OpenEuphoria

Re: 3.0.3

• Posted by Pete Lomax <petelomax at blueyonder.co.uk> May 12, 2007
• 732 views

Derek Parnell wrote:
> I haven't looked into the generated code for it but I guess
> that it's not as efficient as it could be. In fact, I'm not sure that there
> are any optmizations done to the code after it has been generated. Maybe 
> this is an openning for a future enhancement?
I see, I dumped the il to have a look see. These listings are based on 
my hacked 2.5 frontend, but I've checked 3.0.0 and the results seem to
be the same. Derek's routine:

global type boolean (object x)
     if not integer(x) then return FALSE end if
     if x = FALSE then return TRUE end if
     if x = TRUE  then return TRUE end if
     return FALSE
 end type
--il code:
{NOVALUE,99,M_NORMAL,SC_GLOBAL,U_UNUSED,1,"boolean",0,TYPE,52,0,0,-1,0, -- rtn
98
    {IS_AN_INTEGER,99,100 --1: tmp 100 := IS_AN_INTEGER(x)
    ,NOT_IFW,100,10 --4: NOT_IFW tmp 100 (10)
    ,RETURNF,98,94 --7: RETURNF boolean, FALSE
    ,EQUALS_IFW,99,94,17 --10: x EQUALS_IFW FALSE (17)
    ,RETURNF,98,96 --14: RETURNF boolean, TRUE
    ,EQUALS_IFW,99,96,24 --17: x EQUALS_IFW TRUE (24)
    ,RETURNF,98,96 --21: RETURNF boolean, TRUE
    ,RETURNF,98,94 --24: RETURNF boolean, FALSE
    ,BADRETURNF --27: BADRETURNF
    }}

--My routine:
type boo2(object x)
    if integer(x) and (x=FALSE or x=TRUE) then return TRUE end if
    return FALSE
end type
--ilcode:
{NOVALUE,107,M_NORMAL,SC_LOCAL,U_UNUSED,1,"boo2",0,TYPE,1455,0,0,-1,0, -- rtn
106
    {IS_AN_INTEGER,107,108 --1: tmp 108 := IS_AN_INTEGER(x)
    ,SC1_AND_IF,108,108,29 --4: SC1_AND_IF(tmp 108,tmp 108) --> 29
    ,EQUALS,107,94,109 --8: tmp 109 := EQUALS(x,FALSE)
    ,SC1_OR,109,109,23 --12: SC1_OR(tmp 109,tmp 109) --> 23
    ,EQUALS,107,96,110 --16: tmp 110 := EQUALS(x,TRUE)
    ,SC2_OR,110,109 --20: SC2_OR tmp 110,tmp 109
    ,IF,109,29 --23: IF tmp 109 (29)
    ,RETURNF,106,96 --26: RETURNF boo2, TRUE
    ,RETURNF,106,94 --29: RETURNF boo2, FALSE
    ,BADRETURNF --32: BADRETURNF
    }}

Theoretically, the SC1_AND_IF could (in this case) be a plain IF, the 
EQUALS/SC1_OR pair a NOTEQ_IFW to the return TRUE statement, and the 
EQUALS/SC2_OR/IF triple an EQUALS_IFW to the return FALSE statement, 
I think:

{IS_AN_INTEGER,107,108 --1: tmp 108 := IS_AN_INTEGER(x)
    ,IF,108,18 --4: IF tmp 108 (18)
    ,NOTEQ_IFW,107,94,15 --7: x NOTEQ_IFW FALSE (15)
    ,EQUALS_IFW,107,96,18 --11: x EQUALS_IFW TRUE (18)
    ,RETURNF,106,96 --15: RETURNF boo2, TRUE
    ,RETURNF,106,94 --18: RETURNF boo2, FALSE
    ,BADRETURNF --21: BADRETURNF
    }

Easier said than done, no doubt.

Regards,
Pete
PS I should perhaps point out that the latter il obviously ought to beat the
current il for the same source, though I doubt it will improve any on the il
emitted for Derek's code, which looks spot on to me. Both first and last ils
execute the same number of il instructions in all cases. The only reason the
latter is smaller is it encodes only two return statements instead of four.