During AOT compilation, the value of some expressions need to be known at
compile time. The compiler has the ability to statically evaluate expressions
the best it can, but there can be occurrences when an expression cannot be
evaluated statically. For instance, the evaluation could depend on a dynamic
value or syntax is used that the compiler does not understand. Alternatively,
it is possible that an expression could be statically evaluated but the
resulting value would be of an incorrect type.
In these situations, it would be helpful if the compiler could explain why it
is unable to evaluate an expression. To this extend, the static interpreter
in Ivy keeps track of a trail of `DynamicValue`s which follow the path of nodes
that were considered all the way to the node that causes an expression to be
considered dynamic. Up until this commit, this rich trail of information was
not surfaced to a developer so the compiler was of little help to explain
why static evaluation failed, resulting in situations that are hard to debug
and resolve.
This commit adds much more insight to the diagnostic that is produced for static
evaluation errors. For dynamic values, the trail of `DynamicValue` instances
is presented to the user in a meaningful way. If a value is available but not
of the correct type, the type of the resolved value is shown.
Resolves FW-2155
PR Close#37587
Commit 4213e8d5 introduced shim reference tagging into the compiler, and
changed how the `TypeCheckProgramHost` worked under the hood during the
creation of a template type-checking program. This work enabled a more
incremental flow for template type-checking, but unintentionally introduced
several regressions in performance, caused by poor incrementality during
`ts.Program` creation.
1. The `TypeCheckProgramHost` was made to rely on the `ts.CompilerHost` to
retrieve instances of `ts.SourceFile`s from the original program. If the
host does not return the original instance of such files, but instead
creates new instances, this has two negative effects: it incurs
additional parsing time, and it interferes with TypeScript's ability to
reuse information about such files.
2. During the incremental creation of a `ts.Program`, TypeScript compares
the `referencedFiles` of `ts.SourceFile` instances from the old program
with those in the new program. If these arrays differ, TypeScript cannot
fully reuse the old program. The implementation of reference tagging
introduced in 4213e8d5 restores the original `referencedFiles` array
after a `ts.Program` is created, which means that future incremental
operations involving that program will always fail this comparison,
effectively limiting the incrementality TypeScript can achieve.
Problem 1 exacerbates problem 2: if a new `ts.SourceFile` is created by the
host after shim generation has been disabled, it will have an untagged
`referencedFiles` array even if the original file's `referencedFiles` was
not restored, triggering problem 2 when creating the template type-checking
program.
To fix these issues, `referencedFiles` arrays are now restored on the old
`ts.Program` prior to the creation of a new incremental program. This allows
TypeScript to get the most out of reusing the old program's data.
Additionally, the `TypeCheckProgramHost` now uses the original `ts.Program`
to retrieve original instances of `ts.SourceFile`s where possible,
preventing issues when a host would otherwise return fresh instances.
Together, these fixes ensure that program reuse is as incremental as
possible, and tests have been added to verify this for certain scenarios.
An optimization was further added to prevent the creation of a type-checking
`ts.Program` in the first place if no type-checking is necessary.
PR Close#37641
We recently added a transformer to NGC that is responsible for downleveling Angular
decorators and constructor parameter types. The primary goal was to mitigate a
TypeScript limitation/issue that surfaces in Angular projects due to the heavy
reliance on type metadata being captured for DI. Additionally this is a pre-requisite
of making `tsickle` optional in the Angular bazel toolchain.
See: 401ef71ae5 for more context on this.
Another (less important) goal was to make sure that the CLI can re-use
this transformer for its JIT mode compilation. The CLI (as outlined in
the commit mentioned above), already has a transformer for downleveling
constructor parameters. We want to avoid this duplication and exported
the transform through the tooling-private compiler entry-point.
Early experiments in using this transformer over the current one, highlighted
that in JIT, class decorators cannot be downleveled. Angular relies on those
to be invoked immediately for JIT (so that factories etc. are generated upon loading)
The transformer we exposed, always downlevels such class decorators
though, so that would break CLI's JIT mode. We can address the CLI's
needs by adding another flag to skip class decorators. This will allow
us to continue with the goal of de-duplication.
