stricter shapes for setindex

[adienes]

not now ambitious enough to claim this fixes #40018, but let's just say I wrote it while reading the issue

fixes JuliaSparse/SparseArrays.jl#569
updates #45374 (closes?)

[giordano]

not ambitious enough to claim this fixes #40018

By writing that you automatically marked this PR to close that issue anyway. If you don't want to do that, put something between "fixes" and "#40018", like "issue".

[adienes]

I (believe) this implements the idea in #40018 (comment)

I really hope it isn't a performance trainwreck

edit: BitArray needs fixing

[adienes]

JuliaSparse/SparseArrays.jl#637

would fix the sparsearrays test case

looks like there was some previous discussion in #43644

[adienes]

@nanosoldier runtests()

[nanosoldier]

The package evaluation job you requested has completed - possible new issues were detected.
The full report is available.

Report summary

✖ Packages that failed

9 packages failed only on the current version.

• Package has test failures: 1 packages
• Package tests unexpectedly errored: 5 packages
• Test duration exceeded the time limit: 3 packages

16 packages failed on the previous version too.

✔ Packages that passed tests

3 packages passed tests only on the current version.

• Other: 3 packages

58 packages passed tests on the previous version too.

~ Packages that at least loaded

3 packages successfully loaded only on the current version.

• Other: 3 packages

24 packages successfully loaded on the previous version too.

[adienes]

that's better. so there are ~5 real failures, all of which are trying to drop non-trailing singleton dimensions besides Vlasiator.jl, which gets
DimensionMismatch: tried to assign 3×8192 array to 3×32×16×16 destination

FYI I tried to outline the error path, but it looks like there are bootstrap issues with @noinline to annotate functions, so I had to remove it.

[jishnub]

there are bootstrap issues with @noinline

Possibly fixed by using Base.@_noinline_meta instead?

[adienes]

#45374 could be fixed by this PR if it were adapted to pass the axes of the indices to the shape check rather than the size, but maybe that comes with a performance cost?

[oscardssmith]

Triage approves (modulo pkgeval not being clean enough and code review/docs etc). If we get #54903, we should consider further restricting to the broadcast behavior (which doesn't have the vector opt out).

[adienes]

triage also considered the fact that this implementation would allow I or X in an A[I] = X expression to have differently-offset axes so long as the size matches the rule here, and was cautiously ok with that

[jishnub]

Just to clarify: ignoring the axes offset in setindex! seems to be what we already do (#45374), and this PR doesn't change anything here?

[adienes]

correct; this PR should not change the behavior of the MWE there

julia> a = [1:7;];

julia> a[Base.IdentityUnitRange(3:4)] = 2:3
2:3

although it may add documentation clarifying that this is allowed

[mbauman]

I threw a commit on here that tries to document this; it's just meant to be a starting point.

[adienes]

not to add another wrinkle but should this PR also remove the require_one_based_indexing here ?

julia/base/array.jl

Line 1028 in 13f6b3b

require_one_based_indexing(X)

if we are going to really commit to "sizes must match, axes need not" decision

[adienes]

I will need help coordinating with JuliaSparse/SparseArrays.jl#637