| Commit message (Collapse) | Author | Age |
|
|
|
|
| |
RELNOTES: None.
PiperOrigin-RevId: 179425421
|
|
|
|
|
|
|
|
| |
aapt2 always gets the complete, unfiltered resources, and then filters them in
execution. Save time by not uselessly filtering in analysis.
RELNOTES: none
PiperOrigin-RevId: 178397137
|
|
|
|
|
|
|
|
|
|
|
| |
In the next review, to handle issues around density filtering, ResourceFilterFactory will return another object that actually handles filtering. To ensure stuff is named properly, rename ResourceFilter to ResourceFilterFactory now so that the new class can be called ResourceFilter.
This is a straightforward automated refactor, followed with some automated reformatting to make linting happy.
I used the name ResourceFilterFactory, rather than the more concise ResourceFilters, as this class actually contains state (both around what filtering should currently do and about what resources were filtered out) and isn't just a helper class.
RELNOTES: none
PiperOrigin-RevId: 174049618
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Recent changes:
1) Started passing all resources to processors, ignoring the filtered
ResourceContainers, and
2) Started loading an unfiltered LocalResourceContainer into binary resource
processing, in addition to the filtered container.
Fix both of these. To fix the former, we need to split the misleadingly-named
'transitiveResourceRoots' (actually transitive resource and assets artifacts)
into transitive resources and assets.
Update resource filtering tests to catch bugs like these.
Also, rename getters for resource containers to make clear that they are not
getters for resources.
Finally, document some weirdness and partially-completed migrations encountered
as part of investigating these issues, and add appropriate TODOs and
deprecation.
RELNOTES: None
PiperOrigin-RevId: 172929936
|
|
|
|
|
|
|
| |
files.
RELNOTES: None.
PiperOrigin-RevId: 172485548
|
|
|
|
|
| |
RELNOTES: None
PiperOrigin-RevId: 170136145
|
|
|
|
|
|
| |
This is a trivial change with a large file footprint.
PiperOrigin-RevId: 169169864
|
|
|
|
|
|
| |
This is necessary for the upcoming Skylark implementation of param files.
PiperOrigin-RevId: 168744486
|
|
|
|
|
|
|
| |
First commit in series, the second will make the necessary changes to the action.
RELNOTES: None
PiperOrigin-RevId: 168286654
|
|
|
|
|
|
| |
This whole CL was done using IDE refactoring tools and should be safe.
PiperOrigin-RevId: 168275575
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
In converting SpawnAction.Builder (multi-thousand line CL) users directly to CustomCommandLine I didn't like the resulting loss of readability, and the methods didn't feel very discoverable. Unless it's very convenient and readable to use CustomCommandLine, people will resort to non-memory efficient patterns by default. I'm holding that CL for this, which should offer a nicer interface.
This CL removes VectorArg from the API contact surface area, instead creating 64 overloads for every valid combination of parameters. Pretty sad, but the methods dispatch straight to internal helper methods so it's mostly boilerplate to the tune of +400 LOC.
Other changes:
* Change ImmutableCollection -> Collection and copy the args directly into the internal args vector. Saves on collection object overhead and saves users from having to create immutable copies.
* Change some names, notably add -> addAll for collection methods
* Create additional missing overloads
* Fix JavaDoc
RELNOTES: None
PiperOrigin-RevId: 165943879
|
|
|
|
|
|
|
|
|
|
|
|
| |
Allowing add(Object) is too loose and can easily lead to programmer mistakes.
Because of type erasure, we can't use the same overload name for (eg.) add(NestedSet<String>) and add(NestedSet<Artifact>). The API is overhauled to use the same terms everywhere, eg. "add", "addPaths", "addExecPaths". This is similar to how it used to be a few CLs ago.
The API is overhauled to make sure it's consistent for all types. While tedious, the facade methods immediately dispatch to internal helpers, so implementation wise it's not too heavy.
While large, this CL is almost entirely an automated refactor.
PiperOrigin-RevId: 165358287
|
|
|
|
|
|
|
|
|
| |
This enforces certain memory-efficient patterns. For deliberate use of dynamic strings, explicitly named overloads are introduced, with javadoc that guides the programmer into making the right choice.
This CL is a memory no-op on benchmarks, but it tries to prevent backslide by making sure programmers make conscious choices when they construct their command lines.
RELNOTES: None
PiperOrigin-RevId: 165185997
|
|
|
|
|
|
|
|
|
| |
Apart from updating CustomCommandLineTest this CL is entirely automated.
We also sneak in a rename of addFormat -> addFormatted.
