diff options
| -rw-r--r-- | sysdeps/x86_64/multiarch/memcmp-avx2-movbe.S | 98 |
1 files changed, 61 insertions, 37 deletions
diff --git a/sysdeps/x86_64/multiarch/memcmp-avx2-movbe.S b/sysdeps/x86_64/multiarch/memcmp-avx2-movbe.S index a34ea1645d..210c9925b6 100644 --- a/sysdeps/x86_64/multiarch/memcmp-avx2-movbe.S +++ b/sysdeps/x86_64/multiarch/memcmp-avx2-movbe.S @@ -429,22 +429,21 @@ L(page_cross_less_vec): # ifndef USE_AS_WMEMCMP cmpl $8, %edx jae L(between_8_15) + /* Fall through for [4, 7]. */ cmpl $4, %edx - jae L(between_4_7) + jb L(between_2_3) - /* Load as big endian to avoid branches. */ - movzwl (%rdi), %eax - movzwl (%rsi), %ecx - shll $8, %eax - shll $8, %ecx - bswap %eax - bswap %ecx - movzbl -1(%rdi, %rdx), %edi - movzbl -1(%rsi, %rdx), %esi - orl %edi, %eax - orl %esi, %ecx - /* Subtraction is okay because the upper 8 bits are zero. */ - subl %ecx, %eax + movbe (%rdi), %eax + movbe (%rsi), %ecx + shlq $32, %rax + shlq $32, %rcx + movbe -4(%rdi, %rdx), %edi + movbe -4(%rsi, %rdx), %esi + orq %rdi, %rax + orq %rsi, %rcx + subq %rcx, %rax + /* Fast path for return zero. */ + jnz L(ret_nonzero) /* No ymm register was touched. */ ret @@ -457,9 +456,33 @@ L(one_or_less): /* No ymm register was touched. */ ret + .p2align 4,, 5 +L(ret_nonzero): + sbbl %eax, %eax + orl $1, %eax + /* No ymm register was touched. */ + ret + + .p2align 4,, 2 +L(zero): + xorl %eax, %eax + /* No ymm register was touched. */ + ret + .p2align 4 L(between_8_15): -# endif + movbe (%rdi), %rax + movbe (%rsi), %rcx + subq %rcx, %rax + jnz L(ret_nonzero) + movbe -8(%rdi, %rdx), %rax + movbe -8(%rsi, %rdx), %rcx + subq %rcx, %rax + /* Fast path for return zero. */ + jnz L(ret_nonzero) + /* No ymm register was touched. */ + ret +# else /* If USE_AS_WMEMCMP fall through into 8-15 byte case. */ vmovq (%rdi), %xmm1 vmovq (%rsi), %xmm2 @@ -475,16 +498,13 @@ L(between_8_15): VPCMPEQ %xmm1, %xmm2, %xmm2 vpmovmskb %xmm2, %eax subl $0xffff, %eax + /* Fast path for return zero. */ jnz L(return_vec_0) /* No ymm register was touched. */ ret +# endif - .p2align 4 -L(zero): - xorl %eax, %eax - ret - - .p2align 4 + .p2align 4,, 10 L(between_16_31): /* From 16 to 31 bytes. No branch when size == 16. */ vmovdqu (%rsi), %xmm2 @@ -501,11 +521,17 @@ L(between_16_31): VPCMPEQ (%rdi), %xmm2, %xmm2 vpmovmskb %xmm2, %eax subl $0xffff, %eax + /* Fast path for return zero. */ jnz L(return_vec_0) /* No ymm register was touched. */ ret # ifdef USE_AS_WMEMCMP + .p2align 4,, 2 +L(zero): + xorl %eax, %eax + ret + .p2align 4 L(one_or_less): jb L(zero) @@ -520,22 +546,20 @@ L(one_or_less): # else .p2align 4 -L(between_4_7): - /* Load as big endian with overlapping movbe to avoid branches. - */ - movbe (%rdi), %eax - movbe (%rsi), %ecx - shlq $32, %rax - shlq $32, %rcx - movbe -4(%rdi, %rdx), %edi - movbe -4(%rsi, %rdx), %esi - orq %rdi, %rax - orq %rsi, %rcx - subq %rcx, %rax - jz L(zero_4_7) - sbbl %eax, %eax - orl $1, %eax -L(zero_4_7): +L(between_2_3): + /* Load as big endian to avoid branches. */ + movzwl (%rdi), %eax + movzwl (%rsi), %ecx + bswap %eax + bswap %ecx + shrl %eax + shrl %ecx + movzbl -1(%rdi, %rdx), %edi + movzbl -1(%rsi, %rdx), %esi + orl %edi, %eax + orl %esi, %ecx + /* Subtraction is okay because the upper bit is zero. */ + subl %ecx, %eax /* No ymm register was touched. */ ret # endif |