"git clone" that clones from SHA-1 repository, while
GIT_DEFAULT_HASH set to use SHA-256 already, resulted in an
unusable repository that half-claims to be SHA-256 repository
with SHA-1 objects and refs. This has been corrected.
* bc/clone-with-git-default-hash-fix:
builtin/clone: avoid failure with GIT_DEFAULT_HASH
"git commit-graph write" learned to limit the number of bloom
filters that are computed from scratch with the --max-new-filters
option.
* tb/bloom-improvements:
commit-graph: introduce 'commitGraph.maxNewFilters'
builtin/commit-graph.c: introduce '--max-new-filters=<n>'
commit-graph: rename 'split_commit_graph_opts'
bloom: encode out-of-bounds filters as non-empty
bloom/diff: properly short-circuit on max_changes
bloom: use provided 'struct bloom_filter_settings'
bloom: split 'get_bloom_filter()' in two
commit-graph.c: store maximum changed paths
commit-graph: respect 'commitGraph.readChangedPaths'
t/helper/test-read-graph.c: prepare repo settings
commit-graph: pass a 'struct repository *' in more places
t4216: use an '&&'-chain
commit-graph: introduce 'get_bloom_filter_settings()'
More FAQ entries.
* bc/faq-misc:
docs: explain how to deal with files that are always modified
docs: explain why reverts are not always applied on merge
docs: explain why squash merges are broken with long-running branches
Get rid of 'dense' argument that is redundant for every function that has
'struct rev_info *rev' argument as well, as the value of 'dense' passed is
always taken from 'rev->dense_combined_merges' field.
The only place where this was not the case is in 'submodule.c' where
'diff_tree_combined_merge()' was called with '1' for 'dense' argument. However,
at that call the 'revs' instance used is local to the function, and we now just
set 'revs->dense_combined_merges' to 1 in this local instance.
Signed-off-by: Sergey Organov <sorganov@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When add_delta_base_cache() is called with a base that is already in the
cache, no operation is performed. But the check is done after allocating
space for a new entry, so we end up leaking memory on the early return.
In addition, the caller never free()'s the base as it expects the
function to take ownership of it. But the base is not released when we
skip insertion, so it also gets leaked. To fix these problems, move the
allocation of a new entry further down in add_delta_base_cache(), and
free() the base on early return.
Signed-off-by: Matheus Tavares <matheus.bernardino@usp.br>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The third phase of unpack_entry() performs the following sequence in a
loop, until all the deltas enumerated in phase one are applied and the
entry is fully reconstructed:
1. Add the current base entry to the delta base cache
2. Unpack the next delta
3. Patch the unpacked delta on top of the base
When the optional object reading lock is enabled, the above steps will
be performed while holding the lock. However, step 2. momentarily
releases it so that inflation can be performed in parallel for increased
performance. Because the `base` buffer inserted in the cache at 1. is
not duplicated, another thread can potentially free() it while the lock
is released at 2. (e.g. when there is no space left in the cache to
insert another entry). In this case, the later attempt to dereference
`base` at 3. will cause a segmentation fault. This problem was observed
during a multithreaded git-grep execution on a repository with large
objects.
To fix the race condition (and later segmentation fault), let's reorder
the aforementioned steps so that `base` is only added to the cache at
the end. This will prevent the buffer from being released by another
thread while it is still in use. An alternative solution which would not
require the reordering would be to duplicate `base` before inserting it
in the cache. However, as Phil Hord mentioned, memcpy()'ing large bases
can negatively affect performance: in his experiments, this alternative
approach slowed git-grep down by 10% to 20%.
Reported-by: Phil Hord <phil.hord@gmail.com>
Signed-off-by: Matheus Tavares <matheus.bernardino@usp.br>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When a fetch with the --filter argument is made, the configured default
filter is set even if one already exists. This change was made in
5e46139376 ("builtin/fetch: remove unique promisor remote limitation",
2019-06-25) - in particular, changing from:
* If this is the FIRST partial-fetch request, we enable partial
* on this repo and remember the given filter-spec as the default
* for subsequent fetches to this remote.
to:
* If this is a partial-fetch request, we enable partial on
* this repo if not already enabled and remember the given
* filter-spec as the default for subsequent fetches to this
* remote.
