mesa: Sparc's IROUND() optimization is invalid.

We can't use the "fstoi" instruction like this.

Unlike other floating point instructions, "fstoi" always rounds
towards zero no matter what rounding mode the FPU has been set to.

This was validated using the following test program:

--------------------
static inline int iround(float f)
{
   int r;
   __asm__ ("fstoi %1, %0" : "=f" (r) : "f" (f));
   return r;
}
#define IROUND(x)  iround(x)

#define IROUND_REF(f)  ((int) (((f) >= 0.0F) ? ((f) + 0.5F) : ((f) - 0.5F)))

int main(void)
{
	float f = -2.0;

	while (f < 3.0f) {
		int sparc_val = IROUND(f);
		int ref_val = IROUND_REF(f);

		if (sparc_val != ref_val)
			printf("DIFFERENT[%f]: REF==%d SPARC==%d\n",
			       f, ref_val, sparc_val);
		f += 0.1f;
	}

	return 0;
}
--------------------

which prints out things like:

--------------------
DIFFERENT[-1.900000]: REF==-2 SPARC==-1
DIFFERENT[-1.800000]: REF==-2 SPARC==-1
DIFFERENT[-1.700000]: REF==-2 SPARC==-1
DIFFERENT[-1.600000]: REF==-2 SPARC==-1
DIFFERENT[-1.000000]: REF==-1 SPARC==0
DIFFERENT[-0.900000]: REF==-1 SPARC==0
DIFFERENT[-0.800000]: REF==-1 SPARC==0
DIFFERENT[-0.700000]: REF==-1 SPARC==0
DIFFERENT[-0.600000]: REF==-1 SPARC==0
DIFFERENT[0.500000]: REF==1 SPARC==0
DIFFERENT[0.600000]: REF==1 SPARC==0
...
--------------------

So we have to remove Sparc's IROUND() definition, it's wrong.

Signed-off-by: David S. Miller <davem@davemloft.net>

0 files changed, 0 insertions, 0 deletions