RegExp: Remove use of 16-bit immediates on ia32/x64.

RegExp: Remove use of 16-bit immediates on ia32/x64. 
Also check more than one character at a time.
16-bit immediates requires a prefix that changes the length of the instruction. 
This causes predecoder mispredictions and subsequent pipeline stalls.
Also removed redundant "atStart" local variable which is equivalent to startIndex == 0.

[Lasse Reichstein, 2010-05-10 12:29:29]

RegExp optimization.
The one-character first test, before starting to test multiple characters at a
time, is still speculative. 
Measurements are on the way.

[Erik Corry, 2010-05-10 20:23:47]

There's a change in here to handle start-of-string differently.  That should be
mentioned in the commit message, and it would be nice to know its relative
impact in the benchmarks.

LGTM;

http://codereview.chromium.org/1988009/diff/1/2
File src/ia32/regexp-macro-assembler-ia32.cc (right):

http://codereview.chromium.org/1988009/diff/1/2#newcode250
src/ia32/regexp-macro-assembler-ia32.cc:250: for (int i = 0, n = str.length(); i
< n;) {
i = 1 seems better and the x64 version has it.

http://codereview.chromium.org/1988009/diff/1/2#newcode274
src/ia32/regexp-macro-assembler-ia32.cc:274: // flushes.
Is it possible to quote the section of the optimization guide this is based on? 
Intel or AMD?

[Lasse Reichstein, 2010-05-11 07:28:47]

The removal of atStart doesn't change performance noticably. The test code is
still the same, we only save ~five instructions of initialization.

http://codereview.chromium.org/1988009/diff/1/2
File src/ia32/regexp-macro-assembler-ia32.cc (right):

http://codereview.chromium.org/1988009/diff/1/2#newcode250
src/ia32/regexp-macro-assembler-ia32.cc:250: for (int i = 0, n = str.length(); i
< n;) {
Bugger. Good catch!

http://codereview.chromium.org/1988009/diff/1/2#newcode274
src/ia32/regexp-macro-assembler-ia32.cc:274: // flushes.
Quite possible, and done.
(Intel, Section 3.4.2.3, "Length-Changing Prefixes (LCP)").