Re: Value regression

In article <[EMAIL PROTECTED]>, 
[EMAIL PROTECTED] (Iain McClatchie) wrote:

>What the heck are you two talking about?

"restoring values on an exception ... by recomputation" 
rather than by maintaining register values and restoring 
by a simple remapping.

The main benefit would seem to be moving state away 
from the timing critical issue loop (ReOrder Buffers, 
register file, et al.) toward the less timing critical 
commit zone (though there might also be a reduction 
in the quantity of state required).

I receive the impression than Andy Glew is thinking 
about retaining execution traces that could be 
rerun from a smaller set of exception restoration 
points.  I was only thinking about applying a functional 
inverse for simple operations.  If a register can be 
'speculatively' be freed early, the register file/ROB can 
be smaller (and the cost of recomputation could be 
modest because the pipeline is flushed after a 
branch misprediction).

A little clearer than mud now?  (I hope I have not 
drastically misrepresented Mr. Glew's intention.)

Paul A. Clayton
not entirely useless fellow
for mail: de-133t 'Dy5thymicd01t' to generate a 
synonym for 'Depressivedunce' (at aol.com)