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

src/ia32/full-codegen-ia32.cc

Index: src/ia32/full-codegen-ia32.cc
===================================================================
--- src/ia32/full-codegen-ia32.cc	(revision 6283)
+++ src/ia32/full-codegen-ia32.cc	(working copy)
@@ -3351,8 +3351,12 @@
 
 
 void FullCodeGenerator::EmitFastAsciiArrayJoin(ZoneList<Expression*>* args) {
-  Label bailout;
-  Label done;
+  Label bailout, done, one_char_separator, long_separator,
+      non_trivial_array, not_size_one_array, loop, loop_condition,
+      loop_1, loop_1_condition, loop_1a, loop_1a_condition,
+      loop_2, loop_2_entry, loop_2a, loop_2a_condition,
+      loop_3, loop_3_entry, loop_3a, loop_3a_condition,
+      loop_3b, loop_3b_condition;

[Lasse Reichstein, 2011/01/13 09:20:32]

That's a lot of loops. Could they be given more telling names? Or give an
explanation of the algorithm here, so we can see which loops do what.

[William Hesse, 2011/01/13 16:25:59]

Loops eliminated by making the string copy a macro-assembler function.

 
   ASSERT(args->length() == 2);
   // We will leave the separator on the stack until the end of the function.
@@ -3362,28 +3366,28 @@
 
   // 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 result_pos = no_reg;  // Will be eax.
 
   Register index = edi;
 
-  Register current_string_length = ecx;  // Will be ecx when live.
+  Register current_string_length = ecx;
 
   Register current_string = edx;
 
   Register scratch = ebx;
 
-  Register scratch_2 = esi;
-  Register new_padding_chars = scratch_2;
+  Register array_length = esi;
+  Register pos = no_reg;  // Will be esi.
 
-  Operand separator = Operand(esp, 4 * kPointerSize);  // Already pushed.
-  Operand elements = Operand(esp, 3 * kPointerSize);
-  Operand result = Operand(esp, 2 * kPointerSize);
-  Operand padding_chars = Operand(esp, 1 * kPointerSize);
-  Operand array_length = Operand(esp, 0);
-  __ sub(Operand(esp), Immediate(4 * kPointerSize));
+  // Separator operand is already pushed.
+  Operand separator_operand = Operand(esp, 3 * kPointerSize);
+  Operand elements_operand = Operand(esp, 2 * kPointerSize);
+  Operand result_operand = Operand(esp, 1 * kPointerSize);
+  Operand array_length_operand = Operand(esp, 0);
+  __ sub(Operand(esp), Immediate(3 * kPointerSize));
 
-
-  // Check that eax is a JSArray
+  // Check that the array is a JSArray
   __ test(array, Immediate(kSmiTagMask));
   __ j(zero, &bailout);
   __ CmpObjectType(array, JS_ARRAY_TYPE, scratch);
@@ -3394,27 +3398,42 @@
             1 << Map::kHasFastElements);

[Lasse Reichstein, 2011/01/13 09:20:32]

Will Copy-on-Write arrays have fast elements if queried this way?

[William Hesse, 2011/01/13 16:25:59]

Yes, they will (I checked).

   __ 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));
   __ sar(scratch, 1);
-  Label non_trivial;
-  __ j(not_zero, &non_trivial);
-  __ mov(result, Factory::empty_string());
+  __ j(not_zero, &non_trivial_array);
+  __ mov(result_operand, Factory::empty_string());
   __ jmp(&done);
 
-  __ bind(&non_trivial);
+  // Save the array length.
+  __ bind(&non_trivial_array);
+  __ mov(array_length_operand, scratch);
   __ mov(array_length, scratch);
 
+  // Save the FixedArray containing array's elements.
   __ mov(scratch, FieldOperand(array, JSArray::kElementsOffset));
-  __ mov(elements, scratch);
-
   // End of array's live range.
-  result_pos = array;
+  elements = array;
   array = no_reg;
+  __ mov(elements_operand, scratch);
+  __ mov(elements, scratch);
 
 
-  // Check that the separator is a flat ascii string.
-  __ mov(current_string, separator);
+  // Check that all array elements are sequential ascii strings, and

[Lasse Reichstein, 2011/01/13 09:20:32]

ascii -> ASCII.

[William Hesse, 2011/01/13 16:25:59]

Both forms occur in our comments, about 1/3 lowercase, 2/3 uppercase.
I prefer lowercase.
But done.

+  // accumulate the sum of their lengths.

[Lasse Reichstein, 2011/01/13 09:20:32]

Say that the sum is maintained as a Smi value.

[William Hesse, 2011/01/13 16:25:59]

Done.

