https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110087
Alexander Monakov <amonakov at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |amonakov at gcc dot gnu.org
--- Comment #3 from Alexander Monakov <amonakov at gcc dot gnu.org> ---
(In reply to Uroš Bizjak from comment #1)
> BTW: If the result of foo is random, then cmove gets badly predicted.
> Considering the problems with cmove on x86 (even without bad prediction),
> the above optimization can be quite important. Clang does it.
There is no prediction involved in execution of CMOV. It is one ALU uop with
latency 1 on any recent x86, or two uops with latency 1 on Haswell, going back
to Intel Core: https://uops.info/html-instr/CMOVZ_R32_R32.html
If you convert a control dependency to a data dependency with CMOV you may end
up with slower code due to longer dependency chains, but this is not the case
here.
