N5N3

Always return a new `Array` seems a better choice? At least we know that subsequent modifications to the output will never affect the original inputs. If we allow to return...

They should be unrelated, as we are pretty sure that `stack` is not used in `Base` and `InteractiveUtils`. I think this PR is ready to merge. Not sure will there...

But `@code_typed` gives the same result with/without this PR? So I'm assuming there's no performance difference. (Our compiler should be good enough to track this kind of local instability.)

The root cause should come from `intersect_var`, it has a branch that fixing the TypeVar's `ub/lb` eagerly. https://github.com/JuliaLang/julia/blob/e1fa6a51e4142fbf25019b1c95ebc3b5b7f4e8a1/src/subtype.c#L2296-L2297 Once we hit to this branch during `intersect_union`, `env` records the first...

It should not be static. ```julia julia> A = rand(5,12); julia> B = reshape(A', (3,4,5)); julia> C = randn(1,60); julia> D = reshape(C', (3,4,5)); julia> typeof(B) == typeof(D) true julia>...

I know non-static result might cause performance issue. But, the above example shows that we can't derive the dense dimension with the current type information. One possible solution is adding...

Comments added. If the test passed on 1.6, I'll add some test for strided layout.

test added

@maleadt Sorry for what happened. I don't have a available CUDA10.x environment, so I used the CI to verify the unexpected error which is correct in 11.x. At present, the...

Well, my mistake, the drop actrually failed. I want to write something like: ```julia @static if version() >= v"10.2.1" CuFFTArray{T,N} = StridedCuArray{T,N} # only CUDA 11.x pass the strided test...