With Typescript 4, `ts.updateIdentifier` is no longer available.
Calling `ts.updateIdentifier` used to return the same node when
`typeArguments` was `undefined` because `node.typeArguments`
was also `undefined`.
Relevant TS code:
```js
function updateIdentifier(node, typeArguments) {
return node.typeArguments !== typeArguments
? updateNode(createIdentifier(ts.idText(node), typeArguments), node)
: node;
}
```
PR Close#38076
Prior to this change, the unary + and - operators would be parsed as `x - 0`
and `0 - x` respectively. The runtime semantics of these expressions are
equivalent, however they may introduce inaccurate template type checking
errors as the literal type is lost, for example:
```ts
@Component({
template: `<button [disabled]="isAdjacent(-1)"></button>`
})
export class Example {
isAdjacent(direction: -1 | 1): boolean { return false; }
}
```
would incorrectly report a type-check error:
> error TS2345: Argument of type 'number' is not assignable to parameter
of type '-1 | 1'.
Additionally, the translated expression for the unary + operator would be
considered as arithmetic expression with an incompatible left-hand side:
> error TS2362: The left-hand side of an arithmetic operation must be of
type 'any', 'number', 'bigint' or an enum type.
To resolve this issues, the implicit transformation should be avoided.
This commit adds a new unary AST node to represent these expressions,
allowing for more accurate type-checking.
Fixes#20845Fixes#36178
PR Close#37918
We had a couple of places where we were assuming that if a particular
symbol has a value, then it will exist at runtime. This is true in most cases,
but it breaks down for `const` enums.
Fixes#38513.
PR Close#38542
This commit adds a `getTemplateOfComponent` method to the
`TemplateTypeChecker` API, which retrieves the actual nodes parsed and used
by the compiler for template type-checking. This is advantageous for the
language service, which may need to query other APIs in
`TemplateTypeChecker` that require the same nodes used to bind the template
while generating the TCB.
Fixes#38352
PR Close#38355
Similarly to the change we landed in the `@angular/core` reflection
capabilities, we need to make sure that ngcc can detect pass-through
delegate constructors for classes using downleveled ES2015 output.
More details can be found in the preceding commit, and in the issue
outlining the problem: #38453.
Fixes#38453.
PR Close#38463
This commit updates the code to move generated i18n statements into the `consts` field of
ComponentDef to avoid invoking `$localize` function before component initialization (to better
support runtime translations) and also avoid problems with lazy-loading when i18n defs may not
be present in a chunk where it's referenced.
Prior to this change the i18n statements were generated at the top leve:
```
var I18N_0;
if (typeof ngI18nClosureMode !== "undefined" && ngI18nClosureMode) {
var MSG_X = goog.getMsg(“…”);
I18N_0 = MSG_X;
} else {
I18N_0 = $localize('...');
}
defineComponent({
// ...
template: function App_Template(rf, ctx) {
i0.ɵɵi18n(2, I18N_0);
}
});
```
This commit updates the logic to generate the following code instead:
```
defineComponent({
// ...
consts: function() {
var I18N_0;
if (typeof ngI18nClosureMode !== "undefined" && ngI18nClosureMode) {
var MSG_X = goog.getMsg(“…”);
I18N_0 = MSG_X;
} else {
I18N_0 = $localize('...');
}
return [
I18N_0
];
},
template: function App_Template(rf, ctx) {
i0.ɵɵi18n(2, 0);
}
});
```
Also note that i18n template instructions now refer to the `consts` array using an index
(similar to other template instructions).
PR Close#38404
For a template that contains for example `<span *ngIf="first"></span>`
there's no need to render the `NgIf` guard expression, as the child
scope does not have any type-checking statements, so any narrowing
effect of the guard is not applicable.
This seems like a minor improvement, however it reduces the number of
flow-node antecedents that TypeScript needs to keep into account for
such cases, resulting in an overall reduction of type-checking time.
PR Close#38418
The template type-checker would always generate a directive declaration
even if its type was never used. For example, directives without any
input nor output bindings nor exportAs references don't need the
directive to be declared, as its type would never be used.
