x86: Optimize strcmp-avx2.S

Optimization are primarily to the loop logic and how the page cross
logic interacts with the loop.

The page cross logic is at times more expensive for short strings near
the end of a page but not crossing the page. This is done to retest
the page cross conditions with a non-faulty check and to improve the
logic for entering the loop afterwards. This is only particular cases,
however, and is general made up for by more than 10x improvements on
the transition from the page cross -> loop case.

The non-page cross cases are improved most for smaller sizes [0, 128]
and go about even for (128, 4096]. The loop page cross logic is
improved so some more significant speedup is seen there as well.

test-strcmp, test-strncmp, test-wcscmp, and test-wcsncmp all pass.

Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com>

1 files changed, 939 insertions, 651 deletions

diff --git a/sysdeps/x86_64/multiarch/strcmp-avx2.S b/sysdeps/x86_64/multiarch/strcmp-avx2.S
index 9c73b5899d..28d6a0025a 100644
--- a/sysdeps/x86_64/multiarch/strcmp-avx2.S
+++ b/sysdeps/x86_64/multiarch/strcmp-avx2.S
@@ -26,35 +26,57 @@
 
 # define PAGE_SIZE	4096
 
-/* VEC_SIZE = Number of bytes in a ymm register */
+	/* VEC_SIZE = Number of bytes in a ymm register.  */
 # define VEC_SIZE	32
 
-/* Shift for dividing by (VEC_SIZE * 4).  */
-# define DIVIDE_BY_VEC_4_SHIFT	7
-# if (VEC_SIZE * 4) != (1 << DIVIDE_BY_VEC_4_SHIFT)
-#  error (VEC_SIZE * 4) != (1 << DIVIDE_BY_VEC_4_SHIFT)
-# endif
+# define VMOVU	vmovdqu
+# define VMOVA	vmovdqa
 
 # ifdef USE_AS_WCSCMP
-/* Compare packed dwords.  */
+	/* Compare packed dwords.  */
 #  define VPCMPEQ	vpcmpeqd
-/* Compare packed dwords and store minimum.  */
+	/* Compare packed dwords and store minimum.  */
 #  define VPMINU	vpminud
-/* 1 dword char == 4 bytes.  */
+	/* 1 dword char == 4 bytes.  */
 #  define SIZE_OF_CHAR	4
 # else
-/* Compare packed bytes.  */
+	/* Compare packed bytes.  */
 #  define VPCMPEQ	vpcmpeqb
-/* Compare packed bytes and store minimum.  */
+	/* Compare packed bytes and store minimum.  */
 #  define VPMINU	vpminub
-/* 1 byte char == 1 byte.  */
+	/* 1 byte char == 1 byte.  */
 #  define SIZE_OF_CHAR	1
 # endif
 
+# ifdef USE_AS_STRNCMP
+#  define LOOP_REG	r9d
+#  define LOOP_REG64	r9
+
+#  define OFFSET_REG8	r9b
+#  define OFFSET_REG	r9d
+#  define OFFSET_REG64	r9
+# else
+#  define LOOP_REG	edx
+#  define LOOP_REG64	rdx
+
+#  define OFFSET_REG8	dl
+#  define OFFSET_REG	edx
+#  define OFFSET_REG64	rdx
+# endif
+
 # ifndef VZEROUPPER
 #  define VZEROUPPER	vzeroupper
 # endif
 
+# if defined USE_AS_STRNCMP
+#  define VEC_OFFSET	0
+# else
+#  define VEC_OFFSET	(-VEC_SIZE)
+# endif
+
+# define xmmZERO	xmm15
+# define ymmZERO	ymm15
+
 # ifndef SECTION
 #  define SECTION(p)	p##.avx
 # endif
@@ -79,783 +101,1049 @@
    the maximum offset is reached before a difference is found, zero is
    returned.  */
 
-	.section SECTION(.text),"ax",@progbits
-ENTRY (STRCMP)
+	.section SECTION(.text), "ax", @progbits
+ENTRY(STRCMP)
 # ifdef USE_AS_STRNCMP
-	/* Check for simple cases (0 or 1) in offset.  */
+#  ifdef __ILP32__
+	/* Clear the upper 32 bits.  */
+	movl	%edx, %rdx
+#  endif
 	cmp	$1, %RDX_LP
-	je	L(char0)
-	jb	L(zero)
+	/* Signed comparison intentional. We use this branch to also
+	   test cases where length >= 2^63. These very large sizes can be
+	   handled with strcmp as there is no way for that length to
+	   actually bound the buffer.  */
+	jle	L(one_or_less)
 #  ifdef USE_AS_WCSCMP
-#  ifndef __ILP32__
 	movq	%rdx, %rcx
-	/* Check if length could overflow when multiplied by
-	   sizeof(wchar_t). Checking top 8 bits will cover all potential
-	   overflow cases as well as redirect cases where its impossible to
-	   length to bound a valid memory region. In these cases just use
-	   'wcscmp'.  */
+
+	/* Multiplying length by sizeof(wchar_t) can result in overflow.
+	   Check if that is possible. All cases where overflow are possible
+	   are cases where length is large enough that it can never be a
+	   bound on valid memory so just use wcscmp.  */
 	shrq	$56, %rcx
 	jnz	__wcscmp_avx2
+
+	leaq	(, %rdx, 4), %rdx
 #  endif
-	/* Convert units: from wide to byte char.  */
-	shl	$2, %RDX_LP
-#  endif
-	/* Register %r11 tracks the maximum offset.  */
-	mov	%RDX_LP, %R11_LP
 # endif
+	vpxor	%xmmZERO, %xmmZERO, %xmmZERO
 	movl	%edi, %eax
-	xorl	%edx, %edx
-	/* Make %xmm7 (%ymm7) all zeros in this function.  */
-	vpxor	%xmm7, %xmm7, %xmm7
 	orl	%esi, %eax
-	andl	$(PAGE_SIZE - 1), %eax
-	cmpl	$(PAGE_SIZE - (VEC_SIZE * 4)), %eax
-	jg	L(cross_page)
-	/* Start comparing 4 vectors.  */
-	vmovdqu	(%rdi), %ymm1
-	VPCMPEQ	(%rsi), %ymm1, %ymm0
-	VPMINU	%ymm1, %ymm0, %ymm0
-	VPCMPEQ	%ymm7, %ymm0, %ymm0
-	vpmovmskb %ymm0, %ecx
-	testl	%ecx, %ecx
-	je	L(next_3_vectors)
-	tzcntl	%ecx, %edx
+	sall	$20, %eax
+	/* Check if s1 or s2 may cross a page  in next 4x VEC loads.  */
+	cmpl	$((PAGE_SIZE -(VEC_SIZE * 4)) << 20), %eax
+	ja	L(page_cross)
+
+L(no_page_cross):
+	/* Safe to compare 4x vectors.  */
+	VMOVU	(%rdi), %ymm0
+	/* 1s where s1 and s2 equal.  */
+	VPCMPEQ	(%rsi), %ymm0, %ymm1
+	/* 1s at null CHAR.  */
+	VPCMPEQ	%ymm0, %ymmZERO, %ymm2
+	/* 1s where s1 and s2 equal AND not null CHAR.  */
+	vpandn	%ymm1, %ymm2, %ymm1
+
+	/* All 1s -> keep going, any 0s -> return.  */
+	vpmovmskb %ymm1, %ecx
 # ifdef USE_AS_STRNCMP
-	/* Return 0 if the mismatched index (%rdx) is after the maximum
-	   offset (%r11).   */
-	cmpq	%r11, %rdx
-	jae	L(zero)
+	cmpq	$VEC_SIZE, %rdx
+	jbe	L(vec_0_test_len)
 # endif
+
+	/* All 1s represents all equals. incl will overflow to zero in
+	   all equals case. Otherwise 1s will carry until position of first
+	   mismatch.  */
+	incl	%ecx
+	jz	L(more_3x_vec)
+
+	.p2align 4,, 4
+L(return_vec_0):
+	tzcntl	%ecx, %ecx
 # ifdef USE_AS_WCSCMP
+	movl	(%rdi, %rcx), %edx
 	xorl	%eax, %eax
-	movl	(%rdi, %rdx), %ecx
-	cmpl	(%rsi, %rdx), %ecx
-	je	L(return)
-L(wcscmp_return):
+	cmpl	(%rsi, %rcx), %edx
+	je	L(ret0)
 	setl	%al
 	negl	%eax
 	orl	$1, %eax
-L(return):
 # else
-	movzbl	(%rdi, %rdx), %eax
-	movzbl	(%rsi, %rdx), %edx
-	subl	%edx, %eax
+	movzbl	(%rdi, %rcx), %eax
+	movzbl	(%rsi, %rcx), %ecx
+	subl	%ecx, %eax
 # endif
+L(ret0):
 L(return_vzeroupper):
 	ZERO_UPPER_VEC_REGISTERS_RETURN
 
