reduce: unalias arrays for hcat/vcat (map)reduce #58490
Conversation
I was using `reduce(vcat, list_of_arrays)` and then modifying the resulting array. This worked great, except that in the case when `length(list_of_arrays) == 1`, the resulting array would alias the input, resulting in the input being modified and causing havoc down the road. Of course the proper fix for this is better built-in support for immutable arrays (and possibly memory ownership) to encode those assumptions explicitly. However, while this hack is unsatisfying, I do think it is somewhat defensible. One argument `hcat` and `vcat` perform the copy, as do the various optimized special cases for these functions. Of course, this is fragile and would break as soon as someone puts it in a closure, but maybe it saves at least a little bit of headache. I considered a more aggressive change to make the generic fallback `reduce_first(op, x) = op(x)`, but in discussion with Jeff, this was deemed too breaking, since it would add an assumption that `op` has a one-arg form.
In #58490, I said: ``` I considered a more aggressive change to make the generic fallback reduce_first(op, x) = op(x), but in discussion with Jeff, this was deemed too breaking, since it would add an assumption that op has a one-arg form. ``` However, what if we just checked whether `op` does have an appropriate method. This would generalize the fix also. For example, `reduce(+, list_of_arrays)` has the same aliasing issue that I complained about in #58490. In addition, this lets us remove the various reduce_first special cases, since the one arg reduction op has the correct behavior for all existing cases.
|
Or maybe we should do |
|
I like this as is. It also fixes another problem: the specific example in #34380, which isn't about aliasing but rather return values (and types) themselves. It's rather interesting that this crops up so very often with the I think that's a pretty good signal this is warranted and should be a targeted fixup like this. |
Maybe horrible option: reduce_first(op, x) = applicable(op, x) ? op(x) : x |
I had the same thought: #58491 but turns out |
|
Triage thinks that we definitely need this method. |
I was using
reduce(vcat, list_of_arrays)and then modifying the resulting array. This worked great, except that in the case whenlength(list_of_arrays) == 1, the resulting array would alias the input, resulting in the input being modified and causing havoc down the road. Of course the proper fix for this is better built-in support for immutable arrays (and possibly memory ownership) to encode those assumptions explicitly. However, while this hack is unsatisfying, I do think it is somewhat defensible. One argumenthcatandvcatperform the copy, as do the various optimized special cases for these functions. Of course, this is fragile and would break as soon as someone puts it in a closure, but maybe it saves at least a little bit of headache.I considered a more aggressive change to make the generic fallback
reduce_first(op, x) = op(x), but in discussion with Jeff, this was deemed too breaking, since it would add an assumption thatophas a one-arg form.