[turbofan] Inline StringCharCodeAt like Crankshaft did.

This avoids the call to the StringCharCodeAt builtin from
within TurboFan optimized code and instead emits a loop
that does the character load. This (together with previously
reverted CL to the JSCallReducer) almost completely recovers
the performance regression caused when we shipped TurboFan.

Without untrusted code mitigations the benchmark goes from
580ms to roughly 490ms, and with the patch to the JSCallReducer
the time goes down to 280ms, which is very close to what we
had with Crankshaft.

This also renames the LoadFromString helper method in the
EffectControlLinearizer to LoadFromSeqString to make it
clear what it does.

Bug: v8:7326
Change-Id: Ibe0ec1847911a234f244bd8dcec6be18b241fda0
Reviewed-on: https://chromium-review.googlesource.com/873376Reviewed-by: Yang Guo <yangguo@chromium.org>
Commit-Queue: Benedikt Meurer <bmeurer@chromium.org>
Cr-Commit-Position: refs/heads/master@{#50702}

src/builtins/builtins-definitions.h

@@ -92,7 +92,6 @@ namespace internal {
                                                                                \
   /* String helpers */                                                         \
   TFC(StringCharAt, StringAt, 1)                                               \
-  TFC(StringCharCodeAt, StringAt, 1)                                           \
   TFC(StringCodePointAt, StringAt, 1)                                          \
   TFC(StringEqual, Compare, 1)                                                 \
   TFC(StringGreaterThan, Compare, 1)                                           \

src/builtins/builtins-string-gen.cc

@@ -563,19 +563,6 @@ TF_BUILTIN(StringCharAt, StringBuiltinsAssembler) {
   Return(result);
 }
 
-TF_BUILTIN(StringCharCodeAt, StringBuiltinsAssembler) {
-  Node* receiver = Parameter(Descriptor::kReceiver);
-  Node* position = Parameter(Descriptor::kPosition);
-
-  // Load the character code at the {position} from the {receiver}.
-  TNode<Int32T> code = StringCharCodeAt(receiver, position);
-
-  // And return it as TaggedSigned value.
-  // TODO(turbofan): Allow builtins to return values untagged.
-  TNode<Smi> result = SmiFromWord32(code);
-  Return(result);
-}
-
 TF_BUILTIN(StringCodePointAt, StringBuiltinsAssembler) {
   Node* receiver = Parameter(Descriptor::kReceiver);
   Node* position = Parameter(Descriptor::kPosition);

src/compiler/effect-control-linearizer.cc

@@ -2702,15 +2702,139 @@ Node* EffectControlLinearizer::LowerStringCharCodeAt(Node* node) {
   Node* receiver = node->InputAt(0);
   Node* position = node->InputAt(1);
 
