Port and integrate the irregexp engine from V8

runtime/vm/intrinsifier_x64.cc

 // Copyright (c) 2013, the Dart project authors.  Please see the AUTHORS file
 // for details. All rights reserved. Use of this source code is governed by a
 // BSD-style license that can be found in the LICENSE file.
 
 #include "vm/globals.h"  // Needed here to get TARGET_ARCH_X64.
 #if defined(TARGET_ARCH_X64)
 
 #include "vm/intrinsifier.h"
 
 #include "vm/assembler.h"
@@ skipping 1492 matching lines @@
 void Intrinsifier::OneByteString_equality(Assembler* assembler) {
   StringEquality(assembler, kOneByteStringCid);
 }
 
 
 void Intrinsifier::TwoByteString_equality(Assembler* assembler) {
   StringEquality(assembler, kTwoByteStringCid);
 }
 
 
+static void GenerateRegexpSpecializationFor(
+    Assembler* assembler,
+    intptr_t cid,
+    Label* entry,
+    Label* exit,
+    bool jump_kind = Assembler::kFarJump) {
+  ASSERT(cid == kOneByteStringCid || cid == kExternalOneByteStringCid ||
+         cid == kTwoByteStringCid || cid == kExternalTwoByteStringCid);
+
+  __ Bind(entry);
+
+  // The passed string is of class cid.
+
+  __ movq(RAX, FieldAddress(RBX, JSRegExp::function_offset(cid)));
+
+  // Set the subject if it is currently unset.
+
+  __ LoadObject(RCX, Object::null_object(), PP);
+  __ cmpq(RCX, FieldAddress(RBX, JSRegExp::subject_offset(cid)));
+  __ j(NOT_EQUAL, exit, jump_kind);
+
+  __ StoreIntoObject(RBX,

[Vyacheslav Egorov (Google), 2014/10/01 20:13:21]

I really don't think there is any good reason to keep sample subject alive after
we compiled compiled regexp - which means you should really null it after regexp
is compiled.

Otherwise there will be hard to debug memory leaks in the system.

There is a lot of complexity in the system because we are trying to make normal
lazy compilation work for us.

I suggest we just do 

mov target, [RBX + JSRegExp::function_offset(cid)];
cmp target, #null
jne call-target
;; compile target explicitly passing subject as an argument

calltarget:
;; target is created and compiled here

This would allow us to kill a lot of fields that you added just to make current
lazy compilation system to work.

+                     FieldAddress(RBX, JSRegExp::subject_offset(cid)),
+                     RDI,
+                     false);
+
+  __ jmp(exit, jump_kind);
+}
+
+
+void Intrinsifier::JSRegExp_ExecuteMatch(Assembler* assembler) {
+  Label is_one_byte_string;
+  Label is_two_byte_string;
+  Label is_external_one_byte_string;
+  Label is_external_two_byte_string;
+  Label fallthrough;
+
+  static const intptr_t kRegExpParamOffset = + 3 * kWordSize;
+  static const intptr_t kStringParamOffset = + 2 * kWordSize;
+  // const Smi& start_index is located at   (+ 1 * kWordSize).
+
+  // Register assignments are as follows:
+  // RAX: The appropriate specialized matcher function.
+  // RBX: The regexp object.
+  // RCX: Temp.
+  // RDI: Temp, Pointer to the function's code which we then tail-call.
+
+  __ movq(RBX, Address(RSP, kRegExpParamOffset));
+  __ movq(RDI, Address(RSP, kStringParamOffset));
+
+  // Check which string class we have been passed.
+
+  __ CompareClassId(RDI, kOneByteStringCid);
+  __ j(EQUAL, &is_one_byte_string);
+
+  __ CompareClassId(RDI, kTwoByteStringCid);
+  __ j(EQUAL, &is_two_byte_string);
+
+  __ CompareClassId(RDI, kExternalOneByteStringCid);
+  __ j(EQUAL, &is_external_one_byte_string);
+
+  __ CompareClassId(RDI, kExternalTwoByteStringCid);
+  __ j(EQUAL, &is_external_two_byte_string);
+
+  __ Stop("Unexpected class id");
+
+  GenerateRegexpSpecializationFor(assembler, kExternalTwoByteStringCid,
+                                  &is_external_two_byte_string, &fallthrough);
+  GenerateRegexpSpecializationFor(assembler, kExternalOneByteStringCid,
+                                  &is_external_one_byte_string, &fallthrough);
+  GenerateRegexpSpecializationFor(assembler, kTwoByteStringCid,
+                                  &is_two_byte_string, &fallthrough,
+                                  Assembler::kNearJump);
+  GenerateRegexpSpecializationFor(assembler, kOneByteStringCid,
+                                  &is_one_byte_string, &fallthrough,
+                                  Assembler::kNearJump);
+
+  // RAX contains the appropriate specialized function.
+
+  __ Bind(&fallthrough);
+
+  // Registers have been set up for the lazy compile stub at this point.
+  // It expects the function in RAX, the argument descriptor in R10, and
+  // IC-Data in RCX. Irregexp generated functions do not use either the
+  // arguments descriptor or IC-Data, and hence these are not set here.
+
+  // Tail-call the function.
+  __ movq(RDI, FieldAddress(RAX, Function::instructions_offset()));
+  __ addq(RDI, Immediate(Instructions::HeaderSize() - kHeapObjectTag));
+  __ jmp(RDI);
+}
+
+
 // On stack: user tag (+1), return-address (+0).
 void Intrinsifier::UserTag_makeCurrent(Assembler* assembler) {
   // RBX: Isolate.
   Isolate* isolate = Isolate::Current();
   const Immediate& isolate_address =
       Immediate(reinterpret_cast<int64_t>(isolate));
   __ movq(RBX, isolate_address);
   // RAX: Current user tag.
   __ movq(RAX, Address(RBX, Isolate::current_tag_offset()));
   // R10: UserTag.
@@ skipping 30 matching lines @@
   // Set return value to Isolate::current_tag_.
   __ movq(RAX, Address(RBX, Isolate::current_tag_offset()));
   __ ret();
 }
 
 #undef __
 
 }  // namespace dart
 
 #endif  // defined TARGET_ARCH_X64