[PATCH] D109594: [AMDGPU] Initialize LDS pointers after alloca, but before call.

Tue Sep 14 04:24:27 PDT 2021

hsmhsm added a comment.

In D109594#2999421 <https://reviews.llvm.org/D109594#2999421>, @JonChesterfield wrote:

> Modifying exec from IR is right out. I was referring back to the internal discussion on how to implement this where that was briefly considered.
>
> Given we now know more than when the original decision was taken, let us revisit that decision and not split the block. MIR is a point where this and all other uniform stores could be optimised for power consumption by masking exec, without changing IR optimisations.

If you think so,  then let's have an internal discussion by involving all the key folks, and take a decision, I am fine with whatever decision is being taken.

> Moving all alloca to entry was proposed on the mailing list some years ago and implemented in at least one out of tree target. I suspect it is not already done in tree because codegen is more efficient on most CPU targets without it. I doubt it is especially complicated to implement - the entry block dominates the other blocks in the function.

This is just only your suspect as I have my own suspect. We will not be having clear answer to it, until we ask community itself. Let me try that too. We do not know how the out of tree target has implemented it, it could also be just hacky stuff with some serious constraints, but it just work for them.

> I'm not convinced the transform being enabled here is necessary, and have previously outlined a variety of alternatives which you remain unwilling to consider. My interest here is solely in avoiding your "temporary hacks" breaking openmp.

I am not sure what are the  *variety* of alternatives discussed here which are quickly and neatly implementable. Can you again outline them here?

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D109594/new/

https://reviews.llvm.org/D109594