diff options
Diffstat (limited to 'sysdeps/x86_64/strcmp.S')
| -rw-r--r-- | sysdeps/x86_64/strcmp.S | 238 |
1 files changed, 237 insertions, 1 deletions
diff --git a/sysdeps/x86_64/strcmp.S b/sysdeps/x86_64/strcmp.S index 76809373e8..89597841ad 100644 --- a/sysdeps/x86_64/strcmp.S +++ b/sysdeps/x86_64/strcmp.S @@ -128,7 +128,16 @@ libc_hidden_def (__strncasecmp) ENTRY (STRCMP) #ifdef NOT_IN_libc /* Simple version since we can't use SSE registers in ld.so. */ -L(oop): movb (%rdi), %al +#ifdef __CHKP__ + bndcl (%rdi), %bnd0 + bndcl (%rsi), %bnd1 +#endif +L(oop): +#ifdef __CHKP__ + bndcu (%rdi), %bnd0 + bndcu (%rsi), %bnd1 +#endif + movb (%rdi), %al cmpb (%rsi), %al jne L(neq) incq %rdi @@ -177,6 +186,12 @@ END (STRCMP) 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. */ @@ -243,6 +258,10 @@ END (STRCMP) # 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 alignment. @@ -263,6 +282,11 @@ 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): lea 15(%rax), %r9 sub %rcx, %r9 @@ -310,6 +334,10 @@ LABEL(ashr_0): */ .p2align 4 LABEL(loop_ashr_0): +#ifdef __CHKP__ + bndcu -1(%rdi, %rcx), %bnd0 + bndcu -1(%rsi, %rcx), %bnd1 +#endif movdqa (%rsi, %rcx), %xmm1 movdqa (%rdi, %rcx), %xmm2 TOLOWER (%xmm1, %xmm2) @@ -326,6 +354,10 @@ LABEL(loop_ashr_0): jbe LABEL(strcmp_exitz) # endif add $16, %rcx +#ifdef __CHKP__ + bndcu -1(%rdi, %rcx), %bnd0 + bndcu -1(%rsi, %rcx), %bnd1 +#endif movdqa (%rsi, %rcx), %xmm1 movdqa (%rdi, %rcx), %xmm2 TOLOWER (%xmm1, %xmm2) @@ -377,6 +409,15 @@ LABEL(ashr_1): lea 1(%rdi), %r10 and $0xfff, %r10 /* offset into 4K page */ sub $0x1000, %r10 /* subtract 4K pagesize */ +# ifdef __CHKP__ + bndcu -16(%rdi, %rcx), %bnd0 + bndcu -16(%rsi, %rcx), %bnd1 + jmp LABEL(loop_ashr_1) +LABEL(ashr_1_check): + bndcu (%rdi, %rcx), %bnd0 + bndcu (%rsi, %rcx), %bnd1 + jmp LABEL(gobble_ashr_1) +# endif .p2align 4 LABEL(loop_ashr_1): @@ -460,7 +501,11 @@ LABEL(nibble_ashr_1): pxor %xmm0, %xmm0 sub $0x1000, %r10 /* substract 4K from %r10 */ +# ifdef __CHKP__ + jmp LABEL(ashr_1_check) +# else jmp LABEL(gobble_ashr_1) +# endif /* * Once find null char, determine if there is a string mismatch @@ -507,6 +552,15 @@ LABEL(ashr_2): lea 2(%rdi), %r10 and $0xfff, %r10 /* offset into 4K page */ sub $0x1000, %r10 /* subtract 4K pagesize */ +# ifdef __CHKP__ + bndcu -16(%rdi, %rcx), %bnd0 + bndcu -16(%rsi, %rcx), %bnd1 + jmp LABEL(loop_ashr_2) +LABEL(ashr_2_check): + bndcu (%rdi, %rcx), %bnd0 + bndcu (%rsi, %rcx), %bnd1 + jmp LABEL(gobble_ashr_2) +# endif .p2align 4 LABEL(loop_ashr_2): @@ -588,7 +642,11 @@ LABEL(nibble_ashr_2): pxor %xmm0, %xmm0 sub $0x1000, %r10 +# ifdef __CHKP__ + jmp LABEL(ashr_2_check) +# else jmp