fix #23107, ensure Array constructor always makes new arrays

[JeffBezanson]

Fix #23107.

[fredrikekre]

@nanosoldier runbenchmarks(ALL, vs=":master")

[nanosoldier]

Something went wrong when running your job:

NanosoldierError: failed to run benchmarks against primary commit: failed process: Process(`sudo cset shield -e su nanosoldier -- -c ./benchscript.sh`, ProcessExited(1)) [1]

Logs and partial data can be found here
cc @ararslan

[JeffBezanson]

@nanosoldier runbenchmarks(ALL, vs=":master")

[nanosoldier]

Your benchmark job has completed - possible performance regressions were detected. A full report can be found here. cc @ararslan

[carlobaldassi]

We should have the same change applied to BitArrays too, for consistency.
(Currently, BitArray(::BitArray) just falls back to convert, here.)

[nalimilan]

Is this documented somewhere? This is something I wondered when writing CategoricalArrays contructors. Since one-argument constructors fall back to convert, by default they do not make copies. I agree it would make sense for convertnot to make a copy and for constructors to make a copy, but this should be stated clearly in the manual. Then we should review other types in Base (and in our packages) to check that this rule is followed consistently (not only for arrays).

EDIT: Sorry, I had missed the discussion at #23107. I agree with the conclusion there, so this is just a request to document this behavior instead of just silently doing it. :-)

[JeffBezanson]

I'll re-do the manual entries on convert and construct as part of #15120. In progress in #23273.

[JeffBezanson]

I looked into the apparent memory increase in the remotecall benchmarks, but that code doesn't seem to hit the new methods here.

[JeffBezanson]

There are possible remaining cases for SharedArray and IOBuffer here. However, there aren't any constructors and conversions for SharedArray(::SharedArray) currently, so the only matching method is the fallback identity conversion (unlike for Arrays, where whether a copy was made depended on whether the type was different). That will be removed with #15120, and then we can separately decide whether those types should have copy constructors.