diff options
Diffstat (limited to 'sysdeps/x86_64')
| -rw-r--r-- | sysdeps/x86_64/wcscmp.S | 109 |
1 files changed, 64 insertions, 45 deletions
diff --git a/sysdeps/x86_64/wcscmp.S b/sysdeps/x86_64/wcscmp.S index 991ecb2cab..12bfdafd41 100644 --- a/sysdeps/x86_64/wcscmp.S +++ b/sysdeps/x86_64/wcscmp.S @@ -20,6 +20,8 @@ #include <sysdep.h> +/* Note: wcscmp uses signed comparison, not unsighed as in strcmp function. */ + .text ENTRY (wcscmp) /* @@ -76,7 +78,7 @@ L(continue_48_48): jne L(nequal) test %ecx, %ecx jz L(equal) - + movdqu 16(%rdi), %xmm1 movdqu 16(%rsi), %xmm2 pcmpeqd %xmm1, %xmm0 /* Any null double_word? */ @@ -209,21 +211,21 @@ L(continue_00_48): test %ecx, %ecx jnz L(less4_double_words1) - sub (%rsi), %eax - jnz L(return) - + cmp (%rsi), %eax + jne L(nequal) + mov 4(%rdi), %eax - sub 4(%rsi), %eax - jnz L(return) + cmp 4(%rsi), %eax + jne L(nequal) mov 8(%rdi), %eax - sub 8(%rsi), %eax - jnz L(return) + cmp 8(%rsi), %eax + jne L(nequal) mov 12(%rdi), %eax - sub 12(%rsi), %eax - jnz L(return) - + cmp 12(%rsi), %eax + jne L(nequal) + movdqu 16(%rsi), %xmm2 pcmpeqd %xmm2, %xmm0 /* Any null double_word? */ pcmpeqd 16(%rdi), %xmm2 /* compare first 4 double_words for equality */ @@ -530,21 +532,21 @@ L(continue_48_00): test %ecx, %ecx jnz L(less4_double_words1) - sub (%rsi), %eax - jnz L(return) - + cmp (%rsi), %eax + jne L(nequal) + mov 4(%rdi), %eax - sub 4(%rsi), %eax - jnz L(return) + cmp 4(%rsi), %eax + jne L(nequal) mov 8(%rdi), %eax - sub 8(%rsi), %eax - jnz L(return) + cmp 8(%rsi), %eax + jne L(nequal) mov 12(%rdi), %eax - sub 12(%rsi), %eax - jnz L(return) - + cmp 12(%rsi), %eax + jne L(nequal) + movdqu 16(%rdi), %xmm1 pcmpeqd %xmm1, %xmm0 /* Any null double_word? */ pcmpeqd 16(%rsi), %xmm1 /* compare first 4 double_words for equality */ @@ -784,25 +786,29 @@ L(less4_double_words1): test %ecx, %ecx jz L(equal) - mov 12(%rsi), %edx - mov 12(%rdi), %eax - sub %edx, %eax + mov 12(%rsi), %ecx + cmp %ecx, 12(%rdi) + jne L(nequal) + xor %eax, %eax ret .p2align 4 L(less4_double_words): + xor %eax, %eax test %dl, %dl jz L(next_two_double_words) and $15, %dl jz L(second_double_word) mov (%rdi), %eax - sub (%rsi), %eax + cmp (%rsi), %eax + jne L(nequal) ret .p2align 4 L(second_double_word): mov 4(%rdi), %eax - sub 4(%rsi), %eax + cmp 4(%rsi), %eax + jne L(nequal) ret .p2align 4 @@ -810,29 +816,34 @@ L(next_two_double_words): and $15, %dh jz L(fourth_double_word) mov 8(%rdi), %eax - sub 8(%rsi), %eax + cmp 8(%rsi), %eax + jne L(nequal) ret .p2align 4 L(fourth_double_word): mov 12(%rdi), %eax - sub 12(%rsi), %eax + cmp 12(%rsi), %eax + jne L(nequal) ret .p2align 4 L(less4_double_words_16): + xor %eax, %eax test %dl, %dl jz L(next_two_double_words_16) and $15, %dl jz L(second_double_word_16) mov 16(%rdi), %eax - sub 16(%rsi), %eax + cmp 16(%rsi), %eax + jne L(nequal) ret .p2align 4 L(second_double_word_16): mov 20(%rdi), %eax - sub 20(%rsi), %eax + cmp 20(%rsi), %eax + jne L(nequal) ret .p2align 4 @@ -840,29 +851,34 @@ L(next_two_double_words_16): and $15, %dh jz L(fourth_double_word_16) mov 24(%rdi), %eax - sub 24(%rsi), %eax + cmp 24(%rsi), %eax + jne L(nequal) ret .p2align 4 L(fourth_double_word_16): mov 28(%rdi), %eax - sub 28(%rsi), %eax + cmp 28(%rsi), %eax + jne L(nequal) ret .p2align 4 L(less4_double_words_32): + xor %eax, %eax test %dl, %dl jz L(next_two_double_words_32) and $15, %dl jz L(second_double_word_32) mov 32(%rdi), %eax - sub 32(%rsi), %eax + cmp 32(%rsi), %eax + jne L(nequal) ret .p2align 4 L(second_double_word_32): mov 36(%rdi), %eax - sub 36(%rsi), %eax + cmp 36(%rsi), %eax + jne L(nequal) ret .p2align 4 @@ -870,29 +886,34 @@ L(next_two_double_words_32): and $15, %dh jz L(fourth_double_word_32) mov 40(%rdi), %eax - sub 40(%rsi), %eax + cmp 40(%rsi), %eax + jne L(nequal) ret .p2align 4 L(fourth_double_word_32): mov 44(%rdi), %eax - sub 44(%rsi), %eax + cmp 44(%rsi), %eax + jne L(nequal) ret .p2align 4 L(less4_double_words_48): + xor %eax, %eax test %dl, %dl jz L(next_two_double_words_48) and $15, %dl jz L(second_double_word_48) mov 48(%rdi), %eax - sub 48(%rsi), %eax + cmp 48(%rsi), %eax + jne L(nequal) ret .p2align 4 L(second_double_word_48): mov 52(%rdi), %eax - sub 52(%rsi), %eax + cmp 52(%rsi), %eax + jne L(nequal) ret .p2align 4 @@ -900,23 +921,21 @@ L(next_two_double_words_48): and $15, %dh jz L(fourth_double_word_48) mov 56(%rdi), %eax - sub 56(%rsi), %eax + cmp 56(%rsi), %eax + jne L(nequal) ret .p2align 4 L(fourth_double_word_48): mov 60(%rdi), %eax - sub 60(%rsi), %eax - ret - - .p2align 4 -L(return): + cmp 60(%rsi), %eax + jne L(nequal) ret .p2align 4 L(nequal): mov $1, %eax - ja L(nequal_bigger) + jg L(nequal_bigger) neg %eax L(nequal_bigger): |