Fix juliac --trim

[johroj]

This PR is the result a series of fixes I implemented for FFTW to play nicely with juliac --trim. Sort of solves #319. Its not perfect, but I think it should be better than nothing for those of us who depend on FFTW for making meaningful evaluations of --trim.

Library Loading

This temporarily solves the issues surrounding the current way of loading libraries, which should be replaced in the future (as suggested by the comment at https://github.com/JuliaMath/FFTW.jl/blob/master/src/FFTW.jl#L33)

1. Add type parameter to FakeLazyLibrary for type stability.

And two issues related to ccall(::FakeLazyLibrary, ...) forwarding the call to dlopen(::FakeLazyLibrary):

2. Call dlopen in __init__to ensure it gets precompiled (which won't happen by default since it is only called from the C layer).
3. Manually running Base.unsafe_store!(cglobal(:jl_libdl_dlopen_func, Any), dlopen) before calling fft is necessary in all entrypoints for ccall to become aware of dlopen. Theoretically this could be done in FFTW, but I think its best to avoid messing with Julias internal mechanisms for applications outside of experimenting with --trim.

The related code on the Julia side is https://github.com/JuliaLang/julia/blob/f5f4c0174ed51de3811da861c0bf01b4e9f10972/src/runtime_ccall.cpp#L86 and https://github.com/JuliaLang/julia/blob/41e50a77e6f6b9b2cfea9056529224c56923dac6/base/libdl.jl#L488.

Other Type-Instability Fixes

Some additional type instabilities that are probably worth fixing even if the library loading mechanism is updated in the future:

4. Add FFTWPlanDestructor for type stable destruction.
5. Replace an occurence of string(args...) with join([str1, str2, ...]).
6. Use a minimal type-stable equivalent of @sync in spawnloop. I would be happy for a better proposal for how to solve this. The core issue is that spawnloop is precompiled even if only one thread is used. The internal use of a Channel{Any} in @sync makes this problematic. Compared to https://github.com/JuliaLang/julia/blob/29a4bbf950aefa6116b296d48ef966ab36c5ca84/base/task.jl#L584, the only significant difference I can only see is that the real @sync gives a nicer handling and representation of errors. I hope I did not miss anything. Worst case, we could drop this from the PR and leave it to the user to override spawnloop using --trim.

Example

With this PR the following will build with --trim and run without errors:

using FFTW
using Libdl

function @main(ARGS)
    # Must be called before anything from FFTW
    Base.unsafe_store!(cglobal(:jl_libdl_dlopen_func, Any), dlopen)

    x = [1., 2., 3.]
    println(Core.stdout, "rfft([$(join(x, ", "))]) = [$(join(rfft(x), ", "))]")
    return 0
end

[codecov]

Codecov Report

❌ Patch coverage is 95.23810% with 1 line in your changes missing coverage. Please review.
✅ Project coverage is 74.40%. Comparing base (062fddb) to head (b06d563).

Files with missing lines	Patch %	Lines
src/fft.jl	87.50%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #327      +/-   ##
==========================================
+ Coverage   73.64%   74.40%   +0.76%     
==========================================
  Files           5        5              
  Lines         535      547      +12     
==========================================
+ Hits          394      407      +13     
+ Misses        141      140       -1     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.