[PATCH] D142375: [AMDGPU][CodeGen] Combine SDAG and global isel tests for intersect ray intrinsics.

Mon Feb 27 13:24:41 PST 2023

arsenm added a comment.

In D142375#4132791 <https://reviews.llvm.org/D142375#4132791>, @kosarev wrote:

>> No no, all the permutations should be tested. This is not a cosmetic choice, we should be testing these are handled correctly
>
> That's what the comments say, yes. But then if vector versions have only been used to make sure the register operands are where the values actually land, maybe we can test that just by adding some extra arguments > to flattened versions and thus make accidentally correct output sufficiently unlikely?

I don't really follow what this is saying. One reason for the split would have been that SelectionDAG is broken in the trickier cases that require waterfall loops

> Tests like this don't feel to be the right place to test codegen'ing of moves preparing the operands and cluttering tests checking them repeatedly for various kinds of calls probably doesn't make much sense as well?

It's making sure they codegen for any inputs, which would also cover the waterfall cases?

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D142375/new/

https://reviews.llvm.org/D142375