[PATCH] D123548: AMDGPU: Emit metadata for the hidden_multigrid_sync_arg conditionally

[Changpeng Fang via Phabricator via llvm-commits]

cfang marked an inline comment as done.
cfang added inline comments.


================
Comment at: llvm/lib/Target/AMDGPU/AMDGPUAttributor.cpp:545
+      return false;
+    AAPointerInfo::OffsetAndSize OAS(AMDGPU::ImplicitArg::HEAP_PTR_OFFSET, 8);
+    return funcRetrievesImplicitKernelArg(A, OAS);
----------------
b-sumner wrote:
> I expected to see getMultigridSyncArgImplicitArgPosition used here.
Thanks. I was thinking whether we should only limit the change to v5
and not sure how did I end up with the garbage code. Will correct it.

The reason I was not sure for pre-v5 is that printf and hostcall hold the same offset, so it is not possible to use implicitarg_ptr + offset to detect.

To say that, our hostcall_ptr check logic is broken because even though you are doing printf, you may still end up hostcall metadata emission ( for earlier code object versions). 


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D123548/new/

https://reviews.llvm.org/D123548