This commit makes the `TcbOp`s that are responsible for declaring a
directive as optional, such that they are only executed when requested
from another operation.
PR Close#38418
The template type-checker would generate a statement with a call
expression for all DOM elements in a template of the form:
```
const _t1 = document.createElement("div");
```
Profiling has shown that this is a particularly expensive call to
perform type inference on, as TypeScript needs to perform signature
selection of `Document.createElement` and resolve the exact type from
the `HTMLElementTagNameMap`. However, it can be observed that the
statement by itself does not contribute anything to the type-checking
result if `_t1` is not actually used anywhere, which is only rarely the
case---it requires that the element is referenced by its name from
somewhere else in the template. Consequently, the type-checker can skip
generating this statement altogether for most DOM elements.
The effect of this optimization is significant in several phases:
1. Less type-check code to generate
2. Less type-check code to emit and parse again
3. No expensive type inference to perform for the call expression
The effect on phase 3 is the most significant here, as type-checking is
not currently incremental in the sense that only phases 1 and 2 can
be reused from a prior compilation. The actual type-checking of all
templates in phase 3 needs to be repeated on each incremental
compilation, so any performance gains we achieve here are very
beneficial.
PR Close#38418
The compiler does not currently report errors when there's an `@Input()`
for a `private`, `protected`, or `readonly` directive/component class member.
This change adds an option to enable reporting errors when a template
attempts to bind to one of these restricted input fields.
PR Close#38249
Prior to this change, the template type checker would always use a
type-constructor to instantiate a directive. This type-constructor call
serves two purposes:
1. Infer any generic types for the directive instance from the inputs
that are passed in.
2. Type check the inputs that are passed into the directive's inputs.
The first purpose is only relevant when the directive actually has any
generic types and using a type-constructor for these cases inhibits
a type-check performance penalty, as a type-constructor's signature is
quite complex and needs to be generated for each directive.
This commit refactors the generated type-check blocks to only generate
a type-constructor call for directives that have generic types. Type
checking of inputs is achieved by generating individual statements for
all inputs, using assignments into the directive's fields.
Even if a type-constructor is used for type-inference of generic types
will the input checking also be achieved using the individual assignment
statements. This is done to support the rework of the language service,
which will start to extract symbol information from the type-check
blocks.
As a future optimization, it may be possible to reduce the number of
inputs passed into a type-constructor to only those inputs that
contribute the the type-inference of the generics. As this is not a
necessity at the moment this is left as follow-up work.
Closes#38185
PR Close#38249
"Quote expressions" are expressions that start with an identifier followed by a
comma, allowing arbitrary syntax to follow. These kinds of expressions would
throw a an error in the template type checker, which would make them hard to
track down. As quote expressions are not generally used at all, the error would
typically occur for URLs that would inadvertently occur in a binding:
```html
<a [href]="https://example.com"></a>
```
This commit lets such bindings be inferred as the `any` type.
Fixes#36568
Resolves FW-2051
PR Close#37917
In TypeScript 3.8 support was added for type-only imports, which only brings in
the symbol as a type, not their value. The Angular compiler did not yet take
the type-only keyword into account when representing symbols in type positions
as value expressions. The class metadata that the compiler emits would include
the value expression for its parameter types, generating actual imports as
necessary. For type-only imports this should not be done, as it introduces an
actual import of the module that was originally just a type-only import.
This commit lets the compiler deal with type-only imports specially, preventing
a value expression from being created.
Fixes#37900
PR Close#37912
When using the safe navigation operator in a binding expression, a temporary
variable may be used for storing the result of a side-effectful call.