PR Close#37545
Previously, when an entry-point was ignored via an ngcc config, ngcc
would scan sub-directories for sub-entry-points, but would not use the
correct `packagePath`. For example, if `@angular/common` was ignored, it
would look at `@angular/common/http` but incorrectly use
`.../@angular/common/http` as the `packagePath` (instead of
`.../@angular/common`). As a result, it would not retrieve the correct
ngcc config for the actual package.
This commit fixes it by ensuring the correct `packagePath` is used, even
if the primary entry-point corresponding to that path is ignored. In
order to do this, a new return value for `getEntryPointInfo()` is added:
`IGNORED_ENTRY_POINT`. This is used to differentiate between directories
that correspond to no or an incompatible entry-point and those that
correspond to an entry-point that could otherwise be valid but is
explicitly ignored. Consumers of `getEntryPointInfo()` can then use this
info to discard ignored entry-points, but still use the correct
`packagePath` when scanning their sub-directories for secondary
entry-points.
PR Close#37040
Currently Angular internally already handles `InjectionToken` as
predicates for queries. This commit exposes this as public API as
developers already relied on this functionality but currently use
workarounds to satisfy the type constraints (e.g. `as any`).
We intend to make this public as it's low-effort to support, and
it's a significant key part for the use of light-weight tokens as
described in the upcoming guide: https://github.com/angular/angular/pull/36144.
In concrete, applications might use injection tokens over classes
for both optional DI and queries, because otherwise such references
cause classes to be always retained. This was also an issue in View
Engine, but now with Ivy, this pattern became worse, as factories are
directly attached to retained classes (ultimately ending up in the
production bundle, while being unused).
More details in the light-weight token guide and in: https://github.com/angular/angular-cli/issues/16866.
Closes#21152. Related to #36144.
PR Close#37506
In v7 of Angular we removed `tsickle` from the default `ngc` pipeline.
This had the negative potential of breaking ES2015 output and SSR due
to a limitation in TypeScript.
TypeScript by default preserves type information for decorated constructor
parameters when `emitDecoratorMetadata` is enabled. For example,
consider this snippet below:
```
@Directive()
export class MyDirective {
constructor(button: MyButton) {}
}
export class MyButton {}
```
TypeScript would generate metadata for the `MyDirective` class it has
a decorator applied. This metadata would be needed in JIT mode, or
for libraries that provide `MyDirective` through NPM. The metadata would
look as followed:
```
let MyDirective = class MyDir {}
MyDirective = __decorate([
Directive(),
__metadata("design:paramtypes", [MyButton]),
], MyDirective);
let MyButton = class MyButton {}
```
Notice that TypeScript generated calls to `__decorate` and
`__metadata`. These calls are needed so that the Angular compiler
is able to determine whether `MyDirective` is actually an directive,
and what types are needed for dependency injection.
The limitation surfaces in this concrete example because `MyButton`
is declared after the `__metadata(..)` call, while `__metadata`
actually directly references `MyButton`. This is illegal though because
`MyButton` has not been declared at this point. This is due to the
so-called temporal dead zone in JavaScript. Errors like followed will
be reported at runtime when such file/code evaluates:
```
Uncaught ReferenceError: Cannot access 'MyButton' before initialization
```
As noted, this is a TypeScript limitation because ideally TypeScript
shouldn't evaluate `__metadata`/reference `MyButton` immediately.
Instead, it should defer the reference until `MyButton` is actually
declared. This limitation will not be fixed by the TypeScript team
though because it's a limitation as per current design and they will
only revisit this once the tc39 decorator proposal is finalized
(currently stage-2 at time of writing).
Given this wontfix on the TypeScript side, and our heavy reliance on
this metadata in libraries (and for JIT mode), we intend to fix this
from within the Angular compiler by downleveling decorators to static
properties that don't need to evaluate directly. For example:
```
MyDirective.ctorParameters = () => [MyButton];
```
With this snippet above, `MyButton` is not referenced directly. Only
lazily when the Angular runtime needs it. This mitigates the temporal
dead zone issue caused by a limitation in TypeScript's decorator
metadata output. See: https://github.com/microsoft/TypeScript/issues/27519.
