Menu
▾
▴
#295 Optimize division by powers of two.
x / 2^n for a int16_t x
could be implemented like this:
int16_t divbytwo(int16_t x, uint8_t n)
{
bool s = (x < 0);
uint16_t v = x;
do
v >>= 1;
while(--n);
return(s ? -v : v);
}
This would be significantly faster than the currently used generic division.
Philipp
Moreover it is possible to optimize (for speed) any multiplication with constant using shifts, additions and subtractions.
Examples:
x * 7 == {t = x; x <<= 3; x -= t; }
x * 5 == {t = x; x <<= 2; x += t; }
Hi b-s-a,
code for optimizing multiply with 7 and 5 was there for PIC once.
case 5:
emitpcode(POC_MOVFW, popGet(AOP(left),0));
emitpcode(POC_ADDFW, popGet(AOP(left),0)); // W = 2*F
emitpcode(POC_ADDWF, popGet(AOP(left),0)); // F = 3*F
emitpcode(POC_ADDWF, popGet(AOP(left),0)); // F = 5*F
return;
case 7:
but unfortunately was removed.
With recent renaming/moving of the pic tree to the pic14 tree
the history of all the files is lost. Not good.
But the really out of date cvs repository still has the function
genUMult8XLit_16(). See
http://sdcc.cvs.sourceforge.net/viewvc/sdcc/sdcc/src/pic/genarith.c?revision=1.36&view=markup
line 1710
There is another reminiscent of these formerly implemented pic14
optimizations in the pic regression test in sdcc/src/regression/mult1.c
which includes checks for multiplies with
2; 3; 4; 5; 6; 7; 8; 9; 10;11;12;13;14;15;16;17;18;19;20;21;22;23;24;31;32;64;128;
See also
http://sourceforge.net/tracker/?func=detail&aid=2997777&group_id=599&atid=300599
and
http://sourceforge.net/tracker/?func=detail&aid=1307995&group_id=599&atid=350599
RFE #1307995 has a more elegant solution. I thus close this item (and open an new RFE for the suggestion from b-s-a).
Philipp
Sorry, I misinterpreted b-s-a's proposal at first reading. What is proposed by b-s-a is already done by at least the Z80 backend for multiplication with constants.
Philipp