diff options
Diffstat (limited to 'sysdeps/x86_64/multiarch/memchr-sse2.S')
| -rw-r--r-- | sysdeps/x86_64/multiarch/memchr-sse2.S | 363 |
1 files changed, 355 insertions, 8 deletions
diff --git a/sysdeps/x86_64/multiarch/memchr-sse2.S b/sysdeps/x86_64/multiarch/memchr-sse2.S index 2c6fdd41d6..8c561cd687 100644 --- a/sysdeps/x86_64/multiarch/memchr-sse2.S +++ b/sysdeps/x86_64/multiarch/memchr-sse2.S @@ -16,13 +16,360 @@ License along with the GNU C Library; if not, see <https://www.gnu.org/licenses/>. */ -#if IS_IN (libc) -# define memchr __memchr_sse2 +#include <isa-level.h> +#include <sysdep.h> -# undef strong_alias -# define strong_alias(memchr, __memchr) -# undef libc_hidden_builtin_def -# define libc_hidden_builtin_def(memchr) -#endif +/* MINIMUM_X86_ISA_LEVEL <= 2 because there is no V2 implementation + so we need this to build for ISA V2 builds. */ +#if ISA_SHOULD_BUILD (2) + +# ifndef MEMCHR +# define MEMCHR __memchr_sse2 +# endif +# ifdef USE_AS_WMEMCHR +# define PCMPEQ pcmpeqd +# define CHAR_PER_VEC 4 +# else +# define PCMPEQ pcmpeqb +# define CHAR_PER_VEC 16 +# endif + +/* fast SSE2 version with using pmaxub and 64 byte loop */ + + .text +ENTRY(MEMCHR) + movd %esi, %xmm1 + mov %edi, %ecx + +# ifdef __ILP32__ + /* Clear the upper 32 bits. */ + movl %edx, %edx +# endif +# ifdef USE_AS_WMEMCHR + test %RDX_LP, %RDX_LP + jz L(return_null) +# else + punpcklbw %xmm1, %xmm1 + test %RDX_LP, %RDX_LP + jz L(return_null) + punpcklbw %xmm1, %xmm1 +# endif + + and $63, %ecx + pshufd $0, %xmm1, %xmm1 + + cmp $48, %ecx + ja L(crosscache) + + movdqu (%rdi), %xmm0 + PCMPEQ %xmm1, %xmm0 + pmovmskb %xmm0, %eax + test %eax, %eax + + jnz L(matches_1) + sub $CHAR_PER_VEC, %rdx + jbe L(return_null) + add $16, %rdi + and $15, %ecx + and $-16, %rdi +# ifdef USE_AS_WMEMCHR + shr $2, %ecx +# endif + add %rcx, %rdx + sub $(CHAR_PER_VEC * 4), %rdx + jbe L(exit_loop) + jmp L(loop_prolog) + + .p2align 4 +L(crosscache): + and $15, %ecx + and $-16, %rdi + movdqa (%rdi), %xmm0 + + PCMPEQ %xmm1, %xmm0 + /* Check if there is a match. */ + pmovmskb %xmm0, %eax + /* Remove the leading bytes. */ + sar %cl, %eax + test %eax, %eax + je L(unaligned_no_match) + /* Check which byte is a match. */ + bsf %eax, %eax +# ifdef USE_AS_WMEMCHR + mov %eax, %esi + shr $2, %esi + sub %rsi, %rdx +# else + sub %rax, %rdx +# endif + jbe L(return_null) + add %rdi, %rax + add %rcx, %rax + ret + + .p2align 4 +L(unaligned_no_match): + /* "rcx" is less than 16. Calculate "rdx + rcx - 16" by using + "rdx - (16 - rcx)" instead of "(rdx + rcx) - 16" to void + possible addition overflow. */ + neg %rcx + add $16, %rcx +# ifdef USE_AS_WMEMCHR + shr $2, %ecx +# endif + sub %rcx, %rdx + jbe L(return_null) + add $16, %rdi + sub $(CHAR_PER_VEC * 4), %rdx + jbe L(exit_loop) + + .p2align 4 +L(loop_prolog): + movdqa (%rdi), %xmm0 + PCMPEQ %xmm1, %xmm0 + pmovmskb %xmm0, %eax + test %eax, %eax + jnz L(matches) + + movdqa 16(%rdi), %xmm2 + PCMPEQ %xmm1, %xmm2 + pmovmskb %xmm2, %eax + test %eax, %eax + jnz L(matches16) + + movdqa 32(%rdi), %xmm3 + PCMPEQ %xmm1, %xmm3 + pmovmskb %xmm3, %eax + test %eax, %eax + jnz L(matches32) + + movdqa 48(%rdi), %xmm4 + PCMPEQ %xmm1, %xmm4 + add $64, %rdi + pmovmskb %xmm4, %eax + test %eax, %eax + jnz L(matches0) + + test $0x3f, %rdi + jz L(align64_loop) + + sub $(CHAR_PER_VEC * 4), %rdx + jbe L(exit_loop) + + movdqa (%rdi), %xmm0 + PCMPEQ %xmm1, %xmm0 + pmovmskb %xmm0, %eax + test %eax, %eax + jnz L(matches) + + movdqa 16(%rdi), %xmm2 + PCMPEQ %xmm1, %xmm2 + pmovmskb %xmm2, %eax + test %eax, %eax + jnz L(matches16) + + movdqa 32(%rdi), %xmm3 + PCMPEQ %xmm1, %xmm3 + pmovmskb %xmm3, %eax + test %eax, %eax + jnz L(matches32) + + movdqa 48(%rdi), %xmm3 + PCMPEQ %xmm1, %xmm3 + pmovmskb %xmm3, %eax + + add $64, %rdi + test %eax, %eax + jnz L(matches0) + + mov %rdi, %rcx + and $-64, %rdi + and $63, %ecx +# ifdef USE_AS_WMEMCHR + shr $2, %ecx +# endif + add %rcx, %rdx + + .p2align 4 +L(align64_loop): + sub $(CHAR_PER_VEC * 4), %rdx + jbe L(exit_loop) + movdqa (%rdi), %xmm0 + movdqa 16(%rdi), %xmm2 + movdqa 32(%rdi), %xmm3 + movdqa 48(%rdi), %xmm4 + + PCMPEQ %xmm1, %xmm0 + PCMPEQ %xmm1, %xmm2 + PCMPEQ %xmm1, %xmm3 + PCMPEQ %xmm1, %xmm4 -#include "../memchr.S" + pmaxub %xmm0, %xmm3 + pmaxub %xmm2, %xmm4 + pmaxub %xmm3, %xmm4 + pmovmskb %xmm4, %eax + + add $64, %rdi + + test %eax, %eax + jz L(align64_loop) + + sub $64, %rdi + + pmovmskb %xmm0, %eax + test %eax, %eax + jnz L(matches) + + pmovmskb %xmm2, %eax + test %eax, %eax + jnz L(matches16) + + movdqa 32(%rdi), %xmm3 + PCMPEQ %xmm1, %xmm3 + + PCMPEQ 48(%rdi), %xmm1 + pmovmskb %xmm3, %eax + test %eax, %eax + jnz L(matches32) + + pmovmskb %xmm1, %eax + bsf %eax, %eax + lea 48(%rdi, %rax), %rax + ret + + .p2align 4 +L(exit_loop): + add $(CHAR_PER_VEC * 2), %edx + jle L(exit_loop_32) + + movdqa (%rdi), %xmm0 + PCMPEQ %xmm1, %xmm0 + pmovmskb %xmm0, %eax + test %eax, %eax + jnz L(matches) + + movdqa 16(%rdi), %xmm2 + PCMPEQ %xmm1, %xmm2 + pmovmskb %xmm2, %eax + test %eax, %eax + jnz L(matches16) + + movdqa 32(%rdi), %xmm3 + PCMPEQ %xmm1, %xmm3 + pmovmskb %xmm3, %eax + test %eax, %eax + jnz L(matches32_1) + sub $CHAR_PER_VEC, %edx + jle L(return_null) + + PCMPEQ 48(%rdi), %xmm1 + pmovmskb %xmm1, %eax + test %eax, %eax + jnz L(matches48_1) + xor %eax, %eax + ret + + .p2align 4 +L(exit_loop_32): + add $(CHAR_PER_VEC * 2), %edx + movdqa (%rdi), %xmm0 + PCMPEQ %xmm1, %xmm0 + pmovmskb %xmm0, %eax + test %eax, %eax + jnz L(matches_1) + sub $CHAR_PER_VEC, %edx + jbe L(return_null) + + PCMPEQ 16(%rdi), %xmm1 + pmovmskb %xmm1, %eax + test %eax, %eax + jnz L(matches16_1) + xor %eax, %eax + ret + + .p2align 4 +L(matches0): + bsf %eax, %eax + lea -16(%rax, %rdi), %rax + ret + + .p2align 4 +L(matches): + bsf %eax, %eax + add %rdi, %rax + ret + + .p2align 4 +L(matches16): + bsf %eax, %eax + lea 16(%rax, %rdi), %rax + ret + + .p2align 4 +L(matches32): + bsf %eax, %eax + lea 32(%rax, %rdi), %rax + ret + + .p2align 4 +L(matches_1): + bsf %eax, %eax +# ifdef USE_AS_WMEMCHR + mov %eax, %esi + shr $2, %esi + sub %rsi, %rdx +# else + sub %rax, %rdx +# endif + jbe L(return_null) + add %rdi, %rax + ret + + .p2align 4 +L(matches16_1): + bsf %eax, %eax +# ifdef USE_AS_WMEMCHR + mov %eax, %esi + shr $2, %esi + sub %rsi, %rdx +# else + sub %rax, %rdx +# endif + jbe L(return_null) + lea 16(%rdi, %rax), %rax + ret + + .p2align 4 +L(matches32_1): + bsf %eax, %eax +# ifdef USE_AS_WMEMCHR + mov %eax, %esi + shr $2, %esi + sub %rsi, %rdx +# else + sub %rax, %rdx +# endif + jbe L(return_null) + lea 32(%rdi, %rax), %rax + ret + + .p2align 4 +L(matches48_1): + bsf %eax, %eax +# ifdef USE_AS_WMEMCHR + mov %eax, %esi + shr $2, %esi + sub %rsi, %rdx +# else + sub %rax, %rdx +# endif + jbe L(return_null) + lea 48(%rdi, %rax), %rax + ret + + .p2align 4 +L(return_null): + xor %eax, %eax + ret +END(MEMCHR) +#endif |