[PATCH] D82737: [AMDGPU] Insert PS early exit at end of control flow

Mon Jun 29 22:00:30 PDT 2020

critson marked 2 inline comments as done.
critson added a comment.

In D82737#2119798 <https://reviews.llvm.org/D82737#2119798>, @nhaehnle wrote:

> Can you explain the reasoning behind how you decide where to insert SI_KILL_CLEANUP?

I will add this as a comment in the code (because apparently I forgot):
SI_KILL_CLEANUP is inserted immediately after SI_END_CF for control flow where a kill is present.
When SI_KILL_CLEANUP is uniformly reached it will become a test of the EXEC mask immediately after it has been updated by the SI_END_CF to reflect the true state of live lanes.

We could actually blindly insert SI_KILL_CLEANUP after every SI_END_CF, reducing the code complexity, but we'd end up with a lot of early exits that would never be taken.

================
Comment at: llvm/lib/Target/AMDGPU/SIInstructions.td:382-383

+let Defs = [EXEC] in
+def SI_KILL_CLEANUP : SPseudoInstSI <(outs), (ins)>;
+
----------------
nhaehnle wrote:
> Shouldn't this also Use EXEC?
It does, this entire block of instructions is Use EXEC.

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D82737/new/

https://reviews.llvm.org/D82737