Re: [Qemu-devel] [PATCH] fpu/softfloat: check for Inf / x or 0 / x before /0

[Richard Henderson]

On 04/17/2018 12:38 PM, Emilio G. Cota wrote:
> On Tue, Apr 17, 2018 at 22:45:51 +0100, Peter Maydell wrote:
>> On 17 April 2018 at 22:27, Emilio G. Cota <address@hidden> wrote:
>>> BTW I just checked with -t host on an IBM Power8, and we get
>>> the same 1049 flag errors we get with -t soft plus two additional ones:
>>>
>>> +A 0xffb00000, expected: 0x7fa00000, returned: 0x7fa00000, \
>>>   expected exceptions: i, returned: none
>>> +error: flags mismatch for input @ ibm/Basic-Types-Inputs.fptest:382:
>>> +b32A =0 S -> S i
>>
>> That's Abs of an SNaN; the test expects Invalid, which is wrong,
>> because IEEE754 says absolute-value is a "quiet-computational
>> operation" that never signals an exception.
>>
>> What's odd is that we don't report that error for the softfloat
>> implementation! I also don't understand why the expected value
>> isn't just the input value with the sign bit flipped.
> 
> With -t soft we don't handle "abs" and we don't get the error -- we get
> a "not handled" instead.
> Is there a function that we could use for abs? The only ones I've seen
> are floatX_abs() which mask out the sign bit and do nothing else.

Both abs and neg are pure bit operations.  So, yes, floatX_abs (and floatX_chs)
are the softfloat functions for these.

And if fp-test thinks Invalid should be raised, it's wrong.

>>> (...)
>>> +cff 0xffb00000, expected: 0x7ff8000000000000, returned: 
>>> 0x7ff4000000000000, \
>>>   expected exceptions: i, returned: none
>>> +error: flags mismatch for input @ ibm/Basic-Types-Inputs.fptest:26170:
>>> +b32b64cff =0 S -> Q i
>>
>> SNaN conversion from 32 bit to 64 bit. Here I agree
>> with the test -- we should quieten the NaN and raise
>> Invalid -- which implies that the hardware is wrong ?!?
> 
> This passes on an Intel host, and fails on both Power7 and 8 hosts I have
> access to. I don't have the Power ISA spec in front of me, but I hope
> there's something about this specified in it.

IIRC this is unspecified and does vary by implementation.


r~