The compiler uses metadata to represent what it statically knows about
various expressions in a program. Occasionally, expressions in the program
for which metadata is extracted may contain sub-expressions which are not
representable in metadata. One such construct is an arrow function.
The compiler does not always need to understand such expressions completely.
For example, for a provider defined with `useValue`, the compiler does not
need to understand the value at all, only the outer provider definition. In
this case, the compiler employs a technique known as "expression lowering",
where it rewrites the provider expression into one that can be represented
in metadata. Chiefly, this involves extracting out the dynamic part (the
`useValue` expression) into an exported constant.
Lowering is applied through a heuristic, which considers the containing
statement as well as the field name of the expression.
Previously, this heuristic was not completely accurate in the case of
route definitions and the `loadChildren` field, which is lowered. If the
route definition using `loadChildren` existed inside a decorator invocation,
lowering was performed correctly. However, if it existed inside a standalone
variable declaration with an export keyword, the heuristic would conclude
that lowering was unnecessary. For ordinary providers this is true; however
the compiler attempts to fully understand the ROUTES token and thus even if
an array of routes is declared in an exported variable, any `loadChildren`
expressions within still need to be lowered.
This commit enables lowering of already exported variables under a limited
set of conditions (where the initializer expression is of a specific form).
This should enable the use of `loadChildren` in route definitions.
PR Close#30038
Previously, during the evaluation of a function call where no argument
was provided for a parameter that has a default value, the default value
would be taken from the context of the caller, instead of the callee.
This commit fixes the behavior by resolving the default value of a
parameter in the context of the callee.
PR Close#29888
Previously, ngtsc's static evaluator did not take spread operators into
account when evaluating function calls, nor did it handle rest arguments
correctly. This commit adds support for static evaluation of these
language features.
PR Close#29888
Template type-checking is enabled by default in the View Engine compiler.
The feature in Ivy is not quite ready for this yet, so this flag will
temporarily control whether templates are type-checked in ngtsc.
The goal is to remove this flag after rolling out template type-checking in
google3 in Ivy mode, and making sure the feature is as compatible with the
View Engine implementation as possible.
Initially, the default value of the flag will leave checking disabled.
PR Close#29698
This commit adds support for template type-checking a pipe binding which
previously was not handled by the type-checking engine. In compatibility
mode, the arguments to transform() are not checked and the type returned
by a pipe is 'any'. In full type-checking mode, the transform() method's
type signature is used to check the pipe usage and infer the return type
of the pipe.
Testing strategy: TCB tests included.
PR Close#29698
The template type-checking engine previously would assemble a type-checking
program by inserting Type Check Blocks (TCBs) into existing user files. This
approach proved expensive, as TypeScript has to re-parse and re-type-check
those files when processing the type-checking program.
Instead, a far more performant approach is to augment the program with a
single type-checking file, into which all TCBs are generated. Additionally,
type constructors are also inlined into this file.
This is not always possible - both TCBs and type constructors can sometimes
require inlining into user code, particularly if bound generic type
parameters are present, so the approach taken is actually a hybrid. These
operations are inlined if necessary, but are otherwise generated in a single
file.
It is critically important that the original program also include an empty
version of the type-checking file, otherwise the shape of the two programs
will be different and TypeScript will throw away all the old program
information. This leads to a painfully slow type checking pass, on the same
order as the original program creation. A shim to generate this file in the
original program is therefore added.
Testing strategy: this commit is largely a refactor with no externally
observable behavioral differences, and thus no tests are needed.
PR Close#29698
This commit adds support in the template type-checking engine for handling
the logical not operation and the safe navigation operation.
Safe navigation in particular is tricky, as the View Engine implementation
has a rather inconvenient flaw. View Engine checks a safe navigation
operation `a?.b` as:
```typescript
(a != null ? a!.b : null as any)
```
The type of this expression is always 'any', as the false branch of the
ternary has type 'any'. Thus, using null-safe navigation throws away the
type of the result, and breaks type-checking for the rest of the expression.
A flag is introduced in the type-checking configuration to allow Ivy to
mimic this behavior when needed.
Testing strategy: TCB tests included.
PR Close#29698
View Engine's implementation of naive template type-checking is less
advanced than the current Ivy implementation. As a result, Ivy catches lots
of typing bugs which VE does not. As a result, it's necessary to tone down
the Ivy template type-checker in the default case.
This commit introduces a mechanism for doing that, by passing a config to
the template type-checking engine. Through this configuration, particular
checks can be loosened or disabled entirely.
