Deprecate vectorized & in favor of compact broadcast syntax

[Sacha0]

This pull request deprecates all remaining vectorized & methods in favor of compact broadcast syntax. Best!

[musm]

This will lead to a performance regression, since the corresponding .& version will have to check each element?

[Sacha0]

Please see the discussion in #17623, specifically #17623 (comment), #17623 (comment), and #17623 (comment). Best!

[stevengj]

Might be better to hold off on these boolean-operator deprecations for BitArray in 0.6, to be conservative about the potential performance implications.

[Sacha0]

Might be better to hold off on these boolean-operator deprecations for BitArray in 0.6, to be conservative about the potential performance implications.

Thinking along similar lines, I planned to restore the efficient vectorized &, |, and xor methods as broadcast specializations so that they remain accessible without stalling deprecation of the vectorized methods. Would you be on board with that approach as well? Thanks!

[stevengj]

I thought about that, but the problem as usual is that specialized broadcast methods often won't be called due to fusion.

On the other hand, in cases where fusion occurs, you probably don't want the specialized method. For example, on my machine, if I compare

f(A) = (A .> 0.3) .& (A .< 0.7)
g(A) = (A .> 0.3) & (A .< 0.7)

then f is about 30% faster than g for A = rand(10^5) (and allocates less memory) due to fusion, even though g calls the optimized &.

So I guess it makes sense to deprecate & etc., define specialized broadcast methods, and maybe revisit this whole issue later if there are further performance improvements to be had.

(Specialized broadcast methods can also be a bit of a pain because they often introduce method ambiguities, e.g. with the sparse methods.)

[Sacha0]

Cheers, seems we're on the same page. Writing performant generic broadcast methods for BitArray (if necessary) seems the best long-run approach (e.g. as we've done for sparse vectors/matrices); broadcast specializations merely serve as a stopgap that teaches the new pattern while providing access to the former performance. Best!

[Sacha0]

Subsumed by #19791.