git/diffcore-rename.c

1723 lines
51 KiB
C
Raw Normal View History

/*
*
* Copyright (C) 2005 Junio C Hamano
*/
#include "git-compat-util.h"
#include "alloc.h"
#include "diff.h"
#include "diffcore.h"
#include "object-store.h"
#include "hashmap.h"
#include "mem-pool.h"
#include "oid-array.h"
#include "progress.h"
#include "promisor-remote.h"
#include "string-list.h"
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 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. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-11 09:08:47 +00:00
#include "strmap.h"
#include "trace2.h"
/* Table of rename/copy destinations */
static struct diff_rename_dst {
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 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. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-11 09:08:47 +00:00
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;
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 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. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-11 09:08:47 +00:00
/* Mapping from break source pathname to break destination index */
static struct strintmap *break_idx = NULL;
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 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. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-11 09:08:47 +00:00
static struct diff_rename_dst *locate_rename_dst(struct diff_filepair *p)
{
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 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. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-11 09:08:47 +00:00
/* 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.
*/
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 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. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-11 09:08:47 +00:00
static int add_rename_dst(struct diff_filepair *p)
{
ALLOC_GROW(rename_dst, rename_dst_nr + 1, rename_dst_alloc);
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 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. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-11 09:08:47 +00:00
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++;
return 0;
}
/* Table of rename/copy src files */
static struct diff_rename_src {
struct diff_filepair *p;
unsigned short score; /* to remember the break score */
} *rename_src;
static int rename_src_nr, rename_src_alloc;
static void register_rename_src(struct diff_filepair *p)
{
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 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. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-11 09:08:47 +00:00
if (p->broken_pair) {
if (!break_idx) {
break_idx = xmalloc(sizeof(*break_idx));
strintmap_init_with_options(break_idx, -1, NULL, 0);
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 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. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-11 09:08:47 +00:00
}
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;
rename_src_nr++;
}
static int basename_same(struct diff_filespec *src, struct diff_filespec *dst)
{
int src_len = strlen(src->path), dst_len = strlen(dst->path);
while (src_len && dst_len) {
char c1 = src->path[--src_len];
char c2 = dst->path[--dst_len];
if (c1 != c2)
return 0;
if (c1 == '/')
return 1;
}
return (!src_len || src->path[src_len - 1] == '/') &&
(!dst_len || dst->path[dst_len - 1] == '/');
}
struct diff_score {
int src; /* index in rename_src */
int dst; /* index in rename_dst */
unsigned short score;
short name_score;
};
diffcore-rename: use a different prefetch for basename comparisons merge-ort was designed to minimize the amount of data needed and used, and several changes were made to diffcore-rename to take advantage of extra metadata to enable this data minimization (particularly the relevant_sources variable for skipping "irrelevant" renames). This effort obviously succeeded in drastically reducing computation times, but should also theoretically allow partial clones to download much less information. Previously, though, the "prefetch" command used in diffcore-rename had never been modified and downloaded many blobs that were unnecessary for merge-ort. This commit corrects that. When doing basename comparisons, we want to fetch only the objects that will be used for basename comparisons. If after basename fetching this leaves us with no more relevant sources (or no more destinations), then we won't need to do the full inexact rename detection and can skip downloading additional source and destination files. Even if we have to do that later full inexact rename detection, irrelevant sources are culled after basename matching and before the full inexact rename detection, so we can still avoid downloading the blobs for irrelevant sources. Rename prefetch() to inexact_prefetch(), and introduce a new basename_prefetch() to take advantage of this. If we modify the testcase from commit 557ac0350d ("merge-ort: begin performance work; instrument with trace2_region_* calls", 2021-01-23) to pass --sparse --filter=blob:none to the clone command, and use the new trace2 "fetch_count" output from a few commits ago to track both the number of fetch subcommands invoked and the number of objects fetched across all those fetches, then for the mega-renames testcase we observe the following: BEFORE this commit, rebasing 35 patches: strategy # of fetches total # of objects fetched --------- ------------ -------------------------- recursive 62 11423 ort 30 11391 AFTER this commit, rebasing the same 35 patches: ort 32 63 This means that the new code only needs to download less than 2 blobs per patch being rebased. That is especially interesting given that the repository at the start only had approximately half a dozen TOTAL blobs downloaded to start with (because the default sparse-checkout of just the toplevel directory was in use). So, for this particular linux kernel testcase that involved ~26,000 renames on the upstream side (drivers/ -> pilots/) across which 35 patches were being rebased, this change reduces the number of blobs that need to be downloaded by a factor of ~180. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-06-22 08:04:40 +00:00
struct inexact_prefetch_options {
struct repository *repo;
int skip_unmodified;
};
diffcore-rename: use a different prefetch for basename comparisons merge-ort was designed to minimize the amount of data needed and used, and several changes were made to diffcore-rename to take advantage of extra metadata to enable this data minimization (particularly the relevant_sources variable for skipping "irrelevant" renames). This effort obviously succeeded in drastically reducing computation times, but should also theoretically allow partial clones to download much less information. Previously, though, the "prefetch" command used in diffcore-rename had never been modified and downloaded many blobs that were unnecessary for merge-ort. This commit corrects that. When doing basename comparisons, we want to fetch only the objects that will be used for basename comparisons. If after basename fetching this leaves us with no more relevant sources (or no more destinations), then we won't need to do the full inexact rename detection and can skip downloading additional source and destination files. Even if we have to do that later full inexact rename detection, irrelevant sources are culled after basename matching and before the full inexact rename detection, so we can still avoid downloading the blobs for irrelevant sources. Rename prefetch() to inexact_prefetch(), and introduce a new basename_prefetch() to take advantage of this. If we modify the testcase from commit 557ac0350d ("merge-ort: begin performance work; instrument with trace2_region_* calls", 2021-01-23) to pass --sparse --filter=blob:none to the clone command, and use the new trace2 "fetch_count" output from a few commits ago to track both the number of fetch subcommands invoked and the number of objects fetched across all those fetches, then for the mega-renames testcase we observe the following: BEFORE this commit, rebasing 35 patches: strategy # of fetches total # of objects fetched --------- ------------ -------------------------- recursive 62 11423 ort 30 11391 AFTER this commit, rebasing the same 35 patches: ort 32 63 This means that the new code only needs to download less than 2 blobs per patch being rebased. That is especially interesting given that the repository at the start only had approximately half a dozen TOTAL blobs downloaded to start with (because the default sparse-checkout of just the toplevel directory was in use). So, for this particular linux kernel testcase that involved ~26,000 renames on the upstream side (drivers/ -> pilots/) across which 35 patches were being rebased, this change reduces the number of blobs that need to be downloaded by a factor of ~180. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-06-22 08:04:40 +00:00
static void inexact_prefetch(void *prefetch_options)
{
diffcore-rename: use a different prefetch for basename comparisons merge-ort was designed to minimize the amount of data needed and used, and several changes were made to diffcore-rename to take advantage of extra metadata to enable this data minimization (particularly the relevant_sources variable for skipping "irrelevant" renames). This effort obviously succeeded in drastically reducing computation times, but should also theoretically allow partial clones to download much less information. Previously, though, the "prefetch" command used in diffcore-rename had never been modified and downloaded many blobs that were unnecessary for merge-ort. This commit corrects that. When doing basename comparisons, we want to fetch only the objects that will be used for basename comparisons. If after basename fetching this leaves us with no more relevant sources (or no more destinations), then we won't need to do the full inexact rename detection and can skip downloading additional source and destination files. Even if we have to do that later full inexact rename detection, irrelevant sources are culled after basename matching and before the full inexact rename detection, so we can still avoid downloading the blobs for irrelevant sources. Rename prefetch() to inexact_prefetch(), and introduce a new basename_prefetch() to take advantage of this. If we modify the testcase from commit 557ac0350d ("merge-ort: begin performance work; instrument with trace2_region_* calls", 2021-01-23) to pass --sparse --filter=blob:none to the clone command, and use the new trace2 "fetch_count" output from a few commits ago to track both the number of fetch subcommands invoked and the number of objects fetched across all those fetches, then for the mega-renames testcase we observe the following: BEFORE this commit, rebasing 35 patches: strategy # of fetches total # of objects fetched --------- ------------ -------------------------- recursive 62 11423 ort 30 11391 AFTER this commit, rebasing the same 35 patches: ort 32 63 This means that the new code only needs to download less than 2 blobs per patch being rebased. That is especially interesting given that the repository at the start only had approximately half a dozen TOTAL blobs downloaded to start with (because the default sparse-checkout of just the toplevel directory was in use). So, for this particular linux kernel testcase that involved ~26,000 renames on the upstream side (drivers/ -> pilots/) across which 35 patches were being rebased, this change reduces the number of blobs that need to be downloaded by a factor of ~180. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-06-22 08:04:40 +00:00
struct inexact_prefetch_options *options = prefetch_options;
int i;
struct oid_array to_fetch = OID_ARRAY_INIT;
for (i = 0; i < rename_dst_nr; i++) {
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 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. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-11 09:08:47 +00:00
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,
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 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. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-11 09:08:47 +00:00
rename_dst[i].p->two);
}
for (i = 0; i < rename_src_nr; i++) {
if (options->skip_unmodified &&
diff_unmodified_pair(rename_src[i].p))
/*
* The loop in diffcore_rename() will not need these
* blobs, so skip prefetching.
*/
continue;
diff_add_if_missing(options->repo, &to_fetch,
rename_src[i].p->one);
}
promisor_remote_get_direct(options->repo, to_fetch.oid, to_fetch.nr);
oid_array_clear(&to_fetch);
}
static int estimate_similarity(struct repository *r,
struct diff_filespec *src,
struct diff_filespec *dst,
int minimum_score,
struct diff_populate_filespec_options *dpf_opt)
{
/* 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
* 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.
*
* Compare them and return how similar they are, representing
* the score as an integer between 0 and MAX_SCORE.
*
* When there is an exact match, it is considered a better
* match than anything else; the destination does not even
* call into this function in that case.
*/
unsigned long max_size, delta_size, base_size, src_copied, literal_added;
int score;
/* We deal only with regular files. Symlink renames are handled
* only when they are exact matches --- in other words, no edits
* after renaming.
*/
if (!S_ISREG(src->mode) || !S_ISREG(dst->mode))
return 0;
/*
* Need to check that source and destination sizes are
* filled in before comparing them.
*
* If we already have "cnt_data" filled in, we know it's
* all good (avoid checking the size for zero, as that
* is a possible size - we really should have a flag to
* say whether the size is valid or not!)
*/
dpf_opt->check_size_only = 1;
if (!src->cnt_data &&
diff_populate_filespec(r, src, dpf_opt))
return 0;
if (!dst->cnt_data &&
diff_populate_filespec(r, dst, dpf_opt))
return 0;
max_size = ((src->size > dst->size) ? src->size : dst->size);
base_size = ((src->size < dst->size) ? src->size : dst->size);
delta_size = max_size - base_size;
/* We would not consider edits that change the file size so
* drastically. delta_size must be smaller than
* (MAX_SCORE-minimum_score)/MAX_SCORE * min(src->size, dst->size).
*
* Note that base_size == 0 case is handled here already
* and the final score computation below would not have a
* divide-by-zero issue.
*/
if (max_size * (MAX_SCORE-minimum_score) < delta_size * MAX_SCORE)
return 0;
dpf_opt->check_size_only = 0;
if (!src->cnt_data && diff_populate_filespec(r, src, dpf_opt))
return 0;
if (!dst->cnt_data && diff_populate_filespec(r, dst, dpf_opt))
return 0;
if (diffcore_count_changes(r, src, dst,
&src->cnt_data, &dst->cnt_data,
&src_copied, &literal_added))
return 0;
/* How similar are they?
* what percentage of material in dst are from source?
*/
if (!dst->size)
score = 0; /* should not happen */
else
score = (int)(src_copied * MAX_SCORE / max_size);
return score;
}
static void record_rename_pair(int dst_index, int src_index, int score)
{
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 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. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-11 09:08:47 +00:00
struct diff_filepair *src = rename_src[src_index].p;
struct diff_filepair *dst = rename_dst[dst_index].p;
[PATCH] Rename/copy detection fix. The rename/copy detection logic in earlier round was only good enough to show patch output and discussion on the mailing list about the diff-raw format updates revealed many problems with it. This patch fixes all the ones known to me, without making things I want to do later impossible, mostly related to patch reordering. (1) Earlier rename/copy detector determined which one is rename and which one is copy too early, which made it impossible to later introduce diffcore transformers to reorder patches. This patch fixes it by moving that logic to the very end of the processing. (2) Earlier output routine diff_flush() was pruning all the "no-change" entries indiscriminatingly. This was done due to my false assumption that one of the requirements in the diff-raw output was not to show such an entry (which resulted in my incorrect comment about "diff-helper never being able to be equivalent to built-in diff driver"). My special thanks go to Linus for correcting me about this. When we produce diff-raw output, for the downstream to be able to tell renames from copies, sometimes it _is_ necessary to output "no-change" entries, and this patch adds diffcore_prune() function for doing it. (3) Earlier diff_filepair structure was trying to be not too specific about rename/copy operations, but the purpose of the structure was to record one or two paths, which _was_ indeed about rename/copy. This patch discards xfrm_msg field which was trying to be generic for this wrong reason, and introduces a couple of fields (rename_score and rename_rank) that are explicitly specific to rename/copy logic. One thing to note is that the information in a single diff_filepair structure _still_ does not distinguish renames from copies, and it is deliberately so. This is to allow patches to be reordered in later stages. (4) This patch also adds some tests about diff-raw format output and makes sure that necessary "no-change" entries appear on the output. Signed-off-by: Junio C Hamano <junkio@cox.net> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-05-23 04:26:09 +00:00
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 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. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-11 09:08:47 +00:00
if (dst->renamed_pair)
die("internal error: dst already matched.");
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 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. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-11 09:08:47 +00:00
src->one->rename_used++;
src->one->count++;
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 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. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-11 09:08:47 +00:00
rename_dst[dst_index].filespec_to_free = dst->one;
rename_dst[dst_index].is_rename = 1;
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 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. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-11 09:08:47 +00:00
dst->one = src->one;
dst->renamed_pair = 1;
if (!strcmp(dst->one->path, dst->two->path))
dst->score = rename_src[src_index].score;
else
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 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. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-11 09:08:47 +00:00
dst->score = score;
}
/*
* We sort the rename similarity matrix with the score, in descending
* order (the most similar first).
*/
static int score_compare(const void *a_, const void *b_)
{
const struct diff_score *a = a_, *b = b_;
/* sink the unused ones to the bottom */
if (a->dst < 0)
return (0 <= b->dst);
else if (b->dst < 0)
return -1;
if (a->score == b->score)
return b->name_score - a->name_score;
return b->score - a->score;
}
struct file_similarity {
struct hashmap_entry entry;
int index;
struct diff_filespec *filespec;
};
static unsigned int hash_filespec(struct repository *r,
struct diff_filespec *filespec)
{
if (!filespec->oid_valid) {
if (diff_populate_filespec(r, filespec, NULL))
return 0;
hash_object_file(r->hash_algo, filespec->data, filespec->size,
OBJ_BLOB, &filespec->oid);
}
return oidhash(&filespec->oid);
}
static int find_identical_files(struct hashmap *srcs,
int dst_index,
struct diff_options *options)
{
int renames = 0;
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 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. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-11 09:08:47 +00:00
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);
/*
* Find the best source match for specified destination.
*/
p = hashmap_get_entry_from_hash(srcs, hash, NULL,
struct file_similarity, entry);
hashmap_for_each_entry_from(srcs, p, entry) {
int score;
struct diff_filespec *source = p->filespec;
/* False hash collision? */
if (!oideq(&source->oid, &target->oid))
continue;
/* Non-regular files? If so, the modes must match! */
if (!S_ISREG(source->mode) || !S_ISREG(target->mode)) {
if (source->mode != target->mode)
continue;
}
/* Give higher scores to sources that haven't been used already */
score = !source->rename_used;
if (source->rename_used && options->detect_rename != DIFF_DETECT_COPY)
continue;
score += basename_same(source, target);
if (score > best_score) {
best = p;
best_score = score;
if (score == 2)
break;
}
/* Too many identical alternatives? Pick one */
if (!--i)
break;
}
if (best) {
record_rename_pair(dst_index, best->index, MAX_SCORE);
renames++;
}
return renames;
}
static void insert_file_table(struct repository *r,
struct mem_pool *pool,
struct hashmap *table, int index,
struct diff_filespec *filespec)
{
struct file_similarity *entry = mem_pool_alloc(pool, sizeof(*entry));
entry->index = index;
entry->filespec = filespec;
hashmap_entry_init(&entry->entry, hash_filespec(r, filespec));
hashmap_add(table, &entry->entry);
}
/*
* Find exact renames first.