In the past (as noted; before version 7), the Angular compiler by
default used tsickle that already performed this transformation. We
moved the transformation to the CLI for JIT and `ng-packager`, but now
we realize that we can move this all to a single place in the compiler
so that standalone ngc consumers can benefit too, and that we can
disable tsickle in our Bazel `ngc-wrapped` pipeline (that currently
still relies on tsickle to perform this decorator processing).
This transformation also has another positive side-effect of making
Angular application/library code more compatible with server-side
rendering. In principle, TypeScript would also preserve type information
for decorated class members (similar to how it did that for constructor
parameters) at runtime. This becomes an issue when your application
relies on native DOM globals for decorated class member types. e.g.
```
@Input() panelElement: HTMLElement;
```
Your application code would then reference `HTMLElement` directly
whenever the source file is loaded in NodeJS for SSR. `HTMLElement`
does not exist on the server though, so that will become an invalid
reference. One could work around this by providing global mocks for
these DOM symbols, but that doesn't match up with other places where
dependency injection is used for mocking DOM/browser specific symbols.
More context in this issue: #30586. The TL;DR here is that the Angular
compiler does not care about types for these class members, so it won't
ever reference `HTMLElement` at runtime.
Fixes#30106. Fixes#30586. Fixes#30141.
Resolves FW-2196. Resolves FW-2199.
PR Close#37382
When the compiler encounters a function call within an NgModule imports
section, it attempts to resolve it to an NgModule-annotated class by
looking at the function body and evaluating the statements there. This
evaluation can only understand simple functions which have a single
return statement as their body. If the function the user writes is more
complex than that, the compiler won't be able to understand it and
previously the PartialEvaluator would return a "DynamicValue" for
that import.
With this change, in the event the function body resolution fails the
PartialEvaluator will now attempt to use its foreign function resolvers to
determine the correct result from the function's type signtaure instead. If
the function is annotated with a correct ModuleWithProviders type, the
compiler will be able to understand the import without static analysis of
the function body.
PR Close#37126
In some versions of TypeScript, the transformation of synthetic
`$localize` tagged template literals is broken.
See https://github.com/microsoft/TypeScript/issues/38485
We now compute what the expected final output target of the
compilation will be so that we can generate ES5 compliant
`$localize` calls instead of relying upon TS to do the downleveling
for us.
This is a workaround for the TS compiler bug, which could be removed
when this is fixed. But since it only affects ES5 targeted compilations,
which is now not the norm, it has limited impact on the majority of
Angular projects. So this fix can probably be left in indefinitely.
PR Close#36989
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
ASTs for property read and method calls contain information about
the entire span of the expression, including its receiver. Use cases
like a language service and compile error messages may be more
interested in the span of the direct identifier for which the
expression is constructed (i.e. an accessed property). To support this,
this commit adds a `nameSpan` property on
- `PropertyRead`s
- `SafePropertyRead`s
- `PropertyWrite`s
- `MethodCall`s
- `SafeMethodCall`s
The `nameSpan` property already existed for `BindingPipe`s.
This commit also updates usages of these expressions' `sourceSpan`s in
Ngtsc and the langauge service to use `nameSpan`s where appropriate.
PR Close#36826
Some projects include .js source files (via the TypeScript allowJs option).
Previously, the compiler would attempt to tag these files for shims, which
caused errors as the regex used to create shim filenames assumes a .ts file.
This commit fixes the bug by filtering out non-ts files during tagging.
PR Close#36987
Previously in v9, we deprecated the pattern of undecorated base classes
that rely on Angular features. We ran a migration for this in version 9
and will run the same on in version 10 again.
To ensure that projects do not regress and start using the unsupported
pattern again, we report an error in ngtsc if such undecorated classes
are discovered.
We keep the compatibility code enabled in ngcc so that libraries
can be still be consumed, even if they have not been migrated yet.
Resolves FW-2130.
