[Sprint 3] SetIsNoInlinePass supported by adaptive llvm pipeline passes

[whnbaek]

This pull request modifies almost 300 lines, so the rest of the codes including backend optimization will be added in wrap-up phase. I analyze llvm O3 pipelines defined in buildPerModulePipeline and create our adaptive pipelines. These contain inlining process. It is observed that there are some cases where inlining should not occur, i.e. in matmul3 and matmul4. Thus, I add SetIsNoInlinePass which calculates number of colors and decide whether CallInst should be removed by inlining or not. However, there is a problem. After inlining, loop unrolling unrolls loops and increase number of colors of some functions, causing lack of colors. So results of matmul3 and matmul4 looks bad at this point. This would be dealt in wrap-up phase by considering both loop unrolling and pipelines. Moreover, result degradation of rmq1d_naive should be analyzed. Seeming we need an outliner!

Benchmark Results

benchmark	improvement	benchmark	improvement	benchmark	improvement
anagram	25.41%	collatz	63.50%	merge_sort	16.61%
binary_tree	20.08%	floyd	34.52%	prime	33.55%
bitcount1	21.05%	gcd	10.72%	quick_sort	29.28%
bitcount2	18.76%	jenkins_hash	24.19%	radix_sort	23.00%
bitcount3	12.44%	matmul1	9.44%	rmq1d_naive	6.17%
bitcount4	41.78%	matmul2	32.06%	rmq1d_sparsetable	35.53%
bitcount5	15.25%	matmul3	-4.35%	rmq2d_naive	13.48%
bubble_sort	29.58%	matmul4	29.47%	rmq2d_sparsetable	40.96%