Bull Hn Information Systems, Inc.

Jeffrey D Weintraub

Russell Wayne Guenthner

Ronald E Lange

Clinton B Eckard

Bruce E Flocken

Richard L Demers

Wilbur L Stewart

Sleiman Chamoun

In a central processor incorporating at least one co-processor, such as a floating point arithmetic co-processor, in addition to a basic arithmetic logic unit, the problem of rationalizing the contents of the accumulator and supplementary accumulator registers without the burden of speed penalties is addressed and solved. This is achieved by providing input/output access to a common register file and by switching control of the register file to the proper processing unit appropriately. A single, shared accumulator register and a single, shared supplementary accumulator register are included in the stack along with other sharable registers such as address modification registers. Thus, the contents of the accumulator register and the supplementary accumulator register are always up-to-date and available to all processing units in the central processor without the need for first carrying out rationalization steps.

Sharing of register stack by two execution units in a central processor

In order to efficiently recover from a processing error in a central processing trait (CPU) incorporating a cache memory and a basic processing unit, the BPU is provided in duplicate, and all BPU data manipulation operations are performed redundantly. After duplicate data has been obtained from the cache memory and manipulated by the duplicate BPUs, the outputs from the duplicate BPUs are placed on respective master (MRB) and slave (SRB) result busses which are coupled to the cache unit where the results are compared for identity. If the results are not identical, a local error signal is issued. In response to the error signal, the corrupted data is stored into cache, but, before the cache is deliberately frozen, the data is restored again using a segment of the original data withdrawn from the cache memory by the BPUs such that, when the cache is frozen in anticipation of remedial action, the data block whose modification took place during the faulting operation will have been restored to its preprocessing condition. As a result, restart, if possible, can commence at the same point in the process rather than at an earlier point.

Store 'undo' for cache store error recovery

Growing community of inventors

Scottsdale, AZ, United States of America

Jeffrey D Weintraub