Re: GMP code committed -- watch for bugs.

[Mikael Djurfeldt]

Mikael Djurfeldt <address@hidden> writes:

> 2. My first attempt to build Guile has failed.  check_sanity aborts at
>    the long_long LLONG_MAX test.  Unfortunately, I can't look into
>    that right now.

OK.  There seems to be two bugs preventing the long_long tests from
succeeding:

1. num2integral.i.c:INTEGRAL2BIG had a bug in computing the absolute
   value of n when preparing arguments to mpz_import.  I've committed
   a fix for this.

2. The determination of required storage size for a bignum is buggy:
   
ITYPE
NUM2INTEGRAL (SCM num, unsigned long int pos, const char *s_caller)
{

     [...]

              numbits = mpz_sizeinbase (SCM_I_BIG_MPZ (num), 2);
              if (UNSIGNED) numbits++;
              scm_remember_upto_here_1 (num);
              if (numbits > (sizeof (ITYPE) * 8))
                scm_out_of_range (s_caller, num);

Firstly, you probably intended to write "if (!UNSIGNED)" (to make room
for the sign bit).  Secondly, that doesn't work either.  For example,
the absolute value of LLONG_MIN requires 64 bits, even though it is a
signed number.  So, regardless if we have "if (UNSIGNED)" or "if
(!UNSIGNED)" there will always be a case where a valid long_long or
ulong_long will cause an out_of_range error.

Unfortunately, I don't have time to get enough overview of the
situation to fix this bug.  I'll leave that to you.

Best regards,
Mikael