[Mesa-dev] [PATCH] gallivm: use fallback code for mul_hi with llvm >= 7.0

[Roland Scheidegger]

Am 16.10.19 um 01:20 schrieb Dave Airlie:
> On Fri, 30 Aug 2019 at 00:55, Roland Scheidegger <sroland at vmware.com> wrote:
>>
>> Am 29.08.19 um 15:05 schrieb Jose Fonseca:
>>> This change is
>>>
>>>   Reviewed-by: Jose Fonseca <jfonseca at vmware.com>
>>>
>>> Regarding follow up change, do you think the LLVM pattern is sane/doable?
>> Yes, should be doable and not too bad (I did not verify that what we're
>> doing doesn't actually get recognized, since it's theoretically possible
>> some other lowering could produce the pattern, although it seems unlikely).
>> I think though this code isn't hit a lot - it was once used by draw,
>> which is why I noticed the suboptimal code generated and added the
>> optimized version, but nowadays it's just for mulhi, so should be fairly
>> rare in practice.
>>
>>
>>>
>>> If not we should try ask them to reconsider relying strictly upon
>>> pattern matching.  I get the feeling upstream LLVM is throwing the baby
>>> with the water with these changes.  I do understand the advantages of
>>> moving away from vendor specific intrinsics, but I think that for things
>>> which have no natural representation on LLVM base IR, they should add a
>>> vendor-agnostic intrinsic, for example a new "/llvm.mulhi.*"  set of
>>> instrinsics/, as narrow pattern matching is bound to produce performance
>>> cliffs nobody will notice.
>> They did add new generic intrinsics for some things, but not this one
>> indeed.
>> I'm not exactly a big fan of this pattern matching in favor of
>> intrinsics neither, at least if the patterns are non-trivial...
> 
> Btw In working on something else, I found the padd and psub intrinsic
> paths seem to fail at least on LLVM 8.
You mean the signed sse2/avx2 intrinsics for psub/padd?
At a quick glance it seems like llvm 8 actually already removed both the
signed and unsigned intrinsics in the end (initially only the unsigned
versions were removed, which is a lot more noticeable as it is hit
basically in all tests). I think the llvm specific intrinsics should
work in llvm 8 too.
I'll test this...

Roland


> 
> I don't have an example test to show though.
> 
> Dave.
>