diff options
Diffstat (limited to 'sysdeps/x86_64/multiarch/strcmp.S')
| -rw-r--r-- | sysdeps/x86_64/multiarch/strcmp.S | 361 |
1 files changed, 108 insertions, 253 deletions
diff --git a/sysdeps/x86_64/multiarch/strcmp.S b/sysdeps/x86_64/multiarch/strcmp.S index 15148e4f7f..1a315737af 100644 --- a/sysdeps/x86_64/multiarch/strcmp.S +++ b/sysdeps/x86_64/multiarch/strcmp.S @@ -28,9 +28,9 @@ /* calculate left number to compare */ \ lea -16(%rcx, %r11), %r9; \ cmp %r9, %r11; \ - jb LABEL(strcmp_exitz); \ + jb LABEL(strcmp_exitz_sse4_2); \ test %r9, %r9; \ - je LABEL(strcmp_exitz); \ + je LABEL(strcmp_exitz_sse4_2); \ mov %r9, %r11 #define STRCMP_SSE42 __strncmp_sse42 @@ -106,9 +106,9 @@ STRCMP_SSE42: */ #ifdef USE_AS_STRNCMP test %rdx, %rdx - je LABEL(strcmp_exitz) + je LABEL(strcmp_exitz_sse4_2) cmp $1, %rdx - je LABEL(Byte0) + je LABEL(Byte0_sse4_2) mov %rdx, %r11 #endif mov %esi, %ecx @@ -117,9 +117,9 @@ STRCMP_SSE42: and $0x3f, %rcx /* rsi alignment in cache line */ and $0x3f, %rax /* rdi alignment in cache line */ cmp $0x30, %ecx - ja LABEL(crosscache) /* rsi: 16-byte load will cross cache line */ + ja LABEL(crosscache_sse4_2)/* rsi: 16-byte load will cross cache line */ cmp $0x30, %eax - ja LABEL(crosscache) /* rdi: 16-byte load will cross cache line */ + ja LABEL(crosscache_sse4_2)/* rdi: 16-byte load will cross cache line */ movdqu (%rdi), %xmm1 movdqu (%rsi), %xmm2 pxor %xmm0, %xmm0 /* clear %xmm0 for null char checks */ @@ -128,10 +128,10 @@ STRCMP_SSE42: psubb %xmm0, %xmm1 /* packed sub of comparison results*/ pmovmskb %xmm1, %edx sub $0xffff, %edx /* if first 16 bytes are same, edx == 0xffff */ - jnz LABEL(less16bytes) /* If not, find different value or null char */ + jnz LABEL(less16bytes_sse4_2)/* If not, find different value or null char */ #ifdef USE_AS_STRNCMP sub $16, %r11 - jbe LABEL(strcmp_exitz) /* finish comparision */ + jbe LABEL(strcmp_exitz_sse4_2)/* finish comparision */ #endif add $16, %rsi /* prepare to search next 16 bytes */ add $16, %rdi /* prepare to search next 16 bytes */ @@ -142,7 +142,7 @@ STRCMP_SSE42: * below to use. */ .p2align 4 -LABEL(crosscache): +LABEL(crosscache_sse4_2): and $0xfffffffffffffff0, %rsi /* force %rsi is 16 byte aligned */ and $0xfffffffffffffff0, %rdi /* force %rdi is 16 byte aligned */ mov $0xffff, %edx /* for equivalent offset */ @@ -150,15 +150,15 @@ LABEL(crosscache): and $0xf, %ecx /* offset of rsi */ and $0xf, %eax /* offset of rdi */ cmp %eax, %ecx - je LABEL(ashr_0) /* rsi and rdi relative offset same */ - ja LABEL(bigger) + je LABEL(ashr_0_sse4_2) /* rsi and rdi relative offset same */ + ja LABEL(bigger_sse4_2) mov %edx, %r8d /* r8d is offset flag for exit tail */ xchg %ecx, %eax xchg %rsi, %rdi -LABEL(bigger): +LABEL(bigger_sse4_2): lea 15(%rax), %r9 