Teach vrp that THIS pointer and references are non-zero

Teach vrp that THIS pointer and references are non-zero

[Jan Hubicka]

Hi,
with multiple inheritance compiling the testcase bellow, the first call always
leads to call with offseted pointer, while the other call results in if
conditoinal testing if parameter is non-NULL.

This patch teach VRP that THIS pointers and references are non-zero.  I hope
this is true also for fortran and ada, but I think both languages do not really
have NULL.

I think VRP could basically assume all reference pointers to be non-zero, but
I am not sure how to do that with GIMPLE useless conversion rules.

Bootstrapped/regtested x86_64-linux, OK?

Honza

	* g++.dg/tree-ssa/nonzero-3.C: New testcase.
	* tree-vrp.c (nonnull_arg_p): THIS pointers and references are non-zero.
	(gimple_stmt_nonzero_warnv_p): Reference return values are non-zero.
Index: testsuite/g++.dg/tree-ssa/nonzero-3.C
===================================================================
--- testsuite/g++.dg/tree-ssa/nonzero-3.C	(revision 0)
+++ testsuite/g++.dg/tree-ssa/nonzero-3.C	(working copy)
@@ -0,0 +1,22 @@
+/* { dg-do compile } */
+/* { dg-options "-O2 -fdump-tree-vrp1 -fdelete-null-pointer-checks" } */
+struct A {int a;};
+struct B {int b;};
+struct C:A,B {int c;
+  void bar();};
+
+void foo (struct B *);
+void C::bar ()
+{
+  struct C *d = this;
+  foo(this);
+  foo(d);
+}
+void bar (struct C &c)
+{
+  struct C *d = &c;
+  foo(&c);
+  foo(d);
+}
+/* { dg-final { scan-tree-dump-not "if \\(" "vrp1"} } */
+/* { dg-final { cleanup-tree-dump "vrp1" } } */
Index: tree-vrp.c
===================================================================
--- tree-vrp.c	(revision 222016)
+++ tree-vrp.c	(working copy)
@@ -393,6 +393,17 @@ nonnull_arg_p (const_tree arg)
   if (arg == cfun->static_chain_decl)
     return true;
 
+  /* THIS argument of method is always non-NULL.  */
+  if (TREE_CODE (TREE_TYPE (current_function_decl)) == METHOD_TYPE
+      && arg == DECL_ARGUMENTS (current_function_decl)
+      && flag_delete_null_pointer_checks)
+    return true;
+
+  /* Values passed by reference are always non-NULL.  */
+  if (TREE_CODE (TREE_TYPE (arg)) == REFERENCE_TYPE
+      && flag_delete_null_pointer_checks)
+    return true;
+
   fntype = TREE_TYPE (current_function_decl);
   for (attrs = TYPE_ATTRIBUTES (fntype); attrs; attrs = TREE_CHAIN (attrs))
     {
@@ -1216,6 +1227,10 @@ gimple_stmt_nonzero_warnv_p (gimple stmt
 	    && DECL_IS_OPERATOR_NEW (fndecl)
 	    && !TREE_NOTHROW (fndecl))
 	  return true;
+	/* Referneces are alwyas non-NULL.  */
+	if (flag_delete_null_pointer_checks
+	    && TREE_CODE (TREE_TYPE (fndecl)) == REFERENCE_TYPE)
+	  return true;
 	if (flag_delete_null_pointer_checks && 
 	    lookup_attribute ("returns_nonnull",
 			      TYPE_ATTRIBUTES (gimple_call_fntype (stmt))))

Re: Teach vrp that THIS pointer and references are non-zero

[Jakub Jelinek]

On Mon, Apr 13, 2015 at 12:12:49PM +0200, Jan Hubicka wrote:
> @@ -1216,6 +1227,10 @@ gimple_stmt_nonzero_warnv_p (gimple stmt
>  	    && DECL_IS_OPERATOR_NEW (fndecl)
>  	    && !TREE_NOTHROW (fndecl))
>  	  return true;
> +	/* Referneces are alwyas non-NULL.  */

Two typos.

I'll defer the patch review to richi/jason; as far as ubsan is concerned,
the flag_delete_null_pointer_checks guards should hopefully make sure we are
ok.

	Jakub

Re: Teach vrp that THIS pointer and references are non-zero

[Richard Biener]

On Mon, 13 Apr 2015, Jan Hubicka wrote:

> Hi,
> with multiple inheritance compiling the testcase bellow, the first call always
> leads to call with offseted pointer, while the other call results in if
> conditoinal testing if parameter is non-NULL.
> 
> This patch teach VRP that THIS pointers and references are non-zero.  I hope
> this is true also for fortran and ada, but I think both languages do not really
> have NULL.
> 
> I think VRP could basically assume all reference pointers to be non-zero, but
> I am not sure how to do that with GIMPLE useless conversion rules.
> 
> Bootstrapped/regtested x86_64-linux, OK?

Hmm - doesn't ESRA remove unused this eventually so this falls apart?

Thanks,
Richard.

