🚀 go-pugleaf

I'm writing a program in DJGPP that uses fixed-point math.  It needs to
run as quickly as possible.  Currently I'm implementing multiplication
and division with the following functions:

typedef int fixed;

static inline fixed FixMul(fixed a,fixed b)
{
        fixed result;

        asm("movl %1,%%eax;imull %2;shrd $16,%%edx,%%eax;movl %%eax,%0"
         : "=g"(result) : "g"(a),"g"(b) : "%eax","%edx");
        return result;
}

static inline fixed FixDiv(fixed a,fixed b)
{
        fixed result;

        asm("movl %1,%%eax;movl %%eax,%%edx;shll $16,%%eax;sarl $16,%%edx;"
            "idivl %2;movl %%eax,%0"
         : "=g"(result) : "g"(a),"g"(b) : "%eax","%edx");
        return result;
}

This works pretty well, but it would be nice if GCC knew that it could use
%eax as an input and/or output register.  I could eliminate "%eax" from the
list of trashed registers, but it seems like this would confuse GCC into
thinking that %eax was preserved by these functions, leading to potential
problems.  Also, when I tried this, I got instructions like "movl %eax,%eax".

I'm not very well-versed in the intricacies of the asm() statement; is there
a better way to implement these functions?

--
Ben Rudiak-Gould
benrg@uclink.berkeley.edu

"Those are my principles. If you don't like them I have others." --Groucho Marx