This is the compile-time implementation of the `hostDirectives` feature plus a little bit of runtime code to illustrate how the newly-generated code will plug into the runtime. It works by creating a call to the new `ɵɵHostDirectivesFeature` feature whenever a directive has a `hostDirectives` field. Afterwards `ɵɵHostDirectivesFeature` will patch a new function onto the directive definition that will be invoked during directive matching.
For example, if we take the following definition:
```ts
@Directive({
hostDirectives: [HostA, {directive: HostB, inputs: ['input: alias']}]
})
class MyDir {}
```
Will compile to:
```js
MyDir.ɵdir = ɵɵdefineComponent({
features: [ɵɵHostDirectivesFeature([HostA, {
directive: HostB,
inputs: {
input: "alias"
}
}])]
});
```
The template type checking is implemented during directive matching by adding the host directives applied on the host to the array of matched directives whenever the host is matched in a template.
Relates to #8785.
PR Close#46868
Saves us some bytes by not emitting `providers` in `defineInjector`. While the amount of bytes isn't huge, I think that this change is worthwhile, because `ng generate` currently generates `providers: []` with every `NgModule` which users can forget to remove.
PR Close#46301
This commit accounts for the Babel types changes. Some properties
can now also be `undefined` so existing checks/assertions had to
be adjusted to also capture `undefined` (along with `null`).
Additionally, in preparation for a new ECMA proposal, Babel types
seem to have been updated to include private names in object property
keys. This is not necessarily the case for object expressions, but
could be for object patterns (in the future -- when implemented).
More details: https://github.com/babel/babel/pull/14304 and
https://github.com/tc39/proposal-destructuring-private.
PR Close#45967
Excludes styles that resolve to empty strings from the emitted metadata so that they don't result in empty `<style>` tags at runtime.
Fixes#31191.
PR Close#45459
In v14, the partial compilation output of components have changed in a way
that prevents older versions of Angular to compile the partial declarations
correctly.
In particular, we used to emit used directives/components in separate arrays called
`components` and `directives`, and used pipes in a property called `pipes`:
```js
TestComponent.ɵcmp = i0.ɵɵngDeclareComponent({
minVersion: "12.0.0",
version: "13.3.0",
type: TestComponent,
selector: "ng-component",
ngImport: i0,
template: ``,
isInline: true,
directives: [{ type: i1.SomeDir, selector: "[some-dir]" }],
components: [{ type: i1.SomeCmp, selector: "some-cmp" }],
pipes: { 'async': i2.AsyncPipe },
});
```
In the above example, the `version` property indicates which exact compiler
version was used to compile the component, but the `minVersion` allows older
versions of the compiler/Angular linker to "link" the partial declaration to
its final AOT compilation output.
In v14, the used directives, components and pipes are now emitted together
into a single array under the `dependencies` property:
```js
TestComponent.ɵcmp = i0.ɵɵngDeclareComponent({
minVersion: "12.0.0",
version: "13.3.0",
type: TestComponent,
selector: "ng-component",
ngImport: i0,
template: ``,
isInline: true,
dependencies: [
{ kind: "directive", type: i1.SomeDir, selector: "[some-dir]" },
{ kind: "component", type: i1.SomeCmp, selector: "some-cmp" },
{ kind: "pipe", type: i2.AsyncPipe },
],
});
```
This change has been made in support of standalone components, but it does mean
that older compiler versions can no longer link these partial declarations
as desirable as none of the components, directives and pipes would be included
in the AOT-compiled code.
By increasing the `minVersion` property, we hint to older compiler versions that
they are not capable of processing the partial declaration. This allows reporting
an error at compile time instead of resulting in runtime failures due to missing
components, directives and pipes.
PR Close#45782
Before standalone, everything that could appear in an NgModule's `imports`
was relevant to DI, and needed to be emitted in the `imports` of the
generated `InjectorDef` definition. With the introduction of standalone
types, NgModule `imports` can now contain components, directives, and pipes
which are standalone. Only standalone components need to be included in
the `imports` of the generated injector definition - directives and pipes
have no effect on DI. Having them present doesn't cause any errors in the
runtime (they're filtered out by the injector itself) but it does prevent
tree-shaking.