*
* The first round matches up the up-to-date entries,
* and then during the second round we try to match
* cache-dirty entries as well.
*/
static int find_exact_renames(struct diff_options *options,
struct mem_pool *pool)
{
int i, renames = 0;
struct hashmap file_table;
2016-03-30 08:35:07 +00:00
/* Add all sources to the hash table in reverse order, because
* later on they will be retrieved in LIFO order.
*/
hashmap_init(&file_table, NULL, NULL, rename_src_nr);
2016-03-30 08:35:07 +00:00
for (i = rename_src_nr-1; i >= 0; i--)
insert_file_table(options->repo, pool,
&file_table, i,
rename_src[i].p->one);
/* Walk the destinations and find best source match */
for (i = 0; i < rename_dst_nr; i++)
renames += find_identical_files(&file_table, i, options);
/* Free the hash data structure (entries will be freed with the pool) */
hashmap_clear(&file_table);
return renames;
}
struct dir_rename_info {
struct strintmap idx_map;
struct strmap dir_rename_guess;
struct strmap *dir_rename_count;
struct strintmap *relevant_source_dirs;
unsigned setup;
};
static char *get_dirname(const char *filename)
{
char *slash = strrchr(filename, '/');
return slash ? xstrndup(filename, slash - filename) : xstrdup("");
}
static void dirname_munge(char *filename)
{
char *slash = strrchr(filename, '/');
if (!slash)
slash = filename;
*slash = '\0';
}
static const char *get_highest_rename_path(struct strintmap *counts)
{
int highest_count = 0;
const char *highest_destination_dir = NULL;
struct hashmap_iter iter;
struct strmap_entry *entry;
strintmap_for_each_entry(counts, &iter, entry) {
const char *destination_dir = entry->key;
intptr_t count = (intptr_t)entry->value;
if (count > highest_count) {
highest_count = count;
highest_destination_dir = destination_dir;
}
}
return highest_destination_dir;
}
static char *UNKNOWN_DIR = "/"; /* placeholder -- short, illegal directory */
static int dir_rename_already_determinable(struct strintmap *counts)
{
struct hashmap_iter iter;
struct strmap_entry *entry;
int first = 0, second = 0, unknown = 0;
strintmap_for_each_entry(counts, &iter, entry) {
const char *destination_dir = entry->key;
intptr_t count = (intptr_t)entry->value;
if (!strcmp(destination_dir, UNKNOWN_DIR)) {
unknown = count;
} else if (count >= first) {
second = first;
first = count;
} else if (count >= second) {
second = count;
}
}
return first > second + unknown;
}
static void increment_count(struct dir_rename_info *info,
char *old_dir,
char *new_dir)
{
struct strintmap *counts;
struct strmap_entry *e;
/* Get the {new_dirs -> counts} mapping using old_dir */
e = strmap_get_entry(info->dir_rename_count, old_dir);
if (e) {
counts = e->value;
} else {
counts = xmalloc(sizeof(*counts));
strintmap_init_with_options(counts, 0, NULL, 1);
strmap_put(info->dir_rename_count, old_dir, counts);
}
/* Increment the count for new_dir */
strintmap_incr(counts, new_dir, 1);
}
static void update_dir_rename_counts(struct dir_rename_info *info,
struct strintmap *dirs_removed,
const char *oldname,
const char *newname)
{
diffcore-rename: move old_dir/new_dir definition to plug leak old_dir/new_dir are free()'d at the end of update_dir_rename_counts, however if we return early we'll never free those strings. Therefore we should move all new allocations after the possible early return, avoiding a leak. This seems like a fairly recent leak, that started happening since the early-return was added in: 1ad69eb0dc (diffcore-rename: compute dir_rename_counts in stages, 2021-02-27) LSAN output from t0022: Direct leak of 7 byte(s) in 1 object(s) allocated from: #0 0x486804 in strdup ../projects/compiler-rt/lib/asan/asan_interceptors.cpp:452:3 #1 0xa71e48 in xstrdup wrapper.c:29:14 #2 0x7db9c7 in update_dir_rename_counts diffcore-rename.c:464:12 #3 0x7db6ae in find_renames diffcore-rename.c:1062:3 #4 0x7d76c3 in diffcore_rename_extended diffcore-rename.c:1472:18 #5 0x7b4cfc in diffcore_std diff.c:6705:4 #6 0x855e46 in log_tree_diff_flush log-tree.c:846:2 #7 0x856574 in log_tree_diff log-tree.c:955:3 #8 0x856574 in log_tree_commit log-tree.c:986:10 #9 0x9a9c67 in print_commit_summary sequencer.c:1329:7 #10 0x52e623 in cmd_commit builtin/commit.c:1862:3 #11 0x4ce83e in run_builtin git.c:475:11 #12 0x4ccafe in handle_builtin git.c:729:3 #13 0x4cb01c in run_argv git.c:818:4 #14 0x4cb01c in cmd_main git.c:949:19 #15 0x6b3f3d in main common-main.c:52:11 #16 0x7fe397c7a349 in __libc_start_main (/lib64/libc.so.6+0x24349) Direct leak of 7 byte(s) in 1 object(s) allocated from: #0 0x486804 in strdup ../projects/compiler-rt/lib/asan/asan_interceptors.cpp:452:3 #1 0xa71e48 in xstrdup wrapper.c:29:14 #2 0x7db9bc in update_dir_rename_counts diffcore-rename.c:463:12 #3 0x7db6ae in find_renames diffcore-rename.c:1062:3 #4 0x7d76c3 in diffcore_rename_extended diffcore-rename.c:1472:18 #5 0x7b4cfc in diffcore_std diff.c:6705:4 #6 0x855e46 in log_tree_diff_flush log-tree.c:846:2 #7 0x856574 in log_tree_diff log-tree.c:955:3 #8 0x856574 in log_tree_commit log-tree.c:986:10 #9 0x9a9c67 in print_commit_summary sequencer.c:1329:7 #10 0x52e623 in cmd_commit builtin/commit.c:1862:3 #11 0x4ce83e in run_builtin git.c:475:11 #12 0x4ccafe in handle_builtin git.c:729:3 #13 0x4cb01c in run_argv git.c:818:4 #14 0x4cb01c in cmd_main git.c:949:19 #15 0x6b3f3d in main common-main.c:52:11 #16 0x7fe397c7a349 in __libc_start_main (/lib64/libc.so.6+0x24349) SUMMARY: AddressSanitizer: 14 byte(s) leaked in 2 allocation(s). Signed-off-by: Andrzej Hunt <andrzej@ahunt.org> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-07-25 13:08:23 +00:00
char *old_dir;
char *new_dir;
const char new_dir_first_char = newname[0];
int first_time_in_loop = 1;
if (!info->setup)
/*
* info->setup is 0 here in two cases: (1) all auxiliary
* vars (like dirs_removed) were NULL so
* initialize_dir_rename_info() returned early, or (2)
* either break detection or copy detection are active so
* that we never called initialize_dir_rename_info(). In
* the former case, we don't have enough info to know if
* directories were renamed (because dirs_removed lets us
* know about a necessary prerequisite, namely if they were
* removed), and in the latter, we don't care about
* directory renames or find_basename_matches.
*
* This matters because both basename and inexact matching
* will also call update_dir_rename_counts(). In either of
* the above two cases info->dir_rename_counts will not
* have been properly initialized which prevents us from
* updating it, but in these two cases we don't care about
* dir_rename_counts anyway, so we can just exit early.
*/
return;
diffcore-rename: move old_dir/new_dir definition to plug leak old_dir/new_dir are free()'d at the end of update_dir_rename_counts, however if we return early we'll never free those strings. Therefore we should move all new allocations after the possible early return, avoiding a leak. This seems like a fairly recent leak, that started happening since the early-return was added in: 1ad69eb0dc (diffcore-rename: compute dir_rename_counts in stages, 2021-02-27) LSAN output from t0022: Direct leak of 7 byte(s) in 1 object(s) allocated from: #0 0x486804 in strdup ../projects/compiler-rt/lib/asan/asan_interceptors.cpp:452:3 #1 0xa71e48 in xstrdup wrapper.c:29:14 #2 0x7db9c7 in update_dir_rename_counts diffcore-rename.c:464:12 #3 0x7db6ae in find_renames diffcore-rename.c:1062:3 #4 0x7d76c3 in diffcore_rename_extended diffcore-rename.c:1472:18 #5 0x7b4cfc in diffcore_std diff.c:6705:4 #6 0x855e46 in log_tree_diff_flush log-tree.c:846:2 #7 0x856574 in log_tree_diff log-tree.c:955:3 #8 0x856574 in log_tree_commit log-tree.c:986:10 #9 0x9a9c67 in print_commit_summary sequencer.c:1329:7 #10 0x52e623 in cmd_commit builtin/commit.c:1862:3 #11 0x4ce83e in run_builtin git.c:475:11 #12 0x4ccafe in handle_builtin git.c:729:3 #13 0x4cb01c in run_argv git.c:818:4 #14 0x4cb01c in cmd_main git.c:949:19 #15 0x6b3f3d in main common-main.c:52:11 #16 0x7fe397c7a349 in __libc_start_main (/lib64/libc.so.6+0x24349) Direct leak of 7 byte(s) in 1 object(s) allocated from: #0 0x486804 in strdup ../projects/compiler-rt/lib/asan/asan_interceptors.cpp:452:3 #1 0xa71e48 in xstrdup wrapper.c:29:14 #2 0x7db9bc in update_dir_rename_counts diffcore-rename.c:463:12 #3 0x7db6ae in find_renames diffcore-rename.c:1062:3 #4 0x7d76c3 in diffcore_rename_extended diffcore-rename.c:1472:18 #5 0x7b4cfc in diffcore_std diff.c:6705:4 #6 0x855e46 in log_tree_diff_flush log-tree.c:846:2 #7 0x856574 in log_tree_diff log-tree.c:955:3 #8 0x856574 in log_tree_commit log-tree.c:986:10 #9 0x9a9c67 in print_commit_summary sequencer.c:1329:7 #10 0x52e623 in cmd_commit builtin/commit.c:1862:3 #11 0x4ce83e in run_builtin git.c:475:11 #12 0x4ccafe in handle_builtin git.c:729:3 #13 0x4cb01c in run_argv git.c:818:4 #14 0x4cb01c in cmd_main git.c:949:19 #15 0x6b3f3d in main common-main.c:52:11 #16 0x7fe397c7a349 in __libc_start_main (/lib64/libc.so.6+0x24349) SUMMARY: AddressSanitizer: 14 byte(s) leaked in 2 allocation(s). Signed-off-by: Andrzej Hunt <andrzej@ahunt.org> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-07-25 13:08:23 +00:00
old_dir = xstrdup(oldname);
new_dir = xstrdup(newname);
while (1) {
diffcore-rename: only compute dir_rename_count for relevant directories When one side adds files to a directory that the other side renamed, directory rename detection is used to either move the new paths to the newer directory or warn the user about the fact that another path location might be better. If a parent of the given directory had new files added to it, any renames in the current directory are also part of determining where the parent directory is renamed to. Thus, naively, we need to record each rename N times for a path at depth N. However, we can use the additional information added to dirs_removed in the last commit to avoid traversing all N parent directories in many cases. Let's use an example to explain how this works. If we have a path named src/old_dir/a/b/file.c and src/old_dir doesn't exist on one side of history, but the other added a file named src/old_dir/newfile.c, then if one side renamed src/old_dir/a/b/file.c => source/new_dir/a/b/file.c then this file would affect potential directory rename detection counts for src/old_dir/a/b => source/new_dir/a/b src/old_dir/a => source/new_dir/a src/old_dir => source/new_dir src => source adding a weight of 1 to each in dir_rename_counts. However, if src/ exists on both sides of history, then we don't need to track any entries for it in dir_rename_counts. That was implemented previously. What we are adding now, is that if no new files were added to src/old_dir/a or src/old_dir/b, then we don't need to have counts in dir_rename_count for those directories either. In short, we only need to track counts in dir_rename_count for directories whose dirs_removed value is RELEVANT_FOR_SELF. And as soon as we reach a directory that isn't in dirs_removed (signalled by returning the default value of NOT_RELEVANT from strintmap_get()), we can stop looking any further up the directory hierarchy. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-03-13 22:22:04 +00:00
int drd_flag = NOT_RELEVANT;
/* Get old_dir, skip if its directory isn't relevant. */
dirname_munge(old_dir);
if (info->relevant_source_dirs &&
!strintmap_contains(info->relevant_source_dirs, old_dir))
break;
/* Get new_dir */
dirname_munge(new_dir);
/*
* When renaming
* "a/b/c/d/e/foo.c" -> "a/b/some/thing/else/e/foo.c"
* then this suggests that both
* a/b/c/d/e/ => a/b/some/thing/else/e/
* a/b/c/d/ => a/b/some/thing/else/
* so we want to increment counters for both. We do NOT,
* however, also want to suggest that there was the following
* rename:
* a/b/c/ => a/b/some/thing/
* so we need to quit at that point.
*
* Note the when first_time_in_loop, we only strip off the
* basename, and we don't care if that's different.
*/
if (!first_time_in_loop) {
char *old_sub_dir = strchr(old_dir, '\0')+1;
char *new_sub_dir = strchr(new_dir, '\0')+1;
if (!*new_dir) {
/*
* Special case when renaming to root directory,
* i.e. when new_dir == "". In this case, we had
* something like
* a/b/subdir => subdir
* and so dirname_munge() sets things up so that
* old_dir = "a/b\0subdir\0"
* new_dir = "\0ubdir\0"
* We didn't have a '/' to overwrite a '\0' onto
* in new_dir, so we have to compare differently.
*/
if (new_dir_first_char != old_sub_dir[0] ||
strcmp(old_sub_dir+1, new_sub_dir))
break;
} else {
if (strcmp(old_sub_dir, new_sub_dir))
break;
}
}
diffcore-rename: only compute dir_rename_count for relevant directories When one side adds files to a directory that the other side renamed, directory rename detection is used to either move the new paths to the newer directory or warn the user about the fact that another path location might be better. If a parent of the given directory had new files added to it, any renames in the current directory are also part of determining where the parent directory is renamed to. Thus, naively, we need to record each rename N times for a path at depth N. However, we can use the additional information added to dirs_removed in the last commit to avoid traversing all N parent directories in many cases. Let's use an example to explain how this works. If we have a path named src/old_dir/a/b/file.c and src/old_dir doesn't exist on one side of history, but the other added a file named src/old_dir/newfile.c, then if one side renamed src/old_dir/a/b/file.c => source/new_dir/a/b/file.c then this file would affect potential directory rename detection counts for src/old_dir/a/b => source/new_dir/a/b src/old_dir/a => source/new_dir/a src/old_dir => source/new_dir src => source adding a weight of 1 to each in dir_rename_counts. However, if src/ exists on both sides of history, then we don't need to track any entries for it in dir_rename_counts. That was implemented previously. What we are adding now, is that if no new files were added to src/old_dir/a or src/old_dir/b, then we don't need to have counts in dir_rename_count for those directories either. In short, we only need to track counts in dir_rename_count for directories whose dirs_removed value is RELEVANT_FOR_SELF. And as soon as we reach a directory that isn't in dirs_removed (signalled by returning the default value of NOT_RELEVANT from strintmap_get()), we can stop looking any further up the directory hierarchy. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-03-13 22:22:04 +00:00
/*
* Above we suggested that we'd keep recording renames for
* all ancestor directories where the trailing directories
* matched, i.e. for
* "a/b/c/d/e/foo.c" -> "a/b/some/thing/else/e/foo.c"
* we'd increment rename counts for each of
* a/b/c/d/e/ => a/b/some/thing/else/e/
* a/b/c/d/ => a/b/some/thing/else/
* However, we only need the rename counts for directories
* in dirs_removed whose value is RELEVANT_FOR_SELF.
* However, we add one special case of also recording it for
* first_time_in_loop because find_basename_matches() can
* use that as a hint to find a good pairing.
