Blob Blame History Raw
From 93b1c47bd092f8e1444a10b5d6ec20e44d66459a Mon Sep 17 00:00:00 2001
From: Noah Goldstein <goldstein.w.n@gmail.com>
Date: Fri, 23 Apr 2021 15:56:24 -0400
Subject: [PATCH] x86: Optimize strchr-avx2.S

No bug. This commit optimizes strchr-avx2.S. The optimizations are all
small things such as save an ALU in the alignment process, saving a
few instructions in the loop return, saving some bytes in the main
loop, and increasing the ILP in the return cases. test-strchr,
test-strchrnul, test-wcschr, and test-wcschrnul are all passing.

Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com>
(cherry picked from commit ccabe7971f508709d034b63b8672f6f751a3d356)
---
 sysdeps/x86_64/multiarch/strchr-avx2.S | 290 +++++++++++++++----------
 1 file changed, 170 insertions(+), 120 deletions(-)

diff --git a/sysdeps/x86_64/multiarch/strchr-avx2.S b/sysdeps/x86_64/multiarch/strchr-avx2.S
index 919d256c..5884726b 100644
--- a/sysdeps/x86_64/multiarch/strchr-avx2.S
+++ b/sysdeps/x86_64/multiarch/strchr-avx2.S
@@ -49,133 +49,144 @@
 
 	.section SECTION(.text),"ax",@progbits
 ENTRY (STRCHR)
-	movl	%edi, %ecx
-# ifndef USE_AS_STRCHRNUL
-	xorl	%edx, %edx
-# endif
-
 	/* Broadcast CHAR to YMM0.	*/
 	vmovd	%esi, %xmm0
+	movl	%edi, %eax
+	andl	$(PAGE_SIZE - 1), %eax
+	VPBROADCAST	%xmm0, %ymm0
 	vpxor	%xmm9, %xmm9, %xmm9
-	VPBROADCAST %xmm0, %ymm0
 
 	/* Check if we cross page boundary with one vector load.  */
-	andl	$(PAGE_SIZE - 1), %ecx
-	cmpl	$(PAGE_SIZE - VEC_SIZE), %ecx
-	ja  L(cross_page_boundary)
+	cmpl	$(PAGE_SIZE - VEC_SIZE), %eax
+	ja	L(cross_page_boundary)
 
 	/* Check the first VEC_SIZE bytes.	Search for both CHAR and the
 	   null byte.  */
 	vmovdqu	(%rdi), %ymm8
-	VPCMPEQ %ymm8, %ymm0, %ymm1
-	VPCMPEQ %ymm8, %ymm9, %ymm2
+	VPCMPEQ	%ymm8, %ymm0, %ymm1
+	VPCMPEQ	%ymm8, %ymm9, %ymm2
 	vpor	%ymm1, %ymm2, %ymm1
 	vpmovmskb %ymm1, %eax
 	testl	%eax, %eax
-	jz	L(more_vecs)
+	jz	L(aligned_more)
 	tzcntl	%eax, %eax
+# ifndef USE_AS_STRCHRNUL
 	/* Found CHAR or the null byte.	 */
+	cmp	(%rdi, %rax), %CHAR_REG
+	jne	L(zero)
+# endif
 	addq	%rdi, %rax
+	VZEROUPPER_RETURN
+
+	/* .p2align 5 helps keep performance more consistent if ENTRY()
+	   alignment % 32 was either 16 or 0. As well this makes the
+	   alignment % 32 of the loop_4x_vec fixed which makes tuning it
+	   easier.  */
+	.p2align 5
+L(first_vec_x4):
+	tzcntl	%eax, %eax
+	addq	$(VEC_SIZE * 3 + 1), %rdi
 # ifndef USE_AS_STRCHRNUL
-	cmp (%rax), %CHAR_REG
-	cmovne	%rdx, %rax
+	/* Found CHAR or the null byte.	 */
+	cmp	(%rdi, %rax), %CHAR_REG
+	jne	L(zero)
 # endif
-L(return_vzeroupper):
-	ZERO_UPPER_VEC_REGISTERS_RETURN
-
-	.p2align 4
-L(more_vecs):
-	/* Align data for aligned loads in the loop.  */
-	andq	$-VEC_SIZE, %rdi
-L(aligned_more):
-
-	/* Check the next 4 * VEC_SIZE.	 Only one VEC_SIZE at a time
-	   since data is only aligned to VEC_SIZE.	*/
-	vmovdqa	VEC_SIZE(%rdi), %ymm8
-	addq	$VEC_SIZE, %rdi
-	VPCMPEQ %ymm8, %ymm0, %ymm1
-	VPCMPEQ %ymm8, %ymm9, %ymm2
-	vpor	%ymm1, %ymm2, %ymm1
-	vpmovmskb %ymm1, %eax
-	testl	%eax, %eax
-	jnz	L(first_vec_x0)
-
-	vmovdqa	VEC_SIZE(%rdi), %ymm8
-	VPCMPEQ %ymm8, %ymm0, %ymm1
-	VPCMPEQ %ymm8, %ymm9, %ymm2
-	vpor	%ymm1, %ymm2, %ymm1
-	vpmovmskb %ymm1, %eax
-	testl	%eax, %eax
-	jnz	L(first_vec_x1)
-
-	vmovdqa	(VEC_SIZE * 2)(%rdi), %ymm8
-	VPCMPEQ %ymm8, %ymm0, %ymm1
-	VPCMPEQ %ymm8, %ymm9, %ymm2
-	vpor	%ymm1, %ymm2, %ymm1
-	vpmovmskb %ymm1, %eax
-	testl	%eax, %eax
-	jnz	L(first_vec_x2)
-
-	vmovdqa	(VEC_SIZE * 3)(%rdi), %ymm8
-	VPCMPEQ %ymm8, %ymm0, %ymm1
-	VPCMPEQ %ymm8, %ymm9, %ymm2
-	vpor	%ymm1, %ymm2, %ymm1
-	vpmovmskb %ymm1, %eax
-	testl	%eax, %eax
-	jz	L(prep_loop_4x)
+	addq	%rdi, %rax
+	VZEROUPPER_RETURN
 
