[Bug tree-optimization/54634] New: [4.8 Regression] miscompilation with -O3 -ftree-loop-distribution

[Bug tree-optimization/54634] New: [4.8 Regression] miscompilation with -O3 -ftree-loop-distribution

[Joost.VandeVondele at mat dot ethz.ch]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54634

             Bug #: 54634
           Summary: [4.8 Regression] miscompilation with -O3
                    -ftree-loop-distribution
    Classification: Unclassified
           Product: gcc
           Version: 4.8.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: Joost.VandeVondele@mat.ethz.ch


Created attachment 28227
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=28227
testcase sources

The attached sources are miscompiled with current trunk ([trunk revision
191430]) at -O3  -ftree-loop-distribution. To reproduce 

gfortran -O3  -ftree-loop-distribution  -ffree-form other.F mathconstants.F
orbital_pointers.F orbital_symbols.F orbital_transformation_matrices.F main.F ;
./a.out

which outputs wrong values (as compared to -O0) and shows a valgrind warning
(not present at -O0).

The miscompiled file is orbital_transformation_matrices.F, most likely the
routine create_spherical_harmonics (which seems inlined). If I cat at files in
a single .F file, the error also disappears, which might hint at some ipa thing
?

4.7 branch ([gcc-4_7-branch revision 190437]) is doing fine.

[Bug tree-optimization/54634] [4.8 Regression] miscompilation with -O3 -ftree-loop-distribution

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54634

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jakub at gcc dot gnu.org

--- Comment #1 from Jakub Jelinek <jakub at gcc dot gnu.org> 2012-09-20 08:58:02 UTC ---
Retry with PR54629 fix?

[Bug tree-optimization/54634] [4.8 Regression] miscompilation with -O3 -ftree-loop-distribution

[Joost.VandeVondele at mat dot ethz.ch]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54634

--- Comment #2 from Joost VandeVondele <Joost.VandeVondele at mat dot ethz.ch> 2012-09-20 10:15:57 UTC ---
(In reply to comment #1)
> Retry with PR54629 fix?

after applying the patch mentioned above, the testcase still fails. The failure
is also older than the commit mentioned in PR54629

[Bug tree-optimization/54634] [4.8 Regression] miscompilation with -O3 -ftree-loop-distribution

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54634

Richard Guenther <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |wrong-code
             Status|UNCONFIRMED                 |ASSIGNED
   Last reconfirmed|                            |2012-09-20
         AssignedTo|unassigned at gcc dot       |rguenth at gcc dot gnu.org
                   |gnu.org                     |
   Target Milestone|---                         |4.8.0
     Ever Confirmed|0                           |1

--- Comment #3 from Richard Guenther <rguenth at gcc dot gnu.org> 2012-09-20 11:29:51 UTC ---
Mine.

[Bug tree-optimization/54634] [4.8 Regression] miscompilation with -O3 -ftree-loop-distribution

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54634

--- Comment #4 from Richard Guenther <rguenth at gcc dot gnu.org> 2012-09-20 12:26:53 UTC ---
We split

  <bb 89>:
  # s2_207 = PHI <0.0(88), s2_524(91)>
  # prephitmp_1456 = PHI <0(88), k.76_530(91)>
  _496 = prephitmp_1456 * -2;
  _1535 = lx_355 + _496;
  D.3330 = _1535;
  _1533 = binomial (&j, &k);
  _517 = binomial (&ma, &D.3330);
  _518 = _1533 * _517;
  _520 = _518 * 0.0;
  s2_524 = _520 + s2_207;
  D.3330 ={v} {CLOBBER};
  k.76_530 = prephitmp_1456 + 1;
  k = k.76_530;
  if (j.72_383 == prephitmp_1456)
    goto <bb 90>;
  else
    goto <bb 91>;

into bogus pieces:

  <bb 204>:
  # prephitmp_1664 = PHI <k.76_1674(205), 0(88)>
  _1665 = prephitmp_1664 * -2;
  _1666 = lx_355 + _1665;
  D.3330 = _1666;
  D.3330 ={v} {CLOBBER};
  k.76_1674 = prephitmp_1664 + 1;
  if (j.72_383 == prephitmp_1664)
    goto <bb 206>;
  else
    goto <bb 205>;

split for the store D.3330 (huh, well ...), but somehow lost the
binomial call here.

