diff options
Diffstat (limited to 'sysdeps/i386/i686/multiarch/memcmp-ssse3.S')
| -rw-r--r-- | sysdeps/i386/i686/multiarch/memcmp-ssse3.S | 75 |
1 files changed, 70 insertions, 5 deletions
diff --git a/sysdeps/i386/i686/multiarch/memcmp-ssse3.S b/sysdeps/i386/i686/multiarch/memcmp-ssse3.S index bfcf660729..869f37a912 100644 --- a/sysdeps/i386/i686/multiarch/memcmp-ssse3.S +++ b/sysdeps/i386/i686/multiarch/memcmp-ssse3.S @@ -43,8 +43,7 @@ #define BLK2 BLK1+4 #define LEN BLK2+4 #define RETURN_END POP (%edi); POP (%esi); POP (%ebx); ret -#define RETURN RETURN_END; CFI_PUSH (%ebx); CFI_PUSH (%edi); \ - CFI_PUSH (%esi) +#define RETURN RETURN_END; cfi_restore_state; cfi_remember_state .section .text.ssse3,"ax",@progbits ENTRY (MEMCMP) @@ -76,12 +75,13 @@ L(1bytesend): L(zero): mov $0, %eax ret - + ALIGN (4) L(48bytesormore): PUSH (%ebx) PUSH (%esi) PUSH (%edi) + cfi_remember_state movdqu (%eax), %xmm3 movdqu (%edx), %xmm0 movl %eax, %edi @@ -155,7 +155,7 @@ L(shr_0): add $32, %esi sub $0xffff, %edx jnz L(exit) - + lea (%ecx, %edi,1), %eax lea (%ecx, %esi,1), %edx @@ -163,6 +163,8 @@ L(shr_0): POP (%esi) jmp L(less48bytes) + cfi_restore_state + cfi_remember_state ALIGN (4) L(shr_0_gobble): lea -48(%ecx), %ecx @@ -207,6 +209,8 @@ L(shr_0_gobble_loop_next): POP (%esi) jmp L(less48bytes) + cfi_restore_state + cfi_remember_state ALIGN (4) L(shr_1): cmp $80, %ecx @@ -235,6 +239,8 @@ L(shr_1): POP (%esi) jmp L(less48bytes) + cfi_restore_state + cfi_remember_state ALIGN (4) L(shr_1_gobble): sub $32, %ecx @@ -286,6 +292,8 @@ L(shr_1_gobble_next): jmp L(less48bytes) + cfi_restore_state + cfi_remember_state ALIGN (4) L(shr_2): cmp $80, %ecx @@ -314,6 +322,8 @@ L(shr_2): POP (%esi) jmp L(less48bytes) + cfi_restore_state + cfi_remember_state ALIGN (4) L(shr_2_gobble): sub $32, %ecx @@ -364,6 +374,8 @@ L(shr_2_gobble_next): POP (%esi) jmp L(less48bytes) + cfi_restore_state + cfi_remember_state ALIGN (4) L(shr_3): cmp $80, %ecx @@ -392,6 +404,8 @@ L(shr_3): POP (%esi) jmp L(less48bytes) + cfi_restore_state + cfi_remember_state ALIGN (4) L(shr_3_gobble): sub $32, %ecx @@ -442,6 +456,8 @@ L(shr_3_gobble_next): POP (%esi) jmp L(less48bytes) + cfi_restore_state + cfi_remember_state ALIGN (4) L(shr_4): cmp $80, %ecx @@ -470,6 +486,8 @@ L(shr_4): POP (%esi) jmp L(less48bytes) + cfi_restore_state + cfi_remember_state ALIGN (4) L(shr_4_gobble): sub $32, %ecx @@ -520,6 +538,8 @@ L(shr_4_gobble_next): POP (%esi) jmp L(less48bytes) + cfi_restore_state + cfi_remember_state ALIGN (4) L(shr_5): cmp $80, %ecx @@ -548,6 +568,8 @@ L(shr_5): POP (%esi) jmp L(less48bytes) + cfi_restore_state + cfi_remember_state ALIGN (4) L(shr_5_gobble): sub $32, %ecx @@ -598,6 +620,8 @@ L(shr_5_gobble_next): POP (%esi) jmp L(less48bytes) + cfi_restore_state + cfi_remember_state ALIGN (4) L(shr_6): cmp $80, %ecx @@ -626,6 +650,8 @@ L(shr_6): POP (%esi) jmp L(less48bytes) + cfi_restore_state + cfi_remember_state ALIGN (4) L(shr_6_gobble): sub $32, %ecx @@ -676,6 +702,8 @@ L(shr_6_gobble_next): POP (%esi) jmp L(less48bytes) + cfi_restore_state + cfi_remember_state ALIGN (4) L(shr_7): cmp $80, %ecx @@ -704,6 +732,8 @@ L(shr_7): POP (%esi) jmp L(less48bytes) + cfi_restore_state + cfi_remember_state ALIGN (4) L(shr_7_gobble): sub $32, %ecx @@ -754,6 +784,8 @@ L(shr_7_gobble_next): POP (%esi) jmp L(less48bytes) + cfi_restore_state + cfi_remember_state ALIGN (4) L(shr_8): cmp $80, %ecx @@ -782,6 +814,8 @@ L(shr_8): POP (%esi) jmp L(less48bytes) + cfi_restore_state + cfi_remember_state ALIGN (4) L(shr_8_gobble): sub $32, %ecx @@ -832,6 +866,8 @@ L(shr_8_gobble_next): POP (%esi) jmp L(less48bytes) + cfi_restore_state + cfi_remember_state ALIGN (4) L(shr_9): cmp $80, %ecx @@ -860,6 +896,8 @@ L(shr_9): POP (%esi) jmp L(less48bytes) + cfi_restore_state + cfi_remember_state ALIGN (4) L(shr_9_gobble): sub $32, %ecx @@ -910,6 +948,8 @@ L(shr_9_gobble_next): POP (%esi) jmp L(less48bytes) + cfi_restore_state + cfi_remember_state ALIGN (4) L(shr_10): cmp $80, %ecx @@ -938,6 +978,8 @@ L(shr_10): POP (%esi) jmp L(less48bytes) + cfi_restore_state + cfi_remember_state ALIGN (4) L(shr_10_gobble): sub $32, %ecx @@ -988,6 +1030,8 @@ L(shr_10_gobble_next): POP (%esi) jmp L(less48bytes) + cfi_restore_state + cfi_remember_state ALIGN (4) L(shr_11): cmp $80, %ecx @@ -1016,6 +1060,8 @@ L(shr_11): POP (%esi) jmp L(less48bytes) + cfi_restore_state + cfi_remember_state ALIGN (4) L(shr_11_gobble): sub $32, %ecx @@ -1066,6 +1112,8 @@ L(shr_11_gobble_next): POP (%esi) jmp L(less48bytes) + cfi_restore_state + cfi_remember_state ALIGN (4) L(shr_12): cmp $80, %ecx @@ -1094,6 +1142,8 @@ L(shr_12): POP (%esi) jmp L(less48bytes) + cfi_restore_state + cfi_remember_state ALIGN (4) L(shr_12_gobble): sub $32, %ecx @@ -1144,6 +1194,8 @@ L(shr_12_gobble_next): POP (%esi) jmp L(less48bytes) + cfi_restore_state + cfi_remember_state ALIGN (4) L(shr_13): cmp $80, %ecx @@ -1172,6 +1224,8 @@ L(shr_13): POP (%esi) jmp L(less48bytes) + cfi_restore_state + cfi_remember_state ALIGN (4) L(shr_13_gobble): sub $32, %ecx @@ -1222,6 +1276,8 @@ L(shr_13_gobble_next): POP (%esi) jmp L(less48bytes) + cfi_restore_state + cfi_remember_state ALIGN (4) L(shr_14): cmp $80, %ecx @@ -1250,6 +1306,8 @@ L(shr_14): POP (%esi) jmp L(less48bytes) + cfi_restore_state + cfi_remember_state ALIGN (4) L(shr_14_gobble): sub $32, %ecx @@ -1300,6 +1358,8 @@ L(shr_14_gobble_next): POP (%esi) jmp L(less48bytes) + cfi_restore_state + cfi_remember_state ALIGN (4) L(shr_15): cmp $80, %ecx @@ -1328,6 +1388,8 @@ L(shr_15): POP (%esi) jmp L(less48bytes) + cfi_restore_state + cfi_remember_state ALIGN (4) L(shr_15_gobble): sub $32, %ecx @@ -1378,6 +1440,8 @@ L(shr_15_gobble_next): POP (%esi) jmp L(less48bytes) + cfi_restore_state + cfi_remember_state ALIGN (4) L(exit): pmovmskb %xmm1, %ebx @@ -1497,8 +1561,9 @@ L(Byte31): movzbl -9(%edi), %eax movzbl -9(%esi), %edx sub %edx, %eax - RETURN + RETURN_END + CFI_PUSH (%ebx) ALIGN (4) L(more8bytes): cmp $16, %ecx |