[PATCH v1 2/3] x86: Optimize memcmp-avx2-movbe.S

[Noah Goldstein <goldstein.w.n@gmail.com>]

No bug. This commit optimizes memcmp-avx2.S. The optimizations include
adding a new vec compare path for small sizes, reorganizing the entry
control flow, and removing some unnecissary ALU instructions from the
main loop. test-memcmp and test-wmemcmp are both passing.

Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com>
---
 sysdeps/x86_64/multiarch/ifunc-memcmp.h      |   1 +
 sysdeps/x86_64/multiarch/memcmp-avx2-movbe.S | 676 +++++++++++--------
 2 files changed, 396 insertions(+), 281 deletions(-)

diff --git a/sysdeps/x86_64/multiarch/ifunc-memcmp.h b/sysdeps/x86_64/multiarch/ifunc-memcmp.h
index 8bee1aff75..89e2129968 100644
--- a/sysdeps/x86_64/multiarch/ifunc-memcmp.h
+++ b/sysdeps/x86_64/multiarch/ifunc-memcmp.h
@@ -33,6 +33,7 @@ IFUNC_SELECTOR (void)
 
   if (CPU_FEATURE_USABLE_P (cpu_features, AVX2)
       && CPU_FEATURE_USABLE_P (cpu_features, MOVBE)
+      && CPU_FEATURE_USABLE_P (cpu_features, BMI2)
       && CPU_FEATURES_ARCH_P (cpu_features, AVX_Fast_Unaligned_Load))
     {
       if (CPU_FEATURE_USABLE_P (cpu_features, AVX512VL)
diff --git a/sysdeps/x86_64/multiarch/memcmp-avx2-movbe.S b/sysdeps/x86_64/multiarch/memcmp-avx2-movbe.S
index ad0fa962a1..2621ec907a 100644
--- a/sysdeps/x86_64/multiarch/memcmp-avx2-movbe.S
+++ b/sysdeps/x86_64/multiarch/memcmp-avx2-movbe.S
@@ -19,17 +19,23 @@
 #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 < VEC_SIZE
+      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 * VEC_SIZE 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 * VEC_SIZE or less.
+   8. Use 4 vector compares when size is 4 * VEC_SIZE or less.
+   9. Use 8 vector compares when size is 8 * VEC_SIZE or less.  */
+
 
 # include <sysdep.h>
 
@@ -38,8 +44,10 @@
 # endif
 
 # ifdef USE_AS_WMEMCMP
+#  define CHAR_SIZE	4
 #  define VPCMPEQ	vpcmpeqd
 # else
+#  define CHAR_SIZE	1
 #  define VPCMPEQ	vpcmpeqb
 # endif
 
@@ -52,7 +60,7 @@
 # endif
 
 # define VEC_SIZE 32
-# define VEC_MASK ((1 << VEC_SIZE) - 1)
+# define PAGE_SIZE	4096
 
 /* Warning!
            wmemcmp has to use SIGNED comparison for elements.
@@ -71,136 +79,359 @@ ENTRY (MEMCMP)
 	jb	L(less_vec)
 
 	/* From VEC to 2 * VEC.  No branch when size == VEC_SIZE.  */
-	vmovdqu	(%rsi), %ymm2
-	VPCMPEQ (%rdi), %ymm2, %ymm2
-	vpmovmskb %ymm2, %eax
-	subl    $VEC_MASK, %eax
-	jnz	L(first_vec)
+	vmovdqu	(%rsi), %ymm1
+	VPCMPEQ	(%rdi), %ymm1, %ymm1
+	vpmovmskb %ymm1, %eax
+	/* NB: eax must be destination register if going to
+	   L(return_vec_[0,2]). For L(return_vec_3 destination register
+	   must be ecx.  */
+	incl	%eax
+	jnz	L(return_vec_0)
 
 	cmpq	$(VEC_SIZE * 2), %rdx
-	jbe	L(last_vec)
-
-	VPCMPEQ	%ymm0, %ymm0, %ymm0
-	/* More than 2 * VEC.  */
-	cmpq	$(VEC_SIZE * 8), %rdx
-	ja	L(more_8x_vec)
-	cmpq	$(VEC_SIZE * 4), %rdx
-	jb	L(last_4x_vec)
-
-	/* From 4 * VEC to 8 * VEC, inclusively. */
-	vmovdqu	(%rsi), %ymm1
-	VPCMPEQ (%rdi), %ymm1, %ymm1
+	jbe	L(last_1x_vec)
 