sub %rcx, %r9 - lea LABEL(unaligned_table)(%rip), %r10 + lea LABEL(unaligned_table_sse4_2)(%rip), %r10 movslq (%r10, %r9,4), %r9 lea (%r10, %r9), %r10 jmp *%r10 /* jump to corresponding case */ @@ -169,7 +169,7 @@ LABEL(bigger): * n(0~15) n(0~15) 15(15+ n-n) ashr_0 */ .p2align 4 -LABEL(ashr_0): +LABEL(ashr_0_sse4_2): movdqa (%rsi), %xmm1 pxor %xmm0, %xmm0 /* clear %xmm0 for null char check */ @@ -184,7 +184,7 @@ LABEL(ashr_0): * edx must be the same with r9d if in left byte (16-rcx) is equal to * the start from (16-rax) and no null char was seen. */ - jne LABEL(less32bytes) /* mismatch or null char */ + jne LABEL(less32bytes_sse4_2) /* mismatch or null char */ UPDATE_STRNCMP_COUNTER mov $16, %rcx mov $16, %r9 @@ -203,7 +203,7 @@ LABEL(ashr_0_use_sse4_2): jbe LABEL(ashr_0_use_sse4_2_exit) #ifdef USE_AS_STRNCMP sub $16, %r11 - jbe LABEL(strcmp_exitz) + jbe LABEL(strcmp_exitz_sse4_2) #endif movdqa (%rdi,%rdx), %xmm0 @@ -212,17 +212,17 @@ LABEL(ashr_0_use_sse4_2): jbe LABEL(ashr_0_use_sse4_2_exit) #ifdef USE_AS_STRNCMP sub $16, %r11 - jbe LABEL(strcmp_exitz) + jbe LABEL(strcmp_exitz_sse4_2) #endif jmp LABEL(ashr_0_use_sse4_2) .p2align 4 LABEL(ashr_0_use_sse4_2_exit): - jnc LABEL(strcmp_exitz) + jnc LABEL(strcmp_exitz_sse4_2) #ifdef USE_AS_STRNCMP sub %rcx, %r11 - jbe LABEL(strcmp_exitz) + jbe LABEL(strcmp_exitz_sse4_2) #endif lea -16(%rdx, %rcx), %rcx movzbl (%rdi, %rcx), %eax @@ -239,7 +239,7 @@ LABEL(ashr_0_use_sse4_2_exit): * n(15) n -15 0(15 +(n-15) - n) ashr_1 */ .p2align 4 -LABEL(ashr_1): +LABEL(ashr_1_sse4_2): pxor %xmm0, %xmm0 movdqa (%rdi), %xmm2 movdqa (%rsi), %xmm1 @@ -251,7 +251,7 @@ LABEL(ashr_1): shr %cl, %edx /* adjust 0xffff for offset */ shr %cl, %r9d /* adjust for 16-byte offset */ sub %r9d, %edx - jnz LABEL(less32bytes) /* mismatch or null char seen */ + jnz LABEL(less32bytes_sse4_2)/* mismatch or null char seen */ movdqa (%rdi), %xmm3 UPDATE_STRNCMP_COUNTER @@ -279,7 +279,7 @@ LABEL(loop_ashr_1_use_sse4_2): jbe LABEL(use_sse4_2_exit) #ifdef USE_AS_STRNCMP sub $16, %r11 - jbe LABEL(strcmp_exitz) + jbe LABEL(strcmp_exitz_sse4_2) #endif add $16, %rdx @@ -292,7 +292,7 @@ LABEL(loop_ashr_1_use_sse4_2): jbe LABEL(use_sse4_2_exit) #ifdef USE_AS_STRNCMP sub $16, %r11 - jbe LABEL(strcmp_exitz) + jbe LABEL(strcmp_exitz_sse4_2) #endif add $16, %rdx jmp LABEL(loop_ashr_1_use_sse4_2) @@ -318,7 +318,7 @@ LABEL(nibble_ashr_1_use_sse4_2): * n(14~15) n -14 1(15 +(n-14) - n) ashr_2 */ .p2align 4 -LABEL(ashr_2): +LABEL(ashr_2_sse4_2): pxor %xmm0, %xmm0 movdqa (%rdi), %xmm2 movdqa (%rsi), %xmm1 @@ -330,7 +330,7 @@ LABEL(ashr_2): shr %cl, %edx shr %cl, %r9d sub %r9d, %edx - jnz LABEL(less32bytes) + jnz LABEL(less32bytes_sse4_2) movdqa (%rdi), %xmm3 UPDATE_STRNCMP_COUNTER @@ -358,7 +358,7 @@ LABEL(loop_ashr_2_use_sse4_2): jbe