[glibc/release/2.33/master] x86: Optimize memcmp-evex-movbe.S

[glibc/release/2.33/master] x86: Optimize memcmp-evex-movbe.S

[Sunil Pandey]

https://sourceware.org/git/gitweb.cgi?p=glibc.git;h=903190e981e995f9f10063fd717ce72ab1c6cb04

commit 903190e981e995f9f10063fd717ce72ab1c6cb04
Author: Noah Goldstein <goldstein.w.n@gmail.com>
Date:   Mon May 17 13:57:24 2021 -0400

    x86: Optimize memcmp-evex-movbe.S
    
    No bug. This commit optimizes memcmp-evex.S. The optimizations include
    adding a new vec compare path for small sizes, reorganizing the entry
    control flow, removing some unnecissary ALU instructions from the main
    loop, and most importantly replacing the heavy use of vpcmp + kand
    logic with vpxor + vptern. test-memcmp and test-wmemcmp are both
    passing.
    
    Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com>
    Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
    (cherry picked from commit 4ad473e97acdc5f6d811755b67c09f2128a644ce)

Diff:
---
 sysdeps/x86_64/multiarch/memcmp-evex-movbe.S | 710 +++++++++++++++------------
 1 file changed, 408 insertions(+), 302 deletions(-)

diff --git a/sysdeps/x86_64/multiarch/memcmp-evex-movbe.S b/sysdeps/x86_64/multiarch/memcmp-evex-movbe.S
index 9c093972e1..654dc7ac8c 100644
--- a/sysdeps/x86_64/multiarch/memcmp-evex-movbe.S
+++ b/sysdeps/x86_64/multiarch/memcmp-evex-movbe.S
@@ -19,17 +19,22 @@
 #if IS_IN (libc)
 
 /* memcmp/wmemcmp is implemented as:
-   1. For size from 2 to 7 bytes, load as big endian with movbe and bswap
-      to avoid branches.
-   2. Use overlapping compare to avoid branch.
-   3. Use vector compare when size >= 4 bytes for memcmp or size >= 8
-      bytes for wmemcmp.
-   4. If size is 8 * VEC_SIZE or less, unroll the loop.
-   5. Compare 4 * VEC_SIZE at a time with the aligned first memory
+   1. Use ymm vector compares when possible. The only case where
+      vector compares is not possible for when size < CHAR_PER_VEC
+      and loading from either s1 or s2 would cause a page cross.
+   2. For size from 2 to 7 bytes on page cross, load as big endian
+      with movbe and bswap to avoid branches.
+   3. Use xmm vector compare when size >= 4 bytes for memcmp or
+      size >= 8 bytes for wmemcmp.
+   4. Optimistically compare up to first 4 * CHAR_PER_VEC one at a
+      to check for early mismatches. Only do this if its guranteed the
+      work is not wasted.
+   5. If size is 8 * VEC_SIZE or less, unroll the loop.
+   6. Compare 4 * VEC_SIZE at a time with the aligned first memory
       area.
-   6. Use 2 vector compares when size is 2 * VEC_SIZE or less.
-   7. Use 4 vector compares when size is 4 * VEC_SIZE or less.
-   8. Use 8 vector compares when size is 8 * VEC_SIZE or less.  */
+   7. Use 2 vector compares when size is 2 * CHAR_PER_VEC or less.
+   8. Use 4 vector compares when size is 4 * CHAR_PER_VEC or less.
+   9. Use 8 vector compares when size is 8 * CHAR_PER_VEC or less.  */
 
 # include <sysdep.h>
 
@@ -40,11 +45,21 @@
 # define VMOVU		vmovdqu64
 
 # ifdef USE_AS_WMEMCMP
-#  define VPCMPEQ	vpcmpeqd
+#  define CHAR_SIZE	4
+#  define VPCMP	vpcmpd
 # else
-#  define VPCMPEQ	vpcmpeqb
+#  define CHAR_SIZE	1
+#  define VPCMP	vpcmpub
 # endif
 
+# define VEC_SIZE	32
+# define PAGE_SIZE	4096
+# define CHAR_PER_VEC	(VEC_SIZE / CHAR_SIZE)
+
+# define XMM0		xmm16
+# define XMM1		xmm17
+# define XMM2		xmm18
+# define YMM0		ymm16
 # define XMM1		xmm17
 # define XMM2		xmm18
 # define YMM1		ymm17
@@ -54,15 +69,6 @@
 # define YMM5		ymm21
 # define YMM6		ymm22
 
-# define VEC_SIZE 32
-# ifdef USE_AS_WMEMCMP
-#  define VEC_MASK 0xff
-#  define XMM_MASK 0xf
-# else
-#  define VEC_MASK 0xffffffff
-#  define XMM_MASK 0xffff
-# endif
-
 /* Warning!
            wmemcmp has to use SIGNED comparison for elements.
            memcmp has to use UNSIGNED comparison for elemnts.
@@ -70,145 +76,370 @@
 