LABEL(gobble_ashr_2) +# endif .p2align 4 LABEL(ashr_2_exittail): @@ -632,6 +690,15 @@ LABEL(ashr_3): lea 3(%rdi), %r10 and $0xfff, %r10 /* offset into 4K page */ sub $0x1000, %r10 /* subtract 4K pagesize */ +# ifdef __CHKP__ + bndcu -16(%rdi, %rcx), %bnd0 + bndcu -16(%rsi, %rcx), %bnd1 + jmp LABEL(loop_ashr_3) +LABEL(ashr_3_check): + bndcu (%rdi, %rcx), %bnd0 + bndcu (%rsi, %rcx), %bnd1 + jmp LABEL(gobble_ashr_3) +# endif .p2align 4 LABEL(loop_ashr_3): @@ -713,7 +780,11 @@ LABEL(nibble_ashr_3): pxor %xmm0, %xmm0 sub $0x1000, %r10 +# ifdef __CHKP__ + jmp LABEL(ashr_3_check) +# else jmp LABEL(gobble_ashr_3) +# endif .p2align 4 LABEL(ashr_3_exittail): @@ -757,6 +828,15 @@ LABEL(ashr_4): lea 4(%rdi), %r10 and $0xfff, %r10 /* offset into 4K page */ sub $0x1000, %r10 /* subtract 4K pagesize */ +# ifdef __CHKP__ + bndcu -16(%rdi, %rcx), %bnd0 + bndcu -16(%rsi, %rcx), %bnd1 + jmp LABEL(loop_ashr_4) +LABEL(ashr_4_check): + bndcu (%rdi, %rcx), %bnd0 + bndcu (%rsi, %rcx), %bnd1 + jmp LABEL(gobble_ashr_4) +# endif .p2align 4 LABEL(loop_ashr_4): @@ -838,7 +918,11 @@ LABEL(nibble_ashr_4): pxor %xmm0, %xmm0 sub $0x1000, %r10 +# ifdef __CHKP__ + jmp LABEL(ashr_4_check) +# else jmp LABEL(gobble_ashr_4) +# endif .p2align 4 LABEL(ashr_4_exittail): @@ -882,6 +966,15 @@ LABEL(ashr_5): lea 5(%rdi), %r10 and $0xfff, %r10 /* offset into 4K page */ sub $0x1000, %r10 /* subtract 4K pagesize */ +# ifdef __CHKP__ + bndcu -16(%rdi, %rcx), %bnd0 + bndcu -16(%rsi, %rcx), %bnd1 + jmp LABEL(loop_ashr_5) +LABEL(ashr_5_check): + bndcu (%rdi, %rcx), %bnd0 + bndcu (%rsi, %rcx), %bnd1 + jmp LABEL(gobble_ashr_5) +# endif .p2align 4 LABEL(loop_ashr_5): @@ -963,7 +1056,11 @@ LABEL(nibble_ashr_5): pxor %xmm0, %xmm0 sub $0x1000, %r10 +# ifdef __CHKP__ + jmp LABEL(ashr_5_check) +# else jmp LABEL(gobble_ashr_5) +# endif .p2align 4 LABEL(ashr_5_exittail): @@ -1007,6 +1104,15 @@ LABEL(ashr_6): lea 6(%rdi), %r10 and $0xfff, %r10 /* offset into 4K page */ sub $0x1000, %r10 /* subtract 4K pagesize */ +# ifdef __CHKP__ + bndcu -16(%rdi, %rcx), %bnd0 + bndcu -16(%rsi, %rcx), %bnd1 + jmp LABEL(loop_ashr_6) +LABEL(ashr_6_check): + bndcu (%rdi, %rcx), %bnd0 + bndcu (%rsi, %rcx), %bnd1 + jmp LABEL(gobble_ashr_6) +# endif .p2align 4 LABEL(loop_ashr_6): @@ -1088,7 +1194,11 @@ LABEL(nibble_ashr_6): pxor %xmm0, %xmm0 sub $0x1000, %r10 +# ifdef __CHKP__ + jmp LABEL(ashr_6_check) +# else jmp LABEL(gobble_ashr_6) +# endif .p2align 4 LABEL(ashr_6_exittail): @@ -1132,6 +1242,15 @@ LABEL(ashr_7): lea 7(%rdi), %r10 and $0xfff, %r10 /* offset into 4K page */ sub $0x1000, %r10 /* subtract 4K pagesize */ +# ifdef __CHKP__ + bndcu -16(%rdi, %rcx), %bnd0 + bndcu -16(%rsi, %rcx), %bnd1 + jmp