[Bug middle-end/52835] [4.7/4.8 Regression] -O3 wrongly optimizes loop __builtin_memcpy away

["burnus at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52835

Tobias Burnus <burnus at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
           Keywords|                            |wrong-code
   Last reconfirmed|                            |2012-04-02
          Component|fortran                     |middle-end
                 CC|                            |burnus at gcc dot gnu.org
     Ever Confirmed|0                           |1
            Summary|Incorrect code generated by |[4.7/4.8 Regression] -O3
                   |gfortran 4.7.0              |wrongly optimizes loop
                   |                            |__builtin_memcpy away
   Target Milestone|---                         |4.7.1

--- Comment #1 from Tobias Burnus <burnus at gcc dot gnu.org> 2012-04-02 16:11:37 UTC ---
Confirmed - and smells rather like a middle-end issue. (Or a rather convoluted
FE declaration bug.)


Works with my 2011-12-24-r182676 build
Fails with my 2012-02-28-r184613 build
(Note: Those contained some patches, which shouldn'tâ affect the result.)

 * * *

Works with: gfortran -fdump-tree-optimized -funswitch-loops
-fpredictive-commoning -fgcse-after-reload -ftree-vectorize -finline-functions
-fipa-cp-clone -O2
(And fails with -O3 and the -fno-* version of those flags.) Thus, none of the
specific flags which get enabled with -O3 seems too blame.

 * * *

Looking at the optimized dump, the loop
      DO IS=1,MS
         ISU(IS)=0
         ZSPM(IS)=' '
         RSV(IS) =0.0
      ENDDO

gets replaced by:
  __builtin_memset (D.1976_27, 0, 12);
  __builtin_memset (D.1956_11, 0, 12);

That's almost fine, however, the
  ZSPM(is) = ' '
is lacking. With -O2 one finds in the loop:
  __builtin_memcpy (D.2078_113, &"        "[0], 8);

Note that ZSPM is an array of size 3 of character strings of length 8:
  character(len=8) :: ZSPM(MS)
where "MS" is 3.

The best (optimized) version would be something like "memset(ZSPM, " ", 3*8);".


In 4.6, one had (with -O3) and expanded loop with the following for [0], [1]
and [2]:
  D.1697_89 = &MEM[(character(kind=1)[0:D.1598][1:8] *)D.1577_23][0];
  __builtin_memcpy (D.1697_89, &"        "[0], 8);