[llvm-branch-commits] [DirectX] Lower `@llvm.dx.typedBufferLoad` to DXIL ops (PR #104252)

[David Peixotto via llvm-branch-commits]

================
@@ -236,6 +236,59 @@ class OpLowerer {
       lowerToBindAndAnnotateHandle(F);
   }
 
+  void lowerTypedBufferLoad(Function &F) {
+    IRBuilder<> &IRB = OpBuilder.getIRB();
+    Type *Int32Ty = IRB.getInt32Ty();
+
+    replaceFunction(F, [&](CallInst *CI) -> Error {
+      IRB.SetInsertPoint(CI);
+
+      Value *Handle =
+          createTmpHandleCast(CI->getArgOperand(0), OpBuilder.getHandleType());
+      Value *Index0 = CI->getArgOperand(1);
+      Value *Index1 = UndefValue::get(Int32Ty);
+      Type *RetTy = OpBuilder.getResRetType(CI->getType()->getScalarType());
+
+      std::array<Value *, 3> Args{Handle, Index0, Index1};
+      Expected<CallInst *> OpCall =
+          OpBuilder.tryCreateOp(OpCode::BufferLoad, Args, RetTy);
+      if (Error E = OpCall.takeError())
+        return E;
+
+      std::array<Value *, 4> Extracts = {};
+
+      // We've switched the return type from a vector to a struct, but at this
+      // point most vectors have probably already been scalarized. Try to
+      // forward arguments directly rather than inserting into and immediately
+      // extracting from a vector.
+      for (Use &U : make_early_inc_range(CI->uses()))
+        if (auto *EEI = dyn_cast<ExtractElementInst>(U.getUser()))
+          if (auto *Index = dyn_cast<ConstantInt>(EEI->getIndexOperand())) {
+            size_t IndexVal = Index->getZExtValue();
+            assert(IndexVal < 4 && "Index into buffer load out of range");
----------------
dmpots wrote:

Have we thought about how `CheckAccessFullyMapped` is going to fit into this scheme? In dxil, that will lower to an extract of index 4 on the `dx.types.ResRet` struct.

https://github.com/llvm/llvm-project/pull/104252