Revert of [fullcodegen] Remove the hacky %_FastOneByteArrayJoin intrinsic....

Revert of [fullcodegen] Remove the hacky %_FastOneByteArrayJoin intrinsic. (patchset #1 id:1 of https://codereview.chromium.org/1708523002/ )

Reason for revert:
[Sheriff] Speculative revert for adding arm flakes:
https://build.chromium.org/p/client.v8/builders/V8%20Linux%20-%20arm%20-%20sim/builds/5756
https://build.chromium.org/p/client.v8/builders/V8%20Linux%20-%20arm%20-%20sim%20-%20novfp3/builds/3375/

And some more...

Original issue's description:
> [fullcodegen] Remove the hacky %_FastOneByteArrayJoin intrinsic.
>
> This intrinsic was only supported in fullcodegen, and is actually no
> longer relevant for SunSpider peak performance it seems, so let's get
> rid of it and maybe just implement Array.prototype.join with a fast
> path at some point instead.
>
> R=mstarzinger@chromium.org
>
> Committed: https://crrev.com/ccf12b4bede3f1ce3ce14fb33bcc4041525a40af
> Cr-Commit-Position: refs/heads/master@{#34084}

TBR=mstarzinger@chromium.org,bmeurer@chromium.org
# Skipping CQ checks because original CL landed less than 1 days ago.
NOPRESUBMIT=true
NOTREECHECKS=true
NOTRY=true

Review URL: https://codereview.chromium.org/1709693002

Cr-Commit-Position: refs/heads/master@{#34091}

src/bailout-reason.h

@@ -97,6 +97,8 @@ namespace internal {
   V(kImportDeclaration, "Import declaration")                                  \
   V(kIndexIsNegative, "Index is negative")                                     \
   V(kIndexIsTooLarge, "Index is too large")                                    \
+  V(kInlinedRuntimeFunctionFastOneByteArrayJoin,                               \
+    "Inlined runtime function: FastOneByteArrayJoin")                          \
   V(kInliningBailedOut, "Inlining bailed out")                                 \
   V(kInputGPRIsExpectedToHaveUpper32Cleared,                                   \
     "Input GPR is expected to have upper32 cleared")                           \
@@ -127,6 +129,8 @@ namespace internal {
   V(kNativeFunctionLiteral, "Native function literal")                         \
   V(kNeedSmiLiteral, "Need a Smi literal here")                                \
   V(kNoCasesLeft, "No cases left")                                             \
+  V(kNoEmptyArraysHereInEmitFastOneByteArrayJoin,                              \
+    "No empty arrays here in EmitFastOneByteArrayJoin")                        \
   V(kNonInitializerAssignmentToConst, "Non-initializer assignment to const")   \
   V(kNonSmiIndex, "Non-smi index")                                             \
   V(kNonSmiKeyInArrayLiteral, "Non-smi key in array literal")                  \

src/crankshaft/hydrogen.cc

@@ -12900,6 +12900,16 @@ void HOptimizedGraphBuilder::GenerateGetCachedArrayIndex(CallRuntime* call) {
 }
 
 
+void HOptimizedGraphBuilder::GenerateFastOneByteArrayJoin(CallRuntime* call) {
+  // Simply returning undefined here would be semantically correct and even
+  // avoid the bailout. Nevertheless, some ancient benchmarks like SunSpider's
+  // string-fasta would tank, because fullcode contains an optimized version.
+  // Obviously the fullcode => Crankshaft => bailout => fullcode dance is
+  // faster... *sigh*
+  return Bailout(kInlinedRuntimeFunctionFastOneByteArrayJoin);
+}
+
+
 void HOptimizedGraphBuilder::GenerateDebugBreakInOptimizedCode(
     CallRuntime* call) {
   Add<HDebugBreak>();

src/crankshaft/hydrogen.h

@@ -2213,6 +2213,7 @@ class HOptimizedGraphBuilder : public HGraphBuilder, public AstVisitor {
   F(MathPow)                           \
   F(HasCachedArrayIndex)               \
   F(GetCachedArrayIndex)               \
+  F(FastOneByteArrayJoin)              \
   F(DebugBreakInOptimizedCode)         \
   F(StringCharCodeAt)                  \
   F(SubString)                         \

src/full-codegen/arm/full-codegen-arm.cc

@@ -3414,6 +3414,243 @@ void FullCodeGenerator::EmitGetSuperConstructor(CallRuntime* expr) {
 }
 