-	tzcntl	%eax, %eax
-	leaq	(VEC_SIZE * 3)(%rdi, %rax), %rax
 # ifndef USE_AS_STRCHRNUL
-	cmp (%rax), %CHAR_REG
-	cmovne	%rdx, %rax
+L(zero):
+	xorl	%eax, %eax
+	VZEROUPPER_RETURN
 # endif
-	VZEROUPPER
-	ret
+
 
 	.p2align 4
-L(first_vec_x0):
+L(first_vec_x1):
 	tzcntl	%eax, %eax
-	/* Found CHAR or the null byte.	 */
-	addq	%rdi, %rax
+	incq	%rdi
 # ifndef USE_AS_STRCHRNUL
-	cmp (%rax), %CHAR_REG
-	cmovne	%rdx, %rax
+	/* Found CHAR or the null byte.	 */
+	cmp	(%rdi, %rax), %CHAR_REG
+	jne	L(zero)
 # endif
+	addq	%rdi, %rax
 	VZEROUPPER_RETURN
 
 	.p2align 4
-L(first_vec_x1):
+L(first_vec_x2):
 	tzcntl	%eax, %eax
-	leaq	VEC_SIZE(%rdi, %rax), %rax
+	addq	$(VEC_SIZE + 1), %rdi
 # ifndef USE_AS_STRCHRNUL
-	cmp (%rax), %CHAR_REG
-	cmovne	%rdx, %rax
+	/* Found CHAR or the null byte.	 */
+	cmp	(%rdi, %rax), %CHAR_REG
+	jne	L(zero)
 # endif
+	addq	%rdi, %rax
 	VZEROUPPER_RETURN
 
 	.p2align 4
-L(first_vec_x2):
+L(first_vec_x3):
 	tzcntl	%eax, %eax
-	/* Found CHAR or the null byte.	 */
-	leaq	(VEC_SIZE * 2)(%rdi, %rax), %rax
+	addq	$(VEC_SIZE * 2 + 1), %rdi
 # ifndef USE_AS_STRCHRNUL
-	cmp (%rax), %CHAR_REG
-	cmovne	%rdx, %rax
+	/* Found CHAR or the null byte.	 */
+	cmp	(%rdi, %rax), %CHAR_REG
+	jne	L(zero)
 # endif
+	addq	%rdi, %rax
 	VZEROUPPER_RETURN
 
