Fix #19503 (provide a unary minus method specialized for sparse matrices)

[Sacha0]

As @StephenVavasis caught in #19503 (thanks!), I accidentally removed the unary minus method specialized for sparse matrices in #17265. This pull request provides a replacement method. Fixes #19503. Best!

[StefanKarpinski]

Is there any way to test that we don't accidentally delete this again in the future?

[Sacha0]

Is there any way to test that we don't accidentally delete this again in the future?

I was wondering the same. In an earlier PR I somehow checked calling of a specialized method, but I haven't been able to figure out which PR that was. Will spend a bit more time looking. Thanks for reviewing! Edit: Found it, #16979. Update inbound.

[Sacha0]

Now with a test. Thanks!

[simonbyrne]

Could also add something to BaseBenchmarks.jl?

[Sacha0]

Could also add something to BaseBenchmarks.jl?

Could, but not certain that's worthwhile? The test included in this PR should prevent regressions, and a benchmark of the one-liner in this PR distills to a benchmark of map(-, <:Vector) and copy(<:Vector). Thoughts? Thanks!

[tkelman]

always worth adding more benchmarks

[Sacha0]

Cheers, I can prepare a BaseBenchmarks PR at some point if you feel this is worth benchmarking. Question though:

always worth adding more benchmarks

Does that statement not fail at some point? Consider this illustrative hyperbole: Would a benchmark equivalent to @benchmark (x::Int = 0; for k in 1:typemax(Int); x += 1; end) be worthwhile? And with such a benchmark, would an additional benchmark @benchmark (x::Int = 1; for k in 2:typemax(Int); x += 1; end) be worthwhile? Where does one draw the line?

[tkelman]

If it's completely redundant with existing benchmarks then it's probably not worth adding. But I don't think this is, since the tests and dispatch here can change over time and independently tracking the performance of the end-user operation that was reported here is worthwhile.

[Sacha0]

If it's completely redundant with existing benchmarks then it's probably not worth adding. But I don't think this is, since the tests and dispatch here can change over time and independently tracking the performance of the end-user operation that was reported here is worthwhile.

Cheers, I appreciate the explanation!

[Sacha0]

Suggested benchmark submitted: JuliaCI/BaseBenchmarks.jl#45. Thanks!

[Sacha0]

Thanks for reviewing / merging!