Reviewed by John.
- fix <rdar://problem/4687840> 9A241: JavaScript RegExp 25-30x slower than on 10.4.7
I used Shark to figure out what to do. The test case is now 15% faster than with
stock Safari. Some other regular expression cases might still be a few % slower
than before, but the >10x slowdown is now completely gone.
1) Fix slowness caused by setjmp/longjmp by using computed goto instead.
Use GCC extensions - locally declared labels, labels as values, and computed goto -
instead of using setjmp/longjmp to implemement non-recursive version of the regular
expression system. We could probably make this even faster if we reduced the use
of malloc a bit too.
2) Fix slowness caused by allocating heapframe objects by allocating the first
16 of them from the stack.
3) Speed up use of malloc and free in PCRE by making it use fastMalloc and fastFree.
4) Speed up the test case by adding a special case to a UString function.
5) Made a small improvement to the innermost hottest loop of match by hoisting
the conversion from int to pcre_uchar out of the loop.
- JavaScriptCore.xcodeproj/project.pbxproj: Compile FastMallocPCRE.cpp, and don't
compile pcre_globals.c.
- wtf/FastMallocPCRE.cpp: Added. A copy of pcre_globals.c that uses FastMalloc.h.
This is better than code that sets the PCRE allocation globals because by doing it
this way there's guaranteed to be no problem with order of initialization.
- kjs/ustring.cpp: (KJS::UString::spliceSubstringsWithSeparators): Add a fast
special case when this is called for only one subrange and no seaprators. This
was happening a lot in the test case and it seems quite reasonable to optimize this.
- pcre/pcre_exec.c: Create a copy of the RMATCH and RRETURN macros that use goto
instead of setjmp/longjmp. Change code that calls pcre_stack_malloc to first use
storage on the stack inside the match function.
(match): Move initialization of utf8 up a couple lines to avoid "possibly used
uninitialized" warning. Use a local variable so we compare with pcre_uchar instead
of with int inside the inner "find a character" loop.
