Summary:
Trying to improve my previous QE regression on all the perf: https://www.internalfb.com/intern/qe2/ig_ios_listkit_improvements_universe/ig_ios_listkit_improvements_skip_view_section_controller_map/review
My theory is that it's incresing the load in the -visibleObjects call, then we call -[UICollectionView cellForIndexPath:] which is expensive in the `_sectionControllerForCell:`
Thus my new approach is to optimize this -visibleObjects entirely and instead of dealing with visible cells, we can use the `indexPathsForVisibleItems` which can give us the indexpath instead and make it faster. It would be O(N) instead of O(N^2) compared to before.
Reviewed By: calimarkus
Differential Revision: D24663555
fbshipit-source-id: 69fcf2d0b44f8bc06351c92f96c3cbe227c22479
Summary:
Another source-of-truth inconsistency issue.
I kept getting this assertion T65423827 task.
It turned out the the root cause is that we use the `_viewSectionControllerMap` to keep track of the mapping between view and section controller.
Which is dangerous.
Because the UICV's source of truth is the `sectionMap`, now we are adding another source of truth between view <-> sectionController to be `_viewSectionControllerMap`, this is just waiting for bugs to happen.
We should always refers to sectionMap as the source of truth for everything here: since we can get the section index, we can replace all the `-sectionControllerForView` to use `-sectionControllerForSection:` instead, which is safer.
Multiple source of truth is the bug for this unexpected assertion.
Once we can cleanup this, then we dont need to use that weird: `- (void)mapView:(UICollectionReusableView *)view toSectionController:(IGListSectionController *)sectionController` API which we might forget to map, that can cause bug.
Reviewed By: elfredpagan
Differential Revision: D24173911
fbshipit-source-id: 6138cafc0b382fc569d17b14b13a6b50d85d342d
Summary:
## Changes in this pull request
### Checklist
- [x] All tests pass. Demo project builds and runs.
- [x] I added tests, an experiment, or detailed why my change isn't tested.
- [x] I added an entry to the `CHANGELOG.md` for any breaking changes, enhancements, or bug fixes.
- [x] I have reviewed the [contributing guide](https://github.com/Instagram/IGListKit/blob/master/.github/CONTRIBUTING.md)
Pull Request resolved: https://github.com/Instagram/IGListKit/pull/1466
Reviewed By: bdotdub
Differential Revision: D24235336
Pulled By: lorixx
fbshipit-source-id: bfc5fab26ae28ff9306d1b889b82b49c6ae2a636
Summary: Some product callsite just pass in nil object to this API thus keep causing the assertion. It's a bit too noisy and not much value here since we by default would still return `NSNotFound` here.
Reviewed By: stavash
Differential Revision: D24233480
fbshipit-source-id: 38b2eed3a9286bdd0a7eb0467894a13bbdca007a
Summary:
This diff introduces two experimental APIs to make `IGListKit` more usable with Swift values types:
- The `ListIdentifiable` protocol, which provides the equivalent of the diff identifier half of `ListDiffable`. For equality, we just use Swift's native `Equatable`.
- The `ListValueSectionController` base class, which provides a section controller interface for use with `ListIdentifiable` values.
These two APIs work together through the use of a private box class. I'd like to keep this detail hidden from callers, so that product code only deals with native Swift types.
Differential Revision: D23606413
fbshipit-source-id: 23718c508643f165c74550f1515d77448bd42d42
Summary:
Pretty sure in `-invalidateLayout`, we should calculate the `NSIndexPaths` after the delay. Otherwise, those `NSIndexPaths` might be incorrect after the update.
Before this change, running `test_whenInvalidatingInsideBatchUpdate_andRemoveThatSectionController_thatCollectionViewDoesntCrash` crashes with:
> Terminating app due to uncaught exception 'NSInternalInconsistencyException', reason: 'attempting to invalidate an item at an invalid indexPath: <NSIndexPath: 0x9566e79c2077363c> {length = 2, path = 1 - 0} globalIndex: 1 numItems: 1'
Reviewed By: candance
Differential Revision: D23742819
fbshipit-source-id: 39496d04025ff3554673313df3c7516c33c305d4
Summary:
These methods are currently correctly annotated as `nullable`, because they can return `nil`. However, this is not practical for writing actual Swift code that uses the APIs. Overrides of `cellForItem(at:)` //must// return a `UICollectionViewCell`. Without a collection context, there's no good way to get one, so callsites must either:
- Force-unwrap the `collectionContext`.
