[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Qemu-devel] [PATCH v1 08/15] s390x/tcg: Handle SET FPC AND LOAD FPC

From: David Hildenbrand
Subject: Re: [Qemu-devel] [PATCH v1 08/15] s390x/tcg: Handle SET FPC AND LOAD FPC 3-bit BFP rounding modes
Date: Tue, 12 Feb 2019 20:32:47 +0100
User-agent: Mozilla/5.0 (X11; Linux x86_64; rv:60.0) Gecko/20100101 Thunderbird/60.4.0

On 12.02.19 20:07, Richard Henderson wrote:
> On 2/12/19 3:03 AM, David Hildenbrand wrote:
>> We already forward the 3 bits correctly in the translation functions. We
>> also have to handle them properly and check for specification
>> exceptions.
>> Setting an invalid rounding mode (BFP only, all DFP rounding modes)
>> results in a specification exception. Setting unassigned bits in the
>> fpc, results in a specification exception.
>> This fixes LOAD FPC (AND SIGNAL), SET FPC (AND SIGNAL). Also for,
>> SET BFP ROUNDING MODE, 3-bit rounding mode is now explicitly checked.
>> Notes:
>> 1. Use "float_round_to_zero" for now to handle "Round to prepare for
>>    shorter precision". Looking at the PoP "Summary of Rounding and Range
>>    Actions" for BFP. They differ when it comes to tiny values.
>> 2. TCG_CALL_NO_WG is required for sfpc handler, as we now inject
>>    exceptions.
>> We won't be modeling abscence of the "floating-point extension facility"
>> for now, not necessary as most take the facility for granted without
>> checking.
>> z14 PoP, 9-23, "LOAD FPC"
>>     When the floating-point extension facility is
>>     installed, bits 29-31 of the second operand must
>>     specify a valid BFP rounding mode and bits 6-7,
>>     14-15, 24, and 28 must be zero; otherwise, a
>>     specification exception is recognized.
> Reviewed-by: Richard Henderson <address@hidden>
>> +    /*
>> +     * FIXME: we actually want something like round_to_odd, but that does 
>> not
>> +     * support all data types yet.
>> +     */
>> +    float_round_to_zero,
> Yes, you want round_to_odd.  I suppose that's not valid for float128 right 
> now?


as well as


support it.

It's not implemented for round_canonical(), round_to_int(),
roundAndPackInt32(), roundAndPackInt64(), roundAndPackUint64(),
roundAndPackFloat32() (and roundAndPackFloatx80()).

I assume at least 32bit is missing. I can't judge if the other functions
are relevant (x80 clearly not).

> r~



David / dhildenb

reply via email to

[Prev in Thread] Current Thread [Next in Thread]