RegExp: Add support for table-based character class

src/x64/regexp-macro-assembler-x64.cc

Index: src/x64/regexp-macro-assembler-x64.cc
===================================================================
--- src/x64/regexp-macro-assembler-x64.cc	(revision 11134)
+++ src/x64/regexp-macro-assembler-x64.cc	(working copy)
@@ -572,6 +572,42 @@
 }
 
 
+void RegExpMacroAssemblerX64::CheckCharacterInRange(
+    uc16 from,
+    uc16 to,
+    Label* on_in_range) {

[Lasse Reichstein Nielsen, 2012/03/28 14:02:10]

Maybe special-case if 'from' is zero (or even if 'to' is 0xffff)

[Erik Corry, 2012/03/29 14:49:00]

Never happens.

+  __ lea(rax, Operand(current_character(), -from));

[Lasse Reichstein Nielsen, 2012/03/28 14:02:10]

lea->leal (saves a byte).

[Erik Corry, 2012/03/29 14:49:00]

I think you mean on x64.
Fixed.

+  __ cmpl(rax, Immediate(to - from));
+  BranchOrBacktrack(below_equal, on_in_range);
+}
+
+
+void RegExpMacroAssemblerX64::CheckCharacterNotInRange(
+    uc16 from,
+    uc16 to,
+    Label* on_not_in_range) {
+  __ lea(rax, Operand(current_character(), -from));
+  __ cmpl(rax, Immediate(to - from));
+  BranchOrBacktrack(above, on_not_in_range);
+}
+
+
+void RegExpMacroAssemblerX64::CheckBitInTable(
+    Handle<ByteArray> table,
+    Label* on_bit_set) {
+  __ Move(rax, table);
+  Register index = current_character();
+  if (mode_ != ASCII || kTableMask != String::kMaxAsciiCharCode) {
+    __ movq(rbx, current_character());
+    __ and_(rbx, Immediate(kTableMask));

[Lasse Reichstein Nielsen, 2012/03/28 14:02:10]

I'd consider swapping these:
  movq(rbx, Immediate(kTableMask));
  and_(rbx, current_character());
(speculating that the constant load could hoist better).

[Erik Corry, 2012/03/29 14:49:00]

OTOH a mov can be eliminated completely by the register renamer.

[Lasse Reichstein, 2012/03/29 15:25:32]

I don't think the register renamer will do that since the value of
current_character() needs to be retained as well. A move will use a different
physical register.

[Erik Corry, 2012/03/30 07:46:28]

My understanding is that after renaming you have gone from 2-register
instructions to 3-register instructions so the mov disappears.

[Erik Corry, 2012/03/31 20:04:04]

But I tested it and your version is marginally faster.

+    index = rbx;
+  }
+  __ cmpb(FieldOperand(rax, index, times_1, ByteArray::kHeaderSize),
+          Immediate(0));
+  BranchOrBacktrack(not_equal, on_bit_set);
+}
+
+
 bool RegExpMacroAssemblerX64::CheckSpecialCharacterClass(uc16 type,
                                                          Label* on_no_match) {
   // Range checks (c in min..max) are generally implemented by an unsigned