Speed up quoting of JSON strings by allocating a string that is big enough...

Speed up quoting of JSON strings by allocating a string that is big enough
and then trimming it when the length is known.  This way we only have to
traverse the input once.

Committed: http://code.google.com/p/v8/source/detail?r=5951

[Vyacheslav Egorov (Chromium), 2010-12-08 10:51:31]

LGTM

http://codereview.chromium.org/5556012/diff/1/src/runtime.cc
File src/runtime.cc (right):

http://codereview.chromium.org/5556012/diff/1/src/runtime.cc#newcode4596
src/runtime.cc:4596: static const byte JsonQuoteLengths[kQuoteTableLength] = {
I do not like subtle connection between string above and this array.

It's excruciatingly difficult to check correctness (4 in a row vs. 8 in a row).

Can we change it somehow?

http://codereview.chromium.org/5556012/diff/1/src/runtime.cc#newcode4710
src/runtime.cc:4710: write_cursor[0] = replacement[0];
Some replacements have len == 1.

Why we optimistically write two characters for them and them move write cursor
by 1? This is really confusing.

[Erik Corry, 2010-12-08 16:35:56]

http://codereview.chromium.org/5556012/diff/1/src/runtime.cc
File src/runtime.cc (right):

http://codereview.chromium.org/5556012/diff/1/src/runtime.cc#newcode4596
src/runtime.cc:4596: static const byte JsonQuoteLengths[kQuoteTableLength] = {
On 2010/12/08 10:51:31, Vyacheslav Egorov wrote:
> I do not like subtle connection between string above and this array.
> 
> It's excruciatingly difficult to check correctness (4 in a row vs. 8 in a
row).
> 
> Can we change it somehow?
> 
Changed the string above so it is also 8 per line, to make visual checking
easier.

http://codereview.chromium.org/5556012/diff/1/src/runtime.cc#newcode4710
src/runtime.cc:4710: write_cursor[0] = replacement[0];
On 2010/12/08 10:51:31, Vyacheslav Egorov wrote:
> Some replacements have len == 1.
> 
> Why we optimistically write two characters for them and them move write cursor
> by 1? This is really confusing.

It is in order to remove mispredicted branches in this very hot inner loop. 
However measurements show that the extra writes almost cancel out the benefit of
not having mispredicts so I will remove this trick it for clarity.