-  Callable const callable =
-      Builtins::CallableFor(isolate(), Builtins::kStringCharCodeAt);
-  Operator::Properties properties = Operator::kNoThrow | Operator::kNoWrite;
-  CallDescriptor::Flags flags = CallDescriptor::kNoFlags;
-  CallDescriptor* desc = Linkage::GetStubCallDescriptor(
-      isolate(), graph()->zone(), callable.descriptor(), 0, flags, properties,
-      MachineType::TaggedSigned());
-  return __ Call(desc, __ HeapConstant(callable.code()), receiver, position,
-                 __ NoContextConstant());
+  // We need a loop here to properly deal with indirect strings
+  // (SlicedString, ConsString and ThinString).
+  auto loop = __ MakeLoopLabel(MachineRepresentation::kTagged,
+                               MachineRepresentation::kWord32);
+  auto loop_next = __ MakeLabel(MachineRepresentation::kTagged,
+                                MachineRepresentation::kWord32);
+  auto loop_done = __ MakeLabel(MachineRepresentation::kWord32);
+  __ Goto(&loop, receiver, position);
+  __ Bind(&loop);
+  {
+    Node* receiver = loop.PhiAt(0);
+    Node* position = loop.PhiAt(1);
+    Node* receiver_map = __ LoadField(AccessBuilder::ForMap(), receiver);
+    Node* receiver_instance_type =
+        __ LoadField(AccessBuilder::ForMapInstanceType(), receiver_map);
+    Node* receiver_representation = __ Word32And(
+        receiver_instance_type, __ Int32Constant(kStringRepresentationMask));
+
+    // Dispatch on the current {receiver}s string representation.
+    auto if_seqstring = __ MakeLabel();
+    auto if_consstring = __ MakeLabel();
+    auto if_thinstring = __ MakeLabel();
+    auto if_externalstring = __ MakeLabel();
+    auto if_slicedstring = __ MakeLabel();
+    auto if_runtime = __ MakeDeferredLabel();
+    __ GotoIf(__ Word32Equal(receiver_representation,
+                             __ Int32Constant(kSeqStringTag)),
+              &if_seqstring);
+    __ GotoIf(__ Word32Equal(receiver_representation,
+                             __ Int32Constant(kConsStringTag)),
+              &if_consstring);
+    __ GotoIf(__ Word32Equal(receiver_representation,
+                             __ Int32Constant(kThinStringTag)),
+              &if_thinstring);
+    __ GotoIf(__ Word32Equal(receiver_representation,
+                             __ Int32Constant(kExternalStringTag)),
+              &if_externalstring);
+    __ Branch(__ Word32Equal(receiver_representation,
+                             __ Int32Constant(kSlicedStringTag)),
+              &if_slicedstring, &if_runtime);
+
+    __ Bind(&if_seqstring);
+    {
+      Node* receiver_is_onebyte = __ Word32Equal(
+          __ Word32Equal(__ Word32And(receiver_instance_type,
+                                      __ Int32Constant(kStringEncodingMask)),
+                         __ Int32Constant(kTwoByteStringTag)),
+          __ Int32Constant(0));
+      Node* result = LoadFromSeqString(receiver, position, receiver_is_onebyte);
+      __ Goto(&loop_done, result);
+    }
+
+    __ Bind(&if_thinstring);
+    {
+      Node* receiver_actual =
+          __ LoadField(AccessBuilder::ForThinStringActual(), receiver);
+      __ Goto(&loop_next, receiver_actual, position);
+    }
+
+    __ Bind(&if_consstring);
+    {
+      Node* receiver_second =
+          __ LoadField(AccessBuilder::ForConsStringSecond(), receiver);
+      __ GotoIfNot(__ WordEqual(receiver_second, __ EmptyStringConstant()),
+                   &if_runtime);
+      Node* receiver_first =
+          __ LoadField(AccessBuilder::ForConsStringFirst(), receiver);
+      __ Goto(&loop_next, receiver_first, position);
+    }
+
+    __ Bind(&if_externalstring);
+    {
+      // We need to bailout to the runtime for short external strings.
+      __ GotoIf(__ Word32Equal(
+                    __ Word32And(receiver_instance_type,
+                                 __ Int32Constant(kShortExternalStringMask)),
+                    __ Int32Constant(kShortExternalStringTag)),
+                &if_runtime);
+
+      Node* receiver_data = __ LoadField(
+          AccessBuilder::ForExternalStringResourceData(), receiver);
+
+      auto if_onebyte = __ MakeLabel();
+      auto if_twobyte = __ MakeLabel();
+      __ Branch(
+          __ Word32Equal(__ Word32And(receiver_instance_type,
+                                      __ Int32Constant(kStringEncodingMask)),
+                         __ Int32Constant(kTwoByteStringTag)),
+          &if_twobyte, &if_onebyte);
+
+      __ Bind(&if_onebyte);
+      {
+        Node* result = __ Load(MachineType::Uint8(), receiver_data,
+                               ChangeInt32ToIntPtr(position));
+        __ Goto(&loop_done, result);
+      }
+
+      __ Bind(&if_twobyte);
+      {
+        Node* result = __ Load(MachineType::Uint16(), receiver_data,
+                               ChangeInt32ToIntPtr(position));
+        __ Goto(&loop_done, result);
+      }
+    }
+
+    __ Bind(&if_slicedstring);
+    {
+      Node* receiver_offset =
+          __ LoadField(AccessBuilder::ForSlicedStringOffset(), receiver);
+      Node* receiver_parent =
+          __ LoadField(AccessBuilder::ForSlicedStringParent(), receiver);
+      __ Goto(&loop_next, receiver_parent,
+              __ Int32Add(position, ChangeSmiToInt32(receiver_offset)));
+    }
+
+    __ Bind(&if_runtime);
+    {
+      Operator::Properties properties = Operator::kNoDeopt | Operator::kNoThrow;
+      Runtime::FunctionId id = Runtime::kStringCharCodeAt;
+      CallDescriptor const* desc = Linkage::GetRuntimeCallDescriptor(
+          graph()->zone(), id, 2, properties, CallDescriptor::kNoFlags);
+      Node* result = __ Call(
+          desc, __ CEntryStubConstant(1), receiver, ChangeInt32ToSmi(position),
+          __ ExternalConstant(ExternalReference(id, isolate())),
+          __ Int32Constant(2), __ NoContextConstant());
+      __ Goto(&loop_done, ChangeSmiToInt32(result));
+    }
+
+    __ Bind(&loop_next);
+    __ Goto(&loop, loop_next.PhiAt(0), loop_next.PhiAt(1));
+  }
+  __ Bind(&loop_done);
+  return loop_done.PhiAt(0);
 }
 
