[PATCH] Fix up ICE with reduction % (PR tree-optimization/55955)

[PATCH] Fix up ICE with reduction % (PR tree-optimization/55955)

[Jakub Jelinek]

Hi!

For *SHIFT_EXPR/*ROTATE_EXPR with VECTOR_TYPE types, optab_for_tree_code
requires us to decide if we want optab_vector or optab_scalar.  It is likely
not worth it to handle these four, they would fail soon anyway.

Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?

2013-01-14  Jakub Jelinek  <jakub@redhat.com>

	PR tree-optimization/55955
	* tree-vect-loop.c (vectorizable_reduction): Give up early on
	*SHIFT_EXPR and *ROTATE_EXPR codes.

	* gcc.c-torture/compile/pr55955.c: New test.

--- gcc/tree-vect-loop.c.jj	2013-01-11 09:02:55.000000000 +0100
+++ gcc/tree-vect-loop.c	2013-01-14 12:21:29.359018748 +0100
@@ -4776,6 +4776,17 @@ vectorizable_reduction (gimple stmt, gim
     {
       /* 4. Supportable by target?  */
 
+      if (code == LSHIFT_EXPR || code == RSHIFT_EXPR
+	  || code == LROTATE_EXPR || code == RROTATE_EXPR)
+	{
+	  /* Shifts and rotates are only supported by vectorizable_shifts,
+	     not vectorizable_reduction.  */
+          if (dump_enabled_p ())
+	    dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
+			     "unsupported shift or rotation.");
+	  return false;
+	}
+
       /* 4.1. check support for the operation in the loop  */
       optab = optab_for_tree_code (code, vectype_in, optab_default);
       if (!optab)
--- gcc/testsuite/gcc.c-torture/compile/pr55955.c.jj	2013-01-14 12:23:28.273324350 +0100
+++ gcc/testsuite/gcc.c-torture/compile/pr55955.c	2013-01-14 12:23:14.000000000 +0100
@@ -0,0 +1,12 @@
+/* PR tree-optimization/55955 */
+
+int b;
+
+void
+foo (int x)
+{
+  int a;
+  for (a = x; a < 2; a++)
+    for (b = 0; b < 2; b++)
+      *(unsigned short *) 0x100000UL %= 46;
+}

	Jakub

Re: [PATCH] Fix up ICE with reduction % (PR tree-optimization/55955)

[Richard Biener]

On Mon, 14 Jan 2013, Jakub Jelinek wrote:

> Hi!
> 
> For *SHIFT_EXPR/*ROTATE_EXPR with VECTOR_TYPE types, optab_for_tree_code
> requires us to decide if we want optab_vector or optab_scalar.  It is likely
> not worth it to handle these four, they would fail soon anyway.
> 
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?

Ok.

Thanks,
Richard.

> 2013-01-14  Jakub Jelinek  <jakub@redhat.com>
> 
> 	PR tree-optimization/55955
> 	* tree-vect-loop.c (vectorizable_reduction): Give up early on
> 	*SHIFT_EXPR and *ROTATE_EXPR codes.
> 
> 	* gcc.c-torture/compile/pr55955.c: New test.
> 
> --- gcc/tree-vect-loop.c.jj	2013-01-11 09:02:55.000000000 +0100
> +++ gcc/tree-vect-loop.c	2013-01-14 12:21:29.359018748 +0100
> @@ -4776,6 +4776,17 @@ vectorizable_reduction (gimple stmt, gim
>      {
>        /* 4. Supportable by target?  */
>  
> +      if (code == LSHIFT_EXPR || code == RSHIFT_EXPR
> +	  || code == LROTATE_EXPR || code == RROTATE_EXPR)
> +	{
> +	  /* Shifts and rotates are only supported by vectorizable_shifts,
> +	     not vectorizable_reduction.  */
> +          if (dump_enabled_p ())
> +	    dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
> +			     "unsupported shift or rotation.");
> +	  return false;
> +	}
> +
>        /* 4.1. check support for the operation in the loop  */
>        optab = optab_for_tree_code (code, vectype_in, optab_default);
>        if (!optab)
> --- gcc/testsuite/gcc.c-torture/compile/pr55955.c.jj	2013-01-14 12:23:28.273324350 +0100
> +++ gcc/testsuite/gcc.c-torture/compile/pr55955.c	2013-01-14 12:23:14.000000000 +0100
> @@ -0,0 +1,12 @@
> +/* PR tree-optimization/55955 */
> +
> +int b;
> +
> +void
> +foo (int x)
> +{
> +  int a;
> +  for (a = x; a < 2; a++)
> +    for (b = 0; b < 2; b++)
> +      *(unsigned short *) 0x100000UL %= 46;
> +}
> 
> 	Jakub
> 
> 

-- 
Richard Biener <rguenther@suse.de>
SUSE / SUSE Labs
SUSE LINUX Products GmbH - Nuernberg - AG Nuernberg - HRB 16746
GF: Jeff Hawn, Jennifer Guild, Felix Imend