-	.p2align 4
-L(return_vec_size):
-	tzcntl	%ecx, %edx
 # ifdef USE_AS_STRNCMP
-	/* Return 0 if the mismatched index (%rdx + VEC_SIZE) is after
-	   the maximum offset (%r11).  */
-	addq	$VEC_SIZE, %rdx
-	cmpq	%r11, %rdx
-	jae	L(zero)
-#  ifdef USE_AS_WCSCMP
+	.p2align 4,, 8
+L(vec_0_test_len):
+	notl	%ecx
+	bzhil	%edx, %ecx, %eax
+	jnz	L(return_vec_0)
+	/* Align if will cross fetch block.  */
+	.p2align 4,, 2
+L(ret_zero):
 	xorl	%eax, %eax
-	movl	(%rdi, %rdx), %ecx
-	cmpl	(%rsi, %rdx), %ecx
-	jne	L(wcscmp_return)
-#  else
-	movzbl	(%rdi, %rdx), %eax
-	movzbl	(%rsi, %rdx), %edx
-	subl	%edx, %eax
-#  endif
-# else
+	VZEROUPPER_RETURN
+
+	.p2align 4,, 5
+L(one_or_less):
+	jb	L(ret_zero)
 #  ifdef USE_AS_WCSCMP
+	/* 'nbe' covers the case where length is negative (large
+	   unsigned).  */
+	jnbe	__wcscmp_avx2
+	movl	(%rdi), %edx
 	xorl	%eax, %eax
-	movl	VEC_SIZE(%rdi, %rdx), %ecx
-	cmpl	VEC_SIZE(%rsi, %rdx), %ecx
-	jne	L(wcscmp_return)
+	cmpl	(%rsi), %edx
+	je	L(ret1)
+	setl	%al
+	negl	%eax
+	orl	$1, %eax
 #  else
-	movzbl	VEC_SIZE(%rdi, %rdx), %eax
-	movzbl	VEC_SIZE(%rsi, %rdx), %edx
-	subl	%edx, %eax
+	/* 'nbe' covers the case where length is negative (large
+	   unsigned).  */
+
+	jnbe	__strcmp_avx2
+	movzbl	(%rdi), %eax
+	movzbl	(%rsi), %ecx
+	subl	%ecx, %eax
 #  endif
+L(ret1):
+	ret
 # endif
-	VZEROUPPER_RETURN
 
-	.p2align 4
-L(return_2_vec_size):
-	tzcntl	%ecx, %edx
+	.p2align 4,, 10
+L(return_vec_1):
+	tzcntl	%ecx, %ecx
 # ifdef USE_AS_STRNCMP
-	/* Return 0 if the mismatched index (%rdx + 2 * VEC_SIZE) is
-	   after the maximum offset (%r11).  */
-	addq	$(VEC_SIZE * 2), %rdx
-	cmpq	%r11, %rdx
-	jae	L(zero)
-#  ifdef USE_AS_WCSCMP
+	/* rdx must be > CHAR_PER_VEC so save to subtract w.o fear of
+	   overflow.  */
+	addq	$-VEC_SIZE, %rdx
+	cmpq	%rcx, %rdx
+	jbe	L(ret_zero)
+# endif
+# ifdef USE_AS_WCSCMP
+	movl	VEC_SIZE(%rdi, %rcx), %edx
 	xorl	%eax, %eax
-	movl	(%rdi, %rdx), %ecx
-	cmpl	(%rsi, %rdx), %ecx
-	jne	L(wcscmp_return)
-#  else
-	movzbl	(%rdi, %rdx), %eax
-	movzbl	(%rsi, %rdx), %edx
-	subl	%edx, %eax
-#  endif
+	cmpl	VEC_SIZE(%rsi, %rcx), %edx
+	je	L(ret2)
+	setl	%al
+	negl	%eax
+	orl	$1, %eax
 # else
-#  ifdef USE_AS_WCSCMP
-	xorl	%eax, %eax
-	movl	(VEC_SIZE * 2)(%rdi, %rdx), %ecx
-	cmpl	(VEC_SIZE * 2)(%rsi, %rdx), %ecx
-	jne	L(wcscmp_return)
-#  else
-	movzbl	(VEC_SIZE * 2)(%rdi, %rdx), %eax
-	movzbl	(VEC_SIZE * 2)(%rsi, %rdx), %edx
-	subl	%edx, %eax
-#  endif
+	movzbl	VEC_SIZE(%rdi, %rcx), %eax
+	movzbl	VEC_SIZE(%rsi, %rcx), %ecx
+	subl	%ecx, %eax
 # endif
+L(ret2):
 	VZEROUPPER_RETURN
 
-	.p2align 4
-L(return_3_vec_size):
-	tzcntl	%ecx, %edx
+	.p2align 4,, 10
 # ifdef USE_AS_STRNCMP
-	/* Return 0 if the mismatched index (%rdx + 3 * VEC_SIZE) is
-	   after the maximum offset (%r11).  */
-	addq	$(VEC_SIZE * 3), %rdx
-	cmpq	%r11, %rdx
-	jae	L(zero)
-#  ifdef USE_AS_WCSCMP
+L(return_vec_3):
+	salq	$32, %rcx
+# endif
+
+L(return_vec_2):
+# ifndef USE_AS_STRNCMP
+	tzcntl	%ecx, %ecx
+# else
+	tzcntq	%rcx, %rcx
+	cmpq	%rcx, %rdx
+	jbe	L(ret_zero)
+# endif
+
+# ifdef USE_AS_WCSCMP
+	movl	(VEC_SIZE * 2)(%rdi, %rcx), %edx
 	xorl	%eax, %eax
-	movl	(%rdi, %rdx), %ecx
-	cmpl	(%rsi, %rdx), %ecx
-	jne	L(wcscmp_return)
-#  else
-	movzbl	(%rdi, %rdx), %eax
-	movzbl	(%rsi, %rdx), %edx
-	subl	%edx, %eax
-#  endif
+	cmpl	(VEC_SIZE * 2)(%rsi, %rcx), %edx
+	je	L(ret3)
+	setl	%al
+	negl	%eax
+	orl	$1, %eax
 # else
+	movzbl	(VEC_SIZE * 2)(%rdi, %rcx), %eax
+	movzbl	(VEC_SIZE * 2)(%rsi, %rcx), %ecx
+	subl	%ecx, %eax
+# endif
+L(ret3):
+	VZEROUPPER_RETURN
+
+# ifndef USE_AS_STRNCMP
+	.p2align 4,, 10
+L(return_vec_3):
+	tzcntl	%ecx, %ecx
 #  ifdef USE_AS_WCSCMP
+	movl	(VEC_SIZE * 3)(%rdi, %rcx), %edx
 	xorl	%eax, %eax
-	movl	(VEC_SIZE * 3)(%rdi, %rdx), %ecx
-	cmpl	(VEC_SIZE * 3)(%rsi, %rdx), %ecx
-	jne	L(wcscmp_return)
+	cmpl	(VEC_SIZE * 3)(%rsi, %rcx), %edx
+	je	L(ret4)
+	setl	%al
+	negl	%eax
+	orl	$1, %eax
 #  else
-	movzbl	(VEC_SIZE * 3)(%rdi, %rdx), %eax
-	movzbl	(VEC_SIZE * 3)(%rsi, %rdx), %edx
-	subl	%edx, %eax
+	movzbl	(VEC_SIZE * 3)(%rdi, %rcx), %eax
+	movzbl	(VEC_SIZE * 3)(%rsi, %rcx), %ecx
+	subl	%ecx, %eax
 #  endif
-# endif
+L(ret4):
 	VZEROUPPER_RETURN
+# endif
+
+	.p2align 4,, 10
+L(more_3x_vec):
+	/* Safe to compare 4x vectors.  */
+	VMOVU	VEC_SIZE(%rdi), %ymm0
+	VPCMPEQ	VEC_SIZE(%rsi), %ymm0, %ymm1
+	VPCMPEQ	%ymm0, %ymmZERO, %ymm2
+	vpandn	%ymm1, %ymm2, %ymm1
+	vpmovmskb %ymm1, %ecx
+	incl	%ecx
+	jnz	L(return_vec_1)
+
+# ifdef USE_AS_STRNCMP
+	subq	$(VEC_SIZE * 2), %rdx
+	jbe	L(ret_zero)
+# endif
+
+	VMOVU	(VEC_SIZE * 2)(%rdi), %ymm0
+	VPCMPEQ	(VEC_SIZE * 2)(%rsi), %ymm0, %ymm1
+	VPCMPEQ	%ymm0, %ymmZERO, %ymm2
+	vpandn	%ymm1, %ymm2, %ymm1
+	vpmovmskb %ymm1, %ecx
+	incl	%ecx
+	jnz	L(return_vec_2)
+
+	VMOVU	(VEC_SIZE * 3)(%rdi), %ymm0
+	VPCMPEQ	(VEC_SIZE * 3)(%rsi), %ymm0, %ymm1
+	VPCMPEQ	%ymm0, %ymmZERO, %ymm2
+	vpandn	%ymm1, %ymm2, %ymm1
+	vpmovmskb %ymm1, %ecx
+	incl	%ecx
+	jnz	L(return_vec_3)
 
