[PATCH] D92951: AMDGPU: If a store defines (alias) a load, it clobbers the load.

Mon Dec 14 15:40:41 PST 2020

cfang marked 2 inline comments as done.
cfang added inline comments.

================
Comment at: llvm/test/CodeGen/AMDGPU/store-clobbers-load.ll:15
+  %vaddr = bitcast [4 x i32] addrspace(5)* %alloca to <4 x i32> addrspace(5)*
+  %zero = load <4 x i32>, <4 x i32> addrspace(5)* %vaddr, align 16
+  %one = insertelement <4 x i32> %zero, i32 1, i32 1
----------------
arsenm wrote:
> Why are we bothering to annotate a private memory access in the first place? 
This is a good question. Just found that the only place that uses amdgpu.noclobber metadata is
in SITargetLowering::LowerLOAD for global address space.

Should we limit the analysis only to global address space (I am not clear whether there will be future
usage of amdgpu.noclobber)?

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D92951/new/

https://reviews.llvm.org/D92951