[FFTW]: Try removing win64 AVX workaround

[staticfloat]

Just a testing PR to see what happens when we remove the win64 AVX assembler workaround; without this we usually get an assembler error, but perhaps with the right GCC version, we can avoid such issues.

[staticfloat]

@stevengj it looks like with GCC v8, it no longer requires the -fno-asynchronous-unwind-tables compiler flag. I'm deeply suspicious of that flag, so my first debugging step would be to have a user download this file and copy the dlls over the old stuff in the artifacts folder: https://julia-bb-buildcache.s3.amazonaws.com/50f3c78ff741e7718025385195b29e11b5390350/faa8b5976285c59f883f372ba1ae3949a3b31f72f808aca83b23bf351216c1e7/x86_64-w64-mingw32.tar.gz

That will allow us to quickly test whether this flag is actually what's wrong, and if it is, we can just merge this PR and get a new JLL pushed out to everybody.

[giordano]

Maybe want to post to https://discourse.julialang.org/t/serious-slowdown-with-fftw-v1-1-0-and-v1-2-0-on-windows/34726?

[giordano]

It might have worked: https://discourse.julialang.org/t/serious-slowdown-with-fftw-v1-1-0-and-v1-2-0-on-windows/34726/20. Hopefully we don't need to do what was suggested in #534 (comment)

[staticfloat]

Fantastic. We're rolling with it.

[giordano]

This might have broken FFTW on 32-bit ARM: JuliaMath/FFTW.jl#143.

Edit: well, I don't see how since this shouldn't have touched ARM 😕

[staticfloat]

Yeah, that bug report also doesn’t necessarily have anything to do with the JLL.

[giordano]

But it's failing also on Drone after JuliaMath/FFTW.jl#138, which is also very unlikely to have any effect on loading libfftw3. All previous builds on Drone worked, but they were using the JLL before this one.

[giordano]

Apparently pinning the jll before this PR fixed the issue on ARM: JuliaMath/FFTW.jl#143 (comment) 😕