From: Jeff Hostetler <git@jeffhostetler.com>
To: "René Scharfe" <l.s.r@web.de>, git@vger.kernel.org
Cc: gitster@pobox.com, peff@peff.net,
Jeff Hostetler <jeffhost@microsoft.com>
Subject: Re: [PATCH v10 3/3] read-cache: speed up add_index_entry during checkout
Date: Mon, 17 Apr 2017 10:53:06 -0400 [thread overview]
Message-ID: <28ef82b9-ab7a-ccaf-7bae-4eaa332725f8@jeffhostetler.com> (raw)
In-Reply-To: <d9f1acc8-03e3-abc6-af9d-da55a33218fa@web.de>
On 4/15/2017 1:55 PM, René Scharfe wrote:
> Am 14.04.2017 um 21:12 schrieb git@jeffhostetler.com:
>> From: Jeff Hostetler <jeffhost@microsoft.com>
>
> Very nice, especially the perf test! But can we unbundle the different
> optimizations into separate patches with their own performance numbers?
> Candidates IMHO: The change in add_index_entry_with_check(), the first
> hunk in has_dir_name(), the loop shortcuts. That might also help find
> the reason for the slight slowdown of 0006.3 for the kernel repository.
Let me take a look at this and see if it helps.
>> diff --git a/read-cache.c b/read-cache.c
>> index 97f13a1..ba95fbb 100644
>> --- a/read-cache.c
>> +++ b/read-cache.c
>> @@ -910,6 +910,9 @@ int strcmp_offset(const char *s1, const char *s2,
>> size_t *first_change)
>> /*
>> * Do we have another file with a pathname that is a proper
>> * subset of the name we're trying to add?
>> + *
>> + * That is, is there another file in the index with a path
>> + * that matches a sub-directory in the given entry?
>> */
>> static int has_dir_name(struct index_state *istate,
>> const struct cache_entry *ce, int pos, int ok_to_replace)
>> @@ -918,9 +921,51 @@ static int has_dir_name(struct index_state *istate,
>> int stage = ce_stage(ce);
>> const char *name = ce->name;
>> const char *slash = name + ce_namelen(ce);
>> + size_t len_eq_last;
>> + int cmp_last = 0;
>> +
>> + /*
>> + * We are frequently called during an iteration on a sorted
>> + * list of pathnames and while building a new index. Therefore,
>> + * there is a high probability that this entry will eventually
>> + * be appended to the index, rather than inserted in the middle.
>> + * If we can confirm that, we can avoid binary searches on the
>> + * components of the pathname.
>> + *
>> + * Compare the entry's full path with the last path in the index.
>> + */
>> + if (istate->cache_nr > 0) {
>> + cmp_last = strcmp_offset(name,
>> + istate->cache[istate->cache_nr - 1]->name,
>> + &len_eq_last);
>> + if (cmp_last > 0) {
>> + if (len_eq_last == 0) {
>> + /*
>> + * The entry sorts AFTER the last one in the
>> + * index and their paths have no common prefix,
>> + * so there cannot be a F/D conflict.
>> + */
>> + return retval;
>> + } else {
>> + /*
>> + * The entry sorts AFTER the last one in the
>> + * index, but has a common prefix. Fall through
>> + * to the loop below to disect the entry's path
>> + * and see where the difference is.
>> + */
>> + }
>> + } else if (cmp_last == 0) {
>> + /*
>> + * The entry exactly matches the last one in the
>> + * index, but because of multiple stage and CE_REMOVE
>> + * items, we fall through and let the regular search
>> + * code handle it.
>> + */
>> + }
>> + }
>> for (;;) {
>> - int len;
>> + size_t len;
>> for (;;) {
>> if (*--slash == '/')
>> @@ -930,6 +975,66 @@ static int has_dir_name(struct index_state *istate,
>> }
>> len = slash - name;
>> + if (cmp_last > 0) {
>> + /*
>> + * (len + 1) is a directory boundary (including
>> + * the trailing slash). And since the loop is
>> + * decrementing "slash", the first iteration is
>> + * the longest directory prefix; subsequent
>> + * iterations consider parent directories.
>> + */
>> +
>> + if (len + 1 <= len_eq_last) {
>> + /*
>> + * The directory prefix (including the trailing
>> + * slash) also appears as a prefix in the last
>> + * entry, so the remainder cannot collide (because
>> + * strcmp said the whole path was greater).
