gcc floating point bug(?) update + Source

[Christoph Probst]

Hi Juggy.

juggy@gmx.net wrote:

> (2) tstbug.c:
>      When I tried to isolate the error in my software I first wrote this
>      little program. Though I wasn't able to reproduce the exact same
>      behaviour, but some other oddities showed up.


I minimized the tstbug.c code to get the error more obvious:


/--[ code sample gcc-test.c ]-----------------------------------------

#include <math.h>

double a, b, c, d = +3.247778884435419843867975941975601017475128173828125E+00;

int main() {

  a = -3.1493999999999999772626324556767940521240E+01/d;

  if (fabs(a)> fabs(d)) {

    b = 1;

  }

  printf("%+1.40E\n", 31.49 + a * 3.247778884435419843);

  return 0;
}

\---------------------------------------------------------------------



gcc -O1 gcc-test.c ; ./a.out  
-3.9999999999989436366698569713662436697632E-03

gcc -O2 gcc-test.c ; ./a.out
-4.0000000000013340856197530115423433016986E-03


The results differ depending on the compiler optimization level.

One might say this is a normal rounding error, but how would you explain
the following behaviour?  Make little changes to the above code


e.g. assign d inside the main()-function and not before:

------------------------------------8<---------------------------------
double a, b, c, d;

int main() {

  d = +3.247778884435419843867975941975601017475128173828125E+00;
  a = -3.1493999999999999772626324556767940521240E+01/d;

...
------------------------------------8<---------------------------------



OR add a simple printf("test") to the if-statement:

------------------------------------8<---------------------------------
...

  if (fabs(a)> fabs(d)) {
    printf("test");
    b = 1;

  }

...
------------------------------------8<---------------------------------


and the rounding error disappears:


gcc -O1 gcc-test.c ; ./a.out
-3.9999999999989436366698569713662436697632E-03

gcc -O2 gcc-test.c ; ./a.out
-3.9999999999989436366698569713662436697632E-03


Testet on Gentoo Linux, Kernel 2.4.20, gcc version 2.95.3 20010315
Tests with gcc-3.2.1 and tests on FreeBSD machines point out no error
at all ...

HTH
Chris

-- 
"Beware of bugs in the above code; I have only proved it correct,
 not tried it." - Donald Knuth