Date:   Thu, 9 Mar 2023 22:54:38 -0800
From:   Eric Biggers <ebiggers@kernel.org>
To:     Al Viro <viro@zeniv.linux.org.uk>
Cc:     Yangtao Li <frank.li@vivo.com>, tytso@mit.edu,
        adilger.kernel@dilger.ca, linux-ext4@vger.kernel.org,
        linux-kernel@vger.kernel.org
Subject: Re: [PATCH] ext4: convert to DIV_ROUND_UP() in
 mpage_process_page_bufs()
Message-ID: <ZArULrNVaxl7e42r@sol.localdomain>
References: <20230310060734.8780-1-frank.li@vivo.com>
 <ZArLbO1ckmcXwQf0@sol.localdomain>
 <20230310063729.GO3390869@ZenIV>
 <ZArRq/25a5Gt+YMB@sol.localdomain>
 <20230310064612.GQ3390869@ZenIV>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20230310064612.GQ3390869@ZenIV>
Precedence: bulk

On Fri, Mar 10, 2023 at 06:46:12AM +0000, Al Viro wrote:
> On Thu, Mar 09, 2023 at 10:43:55PM -0800, Eric Biggers wrote:
> > On Fri, Mar 10, 2023 at 06:37:29AM +0000, Al Viro wrote:
> > > On Thu, Mar 09, 2023 at 10:17:16PM -0800, Eric Biggers wrote:
> > > > On Fri, Mar 10, 2023 at 02:07:34PM +0800, Yangtao Li wrote:
> > > > > Just for better readability, no code logic change.
> > > > > 
> > > > > Signed-off-by: Yangtao Li <frank.li@vivo.com>
> > > > > ---
> > > > >  fs/ext4/inode.c | 3 +--
> > > > >  1 file changed, 1 insertion(+), 2 deletions(-)
> > > > > 
> > > > > diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c
> > > > > index d251d705c276..d121cde74522 100644
> > > > > --- a/fs/ext4/inode.c
> > > > > +++ b/fs/ext4/inode.c
> > > > > @@ -2218,8 +2218,7 @@ static int mpage_process_page_bufs(struct mpage_da_data *mpd,
> > > > >  {
> > > > >  	struct inode *inode = mpd->inode;
> > > > >  	int err;
> > > > > -	ext4_lblk_t blocks = (i_size_read(inode) + i_blocksize(inode) - 1)
> > > > > -							>> inode->i_blkbits;
> > > > > +	ext4_lblk_t blocks = DIV_ROUND_UP(i_size_read(inode), i_blocksize(inode));
> > > > >  
> > > > 
> > > > Please don't do this.  This makes the code compile down to a division, which is
> > > > far less efficient.  I've verified this by checking the assembly generated.
> > > 
> > > Which compiler is doing that?
> > 
> > $ gcc --version
> > gcc (GCC) 12.2.1 20230201
> > 
> > i_blocksize(inode) is not a constant, so this should not be particularly
> > surprising.  One might hope that a / (1 << b) would be optimized into a >> b,
> > but that doesn't seem to happen.
> 
> It really ought to be a / (1u << b), though...

Sure, that does better:

uint64_t f(uint64_t a, int b)
{
        return a / (1U << b);
}

gcc:
	0000000000000000 <f>:
	   0:	48 89 f8             	mov    %rdi,%rax
	   3:	89 f1                	mov    %esi,%ecx
	   5:	48 d3 e8             	shr    %cl,%rax
	   8:	c3                   	ret

clang:
	0000000000000000 <f>:
	   0:	89 f1                	mov    %esi,%ecx
	   2:	48 89 f8             	mov    %rdi,%rax
	   5:	48 d3 e8             	shr    %cl,%rax
	   8:	c3                   	ret

But with a signed dividend (which is the case here) it gets messed up:

int64_t f(int64_t a, int b)
{
        return a / (1U << b);
}

gcc:
	0000000000000000 <f>:
	   0:	48 89 f8             	mov    %rdi,%rax
	   3:	89 f1                	mov    %esi,%ecx
	   5:	bf 01 00 00 00       	mov    $0x1,%edi
	   a:	d3 e7                	shl    %cl,%edi
	   c:	48 99                	cqto
	   e:	48 f7 ff             	idiv   %rdi
	  11:	c3                   	ret

clang:
	0000000000000000 <f>:
	   0:	89 f1                	mov    %esi,%ecx
	   2:	be 01 00 00 00       	mov    $0x1,%esi
	   7:	d3 e6                	shl    %cl,%esi
	   9:	48 89 f8             	mov    %rdi,%rax
	   c:	48 89 f9             	mov    %rdi,%rcx
	   f:	48 c1 e9 20          	shr    $0x20,%rcx
	  13:	74 06                	je     1b <f+0x1b>
	  15:	48 99                	cqto
	  17:	48 f7 fe             	idiv   %rsi
	  1a:	c3                   	ret
	  1b:	31 d2                	xor    %edx,%edx
	  1d:	f7 f6                	div    %esi
	  1f:	c3                   	ret