Speed up FastAsciiArrayJoin on ia32 by improving hand-written assembly code.

Review URL: http://codereview.chromium.org/6148007

git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@6310 ce2b1a6d-e550-0410-aec6-3dcde31c8c00

src/ia32/codegen-ia32.cc

@@ -6649,38 +6649,41 @@ void CodeGenerator::GenerateIsArray(ZoneList<Expression*>* args) {
 
 
 void CodeGenerator::GenerateFastAsciiArrayJoin(ZoneList<Expression*>* args) {
+  Label bailout, done, one_char_separator, long_separator,
+      non_trivial_array, not_size_one_array, loop, loop_condition,
+      loop_1, loop_1_condition, loop_2, loop_2_entry, loop_3, loop_3_entry;
+
   ASSERT(args->length() == 2);
+  // We will leave the separator on the stack until the end of the function.
   Load(args->at(1));
+  // Load this to eax (= array)
   Load(args->at(0));
   Result array_result = frame_->Pop();
   array_result.ToRegister(eax);
   frame_->SpillAll();
 
-  Label bailout;
-  Label done;
   // All aliases of the same register have disjoint lifetimes.
   Register array = eax;
-  Register result_pos = no_reg;
+  Register elements = no_reg;  // Will be eax.
 
-  Register index = edi;
+  Register index = edx;
 
-  Register current_string_length = ecx;  // Will be ecx when live.
+  Register string_length = ecx;
 
-  Register current_string = edx;
+  Register string = esi;
 
   Register scratch = ebx;
 
-  Register scratch_2 = esi;
+  Register array_length = edi;
-  Register new_padding_chars = scratch_2;
+  Register result_pos = no_reg;  // Will be edi.
 
-  Operand separator = Operand(esp, 4 * kPointerSize);  // Already pushed.
+  // Separator operand is already pushed.
-  Operand elements = Operand(esp, 3 * kPointerSize);
+  Operand separator_operand = Operand(esp, 2 * kPointerSize);
-  Operand result = Operand(esp, 2 * kPointerSize);
+  Operand result_operand = Operand(esp, 1 * kPointerSize);
-  Operand padding_chars = Operand(esp, 1 * kPointerSize);
+  Operand array_length_operand = Operand(esp, 0);
-  Operand array_length = Operand(esp, 0);
+  __ sub(Operand(esp), Immediate(2 * kPointerSize));
-  __ sub(Operand(esp), Immediate(4 * kPointerSize));
+  __ cld();
-
+  // Check that the array is a JSArray
-  // Check that eax is a JSArray
   __ test(array, Immediate(kSmiTagMask));
   __ j(zero, &bailout);
   __ CmpObjectType(array, JS_ARRAY_TYPE, scratch);
@@ -6691,140 +6694,226 @@ void CodeGenerator::GenerateFastAsciiArrayJoin(ZoneList<Expression*>* args) {
             1 << Map::kHasFastElements);
   __ j(zero, &bailout);
 
-  // If the array is empty, return the empty string.
+  // If the array has length zero, return the empty string.
-  __ mov(scratch, FieldOperand(array, JSArray::kLengthOffset));
+  __ mov(array_length, FieldOperand(array, JSArray::kLengthOffset));
-  __ sar(scratch, 1);
+  __ sar(array_length, 1);
-  Label non_trivial;
+  __ j(not_zero, &non_trivial_array);
-  __ j(not_zero, &non_trivial);
+  __ mov(result_operand, Factory::empty_string());
-  __ mov(result, Factory::empty_string());
   __ jmp(&done);
 
-  __ bind(&non_trivial);
+  // Save the array length.
-  __ mov(array_length, scratch);
+  __ bind(&non_trivial_array);
-
+  __ mov(array_length_operand, array_length);
-  __ mov(scratch, FieldOperand(array, JSArray::kElementsOffset));
-  __ mov(elements, scratch);
 
+  // Save the FixedArray containing array's elements.
   // End of array's live range.
-  result_pos = array;
+  elements = array;
+  __ mov(elements, FieldOperand(array, JSArray::kElementsOffset));
   array = no_reg;
 