+	/* Check second VEC no matter what.  */
 	vmovdqu	VEC_SIZE(%rsi), %ymm2
-	VPCMPEQ VEC_SIZE(%rdi), %ymm2, %ymm2
+	VPCMPEQ	VEC_SIZE(%rdi), %ymm2, %ymm2
+	vpmovmskb %ymm2, %eax
+	/* If all 4 VEC where equal eax will be all 1s so incl will
+	   overflow and set zero flag.  */
+	incl	%eax
+	jnz	L(return_vec_1)
 
-	vmovdqu	(VEC_SIZE * 2)(%rsi), %ymm3
-	VPCMPEQ (VEC_SIZE * 2)(%rdi), %ymm3, %ymm3
+	/* Less than 4 * VEC.  */
+	cmpq	$(VEC_SIZE * 4), %rdx
+	jbe	L(last_2x_vec)
 
+	/* Check third and fourth VEC no matter what.  */
+	vmovdqu	(VEC_SIZE * 2)(%rsi), %ymm3
+	VPCMPEQ	(VEC_SIZE * 2)(%rdi), %ymm3, %ymm3
+	vpmovmskb %ymm3, %eax
+	incl	%eax
+	jnz	L(return_vec_2)
 	vmovdqu	(VEC_SIZE * 3)(%rsi), %ymm4
-	VPCMPEQ (VEC_SIZE * 3)(%rdi), %ymm4, %ymm4
+	VPCMPEQ	(VEC_SIZE * 3)(%rdi), %ymm4, %ymm4
+	vpmovmskb %ymm4, %ecx
+	incl	%ecx
+	jnz	L(return_vec_3)
 
-	vpand	%ymm1, %ymm2, %ymm5
-	vpand	%ymm3, %ymm4, %ymm6
-	vpand	%ymm5, %ymm6, %ymm5
+	/* Go to 4x VEC loop.  */
+	cmpq	$(VEC_SIZE * 8), %rdx
+	ja	L(more_8x_vec)
 
-	vptest	%ymm0, %ymm5
-	jnc	L(4x_vec_end)
+	/* Handle remainder of size = 4 * VEC + 1 to 8 * VEC without any
+	   branches.  */
 
+	/* Load first two VEC from s2 before adjusting addresses.  */
+	vmovdqu	-(VEC_SIZE * 4)(%rsi, %rdx), %ymm1
+	vmovdqu	-(VEC_SIZE * 3)(%rsi, %rdx), %ymm2
 	leaq	-(4 * VEC_SIZE)(%rdi, %rdx), %rdi
 	leaq	-(4 * VEC_SIZE)(%rsi, %rdx), %rsi
-	vmovdqu	(%rsi), %ymm1
-	VPCMPEQ (%rdi), %ymm1, %ymm1
 
-	vmovdqu	VEC_SIZE(%rsi), %ymm2
-	VPCMPEQ VEC_SIZE(%rdi), %ymm2, %ymm2
-	vpand	%ymm2, %ymm1, %ymm5
+	/* Wait to load from s1 until addressed adjust due to
+	   unlamination of microfusion with complex address mode.  */
+	VPCMPEQ	(%rdi), %ymm1, %ymm1
+	VPCMPEQ	(VEC_SIZE)(%rdi), %ymm2, %ymm2
 
 	vmovdqu	(VEC_SIZE * 2)(%rsi), %ymm3
-	VPCMPEQ (VEC_SIZE * 2)(%rdi), %ymm3, %ymm3
-	vpand	%ymm3, %ymm5, %ymm5
-
+	VPCMPEQ	(VEC_SIZE * 2)(%rdi), %ymm3, %ymm3
 	vmovdqu	(VEC_SIZE * 3)(%rsi), %ymm4
-	VPCMPEQ (VEC_SIZE * 3)(%rdi), %ymm4, %ymm4
-	vpand	%ymm4, %ymm5, %ymm5
+	VPCMPEQ	(VEC_SIZE * 3)(%rdi), %ymm4, %ymm4
 
