Support other forms of quote node in allocated macro function elision

[vtjnash]

Refs #58057

[odow]

Is this the right fix? x-ref #58057 (comment)

julia> import MathOptInterface as MOI

julia> using Test

julia> function my_is_simply_call(Base.@nospecialize ex)
           Meta.isexpr(ex, :call) || return false
           for a in ex.args
               a isa QuoteNode && continue
               a isa Symbol && continue
               Base.is_self_quoting(a) && continue
               Base.is_quoted(a) && continue
               return false
           end
           return true
       end
my_is_simply_call (generic function with 1 method)

julia> macro my_allocated(ex)
           ex = macroexpand(__module__, ex)
           if !my_is_simply_call(ex)
               ex = :((() -> $ex)())
           end
           pushfirst!(ex.args, Base.GlobalRef(Base, :allocated))
           return esc(ex)
       end
@my_allocated (macro with 1 method)

julia> function test_main_captured_x()
           r = MOI.Nonlinear.OperatorRegistry()
           x = [1.1, 1.0]
           H = zeros(2, 2)
           ret = @my_allocated MOI.Nonlinear.eval_multivariate_hessian(r, :^, H, x)
           x = [1.1, 2.0]
           return ret
       end
test_main_captured_x (generic function with 1 method)

julia> test_main_captured_x()
112

julia> test_main_captured_x()
112

[vtjnash]

Yes, but you're still not allowed to have highly complicated expressions with arbitrary side-effects (MOI.Nonlinear.eval_multivariate_hessian) to benefit from this automatic simple-call simplification

[odow]

but you're still not allowed to have highly complicated expressions with arbitrary side-effects

I'm not sure I understand this part. The code does not allocate if I remove the @allocated macro. Previously, @allocated reported 0. Now it reports > 0? That seems like a break in behavior, even if the fix is intended to allow optimizations for other code.

[JeffBezanson]

Maybe we should just make @allocated officially only operate on the called function, i.e. allocations in argument expressions are not counted. That's a more dramatic change but it's probably what is wanted most of the time.

[vtjnash]

Perhaps most of the time, but it would be a breaking change for several packages (not all packages measure a call expression)

[JeffBezanson]

Fair enough. But the unintuitive thing here is that A.B.C is seen as a name, and not a computation. If you switch it to

                  f = MOI.Nonlinear.eval_multivariate_hessian
                  ret = @my_allocated f(r, :^, H, x)

you get 0 as expected. Maybe the right heuristic is just "is the expression a call"? Then if you really want to measure everything, all you need to do is wrap it in begin...end.

[vtjnash]

Shoot, if it needs a test and all that, this PR isn't worth it anymore. There's too many cases in the wild already doing various things. The PR currently breaks because we don't know that things are globals, and capturing those as arguments fails very badly for inference if they turn out to have been a Module.