-L(prep_loop_4x):
-	/* Align data to 4 * VEC_SIZE.	*/
-	andq	$-(VEC_SIZE * 4), %rdi
+	.p2align 4
+L(aligned_more):
+	/* Align data to VEC_SIZE - 1. This is the same number of
+	   instructions as using andq -VEC_SIZE but saves 4 bytes of code
+	   on x4 check.  */
+	orq	$(VEC_SIZE - 1), %rdi
+L(cross_page_continue):
+	/* Check the next 4 * VEC_SIZE.  Only one VEC_SIZE at a time
+	   since data is only aligned to VEC_SIZE.  */
+	vmovdqa	1(%rdi), %ymm8
+	VPCMPEQ	%ymm8, %ymm0, %ymm1
+	VPCMPEQ	%ymm8, %ymm9, %ymm2
+	vpor	%ymm1, %ymm2, %ymm1
+	vpmovmskb %ymm1, %eax
+	testl	%eax, %eax
+	jnz	L(first_vec_x1)
+
+	vmovdqa	(VEC_SIZE + 1)(%rdi), %ymm8
+	VPCMPEQ	%ymm8, %ymm0, %ymm1
+	VPCMPEQ	%ymm8, %ymm9, %ymm2
+	vpor	%ymm1, %ymm2, %ymm1
+	vpmovmskb %ymm1, %eax
+	testl	%eax, %eax
+	jnz	L(first_vec_x2)
+
+	vmovdqa	(VEC_SIZE * 2 + 1)(%rdi), %ymm8
+	VPCMPEQ	%ymm8, %ymm0, %ymm1
+	VPCMPEQ	%ymm8, %ymm9, %ymm2
+	vpor	%ymm1, %ymm2, %ymm1
+	vpmovmskb %ymm1, %eax
+	testl	%eax, %eax
+	jnz	L(first_vec_x3)
 
+	vmovdqa	(VEC_SIZE * 3 + 1)(%rdi), %ymm8
+	VPCMPEQ	%ymm8, %ymm0, %ymm1
+	VPCMPEQ	%ymm8, %ymm9, %ymm2
+	vpor	%ymm1, %ymm2, %ymm1
+	vpmovmskb %ymm1, %eax
+	testl	%eax, %eax
+	jnz	L(first_vec_x4)
+	/* Align data to VEC_SIZE * 4 - 1.	*/
+	addq	$(VEC_SIZE * 4 + 1), %rdi
+	andq	$-(VEC_SIZE * 4), %rdi
 	.p2align 4
 L(loop_4x_vec):
 	/* Compare 4 * VEC at a time forward.  */
-	vmovdqa	(VEC_SIZE * 4)(%rdi), %ymm5
-	vmovdqa	(VEC_SIZE * 5)(%rdi), %ymm6
-	vmovdqa	(VEC_SIZE * 6)(%rdi), %ymm7
-	vmovdqa	(VEC_SIZE * 7)(%rdi), %ymm8
+	vmovdqa	(%rdi), %ymm5
+	vmovdqa	(VEC_SIZE)(%rdi), %ymm6
+	vmovdqa	(VEC_SIZE * 2)(%rdi), %ymm7
+	vmovdqa	(VEC_SIZE * 3)(%rdi), %ymm8
 
 	/* Leaves only CHARS matching esi as 0.	 */
 	vpxor	%ymm5, %ymm0, %ymm1
@@ -191,63 +202,102 @@ L(loop_4x_vec):
 	VPMINU	%ymm1, %ymm2, %ymm5
 	VPMINU	%ymm3, %ymm4, %ymm6
 
-	VPMINU	%ymm5, %ymm6, %ymm5
+	VPMINU	%ymm5, %ymm6, %ymm6
 
-	VPCMPEQ %ymm5, %ymm9, %ymm5
-	vpmovmskb %ymm5, %eax
+	VPCMPEQ	%ymm6, %ymm9, %ymm6
+	vpmovmskb %ymm6, %ecx
+	subq	$-(VEC_SIZE * 4), %rdi
+	testl	%ecx, %ecx
+	jz	L(loop_4x_vec)
 
-	addq	$(VEC_SIZE * 4), %rdi
-	testl	%eax, %eax
-	jz  L(loop_4x_vec)
 
-	VPCMPEQ %ymm1, %ymm9, %ymm1
+	VPCMPEQ	%ymm1, %ymm9, %ymm1
 	vpmovmskb %ymm1, %eax
 	testl	%eax, %eax
-	jnz	L(first_vec_x0)
+	jnz	L(last_vec_x0)
+
 
