[Mlir-commits] [mlir] [OpenMP][OMPIRBuilder] Add delayed privatization support for `wsloop` (PR #118463)

[Kareem Ergawy]

================
@@ -1030,6 +1029,99 @@ mapInitializationArgs(T loop, LLVM::ModuleTranslation &moduleTranslation,
   }
 }
 
+template <typename OP>
+static LogicalResult
+initReductionVars(OP op, ArrayRef<BlockArgument> reductionArgs,
+                  llvm::IRBuilderBase &builder,
+                  LLVM::ModuleTranslation &moduleTranslation,
+                  llvm::BasicBlock *latestAllocaBlock,
+                  SmallVectorImpl<omp::DeclareReductionOp> &reductionDecls,
+                  SmallVectorImpl<llvm::Value *> &privateReductionVariables,
+                  DenseMap<Value, llvm::Value *> &reductionVariableMap,
+                  llvm::ArrayRef<bool> isByRef,
+                  SmallVectorImpl<DeferredStore> &deferredStores) {
+  if (op.getNumReductionVars() == 0)
+    return success();
+
+  llvm::IRBuilderBase::InsertPointGuard guard(builder);
+
+  builder.SetInsertPoint(latestAllocaBlock->getTerminator());
+  llvm::BasicBlock *initBlock = splitBB(builder, true, "omp.reduction.init");
+  auto allocaIP = llvm::IRBuilderBase::InsertPoint(
+      latestAllocaBlock, latestAllocaBlock->getTerminator()->getIterator());
+  builder.restoreIP(allocaIP);
+  SmallVector<llvm::Value *> byRefVars(op.getNumReductionVars());
+
+  for (unsigned i = 0; i < op.getNumReductionVars(); ++i) {
----------------
ergawy wrote:

Because this creates some allocations. So instead of intermingling these allocas with the intialization logic that follows in the next loop, I just emit these allocations in the alloc region and then the init logic after all allocs are done.

https://github.com/llvm/llvm-project/pull/118463