diffcore-rename improvements

This patch series include 4 small code cleanups, 1 small bug fix (or perhaps just a UI improvement -- it's a very minor issue either way), and 2 performance improvements. The first 6 patches are relatively easy to review, while the last one...is a bit more involved (but provided the biggest performance boost).

These patches came out of the merge-ort work, and are related, but there is no dependency between any of the series.

Elijah Newren (7):
diffcore-rename: avoid usage of global in too_many_rename_candidates()
diffcore-rename: remove unnecessary if-clause
diffcore-rename: rename num_create to num_targets
diffcore-rename: change a few comments to use 'add' instead of
'create'
diffcore-rename: reduce jumpiness in progress counters
diffcore-rename: simplify and accelerate register_rename_src() Accelerate rename_dst setup

 diffcore-rename.c | 225 ++++++++++++++++++++--------------------------
 1 file changed, 96 insertions(+), 129 deletions(-)

base-commit: 3a0b884caba2752da0af626fb2de7d597c844e8b
Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-git-929%2Fnewren%2Fdiffcore-rename-improvements-v1
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-git-929/newren/diffcore-rename-improvements-v1
Pull-Request: https://github.com/git/git/pull/929

gitgitgadget

From: Elijah Newren <[email protected]>

too_many_rename_candidates() got the number of rename targets via an argument to the function, but the number of rename sources via a global variable. That felt rather inconsistent. Pass in the number of rename sources as an argument as well.

While we are at it... We had a local variable, num_src, that served two purposes. Initially it was set to this global value, but later was used for counting a subset of the number of sources. Since we now have a function argument for the former usage, introduce a clearer variable name for the latter usage.

This patch has no behavioral changes; it's just renaming and passing an argument instead of grabbing it from the global namespace. (You may find it easier to view the patch using git diff's --color-words option.)

---
 diffcore-rename.c | 23 +++++++++++------------
 1 file changed, 11 insertions(+), 12 deletions(-)

index d367a6d2443..68ddf51a2a1 100644
@@ -434,12 +434,11 @@ static void record_if_better(struct diff_score m[], struct diff_score *o)
  * 1 if we need to disable inexact rename detection;
  * 2 if we would be under the limit if we were given -C instead of -C -C.
  */
-static int too_many_rename_candidates(int num_create,
+static int too_many_rename_candidates(int num_targets, int num_sources,
 				      struct diff_options *options)
 {
 	int rename_limit = options->rename_limit;
-	int num_src = rename_src_nr;
-	int i;
+	int i, limited_sources;
 
 	options->needed_rename_limit = 0;
 

@@ -447,30 +446,30 @@ static int too_many_rename_candidates(int num_create,
 	 * This basically does a test for the rename matrix not
 	 * growing larger than a "rename_limit" square matrix, ie:
 	 *
-	 *    num_create * num_src > rename_limit * rename_limit
+	 *    num_targets * num_sources > rename_limit * rename_limit
 	 */
 	if (rename_limit <= 0)
 		rename_limit = 32767;
-	if ((num_create <= rename_limit || num_src <= rename_limit) &&
-	    ((uint64_t)num_create * (uint64_t)num_src
+	if ((num_targets <= rename_limit || num_sources <= rename_limit) &&
+	    ((uint64_t)num_targets * (uint64_t)num_sources
 	     <= (uint64_t)rename_limit * (uint64_t)rename_limit))
 		return 0;
 
 	options->needed_rename_limit =
-		num_src > num_create ? num_src : num_create;
+		num_sources > num_targets ? num_sources : num_targets;
 
 	/* Are we running under -C -C? */
 	if (!options->flags.find_copies_harder)
 		return 1;
 
 	/* Would we bust the limit if we were running under -C? */
-	for (num_src = i = 0; i < rename_src_nr; i++) {
+	for (limited_sources = i = 0; i < num_sources; i++) {
 		if (diff_unmodified_pair(rename_src[i].p))
 			continue;
-		num_src++;
+		limited_sources++;
 	}
-	if ((num_create <= rename_limit || num_src <= rename_limit) &&
-	    ((uint64_t)num_create * (uint64_t)num_src
+	if ((num_targets <= rename_limit || limited_sources <= rename_limit) &&
+	    ((uint64_t)num_targets * (uint64_t)limited_sources
 	     <= (uint64_t)rename_limit * (uint64_t)rename_limit))
 		return 2;
 	return 1;

@@ -576,7 +575,7 @@ void diffcore_rename(struct diff_options *options)
 	if (!num_create)
 		goto cleanup;
 
-	switch (too_many_rename_candidates(num_create, options)) {
+	switch (too_many_rename_candidates(num_create, rename_src_nr, options)) {
 	case 1:
 		goto cleanup;
 	case 2:

gitgitgadget

From: Elijah Newren <[email protected]>

diffcore-rename had two different checks of the form

if ((a < limit || b < limit) &&
a * b <= limit * limit)

Since these are all non-negative integers, this can be simplified to

if (a * b <= limit * limit)

The only advantage of the former would be in avoiding a couple multiplications in the rare case that both a and b are BOTH very large. I see no reason for such an optimization given that this code is not in any kind of loop. Prefer code simplicity here and change to the latter form.

---
 diffcore-rename.c | 10 ++++------
 1 file changed, 4 insertions(+), 6 deletions(-)

index 68ddf51a2a1..0f8fce9293e 100644
@@ -450,9 +450,8 @@ static int too_many_rename_candidates(int num_targets, int num_sources,
 	 */
 	if (rename_limit <= 0)
 		rename_limit = 32767;
-	if ((num_targets <= rename_limit || num_sources <= rename_limit) &&
-	    ((uint64_t)num_targets * (uint64_t)num_sources
-	     <= (uint64_t)rename_limit * (uint64_t)rename_limit))
+	if ((uint64_t)num_targets * (uint64_t)num_sources
+	    <= (uint64_t)rename_limit * (uint64_t)rename_limit)
 		return 0;
 
 	options->needed_rename_limit =

@@ -468,9 +467,8 @@ static int too_many_rename_candidates(int num_targets, int num_sources,
 			continue;
 		limited_sources++;
 	}
-	if ((num_targets <= rename_limit || limited_sources <= rename_limit) &&
-	    ((uint64_t)num_targets * (uint64_t)limited_sources
-	     <= (uint64_t)rename_limit * (uint64_t)rename_limit))
+	if ((uint64_t)num_targets * (uint64_t)limited_sources
+	    <= (uint64_t)rename_limit * (uint64_t)rename_limit)
 		return 2;
 	return 1;
 }

gitgitgadget

Makes sense.

If you were really paranoid, you could perform these checks with unsigned_mult_overflows(), but I don't think that it's worth doing so here.

Yeah, I think that reads slightly better too, but on the other hand it does make the diff slightly harder to read. *shrug*.

The right hand side of the expression is the same -- rename_limit * rename_limit -- so it could be stored off (though I don't think there's much point in doing so unless it made the code clearer; this is not remotely close to a hot path). However, in the left hand side, it's not so much that one of the variables has changed since the last check, it's that it uses a different set of variables in the check. In particular, it replaces 'num_sources' with 'limited_sources'.

I've always assumed that the original was for correctness (if a and b are both larger than limit, a*b could end up being smaller than limit*limit when the result of multiplication of the former wraps around but not the latter) ...

... and in no way as an optimization.

So, I dunno.

Ah, so would you be okay replacing these with

if (st_mult(num_targets, limited_sources) <= st_mult(rename_limit,
rename_limit))

?

That'd make the correctness intent far clearer, and allow us to drop the casting as well since st_mult converts to size_t. (If, on the off chance you're on a platform where size_t is 32-bits and the multiplications don't fit in that size, then the requested matrices for computing rename detection won't fit in memory for you.)

It's subtle that the use of size_t is a sensible thing to do, and definitely deserves an in-line comment, but I think that is a good way to go.

Hi Christian, Kaartic, and Sangeeta,

If I've heard correctly, you three are now interested in rename detection and thus diffcore-rename.c. You might find this patch series interesting; even if you don't want to review the details, just looking over some of the commit messages might help orient you to the codebase that you'll be working with in the upcoming months.

From: Elijah Newren <[email protected]>

Files added since the base commit serve as targets for rename detection. While it is true that added files can be thought of as being "created" when they are added IF they have no pairing file that they were renamed from, and it is true we start out not knowing what the pairings are, it seems a little odd to think in terms of "file creation" when we are looking for "file renames". Rename the variable to avoid this minor point of confusion.

---
 diffcore-rename.c | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)

index 0f8fce9293e..655a67759c8 100644
@@ -502,7 +502,7 @@ void diffcore_rename(struct diff_options *options)
 	struct diff_queue_struct outq;
 	struct diff_score *mx;
 	int i, j, rename_count, skip_unmodified = 0;
-	int num_create, dst_cnt;
+	int num_targets, dst_cnt;
 	struct progress *progress = NULL;
 
 	if (!minimum_score)

@@ -567,13 +567,13 @@ void diffcore_rename(struct diff_options *options)
 	 * Calculate how many renames are left (but all the source
 	 * files still remain as options for rename/copies!)
 	 */
-	num_create = (rename_dst_nr - rename_count);
+	num_targets = (rename_dst_nr - rename_count);
 
 	/* All done? */
-	if (!num_create)
+	if (!num_targets)
 		goto cleanup;
 
-	switch (too_many_rename_candidates(num_create, rename_src_nr, options)) {
+	switch (too_many_rename_candidates(num_targets, rename_src_nr, options)) {
 	case 1:
 		goto cleanup;
 	case 2:

@@ -590,7 +590,7 @@ void diffcore_rename(struct diff_options *options)
 				(uint64_t)rename_dst_nr * (uint64_t)rename_src_nr);
 	}
 
-	mx = xcalloc(st_mult(NUM_CANDIDATE_PER_DST, num_create), sizeof(*mx));
+	mx = xcalloc(st_mult(NUM_CANDIDATE_PER_DST, num_targets), sizeof(*mx));
 	for (dst_cnt = i = 0; i < rename_dst_nr; i++) {
 		struct diff_filespec *two = rename_dst[i].two;
 		struct diff_score *m;

gitgitgadget

This is probably subjective.

I've always viewed the rename detection as first collecting a set of deleted paths and a set of created paths, and then trying to find a good mapping from the former into the latter, so I find num_create a lot more intuitive than num_targets.

But the remaining elements in the latter set are counted in the counter "rename_dst_nr", so we clearly are OK to call the elements of the latter set "the destination" (of a rename), which contrasts very well with "the source" (of a rename), which is how the deleted paths are counted with rename_src_nr.

When doing -C and -C -C, the "source" set has not just deleted but also the preimage of the modified paths, so "source" is a more appropriate name than "delete". From that point of view, "destination" is a more appropriate name for "create" because it contrasts well with "source".

You silently renamed num_create to num_targets in 1/7 without justification while adding num_sources. Perhaps we should go back to that step and use num_destinations to match? The result would be using words <dst, src> that pair with each other much better than introducing "target" to an existing mix of <create==dst, src> to make it <target==dst, src>.

Ooh, I like it. num_destinations it is.

From: Elijah Newren <[email protected]>

As with the last commit, it feels odd to talk about "creation" (which to me implies a brand new file, i.e. one not associated with an existing file, and thus one that could not have been a rename) when we are using the files to look for renames. Use the term "addition" instead.

---
 diffcore-rename.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

index 655a67759c8..7270eb6af48 100644
@@ -173,7 +173,7 @@ static int estimate_similarity(struct repository *r,
 {
 	/* src points at a file that existed in the original tree (or
 	 * optionally a file in the destination tree) and dst points
-	 * at a newly created file.  They may be quite similar, in which
+	 * at a newly added file.  They may be quite similar, in which
 	 * case we want to say src is renamed to dst or src is copied into
 	 * dst, and then some edit has been applied to dst.
 	 *

@@ -652,9 +652,9 @@ void diffcore_rename(struct diff_options *options)
 		}
 		else if (!DIFF_FILE_VALID(p->one) && DIFF_FILE_VALID(p->two)) {
 			/*
-			 * Creation
+			 * Addition
 			 *
-			 * We would output this create record if it has
+			 * We would output this add record if it has
 			 * not been turned into a rename/copy already.
 			 */
 			struct diff_rename_dst *dst = locate_rename_dst(p->two);

@@ -664,7 +664,7 @@ void diffcore_rename(struct diff_options *options)
 			}
 			else
 				/* no matching rename/copy source, so
-				 * record this as a creation.
+				 * record this as an addition.
 				 */
 				diff_q(&outq, p);
 		}

gitgitgadget

Hmph, to me "create" and "add" mean the same thing in all the sentences this patch updates.

We first notice the differences in two trees, some paths have contents in both trees so they cannot be creation or deletion. Some paths only exist in one of the trees and they are initially labelled as "created" or "deleted". If there were no request by the caller to detect renames, then that's the end of the story.

Only after we match up a path in the new tree that did not exist in the old tree with another path in the old tree that no longer exists in the new tree, what initially got labelled as creation of one and deletion of another may become a rename.

And in the above two paragraphs, if I replace "to create" with "to add" and "to delete" with "to remove", I do not see how the story becomes more natural or logical or intuitive. Perhaps I am not getting the subtle difference you find in these two words? This looks like a totally meaningless churn to me.

From: Elijah Newren <[email protected]>

Inexact rename detection works by comparing all sources to all destinations, computing similarities, and then finding the best matches among those that are sufficiently similar.

However, it is preceded by exact rename detection that works by checking if there are files with identical hashes. If exact renames are found, we can exclude some files from inexact rename detection.

The inexact rename detection loops over the full set of files, but immediately skips those for which rename_dst[i].is_rename is true and thus doesn't compare any sources to that destination. As such, these paths shouldn't be included in the progress counter.

For the eagle eyed, this change hints at an actual optimization -- the first one I presented at Git Merge 2020. I'll be submitting that optimization later, once the basic merge-ort algorithm has merged.

---
 diffcore-rename.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

index 7270eb6af48..3d637ba4645 100644
@@ -587,7 +587,7 @@ void diffcore_rename(struct diff_options *options)
 	if (options->show_rename_progress) {
 		progress = start_delayed_progress(
 				_("Performing inexact rename detection"),
-				(uint64_t)rename_dst_nr * (uint64_t)rename_src_nr);
+				(uint64_t)num_targets * (uint64_t)rename_src_nr);
 	}
 
 	mx = xcalloc(st_mult(NUM_CANDIDATE_PER_DST, num_targets), sizeof(*mx));

@@ -626,7 +626,8 @@ void diffcore_rename(struct diff_options *options)
 			diff_free_filespec_blob(two);
 		}
 		dst_cnt++;
-		display_progress(progress, (uint64_t)(i+1)*(uint64_t)rename_src_nr);
+		display_progress(progress,
+				 (uint64_t)dst_cnt * (uint64_t)rename_src_nr);
 	}
 	stop_progress(&progress);
 

gitgitgadget

;-) I was hoping that that was the case.

Here, you are using <sources, destinations> as contrasting pair of words, which supports my comment on 1/7 and 3/7 ;-)

From: Elijah Newren <[email protected]>

reigster_rename_src() took pains to create an array in rename_src which was sorted by pathname of the contained diff_filepair. However, the fact that this array was sorted was not needed anywhere, and thus represented wasted time. Simply append to the end of the array, which in a usecase of note saved 45% of diffcore_rename() setup time for me.

Technically, there is one difference in the end result for the code. IF the caller of diffcore-rename provides a set of pairs with a duplicate filename in the sources (an erroneous input condition), the old code would simply ignore the duplicate (without warning or error). The new code will simply swallow it and thus allow multiple pairings for the "same" source file. Checking for this error condition is expensive (basically undoing the optimization) and the only existing callers already avoid passing such an invalid input. If someone really wants to add some error checking here on this input condition, I believe they should add a separate function which can be called before diffcore_rename(), and then other callers that don't need additional checks because of their design (such as merge-ort) can avoid the extra overhead.

Also, note that I dropped the return type on the function since it was unconditionally discarded anyway.

This patch is being submitted in a different order than its original development, but in a large rebase of many commits with lots of renames and with several optimizations to inexact rename detection, diffcore_rename() setup time was a sizeable chunk of overall runtime. This patch dropped execution time of rebasing 35 commits with lots of renames by 2% overall.