(The given filter-spec is "remembered" even if there is already an
existing one.)
This is problematic whenever a lazy fetch is made, because lazy fetches
are made using "git fetch --filter=blob:none", but this will also happen
if the user invokes "git fetch --filter=<filter>" manually. Therefore,
restore the behavior prior to 5e46139376, which writes a filter-spec
only if the current fetch request is the first partial-fetch one (for
that remote).
Signed-off-by: Jonathan Tan <jonathantanmy@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The text tries to say the code accepts many variations that look remotely
like scissors and perforation marks, but gives too little detail for users
to decide what is and what is not taken as a scissors line for themselves.
Instead of describing the heuristics more, just spell out what will always
be accepted, namely "-- >8 --", as it would not help users to give them
more choices and flexibility and be "creative" in their scissors line.
Signed-off-by: Evan Gates <evan.gates@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
On Windows, we use the `vcpkg` project to manage the dependencies, via
`compat/vcbuild/`. Let's make sure that these dependencies are found by
default.
This is needed because we are about to recommend loading the Git
worktree as a folder into Visual Studio, relying on the automatic CMake
support (which would make it relatively cumbersome to adjust the search
path used by CMake manually).
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
By default, Git for Windows does not install its `sh.exe` into the
`PATH`. However, our current `CMakeLists.txt` expects to find a shell
interpreter in the `PATH`.
So let's fall back to looking in the default location where Git for
Windows _does_ install a relatively convenient `sh.exe`:
`C:\Program Files\Git\bin\sh.exe`
Helped-by: Øystein Walle <oystwa@gmail.com>
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
That should be a ":", not a second "=". While at it, refer to the
placeholder "<n>" as "<n>", not "n" (see, e.g., the entry just before
this one).
Signed-off-by: Martin Ågren <martin.agren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We document how `merge.suppressDest` can be used to omit " into <branch
name>" from the title of the merge message. It is true that we omit the
space character before "into", but that lone double quote character
risks ending up on the wrong side of a line break, looking a bit out of
place. This currently happens with, e.g., 80-character terminals.
Drop that leading quoted space. The result should be just as clear about
how this option affects the formatted message.
Signed-off-by: Martin Ågren <martin.agren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Instead of checking for "refs/heads/" using `starts_with()`, then
skipping past "refs/heads/" using `strlen()`, just use `skip_prefix()`.
In `is_worktree_being_rebased()`, we can adjust the indentation while
we're here and lose a pair of parentheses which isn't needed and which
might even make the reader wonder what they're missing and why that
grouping is there.
Signed-off-by: Martin Ågren <martin.agren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
As the commit message of 04a3dfb8b5 ("worktree.c: check whether branch
is bisected in another worktree", 2016-04-22) indicates, the function
`is_worktree_being_bisected()` is based on the older function
`is_worktree_being_rebased()`. This heritage can also be seen in the
name of the variable where we store our return value: It was never
adapted while copy-editing and remains as `found_rebase`.
Rename the variable to make clear that we're looking for a bisect(ion),
nothing else.
Signed-off-by: Martin Ågren <martin.agren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The comment above `add_head_info()` mentions "head_sha1", but it was
renamed to "head_oid" in 0f05154c70 ("worktree: convert struct worktree
to object_id", 2017-10-15). Update the comment.
Signed-off-by: Martin Ågren <martin.agren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We have `strbuf_worktree_ref()`, which works on a strbuf, and a wrapper
for it, `worktree_ref()` which returns a string. We even make this
wrapper available through worktree.h. But it only has a single caller,
sitting right next to it in worktree.c.
Just inline the wrapper into its only caller. This means the caller can
quite naturally reuse a single strbuf. We currently achieve something
similar by having a static strbuf in the wrapper.
Signed-off-by: Martin Ågren <martin.agren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When we have a `struct wt_status_state`, we manually free its `branch`,
`onto` and `detached_from`, or sometimes just one or two of them.
Provide a function `wt_status_state_free_buffers()` which does the
freeing.
