https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116458
--- Comment #6 from Alexander Monakov <amonakov at gcc dot gnu.org> --- As for Valgrind false positive, it handles this SSSE3 code really well and misses the key point by a very narrow margin. We have found = m1 + (m2 << 16); where both m1 and m2 hold 16-bit masks from pmovmskb, so 'found' is just the 32-bit concatenation of those masks. Suppose m1 has uninit bits. In that case, its highest known bit is guaranteed to be non-zero (because our character stream always ends with a newline), in 'm2 << 16' the lower 16 bits are known, and therefore 'found' has a known non-zero bit, making the branch on found != 0 well-defined. If we instead combine the halves with an OR instead of PLUS: found = m1 | (m2 << 16); Valgrind properly computes known bits in 'found' and does not complain. I can see the difficulty in computing known bits after addition in the general case though.
