495848 - ABC Interpreter should use computed-gotos when possible

Status: RESOLVED WONTFIX

(Tamarin Graveyard :: Interpreter, defect)

Description

[Steven Johnson]

Currently, the ABC-only interpreter always dispatches via a switch statement. Using computed-goto on compilers that support it is a measurable speed gain (up to 15% in TT IIRC) and should be simple to implement.

Comment 1

[Edwin Smith]

the abc interpreter can't (by definition) use direct-threaded code.  It could however use indirect threaded code plus computed gotos.

Comment 2

[Lars T Hansen]

(In reply to comment #1)
> the abc interpreter can't (by definition) use direct-threaded code.  It could
> however use indirect threaded code plus computed gotos.

Steven swears it helps.  Probably compiler and CPU dependent (you get rid of at most a range check in the switch, plus a jump to the switch dispatch, plus you get marginally better branch prediction at the end of each instruction).

Comment 3

[Edwin Smith]

agreed, seems like low hanging fruit on GCC or ARMCC even without direct threading.

Comment 4

[Lars T Hansen]

I'm skeptical about the 15% given the current interpreter structure with our very-fat bytecodes (compare to TT's very skinny ones), and given the amount of branching and the presumably relatively simplistic branch prediction hardware on mobile CPUs I'm skeptical about those gains too.  It may be low hanging fruit, but is it edible?  :-)

Comment 5

[Edwin Smith]

I would be extatic if we got 15%, we probably won't.  i think we should keep this open as an option but not work on it until there's evidence it will matter.

Comment 6

[Steven Johnson]

(In reply to comment #5)
> I would be extatic if we got 15%, we probably won't. 

No, but I wouldn't be surprised if it was 2-3%, and it's a small amount of work...