Port inlined version of swap primitive for sorting from ia32 to x64.

Original code review for ia32 version: http://codereview.chromium.org/1709008
Review URL: http://codereview.chromium.org/1858002

git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@4569 ce2b1a6d-e550-0410-aec6-3dcde31c8c00

src/ia32/codegen-ia32.h

@@ -636,7 +636,9 @@ class CodeGenerator: public AstVisitor {
   // Fast support for number to string.
   void GenerateNumberToString(ZoneList<Expression*>* args);
 
-  // Fast swapping of elements.
+  // Fast swapping of elements. Takes three expressions, the object and two
+  // indices. This should only be used if the indices are known to be
+  // non-negative and within bounds of the elements array at the call site.
   void GenerateSwapElements(ZoneList<Expression*>* args);
 
   // Fast call for custom callbacks.

src/x64/codegen-x64.cc

@@ -4487,6 +4487,28 @@ void CodeGenerator::GenerateNumberToString(ZoneList<Expression*>* args) {
 }
 
 
+class DeferredSwapElements: public DeferredCode {
+ public:
+  DeferredSwapElements(Register object, Register index1, Register index2)
+      : object_(object), index1_(index1), index2_(index2) {
+    set_comment("[ DeferredSwapElements");
+  }
+
+  virtual void Generate();
+
+ private:
+  Register object_, index1_, index2_;
+};
+
+
+void DeferredSwapElements::Generate() {
+  __ push(object_);
+  __ push(index1_);
+  __ push(index2_);
+  __ CallRuntime(Runtime::kSwapElements, 3);
+}
+
+
 void CodeGenerator::GenerateSwapElements(ZoneList<Expression*>* args) {
   Comment cmnt(masm_, "[ GenerateSwapElements");
 
@@ -4496,8 +4518,81 @@ void CodeGenerator::GenerateSwapElements(ZoneList<Expression*>* args) {
   Load(args->at(1));
   Load(args->at(2));
 
-  Result result = frame_->CallRuntime(Runtime::kSwapElements, 3);
+  Result index2 = frame_->Pop();
-  frame_->Push(&result);
+  index2.ToRegister();
+
+  Result index1 = frame_->Pop();
+  index1.ToRegister();
+
+  Result object = frame_->Pop();
+  object.ToRegister();
+
+  Result tmp1 = allocator()->Allocate();
+  tmp1.ToRegister();
+  Result tmp2 = allocator()->Allocate();
+  tmp2.ToRegister();
+
+  frame_->Spill(object.reg());
+  frame_->Spill(index1.reg());
+  frame_->Spill(index2.reg());
+
+  DeferredSwapElements* deferred = new DeferredSwapElements(object.reg(),
+                                                            index1.reg(),
+                                                            index2.reg());
+
+  // Fetch the map and check if array is in fast case.
+  // Check that object doesn't require security checks and
+  // has no indexed interceptor.
+  __ CmpObjectType(object.reg(), FIRST_JS_OBJECT_TYPE, tmp1.reg());
+  deferred->Branch(below);
+  __ testb(FieldOperand(tmp1.reg(), Map::kBitFieldOffset),
+           Immediate(KeyedLoadIC::kSlowCaseBitFieldMask));
+  deferred->Branch(not_zero);
+
+  // Check the object's elements are in fast case.
+  __ movq(tmp1.reg(), FieldOperand(object.reg(), JSObject::kElementsOffset));
+  __ CompareRoot(FieldOperand(tmp1.reg(), HeapObject::kMapOffset),
+                 Heap::kFixedArrayMapRootIndex);
+  deferred->Branch(not_equal);
+
+  // Check that both indices are smis.
+  Condition both_smi = __ CheckBothSmi(index1.reg(), index2.reg());
+  deferred->Branch(NegateCondition(both_smi));
+
+  // Bring addresses into index1 and index2.
+  __ SmiToInteger32(index1.reg(), index1.reg());
+  __ lea(index1.reg(), FieldOperand(tmp1.reg(),
+                                    index1.reg(),
+                                    times_pointer_size,
+                                    FixedArray::kHeaderSize));
+  __ SmiToInteger32(index2.reg(), index2.reg());
+  __ lea(index2.reg(), FieldOperand(tmp1.reg(),
+                                    index2.reg(),
+                                    times_pointer_size,
+                                    FixedArray::kHeaderSize));
+
+  // Swap elements.
+  __ movq(object.reg(), Operand(index1.reg(), 0));
+  __ movq(tmp2.reg(), Operand(index2.reg(), 0));
+  __ movq(Operand(index2.reg(), 0), object.reg());
+  __ movq(Operand(index1.reg(), 0), tmp2.reg());
+
+  Label done;
+  __ InNewSpace(tmp1.reg(), tmp2.reg(), equal, &done);
+  // Possible optimization: do a check that both values are Smis
+  // (or them and test against Smi mask.)
+
+  __ movq(tmp2.reg(), tmp1.reg());
+  RecordWriteStub recordWrite1(tmp2.reg(), index1.reg(), object.reg());
+  __ CallStub(&recordWrite1);
+
+  RecordWriteStub recordWrite2(tmp1.reg(), index2.reg(), object.reg());
+  __ CallStub(&recordWrite2);
+
+  __ bind(&done);
+
+  deferred->BindExit();
+  frame_->Push(Factory::undefined_value());
 }
 
