[Python-Dev] RE: Making the CVS tree pass the -Wall test...

[Tim Peters]

>> ./pypcre.c:4524: warning: variable `start_bits' might be
>> clobbered by `longjmp' or `vfork'
>> ./pypcre.c:4525: warning: variable `start_match' might be
>> clobbered by `longjmp' or `vfork'

[Tim sez these are almost always bogus wngs, Martin sez these two are
 legit]

The start_bits wng is bogus because start_bit's only definition points
precede the do-loop:  start_bits will either be NULL or extra->start_bits by
the time the "do" is entered, and can never change after that (yes,
start_bits "appears" in the loop, but not as an lvalue -- it's only
referenced).  So the value setjmp tucks away is necessarily the correct
value for longjmp to restore.  This is a purely local flow question;
knowledge of when a longjmp may occur isn't needed.

I agree the start_match wng is legit.  In my previous life, I tracked down
about a hundred of these, and found only one that was legit.  So pypcre is
as sick as Dragon's half-million lines of speech recognition code <wink>.

"The usual" way to fix this is to slop "volatile" onto the declarations, as
the author already did for three other locals.  That's the second reason
this warning sucks:  it depends on which locals gcc decides to carry in
registers, and that varies from release to release.  And that's the third
reason this warning sucks:  gcc is actually complaining about its *own*
decision to carry something in a register <0.1 wink -- and if it *knows*
it's not safe to carry something in a register, it shouldn't!>.