Testing strategy: TCB tests included.
PR Close#29698
Previously the template type-checking code only considered the metadata of
directive classes actually referenced in the template. If those directives
had base classes, any inputs/outputs/etc of the base classes were not
tracked when generating the TCB. This resulted in bindings to those inputs
being incorrectly attributed to the host component or element.
This commit uses the new metadata package to follow directive inheritance
chains and use the full metadata for a directive for TCB generation.
Testing strategy: Template type-checking tests included.
PR Close#29698
Previously, metadata registration (the recording of collected metadata
during analysis of directives, pipes, and NgModules) was only used to
produce the `LocalModuleScope`, and thus was handled by the
`LocalModuleScopeRegistry`.
However, the template type-checker also needs information about registered
directives, outside of the NgModule scope determinations. Rather than
reuse the scope registry for an unintended purpose, this commit introduces
new abstractions for metadata registration and lookups in a separate
'metadata' package, which the scope registry implements.
This paves the way for a future commit to make use of this metadata for the
template type-checking system.
Testing strategy: this commit is a refactoring which introduces no new
functionality, so existing tests are sufficient.
PR Close#29698
Previously, bindings to [class] and [style] were treated like any other
property binding. That is, they would result in type-checking code that
attempted to write directly to .class or .style on the element node.
This is incorrect, however - the mapping from Angular's [class] and [style]
onto the DOM properties is non-trivial.
For now, this commit avoids the issue by only checking the expressions
themselves and not the assignment to the element properties.
Testing strategy: TCB tests included.
PR Close#29698
Previously the template type-checking engine processed templates in a linear
manner, and could not handle '#' references within a template. One reason
for this is that '#' references are non-linear - a reference can be used
before its declaration. Consider the template:
```html
{{ref.value}}
<input #ref>
```
Accommodating this required refactoring the type-checking code generator to
be able to produce Type Check Block (TCB) code non-linearly. Now, each
template is processed and a list of TCB operations (`TcbOp`s) are created.
Non-linearity is modeled via dependencies between operations, with the
appropriate protection in place for circular dependencies.
Testing strategy: TCB tests included.
PR Close#29698
This commit adds support for the generation of type-checking expressions for
forms which were previously unsupported:
* array literals
* map literals
* keyed property accesses
* non-null assertions
Testing strategy: TCB tests included.
Fixes#29327
FW-1218 #resolve
PR Close#29698
This commit adds a test suite for the Type Check Block generation which
doesn't require running the entire compiler (specifically, it doesn't even
require the creation of a ts.Program).
PR Close#29698
This commit adds registration of AOT compiled NgModules that have 'id'
properties set in their metadata. Such modules have a call to
registerNgModuleType() emitted as part of compilation.
The JIT behavior of this code is already in place.
This is required for module loading systems (such as g3) which rely on
getModuleFactory().
PR Close#29980
Previously, ngtsc would fail to resolve `forwardRef` calls if they
contained additional parenthesis or casts. This commit changes the
behavior to first unwrap the AST nodes to see past such insignificant
nodes, resolving the issue.
Fixes#29639
PR Close#29886
Previously, only static evaluation of `Array.slice` was implemented in
ngtsc's static evaluator. This commit adds support for `Array.concat`.
Closes#29835
PR Close#29887
The config path is an optional argument to `ts.parseJsonConfigFileContent`. When passed, it is added to the returned object as `options.configFilePath`, and `tsc` itself passes it in.
The new TS 3.4 [incremental](https://www.typescriptlang.org/docs/handbook/release-notes/typescript-3-4.html) build functionality relies on this property being present: 025d826339/src/compiler/emitter.ts (L56-L57)
When using The compiler-cli `readConfiguration` the config path option isn't passed, preventing consumers (like @ngtools/webpack) from obtaining a complete config object.
This PR fixes this omission and should allow JIT users of @ngtools/webpack to set the `incremental` option in their tsconfig and have it be used by the TS program.
I tested this in JIT and saw a small decrease in build times in a small project. In AOT the incremental option didn't seem to be used at all, due to how `ngc` uses the TS APIs.
Related to https://github.com/angular/angular-cli/issues/13941.
PR Close#29872
Plural ICU expressions depend on the locale (different languages have different plural forms). Until now the locale was hard coded as `en-US`.
For compatibility reasons, if you use ivy with AOT and bootstrap your app with `bootstrapModule` then the `LOCALE_ID` token will be set automatically for ivy, which is then used to get the correct plural form.
