sizeof(::Array): typeassert ::Int to help abstract type inference

[nsajko]

Fix a regression where abstract return type inference of sizeof(::Array) used to produce Any after Julia v1.11.

[JeffBezanson]

I think instead we can delete some of the redundant Array length methods.

[JeffBezanson]

This should fix all these issues:

--- a/base/essentials.jl
+++ b/base/essentials.jl
@@ -7,9 +7,7 @@ const Callable = Union{Function,Type}
 const Bottom = Union{}
 
 # Define minimal array interface here to help code used in macros:
-length(a::Array{T, 0}) where {T} = 1
 length(a::Array{T, 1}) where {T} = getfield(a, :size)[1]
-length(a::Array{T, 2}) where {T} = (sz = getfield(a, :size); sz[1] * sz[2])

[nsajko]

instead we can delete some of the redundant Array length methods

I know (already opened relevant PRs before), but this seemed like on OK quick fix. Relevant PRs:

• decrease max_methods world-splitting setting for some functions #59377

• Base: bootstrap: eliminate Array-specific length methods #57627

[nsajko]

In particular, an issue with the proposed diff above @JeffBezanson, as pointed out in a comment by @KristofferC in one of the linked PRs of mine, is that it would cause new invalidations. That is why now I'm trying to combine deleting unnecessary length methods with lowering max_methods for length (and other functions).

[nsajko]

However it still might make sense to do the simple ::Int typeasserts where backportable.

[JeffBezanson]

Why does it cause invalidations? I would think adding new methods could not invalidate length(::Array). Do we have examples of this?

If it is indeed a problem, maybe keeping the method for Vector will help.

[nsajko]

You're right, PR #57627 does not seem to cause any invalidation. I'm moving the tests from these few most recent PRs to that PR.