-	.p2align 4
-L(next_3_vectors):
-	vmovdqu	VEC_SIZE(%rdi), %ymm6
-	VPCMPEQ	VEC_SIZE(%rsi), %ymm6, %ymm3
-	VPMINU	%ymm6, %ymm3, %ymm3
-	VPCMPEQ	%ymm7, %ymm3, %ymm3
-	vpmovmskb %ymm3, %ecx
-	testl	%ecx, %ecx
-	jne	L(return_vec_size)
-	vmovdqu	(VEC_SIZE * 2)(%rdi), %ymm5
-	vmovdqu	(VEC_SIZE * 3)(%rdi), %ymm4
-	vmovdqu	(VEC_SIZE * 3)(%rsi), %ymm0
-	VPCMPEQ	(VEC_SIZE * 2)(%rsi), %ymm5, %ymm2
-	VPMINU	%ymm5, %ymm2, %ymm2
-	VPCMPEQ	%ymm4, %ymm0, %ymm0
-	VPCMPEQ	%ymm7, %ymm2, %ymm2
-	vpmovmskb %ymm2, %ecx
-	testl	%ecx, %ecx
-	jne	L(return_2_vec_size)
-	VPMINU	%ymm4, %ymm0, %ymm0
-	VPCMPEQ	%ymm7, %ymm0, %ymm0
-	vpmovmskb %ymm0, %ecx
-	testl	%ecx, %ecx
-	jne	L(return_3_vec_size)
-L(main_loop_header):
-	leaq	(VEC_SIZE * 4)(%rdi), %rdx
-	movl	$PAGE_SIZE, %ecx
-	/* Align load via RAX.  */
-	andq	$-(VEC_SIZE * 4), %rdx
-	subq	%rdi, %rdx
-	leaq	(%rdi, %rdx), %rax
 # ifdef USE_AS_STRNCMP
-	/* Starting from this point, the maximum offset, or simply the
-	   'offset', DECREASES by the same amount when base pointers are
-	   moved forward.  Return 0 when:
-	     1) On match: offset <= the matched vector index.
-	     2) On mistmach, offset is before the mistmatched index.
+	cmpq	$(VEC_SIZE * 2), %rdx
+	jbe	L(ret_zero)
+# endif
+
+# ifdef USE_AS_WCSCMP
+	/* any non-zero positive value that doesn't inference with 0x1.
 	 */
-	subq	%rdx, %r11
-	jbe	L(zero)
-# endif
-	addq	%rsi, %rdx
-	movq	%rdx, %rsi
-	andl	$(PAGE_SIZE - 1), %esi
-	/* Number of bytes before page crossing.  */
-	subq	%rsi, %rcx
-	/* Number of VEC_SIZE * 4 blocks before page crossing.  */
-	shrq	$DIVIDE_BY_VEC_4_SHIFT, %rcx
-	/* ESI: Number of VEC_SIZE * 4 blocks before page crossing.   */
-	movl	%ecx, %esi
-	jmp	L(loop_start)
+	movl	$2, %r8d
 
+# else
+	xorl	%r8d, %r8d
+# endif
+
+	/* The prepare labels are various entry points from the page
+	   cross logic.  */
+L(prepare_loop):
+
+# ifdef USE_AS_STRNCMP
+	/* Store N + (VEC_SIZE * 4) and place check at the begining of
+	   the loop.  */
+	leaq	(VEC_SIZE * 2)(%rdi, %rdx), %rdx
+# endif
+L(prepare_loop_no_len):
+
+	/* Align s1 and adjust s2 accordingly.  */
+	subq	%rdi, %rsi
+	andq	$-(VEC_SIZE * 4), %rdi
+	addq	%rdi, %rsi
+
+# ifdef USE_AS_STRNCMP
+	subq	%rdi, %rdx
+# endif
+
+L(prepare_loop_aligned):
+	/* eax stores distance from rsi to next page cross. These cases
+	   need to be handled specially as the 4x loop could potentially
+	   read memory past the length of s1 or s2 and across a page
+	   boundary.  */
+	movl	$-(VEC_SIZE * 4), %eax
+	subl	%esi, %eax
+	andl	$(PAGE_SIZE - 1), %eax
+
+	/* Loop 4x comparisons at a time.  */
 	.p2align 4
 L(loop):
+
+	/* End condition for strncmp.  */
 # ifdef USE_AS_STRNCMP
-	/* Base pointers are moved forward by 4 * VEC_SIZE.  Decrease
-	   the maximum offset (%r11) by the same amount.  */
-	subq	$(VEC_SIZE * 4), %r11
-	jbe	L(zero)
-# endif
-	addq	$(VEC_SIZE * 4), %rax
-	addq	$(VEC_SIZE * 4), %rdx
-L(loop_start):
-	testl	%esi, %esi
-	leal	-1(%esi), %esi
-	je	L(loop_cross_page)
-L(back_to_loop):
-	/* Main loop, comparing 4 vectors are a time.  */
-	vmovdqa	(%rax), %ymm0
-	vmovdqa	VEC_SIZE(%rax), %ymm3
-	VPCMPEQ	(%rdx), %ymm0, %ymm4
-	VPCMPEQ	VEC_SIZE(%rdx), %ymm3, %ymm1
-	VPMINU	%ymm0, %ymm4, %ymm4
-	VPMINU	%ymm3, %ymm1, %ymm1
-	vmovdqa	(VEC_SIZE * 2)(%rax), %ymm2
-	VPMINU	%ymm1, %ymm4, %ymm0
-	vmovdqa	(VEC_SIZE * 3)(%rax), %ymm3
-	VPCMPEQ	(VEC_SIZE * 2)(%rdx), %ymm2, %ymm5
-	VPCMPEQ	(VEC_SIZE * 3)(%rdx), %ymm3, %ymm6
-	VPMINU	%ymm2, %ymm5, %ymm5
-	VPMINU	%ymm3, %ymm6, %ymm6
-	VPMINU	%ymm5, %ymm0, %ymm0
-	VPMINU	%ymm6, %ymm0, %ymm0
-	VPCMPEQ	%ymm7, %ymm0, %ymm0
-
-	/* Test each mask (32 bits) individually because for VEC_SIZE
-	   == 32 is not possible to OR the four masks and keep all bits
-	   in a 64-bit integer register, differing from SSE2 strcmp
-	   where ORing is possible.  */
-	vpmovmskb %ymm0, %ecx
+	subq	$(VEC_SIZE * 4), %rdx
+	jbe	L(ret_zero)
+# endif
+
+	subq	$-(VEC_SIZE * 4), %rdi
+	subq	$-(VEC_SIZE * 4), %rsi
+
+	/* Check if rsi loads will cross a page boundary.  */
+	addl	$-(VEC_SIZE * 4), %eax
+	jnb	L(page_cross_during_loop)
+
+	/* Loop entry after handling page cross during loop.  */
+L(loop_skip_page_cross_check):
+	VMOVA	(VEC_SIZE * 0)(%rdi), %ymm0
+	VMOVA	(VEC_SIZE * 1)(%rdi), %ymm2
+	VMOVA	(VEC_SIZE * 2)(%rdi), %ymm4
+	VMOVA	(VEC_SIZE * 3)(%rdi), %ymm6
+
+	/* ymm1 all 1s where s1 and s2 equal. All 0s otherwise.  */
+	VPCMPEQ	(VEC_SIZE * 0)(%rsi), %ymm0, %ymm1
+
+	VPCMPEQ	(VEC_SIZE * 1)(%rsi), %ymm2, %ymm3
+	VPCMPEQ	(VEC_SIZE * 2)(%rsi), %ymm4, %ymm5
+	VPCMPEQ	(VEC_SIZE * 3)(%rsi), %ymm6, %ymm7
+
+
+	/* If any mismatches or null CHAR then 0 CHAR, otherwise non-
+	   zero.  */
+	vpand	%ymm0, %ymm1, %ymm1
+
+
+	vpand	%ymm2, %ymm3, %ymm3
+	vpand	%ymm4, %ymm5, %ymm5
+	vpand	%ymm6, %ymm7, %ymm7
+
+	VPMINU	%ymm1, %ymm3, %ymm3
+	VPMINU	%ymm5, %ymm7, %ymm7
+
+	/* Reduce all 0 CHARs for the 4x VEC into ymm7.  */
+	VPMINU	%ymm3, %ymm7, %ymm7
+
+	/* If any 0 CHAR then done.  */
+	VPCMPEQ	%ymm7, %ymmZERO, %ymm7
+	vpmovmskb %ymm7, %LOOP_REG
+	testl	%LOOP_REG, %LOOP_REG
+	jz	L(loop)
+
+	/* Find which VEC has the mismatch of end of string.  */
+	VPCMPEQ	%ymm1, %ymmZERO, %ymm1
+	vpmovmskb %ymm1, %ecx
 	testl	%ecx, %ecx
-	je	L(loop)
-	VPCMPEQ	%ymm7, %ymm4, %ymm0
-	vpmovmskb %ymm0, %edi
-	testl	%edi, %edi
-	je	L(test_vec)
-	tzcntl	%edi, %ecx
+	jnz	L(return_vec_0_end)
+
+
+	VPCMPEQ	%ymm3, %ymmZERO, %ymm3
+	vpmovmskb %ymm3, %ecx
+	testl	%ecx, %ecx
+	jnz	L(return_vec_1_end)
+
+L(return_vec_2_3_end):
 # ifdef USE_AS_STRNCMP
-	cmpq	%rcx, %r11
-	jbe	L(zero)
-#  ifdef USE_AS_WCSCMP
-	movq	%rax, %rsi
+	subq	$(VEC_SIZE * 2), %rdx
+	jbe	L(ret_zero_end)
+# endif
+
+	VPCMPEQ	%ymm5, %ymmZERO, %ymm5
+	vpmovmskb %ymm5, %ecx
+	testl	%ecx, %ecx
+	jnz	L(return_vec_2_end)
+
+	/* LOOP_REG contains matches for null/mismatch from the loop. If
+	   VEC 0,1,and 2 all have no null and no mismatches then mismatch
+	   must entirely be from VEC 3 which is fully represented by
+	   LOOP_REG.  */
+	tzcntl	%LOOP_REG, %LOOP_REG
+
+# ifdef USE_AS_STRNCMP
+	subl	$-(VEC_SIZE), %LOOP_REG
+	cmpq	%LOOP_REG64, %rdx
+	jbe	L(ret_zero_end)
+# endif
+
+# ifdef USE_AS_WCSCMP
+	movl	(VEC_SIZE * 2 - VEC_OFFSET)(%rdi, %LOOP_REG64), %ecx
 	xorl	%eax, %eax
-	movl	(%rsi, %rcx), %edi
-	cmpl	(%rdx, %rcx), %edi
-	jne	L(wcscmp_return)
-#  else
-	movzbl	(%rax, %rcx), %eax
-	movzbl	(%rdx, %rcx), %edx
-	subl	%edx, %eax
-#  endif
+	cmpl	(VEC_SIZE * 2 - VEC_OFFSET)(%rsi, %LOOP_REG64), %ecx
+	je	L(ret5)
+	setl	%al
+	negl	%eax
+	xorl	%r8d, %eax
 # else
-#  ifdef USE_AS_WCSCMP
-	movq	%rax, %rsi
-	xorl	%eax, %eax
-	movl	(%rsi, %rcx), %edi
-	cmpl	(%rdx, %rcx), %edi
-	jne	L(wcscmp_return)
-#  else
-	movzbl	(%rax, %rcx), %eax
-	movzbl	(%rdx, %rcx), %edx
-	subl	%edx, %eax
-#  endif
+	movzbl	(VEC_SIZE * 2 - VEC_OFFSET)(%rdi, %LOOP_REG64), %eax
+	movzbl	(VEC_SIZE * 2 - VEC_OFFSET)(%rsi, %LOOP_REG64), %ecx
+	subl	%ecx, %eax
+	xorl	%r8d, %eax
+	subl	%r8d, %eax
 # endif
+L(ret5):
 	VZEROUPPER_RETURN
 
