The Organization of Lookup Tables in Instruction Memoization

Instruction Memoization is a technique that enables performing multiple-cycle computations in a single cycle by exploiting redundant instructions. The technique is based on the notion of memoing: saving the input and output of previous calculations and using the output if the input is encountered again. Several recent papers have proven that non-trivial speedup (over 10%) can be obtained by using this technique. The speedup can be attributed to three major factors: (i) the percentage of instructions that can beneet from memoization, (ii) the integration of the the Lookup Table (LUT) into the datapath of the processor, and (iii) the percentage of successful lookups. This paper focuses on the latter factor and explores how the organization of a Lookup Table (LUT) innuences the percentage of successful lookups (hit-ratio). The best design for a LUT | which we call a memo-table | that yields the highest hit-ratio is described. The design chosen is based on experimental simulations that minimize the innuence of the other two factors of speedup. After the memo-table design is xed the rst two factors can be dealt with, without having to alter the 1 2 memo-table to accommodate changes in (i) the instruction mix of an application and (ii) changes in the microprocessors datapath. The characteristics of the memo-tables explored are size, set as-sociativity, indexing, replacement methods, and contents (instruction mix). Results of simulations have shown that the optimal design is to use 5 memo-tables each which holds a \family" of instructions. Each memo-table contains 256 entries in sets of 4. Entries are replaced randomly and are indexed by using a mix of bits from the exponent and mantissa of FP numbers or the Least Signiicant Bits of integers. Trivial operations are detected using dedicated circuitry and aren't entered into the memo-tables. Integrating the proposed design into the datapaths of the MIPS R10000 and PPC 604e results in an average 8-10% speedup for applications that heavily utilize multiple-cycle instructions.