Upcoming work to implement import resolution will change the dependencies
of some higher-level classes in ngtsc & ngcc. This necessitates changes in
how these classes are created and the lifecycle of the ts.Program in ngtsc
& ngcc.
To avoid complicating the implementation work with refactoring as a result
of the new dependencies, the refactoring is performed in this commit as a
separate prepatory step.
In ngtsc, the testing harness is modified to allow easier access to some
aspects of the ts.Program.
In ngcc, the main change is that the DecorationAnalyzer is created with the
ts.Program as a constructor parameter. This is not a lifecycle change, as
it was previously created with the ts.TypeChecker which is derived from the
ts.Program anyways. This change requires some reorganization in ngcc to
accommodate, especially in testing harnesses where DecorationAnalyzer is
created manually in a number of specs.
PR Close#27743
A surprising interaction with the MagicString library caused inserted
Ivy definitions to be dropped during the removal of decorators, iff all
decorators on the class could be removed. In that case, the removal
location corresponds with the exact location where Ivy definitions were
inserted into.
This commit moves the removal of decorators to occur before Ivy
definitions are inserted. This effectively avoids the problem, as later
inserted text fragments will be retained by MagicString.
PR Close#27159
Ngcc will now render additional exports for classes that are referenced in
`NgModule` decorated classes, but which were not publicly exported
from an entry-point of the package.
This is important because when ngtsc compiles libraries processed by ngcc
it needs to be able to publcly access decorated classes that are referenced
by `NgModule` decorated classes in order to build templates that use these
classes.
Doing this re-exporting is not without its risks. There are chances that
the class is not exported correctly: there may already be similarly named
exports from the entry-point or the class may be being aliased. But there
is not much more we can do from the point of view of ngcc to workaround
such scenarios. Generally, packages should have been built so that this
approach works.
PR Close#26906
There are a number of variables that need to be passed around
the program, in particular to the renderers, which benefit from being
stored in well defined objects.
The new `EntryPointBundle` structure is a specific format of an entry-point
and contains the compiled `BundleProgram` objects for the source and typings,
if appropriate.
This change helps with future refactoring, where we may need to add new
properties to this object. It allows us to maintain more stable APIs between
the constituent parts of ngcc, rather than passing lots of primitive values
around throughout the program.
PR Close#26906
Previously the ivy definition calls we going directly after the
class constructor function But this meant that the lifecycle
hooks attached to the prototype were ignored by the ngtsc
compiler.
Now the definitions are written to the end of the IIFE block,
just before the return statement.
Closes#26849
PR Close#26856
For each package entry-point there is only one format that
is used to compile the typings files (.d.ts). This will be
either esm2015 or fesm2015 (preferred). So we would not run
any dts processing in the renderer if we are not compiling
the appropriate format.
PR Close#26403
1) The `DecorationAnalyzer now analyzes all source files, rather than just
the entry-point files, which fixes#26183.
2) The `DecoratorAnalyzer` now runs all the `handler.analyze()` calls
across the whole entry-point *before* running `handler.compile()`. This
ensures that dependencies between the decorated classes *within* an
entry-point are known to the handlers when running the compile process.
3) The `Renderer` now does the transformation of the typings (.d.ts) files
which allows us to support packages that only have flat format
entry-points better, and is faster, since we won't parse `.d.ts` files twice.
PR Close#26403
The rendering of typings is not specific to the package
format, so it doesn't make sense to put it in a specific
renderer.
As a result there is no real difference between esm5 and esm2015
renderers, so there is no point in having separate classes.
PR Close#26403
Originally, the ivy_switch mechanism used Bazel genrules to conditionally
compile one TS file or another depending on whether ngc or ngtsc was the
selected compiler. This was done because we wanted to avoid importing
certain modules (and thus pulling them into the build) if Ivy was on or
off. This mechanism had a major drawback: ivy_switch became a bottleneck
in the import graph, as it both imports from many places in the codebase
and is imported by many modules in the codebase. This frequently resulted
in cyclic imports which caused issues both with TS and Closure compilation.
It turns out ngcc needs both code paths in the bundle to perform the switch
during its operation anyway, so import switching was later abandoned. This
means that there's no real reason why the ivy_switch mechanism needed to
operate at the Bazel level, and for the ivy_switch file to be a bottleneck.
This commit removes the Bazel-level ivy_switch mechanism, and introduces
an additional TypeScript transform in ngtsc (and the pass-through tsc
compiler used for testing JIT) to perform the same operation that ngcc
does, and flip the switch during ngtsc compilation. This allows the
ivy_switch file to be removed, and the individual switches to be located
directly next to their consumers in the codebase, greatly mitigating the
circular import issues and making the mechanism much easier to use.
As part of this commit, the tag for marking switched variables was changed
from __PRE_NGCC__ to __PRE_R3__, since it's no longer just ngcc which
flips these tags. Most variables were renamed from R3_* to SWITCH_* as well,
since they're referenced mostly in render2 code.
Test strategy: existing test coverage is more than sufficient - if this
didn't work correctly it would break the hello world and todo apps.
PR Close#26550
Previously we only removed assignments to `Class.decorators = [];`
if the array was not empty.
Now we also remove calls to `__decorate([])`, similarly.
PR Close#26236
Closure compiler requires that the i18n message constants of the form
const MSG_XYZ = goog.getMessage('...');
have names that are unique across an entire compilation, even if the
variables themselves are local to a given module. This means that in
practice these names must be unique in a codebase.
The best way to guarantee this requirement is met is to encode the
relative file name of the file into which the constant is being written
into the constant name itself. This commit implements that solution.
PR Close#25689
Alex Rickabaugh
JoostK
Pete Bacon Darwin