[PATCH] D148855: [SLP]Improve tryToGatherExtractElements by using per-register analysis.

[Alexey Bataev via Phabricator via llvm-commits]

ABataev added a comment.

In D148855#4655968 <https://reviews.llvm.org/D148855#4655968>, @mstorsjo wrote:

> This seems to have caused a misoptimization in ffmpeg for aarch64.
>
> To reproduce, you can follow these steps, on aarch64 Linux:
>
>   $ git clone https://github.com/ffmpeg/ffmpeg
>   $ mkdir ffmpeg-build
>   $ cd ffmpeg-build
>   $ ../ffmpeg/configure --cc=clang --samples=$(pwd)/../fate-samples
>   $ make fate-rsync
>   $ make -j$(nproc) fate-vp9-00-quantizer-18
>
> The misoptimized object file is `libavcodec/vp9dsp_8bpp.o`.
>
> The standalone preprocessed input for that object file is available at https://martin.st/temp/vp9dsp_8bpp-preproc.c, you can reproduce the misoptimization with `clang -target aarch64-linux-gnu -c -O3 vp9dsp_8bpp-preproc.c -o vp9dsp_8bpp.o`.
>
> Can you look into this, and possibly revert if fixing takes some time?

Hi, thanks for the report. Generally speaking, this change does he same, what InstCombiner does with extractelement/insertelement sequences. I'll check what's the cause. Do not know the reason yet, but most probably either TTI cost model problem, or codegen (lowering) problem.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D148855/new/

https://reviews.llvm.org/D148855