With this commit, the generation of `InjectorDef` now filters the `imports`
to exclude directives and pipes as much as possible. It's not _always_
possible because an expression in `imports` may pull in both a directive and
a `ModuleWithProviders` reference, and we have no way of referencing just
the MWP part of that expression. Therefore this is an optimization, not a
rule of `InjectorDef` compilation.
PR Close#45701
Previously, the NgModule handler would resolve the `imports` field as one
unit, producing an array of `Reference`s. With this refactoring, if
`imports` is a literal array, each individual element will be resolved
independently. This will allow filtering in the future at the element level,
since there will be a separate `ts.Expression` for each individual element.
PR Close#45701
This commit bundles tests for standalone components that are possible after
previous implementation commits. Most new tests are compliance tests, but
a test is also included to validate that the template type-checking system
can work with standalone components as well.
PR Close#45672
This commit propagates the `isStandalone` flag for a component, directive,
or pipe during partial compilation of a standalone declaration. This flag
allows the linker to properly process a standalone declaration that it
encounters.
PR Close#45672
Previously, the compiler tracked directives and pipes in template scopes
separately. This commit refactors the scope system to unify them into a
single data structure, disambiguated by a `kind` field.
PR Close#45672
Previously, the compiler would represent template dependencies of a
component in its component definition through separate fields (`directives`,
`pipes`).
This commit refactors the compiler/runtime interface to use a single field
(`dependencies`). The runtime component definition object still has separate
`directiveDefs` and `pipeDefs`, which are calculated from the `dependencies`
when the definition is evaluated.
This change is also reflected in partially compiled declarations. To ensure
compatibility with partially compiled code already on NPM, the linker
will still honor the old form of declaration (with separate fields).
PR Close#45672
This links back each placeholder in a message to the original Angular template span which defines its expression. This is useful for understanding where each placeholder comes from in the context of the full message.
PR Close#45606
Recent changes in `rules_nodejs` caused the test case copy file actions
to be transitioned into the `exec` configuration, resulting in much larger
file paths. These paths break on Windows with the shell argument limit, and
with the path limit, causing errors like:
```
ERROR: C:/users/circleci/ng/packages/compiler-cli/test/compliance/test_cases/BUILD.bazel:9:12: Copying file packages/compiler-cli/test/compliance/test_cases/r3_compiler_compliance/components_and_directives/value_composition/structural_directives_if_directive_def.js failed: (Exit 1): cmd.exe failed: error executing command
cd /d C:/users/circleci/_bazel_circleci/u4uoan2j/execroot/angular
SET PATH=C:\Program Files\Git\usr\bin;C:\Program Files\Git\bin;C:\Windows;C:\Windows\System32;C:\Windows\System32\WindowsPowerShell\v1.0
SET RUNFILES_MANIFEST_ONLY=1
cmd.exe /C bazel-out\x64_windows-opt-exec-2B5CBBC6\bin\packages\compiler-cli\test\compliance\test_cases\test_cases--1973427149-cmd.bat
The system cannot find the path specified
```
https://app.circleci.com/pipelines/github/angular/angular/44038/workflows/4b530cb2-f232-4e1d-b35a-e6e085151d08/jobs/1140017
PR Close#45431
When we have an event listener inside an embedded view, we generate a `restoreView` call which saves the view inside of the LFrame. The problem is that we don't clear it until it gets overwritten which can lead to memory leaks.
These changes rework the generated code in order to generate a `resetView` call which will clear the view from the LFrame.
Fixes#42848.
PR Close#43075
Angular contains an NgModule registry, which allows a user to declare
NgModules with string ids and retrieve them via those ids, using the
`getNgModuleById` API.
Previously, we attempted to structure this registration in a clever fashion
to allow for tree-shaking of registered NgModules (that is, those with ids).
This sort of worked due to the accidental alignment of behaviors from the
different tree-shakers involved. However, this trick relies on the
generation of `.ngfactory` files and how they're specifically processed in
various bundling scenarios. We intend to remove `.ngfactory` files, hence
we can no longer rely on them in this way.
The correct solution here is to recognize that `@NgModule({id})` is
inherently declaring a global side-effect, and such classes should not
really be eligible for tree-shaking in the first place. This commit removes
all the old registration machinery, and standardizes on generating a side-
effectful call to `registerNgModuleType` for NgModules that have ids.
