Techniques for instruction count based compiler performance regression testing are described herein. An aspect includes identifying a subset of a plurality of test cases that are regression suspects for a current build of a compiler, wherein the identifying of a test case of the plurality of test cases as a regression suspect is performed based on an instruction count associated with compiling or running of the test case using the current build of the compiler and a previous build of the compiler. Another aspect includes processing the subset of the plurality of test cases to identify any test cases associated with confirmed regressions.