-	vptest	%ymm0, %ymm5
-	jnc	L(4x_vec_end)
-	xorl	%eax, %eax
+	/* Reduce VEC0 - VEC4.  */
+	vpand	%ymm1, %ymm2, %ymm5
+	vpand	%ymm3, %ymm4, %ymm6
+	vpand	%ymm5, %ymm6, %ymm7
+	vpmovmskb %ymm7, %ecx
+	incl	%ecx
+	jnz	L(return_vec_0_1_2_3)
+	/* NB: eax must be zero to reach here.  */
+	VZEROUPPER_RETURN
+
+	.p2align 4
+L(return_vec_0):
+	tzcntl	%eax, %eax
+# ifdef USE_AS_WMEMCMP
+	movl	(%rdi, %rax), %ecx
+	xorl	%edx, %edx
+	cmpl	(%rsi, %rax), %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
 L(return_vzeroupper):
 	ZERO_UPPER_VEC_REGISTERS_RETURN
 
 	.p2align 4
-L(last_2x_vec):
-	/* From VEC to 2 * VEC.  No branch when size == VEC_SIZE.  */
-	vmovdqu	(%rsi), %ymm2
-	VPCMPEQ (%rdi), %ymm2, %ymm2
-	vpmovmskb %ymm2, %eax
-	subl    $VEC_MASK, %eax
-	jnz	L(first_vec)
+L(return_vec_1):
+	tzcntl	%eax, %eax
+# ifdef USE_AS_WMEMCMP
+	movl	VEC_SIZE(%rdi, %rax), %ecx
+	xorl	%edx, %edx
+	cmpl	VEC_SIZE(%rsi, %rax), %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
+	VZEROUPPER_RETURN
+
+	.p2align 4
+L(return_vec_2):
+	tzcntl	%eax, %eax
+# ifdef USE_AS_WMEMCMP
+	movl	(VEC_SIZE * 2)(%rdi, %rax), %ecx
+	xorl	%edx, %edx
+	cmpl	(VEC_SIZE * 2)(%rsi, %rax), %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
+	VZEROUPPER_RETURN
+
+	/* NB: p2align 5 here to ensure 4x loop is 32 byte aligned.  */
+	.p2align 5
+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):
+	vpmovmskb %ymm1, %eax
+	incl	%eax
+	jnz	L(return_vec_0)
 
