Add check_only option to precompile

[kpamnany]

We want to compile a subset of the possible specializations of a function. To this end, we have a number of manually written precompile statements. Creating this list is, unfortunately, error-prone, and the list is also liable to going stale. Thus we'd like to validate each precompile statement in the list.

The simple answer is, of course, to actually run the precompiles, and we naturally do so, but this takes time.

We would like a relatively quick way to check the validity of a precompile statement.
This is a dev-loop optimization, to allow us to check "is-precompilable" in unit tests.

We can't use hasmethod as it has both false positives (too loose):

julia> hasmethod(sum, (AbstractVector,))
true

julia> precompile(sum, (AbstractVector,))
false

julia> precompile(sum, (AbstractVector,); check_only=true) # <- this PR
false

and also false negatives (too strict):

julia> bar(@nospecialize(x::AbstractVector{Int})) = 42
bar (generic function with 1 method)

julia> hasmethod(bar, (AbstractVector,))
false

julia> precompile(bar, (AbstractVector,))
true

julia> precompile(bar, (AbstractVector,); check_only=true) # <- this PR
true

We can't use hasmethod && isconcretetype as it has false negatives (too strict):

julia> has_concrete_method(f, argtypes) = all(isconcretetype, argtypes) && hasmethod(f, argtypes)
has_concrete_method (generic function with 1 method)

julia> has_concrete_method(bar, (AbstractVector,))
false

julia> has_concrete_method(convert, (Type{Int}, Int32))
false

julia> precompile(convert, (Type{Int}, Int32))
true

julia> precompile(convert, (Type{Int}, Int32); check_only=true)  # <- this PR
true

This PR adds the check_only Bool keyword argument to precompile; if true, the method instance isn't actually compiled. We can, of course, add a new isprecompilable instead, or move things around.

Cc: @nickrobinson251 whose idea this is, @gbaraldi who suggested opening the PR for discussion, and @JeffBezanson who had some thoughts on "compilability".

[gbaraldi]

Maybe add a isprecompilable option that's similar to hasmethod but that uses the jl_get_compile_hint_specialization logic that compile_hint has.

[IanButterworth]

Can you just run the original precompile statements in a --compile=no session?
It still gives the correct false here.

./julia --compile=no -e "Base.@trace_compile @show precompile(sum, (AbstractVector,))"
precompile(sum, (AbstractVector,)) = false

[kpamnany]

Maybe add a isprecompilable option that's similar to hasmethod but that uses the jl_get_compile_hint_specialization logic that compile_hint has.

See #58805.

[kpamnany]

@IanButterworth: starting a new process for this seems like a heavyweight solution.

[IanButterworth]

I thought the benefit of only having to construct your precompile statements outweighed that. Then you know for sure you're testing the right thing.

Maybe something like this instead?

@compile_no include("precompile_statements.jl")

[kpamnany]

Replaced by #58805.