[string] Move String.p.toLowerCase to CSA

This CL migrates the CPP builtin to CSA with fast paths for strings
that can be unpacked to direct one-byte strings. Short strings are
handled directly in CSA, others need to call into C for conversion.

Microbenchmarks for "abcd".toLowerCase() show speedups of 2.5x.

BUG=v8:6353,v8:6344

Review-Url: https://codereview.chromium.org/2859203002
Cr-Commit-Position: refs/heads/master@{#45141}

BUILD.gn

@@ -919,6 +919,7 @@ v8_source_set("v8_builtins_generators") {
     "src/builtins/builtins-ic-gen.cc",
     "src/builtins/builtins-internal-gen.cc",
     "src/builtins/builtins-interpreter-gen.cc",
+    "src/builtins/builtins-intl-gen.cc",
     "src/builtins/builtins-math-gen.cc",
     "src/builtins/builtins-number-gen.cc",
     "src/builtins/builtins-object-gen.cc",
@@ -996,6 +997,10 @@ v8_source_set("v8_builtins_generators") {
     ]
   }
 
+  if (!v8_enable_i18n_support) {
+    sources -= [ "src/builtins/builtins-intl-gen.cc" ]
+  }
+
   configs = [ ":internal_config" ]
 }
 

src/assembler.cc

@@ -91,6 +91,10 @@
 #endif  // Target architecture.
 #endif  // V8_INTERPRETED_REGEXP
 
+#ifdef V8_INTL_SUPPORT
+#include "src/intl.h"
+#endif  // V8_INTL_SUPPORT
+
 namespace v8 {
 namespace internal {
 
@@ -1572,6 +1576,20 @@ ExternalReference ExternalReference::try_internalize_string_function(
       isolate, FUNCTION_ADDR(StringTable::LookupStringIfExists_NoAllocate)));
 }
 
+#ifdef V8_INTL_SUPPORT
+ExternalReference ExternalReference::intl_convert_one_byte_to_lower(
+    Isolate* isolate) {
+  return ExternalReference(
+      Redirect(isolate, FUNCTION_ADDR(ConvertOneByteToLower)));
+}
+
+ExternalReference ExternalReference::intl_to_latin1_lower_table(
+    Isolate* isolate) {
+  uint8_t* ptr = const_cast<uint8_t*>(ToLatin1LowerTable());
+  return ExternalReference(reinterpret_cast<Address>(ptr));
+}
+#endif  // V8_INTL_SUPPORT
+
 // Explicit instantiations for all combinations of 1- and 2-byte strings.
 template ExternalReference
 ExternalReference::search_string_raw<const uint8_t, const uint8_t>(Isolate*);

src/assembler.h

@@ -993,6 +993,11 @@ class ExternalReference BASE_EMBEDDED {
 
   static ExternalReference try_internalize_string_function(Isolate* isolate);
 
+#ifdef V8_INTL_SUPPORT
+  static ExternalReference intl_convert_one_byte_to_lower(Isolate* isolate);
+  static ExternalReference intl_to_latin1_lower_table(Isolate* isolate);
+#endif  // V8_INTL_SUPPORT
+
   template <typename SubjectChar, typename PatternChar>
   static ExternalReference search_string_raw(Isolate* isolate);
 

src/bootstrapper.cc

@@ -4084,7 +4084,7 @@ void Genesis::InitializeGlobal_icu_case_mapping() {
     SetFunction(string_prototype,
                 SimpleCreateFunction(isolate(), name,
                                      Builtins::kStringPrototypeToLowerCaseIntl,
-                                     0, false),
+                                     0, true),
                 name);
   }
   {

src/builtins/builtins-definitions.h

@@ -1004,7 +1004,7 @@ namespace internal {
   BUILTIN_LIST_BASE(CPP, API, TFJ, TFC, TFS, TFH, ASM, DBG)  \
                                                              \
   /* ES #sec-string.prototype.tolowercase */                 \
-  CPP(StringPrototypeToLowerCaseIntl)                        \
+  TFJ(StringPrototypeToLowerCaseIntl, 0)                     \
   /* ES #sec-string.prototype.touppercase */                 \
   CPP(StringPrototypeToUpperCaseIntl)                        \
   /* ES #sec-string.prototype.normalize */                   \

