From: Evandro Menezes <e.menezes@samsung.com>
To: James Greenhalgh <james.greenhalgh@arm.com>
Cc: Wilco Dijkstra <Wilco.Dijkstra@arm.com>,
"gcc-patches@gcc.gnu.org" <gcc-patches@gcc.gnu.org>,
nd <nd@arm.com>, Marcus Shawcroft <Marcus.Shawcroft@arm.com>,
Kyrylo Tkachov <Kyrylo.Tkachov@arm.com>,
richard.earnshaw@arm.com
Subject: Re: [PATCH][AArch64] Replace insn to zero up DF register
Date: Thu, 10 Mar 2016 16:27:00 -0000 [thread overview]
Message-ID: <56E1A061.5080901@samsung.com> (raw)
In-Reply-To: <20160310132327.GA38844@arm.com>
On 03/10/16 07:23, James Greenhalgh wrote:
> On Wed, Mar 09, 2016 at 03:35:43PM -0600, Evandro Menezes wrote:
>> On 03/01/16 13:08, Evandro Menezes wrote:
>>> On 03/01/16 13:02, Wilco Dijkstra wrote:
>>>> Evandro Menezes wrote:
>>>>> The meaning of these attributes are not clear to me. Is there a
>>>>> reference somewhere about which insns are FP or SIMD or neither?
>>>> The meaning should be clear, "fp" is a floating point
>>>> instruction, "simd" a SIMD one
>>>> as defined in ARM-ARM.
>>>>
>>>>> Indeed, I had to add the Y for the f_mcr insn to match it with nosimd.
>>>>> However, I didn't feel that it should be moved to the right, since it's
>>>>> already disparaged. Am I missing something detail?
>>>> It might not matter for this specific case, but I have seen
>>>> reload forcing the very
>>>> first alternative without looking at any costs or preferences -
>>>> as long as it is legal.
>>>> This suggests we need to order alternatives from most preferred
>>>> alternative to least
>>>> preferred one.
>>>>
>>>> I think it is good enough for commit, James?
>>> Methinks that my issue with those attributes is that I'm not as
>>> fluent in AArch64 as I'd like to be.
>>>
>>> Please, feel free to edit the patch changing the order then.
>> Replace insn to zero up SIMD registers
>>
>> gcc/
>> * config/aarch64/aarch64.md
>> (*movhf_aarch64): Add "movi %0, #0" to zero up register.
>> (*movsf_aarch64): Likewise and add "simd" and "fp" attributes.
>> (*movdf_aarch64): Likewise.
>>
>> Swapped the order of the constraints to favor MOVI.
>>
>> Just say the word...
> I'm wondering whether this is appropriate for GCC6 now that we are so late
> in the development cycle. Additionally, I have some comments on your patch:
>
>> diff --git a/gcc/config/aarch64/aarch64.md b/gcc/config/aarch64/aarch64.md
>> index 68676c9..4502a58 100644
>> --- a/gcc/config/aarch64/aarch64.md
>> +++ b/gcc/config/aarch64/aarch64.md
>> @@ -1163,11 +1163,12 @@
>> )
>>
>> (define_insn "*movhf_aarch64"
>> - [(set (match_operand:HF 0 "nonimmediate_operand" "=w, ?r,w,w,m,r,m ,r")
>> - (match_operand:HF 1 "general_operand" "?rY, w,w,m,w,m,rY,r"))]
>> + [(set (match_operand:HF 0 "nonimmediate_operand" "=w,w ,?r,w,w,m,r,m ,r")
>> + (match_operand:HF 1 "general_operand" "Y ,?rY, w,w,m,w,m,rY,r"))]
>> "TARGET_FLOAT && (register_operand (operands[0], HFmode)