- `guard let`/`else fatalError` the `collectionContext`.
- Return a default `UICollectionViewCell`, which just hides the bug.
None of these are good: we don't want to encourage things force-unwraps and `fatalError` in idiomatic product code, because people will see them and get the idea that these patterns are okay to use elsewhere, and they are not. This also happens with `IGListGenericSectionController`'s `object` value: to use it when creating a cell, it must be explicitly unwrapped.
We could make these `nonnull`, since there's no enforcement of the annotations in Objective-C, that feels a bit too far, and it would break existing code that uses the optional API, because you can't force-unwrap (etc.) a non-optional value. Instead, we can use `null_resettable` explicitly. [While it's not covered by name in the Apple docs](https://developer.apple.com/documentation/swift/objective-c_and_c_code_customization/designating_nullability_in_objective-c_apis):
> Without a nullability annotation or with a null_resettable annotation—Imported as implicitly unwrapped optionals
...it bridges the property to Swift as an implicitly-unwrapped optional. I suppose it's implied as the equivalent of "without a nullability annotation".
This works even for `readonly` properties that can't be reset, and it solves both issues: it feels like less of a "100% `nonnull`" guarantee, and it's backwards-compatible with code using optionals.
To demonstrate that this is backwards-compatible with existing Swift code, this test code, which runs through various optional/non-optional APIs, compiles:
```
final class TestSectionController: ListSectionController {
override func cellForItem(at index: Int) -> UICollectionViewCell {
let cell = collectionContext.dequeueReusableCell(for: self, at: index)
let optional = collectionContext?.dequeueReusableCell(for: self, at: index)
if let _ = collectionContext {}
return optional ?? cell
}
}
final class TestGenericSectionController: ListGenericSectionController<NSString> {
override func cellForItem(at index: Int) -> UICollectionViewCell {
let label = UILabel() // imagine it's from a cell
label.text = object as String
label.text = (object as String) + "Suffix" // wouldn't work with optional
label.text = object.map { $0 as String }
if let object = self.object {
label.text = object as String
}
return collectionContext.dequeueReusableCell(for: self, at: index)
}
}
```
Reviewed By: patters, lorixx
Differential Revision: D23603716
fbshipit-source-id: e4750dcfe0072d482951dbf2e9efb1ee3de46884
Summary: If we apply the data updates outside the `performBatchUpdates`'s `updates` block, we need to make sure the layout is up to date. Otherwise, `performBatchUpdates` will try to do a layout before calling `updates` and get the wrong cells. Calling `[self.collectionView numberOfSections]` will update the section/item count, but won't get the cells.
Reviewed By: Haud
Differential Revision: D23604257
fbshipit-source-id: 02bbf0484c46cf6e7a12ed02d45ededb3b84ac19
Summary:
This is currently just `id`, so when you override it, you can't use the local without casting. We should declare it as `ObjectType` so that it gets the specialized type instead.
The implementation just assigns to the generic `_object` property-backing ivar, so no difference in type-safety.
Differential Revision: D23589584
fbshipit-source-id: 3784929f89580fd603fffcf940c7a5b502501bc5
Summary: We don't need to create the `emptyViewForListAdapter` if it's hidden
Reviewed By: patters
Differential Revision: D23429238
fbshipit-source-id: 8d85964c177f53a0e0cc0867339c1cbf0db2ee4e
Summary: We don't actually need to call `[UICollectionVew performBatchUpdates ...]` if we don't have changes.
Reviewed By: patters
Differential Revision: D23386181
fbshipit-source-id: 469a22a35b9ead5181e95f36a8bc59ba36373bbe
Summary:
We're seeing some crashes where the `UICollectionView`'s section count doesn't match the `fromObjets`. Ideally that shouldn't happen, but in case it does, lets fallback to a `reloadData` instead of crashing. Lets `IGFailAssert()` to keep an eye on it.