-	.p2align 4
-L(test_vec):
 # ifdef USE_AS_STRNCMP
-	/* The first vector matched.  Return 0 if the maximum offset
-	   (%r11) <= VEC_SIZE.  */
-	cmpq	$VEC_SIZE, %r11
-	jbe	L(zero)
+	.p2align 4,, 2
+L(ret_zero_end):
+	xorl	%eax, %eax
+	VZEROUPPER_RETURN
 # endif
-	VPCMPEQ	%ymm7, %ymm1, %ymm1
-	vpmovmskb %ymm1, %ecx
-	testl	%ecx, %ecx
-	je	L(test_2_vec)
-	tzcntl	%ecx, %edi
+
+
+	/* The L(return_vec_N_end) differ from L(return_vec_N) in that
+	   they use the value of `r8` to negate the return value. This is
+	   because the page cross logic can swap `rdi` and `rsi`.  */
+	.p2align 4,, 10
 # ifdef USE_AS_STRNCMP
-	addq	$VEC_SIZE, %rdi
-	cmpq	%rdi, %r11
-	jbe	L(zero)
-#  ifdef USE_AS_WCSCMP
-	movq	%rax, %rsi
+L(return_vec_1_end):
+	salq	$32, %rcx
+# endif
+L(return_vec_0_end):
+# ifndef USE_AS_STRNCMP
+	tzcntl	%ecx, %ecx
+# else
+	tzcntq	%rcx, %rcx
+	cmpq	%rcx, %rdx
+	jbe	L(ret_zero_end)
+# endif
+
+# ifdef USE_AS_WCSCMP
+	movl	(%rdi, %rcx), %edx
 	xorl	%eax, %eax
-	movl	(%rsi, %rdi), %ecx
-	cmpl	(%rdx, %rdi), %ecx
-	jne	L(wcscmp_return)
-#  else
-	movzbl	(%rax, %rdi), %eax
-	movzbl	(%rdx, %rdi), %edx
-	subl	%edx, %eax
-#  endif
+	cmpl	(%rsi, %rcx), %edx
+	je	L(ret6)
+	setl	%al
+	negl	%eax
+	xorl	%r8d, %eax
 # else
+	movzbl	(%rdi, %rcx), %eax
+	movzbl	(%rsi, %rcx), %ecx
+	subl	%ecx, %eax
+	xorl	%r8d, %eax
+	subl	%r8d, %eax
+# endif
+L(ret6):
+	VZEROUPPER_RETURN
+
+# ifndef USE_AS_STRNCMP
+	.p2align 4,, 10
+L(return_vec_1_end):
+	tzcntl	%ecx, %ecx
 #  ifdef USE_AS_WCSCMP
-	movq	%rax, %rsi
+	movl	VEC_SIZE(%rdi, %rcx), %edx
 	xorl	%eax, %eax
-	movl	VEC_SIZE(%rsi, %rdi), %ecx
-	cmpl	VEC_SIZE(%rdx, %rdi), %ecx
-	jne	L(wcscmp_return)
+	cmpl	VEC_SIZE(%rsi, %rcx), %edx
+	je	L(ret7)
+	setl	%al
+	negl	%eax
+	xorl	%r8d, %eax
 #  else
-	movzbl	VEC_SIZE(%rax, %rdi), %eax
-	movzbl	VEC_SIZE(%rdx, %rdi), %edx
-	subl	%edx, %eax
+	movzbl	VEC_SIZE(%rdi, %rcx), %eax
+	movzbl	VEC_SIZE(%rsi, %rcx), %ecx
+	subl	%ecx, %eax
+	xorl	%r8d, %eax
+	subl	%r8d, %eax
 #  endif
-# endif
+L(ret7):
 	VZEROUPPER_RETURN
+# endif
 
-	.p2align 4
-L(test_2_vec):
+	.p2align 4,, 10
+L(return_vec_2_end):
+	tzcntl	%ecx, %ecx
 # ifdef USE_AS_STRNCMP
-	/* The first 2 vectors matched.  Return 0 if the maximum offset
-	   (%r11) <= 2 * VEC_SIZE.  */
-	cmpq	$(VEC_SIZE * 2), %r11
-	jbe	L(zero)
+	cmpq	%rcx, %rdx
+	jbe	L(ret_zero_page_cross)
 # endif
-	VPCMPEQ	%ymm7, %ymm5, %ymm5
-	vpmovmskb %ymm5, %ecx
-	testl	%ecx, %ecx
-	je	L(test_3_vec)
-	tzcntl	%ecx, %edi
-# ifdef USE_AS_STRNCMP
-	addq	$(VEC_SIZE * 2), %rdi
-	cmpq	%rdi, %r11
-	jbe	L(zero)
-#  ifdef USE_AS_WCSCMP
-	movq	%rax, %rsi
+# ifdef USE_AS_WCSCMP
+	movl	(VEC_SIZE * 2)(%rdi, %rcx), %edx
 	xorl	%eax, %eax
-	movl	(%rsi, %rdi), %ecx
-	cmpl	(%rdx, %rdi), %ecx
-	jne	L(wcscmp_return)
-#  else
-	movzbl	(%rax, %rdi), %eax
-	movzbl	(%rdx, %rdi), %edx
-	subl	%edx, %eax
-#  endif
+	cmpl	(VEC_SIZE * 2)(%rsi, %rcx), %edx
+	je	L(ret11)
+	setl	%al
+	negl	%eax
+	xorl	%r8d, %eax
 # else
-#  ifdef USE_AS_WCSCMP
-	movq	%rax, %rsi
-	xorl	%eax, %eax
-	movl	(VEC_SIZE * 2)(%rsi, %rdi), %ecx
-	cmpl	(VEC_SIZE * 2)(%rdx, %rdi), %ecx
-	jne	L(wcscmp_return)
-#  else
-	movzbl	(VEC_SIZE * 2)(%rax, %rdi), %eax
-	movzbl	(VEC_SIZE * 2)(%rdx, %rdi), %edx
-	subl	%edx, %eax
-#  endif
+	movzbl	(VEC_SIZE * 2)(%rdi, %rcx), %eax
+	movzbl	(VEC_SIZE * 2)(%rsi, %rcx), %ecx
+	subl	%ecx, %eax
+	xorl	%r8d, %eax
+	subl	%r8d, %eax
 # endif
+L(ret11):
 	VZEROUPPER_RETURN
 