-L(last_vec):
-	/* Use overlapping loads to avoid branches.  */
-	leaq	-VEC_SIZE(%rdi, %rdx), %rdi
-	leaq	-VEC_SIZE(%rsi, %rdx), %rsi
-	vmovdqu	(%rsi), %ymm2
-	VPCMPEQ (%rdi), %ymm2, %ymm2
 	vpmovmskb %ymm2, %eax
-	subl    $VEC_MASK, %eax
-	jnz	L(first_vec)
+	incl	%eax
+	jnz	L(return_vec_1)
+
+	vpmovmskb %ymm3, %eax
+	incl	%eax
+	jnz	L(return_vec_2)
+L(return_vec_3):
+	tzcntl	%ecx, %ecx
+# ifdef USE_AS_WMEMCMP
+	movl	(VEC_SIZE * 3)(%rdi, %rcx), %eax
+	xorl	%edx, %edx
+	cmpl	(VEC_SIZE * 3)(%rsi, %rcx), %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
+	VZEROUPPER_RETURN
+
+	.p2align 4
+L(more_8x_vec):
+	/* Set end of s1 in rdx.  */
+	leaq	-(VEC_SIZE * 4)(%rdi, %rdx), %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):
+	/* rsi has s2 - s1 so get correct address by adding s1 (in rdi).
+	 */
+	vmovdqu	(%rsi, %rdi), %ymm1
+	VPCMPEQ	(%rdi), %ymm1, %ymm1
+
+	vmovdqu	VEC_SIZE(%rsi, %rdi), %ymm2
+	VPCMPEQ	VEC_SIZE(%rdi), %ymm2, %ymm2
+
+	vmovdqu	(VEC_SIZE * 2)(%rsi, %rdi), %ymm3
+	VPCMPEQ	(VEC_SIZE * 2)(%rdi), %ymm3, %ymm3
+
+	vmovdqu	(VEC_SIZE * 3)(%rsi, %rdi), %ymm4
+	VPCMPEQ	(VEC_SIZE * 3)(%rdi), %ymm4, %ymm4
+
+	vpand	%ymm1, %ymm2, %ymm5
+	vpand	%ymm3, %ymm4, %ymm6
+	vpand	%ymm5, %ymm6, %ymm7
+	vpmovmskb %ymm7, %ecx
+	incl	%ecx
+	jnz	L(8x_return_vec_0_1_2_3)
+	subq	$-(VEC_SIZE * 4), %rdi
+	/* Check if s1 pointer at end.  */
+	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.  */
+	vmovdqu	(VEC_SIZE * 2)(%rsi, %rdx), %ymm3
+	cmpl	$(VEC_SIZE * 2), %edi
+	jae	L(8x_last_2x_vec)
+
+	/* Check last 4 VEC.  */
+	vmovdqu	(%rsi, %rdx), %ymm1
+	VPCMPEQ	(%rdx), %ymm1, %ymm1
+
+	vmovdqu	VEC_SIZE(%rsi, %rdx), %ymm2
+	VPCMPEQ	VEC_SIZE(%rdx), %ymm2, %ymm2
+
+	VPCMPEQ	(VEC_SIZE * 2)(%rdx), %ymm3, %ymm3
+
+	vmovdqu	(VEC_SIZE * 3)(%rsi, %rdx), %ymm4
+	VPCMPEQ	(VEC_SIZE * 3)(%rdx), %ymm4, %ymm4
+
+	vpand	%ymm1, %ymm2, %ymm5
+	vpand	%ymm3, %ymm4, %ymm6
+	vpand	%ymm5, %ymm6, %ymm7
+	vpmovmskb %ymm7, %ecx
+	/* Restore s1 pointer to rdi.  */
+	movq	%rdx, %rdi
+	incl	%ecx
+	jnz	L(8x_return_vec_0_1_2_3)
+	/* NB: eax must be zero to reach here.  */
+	VZEROUPPER_RETURN
+
+	/* Only entry is from L(more_8x_vec).  */
+	.p2align 4
+L(8x_last_2x_vec):
+	/* Check second to last VEC. rdx store end pointer of s1 and
+	   ymm3 has already been loaded with second to last VEC from s2.
+	 */
+	VPCMPEQ	(VEC_SIZE * 2)(%rdx), %ymm3, %ymm3
+	vpmovmskb %ymm3, %eax
+	incl	%eax
+	jnz	L(8x_return_vec_2)
+	/* Check last VEC.  */
+	.p2align 4
+L(8x_last_1x_vec):
+	vmovdqu	(VEC_SIZE * 3)(%rsi, %rdx), %ymm4
+	VPCMPEQ	(VEC_SIZE * 3)(%rdx), %ymm4, %ymm4
+	vpmovmskb %ymm4, %eax
+	incl	%eax
+	jnz	L(8x_return_vec_3)
 	VZEROUPPER_RETURN
 
