This is the mail archive of the mailing list for the glibc project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

[PATCH] Do not multiply by zero in bzero.


I noticed that bzero jumps directly at memset. Memset then computes mask
by 64bit multiplication. This is suboptimal in bzero case as mask iz

I emulating first few instructions of memset and simplified it. 

Related question is that we first check if n==1. In my profile this
happens only with 2% of cases. Is there special reason to keep it?

Passes test on x86-64. OK to commit?

	* sysdeps/x86_64/memset.S (memset): Move entry from bzero.
	(bzero): Skip multiplication by zero in memset.

 sysdeps/x86_64/memset.S |   14 ++++++++++----
 1 files changed, 10 insertions(+), 4 deletions(-)

diff --git a/sysdeps/x86_64/memset.S b/sysdeps/x86_64/memset.S
index b393efe..375c284 100644
--- a/sysdeps/x86_64/memset.S
+++ b/sysdeps/x86_64/memset.S
@@ -25,9 +25,15 @@
 #if !defined NOT_IN_libc
-	mov	%rsi,%rdx	/* Adjust parameter.  */
-	xorl	%esi,%esi	/* Fill with 0s.  */
-	jmp	L(memset_entry)
+	mov    %rsi,%r8		/* Adjust parameter.  */
+	xor    %rdx,%rdx	/* Fill with 0s.  */
+	cmp    $0x1,%r8
+	mov    %rdi,%rax  /* memset returns the dest address.  */
+	jne    L(entry_from_bzero)
+	mov    %dl,(%rdi)
+	retq
 weak_alias (__bzero, bzero)
@@ -39,7 +45,6 @@ ENTRY_CHK (__memset_chk)
 END_CHK (__memset_chk)
 ENTRY (memset)
 	cmp    $0x1,%rdx
 	mov    %rdi,%rax	/* memset returns the dest address.  */
 	jne    L(ck2)
@@ -50,6 +55,7 @@ L(ck2):
 	mov    %rdx,%r8
 	movzbq %sil,%rdx
 	imul   %r9,%rdx
 	cmp    $0x90,%r8
 	ja     L(ck_mem_ops_method)

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]