[PATCH] D101380: [AMDGPU] Move insertion of function entry waitcnt later

[Austin Kerbow via Phabricator via llvm-commits]

kerbowa added inline comments.


================
Comment at: llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp:1601
+
+  if (!MFI->isEntryFunction()) {
+    // Wait for any outstanding memory operations that the input registers may
----------------
arsenm wrote:
> I thought we made this explicitly a property of the calling convention, so we could have this not be inserted for future conventions. I guess if this isn't checking the function for that already, it's a preexisting problem
When would that waitcnt be added if it's a property of the CC? It seems like these changes could help avoid it being added twice since they would be combined. 


================
Comment at: llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp:1606
+
+    // TODO: Could insert earlier and schedule more liberally with operations
+    // that only use caller preserved registers.
----------------
rampitec wrote:
> Is this comment still relevant?
It might be, if what it's saying is that we could move this waitcnt later until the function uses callee saved registers.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D101380/new/

https://reviews.llvm.org/D101380