src/builtins/builtins-intl-gen.cc

@@ -0,0 +1,113 @@
+// Copyright 2017 the V8 project authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+#ifndef V8_INTL_SUPPORT
+#error Internationalization is expected to be enabled.
+#endif  // V8_INTL_SUPPORT
+
+#include "src/builtins/builtins-utils-gen.h"
+#include "src/code-stub-assembler.h"
+
+namespace v8 {
+namespace internal {
+
+class IntlBuiltinsAssembler : public CodeStubAssembler {
+ public:
+  explicit IntlBuiltinsAssembler(compiler::CodeAssemblerState* state)
+      : CodeStubAssembler(state) {}
+};
+
+TF_BUILTIN(StringPrototypeToLowerCaseIntl, IntlBuiltinsAssembler) {
+  Node* const maybe_string = Parameter(Descriptor::kReceiver);
+  Node* const context = Parameter(Descriptor::kContext);
+
+  Node* const string =
+      ToThisString(context, maybe_string, "String.prototype.toLowerCase");
+
+  Label call_c(this), return_string(this), runtime(this, Label::kDeferred);
+
+  // Early exit on empty strings.
+  Node* const length = SmiUntag(LoadStringLength(string));
+  GotoIf(IntPtrEqual(length, IntPtrConstant(0)), &return_string);
+
+  // Unpack strings if possible, and bail to runtime unless we get a one-byte
+  // flat string.
+  ToDirectStringAssembler to_direct(
+      state(), string, ToDirectStringAssembler::kDontUnpackSlicedStrings);
+  to_direct.TryToDirect(&runtime);
+
+  Node* const instance_type = to_direct.instance_type();
+  CSA_ASSERT(this,
+             Word32BinaryNot(IsIndirectStringInstanceType(instance_type)));
+  GotoIfNot(IsOneByteStringInstanceType(instance_type), &runtime);
+
+  // For short strings, do the conversion in CSA through the lookup table.
+
+  Node* const dst = AllocateSeqOneByteString(context, length);
+
+  const int kMaxShortStringLength = 24;  // Determined empirically.
+  GotoIf(IntPtrGreaterThan(length, IntPtrConstant(kMaxShortStringLength)),
+         &call_c);
+
+  {
+    Node* const dst_ptr = PointerToSeqStringData(dst);
+    VARIABLE(var_cursor, MachineType::PointerRepresentation(),
+             IntPtrConstant(0));
+
+    Node* const start_address = to_direct.PointerToData(&call_c);
+    Node* const end_address = IntPtrAdd(start_address, length);
+
+    Node* const to_lower_table_addr = ExternalConstant(
+        ExternalReference::intl_to_latin1_lower_table(isolate()));
+
+    VariableList push_vars({&var_cursor}, zone());
+    BuildFastLoop(
+        push_vars, start_address, end_address,
+        [=, &var_cursor](Node* current) {
+          Node* c = ChangeInt32ToIntPtr(Load(MachineType::Uint8(), current));
+          Node* lower = Load(MachineType::Uint8(), to_lower_table_addr, c);
+          StoreNoWriteBarrier(MachineRepresentation::kWord8, dst_ptr,
+                              var_cursor.value(), lower);
+          Increment(var_cursor);
+        },
+        kCharSize, INTPTR_PARAMETERS, IndexAdvanceMode::kPost);
+
+    // All lower-case.
+    Return(dst);
+  }
+
+  // Call into C for case conversion. The signature is:
+  // Object* ConvertOneByteToLower(String* src, String* dst, Isolate* isolate);
+  BIND(&call_c);
+  {
+    Node* const src = to_direct.string();
+
+    Node* const function_addr = ExternalConstant(
+        ExternalReference::intl_convert_one_byte_to_lower(isolate()));
+    Node* const isolate_ptr =
+        ExternalConstant(ExternalReference::isolate_address(isolate()));
+
+    MachineType type_ptr = MachineType::Pointer();
+    MachineType type_tagged = MachineType::AnyTagged();
+
+    Node* const result =
+        CallCFunction3(type_tagged, type_tagged, type_tagged, type_ptr,
+                       function_addr, src, dst, isolate_ptr);
+
+    Return(result);
+  }
+
+  BIND(&return_string);
+  Return(string);
+
+  BIND(&runtime);
+  {
+    Node* const result =
+        CallRuntime(Runtime::kStringToLowerCaseIntl, context, string);
+    Return(result);
+  }
+}
+
+}  // namespace internal
+}  // namespace v8

