Optimizing Instruction Scheduling and Register Allocation for Register-File-Connected Clustered VLIW Architectures

Algorithm 1

Predecision phase.

Input:

 Data Dependence Graph, 𝐺={𝑁,𝐸};

 Ready list, 𝑅=𝑁;

Output:

 Preliminary cluster assignment decision for each instruction;

(1)   while (𝑅≠𝜙) do

(2)    Perform ASAP scheduling and ALAP scheduling

(3)    Select Instruction i from 𝑅

(4)    Find the neighborhood of Instruction i,

   𝐵(𝑖)=𝑗|𝑗≠𝑖,𝑗∈𝑁,andedge<𝑖,𝑗>𝑜𝑟<𝑗,𝑖>∈𝐸

(5)    Calculate GF Array for Instruction i

   𝛿=1𝑗,𝑥1,if𝑗isscheduledincluster𝑥𝑁_𝑐𝑗,if𝑗isnotscheduled0,otherwise

   1𝜆(𝑗,𝑥,𝑦)=𝑛,if𝑗hasnactiveinter-clusterdatacommunicationat𝑆𝑐ℎ𝑒𝑑𝑢𝑙𝑒-𝑃𝑜𝑖𝑛𝑡(𝑥,𝑦)𝑁_𝑐(𝑗),if𝑗isnotscheduled0,otherwise

   𝑇𝑡(𝑗)=𝑘,if𝑗isscheduledinclockcycle𝑘_𝑒(𝑗),if𝑗isnotscheduled

   𝑊(𝑖,𝑗,𝑥,𝑦)=𝑡(𝑗)−𝑦

   GF(𝑖,𝑥,𝑦)=∑_𝑗𝛿(𝑗,𝑥)−𝜆(𝑗,𝑥,𝑦)𝑊(𝑖,𝑗,𝑥,𝑦),

    𝑗∈𝐵(𝑖),𝑇_𝑒(𝑖)≤𝑦≤𝑇_𝑙(𝑖)

(6)    Calculate RF Array for Instruction i

   𝑀(𝑗)=𝑇_𝑙𝑗−𝑇_𝑒𝑗+1

   1𝛾(𝑗,𝑥,𝑦)=1,if𝑗isscheduledincluster𝑥cycley𝑀𝑗𝑁_𝑐(𝑗),if𝑗isnotscheduled0,otherwise

   RF(𝑖,𝑥,𝑦)=∑_𝑗𝛾(𝑗,𝑥,𝑦)+𝛽(𝑥,𝑦),

    𝑗≠𝑖,𝑗∈𝑁,𝑇_𝑒(𝑗)≤𝑦≤𝑇_𝑙(𝑗),𝑇_𝑒(𝑖)≤𝑦≤𝑇_𝑙(𝑖)

(7)    Normalize GF Array and RF Array

(8)    Calculate the BF Array for Instruction i

   BF(𝑖,𝑥,𝑦)=GF(𝑖,𝑥,𝑦)−RF(𝑖,𝑥,𝑦)

(9)    Pre-schedule Instruction i to Schedule-Point (𝑝,𝑞) with the maximize value of BF Array

(10)    Delete Instruction i from R.

(11) end while