[Bug fortran/69368] [6 Regression] spec2006 test case 416.gamess fails with the g++ 6.0 compiler starting with r232508

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69368

--- Comment #64 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Dominique d'Humieres from comment #61)
> Another oddity of the "optimization" introduced by r232508 is the following
> test (borrowed from https://gcc.gnu.org/ml/gcc-patches/2016-02/msg01356.html)
> 
> [Book15] f90/bug% cat pr69368_1_a.f90
>       SUBROUTINE FOO
>       IMPLICIT DOUBLE PRECISION (X)
>       INTEGER J
> !      COMMON /MYCOMMON / X(1031)
>       COMMON /MYCOMMON / X(1)
>       DO 10 J=1,1024
>          X(J+1)=X(J+7)
>   10  CONTINUE
>       RETURN
>       END
> [Book15] f90/bug% cat pr69368_1_b.f90
>       IMPLICIT DOUBLE PRECISION (X)
>       COMMON /MYCOMMON/ X(1031)
>       DO I=1,1031
>       X(I)=I
>       END DO
>       call FOO()
>       print *, X(1025)
>       IF (X(1025).NE.1031) CALL ABORT
>       END
> 
> The test fails if pr69368_1_a.f90 is compiled with -O2 (FOO is compiled as a
> simple RETURN). Now if I replace
> 
>       COMMON /MYCOMMON / X(1)
> 
> with
> 
>       COMMON /MYCOMMON / X(2)
> 
> I get at compile tile with -O2
> 
> pr69368_1_a_1.f90:7:0:
> 
>           X(J+1)=X(J+7)
>  
> Warning: iteration 1 invokes undefined behavior
> [-Waggressive-loop-optimizations]
> pr69368_1_a_1.f90:6:0:
> 
>        DO 10 J=1,1024
>  
> note: within this loop
> 
> and the test succeeds. Any reason why there is no warning for "COMMON
> /MYCOMMON / X(1)"?

Because it's not a loop optimization that optimizes it.

> Also the following invalid code
> 
> [Book15] f90/bug% cat pr69368_1_c.f90
>       SUBROUTINE FOO(X)
>       IMPLICIT DOUBLE PRECISION (X)
>       INTEGER J
>       DIMENSION X(1)
>       DO 10 J=1,1024
>          X(J+1)=X(J+7)
>   10  CONTINUE
>       RETURN
>       END
> [Book15] f90/bug% cat pr69368_1_d.f90
>       IMPLICIT DOUBLE PRECISION (X)
>       DIMENSION X(1031)
>       DO I=1,1031
>       X(I)=I
>       END DO
>       call FOO(X)
>       print *, X(1025)
>       IF (X(1025).NE.1031.0) CALL ABORT
>       END
> 
> gives the expected result.

Well, undefined behavior doesn't guarantee a "not expected" result ;)