*/
if (dirs_removed)
drd_flag = strintmap_get(dirs_removed, old_dir);
if (drd_flag == RELEVANT_FOR_SELF || first_time_in_loop)
increment_count(info, old_dir, new_dir);
diffcore-rename: only compute dir_rename_count for relevant directories When one side adds files to a directory that the other side renamed, directory rename detection is used to either move the new paths to the newer directory or warn the user about the fact that another path location might be better. If a parent of the given directory had new files added to it, any renames in the current directory are also part of determining where the parent directory is renamed to. Thus, naively, we need to record each rename N times for a path at depth N. However, we can use the additional information added to dirs_removed in the last commit to avoid traversing all N parent directories in many cases. Let's use an example to explain how this works. If we have a path named src/old_dir/a/b/file.c and src/old_dir doesn't exist on one side of history, but the other added a file named src/old_dir/newfile.c, then if one side renamed src/old_dir/a/b/file.c => source/new_dir/a/b/file.c then this file would affect potential directory rename detection counts for src/old_dir/a/b => source/new_dir/a/b src/old_dir/a => source/new_dir/a src/old_dir => source/new_dir src => source adding a weight of 1 to each in dir_rename_counts. However, if src/ exists on both sides of history, then we don't need to track any entries for it in dir_rename_counts. That was implemented previously. What we are adding now, is that if no new files were added to src/old_dir/a or src/old_dir/b, then we don't need to have counts in dir_rename_count for those directories either. In short, we only need to track counts in dir_rename_count for directories whose dirs_removed value is RELEVANT_FOR_SELF. And as soon as we reach a directory that isn't in dirs_removed (signalled by returning the default value of NOT_RELEVANT from strintmap_get()), we can stop looking any further up the directory hierarchy. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-03-13 22:22:04 +00:00
first_time_in_loop = 0;
if (drd_flag == NOT_RELEVANT)
break;
/* If we hit toplevel directory ("") for old or new dir, quit */
if (!*old_dir || !*new_dir)
break;
}
/* Free resources we don't need anymore */
free(old_dir);
free(new_dir);
}
static void initialize_dir_rename_info(struct dir_rename_info *info,
struct strintmap *relevant_sources,
struct strintmap *dirs_removed,
merge-ort, diffcore-rename: employ cached renames when possible When there are many renames between the old base of a series of commits and the new base, the way sequencer.c, merge-recursive.c, and diffcore-rename.c have traditionally split the work resulted in redetecting the same renames with each and every commit being transplanted. To address this, the last several commits have been creating a cache of rename detection results, determining when it was safe to use such a cache in subsequent merge operations, adding helper functions, and so on. See the previous half dozen commit messages for additional discussion of this optimization, particularly the message a few commits ago entitled "add code to check for whether cached renames can be reused". This commit finally ties all of that work together, modifying the merge algorithm to make use of these cached renames. For the testcases mentioned in commit 557ac0350d ("merge-ort: begin performance work; instrument with trace2_region_* calls", 2020-10-28), this change improves the performance as follows: Before After no-renames: 5.665 s ± 0.129 s 5.622 s ± 0.059 s mega-renames: 11.435 s ± 0.158 s 10.127 s ± 0.073 s just-one-mega: 494.2 ms ± 6.1 ms 500.3 ms ± 3.8 ms That's a fairly small improvement, but mostly because the previous optimizations were so effective for these particular testcases; this optimization only kicks in when the others don't. If we undid the basename-guided rename detection and skip-irrelevant-renames optimizations, then we'd see that this series by itself improved performance as follows: Before Basename Series After Just This Series no-renames: 13.815 s ± 0.062 s 5.697 s ± 0.080 s mega-renames: 1799.937 s ± 0.493 s 205.709 s ± 0.457 s Since this optimization kicks in to help accelerate cases where the previous optimizations do not apply, this last comparison shows that this cached-renames optimization has the potential to help signficantly in cases that don't meet the requirements for the other optimizations to be effective. The changes made in this optimization also lay some important groundwork for a future optimization around having collect_merge_info() avoid recursing into subtrees in more cases. However, for this optimization to be effective, merge_switch_to_result() should only be called when the rebase or cherry-pick operation has either completed or hit a case where the user needs to resolve a conflict or edit the result. If it is called after every commit, as sequencer.c does, then the working tree and index are needlessly updated with every commit and the cached metadata is tossed, defeating this optimization. Refactoring sequencer.c to only call merge_switch_to_result() at the end of the operation is a bigger undertaking, and the practical benefits of this optimization will not be realized until that work is performed. Since `test-tool fast-rebase` only updates at the end of the operation, it was used to obtain the timings above. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-05-20 06:09:41 +00:00
struct strmap *dir_rename_count,
struct strmap *cached_pairs)
{
struct hashmap_iter iter;
struct strmap_entry *entry;
int i;
if (!dirs_removed && !relevant_sources) {
info->setup = 0;
return;
}
info->setup = 1;
info->dir_rename_count = dir_rename_count;
if (!info->dir_rename_count) {
info->dir_rename_count = xmalloc(sizeof(*dir_rename_count));
strmap_init(info->dir_rename_count);
}
strintmap_init_with_options(&info->idx_map, -1, NULL, 0);
strmap_init_with_options(&info->dir_rename_guess, NULL, 0);
/* Setup info->relevant_source_dirs */
info->relevant_source_dirs = NULL;
if (dirs_removed || !relevant_sources) {
info->relevant_source_dirs = dirs_removed; /* might be NULL */
} else {
info->relevant_source_dirs = xmalloc(sizeof(struct strintmap));
strintmap_init(info->relevant_source_dirs, 0 /* unused */);
strintmap_for_each_entry(relevant_sources, &iter, entry) {
char *dirname = get_dirname(entry->key);
if (!dirs_removed ||
strintmap_contains(dirs_removed, dirname))
strintmap_set(info->relevant_source_dirs,
dirname, 0 /* value irrelevant */);
free(dirname);
}
}
/*
* Loop setting up both info->idx_map, and doing setup of
* info->dir_rename_count.
*/
for (i = 0; i < rename_dst_nr; ++i) {
/*
* For non-renamed files, make idx_map contain mapping of
* filename -> index (index within rename_dst, that is)
*/
if (!rename_dst[i].is_rename) {
char *filename = rename_dst[i].p->two->path;
strintmap_set(&info->idx_map, filename, i);
continue;
}
/*
* For everything else (i.e. renamed files), make
* dir_rename_count contain a map of a map:
* old_directory -> {new_directory -> count}
* In other words, for every pair look at the directories for
* the old filename and the new filename and count how many
* times that pairing occurs.
*/
update_dir_rename_counts(info, dirs_removed,
rename_dst[i].p->one->path,
rename_dst[i].p->two->path);
}
merge-ort, diffcore-rename: employ cached renames when possible When there are many renames between the old base of a series of commits and the new base, the way sequencer.c, merge-recursive.c, and diffcore-rename.c have traditionally split the work resulted in redetecting the same renames with each and every commit being transplanted. To address this, the last several commits have been creating a cache of rename detection results, determining when it was safe to use such a cache in subsequent merge operations, adding helper functions, and so on. See the previous half dozen commit messages for additional discussion of this optimization, particularly the message a few commits ago entitled "add code to check for whether cached renames can be reused". This commit finally ties all of that work together, modifying the merge algorithm to make use of these cached renames. For the testcases mentioned in commit 557ac0350d ("merge-ort: begin performance work; instrument with trace2_region_* calls", 2020-10-28), this change improves the performance as follows: Before After no-renames: 5.665 s ± 0.129 s 5.622 s ± 0.059 s mega-renames: 11.435 s ± 0.158 s 10.127 s ± 0.073 s just-one-mega: 494.2 ms ± 6.1 ms 500.3 ms ± 3.8 ms That's a fairly small improvement, but mostly because the previous optimizations were so effective for these particular testcases; this optimization only kicks in when the others don't. If we undid the basename-guided rename detection and skip-irrelevant-renames optimizations, then we'd see that this series by itself improved performance as follows: Before Basename Series After Just This Series no-renames: 13.815 s ± 0.062 s 5.697 s ± 0.080 s mega-renames: 1799.937 s ± 0.493 s 205.709 s ± 0.457 s Since this optimization kicks in to help accelerate cases where the previous optimizations do not apply, this last comparison shows that this cached-renames optimization has the potential to help signficantly in cases that don't meet the requirements for the other optimizations to be effective. The changes made in this optimization also lay some important groundwork for a future optimization around having collect_merge_info() avoid recursing into subtrees in more cases. However, for this optimization to be effective, merge_switch_to_result() should only be called when the rebase or cherry-pick operation has either completed or hit a case where the user needs to resolve a conflict or edit the result. If it is called after every commit, as sequencer.c does, then the working tree and index are needlessly updated with every commit and the cached metadata is tossed, defeating this optimization. Refactoring sequencer.c to only call merge_switch_to_result() at the end of the operation is a bigger undertaking, and the practical benefits of this optimization will not be realized until that work is performed. Since `test-tool fast-rebase` only updates at the end of the operation, it was used to obtain the timings above. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-05-20 06:09:41 +00:00
/* Add cached_pairs to counts */
strmap_for_each_entry(cached_pairs, &iter, entry) {
const char *old_name = entry->key;
const char *new_name = entry->value;
if (!new_name)
/* known delete; ignore it */
continue;
update_dir_rename_counts(info, dirs_removed, old_name, new_name);
}
/*
* Now we collapse
* dir_rename_count: old_directory -> {new_directory -> count}
* down to
* dir_rename_guess: old_directory -> best_new_directory
* where best_new_directory is the one with the highest count.
*/
strmap_for_each_entry(info->dir_rename_count, &iter, entry) {
/* entry->key is source_dir */
struct strintmap *counts = entry->value;
char *best_newdir;
best_newdir = xstrdup(get_highest_rename_path(counts));
strmap_put(&info->dir_rename_guess, entry->key,
best_newdir);
}
}
void partial_clear_dir_rename_count(struct strmap *dir_rename_count)
{
struct hashmap_iter iter;
struct strmap_entry *entry;
strmap_for_each_entry(dir_rename_count, &iter, entry) {
struct strintmap *counts = entry->value;
strintmap_clear(counts);
}
strmap_partial_clear(dir_rename_count, 1);
}
static void cleanup_dir_rename_info(struct dir_rename_info *info,
struct strintmap *dirs_removed,
int keep_dir_rename_count)
{
struct hashmap_iter iter;
struct strmap_entry *entry;
struct string_list to_remove = STRING_LIST_INIT_NODUP;
int i;
if (!info->setup)
return;
/* idx_map */
strintmap_clear(&info->idx_map);
/* dir_rename_guess */
strmap_clear(&info->dir_rename_guess, 1);
/* relevant_source_dirs */
if (info->relevant_source_dirs &&
info->relevant_source_dirs != dirs_removed) {
strintmap_clear(info->relevant_source_dirs);
FREE_AND_NULL(info->relevant_source_dirs);
}
/* dir_rename_count */
if (!keep_dir_rename_count) {
partial_clear_dir_rename_count(info->dir_rename_count);
strmap_clear(info->dir_rename_count, 1);
FREE_AND_NULL(info->dir_rename_count);
return;
}
/*
* Although dir_rename_count was passed in
* diffcore_rename_extended() and we want to keep it around and
* return it to that caller, we first want to remove any counts in
* the maps associated with UNKNOWN_DIR entries and any data
* associated with directories that weren't renamed.
*/
strmap_for_each_entry(info->dir_rename_count, &iter, entry) {
const char *source_dir = entry->key;
struct strintmap *counts = entry->value;
if (!strintmap_get(dirs_removed, source_dir)) {
string_list_append(&to_remove, source_dir);
strintmap_clear(counts);
continue;
}
if (strintmap_contains(counts, UNKNOWN_DIR))
strintmap_remove(counts, UNKNOWN_DIR);
}
for (i = 0; i < to_remove.nr; ++i)
strmap_remove(info->dir_rename_count,
to_remove.items[i].string, 1);
string_list_clear(&to_remove, 0);
}
static const char *get_basename(const char *filename)
{
/*
* gitbasename() has to worry about special drives, multiple
* directory separator characters, trailing slashes, NULL or
* empty strings, etc. We only work on filenames as stored in
* git, and thus get to ignore all those complications.
*/
const char *base = strrchr(filename, '/');
return base ? base + 1 : filename;
}
static int idx_possible_rename(char *filename, struct dir_rename_info *info)
diffcore-rename: use directory rename guided basename comparisons A previous commit noted that it is very common for people to move files across directories while keeping their filename the same. The last few commits took advantage of this and showed that we can accelerate rename detection significantly using basenames; since files with the same basename serve as likely rename candidates, we can check those first and remove them from the rename candidate pool if they are sufficiently similar. Unfortunately, the previous optimization was limited by the fact that the remaining basenames after exact rename detection are not always unique. Many repositories have hundreds of build files with the same name (e.g. Makefile, .gitignore, build.gradle, etc.), and may even have hundreds of source files with the same name. (For example, the linux kernel has 100 setup.c, 87 irq.c, and 112 core.c files. A repository at $DAYJOB has a lot of ObjectFactory.java and Plugin.java files). For these files with non-unique basenames, we are faced with the task of attempting to determine or guess which directory they may have been relocated to. Such a task is precisely the job of directory rename detection. However, there are two catches: (1) the directory rename detection code has traditionally been part of the merge machinery rather than diffcore-rename.c, and (2) directory rename detection currently runs after regular rename detection is complete. The 1st catch is just an implementation issue that can be overcome by some code shuffling. The 2nd requires us to add a further approximation: we only have access to exact renames at this point, so we need to do directory rename detection based on just exact renames. In some cases we won't have exact renames, in which case this extra optimization won't apply. We also choose to not apply the optimization unless we know that the underlying directory was removed, which will require extra data to be passed in to diffcore_rename_extended(). Also, even if we get a prediction about which directory a file may have relocated to, we will still need to check to see if there is a file in the predicted directory, and then compare the two files to see if they meet the higher min_basename_score threshold required for marking the two files as renames. This commit introduces an idx_possible_rename() function which will do this directory rename detection for us and give us the index within rename_dst of the resulting filename. For now, this function is hardcoded to return -1 (not found) and just hooks up how its results would be used once we have a more complete implementation in place. Reviewed-by: Derrick Stolee <dstolee@microsoft.com> Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-02-27 00:30:39 +00:00
{
/*
* Our comparison of files with the same basename (see
* find_basename_matches() below), is only helpful when after exact
* rename detection we have exactly one file with a given basename
* among the rename sources and also only exactly one file with
* that basename among the rename destinations. When we have
* multiple files with the same basename in either set, we do not
* know which to compare against. However, there are some
* filenames that occur in large numbers (particularly
* build-related filenames such as 'Makefile', '.gitignore', or
* 'build.gradle' that potentially exist within every single
* subdirectory), and for performance we want to be able to quickly
* find renames for these files too.
*
* The reason basename comparisons are a useful heuristic was that it
* is common for people to move files across directories while keeping
* their filename the same. If we had a way of determining or even
* making a good educated guess about which directory these non-unique
* basename files had moved the file to, we could check it.
* Luckily...
*
* When an entire directory is in fact renamed, we have two factors
* helping us out:
* (a) the original directory disappeared giving us a hint
* about when we can apply an extra heuristic.
* (a) we often have several files within that directory and
* subdirectories that are renamed without changes
* So, rules for a heuristic:
* (0) If there basename matches are non-unique (the condition under
* which this function is called) AND
* (1) the directory in which the file was found has disappeared
* (i.e. dirs_removed is non-NULL and has a relevant entry) THEN
* (2) use exact renames of files within the directory to determine
* where the directory is likely to have been renamed to. IF
* there is at least one exact rename from within that
* directory, we can proceed.
* (3) If there are multiple places the directory could have been
* renamed to based on exact renames, ignore all but one of them.
* Just use the destination with the most renames going to it.
* (4) Check if applying that directory rename to the original file
* would result in a destination filename that is in the
* potential rename set. If so, return the index of the
* destination file (the index within rename_dst).
* (5) Compare the original file and returned destination for
* similarity, and if they are sufficiently similar, record the
* rename.
*
* This function, idx_possible_rename(), is only responsible for (4).
* The conditions/steps in (1)-(3) are handled via setting up
* dir_rename_count and dir_rename_guess in
* initialize_dir_rename_info(). Steps (0) and (5) are handled by
* the caller of this function.