LABEL(use_sse4_2_exit) #ifdef USE_AS_STRNCMP sub $16, %r11 - jbe LABEL(strcmp_exitz) + jbe LABEL(strcmp_exitz_sse4_2) #endif add $16, %rdx @@ -371,7 +371,7 @@ LABEL(loop_ashr_2_use_sse4_2): jbe LABEL(use_sse4_2_exit) #ifdef USE_AS_STRNCMP sub $16, %r11 - jbe LABEL(strcmp_exitz) + jbe LABEL(strcmp_exitz_sse4_2) #endif add $16, %rdx jmp LABEL(loop_ashr_2_use_sse4_2) @@ -397,7 +397,7 @@ LABEL(nibble_ashr_2_use_sse4_2): * n(13~15) n -13 2(15 +(n-13) - n) ashr_3 */ .p2align 4 -LABEL(ashr_3): +LABEL(ashr_3_sse4_2): pxor %xmm0, %xmm0 movdqa (%rdi), %xmm2 movdqa (%rsi), %xmm1 @@ -409,7 +409,7 @@ LABEL(ashr_3): shr %cl, %edx shr %cl, %r9d sub %r9d, %edx - jnz LABEL(less32bytes) + jnz LABEL(less32bytes_sse4_2) movdqa (%rdi), %xmm3 UPDATE_STRNCMP_COUNTER @@ -437,7 +437,7 @@ LABEL(loop_ashr_3_use_sse4_2): jbe LABEL(use_sse4_2_exit) #ifdef USE_AS_STRNCMP sub $16, %r11 - jbe LABEL(strcmp_exitz) + jbe LABEL(strcmp_exitz_sse4_2) #endif add $16, %rdx @@ -450,7 +450,7 @@ LABEL(loop_ashr_3_use_sse4_2): jbe LABEL(use_sse4_2_exit) #ifdef USE_AS_STRNCMP sub $16, %r11 - jbe LABEL(strcmp_exitz) + jbe LABEL(strcmp_exitz_sse4_2) #endif add $16, %rdx jmp LABEL(loop_ashr_3_use_sse4_2) @@ -476,7 +476,7 @@ LABEL(nibble_ashr_3_use_sse4_2): * n(12~15) n -12 3(15 +(n-12) - n) ashr_4 */ .p2align 4 -LABEL(ashr_4): +LABEL(ashr_4_sse4_2): pxor %xmm0, %xmm0 movdqa (%rdi), %xmm2 movdqa (%rsi), %xmm1 @@ -488,7 +488,7 @@ LABEL(ashr_4): shr %cl, %edx shr %cl, %r9d sub %r9d, %edx - jnz LABEL(less32bytes) + jnz LABEL(less32bytes_sse4_2) movdqa (%rdi), %xmm3 UPDATE_STRNCMP_COUNTER @@ -517,7 +517,7 @@ LABEL(loop_ashr_4_use_sse4_2): jbe LABEL(use_sse4_2_exit) #ifdef USE_AS_STRNCMP sub $16, %r11 - jbe LABEL(strcmp_exitz) + jbe LABEL(strcmp_exitz_sse4_2) #endif add $16, %rdx @@ -530,7 +530,7 @@ LABEL(loop_ashr_4_use_sse4_2): jbe LABEL(use_sse4_2_exit) #ifdef USE_AS_STRNCMP sub $16, %r11 - jbe LABEL(strcmp_exitz) + jbe LABEL(strcmp_exitz_sse4_2) #endif add $16, %rdx jmp LABEL(loop_ashr_4_use_sse4_2) @@ -556,7 +556,7 @@ LABEL(nibble_ashr_4_use_sse4_2): * n(11~15) n - 11 4(15 +(n-11) - n) ashr_5 */ .p2align 4 -LABEL(ashr_5): +LABEL(ashr_5_sse4_2): pxor %xmm0, %xmm0 movdqa (%rdi), %xmm2 movdqa (%rsi), %xmm1 @@ -568,7 +568,7 @@ LABEL(ashr_5): shr %cl, %edx shr %cl, %r9d sub %r9d, %edx - jnz LABEL(less32bytes) + jnz LABEL(less32bytes_sse4_2) movdqa (%rdi), %xmm3 UPDATE_STRNCMP_COUNTER @@ -597,7 +597,7 @@ LABEL(loop_ashr_5_use_sse4_2): jbe LABEL(use_sse4_2_exit) #ifdef USE_AS_STRNCMP sub $16, %r11 - jbe LABEL(strcmp_exitz) + jbe LABEL(strcmp_exitz_sse4_2) #endif add $16, %rdx @@ -611,7 +611,7 @@ LABEL(loop_ashr_5_use_sse4_2): jbe LABEL(use_sse4_2_exit) #ifdef USE_AS_STRNCMP sub $16, %r11 - jbe LABEL(strcmp_exitz) + jbe LABEL(strcmp_exitz_sse4_2) #endif add $16, %rdx