 
-  // Check that the separator is a flat ascii string.
+  // Check that all array elements are sequential ASCII strings, and
-  __ mov(current_string, separator);
+  // accumulate the sum of their lengths, as a smi-encoded value.
-  __ test(current_string, Immediate(kSmiTagMask));
+  __ Set(index, Immediate(0));
-  __ j(zero, &bailout);
+  __ Set(string_length, Immediate(0));
-  __ mov(scratch, FieldOperand(current_string, HeapObject::kMapOffset));
-  __ mov_b(scratch, FieldOperand(scratch, Map::kInstanceTypeOffset));
-  __ and_(scratch, Immediate(
-      kIsNotStringMask | kStringEncodingMask | kStringRepresentationMask));
-  __ cmp(scratch, kStringTag | kAsciiStringTag | kSeqStringTag);
-  __ j(not_equal, &bailout);
-  // If the separator is the empty string, replace it with NULL.
-  // The test for NULL is quicker than the empty string test, in a loop.
-  __ cmp(FieldOperand(current_string, SeqAsciiString::kLengthOffset),
-         Immediate(0));
-  Label separator_checked;
-  __ j(not_zero, &separator_checked);
-  __ mov(separator, Immediate(0));
-  __ bind(&separator_checked);
-
-  // Check that elements[0] is a flat ascii string, and copy it in new space.
-  __ mov(scratch, elements);
-  __ mov(current_string, FieldOperand(scratch, FixedArray::kHeaderSize));
-  __ test(current_string, Immediate(kSmiTagMask));
-  __ j(zero, &bailout);
-  __ mov(scratch, FieldOperand(current_string, HeapObject::kMapOffset));
-  __ mov_b(scratch, FieldOperand(scratch, Map::kInstanceTypeOffset));
-  __ and_(scratch, Immediate(
-      kIsNotStringMask | kStringEncodingMask | kStringRepresentationMask));
-  __ cmp(scratch, kStringTag | kAsciiStringTag | kSeqStringTag);
-  __ j(not_equal, &bailout);
-
-  // Allocate space to copy it.  Round up the size to the alignment granularity.
-  __ mov(current_string_length,
-         FieldOperand(current_string, String::kLengthOffset));
-  __ shr(current_string_length, 1);
-
-  // Live registers and stack values:
-  //   current_string_length: length of elements[0].
-
-  // New string result in new space = elements[0]
-  __ AllocateAsciiString(result_pos, current_string_length, scratch_2,
-                         index, no_reg, &bailout);
-  __ mov(result, result_pos);
-
-  // Adjust current_string_length to include padding bytes at end of string.
-  // Keep track of the number of padding bytes.
-  __ mov(new_padding_chars, current_string_length);
-  __ add(Operand(current_string_length), Immediate(kObjectAlignmentMask));
-  __ and_(Operand(current_string_length), Immediate(~kObjectAlignmentMask));
-  __ sub(new_padding_chars, Operand(current_string_length));
-  __ neg(new_padding_chars);
-  __ mov(padding_chars, new_padding_chars);
-
-  Label copy_loop_1_done;
-  Label copy_loop_1;
-  __ test(current_string_length, Operand(current_string_length));
-  __ j(zero, &copy_loop_1_done);
-  __ bind(&copy_loop_1);
-  __ sub(Operand(current_string_length), Immediate(kPointerSize));
-  __ mov(scratch, FieldOperand(current_string, current_string_length,
-                               times_1, SeqAsciiString::kHeaderSize));
-  __ mov(FieldOperand(result_pos, current_string_length,
-                      times_1, SeqAsciiString::kHeaderSize),
-         scratch);
-  __ j(not_zero, &copy_loop_1);
-  __ bind(&copy_loop_1_done);
-
-  __ mov(index, Immediate(1));
   // Loop condition: while (index < length).
-  Label loop;
+  // Live loop registers: index, array_length, string,
+  //                      scratch, string_length, elements.
+  __ jmp(&loop_condition);
   __ bind(&loop);
-  __ cmp(index, array_length);
+  __ cmp(index, Operand(array_length));
   __ j(greater_equal, &done);
 
-  // If the separator is the empty string, signalled by NULL, skip it.
+  __ mov(string, FieldOperand(elements, index,
-  Label separator_done;
-  __ mov(current_string, separator);
-  __ test(current_string, Operand(current_string));
-  __ j(zero, &separator_done);
-
-  // Append separator to result.  It is known to be a flat ascii string.
-  __ AppendStringToTopOfNewSpace(current_string, current_string_length,
-                                 result_pos, scratch, scratch_2, result,
-                                 padding_chars, &bailout);
-  __ bind(&separator_done);
-
-  // Add next element of array to the end of the result.
-  // Get current_string = array[index].
-  __ mov(scratch, elements);
-  __ mov(current_string, FieldOperand(scratch, index,
                                       times_pointer_size,
                                       FixedArray::kHeaderSize));
-  // If current != flat ascii string drop result, return undefined.
+  __ test(string, Immediate(kSmiTagMask));
-  __ test(current_string, Immediate(kSmiTagMask));
   __ j(zero, &bailout);
-  __ mov(scratch, FieldOperand(current_string, HeapObject::kMapOffset));
+  __ mov(scratch, FieldOperand(string, HeapObject::kMapOffset));
-  __ mov_b(scratch, FieldOperand(scratch, Map::kInstanceTypeOffset));
+  __ movzx_b(scratch, FieldOperand(scratch, Map::kInstanceTypeOffset));
+  __ and_(scratch, Immediate(
+      kIsNotStringMask | kStringEncodingMask | kStringRepresentationMask));
+  __ cmp(scratch, kStringTag | kAsciiStringTag | kSeqStringTag);
+  __ j(not_equal, &bailout);
+  __ add(string_length,
+         FieldOperand(string, SeqAsciiString::kLengthOffset));
+  __ j(overflow, &bailout);
+  __ add(Operand(index), Immediate(1));
+  __ bind(&loop_condition);
+  __ cmp(index, Operand(array_length));
+  __ j(less, &loop);
+
+  // If array_length is 1, return elements[0], a string.
+  __ cmp(array_length, 1);
+  __ j(not_equal, &not_size_one_array);
+  __ mov(scratch, FieldOperand(elements, FixedArray::kHeaderSize));
+  __ mov(result_operand, scratch);
+  __ jmp(&done);
+
+  __ bind(&not_size_one_array);
+
+  // End of array_length live range.
+  result_pos = array_length;
+  array_length = no_reg;
+
+  // Live registers:
+  // string_length: Sum of string lengths, as a smi.
+  // elements: FixedArray of strings.
+
+  // Check that the separator is a flat ASCII string.
+  __ mov(string, separator_operand);
+  __ test(string, Immediate(kSmiTagMask));
+  __ j(zero, &bailout);
+  __ mov(scratch, FieldOperand(string, HeapObject::kMapOffset));
+  __ movzx_b(scratch, FieldOperand(scratch, Map::kInstanceTypeOffset));
   __ and_(scratch, Immediate(
       kIsNotStringMask | kStringEncodingMask | kStringRepresentationMask));
   __ cmp(scratch, kStringTag | kAsciiStringTag | kSeqStringTag);
   __ j(not_equal, &bailout);
 