>> || aarch64_reg_or_fp_zero (operands[1], HFmode))"
>> "@
>> + movi\\t%0.4h, #0
>> mov\\t%0.h[0], %w1
>> umov\\t%w0, %1.h[0]
>> mov\\t%0.h[0], %1.h[0]
>> @@ -1176,18 +1177,19 @@
>> ldrh\\t%w0, %1
>> strh\\t%w1, %0
>> mov\\t%w0, %w1"
>> - [(set_attr "type" "neon_from_gp,neon_to_gp,neon_move,\
>> + [(set_attr "type" "neon_move,neon_from_gp,neon_to_gp,neon_move,\
>> f_loads,f_stores,load1,store1,mov_reg")
>> - (set_attr "simd" "yes,yes,yes,*,*,*,*,*")
>> - (set_attr "fp" "*,*,*,yes,yes,*,*,*")]
>> + (set_attr "simd" "yes,yes,yes,yes,*,*,*,*,*")
>> + (set_attr "fp" "*,*,*,*,yes,yes,*,*,*")]
>> )
>>
>> (define_insn "*movsf_aarch64"
>> - [(set (match_operand:SF 0 "nonimmediate_operand" "=w, ?r,w,w ,w,m,r,m ,r")
>> - (match_operand:SF 1 "general_operand" "?rY, w,w,Ufc,m,w,m,rY,r"))]
>> + [(set (match_operand:SF 0 "nonimmediate_operand" "=w,w ,?r,w,w ,w,m,r,m ,r")
>> + (match_operand:SF 1 "general_operand" "Y ,?rY, w,w,Ufc,m,w,m,rY,r"))]
>> "TARGET_FLOAT && (register_operand (operands[0], SFmode)
>> || aarch64_reg_or_fp_zero (operands[1], SFmode))"
>> "@
>> + movi\\t%0.2s, #0
>> fmov\\t%s0, %w1
>> fmov\\t%w0, %s1
>> fmov\\t%s0, %s1
>> @@ -1197,16 +1199,19 @@
>> ldr\\t%w0, %1
>> str\\t%w1, %0
>> mov\\t%w0, %w1"
>> - [(set_attr "type" "f_mcr,f_mrc,fmov,fconsts,\
>> - f_loads,f_stores,load1,store1,mov_reg")]
>> + [(set_attr "type" "neon_move,f_mcr,f_mrc,fmov,fconsts,\
>> + f_loads,f_stores,load1,store1,mov_reg")
>> + (set_attr "simd" "yes,*,*,*,*,*,*,*,*,*")
>> + (set_attr "fp" "*,*,*,yes,yes,yes,yes,*,*,*")]
>> )
> This fp attribute looks wrong to me. The two fmov instructions that move
> between core and FP registers should be tagged "yes". However, this is
> irrelevant as the whole pattern is guarded by TARGET_FLOAT.
>
> It would be clearer to drop the FP attribute entirely, so as not to give
> the erroneous impression that some alternatives in this insn are enabled
> for !TARGET_FLOAT.
You mean to remove the FP attribute from all, HF, SF, DF, TF?
>> (define_insn "*movdf_aarch64"
>> - [(set (match_operand:DF 0 "nonimmediate_operand" "=w, ?r,w,w ,w,m,r,m ,r")
>> - (match_operand:DF 1 "general_operand" "?rY, w,w,Ufc,m,w,m,rY,r"))]
>> + [(set (match_operand:DF 0 "nonimmediate_operand" "=w,w ,?r,w,w ,w,m,r,m ,r")
>> + (match_operand:DF 1 "general_operand" "Y ,?rY, w,w,Ufc,m,w,m,rY,r"))]
>> "TARGET_FLOAT && (register_operand (operands[0], DFmode)
>> || aarch64_reg_or_fp_zero (operands[1], DFmode))"
>> "@
>> + movi\\t%d0, #0
>> fmov\\t%d0, %x1
>> fmov\\t%x0, %d1
>> fmov\\t%d0, %d1
>> @@ -1216,8 +1221,10 @@
>> ldr\\t%x0, %1
>> str\\t%x1, %0
>> mov\\t%x0, %x1"
>> - [(set_attr "type" "f_mcr,f_mrc,fmov,fconstd,\
>> - f_loadd,f_stored,load1,store1,mov_reg")]
>> + [(set_attr "type" "neon_move,f_mcr,f_mrc,fmov,fconstd,\
>> + f_loadd,f_stored,load1,store1,mov_reg")
>> + (set_attr "simd" "yes,*,*,*,*,*,*,*,*,*")
>> + (set_attr "fp" "*,*,*,yes,yes,yes,yes,*,*,*")]
> Likewise here.
