Re: [Maxima-discuss] scale-float with sbcl & ccl

[Raymond T. <toy...@gm...>]

On 12/5/25 3:51 PM, Richard Fateman wrote:
> There are nuances in the IEEE standard that are not widely used and 
> sometimes
> were left out of the software support (for laziness, lack of 
> understanding, uncertainty
> of whether this would catch on?) and
> were, certainly in the past, left out of some hardware. (vendors would 
> say they support
> the IEEE format, not mentioning lack of support of some operations.)
> Denormalized numbers are certainly in the category of not-widely-used.
> Common Lisp Standard support for floats was deliberately 
> underspecified -- IEEE floats
> were new at the time.

I think it's safe to say that any x86 or arm processor supports 
IEEE-754.  Handling of NaN was never well-specified.  (Agner Fog has an 
excellent article about this.)

Some older (and newer?) x86 chips had really awful performance with 
denormals, but I think this is mostly fixed.  I understand arm is quite 
good and operations with denormals are about twice the cost of normal 
floats.

>
> Curious --  is there a call for this in some important place, say in a 
> math library routine??

I don't know, but if you poke around fdlibm, you can see the code 
handles denormals.  Presumably other math libraries also handle 
denormals; I haven't checked those.


>
> RJF
>
> On Fri, Dec 5, 2025 at 3:39 PM Raymond Toy <toy...@gm...> wrote:
>
>
>     On 12/5/25 3:24 PM, David Scherfgen wrote:
>>     This is a known bug in SBCL:
>>     https://bugs.launchpad.net/sbcl/+bug/2000178
>     Almost 3 years old.  I guess it's not important enough to anyone.
>>
>>     Raymond Toy <toy...@gm...> schrieb am Fr., 5. Dez. 2025,
>>     16:00:
>>
>>
>>         On 12/5/25 10:21 AM, Barton Willis via Maxima-discuss wrote:
>>>
>>>         For subnormal floats,  the SBCL CL function scale-float is
>>>         possibly buggy:
>>>
>>>         SBCL 2.4.7:
>>>
>>>         Define Maxima function that calls the CL function scale-float
>>>
>>>         (%i1) :lisp(defun $scale_float (x m) (scale-float x m))
>>>
>>>         (%i1) x : 2.0^(-1024);
>>>         (%o1)                       5.562684646268004e-309
>>>
>>>         This is a SBCL bug, I think:  surely scaling by zero should
>>>         be an identity?
>>>
>>>         (%i2) scale_float(x,0);
>>>         (%o2)                       2.7813423231340017e-308
>>>
>>>         (%i3) x : 2.0^(-1023);
>>>         (%o3)                       1.1125369292536007e-308
>>         This is kinda surprising.  Cmucl gets this right. So does
>>         ecl.  Kinda funny too; it's not obvious what's happening
>>         here.  `integer-decode-float` for the scaled result is
>>         `#x14000000000000` and an exponent of -1074, instead of the
>>         expected `#x10000000000000` with exponent -1076.
>>
>>         _______________________________________________
>>         Maxima-discuss mailing list
>>         Max...@li...
>>         https://lists.sourceforge.net/lists/listinfo/maxima-discuss
>>
>     _______________________________________________
>     Maxima-discuss mailing list
>     Max...@li...
>     https://lists.sourceforge.net/lists/listinfo/maxima-discuss
>