Generated code for substring slices in x64 and arm.

src/arm/code-stubs-arm.cc

Index: src/arm/code-stubs-arm.cc
diff --git a/src/arm/code-stubs-arm.cc b/src/arm/code-stubs-arm.cc
index 09d2c170c2e85b8bc7f3f0e58fcda7e1be6454b1..6561e696c469d5485a028c020d9f28ae525ae13c 100644
--- a/src/arm/code-stubs-arm.cc
+++ b/src/arm/code-stubs-arm.cc
@@ -5468,11 +5468,6 @@ void SubStringStub::Generate(MacroAssembler* masm) {
   Register to = r6;
   Register from = r7;
 
-  if (FLAG_string_slices) {
-    __ nop(0);  // Jumping as first instruction would crash the code generation.
-    __ jmp(&runtime);
-  }
-
   __ Ldrd(to, from, MemOperand(sp, kToOffset));
   STATIC_ASSERT(kFromOffset == kToOffset + 4);
   STATIC_ASSERT(kSmiTag == 0);
@@ -5500,16 +5495,28 @@ void SubStringStub::Generate(MacroAssembler* masm) {
   // r7 (a.k.a. from): from offset (smi)
 
   // Make sure first argument is a sequential (or flat) string.
-  __ ldr(r5, MemOperand(sp, kStringOffset));
+  __ ldr(r0, MemOperand(sp, kStringOffset));
   STATIC_ASSERT(kSmiTag == 0);
-  __ JumpIfSmi(r5, &runtime);
-  Condition is_string = masm->IsObjectStringType(r5, r1);
+  __ JumpIfSmi(r0, &runtime);
+  Condition is_string = masm->IsObjectStringType(r0, r1);
   __ b(NegateCondition(is_string), &runtime);
 
+  // Short-cut for the case of trivial substring.
+  Label return_r0;
+  __ ldr(r4, FieldMemOperand(r0, String::kLengthOffset));
+  __ cmp(r2, Operand(r4, ASR, 1));
+  __ b(ge, &return_r0);

[Erik Corry, 2011/08/31 14:14:13]

I think eq is clearer than ge.  A comment to the effect that r2 is the length
would be nice.

+
+  Label create_slice;
+  if (FLAG_string_slices) {
+    __ cmp(r2, Operand(SlicedString::kMinLength));
+    __ b(ge, &create_slice);
+  }
+
+  // r0: original string
   // r1: instance type
   // r2: length
   // r3: from index (untagged smi)
-  // r5: string
   // r6 (a.k.a. to): to (smi)
   // r7 (a.k.a. from): from offset (smi)
   Label seq_string;
@@ -5523,8 +5530,8 @@ void SubStringStub::Generate(MacroAssembler* masm) {
   // Cons string. Try to recurse (once) on the first substring.
   // (This adds a little more generality than necessary to handle flattened
   // cons strings, but not much).
-  __ ldr(r5, FieldMemOperand(r5, ConsString::kFirstOffset));
-  __ ldr(r4, FieldMemOperand(r5, HeapObject::kMapOffset));
+  __ ldr(r0, FieldMemOperand(r0, ConsString::kFirstOffset));
+  __ ldr(r4, FieldMemOperand(r0, HeapObject::kMapOffset));
   __ ldrb(r1, FieldMemOperand(r4, Map::kInstanceTypeOffset));
   __ tst(r1, Operand(kStringRepresentationMask));
   STATIC_ASSERT(kSeqStringTag == 0);
@@ -5533,21 +5540,21 @@ void SubStringStub::Generate(MacroAssembler* masm) {
   // Definitly a sequential string.
   __ bind(&seq_string);
 
-  // r1: instance type.
+  // r0: original string
+  // r1: instance type
   // r2: length
   // r3: from index (untaged smi)

[Erik Corry, 2011/08/31 14:14:13]

untaged -> untagged
also at least one other place.

-  // r5: string
   // r6 (a.k.a. to): to (smi)
   // r7 (a.k.a. from): from offset (smi)
-  __ ldr(r4, FieldMemOperand(r5, String::kLengthOffset));
+  __ ldr(r4, FieldMemOperand(r0, String::kLengthOffset));
   __ cmp(r4, Operand(to));
   __ b(lt, &runtime);  // Fail if to > length.
   to = no_reg;
 
-  // r1: instance type.
-  // r2: result string length.
+  // r0: original string

[Erik Corry, 2011/08/31 14:14:13]

Original string or left hand side of the original cons string.

