Some debug methods may not be splitted out in harfbuzz |
||||||
Issue descriptionChrome Version: master OS: checked on Android Chrome. What steps will reproduce the problem? (1) Build Android Chrome (2) objdump the libfreetype_harfbuzz.cr.so (3) lots of blx hb_auto_trace_t<0, bool> remains. What is the expected result? hb_auto_trace_t is debug print method and it is empty on the release build. Should be removed by compiler. What happens instead? Please use labels and text to provide additional information. I happened to know hb_auto_trace_t remains on libharfbuzz_ng.so in Android. Similar things may happen in Chrome for Android. Please check this. FYI: here is a repro step on AOSP. https://docs.google.com/document/d/1_Aj29PjIxt10blwTxVryhHw8QQCAlqv3ZBtV6A_lFkU/edit#
,
Nov 3 2017
So, this sounds like a clang optimization failure bug to me. How should we proceed? Here's another suggested patch for HarfBuzz: https://github.com/behdad/harfbuzz/pull/605
,
Nov 3 2017
+pcc in case he's got an idea about whether it's worth following up with clang.
,
Nov 3 2017
I'm fixing this upstream in HarfBuzz, but think it's worth following up with clang as well.
,
Nov 5 2017
Fixed: https://github.com/behdad/harfbuzz/commit/5aad81943329df199501e9473e2cc39f9d4421a1 I'll make a release tomorrow (Monday).
,
Nov 27 2017
Behdad - Any chance you could roll harfbuzz so that we could realize these savings in Chrome? I'd be happy to roll it as well, but have no idea how to test that it works (is compiling and rendering a web page enough?)
,
Nov 28 2017
Dominik does that. Note that we will eventually get there. It might be blocked by some other issues.
,
Sep 18
This has been for a while. |
||||||
►
Sign in to add a comment |
||||||
Comment 1 by behdad@chromium.org
, Nov 3 2017