suboptimal return type inference for Base.setindex

[nsajko]

julia> Core.Compiler.return_type(Base.setindex, Tuple{Tuple,Nothing,Int}) <: Tuple
false

julia> versioninfo()
Julia Version 1.12.0-DEV.495
Commit 3c966a5107a (2024-05-09 11:53 UTC)
Build Info:
  Official https://julialang.org/ release
Platform Info:
  OS: Linux (x86_64-linux-gnu)
  CPU: 8 × AMD Ryzen 3 5300U with Radeon Graphics
  WORD_SIZE: 64
  LLVM: libLLVM-17.0.6 (ORCJIT, znver2)
Threads: 1 default, 0 interactive, 1 GC (on 8 virtual cores)

FWIW the type inference for Base.setindex was better on v1.7.3.

A quick fix, for Base.setindex at least, is adding a ::Tuple type assertion for the Base._setindex return value.

However, I'm not sure why does the inference fail, actually? We do have:

Core.Compiler.return_type(ntuple, Tuple{Any,Val}) <: Tuple

Cthulhu.descend shows that, in Base._setindex, the return value gets inferred as Any even though the Val{N}() gets inferred as Val. Weird.

[martinholters]

julia> methods(ntuple, Tuple{Any,Val})
# 5 methods for generic function "ntuple" from Base:
 [1] ntuple(f, ::Val{3})
     @ ntuple.jl:50
 [2] ntuple(f, ::Val{2})
     @ ntuple.jl:49
 [3] ntuple(f, ::Val{1})
     @ ntuple.jl:48
 [4] ntuple(f, ::Val{0})
     @ ntuple.jl:47
 [5] ntuple(f::F, ::Val{N}) where {F, N}
     @ ntuple.jl:69

Five methods is too much for inference which therefore bails out. Which is quite unfortunate. Maybe it's worth renaming those to _ntuple and adding ntuple(f::F, ::Val{N}) where {F,N} = _ntuple(f, Val{N}())::Tuple{Vararg{Any,N}}?

[KristofferC]

#54528 is an alternative to that

[martinholters]

👍 Yes, that looks like good idea to me.