From mboxrd@z Thu Jan 1 00:00:00 1970 From: "J. R. Okajima" Subject: Re: Big git diff speedup by avoiding x86 "fast string" memcmp Date: Wed, 15 Dec 2010 14:57:05 +0900 Message-ID: <6910.1292392625@jrobl> References: <20101209070938.GA3949@amd> <19324.1291990997@jrobl> <20101213014553.GA6522@amd> <9580.1292225351@jrobl> <12853.1292353313@jrobl> Cc: Nick Piggin , Linus Torvalds , linux-arch@vger.kernel.org, x86@kernel.org, linux-kernel@vger.kernel.org, linux-fsdevel@vger.kernel.org To: Nick Piggin Return-path: In-Reply-To: Sender: linux-arch-owner@vger.kernel.org List-Id: linux-fsdevel.vger.kernel.org Thank you for explanation. I understood about padding zero tail bytes as an improvement or a variation of the 'long *' approach. But I didn't write about it in my previous mail. If I had written, you might not need to describe again. I have also compared "repe; cmpsb" and __builtin_memcmp "in userspace". And __builtin_memcmp shows faster and it looks like an approach of mixing pointers. Although the code size grows, in source it will be simple #define and it is already done in x86_32. Do you (or any other x86 experts) think it is worth to try? J. R. Okajima Nick Piggin: > What I meant is that a "normal" memcmp that does long * memory > operations is not a good idea, because it requires code and branches > to handle the tail of the string. When average string lengths are less > than 16 bytes, it hardly seems wothwhile. It will just get more > mispredicts and bigger icache footprint. > > However instead of a normal memcmp, we could actually pad dentry > names out to sizeof(long) with zeros, and take advantage of that with > a memcmp that does not have to handle tails -- it would operate > entirely with longs. > > That would avoid icache and branch regressions, and might speed up > the operation on some architectures. I just doubted whether it would > show an improvement to be worth doing at all. If it does, I'm all for it.