Ported SubStringStub to X64.

Review URL: http://codereview.chromium.org/555049

git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@3683 ce2b1a6d-e550-0410-aec6-3dcde31c8c00

src/x64/assembler-x64.cc

@@ -1537,6 +1537,40 @@ void Assembler::movzxwl(Register dst, const Operand& src) {
 }
 
 
+void Assembler::repmovsb() {
+  EnsureSpace ensure_space(this);
+  last_pc_ = pc_;
+  emit(0xF3);
+  emit(0xA4);
+}
+
+
+void Assembler::repmovsw() {
+  EnsureSpace ensure_space(this);
+  last_pc_ = pc_;
+  emit(0x66);  // Operand size override.
+  emit(0xF3);
+  emit(0xA4);
+}
+
+
+void Assembler::repmovsl() {
+  EnsureSpace ensure_space(this);
+  last_pc_ = pc_;
+  emit(0xF3);
+  emit(0xA5);
+}
+
+
+void Assembler::repmovsq() {
+  EnsureSpace ensure_space(this);
+  last_pc_ = pc_;
+  emit(0xF3);
+  emit_rex_64();
+  emit(0xA5);
+}
+
+
 void Assembler::mul(Register src) {
   EnsureSpace ensure_space(this);
   last_pc_ = pc_;

src/x64/assembler-x64.h

@@ -574,6 +574,13 @@ class Assembler : public Malloced {
   void movzxwq(Register dst, const Operand& src);
   void movzxwl(Register dst, const Operand& src);
 
+  // Repeated moves.
+
+  void repmovsb();
+  void repmovsw();
+  void repmovsl();
+  void repmovsq();
+
   // New x64 instruction to load from an immediate 64-bit pointer into RAX.
   void load_rax(void* ptr, RelocInfo::Mode rmode);
   void load_rax(ExternalReference ext);

src/x64/codegen-x64.cc

@@ -3942,7 +3942,8 @@ void CodeGenerator::GenerateSubString(ZoneList<Expression*>* args) {
   Load(args->at(1));
   Load(args->at(2));
 
-  Result answer = frame_->CallRuntime(Runtime::kSubString, 3);
+  SubStringStub stub;
+  Result answer = frame_->CallStub(&stub, 3);
   frame_->Push(&answer);
 }
 
@@ -8138,11 +8139,11 @@ void StringAddStub::Generate(MacroAssembler* masm) {
 }
 
 
-void StringAddStub::GenerateCopyCharacters(MacroAssembler* masm,
-                                           Register dest,
-                                           Register src,
-                                           Register count,
-                                           bool ascii) {
+void StringStubBase::GenerateCopyCharacters(MacroAssembler* masm,
+                                            Register dest,
+                                            Register src,
+                                            Register count,
+                                            bool ascii) {
   Label loop;
   __ bind(&loop);
   // This loop just copies one character at a time, as it is only used for very
@@ -8163,6 +8164,174 @@ void StringAddStub::GenerateCopyCharacters(MacroAssembler* masm,
 }
 
