It should be possible. You just need to keep a note of the cycles elapsed by each compiled opcode. This need only be synchronized every branch or every hardware interaction. The cycle counting will be compiled into the dynamic code, just like the rest of the code.

I'm a little more sceptical than kode54. A given piece of code may take a different number of cycles to run depending on things like the phases of various clocks and counters, the state of the prefetch queue and bus when it starts, and even the values the code is operating on. So just keeping a count of elapsed cycles isn't enough. Not saying it's impossible but it's going to be really difficult.

Well, I only mentioned keeping counts, because most N64 emulators around now use counters even in their recompiler mode, so they can keep accurate timers, since they need timers for interrupts and video frame timing. The only real thing in N64 emulation that's kind of wonky, and is better handled by some emulators, is RSP timing, since many of them just execute a whole RSP task list instantly instead of attempting to cycle time the RSP opcodes.

I do not think it is do-able, in the sense that you will not really end up with both a faster emulator and a cycle-accurate one. At least not in the way other emulators use recompilation.

I took at stab at it too in CAPE. The best I could think of is to have DYNAREC for microcode (NOT 8086 instructions), which is what CAPE breaks down all instruction into. There is microcode for other parts of the CPU not just instructions like EA, BIU etc. All those "microcode" steps can be dynamically recompiled. However I am doubtful that the effort put in (which is considerable) would yield that much faster emulator.

That being said, it would be fun to try.