nothrow changes

nothrow changes

[Jeffrey A Law]

All the nothrow work (as a whole) has broken nonlocal gotos.

Basically we are setting TREE_NOTHROW to a nonzero value on the function decl
for the nested function that has a computed goto.

When we later expand a call to that function we see TREE_NOTHROW is set
and set the EH region on the generated CALL_INSN to -1, so flow thinks
the call can not do a nonlocal goto or throw.

This results in flow computing an incorrect cfg and deleting code that
it believes is unreachable, but is actually reachable via the nonlocal
goto in the a nested function.

execute/920428-2.c shows this problem.  I've tested it on ia32, ppc & the PA
all with the same basic results.

The mis-compiled function in question:
s(i)
{
  if(i>0)
    {
      __label__ l1;

      int f(int i)
        {
          if(i==2)
            goto l1;
          return 0;
        }

     return f(i);
    l1:;
    }
  return 1;
}

We set TREE_NOTHROW on the function decl for "f".  So when "s" calls "f"
we do not think that the call could do a non-local goto to nonlocal goto
receiver for label "l1" within "s".

So we delete the nonlocal goto receiver, which ultimately results in either
a core dump, an incorrect return value or other nasties.

I'm not entirely sure how you expected this stuff to work for nonlocal
gotos, so I haven't tried to fix it myself.  I would appreciate it if you
would look into this problem.

Thanks,
jeff

Patch Re: nothrow changes

[Jason Merrill]

>>>>> Jeffrey A Law <law@cygnus.com> writes:

 > All the nothrow work (as a whole) has broken nonlocal gotos.

 > When we later expand a call to that function we see TREE_NOTHROW is set
 > and set the EH region on the generated CALL_INSN to -1, so flow thinks
 > the call can not do a nonlocal goto or throw.

Oops.

2000-03-15  Jason Merrill  <jason@casey.cygnus.com>

	* calls.c (emit_call_1): Nothrow functions can still have nonlocal
	gotos.

Index: calls.c
===================================================================
RCS file: /cvs/gcc/egcs/gcc/calls.c,v
retrieving revision 1.91
diff -c -p -r1.91 calls.c
*** calls.c	2000/03/09 20:34:51	1.91
--- calls.c	2000/03/16 07:31:51
*************** emit_call_1 (funexp, fndecl, funtype, st
*** 495,501 ****
    /* If this call can't throw, attach a REG_EH_REGION reg note to that
       effect.  */
    if (nothrow)
!     REG_NOTES (call_insn) = gen_rtx_EXPR_LIST (REG_EH_REGION, GEN_INT (-1),
  					       REG_NOTES (call_insn));
  
    /* Restore this now, so that we do defer pops for this call's args
--- 495,501 ----
    /* If this call can't throw, attach a REG_EH_REGION reg note to that
       effect.  */
    if (nothrow)
!     REG_NOTES (call_insn) = gen_rtx_EXPR_LIST (REG_EH_REGION, const0_rtx,
  					       REG_NOTES (call_insn));
  
    /* Restore this now, so that we do defer pops for this call's args

Re: Patch Re: nothrow changes

[Jeffrey A Law]

  In message < u91z5bjrk5.fsf@yorick.cygnus.com >you write:
  > Oops.
  > 
  > 2000-03-15  Jason Merrill  <jason@casey.cygnus.com>
  > 
  > 	* calls.c (emit_call_1): Nothrow functions can still have nonlocal
  > 	gotos.
Thanks.  I was looking completely elsewhere and would have mucked it up
badly had I tried to fix this.  Thanks again!

jeff