The callers are still aware of these fields, e.g., they check whether
`branch` was populated or not. But this way, they don't need to know
about *all* of them, and if `struct wt_status_state` gets more fields,
they will not need to learn to free them.
Users of `struct wt_status` (which contains a `wt_status_state`) already
have `wt_status_collect_free_buffers()` (corresponding to
`wt_status_collect()`) which we can also teach to use this new helper.
Finally, note that we're currently leaving dangling pointers behind.
Some callers work on a stack-allocated struct, where this is obviously
ok. But for the users of `run_status()` in builtin/commit.c, there are
ample opportunities for someone to mistakenly use those dangling
pointers. We seem to be ok for now, but it's a use-after-free waiting to
happen. Let's leave NULL-pointers behind instead.
Signed-off-by: Martin Ågren <martin.agren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We pass around a `FILE *` in the `struct wt_status` and almost always
print to it. But in a few places, we write to `stdout` instead, either
explicitly through `fprintf(stdout, ...)` or implicitly with
`printf(...)` (and a few `putchar(...)`).
Always be explicit about writing to `s->fp`. To the best of my
understanding, this never mattered in practice because these spots are
involved in various forms of `git status` which always end up at
standard output anyway. When we do write to another file, it's because
we're creating a commit message template, and these code paths aren't
involved.
But let's be consistent to help future readers and avoid future bugs.
Signed-off-by: Martin Ågren <martin.agren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
`abbrev_sha1_in_line()` uses a `struct object_id oid` and should be
fully prepared to handle non-SHA1 object ids. Rename it to
`abbrev_oid_in_line()`.
A few comments in `wt_status_get_detached_from()` mention "sha1". The
variable they refer to was renamed in e86ab2c1cd ("wt-status: convert to
struct object_id", 2017-02-21). Update the comments to reference "oid"
instead.
Signed-off-by: Martin Ågren <martin.agren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Now that shortlog supports reading from trailers, it can be useful to
combine counts from multiple trailers, or between trailers and authors.
This can be done manually by post-processing the output from multiple
runs, but it's non-trivial to make sure that each name/commit pair is
counted only once.
This patch teaches shortlog to accept multiple --group options on the
command line, and pull data from all of them. That makes it possible to
run:
git shortlog -ns --group=author --group=trailer:co-authored-by
to get a shortlog that counts authors and co-authors equally.
The implementation is mostly straightforward. The "group" enum becomes a
bitfield, and the trailer key becomes a list. I didn't bother
implementing the multi-group semantics for reading from stdin. It would
be possible to do, but the existing matching code makes it awkward, and
I doubt anybody cares.
The duplicate suppression we used for trailers now covers authors and
committers as well (though in non-trailer single-group mode we can skip
the hash insertion and lookup, since we only see one value per commit).
There is one subtlety: we now care about the case when no group bit is
set (in which case we default to showing the author). The caller in
builtin/log.c needs to be adapted to ask explicitly for authors, rather
than relying on shortlog_init(). It would be possible with some
gymnastics to make this keep working as-is, but it's not worth it for a
single caller.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Trailers don't necessarily contain name/email identity values, so
shortlog has so far treated them as opaque strings. However, since many
trailers do contain identities, it's useful to treat them as such when
they can be parsed. That lets "-e" work as usual, as well as mailmap.
When they can't be parsed, we'll continue with the old behavior of
treating them as a single string (there's no new test for that here,
since the existing tests cover a trailer like this).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This function is actually useful for parsing any identity, whether from
stdin or not. We'll need it for handling trailers.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The current documentation is vague about what happens with
--group=trailer:signed-off-by when we see a commit with:
Signed-off-by: One
Signed-off-by: Two
Signed-off-by: One
We clearly should credit both "One" and "Two", but should "One" get
credited twice? The current code does so, but mostly because that was
the easiest thing to do. It's probably more useful to count each commit
at most once. This will become especially important when we allow
values from multiple sources in a future patch.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
If a project uses commit trailers, this patch lets you use
shortlog to see who is performing each action. For example,
running:
git shortlog -ns --group=trailer:reviewed-by
in git.git shows who has reviewed. You can even use a custom
format to see things like who has helped whom:
git shortlog --format="...helped %an (%ad)" \
--group=trailer:helped-by
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The trailer code knows how to parse out the trailers and re-format them,
but there's no easy way to iterate over the trailers (you can use
trailer_info, but you have to then do a bunch of extra parsing).