For example, the following template uses a pipe and a safe property access:
```html
<app-person-view [enabled]="enabled" [firstName]="(person$ | async)?.name"></app-person-view>
```
The result of the pipe evaluation is stored in a temporary to be able to check
whether it is present. The temporary variable needs to be declared in a separate
statement and this would also cause the full expression itself to be pulled out
into a separate statement. This would compile into the following
pseudo-code instructions:
```js
var temp = null;
var firstName = (temp = pipe('async', ctx.person$)) == null ? null : temp.name;
property('enabled', ctx.enabled)('firstName', firstName);
```
Notice that the pipe evaluation happens before evaluating the `enabled` binding,
such that the runtime's internal binding index would correspond with `enabled`,
not `firstName`. This introduces a problem when the pipe uses `WrappedValue` to
force a change to be detected, as the runtime would then mark the binding slot
corresponding with `enabled` as dirty, instead of `firstName`. This results
in the `enabled` binding to be updated, triggering setters and affecting how
`OnChanges` is called.
In the pseudo-code above, the intermediate `firstName` variable is not strictly
necessary---it only improved readability a bit---and emitting it inline with
the binding itself avoids the out-of-order execution of the pipe:
```js
var temp = null;
property('enabled', ctx.enabled)
('firstName', (temp = pipe('async', ctx.person$)) == null ? null : temp.name);
```
This commit introduces a new `BindingForm` that results in the above code to be
generated and adds compiler and acceptance tests to verify the proper behavior.
Fixes#37194
PR Close#37911
When we were outputting class members for `setClassMetadata` calls,
we were using the string representation of the member name. This can
lead to us generating invalid code when the name contains dashes and
is quoted (e.g. `@Output() 'has-dashes' = new EventEmitter()`), because
the quotes will be stripped for the string representation.
These changes fix the issue by using the original name AST node that was
used for the declaration and which knows whether it's supposed to be
quoted or not.
Fixes#38311.
PR Close#38387
For attribute bindings that target a directive's input, the template
type checker is able to verify that the type of the input expression is
compatible with the directive's declaration for said input. This
checking adheres to the `strictNullChecks` flag as configured in the
TypeScript compilation, such that errors are reported for expressions
that include `undefined` or `null` in their type if the input's
declaration does not include those types.
There was a bug with this level of type-checking for directives that
also declare coercion members, where binding an expression that includes
the `undefined` type to a directive's input that does not include the
`undefined` type would not be reported as error.
This commit fixes the bug by changing the type-constructor in type-check
code to use an intersection type of regular inputs and coerced inputs,
instead of a union type. The union type would inadvertently allow
`undefined` types to be assigned into the regular inputs, as that would
still satisfy the characteristics of a union type.
As a result of this change, you may start to see build failures if
`strictTemplates` is enabled and `strictInputTypes` is not disabled.
These errors are legitimate and some action is required to achieve a
successful build:
1. Update the templates for which an error is reported and introduce the
non-null assertion operator at the end of the expression. This
removes the `undefined` type from the expression's type, making it
appear as a valid assignment.
2. Disable `strictNullInputTypes` in the compiler options. This will
implicitly add the non-null assertion operators similar to option 1,
but all templates in the compilation are affected.
3. Update the directive's input declaration to include the `undefined`
type, if the directive is not implemented in an external library.
PR Close#38273
Roll forward of #38147.
This allows Closure compiler to tree shake unused constructor calls to `NgModuleFactory`, which is otherwise considered
side-effectful. The Angular compiler generates factory objects which are exported but typically not used, as they are
only needed for compatibility with View Engine. This results in top-level constructor calls, such as:
```typescript
export const FooNgFactory = new NgModuleFactory(Foo);
```
`NgModuleFactory` has a side-effecting constructor, so this statement cannot be tree shaken, even if `FooNgFactory` is
never imported. The `NgModuleFactory` continues to reference its associated `NgModule` and prevents the module and all
its unused dependencies from being tree shaken, making Closure builds significantly larger than necessary.
The fix here is to wrap `NgModuleFactory` constructor with `noSideEffects(() => /* ... */)`, which tricks the Closure
compiler into assuming that the invoked function has no side effects. This allows it to tree-shake unused
`NgModuleFactory()` constructors when they aren't imported. Since the factory can be removed, the module can also be
removed (if nothing else references it), thus tree shaking unused dependencies as expected.