src/builtins/builtins-intl.cc

@@ -16,13 +16,6 @@
 namespace v8 {
 namespace internal {
 
-BUILTIN(StringPrototypeToLowerCaseIntl) {
-  HandleScope scope(isolate);
-  TO_THIS_STRING(string, "String.prototype.toLowerCase");
-  string = String::Flatten(string);
-  return ConvertCase(string, false, isolate);
-}
-
 BUILTIN(StringPrototypeToUpperCaseIntl) {
   HandleScope scope(isolate);
   TO_THIS_STRING(string, "String.prototype.toUpperCase");

src/code-stub-assembler.cc

@@ -1238,6 +1238,16 @@ Node* CodeStubAssembler::LoadStringLength(Node* object) {
   return LoadObjectField(object, String::kLengthOffset);
 }
 
+Node* CodeStubAssembler::PointerToSeqStringData(Node* seq_string) {
+  CSA_ASSERT(this, IsString(seq_string));
+  CSA_ASSERT(this,
+             IsSequentialStringInstanceType(LoadInstanceType(seq_string)));
+  STATIC_ASSERT(SeqOneByteString::kHeaderSize == SeqTwoByteString::kHeaderSize);
+  return IntPtrAdd(
+      BitcastTaggedToWord(seq_string),
+      IntPtrConstant(SeqOneByteString::kHeaderSize - kHeapObjectTag));
+}
+
 Node* CodeStubAssembler::LoadJSValueValue(Node* object) {
   CSA_ASSERT(this, IsJSValue(object));
   return LoadObjectField(object, JSValue::kValueOffset);
@@ -3137,6 +3147,13 @@ Node* CodeStubAssembler::IsConsStringInstanceType(Node* instance_type) {
       Int32Constant(kConsStringTag));
 }
 
+Node* CodeStubAssembler::IsIndirectStringInstanceType(Node* instance_type) {
+  CSA_ASSERT(this, IsStringInstanceType(instance_type));
+  STATIC_ASSERT(kIsIndirectStringMask == 0x1);
+  STATIC_ASSERT(kIsIndirectStringTag == 0x1);
+  return Word32And(instance_type, Int32Constant(kIsIndirectStringMask));
+}
+
 Node* CodeStubAssembler::IsExternalStringInstanceType(Node* instance_type) {
   CSA_ASSERT(this, IsStringInstanceType(instance_type));
   return Word32Equal(
@@ -3660,12 +3677,13 @@ Node* CodeStubAssembler::SubString(Node* context, Node* string, Node* from,
 }
 
 ToDirectStringAssembler::ToDirectStringAssembler(
-    compiler::CodeAssemblerState* state, Node* string)
+    compiler::CodeAssemblerState* state, Node* string, Flags flags)
     : CodeStubAssembler(state),
       var_string_(this, MachineRepresentation::kTagged, string),
       var_instance_type_(this, MachineRepresentation::kWord32),
       var_offset_(this, MachineType::PointerRepresentation()),
-      var_is_external_(this, MachineRepresentation::kWord32) {
+      var_is_external_(this, MachineRepresentation::kWord32),
+      flags_(flags) {
   CSA_ASSERT(this, TaggedIsNotSmi(string));
   CSA_ASSERT(this, IsString(string));
 