-	.p2align 4
-L(test_3_vec):
+
+	/* Page cross in rsi in next 4x VEC.  */
+
+	/* TODO: Improve logic here.  */
+	.p2align 4,, 10
+L(page_cross_during_loop):
+	/* eax contains [distance_from_page - (VEC_SIZE * 4)].  */
+
+	/* Optimistically rsi and rdi and both aligned inwhich case we
+	   don't need any logic here.  */
+	cmpl	$-(VEC_SIZE * 4), %eax
+	/* Don't adjust eax before jumping back to loop and we will
+	   never hit page cross case again.  */
+	je	L(loop_skip_page_cross_check)
+
+	/* Check if we can safely load a VEC.  */
+	cmpl	$-(VEC_SIZE * 3), %eax
+	jle	L(less_1x_vec_till_page_cross)
+
+	VMOVA	(%rdi), %ymm0
+	VPCMPEQ	(%rsi), %ymm0, %ymm1
+	VPCMPEQ	%ymm0, %ymmZERO, %ymm2
+	vpandn	%ymm1, %ymm2, %ymm1
+	vpmovmskb %ymm1, %ecx
+	incl	%ecx
+	jnz	L(return_vec_0_end)
+
+	/* if distance >= 2x VEC then eax > -(VEC_SIZE * 2).  */
+	cmpl	$-(VEC_SIZE * 2), %eax
+	jg	L(more_2x_vec_till_page_cross)
+
+	.p2align 4,, 4
+L(less_1x_vec_till_page_cross):
+	subl	$-(VEC_SIZE * 4), %eax
+	/* Guranteed safe to read from rdi - VEC_SIZE here. The only
+	   concerning case is first iteration if incoming s1 was near start
+	   of a page and s2 near end. If s1 was near the start of the page
+	   we already aligned up to nearest VEC_SIZE * 4 so gurnateed safe
+	   to read back -VEC_SIZE. If rdi is truly at the start of a page
+	   here, it means the previous page (rdi - VEC_SIZE) has already
+	   been loaded earlier so must be valid.  */
+	VMOVU	-VEC_SIZE(%rdi, %rax), %ymm0
+	VPCMPEQ	-VEC_SIZE(%rsi, %rax), %ymm0, %ymm1
+	VPCMPEQ	%ymm0, %ymmZERO, %ymm2
+	vpandn	%ymm1, %ymm2, %ymm1
+	vpmovmskb %ymm1, %ecx
+
+	/* Mask of potentially valid bits. The lower bits can be out of
+	   range comparisons (but safe regarding page crosses).  */
+	movl	$-1, %r10d
+	shlxl	%esi, %r10d, %r10d
+	notl	%ecx
+
 # ifdef USE_AS_STRNCMP
-	/* The first 3 vectors matched.  Return 0 if the maximum offset
-	   (%r11) <= 3 * VEC_SIZE.  */
-	cmpq	$(VEC_SIZE * 3), %r11
-	jbe	L(zero)
-# endif
-	VPCMPEQ	%ymm7, %ymm6, %ymm6
-	vpmovmskb %ymm6, %esi
-	tzcntl	%esi, %ecx
+	cmpq	%rax, %rdx
+	jbe	L(return_page_cross_end_check)
+# endif
+	movl	%eax, %OFFSET_REG
+	addl	$(PAGE_SIZE - VEC_SIZE * 4), %eax
+
+	andl	%r10d, %ecx
+	jz	L(loop_skip_page_cross_check)
+
+	.p2align 4,, 3
+L(return_page_cross_end):
+	tzcntl	%ecx, %ecx
+
 # ifdef USE_AS_STRNCMP
-	addq	$(VEC_SIZE * 3), %rcx
-	cmpq	%rcx, %r11
-	jbe	L(zero)
-#  ifdef USE_AS_WCSCMP
-	movq	%rax, %rsi
-	xorl	%eax, %eax
-	movl	(%rsi, %rcx), %esi
-	cmpl	(%rdx, %rcx), %esi
-	jne	L(wcscmp_return)
-#  else
-	movzbl	(%rax, %rcx), %eax
-	movzbl	(%rdx, %rcx), %edx
-	subl	%edx, %eax
-#  endif
+	leal	-VEC_SIZE(%OFFSET_REG64, %rcx), %ecx
+L(return_page_cross_cmp_mem):
 # else
-#  ifdef USE_AS_WCSCMP
-	movq	%rax, %rsi
+	addl	%OFFSET_REG, %ecx
+# endif
+# ifdef USE_AS_WCSCMP
+	movl	VEC_OFFSET(%rdi, %rcx), %edx
 	xorl	%eax, %eax
-	movl	(VEC_SIZE * 3)(%rsi, %rcx), %esi
-	cmpl	(VEC_SIZE * 3)(%rdx, %rcx), %esi
-	jne	L(wcscmp_return)
-#  else
-	movzbl	(VEC_SIZE * 3)(%rax, %rcx), %eax
-	movzbl	(VEC_SIZE * 3)(%rdx, %rcx), %edx
-	subl	%edx, %eax
-#  endif
+	cmpl	VEC_OFFSET(%rsi, %rcx), %edx
+	je	L(ret8)
+	setl	%al
+	negl	%eax
+	xorl	%r8d, %eax
+# else
+	movzbl	VEC_OFFSET(%rdi, %rcx), %eax
+	movzbl	VEC_OFFSET(%rsi, %rcx), %ecx
+	subl	%ecx, %eax
+	xorl	%r8d, %eax
+	subl	%r8d, %eax
 # endif
+L(ret8):
 	VZEROUPPER_RETURN
 
-	.p2align 4
-L(loop_cross_page):
-	xorl	%r10d, %r10d
-	movq	%rdx, %rcx
-	/* Align load via RDX.  We load the extra ECX bytes which should
-	   be ignored.  */
-	andl	$((VEC_SIZE * 4) - 1), %ecx
-	/* R10 is -RCX.  */
-	subq	%rcx, %r10
-
-	/* This works only if VEC_SIZE * 2 == 64. */
-# if (VEC_SIZE * 2) != 64
-#  error (VEC_SIZE * 2) != 64
-# endif
-
-	/* Check if the first VEC_SIZE * 2 bytes should be ignored.  */
-	cmpl	$(VEC_SIZE * 2), %ecx
-	jge	L(loop_cross_page_2_vec)
-
-	vmovdqu	(%rax, %r10), %ymm2
-	vmovdqu	VEC_SIZE(%rax, %r10), %ymm3
-	VPCMPEQ	(%rdx, %r10), %ymm2, %ymm0
-	VPCMPEQ	VEC_SIZE(%rdx, %r10), %ymm3, %ymm1
-	VPMINU	%ymm2, %ymm0, %ymm0
-	VPMINU	%ymm3, %ymm1, %ymm1
-	VPCMPEQ	%ymm7, %ymm0, %ymm0
-	VPCMPEQ	%ymm7, %ymm1, %ymm1
-
-	vpmovmskb %ymm0, %edi
-	vpmovmskb %ymm1, %esi
-
-	salq	$32, %rsi
-	xorq	%rsi, %rdi
-
-	/* Since ECX < VEC_SIZE * 2, simply skip the first ECX bytes.  */
-	shrq	%cl, %rdi
-
-	testq	%rdi, %rdi
-	je	L(loop_cross_page_2_vec)
-	tzcntq	%rdi, %rcx
 # ifdef USE_AS_STRNCMP
-	cmpq	%rcx, %r11
-	jbe	L(zero)
-#  ifdef USE_AS_WCSCMP
-	movq	%rax, %rsi
+	.p2align 4,, 10
+L(return_page_cross_end_check):
+	tzcntl	%ecx, %ecx
+	leal	-VEC_SIZE(%rax, %rcx), %ecx
+	cmpl	%ecx, %edx
+	ja	L(return_page_cross_cmp_mem)
 	xorl	%eax, %eax
-	movl	(%rsi, %rcx), %edi
-	cmpl	(%rdx, %rcx), %edi
-	jne	L(wcscmp_return)
-#  else
-	movzbl	(%rax, %rcx), %eax
-	movzbl	(%rdx, %rcx), %edx
-	subl	%edx, %eax
-#  endif
-# else
-#  ifdef USE_AS_WCSCMP
-	movq	%rax, %rsi
-	xorl	%eax, %eax
-	movl	(%rsi, %rcx), %edi
-	cmpl	(%rdx, %rcx), %edi
-	jne	L(wcscmp_return)
-#  else
-	movzbl	(%rax, %rcx), %eax
-	movzbl	(%rdx, %rcx), %edx
-	subl	%edx, %eax
-#  endif
-# endif
 	VZEROUPPER_RETURN
+# endif
 
-	.p2align 4
-L(loop_cross_page_2_vec):
-	/* The first VEC_SIZE * 2 bytes match or are ignored.  */
-	vmovdqu	(VEC_SIZE * 2)(%rax, %r10), %ymm2
-	vmovdqu	(VEC_SIZE * 3)(%rax, %r10), %ymm3
-	VPCMPEQ	(VEC_SIZE * 2)(%rdx, %r10), %ymm2, %ymm5
-	VPMINU	%ymm2, %ymm5, %ymm5
-	VPCMPEQ	(VEC_SIZE * 3)(%rdx, %r10), %ymm3, %ymm6
-	VPCMPEQ	%ymm7, %ymm5, %ymm5
-	VPMINU	%ymm3, %ymm6, %ymm6
-	VPCMPEQ	%ymm7, %ymm6, %ymm6
-
-	vpmovmskb %ymm5, %edi
-	vpmovmskb %ymm6, %esi
-
-	salq	$32, %rsi
-	xorq	%rsi, %rdi
 
