Re: ifcvt too eager at verifing flow info?

[Alexandre Oliva <aoliva at cygnus dot com>]

On May 20, 2000, Richard Henderson <rth@cygnus.com> wrote:

> On Sat, May 20, 2000 at 02:43:23PM -0300, Alexandre Oliva wrote:

>> In the last case, verify_flow_info detects a real bug in ifcvt:
>> it introduces a jump_insn without introducing a basic block note...

> Eh?  Where'd the jump come from?  Is it by chance emitted by the
> back end when implementing setcc or something?

Here's the scenario.  Given this source code:

unsigned long long
f2 (x)
{
  return ((unsigned long long) (x != 0) - 3) / 2;
}

and a cross compiler to mn10300-elf, the ce pass is attempting to
convert the sequence `r = 0x7ff...ff; if (x == 0) r = 0x7ff...fe;' to
avoid the branch using noce_try_store_flag_constants().
emit_noce_store_flag() succeeds, thus noce_try_store... enters the
case:

      /* if (test) x = a; else x = b;
	 =>   x = (-(test != 0) & (b - a)) + a;  */

However, when expanding the add, after the and, ends up deciding to
store the carry using emit_store_flag_force.  But then, because
emit_store_flag fails, it ends up calling do_compare_rtx_and_jump,
that generates a jump for which no BB is introduced in the CFG.

I'm too tired to try to do something about it now.  Maybe tomorrow...

-- 
Alexandre Oliva    Enjoy Guaranį, see http://www.ic.unicamp.br/~oliva/
Cygnus Solutions, a Red Hat company        aoliva@{redhat, cygnus}.com
Free Software Developer and Evangelist    CS PhD student at IC-Unicamp
oliva@{lsd.ic.unicamp.br, gnu.org}   Write to mailing lists, not to me