From: Jeff Law <jeffreyalaw@gmail.com>
To: Richard Biener <richard.guenther@gmail.com>, Jeff Law <jlaw@tachyum.com>
Cc: GCC Patches <gcc-patches@gcc.gnu.org>
Subject: Re: [RFA] Attach MEM_EXPR information when flushing BLKmode args to the stack
Date: Thu, 8 Jul 2021 13:19:21 -0600 [thread overview]
Message-ID: <4334b028-0fe3-c6df-1cc0-a9bf8b8b453c@gmail.com> (raw)
In-Reply-To: <CAFiYyc3pyAWJmNRhE3QrVBNWgRmVJ0m_VHErkzorK_JHpB+4pw@mail.gmail.com>
On 7/5/2021 5:17 AM, Richard Biener via Gcc-patches wrote:
> On Fri, Jul 2, 2021 at 6:13 PM Jeff Law <jlaw@tachyum.com> wrote:
>>
>> This is a minor missed optimization we found with our internal port.
>>
>> Given this code:
>>
>> typedef struct {short a; short b;} T;
>>
>> extern void g1();
>>
>> void f(T s)
>> {
>> if (s.a < 0)
>> g1();
>> }
>>
>>
>> "s" is passed in a register, but it's still a BLKmode object because the
>> alignment of T is smaller than the alignment that an integer of the same
>> size would have (16 bits vs 32 bits).
>>
>>
>> Because "s" is BLKmode function.c is going to store it into a stack slot
>> and we'll load it from the that slot for each reference. So on the v850
>> (just to pick a port that likely has the same behavior we're seeing) we
>> have this RTL from CSE2:
>>
>>
>> (insn 2 4 3 2 (set (mem/c:SI (plus:SI (reg/f:SI 34 .fp)
>> (const_int -4 [0xfffffffffffffffc])) [2 S4 A32])
>> (reg:SI 6 r6)) "j.c":6:1 7 {*movsi_internal}
>> (expr_list:REG_DEAD (reg:SI 6 r6)
>> (nil)))
>> (note 3 2 8 2 NOTE_INSN_FUNCTION_BEG)
>> (insn 8 3 9 2 (set (reg:HI 44 [ s.a ])
>> (mem/c:HI (plus:SI (reg/f:SI 34 .fp)
>> (const_int -4 [0xfffffffffffffffc])) [1 s.a+0 S2 A32]))
>> "j.c":7:5 3 {*movhi_internal}
>> (nil))
>> (insn 9 8 10 2 (parallel [
>> (set (reg:SI 45)
>> (ashift:SI (subreg:SI (reg:HI 44 [ s.a ]) 0)
>> (const_int 16 [0x10])))
>> (clobber (reg:CC 32 psw))
>> ]) "j.c":7:5 94 {ashlsi3_clobber_flags}
>> (expr_list:REG_DEAD (reg:HI 44 [ s.a ])
>> (expr_list:REG_UNUSED (reg:CC 32 psw)
>> (nil))))
>> (insn 10 9 11 2 (parallel [
>> (set (reg:SI 43)
>> (ashiftrt:SI (reg:SI 45)
>> (const_int 16 [0x10])))
>> (clobber (reg:CC 32 psw))
>> ]) "j.c":7:5 104 {ashrsi3_clobber_flags}
>> (expr_list:REG_DEAD (reg:SI 45)
>> (expr_list:REG_UNUSED (reg:CC 32 psw)
>> (nil))))
>>
>>
>> Insn 2 is the store into the stack. insn 8 is the load for s.a in the
>> conditional. DSE1 replaces the MEM in insn 8 with (reg 6) since (reg 6)
>> has the value we want. After that the store at insn 2 is dead. Sadly
>> DSE never removes the store.
>>
>> The problem is RTL DSE considers a store with no MEM_EXPR as escaping,
>> which keeps the MEM live. The lack of a MEM_EXPR is due to call to
>> change_address to twiddle the mode on the MEM for the store at insn 2.
>> It should be safe to copy the MEM_EXPR (which should always be a
>> PARM_DECL) from the original memory to the memory returned by
>> change_address. Doing so results in DSE1 removing the store at insn 2.
>>
>> It would be nice to remove the stack setup/teardown. I'm not offhand
>> aware of mechanisms to remove the setup/teardown after we've already
>> allocated a slot, even if the slot is no longer used.
>>
>> Bootstrapped and regression tested on x86, though I don't think that's a
>> particularly useful test. So I also ran it through my tester across
>> those pesky embedded targets without regressions as well.
>>
>> I didn't include a test simply because I didn't want to have an insane
>> target selector. I guess if we really wanted a test we could look after
>> DSE1 is done and verify there aren't any MEMs left at all. Willing to
>> try that if the consensus is we want this tested.
>>
>> OK for the trunk?
> I wonder why the code doesn't use adjust_address instead? That
> handles most cases already and the code doesn't change the
> address but just the mode (and access size)?
Yea, adjust_address seems to work fine. I'm spinning that in my tester
at the moment.
Jeff
next prev parent reply other threads:[~2021-07-08 19:19 UTC|newest]
Thread overview: 6+ messages / expand[flat|nested] mbox.gz Atom feed top
2021-07-02 16:13 Jeff Law
2021-07-05 11:17 ` Richard Biener
2021-07-06 4:01 ` Jeff Law
2021-07-08 19:19 ` Jeff Law [this message]
2021-07-09 2:39 ` [RFA] Attach MEM_EXPR information when flushing BLKmode args to the stack - V2 Jeff Law
2021-07-09 6:44 ` Richard Biener
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=4334b028-0fe3-c6df-1cc0-a9bf8b8b453c@gmail.com \
--to=jeffreyalaw@gmail.com \
--cc=gcc-patches@gcc.gnu.org \
--cc=jlaw@tachyum.com \
--cc=richard.guenther@gmail.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).