If you use JIT, you need to define the `LOCALE_ID` provider on the module that you bootstrap.
For `TestBed` you can use either `configureTestingModule` or `overrideProvider` to define that provider.
If you don't use the compat mode and start your app with `renderComponent` you need to call `ɵsetLocaleId` manually to define the `LOCALE_ID` before bootstrap. We expect this to change once we start adding the new i18n APIs, so don't rely on this function (there's a reason why it's a private export).
PR Close#29249
The `Δ` caused issue with other infrastructure, and we are temporarily
changing it to `ɵɵ`.
This commit also patches ts_api_guardian_test and AIO to understand `ɵɵ`.
PR Close#29850
Previously, if a matching rootDir ended with a slash then the path
returned from `logicalPathOfFile()` would not start with a slash,
which is inconsistent.
PR Close#29627
The defineInjector function specifies its providers and imports array to
be optional, so if no providers/imports are present these keys may be
omitted. This commit updates the compiler to only generate the keys when
necessary.
PR Close#29598
Prior to this change, a module's imports and exports would be used verbatim
as an injectors' imports. This is detrimental for tree-shaking, as a
module's exports could reference declarations that would then prevent such
declarations from being eligible for tree-shaking.
Since an injector actually only needs NgModule references as its imports,
we may safely filter out any declarations from the list of module exports.
This makes them eligible for tree-shaking once again.
PR Close#29598
Prior to this change, all module metadata would be included in the
`defineNgModule` call that is set as the `ngModuleDef` field of module
types. Part of the metadata is scope information like declarations,
imports and exports that is used for computing the transitive module
scope in JIT environments, preventing those references from being
tree-shaken for production builds.
This change moves the metadata for scope computations to a pure function
call that patches the scope references onto the module type. Because the
function is marked pure, it may be tree-shaken out during production builds
such that references to declarations and exports are dropped, which in turn
allows for tree-shaken any declaration that is not otherwise referenced.
Fixes#28077, FW-1035
PR Close#29598
In ES2015, classes could have been emitted as a variable declaration
initialized with a class expression. In certain situations, an intermediary
variable suffixed with `_1` is present such that the variable
declaration's initializer becomes a binary expression with its rhs being
the class expression, and its lhs being the identifier of the intermediate
variable. This structure was not recognized, resulting in such classes not
being considered as a class in `Esm2015ReflectionHost`.
As a consequence, the analysis of functions/methods that return a
`ModuleWithProviders` object did not take the methods of such classes into
account.
Another edge-case with such intermediate variable was that static
properties would not be considered as class members. A testcase was added
to prevent regressions.
Fixes#29078
PR Close#29119
Currently there is no support in ngtsc for imports of the form:
```
import * as core from `@angular/core`
export function forRoot(): core.ModuleWithProviders;
```
This commit modifies the `ReflectionHost.getImportOfIdentifier(id)`
method, so that it supports this kind of return type.
PR Close#27675
This commit introduces a mechanism for incremental compilation to the ngtsc
compiler.
Previously, incremental information was used in the construction of the
ts.Program for subsequent compilations, but was not used in ngtsc itself.
This commit adds an IncrementalState class, which tracks state between ngtsc
compilations. Currently, this supports skipping the TypeScript emit step
when the compiler can prove the contents of emit have not changed.
This is implemented for @Injectables as well as for files which don't
contain any Angular decorated types. These are the only files which can be
proven to be safe today.
See ngtsc/incremental/README.md for more details.
PR Close#29380
This commit adds support for compiling the same program repeatedly in a way
that's similar to how incremental builds work in a tool such as the CLI.
* support is added to the compiler entrypoint for reuse of the Program
object between compilations. This is the basis of the compiler's
incremental compilation model.
* support is added to wrap the CompilerHost the compiler creates and cache
ts.SourceFiles in between compilations.
* support is added to track when files are emitted, for assertion purposes.
* an 'exclude' section is added to the base tsconfig to prevent .d.ts
outputs from the first compilation from becoming inputs to any subsequent
compilations.
PR Close#29380
This commit adds a `tracePerformance` option for tsconfig.json. When
specified, it causes a JSON file with timing information from the ngtsc
compiler to be emitted at the specified path.
This tracing system is used to instrument the analysis/emit phases of
compilation, and will be useful in debugging future integration work with
@angular/cli.
See ngtsc/perf/README.md for more details.
