4 ms·

And if you want a better performance, you can replace this switch with computed gotos (supported at least by gcc and clang).

by sklogic 11y ago

simias 11y ago
You should first check if the compiler doesn't do it for you. Otherwise you're just making your code harder to maintain and less portable for zero gain. In my experience with similar code you can get GCC and LLVM to generate a proper jump table if you're careful to make your code as regular as possible in order for the optimizer to understand what you're doing. Case in point: https://github.com/simias/psx-rs/blob/master/src/cpu/mod.rs#L325-L400 https://github.com/simias/psx-rs/blob/master/src/cpu/mod.rs#... Here LLVM manages to inline all the functions and generate a proper jump table. IIRC I tried getting the function address in the "match" and calling it afterwards but then the compiler wouldn't optimize it properly anymore.
- sklogic 11y ago
  > You should first check if the compiler doesn't do it for you. No, compiler cannot rewrite your bytecode into a threaded code. > Otherwise you're just making your code harder to maintain and less portable for zero gain. No. Threaded code is almost always faster, even if run only once. > a proper jump table I'm not talking about jump tables. I'm talking about replacing the bytecode indices with jump addresses once (yes, with a precalculated table) and then doing an indirect jump for each translated opcode from the input stream. An example: https://github.com/ocaml/ocaml/blob/trunk/byterun/interp.c https://github.com/ocaml/ocaml/blob/trunk/byterun/interp.c (see the Next macro definition)
  - simias 11y ago
    I don't follow you, what do computed gotos have to do with threaded code? I'm not sure I understand the code you've linked but "goto (void )(jumptbl_base + *pc++)" looks very much like my big "match" statement except that I match only a subset of the 32 bit instruction. I don't see why the compiler wouldn't be able to rewrite a bit switch/match statement as a jump table without having to spell it out directly. > I'm not talking about jump tables. I'm talking about replacing the bytecode indices with jump addresses once (yes, with a precalculated table) and then doing an indirect jump for each translated opcode from the input stream. Now you've lost me completely.
    - sklogic 11y ago
      > I don't follow you, what do computed gotos have to do with threaded code? 1. Build a jump table 2. Read a bytecode 3. Replace all the instruction codes in the bytecode stream with values from the jump table (i.e., build a threaded code from your bytecode). 4. In a loop, take the next value at PC and jump to it - it's already an address. No table lookups (just once, when you load your bytecode). So it's much faster than any switch can be. EDIT: probably you did not notice that for most of the targets jumptbl_base = 0, and an indirection is removed.
      - arcatek 11y ago
        Even without inlining? Since the compiler cannot know what instruction will be called, it cannot inline it in the loop.
        sklogic 11y ago
        Why do you want any inlining? It's a threaded code. You know an address of each instruction handler, so you can replace an opcode with this address and eliminate a switch and any table lookups altogether.
      - simias 11y ago
        Oh I think I get it, thanks (although I still don't understand what "threaded" means in this context). The only downside I can see is that you'd significantly increase the size of the code. On a 64bit architecture you trade each byte for a 64bit address, effectively multiplying by eight the footprint in the data cache. A 256 entry LUT on the other hand will fit snugly in cache and the lookup shouldn't be very costly. Also if I understood you correctly what you're proposing doesn't have much to do with the "computed gotos" extension.
        sklogic 11y ago
        > I still don't understand what "threaded" means in this context https://en.wikipedia.org/wiki/Threaded_code#Indirect_threading https://en.wikipedia.org/wiki/Threaded_code#Indirect_threadi... > The only downside I can see is that you'd significantly increase the size of the code This is why adding the jumptbl_base on 64-bit platforms. Pointers are still 32-bit, just with an added offset (and a tiny overhead). > doesn't have much to do with the "computed gotos" extension You cannot implement indirect threaded code without a computed goto. You can implement a direct threaded code, of course (generating jump or call instructions directly), but this is a totally different thing.
- plq 11y ago
  > You should first check if the compiler doesn't do it for you. It probably doesn't, because otherwise Intel wouldn't have submitted this patch that adds computed goto support to CPython: http://permalink.gmane.org/gmane.comp.python.devel/153401 http://permalink.gmane.org/gmane.comp.python.devel/153401
  - simias 11y ago
    I don't know the specifics but from the linked thread at https://bugs.python.org/issue4753 https://bugs.python.org/issue4753 : Having a jump table in itself isn't special (the compiler does exactly that when compiling the switch() statement). What's special is that a dedicated indirect jump instruction at the end of each opcode helps the CPU make a separate prediction for which opcode follows the other one, which is not possible with a switch statement where the jump instruction is shared by all opcodes. I believe that's where most of the speedup comes from. So I don't think you can conclude that the compiler "probably doesn't" generate the jump table, just that you might do a better job yourself if you know what you're doing.
  - xigency 11y ago
    The same performance can easily be achieved by simple compiler optimizations like inlining functions, unrolling loops, peephole optimizations, etc. The fact that your compiler doesn't support the same performance doesn't necessarily mean to re implement things in a confusing way. Another way of achieving the exact same result would be to use an array of function pointers. I think the more interesting dilemma is when you have a more complex VM possibly operating on an abstract syntax tree instead of just a bytecode stream, and the ways to optimize code (in C) while maintaining a proper stack. After figuring out how to make a switch statement work, there are a ton of more pressing organizational concerns like how to make function trampolining work without any overhead. Another way to make it faster would be to have it interpret * two * op-codes at once, but hey, that wouldn't do anything except save two cycles and explode the code size to a square of what it was.
    - sklogic 11y ago
      Yes, that python patch is only halfway to a proper indirect threaded dispatch, it still got a lookup table. See the OCaml bytecode VM for a really high performance example of this approach. > a more complex VM possibly operating on an abstract syntax tree Don't do it.
- arcatek 11y ago
  Eh, I was reading your psx guide a moment ago :) Small world!

