Currently the TS generated output (not JIT), breaks at runtime because
we emit all variables using the `Final` modifier. i.e. constants.
This breaks shallow references when we run our acceptance signal
tests using AOT.
Currently, a listener on an element containing a dash, will result in
runtime errors because the function name will be generated using a dash.
e.g.
```
function MyApp_Template_some-comp_bla_0() {}
```
throwing with a syntax error due to the dash. We fix this by re-using
the sanitize identifier function from the current template definition
builder.
Compiler portion for supporting content queries in signal components.
This is based on the preparation work, for detecting such queries. See
previous commits.
With the recent changes for local refs, we now also generate all
variables in the create block.
This works overall, but we end up with incorrect variable optimization
because calls of `nextContext` in listener functions end up tricking
the optimization logic into thinking that the outer `nextContext` calls
in the create block are _relevant_ for the listener function body.
e.g.
```
if (create) {
const _bla = nextContext();
i0.listener(() => {
restoreView();
const _ctx1 = nextContext();
/* do somth with _ctx1 */
});
}
```
`_bla` can safely be removed here, even though a `ContextRead/Write` has
occurred inside the listener operation. We process listeners separately.
We recently changed the pure function code to look for
all operations in a view. This ended up causing pure functions
to be generated for expressions in listeners. This might be fine
but diverges from the existing TDB and breaks compliance tests.
Culprit commit: https://github.com/angular/angular/commit/a61e41c88049eb12a0b23b1e7fa49a16b0eb91d7
New implementation of type-checking for signal inputs. Much simpler, and
without causing signficant type check block changes that would require
more changes to the language service completion logic / or rewrites of
more than 50+ handwritten code output tests for type check blocks.
See previous commits for other explored solutions.
NOTE: This is very experimental and manual type check block code is not
updated. All of this complexity arises from the fact of attempting to
support bindings where they go to the same directive with multiple
inputs of the same name.
Potentially we can deprecate this behavior for input signal based
components and avoid this complexity. Alternative solution is to simply
repeat expression diagnostics if a binding goes to two same fields.
This commit adds support for local refs in signal property bindings, by
generating variables in the creation block. To implement this support:
* a new instruction shallowReference() is introduced, which is not tied to
the current view context.
* variables are now generated for creation mode as well as update mode
* a new template pipeline pass (`phaseCreationVarColocation`) handles
ordering of reference variable declaration and assignment in the creation
block (guaranteeing references are only read after declarations).
Previously the compiler output AST did not support using
`o.BinaryExpression` for assignments (as TypeScript does). Instead
`o.WriteVarExpr` is defined, which only allows writing to string named
variables.
This commit adds the `ConsumesVarsTrait` to `PropertyCreateOp` which causes
vars to be counted for use in the virtual instructions it might add. This
allows the DOM test to pass.
Previously the template pipeline would prefer to use a local variable
definition in favor of ctx. This commit improves the logic to prefer `ctx`
for the root view even if a variable for the context is available.
⚠️Disclaimer⚠️ this PR implements syntax that is still in an open RFC. It will be adjusted once the RFC is closed.
These changes implement the `BlockGroup` and `Block` AST nodes that will then be used to generate instructions based on the new syntax. A `BlockGroup` is a container for `Block` instances. The first block of a block is always implicit and required while any subsequent blocks are optional.
PR Close#50953
⚠️Disclaimer⚠️ this PR implements syntax that is still in an open RFC. It will be adjusted once the RFC is closed.
These changes extend the lexer to recognize the concepts of a block group (`{#foo paramA; paramB}{/foo}`) and a block (`{:foo paramA; paramB;}`) which will be useful later on for the control flow and defer proposals. Block groups can be used anywhere and require a closing tag while block can only be used inside of a block.
The idea is that in the next PRs the markup AST will be expanded to have some more specialized node like `ConditionalBlock` or `DeferBlock` which will then be turned into instructions.
PR Close#50895
This commit adds the ability to generate attribute instructions as a result of property bindings such as `[attr.foo]='bar'` or `attr.foo='{{bar}}'`. "Singleton" interpolations, such as the previous example, will also be transformed into a simple `attribute` instruction.
PR Close#50818
The new interface is discrete-unioned with the existing interface to cover the cases for local and global (i.e., full and partial) compilation modes.
This change of interface required some adjustmeents cross repo which explains the changes made to other files.
PR Close#50577
Previously, the template pipeline save/restore view logic only added the
save/restore operation in listeners inside embedded views. However, this
operation is also needed if local refs are accessed within a listener body.
This commit updates the logic to detect more accurately whether save/restore
is needed.
PR Close#50834
The expression `a()?.b` should expand into `(tmp = a()) === null ? null : tmp.b`, in order to avoid calling the function `a()` twice.
This commit modifies the null-safe-expansion algorithm to emit temporary assignments, and provides the reification code to actually generate the declarations, assignments, and reads.
Note also that, with our bottom-up algorithm, there are some tricky cases when a function call exists inside an indexed access, such as `f1()?.[f2()?.a]?.b`. We add some special logic to avoid generating a double-assignment to the temporary storing the result of `f2()`.