There is some risk here that NgModules with unnecessary `id`s may not
tree-shake as a result of this change, whereas they would have in previous
circumstances. The fix here should be to remove the `id` if it's not needed.
Specifying an `id` is a request that the NgModule be retained regardless of
any other references, in case it is later looked up by string id.
PR Close#45024
The `@angular/localize` package depends on a version of Babel that is two years
old, so this commit updates to the latest version.
Some changes were made to the linker and compliance tests to account for slight
changes in source maps, along with a few code updates because of changes to
the typings of Babel.
PR Close#44931
In templates with several levels of nested nodes, it's common for several `elementStart`/`elementEnd` instructions to show up in a row which can be optimized away.
These changes add chaining support for `elementStart`, `elementEnd`, `elementContainerStart` and `elementContainerEnd` to shave off some bytes when possible.
PR Close#44994
Refs http://b/214103351.
This happens if a user writes `<span i18n>Message</span>`. This is accepted as an internationalized message, but without a description. JSCompiler will throw an error in this situation because descriptions are generally required. Now, the Angular compiler will generate a suppression annotation so JSCompiler allows the syntax. This will ease an internal migration to JSCompiler-based i18n.
PR Close#44787
An `ng-template` with an inline template (i.e. has a structural
directive) would previously not get an `undefined` `tagName` because the
logic assumed the element would be `t.Element` or `t.Content` and read
the tag name from the `name` property. For a `t.Template`, this exists
instead on the `t.tagName`. The final result would be an `tagName` of `undefined`
for the parent `t.Template`, causing failures in the indexer downstream.
This `undefined` value is actually expected in the renderer code, even
though the type does not specify this possibility. This change updates
the type of `tagName` to be `string|null` and explicitly handles the
case where there is a structural directive on an `ng-template`. You can
see how the two are differentiated in the compliance code that was
modified in this commit.
PR Close#44788
When a partially compiled component or directive is "linked" in JIT mode, the body
of its declaration is evaluated by the JavaScript runtime. If a class is referenced
in a query (e.g. `ViewQuery` or `ContentQuery`) but its definition is later in the
file, then the reference must be wrapped in a `forwardRef()` call.
Previously, query predicates were not wrapped correctly in partial declarations
causing the code to crash at runtime. In AOT mode, this code is never evaluated
but instead transformed as part of the build, so this bug did not become apparent
until Angular Material started running JIT mode tests on its distributable output.
This change fixes this problem by noting when queries are wrapped in `forwardRef()`
calls and ensuring that this gets passed through to partial compilation declarations
and then suitably stripped during linking.
See https://github.com/angular/components/pull/23882 and https://github.com/angular/components/issues/23907
PR Close#44113
When a safe method call such as `person?.getName()` is used, the
compiler would generate invalid code if the argument list also contained
a safe method call. For example, the following code:
```
person?.getName(config?.get('title').enabled)
```
would generate
```
let tmp;
ctx.person == null ? null : ctx.person.getName((tmp = tmp) == null ?
null : tmp.enabled)
```
Notice how the call to `config.get('title')` has completely disappeared,
with `(tmp = tmp)` having taken its place.
The issue occurred due to how the argument list would be converted
from expression AST to output AST twice. First, the outer safe method
call would first convert its arguments list. This resulted in a
temporary being allocated for `config.get('title')`, which was stored in
the internal `_resultMap`. Only after the argument list has been
converted would the outer safe method call realize that it should be
guarded by a safe access of `person`, entering the `convertSafeAccess`
procedure to convert itself. This would convert the argument list once
again, but this time the `_resultMap` would already contain the
temporary `tmp` for `config?.get('title')`. Consequently, the safe
method in the argument list would be emitted as `tmp`.
This commit fixes the issue by ensuring that nodes are only converted
once.
Closes#44069
PR Close#44088
This reverts commit bba0a87055.
The reason for rollback: this change is breaking some targets in Google's codebase when there is no attribute value is displayed (attr.aria-label) when translated.
PR Close#43882
While fully dynamic bound properties (and attributes) cannot be marked for localization, properties that only contain interpolation can.