The one notable edge case is for lazy loaded modules. Internally, lazy loading is done as a side effect when the lazy
script is evaluated. For Angular, this side effect is registering the `NgModule`. In Ivy this is done by the
`NgModuleFactory` constructor, so lazy loaded modules **cannot** have their top-level `NgModuleFactory` constructor
call tree shaken. We handle this case by looking for the `id` field on `@NgModule` annotations. All lazy loaded modules
include an `id`. When this `id` is found, the `NgModuleFactory` is generated **without** with `noSideEffects()` call,
so Closure will not tree shake it and the module will lazy-load correctly.
PR Close#38320
This introduces a new `ModuleInfo` interface to represent some of the statically analyzed data from an `NgModule`. This
gets passed into transforms to give them more context around a given `NgModule` in the compilation.
PR Close#38320
The `TscPlugin` interface using a type of `ts.CompilerHost&Partial<UnifiedModulesHost>` for the `host` parameter
of the `wrapHost` method. However, prior to this change, the interface implementing `NgTscPlugin` class used a
type of `ts.CompilerHost&UnifiedModulesHost` for the parameter. This change corrects the inconsistency and
allows `UnifiedModulesHost` members to be optional when using the `NgtscPlugin`.
PR Close#38004
Currently the `getInheritedFactory` function is implemented to allow
closure to remove the call if the base factory is unused. However, this
method does not work with terser. By adding the PURE annotation,
terser will also be able to remove the call when unused.
PR Close#38291
This reverts commit 7f8c2225f2.
This commit caused test failures internally, which were traced back to the
optimizer removing NgModuleFactory constructor calls when those calls caused
side-effectful registration of NgModules by their ids.
PR Close#38303
This commit disables one TypeChecker test (added as a part of
https://github.com/angular/angular/pull/38105) which make assertions about the filename while
running on Windows.
Such assertions are currently suffering from a case sensitivity issue.
PR Close#38294
This allows Closure compiler to tree shake unused constructor calls to `NgModuleFactory`, which is otherwise considered
side-effectful. The Angular compiler generates factory objects which are exported but typically not used, as they are
only needed for compatibility with View Engine. This results in top-level constructor calls, such as:
```typescript
export const FooNgFactory = new NgModuleFactory(Foo);
```
`NgModuleFactory` has a side-effecting constructor, so this statement cannot be tree shaken, even if `FooNgFactory` is
never imported. The `NgModuleFactory` continues to reference its associated `NgModule` and prevents the module and all
its unused dependencies from being tree shaken. This effectively prevents all components from being tree shaken, making
Closure builds significantly larger than they should be.
The fix here is to wrap `NgModuleFactory` constructor with `noSideEffects(() => /* ... */)`, which tricks the Closure
compiler into assuming that the invoked function has no side effects. This allows it to tree-shake unused
`NgModuleFactory()` constructors when they aren't imported. Since the factory can be removed, the module can also be
removed (if nothing else references it), thus tree shaking unused components as expected.
PR Close#38147
Large strings constants are now wrapped in a function which is called whenever used. This works around a unique
limitation of Closure, where it will **always** inline string literals at **every** usage, regardless of how large the
string literal is or how many times it is used.The workaround is to use a function rather than a string literal.
Closure has differently inlining semantics for functions, where it will check the length of the function and the number
of times it is used before choosing to inline it. By using a function, `ngtsc` makes Closure more conservative about
inlining large strings, and avoids blowing up the bundle size.This optimization is only used if the constant is a large
string. A wrapping function is not included for other use cases, since it would just increase the bundle size and add
unnecessary runtime performance overhead.
PR Close#38253
This commit adds a method `getDiagnosticsForComponent` to the
`TemplateTypeChecker`, which does the minimum amount of work to retrieve
diagnostics for a single component.
With the normal `ReusedProgramStrategy` this offers virtually no improvement
over the standard `getDiagnosticsForFile` operation, but if the
`TypeCheckingProgramStrategy` supports separate shims for each component,
this operation can yield a faster turnaround for components that are
declared in files with many other components.