---
 diffcore-rename.c | 30 +++---------------------------
 1 file changed, 3 insertions(+), 27 deletions(-)

index 3d637ba4645..816d2fbac44 100644
@@ -76,36 +76,12 @@ static struct diff_rename_src {
 } *rename_src;
 static int rename_src_nr, rename_src_alloc;
 
-static struct diff_rename_src *register_rename_src(struct diff_filepair *p)
+static void register_rename_src(struct diff_filepair *p)
 {
-	int first, last;
-	struct diff_filespec *one = p->one;
-	unsigned short score = p->score;
-
-	first = 0;
-	last = rename_src_nr;
-	while (last > first) {
-		int next = first + ((last - first) >> 1);
-		struct diff_rename_src *src = &(rename_src[next]);
-		int cmp = strcmp(one->path, src->p->one->path);
-		if (!cmp)
-			return src;
-		if (cmp < 0) {
-			last = next;
-			continue;
-		}
-		first = next+1;
-	}
-
-	/* insert to make it at "first" */
 	ALLOC_GROW(rename_src, rename_src_nr + 1, rename_src_alloc);
+	rename_src[rename_src_nr].p = p;
+	rename_src[rename_src_nr].score = p->score;
 	rename_src_nr++;
-	if (first < rename_src_nr)
-		MOVE_ARRAY(rename_src + first + 1, rename_src + first,
-			   rename_src_nr - first - 1);
-	rename_src[first].p = p;
-	rename_src[first].score = score;
-	return &(rename_src[first]);
 }
 
 static int basename_same(struct diff_filespec *src, struct diff_filespec *dst)

gitgitgadget

s/IF/If ?

It seems like this is currently impossible to trigger, making any extra (expensive) checking of it worthless, no?

Neat!

Indeed; will fix.

I believe that's what it currently amounts to, and I debated just ripping the paragraph out.

However, a natural question that can easily arise for current reviewers or future readers of the patch is why was there ever sorting in the first place if the sorting isn't used? That question came up for me, and I dug into it. Sorting is also used with rename_dst in nearby code in the file, and there are a few reasons for it there. The quick indexing that rename_dst needs doesn't apply to rename_src, but it's not as obvious whether the broken-trees-with-duplicate-entries rationale applies or not. I spent a while digging into it. (And it's possible that I didn't correctly check the callers or that in the seven months since I wrote this message someone added another caller of this code that does pass multiple diff_filepair-s for the "same" source file.) Anyway, this paragraph exists because I had to go down a goose chase to answer this natural question, and I wanted to provide an answer to anyone else asking the same question. Also, in the off chance that anyone did want to add callers that passed multiple copies of any source file, I wanted to point out that this modified algorithm would result in a slight behavioral difference, but that otherwise the modified algorithm gives identical results. (And if some future reader stumbled on the paragraph because they had made such a change, I wanted to provide a quick suggestion of how to get what they wanted without adversely affecting performance.)

I hope that helps. I'm sorry if my worrying about these cases and discussing them made the patch harder to read or review, but I feel like diffcore-rename is one of those low-level components you want to be careful with. diffcore-rename is one of those parts of the code where a bug in it might not result in a directly observable breakage to users but in some secondary or tertiary side-effect showing weird results (e.g. in a merge you won't necessarily see that A was renamed to B, instead you get a three-way content merge of original A, other A, and new B -- or don't get a three-way content merge you might expect).

Thanks for taking a look!

:-). In fairness, I did read it with _emphasis_, so it at least gave me a laugh!

No, no need to be sorry: this is exactly the sort of discussion I had hoped to provoke by sanity checking my understanding of what you had written. I think that the paragraph does belong in the patch message, and interested readers can find our discussion here.

Thanks,
Taylor

register?

I originally started writing "I do not recall when the sortedness stopped mattering", until I realized you wrote "anywhere" not "anymore".

I do not think of any other reason than we wanted to notice and deal with the duplicated input. We do not look up the list of rename sources by pathname. So if we were sorting it, it is to prevent such a bug from breaking the rename machinery. What you call "technically the behaviour is different" is "removing the safety".

I do not offhand know which caller might give us such an input in the current code, so it may be entirely a safe thing to do. Besides, we use a hashmap of rename sources when computing exact renames, so even if we need to notice and/or avoid duplicates, we do not have to have register_rename_src() build the table with an insertion sort as a way to do so.

From: Elijah Newren <[email protected]>

register_rename_src() simply references the passed pair inside rename_src. In contrast, add_rename_dst() did something entirely different for rename_dst. Instead of copying the passed pair, it made a copy of the second diff_filespec from the passed pair, referenced it, and then set the diff_rename_dst.pair field to NULL. Later, when a pairing is found, record_rename_pair() allocated a full diff_filepair via diff_queue() and pointed its src and dst fields at the appropriate diff_filespecs. This contrast between register_rename_src() for the rename_src data structure and add_rename_dst() for the rename_dst data structure is oddly inconsistent and requires more memory and work than necessary. Let's just reference the original diff_filepair in rename_dst as-is, just as we do with rename_src. Add a new rename_dst.is_rename field, since the rename_dst.p field is never NULL unlike the old rename_dst.pair field.

Taking advantage of this change and the fact that same-named paths will be adjacent, we can get rid of the sorting of the array and most of the lookups on it, allowing us to instead just append as we go. However, there is one remaining reason to still keep locate_rename_dst():
handling broken pairs (i.e. when break detection is on). Those are somewhat rare, but we can set up a simple strintmap to get the map between the source and the index. Doing that allows us to still have a fast lookup without sorting the rename_dst array. Since the sorting had been done in a weakly quadratic manner, when many renames are involved this time could add up.

There is still a strcmp() in add_rename_dst() that I feel is unnecessary, because I disagree with a regression test in t4058. However, I'm trying to increase performance without changing behavior, so I'll leave that fight for a different day (a TODO has been left in the code about it).

This patch is being submitted in a different order than its original development, but in a large rebase of many commits with lots of renames and with several optimizations to inexact rename detection, both setup time and write back to output queue time from diffcore_rename() were sizeable chunks of overall runtime. This patch accelerated the setup time by about 65%, and final write back to the output queue time by about 50%, resulting in an overall drop of 3.5% on the execution time of rebasing a few dozen patches.

---
 diffcore-rename.c | 161 ++++++++++++++++++++++------------------------
 1 file changed, 77 insertions(+), 84 deletions(-)

index 816d2fbac44..fb3c2817c6b 100644
@@ -5,67 +5,64 @@
 #include "cache.h"
 #include "diff.h"
 #include "diffcore.h"
-#include "object-store.h"
 #include "hashmap.h"
+#include "object-store.h"
 #include "progress.h"
 #include "promisor-remote.h"
+#include "strmap.h"
 
 /* Table of rename/copy destinations */
 
 static struct diff_rename_dst {
-	struct diff_filespec *two;
-	struct diff_filepair *pair;
+	struct diff_filepair *p;
+	struct diff_filespec *filespec_to_free;
+	int is_rename; /* false -> just a create; true -> rename or copy */
 } *rename_dst;
 static int rename_dst_nr, rename_dst_alloc;
+/* Mapping from break source pathname to break destination index */
+static struct strintmap *break_idx = NULL;
 
-static int find_rename_dst(struct diff_filespec *two)
-{
-	int first, last;
-
-	first = 0;
-	last = rename_dst_nr;
-	while (last > first) {
-		int next = first + ((last - first) >> 1);
-		struct diff_rename_dst *dst = &(rename_dst[next]);
-		int cmp = strcmp(two->path, dst->two->path);
-		if (!cmp)
-			return next;
-		if (cmp < 0) {
-			last = next;
-			continue;
-		}
-		first = next+1;
-	}
-	return -first - 1;
-}
-
-static struct diff_rename_dst *locate_rename_dst(struct diff_filespec *two)
+static struct diff_rename_dst *locate_rename_dst(struct diff_filepair *p)
 {
-	int ofs = find_rename_dst(two);
-	return ofs < 0 ? NULL : &rename_dst[ofs];
+	/* Lookup by p->ONE->path */
+	int idx = break_idx ? strintmap_get(break_idx, p->one->path) : -1;
+	return (idx == -1) ? NULL : &rename_dst[idx];
 }
 
 /*
  * Returns 0 on success, -1 if we found a duplicate.
  */
-static int add_rename_dst(struct diff_filespec *two)
+static int add_rename_dst(struct diff_filepair *p)
 {
-	int first = find_rename_dst(two);
-
-	if (first >= 0)
+	/*
+	 * See t4058; trees might have duplicate entries.  I think
+	 * trees with duplicate entries should be ignored and we
+	 * should just leave rename detection on; while the results
+	 * may be slightly harder to understand, that's merely a
+	 * result of the underlying tree making no sense.  But I
+	 * believe everything still works fine, the results do still
+	 * make sense, and the extra overhead of doing this checking
+	 * for a few historical weird trees from long ago seems like
+	 * the dog wagging the tail to me.
+	 *
+	 * However: I don't feel like fighting that battle right now.
+	 * For now, to keep the regression test passing, we have to
+	 * detect it.  Since the diff machinery passes these to us in
+	 * adjacent pairs, we just need to check to see if our name
+	 * matches the previous one.
+	 *
+	 * TODO: Dispense with this test, rip out the test in t4058, and
+	 * lobby folks for the change.
+	 */
+	if (rename_dst_nr > 0 &&
+	    !strcmp(rename_dst[rename_dst_nr-1].p->two->path, p->two->path))
 		return -1;
-	first = -first - 1;
 
-	/* insert to make it at "first" */
 	ALLOC_GROW(rename_dst, rename_dst_nr + 1, rename_dst_alloc);
+	rename_dst[rename_dst_nr].p = p;
+	rename_dst[rename_dst_nr].filespec_to_free = NULL;
+	rename_dst[rename_dst_nr].is_rename = 0;
 	rename_dst_nr++;
-	if (first < rename_dst_nr)
-		MOVE_ARRAY(rename_dst + first + 1, rename_dst + first,
-			   rename_dst_nr - first - 1);
-	rename_dst[first].two = alloc_filespec(two->path);
-	fill_filespec(rename_dst[first].two, &two->oid, two->oid_valid,
-		      two->mode);
-	rename_dst[first].pair = NULL;
 	return 0;
 }
 

@@ -78,6 +75,14 @@ static int rename_src_nr, rename_src_alloc;
 
 static void register_rename_src(struct diff_filepair *p)
 {
+	if (p->broken_pair) {
+		if (!break_idx) {
+			break_idx = xmalloc(sizeof(*break_idx));
+			strintmap_init(break_idx, -1);
+		}
+		strintmap_set(break_idx, p->one->path, rename_dst_nr);
+	}
+
 	ALLOC_GROW(rename_src, rename_src_nr + 1, rename_src_alloc);
 	rename_src[rename_src_nr].p = p;
 	rename_src[rename_src_nr].score = p->score;

@@ -117,14 +122,14 @@ static void prefetch(void *prefetch_options)
 	struct oid_array to_fetch = OID_ARRAY_INIT;
 
 	for (i = 0; i < rename_dst_nr; i++) {
-		if (rename_dst[i].pair)
+		if (rename_dst[i].p->renamed_pair)
 			/*
 			 * The loop in diffcore_rename() will not need these
 			 * blobs, so skip prefetching.
 			 */
 			continue; /* already found exact match */
 		diff_add_if_missing(options->repo, &to_fetch,
-				    rename_dst[i].two);
+				    rename_dst[i].p->two);
 	}
 	for (i = 0; i < rename_src_nr; i++) {
 		if (options->skip_unmodified &&

@@ -234,26 +239,24 @@ static int estimate_similarity(struct repository *r,
 
 static void record_rename_pair(int dst_index, int src_index, int score)
 {
-	struct diff_filespec *src, *dst;
-	struct diff_filepair *dp;
+	struct diff_filepair *src = rename_src[src_index].p;
+	struct diff_filepair *dst = rename_dst[dst_index].p;
 
-	if (rename_dst[dst_index].pair)
+	if (dst->renamed_pair)
 		die("internal error: dst already matched.");
 
-	src = rename_src[src_index].p->one;
-	src->rename_used++;
-	src->count++;
+	src->one->rename_used++;
+	src->one->count++;
 
-	dst = rename_dst[dst_index].two;
-	dst->count++;
+	rename_dst[dst_index].filespec_to_free = dst->one;
+	rename_dst[dst_index].is_rename = 1;
 
-	dp = diff_queue(NULL, src, dst);
-	dp->renamed_pair = 1;
-	if (!strcmp(src->path, dst->path))
-		dp->score = rename_src[src_index].score;
+	dst->one = src->one;
+	dst->renamed_pair = 1;
+	if (!strcmp(dst->one->path, dst->two->path))
+		dst->score = rename_src[src_index].score;
 	else
-		dp->score = score;
-	rename_dst[dst_index].pair = dp;
+		dst->score = score;
 }
 
 /*

@@ -299,7 +302,7 @@ static int find_identical_files(struct hashmap *srcs,
 				struct diff_options *options)
 {
 	int renames = 0;
-	struct diff_filespec *target = rename_dst[dst_index].two;
+	struct diff_filespec *target = rename_dst[dst_index].p->two;
 	struct file_similarity *p, *best = NULL;
 	int i = 100, best_score = -1;
 	unsigned int hash = hash_filespec(options->repo, target);

@@ -460,7 +463,7 @@ static int find_renames(struct diff_score *mx, int dst_cnt, int minimum_score, i
 		    (mx[i].score < minimum_score))
 			break; /* there is no more usable pair. */
 		dst = &rename_dst[mx[i].dst];
-		if (dst->pair)
+		if (dst->is_rename)
 			continue; /* already done, either exact or fuzzy. */
 		if (!copies && rename_src[mx[i].src].p->one->rename_used)
 			continue;

@@ -495,7 +498,7 @@ void diffcore_rename(struct diff_options *options)
 			else if (!options->flags.rename_empty &&
 				 is_empty_blob_oid(&p->two->oid))
 				continue;
-			else if (add_rename_dst(p->two) < 0) {
+			else if (add_rename_dst(p) < 0) {
 				warning("skipping rename detection, detected"
 					" duplicate destination '%s'",
 					p->two->path);

@@ -568,10 +571,10 @@ void diffcore_rename(struct diff_options *options)
 
 	mx = xcalloc(st_mult(NUM_CANDIDATE_PER_DST, num_targets), sizeof(*mx));
 	for (dst_cnt = i = 0; i < rename_dst_nr; i++) {
-		struct diff_filespec *two = rename_dst[i].two;
+		struct diff_filespec *two = rename_dst[i].p->two;
 		struct diff_score *m;
 
-		if (rename_dst[i].pair)
+		if (rename_dst[i].is_rename)
 			continue; /* dealt with exact match already. */
 
 		m = &mx[dst_cnt * NUM_CANDIDATE_PER_DST];

@@ -628,22 +631,8 @@ void diffcore_rename(struct diff_options *options)
 			diff_q(&outq, p);
 		}
 		else if (!DIFF_FILE_VALID(p->one) && DIFF_FILE_VALID(p->two)) {
-			/*
-			 * Addition
-			 *
-			 * We would output this add record if it has
-			 * not been turned into a rename/copy already.
-			 */
-			struct diff_rename_dst *dst = locate_rename_dst(p->two);
-			if (dst && dst->pair) {
-				diff_q(&outq, dst->pair);
-				pair_to_free = p;
-			}
-			else
-				/* no matching rename/copy source, so
-				 * record this as an addition.
-				 */
-				diff_q(&outq, p);
+			/* Addition */
+			diff_q(&outq, p);
 		}
 		else if (DIFF_FILE_VALID(p->one) && !DIFF_FILE_VALID(p->two)) {
 			/*

@@ -664,8 +653,9 @@ void diffcore_rename(struct diff_options *options)
 			 */
 			if (DIFF_PAIR_BROKEN(p)) {
 				/* broken delete */
-				struct diff_rename_dst *dst = locate_rename_dst(p->one);
-				if (dst && dst->pair)
+				struct diff_rename_dst *dst = locate_rename_dst(p);
+				assert(dst);
+				if (dst->is_rename)
 					/* counterpart is now rename/copy */
 					pair_to_free = p;
 			}

@@ -675,16 +665,14 @@ void diffcore_rename(struct diff_options *options)
 					pair_to_free = p;
 			}
 
-			if (pair_to_free)
-				;
-			else
+			if (!pair_to_free)
 				diff_q(&outq, p);
 		}
 		else if (!diff_unmodified_pair(p))
 			/* all the usual ones need to be kept */
 			diff_q(&outq, p);
 		else
-			/* no need to keep unmodified pairs */
+			/* no need to keep unmodified pairs; FIXME: remove earlier? */
 			pair_to_free = p;
 
 		if (pair_to_free)

@@ -697,11 +685,16 @@ void diffcore_rename(struct diff_options *options)
 	diff_debug_queue("done collapsing", q);
 
 	for (i = 0; i < rename_dst_nr; i++)
-		free_filespec(rename_dst[i].two);
+		if (rename_dst[i].filespec_to_free)
+			free_filespec(rename_dst[i].filespec_to_free);
 
 	FREE_AND_NULL(rename_dst);
 	rename_dst_nr = rename_dst_alloc = 0;
 	FREE_AND_NULL(rename_src);
 	rename_src_nr = rename_src_alloc = 0;
+	if (break_idx) {
+		strintmap_clear(break_idx);
+		FREE_AND_NULL(break_idx);
+	}
 	return;
 }

gitgitgadget

Um, well...I guess I couldn't help myself from alphabetizing the #include list (at least the portion after the initial "cache.h" or "git-compat-util.h" must come first). ;-)

I probably should have done it in a separate patch.

