Re: [PATCH 4/4] x86: provide a 128-bit VBROADCASTSD pseudo

[Jan Beulich <jbeulich@suse.com>]

On 16.06.2023 18:59, H.J. Lu wrote:
> On Fri, Jun 16, 2023 at 12:32 AM Jan Beulich <jbeulich@suse.com> wrote:
>> VBROADCASTSD not supporting 128-bit destinations in any of their AVX,
>> AVX2, or AVX512F incarnations is presumably because of VMOVDDUP
>> precisely supporting this very operation. (It is therefore different
>> from e.g. VPBROADCASTQ, which has no exact equivalent.) Still its
>> absence has led to people using VPBROADCASTQ as substitution; this could
>> have been avoided if such a pseudo had been supported from the very
>> beginning.
>>
>> Note that the pseudos try to match what the real instructions would have
>> used as closely as possible, i.e. VexW0 instead of VexWIG for the AVX
>> and AVX2 forms as well as AVX2 in the first place for the register
>> source form.
>> ---
>> For being the first example of us supplying such, this is partly RFC. On
>> top of that a question is also whether to indeed have split AVX/AVX2
>> templates, when in principle one (allowing for both memory and register
>> source) could do.
>>
> 
> I don't think assembler should invent such instructions.

May I ask about the "why" behind this? If such a pseudo had been there
from the beginning, an admittedly minor mistake like that corrected by
gcc commit a4df0ce78d6f likely wouldn't have been made, because no
special casing of V2DFmode would have been necessary in the first place.

Jan