Angular View Engine uses global knowledge to compile the following code:
```typescript
export class Base {
constructor(private vcr: ViewContainerRef) {}
}
@Directive({...})
export class Dir extends Base {
// constructor inherited from base
}
```
Here, `Dir` extends `Base` and inherits its constructor. To create a `Dir`
the arguments to this inherited constructor must be obtained via dependency
injection. View Engine is able to generate a correct factory for `Dir` to do
this because via metadata it knows the arguments of `Base`'s constructor,
even if `Base` is declared in a different library.
In Ivy, DI is entirely a runtime concept. Currently `Dir` is compiled with
an ngDirectiveDef field that delegates its factory to `getInheritedFactory`.
This looks for some kind of factory function on `Base`, which comes up
empty. This case looks identical to an inheritance chain with no
constructors, which works today in Ivy.
Both of these cases will now become an error in this commit. If a decorated
class inherits from an undecorated base class, a diagnostic is produced
informing the user of the need to either explicitly declare a constructor or
to decorate the base class.
PR Close#34460
Adds a compilation error if the consumer tries to pass in an undecorated class into the `providers` of an `NgModule`, or the `providers`/`viewProviders` arrays of a `Directive`/`Component`.
PR Close#34460
The function `makeTemplateDiagnostic` was accepting an error code of type
`number`, making it easy to accidentally pass an `ErrorCode` directly and
not convert it to an Angular diagnostic code first.
This commit refactors `makeTemplateDiagnostic` to accept `ErrorCode` up
front, and convert it internally. This is less error-prone.
PR Close#34460
Previously, ngtsc would perform scope analysis (which directives/pipes are
available inside a component's template) and template type-checking of that
template as separate steps. If a component's scope was somehow invalid (e.g.
its NgModule imported something which wasn't another NgModule), the
component was treated as not having a scope. This meant that during template
type-checking, errors would be produced for any invalid expressions/usage of
other components that should have been in the scope.
This commit changes ngtsc to skip template type-checking of a component if
its scope is erroneous (as opposed to not present in the first place). Thus,
users aren't overwhelmed with diagnostic errors for the template and are
only informed of the root cause of the problem: an invalid NgModule scope.
Fixes#33849
PR Close#34460
Previously each NgModule trait checked its own scope for valid declarations
during 'resolve'. This worked, but caused the LocalModuleScopeRegistry to
declare that NgModule scopes were valid even if they contained invalid
declarations.
This commit moves the generation of diagnostic errors to the
LocalModuleScopeRegistry where it belongs. Now the registry can consider an
NgModule's scope to be invalid if it contains invalid declarations.
PR Close#34460
The template type checker generates TypeScript expressions for any
expression that occurs in a template, so that TypeScript can check it
and produce errors. Some expressions as they occur in a template may be
translated into TypeScript code multiple times, for instance a binding
to a directive input that has a template guard.
One example would be the `NgIf` directive, which has a template guard to
narrow the type in the template as appropriate. Given the following
template:
```typescript
@Component({
template: '<div *ngIf="person">{{ person.name }}</div>'
})
class AppComponent {
person?: { name: string };
}
```
A type check block (TCB) with roughly the following structure is
created:
```typescript
function tcb(ctx: AppComponent) {
const t1 = NgIf.ngTypeCtor({ ngIf: ctx.person });
if (ctx.person) {
"" + ctx.person.name;
}
}
```
Notice how the `*ngIf="person"` binding is present twice: once in the
type constructor call and once in the `if` guard. As such, TypeScript
will check both instances and would produce duplicate errors, if any
were found.
Another instance is when the safe navigation operator is used, where an
expression such as `person?.name` is emitted into the TCB as
`person != null ? person!.name : undefined`. As can be seen, the
left-hand side expression `person` occurs twice in the TCB.
This commit adds the ability to insert markers into the TCB that
indicate that any errors within the expression should be ignored. This
is similar to `@ts-ignore`, however it can be applied more granularly.
PR Close#34417
Previously, the type checker would compute an absolute source span by
combining an expression AST node's `ParseSpan` (relative to the start of
the expression) together with the absolute offset of the expression as
represented in a `ParseSourceSpan`, to arrive at a span relative to the
start of the file. This information is now directly available on an
expression AST node in the `AST.sourceSpan` property, which can be used
instead.