-  // Append current to the result.
+  // Add (separator length times array_length) - separator length
-  __ AppendStringToTopOfNewSpace(current_string, current_string_length,
+  // to string_length.
-                                 result_pos, scratch, scratch_2, result,
+  __ mov(scratch, separator_operand);
-                                 padding_chars, &bailout);
+  __ mov(scratch, FieldOperand(scratch, SeqAsciiString::kLengthOffset));
+  __ sub(string_length, Operand(scratch));  // May be negative, temporarily.
+  __ imul(scratch, array_length_operand);
+  __ j(overflow, &bailout);
+  __ add(string_length, Operand(scratch));
+  __ j(overflow, &bailout);
+
+  __ shr(string_length, 1);
+  // Live registers and stack values:
+  //   string_length
+  //   elements
+  __ AllocateAsciiString(result_pos, string_length, scratch,
+                         index, string, &bailout);
+  __ mov(result_operand, result_pos);
+  __ lea(result_pos, FieldOperand(result_pos, SeqAsciiString::kHeaderSize));
+
+
+  __ mov(string, separator_operand);
+  __ cmp(FieldOperand(string, SeqAsciiString::kLengthOffset),
+         Immediate(Smi::FromInt(1)));
+  __ j(equal, &one_char_separator);
+  __ j(greater, &long_separator);
+
+
+  // Empty separator case
+  __ mov(index, Immediate(0));
+  __ jmp(&loop_1_condition);
+  // Loop condition: while (index < length).
+  __ bind(&loop_1);
+  // Each iteration of the loop concatenates one string to the result.
+  // Live values in registers:
+  //   index: which element of the elements array we are adding to the result.
+  //   result_pos: the position to which we are currently copying characters.
+  //   elements: the FixedArray of strings we are joining.
+
+  // Get string = array[index].
+  __ mov(string, FieldOperand(elements, index,
+                              times_pointer_size,
+                              FixedArray::kHeaderSize));
+  __ mov(string_length,
+         FieldOperand(string, String::kLengthOffset));
+  __ shr(string_length, 1);
+  __ lea(string,
+         FieldOperand(string, SeqAsciiString::kHeaderSize));
+  __ CopyBytes(string, result_pos, string_length, scratch);
   __ add(Operand(index), Immediate(1));
-  __ jmp(&loop);  // End while (index < length).
+  __ bind(&loop_1_condition);
+  __ cmp(index, array_length_operand);
+  __ j(less, &loop_1);  // End while (index < length).
+  __ jmp(&done);
+
+
+
+  // One-character separator case
+  __ bind(&one_char_separator);
+  // Replace separator with its ascii character value.
+  __ mov_b(scratch, FieldOperand(string, SeqAsciiString::kHeaderSize));
+  __ mov_b(separator_operand, scratch);
+
+  __ Set(index, Immediate(0));
+  // Jump into the loop after the code that copies the separator, so the first
+  // element is not preceded by a separator
+  __ jmp(&loop_2_entry);
+  // Loop condition: while (index < length).
+  __ bind(&loop_2);
+  // Each iteration of the loop concatenates one string to the result.
+  // Live values in registers:
+  //   index: which element of the elements array we are adding to the result.
+  //   result_pos: the position to which we are currently copying characters.
+
+  // Copy the separator character to the result.
+  __ mov_b(scratch, separator_operand);
+  __ mov_b(Operand(result_pos, 0), scratch);
+  __ inc(result_pos);
+
+  __ bind(&loop_2_entry);
+  // Get string = array[index].
+  __ mov(string, FieldOperand(elements, index,
+                              times_pointer_size,
+                              FixedArray::kHeaderSize));
+  __ mov(string_length,
+         FieldOperand(string, String::kLengthOffset));
+  __ shr(string_length, 1);
+  __ lea(string,
+         FieldOperand(string, SeqAsciiString::kHeaderSize));
+  __ CopyBytes(string, result_pos, string_length, scratch);
+  __ add(Operand(index), Immediate(1));
+
+  __ cmp(index, array_length_operand);
+  __ j(less, &loop_2);  // End while (index < length).
+  __ jmp(&done);
+
+
+  // Long separator case (separator is more than one character).
+  __ bind(&long_separator);
+
+  __ Set(index, Immediate(0));
+  // Jump into the loop after the code that copies the separator, so the first
+  // element is not preceded by a separator
+  __ jmp(&loop_3_entry);
+  // Loop condition: while (index < length).
+  __ bind(&loop_3);
+  // Each iteration of the loop concatenates one string to the result.
+  // Live values in registers:
+  //   index: which element of the elements array we are adding to the result.
+  //   result_pos: the position to which we are currently copying characters.
+
+  // Copy the separator to the result.
+  __ mov(string, separator_operand);
+  __ mov(string_length,
+         FieldOperand(string, String::kLengthOffset));
+  __ shr(string_length, 1);
+  __ lea(string,
+         FieldOperand(string, SeqAsciiString::kHeaderSize));
+  __ CopyBytes(string, result_pos, string_length, scratch);
+
+  __ bind(&loop_3_entry);
+  // Get string = array[index].
+  __ mov(string, FieldOperand(elements, index,
+                              times_pointer_size,
+                              FixedArray::kHeaderSize));
+  __ mov(string_length,
+         FieldOperand(string, String::kLengthOffset));
+  __ shr(string_length, 1);
+  __ lea(string,
+         FieldOperand(string, SeqAsciiString::kHeaderSize));
+  __ CopyBytes(string, result_pos, string_length, scratch);
+  __ add(Operand(index), Immediate(1));
+
+  __ cmp(index, array_length_operand);
+  __ j(less, &loop_3);  // End while (index < length).
+  __ jmp(&done);
+
 