@@ -3722,16 +3740,20 @@ Node* ToDirectStringAssembler::TryToDirect(Label* if_bailout) {
   // Sliced string. Fetch parent and correct start index by offset.
   BIND(&if_issliced);
   {
-    Node* const string = var_string_.value();
-    Node* const sliced_offset =
-        LoadAndUntagObjectField(string, SlicedString::kOffsetOffset);
-    var_offset_.Bind(IntPtrAdd(var_offset_.value(), sliced_offset));
+    if (flags_ & kDontUnpackSlicedStrings) {
+      Goto(if_bailout);
+    } else {
+      Node* const string = var_string_.value();
+      Node* const sliced_offset =
+          LoadAndUntagObjectField(string, SlicedString::kOffsetOffset);
+      var_offset_.Bind(IntPtrAdd(var_offset_.value(), sliced_offset));
 
-    Node* const parent = LoadObjectField(string, SlicedString::kParentOffset);
-    var_string_.Bind(parent);
-    var_instance_type_.Bind(LoadInstanceType(parent));
+      Node* const parent = LoadObjectField(string, SlicedString::kParentOffset);
+      var_string_.Bind(parent);
+      var_instance_type_.Bind(LoadInstanceType(parent));
 
-    Goto(&dispatch);
+      Goto(&dispatch);
+    }
   }
 
   // Thin string. Fetch the actual string.

src/code-stub-assembler.h

@@ -420,6 +420,8 @@ class V8_EXPORT_PRIVATE CodeStubAssembler : public compiler::CodeAssembler {
 
   // Load length field of a String object.
   Node* LoadStringLength(Node* object);
+  // Loads a pointer to the sequential String char array.
+  Node* PointerToSeqStringData(Node* seq_string);
   // Load value field of a JSValue object.
   Node* LoadJSValueValue(Node* object);
   // Load value field of a WeakCell object.
@@ -742,6 +744,7 @@ class V8_EXPORT_PRIVATE CodeStubAssembler : public compiler::CodeAssembler {
   Node* IsShortExternalStringInstanceType(Node* instance_type);
   Node* IsSequentialStringInstanceType(Node* instance_type);
   Node* IsConsStringInstanceType(Node* instance_type);
+  Node* IsIndirectStringInstanceType(Node* instance_type);
   Node* IsString(Node* object);
   Node* IsJSObject(Node* object);
   Node* IsJSGlobalProxy(Node* object);
@@ -1531,19 +1534,28 @@ class ToDirectStringAssembler : public CodeStubAssembler {
   enum StringPointerKind { PTR_TO_DATA, PTR_TO_STRING };
 
  public:
-  explicit ToDirectStringAssembler(compiler::CodeAssemblerState* state,
-                                   Node* string);
+  enum Flag {
+    kDontUnpackSlicedStrings = 1 << 0,
+  };
+  typedef base::Flags<Flag> Flags;
+
+  ToDirectStringAssembler(compiler::CodeAssemblerState* state, Node* string,
+                          Flags flags = Flags());
 
   // Converts flat cons, thin, and sliced strings and returns the direct
   // string. The result can be either a sequential or external string.
+  // Jumps to if_bailout if the string if the string is indirect and cannot
+  // be unpacked.
   Node* TryToDirect(Label* if_bailout);
 
   // Returns a pointer to the beginning of the string data.
+  // Jumps to if_bailout if the external string cannot be unpacked.
   Node* PointerToData(Label* if_bailout) {
     return TryToSequential(PTR_TO_DATA, if_bailout);
   }
 
   // Returns a pointer that, offset-wise, looks like a String.
+  // Jumps to if_bailout if the external string cannot be unpacked.
   Node* PointerToString(Label* if_bailout) {
     return TryToSequential(PTR_TO_STRING, if_bailout);
   }
@@ -1560,6 +1572,8 @@ class ToDirectStringAssembler : public CodeStubAssembler {
   Variable var_instance_type_;
   Variable var_offset_;
   Variable var_is_external_;
+
+  const Flags flags_;
 };
 
