================
@@ -594,45 +594,85 @@ convertOmpOrderedRegion(Operation &opInst,
llvm::IRBuilderBase &builder,
/// Allocate space for privatized reduction variables.
template <typename T>
-static void allocByValReductionVars(
- T loop, ArrayRef<BlockArgument> reductionArgs, llvm::IRBuilderBase
&builder,
- LLVM::ModuleTranslation &moduleTranslation,
- llvm::OpenMPIRBuilder::InsertPointTy &allocaIP,
- SmallVectorImpl<omp::DeclareReductionOp> &reductionDecls,
- SmallVectorImpl<llvm::Value *> &privateReductionVariables,
- DenseMap<Value, llvm::Value *> &reductionVariableMap,
- llvm::ArrayRef<bool> isByRefs) {
+static LogicalResult
+allocReductionVars(T loop, ArrayRef<BlockArgument> reductionArgs,
+ llvm::IRBuilderBase &builder,
+ LLVM::ModuleTranslation &moduleTranslation,
+ llvm::OpenMPIRBuilder::InsertPointTy &allocaIP,
+ SmallVectorImpl<omp::DeclareReductionOp> &reductionDecls,
+ SmallVectorImpl<llvm::Value *> &privateReductionVariables,
+ DenseMap<Value, llvm::Value *> &reductionVariableMap,
+ llvm::ArrayRef<bool> isByRefs) {
llvm::IRBuilderBase::InsertPointGuard guard(builder);
builder.SetInsertPoint(allocaIP.getBlock()->getTerminator());
+ // delay creating stores until after all allocas
+ SmallVector<std::pair<llvm::Value *, llvm::Value *>> storesToCreate;
+ storesToCreate.reserve(loop.getNumReductionVars());
+
for (std::size_t i = 0; i < loop.getNumReductionVars(); ++i) {
- if (isByRefs[i])
- continue;
- llvm::Value *var = builder.CreateAlloca(
- moduleTranslation.convertType(reductionDecls[i].getType()));
- moduleTranslation.mapValue(reductionArgs[i], var);
- privateReductionVariables[i] = var;
- reductionVariableMap.try_emplace(loop.getReductionVars()[i], var);
+ Region &allocRegion = reductionDecls[i].getAllocRegion();
+ if (isByRefs[i]) {
+ if (allocRegion.empty())
----------------
tblah wrote:
The alloc region is optional. If it isn't included it could still be included
in the initialization region as normal. This could happen for example if there
is no part of allocation that is on the stack (because we don't want a call to
malloc mixed into the middle of allocas).
https://github.com/llvm/llvm-project/pull/102524
_______________________________________________
llvm-branch-commits mailing list
llvm-branch-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-branch-commits