   __ bind(&bailout);
-  __ mov(result, Factory::undefined_value());
+  __ mov(result_operand, Factory::undefined_value());
   __ bind(&done);
-  __ mov(eax, result);
+  __ mov(eax, result_operand);
   // Drop temp values from the stack, and restore context register.
-  __ add(Operand(esp), Immediate(4 * kPointerSize));
+  __ add(Operand(esp), Immediate(2 * kPointerSize));
 
   __ mov(esi, Operand(ebp, StandardFrameConstants::kContextOffset));
   frame_->Drop(1);

src/ia32/full-codegen-ia32.cc

@@ -3351,39 +3351,37 @@ void FullCodeGenerator::EmitGetCachedArrayIndex(ZoneList<Expression*>* args) {
 
 
 void FullCodeGenerator::EmitFastAsciiArrayJoin(ZoneList<Expression*>* args) {
-  Label bailout;
+  Label bailout, done, one_char_separator, long_separator,
-  Label done;
+      non_trivial_array, not_size_one_array, loop, loop_condition,
+      loop_1, loop_1_condition, loop_2, loop_2_entry, loop_3, loop_3_entry;
 
   ASSERT(args->length() == 2);
   // We will leave the separator on the stack until the end of the function.
   VisitForStackValue(args->at(1));
   // Load this to eax (= array)
   VisitForAccumulatorValue(args->at(0));
-
   // All aliases of the same register have disjoint lifetimes.
   Register array = eax;
-  Register result_pos = no_reg;
+  Register elements = no_reg;  // Will be eax.
 
-  Register index = edi;
+  Register index = edx;
 
-  Register current_string_length = ecx;  // Will be ecx when live.
+  Register string_length = ecx;
 
-  Register current_string = edx;
+  Register string = esi;
 
   Register scratch = ebx;
 
-  Register scratch_2 = esi;
+  Register array_length = edi;
-  Register new_padding_chars = scratch_2;
+  Register result_pos = no_reg;  // Will be edi.
 
-  Operand separator = Operand(esp, 4 * kPointerSize);  // Already pushed.
+  // Separator operand is already pushed.
-  Operand elements = Operand(esp, 3 * kPointerSize);
+  Operand separator_operand = Operand(esp, 2 * kPointerSize);
-  Operand result = Operand(esp, 2 * kPointerSize);
+  Operand result_operand = Operand(esp, 1 * kPointerSize);
-  Operand padding_chars = Operand(esp, 1 * kPointerSize);
+  Operand array_length_operand = Operand(esp, 0);
-  Operand array_length = Operand(esp, 0);
+  __ sub(Operand(esp), Immediate(2 * kPointerSize));
-  __ sub(Operand(esp), Immediate(4 * kPointerSize));
+  __ cld();
-
+  // Check that the array is a JSArray
-
-  // Check that eax is a JSArray
   __ test(array, Immediate(kSmiTagMask));
   __ j(zero, &bailout);
   __ CmpObjectType(array, JS_ARRAY_TYPE, scratch);
@@ -3394,140 +3392,226 @@ void FullCodeGenerator::EmitFastAsciiArrayJoin(ZoneList<Expression*>* args) {
             1 << Map::kHasFastElements);
   __ j(zero, &bailout);
 
