tail
+	cmp  $0x20, %edx
+	jb .Lhandle_tail
 
-	.p2align 4
-.Lloop_64:
 	/*
-	 * We decrement the loop index here - and the zero-flag is
-	 * checked at the end of the loop (instructions inbetween do
-	 * not change the zero flag):
+	 * We check whether memory false dependece could occur,
+	 * then jump to corresponding copy mode.
 	 */
-	decl %ecx
+	cmp  %dil, %sil
+	jl .Lcopy_backward
+	subl $0x20, %edx
+.Lcopy_forward_loop:
+	subq $0x20,	%rdx
 
 	/*
-	 * Move in blocks of 4x16 bytes:
+	 * Move in blocks of 4x8 bytes:
 	 */
-	movq 0*8(%rsi),		%r11
-	movq 1*8(%rsi),		%r8
-	movq %r11,		0*8(%rdi)
-	movq %r8,		1*8(%rdi)
-
-	movq 2*8(%rsi),		%r9
-	movq 3*8(%rsi),		%r10
-	movq %r9,		2*8(%rdi)
-	movq %r10,		3*8(%rdi)
-
-	movq 4*8(%rsi),		%r11
-	movq 5*8(%rsi),		%r8
-	movq %r11,		4*8(%rdi)
-	movq %r8,		5*8(%rdi)
-
-	movq 6*8(%rsi),		%r9
-	movq 7*8(%rsi),		%r10
-	movq %r9,		6*8(%rdi)
-	movq %r10,		7*8(%rdi)
-
-	leaq 64(%rsi), %rsi
-	leaq 64(%rdi), %rdi
-
-	jnz  .Lloop_64
+	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:
-	movl %edx, %ecx
-	andl  $63, %ecx
-	shrl   $3, %ecx
-	jz   .Lhandle_7
+	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
-.Lloop_8:
-	decl %ecx
-	movq (%rsi),		%r8
-	movq %r8,		(%rdi)
-	leaq 8(%rdi),		%rdi
-	leaq 8(%rsi),		%rsi
-	jnz  .Lloop_8
-
-.Lhandle_7:
-	movl %edx, %ecx
-	andl $7, %ecx
-	jz .Lend
+.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 %ecx
+	decl %edx
 	jnz .Lloop_1
 
 .Lend:
-	ret
+	retq
 	CFI_ENDPROC
 ENDPROC(memcpy)
 ENDPROC(__memcpy)
-- 
cgit v1.2.3


From 3b4b682becdfa9f42321aa024d5cc84f71f06d8c Mon Sep 17 00:00:00 2001
From: Ma Ling <ling.ma@intel.com>
Date: Fri, 17 Sep 2010 03:12:40 +0800
Subject: x86, mem: Optimize memmove for small size and unaligned cases

movs instruction will combine data to accelerate moving data,
however we need to concern two cases about it.

1. movs instruction need long lantency to startup,
   so here we use general mov instruction to copy data.
2. movs instruction is not good for unaligned case,
   even if src offset is 0x10, dest offset is 0x0,
   we avoid and handle the case by general mov instruction.

Signed-off-by: Ma Ling <ling.ma@intel.com>
LKML-Reference: <1284664360-6138-1-git-send-email-ling.ma@intel.com>
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
---
 arch/x86/lib/memcpy_32.c  | 213 ++++++++++++++++++++++++++++++++++++-------
 arch/x86/lib/memmove_64.c | 225 ++++++++++++++++++++++++++++++++++++----------
 2 files changed, 362 insertions(+), 76 deletions(-)

