[PATCH] D135078: [AArch64][GlobalISel] Fold away lowered vector sign-extend of vector compares

Mon Oct 3 15:02:39 PDT 2022

aemerson added a comment.

In D135078#3831868 <https://reviews.llvm.org/D135078#3831868>, @arsenm wrote:

> In D135078#3831867 <https://reviews.llvm.org/D135078#3831867>, @arsenm wrote:
>
>> In D135078#3830895 <https://reviews.llvm.org/D135078#3830895>, @aemerson wrote:
>>
>>> In D135078#3830892 <https://reviews.llvm.org/D135078#3830892>, @arsenm wrote:
>>>
>>>> This should have been caught and combined into g_sext_inreg, which should have folded out based on computeNumSignBits
>>>
>>> The only place we have an opportunity to catch the extend before it gets legalized into shifts is in the artifact combiner, and this is just a code quality optimization so it shouldn't go there.
>>
>> The shifts themselves combine into g_sext_inreg
>
> Plus there's nothing AArch64 specific about this. You're not getting the sext_inreg with the base set of combines because the aarch64 rule is to always lower sext_inreg

Ah I see. I’ll look into that tomorrow.

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D135078/new/

https://reviews.llvm.org/D135078