Re: [PATCH v8] refs.c: SSE2 optimizations for check_refname_component

From: David Turner <dturner@twopensource.com>
Date: Jun 17, 2014, 00:51 UTC
Message-ID: <1402966317.5629.93.camel@stross>
In-Reply-To: <xmqqha3ko0vs.fsf@gitster.dls.corp.google.com>

On Mon, 2014-06-16 at 17:06 -0700, Junio C Hamano wrote:

Show 19 quoted lines

> David Turner <dturner@twopensource.com> writes:
> 
> > Optimize check_refname_component using SSE2 on x86_64.
> >
> > git rev-parse HEAD is a good test-case for this, since it does almost
> > nothing except parse refs.  For one particular repo with about 60k
> > refs, almost all packed, the timings are:
> >
> > Look up table: 29 ms
> > SSE2:          23 ms
> >
> > This cuts about 20% off of the runtime.
> >
> > The configure.ac changes include code from the GNU C Library written
> > by Joseph S. Myers <joseph at codesourcery dot com>.
> >
> > Ondřej Bílka <neleai@seznam.cz> suggested an SSE2 approach to the
> 
> One e-mail address is obfuscated while the other not; intended?

Yes (although I'll happily obfuscate or remove Ondřej's if he prefers); Joseph's was obfuscated because that's how I found it in the source material.

Show 18 quoted lines

> > substring searches, which netted a speed boost over the SSE4.2 code I
> > had initially written.
> >
> > Signed-off-by: David Turner <dturner@twitter.com>
> > ---
> > diff --git a/git-compat-util.h b/git-compat-util.h
> > index f6d3a46..291d46b 100644
> > --- a/git-compat-util.h
> > +++ b/git-compat-util.h
> > @@ -668,6 +668,16 @@ void git_qsort(void *base, size_t nmemb, size_t size,
> >  #endif
> >  #endif
> >  
> > +#if defined(__GNUC__) && defined(__x86_64__)
> > +#include <emmintrin.h>
> > +/* This is the system memory page size; it's used so that we can read
> 
> Style (there are other instances of the same kind).

Will fix.

Show 21 quoted lines

> > +static int check_refname_component_trailer(const char *cp, const char *refname, int flags)
> > +{
> > +	if (cp == refname)
> > +		return 0; /* Component has zero length. */
> > +	if (refname[0] == '.') {
> > +		if (!(flags & REFNAME_DOT_COMPONENT))
> > +			return -1; /* Component starts with '.'. */
> > +		/*
> > +		 * Even if leading dots are allowed, don't allow "."
> > +		 * as a component (".." is prevented by a rule above).
> > +		 */
> > +		if (refname[1] == '\0')
> > +			return -1; /* Component equals ".". */
> > +	}
> > +	if (cp - refname >= 5 && !memcmp(cp - 5, ".lock", 5))
> > +		return -1; /* Refname ends with ".lock". */
> 
> This is merely a moved code that retained the same comment, but it
> is more like "the current refname component ends with .lock", I
> suspect.  In other words, we do not allow "refs/heads/foo.lock/bar".
> Am I reading the patch correctly?

This code didn't actually need to be moved; I'll move it back.

But it now occurs to me that the SSE2 code does not actually check for .lock at the end of each component, but only at the end of string.

I don't know why we don't allow refs/heads/foo.lock/bar. I can understand not allowing refs/heads/foo.lock, since it might be a lock file. But I have never heard of a lock directory like this.

I'm happy to replicate what the existing code does, but do you think it is correct?

> Style (there are other instances of the same kind).  No SP between
> function/macro name and opening parenthesis.

Will fix.

Show 14 quoted lines

> > +	if (refname[0] == '.') {
> > +		if (refname[1] == '/' || refname[1] == '\0')
> > +			return -1;
> > +		if (!(flags & REFNAME_DOT_COMPONENT))
> > +			return -1;
> > +	}
> > +	while(1) {
> > +		__m128i tmp, tmp1, result;
> > +		uint64_t mask;
> > +
> > +		if ((uintptr_t) cp % PAGE_SIZE > PAGE_SIZE - SSE_VECTOR_BYTES  - 1)
> 
> OK, so we make sure we do not overrun by reading too much near the
> end of the page, as the next page might be unmapped.

Exactly.

> I am showing my ignorance but does cp (i.e. refname) upon entry to
> this function need to be aligned in some way?

Nope -- _mm_loadu_si128 does an unaligned load.