[Bug target/95784] Failure to optimize usage of __builtin_add_overflow with return statement properly

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95784

--- Comment #9 from Richard Biener <rguenth at gcc dot gnu.org> ---
I wouldn't be surprised if a version with a branch is faster even with each
of the branches mispredicted.  cmovs are weird beasts but since they
are not dependent on each other their latency at least shouldn't add up here
so LLVMs two cmovs shouldnt be worse off than GCCs one cmov.  You'd need to
compare against a variant without any cmov.