See https://github.com/flutter/flutter/issues/159154.
See https://github.com/flutter/flutter/pull/159169.
Before this PR, it appeared we were accidentally leaking (keeping
active) `flutter_tester` instances (or any test device) after `flutter
run` completion, even if the runner was not explicitly detached. I
_think_ this is a bug, but I'll check with the tools team and possibly
@jonahwilliams before finalizing this.
/cc @jason-simmons
---------
Co-authored-by: Andrew Kolos <andrewrkolos@gmail.com>
The dart-lang/native repository contains a `Target` class that is almost
not needed anymore. The remaining uses are mainly due to kernel asset
mapping (which we may be able to remove in the future).
This PR removes usages of that `Target` (in favor of `Architecture`)
class in most places in flutter tools.
This makes the code also cleaner because we no longer have an implicit
assumption that
a `List<Target>` all belong to the same operating system.
Closes https://github.com/flutter/flutter/issues/20949.
Signals (such as SIGTERM or SIGKILL) end up flowing through
`exitWithHooks`, which in turn, after running hooks, call `exit().` That
means, as a result, any `try { } finally { }` guarded execution may
_not_ run, which happens to also be how `flutter_tester` instances are
cleaned up if they have not terminated.
This PR adds in-progress `flutter_tester` runs (or any platform
`flutter_platform` supports) to the shutdown hooks, guaranteeing that
the finalizers (which in turn, kill the process) are _always_ executed
as long as either the test completes, _or_ `exitWithHooks` is called.
The existing integration tests (`integration.shard/test_test.dart`)
still pass as well.
It's possible that the tool can be in the process of shutting down,
which could result in the temp directory being deleted after the
shutdown hooks run before we check if `output` exists. If this happens,
we shouldn't crash but just carry on as if no devices were found as the
tool will exit on its own.
Fixes https://github.com/flutter/flutter/issues/141892
---------
Co-authored-by: Andrew Kolos <andrewrkolos@gmail.com>
Currently the `NativeAsset` target in flutter tools is responsible for
two things:
* performing the dart build (in the app as well as all transitive pub
dependencies)
* taking output (shared libraries) from this build and copying them
around
This intermingling of responsibilities leads to more complex code and
potentially unnecessary work: If the source code changed (e.g. `.c`
files change) we have to run the dart build again. But doing so may
result in the same shared libraries (e.g. adding comments to the `.c`
code). Currently we're going to copy the shared libraries despite them
having not changed, which then may cause upstream things to be dirtied
(if it's based on timestamp of files) and re-built.
Instead this PR splits this `NativeAsset` into the two orthogonal pieces
* `DartBuild` target that is responsible for the dart build
* `InstallCodeAssets` that is responsible for copying shared libraries
to the right place and producing a `native_assets.yaml`.
This decoupling is also preparation for a future where a dart build can
produce other kinds of assets (e.g. data assets) and is used in the web
build as well. (The web build would use `DartBuild` but not
`InstalCodeAssets`).
After the dart build is done, the flutter tool has to bundle the
produced shared libraries, which it does that by copying them around.
Though the code assumed that all code assets are shared libraries to be
bundled, whereas in fact one can have code assets without any actual
code (ones that are installed on the target system already or artificial
code assets whose symbols get resolved from executable / process).
=> Using non-bundled code assets currently results in null pointer
exceptions and/or cast errors.
=> We update the copy code to only operate on code assets that have a
shared library to bundle.
We also update the copy routines by removing copy&past'ed - but slightly
different - printing code into the shared caller function.
In release builds linking of native assets is enabled. The build step is
only a temprary step, it's output is given to the link step which then
returns all final assets (effectively a map-reduce system). Assets that
aren't sent to a specific linker could be conceptually viewed as sent to
a linker that will emit it's input as-is.
=> The code currently took output of build & link step and therefore
accumulated assets that aren't explicitly sent to a linker twice.
=> This led to performing work twice for those (e.g. copying them twice)
This PR changes this such that if linking mode is enabled, we only rely
on the output of the link phase.
That in return means many tests that mock the native asset builds need
to be updated to mock the output of the link phase.
This started happening after moving DDS to launch from `dart
development-service` rather than `DartDevelopmentService` (see
33b402d24c2131dcf97e272ec3d59c0249762e7f) . This state error was
originally meant to be thrown when some string parsing failed, but is
currently wrapping the `DartDevelopmentServiceException`.
Fixes https://github.com/flutter/flutter/issues/158537
I was poking around in logger.dart when I noticed a few classes have unnecessary exposed member variables. This PR reduces visibility of these in the pursuit of making these classes slightly easier to grok.
