disable AVX2/AVX3 gather

[Andarwinux]

Is it possible to add an option to disable the use of gather instructions for AVX2/AVX3 targets? I've noticed that even building libjxl with -mno-gather still comes with some gather instructions, so it looks like highway is using explicit gather intrinsics somewhere.

Since newer microcode enforces GDS mitigation, this can cause performance issues on skylake~tigerlake.

[jan-wassenberg]

Yes, JPEG XL includes a GatherIndex. I believe it's in the encode side, are we sure this is time-critical, and that the replacement (scalar fallback?) would be better? The last time I looked at this, the CPU mitigation cost about 1.3x perf, and gathers embedded within vector code are still better than scalarizing.

[Andarwinux]

Well, since I have no control over whether gather is enabled or not, I can't evaluate its impact on libjxl performance. But according to the sample in llvm/llvm-project#70259, on my 11800H, gather causes a 1.58x slowdown. (1.65s > 2.6s, for reference, 1.2s with gather enabled without updated microcode)

[jan-wassenberg]

Thanks for the data point. It is an extreme case because usually we have considerable other work going on around the gather which allows hiding its latency.
But fair enough, we can add a flag - will get to that early next week.

[Andarwinux]

Thanks! And if possible, this flag/macro should only affect AVX2/AVX3/AVX3_DL targets, since SPR/ZEN4/AVX10_2 are unaffected by GDS.