Convert all of the callers of OtherModuleProvider/OtherModuleHasProvider
to use the type-safe android.OtherModuleProvider API.
Bug: 316410648
Test: builds
Change-Id: Id77f514d68761a262d9ea830a601dbed804bbbe5
Remove the ConvertWithBp2build implementations from all the module
types, along with the related code.
Bug: 315353489
Test: m blueprint_tests
Change-Id: I212672286686a318893bc7348ddd5a5ec51e77a7
Afdo creates more variants to build, and optimizing LTO is costly to
perform. Turn off these two optimizations for eng builds for a faster
build speed.
This reduces total C/C++ invocation time for aosp_arm64-eng by 6.28%.
Test: presubmit
Bug: 307753064
Change-Id: Ibac4269c66a64e896dba2074b607d71a2da37546
This reverts commit febf8b9bdf.
We are still seeing flaky performance alerts from microbenchmarks even
with this change, and this has caused binary size / memory usage
increase. Let's add this flag back and observe whether performance
regression reports become more frequent.
Change-Id: I0fec1296b1f49b21b59d81add677f6a926ee88dc
FSAFDO (Flow Sensitive AutoFDO) is an improvement over AutoFDO. It
attempts to capture and represent the flow-sensitive profile in LLVM
optimization passes with the mechanism to expose and iteratively
reload/match the profile with better precision.
Enable this for all AFDO projects so that we can capture data from the
field, and then evaluate the performance gain.
Test: presubmit
Bug: 304282732
Change-Id: I0de6b4071e4f0a75500f3ffd5171a7be5117ec86
-fprofile-sample-accurate allows the compiler to de-optimize cold code
paths according to the profile. However we have a number of micro-
benchmarks that tests these paths and we often trigger regression
reports from those benchmarks.
Turning off the flag mitigates the issue at a cost of ~2MB increase in
system image size. But the more deterministic codegen could reduce the
differential OTA size.
Test: presubmit
Bug: 283235002
Change-Id: I1016f573f21a60223a97c64f5818f0e7e66e41e3
We used to enable unique names only when a profile is available.
But A profile generated for a binary without unique names doesn't
work well when building a binary with unique names. To avoid a
chicken-and-egg problem, this CL enables the flag when afdo=true,
whether a profile exists or not.
Bug: 241523910
Test: build
Change-Id: I74d834510d81d2db76e34d0488c74a60e1fcecd5
This enables turnning off -fprofile-sample-accurate on a per project
basis.
Bug: 274725934
Bug: 273415627
Test: presubmit
Change-Id: I15060950f7b55e51069cb00b66512c7e94981f1b
Ignore-AOSP-First: The parent CL is internal
Bug: 267229065
Test: go test
Change-Id: Ia14679285b92f3f14ff269392a61f978c71311b2
Merged-In: Ia14679285b92f3f14ff269392a61f978c71311b2
Introducing fdo_profile module type to reimplement the afdo support in cc moduels. This change allows the feature to be compatible with Bazel migration.
How it works:
PreDepsMutators:
* BeginMutator: If non-static cc modules sets afdo prop, search and add corresponding fdo_profile module as a dep with fdoProfileTag
* fdoProfileMutator:
* If in fdo_profile module, set FdoProfileProvider with full path to profile
* If in cc module, read FdoProfileProvider from dep with fdoProfileTag and set FdoProfileInfo.Path to FdoProfilePath field
PostDepsMutators:
* afdoDepsMutator: If a module has FdoProfilePath set, walk to its static deps and set itself to the deps' AfdoRdeps
* afdoMutator: If a static dep has AfdoRDeps set, create afdo variant.
Ignore-AOSP-First: Other CLs in the same topic are internal-only
Test: go test
Bug: b/267229065
Change-Id: I687d798a02d9743c92804fea36fb4ae3a7a0e5e3
Merged-In: I687d798a02d9743c92804fea36fb4ae3a7a0e5e3
AfdoAdditionalProfileDirs is unconnected to build/make/core/soong_config.mk and is currently unused. Removing AfdoAdditionalProfileDirs will cut down on complexity when we migrate the support of auto-fdo profiles to Bazel cc rules.
b/258684450 proposes an alternative to support the use cases that were meant to be supported by AfdoAdditionalProfileDirs.
Test: existing go tests
Bug: 253540178
Change-Id: I024e48598a560717a1d6c19b16bde3fe65ff1961
AutoFDO uses a function’s symbol name to look up profile data but
functions with internal linkages do not have globally unique
identifiers. This can lead to name conflicts leading to missing profiles
and performance. We are enabling compiler option
-funique-internal-linkage-names for AFDO builds that will uniquefy
internal linkage symbols making this problem go away.
This CL might regress the performance due to mismatched profiles. It
will recover once we update the profiles.
Test: presubmit
Bug: 218448389
Change-Id: I92a2b6f72b52faab79aec9148f1962111bad7232
The static() property is true for both static libraries and static
binaries. We should allow static binaries to enable AFDO.
Test: Enable AFDO for linker
Change-Id: I409fc9f764ea878f01dc42bb2e5c075219403009
A new configuration rule for sampling PGO. This differs from the
original pgo.go rule in the following ways:
* Automatic propagation to static dependencies
* Simpler configuration (just put `afdo: true` to optimsation targets)
http://go/android-afdo
Test: build
Bug: 79161490
Change-Id: Ie194824cd523bca19e10ced41d2078fc598f13b3