Make it easier to avoid StackOverflowErrors with promotion and convert

[timholy]

This guards against errors that can lead to stack overflows. It resolves a mutually-recursive interaction between T(x) and convert(T, x) for T==Union{}, and more significantly introduces a new function, promote_noncircular(x, y, z...), that throws an error if promotion didn't change at least one of the types. That prevents stack overflows for definitions like

foo{T}(x::T, y::T) = # "real" definition of foo
foo(x, y) = foo(promote_noncircular(x, y)...)

promote_noncircular is somewhat similar to (but more general than) promote_to_supertype. It is not quite as @pure, so the existing mechanism is also retained. At some point we may want to think about exporting this (or a variant), but for now this is a purely internal method.

Fixes #12007, fixes #10326, fixes #15736

[timholy]

Also worth noting that we still have good codegen:

julia> @code_llvm flipsign(1,Int8(1))

define i64 @julia_flipsign_62360(i64, i8) #0 !dbg !5 {
top:
  %2 = sext i8 %1 to i64
  %3 = ashr i64 %2, 63
  %4 = add i64 %3, %0
  %5 = xor i64 %4, %3
  ret i64 %5
}

Still, we should @nanosoldier runbenchmarks(ALL, vs = ":master").

[nanosoldier]

Your benchmark job has completed - possible performance regressions were detected. A full report can be found here. cc @jrevels

[timholy]

I can't reproduce any of those regressions, so I think this is good to go. Merge? Perhaps the name promote_noncircular could use some bikeshedding, particularly since I suspect packages will want to call it (exported or not). Certainly some of mine will 😄.

[pabloferz]

safe_promote?

[timholy]

That suggestion has merit, and I even thought of it ahead of time. I went with promote_noncircular as being a bit more specific about how it was being safe. But I could easily be persuaded to change, if most others prefer it.

[KristofferC]

I think safe is a bit too strong and doesn't really say what is unsafe with the normal convert so I like noncircular a bit more.

[JeffBezanson]

On my branch, the convert error doesn't work, since calling convert(Union{}, x) is ambiguous.

[timholy]

Sorry to steal a couple of issue-closings from you, but presumably you'll close so many when that merges you won't notice 😉. (Obviously feel free to modify those tests as needed.)