+  // r1: instance type
+  // r2: result string length
   // r3: from index (untaged smi)
-  // r5: string.
   // r7 (a.k.a. from): from offset (smi)
   // Check for flat ASCII string.
   Label non_ascii_flat;
@@ -5561,73 +5568,70 @@ void SubStringStub::Generate(MacroAssembler* masm) {
 
   // Sub string of length 2 requested.
   // Get the two characters forming the sub string.
-  __ add(r5, r5, Operand(r3));
-  __ ldrb(r3, FieldMemOperand(r5, SeqAsciiString::kHeaderSize));
-  __ ldrb(r4, FieldMemOperand(r5, SeqAsciiString::kHeaderSize + 1));
+  __ add(r0, r0, Operand(r3));
+  __ ldrb(r3, FieldMemOperand(r0, SeqAsciiString::kHeaderSize));
+  __ ldrb(r4, FieldMemOperand(r0, SeqAsciiString::kHeaderSize + 1));
 
   // Try to lookup two character string in symbol table.
   Label make_two_character_string;
   StringHelper::GenerateTwoCharacterSymbolTableProbe(
       masm, r3, r4, r1, r5, r6, r7, r9, &make_two_character_string);
   Counters* counters = masm->isolate()->counters();
-  __ IncrementCounter(counters->sub_string_native(), 1, r3, r4);
-  __ add(sp, sp, Operand(3 * kPointerSize));
-  __ Ret();
+  __ jmp(&return_r0);
 
   // r2: result string length.
   // r3: two characters combined into halfword in little endian byte order.
   __ bind(&make_two_character_string);
   __ AllocateAsciiString(r0, r2, r4, r5, r9, &runtime);
   __ strh(r3, FieldMemOperand(r0, SeqAsciiString::kHeaderSize));
-  __ IncrementCounter(counters->sub_string_native(), 1, r3, r4);
-  __ add(sp, sp, Operand(3 * kPointerSize));
-  __ Ret();
+  __ jmp(&return_r0);
 
   __ bind(&result_longer_than_two);
 
+  // Locate 'from' character of string.
+  __ add(r5, r0, Operand(SeqAsciiString::kHeaderSize - kHeapObjectTag));
+  __ add(r5, r5, Operand(from, ASR, 1));
+
   // Allocate the result.
   __ AllocateAsciiString(r0, r2, r3, r4, r1, &runtime);
 
-  // r0: result string.
-  // r2: result string length.
-  // r5: string.
+  // r0: result string
+  // r2: result string length
+  // r5: first character of sub string to copy

[Erik Corry, 2011/08/31 14:14:13]

sub string -> substring

   // r7 (a.k.a. from): from offset (smi)
   // Locate first character of result.
   __ add(r1, r0, Operand(SeqAsciiString::kHeaderSize - kHeapObjectTag));
-  // Locate 'from' character of string.
-  __ add(r5, r5, Operand(SeqAsciiString::kHeaderSize - kHeapObjectTag));
-  __ add(r5, r5, Operand(from, ASR, 1));
 
-  // r0: result string.
-  // r1: first character of result string.
-  // r2: result string length.
-  // r5: first character of sub string to copy.
+  // r0: result string
+  // r1: first character of result string
+  // r2: result string length
+  // r5: first character of sub string to copy
   STATIC_ASSERT((SeqAsciiString::kHeaderSize & kObjectAlignmentMask) == 0);
   StringHelper::GenerateCopyCharactersLong(masm, r1, r5, r2, r3, r4, r6, r7, r9,
                                            COPY_ASCII | DEST_ALWAYS_ALIGNED);
-  __ IncrementCounter(counters->sub_string_native(), 1, r3, r4);
-  __ add(sp, sp, Operand(3 * kPointerSize));
-  __ Ret();
+  __ jmp(&return_r0);
 
   __ bind(&non_ascii_flat);
-  // r2: result string length.
-  // r5: string.
+  // r0: original string
+  // r2: result string length
   // r7 (a.k.a. from): from offset (smi)
   // Check for flat two byte string.
 
+  // Locate 'from' character of string.
+  __ add(r5, r0, Operand(SeqTwoByteString::kHeaderSize - kHeapObjectTag));
+  // As "from" is a smi it is 2 times the value which matches the size of a two
+  // byte character.

[Erik Corry, 2011/08/31 14:14:13]

Add a static assert that the kSmiTag is 0 and kSmiTagSize is 1.

+  __ add(r5, r5, Operand(from));
+
   // Allocate the result.
   __ AllocateTwoByteString(r0, r2, r1, r3, r4, &runtime);
 
-  // r0: result string.
-  // r2: result string length.
-  // r5: string.
+  // r0: result string
+  // r2: result string length
+  // r5: first character of sub string to copy

[Erik Corry, 2011/08/31 14:14:13]

sub string -> substring

   // Locate first character of result.
   __ add(r1, r0, Operand(SeqTwoByteString::kHeaderSize - kHeapObjectTag));
-  // Locate 'from' character of string.
-  __ add(r5, r5, Operand(SeqTwoByteString::kHeaderSize - kHeapObjectTag));
-  // As "from" is a smi it is 2 times the value which matches the size of a two
-  // byte character.
-  __ add(r5, r5, Operand(from));
+
   from = no_reg;
 