PR Close#36921
We can remove all of the entry point resolution configuration from the package.json
in our source code as ng_package rule adds the properties automatically and correctly
configures them.
This change simplifies our code base but doesn't have any impact on the package.json
in the distributed npm_packages.
PR Close#36944
This optimization builds on a lot of prior work to finally make type-
checking of templates incremental.
Incrementality requires two main components:
- the ability to reuse work from a prior compilation.
- the ability to know when changes in the current program invalidate that
prior work.
Prior to this commit, on every type-checking pass the compiler would
generate new .ngtypecheck files for each original input file in the program.
1. (Build #1 main program): empty .ngtypecheck files generated for each
original input file.
2. (Build #1 type-check program): .ngtypecheck contents overridden for those
which have corresponding components that need type-checked.
3. (Build #2 main program): throw away old .ngtypecheck files and generate
new empty ones.
4. (Build #2 type-check program): same as step 2.
With this commit, the `IncrementalDriver` now tracks template type-checking
_metadata_ for each input file. The metadata contains information about
source mappings for generated type-checking code, as well as some
diagnostics which were discovered at type-check analysis time. The actual
type-checking code is stored in the TypeScript AST for type-checking files,
which is now re-used between programs as follows:
1. (Build #1 main program): empty .ngtypecheck files generated for each
original input file.
2. (Build #1 type-check program): .ngtypecheck contents overridden for those
which have corresponding components that need type-checked, and the
metadata registered in the `IncrementalDriver`.
3. (Build #2 main program): The `TypeCheckShimGenerator` now reuses _all_
.ngtypecheck `ts.SourceFile` shims from build #1's type-check program in
the construction of build #2's main program. Some of the contents of
these files might be stale (if a component's template changed, for
example), but wholesale reuse here prevents unnecessary changes in the
contents of the program at this point and makes TypeScript's job a lot
easier.
4. (Build #2 type-check program): For those input files which have not
"logically changed" (meaning components within are semantically the same
as they were before), the compiler will re-use the type-check file
metadata from build #1, and _not_ generate a new .ngtypecheck shim.
For components which have logically changed or where the previous
.ngtypecheck contents cannot otherwise be reused, code generation happens
as before.
PR Close#36211
As a performance optimization, this commit splits the single
__ngtypecheck__.ts file which was previously added to the user's program as
a container for all template type-checking code into multiple .ngtypecheck
shim files, one for each original file in the user's program.
In larger applications, the generation, parsing, and checking of this single
type-checking file was a huge performance bottleneck, with the file often
exceeding 1 MB in text content. Particularly in incremental builds,
regenerating this single file for the entire application proved especially
expensive.
This commit introduces a new strategy for template type-checking code which
makes use of a new interface, the `TypeCheckingProgramStrategy`. This
interface abstracts the process of creating a new `ts.Program` to type-check
a particular compilation, and allows the mechanism there to be kept separate
from the more complex logic around dealing with multiple .ngtypecheck files.
A new `TemplateTypeChecker` hosts that logic and interacts with the
`TypeCheckingProgramStrategy` to actually generate and return diagnostics.
The `TypeCheckContext` class, previously the workhorse of template type-
checking, is now solely focused on collecting and generating type-checking
file contents.
A side effect of implementing the new `TypeCheckingProgramStrategy` in this
way is that the API is designed to be suitable for use by the Angular
Language Service as well. The LS also needs to type-check components, but
has its own method for constructing a `ts.Program` with type-checking code.
Note that this commit does not make the actual checking of templates at all
_incremental_ just yet. That will happen in a future commit.
PR Close#36211
Shim generation was built on a lie.
Shims are files added to the program which aren't original files authored by
the user, but files authored effectively by the compiler. These fall into
two categories: files which will be generated (like the .ngfactory shims we
generate for View Engine compatibility) as well as files used internally in
compilation (like the __ng_typecheck__.ts file).
Previously, shim generation was driven by the `rootFiles` passed to the
compiler as input. These are effectively the `files` listed in the
`tsconfig.json`. Each shim generator (e.g. the `FactoryGenerator`) would
examine the `rootFiles` and produce a list of shim file names which it would
be responsible for generating. These names would then be added to the
`rootFiles` when the program was created.