<details>
<summary> Pre-launch checklist </summary>
</details>
Cleans up https://github.com/flutter/flutter/pull/155800. In summary, `ResidentRunner`/`FlutterDevice` have branching behavior around logging that depends on the type of `DeviceLogReader` on the `FlutterDevice` instance. Let's instead move this behavior to the `DeviceLogReader` implementations.
My apologies for the large diff. Much of this is a refactor that was a bit too difficult to separate into its own commits.
Here are the main two changes
* Replaces the mutable `connectedVmService` field on the `DeviceLogReader` class with a new method `provideVmService`. This serves largely the same purpose as the mutable field, but it allows for asynchronous code. This is where we put the logic that used to exist in `FlutterDevice.tryInitLogReader`.
* Removes the `tryInitLogReader` method from `FlutterDevice`. This method served to set the `appPid` field on the `FlutterDevice`'s `DeviceLogReader` instance. This was only used in the case of Android to filter out logs unrelated to the flutter app coming from the device, so we can move this logic to `AdbLogReader`'s implementation of `provideVmService`.
Closes https://github.com/flutter/flutter/issues/158012.
This is (effectively) a user-facing NOP, which is exchanging an
on-by-default command-line argument (`--implicit-pubspec-resolution`)
for an off-by-default global feature flag
(`explicit-package-dependencies`). It matches the mental model better,
is less painstaking to maintain and feed throughout, and will be easier
to globally flip on/off in a future PR.
---------
Co-authored-by: Andrew Kolos <andrewrkolos@gmail.com>
In the future, it will be possible for Swift Package Manager to be enabled on one but not all platforms (see https://github.com/flutter/flutter/issues/151567#issuecomment-2455941279).
This moves the `usesSwiftPackageManager` property from the platform-agnostic `Project` to the platform-specific `IosProject` and `MacOSProject`.
This will allow the `IosProject` and `MacOSProject` to return different values for `usesSwiftPackageManager` in the future. For now, both of these projects will always return the same value.
Part of https://github.com/flutter/flutter/issues/151567
In the future, it will be possible for Swift Package Manager to be enabled on one but not all platforms (see https://github.com/flutter/flutter/issues/151567#issuecomment-2455941279).
This updates the `.flutter-plugin-dependencies` file format to separate iOS's and macOS's SwiftPM enablement. For now, these platforms will always have the same value.
This `.flutter-plugin-dependencies` file is read by our CocoaPods scripts to determine whether we should use CocoaPods or not to inject plugins.
Part of https://github.com/flutter/flutter/issues/151567
Fixes https://github.com/flutter/flutter/issues/153777.
To summarize that issue, `ErrorHandlingFileSystem.systemTempDirectory` calls [`LocalFileSystem.systemTempDirectory`](45c8881eb2/packages/flutter_tools/lib/src/base/file_system.dart (L229)), which makes a `Directory.createSync` call, which can throw exceptions that _should_ be handled and result in a graceful tool exit (e.g. insufficient storage). However, we aren't catching those, hence this issue.
All we need to do is wrap that call with the `FileSystemException`-handling logic we already have in the tool. See the diff.
I don't think I'll be cherry-picking this since 1) it's not an extremely common crash and 2) users can probably pick apart the crash message and figure out that they need to clear some storage space to proceed.
<details>
<summary> Pre-launch checklist </summary>
</details>
Removes duplicated constants and ensures consistency by using package:vm_service as a source of truth for RPC error codes for requests made with package:vm_service.
Work towards https://github.com/flutter/flutter/issues/56591.
I explicitly want an LGTM from @andrewkolos @jmagman @jonahwilliams before merging.
---
After this PR, `<Plugin>.isDevDependency` is resolved based on the following logic, IFF:
- The plugin comes from a package _A_ listed in the app's package's `dev_dependencies: ...`
- The package _A_ is not a normal dependency of any transitive non-dev dependency of the app
See [`compute_dev_dependencies_test.dart`](51676093a3/packages/flutter_tools/test/general.shard/compute_dev_dependencies_test.dart) for probably the best specification of this behavior.
We (still) do not write the property to disk (i.e. it never makes it to `.flutter-plugins-dependencies`), so there is no impact to build artifacts at this time; that would come in a follow-up PR (and then follow-up follow-up PRs for the various build systems in both Gradle and Xcode to actually use that value to omit dependencies).
Some tests had to be updated; for the most part it was updating the default `ProcessManager` because a call to `dart pub deps --json` is now made in code that computes what plugins are available, but there should be no change in behavior.