(limited to 'arch')

diff --git a/arch/x86/lib/memcpy_32.c b/arch/x86/lib/memcpy_32.c
index 81130d477ee..b908a59eccf 100644
--- a/arch/x86/lib/memcpy_32.c
+++ b/arch/x86/lib/memcpy_32.c
@@ -22,36 +22,187 @@ EXPORT_SYMBOL(memset);
 
 void *memmove(void *dest, const void *src, size_t n)
 {
-	int d0, d1, d2;
-
-	if (dest < src) {
-		if ((dest + n) < src)
-			 return memcpy(dest, src, n);
-		else
-			__asm__ __volatile__(
-				"rep\n\t"
-				"movsb\n\t"
-				: "=&c" (d0), "=&S" (d1), "=&D" (d2)
-				:"0" (n),
-				 "1" (src),
-				 "2" (dest)
-				:"memory");
-	} else {
-		if((src + n) < dest)
-			return memcpy(dest, src, n);
-		else
-			__asm__ __volatile__(
-				"std\n\t"
-				"rep\n\t"
-				"movsb\n\t"
-				"cld"
-				: "=&c" (d0), "=&S" (d1), "=&D" (d2)
-				:"0" (n),
-				 "1" (n-1+src),
-				 "2" (n-1+dest)
-				:"memory");
-	}
-
-	return dest;
+	int d0,d1,d2,d3,d4,d5;
+	char *ret = dest;
+
+	__asm__ __volatile__(
+		/* Handle more 16bytes in loop */
+		"cmp $0x10, %0\n\t"
+		"jb	1f\n\t"
+
+		/* Decide forward/backward copy mode */
+		"cmp %2, %1\n\t"
+		"jb	2f\n\t"
+
+		/*
+		 * movs instruction have many startup latency
+		 * so we handle small size by general register.
+		 */
+		"cmp  $680, %0\n\t"
+		"jb 3f\n\t"
+		/*
+		 * movs instruction is only good for aligned case.
+		 */
+		"mov %1, %3\n\t"
+		"xor %2, %3\n\t"
+		"and $0xff, %3\n\t"
+		"jz 4f\n\t"
+		"3:\n\t"
+		"sub $0x10, %0\n\t"
+
+		/*
+		 * We gobble 16byts forward in each loop.
+		 */
+		"3:\n\t"
+		"sub $0x10, %0\n\t"
+		"mov 0*4(%1), %3\n\t"
+		"mov 1*4(%1), %4\n\t"
+		"mov  %3, 0*4(%2)\n\t"
+		"mov  %4, 1*4(%2)\n\t"
+		"mov 2*4(%1), %3\n\t"
+		"mov 3*4(%1), %4\n\t"
+		"mov  %3, 2*4(%2)\n\t"
+		"mov  %4, 3*4(%2)\n\t"
+		"lea  0x10(%1), %1\n\t"
+		"lea  0x10(%2), %2\n\t"
+		"jae 3b\n\t"
+		"add $0x10, %0\n\t"
+		"jmp 1f\n\t"
+
+		/*
+		 * Handle data forward by movs.
+		 */
+		".p2align 4\n\t"
+		"4:\n\t"
+		"mov -4(%1, %0), %3\n\t"
+		"lea -4(%2, %0), %4\n\t"
+		"shr $2, %0\n\t"
+		"rep movsl\n\t"
+		"mov %3, (%4)\n\t"
+		"jmp 11f\n\t"
+		/*
+		 * Handle data backward by movs.
+		 */
+		".p2align 4\n\t"
+		"6:\n\t"
+		"mov (%1), %3\n\t"
+		"mov %2, %4\n\t"
+		"lea -4(%1, %0), %1\n\t"
+		"lea -4(%2, %0), %2\n\t"
+		"shr $2, %0\n\t"
+		"std\n\t"
+		"rep movsl\n\t"
+		"mov %3,(%4)\n\t"
+		"cld\n\t"
+		"jmp 11f\n\t"
+
+		/*
