In the softfloat.c file there is a function called float64_add(). What I need
is a way to pack and canonicalize a FloatParts value. There doesn't appear to
be any way. This is what I have tried so far, but the results are not good.
The intermediate_result and the rounded_result should be close to each other.
What currently happens is the values are very far from each other.
float64 __attribute__((flatten)) float64_add(float64 a, float64 b,
float_status *status)
{
FloatParts pa = float64_unpack_canonical(a, status);
FloatParts pb = float64_unpack_canonical(b, status);
FloatParts pr = addsub_floats(pa, pb, false, status);
FloatParts intermediate_parts = canonicalize(pr, &float64_params, status);
float64 intermediate_result = float64_pack_raw(intermediate_parts);
float64 rounded_result = float64_round_pack_canonical(pr, status);
if (intermediate_result != rounded_result) {
printf("a = 0x%" PRIx64 "\n", a);
printf("b = 0x%" PRIx64 "\n", b);
printf("intermediate result: 0x%" PRIx64 "\n", intermediate_result);
printf("rounded result: 0x%" PRIx64 "\n", rounded_result);
}
return rounded_result;
}
Here is some example output:
a = 0x40b11f0000000000 4383
b = 0x4076d00000000000 365
intermediate result: 0x40d0000000000000 16384
rounded result: 0x40b28c0000000000 4748
Here the intermediate result is way bigger than the rounded result.
What I hope to do is implement a PowerPC CPU flag that indicates if an rounded
result is inexact.
