[Bug debug/51650] [4.7 regression] LTO ICE in dwarf2out_finish, at dwarf2out.c:22501 while building libxul

["rguenth at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51650

Richard Guenther <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |hubicka at gcc dot gnu.org

--- Comment #15 from Richard Guenther <rguenth at gcc dot gnu.org> 2011-12-23 10:41:02 UTC ---
(In reply to comment #13)
> Reduced (it almost identical to the first testcase):
> 
> struct T;
> class C
> {
> public:
>     typedef ::T T;
>     virtual void E();
>     static T *m ()
>       {
>         static T *d;
>         return d;
>       }
> };
> int
> fn ()
> {
>   C::m ();
> }
> int main() {}

>From the C++ frontend the sequence is that we call debug_hooks->function_decl
for C::m from rest_of_handle_final which creates the type DIE for T.  With
LTO everything but main is optimized out, but we still emit debug info for
C::m::d via emit_debug_global_declarations () because we think C::m::d is
still live (for some reason ...).

That is, without LTO but with -fwhole-program we do not output
_ZZN1C1mEvE1d but with -flto we do (but we do not output C::m in either
case).  That seems to be a missed optimization (at least), probably
caused by applying the whole-program assumption late at WPA time:


Marking local functions: fn m


Marking externally visible functions: main


Marking externally visible variables:


Needed variables: d

why is d needed?

Reclaiming functions: fn m


cgraph pre whole_program_function_and_variable_visibility time:

main/2 @0x7ffff5a29900 (asm: main) availability:available analyzed reachable
body externally_visible prevailing_def finalized
  called by: 
  calls: 
  References: 
  Refering this function: 
fn/1 @0x7ffff5a297e0 (asm: _Z2fnv) availability:available analyzed reachable
body externally_visible prevailing_def_ironly finalized
  called by: 
  calls: m/0 (1.00 per call) 
  References: 
  Refering this function: 
m/0 @0x7ffff5a296c0 (asm: _ZN1C1mEv) availability:available analyzed reachable
body externally_visible finalized
  called by: fn/1 (1.00 per call) 
  calls: 
  References:  var:d (read)
  Refering this function: 

it seems that m/0 lacks a resolution, the resolution file has

1
t.o 3
195 67dffb6b12d80baf PREVAILING_DEF_IRONLY _Z2fnv
200 67dffb6b12d80baf PREVAILING_DEF main
202 67dffb6b12d80baf PREVAILING_DEF_IRONLY _ZZN1C1mEvE1d

hm.  And the function decl for m looks like

 <function_decl 0x7ffff5b60c00 m
    type <function_type 0x7ffff5b5f7e0
...
    addressable used nothrow public static weak autoinline QI defer-output
...

and it is DECL_COMDAT, so we do not output it to the LTO IL symtab and
thus do not get a resolution for it from the first loop.  In the
2nd loop we do not output it because the node isn't DECL_EXTERNAL either.

Why do we never output comdat unsharable functions and thus do not get
a resolution (which would be valid even if we end up unsharing them)?