+  __ Set(index, Immediate(0));
+  __ Set(current_string_length, Immediate(0));
+  // Loop condition: while (index < length).
+  // Live loop registers: index, array_length, current_string,
+  //                      scratch, current_string_length, elements.
+  __ jmp(&loop_condition);
+  __ bind(&loop);
+  __ cmp(index, Operand(array_length));
+  __ j(greater_equal, &done);
+
+  __ mov(current_string, FieldOperand(elements, index,
+                                      times_pointer_size,
+                                      FixedArray::kHeaderSize));
   __ test(current_string, Immediate(kSmiTagMask));
   __ j(zero, &bailout);
   __ mov(scratch, FieldOperand(current_string, HeapObject::kMapOffset));
@@ -3423,18 +3442,35 @@
       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);
+  __ add(current_string_length,
+         FieldOperand(current_string, SeqAsciiString::kLengthOffset));
+  __ j(overflow, &bailout);
+  __ add(Operand(index), Immediate(1));
+  __ bind(&loop_condition);
+  __ cmp(index, Operand(array_length));
+  __ j(less, &loop);
 
-  // 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));
+  // 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 elements live range
+  result_pos = elements;
+  elements = no_reg;
+
+  // End of array_length live range.
+  pos = array_length;
+  array_length = no_reg;
+
+  // Live register:
+  // current_string_length: Sum of string lengths, as a smi.
+
+  // Check that the separator is a flat ascii string.

[Lasse Reichstein, 2011/01/13 09:20:32]

ASCII.

[William Hesse, 2011/01/13 16:25:59]

Done.

+  __ mov(current_string, separator_operand);
   __ test(current_string, Immediate(kSmiTagMask));
   __ j(zero, &bailout);
   __ mov(scratch, FieldOperand(current_string, HeapObject::kMapOffset));
@@ -3444,90 +3480,224 @@
   __ 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));
+  // Add (separator length times array_length) - separator length
+  // to current_string_length.
+  __ mov(scratch, separator_operand);
+  __ mov(scratch, FieldOperand(scratch, SeqAsciiString::kLengthOffset));
+  __ sub(current_string_length, Operand(scratch));

[Lasse Reichstein, 2011/01/13 09:20:32]

Perhaps worth noting that this can (and may) give a negative result.

[William Hesse, 2011/01/13 16:25:59]

Done.

+  __ imul(scratch, array_length_operand);
+  __ j(overflow, &bailout);
+  __ add(current_string_length, Operand(scratch));
+  __ j(overflow, &bailout);
+
   __ 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);
+  __ AllocateAsciiString(result_pos, current_string_length, pos,
+                         index, current_string, &bailout);
+  __ mov(result_operand, result_pos);
+  __ lea(result_pos, FieldOperand(result_pos, SeqAsciiString::kHeaderSize));
 
-  // 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(current_string, separator_operand);
+  __ cmp(FieldOperand(current_string, SeqAsciiString::kLengthOffset),
+         Immediate(Smi::FromInt(1)));
+  __ j(equal, &one_char_separator);
+  __ j(greater, &long_separator);
 
-  __ mov(index, Immediate(1));
+
+  // Empty separator case
+  __ mov(index, Immediate(0));
+  __ jmp(&loop_1_condition);
   // Loop condition: while (index < length).
-  Label loop;
-  __ bind(&loop);
-  __ cmp(index, array_length);
-  __ j(greater_equal, &done);
+  __ 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.
 
-  // If the separator is the empty string, signalled by NULL, skip it.
-  Label separator_done;
-  __ mov(current_string, separator);
-  __ test(current_string, Operand(current_string));
-  __ j(zero, &separator_done);
+  // Get current_string = array[index].
+  __ mov(scratch, elements_operand);
+  __ mov(current_string, FieldOperand(scratch, index,
+                                      times_pointer_size,
+                                      FixedArray::kHeaderSize));
+  __ mov(current_string_length,
+         FieldOperand(current_string, String::kLengthOffset));
+  __ shr(current_string_length, 1);
+  __ lea(current_string,
+         FieldOperand(current_string, SeqAsciiString::kHeaderSize));
+  __ Set(pos, Immediate(0));
+  __ jmp(&loop_1a_condition);
 
-  // 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);
+  __ bind(&loop_1a);
+  // Live registers:
+  //   current_string: the start of the current string's data.
+  //   pos: the index into the currently copied character of current_string.
+  //   current_string_length: the length of the current string.
+  //   scratch: the byte being copied.
+  //   result_pos: the position to copy the current string to.
+  //   index: which element of the array is being copied (live across loop).
+  __ mov_b(scratch, Operand(current_string, pos, times_1, 0));
+  __ mov_b(Operand(result_pos, pos, times_1, 0), scratch);
+  __ inc(pos);
+  __ bind(&loop_1a_condition);
+  __ cmp(pos, Operand(current_string_length));
+  __ j(less, &loop_1a);
 
