linux-kernel - Re: [PATCH 1/1] mm/rmap: optimize MM-ID mapcount handling with union

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <e3c635ed-7d4b-4d07-80ad-75be950b803c@redhat.com>
Date: Mon, 21 Apr 2025 09:40:20 +0200
From: David Hildenbrand <david@...hat.com>
To: Lance Yang <lance.yang@...ux.dev>, Lance Yang <ioworker0@...il.com>,
 akpm@...ux-foundation.org
Cc: mingzhe.yang@...com, willy@...radead.org, ziy@...dia.com,
 mhocko@...e.com, vbabka@...e.cz, surenb@...gle.com, linux-mm@...ck.org,
 jackmanb@...gle.com, hannes@...xchg.org, linux-kernel@...r.kernel.org
Subject: Re: [PATCH 1/1] mm/rmap: optimize MM-ID mapcount handling with union

>>
>> Are we sure the compiler cannot optimize that itself?
>>
>> On x86-64 I get with gcc 14.2.1:
>>
>> ; folio->_mm_id_mapcount[0] = -1;
>>
>>   3f2f: 48 c7 42 60 ff ff ff ff movq $-0x1, 0x60(%rdx)
>>
>> Which should be a quadword (64bit) setting, so exactly what you want to achieve.
> 
> 
> Yeah, the compiler should be as smart as we expect it to be.
> However, it seems that gcc 4.8.5 doesn't behave as expected
> with the -O2 optimization level on the x86-64 test machine.
> 
> struct folio_array {
>      int _mm_id_mapcount[2];
> };
> 
> void init_array(struct folio_array *f) {
>      f->_mm_id_mapcount[0] = -1;
>      f->_mm_id_mapcount[1] = -1;
> }
> 
> 0000000000000000 <init_array>:
>     0:	c7 07 ff ff ff ff    	movl   $0xffffffff,(%rdi)
>     6:	c7 47 04 ff ff ff ff 	movl   $0xffffffff,0x4(%rdi)
>     d:	c3                   	retq
> 
> ---
> 
> struct folio_union {
>      union {
>          int _mm_id_mapcount[2];
>          unsigned long _mm_id_mapcounts;
>      };
> };
> 
> void init_union(struct folio_union *f) {
>      f->_mm_id_mapcounts = -1UL;
> }
> 
> 0000000000000010 <init_union>:
>    10:	48 c7 07 ff ff ff ff 	movq   $0xffffffffffffffff,(%rdi)
>    17:	c3                   	retq
> 
> Hmm... I'm not sure if it's valuable for those compilers that
> are not very new.

Yeah, we shouldn't care about performance with rusty old compilers, 
especially if the gain would likely not even be measurable.

Note that even Linux requires 5.1 ever since 2021. GCC seems to 
implement this optimization starting with 7.1 (at least when playing 
with the compiler explorer).

-- 
Cheers,

David / dhildenb