<html>
<head>
<meta http-equiv="Content-Type" content="text/html; charset=utf-8">
</head>
<body text="#000000" bgcolor="#FFFFFF">
<p>Hi</p>
<p>I object to one component of Gareth's proposal - to make
log2(0)=0. The problem lies not with what Gareth wants to do with
it, but what other people will use it for once it becomes
available. log2(0) is undefined (and undefinable, as it is not a
computable number), the appropriate choice for log2(0) is to make
it Not-A-Number (NaN). <br>
</p>
<p align="center">FLog2(0) = NaN = CLog2. <br>
</p>
<p>Such a choice would avoid the mess Gustavson got himself into
when he <a moz-do-not-send="true"
href="www.johngustafson.net/pdfs/BeatingFloatingPoint.pdf">mapped</a>
0 and 1/0 onto the same number - a mapping that has many
advantages for computing, but eventually <a
moz-do-not-send="true" href="https://arxiv.org/pdf/1701.00722">destroys
computability</a>. To a lesser degree, this mess is already
present in the IEEE 754 standard for floating-point arithmetic,
and thus led to, to put it mildly, <a moz-do-not-send="true"
href="www.itu.dk/%7Esestoft/bachelor/IEEE754_article.pdf">computing
difficulties</a>, difficulties that many programmers gloss over
- or simply ignore.<br>
</p>
I will have to say more about this when I am ready to file a bug
report on floating point exceptions, since Gareth's proposal has
deep links to how floating point numbers are defined - and why they
were defined such.<br>
<br>
Wolf<br>
<br>
<div class="moz-cite-prefix">On 13/06/2018 00:42, J. Gareth Moreton
wrote:<br>
</div>
<blockquote type="cite"
cite="mid:2794.1528807356@web-cluster.fastnet.co.uk">
<div>
<style> BODY { font-family:Arial, Helvetica, sans-serif;font-size:12px; }</style>Hi
everyone,</div>
<div><br>
</div>
<div>Sorry to flood the mailing list again with more ideas and
experiments.<br>
<br>
I would like to propose introducing a new pair of in-built
functions for the compiler.<br>
<br>
function FLog2(X: Cardinal): Cardinal;<br>
function CLog2(X: Cardinal): Cardinal;<br>
<br>
FLog2 returns the base-2 logarithm of X, rounded down, while
CLog2 returns the base-2 logarithm of X, rounded up.<br>
<br>
To give examples where these functions could be useful, FLog2(X)
+ 1 indicates the minimum number of bits required to store X as
an unsigned integer, while CLog2(X) is equal to the maximum
number of iterations required for a binary search of an X-sized
list.<br>
</div>
</blockquote>
Given the stated purpose, I could withdraw my objection if any
reference to logarithm was removed from the function and its name.
Then Gareth would be free to create his function any way he likes
and assign to it the properties he chooses. The only requirement
left then would be to state in the comments around the function what
it is supposed to achieve, as a deterrence to mis-use and guidance
to future maintainers, who may not think the same as Gareth does.<br>
<blockquote type="cite"
cite="mid:2794.1528807356@web-cluster.fastnet.co.uk">
<div>
Why should they be in-built though? With the binary search
example, the size of the list is sometimes known at compile
time, hence is a constant - therefore, its logarithm is also a
constant. By pre-calculating the logarithm using the in-built
function, it can be used to aid optimization such as loop
unrolling. It also makes them easier to inline, where FLog2(X)
on x86_64-win64 translates to a single line of assembly
language: BSR EAX, ECX (unless X is zero, in which case ZF is
set and EAX is undefined).<br>
<br>
If there had to be a slight nuance though, it's what happens if
X = 0, since log(0) = -oo</div>
</blockquote>
This statement is false. log(0) is not infinity. To obtain a
numerical value for log(0) by e.g. Taylor series expansion, at one
stage you have to divide by zero since the differential<br>
<div align="center">(d ln x )/ d x = 1/x.<br>
</div>
And since 1/0 is not an element of the set of computable numbers,
log(0) is not either. The only valid assignment can be log(0):=NaN,
for any base.<br>
<blockquote type="cite"
cite="mid:2794.1528807356@web-cluster.fastnet.co.uk">
<div>, which cannot be stored as an integer and may cause problems
with the compiler. I would propose it return 0 as a special
case, as this will fix most problems when it comes to loops and
storage, and also ensure that FLog2 and CLog2 are "entire
functions". To give an optimised example of FLog(2) in x86-64
assembly language:<br>
<br>
XOR EDX, EDX<br>
BSR EAX, ECX // ZF is set if ECX is zero<br>
CMOVZ EAX, EDX // Zero (EDX) written to result if ZF is set.<br>
<br>
Some kind of deep optimization could be used if the input is
known to be non-zero and remove the instructions either side of
BSR.<br>
<br>
(Alternative names: FILog2 and CILog2, to indicate they work on
integers and to distinguish them from versions that work with
floating-point numbers)<br>
<br>
Gareth<br>
</div>
<!--'"--><br>
<fieldset class="mimeAttachmentHeader"></fieldset>
<br>
<pre wrap="">_______________________________________________
fpc-devel maillist - <a class="moz-txt-link-abbreviated" href="mailto:fpc-devel@lists.freepascal.org">fpc-devel@lists.freepascal.org</a>
<a class="moz-txt-link-freetext" href="http://lists.freepascal.org/cgi-bin/mailman/listinfo/fpc-devel">http://lists.freepascal.org/cgi-bin/mailman/listinfo/fpc-devel</a>
</pre>
</blockquote>
<br>
</body>
</html>