 
@@ -8341,6 +8436,12 @@ void NumberToStringStub::Generate(MacroAssembler* masm) {
 }
 
 
+void RecordWriteStub::Generate(MacroAssembler* masm) {
+  masm->RecordWriteHelper(object_, addr_, scratch_);
+  masm->ret(0);
+}
+
+
 static int NegativeComparisonResult(Condition cc) {
   ASSERT(cc != equal);
   ASSERT((cc == less) || (cc == less_equal)

src/x64/codegen-x64.h

@@ -591,7 +591,9 @@ class CodeGenerator: public AstVisitor {
   // Fast support for number to string.
   void GenerateNumberToString(ZoneList<Expression*>* args);
 
-  // Fast swapping of elements.
+  // Fast swapping of elements. Takes three expressions, the object and two
+  // indices. This should only be used if the indices are known to be
+  // non-negative and within bounds of the elements array at the call site.
   void GenerateSwapElements(ZoneList<Expression*>* args);
 
   // Fast call for custom callbacks.
@@ -1011,6 +1013,42 @@ class NumberToStringStub: public CodeStub {
 };
 
 
+class RecordWriteStub : public CodeStub {
+ public:
+  RecordWriteStub(Register object, Register addr, Register scratch)
+      : object_(object), addr_(addr), scratch_(scratch) { }
+
+  void Generate(MacroAssembler* masm);
+
+ private:
+  Register object_;
+  Register addr_;
+  Register scratch_;
+
+#ifdef DEBUG
+  void Print() {
+    PrintF("RecordWriteStub (object reg %d), (addr reg %d), (scratch reg %d)\n",
+           object_.code(), addr_.code(), scratch_.code());
+  }
+#endif
+
+  // Minor key encoding in 12 bits of three registers (object, address and
+  // scratch) OOOOAAAASSSS.
+  class ScratchBits : public BitField<uint32_t, 0, 4> {};
+  class AddressBits : public BitField<uint32_t, 4, 4> {};
+  class ObjectBits : public BitField<uint32_t, 8, 4> {};
+
+  Major MajorKey() { return RecordWrite; }
+
+  int MinorKey() {
+    // Encode the registers.
+    return ObjectBits::encode(object_.code()) |
+           AddressBits::encode(addr_.code()) |
+           ScratchBits::encode(scratch_.code());
+  }
+};
+
+
 } }  // namespace v8::internal
 
 #endif  // V8_X64_CODEGEN_X64_H_

src/x64/macro-assembler-x64.cc

@@ -72,35 +72,34 @@ void MacroAssembler::StackLimitCheck(Label* on_stack_overflow) {
 }
 
 
-static void RecordWriteHelper(MacroAssembler* masm,
+void MacroAssembler::RecordWriteHelper(Register object,
-                              Register object,
+                                       Register addr,
-                              Register addr,
+                                       Register scratch) {
-                              Register scratch) {
   Label fast;
 
   // Compute the page start address from the heap object pointer, and reuse
   // the 'object' register for it.
   ASSERT(is_int32(~Page::kPageAlignmentMask));
-  masm->and_(object,
+  and_(object,
-             Immediate(static_cast<int32_t>(~Page::kPageAlignmentMask)));
+       Immediate(static_cast<int32_t>(~Page::kPageAlignmentMask)));
   Register page_start = object;
 