EDIT: Turns out `UICollectionView` will sometimes already do that for item-level inconsistencies! Example error message:
> The number of items contained in an existing section after the update (9) must be equal to the number of items contained in that section before the update (10), plus or minus the number of items inserted or deleted from that section (0 inserted, 0 deleted) and plus or minus the number of items moved into or out of that section (0 moved in, 0 moved out). - will perform reloadData.
But it still throws a `NSInternalInconsistencyException` for section-level inconsistencies.
Reviewed By: patters
Differential Revision: D23386178
fbshipit-source-id: a9930b619701cffd37d0d6e2bcb1268ef7c9d0ce
Summary:
It's easy to create a retain cycle or just forget to release a pointer, so lets test that the expendable parts of `IGListExperimentalAdapterUpdater` get deallocated after an update.
Lets also try to group related tests together. It's getting kind of hard to tell which tests already exist.
Reviewed By: patters
Differential Revision: D23386180
fbshipit-source-id: 90f0e6c7532287ee245e788dd752d45f368dc27e
Summary:
Changing the `IGListAdapter`'s `collectionView` or `dataSource` isn't safe. For example:
* If we change the `dataSource`, the `collectionView` will be out of sync with the `IGListAdapter`'s data and can easily crash.
* If we change the `collectionView` or `dataSource` during background diffing, the diffing result will be out of date by the time we get back to to the main main thread.
To fix this:
* On `[IGListAdapter setDataSource: ...]`, lets also change the `UICollectionView`'s `dataSource` to invalidate any section/item counts. We don't have to call `[UICollectionView reloadData]` just yet, we just want to let the collectionView know that the counts might have changed.
* On `[IGListAdapter setDataSource: ...]` and `[IGListAdapter setCollectionView:...]`, we should cancel any pending or on-going updates. The tricky part is that we should still execute the `itemUpdateBlocks` and `completionBlocks` in the right order.
I kind of want to make the `collectionView` readonly, but I think it would be too large of a public API change at this point.
Reviewed By: patters
Differential Revision: D23151919
fbshipit-source-id: 61cf025127706acaf22f153ec148ac0ea575bc98
Summary: Same as last diff, but for `IGListBatchUpdateTransaction`
Reviewed By: patters
Differential Revision: D23151920
fbshipit-source-id: 8941791b7870f34077a5e2beded2e913b453fbbf
Summary:
Using blocks can be nice, but here it causes a couple problems:
* The order of execution is a bit hard to follow. For example, in `-begin`, the very first thing we see defined is `executeCompletionBlocks` which is the last thing actually called. `IGListReloadTransaction` isn't too complex, but it gets even harder to follow in `IGListBatchUpdateTransaction`.
* It makes crash and assert reports hard to understand, because the stack includes mostly block names.
So lets turn blocks into methods.
Reviewed By: patters
Differential Revision: D23151918
fbshipit-source-id: 3ac4c77e9978c2700fd6744f084f624d27f0d300
Summary:
Lets move things to the right transaction:
* `performBatchUpdate` path to `IGListBatchUpdateTransaction`
* `reloadData` path to `IGListReloadTransaction`
Reviewed By: patters
Differential Revision: D23145770
fbshipit-source-id: e80fc05d2783e165354a147453083b449c92a61c
Summary: Now, lets move pending changes to a separate object `IGListUpdateTransactionBuilder`, which can we discarded once the update actually starts.
Reviewed By: patters
Differential Revision: D23145774
fbshipit-source-id: e12de178de33497f476972a9ad89ebb4e8d413ab
Summary:
Before adding `transactions`, lets clarify what information is collected before vs during the update. We can split apart `IGListBatchUpdates`.
* Before the update
* Collect the update blocks in `itemUpdateBlocks` which will be executed later during the update.
* Collect completion blocks in the same `completionBlocks` list as the other updates.
* During the update
* Collect item updates in `IGListItemUpdatesCollector`, like inserts and deletes.