+		 * Start to prepare for backward copy.
+		 */
+		".p2align 4\n\t"
+		"2:\n\t"
+		"cmp  $680, %0\n\t"
+		"jb 5f\n\t"
+		"mov %1, %3\n\t"
+		"xor %2, %3\n\t"
+		"and $0xff, %3\n\t"
+		"jz 6b\n\t"
+
+		/*
+		 * Calculate copy position to tail.
+		 */
+		"5:\n\t"
+		"add %0, %1\n\t"
+		"add %0, %2\n\t"
+		"sub $0x10, %0\n\t"
+
+		/*
+		 * We gobble 16byts backward in each loop.
+		 */
+		"7:\n\t"
+		"sub $0x10, %0\n\t"
+
+		"mov -1*4(%1), %3\n\t"
+		"mov -2*4(%1), %4\n\t"
+		"mov  %3, -1*4(%2)\n\t"
+		"mov  %4, -2*4(%2)\n\t"
+		"mov -3*4(%1), %3\n\t"
+		"mov -4*4(%1), %4\n\t"
+		"mov  %3, -3*4(%2)\n\t"
+		"mov  %4, -4*4(%2)\n\t"
+		"lea  -0x10(%1), %1\n\t"
+		"lea  -0x10(%2), %2\n\t"
+		"jae 7b\n\t"
+		/*
+		 * Calculate copy position to head.
+		 */
+		"add $0x10, %0\n\t"
+		"sub %0, %1\n\t"
+		"sub %0, %2\n\t"
+
+		/*
+		 * Move data from 8 bytes to 15 bytes.
+		 */
+		".p2align 4\n\t"
+		"1:\n\t"
+		"cmp $8, %0\n\t"
+		"jb 8f\n\t"
+		"mov 0*4(%1), %3\n\t"
+		"mov 1*4(%1), %4\n\t"
+		"mov -2*4(%1, %0), %5\n\t"
+		"mov -1*4(%1, %0), %1\n\t"
+
+		"mov  %3, 0*4(%2)\n\t"
+		"mov  %4, 1*4(%2)\n\t"
+		"mov  %5, -2*4(%2, %0)\n\t"
+		"mov  %1, -1*4(%2, %0)\n\t"
+		"jmp 11f\n\t"
+
+		/*
+		 * Move data from 4 bytes to 7 bytes.
+		 */
+		".p2align 4\n\t"
+		"8:\n\t"
+		"cmp $4, %0\n\t"
+		"jb 9f\n\t"
+		"mov 0*4(%1), %3\n\t"
+		"mov -1*4(%1, %0), %4\n\t"
+		"mov  %3, 0*4(%2)\n\t"
+		"mov  %4, -1*4(%2, %0)\n\t"
+		"jmp 11f\n\t"
+
+		/*
+		 * Move data from 2 bytes to 3 bytes.
+		 */
+		".p2align 4\n\t"
+		"9:\n\t"
+		"cmp $2, %0\n\t"
+		"jb 10f\n\t"
+		"movw 0*2(%1), %%dx\n\t"
+		"movw -1*2(%1, %0), %%bx\n\t"
+		"movw %%dx, 0*2(%2)\n\t"
+		"movw %%bx, -1*2(%2, %0)\n\t"
+		"jmp 11f\n\t"
+
+		/*
+		 * Move data for 1 byte.
+		 */
+		".p2align 4\n\t"
+		"10:\n\t"
+		"cmp $1, %0\n\t"
+		"jb 11f\n\t"
+		"movb (%1), %%cl\n\t"
+		"movb %%cl, (%2)\n\t"
+		".p2align 4\n\t"
+		"11:"
+		: "=&c" (d0), "=&S" (d1), "=&D" (d2),
+		  "=r" (d3),"=r" (d4), "=r"(d5)
+		:"0" (n),
+		 "1" (src),
+		 "2" (dest)
+		:"memory");
+
+	return ret;
+
 }
 EXPORT_SYMBOL(memmove);
diff --git a/arch/x86/lib/memmove_64.c b/arch/x86/lib/memmove_64.c
index ecacc4b3d9e..6d0f0ec41b3 100644
--- a/arch/x86/lib/memmove_64.c
+++ b/arch/x86/lib/memmove_64.c
@@ -8,50 +8,185 @@
 #undef memmove
 void *memmove(void *dest, const void *src, size_t count)
 {
-	unsigned long d0, d1, d2, d3;