 
+void StringStubBase::GenerateCopyCharactersREP(MacroAssembler* masm,
+                                               Register dest,
+                                               Register src,
+                                               Register count,
+                                               bool ascii) {
+  // Copy characters using rep movs of doublewords. Align destination on 4 byte
+  // boundary before starting rep movs. Copy remaining characters after running
+  // rep movs.
+  ASSERT(dest.is(rdi));  // rep movs destination
+  ASSERT(src.is(rsi));  // rep movs source
+  ASSERT(count.is(rcx));  // rep movs count
+
+  // Nothing to do for zero characters.
+  Label done;
+  __ testq(count, count);
+  __ j(zero, &done);
+
+  // Make count the number of bytes to copy.
+  if (!ascii) {
+    ASSERT_EQ(2, sizeof(uc16));  // NOLINT
+    __ addq(count, count);
+  }
+
+  // Don't enter the rep movs if there are less than 4 bytes to copy.
+  Label last_bytes;
+  __ testq(count, Immediate(~7));
+  __ j(zero, &last_bytes);
+
+  // Copy from edi to esi using rep movs instruction.
+  __ movq(kScratchRegister, count);
+  __ sar(count, Immediate(3));  // Number of doublewords to copy.
+  __ repmovsq();
+
+  // Find number of bytes left.
+  __ movq(count, kScratchRegister);
+  __ and_(count, Immediate(7));
+
+  // Check if there are more bytes to copy.
+  __ bind(&last_bytes);
+  __ testq(count, count);
+  __ j(zero, &done);
+
+  // Copy remaining characters.
+  Label loop;
+  __ bind(&loop);
+  __ movb(kScratchRegister, Operand(src, 0));
+  __ movb(Operand(dest, 0), kScratchRegister);
+  __ addq(src, Immediate(1));
+  __ addq(dest, Immediate(1));
+  __ subq(count, Immediate(1));
+  __ j(not_zero, &loop);
+
+  __ bind(&done);
+}
+
+
+void SubStringStub::Generate(MacroAssembler* masm) {
+  Label runtime;
+
+  // Stack frame on entry.
+  //  rsp[0]: return address
+  //  rsp[8]: to
+  //  rsp[16]: from
+  //  rsp[24]: string
+
+  const int kToOffset = 1 * kPointerSize;
+  const int kFromOffset = kToOffset + kPointerSize;
+  const int kStringOffset = kFromOffset + kPointerSize;
+  const int kArgumentsSize = (kStringOffset + kPointerSize) - kToOffset;
+
+  // Make sure first argument is a string.
+  __ movq(rax, Operand(rsp, kStringOffset));
+  ASSERT_EQ(0, kSmiTag);
+  __ testl(rax, Immediate(kSmiTagMask));
+  __ j(zero, &runtime);
+  Condition is_string = masm->IsObjectStringType(rax, rbx, rbx);
+  __ j(NegateCondition(is_string), &runtime);
+
+  // rax: string
+  // rbx: instance type
+  // Calculate length of sub string using the smi values.
+  __ movq(rcx, Operand(rsp, kToOffset));
+  __ movq(rdx, Operand(rsp, kFromOffset));
+  __ JumpIfNotBothPositiveSmi(rcx, rdx, &runtime);
+
+  __ SmiSub(rcx, rcx, rdx, NULL);  // Overflow doesn't happen.
+  __ j(negative, &runtime);
+  // Handle sub-strings of length 2 and less in the runtime system.
+  __ SmiToInteger32(rcx, rcx);
+  __ cmpl(rcx, Immediate(2));
+  __ j(below_equal, &runtime);
+
+  // rax: string
+  // rbx: instance type
+  // rcx: result string length
+  // Check for flat ascii string
+  Label non_ascii_flat;
+  __ and_(rbx, Immediate(kStringRepresentationMask | kStringEncodingMask));
+  __ cmpb(rbx, Immediate(kSeqStringTag | kAsciiStringTag));
+  __ j(not_equal, &non_ascii_flat);
+
+  // Allocate the result.
+  __ AllocateAsciiString(rax, rcx, rbx, rdx, rdi, &runtime);
+
+  // rax: result string
+  // rcx: result string length
+  __ movq(rdx, rsi);  // esi used by following code.
+  // Locate first character of result.
+  __ lea(rdi, FieldOperand(rax, SeqAsciiString::kHeaderSize));
+  // Load string argument and locate character of sub string start.
+  __ movq(rsi, Operand(rsp, kStringOffset));
+  __ movq(rbx, Operand(rsp, kFromOffset));
+  {
+    SmiIndex smi_as_index = masm->SmiToIndex(rbx, rbx, times_1);
+    __ lea(rsi, Operand(rsi, smi_as_index.reg, smi_as_index.scale,
+                        SeqAsciiString::kHeaderSize - kHeapObjectTag));
+  }
+
+  // rax: result string
+  // rcx: result length
+  // rdx: original value of rsi
+  // rdi: first character of result
+  // rsi: character of sub string start
+  GenerateCopyCharactersREP(masm, rdi, rsi, rcx, true);
+  __ movq(rsi, rdx);  // Restore rsi.
+  __ IncrementCounter(&Counters::sub_string_native, 1);
+  __ ret(kArgumentsSize);
+
+  __ bind(&non_ascii_flat);
+  // rax: string
+  // rbx: instance type & kStringRepresentationMask | kStringEncodingMask
+  // rcx: result string length
+  // Check for sequential two byte string
+  __ cmpb(rbx, Immediate(kSeqStringTag | kTwoByteStringTag));
+  __ j(not_equal, &runtime);
+
+  // Allocate the result.
+  __ AllocateTwoByteString(rax, rcx, rbx, rdx, rdi, &runtime);
+
+  // rax: result string
+  // rcx: result string length
+  __ movq(rdx, rsi);  // esi used by following code.
+  // Locate first character of result.
+  __ lea(rdi, FieldOperand(rax, SeqTwoByteString::kHeaderSize));
+  // Load string argument and locate character of sub string start.
+  __ movq(rsi, Operand(rsp, kStringOffset));
+  __ movq(rbx, Operand(rsp, kFromOffset));
+  {
+    SmiIndex smi_as_index = masm->SmiToIndex(rbx, rbx, times_2);
+    __ lea(rsi, Operand(rsi, smi_as_index.reg, smi_as_index.scale,
+                        SeqAsciiString::kHeaderSize - kHeapObjectTag));
+  }
+
+  // rax: result string
+  // rcx: result length
+  // rdx: original value of rsi
+  // rdi: first character of result
+  // rsi: character of sub string start
+  GenerateCopyCharactersREP(masm, rdi, rsi, rcx, false);
+  __ movq(rsi, rdx);  // Restore esi.
+  __ IncrementCounter(&Counters::sub_string_native, 1);
+  __ ret(kArgumentsSize);
+
+  // Just jump to runtime to create the sub string.
+  __ bind(&runtime);
+  __ TailCallRuntime(ExternalReference(Runtime::kSubString), 3, 1);
+}
+
 