 	.p2align 4
-L(first_vec):
-	/* A byte or int32 is different within 16 or 32 bytes.  */
-	tzcntl	%eax, %ecx
+L(last_2x_vec):
+	/* Check second to last VEC.  */
+	vmovdqu	-(VEC_SIZE * 2)(%rsi, %rdx), %ymm1
+	VPCMPEQ	-(VEC_SIZE * 2)(%rdi, %rdx), %ymm1, %ymm1
+	vpmovmskb %ymm1, %eax
+	incl	%eax
+	jnz	L(return_vec_1_end)
+	/* Check last VEC.  */
+L(last_1x_vec):
+	vmovdqu	-(VEC_SIZE * 1)(%rsi, %rdx), %ymm1
+	VPCMPEQ	-(VEC_SIZE * 1)(%rdi, %rdx), %ymm1, %ymm1
+	vpmovmskb %ymm1, %eax
+	incl	%eax
+	jnz	L(return_vec_0_end)
+	VZEROUPPER_RETURN
+
+	.p2align 4
+L(8x_return_vec_2):
+	subq	$VEC_SIZE, %rdx
+L(8x_return_vec_3):
+	tzcntl	%eax, %eax
+	addq	%rdx, %rax
 # ifdef USE_AS_WMEMCMP
-	xorl	%eax, %eax
-	movl	(%rdi, %rcx), %edx
-	cmpl	(%rsi, %rcx), %edx
-L(wmemcmp_return):
-	setl	%al
-	negl	%eax
-	orl	$1, %eax
+	movl	(VEC_SIZE * 3)(%rax), %ecx
+	xorl	%edx, %edx
+	cmpl	(VEC_SIZE * 3)(%rsi, %rax), %ecx
+	setg	%dl
+	leal	-1(%rdx, %rdx), %eax
 # else
-	movzbl	(%rdi, %rcx), %eax
-	movzbl	(%rsi, %rcx), %edx
-	sub	%edx, %eax
+	movzbl	(VEC_SIZE * 3)(%rsi, %rax), %ecx
+	movzbl	(VEC_SIZE * 3)(%rax), %eax
+	subl	%ecx, %eax
 # endif
 	VZEROUPPER_RETURN
 
-# ifdef USE_AS_WMEMCMP
 	.p2align 4
-L(4):
-	xorl	%eax, %eax
-	movl	(%rdi), %edx
-	cmpl	(%rsi), %edx
-	jne	L(wmemcmp_return)
-	ret
+L(return_vec_1_end):
+	tzcntl	%eax, %eax
+	addl	%edx, %eax
+# ifdef USE_AS_WMEMCMP
+	movl	-(VEC_SIZE * 2)(%rdi, %rax), %ecx
+	xorl	%edx, %edx
+	cmpl	-(VEC_SIZE * 2)(%rsi, %rax), %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
+	VZEROUPPER_RETURN
+
 	.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
-	ret
+L(return_vec_0_end):
+	tzcntl	%eax, %eax
+	addl	%edx, %eax
+# ifdef USE_AS_WMEMCMP
+	movl	-VEC_SIZE(%rdi, %rax), %ecx
+	xorl	%edx, %edx
+	cmpl	-VEC_SIZE(%rsi, %rax), %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
+	VZEROUPPER_RETURN
 
 	.p2align 4
-L(exit):
-	ret
+L(less_vec):
+	/* Check if one or less CHAR. This is necessary for size = 0 but
+	   is also faster for size = CHAR_SIZE.  */
+	cmpl	$CHAR_SIZE, %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.  */
+	vmovdqu	(%rsi), %ymm2
+	VPCMPEQ	(%rdi), %ymm2, %ymm2
+	vpmovmskb %ymm2, %eax
+	incl	%eax
+	/* Result will be zero if s1 and s2 match. Otherwise first set
+	   bit will be first mismatch.  */
+	bzhil	%edx, %eax, %edx
+	jnz	L(return_vec_0)
+	xorl	%eax, %eax
+	VZEROUPPER_RETURN
 