> Honza
> 
> 	* g++.dg/tree-ssa/nonzero-3.C: New testcase.
> 	* tree-vrp.c (nonnull_arg_p): THIS pointers and references are non-zero.
> 	(gimple_stmt_nonzero_warnv_p): Reference return values are non-zero.
> Index: testsuite/g++.dg/tree-ssa/nonzero-3.C
> ===================================================================
> --- testsuite/g++.dg/tree-ssa/nonzero-3.C	(revision 0)
> +++ testsuite/g++.dg/tree-ssa/nonzero-3.C	(working copy)
> @@ -0,0 +1,22 @@
> +/* { dg-do compile } */
> +/* { dg-options "-O2 -fdump-tree-vrp1 -fdelete-null-pointer-checks" } */
> +struct A {int a;};
> +struct B {int b;};
> +struct C:A,B {int c;
> +  void bar();};
> +
> +void foo (struct B *);
> +void C::bar ()
> +{
> +  struct C *d = this;
> +  foo(this);
> +  foo(d);
> +}
> +void bar (struct C &c)
> +{
> +  struct C *d = &c;
> +  foo(&c);
> +  foo(d);
> +}
> +/* { dg-final { scan-tree-dump-not "if \\(" "vrp1"} } */
> +/* { dg-final { cleanup-tree-dump "vrp1" } } */
> Index: tree-vrp.c
> ===================================================================
> --- tree-vrp.c	(revision 222016)
> +++ tree-vrp.c	(working copy)
> @@ -393,6 +393,17 @@ nonnull_arg_p (const_tree arg)
>    if (arg == cfun->static_chain_decl)
>      return true;
>  
> +  /* THIS argument of method is always non-NULL.  */
> +  if (TREE_CODE (TREE_TYPE (current_function_decl)) == METHOD_TYPE
> +      && arg == DECL_ARGUMENTS (current_function_decl)
> +      && flag_delete_null_pointer_checks)
> +    return true;
> +
> +  /* Values passed by reference are always non-NULL.  */
> +  if (TREE_CODE (TREE_TYPE (arg)) == REFERENCE_TYPE
> +      && flag_delete_null_pointer_checks)
> +    return true;
> +
>    fntype = TREE_TYPE (current_function_decl);
>    for (attrs = TYPE_ATTRIBUTES (fntype); attrs; attrs = TREE_CHAIN (attrs))
>      {
> @@ -1216,6 +1227,10 @@ gimple_stmt_nonzero_warnv_p (gimple stmt
>  	    && DECL_IS_OPERATOR_NEW (fndecl)
>  	    && !TREE_NOTHROW (fndecl))
>  	  return true;
> +	/* Referneces are alwyas non-NULL.  */
> +	if (flag_delete_null_pointer_checks
> +	    && TREE_CODE (TREE_TYPE (fndecl)) == REFERENCE_TYPE)
> +	  return true;
>  	if (flag_delete_null_pointer_checks && 
>  	    lookup_attribute ("returns_nonnull",
>  			      TYPE_ATTRIBUTES (gimple_call_fntype (stmt))))
> 
> 

-- 
Richard Biener <rguenther@suse.de>
SUSE LINUX GmbH, GF: Felix Imendoerffer, Jane Smithard, Jennifer Guild,
Dilip Upmanyu, Graham Norton HRB 21284 (AG Nuernberg)

Re: Teach vrp that THIS pointer and references are non-zero

[Jan Hubicka]

> On Mon, 13 Apr 2015, Jan Hubicka wrote:
> 
> > Hi,
> > with multiple inheritance compiling the testcase bellow, the first call always
> > leads to call with offseted pointer, while the other call results in if
> > conditoinal testing if parameter is non-NULL.
> > 
> > This patch teach VRP that THIS pointers and references are non-zero.  I hope
> > this is true also for fortran and ada, but I think both languages do not really
> > have NULL.
> > 
> > I think VRP could basically assume all reference pointers to be non-zero, but
> > I am not sure how to do that with GIMPLE useless conversion rules.
> > 
> > Bootstrapped/regtested x86_64-linux, OK?
> 
> Hmm - doesn't ESRA remove unused this eventually so this falls apart?

It changes METHOD_TYPE to FUNCTION_TYPE then.

Honza

Re: Teach vrp that THIS pointer and references are non-zero

[Richard Biener]

On Mon, 13 Apr 2015, Jan Hubicka wrote:

> > On Mon, 13 Apr 2015, Jan Hubicka wrote:
> > 
> > > Hi,
> > > with multiple inheritance compiling the testcase bellow, the first call always
> > > leads to call with offseted pointer, while the other call results in if
> > > conditoinal testing if parameter is non-NULL.
> > > 
> > > This patch teach VRP that THIS pointers and references are non-zero.  I hope
> > > this is true also for fortran and ada, but I think both languages do not really
> > > have NULL.
> > > 
> > > I think VRP could basically assume all reference pointers to be non-zero, but
> > > I am not sure how to do that with GIMPLE useless conversion rules.
> > > 
> > > Bootstrapped/regtested x86_64-linux, OK?
> > 
> > Hmm - doesn't ESRA remove unused this eventually so this falls apart?
> 
> It changes METHOD_TYPE to FUNCTION_TYPE then.

I see.  Patch is ok then with Jakubs fixes.

Thanks,
Richard.

Re: Teach vrp that THIS pointer and references are non-zero

[Jason Merrill]

Yep, looks good.

Jason