 void StringCompareStub::GenerateCompareFlatAsciiStrings(MacroAssembler* masm,
                                                         Register left,

src/x64/codegen-x64.h

@@ -713,6 +713,29 @@ class GenericBinaryOpStub: public CodeStub {
 };
 
 
+class StringStubBase: public CodeStub {
+ public:
+  // Generate code for copying characters using a simple loop. This should only
+  // be used in places where the number of characters is small and the
+  // additional setup and checking in GenerateCopyCharactersREP adds too much
+  // overhead. Copying of overlapping regions is not supported.
+  void GenerateCopyCharacters(MacroAssembler* masm,
+                              Register dest,
+                              Register src,
+                              Register count,
+                              bool ascii);
+
+  // Generate code for copying characters using the rep movs instruction.
+  // Copies rcx characters from rsi to rdi. Copying of overlapping regions is
+  // not supported.
+  void GenerateCopyCharactersREP(MacroAssembler* masm,
+                                 Register dest,     // Must be rdi.
+                                 Register src,      // Must be rsi.
+                                 Register count,    // Must be rcx.
+                                 bool ascii);
+};
+
+
 // Flag that indicates how to generate code for the stub StringAddStub.
 enum StringAddFlags {
   NO_STRING_ADD_FLAGS = 0,
@@ -720,7 +743,7 @@ enum StringAddFlags {
 };
 
 
-class StringAddStub: public CodeStub {
+class StringAddStub: public StringStubBase {
  public:
   explicit StringAddStub(StringAddFlags flags) {
     string_check_ = ((flags & NO_STRING_CHECK_IN_STUB) == 0);
@@ -732,17 +755,23 @@ class StringAddStub: public CodeStub {
 
   void Generate(MacroAssembler* masm);
 
