It looks like you did not touch the code paths handling rational infinity, like big(1//0). It does not seem to affect the usual +, -, * and / operations but the inplace versions are affected: for instance

julia> a = big(1//0);

julia> b = big(0//1);

julia> Base.GMP.MPQ.sub!(b, a);

julia> b
0//1

when I'd expect it to be -1//0.
Those inplace functions are not exported nor documented and it has no impact on usual operations so I don't know if it's a blocker, but it feels a bit dangerous.

@Liozou thanks for your comment! I fixed inplace for infinities, and added some tests

Just a very simple benchmark of how much speed up is possible using preallocated Rational for result and inplace operations, (note that these are very noisy)

for x, y, z of Rational{BigInt}

for x, y of Int

using BenchmarkTools

z, x, y = Rational{BigInt}(0, 1), Rational{BigInt}(123456789, 3456789), Rational{BigInt}(2323223, 11991398)

@btime +($x, $y)

@btime *($x, $y)

@btime Base.GMP.MPQ.add!($z, $x, $y)

@btime Base.GMP.MPQ.mul!($z, $x, $y)

@info "###########"

x, y = 12345, 3423437

@btime Base.unsafe_rational(BigInt, $x, $y)

@btime Base.GMP.MPQ.set_si!($z, $x, $y)

x = BigInt(2123434)

@btime Base.GMP.MPQ.set_z!($z, $x)

@Liozou thank you a lot for your comments !

Do you think the need to store rat::Rational{BigInt} in _MPQ can be relaxed ? If so, we can probably do

@eval $op(a::$T) = $op!(unsafe_rational(BigInt(), BigInt()), a)

more efficient by using set_si!(_MPQ(), 0, 1) directly instead of unsafe_rational

In the definition of _MPQ, the need for the rat field is indicated to be related to GC, which means (I assume) that otherwise the GC does not know that the pointers in the _MPQ struct refer to the BigInts in rat, so it may corrupt those structs. I'm no GC guru so maybe there is a way to avoid that, but I don't see it...

I think one way to completely bypass the issue consists in not using Rational{BigInt} but directly the gmpq object: for this, I direct you towards my BigRationals.jl library, which is a very basic wrapper around GMPQ. But that's orthogonal to this PR.

There is a more in-depth discussion about this in #9826