| 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
| ||
|
Message-ID: <CALx6S35OYdrFFVvYid7hA6iHJtW9B=FOuwR4L+EA6TWk5sP7qw@mail.gmail.com>
Date: Wed, 9 Mar 2016 16:58:49 -0800
From: Tom Herbert <tom@...bertland.com>
To: Joe Perches <joe@...ches.com>
Cc: Alexander Duyck <alexander.duyck@...il.com>,
Alexander Duyck <aduyck@...antis.com>,
Netdev <netdev@...r.kernel.org>,
David Miller <davem@...emloft.net>
Subject: Re: [net-next PATCH] csum: Update csum_block_add to use rotate
instead of byteswap
On Wed, Mar 9, 2016 at 4:18 PM, Joe Perches <joe@...ches.com> wrote:
> On Wed, 2016-03-09 at 08:08 -0800, Alexander Duyck wrote:
>> On Tue, Mar 8, 2016 at 10:31 PM, Tom Herbert <tom@...bertland.com> wrote:
>> > I took a look inlining these.
>> >
>> > #define rol32(V, X) ({ \
>> > int word = V; \
>> > if (__builtin_constant_p(X)) \
>> > asm("roll $" #X ",%[word]\n\t" \
>> > : [word] "=r" (word)); \
>> > else \
>> > asm("roll %%cl,%[word]\n\t" \
>> > : [word] "=r" (word) \
>> > : "c" (X)); \
>> > word; \
>> > })
>> >
>> > With this I'm seeing a nice speedup in jhash which uses a lot of rol32s...
>> Is gcc really not converting the rol32 calls into rotates?
>
> No, it is.
>
> The difference in the object code with the asm for instance is:
>
> (old, compiled with gcc 5.3.1)
>
> <jhash_2words.constprop.5>:
> 84e: 81 ee 09 41 52 21 sub $0x21524109,%esi
> 854: 81 ef 09 41 52 21 sub $0x21524109,%edi
> 85a: 55 push %rbp
> 85b: 89 f0 mov %esi,%eax
> 85d: 89 f2 mov %esi,%edx
> 85f: 48 ff 05 00 00 00 00 incq 0x0(%rip) # 866 <jhash_2words.constprop.5+0x18>
> 866: c1 c2 0e rol $0xe,%edx
> 869: 35 f7 be ad de xor $0xdeadbef7,%eax
> 86e: 48 89 e5 mov %rsp,%rbp
> 871: 29 d0 sub %edx,%eax
> 873: 48 ff 05 00 00 00 00 incq 0x0(%rip) # 87a <jhash_2words.constprop.5+0x2c>
> 87a: 48 ff 05 00 00 00 00 incq 0x0(%rip) # 881 <jhash_2words.constprop.5+0x33>
> 881: 89 c2 mov %eax,%edx
> 883: 31 c7 xor %eax,%edi
> 885: c1 c2 0b rol $0xb,%edx
> 888: 29 d7 sub %edx,%edi
> 88a: 89 fa mov %edi,%edx
> 88c: 31 fe xor %edi,%esi
> 88e: c1 ca 07 ror $0x7,%edx
> 891: 29 d6 sub %edx,%esi
> 893: 89 f2 mov %esi,%edx
> 895: 31 f0 xor %esi,%eax
> 897: c1 c2 10 rol $0x10,%edx
> 89a: 29 d0 sub %edx,%eax
> 89c: 89 c2 mov %eax,%edx
> 89e: 31 c7 xor %eax,%edi
> 8a0: c1 c2 04 rol $0x4,%edx
> 8a3: 29 d7 sub %edx,%edi
> 8a5: 31 fe xor %edi,%esi
> 8a7: c1 c7 0e rol $0xe,%edi
> 8aa: 29 fe sub %edi,%esi
> 8ac: 31 f0 xor %esi,%eax
> 8ae: c1 ce 08 ror $0x8,%esi
> 8b1: 29 f0 sub %esi,%eax
> 8b3: 5d pop %rbp
> 8b4: c3 retq
>
> vs Tom's asm
>
> 000000000000084e <jhash_2words.constprop.5>:
> 84e: 81 ee 09 41 52 21 sub $0x21524109,%esi
> 854: 8d 87 f7 be ad de lea -0x21524109(%rdi),%eax
> 85a: 55 push %rbp
> 85b: 89 f2 mov %esi,%edx
> 85d: 48 ff 05 00 00 00 00 incq 0x0(%rip) # 864 <jhash_2words.constprop.5+0x16>
> 864: 48 ff 05 00 00 00 00 incq 0x0(%rip) # 86b <jhash_2words.constprop.5+0x1d>
> 86b: 81 f2 f7 be ad de xor $0xdeadbef7,%edx
> 871: 48 89 e5 mov %rsp,%rbp
> 874: c1 c1 0e rol $0xe,%ecx
> 877: 29 ca sub %ecx,%edx
> 879: 31 d0 xor %edx,%eax
> 87b: c1 c7 0b rol $0xb,%edi
> 87e: 29 f8 sub %edi,%eax
> 880: 48 ff 05 00 00 00 00 incq 0x0(%rip) # 887 <jhash_2words.constprop.5+0x39>
> 887: 31 c6 xor %eax,%esi
> 889: c1 c7 19 rol $0x19,%edi
> 88c: 29 fe sub %edi,%esi
> 88e: 31 f2 xor %esi,%edx
> 890: c1 c7 10 rol $0x10,%edi
> 893: 29 fa sub %edi,%edx
> 895: 31 d0 xor %edx,%eax
> 897: c1 c7 04 rol $0x4,%edi
> 89a: 29 f8 sub %edi,%eax
> 89c: 31 f0 xor %esi,%eax
> 89e: 29 c8 sub %ecx,%eax
> 8a0: 31 d0 xor %edx,%eax
> 8a2: 5d pop %rbp
> 8a3: c1 c2 18 rol $0x18,%edx
> 8a6: 29 d0 sub %edx,%eax
> 8a8: c3 retq
>
>> If we need this type of code in order to get the rotates to occur as
>> expected then maybe we need to look at doing arch specific versions of
>> the functions in bitops.h in order to improve the performance since I
>> know these calls are used in some performance critical paths such as
>> crypto and hashing.
>
> Yeah, maybe, but why couldn't gcc generate similar code
> as Tom's asm? (modulo the ripple reducing ror vs rol uses
> when the shift is > 16
I see gcc doing that now, not sure why I was seeing differences before....
Powered by blists - more mailing lists