It's identical for runtime correctness, yes. But the primary point isn't what happens at runtime, but what happens when future code readers come along. If I only keep the "if (dst->is_rename)" line that comes after without the assert, then someone in the future will come along (maybe even myself) and think "the original author wasn't being careful here; they should change this to a check on (dst && dst->is_rename)" (because honestly, it is the kind of mistake I'm prone to make). However, if they were to do so, and some bug gets introduced so that locate_rename_dst() returns a NULL for a pair of interest, then they've masked a bug in the algorithm and made it fail in harder-to-detect-and-track-down ways. I wanted it to be clear that dst == NULL is unacceptable. I guess I could have marked it with BUG(), but between an assertion and a NULL-pointer indirection, I figured that code aborting under that condition was pretty well covered. :-)

assert() often gets turned into no-op, so it is more like leaving a comment in the code. A comment or BUG("text") can describe not just the fact that reaching this point with dst==NULL is a bug in the caller but also why it is a bug (e.g. how NULL-dst would screw up the computation downstream from this point), but an assert() cannot.

This patch series includes 3 small code cleanups, 1 small bug fix (or perhaps just a UI improvement -- it's a very minor issue either way), more testcases, and 3 performance improvements. The first 7 patches are relatively easy to review, while the second to last one...is a bit more involved (but provided the biggest performance boost).

Changes since v1:

* Inserted a patch at the beginning to rename num_create to num_destinations, as suggested by Junio. Also simplifies patch 2.
* Used st_mult() to simplify the limit check
* Dropped a patch Junio didn't care for
* Inserted two new patches, 5 & 6, to add new testcases motivated by Junio's comments about register_rename_src() and correctness in the face of duplicate tree entries.
* Modified patch 7 so that register_rename_src() behaves identically to current code, even when trees have duplicate entries
* Added a new patch 9 which rips out the unnecessary duplicate entry checks for BOTH rename_src and rename_dst, and documents why they are safe to remove. This is also supported by the new testcases in patches 5 and 6. I also ran t4058 (which patches 5 and 6 add to) under valgrind afterwards and verified that no errors from diffcore-rename were reported (valgrind did report problems from cache-tree and unpack-trees, as documented in patch 6, but that was true before this patch series and probably has been that way for years or decades).
* Miscellaneous small cleanups highlighted by Taylor and Junio

Elijah Newren (9):
diffcore-rename: rename num_create to num_destinations
diffcore-rename: avoid usage of global in too_many_rename_candidates()
diffcore-rename: simplify limit check
diffcore-rename: reduce jumpiness in progress counters
t4058: add more tests and documentation for duplicate tree entry
handling
t4058: explore duplicate tree entry handling in a bit more detail
diffcore-rename: simplify and accelerate register_rename_src()
diffcore-rename: accelerate rename_dst setup
diffcore-rename: remove unneccessary duplicate entry checks

 diffcore-rename.c          | 209 ++++++++++++++-----------------------
 t/t4058-diff-duplicates.sh | 114 +++++++++++++++++++-
 2 files changed, 192 insertions(+), 131 deletions(-)

base-commit: 3a0b884caba2752da0af626fb2de7d597c844e8b
Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-git-929%2Fnewren%2Fdiffcore-rename-improvements-v2
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-git-929/newren/diffcore-rename-improvements-v2
Pull-Request: https://github.com/git/git/pull/929

Range-diff vs v1:

3: 30381addc5c ! 1: 428d8204894 diffcore-rename: rename num_create to num_targets
@@ Metadata
Author: Elijah Newren <[email protected]>

## Commit message ##
- diffcore-rename: rename num_create to num_targets
+ diffcore-rename: rename num_create to num_destinations

- Files added since the base commit serve as targets for rename detection.
- While it is true that added files can be thought of as being "created"
- when they are added IF they have no pairing file that they were renamed
- from, and it is true we start out not knowing what the pairings are, it
- seems a little odd to think in terms of "file creation" when we are
- looking for "file renames". Rename the variable to avoid this minor
- point of confusion.
+ Our main data structures are rename_src and rename_dst. For counters of
+ these data structures, num_sources and num_destinations seem natural;
+ definitely more so than using num_create for the latter.

## diffcore-rename.c ##
+@@ diffcore-rename.c: static void record_if_better(struct diff_score m[], struct diff_score *o)
+ * 1 if we need to disable inexact rename detection;
+ * 2 if we would be under the limit if we were given -C instead of -C -C.
+ */
+-static int too_many_rename_candidates(int num_create,
++static int too_many_rename_candidates(int num_destinations,
+ struct diff_options *options)
+ {
+ int rename_limit = options->rename_limit;
+@@ diffcore-rename.c: static int too_many_rename_candidates(int num_create,
+ * This basically does a test for the rename matrix not
+ * growing larger than a "rename_limit" square matrix, ie:
+ *
+- * num_create * num_src > rename_limit * rename_limit
++ * num_destinations * num_src > rename_limit * rename_limit
+ */
+ if (rename_limit <= 0)
+ rename_limit = 32767;
+- if ((num_create <= rename_limit || num_src <= rename_limit) &&
+- ((uint64_t)num_create * (uint64_t)num_src
++ if ((num_destinations <= rename_limit || num_src <= rename_limit) &&
++ ((uint64_t)num_destinations * (uint64_t)num_src
+ <= (uint64_t)rename_limit * (uint64_t)rename_limit))
+ return 0;
+
+ options->needed_rename_limit =
+- num_src > num_create ? num_src : num_create;
++ num_src > num_destinations ? num_src : num_destinations;
+
+ /* Are we running under -C -C? */
+ if (!options->flags.find_copies_harder)
+@@ diffcore-rename.c: static int too_many_rename_candidates(int num_create,
+ continue;
+ num_src++;
+ }
+- if ((num_create <= rename_limit || num_src <= rename_limit) &&
+- ((uint64_t)num_create * (uint64_t)num_src
++ if ((num_destinations <= rename_limit || num_src <= rename_limit) &&
++ ((uint64_t)num_destinations * (uint64_t)num_src
+ <= (uint64_t)rename_limit * (uint64_t)rename_limit))
+ return 2;
+ return 1;
@@ diffcore-rename.c: void diffcore_rename(struct diff_options *options)
struct diff_queue_struct outq;
struct diff_score *mx;
int i, j, rename_count, skip_unmodified = 0;
- int num_create, dst_cnt;
-+ int num_targets, dst_cnt;
++ int num_destinations, dst_cnt;
struct progress *progress = NULL;

if (!minimum_score)
@@ diffcore-rename.c: void diffcore_rename(struct diff_options *options)
* files still remain as options for rename/copies!)
*/
- num_create = (rename_dst_nr - rename_count);
-+ num_targets = (rename_dst_nr - rename_count);
++ num_destinations = (rename_dst_nr - rename_count);

/* All done? */
- if (!num_create)
-+ if (!num_targets)
++ if (!num_destinations)
goto cleanup;

-- switch (too_many_rename_candidates(num_create, rename_src_nr, options)) {
-+ switch (too_many_rename_candidates(num_targets, rename_src_nr, options)) {
+- switch (too_many_rename_candidates(num_create, options)) {
++ switch (too_many_rename_candidates(num_destinations, options)) {
case 1:
goto cleanup;
case 2:
@@ diffcore-rename.c: void diffcore_rename(struct diff_options *options)
}

- mx = xcalloc(st_mult(NUM_CANDIDATE_PER_DST, num_create), sizeof(*mx));
-+ mx = xcalloc(st_mult(NUM_CANDIDATE_PER_DST, num_targets), sizeof(*mx));
++ mx = xcalloc(st_mult(NUM_CANDIDATE_PER_DST, num_destinations),
++ sizeof(*mx));
for (dst_cnt = i = 0; i < rename_dst_nr; i++) {
struct diff_filespec *two = rename_dst[i].two;
struct diff_score *m;
1: 95a3fcb8be0 ! 2: fc62f4c4f89 diffcore-rename: avoid usage of global in too_many_rename_candidates()

     @@ Metadata
       ## Commit message ##
          diffcore-rename: avoid usage of global in too_many_rename_candidates()
      
     -    too_many_rename_candidates() got the number of rename targets via an
     -    argument to the function, but the number of rename sources via a global
     -    variable.  That felt rather inconsistent.  Pass in the number of rename
     -    sources as an argument as well.
     +    too_many_rename_candidates() got the number of rename destinations via
     +    an argument to the function, but the number of rename sources via a
     +    global variable.  That felt rather inconsistent.  Pass in the number of
     +    rename sources as an argument as well.
      
          While we are at it... We had a local variable, num_src, that served two
     -    purposes.  Initially it was set to this global value, but later was used
     +    purposes.  Initially it was set to the global value, but later was used
          for counting a subset of the number of sources.  Since we now have a
          function argument for the former usage, introduce a clearer variable
          name for the latter usage.
     @@ diffcore-rename.c: static void record_if_better(struct diff_score m[], struct di
        * 1 if we need to disable inexact rename detection;
        * 2 if we would be under the limit if we were given -C instead of -C -C.
        */
     --static int too_many_rename_candidates(int num_create,
     -+static int too_many_rename_candidates(int num_targets, int num_sources,
     +-static int too_many_rename_candidates(int num_destinations,
     ++static int too_many_rename_candidates(int num_destinations, int num_sources,
       				      struct diff_options *options)
       {
       	int rename_limit = options->rename_limit;
     @@ diffcore-rename.c: static void record_if_better(struct diff_score m[], struct di
       
       	options->needed_rename_limit = 0;
       
     -@@ diffcore-rename.c: static int too_many_rename_candidates(int num_create,
     +@@ diffcore-rename.c: static int too_many_rename_candidates(int num_destinations,
       	 * This basically does a test for the rename matrix not
       	 * growing larger than a "rename_limit" square matrix, ie:
       	 *
     --	 *    num_create * num_src > rename_limit * rename_limit
     -+	 *    num_targets * num_sources > rename_limit * rename_limit
     +-	 *    num_destinations * num_src > rename_limit * rename_limit
     ++	 *    num_destinations * num_sources > rename_limit * rename_limit
       	 */
       	if (rename_limit <= 0)
       		rename_limit = 32767;
     --	if ((num_create <= rename_limit || num_src <= rename_limit) &&
     --	    ((uint64_t)num_create * (uint64_t)num_src
     -+	if ((num_targets <= rename_limit || num_sources <= rename_limit) &&
     -+	    ((uint64_t)num_targets * (uint64_t)num_sources
     +-	if ((num_destinations <= rename_limit || num_src <= rename_limit) &&
     +-	    ((uint64_t)num_destinations * (uint64_t)num_src
     ++	if ((num_destinations <= rename_limit || num_sources <= rename_limit) &&
     ++	    ((uint64_t)num_destinations * (uint64_t)num_sources
       	     <= (uint64_t)rename_limit * (uint64_t)rename_limit))
       		return 0;
       
       	options->needed_rename_limit =
     --		num_src > num_create ? num_src : num_create;
     -+		num_sources > num_targets ? num_sources : num_targets;
     +-		num_src > num_destinations ? num_src : num_destinations;
     ++		num_sources > num_destinations ? num_sources : num_destinations;
       
       	/* Are we running under -C -C? */
       	if (!options->flags.find_copies_harder)
     @@ diffcore-rename.c: static int too_many_rename_candidates(int num_create,
      -		num_src++;
      +		limited_sources++;
       	}
     --	if ((num_create <= rename_limit || num_src <= rename_limit) &&
     --	    ((uint64_t)num_create * (uint64_t)num_src
     -+	if ((num_targets <= rename_limit || limited_sources <= rename_limit) &&
     -+	    ((uint64_t)num_targets * (uint64_t)limited_sources
     +-	if ((num_destinations <= rename_limit || num_src <= rename_limit) &&
     +-	    ((uint64_t)num_destinations * (uint64_t)num_src
     ++	if ((num_destinations <= rename_limit || limited_sources <= rename_limit) &&
     ++	    ((uint64_t)num_destinations * (uint64_t)limited_sources
       	     <= (uint64_t)rename_limit * (uint64_t)rename_limit))
       		return 2;
       	return 1;
      @@ diffcore-rename.c: void diffcore_rename(struct diff_options *options)
     - 	if (!num_create)
     + 	if (!num_destinations)
       		goto cleanup;
       
     --	switch (too_many_rename_candidates(num_create, options)) {
     -+	switch (too_many_rename_candidates(num_create, rename_src_nr, options)) {
     +-	switch (too_many_rename_candidates(num_destinations, options)) {
     ++	switch (too_many_rename_candidates(num_destinations, rename_src_nr,
     ++					   options)) {
       	case 1:
       		goto cleanup;
       	case 2:

2: f96bb36930a ! 3: 7214fa73fdd diffcore-rename: remove unnecessary if-clause
@@ Metadata
Author: Elijah Newren <[email protected]>

## Commit message ##
- diffcore-rename: remove unnecessary if-clause
+ diffcore-rename: simplify limit check

diffcore-rename had two different checks of the form

if ((a < limit || b < limit) &&
a * b <= limit * limit)

- Since these are all non-negative integers, this can be simplified to
+ This can be simplified to

- if (a * b <= limit * limit)
+ if (st_mult(a, b) <= st_mult(limit, limit))

- The only advantage of the former would be in avoiding a couple
- multiplications in the rare case that both a and b are BOTH very large.
- I see no reason for such an optimization given that this code is not in
- any kind of loop. Prefer code simplicity here and change to the latter
- form.
+ which makes it clearer how we are checking for overflow, and makes it
+ much easier to parse given the drop from 8 to 4 variable appearances.

