[PATCH] D107344: [SLP]Do not emit extra shuffle for insertelements vectorization.

Alexey Bataev via Phabricator via llvm-commits llvm-commits at lists.llvm.org
Wed Aug 4 09:51:38 PDT 2021


ABataev added a comment.

In D107344#2925936 <https://reviews.llvm.org/D107344#2925936>, @srj wrote:

> In D107344#2925918 <https://reviews.llvm.org/D107344#2925918>, @ABataev wrote:
>
>> In D107344#2925891 <https://reviews.llvm.org/D107344#2925891>, @srj wrote:
>>
>>> This appears to have injected a failure into Halide codegen; I'll be providing a repro case shortly.
>>
>> Thanks for the report, waiting for the repro.
>
> Still trying to get something self-contained for  you, but the difference in codegen is something like:
>
> before:
>
>   %193 = shufflevector <2 x double> %179, <2 x double> poison, <4 x i32> <i32 0, i32 1, i32 undef, i32 undef>
>   ...
>   %200 = insertelement <4 x double> %193, double 1.000000e+02, i32 3
>   %201 = insertelement <4 x double> %200, double %184, i32 2
>
> after:
>
>   %193 = shufflevector <2 x double> %179, <2 x double> poison, <4 x i32> <i32 0, i32 1, i32 undef, i32 undef>
>   ...
>   %200 = insertelement <4 x double> %193, double %184, i32 2

Hmm, very strange, the patch removes extra shuffle instruction, it has nothing to do with the insertelements. Waiting for the reproducer for investigation.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D107344/new/

https://reviews.llvm.org/D107344



More information about the llvm-commits mailing list