From 72441af7c4e3bde33cdf7edafcf09c227d5d5296 Mon Sep 17 00:00:00 2001 From: Kirill Smelkov Date: Mon, 7 Apr 2014 01:46:26 +0400 Subject: tree-diff: rework diff_tree() to generate diffs for multiparent cases as well MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Previously diff_tree(), which is now named ll_diff_tree_sha1(), was generating diff_filepair(s) for two trees t1 and t2, and that was usually used for a commit as t1=HEAD~, and t2=HEAD - i.e. to see changes a commit introduces. In Git, however, we have fundamentally built flexibility in that a commit can have many parents - 1 for a plain commit, 2 for a simple merge, but also more than 2 for merging several heads at once. For merges there is a so called combine-diff, which shows diff, a merge introduces by itself, omitting changes done by any parent. That works through first finding paths, that are different to all parents, and then showing generalized diff, with separate columns for +/- for each parent. The code lives in combine-diff.c . There is an impedance mismatch, however, in that a commit could generally have any number of parents, and that while diffing trees, we divide cases for 2-tree diffs and more-than-2-tree diffs. I mean there is no special casing for multiple parents commits in e.g. revision-walker . That impedance mismatch *hurts* *performance* *badly* for generating combined diffs - in "combine-diff: optimize combine_diff_path sets intersection" I've already removed some slowness from it, but from the timings provided there, it could be seen, that combined diffs still cost more than an order of magnitude more cpu time, compared to diff for usual commits, and that would only be an optimistic estimate, if we take into account that for e.g. linux.git there is only one merge for several dozens of plain commits. That slowness comes from the fact that currently, while generating combined diff, a lot of time is spent computing diff(commit,commit^2) just to only then intersect that huge diff to almost small set of files from diff(commit,commit^1). That's because at present, to compute combine-diff, for first finding paths, that "every parent touches", we use the following combine-diff property/definition: D(A,P1...Pn) = D(A,P1) ^ ... ^ D(A,Pn) (w.r.t. paths) where D(A,P1...Pn) is combined diff between commit A, and parents Pi and D(A,Pi) is usual two-tree diff Pi..A So if any of that D(A,Pi) is huge, tracting 1 n-parent combine-diff as n 1-parent diffs and intersecting results will be slow. And usually, for linux.git and other topic-based workflows, that D(A,P2) is huge, because, if merge-base of A and P2, is several dozens of merges (from A, via first parent) below, that D(A,P2) will be diffing sum of merges from several subsystems to 1 subsystem. The solution is to avoid computing n 1-parent diffs, and to find changed-to-all-parents paths via scanning A's and all Pi's trees simultaneously, at each step comparing their entries, and based on that comparison, populate paths result, and deduce we could *skip* *recursing* into subdirectories, if at least for 1 parent, sha1 of that dir tree is the same as in A. That would save us from doing significant amount of needless work. Such approach is very similar to what diff_tree() does, only there we deal with scanning only 2 trees simultaneously, and for n+1 tree, the logic is a bit more complex: D(T,P1...Pn) calculation scheme ------------------------------- D(T,P1...Pn) = D(T,P1) ^ ... ^ D(T,Pn) (regarding resulting paths set) D(T,Pj) - diff between T..Pj D(T,P1...Pn) - combined diff from T to parents P1,...,Pn We start from all trees, which are sorted, and compare their entries in lock-step: T P1 Pn - - - |t| |p1| |pn| |-| |--| ... |--| imin = argmin(p1...pn) | | | | | | |-| |--| |--| |.