Consider the following very quirky Angular template, which has both an i18n attribute binding and a property binding to `in`:
```
<cmp [in]="foo" in="bar" i18n-in />
```
What would you expect the above template to do? `TemplateDefinitionBuilder` will emit the following Ivy instructions:
```
// Element constant attributes
consts: () => {
__i18nMsg__('bar', [], {}, {})
return [["in", i18n_0, __AttributeMarker.I18n__, "in"]];
}
// ...
function MyComponent_Template(rf, ctx) {
if (rf & 1) {
// Create mode
i0.ɵɵelement(0, "cmp", 0);
}
}
```
This makes some sense -- we create a single element, and attach an i18n message to the `in` attribute. But is this actually correct? Notice that the property binding is completely missing!
Indeed, Template Pipeline actually produces this code:
```
// Element constant attributes
consts: () => {
__i18nMsg__('bar', [], {}, {})
return [["in", i18n_0, __AttributeMarker.I18n__, "in"]];
}
// ...
function MyComponent_Template(rf, ctx) {
if (rf & 1) {
// Create mode
i0.ɵɵelement(0, "cmp", 0);
} else if (rf & 2) {
// Update mode
i0.ɵɵproperty("in", ctx.foo);
}
}
```
Aha! There's the property binding! Arguably, this is a bug in `TemplateDefinitionBuilder`, but after some discussion on Slack, we have decided to ban this practice in a future Angular version.
For now, we allow Template Pipeline to have slightly different output, but print an error to warn the user of the issue.
PR Close#54063
Fixes that `@defer` blocks weren't recognizing default imports and generating the proper code for them. Default symbols need to be accessed through the `default` property in the `import` statement, rather than by their name.
PR Close#53695
One of the earlier commits separated one-way and two-way bindings which ended up breaking some internal targets, because it changed the assignment order. These changes bring back the old order.
PR Close#54154
In one of the earlier commits, the logic that appends `=$event` before parsing two-way bindings was removed and some validation was added to prevent unassignable expressions from being used. This ended up being problematic, because previously the parser was incorrectly allowing some invalid expressions which users came to depend on. For example, it transformed `[(value)]="a && a.b"` to `a && (a.b = $event)`.
These changes add some special cases for the common breakages that came up during the TGP.
PR Close#54154
Updates the template definition builder to emit the new format for the listener side of two-way bindings.
```js
// Before
listener("ngModelChange", function($event) {
return ctx.name = $event;
});
// After
ɵɵtwoWayListener("ngModelChange", function($event) {
ɵɵtwoWayBindingSet(ctx.name, $event) || (ctx.name = $event);
return $event;
});
```
PR Close#54154
Currently the listener side two-way listeners are parsed by appending `=$event` to the raw expression. This is problematic, because:
1. It can interfere with other expressions (see #37809).
2. It can lead to confusing error messages because users will see code that they didn't write.
3. It doesn't allow us to further manipulate the expression.
These changes remove the logic that appends `=$event` to resolve the issue. There's also some new logic that checks the expression after it has been parsed to ensure that the result is an assignable expression.
Subsequent commits will update the code that emits the expression to add back the `$event` assignment where it's needed.
PR Close#54154
Adds the following new instructions:
* `twoWayBindingSet` - used to assign values inside of the listener side of a two-way binding. Currently a noop, but will come into play later.
* `twoWayListener` - used to bind a two-way listener. Currently calls directly into `listener`, but it may be useful in the future.
PR Close#54154
Currently all the members of `_ParseAST` are public, even though they're all used only within the class. This change marks them as private so that it's explicit which ones are intended to be used outside the class.
PR Close#54154
Reworks the compiler so that it generates a `twoWayProperty` instruction, instead of `property`, for the property side of a two-way binding. Currently the new instruction passes through to `property`, but it'll have some two-way-binding-specific logic in subsequent PRs.
PR Close#54154
During the template parsing stage two-way bindings are split up into a property and event binding. All the downstream code treats these binding the same as their one-way equivalents. For some future work we'll have to distinguish between the two so these changes update the `BoundElementProperty.type` and `ParsedEvent.type` to include a `TwoWay` type. All existing call-sites have been updated to treat `TwoWay` the same as `Property`/`Regular`, but more specialized logic will be added in the future.
PR Close#54065
Previously, defer deps fns names were only prefixed with the component name, meaning that distinct deps fns in the same component would produce a name collision. Now, we take into account the entire template function name when naming inner deps fns.
PR Close#54060
The Template Pipeline is a brand new backend for the Angular compiler, replacing `TemplateDefinitionBuilder`. It generates the Ivy instructions corresponding to an input template (or host binding). The Template Pipeline has an all-new design based on an intermediate representation compiled over many phases, which will allow us to experiment with compiler changes more easily in the future.
With this commit, the template pipeline can now be enabled in any project via the `useTemplatePipeline` TSConfig option. However, it is still disabled by default.
PR Close#54057
In #53591, Andrew added local compliation support for defer blocks. However, this requires the ability to emit pre-generated static defer deps functions. We now also support that feature in Template Pipeline.
