Mirror of forum.nim-lang.org

63 :: compile case statements as jump tables in the C codegen

• adrianv (orginal) [2012-07-29T13:24:45+02:00] view original

  again me ;-)

  I read a discussion on the D forum about using jump tables. I think the nimrod case statement over an enum could be a nice possibility to use these jump tables in the c codegen as well. When all possible values of an enum are handled in the case (no else allowed), then this case could be translated into a jump table instead of a switch. This should give a better performance because of a better use of the branch prediction. Maybe something for version 0.9++ ?

---

• Araq (orginal) [2012-07-29T13:40:09+02:00] view original

  The compiler already hints Visual C++ by generating a default: __assume(0); should the case be exhaustive.

  It should be faster to order by statistics so that common enum values come first and then use the linearScanEnd pragma to get very efficient dispatching.

  Computed gotos may be implemented some day. Make a feature request, its priority may be based on some benchmark results. (Hint, hint ;-) )

---

• Araq (orginal) [2013-08-16T14:32:00+02:00] view original

  Update: This feature has been implemented. It only works in combination with while true loops because this allows for code duplication of the dispatching which improves the CPU's branch prediction:

  # bytecode interpreter loop
  while true:
     {.computedGoto.}
     case code[pc]
     of jump: pc = jumpTarget
     of addi: ...
     ...
  Run

---

• adrianv (orginal) [2013-08-16T20:15:57+02:00] view original

  I tested it already - in my test case I got a performance increase of about 30%. There seems to be a problem if instructions are before (problem here) or after the case (no problem here). I will post a bug report when I am sure.

  Wish I had more time to program in nimrod - but I do my best to convince everybody I talk to.

---