In the past, dexpreopt for boot jars was very inflexible, and it was
incredibly hard to make a change that is as simple as adding a jar to a
boot image. Boot image generation was handled by
"platform_bootclasspath" and "bootclasspath_fragment" separately. This
caused not only code duplication but also the inflexiblity as such a
design did not fit today's use cases, where a boot image may take jars
from multiple mainline modules and the platform, and a mainline module
can contribute to multiple boot images. The design casued a huge
maintenance burden as any change to the boot image cost multi-week
efforts.
In recent years, efforts have been made to improve this a bit by a bit.
This change is another step towards making dexpreopt reasonable.
After this change, all boot images are generated by "dex_bootjars",
which is in build/soong and is therefore available on both the full
source tree and the thin manifest (master-art). The change decouples
profile generation/extraction from boot image generation. Profiles for
mainline modules are still handled by "bootclasspath_fragment"
because they need to be packed into APEXes when building mainline
modules and extracted from APEXes whem building the system image from
prebuilt modules. Boot images are not handled by
"bootclasspath_fragment" anymore.
Bug: 290583827
Test: m (all existing tests are still passing)
Test: Manually checked that the boot images are exactly the same as
before.
Change-Id: Ib5a5f401bee334ffcab5c26618e0c8888b84575a
We want to measure how much of build time is spent on writing large
zips to disk. We can do this by doing the rest of the work but
skipping the write itself.
Some local and trivial testing shows this is probably a lot of the
cost of these rules:
$ python -m timeit -s "import os" -r 10 "os.system('soong_zip -write_if_changed -o foo.zip -D prebuilts/sdk')"
1 loop, best of 10: 8.37 sec per loop
$ python -m timeit -s "import os" -r 10 "os.system('soong_zip -o foo.zip -D prebuilts/sdk')"
1 loop, best of 10: 5.39 sec per loop
$ python -m timeit -s "import os" -r 10 "os.system('soong_zip -n -o foo.zip -D prebuilts/sdk')"
1 loop, best of 10: 3.66 sec per loop
Bug: None
Test: treehugger
Change-Id: I43acdb08d0e00efaba9e5d7add972a7ec41646d2
product SBOM of cuttlefish devices, ignore these diffs for now in the
SBOM integration test to fix the CI breakage.
Bug: 291073898
Test: build/soong/tests/sbom_test.sh
Change-Id: Idf5569f67642784ee422847301c59dddc8ff0907
* changes:
Respect package boundaries in bp2build conversion of go modules
Create a temporary denylist for go binaries used in mixed builds
Partial bp2build conversion of blueprint_go_binary
Partial bp2build conversion of bootstratp_go_package
A heuristics is implemented in the build/bazel/scripts/bp2build_progress/bp2build_progress.py script to determine the reason for why a module is not converted.
Test: b build //build/bazel/scripts/bp2build_progress:bp2build_progress
Change-Id: I9d4eaf309dbf26bbb6de18e1af0d9cdc8fe09e94
bp2build's codegen context does not implement
BazelPathConversionContext. To reuse the utility function
transformPackagePaths, update its signature
(Also make deps of go_library unique to make the conversion resilient)
Test: go test ./bp2build
Change-Id: I126b1057d2b26bc6c7d3be2780f1b62d28323cf0
This allows us to rollout building _some_ go targets using rules_go
without affecting mixed builds.
Test: Presubmit
Bug: 284483729
Change-Id: I0ccb4c9b90614369147a380f44f7ae372ef9396e
The Java compiler may generate sources when processing annotations.
These sources are currently not included in the IDE project model, which
results in unresolved symbols when using Hilt and other annotation
processors. These sources are not passed to the Java compiler for
compilation, so they cannot be handled like other Java module sources.
This change archives the generated sources and appends them to the
existing srcjar list in module_bp_java_deps.json.
Test: go test .
Test: Verified that module_bp_java_deps.json contained srcjar entry
Test: with path/to/android_common/javac/anno.srcjar.
Test: Verified that IntelliJ can resolve generated symbols after
Test: invoking aidegen.
Change-Id: I6a3011f9140fd25f785178baee0cd6074ad94600
metalava no longer prints an ASCII banner, and has removed its
--no-banner argument. Update all call sites accordingly.
Test: presubmit
Bug: 286023667
Change-Id: I61cf5f1125dfbd93496e6a59eb1ea62886b32046
This module type does not implement android.Module, and therefore we
cannot write a conventional bp2build converter for this module type.
Instead this has been special-cased inside bp2build/build_conversion.go.