-	if (dest < src) {
-		if ((dest + count) < src)
-			 return memcpy(dest, src, count);
-		else
-			__asm__ __volatile__(
-				"movq %0, %3\n\t"
-				"shr $3, %0\n\t"
-				"andq $7, %3\n\t"
-				"rep\n\t"
-				"movsq\n\t"
-				"movq %3, %0\n\t"
-				"rep\n\t"
-				"movsb"
-				: "=&c" (d0), "=&S" (d1), "=&D" (d2), "=r" (d3)
-				:"0" (count),
-				 "1" (src),
-				 "2" (dest)
-				:"memory");
-	} else {
-		if((src + count) < dest)
-			return memcpy(dest, src, count);
-		else
-			__asm__ __volatile__(
-				"movq %0, %3\n\t"
-				"lea -8(%1, %0), %1\n\t"
-				"lea -8(%2, %0), %2\n\t"
-				"shr $3, %0\n\t"
-				"andq $7, %3\n\t"
-				"std\n\t"
-				"rep\n\t"
-				"movsq\n\t"
-				"lea 7(%1), %1\n\t"
-				"lea 7(%2), %2\n\t"
-				"movq %3, %0\n\t"
-				"rep\n\t"
-				"movsb\n\t"
-				"cld"
-				: "=&c" (d0), "=&S" (d1), "=&D" (d2), "=r" (d3)
-				:"0" (count),
-				 "1" (src),
-				 "2" (dest)
-				:"memory");
-	}
-	return dest;
+	unsigned long d0,d1,d2,d3,d4,d5,d6,d7;
+	char *ret;
+
+	__asm__ __volatile__(
+		/* Handle more 32bytes in loop */
+		"mov %2, %3\n\t"
+		"cmp $0x20, %0\n\t"
+		"jb	1f\n\t"
+
+		/* Decide forward/backward copy mode */
+		"cmp %2, %1\n\t"
+		"jb	2f\n\t"
+
+		/*
+		 * movsq instruction have many startup latency
+		 * so we handle small size by general register.
+		 */
+		"cmp  $680, %0\n\t"
+		"jb 3f\n\t"
+		/*
+		 * movsq instruction is only good for aligned case.
+		 */
+		"cmpb %%dil, %%sil\n\t"
+		"je 4f\n\t"
+		"3:\n\t"
+		"sub $0x20, %0\n\t"
+		/*
+		 * We gobble 32byts forward in each loop.
+		 */
+		"5:\n\t"
+		"sub $0x20, %0\n\t"
+		"movq 0*8(%1), %4\n\t"
+		"movq 1*8(%1), %5\n\t"
+		"movq 2*8(%1), %6\n\t"
+		"movq 3*8(%1), %7\n\t"
+		"leaq 4*8(%1), %1\n\t"
+
+		"movq %4, 0*8(%2)\n\t"
+		"movq %5, 1*8(%2)\n\t"
+		"movq %6, 2*8(%2)\n\t"
+		"movq %7, 3*8(%2)\n\t"
+		"leaq 4*8(%2), %2\n\t"
+		"jae 5b\n\t"
+		"addq $0x20, %0\n\t"
+		"jmp 1f\n\t"
+		/*
+		 * Handle data forward by movsq.
+		 */
+		".p2align 4\n\t"
+		"4:\n\t"
+		"movq %0, %8\n\t"
+		"movq -8(%1, %0), %4\n\t"
+		"lea -8(%2, %0), %5\n\t"
+		"shrq $3, %8\n\t"
+		"rep movsq\n\t"
+		"movq %4, (%5)\n\t"
+		"jmp 13f\n\t"
+		/*
+		 * Handle data backward by movsq.
+		 */
+		".p2align 4\n\t"
+		"7:\n\t"
+		"movq %0, %8\n\t"
+		"movq (%1), %4\n\t"
+		"movq %2, %5\n\t"
+		"leaq -8(%1, %0), %1\n\t"
+		"leaq -8(%2, %0), %2\n\t"
