From @andk

I think this bugreport deserves to be a quiz. Test your intuition.

Sample string to test the regexps against​:

  sprintf "%scould%snot%sopen%s", ("x"x10000)x4;

Regexp 1​:

  qr!.*(?​:could ?not (?​:open|connect|find))!

Regexp 2 (the difference is it's case insensitive)​:

  qr!.*(?i​:could ?not (?​:open|connect|find))!

Note​: neither matches and this is correct.

Question​: which is faster and by how much?

Answer​: regexp 1 is much more than 100 times slower than the regexp 2.

On my machine this program takes 1-2 wallclock seconds​:

  time $p -le 'my $x = sprintf "%scould%snot%sopen%s", ("x"x10000)x4;print $x =~ m!.*(?​:could ?not (?​:open|connect|find))! ? "not ok" : "ok";'

And on all those perls it takes less than 0.01 wallclock seconds when I
add the "i".

Historical evidence​: this is not a regression, at least not since 5.6. I
have witnessed the same slowness on all my perls between 5.6.1 and
bleadperl.

Competition evidence​: with 'use re​::engine​::PCRE;' both regexps are
fast.

Enjoy,
--
andreas

From @iabyn

On Sat, Jul 19, 2014 at 02​:21​:03PM -0700, Andreas J. Koenig via RT wrote​:

sprintf "%scould%snot%sopen%s", ("x"x10000)x4;

Regexp 1​:

qr!.*(?​:could ?not (?​:open|connect|find))!

Regexp 2 (the difference is it's case insensitive)​:

qr!.*(?i​:could ?not (?​:open|connect|find))!

The second is faster because it disables intuiting, which causes it to
take a different path through regexec_flags(), which causes it to fail
after the first call to regtry(), rather than re-trying regtry at
positions 1,2,3,... and eventually failing.

I don't know know what the fix is, but this is tied up with how intuit()
and the parent regex_flags() handle all the various flags (like ANCH_MBOL,
and whether that MBOL is a fake one due the presence of a leading '.*'
etc).

Its all a bit of a mess, and I intend to look at this properly when I get
round to doing my second batch of intuit()/regexec_flags() refactoring
sometime this year.

--
Modern art​:
  "That's easy, I could have done that!"
  "Ah, but you didn't!"

The RT System itself - Status changed from 'new' to 'open'

I can reproduce that on 5.18.4 but I can't on 5.30.2 so it definitely was fixed at some point.