PR Close#38105
Previously, a stable template id was implemented for each component in a
file. This commit adds this id to each `TemplateDiagnostic` generated from
the template type-checker, so it can potentially be used for filtration.
PR Close#38105
This commit adds an `overrideComponentTemplate` operation to the template
type-checker. This operation changes the template used during template
type-checking operations.
Overriding a template causes any previous work for it to be discarded, and
the template type-checking engine will regenerate the TCB for that template
on the next request.
This operation can be used by a consumer such as the language service to
get rapid feedback or diagnostics as the user is editing a template file,
without the need for a full incremental build iteration.
Closes#38058
PR Close#38105
Previously, the `TemplateTypeChecker` abstraction allowed fetching
diagnostics for a single file, but under the hood would generate type
checking code for the entire program to satisfy the request.
With this commit, an `OptimizeFor` hint is passed to `getDiagnosticsForFile`
which indicates whether the user intends to request diagnostics for the
whole program or is truly interested in just the single file. If the latter,
the `TemplateTypeChecker` can perform only the work needed to produce
diagnostics for just that file, thus returning answers more efficiently.
PR Close#38105
The template type-checking engine relies on the abstraction interface
`TypeCheckingProgramStrategy` to create updated `ts.Program`s for
template type-checking. The basic API is that the type-checking engine
requests changes to certain files in the program, and the strategy provides
an updated `ts.Program`.
Typically, such changes are made to 'ngtypecheck' shim files, but certain
conditions can cause template type-checking to require "inline" operations,
which change user .ts files instead. The strategy used by 'ngc' (the
`ReusedProgramStrategy`) supports these kinds of updates, but other clients
such as the language service might not always support modifying user files.
To accommodate this, the `TypeCheckingProgramStrategy` interface was
modified to include a `supportsInlineOperations` flag. If an implementation
specifies `false` for inline support, the template type-checking system will
return diagnostics on components which would otherwise require inline
operations.
Closes#38059
PR Close#38105
This commit significantly refactors the 'typecheck' package to introduce a
new abstraction, the `TemplateTypeChecker`. To achieve this:
* a 'typecheck:api' package is introduced, containing common interfaces that
consumers of the template type-checking infrastructure can depend on
without incurring a dependency on the template type-checking machinery as
a whole.
* interfaces for `TemplateTypeChecker` and `TypeCheckContext` are introduced
which contain the abstract operations supported by the implementation
classes `TemplateTypeCheckerImpl` and `TypeCheckContextImpl` respectively.
* the `TemplateTypeChecker` interface supports diagnostics on a whole
program basis to start with, but the implementation is purposefully
designed to support incremental diagnostics at a per-file or per-component
level.
* `TemplateTypeChecker` supports direct access to the type check block of a
component.
* the testing utility is refactored to be a lot more useful, and new tests
are added for the new abstraction.
PR Close#38105
Previously in the template type-checking engine, it was assumed that every
input file would have an associated type-checking shim. The type check block
code for all components in the input file would be generated into this shim.
This is fine for whole-program type checking operations, but to support the
language service's requirements for low latency, it would be ideal to be
able to check a single component in isolation, especially if the component
is declared along with many others in a single file.
This commit removes the assumption that the file/shim mapping is 1:1, and
introduces the concept of component-to-shim mapping. Any
`TypeCheckingProgramStrategy` must provide such a mapping.
To achieve this:
* type checking record information is now split into file-level data as
well as per-shim data.
* components are now assigned a stable `TemplateId` which is unique to the
file in which they're declared.
PR Close#38105
When the `NgIf` directive is used in a template, its context variables
can be used to capture the bound value. This is sometimes used in
complex expressions, where the resulting value is captured in a
context variable. There's two syntax forms available:
1. Binding to `NgIfContext.ngIf` using the `as` syntax:
```html
<span *ngIf="enabled && user as u">{{u.name}}</span>
```
2. Binding to `NgIfContext.$implicit` using the `let` syntax:
```html
<span *ngIf="enabled && user; let u">{{u.name}}</span>
```
Because of the semantics of `ngIf`, it is known that the captured
context variable is truthy, however the template type checker
would not consider them as such and still report errors when
`strict` is enabled.
