[wasm-gc] Polymorphic inlining for call_ref

When call_ref has seen more than one call target, we now support
inlining all of them (constrained by budget/heuristics).

Bug: v8:7748,v8:12166
Change-Id: Iae16e74da1bad5e7a117f70efb6c61b3f39f832c
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/3650607Reviewed-by: Manos Koukoutos <manoskouk@chromium.org>
Commit-Queue: Jakob Kummerow <jkummerow@chromium.org>
Cr-Commit-Position: refs/heads/main@{#80602}

src/builtins/wasm.tq

@@ -52,6 +52,8 @@ const kAnyTableType: constexpr int31
     generates 'wasm::kWasmAnyRef.raw_bit_field()';
 const kAnyNonNullTableType: constexpr int31
     generates 'wasm::kWasmAnyNonNullableRef.raw_bit_field()';
+const kMaxPolymorphism:
+    constexpr int31 generates 'wasm::kMaxPolymorphism';
 
 extern macro WasmBuiltinsAssembler::LoadInstanceFromFrame(): WasmInstanceObject;
 
@@ -496,7 +498,7 @@ macro GetTargetAndInstance(funcref: WasmInternalFunction): TargetAndInstance {
 // - monomorphic: (funcref, count (smi)). The second slot is a counter for how
 //   often the funcref in the first slot has been seen.
 // - polymorphic: (fixed_array, <unused>). In this case, the array
-//   contains 2..4 pairs (funcref, count (smi)) (like monomorphic data).
+//   contains 2..kMaxPolymorphism pairs (funcref, count (smi))
 // - megamorphic: ("megamorphic" sentinel, <unused>)
 //
 // TODO(rstz): The counter might overflow if it exceeds the range of a Smi.
@@ -533,7 +535,8 @@ builtin CallRefIC(
   } else if (Is<FixedArray>(value)) {
     // Polymorphic miss.
     const entries = UnsafeCast<FixedArray>(value);
-    if (entries.length == SmiConstant(8)) {  // 4 entries, 2 slots each.
+    const kMaxSlots = kMaxPolymorphism * 2;  // 2 slots per entry.
+    if (entries.length == SmiConstant(kMaxSlots)) {
       vector.objects[index] = ic::kMegamorphicSymbol;
       vector.objects[index + 1] = ic::kMegamorphicSymbol;
     } else {

src/compiler/wasm-compiler.cc

@@ -2962,9 +2962,11 @@ Node* WasmGraphBuilder::BuildCallRef(const wasm::FunctionSig* real_sig,
   return call;
 }
 
-void WasmGraphBuilder::CompareToInternalFunctionAtIndex(
-    Node* func_ref, uint32_t function_index, Node** success_control,
-    Node** failure_control) {
+void WasmGraphBuilder::CompareToInternalFunctionAtIndex(Node* func_ref,
+                                                        uint32_t function_index,
+                                                        Node** success_control,
+                                                        Node** failure_control,
+                                                        bool is_last_case) {
   // Since we are comparing to a function reference, it is guaranteed that
   // instance->wasm_internal_functions() has been initialized.
   Node* internal_functions = gasm_->LoadImmutable(
@@ -2974,8 +2976,9 @@ void WasmGraphBuilder::CompareToInternalFunctionAtIndex(
   Node* function_ref_at_index = gasm_->LoadFixedArrayElement(
       internal_functions, gasm_->IntPtrConstant(function_index),
       MachineType::AnyTagged());
+  BranchHint hint = is_last_case ? BranchHint::kTrue : BranchHint::kNone;
   gasm_->Branch(gasm_->TaggedEqual(function_ref_at_index, func_ref),
-                success_control, failure_control, BranchHint::kTrue);
+                success_control, failure_control, hint);
 }
 
