The same thing happened to me the first time I tried to benchmark my code on M1.

In my case I was building using Bazel. Bazel was running under Rosetta because they don't release a darwin_arm64 build yet. I didn't realize the resulting code was also built for x86-64.

I tried explicitly passing -march but the compiler rejected this, saying it was an unknown architecture. After some experimentation, it appears that when you exec clang from a Rosetta binary it puts it in a mode where it only knows how to build x86.