-	xorl	%r8d, %r8d
-	/* If ECX > VEC_SIZE * 2, skip ECX - (VEC_SIZE * 2) bytes.  */
-	subl	$(VEC_SIZE * 2), %ecx
-	jle	1f
-	/* Skip ECX bytes.  */
-	shrq	%cl, %rdi
-	/* R8 has number of bytes skipped.  */
-	movl	%ecx, %r8d
-1:
-	/* Before jumping back to the loop, set ESI to the number of
-	   VEC_SIZE * 4 blocks before page crossing.  */
-	movl	$(PAGE_SIZE / (VEC_SIZE * 4) - 1), %esi
-
-	testq	%rdi, %rdi
+	.p2align 4,, 10
+L(more_2x_vec_till_page_cross):
+	/* If more 2x vec till cross we will complete a full loop
+	   iteration here.  */
+
+	VMOVU	VEC_SIZE(%rdi), %ymm0
+	VPCMPEQ	VEC_SIZE(%rsi), %ymm0, %ymm1
+	VPCMPEQ	%ymm0, %ymmZERO, %ymm2
+	vpandn	%ymm1, %ymm2, %ymm1
+	vpmovmskb %ymm1, %ecx
+	incl	%ecx
+	jnz	L(return_vec_1_end)
+
 # ifdef USE_AS_STRNCMP
-	/* At this point, if %rdi value is 0, it already tested
-	   VEC_SIZE*4+%r10 byte starting from %rax. This label
-	   checks whether strncmp maximum offset reached or not.  */
-	je	L(string_nbyte_offset_check)
-# else
-	je	L(back_to_loop)
+	cmpq	$(VEC_SIZE * 2), %rdx
+	jbe	L(ret_zero_in_loop_page_cross)
 # endif
-	tzcntq	%rdi, %rcx
-	addq	%r10, %rcx
-	/* Adjust for number of bytes skipped.  */
-	addq	%r8, %rcx
+
+	subl	$-(VEC_SIZE * 4), %eax
+
+	/* Safe to include comparisons from lower bytes.  */
+	VMOVU	-(VEC_SIZE * 2)(%rdi, %rax), %ymm0
+	VPCMPEQ	-(VEC_SIZE * 2)(%rsi, %rax), %ymm0, %ymm1
+	VPCMPEQ	%ymm0, %ymmZERO, %ymm2
+	vpandn	%ymm1, %ymm2, %ymm1
+	vpmovmskb %ymm1, %ecx
+	incl	%ecx
+	jnz	L(return_vec_page_cross_0)
+
+	VMOVU	-(VEC_SIZE * 1)(%rdi, %rax), %ymm0
+	VPCMPEQ	-(VEC_SIZE * 1)(%rsi, %rax), %ymm0, %ymm1
+	VPCMPEQ	%ymm0, %ymmZERO, %ymm2
+	vpandn	%ymm1, %ymm2, %ymm1
+	vpmovmskb %ymm1, %ecx
+	incl	%ecx
+	jnz	L(return_vec_page_cross_1)
+
 # ifdef USE_AS_STRNCMP
-	addq	$(VEC_SIZE * 2), %rcx
-	subq	%rcx, %r11
-	jbe	L(zero)
-#  ifdef USE_AS_WCSCMP
-	movq	%rax, %rsi
+	/* Must check length here as length might proclude reading next
+	   page.  */
+	cmpq	%rax, %rdx
+	jbe	L(ret_zero_in_loop_page_cross)
+# endif
+
+	/* Finish the loop.  */
+	VMOVA	(VEC_SIZE * 2)(%rdi), %ymm4
+	VMOVA	(VEC_SIZE * 3)(%rdi), %ymm6
+
+	VPCMPEQ	(VEC_SIZE * 2)(%rsi), %ymm4, %ymm5
+	VPCMPEQ	(VEC_SIZE * 3)(%rsi), %ymm6, %ymm7
+	vpand	%ymm4, %ymm5, %ymm5
+	vpand	%ymm6, %ymm7, %ymm7
+	VPMINU	%ymm5, %ymm7, %ymm7
+	VPCMPEQ	%ymm7, %ymmZERO, %ymm7
+	vpmovmskb %ymm7, %LOOP_REG
+	testl	%LOOP_REG, %LOOP_REG
+	jnz	L(return_vec_2_3_end)
+
+	/* Best for code size to include ucond-jmp here. Would be faster
+	   if this case is hot to duplicate the L(return_vec_2_3_end) code
+	   as fall-through and have jump back to loop on mismatch
+	   comparison.  */
+	subq	$-(VEC_SIZE * 4), %rdi
+	subq	$-(VEC_SIZE * 4), %rsi
+	addl	$(PAGE_SIZE - VEC_SIZE * 8), %eax
+# ifdef USE_AS_STRNCMP
+	subq	$(VEC_SIZE * 4), %rdx
+	ja	L(loop_skip_page_cross_check)
+L(ret_zero_in_loop_page_cross):
 	xorl	%eax, %eax
-	movl	(%rsi, %rcx), %edi
-	cmpl	(%rdx, %rcx), %edi
-	jne	L(wcscmp_return)
-#  else
-	movzbl	(%rax, %rcx), %eax
-	movzbl	(%rdx, %rcx), %edx
-	subl	%edx, %eax
-#  endif
+	VZEROUPPER_RETURN
 # else
-#  ifdef USE_AS_WCSCMP
-	movq	%rax, %rsi
-	xorl	%eax, %eax
-	movl	(VEC_SIZE * 2)(%rsi, %rcx), %edi
-	cmpl	(VEC_SIZE * 2)(%rdx, %rcx), %edi
-	jne	L(wcscmp_return)
-#  else
-	movzbl	(VEC_SIZE * 2)(%rax, %rcx), %eax
-	movzbl	(VEC_SIZE * 2)(%rdx, %rcx), %edx
-	subl	%edx, %eax
-#  endif
+	jmp	L(loop_skip_page_cross_check)
 # endif
-	VZEROUPPER_RETURN
 
+
+	.p2align 4,, 10
+L(return_vec_page_cross_0):
+	addl	$-VEC_SIZE, %eax
+L(return_vec_page_cross_1):
+	tzcntl	%ecx, %ecx
 # ifdef USE_AS_STRNCMP
-L(string_nbyte_offset_check):
-	leaq	(VEC_SIZE * 4)(%r10), %r10
-	cmpq	%r10, %r11
-	jbe	L(zero)
-	jmp	L(back_to_loop)
+	leal	-VEC_SIZE(%rax, %rcx), %ecx
+	cmpq	%rcx, %rdx
+	jbe	L(ret_zero_in_loop_page_cross)
+# else
+	addl	%eax, %ecx
 # endif
 
-	.p2align 4
-L(cross_page_loop):
-	/* Check one byte/dword at a time.  */
 # ifdef USE_AS_WCSCMP
-	cmpl	%ecx, %eax
+	movl	VEC_OFFSET(%rdi, %rcx), %edx
+	xorl	%eax, %eax
+	cmpl	VEC_OFFSET(%rsi, %rcx), %edx
+	je	L(ret9)
+	setl	%al
+	negl	%eax
+	xorl	%r8d, %eax
 # else
+	movzbl	VEC_OFFSET(%rdi, %rcx), %eax
+	movzbl	VEC_OFFSET(%rsi, %rcx), %ecx
 	subl	%ecx, %eax
+	xorl	%r8d, %eax
+	subl	%r8d, %eax
 # endif
-	jne	L(different)
-	addl	$SIZE_OF_CHAR, %edx
-	cmpl	$(VEC_SIZE * 4), %edx
-	je	L(main_loop_header)
-# ifdef USE_AS_STRNCMP
-	cmpq	%r11, %rdx
-	jae	L(zero)
+L(ret9):
+	VZEROUPPER_RETURN
+
+
+	.p2align 4,, 10
+L(page_cross):
+# ifndef USE_AS_STRNCMP
+	/* If both are VEC aligned we don't need any special logic here.
+	   Only valid for strcmp where stop condition is guranteed to be
+	   reachable by just reading memory.  */
+	testl	$((VEC_SIZE - 1) << 20), %eax
+	jz	L(no_page_cross)
 # endif
+
+	movl	%edi, %eax
+	movl	%esi, %ecx
+	andl	$(PAGE_SIZE - 1), %eax
+	andl	$(PAGE_SIZE - 1), %ecx
+
+	xorl	%OFFSET_REG, %OFFSET_REG
+
+	/* Check which is closer to page cross, s1 or s2.  */
+	cmpl	%eax, %ecx
+	jg	L(page_cross_s2)
+
+	/* The previous page cross check has false positives. Check for
+	   true positive as page cross logic is very expensive.  */
+	subl	$(PAGE_SIZE - VEC_SIZE * 4), %eax
+	jbe	L(no_page_cross)
+
+	/* Set r8 to not interfere with normal return value (rdi and rsi
+	   did not swap).  */
 # ifdef USE_AS_WCSCMP
-	movl	(%rdi, %rdx), %eax
-	movl	(%rsi, %rdx), %ecx
+	/* any non-zero positive value that doesn't inference with 0x1.
+	 */
+	movl	$2, %r8d
 # else
-	movzbl	(%rdi, %rdx), %eax
-	movzbl	(%rsi, %rdx), %ecx
+	xorl	%r8d, %r8d
 # endif
-	/* Check null char.  */
-	testl	%eax, %eax
-	jne	L(cross_page_loop)
-	/* Since %eax == 0, subtract is OK for both SIGNED and UNSIGNED
-	   comparisons.  */
-	subl	%ecx, %eax
-# ifndef USE_AS_WCSCMP
-L(different):
+
+	/* Check if less than 1x VEC till page cross.  */
+	subl	$(VEC_SIZE * 3), %eax
+	jg	L(less_1x_vec_till_page)
+
+	/* If more than 1x VEC till page cross, loop throuh safely
+	   loadable memory until within 1x VEC of page cross.  */
+
+	.p2align 4,, 10
+L(page_cross_loop):
+
+	VMOVU	(%rdi, %OFFSET_REG64), %ymm0
+	VPCMPEQ	(%rsi, %OFFSET_REG64), %ymm0, %ymm1
+	VPCMPEQ	%ymm0, %ymmZERO, %ymm2
+	vpandn	%ymm1, %ymm2, %ymm1
+	vpmovmskb %ymm1, %ecx
+	incl	%ecx
+
+	jnz	L(check_ret_vec_page_cross)
+	addl	$VEC_SIZE, %OFFSET_REG
+# ifdef USE_AS_STRNCMP
+	cmpq	%OFFSET_REG64, %rdx
+	jbe	L(ret_zero_page_cross)
 # endif
-	VZEROUPPER_RETURN
+	addl	$VEC_SIZE, %eax
+	jl	L(page_cross_loop)
+
+	subl	%eax, %OFFSET_REG
+	/* OFFSET_REG has distance to page cross - VEC_SIZE. Guranteed
+	   to not cross page so is safe to load. Since we have already
+	   loaded at least 1 VEC from rsi it is also guranteed to be safe.
+	 */
+
+	VMOVU	(%rdi, %OFFSET_REG64), %ymm0
+	VPCMPEQ	(%rsi, %OFFSET_REG64), %ymm0, %ymm1
+	VPCMPEQ	%ymm0, %ymmZERO, %ymm2
+	vpandn	%ymm1, %ymm2, %ymm1
+	vpmovmskb %ymm1, %ecx
+
+# ifdef USE_AS_STRNCMP
+	leal	VEC_SIZE(%OFFSET_REG64), %eax
+	cmpq	%rax, %rdx
+	jbe	L(check_ret_vec_page_cross2)
+	addq	%rdi, %rdx
+# endif
+	incl	%ecx
+	jz	L(prepare_loop_no_len)
 