## diffcore-rename.c ##
-@@ diffcore-rename.c: static int too_many_rename_candidates(int num_targets, int num_sources, +@@ diffcore-rename.c: static int too_many_rename_candidates(int num_destinations, int num_sources,
+ * growing larger than a "rename_limit" square matrix, ie:
+ *
+ * num_destinations * num_sources > rename_limit * rename_limit
++ *
++ * We use st_mult() to check overflow conditions; in the
++ * exceptional circumstance that size_t isn't large enough to hold
++ * the multiplication, the system won't be able to allocate enough
++ * memory for the matrix anyway.
*/
if (rename_limit <= 0)
rename_limit = 32767;
-- if ((num_targets <= rename_limit || num_sources <= rename_limit) &&
-- ((uint64_t)num_targets * (uint64_t)num_sources
+- if ((num_destinations <= rename_limit || num_sources <= rename_limit) &&
+- ((uint64_t)num_destinations * (uint64_t)num_sources
- <= (uint64_t)rename_limit * (uint64_t)rename_limit))
-+ if ((uint64_t)num_targets * (uint64_t)num_sources
-+ <= (uint64_t)rename_limit * (uint64_t)rename_limit)
++ if (st_mult(num_destinations, num_sources)
++ <= st_mult(rename_limit, rename_limit))
return 0;

options->needed_rename_limit =
-@@ diffcore-rename.c: static int too_many_rename_candidates(int num_targets, int num_sources, +@@ diffcore-rename.c: static int too_many_rename_candidates(int num_destinations, int num_sources,
continue;
limited_sources++;
}
-- if ((num_targets <= rename_limit || limited_sources <= rename_limit) &&
-- ((uint64_t)num_targets * (uint64_t)limited_sources
+- if ((num_destinations <= rename_limit || limited_sources <= rename_limit) &&
+- ((uint64_t)num_destinations * (uint64_t)limited_sources
- <= (uint64_t)rename_limit * (uint64_t)rename_limit))
-+ if ((uint64_t)num_targets * (uint64_t)limited_sources
-+ <= (uint64_t)rename_limit * (uint64_t)rename_limit)
++ if (st_mult(num_destinations, limited_sources)
++ <= st_mult(rename_limit, rename_limit))
return 2;
return 1;
}

  4:  77ca4656ed4 <  -:  ----------- diffcore-rename: change a few comments to use 'add' instead of 'create'
  5:  e8257516c1f !  4:  bf3581b1ac1 diffcore-rename: reduce jumpiness in progress counters
     @@ diffcore-rename.c: void diffcore_rename(struct diff_options *options)
       		progress = start_delayed_progress(
       				_("Performing inexact rename detection"),
      -				(uint64_t)rename_dst_nr * (uint64_t)rename_src_nr);
     -+				(uint64_t)num_targets * (uint64_t)rename_src_nr);
     ++				(uint64_t)num_destinations * (uint64_t)rename_src_nr);
       	}
       
     - 	mx = xcalloc(st_mult(NUM_CANDIDATE_PER_DST, num_targets), sizeof(*mx));
     + 	mx = xcalloc(st_mult(NUM_CANDIDATE_PER_DST, num_destinations),
      @@ diffcore-rename.c: void diffcore_rename(struct diff_options *options)
       			diff_free_filespec_blob(two);
       		}
  -:  ----------- >  5:  9a4a3159acf t4058: add more tests and documentation for duplicate tree entry handling
  -:  ----------- >  6:  8db27892c59 t4058: explore duplicate tree entry handling in a bit more detail
  6:  306a48820dd !  7:  a58639b2927 diffcore-rename: simplify and accelerate register_rename_src()

@@ Metadata
## Commit message ##
diffcore-rename: simplify and accelerate register_rename_src()

- reigster_rename_src() took pains to create an array in rename_src which
- was sorted by pathname of the contained diff_filepair. However, the
- fact that this array was sorted was not needed anywhere, and thus
- represented wasted time. Simply append to the end of the array, which
- in a usecase of note saved 45% of diffcore_rename() setup time for me.
-
- Technically, there is one difference in the end result for the code. IF
- the caller of diffcore-rename provides a set of pairs with a duplicate
- filename in the sources (an erroneous input condition), the old code
- would simply ignore the duplicate (without warning or error). The new
- code will simply swallow it and thus allow multiple pairings for the
- "same" source file. Checking for this error condition is expensive
- (basically undoing the optimization) and the only existing callers
- already avoid passing such an invalid input. If someone really wants to
- add some error checking here on this input condition, I believe they
- should add a separate function which can be called before
- diffcore_rename(), and then other callers that don't need additional
- checks because of their design (such as merge-ort) can avoid the extra
- overhead.
+ register_rename_src() took pains to create an array in rename_src which
+ was sorted by pathname of the contained diff_filepair. The sorting was
+ entirely unnecessary since callers pass filepairs to us in sorted
+ order. We can simply append to the end of the rename_src array,
+ speeding up diffcore_rename() setup time.

Also, note that I dropped the return type on the function since it was unconditionally discarded anyway.
@@ diffcore-rename.c: static struct diff_rename_src {
- }
- first = next+1;
- }
--
++ /*
++ * If we have multiple entries at the same path in the source tree
++ * (an invalid tree, to be sure), avoid using more more than one
++ * such entry in rename detection. Once upon a time, doing so
++ * caused segfaults; see commit 25d5ea410f ("[PATCH] Redo
++ * rename/copy detection logic.", 2005-05-24).
++ */
++ if (rename_src_nr > 0 &&
++ !strcmp(rename_src[rename_src_nr-1].p->one->path, p->one->path))
++ return;
+
- /* insert to make it at "first" */
ALLOC_GROW(rename_src, rename_src_nr + 1, rename_src_alloc);
+ rename_src[rename_src_nr].p = p;
7: 3e83b51628b ! 8: 85714e1583d Accelerate rename_dst setup
@@ Metadata
Author: Elijah Newren <[email protected]>

## Commit message ##
- Accelerate rename_dst setup
+ diffcore-rename: accelerate rename_dst setup

register_rename_src() simply references the passed pair inside rename_src. In contrast, add_rename_dst() did something entirely
@@ Commit message
been done in a weakly quadratic manner, when many renames are involved this time could add up.

- There is still a strcmp() in add_rename_dst() that I feel is
- unnecessary, because I disagree with a regression test in t4058.
- However, I'm trying to increase performance without changing behavior,
- so I'll leave that fight for a different day (a TODO has been left in
- the code about it).
+ There is still a strcmp() in add_rename_dst() that I have left in place
+ to make it easier to verify that the algorithm has the same results.
+ This strcmp() is there to check for duplicate destination entries (which
+ was the easiest way at the time to avoid segfaults in the
+ diffcore-rename code when trees had multiple entries at a given path).
+ The underlying double free()s are no longer an issue with the new
+ algorithm, but that can be addressed in a subsequent commit.

This patch is being submitted in a different order than its original development, but in a large rebase of many commits with lots of renames
@@ Commit message

## diffcore-rename.c ##
@@
- #include "cache.h"
- #include "diff.h"
- #include "diffcore.h"
--#include "object-store.h"
#include "hashmap.h"
-+#include "object-store.h"
#include "progress.h"
#include "promisor-remote.h"
+#include "strmap.h"
@@ diffcore-rename.c
-
- if (first >= 0)
+ /*
-+ * See t4058; trees might have duplicate entries. I think
-+ * trees with duplicate entries should be ignored and we
-+ * should just leave rename detection on; while the results
-+ * may be slightly harder to understand, that's merely a
-+ * result of the underlying tree making no sense. But I
-+ * believe everything still works fine, the results do still
-+ * make sense, and the extra overhead of doing this checking
-+ * for a few historical weird trees from long ago seems like
-+ * the dog wagging the tail to me.
-+ *
-+ * However: I don't feel like fighting that battle right now.
-+ * For now, to keep the regression test passing, we have to
-+ * detect it. Since the diff machinery passes these to us in
-+ * adjacent pairs, we just need to check to see if our name
-+ * matches the previous one.
-+ *
-+ * TODO: Dispense with this test, rip out the test in t4058, and
-+ * lobby folks for the change.
++ * If we have multiple entries at the same path in the destination
++ * tree (an invalid tree, to be sure), turn off rename detection
++ * entirely. Once upon a time, diffcore-rename had double free()
++ * issues due to such duplicate paths, resulting in segfaults. See
++ * commit 4d6be03b95 ("diffcore-rename: avoid processing duplicate
++ * destinations", 2015-02-26).
+ */
+ if (rename_dst_nr > 0 &&
+ !strcmp(rename_dst[rename_dst_nr-1].p->two->path, p->two->path))
@@ diffcore-rename.c
return 0;
}

-@@ diffcore-rename.c: static int rename_src_nr, rename_src_alloc; +@@ diffcore-rename.c: static void register_rename_src(struct diff_filepair *p)
+ !strcmp(rename_src[rename_src_nr-1].p->one->path, p->one->path))
+ return;

- static void register_rename_src(struct diff_filepair *p)
- {
+ if (p->broken_pair) {
+ if (!break_idx) {
+ break_idx = xmalloc(sizeof(*break_idx));
@@ diffcore-rename.c: void diffcore_rename(struct diff_options *options)
" duplicate destination '%s'", p->two->path);
@@ diffcore-rename.c: void diffcore_rename(struct diff_options *options)
-
- mx = xcalloc(st_mult(NUM_CANDIDATE_PER_DST, num_targets), sizeof(*mx));
+ mx = xcalloc(st_mult(NUM_CANDIDATE_PER_DST, num_destinations),
+ sizeof(*mx));
for (dst_cnt = i = 0; i < rename_dst_nr; i++) {
- struct diff_filespec *two = rename_dst[i].two;
+ struct diff_filespec *two = rename_dst[i].p->two;
@@ diffcore-rename.c: void diffcore_rename(struct diff_options *options)
}
else if (!DIFF_FILE_VALID(p->one) && DIFF_FILE_VALID(p->two)) {
- /*
-- * Addition
+- * Creation
- *
-- * We would output this add record if it has
+- * We would output this create record if it has
- * not been turned into a rename/copy already.
- */
- struct diff_rename_dst *dst = locate_rename_dst(p->two);
@@ diffcore-rename.c: void diffcore_rename(struct diff_options *options)
- }
- else
- /* no matching rename/copy source, so
-- * record this as an addition.
+- * record this as a creation.
- */
- diff_q(&outq, p);
-+ /* Addition */
++ /* Creation */
+ diff_q(&outq, p);
}
else if (DIFF_FILE_VALID(p->one) && !DIFF_FILE_VALID(p->two)) {
@@ diffcore-rename.c: void diffcore_rename(struct diff_options *options)
- struct diff_rename_dst *dst = locate_rename_dst(p->one);
- if (dst && dst->pair)
+ struct diff_rename_dst *dst = locate_rename_dst(p);
-+ assert(dst);
++ if (!dst)
++ BUG("tracking failed somehow; failed to find associated dst for broken pair");
+ if (dst->is_rename)
/* counterpart is now rename/copy */ pair_to_free = p;
-: ----------- > 9: af256b1c534 diffcore-rename: remove unneccessary duplicate entry checks

gitgitgadget

From: Elijah Newren <[email protected]>

Our main data structures are rename_src and rename_dst. For counters of these data structures, num_sources and num_destinations seem natural; definitely more so than using num_create for the latter.

---
 diffcore-rename.c | 25 +++++++++++++------------
 1 file changed, 13 insertions(+), 12 deletions(-)

index d367a6d2443..15a98f566e4 100644
@@ -434,7 +434,7 @@ static void record_if_better(struct diff_score m[], struct diff_score *o)
  * 1 if we need to disable inexact rename detection;
  * 2 if we would be under the limit if we were given -C instead of -C -C.
  */
-static int too_many_rename_candidates(int num_create,
+static int too_many_rename_candidates(int num_destinations,
 				      struct diff_options *options)
 {
 	int rename_limit = options->rename_limit;

@@ -447,17 +447,17 @@ static int too_many_rename_candidates(int num_create,
 	 * This basically does a test for the rename matrix not
 	 * growing larger than a "rename_limit" square matrix, ie:
 	 *
-	 *    num_create * num_src > rename_limit * rename_limit
+	 *    num_destinations * num_src > rename_limit * rename_limit
 	 */
 	if (rename_limit <= 0)
 		rename_limit = 32767;
-	if ((num_create <= rename_limit || num_src <= rename_limit) &&
-	    ((uint64_t)num_create * (uint64_t)num_src
+	if ((num_destinations <= rename_limit || num_src <= rename_limit) &&
+	    ((uint64_t)num_destinations * (uint64_t)num_src
 	     <= (uint64_t)rename_limit * (uint64_t)rename_limit))
 		return 0;
 
 	options->needed_rename_limit =
-		num_src > num_create ? num_src : num_create;
+		num_src > num_destinations ? num_src : num_destinations;
 
 	/* Are we running under -C -C? */
 	if (!options->flags.find_copies_harder)

@@ -469,8 +469,8 @@ static int too_many_rename_candidates(int num_create,
 			continue;
 		num_src++;
 	}
-	if ((num_create <= rename_limit || num_src <= rename_limit) &&
-	    ((uint64_t)num_create * (uint64_t)num_src
+	if ((num_destinations <= rename_limit || num_src <= rename_limit) &&
+	    ((uint64_t)num_destinations * (uint64_t)num_src
 	     <= (uint64_t)rename_limit * (uint64_t)rename_limit))
 		return 2;
 	return 1;

@@ -505,7 +505,7 @@ void diffcore_rename(struct diff_options *options)
 	struct diff_queue_struct outq;
 	struct diff_score *mx;
 	int i, j, rename_count, skip_unmodified = 0;
-	int num_create, dst_cnt;
+	int num_destinations, dst_cnt;
 	struct progress *progress = NULL;
 
 	if (!minimum_score)

@@ -570,13 +570,13 @@ void diffcore_rename(struct diff_options *options)
 	 * Calculate how many renames are left (but all the source
 	 * files still remain as options for rename/copies!)
 	 */
-	num_create = (rename_dst_nr - rename_count);
+	num_destinations = (rename_dst_nr - rename_count);
 
 	/* All done? */
-	if (!num_create)
+	if (!num_destinations)
 		goto cleanup;
 
-	switch (too_many_rename_candidates(num_create, options)) {
+	switch (too_many_rename_candidates(num_destinations, options)) {
 	case 1:
 		goto cleanup;
 	case 2:

@@ -593,7 +593,8 @@ void diffcore_rename(struct diff_options *options)
 				(uint64_t)rename_dst_nr * (uint64_t)rename_src_nr);
 	}
 
-	mx = xcalloc(st_mult(NUM_CANDIDATE_PER_DST, num_create), sizeof(*mx));
+	mx = xcalloc(st_mult(NUM_CANDIDATE_PER_DST, num_destinations),
+		     sizeof(*mx));
 	for (dst_cnt = i = 0; i < rename_dst_nr; i++) {
 		struct diff_filespec *two = rename_dst[i].two;
 		struct diff_score *m;

gitgitgadget

From: Elijah Newren <[email protected]>

too_many_rename_candidates() got the number of rename destinations via an argument to the function, but the number of rename sources via a global variable. That felt rather inconsistent. Pass in the number of rename sources as an argument as well.

While we are at it... We had a local variable, num_src, that served two purposes. Initially it was set to the global value, but later was used for counting a subset of the number of sources. Since we now have a function argument for the former usage, introduce a clearer variable name for the latter usage.

This patch has no behavioral changes; it's just renaming and passing an argument instead of grabbing it from the global namespace. (You may find it easier to view the patch using git diff's --color-words option.)

---
 diffcore-rename.c | 24 ++++++++++++------------
 1 file changed, 12 insertions(+), 12 deletions(-)

index 15a98f566e4..1d6675c040d 100644
@@ -434,12 +434,11 @@ static void record_if_better(struct diff_score m[], struct diff_score *o)
  * 1 if we need to disable inexact rename detection;
  * 2 if we would be under the limit if we were given -C instead of -C -C.
  */
-static int too_many_rename_candidates(int num_destinations,
+static int too_many_rename_candidates(int num_destinations, int num_sources,
 				      struct diff_options *options)
 {
 	int rename_limit = options->rename_limit;
-	int num_src = rename_src_nr;
-	int i;
+	int i, limited_sources;
 
 	options->needed_rename_limit = 0;
 

@@ -447,30 +446,30 @@ static int too_many_rename_candidates(int num_destinations,
 	 * This basically does a test for the rename matrix not
 	 * growing larger than a "rename_limit" square matrix, ie:
 	 *
-	 *    num_destinations * num_src > rename_limit * rename_limit
+	 *    num_destinations * num_sources > rename_limit * rename_limit
 	 */
 	if (rename_limit <= 0)
 		rename_limit = 32767;
-	if ((num_destinations <= rename_limit || num_src <= rename_limit) &&
-	    ((uint64_t)num_destinations * (uint64_t)num_src
+	if ((num_destinations <= rename_limit || num_sources <= rename_limit) &&
+	    ((uint64_t)num_destinations * (uint64_t)num_sources
 	     <= (uint64_t)rename_limit * (uint64_t)rename_limit))
 		return 0;
 
 	options->needed_rename_limit =
-		num_src > num_destinations ? num_src : num_destinations;
+		num_sources > num_destinations ? num_sources : num_destinations;
 
 	/* Are we running under -C -C? */
 	if (!options->flags.find_copies_harder)
 		return 1;
 
 	/* Would we bust the limit if we were running under -C? */
-	for (num_src = i = 0; i < rename_src_nr; i++) {
+	for (limited_sources = i = 0; i < num_sources; i++) {
 		if (diff_unmodified_pair(rename_src[i].p))
 			continue;
-		num_src++;
+		limited_sources++;
 	}
-	if ((num_destinations <= rename_limit || num_src <= rename_limit) &&
-	    ((uint64_t)num_destinations * (uint64_t)num_src
+	if ((num_destinations <= rename_limit || limited_sources <= rename_limit) &&
+	    ((uint64_t)num_destinations * (uint64_t)limited_sources
 	     <= (uint64_t)rename_limit * (uint64_t)rename_limit))
 		return 2;
 	return 1;

@@ -576,7 +575,8 @@ void diffcore_rename(struct diff_options *options)
 	if (!num_destinations)
 		goto cleanup;
 
-	switch (too_many_rename_candidates(num_destinations, options)) {
+	switch (too_many_rename_candidates(num_destinations, rename_src_nr,
+					   options)) {
 	case 1:
 		goto cleanup;
 	case 2:

gitgitgadget

From: Elijah Newren <[email protected]>

diffcore-rename had two different checks of the form

if ((a < limit || b < limit) &&
a * b <= limit * limit)

This can be simplified to

    if (st_mult(a, b) <= st_mult(limit, limit))

which makes it clearer how we are checking for overflow, and makes it much easier to parse given the drop from 8 to 4 variable appearances.