| |. | |. | . . . . . . at any time there could be 3 cases: 1) t < p[imin]; 2) t > p[imin]; 3) t = p[imin]. Schematic deduction of what every case means, and what to do, follows: 1) t < p[imin] -> ∀j t ∉ Pj -> "+t" ∈ D(T,Pj) -> D += "+t"; t↓ 2) t > p[imin] 2.1) ∃j: pj > p[imin] -> "-p[imin]" ∉ D(T,Pj) -> D += ø; ∀ pi=p[imin] pi↓ 2.2) ∀i pi = p[imin] -> pi ∉ T -> "-pi" ∈ D(T,Pi) -> D += "-p[imin]"; ∀i pi↓ 3) t = p[imin] 3.1) ∃j: pj > p[imin] -> "+t" ∈ D(T,Pj) -> only pi=p[imin] remains to investigate 3.2) pi = p[imin] -> investigate δ(t,pi) | | v 3.1+3.2) looking at δ(t,pi) ∀i: pi=p[imin] - if all != ø -> ⎧δ(t,pi) - if pi=p[imin] -> D += ⎨ ⎩"+t" - if pi>p[imin] in any case t↓ ∀ pi=p[imin] pi↓ ~ For comparison, here is how diff_tree() works: D(A,B) calculation scheme ------------------------- A B - - |a| |b| a < b -> a ∉ B -> D(A,B) += +a a↓ |-| |-| a > b -> b ∉ A -> D(A,B) += -b b↓ | | | | a = b -> investigate δ(a,b) a↓ b↓ |-| |-| |.| |.| . . . . ~~~~~~~~ This patch generalizes diff tree-walker to work with arbitrary number of parents as described above - i.e. now there is a resulting tree t, and some parents trees tp[i] i=[0..nparent). The generalization builds on the fact that usual diff D(A,B) is by definition the same as combined diff D(A,[B]), so if we could rework the code for common case and make it be not slower for nparent=1 case, usual diff(t1,t2) generation will not be slower, and multiparent diff tree-walker would greatly benefit generating combine-diff. What we do is as follows: 1) diff tree-walker ll_diff_tree_sha1() is internally reworked to be a paths generator (new name diff_tree_paths()), with each generated path being `struct combine_diff_path` with info for path, new sha1,mode and for every parent which sha1,mode it was in it. 2) From that info, we can still generate usual diff queue with struct diff_filepairs, via "exporting" generated combine_diff_path, if we know we run for nparent=1 case. (see emit_diff() which is now named emit_diff_first_parent_only()) 3) In order for diff_can_quit_early(), which checks DIFF_OPT_TST(opt, HAS_CHANGES)) to work, that exporting have to be happening not in bulk, but incrementally, one diff path at a time. For such consumers, there is a new callback in diff_options introduced: ->pathchange(opt, struct combine_diff_path *) which, if set to !NULL, is called for every generated path. (see new compat ll_diff_tree_sha1() wrapper around new paths generator for setup) 4) The paths generation itself, is reworked from previous ll_diff_tree_sha1() code according to "D(A,P1...Pn) calculation scheme" provided above: On the start we allocate [nparent] arrays in place what was earlier just for one parent tree. then we just generalize loops, and comparison according to the algorithm. Some notes(*): 1) alloca(), for small arrays, is used for "runs not slower for nparent=1 case than before" goal - if we change it to xmalloc()/free() the timings get ~1% worse. For alloca() we use just-introduced xalloca/xalloca_free compatibility wrappers, so it should not be a portability problem. 2) For every parent tree, we need to keep a tag, whether entry from that parent equals to entry from minimal parent. For performance reasons I'm keeping that tag in entry's mode field in unused bit - see S_IFXMIN_NEQ. Not doing so, we'd need to alloca another [nparent] array, which hurts performance. 3) For emitted paths, memory could be reused, if we know the path was processed via callback and will not be needed later. We use efficient hand-made realloc-style path_appendnew(), that saves us from ~1-1.5% of potential additional slowdown. 4) goto(s) are used in several places, as the code executes a little bit faster with lowered register pressure. Also - we should now check for FIND_COPIES_HARDER not only when two entries names are the same, and their hashes are equal, but also for a case, when a path was removed from some of all parents having it. The reason is, if we don't, that path won't be emitted at all (see "a > xi" case), and we'll just skip it, and FIND_COPIES_HARDER wants all paths - with diff or without - to be emitted, to be later analyzed for being copies sources. The new check is only necessary for nparent >1, as for nparent=1 case xmin_eqtotal always =1 =nparent, and a path is always added to diff as removal. ~~~~~~~~ Timings for # without -c, i.e. testing only nparent=1 case `git log --raw --no-abbrev --no-renames` before and after the patch are as follows: navy.git linux.git v3.10..v3.11 before 0.611s 1.889s after 0.619s 1.907s slowdown 1.3% 0.9% This