Re: Folk Theorem: Assemblers are superior to Compilers

> Dave Gillespie <synaptx!daveg@uunet.UU.NET> writes:


> My favorite example comes from an i860 loop we wrote here,
> which involved a bunch of multipy-adds followed by something
> of the form max(min(x, limit), -limit), i.e., constrain
> abs(x) <= limit. Branches are awful on the i860...


I am not sure why a compiler would want to use branches. Noting that


max(x,limit) = max(x-limit,0) + limit


it suffices to evaluate max(z,0) up to a couple of arithmetic
operations. Now,


max(z,0) = z & (z > 0)


where "&" means "bitwise and", and "z > 0" is a word of all ones if z
is greater than zero and a word of all zeros if it is not. This may be
computed by propagating the sign bit of z through an entire word and
logically negating it. In C terms this leads to the following function
(for 32 long ints and floats with the sign bit at the left end)


float max(float x, float y) {
                    union fl {float f; long l;} z;
z.f = x - y;
z.l &= ~z.l >> 31;
return z.f + y;}


I am not claiming that this is faster than your trick, but the C
compiler *could* generate code without any branches.


(Caveat: I checked my method on a VAX, where the sign bit is in the
middle of word. I have not directly checked the version given here).
--

Return to the comp.compilers page.
Search the comp.compilers archives again.