PR Close#54043
Similar to signal-based inputs, we support signal-based queries in JIT
by expecting a decorator to be added. This is a consequence of the
design, given that JIT requires query declaration information before
the class is initialized- but ironically there is no way to collect this
information without instantiating the class.
A JIT transform in the Angular CLI will automatically generate these
decorators for testing.
PR Close#54019
Collapses multiple sibling query advance statements into single
query advance invocations. This will help reducing generated code
for directives/components with many queries.
PR Close#54019
Previously, if an ICU was inside a nested i18n root, it would use the nested root to calculate whether it should be applied. Now, we use the root i18n block.
PR Close#54026
This commit ensures that libraries can use signal-based queries, and the
partial compilation output will capture their metadata.
The linker is updated to support parsing this.
Two notes:
1. Older linker versions are not capable of parsing this, so the minimum
version for signal-based queries is adjusted when such are used.
2. We only emit `isSignal` metadata for queries when signal queries are
used. This enables libraries to continue supporting older linker
versions, if signal-based queries are not used.
PR Close#53978
This commit uses the initializer API recognition that we built for
signal-based inputs, and teaches the compiler to recognize class members
that refer to `viewChild`, `viewChildren`, `contentChild` or
`contentChildren`. Those will declare signal-based view or content queries.
PR Close#53978
This commit introduces the compiler output generation for signal-based
queries. Signal-based queries will have new creation-mode instructions
and update instructions to advance the current query indices in the
global shared context.
An output like the following is the expected output for signal-based
queries:
```
i0.ɵɵdefineComponent({
viewQuery: function App_Query(rf, ctx) {
if (rf & 1) {
i0.ɵɵviewQuery(ctx.d, _c0, 5);
i0.ɵɵviewQuerySignal(ctx.ds1, _c0, 5);
i0.ɵɵviewQuerySignal(ctx.ds2, _c0, 5);
}
if (rf & 2) {
let _t;
// only change-detected queries need explicit refresh
i0.ɵɵqueryRefresh(_t = i0.ɵɵloadQuery()) && (ctx.d = _t.first);
// we bump up current query index by 2 positions since there are 2 signal-based queries
i0.ɵɵqueryAdvance(2);
}
…
},
…
});
```
Note: For now, the collapsing of multiple advance instructions is not
implemented. This will be a follow-up.
Note 2: A couple of query helpers are now in their own file. This makes
it easier to focus on query-specific compiler code. The new function is
called `createQueryCreateCall`, which is a modified variant of the
existing function that previously only generated query parameters.
PR Close#53978
This commit adds extra logic to produce a diagnostic in case `@Component.deferredImports` contain types from imports that also bring eager symbols. This would result in retaining a regular import and generating a dynamic import, which would not allow to defer-load dependencies.
PR Close#53899
This commit updates the logic of the `TemplateDefinitionBuilder` to support local compilation and generate a single dependency function for all explicitly deferred deps within a component.
PR Close#53591
As part of testing we did accidentally use `bitwiseAnd` for the input
flags, given we started without an extra flag for `HasTransform`.
This commit teaches the compiler to support emitting bitwise OR
and uses it when combining input flags, fully re-enabling transforms
for signal components after the new flag mechanism was introduced in
previous commits.
PR Close#53808
This commit changes the `HasTransform` flag to be only concerned with
decorator inputs. This allows us to automatically detect signal input
transforms without reliance on the flag, resulting in less complexity in
the compiler (as outlined in the design doc) and various other places,
while it also allows us to simplify JIT support for signal inputs
because there would be no need to capture the "hasTransform" state in
the decorator so that JIT can generate the according input flags.
`isSignal` will still persist as an input flag to allow for monomorphic
and highly efficient distinguishing at runtime, whether an input is
signal based or not. JIT transform will also need to propagate this
information to the runtime somehow.
PR Close#53808
We are adding internal support for declaring signal inputs via the
`@Input` decorator. This is needed for JIT unit testing, or JIT
applications.
In JIT, Angular is not able to recognize signal inputs due to the
lack of static reflection metadata. Decorators attach their information
on the class- without it needing to be instantiated. This allows Angular
to know inputs when preparing/generating the directive definition. With
signal inputs this is not possible- so we need a way to tell Angular
about inputs for JIT applications. We've decided that this is not
something users should have to deal with, so a transform will be added
in a follow-up that will automatically derive/and add the decorators
for signal inputs when requested in JIT environments.
PR Close#53808
We generate `advance` instructions before most update instructions and the majority of `advance` calls are advancing by one. We can save some bytes for the most common case by omitting the parameter for `advance(1)` altogether.
PR Close#53845
Instead of computing the bit input flags at compile-time and inling
the final bit flag number, we will use the `InputFlags` enum directly.
This is a little more code in the compiler side, but will allow us to
have better debuggable development code, and also prevents problems
where runtime flag bitmasks differ from the compiler flag bitmasks.
This is in practice a noop for optimized applications as the enum values
would be inlined anyway. This matches existing compiler emit for e.g.
change detection strategy, or view encapsulation enums.