 
+void FullCodeGenerator::EmitFastOneByteArrayJoin(CallRuntime* expr) {
+  Label bailout, done, one_char_separator, long_separator, non_trivial_array,
+      not_size_one_array, loop, empty_separator_loop, one_char_separator_loop,
+      one_char_separator_loop_entry, long_separator_loop;
+  ZoneList<Expression*>* args = expr->arguments();
+  DCHECK(args->length() == 2);
+  VisitForStackValue(args->at(1));
+  VisitForAccumulatorValue(args->at(0));
+
+  // All aliases of the same register have disjoint lifetimes.
+  Register array = r0;
+  Register elements = no_reg;  // Will be r0.
+  Register result = no_reg;  // Will be r0.
+  Register separator = r1;
+  Register array_length = r2;
+  Register result_pos = no_reg;  // Will be r2
+  Register string_length = r3;
+  Register string = r4;
+  Register element = r5;
+  Register elements_end = r6;
+  Register scratch = r9;
+
+  // Separator operand is on the stack.
+  __ pop(separator);
+
+  // Check that the array is a JSArray.
+  __ JumpIfSmi(array, &bailout);
+  __ CompareObjectType(array, scratch, array_length, JS_ARRAY_TYPE);
+  __ b(ne, &bailout);
+
+  // Check that the array has fast elements.
+  __ CheckFastElements(scratch, array_length, &bailout);
+
+  // If the array has length zero, return the empty string.
+  __ ldr(array_length, FieldMemOperand(array, JSArray::kLengthOffset));
+  __ SmiUntag(array_length, SetCC);
+  __ b(ne, &non_trivial_array);
+  __ LoadRoot(r0, Heap::kempty_stringRootIndex);
+  __ b(&done);
+
+  __ bind(&non_trivial_array);
+
+  // Get the FixedArray containing array's elements.
+  elements = array;
+  __ ldr(elements, FieldMemOperand(array, JSArray::kElementsOffset));
+  array = no_reg;  // End of array's live range.
+
+  // Check that all array elements are sequential one-byte strings, and
+  // accumulate the sum of their lengths, as a smi-encoded value.
+  __ mov(string_length, Operand::Zero());
+  __ add(element,
+         elements, Operand(FixedArray::kHeaderSize - kHeapObjectTag));
+  __ add(elements_end, element, Operand(array_length, LSL, kPointerSizeLog2));
+  // Loop condition: while (element < elements_end).
+  // Live values in registers:
+  //   elements: Fixed array of strings.
+  //   array_length: Length of the fixed array of strings (not smi)
+  //   separator: Separator string
+  //   string_length: Accumulated sum of string lengths (smi).
+  //   element: Current array element.
+  //   elements_end: Array end.
+  if (generate_debug_code_) {
+    __ cmp(array_length, Operand::Zero());
+    __ Assert(gt, kNoEmptyArraysHereInEmitFastOneByteArrayJoin);
+  }
+  __ bind(&loop);
+  __ ldr(string, MemOperand(element, kPointerSize, PostIndex));
+  __ JumpIfSmi(string, &bailout);
+  __ ldr(scratch, FieldMemOperand(string, HeapObject::kMapOffset));
+  __ ldrb(scratch, FieldMemOperand(scratch, Map::kInstanceTypeOffset));
+  __ JumpIfInstanceTypeIsNotSequentialOneByte(scratch, scratch, &bailout);
+  __ ldr(scratch, FieldMemOperand(string, SeqOneByteString::kLengthOffset));
+  __ add(string_length, string_length, Operand(scratch), SetCC);
+  __ b(vs, &bailout);
+  __ cmp(element, elements_end);
+  __ b(lt, &loop);
+
+  // If array_length is 1, return elements[0], a string.
+  __ cmp(array_length, Operand(1));
+  __ b(ne, &not_size_one_array);
+  __ ldr(r0, FieldMemOperand(elements, FixedArray::kHeaderSize));
+  __ b(&done);
+
+  __ bind(&not_size_one_array);
+
+  // Live values in registers:
+  //   separator: Separator string
+  //   array_length: Length of the array.
+  //   string_length: Sum of string lengths (smi).
+  //   elements: FixedArray of strings.
+
+  // Check that the separator is a flat one-byte string.
+  __ JumpIfSmi(separator, &bailout);
+  __ ldr(scratch, FieldMemOperand(separator, HeapObject::kMapOffset));
+  __ ldrb(scratch, FieldMemOperand(scratch, Map::kInstanceTypeOffset));
+  __ JumpIfInstanceTypeIsNotSequentialOneByte(scratch, scratch, &bailout);
+
+  // Add (separator length times array_length) - separator length to the
+  // string_length to get the length of the result string. array_length is not
+  // smi but the other values are, so the result is a smi
+  __ ldr(scratch, FieldMemOperand(separator, SeqOneByteString::kLengthOffset));
+  __ sub(string_length, string_length, Operand(scratch));
+  __ smull(scratch, ip, array_length, scratch);
+  // Check for smi overflow. No overflow if higher 33 bits of 64-bit result are
+  // zero.
+  __ cmp(ip, Operand::Zero());
+  __ b(ne, &bailout);
+  __ tst(scratch, Operand(0x80000000));
+  __ b(ne, &bailout);
+  __ add(string_length, string_length, Operand(scratch), SetCC);
+  __ b(vs, &bailout);
+  __ SmiUntag(string_length);
+
+  // Bailout for large object allocations.
+  __ cmp(string_length, Operand(Page::kMaxRegularHeapObjectSize));
+  __ b(gt, &bailout);
+
+  // Get first element in the array to free up the elements register to be used
+  // for the result.
+  __ add(element,
+         elements, Operand(FixedArray::kHeaderSize - kHeapObjectTag));
+  result = elements;  // End of live range for elements.
+  elements = no_reg;
+  // Live values in registers:
+  //   element: First array element
+  //   separator: Separator string
+  //   string_length: Length of result string (not smi)
+  //   array_length: Length of the array.
+  __ AllocateOneByteString(result, string_length, scratch,
+                           string,        // used as scratch
+                           elements_end,  // used as scratch
+                           &bailout);
+  // Prepare for looping. Set up elements_end to end of the array. Set
+  // result_pos to the position of the result where to write the first
+  // character.
+  __ add(elements_end, element, Operand(array_length, LSL, kPointerSizeLog2));
+  result_pos = array_length;  // End of live range for array_length.
+  array_length = no_reg;
+  __ add(result_pos,
+         result,
+         Operand(SeqOneByteString::kHeaderSize - kHeapObjectTag));
+
+  // Check the length of the separator.
+  __ ldr(scratch, FieldMemOperand(separator, SeqOneByteString::kLengthOffset));
+  __ cmp(scratch, Operand(Smi::FromInt(1)));
+  __ b(eq, &one_char_separator);
+  __ b(gt, &long_separator);
+
+  // Empty separator case
+  __ bind(&empty_separator_loop);
+  // Live values in registers:
+  //   result_pos: the position to which we are currently copying characters.
+  //   element: Current array element.
+  //   elements_end: Array end.
+
+  // Copy next array element to the result.
+  __ ldr(string, MemOperand(element, kPointerSize, PostIndex));
+  __ ldr(string_length, FieldMemOperand(string, String::kLengthOffset));
+  __ SmiUntag(string_length);
+  __ add(string,
+         string,
+         Operand(SeqOneByteString::kHeaderSize - kHeapObjectTag));
+  __ CopyBytes(string, result_pos, string_length, scratch);
+  __ cmp(element, elements_end);
+  __ b(lt, &empty_separator_loop);  // End while (element < elements_end).
+  DCHECK(result.is(r0));
+  __ b(&done);
+
+  // One-character separator case
+  __ bind(&one_char_separator);
+  // Replace separator with its one-byte character value.
+  __ ldrb(separator, FieldMemOperand(separator, SeqOneByteString::kHeaderSize));
+  // Jump into the loop after the code that copies the separator, so the first
+  // element is not preceded by a separator
+  __ jmp(&one_char_separator_loop_entry);
+
+  __ bind(&one_char_separator_loop);
+  // Live values in registers:
+  //   result_pos: the position to which we are currently copying characters.
+  //   element: Current array element.
+  //   elements_end: Array end.
+  //   separator: Single separator one-byte char (in lower byte).
+
+  // Copy the separator character to the result.
+  __ strb(separator, MemOperand(result_pos, 1, PostIndex));
+
+  // Copy next array element to the result.
+  __ bind(&one_char_separator_loop_entry);
+  __ ldr(string, MemOperand(element, kPointerSize, PostIndex));
+  __ ldr(string_length, FieldMemOperand(string, String::kLengthOffset));
+  __ SmiUntag(string_length);
+  __ add(string,
+         string,
+         Operand(SeqOneByteString::kHeaderSize - kHeapObjectTag));
+  __ CopyBytes(string, result_pos, string_length, scratch);
+  __ cmp(element, elements_end);
+  __ b(lt, &one_char_separator_loop);  // End while (element < elements_end).
+  DCHECK(result.is(r0));
+  __ b(&done);
+
+  // Long separator case (separator is more than one character). Entry is at the
+  // label long_separator below.
+  __ bind(&long_separator_loop);
+  // Live values in registers:
+  //   result_pos: the position to which we are currently copying characters.
+  //   element: Current array element.
+  //   elements_end: Array end.
+  //   separator: Separator string.
+
+  // Copy the separator to the result.
+  __ ldr(string_length, FieldMemOperand(separator, String::kLengthOffset));
+  __ SmiUntag(string_length);
+  __ add(string,
+         separator,
+         Operand(SeqOneByteString::kHeaderSize - kHeapObjectTag));
+  __ CopyBytes(string, result_pos, string_length, scratch);
+
+  __ bind(&long_separator);
+  __ ldr(string, MemOperand(element, kPointerSize, PostIndex));
+  __ ldr(string_length, FieldMemOperand(string, String::kLengthOffset));
+  __ SmiUntag(string_length);
+  __ add(string,
+         string,
+         Operand(SeqOneByteString::kHeaderSize - kHeapObjectTag));
+  __ CopyBytes(string, result_pos, string_length, scratch);
+  __ cmp(element, elements_end);
+  __ b(lt, &long_separator_loop);  // End while (element < elements_end).
+  DCHECK(result.is(r0));
+  __ b(&done);
+
+  __ bind(&bailout);
+  __ LoadRoot(r0, Heap::kUndefinedValueRootIndex);
+  __ bind(&done);
+  context()->Plug(r0);
+}
+
+
 void FullCodeGenerator::EmitDebugIsActive(CallRuntime* expr) {
   DCHECK(expr->arguments()->length() == 0);
   ExternalReference debug_is_active =

src/full-codegen/arm64/full-codegen-arm64.cc

@@ -3221,6 +3221,226 @@ void FullCodeGenerator::EmitGetSuperConstructor(CallRuntime* expr) {
 }
 
