513694
From 017773f93b0e41f3b164e5db86d0c7b7f75675e9 Mon Sep 17 00:00:00 2001
513694
From: Noah Goldstein <goldstein.w.n@gmail.com>
513694
Date: Wed, 23 Mar 2022 16:57:36 -0500
513694
Subject: [PATCH] x86: Optimize str{n}casecmp TOLOWER logic in strcmp.S
513694
513694
Slightly faster method of doing TOLOWER that saves an
513694
instruction.
513694
513694
Also replace the hard coded 5-byte no with .p2align 4. On builds with
513694
CET enabled this misaligned entry to strcasecmp.
513694
513694
geometric_mean(N=40) of all benchmarks New / Original: .894
513694
513694
All string/memory tests pass.
513694
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
513694
513694
(cherry picked from commit 670b54bc585ea4a94f3b2e9272ba44aa6b730b73)
513694
---
513694
 sysdeps/x86_64/strcmp.S | 64 +++++++++++++++++++----------------------
513694
 1 file changed, 29 insertions(+), 35 deletions(-)
513694
513694
diff --git a/sysdeps/x86_64/strcmp.S b/sysdeps/x86_64/strcmp.S
513694
index aa6df898..f454ce5b 100644
513694
--- a/sysdeps/x86_64/strcmp.S
513694
+++ b/sysdeps/x86_64/strcmp.S
513694
@@ -78,9 +78,8 @@ ENTRY2 (__strcasecmp)
513694
 	movq	__libc_tsd_LOCALE@gottpoff(%rip),%rax
513694
 	mov	%fs:(%rax),%RDX_LP
513694
 
513694
-	// XXX 5 byte should be before the function
513694
-	/* 5-byte NOP.  */
513694
-	.byte	0x0f,0x1f,0x44,0x00,0x00
513694
+	/* Either 1 or 5 bytes (dependeing if CET is enabled).  */
513694
+	.p2align 4
513694
 END2 (__strcasecmp)
513694
 # ifndef NO_NOLOCALE_ALIAS
513694
 weak_alias (__strcasecmp, strcasecmp)
513694
@@ -97,9 +96,8 @@ ENTRY2 (__strncasecmp)
513694
 	movq	__libc_tsd_LOCALE@gottpoff(%rip),%rax
513694
 	mov	%fs:(%rax),%RCX_LP
513694
 
513694
-	// XXX 5 byte should be before the function
513694
-	/* 5-byte NOP.  */
513694
-	.byte	0x0f,0x1f,0x44,0x00,0x00
513694
+	/* Either 1 or 5 bytes (dependeing if CET is enabled).  */
513694
+	.p2align 4
513694
 END2 (__strncasecmp)
513694
 # ifndef NO_NOLOCALE_ALIAS
513694
 weak_alias (__strncasecmp, strncasecmp)
513694
@@ -149,22 +147,22 @@ ENTRY (STRCMP)
513694
 #if defined USE_AS_STRCASECMP_L || defined USE_AS_STRNCASECMP_L
513694
 	.section .rodata.cst16,"aM",@progbits,16
513694
 	.align 16
513694
-.Lbelowupper:
513694
-	.quad	0x4040404040404040
513694
-	.quad	0x4040404040404040
513694
-.Ltopupper:
513694
-	.quad	0x5b5b5b5b5b5b5b5b
513694
-	.quad	0x5b5b5b5b5b5b5b5b
513694
-.Ltouppermask:
513694
+.Llcase_min:
513694
+	.quad	0x3f3f3f3f3f3f3f3f
513694
+	.quad	0x3f3f3f3f3f3f3f3f
513694
+.Llcase_max:
513694
+	.quad	0x9999999999999999
513694
+	.quad	0x9999999999999999
513694
+.Lcase_add:
513694
 	.quad	0x2020202020202020
513694
 	.quad	0x2020202020202020
513694
 	.previous
513694
-	movdqa	.Lbelowupper(%rip), %xmm5
513694
-# define UCLOW_reg %xmm5
513694
-	movdqa	.Ltopupper(%rip), %xmm6
513694
-# define UCHIGH_reg %xmm6
513694
-	movdqa	.Ltouppermask(%rip), %xmm7
513694
-# define LCQWORD_reg %xmm7
513694
+	movdqa	.Llcase_min(%rip), %xmm5
513694
+# define LCASE_MIN_reg %xmm5
513694
+	movdqa	.Llcase_max(%rip), %xmm6
513694
+# define LCASE_MAX_reg %xmm6
513694
+	movdqa	.Lcase_add(%rip), %xmm7
513694
+# define CASE_ADD_reg %xmm7
513694
 #endif
513694
 	cmp	$0x30, %ecx
513694
 	ja	LABEL(crosscache)	/* rsi: 16-byte load will cross cache line */
513694
@@ -175,22 +173,18 @@ ENTRY (STRCMP)
513694
 	movhpd	8(%rdi), %xmm1
513694
 	movhpd	8(%rsi), %xmm2
513694
 #if defined USE_AS_STRCASECMP_L || defined USE_AS_STRNCASECMP_L
513694
-# define TOLOWER(reg1, reg2) \
513694
-	movdqa	reg1, %xmm8;					\
513694
-	movdqa	UCHIGH_reg, %xmm9;				\
513694
-	movdqa	reg2, %xmm10;					\
513694
-	movdqa	UCHIGH_reg, %xmm11;				\
513694
-	pcmpgtb	UCLOW_reg, %xmm8;				\
513694
-	pcmpgtb	reg1, %xmm9;					\
513694
-	pcmpgtb	UCLOW_reg, %xmm10;				\
513694
-	pcmpgtb	reg2, %xmm11;					\
513694
-	pand	%xmm9, %xmm8;					\
513694
-	pand	%xmm11, %xmm10;					\
513694
-	pand	LCQWORD_reg, %xmm8;				\
513694
-	pand	LCQWORD_reg, %xmm10;				\
513694
-	por	%xmm8, reg1;					\
513694
-	por	%xmm10, reg2
513694
-	TOLOWER (%xmm1, %xmm2)
513694
+#  define TOLOWER(reg1, reg2) \
513694
+	movdqa	LCASE_MIN_reg, %xmm8;					\
513694
+	movdqa	LCASE_MIN_reg, %xmm9;					\
513694
+	paddb	reg1, %xmm8;					\
513694
+	paddb	reg2, %xmm9;					\
513694
+	pcmpgtb	LCASE_MAX_reg, %xmm8;				\
513694
+	pcmpgtb	LCASE_MAX_reg, %xmm9;				\
513694
+	pandn	CASE_ADD_reg, %xmm8;					\
513694
+	pandn	CASE_ADD_reg, %xmm9;					\
513694
+	paddb	%xmm8, reg1;					\
513694
+	paddb	%xmm9, reg2
513694
+	TOLOWER	(%xmm1, %xmm2)
513694
 #else
513694
 # define TOLOWER(reg1, reg2)
513694
 #endif
513694
-- 
513694
GitLab
513694