Thread Links Date Links
Thread Prev Thread Next Thread Index Date Prev Date Next Date Index

Re: Comparisons and decorations



> From: "Nate Hayes" <nh@xxxxxxxxxxxxxxxxx>
> To: "Dan Zuras Intervals" <intervals08@xxxxxxxxxxxxxx>,
> 	"John Pryce" <j.d.pryce@xxxxxxxxxxxx>
> Cc: <stds-1788@xxxxxxxxxxxxxxxxx>
> Subject: Re: Comparisons and decorations
> Date: Sat, 25 Sep 2010 20:09:59 -0500
> 
> Dan,
> 
> I think yours is best so far, and agree with problems
> of using nextDown and nextUp for reasons you mention.

	Thanks.
	Somebody else'll beat it sooner or later.
	These things are so fleeting... :-)

> 
> My only tweak is that I might not use short-circuit
> conditionals, since this introduces branching. I
> suspect the performance hit from the branch will be 
> bigger than just performing all four comparisons.

	Sometimes.  Sometimes not.

	More & more these days conditionals (even including
	short-circuiting) are being implemented as either
	setting a condition code on the fly or a conditional
	store.  The short circuit is when the following
	conditions are conditioned on the result of earlier
	ones.

	Some of these architectures are more efficient at
	this than others.  Some allow other instructions
	to slip into the unused execution slot more often
	than not.  Some don't.

	But almost everyone is dealing with the problems
	of conditional branches to one extent or another.

	Even Intel is getting involved in some interesting
	microarchitectures.  Its amazing what you can do
	with a 40 year old architecture when you put your
	mind to it.
	
	Well, that & a couple billion dollars development. :-)


				Dan

> 
> For me, one downside of all these proposals is they
> would break Jurgen's overlapping operator, which is
> something I've grown fond of.
> 
> Nate
>