+	.p2align 4,, 4
+L(ret_vec_page_cross):
+# ifndef USE_AS_STRNCMP
+L(check_ret_vec_page_cross):
+# endif
+	tzcntl	%ecx, %ecx
+	addl	%OFFSET_REG, %ecx
+L(ret_vec_page_cross_cont):
 # ifdef USE_AS_WCSCMP
-	.p2align 4
-L(different):
-	/* Use movl to avoid modifying EFLAGS.  */
-	movl	$0, %eax
+	movl	(%rdi, %rcx), %edx
+	xorl	%eax, %eax
+	cmpl	(%rsi, %rcx), %edx
+	je	L(ret12)
 	setl	%al
 	negl	%eax
-	orl	$1, %eax
-	VZEROUPPER_RETURN
+	xorl	%r8d, %eax
+# else
+	movzbl	(%rdi, %rcx), %eax
+	movzbl	(%rsi, %rcx), %ecx
+	subl	%ecx, %eax
+	xorl	%r8d, %eax
+	subl	%r8d, %eax
 # endif
+L(ret12):
+	VZEROUPPER_RETURN
 
 # ifdef USE_AS_STRNCMP
-	.p2align 4
-L(zero):
+	.p2align 4,, 10
+L(check_ret_vec_page_cross2):
+	incl	%ecx
+L(check_ret_vec_page_cross):
+	tzcntl	%ecx, %ecx
+	addl	%OFFSET_REG, %ecx
+	cmpq	%rcx, %rdx
+	ja	L(ret_vec_page_cross_cont)
+	.p2align 4,, 2
+L(ret_zero_page_cross):
 	xorl	%eax, %eax
 	VZEROUPPER_RETURN
+# endif
 
-	.p2align 4
-L(char0):
-#  ifdef USE_AS_WCSCMP
-	xorl	%eax, %eax
-	movl	(%rdi), %ecx
-	cmpl	(%rsi), %ecx
-	jne	L(wcscmp_return)
-#  else
-	movzbl	(%rsi), %ecx
-	movzbl	(%rdi), %eax
-	subl	%ecx, %eax
-#  endif
-	VZEROUPPER_RETURN
+	.p2align 4,, 4
+L(page_cross_s2):
+	/* Ensure this is a true page cross.  */
+	subl	$(PAGE_SIZE - VEC_SIZE * 4), %ecx
+	jbe	L(no_page_cross)
+
+
+	movl	%ecx, %eax
+	movq	%rdi, %rcx
+	movq	%rsi, %rdi
+	movq	%rcx, %rsi
+
+	/* set r8 to negate return value as rdi and rsi swapped.  */
+# ifdef USE_AS_WCSCMP
+	movl	$-4, %r8d
+# else
+	movl	$-1, %r8d
 # endif
+	xorl	%OFFSET_REG, %OFFSET_REG
 
-	.p2align 4
-L(last_vector):
-	addq	%rdx, %rdi
-	addq	%rdx, %rsi
+	/* Check if more than 1x VEC till page cross.  */
+	subl	$(VEC_SIZE * 3), %eax
+	jle	L(page_cross_loop)
+
+	.p2align 4,, 6
+L(less_1x_vec_till_page):
+	/* Find largest load size we can use.  */
+	cmpl	$16, %eax
+	ja	L(less_16_till_page)
+
+	VMOVU	(%rdi), %xmm0
+	VPCMPEQ	(%rsi), %xmm0, %xmm1
+	VPCMPEQ	%xmm0, %xmmZERO, %xmm2
+	vpandn	%xmm1, %xmm2, %xmm1
+	vpmovmskb %ymm1, %ecx
+	incw	%cx
+	jnz	L(check_ret_vec_page_cross)
+	movl	$16, %OFFSET_REG
 # ifdef USE_AS_STRNCMP
-	subq	%rdx, %r11
+	cmpq	%OFFSET_REG64, %rdx
+	jbe	L(ret_zero_page_cross_slow_case0)
+	subl	%eax, %OFFSET_REG
+# else
+	/* Explicit check for 16 byte alignment.  */
+	subl	%eax, %OFFSET_REG
+	jz	L(prepare_loop)
 # endif
-	tzcntl	%ecx, %edx
+
+	VMOVU	(%rdi, %OFFSET_REG64), %xmm0
+	VPCMPEQ	(%rsi, %OFFSET_REG64), %xmm0, %xmm1
+	VPCMPEQ	%xmm0, %xmmZERO, %xmm2
+	vpandn	%xmm1, %xmm2, %xmm1
+	vpmovmskb %ymm1, %ecx
+	incw	%cx
+	jnz	L(check_ret_vec_page_cross)
+
 # ifdef USE_AS_STRNCMP
-	cmpq	%r11, %rdx
-	jae	L(zero)
+	addl	$16, %OFFSET_REG
+	subq	%OFFSET_REG64, %rdx
+	jbe	L(ret_zero_page_cross_slow_case0)
+	subq	$-(VEC_SIZE * 4), %rdx
+
+	leaq	-(VEC_SIZE * 4)(%rdi, %OFFSET_REG64), %rdi
+	leaq	-(VEC_SIZE * 4)(%rsi, %OFFSET_REG64), %rsi
+# else
+	leaq	(16 - VEC_SIZE * 4)(%rdi, %OFFSET_REG64), %rdi
+	leaq	(16 - VEC_SIZE * 4)(%rsi, %OFFSET_REG64), %rsi
 # endif
-# ifdef USE_AS_WCSCMP
+	jmp	L(prepare_loop_aligned)
+
+# ifdef USE_AS_STRNCMP
+	.p2align 4,, 2
+L(ret_zero_page_cross_slow_case0):
 	xorl	%eax, %eax
-	movl	(%rdi, %rdx), %ecx
-	cmpl	(%rsi, %rdx), %ecx
-	jne	L(wcscmp_return)
-# else
-	movzbl	(%rdi, %rdx), %eax
-	movzbl	(%rsi, %rdx), %edx
-	subl	%edx, %eax
+	ret
 # endif
-	VZEROUPPER_RETURN
 
-	/* Comparing on page boundary region requires special treatment:
-	   It must done one vector at the time, starting with the wider
-	   ymm vector if possible, if not, with xmm. If fetching 16 bytes
-	   (xmm) still passes the boundary, byte comparison must be done.
-	 */
-	.p2align 4
-L(cross_page):
-	/* Try one ymm vector at a time.  */
-	cmpl	$(PAGE_SIZE - VEC_SIZE), %eax
-	jg	L(cross_page_1_vector)
-L(loop_1_vector):
-	vmovdqu	(%rdi, %rdx), %ymm1
-	VPCMPEQ	(%rsi, %rdx), %ymm1, %ymm0
-	VPMINU	%ymm1, %ymm0, %ymm0
-	VPCMPEQ	%ymm7, %ymm0, %ymm0
-	vpmovmskb %ymm0, %ecx
-	testl	%ecx, %ecx
-	jne	L(last_vector)
 
-	addl	$VEC_SIZE, %edx
+	.p2align 4,, 10
+L(less_16_till_page):
+	/* Find largest load size we can use.  */
+	cmpl	$24, %eax
+	ja	L(less_8_till_page)
 
-	addl	$VEC_SIZE, %eax
-# ifdef USE_AS_STRNCMP
-	/* Return 0 if the current offset (%rdx) >= the maximum offset
-	   (%r11).  */
-	cmpq	%r11, %rdx
-	jae	L(zero)
-# endif
-	cmpl	$(PAGE_SIZE - VEC_SIZE), %eax
-	jle	L(loop_1_vector)
-L(cross_page_1_vector):
-	/* Less than 32 bytes to check, try one xmm vector.  */
-	cmpl	$(PAGE_SIZE - 16), %eax
-	jg	L(cross_page_1_xmm)
-	vmovdqu	(%rdi, %rdx), %xmm1
-	VPCMPEQ	(%rsi, %rdx), %xmm1, %xmm0
-	VPMINU	%xmm1, %xmm0, %xmm0
-	VPCMPEQ	%xmm7, %xmm0, %xmm0
-	vpmovmskb %xmm0, %ecx
-	testl	%ecx, %ecx
-	jne	L(last_vector)
+	vmovq	(%rdi), %xmm0
+	vmovq	(%rsi), %xmm1
+	VPCMPEQ	%xmm0, %xmmZERO, %xmm2
+	VPCMPEQ	%xmm1, %xmm0, %xmm1
+	vpandn	%xmm1, %xmm2, %xmm1
+	vpmovmskb %ymm1, %ecx
+	incb	%cl
+	jnz	L(check_ret_vec_page_cross)
 
