On Tue, Feb 05, 2002 at 10:54:42PM +0800, Zhang Fuxin wrote:
> hi,
> It seems gcc with optimization > -O2 will produce incorrect code for fp code:
> I find this example when tracking down the libm-test failures.
> ( expf(NaN) == NaN will report an extra "invalid operation" exception )
> The faulting code snippet is:
> float __my_expf(float x) /* wrapper expf */
> {
> float z;
> unsigned int hx;
> z = __ieee754_expf(x);
> if(_LIB_VERSION == _IEEE_) return z;
> if(__finitef(x)) {
> if(x>o_threshold)
> return 1.0;
> else if(x<u_threshold)
> return 2.0;
> printf("haha\n");
> }
> return z;
> }
>
> with -O2(or higher),some statements inside "if (__finitef(x))" are put before
> testing the return value of __finitef(x),one of which is a c.lt.s that cause
> the extra "invalid operation" exception being raised.
>
This bug seem fixed in gcc 3.1.
H.J.
