94eca17
commit df5de87260dba479873b2850bbe5c0b81c2376f6
94eca17
Author: Noah Goldstein <goldstein.w.n@gmail.com>
94eca17
Date:   Fri Apr 15 12:28:01 2022 -0500
94eca17
94eca17
    x86: Cleanup page cross code in memcmp-avx2-movbe.S
94eca17
    
94eca17
    Old code was both inefficient and wasted code size. New code (-62
94eca17
    bytes) and comparable or better performance in the page cross case.
94eca17
    
94eca17
    geometric_mean(N=20) of page cross cases New / Original: 0.960
94eca17
    
94eca17
    size, align0, align1, ret, New Time/Old Time
94eca17
       1,   4095,      0,   0,             1.001
94eca17
       1,   4095,      0,   1,             0.999
94eca17
       1,   4095,      0,  -1,               1.0
94eca17
       2,   4094,      0,   0,               1.0
94eca17
       2,   4094,      0,   1,               1.0
94eca17
       2,   4094,      0,  -1,               1.0
94eca17
       3,   4093,      0,   0,               1.0
94eca17
       3,   4093,      0,   1,               1.0
94eca17
       3,   4093,      0,  -1,               1.0
94eca17
       4,   4092,      0,   0,             0.987
94eca17
       4,   4092,      0,   1,               1.0
94eca17
       4,   4092,      0,  -1,               1.0
94eca17
       5,   4091,      0,   0,             0.984
94eca17
       5,   4091,      0,   1,             1.002
94eca17
       5,   4091,      0,  -1,             1.005
94eca17
       6,   4090,      0,   0,             0.993
94eca17
       6,   4090,      0,   1,             1.001
94eca17
       6,   4090,      0,  -1,             1.003
94eca17
       7,   4089,      0,   0,             0.991
94eca17
       7,   4089,      0,   1,               1.0
94eca17
       7,   4089,      0,  -1,             1.001
94eca17
       8,   4088,      0,   0,             0.875
94eca17
       8,   4088,      0,   1,             0.881
94eca17
       8,   4088,      0,  -1,             0.888
94eca17
       9,   4087,      0,   0,             0.872
94eca17
       9,   4087,      0,   1,             0.879
94eca17
       9,   4087,      0,  -1,             0.883
94eca17
      10,   4086,      0,   0,             0.878
94eca17
      10,   4086,      0,   1,             0.886
94eca17
      10,   4086,      0,  -1,             0.873
94eca17
      11,   4085,      0,   0,             0.878
94eca17
      11,   4085,      0,   1,             0.881
94eca17
      11,   4085,      0,  -1,             0.879
94eca17
      12,   4084,      0,   0,             0.873
94eca17
      12,   4084,      0,   1,             0.889
94eca17
      12,   4084,      0,  -1,             0.875
94eca17
      13,   4083,      0,   0,             0.873
94eca17
      13,   4083,      0,   1,             0.863
94eca17
      13,   4083,      0,  -1,             0.863
94eca17
      14,   4082,      0,   0,             0.838
94eca17
      14,   4082,      0,   1,             0.869
94eca17
      14,   4082,      0,  -1,             0.877
94eca17
      15,   4081,      0,   0,             0.841
94eca17
      15,   4081,      0,   1,             0.869
94eca17
      15,   4081,      0,  -1,             0.876
94eca17
      16,   4080,      0,   0,             0.988
94eca17
      16,   4080,      0,   1,              0.99
94eca17
      16,   4080,      0,  -1,             0.989
94eca17
      17,   4079,      0,   0,             0.978
94eca17
      17,   4079,      0,   1,             0.981
94eca17
      17,   4079,      0,  -1,              0.98
94eca17
      18,   4078,      0,   0,             0.981
94eca17
      18,   4078,      0,   1,              0.98
94eca17
      18,   4078,      0,  -1,             0.985
94eca17
      19,   4077,      0,   0,             0.977
94eca17
      19,   4077,      0,   1,             0.979
94eca17
      19,   4077,      0,  -1,             0.986
94eca17
      20,   4076,      0,   0,             0.977
94eca17
      20,   4076,      0,   1,             0.986
94eca17
      20,   4076,      0,  -1,             0.984
94eca17
      21,   4075,      0,   0,             0.977
94eca17
      21,   4075,      0,   1,             0.983
94eca17
      21,   4075,      0,  -1,             0.988
94eca17
      22,   4074,      0,   0,             0.983
94eca17
      22,   4074,      0,   1,             0.994
94eca17
      22,   4074,      0,  -1,             0.993
94eca17
      23,   4073,      0,   0,              0.98
94eca17
      23,   4073,      0,   1,             0.992
94eca17
      23,   4073,      0,  -1,             0.995
94eca17
      24,   4072,      0,   0,             0.989
94eca17
      24,   4072,      0,   1,             0.989
94eca17
      24,   4072,      0,  -1,             0.991
94eca17
      25,   4071,      0,   0,              0.99
94eca17
      25,   4071,      0,   1,             0.999
94eca17
      25,   4071,      0,  -1,             0.996
94eca17
      26,   4070,      0,   0,             0.993
94eca17
      26,   4070,      0,   1,             0.995
94eca17
      26,   4070,      0,  -1,             0.998
94eca17
      27,   4069,      0,   0,             0.993
94eca17
      27,   4069,      0,   1,             0.999
94eca17
      27,   4069,      0,  -1,               1.0
94eca17
      28,   4068,      0,   0,             0.997
94eca17
      28,   4068,      0,   1,               1.0
94eca17
      28,   4068,      0,  -1,             0.999
94eca17
      29,   4067,      0,   0,             0.996
94eca17
      29,   4067,      0,   1,             0.999
94eca17
      29,   4067,      0,  -1,             0.999
94eca17
      30,   4066,      0,   0,             0.991
94eca17
      30,   4066,      0,   1,             1.001
94eca17
      30,   4066,      0,  -1,             0.999
94eca17
      31,   4065,      0,   0,             0.988
94eca17
      31,   4065,      0,   1,             0.998
94eca17
      31,   4065,      0,  -1,             0.998
94eca17
    Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
94eca17
    