jmp LABEL(loop_ashr_5_use_sse4_2) @@ -637,7 +637,7 @@ LABEL(nibble_ashr_5_use_sse4_2): * n(10~15) n - 10 5(15 +(n-10) - n) ashr_6 */ .p2align 4 -LABEL(ashr_6): +LABEL(ashr_6_sse4_2): pxor %xmm0, %xmm0 movdqa (%rdi), %xmm2 movdqa (%rsi), %xmm1 @@ -649,7 +649,7 @@ LABEL(ashr_6): shr %cl, %edx shr %cl, %r9d sub %r9d, %edx - jnz LABEL(less32bytes) + jnz LABEL(less32bytes_sse4_2) movdqa (%rdi), %xmm3 UPDATE_STRNCMP_COUNTER @@ -678,7 +678,7 @@ LABEL(loop_ashr_6_use_sse4_2): jbe LABEL(use_sse4_2_exit) #ifdef USE_AS_STRNCMP sub $16, %r11 - jbe LABEL(strcmp_exitz) + jbe LABEL(strcmp_exitz_sse4_2) #endif add $16, %rdx @@ -691,7 +691,7 @@ LABEL(loop_ashr_6_use_sse4_2): jbe LABEL(use_sse4_2_exit) #ifdef USE_AS_STRNCMP sub $16, %r11 - jbe LABEL(strcmp_exitz) + jbe LABEL(strcmp_exitz_sse4_2) #endif add $16, %rdx jmp LABEL(loop_ashr_6_use_sse4_2) @@ -717,7 +717,7 @@ LABEL(nibble_ashr_6_use_sse4_2): * n(9~15) n - 9 6(15 +(n - 9) - n) ashr_7 */ .p2align 4 -LABEL(ashr_7): +LABEL(ashr_7_sse4_2): pxor %xmm0, %xmm0 movdqa (%rdi), %xmm2 movdqa (%rsi), %xmm1 @@ -729,7 +729,7 @@ LABEL(ashr_7): shr %cl, %edx shr %cl, %r9d sub %r9d, %edx - jnz LABEL(less32bytes) + jnz LABEL(less32bytes_sse4_2) movdqa (%rdi), %xmm3 UPDATE_STRNCMP_COUNTER @@ -758,7 +758,7 @@ LABEL(loop_ashr_7_use_sse4_2): jbe LABEL(use_sse4_2_exit) #ifdef USE_AS_STRNCMP sub $16, %r11 - jbe LABEL(strcmp_exitz) + jbe LABEL(strcmp_exitz_sse4_2) #endif add $16, %rdx @@ -771,7 +771,7 @@ LABEL(loop_ashr_7_use_sse4_2): jbe LABEL(use_sse4_2_exit) #ifdef USE_AS_STRNCMP sub $16, %r11 - jbe LABEL(strcmp_exitz) + jbe LABEL(strcmp_exitz_sse4_2) #endif add $16, %rdx jmp LABEL(loop_ashr_7_use_sse4_2) @@ -797,7 +797,7 @@ LABEL(nibble_ashr_7_use_sse4_2): * n(8~15) n - 8 7(15 +(n - 8) - n) ashr_8 */ .p2align 4 -LABEL(ashr_8): +LABEL(ashr_8_sse4_2): pxor %xmm0, %xmm0 movdqa (%rdi), %xmm2 movdqa (%rsi), %xmm1 @@ -809,7 +809,7 @@ LABEL(ashr_8): shr %cl, %edx shr %cl, %r9d sub %r9d, %edx - jnz LABEL(less32bytes) + jnz LABEL(less32bytes_sse4_2) movdqa (%rdi), %xmm3 UPDATE_STRNCMP_COUNTER @@ -838,7 +838,7 @@ LABEL(loop_ashr_8_use_sse4_2): jbe LABEL(use_sse4_2_exit) #ifdef USE_AS_STRNCMP sub $16, %r11 - jbe LABEL(strcmp_exitz) + jbe LABEL(strcmp_exitz_sse4_2) #endif add $16, %rdx @@ -851,7 +851,7 @@ LABEL(loop_ashr_8_use_sse4_2): jbe LABEL(use_sse4_2_exit) #ifdef USE_AS_STRNCMP sub $16, %r11 - jbe LABEL(strcmp_exitz) + jbe LABEL(strcmp_exitz_sse4_2) #endif add $16, %rdx jmp LABEL(loop_ashr_8_use_sse4_2) @@ -877,7 +877,7 @@ LABEL(nibble_ashr_8_use_sse4_2): * n(7~15) n - 7 8(15 +(n - 7) - n) ashr_9 */ .p2align 4 -LABEL(ashr_9): +LABEL(ashr_9_sse4_2): pxor %xmm0, %xmm0 movdqa (%rdi), %xmm2 movdqa (%rsi), %xmm1 @@ -889,7 +889,7 @@ LABEL(ashr_9): shr %cl, %edx shr %cl, %r9d sub %r9d, %edx - jnz