Nowadays, I’m researching about a compiler problem - “Phase Ordering Problem”.
I would like to solve this problem ( or make better than O3) by using static analyze the llvm IR.
My method is: Get the passes of “-O3" and reorder them by random selection. Get many of the random passes and generate the llvm IR by these passes from a single source IR.
The passes of O3: A, B, C
get the pass,
random 1 passes: A, C, B
random 2 passes: B, C, A
random 3 passes: C, A, B
apply these pass on a identical source IR and get ‘four’ files: O3.ll, random1.ll, random2.ll, random3.ll
How could I compare the performance of four of them by analyze the IR code ?
Now I’m using the fit function like this formula  and calculate all of the IR I generate. And then get the max fit value of them.
The one which has the max fit value is better than “-O3” but only faster about 10% or less. It is not enough for me. And the variables in the fit function I created are just selected by my experiments and I couldn’t find the paper or relative information about this issue.
I hope someone can help me to figure out or just give me a keyword to let me search and study.
 (the lines of vectorbody blocks - store instconut - load instcount) / (instruction cost + instruction count)