[llvm-bugs] [Bug 33921] New: ARM NEON vectorization miscompiles code in 5.0

via llvm-bugs llvm-bugs at lists.llvm.org
Tue Jul 25 04:17:55 PDT 2017


https://bugs.llvm.org/show_bug.cgi?id=33921

            Bug ID: 33921
           Summary: ARM NEON vectorization miscompiles code in 5.0
           Product: libraries
           Version: 5.0
          Hardware: PC
                OS: All
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: Backend: ARM
          Assignee: unassignedbugs at nondot.org
          Reporter: martin at martin.st
                CC: llvm-bugs at lists.llvm.org

Starting with SVN r306803, some C code that wasn't vectorized before now is
vectorized. This seems to break e.g. some VP9 inverse transforms in libavcodec,
in code like this:

https://git.libav.org/?p=libav.git;a=blob;f=libavcodec/vp9dsp.c;h=7f863943284b8970744c81b0b7d749e5c9cf8bdf;hb=2b1324bd167553f49736e4eaa94f96da9982925e#l947
with the idct8_1d and iadst8_1d functions. If this file is built with
-fno-tree-vectorize, the issue disappears.

This regressed when upgrading from SVN r305659 to the 5.0 branch point:
https://fate.libav.org/armv7-win32-clang-5.0

Bisecting this leads to SVN r306803. The same issue can also be reproduced when
targeting linux, not only windows. It should be reproducable by building libav
and running "make fate-checkasm-vp9dsp", which should pass unless this is
miscompiled.

I will dig into it further, hopefully soon, to fill in with a better
reproducable testcase of the issue.

-- 
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20170725/dfb95a11/attachment.html>


More information about the llvm-bugs mailing list