Get tests to pass when running Julia with --compiled-modules=no.

[chriselrod]

This fixes JuliaSIMD/LoopVectorization.jl#192 (comment) for me: Trixi's tests pass, as do ArrayInterface's after starting Julia with --compiled-modules=no.

Normally, all functions in a module have the same world age, so the order in which generated functions are defined with respect to everything else is irrelevant. But not with --compiled-modules=no, which causes world age errors whenever an @generated function calls a method defined later.

I addressed this the lazy way, but just searching for all the @generated functions calling non-base methods and sticking them in a file at the end.

We could certainly do a lot better organizationally.

@Tokazama I'm not sure why it isn't letting me request you as a reviewer.

[codecov]

Codecov Report

Merging #114 (c82c7c5) into master (db43f3f) will not change coverage.
The diff coverage is 100.00%.

@@           Coverage Diff           @@
##           master     #114   +/-   ##
=======================================
  Coverage   84.26%   84.26%           
=======================================
  Files           8        8           
  Lines        1506     1506           
=======================================
  Hits         1269     1269           
  Misses        237      237           

Impacted Files	Coverage Δ
src/ArrayInterface.jl	84.81% <ø> (ø)
src/stridelayout.jl	76.38% <ø> (-0.75%)	⬇️
src/dimensions.jl	89.59% <100.00%> (+0.43%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update db43f3f...44b06f2. Read the comment docs.

[Tokazama]

We could certainly do a lot better organizationally

Agreed. Did you have any suggestions? I've been trying to move away from having one gigantic file but I probably complicated things in the process

[chriselrod]

Agreed. Did you have any suggestions? I've been trying to move away from having one gigantic file but I probably complicated things in the process

I think breaking it up into indexing/dimensions/etc is fine, but breaking things up into "is it a @generated function", which this PR did, is not. Except that it fixes world age problems introduced by using --compiled-modules=no.

However, I now deleted generatedfuncs.jl, moving most of the code back, with a much more minor reordering to make sure we don't get world age issues under --compiled-modules=no. @generated definitions should all be following the definitions they're using now, or at least all of them covered by our unit tests.

So this PR is in much better shape/much less objectionable now. I'll merge and make a new release fairly soon.