[PATCH] D39731: AMDGPU: Don't use MUBUF vaddr if address may overflow

[Matt Arsenault via Phabricator via llvm-commits]

arsenm added inline comments.


================
Comment at: lib/Target/AMDGPU/SIISelLowering.cpp:170
   setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
+  //setOperationAction(ISD::FrameIndex, MVT::i32, Custom);
   setOperationAction(ISD::ConstantPool, MVT::v2i64, Expand);
----------------
rampitec wrote:
> How does LowerOperation work without this?
This isn't custom lowered, this is leftover junk.


================
Comment at: test/CodeGen/AMDGPU/amdgpu.private-memory.ll:395
+; SI: v_add_i32_e32 [[ADD_OFFSET:v[0-9]+]], vcc, 5,
+; SI: buffer_load_dword v{{[0-9]+}}, [[ADD_OFFSET:v[0-9]+]], s[{{[0-9]+:[0-9]+}}], s{{[0-9]+}} offen ;
 define amdgpu_kernel void @ptrtoint(i32 addrspace(1)* %out, i32 %a, i32 %b) #0 {
----------------
rampitec wrote:
> It looks like regression to me.
Yes, this is the point. This needs to be more conservative


https://reviews.llvm.org/D39731