PR Close#34417
Now that the source to typings matching is able to handle
aliasing of exports, there is no need to handle aliases in private
declarations analysis.
These were originally added to cope when the typings files had
to use the name that the original source files used when exporting.
PR Close#34254
Previously the identifiers used in the typings files were the same as
those used in the source files.
When the typings files and the source files do not match exactly, e.g.
when one of them is flattened, while the other is a deep tree, it is
possible for identifiers to be renamed.
This commit ensures that the correct identifier is used in typings files
when the typings file does not export the same name as the source file.
Fixes https://github.com/angular/ngcc-validation/pull/608
PR Close#34254
The naïve matching algorithm we previously used to match declarations in
source files to declarations in typings files was based only on the name
of the thing being declared. This did not handle cases where the declared
item had been exported via an alias - a common scenario when one of the two
file sets (source or typings) has been flattened, while the other has not.
The new algorithm tries to overcome this by creating two maps of export
name to declaration (i.e. `Map<string, ts.Declaration>`).
One for the source files and one for the typings files.
It then joins these two together by matching export names, resulting in a
new map that maps source declarations to typings declarations directly
(i.e. `Map<ts.Declaration, ts.Declaration>`).
This new map can handle the declaration names being different between the
source and typings as long as they are ultimately both exported with the
same alias name.
Further more, there is one map for "public exports", i.e. exported via the
root of the source tree (the entry-point), and another map for "private
exports", which are exported from individual files in the source tree but
not necessarily from the root. This second map can be used to "guess"
the mapping between exports in a deep (non-flat) file tree, which can be
used by ngcc to add required private exports to the entry-point.
Fixes#33593
PR Close#34254
In TS we can re-export imports using statements of the form:
```
export * from 'some-import';
```
This is downleveled in UMD to:
```
function factory(exports, someImport) {
function __export(m) {
for (var p in m) if (!exports.hasOwnProperty(p)) exports[p] = m[p];
}
__export(someImport);
}
```
This commit adds support for this.
PR Close#34254
In TS we can re-export imports using statements of the form:
```
export * from 'some-import';
```
This can be downleveled in CommonJS to either:
```
__export(require('some-import'));
```
or
```
var someImport = require('some-import');
__export(someImport);
```
Previously we only supported the first downleveled version.
This commit adds support for the second version.
PR Close#34254
Previously individual properties of the src bundle program were
passed to the reflection host constructors. But going forward,
more properties will be required. To prevent the signature getting
continually larger and more unwieldy, this change just passes the
whole src bundle to the constructor, allowing it to extract what it
needs.
PR Close#34254
When a HTML Ast containing an Attribute node is converted to a Template Ast,
the attribute node might get dropped from the Template Ast path.
This is because the AttrNode is not even in the Template Ast to begin with.
In this case, we manually fix the path by converting the Attribute node
to a AttrAst node and appending it to the path.
This allows the `ExpressionVisitor` to properly visit the leaf node in the
TemplateAst path. We no longer need to visit the `Element` and look for
attributes.
PR Close#34459
Prior to this change, the ExpressionChangedAfterChecked error thrown in Ivy was missing useful information that was available in View Engine, specifically: missing property name for proprty bindings and also the content of the entire property interpolation (only a changed value was displayed) if one of expressions was changed unexpectedly. This commit improves the error message by including the mentioned information into the error text.
PR Close#34381
Previously, bound events were incorrectly bound to directives with
inputs matching the bound event attribute. This fixes that so bound
events can only be bound to directives with matching outputs.
Adds tests for all kinds of directive matching on bound attributes.
PR Close#31938
This is not expected to have any noticeable perf impact, but it wasteful
nonetheless (and annoying when stepping through the code while debugging
`ngtsc`/`ngcc`).
PR Close#34441
This commit adds three previously missing validations to
NgModule.declarations:
1. It checks that declared classes are actually within the current
compilation.
2. It checks that declared classes are directives, components, or pipes.
3. It checks that classes are declared in at most one NgModule.
PR Close#34404
NgModel internally coerces any arbitrary value that will assigned
to the `disabled` `@Input` to a boolean. This has been done to
support the common case where developers set the disabled attribute
without a value. For example:
```html
<input type="checkbox" [(ngModel)]="value" disabled>
```
This worked in View Engine without any errors because inputs were
not strictly checked. In Ivy though, developers can opt-in into
strict template type checking where the attribute would be flagged.
