http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47311
--- Comment #17 from Richard Henderson <rth at gcc dot gnu.org> 2011-01-17 18:40:05 UTC --- (In reply to comment #16) > ==5267== Invalid read of size 8 > ==5267== at 0x11E4E24: search_line_sse42(unsigned char const*, unsigned It's fine. (1) Note that we've already verified above that the address is aligned and (2) that we will in fact find the end-of-line marker that we expect before we examine data beyond END. I suppose valgrind expects us to update EDX in > 454│ /* Main loop, processing 16 bytes at a time. By doing the whole loop > 455│ in inline assembly, we can make proper use of the flags set. */ > 456│ __asm ( "sub $16, %1\n" > 457│ " .balign 16\n" > 458│ "0: add $16, %1\n" > 459│ " %vpcmpestri $0, (%1), %2\n" > 460│ " jnc 0b" > 461│ : "=&c"(index), "+r"(s) > 462├> : "x"(search), "a"(4), "d"(16)); so that it can tell that we're examining only the first EDX bytes of the value read, but that would require a bit of useless arithmetic here. If there's some easy way we can add a valgrind exception for these functions to the gcc source tree, I'd be open to do so. But I also don't want to figure out how to do that myself.
