[Bug middle-end/42181] [4.5 Regression][graphite] -fgraphite-identity miscompiles air.f90

[spop at gcc dot gnu dot org]

------- Comment #23 from spop at gcc dot gnu dot org  2010-03-09 22:31 -------
Thanks for reducing this testcase.

On the further reduced kernel:

      SUBROUTINE SPECTOP(Dr,N)
      IMPLICIT REAL*8(A-H,o-Z)
      DIMENSION d1(0:32,0:32) , Dr(0:32,0:32) , x(0:32)
      REAL*8 Dr

      DO k = 1 , N-1
         d1(k,0) = fctr1/(t*(x(k)-x(0)))
         fctr = -fctr1
         DO j = 1 , N-1
            if (j.ne.k) d1(k,j) = fctr/(x(k)-x(j))
            fctr = -fctr
         ENDDO
         d1(k,N) = fctr/(t*(x(k)-x(N)))
         fctr1 = -fctr1
      ENDDO

      Dr(k,j) = d1(N-k,N-j)
      CONTINUE
      END

We have this CLAST generated by CLooG:

if (T_10 >= 3) {
  S4(0) ;
  S5(0,0) ;
  S7(0,0) ;
  for (scat_3=1;scat_3<=T_10-2;scat_3++) {
    S5(0,scat_3) ;
    S6(0,scat_3) ;
    S7(0,scat_3) ;
  }
  S9(0) ;
  S14(0) ;
}
for (scat_1=1;scat_1<=T_10-3;scat_1++) {
  S4(scat_1) ;
  S5(scat_1,scat_1) ;
  S7(scat_1,scat_1) ;
  for (scat_3=scat_1+1;scat_3<=T_10-2;scat_3++) {
    S5(scat_1,scat_3) ;
    S6(scat_1,scat_3) ;
    S7(scat_1,scat_3) ;
  }
  for (scat_3=0;scat_3<=scat_1-1;scat_3++) {
    S5(scat_1,scat_3) ;
    S6(scat_1,scat_3) ;
    S7(scat_1,scat_3) ;
  }
  S9(scat_1) ;
  S14(scat_1) ;
}
if (T_10 >= 3) {
  S4(T_10-2) ;
  S5(T_10-2,T_10-2) ;
  S7(T_10-2,T_10-2) ;
  for (scat_3=0;scat_3<=T_10-3;scat_3++) {
    S5(T_10-2,scat_3) ;
    S6(T_10-2,scat_3) ;
    S7(T_10-2,scat_3) ;
  }
  S9(T_10-2) ;
  S14(T_10-2) ;
}
if (T_10 == 2) {
  S4(0) ;
  S5(0,0) ;
  S7(0,0) ;
  S9(0) ;
  S14(0) ;
}

Where T_10 stands for the parameter N,
S4 stands for
         d1(k,0) = fctr1/(t*(x(k)-x(0)))
         fctr = -fctr1
S5 stands for "if (j.ne.k)"
S6 stands for "d1(k,j) = fctr/(x(k)-x(j))"
S7 stands for "fctr = -fctr".
S14 stands for
         d1(k,N) = fctr/(t*(x(k)-x(N)))
         fctr1 = -fctr1

So the error seems to be that we are splitting the index scat_3 into
two ranges that are not ordered in the same way as in the original
loop nest:

for (scat_1=1;scat_1<=T_10-3;scat_1++) {
  S4(scat_1) ;
  S5(scat_1,scat_1) ;
  S7(scat_1,scat_1) ;
  for (scat_3=scat_1+1;scat_3<=T_10-2;scat_3++) {
    S5(scat_1,scat_3) ;
    S6(scat_1,scat_3) ;
    S7(scat_1,scat_3) ;
  }
  for (scat_3=0;scat_3<=scat_1-1;scat_3++) {
    S5(scat_1,scat_3) ;
    S6(scat_1,scat_3) ;
    S7(scat_1,scat_3) ;
  }
  S9(scat_1) ;
  S14(scat_1) ;
}

We are executing the range [scat_1+1, T_10-2] before executing the
range [0, scat_1-1].


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=42181