Re: swap (a, b) suggestion

[Alan Lehotsky <lehotsky at tiac dot net>]

At 8:22 AM +0100 1/4/98, Benny Amorsen wrote:

 * >>>>> "RH" == Richard Henderson <rth@cygnus.com> writes:
 *
 * RH> On Sat, Jan 03, 1998 at 06:11:23PM -0500, Eric Buddington wrote:
 * >> For integer values, the following works, and is significantly
 * >> faster:
 * >>
 * >> template <class T>
 * >> inline void swap(T& a, T& b) {
 * >>   a ^= b;
 * >>   b ^= a;
 * >>   a ^= b;
 * >> }
 *
 * RH> If it is faster at all I'll warrent it is only so for depressingly
 * RH> register starved machines.
 *
 * Perhaps the optimization should be made in a different place -- Make
 * the optimizer recognize the sequence "tmp = a; a = b; b = tmp" and do
 * it the most efficient way. The instruction combiner or the peephole
 * optimizer could be made to catch this.
 *
 *
 * Benny

That's the way I did this in my enhancements to the SHARC DSP port
of GCC.  It has an instruction that allows parallel register
assignments, so I can do

		ra=pass rb, rb=ra;

which in one instruction does the swap.  I suppose for completeness sake, I
ought to add a pattern for the three XORs and reduce them to the single
instruction too :-)

-- Al Lehotsky

------------------------------------------------------------------------

		    Quality Software Management
		http://www.tiac.net/users/lehotsky
			lehotsky@tiac.net
			(617)862-5418 Voice
			(617)674-1096 FAX

	Software Process Improvement and Management Consulting
	     Language Design and Compiler Implementation