Am Montag, 25. September 2006 00:06 schrieb Karl Forner: > On 9/15/06, Leopold Toetsch <[EMAIL PROTECTED]> wrote: > > - There are 2 use cases for register_move > > a) optimized recursive tail calls > > b) simple function call argument passing compiled to native code in JIT > > core
> So in that case ( a or b ?), the src_regs are always (0,1,2,..,n) ? > If so I'm sure that we can use this property to optimize the code. b) isn't yet using Parrot_register_move(). Therefore it isn't given that src_regs are always register positions. > The problem is that I have no clue of JIT, registers and so on. But could > you give an example of code that would > lead to that case ? Let's forget that case. JIT code is only created for simple and matching arguments. > > Again, we have to check if any argument conversions could take place and > > then > > turn off the optimization in case. > > ok, so where would the check take place ? > I can imagine an extra argument indicating whether or not doing this extra > optimization, or two different functions/implementations, > or more infomration about the registers allowing the function to decide by > itself to do or not the optimization on a register-couple basis .... No. There have to be an extra check before tailcall optimization is done. When there's argument conversion, which involves P registers, Parrot_register_move() will not be used at all. > So is there anything more I could do on this subject ? Well, you could have a look at recursive_tail_call() in compilers/imcc/pcc.c and improve the checks, if the tailcall can be optimized. > Karl leo
