0.7 / 1.0 ?= 0.52

[Hans Aberg]

On 15 Jun 2012, at 15:07, Torbjorn Granlund wrote:

> Hans Aberg <haberg-1 at telia.com> writes:
> 
>> The division produce a funny bug. What do you think?
>> 
>>   mpf_t a;
>>   mpf_init(a);
>>   mpf_set_d(a, 0.7);
>>   mpf_t b;
>>   mpf_init(b);
>>   mpf_set_d(b, 1.0);
>>   mpf_t l;
>>   mpf_init(l);
>> 
>>   gmp_printf ("%.*Ff \n", 5, a);		--- 0.70000
>>   gmp_printf ("%.*Ff \n", 5, b);		--- 1.00000
>> 
>>   mpf_div(l, a, b);
>>   gmp_printf ("%.*Ff", 5, l);			--- 0.52502
> 
>  This example works on Mac OS X 10.7.4 with GMP 5.0.4 compiled directly
>  from source, using the llvm-gcc 4.2.1 you indicate, and the example
>  compiled using this compiler, clang 3.1 also from Xcode 4.3.2, and gcc
>  4.7.0 compiled from source using this llvm-gcc.
> 
> If llvm-gcc compiles that code, it is more broken than I could ever have
> imagined; I knew its code generator and optimiser were badly broken, but
> now its parser thinks the above code is valid C…!

I had to wrap it into a function, and change the Haskell style '--' comments into C/C++ style (see below). Then just use any of the compilers mentioned above. 

Hans


---- divide-test.c ----
#include <gmp.h>

int main() {
  mpf_t a;
  mpf_init(a);
  mpf_set_d(a, 0.7);
  mpf_t b;
  mpf_init(b);
  mpf_set_d(b, 1.0);
  mpf_t l;
  mpf_init(l);
  
  gmp_printf("%.*Ff \n", 5, a);		// Prints 0.70000
  gmp_printf ("%.*Ff \n", 5, b);	// Prints 1.00000
  
  mpf_div(l, a, b);
  gmp_printf ("%.*Ff", 5, l);		// Prints 0.70000

  return 0;
}
----