* Collect new completion blocks (in case a section-controller schedules an update in -didUpdateToObject) in a separete list `inUpdateCompletionBlocks`
Reviewed By: patters
Differential Revision: D23145778
fbshipit-source-id: cae2c0689ac1ef0d93e3c04856b0495856e305b6
Summary:
Lets implement `performExperimentalUpdateAnimated`. There's a couple of things happening:
* We're now using the `IGListTransitionData` container object
* We don't need `pendingTransitionToObjects` anymore because we query the `fromObjects` just before performing the diffing.
* We should update the unit tests to use `performExperimentalUpdateAnimated` since the regular `performUpdateAnimated` will now fail.
Reviewed By: patters
Differential Revision: D23145781
fbshipit-source-id: 0b896e918241e709c5eb23f56a6b7323521a2222
Summary:
There's a few issues with `IGListAdapterUpdater` which would be difficult to safely fix "in place", so lets create a duplicate updater that we can test separatly.
Ugh, copy paste?
* The alternative would be to change `IGListAdapterUpdater` directly and have lots of `if/else` branching. I've tried it and it's hard to follow and easy to break. By creating a separate class, we do run the risk of having 2 diverging updaters, but I think it's worth the risk since it rarely gets changed. I'll try to ship (or burn) this new updater quickly.
Why duplicate `IGListAdapterUpdater` rather then starting from scratch?
* I want to take advantage of the existing unit tests. I can make small incremental changes and make sure the tests still pass.
* There's a lot going on in `IGListAdapterUpdater` and it's easier to refactor it by moving things around, rather then writing it from nothing.
Why does `IGListAdapterUpdater` need a clean up?
* Check out the first diff in the stack or task T74605897.
Reviewed By: patters
Differential Revision: D23145777
fbshipit-source-id: 360e980a89a5681ce38d1b5f6f1ca035eb1eb195
Summary:
For diffing, `IGListAdapterUpdater` uses the full `fromObjects` instead of `validObjects`, which can be different if we're missing `IGListSectionController`. The diffing results won't match what the `IGListAdapter` tells the `UICollectionView`, so we can crash.
To fix this, lets try to generate the `IGListSectionController` before performing the diffing, and keep them in a container object (`IGListTransitionData`) alonside the `fromObjects` and `toObjects`. This might also unblock other cool features in the future, like giving `IGListSectionControllers` time for background work, but that's for another day.
Since this change is relatively large, lets create a temporary protocol `IGListUpdatingDelegateExperimental` which will have a slightly different API from `IGListUpdatingDelegate`. When/if the new implementation ships, we can merge both protocols.
Why not edit `IGListAdapterUpdater` directly?
* I don't want to mess up exiting updaters just yet.
Reviewed By: patters
Differential Revision: D23145776
fbshipit-source-id: 21642cafa64e2a26a173e2ba683ef5c6cede17d7
Summary: Create temporary protocol for `IGListAdapterUpdater` to test a different implementation. The idea is to ship (or burn) the new implementation before the next 5.0.0 version release.
Reviewed By: patters
Differential Revision: D23145772
fbshipit-source-id: ede312c4a1289a34f73c14415a9bca40033fe925
Summary: We pass `experiments` to places that don't need it, so lets clean that up.
Reviewed By: patters
Differential Revision: D23145782
fbshipit-source-id: affcfe0f38d1b8e544940af89e5692dbdc36dd76
Summary:
On `IGListAdapater` data update:
1) Pass the capacity count, so that arrays don't have to re-size.
2) Avoid using a set, so that we don't need to deal with hashes and equality. The updater should have dealt with duplicates already.
Reviewed By: patters
Differential Revision: D23145771
fbshipit-source-id: 2ed93231e15ddcd66cfe4d1f7384c563c77caa8e
Summary:
Issue: Invalidating the layout of a `IGListSectionController` that's been removed from the map crashes, because we're passing `NSNotFound` to `[UICollectionView numberOfItemsInSection:...]`. This can easily happen if the section-controller invalidates its layout in a `performBatchAnimated` completion block, but the associated object was deleted during the update or after a full reload.
Fix: Like the rest of `IGListAdapter`, lets check for a valid `NSInteger` and bail early if we don't have it.
