This is the mail archive of the
mailing list for the glibc project.
Re: Variations of memset()
On Fri, Aug 4, 2017 at 12:54 PM, Carlos O'Donell <email@example.com> wrote:
> On 08/04/2017 03:11 PM, Carlos O'Donell wrote:
>> On 08/04/2017 03:02 PM, Matthew Wilcox wrote:
>>> Here's the sample usage from the symbios driver:
>>> - for (i = 0 ; i < 64 ; i++)
>>> - tp->luntbl[i] = cpu_to_scr(vtobus(&np->badlun_sa));
>>> + memset32(tp->luntbl, cpu_to_scr(vtobus(&np->badlun_sa)), 64);
>>> I expect a lot of users would be of this type; simply replacing the
>>> explicit for-loop equivalent with a library call.
>> Have you measured the performance of this kind of conversion when using a
>> simple application and a library implementing your various memset routines?
>> In the kernel is one thing, outside of the kernel we have dynamic linking
>> and no-inling across that shared object boundary.
> I want to reiterate that measuring the performance of various options in
> userspace is going to be relevant (particularly when they vary from the kernel):
> * Application doing the naive loop above (-O0).
> * Application doing the naive loop above ([-O2,-O3] + <vectorize options>).
> * Application calling memset32 (-O0)
> * Application calling memset32 (-O3)
> <vectorize options>="-ftree-vectorize [-msse2,-mavx] -fopt-info-missed=missed.all"
> You need to split the memset32 into another DSO to simulate this accurately.
These functions aren't very useful for x86-64 where wmemset,
aka, memset32, is implemented with memset:
Dump of assembler code for function __wmemset_sse2_unaligned:
0x0000000000000020 <+0>: shl $0x2,%rdx
0x0000000000000024 <+4>: movd %esi,%xmm0
0x0000000000000028 <+8>: mov %rdi,%rax
0x000000000000002b <+11>: pshufd $0x0,%xmm0,%xmm0
0x0000000000000030 <+16>: jmp 0x64 <__memset_sse2_unaligned+20>
End of assembler dump.