From: Kyrill Tkachov <kyrylo.tkachov@arm.com>
To: Bernd Schmidt <bschmidt@redhat.com>,
GCC Patches <gcc-patches@gcc.gnu.org>
Cc: Jeff Law <law@redhat.com>
Subject: Re: [PATCH][RTL-ree] PR rtl-optimization/68194: Restrict copy instruction in presence of conditional moves
Date: Tue, 17 Nov 2015 10:17:00 -0000 [thread overview]
Message-ID: <564AFEB9.7040703@arm.com> (raw)
In-Reply-To: <564AF80C.3010605@arm.com>
On 17/11/15 09:49, Kyrill Tkachov wrote:
>
> On 17/11/15 09:08, Kyrill Tkachov wrote:
>> Hi Bernd,
>>
>> On 16/11/15 18:40, Bernd Schmidt wrote:
>>> On 11/16/2015 03:07 PM, Kyrill Tkachov wrote:
>>>
>>>> I've explained in the comments in the patch what's going on but the
>>>> short version is trying to change the destination of a defining insn
>>>> that feeds into an extend insn is not valid if the defining insn
>>>> doesn't feed directly into the extend insn. In the ree pass the only
>>>> way this can happen is if there is an intermediate conditional move
>>>> that the pass tries to handle in a special way. An equivalent fix
>>>> would have been to check on that path (when copy_needed in
>>>> combine_reaching_defs is true) that the state->copies_list vector
>>>> (that contains the conditional move insns feeding into the extend
>>>> insn) is empty.
>>>
>>> I ran this through gdb, and I think I see what's going on. For reference, here's a comment from the source:
>>>
>>> /* Considering transformation of
>>> (set (reg1) (expression))
>>> ...
>>> (set (reg2) (any_extend (reg1)))
>>>
>>> into
>>>
>>> (set (reg2) (any_extend (expression)))
>>> (set (reg1) (reg2))
>>> ... */
>>>
>>> I was thinking that another possible fix would be to also check !reg_used_between_p for reg1 to ensure it's not used. I'm thinking this might be a little clearer - what is your opinion?
>>
>> Yes, I had considered that as well. It should be equivalent. I didn't use !reg_used_between_p because I thought
>> it'd be more expensive than checking reg_overlap_mentioned_p since we must iterate over a number of instructions
>> and call reg_overlap_mentioned_p on each one. But I suppose this case is rare enough that it wouldn't make any
>> measurable difference.
>
> Actually, I tried it out. And while a check reg_used_between_p fixed the testcase, it caused code quality regressions
> on aarch64. Seems it's too aggressive in restricting ree.
>
> I'll have a closer look.
Ok, so the testcases that regress code-quality-wise on aarch64 look like this before ree:
(insn 48 57 49 7 (set (reg:SI 7 x7)
(zero_extend:SI (mem:QI (plus:DI (reg/f:DI 1 x1)
(const_int 2 [0x2]))))))
(insn 49 48 52 7 (set (reg/v:SI 2 x2)
(reg:SI 7 x7)))
(insn 52 49 53 7 (set (reg:DI 8 x8)
(zero_extend:DI (reg:SI 7 x7))))
ree wants to transform this into:
(insn 48 57 296 7 (set (reg:DI 8 x8)
(zero_extend:DI (mem:QI (plus:DI (reg/f:DI 1 x1)
(const_int 2 [0x2]))))))
(insn 296 48 49 7 (set (reg:DI 7 x7)
(reg:DI 8 x8)))
(insn 49 296 53 7 (set (reg/v:SI 2 x2)
(reg:SI 7 x7)))
which is valid, but we reject that with the reg_used_between_p check because x7 is used in
the intermediate insn 49. Note that no conditional move is present here.
So, I think that the crucial element here is that the destination of the def_insn should
feed directly into the extend, and that is what my original patch was testing for.
So, I'd like to keep my original proposed patch as is, although I think I'll add a couple of
testcases from the duplicate PRs to it for the testsuite.
What do you think?
Thanks,
Kyrill
>
> Kyrill
>
>>
>> Would you prefer to use !reg_used_between_p here?
>>
>>>
>>> The added comment could lead to some confusion since it's placed in front of an existing if statement that also tests a different condition. Also, if we go with your fix,
>>>
>>>> + || !reg_overlap_mentioned_p (tmp_reg, SET_SRC (PATTERN (cand->insn))))
>>>
>>> Shouldn't this really be !rtx_equal_p?
>>>
>>
>> Maybe, will it behave the right way if the two regs have different modes or when subregs are involved?
>> (will we even hit such a case in this path?)
>>
>> Thanks,
>> Kyrill
>>>
>>> Bernd
>>>
>>
>
next prev parent reply other threads:[~2015-11-17 10:17 UTC|newest]
Thread overview: 15+ messages / expand[flat|nested] mbox.gz Atom feed top
2015-11-16 14:08 Kyrill Tkachov
2015-11-16 18:41 ` Bernd Schmidt
2015-11-17 9:08 ` Kyrill Tkachov
2015-11-17 9:49 ` Kyrill Tkachov
2015-11-17 10:17 ` Kyrill Tkachov [this message]
2015-11-17 12:10 ` Bernd Schmidt
2015-11-17 13:03 ` Kyrill Tkachov
2015-11-17 23:11 ` Bernd Schmidt
2015-11-18 9:11 ` Kyrill Tkachov
2015-11-19 10:28 ` Kyrill Tkachov
2015-11-20 1:41 ` Bernd Schmidt
2015-11-20 9:16 ` Kyrill Tkachov
2015-11-23 15:12 ` Kyrill Tkachov
2015-11-24 13:33 ` Kyrill Tkachov
2015-11-24 13:42 ` Bernd Schmidt
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=564AFEB9.7040703@arm.com \
--to=kyrylo.tkachov@arm.com \
--cc=bschmidt@redhat.com \
--cc=gcc-patches@gcc.gnu.org \
--cc=law@redhat.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).