Browse Prior Art Database

Overlapping Instruction Unit Delay With Instruction Prefetching

IP.com Disclosure Number: IPCOM000063616D
Original Publication Date: 1985-Mar-01
Included in the Prior Art Database: 2005-Feb-18

Publishing Venue

IBM

Related People

Authors:
Pomerene, JH Puzak, TR Rechtschaffen, RN Rao, GS [+details]

Abstract

In many high performance pipelined machines, the occurrence of certain conditions causes the Instruction (I) Unit to stop until the condition is cleared, e.g., Address Generation Interlocks, Queue Full condition, All Operand Address Registers busy condition, etc. In some of these situations the opportunity exists for prefetching the target stream of a branch instruction: L 4, 8(6) L 5, 16(4) BC F, 80(6). If the target stream is prefetched, then two potential benefits accrue when the I Unit eventually continues. 1) The BC and the target instruction can be decoded on successive cycles, thereby eliminating the taken branch delay. 2) The additional opportunity for instruction fetching can reduce the delay due to Empty I Buffer.