All temporary directory start with `flutter_` and have their random component separated from the name by a period, as in `flutter_test_bundle.YFYQMY`.
I've tried to find some of the places where we didn't cleanly delete temporary directories, too. This greatly reduces, though it does not entirely eliminate, the directories we leave behind when running tests, especially `flutter_tools` tests.
While I was at it I standardized on `tempDir` as the variable name for temporary directories, since it was the most common, removing occurrences of `temp` and `tmp`, among others.
Also I factored out some common code that used to catch exceptions that happen on Windows, and made more places use that pattern.
* Upgrade everything except matcher.
* Roll matcher (and test)
* Adjust tests that depend on flutter:test directly to depend on a shim
* Require use of package:test shim and remove other references to package:test
Disallow calling stop() or cancel() multiple times. This means that
when you use startProgress you have to more carefully think about what
exactly is going on.
Properly cancel startProgress in non-ANSI situations, so that
back-to-back startProgress calls all render to the console.
Disallow calling stop() or cancel() multiple times. This means that
when you use startProgress you have to more carefully think about what
exactly is going on.
Properly cancel startProgress in non-ANSI situations, so that
back-to-back startProgress calls all render to the console.
fuchsia_tester.dart still assumes Dart 1. Previously, it ran tests directly
from source, flutter_platform.dart automatically runs a kernel compile when
operating in Dart 2 mode, but this assumes a functional Dart SDK is available
in the artifacts directly, and fuchsia_tester.dart mocks out the artifacts
directory with an empty temp dir.
Remaining work is:
1. Get the frontend server building as a dependency on Fuchsia.
2. Patch fuchsia_tester.dart to use a valid Dart SDK and frontend server.
This also reverts migration to Dart 2 typedef syntax.
This reverts commit 6c56bb2. (#18362)
This reverts commit 3daebd0. (#18316)
Now that Dart 1 is turned off, reapplying my change to turn on the prefer_generic_function_type_aliases analysis option, and fix all the typedefs to Dart 2 preferred syntax.
Also eliminated the unused analysis_options_repo.yaml file and turned on public_member_api_docs in analysys_options.yaml.
No logic changes, just changing the typedef syntax for all typedefs, and updating analysis options.
* It's time to #deleteDart1 (#18293)
Eliminates support for Dart 1 in flutter_tools, and drops our Dart 1
benchmarks. All commands now run in Dart 1 mode only.
Eliminates --preview-dart-2 / --no-preview-dart-2 support.
* Fix indentation, remove no longer necessary .toList()
* Only push udpated kernel if >0 invalidated srcs
Eliminates support for Dart 1 in flutter_tools, and drops our Dart 1
benchmarks. All commands now run in Dart 1 mode only.
Eliminates --preview-dart-2 / --no-preview-dart-2 support.
Previously, in non-release (i.e. profile) AOT builds, we were setting
--no-checked and --conditional_directives flags. --no-checked is the
default, and we don't make use of conditional directives in Flutter.
Allows users of Fingerprinter to filter the set of paths collected from
the explicitly-specified paths and those collected from depfiles.
In some cases, depfiles are emitted with files that are not present on
the local disk (e.g. the frontend compiler currently emits buildbot
paths for the dart core libraries and dart:ui). These files will not
materially affect whether we need to re-run a build action for which
they are inputs, since they're not present in the filesystem and
therefore cannot change.
This replaces the --prefer-shared-library flag, which falls back to
regular (non-shared-lib) compile if the NDK is not found, with the
--build-shared-library flag, which exits with an error message if the
NDK is not found.
This simplifies the set of allowed code paths through AOT compile,
resulting in better testability and easier-to-follow logic. It also
results in more predictable behaviour for continuous integration and
other scenarios.
This change adds support for armv7, arm64, and universal iOS apps.
This change eliminates iOS target architecture hardcoding (previously
arm64 only) and uses the target architecture(s) specified in Xcode's
ARCHS setting ('Architectures' in Xcode Build Settings).