---
 diffcore-rename.c | 15 +++++++++------
 1 file changed, 9 insertions(+), 6 deletions(-)

index 1d6675c040d..16553ab259f 100644
@@ -447,12 +447,16 @@ static int too_many_rename_candidates(int num_destinations, int num_sources,
 	 * growing larger than a "rename_limit" square matrix, ie:
 	 *
 	 *    num_destinations * num_sources > rename_limit * rename_limit
+	 *
+	 * We use st_mult() to check overflow conditions; in the
+	 * exceptional circumstance that size_t isn't large enough to hold
+	 * the multiplication, the system won't be able to allocate enough
+	 * memory for the matrix anyway.
 	 */
 	if (rename_limit <= 0)
 		rename_limit = 32767;
-	if ((num_destinations <= rename_limit || num_sources <= rename_limit) &&
-	    ((uint64_t)num_destinations * (uint64_t)num_sources
-	     <= (uint64_t)rename_limit * (uint64_t)rename_limit))
+	if (st_mult(num_destinations, num_sources)
+	    <= st_mult(rename_limit, rename_limit))
 		return 0;
 
 	options->needed_rename_limit =

@@ -468,9 +472,8 @@ static int too_many_rename_candidates(int num_destinations, int num_sources,
 			continue;
 		limited_sources++;
 	}
-	if ((num_destinations <= rename_limit || limited_sources <= rename_limit) &&
-	    ((uint64_t)num_destinations * (uint64_t)limited_sources
-	     <= (uint64_t)rename_limit * (uint64_t)rename_limit))
+	if (st_mult(num_destinations, limited_sources)
+	    <= st_mult(rename_limit, rename_limit))
 		return 2;
 	return 1;
 }

gitgitgadget

From: Elijah Newren <[email protected]>

Inexact rename detection works by comparing all sources to all destinations, computing similarities, and then finding the best matches among those that are sufficiently similar.

However, it is preceded by exact rename detection that works by checking if there are files with identical hashes. If exact renames are found, we can exclude some files from inexact rename detection.

The inexact rename detection loops over the full set of files, but immediately skips those for which rename_dst[i].is_rename is true and thus doesn't compare any sources to that destination. As such, these paths shouldn't be included in the progress counter.

For the eagle eyed, this change hints at an actual optimization -- the first one I presented at Git Merge 2020. I'll be submitting that optimization later, once the basic merge-ort algorithm has merged.

---
 diffcore-rename.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

index 16553ab259f..55a188abcc3 100644
@@ -593,7 +593,7 @@ void diffcore_rename(struct diff_options *options)
 	if (options->show_rename_progress) {
 		progress = start_delayed_progress(
 				_("Performing inexact rename detection"),
-				(uint64_t)rename_dst_nr * (uint64_t)rename_src_nr);
+				(uint64_t)num_destinations * (uint64_t)rename_src_nr);
 	}
 
 	mx = xcalloc(st_mult(NUM_CANDIDATE_PER_DST, num_destinations),

@@ -633,7 +633,8 @@ void diffcore_rename(struct diff_options *options)
 			diff_free_filespec_blob(two);
 		}
 		dst_cnt++;
-		display_progress(progress, (uint64_t)(i+1)*(uint64_t)rename_src_nr);
+		display_progress(progress,
+				 (uint64_t)dst_cnt * (uint64_t)rename_src_nr);
 	}
 	stop_progress(&progress);
 

gitgitgadget

From: Elijah Newren <[email protected]>

Commit 4d6be03b95 ("diffcore-rename: avoid processing duplicate destinations", 2015-02-26) added t4058 to demonstrate that a workaround it added to avoid double frees (namely to just turn off rename detection when trees had duplicate entries) would indeed avoid segfaults. The tests, though, give the impression that the expected diffs are "correct" when in reality they are just "don't segfault, and do something semi-reasonable under the circumstances". Add some notes to make this clearer.

Also, commit 25d5ea410f ("[PATCH] Redo rename/copy detection logic.", 2005-05-24) added a similar workaround to avoid segfaults, but for rename_src rather than rename_dst. I do not see any tests in the testsuite to cover the collision detection of entries limited to the source side, so add a couple.

---
 t/t4058-diff-duplicates.sh | 47 ++++++++++++++++++++++++++++++++++++--
 1 file changed, 45 insertions(+), 2 deletions(-)

index c24ee175ef0..bd685089561 100755
@@ -1,5 +1,14 @@
 #!/bin/sh
 
+# NOTICE:
+#   This testsuite does a number of diffs and checks that the output match.
+#   However, it is a "garbage in, garbage out" situation; the trees have
+#   duplicate entries for individual paths, and it results in diffs that do
+#   not make much sense.  As such, it is not clear that the diffs are
+#   "correct".  The primary purpose of these tests was to verify that
+#   diff-tree does not segfault, but there is perhaps some value in ensuring
+#   that the diff output isn't wildly unreasonable.
+
 test_description='test tree diff when trees have duplicate entries'
 . ./test-lib.sh
 

@@ -57,7 +66,16 @@ test_expect_success 'create trees with duplicate entries' '
 	git tag two $outer_two
 '
 
-test_expect_success 'diff-tree between trees' '
+test_expect_success 'create tree without duplicate entries' '
+	blob_one=$(echo one | git hash-object -w --stdin) &&
+	outer_three=$(make_tree \
+		100644 renamed $blob_one
+	) &&
+	git tag three $outer_three
+'
+
+test_expect_success 'diff-tree between duplicate trees' '
+	# See NOTICE at top of file
 	{
 		printf ":000000 100644 $ZERO_OID $blob_two A\touter/inner\n" &&
 		printf ":000000 100644 $ZERO_OID $blob_two A\touter/inner\n" &&

@@ -71,9 +89,34 @@ test_expect_success 'diff-tree between trees' '
 '
 
 test_expect_success 'diff-tree with renames' '
-	# same expectation as above, since we disable rename detection
+	# See NOTICE at top of file.
 	git diff-tree -M -r --no-abbrev one two >actual &&
 	test_cmp expect actual
 '
 
+test_expect_success 'diff-tree FROM duplicate tree' '
+	# See NOTICE at top of file.
+	{
+		printf ":100644 000000 $blob_one $ZERO_OID D\touter/inner\n" &&
+		printf ":100644 000000 $blob_two $ZERO_OID D\touter/inner\n" &&
+		printf ":100644 000000 $blob_two $ZERO_OID D\touter/inner\n" &&
+		printf ":100644 000000 $blob_two $ZERO_OID D\touter/inner\n" &&
+		printf ":000000 100644 $ZERO_OID $blob_one A\trenamed\n"
+	} >expect &&
+	git diff-tree -r --no-abbrev one three >actual &&
+	test_cmp expect actual
+'
+
+test_expect_success 'diff-tree FROM duplicate tree, with renames' '
+	# See NOTICE at top of file.
+	{
+		printf ":100644 000000 $blob_two $ZERO_OID D\touter/inner\n" &&
+		printf ":100644 000000 $blob_two $ZERO_OID D\touter/inner\n" &&
+		printf ":100644 000000 $blob_two $ZERO_OID D\touter/inner\n" &&
+		printf ":100644 100644 $blob_one $blob_one R100\touter/inner\trenamed\n"
+	} >expect &&
+	git diff-tree -M -r --no-abbrev one three >actual &&
+	test_cmp expect actual
+'
+
 test_done

gitgitgadget

From: Elijah Newren <[email protected]>

While creating the last commit, I found a number of other cases where git would segfault when faced with trees that have duplicate entries. None of these segfaults are in the diffcore-rename code (they all occur in cache-tree and unpack-trees). Further, to my knowledge, no one has ever been adversely affected by these bugs, and given that it has been 15 years and folks have fixed a few other issues with historical duplicate entries (as noted in the last commit), I am not sure we will ever run into anyone having problems with these. So I am not sure these are worth fixing, but it doesn't hurt to at least document these failures in the same test file that is concerned with duplicate tree entries.

---
 t/t4058-diff-duplicates.sh | 67 ++++++++++++++++++++++++++++++++++++++
 1 file changed, 67 insertions(+)

index bd685089561..ad3f37d388d 100755
@@ -119,4 +119,71 @@ test_expect_success 'diff-tree FROM duplicate tree, with renames' '
 	test_cmp expect actual
 '
 
+test_expect_success 'create a few commits' '
+	git commit-tree -m "Duplicate Entries" two^{tree} >commit_id &&
+	git branch base $(cat commit_id) &&
+
+	git commit-tree -p $(cat commit_id) -m "Just one" three^{tree} >up &&
+	git branch update $(cat up) &&
+
+	git commit-tree -p $(cat up) -m "Back to weird" two^{tree} >final &&
+	git branch final $(cat final) &&
+
+	rm commit_id up final
+'
+
+test_expect_failure 'git read-tree does not segfault' '
+	test_when_finished rm .git/index.lock &&
+	test_might_fail git read-tree --reset base
+'
+
+test_expect_failure 'reset --hard does not segfault' '
+	test_when_finished rm .git/index.lock &&
+	git checkout base &&
+	test_might_fail git reset --hard
+'
+
+test_expect_failure 'git diff HEAD does not segfault' '
+	git checkout base &&
+	GIT_TEST_CHECK_CACHE_TREE=false &&
+	git reset --hard &&
+	test_might_fail git diff HEAD
+'
+
+test_expect_failure 'can switch to another branch when status is empty' '
+	git clean -ffdqx &&
+	git status --porcelain -uno >actual &&
+	test_must_be_empty actual &&
+	git checkout update
+'
+
+test_expect_success 'forcibly switch to another branch, verify status empty' '
+	git checkout -f update &&
+	git status --porcelain -uno >actual &&
+	test_must_be_empty actual
+'
+
+test_expect_success 'fast-forward from non-duplicate entries to duplicate' '
+	git merge final
+'
+
+test_expect_failure 'clean status, switch branches, status still clean' '
+	git status --porcelain -uno >actual &&
+	test_must_be_empty actual &&
+	git checkout base &&
+	git status --porcelain -uno >actual &&
+	test_must_be_empty actual
+'
+
+test_expect_success 'switch to base branch and force status to be clean' '
+	git checkout base &&
+	GIT_TEST_CHECK_CACHE_TREE=false git reset --hard &&
+	git status --porcelain -uno >actual &&
+	test_must_be_empty actual
+'
+
+test_expect_failure 'fast-forward from duplicate entries to non-duplicate' '
+	git merge update
+'
+
 test_done

gitgitgadget

From: Elijah Newren <[email protected]>

register_rename_src() took pains to create an array in rename_src which was sorted by pathname of the contained diff_filepair. The sorting was entirely unnecessary since callers pass filepairs to us in sorted order. We can simply append to the end of the rename_src array, speeding up diffcore_rename() setup time.

Also, note that I dropped the return type on the function since it was unconditionally discarded anyway.

This patch is being submitted in a different order than its original development, but in a large rebase of many commits with lots of renames and with several optimizations to inexact rename detection, diffcore_rename() setup time was a sizeable chunk of overall runtime. This patch dropped execution time of rebasing 35 commits with lots of renames by 2% overall.

---
 diffcore-rename.c | 39 +++++++++++++--------------------------
 1 file changed, 13 insertions(+), 26 deletions(-)

index 55a188abcc3..a215421a9cb 100644
@@ -76,36 +76,23 @@ static struct diff_rename_src {
 } *rename_src;
 static int rename_src_nr, rename_src_alloc;
 
-static struct diff_rename_src *register_rename_src(struct diff_filepair *p)
+static void register_rename_src(struct diff_filepair *p)
 {
-	int first, last;
-	struct diff_filespec *one = p->one;
-	unsigned short score = p->score;
-
-	first = 0;
-	last = rename_src_nr;
-	while (last > first) {
-		int next = first + ((last - first) >> 1);
-		struct diff_rename_src *src = &(rename_src[next]);
-		int cmp = strcmp(one->path, src->p->one->path);
-		if (!cmp)
-			return src;
-		if (cmp < 0) {
-			last = next;
-			continue;
-		}
-		first = next+1;
-	}
+	/*
+	 * If we have multiple entries at the same path in the source tree
+	 * (an invalid tree, to be sure), avoid using more more than one
+	 * such entry in rename detection.  Once upon a time, doing so
+	 * caused segfaults; see commit 25d5ea410f ("[PATCH] Redo
+	 * rename/copy detection logic.", 2005-05-24).
+	 */
+	if (rename_src_nr > 0 &&
+	    !strcmp(rename_src[rename_src_nr-1].p->one->path, p->one->path))
+		return;
 
-	/* insert to make it at "first" */
 	ALLOC_GROW(rename_src, rename_src_nr + 1, rename_src_alloc);
+	rename_src[rename_src_nr].p = p;
+	rename_src[rename_src_nr].score = p->score;
 	rename_src_nr++;
-	if (first < rename_src_nr)
-		MOVE_ARRAY(rename_src + first + 1, rename_src + first,
-			   rename_src_nr - first - 1);
-	rename_src[first].p = p;
-	rename_src[first].score = score;
-	return &(rename_src[first]);
 }
 
 static int basename_same(struct diff_filespec *src, struct diff_filespec *dst)

gitgitgadget

From: Elijah Newren <[email protected]>

register_rename_src() simply references the passed pair inside rename_src. In contrast, add_rename_dst() did something entirely different for rename_dst. Instead of copying the passed pair, it made a copy of the second diff_filespec from the passed pair, referenced it, and then set the diff_rename_dst.pair field to NULL. Later, when a pairing is found, record_rename_pair() allocated a full diff_filepair via diff_queue() and pointed its src and dst fields at the appropriate diff_filespecs. This contrast between register_rename_src() for the rename_src data structure and add_rename_dst() for the rename_dst data structure is oddly inconsistent and requires more memory and work than necessary. Let's just reference the original diff_filepair in rename_dst as-is, just as we do with rename_src. Add a new rename_dst.is_rename field, since the rename_dst.p field is never NULL unlike the old rename_dst.pair field.

Taking advantage of this change and the fact that same-named paths will be adjacent, we can get rid of the sorting of the array and most of the lookups on it, allowing us to instead just append as we go. However, there is one remaining reason to still keep locate_rename_dst():
handling broken pairs (i.e. when break detection is on). Those are somewhat rare, but we can set up a simple strintmap to get the map between the source and the index. Doing that allows us to still have a fast lookup without sorting the rename_dst array. Since the sorting had been done in a weakly quadratic manner, when many renames are involved this time could add up.

There is still a strcmp() in add_rename_dst() that I have left in place to make it easier to verify that the algorithm has the same results. This strcmp() is there to check for duplicate destination entries (which was the easiest way at the time to avoid segfaults in the diffcore-rename code when trees had multiple entries at a given path). The underlying double free()s are no longer an issue with the new algorithm, but that can be addressed in a subsequent commit.

This patch is being submitted in a different order than its original development, but in a large rebase of many commits with lots of renames and with several optimizations to inexact rename detection, both setup time and write back to output queue time from diffcore_rename() were sizeable chunks of overall runtime. This patch accelerated the setup time by about 65%, and final write back to the output queue time by about 50%, resulting in an overall drop of 3.5% on the execution time of rebasing a few dozen patches.