 Node* WasmGraphBuilder::CallRef(const wasm::FunctionSig* real_sig,

src/compiler/wasm-compiler.h

@@ -371,7 +371,8 @@ class WasmGraphBuilder {
 
   void CompareToInternalFunctionAtIndex(Node* func_ref, uint32_t function_index,
                                         Node** success_control,
-                                        Node** failure_control);
+                                        Node** failure_control,
+                                        bool is_last_case);
 
   void BrOnNull(Node* ref_object, Node** non_null_node, Node** null_node);
 

src/compiler/wasm-inlining.cc

@@ -94,11 +94,12 @@ Reduction WasmInliner::ReduceCall(Node* call) {
 }
 
 bool SmallEnoughToInline(size_t current_graph_size, uint32_t candidate_size) {
-  if (WasmInliner::graph_size_allows_inlining(current_graph_size)) {
+  if (WasmInliner::graph_size_allows_inlining(current_graph_size +
+                                              candidate_size)) {
     return true;
   }
   // For truly tiny functions, let's be a bit more generous.
-  return candidate_size < 10 &&
+  return candidate_size <= 12 &&
          WasmInliner::graph_size_allows_inlining(current_graph_size - 100);
 }
 

src/compiler/wasm-inlining.h

@@ -58,8 +58,8 @@ class WasmInliner final : public AdvancedReducer {
   Reduction Reduce(Node* node) final;
   void Finalize() final;
 
-  static bool graph_size_allows_inlining(size_t initial_graph_size) {
-    return initial_graph_size < 5000;
+  static bool graph_size_allows_inlining(size_t graph_size) {
+    return graph_size < FLAG_wasm_inlining_budget;
   }
 
  private:

src/flags/flag-definitions.h

@@ -1108,11 +1108,8 @@ DEFINE_BOOL(wasm_math_intrinsics, true,
 DEFINE_BOOL(
     wasm_inlining, false,
     "enable inlining of wasm functions into wasm functions (experimental)")
-DEFINE_SIZE_T(
-    wasm_inlining_budget_factor, 75000,
-    "maximum allowed size to inline a function is given by {n / caller size}")
-DEFINE_SIZE_T(wasm_inlining_max_size, 1000,
-              "maximum size of a function that can be inlined, in TF nodes")
+DEFINE_SIZE_T(wasm_inlining_budget, 9000,
+              "maximum graph size (in TF nodes) that allows inlining more")
 DEFINE_BOOL(wasm_speculative_inlining, false,
             "enable speculative inlining of call_ref targets (experimental)")
 DEFINE_BOOL(trace_wasm_inlining, false, "trace wasm inlining")

