[Bug fortran/54932] Invalid loop code generated by Fortran FE for loops with bounds in HIGH(type)

["dominiq at lps dot ens.fr" <gcc-bugzilla@gcc.gnu.org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54932

--- Comment #12 from Dominique d'Humieres <dominiq at lps dot ens.fr> 2013-02-01 13:59:11 UTC ---
(In reply to comment #11)
> > > Thus, I close the bug as INVALID.
> > ... in wich case could you, please, update the testcase to be valid and remove
> > the XFAIL I introduced?
>
> We jump through some hoops in or DO loop code generation to execute
> a loop until HUGE(i) in a way that somebody who did not read the
> standard well might expect, but which is actually invalid.
>
> If we do not do this any more, then we can probably simplify our DO
> loops considerably.

This is probably too late for 4.8.0. The following patch takes advantage of the
new option -fno-aggressive-loop-optimizations to remove the xfail (and the two
XPASS at -O0 and -O1):

--- /opt/gcc/_gcc_clean/gcc/testsuite/gfortran.dg/do_1.f90    2012-10-18
00:34:47.000000000 +0200
+++ /opt/gcc/work/gcc/testsuite/gfortran.dg/do_1.f90    2013-01-31
22:02:16.000000000 +0100
@@ -1,5 +1,5 @@
-! { dg-do run { xfail *-*-* } }
-! XFAIL is tracked in PR 54932
+! { dg-do run }
+! { dg-additional-options "-fno-aggressive-loop-optimizations" }
 ! Program to check corner cases for DO statements.
 program do_1
   implicit none

Tested on x86_64-apple-darwin10 and powerpc-apple-darwin9.

> We should also warn about invalid code in the FE.

One example is

program test_real_do
    integer n
    integer x
    n=0
    do x=0.0,10.0,0.1
        n=n+1
    end do
    print *, "N is ", n
end program test_real_do

which on x86_64-apple-darwin10 gives

Program received signal SIGFPE: Floating-point exception - erroneous arithmetic
operation.

when executed. This is due to the division by zero in

    x = 0;
    countm1.0 = 10 / 0;
    while (1)
      {
        {
          unsigned int countm1t.1;

          n = n + 1;
          L.1:;
          x = NON_LVALUE_EXPR <x>;
          countm1t.1 = countm1.0;
          countm1.0 = countm1.0 + 4294967295;
          if (countm1t.1 == 0) goto L.2;
        }
      }
    L.2:;

The FE should detect that the stride is zero and output an error.

Note that on powerpc-apple-darwin9, the same test (with the same
dump-tree-original) runs and outputs

 N is            1

!?-(