From: Charles Baylis <charles.baylis@linaro.org>
To: Michael Collison <michael.collison@linaro.org>
Cc: Kyrill Tkachov <kyrylo.tkachov@arm.com>,
GCC Patches <gcc-patches@gcc.gnu.org>,
Ramana Radhakrishnan <Ramana.Radhakrishnan@arm.com>
Subject: Re: [ARM] Use vector wide add for mixed-mode adds
Date: Wed, 21 Oct 2015 15:14:00 -0000 [thread overview]
Message-ID: <CADnVucDr3e3Vgd5uCpKrz49HJi+ZoGBAoVwTErN5kQMYs4LxKg@mail.gmail.com> (raw)
In-Reply-To: <5625F31B.5060203@linaro.org>
On 20 October 2015 at 08:54, Michael Collison
<michael.collison@linaro.org> wrote:
> I want to ask a question about existing patterns in neon.md that utilize the
> vec_select and all the lanes as my example does: Why are the following
> pattern not matched if the target is big endian?
> (define_insn "neon_vec_unpack<US>_lo_<mode>"
> [(set (match_operand:<V_unpack> 0 "register_operand" "=w")
> (SE:<V_unpack> (vec_select:<V_HALF>
> (match_operand:VU 1 "register_operand" "w")
> (match_operand:VU 2 "vect_par_constant_low" ""))))]
> "TARGET_NEON && !BYTES_BIG_ENDIAN"
> "vmovl.<US><V_sz_elem> %q0, %e1"
> [(set_attr "type" "neon_shift_imm_long")]
> )
>
> (define_insn "neon_vec_unpack<US>_hi_<mode>"
> [(set (match_operand:<V_unpack> 0 "register_operand" "=w")
> (SE:<V_unpack> (vec_select:<V_HALF>
> (match_operand:VU 1 "register_operand" "w")
> (match_operand:VU 2 "vect_par_constant_high" ""))))]
> "TARGET_NEON && !BYTES_BIG_ENDIAN"
> "vmovl.<US><V_sz_elem> %q0, %f1"
> [(set_attr "type" "neon_shift_imm_long")]
>
> These patterns are similar to the new patterns I am adding and I am
> wondering if my patterns should exclude BYTES_BIG_ENDIAN?
These patterns use %e and %f to access the low and high part of the
input operand - so %e is used to match the use of _lo in the pattern
name, and vect_par_constant_low, and %f with _hi and
vect_par_constant_high. For big-endian, the use of %e and %f would
need to be swapped.
Looking at the patch you posted last month (possibly not the latest version?):
This is a pattern which is supposed to act on the low part of the
input vector, hence _lo in the name:
+(define_insn "vec_sel_widen_ssum_lo<VQI:mode><VW:mode>3"
+ [(set (match_operand:<VW:V_widen> 0 "s_register_operand" "=w")
+ (plus:<VW:V_widen> (sign_extend:<VW:V_widen> (vec_select:VW
(match_operand:VQI 1 "s_register_operand" "%w")
+ (match_operand:VQI 2 "vect_par_constant_low" "")))
+ (match_operand:<VW:V_widen> 3 "s_register_operand" "0")))]
+ "TARGET_NEON"
+ "vaddw.<V_s_elem>\t%q0, %q3, %e1"
Here, using %e1 carries an implicit assumption that the low part of
the input vector is in the lowest numbered of the pair of D registers,
which is only true on little-endian.
This is a bit ugly (and untested) but perhaps something like this
would fix the problem
{
return BYTES_BIG_ENDIAN ? "vaddw.<V_s_elem>\t%q0, %q3, %f1" :
"vaddw.<V_s_elem>\t%q0, %q3, %e1";
}
+ [(set_attr "type" "neon_add_widen")
+ (set_attr "length" "8")]
+)
Similarly, here. Pattern is _hi, register is %f1:
+(define_insn "vec_sel_widen_ssum_hi<VQI:mode><VW:mode>3"
+ [(set (match_operand:<VW:V_widen> 0 "s_register_operand" "=w")
+ (plus:<VW:V_widen> (sign_extend:<VW:V_widen> (vec_select:VW
(match_operand:VQI 1 "s_register_operand" "%w")
+ (match_operand:VQI 2 "vect_par_constant_high" "")))
+ (match_operand:<VW:V_widen> 3 "s_register_operand" "0")))]
+ "TARGET_NEON"
+ "vaddw.<V_s_elem>\t%q0, %q3, %f1"
+ [(set_attr "type" "neon_add_widen")
+ (set_attr "length" "8")]
+)
However, as far as I can see, there isn't an endianness dependency in
widen_ssum<mode>3/widen_usum<mode>3 because both halves of the vector
are used and added together.
Hope this helps
Charles
next prev parent reply other threads:[~2015-10-21 15:05 UTC|newest]
Thread overview: 15+ messages / expand[flat|nested] mbox.gz Atom feed top
2015-09-23 2:40 Michael Collison
2015-09-23 8:59 ` Kyrill Tkachov
2015-10-01 10:05 ` Michael Collison
2015-10-08 11:02 ` Kyrill Tkachov
2015-10-20 8:11 ` Michael Collison
2015-10-21 15:14 ` Charles Baylis [this message]
-- strict thread matches above, loose matches on Subject: below --
2015-11-30 6:59 Michael Collison
2015-12-10 15:09 ` Kyrill Tkachov
2015-12-17 0:02 ` Michael Collison
2016-02-09 16:27 ` Kyrill Tkachov
2016-02-15 6:32 ` Michael Collison
2015-08-18 8:02 Michael Collison
2015-08-18 13:46 ` Ramana Radhakrishnan
2015-08-23 4:16 ` Michael Collison
2015-08-24 8:37 ` Ramana Radhakrishnan
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=CADnVucDr3e3Vgd5uCpKrz49HJi+ZoGBAoVwTErN5kQMYs4LxKg@mail.gmail.com \
--to=charles.baylis@linaro.org \
--cc=Ramana.Radhakrishnan@arm.com \
--cc=gcc-patches@gcc.gnu.org \
--cc=kyrylo.tkachov@arm.com \
--cc=michael.collison@linaro.org \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).