-  void GenerateCopyCharacters(MacroAssembler* masm,
-                              Register desc,
-                              Register src,
-                              Register count,
-                              bool ascii);
-
   // Should the stub check whether arguments are strings?
   bool string_check_;
 };
 
 
+class SubStringStub: public StringStubBase {
+ public:
+  SubStringStub() {}
+
+ private:
+  Major MajorKey() { return SubString; }
+  int MinorKey() { return 0; }
+
+  void Generate(MacroAssembler* masm);
+};
+
+
 class StringCompareStub: public CodeStub {
  public:
   explicit StringCompareStub() {}

src/x64/disasm-x64.cc

@@ -114,6 +114,10 @@ static ByteMnemonic zero_operands_instr[] = {
   { 0x9E, UNSET_OP_ORDER, "sahf" },
   { 0x99, UNSET_OP_ORDER, "cdq" },
   { 0x9B, UNSET_OP_ORDER, "fwait" },
+  { 0xA4, UNSET_OP_ORDER, "movs" },
+  { 0xA5, UNSET_OP_ORDER, "movs" },
+  { 0xA6, UNSET_OP_ORDER, "cmps" },
+  { 0xA7, UNSET_OP_ORDER, "cmps" },
   { -1, UNSET_OP_ORDER, "" }
 };
 
@@ -157,6 +161,16 @@ enum InstructionType {
 };
 
 
+enum Prefixes {
+  ESCAPE_PREFIX = 0x0F,
+  OPERAND_SIZE_OVERRIDE_PREFIX = 0x66,
+  ADDRESS_SIZE_OVERRIDE_PREFIX = 0x67,
+  REPNE_PREFIX = 0xF2,
+  REP_PREFIX = 0xF3,
+  REPEQ_PREFIX = REP_PREFIX
+};
+
+
 struct InstructionDesc {
   const char* mnem;
   InstructionType type;
@@ -1128,12 +1142,12 @@ int DisassemblerX64::InstructionDecode(v8::internal::Vector<char> out_buffer,
   // Scan for prefixes.
   while (true) {
     current = *data;
-    if (current == 0x66) {  // Group 3 prefix.
+    if (current == OPERAND_SIZE_OVERRIDE_PREFIX) {  // Group 3 prefix.
       operand_size_ = current;
     } else if ((current & 0xF0) == 0x40) {  // REX prefix.
       setRex(current);
       if (rex_w()) AppendToBuffer("REX.W ");
-    } else if ((current & 0xFE) == 0xF2) {  // Group 1 prefix.
+    } else if ((current & 0xFE) == 0xF2) {  // Group 1 prefix (0xF2 or 0xF3).
       group_1_prefix_ = current;
     } else {  // Not a prefix - an opcode.
       break;
@@ -1145,7 +1159,17 @@ int DisassemblerX64::InstructionDecode(v8::internal::Vector<char> out_buffer,
   byte_size_operand_ = idesc.byte_size_operation;
   switch (idesc.type) {
     case ZERO_OPERANDS_INSTR:
-      AppendToBuffer(idesc.mnem);
+      if (current >= 0xA4 && current <= 0xA7) {
+        // String move or compare operations.
+        if (group_1_prefix_ == REP_PREFIX) {
+          // REP.
+          AppendToBuffer("rep ");
+        }
+        if (rex_w()) AppendToBuffer("REX.W ");
+        AppendToBuffer("%s%c", idesc.mnem, operand_size_code());
+      } else {
+        AppendToBuffer("%s", idesc.mnem, operand_size_code());
+      }
       data++;
       break;
 

src/x64/macro-assembler-x64.cc

@@ -581,6 +581,20 @@ Condition MacroAssembler::CheckBothSmi(Register first, Register second) {
 }
 