*/
char *old_dir, *new_dir;
struct strbuf new_path = STRBUF_INIT;
int idx;
if (!info->setup)
return -1;
old_dir = get_dirname(filename);
new_dir = strmap_get(&info->dir_rename_guess, old_dir);
free(old_dir);
if (!new_dir)
return -1;
strbuf_addstr(&new_path, new_dir);
strbuf_addch(&new_path, '/');
strbuf_addstr(&new_path, get_basename(filename));
idx = strintmap_get(&info->idx_map, new_path.buf);
strbuf_release(&new_path);
return idx;
diffcore-rename: use directory rename guided basename comparisons A previous commit noted that it is very common for people to move files across directories while keeping their filename the same. The last few commits took advantage of this and showed that we can accelerate rename detection significantly using basenames; since files with the same basename serve as likely rename candidates, we can check those first and remove them from the rename candidate pool if they are sufficiently similar. Unfortunately, the previous optimization was limited by the fact that the remaining basenames after exact rename detection are not always unique. Many repositories have hundreds of build files with the same name (e.g. Makefile, .gitignore, build.gradle, etc.), and may even have hundreds of source files with the same name. (For example, the linux kernel has 100 setup.c, 87 irq.c, and 112 core.c files. A repository at $DAYJOB has a lot of ObjectFactory.java and Plugin.java files). For these files with non-unique basenames, we are faced with the task of attempting to determine or guess which directory they may have been relocated to. Such a task is precisely the job of directory rename detection. However, there are two catches: (1) the directory rename detection code has traditionally been part of the merge machinery rather than diffcore-rename.c, and (2) directory rename detection currently runs after regular rename detection is complete. The 1st catch is just an implementation issue that can be overcome by some code shuffling. The 2nd requires us to add a further approximation: we only have access to exact renames at this point, so we need to do directory rename detection based on just exact renames. In some cases we won't have exact renames, in which case this extra optimization won't apply. We also choose to not apply the optimization unless we know that the underlying directory was removed, which will require extra data to be passed in to diffcore_rename_extended(). Also, even if we get a prediction about which directory a file may have relocated to, we will still need to check to see if there is a file in the predicted directory, and then compare the two files to see if they meet the higher min_basename_score threshold required for marking the two files as renames. This commit introduces an idx_possible_rename() function which will do this directory rename detection for us and give us the index within rename_dst of the resulting filename. For now, this function is hardcoded to return -1 (not found) and just hooks up how its results would be used once we have a more complete implementation in place. Reviewed-by: Derrick Stolee <dstolee@microsoft.com> Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-02-27 00:30:39 +00:00
}
diffcore-rename: use a different prefetch for basename comparisons merge-ort was designed to minimize the amount of data needed and used, and several changes were made to diffcore-rename to take advantage of extra metadata to enable this data minimization (particularly the relevant_sources variable for skipping "irrelevant" renames). This effort obviously succeeded in drastically reducing computation times, but should also theoretically allow partial clones to download much less information. Previously, though, the "prefetch" command used in diffcore-rename had never been modified and downloaded many blobs that were unnecessary for merge-ort. This commit corrects that. When doing basename comparisons, we want to fetch only the objects that will be used for basename comparisons. If after basename fetching this leaves us with no more relevant sources (or no more destinations), then we won't need to do the full inexact rename detection and can skip downloading additional source and destination files. Even if we have to do that later full inexact rename detection, irrelevant sources are culled after basename matching and before the full inexact rename detection, so we can still avoid downloading the blobs for irrelevant sources. Rename prefetch() to inexact_prefetch(), and introduce a new basename_prefetch() to take advantage of this. If we modify the testcase from commit 557ac0350d ("merge-ort: begin performance work; instrument with trace2_region_* calls", 2021-01-23) to pass --sparse --filter=blob:none to the clone command, and use the new trace2 "fetch_count" output from a few commits ago to track both the number of fetch subcommands invoked and the number of objects fetched across all those fetches, then for the mega-renames testcase we observe the following: BEFORE this commit, rebasing 35 patches: strategy # of fetches total # of objects fetched --------- ------------ -------------------------- recursive 62 11423 ort 30 11391 AFTER this commit, rebasing the same 35 patches: ort 32 63 This means that the new code only needs to download less than 2 blobs per patch being rebased. That is especially interesting given that the repository at the start only had approximately half a dozen TOTAL blobs downloaded to start with (because the default sparse-checkout of just the toplevel directory was in use). So, for this particular linux kernel testcase that involved ~26,000 renames on the upstream side (drivers/ -> pilots/) across which 35 patches were being rebased, this change reduces the number of blobs that need to be downloaded by a factor of ~180. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-06-22 08:04:40 +00:00
struct basename_prefetch_options {
struct repository *repo;
struct strintmap *relevant_sources;
struct strintmap *sources;
struct strintmap *dests;
struct dir_rename_info *info;
};
static void basename_prefetch(void *prefetch_options)
{
struct basename_prefetch_options *options = prefetch_options;
struct strintmap *relevant_sources = options->relevant_sources;
struct strintmap *sources = options->sources;
struct strintmap *dests = options->dests;
struct dir_rename_info *info = options->info;
int i;
struct oid_array to_fetch = OID_ARRAY_INIT;
/*
* TODO: The following loops mirror the code/logic from
* find_basename_matches(), though not quite exactly. Maybe
* abstract the iteration logic out somehow?
*/
for (i = 0; i < rename_src_nr; ++i) {
char *filename = rename_src[i].p->one->path;
const char *base = NULL;
intptr_t src_index;
intptr_t dst_index;
/* Skip irrelevant sources */
if (relevant_sources &&
!strintmap_contains(relevant_sources, filename))
continue;
/*
* If the basename is unique among remaining sources, then
* src_index will equal 'i' and we can attempt to match it
* to a unique basename in the destinations. Otherwise,
* use directory rename heuristics, if possible.
*/
base = get_basename(filename);
src_index = strintmap_get(sources, base);
assert(src_index == -1 || src_index == i);
if (strintmap_contains(dests, base)) {
struct diff_filespec *one, *two;
/* Find a matching destination, if possible */
dst_index = strintmap_get(dests, base);
if (src_index == -1 || dst_index == -1) {
src_index = i;
dst_index = idx_possible_rename(filename, info);
}
if (dst_index == -1)
continue;
/* Ignore this dest if already used in a rename */
if (rename_dst[dst_index].is_rename)
continue; /* already used previously */
one = rename_src[src_index].p->one;
two = rename_dst[dst_index].p->two;
/* Add the pairs */
diff_add_if_missing(options->repo, &to_fetch, two);
diff_add_if_missing(options->repo, &to_fetch, one);
}
}
promisor_remote_get_direct(options->repo, to_fetch.oid, to_fetch.nr);
oid_array_clear(&to_fetch);
}
static int find_basename_matches(struct diff_options *options,
int minimum_score,
struct dir_rename_info *info,
struct strintmap *relevant_sources,
struct strintmap *dirs_removed)
{
diffcore-rename: complete find_basename_matches() It is not uncommon in real world repositories for the majority of file renames to not change the basename of the file; i.e. most "renames" are just a move of files into different directories. We can make use of this to avoid comparing all rename source candidates with all rename destination candidates, by first comparing sources to destinations with the same basenames. If two files with the same basename are sufficiently similar, we record the rename; if not, we include those files in the more exhaustive matrix comparison. This means we are adding a set of preliminary additional comparisons, but for each file we only compare it with at most one other file. For example, if there was a include/media/device.h that was deleted and a src/module/media/device.h that was added, and there are no other device.h files in the remaining sets of added and deleted files after exact rename detection, then these two files would be compared in the preliminary step. This commit does not yet actually employ this new optimization, it merely adds a function which can be used for this purpose. The next commit will do the necessary plumbing to make use of it. Note that this optimization might give us different results than without the optimization, because it's possible that despite files with the same basename being sufficiently similar to be considered a rename, there's an even better match between files without the same basename. I think that is okay for four reasons: (1) it's easy to explain to the users what happened if it does ever occur (or even for them to intuitively figure out), (2) as the next patch will show it provides such a large performance boost that it's worth the tradeoff, and (3) it's somewhat unlikely that despite having unique matching basenames that other files serve as better matches. Reason (4) takes a full paragraph to explain... If the previous three reasons aren't enough, consider what rename detection already does. Break detection is not the default, meaning that if files have the same _fullname_, then they are considered related even if they are 0% similar. In fact, in such a case, we don't even bother comparing the files to see if they are similar let alone comparing them to all other files to see what they are most similar to. Basically, we override content similarity based on sufficient filename similarity. Without the filename similarity (currently implemented as an exact match of filename), we swing the pendulum the opposite direction and say that filename similarity is irrelevant and compare a full N x M matrix of sources and destinations to find out which have the most similar contents. This optimization just adds another form of filename similarity comparison, but augments it with a file content similarity check as well. Basically, if two files have the same basename and are sufficiently similar to be considered a rename, mark them as such without comparing the two to all other rename candidates. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-02-14 07:51:48 +00:00
/*
* When I checked in early 2020, over 76% of file renames in linux
* just moved files to a different directory but kept the same
* basename. gcc did that with over 64% of renames, gecko did it
* with over 79%, and WebKit did it with over 89%.
*
* Therefore we can bypass the normal exhaustive NxM matrix
* comparison of similarities between all potential rename sources
* and destinations by instead using file basename as a hint (i.e.
* the portion of the filename after the last '/'), checking for
* similarity between files with the same basename, and if we find
* a pair that are sufficiently similar, record the rename pair and
* exclude those two from the NxM matrix.
*
* This *might* cause us to find a less than optimal pairing (if
* there is another file that we are even more similar to but has a
* different basename). Given the huge performance advantage
* basename matching provides, and given the frequency with which
* people use the same basename in real world projects, that's a
* trade-off we are willing to accept when doing just rename
* detection.
*
* If someone wants copy detection that implies they are willing to
* spend more cycles to find similarities between files, so it may
* be less likely that this heuristic is wanted. If someone is
* doing break detection, that means they do not want filename
* similarity to imply any form of content similiarity, and thus
* this heuristic would definitely be incompatible.
*/
int i, renames = 0;
struct strintmap sources;
struct strintmap dests;
struct diff_populate_filespec_options dpf_options = {
.check_binary = 0,
.missing_object_cb = NULL,
.missing_object_data = NULL
};
diffcore-rename: use a different prefetch for basename comparisons merge-ort was designed to minimize the amount of data needed and used, and several changes were made to diffcore-rename to take advantage of extra metadata to enable this data minimization (particularly the relevant_sources variable for skipping "irrelevant" renames). This effort obviously succeeded in drastically reducing computation times, but should also theoretically allow partial clones to download much less information. Previously, though, the "prefetch" command used in diffcore-rename had never been modified and downloaded many blobs that were unnecessary for merge-ort. This commit corrects that. When doing basename comparisons, we want to fetch only the objects that will be used for basename comparisons. If after basename fetching this leaves us with no more relevant sources (or no more destinations), then we won't need to do the full inexact rename detection and can skip downloading additional source and destination files. Even if we have to do that later full inexact rename detection, irrelevant sources are culled after basename matching and before the full inexact rename detection, so we can still avoid downloading the blobs for irrelevant sources. Rename prefetch() to inexact_prefetch(), and introduce a new basename_prefetch() to take advantage of this. If we modify the testcase from commit 557ac0350d ("merge-ort: begin performance work; instrument with trace2_region_* calls", 2021-01-23) to pass --sparse --filter=blob:none to the clone command, and use the new trace2 "fetch_count" output from a few commits ago to track both the number of fetch subcommands invoked and the number of objects fetched across all those fetches, then for the mega-renames testcase we observe the following: BEFORE this commit, rebasing 35 patches: strategy # of fetches total # of objects fetched --------- ------------ -------------------------- recursive 62 11423 ort 30 11391 AFTER this commit, rebasing the same 35 patches: ort 32 63 This means that the new code only needs to download less than 2 blobs per patch being rebased. That is especially interesting given that the repository at the start only had approximately half a dozen TOTAL blobs downloaded to start with (because the default sparse-checkout of just the toplevel directory was in use). So, for this particular linux kernel testcase that involved ~26,000 renames on the upstream side (drivers/ -> pilots/) across which 35 patches were being rebased, this change reduces the number of blobs that need to be downloaded by a factor of ~180. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-06-22 08:04:40 +00:00
struct basename_prefetch_options prefetch_options = {
.repo = options->repo,
diffcore-rename: use a different prefetch for basename comparisons merge-ort was designed to minimize the amount of data needed and used, and several changes were made to diffcore-rename to take advantage of extra metadata to enable this data minimization (particularly the relevant_sources variable for skipping "irrelevant" renames). This effort obviously succeeded in drastically reducing computation times, but should also theoretically allow partial clones to download much less information. Previously, though, the "prefetch" command used in diffcore-rename had never been modified and downloaded many blobs that were unnecessary for merge-ort. This commit corrects that. When doing basename comparisons, we want to fetch only the objects that will be used for basename comparisons. If after basename fetching this leaves us with no more relevant sources (or no more destinations), then we won't need to do the full inexact rename detection and can skip downloading additional source and destination files. Even if we have to do that later full inexact rename detection, irrelevant sources are culled after basename matching and before the full inexact rename detection, so we can still avoid downloading the blobs for irrelevant sources. Rename prefetch() to inexact_prefetch(), and introduce a new basename_prefetch() to take advantage of this. If we modify the testcase from commit 557ac0350d ("merge-ort: begin performance work; instrument with trace2_region_* calls", 2021-01-23) to pass --sparse --filter=blob:none to the clone command, and use the new trace2 "fetch_count" output from a few commits ago to track both the number of fetch subcommands invoked and the number of objects fetched across all those fetches, then for the mega-renames testcase we observe the following: BEFORE this commit, rebasing 35 patches: strategy # of fetches total # of objects fetched --------- ------------ -------------------------- recursive 62 11423 ort 30 11391 AFTER this commit, rebasing the same 35 patches: ort 32 63 This means that the new code only needs to download less than 2 blobs per patch being rebased. That is especially interesting given that the repository at the start only had approximately half a dozen TOTAL blobs downloaded to start with (because the default sparse-checkout of just the toplevel directory was in use). So, for this particular linux kernel testcase that involved ~26,000 renames on the upstream side (drivers/ -> pilots/) across which 35 patches were being rebased, this change reduces the number of blobs that need to be downloaded by a factor of ~180. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-06-22 08:04:40 +00:00
.relevant_sources = relevant_sources,
.sources = &sources,
.dests = &dests,
.info = info
};
/*
* Create maps of basename -> fullname(s) for remaining sources and
* dests.