 
+void FullCodeGenerator::EmitFastOneByteArrayJoin(CallRuntime* expr) {
+  ASM_LOCATION("FullCodeGenerator::EmitFastOneByteArrayJoin");
+
+  ZoneList<Expression*>* args = expr->arguments();
+  DCHECK(args->length() == 2);
+  VisitForStackValue(args->at(1));
+  VisitForAccumulatorValue(args->at(0));
+
+  Register array = x0;
+  Register result = x0;
+  Register elements = x1;
+  Register element = x2;
+  Register separator = x3;
+  Register array_length = x4;
+  Register result_pos = x5;
+  Register map = x6;
+  Register string_length = x10;
+  Register elements_end = x11;
+  Register string = x12;
+  Register scratch1 = x13;
+  Register scratch2 = x14;
+  Register scratch3 = x7;
+  Register separator_length = x15;
+
+  Label bailout, done, one_char_separator, long_separator,
+      non_trivial_array, not_size_one_array, loop,
+      empty_separator_loop, one_char_separator_loop,
+      one_char_separator_loop_entry, long_separator_loop;
+
+  // The separator operand is on the stack.
+  __ Pop(separator);
+
+  // Check that the array is a JSArray.
+  __ JumpIfSmi(array, &bailout);
+  __ JumpIfNotObjectType(array, map, scratch1, JS_ARRAY_TYPE, &bailout);
+
+  // Check that the array has fast elements.
+  __ CheckFastElements(map, scratch1, &bailout);
+
+  // If the array has length zero, return the empty string.
+  // Load and untag the length of the array.
+  // It is an unsigned value, so we can skip sign extension.
+  // We assume little endianness.
+  __ Ldrsw(array_length,
+           UntagSmiFieldMemOperand(array, JSArray::kLengthOffset));
+  __ Cbnz(array_length, &non_trivial_array);
+  __ LoadRoot(result, Heap::kempty_stringRootIndex);
+  __ B(&done);
+
+  __ Bind(&non_trivial_array);
+  // Get the FixedArray containing array's elements.
+  __ Ldr(elements, FieldMemOperand(array, JSArray::kElementsOffset));
+
+  // Check that all array elements are sequential one-byte strings, and
+  // accumulate the sum of their lengths.
+  __ Mov(string_length, 0);
+  __ Add(element, elements, FixedArray::kHeaderSize - kHeapObjectTag);
+  __ Add(elements_end, element, Operand(array_length, LSL, kPointerSizeLog2));
+  // Loop condition: while (element < elements_end).
+  // Live values in registers:
+  //   elements: Fixed array of strings.
+  //   array_length: Length of the fixed array of strings (not smi)
+  //   separator: Separator string
+  //   string_length: Accumulated sum of string lengths (not smi).
+  //   element: Current array element.
+  //   elements_end: Array end.
+  if (FLAG_debug_code) {
+    __ Cmp(array_length, 0);
+    __ Assert(gt, kNoEmptyArraysHereInEmitFastOneByteArrayJoin);
+  }
+  __ Bind(&loop);
+  __ Ldr(string, MemOperand(element, kPointerSize, PostIndex));
+  __ JumpIfSmi(string, &bailout);
+  __ Ldr(scratch1, FieldMemOperand(string, HeapObject::kMapOffset));
+  __ Ldrb(scratch1, FieldMemOperand(scratch1, Map::kInstanceTypeOffset));
+  __ JumpIfInstanceTypeIsNotSequentialOneByte(scratch1, scratch2, &bailout);
+  __ Ldrsw(scratch1,
+           UntagSmiFieldMemOperand(string, SeqOneByteString::kLengthOffset));
+  __ Adds(string_length, string_length, scratch1);
+  __ B(vs, &bailout);
+  __ Cmp(element, elements_end);
+  __ B(lt, &loop);
+
+  // If array_length is 1, return elements[0], a string.
+  __ Cmp(array_length, 1);
+  __ B(ne, &not_size_one_array);
+  __ Ldr(result, FieldMemOperand(elements, FixedArray::kHeaderSize));
+  __ B(&done);
+
+  __ Bind(&not_size_one_array);
+
+  // Live values in registers:
+  //   separator: Separator string
+  //   array_length: Length of the array (not smi).
+  //   string_length: Sum of string lengths (not smi).
+  //   elements: FixedArray of strings.
+
+  // Check that the separator is a flat one-byte string.
+  __ JumpIfSmi(separator, &bailout);
+  __ Ldr(scratch1, FieldMemOperand(separator, HeapObject::kMapOffset));
+  __ Ldrb(scratch1, FieldMemOperand(scratch1, Map::kInstanceTypeOffset));
+  __ JumpIfInstanceTypeIsNotSequentialOneByte(scratch1, scratch2, &bailout);
+
+  // Add (separator length times array_length) - separator length to the
+  // string_length to get the length of the result string.
+  // Load the separator length as untagged.
+  // We assume little endianness, and that the length is positive.
+  __ Ldrsw(separator_length,
+           UntagSmiFieldMemOperand(separator,
+                                   SeqOneByteString::kLengthOffset));
+  __ Sub(string_length, string_length, separator_length);
+  __ Umaddl(string_length, array_length.W(), separator_length.W(),
+            string_length);
+
+  // Bailout for large object allocations.
+  __ Cmp(string_length, Page::kMaxRegularHeapObjectSize);
+  __ B(gt, &bailout);
+
+  // Get first element in the array.
+  __ Add(element, elements, FixedArray::kHeaderSize - kHeapObjectTag);
+  // Live values in registers:
+  //   element: First array element
+  //   separator: Separator string
+  //   string_length: Length of result string (not smi)
+  //   array_length: Length of the array (not smi).
+  __ AllocateOneByteString(result, string_length, scratch1, scratch2, scratch3,
+                           &bailout);
+
+  // Prepare for looping. Set up elements_end to end of the array. Set
+  // result_pos to the position of the result where to write the first
+  // character.
+  // TODO(all): useless unless AllocateOneByteString trashes the register.
+  __ Add(elements_end, element, Operand(array_length, LSL, kPointerSizeLog2));
+  __ Add(result_pos, result, SeqOneByteString::kHeaderSize - kHeapObjectTag);
+
+  // Check the length of the separator.
+  __ Cmp(separator_length, 1);
+  __ B(eq, &one_char_separator);
+  __ B(gt, &long_separator);
+
+  // Empty separator case
+  __ Bind(&empty_separator_loop);
+  // Live values in registers:
+  //   result_pos: the position to which we are currently copying characters.
+  //   element: Current array element.
+  //   elements_end: Array end.
+
+  // Copy next array element to the result.
+  __ Ldr(string, MemOperand(element, kPointerSize, PostIndex));
+  __ Ldrsw(string_length,
+           UntagSmiFieldMemOperand(string, String::kLengthOffset));
+  __ Add(string, string, SeqOneByteString::kHeaderSize - kHeapObjectTag);
+  __ CopyBytes(result_pos, string, string_length, scratch1);
+  __ Cmp(element, elements_end);
+  __ B(lt, &empty_separator_loop);  // End while (element < elements_end).
+  __ B(&done);
+
+  // One-character separator case
+  __ Bind(&one_char_separator);
+  // Replace separator with its one-byte character value.
+  __ Ldrb(separator, FieldMemOperand(separator, SeqOneByteString::kHeaderSize));
+  // Jump into the loop after the code that copies the separator, so the first
+  // element is not preceded by a separator
+  __ B(&one_char_separator_loop_entry);
+
+  __ Bind(&one_char_separator_loop);
+  // Live values in registers:
+  //   result_pos: the position to which we are currently copying characters.
+  //   element: Current array element.
+  //   elements_end: Array end.
+  //   separator: Single separator one-byte char (in lower byte).
+
+  // Copy the separator character to the result.
+  __ Strb(separator, MemOperand(result_pos, 1, PostIndex));
+
+  // Copy next array element to the result.
+  __ Bind(&one_char_separator_loop_entry);
+  __ Ldr(string, MemOperand(element, kPointerSize, PostIndex));
+  __ Ldrsw(string_length,
+           UntagSmiFieldMemOperand(string, String::kLengthOffset));
+  __ Add(string, string, SeqOneByteString::kHeaderSize - kHeapObjectTag);
+  __ CopyBytes(result_pos, string, string_length, scratch1);
+  __ Cmp(element, elements_end);
+  __ B(lt, &one_char_separator_loop);  // End while (element < elements_end).
+  __ B(&done);
+
+  // Long separator case (separator is more than one character). Entry is at the
+  // label long_separator below.
+  __ Bind(&long_separator_loop);
+  // Live values in registers:
+  //   result_pos: the position to which we are currently copying characters.
+  //   element: Current array element.
+  //   elements_end: Array end.
+  //   separator: Separator string.
+
+  // Copy the separator to the result.
+  // TODO(all): hoist next two instructions.
+  __ Ldrsw(string_length,
+           UntagSmiFieldMemOperand(separator, String::kLengthOffset));
+  __ Add(string, separator, SeqOneByteString::kHeaderSize - kHeapObjectTag);
+  __ CopyBytes(result_pos, string, string_length, scratch1);
+
+  __ Bind(&long_separator);
+  __ Ldr(string, MemOperand(element, kPointerSize, PostIndex));
+  __ Ldrsw(string_length,
+           UntagSmiFieldMemOperand(string, String::kLengthOffset));
+  __ Add(string, string, SeqOneByteString::kHeaderSize - kHeapObjectTag);
+  __ CopyBytes(result_pos, string, string_length, scratch1);
+  __ Cmp(element, elements_end);
+  __ B(lt, &long_separator_loop);  // End while (element < elements_end).
+  __ B(&done);
+
+  __ Bind(&bailout);
+  // Returning undefined will force slower code to handle it.
+  __ LoadRoot(result, Heap::kUndefinedValueRootIndex);
+  __ Bind(&done);
+  context()->Plug(result);
+}
+
+
 void FullCodeGenerator::EmitDebugIsActive(CallRuntime* expr) {
   DCHECK(expr->arguments()->length() == 0);
   ExternalReference debug_is_active =

