Avoid core dump on some paren'd regexp matches

Avoid core dump on some paren'd regexp matches

In article <199706260526.XAA01060@lunkwill.ml.org> Jason wrote:
:This script causes Perl to dump core with a segmentation fault under
:Linux as well as HP-UX.  Here's the script:
:
:#!/usr/local/bin/perl
:@justalist = ("foo\nbar" =~ /(\s|(foo)|(bar))*/ );

It does the same under 5.004_01. The reason is that on the second match
it tried to match (foo) and succeeded, leaving startp[2] and endp[2]
pointing to the beginning and end of the matched 'foo'. On the third
match, it tried to match (foo) and failed; in doing so, it overwrote
startp[2] with the startpoint it was trying to match ('bar'), but left
endp[2] unaltered.

If that third match had failed, no problem would occur - it would
restore startp[] and endp[] from saved copies. However, because the
third match then succeeded on the final alternate the modified
startp[] and endp[] were retained, leaving a mismatched pair of values
for $2.

The solution depends on what the answer should be - one interpretation
is that, since (foo) failed to match the last time it was tried, the
results should be ('bar', undef, 'bar'). The first patch below effects
this. Alternatively, you could say that it was more correct and/or
more useful for it to return the last successful match on (foo), in
which case you want the rather more complicated second patch below.

I'm not an expert on this stuff - Ilya, can you take a look at these
patches and tell me how broken they are, please?

My own feeling is that the second interpretation is more useful, but
I have much less confidence in the completeness of my patch for this.
No test cases supplied at this stage: Jason's testcase above should
suffice for the moment. Perl passes all tests here with either patch.

p5p-msgid: 199706261236.NAA03472@crypt.compulink.co.uk