Adding to execute_region_op some missing support

Author: ddubov100

mlir/include/mlir/Dialect/SCF/IR/SCFOps.td

@@ -77,7 +77,9 @@ def ConditionOp : SCF_Op<"condition", [
 //===----------------------------------------------------------------------===//
 
 def ExecuteRegionOp : SCF_Op<"execute_region", [
-    DeclareOpInterfaceMethods<RegionBranchOpInterface>]> {
+    DeclareOpInterfaceMethods<RegionBranchOpInterface>,
+    RecursiveMemoryEffects,
+    DeclareOpInterfaceMethods<MemoryEffectsOpInterface>]> {//, SingleBlockImplicitTerminator<"scf::YieldOp">]> { //, RecursiveMemoryEffects]> {
   let summary = "operation that executes its region exactly once";
   let description = [{
     The `scf.execute_region` operation is used to allow multiple blocks within SCF

mlir/lib/Dialect/SCF/IR/SCF.cpp

@@ -291,9 +291,93 @@ struct MultiBlockExecuteInliner : public OpRewritePattern<ExecuteRegionOp> {
   }
 };
 
+// Pattern to eliminate ExecuteRegionOp results when it only forwards external values.
+// It operates only on execute regions with single terminator yield operation.
+struct ExecuteRegionForwardingEliminator
+    : public OpRewritePattern<ExecuteRegionOp> {
+  using OpRewritePattern<ExecuteRegionOp>::OpRewritePattern;
+
+  LogicalResult matchAndRewrite(ExecuteRegionOp op,
+                                PatternRewriter &rewriter) const override {
+    if (op.getNumResults() == 0)
+      return failure();
+
+    SmallVector<Operation *> yieldOps;
+    for (Block &block : op.getRegion()) {
+      if (auto yield = dyn_cast<scf::YieldOp>(block.getTerminator())) {
+        if (yield.getResults().empty())
+          continue;
+        yieldOps.push_back(yield.getOperation());
+      }
+    }
+
+    if (yieldOps.size() != 1)
+      return failure();
+
+    auto yieldOp = dyn_cast<scf::YieldOp>(yieldOps.front());
+    auto yieldedValues = yieldOp.getOperands();
+    // Check if all yielded values are from outside the region
+    bool allExternal = true;
+    for (Value yieldedValue : yieldedValues) {
+      if (isValueFromInsideRegion(yieldedValue, op)) {
+        allExternal = false;
+        break;
+      }
+    }
+
+    if (!allExternal)
+      return failure();
+
+    // All yielded values are external - create a new execute_region with no
+    // results.
+    auto newOp = rewriter.create<ExecuteRegionOp>(op.getLoc(), TypeRange{});
+    newOp->setAttrs(op->getAttrs());
+
+    // Move the region content to the new operation
+    newOp.getRegion().takeBody(op.getRegion());
+
+    // Replace the yield operation with a new yield operation with no results.
+    rewriter.setInsertionPoint(yieldOp);
+    rewriter.eraseOp(yieldOp);
+    rewriter.create<scf::YieldOp>(yieldOp.getLoc());
+
+    // Replace the old operation with the external values directly.
+    rewriter.replaceOp(op, yieldedValues);
+    return success();
+  }
+
+private:
+  bool isValueFromInsideRegion(Value value,
+                              ExecuteRegionOp executeRegionOp) const {
+    // Check if the value is defined within the execute_region
+    if (Operation *defOp = value.getDefiningOp()) {
+      return executeRegionOp.getRegion().isAncestor(defOp->getParentRegion());
+    }
+ 
+    // If it's a block argument, check if it's from within the region
+    if (BlockArgument blockArg = dyn_cast<BlockArgument>(value)) {
+      return executeRegionOp.getRegion().isAncestor(blockArg.getParentRegion());
+    }
+ 
+    return false; // Value is from outside the region
+  }
+};
+
 void ExecuteRegionOp::getCanonicalizationPatterns(RewritePatternSet &results,
                                                   MLIRContext *context) {
-  results.add<SingleBlockExecuteInliner, MultiBlockExecuteInliner>(context);
+  results.add<SingleBlockExecuteInliner, MultiBlockExecuteInliner, ExecuteRegionForwardingEliminator>(
+                context);
+}
+
+void ExecuteRegionOp::getEffects(
+    SmallVectorImpl<SideEffects::EffectInstance<MemoryEffects::Effect>>
+        &effects) {
+  if (!getNoInline())
+    return;
+  // In case there is attribute no_inline we want the region not to be inlined
+  // into the parent operation.
+  effects.emplace_back(MemoryEffects::Write::get(),
+                       SideEffects::DefaultResource::get());
 }
 
 void ExecuteRegionOp::getSuccessorRegions(

mlir/test/Dialect/SCF/canonicalize.mlir

@@ -1604,6 +1604,28 @@ func.func @func_execute_region_inline_multi_yield() {
 
 // -----
 
+// CHECK-LABEL:   func.func private @canonicalize_execute_region_yeilding_external_value(
+// CHECK-SAME:                                                                           %[[VAL_0:.*]]: tensor<1x120xui8>) -> tensor<1x60xui8> {
+// CHECK:           %[[VAL_1:.*]] = memref.alloc() {alignment = 64 : i64} : memref<1x60xui8>
+// CHECK:           scf.execute_region no_inline {
+// CHECK:             scf.yield
+// CHECK:           }
+// CHECK:           %[[VAL_2:.*]] = bufferization.to_tensor %[[VAL_1]] : memref<1x60xui8> to tensor<1x60xui8>
+// CHECK:           return %[[VAL_2]] : tensor<1x60xui8>
+// CHECK:         }
+
+func.func private @canonicalize_execute_region_yeilding_external_value(%arg0: tensor<1x120xui8>) -> tensor<1x60xui8> {
+  %alloc = memref.alloc() {alignment = 64 : i64} : memref<1x60xui8>
+  %0 = bufferization.to_tensor %alloc : memref<1x60xui8> to tensor<1x60xui8>
+  %1 = scf.execute_region -> memref<1x60xui8> no_inline {
+    scf.yield %alloc : memref<1x60xui8>
+  }
+  %2 = bufferization.to_tensor %1 : memref<1x60xui8> to tensor<1x60xui8>
+  return %2 : tensor<1x60xui8>
+}
+
+// -----
+
 // CHECK-LABEL: func @canonicalize_parallel_insert_slice_indices(
 //  CHECK-SAME:     %[[arg0:.*]]: tensor<1x5xf32>, %[[arg1:.*]]: tensor<?x?xf32>
 func.func @canonicalize_parallel_insert_slice_indices(