Bug 16582 – Move optimization (like firstByte) branches outside jsRegExpExecute match loop

• Status: RESOLVED
• Resolution: INVALID
• Component: JavaScriptCore

[Eric Seidel (no email)]

Move optimization (like firstByte) branches outside jsRegExpExecute match loop

There is no reason for us to test the firstByte optimization each time through the inner jsRegExpExecute loop.  firstByte is determined outside of that loop.  We should just have different static inline functions (or a template function) for the various different loop cases, and call the right function to handle an optimized inner loop.

See also bug 16581 for other issues with this function.

[Eric Seidel (no email)]

We're spending about 0.1% of total time (across all tests) on sunspider in these optimization branches.

[Eric Seidel (no email)]

I tried hard to avoid solving bug 16581 at the same time... but I think I mostly just made things messier by avoiding it.  Ick.  I guess the one crucial piece which stands in my way before solving bug 16581 is understanding how the offset vector "scratch space" (the last 1/3rd of the vector) is used during execution.  It was not clear to me last time I looked at the code.

[Eric Seidel (no email)]

Created attachment 18084 [details]
More cleanup to jsRegExpExecute

 JavaScriptCore/pcre/pcre_exec.cpp |  262 ++++++++++++++++++++-----------------
 1 files changed, 140 insertions(+), 122 deletions(-)

[Eric Seidel (no email)]

Created attachment 18085 [details]
Compiles and runs, 4 regressions

 JavaScriptCore/pcre/pcre_exec.cpp |  432 ++++++++++++++++++++++---------------
 1 files changed, 259 insertions(+), 173 deletions(-)

[Eric Seidel (no email)]

Created attachment 18086 [details]
Continue moving more code into class structures

 JavaScriptCore/pcre/pcre_exec.cpp |   33 ++++++++++++++++++---------------
 1 files changed, 18 insertions(+), 15 deletions(-)

[Eric Seidel (no email)]

So this still has 4 regressions.  I'm not sure why.  The code is better, and probably faster... but in an ugly middle state where it's not quite using classes like it should.  I'm a bit worn from hacking on this patch for so long, so I will plan to come back to it later this week and fix the regressions.

[Alexey Proskuryakov]

Seems obsolete, one can Shark again to find what to optimize in JavaScriptCore.