Date:	Thu, 03 Nov 1994 16:10:29 +0100
From:	Peter Csizmadia <cspt AT ludens DOT elte DOT hu>
To:	djgpp AT sun DOT soe DOT clarkson DOT edu
Subject:	2.4/3-2.4/3!=0

main()
{
   register double a1=2.4, b1=3.0;
   volatile double a2=2.4, b2=3.0;
   printf("2.4/3-2.4/3 = %le\n", a1/b1-a2/b2);
}

gcc a.c -lm -O1
go32 a.out
2.4/3-2.4/3 = -3.70255041903800e-17

Is it a normal effect of -O1? (a1=a2, b1=b2, a1/b1!=a2/b2)

Other problem:

#include <stdio.h>
#include <math.h>

main()
{
   volatile double a = 1.54734129e8;
   double b = sqrt(1+a*a);
   double x = a/b;
   printf("1-x=%le\n", 1.0-x);
   printf("1-x=%le\n", 1.0-x);
   printf("x%s1\n", (x==1.0)? "==":"!=");
   printf("1-x=%le\n", 1.0-x);
}

I compiled it with DJGPP 1.12.maint2:
gcc f.c -o f -lm -O1
gcc f.C -o fpp -lm -O1

These are the results on a 386 with a coprocessor:

go32 f
1-x=2.08708918203460e-17
1-x=2.08708918203460e-17
x==1					(?)
1-x=0.00000000000000e+00

go32 fpp
1-x=-2.09251019289702e-17		(it should be >=0)
1-x=-2.09251019289702e-17
x==1					(?)
1-x=0.00000000000000e+00

The numbers are a bit different with emu387, but the problem is the same.
I think one (at least) of the optimizations turned on by -O1 is buggy, but 
don't know which. And I have no idea about the cause of the difference  
between C and C++ compiling.

- Raw text -