-	addl	$16, %edx
-# ifndef USE_AS_WCSCMP
-	addl	$16, %eax
+
+# ifdef USE_AS_STRNCMP
+	cmpq	$8, %rdx
+	jbe	L(ret_zero_page_cross_slow_case0)
 # endif
+	movl	$24, %OFFSET_REG
+	/* Explicit check for 16 byte alignment.  */
+	subl	%eax, %OFFSET_REG
+
+
+
+	vmovq	(%rdi, %OFFSET_REG64), %xmm0
+	vmovq	(%rsi, %OFFSET_REG64), %xmm1
+	VPCMPEQ	%xmm0, %xmmZERO, %xmm2
+	VPCMPEQ	%xmm1, %xmm0, %xmm1
+	vpandn	%xmm1, %xmm2, %xmm1
+	vpmovmskb %ymm1, %ecx
+	incb	%cl
+	jnz	L(check_ret_vec_page_cross)
+
 # ifdef USE_AS_STRNCMP
-	/* Return 0 if the current offset (%rdx) >= the maximum offset
-	   (%r11).  */
-	cmpq	%r11, %rdx
-	jae	L(zero)
-# endif
-
-L(cross_page_1_xmm):
-# ifndef USE_AS_WCSCMP
-	/* Less than 16 bytes to check, try 8 byte vector.  NB: No need
-	   for wcscmp nor wcsncmp since wide char is 4 bytes.   */
-	cmpl	$(PAGE_SIZE - 8), %eax
-	jg	L(cross_page_8bytes)
-	vmovq	(%rdi, %rdx), %xmm1
-	vmovq	(%rsi, %rdx), %xmm0
-	VPCMPEQ	%xmm0, %xmm1, %xmm0
-	VPMINU	%xmm1, %xmm0, %xmm0
-	VPCMPEQ	%xmm7, %xmm0, %xmm0
-	vpmovmskb %xmm0, %ecx
-	/* Only last 8 bits are valid.  */
-	andl	$0xff, %ecx
-	testl	%ecx, %ecx
-	jne	L(last_vector)
+	addl	$8, %OFFSET_REG
+	subq	%OFFSET_REG64, %rdx
+	jbe	L(ret_zero_page_cross_slow_case0)
+	subq	$-(VEC_SIZE * 4), %rdx
 
-	addl	$8, %edx
-	addl	$8, %eax
+	leaq	-(VEC_SIZE * 4)(%rdi, %OFFSET_REG64), %rdi
+	leaq	-(VEC_SIZE * 4)(%rsi, %OFFSET_REG64), %rsi
+# else
+	leaq	(8 - VEC_SIZE * 4)(%rdi, %OFFSET_REG64), %rdi
+	leaq	(8 - VEC_SIZE * 4)(%rsi, %OFFSET_REG64), %rsi
+# endif
+	jmp	L(prepare_loop_aligned)
+
+
+	.p2align 4,, 10
+L(less_8_till_page):
+# ifdef USE_AS_WCSCMP
+	/* If using wchar then this is the only check before we reach
+	   the page boundary.  */
+	movl	(%rdi), %eax
+	movl	(%rsi), %ecx
+	cmpl	%ecx, %eax
+	jnz	L(ret_less_8_wcs)
 #  ifdef USE_AS_STRNCMP
-	/* Return 0 if the current offset (%rdx) >= the maximum offset
-	   (%r11).  */
-	cmpq	%r11, %rdx
-	jae	L(zero)
+	addq	%rdi, %rdx
+	/* We already checked for len <= 1 so cannot hit that case here.
+	 */
 #  endif
+	testl	%eax, %eax
+	jnz	L(prepare_loop_no_len)
+	ret
 
-L(cross_page_8bytes):
-	/* Less than 8 bytes to check, try 4 byte vector.  */
-	cmpl	$(PAGE_SIZE - 4), %eax
-	jg	L(cross_page_4bytes)
-	vmovd	(%rdi, %rdx), %xmm1
-	vmovd	(%rsi, %rdx), %xmm0
-	VPCMPEQ	%xmm0, %xmm1, %xmm0
-	VPMINU	%xmm1, %xmm0, %xmm0
-	VPCMPEQ	%xmm7, %xmm0, %xmm0
-	vpmovmskb %xmm0, %ecx
-	/* Only last 4 bits are valid.  */
-	andl	$0xf, %ecx
-	testl	%ecx, %ecx
-	jne	L(last_vector)
+	.p2align 4,, 8
+L(ret_less_8_wcs):
+	setl	%OFFSET_REG8
+	negl	%OFFSET_REG
+	movl	%OFFSET_REG, %eax
+	xorl	%r8d, %eax
+	ret
+
+# else
+
+	/* Find largest load size we can use.  */
+	cmpl	$28, %eax
+	ja	L(less_4_till_page)
+
+	vmovd	(%rdi), %xmm0
+	vmovd	(%rsi), %xmm1
+	VPCMPEQ	%xmm0, %xmmZERO, %xmm2
+	VPCMPEQ	%xmm1, %xmm0, %xmm1
+	vpandn	%xmm1, %xmm2, %xmm1
+	vpmovmskb %ymm1, %ecx
+	subl	$0xf, %ecx
+	jnz	L(check_ret_vec_page_cross)
 
-	addl	$4, %edx
 #  ifdef USE_AS_STRNCMP
-	/* Return 0 if the current offset (%rdx) >= the maximum offset
-	   (%r11).  */
-	cmpq	%r11, %rdx
-	jae	L(zero)
+	cmpq	$4, %rdx
+	jbe	L(ret_zero_page_cross_slow_case1)
 #  endif
+	movl	$28, %OFFSET_REG
+	/* Explicit check for 16 byte alignment.  */
+	subl	%eax, %OFFSET_REG
 
-L(cross_page_4bytes):
-# endif
-	/* Less than 4 bytes to check, try one byte/dword at a time.  */
-# ifdef USE_AS_STRNCMP
-	cmpq	%r11, %rdx
-	jae	L(zero)
-# endif
-# ifdef USE_AS_WCSCMP
-	movl	(%rdi, %rdx), %eax
-	movl	(%rsi, %rdx), %ecx
-# else
-	movzbl	(%rdi, %rdx), %eax
-	movzbl	(%rsi, %rdx), %ecx
-# endif
-	testl	%eax, %eax
-	jne	L(cross_page_loop)
+
+
+	vmovd	(%rdi, %OFFSET_REG64), %xmm0
+	vmovd	(%rsi, %OFFSET_REG64), %xmm1
+	VPCMPEQ	%xmm0, %xmmZERO, %xmm2
+	VPCMPEQ	%xmm1, %xmm0, %xmm1
+	vpandn	%xmm1, %xmm2, %xmm1
+	vpmovmskb %ymm1, %ecx
+	subl	$0xf, %ecx
+	jnz	L(check_ret_vec_page_cross)
+
+#  ifdef USE_AS_STRNCMP
+	addl	$4, %OFFSET_REG
+	subq	%OFFSET_REG64, %rdx
+	jbe	L(ret_zero_page_cross_slow_case1)
+	subq	$-(VEC_SIZE * 4), %rdx
+
+	leaq	-(VEC_SIZE * 4)(%rdi, %OFFSET_REG64), %rdi
+	leaq	-(VEC_SIZE * 4)(%rsi, %OFFSET_REG64), %rsi
+#  else
+	leaq	(4 - VEC_SIZE * 4)(%rdi, %OFFSET_REG64), %rdi
+	leaq	(4 - VEC_SIZE * 4)(%rsi, %OFFSET_REG64), %rsi
+#  endif
+	jmp	L(prepare_loop_aligned)
+
+#  ifdef USE_AS_STRNCMP
+	.p2align 4,, 2
+L(ret_zero_page_cross_slow_case1):
+	xorl	%eax, %eax
+	ret
+#  endif
+
+	.p2align 4,, 10
+L(less_4_till_page):
+	subq	%rdi, %rsi
+	/* Extremely slow byte comparison loop.  */
+L(less_4_loop):
+	movzbl	(%rdi), %eax
+	movzbl	(%rsi, %rdi), %ecx
 	subl	%ecx, %eax
-	VZEROUPPER_RETURN
-END (STRCMP)
+	jnz	L(ret_less_4_loop)
+	testl	%ecx, %ecx
+	jz	L(ret_zero_4_loop)
+#  ifdef USE_AS_STRNCMP
+	decq	%rdx
+	jz	L(ret_zero_4_loop)
+#  endif
+	incq	%rdi
+	/* end condition is reach page boundary (rdi is aligned).  */
+	testl	$31, %edi
+	jnz	L(less_4_loop)
+	leaq	-(VEC_SIZE * 4)(%rdi, %rsi), %rsi
+	addq	$-(VEC_SIZE * 4), %rdi
+#  ifdef USE_AS_STRNCMP
+	subq	$-(VEC_SIZE * 4), %rdx
+#  endif
+	jmp	L(prepare_loop_aligned)
+
+L(ret_zero_4_loop):
+	xorl	%eax, %eax
+	ret
+L(ret_less_4_loop):
+	xorl	%r8d, %eax
+	subl	%r8d, %eax
+	ret
+# endif
+END(STRCMP)
 #endif