[Bug c++/32182] [4.2 Regression] -fstrict-aliasing optimizations cause constructor not to run for object causing segfault

[rguenth at gcc dot gnu dot org]

------- Comment #12 from rguenth at gcc dot gnu dot org  2007-06-02 11:35 
-------
Confirmed.  Actually compiling Wrapper.cxx with -fstrict-aliasing is enough to
trigger the failure.

In getItem() the difference is

 <bb 2>:
-  D.3769 = &_local_result + 4B;
-  this = (struct BaseClass *) D.3769;
-  this->d_self = 0B;
+  D.3821 = &_local_result + 4B;
+  this = (struct BaseClass *) D.3821;
   this = (struct NextClass *) &_local_result;
   iftmp.0 = (int (*__vtbl_ptr_type) (void) *) _ZTT9TestClass[2];
-  this->_vptr.NextClass = iftmp.0;
   _local_result.D.2186._vptr.NextClass = &_ZTV9TestClass[4];
   this->_vptr.NextClass = &_ZTV9TestClass[4];
   D.2853 = at (cont->d_cont, ind) [return slot optimization];

in addItem()

@@ -547,14 +541,12 @@
   tmp.D.2186._vptr.NextClass = &_ZTV9TestClass[4];
   this.9 = (struct NextClass *) &tmp;
   this.9->_vptr.NextClass = (int (*__vtbl_ptr_type) (void) *)
_ZTT9TestClass[2]
;
-  D.4100 = &tmp + 4B;
-  this = (struct BaseClass *) D.4100;
-  D.4135 = this->d_self;
-  if (D.4135 != 0B) goto <L21>; else goto <L12>;
+  D.4160 = &tmp + 4B;
+  D.4195 = ((struct BaseClass *) D.4160)->d_self;
+  if (D.4195 != 0B) goto <L21>; else goto <L12>;

 <L21>:;
-  deleteRef (D.4135);
-  this->d_self = 0B;
+  deleteRef (D.4195);

 <L12>:;
   <<<exception object>>> = save_eptr.48;
@@ -565,14 +557,12 @@
   tmp.D.2186._vptr.NextClass = &_ZTV9TestClass[4];
   this.9 = (struct NextClass *) &tmp;
   this.9->_vptr.NextClass = (int (*__vtbl_ptr_type) (void) *)
_ZTT9TestClass[2]
;
-  D.4155 = &tmp + 4B;
-  this = (struct BaseClass *) D.4155;
-  D.4190 = this->d_self;
-  if (D.4190 != 0B) goto <L39>; else goto <L4>;
+  D.4215 = &tmp + 4B;
+  D.4250 = ((struct BaseClass *) D.4215)->d_self;
+  if (D.4250 != 0B) goto <L39>; else goto <L4>;

 <L39>:;
-  deleteRef (D.4190);
-  this->d_self = 0B;
+  deleteRef (D.4250);


Note that making the inheritance non-virtual and fixing up
TestClass::TestClass(TestClass::ior_t*) to initialize NextClass instead
of BaseClass fixes the problem as well.

So this may be a C++ frontend problem with virtual inheritance or an
invalid testcase as well.

(What happens if you initialize BaseClass from TestClass, but not NextClass
-- if NextClass is default constructed then it will default construct
BaseClass as well?  C++ language lawyer question.)


-- 

rguenth at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
     Ever Confirmed|0                           |1
           Keywords|                            |alias, wrong-code
   Last reconfirmed|0000-00-00 00:00:00         |2007-06-02 11:35:26
               date|                            |
            Summary|-fstrict-aliasing           |[4.2 Regression] -fstrict-
                   |optimizations cause         |aliasing optimizations cause
                   |constructor not to run for  |constructor not to run for
                   |object causing segfault     |object causing segfault


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=32182