src/wasm/graph-builder-interface.cc

@@ -127,8 +127,15 @@ class WasmGraphBuildingInterface {
       base::MutexGuard mutex_guard(&feedbacks.mutex);
       auto feedback = feedbacks.feedback_for_function.find(func_index_);
       if (feedback != feedbacks.feedback_for_function.end()) {
+        // This creates a copy of the vector, which is cheaper than holding on
+        // to the mutex throughout graph building.
         type_feedback_ = feedback->second.feedback_vector;
-        builder_->ReserveCallCounts(type_feedback_.size());
+        // Preallocate space for storing call counts to save Zone memory.
+        int total_calls = 0;
+        for (size_t i = 0; i < type_feedback_.size(); i++) {
+          total_calls += type_feedback_[i].num_cases();
+        }
+        builder_->ReserveCallCounts(static_cast<size_t>(total_calls));
         // We need to keep the feedback in the module to inline later. However,
         // this means we are stuck with it forever.
         // TODO(jkummerow): Reconsider our options here.
@@ -644,7 +651,8 @@ class WasmGraphBuildingInterface {
     int maybe_call_count = -1;
     if (FLAG_wasm_speculative_inlining && type_feedback_.size() > 0) {
       const CallSiteFeedback& feedback = next_call_feedback();
-      maybe_call_count = feedback.absolute_call_frequency;
+      DCHECK_EQ(feedback.num_cases(), 1);
+      maybe_call_count = feedback.call_count(0);
     }
     DoCall(decoder, CallInfo::CallDirect(imm.index, maybe_call_count), imm.sig,
            args, returns);
@@ -656,7 +664,8 @@ class WasmGraphBuildingInterface {
     int maybe_call_count = -1;
     if (FLAG_wasm_speculative_inlining && type_feedback_.size() > 0) {
       const CallSiteFeedback& feedback = next_call_feedback();
-      maybe_call_count = feedback.absolute_call_frequency;
+      DCHECK_EQ(feedback.num_cases(), 1);
+      maybe_call_count = feedback.call_count(0);
     }
     DoReturnCall(decoder, CallInfo::CallDirect(imm.index, maybe_call_count),
                  imm.sig, args);
@@ -683,14 +692,11 @@ class WasmGraphBuildingInterface {
   void CallRef(FullDecoder* decoder, const Value& func_ref,
                const FunctionSig* sig, uint32_t sig_index, const Value args[],
                Value returns[]) {
-    int maybe_feedback = -1;
-    int maybe_call_count = -1;
+    const CallSiteFeedback* feedback = nullptr;
     if (FLAG_wasm_speculative_inlining && type_feedback_.size() > 0) {
-      const CallSiteFeedback& feedback = next_call_feedback();
-      maybe_feedback = feedback.function_index;
-      maybe_call_count = feedback.absolute_call_frequency;
+      feedback = &next_call_feedback();
     }
-    if (maybe_feedback == -1) {
+    if (feedback == nullptr || feedback->num_cases() == 0) {
       DoCall(decoder, CallInfo::CallRef(func_ref, NullCheckFor(func_ref.type)),
              sig, args, returns);
       return;
@@ -698,11 +704,18 @@ class WasmGraphBuildingInterface {
 
     // Check for equality against a function at a specific index, and if
     // successful, just emit a direct call.
-    DCHECK_GE(maybe_feedback, 0);
-    const uint32_t expected_function_index = maybe_feedback;
+    int num_cases = feedback->num_cases();
+    std::vector<TFNode*> control_args;
+    std::vector<TFNode*> effect_args;
+    std::vector<Value*> returns_values;
+    control_args.reserve(num_cases + 1);
+    effect_args.reserve(num_cases + 2);
+    returns_values.reserve(num_cases);
+    for (int i = 0; i < num_cases; i++) {
+      const uint32_t expected_function_index = feedback->function_index(i);
 
       if (FLAG_trace_wasm_speculative_inlining) {
-      PrintF("[Function #%d call #%d: graph support for inlining target #%d]\n",
+        PrintF("[Function #%d call #%d: graph support for inlining #%d]\n",
                func_index_, feedback_instruction_index_ - 1,
                expected_function_index);
       }
@@ -711,7 +724,7 @@ class WasmGraphBuildingInterface {
       TFNode* failure_control;
       builder_->CompareToInternalFunctionAtIndex(
           func_ref.node, expected_function_index, &success_control,
-        &failure_control);
+          &failure_control, i == num_cases - 1);
       TFNode* initial_effect = effect();
 
       builder_->SetControl(success_control);
@@ -719,49 +732,52 @@ class WasmGraphBuildingInterface {
       Value* returns_direct =
           decoder->zone()->NewArray<Value>(sig->return_count());
       DoCall(decoder,
-           CallInfo::CallDirect(expected_function_index, maybe_call_count),
+             CallInfo::CallDirect(expected_function_index,
+                                  feedback->call_count(i)),
              decoder->module_->signature(sig_index), args, returns_direct);
-    TFNode* control_direct = control();
-    TFNode* effect_direct = effect();
+      control_args.push_back(control());
+      effect_args.push_back(effect());
+      returns_values.push_back(returns_direct);
 
       builder_->SetEffectControl(initial_effect, failure_control);
       ssa_env_->effect = initial_effect;
       ssa_env_->control = failure_control;
+    }
     Value* returns_ref = decoder->zone()->NewArray<Value>(sig->return_count());
     DoCall(decoder, CallInfo::CallRef(func_ref, NullCheckFor(func_ref.type)),
            sig, args, returns_ref);
 
