Feedback-ID: id4a34324:Fastmail
Date: Fri, 13 Oct 2023 07:55:10 -0700
From: Andres Freund <andres@anarazel.de>
To: Dmitry Dolgov <9erthalion6@gmail.com>
Cc: Thomas Munro <thomas.munro@gmail.com>,
	Devrim =?utf-8?B?R8O8bmTDvHo=?= <devrim@gunduz.org>,
	PostgreSQL Hackers <pgsql-hackers@lists.postgresql.org>,
	Fabien COELHO <coelho@cri.ensmp.fr>,
	Ronan Dunklau <ronan.dunklau@aiven.io>
Subject: Re: LLVM 16 (opaque pointers)
Message-ID: <20231013145510.aq3qiqig6rw5x72v@awork3.anarazel.de>
References: <CA+hUKGKNX_=f+1C4r06WETKTq0G4Z_7q4L4Fxn5WWpMycDj9Fw@mail.gmail.com>
 <CA+hUKG+WXznXCyTgCADd=HWkP9Qksa6chd7L=GCnZo-MBgg9Lg@mail.gmail.com>
 <7d82e672-ed84-4fb1-b686-d477a7e2a043@mines-paristech.fr>
 <CA+hUKG+UQq_qZWuAd8Mv=RpL4sACEb=4UNgNL2--=xZHixVspg@mail.gmail.com>
 <34036e05460e871024873d7542ceff0b2eaac13c.camel@gunduz.org>
 <CA+hUKGLdgmOGFUEefFGCaofPCsPj4WoCFJ-NpFTMjOFQnaog_w@mail.gmail.com>
 <CA+hUKGJnG8jJJkyULpg1xTxzNtL2FreS4C16TFcRLwu8RR1yUQ@mail.gmail.com>
 <20231012233120.chrw72oxc2pg3vpi@awork3.anarazel.de>
 <20231013090621.jmnlnij5slwiwwym@erthalion.local>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20231013090621.jmnlnij5slwiwwym@erthalion.local>
Archived-At: <https://www.postgresql.org/message-id/20231013145510.aq3qiqig6rw5x72v%40awork3.anarazel.de>
Precedence: bulk

Hi,

On 2023-10-13 11:06:21 +0200, Dmitry Dolgov wrote:
> > On Thu, Oct 12, 2023 at 04:31:20PM -0700, Andres Freund wrote:
> > I also don't think we should add the mem2reg pass outside of -O0 - running it
> > after a real optimization pipeline doesn't seem useful and might even make the
> > code worse? mem2reg is included in default<O1> (and obviously also in O3).
> 
> My understanding was that while mem2reg is included everywhere above
> -O0, this set of passes won't hurt. But yeah, if you say it could
> degrade the final result, it's better to not do this. I'll update this
> part.

It's indeed included anywhere above that, but adding it explicitly to the
schedule means it's excuted twice:

echo 'int foo(int a) { return a / 343; }' | clang-16 -emit-llvm -x c -c -o - -S -|sed -e 's/optnone//'|opt-17 -debug-pass-manager -passes='default<O1>,mem2reg' -o /dev/null 2>&1|grep Promote
Running pass: PromotePass on foo (2 instructions)
Running pass: PromotePass on foo (2 instructions)

The second one is in a point in the pipeline where it doesn't help. It also
requires another analysis pass to be executed unnecessarily.

Greetings,

Andres Freund