Hacker News Reader

Comment by IAmLiterallyAB | Hacker News Reader

Surely emulating ARM on ARM would be faster / easier to JIT? Or at least it seems that way

Actually, no, mips is an easier JIT target than arm as well. Source: I've written both ARM-to-thumb1 and MIPS-to-thumb1 JITs

Fascinating!

The main hang-up is that ARM uses PC as a general register a lot and in ways that make translating ARM instrs rather messy.

    ADD R0, SP, PC, ROR SP

is entirely valid, even if nonsensical, instruction. But you must translate all valid inputs, else you risk breaking things. That may be a contrived example, but here is a common one: if one has a jumptable of relative offsets somewhere, pointed to by R10, even this is valid:

   LDR R0, [R10, R0, LSL #2]
   ADD PC, R10, R0

That gets messy to translate

Interesting, I hadn't thought about this. Is the issue that the JIT output is likely to be a different number of bytes away? tbb/tbh seems like a more common version of that problem, TBH.

As I understand it, this kind of thing was a big problem for ARM in the mid-90s when they finally wrote the ARM ARM and outlawed things like ldmia r2!, {r0-r4}.

Different number of bytes out than in is not an issue. Efficiently translating such constructs in the general case is hard. Imagine what it would look like.

I have some ideas, but I guess I should try writing a JITting emulator instead of asking you to debug them :)

Try it - it is a fun project.

Thanks for the encouragement! I admit to being a bit intimidated by it.

Reply on news.ycombinator.com