src/full-codegen/full-codegen.h

@@ -535,6 +535,7 @@ class FullCodeGenerator: public AstVisitor {
   F(HasCachedArrayIndex)                \
   F(GetCachedArrayIndex)                \
   F(GetSuperConstructor)                \
+  F(FastOneByteArrayJoin)               \
   F(GeneratorNext)                      \
   F(GeneratorReturn)                    \
   F(GeneratorThrow)                     \

src/full-codegen/ia32/full-codegen-ia32.cc

@@ -3304,6 +3304,275 @@ void FullCodeGenerator::EmitGetSuperConstructor(CallRuntime* expr) {
 }
 
 
+void FullCodeGenerator::EmitFastOneByteArrayJoin(CallRuntime* expr) {
+  Label bailout, done, one_char_separator, long_separator,
+      non_trivial_array, not_size_one_array, loop,
+      loop_1, loop_1_condition, loop_2, loop_2_entry, loop_3, loop_3_entry;
+
+  ZoneList<Expression*>* args = expr->arguments();
+  DCHECK(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 elements = no_reg;  // Will be eax.
+
+  Register index = edx;
+
+  Register string_length = ecx;
+
+  Register string = esi;
+
+  Register scratch = ebx;
+
+  Register array_length = edi;
+  Register result_pos = no_reg;  // Will be edi.
+
+  // Separator operand is already pushed.
+  Operand separator_operand = Operand(esp, 2 * kPointerSize);
+  Operand result_operand = Operand(esp, 1 * kPointerSize);
+  Operand array_length_operand = Operand(esp, 0);
+  __ sub(esp, Immediate(2 * kPointerSize));
+  __ cld();
+  // Check that the array is a JSArray
+  __ JumpIfSmi(array, &bailout);
+  __ CmpObjectType(array, JS_ARRAY_TYPE, scratch);
+  __ j(not_equal, &bailout);
+
+  // Check that the array has fast elements.
+  __ CheckFastElements(scratch, &bailout);
+
+  // If the array has length zero, return the empty string.
+  __ mov(array_length, FieldOperand(array, JSArray::kLengthOffset));
+  __ SmiUntag(array_length);
+  __ j(not_zero, &non_trivial_array);
+  __ mov(result_operand, isolate()->factory()->empty_string());
+  __ jmp(&done);
+
+  // Save the array length.
+  __ bind(&non_trivial_array);
+  __ mov(array_length_operand, array_length);
+
+  // Save the FixedArray containing array's elements.
+  // End of array's live range.
+  elements = array;
+  __ mov(elements, FieldOperand(array, JSArray::kElementsOffset));
+  array = no_reg;
+
+
+  // Check that all array elements are sequential one-byte strings, and
+  // accumulate the sum of their lengths, as a smi-encoded value.
+  __ Move(index, Immediate(0));
+  __ Move(string_length, Immediate(0));
+  // Loop condition: while (index < length).
+  // Live loop registers: index, array_length, string,
+  //                      scratch, string_length, elements.
+  if (generate_debug_code_) {
+    __ cmp(index, array_length);
+    __ Assert(less, kNoEmptyArraysHereInEmitFastOneByteArrayJoin);
+  }
+  __ bind(&loop);
+  __ mov(string, FieldOperand(elements,
+                              index,
+                              times_pointer_size,
+                              FixedArray::kHeaderSize));
+  __ JumpIfSmi(string, &bailout);
+  __ mov(scratch, FieldOperand(string, HeapObject::kMapOffset));
+  __ movzx_b(scratch, FieldOperand(scratch, Map::kInstanceTypeOffset));
+  __ and_(scratch, Immediate(
+      kIsNotStringMask | kStringEncodingMask | kStringRepresentationMask));
+  __ cmp(scratch, kStringTag | kOneByteStringTag | kSeqStringTag);
+  __ j(not_equal, &bailout);
+  __ add(string_length,
+         FieldOperand(string, SeqOneByteString::kLengthOffset));
+  __ j(overflow, &bailout);
+  __ add(index, Immediate(1));
+  __ cmp(index, 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 one-byte string.
+  __ mov(string, separator_operand);
+  __ JumpIfSmi(string, &bailout);
+  __ mov(scratch, FieldOperand(string, HeapObject::kMapOffset));
+  __ movzx_b(scratch, FieldOperand(scratch, Map::kInstanceTypeOffset));
+  __ and_(scratch, Immediate(
+      kIsNotStringMask | kStringEncodingMask | kStringRepresentationMask));
+  __ cmp(scratch, kStringTag | kOneByteStringTag | kSeqStringTag);
+  __ j(not_equal, &bailout);
+
+  // Add (separator length times array_length) - separator length
+  // to string_length.
+  __ mov(scratch, separator_operand);
+  __ mov(scratch, FieldOperand(scratch, SeqOneByteString::kLengthOffset));
+  __ sub(string_length, scratch);  // May be negative, temporarily.
+  __ imul(scratch, array_length_operand);
+  __ j(overflow, &bailout);
+  __ add(string_length, scratch);
+  __ j(overflow, &bailout);
+
+  __ shr(string_length, 1);
+
+  // Bailout for large object allocations.
+  __ cmp(string_length, Page::kMaxRegularHeapObjectSize);
+  __ j(greater, &bailout);
+
+  // Live registers and stack values:
+  //   string_length
+  //   elements
+  __ AllocateOneByteString(result_pos, string_length, scratch, index, string,
+                           &bailout);
+  __ mov(result_operand, result_pos);
+  __ lea(result_pos, FieldOperand(result_pos, SeqOneByteString::kHeaderSize));
+
+
+  __ mov(string, separator_operand);
+  __ cmp(FieldOperand(string, SeqOneByteString::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, SeqOneByteString::kHeaderSize));
+  __ CopyBytes(string, result_pos, string_length, scratch);
+  __ add(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 one-byte character value.
+  __ mov_b(scratch, FieldOperand(string, SeqOneByteString::kHeaderSize));
+  __ mov_b(separator_operand, scratch);
+
+  __ Move(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, SeqOneByteString::kHeaderSize));
+  __ CopyBytes(string, result_pos, string_length, scratch);
+  __ add(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);
+
+  __ Move(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, SeqOneByteString::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, SeqOneByteString::kHeaderSize));
+  __ CopyBytes(string, result_pos, string_length, scratch);
+  __ add(index, Immediate(1));
+
+  __ cmp(index, array_length_operand);
+  __ j(less, &loop_3);  // End while (index < length).
+  __ jmp(&done);
+
+
+  __ bind(&bailout);
+  __ mov(result_operand, isolate()->factory()->undefined_value());
+  __ bind(&done);
+  __ mov(eax, result_operand);
+  // Drop temp values from the stack, and restore context register.
+  __ add(esp, Immediate(3 * kPointerSize));
+
+  __ mov(esi, Operand(ebp, StandardFrameConstants::kContextOffset));
+  context()->Plug(eax);
+}
+
+
 void FullCodeGenerator::EmitDebugIsActive(CallRuntime* expr) {
   DCHECK(expr->arguments()->length() == 0);
   ExternalReference debug_is_active =

src/full-codegen/mips/full-codegen-mips.cc

@@ -3416,6 +3416,238 @@ void FullCodeGenerator::EmitGetSuperConstructor(CallRuntime* expr) {
 }
 