 	.section .text.evex,"ax",@progbits
 ENTRY (MEMCMP)
-# ifdef USE_AS_WMEMCMP
-	shl	$2, %RDX_LP
-# elif defined __ILP32__
+# ifdef __ILP32__
 	/* Clear the upper 32 bits.  */
 	movl	%edx, %edx
 # endif
-	cmp	$VEC_SIZE, %RDX_LP
+	cmp	$CHAR_PER_VEC, %RDX_LP
 	jb	L(less_vec)
 
 	/* From VEC to 2 * VEC.  No branch when size == VEC_SIZE.  */
-	VMOVU	(%rsi), %YMM2
-	VPCMPEQ (%rdi), %YMM2, %k1
+	VMOVU	(%rsi), %YMM1
+	/* Use compare not equals to directly check for mismatch.  */
+	VPCMP	$4, (%rdi), %YMM1, %k1
 	kmovd	%k1, %eax
-	subl    $VEC_MASK, %eax
-	jnz	L(first_vec)
-
-	cmpq	$(VEC_SIZE * 2), %rdx
-	jbe	L(last_vec)
-
-	/* More than 2 * VEC.  */
-	cmpq	$(VEC_SIZE * 8), %rdx
-	ja	L(more_8x_vec)
-	cmpq	$(VEC_SIZE * 4), %rdx
-	jb	L(last_4x_vec)
+	/* NB: eax must be destination register if going to
+	   L(return_vec_[0,2]). For L(return_vec_3 destination register
+	   must be ecx.  */
+	testl	%eax, %eax
+	jnz	L(return_vec_0)
 
-	/* From 4 * VEC to 8 * VEC, inclusively. */
-	VMOVU	(%rsi), %YMM1
-	VPCMPEQ (%rdi), %YMM1, %k1
+	cmpq	$(CHAR_PER_VEC * 2), %rdx
+	jbe	L(last_1x_vec)
 
+	/* Check second VEC no matter what.  */
 	VMOVU	VEC_SIZE(%rsi), %YMM2
-	VPCMPEQ VEC_SIZE(%rdi), %YMM2, %k2
+	VPCMP	$4, VEC_SIZE(%rdi), %YMM2, %k1
+	kmovd	%k1, %eax
+	testl	%eax, %eax
+	jnz	L(return_vec_1)
+
+	/* Less than 4 * VEC.  */
+	cmpq	$(CHAR_PER_VEC * 4), %rdx
+	jbe	L(last_2x_vec)
 
+	/* Check third and fourth VEC no matter what.  */
 	VMOVU	(VEC_SIZE * 2)(%rsi), %YMM3
-	VPCMPEQ (VEC_SIZE * 2)(%rdi), %YMM3, %k3
+	VPCMP	$4, (VEC_SIZE * 2)(%rdi), %YMM3, %k1
+	kmovd	%k1, %eax
+	testl	%eax, %eax
+	jnz	L(return_vec_2)
 
 	VMOVU	(VEC_SIZE * 3)(%rsi), %YMM4
-	VPCMPEQ (VEC_SIZE * 3)(%rdi), %YMM4, %k4
+	VPCMP	$4, (VEC_SIZE * 3)(%rdi), %YMM4, %k1
+	kmovd	%k1, %ecx
+	testl	%ecx, %ecx
+	jnz	L(return_vec_3)
 
-	kandd	%k1, %k2, %k5
-	kandd	%k3, %k4, %k6
-	kandd	%k5, %k6, %k6
+	/* Zero YMM0. 4x VEC reduction is done with vpxor + vtern so
+	   compare with zero to get a mask is needed.  */
+	vpxorq	%XMM0, %XMM0, %XMM0
 
-	kmovd	%k6, %eax
-	cmpl	$VEC_MASK, %eax
-	jne	L(4x_vec_end)
+	/* Go to 4x VEC loop.  */
+	cmpq	$(CHAR_PER_VEC * 8), %rdx
+	ja	L(more_8x_vec)
 
-	leaq	-(4 * VEC_SIZE)(%rdi, %rdx), %rdi
-	leaq	-(4 * VEC_SIZE)(%rsi, %rdx), %rsi
-	VMOVU	(%rsi), %YMM1
-	VPCMPEQ (%rdi), %YMM1, %k1
+	/* Handle remainder of size = 4 * VEC + 1 to 8 * VEC without any
+	   branches.  */
 
