[Bug tree-optimization/33245] Missed opportunities for vectorization due to invariant condition

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

------- Comment #1 from dorit at gcc dot gnu dot org  2007-08-31 13:39 -------
(In reply to comment #0)
> The innermost loop in "j" cannot be vectorized because of the
> irregular code in that loop, i.e. the condition "IF ( l.NE.k )".  But
> the cond expression is invariant in that loop, so the whole condition
> can be hoisted outside that loop, versioning the loop, and potentially
> allowing the vectorization of the innermost loop.

if you use -O3 the condition *is* taken out of the loop by loop-unswitch (at
least that's what I see with revision 127623).

>       SUBROUTINE DGEFA(A,Lda,N,Ipvt,Info)
>       INTEGER Lda , N , Ipvt(*) , Info
>       DOUBLE PRECISION A(Lda,*)
>       DOUBLE PRECISION t
>       INTEGER IDAMAX , j , k , kp1 , l , nm1
>       Info = 0
>       nm1 = N - 1
>       IF ( nm1.GE.1 ) THEN
>          DO k = 1 , nm1
>             kp1 = k + 1
>             l = IDAMAX(N-k+1,A(k,k),1) + k - 1
>             Ipvt(k) = l
>             IF ( A(l,k).EQ.0.0D0 ) THEN
>                Info = k
>             ELSE
>                IF ( l.NE.k ) THEN
>                   t = A(l,k)
>                   A(l,k) = A(k,k)
>                   A(k,k) = t
>                ENDIF
>                t = -1.0D0/A(k,k)
>                CALL DSCAL(N-k,t,A(k+1,k),1)
>                DO j = kp1 , N
>                   t = A(l,j)
>                   IF ( l.NE.k ) THEN
>                      A(l,j) = A(k,j)
>                      A(k,j) = t
>                   ENDIF
>                   CALL DAXPY(N-k,t,A(k+1,k),1,A(k+1,j),1)
>                ENDDO
>             ENDIF
>          ENDDO
>       ENDIF
>       Ipvt(N) = N
>       IF ( A(N,N).EQ.0.0D0 ) Info = N
>       CONTINUE
>       END
> The result of the vectorizer on this testcase is:
> /home/seb/ex/linpk.f90:24: note: not vectorized: too many BBs in loop.
> /home/seb/ex/linpk.f90:24: note: bad loop form.
> /home/seb/ex/linpk.f90:1: note: vectorized 0 loops in function.
> Okay, if I'm versioning that loop by hand, I get the same error due to
> the PRE as for capacita.f90: the PRE inserts in the loop->latch block
> some code: 
>       <bb 11>:
>         # VUSE <PARM_NOALIAS.16_252> { PARM_NOALIAS.16 }
>         pretmp.47_297 = *n_13(D);
>         goto <bb 10>;

Looks like -fno-tree-pre is not enough, because if PRE doesn't do it, then sink
does it. When I use "-O3 -ftree-vectorize -msse2 -fno-tree-pre -fno-tree-sink"
I get the dataref problem you report below, without manual modifications to the
code

> And with PRE disabled, the fail occurs in the data ref analysis:
> ./linpk_corrected.f90:26: note: not vectorized: data ref analysis failed t.8_70
> = (*a_25(D))[D.1406_69]
> ./linpk_corrected.f90:26: note: bad data references.

Just for the record, this is the dataref problem that the dataref analyzer
reports:

Creating dr for t
analyze_innermost: (analyze_scalar_evolution
  (loop_nb = 3)
  (scalar = &t)
(get_scalar_evolution
  (scalar = &t)
  (scalar_evolution = ))
)
success.
        base_address: &t
        offset from base address: 0
        constant offset from base address: 0
        step: 0
        aligned to: 128
        base_object: t
        symbol tag: t
        FAILED as dr address is invariant


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33245