[llvm-branch-commits] [llvm] [AMDGPU] Return two MMOs for load-to-lds and store-from-lds intrinsics (PR #175845)

[Nicolai Hähnle via llvm-branch-commits]

================
@@ -1432,8 +1432,24 @@ void SITargetLowering::getTgtMemIntrinsic(SmallVectorImpl<IntrinsicInfo> &Infos,
       case Intrinsic::amdgcn_struct_buffer_load_lds:
       case Intrinsic::amdgcn_struct_ptr_buffer_load_lds: {
         unsigned Width = cast<ConstantInt>(CI.getArgOperand(2))->getZExtValue();
+
+        // Entry 0: Load from buffer.
+        // Don't set an offset, since the pointer value always represents the
+        // base of the buffer.
         Info.memVT = EVT::getIntegerVT(CI.getContext(), Width * 8);
-        Info.ptrVal = CI.getArgOperand(1);
+        Info.flags &= ~MachineMemOperand::MOStore;
+        Infos.push_back(Info);
+
+        // Entry 1: Store to LDS.
+        // Instruction offset is applied, and an additional per-lane offset
+        // which we simulate using a larger memory type.
+        Info.memVT = EVT::getIntegerVT(CI.getContext(), Width * 8 * 64);
----------------
nhaehnle wrote:

Ugh, you're right. To be precise, it may be either 32 or 64.

https://github.com/llvm/llvm-project/pull/175845