-	VMOVU	VEC_SIZE(%rsi), %YMM2
-	VPCMPEQ VEC_SIZE(%rdi), %YMM2, %k2
-	kandd	%k1, %k2, %k5
+	/* Load first two VEC from s2 before adjusting addresses.  */
+	VMOVU	-(VEC_SIZE * 4)(%rsi, %rdx, CHAR_SIZE), %YMM1
+	VMOVU	-(VEC_SIZE * 3)(%rsi, %rdx, CHAR_SIZE), %YMM2
+	leaq	-(4 * VEC_SIZE)(%rdi, %rdx, CHAR_SIZE), %rdi
+	leaq	-(4 * VEC_SIZE)(%rsi, %rdx, CHAR_SIZE), %rsi
+
+	/* Wait to load from s1 until addressed adjust due to
+	   unlamination of microfusion with complex address mode.  */
+
+	/* vpxor will be all 0s if s1 and s2 are equal. Otherwise it
+	   will have some 1s.  */
+	vpxorq	(%rdi), %YMM1, %YMM1
+	vpxorq	(VEC_SIZE)(%rdi), %YMM2, %YMM2
 
 	VMOVU	(VEC_SIZE * 2)(%rsi), %YMM3
-	VPCMPEQ (VEC_SIZE * 2)(%rdi), %YMM3, %k3
-	kandd	%k3, %k5, %k5
+	vpxorq	(VEC_SIZE * 2)(%rdi), %YMM3, %YMM3
+	/* Or together YMM1, YMM2, and YMM3 into YMM3.  */
+	vpternlogd $0xfe, %YMM1, %YMM2, %YMM3
 
 	VMOVU	(VEC_SIZE * 3)(%rsi), %YMM4
-	VPCMPEQ (VEC_SIZE * 3)(%rdi), %YMM4, %k4
-	kandd	%k4, %k5, %k5
+	/* Ternary logic to xor (VEC_SIZE * 3)(%rdi) with YMM4 while
+	   oring with YMM3. Result is stored in YMM4.  */
+	vpternlogd $0xde, (VEC_SIZE * 3)(%rdi), %YMM3, %YMM4
+	/* Compare YMM4 with 0. If any 1s s1 and s2 don't match.  */
+	VPCMP	$4, %YMM4, %YMM0, %k1
+	kmovd	%k1, %ecx
+	testl	%ecx, %ecx
+	jnz	L(return_vec_0_1_2_3)
+	/* NB: eax must be zero to reach here.  */
+	ret
 
-	kmovd	%k5, %eax
-	cmpl	$VEC_MASK, %eax
-	jne	L(4x_vec_end)
-	xorl	%eax, %eax
+	/* NB: aligning 32 here allows for the rest of the jump targets
+	   to be tuned for 32 byte alignment. Most important this ensures
+	   the L(more_8x_vec) loop is 32 byte aligned.  */
+	.p2align 5
+L(less_vec):
+	/* Check if one or less CHAR. This is necessary for size = 0 but
+	   is also faster for size = CHAR_SIZE.  */
+	cmpl	$1, %edx
+	jbe	L(one_or_less)
+
+	/* Check if loading one VEC from either s1 or s2 could cause a
+	   page cross. This can have false positives but is by far the
+	   fastest method.  */
+	movl	%edi, %eax
+	orl	%esi, %eax
+	andl	$(PAGE_SIZE - 1), %eax
+	cmpl	$(PAGE_SIZE - VEC_SIZE), %eax
+	jg	L(page_cross_less_vec)
+
+	/* No page cross possible.  */
+	VMOVU	(%rsi), %YMM2
+	VPCMP	$4, (%rdi), %YMM2, %k1
+	kmovd	%k1, %eax
+	/* Create mask in ecx for potentially in bound matches.  */
+	bzhil	%edx, %eax, %eax
+	jnz	L(return_vec_0)
 	ret
 
 	.p2align 4
-L(last_2x_vec):
-	/* From VEC to 2 * VEC.  No branch when size == VEC_SIZE.  */
-	VMOVU	(%rsi), %YMM2
-	VPCMPEQ (%rdi), %YMM2, %k2
-	kmovd	%k2, %eax
-	subl    $VEC_MASK, %eax
-	jnz	L(first_vec)
+L(return_vec_0):
+	tzcntl	%eax, %eax
+# ifdef USE_AS_WMEMCMP
+	movl	(%rdi, %rax, CHAR_SIZE), %ecx
+	xorl	%edx, %edx
+	cmpl	(%rsi, %rax, CHAR_SIZE), %ecx
+	/* NB: no partial register stall here because xorl zero idiom
+	   above.  */
+	setg	%dl
+	leal	-1(%rdx, %rdx), %eax
+# else
+	movzbl	(%rsi, %rax), %ecx
+	movzbl	(%rdi, %rax), %eax
+	subl	%ecx, %eax
+# endif
+	ret
 