timings show we did no harm to usual diff(tree1,tree2) generation. From the table we can see that we actually did ~1% slowdown, but I think I've "earned" that 1% in the previous patch ("tree-diff: reuse base str(buf) memory on sub-tree recursion", HEAD~~) so for nparent=1 case, net timings stays approximately the same. The output also stayed the same. (*) If we revert 1)-4) to more usual techniques, for nparent=1 case, we'll get ~2-2.5% of additional slowdown, which I've tried to avoid, as "do no harm for nparent=1 case" rule. For linux.git, combined diff will run an order of magnitude faster and appropriate timings will be provided in the next commit, as we'll be taking advantage of the new diff tree-walker for combined-diff generation there. P.S. and combined diff is not some exotic/for-play-only stuff - for example for a program I write to represent Git archives as readonly filesystem, there is initial scan with `git log --reverse --raw --no-abbrev --no-renames -c` to extract log of what was created/changed when, as a result building a map {} sha1 -> in which commit (and date) a content was added that `-c` means also show combined diff for merges, and without them, if a merge is non-trivial (merges changes from two parents with both having separate changes to a file), or an evil one, the map will not be full, i.e. some valid sha1 would be absent from it. That case was my initial motivation for combined diffs speedup. Signed-off-by: Kirill Smelkov Signed-off-by: Junio C Hamano --- diff.c | 1 + 1 file changed, 1 insertion(+) (limited to 'diff.c') diff --git a/diff.c b/diff.c index 8e4a6a9105..cda4aa8f85 100644 --- a/diff.c +++ b/diff.c @@ -3216,6 +3216,7 @@ void diff_setup(struct diff_options *options) options->context = diff_context_default; DIFF_OPT_SET(options, RENAME_EMPTY); + /* pathchange left =NULL by default */ options->change = diff_change; options->add_remove = diff_addremove; options->use_color = diff_use_color_default; -- cgit v1.2.3 From 7195fbfaf5a539b8e8358097e02b63991e78a565 Mon Sep 17 00:00:00 2001 From: Kirill Smelkov Date: Mon, 24 Feb 2014 20:21:51 +0400 Subject: combine-diff: speed it up, by using multiparent diff tree-walker directly As was recently shown in "combine-diff: optimize combine_diff_path sets intersection", combine-diff runs very slowly. In that commit we optimized paths sets intersection, but that accounted only for ~ 25% of the slowness, and as my tracing showed, for linux.git v3.10..v3.11, for merges a lot of time is spent computing diff(commit,commit^2) just to only then intersect that huge diff to almost small set of files from diff(commit,commit^1). In previous commit, we described the problem in more details, and reworked the diff tree-walker to be general one - i.e. to work in multiple parent case too. Now is the time to take advantage of it for finding paths for combine diff. The implementation is straightforward - if we know, we can get generated diff paths directly, and at present that means no diff filtering or rename/copy detection was requested(*), we can call multiparent tree-walker directly and get ready paths. (*) because e.g. at present, all diffcore transformations work on diff_filepair queues, but in the future, that limitation can be lifted, if filters would operate directly on combine_diff_paths. Timings for `git log --raw --no-abbrev --no-renames` without `-c` ("git log") and with `-c` ("git log -c") and with `-c --merges` ("git log -c --merges") before and after the patch are as follows: linux.git v3.10..v3.11 log log -c log -c --merges before 1.9s 16.4s 15.2s after 1.9s 2.4s 1.1s The result stayed the same. Signed-off-by: Kirill Smelkov Signed-off-by: Junio C Hamano --- combine-diff.c | 88 ++++++++++++++++++++++++++++++++++++++++++++++++++++++---- diff.c | 1 + 2 files changed, 84 insertions(+), 5 deletions(-) (limited to 'diff.c') diff --git a/combine-diff.c b/combine-diff.c index 1732dfd110..12764fb733 100644 --- a/combine-diff.c +++ b/combine-diff.c @@ -1303,7 +1303,7 @@ static const char *path_path(void *obj) /* find set of paths that every parent touches */ -static struct combine_diff_path *find_paths(const unsigned char *sha1, +static