From 5905e7b3e29139dbef84c065ca39315485f497e1 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Ond=C5=99ej=20B=C3=ADlka?= Date: Wed, 11 Sep 2013 17:07:38 +0200 Subject: Faster strchr implementation. --- sysdeps/x86_64/strchr.S | 191 +++++++++++++++++++++++++++++++++++++++--------- 1 file changed, 157 insertions(+), 34 deletions(-) (limited to 'sysdeps/x86_64/strchr.S') diff --git a/sysdeps/x86_64/strchr.S b/sysdeps/x86_64/strchr.S index d89f1eb..1900b37 100644 --- a/sysdeps/x86_64/strchr.S +++ b/sysdeps/x86_64/strchr.S @@ -19,51 +19,174 @@ #include +# ifndef ALIGN +# define ALIGN(n) .p2align n +# endif + .text ENTRY (strchr) movd %esi, %xmm1 - movq %rdi, %rcx - punpcklbw %xmm1, %xmm1 - andq $~15, %rdi - pxor %xmm2, %xmm2 + movl %edi, %eax + andl $4095, %eax punpcklbw %xmm1, %xmm1 - orl $0xffffffff, %esi - movdqa (%rdi), %xmm0 + cmpl $4032, %eax + punpcklwd %xmm1, %xmm1 pshufd $0, %xmm1, %xmm1 - subq %rdi, %rcx - movdqa %xmm0, %xmm3 - leaq 16(%rdi), %rdi + jg L(cross_page) + movdqu (%rdi), %xmm0 + pxor %xmm3, %xmm3 + movdqa %xmm0, %xmm4 pcmpeqb %xmm1, %xmm0 - pcmpeqb %xmm2, %xmm3 - shl %cl, %esi - pmovmskb %xmm0, %edx - pmovmskb %xmm3, %ecx - andl %esi, %edx - andl %esi, %ecx - orl %edx, %ecx - jnz 1f + pcmpeqb %xmm3, %xmm4 + por %xmm4, %xmm0 + pmovmskb %xmm0, %eax + test %eax, %eax + je L(next_48_bytes) + bsf %eax, %eax +#ifdef AS_STRCHRNUL + leaq (%rdi,%rax), %rax +#else + movl $0, %edx + leaq (%rdi,%rax), %rax + cmpb %sil, (%rax) + cmovne %rdx, %rax +#endif + ret -2: movdqa (%rdi), %xmm0 - leaq 16(%rdi), %rdi - movdqa %xmm0, %xmm3 + ALIGN(3) + L(next_48_bytes): + movdqu 16(%rdi), %xmm0 + movdqa %xmm0, %xmm4 pcmpeqb %xmm1, %xmm0 - pcmpeqb %xmm2, %xmm3 - pmovmskb %xmm0, %edx - pmovmskb %xmm3, %ecx - orl %edx, %ecx - jz 2b + pcmpeqb %xmm3, %xmm4 + por %xmm4, %xmm0 + pmovmskb %xmm0, %ecx + movdqu 32(%rdi), %xmm0 + movdqa %xmm0, %xmm4 + pcmpeqb %xmm1, %xmm0 + salq $16, %rcx + pcmpeqb %xmm3, %xmm4 + por %xmm4, %xmm0 + pmovmskb %xmm0, %eax + movdqu 48(%rdi), %xmm0 + pcmpeqb %xmm0, %xmm3 + salq $32, %rax + pcmpeqb %xmm1, %xmm0 + orq %rcx, %rax + por %xmm3, %xmm0 + pmovmskb %xmm0, %ecx + salq $48, %rcx + orq %rcx, %rax + testq %rax, %rax + jne L(return) +L(loop_start): + /* We use this alignment to force loop be aligned to 8 but not + 16 bytes. This gives better sheduling on AMD processors. */ + ALIGN(4) + pxor %xmm6, %xmm6 + andq $-64, %rdi + ALIGN(3) +L(loop64): + addq $64, %rdi + movdqa (%rdi), %xmm5 + movdqa 16(%rdi), %xmm2 + movdqa 32(%rdi), %xmm3 + pxor %xmm1, %xmm5 + movdqa 48(%rdi), %xmm4 + pxor %xmm1, %xmm2 + pxor %xmm1, %xmm3 + pminub (%rdi), %xmm5 + pxor %xmm1, %xmm4 + pminub 16(%rdi), %xmm2 + pminub 32(%rdi), %xmm3 + pminub %xmm2, %xmm5 + pminub 48(%rdi), %xmm4 + pminub %xmm3, %xmm5 + pminub %xmm4, %xmm5 + pcmpeqb %xmm6, %xmm5 + pmovmskb %xmm5, %eax + + testl %eax, %eax + je L(loop64) + + movdqa (%rdi), %xmm5 + movdqa %xmm5, %xmm0 + pcmpeqb %xmm1, %xmm5 + pcmpeqb %xmm6, %xmm0 + por %xmm0, %xmm5 + pcmpeqb %xmm6, %xmm2 + pcmpeqb %xmm6, %xmm3 + pcmpeqb %xmm6, %xmm4 + + pmovmskb %xmm5, %ecx + pmovmskb %xmm2, %eax + salq $16, %rax + pmovmskb %xmm3, %r8d + pmovmskb %xmm4, %edx + salq $32, %r8 + orq %r8, %rax + orq %rcx, %rax + salq $48, %rdx + orq %rdx, %rax + ALIGN(3) +L(return): + bsfq %rax, %rax +#ifdef AS_STRCHRNUL + leaq (%rdi,%rax), %rax +#else + movl $0, %edx + leaq (%rdi,%rax), %rax + cmpb %sil, (%rax) + cmovne %rdx, %rax +#endif + ret + ALIGN(4) + +L(cross_page): + movq %rdi, %rdx + pxor %xmm2, %xmm2 + andq $-64, %rdx + movdqa %xmm1, %xmm0 + movdqa (%rdx), %xmm3 + movdqa %xmm3, %xmm4 + pcmpeqb %xmm1, %xmm3 + pcmpeqb %xmm2, %xmm4 + por %xmm4, %xmm3 + pmovmskb %xmm3, %r8d + movdqa 16(%rdx), %xmm3 + movdqa %xmm3, %xmm4 + pcmpeqb %xmm1, %xmm3 + pcmpeqb %xmm2, %xmm4 + por %xmm4, %xmm3 + pmovmskb %xmm3, %eax + movdqa 32(%rdx), %xmm3 + movdqa %xmm3, %xmm4 + pcmpeqb %xmm1, %xmm3 + salq $16, %rax + pcmpeqb %xmm2, %xmm4 + por %xmm4, %xmm3 + pmovmskb %xmm3, %r9d + movdqa 48(%rdx), %xmm3 + pcmpeqb %xmm3, %xmm2 + salq $32, %r9 + pcmpeqb %xmm3, %xmm0 + orq %r9, %rax + orq %r8, %rax + por %xmm2, %xmm0 + pmovmskb %xmm0, %ecx + salq $48, %rcx + orq %rcx, %rax + movl %edi, %ecx + subb %dl, %cl + shrq %cl, %rax + testq %rax, %rax + jne L(return) + jmp L(loop_start) -1: bsfl %edx, %edx - jz 4f - bsfl %ecx, %ecx - leaq -16(%rdi,%rdx), %rax - cmpl %edx, %ecx - je 5f -4: xorl %eax, %eax -5: ret END (strchr) +#ifndef AS_STRCHRNUL weak_alias (strchr, index) libc_hidden_builtin_def (strchr) - +#endif -- cgit v1.1