[turbofan] Introduce fix up phase after memory optimization

On arm processors, we need to have some kind of subexpression
elimination running after memory optimizer, so that we reuse index
calculation for loads and stores. This CL introduces a small cleanup
phase after memory optimizer.

Bug: chromium:947225
Change-Id: Ifda7b348d968d58f31947a4ba139863059f4112d
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/1547664Reviewed-by: Benedikt Meurer <bmeurer@chromium.org>
Commit-Queue: Jaroslav Sevcik <jarin@chromium.org>
Cr-Commit-Position: refs/heads/master@{#60602}

src/compiler/pipeline.cc

@@ -1547,6 +1547,21 @@ struct LateOptimizationPhase {
   }
 };
 
+struct MachineOperatorOptimizationPhase {
+  static const char* phase_name() { return "V8.TFMachineOperatorOptimization"; }
+
+  void Run(PipelineData* data, Zone* temp_zone) {
+    GraphReducer graph_reducer(temp_zone, data->graph(),
+                               data->jsgraph()->Dead());
+    ValueNumberingReducer value_numbering(temp_zone, data->graph()->zone());
+    MachineOperatorReducer machine_reducer(&graph_reducer, data->jsgraph());
+
+    AddReducer(data, &graph_reducer, &machine_reducer);
+    AddReducer(data, &graph_reducer, &value_numbering);
+    graph_reducer.ReduceGraph();
+  }
+};
+
 struct CsaOptimizationPhase {
   static const char* phase_name() { return "V8.CSAOptimization"; }
 
@@ -2114,6 +2129,12 @@ bool PipelineImpl::OptimizeGraph(Linkage* linkage) {
   Run<MemoryOptimizationPhase>();
   RunPrintAndVerify(MemoryOptimizationPhase::phase_name(), true);
 
+  // Run value numbering and machine operator reducer to optimize load/store
+  // address computation (in particular, reuse the address computation whenever
+  // possible).
+  Run<MachineOperatorOptimizationPhase>();
+  RunPrintAndVerify(MachineOperatorOptimizationPhase::phase_name(), true);
+
   data->source_positions()->RemoveDecorator();
   if (data->info()->trace_turbo_json_enabled()) {
     data->node_origins()->RemoveDecorator();