https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116896
--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Note, we have a pattern matcher for floating point spaceship (that has without
-ffast-math 4 cases), but not for integer. And the code we emit for lhs <=>
rhs matches
what we emit for
if (lhs == rhs) {
return std::strong_ordering::equal;
} else if (lhs < rhs) {
return std::strong_ordering::less;
} else {
return std::strong_ordering::greater;
}
Seems clang 19 also emitted different sequences for the two different manual
cases but used cmov nevertheless instead of the conditional jump (as I said,
that can be a win or can be bad pessimization, depending on the exact compared
values).
That said, clang trunk seems to emit
movl (%rdi), %eax
cmpl (%rsi), %eax
setl %al
setg %cl
subb %al, %cl
movsbl %cl, %eax
for both, so something without jumps and without conditional moves. I guess we
should check out how does that work for latency and throughput on different
CPUs.