This is because the `NgModel#isDisabled` property type-wise only
accepts a `boolean`. To ensure that the common pattern described
above can still be used, and to reflect the actual runtime behavior,
we should add an acceptance member that makes it work without type
checking errors.
Using a coercion member means that this is not a breaking change.
PR Close#34438
Given the following HTML and cursor position:
```
<div c|></div>
^ cursor is here
```
Note that the cursor is **after** the attribute `c`.
Under the current implementation, only `Element` is included in the
path. Instead, it should be `Element -> Attribute`.
This bug occurs only for cases where the cursor is right after the Node,
and it is because the `end` position of the span is excluded from the search.
Instead, the `end` position should be included.
PR Close#34440
`let` and `of` should be considered reserved keywords in template syntax
and thus should not be part of the autocomplete suggestions.
For reference, TypeScript does not provide such completions.
This commit removes these results and cleans up the code.
PR Close#34434
The ordering matters because we don't currently throw if multiple
configurations are provided (i.e. provider has *both* useExisting and
useFactory). We should actually throw an error in this case, but to
avoid another breaking change in v9, this PR simply aligns the Ivy
behavior with ViewEngine.
PR Close#34433
In the TransitionAnimationEngine we keep track of the existing elements with animations and we clear the cached data when they're removed. We also have some logic where we transition away the child elements when a parent is removed, however in that case we never cleared the cached element data which resulted in a memory leak. The leak is particularly visible in Material where whenever there's an animated overlay with a component inside of it that has an animation, the child component would always be retained in memory.
Fixes#25744.
PR Close#34409
Prior to ivy, undefined values passed in an object to the
ngStyle directive were ignored. Restore this behavior by
ignoring keys that point to undefined values.
closes#34310
PR Close#34422
Expressions in an inline template binding are improperly recorded as
spaning an offset calculated from the start of the template binding
attribute key, whereas they should be calculated from the start of the
attribute value, which contains the actual binding AST.
PR Close#31813
There were some extra examples for `downgradeComponent()` in the upgrade
guide. Added a link to the relevant section of the guide in the
`downgradeComponent()` docs.
Fixes#31584
PR Close#34406
Currently we only run Saucelabs on PRs using the legacy View Engine
build. Switching that build to Ivy is not trivial and there are various
options:
1. Updating the R3 switches to use POST_R3 by default. At first glance,
this doesn't look easy because the current ngtsc switch logic seems to
be unidirectional (only PRE_R3 to POST_R3).
2. Updating the legacy setup to run with Ivy. This sounds like the easiest
solution at first.. but it turns out to be way more complicated. Packages
would need to be built with ngtsc using legacy tools (i.e. first building
the compiler-cli; and then building packages) and View Engine only tests
would need to be determined and filtered out. Basically it will result in
re-auditing all test targets. This is contradictory to the fact that we have
this information in Bazel already.
3. Creating a new job that runs tests on Saucelabs with Bazel. We specify
fine-grained test targets that should run. This would be a good start
(e.g. acceptance tests) and also would mean that we do not continue maintaining
the legacy setup..
This commit implements the third option as it allows us to move forward
with the general Bazel migration. We don't want to spend too much time
on our legacy setup since it will be removed anyway in the future.
PR Close#34277
When we log DI errors we get the name of the provider via `SomeClass.name`. In IE functions that inherit from other functions don't have their own `name`, but they take the `name` from the lowest parent in the chain, before `Function`. I've added some changes to fall back to parsing out the function name from the function's string form.
PR Close#34305
The way definitions are added in JIT mode is through `Object.defineProperty`, but the problem is that in IE10 properties defined through `defineProperty` won't be inherited which means that inheriting injectable definitions no longer works. These changes add a workaround only for JIT mode where we define a fallback method for retrieving the definition. This isn't ideal, but it should only be required until v10 where we'll no longer support inheriting injectable definitions from undecorated classes.
PR Close#34305
In JIT mode we use `__proto__` when reading constructor parameter metadata, however it's not supported on IE10. These changes switch to using `Object.getPrototypeOf` instead.
PR Close#34305
We've got some tests that assert that the generate DOM looks correct. The problem is that IE changes the attribute order in `innerHTML` which caused the tests to fail. I've reworked the relevant tests not to assert directly against `innerHTML`.
