optimization bug in g++ 4.3.3

[Tim Prince]

Ed Boas wrote:
> The following code:
> 
> ----------
> #include <stdio.h>
> #include <inttypes.h>
> 
> // Rounds in current rounding mode, which is "round to closest
> integer" by default on Pentium & Xeon
> // Valid for |fval| <= 2^22 - 1, that is, -4194303 <= fval <= 4914303
> // based on code from www.lomont.org
> inline int32_t fast_round(float x) {
>  x+=12582912; // 2^23+2^22
>  int32_t i=*(int32_t*)&x;
>  i&=8388607; // 2^23-1;
>  i-=4194304; // 2^22
>  return i;
> }
> 
> int main(void) {
>  printf("%f rounds to %i\n", 1.4f, fast_round(1.4f));
>  return 0;
> }
> ----------
> 
> produces the correct output when compiled using g++ -O:
> 
> 1.400000 rounds to 1
> 
> However, when it is compiled with g++ -O2, it prints:
> 
> 1.400000 rounds to -4194304
> 


$ gcc -Wall  -O2  ed.c
ed.c: In function 'int32_t fast_round(float)':
ed.c:9:25: warning: dereferencing type-punned pointer will break 
strict-aliasing
  rules

TCPRINCE@tcprince-MOBL5 /cygdrive/c/Documents and Settings/tcprince/My 
Documents
/tim/tim/src/net
$ ./a
1.400000 rounds to 1

gcc 4.5 doesn't show the problem of gcc 4.3.2 on this broken source 
code.  You invoke Undefined Behavior with a past release of gcc, you get 
what you paid for.  Who's to say what is "correct output?"

Set -no-strict-aliasing, the warning goes away, and even gcc 4.3 gives 
your "correct output."