We recently reworked our `ng_rollup_bundle` rule to no longer output
ESM5 and to optimize applications properly (previously applications were
not optimized properly due to incorrect build optimizer setup).
This change meant that a lot of symbols have been removed from the
golden correctly. See: fd65958b88
Unfortunately though, a few symbols have been accidentally removed
because they are now part of the bundle as ES2015 classes which the
symbol extractor does not pick up. This commit fixes the symbol
extractor to capture ES2015 classes. We also update the golden to
reflect this change.
PR Close#38093
Currently we read lifecycle hooks eagerly during `ɵɵdefineComponent`.
The result is that it is not possible to do any sort of meta-programing
such as mixins or adding lifecycle hooks using custom decorators since
any such code executes after `ɵɵdefineComponent` has extracted the
lifecycle hooks from the prototype. Additionally the behavior is
inconsistent between AOT and JIT mode. In JIT mode overriding lifecycle
hooks is possible because the whole `ɵɵdefineComponent` is placed in
getter which is executed lazily. This is because JIT mode must compile a
template which can be specified as `templateURL` and those we are
waiting for its resolution.
- `+` `ɵɵdefineComponent` becomes smaller as it no longer needs to copy
lifecycle hooks from prototype to `ComponentDef`
- `-` `ɵɵNgOnChangesFeature` feature is now always included with the
codebase as it is no longer tree shakable.
Previously we have read lifecycle hooks from prototype in the
`ɵɵdefineComponent` so that lifecycle hook access would be monomorphic.
This decision was made before we had `T*` data structures. By not
reading the lifecycle hooks we are moving the megamorhic read form
`ɵɵdefineComponent` to instructions. However, the reads happen on
`firstTemplatePass` only and are subsequently cached in the `T*` data
structures. The result is that the overall performance should be same
(or slightly better as the intermediate `ComponentDef` has been
removed.)
- [ ] Remove `ɵɵNgOnChangesFeature` from compiler. (It will no longer
be a feature.)
- [ ] Discuss the future of `Features` as they hinder meta-programing.
Fix#30497
PR Close#35464
Interestingly enough, our rollup bundle optimization pipeline
did not work properly before 1b827b058e5060963590628d4735e6ac83c6dfdd.
Unused declarations were not elided because build optimizer did not
consider the Angular packages as side-effect free. Build optimizer has
a hard-coded list of Angular packages that are considered side-effect
free. Though this one did not match in the old version of the rollup
bundle rule, as internal sources were resolved through their resolved
bazel-out paths. Hence build optimizer could not detect the known
Angular framework packages. Now though, since we leverage the
Bazel-idiomatic `@bazel/rollup` implementation, sources are resolved
through linked `node_modules`, and build optimizer is able to properly
detect files as side-effect free.
PR Close#37623
Remove `looseIdentical` implementation and instead use the ES2015 `Object.is` in its place.
They behave exactly the same way except for `+0`/`-0`.
`looseIdentical(+0, -0)` => `true`
`Object.is(+0, -0)` => `false`
Other than the difference noted above, this is not be a breaking change because:
1. `looseIdentical` is a private API
2. ES2015 is listed as a mandatory polyfill in the [browser support
guide](https://angular.io/guide/browser-support#mandatory-polyfills)
3. Also note that `Ivy` already uses `Object.is` in `bindingUpdated`.
PR Close#37191
Dynamic embedded views were conceptually different from inline embedded views, but we have since
removed the inline embedded views so we now only have "embedded views".
See related refactoring work to remove inline embedded views in #34715
and #37073.
PR Close#37117
The ActiveIndexFlag is no longer needed because we no longer have "inline embedded views".
There is only one type of embedded view so we do not need complex tracking for
inline embedded views.
HAS_TRANSPLANTED_VIEWS now takes the place of the ACTIVE_INDEX slot as a
simple boolean rather than being a shifted flag inside the ACTIVE_INDEX bits.
PR Close#37073
In 420b9be1c1 all style-based sanitization code was
disabled because modern browsers no longer allow for javascript expressions within
CSS. This patch is a follow-up patch which removes all traces of style sanitization
code (both instructions and runtime logic) for the `[style]` and `[style.prop]` bindings.