-L(last_vec):
-	/* Use overlapping loads to avoid branches.  */
-	leaq	-VEC_SIZE(%rdi, %rdx), %rdi
-	leaq	-VEC_SIZE(%rsi, %rdx), %rsi
-	VMOVU	(%rsi), %YMM2
-	VPCMPEQ (%rdi), %YMM2, %k2
-	kmovd	%k2, %eax
-	subl    $VEC_MASK, %eax
-	jnz	L(first_vec)
+	/* NB: No p2align necessary. Alignment  % 16 is naturally 1
+	   which is good enough for a target not in a loop.  */
+L(return_vec_1):
+	tzcntl	%eax, %eax
+# ifdef USE_AS_WMEMCMP
+	movl	VEC_SIZE(%rdi, %rax, CHAR_SIZE), %ecx
+	xorl	%edx, %edx
+	cmpl	VEC_SIZE(%rsi, %rax, CHAR_SIZE), %ecx
+	setg	%dl
+	leal	-1(%rdx, %rdx), %eax
+# else
+	movzbl	VEC_SIZE(%rsi, %rax), %ecx
+	movzbl	VEC_SIZE(%rdi, %rax), %eax
+	subl	%ecx, %eax
+# endif
 	ret
 
-	.p2align 4
-L(first_vec):
-	/* A byte or int32 is different within 16 or 32 bytes.  */
-	tzcntl	%eax, %ecx
+	/* NB: No p2align necessary. Alignment  % 16 is naturally 2
+	   which is good enough for a target not in a loop.  */
+L(return_vec_2):
+	tzcntl	%eax, %eax
 # ifdef USE_AS_WMEMCMP
-	xorl	%eax, %eax
-	movl	(%rdi, %rcx, 4), %edx
-	cmpl	(%rsi, %rcx, 4), %edx
-L(wmemcmp_return):
-	setl	%al
-	negl	%eax
-	orl	$1, %eax
+	movl	(VEC_SIZE * 2)(%rdi, %rax, CHAR_SIZE), %ecx
+	xorl	%edx, %edx
+	cmpl	(VEC_SIZE * 2)(%rsi, %rax, CHAR_SIZE), %ecx
+	setg	%dl
+	leal	-1(%rdx, %rdx), %eax
 # else
-	movzbl	(%rdi, %rcx), %eax
-	movzbl	(%rsi, %rcx), %edx
-	sub	%edx, %eax
+	movzbl	(VEC_SIZE * 2)(%rsi, %rax), %ecx
+	movzbl	(VEC_SIZE * 2)(%rdi, %rax), %eax
+	subl	%ecx, %eax
 # endif
 	ret
 
