This is the mail archive of the libc-alpha@sourceware.org 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]

Re: [PATCH 1.5/2] Fix strrchr regression.


On Mon, Aug 05, 2013 at 07:33:46PM +0200, OndÅej BÃlka wrote:
> Hi, 
>
... 
> To get reliable results I added strrchr to my profiler. You can clearly
> see asymptotic behavior of strrchr here.
> http://kam.mff.cuni.cz/~ondra/benchmark_string/strrchr_profile.html
> 
> Actually there are two regressions here. First one is that strchr_sse42
> is also about 14% asymptotically slower in sandy/ivy bridge and around
> 3% slower for fx10. 
> 
I added optimized strrchr version that decects end by trick from strlen and 
tested by my benchmark. It gives another 10% for all architectures
except old athlon over __strrchr_no_bsf.

I do not know yet how write optimal header. It looks that strchr_sse2
simpler header is faster on result_gcc workload. However I need more
data as strrchr there is used sparingly. 

Results at
http://kam.mff.cuni.cz/~ondra/benchmark_string/strrchr_profile.html
are updated and benchmark program is here:
http://kam.mff.cuni.cz/~ondra/strrchr_profile060813.tar.bz2


My new variant is following (ask if you do want a template):

	.file	"strrchr.c"
	.text
	.p2align 4,,15
	.globl	strrchr_new
	.type	strrchr_new, @function
strrchr_new:
.LFB521:
	.cfi_startproc
	movd	%esi, %xmm1
	movq	%rdi, %rax
	andl	$4095, %eax
	punpcklbw	%xmm1, %xmm1
	cmpq	$4032, %rax
	punpcklwd	%xmm1, %xmm1
	pshufd	$0, %xmm1, %xmm1
	ja	.L2
	movdqu	(%rdi), %xmm0
	pxor	%xmm2, %xmm2
	movdqa	%xmm0, %xmm3
	pcmpeqb	%xmm1, %xmm0
	pcmpeqb	%xmm2, %xmm3
	pmovmskb	%xmm0, %ecx
	pmovmskb	%xmm3, %edx
	testq	%rdx, %rdx
	je	.L3
	leaq	-1(%rdx), %rax
	xorq	%rdx, %rax
	andq	%rcx, %rax
	je	.L4
#APP
# 47 "strrchr.c" 1
	bsrq	%rax, %rax
# 0 "" 2
#NO_APP
	addq	%rdi, %rax
	ret
	.p2align 4,,10
	.p2align 3
.L2:
	movq	%rdi, %rcx
	pxor	%xmm0, %xmm0
	andq	$-64, %rcx
	movdqu	(%rcx), %xmm5
	movdqa	%xmm5, %xmm6
	movdqu	16(%rcx), %xmm4
	pcmpeqb	%xmm1, %xmm5
	pcmpeqb	%xmm0, %xmm6
	movdqu	32(%rcx), %xmm3
	pmovmskb	%xmm6, %esi
	movdqa	%xmm4, %xmm6
	movdqu	48(%rcx), %xmm2
	pcmpeqb	%xmm1, %xmm4
	pcmpeqb	%xmm0, %xmm6
	pmovmskb	%xmm6, %eax
	movdqa	%xmm3, %xmm6
	pcmpeqb	%xmm1, %xmm3
	pcmpeqb	%xmm0, %xmm6
	cltq
	pcmpeqb	%xmm2, %xmm0
	salq	$16, %rax
	pmovmskb	%xmm3, %r9d
	pmovmskb	%xmm6, %r8d
	pmovmskb	%xmm0, %edx
	salq	$32, %r9
	salq	$32, %r8
	pcmpeqb	%xmm1, %xmm2
	orq	%r8, %rax
	salq	$48, %rdx
	pmovmskb	%xmm5, %r8d
	orq	%rsi, %rax
	pmovmskb	%xmm2, %esi
	orq	%rdx, %rax
	pmovmskb	%xmm4, %edx
	salq	$48, %rsi
	salq	$16, %rdx
	orq	%r9, %rdx
	orq	%r8, %rdx
	orq	%rsi, %rdx
	movl	%edi, %esi
	subl	%ecx, %esi