---
 diffcore-rename.c | 148 ++++++++++++++++++++--------------------------
 1 file changed, 65 insertions(+), 83 deletions(-)

index a215421a9cb..2a8fcd52928 100644
@@ -9,63 +9,48 @@
 #include "hashmap.h"
 #include "progress.h"
 #include "promisor-remote.h"
+#include "strmap.h"
 
 /* Table of rename/copy destinations */
 
 static struct diff_rename_dst {
-	struct diff_filespec *two;
-	struct diff_filepair *pair;
+	struct diff_filepair *p;
+	struct diff_filespec *filespec_to_free;
+	int is_rename; /* false -> just a create; true -> rename or copy */
 } *rename_dst;
 static int rename_dst_nr, rename_dst_alloc;
+/* Mapping from break source pathname to break destination index */
+static struct strintmap *break_idx = NULL;
 
-static int find_rename_dst(struct diff_filespec *two)
-{
-	int first, last;
-
-	first = 0;
-	last = rename_dst_nr;
-	while (last > first) {
-		int next = first + ((last - first) >> 1);
-		struct diff_rename_dst *dst = &(rename_dst[next]);
-		int cmp = strcmp(two->path, dst->two->path);
-		if (!cmp)
-			return next;
-		if (cmp < 0) {
-			last = next;
-			continue;
-		}
-		first = next+1;
-	}
-	return -first - 1;
-}
-
-static struct diff_rename_dst *locate_rename_dst(struct diff_filespec *two)
+static struct diff_rename_dst *locate_rename_dst(struct diff_filepair *p)
 {
-	int ofs = find_rename_dst(two);
-	return ofs < 0 ? NULL : &rename_dst[ofs];
+	/* Lookup by p->ONE->path */
+	int idx = break_idx ? strintmap_get(break_idx, p->one->path) : -1;
+	return (idx == -1) ? NULL : &rename_dst[idx];
 }
 
 /*
  * Returns 0 on success, -1 if we found a duplicate.
  */
-static int add_rename_dst(struct diff_filespec *two)
+static int add_rename_dst(struct diff_filepair *p)
 {
-	int first = find_rename_dst(two);
-
-	if (first >= 0)
+	/*
+	 * If we have multiple entries at the same path in the destination
+	 * tree (an invalid tree, to be sure), turn off rename detection
+	 * entirely.  Once upon a time, diffcore-rename had double free()
+	 * issues due to such duplicate paths, resulting in segfaults.  See
+	 * commit 4d6be03b95 ("diffcore-rename: avoid processing duplicate
+	 * destinations", 2015-02-26).
+	 */
+	if (rename_dst_nr > 0 &&
+	    !strcmp(rename_dst[rename_dst_nr-1].p->two->path, p->two->path))
 		return -1;
-	first = -first - 1;
 
-	/* insert to make it at "first" */
 	ALLOC_GROW(rename_dst, rename_dst_nr + 1, rename_dst_alloc);
+	rename_dst[rename_dst_nr].p = p;
+	rename_dst[rename_dst_nr].filespec_to_free = NULL;
+	rename_dst[rename_dst_nr].is_rename = 0;
 	rename_dst_nr++;
-	if (first < rename_dst_nr)
-		MOVE_ARRAY(rename_dst + first + 1, rename_dst + first,
-			   rename_dst_nr - first - 1);
-	rename_dst[first].two = alloc_filespec(two->path);
-	fill_filespec(rename_dst[first].two, &two->oid, two->oid_valid,
-		      two->mode);
-	rename_dst[first].pair = NULL;
 	return 0;
 }
 

@@ -89,6 +74,14 @@ static void register_rename_src(struct diff_filepair *p)
 	    !strcmp(rename_src[rename_src_nr-1].p->one->path, p->one->path))
 		return;
 
+	if (p->broken_pair) {
+		if (!break_idx) {
+			break_idx = xmalloc(sizeof(*break_idx));
+			strintmap_init(break_idx, -1);
+		}
+		strintmap_set(break_idx, p->one->path, rename_dst_nr);
+	}
+
 	ALLOC_GROW(rename_src, rename_src_nr + 1, rename_src_alloc);
 	rename_src[rename_src_nr].p = p;
 	rename_src[rename_src_nr].score = p->score;

@@ -128,14 +121,14 @@ static void prefetch(void *prefetch_options)
 	struct oid_array to_fetch = OID_ARRAY_INIT;
 
 	for (i = 0; i < rename_dst_nr; i++) {
-		if (rename_dst[i].pair)
+		if (rename_dst[i].p->renamed_pair)
 			/*
 			 * The loop in diffcore_rename() will not need these
 			 * blobs, so skip prefetching.
 			 */
 			continue; /* already found exact match */
 		diff_add_if_missing(options->repo, &to_fetch,
-				    rename_dst[i].two);
+				    rename_dst[i].p->two);
 	}
 	for (i = 0; i < rename_src_nr; i++) {
 		if (options->skip_unmodified &&

@@ -245,26 +238,24 @@ static int estimate_similarity(struct repository *r,
 
 static void record_rename_pair(int dst_index, int src_index, int score)
 {
-	struct diff_filespec *src, *dst;
-	struct diff_filepair *dp;
+	struct diff_filepair *src = rename_src[src_index].p;
+	struct diff_filepair *dst = rename_dst[dst_index].p;
 
-	if (rename_dst[dst_index].pair)
+	if (dst->renamed_pair)
 		die("internal error: dst already matched.");
 
-	src = rename_src[src_index].p->one;
-	src->rename_used++;
-	src->count++;
+	src->one->rename_used++;
+	src->one->count++;
 
-	dst = rename_dst[dst_index].two;
-	dst->count++;
+	rename_dst[dst_index].filespec_to_free = dst->one;
+	rename_dst[dst_index].is_rename = 1;
 
-	dp = diff_queue(NULL, src, dst);
-	dp->renamed_pair = 1;
-	if (!strcmp(src->path, dst->path))
-		dp->score = rename_src[src_index].score;
+	dst->one = src->one;
+	dst->renamed_pair = 1;
+	if (!strcmp(dst->one->path, dst->two->path))
+		dst->score = rename_src[src_index].score;
 	else
-		dp->score = score;
-	rename_dst[dst_index].pair = dp;
+		dst->score = score;
 }
 
 /*

@@ -310,7 +301,7 @@ static int find_identical_files(struct hashmap *srcs,
 				struct diff_options *options)
 {
 	int renames = 0;
-	struct diff_filespec *target = rename_dst[dst_index].two;
+	struct diff_filespec *target = rename_dst[dst_index].p->two;
 	struct file_similarity *p, *best = NULL;
 	int i = 100, best_score = -1;
 	unsigned int hash = hash_filespec(options->repo, target);

@@ -476,7 +467,7 @@ static int find_renames(struct diff_score *mx, int dst_cnt, int minimum_score, i
 		    (mx[i].score < minimum_score))
 			break; /* there is no more usable pair. */
 		dst = &rename_dst[mx[i].dst];
-		if (dst->pair)
+		if (dst->is_rename)
 			continue; /* already done, either exact or fuzzy. */
 		if (!copies && rename_src[mx[i].src].p->one->rename_used)
 			continue;

@@ -511,7 +502,7 @@ void diffcore_rename(struct diff_options *options)
 			else if (!options->flags.rename_empty &&
 				 is_empty_blob_oid(&p->two->oid))
 				continue;
-			else if (add_rename_dst(p->two) < 0) {
+			else if (add_rename_dst(p) < 0) {
 				warning("skipping rename detection, detected"
 					" duplicate destination '%s'",
 					p->two->path);

@@ -586,10 +577,10 @@ void diffcore_rename(struct diff_options *options)
 	mx = xcalloc(st_mult(NUM_CANDIDATE_PER_DST, num_destinations),
 		     sizeof(*mx));
 	for (dst_cnt = i = 0; i < rename_dst_nr; i++) {
-		struct diff_filespec *two = rename_dst[i].two;
+		struct diff_filespec *two = rename_dst[i].p->two;
 		struct diff_score *m;
 
-		if (rename_dst[i].pair)
+		if (rename_dst[i].is_rename)
 			continue; /* dealt with exact match already. */
 
 		m = &mx[dst_cnt * NUM_CANDIDATE_PER_DST];

@@ -646,22 +637,8 @@ void diffcore_rename(struct diff_options *options)
 			diff_q(&outq, p);
 		}
 		else if (!DIFF_FILE_VALID(p->one) && DIFF_FILE_VALID(p->two)) {
-			/*
-			 * Creation
-			 *
-			 * We would output this create record if it has
-			 * not been turned into a rename/copy already.
-			 */
-			struct diff_rename_dst *dst = locate_rename_dst(p->two);
-			if (dst && dst->pair) {
-				diff_q(&outq, dst->pair);
-				pair_to_free = p;
-			}
-			else
-				/* no matching rename/copy source, so
-				 * record this as a creation.
-				 */
-				diff_q(&outq, p);
+			/* Creation */
+			diff_q(&outq, p);
 		}
 		else if (DIFF_FILE_VALID(p->one) && !DIFF_FILE_VALID(p->two)) {
 			/*

@@ -682,8 +659,10 @@ void diffcore_rename(struct diff_options *options)
 			 */
 			if (DIFF_PAIR_BROKEN(p)) {
 				/* broken delete */
-				struct diff_rename_dst *dst = locate_rename_dst(p->one);
-				if (dst && dst->pair)
+				struct diff_rename_dst *dst = locate_rename_dst(p);
+				if (!dst)
+					BUG("tracking failed somehow; failed to find associated dst for broken pair");
+				if (dst->is_rename)
 					/* counterpart is now rename/copy */
 					pair_to_free = p;
 			}

@@ -693,16 +672,14 @@ void diffcore_rename(struct diff_options *options)
 					pair_to_free = p;
 			}
 
-			if (pair_to_free)
-				;
-			else
+			if (!pair_to_free)
 				diff_q(&outq, p);
 		}
 		else if (!diff_unmodified_pair(p))
 			/* all the usual ones need to be kept */
 			diff_q(&outq, p);
 		else
-			/* no need to keep unmodified pairs */
+			/* no need to keep unmodified pairs; FIXME: remove earlier? */
 			pair_to_free = p;
 
 		if (pair_to_free)

@@ -715,11 +692,16 @@ void diffcore_rename(struct diff_options *options)
 	diff_debug_queue("done collapsing", q);
 
 	for (i = 0; i < rename_dst_nr; i++)
-		free_filespec(rename_dst[i].two);
+		if (rename_dst[i].filespec_to_free)
+			free_filespec(rename_dst[i].filespec_to_free);
 
 	FREE_AND_NULL(rename_dst);
 	rename_dst_nr = rename_dst_alloc = 0;
 	FREE_AND_NULL(rename_src);
 	rename_src_nr = rename_src_alloc = 0;
+	if (break_idx) {
+		strintmap_clear(break_idx);
+		FREE_AND_NULL(break_idx);
+	}
 	return;
 }

gitgitgadget

From: Elijah Newren <[email protected]>

Commit 25d5ea410f ("[PATCH] Redo rename/copy detection logic.", 2005-05-24) added a duplicate entry check on rename_src in order to avoid segfaults; the code at the time was prone to double free()s and an easy way to avoid it was just to turn off rename detection for any duplicate entries. Note that the form of the check was modified two commits ago in this series.

Similarly, commit 4d6be03b95 ("diffcore-rename: avoid processing duplicate destinations", 2015-02-26) added a duplicate entry check on rename_dst for the exact same reason -- the code was prone to double free()s, and an easy way to avoid it was just to turn off rename detection entirely. Note that the form of the check was modified in the commit just before this one.

In the original code in both places, the code was dealing with individual diff_filespecs and trying to match things up, instead of just keeping the original diff_filepairs around as we do now. The intervening change in structure has fixed the accounting problems and the associated double free()s that used to occur, and thus we already have a better fix. As such, we can remove the band-aid checks for duplicate entries.

Due to the last two patches, the diffcore_rename() setup is no longer a sizeable chunk of overall runtime. Thus, in a large rebase of many commits with lots of renames and several optimizations to inexact rename detection, this patch only speeds up the overall code by about half a percent or so and is pretty close to the run-to-run variability making it hard to get an exact measurement. However, with some trace2 regions around the setup code in diffcore_rename() so that I can focus on just it, I measure that this patch consistently saves almost a third of the remaining time spent in diffcore_rename() setup.

---
 diffcore-rename.c          | 23 -----------------------
 t/t4058-diff-duplicates.sh |  2 +-
 2 files changed, 1 insertion(+), 24 deletions(-)

index 2a8fcd52928..90db9ebd6d0 100644
@@ -34,18 +34,6 @@ static struct diff_rename_dst *locate_rename_dst(struct diff_filepair *p)
  */
 static int add_rename_dst(struct diff_filepair *p)
 {
-	/*
-	 * If we have multiple entries at the same path in the destination
-	 * tree (an invalid tree, to be sure), turn off rename detection
-	 * entirely.  Once upon a time, diffcore-rename had double free()
-	 * issues due to such duplicate paths, resulting in segfaults.  See
-	 * commit 4d6be03b95 ("diffcore-rename: avoid processing duplicate
-	 * destinations", 2015-02-26).
-	 */
-	if (rename_dst_nr > 0 &&
-	    !strcmp(rename_dst[rename_dst_nr-1].p->two->path, p->two->path))
-		return -1;
-
 	ALLOC_GROW(rename_dst, rename_dst_nr + 1, rename_dst_alloc);
 	rename_dst[rename_dst_nr].p = p;
 	rename_dst[rename_dst_nr].filespec_to_free = NULL;

@@ -63,17 +51,6 @@ static int rename_src_nr, rename_src_alloc;
 
 static void register_rename_src(struct diff_filepair *p)
 {
-	/*
-	 * If we have multiple entries at the same path in the source tree
-	 * (an invalid tree, to be sure), avoid using more more than one
-	 * such entry in rename detection.  Once upon a time, doing so
-	 * caused segfaults; see commit 25d5ea410f ("[PATCH] Redo
-	 * rename/copy detection logic.", 2005-05-24).
-	 */
-	if (rename_src_nr > 0 &&
-	    !strcmp(rename_src[rename_src_nr-1].p->one->path, p->one->path))
-		return;
-
 	if (p->broken_pair) {
 		if (!break_idx) {
 			break_idx = xmalloc(sizeof(*break_idx));

index ad3f37d388d..54614b814db 100755
@@ -91,7 +91,7 @@ test_expect_success 'diff-tree between duplicate trees' '
 test_expect_success 'diff-tree with renames' '
 	# See NOTICE at top of file.
 	git diff-tree -M -r --no-abbrev one two >actual &&
-	test_cmp expect actual
+	test_must_be_empty actual
 '
 
 test_expect_success 'diff-tree FROM duplicate tree' '

gitgitgadget

About the subject:

s/unneccessary/unnecessary/

Thanks, I'll remove the unnecessary double c.

This patch series includes 3 small code cleanups, 1 small bug fix (or perhaps just a UI improvement -- it's a very minor issue either way), more testcases, and 3 performance improvements. The first 7 patches are relatively easy to review, while the second to last one...is a bit more involved (but provided the biggest performance boost).

Changes since v2:

* Fixed spelling error: remove unnecessary double c in 'unneccessary' (thanks to Christian for spotting)

Elijah Newren (9):
diffcore-rename: rename num_create to num_destinations
diffcore-rename: avoid usage of global in too_many_rename_candidates()
diffcore-rename: simplify limit check
diffcore-rename: reduce jumpiness in progress counters
t4058: add more tests and documentation for duplicate tree entry
handling
t4058: explore duplicate tree entry handling in a bit more detail
diffcore-rename: simplify and accelerate register_rename_src()
diffcore-rename: accelerate rename_dst setup
diffcore-rename: remove unnecessary duplicate entry checks

 diffcore-rename.c          | 209 ++++++++++++++-----------------------
 t/t4058-diff-duplicates.sh | 114 +++++++++++++++++++-
 2 files changed, 192 insertions(+), 131 deletions(-)

base-commit: 3a0b884caba2752da0af626fb2de7d597c844e8b
Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-git-929%2Fnewren%2Fdiffcore-rename-improvements-v3
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-git-929/newren/diffcore-rename-improvements-v3
Pull-Request: https://github.com/git/git/pull/929

Range-diff vs v2:

  1:  428d8204894 =  1:  428d8204894 diffcore-rename: rename num_create to num_destinations
  2:  fc62f4c4f89 =  2:  fc62f4c4f89 diffcore-rename: avoid usage of global in too_many_rename_candidates()
  3:  7214fa73fdd =  3:  7214fa73fdd diffcore-rename: simplify limit check
  4:  bf3581b1ac1 =  4:  bf3581b1ac1 diffcore-rename: reduce jumpiness in progress counters
  5:  9a4a3159acf =  5:  9a4a3159acf t4058: add more tests and documentation for duplicate tree entry handling
  6:  8db27892c59 =  6:  8db27892c59 t4058: explore duplicate tree entry handling in a bit more detail
  7:  a58639b2927 =  7:  a58639b2927 diffcore-rename: simplify and accelerate register_rename_src()
  8:  85714e1583d =  8:  85714e1583d diffcore-rename: accelerate rename_dst setup
  9:  af256b1c534 !  9:  723b0e55e75 diffcore-rename: remove unneccessary duplicate entry checks

@@ Metadata
Author: Elijah Newren <[email protected]>

## Commit message ##
- diffcore-rename: remove unneccessary duplicate entry checks
+ diffcore-rename: remove unnecessary duplicate entry checks

Commit 25d5ea410f ("[PATCH] Redo rename/copy detection logic.", 2005-05-24) added a duplicate entry check on rename_src in order to

gitgitgadget

From: Elijah Newren <[email protected]>

Our main data structures are rename_src and rename_dst. For counters of these data structures, num_sources and num_destinations seem natural; definitely more so than using num_create for the latter.