PR Close#36965
Only refresh transplanted views at the insertion location in Ivy.
Previously, Ivy would check transplanted views at both the insertion and
declaration points. This is achieved by adding a marker to the insertion
tree when we encounter a transplanted view that needs to be refreshed at
its declaration. We use this marker as an extra indication that we still
need to descend and refresh those transplanted views at their insertion
locations even if the insertion view and/or its parents are not dirty.
This change fixes several issues:
* Transplanted views refreshed twice if both insertion and declaration
are dirty. This could be an error if the insertion component changes
result in data not being available to the transplanted view because it
is slated to be removed.
* CheckAlways transplanted views not refreshed if shielded by
non-dirty OnPush (fixes#35400)
* Transplanted views still refreshed when insertion tree is detached
(fixes#21324)
PR Close#35968
Prior to this change, animations-related runtime logic assumed that the @HostBinding and @HostListener with synthetic (animations) props are used for Components only. However having @HostBinding and @HostListener with synthetic props on Directives is also supported by View Engine. This commit updates the logic to select correct renderer to execute instructions (current renderer for Directives and sub-component renderer for Components).
This PR resolves#35501.
PR Close#35568
Prior to this commit unbound attributes were treated as possible inputs to structural directives. Since structural directives can only accepts inputs defined using microsyntax expression (e.g. `<div *dir="exp">`), such unbound attributes should not be considered as inputs. This commit aligns Ivy and View Engine behavior and avoids using unbound attributes as inputs to structural directives.
PR Close#36441
This commit performs a few updates to internal functions that would be required in upcoming changes to support synthetic host bindings in Directives.
* the `elementPropertyInternal` function was refactored to accept renderer as an argument (prior to that, there was a function that loads the renderer in some specific way for animation bindings)
* `elementPropertyInternal`, `elementAttributeInternal` and `listenerInternal` functions were updated to have a fixed set of arguments (for better performance)
* `elementPropertyInternal` and `elementAttributeInternal` functions were updated to take `tNode` as an argument instead of passing node index (that was used to retrieve `tNode` internally), in some cases we already have `tNode` available or we can retrieve it from the state
The refactoring was triggered by the need to pass different renderers to the `elementPropertyInternal` to support synthetic host bindings in Directives (see this comment for additional context: https://github.com/angular/angular/pull/35568/files#r388034584).
PR Close#35884
Root cause is that for perf reasons we cache `LFrame` so that we don't have to allocate it all the time. To be extra fast we clear the `LFrame` on `enterView()` rather that on `leaveView()`. The implication of this strategy is that the deepest `LFrame` will retain objects until the `LFrame` allocation depth matches the deepest object.
The fix is to simply clear the `LFrame` on `leaveView()` rather then on `enterView()`
Fix#35148
PR Close#35156
- Adds `TView` into `LFrame`, read the `TView` from `LView` on `enterView`.
- Before this change the `TView` was ofter looked up from `LView` as `lView[TVIEW]`. This is suboptimal since reading from an Array, requires that the read checks array size before the read. This means that such a read has a much higher cost than reading from the property directly. By passing in the `TView` explicitly it makes the code more explicit and faster.
- Some rearrangements of arguments so that `TView` would come before `LView` for consistency.
PR Close#35069
Inside `*ngFor` the second run of the styling instructions can get into situation where it tries to read a value from a binding which has not yet executed. As a result the read is `NO_CHANGE` value and subsequent property read cause an exception as it is of wrong type.
Fix#35118
PR Close#35133
This change changes the priority order of static styling.