Reviewed By: mariajayne
Differential Revision: D22588010
fbshipit-source-id: bf0c26f313795bb10682ef6b660d6ea8a7ce9f1e
Summary: Wrap `[UICollectionView performBatchUpdates ...]` so that in case it crashes, the first app symbol will not be a block. A block name includes the line number, which means if you change the block line number, it will be categorized as a different crash. This makes tracking crashes across multiple app-versions a pain.
Differential Revision: D22398750
fbshipit-source-id: 90643e7eafe76b07e7022111f183e4734fd8a347
Summary: To prepare testing `Foundation` vs `IGListKit` diffing, lets measure how long it takes.
Reviewed By: Haud
Differential Revision: D22295546
fbshipit-source-id: 8023717f66ea68cbc24981272e8c134dd893274a
Summary: This header refers to `UICollectionViewCell` but does not import UIKit.
Reviewed By: candance
Differential Revision: D22340609
fbshipit-source-id: 8d03a48d363ecf3b00b29f89828d83ab44615fd4
Summary: The duration of an update depends a lot of if it's animated, so lets pass that info to the delegate.
Reviewed By: lorixx
Differential Revision: D22265405
fbshipit-source-id: 4d164447384b84eb6f41c7f8365d0b28670e1372
Summary: We call `[collectionView layoutIfNeeded]` before performing updates, but it's not clear why. It leads to some strange animation issues when performing updates while scrolling. So lets try to skip it?
Reviewed By: Haud
Differential Revision: D22244053
fbshipit-source-id: 0c01a73860ebc16f430fed04d5c29a5bde038286
Summary: This experiment shipped so we can remove it. We still need a way to disable it with `allowsReloadingOnTooManyUpdates` in case we don't want to resign the first responder on large updates.
Reviewed By: Haud
Differential Revision: D22219238
fbshipit-source-id: 98e78f3ed040809db6c4b4c8da8fd0e976aca0a2
Summary: If we fallback to `reloadData`, we don't queue the next update. That means if you call `performUpdate` multiple times quickly, there's a chance that the last update doesn't take effect.
Reviewed By: Haud
Differential Revision: D22219237
fbshipit-source-id: 167e3428859a0194f8f8c57624504edd531480df
Summary:
Issue: There's a couple of situations where `IGListAdapterUpdater` updates the `collectionView` without notifying the `IGListAdapterUpdater`.
* If we call `reloadDataFallback()` because of a missing window, there's no delegate calls.
* If we call `reloadDataFallback()` because of too many diff updates, we call `willPerformBatchUpdatesWithCollectionView` but not `didPerformBatchUpdates`.
Fix: Lets clean up the delegate calls
* If we `[UICollectioView performBatchUpdates:]` lets call `willPerformBatchUpdatesWithCollectionView` & `didPerformBatchUpdates`
* If we fallback to `[UICollectionView reloadData]` lets call `willReloadDataWithCollectionView` & `didReloadDataWithCollectionView`
* If we fallback to doing nothing, lets call `didFinishWithoutUpdates`
Reviewed By: Haud
Differential Revision: D22219236
fbshipit-source-id: 1afb4df8dbbaf4725424027bb52c1a5cc5100b30
Summary: Actually this won't work and causes even more crashes. For example, if we reload cell 0 and insert at 0, we end up with just a single delete & insert at 0, which causes an inconsistency exception. `IGListApplyUpdatesToCollectionView` already takes care of avoiding multiple reloads of the same `NSIndexPath` by using a `NSMutableSet` when collecting the `reloadDeletePaths`.
Reviewed By: apadalko
Differential Revision: D21822502
fbshipit-source-id: c2382951ffe013f82c1de2492286d2911f40444d
Summary: Ship this experiment. `IGListCollectionViewLayout` should get the section/index counts via `UICollectionView` to stay in sync, instead of the `dataSource`.
Reviewed By: apadalko
Differential Revision: D21789871
fbshipit-source-id: 9fe069bd793b36680cd8562e69c7367d69a11ec7
Summary:
= Context =
Thanks to the awesome Comparison View analysis, I am able to pinpoint where the scroll perf regression is for the single section inline update QE.
