Change CollectCallFeedback to take the receiver as LazyNode

This CL makes a call to CollectCallFeedback from Ignition load the receiver lazily, which may improve the performance of Octane/{Richards, Delta-Blue} with --no-opt for 1-3%.

Bug: chromium:1207349
Change-Id: I5fb09d5c5662ef8714acf71dd7341d3164b44f93
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2895358Reviewed-by: Ross McIlroy <rmcilroy@chromium.org>
Reviewed-by: Georg Neis <neis@chromium.org>
Commit-Queue: Fanchen Kong <fanchen.kong@intel.com>
Cr-Commit-Position: refs/heads/master@{#74580}

src/builtins/base.tq

@@ -141,6 +141,7 @@ intrinsic %MakeLazy<T: type, A1: type, A2: type, A3: type>(
 // template, but Torque doesn't understand how to use templates for extern
 // macros, so just add whatever overload definitions you need here.
 extern macro RunLazy(Lazy<Smi>): Smi;
+extern macro RunLazy(Lazy<JSAny>): JSAny;
 
 // A Smi value containing a bitfield struct as its integer data.
 @useParentTypeChecker type SmiTagged<T : type extends uint31> extends Smi;

src/builtins/builtins-call-gen.cc

@@ -113,7 +113,8 @@ TF_BUILTIN(Call_ReceiverIsNullOrUndefined_WithFeedback,
   auto feedback_vector = Parameter<FeedbackVector>(Descriptor::kFeedbackVector);
   auto slot = UncheckedParameter<UintPtrT>(Descriptor::kSlot);
   auto receiver = Parameter<Object>(Descriptor::kReceiver);
-  CollectCallFeedback(target, receiver, context, feedback_vector, slot);
+  CollectCallFeedback(
+      target, [=] { return receiver; }, context, feedback_vector, slot);
   TailCallBuiltin(Builtins::kCall_ReceiverIsNullOrUndefined, context, target,
                   argc);
 }
@@ -126,7 +127,8 @@ TF_BUILTIN(Call_ReceiverIsNotNullOrUndefined_WithFeedback,
   auto feedback_vector = Parameter<FeedbackVector>(Descriptor::kFeedbackVector);
   auto slot = UncheckedParameter<UintPtrT>(Descriptor::kSlot);
   auto receiver = Parameter<Object>(Descriptor::kReceiver);
-  CollectCallFeedback(target, receiver, context, feedback_vector, slot);
+  CollectCallFeedback(
+      target, [=] { return receiver; }, context, feedback_vector, slot);
   TailCallBuiltin(Builtins::kCall_ReceiverIsNotNullOrUndefined, context, target,
                   argc);
 }
@@ -138,7 +140,8 @@ TF_BUILTIN(Call_ReceiverIsAny_WithFeedback, CallOrConstructBuiltinsAssembler) {
   auto feedback_vector = Parameter<FeedbackVector>(Descriptor::kFeedbackVector);
   auto slot = UncheckedParameter<UintPtrT>(Descriptor::kSlot);
   auto receiver = Parameter<Object>(Descriptor::kReceiver);
-  CollectCallFeedback(target, receiver, context, feedback_vector, slot);
+  CollectCallFeedback(
+      target, [=] { return receiver; }, context, feedback_vector, slot);
   TailCallBuiltin(Builtins::kCall_ReceiverIsAny, context, target, argc);
 }
 