-	VPCMPEQ %ymm2, %ymm9, %ymm2
+	VPCMPEQ	%ymm5, %ymm9, %ymm2
 	vpmovmskb %ymm2, %eax
 	testl	%eax, %eax
-	jnz	L(first_vec_x1)
+	jnz	L(last_vec_x1)
+
+	VPCMPEQ	%ymm3, %ymm9, %ymm3
+	vpmovmskb %ymm3, %eax
+	/* rcx has combined result from all 4 VEC. It will only be used
+	   if the first 3 other VEC all did not contain a match.  */
+	salq	$32, %rcx
+	orq	%rcx, %rax
+	tzcntq	%rax, %rax
+	subq	$(VEC_SIZE * 2), %rdi
+# ifndef USE_AS_STRCHRNUL
+	/* Found CHAR or the null byte.	 */
+	cmp	(%rdi, %rax), %CHAR_REG
+	jne	L(zero_end)
+# endif
+	addq	%rdi, %rax
+	VZEROUPPER_RETURN
+
+
+	.p2align 4
+L(last_vec_x0):
+	tzcntl	%eax, %eax
+	addq	$-(VEC_SIZE * 4), %rdi
+# ifndef USE_AS_STRCHRNUL
+	/* Found CHAR or the null byte.	 */
+	cmp	(%rdi, %rax), %CHAR_REG
+	jne	L(zero_end)
+# endif
+	addq	%rdi, %rax
+	VZEROUPPER_RETURN
 
-	VPCMPEQ %ymm3, %ymm9, %ymm3
-	VPCMPEQ %ymm4, %ymm9, %ymm4
-	vpmovmskb %ymm3, %ecx
-	vpmovmskb %ymm4, %eax
-	salq	$32, %rax
-	orq %rcx, %rax
-	tzcntq  %rax, %rax
-	leaq	(VEC_SIZE * 2)(%rdi, %rax), %rax
 # ifndef USE_AS_STRCHRNUL
-	cmp (%rax), %CHAR_REG
-	cmovne	%rdx, %rax
+L(zero_end):
+	xorl	%eax, %eax
+	VZEROUPPER_RETURN
 # endif
-	VZEROUPPER
-	ret
+
+	.p2align 4
+L(last_vec_x1):
+	tzcntl	%eax, %eax
+	subq	$(VEC_SIZE * 3), %rdi
+# ifndef USE_AS_STRCHRNUL
+	/* Found CHAR or the null byte.	 */
+	cmp	(%rdi, %rax), %CHAR_REG
+	jne	L(zero_end)
+# endif
+	addq	%rdi, %rax
+	VZEROUPPER_RETURN
+
 
 	/* Cold case for crossing page with first load.	 */
 	.p2align 4
 L(cross_page_boundary):
-	andq	$-VEC_SIZE, %rdi
-	andl	$(VEC_SIZE - 1), %ecx
-
-	vmovdqa	(%rdi), %ymm8
-	VPCMPEQ %ymm8, %ymm0, %ymm1
-	VPCMPEQ %ymm8, %ymm9, %ymm2
+	movq	%rdi, %rdx
+	/* Align rdi to VEC_SIZE - 1.  */
+	orq	$(VEC_SIZE - 1), %rdi
+	vmovdqa	-(VEC_SIZE - 1)(%rdi), %ymm8
+	VPCMPEQ	%ymm8, %ymm0, %ymm1
+	VPCMPEQ	%ymm8, %ymm9, %ymm2
 	vpor	%ymm1, %ymm2, %ymm1
 	vpmovmskb %ymm1, %eax
-	/* Remove the leading bits.	 */
-	sarxl	%ecx, %eax, %eax
+	/* Remove the leading bytes. sarxl only uses bits [5:0] of COUNT
+	   so no need to manually mod edx.  */
+	sarxl	%edx, %eax, %eax
 	testl	%eax, %eax
-	jz	L(aligned_more)
+	jz	L(cross_page_continue)
 	tzcntl	%eax, %eax
-	addq	%rcx, %rdi
-	addq	%rdi, %rax
 # ifndef USE_AS_STRCHRNUL
-	cmp (%rax), %CHAR_REG
-	cmovne	%rdx, %rax
+	xorl	%ecx, %ecx
+	/* Found CHAR or the null byte.	 */
+	cmp	(%rdx, %rax), %CHAR_REG
+	leaq	(%rdx, %rax), %rax
+	cmovne	%rcx, %rax
+# else
+	addq	%rdx, %rax
 # endif
-	VZEROUPPER_RETURN
+L(return_vzeroupper):
+	ZERO_UPPER_VEC_REGISTERS_RETURN
 
 END (STRCHR)
 # endif
-- 
GitLab