[llvm] [LoopVectorizer] Add support for partial reductions (PR #92418)

[Graham Hunter via llvm-commits]

huntergr-arm wrote:

> If I'm understanding correctly, a "partial reduction" is just a slightly different way of generating code for a reduction? Basically, instead of performing the reduction using a number of lanes equal to the vector factor, you combine some of the lanes each iteration. Usually, this wouldn't really be profitable unless you have a register pressure problem. But in very specific cases, you can use specialized instructions that do horizontal sums, in which case it's extremely profitable. (This is why the testcase is called "partial-reduced-sdot.ll", I assume.)
> 
> It seems a bit weird to me to introduce a new intrinsic that, in the general case, isn't actually a natively supported operation on any target.

Hi,

Yes, it's effectively a way of representing a reduction that allows us to vectorize with a wider VF than we normally would, since the IR extends the elements loaded from memory. For the AArch64 instructions we're targeting (sdot, udot, etc.) the extension is part of the instruction; e.g. sdot of two <vscale x 16 x i8> inputs results in a <vscale x 4 x i32> output. While this may be interesting for some actual dot products in SLP vectorization, for this patch we're just interesting in increasing our VF where possible.

I posted PRs last year for a different approach which only widened the VF in LoopVec and pattern-matched to aarch64-specific dot product instructions in a target-specific pass. There was no real interest in those PRs and I was asked to consider a different approach. Nick has now implemented the suggested approach.

(obsolete LoopVec pass to widen VF: https://github.com/llvm/llvm-project/pull/69587)
(obsolete AArch64-specific target pass to pattern match the resulting IR: https://github.com/llvm/llvm-project/pull/69583)

@paulwalker-arm 's RFC for the alternative: https://discourse.llvm.org/t/rfc-is-a-more-expressive-way-to-represent-reductions-useful/74929

https://github.com/llvm/llvm-project/pull/92418