New issue
Advanced search Search tips

Issue 648300 link

Starred by 1 user

Issue metadata

Status: Fixed
Owner:
Closed: Sep 2016
Cc:
EstimatedDays: ----
NextAction: ----
OS: Linux
Pri: 3
Type: Bug

Blocking:
issue 645953



Sign in to add a comment

Several clang tests fail on linux clang upload bot stage 1

Project Member Reported by thakis@chromium.org, Sep 19 2016

Issue description

Several clang tests fail on linux clang upload bot stage 1:

********************
Testing: 0 .. 10.. 20.. 30.. 40.. 50.. 60.. 70.. 80.. 90.. 
Testing Time: 41.11s
********************
Failing Tests (3):
    LLVM :: Transforms/LoopVectorize/X86/gather_scatter.ll
    LLVM :: Transforms/LoopVectorize/X86/reg-usage.ll
    LLVM :: Transforms/LoopVectorize/X86/scatter_crash.ll


I can repro by running  tools/clang/scripts/package.py locally (at clang-281889-1), but I can't repro in my normal llvm build dir at that same revision. Maybe this only fails when using third_party/llvm-build-tools/gcc482precise/bin/gcc --version as host compiler (that's 4.8.2; my system gcc is 4.8.4)?


See linux_upload_clang on ps1 of https://codereview.chromium.org/2349183003
 

Comment 1 by h...@chromium.org, Sep 19 2016

Pretty random guess, but r281851 has both "LoopVectorize" and "GEP" in it..

Comment 2 by thakis@chromium.org, Sep 19 2016

I tried at 281850-1 locally and those still fail

Comment 3 by thakis@chromium.org, Sep 19 2016

Current regression range 280900:280900. afk for a bit, will bisect more once back.

Comment 4 by thakis@chromium.org, Sep 19 2016

err  280900:280980

Comment 5 by thakis@chromium.org, Sep 19 2016

280979 is where this starts happening

Comment 7 by thakis@chromium.org, Sep 19 2016

Filed https://llvm.org/bugs/show_bug.cgi?id=30450 -- not sure what to do about this though.

Comment 8 by h...@chromium.org, Sep 19 2016

If "NumOperands not initialized" as the assert says, maybe valgrind can give a clue?
Project Member

Comment 9 by bugdroid1@chromium.org, Sep 20 2016

The following revision refers to this bug:
  https://chromium.googlesource.com/chromium/src.git/+/6e96f3f9582481230cee3a6e8987cc877940fab6

commit 6e96f3f9582481230cee3a6e8987cc877940fab6
Author: Nico Weber <thakis@chromium.org>
Date: Tue Sep 20 00:08:30 2016

clang bots: Use gcc 4.8.5 instead of 4.8.2 as stage1 compiler.

For reasons not fully understood, some of clang's tests started failing
when building clang with gcc 4.8.2 but they pass with 4.8.4 or 4.8.5.
Since we do bootstrap builds when building clang binaries, the stage1
host compiler doesn't matter much to us, so just flip the bots to
gcc 4.8.5 and move on with life.

I built 4.8.5 (in a precise vm) with:

  contrib/download_prerequisites
  mkdir build && cd build
  ../configure --prefix=$HOME/gcc485precise --enable-languages=c,c++ --enable-checking=release
  time make -j10
  make install-strip
  tar czf gcc485precise.tgz gcc485precise
  cd ~
  ~/gsutil/gsutil cp -n -a public-read gcc485precise.tgz gs://chromium-browser-clang/tools/gcc485precise.tgz

i.e. I switched from `make install` to `make install-strip`, which also reduces
download size from 125MB to 30 MB.

BUG= 648300 
TBR=hans

Review URL: https://codereview.chromium.org/2354583004 .

Cr-Commit-Position: refs/heads/master@{#419607}

[modify] https://crrev.com/6e96f3f9582481230cee3a6e8987cc877940fab6/tools/clang/scripts/update.py

Owner: thakis@chromium.org
Status: Fixed (was: Untriaged)
Switching the linux bots to use gcc4.8.5 fixed things: https://build.chromium.org/p/tryserver.chromium.linux/builders/linux_upload_clang/builds/86

It's still broken for others using gcc 4.8.4 (see upstream bug), but for us this is worked around.

This means a slightly newer libstdc++ is bundled with clangs from now on, but hopefully that won't cause issues. We'll see.

Sign in to add a comment