LABEL(less32bytes) + jnz LABEL(less32bytes_sse4_2) movdqa (%rdi), %xmm3 UPDATE_STRNCMP_COUNTER @@ -919,7 +919,7 @@ LABEL(loop_ashr_9_use_sse4_2): jbe LABEL(use_sse4_2_exit) #ifdef USE_AS_STRNCMP sub $16, %r11 - jbe LABEL(strcmp_exitz) + jbe LABEL(strcmp_exitz_sse4_2) #endif add $16, %rdx @@ -932,7 +932,7 @@ LABEL(loop_ashr_9_use_sse4_2): jbe LABEL(use_sse4_2_exit) #ifdef USE_AS_STRNCMP sub $16, %r11 - jbe LABEL(strcmp_exitz) + jbe LABEL(strcmp_exitz_sse4_2) #endif add $16, %rdx jmp LABEL(loop_ashr_9_use_sse4_2) @@ -958,7 +958,7 @@ LABEL(nibble_ashr_9_use_sse4_2): * n(6~15) n - 6 9(15 +(n - 6) - n) ashr_10 */ .p2align 4 -LABEL(ashr_10): +LABEL(ashr_10_sse4_2): pxor %xmm0, %xmm0 movdqa (%rdi), %xmm2 movdqa (%rsi), %xmm1 @@ -970,7 +970,7 @@ LABEL(ashr_10): shr %cl, %edx shr %cl, %r9d sub %r9d, %edx - jnz LABEL(less32bytes) + jnz LABEL(less32bytes_sse4_2) movdqa (%rdi), %xmm3 UPDATE_STRNCMP_COUNTER @@ -999,7 +999,7 @@ LABEL(loop_ashr_10_use_sse4_2): jbe LABEL(use_sse4_2_exit) #ifdef USE_AS_STRNCMP sub $16, %r11 - jbe LABEL(strcmp_exitz) + jbe LABEL(strcmp_exitz_sse4_2) #endif add $16, %rdx @@ -1012,7 +1012,7 @@ LABEL(loop_ashr_10_use_sse4_2): jbe LABEL(use_sse4_2_exit) #ifdef USE_AS_STRNCMP sub $16, %r11 - jbe LABEL(strcmp_exitz) + jbe LABEL(strcmp_exitz_sse4_2) #endif add $16, %rdx jmp LABEL(loop_ashr_10_use_sse4_2) @@ -1038,7 +1038,7 @@ LABEL(nibble_ashr_10_use_sse4_2): * n(5~15) n - 5 10(15 +(n - 5) - n) ashr_11 */ .p2align 4 -LABEL(ashr_11): +LABEL(ashr_11_sse4_2): pxor %xmm0, %xmm0 movdqa (%rdi), %xmm2 movdqa (%rsi), %xmm1 @@ -1050,7 +1050,7 @@ LABEL(ashr_11): shr %cl, %edx shr %cl, %r9d sub %r9d, %edx - jnz LABEL(less32bytes) + jnz LABEL(less32bytes_sse4_2) movdqa (%rdi), %xmm3 UPDATE_STRNCMP_COUNTER @@ -1079,7 +1079,7 @@ LABEL(loop_ashr_11_use_sse4_2): jbe LABEL(use_sse4_2_exit) #ifdef USE_AS_STRNCMP sub $16, %r11 - jbe LABEL(strcmp_exitz) + jbe LABEL(strcmp_exitz_sse4_2) #endif add $16, %rdx @@ -1092,7 +1092,7 @@ LABEL(loop_ashr_11_use_sse4_2): jbe LABEL(use_sse4_2_exit) #ifdef USE_AS_STRNCMP sub $16, %r11 - jbe LABEL(strcmp_exitz) + jbe LABEL(strcmp_exitz_sse4_2) #endif add $16, %rdx jmp LABEL(loop_ashr_11_use_sse4_2) @@ -1118,7 +1118,7 @@ LABEL(nibble_ashr_11_use_sse4_2): * n(4~15) n - 4 11(15 +(n - 4) - n) ashr_12 */ .p2align 4 -LABEL(ashr_12): +LABEL(ashr_12_sse4_2): pxor %xmm0, %xmm0 movdqa (%rdi), %xmm2 movdqa (%rsi), %xmm1 @@ -1130,7 +1130,7 @@ LABEL(ashr_12): shr %cl, %edx shr %cl, %r9d sub %r9d, %edx - jnz LABEL(less32bytes) + jnz LABEL(less32bytes_sse4_2) movdqa (%rdi), %xmm3 UPDATE_STRNCMP_COUNTER @@ -1159,7 +1159,7 @@ LABEL(loop_ashr_12_use_sse4_2): jbe LABEL(use_sse4_2_exit) #ifdef USE_AS_STRNCMP sub $16, %r11 - jbe LABEL(strcmp_exitz) + jbe LABEL(strcmp_exitz_sse4_2) #endif add $16, %rdx @@ -1172,7 +1172,7 @@ LABEL(loop_ashr_12_use_sse4_2): jbe LABEL(use_sse4_2_exit) #ifdef USE_AS_STRNCMP sub $16, %r11 - jbe LABEL(strcmp_exitz) + jbe LABEL(strcmp_exitz_sse4_2) #endif add $16, %rdx jmp LABEL(loop_ashr_12_use_sse4_2) @@ -1198,7 +1198,7 @@ LABEL(nibble_ashr_12_use_sse4_2): * n(3~15) n - 3 12(15 +(n - 3) - n) ashr_13 */ .p2align 4 -LABEL(ashr_13): +LABEL(ashr_13_sse4_2): pxor %xmm0, %xmm0 movdqa (%rdi), %xmm2 movdqa (%rsi), %xmm1 @@ -1210,7 +1210,7 @@ LABEL(ashr_13): shr %cl, %edx shr %cl, %r9d sub %r9d, %edx - jnz LABEL(less32bytes) + jnz LABEL(less32bytes_sse4_2) movdqa (%rdi), %xmm3 UPDATE_STRNCMP_COUNTER @@ -1240,7 +1240,7 @@ LABEL(loop_ashr_13_use_sse4_2): jbe LABEL(use_sse4_2_exit) #ifdef USE_AS_STRNCMP sub $16, %r11 - jbe LABEL(strcmp_exitz) + jbe LABEL(strcmp_exitz_sse4_2) #endif add $16, %rdx @@ -1253,7 +1253,7 @@ LABEL(loop_ashr_13_use_sse4_2): jbe LABEL(use_sse4_2_exit) #ifdef USE_AS_STRNCMP sub $16, %r11 - jbe LABEL(strcmp_exitz) + jbe LABEL(strcmp_exitz_sse4_2) #endif add $16, %rdx jmp LABEL(loop_ashr_13_use_sse4_2) @@ -1279,7 +1279,7 @@ LABEL(nibble_ashr_13_use_sse4_2): * n(2~15) n - 2 13(15 +(n - 2) - n) ashr_14 */ .p2align 4 -LABEL(ashr_14): +LABEL(ashr_14_sse4_2): pxor %xmm0, %xmm0 movdqa (%rdi), %xmm2 movdqa (%rsi), %xmm1 @@ -1291,7 +1291,7 @@ LABEL(ashr_14): shr %cl, %edx shr %cl, %r9d sub %r9d, %edx - jnz LABEL(less32bytes) + jnz LABEL(less32bytes_sse4_2) movdqa (%rdi), %xmm3 UPDATE_STRNCMP_COUNTER @@ -1321,7 +1321,7 @@ LABEL(loop_ashr_14_use_sse4_2): jbe LABEL(use_sse4_2_exit) #ifdef USE_AS_STRNCMP sub $16, %r11 - jbe LABEL(strcmp_exitz) + jbe LABEL(strcmp_exitz_sse4_2) #endif add $16, %rdx @@ -1334,7 +1334,7 @@ LABEL(loop_ashr_14_use_sse4_2): jbe LABEL(use_sse4_2_exit) #ifdef USE_AS_STRNCMP sub $16, %r11 - jbe LABEL(strcmp_exitz) + jbe LABEL(strcmp_exitz_sse4_2) #endif add $16, %rdx jmp LABEL(loop_ashr_14_use_sse4_2) @@ -1360,7 +1360,7 @@ LABEL(nibble_ashr_14_use_sse4_2): * n(1~15) n - 1 14(15 +(n - 1) - n) ashr_15 */ .p2align 4 -LABEL(ashr_15): +LABEL(ashr_15_sse4_2): pxor %xmm0, %xmm0 movdqa (%rdi), %xmm2 movdqa (%rsi), %xmm1 @@ -1372,7 +1372,7 @@ LABEL(ashr_15): shr %cl, %edx shr %cl, %r9d sub %r9d, %edx - jnz LABEL(less32bytes) + jnz LABEL(less32bytes_sse4_2) movdqa (%rdi), %xmm3 @@ -1404,7 +1404,7 @@ LABEL(loop_ashr_15_use_sse4_2): jbe LABEL(use_sse4_2_exit) #ifdef USE_AS_STRNCMP sub $16, %r11 - jbe LABEL(strcmp_exitz) + jbe LABEL(strcmp_exitz_sse4_2) #endif add $16, %rdx @@ -1417,7 +1417,7 @@ LABEL(loop_ashr_15_use_sse4_2): jbe LABEL(use_sse4_2_exit) #ifdef USE_AS_STRNCMP sub $16, %r11 - jbe LABEL(strcmp_exitz) + jbe LABEL(strcmp_exitz_sse4_2) #endif add $16, %rdx jmp LABEL(loop_ashr_15_use_sse4_2) @@ -1439,56 +1439,37 @@ LABEL(nibble_ashr_use_sse4_2_exit): pcmpistri $0x1a,(%rsi,%rdx), %xmm0 .p2align 4 LABEL(use_sse4_2_exit): - jnc LABEL(strcmp_exitz) + jnc LABEL(strcmp_exitz_sse4_2) #ifdef USE_AS_STRNCMP sub %rcx, %r11 - jbe LABEL(strcmp_exitz) + jbe LABEL(strcmp_exitz_sse4_2) #endif add %rcx, %rdx lea -16(%rdi, %r9), %rdi movzbl (%rdi, %rdx), %eax movzbl (%rsi, %rdx), %edx test %r8d, %r8d - jz LABEL(use_sse4_2_ret) + jz LABEL(use_sse4_2_ret_sse4_2) xchg %eax, %edx -LABEL(use_sse4_2_ret): +LABEL(use_sse4_2_ret_sse4_2): sub %edx, %eax ret -#if 0 - /* This code was in the origial submission but isn't used. - --drepper */ - .p2align 4 -LABEL(aftertail): - pcmpeqb %xmm3, %xmm1 - psubb %xmm0, %xmm1 - pmovmskb %xmm1, %edx - not %edx - - .p2align 4 -LABEL(exit): - lea -16(%r9, %rcx), %rax /* locate the exact offset for rdi */ -#endif - -LABEL(less32bytes): +LABEL(less32bytes_sse4_2): lea (%rdi, %rax), %rdi /* locate the exact address for first operand(rdi) */ lea (%rsi, %rcx), %rsi /* locate the exact address for second operand(rsi) */ test %r8d, %r8d - jz LABEL(ret) + jz LABEL(ret_sse4_2) xchg %rsi, %rdi /* recover original order according to flag(%r8d) */ .p2align 4 -LABEL(ret): -LABEL(less16bytes): - /* - * Check to see if BSF is fast on this processor. If not, use a different - * exit tail. - */ +LABEL(ret_sse4_2): +LABEL(less16bytes_sse4_2): bsf %rdx, %rdx /* find and store bit index in %rdx */ #ifdef USE_AS_STRNCMP sub %rdx, %r11 - jbe LABEL(strcmp_exitz) + jbe LABEL(strcmp_exitz_sse4_2) #endif movzbl (%rsi, %rdx), %ecx movzbl (%rdi, %rdx), %eax @@ -1496,164 +1477,40 @@ LABEL(less16bytes): sub %ecx, %eax ret -LABEL(strcmp_exitz): +LABEL(strcmp_exitz_sse4_2): xor %eax, %eax ret .p2align 4 -LABEL(Byte0): - /* - * never need to handle byte 0 for strncmpy -#ifdef USE_AS_STRNCMP - sub $0, %r11 - jbe LABEL(strcmp_exitz) -#endif - */ +LABEL(Byte0_sse4_2): movzx (%rsi), %ecx movzx (%rdi), %eax sub %ecx, %eax ret - - .p2align 4 -LABEL(Byte1): - -#ifdef USE_AS_STRNCMP - sub $1, %r11 - jbe LABEL(strcmp_exitz) -#endif - movzx 1(%rsi), %ecx - movzx 1(%rdi), %eax - - sub %ecx, %eax - ret - - .p2align 4 -LABEL(Byte2): - -#ifdef USE_AS_STRNCMP - sub $2, %r11 - jbe LABEL(strcmp_exitz) -#endif - movzx 2(%rsi), %ecx - movzx 2(%rdi), %eax - - sub %ecx, %eax - ret - - .p2align 4 -LABEL(Byte3): - -#ifdef USE_AS_STRNCMP - sub $3, %r11 - jbe LABEL(strcmp_exitz) -#endif - movzx 3(%rsi), %ecx - movzx 3(%rdi), %eax - - sub %ecx, %eax - ret - - .p2align 4 -LABEL(Byte4): - -#ifdef USE_AS_STRNCMP - sub $4, %r11 - jbe LABEL(strcmp_exitz) -#endif - movzx 4(%rsi), %ecx - movzx 4(%rdi), %eax - - sub %ecx, %eax - ret - - .p2align 4 -LABEL(Byte5): - -#ifdef USE_AS_STRNCMP - sub $5, %r11 - jbe LABEL(strcmp_exitz) -#endif - movzx 5(%rsi), %ecx - movzx 5(%rdi), %eax - - sub %ecx, %eax - ret - - .p2align 4 -LABEL(Byte6): - -#ifdef USE_AS_STRNCMP - sub $6, %r11 - jbe LABEL(strcmp_exitz) -#endif - movzx 6(%rsi), %ecx - movzx 6(%rdi), %eax - - sub %ecx, %eax - ret - - .p2align 4 -LABEL(next_8_bytes): - add $8, %rdi - add $8, %rsi -#ifdef USE_AS_STRNCMP - sub $8, %r11 - jbe LABEL(strcmp_exitz) -#endif - test $0x01, %dh - jnz LABEL(Byte0) - - test $0x02, %dh - jnz LABEL(Byte1) - - test $0x04, %dh - jnz LABEL(Byte2) - - test $0x08, %dh - jnz LABEL(Byte3) - - test $0x10, %dh - jnz LABEL(Byte4) - - test $0x20, %dh - jnz LABEL(Byte5) - - test $0x40, %dh - jnz LABEL(Byte6) - -#ifdef USE_AS_STRNCMP - sub $7, %r11 - jbe LABEL(strcmp_exitz) -#endif - movzx 7(%rsi), %ecx - movzx 7(%rdi), %eax - - sub %ecx, %eax - ret cfi_endproc .size STRCMP_SSE42, .-STRCMP_SSE42 /* Put all SSE 4.2 functions together. */ .section .rodata.sse4.2,"a",@progbits - .p2align 4 -LABEL(unaligned_table): - .int LABEL(ashr_1) - LABEL(unaligned_table) - .int LABEL(ashr_2) - LABEL(unaligned_table) - .int LABEL(ashr_3) - LABEL(unaligned_table) - .int LABEL(ashr_4) - LABEL(unaligned_table) - .int LABEL(ashr_5) - LABEL(unaligned_table) - .int LABEL(ashr_6) - LABEL(unaligned_table) - .int LABEL(ashr_7) - LABEL(unaligned_table) - .int LABEL(ashr_8) - LABEL(unaligned_table) - .int LABEL(ashr_9) - LABEL(unaligned_table) - .int LABEL(ashr_10) - LABEL(unaligned_table) - .int LABEL(ashr_11) - LABEL(unaligned_table) - .int LABEL(ashr_12) - LABEL(unaligned_table) - .int LABEL(ashr_13) - LABEL(unaligned_table) - .int LABEL(ashr_14) - LABEL(unaligned_table) - .int LABEL(ashr_15) - LABEL(unaligned_table) - .int LABEL(ashr_0) - LABEL(unaligned_table) + .p2align 3 +LABEL(unaligned_table_sse4_2): + .int LABEL(ashr_1_sse4_2) - LABEL(unaligned_table_sse4_2) + .int LABEL(ashr_2_sse4_2) - LABEL(unaligned_table_sse4_2) + .int LABEL(ashr_3_sse4_2) - LABEL(unaligned_table_sse4_2) + .int LABEL(ashr_4_sse4_2) - LABEL(unaligned_table_sse4_2) + .int LABEL(ashr_5_sse4_2) - LABEL(unaligned_table_sse4_2) + .int LABEL(ashr_6_sse4_2) - LABEL(unaligned_table_sse4_2) + .int LABEL(ashr_7_sse4_2) - LABEL(unaligned_table_sse4_2) + .int LABEL(ashr_8_sse4_2) - LABEL(unaligned_table_sse4_2) + .int LABEL(ashr_9_sse4_2) - LABEL(unaligned_table_sse4_2) + .int LABEL(ashr_10_sse4_2) - LABEL(unaligned_table_sse4_2) + .int LABEL(ashr_11_sse4_2) - LABEL(unaligned_table_sse4_2) + .int LABEL(ashr_12_sse4_2) - LABEL(unaligned_table_sse4_2) + .int LABEL(ashr_13_sse4_2) - LABEL(unaligned_table_sse4_2) + .int LABEL(ashr_14_sse4_2) - LABEL(unaligned_table_sse4_2) + .int LABEL(ashr_15_sse4_2) - LABEL(unaligned_table_sse4_2) + .int LABEL(ashr_0_sse4_2) - LABEL(unaligned_table_sse4_2) # undef ENTRY @@ -1673,6 +1530,4 @@ LABEL(unaligned_table): .globl __GI_STRCMP; __GI_STRCMP = STRCMP_SSE2 #endif -#ifndef USE_AS_STRNCMP #include "../strcmp.S" -#endif |