After the move to PCRE2, optimizing patterns has been a thorn in the
side due to the fact that PCRE2's JIT compiler modifies the pattern
object itself (instead of returning a new set of data, like PCRE1
did). To make this fit with the existing behavior, a read/write
lock was introduced, with the read part locking when matching and
the write when compiling (or JIT-compiling) the pattern.
This locking strategy however introduced a performance issue,
as we needed:
* to acquire a write lock to compile/optimize the pattern (incl. the
common case where the pattern was already compiled, so bailing out
immediately);
* to acquire a read lock during the actual match, to prevent
some other thread from optimizing the pattern under our nose.
This was due to the "lazy" optimization policy of QRegularExpression
-- optimize a pattern after a certain number of usages. The
excessive amount of locking effectively limited scalability.
Simplify the code, and drop that policy altogether: since JIT
compiling in PCRE2 is faster and pretty much "always recommended",
just always do it for any pattern (unless it gets disabled via
env variables) when compiling it.
This allows to go back to a plain QMutex, and now the actual
matching doesn't require acquiring any locks any longer. Of course,
there is still a mutex acquired just before matching for checking
whether the pattern needs recompiling in the first place; this can
probably be further optimized via double-checked locking (using
atomics), but not doing it right now.
This shift makes a couple of pattern options controlling
optimization useless, and allows to centralize the 3
QRegularExpression tests (which were actually the very same test,
just setting slightly different optimizations strategies).
While at it, install a stress-test for threading, with the idea
of running it under TSAN or helgrind to catch bugs in
QRegularExpression's locking.
[ChangeLog][Important Behavior Changes][QRegularExpression] Regular
expressions are now automatically optimized (including JIT
compiling) on their first usage. The pattern options
OptimizeOnFirstUsageOption and DontAutomaticallyOptimizeOption no
longer have any effect, and will get removed in a future version of
Qt. QRegularExpression::optimize() can be still used to compile and
optimize the regular expression in advance (before any match), if
needed.
Task-number: QTBUG-66781
Change-Id: Ia0e97208ae78255fe811b78029ed01c204e47bd2
Reviewed-by: David Faure <david.faure@kdab.com>
06af9a1e38