+	.p2align 4
+L(8x_return_vec_0_1_2_3):
+	/* Returning from L(more_8x_vec) requires restoring rsi.  */
+	addq	%rdi, %rsi
+L(return_vec_0_1_2_3):
+	VPCMP	$4, %YMM1, %YMM0, %k0
+	kmovd	%k0, %eax
+	testl	%eax, %eax
+	jnz	L(return_vec_0)
+
+	VPCMP	$4, %YMM2, %YMM0, %k0
+	kmovd	%k0, %eax
+	testl	%eax, %eax
+	jnz	L(return_vec_1)
+
+	VPCMP	$4, %YMM3, %YMM0, %k0
+	kmovd	%k0, %eax
+	testl	%eax, %eax
+	jnz	L(return_vec_2)
+L(return_vec_3):
+	tzcntl	%ecx, %ecx
 # ifdef USE_AS_WMEMCMP
+	movl	(VEC_SIZE * 3)(%rdi, %rcx, CHAR_SIZE), %eax
+	xorl	%edx, %edx
+	cmpl	(VEC_SIZE * 3)(%rsi, %rcx, CHAR_SIZE), %eax
+	setg	%dl
+	leal	-1(%rdx, %rdx), %eax
+# else
+	movzbl	(VEC_SIZE * 3)(%rdi, %rcx), %eax
+	movzbl	(VEC_SIZE * 3)(%rsi, %rcx), %ecx
+	subl	%ecx, %eax
+# endif
+	ret
+
 	.p2align 4
-L(4):
-	xorl	%eax, %eax
-	movl	(%rdi), %edx
-	cmpl	(%rsi), %edx
-	jne	L(wmemcmp_return)
+L(more_8x_vec):
+	/* Set end of s1 in rdx.  */
+	leaq	-(VEC_SIZE * 4)(%rdi, %rdx, CHAR_SIZE), %rdx
+	/* rsi stores s2 - s1. This allows loop to only update one
+	   pointer.  */
+	subq	%rdi, %rsi
+	/* Align s1 pointer.  */
+	andq	$-VEC_SIZE, %rdi
+	/* Adjust because first 4x vec where check already.  */
+	subq	$-(VEC_SIZE * 4), %rdi
+	.p2align 4
+L(loop_4x_vec):
+	VMOVU	(%rsi, %rdi), %YMM1
+	vpxorq	(%rdi), %YMM1, %YMM1
+
+	VMOVU	VEC_SIZE(%rsi, %rdi), %YMM2
+	vpxorq	VEC_SIZE(%rdi), %YMM2, %YMM2
+
+	VMOVU	(VEC_SIZE * 2)(%rsi, %rdi), %YMM3
+	vpxorq	(VEC_SIZE * 2)(%rdi), %YMM3, %YMM3
+	vpternlogd $0xfe, %YMM1, %YMM2, %YMM3
+
+	VMOVU	(VEC_SIZE * 3)(%rsi, %rdi), %YMM4
+	vpternlogd $0xde, (VEC_SIZE * 3)(%rdi), %YMM3, %YMM4
+	VPCMP	$4, %YMM4, %YMM0, %k1
+	kmovd	%k1, %ecx
+	testl	%ecx, %ecx
+	jnz	L(8x_return_vec_0_1_2_3)
+	subq	$-(VEC_SIZE * 4), %rdi
+	cmpq	%rdx, %rdi
+	jb	L(loop_4x_vec)
+
+	subq	%rdx, %rdi
+	/* rdi has 4 * VEC_SIZE - remaining length.  */
+	cmpl	$(VEC_SIZE * 3), %edi
+	jae	L(8x_last_1x_vec)
+	/* Load regardless of branch.  */
+	VMOVU	(VEC_SIZE * 2)(%rsi, %rdx), %YMM3
+	cmpl	$(VEC_SIZE * 2), %edi
+	jae	L(8x_last_2x_vec)
+
+	VMOVU	(%rsi, %rdx), %YMM1
+	vpxorq	(%rdx), %YMM1, %YMM1
+
+	VMOVU	VEC_SIZE(%rsi, %rdx), %YMM2
+	vpxorq	VEC_SIZE(%rdx), %YMM2, %YMM2
+
+	vpxorq	(VEC_SIZE * 2)(%rdx), %YMM3, %YMM3
+	vpternlogd $0xfe, %YMM1, %YMM2, %YMM3
+
+	VMOVU	(VEC_SIZE * 3)(%rsi, %rdx), %YMM4
+	vpternlogd $0xde, (VEC_SIZE * 3)(%rdx), %YMM3, %YMM4
+	VPCMP	$4, %YMM4, %YMM0, %k1
+	kmovd	%k1, %ecx
+	/* Restore s1 pointer to rdi.  */
+	movq	%rdx, %rdi
+	testl	%ecx, %ecx
+	jnz	L(8x_return_vec_0_1_2_3)
+	/* NB: eax must be zero to reach here.  */
+	ret
+
+	/* Only entry is from L(more_8x_vec).  */
+	.p2align 4
+L(8x_last_2x_vec):
+	VPCMP	$4, (VEC_SIZE * 2)(%rdx), %YMM3, %k1
+	kmovd	%k1, %eax
+	testl	%eax, %eax
+	jnz	L(8x_return_vec_2)
+	/* Naturally aligned to 16 bytes.  */
+L(8x_last_1x_vec):
+	VMOVU	(VEC_SIZE * 3)(%rsi, %rdx), %YMM1
+	VPCMP	$4, (VEC_SIZE * 3)(%rdx), %YMM1, %k1
+	kmovd	%k1, %eax
+	testl	%eax, %eax
+	jnz	L(8x_return_vec_3)
+	ret
+
+	.p2align 4
+L(last_2x_vec):
+	/* Check second to last VEC.  */
+	VMOVU	-(VEC_SIZE * 2)(%rsi, %rdx, CHAR_SIZE), %YMM1
+	VPCMP	$4, -(VEC_SIZE * 2)(%rdi, %rdx, CHAR_SIZE), %YMM1, %k1
+	kmovd	%k1, %eax
+	testl	%eax, %eax
+	jnz	L(return_vec_1_end)
+
+	/* Check last VEC.  */
+	.p2align 4
+L(last_1x_vec):
+	VMOVU	-(VEC_SIZE * 1)(%rsi, %rdx, CHAR_SIZE), %YMM1
+	VPCMP	$4, -(VEC_SIZE * 1)(%rdi, %rdx, CHAR_SIZE), %YMM1, %k1
+	kmovd	%k1, %eax
+	testl	%eax, %eax
+	jnz	L(return_vec_0_end)
 	ret
+
+	.p2align 4
+L(8x_return_vec_2):
+	subq	$VEC_SIZE, %rdx
+L(8x_return_vec_3):
+	tzcntl	%eax, %eax
+# ifdef USE_AS_WMEMCMP
+	leaq	(%rdx, %rax, CHAR_SIZE), %rax
+	movl	(VEC_SIZE * 3)(%rax), %ecx
+	xorl	%edx, %edx
+	cmpl	(VEC_SIZE * 3)(%rsi, %rax), %ecx
+	setg	%dl
+	leal	-1(%rdx, %rdx), %eax
 # else
+	addq	%rdx, %rax
+	movzbl	(VEC_SIZE * 3)(%rsi, %rax), %ecx
+	movzbl	(VEC_SIZE * 3)(%rax), %eax
+	subl	%ecx, %eax
+# endif
+	ret
+
 	.p2align 4
-L(between_4_7):
-	/* Load as big endian with overlapping movbe to avoid branches.  */
-	movbe	(%rdi), %eax
-	movbe	(%rsi), %ecx
-	shlq	$32, %rax
-	shlq	$32, %rcx
-	movbe	-4(%rdi, %rdx), %edi
-	movbe	-4(%rsi, %rdx), %esi
-	orq	%rdi, %rax
-	orq	%rsi, %rcx
-	subq	%rcx, %rax
-	je	L(exit)
-	sbbl	%eax, %eax
-	orl	$1, %eax
+L(return_vec_0_end):
+	tzcntl	%eax, %eax
+	addl	%edx, %eax
+# ifdef USE_AS_WMEMCMP
+	movl	-VEC_SIZE(%rdi, %rax, CHAR_SIZE), %ecx
+	xorl	%edx, %edx
+	cmpl	-VEC_SIZE(%rsi, %rax, CHAR_SIZE), %ecx
+	setg	%dl
+	leal	-1(%rdx, %rdx), %eax
+# else
+	movzbl	-VEC_SIZE(%rsi, %rax), %ecx
+	movzbl	-VEC_SIZE(%rdi, %rax), %eax
+	subl	%ecx, %eax
+# endif
 	ret
 