-  // Add next element of array to the end of the result.
+  __ add(result_pos, Operand(current_string_length));
+  __ add(Operand(index), Immediate(1));
+  __ 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(current_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 current_string = array[index].
-  __ mov(scratch, elements);
+  __ mov(scratch, elements_operand);
   __ mov(current_string, FieldOperand(scratch, index,
                                       times_pointer_size,
                                       FixedArray::kHeaderSize));
-  // If current != flat ascii string drop result, return undefined.
-  __ 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);
+  __ mov(current_string_length,
+         FieldOperand(current_string, String::kLengthOffset));
+  __ shr(current_string_length, 1);
+  __ lea(current_string,
+         FieldOperand(current_string, SeqAsciiString::kHeaderSize));
+  __ Set(pos, Immediate(0));
+  __ jmp(&loop_2a_condition);
 
-  // Append current to the result.
-  __ AppendStringToTopOfNewSpace(current_string, current_string_length,
-                                 result_pos, scratch, scratch_2, result,
-                                 padding_chars, &bailout);
+  __ bind(&loop_2a);
+  // Live registers:
+  //   current_string: the start of the current string's data.
+  //   pos: the index into the currently copied character of current_string.
+  //   current_string_length: the length of the current string.
+  //   scratch: the byte being copied.
+  //   result_pos: the position to copy the current string to.
+  //   index: which element of the array is being copied (live across loop).
+  __ mov_b(scratch, Operand(current_string, pos, times_1, 0));
+  __ mov_b(Operand(result_pos, pos, times_1, 0), scratch);
+  __ inc(pos);
+  __ bind(&loop_2a_condition);
+  __ cmp(pos, Operand(current_string_length));
+  __ j(less, &loop_2a);
+
+  __ add(result_pos, Operand(current_string_length));
   __ add(Operand(index), Immediate(1));
-  __ jmp(&loop);  // End while (index < length).
 
+  __ 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(current_string, separator_operand);
+  __ mov(current_string_length,
+         FieldOperand(current_string, String::kLengthOffset));
+  __ shr(current_string_length, 1);
+  __ lea(current_string,
+         FieldOperand(current_string, SeqAsciiString::kHeaderSize));
+  __ Set(pos, Immediate(0));
+  __ jmp(&loop_3a_condition);
+
+  __ bind(&loop_3a);
+  // Live registers:
+  //   current_string: the start of the separator's data.
+  //   pos: the index into the currently copied character of separator.
+  //   current_string_length: the length of the separator.
+  //   scratch: the byte being copied.
+  //   result_pos: the position to copy the separator to.
+  //   index: which element of the array is being copied (live across loop).
+  __ mov_b(scratch, Operand(current_string, pos, times_1, 0));
+  __ mov_b(Operand(result_pos, pos, times_1, 0), scratch);
+  __ inc(pos);
+  __ bind(&loop_3a_condition);
+  __ cmp(pos, Operand(current_string_length));
+  __ j(less, &loop_3a);
+
+  __ add(result_pos, Operand(current_string_length));
+
+  __ bind(&loop_3_entry);
+  // Get current_string = array[index].
+  __ mov(scratch, elements_operand);
+  __ mov(current_string, FieldOperand(scratch, index,
+                                      times_pointer_size,
+                                      FixedArray::kHeaderSize));
+  __ mov(current_string_length,
+         FieldOperand(current_string, String::kLengthOffset));
+  __ shr(current_string_length, 1);
+  __ lea(current_string,
+         FieldOperand(current_string, SeqAsciiString::kHeaderSize));
+  __ Set(pos, Immediate(0));
+  __ jmp(&loop_3b_condition);
+
+  __ bind(&loop_3b);
+  // Live registers:
+  //   current_string: the start of the current string's data.
+  //   pos: the index into the currently copied character of current_string.
+  //   current_string_length: the length of the current string.
+  //   scratch: the byte being copied.
+  //   result_pos: the position to copy the current string to.
+  //   index: which element of the array is being copied (live across loop).
+  __ mov_b(scratch, Operand(current_string, pos, times_1, 0));
+  __ mov_b(Operand(result_pos, pos, times_1, 0), scratch);
+  __ inc(pos);
+  __ bind(&loop_3b_condition);
+  __ cmp(pos, Operand(current_string_length));
+  __ j(less, &loop_3b);
+
+  __ add(result_pos, Operand(current_string_length));
+  __ 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(4 * kPointerSize));
 
   __ mov(esi, Operand(ebp, StandardFrameConstants::kContextOffset));
   context()->Plug(eax);