Add @inline for do blocks

[tkf]

This PR makes @inline work with do blocks. This is very useful for writing high-performance code with higher-order functions such as foreach, foldl, and reduce.

[StefanKarpinski]

Alternately, should do blocks be automatically marked for probably inlining?

[JeffBezanson]

Alternately, should do blocks be automatically marked for probably inlining?

I would prefer not to do that, since many uses of do blocks don't need that level of performance, and it will add compile time cost. It's easier to mark cases where you need performance than to mark cases where you don't need performance, especially since the only symptom would be gradual degradation in compile times.

[tkf]

I think a long term solution would be to have call-site inlining as the caller has a good guess for what should be inlined in this case (e.g., foreach requires inlining but open does not).

[JeffBezanson]

Yes, call-site inlining would be good. Now that you mention it, the syntax in this PR conflicts with it --- @inline f() do ... end would mean f should be inlined if we had that feature.

[StefanKarpinski]

IIRC, @vtjnash claimed that call site inlining would be "fairly trivial to implement" like three yrs ago.

[tkf]

Maybe inline both is what you'd want in this case? For finer control, how about

Normal call

f(a, b) do c, d
end

Inline only f:

@inline(f(a, b)) do c, d
end

Inline only do block:

f(a, b) do c, d
    @inline  # just like @_inline_meta
end

[aviatesk]

Superseded by #41312.