 	.p2align 4
-L(exit):
+L(return_vec_1_end):
+	tzcntl	%eax, %eax
+	addl	%edx, %eax
+# ifdef USE_AS_WMEMCMP
+	movl	-(VEC_SIZE * 2)(%rdi, %rax, CHAR_SIZE), %ecx
+	xorl	%edx, %edx
+	cmpl	-(VEC_SIZE * 2)(%rsi, %rax, CHAR_SIZE), %ecx
+	setg	%dl
+	leal	-1(%rdx, %rdx), %eax
+# else
+	movzbl	-(VEC_SIZE * 2)(%rsi, %rax), %ecx
+	movzbl	-(VEC_SIZE * 2)(%rdi, %rax), %eax
+	subl	%ecx, %eax
+# endif
 	ret
 
+
 	.p2align 4
+L(page_cross_less_vec):
+	/* if USE_AS_WMEMCMP it can only be 0, 4, 8, 12, 16, 20, 24, 28
+	   bytes.  */
+	cmpl	$(16 / CHAR_SIZE), %edx
+	jae	L(between_16_31)
+# ifndef USE_AS_WMEMCMP
+	cmpl	$8, %edx
+	jae	L(between_8_15)
+	cmpl	$4, %edx
+	jae	L(between_4_7)
 L(between_2_3):
 	/* Load as big endian to avoid branches.  */
 	movzwl	(%rdi), %eax
@@ -217,224 +448,99 @@ L(between_2_3):
 	shll	$8, %ecx
 	bswap	%eax
 	bswap	%ecx
-	movb	-1(%rdi, %rdx), %al
-	movb	-1(%rsi, %rdx), %cl
+	movzbl	-1(%rdi, %rdx), %edi
+	movzbl	-1(%rsi, %rdx), %esi
+	orl	%edi, %eax
+	orl	%esi, %ecx
 	/* Subtraction is okay because the upper 8 bits are zero.  */
 	subl	%ecx, %eax
 	ret
-
 	.p2align 4
-L(1):
-	movzbl	(%rdi), %eax
+L(one_or_less):
+	jb	L(zero)
 	movzbl	(%rsi), %ecx
+	movzbl	(%rdi), %eax
 	subl	%ecx, %eax
 	ret
-# endif
-
-	.p2align 4
-L(zero):
-	xorl	%eax, %eax
-	ret
 
 	.p2align 4
-L(less_vec):
-# ifdef USE_AS_WMEMCMP
-	/* It can only be 0, 4, 8, 12, 16, 20, 24, 28 bytes.  */
-	cmpb	$4, %dl
-	je	L(4)
-	jb	L(zero)
-# else
-	cmpb	$1, %dl
-	je	L(1)
-	jb	L(zero)
-	cmpb	$4, %dl
-	jb	L(between_2_3)
-	cmpb	$8, %dl
-	jb	L(between_4_7)
+L(between_8_15):
 # endif
-	cmpb	$16, %dl
-	jae	L(between_16_31)
-	/* It is between 8 and 15 bytes.  */
+	/* If USE_AS_WMEMCMP fall through into 8-15 byte case.  */
 	vmovq	(%rdi), %XMM1
 	vmovq	(%rsi), %XMM2
-	VPCMPEQ %XMM1, %XMM2, %k2
-	kmovw	%k2, %eax
-	subl    $XMM_MASK, %eax
-	jnz	L(first_vec)
+	VPCMP	$4, %XMM1, %XMM2, %k1
+	kmovd	%k1, %eax
+	testl	%eax, %eax
+	jnz	L(return_vec_0)
 	/* Use overlapping loads to avoid branches.  */
-	leaq	-8(%rdi, %rdx), %rdi
-	leaq	-8(%rsi, %rdx), %rsi
+	leaq	-8(%rdi, %rdx, CHAR_SIZE), %rdi
+	leaq	-8(%rsi, %rdx, CHAR_SIZE), %rsi
 	vmovq	(%rdi), %XMM1
 	vmovq	(%rsi), %XMM2
-	VPCMPEQ %XMM1, %XMM2, %k2
-	kmovw	%k2, %eax
-	subl    $XMM_MASK, %eax
-	jnz	L(first_vec)
+	VPCMP	$4, %XMM1, %XMM2, %k1
+	kmovd	%k1, %eax
+	testl	%eax, %eax
+	jnz	L(return_vec_0)
 	ret
 