   // r0: result string.
@@ -5637,6 +5641,71 @@ void SubStringStub::Generate(MacroAssembler* masm) {
   STATIC_ASSERT((SeqTwoByteString::kHeaderSize & kObjectAlignmentMask) == 0);
   StringHelper::GenerateCopyCharactersLong(
       masm, r1, r5, r2, r3, r4, r6, r7, r9, DEST_ALWAYS_ALIGNED);
+  __ jmp(&return_r0);
+
+  if (FLAG_string_slices) {
+    __ bind(&create_slice);
+    // r0: original string
+    // r1: instance type
+    // r2: length
+    // r3: from index (untagged smi)
+    // r6 (a.k.a. to): to (smi)
+    // r7 (a.k.a. from): from offset (smi)
+    Label allocate_slice, sliced_string, seq_string;
+    STATIC_ASSERT(kSeqStringTag == 0);
+    __ tst(r1, Operand(kStringRepresentationMask));
+    __ b(eq, &seq_string);
+    STATIC_ASSERT(kIsIndirectStringMask == (kSlicedStringTag & kConsStringTag));

[antonm, 2011/08/31 13:08:42]

those constants might be worth lifting into objects.h

+    STATIC_ASSERT(kIsIndirectStringMask != 0);
+    __ tst(r1, Operand(kIsIndirectStringMask));
+    // External string.  Jump to runtime.
+    __ b(eq, &runtime);
+
+    __ tst(r1, Operand(kSlicedNotConsMask));
+    __ b(ne, &sliced_string);
+    // Cons string.  Check whether it is flat, then fetch first part.

[Erik Corry, 2011/08/31 14:14:13]

How do we get here if it's a cons string?

[Yang, 2011/08/31 15:58:16]

A cons string's string representation bits has the kIsIndirectStringMask set
(both cons and slices are "indirect strings"), therefore it is not bailed out to
runtime, and further checked whether it has the kSlicedNotConsMask bit set.
Since that bit is not set, we do not jump to sliced_string, but proceed with the
part that handles cons strings.

+    __ ldr(r5, FieldMemOperand(r0, ConsString::kSecondOffset));
+    __ LoadRoot(r9, Heap::kEmptyStringRootIndex);
+    __ cmp(r5, r9);
+    __ b(ne, &runtime);
+    __ ldr(r5, FieldMemOperand(r0, ConsString::kFirstOffset));
+    __ jmp(&allocate_slice);
+
+    __ bind(&sliced_string);
+    // Sliced string.  Fetch parent and correct start index by offset.

[Erik Corry, 2011/08/31 14:14:13]

Do you have test coverage here?  I think we already bailed to runtime in this
case.

[Yang, 2011/08/31 15:58:16]

We only bail out in the case when the original string is a slice if the
substring length is shorter than SlicedString::kMinLength. In that case we do
not create a slice, but create a seq string, to which we copy the content.

If the substring length is longer or equal SlicedString::kMinLength and
FLAG_string_slices==true, then the code jumps to the label create_slice. 

I added a cctest for slice of slice to check the string shapes.

+    __ ldr(r5, FieldMemOperand(r0, SlicedString::kOffsetOffset));
+    __ add(r7, r7, r5);
+    __ ldr(r5, FieldMemOperand(r0, SlicedString::kParentOffset));
+    __ jmp(&allocate_slice);
+
+    __ bind(&seq_string);
+    // Sequential string.  Just move string to the right register.
+    __ mov(r5, r0);
+
+    __ bind(&allocate_slice);
+    // r1: instance type of original string
+    // r2: length
+    // r5: underlying subject string
+    // r7 (a.k.a. from): from offset (smi)
+    // Allocate new sliced string.  At this point we do not reload the instance
+    // type including the string encoding because we simply rely on the info
+    // provided by the original string.  It does not matter if the original
+    // string's encoding is wrong because we always have to recheck encoding of
+    // the newly created string's parent anyways due to externalized strings.
+    Label two_byte_slice, set_slice_header;
+    STATIC_ASSERT(kAsciiStringTag != 0);
+    __ tst(r1, Operand(kAsciiStringTag));
+    __ b(eq, &two_byte_slice);
+    __ AllocateAsciiSlicedString(r0, r2, r3, r4, &runtime);
+    __ jmp(&set_slice_header);
+    __ bind(&two_byte_slice);
+    __ AllocateTwoByteSlicedString(r0, r2, r3, r4, &runtime);
+    __ bind(&set_slice_header);
+    __ str(r7, FieldMemOperand(r0, SlicedString::kOffsetOffset));
+    __ str(r5, FieldMemOperand(r0, SlicedString::kParentOffset));
+  }
+
+  __ bind(&return_r0);
   __ IncrementCounter(counters->sub_string_native(), 1, r3, r4);
   __ add(sp, sp, Operand(3 * kPointerSize));
   __ Ret();