This commit ensure that attribute bindings that only contain interpolation can also be marked for localization.
Closes#43260
PR Close#43815
When extracting i18n messages from templates, ICU messages are split out from the
message that contains them. This can make it difficult in the translation files to match up
the two messages, especially if the ICU is reused in multiple placeholders.
This commit builds on top of the previous one to expose the message ID of ICU messages
from the ICU placeholders as additional metadata in the `$localize` tagged strings.
Now the metablock following any placeholder can also contain the associated ID
delimited from the placeholder name by `@@`.
Fixes#17506
PR Close#43534
When specifying the `deps` array in the `@Injectable` decorator to
inject dependencies into the injectable's factory function, it should
be possible to use an array literal to configure how the dependency
should be resolved by the DI system.
For example, the following example is allowed:
```ts
@Injectable({
providedIn: 'root',
useFactory: a => new AppService(a),
deps: [[new Optional(), 'a']],
})
export class AppService {
constructor(a) {}
}
```
Here, the `'a'` string token should be injected as optional. However,
the AOT compiler incorrectly used the array literal itself as injection
token, resulting in a failure at runtime. Only if the token were to be
provided using `[new Optional(), new Inject('a')]` would it work
correctly.
This commit fixes the issue by using the last non-decorator in the
array literal as the token value, instead of the array literal itself.
Note that this is a loose interpretation of array literals: if a token
is omitted from the array literal then the array literal itself is used
as token, but any decorator such as `new Optional()` would still have
been applied. When there's multiple tokens in the list then only the
last one will be used as actual token, any prior tokens are silently
ignored. This behavior mirrors the JIT interpretation so is kept as is
for now, but may benefit from some stricter checking and better error
reporting in the future.
Fixes#42987
PR Close#43226
Previously, the way templates were tokenized meant that we lost information
about the location of interpolations if the template contained encoded HTML
entities. This meant that the mapping back to the source interpolated strings
could be offset incorrectly.
Also, the source-span assigned to an i18n message did not include leading
whitespace. This confused the output source-mappings so that the first text
nodes of the message stopped at the first non-whitespace character.
This commit makes use of the previous refactorings, where more fine grain
information was provided in text tokens, to enable the parser to identify
the location of the interpolations in the original source more accurately.
Fixes#41034
PR Close#43132
The lexer now splits encoded entity tokens out from text and attribute value tokens.
Previously encoded entities would be decoded and the decoded value would be
included as part of the text token of the surrounding text. Now the entities
have their own tokens. There are two scenarios: text and attribute values.
Previously the contents of `<div>Hello & goodbye</div>` would be a single
TEXT token. Now it will be three tokens:
```
TEXT: "Hello "
ENCODED_ENTITY: "&", "&"
TEXT: " goodbye"
```
Previously the attribute value in `<div title="Hello & goodbye">` would be
a single text token. Now it will be three tokens:
```
ATTR_VALUE_TEXT: "Hello "
ENCODED_ENTITY: "&", "&"
ATTR_VALUE_TEXT: " goodbye"
```
- ENCODED_ENTITY tokens have two parts: "decoded" and "encoded".
- ENCODED_ENTITY tokens are always preceded and followed by either TEXT tokens
or ATTR_VALUE_TEXT tokens, depending upon the context, even if they represent
an empty string.
The HTML parser has been modified to recombine these tokens to allow this
refactoring to have limited effect in this commit. Further refactorings
to use these new tokens will follow in subsequent commits.
PR Close#43132
Previously with View Engine output, the `enableResourceInlining` option
could be set to inline external templates and styles (also for the
resulting `.metadata.json` files). We want to do the same for the Ivy
compilation pipeline (regardless of the compilation mode). The full
compilation definitions, and partial declarations currently already
inline resources in a way that no external requests need to be made.
Although there is one exception currently. These are the calls for
setting class metadata (for testbed overrides). This commit updates
the set class metadata calls (for both partial and full compilation)
to always inline resources. This means that libraries do not need
to start shipping external styles/templates just for the
`setClassMetadata` calls.
Note: Only doing this for partial compilation has been considered, but
it seems like it would be simpler implementation-wise to do this for
full compilation as well. Given the external resources are already
inlined (through their `ecmp` definitions), it seems acceptable (or
even more aligned) to do the same for the set class metadata calls.
