https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92578
--- Comment #2 from Uroš Bizjak <ubizjak at gmail dot com> --- (In reply to Richard Biener from comment #1) > With newcnt-=2 you get > > movl %edx, %r8d > movl %esi, %eax > leal -2(%rsi), %edx > cmpl %r8d, %edi > cmove %edx, %eax > ret > > so it's if-conversions preference to generate setCC which I'm not sure is > really worse here, is it? We're saving a zero extend and some reg-reg moves. This if-conversion is intentional, since cmove needs well predicted comparison, which is not always the case. > I'm sure we have a duplicate report about this. PR 79389 shows devastating SciMark2.0 results with cmove for random values of compare operands. Related to PR 56309.
