Date: Mon, 3 Nov 1997 10:26:50 +0200 (IST)
From: Eli Zaretskii <eliz AT is DOT elta DOT co DOT il>
To: Mr Kevin Yeung <keviny AT tst DOT hk DOT super DOT net>
cc: djgpp AT delorie DOT com
Subject: Re: DJGPP failed library tests?
In-Reply-To: <63ia44$hoo$1@tst.hk.super.net>
Message-ID: <Pine.SUN.3.91.971103102626.14641C-100000@is>
MIME-Version: 1.0
Content-Type: TEXT/PLAIN; charset=US-ASCII
Precedence: bulk


On 2 Nov 1997, Mr Kevin Yeung wrote:

>   assert(sprintf(buf, "%E|%.2f|%Lg", 1.1e20, -3.346, .02L) == 23);
>   assert(strcmp(buf, "1.100000E+20|-3.35|0.02") == 0);
>   assert(sscanf(buf, "%e|%lg|%Lf", &fl, &db, &ld) == 3);
>   assert(fabs(fl - 1.1e20) / 1.1e20 < 4 * FLT_EPSILON);
>   assert(fabs(db + 3.35) / 3.35 < 4 * DBL_EPSILON);
>   assert(fabs(ld - 0.02) / 0.02 < 4 * LDBL_EPSILON);
>   return(0);
> }
> 
> It seems to be the problem of the string functions rather than that of the
> math functions.

You need to be more careful with your constants.  Replace this line:

>   assert(fabs(ld - 0.02) / 0.02 < 4 * LDBL_EPSILON);

with this:

>   assert(fabs(ld - 0.02L) / 0.02 < 4 * LDBL_EPSILON);

(note the `L' after 0.02), and it will work as expected.  When you use
0.02, the code uses double constant 0.02, which loses accuracy.  As a
matter of principle, you should always use the same types when dealing
with floating-point code.  Don't mix floats with doubles and doubles
with long doubles unless you both absolutely have to and know what you
are doing.

> But the maths functions don't trap EDOM errors is for
> sure.

Please post your code.  I don't understand what does ``don't trap EDOM
errors'' mean.

In general, you need to link with -lm to get more compatible math
library, and if you want your program to set errno rather than crash
in certain cases (like sqrt(-1)), you need either to use `_control87'
library function to set certain bits of the x87 processor so that it
doesn't generate exceptions, or install a handler for signal SIGFPE.