-    TFNode* control_ref = control();
-    TFNode* effect_ref = effect();
+    control_args.push_back(control());
+    TFNode* control = builder_->Merge(num_cases + 1, control_args.data());
 
-    TFNode* control_args[] = {control_direct, control_ref};
-    TFNode* control = builder_->Merge(2, control_args);
-
-    TFNode* effect_args[] = {effect_direct, effect_ref, control};
-    TFNode* effect = builder_->EffectPhi(2, effect_args);
+    effect_args.push_back(effect());
+    effect_args.push_back(control);
+    TFNode* effect = builder_->EffectPhi(num_cases + 1, effect_args.data());
 
     ssa_env_->control = control;
     ssa_env_->effect = effect;
     builder_->SetEffectControl(effect, control);
 
     for (uint32_t i = 0; i < sig->return_count(); i++) {
-      TFNode* phi_args[] = {returns_direct[i].node, returns_ref[i].node,
-                            control};
-      returns[i].node = builder_->Phi(sig->GetReturn(i), 2, phi_args);
+      std::vector<TFNode*> phi_args;
+      for (int j = 0; j < num_cases; j++) {
+        phi_args.push_back(returns_values[j][i].node);
+      }
+      phi_args.push_back(returns_ref[i].node);
+      phi_args.push_back(control);
+      returns[i].node =
+          builder_->Phi(sig->GetReturn(i), num_cases + 1, phi_args.data());
     }
   }
 
   void ReturnCallRef(FullDecoder* decoder, const Value& func_ref,
                      const FunctionSig* sig, uint32_t sig_index,
                      const Value args[]) {
-    int maybe_feedback = -1;
-    int maybe_call_count = -1;
+    const CallSiteFeedback* feedback = nullptr;
     if (FLAG_wasm_speculative_inlining && type_feedback_.size() > 0) {
-      const CallSiteFeedback& feedback = next_call_feedback();
-      maybe_feedback = feedback.function_index;
-      maybe_call_count = feedback.absolute_call_frequency;
+      feedback = &next_call_feedback();
     }
-    if (maybe_feedback == -1) {
+    if (feedback == nullptr || feedback->num_cases() == 0) {
       DoReturnCall(decoder,
                    CallInfo::CallRef(func_ref, NullCheckFor(func_ref.type)),
                    sig, args);
@@ -770,11 +786,12 @@ class WasmGraphBuildingInterface {
 
     // Check for equality against a function at a specific index, and if
     // successful, just emit a direct call.
-    DCHECK_GE(maybe_feedback, 0);
-    const uint32_t expected_function_index = maybe_feedback;
+    int num_cases = feedback->num_cases();
+    for (int i = 0; i < num_cases; i++) {
+      const uint32_t expected_function_index = feedback->function_index(i);
 
       if (FLAG_trace_wasm_speculative_inlining) {
-      PrintF("[Function #%d call #%d: graph support for inlining target #%d]\n",
+        PrintF("[Function #%d call #%d: graph support for inlining #%d]\n",
                func_index_, feedback_instruction_index_ - 1,
                expected_function_index);
       }
@@ -783,19 +800,21 @@ class WasmGraphBuildingInterface {
       TFNode* failure_control;
       builder_->CompareToInternalFunctionAtIndex(
           func_ref.node, expected_function_index, &success_control,
-        &failure_control);
+          &failure_control, i == num_cases - 1);
       TFNode* initial_effect = effect();
 
       builder_->SetControl(success_control);
       ssa_env_->control = success_control;
-    DoReturnCall(
-        decoder,
-        CallInfo::CallDirect(expected_function_index, maybe_call_count), sig,
-        args);
+      DoReturnCall(decoder,
+                   CallInfo::CallDirect(expected_function_index,
+                                        feedback->call_count(i)),
+                   sig, args);
 
