A Result Forwarding Mechanism for Asynchronous Pipelined Systems
D.A. Gilbert, J.D. Garside
Abstract
Modern, fast microprocessors are deeply pipelined to enhance their
performance. Thus they cannot afford to wait for each instruction to
complete before starting the next. When inter-instruction dependencies are
encountered it is essential that data are forwarded from their point of
production to where they are needed as rapidly as possible. This has been a
problem in asynchronous processors because of the lack of synchronisation
between the units producing and consuming the data. This paper presents a
solution to this problem.
The mechanism described allows the depth of speculative execution to be
increased, improving memory efficiency by hiding the load latency yet still
allowing exact data aborts.
Full
paper (43K compressed).