 
+void FullCodeGenerator::EmitFastOneByteArrayJoin(CallRuntime* expr) {
+  Label bailout, done, one_char_separator, long_separator,
+      non_trivial_array, not_size_one_array, loop,
+      empty_separator_loop, one_char_separator_loop,
+      one_char_separator_loop_entry, long_separator_loop;
+  ZoneList<Expression*>* args = expr->arguments();
+  DCHECK(args->length() == 2);
+  VisitForStackValue(args->at(1));
+  VisitForAccumulatorValue(args->at(0));
+
+  // All aliases of the same register have disjoint lifetimes.
+  Register array = v0;
+  Register elements = no_reg;  // Will be v0.
+  Register result = no_reg;  // Will be v0.
+  Register separator = a1;
+  Register array_length = a2;
+  Register result_pos = no_reg;  // Will be a2.
+  Register string_length = a3;
+  Register string = t0;
+  Register element = t1;
+  Register elements_end = t2;
+  Register scratch1 = t3;
+  Register scratch2 = t5;
+  Register scratch3 = t4;
+
+  // Separator operand is on the stack.
+  __ pop(separator);
+
+  // Check that the array is a JSArray.
+  __ JumpIfSmi(array, &bailout);
+  __ GetObjectType(array, scratch1, scratch2);
+  __ Branch(&bailout, ne, scratch2, Operand(JS_ARRAY_TYPE));
+
+  // Check that the array has fast elements.
+  __ CheckFastElements(scratch1, scratch2, &bailout);
+
+  // If the array has length zero, return the empty string.
+  __ lw(array_length, FieldMemOperand(array, JSArray::kLengthOffset));
+  __ SmiUntag(array_length);
+  __ Branch(&non_trivial_array, ne, array_length, Operand(zero_reg));
+  __ LoadRoot(v0, Heap::kempty_stringRootIndex);
+  __ Branch(&done);
+
+  __ bind(&non_trivial_array);
+
+  // Get the FixedArray containing array's elements.
+  elements = array;
+  __ lw(elements, FieldMemOperand(array, JSArray::kElementsOffset));
+  array = no_reg;  // End of array's live range.
+
+  // Check that all array elements are sequential one-byte strings, and
+  // accumulate the sum of their lengths, as a smi-encoded value.
+  __ mov(string_length, zero_reg);
+  __ Addu(element,
+          elements, Operand(FixedArray::kHeaderSize - kHeapObjectTag));
+  __ Lsa(elements_end, element, array_length, kPointerSizeLog2);
+  // Loop condition: while (element < elements_end).
+  // Live values in registers:
+  //   elements: Fixed array of strings.
+  //   array_length: Length of the fixed array of strings (not smi)
+  //   separator: Separator string
+  //   string_length: Accumulated sum of string lengths (smi).
+  //   element: Current array element.
+  //   elements_end: Array end.
+  if (generate_debug_code_) {
+    __ Assert(gt, kNoEmptyArraysHereInEmitFastOneByteArrayJoin, array_length,
+              Operand(zero_reg));
+  }
+  __ bind(&loop);
+  __ lw(string, MemOperand(element));
+  __ Addu(element, element, kPointerSize);
+  __ JumpIfSmi(string, &bailout);
+  __ lw(scratch1, FieldMemOperand(string, HeapObject::kMapOffset));
+  __ lbu(scratch1, FieldMemOperand(scratch1, Map::kInstanceTypeOffset));
+  __ JumpIfInstanceTypeIsNotSequentialOneByte(scratch1, scratch2, &bailout);
+  __ lw(scratch1, FieldMemOperand(string, SeqOneByteString::kLengthOffset));
+  __ AddBranchOvf(string_length, string_length, Operand(scratch1), &bailout);
+  __ Branch(&loop, lt, element, Operand(elements_end));
+
+  // If array_length is 1, return elements[0], a string.
+  __ Branch(&not_size_one_array, ne, array_length, Operand(1));
+  __ lw(v0, FieldMemOperand(elements, FixedArray::kHeaderSize));
+  __ Branch(&done);
+
+  __ bind(&not_size_one_array);
+
+  // Live values in registers:
+  //   separator: Separator string
+  //   array_length: Length of the array.
+  //   string_length: Sum of string lengths (smi).
+  //   elements: FixedArray of strings.
+
+  // Check that the separator is a flat one-byte string.
+  __ JumpIfSmi(separator, &bailout);
+  __ lw(scratch1, FieldMemOperand(separator, HeapObject::kMapOffset));
+  __ lbu(scratch1, FieldMemOperand(scratch1, Map::kInstanceTypeOffset));
+  __ JumpIfInstanceTypeIsNotSequentialOneByte(scratch1, scratch2, &bailout);
+
+  // Add (separator length times array_length) - separator length to the
+  // string_length to get the length of the result string. array_length is not
+  // smi but the other values are, so the result is a smi.
+  __ lw(scratch1, FieldMemOperand(separator, SeqOneByteString::kLengthOffset));
+  __ Subu(string_length, string_length, Operand(scratch1));
+  __ Mul(scratch3, scratch2, array_length, scratch1);
+  // Check for smi overflow. No overflow if higher 33 bits of 64-bit result are
+  // zero.
+  __ Branch(&bailout, ne, scratch3, Operand(zero_reg));
+  __ And(scratch3, scratch2, Operand(0x80000000));
+  __ Branch(&bailout, ne, scratch3, Operand(zero_reg));
+  __ AddBranchOvf(string_length, string_length, Operand(scratch2), &bailout);
+  __ SmiUntag(string_length);
+
+  // Bailout for large object allocations.
+  __ Branch(&bailout, gt, string_length,
+            Operand(Page::kMaxRegularHeapObjectSize));
+
+  // Get first element in the array to free up the elements register to be used
+  // for the result.
+  __ Addu(element,
+          elements, Operand(FixedArray::kHeaderSize - kHeapObjectTag));
+  result = elements;  // End of live range for elements.
+  elements = no_reg;
+  // Live values in registers:
+  //   element: First array element
+  //   separator: Separator string
+  //   string_length: Length of result string (not smi)
+  //   array_length: Length of the array.
+  __ AllocateOneByteString(result, string_length, scratch1, scratch2,
+                           elements_end, &bailout);
+  // Prepare for looping. Set up elements_end to end of the array. Set
+  // result_pos to the position of the result where to write the first
+  // character.
+  __ Lsa(elements_end, element, array_length, kPointerSizeLog2);
+  result_pos = array_length;  // End of live range for array_length.
+  array_length = no_reg;
+  __ Addu(result_pos,
+          result,
+          Operand(SeqOneByteString::kHeaderSize - kHeapObjectTag));
+
+  // Check the length of the separator.
+  __ lw(scratch1, FieldMemOperand(separator, SeqOneByteString::kLengthOffset));
+  __ li(at, Operand(Smi::FromInt(1)));
+  __ Branch(&one_char_separator, eq, scratch1, Operand(at));
+  __ Branch(&long_separator, gt, scratch1, Operand(at));
+
+  // Empty separator case.
+  __ bind(&empty_separator_loop);
+  // Live values in registers:
+  //   result_pos: the position to which we are currently copying characters.
+  //   element: Current array element.
+  //   elements_end: Array end.
+
+  // Copy next array element to the result.
+  __ lw(string, MemOperand(element));
+  __ Addu(element, element, kPointerSize);
+  __ lw(string_length, FieldMemOperand(string, String::kLengthOffset));
+  __ SmiUntag(string_length);
+  __ Addu(string, string, SeqOneByteString::kHeaderSize - kHeapObjectTag);
+  __ CopyBytes(string, result_pos, string_length, scratch1);
+  // End while (element < elements_end).
+  __ Branch(&empty_separator_loop, lt, element, Operand(elements_end));
+  DCHECK(result.is(v0));
+  __ Branch(&done);
+
+  // One-character separator case.
+  __ bind(&one_char_separator);
+  // Replace separator with its one-byte character value.
+  __ lbu(separator, FieldMemOperand(separator, SeqOneByteString::kHeaderSize));
+  // Jump into the loop after the code that copies the separator, so the first
+  // element is not preceded by a separator.
+  __ jmp(&one_char_separator_loop_entry);
+
+  __ bind(&one_char_separator_loop);
+  // Live values in registers:
+  //   result_pos: the position to which we are currently copying characters.
+  //   element: Current array element.
+  //   elements_end: Array end.
+  //   separator: Single separator one-byte char (in lower byte).
+
+  // Copy the separator character to the result.
+  __ sb(separator, MemOperand(result_pos));
+  __ Addu(result_pos, result_pos, 1);
+
+  // Copy next array element to the result.
+  __ bind(&one_char_separator_loop_entry);
+  __ lw(string, MemOperand(element));
+  __ Addu(element, element, kPointerSize);
+  __ lw(string_length, FieldMemOperand(string, String::kLengthOffset));
+  __ SmiUntag(string_length);
+  __ Addu(string, string, SeqOneByteString::kHeaderSize - kHeapObjectTag);
+  __ CopyBytes(string, result_pos, string_length, scratch1);
+  // End while (element < elements_end).
+  __ Branch(&one_char_separator_loop, lt, element, Operand(elements_end));
+  DCHECK(result.is(v0));
+  __ Branch(&done);
+
+  // Long separator case (separator is more than one character). Entry is at the
+  // label long_separator below.
+  __ bind(&long_separator_loop);
+  // Live values in registers:
+  //   result_pos: the position to which we are currently copying characters.
+  //   element: Current array element.
+  //   elements_end: Array end.
+  //   separator: Separator string.
+
+  // Copy the separator to the result.
+  __ lw(string_length, FieldMemOperand(separator, String::kLengthOffset));
+  __ SmiUntag(string_length);
+  __ Addu(string,
+          separator,
+          Operand(SeqOneByteString::kHeaderSize - kHeapObjectTag));
+  __ CopyBytes(string, result_pos, string_length, scratch1);
+
+  __ bind(&long_separator);
+  __ lw(string, MemOperand(element));
+  __ Addu(element, element, kPointerSize);
+  __ lw(string_length, FieldMemOperand(string, String::kLengthOffset));
+  __ SmiUntag(string_length);
+  __ Addu(string, string, SeqOneByteString::kHeaderSize - kHeapObjectTag);
+  __ CopyBytes(string, result_pos, string_length, scratch1);
+  // End while (element < elements_end).
+  __ Branch(&long_separator_loop, lt, element, Operand(elements_end));
+  DCHECK(result.is(v0));
+  __ Branch(&done);
+
+  __ bind(&bailout);
+  __ LoadRoot(v0, Heap::kUndefinedValueRootIndex);
+  __ bind(&done);
+  context()->Plug(v0);
+}
+
+
 void FullCodeGenerator::EmitDebugIsActive(CallRuntime* expr) {
   DCHECK(expr->arguments()->length() == 0);
   ExternalReference debug_is_active =

