Make gcd/lcm effect-free by using LazyString

[tkf]

Before:

julia> Base.infer_effects(gcd, Tuple{Int,Int})
(?c,?e,?n,?t)′

After:

julia> Base.infer_effects(gcd, Tuple{Int,Int})
(?c,+e,?n,?t)

When playing with auto-parallelization (#43910, https://github.com/JuliaFolds/ParallelMagics.jl), I wanted to use gcd in a demo but it didn't work. I think it's a relatively simple change and also in general it's nice to have. But it's a relatively low-stakes "fix" for me and so I'm OK to not do this if there are some potential problems.

[oscardssmith]

Does this increase the amount of codegen for errors? Should we start reporting the value for all our errors?

[tkf]

Does this increase the amount of codegen for errors?

I think the main difference is a construction of a string vs a construction of a tuple. I'd guess they are not very different and also negligible. But I don't know exactly about it and also how this part of code is used.

I forgot to post the OP initially but, as I just mentioned above, it's really just a nice-to-have change for me.

[tkf]

This invokelatest was added in #30566 for reducing the latency. But, since with LazyString we also avoid the compilation of print of each interpolated values, I'd guess it has a similar effect?

Though I wonder if we need @nospecialize in the constructor?

julia/base/strings/lazy.jl

Line 28 in 0deb326

LazyString(args...) = new(args)

[aviatesk]

Yes, LazyString should hide the inference cost for printing values at a time when we compile the parent code containing LazyString constructor (and the deferred printing will essentially be dynamically dispatched as like invokelatest).

We generally don't need @nospecialize annotation when a generic function has only a few # of methods and the method body is very simple. But it won't hurt anything (at least at this moment) if we add @nospecialize for this constructor anyway.

[aviatesk]

Do we want to use lazy"..." syntax for constructing LazyString, e.g. lazy"$x $op $y overflowed for type $(typeof(x))"?

[tkf]

I'm pretty sure I tried it but the macro didn't work at this phase during bootstrap.