|
From: | Dave N6NZ |
Subject: | Re: [avr-libc-dev] 64-bit doubles |
Date: | Sat, 20 Dec 2008 22:27:02 -0800 |
User-agent: | Thunderbird 1.5 (X11/20051201) |
Ron Kreymborg wrote:
Isn't there some compiler infrastructure that would have to happen also?Yes, we need to have both the library support, as well as the compiler support. The compiler work might be easier than the library work, but this is just conjecture. EricFor the moment I am thinking of implementing fp doubles as the typedef"longfp" in long longs.
Well, I learned something. I've never wanted long long's in my AVR code, so I never went looking for them :) So now I know they are there.
As Preston has mentioned, this would be outside the compiler. They would have the IEEE format, ie 52-bit mantissa and 11-bit exponent. Obviously not the most efficient but very testable.
Testable? Well.... yes and no :)Yes, in that since you are choosing the IEEE format many machines will give you most of the expected results that you need, modulo a small amount of errata.
No, in that the test space is enormous. Even if you can get expected results easily, it still takes a lot of careful test design to select the stimulus effectively.
-dave
Ron
[Prev in Thread] | Current Thread | [Next in Thread] |