Add workspace/buffer support to sorting

[LilithHafner]

This PR adds an additional workspace argument to most internal sorting functions and allows the end-user to pass one in with the keyword argument workspace (undocumented).

There are two types of workspaces: nothing, and an AbstractVector. nothing indicates that sorting functions should allocate new workspaces as needed and discard them when done. An AbstractVector indicates that sorting functions should use that vector, resizing it larger if needed. Wherever possible, sorting functions can handle receiving nothing, and AbstractVector with appropriate eltype, or no argument at all.

This PR also utilizes that feature to implement solution 3 of #45326 and therefore closes #45326.

As is, this PR exposes a new public-facing keyword argument for (almost) all sorting functions:

sort(x, workspace=y)

Is there a standard name for this optional keyword argument?

[pcjentsch]

Is there a standard name for this optional keyword argument?

I think SciML usually calls it cache.

[LilithHafner]

I crossposted to discourse and am crossposting Jar1's response back

I usually think of a “cache” as being a cache of values, rather than a location with empty available memory.

Perhaps “workspace” or “buffer” works better.

[LilithHafner]

There's likely still room for improvement by letting the workspace function depend on OffsetArrays, but this is already pretty good:

@btime sort(x; dims=1) setup=(x=rand(Int, 3000, 3000)) evals=1;
v1.7.2 => 357.704 ms (5 allocations: 68.66 MiB)
v1.9.0 => 10.980 s (12005 allocations: 201.24 GiB)
PR     => 229.195 ms (3018 allocations: 150.60 MiB)

[LilithHafner]

Bump :)

[musm]

I would've preferred the keyword name buffer for this. Looking over the PR I kept expecting workspace to denote something related to the current session's work space of variables.

Is the intention to keep this as an internal undocumented feature?

[StefanKarpinski]

Agree that buffer would be a clearer name. I thought this might be about implementing variable clearing.

[LilithHafner]

Both seem fine to me. There is no reason we have to stick to "workspace" just because this merged. Googling definitions yields:

Buffer:
a temporary memory area in which data is stored while it is being processed or transferred, especially one used while streaming video or downloading audio.

Workspace:
a memory storage facility for temporary use.

Scratchspace:
Scratch space is space on the hard disk drive that is dedicated for storage of temporary user data

As long as we reach consensus by the time we start 1.9 prereleases this shouldn't be a problem.

[LilithHafner]

FWIW I just talked with a CS researcher who doesn't use Julia. She prefers "workspace".

[LilithHafner]

Another CS researcher who doesn't use Julia prefers Workspace over Cache, Buffer, and Scratchspace, saying "I feel like workspace is a more intuitive word for non-computer scientist people to understand, if you have to use something"

[musm]

I think that's arguable, but I thought I'd share my reaction looking through this PR. I'd hazard that throughout most of the Julia code base buffer is used to denote this concept.

Workspace to me denotes something entirely different. For example, in MATLAB (https://www.mathworks.com/help/matlab/ref/workspace.html) and VS Code https://code.visualstudio.com/docs/editor/workspaces. This was why I was thrown off initially. Apparently, there is a precedent from Fortran to use workspace for the concept, but that almost feels to me a little antiquated.

[LilithHafner]

Unless anyone else chimes in in favor of keeping workspace I'll do a renaming workspace => buffer once an opportunity arises that avoids excessive merge conflicts with ongoing sorting PRs.

[LilithHafner]

The renaming pull request is here

[LilithHafner]

As this is internal, I don't think it needs news or a compat annotation.