Hacker News Reader

Comment by mgaunard | Hacker News Reader

I find it sad that you need to use inline asm for this.

Why isn't GCC providing a built-in to reliably use cmov? Every time I use ?: it's a lottery whether I get it or not. I've had to resort to using the slower bitwise logic to get branch-free code reliably.

View on HN

I think I’ve observed reliably-emitted CMOVs when using equivalent (but contrived) arithmetic branchless idioms, e.g. instead of (a ? b : c) write (-!!a & b | -!a & c) (yes, this is rather repulsive, but less so if you think about the negation as turning C-style booleans into Forth/BASIC/mask-style booleans).

It should just be a builtin.

that is what I meant by "the slower bitwise logic".

I didn't know GCC was able to optimize that down to cmov.

It’s not, it turns out, I misremembered :( Clang 13 does compile this down to a CMOV in my tests, but GCC 11.1 just translates it as is even with -mtune=native (on Haswell).

It would be great to have builtin for CMOV. But I also want GCC to utilize the hack with the SBB instruction when I write something like:

(unsigned)a < (unsigned)b ? ~0 : 0

Instead of three instructions what it generates.

It does use cmp + sbb? https://godbolt.org/z/hnvqrcTcW

Cool, but it doesn't generate SBB in real code, but instead gives me something like this:

    xorl %r11d, %r11d
    cmpl %r12d, %eax
    seta %r11b

And if I do both of these, side by side:

    tmp = rc.code < rc_bound ? ~0 : 0;
    if (rc.code < rc_bound) rc.range = rc_bound;

Then GCC generates an if-else branch.

Thus, inline assembly is the only option so far to prevent the compiler from doing something dumb.

Reply on news.ycombinator.com