Current priority:
```
(least priority)
- Static
- Component
- Directives
- Template
- Dynamic Binding
- Component
- Map/Interpolation
- Property
- Directives
- Map/Interpolation
- Property
- Template
- Map/Interpolation
- Property
(highest priority)
```
The issue with the above priority is this use case:
```
<div style="color: red;" directive-which-sets-color-blue>
```
In the above case the directive will win and the resulting color will be `blue`. However a small change of adding interpolation to the example like so. (Style interpolation is coming in https://github.com/angular/angular/pull/34202)
```
<div style="color: red; width: {{exp}}px" directive-which-sets-color-blue>
```
Changes the priority from static binding to interpolated binding which means now the resulting color is `red`. It is very surprising that adding an unrelated interpolation and style can change the `color` which was not changed. To fix that we need to make sure that the static values are associated with priority of the source (directive or template) where they were declared. The new resulting priority is:
```
(least priority)
- Component
- Static
- Map/Interpolation
- Property
- Directives
- Static
- Map/Interpolation
- Property
- Template
- Static
- Map/Interpolation
- Property
(highest priority)
```
PR Close#34938
Previously we would write to class/style as strings `element.className` and `element.style.cssText`. Turns out that approach is good for initial render but not good for updates. Updates using this approach are problematic because we have to check to see if there was an out of bound write to style and than perform reconciliation. This also requires the browser to bring up CSS parser which is expensive.
Another problem with old approach is that we had to queue the DOM writes and flush them twice. Once on element advance instruction and once in `hostBindings`. The double flushing is expensive but it also means that a directive can observe that styles are not yet written (they are written after directive executes.)
The new approach uses `element.classList.add/remove` and `element.style.setProperty/removeProperty` API for updates only (it continues to use `element.className` and `element.style.cssText` for initial render as it is cheaper.) The other change is that the styling changes are applied immediately (no queueing). This means that it is the instruction which computes priority. In some circumstances it may result in intermediate writes which are than overwritten with new value. (This should be rare)
Overall this change deletes most of the previous code and replaces it with new simplified implement. The simplification results in code savings.
PR Close#34804
NOTE: This change must be reverted with previous deletes so that it code remains in build-able state.
This change deletes old styling code and replaces it with a simplified styling algorithm.
The mental model for the new algorithm is:
- Create a linked list of styling bindings in the order of priority. All styling bindings ere executed in compiled order and than a linked list of bindings is created in priority order.
- Flush the style bindings at the end of `advance()` instruction. This implies that there are two flush events. One at the end of template `advance` instruction in the template. Second one at the end of `hostBindings` `advance` instruction when processing host bindings (if any).
- Each binding instructions effectively updates the string to represent the string at that location. Because most of the bindings are additive, this is a cheap strategy in most cases. In rare cases the strategy requires removing tokens from the styling up to this point. (We expect that to be rare case)S Because, the bindings are presorted in the order of priority, it is safe to resume the processing of the concatenated string from the last change binding.
PR Close#34616
This change moves information from instructions to declarative position:
- `ɵɵallocHostVars(vars)` => `DirectiveDef.hostVars`
- `ɵɵelementHostAttrs(attrs)` => `DirectiveDef.hostAttrs`
When merging directives it is necessary to know about `hostVars` and `hostAttrs`. Before this change the information was stored in the `hostBindings` function. This was problematic, because in order to get to the information the `hostBindings` would have to be executed. In order for `hostBindings` to be executed the directives would have to be instantiated. This means that the directive instantiation would happen before we had knowledge about the `hostAttrs` and as a result the directive could observe in the constructor that not all of the `hostAttrs` have been applied. This further complicates the runtime as we have to apply `hostAttrs` in parts over many invocations.
`ɵɵallocHostVars` was unnecessarily complicated because it would have to update the `LView` (and Blueprint) while existing directives are already executing. By moving it out of `hostBindings` function we can access it statically and we can create correct `LView` (and Blueprint) in a single pass.
This change only changes how the instructions are generated, but does not change the runtime much. (We cheat by emulating the old behavior by calling `ɵɵallocHostVars` and `ɵɵelementHostAttrs`) Subsequent change will refactor the runtime to take advantage of the static information.
PR Close#34683
This refactorings clearly separates the first and subsequent creation execution
of the `template` instruction. This approach has the following benefits:
- it is clear what happens during the first vs. subsequent executions;
- we can avoid several memory reads and checks after the first creation pass
(there is measurable performance improvement on various benchmarks);
- the template instructions becomes smaller and should become a candidate
for optimisations / inlining faster;
PR Close#33856
The `ngInjectableDef` property was renamed to `ɵprov`, but core must
still support both because there are published libraries that use the
older term.