@@ -483,13 +486,15 @@ void CallOrConstructBuiltinsAssembler::CallReceiver(
   auto target = Parameter<Object>(Descriptor::kFunction);
   auto context = LoadContextFromBaseline();
   auto feedback_vector = LoadFeedbackVectorFromBaseline();
-  TNode<Object> receiver;
+  LazyNode<Object> receiver = [=] {
     if (maybe_receiver) {
-    receiver = *maybe_receiver;
+      return *maybe_receiver;
     } else {
       CodeStubArguments args(this, argc);
-    receiver = args.GetReceiver();
+      return args.GetReceiver();
     }
+  };
+
   CollectCallFeedback(target, receiver, context, feedback_vector, slot);
   TailCallBuiltin(id, context, target, argc);
 }
@@ -510,7 +515,8 @@ TF_BUILTIN(CallWithArrayLike_WithFeedback, CallOrConstructBuiltinsAssembler) {
   auto feedback_vector = Parameter<FeedbackVector>(Descriptor::kFeedbackVector);
   auto slot = UncheckedParameter<UintPtrT>(Descriptor::kSlot);
   auto receiver = Parameter<Object>(Descriptor::kReceiver);
-  CollectCallFeedback(target, receiver, context, feedback_vector, slot);
+  CollectCallFeedback(
+      target, [=] { return receiver; }, context, feedback_vector, slot);
   CallOrConstructWithArrayLike(target, new_target, arguments_list, context);
 }
 
@@ -532,8 +538,9 @@ TF_BUILTIN(CallWithSpread_Baseline, CallOrConstructBuiltinsAssembler) {
   auto feedback_vector = LoadFeedbackVectorFromBaseline();
   auto slot = UncheckedParameter<UintPtrT>(Descriptor::kSlot);
   CodeStubArguments args(this, args_count);
-  auto receiver = args.GetReceiver();
-  CollectCallFeedback(target, receiver, context, feedback_vector, slot);
+  CollectCallFeedback(
+      target, [=] { return args.GetReceiver(); }, context, feedback_vector,
+      slot);
   CallOrConstructWithSpread(target, new_target, spread, args_count, context);
 }
 
@@ -546,7 +553,8 @@ TF_BUILTIN(CallWithSpread_WithFeedback, CallOrConstructBuiltinsAssembler) {
   auto feedback_vector = Parameter<FeedbackVector>(Descriptor::kFeedbackVector);
   auto slot = UncheckedParameter<UintPtrT>(Descriptor::kSlot);
   auto receiver = Parameter<Object>(Descriptor::kReceiver);
-  CollectCallFeedback(target, receiver, context, feedback_vector, slot);
+  CollectCallFeedback(
+      target, [=] { return receiver; }, context, feedback_vector, slot);
   CallOrConstructWithSpread(target, new_target, spread, args_count, context);
 }
 

src/builtins/ic-callable.tq

@@ -89,7 +89,7 @@ macro SetCallFeedbackContent(implicit context: Context)(
 }
 
 macro CollectCallFeedback(
-    maybeTarget: JSAny, maybeReceiver: JSAny, context: Context,
+    maybeTarget: JSAny, maybeReceiver: Lazy<JSAny>, context: Context,
     maybeFeedbackVector: Undefined|FeedbackVector, slotId: uintptr): void {
   // TODO(v8:9891): Remove this assert once all callers are ported to Torque.
   // This assert ensures correctness of maybeFeedbackVector's type which can
@@ -112,7 +112,7 @@ macro CollectCallFeedback(
         TaggedEqualPrototypeApplyFunction(maybeTarget)) {
       // If the Receiver is recorded and the target is Function.prototype.apply,
       // check whether we can stay monomorphic based on the receiver.
-      if (IsMonomorphic(feedback, maybeReceiver)) {
+      if (IsMonomorphic(feedback, RunLazy(maybeReceiver))) {
         return;
       } else {
         // If not, reinitialize the feedback with target.
@@ -149,7 +149,7 @@ macro CollectCallFeedback(
   } label TryInitializeAsMonomorphic {
     let recordedFunction = maybeTarget;
     if (TaggedEqualPrototypeApplyFunction(maybeTarget)) {
-      recordedFunction = maybeReceiver;
+      recordedFunction = RunLazy(maybeReceiver);
       SetCallFeedbackContent(
           feedbackVector, slotId, CallFeedbackContent::kReceiver);
     } else {

src/builtins/ic.tq

@@ -8,7 +8,7 @@ namespace ic {
 
 @export
 macro CollectCallFeedback(
-    maybeTarget: JSAny, maybeReceiver: JSAny, context: Context,
+    maybeTarget: JSAny, maybeReceiver: Lazy<JSAny>, context: Context,
     maybeFeedbackVector: Undefined|FeedbackVector, slotId: uintptr): void {
   callable::CollectCallFeedback(
       maybeTarget, maybeReceiver, context, maybeFeedbackVector, slotId);

src/builtins/iterator.tq

@@ -97,13 +97,18 @@ transitioning builtin CreateAsyncFromSyncIteratorBaseline(syncIterator: JSAny):
   return CreateAsyncFromSyncIterator(context, syncIterator);
 }
 
