← Back to context

Comment by cogman10

16 days ago

What's strange is I'm finding that gcc really struggles to correctly optimize this.

This was my function

    for (auto v : array) {
        if (v != 0)
            return false;
    }
    return true;

clang emits basically the same thing yours does. But gcc ends up just really struggling to vectorize for large numbers of array.

Here's gcc for 42 elements:

https://godbolt.org/z/sjz7xd8Gs

and here's clang for 42 elements:

https://godbolt.org/z/frvbhrnEK

Very bizarre. Clang pretty readily sees that it can use SIMD instructions and really optimizes this while GCC really struggles to want to use it. I've even seen strange output where GCC will emit SIMD instructions for the first loop and then falls back on regular x86 compares for the rest.

Edit: Actually, it looks like for large enough array sizes, it flips. At 256 elements, gcc ends up emitting simd instructions while clang does pure x86. So strange.

4 comments

cogman10

Reply
maccard  15 days ago

Writing a micro benchmark is an academic exercise. You end up benchmarking in isolation which only tells you is your function faster in that exact scenario. Something which is faster in isolation in a microbenchmark can be slower when put in a real workload because vextoising is likely to have way more of an impact than anything else. Similarly, if you parallelise it, you introduce a whole new category of ways to compare.

  • cogman10  15 days ago

    This isn't a microbenchmark. In fact, I haven't even bothered to benchmark it (perhaps the non-simd version actually is faster?)

    This is purely me looking at the emitted assembly and being surprised at when the compilers decide to deploy it and not deploy it. It may be the case that the SIMD instructions are in fact slower even though they should theoretically end up faster.

    Both compilers are simply using heuristics to determine when it's fruitful to deploy SIMD instructions.