This commit updates `NgIf`'s context guard to make the types of the
context variables truthy, avoiding the issue.
Based on https://github.com/angular/angular/pull/35125
PR Close#36627
The current implementation of the TypeScriptReflectionHost does not account for members that
are string literals, i.e. `class A { 'string-literal-prop': string; }`
PR Close#38226
Prior to this commit, duplicated styles defined in multiple components in the same file were not
shared between components, thus causing extra payload size. This commit updates compiler logic to
use `ConstantPool` for the styles (while generating the `styles` array on component def), which
enables styles sharing when needed (when duplicates styles are present).
Resolves#38204.
PR Close#38213
This commit splits the transformation into 2 separate steps: Ivy compilation and actual transformation
of corresponding TS nodes. This is needed to have all `o.Expression`s generated before any TS transforms
happen. This allows `ConstantPool` to properly identify expressions that can be shared across multiple
components declared in the same file.
Resolves#38203.
PR Close#38213
This commit updates synthetic host property and listener instruction names to better align with other instructions.
The `ɵɵupdateSyntheticHostBinding` instruction was renamed to `ɵɵsyntheticHostProperty` (to match the `ɵɵhostProperty`
instruction name) and `ɵɵcomponentHostSyntheticListener` was renamed to `ɵɵsyntheticHostListener` since this
instruction is generated for both Components and Directives (so 'component' is removed from the name).
This PR is a followup after PR #35568.
PR Close#37145
Fixes the following issues related to how we validate properties during JIT:
- The invalid property warning was printing `null` as the node name
for `ng-content`. The problem is that when generating a template from
`ng-content` we weren't capturing the node name.
- We weren't running property validation on `ng-container` at all.
This used to be supported on ViewEngine and seems like an oversight.
In the process of making these changes, I found and cleaned up a
few places where we were passing in `LView` unnecessarily.
PR Close#37773
In CLI v10 there was a move to use the new solution-style tsconfig
which became available in TS 3.9.
The result of this is that the standard tsconfig.json no longer contains
important information such as "paths" mappings, which ngcc might need to
correctly compute dependencies.
ngcc (and ngc and tsc) infer the path to tsconfig.json if not given an
explicit tsconfig file-path. But now that means it infers the solution
tsconfig rather than one that contains the useful information it used to
get.
This commit logs a warning in this case to inform the developer
that they might not have meant to load this tsconfig and offer
alternative options.
Fixes#36386
PR Close#38003
The `fs.relative()` method assumed that the file-system is a single tree,
which is not the case in Windows, where you can have multiple drives,
e.g. `C:`, `D:` etc.
This commit changes `fs.relative()` so that it no longer forces the result
to be a `PathSegment` and then flows that refactoring through the rest of
the compiler-cli (and ngcc). The main difference is that now, in some cases,
we needed to check whether the result is "rooted", i.e an `AbsoluteFsPath`,
rather than a `PathSegment`, before using it.
Fixes#36777
PR Close#37959
Builds on top of #34655 to support more cases that could be using a pipe inside host bindings (e.g. ternary expressions or function calls).
Fixes#37610.
PR Close#37883
The `ng_module` rule supports the generation of flat module bundles. In
View Engine, information about this flat module bundle is exposed
as a Bazel provider. This is helpful as other rules like `ng_package`
could rely on this information to determine entry-points for the APF.
With Ivy this currently does not work because the flat module
information is not exposed in the provider. The reason for this is
unclear. We should also provide this information in Ivy so that rules
like `ng_package` can also determine the correct entry-points when a
package is built specifically with `--config=ivy`.
PR Close#36971
The ngtsc testing packages for file_system and logging were missing from the bazel deps rules, which means that they were not included in the releases
PR Close#37977
Currently when the `plural` or `select` keywords in an ICU contain trailing spaces (e.g. `{count, select , ...}`), these spaces are also included into the key names in ICU vars (e.g. "VAR_SELECT "). These trailing spaces are not desirable, since they will later be converted into `_` symbols while normalizing placeholder names, thus causing mismatches at runtime (i.e. placeholder will not be replaced with the correct value). This commit updates the code to trim these spaces while generating an object with placeholders, to make sure the runtime logic can replace these placeholders with the right values.