+macro GetLazyReceiver(receiver: JSAny): JSAny {
+  return receiver;
+}
+
 transitioning builtin CallIteratorWithFeedback(
     context: Context, receiver: JSAny, iteratorMethod: JSAny, callSlot: Smi,
     feedback: Undefined|FeedbackVector): JSAny {
   // TODO(v8:10047): Use TaggedIndex here once TurboFan supports it.
   const callSlotUnTagged: uintptr = Unsigned(SmiUntag(callSlot));
   ic::CollectCallFeedback(
-      iteratorMethod, receiver, context, feedback, callSlotUnTagged);
+      iteratorMethod, %MakeLazy<JSAny, JSAny>('GetLazyReceiver', receiver),
+      context, feedback, callSlotUnTagged);
   const iteratorCallable: Callable = Cast<Callable>(iteratorMethod)
       otherwise ThrowCalledNonCallable(iteratorMethod);
   return Call(context, iteratorCallable, receiver);

src/interpreter/interpreter-assembler.cc

@@ -803,7 +803,7 @@ void InterpreterAssembler::CallJSWithSpreadAndDispatch(
     TNode<UintPtrT> slot_id, TNode<HeapObject> maybe_feedback_vector) {
   DCHECK(Bytecodes::MakesCallAlongCriticalPath(bytecode_));
   DCHECK_EQ(Bytecodes::GetReceiverMode(bytecode_), ConvertReceiverMode::kAny);
-  TNode<Object> receiver = LoadRegisterAtOperandIndex(1);
+  LazyNode<Object> receiver = [=] { return LoadRegisterAtOperandIndex(1); };
   CollectCallFeedback(function, receiver, context, maybe_feedback_vector,
                       slot_id);
   Comment("call using CallWithSpread builtin");

src/interpreter/interpreter-generator.cc

@@ -1365,10 +1365,11 @@ class InterpreterJSCallAssembler : public InterpreterAssembler {
   // Generates code to perform a JS call that collects type feedback.
   void JSCall(ConvertReceiverMode receiver_mode) {
     TNode<Object> function = LoadRegisterAtOperandIndex(0);
-    TNode<Object> receiver =
-        receiver_mode == ConvertReceiverMode::kNullOrUndefined
+    LazyNode<Object> receiver = [=] {
+      return receiver_mode == ConvertReceiverMode::kNullOrUndefined
                  ? UndefinedConstant()
                  : LoadRegisterAtOperandIndex(1);
+    };
     RegListNodePair args = GetRegisterListAtOperandIndex(1);
     TNode<UintPtrT> slot_id = BytecodeOperandIdx(3);
     TNode<HeapObject> maybe_feedback_vector = LoadFeedbackVector();
@@ -1404,10 +1405,11 @@ class InterpreterJSCallAssembler : public InterpreterAssembler {
         kFirstArgumentOperandIndex + kReceiverAndArgOperandCount;
 
     TNode<Object> function = LoadRegisterAtOperandIndex(0);
-    TNode<Object> receiver =
-        receiver_mode == ConvertReceiverMode::kNullOrUndefined
+    LazyNode<Object> receiver = [=] {
+      return receiver_mode == ConvertReceiverMode::kNullOrUndefined
                  ? UndefinedConstant()
                  : LoadRegisterAtOperandIndex(1);
+    };
     TNode<UintPtrT> slot_id = BytecodeOperandIdx(kSlotOperandIndex);
     TNode<HeapObject> maybe_feedback_vector = LoadFeedbackVector();
     TNode<Context> context = GetContext();