>
> As to whether this is OK for GCC6, this doesn't look like a regression to me,
> we've generated FMOV since the dawn of the port. The patch has performance
> implications (arguably trivial ones, but still performance implications)
> for generic code generation, and release is coming soon - so I'd rather
> keep this back until GCC 7 opens.
>
> Remove the FP attributes, and this patch is OK for GCC 7.
>
> I'm happy to be overruled on this by Marcus or Richard if they feel we
> should take this patch now, but my opinion is that waiting is more in the
> spirit of using Stage 4 to stabilise the compiler.
I agree to postpone until GCC 7.
[AArch64] Replace insn to zero up SIMD registers
gcc/
* config/aarch64/aarch64.md
(*movhf_aarch64): Add "movi %0, #0" to zero up register.
(*movsf_aarch64): Likewise and add "simd" attributes.
(*movdf_aarch64): Likewise.
This patch removes the FP attributes from the HF, SF, DF, TF moves.
Thank you,
--
Evandro Menezes
next prev parent reply other threads:[~2016-03-10 16:27 UTC|newest]
Thread overview: 27+ messages / expand[flat|nested] mbox.gz Atom feed top
2016-01-22 13:52 Wilco Dijkstra
2016-01-27 23:14 ` Evandro Menezes
2016-02-26 12:37 ` Wilco Dijkstra
2016-02-26 22:43 ` Evandro Menezes
2016-02-29 18:07 ` Wilco Dijkstra
2016-02-29 23:11 ` Evandro Menezes
2016-03-01 19:02 ` Wilco Dijkstra
2016-03-01 19:08 ` Evandro Menezes
2016-03-09 21:36 ` Evandro Menezes
2016-03-10 13:23 ` James Greenhalgh
2016-03-10 16:27 ` Evandro Menezes [this message]
2016-03-10 16:32 ` Evandro Menezes
2016-03-10 16:37 ` James Greenhalgh
2016-04-25 19:20 ` [PATCH][AArch64] Replace insn to zero up SIMD registers Evandro Menezes
2016-04-26 13:25 ` Wilco Dijkstra
2016-04-27 19:55 ` Evandro Menezes
-- strict thread matches above, loose matches on Subject: below --
2015-10-19 23:41 [PATCH][AArch64] Replace insn to zero up DF register Evandro Menezes
2015-10-19 23:52 ` Andrew Pinski
2015-10-20 0:33 ` Andrew Pinski
2015-10-20 14:46 ` Andrew Pinski
2015-10-28 18:49 ` Evandro Menezes
2015-10-30 10:26 ` Marcus Shawcroft
2015-11-09 22:59 ` Evandro Menezes
2015-12-03 21:01 ` Evandro Menezes
2015-11-19 22:01 ` Evandro Menezes
2015-12-16 21:30 ` Evandro Menezes
2016-01-13 0:06 ` Evandro Menezes
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=56E1A061.5080901@samsung.com \
--to=e.menezes@samsung.com \
--cc=Kyrylo.Tkachov@arm.com \
--cc=Marcus.Shawcroft@arm.com \
--cc=Wilco.Dijkstra@arm.com \
--cc=gcc-patches@gcc.gnu.org \
--cc=james.greenhalgh@arm.com \
--cc=nd@arm.com \
--cc=richard.earnshaw@arm.com \
/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).