*/
strintmap_init_with_options(&sources, -1, NULL, 0);
strintmap_init_with_options(&dests, -1, NULL, 0);
for (i = 0; i < rename_src_nr; ++i) {
char *filename = rename_src[i].p->one->path;
const char *base;
/* exact renames removed in remove_unneeded_paths_from_src() */
assert(!rename_src[i].p->one->rename_used);
/* Record index within rename_src (i) if basename is unique */
base = get_basename(filename);
if (strintmap_contains(&sources, base))
strintmap_set(&sources, base, -1);
else
strintmap_set(&sources, base, i);
}
for (i = 0; i < rename_dst_nr; ++i) {
char *filename = rename_dst[i].p->two->path;
const char *base;
if (rename_dst[i].is_rename)
continue; /* involved in exact match already. */
/* Record index within rename_dst (i) if basename is unique */
base = get_basename(filename);
if (strintmap_contains(&dests, base))
strintmap_set(&dests, base, -1);
else
strintmap_set(&dests, base, i);
}
if (options->repo == the_repository && repo_has_promisor_remote(the_repository)) {
diffcore-rename: use a different prefetch for basename comparisons merge-ort was designed to minimize the amount of data needed and used, and several changes were made to diffcore-rename to take advantage of extra metadata to enable this data minimization (particularly the relevant_sources variable for skipping "irrelevant" renames). This effort obviously succeeded in drastically reducing computation times, but should also theoretically allow partial clones to download much less information. Previously, though, the "prefetch" command used in diffcore-rename had never been modified and downloaded many blobs that were unnecessary for merge-ort. This commit corrects that. When doing basename comparisons, we want to fetch only the objects that will be used for basename comparisons. If after basename fetching this leaves us with no more relevant sources (or no more destinations), then we won't need to do the full inexact rename detection and can skip downloading additional source and destination files. Even if we have to do that later full inexact rename detection, irrelevant sources are culled after basename matching and before the full inexact rename detection, so we can still avoid downloading the blobs for irrelevant sources. Rename prefetch() to inexact_prefetch(), and introduce a new basename_prefetch() to take advantage of this. If we modify the testcase from commit 557ac0350d ("merge-ort: begin performance work; instrument with trace2_region_* calls", 2021-01-23) to pass --sparse --filter=blob:none to the clone command, and use the new trace2 "fetch_count" output from a few commits ago to track both the number of fetch subcommands invoked and the number of objects fetched across all those fetches, then for the mega-renames testcase we observe the following: BEFORE this commit, rebasing 35 patches: strategy # of fetches total # of objects fetched --------- ------------ -------------------------- recursive 62 11423 ort 30 11391 AFTER this commit, rebasing the same 35 patches: ort 32 63 This means that the new code only needs to download less than 2 blobs per patch being rebased. That is especially interesting given that the repository at the start only had approximately half a dozen TOTAL blobs downloaded to start with (because the default sparse-checkout of just the toplevel directory was in use). So, for this particular linux kernel testcase that involved ~26,000 renames on the upstream side (drivers/ -> pilots/) across which 35 patches were being rebased, this change reduces the number of blobs that need to be downloaded by a factor of ~180. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-06-22 08:04:40 +00:00
dpf_options.missing_object_cb = basename_prefetch;
dpf_options.missing_object_data = &prefetch_options;
}
diffcore-rename: complete find_basename_matches() It is not uncommon in real world repositories for the majority of file renames to not change the basename of the file; i.e. most "renames" are just a move of files into different directories. We can make use of this to avoid comparing all rename source candidates with all rename destination candidates, by first comparing sources to destinations with the same basenames. If two files with the same basename are sufficiently similar, we record the rename; if not, we include those files in the more exhaustive matrix comparison. This means we are adding a set of preliminary additional comparisons, but for each file we only compare it with at most one other file. For example, if there was a include/media/device.h that was deleted and a src/module/media/device.h that was added, and there are no other device.h files in the remaining sets of added and deleted files after exact rename detection, then these two files would be compared in the preliminary step. This commit does not yet actually employ this new optimization, it merely adds a function which can be used for this purpose. The next commit will do the necessary plumbing to make use of it. Note that this optimization might give us different results than without the optimization, because it's possible that despite files with the same basename being sufficiently similar to be considered a rename, there's an even better match between files without the same basename. I think that is okay for four reasons: (1) it's easy to explain to the users what happened if it does ever occur (or even for them to intuitively figure out), (2) as the next patch will show it provides such a large performance boost that it's worth the tradeoff, and (3) it's somewhat unlikely that despite having unique matching basenames that other files serve as better matches. Reason (4) takes a full paragraph to explain... If the previous three reasons aren't enough, consider what rename detection already does. Break detection is not the default, meaning that if files have the same _fullname_, then they are considered related even if they are 0% similar. In fact, in such a case, we don't even bother comparing the files to see if they are similar let alone comparing them to all other files to see what they are most similar to. Basically, we override content similarity based on sufficient filename similarity. Without the filename similarity (currently implemented as an exact match of filename), we swing the pendulum the opposite direction and say that filename similarity is irrelevant and compare a full N x M matrix of sources and destinations to find out which have the most similar contents. This optimization just adds another form of filename similarity comparison, but augments it with a file content similarity check as well. Basically, if two files have the same basename and are sufficiently similar to be considered a rename, mark them as such without comparing the two to all other rename candidates. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-02-14 07:51:48 +00:00
/* Now look for basename matchups and do similarity estimation */
diffcore-rename: use directory rename guided basename comparisons A previous commit noted that it is very common for people to move files across directories while keeping their filename the same. The last few commits took advantage of this and showed that we can accelerate rename detection significantly using basenames; since files with the same basename serve as likely rename candidates, we can check those first and remove them from the rename candidate pool if they are sufficiently similar. Unfortunately, the previous optimization was limited by the fact that the remaining basenames after exact rename detection are not always unique. Many repositories have hundreds of build files with the same name (e.g. Makefile, .gitignore, build.gradle, etc.), and may even have hundreds of source files with the same name. (For example, the linux kernel has 100 setup.c, 87 irq.c, and 112 core.c files. A repository at $DAYJOB has a lot of ObjectFactory.java and Plugin.java files). For these files with non-unique basenames, we are faced with the task of attempting to determine or guess which directory they may have been relocated to. Such a task is precisely the job of directory rename detection. However, there are two catches: (1) the directory rename detection code has traditionally been part of the merge machinery rather than diffcore-rename.c, and (2) directory rename detection currently runs after regular rename detection is complete. The 1st catch is just an implementation issue that can be overcome by some code shuffling. The 2nd requires us to add a further approximation: we only have access to exact renames at this point, so we need to do directory rename detection based on just exact renames. In some cases we won't have exact renames, in which case this extra optimization won't apply. We also choose to not apply the optimization unless we know that the underlying directory was removed, which will require extra data to be passed in to diffcore_rename_extended(). Also, even if we get a prediction about which directory a file may have relocated to, we will still need to check to see if there is a file in the predicted directory, and then compare the two files to see if they meet the higher min_basename_score threshold required for marking the two files as renames. This commit introduces an idx_possible_rename() function which will do this directory rename detection for us and give us the index within rename_dst of the resulting filename. For now, this function is hardcoded to return -1 (not found) and just hooks up how its results would be used once we have a more complete implementation in place. Reviewed-by: Derrick Stolee <dstolee@microsoft.com> Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-02-27 00:30:39 +00:00
for (i = 0; i < rename_src_nr; ++i) {
char *filename = rename_src[i].p->one->path;
const char *base = NULL;
intptr_t src_index;
diffcore-rename: complete find_basename_matches() It is not uncommon in real world repositories for the majority of file renames to not change the basename of the file; i.e. most "renames" are just a move of files into different directories. We can make use of this to avoid comparing all rename source candidates with all rename destination candidates, by first comparing sources to destinations with the same basenames. If two files with the same basename are sufficiently similar, we record the rename; if not, we include those files in the more exhaustive matrix comparison. This means we are adding a set of preliminary additional comparisons, but for each file we only compare it with at most one other file. For example, if there was a include/media/device.h that was deleted and a src/module/media/device.h that was added, and there are no other device.h files in the remaining sets of added and deleted files after exact rename detection, then these two files would be compared in the preliminary step. This commit does not yet actually employ this new optimization, it merely adds a function which can be used for this purpose. The next commit will do the necessary plumbing to make use of it. Note that this optimization might give us different results than without the optimization, because it's possible that despite files with the same basename being sufficiently similar to be considered a rename, there's an even better match between files without the same basename. I think that is okay for four reasons: (1) it's easy to explain to the users what happened if it does ever occur (or even for them to intuitively figure out), (2) as the next patch will show it provides such a large performance boost that it's worth the tradeoff, and (3) it's somewhat unlikely that despite having unique matching basenames that other files serve as better matches. Reason (4) takes a full paragraph to explain... If the previous three reasons aren't enough, consider what rename detection already does. Break detection is not the default, meaning that if files have the same _fullname_, then they are considered related even if they are 0% similar. In fact, in such a case, we don't even bother comparing the files to see if they are similar let alone comparing them to all other files to see what they are most similar to. Basically, we override content similarity based on sufficient filename similarity. Without the filename similarity (currently implemented as an exact match of filename), we swing the pendulum the opposite direction and say that filename similarity is irrelevant and compare a full N x M matrix of sources and destinations to find out which have the most similar contents. This optimization just adds another form of filename similarity comparison, but augments it with a file content similarity check as well. Basically, if two files have the same basename and are sufficiently similar to be considered a rename, mark them as such without comparing the two to all other rename candidates. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-02-14 07:51:48 +00:00
intptr_t dst_index;
/* Skip irrelevant sources */
if (relevant_sources &&
!strintmap_contains(relevant_sources, filename))
continue;
diffcore-rename: use directory rename guided basename comparisons A previous commit noted that it is very common for people to move files across directories while keeping their filename the same. The last few commits took advantage of this and showed that we can accelerate rename detection significantly using basenames; since files with the same basename serve as likely rename candidates, we can check those first and remove them from the rename candidate pool if they are sufficiently similar. Unfortunately, the previous optimization was limited by the fact that the remaining basenames after exact rename detection are not always unique. Many repositories have hundreds of build files with the same name (e.g. Makefile, .gitignore, build.gradle, etc.), and may even have hundreds of source files with the same name. (For example, the linux kernel has 100 setup.c, 87 irq.c, and 112 core.c files. A repository at $DAYJOB has a lot of ObjectFactory.java and Plugin.java files). For these files with non-unique basenames, we are faced with the task of attempting to determine or guess which directory they may have been relocated to. Such a task is precisely the job of directory rename detection. However, there are two catches: (1) the directory rename detection code has traditionally been part of the merge machinery rather than diffcore-rename.c, and (2) directory rename detection currently runs after regular rename detection is complete. The 1st catch is just an implementation issue that can be overcome by some code shuffling. The 2nd requires us to add a further approximation: we only have access to exact renames at this point, so we need to do directory rename detection based on just exact renames. In some cases we won't have exact renames, in which case this extra optimization won't apply. We also choose to not apply the optimization unless we know that the underlying directory was removed, which will require extra data to be passed in to diffcore_rename_extended(). Also, even if we get a prediction about which directory a file may have relocated to, we will still need to check to see if there is a file in the predicted directory, and then compare the two files to see if they meet the higher min_basename_score threshold required for marking the two files as renames. This commit introduces an idx_possible_rename() function which will do this directory rename detection for us and give us the index within rename_dst of the resulting filename. For now, this function is hardcoded to return -1 (not found) and just hooks up how its results would be used once we have a more complete implementation in place. Reviewed-by: Derrick Stolee <dstolee@microsoft.com> Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-02-27 00:30:39 +00:00
/*
* If the basename is unique among remaining sources, then
* src_index will equal 'i' and we can attempt to match it
* to a unique basename in the destinations. Otherwise,
* use directory rename heuristics, if possible.
*/
base = get_basename(filename);
src_index = strintmap_get(&sources, base);
assert(src_index == -1 || src_index == i);
if (strintmap_contains(&dests, base)) {
diffcore-rename: complete find_basename_matches() It is not uncommon in real world repositories for the majority of file renames to not change the basename of the file; i.e. most "renames" are just a move of files into different directories. We can make use of this to avoid comparing all rename source candidates with all rename destination candidates, by first comparing sources to destinations with the same basenames. If two files with the same basename are sufficiently similar, we record the rename; if not, we include those files in the more exhaustive matrix comparison. This means we are adding a set of preliminary additional comparisons, but for each file we only compare it with at most one other file. For example, if there was a include/media/device.h that was deleted and a src/module/media/device.h that was added, and there are no other device.h files in the remaining sets of added and deleted files after exact rename detection, then these two files would be compared in the preliminary step. This commit does not yet actually employ this new optimization, it merely adds a function which can be used for this purpose. The next commit will do the necessary plumbing to make use of it. Note that this optimization might give us different results than without the optimization, because it's possible that despite files with the same basename being sufficiently similar to be considered a rename, there's an even better match between files without the same basename. I think that is okay for four reasons: (1) it's easy to explain to the users what happened if it does ever occur (or even for them to intuitively figure out), (2) as the next patch will show it provides such a large performance boost that it's worth the tradeoff, and (3) it's somewhat unlikely that despite having unique matching basenames that other files serve as better matches. Reason (4) takes a full paragraph to explain... If the previous three reasons aren't enough, consider what rename detection already does. Break detection is not the default, meaning that if files have the same _fullname_, then they are considered related even if they are 0% similar. In fact, in such a case, we don't even bother comparing the files to see if they are similar let alone comparing them to all other files to see what they are most similar to. Basically, we override content similarity based on sufficient filename similarity. Without the filename similarity (currently implemented as an exact match of filename), we swing the pendulum the opposite direction and say that filename similarity is irrelevant and compare a full N x M matrix of sources and destinations to find out which have the most similar contents. This optimization just adds another form of filename similarity comparison, but augments it with a file content similarity check as well. Basically, if two files have the same basename and are sufficiently similar to be considered a rename, mark them as such without comparing the two to all other rename candidates. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-02-14 07:51:48 +00:00
struct diff_filespec *one, *two;
int score;
diffcore-rename: use directory rename guided basename comparisons A previous commit noted that it is very common for people to move files across directories while keeping their filename the same. The last few commits took advantage of this and showed that we can accelerate rename detection significantly using basenames; since files with the same basename serve as likely rename candidates, we can check those first and remove them from the rename candidate pool if they are sufficiently similar. Unfortunately, the previous optimization was limited by the fact that the remaining basenames after exact rename detection are not always unique. Many repositories have hundreds of build files with the same name (e.g. Makefile, .gitignore, build.gradle, etc.), and may even have hundreds of source files with the same name. (For example, the linux kernel has 100 setup.c, 87 irq.c, and 112 core.c files. A repository at $DAYJOB has a lot of ObjectFactory.java and Plugin.java files). For these files with non-unique basenames, we are faced with the task of attempting to determine or guess which directory they may have been relocated to. Such a task is precisely the job of directory rename detection. However, there are two catches: (1) the directory rename detection code has traditionally been part of the merge machinery rather than diffcore-rename.c, and (2) directory rename detection currently runs after regular rename detection is complete. The 1st catch is just an implementation issue that can be overcome by some code shuffling. The 2nd requires us to add a further approximation: we only have access to exact renames at this point, so we need to do directory rename detection based on just exact renames. In some cases we won't have exact renames, in which case this extra optimization won't apply. We also choose to not apply the optimization unless we know that the underlying directory was removed, which will require extra data to be passed in to diffcore_rename_extended(). Also, even if we get a prediction about which directory a file may have relocated to, we will still need to check to see if there is a file in the predicted directory, and then compare the two files to see if they meet the higher min_basename_score threshold required for marking the two files as renames. This commit introduces an idx_possible_rename() function which will do this directory rename detection for us and give us the index within rename_dst of the resulting filename. For now, this function is hardcoded to return -1 (not found) and just hooks up how its results would be used once we have a more complete implementation in place. Reviewed-by: Derrick Stolee <dstolee@microsoft.com> Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-02-27 00:30:39 +00:00
/* Find a matching destination, if possible */
dst_index = strintmap_get(&dests, base);
if (src_index == -1 || dst_index == -1) {
src_index = i;
dst_index = idx_possible_rename(filename, info);
diffcore-rename: use directory rename guided basename comparisons A previous commit noted that it is very common for people to move files across directories while keeping their filename the same. The last few commits took advantage of this and showed that we can accelerate rename detection significantly using basenames; since files with the same basename serve as likely rename candidates, we can check those first and remove them from the rename candidate pool if they are sufficiently similar. Unfortunately, the previous optimization was limited by the fact that the remaining basenames after exact rename detection are not always unique. Many repositories have hundreds of build files with the same name (e.g. Makefile, .gitignore, build.gradle, etc.), and may even have hundreds of source files with the same name. (For example, the linux kernel has 100 setup.c, 87 irq.c, and 112 core.c files. A repository at $DAYJOB has a lot of ObjectFactory.java and Plugin.java files). For these files with non-unique basenames, we are faced with the task of attempting to determine or guess which directory they may have been relocated to. Such a task is precisely the job of directory rename detection. However, there are two catches: (1) the directory rename detection code has traditionally been part of the merge machinery rather than diffcore-rename.c, and (2) directory rename detection currently runs after regular rename detection is complete. The 1st catch is just an implementation issue that can be overcome by some code shuffling. The 2nd requires us to add a further approximation: we only have access to exact renames at this point, so we need to do directory rename detection based on just exact renames. In some cases we won't have exact renames, in which case this extra optimization won't apply. We also choose to not apply the optimization unless we know that the underlying directory was removed, which will require extra data to be passed in to diffcore_rename_extended(). Also, even if we get a prediction about which directory a file may have relocated to, we will still need to check to see if there is a file in the predicted directory, and then compare the two files to see if they meet the higher min_basename_score threshold required for marking the two files as renames. This commit introduces an idx_possible_rename() function which will do this directory rename detection for us and give us the index within rename_dst of the resulting filename. For now, this function is hardcoded to return -1 (not found) and just hooks up how its results would be used once we have a more complete implementation in place. Reviewed-by: Derrick Stolee <dstolee@microsoft.com> Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-02-27 00:30:39 +00:00
