diff options
Diffstat (limited to 'sysdeps/x86_64/memcmp.S')
| -rw-r--r-- | sysdeps/x86_64/memcmp.S | 85 |
1 files changed, 85 insertions, 0 deletions
diff --git a/sysdeps/x86_64/memcmp.S b/sysdeps/x86_64/memcmp.S index d5c072c7f4..77a7bcaafe 100644 --- a/sysdeps/x86_64/memcmp.S +++ b/sysdeps/x86_64/memcmp.S @@ -23,6 +23,11 @@ ENTRY (memcmp) test %rdx, %rdx jz L(finz) +#ifdef __CHKP__ + bndcl (%rdi), %bnd0 + bndcl (%rsi), %bnd1 +#endif + pxor %xmm0, %xmm0 cmpq $1, %rdx jle L(finr1b) subq %rdi, %rsi @@ -86,6 +91,10 @@ L(s16b): .p2align 4,, 4 L(finr1b): +#ifdef __CHKP__ + bndcu (%rdi), %bnd0 + bndcu (%rsi), %bnd1 +#endif movzbl (%rdi), %eax movzbl (%rsi), %edx L(finz1): @@ -132,6 +141,10 @@ L(gt32): andq $15, %r8 jz L(16am) /* Both pointers may be misaligned. */ +#ifdef __CHKP__ + bndcu (%rdi), %bnd0 + bndcu (%rdi, %rsi), %bnd1 +#endif movdqu (%rdi), %xmm1 movdqu (%rdi, %rsi), %xmm0 pcmpeqb %xmm0, %xmm1 @@ -146,6 +159,10 @@ L(16am): jz L(ATR) testq $16, %rdi jz L(A32) +#ifdef __CHKP__ + bndcu (%rdi), %bnd0 + bndcu (%rdi, %rsi), %bnd1 +#endif movdqu (%rdi, %rsi), %xmm0 pcmpeqb (%rdi), %xmm0 pmovmskb %xmm0, %edx @@ -160,6 +177,10 @@ L(A32): /* Pre-unroll to be ready for unrolled 64B loop. */ testq $32, %rdi jz L(A64) +#ifdef __CHKP__ + bndcu (%rdi), %bnd0 + bndcu (%rdi, %rsi), %bnd1 +#endif movdqu (%rdi,%rsi), %xmm0 pcmpeqb (%rdi), %xmm0 pmovmskb %xmm0, %edx @@ -167,6 +188,10 @@ L(A32): jnz L(neq) addq $16, %rdi +#ifdef __CHKP__ + bndcu (%rdi), %bnd0 + bndcu (%rdi, %rsi), %bnd1 +#endif movdqu (%rdi,%rsi), %xmm0 pcmpeqb (%rdi), %xmm0 pmovmskb %xmm0, %edx @@ -181,6 +206,10 @@ L(A64): jge L(mt32) L(A64main): +#ifdef __CHKP__ + bndcu (%rdi), %bnd0 + bndcu (%rdi, %rsi), %bnd1 +#endif movdqu (%rdi,%rsi), %xmm0 pcmpeqb (%rdi), %xmm0 pmovmskb %xmm0, %edx @@ -188,6 +217,10 @@ L(A64main): jnz L(neq) addq $16, %rdi +#ifdef __CHKP__ + bndcu (%rdi), %bnd0 + bndcu (%rdi, %rsi), %bnd1 +#endif movdqu (%rdi,%rsi), %xmm0 pcmpeqb (%rdi), %xmm0 pmovmskb %xmm0, %edx @@ -195,6 +228,10 @@ L(A64main): jnz L(neq) addq $16, %rdi +#ifdef __CHKP__ + bndcu (%rdi), %bnd0 + bndcu (%rdi, %rsi), %bnd1 +#endif movdqu (%rdi,%rsi), %xmm0 pcmpeqb (%rdi), %xmm0 pmovmskb %xmm0, %edx @@ -202,6 +239,10 @@ L(A64main): jnz L(neq) addq $16, %rdi +#ifdef __CHKP__ + bndcu (%rdi), %bnd0 + bndcu (%rdi, %rsi), %bnd1 +#endif movdqu (%rdi,%rsi), %xmm0 pcmpeqb (%rdi), %xmm0 pmovmskb %xmm0, %edx @@ -219,6 +260,10 @@ L(mt32): jge L(mt16) L(A32main): +#ifdef __CHKP__ + bndcu (%rdi), %bnd0 + bndcu (%rdi, %rsi), %bnd1 +#endif movdqu (%rdi,%rsi), %xmm0 pcmpeqb (%rdi), %xmm0 pmovmskb %xmm0, %edx @@ -226,6 +271,10 @@ L(A32main): jnz L(neq) addq $16, %rdi +#ifdef __CHKP__ + bndcu (%rdi), %bnd0 + bndcu (%rdi, %rsi), %bnd1 +#endif movdqu (%rdi,%rsi), %xmm0 pcmpeqb (%rdi), %xmm0 pmovmskb %xmm0, %edx @@ -258,6 +307,10 @@ L(ATR): testq $16, %rdi jz L(ATR32) +#ifdef __CHKP__ + bndcu (%rdi), %bnd0 + bndcu (%rdi, %rsi), %bnd1 +#endif movdqa (%rdi,%rsi), %xmm0 pcmpeqb (%rdi), %xmm0 pmovmskb %xmm0, %edx @@ -273,6 +326,10 @@ L(ATR32): testq $32, %rdi jz L(ATR64) +#ifdef __CHKP__ + bndcu (%rdi), %bnd0 + bndcu (%rdi, %rsi), %bnd1 +#endif movdqa (%rdi,%rsi), %xmm0 pcmpeqb (%rdi), %xmm0 pmovmskb %xmm0, %edx @@ -280,6 +337,10 @@ L(ATR32): jnz L(neq) addq $16, %rdi +#ifdef __CHKP__ + bndcu (%rdi), %bnd0 + bndcu (%rdi, %rsi), %bnd1 +#endif movdqa (%rdi,%rsi), %xmm0 pcmpeqb (%rdi), %xmm0 pmovmskb %xmm0, %edx @@ -292,6 +353,10 @@ L(ATR64): je L(mt32) L(ATR64main): +#ifdef __CHKP__ + bndcu (%rdi), %bnd0 + bndcu (%rdi, %rsi), %bnd1 +#endif movdqa (%rdi,%rsi), %xmm0 pcmpeqb (%rdi), %xmm0 pmovmskb %xmm0, %edx @@ -299,6 +364,10 @@ L(ATR64main): jnz L(neq) addq $16, %rdi +#ifdef __CHKP__ + bndcu (%rdi), %bnd0 + bndcu (%rdi, %rsi), %bnd1 +#endif movdqa (%rdi,%rsi), %xmm0 pcmpeqb (%rdi), %xmm0 pmovmskb %xmm0, %edx @@ -306,6 +375,10 @@ L(ATR64main): jnz L(neq) addq $16, %rdi +#ifdef __CHKP__ + bndcu (%rdi), %bnd0 + bndcu (%rdi, %rsi), %bnd1 +#endif movdqa (%rdi,%rsi), %xmm0 pcmpeqb (%rdi), %xmm0 pmovmskb %xmm0, %edx @@ -313,6 +386,10 @@ L(ATR64main): jnz L(neq) addq $16, %rdi +#ifdef __CHKP__ + bndcu (%rdi), %bnd0 + bndcu (%rdi, %rsi), %bnd1 +#endif movdqa (%rdi,%rsi), %xmm0 pcmpeqb (%rdi), %xmm0 pmovmskb %xmm0, %edx @@ -328,6 +405,10 @@ L(ATR64main): jge L(mt16) L(ATR32res): +#ifdef __CHKP__ + bndcu (%rdi), %bnd0 + bndcu (%rdi, %rsi), %bnd1 +#endif movdqa (%rdi,%rsi), %xmm0 pcmpeqb (%rdi), %xmm0 pmovmskb %xmm0, %edx @@ -335,6 +416,10 @@ L(ATR32res): jnz L(neq) addq $16, %rdi +#ifdef __CHKP__ + bndcu (%rdi), %bnd0 + bndcu (%rdi, %rsi), %bnd1 +#endif movdqa (%rdi,%rsi), %xmm0 pcmpeqb (%rdi), %xmm0 pmovmskb %xmm0, %edx |