cselim tree optimizer generates incorrect code

[Stéphane Glondu]

I think gcc (4.6.3-1ubuntu5, amd64) generates incorrect code with the attached test case.

The expected output is "m = 8241071186". However, when compiled with "-O3", it outputs "m = 1". Looking further, it seems that the following lines:

       *m = 1UL;
      ularith_div_2ul_ul_ul_r (m, 0UL, 1UL, s);

are being swapped in the assembly output.

The output is correct when compiled with "-O2" or with "-O3 -fno-tree-cselim".

On Debian (version 4.6.3-1), the output is always correct. I have reports that the output is correct with upstream gcc compiled on Ubuntu, so this behaviour looks Ubuntu-specific.

[Stéphane Glondu (glondu)]

The version in quantal (4.6.3-8ubuntu1) is not affected.

[Stéphane Glondu (glondu)]

Sorry, I meant the default gcc (which is gcc-4.7). With gcc-4.6, I observe the same behaviour.

[Stéphane Glondu (glondu)]

Same issue in oneiric (4.6.1-9ubuntu3).

[Stéphane Glondu (glondu)]

I've tried various historic versions of the gcc-4.6 package in a clean precise chroot.

The first faulty version appears to be 4.6.0-7ubuntu1. The previous version, 4.6.0-6ubuntu3, generates correct code.

Moreover, the last version (4.6.3-1ubuntu5) with all Linaro patches out generates correct code.

[Michael Hope (michaelh1)]

Thank you for the bug report. I've confirmed this with gcc-linaro-4.6+bzr106903 on x86_64:

michaelh@crucis:~/linaro/bugs$ ~/linaro/gcc/build/native-4.6/install/bin/gcc -O3 testcase3.c
michaelh@crucis:~/linaro/bugs$ ./a.out
m = 1

The fault is probably introduced by r106733 which added condtional store sinking support.

I've set it to high priority as bad code is generated. Likely trunk fixes include r113242, r111593, r110,443, and r114986 of lp:gcc-linaro/4.7.

[Matthew Gretton-Dann (matthew-gretton-dann)]

Confirmed reverting lp:gcc-linaro/4.6 r106733 fixes the issue.

[Matthew Gretton-Dann (matthew-gretton-dann)]

The revisions Michael suggested do not fix the issue - either singly or when applied on top of each other.

[Matthew Gretton-Dann (matthew-gretton-dann)]

Investigations lead me to believe this is upstream bug PR51987 (http://gcc.gnu.org/PR51987) made visible because of the conditional store sinking support being added.

I believe upstream fix svn r183524 fixes the issue. This is currently under test

[Matthias Klose (doko)]

the backport fixes the issue, no regressions running the testsuite. however the testcase fails, because 4.6 doesn't know about ia32, so s/ia32/ilp32/

[Launchpad Janitor (janitor)]

This bug was fixed in the package gcc-4.6 - 4.6.3-10ubuntu1

---------------
gcc-4.6 (4.6.3-10ubuntu1) quantal; urgency=low

  * Merge with Debian.

gcc-4.6 (4.6.3-10) unstable; urgency=low

  * Update to SVN 20120918 (r191439) from the gcc-4_6-branch.
    - Fix PR c/54552 (ice on valid), PR c/54103 (ice on valid),
      PR target/54536 (AVR), PR middle-end/54515 (ice on valid),
      PR target/45070 (ARM, wrong code), PR target/54220 (AVR),
      PR driver/54335, PR rtl-optimization/54369 (mips, wrong code),
      PR c++/54511 (ice on valid), PR fortran/54225 (ice on invalid),
      PR fortran/53306 (ice on invalid), PR fortran/54556 (wrong code),
      PR fortran/54208 (rejects valid).

  [ Nobuhiro Iwamatsu ]
  * Remove sh4-enable-ieee.diff, -mieee enabled by default. Closes: #685974.

  [ Matthias Klose ]
  * Fix PR tree-optimization/51987, backport from the trunk, Linaro only
   (Matthew Gretton-Dann). LP: #1029454.
 -- Matthias Klose <email address hidden> Tue, 18 Sep 2012 22:44:00 +0200

[Steve Langasek (vorlon)]

The Precise Pangolin has reached end of life, so this bug will not be fixed for that release