	movl	%esi, %ecx
	shrq	%cl, %rax
	shrq	%cl, %rdx
	testq	%rax, %rax
	je	.L6
	leaq	-1(%rax), %rcx
	xorq	%rax, %rcx
	andq	%rcx, %rdx
	je	.L4
.L33:
#APP
# 47 "strrchr.c" 1
	bsrq	%rdx, %rdx
# 0 "" 2
#NO_APP
	leaq	(%rdi,%rdx), %rax
	ret
	.p2align 4,,10
	.p2align 3
.L6:
	xorl	%eax, %eax
	testq	%rdx, %rdx
	jne	.L35
.L8:
	addq	$64, %rdi
	pxor	%xmm7, %xmm7
	andq	$-64, %rdi
	jmp	.L13
	.p2align 4,,10
	.p2align 3
.L11:
	testq	%rdx, %rdx
	je	.L12
#APP
# 47 "strrchr.c" 1
	bsrq	%rdx, %rdx
# 0 "" 2
#NO_APP
	leaq	(%rdi,%rdx), %rax
.L12:
	addq	$64, %rdi
.L13:
	movdqa	32(%rdi), %xmm3
	pxor	%xmm6, %xmm6
	movdqa	48(%rdi), %xmm2
	movdqa	%xmm3, %xmm0
	movdqa	16(%rdi), %xmm4
	pminub	%xmm2, %xmm0
	movdqa	(%rdi), %xmm5
	pminub	%xmm4, %xmm0
	pminub	%xmm5, %xmm0
	pcmpeqb	%xmm7, %xmm0
	pmovmskb	%xmm0, %ecx
	movdqa	%xmm5, %xmm0
	pcmpeqb	%xmm1, %xmm0
	pmovmskb	%xmm0, %r8d
	movdqa	%xmm4, %xmm0
	pcmpeqb	%xmm1, %xmm0
	pmovmskb	%xmm0, %edx
	movdqa	%xmm3, %xmm0
	pcmpeqb	%xmm1, %xmm0
	salq	$16, %rdx
	pmovmskb	%xmm0, %r9d
	movdqa	%xmm2, %xmm0
	pcmpeqb	%xmm1, %xmm0
	salq	$32, %r9
	orq	%r9, %rdx
	pmovmskb	%xmm0, %esi
	orq	%r8, %rdx
	salq	$48, %rsi
	orq	%rsi, %rdx
	testl	%ecx, %ecx
	je	.L11
	pcmpeqb	%xmm6, %xmm4
	pcmpeqb	%xmm6, %xmm3
	pcmpeqb	%xmm6, %xmm5
	pmovmskb	%xmm4, %ecx
	pmovmskb	%xmm3, %r9d
	pcmpeqb	%xmm6, %xmm2
	pmovmskb	%xmm5, %r8d
	salq	$16, %rcx
	salq	$32, %r9
	pmovmskb	%xmm2, %esi
	orq	%r9, %rcx
	orq	%r8, %rcx
	salq	$48, %rsi
	orq	%rsi, %rcx
	leaq	-1(%rcx), %rsi
	xorq	%rcx, %rsi
	andq	%rsi, %rdx
	jne	.L33
	rep
	ret
	.p2align 4,,10
	.p2align 3
.L3:
	movdqu	16(%rdi), %xmm4
	movdqa	%xmm4, %xmm5
	movdqu	32(%rdi), %xmm3
	pcmpeqb	%xmm1, %xmm4
	pcmpeqb	%xmm2, %xmm5
	movdqu	48(%rdi), %xmm0
	pmovmskb	%xmm5, %edx
	movdqa	%xmm3, %xmm5
	pcmpeqb	%xmm1, %xmm3
	pcmpeqb	%xmm2, %xmm5
	salq	$16, %rdx
	pmovmskb	%xmm3, %r8d
	pmovmskb	%xmm5, %eax
	pcmpeqb	%xmm0, %xmm2
	salq	$32, %r8
	salq	$32, %rax
	pcmpeqb	%xmm1, %xmm0
	orq	%rdx, %rax
	pmovmskb	%xmm4, %edx
	pmovmskb	%xmm2, %esi
	salq	$48, %rsi
	salq	$16, %rdx
	orq	%r8, %rdx
	orq	%rcx, %rdx
	pmovmskb	%xmm0, %ecx
	salq	$48, %rcx
	orq	%rcx, %rdx
	orq	%rsi, %rax
	je	.L6
	leaq	-1(%rax), %rcx
	xorq	%rax, %rcx
	andq	%rcx, %rdx
	jne	.L36
	.p2align 4,,10
	.p2align 3
.L4:
	xorl	%eax, %eax
	ret
	.p2align 4,,10
	.p2align 3
.L35:
#APP
# 47 "strrchr.c" 1
	bsrq	%rdx, %rax
# 0 "" 2
#NO_APP
	addq	%rdi, %rax
	jmp	.L8
	.p2align 4,,10
	.p2align 3
.L36:
#APP
# 47 "strrchr.c" 1
	bsrq	%rdx, %rax
# 0 "" 2
#NO_APP
	addq	%rdi, %rax
	ret
	.cfi_endproc
.LFE521:
	.size	strrchr_new, .-strrchr_new
	.ident	"GCC: (Debian 4.7.1-2) 4.7.1"
	.section	.note.GNU-stack,"",@progbits


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