PR Close#43178
The compliance tests can check source-map segments against expectations
encoded into the expectation files. Previously, the encoding of the expected
segment was only delimited by whitespace, but this made it difficult to identify
segments that started or ended with whitespace.
Now these segment expectations are wrapped in double-quotes which makes
it easier to read and understand the expectation files.
PR Close#43129
Previously, the way templates were tokenized meant that we lost information
about the location of interpolations if the template contained encoded HTML
entities. This meant that the mapping back to the source interpolated strings
could be offset incorrectly.
Also, the source-span assigned to an i18n message did not include leading
whitespace. This confused the output source-mappings so that the first text
nodes of the message stopped at the first non-whitespace character.
This commit makes use of the previous refactorings, where more fine grain
information was provided in text tokens, to enable the parser to identify
the location of the interpolations in the original source more accurately.
Fixes#41034
PR Close#42062
The compliance tests can check source-map segments against expectations
encoded into the expectation files. Previously, the encoding of the expected
segment was only delimited by whitespace, but this made it difficult to identify
segments that started or ended with whitespace.
Now these segment expectations are wrapped in double-quotes which makes
it easier to read and understand the expectation files.
PR Close#42062
Currently unless a listener inside of an embedded view tries to reference something from the parent view, or if the reference is a local ref, we don't generate the view restoration instructions and we allow for the value to be picked up from the context object in the function parameters. The problem is that the listener is only run during creation mode and the context object may have been swapped out afterwards.
These changes fix the issue by always generating the view restoration instructions for listeners inside templates.
Fixes#42698.
PR Close#42755
As of ES2021, JavaScript allows using underscores as separators inside numbers, in order to make them more readable (e.g. `1_000_000` vs `1000000`). TypeScript has had support for separators for a while so these changes expand the template parser to handle them as well.
PR Close#42672
Adds support for shorthand property declarations inside Angular templates. E.g. doing `{foo, bar}` instead of `{foo: foo, bar: bar}`.
Fixes#10277.
PR Close#42421
If an implcit receiver is accessed in a listener inside of an `ng-template`, we generate some extra code in order to ensure that we're assigning to the correct object. The problem is that the logic wasn't covering keyed writes which caused it to write to the wrong object and throw an assertion error at runtime.
These changes expand the logic to cover keyed writes.
Fixes#41267.
PR Close#42603
Updates the compiler-cli compliance goldens. The golden updates are
required due to a regression in TypeScript 4.3 that causes the emitter
to not incorrectly preserve lines when emitting a node list.
This can be reverted once https://github.com/microsoft/TypeScript/pull/44070
is available.
PR Close#42022
Currently we support safe property (`a?.b`) and method (`a?.b()`) accesses, but we don't handle safe keyed reads (`a?.[0]`) which is inconsistent. These changes expand the compiler in order to support safe key read expressions as well.
PR Close#41911
When `checkTypeOfPipes` is set to `false`, our TCB currently generates
the a statement like the following when pipes appear in the template:
`(_pipe1 as any).transform(args)`
This did enable us to get _some_ information from the Language Service
about pipes in this case because we still had access to the pipe
instance. However, because it is immediately cast to `any`, we cannot
get type information about the transform access. That means actions like "go to
definition", "find references", "quick info", etc. will return
incomplete information or fail altogether.
Instead, this commit changes the TCB to generate `(_pipe1.transform as any)(args)`.
This gives us the ability to get complete information for the LS
operations listed above.
PR Close#40523
This commit changes the reference emitters in the Ivy compiler to prefer
non-aliased exports if they exist. This avoids selecting "private
exports" that may not be stable, e.g. the reexports that have been added
by the View Engine compiler. Such reexports are not stable and are
therefore not suitable to be emitted into partial compilations, as the
output of partial compilations should only reference stable symbols
from upstream libraries.
An alternative solution has been considered where ViewEngine-generated
exports would gain a certain prefix, such that the Ivy compiler could
just exclude those exports (see #41443). However, that solution would
be insufficient in case a library is built using partial compilation and
while depending itself on a VE-compiled library from earlier versions of
Angular, where the magic prefix would be missing. For such libraries,
ngcc would have generated reexports using the declared name if not already
present so this change does result in choosing the correct export.