LABEL(loop_ashr_7) +LABEL(ashr_7_check): + bndcu (%rdi, %rcx), %bnd0 + bndcu (%rsi, %rcx), %bnd1 + jmp LABEL(gobble_ashr_7) +# endif .p2align 4 LABEL(loop_ashr_7): @@ -1213,7 +1332,11 @@ LABEL(nibble_ashr_7): pxor %xmm0, %xmm0 sub $0x1000, %r10 +# ifdef __CHKP__ + jmp LABEL(ashr_7_check) +# else jmp LABEL(gobble_ashr_7) +# endif .p2align 4 LABEL(ashr_7_exittail): @@ -1257,6 +1380,15 @@ LABEL(ashr_8): lea 8(%rdi), %r10 and $0xfff, %r10 /* offset into 4K page */ sub $0x1000, %r10 /* subtract 4K pagesize */ +# ifdef __CHKP__ + bndcu -16(%rdi, %rcx), %bnd0 + bndcu -16(%rsi, %rcx), %bnd1 + jmp LABEL(loop_ashr_8) +LABEL(ashr_8_check): + bndcu (%rdi, %rcx), %bnd0 + bndcu (%rsi, %rcx), %bnd1 + jmp LABEL(gobble_ashr_8) +# endif .p2align 4 LABEL(loop_ashr_8): @@ -1338,7 +1470,11 @@ LABEL(nibble_ashr_8): pxor %xmm0, %xmm0 sub $0x1000, %r10 +# ifdef __CHKP__ + jmp LABEL(ashr_8_check) +# else jmp LABEL(gobble_ashr_8) +# endif .p2align 4 LABEL(ashr_8_exittail): @@ -1382,6 +1518,15 @@ LABEL(ashr_9): lea 9(%rdi), %r10 and $0xfff, %r10 /* offset into 4K page */ sub $0x1000, %r10 /* subtract 4K pagesize */ +# ifdef __CHKP__ + bndcu -16(%rdi, %rcx), %bnd0 + bndcu -16(%rsi, %rcx), %bnd1 + jmp LABEL(loop_ashr_9) +LABEL(ashr_9_check): + bndcu (%rdi, %rcx), %bnd0 + bndcu (%rsi, %rcx), %bnd1 + jmp LABEL(gobble_ashr_9) +# endif .p2align 4 LABEL(loop_ashr_9): @@ -1463,7 +1608,11 @@ LABEL(nibble_ashr_9): pxor %xmm0, %xmm0 sub $0x1000, %r10 +# ifdef __CHKP__ + jmp LABEL(ashr_9_check) +# else jmp LABEL(gobble_ashr_9) +# endif .p2align 4 LABEL(ashr_9_exittail): @@ -1507,6 +1656,15 @@ LABEL(ashr_10): lea 10(%rdi), %r10 and $0xfff, %r10 /* offset into 4K page */ sub $0x1000, %r10 /* subtract 4K pagesize */ +# ifdef __CHKP__ + bndcu -16(%rdi, %rcx), %bnd0 + bndcu -16(%rsi, %rcx), %bnd1 + jmp LABEL(loop_ashr_10) +LABEL(ashr_10_check): + bndcu (%rdi, %rcx), %bnd0 + bndcu (%rsi, %rcx), %bnd1 + jmp LABEL(gobble_ashr_10) +# endif .p2align 4 LABEL(loop_ashr_10): @@ -1588,7 +1746,11 @@ LABEL(nibble_ashr_10): pxor %xmm0, %xmm0 sub $0x1000, %r10 +# ifdef __CHKP__ + jmp LABEL(ashr_10_check) +# else jmp LABEL(gobble_ashr_10) +# endif .p2align 4 LABEL(ashr_10_exittail): @@ -1632,6 +1794,15 @@ LABEL(ashr_11): lea 11(%rdi), %r10 and $0xfff, %r10 /* offset into 4K page */ sub $0x1000, %r10 /* subtract 4K pagesize */ +# ifdef __CHKP__ + bndcu -16(%rdi, %rcx), %bnd0 + bndcu -16(%rsi, %rcx), %bnd1 + jmp LABEL(loop_ashr_11) +LABEL(ashr_11_check): + bndcu (%rdi, %rcx), %bnd0 + bndcu (%rsi, %rcx), %bnd1 + jmp LABEL(gobble_ashr_11) +# endif .p2align 4 LABEL(loop_ashr_11): @@ -1713,7 +1884,11 @@ LABEL(nibble_ashr_11): pxor %xmm0, %xmm0 sub $0x1000, %r10 +# ifdef __CHKP__ + jmp