RELNOTES: None
PiperOrigin-RevId: 164870140
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Instead of having custom ArgvFragments for every combination of desired things, we make a combined "interpreter" of argvs. This saves memory and simplifies things as we do not have to allocate a strategy instance per call to args (instead pushing a single shared instance, followed by the args).
The generic interpreter does have a lot of branching compared to the bespoke implementations, but because the branch is always the same for long stretches the branch predictor should easily be able to handle it with minimal overhead (~1 cycle per branch IIRC).
This CL also elevates that we either want a NestedSet or an ImmutableCollection to the surface of the API, so consumers understand the cost when they call it with a non-immutable collection. Most of the changes in clients is due to this.
To cut down on CL churn, @Deprecated forwarding methods are added to CustomCommandLine. These will be removed in a separate CL using IDE inlining.
RELNOTES: None
PiperOrigin-RevId: 164725370
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
*** Reason for rollback ***
Updated to avoid https://github.com/bazelbuild/bazel/issues/3501
This is a rollback of a rollback, with additional
modifications to BazelConfiguration.java to fix
https://github.com/bazelbuild/bazel/issues/3501,
the issue that was the reason we rolled back the
original change.
The additional updates serve to propagate the
client's TMP and TEMP envvars to the action, which
is a short-term solution to allow actions have a
TMP/TEMP envvar on Windows. They need at least one
of those to create temp directories.
The long-term solution is to set a value for TMP
or TEMP in the executor just before executing the
actions, so the TMP/TEMP would not be part of the
action key.
All of this only affects Bazel on Windows.
*** Original change description ***
Automated rollback of commit 0abf5fa2d64c76def5a8fa0f960b73ce0566af4d.
*** Reason for rollback ***
Breaks Bazel CI (https://github.com/bazelbuild/bazel/issues/3501)
*** Original change description ***
Android BusyBox: actions use the default shell env
SpawnActions that run the Android BusyBox now use
the default shell environment.
This has the following benefits:
- Bazel propagates the PATH, TMPDIR envvars to the
action
- Bazel propagates the --action_env envvars to the
action
This allows the Bazel client to pass
--action_env=TMP or --action_env=TEMP (whichever
of the envvars is defined) to the server, so the
BusyBox actions will have TMP/TEMP set (to the
same value as the clientenv), so they can create
temp directories using
java.nio.file.Files.createTempDirectory.
This method seems to be calling the GetTempPath
WinAPI function, which needs the TMP or TEMP
envvar, otherwise it falls back to returning
c:\windows which is non-writable.
There's one drawback of using the default shell
environment, although @ulfjack is working on it:
- PATH is now also part of the action's cache key.
However in a single-machine environment (no
remote execution) and assuming PATH isn't likely
to change between builds, this probably doesn't
poision the action cache in practice.
This change is a short-term solution. Propagating
the client env's TMP/TEMP means we make that part
of the action's cache key.
The ideal long-term solution will be to not
propagate this envvar, and instead let the
execution strategy set it to some
client-env-independent value.
See https://github.com/bazelbuild/bazel/issues/3264
Change-Id: I756a4203b5d86c881bc36cc089e35cde0d419914
RELNOTES: none
PiperOrigin-RevId: 164696600
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
*** Reason for rollback ***
Breaks Bazel CI (https://github.com/bazelbuild/bazel/issues/3501)
*** Original change description ***
Android BusyBox: actions use the default shell env
SpawnActions that run the Android BusyBox now use
the default shell environment.
This has the following benefits:
- Bazel propagates the PATH, TMPDIR envvars to the
action
- Bazel propagates the --action_env envvars to the
action
This allows the Bazel client to pass
--action_env=TMP or --action_env=TEMP (whichever
of the envvars is defined) to the server, so the
BusyBox actions will have TMP/TEMP set...
***
PiperOrigin-RevId: 164126020
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
SpawnActions that run the Android BusyBox now use
the default shell environment.
This has the following benefits:
- Bazel propagates the PATH, TMPDIR envvars to the
action
- Bazel propagates the --action_env envvars to the
action
This allows the Bazel client to pass
--action_env=TMP or --action_env=TEMP (whichever
of the envvars is defined) to the server, so the
BusyBox actions will have TMP/TEMP set (to the
same value as the clientenv), so they can create
temp directories using
java.nio.file.Files.createTempDirectory.
This method seems to be calling the GetTempPath
WinAPI function, which needs the TMP or TEMP
envvar, otherwise it falls back to returning
c:\windows which is non-writable.
There's one drawback of using the default shell
environment, although @ulfjack is working on it:
- PATH is now also part of the action's cache key.
However in a single-machine environment (no
remote execution) and assuming PATH isn't likely
to change between builds, this probably doesn't
poision the action cache in practice.
