[release-0.5] Backport #19015

[timholy]

This is against release-0.5. CC @tkelman, @nalimilan.

[tkelman]

missing the pr reference in the commit message

[timholy]

Each references the "parent" commit separately (I used cherry-pick -x). Should I have done it differently? How do you recommend fixing?

[tkelman]

commit --amend on the last one and add "ref #19015"

[timholy]

As a preemptive strike, I just did a grep on my installed packages. It looks like the following call Base.reduced_dims and/or Base.reduced_dims0 directly:

• DataArrays
• Images
• PackageEvaluator
• StatsBase

With this PR, the return type is now an Indices tuple rather than a Dims tuple, so it's almost certain to cause a breakage.

If that's too much breakage, there's an easy fix: make the internal function _reduced_dims (which returns an Indices tuple) and then have reduced_dims unwrap any OneTo tuple as a Dims tuple. That should (theoretically) make this "invisible." Would you prefer that?

[timholy]

Alternatively, we could rename the core function reduced_indices on julia-0.6, define a deprecated reduced_dims to wrap-and-unwrap Indices<--->Dims, and then having a version of that change join this one before backporting to 0.5. The advantage of the latter strategy is that there would be some testing of this last piece on 0.6 first.

[timholy]

See #19103.

[timholy]

OK, commit message edited to link to the original PR, and #19103 added. In theory this should be 100% non-breaking.

[tkelman]

Thanks. I'll run pkgeval on a separate PR, hopefully this won't break anything but let's see.

[timholy]

Probably want #19196 too.

[tkelman]

That only touched deprecated.jl, which this didn't

[timholy]

Right, but the same issue might apply. I haven't checked carefully, though. (This had some of the deprecated functions without the deprecations.)