 Node* EffectControlLinearizer::LowerStringCodePointAt(Node* node) {
@@ -2728,8 +2852,8 @@ Node* EffectControlLinearizer::LowerStringCodePointAt(Node* node) {
                  __ NoContextConstant());
 }
 
-Node* EffectControlLinearizer::LoadFromString(Node* receiver, Node* position,
-                                              Node* is_one_byte) {
+Node* EffectControlLinearizer::LoadFromSeqString(Node* receiver, Node* position,
+                                                 Node* is_one_byte) {
   auto one_byte_load = __ MakeLabel();
   auto done = __ MakeLabel(MachineRepresentation::kWord32);
   __ GotoIf(is_one_byte, &one_byte_load);
@@ -2756,7 +2880,7 @@ Node* EffectControlLinearizer::LowerSeqStringCharCodeAt(Node* node) {
       __ Word32And(instance_type, __ Int32Constant(kStringEncodingMask)),
       __ Int32Constant(kOneByteStringTag));
 
-  return LoadFromString(receiver, position, is_one_byte);
+  return LoadFromSeqString(receiver, position, is_one_byte);
 }
 
 Node* EffectControlLinearizer::LowerSeqStringCodePointAt(
@@ -2770,7 +2894,7 @@ Node* EffectControlLinearizer::LowerSeqStringCodePointAt(
       __ Word32And(instance_type, __ Int32Constant(kStringEncodingMask)),
       __ Int32Constant(kOneByteStringTag));
 
-  Node* first_char_code = LoadFromString(receiver, position, is_one_byte);
+  Node* first_char_code = LoadFromSeqString(receiver, position, is_one_byte);
 
   auto return_result = __ MakeLabel(MachineRepresentation::kWord32);
 
@@ -2787,7 +2911,8 @@ Node* EffectControlLinearizer::LowerSeqStringCodePointAt(
   __ GotoIf(next_position_in_range, &return_result, first_char_code);
 
   // Load second character code.
-  Node* second_char_code = LoadFromString(receiver, next_position, is_one_byte);
+  Node* second_char_code =
+      LoadFromSeqString(receiver, next_position, is_one_byte);
   // Check if first character code is outside of interval [0xD800, 0xDBFF].
   Node* second_out =
       __ Word32Equal(__ Word32And(second_char_code, __ Int32Constant(0xFC00)),
@@ -3218,10 +3343,14 @@ Node* EffectControlLinearizer::AllocateHeapNumberWithValue(Node* value) {
 }
 
 Node* EffectControlLinearizer::ChangeInt32ToSmi(Node* value) {
+  return __ WordShl(ChangeInt32ToIntPtr(value), SmiShiftBitsConstant());
+}
+
+Node* EffectControlLinearizer::ChangeInt32ToIntPtr(Node* value) {
   if (machine()->Is64()) {
     value = __ ChangeInt32ToInt64(value);
   }
-  return __ WordShl(value, SmiShiftBitsConstant());
+  return value;
 }
 
 Node* EffectControlLinearizer::ChangeIntPtrToInt32(Node* value) {

src/compiler/effect-control-linearizer.h

@@ -176,13 +176,14 @@ class V8_EXPORT_PRIVATE EffectControlLinearizer {
   Node* IsElementsKindGreaterThan(Node* kind, ElementsKind reference_kind);
 
   Node* ChangeInt32ToSmi(Node* value);
+  Node* ChangeInt32ToIntPtr(Node* value);
   Node* ChangeIntPtrToInt32(Node* value);
   Node* ChangeUint32ToUintPtr(Node* value);
   Node* ChangeUint32ToSmi(Node* value);
   Node* ChangeSmiToIntPtr(Node* value);
   Node* ChangeSmiToInt32(Node* value);
   Node* ObjectIsSmi(Node* value);
-  Node* LoadFromString(Node* receiver, Node* position, Node* is_one_byte);
+  Node* LoadFromSeqString(Node* receiver, Node* position, Node* is_one_byte);
 
   Node* SmiMaxValueConstant();
   Node* SmiShiftBitsConstant();

src/compiler/simplified-lowering.cc

@@ -2352,9 +2352,8 @@ class RepresentationSelector {
             NodeProperties::ChangeOp(node, simplified()->SeqStringCharCodeAt());
           }
         } else {
-          // TODO(turbofan): Allow builtins to return untagged values.
           VisitBinop(node, UseInfo::AnyTagged(), UseInfo::TruncatingWord32(),
-                     MachineRepresentation::kTaggedSigned);
+                     MachineRepresentation::kWord32);
         }
         return;
       }