A bug was found in freebios2/src/northbridge/via/vt8601/raminit.c

[Eric W. Biederman]

ron minnich <rminnich at lanl.gov> writes:

> On 6 Jul 2004, Eric W. Biederman wrote:
> 
> > Hmm..  Using bytes does not help the register pressure and in fact
> > slightly increases code size.  If the byte register instructions
> > on x86 were more symmetric this might be different, but...
> 
> sad that they don't work well. 

They work well unless you are register starved, and
since romcc is always register starved they don't work well
in that context.  There are no general register to register
move instructions to let me move a byte register into a longer
register and vice versa.  So generally I don't even bother with the
byte registers.

> > I cannot think of an architecture where using bytes would be better.
> > The increase in code size is because of the need to clamp the values
> > at their maximum so code like this fails properly.  
> 
> They exist :-)

Where bytes are better for register pressure? 

Beyond the theoretical case where something could be built I would
love to see one.  Note this is limited to treating any instruction
set as a load/store architecture, because register to memory
instructions can't be used.  Which is the primary thing that
cripples the x86 case and bytes.
 
> thanks for the tip, this is very good to know.

There are 2 important cases for registers.
1) Using them.  
2) Storing data in them.

When you are using a register it is best to have the value unpacked
and to just use that register.

When you are simply storing data in a register the trade offs change
because you can afford a second register to hold the value when
you take it out and modify it.

For optimizing the storage case I have implemented bitfields. When
they are unpacked and being manipulated they need an extra.  In
the cases where registers are tight bitfields could be handy,
in structures could be handy for passing around multiple values.
A loop counter usually does not fit that description though.
Especially not when you don't have something else to put in the same
register.  

Since values are mostly used and not stored in registers I leave
it up to the programmer to implement the storage optimization case.
Although by implementing bitfields I am as helpful as I can be.

Eric