PR Close#29380
Currently, ngtsc decides to use remote scoping if the compilation of a
component may create a cyclic import. This happens if there are two
components in a scope (say, A and B) and A directly uses B. During
compilation of B ngtsc will then note that if B were to use A, a cycle would
be generated, and so it will opt to use remote scoping for B.
ngtsc already uses the R3TargetBinder to correctly track the imports that
are actually required, for future cycle tracking. This commit expands that
usage to not trigger remote scoping unless B actually does consume A in its
template.
PR Close#29404
Currently when building an Angular project with `ngtsc`
and `flatModuleOutFile` enabled, the Ngtsc build will fail
if there are multiple source files as root file names.
Ngtsc and NGC currently determine the entry-point for multiple
root file names by looking for files ending with `/index.ts`.
This functionality is technically deprecated, but still supported
and currently breaks on Windows as the root file names are not
guaranteed to be normalized POSIX-like paths.
In order to make this logic more reliable in the future, this commit
also switches the shim generators and entry-point logic to the branded
path types. This ensures that we don't break this in the future.
PR Close#29453
Previously, several `ngtsc` and `ngcc` APIs dealing with class
declaration nodes used inconsistent types. For example, some methods of
the `DecoratorHandler` interface expected a `ts.Declaration` argument,
but actual `DecoratorHandler` implementations specified a stricter
`ts.ClassDeclaration` type.
As a result, the stricter methods would operate under the incorrect
assumption that their arguments were of type `ts.ClassDeclaration`,
while the actual arguments might be of different types (e.g. `ngcc`
would call them with `ts.FunctionDeclaration` or
`ts.VariableDeclaration` arguments, when compiling ES5 code).
Additionally, since we need those class declarations to be referenced in
other parts of the program, `ngtsc`/`ngcc` had to either repeatedly
check for `ts.isIdentifier(node.name)` or assume there was a `name`
identifier and use `node.name!`. While this assumption happens to be
true in the current implementation, working around type-checking is
error-prone (e.g. the assumption might stop being true in the future).
This commit fixes this by introducing a new type to be used for such
class declarations (`ts.Declaration & {name: ts.Identifier}`) and using
it consistently throughput the code.
PR Close#29209
Previously, only directives and services with generic type parameters
would emit `any` as generic type when emitting Ivy metadata into .d.ts
files. Pipes can also have generic type parameters but did not emit
`any` for all type parameters, resulting in the omission of those
parameters which causes compilation errors.
This commit adds support for pipes with generic type arguments and emits
`any` as generic type in the Ivy metadata.
Fixes#29400
PR Close#29403
Previously we only compiled the typings files, in ngcc, if there was
an ES2015 formatted bundle avaiable. This turns out to be an artificial
constraint and we can also support typings compilation via ES5 formats
too.
This commit changes the ngcc compiler to attempt typings compilation
via ES5 if necessary. The order of the formats to consider is now:
FESM2015, FESM5, ESM2015, ESM5.
FW-1122
PR Close#29092
This fix is for a bug in the ngtsc PartialEvaluator, which statically
evaluates expressions.
Sometimes, evaluating a reference requires resolving a function which is
declared in another module, and thus no function body is available. To
support this case, the PartialEvaluator has the concept of a foreign
function resolver.
This allows the interpretation of expressions like:
const router = RouterModule.forRoot([]);
even though the definition of the 'forRoot' function has no body. In
ngtsc today, this will be resolved to a Reference to RouterModule itself,
via the ModuleWithProviders foreign function resolver.
However, the PartialEvaluator also associates any Identifiers in the path
of this resolution with the Reference. This is done so that if the user
writes
const x = imported.y;
'x' can be generated as a local identifier instead of adding an import for
'y'.
This was at the heart of a bug. In the above case with 'router', the
PartialEvaluator added the identifier 'router' to the Reference generated
(through FFR) to RouterModule.
This is not correct. References that result from FFR expressions may not
have the same value at runtime as they do at compile time (indeed, this is
not the case for ModuleWithProviders). The Reference generated via FFR is
"synthetic" in the sense that it's constructed based on a useful
interpretation of the code, not an accurate representation of the runtime
value. Therefore, it may not be legal to refer to the Reference via the
'router' identifier.
This commit adds the ability to mark such a Reference as 'synthetic', which
allows the PartialEvaluator to not add the 'router' identifier down the
line. Tests are included for both the PartialEvaluator itself as well as the
resultant buggy behavior in ngtsc overall.
PR Close#29387