And if you want a better performance, you can replace this switch with computed gotos (supported at least by gcc and clang).

by sklogic 11y ago

And if you want a better performance, you can replace this switch with computed gotos (supported at least by gcc and clang).

simias 11y ago

You should first check if the compiler doesn't do it for you. Otherwise you're just making your code harder to maintain and less portable for zero gain. In my experience with similar code you can get GCC and LLVM to generate a proper jump table if you're careful to make your code as regular as possible in order for the optimizer to understand what you're doing. Case in point: https://github.com/simias/psx-rs/blob/master/src/cpu/mod.rs#L325-L400 https://github.com/simias/psx-rs/blob/master/src/cpu/mod.rs#... Here LLVM manages to inline all the functions and generate a proper jump table. IIRC I tried getting the function address in the "match" and calling it afterwards but then the compiler wouldn't optimize it properly anymore.

sklogic 11y ago
> You should first check if the compiler doesn't do it for you. No, compiler cannot rewrite your bytecode into a threaded code. > Otherwise you're just making your code harder to maintain and less portable for zero gain. No. Threaded code is almost always faster, even if run only once. > a proper jump table I'm not talking about jump tables. I'm talking about replacing the bytecode indices with jump addresses once (yes, with a precalculated table) and then doing an indirect jump for each translated opcode from the input stream. An example: https://github.com/ocaml/ocaml/blob/trunk/byterun/interp.c https://github.com/ocaml/ocaml/blob/trunk/byterun/interp.c (see the Next macro definition)
- simias 11y ago
  I don't follow you, what do computed gotos have to do with threaded code? I'm not sure I understand the code you've linked but "goto (void )(jumptbl_base + *pc++)" looks very much like my big "match" statement except that I match only a subset of the 32 bit instruction. I don't see why the compiler wouldn't be able to rewrite a bit switch/match statement as a jump table without having to spell it out directly. > I'm not talking about jump tables. I'm talking about replacing the bytecode indices with jump addresses once (yes, with a precalculated table) and then doing an indirect jump for each translated opcode from the input stream. Now you've lost me completely.
  - sklogic 11y ago
    > I don't follow you, what do computed gotos have to do with threaded code? 1. Build a jump table 2. Read a bytecode 3. Replace all the instruction codes in the bytecode stream with values from the jump table (i.e., build a threaded code from your bytecode). 4. In a loop, take the next value at PC and jump to it - it's already an address. No table lookups (just once, when you load your bytecode). So it's much faster than any switch can be. EDIT: probably you did not notice that for most of the targets jumptbl_base = 0, and an indirection is removed.
    - arcatek 11y ago
      Even without inlining? Since the compiler cannot know what instruction will be called, it cannot inline it in the loop.
      - sklogic 11y ago
        Why do you want any inlining? It's a threaded code. You know an address of each instruction handler, so you can replace an opcode with this address and eliminate a switch and any table lookups altogether.
    - simias 11y ago
