Recent LLVM trunk builds seem to have broken predicated_store_load

[steven-johnson]

Haven't investigated yet, just opening an issue for tracking.

[abadams]

I noticed massive performance issues with masked vector store instructions on x86. It's not supposed to be non-temporal anymore, but subsequent loads of an address that you do a masked store to take forever, so it seems like it has left cache. I just turned them off entirely in the autoscheduler branch and was going to propose turning them off in master for x86 for now.

[steven-johnson]

propose turning them off in master for x86 for now

+1, make it so

[steven-johnson]

I just turned them off entirely in the autoscheduler branch

Did you do that locally? I don't see it in the branch.

[abadams]

https://github.com/halide/Halide/blob/standalone_autoscheduler/src/VectorizeLoops.cpp#L254

[steven-johnson]

With that in place, the branch is still failing; e.g. #3535 which is up to date with standalone_autoscheduler still fails predicated_load_store

[abadams]

It fails it because it's not generating any predicated stores, and the test asserts that it does.

[steven-johnson]

d'oh

[steven-johnson]

Looks like we never followed up on this -- probably worth revisiting to see if the 'recent LLVM trunk breakage' is still a thing.

[steven-johnson]

I'm going to assume this is long-since-stale