broadcast: disable nospecialize logic for outer method signature

[vtjnash]

This removes the dependence on inlining for performance, so we also remove @inline, since it can harm performance.

@nanosoldier runbenchmarks("broadcast" || "sparse" || "array", vs=":master")

[mbauman]

This looks like it updates and replaces #35675? Does it work around the problem of the mutable Ref?

[vtjnash]

Ah, I had forgotten #35675 existed. I think we need to stop preventing progress on this bug. We seen these @inline substantially harm performance in many cases, but in a few cases lead to trivial allocations, which should typically not be considered a blocking issue.

[vtjnash]

I can put up a PR to add support in broadcast for Some (currently an error). And later we can also merge the PR that adds support for FillArray (currently not yet defined).

[nanosoldier]

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

[KristofferC]

bump

Edit: Ok, I read https://github.com/JuliaLang/julia/pull/35675#issuecomment-624222014 etc now...

[vtjnash]

It would be good to merge, so that folks had to actually merge the fix for that problem and not ignore them any longer

[mbauman]

I mean, I'm all for this, but we should know what the tradeoffs are and make an intentional decision that the good outweighs the bad.

@KristofferC do you have a branch with conflicts resolved against which we can run nanosoldier again?