       builder_->SetEffectControl(initial_effect, failure_control);
       ssa_env_->effect = initial_effect;
       ssa_env_->control = failure_control;
+    }
+
     DoReturnCall(decoder,
                  CallInfo::CallRef(func_ref, NullCheckFor(func_ref.type)), sig,
                  args);
@@ -1464,7 +1483,7 @@ class WasmGraphBuildingInterface {
   std::vector<compiler::WasmLoopInfo> loop_infos_;
   InlinedStatus inlined_status_;
   // The entries in {type_feedback_} are indexed by the position of feedback-
-  // consuming instructions (currently only call_ref).
+  // consuming instructions (currently only calls).
   int feedback_instruction_index_ = 0;
   std::vector<CallSiteFeedback> type_feedback_;
 

src/wasm/module-compiler.cc

@@ -1284,13 +1284,13 @@ class TransitiveTypeFeedbackProcessor {
  private:
   void Process(int func_index);
 
-  void EnqueueCallees(std::vector<CallSiteFeedback> feedback) {
+  void EnqueueCallees(const std::vector<CallSiteFeedback>& feedback) {
     for (size_t i = 0; i < feedback.size(); i++) {
-      int func = feedback[i].function_index;
-      // Nothing to do for non-inlineable (e.g. megamorphic) calls.
-      if (func == -1) continue;
+      const CallSiteFeedback& csf = feedback[i];
+      for (int j = 0; j < csf.num_cases(); j++) {
+        int func = csf.function_index(j);
         // Don't spend time on calls that have never been executed.
-      if (feedback[i].absolute_call_frequency == 0) continue;
+        if (csf.call_count(j) == 0) continue;
         // Don't recompute feedback that has already been processed.
         auto existing = feedback_for_function_.find(func);
         if (existing != feedback_for_function_.end() &&
@@ -1300,117 +1300,134 @@ class TransitiveTypeFeedbackProcessor {
         queue_.insert(func);
       }
     }
+  }
 
   Handle<WasmInstanceObject> instance_;
   std::map<uint32_t, FunctionTypeFeedback>& feedback_for_function_;
   std::unordered_set<int> queue_;
 };
 