There is one major deviation between Soong and Bazel's
go_binary/go_library. Soong
collects the deps in the transitve closure and puts them on compile/link
paths. Bazel OTOH, requires the direct imports to be listed in deps of the binary
explicitly (AFAIK). Since bp2build cannot determine the list of direct
imports from the list of transitive deps, put all the transitive deps in
`deps`
Test: unit tests
Test: TH
Bug: 284483729
Change-Id: I004aaf8607fef1697a0d9e7d018ad657b67778ac
This module type does not implement android.Module, and therefore we
cannot write a conventional bp2build converter for this module type.
Instead this has been special-cased inside bp2build/build_conversion.go.
Because of the above, we also do not have access to useful functions
available in the ctx object of ConvertWithBp2build. This includes
1. Finding the package (directory) of a dep. This requires getting a
handle of the underlying module from its name (string). To solve, we
do a pre-visit to collect this information. This did not increase the
wall time. On my machine, `m bp2build --skip-soong-tests` takes ~14s
before and after this CL
2. Converting srcs to labels. This requires glob and package boundary
resolution. This CL introduces a partial implementation for this
function. (glob patterns are not used in go tools)
For (1), I considered creating a `ModuleFromName` on
`blueprint.Context` instead of a pre-run, but this increased the time to ~27s.
Test: unit tests
Test: TH
Bug: 284483729
Change-Id: Ifeb029103d14947352556dba295591dd7038b090
* changes:
Change the profile path on host.
Extract duplicate code to common helper functions.
Fix dumpOatRules.
Remove Modules() from BootclasspathFragmentApexContentInfo.
Fix some tests for dexpreopt and remove unnecessary tests.
This is to decouple profile generation from image configs. A boot image
profile is either for a mainline module or for the platform, which is
orthogonal to boot images.
Bug: 290583827
Test: m nothing
Change-Id: I0918f2fa945a2af1839f1a08ecede331c64d0317
Also, fall back to using a default name for the dexpreopt directory if
we are not building for Android.
Bug: 290583827
Test: m nothing
Change-Id: I3fc6ff9142a2dcdf995796f75891b242fe2848d0
Instead of rolling assets and manifests up through each static lib,
provide them as a DepSet. This will help with the next patch, which
needs to pass all the transitive manifests and R.txt files together.
Test: app_test.go
Change-Id: Id8b3aa2bed3771e82ab6bde192c9b43baa38b54c
This method generates a build rule that runs oatdump for debugging
purposes.
- Pass "--runtime-arg -Xgc:CMC" to oatdump if UFFD GC is enabled.
- Build the phony name from the image name to avoid conflicts when using
this method for multiple boot images.
Bug: 290583827
Test: m dump-oat-boot
Change-Id: I7adc97cb85b571486b78f173fb80ff7da2c4bfe4
The Modules() method was for the ART boot image in the ART APEX. It's no
longer needed because we don't install the ART boot image in the ART
APEX anymore.
Bug: 290583827
Test: m nothing
Change-Id: I674a5fa1ed18908413a64129df5947c7d26e638e
This CL is to prepare for the changes in http://r.android.com/2652081.
See the description of that CL for the reasons.
Some tests were set up in a fragile way and were easy to be broken when
the implementation changes. This CL is a pure test change that fixes
them.
This CL also removes TestNoUpdatableJarsInBootImage. That test checks
what should go to the ART boot image and what should go to the
platform one, which does not apply today because boot images are not
associated with mainline modules anymore: a boot image may take jars
from multiple mainline modules and the platform, and a mainline module
can contribute to multiple boot images. In practice, we have ART jars in
the platform boot image, and we are going to add core-icu4j and
consrypt to the ART boot image, which is now for testing only.
Bug: 290583827
Test: m nothing
Change-Id: I22c45cbf6f853b030b68edb51197854e9c53a02e
Build C/C++ targets with ThinLTO with "--lto-O0" by default. This takes
advantage of better dead code elimination and CFG simplification due to
ThinLTO's global view of the program, but do not enable the costly LTO
optimizations. This also makes builds faster because we can avoid doing
codegen for the dead code.
Code size for panther-userdebug:
/system/bin /system/lib /system/lib64
Original 57648 86264 181700
Global ThinLTO 55764 84916 175728
-3.27% -1.56% -3.29%
Build time for aosp_arm64-userdebug:
clang time linker time
Original 56993.87s 1712.36s
Global ThinLTO 52839.18s 3114.93s
saving 4.68% build time in total.
Bug: 169004486
Test: presubmit
Change-Id: Idb0f4675ca4750b8b12b24f4679579419d3448c8
