DESCRIPTION: This problem arises if you mix the use of this two ways of generating random number: the srand()/rand() way and the srandom()/random() function family. The issue is that srand() is somehow initializing the seed for random() function, but it should be initialized *only* with srandom(). The expected behavior is random() generator not modifying its seed until done by srandom(), the current behavior is srand() changing that seed.

IMPACT: As happened to me with a scientific code, it could imply a portability problem as this is working properly (from my humble point of view) in other implementations of the library (I tested the different behavior between GNU glibc 2.12 packaged in a Scientific Linux and the BSD/MacOS-Darwin one). I detected also the problem in GNU implementation of cstdlib with C++.

HOW TO REPRODUCE: It's very easy, this is a test program:

---- START CODE ----
// Test random()
#include <stdio.h>
#include <stdlib.h>

int main() {

  int i;

  srand((unsigned)time(NULL));

  for (i=1;i<11;i++) {
	printf("%g \n",(random() / (double)0x7fffffff));
  }
}
---- CODE END ----

This is what you get with glibc-headers-2.12-1.80:

$ ./rndtest 
0.0940072
0.269949
0.729428
0.48224
0.0592921
0.4699
0.132521
0.973561
0.532405
0.167931
$ ./rndtest 
0.954241
0.686682
0.0374202
0.304717
0.497409
0.197191
0.496588
0.273147
0.345674
0.756533

This is the expected (what you get in the mentioned other implementation of stdlib.c and cstdlib):

> ./rndtest 
0.840188 
0.394383 
0.783099 
0.79844 
0.911647 
0.197551 
0.335223 
0.76823 
0.277775 
0.55397 
> ./rndtest
0.840188 
0.394383 
0.783099 
0.79844 
0.911647 
0.197551 
0.335223 
0.76823 
0.277775 
0.55397 

Many thanks.