PR Close#37866
Incremental compilation allows for the output state of one compilation to be
reused as input to the next compilation. This involves retaining references
to instances from prior compilations, which must be done carefully to avoid
memory leaks.
This commit fixes such a leak with a complicated retention chain:
* `TrackedIncrementalBuildStrategy` unnecessarily hangs on to the previous
`IncrementalDriver` (state of the previous compilation) once the current
compilation completes.
In general this is unnecessary, but should be safe as long as the chain
only goes back one level - if the `IncrementalDriver` doesn't retain any
previous `TrackedIncrementalBuildStrategy` instances. However, this does
happen:
* `NgCompiler` indirectly causes retention of previous `NgCompiler`
instances (and thus previous `TrackedIncrementalBuildStrategy` instances)
through accidental capture of the `this` context in a closure created in
its constructor. This closure is wrapped in a `ts.ModuleResolutionCache`
used to create a `ModuleResolver` class, which is passed to the program's
`TraitCompiler` on construction.
* The `IncrementalDriver` retains a reference to the `TraitCompiler` of the
previous compilation, completing the reference chain.
The final retention chain thus looks like:
* `TrackedIncrementalBuildStrategy` of current program
* `.previous`: `IncrementalDriver` of previous program
* `.lastGood.traitCompiler`: `TraitCompiler`
* `.handlers[..].moduleResolver.moduleResolutionCache`: cache
* (via `getCanonicalFileName` closure): `NgCompiler`
* `.incrementalStrategy`: `TrackedIncrementalBuildStrategy` of previous
program.
The closure link is the "real" leak here. `NgCompiler` is creating a closure
for `getCanonicalFileName`, delegating to its
`this.adapter.getCanonicalFileName`, for the purposes of creating a
`ts.ModuleResolutionCache`. The fact that the closure references
`NgCompiler` thus eventually causes previous `NgCompiler` iterations to be
retained. This is also potentially problematic due to the shared nature of
`ts.ModuleResolutionCache`, which is potentially retained across multiple
compilations intentionally.
This commit fixes the first two links in the retention chain: the build
strategy is patched to not retain a `previous` pointer, and the `NgCompiler`
is patched to not create a closure in the first place, but instead pass a
bound function. This ensures that the `NgCompiler` does not retain previous
instances of itself in the first place, even if the build strategy does
end up retaining the previous incremental state unnecessarily.
The third link (`IncrementalDriver` unnecessarily retaining the whole
`TraitCompiler`) is not addressed in this commit as it's a more
architectural problem that will require some refactoring. However, the leak
potential of this retention is eliminated thanks to fixing the first two
issues.
PR Close#37835
When ngcc creates an entry-point program, the `allowJs` option is enabled
in order to operate on the JavaScript source files of the entry-point.
A side-effect of this approach is that external modules that don't ship
declaration files will also have their JavaScript source files loaded
into the program, as the `allowJs` flag allows for them to be imported.
This may pose an issue in certain edge cases, where ngcc would inadvertently
operate on these external modules. This can introduce all sorts of undesirable
behavior and incompatibilities, e.g. the reflection host that is selected for
the entry-point's format could be incompatible with that of the external
module's JavaScript bundles.
To avoid these kinds of issues, module resolution that would resolve to
a JavaScript file located outside of the package will instead be rejected,
as if the file would not exist. This would have been the behavior when
`allowJs` is set to false, which is the case in typical Angular compilations.
Fixes#37508
PR Close#37596
Alan Agius
JoostK
crisbeto
Alex Rickabaugh
Joey Perrott
Paul Gschwendtner
Andrew Kushnir
Ahn
Andrew Scott
Doug Parker
Charles Lyding
Andrea Canciani
Pete Bacon Darwin
Olivier Combe