We are only interested in such properties that are defined directly on
the type being injected, not on base classes. So there is a check that
the defintion is specifically for the given type.
Previously if you tried to inject a class that had `ngInjectableDef` but
also inherited `ɵprov` then the check would fail on the `ɵprov` property
and never even try the `ngInjectableDef` property resulting in a failed
injection.
This commit fixes this by attempting to find each of the properties
independently.
Fixes https://github.com/angular/ngcc-validation/pull/526
PR Close#33732
Most of the use of `document` in the framework is within
the DI so they just inject the `DOCUMENT` token and are done.
Ivy is special because it does not rely upon the DI and must
get hold of the document some other way. There are a limited
number of places relevant to ivy that currently consume a global
document object.
The solution is modelled on the `LOCALE_ID` approach, which has
`getLocaleId()` and `setLocaleId()` top-level functions for ivy (see
`core/src/render3/i18n.ts`). In the rest of Angular (i.e. using DI) the
`LOCALE_ID` token has a provider that also calls setLocaleId() to
ensure that ivy has the same value.
This commit defines `getDocument()` and `setDocument() `top-level
functions for ivy. Wherever ivy needs the global `document`, it calls
`getDocument()` instead. Each of the platforms (e.g. Browser, Server,
WebWorker) have providers for `DOCUMENT`. In each of those providers
they also call `setDocument()` accordingly.
Fixes#33651
PR Close#33712
When debugging `LView`s it is easy to get lost since all of them have
the same name. This change does three things:
1. It makes `TView` have an explicit type:
- `Host`: for the top level `TView` for bootstrap
- `Component`: for the `TView` which represents components template
- `Embedded`: for the `TView` which represents an embedded template
2. It changes the name of `LView` to `LHostView`, `LComponentView`, and
`LEmbeddedView` depending on the `TView` type.
3. For `LComponentView` and `LEmbeddedView` we also append the name of
of the `context` constructor. The result is that we have `LView`s which
are name as: `LComponentView_MyComponent` and `LEmbeddedView_NgIfContext`.
The above changes will make it easier to understand the structure of the
application when debugging.
NOTE: All of these are behind `ngDevMode` and will get removed in
production application.
PR Close#33449
This patch gets rid of the configuration settings present in the
`TStylingContext` array that is used within the styling algorithm
for `[style]`, `[style.prop]`, `[class]` and `[class.name]` bindings.
These configurations now all live inside of the `TNodeFlags`.
PR Close#33540
This patch removes the need to lock the style and class context
instances to track when bindings can be added. What happens now is
that the `tNode.firstUpdatePass` is used to track when bindings are
registered on the context instances.
PR Close#33521
We already store a reference to a native host of a component
view so we can drop the getHostNative utility function (that
was getting the same reference from another data structure).
PR Close#33554
Before this change instantiating multiple directives on the same
host node would result in repeated RNode retrieval and patching.
This commint re-organises code around directive instance creation
so the host node processing (common to all directives) happens
once and only once.
As the additional benefit the directive instantiation logic gets
centralised in one function (at the expense of patching logic
duplication for root node).
PR Close#33322
`bindingIndex` stores the current location of the bindings in the
template function. Because it used to be stored in `LView` that `LView`
was not reentrant. This could happen if a binding was a getter and had
a side-effect of calling `detectChanges()`.
By moving the `bindingIndex` to `LFrame` where all of the global state
is kept in reentrant way we correct the issue.
PR Close#33235
This patch ensures that the `[style]` and `[class]` based bindings
are directly applied to an element's style and className attributes.
This patch optimizes the algorithm so that it...
- Doesn't construct an update an instance of `StylingMapArray` for
`[style]` and `[class]` bindings
- Doesn't apply `[style]` and `[class]` based entries using
`classList` and `style` (direct attributes are used instead)
- Doesn't split or iterate over all string-based tokens in a
string value obtained from a `[class]` binding.
This patch speeds up the following cases:
- `<div [class]>` and `<div class="..." [class]>`
- `<div [style]>` and `<div style="..." [style]>`
The overall speec increase is by over 5x.
