revert

Author: liqiangxl

csrc/alias_analysis.cpp

@@ -288,25 +288,13 @@ void AliasFinder::handle(const SliceOp* slice) {
       out, in, Layout(std::move(out_allocation), std::move(out_contiguity)));
 }
 
-// Only consider broadcast aliasing when IO tensor is involved.
-// Intermediate broadcasts will be fused with other ops and don't need explicit
-// alias handling. Limiting to fusion boundaries avoids unnecessary allocation
-// domain changes on intermediate tensors, which may trigger transpose scheduler
-// when pointwise is preferred. For example, when a normalization kernel is
-// segmented, we prefer reduction + pointwise instead of reduction + transpose.
-// See SmemPersistentNotSupportedIn3DReduction.
 void AliasFinder::handle(const BroadcastOp* bcast) {
   auto* in = dynamic_cast<TensorView*>(bcast->in());
   if (in == nullptr) {
     return;
   }
   auto* out = bcast->out()->as<TensorView>();
 
-  // No alias analysis needed if no IO tensors are involved
-  if (!out->isFusionOutput() && !in->isFusionInput()) {
-    return;
-  }
-
   std::optional<Layout> out_layout =
       mapInLayoutToOutRoot(analysis_.preferredLayout(in), in, out);
   if (!out_layout.has_value()) {

tests/cpp/test_persistent_buffer.cpp

@@ -1257,13 +1257,6 @@ TEST_F(PersistentBufferTest, SmemPersistentNotSupportedIn3DReduction) {
   // persistent is not supported yet for 3D reduction.
   EXPECT_TRUE(executor_cache.getMostRecentKernelRuntime()->isSegmented());
 
-  // expect reduction and pointwise scheduler
-  EXPECT_THAT(
-      executor_cache.getMostRecentKernelRuntime()->fusionSegments()->groups(),
-      UnorderedElementsAre(
-          HeuristicIs(SchedulerType::PointWise),
-          HeuristicIs(SchedulerType::Reduction)));
-
   testValidate(executor_cache.fusion(), cg_outputs, {t0}, __LINE__, __FILE__);
 }