[Mlir-commits] [mlir] [MLIR][SCF] Add an API to fuse consumer to a producer within scf loop (PR #88712)

[llvmlistbot at llvm.org]

================
@@ -1100,6 +1102,398 @@ mlir::scf::tileConsumerAndFuseProducersUsingSCF(
                                    replacements};
 }
 
+//===----------------------------------------------------------------------===//
+// tileAndFuseConsumerUsingSCF implementation.
+//===----------------------------------------------------------------------===//
+
+/// A utility function that checks whether the passed value has only one user.
+/// In case the defining operation is a tensor.insert_slice, it checks if the
+/// user is scf.yield.
+static LogicalResult checkAssumptionForFusingConsumer(Value result) {
+  Value::use_range uses = result.getUses();
+  if (!llvm::hasSingleElement(uses)) {
+    LLVM_DEBUG(llvm::dbgs() << "Too many uses of the candidate slice op\n");
+    return failure();
+  }
+  OpOperand &operandUse = (*uses.begin());
+  Operation *userOp = operandUse.getOwner();
+  if (!isa<scf::YieldOp>(userOp)) {
+    LLVM_DEBUG(llvm::dbgs()
+               << "Expected scf.yield to be the only user, but got -> "
+               << (*userOp));
+    return failure();
+  }
+  if (result.getDefiningOp()->getBlock() != userOp->getBlock()) {
+    LLVM_DEBUG(llvm::dbgs() << "Expected tensor.insert_slice and scf.yield to "
+                               "be in the same block\n");
+    return failure();
+  }
+  return success();
+}
+
+/// Fetch the first untiled consumer of a scf.for's result which is yielded by
+/// a tensor.insert_slice. This function makes the following assumptions :-
+/// 1.  tensor.insert_slice has scf.yield as its only user.
+/// 2.  scf.for's corresponding result has only one use.
+static OpOperand *
+getUntiledConsumerFromSlice(tensor::InsertSliceOp candidateSliceOp) {
+  Value sliceResult = candidateSliceOp.getResult();
+  if (failed(checkAssumptionForFusingConsumer(candidateSliceOp.getResult()))) {
+    return nullptr;
+  }
+  // Step 1. Fetch the corresponding output.
+  OpOperand &yieldOpOperand = (*sliceResult.getUses().begin());
+  unsigned resultNumber = yieldOpOperand.getOperandNumber();
+  // Step 2. Check containing op is scf.for.
+  Operation *containingOp = candidateSliceOp->getParentOp();
+  auto forOp = dyn_cast<scf::ForOp>(containingOp);
+  if (!forOp) {
+    return nullptr;
+  }
+  Value resultingValue = forOp->getResult(resultNumber);
+
+  // Step 3. Check resulting value of scf.for has exactly one use.
+  if (!llvm::hasSingleElement(resultingValue.getUses())) {
+    return nullptr;
+  }
+
+  // Step 4. Get uses.
+  OpOperand &operand = (*resultingValue.getUses().begin());
+  Operation *consumerOp = operand.getOwner();
+  // TODO: We have to init result of consumer before scf.for, use
+  //       DestinationStyleOpInterface to get result shape from init for now.
+  //       Add support for other op such as op has InferTypeOpInterface.
+  if (!isa<TilingInterface>(consumerOp) ||
+      !isa<DestinationStyleOpInterface>(consumerOp)) {
+    return nullptr;
+  }
+  if (containingOp->getBlock() != consumerOp->getBlock()) {
+    return nullptr;
+  }
+  return &operand;
+}
+
+/// Fetch the first untiled consumer of a scf.forall's result which is yielded
+/// by a tensor.parallel_insert_slice.
+static OpOperand *
----------------
MaheshRavishankar wrote:

Nit: It might be better to just return a  `FailureOr<OpOperand*>`  instead of returning nullptr.

https://github.com/llvm/llvm-project/pull/88712