      Oh I think I get it, thanks (although I still don't understand what "threaded" means in this context). The only downside I can see is that you'd significantly increase the size of the code. On a 64bit architecture you trade each byte for a 64bit address, effectively multiplying by eight the footprint in the data cache. A 256 entry LUT on the other hand will fit snugly in cache and the lookup shouldn't be very costly. Also if I understood you correctly what you're proposing doesn't have much to do with the "computed gotos" extension.
      - sklogic 11y ago
        > I still don't understand what "threaded" means in this context https://en.wikipedia.org/wiki/Threaded_code#Indirect_threading https://en.wikipedia.org/wiki/Threaded_code#Indirect_threadi... > The only downside I can see is that you'd significantly increase the size of the code This is why adding the jumptbl_base on 64-bit platforms. Pointers are still 32-bit, just with an added offset (and a tiny overhead). > doesn't have much to do with the "computed gotos" extension You cannot implement indirect threaded code without a computed goto. You can implement a direct threaded code, of course (generating jump or call instructions directly), but this is a totally different thing.
plq 11y ago
> You should first check if the compiler doesn't do it for you. It probably doesn't, because otherwise Intel wouldn't have submitted this patch that adds computed goto support to CPython: http://permalink.gmane.org/gmane.comp.python.devel/153401 http://permalink.gmane.org/gmane.comp.python.devel/153401
- simias 11y ago
  I don't know the specifics but from the linked thread at https://bugs.python.org/issue4753 https://bugs.python.org/issue4753 : Having a jump table in itself isn't special (the compiler does exactly that when compiling the switch() statement). What's special is that a dedicated indirect jump instruction at the end of each opcode helps the CPU make a separate prediction for which opcode follows the other one, which is not possible with a switch statement where the jump instruction is shared by all opcodes. I believe that's where most of the speedup comes from. So I don't think you can conclude that the compiler "probably doesn't" generate the jump table, just that you might do a better job yourself if you know what you're doing.
- xigency 11y ago
  The same performance can easily be achieved by simple compiler optimizations like inlining functions, unrolling loops, peephole optimizations, etc. The fact that your compiler doesn't support the same performance doesn't necessarily mean to re implement things in a confusing way. Another way of achieving the exact same result would be to use an array of function pointers. I think the more interesting dilemma is when you have a more complex VM possibly operating on an abstract syntax tree instead of just a bytecode stream, and the ways to optimize code (in C) while maintaining a proper stack. After figuring out how to make a switch statement work, there are a ton of more pressing organizational concerns like how to make function trampolining work without any overhead. Another way to make it faster would be to have it interpret * two * op-codes at once, but hey, that wouldn't do anything except save two cycles and explode the code size to a square of what it was.
  - sklogic 11y ago
    Yes, that python patch is only halfway to a proper indirect threaded dispatch, it still got a lookup table. See the OCaml bytecode VM for a really high performance example of this approach. > a more complex VM possibly operating on an abstract syntax tree Don't do it.
arcatek 11y ago
Eh, I was reading your psx guide a moment ago :) Small world!