_/cc @jonasfj @sigurdm for FYI only (we talked on an internal thread about this; see https://github.com/dart-lang/sdk/issues/56968)._
_/cc @camsim99 @cbracken @johnmccutchan for visibility on the change._
Almost all of the code is just adopting to changes to the APIs of
`package:native_assets_builder`, `package:native_assets_cli` and
`package:native_toolchain_c`
There's only two semantic changes
* Removes a test that checks for a verification error if a build hook
produces a static library if the preferred linking mode is dynamic:
=> The test is written in a very hacky way. By monkey patching the build
config.json that flutter build actually made. This monkey patching
relies on package:cli_config which is now no longer used.
=> The actual code that checks for this mismatch lives in
dart-lang/native repository and is tested there. So there's really no
need to duplicate that.
* The `package:native_assets_builder` no longer knows about code assets.
This is something a user of that package (e.g. flutter tools) adds. Now
the dry-run functionality will invoke build hooks who produce code
assets without an architecture.
=> The `package:native_assets_builder` used to expand such a code asset
to N different code assets (one for each supported architecture)
=> This logic was now moved to flutter tools. => In the near future
we're going to this dry-run complexity, which will then also get rid of
this uglyness (of expanding to all archs of an OS).
While doing some hacking on `Cache` in https://github.com/flutter/flutter/pull/158081, I noticed that [`Cache.test`](de93182753/packages/flutter_tools/lib/src/cache.dart (L139)) allows the caller to tell Cache to use some given directory as the flutter root (instead of depending on the static global [`Cache.flutterRoot`](4f3976a4f2/packages/flutter_tools/lib/src/cache.dart (L206))). This has a default value, `/cache`. However, `/cache` is an unintuitive name for the root directory of a Flutter installation.
This led to confusion when updating some tests. I wanted to create `/bin/cache/engine-dart-sdk.stamp` for tests, but in reality I needed to create `/cache/bin/cache/engine-dart-sdk.stamp`.
This PR changes this default to the current directory of the file system (which I'm guessing is `/` for all intents and purposes).
<details>
<summary> Pre-launch checklist </summary>
</details>
This is to handle Google testing failures for https://github.com/flutter/flutter/pull/158126.
We want to update G3 to provide this parameter before merging the full change.
Testing is not needed because the change is no-op.
Work towards https://github.com/flutter/flutter/issues/157819. **No behavior changes as a result of this PR**.
Based on a proof of concept by @jonahwilliams (https://github.com/flutter/flutter/pull/157818).
The existence of this flag (which for the time being, defaults to `true`) implies the following:
1. The (legacy, deprecated) `.flutter-plugins` file is not generated:
https://docs.flutter.dev/release/breaking-changes/flutter-plugins-configuration
2. The (legacy, deprecated) `package:flutter_gen` is not synthetically generated:
https://github.com/flutter/website/pull/11343
(awaiting website approvers, but owners approve this change)
This change creates `useImplicitPubspecResolution` and plumbs it through as a required variable, parsing it from a `FlutterCommand.globalResults` where able. In tests, I've defaulted the value to `true` 100% of the time - except for places where the value itself is acted on directly, in which case there are true and false test-cases (e.g. localization and i10n based classes and functions).
I'm not extremely happy this needed to change 50+ files, but is sort of a result of how inter-connected many of the elements of the tools are. I believe keeping this as an explicit (flagged) argument will be our best way to ensure the default behavior changes consistently and that tests are running as expected.
These are the versions we use in test, as of https://github.com/flutter/flutter/pull/157617.
Motivated by noticing a warning with the old template version:
```
This Android Gradle plugin (8.1.0) was tested up to compileSdk = 33 (and compileSdkPreview = "UpsideDownCakePrivacySandbox").
You are strongly encouraged to update your project to use a newer
Android Gradle plugin that has been tested with compileSdk = 35.
```
Documents when to use `throwToolExit` and how to use it.
Replaces every invocation of `throw ToolExit` with `throwToolExit` and makes the former impossible; this is so that every user will at least (hypothetically) have the chance to read the documentation attached to `throwToolExit` (and if we change parameters in the future they will all flow through one place).
Reverts: flutter/flutter#157032
Initiated by: gmackall
Reason for reverting: https://github.com/flutter/flutter/pull/157032#issuecomment-2436336078
Original PR Author: gmackall
Reviewed By: {reidbaker, bartekpacia}
This change reverts the following previous change:
I recently noticed the following log when building an app in verbose mode:
```
This Android Gradle plugin (8.1.0) was tested up to compileSdk = 33 (and compileSdkPreview = "UpsideDownCakePrivacySandbox").
You are strongly encouraged to update your project to use a newer
Android Gradle plugin that has been tested with compileSdk = 35.
```
It looks like AGP would like us to use a newer AGP version if we want to use compileSdk 35 (which we do). This pr upgrades the tests, in advance of updating the templates.