 	.p2align 4
-L(between_2_3):
+L(page_cross_less_vec):
+	/* if USE_AS_WMEMCMP it can only be 0, 4, 8, 12, 16, 20, 24, 28
+	   bytes.  */
+	cmpl	$16, %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)
+
 	/* Load as big endian to avoid branches.  */
 	movzwl	(%rdi), %eax
 	movzwl	(%rsi), %ecx
@@ -208,223 +439,106 @@ 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
+	/* No ymm register was touched.  */
 	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
+	/* No ymm register was touched.  */
 	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, %xmm2
+	VPCMPEQ	%xmm1, %xmm2, %xmm2
 	vpmovmskb %xmm2, %eax
-	subl    $0xffff, %eax
-	jnz	L(first_vec)
+	subl	$0xffff, %eax
+	jnz	L(return_vec_0)
 	/* Use overlapping loads to avoid branches.  */
 	leaq	-8(%rdi, %rdx), %rdi
 	leaq	-8(%rsi, %rdx), %rsi
 	vmovq	(%rdi), %xmm1
 	vmovq	(%rsi), %xmm2
-	VPCMPEQ %xmm1, %xmm2, %xmm2
+	VPCMPEQ	%xmm1, %xmm2, %xmm2
 	vpmovmskb %xmm2, %eax
-	subl    $0xffff, %eax
-	jnz	L(first_vec)
+	subl	$0xffff, %eax
+	jnz	L(return_vec_0)
+	/* No ymm register was touched.  */
+	ret
+
+	.p2align 4
+L(zero):
+	xorl	%eax, %eax
 	ret
 
 	.p2align 4
 L(between_16_31):
 	/* From 16 to 31 bytes.  No branch when size == 16.  */
 	vmovdqu	(%rsi), %xmm2
-	VPCMPEQ (%rdi), %xmm2, %xmm2
+	VPCMPEQ	(%rdi), %xmm2, %xmm2
 	vpmovmskb %xmm2, %eax
-	subl    $0xffff, %eax
-	jnz	L(first_vec)
+	subl	$0xffff, %eax
+	jnz	L(return_vec_0)
 
 	/* Use overlapping loads to avoid branches.  */
+
+	vmovdqu	-16(%rsi, %rdx), %xmm2
 	leaq	-16(%rdi, %rdx), %rdi
 	leaq	-16(%rsi, %rdx), %rsi
-	vmovdqu	(%rsi), %xmm2
-	VPCMPEQ (%rdi), %xmm2, %xmm2
+	VPCMPEQ	(%rdi), %xmm2, %xmm2
 	vpmovmskb %xmm2, %eax
-	subl    $0xffff, %eax
-	jnz	L(first_vec)
+	subl	$0xffff, %eax
+	jnz	L(return_vec_0)
+	/* No ymm register was touched.  */
 	ret
 
