Re: [PATCH] x86: make better use of VBROADCASTSS / VPBROADCASTD

[Jan Beulich <jbeulich@suse.com>]

On 15.06.2023 07:23, Hongtao Liu wrote:
> On Wed, Jun 14, 2023 at 5:03 PM Jan Beulich <jbeulich@suse.com> wrote:
>>
>> On 14.06.2023 09:41, Hongtao Liu wrote:
>>> On Wed, Jun 14, 2023 at 1:58 PM Jan Beulich via Gcc-patches
>>> <gcc-patches@gcc.gnu.org> wrote:
>>>>
>>>> ... in vec_dupv4sf / *vec_dupv4si. The respective broadcast insns are
>>>> never longer (yet sometimes shorter) than the corresponding VSHUFPS /
>>>> VPSHUFD, due to the immediate operand of the shuffle insns balancing the
>>>> need for VEX3 in the broadcast ones. When EVEX encoding is required the
>>>> broadcast insns are always shorter.
>>>>
>>>> Add two new alternatives each, one covering the AVX2 case and one
>>>> covering AVX512.
>>> I think you can just change assemble output for this first alternative
>>> when TARGET_AVX2, use vbroadcastss, else use vshufps since
>>> vbroadcastss only accept register operand when TARGET_AVX2. And no
>>> need to support 2 extra alternatives which doesn't make sense just
>>> make RA more confused about the same meaning of different
>>> alternatives.
>>
>> You mean by switching from "@ ..." to C code using "switch
>> (which_alternative)"? I can do that, sure. Yet that'll make for a
>> more complicated "length_immediate" attribute then. Would be nice
> Yes, you can also do something like
>    (set (attr "length_immediate")
>      (cond [(eq_attr "alternative" "0")
>                    (if_then_else (match_test "TARGET_AVX2)
>                     (const_string "")
>                    (const_string "1"))
>                 ...]

Yes, that's along the lines of what I was thinking of. I'm uncertain
about one aspect of what you spelled out above, though: What is the
meaning of the empty string in (const_string "")? Shouldn't this be
"0" or "*"?

>> But that'll be for vec_dupv4sf only, as vec_dupv4si is subtly
>> different.
> Yes, but can we use vpbroadcastd for vec_dupv4si similarly?

Well, the use there is similar, but the folding with the shuffle
alternative won't be possible, because of the new first alternative
also allowing m for the source, when the shuffle one allows for only
Yv. The extra m is pointless to have in vec_dupv4sf (because a later
alternative with a wider ISA [avx] has it already), while in
vec_dupv4si the similar later alternative resolves to vbroadcastss,
not vpbroadcastd. I should be able to fold the two vpbroadcastd
alternatives, along the lines of what I've done in the vec_dupv2di
patch just sent. (As I just realized the m in what are alternatives
1 each in patch v1 is pointless, since already taken care of by
other alternatives.)

Jan