fix #31674, error when storing nonzeros into structural zeros with .= #31678
Conversation
Previously, broadcasted assignment (`.=`) would happily ignore all nonstructured portions of the destination, regardless of whether the broadcasted expression would actually evaluate to zero or not. This changes these in-place methods to use the same infrastructure that out-of-place broadcast uses to determine the result type. If we are unsure of the structural properties of the output, we fall back to the generic implementation, which will attempt to store into every single location of the destination -- including those structural zeros. Thus we now error in cases where we generate nonzeros in those locations.
| T = Tridiagonal(rand(N - 1), rand(N), rand(N - 1)) | ||
| ◣ = LowerTriangular(rand(N,N)) |
|
Why is it necessary to add special checks for this instead? I would assume that, in general, in-place broadcasting would automatically fail when there are structural zeros with only whitelisted special cases (zero preserving mapings) would work. |
|
Yes, the general broadcast methods just do the right thing. An alternative PR would be to simply remove all these But these |
|
Should this be merged? |
|
Yes please! This is good by me; I'm probably just overly conservative about waiting for express approval before merging my own PRs. |
…#31678) Previously, broadcasted assignment (`.=`) would happily ignore all nonstructured portions of the destination, regardless of whether the broadcasted expression would actually evaluate to zero or not. This changes these in-place methods to use the same infrastructure that out-of-place broadcast uses to determine the result type. If we are unsure of the structural properties of the output, we fall back to the generic implementation, which will attempt to store into every single location of the destination -- including those structural zeros. Thus we now error in cases where we generate nonzeros in those locations. (cherry picked from commit 6bd3967)
Previously, broadcasted assignment (
.=) would happily ignore all nonstructured portions of the destination, regardless of whether the broadcasted expression would actually evaluate to zero or not. This changes these in-place methods to use the same infrastructure that out-of-place broadcast uses to determine the result type. If we are unsure of the structural properties of the output, we fall back to the generic implementation, which will attempt to store into every single location of the destination -- including those structural zeros. Thus we now error in cases where we generate nonzeros in those locations.