[PATCH] D37065: Ensure standard pseudo instructions (TargetOpcode::*) are compatible with guessInstructionProperties=0

[Alex Bradbury via Phabricator via llvm-commits]

asb added a comment.

In https://reviews.llvm.org/D37065#904038, @qcolombet wrote:

> Hi Alex,
>
> I see that we stuck to hasSideEffects = 1 for BUNDLE, EH_LABEL and so on (the ones I marked with Ditto at some point).
>  Are we going to do the same thing as PHIs here, or do we plan to leave them with that flag ON?
>
> I am not sure it is the right thing to do, but assuming it is, could you add a comment before each of them to explain why they have side effects.


My intention is to commit this as a (hopefully) no-op change, then land further patches which change the current behaviour to the desired behaviour (see https://reviews.llvm.org/D37097 and https://reviews.llvm.org/D37230). The only problem with the BUNDLE update is that I find it very time consuming to craft and verify updates to the AMDGPU tests, so can't guarantee I'll have the time to finish it off. I'm committed to landing patches to the other instructions though.

I'll update this patch to add FIXMEs/TODOs for the instructions we believe should have hasSideEffects = 0.


https://reviews.llvm.org/D37065