macrocall: Specialize macro invocations like normal function calls

[topolarity]

Inspired by #57781 (kudos to @nsajko for another good find)

jl_invoke behaves like Expr(:invoke, ...) not Core.invoke, so this was bypassing the default specialization for a macro and only ever invoked de-specialized MethodInstances.

Change this to jl_apply so that we specialize as normal, which should dramatically improve inference and not lead to much extra code since our AST is very homogeneous (Expr + typeof.(literals), IIUC).

[vtjnash]

only ever invoked de-specialized MethodInstances

That was sort of the point though, since inference is supposed to reject these also anyways, so attempted specialization here is expected to be just wasted effort

[topolarity]

inference is supposed to reject these also anyways, so attempted specialization here is expected to be just wasted effort

Why the invalidations in #57781 then?

[nsajko]

This PR doesn't seem to prevent any invalidation - the invalidations are exactly the same before and after applying this change on top of current master, 6eef79c.

[nsajko]

Furthermore, I can confirm the invalidation of macro MethodInstances, such as "MethodInstance for var\"@ip_str\"(::LineNumberNode, ::Module, ::Any)", is still present on master (and with this PR).

[JeffBezanson]

only ever invoked de-specialized MethodInstances

I'm not sure why the code is written this way (seems to be from c79e34c) but I don't think this is true? This code path does end up doing specialization. But then there is a check for functions starting with '@', and we skip inference in that case. But then I'm not sure where the backedges come from. There must be another code path that ends up invoking inference. But anyway, you could try disabling the '@' check in gf.c.

[JeffBezanson]

Oh we also insert nospecialize on macro arguments in lowering.

[topolarity]

Oh we also insert nospecialize on macro arguments in lowering.

Ah yep - Looks like that was auto-qualifying any macro for the "single specialization" pre-compile hint: #57833