Implemented FastAsciiStringJoin in X64 full codegen.

Review URL: http://codereview.chromium.org/6729016

git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@7345 ce2b1a6d-e550-0410-aec6-3dcde31c8c00

src/x64/assembler-x64.cc

@@ -944,6 +944,12 @@ void Assembler::clc() {
   emit(0xF8);
 }
 
+void Assembler::cld() {
+  EnsureSpace ensure_space(this);
+  last_pc_ = pc_;
+  emit(0xFC);
+}
+
 void Assembler::cdq() {
   EnsureSpace ensure_space(this);
   last_pc_ = pc_;

src/x64/assembler-x64.h

@@ -1137,6 +1137,7 @@ class Assembler : public AssemblerBase {
 
   // Miscellaneous
   void clc();
+  void cld();
   void cpuid();
   void hlt();
   void int3();

src/x64/macro-assembler-x64.cc

@@ -909,9 +909,9 @@ void MacroAssembler::Integer64PlusConstantToSmi(Register dst,
                                                 Register src,
                                                 int constant) {
   if (dst.is(src)) {
-    addq(dst, Immediate(constant));
+    addl(dst, Immediate(constant));
   } else {
-    lea(dst, Operand(src, constant));
+    leal(dst, Operand(src, constant));
   }
   shl(dst, Immediate(kSmiShift));
 }
@@ -1245,12 +1245,10 @@ void MacroAssembler::SmiAdd(Register dst,
   // No overflow checking. Use only when it's known that
   // overflowing is impossible.
   ASSERT(!dst.is(src2));
-  if (dst.is(src1)) {
-    addq(dst, src2);
-  } else {
+  if (!dst.is(src1)) {
     movq(dst, src1);
-    addq(dst, src2);
   }
+  addq(dst, src2);
   Assert(no_overflow, "Smi addition overflow");
 }
 
@@ -1259,12 +1257,10 @@ void MacroAssembler::SmiSub(Register dst, Register src1, Register src2) {
   // No overflow checking. Use only when it's known that
   // overflowing is impossible (e.g., subtracting two positive smis).
   ASSERT(!dst.is(src2));
-  if (dst.is(src1)) {
-    subq(dst, src2);
-  } else {
+  if (!dst.is(src1)) {
     movq(dst, src1);
-    subq(dst, src2);
   }
+  subq(dst, src2);
   Assert(no_overflow, "Smi subtraction overflow");
 }
 
@@ -1274,12 +1270,10 @@ void MacroAssembler::SmiSub(Register dst,
                             const Operand& src2) {
   // No overflow checking. Use only when it's known that
   // overflowing is impossible (e.g., subtracting two positive smis).
-  if (dst.is(src1)) {
-    subq(dst, src2);
-  } else {
+  if (!dst.is(src1)) {
     movq(dst, src1);
-    subq(dst, src2);
   }
+  subq(dst, src2);
   Assert(no_overflow, "Smi subtraction overflow");
 }
 
@@ -1466,6 +1460,13 @@ SmiIndex MacroAssembler::SmiToNegativeIndex(Register dst,
 }
 
 
+void MacroAssembler::AddSmiField(Register dst, const Operand& src) {
+  ASSERT_EQ(0, kSmiShift % kBitsPerByte);
+  addl(dst, Operand(src, kSmiShift / kBitsPerByte));
+}
+
+
+
 void MacroAssembler::Move(Register dst, Register src) {
   if (!dst.is(src)) {
     movq(dst, src);
@@ -2701,6 +2702,70 @@ void MacroAssembler::AllocateAsciiConsString(Register result,
 }
 