Let's add an iteration interface that makes this easy to do.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
In preparation for adding more grouping types, let's refactor the
committer/author grouping code and add a user-facing option that binds
them together. In particular:
- the main option is now "--group", to make it clear
that the various group types are mutually exclusive. The
"--committer" option is an alias for "--group=committer".
- we keep an enum rather than a binary flag, to prepare
for more values
- we prefer switch statements to ternary assignment, since
other group types will need more custom code
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The completion tests used that name unnecessarily, and it is a
non-inclusive term, so let's avoid using it here.
Since three of the touched test cases make use of the fact that two of
the branch names (`master` and `maint`) start with the same letter (or
even with the same two letters), we choose to replace the use of
`master` by a name that also has that property: `main`.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The term `master` has a loaded history that serves as a constant
reminder of racial injustice. The Git project has no desire to
perpetuate this and already started avoiding it.
The test suite uses variations of this name for branches other than the
default one. Apart from t3200, where we just addressed this in the
previous commit, those instances can be renamed in an automated manner
because they do not require any changes outside of the test script, so
let's do that.
Seeing as the touched branches have very little (if anything) to do with
the default branch, we choose to use a completely separate naming
scheme: `topic_<number>` (it cannot be `topic-<number>` because t5515
uses the `test_oid` machinery with the term, and that machinery uses
shell variables internally, whose names cannot contain dashes).
This trick was performed by this (GNU) sed invocation:
$ sed -i 's/master\([a-z0-9]\)/topic_\1/g' t/t*.sh
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
21bf933928 (ref-filter: allow merged and no-merged filters, 2020-09-15)
added an early return to reach_filter(). Avoid leaking the memory of a
then unused array by postponing its allocation until we know we need it.
Signed-off-by: René Scharfe <l.s.r@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Currently only the long version (--source=) supports completion.
Add completion support to the short (-s) option too.
Signed-off-by: Ákos Uzonyi <uzonyi.akos@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
In both _git_checkout and _git_switch a new "prevword" variable were
introduced, however the "prev" variable already contains the last word.
The "prevword" variable is replaced with "prev", and the case is moved
to the beginning of the function, like it's done in many other places
(e.g. _git_commit). Also the indentaion of the case is fixed.
Signed-off-by: Ákos Uzonyi <uzonyi.akos@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
"diff-highlight" (in contrib/) had a logic to flush its output upon
seeing a blank line but the way it detected a blank line was broken.
* jk/diff-highlight-blank-match-fix:
diff-highlight: correctly match blank lines for flush
"git push" that wants to be atomic and wants to send push
certificate learned not to prepare and sign the push certificate
when it fails the local check (hence due to atomicity it is known
that no certificate is needed).
* hx/push-atomic-with-cert:
send-pack: run GPG after atomic push checking
The "unshelve" subcommand of "git p4" used incorrectly used
commit^N where it meant to say commit~N to name the Nth generation
ancestor, which has been corrected.
* ld/p4-unshelve-fix:
git-p4: use HEAD~$n to find parent commit for unshelve
git-p4 unshelve: adding a commit breaks git-p4 unshelve
"git receive-pack" that accepts requests by "git push" learned to
outsource most of the ref updates to the new "proc-receive" hook.
* jx/proc-receive-hook:
doc: add documentation for the proc-receive hook
transport: parse report options for tracking refs
t5411: test updates of remote-tracking branches
receive-pack: new config receive.procReceiveRefs
doc: add document for capability report-status-v2
New capability "report-status-v2" for git-push
receive-pack: feed report options to post-receive
receive-pack: add new proc-receive hook
t5411: add basic test cases for proc-receive hook
transport: not report a non-head push as a branch
A "git gc"'s big brother has been introduced to take care of more
repository maintenance tasks, not limited to the object database
cleaning.
