Extend code_typed to be able to debug constant prop

[Keno]

I'm currently debugging code that uses a lot of arrays with
dimensions in their type parameters. This type of code heavily
relies on constant propagation to lift dimensions from the value
domain into the type domain. Unfortunately, it's a bit hard to
discover what exactly causes inference to drop information from
constants, because there's no way to feed in constants for a
particular invocation using code_typed. This is a quick hack to
remidy that, by making $-interpolated expressions available
as constants to type inference, e.g.

julia> @code_typed 1+1
CodeInfo(
53 1 ─ %1 = (Base.add_int)(x, y)::Int64
   └──      return %1
) => Int64

julia> @code_typed $(1)+$(1)
CodeInfo(
53 1 ─     return 2
) => Int64

This is missing handling for a bunch of the other cases in code_typed, which is why this is WIP. I plan to put those in and finish this up once I've confirmed this is useful, but figured I'd put it up in the meantime.

[vtjnash]

This capability does seem useful. I'd like to base it on top of #28955 though, so that hopefully the internal API can be better.

[Keno]

Yes, I agree. This also needs to be cleaned up a bit, but I've been using it quite extensively and it's been very useful.

[Keno]

Rebased on top of current master (i.e. including #28955), extended to handle a few extra syntax forms. Additionally, the Symbol argument in get/setproperty! now gets automatically consted (only in the syntax form of course).

[Keno]

In any case, this is good to go from my perspective. Any further review comments? Otherwise I'll plan to merge in a day or so.

[vtjnash]

I'm not sure I'll have a chance to finish review this week. Can you ping me again next week, unless someone else can make sure the _gen_call_with_extracted_types are correct and won't make it difficult to improve this code in the future (#20159 is the only open issue I can find to reference, so maybe we didn't open / prematurely closed the others)?

[vtjnash]

There appear to be a couple issues with this (invalid assumptions about the structure of types, typeof instead of Core.Typeof). Would it be possible to format this better too? The deeply nested conditionals and indentation are hard to follow.

[Keno]

@vtjnash ping. Other than the obvious Typeof, could you suggest what the correct way to write this is?

[jrevels]

Bump. @vtjnash ping again 🙂

[vchuravy]

Bump. This would be quite useful to have for https://github.com/vchuravy/Cthulhu.jl!

[User-764Q]

This issue is a few years old now, is it worth keeping it open or can it be closed?

Matt.

[timholy]

Cthulhu is now quite good for debugging constprop, and that seems like a fine place for me. But I'll let @Keno decide whether to close this.

[vchuravy]

I think the question is do we want to expose entering type-inference with const-args on the initial frame?

code_typed((Float64,)) do (x,)
   sin(1.0) + x
end

Is the current "canonical" workaround.
cc: @aviatesk