 
+// Copy memory, byte-by-byte, from source to destination.  Not optimized for
+// long or aligned copies.  The contents of scratch and length are destroyed.
+// Destination is incremented by length, source, length and scratch are
+// clobbered.
+// A simpler loop is faster on small copies, but slower on large ones.
+// The cld() instruction must have been emitted, to set the direction flag(),
+// before calling this function.
+void MacroAssembler::CopyBytes(Register destination,
+                               Register source,
+                               Register length,
+                               int min_length,
+                               Register scratch) {
+  ASSERT(min_length >= 0);
+  if (FLAG_debug_code) {
+    cmpl(length, Immediate(min_length));
+    Assert(greater_equal, "Invalid min_length");
+  }
+  Label loop, done, short_string, short_loop;
+
+  const int kLongStringLimit = 20;
+  if (min_length <= kLongStringLimit) {
+    cmpl(length, Immediate(kLongStringLimit));
+    j(less_equal, &short_string);
+  }
+
+  ASSERT(source.is(rsi));
+  ASSERT(destination.is(rdi));
+  ASSERT(length.is(rcx));
+
+  // Because source is 8-byte aligned in our uses of this function,
+  // we keep source aligned for the rep movs operation by copying the odd bytes
+  // at the end of the ranges.
+  movq(scratch, length);
+  shrl(length, Immediate(3));
+  repmovsq();
+  // Move remaining bytes of length.
+  andl(scratch, Immediate(0x7));
+  movq(length, Operand(source, scratch, times_1, -8));
+  movq(Operand(destination, scratch, times_1, -8), length);
+  addq(destination, scratch);
+
+  if (min_length <= kLongStringLimit) {
+    jmp(&done);
+
+    bind(&short_string);
+    if (min_length == 0) {
+      testl(length, length);
+      j(zero, &done);
+    }
+    lea(scratch, Operand(destination, length, times_1, 0));
+
+    bind(&short_loop);
+    movb(length, Operand(source, 0));
+    movb(Operand(destination, 0), length);
+    incq(source);
+    incq(destination);
+    cmpq(destination, scratch);
+    j(not_equal, &short_loop);
+
+    bind(&done);
+  }
+}
+
+
 void MacroAssembler::LoadContext(Register dst, int context_chain_length) {
   if (context_chain_length > 0) {
     // Move up the chain of contexts to the context containing the slot.

src/x64/macro-assembler-x64.h

@@ -470,6 +470,11 @@ class MacroAssembler: public Assembler {
               Register src1,
               Register src2,
               LabelType* on_not_smi_result);
+  template <typename LabelType>
+  void SmiAdd(Register dst,
+              Register src1,
+              const Operand& src2,
+              LabelType* on_not_smi_result);
 
   void SmiAdd(Register dst,
               Register src1,
@@ -590,6 +595,10 @@ class MacroAssembler: public Assembler {
   // Converts a positive smi to a negative index.
   SmiIndex SmiToNegativeIndex(Register dst, Register src, int shift);
 
+  // Add the value of a smi in memory to an int32 register.
+  // Sets flags as a normal add.
+  void AddSmiField(Register dst, const Operand& src);
+
   // Basic Smi operations.
   void Move(Register dst, Smi* source) {
     LoadSmiConstant(dst, source);
@@ -1022,6 +1031,18 @@ class MacroAssembler: public Assembler {
 
   Handle<Object> CodeObject() { return code_object_; }
 
+  // Copy length bytes from source to destination.
+  // Uses scratch register internally (if you have a low-eight register
+  // free, do use it, otherwise kScratchRegister will be used).
+  // The min_length is a minimum limit on the value that length will have.
+  // The algorithm has some special cases that might be omitted if the string
+  // is known to always be long.
+  void CopyBytes(Register destination,
+                 Register source,
+                 Register length,
+                 int min_length = 0,
+                 Register scratch = kScratchRegister);
+
 
   // ---------------------------------------------------------------------------
   // StatsCounter support
@@ -1272,6 +1293,26 @@ void MacroAssembler::SmiAdd(Register dst,
 }
 
 
+template <typename LabelType>
+void MacroAssembler::SmiAdd(Register dst,
+                            Register src1,
+                            const Operand& src2,
+                            LabelType* on_not_smi_result) {
+  ASSERT_NOT_NULL(on_not_smi_result);
+  if (dst.is(src1)) {
+    movq(kScratchRegister, src1);
+    addq(kScratchRegister, src2);
+    j(overflow, on_not_smi_result);
+    movq(dst, kScratchRegister);
+  } else {
+    ASSERT(!src2.AddressUsesRegister(dst));
+    movq(dst, src1);
+    addq(dst, src2);
+    j(overflow, on_not_smi_result);
+  }
+}
+
+
 template <typename LabelType>
 void MacroAssembler::SmiSub(Register dst,
                             Register src1,