* ds/maintenance-part-1:
maintenance: add trace2 regions for task execution
maintenance: add auto condition for commit-graph task
maintenance: use pointers to check --auto
maintenance: create maintenance.<task>.enabled config
maintenance: take a lock on the objects directory
maintenance: add --task option
maintenance: add commit-graph task
maintenance: initialize task array
maintenance: replace run_auto_gc()
maintenance: add --quiet option
maintenance: create basic maintenance runner
The object name written to this file is not exposed to end-users and
the only reader of this file immediately expands it back to a full
object name. Stop abbreviating while writing, and expect a full
object name while reading, which simplifies the code a bit.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Because these constructs can be used to parse user input to be
passed to rev-list --objects, e.g.
range=$(git rev-parse v1.0..v2.0) &&
git rev-list --objects $range | git pack-objects --stdin
the endpoints (v1.0 and v2.0 in the example) are shown without
peeling them to underlying commits, even when they are annotated
tags. Make sure it stays that way.
While at it, ensure "rev-parse A...B" also keeps the endpoints A and
B unpeeled, even though the negative side (i.e. the merge-base
between A and B) has to become a commit.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Protocol v2 became the default in v2.26.0 via 684ceae32d (fetch: default
to protocol version 2, 2019-12-23). More widespread use turned up a
regression in negotiation. That was fixed in v2.27.0 via 4fa3f00abb
(fetch-pack: in protocol v2, in_vain only after ACK, 2020-04-27), but we
also reverted the default to v0 as a precuation in 11c7f2a30b (Revert
"fetch: default to protocol version 2", 2020-04-22).
In v2.28.0, we re-enabled it for experimental users with 3697caf4b9
(config: let feature.experimental imply protocol.version=2, 2020-05-20)
and haven't heard any complaints. v2.28 has only been out for 2 months,
but I'd generally expect people turning on feature.experimental to also
stay pretty up-to-date. So we're not likely to collect much more data by
waiting. In addition, we have no further reports from people running
v2.26.0, and of course some people have been setting protocol.version
manually for ages.
Let's move forward with v2 as the default again. It's possible there are
still lurking bugs, but we won't know until it gets more widespread use.
And we can find and squash them just like any other bug at this point.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
As of recent Visual Studio versions, CMake support is built-in:
https://docs.microsoft.com/en-us/cpp/build/cmake-projects-in-visual-studio?view=vs-2019
All that needs to be done is to open the worktree as a folder, and
Visual Studio will find the `CMakeLists.txt` file and automatically
generate the project files.
Let's ignore the entirety of those generated files.
Helped-by: Đoàn Trần Công Danh <congdanhqx@gmail.com>
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We already match "committer", and we're about to start
matching more things. Let's use a more neutral variable to
avoid confusion.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
In 06f5608c14 (bisect--helper: `bisect_start` shell function
partially in C, 2019-01-02), we changed the following shell
code:
- rev=$(git rev-parse -q --verify "$arg^{commit}") || {
- test $has_double_dash -eq 1 &&
- die "$(eval_gettext "'\$arg' does not appear to be a valid revision")"
- break
- }
- revs="$revs $rev"
into:
+ char *commit_id = xstrfmt("%s^{commit}", arg);
+ if (get_oid(commit_id, &oid) && has_double_dash)
+ die(_("'%s' does not appear to be a valid "
+ "revision"), arg);
+
+ string_list_append(&revs, oid_to_hex(&oid));
+ free(commit_id);
In case of an invalid "arg" when "has_double_dash" is false, the old
code would "break" out of the argument loop.
In the new C code though, `oid_to_hex(&oid)` is unconditonally
appended to "revs". This is wrong first because "oid" is junk as
`get_oid(commit_id, &oid)` failed and second because it doesn't break
out of the argument loop.
Not breaking out of the argument loop means that "arg" is then not
treated as a path restriction (which is wrong).
Signed-off-by: Christian Couder <chriscool@tuxfamily.org>
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
A user who understands enough to set pull.ff does not need additional
instructions.
