Re: [Bug c++/14639] New: Incorrect emission of unused compiler-generateddestructor at -O2

[Mark Mitchell <mark at codesourcery dot com>]

Jan Hubicka wrote:

> > Consider the following file:
> > class AAA {
> > public:
> >  virtual ~AAA();
> > };
> >
> > class BBB {
> > public:
> >  virtual ~BBB();
> > };
> >
> > class xyz : public AAA, public BBB {
> > public:
> >  xyz() : AAA(), BBB() { }
> > };
> >
> >
> >
> >
> > Class xyz has no key method, so its vtable should be emitted only in translation units where it's used.  
> > With "gcc -O0 -c foo.cc" I get what I expect: the compiler doesn't emit _ZTV3xyz, _ZN3xyzD0Ev, or 
> > anything else.  With "gcc -O2 -c foo.cc" I get incorrect behavior: the compiler emits xyz's destructor 
> > and vtable, thunks, and all of the things in AAA and BBB that are required to support them.
> >
> > This is a regression.  In 3.2.2, O0 and O2 both behave correctly.
> >    
>
> This can be avoided by disabling following mark_referenced call:
>
> Index: cp/method.c
> ===================================================================
> RCS file: /cvs/gcc/gcc/gcc/cp/method.c,v
> retrieving revision 1.279
> diff -c -3 -p -r1.279 method.c
> *** cp/method.c	12 Mar 2004 17:09:01 -0000	1.279
> --- cp/method.c	20 Mar 2004 15:27:33 -0000
> *************** use_thunk (tree thunk_fndecl, bool emit_
> *** 359,365 ****
>       this translation unit.  */
>    TREE_ADDRESSABLE (function) = 1;
>    mark_used (function);
> !   mark_referenced (DECL_ASSEMBLER_NAME (function));
>    if (!emit_p)
>      return;
>  
> --- 359,365 ----
>       this translation unit.  */
>    TREE_ADDRESSABLE (function) = 1;
>    mark_used (function);
> !   /*mark_referenced (DECL_ASSEMBLER_NAME (function));*/
>    if (!emit_p)
>      return;
>  
> The problem is that C++ frotnend manages thunks "on the side" and does
> not show to backend what DECLs will be used by thunks and what thunks
> will be emit.  This is dealt with handling the depenedencies
> conservativly.   This line comes from replacing DECL_REFERENCED.* = 1 by
> mark_referenced call so I am not quite sure it is still necesary
> (testsuite passes without this line).
>
> However we likely need some way to explain backend the hidden
> dependencies.  Would it be possible to explain me why exactly the thunk
> needs to be output and what DECLs will it reference?
> (I know that the thunks are associated to functions and are output at
> the same time the function is output and they become needed by
> use_thunk, but this all can be overly conservative now)
>  
Every thunk references exactly one function.  The thunk is just a little 
stub that adjusts one incoming parameter, and then transfers control to 
the referenced function.

From a FUNCTION_DECL, you can look at DECL_THUNKS to find all of its 
thunks.  From a thunk, you can look at THUNK_TARGET to find the function 
to which it transfers control.  It does make sense to me that we call 
mark_referenced in the case that emit_p is zero.  If emit_p is 1, that 
does make sense.  In the emit_p == 0 case, we should wait to emit the 
thunk until we know that we need to emit the vtable referencing it, 
which is what we used to do.

--
Mark Mitchell
CodeSourcery, LLC
(916) 791-8304
mark@codesourcery.com