This change is a short-term solution. Propagating
the client env's TMP/TEMP means we make that part
of the action's cache key.
The ideal long-term solution will be to not
propagate this envvar, and instead let the
execution strategy set it to some
client-env-independent value.
See https://github.com/bazelbuild/bazel/issues/3264
Change-Id: I756a4203b5d86c881bc36cc089e35cde0d419914
PiperOrigin-RevId: 164114502
|
|
|
|
|
|
|
|
|
|
| |
Consumers using spawn action builder now have access to handy overloads that behind the scene do a lazy String.format. In 95% of cases progress messages are expressible as 0, 1, or 2 argument String.formats.
This saves memory because the format string is constant and shared between all actions, and the captured subjects are usually live on the heap anyway (eg. labels).
Skylark still computes its progress messages eagerly. If we want similar savings there I'd have to follow up with a Skylark proposal.
PiperOrigin-RevId: 164068816
|
|
|
|
|
| |
RELNOTES: None
PiperOrigin-RevId: 162613289
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This is a partial rollback.
I restored the old flags (with PathListConverter),
added the new ones, added deprecation warnings and
annotations, and added code to concat the values
of new and old.
Existing released versions of Bazel should keep
working with the new code until 2018-01-31, i.e.
for the next ~6 months. After that I'll remove the
old flags and the PathListConverter.
*** Reason for rollback ***
Fixed broken integration with release Blaze.
*** Original change description ***
Automated rollback of commit 5752463ece84ebb4fb074888cba57412ab8d86b3.
*** Reason for rollback ***
Broke too many targets.
***
RELNOTES: none
PiperOrigin-RevId: 162587548
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
*** Reason for rollback ***
Caused memory regression.
*** Original change description ***
CustomCommandLine.Builder: clean up its interface
In this commit:
- remove unused methods and classes
- turn CustomCommandLine.ArgvFragment into an
interface
- remove the
CustomCommandLine.TreeFileArtifactArgvFragment
abstract class; it only had one remaining
subclass
- add @Nullable annotations where nulls are fine
- add Precondition checks for non-nullable args
- simplify the interface by removing add* methods
that can be composed of other add* methods; this
makes it easier to see...
***
RELNOTES: none
PiperOrigin-RevId: 162320031
|
|
|
|
| |
PiperOrigin-RevId: 162194755
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This commit:
- deprecates PathListConverter
- removes ExistingPathListConverter because it was
not used in production, only tests
- deprecated List<Path> type flags that use
PathListConverter
- introduces new List<Path> type flags next to the
deprecated ones that use @Options.allowMultiple
and convert with PathConverter; the new and old
lists are concatenated, yielding the flag value
PathListConverter and all of its occurrences
should be removed after 2018-01-31 (about 6 months
from now, which is a safe enough timeframe for
everyone to upgrade Bazel so it uses the new-style
flags).
Reason for deprecation is that colon-separated
path lists don't work on Windows because paths
have colons in them.
Since the Android BusyBox is not intended to be
executed by users but by Bazel only, there's no
release notes necessary.
See https://github.com/bazelbuild/bazel/issues/3264
RELNOTES: none
PiperOrigin-RevId: 162193998
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
In this commit:
- remove unused methods and classes
- turn CustomCommandLine.ArgvFragment into an
interface
- remove the
CustomCommandLine.TreeFileArtifactArgvFragment
abstract class; it only had one remaining
subclass
- add @Nullable annotations where nulls are fine
- add Precondition checks for non-nullable args
- simplify the interface by removing add* methods
that can be composed of other add* methods; this
makes it easier to see what the callers do with
the Builder
- remove add* methods that add a single argument
followed by a list of other elements (or a
joined string of them); these had a bug in that
they didn't check if the collection was empty
(only that it was not null), and if it was empty
then the single argument was still added though
it was not followed by any value
- fix call sites of add* methods where we
previously could have added a flag with an empty
collection
- audit every affected call site
RELNOTES: none
PiperOrigin-RevId: 161957521
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Rollback of commit df366408188f0451bae9b2ed079c795a4beb2e2b.
In addition to undoing the rollback of my previous change, fix the bugs it
introduced and add tests for those bugs.
Always ignore empty filters. Empty filters are always useless or
counterproductive. Before the original change, empty filters as a single item
within the list of filters (e.g., ["en", ""]) were ignored, but empty filters
as a portion of a string in the list (e.g., ["en,"]) were not. I can't imagine
any reason people would actually want the empty filter (if it were handled
correctly, it would effectively tell Bazel to just ignore every other filter
the user passed in).