---
 diffcore-rename.c | 25 +++++++++++++------------
 1 file changed, 13 insertions(+), 12 deletions(-)

index d367a6d2443..15a98f566e4 100644
@@ -434,7 +434,7 @@ static void record_if_better(struct diff_score m[], struct diff_score *o)
  * 1 if we need to disable inexact rename detection;
  * 2 if we would be under the limit if we were given -C instead of -C -C.
  */
-static int too_many_rename_candidates(int num_create,
+static int too_many_rename_candidates(int num_destinations,
 				      struct diff_options *options)
 {
 	int rename_limit = options->rename_limit;

@@ -447,17 +447,17 @@ static int too_many_rename_candidates(int num_create,
 	 * This basically does a test for the rename matrix not
 	 * growing larger than a "rename_limit" square matrix, ie:
 	 *
-	 *    num_create * num_src > rename_limit * rename_limit
+	 *    num_destinations * num_src > rename_limit * rename_limit
 	 */
 	if (rename_limit <= 0)
 		rename_limit = 32767;
-	if ((num_create <= rename_limit || num_src <= rename_limit) &&
-	    ((uint64_t)num_create * (uint64_t)num_src
+	if ((num_destinations <= rename_limit || num_src <= rename_limit) &&
+	    ((uint64_t)num_destinations * (uint64_t)num_src
 	     <= (uint64_t)rename_limit * (uint64_t)rename_limit))
 		return 0;
 
 	options->needed_rename_limit =
-		num_src > num_create ? num_src : num_create;
+		num_src > num_destinations ? num_src : num_destinations;
 
 	/* Are we running under -C -C? */
 	if (!options->flags.find_copies_harder)

@@ -469,8 +469,8 @@ static int too_many_rename_candidates(int num_create,
 			continue;
 		num_src++;
 	}
-	if ((num_create <= rename_limit || num_src <= rename_limit) &&
-	    ((uint64_t)num_create * (uint64_t)num_src
+	if ((num_destinations <= rename_limit || num_src <= rename_limit) &&
+	    ((uint64_t)num_destinations * (uint64_t)num_src
 	     <= (uint64_t)rename_limit * (uint64_t)rename_limit))
 		return 2;
 	return 1;

@@ -505,7 +505,7 @@ void diffcore_rename(struct diff_options *options)
 	struct diff_queue_struct outq;
 	struct diff_score *mx;
 	int i, j, rename_count, skip_unmodified = 0;
-	int num_create, dst_cnt;
+	int num_destinations, dst_cnt;
 	struct progress *progress = NULL;
 
 	if (!minimum_score)

@@ -570,13 +570,13 @@ void diffcore_rename(struct diff_options *options)
 	 * Calculate how many renames are left (but all the source
 	 * files still remain as options for rename/copies!)
 	 */
-	num_create = (rename_dst_nr - rename_count);
+	num_destinations = (rename_dst_nr - rename_count);
 
 	/* All done? */
-	if (!num_create)
+	if (!num_destinations)
 		goto cleanup;
 
-	switch (too_many_rename_candidates(num_create, options)) {
+	switch (too_many_rename_candidates(num_destinations, options)) {
 	case 1:
 		goto cleanup;
 	case 2:

@@ -593,7 +593,8 @@ void diffcore_rename(struct diff_options *options)
 				(uint64_t)rename_dst_nr * (uint64_t)rename_src_nr);
 	}
 
-	mx = xcalloc(st_mult(NUM_CANDIDATE_PER_DST, num_create), sizeof(*mx));
+	mx = xcalloc(st_mult(NUM_CANDIDATE_PER_DST, num_destinations),
+		     sizeof(*mx));
 	for (dst_cnt = i = 0; i < rename_dst_nr; i++) {
 		struct diff_filespec *two = rename_dst[i].two;
 		struct diff_score *m;

gitgitgadget

From: Elijah Newren <[email protected]>

too_many_rename_candidates() got the number of rename destinations via an argument to the function, but the number of rename sources via a global variable. That felt rather inconsistent. Pass in the number of rename sources as an argument as well.

While we are at it... We had a local variable, num_src, that served two purposes. Initially it was set to the global value, but later was used for counting a subset of the number of sources. Since we now have a function argument for the former usage, introduce a clearer variable name for the latter usage.

This patch has no behavioral changes; it's just renaming and passing an argument instead of grabbing it from the global namespace. (You may find it easier to view the patch using git diff's --color-words option.)

---
 diffcore-rename.c | 24 ++++++++++++------------
 1 file changed, 12 insertions(+), 12 deletions(-)

index 15a98f566e4..1d6675c040d 100644
@@ -434,12 +434,11 @@ static void record_if_better(struct diff_score m[], struct diff_score *o)
  * 1 if we need to disable inexact rename detection;
  * 2 if we would be under the limit if we were given -C instead of -C -C.
  */
-static int too_many_rename_candidates(int num_destinations,
+static int too_many_rename_candidates(int num_destinations, int num_sources,
 				      struct diff_options *options)
 {
 	int rename_limit = options->rename_limit;
-	int num_src = rename_src_nr;
-	int i;
+	int i, limited_sources;
 
 	options->needed_rename_limit = 0;
 

@@ -447,30 +446,30 @@ static int too_many_rename_candidates(int num_destinations,
 	 * This basically does a test for the rename matrix not
 	 * growing larger than a "rename_limit" square matrix, ie:
 	 *
-	 *    num_destinations * num_src > rename_limit * rename_limit
+	 *    num_destinations * num_sources > rename_limit * rename_limit
 	 */
 	if (rename_limit <= 0)
 		rename_limit = 32767;
-	if ((num_destinations <= rename_limit || num_src <= rename_limit) &&
-	    ((uint64_t)num_destinations * (uint64_t)num_src
+	if ((num_destinations <= rename_limit || num_sources <= rename_limit) &&
+	    ((uint64_t)num_destinations * (uint64_t)num_sources
 	     <= (uint64_t)rename_limit * (uint64_t)rename_limit))
 		return 0;
 
 	options->needed_rename_limit =
-		num_src > num_destinations ? num_src : num_destinations;
+		num_sources > num_destinations ? num_sources : num_destinations;
 
 	/* Are we running under -C -C? */
 	if (!options->flags.find_copies_harder)
 		return 1;
 
 	/* Would we bust the limit if we were running under -C? */
-	for (num_src = i = 0; i < rename_src_nr; i++) {
+	for (limited_sources = i = 0; i < num_sources; i++) {
 		if (diff_unmodified_pair(rename_src[i].p))
 			continue;
-		num_src++;
+		limited_sources++;
 	}
-	if ((num_destinations <= rename_limit || num_src <= rename_limit) &&
-	    ((uint64_t)num_destinations * (uint64_t)num_src
+	if ((num_destinations <= rename_limit || limited_sources <= rename_limit) &&
+	    ((uint64_t)num_destinations * (uint64_t)limited_sources
 	     <= (uint64_t)rename_limit * (uint64_t)rename_limit))
 		return 2;
 	return 1;

@@ -576,7 +575,8 @@ void diffcore_rename(struct diff_options *options)
 	if (!num_destinations)
 		goto cleanup;
 
-	switch (too_many_rename_candidates(num_destinations, options)) {
+	switch (too_many_rename_candidates(num_destinations, rename_src_nr,
+					   options)) {
 	case 1:
 		goto cleanup;
 	case 2:

gitgitgadget

From: Elijah Newren <[email protected]>

diffcore-rename had two different checks of the form

if ((a < limit || b < limit) &&
a * b <= limit * limit)

This can be simplified to

    if (st_mult(a, b) <= st_mult(limit, limit))

which makes it clearer how we are checking for overflow, and makes it much easier to parse given the drop from 8 to 4 variable appearances.

---
 diffcore-rename.c | 15 +++++++++------
 1 file changed, 9 insertions(+), 6 deletions(-)

index 1d6675c040d..16553ab259f 100644
@@ -447,12 +447,16 @@ static int too_many_rename_candidates(int num_destinations, int num_sources,
 	 * growing larger than a "rename_limit" square matrix, ie:
 	 *
 	 *    num_destinations * num_sources > rename_limit * rename_limit
+	 *
+	 * We use st_mult() to check overflow conditions; in the
+	 * exceptional circumstance that size_t isn't large enough to hold
+	 * the multiplication, the system won't be able to allocate enough
+	 * memory for the matrix anyway.
 	 */
 	if (rename_limit <= 0)
 		rename_limit = 32767;
-	if ((num_destinations <= rename_limit || num_sources <= rename_limit) &&
-	    ((uint64_t)num_destinations * (uint64_t)num_sources
-	     <= (uint64_t)rename_limit * (uint64_t)rename_limit))
+	if (st_mult(num_destinations, num_sources)
+	    <= st_mult(rename_limit, rename_limit))
 		return 0;
 
 	options->needed_rename_limit =

@@ -468,9 +472,8 @@ static int too_many_rename_candidates(int num_destinations, int num_sources,
 			continue;
 		limited_sources++;
 	}
-	if ((num_destinations <= rename_limit || limited_sources <= rename_limit) &&
-	    ((uint64_t)num_destinations * (uint64_t)limited_sources
-	     <= (uint64_t)rename_limit * (uint64_t)rename_limit))
+	if (st_mult(num_destinations, limited_sources)
+	    <= st_mult(rename_limit, rename_limit))
 		return 2;
 	return 1;
 }

gitgitgadget

Hi all,

First post on Git mailing list, to provide a comment on a patch. Hope this works.

In cases where the rename_limit is already greater than 65535, the st_mult(rename_limit, rename_limit) multiplication overflows and process halts. But I don't think an 'overflow error' is very helpful for the users in understanding what is wrong with their configuration; i.e. diff.renameLimit documentation says nothing about a 'maximum allowed value'. I would either clamp it to a reasonable range, or inform the users about the limits, or maybe both.

Cheers,

Basar

Out of curiosity, what system are you on? And how many renames do you actually have?

We used to clamp to 32767, but one specific repository needed values larger than that, in the range of ~50000. However, due to a number of rename detection related optimizations added since then, the git of today can handle that same particular repository and support full rename detection with a rename limit under 1000 for merge/cherry-pick (sorry, forgot the exact numbers), and probably less than 10000 for diff (just guestimating; I don't want to go and check).

Anyway, all that aside, I don't see any such overflow for rename_limit being 2**16; we can push it much farther:

    size_t a = 4294967295;   /*  2**32 -1  */
    size_t b = a;
    size_t c = st_mult(a, b);
    printf("%"PRIuMAX" = %"PRIuMAX" * %"PRIuMAX"\n", c, a, b);

Output:

    18446744065119617025 = 4294967295 * 4294967295

Adding one to the value of a results in:

    fatal: size_t overflow: 4294967296 * 4294967296

That sounds reasonable, but I'm not sure it's worth the effort in practice. 2**32 - 1 is so impractically large for a rename_limit that I don't see why anyone would need a value even remotely close to that level (and I wouldn't at all be surprised if other things in git scaling broke before we even got to that point).

Perhaps you're dealing with a 32-bit system? Even then, the repository that hit this was about 6.5GB packed .git history; and you might need to be a lot larger than that given the optimization since then in order to hit this. Can 32-bit systems even handle that size of repository without dying in several other ways?

I am on a 64-bit Windows 10; but following up on your question, it became obvious that these limits have something to do with 'which git executable' I was dealing with. This problem surfaced when I was working on Visual Studio 2019, and was trying to rename not more than 10 files. My config had 999999 as the renameLimit, and VS2019 showed the 'fatal error' in its git output. However, git bash was all fine with listing the renamed files. And the difference between these scenarios turned out to be, yes, different git executables. VS2019 has its own copy of git which is 32-bit, and it hits this 999999 * 999999 overflow; whereas *my* copy of git used in bash is 64-bit which does not have that overflow problem.

Good point, but the system aside, 2**16 - 1 = 65535 would remain to be the limit for the 32-bit git executables, wherever they are used. Therefore, maybe there is a point to curb it, or mention this somewhere, as I have said before.

Basar

Ah, thanks for the extra info.

Fair enough. If someone wants to contribute a patch to either provide a nicer error message if the value is set too high, or to clamp it with a warning, and in either case to make sure the too-large check is specific to 32-bit systems (or uses appropriately different limits on 32-bit and 64-bit systems), then that would be a welcome contribution.

From: Elijah Newren <[email protected]>

Inexact rename detection works by comparing all sources to all destinations, computing similarities, and then finding the best matches among those that are sufficiently similar.

However, it is preceded by exact rename detection that works by checking if there are files with identical hashes. If exact renames are found, we can exclude some files from inexact rename detection.

The inexact rename detection loops over the full set of files, but immediately skips those for which rename_dst[i].is_rename is true and thus doesn't compare any sources to that destination. As such, these paths shouldn't be included in the progress counter.

For the eagle eyed, this change hints at an actual optimization -- the first one I presented at Git Merge 2020. I'll be submitting that optimization later, once the basic merge-ort algorithm has merged.

---
 diffcore-rename.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

index 16553ab259f..55a188abcc3 100644
@@ -593,7 +593,7 @@ void diffcore_rename(struct diff_options *options)
 	if (options->show_rename_progress) {
 		progress = start_delayed_progress(
 				_("Performing inexact rename detection"),
-				(uint64_t)rename_dst_nr * (uint64_t)rename_src_nr);
+				(uint64_t)num_destinations * (uint64_t)rename_src_nr);
 	}
 
 	mx = xcalloc(st_mult(NUM_CANDIDATE_PER_DST, num_destinations),

@@ -633,7 +633,8 @@ void diffcore_rename(struct diff_options *options)
 			diff_free_filespec_blob(two);
 		}
 		dst_cnt++;
-		display_progress(progress, (uint64_t)(i+1)*(uint64_t)rename_src_nr);
+		display_progress(progress,
+				 (uint64_t)dst_cnt * (uint64_t)rename_src_nr);
 	}
 	stop_progress(&progress);
 

gitgitgadget

From: Elijah Newren <[email protected]>

Commit 4d6be03b95 ("diffcore-rename: avoid processing duplicate destinations", 2015-02-26) added t4058 to demonstrate that a workaround it added to avoid double frees (namely to just turn off rename detection when trees had duplicate entries) would indeed avoid segfaults. The tests, though, give the impression that the expected diffs are "correct" when in reality they are just "don't segfault, and do something semi-reasonable under the circumstances". Add some notes to make this clearer.

Also, commit 25d5ea410f ("[PATCH] Redo rename/copy detection logic.", 2005-05-24) added a similar workaround to avoid segfaults, but for rename_src rather than rename_dst. I do not see any tests in the testsuite to cover the collision detection of entries limited to the source side, so add a couple.

