[ANN] LoopVectorization

[yuyichao]

@Elrod

Three examples:

I am surprised that llvm left such optimization on the table, though for most (all?) of it, it again seems that LLVM does know what’s the offsets are but didn’t manage to emit the optimal code to generate it. It can certainly be added to the cases were better codgen is needed but I don’t think this info has ever be hidden from LLVM IR.

@Elrod

If you like to waste time with micro-optimizations as much as I do, AVX512 is fun because of the extra options it provides and large possible reward when you properly leverage it (that large reward being just as much a function of the compiler’s poor AVX512 performance as you note as it is the larger vectors; the opportunity cost of hand optimizing / autogenerating optimized code is what the compiler would have done).

I like to do it when it matters for me ;-p… So yes when I use AVX512 capable hardware as my main computer.

@Elrod

This has now been fixed on LLVM master, after I asked for help getting rid of those vmovapd s online. That means on Julia versions with an LLVM using that commit, I can quit using asm call . I think that’ll be starting with LLVM 11?

As far as my concern, you are very welcome to submit that patch to be included in our LLVM build. I basically did that for all of the few LLVM performance related issues that I’ve fixed/experienced, and ususally before they are merged upstream. There’s no need to wait for a new LLVM release to be used.

---

And I’ll also add that doing these pattern matching with LLVM IR is not hard. In fact I always found it to be easier than working with julia IR. The switch to SSA for julia IR have cerntainly helpped (though that’s not the level this package uses) but LLVM still have WAY more utilities to do all of these analysis.