  <bb 207>:
  # prephitmp_1678 = PHI <k.76_1688(208), 0(206)>
  k.76_1688 = prephitmp_1678 + 1;
  k = k.76_1688;
  if (j.72_383 == prephitmp_1678)
    goto <bb 209>;
  else
    goto <bb 208>;

well ... likewise.

  <bb 89>:
  # s2_207 = PHI <0.0(209), s2_524(91)>
  # prephitmp_1456 = PHI <0(209), k.76_530(91)>
  _1533 = binomial (&j, &k);
  _517 = binomial (&ma, &D.3330);
  _518 = _1533 * _517;
  _520 = _518 * 0.0;
  s2_524 = _520 + s2_207;
  k.76_530 = prephitmp_1456 + 1;
  if (j.72_383 == prephitmp_1456)
    goto <bb 90>;
  else
    goto <bb 91>;

finally all the calls.  Investigating some more.

Ah, binomial () is pure.  The RDG is bogus, it does not consider
calls that merely read unknown memory.  That is because
get_references_in_stmt handles calls by returning true/false only ...
in this case it doesn't add any data references.

So it's a very old problem that bites us now.

*sigh*

[Bug tree-optimization/54634] [4.8 Regression] miscompilation with -O3 -ftree-loop-distribution

[Joost.VandeVondele at mat dot ethz.ch]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54634

--- Comment #5 from Joost VandeVondele <Joost.VandeVondele at mat dot ethz.ch> 2012-09-20 13:06:50 UTC ---
(In reply to comment #4)
> Ah, binomial () is pure.

In this case, it was presumably triggered by Tobias' changes for PR54389.
binomial() has not been declared pure in the source, but most likely correctly
declared 'implicitly pure' but the Fortran frontend.

[Bug tree-optimization/54634] [4.8 Regression] miscompilation with -O3 -ftree-loop-distribution

[rguenther at suse dot de]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54634

--- Comment #6 from rguenther at suse dot de <rguenther at suse dot de> 2012-09-20 13:43:33 UTC ---
On Thu, 20 Sep 2012, Joost.VandeVondele at mat dot ethz.ch wrote:

> 
> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54634
> 
> --- Comment #5 from Joost VandeVondele <Joost.VandeVondele at mat dot ethz.ch> 2012-09-20 13:06:50 UTC ---
> (In reply to comment #4)
> > Ah, binomial () is pure.
> 
> In this case, it was presumably triggered by Tobias' changes for PR54389.
> binomial() has not been declared pure in the source, but most likely correctly
> declared 'implicitly pure' but the Fortran frontend.

Btw, it's just what triggers the latent bug in data dependence analysis.
I am testing

Index: gcc/tree-data-ref.c
===================================================================
--- gcc/tree-data-ref.c (revision 191561)
+++ gcc/tree-data-ref.c (working copy)
@@ -4307,10 +4307,10 @@ get_references_in_stmt (gimple stmt, VEC
   *references = NULL;

   /* ASM_EXPR and CALL_EXPR may embed arbitrary side effects.
-     Calls have side-effects, except those to const or pure
-     functions.  */
+     As we cannot model data-references to not spelled out
+     accesses give up if they may occur.  */
   if ((stmt_code == GIMPLE_CALL
-       && !(gimple_call_flags (stmt) & (ECF_CONST | ECF_PURE)))
+       && !(gimple_call_flags (stmt) & ECF_CONST))
       || (stmt_code == GIMPLE_ASM
          && (gimple_asm_volatile_p (stmt) || gimple_vuse (stmt))))
     clobbers_memory = true;

currently.

[Bug tree-optimization/54634] [4.8 Regression] miscompilation with -O3 -ftree-loop-distribution

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54634

--- Comment #7 from Richard Guenther <rguenth at gcc dot gnu.org> 2012-09-20 14:46:41 UTC ---
Author: rguenth
Date: Thu Sep 20 14:46:32 2012
New Revision: 191567

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=191567
Log:
2012-09-20  Richard Guenther  <rguenther@suse.de>

    PR tree-optimization/54634
    * tree-data-ref.c (get_references_in_stmt): For now give
    up for pure functions.

Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/tree-data-ref.c

[Bug tree-optimization/54634] [4.8 Regression] miscompilation with -O3 -ftree-loop-distribution

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54634

Richard Guenther <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |RESOLVED
         Resolution|                            |FIXED

--- Comment #8 from Richard Guenther <rguenth at gcc dot gnu.org> 2012-09-20 14:48:57 UTC ---
Fixed.