I just got a random idea - have anyone tried to make an x86 emulator, like Bochs, that runs on bare hardware and passes all i/o to the host and back while running a guest OS?

The code itself will occupy a block at the end of physical RAM, hiding it from the emulated os, that runs in actual memory. You can connect a debugger to it thru a serial port, or something, and stop or break the execution at any point, or log the io ports activity. That can be used for reverse engineering vendor hardware drivers, or just for OS debugging.

What kind of problems could there be?
I tried the concept successfully with VESA bios calls from custom PM OS, at least for that the speed does not matter. What about other hardware? Would it stand slow I/O?
Any architectural show-stoppers?

And, most importantly, was it done before, and does anyone need something like that?

There are lots of emulators and VMs available. If you want to start a project in asm then that would be good. But if you are looking for a ready solution then why does Bochs not do what you want? What do you need extra?

Because, Bochs runs in an OS.
I'm talking about an emulator that runs on nothing but bare hardware, and emulates nothing but the CPU.
I'm thinking of starting a project, and checking for reinvented bicycle scenario.

Artlav, you probably want to check Nouveau, they have some tools to do hardware tracing.

If you need to eliminate the OS the perhaps the way to go is to use a hypervisor. That way you can partition your memory so that the OS can't see or change your code and data.