 
+Condition MacroAssembler::CheckBothPositiveSmi(Register first,
+                                               Register second) {
+  if (first.is(second)) {
+    return CheckPositiveSmi(first);
+  }
+  movl(kScratchRegister, first);
+  orl(kScratchRegister, second);
+  rol(kScratchRegister, Immediate(1));
+  testl(kScratchRegister, Immediate(0x03));
+  return zero;
+}
+
+
+
 Condition MacroAssembler::CheckEitherSmi(Register first, Register second) {
   if (first.is(second)) {
     return CheckSmi(first);
@@ -660,7 +674,17 @@ void MacroAssembler::SmiSub(Register dst,
                             Register src2,
                             Label* on_not_smi_result) {
   ASSERT(!dst.is(src2));
-  if (dst.is(src1)) {
+  if (on_not_smi_result == NULL) {
+    // 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 {
+      movq(dst, src1);
+      subq(dst, src2);
+    }
+    Assert(no_overflow, "Smi substraction onverflow");
+  } else if (dst.is(src1)) {
     subq(dst, src2);
     Label smi_result;
     j(no_overflow, &smi_result);
@@ -1292,6 +1316,14 @@ void MacroAssembler::JumpIfNotBothSmi(Register src1, Register src2,
 }
 
 
+void MacroAssembler::JumpIfNotBothPositiveSmi(Register src1, Register src2,
+                                              Label* on_not_both_smi) {
+  Condition both_smi = CheckBothPositiveSmi(src1, src2);
+  j(NegateCondition(both_smi), on_not_both_smi);
+}
+
+
+
 void MacroAssembler::JumpIfNotBothSequentialAsciiStrings(Register first_object,
                                                          Register second_object,
                                                          Register scratch1,
@@ -1517,6 +1549,17 @@ void MacroAssembler::CmpInstanceType(Register map, InstanceType type) {
 }
 
 
+Condition MacroAssembler::IsObjectStringType(Register heap_object,
+                                             Register map,
+                                             Register instance_type) {
+  movq(map, FieldOperand(heap_object, HeapObject::kMapOffset));
+  movzxbq(instance_type, FieldOperand(map, Map::kInstanceTypeOffset));
+  ASSERT(kNotStringTag != 0);
+  testb(instance_type, Immediate(kIsNotStringMask));
+  return zero;
+}
+
+
 void MacroAssembler::TryGetFunctionPrototype(Register function,
                                              Register result,
                                              Label* miss) {

src/x64/macro-assembler-x64.h

@@ -207,6 +207,9 @@ class MacroAssembler: public Assembler {
   // Are both values tagged smis.
   Condition CheckBothSmi(Register first, Register second);
 
+  // Are both values tagged smis.
+  Condition CheckBothPositiveSmi(Register first, Register second);
+
   // Are either value a tagged smi.
   Condition CheckEitherSmi(Register first, Register second);
 
@@ -248,6 +251,10 @@ class MacroAssembler: public Assembler {
   // Jump if either or both register are not smi values.
   void JumpIfNotBothSmi(Register src1, Register src2, Label* on_not_both_smi);
 
+  // Jump if either or both register are not positive smi values.
+  void JumpIfNotBothPositiveSmi(Register src1, Register src2,
+                                Label* on_not_both_smi);
+
   // Operations on tagged smi values.
 
   // Smis represent a subset of integers. The subset is always equivalent to
@@ -452,6 +459,15 @@ class MacroAssembler: public Assembler {
   // Always use unsigned comparisons: above and below, not less and greater.
   void CmpInstanceType(Register map, InstanceType type);
 
+  // Check if the object in register heap_object is a string. Afterwards the
+  // register map contains the object map and the register instance_type
+  // contains the instance_type. The registers map and instance_type can be the
+  // same in which case it contains the instance type afterwards. Either of the
+  // registers map and instance_type can be the same as heap_object.
+  Condition IsObjectStringType(Register heap_object,
+                               Register map,
+                               Register instance_type);
+
   // FCmp is similar to integer cmp, but requires unsigned
   // jcc instructions (je, ja, jae, jb, jbe, je, and jz).
   void FCmp();