Finally, there are opportunities to reuse the same temporary in expressions like `a?.[f()]?.[f()]`. We save this for the next commit.
PR Close#50688
We have a lint rule configured that enforces that any abstract member
implementation uses an explicit `override` identifier. This ensures that
downstream classes will have errors if the parent abstract class
suddenly removes the abstract member.
The lint rule, living in the dev-infra repository, occasionally does
miss some places due to a temporary TS version mismatch that causes
syntax kind indices to be different. Looks like we are now matching
again and there is a new lint failure that got introduced recently. This
commit fixes that error.
PR Close#50772
If a library is compiling with Angular v16.1.0, the library will break
for users that are still on Angular v16.0.x. This happens because the
`DirectiveDeclaration` or `ComponentDeclaration` types are not expecting
an extra field for `signals` metadata. This field was only added to the
generic types in `16.1.0`- so compilations fail with errors like this:
```
Error: node_modules/@angular/material/icon/index.d.ts:204:18 -
error TS2707: Generic type 'ɵɵComponentDeclaration' requires between 7 and 9 type arguments.
```
To fix this, we quickly roll back the code for inserting this metadata
field. That way, libraries remain compatible with all v16.x framework
versions.
We continue to include the `signals` metadata if `signals: true` is set.
This is not public API anyway right now- so cannot happen- but imagine
we expose some signal APIs in e.g. 16.2.x, then we'd need this metadata
and can reasonably expect signal-component library users to use a more
recent framework core version.
PR Close#50714
Angular's null-safe access operators differ from Javascript's built-in semantics, in that they short-circuit to `null` instead of `undefined`. This necessitates providing a custom transformation, instead of relying on Typescript or Javascript itseld.
The old TemplateDefinitionBuilder uses a top-down approach based on the Visitor pattern, in which it recursively extracts the left-most safe access, and hoists it to a null check at the top. See `expression_converter.ts` for details.
In this commit, we replace that approach with a new bottom-up algorithm, as part of the template pipeline. This requires an intermediate expression type to represent the not-yet-expanded ternary operators, and is split into its own pass.
Null-safe function calls are not yet implemented, since they will rely on a future temporary variable allocation pass.
Co-authored-by: Alex Rickabaugh <alxhub@users.noreply.github.com>
PR Close#50594
Create a pass that expands nullish coalescing operators into null checks.
This is not yet finished because we need to emit temporary variable assignments, which we will do using a future temporary variable allocation pass. Also, TemplateDefinitionBuilder is a bit quirky, and we still need to exactly match its behavior.
Nevertheless, this is good enough to prevent the diffs from getting ruined as a result of nullish coalescing operations.
PR Close#50594
It is sometimes useful to clone an expression tree, in order to copy it and mutate it in a phase, without affecting other subtrees due to the copy-by-reference.
PR Close#50594
Refactor attribute and property binding ingestion and add an attribute extraction phase
Co-authored-by: Alex Rickabaugh <alxhub@users.noreply.github.com>
Co-authored-by: Dylan Hunn <dylhunn@users.noreply.github.com>
Only add the value to the ElementAttributes map for style and attribute kinds
Other kinds should not have their value represented in the consts array
Add missing attribute ingesiton for templates
Unify how template and element bindings are ingested
This resolves the issue of missing listener attributes on templates. In
order to avoid emitting extraneous instructions, listener ops on
templates are stripped in the attribute extraction phase instead.
Handle different binding types separately in ingest
Cleanup code and comments
Disable test that fails on new explicit error.
Previously the test was passing because ingestPropertyBinding treated
attribute bindings as normal bindings which happened to be ok for the
particular test. Now there's an explicit error that attrbiute bindings
aren't yet handled which causes the test to fail
Address feeback
PR Close#50664
Fixes an error that surfaced in #50580 where the compiler was throwing an error in JIT mode when reading the result of `compileDirectiveDeclaration`. It is caused by the fact that input transform functions were being passed around directly, instead of being wrapped in an AST node.
PR Close#50600
This commit adds end-to-end support for pipes in the template pipeline. This
support works across multiple steps:
1. Pipes are first ingested as `ir.PipeBindingExpr`s during the ingest step.
2. A "pipe creation" phase inserts operations to instantiate each required
pipe, based on the presence of those `ir.PipeBindingExpr`s.
3. A "variadic pipe" phase transforms pipes with more than 4 arguments into
variadic pipe bindings, which use a literal array argument. This literal
array will be later memoized into a pure function invocation.
4. A special phase (`phaseAlignPipeVariadicVarOffset`) reconciles a
difference in variable slot assignment logic between the template pipeline
and the `TemplateDefinitionBuilder`, to ensure that the pipeline output can
pass the existing tests. This phase should not affect runtime semantics and
can be dropped once matching output is no longer necessary.
5. Reification emits pipe instructions based on the argument count.
PR Close#50118
The logic for `insertBefore` in template pipeline operation lists has a bug
when inserting at the end of a list. This commit fixes the safety assertions
to be more accurate.
PR Close#50118
Paul Gschwendtner