I gave up on trying to make gcc emit branchless code on amd64 and switched to inline asm.

[vitaut@mastodon.social]

Clang has no trouble whatsoever, consistently delivering 10% better perf.

[vitaut@mastodon.social]

Actually I did manage to get branchless version of gcc with masks and bitwise arithmetic but it was shit.

[vitaut@mastodon.social]

Aren't compilers supposed to be good at optimizing code?

[sabena@mastodon.social]

@vitaut I have found clang will emit branchless code come hell or high water. Even in situations where the branchless code is horrible.

[vitaut@mastodon.social]

@sabena Ideally it should be controllable

[vitaut@mastodon.social]

@sabena In case it wasn’t clear I wasn’t generalizing but speaking about my specific case.

[sabena@mastodon.social]

@vitaut yeah, it should. Does GCC get it with pgo data?

[vitaut@mastodon.social]

@sabena Maybe but it seems like even more pain that a few lines of inline asm.

[sabena@mastodon.social]

@vitaut yeah, true. You should be able to annotate with both [[likely]] and [[unlikely]] at the same time tbh

[shac@ioc.exchange]

@vitaut Is the clang code better on average or is it better for your benchmark? A branch is not necessarily bad, especially if it’s short and predictable. Your benchmark may be stressing the core in a way that makes the branch-based version underperform, while it might do better in a one-shot situation.