>> + *
>> + * EQ: last: xxx/A
>> + * this: xxx/B
>> + *
>> + * LT: last: xxx/file_A
>> + * this: xxx/file_B
>> + */
>> + return retval;
>> + }
>> +
>> + if (len > len_eq_last) {
>> + /*
>> + * This part of the directory prefix (excluding
>> + * the trailing slash) is longer than the known
>> + * equal portions, so this sub-directory cannot
>> + * collide with a file.
>> + *
>> + * GT: last: xxxA
>> + * this: xxxB/file
>> + */
>> + return retval;
>> + }
>> +
>
> At this point len and len_eq_last are equal -- otherwise one of the two
> previous conditions would have triggered. Silly question: Is this
> necessary for the following shortcut to work? Removing the two checks
> above doesn't seem to affect performance very much, and at least the
> test suite still passes..
Both of these are highly dependent on the shape of the
pathnames in the tree. Let me try to quantify this.
>
>> + if (ce_namelen(istate->cache[istate->cache_nr - 1]) > len) {
>> + /*
>> + * The directory prefix lines up with part of
>> + * a longer file or directory name, but sorts
>> + * after it, so this sub-directory cannot
>> + * collide with a file.
>> + *
>> + * last: xxx/yy-file (because '-' sorts before '/')
>> + * this: xxx/yy/abc
>> + */
>> + return retval;
>> + }
>
> istate->cache_nr can be zero if remove_index_entry_at() had been called
> in a previous iteration, resulting in a segfault. Checking right here
> is probably the easiest way out; not sure if exiting early when the
> index becomes empty would be better.
yeah, i guess remove_...() could delete them all and this test
would be ill-defined. Let me fix that.
>
>> +
>> + /*
>> + * This is a possible collision. Fall through and
>> + * let the regular search code handle it.
>> + *
>> + * last: xxx
>> + * this: xxx/file
>> + */
>> + }
>> +
>> pos = index_name_stage_pos(istate, name, len, stage);
>> if (pos >= 0) {
>> /*
>> @@ -1021,7 +1126,16 @@ static int add_index_entry_with_check(struct
>> index_state *istate, struct cache_e
>> if (!(option & ADD_CACHE_KEEP_CACHE_TREE))
>> cache_tree_invalidate_path(istate, ce->name);
>> - pos = index_name_stage_pos(istate, ce->name, ce_namelen(ce),
>> ce_stage(ce));
>> +
>> + /*
>> + * If this entry's path sorts after the last entry in the index,
>> + * we can avoid searching for it.
>> + */
>> + if (istate->cache_nr > 0 &&
>> + strcmp(ce->name, istate->cache[istate->cache_nr - 1]->name) > 0)
>> + pos = -istate->cache_nr - 1;
>> + else
>> + pos = index_name_stage_pos(istate, ce->name, ce_namelen(ce),
>> ce_stage(ce));
>> /* existing match? Just replace it. */
>> if (pos >= 0) {
>>
next prev parent reply other threads:[~2017-04-17 14:53 UTC|newest]
Thread overview: 7+ messages / expand[flat|nested] mbox.gz Atom feed top
2017-04-14 19:12 [PATCH v10 0/3] read-cache: speed up add_index_entry git
2017-04-14 19:12 ` [PATCH v10 1/3] read-cache: add strcmp_offset function git
2017-04-14 19:12 ` [PATCH v10 2/3] p0006-read-tree-checkout: perf test to time read-tree git
2017-04-14 19:12 ` [PATCH v10 3/3] read-cache: speed up add_index_entry during checkout git
2017-04-15 17:55 ` René Scharfe
2017-04-17 14:53 ` Jeff Hostetler [this message]
2017-04-18 16:19 ` Jeff Hostetler
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
List information: http://vger.kernel.org/majordomo-info.html
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=28ef82b9-ab7a-ccaf-7bae-4eaa332725f8@jeffhostetler.com \
--to=git@jeffhostetler.com \
--cc=git@vger.kernel.org \
--cc=gitster@pobox.com \
--cc=jeffhost@microsoft.com \
--cc=l.s.r@web.de \
--cc=peff@peff.net \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
Code repositories for project(s) associated with this public inbox
https://80x24.org/mirrors/git.git
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).