The following code is miscompiled at -O3 with trunk and 4.5.0: ! From forall_12.f90 ! Fails with loop reversal at -O3 ! character(len=1) :: b(4) = (/"1","2","3","4"/), c(4) c = b i = 1 ! This statement must be here for the abort below b(1:3)(i:i) = b(2:4)(i:i) b = c b(4:2:-1)(i:i) = b(3:1:-1)(i:i) ! This fails. If the condition is printed, the result is F F F F if (any (b .ne. (/"1","1","2","3"/))) i = 2 print *, b print *, b .ne. (/"1","1","2","3"/) if (i == 2) call abort end [macbook] f90/bug% gfc -O3 forall_12_db.f90 [macbook] f90/bug% a.out 1123 F F F F Abort Abort is not called if the test 'if (any ...' is replaced with if (any (b .ne. (/"1","1","2","4"/))) i = 2 As spotted by Tobias Burnus, the problem comes from ... static character(kind=1)[1:1] * A.5[4] = {&"1"[1]{lb: 1 sz: 1}, &"1"[1]{lb: 1 sz: 1}, &"2"[1]{lb: 1 sz: 1}, &"3"[1]{lb: 1 sz: 1}}; ... <bb 5>: D.1645_85 = A.5[3]; D.1646_86 = (*D.1645_85)[1]{lb: 1 sz: 1}; if (D.1646_86 != D.1596_154) goto <bb 7>; else goto <bb 6>; <bb 6>: ...
Fails also at -O2 -funroll-loops. I can't see anything wrong with the tree-level code.
It is caused by revision 152236: http://gcc.gnu.org/ml/gcc-cvs/2009-09/msg00987.html
This goes wrong with -O2 -funroll-loops already during tree opts. In vrp2 we have: b[2][1]{lb: 1 sz: 1} = D.1599_148; b[3][1]{lb: 1 sz: 1} = D.1599_154; D.1635_152 = 2; D.1636_151 = &b[2][1]{lb: 1 sz: 1}; D.1637_146 = 3; D.1638_145 = &b[3][1]{lb: 1 sz: 1}; __builtin_memcpy (D.1638_145, D.1636_151, 1); ... D.1647_84 = b[3][1]{lb: 1 sz: 1}; D.1648_85 = A.5[3]; D.1649_86 = (*D.1648_85)[1]{lb: 1 sz: 1}; if (D.1647_84 != D.1649_86) but in dom2: b[2][1]{lb: 1 sz: 1} = D.1599_148; b[3][1]{lb: 1 sz: 1} = D.1599_154; D.1635_152 = 2; D.1636_151 = &b[2][1]{lb: 1 sz: 1}; D.1637_146 = 3; D.1638_145 = &b[3][1]{lb: 1 sz: 1}; D.1680_170 = b[2][1]{lb: 1 sz: 1}; b[3][1]{lb: 1 sz: 1} = D.1680_170; ... D.1647_84 = D.1599_154; D.1648_85 = A.5[3]; D.1649_86 = (*D.1648_85)[1]{lb: 1 sz: 1}; if (D.1649_86 != D.1599_154) b[3] at that point isn't D.1599_154 though, but D.1599_148 (== D.1680_170).
Indeed. Mine.
Subject: Bug 44592 Author: matz Date: Mon Jun 28 15:14:31 2010 New Revision: 161496 URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=161496 Log: PR middle-end/44592 * gimple-fold.c (gimplify_and_update_call_from_tree): Maintain proper VDEF chain for intermediate stores in the sequence. testsuite/ PR middle-end/44592 * gfortran.dg/pr44592.f90: New test. Added: trunk/gcc/testsuite/gfortran.dg/pr44592.f90 Modified: trunk/gcc/ChangeLog trunk/gcc/gimple-fold.c trunk/gcc/testsuite/ChangeLog
Fixed for 4.6, waiting a bit for 4.5.
> Fixed for 4.6, waiting a bit for 4.5. Revision 161496 caused pr44699.
> > Fixed for 4.6, waiting a bit for 4.5. > > Revision 161496 caused pr44699. I think the PRINTs are of no use in the testsuite for gfortran.dg/pr44592.f90. I'ld suggest to apply the following patch: --- ../_clean/gcc/testsuite/gfortran.dg/pr44592.f90 2010-06-28 17:51:41.000000000 +0200 +++ gcc/testsuite/gfortran.dg/pr44592.f90 2010-07-01 14:12:38.000000000 +0200 @@ -13,8 +13,5 @@ b(4:2:-1)(i:i) = b(3:1:-1)(i:i) ! This fails. If the condition is printed, the result is F F F F - if (any (b .ne. (/"1","1","2","3"/))) i = 2 - print *, b - print *, b .ne. (/"1","1","2","3"/) - if (i == 2) call abort + if (any (b .ne. (/"1","1","2","3"/))) call abort end
The important question is if the testcase still ICEs with the fix reverted when you do that change.
> The important question is if the testcase still ICEs with the fix reverted when > you do that change. The test aborts with revisions before 161496.
GCC 4.5.1 is being released, adjusting target milestone.
What are the plans with is pr? Will it be back-ported to 4.5?
On Sun, 10 Oct 2010, dominiq at lps dot ens.fr wrote: > http://gcc.gnu.org/bugzilla/show_bug.cgi?id=44592 > > --- Comment #12 from Dominique d'Humieres <dominiq at lps dot ens.fr> 2010-10-10 19:36:15 UTC --- > What are the plans with is pr? Will it be back-ported to 4.5? There were various fallouts of the fix, including one fixed only last week. But yes, a backport might eventually happen
I will look at backporting the 4.6 variant.
GCC 4.5.2 is being released, adjusting target milestone.
Author: rguenth Date: Mon Jan 17 11:31:10 2011 New Revision: 168894 URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=168894 Log: 2011-01-17 Richard Guenther <rguenther@suse.de> Backport from mainline PR tree-optimization/47286 * tree-ssa-structalias.c (new_var_info): Register variables are global. * gcc.dg/tree-ssa/pr47286.c: New testcase. PR tree-optimization/44592 * tree-ssa-ccp.c (gimplify_and_update_call_from_tree): Copy from trunk. * gfortran.dg/pr44592.f90: New testcase. Added: branches/gcc-4_5-branch/gcc/testsuite/gcc.dg/tree-ssa/pr47286.c branches/gcc-4_5-branch/gcc/testsuite/gfortran.dg/pr44592.f90 Modified: branches/gcc-4_5-branch/gcc/ChangeLog branches/gcc-4_5-branch/gcc/testsuite/ChangeLog branches/gcc-4_5-branch/gcc/tree-ssa-ccp.c branches/gcc-4_5-branch/gcc/tree-ssa-structalias.c
Fixed.