PR Close#53571
This commit introduces a new enum for capturing additional metadata
about inputs. Called `InputFlags`. These will be built up at compile
time and then propagated into the runtime logic, in a way that does
not require additional lookup dictionaries data structures, or
additional memory allocations for "common inputs" that do not have any flags.
The flags will incorporate information on whether an input is signal
based. This can then be used to avoid megamorphic accesses when such
input is set- as we'd not need to check the input field value. This also
avoids cases where an input signal may be used as initial value for an
input (as we'd not incorrectly detect the input as a signal input then).
The new metadata emit will be useful for incorporating additional
metadata for inputs, such as whether they are required etc (although
required inputs are a build-time only construct right now- but this is a
good illustration of why input flags can be useful). An alternative
could have been to have an additional boolean entry for signal inputs,
but allocating a number with more flexible input flags seems more future
proof and more reasonable andreadable.
More information on the megamorphic access when updating an input
signal
https://docs.google.com/document/d/1FpnFruviKb6BFTQfMAP2AMEqEB0FI7z-3mT_qm7lzX8/edit.
PR Close#53571
In #52931, Kristiyan fixed a TemplateDefinitionBuilder bug in which derived alias variables in for loops (`$even`, `$first`, etc) were referring to the wrong level of nested `@for` block. (These variables are unique because they become inlined expressions, and are not "real" context variables.) He fixed this by appending level information to the generated alias name.
Template Pipeline actually suffered from the same bug. We fix it in a very similar way -- in particular, whenever these derived context variables are used, we make them depend on versions of `$index` and `$count` that have been suffixed with the xref of the enclosing repeater.
I have added a few more pipeline goldens, because we are not quite as clever as TDB about only generating the duplicate suffixed index and count variables when inside nested loops. This is fine, since in the long run, we want to refactor it more fundamentally.
I have also added a TODO to fix this more rigorously. In particular, it would be nice if we had proper support for shadowed variables, as well as unlimited levels of variables depending on one another.
PR Close#53662
Template pipeline previously mangled CSS property names like
`--camelCase` when used in host style bindings. Note: It still *does*
mangle these names in static style attrs, both in host bindings and on
elements. This is clearly wrong, but is consistent with what TDB does
today.
PR Close#53665
It's possible for attributes to have a namespace, we need to handle this
possiblity for both attribute instructions and attributes extracted to
the consts array.
PR Close#53646
The way we were handling ICU placeholders was not compatible with using
interpolations on attributes of elements inside the ICU. This change
refactors the handling of ICU placeholders and unifies the way
expression and tag placeholders work inside ICUs.
The new approach modifies the ingest logic to add the placeholder on to
the TextOp rather than the TextInterpolationOp. This is because, in
ICUs, we may need multiple i18n expressions created from the
interpolation expressions to roll up into the same placeholder. ICUs
essentially do the interpolation at compile time, combining the static
strings with special placeholder strings that represent the expression
values.
PR Close#53643
Consider a case when an explicit `this` read is inside a template with a context that also provides the variable name being read:
```
<ng-template let-a>{{this.a}}</ng-template>
```
Clearly, `this.a` should refer to the class property `a`. However, in today's Angular, `this.a` will refer to `let-a` on the template context.
Amazingly, both TemplateDefinitionBuilder and the Typecheck block have the same bug, and are consistent with each other! This is because `ImplicitReceiver` extends `ThisReceiver` in the parser AST, which is an insane gotcha.
In this commit, I patch the template pipeline to emulate this behavior as well.
To actually fix this nastiness, we have to:
- Update `ingest.ts` in the Template Pipeline (see the corresponding comment)
- Check `type_check_block.ts` in the Typecheck block code (see the corresponding comment)
- Turn off legacy TemplateDefinitionBuilder
- Fix g3, and release in a major version
PR Close#53594
`ng-content` elements, and thus their corresponding projection instructions, can have many attributes on them. Some of these attributes may result in special behavior. For example, `ngProjectAs` and `i18n-foo` both result in special const collection, into the approprate BindingKind slot in the const array. Additionally, `i18n-foo` needs to recieve all the additional i18n attribute processing.
We solve this by subjecting `ng-content` attributes to all the same pipeline logic that applies to attributes on elements, and then allow the element const collection phase to collect them.
PR Close#53594
For regular templates, any listener will have its name const collected into the bindings section of the element consts.
In contrast, host bindings omit listener names from their hostAttrs. This is a strange and inconsistent behavior, so we hide it behind a compatiblity mode flag.
PR Close#53594
We has some special behavior for naming identifiers in Template Pipline, for the sake of compatibility with TDB's source maps tests. However, this has the potential to cause a variable name collision in a particular special case (when the identifier is `ctx`). We add a special check for this, and also tuck all the backwards-compatible naming code inside a compatibility block.
PR Close#53594
It's possible for the user to create a host attrbiute binding with a
name that makes it _look_ like a class binding `{['class.foo']: ''}`, we
were previously treating these as actual class property bindings. This
change fixes the logic so that only true property bindings cam be
converted to class property bindings.
Note: A user who added an attribute like the above almost certainly
intended to create an actual class property binding. It would be nice if
we could add a diagnostic to warn them about this.
PR Close#53626
