I think they're asking how the code in the Go runtime (not the compiler, that being an interesting but also maybe non-obvious distinction!) that implements the garbage collector, a core feature of the language, avoids needing the garbage collector to already exist to be able to run, being written in the language that it's a core feature of. I suspect the answer is just something like "by very carefully not using language features that might tempt the compiler to emit something that requires an allocation". I think it's a fair question as it's not really obvious that that's possible--do you just avoid calling make() and new() and forming pointers to local variables that might escape? Do you need to run on a magical goroutine that won't try to grow its stack with gc-allocated segments? Can you still use slices (probably yes, just not append() or the literal syntax), closures (probably only trivial ones without local captures?), maps (probably no)...?

I think the relevant code is https://github.com/golang/go/blob/master/src/runtime/mgc.go and adjacent files. I see some annotations like //go:systemstack, //go:nosplit, //go:nowritebarrier that are probably relevant but I wouldn't know if there's any other specific requirements for that code.