[OE-core] [PATCH] ARMv8-A: Add tune for AArch32 state and AArch64 state

[Andre McCurdy]

On Wed, Mar 30, 2016 at 11:25 AM, Khem Raj <raj.khem at gmail.com> wrote:
>
>> On Mar 30, 2016, at 11:10 AM, Andre McCurdy <armccurdy at gmail.com> wrote:
>>
>> On Wed, Mar 30, 2016 at 9:36 AM, Khem Raj <raj.khem at gmail.com> wrote:
>>>
>>>> On Mar 30, 2016, at 9:03 AM, Dragomir Daniel <daniel.dragomir at windriver.com> wrote:
>>>>
>>>> On 03/29/2016 07:07 PM, Khem Raj wrote:
>>>>>> On Mar 28, 2016, at 3:10 PM, Phil Blundell <pb at pbcl.net> wrote:
>>>>>>
>>>>>> On Mon, 2016-03-28 at 18:29 +0300, Daniel Dragomir wrote:
>>>>>>> For AArch64 state, the default tune is aarch32 and include which
>>>>>>>   include just the aarch64 feature.
>>>>>> I don't really understand what this sentence is trying to say.  Can you
>>>>>> re-phrase it so as to be more accessible to non-experts?
>>>>>>
>>>>>>> meta/conf/machine/include/arm/arch-arm64.inc       |  36 -------
>>>>>>> meta/conf/machine/include/arm/arch-armv8.inc       |   1 -
>>>>>> If you are deleting existing files, please mention that in the commit
>>>>>> message.  And in particular, if you are removing a file that supports
>>>>>> generic ARMv8 (if imperfectly) and replacing it with something that is
>>>>>> specific to ARMv8-A, please explain why this is a good thing.
>>>>>>
>>>>>>> +TUNEVALID[crc] = "Enable CRC instructions for ARMv8-a"
>>>>>>> +ARMPKGSFX_FPU .= "${@bb.utils.contains('TUNE_FEATURES', 'crc', '-crc', '', d)}"
>>>>>>> +ARMPKGSFX_FPU_64 = "${@bb.utils.contains('TUNE_FEATURES', 'crc', '-crc', '', d)}"
>>>>>> Why is this involved with ARMPKGSFX_FPU?  The crc instructions are not
>>>>>> related to the fpu.  Also, the fact that you need to duplicate both this
>>>>>> and the crypto one for both FPU and FPU_64 seems like an indication that
>>>>>> something elsewhere is misdesigned.
>>>>>>
>>>>>>> +TUNEVALID[fp-armv8] = "Enable ARMv8 Vector Floating Point unit."
>>>>>>> +ARMPKGSFX_FPU .= "${@bb.utils.contains('TUNE_FEATURES', 'fp-armv8', '-fp-armv8', '', d)}"
>>>>>> I don't entirely understand why this one doesn't have an FPU_64
>>>>>> equivalent.  Are you always enabling vfp for A64?
>>>>>>
>>>>>>> +MACHINEOVERRIDES =. "${@bb.utils.contains('TUNE_FEATURES', 'aarch32', 'aarch32:', '' ,d)}"
>>>>>>> +MACHINEOVERRIDES .= "${@bb.utils.contains('TUNE_FEATURES', 'aarch64', ':aarch64', '' ,d)}"
>>>>>> As previously discussed, I am not wild about "aarch32" as the name of an
>>>>>> override which really means "ARMv8 in AArch32 state".  I know there is a
>>>>>> school of thought that says that the execution state of older cpus is
>>>>>> not AArch32 even if in practice it is indistinguishable, but this
>>>>>> doesn't seem to match either the expectations that a rational user of OE
>>>>>> is likely to have, or the information in the published literature.
>>>>> Don’t disagree with technical argument I said that previously too. All I am trying
>>>>> to say is lets take this opportunity to simplify arm tunes starting with armv8
>>>>> we have this opportunity and what you suggest will keep the thread alive with
>>>>> prior tunes. With armv8 we should match what we do for x86 and x86_64 ideally.
>>>>
>>>> We'll need to take a decision on this and continue the work on tunes.
>>>>
>>>> What is the best way to name the armv8a tunes:
>>>> aarch32 and aarch64 or armv8a32 and armv8a64?
>>>
>>> We are catering armv8 here and not 64bit or 32bit arm. Thats the difference.
>>> may be those could be added in a common file?
>>> its fine to call them armv8a32 and armv8a64 it doesnt matter
>>> but break the inclusion of .inc files that will drag it
>>> into pre-armv8 world and define a consistent 32bit tunes
>>> such that when we think armv8, we get three options
>>> aarch64, aarch32, aarch32t2,
>>
>> All armv8a CPUs support Thumb2, so there's no need to distinguish
>> between aarch32 and aarch32t2.
>
> I was more talking about execution states than packaging arches.

Either way it shouldn't be a concern for the CPU tuning files.
Building Thumb2 for an armv8a CPU is really just an extension of
"optimise for size" and we don't include -Os, -O2, etc options in the
CPU tuning files.

>> We've made that mistake historically for armv6 and armv7 but we
>> shouldn't keep carrying it forward.
>>
>>> all the vfp options
>>> may be folder into these three and also neon and others
>>
>> For most cores from Cortex A9 onwards NEON and VFP are optional, so
>> hardcoded assumptions won't work.
>
> Is it really ‘onwards’ or just cortexa9 was an exception. Can you
> point to some documentation on it ?

Nothing very concrete. The ARM Cortex-A Series Programmer’s Guide for
ARMv8-A mentions:

"Both floating-point and NEON are required in all standard ARMv8
implementations. However, implementations targeting specialized
markets may support the following combinations:

 - No NEON or floating-point.
 - Full floating-point and SIMD support with exception trapping.
 - Full floating-point and SIMD support without exception trapping.

http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.den0024a/CJHECGIH.html

>>
>>> Can we reach that somehow ?
>>>
>>>
>>>>
>>>> Who else need to express his opinion about this and make the decision?
>>>>
>>>> Daniel
>>>>
>>>>>
>>>>>>> +ARMPKGARCH_tune-aarch64            ?= "aarch64"
>>>>>> This seems a tiny bit short-sighted since presumably there will be an
>>>>>> ARMv9 (or ARMv8.2) at some point.  What will ARMPKGARCH for A64 be then?
>>>>>>
>>>>>>> +BASE_LIB_tune-aarch64            = "lib64"
>>>>>> This seems like more a matter of distro policy and I'm not sure it
>>>>>> belongs in a tune file.  If it does then can't you rely on the aarch64
>>>>>> MACHINEOVERRIDE and just write "BASE_LIB_aarch64" rather than having to
>>>>>> duplicate this for all four of the tunes (and the -be equivalents)?
>>>>>>
>>>>>>> +# Duplicated from arch-arm.inc
>>>>>> Please add a comment explaining why you can't include arch-arm.inc.
>>>>>>
>>>>>> p.
>>>>>>
>>>>>>
>>>>>> --
>>>>>> _______________________________________________
>>>>>> Openembedded-core mailing list
>>>>>> Openembedded-core at lists.openembedded.org
>>>>>> http://lists.openembedded.org/mailman/listinfo/openembedded-core
>>>>
>>>
>>>
>>> --
>>> _______________________________________________
>>> Openembedded-core mailing list
>>> Openembedded-core at lists.openembedded.org
>>> http://lists.openembedded.org/mailman/listinfo/openembedded-core
>>>
>