 	.p2align 4
-L(between_16_31):
-	/* From 16 to 31 bytes.  No branch when size == 16.  */
-	VMOVU	(%rsi), %XMM2
-	VPCMPEQ (%rdi), %XMM2, %k2
-	kmovw	%k2, %eax
-	subl    $XMM_MASK, %eax
-	jnz	L(first_vec)
-
-	/* Use overlapping loads to avoid branches.  */
-	leaq	-16(%rdi, %rdx), %rdi
-	leaq	-16(%rsi, %rdx), %rsi
-	VMOVU	(%rsi), %XMM2
-	VPCMPEQ (%rdi), %XMM2, %k2
-	kmovw	%k2, %eax
-	subl    $XMM_MASK, %eax
-	jnz	L(first_vec)
+L(zero):
+	xorl	%eax, %eax
 	ret
 
 	.p2align 4
-L(more_8x_vec):
-	/* More than 8 * VEC.  Check the first VEC.  */
-	VMOVU	(%rsi), %YMM2
-	VPCMPEQ (%rdi), %YMM2, %k2
-	kmovd	%k2, %eax
-	subl    $VEC_MASK, %eax
-	jnz	L(first_vec)
-
-	/* Align the first memory area for aligned loads in the loop.
-	   Compute how much the first memory area is misaligned.  */
-	movq	%rdi, %rcx
-	andl	$(VEC_SIZE - 1), %ecx
-	/* Get the negative of offset for alignment.  */
-	subq	$VEC_SIZE, %rcx
-	/* Adjust the second memory area.  */
-	subq	%rcx, %rsi
-	/* Adjust the first memory area which should be aligned now.  */
-	subq	%rcx, %rdi
-	/* Adjust length.  */
-	addq	%rcx, %rdx
-
-L(loop_4x_vec):
-	/* Compare 4 * VEC at a time forward.  */
-	VMOVU	(%rsi), %YMM1
-	VPCMPEQ (%rdi), %YMM1, %k1
-
-	VMOVU	VEC_SIZE(%rsi), %YMM2
-	VPCMPEQ VEC_SIZE(%rdi), %YMM2, %k2
-	kandd	%k2, %k1, %k5
-
-	VMOVU	(VEC_SIZE * 2)(%rsi), %YMM3
-	VPCMPEQ (VEC_SIZE * 2)(%rdi), %YMM3, %k3
-	kandd	%k3, %k5, %k5
-
-	VMOVU	(VEC_SIZE * 3)(%rsi), %YMM4
-	VPCMPEQ (VEC_SIZE * 3)(%rdi), %YMM4, %k4
-	kandd	%k4, %k5, %k5
-
-	kmovd	%k5, %eax
-	cmpl	$VEC_MASK, %eax
-	jne	L(4x_vec_end)
-
-	addq	$(VEC_SIZE * 4), %rdi
-	addq	$(VEC_SIZE * 4), %rsi
-
-	subq	$(VEC_SIZE * 4), %rdx
-	cmpq	$(VEC_SIZE * 4), %rdx
-	jae	L(loop_4x_vec)
-
-	/* Less than 4 * VEC.  */
-	cmpq	$VEC_SIZE, %rdx
-	jbe	L(last_vec)
-	cmpq	$(VEC_SIZE * 2), %rdx
-	jbe	L(last_2x_vec)
-
-L(last_4x_vec):
-	/* From 2 * VEC to 4 * VEC. */
-	VMOVU	(%rsi), %YMM2
-	VPCMPEQ (%rdi), %YMM2, %k2
-	kmovd	%k2, %eax
-	subl    $VEC_MASK, %eax
-	jnz	L(first_vec)
-
-	addq	$VEC_SIZE, %rdi
-	addq	$VEC_SIZE, %rsi
-	VMOVU	(%rsi), %YMM2
-	VPCMPEQ (%rdi), %YMM2, %k2
-	kmovd	%k2, %eax
-	subl    $VEC_MASK, %eax
-	jnz	L(first_vec)
+L(between_16_31):
+	/* From 16 to 31 bytes.  No branch when size == 16.  */
+	VMOVU	(%rsi), %XMM2
+	VPCMP	$4, (%rdi), %XMM2, %k1
+	kmovd	%k1, %eax
+	testl	%eax, %eax
+	jnz	L(return_vec_0)
 