src/full-codegen/mips64/full-codegen-mips64.cc

@@ -3422,6 +3422,240 @@ void FullCodeGenerator::EmitGetSuperConstructor(CallRuntime* expr) {
 }
 
 
+void FullCodeGenerator::EmitFastOneByteArrayJoin(CallRuntime* expr) {
+  Label bailout, done, one_char_separator, long_separator,
+      non_trivial_array, not_size_one_array, loop,
+      empty_separator_loop, one_char_separator_loop,
+      one_char_separator_loop_entry, long_separator_loop;
+  ZoneList<Expression*>* args = expr->arguments();
+  DCHECK(args->length() == 2);
+  VisitForStackValue(args->at(1));
+  VisitForAccumulatorValue(args->at(0));
+
+  // All aliases of the same register have disjoint lifetimes.
+  Register array = v0;
+  Register elements = no_reg;  // Will be v0.
+  Register result = no_reg;  // Will be v0.
+  Register separator = a1;
+  Register array_length = a2;
+  Register result_pos = no_reg;  // Will be a2.
+  Register string_length = a3;
+  Register string = a4;
+  Register element = a5;
+  Register elements_end = a6;
+  Register scratch1 = a7;
+  Register scratch2 = t1;
+  Register scratch3 = t0;
+
+  // Separator operand is on the stack.
+  __ pop(separator);
+
+  // Check that the array is a JSArray.
+  __ JumpIfSmi(array, &bailout);
+  __ GetObjectType(array, scratch1, scratch2);
+  __ Branch(&bailout, ne, scratch2, Operand(JS_ARRAY_TYPE));
+
+  // Check that the array has fast elements.
+  __ CheckFastElements(scratch1, scratch2, &bailout);
+
+  // If the array has length zero, return the empty string.
+  __ ld(array_length, FieldMemOperand(array, JSArray::kLengthOffset));
+  __ SmiUntag(array_length);
+  __ Branch(&non_trivial_array, ne, array_length, Operand(zero_reg));
+  __ LoadRoot(v0, Heap::kempty_stringRootIndex);
+  __ Branch(&done);
+
+  __ bind(&non_trivial_array);
+
+  // Get the FixedArray containing array's elements.
+  elements = array;
+  __ ld(elements, FieldMemOperand(array, JSArray::kElementsOffset));
+  array = no_reg;  // End of array's live range.
+
+  // Check that all array elements are sequential one-byte strings, and
+  // accumulate the sum of their lengths, as a smi-encoded value.
+  __ mov(string_length, zero_reg);
+  __ Daddu(element,
+          elements, Operand(FixedArray::kHeaderSize - kHeapObjectTag));
+  __ Dlsa(elements_end, element, array_length, kPointerSizeLog2);
+  // Loop condition: while (element < elements_end).
+  // Live values in registers:
+  //   elements: Fixed array of strings.
+  //   array_length: Length of the fixed array of strings (not smi)
+  //   separator: Separator string
+  //   string_length: Accumulated sum of string lengths (smi).
+  //   element: Current array element.
+  //   elements_end: Array end.
+  if (generate_debug_code_) {
+    __ Assert(gt, kNoEmptyArraysHereInEmitFastOneByteArrayJoin, array_length,
+              Operand(zero_reg));
+  }
+  __ bind(&loop);
+  __ ld(string, MemOperand(element));
+  __ Daddu(element, element, kPointerSize);
+  __ JumpIfSmi(string, &bailout);
+  __ ld(scratch1, FieldMemOperand(string, HeapObject::kMapOffset));
+  __ lbu(scratch1, FieldMemOperand(scratch1, Map::kInstanceTypeOffset));
+  __ JumpIfInstanceTypeIsNotSequentialOneByte(scratch1, scratch2, &bailout);
+  __ ld(scratch1, FieldMemOperand(string, SeqOneByteString::kLengthOffset));
+  __ DadduAndCheckForOverflow(string_length, string_length, scratch1, scratch3);
+  __ BranchOnOverflow(&bailout, scratch3);
+  __ Branch(&loop, lt, element, Operand(elements_end));
+
+  // If array_length is 1, return elements[0], a string.
+  __ Branch(&not_size_one_array, ne, array_length, Operand(1));
+  __ ld(v0, FieldMemOperand(elements, FixedArray::kHeaderSize));
+  __ Branch(&done);
+
+  __ bind(&not_size_one_array);
+
+  // Live values in registers:
+  //   separator: Separator string
+  //   array_length: Length of the array.
+  //   string_length: Sum of string lengths (smi).
+  //   elements: FixedArray of strings.
+
+  // Check that the separator is a flat one-byte string.
+  __ JumpIfSmi(separator, &bailout);
+  __ ld(scratch1, FieldMemOperand(separator, HeapObject::kMapOffset));
+  __ lbu(scratch1, FieldMemOperand(scratch1, Map::kInstanceTypeOffset));
+  __ JumpIfInstanceTypeIsNotSequentialOneByte(scratch1, scratch2, &bailout);
+
+  // Add (separator length times array_length) - separator length to the
+  // string_length to get the length of the result string. array_length is not
+  // smi but the other values are, so the result is a smi.
+  __ ld(scratch1, FieldMemOperand(separator, SeqOneByteString::kLengthOffset));
+  __ Dsubu(string_length, string_length, Operand(scratch1));
+  __ SmiUntag(scratch1);
+  __ Dmul(scratch2, array_length, scratch1);
+  // Check for smi overflow. No overflow if higher 33 bits of 64-bit result are
+  // zero.
+  __ dsra32(scratch1, scratch2, 0);
+  __ Branch(&bailout, ne, scratch2, Operand(zero_reg));
+  __ SmiUntag(string_length);
+  __ AdduAndCheckForOverflow(string_length, string_length, scratch2, scratch3);
+  __ BranchOnOverflow(&bailout, scratch3);
+
+  // Bailout for large object allocations.
+  __ Branch(&bailout, gt, string_length,
+            Operand(Page::kMaxRegularHeapObjectSize));
