This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [PATCH] Fix non-INTEGER_CST step vectorization (PR tree-optimization/57741)
- From: Richard Biener <rguenther at suse dot de>
- To: Jakub Jelinek <jakub at redhat dot com>, Jakub Jelinek <jakub at redhat dot com>
- Cc: gcc-patches at gcc dot gnu dot org, Pat Haugen <pthaugen at linux dot vnet dot ibm dot com>
- Date: Tue, 02 Jul 2013 13:21:55 +0200
- Subject: Re: [PATCH] Fix non-INTEGER_CST step vectorization (PR tree-optimization/57741)
- References: <20130628062629 dot GK2336 at tucnak dot redhat dot com>
Jakub Jelinek <jakub@redhat.com> wrote:
>Hi!
>
>My recent patch introduced a regression as shown by the attached
>testcases, there was nothing actually checking the type of the
>induction,
>but because we allowed only INTEGER_CST steps that implied
>integral/pointer
>types only. With -ffast-math I don't see a reason why we can't handle
>floating point types the same, so the patch adds type checks to
>vect_is_simple_iv_evolution so that we won't be surprised by fixed
>point,
>vector (and whatever else comes later) inductions, and handles the
>REAL_TYPE
>steps in get_initial_def_for_induction.
>
>Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
Ok.
Thanks,
Richard.
>2013-06-28 Jakub Jelinek <jakub@redhat.com>
>
> PR tree-optimization/57741
> * tree-vect-loop.c (vect_is_simple_iv_evolution): Disallow
> non-INTEGRAL_TYPE_P non-SCALAR_FLOAT_TYPE_P SSA_NAME step_exprs,
> or SCALAR_FLOAT_TYPE_P SSA_NAMEs if !flag_associative_math.
> Allow REAL_CST step_exprs if flag_associative_math.
> (get_initial_def_for_induction): Handle SCALAR_FLOAT_TYPE_P step_expr.
>
> * gcc.dg/vect/pr57741-1.c: New test.
> * gcc.dg/vect/pr57741-2.c: New test.
> * gcc.dg/vect/pr57741-3.c: New test.
>
>--- gcc/tree-vect-loop.c.jj 2013-06-25 14:34:33.000000000 +0200
>+++ gcc/tree-vect-loop.c 2013-06-28 01:07:42.524553908 +0200
>@@ -538,7 +538,12 @@ vect_is_simple_iv_evolution (unsigned lo
> if (TREE_CODE (step_expr) != INTEGER_CST
> && (TREE_CODE (step_expr) != SSA_NAME
> || ((bb = gimple_bb (SSA_NAME_DEF_STMT (step_expr)))
>- && flow_bb_inside_loop_p (get_loop (cfun, loop_nb), bb))))
>+ && flow_bb_inside_loop_p (get_loop (cfun, loop_nb), bb))
>+ || (!INTEGRAL_TYPE_P (TREE_TYPE (step_expr))
>+ && (!SCALAR_FLOAT_TYPE_P (TREE_TYPE (step_expr))
>+ || !flag_associative_math)))
>+ && (TREE_CODE (step_expr) != REAL_CST
>+ || !flag_associative_math))
> {
> if (dump_enabled_p ())
> dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
>@@ -3276,7 +3281,13 @@ get_initial_def_for_induction (gimple iv
> {
> /* iv_loop is the loop to be vectorized. Generate:
> vec_step = [VF*S, VF*S, VF*S, VF*S] */
>- expr = build_int_cst (TREE_TYPE (step_expr), vf);
>+ if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (step_expr)))
>+ {
>+ expr = build_int_cst (integer_type_node, vf);
>+ expr = fold_convert (TREE_TYPE (step_expr), expr);
>+ }
>+ else
>+ expr = build_int_cst (TREE_TYPE (step_expr), vf);
> new_name = fold_build2 (MULT_EXPR, TREE_TYPE (step_expr),
> expr, step_expr);
> if (TREE_CODE (step_expr) == SSA_NAME)
>@@ -3339,7 +3350,13 @@ get_initial_def_for_induction (gimple iv
> gcc_assert (!nested_in_vect_loop);
>
