diff options
Diffstat (limited to 'sysdeps/x86_64/multiarch/strcmp-sse42.S')
| -rw-r--r-- | sysdeps/x86_64/multiarch/strcmp-sse42.S | 238 |
1 files changed, 238 insertions, 0 deletions
diff --git a/sysdeps/x86_64/multiarch/strcmp-sse42.S b/sysdeps/x86_64/multiarch/strcmp-sse42.S index c84f1c2b31..edfa915707 100644 --- a/sysdeps/x86_64/multiarch/strcmp-sse42.S +++ b/sysdeps/x86_64/multiarch/strcmp-sse42.S @@ -127,6 +127,14 @@ STRCMP_SSE42: je LABEL(Byte0) mov %rdx, %r11 #endif + +#ifdef __CHKP__ + bndcl (%rdi), %bnd0 + bndcu (%rdi), %bnd0 + bndcl (%rsi), %bnd1 + bndcu (%rsi), %bnd1 +#endif + mov %esi, %ecx mov %edi, %eax /* Use 64bit AND here to avoid long NOP padding. */ @@ -210,6 +218,10 @@ LABEL(touppermask): #endif add $16, %rsi /* prepare to search next 16 bytes */ add $16, %rdi /* prepare to search next 16 bytes */ +#ifdef __CHKP__ + bndcu (%rdi), %bnd0 + bndcu (%rsi), %bnd1 +#endif /* * Determine source and destination string offsets from 16-byte @@ -231,6 +243,11 @@ LABEL(crosscache): mov %edx, %r8d /* r8d is offset flag for exit tail */ xchg %ecx, %eax xchg %rsi, %rdi +#ifdef __CHKP__ + bndmov %bnd0, %bnd2 + bndmov %bnd1, %bnd0 + bndmov %bnd2, %bnd1 +#endif LABEL(bigger): movdqa (%rdi), %xmm2 movdqa (%rsi), %xmm1 @@ -280,6 +297,10 @@ LABEL(ashr_0): mov %rcx, %rdx /* only for offset of sse4 instruction loop*/ .p2align 4 LABEL(ashr_0_use): +#ifdef __CHKP__ + bndcu -1(%rdi, %rdx), %bnd0 + bndcu -1(%rsi, %rdx), %bnd1 +#endif movdqa (%rdi,%rdx), %xmm0 #if !defined USE_AS_STRCASECMP_L && !defined USE_AS_STRNCASECMP_L pcmpistri $0x1a,(%rsi,%rdx), %xmm0 @@ -295,6 +316,10 @@ LABEL(ashr_0_use): jbe LABEL(strcmp_exitz) #endif +#ifdef __CHKP__ + bndcu -1(%rdi, %rdx), %bnd0 + bndcu -1(%rsi, %rdx), %bnd1 +#endif movdqa (%rdi,%rdx), %xmm0 #if !defined USE_AS_STRCASECMP_L && !defined USE_AS_STRNCASECMP_L pcmpistri $0x1a,(%rsi,%rdx), %xmm0 @@ -320,6 +345,10 @@ LABEL(ashr_0_exit_use): jbe LABEL(strcmp_exitz) #endif lea -16(%rdx, %rcx), %rcx +#ifdef __CHKP__ + bndcu -1(%rdi, %rcx), %bnd0 + bndcu -1(%rsi, %rcx), %bnd1 +#endif movzbl (%rdi, %rcx), %eax movzbl (%rsi, %rcx), %edx #if defined USE_AS_STRCASECMP_L || defined USE_AS_STRNCASECMP_L @@ -362,6 +391,15 @@ LABEL(ashr_1): and $0xfff, %r10 /* offset into 4K page */ sub $0x1000, %r10 /* subtract 4K pagesize */ mov %rcx, %rdx /* only for offset of sse4 instruction loop*/ +#ifdef __CHKP__ + bndcu -16(%rdi, %rdx), %bnd0 + bndcu -16(%rsi, %rdx), %bnd1 + jmp LABEL(loop_ashr_1_use) +LABEL(ashr_1_check): + bndcu (%rdi, %rdx), %bnd0 + bndcu (%rsi, %rdx), %bnd1 + jmp LABEL(nibble_ashr_1_restart_use) +#endif .p2align 4 LABEL(loop_ashr_1_use): @@ -416,7 +454,11 @@ LABEL(nibble_ashr_1_use): jae LABEL(nibble_ashr_exit_use) #endif cmp $14, %ecx +#ifdef __CHKP__ + ja LABEL(ashr_1_check) +#else ja LABEL(nibble_ashr_1_restart_use) +#endif jmp LABEL(nibble_ashr_exit_use) @@ -450,6 +492,15 @@ LABEL(ashr_2): and $0xfff, %r10 /* offset into 4K page */ sub $0x1000, %r10 /* subtract 4K pagesize */ mov %rcx, %rdx /* only for offset of sse4 instruction loop*/ +#ifdef __CHKP__ + bndcu -16(%rdi, %rdx), %bnd0 + bndcu -16(%rsi, %rdx), %bnd1 + jmp LABEL(loop_ashr_2_use) +LABEL(ashr_2_check): + bndcu (%rdi, %rdx), %bnd0 + bndcu (%rsi, %rdx), %bnd1 + jmp LABEL(nibble_ashr_2_restart_use) +#endif .p2align 4 LABEL(loop_ashr_2_use): @@ -504,7 +555,11 @@ LABEL(nibble_ashr_2_use): jae LABEL(nibble_ashr_exit_use) #endif cmp $13, %ecx +#ifdef __CHKP__ + ja LABEL(ashr_2_check) +#else ja LABEL(nibble_ashr_2_restart_use) +#endif jmp LABEL(nibble_ashr_exit_use) @@ -539,6 +594,15 @@ LABEL(ashr_3): and $0xfff, %r10 /* offset into 4K page */ sub $0x1000, %r10 /* subtract 4K pagesize */ mov %rcx, %rdx /* only for offset of sse4 instruction loop*/ +#ifdef __CHKP__ + bndcu -16(%rdi, %rdx), %bnd0 + bndcu -16(%rsi, %rdx), %bnd1 + jmp LABEL(loop_ashr_3_use) +LABEL(ashr_3_check): + bndcu (%rdi, %rdx), %bnd0 + bndcu (%rsi, %rdx), %bnd1 + jmp LABEL(nibble_ashr_3_restart_use) +#endif LABEL(loop_ashr_3_use): add $16, %r10 @@ -592,7 +656,11 @@ LABEL(nibble_ashr_3_use): jae LABEL(nibble_ashr_exit_use) #endif cmp $12, %ecx +#ifdef __CHKP__ + ja LABEL(ashr_3_check) +#else ja LABEL(nibble_ashr_3_restart_use) +#endif jmp LABEL(nibble_ashr_exit_use) @@ -627,6 +695,15 @@ LABEL(ashr_4): and $0xfff, %r10 /* offset into 4K page */ sub $0x1000, %r10 /* subtract 4K pagesize */ mov %rcx, %rdx /* only for offset of sse4 instruction loop*/ +#ifdef __CHKP__ + bndcu -16(%rdi, %rdx), %bnd0 + bndcu -16(%rsi, %rdx), %bnd1 + jmp LABEL(loop_ashr_4_use) +LABEL(ashr_4_check): + bndcu (%rdi, %rdx), %bnd0 + bndcu (%rsi, %rdx), %bnd1 + jmp LABEL(nibble_ashr_4_restart_use) +#endif .p2align 4 LABEL(loop_ashr_4_use): @@ -681,7 +758,11 @@ LABEL(nibble_ashr_4_use): jae LABEL(nibble_ashr_exit_use) #endif cmp $11, %ecx +#ifdef __CHKP__ + ja LABEL(ashr_4_check) +#else ja LABEL(nibble_ashr_4_restart_use) +#endif jmp LABEL(nibble_ashr_exit_use) @@ -716,6 +797,15 @@ LABEL(ashr_5): and $0xfff, %r10 /* offset into 4K page */ sub $0x1000, %r10 /* subtract 4K pagesize */ mov %rcx, %rdx /* only for offset of sse4 instruction loop*/ +#ifdef __CHKP__ + bndcu -16(%rdi, %rdx), %bnd0 + bndcu -16(%rsi, %rdx), %bnd1 + jmp LABEL(loop_ashr_5_use) +LABEL(ashr_5_check): + bndcu (%rdi, %rdx), %bnd0 + bndcu (%rsi, %rdx), %bnd1 + jmp LABEL(nibble_ashr_5_restart_use) +#endif .p2align 4 LABEL(loop_ashr_5_use): @@ -771,7 +861,11 @@ LABEL(nibble_ashr_5_use): jae LABEL(nibble_ashr_exit_use) #endif cmp $10, %ecx +#ifdef __CHKP__ + ja LABEL(ashr_5_check) +#else ja LABEL(nibble_ashr_5_restart_use) +#endif jmp LABEL(nibble_ashr_exit_use) @@ -806,6 +900,15 @@ LABEL(ashr_6): and $0xfff, %r10 /* offset into 4K page */ sub $0x1000, %r10 /* subtract 4K pagesize */ mov %rcx, %rdx /* only for offset of sse4 instruction loop*/ +#ifdef __CHKP__ + bndcu -16(%rdi, %rdx), %bnd0 + bndcu -16(%rsi, %rdx), %bnd1 + jmp LABEL(loop_ashr_6_use) +LABEL(ashr_6_check): + bndcu (%rdi, %rdx), %bnd0 + bndcu (%rsi, %rdx), %bnd1 + jmp LABEL(nibble_ashr_6_restart_use) +#endif .p2align 4 LABEL(loop_ashr_6_use): @@ -860,7 +963,11 @@ LABEL(nibble_ashr_6_use): jae LABEL(nibble_ashr_exit_use) #endif cmp $9, %ecx +#ifdef __CHKP__ + ja LABEL(ashr_6_check) +#else ja LABEL(nibble_ashr_6_restart_use) +#endif jmp LABEL(nibble_ashr_exit_use) @@ -895,6 +1002,15 @@ LABEL(ashr_7): and $0xfff, %r10 /* offset into 4K page */ sub $0x1000, %r10 /* subtract 4K pagesize */ mov %rcx, %rdx /* only for offset of sse4 instruction loop*/ +#ifdef __CHKP__ + bndcu -16(%rdi, %rdx), %bnd0 + bndcu -16(%rsi, %rdx), %bnd1 + jmp LABEL(loop_ashr_7_use) +LABEL(ashr_7_check): + bndcu (%rdi, %rdx), %bnd0 + bndcu (%rsi, %rdx), %bnd1 + jmp LABEL(nibble_ashr_7_restart_use) +#endif .p2align 4 LABEL(loop_ashr_7_use): @@ -949,7 +1065,11 @@ LABEL(nibble_ashr_7_use): jae LABEL(nibble_ashr_exit_use) #endif cmp $8, %ecx +#ifdef __CHKP__ + ja LABEL(ashr_7_check) +#else ja LABEL(nibble_ashr_7_restart_use) +#endif jmp LABEL(nibble_ashr_exit_use) @@ -984,6 +1104,15 @@ LABEL(ashr_8): and $0xfff, %r10 /* offset into 4K page */ sub $0x1000, %r10 /* subtract 4K pagesize */ mov %rcx, %rdx /* only for offset of sse4 instruction loop*/ +#ifdef __CHKP__ + bndcu -16(%rdi, %rdx), %bnd0 + bndcu -16(%rsi, %rdx), %bnd1 + jmp LABEL(loop_ashr_8_use) +LABEL(ashr_8_check): + bndcu (%rdi, %rdx), %bnd0 + bndcu (%rsi, %rdx), %bnd1 + jmp LABEL(nibble_ashr_8_restart_use) +#endif .p2align 4 LABEL(loop_ashr_8_use): @@ -1038,7 +1167,11 @@ LABEL(nibble_ashr_8_use): jae LABEL(nibble_ashr_exit_use) #endif cmp $7, %ecx +#ifdef __CHKP__ + ja LABEL(ashr_8_check) +#else ja LABEL(nibble_ashr_8_restart_use) +#endif jmp LABEL(nibble_ashr_exit_use) @@ -1073,6 +1206,15 @@ LABEL(ashr_9): and $0xfff, %r10 /* offset into 4K page */ sub $0x1000, %r10 /* subtract 4K pagesize */ mov %rcx, %rdx /* only for offset of sse4 instruction loop*/ +#ifdef __CHKP__ + bndcu -16(%rdi, %rdx), %bnd0 + bndcu -16(%rsi, %rdx), %bnd1 + jmp