 #ifdef DEBUG

src/external-reference-table.cc

@@ -250,6 +250,12 @@ void ExternalReferenceTable::AddReferences(Isolate* isolate) {
       "libc_memset");
   Add(ExternalReference::try_internalize_string_function(isolate).address(),
       "try_internalize_string_function");
+#ifdef V8_INTL_SUPPORT
+  Add(ExternalReference::intl_convert_one_byte_to_lower(isolate).address(),
+      "intl_convert_one_byte_to_lower");
+  Add(ExternalReference::intl_to_latin1_lower_table(isolate).address(),
+      "intl_to_latin1_lower_table");
+#endif  // V8_INTL_SUPPORT
   Add(ExternalReference::search_string_raw<const uint8_t, const uint8_t>(
           isolate)
           .address(),

src/intl.cc

@@ -141,6 +141,8 @@ inline int FindFirstUpperOrNonAscii(Handle<String> s, int length) {
 
 }  // namespace
 
+const uint8_t* ToLatin1LowerTable() { return &kToLower[0]; }
+
 const UChar* GetUCharBufferFromFlat(const String::FlatContent& flat,
                                     std::unique_ptr<uc16[]>* dest,
                                     int32_t length) {
@@ -201,6 +203,41 @@ MUST_USE_RESULT Object* LocaleConvertCase(Handle<String> s, Isolate* isolate,
   return *s;
 }
 
+// A stripped-down version of ConvertToLower that can only handle flat one-byte
+// strings and does not allocate.
+// Called from TF builtins.
+MUST_USE_RESULT Object* ConvertOneByteToLower(String* src, String* dst,
+                                              Isolate* isolate) {
+  DCHECK_EQ(src->length(), dst->length());
+  DCHECK(src->IsOneByteRepresentation());
+  DCHECK(src->IsFlat());
+  DCHECK(dst->IsSeqOneByteString());
+
+  DisallowHeapAllocation no_gc;
+
+  const int length = src->length();
+
+  const uint8_t* src_data = src->GetFlatContent().ToOneByteVector().start();
+  uint8_t* dst_data = SeqOneByteString::cast(dst)->GetChars();
+
+  bool has_changed_character = false;
+  int index_to_first_unprocessed = FastAsciiConvert<true>(
+      reinterpret_cast<char*>(dst_data),
+      reinterpret_cast<const char*>(src_data), length, &has_changed_character);
+
+  if (index_to_first_unprocessed == length) {
+    return has_changed_character ? dst : src;
+  }
+
+  // If not ASCII, we keep the result up to index_to_first_unprocessed and
+  // process the rest.
+  for (int index = index_to_first_unprocessed; index < length; ++index) {
+    dst_data[index] = ToLatin1Lower(static_cast<uint16_t>(src_data[index]));
+  }
+
+  return dst;
+}
+
 MUST_USE_RESULT Object* ConvertToLower(Handle<String> s, Isolate* isolate) {
   if (!s->HasOnlyOneByteChars()) {
     // Use a slower implementation for strings with characters beyond U+00FF.
@@ -230,36 +267,27 @@ MUST_USE_RESULT Object* ConvertToLower(Handle<String> s, Isolate* isolate) {
   Handle<SeqOneByteString> result =
       isolate->factory()->NewRawOneByteString(length).ToHandleChecked();
 
