[PATCH] Fix PR88775

[PATCH] Fix PR88775

[Richard Biener]

I am testing the following patch teaching VRP predicate analysis about

  __x.5_4 = (long unsigned int) "hello";
  __y.6_5 = (long unsigned int) _3;
  if (__x.5_4 != __y.6_5)

so that we know sth about the relation of the converted entities.
This appearantly (didn't back out other stuff) helps PR88775
after Jakubs changes to libstdc++ (before his changes a related
VN patch helped which meanwhile shows miscompiling
20_util/function_objects/comparisons_pointer.cc...).

I now see DOM performing all required optimization thanks to it
using the EVRP machinery.

Bootstrap & regtest running on x86_64-unknown-linux-gnu.

OK for trunk?

Any idea how we can have a reliable testcase for the std::string
optimization?

Thanks,
Richard.

From 78a345845651565daac295f8dfbfc64cf5e8ccf3 Mon Sep 17 00:00:00 2001
From: Richard Guenther <rguenther@suse.de>
Date: Thu, 10 Jan 2019 14:34:22 +0100
Subject: [PATCH] fix-pr88775-2

2019-01-10  Richard Biener  <rguenther@suse.de>

	PR tree-optimization/88775
	* tree-vrp.c (register_edge_assert_for_2): Register asserts
	from (T) a CMP (T) b.

diff --git a/gcc/tree-vrp.c b/gcc/tree-vrp.c
index 8d18e19d6e4..1efb907ae5e 100644
--- a/gcc/tree-vrp.c
+++ b/gcc/tree-vrp.c
@@ -3247,6 +3247,42 @@ register_edge_assert_for_2 (tree name, edge e,
 	    }
 	}
     }
+
+  /* For things like (T)a CMP (T)b register asserts for a CMP b if possible.  */
+  if (TREE_CODE_CLASS (comp_code) == tcc_comparison
+      && TREE_CODE (val) == SSA_NAME
+      && (INTEGRAL_TYPE_P (TREE_TYPE (val))
+	  || POINTER_TYPE_P (TREE_TYPE (val))))
+    {
+      gassign *def1 = dyn_cast <gassign *> (SSA_NAME_DEF_STMT (name));
+      gassign *def2 = dyn_cast <gassign *> (SSA_NAME_DEF_STMT (val));
+      if (def1 && def2
+	  && CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (def1))
+	  && CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (def2))
+	  && types_compatible_p (TREE_TYPE (gimple_assign_rhs1 (def1)),
+				 TREE_TYPE (gimple_assign_rhs1 (def2)))
+	  && (TYPE_PRECISION (TREE_TYPE (gimple_assign_rhs1 (def1)))
+	      <= TYPE_PRECISION (TREE_TYPE (val)))
+	  && (comp_code == EQ_EXPR
+	      || comp_code == NE_EXPR
+	      || (TYPE_SIGN (TREE_TYPE (val))
+		  == TYPE_SIGN (TREE_TYPE (gimple_assign_rhs1 (def1))))
+	      || (TYPE_UNSIGNED (TREE_TYPE (gimple_assign_rhs1 (def1)))
+		  && (TYPE_PRECISION (TREE_TYPE (gimple_assign_rhs1 (def1)))
+		      < TYPE_PRECISION (TREE_TYPE (val))))))
+	{
+	  tree op0 = gimple_assign_rhs1 (def1);
+	  tree op1 = gimple_assign_rhs1 (def2);
+	  enum tree_code alt_comp_code = comp_code;
+	  if (TREE_CODE (op0) != SSA_NAME)
+	    {
+	      alt_comp_code = swap_tree_comparison (alt_comp_code);
+	      std::swap (op0, op1);
+	    }
+	  if (TREE_CODE (op0) == SSA_NAME)
+	    add_assert_info (asserts, op0, op0, alt_comp_code, op1);
+	}
+    }
 }
 
 /* OP is an operand of a truth value expression which is known to have

Re: [PATCH] Fix PR88775

[Jeff Law]

On 1/10/19 6:44 AM, Richard Biener wrote:
> 
> I am testing the following patch teaching VRP predicate analysis about
> 
>   __x.5_4 = (long unsigned int) "hello";
>   __y.6_5 = (long unsigned int) _3;
>   if (__x.5_4 != __y.6_5)
> 
> so that we know sth about the relation of the converted entities.
> This appearantly (didn't back out other stuff) helps PR88775
> after Jakubs changes to libstdc++ (before his changes a related
> VN patch helped which meanwhile shows miscompiling
> 20_util/function_objects/comparisons_pointer.cc...).
> 
> I now see DOM performing all required optimization thanks to it
> using the EVRP machinery.
> 
> Bootstrap & regtest running on x86_64-unknown-linux-gnu.
> 
> OK for trunk?
> 
> Any idea how we can have a reliable testcase for the std::string
> optimization?
> 
> Thanks,
> Richard.
> 
> From 78a345845651565daac295f8dfbfc64cf5e8ccf3 Mon Sep 17 00:00:00 2001
> From: Richard Guenther <rguenther@suse.de>
> Date: Thu, 10 Jan 2019 14:34:22 +0100
> Subject: [PATCH] fix-pr88775-2
> 
> 2019-01-10  Richard Biener  <rguenther@suse.de>
> 
> 	PR tree-optimization/88775
> 	* tree-vrp.c (register_edge_assert_for_2): Register asserts
> 	from (T) a CMP (T) b.
Seems reasonable to me.  It looks like a slight generalization of code
I've seen elsewhere, but can't find at the moment.

Jeff