By: anon1 (anon.delete@this.anon.com), May 16, 2022 4:44 am
Room: Moderated Discussions
Michael S (already5chosen.delete@this.yahoo.com) on May 14, 2022 12:27 pm wrote:
> After looking at code, generated by LLVM autovectorizer last year, I am more that a somewhat
> doubtful. To say that a year ago they were bad would be an undeserving compliment.
> "Horrendously stupid" describes a situation more adequately.
Are you talking about SVE or about legacy SIMD ISAs? I can imagine that autovectorizing for SVE2 is probably simpler, given that they give you loops and direct scatter/gather. One doesn't need to think of clever tricks how to abuse the SIMD instructions, at least not as much.
> After looking at code, generated by LLVM autovectorizer last year, I am more that a somewhat
> doubtful. To say that a year ago they were bad would be an undeserving compliment.
> "Horrendously stupid" describes a situation more adequately.
Are you talking about SVE or about legacy SIMD ISAs? I can imagine that autovectorizing for SVE2 is probably simpler, given that they give you loops and direct scatter/gather. One doesn't need to think of clever tricks how to abuse the SIMD instructions, at least not as much.