[Bug middle-end/51644] [4.7 Regression] va_list vs. warning: ‘noreturn’ function does return is not fixable

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51644

--- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> 2011-12-21 09:51:45 UTC ---
Caused by r181172 aka TREE_CLOBBER_P.  The problem is that we have before
*.cfg:
  __builtin_va_start (&args, 0);
  b (s, &args);
  __builtin_va_end (&args);
  finally_tmp.0 = 1;
  <D.2917>:
  args = {CLOBBER};
  switch (finally_tmp.0) <default: <D.2921>, default: <D.2921>, case 1:
<D.2918>>
  <D.2918>:
  goto <D.2919>;
  <D.2919>:
  return;
  <D.2920>:
  finally_tmp.0 = 0;
  goto <D.2917>;
  <D.2921>:
  resx 1
and *.cfg pass obviously optimizes away anything after the noreturn call:
<bb 2>:
  __builtin_va_start (&args, 0);
  b (s, &args);
<L5>:
  return;
<L3>:
  finally_tmp.0 = 0;
  args = {CLOBBER};
  switch (finally_tmp.0) <default: <L4>, case 1: <L5>>
<L4>:
  resx 1
but the warn_function_return pass is immediately after that (and, this is -O0
anyway), so nothing optimizes the switch away into simple fall through into the
resx 1 block.  At -O and above we don't warn, because the *.eh pass before
*.cfg pass optimizes this.

So, IMHO to fix this, we need to do 2 things:
1) in decide_copy_try_finally reconsider
     if (!optimize)
    return ndests == 1;
   if finally contains only gimple_clobber_p stmts
2) estimate_num_insns should return 0 for gimple_clobber_p stmts