Patchwork Fix vectorizable_reduction (PR tree-optimization/56350)

login
register
mail settings
Submitter Jakub Jelinek
Date Feb. 18, 2013, 6:56 p.m.
Message ID <20130218185646.GZ1215@tucnak.zalov.cz>
Download mbox | patch
Permalink /patch/221451/
State New
Headers show

Comments

Jakub Jelinek - Feb. 18, 2013, 6:56 p.m.
Hi!

On the following testcase, a = a / 8; looks like reasonable reduction
statement, but we pattern recognize it as
patt_1 = a < 0 ? 7 : 0;
patt_2 = a + patt_1;
a = patt2 >> 3;
and in the first pattern stmt the reduction operand is in the condition of
COND_EXPR, not operands as vectorizable_reduction was asserting.  Generally,
it couldn't be used anywhere in a pattern stmt, it could be just some
preparation statement.

Fixed thusly, bootstrapped/regtested on x86_64-linux and i686-linux, ok for
trunk?

2013-02-18  Jakub Jelinek  <jakub@redhat.com>

	PR tree-optimization/56350
	* tree-vect-loop.c (vectorizable_reduction): If orig_stmt, return false
	if haven't found reduction or nested cycle operand, rather than
	asserting we must find it.

	* gcc.dg/pr56350.c: New test.


	Jakub
Richard Guenther - Feb. 19, 2013, 9:01 a.m.
On Mon, 18 Feb 2013, Jakub Jelinek wrote:

> Hi!
> 
> On the following testcase, a = a / 8; looks like reasonable reduction
> statement, but we pattern recognize it as
> patt_1 = a < 0 ? 7 : 0;
> patt_2 = a + patt_1;
> a = patt2 >> 3;
> and in the first pattern stmt the reduction operand is in the condition of
> COND_EXPR, not operands as vectorizable_reduction was asserting.  Generally,
> it couldn't be used anywhere in a pattern stmt, it could be just some
> preparation statement.
> 
> Fixed thusly, bootstrapped/regtested on x86_64-linux and i686-linux, ok for
> trunk?

Ok.

Thanks,
Richard.

> 2013-02-18  Jakub Jelinek  <jakub@redhat.com>
> 
> 	PR tree-optimization/56350
> 	* tree-vect-loop.c (vectorizable_reduction): If orig_stmt, return false
> 	if haven't found reduction or nested cycle operand, rather than
> 	asserting we must find it.
> 
> 	* gcc.dg/pr56350.c: New test.
> 
> --- gcc/tree-vect-loop.c.jj	2013-01-15 10:30:19.000000000 +0100
> +++ gcc/tree-vect-loop.c	2013-02-18 11:32:30.306566358 +0100
> @@ -4692,7 +4692,7 @@ vectorizable_reduction (gimple stmt, gim
>       The last use is the reduction variable.  In case of nested cycle this
>       assumption is not true: we use reduc_index to record the index of the
>       reduction variable.  */
> -  for (i = 0; i < op_type-1; i++)
> +  for (i = 0; i < op_type - 1; i++)
>      {
>        /* The condition of COND_EXPR is checked in vectorizable_condition().  */
>        if (i == 0 && code == COND_EXPR)
> @@ -4724,11 +4724,18 @@ vectorizable_reduction (gimple stmt, gim
>    if (!vectype_in)
>      vectype_in = tem;
>    gcc_assert (is_simple_use);
> -  gcc_assert (dt == vect_reduction_def
> -              || dt == vect_nested_cycle
> -              || ((dt == vect_internal_def || dt == vect_external_def
> -                   || dt == vect_constant_def || dt == vect_induction_def)
> -                   && nested_cycle && found_nested_cycle_def));
> +  if (!(dt == vect_reduction_def
> +	|| dt == vect_nested_cycle
> +	|| ((dt == vect_internal_def || dt == vect_external_def
> +	     || dt == vect_constant_def || dt == vect_induction_def)
> +	    && nested_cycle && found_nested_cycle_def)))
> +    {
> +      /* For pattern recognized stmts, orig_stmt might be a reduction,
> +	 but some helper statements for the pattern might not, or
> +	 might be COND_EXPRs with reduction uses in the condition.  */
> +      gcc_assert (orig_stmt);
> +      return false;
> +    }
>    if (!found_nested_cycle_def)
>      reduc_def_stmt = def_stmt;
>  
> --- gcc/testsuite/gcc.dg/pr56350.c.jj	2013-02-18 11:30:12.117327599 +0100
> +++ gcc/testsuite/gcc.dg/pr56350.c	2013-02-18 11:29:59.000000000 +0100
> @@ -0,0 +1,13 @@
> +/* PR tree-optimization/56350 */
> +/* { dg-do compile } */
> +/* { dg-options "-O -ftree-vectorize" } */
> +
> +int a, b, c;
> +
> +void
> +f (void)
> +{
> +  for (; c; c++)
> +    for (b = 0; b < 2; b++)
> +      a /= 8;
> +}
> 
> 	Jakub
> 
>

Patch

--- gcc/tree-vect-loop.c.jj	2013-01-15 10:30:19.000000000 +0100
+++ gcc/tree-vect-loop.c	2013-02-18 11:32:30.306566358 +0100
@@ -4692,7 +4692,7 @@  vectorizable_reduction (gimple stmt, gim
      The last use is the reduction variable.  In case of nested cycle this
      assumption is not true: we use reduc_index to record the index of the
      reduction variable.  */
-  for (i = 0; i < op_type-1; i++)
+  for (i = 0; i < op_type - 1; i++)
     {
       /* The condition of COND_EXPR is checked in vectorizable_condition().  */
       if (i == 0 && code == COND_EXPR)
@@ -4724,11 +4724,18 @@  vectorizable_reduction (gimple stmt, gim
   if (!vectype_in)
     vectype_in = tem;
   gcc_assert (is_simple_use);
-  gcc_assert (dt == vect_reduction_def
-              || dt == vect_nested_cycle
-              || ((dt == vect_internal_def || dt == vect_external_def
-                   || dt == vect_constant_def || dt == vect_induction_def)
-                   && nested_cycle && found_nested_cycle_def));
+  if (!(dt == vect_reduction_def
+	|| dt == vect_nested_cycle
+	|| ((dt == vect_internal_def || dt == vect_external_def
+	     || dt == vect_constant_def || dt == vect_induction_def)
+	    && nested_cycle && found_nested_cycle_def)))
+    {
+      /* For pattern recognized stmts, orig_stmt might be a reduction,
+	 but some helper statements for the pattern might not, or
+	 might be COND_EXPRs with reduction uses in the condition.  */
+      gcc_assert (orig_stmt);
+      return false;
+    }
   if (!found_nested_cycle_def)
     reduc_def_stmt = def_stmt;
 
--- gcc/testsuite/gcc.dg/pr56350.c.jj	2013-02-18 11:30:12.117327599 +0100
+++ gcc/testsuite/gcc.dg/pr56350.c	2013-02-18 11:29:59.000000000 +0100
@@ -0,0 +1,13 @@ 
+/* PR tree-optimization/56350 */
+/* { dg-do compile } */
+/* { dg-options "-O -ftree-vectorize" } */
+
+int a, b, c;
+
+void
+f (void)
+{
+  for (; c; c++)
+    for (b = 0; b < 2; b++)
+      a /= 8;
+}