-	.p2align 4
-L(more_8x_vec):
-	/* More than 8 * VEC.  Check the first VEC.  */
-	vmovdqu	(%rsi), %ymm2
-	VPCMPEQ (%rdi), %ymm2, %ymm2
-	vpmovmskb %ymm2, %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.  */
-	vmovdqu	(%rsi), %ymm1
-	VPCMPEQ (%rdi), %ymm1, %ymm1
-
-	vmovdqu	VEC_SIZE(%rsi), %ymm2
-	VPCMPEQ VEC_SIZE(%rdi), %ymm2, %ymm2
-	vpand	%ymm2, %ymm1, %ymm5
-
-	vmovdqu	(VEC_SIZE * 2)(%rsi), %ymm3
-	VPCMPEQ (VEC_SIZE * 2)(%rdi), %ymm3, %ymm3
-	vpand	%ymm3, %ymm5, %ymm5
-
-	vmovdqu	(VEC_SIZE * 3)(%rsi), %ymm4
-	VPCMPEQ (VEC_SIZE * 3)(%rdi), %ymm4, %ymm4
-	vpand	%ymm4, %ymm5, %ymm5
-
-	vptest	%ymm0, %ymm5
-	jnc	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. */
-	vmovdqu	(%rsi), %ymm2
-	VPCMPEQ (%rdi), %ymm2, %ymm2
-	vpmovmskb %ymm2, %eax
-	subl    $VEC_MASK, %eax
-	jnz	L(first_vec)
-
-	addq	$VEC_SIZE, %rdi
-	addq	$VEC_SIZE, %rsi
-	vmovdqu	(%rsi), %ymm2
-	VPCMPEQ (%rdi), %ymm2, %ymm2
-	vpmovmskb %ymm2, %eax
-	subl    $VEC_MASK, %eax
-	jnz	L(first_vec)
-
-	/* Use overlapping loads to avoid branches.  */
-	leaq	-(3 * VEC_SIZE)(%rdi, %rdx), %rdi
-	leaq	-(3 * VEC_SIZE)(%rsi, %rdx), %rsi
-	vmovdqu	(%rsi), %ymm2
-	VPCMPEQ (%rdi), %ymm2, %ymm2
-	vpmovmskb %ymm2, %eax
-	subl    $VEC_MASK, %eax
-	jnz	L(first_vec)
-
-	addq	$VEC_SIZE, %rdi
-	addq	$VEC_SIZE, %rsi
-	vmovdqu	(%rsi), %ymm2
-	VPCMPEQ (%rdi), %ymm2, %ymm2
-	vpmovmskb %ymm2, %eax
-	subl    $VEC_MASK, %eax
-	jnz	L(first_vec)
-	VZEROUPPER_RETURN
-
-	.p2align 4
-L(4x_vec_end):
-	vpmovmskb %ymm1, %eax
-	subl	$VEC_MASK, %eax
-	jnz	L(first_vec)
-	vpmovmskb %ymm2, %eax
-	subl	$VEC_MASK, %eax
-	jnz	L(first_vec_x1)
-	vpmovmskb %ymm3, %eax
-	subl	$VEC_MASK, %eax
-	jnz	L(first_vec_x2)
-	vpmovmskb %ymm4, %eax
-	subl	$VEC_MASK, %eax
-	tzcntl	%eax, %ecx
 # ifdef USE_AS_WMEMCMP
-	xorl	%eax, %eax
-	movl	(VEC_SIZE * 3)(%rdi, %rcx), %edx
-	cmpl	(VEC_SIZE * 3)(%rsi, %rcx), %edx
-	jmp	L(wmemcmp_return)
-# else
-	movzbl	(VEC_SIZE * 3)(%rdi, %rcx), %eax
-	movzbl	(VEC_SIZE * 3)(%rsi, %rcx), %edx
-	sub	%edx, %eax
-# endif
-	VZEROUPPER_RETURN
-
 	.p2align 4
-L(first_vec_x1):
-	tzcntl	%eax, %ecx
-# ifdef USE_AS_WMEMCMP
-	xorl	%eax, %eax
-	movl	VEC_SIZE(%rdi, %rcx), %edx
-	cmpl	VEC_SIZE(%rsi, %rcx), %edx
-	jmp	L(wmemcmp_return)
+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
+	/* No ymm register was touched.  */
+	ret
 # else
-	movzbl	VEC_SIZE(%rdi, %rcx), %eax
-	movzbl	VEC_SIZE(%rsi, %rcx), %edx
-	sub	%edx, %eax
-# endif
-	VZEROUPPER_RETURN
 
 	.p2align 4
-L(first_vec_x2):
-	tzcntl	%eax, %ecx
-# ifdef USE_AS_WMEMCMP
-	xorl	%eax, %eax
-	movl	(VEC_SIZE * 2)(%rdi, %rcx), %edx
-	cmpl	(VEC_SIZE * 2)(%rsi, %rcx), %edx
-	jmp	L(wmemcmp_return)
-# else
-	movzbl	(VEC_SIZE * 2)(%rdi, %rcx), %eax
-	movzbl	(VEC_SIZE * 2)(%rsi, %rcx), %edx
-	sub	%edx, %eax
+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):
+	/* No ymm register was touched.  */
+	ret
 # endif
-	VZEROUPPER_RETURN
+
 END (MEMCMP)
 #endif
-- 
2.29.2