Patch to allow more granular aborting.

[Torbjorn Granlund]

[This thread began at gmp-devel.]

Theo Schlossnagle <jesus at omniti.com> writes:

  On Tuesday, March 15, 2011, Torbjorn Granlund <tg at gmplib.org> wrote:
  > Theo Schlossnagle <jesus at omniti.com> writes:
  >
  >   I'm confused now.  In a previous mail, you suggested I use the builtin
  >   libgmp mechanism to replace the allocators.  It appears libgmp indeed
  >   has replaceable allocators, but then when it goes to allocate they are
  >   not used. I results in I me being unable to wrap realloc and throw and
  >   exception on NULL return.
  >
  > GMP only allocates (reallocates) things when memory (more memory) is
  > needed, as Niels pointed out.
  >
  > If you want allocation (reallocation) using the replaceable allocators
  > to take place also when memory is not needed, then you probably need to
  > maintain your custom GMP version.
  >
  > Your statement that "when [GMP] goes to allocate [the custom allocators]
  > are not used" is simply a misunderstanding of GMP source code.
  
  I call mpz_n_pow_ui. Libgmp performs an allocation directly using
  mpz_realloc instead of the reallocation function I set. Therefore I
  cannot change the way in which libgmp handles OOM. This is my problem.
  It appears that the mpz_n_pow_ui is not the only function that behaves
  this way. There is a problem.

If that is really the case, it is a bug.  Please read the page in the
manual about bug reporting, and then send a report to the gmp-bugs list.
  
  Handling these conditions from a signal handler is not portably safe.
  I proposed a solution to the problem. An explanation that I simply
  don't understand the source code is not helpful. Please instead
  propose a solution that allows me to use the stock version of libgmp.
  This problem is not unique to me. It is a problem for anyone
  attempting to use libgmp in a long running process that takes user
  input. (lua embed, for example)
  
I agree that error handling in GMP is not good.  This issue has been
discussed several times on this list.  Unfortunately, the C language
does not provide any good solutions to the problem.  People have pointed
out why your patch does not help, I will not repeat it.

-- 
Torbjörn