From mboxrd@z Thu Jan  1 00:00:00 1970
X-Spam-Checker-Version: SpamAssassin 3.4.4 (2020-01-24) on polar.synack.me
X-Spam-Level: 
X-Spam-Status: No, score=-0.3 required=5.0 tests=BAYES_00,
	REPLYTO_WITHOUT_TO_CC autolearn=no autolearn_force=no version=3.4.4
Path: 
 buffer1.nntp.dca1.giganews.com!border1.nntp.dca1.giganews.com!nntp.giganews.com!goblin2!goblin.stu.neva.ru!aioe.org!.POSTED!not-for-mail
From: "Nasser M. Abbasi" <nma@12000.org>
Newsgroups: comp.lang.ada
Subject: Re: getting same output as gfortran, long_float
Date: Thu, 30 Apr 2015 17:37:13 -0500
Organization: Aioe.org NNTP Server
Message-ID: <mhuaqp$2jk$1@speranza.aioe.org>
References: <mhu65r$ntj$1@speranza.aioe.org>
 <1kxou0nloqg9c$.1x0itzgdrlosm$.dlg@40tude.net>
Reply-To: nma@12000.org
NNTP-Posting-Host: CV72NQ0GT7rQd8cP1ZYi/A.user.speranza.aioe.org
Mime-Version: 1.0
Content-Type: text/plain; charset=windows-1252; format=flowed
Content-Transfer-Encoding: 7bit
X-Complaints-To: abuse@aioe.org
User-Agent: Mozilla/5.0 (Windows NT 6.1; WOW64;
 rv:31.0) Gecko/20100101 Thunderbird/31.6.0
X-Notice: Filtered by postfilter v. 0.8.2
Xref: number.nntp.giganews.com comp.lang.ada:192977
Date: 2015-04-30T17:37:13-05:00
List-Id: <comp.lang.ada>

On 4/30/2015 5:08 PM, Dmitry A. Kazakov wrote:
>
> I don't know what FORTRAN does, but assuming, you are using Intel, the
> longest floating-point available there is not, as you seem assume, of 16
> bytes (quadruple-precision), but so-called extended precision, which is not
> that long (10 bytes, namely).
>

Yes, I do not know either how gfortran does it. I am on intel,
using

>gfortran --version
GNU Fortran (Ubuntu 4.8.2-19ubuntu1) 4.8.2

It might be linked to libquadmath? I see article on this

http://glennklockwood.blogspot.com/2014/02/linux-perf-libquadmath-and-gfortrans.html

"Executive Summary: libquadmath was introduced in GFortran 4.6
which fundamentally changed what the -fdefault-real-8 switch does"

thanks for your input below, will look at it:

> AFAIK, GNAT on x86 maps:
>
> Float to single precision float (4 bytes)
> Long_Float to double precision float (8 bytes)
> Long_Long_Float to extended precision float (10 bytes, effectively)
>
> You can try this program to see how many binary mantissa bits are available
> for each type:
> -----------------------------
> with Text_IO; use Text_IO;
> procedure Test is
> begin
>     Put_Line ("Float mantissa:" & Integer'Image (Float'Machine_Mantissa));
>     Put_Line ("Long Float mantissa:" & Integer'Image
> (Long_Float'Machine_Mantissa));
>     Put_Line ("Long Long Float mantissa:" & Integer'Image
> (Long_Long_Float'Machine_Mantissa));
> end Test;
> --------------------------
> On an Intel machine it will print:
>
> Float mantissa: 24
> Long_Float mantissa: 53
> Long_Long_Float mantissa: 64
>
> If FORTRAN supports longer floating point numbers (e.g.
> quadruple-precision) then it must emulate them since Intel hardware does
> not have them. That would make computations quite slow and you should
> consider if you really need that many significant digits.
>
> If GNAT allowed floating-point emulation, you could declare a custom type
> of required precision:
>
>     type My_Float is digits 34; -- 112 * lg(2)
>
> Which is advisable anyway in order to make your program portable. Using
> built-in types is a bad idea and poor taste in most cases.
>
> Since GNAT does support emulation, at least not with the standard compiler
> switches, you need a library for doing this.
>