+class FeedbackMaker {
+ public:
+  FeedbackMaker(WasmInstanceObject instance, int func_index_, int num_calls)
+      : instance_(instance),
+        num_imported_functions_(
+            static_cast<int>(instance.module()->num_imported_functions)),
+        targets_cache_(kMaxPolymorphism),
+        counts_cache_(kMaxPolymorphism) {
+    result_.reserve(num_calls);
+  }
+
+  void AddCandidate(Object maybe_function, int count) {
+    if (!maybe_function.IsWasmInternalFunction()) return;
+    WasmInternalFunction function = WasmInternalFunction::cast(maybe_function);
+    if (!WasmExportedFunction::IsWasmExportedFunction(function.external())) {
+      return;
+    }
+    WasmExportedFunction target =
+        WasmExportedFunction::cast(function.external());
+    if (target.instance() != instance_) return;
+    if (target.function_index() < num_imported_functions_) return;
+    AddCall(target.function_index(), count);
+  }
+
+  void AddCall(int target, int count) {
+    // Keep the cache sorted (using insertion-sort), highest count first.
+    int insertion_index = 0;
+    while (insertion_index < cache_usage_ &&
+           counts_cache_[insertion_index] >= count) {
+      insertion_index++;
+    }
+    for (int shifted_index = cache_usage_ - 1; shifted_index >= insertion_index;
+         shifted_index--) {
+      targets_cache_[shifted_index + 1] = targets_cache_[shifted_index];
+      counts_cache_[shifted_index + 1] = counts_cache_[shifted_index];
+    }
+    targets_cache_[insertion_index] = target;
+    counts_cache_[insertion_index] = count;
+    cache_usage_++;
+  }
+
+  void FinalizeCall() {
+    if (cache_usage_ == 0) {
+      result_.emplace_back();
+    } else if (cache_usage_ == 1) {
+      if (FLAG_trace_wasm_speculative_inlining) {
+        PrintF("[Function #%d call_ref #%zu inlineable (monomorphic)]\n",
+               func_index_, result_.size());
+      }
+      result_.emplace_back(targets_cache_[0], counts_cache_[0]);
+    } else {
+      if (FLAG_trace_wasm_speculative_inlining) {
+        PrintF("[Function #%d call_ref #%zu inlineable (polymorphic %d)]\n",
+               func_index_, result_.size(), cache_usage_);
+      }
+      CallSiteFeedback::PolymorphicCase* polymorphic =
+          new CallSiteFeedback::PolymorphicCase[cache_usage_];
+      for (int i = 0; i < cache_usage_; i++) {
+        polymorphic[i].function_index = targets_cache_[i];
+        polymorphic[i].absolute_call_frequency = counts_cache_[i];
+      }
+      result_.emplace_back(polymorphic, cache_usage_);
+    }
+    cache_usage_ = 0;
+  }
+
+  std::vector<CallSiteFeedback>&& GetResult() { return std::move(result_); }
+
+ private:
+  WasmInstanceObject instance_;
+  std::vector<CallSiteFeedback> result_;
+  int num_imported_functions_;
+  int func_index_;
+  int cache_usage_{0};
+  std::vector<int> targets_cache_;
+  std::vector<int> counts_cache_;
+};
+
 void TransitiveTypeFeedbackProcessor::Process(int func_index) {
   int which_vector = declared_function_index(instance_->module(), func_index);
   Object maybe_feedback = instance_->feedback_vectors().get(which_vector);
   if (!maybe_feedback.IsFixedArray()) return;
   FixedArray feedback = FixedArray::cast(maybe_feedback);
-  std::vector<CallSiteFeedback> result(feedback.length() / 2);
-  int imported_functions =
-      static_cast<int>(instance_->module()->num_imported_functions);
   WasmModuleObject module_object = instance_->module_object();
   const NativeModule* native_module = module_object.native_module();
   const WasmModule* module = native_module->module();
   const std::vector<uint32_t>& call_direct_targets(
       module->type_feedback.feedback_for_function[func_index].call_targets);
+  FeedbackMaker fm(*instance_, func_index, feedback.length() / 2);
   for (int i = 0; i < feedback.length(); i += 2) {
     Object value = feedback.get(i);
-    if (value.IsWasmInternalFunction() &&
-        WasmExportedFunction::IsWasmExportedFunction(
-            WasmInternalFunction::cast(value).external())) {
-      // Monomorphic, and the internal function points to a wasm-generated
-      // external function (WasmExportedFunction). Mark the target for inlining
-      // if it's defined in the same module.
-      WasmExportedFunction target = WasmExportedFunction::cast(
-          WasmInternalFunction::cast(value).external());
-      if (target.instance() == *instance_ &&
-          target.function_index() >= imported_functions) {
