Instruction Buffers

Abstract

In dealing with operand accessing in earlier chapters we considered various techniques used to overcome the disparity between processing speed and main store accessing rate. This problem also impinges on instruction accessing, since for efficient operation instructions must also be supplied to the processor at a rate matching its execution rate. In the case of instruction accessing, however, the problem is ameliorated by the fact that most instructions are obeyed sequentially and the main store word size is normally such that one word fetched from main store can contain several instructions. Furthermore, with an interleaved store, successive accesses for sequential instructions reference each stack in turn and are not held up by cycle time effects. Thus store requests can be made in advance of the corresponding instruction being required and the replies buffered until they are needed for execution. This pre-fetching technique is used in almost all high performance pipelined processors. A significant proportion of instructions cause control transfers, however, and each such transfer requires a request to be made to the store for a new sequence of instructions. Thus although the accessing rate for instructions can normally be matched satisfactorily to the processing rate, the access time for the first instruction of a new sequence can result in a long delay to the processor. Techniques for overcoming this problem rely on the fact that the cause of many control transfers is a branch back from the end to the start of a loop of instructions, and loop-catching buffers are incorporated into a number of processors. In this chapter we shall consider the instruction accessing and buffering techniques used in the IBM System/360 Model 195, the CDC 6600 and 7600, MU5 and the CRAY-1.