94eca17
    (cherry picked from commit 23102686ec67b856a2d4fd25ddaa1c0b8d175c4f)
94eca17
94eca17
diff --git a/sysdeps/x86_64/multiarch/memcmp-avx2-movbe.S b/sysdeps/x86_64/multiarch/memcmp-avx2-movbe.S
94eca17
index 2621ec907aedb781..ec9cf0852edf216d 100644
94eca17
--- a/sysdeps/x86_64/multiarch/memcmp-avx2-movbe.S
94eca17
+++ b/sysdeps/x86_64/multiarch/memcmp-avx2-movbe.S
94eca17
@@ -429,22 +429,21 @@ L(page_cross_less_vec):
94eca17
 # ifndef USE_AS_WMEMCMP
94eca17
 	cmpl	$8, %edx
94eca17
 	jae	L(between_8_15)
94eca17
+	/* Fall through for [4, 7].  */
94eca17
 	cmpl	$4, %edx
94eca17
-	jae	L(between_4_7)
94eca17
+	jb	L(between_2_3)
94eca17
 
94eca17
-	/* Load as big endian to avoid branches.  */
94eca17
-	movzwl	(%rdi), %eax
94eca17
-	movzwl	(%rsi), %ecx
94eca17
-	shll	$8, %eax
94eca17
-	shll	$8, %ecx
94eca17
-	bswap	%eax
94eca17
-	bswap	%ecx
94eca17
-	movzbl	-1(%rdi, %rdx), %edi
94eca17
-	movzbl	-1(%rsi, %rdx), %esi
94eca17
-	orl	%edi, %eax
94eca17
-	orl	%esi, %ecx
94eca17
-	/* Subtraction is okay because the upper 8 bits are zero.  */
94eca17
-	subl	%ecx, %eax
94eca17
+	movbe	(%rdi), %eax
94eca17
+	movbe	(%rsi), %ecx
94eca17
+	shlq	$32, %rax
94eca17
+	shlq	$32, %rcx
94eca17
+	movbe	-4(%rdi, %rdx), %edi
94eca17
+	movbe	-4(%rsi, %rdx), %esi
94eca17
+	orq	%rdi, %rax
94eca17
+	orq	%rsi, %rcx
94eca17
+	subq	%rcx, %rax
94eca17
+	/* Fast path for return zero.  */
94eca17
+	jnz	L(ret_nonzero)
94eca17
 	/* No ymm register was touched.  */
94eca17
 	ret
94eca17
 
94eca17
@@ -457,9 +456,33 @@ L(one_or_less):
94eca17
 	/* No ymm register was touched.  */
94eca17
 	ret
94eca17
 
