On Thu, 15 Dec 2011, Ira Rosen wrote:
>
>
> Jakub Jelinek <ja...@redhat.com> wrote on 15/12/2011 09:02:57 AM:
>
> > On Thu, Dec 15, 2011 at 08:32:26AM +0200, Ira Rosen wrote:
> > > > + cond = build2 (LT_EXPR, boolean_type_node, oprnd0, build_int_cst
> > > > (itype, 0));
> > > > + gsi = gsi_for_stmt (last_stmt);
> > > > + if (rhs_code == TRUNC_DIV_EXPR)
> > > > + {
> > > > + tree var = vect_recog_temp_ssa_var (itype, NULL);
> > > > + def_stmt
> > > > + = gimple_build_assign_with_ops3 (COND_EXPR, var, cond,
> > > > + fold_build2 (MINUS_EXPR, itype,
> > > > + oprnd1,
> > > > + build_int_cst (itype,
> > > > + 1)),
> > > > + build_int_cst (itype, 0));
> > > > + gsi_insert_before (&gsi, def_stmt, GSI_SAME_STMT);
> > >
> > > Hmm, you are inserting pattern stmts. This was causing some mess in the
> > > past as explained here
> > > http://gcc.gnu.org/ml/gcc-patches/2011-06/msg00801.html. Maybe you can
> use
> > > STMT_VINFO_PATTERN_DEF_STMT to keep a chain of def_stmts?
> >
> > Yes, I know, but STMT_VINFO_PATTERN_DEF_STMT contains a single gimple
> stmt,
> > while I need here several def stmts.
>
> > + S3 y_t = b_t < 0 ? N - 1 : 0;
> > + S2 x_t = b_t + y_t;
> > + S1' a_t = x_t >> log2 (N);
>
> I was talking about putting S3 in STMT_VINFO_PATTERN_DEF_STMT of S2.
>
> > I think it can't cause problems in
> > this case, the stmts will be easily DCEd.
>
> But it's really ugly to insert part of pattern sequence, don't you think?
It indeed is. The issue in the past was ICEing with -fno-tree-dce
when the pattern stmts did not have regular RTL expansion support
and the vectorization didn't trigger in the end.
We should really avoid inserting the pattern stmts.
Richard.
