After some promising results with a proof-of-concept implementation, I'm optimistic we can rewrite integer-gmp to use only non-allocating GMP lib functions without suffering from serious regressions.

If successful, this would

• allow to avoid the custom GMP allocator hack, and thus
• avoid issues when linking against other C libraries using GMP,
• simplify code, as we would perform all heap allocations in Haskell code (and never inside Cmm/C code as its done now),
• and finally maybe even remove a few more superfluous temporary heap allocations.

see also wiki:Design/IntegerGmp2