getauxv and ARMv7 platform detection

getauxv and ARMv7 platform detection

[Jeffrey Walton]

Hi Everyone,

I just noticed the following code was failing on ARMv7 platforms:

// For older machines, like GCC 4.3 on ARMv6
#ifndef HWCAP_ARMv7
# define HWCAP_ARMv7 (1 << 29)
#endif

inline bool CPU_QueryARMv7()
{
    if (getauxval(AT_HWCAP) & HWCAP_ARMv7) != 0)
        return true;
    return false;
}

Looking through Torvalds GitHub [1] I don't see a reliable define for
HWCAP_ARMv7 in hwcap.h.

I know I can fallback to NEON, but that misses earlier devices that
are ARMv7 but lacks NEON.

Searching is turning up a lot of spurious noise (for Aarch64/ARMv8),
and one hit for Mozilla working around the same problem [2].

My question is, how do we use getauxval to detect ARMv7 platforms?

Thanks in advance.

[1] https://github.com/torvalds/linux/blob/master/arch/arm/include/uapi/asm/hwcap.h
[2] https://hg.mozilla.org/releases/mozilla-aurora/rev/522951ff7046fd2bdf3916c973fbfde27be0af31

Re: getauxv and ARMv7 platform detection

[Florian Weimer]

* Jeffrey Walton:

> My question is, how do we use getauxval to detect ARMv7 platforms?

What is the actual problem you are trying to solve?

Thanks,
Florian

Re: getauxv and ARMv7 platform detection

[Jeffrey Walton]

On Tue, May 21, 2019 at 2:03 AM Florian Weimer <fweimer@redhat.com> wrote:
>
> * Jeffrey Walton:
>
> > My question is, how do we use getauxval to detect ARMv7 platforms?
>
> What is the actual problem you are trying to solve?

I'm trying to detect if the platform is ARMv7 at runtime.

Jeff

Re: getauxv and ARMv7 platform detection

[Florian Weimer]

* Jeffrey Walton:

> On Tue, May 21, 2019 at 2:03 AM Florian Weimer <fweimer@redhat.com> wrote:
>>
>> * Jeffrey Walton:
>>
>> > My question is, how do we use getauxval to detect ARMv7 platforms?
>>
>> What is the actual problem you are trying to solve?
>
> I'm trying to detect if the platform is ARMv7 at runtime.

Yes, but why?  Just to print a diagnostic?

Thanks,
Florian

Re: getauxv and ARMv7 platform detection

[Jeffrey Walton]

On Tue, May 21, 2019 at 2:31 AM Florian Weimer <fweimer@redhat.com> wrote:
>
> * Jeffrey Walton:
>
> > On Tue, May 21, 2019 at 2:03 AM Florian Weimer <fweimer@redhat.com> wrote:
> >>
> >> * Jeffrey Walton:
> >>
> >> > My question is, how do we use getauxval to detect ARMv7 platforms?
> >>
> >> What is the actual problem you are trying to solve?
> >
> > I'm trying to detect if the platform is ARMv7 at runtime.
>
> Yes, but why?  Just to print a diagnostic?

We switch to ARMv7 and NEON implementations at runtime if the CPU
supports it. We can usually achieve 2x or 3x speedups over
ARMv{4|5|6}.

We've got SIGILL probes in place but they are an expensive fallback.
We prefer a faster and cleaner solution on Linux like getauxv().

We don't control how the library is built. Users and distros do their
own things. About the only thing we can count on is, -march=native is
useless (and sometimes segfaults the compiler), -march=XXX is missing
and -mfpu=YYY is missing.

Jeff

Re: getauxv and ARMv7 platform detection

[Adhemerval Zanella]

On 21/05/2019 03:45, Jeffrey Walton wrote:
> On Tue, May 21, 2019 at 2:31 AM Florian Weimer <fweimer@redhat.com> wrote:
>>
>> * Jeffrey Walton:
>>
>>> On Tue, May 21, 2019 at 2:03 AM Florian Weimer <fweimer@redhat.com> wrote:
>>>>
>>>> * Jeffrey Walton:
>>>>
>>>>> My question is, how do we use getauxval to detect ARMv7 platforms?
>>>>
>>>> What is the actual problem you are trying to solve?
>>>
>>> I'm trying to detect if the platform is ARMv7 at runtime.
>>
>> Yes, but why?  Just to print a diagnostic?
> 
> We switch to ARMv7 and NEON implementations at runtime if the CPU
> supports it. We can usually achieve 2x or 3x speedups over
> ARMv{4|5|6}.
> 
> We've got SIGILL probes in place but they are an expensive fallback.
> We prefer a faster and cleaner solution on Linux like getauxv().
> 
> We don't control how the library is built. Users and distros do their
> own things. About the only thing we can count on is, -march=native is
> useless (and sometimes segfaults the compiler), -march=XXX is missing
> and -mfpu=YYY is missing.
> 

Unfortunately hwcap does not export this information, you will need to
parse /proc/cpuinfo 'CPU architecture' field and take some issues
in consideration.  Check google cpu_features project [1], it provides
such information through a library.

[1] https://github.com/google/cpu_features/blob/master/src/cpuinfo_arm.c