}
if (dst_index == -1)
continue;
/* Ignore this dest if already used in a rename */
if (rename_dst[dst_index].is_rename)
continue; /* already used previously */
diffcore-rename: complete find_basename_matches() It is not uncommon in real world repositories for the majority of file renames to not change the basename of the file; i.e. most "renames" are just a move of files into different directories. We can make use of this to avoid comparing all rename source candidates with all rename destination candidates, by first comparing sources to destinations with the same basenames. If two files with the same basename are sufficiently similar, we record the rename; if not, we include those files in the more exhaustive matrix comparison. This means we are adding a set of preliminary additional comparisons, but for each file we only compare it with at most one other file. For example, if there was a include/media/device.h that was deleted and a src/module/media/device.h that was added, and there are no other device.h files in the remaining sets of added and deleted files after exact rename detection, then these two files would be compared in the preliminary step. This commit does not yet actually employ this new optimization, it merely adds a function which can be used for this purpose. The next commit will do the necessary plumbing to make use of it. Note that this optimization might give us different results than without the optimization, because it's possible that despite files with the same basename being sufficiently similar to be considered a rename, there's an even better match between files without the same basename. I think that is okay for four reasons: (1) it's easy to explain to the users what happened if it does ever occur (or even for them to intuitively figure out), (2) as the next patch will show it provides such a large performance boost that it's worth the tradeoff, and (3) it's somewhat unlikely that despite having unique matching basenames that other files serve as better matches. Reason (4) takes a full paragraph to explain... If the previous three reasons aren't enough, consider what rename detection already does. Break detection is not the default, meaning that if files have the same _fullname_, then they are considered related even if they are 0% similar. In fact, in such a case, we don't even bother comparing the files to see if they are similar let alone comparing them to all other files to see what they are most similar to. Basically, we override content similarity based on sufficient filename similarity. Without the filename similarity (currently implemented as an exact match of filename), we swing the pendulum the opposite direction and say that filename similarity is irrelevant and compare a full N x M matrix of sources and destinations to find out which have the most similar contents. This optimization just adds another form of filename similarity comparison, but augments it with a file content similarity check as well. Basically, if two files have the same basename and are sufficiently similar to be considered a rename, mark them as such without comparing the two to all other rename candidates. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-02-14 07:51:48 +00:00
/* Estimate the similarity */
one = rename_src[src_index].p->one;
two = rename_dst[dst_index].p->two;
score = estimate_similarity(options->repo, one, two,
minimum_score, &dpf_options);
diffcore-rename: complete find_basename_matches() It is not uncommon in real world repositories for the majority of file renames to not change the basename of the file; i.e. most "renames" are just a move of files into different directories. We can make use of this to avoid comparing all rename source candidates with all rename destination candidates, by first comparing sources to destinations with the same basenames. If two files with the same basename are sufficiently similar, we record the rename; if not, we include those files in the more exhaustive matrix comparison. This means we are adding a set of preliminary additional comparisons, but for each file we only compare it with at most one other file. For example, if there was a include/media/device.h that was deleted and a src/module/media/device.h that was added, and there are no other device.h files in the remaining sets of added and deleted files after exact rename detection, then these two files would be compared in the preliminary step. This commit does not yet actually employ this new optimization, it merely adds a function which can be used for this purpose. The next commit will do the necessary plumbing to make use of it. Note that this optimization might give us different results than without the optimization, because it's possible that despite files with the same basename being sufficiently similar to be considered a rename, there's an even better match between files without the same basename. I think that is okay for four reasons: (1) it's easy to explain to the users what happened if it does ever occur (or even for them to intuitively figure out), (2) as the next patch will show it provides such a large performance boost that it's worth the tradeoff, and (3) it's somewhat unlikely that despite having unique matching basenames that other files serve as better matches. Reason (4) takes a full paragraph to explain... If the previous three reasons aren't enough, consider what rename detection already does. Break detection is not the default, meaning that if files have the same _fullname_, then they are considered related even if they are 0% similar. In fact, in such a case, we don't even bother comparing the files to see if they are similar let alone comparing them to all other files to see what they are most similar to. Basically, we override content similarity based on sufficient filename similarity. Without the filename similarity (currently implemented as an exact match of filename), we swing the pendulum the opposite direction and say that filename similarity is irrelevant and compare a full N x M matrix of sources and destinations to find out which have the most similar contents. This optimization just adds another form of filename similarity comparison, but augments it with a file content similarity check as well. Basically, if two files have the same basename and are sufficiently similar to be considered a rename, mark them as such without comparing the two to all other rename candidates. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-02-14 07:51:48 +00:00
/* If sufficiently similar, record as rename pair */
if (score < minimum_score)
continue;
record_rename_pair(dst_index, src_index, score);
renames++;
update_dir_rename_counts(info, dirs_removed,
one->path, two->path);
diffcore-rename: complete find_basename_matches() It is not uncommon in real world repositories for the majority of file renames to not change the basename of the file; i.e. most "renames" are just a move of files into different directories. We can make use of this to avoid comparing all rename source candidates with all rename destination candidates, by first comparing sources to destinations with the same basenames. If two files with the same basename are sufficiently similar, we record the rename; if not, we include those files in the more exhaustive matrix comparison. This means we are adding a set of preliminary additional comparisons, but for each file we only compare it with at most one other file. For example, if there was a include/media/device.h that was deleted and a src/module/media/device.h that was added, and there are no other device.h files in the remaining sets of added and deleted files after exact rename detection, then these two files would be compared in the preliminary step. This commit does not yet actually employ this new optimization, it merely adds a function which can be used for this purpose. The next commit will do the necessary plumbing to make use of it. Note that this optimization might give us different results than without the optimization, because it's possible that despite files with the same basename being sufficiently similar to be considered a rename, there's an even better match between files without the same basename. I think that is okay for four reasons: (1) it's easy to explain to the users what happened if it does ever occur (or even for them to intuitively figure out), (2) as the next patch will show it provides such a large performance boost that it's worth the tradeoff, and (3) it's somewhat unlikely that despite having unique matching basenames that other files serve as better matches. Reason (4) takes a full paragraph to explain... If the previous three reasons aren't enough, consider what rename detection already does. Break detection is not the default, meaning that if files have the same _fullname_, then they are considered related even if they are 0% similar. In fact, in such a case, we don't even bother comparing the files to see if they are similar let alone comparing them to all other files to see what they are most similar to. Basically, we override content similarity based on sufficient filename similarity. Without the filename similarity (currently implemented as an exact match of filename), we swing the pendulum the opposite direction and say that filename similarity is irrelevant and compare a full N x M matrix of sources and destinations to find out which have the most similar contents. This optimization just adds another form of filename similarity comparison, but augments it with a file content similarity check as well. Basically, if two files have the same basename and are sufficiently similar to be considered a rename, mark them as such without comparing the two to all other rename candidates. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-02-14 07:51:48 +00:00
/*
* Found a rename so don't need text anymore; if we
* didn't find a rename, the filespec_blob would get
* re-used when doing the matrix of comparisons.
*/
diff_free_filespec_blob(one);
diff_free_filespec_blob(two);
}
}
strintmap_clear(&sources);
strintmap_clear(&dests);
diffcore-rename: complete find_basename_matches() It is not uncommon in real world repositories for the majority of file renames to not change the basename of the file; i.e. most "renames" are just a move of files into different directories. We can make use of this to avoid comparing all rename source candidates with all rename destination candidates, by first comparing sources to destinations with the same basenames. If two files with the same basename are sufficiently similar, we record the rename; if not, we include those files in the more exhaustive matrix comparison. This means we are adding a set of preliminary additional comparisons, but for each file we only compare it with at most one other file. For example, if there was a include/media/device.h that was deleted and a src/module/media/device.h that was added, and there are no other device.h files in the remaining sets of added and deleted files after exact rename detection, then these two files would be compared in the preliminary step. This commit does not yet actually employ this new optimization, it merely adds a function which can be used for this purpose. The next commit will do the necessary plumbing to make use of it. Note that this optimization might give us different results than without the optimization, because it's possible that despite files with the same basename being sufficiently similar to be considered a rename, there's an even better match between files without the same basename. I think that is okay for four reasons: (1) it's easy to explain to the users what happened if it does ever occur (or even for them to intuitively figure out), (2) as the next patch will show it provides such a large performance boost that it's worth the tradeoff, and (3) it's somewhat unlikely that despite having unique matching basenames that other files serve as better matches. Reason (4) takes a full paragraph to explain... If the previous three reasons aren't enough, consider what rename detection already does. Break detection is not the default, meaning that if files have the same _fullname_, then they are considered related even if they are 0% similar. In fact, in such a case, we don't even bother comparing the files to see if they are similar let alone comparing them to all other files to see what they are most similar to. Basically, we override content similarity based on sufficient filename similarity. Without the filename similarity (currently implemented as an exact match of filename), we swing the pendulum the opposite direction and say that filename similarity is irrelevant and compare a full N x M matrix of sources and destinations to find out which have the most similar contents. This optimization just adds another form of filename similarity comparison, but augments it with a file content similarity check as well. Basically, if two files have the same basename and are sufficiently similar to be considered a rename, mark them as such without comparing the two to all other rename candidates. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-02-14 07:51:48 +00:00
return renames;
}
#define NUM_CANDIDATE_PER_DST 4
static void record_if_better(struct diff_score m[], struct diff_score *o)
{
int i, worst;
/* find the worst one */
worst = 0;
for (i = 1; i < NUM_CANDIDATE_PER_DST; i++)
if (score_compare(&m[i], &m[worst]) > 0)
worst = i;
/* is it better than the worst one? */
if (score_compare(&m[worst], o) > 0)
m[worst] = *o;
}
/*
* Returns:
* 0 if we are under the limit;
* 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, int num_sources,
struct diff_options *options)
{
int rename_limit = options->rename_limit;
int i, limited_sources;
options->needed_rename_limit = 0;
/*
* This basically does a test for the rename matrix not
* 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)
return 0; /* treat as unlimited */
if (st_mult(num_destinations, num_sources)
<= st_mult(rename_limit, rename_limit))
return 0;
options->needed_rename_limit =
num_sources > num_destinations ? num_sources : num_destinations;
/* Are we running under -C -C? */
diff: make struct diff_flags members lowercase Now that the flags stored in struct diff_flags are being accessed directly and not through macros, change all struct members from being uppercase to lowercase. This conversion is done using the following semantic patch: @@ expression E; @@ - E.RECURSIVE + E.recursive @@ expression E; @@ - E.TREE_IN_RECURSIVE + E.tree_in_recursive @@ expression E; @@ - E.BINARY + E.binary @@ expression E; @@ - E.TEXT + E.text @@ expression E; @@ - E.FULL_INDEX + E.full_index @@ expression E; @@ - E.SILENT_ON_REMOVE + E.silent_on_remove @@ expression E; @@ - E.FIND_COPIES_HARDER + E.find_copies_harder @@ expression E; @@ - E.FOLLOW_RENAMES + E.follow_renames @@ expression E; @@ - E.RENAME_EMPTY + E.rename_empty @@ expression E; @@ - E.HAS_CHANGES + E.has_changes @@ expression E; @@ - E.QUICK + E.quick @@ expression E; @@ - E.NO_INDEX + E.no_index @@ expression E; @@ - E.ALLOW_EXTERNAL + E.allow_external @@ expression E; @@ - E.EXIT_WITH_STATUS + E.exit_with_status @@ expression E; @@ - E.REVERSE_DIFF + E.reverse_diff @@ expression E; @@ - E.CHECK_FAILED + E.check_failed @@ expression E; @@ - E.RELATIVE_NAME + E.relative_name @@ expression E; @@ - E.IGNORE_SUBMODULES + E.ignore_submodules @@ expression E; @@ - E.DIRSTAT_CUMULATIVE + E.dirstat_cumulative @@ expression E; @@ - E.DIRSTAT_BY_FILE + E.dirstat_by_file @@ expression E; @@ - E.ALLOW_TEXTCONV + E.allow_textconv @@ expression E; @@ - E.TEXTCONV_SET_VIA_CMDLINE + E.textconv_set_via_cmdline @@ expression E; @@ - E.DIFF_FROM_CONTENTS + E.diff_from_contents @@ expression E; @@ - E.DIRTY_SUBMODULES + E.dirty_submodules @@ expression E; @@ - E.IGNORE_UNTRACKED_IN_SUBMODULES + E.ignore_untracked_in_submodules @@ expression E; @@ - E.IGNORE_DIRTY_SUBMODULES + E.ignore_dirty_submodules @@ expression E; @@ - E.OVERRIDE_SUBMODULE_CONFIG + E.override_submodule_config @@ expression E; @@ - E.DIRSTAT_BY_LINE + E.dirstat_by_line @@ expression E; @@ - E.FUNCCONTEXT + E.funccontext @@ expression E; @@ - E.PICKAXE_IGNORE_CASE + E.pickaxe_ignore_case @@ expression E; @@ - E.DEFAULT_FOLLOW_RENAMES + E.default_follow_renames Signed-off-by: Brandon Williams <bmwill@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-10-31 18:19:11 +00:00
if (!options->flags.find_copies_harder)
return 1;
/* Would we bust the limit if we were running under -C? */
for (limited_sources = i = 0; i < num_sources; i++) {
if (diff_unmodified_pair(rename_src[i].p))
continue;
limited_sources++;
}
if (st_mult(num_destinations, limited_sources)
<= st_mult(rename_limit, rename_limit))
return 2;
return 1;
}
static int find_renames(struct diff_score *mx,
int dst_cnt,
int minimum_score,
int copies,
struct dir_rename_info *info,
struct strintmap *dirs_removed)
{
int count = 0, i;
for (i = 0; i < dst_cnt * NUM_CANDIDATE_PER_DST; i++) {
struct diff_rename_dst *dst;
if ((mx[i].dst < 0) ||
(mx[i].score < minimum_score))
break; /* there is no more usable pair. */
dst = &rename_dst[mx[i].dst];
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 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. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-11 09:08:47 +00:00
if (dst->is_rename)
continue; /* already done, either exact or fuzzy. */
if (!copies && rename_src[mx[i].src].p->one->rename_used)
continue;
record_rename_pair(mx[i].dst, mx[i].src, mx[i].score);
count++;
update_dir_rename_counts(info, dirs_removed,
rename_src[mx[i].src].p->one->path,
rename_dst[mx[i].dst].p->two->path);
}
return count;
}
diffcore-rename: enable filtering possible rename sources Add the ability to diffcore_rename_extended() to allow external callers to declare that they only need renames detected for a subset of source files, and use that information to skip detecting renames for them. There are two important pieces to this optimization that may not be obvious at first glance: * We do not require callers to just filter the filepairs out to remove the non-relevant sources, because exact rename detection is fast and when it finds a match it can remove both a source and a destination whereas the relevant_sources filter can only remove a source. * We need to filter out the source pairs in a preliminary pass instead of adding a strset_contains(relevant_sources, one->path) check within the nested matrix loop. The reason for that is if we have 30k renames, doing 30k * 30k = 900M strset_contains() calls becomes extraordinarily expensive and defeats the performance gains from this change; we only want to do 30k such calls instead. If callers pass NULL for relevant_sources, that is special cases to treat all sources as relevant. Since all callers currently pass NULL, this optimization does not yet have any effect. Subsequent commits will have merge-ort compute a set of relevant_sources to restrict which sources we detect renames for, and have merge-ort pass that set of relevant_sources to diffcore_rename_extended(). A note about filtering order: Some may be curious why we don't filter out irrelevant sources at the same time we filter out exact renames. While that technically could be done at this point, there are two reasons to defer it: First, was to reinforce a lesson that was too easy to forget. As I mentioned above, in the past I filtered irrelevant sources out before exact rename checking, and then discovered that exact renames' ability to remove both sources and destinations was an important consideration and thus doing the filtering after exact rename checking would speed things up. Then at some point I realized that basename matching could also remove both sources and destinations, and decided to put irrelevant source filtering after basename filtering. That slowed things down a lot. But, despite learning about this important ordering, in later restructuring I forgot and made the same mistake of putting the filtering after basename guided rename detection again. So, I have this series of patches structured to do the irrelevant filtering last to start to show how much extra it costs, and then add relevant filtering in to find_basename_matches() to show how much it speeds things up. Basically, it's a way to reinforce something that apparently was too easy to forget, and make sure the commit messages record this lesson. Second, the items in the "relevant_sources" in this patch series will include all sources that *might be* relevant. It has to be conservative and catch anything that might need a rename, but in the patch series after this one we'll find ways to weed out more of the *might be* relevant ones. Unfortunately, merge-ort does not have sufficient information to weed those ones out, and there isn't enough information at the time of filtering exact renames out to remove the extra ones either. It has to be deferred. So the deferral is in part to simplify some later additions. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-03-11 00:38:24 +00:00
static void remove_unneeded_paths_from_src(int detecting_copies,
struct strintmap *interesting)
{
int i, new_num_src;
diffcore-rename: enable filtering possible rename sources Add the ability to diffcore_rename_extended() to allow external callers to declare that they only need renames detected for a subset of source files, and use that information to skip detecting renames for them. There are two important pieces to this optimization that may not be obvious at first glance: * We do not require callers to just filter the filepairs out to remove the non-relevant sources, because exact rename detection is fast and when it finds a match it can remove both a source and a destination whereas the relevant_sources filter can only remove a source. * We need to filter out the source pairs in a preliminary pass instead of adding a strset_contains(relevant_sources, one->path) check within the nested matrix loop. The reason for that is if we have 30k renames, doing 30k * 30k = 900M strset_contains() calls becomes extraordinarily expensive and defeats the performance gains from this change; we only want to do 30k such calls instead. If callers pass NULL for relevant_sources, that is special cases to treat all sources as relevant. Since all callers currently pass NULL, this optimization does not yet have any effect. Subsequent commits will have merge-ort compute a set of relevant_sources to restrict which sources we detect renames for, and have merge-ort pass that set of relevant_sources to diffcore_rename_extended(). A note about filtering order: Some may be curious why we don't filter out irrelevant sources at the same time we filter out exact renames. While that technically could be done at this point, there are two reasons to defer it: First, was to reinforce a lesson that was too easy to forget. As I mentioned above, in the past I filtered irrelevant sources out before exact rename checking, and then discovered that exact renames' ability to remove both sources and destinations was an important consideration and thus doing the filtering after exact rename checking would speed things up. Then at some point I realized that basename matching could also remove both sources and destinations, and decided to put irrelevant source filtering after basename filtering. That slowed things down a lot. But, despite learning about this important ordering, in later restructuring I forgot and made the same mistake of putting the filtering after basename guided rename detection again. So, I have this series of patches structured to do the irrelevant filtering last to start to show how much extra it costs, and then add relevant filtering in to find_basename_matches() to show how much it speeds things up. Basically, it's a way to reinforce something that apparently was too easy to forget, and make sure the commit messages record this lesson. Second, the items in the "relevant_sources" in this patch series will include all sources that *might be* relevant. It has to be conservative and catch anything that might need a rename, but in the patch series after this one we'll find ways to weed out more of the *might be* relevant ones. Unfortunately, merge-ort does not have sufficient information to weed those ones out, and there isn't enough information at the time of filtering exact renames out to remove the extra ones either. It has to be deferred. So the deferral is in part to simplify some later additions. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-03-11 00:38:24 +00:00
if (detecting_copies && !interesting)
return; /* nothing to remove */
if (break_idx)
return; /* culling incompatible with break detection */
/*
* Note on reasons why we cull unneeded sources but not destinations:
* 1) Pairings are stored in rename_dst (not rename_src), which we
* need to keep around. So, we just can't cull rename_dst even
* if we wanted to. But doing so wouldn't help because...
