diff options
author | Nicolas Geoffray <ngeoffray@google.com> | 2020-09-16 09:46:58 +0100 |
---|---|---|
committer | Nicolas Geoffray <ngeoffray@google.com> | 2020-10-07 08:32:52 +0000 |
commit | 8d34a182fea1b24f7b8361b55e930cb953cf3fb2 (patch) | |
tree | 4f5ed9d9ac417dfd69fd18f64412b2272c448e05 /compiler/optimizing/code_generator.cc | |
parent | 8ecbc4e844fc3b73e6a5c5151eda914d53297180 (diff) |
Change interface conflict stub to take the interface method.
To avoid doing dex cache lookup, pass the interface method instead. This
costs a few hundred KBs on speed compiled APKs (< 0.5% code size), but
improves performance when hitting a conflict (as seen on dogfood data).
For nterp, we currently pass the conflict method instead of the
interface method. We need to handle default methods before optimizing
it.
This removes our last use of dex cache in compiled code. A follow-up CL
will remove the NeedsDexCacheOfDeclaringClass from HInvokeInterface.
Test: test.py
Change-Id: I3cdd4543ad7d904b3e81950af46a48a48af6991a
Diffstat (limited to 'compiler/optimizing/code_generator.cc')
-rw-r--r-- | compiler/optimizing/code_generator.cc | 3 |
1 files changed, 1 insertions, 2 deletions
diff --git a/compiler/optimizing/code_generator.cc b/compiler/optimizing/code_generator.cc index 9ff1f73de4..c2ae0e0632 100644 --- a/compiler/optimizing/code_generator.cc +++ b/compiler/optimizing/code_generator.cc @@ -923,8 +923,7 @@ uint32_t CodeGenerator::GetBootImageOffset(HLoadString* load_string) NO_THREAD_S return GetBootImageOffsetImpl(string.Ptr(), ImageHeader::kSectionObjects); } -uint32_t CodeGenerator::GetBootImageOffset(HInvokeStaticOrDirect* invoke) { - DCHECK_EQ(invoke->GetMethodLoadKind(), MethodLoadKind::kBootImageRelRo); +uint32_t CodeGenerator::GetBootImageOffset(HInvoke* invoke) { ArtMethod* method = invoke->GetResolvedMethod(); DCHECK(method != nullptr); return GetBootImageOffsetImpl(method, ImageHeader::kSectionArtMethods); |