[Top][All Lists]

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

Re: [Qemu-ppc] [Qemu-devel] [PATCH] ppc: Three floating point fixes

From: Paul Clarke
Subject: Re: [Qemu-ppc] [Qemu-devel] [PATCH] ppc: Three floating point fixes
Date: Mon, 19 Aug 2019 12:13:34 -0500
User-agent: Mozilla/5.0 (X11; Linux x86_64; rv:60.0) Gecko/20100101 Thunderbird/60.8.0

On 8/19/19 1:44 AM, Aleksandar Markovic wrote:
> 19.08.2019. 08.30, "David Gibson" <address@hidden> је
> написао/ла:
>> On Sun, Aug 18, 2019 at 10:59:01PM +0200, Aleksandar Markovic wrote:
>>> 18.08.2019. 10.10, "Richard Henderson" <address@hidden> је
>>> написао/ла:
>>>> On 8/16/19 11:59 PM, Aleksandar Markovic wrote:
>>>>>> From: "Paul A. Clarke" <address@hidden>
>>>> ...
>>>>>>   ISA 3.0B has xscvdpspn leaving its result in word 1 of the target
>>>>> register,
>>>>>>   and mffprwz expecting its input to come from word 0 of the source
>>>>> register.
>>>>>>   This sequence fails with QEMU, as a shift is required between
> those
>>> two
>>>>>>   instructions.  However, the hardware splats the result to both
> word 0
>>>>> and
>>>>>>   word 1 of its output register, so the shift is not necessary.
>>>>>>   Expect a future revision of the ISA to specify this behavior.
>>>>> Hmmm... Isn't this a gcc bug (using undocumented hardware feature),
>>> given
>>>>> everything you said here?
>>>> The key here is "expect a future revision of the ISA to specify this
>>> behavior".
>>>> It's clearly within IBM's purview to adjust the specification to
> document
>>> a
>>>> previously undocumented hardware feature.
>>> By no means, yes, the key is in ISA documentation. But, the impression
> that
>>> full original commit message conveys is that the main reason for change
> is
>>> gcc behavior. If we accepted in general that gcc behavior determines
>>> behavior, I am afraid we would be on a very slippery slope - therefore I
>>> think the commit message (and possible code comment) should, in my
> opinion,
>>> mention ISA docs as the central reason for change. Paul, is there any
>>> tentative release date of the new ISA specification?
>> It's not really a question of gcc behaviour, it's a question of actual
>> cpu behaviour versus ISA document.  Which one qemu should follow is
>> somewhat debatable, but it sounds here like the ISA will be updated to
>> match the cpu, which weights it heavily in favour of mimicing the
>> actual cpu.
> This sounds right to me.

Thanks for the reviews and great discussion.

While not yet part of a published version of the ISA, I did find the behavior 
documented in the User's Manuals for the POWER8 and POWER9 processors:

"Public Documents"
- "POWER9 Processor User's Manual"
- "POWER8 Processor User's Manual for the SCM"

POWER9 Processor User's Manual 
4. Power Architecture Compliance
4.3 Floating-Point Processor (FP, VMX, and VSX)
4.3.7 Floating-Point Invalid Forms and Undefined Conditions

POWER8 Processor User's Manual for the Single-Chip Module
3. Power Architecture Compliance
3.2 Floating-Point Processor (FP, VMX, and VSX)
3.2.9 Floating-Point Invalid Forms and Undefined Conditions

In a bullet:
- VSX scalar convert from double-precision to single-precision (xscvdpsp, 
VSR[32:63] is set to VSR[0:31].

I have not confirmed when the new revision of the ISA will be published, but 
it's on somebody's "to do" list.

I will respin the patch as 3 independent patches, and include a reference to 
the above documents for the change under discussion here.  (The other two 
changes may take a bit more time, because like David, I find the FPU emulation 
code cryptic.  :-/  )

These issues were found while running Glibc's test suite for "math", and there 
are still a *LOT* of QEMU-only FAILs, so I may be back again with suggested 
fixes or questions.  :-)


reply via email to

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