*
* 2) There is a matrix pairwise comparison that follows the
* "Performing inexact rename detection" progress message.
* Iterating over the destinations is done in the outer loop,
* hence we only iterate over each of those once and we can
* easily skip the outer loop early if the destination isn't
* relevant. That's only one check per destination path to
* skip.
*
* By contrast, the sources are iterated in the inner loop; if
* we check whether a source can be skipped, then we'll be
* checking it N separate times, once for each destination.
* We don't want to have to iterate over known-not-needed
* sources N times each, so avoid that by removing the sources
* from rename_src here.
*/
for (i = 0, new_num_src = 0; i < rename_src_nr; i++) {
diffcore-rename: enable filtering possible rename sources Add the ability to diffcore_rename_extended() to allow external callers to declare that they only need renames detected for a subset of source files, and use that information to skip detecting renames for them. There are two important pieces to this optimization that may not be obvious at first glance: * We do not require callers to just filter the filepairs out to remove the non-relevant sources, because exact rename detection is fast and when it finds a match it can remove both a source and a destination whereas the relevant_sources filter can only remove a source. * We need to filter out the source pairs in a preliminary pass instead of adding a strset_contains(relevant_sources, one->path) check within the nested matrix loop. The reason for that is if we have 30k renames, doing 30k * 30k = 900M strset_contains() calls becomes extraordinarily expensive and defeats the performance gains from this change; we only want to do 30k such calls instead. If callers pass NULL for relevant_sources, that is special cases to treat all sources as relevant. Since all callers currently pass NULL, this optimization does not yet have any effect. Subsequent commits will have merge-ort compute a set of relevant_sources to restrict which sources we detect renames for, and have merge-ort pass that set of relevant_sources to diffcore_rename_extended(). A note about filtering order: Some may be curious why we don't filter out irrelevant sources at the same time we filter out exact renames. While that technically could be done at this point, there are two reasons to defer it: First, was to reinforce a lesson that was too easy to forget. As I mentioned above, in the past I filtered irrelevant sources out before exact rename checking, and then discovered that exact renames' ability to remove both sources and destinations was an important consideration and thus doing the filtering after exact rename checking would speed things up. Then at some point I realized that basename matching could also remove both sources and destinations, and decided to put irrelevant source filtering after basename filtering. That slowed things down a lot. But, despite learning about this important ordering, in later restructuring I forgot and made the same mistake of putting the filtering after basename guided rename detection again. So, I have this series of patches structured to do the irrelevant filtering last to start to show how much extra it costs, and then add relevant filtering in to find_basename_matches() to show how much it speeds things up. Basically, it's a way to reinforce something that apparently was too easy to forget, and make sure the commit messages record this lesson. Second, the items in the "relevant_sources" in this patch series will include all sources that *might be* relevant. It has to be conservative and catch anything that might need a rename, but in the patch series after this one we'll find ways to weed out more of the *might be* relevant ones. Unfortunately, merge-ort does not have sufficient information to weed those ones out, and there isn't enough information at the time of filtering exact renames out to remove the extra ones either. It has to be deferred. So the deferral is in part to simplify some later additions. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-03-11 00:38:24 +00:00
struct diff_filespec *one = rename_src[i].p->one;
/*
* renames are stored in rename_dst, so if a rename has
* already been detected using this source, we can just
* remove the source knowing rename_dst has its info.
*/
diffcore-rename: enable filtering possible rename sources Add the ability to diffcore_rename_extended() to allow external callers to declare that they only need renames detected for a subset of source files, and use that information to skip detecting renames for them. There are two important pieces to this optimization that may not be obvious at first glance: * We do not require callers to just filter the filepairs out to remove the non-relevant sources, because exact rename detection is fast and when it finds a match it can remove both a source and a destination whereas the relevant_sources filter can only remove a source. * We need to filter out the source pairs in a preliminary pass instead of adding a strset_contains(relevant_sources, one->path) check within the nested matrix loop. The reason for that is if we have 30k renames, doing 30k * 30k = 900M strset_contains() calls becomes extraordinarily expensive and defeats the performance gains from this change; we only want to do 30k such calls instead. If callers pass NULL for relevant_sources, that is special cases to treat all sources as relevant. Since all callers currently pass NULL, this optimization does not yet have any effect. Subsequent commits will have merge-ort compute a set of relevant_sources to restrict which sources we detect renames for, and have merge-ort pass that set of relevant_sources to diffcore_rename_extended(). A note about filtering order: Some may be curious why we don't filter out irrelevant sources at the same time we filter out exact renames. While that technically could be done at this point, there are two reasons to defer it: First, was to reinforce a lesson that was too easy to forget. As I mentioned above, in the past I filtered irrelevant sources out before exact rename checking, and then discovered that exact renames' ability to remove both sources and destinations was an important consideration and thus doing the filtering after exact rename checking would speed things up. Then at some point I realized that basename matching could also remove both sources and destinations, and decided to put irrelevant source filtering after basename filtering. That slowed things down a lot. But, despite learning about this important ordering, in later restructuring I forgot and made the same mistake of putting the filtering after basename guided rename detection again. So, I have this series of patches structured to do the irrelevant filtering last to start to show how much extra it costs, and then add relevant filtering in to find_basename_matches() to show how much it speeds things up. Basically, it's a way to reinforce something that apparently was too easy to forget, and make sure the commit messages record this lesson. Second, the items in the "relevant_sources" in this patch series will include all sources that *might be* relevant. It has to be conservative and catch anything that might need a rename, but in the patch series after this one we'll find ways to weed out more of the *might be* relevant ones. Unfortunately, merge-ort does not have sufficient information to weed those ones out, and there isn't enough information at the time of filtering exact renames out to remove the extra ones either. It has to be deferred. So the deferral is in part to simplify some later additions. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-03-11 00:38:24 +00:00
if (!detecting_copies && one->rename_used)
continue;
/* If we don't care about the source path, skip it */
if (interesting && !strintmap_contains(interesting, one->path))
continue;
if (new_num_src < i)
memcpy(&rename_src[new_num_src], &rename_src[i],
sizeof(struct diff_rename_src));
new_num_src++;
}
rename_src_nr = new_num_src;
}
diffcore-rename: take advantage of "majority rules" to skip more renames In directory rename detection (when a directory is removed on one side of history and the other side adds new files to that directory), we work to find where the greatest number of files within that directory were renamed to so that the new files can be moved with the majority of the files. Naively, we can just do this by detecting renames for *all* files within the removed/renamed directory, looking at all the destination directories where files within that directory were moved, and if there is more than one such directory then taking the one with the greatest number of files as the directory where the old directory was renamed to. However, sometimes there are enough renames from exact rename detection or basename-guided rename detection that we have enough information to determine the majority winner already. Add a function meant to compute whether particular renames are still needed based on this majority rules check. The next several commits will then add the necessary infrastructure to get the information we need to compute which additional rename sources we can skip. An important side note for future further optimization: There is a possible improvement to this optimization that I have not yet attempted and will not be included in this series of patches: we could first check whether exact renames provide enough information for us to determine directory renames, and avoid doing basename-guided rename detection on some or all of the RELEVANT_LOCATION files within those directories. In effect, this variant would mean doing the handle_early_known_dir_renames() both after exact rename detection and again after basename-guided rename detection, though it would also mean decrementing the number of "unknown" renames for each rename we found from basename-guided rename detection. Adding this additional check for skippable renames right after exact rename detection might turn out to be valuable, especially for partial clones where it might allow us to download certain source files entirely. However, this particular optimization was actually the last one I did in original implementation order, and by the time I implemented this idea, every testcase I had was sufficiently fast that further optimization was unwarranted. If future testcases arise that tax rename detection more heavily (or perhaps partial clones can benefit from avoiding loading more objects), it may be worth implementing this more involved variant. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-03-13 22:22:01 +00:00
static void handle_early_known_dir_renames(struct dir_rename_info *info,
struct strintmap *relevant_sources,
struct strintmap *dirs_removed)
diffcore-rename: take advantage of "majority rules" to skip more renames In directory rename detection (when a directory is removed on one side of history and the other side adds new files to that directory), we work to find where the greatest number of files within that directory were renamed to so that the new files can be moved with the majority of the files. Naively, we can just do this by detecting renames for *all* files within the removed/renamed directory, looking at all the destination directories where files within that directory were moved, and if there is more than one such directory then taking the one with the greatest number of files as the directory where the old directory was renamed to. However, sometimes there are enough renames from exact rename detection or basename-guided rename detection that we have enough information to determine the majority winner already. Add a function meant to compute whether particular renames are still needed based on this majority rules check. The next several commits will then add the necessary infrastructure to get the information we need to compute which additional rename sources we can skip. An important side note for future further optimization: There is a possible improvement to this optimization that I have not yet attempted and will not be included in this series of patches: we could first check whether exact renames provide enough information for us to determine directory renames, and avoid doing basename-guided rename detection on some or all of the RELEVANT_LOCATION files within those directories. In effect, this variant would mean doing the handle_early_known_dir_renames() both after exact rename detection and again after basename-guided rename detection, though it would also mean decrementing the number of "unknown" renames for each rename we found from basename-guided rename detection. Adding this additional check for skippable renames right after exact rename detection might turn out to be valuable, especially for partial clones where it might allow us to download certain source files entirely. However, this particular optimization was actually the last one I did in original implementation order, and by the time I implemented this idea, every testcase I had was sufficiently fast that further optimization was unwarranted. If future testcases arise that tax rename detection more heavily (or perhaps partial clones can benefit from avoiding loading more objects), it may be worth implementing this more involved variant. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-03-13 22:22:01 +00:00
{
/*
* Directory renames are determined via an aggregate of all renames
* under them and using a "majority wins" rule. The fact that
* "majority wins", though, means we don't need all the renames
* under the given directory, we only need enough to ensure we have
* a majority.
diffcore-rename: take advantage of "majority rules" to skip more renames In directory rename detection (when a directory is removed on one side of history and the other side adds new files to that directory), we work to find where the greatest number of files within that directory were renamed to so that the new files can be moved with the majority of the files. Naively, we can just do this by detecting renames for *all* files within the removed/renamed directory, looking at all the destination directories where files within that directory were moved, and if there is more than one such directory then taking the one with the greatest number of files as the directory where the old directory was renamed to. However, sometimes there are enough renames from exact rename detection or basename-guided rename detection that we have enough information to determine the majority winner already. Add a function meant to compute whether particular renames are still needed based on this majority rules check. The next several commits will then add the necessary infrastructure to get the information we need to compute which additional rename sources we can skip. An important side note for future further optimization: There is a possible improvement to this optimization that I have not yet attempted and will not be included in this series of patches: we could first check whether exact renames provide enough information for us to determine directory renames, and avoid doing basename-guided rename detection on some or all of the RELEVANT_LOCATION files within those directories. In effect, this variant would mean doing the handle_early_known_dir_renames() both after exact rename detection and again after basename-guided rename detection, though it would also mean decrementing the number of "unknown" renames for each rename we found from basename-guided rename detection. Adding this additional check for skippable renames right after exact rename detection might turn out to be valuable, especially for partial clones where it might allow us to download certain source files entirely. However, this particular optimization was actually the last one I did in original implementation order, and by the time I implemented this idea, every testcase I had was sufficiently fast that further optimization was unwarranted. If future testcases arise that tax rename detection more heavily (or perhaps partial clones can benefit from avoiding loading more objects), it may be worth implementing this more involved variant. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-03-13 22:22:01 +00:00
*/
diffcore-rename: determine which relevant_sources are no longer relevant As noted a few commits ago ("diffcore-rename: only compute dir_rename_count for relevant directories"), when a source file rename is used as part of directory rename detection, we need to increment counts for each ancestor directory in dirs_removed with value RELEVANT_FOR_SELF. However, a few commits ago ("diffcore-rename: check if we have enough renames for directories early on"), we may have downgraded all relevant ancestor directories from RELEVANT_FOR_SELF to RELEVANT_FOR_ANCESTOR. For a given file, if no ancestor directory is found in dirs_removed with a value of RELEVANT_FOR_SELF, then we can downgrade relevant_source[PATH] from RELEVANT_LOCATION to RELEVANT_NO_MORE. This means we can skip detecting a rename for that particular path (and any other paths in the same directory). For the testcases mentioned in commit 557ac0350d ("merge-ort: begin performance work; instrument with trace2_region_* calls", 2020-10-28), this change improves the performance as follows: Before After no-renames: 5.680 s ± 0.096 s 5.665 s ± 0.129 s mega-renames: 13.812 s ± 0.162 s 11.435 s ± 0.158 s just-one-mega: 506.0 ms ± 3.9 ms 494.2 ms ± 6.1 ms While this improvement looks rather modest for these testcases (because all the previous optimizations were sufficient to nearly remove all time spent in rename detection already), consider this alternative testcase tweaked from the ones in commit 557ac0350d as follows <Same initial setup as commit 557ac0350d, then...> $ git switch -c add-empty-file v5.5 $ >drivers/gpu/drm/i915/new-empty-file $ git add drivers/gpu/drm/i915/new-empty-file $ git commit -m "new file" $ git switch 5.4-rename $ git cherry-pick --strategy=ort add-empty-file For this testcase, we see the following improvement: Before After pick-empty: 1.936 s ± 0.024 s 688.1 ms ± 4.2 ms So roughly a factor of 3 speedup. At $DAYJOB, there was a particular repository and cherry-pick that inspired this optimization; for that case I saw a speedup factor of 7 with this optimization. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-03-13 22:22:08 +00:00
int i, new_num_src;
struct hashmap_iter iter;
struct strmap_entry *entry;
if (!dirs_removed || !relevant_sources)
return; /* nothing to cull */
if (break_idx)
return; /* culling incompatbile with break detection */
/*
* Supplement dir_rename_count with number of potential renames,
* marking all potential rename sources as mapping to UNKNOWN_DIR.
*/
for (i = 0; i < rename_src_nr; i++) {
char *old_dir;
struct diff_filespec *one = rename_src[i].p->one;
/*
* sources that are part of a rename will have already been
* removed by a prior call to remove_unneeded_paths_from_src()
*/
assert(!one->rename_used);
old_dir = get_dirname(one->path);
while (*old_dir != '\0' &&
NOT_RELEVANT != strintmap_get(dirs_removed, old_dir)) {
char *freeme = old_dir;
increment_count(info, old_dir, UNKNOWN_DIR);
old_dir = get_dirname(old_dir);
/* Free resources we don't need anymore */
free(freeme);
}
/*
* old_dir and new_dir free'd in increment_count, but
* get_dirname() gives us a new pointer we need to free for
* old_dir. Also, if the loop runs 0 times we need old_dir
* to be freed.
*/
free(old_dir);
}
/*
* For any directory which we need a potential rename detected for
* (i.e. those marked as RELEVANT_FOR_SELF in dirs_removed), check
* whether we have enough renames to satisfy the "majority rules"
* requirement such that detecting any more renames of files under
* it won't change the result. For any such directory, mark that
* we no longer need to detect a rename for it. However, since we
* might need to still detect renames for an ancestor of that
* directory, use RELEVANT_FOR_ANCESTOR.
