Could you explain briefly what were the issues before this PR?

@bew GMP arithmetics highly depends on LibGMP::Long and LibGMP::ULong types to provide fast paths with integers of supported types.

LibGMP built with 32bit Long/ULong on 32bit machines, otherwise 64bit (for now).

So we must take in care it's sizeof to provide correct behavior.

See also 853edff and #4653 (it's a long story)

@akzhan Is there a spec that was failing in 32 bits? I ask because these changes, without accompanying specs, seem like they aren't doing anything.

@asterite Just tested BigInt.+ on current master (#4931):

Failures:
  1) BigInt adds
     Failure/Error: (5.to_big_i + Int64::MAX).should be > Int64::MAX.to_big_i
       Expected 4294967300_big_i to be > 9223372036854775807
     # spec/std/big/big_int_spec.cr:84
Finished in 11.78 seconds
5281 examples, 1 failures, 0 errors, 4 pending
Failed examples:
crystal spec spec/std/big/big_int_spec.cr:79 # BigInt adds

Will add this spec to this branch.

It seems like you're changing + - * / so there should be specs for all of those that beak on 32bit.

@akzhan Awesome! Thank you :-)

LGTM 👍

Please remove the changelog until we decide #4936.

Done ( and squashed into two logical parts).