diff mbox

[C++] PR 60955

Message ID 54889DF2.7060704@oracle.com
State New
Headers show

Commit Message

Paolo Carlini Dec. 10, 2014, 7:24 p.m. UTC 
Hi,

this regression, a spurious warning about taking the address of a 
register parameter, happens in C++14 mode due to the use of 
force_paren_expr, called by finish_parenthesized_expr, which ends up 
calling build_static_cast. Manuel mentioned in the audit trail that 
TREE_NO_WARNING can be used for DECLs too, and indeed I noticed today 
that we are *already* using it for EXPRs, at the beginning of 
finish_parenthesized_expr. I experimented a bit with restricting the 
setting, eg, to PARM_DECLs only, but in fact we have an identical issue 
for, eg, register VAR_DECLs. Tested x86_64-linux.

Thanks,
Paolo.

////////////////////////////
/cp
2014-12-10  Paolo Carlini  <paolo.carlini@oracle.com>

	PR c++/60955
	* semantics.c (finish_parenthesized_expr): Set TREE_NO_WARNING on
	non-EXPRs too.
	* typeck.c (cxx_mark_addressable): Check TREE_NO_WARNING.

/testsuite
2014-12-10  Paolo Carlini  <paolo.carlini@oracle.com>

	PR c++/60955
	* g++.dg/warn/register-parm-1.C: New.

Comments

Paolo Carlini Dec. 17, 2014, 6:45 p.m. UTC | #1 
Hi,

gently pinging this....

     https://gcc.gnu.org/ml/gcc-patches/2014-12/msg00961.html

On 12/10/2014 08:24 PM, Paolo Carlini wrote:
> Hi,
>
> this regression, a spurious warning about taking the address of a 
> register parameter, happens in C++14 mode due to the use of 
> force_paren_expr, called by finish_parenthesized_expr, which ends up 
> calling build_static_cast. Manuel mentioned in the audit trail that 
> TREE_NO_WARNING can be used for DECLs too, and indeed I noticed today 
> that we are *already* using it for EXPRs, at the beginning of 
> finish_parenthesized_expr. I experimented a bit with restricting the 
> setting, eg, to PARM_DECLs only, but in fact we have an identical 
> issue for, eg, register VAR_DECLs. Tested x86_64-linux.
Thanks,
Paolo.
Jason Merrill Dec. 17, 2014, 8:37 p.m. UTC | #2 
I'm uncomfortable with setting TREE_NO_WARNING on a decl just because we 
don't want a warning for one particular use of it.  How about 
suppressing warnings across the call to build_static_cast?

Jason
diff mbox

Patch

Index: cp/semantics.c
===================================================================
--- cp/semantics.c	(revision 218586)
+++ cp/semantics.c	(working copy)
@@ -1674,10 +1674,13 @@  force_paren_expr (tree expr)
 tree
 finish_parenthesized_expr (tree expr)
 {
-  if (EXPR_P (expr))
-    /* This inhibits warnings in c_common_truthvalue_conversion.  */
-    TREE_NO_WARNING (expr) = 1;
+  if (expr == error_mark_node)
+    return error_mark_node;
 
+  /* Inhibit warnings in c_common_truthvalue_conversion and in
+     cxx_mark_addressable.  */
+  TREE_NO_WARNING (expr) = 1;
+
   if (TREE_CODE (expr) == OFFSET_REF
       || TREE_CODE (expr) == SCOPE_REF)
     /* [expr.unary.op]/3 The qualified id of a pointer-to-member must not be
Index: cp/typeck.c
===================================================================
--- cp/typeck.c	(revision 218586)
+++ cp/typeck.c	(working copy)
@@ -6068,7 +6068,8 @@  cxx_mark_addressable (tree exp)
 		  ("address of explicit register variable %qD requested", x);
 		return false;
 	      }
-	    else if (extra_warnings)
+	    else if (extra_warnings
+		     && !TREE_NO_WARNING (x))
 	      warning
 		(OPT_Wextra, "address requested for %qD, which is declared %<register%>", x);
 	  }
Index: testsuite/g++.dg/warn/register-parm-1.C
===================================================================
--- testsuite/g++.dg/warn/register-parm-1.C	(revision 0)
+++ testsuite/g++.dg/warn/register-parm-1.C	(working copy)
@@ -0,0 +1,9 @@ 
+// PR c++/60955
+// { dg-options "-Wextra" }
+
+unsigned int erroneous_warning(register int a) {
+    if ((a) & 0xff) return 1; else return 0;
+}
+unsigned int no_erroneous_warning(register int a) {
+    if (a & 0xff) return 1; else return 0;
+}