Use missing instead of nothing for dynamic values
#232
Conversation
|
For some reason all tests are failing but at different places for different Julia versions. I'll have to do some more poking around. |
This should fix some (but not all) of them: #230 |
|
Odd that this is getting an ambiguity error on I'm not getting that locally, but I am getting a lot of obviously-related failures from |
|
I assumed that |
…d === for comparisons, and Static.find_first_eq still returns nothing on failure, like Base.findfirst
|
The lu ambiguities are because it was using an old version of StaticArrays, before: We could add a test/Project.toml to restrict the version it tests with. |
Codecov Report
@@ Coverage Diff @@
## master #232 +/- ##
==========================================
+ Coverage 88.97% 89.01% +0.03%
==========================================
Files 11 11
Lines 1751 1730 -21
==========================================
- Hits 1558 1540 -18
+ Misses 193 190 -3
Continue to review full report at Codecov.
|
|
Doc tests fail, but should pass once the pkg server updates and has Static 0.5 available. |
|
Since this is breaking, do we want to wait to register this until we manage other breaking changes like moving BandedMatrixIndex out |
Sure, no rush. |
|
I just moved the banded matrix stuff into requires so that when we move to other packages nothing should change. |
|
Is this breaking change really worth to make? The whole SciML ecosystem is still forced on v3 and not able to use new features. |
|
It's been too long without downstream PRs. Let's revert it today. |
|
Is the problem just with getting LoopVectorization updated or do we need to go through the whole ecosystem? I'd be surprised if many packages had to change anything to accommodate this. |
#232 was found to be a huge mistake. For interface checks, you want `nothing` instead of `missing` because you want to handle the cases. The issue with `missing` is that it propagates, `missing + 1 == missing`. We don't want that here, we want clear and precise error messages at the spot where the `nothing` is found if it's not supposed to be there and it's supposed to be handled. That is antithetical to most of its usage here. So this is a strong revert with a breaking update, which should fix downstream libraries (LoopVectorization, OrdinaryDiffEq, etc.) which were never received the proper downstream PRs from the original change anyways.
No description provided.