Design of Instruction Analyzer with Semantic-Based Loop Unrolling Mechanism in the Hyperscalar Architecture

Abstract

Nowadays ILP processors can’t analyze the semantic information of instruction thread to change instruction series automatically for increasing ILP degree. High performance required programs such as image processing or machine learning contain a lot of loop structure. Loop structure will be bounded with the instruction number of one basic block. That cause processors are hard to enhance the computing efficiency. The characteristics of the loop structure in the program are as follows: (1) Instruction will be fetched from cache and be decoded repeatedly. (2) The issued instructions are bounded by the loop body. (3) There is data dependence between iterations. These factors will get worse the poor ILP in the loop codes. In this paper, we propose an architecture called semantic-based dynamic loop unrolling mechanism. The proposed architecture can buffer the instruction series of nested loop, unroll it automatically by analyzing the instruction flow to find the loop body with the semantic of loop instructions, store them to the instruction buffer, and dispatch them to target the processor cores. The proposed architecture consists of three units: loop detect unit (LDU), unrolling control unit (UCU) and loop unrolling unit (LUU). LDU will parse the semantic of instructions to find the closed interval of the loop body instructions. UCU will control LUU in the whole process. LUU will unroll the loop based on the information collected by LDU. Loop controller will handle the complementation overhead for branch miss prediction and the loop finish-up codes. The verifications use ARM instructions generated by Keil $\mu$Vision5 compiler. The results show that eliminating iteration dependence can improve ILP by 140% to 180%.