CompilersFinal

Approach

We identified 16 loop optimizations passes, which give rise to 16! permutations. As such, it is infeasible to test all possible orderings. Instead, we use an iterative approach. We first test all permuations of length 4 of the 16 passes, which yields 43,680 unique passes. Then we take the top five results and generate all permuations of length 4 with the remaning 12 passes, which yields $5 \cdot 11,880 = 59,400$ unique passe. This process is continued

# Results from Previous Iteration	# Remaining Passes	# Unique passes
0	16	43,680
5	12	59,40
25	8	42,000
1,000	4	24,000

After this, we take the top result out of the $1 + 1 + 1 + 24,000 = 24,003$ choices.

Reproducing Results

We use the cortado machines via SLURM for all tests.

Generating Optimization Pass Permutations

To generate the permutations for $n$ remaining passes, run

python3 create_optimizations_list <path to file containing best optimizations from previous iteration>

For the case of 16 remaning passes, there is no previous iteration, so pass an empty file. Results are saved as optimization_permutations_{n}.txt.

Running Permuations via SLURM

The runtests.sh <start> <end> script will run the benchmarks on the permutations start through end (inclusive). Permutation $i$ is the $i$ th line of optimization_permutations_{n}.txt (1-indexed). Note that run_test.sh must be manually modified to use the correct optimization_permutations_{n}.txt file.

To run using SLURM, use submit_jobs.sh <start> <end>. This will batch submit SLURM jobs to run permutations through inclusively where each job runs a fixed number of permuations. Again, this file must be manually modified to set the correct bounds.

Running Individually

To run the GCC loops, TSVC, or LCALS benchmark, use the following as a guideline:

# GCC Loops
cd benchmarks/gcc_loops   # GCC loops
cd benchmarks/TSVC_2/src  # TSVC
cd benchmarks/lcals       # LCALS

make OPTFLAGS="{passes to run}"
make size
make run

To run the NPB benchmark, use the following as a guideline

cd benchmarks/NPB-SER

make suite OPTFLAGS="{passes to run}"
make size
make run

Evaluating Permuations

All results are saved in the temp_results folder. Run python3 aggregate_results.py to aggregate them into a single results.csv file. Then manually modify score_results.py to take the top $n$ results along with the proper file pathes for the results file and the optimization_permutations_{n}.txt file. Run python3 score_results.py and the top permutations will be saved into the filename set via best_file.