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

[Krzysztof Drewniak via llvm-branch-commits]

================
@@ -1626,35 +1642,64 @@ void SITargetLowering::getTgtMemIntrinsic(SmallVectorImpl<IntrinsicInfo> &Infos,
   case Intrinsic::amdgcn_cluster_load_async_to_lds_b32:
   case Intrinsic::amdgcn_cluster_load_async_to_lds_b64:
   case Intrinsic::amdgcn_cluster_load_async_to_lds_b128: {
+    // Entry 0: Load from source (global/flat).
     Info.opc = ISD::INTRINSIC_VOID;
     Info.memVT = EVT::getIntegerVT(CI.getContext(), getIntrMemWidth(IntrID));
-    Info.ptrVal = CI.getArgOperand(1);
-    Info.flags |= MachineMemOperand::MOLoad | MachineMemOperand::MOStore;
+    Info.ptrVal = CI.getArgOperand(0); // Global pointer
+    Info.offset = cast<ConstantInt>(CI.getArgOperand(2))->getSExtValue();
+    Info.flags |= MachineMemOperand::MOLoad;
+    Infos.push_back(Info);
+
+    // Entry 1: Store to LDS (same offset).
+    Info.flags &= ~MachineMemOperand::MOLoad;
+    Info.flags |= MachineMemOperand::MOStore;
+    Info.ptrVal = CI.getArgOperand(1); // LDS pointer
     Infos.push_back(Info);
     return;
   }
   case Intrinsic::amdgcn_global_store_async_from_lds_b8:
   case Intrinsic::amdgcn_global_store_async_from_lds_b32:
   case Intrinsic::amdgcn_global_store_async_from_lds_b64:
   case Intrinsic::amdgcn_global_store_async_from_lds_b128: {
+    // Entry 0: Load from LDS.
     Info.opc = ISD::INTRINSIC_VOID;
     Info.memVT = EVT::getIntegerVT(CI.getContext(), getIntrMemWidth(IntrID));
-    Info.ptrVal = CI.getArgOperand(0);
-    Info.flags |= MachineMemOperand::MOLoad | MachineMemOperand::MOStore;
+    Info.ptrVal = CI.getArgOperand(1); // LDS pointer
+    Info.offset = cast<ConstantInt>(CI.getArgOperand(2))->getSExtValue();
+    Info.flags |= MachineMemOperand::MOLoad;
+    Infos.push_back(Info);
+
+    // Entry 1: Store to global (same offset).
+    Info.flags &= ~MachineMemOperand::MOLoad;
+    Info.flags |= MachineMemOperand::MOStore;
+    Info.ptrVal = CI.getArgOperand(0); // Global pointer
     Infos.push_back(Info);
     return;
   }
   case Intrinsic::amdgcn_load_to_lds:
   case Intrinsic::amdgcn_global_load_lds: {
-    Info.opc = ISD::INTRINSIC_VOID;
     unsigned Width = cast<ConstantInt>(CI.getArgOperand(2))->getZExtValue();
-    Info.memVT = EVT::getIntegerVT(CI.getContext(), Width * 8);
-    Info.ptrVal = CI.getArgOperand(1);
-    Info.flags |= MachineMemOperand::MOLoad | MachineMemOperand::MOStore;
     auto *Aux = cast<ConstantInt>(CI.getArgOperand(CI.arg_size() - 1));
-    if (Aux->getZExtValue() & AMDGPU::CPol::VOLATILE)
+    bool IsVolatile = Aux->getZExtValue() & AMDGPU::CPol::VOLATILE;
+
+    // Entry 0: Load from source (global/flat).
+    Info.opc = ISD::INTRINSIC_VOID;
+    Info.memVT = EVT::getIntegerVT(CI.getContext(), Width * 8);
+    Info.ptrVal = CI.getArgOperand(0); // Source pointer
+    Info.offset = cast<ConstantInt>(CI.getArgOperand(3))->getSExtValue();
+    Info.flags |= MachineMemOperand::MOLoad;
+    if (IsVolatile)
       Info.flags |= MachineMemOperand::MOVolatile;
     Infos.push_back(Info);
+
+    // Entry 1: Store to LDS.
+    // Same offset from the instruction, but an additional per-lane offset is
+    // added. Represent that using a wider memory type.
+    Info.memVT = EVT::getIntegerVT(CI.getContext(), Width * 8 * 64);
----------------
krzysz00 wrote:

simiarly

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