---
 t/t4058-diff-duplicates.sh | 47 ++++++++++++++++++++++++++++++++++++--
 1 file changed, 45 insertions(+), 2 deletions(-)

index c24ee175ef0..bd685089561 100755
@@ -1,5 +1,14 @@
 #!/bin/sh
 
+# NOTICE:
+#   This testsuite does a number of diffs and checks that the output match.
+#   However, it is a "garbage in, garbage out" situation; the trees have
+#   duplicate entries for individual paths, and it results in diffs that do
+#   not make much sense.  As such, it is not clear that the diffs are
+#   "correct".  The primary purpose of these tests was to verify that
+#   diff-tree does not segfault, but there is perhaps some value in ensuring
+#   that the diff output isn't wildly unreasonable.
+
 test_description='test tree diff when trees have duplicate entries'
 . ./test-lib.sh
 

@@ -57,7 +66,16 @@ test_expect_success 'create trees with duplicate entries' '
 	git tag two $outer_two
 '
 
-test_expect_success 'diff-tree between trees' '
+test_expect_success 'create tree without duplicate entries' '
+	blob_one=$(echo one | git hash-object -w --stdin) &&
+	outer_three=$(make_tree \
+		100644 renamed $blob_one
+	) &&
+	git tag three $outer_three
+'
+
+test_expect_success 'diff-tree between duplicate trees' '
+	# See NOTICE at top of file
 	{
 		printf ":000000 100644 $ZERO_OID $blob_two A\touter/inner\n" &&
 		printf ":000000 100644 $ZERO_OID $blob_two A\touter/inner\n" &&

@@ -71,9 +89,34 @@ test_expect_success 'diff-tree between trees' '
 '
 
 test_expect_success 'diff-tree with renames' '
-	# same expectation as above, since we disable rename detection
+	# See NOTICE at top of file.
 	git diff-tree -M -r --no-abbrev one two >actual &&
 	test_cmp expect actual
 '
 
+test_expect_success 'diff-tree FROM duplicate tree' '
+	# See NOTICE at top of file.
+	{
+		printf ":100644 000000 $blob_one $ZERO_OID D\touter/inner\n" &&
+		printf ":100644 000000 $blob_two $ZERO_OID D\touter/inner\n" &&
+		printf ":100644 000000 $blob_two $ZERO_OID D\touter/inner\n" &&
+		printf ":100644 000000 $blob_two $ZERO_OID D\touter/inner\n" &&
+		printf ":000000 100644 $ZERO_OID $blob_one A\trenamed\n"
+	} >expect &&
+	git diff-tree -r --no-abbrev one three >actual &&
+	test_cmp expect actual
+'
+
+test_expect_success 'diff-tree FROM duplicate tree, with renames' '
+	# See NOTICE at top of file.
+	{
+		printf ":100644 000000 $blob_two $ZERO_OID D\touter/inner\n" &&
+		printf ":100644 000000 $blob_two $ZERO_OID D\touter/inner\n" &&
+		printf ":100644 000000 $blob_two $ZERO_OID D\touter/inner\n" &&
+		printf ":100644 100644 $blob_one $blob_one R100\touter/inner\trenamed\n"
+	} >expect &&
+	git diff-tree -M -r --no-abbrev one three >actual &&
+	test_cmp expect actual
+'
+
 test_done

gitgitgadget

From: Elijah Newren <[email protected]>

While creating the last commit, I found a number of other cases where git would segfault when faced with trees that have duplicate entries. None of these segfaults are in the diffcore-rename code (they all occur in cache-tree and unpack-trees). Further, to my knowledge, no one has ever been adversely affected by these bugs, and given that it has been 15 years and folks have fixed a few other issues with historical duplicate entries (as noted in the last commit), I am not sure we will ever run into anyone having problems with these. So I am not sure these are worth fixing, but it doesn't hurt to at least document these failures in the same test file that is concerned with duplicate tree entries.

---
 t/t4058-diff-duplicates.sh | 67 ++++++++++++++++++++++++++++++++++++++
 1 file changed, 67 insertions(+)

index bd685089561..ad3f37d388d 100755
@@ -119,4 +119,71 @@ test_expect_success 'diff-tree FROM duplicate tree, with renames' '
 	test_cmp expect actual
 '
 
+test_expect_success 'create a few commits' '
+	git commit-tree -m "Duplicate Entries" two^{tree} >commit_id &&
+	git branch base $(cat commit_id) &&
+
+	git commit-tree -p $(cat commit_id) -m "Just one" three^{tree} >up &&
+	git branch update $(cat up) &&
+
+	git commit-tree -p $(cat up) -m "Back to weird" two^{tree} >final &&
+	git branch final $(cat final) &&
+
+	rm commit_id up final
+'
+
+test_expect_failure 'git read-tree does not segfault' '
+	test_when_finished rm .git/index.lock &&
+	test_might_fail git read-tree --reset base
+'
+
+test_expect_failure 'reset --hard does not segfault' '
+	test_when_finished rm .git/index.lock &&
+	git checkout base &&
+	test_might_fail git reset --hard
+'
+
+test_expect_failure 'git diff HEAD does not segfault' '
+	git checkout base &&
+	GIT_TEST_CHECK_CACHE_TREE=false &&
+	git reset --hard &&
+	test_might_fail git diff HEAD
+'
+
+test_expect_failure 'can switch to another branch when status is empty' '
+	git clean -ffdqx &&
+	git status --porcelain -uno >actual &&
+	test_must_be_empty actual &&
+	git checkout update
+'
+
+test_expect_success 'forcibly switch to another branch, verify status empty' '
+	git checkout -f update &&
+	git status --porcelain -uno >actual &&
+	test_must_be_empty actual
+'
+
+test_expect_success 'fast-forward from non-duplicate entries to duplicate' '
+	git merge final
+'
+
+test_expect_failure 'clean status, switch branches, status still clean' '
+	git status --porcelain -uno >actual &&
+	test_must_be_empty actual &&
+	git checkout base &&
+	git status --porcelain -uno >actual &&
+	test_must_be_empty actual
+'
+
+test_expect_success 'switch to base branch and force status to be clean' '
+	git checkout base &&
+	GIT_TEST_CHECK_CACHE_TREE=false git reset --hard &&
+	git status --porcelain -uno >actual &&
+	test_must_be_empty actual
+'
+
+test_expect_failure 'fast-forward from duplicate entries to non-duplicate' '
+	git merge update
+'
+
 test_done

gitgitgadget

From: Elijah Newren <[email protected]>

register_rename_src() took pains to create an array in rename_src which was sorted by pathname of the contained diff_filepair. The sorting was entirely unnecessary since callers pass filepairs to us in sorted order. We can simply append to the end of the rename_src array, speeding up diffcore_rename() setup time.

Also, note that I dropped the return type on the function since it was unconditionally discarded anyway.

This patch is being submitted in a different order than its original development, but in a large rebase of many commits with lots of renames and with several optimizations to inexact rename detection, diffcore_rename() setup time was a sizeable chunk of overall runtime. This patch dropped execution time of rebasing 35 commits with lots of renames by 2% overall.

---
 diffcore-rename.c | 39 +++++++++++++--------------------------
 1 file changed, 13 insertions(+), 26 deletions(-)

index 55a188abcc3..a215421a9cb 100644
@@ -76,36 +76,23 @@ static struct diff_rename_src {
 } *rename_src;
 static int rename_src_nr, rename_src_alloc;
 
-static struct diff_rename_src *register_rename_src(struct diff_filepair *p)
+static void register_rename_src(struct diff_filepair *p)
 {
-	int first, last;
-	struct diff_filespec *one = p->one;
-	unsigned short score = p->score;
-
-	first = 0;
-	last = rename_src_nr;
-	while (last > first) {
-		int next = first + ((last - first) >> 1);
-		struct diff_rename_src *src = &(rename_src[next]);
-		int cmp = strcmp(one->path, src->p->one->path);
-		if (!cmp)
-			return src;
-		if (cmp < 0) {
-			last = next;
-			continue;
-		}
-		first = next+1;
-	}
+	/*
+	 * If we have multiple entries at the same path in the source tree
+	 * (an invalid tree, to be sure), avoid using more more than one
+	 * such entry in rename detection.  Once upon a time, doing so
+	 * caused segfaults; see commit 25d5ea410f ("[PATCH] Redo
+	 * rename/copy detection logic.", 2005-05-24).
+	 */
+	if (rename_src_nr > 0 &&
+	    !strcmp(rename_src[rename_src_nr-1].p->one->path, p->one->path))
+		return;
 
-	/* insert to make it at "first" */
 	ALLOC_GROW(rename_src, rename_src_nr + 1, rename_src_alloc);
+	rename_src[rename_src_nr].p = p;
+	rename_src[rename_src_nr].score = p->score;
 	rename_src_nr++;
-	if (first < rename_src_nr)
-		MOVE_ARRAY(rename_src + first + 1, rename_src + first,
-			   rename_src_nr - first - 1);
-	rename_src[first].p = p;
-	rename_src[first].score = score;
-	return &(rename_src[first]);
 }
 
 static int basename_same(struct diff_filespec *src, struct diff_filespec *dst)

gitgitgadget

From: Elijah Newren <[email protected]>

register_rename_src() simply references the passed pair inside rename_src. In contrast, add_rename_dst() did something entirely different for rename_dst. Instead of copying the passed pair, it made a copy of the second diff_filespec from the passed pair, referenced it, and then set the diff_rename_dst.pair field to NULL. Later, when a pairing is found, record_rename_pair() allocated a full diff_filepair via diff_queue() and pointed its src and dst fields at the appropriate diff_filespecs. This contrast between register_rename_src() for the rename_src data structure and add_rename_dst() for the rename_dst data structure is oddly inconsistent and requires more memory and work than necessary. Let's just reference the original diff_filepair in rename_dst as-is, just as we do with rename_src. Add a new rename_dst.is_rename field, since the rename_dst.p field is never NULL unlike the old rename_dst.pair field.

Taking advantage of this change and the fact that same-named paths will be adjacent, we can get rid of the sorting of the array and most of the lookups on it, allowing us to instead just append as we go. However, there is one remaining reason to still keep locate_rename_dst():
handling broken pairs (i.e. when break detection is on). Those are somewhat rare, but we can set up a simple strintmap to get the map between the source and the index. Doing that allows us to still have a fast lookup without sorting the rename_dst array. Since the sorting had been done in a weakly quadratic manner, when many renames are involved this time could add up.

There is still a strcmp() in add_rename_dst() that I have left in place to make it easier to verify that the algorithm has the same results. This strcmp() is there to check for duplicate destination entries (which was the easiest way at the time to avoid segfaults in the diffcore-rename code when trees had multiple entries at a given path). The underlying double free()s are no longer an issue with the new algorithm, but that can be addressed in a subsequent commit.

This patch is being submitted in a different order than its original development, but in a large rebase of many commits with lots of renames and with several optimizations to inexact rename detection, both setup time and write back to output queue time from diffcore_rename() were sizeable chunks of overall runtime. This patch accelerated the setup time by about 65%, and final write back to the output queue time by about 50%, resulting in an overall drop of 3.5% on the execution time of rebasing a few dozen patches.

---
 diffcore-rename.c | 148 ++++++++++++++++++++--------------------------
 1 file changed, 65 insertions(+), 83 deletions(-)

index a215421a9cb..2a8fcd52928 100644
@@ -9,63 +9,48 @@
 #include "hashmap.h"
 #include "progress.h"
 #include "promisor-remote.h"
+#include "strmap.h"
 
 /* Table of rename/copy destinations */
 
 static struct diff_rename_dst {
-	struct diff_filespec *two;
-	struct diff_filepair *pair;
+	struct diff_filepair *p;
+	struct diff_filespec *filespec_to_free;
+	int is_rename; /* false -> just a create; true -> rename or copy */
 } *rename_dst;
 static int rename_dst_nr, rename_dst_alloc;
+/* Mapping from break source pathname to break destination index */
+static struct strintmap *break_idx = NULL;
 
-static int find_rename_dst(struct diff_filespec *two)
-{
-	int first, last;
-
-	first = 0;
-	last = rename_dst_nr;
-	while (last > first) {
-		int next = first + ((last - first) >> 1);
-		struct diff_rename_dst *dst = &(rename_dst[next]);
-		int cmp = strcmp(two->path, dst->two->path);
-		if (!cmp)
-			return next;
-		if (cmp < 0) {
-			last = next;
-			continue;
-		}
-		first = next+1;
-	}
-	return -first - 1;
-}
-
-static struct diff_rename_dst *locate_rename_dst(struct diff_filespec *two)
+static struct diff_rename_dst *locate_rename_dst(struct diff_filepair *p)
 {
-	int ofs = find_rename_dst(two);
-	return ofs < 0 ? NULL : &rename_dst[ofs];
+	/* Lookup by p->ONE->path */
+	int idx = break_idx ? strintmap_get(break_idx, p->one->path) : -1;
+	return (idx == -1) ? NULL : &rename_dst[idx];
 }
 
 /*
  * Returns 0 on success, -1 if we found a duplicate.
  */
-static int add_rename_dst(struct diff_filespec *two)
+static int add_rename_dst(struct diff_filepair *p)
 {
-	int first = find_rename_dst(two);
-
-	if (first >= 0)
+	/*
+	 * If we have multiple entries at the same path in the destination
+	 * tree (an invalid tree, to be sure), turn off rename detection
+	 * entirely.  Once upon a time, diffcore-rename had double free()
+	 * issues due to such duplicate paths, resulting in segfaults.  See
+	 * commit 4d6be03b95 ("diffcore-rename: avoid processing duplicate
+	 * destinations", 2015-02-26).
+	 */
+	if (rename_dst_nr > 0 &&
+	    !strcmp(rename_dst[rename_dst_nr-1].p->two->path, p->two->path))
 		return -1;
-	first = -first - 1;
 
-	/* insert to make it at "first" */
 	ALLOC_GROW(rename_dst, rename_dst_nr + 1, rename_dst_alloc);
+	rename_dst[rename_dst_nr].p = p;
+	rename_dst[rename_dst_nr].filespec_to_free = NULL;
+	rename_dst[rename_dst_nr].is_rename = 0;
 	rename_dst_nr++;
-	if (first < rename_dst_nr)
-		MOVE_ARRAY(rename_dst + first + 1, rename_dst + first,
-			   rename_dst_nr - first - 1);
-	rename_dst[first].two = alloc_filespec(two->path);
-	fill_filespec(rename_dst[first].two, &two->oid, two->oid_valid,
-		      two->mode);
-	rename_dst[first].pair = NULL;
 	return 0;
 }
 

@@ -89,6 +74,14 @@ static void register_rename_src(struct diff_filepair *p)
 	    !strcmp(rename_src[rename_src_nr-1].p->one->path, p->one->path))
 		return;
 
+	if (p->broken_pair) {
+		if (!break_idx) {
+			break_idx = xmalloc(sizeof(*break_idx));
+			strintmap_init(break_idx, -1);
+		}
+		strintmap_set(break_idx, p->one->path, rename_dst_nr);
+	}
+
 	ALLOC_GROW(rename_src, rename_src_nr + 1, rename_src_alloc);
 	rename_src[rename_src_nr].p = p;
 	rename_src[rename_src_nr].score = p->score;

@@ -128,14 +121,14 @@ static void prefetch(void *prefetch_options)
 	struct oid_array to_fetch = OID_ARRAY_INIT;
 
 	for (i = 0; i < rename_dst_nr; i++) {
-		if (rename_dst[i].pair)
+		if (rename_dst[i].p->renamed_pair)
 			/*
 			 * The loop in diffcore_rename() will not need these
 			 * blobs, so skip prefetching.
 			 */
 			continue; /* already found exact match */
 		diff_add_if_missing(options->repo, &to_fetch,
-				    rename_dst[i].two);
+				    rename_dst[i].p->two);
 	}
 	for (i = 0; i < rename_src_nr; i++) {
 		if (options->skip_unmodified &&

@@ -245,26 +238,24 @@ static int estimate_similarity(struct repository *r,
 
 static void record_rename_pair(int dst_index, int src_index, int score)
 {
-	struct diff_filespec *src, *dst;
-	struct diff_filepair *dp;
+	struct diff_filepair *src = rename_src[src_index].p;
+	struct diff_filepair *dst = rename_dst[dst_index].p;
 
-	if (rename_dst[dst_index].pair)
+	if (dst->renamed_pair)
 		die("internal error: dst already matched.");
 
-	src = rename_src[src_index].p->one;
-	src->rename_used++;
-	src->count++;
+	src->one->rename_used++;
+	src->one->count++;
 
-	dst = rename_dst[dst_index].two;
-	dst->count++;
+	rename_dst[dst_index].filespec_to_free = dst->one;
+	rename_dst[dst_index].is_rename = 1;
 
-	dp = diff_queue(NULL, src, dst);
-	dp->renamed_pair = 1;
-	if (!strcmp(src->path, dst->path))
-		dp->score = rename_src[src_index].score;
+	dst->one = src->one;
+	dst->renamed_pair = 1;
+	if (!strcmp(dst->one->path, dst->two->path))
+		dst->score = rename_src[src_index].score;
 	else
-		dp->score = score;
-	rename_dst[dst_index].pair = dp;
+		dst->score = score;
 }
 
 /*