This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [PATCH] Fix PR28900, non-gimple from tree-ifcvt
On Sun, 3 Sep 2006, Andrew Pinski wrote:
> On Sun, 2006-09-03 at 16:30 +0200, Richard Guenther wrote:
> > This fixes PR28900 by ensuring to properly gimplify trees created
> > by collecting conditions.
> >
> > Bootstrapped and tested on x86_64-unknown-linux-gnu.
> >
> > Ok for mainline?
> > Thanks,
> > Richard.
> >
> > :ADDPATCH middle-end:
> >
> > 2006-09-03 Richard Guenther <rguenther@suse.de>
> >
> > PR tree-optimization/28900
> > * tree-if-conv.c (find_phi_replacement_condition): Gimplify
> > compound conditional before creating COND_EXPR condition.
> >
> > * gcc.dg/torture/pr28900.c: New testcase
> >
> > Index: tree-if-conv.c
> > ===================================================================
> > *** tree-if-conv.c (revision 116660)
> > --- tree-if-conv.c (working copy)
> > *************** find_phi_replacement_condition (struct l
> > *** 666,672 ****
> > {
> > basic_block first_bb = NULL;
> > basic_block second_bb = NULL;
> > ! tree tmp_cond;
> >
> > gcc_assert (EDGE_COUNT (bb->preds) == 2);
> > first_bb = (EDGE_PRED (bb, 0))->src;
> > --- 666,672 ----
> > {
> > basic_block first_bb = NULL;
> > basic_block second_bb = NULL;
> > ! tree tmp_cond, new_stmts;
> >
> > gcc_assert (EDGE_COUNT (bb->preds) == 2);
> > first_bb = (EDGE_PRED (bb, 0))->src;
> > *************** find_phi_replacement_condition (struct l
> > *** 732,737 ****
> > --- 732,740 ----
> > value as condition. Various targets use different means to communicate
> > condition in vector compare operation. Using gimple value allows compiler
> > to emit vector compare and select RTL without exposing compare's result. */
> > + *cond = force_gimple_operand (*cond, &new_stmts, false, NULL_TREE);
> > + if (new_stmts)
> > + bsi_insert_before (bsi, new_stmts, BSI_SAME_STMT);
> > if (!is_gimple_reg (*cond) && !is_gimple_condexpr (*cond))
> > {
> > tree new_stmt;
>
> This force_gimple_operand should be done inside the if statement of the
> next check which makes it easier to understand why you are forcing it
> and should save some compile time too.
No, this doesn't work. A more consistent way would be to use
force_gimple_operand (, , true, ), but then ifcvt doesn't look through
conditions DEF appearantly, so a lot of ifcvt tests regress.
Richard.
--
Richard Guenther <rguenther@suse.de>
Novell / SUSE Labs