+
+  // Get first element in the array to free up the elements register to be used
+  // for the result.
+  __ Daddu(element,
+          elements, Operand(FixedArray::kHeaderSize - kHeapObjectTag));
+  result = elements;  // End of live range for elements.
+  elements = no_reg;
+  // Live values in registers:
+  //   element: First array element
+  //   separator: Separator string
+  //   string_length: Length of result string (not smi)
+  //   array_length: Length of the array.
+  __ AllocateOneByteString(result, string_length, scratch1, scratch2,
+                           elements_end, &bailout);
+  // Prepare for looping. Set up elements_end to end of the array. Set
+  // result_pos to the position of the result where to write the first
+  // character.
+  __ Dlsa(elements_end, element, array_length, kPointerSizeLog2);
+  result_pos = array_length;  // End of live range for array_length.
+  array_length = no_reg;
+  __ Daddu(result_pos,
+          result,
+          Operand(SeqOneByteString::kHeaderSize - kHeapObjectTag));
+
+  // Check the length of the separator.
+  __ ld(scratch1, FieldMemOperand(separator, SeqOneByteString::kLengthOffset));
+  __ li(at, Operand(Smi::FromInt(1)));
+  __ Branch(&one_char_separator, eq, scratch1, Operand(at));
+  __ Branch(&long_separator, gt, scratch1, Operand(at));
+
+  // Empty separator case.
+  __ bind(&empty_separator_loop);
+  // Live values in registers:
+  //   result_pos: the position to which we are currently copying characters.
+  //   element: Current array element.
+  //   elements_end: Array end.
+
+  // Copy next array element to the result.
+  __ ld(string, MemOperand(element));
+  __ Daddu(element, element, kPointerSize);
+  __ ld(string_length, FieldMemOperand(string, String::kLengthOffset));
+  __ SmiUntag(string_length);
+  __ Daddu(string, string, SeqOneByteString::kHeaderSize - kHeapObjectTag);
+  __ CopyBytes(string, result_pos, string_length, scratch1);
+  // End while (element < elements_end).
+  __ Branch(&empty_separator_loop, lt, element, Operand(elements_end));
+  DCHECK(result.is(v0));
+  __ Branch(&done);
+
+  // One-character separator case.
+  __ bind(&one_char_separator);
+  // Replace separator with its one-byte character value.
+  __ lbu(separator, FieldMemOperand(separator, SeqOneByteString::kHeaderSize));
+  // Jump into the loop after the code that copies the separator, so the first
+  // element is not preceded by a separator.
+  __ jmp(&one_char_separator_loop_entry);
+
+  __ bind(&one_char_separator_loop);
+  // Live values in registers:
+  //   result_pos: the position to which we are currently copying characters.
+  //   element: Current array element.
+  //   elements_end: Array end.
+  //   separator: Single separator one-byte char (in lower byte).
+
+  // Copy the separator character to the result.
+  __ sb(separator, MemOperand(result_pos));
+  __ Daddu(result_pos, result_pos, 1);
+
+  // Copy next array element to the result.
+  __ bind(&one_char_separator_loop_entry);
+  __ ld(string, MemOperand(element));
+  __ Daddu(element, element, kPointerSize);
+  __ ld(string_length, FieldMemOperand(string, String::kLengthOffset));
+  __ SmiUntag(string_length);
+  __ Daddu(string, string, SeqOneByteString::kHeaderSize - kHeapObjectTag);
+  __ CopyBytes(string, result_pos, string_length, scratch1);
+  // End while (element < elements_end).
+  __ Branch(&one_char_separator_loop, lt, element, Operand(elements_end));
+  DCHECK(result.is(v0));
+  __ Branch(&done);
+
+  // Long separator case (separator is more than one character). Entry is at the
+  // label long_separator below.
+  __ bind(&long_separator_loop);
+  // Live values in registers:
+  //   result_pos: the position to which we are currently copying characters.
+  //   element: Current array element.
+  //   elements_end: Array end.
+  //   separator: Separator string.
+
+  // Copy the separator to the result.
+  __ ld(string_length, FieldMemOperand(separator, String::kLengthOffset));
+  __ SmiUntag(string_length);
+  __ Daddu(string,
+          separator,
+          Operand(SeqOneByteString::kHeaderSize - kHeapObjectTag));
+  __ CopyBytes(string, result_pos, string_length, scratch1);
+
+  __ bind(&long_separator);
+  __ ld(string, MemOperand(element));
+  __ Daddu(element, element, kPointerSize);
+  __ ld(string_length, FieldMemOperand(string, String::kLengthOffset));
+  __ SmiUntag(string_length);
+  __ Daddu(string, string, SeqOneByteString::kHeaderSize - kHeapObjectTag);
+  __ CopyBytes(string, result_pos, string_length, scratch1);
+  // End while (element < elements_end).
+  __ Branch(&long_separator_loop, lt, element, Operand(elements_end));
+  DCHECK(result.is(v0));
+  __ Branch(&done);
+
+  __ bind(&bailout);
+  __ LoadRoot(v0, Heap::kUndefinedValueRootIndex);
+  __ bind(&done);
+  context()->Plug(v0);
+}
+
+
 void FullCodeGenerator::EmitDebugIsActive(CallRuntime* expr) {
   DCHECK(expr->arguments()->length() == 0);
   ExternalReference debug_is_active =

src/full-codegen/x87/full-codegen-x87.cc

@@ -3296,6 +3296,275 @@ void FullCodeGenerator::EmitGetSuperConstructor(CallRuntime* expr) {
 }
 
