[flang-commits] [flang] [llvm] [mlir] [openmp] [Flang][OpenMP] Lower scan directive and inscan reduction modifier (PR #206747)

[CHANDRA GHALE via flang-commits]

================
@@ -5489,17 +5525,63 @@ OpenMPIRBuilder::InsertPointOrErrorTy OpenMPIRBuilder::emitScanReduction(
         ScanRedInfo->Span,
         llvm::ConstantInt::get(ScanRedInfo->Span->getType(), 1));
     Builder.SetInsertPoint(InputBB);
-    Builder.CreateBr(LoopBB);
+    // Branch around the orig-val seed load and combine when the loop has zero
+    // logical iterations. When Span == 0 no scan-buffer entry is populated, so
+    // loading buffer[0] and feeding it to a (possibly input-sensitive
+    // user-defined) reduction combiner would be undefined behavior. In that
+    // case skip straight to the exit; the finals' zero-trip guard preserves
+    // orig-val as the result.
+    llvm::Value *HasElem = Builder.CreateICmpUGT(
+        ScanRedInfo->Span, llvm::ConstantInt::get(IndexTy, 0));
+    llvm::BasicBlock *SeedBB =
+        BasicBlock::Create(CurFn->getContext(), "omp.scan.seed", CurFn, ExitBB);
+    Builder.CreateCondBr(HasElem, SeedBB, ExitBB);
+    Builder.SetInsertPoint(SeedBB);
+    // Combine the original variable's incoming value (orig-val) into the first
+    // buffer element before computing the prefix sum. Per the OpenMP scan
+    // semantics orig-val is a single prefix element that must be reflected in
+    // every inclusive/exclusive scan result and in the final reduction value.
+    // Folding it in here lets the log-scan prefix computation propagate it into
+    // every element (and into the finals' read of `buffer[Span]`), regardless
+    // of whether an iteration's input phase accumulates into or overwrites the
+    // reduction variable. Span > 0 is guaranteed here, so buffer[1] is valid.
+    for (ReductionInfo RedInfo : ReductionInfos) {
+      Value *ReductionVal = RedInfo.PrivateVariable;
+      Value *BuffPtr = (*(ScanRedInfo->ScanBuffPtrs))[ReductionVal];
+      Value *Buff = Builder.CreateLoad(Builder.getPtrTy(), BuffPtr);
+      Type *DestTy = RedInfo.ElementType;
+      Value *ElemPtr = Builder.CreateInBoundsGEP(
----------------
chandraghale wrote:

You're right — the combiner operands were reversed at both sites. ReductionGen's first operand is omp_out (the accumulator), so the earlier prefix element must go first; I've swapped them in the orig-val seed (OrigVal, Elem) and the prefix computation (RHS, LHS). Verified with your left reducer (now b=10 10 10, x=10) and added a noncommutative sub-combiner lit test so +'s commutativity can't mask this again


https://github.com/llvm/llvm-project/pull/206747