[SwayamBorateIIT]

darray_copyto! builds the chunk range it iterates over from the element range, as
fld1(start):fld1(stop). An empty element range does not necessarily give an empty
chunk range, and when it does not, the loop indexes outside the chunk grid:

julia> A = distribute(zeros(16, 16), Blocks(4, 4));

julia> B = distribute(ones(16, 16), Blocks(4, 4));

julia> A[18:17, 1] = @view B[18:17, 1]
ERROR: BoundsError: attempt to access 4×4 Matrix{DTask} at index [5, 1]
[...]
 [21] darray_copyto!(B::DMatrix{...}, A::DMatrix{...}, Binds::Tuple{UnitRange{Int64}, Int64}, Ainds::Tuple{UnitRange{Int64}, Int64})
    @ Dagger ~/src/Dagger.jl/src/array/copy.jl:151
 [22] copyto!
    @ ~/src/Dagger.jl/src/array/copy.jl:209 [inlined]
 [23] setindex!(::DMatrix{...}, ::SubArray{...}, ::UnitRange{Int64}, ::Int64)
    @ Dagger ~/src/Dagger.jl/src/array/indexing.jl:141

18:17 gives fld1(18, 4):fld1(17, 4) = 5:5, which is not empty, so
B.chunks[5, 1] is read from a 4x4 grid. Whether it throws depends on where the two
ends happen to land, so neighbouring cases behave differently:

destination range	chunk range	result
1:0, 5:4, 16:15	empty	ok
17:16	5:4, empty	ok
18:17, 19:18, 20:19	5:5, not empty	BoundsError

This is reachable from ordinary code: LinearAlgebra's generic tril/triu fill
their result column by column with
A[max(firstindex(A,1), col-k):lastindex(A,1), col] = @view M[rows, col], and that
range is empty and starts past the end for any band offset past the corner of the
matrix.

The fix returns before the chunk ranges are computed when the destination range is
empty, since there is nothing to copy.

Testing

Added a testset covering both the previously-throwing and the previously-working
empty ranges, in rows and in columns.

What I ran locally (--test array/copyto -p 1 -t 1):

• on this branch: 658 pass.
• on master: 650 pass — the baseline.
• on this branch with src/array/copy.jl reverted to master: 655 pass, 3 errored,
  on rows=18:17, rows=20:19 and cols=18:17, each the BoundsError above, with the
  five neighbouring empty ranges still passing. So the new tests fail without the fix
  rather than passing for an unrelated reason.

I have not run the full suite: the GPU and streaming files exhaust memory on my machine,
so I am relying on CI for the rest.

---

The change and tests were written with the help of Claude Code (Opus 5). I reviewed the
diff, reproduced the bug, and ran the tests above myself.

[jpsamaroo]

Thanks so much!