StackOverflow when inferring Base.vect with many (>1000) arguments

[topolarity]

julia> x = 1.0
           # foo() = [x, 1.0, 2.0, 3.0, ...]
julia> @eval foo() = [x, $([Float64(i) for i=1:1000]...)]
foo (generic function with 1 method)

julia> foo()
Internal error: stack overflow in type inference of foo().
This might be caused by recursion over very long tuples or argument lists.

The problematic inference stack looks (roughly) like this:

1. Base.vect(::Any, ::Float64 (1000 times)...)
2. Base.promote_typeof(::Any, ::Float64, ::Float64, ::Float64, ::Float64 (1000 times)...)
3. Base.promote_typeof(::Any, ::Float64 (999 times)...)
4. Base.promote_typeof(::Any, ::Float64 (998 times)...)
5. Base.promote_typeof(::Any, ::Float64 (997 times)...)
   etc... (996 more times)

Code like this can apparently end up generated by MTK (SymbolicUtils.Code.create_array), which is where I encountered this.

[topolarity]

I think the strategy to solve this down-stream will be to:

• inline this reduction (e.g. promote_type(promote_type(promote_type(A, B), C), D)), and/or
• change the reduction strategy to minimize recursion depth (a parallel reduction would yield O(log(N)) recursion depth)

Both avoid the need for inference to descend large recursive function hierarchies
(and in theory, runtime performance should also be better)

[topolarity]

It also seems like inference should have better limits for this so that afoldl doesn't grow an arbitrary-length inference stack in the first place.

[vtjnash]

What version are you using? This seems likely to have been fixed already by #53665 as there is no afoldl in your backtrace, but there is on master

[oscardssmith]

I can confirm this works on master.

[topolarity]

Yeah, hit this on 1.10 - Sorry about that, should have checked on master but it's been a lot of different Julia versions this week 😅