-  // If the array is empty, return the empty string.
+  // If the array has length zero, return the empty string.
-  __ mov(scratch, FieldOperand(array, JSArray::kLengthOffset));
+  __ mov(array_length, FieldOperand(array, JSArray::kLengthOffset));
-  __ sar(scratch, 1);
+  __ sar(array_length, 1);
-  Label non_trivial;
+  __ j(not_zero, &non_trivial_array);
-  __ j(not_zero, &non_trivial);
+  __ mov(result_operand, Factory::empty_string());
-  __ mov(result, Factory::empty_string());
   __ jmp(&done);
 
-  __ bind(&non_trivial);
+  // Save the array length.
-  __ mov(array_length, scratch);
+  __ bind(&non_trivial_array);
-
+  __ mov(array_length_operand, array_length);
-  __ mov(scratch, FieldOperand(array, JSArray::kElementsOffset));
-  __ mov(elements, scratch);
 
+  // Save the FixedArray containing array's elements.
   // End of array's live range.
-  result_pos = array;
+  elements = array;
+  __ mov(elements, FieldOperand(array, JSArray::kElementsOffset));
   array = no_reg;
 
 
-  // Check that the separator is a flat ascii string.
+  // Check that all array elements are sequential ASCII strings, and
-  __ mov(current_string, separator);
+  // accumulate the sum of their lengths, as a smi-encoded value.
-  __ test(current_string, Immediate(kSmiTagMask));
+  __ Set(index, Immediate(0));
-  __ j(zero, &bailout);
+  __ Set(string_length, Immediate(0));
-  __ mov(scratch, FieldOperand(current_string, HeapObject::kMapOffset));
-  __ mov_b(scratch, FieldOperand(scratch, Map::kInstanceTypeOffset));
-  __ and_(scratch, Immediate(
-      kIsNotStringMask | kStringEncodingMask | kStringRepresentationMask));
-  __ cmp(scratch, kStringTag | kAsciiStringTag | kSeqStringTag);
-  __ j(not_equal, &bailout);
-  // If the separator is the empty string, replace it with NULL.
-  // The test for NULL is quicker than the empty string test, in a loop.
-  __ cmp(FieldOperand(current_string, SeqAsciiString::kLengthOffset),
-         Immediate(0));
-  Label separator_checked;
-  __ j(not_zero, &separator_checked);
-  __ mov(separator, Immediate(0));
-  __ bind(&separator_checked);
-
-  // Check that elements[0] is a flat ascii string, and copy it in new space.
-  __ mov(scratch, elements);
-  __ mov(current_string, FieldOperand(scratch, FixedArray::kHeaderSize));
-  __ test(current_string, Immediate(kSmiTagMask));
-  __ j(zero, &bailout);
-  __ mov(scratch, FieldOperand(current_string, HeapObject::kMapOffset));
-  __ mov_b(scratch, FieldOperand(scratch, Map::kInstanceTypeOffset));
-  __ and_(scratch, Immediate(
-      kIsNotStringMask | kStringEncodingMask | kStringRepresentationMask));
-  __ cmp(scratch, kStringTag | kAsciiStringTag | kSeqStringTag);
-  __ j(not_equal, &bailout);
-
-  // Allocate space to copy it.  Round up the size to the alignment granularity.
-  __ mov(current_string_length,
-         FieldOperand(current_string, String::kLengthOffset));
-  __ shr(current_string_length, 1);
-
-  // Live registers and stack values:
-  //   current_string_length: length of elements[0].
-
-  // New string result in new space = elements[0]
-  __ AllocateAsciiString(result_pos, current_string_length, scratch_2,
-                         index, no_reg, &bailout);
-  __ mov(result, result_pos);
-
-  // Adjust current_string_length to include padding bytes at end of string.
-  // Keep track of the number of padding bytes.
-  __ mov(new_padding_chars, current_string_length);
-  __ add(Operand(current_string_length), Immediate(kObjectAlignmentMask));
-  __ and_(Operand(current_string_length), Immediate(~kObjectAlignmentMask));
-  __ sub(new_padding_chars, Operand(current_string_length));
-  __ neg(new_padding_chars);
-  __ mov(padding_chars, new_padding_chars);
-
-  Label copy_loop_1_done;
-  Label copy_loop_1;
-  __ test(current_string_length, Operand(current_string_length));
-  __ j(zero, &copy_loop_1_done);
-  __ bind(&copy_loop_1);
-  __ sub(Operand(current_string_length), Immediate(kPointerSize));
-  __ mov(scratch, FieldOperand(current_string, current_string_length,
-                               times_1, SeqAsciiString::kHeaderSize));
-  __ mov(FieldOperand(result_pos, current_string_length,
-                      times_1, SeqAsciiString::kHeaderSize),
-         scratch);
-  __ j(not_zero, &copy_loop_1);
-  __ bind(&copy_loop_1_done);
-
-  __ mov(index, Immediate(1));
   // Loop condition: while (index < length).
-  Label loop;
+  // Live loop registers: index, array_length, string,
+  //                      scratch, string_length, elements.
+  __ jmp(&loop_condition);
   __ bind(&loop);
-  __ cmp(index, array_length);
+  __ cmp(index, Operand(array_length));
   __ j(greater_equal, &done);
 
