Re: [qemu-s390x] [Qemu-devel] [PATCH v1 33/33] s390x/tcg: Implement VECTOR UNPACK *

[David Hildenbrand]

On 28.02.19 10:28, David Hildenbrand wrote:
> On 28.02.19 01:03, Richard Henderson wrote:
>> On 2/26/19 3:39 AM, David Hildenbrand wrote:
>>> Combine all variant in a single handler. As source and destination
>>> have different element sizes, we can't use gvec expansion. Expand
>>> manually. Also watch out for overlapping source and destination and
>>> use a temporary register in that case.
>>>
>>> Signed-off-by: David Hildenbrand <address@hidden>
>>> ---
>>>  target/s390x/insn-data.def      |  8 +++++++
>>>  target/s390x/translate_vx.inc.c | 41 +++++++++++++++++++++++++++++++++
>>>  2 files changed, 49 insertions(+)
>>
>> This works as is, so
>> Reviewed-by: Richard Henderson <address@hidden>
>>
>> But the same comment applies wrt iteration order and not needing a temporary.
>> High unpack can iterate backward, while low unpack can iterate forward, with 
>> no
>> lost data.
> 
> I'll fix that right away. I guess vector pack cannot be handled like this.
> 
> The only way to get rid of the temporary would be to load both elements
> from v2 and v3 and then writing the two (half sized) elements in v1.
> 
> I'll have a look.

Hmm, as v2 and v3 are handled concatenated it is not that easy. I am not
sure if we can handle this without a temporary vector.

I thought about packing them first interleaved

v2 = [v2e0, v2e1]
v3 = [v3e0, ve31]
v1 = [v2e0_packed, v3e0_packed, v2e1_packed, v3e1_packed]

And then restoring the right order

v1 = [v2e0_packed, v2e1_packed, v3e0_packed, v3e1_packed]

But than the second operation seems to be the problem. That shuffling
would have to be hard coded as far as I can see. (shuffling with MO_8 is
nasty -> 14 element shave to be exchanged, in my opinion needing
eventually 14 temporary variables)

Of course, we can also simply detect duplicates and if so, call into a
helper.

-- 

Thanks,

David / dhildenb