LABEL(ashr_11_check) +# else jmp LABEL(gobble_ashr_11) +# endif .p2align 4 LABEL(ashr_11_exittail): @@ -1757,6 +1932,15 @@ LABEL(ashr_12): lea 12(%rdi), %r10 and $0xfff, %r10 /* offset into 4K page */ sub $0x1000, %r10 /* subtract 4K pagesize */ +# ifdef __CHKP__ + bndcu -16(%rdi, %rcx), %bnd0 + bndcu -16(%rsi, %rcx), %bnd1 + jmp LABEL(loop_ashr_12) +LABEL(ashr_12_check): + bndcu (%rdi, %rcx), %bnd0 + bndcu (%rsi, %rcx), %bnd1 + jmp LABEL(gobble_ashr_12) +# endif .p2align 4 LABEL(loop_ashr_12): @@ -1838,7 +2022,11 @@ LABEL(nibble_ashr_12): pxor %xmm0, %xmm0 sub $0x1000, %r10 +# ifdef __CHKP__ + jmp LABEL(ashr_12_check) +# else jmp LABEL(gobble_ashr_12) +# endif .p2align 4 LABEL(ashr_12_exittail): @@ -1882,6 +2070,15 @@ LABEL(ashr_13): lea 13(%rdi), %r10 and $0xfff, %r10 /* offset into 4K page */ sub $0x1000, %r10 /* subtract 4K pagesize */ +# ifdef __CHKP__ + bndcu -16(%rdi, %rcx), %bnd0 + bndcu -16(%rsi, %rcx), %bnd1 + jmp LABEL(loop_ashr_13) +LABEL(ashr_13_check): + bndcu (%rdi, %rcx), %bnd0 + bndcu (%rsi, %rcx), %bnd1 + jmp LABEL(gobble_ashr_13) +# endif .p2align 4 LABEL(loop_ashr_13): @@ -1963,7 +2160,11 @@ LABEL(nibble_ashr_13): pxor %xmm0, %xmm0 sub $0x1000, %r10 +# ifdef __CHKP__ + jmp LABEL(ashr_13_check) +# else jmp LABEL(gobble_ashr_13) +# endif .p2align 4 LABEL(ashr_13_exittail): @@ -2007,6 +2208,15 @@ LABEL(ashr_14): lea 14(%rdi), %r10 and $0xfff, %r10 /* offset into 4K page */ sub $0x1000, %r10 /* subtract 4K pagesize */ +# ifdef __CHKP__ + bndcu -16(%rdi, %rcx), %bnd0 + bndcu -16(%rsi, %rcx), %bnd1 + jmp LABEL(loop_ashr_14) +LABEL(ashr_14_check): + bndcu (%rdi, %rcx), %bnd0 + bndcu (%rsi, %rcx), %bnd1 + jmp LABEL(gobble_ashr_14) +# endif .p2align 4 LABEL(loop_ashr_14): @@ -2088,7 +2298,11 @@ LABEL(nibble_ashr_14): pxor %xmm0, %xmm0 sub $0x1000, %r10 +# ifdef __CHKP__ + jmp LABEL(ashr_14_check) +# else jmp LABEL(gobble_ashr_14) +# endif .p2align 4 LABEL(ashr_14_exittail): @@ -2134,6 +2348,15 @@ LABEL(ashr_15): and $0xfff, %r10 /* offset into 4K page */ sub $0x1000, %r10 /* subtract 4K pagesize */ +# ifdef __CHKP__ + bndcu -16(%rdi, %rcx), %bnd0 + bndcu -16(%rsi, %rcx), %bnd1 + jmp LABEL(loop_ashr_15) +LABEL(ashr_15_check): + bndcu (%rdi, %rcx), %bnd0 + bndcu (%rsi, %rcx), %bnd1 + jmp LABEL(gobble_ashr_15) +# endif .p2align 4 LABEL(loop_ashr_15): @@ -2215,7 +2438,11 @@ LABEL(nibble_ashr_15): pxor %xmm0, %xmm0 sub $0x1000, %r10 +# ifdef __CHKP__ + jmp LABEL(ashr_15_check) +# else jmp LABEL(gobble_ashr_15) +# endif .p2align 4 LABEL(ashr_15_exittail): @@ -2240,6 +2467,11 @@ 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): @@ -2250,6 +2482,10 @@ 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 |