CV2 link: https://fburl.com/cv/fttkv99b
Also see from the graph that the these call stacks are pretty stat-sig,
https://pxl.cl/14D9M
And from the following stack stats, it's pretty clear that the heaviest one is calling from `-[IGListBindingSingleSectionController didUpdateToObject:]`, Approx. Frame Percentage is 24% to 1% compared to Control test group for my QE.
https://pxl.cl/14Db0
= Root cause =
Basically in my new QE, whenever IGListKit update, the -didUpdateToObject: is called on every single IGListSectionController, which in our case, we would call -[-configureCell:withViewModel:] all every single cells. And it's actually pretty expensive for any cell configuration:
https://pxl.cl/14DbZ
In control group, we never trigger any cell "Re-binding" thus no heavy impact. However the way control group works is, that we always trigger another cell creation and replace with the existing one. IGListKit would still check -isEqualToDiffable: to decide whether the current cell needs a reload, the logic is inside IGListAdapterUpdater.
= Solution =
It's easy, just do a simple `-isEqualtoDiffable:` check before updating the _item, inside `[-IGListBindingSingleSectionController didUpdateToObject:]` method.
Differential Revision: D20909174
fbshipit-source-id: 59d57fc8ddda7210cd1ae333942a345025cf1ee3
Summary: This allows us to potentially enable `-Wundef` in the future. Behavior is the same, undefined macros are false.
Differential Revision: D20906603
fbshipit-source-id: d5798861f0c9dfcd7a87936d575621d52c0ef7c7
Summary: We add a boolean isDisplayingCell method to check the existence of displayingCell in IGListBindingSingleSectionController
Reviewed By: chritto
Differential Revision: D20649871
fbshipit-source-id: f626cfdedff907604514c485d11362cd9ba12b99
Summary:
A common cause of `NSInternalInconsistencyException` is when a section controller tries to access a cell or modify the `UICollectionView` in `-didUpdateToObject`. If the collection view doesn't have data yet, this will kick off the regular calls (ex: `-numberOfSectionsInCollectionView`, `-numberOfItemsInSection`), which won't have the right information yet because the `IGListAdpater` is in the middle of updating the object list. Eventually, this can lead to a crash, which are hard to debug because the callstack doesn't include the call that caused the premature update.
Lets add an assert to catch when this happens.
Reviewed By: candance
Differential Revision: D20527372
fbshipit-source-id: 4ad0b2d89e5b126d6ae888a5c3b1d980945b7a4c
Summary:
**Context**
Recently an IGListKit experiment was introduced, `IGListExperimentPerformUpdatesWithoutDeferringCATransactionCommit`. When enabled, we opt to no longer defer a call to `[CATransaction commit]`, because it could feasibly end a different transaction than intended. In practice, this was leading to issues with `UIViewPropertyAnimator`, where deffered commits were ending in-progress animators.
**This Change**
The results we have seen from enabling this fix show no changes to performance and stability, so this seems safe to roll out. This change removes the experiment, and enables the new, non-deffered behavior.
Reviewed By: lorixx
Differential Revision: D20120169
fbshipit-source-id: 0473652020a250d67b02b860fb74c73e43615aef
Summary:
We were using the iVar version of the `_displayDelegate`, but turned out a lot of product callsite just manually override the `displayDelegate` instead of use `self.displayDelegate = self`.
Let's use the self.displayDelegate instead.
Reviewed By: bdotdub
Differential Revision: D20037298
fbshipit-source-id: fa860adfa88cd088f39718279983acd32e90b478
Summary:
There is a KP in IGListBindingSingleSectionController that we need to override the `displayDelegate` in order to track the displayingCell for performant update, when the section controller is not visible.
With the newly refactoring in how IGListSectionController handles the displaying events, IGListBindingSingleSectionController can now override the willDisplay/endDisplay calls without manually adding itself as the `displayDelegate`, the behavior works as before.
Differential Revision: D19973670
fbshipit-source-id: 1f9fec1bf88aa8755c163c636ebb49b377e6873c