 	/* Use overlapping loads to avoid branches.  */
-	leaq	-(3 * VEC_SIZE)(%rdi, %rdx), %rdi
-	leaq	-(3 * VEC_SIZE)(%rsi, %rdx), %rsi
-	VMOVU	(%rsi), %YMM2
-	VPCMPEQ (%rdi), %YMM2, %k2
-	kmovd	%k2, %eax
-	subl    $VEC_MASK, %eax
-	jnz	L(first_vec)
 
-	addq	$VEC_SIZE, %rdi
-	addq	$VEC_SIZE, %rsi
-	VMOVU	(%rsi), %YMM2
-	VPCMPEQ (%rdi), %YMM2, %k2
-	kmovd	%k2, %eax
-	subl    $VEC_MASK, %eax
-	jnz	L(first_vec)
-	ret
-
-	.p2align 4
-L(4x_vec_end):
+	VMOVU	-16(%rsi, %rdx, CHAR_SIZE), %XMM2
+	leaq	-16(%rdi, %rdx, CHAR_SIZE), %rdi
+	leaq	-16(%rsi, %rdx, CHAR_SIZE), %rsi
+	VPCMP	$4, (%rdi), %XMM2, %k1
 	kmovd	%k1, %eax
-	subl	$VEC_MASK, %eax
-	jnz	L(first_vec)
-	kmovd	%k2, %eax
-	subl	$VEC_MASK, %eax
-	jnz	L(first_vec_x1)
-	kmovd	%k3, %eax
-	subl	$VEC_MASK, %eax
-	jnz	L(first_vec_x2)
-	kmovd	%k4, %eax
-	subl	$VEC_MASK, %eax
-	tzcntl	%eax, %ecx
-# ifdef USE_AS_WMEMCMP
-	xorl	%eax, %eax
-	movl	(VEC_SIZE * 3)(%rdi, %rcx, 4), %edx
-	cmpl	(VEC_SIZE * 3)(%rsi, %rcx, 4), %edx
-	jmp	L(wmemcmp_return)
-# else
-	movzbl	(VEC_SIZE * 3)(%rdi, %rcx), %eax
-	movzbl	(VEC_SIZE * 3)(%rsi, %rcx), %edx
-	sub	%edx, %eax
-# endif
+	testl	%eax, %eax
+	jnz	L(return_vec_0)
 	ret
 
-	.p2align 4
-L(first_vec_x1):
-	tzcntl	%eax, %ecx
 # ifdef USE_AS_WMEMCMP
-	xorl	%eax, %eax
-	movl	VEC_SIZE(%rdi, %rcx, 4), %edx
-	cmpl	VEC_SIZE(%rsi, %rcx, 4), %edx
-	jmp	L(wmemcmp_return)
-# else
-	movzbl	VEC_SIZE(%rdi, %rcx), %eax
-	movzbl	VEC_SIZE(%rsi, %rcx), %edx
-	sub	%edx, %eax
-# endif
+	.p2align 4
+L(one_or_less):
+	jb	L(zero)
+	movl	(%rdi), %ecx
+	xorl	%edx, %edx
+	cmpl	(%rsi), %ecx
+	je	L(zero)
+	setg	%dl
+	leal	-1(%rdx, %rdx), %eax
 	ret
+# else
 
 	.p2align 4
-L(first_vec_x2):
-	tzcntl	%eax, %ecx
-# ifdef USE_AS_WMEMCMP
-	xorl	%eax, %eax
-	movl	(VEC_SIZE * 2)(%rdi, %rcx, 4), %edx
-	cmpl	(VEC_SIZE * 2)(%rsi, %rcx, 4), %edx
-	jmp	L(wmemcmp_return)
-# else
-	movzbl	(VEC_SIZE * 2)(%rdi, %rcx), %eax
-	movzbl	(VEC_SIZE * 2)(%rsi, %rcx), %edx
-	sub	%edx, %eax
-# endif
+L(between_4_7):
+	/* Load as big endian with overlapping movbe to avoid branches.
+	 */
+	movbe	(%rdi), %eax
+	movbe	(%rsi), %ecx
+	shlq	$32, %rax
+	shlq	$32, %rcx
+	movbe	-4(%rdi, %rdx), %edi
+	movbe	-4(%rsi, %rdx), %esi
+	orq	%rdi, %rax
+	orq	%rsi, %rcx
+	subq	%rcx, %rax
+	jz	L(zero_4_7)
+	sbbl	%eax, %eax
+	orl	$1, %eax
+L(zero_4_7):
 	ret
+# endif
+
 END (MEMCMP)
 #endif