-  // If the separator is the empty string, signalled by NULL, skip it.
+  __ mov(string, FieldOperand(elements, index,
-  Label separator_done;
-  __ mov(current_string, separator);
-  __ test(current_string, Operand(current_string));
-  __ j(zero, &separator_done);
-
-  // Append separator to result.  It is known to be a flat ascii string.
-  __ AppendStringToTopOfNewSpace(current_string, current_string_length,
-                                 result_pos, scratch, scratch_2, result,
-                                 padding_chars, &bailout);
-  __ bind(&separator_done);
-
-  // Add next element of array to the end of the result.
-  // Get current_string = array[index].
-  __ mov(scratch, elements);
-  __ mov(current_string, FieldOperand(scratch, index,
                                       times_pointer_size,
                                       FixedArray::kHeaderSize));
-  // If current != flat ascii string drop result, return undefined.
+  __ test(string, Immediate(kSmiTagMask));
-  __ test(current_string, Immediate(kSmiTagMask));
   __ j(zero, &bailout);
-  __ mov(scratch, FieldOperand(current_string, HeapObject::kMapOffset));
+  __ mov(scratch, FieldOperand(string, HeapObject::kMapOffset));
-  __ mov_b(scratch, FieldOperand(scratch, Map::kInstanceTypeOffset));
+  __ movzx_b(scratch, FieldOperand(scratch, Map::kInstanceTypeOffset));
+  __ and_(scratch, Immediate(
+      kIsNotStringMask | kStringEncodingMask | kStringRepresentationMask));
+  __ cmp(scratch, kStringTag | kAsciiStringTag | kSeqStringTag);
+  __ j(not_equal, &bailout);
+  __ add(string_length,
+         FieldOperand(string, SeqAsciiString::kLengthOffset));
+  __ j(overflow, &bailout);
+  __ add(Operand(index), Immediate(1));
+  __ bind(&loop_condition);
+  __ cmp(index, Operand(array_length));
+  __ j(less, &loop);
+
+  // If array_length is 1, return elements[0], a string.
+  __ cmp(array_length, 1);
+  __ j(not_equal, &not_size_one_array);
+  __ mov(scratch, FieldOperand(elements, FixedArray::kHeaderSize));
+  __ mov(result_operand, scratch);
+  __ jmp(&done);
+
+  __ bind(&not_size_one_array);
+
+  // End of array_length live range.
+  result_pos = array_length;
+  array_length = no_reg;
+
+  // Live registers:
+  // string_length: Sum of string lengths, as a smi.
+  // elements: FixedArray of strings.
+
+  // Check that the separator is a flat ASCII string.
+  __ mov(string, separator_operand);
+  __ test(string, Immediate(kSmiTagMask));
+  __ j(zero, &bailout);
+  __ mov(scratch, FieldOperand(string, HeapObject::kMapOffset));
+  __ movzx_b(scratch, FieldOperand(scratch, Map::kInstanceTypeOffset));
   __ and_(scratch, Immediate(
       kIsNotStringMask | kStringEncodingMask | kStringRepresentationMask));
   __ cmp(scratch, kStringTag | kAsciiStringTag | kSeqStringTag);
   __ j(not_equal, &bailout);
 