+  if (s->IsOneByteRepresentation()) {
+    return ConvertOneByteToLower(*s, *result, isolate);
+  }
+
   DisallowHeapAllocation no_gc;
   DCHECK(s->IsFlat());
+  DCHECK(s->IsTwoByteRepresentation());
   String::FlatContent flat = s->GetFlatContent();
-  uint8_t* dest = result->GetChars();
-  if (flat.IsOneByte()) {
-    const uint8_t* src = flat.ToOneByteVector().start();
-    bool has_changed_character = false;
-    index_to_first_unprocessed = FastAsciiConvert<true>(
-        reinterpret_cast<char*>(dest), reinterpret_cast<const char*>(src),
-        length, &has_changed_character);
-    // If not ASCII, we keep the result up to index_to_first_unprocessed and
-    // process the rest.
-    if (index_to_first_unprocessed == length)
-      return has_changed_character ? *result : *s;
+  DCHECK(flat.IsTwoByte());
 
-    for (int index = index_to_first_unprocessed; index < length; ++index) {
-      dest[index] = ToLatin1Lower(static_cast<uint16_t>(src[index]));
-    }
-  } else {
-    if (index_to_first_unprocessed == length) {
-      DCHECK(!is_short);
-      index_to_first_unprocessed = FindFirstUpperOrNonAscii(s, length);
-    }
-    // Nothing to do if the string is all ASCII with no uppercase.
-    if (index_to_first_unprocessed == length) return *s;
-    const uint16_t* src = flat.ToUC16Vector().start();
-    CopyChars(dest, src, index_to_first_unprocessed);
-    for (int index = index_to_first_unprocessed; index < length; ++index) {
-      dest[index] = ToLatin1Lower(static_cast<uint16_t>(src[index]));
-    }
+  uint8_t* dest = result->GetChars();
+  if (index_to_first_unprocessed == length) {
+    DCHECK(!is_short);
+    index_to_first_unprocessed = FindFirstUpperOrNonAscii(s, length);
+  }
+  // Nothing to do if the string is all ASCII with no uppercase.
+  if (index_to_first_unprocessed == length) return *s;
+  const uint16_t* src = flat.ToUC16Vector().start();
+  CopyChars(dest, src, index_to_first_unprocessed);
+  for (int index = index_to_first_unprocessed; index < length; ++index) {
+    dest[index] = ToLatin1Lower(static_cast<uint16_t>(src[index]));
   }
 
   return *result;

src/intl.h

@@ -30,6 +30,11 @@ MUST_USE_RESULT Object* ConvertToUpper(Handle<String> s, Isolate* isolate);
 MUST_USE_RESULT Object* ConvertCase(Handle<String> s, bool is_upper,
                                     Isolate* isolate);
 
+MUST_USE_RESULT Object* ConvertOneByteToLower(String* src, String* dst,
+                                              Isolate* isolate);
+
+const uint8_t* ToLatin1LowerTable();
+
 // ICUTimezoneCache calls out to ICU for TimezoneCache
 // functionality in a straightforward way.
 class ICUTimezoneCache : public base::TimezoneCache {

src/v8.gyp

@@ -528,6 +528,7 @@
         'builtins/builtins-string.cc',
         'builtins/builtins-string-gen.cc',
         'builtins/builtins-intl.cc',
+        'builtins/builtins-intl-gen.cc',
         'builtins/builtins-symbol.cc',
         'builtins/builtins-symbol-gen.cc',
         'builtins/builtins-typedarray.cc',
@@ -1836,6 +1837,7 @@
         }, {  # v8_enable_i18n_support==0
           'sources!': [
             'builtins/builtins-intl.cc',
+            'builtins/builtins-intl-gen.cc',
             'intl.cc',
             'intl.h',
             'objects/intl-objects.cc',

test/mjsunit/string-case.js

@@ -59,8 +59,9 @@ function test(length) {
     strLower += String.fromCharCode(charCodeToLower(c));
     strUpper += String.fromCharCode(charCodeToUpper(c));
   }
-  %FlattenString(strLower);
-  %FlattenString(strUpper);
+  str = %FlattenString(str);
+  strLower = %FlattenString(strLower);
+  strUpper = %FlattenString(strUpper);
   // Sequential string.
   assertEquals(strLower, str.toLowerCase());
   assertEquals(strUpper, str.toUpperCase());