[clang] [OpenACC][CIR] Implement 'reduction' combiner lowering for 5 ops (PR #162906)

[Andy Kaylor via cfe-commits]

================
@@ -527,16 +527,140 @@ void OpenACCRecipeBuilderBase::createFirstprivateRecipeCopy(
 // doesn't restore it aftewards.
 void OpenACCRecipeBuilderBase::createReductionRecipeCombiner(
     mlir::Location loc, mlir::Location locEnd, mlir::Value mainOp,
-    mlir::acc::ReductionRecipeOp recipe, size_t numBounds) {
+    mlir::acc::ReductionRecipeOp recipe, size_t numBounds, QualType origType,
+    llvm::ArrayRef<OpenACCReductionRecipe::CombinerRecipe> combinerRecipes) {
   mlir::Block *block =
       createRecipeBlock(recipe.getCombinerRegion(), mainOp.getType(), loc,
                         numBounds, /*isInit=*/false);
   builder.setInsertionPointToEnd(&recipe.getCombinerRegion().back());
   CIRGenFunction::LexicalScope ls(cgf, loc, block);
 
-  mlir::BlockArgument lhsArg = block->getArgument(0);
+  mlir::Value lhsArg = block->getArgument(0);
+  mlir::Value rhsArg = block->getArgument(1);
+  llvm::MutableArrayRef<mlir::BlockArgument> boundsRange =
+      block->getArguments().drop_front(2);
+
+  if (llvm::any_of(combinerRecipes, [](auto &r) { return r.Op == nullptr; })) {
+    cgf.cgm.errorNYI(loc, "OpenACC Reduction combiner not generated");
+    mlir::acc::YieldOp::create(builder, locEnd, block->getArgument(0));
+    return;
+  }
+
+  // apply the bounds so that we can get our bounds emitted correctly.
+  for (mlir::BlockArgument boundArg : llvm::reverse(boundsRange))
+    std::tie(lhsArg, rhsArg) =
+        createBoundsLoop(lhsArg, rhsArg, boundArg, loc, /*inverse=*/false);
+
+  // Emitter for when we know this isn't a struct or array we have to loop
+  // through. This should work for the 'field' once the get-element call has
+  // been made.
+  auto emitSingleCombiner =
+      [&](mlir::Value lhsArg, mlir::Value rhsArg,
+          const OpenACCReductionRecipe::CombinerRecipe &combiner) {
+        mlir::Type elementTy =
+            mlir::cast<cir::PointerType>(lhsArg.getType()).getPointee();
+        CIRGenFunction::DeclMapRevertingRAII declMapRAIILhs{cgf, combiner.LHS};
+        cgf.setAddrOfLocalVar(
+            combiner.LHS, Address{lhsArg, elementTy,
+                                  cgf.getContext().getDeclAlign(combiner.LHS)});
+        CIRGenFunction::DeclMapRevertingRAII declMapRAIIRhs{cgf, combiner.RHS};
+        cgf.setAddrOfLocalVar(
+            combiner.RHS, Address{rhsArg, elementTy,
+                                  cgf.getContext().getDeclAlign(combiner.RHS)});
+
+        [[maybe_unused]] mlir::LogicalResult stmtRes =
+            cgf.emitStmt(combiner.Op, /*useCurrentScope=*/true);
+      };
+
+  // Emitter for when we know this is either a non-array or element of an array
+  // (which also shouldn't be an array type?). This function should generate the
+  // loop to do this on each individual array or struct element (if necessary).
----------------
andykaylor wrote:

I don't understand the comment. If we know this is a non-array, what does the part about a loop to do this on each individual array element mean?

https://github.com/llvm/llvm-project/pull/162906