gmp-4.1.2 assertion failure

[Torbjorn Granlund]

Kevin Ryde <user42 at zip.com.au> writes:

  > +  {
  > +    /* this block is equivalent to 
  > +       ralloc = (bsize * GMP_NUMB_BITS - cnt + GMP_NAIL_BITS) * e / GMP_NUMB_BITS + 5;
  > +       MPZ_REALLOC(r,ralloc+rtwos_limbs);
  > +       but without any overflow problems
  > +     */
  
  This is not too pretty, but I suppose there's no alternative.
  
If this is about the mpz_*pow* size overflow, the suggested 30 line
replacement is not OK.  I would like to look for alternatives.

We don't need to cap result size at 2^32 (for a 32-bit machine).
It is OK to cap it at 2^29 or less, since that will fill the
entire memory.

  I wouldn't mind seeing an ASSERT which does the calculation another
  way, like dividing ULONG_MAX by the bsize stuff to work out a limit
  for the exponent.

Division by a non-invariant takes 10 to 100 cycles on current machines.

-- 
Torbjörn