We should performance test both compilation time and generated binary runtime with -use-spirv-backend and -no-use-spirv-backend to verify there are no critical regressions with the new backend.
We could use the Comgr tests to measure compilation times for a few examples (via AMD_COMGR_TIME_STATISTICS). Or use command line invocations of clang with the flag on/off.
And we can try to use the LLVM test suite / hip-tests to measure runtime performance.