Re: Fwd: gm2 frontend in gcc 14 (fedora 40) bug

[Michael Riedl]

Gerlad,

a definition like

CARDINAL i,N;

BEGIN

           N := somevalue;

            FOR i:=N TO 0 BY -1 DO <loop body>  END;

is totally OK - that is nothing other than e.g.

          i:=N+1; WHILE (i > 0) DO DEC(i); <loop body> END;

hence I did not see a principle issue in your code ...

In most of my code I use CARDINALS if the value should never be negative 
- only the code I sent you is using INTEGERS because of the Fortran 
original.

Gruß

Michael


Am 30.04.24 um 21:48 schrieb Gerald Kasner:
> Hi Gaius,
>
> more on this...if one changes the type of the loop variables from
> cardinal to integer, the program runs fine.
>
> For me it, seems a conflict of types in a loop variable:
>
> In my understanding, I can count down a cardinal variable down to zero
> by a (noncardinal negative step) down to zero.
> Am I wrong?
>
> Best, Gerald
>
> Am 30.04.24 um 21:15 schrieb Gerald Kasner:
> > Hi Gaius,
> >
> > I nailed it down to the back substitution loop over i:
> > (The old fashioned way)
> > The loop does not go down to 1 but exits at (in this case) at i=4.
> >
> > I indicated the section in the attached code.
> > I hope this helps.
> >
> > Best, Gerald
> >
> >
> > Am 29.04.24 um 19:59 schrieb Gaius Mulley:
> > > Gerald Kasner <gerald.kasner@ovgu.de> writes:
> > >
> > > > Hi Gaius,
> > > >
> > > >
> > > > a quick test with gm2 (GCC) 14.0.1 20240411 (Red Hat 14.0.1-0) gave
> > > >
> > > > wrong results with a simple Gaussian elimination, while older 
> > > > gm2-frontends
> > > >
> > > > (gm2 (GCC) 13.2.1 20240316 (Red Hat 13.2.1-7)) give correct results
> > > >
> > > > test compiled with:
> > > >
> > > > gm2 -O2 -fiso testgauss.mod -lm
> > > >
> > > > Result with version 13.2.1 :
> > > >
> > > >        125.000000000000369482222595252097
> > > >      -2880.000000000007275957614183425903
> > > >      14490.000000000032741809263825416565
> > > >     -24640.000000000047293724492192268372
> > > >      13230.000000000023646862246096134186
> > > >
> > > > Result with 14.0.1:
> > > >
> > > >           1.000000000000000000000000000000
> > > >          1.500000000000000000000000000000
> > > >          1.166666666666666740681534975010
> > > >     -24640.000000000047293724492192268372
> > > >      13230.000000000023646862246096134186
> > > >
> > > > With  pivoted Gaussian elimination, the first three are NaN's.
> > > >
> > > > Any idea's ???
> > > >
> > > > testgauss.mod is attached here.
> > > >
> > > > Regards, Gerald
> > > Hi Gerald,
> > >
> > > many thanks for the bug report - now under investigation.  I tried the
> > > usual trick of:
> > >
> > > $ gm2 -fiso -fm2-whole-program -fsoft-check-all -O0 testgauss.mod
> > >
> > > but no runtime error was detected - I'm going to try check pointing 
> > > next
> > > (against gm2-13),
> > >
> > > regards,
> > > Gaius