remaining libjava/verify_local_live_at_start failures

[Jan Hubicka]

> On Wed, Mar 27, 2002 at 02:52:02PM +0100, Jan Hubicka wrote:
> > > If *anything* within a single basic block gets different answers
> > > when run multiple times, something is wrong.
> > 
> > Erm, what exactly you do mean?
> 
> Exactly that.  When SCAN_DEAD is set, we pretend that the dead
> code has in fact been removed, and compute life accordingly.

I finally see. We do not precalculate local_live bitmap, instead we re-scan
allways the whole block and take into account the presence of dead instructions.
I will return to that testcase again this weekend.

Thanks for enlightening me.
Honza
> 
> > BTW why the dead store removal is integrated in register liveness analysis?
> 
> Why not?  The one affects the other.
> 
> 
> 
> r~