+		"shrq $3, %8\n\t"
+		"std\n\t"
+		"rep movsq\n\t"
+		"cld\n\t"
+		"movq %4, (%5)\n\t"
+		"jmp 13f\n\t"
+
+		/*
+		 * Start to prepare for backward copy.
+		 */
+		".p2align 4\n\t"
+		"2:\n\t"
+		"cmp $680, %0\n\t"
+		"jb 6f \n\t"
+		"cmp %%dil, %%sil\n\t"
+		"je 7b \n\t"
+		"6:\n\t"
+		/*
+		 * Calculate copy position to tail.
+		 */
+		"addq %0, %1\n\t"
+		"addq %0, %2\n\t"
+		"subq $0x20, %0\n\t"
+		/*
+		 * We gobble 32byts backward in each loop.
+		 */
+		"8:\n\t"
+		"subq $0x20, %0\n\t"
+		"movq -1*8(%1), %4\n\t"
+		"movq -2*8(%1), %5\n\t"
+		"movq -3*8(%1), %6\n\t"
+		"movq -4*8(%1), %7\n\t"
+		"leaq -4*8(%1), %1\n\t"
+
+		"movq %4, -1*8(%2)\n\t"
+		"movq %5, -2*8(%2)\n\t"
+		"movq %6, -3*8(%2)\n\t"
+		"movq %7, -4*8(%2)\n\t"
+		"leaq -4*8(%2), %2\n\t"
+		"jae 8b\n\t"
+		/*
+		 * Calculate copy position to head.
+		 */
+		"addq $0x20, %0\n\t"
+		"subq %0, %1\n\t"
+		"subq %0, %2\n\t"
+		"1:\n\t"
+		"cmpq $16, %0\n\t"
+		"jb 9f\n\t"
+		/*
+		 * Move data from 16 bytes to 31 bytes.
+		 */
+		"movq 0*8(%1), %4\n\t"
+		"movq 1*8(%1), %5\n\t"
+		"movq -2*8(%1, %0), %6\n\t"
+		"movq -1*8(%1, %0), %7\n\t"
+		"movq %4, 0*8(%2)\n\t"
+		"movq %5, 1*8(%2)\n\t"
+		"movq %6, -2*8(%2, %0)\n\t"
+		"movq %7, -1*8(%2, %0)\n\t"
+		"jmp 13f\n\t"
+		".p2align 4\n\t"
+		"9:\n\t"
+		"cmpq $8, %0\n\t"
+		"jb 10f\n\t"
+		/*
+		 * Move data from 8 bytes to 15 bytes.
+		 */
+		"movq 0*8(%1), %4\n\t"
+		"movq -1*8(%1, %0), %5\n\t"
+		"movq %4, 0*8(%2)\n\t"
+		"movq %5, -1*8(%2, %0)\n\t"
+		"jmp 13f\n\t"
+		"10:\n\t"
+		"cmpq $4, %0\n\t"
+		"jb 11f\n\t"
+		/*
+		 * Move data from 4 bytes to 7 bytes.
+		 */
+		"movl (%1), %4d\n\t"
+		"movl -4(%1, %0), %5d\n\t"
+		"movl %4d, (%2)\n\t"
+		"movl %5d, -4(%2, %0)\n\t"
+		"jmp 13f\n\t"
+		"11:\n\t"
+		"cmp $2, %0\n\t"
+		"jb 12f\n\t"
+		/*
+		 * Move data from 2 bytes to 3 bytes.
+		 */
+		"movw (%1), %4w\n\t"
+		"movw -2(%1, %0), %5w\n\t"
+		"movw %4w, (%2)\n\t"
+		"movw %5w, -2(%2, %0)\n\t"
+		"jmp 13f\n\t"
+		"12:\n\t"
+		"cmp $1, %0\n\t"
+		"jb 13f\n\t"
+		/*
+		 * Move data for 1 byte.
+		 */
+		"movb (%1), %4b\n\t"
+		"movb %4b, (%2)\n\t"
+		"13:\n\t"
+		: "=&d" (d0), "=&S" (d1), "=&D" (d2), "=&a" (ret) ,
+		  "=r"(d3), "=r"(d4), "=r"(d5), "=r"(d6), "=&c" (d7)
+		:"0" (count),
+		 "1" (src),
+		 "2" (dest)
+		:"memory");
+
+		return ret;
+
 }
 EXPORT_SYMBOL(memmove);
-- 
cgit v1.2.3