   // Compute the bit addr in the remembered set/index of the pointer in the
   // page. Reuse 'addr' as pointer_offset.
-  masm->subq(addr, page_start);
+  subq(addr, page_start);
-  masm->shr(addr, Immediate(kPointerSizeLog2));
+  shr(addr, Immediate(kPointerSizeLog2));
   Register pointer_offset = addr;
 
   // If the bit offset lies beyond the normal remembered set range, it is in
   // the extra remembered set area of a large object.
-  masm->cmpq(pointer_offset, Immediate(Page::kPageSize / kPointerSize));
+  cmpq(pointer_offset, Immediate(Page::kPageSize / kPointerSize));
-  masm->j(less, &fast);
+  j(less, &fast);
 
   // Adjust 'page_start' so that addressing using 'pointer_offset' hits the
   // extra remembered set after the large object.
 
   // Load the array length into 'scratch'.
-  masm->movl(scratch,
+  movl(scratch,
              Operand(page_start,
                      Page::kObjectStartOffset + FixedArray::kLengthOffset));
   Register array_length = scratch;
@@ -111,7 +110,7 @@ static void RecordWriteHelper(MacroAssembler* masm,
   // Add the delta between the end of the normal RSet and the start of the
   // extra RSet to 'page_start', so that addressing the bit using
   // 'pointer_offset' hits the extra RSet words.
-  masm->lea(page_start,
+  lea(page_start,
             Operand(page_start, array_length, times_pointer_size,
                     Page::kObjectStartOffset + FixedArray::kHeaderSize
                         - Page::kRSetEndOffset));
@@ -120,50 +119,8 @@ static void RecordWriteHelper(MacroAssembler* masm,
   // to limit code size. We should probably evaluate this decision by
   // measuring the performance of an equivalent implementation using
   // "simpler" instructions
-  masm->bind(&fast);
+  bind(&fast);
-  masm->bts(Operand(page_start, Page::kRSetOffset), pointer_offset);
+  bts(Operand(page_start, Page::kRSetOffset), pointer_offset);
-}
-
-
-class RecordWriteStub : public CodeStub {
- public:
-  RecordWriteStub(Register object, Register addr, Register scratch)
-      : object_(object), addr_(addr), scratch_(scratch) { }
-
-  void Generate(MacroAssembler* masm);
-
- private:
-  Register object_;
-  Register addr_;
-  Register scratch_;
-
-#ifdef DEBUG
-  void Print() {
-    PrintF("RecordWriteStub (object reg %d), (addr reg %d), (scratch reg %d)\n",
-           object_.code(), addr_.code(), scratch_.code());
-  }
-#endif
-
-  // Minor key encoding in 12 bits of three registers (object, address and
-  // scratch) OOOOAAAASSSS.
-  class ScratchBits : public BitField<uint32_t, 0, 4> {};
-  class AddressBits : public BitField<uint32_t, 4, 4> {};
-  class ObjectBits : public BitField<uint32_t, 8, 4> {};
-
-  Major MajorKey() { return RecordWrite; }
-
-  int MinorKey() {
-    // Encode the registers.
-    return ObjectBits::encode(object_.code()) |
-           AddressBits::encode(addr_.code()) |
-           ScratchBits::encode(scratch_.code());
-  }
-};
-
-
-void RecordWriteStub::Generate(MacroAssembler* masm) {
-  RecordWriteHelper(masm, object_, addr_, scratch_);
-  masm->ret(0);
 }
 
 
@@ -279,7 +236,7 @@ void MacroAssembler::RecordWriteNonSmi(Register object,
     // If we are already generating a shared stub, not inlining the
     // record write code isn't going to save us any memory.
     if (generating_stub()) {
-      RecordWriteHelper(this, object, dst, scratch);
+      RecordWriteHelper(object, dst, scratch);
     } else {
       RecordWriteStub stub(object, dst, scratch);
       CallStub(&stub);

src/x64/macro-assembler-x64.h

@@ -66,6 +66,10 @@ class MacroAssembler: public Assembler {
   // ---------------------------------------------------------------------------
   // GC Support
 
+  void RecordWriteHelper(Register object,
+                         Register addr,
+                         Register scratch);
+
   // Check if object is in new space. The condition cc can be equal or
   // not_equal. If it is equal a jump will be done if the object is on new
   // space. The register scratch can be object itself, but it will be clobbered.