Issue metadata
Sign in to add a comment
|
1.3% regression in sizes/libcronet.so on android_cronet_arm64_builder at 416818:416819 |
||||||||||||||||||
Issue descriptionPerformance dashboard identified a 1.3% regression in sizes/libcronet.so on android_cronet_arm64_builder at revision range 416818:416819. Graph: https://chromeperf.appspot.com/report?masters=ChromiumAndroid&bots=android_cronet_arm64_builder&tests=sizes%2Flibcronet.so&checked=libcronet.so%2Clibcronet.so_ref%2Cref&rev=416819 This is due to the NDK roll.
,
Sep 12 2016
I've been looking at this some this morning. Many of the object files got significantly larger when compiled with the version of gcc in NDK r12b (a 2015 4.9.x build) compared to the same files compiled with the version of gcc in NDK r10e (a 2014 4.9.x build). No compile flags changed in the switch. I'm going to spend some time looking at the gcc release notes since 4.9 when I get a chance.
,
Sep 12 2016
of note, though: gcc is deprecated in favor of clang in the NDK as of r11: https://developer.android.com/ndk/downloads/revision_history.html I tried building w/ clang as well, but neither the version in the NDK nor the version in //third_party/llvm-build/Release+Asserts/ work, even w/ a few gn changes. We could look at this further, though.
,
Sep 12 2016
We have an android clang bot, so the version in third_party/llvm-build/Release+Asserts definitely should work (it has bot coverage on some archs).
,
Sep 12 2016
hmm, maybe I'm not configuring it correctly then.
,
Sep 12 2016
Just `is_clang = true` in your args.gn should be enough. Here's one bot I know of (https://build.chromium.org/p/chromium.fyi/builders/ClangToTAndroidASan -- currently red 'cause that one uses trunk clang and trunk clang is apparently broken, but we have some bots on the internal waterfall using pinned clang too I think.)
,
Sep 12 2016
On arm, that's true. On arm64, it isn't: - boringssl's assembly complains. Fixable by changing https://codesearch.chromium.org/chromium/src/third_party/boringssl/BUILD.gn?rcl=0&l=73 s.t. we use -fno-integrated-as on arm64 - base/strings/string_util.h complains about macro expansion in the snprintf declaration in //base/strings/string_util.h. Still looking for a fix.
,
Sep 12 2016
arm64 + clang is bug 539781
,
Sep 12 2016
Ah. Starred.
,
Sep 12 2016
(See also thread "[chromium-dev] Chromium browser Clang build - any progress ?" from a while ago. As far as I understand, this used to roughly work at some point earlier this year -- but we don't have bots for that.)
,
Nov 18 2016
,
May 6 2017
This regression has been included in a stable release and that stable channel is now deprecated. I'm closing this so that we won't have unresponsive performance regressions in the future. |
|||||||||||||||||||
►
Sign in to add a comment |
|||||||||||||||||||
Comment 1 by jbudorick@chromium.org
, Sep 7 2016