search

Quuxplusone / LLVMBugzillaTest

0 stars 0 forks source link

polly optimizer causes opus tests in ffmpeg to fail #28567

Open Quuxplusone opened 7 years ago

Quuxplusone commented 7 years ago
Bugzilla Link PR28568
Status NEW
Importance P normal
Reported by Jeremy Huddleston Sequoia (jeremyhu@apple.com)
Reported on 2016-07-15 01:31:09 -0700
Last modified on 2018-10-09 01:05:38 -0700
Version unspecified
Hardware PC All
CC johannes@jdoerfert.de, llvm-bugs@lists.llvm.org, tobias@grosser.es, zamazan4ik@tut.by
Fixed by commit(s)
Attachments
Blocks
Blocked by
See also 
Check out http://fate.ffmpeg.org/?query=os:darwin//&sort=cc

Recent llvm trunk is failing a few tests when ffmpeg is compiled with:
-O3 -polly -polly-vectorizer=stripmine
or
-O3 -polly

Disabling the optimizer causes the tests to pass:
-O3 -polly -polly-optimizer=none

The failing tests are:

opus-testvector01
opus-testvector05
opus-testvector06
opus-testvector07
opus-testvector08
opus-testvector09
opus-testvector10
opus-testvector11
opus-tron.6ch.tinypkts
Quuxplusone commented 7 years ago 
Hi Jeremy,

thank you for reporting this bug. It seems we currently still have some
correctness issues either because of or exposed by invariant load hoisting
code. I temporarily disabled it in https://llvm.org/svn/llvm-
project/polly/trunk@278681, which -- as I just verified -- indeed causes the
latest version of polly to compile FFMPEG without introducing correctness
issues. The miscompiles are unfortunately still visible with -polly-invariant-
load-hoisting. Before re-enabling this option, we need to make sure that this
bug report gets resolved. I leave this bug open to make sure we don't forget.

For anyone interested in working on this bug, http://llvm.org/PR28985 is a
similar issue which might possibly be related and which has already been
reduced. So starting off from this issue is probably a good idea.

Best,
Tobias
Quuxplusone commented 5 years ago

What is the current status of the bug?