[Mlir-commits] [mlir] Refactor LoopFuseSiblingOp and support parallel fusion (PR #94391)

[Oleksandr Alex Zinenko]

================
@@ -618,6 +618,45 @@ void ForOp::getSuccessorRegions(RegionBranchPoint point,
 
 SmallVector<Region *> ForallOp::getLoopRegions() { return {&getRegion()}; }
 
+FailureOr<LoopLikeOpInterface> ForallOp::replaceWithAdditionalYields(
+    RewriterBase &rewriter, ValueRange newInitOperands,
+    bool replaceInitOperandUsesInLoop,
+    const NewYieldValuesFn &newYieldValuesFn) {
+  // Create a new loop before the existing one, with the extra operands.
+  OpBuilder::InsertionGuard g(rewriter);
+  rewriter.setInsertionPoint(getOperation());
+  SmallVector<Value> inits(getOutputs());
+  inits.append(newInitOperands.begin(), newInitOperands.end());
+  scf::ForallOp newLoop = rewriter.create<scf::ForallOp>(
+      getLoc(), getMixedLowerBound(), getMixedUpperBound(), getMixedStep(),
+      inits, getMapping());
+
+  rewriter.eraseOp(newLoop.getTerminator());
----------------
ftynse wrote:

I don't remember if it's the case here, but for some loops the terminator is only created when there are no results in the loop.

https://github.com/llvm/llvm-project/pull/94391