[Bug ipa/104303] [12 regression] gnatmake is miscompiled by IPA/modref

[hubicka at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104303

--- Comment #5 from Jan Hubicka <hubicka at gcc dot gnu.org> ---
So what modref see on calle is:
void concat5_pkg1.make_failed (struct  s)                                       
{                                                                               
  struct string___XUB * s$P_BOUNDS;                                             

  <bb 2> [local count: 1073741824]:                                             
  concat5_pkg2.compare (s);                                                     
  return;                                                                       

}                                                                               
and it believes that it is safe to only consider global memory accesses since
the parameter escapes because we have no EAF flags attached to
concat5_pkg1.make_failes nor to concat5_pkg1.compare.

Now alias sees:
ESCAPED, points-to non-local, points-to NULL, points-to vars: { D.5011 D.5040 }
(escaped)
....
  S9b.6_22 = .builtin_alloca_with_align (_5, 8);                                
  (*S9b.6_22)[1 ...]{lb: 1 sz: 1} = "option ";                                  
  _6 = MAX_EXPR <L7b_19, 7>;                                                    
  _7 = (sizetype) _6;                                                           
  _8 = _7 + 18446744073709551609;                                               
  _9 = s.P_ARRAY;                                                               
  _10 = &(*S9b.6_22)[8 ...]{lb: 1 sz: 1};                                       
  .builtin_memcpy (_10, _9, _8);                                                
  _11 = iftmp.5_14 + 8;                                                         
  _12 = (sizetype) _11;                                                         
  VIEW_CONVERT_EXPR<character[1:23]>((*S9b.6_22)[_12 ...]{lb: 1 sz: 1}) = "
should start with \'--\'";
  D.5010.P_ARRAY = S9b.6_22;                                                    
  D.5011.LB0 = 1;                                                               
  D.5011.UB0 = L8b_20;                                                          
  D.5010.P_BOUNDS = &D.5011;                                                    
  concat5_pkg1.make_failed (D.5010);                                            

D.5040 is points-to var of the alloca call. I will have to single-step tomorrow
to figure out why the oracle thinks that it can not be aliased by global memory
when it is seen as escaping...