-        if (FLAG_trace_wasm_speculative_inlining) {
-          PrintF("[Function #%d call_ref #%d inlineable (monomorphic)]\n",
-                 func_index, i / 2);
-        }
-        int32_t count = Smi::cast(feedback.get(i + 1)).value();
-        result[i / 2] = {target.function_index(), count};
-        continue;
-      }
+    if (value.IsWasmInternalFunction()) {
+      // Monomorphic.
+      int count = Smi::cast(feedback.get(i + 1)).value();
+      fm.AddCandidate(value, count);
     } else if (value.IsFixedArray()) {
-      // Polymorphic. Pick a target for inlining if there is one that was
-      // seen for most calls, and matches the requirements of the monomorphic
-      // case.
+      // Polymorphic.
       FixedArray polymorphic = FixedArray::cast(value);
-      size_t total_count = 0;
       for (int j = 0; j < polymorphic.length(); j += 2) {
-        total_count += Smi::cast(polymorphic.get(j + 1)).value();
-      }
-      int found_target = -1;
-      int found_count = -1;
-      double best_frequency = 0;
-      for (int j = 0; j < polymorphic.length(); j += 2) {
-        int32_t this_count = Smi::cast(polymorphic.get(j + 1)).value();
-        double frequency = static_cast<double>(this_count) / total_count;
-        if (frequency > best_frequency) best_frequency = frequency;
-        if (frequency < 0.8) continue;
-
-        // We reject this polymorphic entry if:
-        // - it is not defined,
-        // - it is not a wasm-defined function (WasmExportedFunction)
-        // - it was not defined in this module.
-        if (!polymorphic.get(j).IsWasmInternalFunction()) continue;
-        WasmInternalFunction internal =
-            WasmInternalFunction::cast(polymorphic.get(j));
-        if (!WasmExportedFunction::IsWasmExportedFunction(
-                internal.external())) {
-          continue;
-        }
-        WasmExportedFunction target =
-            WasmExportedFunction::cast(internal.external());
-        if (target.instance() != *instance_ ||
-            target.function_index() < imported_functions) {
-          continue;
-        }
-
-        found_target = target.function_index();
-        found_count = static_cast<int>(this_count);
-        if (FLAG_trace_wasm_speculative_inlining) {
-          PrintF("[Function #%d call_ref #%d inlineable (polymorphic %f)]\n",
-                 func_index, i / 2, frequency);
-        }
-        break;
-      }
-      if (found_target >= 0) {
-        result[i / 2] = {found_target, found_count};
-        continue;
-      } else if (FLAG_trace_wasm_speculative_inlining) {
-        PrintF("[Function #%d call_ref #%d: best frequency %f]\n", func_index,
-               i / 2, best_frequency);
+        Object function = polymorphic.get(j);
+        int count = Smi::cast(polymorphic.get(j + 1)).value();
+        fm.AddCandidate(function, count);
       }
     } else if (value.IsSmi()) {
       // Uninitialized, or a direct call collecting call count.
       uint32_t target = call_direct_targets[i / 2];
       if (target != FunctionTypeFeedback::kNonDirectCall) {
         int count = Smi::cast(value).value();
-        if (FLAG_trace_wasm_speculative_inlining) {
-          PrintF("[Function #%d call_direct #%d: frequency %d]\n", func_index,
-                 i / 2, count);
-        }
-        result[i / 2] = {static_cast<int>(target), count};
-        continue;
+        fm.AddCall(static_cast<int>(target), count);
+      } else if (FLAG_trace_wasm_speculative_inlining) {
+        PrintF("[Function #%d call #%d: uninitialized]\n", func_index, i / 2);
       }
+    } else if (FLAG_trace_wasm_speculative_inlining) {
+      if (value ==
+          ReadOnlyRoots(instance_->GetIsolate()).megamorphic_symbol()) {
+        PrintF("[Function #%d call #%d: megamorphic]\n", func_index, i / 2);
       }
-    // If we fall through to here, then this call isn't eligible for inlining.
-    // Possible reasons: uninitialized or megamorphic feedback; or monomorphic
-    // or polymorphic that didn't meet our requirements.
-    if (FLAG_trace_wasm_speculative_inlining) {
-      PrintF("[Function #%d call #%d *not* inlineable]\n", func_index, i / 2);
     }
-    result[i / 2] = {-1, -1};
+    fm.FinalizeCall();
   }
+  std::vector<CallSiteFeedback> result(fm.GetResult());
   EnqueueCallees(result);
   feedback_for_function_[func_index].feedback_vector = std::move(result);
 }

