Re: User directed Function Multiversioning via Function Overloading (issue5752064)

[Sriraman Tallam <tmsriram at google dot com>]

On Wed, May 2, 2012 at 11:04 AM, H.J. Lu <hjl.tools@gmail.com> wrote:
> On Wed, May 2, 2012 at 10:44 AM, Sriraman Tallam <tmsriram@google.com> wrote:
>>>>>
>>>>> 1. ?Since AVX > SSE4 > SSSE3 > SSE3 > SSE2 > SSE, with
>>>>> foo for AVX and SSE3, on AVX processors, which foo will be
>>>>> selected?
>>>>
>>>> foo for AVX will get called since that appears ahead.
>>>>
>>>> The dispatching is done in the same order in which the functions are
>>>> specified. If, potentially, two foo versions can be dispatched for an
>>>> architecture, the first foo will get called. ?There is no way right
>>>> now to specify the order in which the dispatching should be done.
>>>
>>> This is very fragile. ?We know ISAs and processors. ?The source
>>> order should be irrelevant.
>>
>> I am not sure it is always possible keep this dispatching unambiguous
>> to the user. It might be better to let the user specify a priority for
>> each version to control the order of dispatching.
>>
>> ?Still, one way to implement what you said is to assign a significance
>> number to each ISA, where the number of sse4 > sse, for instance.
>> Then, the dispatching can be done in the descending order of
>> significance. What do you think?
>
> This sounds reasonable. ?You should also take processor into
> account when doing this.
>
>> I thought about this earlier and I was thinking along the lines of
>> letting the user specify a priority for each version, when there is
>> ambiguity.
>>
>>>
>>>>
>>>>> 2. ?I don't see any tests for __builtin_cpu_supports ("XXX")
>>>>> nor __builtin_cpu_is ("XXX"). ?I think you need tests for
>>>>> them.
>>>>
>>>> This is already there as part of the previous CPU detection patch that
>>>> was submitted. Please see gcc.target/i386/builtin_target.c. Did you
>>>> want something else?
>>>
>>> gcc.target/i386/builtin_target.c doesn't test if __builtin_cpu_supports ("XXX")
>>> and __builtin_cpu_is ("XXX") are implemented correctly.
>>
>> Oh, you mean like doing a CPUID again in the test case itself and checking, ok.
>>
>
> Yes. BTW, ?I think you should also add FMA support to
> config/i386/i386-cpuinfo.c.

I am preparing a patch for this. I will send it your way soon enough.

Thanks,
-Sri.

>
> Thanks.
>
> --
> H.J.