[PATCH] D76230: [AMDGPU] Enable SEXT divergence driven selection.

Tue Mar 17 08:31:30 PDT 2020

arsenm added a comment.

In D76230#1924541 <https://reviews.llvm.org/D76230#1924541>, @alex-t wrote:

> In D76230#1924449 <https://reviews.llvm.org/D76230#1924449>, @arsenm wrote:
>
> > Test name is misleading, this should just go in the realgar sext test. Also I would expect this to already be well covered
>
>
> This is not covered properly.  Most of the tests except those I have already updated assume that we select everything to scalars and rely on the latest SIFixSGPRCopies moveToVALU. 
>  Also, their names are misleading as well:
>
>   sext-eliminate.ll - eliminate what?

This is an optimization test

>   sext-in-reg - it contains a couple of tests that are not scalar but they're testing some special conditions.

sext_inreg is a different opcode. You're looking for sign_extend.ll

> My test checks specifically the functionality added by the change - proper selection of i16 to i32 uniform/divergent and i16 to i64 uniform/divergent.
>  As for the name: does **sext-divergence-driven-selection.ll ** looks better?
>  Also, I would prefer to have the tests checking divergence driven ISel separately until we remove moveToVALU path completely. Otherwise I always should expect someone accidentally removing selection predicates that generates no errors but silently go through the moveToVALU.

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D76230/new/

https://reviews.llvm.org/D76230