Well, the issue with dynarec is that it completely destroys any cycle-accuracy. I want to keep the emulator capable of doing that(and Dosbox doesn't have that issue, as it runs instructions in blocks anyway).
Currently the main issue preventing speed is the sheer amount of memory and ROM read/write operations itself. So that means that the BIU and mostly instruction fetching itself (which is reading memory most of the time) is the main bottleneck. The memory reads themselves are already optimized to be only a single call when the address is byte/word/dword aligned(it then uses direct memory access to read/write data to/from the CPU cache or write buffer(for writes only) using a simple pointer dereference). Unaligned dword data is handled as words if possible, bytes otherwise.
But even with all that, the profiler still reports the memory accesses themselves being the hot path.