I have seen that, I even replied to his post. Whether the gcc.c compile benchmark qualifies as a real-world comparison is debatable (I tend to say no). But a problem is that it is not a real comparison because the outputs of the three compile runs are different from each other. One gives you 32 bit x86 code, the second one 64 bit x86_64 and the third one 32 bit x32 code. Different compiler optimization techniques might be used for the different architectures and the linker will also have to perform different tasks.

I tried to do some video encode related benchmarks but ran into the problem that ffmpeg must disable assembly optimizations for x32, so the comparison would only become fair if I artificially hampered x86_64 performance by disabling x86_64 asm (which no real world user would do).