PR Close#34305
We have a couple of cases where we use something like `typeof Node === 'function'` to figure out whether we're in a worker context. This works in most browsers, but IE returns `object` instead of `function`. I've updated all the usages to account for it.
PR Close#34305
In `DebugElement.attributes` we return all of the attributes from the underlying DOM node. Most browsers change the attribute names to lower case, but IE preserves the case and since we use camel-cased attributes, the return value was inconsitent. I've changed it to always lower case the attribute names.
PR Close#34305
While sanitizing on browsers that don't support the `template` element (pretty much only IE), we create an inert document and we insert content into it via `document.body.innerHTML = unsafeHTML`. The problem is that IE appears to parse the HTML passed to `innerHTML` differently, depending on whether the element has been inserted into a document or not. In particular, it seems to split some strings into multiple text nodes, which would've otherwise been a single node. This ended up throwing off some of the i18n code down the line and causing a handful of failures. I've worked around it by creating a new inert `body` element into which the HTML would be inserted.
PR Close#34305
A quirk of the Angular template parser is that when parsing templates in the
"default" mode, with options specified by the user, the source mapping
information in the template AST may be inaccurate. As a result, the compiler
parses the template twice: once for "emit" and once to produce an AST with
accurate sourcemaps for diagnostic production.
Previously, only the first parse was performed during analysis. The second
parse occurred during the template type-checking phase, just in time to
produce the template type-checking file.
However, with the reuse of analysis results during incremental builds, it
makes more sense to do the diagnostic parse eagerly during analysis so that
the work isn't unnecessarily repeated in subsequent builds. This commit
refactors the `ComponentDecoratorHandler` to do both parses eagerly, which
actually cleans up some complexity around template parsing as well.
PR Close#34334
Avoids the usage of array destructuring, as it introduces calls to
a `__values` helper function in ES5 that has a relatively high
performance impact. This shaves off roughly 130ms of CPU time for a
large compilation with big templates that uses i18n.
PR Close#34332
The template parser has a certain interpolation config associated with
it and builds a regular expression each time it needs to extract the
interpolations from an input string. Since the interpolation config is
typically the default of `{{` and `}}`, the regular expression doesn't
have to be recreated each time. Therefore, this commit creates only a
single regular expression instance that is used for the default
configuration.
In a large compilation unit with big templates, computing the regular
expression took circa 275ms. This change reduces this to effectively
zero.
PR Close#34332
On a large compilation unit with big templates, the total time spent in
the `PlainCharacterCursor` constructor was 470ms. This commit applies
two optimizations to reduce this time:
1. Avoid the object spread operator within the constructor, as the
generated `__assign` helper in the emitted UMD bundle (ES5) does not
optimize well compared to a hardcoded object literal. This results in a
significant performance improvement. Because of the straight-forward
object literal, the VM is now much better able to optimize the memory
allocations which makes a significant difference as the
`PlainCharacterCursor` constructor is called in tight loops.
2. Reduce the number of `CharacterCursor` clones. Although cloning
itself is now much faster because of the optimization above, several
clone operations were not necessary.
Combined, these changes reduce the total time spent in the
`PlainCharacterCursor` constructor to just 10ms.
PR Close#34332
During TypeScript module resolution, a lot of filesystem requests are
done. This is quite an expensive operation, so a module resolution cache
can be used to speed up the process significantly.
This commit lets the Ivy compiler perform all module resolution with a
module resolution cache. Note that the module resolution behavior can be
changed with a custom compiler host, in which case that custom host
implementation is responsible for caching. In the case of the Angular
CLI a custom compiler host with proper module resolution caching is
already in place, so the CLI already has this optimization.
PR Close#34332
The export scope of NgModules from external compilations units, as
present in .d.ts declarations, does not change during a compilation so
can be easily shared. There was already a cache but the computed export
scope was not actually stored there. This commit fixes that.
PR Close#34332
crisbeto
Alex Rickabaugh
JoostK
Pete Bacon Darwin
Martin Probst
Keen Yee Liau
Keen Yee Liau
Andrew Kushnir
Ayaz Hafiz
George Kalpakas
Paul Gschwendtner
Andrew Scott
Michael Nahkies
ajitsinghkaler
Pawel Kozlowski