 
+void FullCodeGenerator::EmitFastOneByteArrayJoin(CallRuntime* expr) {
+  Label bailout, done, one_char_separator, long_separator,
+      non_trivial_array, not_size_one_array, loop,
+      loop_1, loop_1_condition, loop_2, loop_2_entry, loop_3, loop_3_entry;
+
+  ZoneList<Expression*>* args = expr->arguments();
+  DCHECK(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 elements = no_reg;  // Will be eax.
+
+  Register index = edx;
+
+  Register string_length = ecx;
+
+  Register string = esi;
+
+  Register scratch = ebx;
+
+  Register array_length = edi;
+  Register result_pos = no_reg;  // Will be edi.
+
+  // Separator operand is already pushed.
+  Operand separator_operand = Operand(esp, 2 * kPointerSize);
+  Operand result_operand = Operand(esp, 1 * kPointerSize);
+  Operand array_length_operand = Operand(esp, 0);
+  __ sub(esp, Immediate(2 * kPointerSize));
+  __ cld();
+  // Check that the array is a JSArray
+  __ JumpIfSmi(array, &bailout);
+  __ CmpObjectType(array, JS_ARRAY_TYPE, scratch);
+  __ j(not_equal, &bailout);
+
+  // Check that the array has fast elements.
+  __ CheckFastElements(scratch, &bailout);
+
+  // If the array has length zero, return the empty string.
+  __ mov(array_length, FieldOperand(array, JSArray::kLengthOffset));
+  __ SmiUntag(array_length);
+  __ j(not_zero, &non_trivial_array);
+  __ mov(result_operand, isolate()->factory()->empty_string());
+  __ jmp(&done);
+
+  // Save the array length.
+  __ bind(&non_trivial_array);
+  __ mov(array_length_operand, array_length);
+
+  // Save the FixedArray containing array's elements.
+  // End of array's live range.
+  elements = array;
+  __ mov(elements, FieldOperand(array, JSArray::kElementsOffset));
+  array = no_reg;
+
+
+  // Check that all array elements are sequential one-byte strings, and
+  // accumulate the sum of their lengths, as a smi-encoded value.
+  __ Move(index, Immediate(0));
+  __ Move(string_length, Immediate(0));
+  // Loop condition: while (index < length).
+  // Live loop registers: index, array_length, string,
+  //                      scratch, string_length, elements.
+  if (generate_debug_code_) {
+    __ cmp(index, array_length);
+    __ Assert(less, kNoEmptyArraysHereInEmitFastOneByteArrayJoin);
+  }
+  __ bind(&loop);
+  __ mov(string, FieldOperand(elements,
+                              index,
+                              times_pointer_size,
+                              FixedArray::kHeaderSize));
+  __ JumpIfSmi(string, &bailout);
+  __ mov(scratch, FieldOperand(string, HeapObject::kMapOffset));
+  __ movzx_b(scratch, FieldOperand(scratch, Map::kInstanceTypeOffset));
+  __ and_(scratch, Immediate(
+      kIsNotStringMask | kStringEncodingMask | kStringRepresentationMask));
+  __ cmp(scratch, kStringTag | kOneByteStringTag | kSeqStringTag);
+  __ j(not_equal, &bailout);
+  __ add(string_length,
+         FieldOperand(string, SeqOneByteString::kLengthOffset));
+  __ j(overflow, &bailout);
+  __ add(index, Immediate(1));
+  __ cmp(index, 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 one-byte string.
+  __ mov(string, separator_operand);
+  __ JumpIfSmi(string, &bailout);
+  __ mov(scratch, FieldOperand(string, HeapObject::kMapOffset));
+  __ movzx_b(scratch, FieldOperand(scratch, Map::kInstanceTypeOffset));
+  __ and_(scratch, Immediate(
+      kIsNotStringMask | kStringEncodingMask | kStringRepresentationMask));
+  __ cmp(scratch, kStringTag | kOneByteStringTag | kSeqStringTag);
+  __ j(not_equal, &bailout);
+
+  // Add (separator length times array_length) - separator length
+  // to string_length.
+  __ mov(scratch, separator_operand);
+  __ mov(scratch, FieldOperand(scratch, SeqOneByteString::kLengthOffset));
+  __ sub(string_length, scratch);  // May be negative, temporarily.
+  __ imul(scratch, array_length_operand);
+  __ j(overflow, &bailout);
+  __ add(string_length, scratch);
+  __ j(overflow, &bailout);
+
+  __ shr(string_length, 1);
+
+  // Bailout for large object allocations.
+  __ cmp(string_length, Page::kMaxRegularHeapObjectSize);
+  __ j(greater, &bailout);
+
+  // Live registers and stack values:
+  //   string_length
+  //   elements
+  __ AllocateOneByteString(result_pos, string_length, scratch, index, string,
+                           &bailout);
+  __ mov(result_operand, result_pos);
+  __ lea(result_pos, FieldOperand(result_pos, SeqOneByteString::kHeaderSize));
+
+
+  __ mov(string, separator_operand);
+  __ cmp(FieldOperand(string, SeqOneByteString::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, SeqOneByteString::kHeaderSize));
+  __ CopyBytes(string, result_pos, string_length, scratch);
+  __ add(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 one-byte character value.
+  __ mov_b(scratch, FieldOperand(string, SeqOneByteString::kHeaderSize));
+  __ mov_b(separator_operand, scratch);
+
+  __ Move(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, SeqOneByteString::kHeaderSize));
+  __ CopyBytes(string, result_pos, string_length, scratch);
+  __ add(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);
+
+  __ Move(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, SeqOneByteString::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, SeqOneByteString::kHeaderSize));
+  __ CopyBytes(string, result_pos, string_length, scratch);
+  __ add(index, Immediate(1));
+
+  __ cmp(index, array_length_operand);
+  __ j(less, &loop_3);  // End while (index < length).
+  __ jmp(&done);
+
+
+  __ bind(&bailout);
+  __ mov(result_operand, isolate()->factory()->undefined_value());
+  __ bind(&done);
+  __ mov(eax, result_operand);
+  // Drop temp values from the stack, and restore context register.
+  __ add(esp, Immediate(3 * kPointerSize));
+
+  __ mov(esi, Operand(ebp, StandardFrameConstants::kContextOffset));
+  context()->Plug(eax);
+}
+
+
 void FullCodeGenerator::EmitDebugIsActive(CallRuntime* expr) {
   DCHECK(expr->arguments()->length() == 0);
   ExternalReference debug_is_active =

src/js/array.js

@@ -209,6 +209,8 @@ function Join(array, length, separator, convert) {
         elements[elements_length++] = e;
       }
       elements.length = elements_length;
+      var result = %_FastOneByteArrayJoin(elements, '');
+      if (!IS_UNDEFINED(result)) return result;
       return %StringBuilderConcat(elements, elements_length, '');
     }
     // Non-empty separator case.
@@ -231,6 +233,9 @@ function Join(array, length, separator, convert) {
         elements[i] = e;
       }
     }
+    var result = %_FastOneByteArrayJoin(elements, separator);
+    if (!IS_UNDEFINED(result)) return result;
+
     return %StringBuilderJoin(elements, length, separator);
   } finally {
     // Make sure to remove the last element of the visited array no
@@ -442,6 +447,9 @@ function InnerArrayJoin(separator, array, length) {
     separator = TO_STRING(separator);
   }
 
+  var result = %_FastOneByteArrayJoin(array, separator);
+  if (!IS_UNDEFINED(result)) return result;
+
   // Fast case for one-element arrays.
   if (length === 1) {
     var e = array[0];

src/runtime/runtime-array.cc

@@ -489,6 +489,15 @@ RUNTIME_FUNCTION(Runtime_GetCachedArrayIndex) {
 }
 
 
+RUNTIME_FUNCTION(Runtime_FastOneByteArrayJoin) {
+  SealHandleScope shs(isolate);
+  DCHECK(args.length() == 2);
+  // Returning undefined means that this fast path fails and one has to resort
+  // to a slow path.
+  return isolate->heap()->undefined_value();
+}
+
+
 RUNTIME_FUNCTION(Runtime_ArraySpeciesConstructor) {
   HandleScope scope(isolate);
   DCHECK(args.length() == 1);

src/runtime/runtime.h

@@ -51,6 +51,7 @@ namespace internal {
   F(GetCachedArrayIndex, 1, 1)       \
   F(FixedArrayGet, 2, 1)             \
   F(FixedArraySet, 3, 1)             \
+  F(FastOneByteArrayJoin, 2, 1)      \
   F(ArraySpeciesConstructor, 1, 1)