The fatal flaw here is that `rootFiles` does not always account for all of
the files in the program. In fact, it's quite rare that it does. Users don't
typically specify every file directly in `files`. Instead, they rely on
TypeScript, during program creation, starting with a few root files and
transitively discovering all of the files in the program.
This happens, however, during `ts.createProgram`, which is too late to add
new files to the `rootFiles` list.
As a result, shim generation was only including shims for files actually
listed in the `tsconfig.json` file, and not for the transitive set of files
in the user's program as it should.
This commit completely rewrites shim generation to use a different technique
for adding files to the program, inspired by View Engine's shim generator.
In this new technique, as the program is being created and `ts.SourceFile`s
are being requested from the `NgCompilerHost`, shims for those files are
generated and a reference to them is patched onto the original file's
`ts.SourceFile.referencedFiles`. This causes TS to think that the original
file references the shim, and causes the shim to be included in the program.
The original `referencedFiles` array is saved and restored after program
creation, hiding this little hack from the rest of the system.
The new shim generation engine differentiates between two kinds of shims:
top-level shims (such as the flat module entrypoint file and
__ng_typecheck__.ts) and per-file shims such as ngfactory or ngsummary
files. The former are included via `rootFiles` as before, the latter are
included via the `referencedFiles` of their corresponding original files.
As a result of this change, shims are now correctly generated for all files
in the program, not just the ones named in `tsconfig.json`.
A few mitigating factors prevented this bug from being realized until now:
* in g3, `files` does include the transitive closure of files in the program
* in CLI apps, shims are not really used
This change also makes use of a novel technique for associating information
with source files: the use of an `NgExtension` `Symbol` to patch the
information directly onto the AST object. This is used in several
circumstances:
* For shims, metadata about a `ts.SourceFile`'s status as a shim and its
origins are held in the extension data.
* For original files, the original `referencedFiles` are stashed in the
extension data for later restoration.
The main benefit of this technique is a lot less bookkeeping around `Map`s
of `ts.SourceFile`s to various kinds of data, which need to be tracked/
invalidated as part of incremental builds.
This technique is based on designs used internally in the TypeScript
compiler and is serving as a prototype of this design in ngtsc. If it works
well, it could have benefits across the rest of the compiler.
PR Close#36211
The compiler needs to track the dependencies of a component, including any
NgModules which happen to be present in a component's scope. If an upstream
NgModule changes, any downstream components need to have their templates
re-compiled and re-typechecked.
Previously, the compiler handled this well for the A -> B -> C case where
module A imports module B which re-exports module C. However, it fell apart
in the A -> B -> C -> D case, because previously tracking focused on changes
to components/directives in the scope, and not NgModules specifically.
This commit introduces logic to track which NgModules contributed to a given
scope, and treat them as dependencies of any components within.
This logic also contains a bug, which is intentional for now. It
purposefully does not track transitive dependencies of the NgModules which
contribute to a scope. If it did, using the current dependency system, this
would treat all components and directives (even those not exported into the
scope) as dependencies, causing a major performance bottleneck. Only those
dependencies which contributed to the module's export scope should be
considered, but the current system is incapable of making this distinction.
This will be fixed at a later date.
PR Close#36211
The html parser already normalizes line endings (converting `\r\n` to `\n`)
for most text in templates but it was missing the expressions of ICU expansions.
In ViewEngine backticked literal strings, used to define inline templates,
were already normalized by the TypeScript parser.
In Ivy we are parsing the raw text of the source file directly so the line
endings need to be manually normalized.
This change ensures that inline templates have the line endings of ICU
expression normalized correctly, which matches the ViewEngine.
In ViewEngine external templates, defined in HTML files, the behavior was
different, since TypeScript was not normalizing the line endings.
Specifically, ICU expansion "expressions" are not being normalized.
This is a problem because it means that i18n message ids can be different on
different machines that are setup with different line ending handling,
or if the developer moves a template from inline to external or vice versa.