Signed-off-by: Alex Henrie <alexhenrie24@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The command reads list of object names to place on the ignore list
either from the command line or from a file, but they are not
checked with their object type (those read from the file are not
even checked for object existence).
Extend the oidset_parse_file() API and allow it to take a callback
that can be used to die (e.g. when an inappropriate input is read)
or modify the object name read (e.g. when a tag pointing at a commit
is read, and the caller wants a commit object name), and use it in
the code that handles ignore list.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Only skip diffstats when both oids are valid and identical. This check
was causing both false-positives (files included in diffstats with no
actual changes (0 lines modified) and false-negatives (showing 0 lines
modified in stats when files had actually changed).
Also replaced same_contents with may_differ to avoid confusion.
Signed-off-by: Thomas Guyot-Sionnest <tguyot@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This reverts commit f39ad38410.
That commit was trying to silence a type-punning warning on older
versions of gcc. However, its analysis was all wrong. I didn't notice
that we _were_ in fact type-punning because there are two versions of
put_be32(): one that uses casts and unaligned loads, and another that
uses bitshifts. I looked at the latter, but on my platform we were
defaulting to the former.
However, as of the previous commit, we'll always use the bitshift
version. So we can drop this hackery to avoid the warning, making the
code slightly cleaner.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Our put_be32() routine and its variants (get_be32(), put_be64(), etc)
has two implementations: on some platforms we cast memory in place and
use nothl()/htonl(), which can cause unaligned memory access. And on
others, we pick out the individual bytes using bitshifts.
This introduces extra complexity, and sometimes causes compilers to
generate warnings about type-punning. And it's not clear there's any
performance advantage.
This split goes back to 660231aa97 (block-sha1: support for
architectures with memory alignment restrictions, 2009-08-12). The
unaligned versions were part of the original block-sha1 code in
d7c208a92e (Add new optimized C 'block-sha1' routines, 2009-08-05),
which says it is:
Based on the mozilla SHA1 routine, but doing the input data accesses a
word at a time and with 'htonl()' instead of loading bytes and shifting.
Back then, Linus provided timings versus the mozilla code which showed a
27% improvement:
https://lore.kernel.org/git/alpine.LFD.2.01.0908051545000.3390@localhost.localdomain/
However, the unaligned loads were either not the useful part of that
speedup, or perhaps compilers and processors have changed since then.
Here are times for computing the sha1 of 4GB of random data, with and
without -DNO_UNALIGNED_LOADS (and BLK_SHA1=1, of course). This is with
gcc 10, -O2, and the processor is a Core i9-9880H.
[stock]
Benchmark #1: t/helper/test-tool sha1 <foo.rand
Time (mean ± σ): 6.638 s ± 0.081 s [User: 6.269 s, System: 0.368 s]
Range (min … max): 6.550 s … 6.841 s 10 runs
[-DNO_UNALIGNED_LOADS]
Benchmark #1: t/helper/test-tool sha1 <foo.rand
Time (mean ± σ): 6.418 s ± 0.015 s [User: 6.058 s, System: 0.360 s]
Range (min … max): 6.394 s … 6.447 s 10 runs
And here's the same test run on an AMD A8-7600, using gcc 8.
[stock]
Benchmark #1: t/helper/test-tool sha1 <foo.rand
Time (mean ± σ): 11.721 s ± 0.113 s [User: 10.761 s, System: 0.951 s]
Range (min … max): 11.509 s … 11.861 s 10 runs
[-DNO_UNALIGNED_LOADS]
Benchmark #1: t/helper/test-tool sha1 <foo.rand
Time (mean ± σ): 11.744 s ± 0.066 s [User: 10.807 s, System: 0.928 s]
Range (min … max): 11.637 s … 11.863 s 10 runs
So the unaligned loads don't seem to help much, and actually make things
worse. It's possible there are platforms where they provide more
benefit, but:
- the non-x86 platforms for which we use this code are old and obscure
(powerpc and s390).
- the main caller that cares about performance is block-sha1. But
these days it is rarely used anyway, in favor of sha1dc (which is
already much slower, and nobody seems to have cared that much).
Let's just drop unaligned versions entirely in the name of simplicity.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