> /* Create the vector that holds the step of the induction. */
>- expr = build_int_cst (TREE_TYPE (step_expr), nunits);
>+ if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (step_expr)))
>+ {
>+ expr = build_int_cst (integer_type_node, nunits);
>+ expr = fold_convert (TREE_TYPE (step_expr), expr);
>+ }
>+ else
>+ expr = build_int_cst (TREE_TYPE (step_expr), nunits);
> new_name = fold_build2 (MULT_EXPR, TREE_TYPE (step_expr),
> expr, step_expr);
> if (TREE_CODE (step_expr) == SSA_NAME)
>--- gcc/testsuite/gcc.dg/vect/pr57741-1.c.jj 2013-06-27
>23:29:07.322359740 +0200
>+++ gcc/testsuite/gcc.dg/vect/pr57741-1.c 2013-06-27 23:54:18.216533899
>+0200
>@@ -0,0 +1,21 @@
>+/* PR tree-optimization/57741 */
>+/* { dg-do compile } */
>+
>+void
>+foo (float *p, float *q, float x)
>+{
>+ int i;
>+ float f = 1.0f, g = 2.0f;
>+ for (i = 0; i < 1024; i++)
>+ {
>+ *p++ = f;
>+ f += x;
>+ }
>+ for (i = 0; i < 1024; i++)
>+ {
>+ *q++ = g;
>+ g += 0.5f;
>+ }
>+}
>+
>+/* { dg-final { cleanup-tree-dump "vect" } } */
>--- gcc/testsuite/gcc.dg/vect/pr57741-2.c.jj 2013-06-27
>23:44:40.846538237 +0200
>+++ gcc/testsuite/gcc.dg/vect/pr57741-2.c 2013-06-27 23:50:23.552498840
>+0200
>@@ -0,0 +1,44 @@
>+/* PR tree-optimization/57741 */
>+/* { dg-do run } */
>+/* { dg-additional-options "-ffast-math" } */
>+
>+#include "tree-vect.h"
>+
>+extern void abort (void);
>+
>+__attribute__((noinline, noclone)) void
>+foo (float *p, float *q, float x)
>+{
>+ int i;
>+ p = (float *) __builtin_assume_aligned (p, 32);
>+ q = (float *) __builtin_assume_aligned (q, 32);
>+ float f = 1.0f, g = 2.0f;
>+ for (i = 0; i < 1024; i++)
>+ {
>+ *p++ = f;
>+ f += x;
>+ }
>+ for (i = 0; i < 1024; i++)
>+ {
>+ *q++ = g;
>+ g += 0.5f;
>+ }
>+}
>+
>+float p[1024] __attribute__((aligned (32))) = { 17.0f };
>+float q[1024] __attribute__((aligned (32))) = { 17.0f };
>+
>+int
>+main ()
>+{
>+ int i;
>+ check_vect ();
>+ foo (p, q, 1.5f);
>+ for (i = 0; i < 1024; i++)
>+ if (p[i] != 1.0f + i * 1.5f || q[i] != 2.0f + i * 0.5f)
>+ abort ();
>+ return 0;
>+}
>+
>+/* { dg-final { scan-tree-dump-times "vectorized 2 loop" 1 "vect" } }
>*/
>+/* { dg-final { cleanup-tree-dump "vect" } } */
>--- gcc/testsuite/gcc.dg/vect/pr57741-3.c.jj 2013-06-28
>01:08:29.530334797 +0200
>+++ gcc/testsuite/gcc.dg/vect/pr57741-3.c 2013-06-28 01:11:31.365141459
>+0200
>@@ -0,0 +1,42 @@
>+/* PR tree-optimization/57741 */
>+/* { dg-do run } */
>+/* { dg-additional-options "-ffast-math" } */
>+
>+#include "tree-vect.h"
>+
>+extern void abort (void);
>+
>+float p[1024] __attribute__((aligned (32))) = { 17.0f };
>+float q[1024] __attribute__((aligned (32))) = { 17.0f };
>+char r[1024] __attribute__((aligned (32))) = { 1 };
>+
>+__attribute__((noinline, noclone)) void
>+foo (float x)
>+{
>+ int i;
>+ float f = 1.0f, g = 2.0f;
>+ for (i = 0; i < 1024; i++)
>+ {
>+ p[i] = f;
>+ f += x;
>+ q[i] = g;
>+ g += 0.5f;
>+ r[i]++;
>+ }
>+}
>+
>+int
>+main ()
>+{
>+ int i;
>+ check_vect ();
>+ r[0] = 0;
>+ foo (1.5f);
>+ for (i = 0; i < 1024; i++)
>+ if (p[i] != 1.0f + i * 1.5f || q[i] != 2.0f + i * 0.5f || r[i] !=
>1)
>+ abort ();
>+ return 0;
>+}
>+
>+/* { dg-final { scan-tree-dump-times "vectorized 1 loop" 1 "vect" } }
>*/
>+/* { dg-final { cleanup-tree-dump "vect" } } */
>
> Jakub