Re: [boost] Looking for some "real world" extendedprecision integer arithmetic tests

Subject: Re: [boost] Looking for some "real world" extendedprecision integer arithmetic tests
From: John Maddock (boost.regex_at_[hidden])
Date: 2012-01-26 13:44:57

> That's a little alarming actually.  If the integer is fixed length that 
> means it may overflow and of course will also truncate in division since 
> it is integer.  Just like the compiler, we can't apply an optimization 
> that changes the result the code may produce when overflow and truncation 
> are considered as part of the normal behavior of the type.  Sometimes 
> arithmetic is written specifically with overflow and truncation in mind. 
> Since any arithmetic operation may overflow this can really tie your 
> hands.  For numerical uses of fixed int we may assume that the programmer 
> designed the code in such a way that overflow should never happen, but for 
> behavioral simulation of hardware with fixed width integer units the 
> correct overflow behavior is the whole point.

Very good point!

Looks like I'll have to write that other frontend then....

> By the way, I'm really excited about the expression template frontend.  I 
> like the elimination of unneeded copies and cutting down on the number of 
> temporaries.  I also think it is really cool that you are able to do the 
> arithmetic transformations on the expression template trees.  Are you 
> using proto?

No.  I started off using Proto, but compile times were so long it pretty 
much melted down my laptop :-(

After some discussion on the list here I switched to a home-rolled 
expression-template implementation that was designed to minimize the number 
of templates instantiated (not to mention the amount of meta-programming 
involved).  Even so, compile times can still be quite long if you 
instantiate lots of different expressions, but it is at least 
manageable/usable now.

John.