Learning to Restrict Test Range for Compiler Test

Abstract

it is a tremendous challenge to guarantee the correctness of compilers in a limited time, especially when the compiler product is immature. It is necessary to restrict the test range to avoid over-testing, since our compiler products are usually delivered for specific domain. We perform feature extraction on user code through machine learning and use feature information for fuzzy test case generation. The probability of bugs detected has been improved 3.7 times and case size has been reduced 70%.