🚀 go-pugleaf

Does anyone know how to do fast 32-bit fixed point multiplication on a
386? The numbers involved are stored with the high 16 bits before the
decimal point and the low 16 bits after it, and signed. I wrote the
following function in C, and compiled it under Zortech C version 3, and
it ended up as about 50 machine instructions, which on my 33 MHz 386
works out as about 100,000 multiplications per second. Considering that
the Zortech single-precision floating-point libraries can get the same
machine (without floating-point hardware) to ~200 Kflops, there would
seem to be little point in using fixed point, and I can't figure out how
to shave more than a few clock cycles off the compiler's code. If
anyone knows of a faster solution, it would be much appreciated.

typedef long num;
 
num mul (num x,num y)
{
	long xh,xl,yh,yl,result;
	int sign;
 
	sign = 0;
	if (x < 0)
	{
		x = -x;
		sign = 1;
	}
	if (y < 0)
	{
		y = -y;
		sign = 1 - sign;
	}
	xh = x >> 16;
	xl = x & 0xffff;
	yh = y >> 16;
	yl = y & 0xffff;
	result = (xh*yh << 16) + (xh*yl) + (yh*xl) + (xl*yl >> 16);
	if (sign)
		return -result;
	return result;
}


--
"To summarize the summary of the summary: people are a problem"
Russell Wallace, Trinity College, Dublin
rwallace@unix1.tcd.ie