The test results are completely bogus. Most of tests are compiled w/o any optimizations at all. What's the point of comparing the unoptimized binaries?
The only "good" test is N-Queens, which is compiled with -O3. Everything else is compiled effectively with -O0.
C-Ray is bogus as well, because clang does not support -march=native and thus yields slow armv4 code in this case.