Because ngcc always generates reexports using the declared name even if
an aliased export is present, this change causes those ngcc-generated
exports to be chosen in downstream libraries using partial compilation.
This is unfortunate as it means that the declared names become
effectively public even if the library author was intentionally
exporting it using an alias. This commit does not address this problem;
it is expected that this should not result in widespread issues across
the library ecosystem.
Fixes#41277
PR Close#41866
We have a check that determines whether to generate property binding instructions for an `ng-template`. The check looks at whether the tag name is exactly `ng-template`, but the problem is that if the tag is placed in a non-HTML namespace (e.g. `svg`), the tag name will actually be `:namespace:ng-template` and the check will fail.
These changes resolve the issue by looking at the tag name without the namespace.
Fixes#41308.
PR Close#41669
Currently if a component defines a template inline, but not through a
string literal, the partial compilation references the template expression
as is. This is problematic because the component declaration can no longer
be processed by the linker later as there is no static interpretation. e.g.
```js
const myTemplate = `...`;
TestCmp.ɵcmp = i0.ɵɵngDeclareComponent({
version: "0.0.0-PLACEHOLDER",
type: TestCmp,
selector: "test-cmp",
ngImport: i0,
template: myTemplate,
isInline: true
});
```
To fix this, we use the the resolved template in such cases so that
the linker can process the template/component declaration as expected.
PR Close#41583
This is follow-up from #41437 and it reduces the amount of code we generate for safe property accesses (`a?.b`) and nullish coalescing (`a ?? b`) by:
1. Reusing variables in nested nullish coalescing expressions.
2. Not initializing temporary variables to `null`. The way our code is generated means that the value will always be overwritten before we compare against it so the initializer didn't really matter.
Fixes#41491.
PR Close#41563
Previously, it was not possible to block a partial-linker from trying to
process a declaration that was defined in a newer version of Angular than
that of the partial-linker. For example, if a partial-linker was published as
part of version 12.0.0, there was no way for a partially-compiled declaration
compiled via version 13.0.0 to tell the 12.0.0 linker that it would be invalid
to attempt to process it.
This commit adds a new `minVersion` property to partial-declarations, which is
interpreted as the "minimum partial-linker version" that can process this
declaration. When selecting a partial-linker for such a declaration, the known
linker version ranges are checked to find the most recent linker whose version
range has an overlap with the interpreted declaration range.
This approach allows us to set a minimum version for a declaration, which
can inform an old partial-linker that will it not be able to accurately
process the declaration.
Note that any pre-release part to versions are ignored in this selection
process.
The file-linker can be configured, via the `unknownDeclarationVersionHandling`
property of `LinkerOptions`, to handle such a situation in one of three ways:
- `error` - the version mismatch is a fatal error
- `warn` - a warning is sent to the logger but the most recent partial-linker
will attempt to process the declaration anyway.
- `ignore` - the most recent partial-linker will, silently, attempt to process
the declaration.
The default is to throw an error.
Closes#41497
PR Close#41578
There were three options being made available to users of the linker:
- ` enableI18nLegacyMessageIdFormat`
- `i18nNormalizeLineEndingsInICUs`
- ` i18nUseExternalIds`
None of these should actually be configurable at linking time
because partially-linked libraries have tighter restrictions on
what i18n options can be used.
This commit removes those options from the `LinkerOptions` interface.
It was considered to add a check for backwards compatibilty to ensure
that if these options were being passed, and were different to the expected
defaults, we would throw an informative error. But from looking at the
Angular CLI (the only known client of the linker) it has never been setting
these options so they have already always been set to the defaults.
BREAKING CHANGE:
Linked libraries no longer generate legacy i18n message ids. Any downstream
application that provides translations for these messages, will need to
migrate their message ids using the `localize-migrate` command line tool.
Closes#40673
PR Close#41554
This commit refactors the generated code for class metadata in partial
compilation mode. Instead of emitting class metadata into a top-level
`ɵsetClassMetadata` call guarded by `ngDevMode` flags, the class
metadata is now declared using a top-level `ɵɵngDeclareClassMetadata`
call.
PR Close#41200