Since it makes more sense with the new code, represent the stringified filters
as a list of strings, rather than a single string of comma-seperated values.
Manually trim whitespace from each token. Before the original change, the code
trimmed whitespace following commas (e.g., ["en, es"] -> ["en,es"]) but not
otherwise. If we're allowing whitespace in filter strings anyway, there doesn't
seem to be any reason to allow it in some places but not others.
--
PiperOrigin-RevId: 151128685
MOS_MIGRATED_REVID=151128685
|
|
|
|
|
|
| |
--
PiperOrigin-RevId: 151000602
MOS_MIGRATED_REVID=151000602
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
In addition to filtering android_binary resources by
resource_configuration_filters, we also filter by densities. Doing this in
analysis rather than execution should also result in a speed-up as there's no
need to copy files unwanted for actions to use.
This behavior is controlled by the same object and flags that already
control resource configuration filtering, simplifying the code.
--
PiperOrigin-RevId: 150871620
MOS_MIGRATED_REVID=150871620
|
|
|
|
|
|
|
|
|
| |
This is read by PathListConverter which uses java.io.File#pathSeparator which
is already platform-specific.
--
PiperOrigin-RevId: 150008708
MOS_MIGRATED_REVID=150008708
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Before this change, resource_configuration_filters were only used to filter
resources out of android_binary targets during resource processing. Instead, we
can filter these unwanted resources out during analysis. This means we won't
have to copy these resources for use in execution, and that execution itself
will have a smaller set of inputs. This should produce a speed-up in
android_binary targets with resource_configuration_filters.
Also add a flag that controls this behavior. Currently, the flag defaults to
not doing this prefiltering. We'll change that after further testing.
Also, do some related cleanup:
- Replace repeated used of the constant "resource_configuration_filters" with
references to the new class
- Pass around instances of the new class rather that List<String> objects meant
to represent ResourceConfigurationFilters
--
PiperOrigin-RevId: 148353501
MOS_MIGRATED_REVID=148353501
|
|
|
|
|
|
|
|
|
|
| |
targets when both resource shrinking and Proguard are enabled, as it is required by resource shrinking to work properly.
RELNOTES: Automatically generate Proguard mapping when resource shrinking and Proguard are enabled.
--
PiperOrigin-RevId: 148349205
MOS_MIGRATED_REVID=148349205
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
*** Reason for rollback ***
Rolling forward with fixes for the incremental tool.
*** Original change description ***
Automated [] rollback of commit d11d510c571b10787856395709f9ad945ca70bb2.
*** Reason for rollback ***
--
PiperOrigin-RevId: 146940409
MOS_MIGRATED_REVID=146940409
|
|
|
|
|
|
|
|
| |
*** Reason for rollback ***
--
PiperOrigin-RevId: 146820790
MOS_MIGRATED_REVID=146820790
|
|
|
|
|
|
|
|
|
| |
This makes the code simpler as well as reducing the number of targets to build.
It also makes testing and profiling different action strategies vastly easier.
--
PiperOrigin-RevId: 146812659
MOS_MIGRATED_REVID=146812659
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This merges the AndroidResourceContainerBuilder (which it's not even clear is related to
the nested ResourceContainer!) into the newly generated ResourceContainer.Builder.
It also seemed ridiculous for ResourceContainer to get so large and still be subordinate
to AndroidResourcesProvider, especially when it's getting passed around in a lot of other
places (look how many imports needed fixing!). This CL makes it its own top level class.
This allows for easy modification of an existing instance: call toBuilder on it, set
the properties you want set, and then call build.
--
PiperOrigin-RevId: 143574468
MOS_MIGRATED_REVID=143574468
|
|
|
|
|
| |
--
MOS_MIGRATED_REVID=134118816
|
|
|
|
|
|
|
|
| |
* Re-add proguard mapping file handling to ResourceShrinker to handle the fully Proguarded code, potentially including an obfuscation step.
* Update android_binary to provide the resource shrinker with the standard Proguard jar and obfuscation mapping instead of a custom built shrunk jar.
--
MOS_MIGRATED_REVID=128476602
|
|
|
|
|
|
|
| |
by the Gradle resource shrinker in resources.txt.
--
MOS_MIGRATED_REVID=126420534
|
|
|
|
|
|
|
| |
packages from manifests. This prevents resources from being removed when they have been generated into a package other than the one specified in their target's manifest.
--
MOS_MIGRATED_REVID=119294591
|
|
and specify --experimental_android_resource_shrinking on the command line.
RELNOTES: Specifying --experimental_android_resource_shrinking on the command line will enable a resource shrinking pass on android_binary targets that already use Proguard.
--
MOS_MIGRATED_REVID=116572863
|