On 11 November 2011 17:32, Jakub Jelinek <ja...@redhat.com> wrote:
> Hi!
Hi,
>
> Removing the scalar call in vectorizable_call for SLP vectorization
> is too early, when another SLP instance refers to the same scalar call,
> we'll ICE because that stmt doesn't have bb anymore or gsi_for_stmt
> doesn't succeed for it.
>
> Fixed by postponing replacement of calls with zeroing of lhs for later
> in the SLP case.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
>
> 2011-11-11 Jakub Jelinek <ja...@redhat.com>
>
> PR tree-optimization/51058
> * tree-vect-slp.c (vect_remove_slp_scalar_calls): New function.
> (vect_schedule_slp): Call it.
> * tree-vect-stmts.c (vectorizable_call): If slp_node != NULL,
> don't replace scalar calls with clearing of their lhs here.
I think it's rhs.
>
> * gcc.dg/vect/fast-math-vect-call-1.c: Add f4 test.
> * gfortran.fortran-torture/compile/pr51058.f90: New test.
>
> --- gcc/tree-vect-slp.c.jj 2011-11-10 18:09:12.000000000 +0100
> +++ gcc/tree-vect-slp.c 2011-11-11 13:18:42.157292895 +0100
> @@ -2898,6 +2898,46 @@ vect_schedule_slp_instance (slp_tree nod
> return is_store;
> }
>
> +/* Replace scalar calls from SLP node NODE with clearing of their lhs.
Here too.
> + For loop vectorization this is done in vectorizable_call, but for SLP
> + it needs to be deferred until end of vect_schedule_slp, because multiple
> + SLP instances may refer to the same scalar stmt. */
> +
> +static void
> +vect_remove_slp_scalar_calls (slp_tree node)
> +{
...
> --- gcc/testsuite/gfortran.fortran-torture/compile/pr51058.f90.jj
> 2011-11-11 13:26:14.665615842 +0100
> +++ gcc/testsuite/gfortran.fortran-torture/compile/pr51058.f90 2011-11-11
> 13:25:50.000000000 +0100
> @@ -0,0 +1,18 @@
> +! PR tree-optimization/51058
> +! { dg-do compile }
> +subroutine pr51058(n, u, v, w, z)
> + double precision :: x(3,-2:16384), y(3,-2:16384), b, u, v, w, z
> + integer :: i, n
> + common /c/ x, y
> + do i = 1, n
> + b = u * int(x(1,i)) + sign(z,x(1,i))
> + x(1,i) = x(1,i) - b
> + y(1,i) = y(1,i) - b
> + b = v * int(x(2,i)) + sign(z,x(2,i))
> + x(2,i) = x(2,i) - b
> + y(2,i) = y(2,i) - b
> + b = w * int(x(3,i)) + sign(z,x(3,i))
> + x(3,i) = x(3,i) - b
> + y(3,i) = y(3,i) - b
> + end do
> +end subroutine
Please add
! { dg-final { cleanup-tree-dump "vect" } }
OK otherwise.
Thanks,
Ira
>
> Jakub
>