The goal is always to normalize line endings, whether inline or external.
But this would be a breaking change since it would change i18n message ids
that have been previously computed. Therefore this commit aligns the ivy
template parsing to have the same "buggy" behavior for external templates.
There is now a compiler option `i18nNormalizeLineEndingsInICUs`, which
if set to `true` will ensure the correct non-buggy behavior. For the time
being this option defaults to `false` to ensure backward compatibility while
allowing opt-in to the desired behavior. This option's default will be
flipped in a future breaking change release.
Further, when this option is set to `false`, any ICU expression tokens,
which have not been normalized, are added to the `ParseResult` from the
`HtmlParser.parse()` method. In the future, this collection of tokens could
be used to diagnose and encourage developers to migrate their i18n message
ids. See FW-2106.
Closes#36725
PR Close#36741
Prior to this change, there was a problem while matching template attributes, which mistakenly took i18n attributes (that might be present in attrs array after template ones) into account. This commit updates the logic to avoid template attribute matching logic from entering the i18n section and as a result this also allows generating proper i18n attributes sections instead of keeping these attribute in plain form (with their values) in attribute arrays.
PR Close#36422
When the compiler needs to convert a type reference to a value
expression, it may encounter a type that refers to a namespaced symbol.
Such namespaces need to be handled specially as there's various forms
available. Consider a namespace named "ns":
1. One can refer to a namespace by itself: `ns`. A namespace is only
allowed to be used in a type position if it has been merged with a
class, but even if this is the case it may not be possible to convert
that type into a value expression depending on the import form. More
on this later (case a below)
2. One can refer to a type within the namespace: `ns.Foo`. An import
needs to be generated to `ns`, from which the `Foo` property can then
be read.
3. One can refer to a type in a nested namespace within `ns`:
`ns.Foo.Bar` and possibly even deeper nested. The value
representation is similar to case 2, but includes additional property
accesses.
The exact strategy of how to deal with these cases depends on the type
of import used. There's two flavors available:
a. A namespaced import like `import * as ns from 'ns';` that creates
a local namespace that is irrelevant to the import that needs to be
generated (as said import would be used instead of the original
import).
If the local namespace "ns" itself is referred to in a type position,
it is invalid to convert it into a value expression. Some JavaScript
libraries publish a value as default export using `export = MyClass;`
syntax, however it is illegal to refer to that value using "ns".
Consequently, such usage in a type position *must* be accompanied by
an `@Inject` decorator to provide an explicit token.
b. An explicit namespace declaration within a module, that can be
imported using a named import like `import {ns} from 'ns';` where the
"ns" module declares a namespace using `declare namespace ns {}`.
In this case, it's the namespace itself that needs to be imported,
after which any qualified references into the namespace are converted
into property accesses.
Before this change, support for namespaces in the type-to-value
conversion was limited and only worked correctly for a single qualified
name using a namespace import (case 2a). All other cases were either
producing incorrect code or would crash the compiler (case 1a).
Crashing the compiler is not desirable as it does not indicate where
the issue is. Moreover, the result of a type-to-value conversion is
irrelevant when an explicit injection token is provided using `@Inject`,
so referring to a namespace in a type position (case 1) could still be
valid.
This commit introduces logic to the type-to-value conversion to be able
to properly deal with all type references to namespaced symbols.
Fixes#36006
Resolves FW-1995
PR Close#36106
1. update jasmine to 3.5
2. update @types/jasmine to 3.5
3. update @types/jasminewd2 to 2.0.8
Also fix several cases, the new jasmine 3 will help to create test cases correctly,
such as in the `jasmine 2.x` version, the following case will pass
```
expect(1 == 2);
```
But in jsamine 3, the case will need to be
```
expect(1 == 2).toBeTrue();
```
PR Close#34625
JoostK
Alex Rickabaugh
Paul Gschwendtner
George Kalpakas
Alan Agius
Joey Perrott
Andrew Kushnir
Pete Bacon Darwin
Matias Niemelä
Ayaz Hafiz
Igor Minar
JiaLiPassion