For universal binaries, set ARCHS to its default value, $(ARCHS_STANDARD).
Note that after changing the architecture in Xcode, developers should
run 'pod install' from the ios subdirectory of their project. A separate
change (that will land before this one) will add support for
automatically detecting project file and Podfile changes and re-running
pod install if necessary.
This change also adds an --ios-arch option to flutter build aot. In iOS
AOT builds (in profile and release mode), this dictates which
architectures are built into App.framework. This flag should generally
be unnecessary to set manually since flutter build aot is typically only
invoked internally by flutter itself.
Adds a Fingerprinter utility class that can be used to compute unique
fingerprints for a set of input paths and build options, compare to the
output of a previous run, and skip the build action if no inputs or
options have changed. The existing Fingerprint class still does all the
heavy lifting. Fingerprinter adds common operations such as
reading/writing/comparing fingerprints and parsing depfiles.
This migrates existing uses of Fingerprint over to Fingerprinter.
This also adds better fingerprinting to AOT snapshotting, which
previously failed to include several options in its fingerprint
(--preview-dart-2, --prefer-shared-library).
Previously, we were incorrectly passing --vm_snapshot_data and
--isolate_snapshot_data options to gen_snapshot in assembly AOT builds.
These only make sense in AOT blob snapshot mode (alongside
--vm_snapshot_instructions and --isolate_snapshot_instructions).
There's very little code-sharing between the two, and what little there
is is concentrated in the GenSnapshotClass and the fingerprint
reading/writing utility methods.
iOS debug builds always run in interpreted mode whether on device or on
simulator. In both cases, we can skip snapshotting and link against an
empty App.framework. Previously, we did this for iOS simulator builds.
This does the same for device builds.
Previously, debug iOS builds used gen_snapshot to generate a core
snapshot, then used 'xxd' to generate C files containing the snapshot
data in buffers named kDartVmSnapshotData and kDartIsolateSnapshotData,
which are then compiled/linked into App.framework. This is unnecessary
since the VM compiled into Flutter.framework already contains this data.
Bugfix: Moves AOT snapshot input verification past where the last input
is added to the inputs list.
Cleanup:
* Extracts _isValidAotPlatform method.
* Moves non-platform-specific logic to the top.
* Moves variable declaration closer to first use, and inlines to a
narrower scope where possible.
This relands #17136, which was reverted in #17142 due to breakage in
on-device iOS debug builds.
Bugfix: Moves AOT snapshot input verification past where the last input
is added to the inputs list.
Cleanup:
* Extracts _isValidAotPlatform method.
* Moves non-platform-specific logic to the top.
* Moves variable declaration closer to first use, and inlines to a
narrower scope where possible.
Moves the kernel compile step to the beginning of the AOT build in a
separate method. This is pre-factoring for iOS universal builds where
the kernel build happens once, but we then snapshot twice: once for
armv7 and once for arm64.
This also writes dependencies to build/kernel_compile.d rather than
build/aot/snapshot.d, since that is immediately overwritten by
gen_snapshot.
This is required for iOS debug builds, but unused otherwise. In theory,
Android debug builds could be run in this mode, but this is historically
untested and adds unnecessary complexity to the code. If ad-hoc testing
is required, it can be patched in when necessary.
This re-lands 8c4f0c0d21 with a fix to
xcode_backend.sh to eliminate the use of --interpreter.
This is required for iOS debug builds, but unused otherwise. In theory,
Android debug builds could be run in this mode, but this is historically
untested and adds unnecessary complexity to the code. If ad-hoc testing
is required, it can be patched in when necessary.
This simplifies tests by allowing the test author to express the outputs
they'd like the snapshotter to generate without creating a whole new
instance and context with the additional overrides.
Also sets the default set of outputs to none, setting them instead in
the setUp function for script snapshotting.