src/wasm/wasm-constants.h

@@ -186,6 +186,9 @@ constexpr uint32_t kGenericWrapperBudget = 1000;
 // gives up some module size for faster access to the supertypes.
 constexpr uint32_t kMinimumSupertypeArraySize = 3;
 
+// Maximum number of call targets tracked per call.
+constexpr int kMaxPolymorphism = 4;
+
 #if V8_TARGET_ARCH_X64
 constexpr int32_t kOSRTargetOffset = 5 * kSystemPointerSize;
 #endif

src/wasm/wasm-module.h

@@ -399,9 +399,79 @@ struct V8_EXPORT_PRIVATE WasmDebugSymbols {
   WireBytesRef external_url;
 };
 
-struct CallSiteFeedback {
+class CallSiteFeedback {
+ public:
+  struct PolymorphicCase {
     int function_index;
     int absolute_call_frequency;
+  };
+
+  // Regular constructor: uninitialized/unknown, monomorphic, or polymorphic.
+  CallSiteFeedback() : index_or_count_(-1), frequency_or_ool_(0) {}
+  CallSiteFeedback(int function_index, int call_count)
+      : index_or_count_(function_index), frequency_or_ool_(call_count) {}
+  CallSiteFeedback(PolymorphicCase* polymorphic_cases, int num_cases)
+      : index_or_count_(-num_cases),
+        frequency_or_ool_(reinterpret_cast<intptr_t>(polymorphic_cases)) {}
+
+  // Copying and assignment: prefer moving, as it's cheaper.
+  // The code below makes sure external polymorphic storage is copied and/or
+  // freed as appropriate.
+  CallSiteFeedback(const CallSiteFeedback& other) V8_NOEXCEPT { *this = other; }
+  CallSiteFeedback(CallSiteFeedback&& other) V8_NOEXCEPT { *this = other; }
+  CallSiteFeedback& operator=(const CallSiteFeedback& other) V8_NOEXCEPT {
+    index_or_count_ = other.index_or_count_;
+    if (other.is_polymorphic()) {
+      int num_cases = other.num_cases();
+      PolymorphicCase* polymorphic = new PolymorphicCase[num_cases];
+      for (int i = 0; i < num_cases; i++) {
+        polymorphic[i].function_index = other.function_index(i);
+        polymorphic[i].absolute_call_frequency = other.call_count(i);
+      }
+      frequency_or_ool_ = reinterpret_cast<intptr_t>(polymorphic);
+    } else {
+      frequency_or_ool_ = other.frequency_or_ool_;
+    }
+    return *this;
+  }
+  CallSiteFeedback& operator=(CallSiteFeedback&& other) V8_NOEXCEPT {
+    if (this != &other) {
+      index_or_count_ = other.index_or_count_;
+      frequency_or_ool_ = other.frequency_or_ool_;
+      other.frequency_or_ool_ = 0;
+    }
+    return *this;
+  }
+
+  ~CallSiteFeedback() {
+    if (is_polymorphic()) delete[] polymorphic_storage();
+  }
+
+  int num_cases() const {
+    if (is_monomorphic()) return 1;
+    if (is_invalid()) return 0;
+    return -index_or_count_;
+  }
+  int function_index(int i) const {
+    DCHECK(!is_invalid());
+    if (is_monomorphic()) return index_or_count_;
+    return polymorphic_storage()[i].function_index;
+  }
+  int call_count(int i) const {
+    if (index_or_count_ >= 0) return static_cast<int>(frequency_or_ool_);
+    return polymorphic_storage()[i].absolute_call_frequency;
+  }
+
+ private:
+  bool is_monomorphic() const { return index_or_count_ >= 0; }
+  bool is_polymorphic() const { return index_or_count_ <= -2; }
+  bool is_invalid() const { return index_or_count_ == -1; }
+  const PolymorphicCase* polymorphic_storage() const {
+    return reinterpret_cast<PolymorphicCase*>(frequency_or_ool_);
+  }
+
+  int index_or_count_;
+  intptr_t frequency_or_ool_;
 };
 struct FunctionTypeFeedback {
   std::vector<CallSiteFeedback> feedback_vector;