94eca17
+	.p2align 4,, 5
94eca17
+L(ret_nonzero):
94eca17
+	sbbl	%eax, %eax
94eca17
+	orl	$1, %eax
94eca17
+	/* No ymm register was touched.  */
94eca17
+	ret
94eca17
+
94eca17
+	.p2align 4,, 2
94eca17
+L(zero):
94eca17
+	xorl	%eax, %eax
94eca17
+	/* No ymm register was touched.  */
94eca17
+	ret
94eca17
+
94eca17
 	.p2align 4
94eca17
 L(between_8_15):
94eca17
-# endif
94eca17
+	movbe	(%rdi), %rax
94eca17
+	movbe	(%rsi), %rcx
94eca17
+	subq	%rcx, %rax
94eca17
+	jnz	L(ret_nonzero)
94eca17
+	movbe	-8(%rdi, %rdx), %rax
94eca17
+	movbe	-8(%rsi, %rdx), %rcx
94eca17
+	subq	%rcx, %rax
94eca17
+	/* Fast path for return zero.  */
94eca17
+	jnz	L(ret_nonzero)
94eca17
+	/* No ymm register was touched.  */
94eca17
+	ret
94eca17
+# else
94eca17
 	/* If USE_AS_WMEMCMP fall through into 8-15 byte case.  */
94eca17
 	vmovq	(%rdi), %xmm1
94eca17
 	vmovq	(%rsi), %xmm2
94eca17
@@ -475,16 +498,13 @@ L(between_8_15):
94eca17
 	VPCMPEQ	%xmm1, %xmm2, %xmm2
94eca17
 	vpmovmskb %xmm2, %eax
94eca17
 	subl	$0xffff, %eax
94eca17
+	/* Fast path for return zero.  */
94eca17
 	jnz	L(return_vec_0)
94eca17
 	/* No ymm register was touched.  */
94eca17
 	ret
94eca17
+# endif
94eca17
 
94eca17
-	.p2align 4
94eca17
-L(zero):
94eca17
-	xorl	%eax, %eax
94eca17
-	ret
94eca17
-
94eca17
-	.p2align 4
94eca17
+	.p2align 4,, 10
94eca17
 L(between_16_31):
94eca17
 	/* From 16 to 31 bytes.  No branch when size == 16.  */
94eca17
 	vmovdqu	(%rsi), %xmm2
94eca17
@@ -501,11 +521,17 @@ L(between_16_31):
94eca17
 	VPCMPEQ	(%rdi), %xmm2, %xmm2
94eca17
 	vpmovmskb %xmm2, %eax
94eca17
 	subl	$0xffff, %eax
94eca17
+	/* Fast path for return zero.  */
94eca17
 	jnz	L(return_vec_0)
94eca17
 	/* No ymm register was touched.  */
94eca17
 	ret
94eca17
 
94eca17
 # ifdef USE_AS_WMEMCMP
94eca17
+	.p2align 4,, 2
94eca17
+L(zero):
94eca17
+	xorl	%eax, %eax
94eca17
+	ret
94eca17
+
94eca17
 	.p2align 4
94eca17
 L(one_or_less):
94eca17
 	jb	L(zero)
94eca17
@@ -520,22 +546,20 @@ L(one_or_less):
94eca17
 # else
94eca17
 
94eca17
 	.p2align 4
94eca17
-L(between_4_7):
94eca17
-	/* Load as big endian with overlapping movbe to avoid branches.
94eca17
-	 */
94eca17
-	movbe	(%rdi), %eax
94eca17
-	movbe	(%rsi), %ecx
94eca17
-	shlq	$32, %rax
94eca17
-	shlq	$32, %rcx
94eca17
-	movbe	-4(%rdi, %rdx), %edi
94eca17
-	movbe	-4(%rsi, %rdx), %esi
94eca17
-	orq	%rdi, %rax
94eca17
-	orq	%rsi, %rcx
94eca17
-	subq	%rcx, %rax
94eca17
-	jz	L(zero_4_7)
94eca17
-	sbbl	%eax, %eax
94eca17
-	orl	$1, %eax
94eca17
-L(zero_4_7):
94eca17
+L(between_2_3):
94eca17
+	/* Load as big endian to avoid branches.  */
94eca17
+	movzwl	(%rdi), %eax
94eca17
+	movzwl	(%rsi), %ecx
94eca17
+	bswap	%eax
94eca17
+	bswap	%ecx
94eca17
+	shrl	%eax
94eca17
+	shrl	%ecx
94eca17
+	movzbl	-1(%rdi, %rdx), %edi
94eca17
+	movzbl	-1(%rsi, %rdx), %esi
94eca17
+	orl	%edi, %eax
94eca17
+	orl	%esi, %ecx
94eca17
+	/* Subtraction is okay because the upper bit is zero.  */
94eca17
+	subl	%ecx, %eax
94eca17
 	/* No ymm register was touched.  */
94eca17
 	ret
94eca17
 # endif