Re: PING: [PATCH] Fix PR 20972

[Bernd Schmidt <bernds_cb1 at t-online dot de>]

Josh Conner wrote:
> On Apr 27, 2005, at 3:40 AM, Bernd Schmidt wrote:
>
> > Josh Conner wrote:
> >
> > > In the register allocator, change post-modified registers to be 
> > > considered 'clobbered', instead of 'set'.
> > >     http://gcc.gnu.org/ml/gcc-patches/2005-04/msg01291.html
> >
> >
> > I doubt that this is the right thing to do.  Could you post more 
> > analysis (initial RTL and RTL before and after register allocation)?
> >
> > I'd expect the same assumption is made by a number of other places in 
> > the compiler.

[...]
> (rN corresponds to virtual register 110 and rM to 104 in the attached 
> dumps)
>
> And the question is: should rN and rM be allowed to share a register if 
> rM is not initialized in a previous instruction (it isn't live coming 
> into this instruction)?

Vlad, this appears to be another side effect of your accurate liveness 
analysis: http://gcc.gnu.org/ml/gcc-patches/2004-05/msg01038.html

As far as I can tell the idea of that patch appears to be that a 
register that appears to be uninitialized on a path can't conflict with 
one that is initialized.  However, a register may not only appear to be 
uninitialized, it can actually be used uninitialized, as in Josh's 
testcase.  By sharing the same register for two different inputs we 
could generate RTL that's invalid.  In this case we have an assembler 
warning, but I imagine there could be situations where the compiler 
itself aborts.

The solution I'd prefer is to make sure that if two registers are used 
as inputs in the same instruction, they must conflict, regardless of 
what the liveness analysis thinks.  (We might want to go back to the 
recent find_dummy_reload issue to see if that would also be fixed by 
such a change).


Bernd