-  // Append current to the result.
+  // Add (separator length times array_length) - separator length
-  __ AppendStringToTopOfNewSpace(current_string, current_string_length,
+  // to string_length.
-                                 result_pos, scratch, scratch_2, result,
+  __ mov(scratch, separator_operand);
-                                 padding_chars, &bailout);
+  __ mov(scratch, FieldOperand(scratch, SeqAsciiString::kLengthOffset));
+  __ sub(string_length, Operand(scratch));  // May be negative, temporarily.
+  __ imul(scratch, array_length_operand);
+  __ j(overflow, &bailout);
+  __ add(string_length, Operand(scratch));
+  __ j(overflow, &bailout);
+
+  __ shr(string_length, 1);
+  // Live registers and stack values:
+  //   string_length
+  //   elements
+  __ AllocateAsciiString(result_pos, string_length, scratch,
+                         index, string, &bailout);
+  __ mov(result_operand, result_pos);
+  __ lea(result_pos, FieldOperand(result_pos, SeqAsciiString::kHeaderSize));
+
+
+  __ mov(string, separator_operand);
+  __ cmp(FieldOperand(string, SeqAsciiString::kLengthOffset),
+         Immediate(Smi::FromInt(1)));
+  __ j(equal, &one_char_separator);
+  __ j(greater, &long_separator);
+
+
+  // Empty separator case
+  __ mov(index, Immediate(0));
+  __ jmp(&loop_1_condition);
+  // Loop condition: while (index < length).
+  __ bind(&loop_1);
+  // Each iteration of the loop concatenates one string to the result.
+  // Live values in registers:
+  //   index: which element of the elements array we are adding to the result.
+  //   result_pos: the position to which we are currently copying characters.
+  //   elements: the FixedArray of strings we are joining.
+
+  // Get string = array[index].
+  __ mov(string, FieldOperand(elements, index,
+                              times_pointer_size,
+                              FixedArray::kHeaderSize));
+  __ mov(string_length,
+         FieldOperand(string, String::kLengthOffset));
+  __ shr(string_length, 1);
+  __ lea(string,
+         FieldOperand(string, SeqAsciiString::kHeaderSize));
+  __ CopyBytes(string, result_pos, string_length, scratch);
   __ add(Operand(index), Immediate(1));
-  __ jmp(&loop);  // End while (index < length).
+  __ bind(&loop_1_condition);
+  __ cmp(index, array_length_operand);
+  __ j(less, &loop_1);  // End while (index < length).
+  __ jmp(&done);
+
+
+
+  // One-character separator case
+  __ bind(&one_char_separator);
+  // Replace separator with its ascii character value.
+  __ mov_b(scratch, FieldOperand(string, SeqAsciiString::kHeaderSize));
+  __ mov_b(separator_operand, scratch);
+
+  __ Set(index, Immediate(0));
+  // Jump into the loop after the code that copies the separator, so the first
+  // element is not preceded by a separator
+  __ jmp(&loop_2_entry);
+  // Loop condition: while (index < length).
+  __ bind(&loop_2);
+  // Each iteration of the loop concatenates one string to the result.
+  // Live values in registers:
+  //   index: which element of the elements array we are adding to the result.
+  //   result_pos: the position to which we are currently copying characters.
+
+  // Copy the separator character to the result.
+  __ mov_b(scratch, separator_operand);
+  __ mov_b(Operand(result_pos, 0), scratch);
+  __ inc(result_pos);
+
+  __ bind(&loop_2_entry);
+  // Get string = array[index].
+  __ mov(string, FieldOperand(elements, index,
+                              times_pointer_size,
+                              FixedArray::kHeaderSize));
+  __ mov(string_length,
+         FieldOperand(string, String::kLengthOffset));
+  __ shr(string_length, 1);
+  __ lea(string,
+         FieldOperand(string, SeqAsciiString::kHeaderSize));
+  __ CopyBytes(string, result_pos, string_length, scratch);
+  __ add(Operand(index), Immediate(1));
+
+  __ cmp(index, array_length_operand);
+  __ j(less, &loop_2);  // End while (index < length).
+  __ jmp(&done);
+
+
+  // Long separator case (separator is more than one character).
+  __ bind(&long_separator);
+
+  __ Set(index, Immediate(0));
+  // Jump into the loop after the code that copies the separator, so the first
+  // element is not preceded by a separator
+  __ jmp(&loop_3_entry);
+  // Loop condition: while (index < length).
+  __ bind(&loop_3);
+  // Each iteration of the loop concatenates one string to the result.
+  // Live values in registers:
+  //   index: which element of the elements array we are adding to the result.
+  //   result_pos: the position to which we are currently copying characters.
+
+  // Copy the separator to the result.
+  __ mov(string, separator_operand);
+  __ mov(string_length,
+         FieldOperand(string, String::kLengthOffset));
+  __ shr(string_length, 1);
+  __ lea(string,
+         FieldOperand(string, SeqAsciiString::kHeaderSize));
+  __ CopyBytes(string, result_pos, string_length, scratch);
+
+  __ bind(&loop_3_entry);
+  // Get string = array[index].
+  __ mov(string, FieldOperand(elements, index,
+                              times_pointer_size,
+                              FixedArray::kHeaderSize));
+  __ mov(string_length,
+         FieldOperand(string, String::kLengthOffset));
+  __ shr(string_length, 1);
+  __ lea(string,
+         FieldOperand(string, SeqAsciiString::kHeaderSize));
+  __ CopyBytes(string, result_pos, string_length, scratch);
+  __ add(Operand(index), Immediate(1));
+
+  __ cmp(index, array_length_operand);
+  __ j(less, &loop_3);  // End while (index < length).
+  __ jmp(&done);
+
 
   __ bind(&bailout);
-  __ mov(result, Factory::undefined_value());
+  __ mov(result_operand, Factory::undefined_value());
   __ bind(&done);
-  __ mov(eax, result);
+  __ mov(eax, result_operand);
   // Drop temp values from the stack, and restore context register.
-  __ add(Operand(esp), Immediate(5 * kPointerSize));
+  __ add(Operand(esp), Immediate(3 * kPointerSize));
 
   __ mov(esi, Operand(ebp, StandardFrameConstants::kContextOffset));
   context()->Plug(eax);

src/ia32/macro-assembler-ia32.cc

@@ -877,55 +877,53 @@ void MacroAssembler::AllocateAsciiConsString(Register result,
       Immediate(Factory::cons_ascii_string_map()));
 }
 
