[turbofan] Improve the store-store elimination.

It can now deal with multiple objects at the same time (but no
aliasing), and it propagates store information upwards across effect
chain splits.

R=jarin
BUG=

Review-Url: https://codereview.chromium.org/2159303002
Cr-Commit-Position: refs/heads/master@{#38509}

src/compiler/pipeline.cc

@@ -1058,12 +1058,36 @@ struct EffectControlLinearizationPhase {
   }
 };
 
+// The store-store elimination greatly benefits from doing a common operator
+// reducer just before it, to eliminate conditional deopts with a constant
+// condition.
+
+struct DeadCodeEliminationPhase {
+  static const char* phase_name() { return "common operator reducer"; }
+
+  void Run(PipelineData* data, Zone* temp_zone) {
+    // Run the common operator reducer.
+    JSGraphReducer graph_reducer(data->jsgraph(), temp_zone);
+    DeadCodeElimination dead_code_elimination(&graph_reducer, data->graph(),
+                                              data->common());
+    CommonOperatorReducer common_reducer(&graph_reducer, data->graph(),
+                                         data->common(), data->machine());
+    AddReducer(data, &graph_reducer, &dead_code_elimination);
+    AddReducer(data, &graph_reducer, &common_reducer);
+    graph_reducer.ReduceGraph();
+  }
+};
+
 struct StoreStoreEliminationPhase {
-  static const char* phase_name() { return "Store-store elimination"; }
+  static const char* phase_name() { return "store-store elimination"; }
 
   void Run(PipelineData* data, Zone* temp_zone) {
-    StoreStoreElimination store_store_elimination(data->jsgraph(), temp_zone);
-    store_store_elimination.Run();
+    GraphTrimmer trimmer(temp_zone, data->graph());
+    NodeVector roots(temp_zone);
+    data->jsgraph()->GetCachedNodes(&roots);
+    trimmer.TrimGraph(roots.begin(), roots.end());
+
+    StoreStoreElimination::Run(data->jsgraph(), temp_zone);
   }
 };
 
@@ -1554,6 +1578,9 @@ bool PipelineImpl::OptimizeGraph(Linkage* linkage) {
   Run<EffectControlLinearizationPhase>();
   RunPrintAndVerify("Effect and control linearized", true);
 
+  Run<DeadCodeEliminationPhase>();
+  RunPrintAndVerify("Common operator reducer", true);
+
   if (FLAG_turbo_store_elimination) {
     Run<StoreStoreEliminationPhase>();
     RunPrintAndVerify("Store-store elimination", true);

src/compiler/store-store-elimination.h

@@ -6,31 +6,16 @@
 #define V8_COMPILER_STORE_STORE_ELIMINATION_H_
 
 #include "src/compiler/common-operator.h"
-#include "src/compiler/graph-reducer.h"
 #include "src/compiler/js-graph.h"
+#include "src/zone-containers.h"
 
 namespace v8 {
 namespace internal {
 namespace compiler {
 
-// Forward declarations.
-class CommonOperatorBuilder;
-class JSGraph;
-
 class StoreStoreElimination final {
  public:
-  StoreStoreElimination(JSGraph* js_graph, Zone* temp_zone);
-  ~StoreStoreElimination();
-  void Run();
-
- private:
-  static bool IsEligibleNode(Node* node);
-  void ReduceEligibleNode(Node* node);
-  JSGraph* jsgraph() const { return jsgraph_; }
-  Zone* temp_zone() const { return temp_zone_; }
-
-  JSGraph* const jsgraph_;
-  Zone* const temp_zone_;
+  static void Run(JSGraph* js_graph, Zone* temp_zone);
 };
 
 }  // namespace compiler

src/flag-definitions.h

@@ -356,6 +356,8 @@ DEFINE_STRING(trace_phase, "HLZ", "trace generated IR for specified phases")
 DEFINE_BOOL(trace_inlining, false, "trace inlining decisions")
 DEFINE_BOOL(trace_load_elimination, false, "trace load elimination")
 DEFINE_BOOL(trace_store_elimination, false, "trace store elimination")
+DEFINE_BOOL(turbo_verify_store_elimination, false,
+            "verify store elimination more rigorously")
 DEFINE_BOOL(trace_alloc, false, "trace register allocator")
 DEFINE_BOOL(trace_all_uses, false, "trace all use positions")
 DEFINE_BOOL(trace_range, false, "trace range analysis")
@@ -425,7 +427,6 @@ DEFINE_BOOL(omit_map_checks_for_leaf_maps, true,
 // Flags for TurboFan.
 DEFINE_BOOL(turbo, false, "enable TurboFan compiler")
 DEFINE_IMPLICATION(turbo, turbo_asm_deoptimization)
-DEFINE_IMPLICATION(turbo, turbo_store_elimination)
 DEFINE_IMPLICATION(turbo, turbo_loop_peeling)
 DEFINE_BOOL(turbo_from_bytecode, false, "enable building graphs from bytecode")
 DEFINE_BOOL(turbo_sp_frame_access, false,
@@ -487,7 +488,7 @@ DEFINE_BOOL(turbo_instruction_scheduling, false,
             "enable instruction scheduling in TurboFan")
 DEFINE_BOOL(turbo_stress_instruction_scheduling, false,
             "randomly schedule instructions to stress dependency tracking")
-DEFINE_BOOL(turbo_store_elimination, false,
+DEFINE_BOOL(turbo_store_elimination, true,
             "enable store-store elimination in TurboFan")
 
 // Flags to help platform porters