LABEL(loop_ashr_9_use) +LABEL(ashr_9_check): + bndcu (%rdi, %rdx), %bnd0 + bndcu (%rsi, %rdx), %bnd1 + jmp LABEL(nibble_ashr_9_restart_use) +#endif .p2align 4 LABEL(loop_ashr_9_use): @@ -1128,7 +1270,11 @@ LABEL(nibble_ashr_9_use): jae LABEL(nibble_ashr_exit_use) #endif cmp $6, %ecx +#ifdef __CHKP__ + ja LABEL(ashr_9_check) +#else ja LABEL(nibble_ashr_9_restart_use) +#endif jmp LABEL(nibble_ashr_exit_use) @@ -1163,6 +1309,15 @@ LABEL(ashr_10): and $0xfff, %r10 /* offset into 4K page */ sub $0x1000, %r10 /* subtract 4K pagesize */ mov %rcx, %rdx /* only for offset of sse4 instruction loop*/ +#ifdef __CHKP__ + bndcu -16(%rdi, %rdx), %bnd0 + bndcu -16(%rsi, %rdx), %bnd1 + jmp LABEL(loop_ashr_10_use) +LABEL(ashr_10_check): + bndcu (%rdi, %rdx), %bnd0 + bndcu (%rsi, %rdx), %bnd1 + jmp LABEL(nibble_ashr_10_restart_use) +#endif .p2align 4 LABEL(loop_ashr_10_use): @@ -1217,7 +1372,11 @@ LABEL(nibble_ashr_10_use): jae LABEL(nibble_ashr_exit_use) #endif cmp $5, %ecx +#ifdef __CHKP__ + ja LABEL(ashr_10_check) +#else ja LABEL(nibble_ashr_10_restart_use) +#endif jmp LABEL(nibble_ashr_exit_use) @@ -1252,6 +1411,15 @@ LABEL(ashr_11): and $0xfff, %r10 /* offset into 4K page */ sub $0x1000, %r10 /* subtract 4K pagesize */ mov %rcx, %rdx /* only for offset of sse4 instruction loop*/ +#ifdef __CHKP__ + bndcu -16(%rdi, %rdx), %bnd0 + bndcu -16(%rsi, %rdx), %bnd1 + jmp LABEL(loop_ashr_11_use) +LABEL(ashr_11_check): + bndcu (%rdi, %rdx), %bnd0 + bndcu (%rsi, %rdx), %bnd1 + jmp LABEL(nibble_ashr_11_restart_use) +#endif .p2align 4 LABEL(loop_ashr_11_use): @@ -1306,7 +1474,11 @@ LABEL(nibble_ashr_11_use): jae LABEL(nibble_ashr_exit_use) #endif cmp $4, %ecx +#ifdef __CHKP__ + ja LABEL(ashr_11_check) +#else ja LABEL(nibble_ashr_11_restart_use) +#endif jmp LABEL(nibble_ashr_exit_use) @@ -1341,6 +1513,15 @@ LABEL(ashr_12): and $0xfff, %r10 /* offset into 4K page */ sub $0x1000, %r10 /* subtract 4K pagesize */ mov %rcx, %rdx /* only for offset of sse4 instruction loop*/ +#ifdef __CHKP__ + bndcu -16(%rdi, %rdx), %bnd0 + bndcu -16(%rsi, %rdx), %bnd1 + jmp LABEL(loop_ashr_12_use) +LABEL(ashr_12_check): + bndcu (%rdi, %rdx), %bnd0 + bndcu (%rsi, %rdx), %bnd1 + jmp LABEL(nibble_ashr_12_restart_use) +#endif .p2align 4 LABEL(loop_ashr_12_use): @@ -1395,7 +1576,11 @@ LABEL(nibble_ashr_12_use): jae LABEL(nibble_ashr_exit_use) #endif cmp $3, %ecx +#ifdef __CHKP__ + ja LABEL(ashr_12_check) +#else ja LABEL(nibble_ashr_12_restart_use) +#endif jmp LABEL(nibble_ashr_exit_use) @@ -1431,6 +1616,15 @@ LABEL(ashr_13): sub $0x1000, %r10 /* subtract 4K pagesize */ mov %rcx, %rdx /* only for offset of sse4 instruction loop*/ +#ifdef __CHKP__ + bndcu -16(%rdi, %rdx), %bnd0 + bndcu -16(%rsi, %rdx), %bnd1 + jmp LABEL(loop_ashr_13_use) +LABEL(ashr_13_check): + bndcu (%rdi, %rdx), %bnd0 + bndcu (%rsi, %rdx), %bnd1 + jmp LABEL(nibble_ashr_13_restart_use) +#endif .p2align 4 LABEL(loop_ashr_13_use): @@ -1485,7 +1679,11 @@ LABEL(nibble_ashr_13_use): jae LABEL(nibble_ashr_exit_use) #endif cmp $2, %ecx +#ifdef __CHKP__ + ja LABEL(ashr_13_check) +#else ja LABEL(nibble_ashr_13_restart_use) +#endif jmp LABEL(nibble_ashr_exit_use) @@ -1521,6 +1719,15 @@ LABEL(ashr_14): sub $0x1000, %r10 /* subtract 4K pagesize */ mov %rcx, %rdx /* only for offset of sse4 instruction loop*/ +#ifdef __CHKP__ + bndcu -16(%rdi, %rdx), %bnd0 + bndcu -16(%rsi, %rdx), %bnd1 + jmp LABEL(loop_ashr_14_use) +LABEL(ashr_14_check): + bndcu (%rdi, %rdx), %bnd0 + bndcu (%rsi, %rdx), %bnd1 + jmp LABEL(nibble_ashr_14_restart_use) +#endif .p2align 4 LABEL(loop_ashr_14_use): @@ -1575,7 +1782,11 @@ LABEL(nibble_ashr_14_use): jae LABEL(nibble_ashr_exit_use) #endif cmp $1, %ecx +#ifdef __CHKP__ + ja LABEL(ashr_14_check) +#else ja LABEL(nibble_ashr_14_restart_use) +#endif jmp LABEL(nibble_ashr_exit_use) @@ -1613,6 +1824,15 @@ LABEL(ashr_15): sub $0x1000, %r10 /* subtract 4K pagesize */ mov %rcx, %rdx /* only for offset of sse4 instruction loop*/ +#ifdef __CHKP__ + bndcu -16(%rdi, %rdx), %bnd0 + bndcu -16(%rsi, %rdx), %bnd1 + jmp LABEL(loop_ashr_15_use) +LABEL(ashr_15_check): + bndcu (%rdi, %rdx), %bnd0 + bndcu (%rsi, %rdx), %bnd1 + jmp LABEL(nibble_ashr_15_restart_use) +#endif .p2align 4 LABEL(loop_ashr_15_use): @@ -1667,7 +1887,11 @@ LABEL(nibble_ashr_15_use): jae LABEL(nibble_ashr_exit_use) #endif cmp $0, %ecx +#ifdef __CHKP__ + ja LABEL(ashr_15_check) +#else ja LABEL(nibble_ashr_15_restart_use) +#endif LABEL(nibble_ashr_exit_use): #if !defined USE_AS_STRCASECMP_L && !defined USE_AS_STRNCASECMP_L @@ -1691,6 +1915,11 @@ LABEL(exit_use): test %r8d, %r8d jz LABEL(ret_use) xchg %eax, %edx +#ifdef __CHKP__ + bndmov %bnd0, %bnd2 + bndmov %bnd1, %bnd0 + bndmov %bnd2, %bnd1 +#endif LABEL(ret_use): #if defined USE_AS_STRCASECMP_L || defined USE_AS_STRNCASECMP_L leaq _nl_C_LC_CTYPE_tolower+128*4(%rip), %rcx @@ -1707,6 +1936,11 @@ LABEL(less32bytes): test %r8d, %r8d jz LABEL(ret) xchg %rsi, %rdi /* recover original order according to flag(%r8d) */ +#ifdef __CHKP__ + bndmov %bnd0, %bnd2 + bndmov %bnd1, %bnd0 + bndmov %bnd2, %bnd1 +#endif .p2align 4 LABEL(ret): @@ -1717,6 +1951,10 @@ LABEL(less16bytes): sub %rdx, %r11 jbe LABEL(strcmp_exitz) #endif +#ifdef __CHKP__ + bndcu (%rdi, %rdx), %bnd0 + bndcu (%rsi, %rdx), %bnd1 +#endif movzbl (%rsi, %rdx), %ecx movzbl (%rdi, %rdx), %eax |