PR Close#33336
`LFrame` stores information specifice to the current `LView` As the code
enters and leaves `LView`s we use `enterView()` and `leaveView()`
respectively to build a a stack of `LFrame`s. This allows us to easily
restore the previous `LView` instruction state.
PR Close#33178
Until now, the template type checker has not checked any of the event
bindings that could be present on an element, for example
```
<my-cmp
(changed)="handleChange($event)"
(click)="handleClick($event)"></my-cmp>
```
has two event bindings: the `change` event corresponding with an
`@Output()` on the `my-cmp` component and the `click` DOM event.
This commit adds functionality to the template type checker in order to
type check both kind of event bindings. This means that the correctness
of the bindings expressions, as well as the type of the `$event`
variable will now be taken into account during template type checking.
Resolves FW-1598
PR Close#33125
Turns out that writing to global state is more expensive than writing to
a property on an object.
Slower:
````
let count = 0;
function increment() {
count++;
}
```
Faster:
````
const state = {
count: 0
};
function increment() {
state.count++;
}
```
This change moves all of the instruction state into a single state object.
`noop_change_detection` benchmark
Pre refactoring: 16.7 us
Post refactoring: 14.523 us (-13.3%)
PR Close#33093
Prior to this commit, all `className` inputs were not set because the runtime code assumed that the `classMap` instruction is only generated for `[class]` bindings. However the `[className]` binding also produces the same `classMap`, thus the code needs to distinguish between `class` and `className`. This commit adds extra logic to select the right input name and also throws an error in case `[class]` and `[className]` bindings are used on the same element simultaneously.
PR Close#33188
Injectable defs are not considered public API, so the property
that contains them should be prefixed with Angular's marker
for "private" ('ɵ') to discourage apps from relying on def
APIs directly.
This commit adds the prefix and shortens the name from
ngInjectableDef to "prov" (for "provider", since injector defs
are known as "inj"). This is because property names cannot
be minified by Uglify without turning on property mangling
(which most apps have turned off) and are thus size-sensitive.
PR Close#33151
Prior to this patch, if a map-class binding is applied directly then
that value will be incorrectly provided a sanitizer even if there is no
sanitization present for an element.
PR Close#33154
Directive defs are not considered public API, so the property
that contains them should be prefixed with Angular's marker
for "private" ('ɵ') to discourage apps from relying on def
APIs directly.
This commit adds the prefix and shortens the name from
ngDirectiveDef to dir. This is because property names
cannot be minified by Uglify without turning on property
mangling (which most apps have turned off) and are thus
size-sensitive.
Note that the other "defs" (ngFactoryDef, etc) will be
prefixed and shortened in follow-up PRs, in an attempt to
limit how large and conflict-y this change is.
PR Close#33110
Component defs are not considered public API, so the property
that contains them should be prefixed with Angular's marker
for "private" ('ɵ') to discourage apps from relying on def
APIs directly.
This commit adds the prefix and shortens the name from
`ngComponentDef` to `cmp`. This is because property names
cannot be minified by Uglify without turning on property
mangling (which most apps have turned off) and are thus
size-sensitive.
Note that the other "defs" (ngDirectiveDef, etc) will be
prefixed and shortened in follow-up PRs, in an attempt to
limit how large and conflict-y this change is.
PR Close#33088
Prior to this fix, whenever a style or class binding is present, the
binding application process would require an instance of `TStylingContext`
to be built regardless of whether or not any binding resolution is needed
(just so that it knows whether or not there are any collisions).
This check is, however, unnecessary because if (and only if) there
are directives present on the element then are collisions possible.
This patch removes the need for style/class bindings to register
themselves on to a `TStylingContext` if there are no directives and
present on an element. This means that all map and prop-based
style/class bindings are applied as soon as bindings are updated on
an element.
PR Close#32919
Paul Gschwendtner
Misko Hevery
Andrew Scott
Joey Perrott
Matias Niemelä
Andrew Kushnir
Greg Magolan
Pawel Kozlowski
Pete Bacon Darwin
crisbeto
JoostK
Kara Erickson