struct combine_diff_path *find_paths_generic(const unsigned char *sha1, const struct sha1_array *parents, struct diff_options *opt) { struct combine_diff_path *paths = NULL; @@ -1316,6 +1316,7 @@ static struct combine_diff_path *find_paths(const unsigned char *sha1, /* tell diff_tree to emit paths in sorted (=tree) order */ opt->orderfile = NULL; + /* D(A,P1...Pn) = D(A,P1) ^ ... ^ D(A,Pn) (wrt paths) */ for (i = 0; i < num_parent; i++) { /* * show stat against the first parent even when doing @@ -1346,6 +1347,35 @@ static struct combine_diff_path *find_paths(const unsigned char *sha1, } +/* + * find set of paths that everybody touches, assuming diff is run without + * rename/copy detection, etc, comparing all trees simultaneously (= faster). + */ +static struct combine_diff_path *find_paths_multitree( + const unsigned char *sha1, const struct sha1_array *parents, + struct diff_options *opt) +{ + int i, nparent = parents->nr; + const unsigned char **parents_sha1; + struct combine_diff_path paths_head; + struct strbuf base; + + parents_sha1 = xmalloc(nparent * sizeof(parents_sha1[0])); + for (i = 0; i < nparent; i++) + parents_sha1[i] = parents->sha1[i]; + + /* fake list head, so worker can assume it is non-NULL */ + paths_head.next = NULL; + + strbuf_init(&base, PATH_MAX); + diff_tree_paths(&paths_head, sha1, parents_sha1, nparent, &base, opt); + + strbuf_release(&base); + free(parents_sha1); + return paths_head.next; +} + + void diff_tree_combined(const unsigned char *sha1, const struct sha1_array *parents, int dense, @@ -1355,6 +1385,7 @@ void diff_tree_combined(const unsigned char *sha1, struct diff_options diffopts; struct combine_diff_path *p, *paths; int i, num_paths, needsep, show_log_first, num_parent = parents->nr; + int need_generic_pathscan; /* nothing to do, if no parents */ if (!num_parent) @@ -1377,11 +1408,58 @@ void diff_tree_combined(const unsigned char *sha1, /* find set of paths that everybody touches * - * NOTE find_paths() also handles --stat, as it computes - * diff(sha1,parent_i) for all i to do the job, specifically - * for parent0. + * NOTE + * + * Diffcore transformations are bound to diff_filespec and logic + * comparing two entries - i.e. they do not apply directly to combine + * diff. + * + * If some of such transformations is requested - we launch generic + * path scanning, which works significantly slower compared to + * simultaneous all-trees-in-one-go scan in find_paths_multitree(). + * + * TODO some of the filters could be ported to work on + * combine_diff_paths - i.e. all functionality that skips paths, so in + * theory, we could end up having only multitree path scanning. + * + * NOTE please keep this semantically in sync with diffcore_std() */ - paths = find_paths(sha1, parents, &diffopts); + need_generic_pathscan = opt->skip_stat_unmatch || + DIFF_OPT_TST(opt, FOLLOW_RENAMES) || + opt->break_opt != -1 || + opt->detect_rename || + opt->pickaxe || + opt->filter; + + + if (need_generic_pathscan) { + /* + * NOTE generic case also handles --stat, as it computes + * diff(sha1,parent_i) for all i to do the job, specifically + * for parent0. + */ + paths = find_paths_generic(sha1, parents, &diffopts); + } + else { + int stat_opt; + paths = find_paths_multitree(sha1, parents, &diffopts); + + /* + * show stat against the first parent even + * when doing combined diff. + */ + stat_opt = (opt->output_format & + (DIFF_FORMAT_NUMSTAT|DIFF_FORMAT_DIFFSTAT)); + if (stat_opt) { + diffopts.output_format = stat_opt; + + diff_tree_sha1(parents->sha1[0], sha1, "", &diffopts); + diffcore_std(&diffopts); + if (opt->orderfile) + diffcore_order(opt->orderfile); + diff_flush(&diffopts); + } + } /* find out number of surviving paths */ for (num_paths = 0, p = paths; p; p = p->next) diff --git a/diff.c b/diff.c index cda4aa8f85..f2fff4667d 100644 --- a/diff.c +++ b/diff.c @@ -4764,6 +4764,7 @@ void diffcore_fix_diff_index(struct diff_options *options) void diffcore_std(struct diff_options *options) { + /* NOTE please keep the following in sync with diff_tree_combined() */ if (options->skip_stat_unmatch) diffcore_skip_stat_unmatch(options); if (!options->found_follow) { -- cgit v1.2.3