-// All registers must be distinct.  Only current_string needs valid contents
-// on entry.  All registers may be invalid on exit.  result_operand is
-// unchanged, padding_chars is updated correctly.
-void MacroAssembler::AppendStringToTopOfNewSpace(
-    Register current_string,  // Tagged pointer to string to copy.
-    Register current_string_length,
-    Register result_pos,
-    Register scratch,
-    Register new_padding_chars,
-    Operand operand_result,
-    Operand operand_padding_chars,
-    Label* bailout) {
-  mov(current_string_length,
-      FieldOperand(current_string, String::kLengthOffset));
-  shr(current_string_length, 1);
-  sub(current_string_length, operand_padding_chars);
-  mov(new_padding_chars, current_string_length);
-  add(Operand(current_string_length), Immediate(kObjectAlignmentMask));
-  and_(Operand(current_string_length), Immediate(~kObjectAlignmentMask));
-  sub(new_padding_chars, Operand(current_string_length));
-  neg(new_padding_chars);
-  // We need an allocation even if current_string_length is 0, to fetch
-  // result_pos.  Consider using a faster fetch of result_pos in that case.
-  AllocateInNewSpace(current_string_length, result_pos, scratch, no_reg,
-                     bailout, NO_ALLOCATION_FLAGS);
-  sub(result_pos, operand_padding_chars);
-  mov(operand_padding_chars, new_padding_chars);
 
-  Register scratch_2 = new_padding_chars;  // Used to compute total length.
+// Copy memory, byte-by-byte, from source to destination.  Not optimized for
-  // Copy string to the end of result.
+// long or aligned copies.  The contents of scratch and length are destroyed.
-  mov(current_string_length,
+// Source and destination are incremented by length.
-      FieldOperand(current_string, String::kLengthOffset));
+// Many variants of movsb, loop unrolling, word moves, and indexed operands
-  mov(scratch, operand_result);
+// have been tried here already, and this is fastest.
-  mov(scratch_2, current_string_length);
+// A simpler loop is faster on small copies, but 30% slower on large ones.
-  add(scratch_2, FieldOperand(scratch, String::kLengthOffset));
+// The cld() instruction must have been emitted, to set the direction flag(),
-  mov(FieldOperand(scratch, String::kLengthOffset), scratch_2);
+// before calling this function.
-  shr(current_string_length, 1);
+void MacroAssembler::CopyBytes(Register source,
-  lea(current_string,
+                               Register destination,
-      FieldOperand(current_string, SeqAsciiString::kHeaderSize));
+                               Register length,
-  // Loop condition: while (--current_string_length >= 0).
+                               Register scratch) {
-  Label copy_loop;
+  Label loop, done, short_string, short_loop;
-  Label copy_loop_entry;
+  // Experimentation shows that the short string loop is faster if length < 10.
-  jmp(&copy_loop_entry);
+  cmp(Operand(length), Immediate(10));
-  bind(&copy_loop);
+  j(less_equal, &short_string);
-  mov_b(scratch, Operand(current_string, current_string_length, times_1, 0));
+
-  mov_b(Operand(result_pos, current_string_length, times_1, 0), scratch);
+  ASSERT(source.is(esi));
-  bind(&copy_loop_entry);
+  ASSERT(destination.is(edi));
-  sub(Operand(current_string_length), Immediate(1));
+  ASSERT(length.is(ecx));
-  j(greater_equal, &copy_loop);
+
+  // Because source is 4-byte aligned in our uses of this function,
+  // we keep source aligned for the rep_movs call by copying the odd bytes
+  // at the end of the ranges.
+  mov(scratch, Operand(source, length, times_1, -4));
+  mov(Operand(destination, length, times_1, -4), scratch);
+  mov(scratch, ecx);
+  shr(ecx, 2);
+  rep_movs();
+  and_(Operand(scratch), Immediate(0x3));
+  add(destination, Operand(scratch));
+  jmp(&done);
+
+  bind(&short_string);
+  test(length, Operand(length));
+  j(zero, &done);
+
+  bind(&short_loop);
+  mov_b(scratch, Operand(source, 0));
+  mov_b(Operand(destination, 0), scratch);
+  inc(source);
+  inc(destination);
+  dec(length);
+  j(not_zero, &short_loop);
+
+  bind(&done);
 }
 
 

src/ia32/macro-assembler-ia32.h

@@ -386,22 +386,13 @@ class MacroAssembler: public Assembler {
                                Register scratch2,
                                Label* gc_required);
 
-  // All registers must be distinct.  Only current_string needs valid contents
+  // Copy memory, byte-by-byte, from source to destination.  Not optimized for
-  // on entry.  All registers may be invalid on exit.  result_operand is
+  // long or aligned copies.
-  // unchanged, padding_chars is updated correctly.
+  // The contents of index and scratch are destroyed.
-  // The top of new space must contain a sequential ascii string with
+  void CopyBytes(Register source,
-  // padding_chars bytes free in its top word.  The sequential ascii string
+                 Register destination,
-  // current_string is concatenated to it, allocating the necessary amount
+                 Register length,
-  // of new memory.
+                 Register scratch);
-  void AppendStringToTopOfNewSpace(
-      Register current_string,  // Tagged pointer to string to copy.
-      Register current_string_length,
-      Register result_pos,
-      Register scratch,
-      Register new_padding_chars,
-      Operand operand_result,
-      Operand operand_padding_chars,
-      Label* bailout);
 
   // ---------------------------------------------------------------------------
   // Support functions.