msieve with GMP 6.2.0

[ryanp]

Has anyone successfully compiled and used msieve with the latest GMP release, 6.2.0?

I'm seeing segfaults during relation filtering. Some gdb debugging tracked the problem to the "if" block of this method in include/gmp_xface.h:

static INLINE void uint64_2gmp(uint64 src, mpz_t dest) {

#if GMP_LIMB_BITS == 64
        dest->_mp_d[0] = src;
        dest->_mp_size = (src ? 1 : 0);
#else
        /* mpz_import is terribly slow */
        mpz_set_ui(dest, (uint32)(src >> 32));
        mpz_mul_2exp(dest, dest, 32);
        mpz_add_ui(dest, dest, (uint32)src);
#endif
}

Commenting out the #if part (just using the mpz_* calls) seems to get things working, but I don't know if this fix is correct, or optimal.

[bsquared]

Originally Posted by ryanp

Has anyone successfully compiled and used msieve with the latest GMP release, 6.2.0?

I'm seeing segfaults during relation filtering. Some gdb debugging tracked the problem to the "if" block of this method in include/gmp_xface.h:

static INLINE void uint64_2gmp(uint64 src, mpz_t dest) {

#if GMP_LIMB_BITS == 64
        dest->_mp_d[0] = src;
        dest->_mp_size = (src ? 1 : 0);
#else
        /* mpz_import is terribly slow */
        mpz_set_ui(dest, (uint32)(src >> 32));
        mpz_mul_2exp(dest, dest, 32);
        mpz_add_ui(dest, dest, (uint32)src);
#endif
}

Commenting out the #if part (just using the mpz_* calls) seems to get things working, but I don't know if this fix is correct, or optimal.

Yeah, I ran across this and came up with this fix:
https://www.mersenneforum.org/showpo...7&postcount=81

I suspect the problem comes about because of this from the gmp changelog:

Internal representation of the mpz_t variables now supports lazy allocation; memory is allocated only when a value is stored.

I take that to mean that the _mp_d internals of an mpz_t now may not be actually allocated if accessed directly.

[ryanp]

Originally Posted by bsquared

Yeah, I ran across this and came up with this fix:
https://www.mersenneforum.org/showpo...7&postcount=81

I suspect the problem comes about because of this from the gmp changelog:


I take that to mean that the _mp_d internals of an mpz_t now may not be actually allocated if accessed directly.

Thanks!