diff options
Diffstat (limited to 'sysdeps/i386/i686/multiarch/strlen-sse2.S')
| -rw-r--r-- | sysdeps/i386/i686/multiarch/strlen-sse2.S | 67 |
1 files changed, 66 insertions, 1 deletions
diff --git a/sysdeps/i386/i686/multiarch/strlen-sse2.S b/sysdeps/i386/i686/multiarch/strlen-sse2.S index a4f2806cfe..3d0743ec88 100644 --- a/sysdeps/i386/i686/multiarch/strlen-sse2.S +++ b/sysdeps/i386/i686/multiarch/strlen-sse2.S @@ -41,7 +41,10 @@ # define PUSH(REG) pushl REG; CFI_PUSH (REG) # define POP(REG) popl REG; CFI_POP (REG) # undef RETURN -# define RETURN POP (%edi); CFI_PUSH(%edi); ret +# define RETURN \ + mov STR+4(%esp),%edx; \ + bndcu -1(%edx,%eax), %bnd0; \ + POP (%edi); CFI_PUSH(%edi); ret # endif # ifndef STRLEN @@ -51,12 +54,19 @@ atom_text_section ENTRY (STRLEN) mov STR(%esp), %edx +# ifdef __CHKP__ + bndldx STR(%esp,%edx,1), %bnd0 +# endif # ifdef USE_AS_STRNLEN PUSH (%edi) movl LEN(%esp), %edi sub $4, %edi jbe L(len_less4_prolog) # endif +# ifdef __CHKP__ + bndcl (%edx),%bnd0 + bndcu (%edx),%bnd0 +# endif # endif xor %eax, %eax cmpb $0, (%edx) @@ -122,6 +132,9 @@ ENTRY (STRLEN) jbe L(len_less64) # endif +#ifdef __CHKP__ + bndcu (%eax), %bnd0 +#endif pcmpeqb (%eax), %xmm0 pmovmskb %xmm0, %edx pxor %xmm1, %xmm1 @@ -129,6 +142,9 @@ ENTRY (STRLEN) lea 16(%eax), %eax jnz L(exit) +#ifdef __CHKP__ + bndcu (%eax), %bnd0 +#endif pcmpeqb (%eax), %xmm1 pmovmskb %xmm1, %edx pxor %xmm2, %xmm2 @@ -136,6 +152,9 @@ ENTRY (STRLEN) lea 16(%eax), %eax jnz L(exit) +#ifdef __CHKP__ + bndcu (%eax), %bnd0 +#endif pcmpeqb (%eax), %xmm2 pmovmskb %xmm2, %edx pxor %xmm3, %xmm3 @@ -143,6 +162,9 @@ ENTRY (STRLEN) lea 16(%eax), %eax jnz L(exit) +#ifdef __CHKP__ + bndcu (%eax), %bnd0 +#endif pcmpeqb (%eax), %xmm3 pmovmskb %xmm3, %edx test %edx, %edx @@ -154,24 +176,36 @@ ENTRY (STRLEN) jbe L(len_less64) # endif +#ifdef __CHKP__ + bndcu (%eax), %bnd0 +#endif pcmpeqb (%eax), %xmm0 pmovmskb %xmm0, %edx test %edx, %edx lea 16(%eax), %eax jnz L(exit) +#ifdef __CHKP__ + bndcu (%eax), %bnd0 +#endif pcmpeqb (%eax), %xmm1 pmovmskb %xmm1, %edx test %edx, %edx lea 16(%eax), %eax jnz L(exit) +#ifdef __CHKP__ + bndcu (%eax), %bnd0 +#endif pcmpeqb (%eax), %xmm2 pmovmskb %xmm2, %edx test %edx, %edx lea 16(%eax), %eax jnz L(exit) +#ifdef __CHKP__ + bndcu (%eax), %bnd0 +#endif pcmpeqb (%eax), %xmm3 pmovmskb %xmm3, %edx test %edx, %edx @@ -183,24 +217,36 @@ ENTRY (STRLEN) jbe L(len_less64) # endif +#ifdef __CHKP__ + bndcu (%eax), %bnd0 +#endif pcmpeqb (%eax), %xmm0 pmovmskb %xmm0, %edx test %edx, %edx lea 16(%eax), %eax jnz L(exit) +#ifdef __CHKP__ + bndcu (%eax), %bnd0 +#endif pcmpeqb (%eax), %xmm1 pmovmskb %xmm1, %edx test %edx, %edx lea 16(%eax), %eax jnz L(exit) +#ifdef __CHKP__ + bndcu (%eax), %bnd0 +#endif pcmpeqb (%eax), %xmm2 pmovmskb %xmm2, %edx test %edx, %edx lea 16(%eax), %eax jnz L(exit) +#ifdef __CHKP__ + bndcu (%eax), %bnd0 +#endif pcmpeqb (%eax), %xmm3 pmovmskb %xmm3, %edx test %edx, %edx @@ -212,24 +258,36 @@ ENTRY (STRLEN) jbe L(len_less64) # endif +#ifdef __CHKP__ + bndcu (%eax), %bnd0 +#endif pcmpeqb (%eax), %xmm0 pmovmskb %xmm0, %edx test %edx, %edx lea 16(%eax), %eax jnz L(exit) +#ifdef __CHKP__ + bndcu (%eax), %bnd0 +#endif pcmpeqb (%eax), %xmm1 pmovmskb %xmm1, %edx test %edx, %edx lea 16(%eax), %eax jnz L(exit) +#ifdef __CHKP__ + bndcu (%eax), %bnd0 +#endif pcmpeqb (%eax), %xmm2 pmovmskb %xmm2, %edx test %edx, %edx lea 16(%eax), %eax jnz L(exit) +#ifdef __CHKP__ + bndcu (%eax), %bnd0 +#endif pcmpeqb (%eax), %xmm3 pmovmskb %xmm3, %edx test %edx, %edx @@ -250,6 +308,9 @@ L(aligned_64_loop): sub $64, %edi jbe L(len_less64) # endif +#ifdef __CHKP__ + bndcu (%eax), %bnd0 +#endif movaps (%eax), %xmm0 movaps 16(%eax), %xmm1 movaps 32(%eax), %xmm2 @@ -535,6 +596,10 @@ L(len_less4_prolog): add $4, %edi jz L(exit_tail0) +# ifdef __CHKP__ + bndcl (%edx),%bnd0 + bndcu (%edx),%bnd0 +# endif cmpb $0, (%edx) jz L(exit_tail0) cmp $1, %edi |