*/
strmap_for_each_entry(info->dir_rename_count, &iter, entry) {
/* entry->key is source_dir */
struct strintmap *counts = entry->value;
if (strintmap_get(dirs_removed, entry->key) ==
RELEVANT_FOR_SELF &&
dir_rename_already_determinable(counts)) {
strintmap_set(dirs_removed, entry->key,
RELEVANT_FOR_ANCESTOR);
}
}
diffcore-rename: determine which relevant_sources are no longer relevant As noted a few commits ago ("diffcore-rename: only compute dir_rename_count for relevant directories"), when a source file rename is used as part of directory rename detection, we need to increment counts for each ancestor directory in dirs_removed with value RELEVANT_FOR_SELF. However, a few commits ago ("diffcore-rename: check if we have enough renames for directories early on"), we may have downgraded all relevant ancestor directories from RELEVANT_FOR_SELF to RELEVANT_FOR_ANCESTOR. For a given file, if no ancestor directory is found in dirs_removed with a value of RELEVANT_FOR_SELF, then we can downgrade relevant_source[PATH] from RELEVANT_LOCATION to RELEVANT_NO_MORE. This means we can skip detecting a rename for that particular path (and any other paths in the same directory). For the testcases mentioned in commit 557ac0350d ("merge-ort: begin performance work; instrument with trace2_region_* calls", 2020-10-28), this change improves the performance as follows: Before After no-renames: 5.680 s ± 0.096 s 5.665 s ± 0.129 s mega-renames: 13.812 s ± 0.162 s 11.435 s ± 0.158 s just-one-mega: 506.0 ms ± 3.9 ms 494.2 ms ± 6.1 ms While this improvement looks rather modest for these testcases (because all the previous optimizations were sufficient to nearly remove all time spent in rename detection already), consider this alternative testcase tweaked from the ones in commit 557ac0350d as follows <Same initial setup as commit 557ac0350d, then...> $ git switch -c add-empty-file v5.5 $ >drivers/gpu/drm/i915/new-empty-file $ git add drivers/gpu/drm/i915/new-empty-file $ git commit -m "new file" $ git switch 5.4-rename $ git cherry-pick --strategy=ort add-empty-file For this testcase, we see the following improvement: Before After pick-empty: 1.936 s ± 0.024 s 688.1 ms ± 4.2 ms So roughly a factor of 3 speedup. At $DAYJOB, there was a particular repository and cherry-pick that inspired this optimization; for that case I saw a speedup factor of 7 with this optimization. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-03-13 22:22:08 +00:00
for (i = 0, new_num_src = 0; i < rename_src_nr; i++) {
struct diff_filespec *one = rename_src[i].p->one;
int val;
val = strintmap_get(relevant_sources, one->path);
/*
* sources that were not found in relevant_sources should
* have already been removed by a prior call to
* remove_unneeded_paths_from_src()
*/
assert(val != -1);
if (val == RELEVANT_LOCATION) {
int removable = 1;
char *dir = get_dirname(one->path);
while (1) {
char *freeme = dir;
int res = strintmap_get(dirs_removed, dir);
/* Quit if not found or irrelevant */
if (res == NOT_RELEVANT)
break;
/* If RELEVANT_FOR_SELF, can't remove */
if (res == RELEVANT_FOR_SELF) {
removable = 0;
break;
}
/* Else continue searching upwards */
assert(res == RELEVANT_FOR_ANCESTOR);
dir = get_dirname(dir);
free(freeme);
}
free(dir);
if (removable) {
strintmap_set(relevant_sources, one->path,
RELEVANT_NO_MORE);
continue;
}
}
if (new_num_src < i)
memcpy(&rename_src[new_num_src], &rename_src[i],
sizeof(struct diff_rename_src));
new_num_src++;
}
rename_src_nr = new_num_src;
diffcore-rename: take advantage of "majority rules" to skip more renames In directory rename detection (when a directory is removed on one side of history and the other side adds new files to that directory), we work to find where the greatest number of files within that directory were renamed to so that the new files can be moved with the majority of the files. Naively, we can just do this by detecting renames for *all* files within the removed/renamed directory, looking at all the destination directories where files within that directory were moved, and if there is more than one such directory then taking the one with the greatest number of files as the directory where the old directory was renamed to. However, sometimes there are enough renames from exact rename detection or basename-guided rename detection that we have enough information to determine the majority winner already. Add a function meant to compute whether particular renames are still needed based on this majority rules check. The next several commits will then add the necessary infrastructure to get the information we need to compute which additional rename sources we can skip. An important side note for future further optimization: There is a possible improvement to this optimization that I have not yet attempted and will not be included in this series of patches: we could first check whether exact renames provide enough information for us to determine directory renames, and avoid doing basename-guided rename detection on some or all of the RELEVANT_LOCATION files within those directories. In effect, this variant would mean doing the handle_early_known_dir_renames() both after exact rename detection and again after basename-guided rename detection, though it would also mean decrementing the number of "unknown" renames for each rename we found from basename-guided rename detection. Adding this additional check for skippable renames right after exact rename detection might turn out to be valuable, especially for partial clones where it might allow us to download certain source files entirely. However, this particular optimization was actually the last one I did in original implementation order, and by the time I implemented this idea, every testcase I had was sufficiently fast that further optimization was unwarranted. If future testcases arise that tax rename detection more heavily (or perhaps partial clones can benefit from avoiding loading more objects), it may be worth implementing this more involved variant. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-03-13 22:22:01 +00:00
}
static void free_filespec_data(struct diff_filespec *spec)
{
if (!--spec->count)
diff_free_filespec_data(spec);
}
static void pool_free_filespec(struct mem_pool *pool,
struct diff_filespec *spec)
{
if (!pool) {
free_filespec(spec);
return;
}
/*
* Similar to free_filespec(), but only frees the data. The spec
* itself was allocated in the pool and should not be individually
* freed.
*/
free_filespec_data(spec);
}
void pool_diff_free_filepair(struct mem_pool *pool,
struct diff_filepair *p)
{
if (!pool) {
diff_free_filepair(p);
return;
}
/*
* Similar to diff_free_filepair() but only frees the data from the
* filespecs; not the filespecs or the filepair which were
* allocated from the pool.
*/
free_filespec_data(p->one);
free_filespec_data(p->two);
}
void diffcore_rename_extended(struct diff_options *options,
struct mem_pool *pool,
struct strintmap *relevant_sources,
struct strintmap *dirs_removed,
merge-ort, diffcore-rename: employ cached renames when possible When there are many renames between the old base of a series of commits and the new base, the way sequencer.c, merge-recursive.c, and diffcore-rename.c have traditionally split the work resulted in redetecting the same renames with each and every commit being transplanted. To address this, the last several commits have been creating a cache of rename detection results, determining when it was safe to use such a cache in subsequent merge operations, adding helper functions, and so on. See the previous half dozen commit messages for additional discussion of this optimization, particularly the message a few commits ago entitled "add code to check for whether cached renames can be reused". This commit finally ties all of that work together, modifying the merge algorithm to make use of these cached renames. For the testcases mentioned in commit 557ac0350d ("merge-ort: begin performance work; instrument with trace2_region_* calls", 2020-10-28), this change improves the performance as follows: Before After no-renames: 5.665 s ± 0.129 s 5.622 s ± 0.059 s mega-renames: 11.435 s ± 0.158 s 10.127 s ± 0.073 s just-one-mega: 494.2 ms ± 6.1 ms 500.3 ms ± 3.8 ms That's a fairly small improvement, but mostly because the previous optimizations were so effective for these particular testcases; this optimization only kicks in when the others don't. If we undid the basename-guided rename detection and skip-irrelevant-renames optimizations, then we'd see that this series by itself improved performance as follows: Before Basename Series After Just This Series no-renames: 13.815 s ± 0.062 s 5.697 s ± 0.080 s mega-renames: 1799.937 s ± 0.493 s 205.709 s ± 0.457 s Since this optimization kicks in to help accelerate cases where the previous optimizations do not apply, this last comparison shows that this cached-renames optimization has the potential to help signficantly in cases that don't meet the requirements for the other optimizations to be effective. The changes made in this optimization also lay some important groundwork for a future optimization around having collect_merge_info() avoid recursing into subtrees in more cases. However, for this optimization to be effective, merge_switch_to_result() should only be called when the rebase or cherry-pick operation has either completed or hit a case where the user needs to resolve a conflict or edit the result. If it is called after every commit, as sequencer.c does, then the working tree and index are needlessly updated with every commit and the cached metadata is tossed, defeating this optimization. Refactoring sequencer.c to only call merge_switch_to_result() at the end of the operation is a bigger undertaking, and the practical benefits of this optimization will not be realized until that work is performed. Since `test-tool fast-rebase` only updates at the end of the operation, it was used to obtain the timings above. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-05-20 06:09:41 +00:00
struct strmap *dir_rename_count,
struct strmap *cached_pairs)
{
int detect_rename = options->detect_rename;
int minimum_score = options->rename_score;
struct diff_queue_struct *q = &diff_queued_diff;
struct diff_queue_struct outq;
struct diff_score *mx;
int i, j, rename_count, skip_unmodified = 0;
int num_destinations, dst_cnt;
int num_sources, want_copies;
struct progress *progress = NULL;
struct mem_pool local_pool;
struct dir_rename_info info;
struct diff_populate_filespec_options dpf_options = {
.check_binary = 0,
.missing_object_cb = NULL,
.missing_object_data = NULL
};
diffcore-rename: use a different prefetch for basename comparisons merge-ort was designed to minimize the amount of data needed and used, and several changes were made to diffcore-rename to take advantage of extra metadata to enable this data minimization (particularly the relevant_sources variable for skipping "irrelevant" renames). This effort obviously succeeded in drastically reducing computation times, but should also theoretically allow partial clones to download much less information. Previously, though, the "prefetch" command used in diffcore-rename had never been modified and downloaded many blobs that were unnecessary for merge-ort. This commit corrects that. When doing basename comparisons, we want to fetch only the objects that will be used for basename comparisons. If after basename fetching this leaves us with no more relevant sources (or no more destinations), then we won't need to do the full inexact rename detection and can skip downloading additional source and destination files. Even if we have to do that later full inexact rename detection, irrelevant sources are culled after basename matching and before the full inexact rename detection, so we can still avoid downloading the blobs for irrelevant sources. Rename prefetch() to inexact_prefetch(), and introduce a new basename_prefetch() to take advantage of this. If we modify the testcase from commit 557ac0350d ("merge-ort: begin performance work; instrument with trace2_region_* calls", 2021-01-23) to pass --sparse --filter=blob:none to the clone command, and use the new trace2 "fetch_count" output from a few commits ago to track both the number of fetch subcommands invoked and the number of objects fetched across all those fetches, then for the mega-renames testcase we observe the following: BEFORE this commit, rebasing 35 patches: strategy # of fetches total # of objects fetched --------- ------------ -------------------------- recursive 62 11423 ort 30 11391 AFTER this commit, rebasing the same 35 patches: ort 32 63 This means that the new code only needs to download less than 2 blobs per patch being rebased. That is especially interesting given that the repository at the start only had approximately half a dozen TOTAL blobs downloaded to start with (because the default sparse-checkout of just the toplevel directory was in use). So, for this particular linux kernel testcase that involved ~26,000 renames on the upstream side (drivers/ -> pilots/) across which 35 patches were being rebased, this change reduces the number of blobs that need to be downloaded by a factor of ~180. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-06-22 08:04:40 +00:00
struct inexact_prefetch_options prefetch_options = {
.repo = options->repo
};
merge-ort: begin performance work; instrument with trace2_region_* calls Add some timing instrumentation for both merge-ort and diffcore-rename; I used these to measure and optimize performance in both, and several future patch series will build on these to reduce the timings of some select testcases. === Setup === The primary testcase I used involved rebasing a random topic in the linux kernel (consisting of 35 patches) against an older version. I added two variants, one where I rename a toplevel directory, and another where I only rebase one patch instead of the whole topic. The setup is as follows: $ git clone git://git.kernel.org/pub/scm/linux/kernel/git/stable/linux-stable.git $ git branch hwmon-updates fd8bdb23b91876ac1e624337bb88dc1dcc21d67e $ git branch hwmon-just-one fd8bdb23b91876ac1e624337bb88dc1dcc21d67e~34 $ git branch base 4703d9119972bf586d2cca76ec6438f819ffa30e $ git switch -c 5.4-renames v5.4 $ git mv drivers pilots # Introduce over 26,000 renames $ git commit -m "Rename drivers/ to pilots/" $ git config merge.renameLimit 30000 $ git config merge.directoryRenames true === Testcases === Now with REBASE standing for either "git rebase [--merge]" (using merge-recursive) or "test-tool fast-rebase" (using merge-ort), the testcases are: Testcase #1: no-renames $ git checkout v5.4^0 $ REBASE --onto HEAD base hwmon-updates Note: technically the name is misleading; there are some renames, but very few. Rename detection only takes about half the overall time. Testcase #2: mega-renames $ git checkout 5.4-renames^0 $ REBASE --onto HEAD base hwmon-updates Testcase #3: just-one-mega $ git checkout 5.4-renames^0 $ REBASE --onto HEAD base hwmon-just-one === Timing results === Overall timings, using hyperfine (1 warmup run, 3 runs for mega-renames, 10 runs for the other two cases): merge-recursive merge-ort no-renames: 18.912 s ± 0.174 s 14.263 s ± 0.053 s mega-renames: 5964.031 s ± 10.459 s 5504.231 s ± 5.150 s just-one-mega: 149.583 s ± 0.751 s 158.534 s ± 0.498 s A single re-run of each with some breakdowns: --- no-renames --- merge-recursive merge-ort overall runtime: 19.302 s 14.257 s inexact rename detection: 7.603 s 7.906 s everything else: 11.699 s 6.351 s --- mega-renames --- merge-recursive merge-ort overall runtime: 5950.195 s 5499.672 s inexact rename detection: 5746.309 s 5487.120 s everything else: 203.886 s 17.552 s --- just-one-mega --- merge-recursive merge-ort overall runtime: 151.001 s 158.582 s inexact rename detection: 143.448 s 157.835 s everything else: 7.553 s 0.747 s === Timing observations === 0) Maximum speedup The "everything else" row represents the maximum speedup we could achieve if we were to somehow infinitely parallelize inexact rename detection, but leave everything else alone. The fact that this is so much smaller than the real runtime (even in the case with virtually no renames) makes it clear just how overwhelmingly large the time spent on rename detection can be. 1) no-renames 1a) merge-ort is faster than merge-recursive, which is nice. However, this still should not be considered good enough. Although the "merge" backend to rebase (merge-recursive) is sometimes faster than the "apply" backend, this is one of those cases where it is not. In fact, even merge-ort is slower. The "apply" backend can complete this testcase in 6.940 s ± 0.485 s which is about 2x faster than merge-ort and 3x faster than merge-recursive. One goal of the merge-ort performance work will be to make it faster than git-am on this (and similar) testcases. 2) mega-renames 2a) Obviously rename detection is a huge cost; it's where most the time is spent. We need to cut that down. If we could somehow infinitely parallelize it and drive its time to 0, the merge-recursive time would drop to about 204s, and the merge-ort time would drop to about 17s. I think this particular stat shows I've subtly baked a couple performance improvements into merge-ort and into fast-rebase already. 3) just-one-mega 3a) not much to say here, it just gives some flavor for how rebasing only one patch compares to rebasing 35. === Goals === This patch is obviously just the beginning. Here are some of my goals that this measurement will help us achieve: * Drive the cost of rename detection down considerably for merges * After the above has been achieved, see if there are other slowness factors (which would have previously been overshadowed by rename detection costs) which we can then focus on and also optimize. * Ensure our rebase testcase that requires little rename detection is noticeably faster with merge-ort than with apply-based rebase. Signed-off-by: Elijah Newren <newren@gmail.com> Acked-by: Taylor Blau <ttaylorr@github.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-01-24 06:01:12 +00:00
trace2_region_enter("diff", "setup", options->repo);
info.setup = 0;
assert(!dir_rename_count || strmap_empty(dir_rename_count));
want_copies = (detect_rename == DIFF_DETECT_COPY);
if (dirs_removed && (break_idx || want_copies))
BUG("dirs_removed incompatible with break/copy detection");
diffcore-rename: enable filtering possible rename sources Add the ability to diffcore_rename_extended() to allow external callers to declare that they only need renames detected for a subset of source files, and use that information to skip detecting renames for them. There are two important pieces to this optimization that may not be obvious at first glance: * We do not require callers to just filter the filepairs out to remove the non-relevant sources, because exact rename detection is fast and when it finds a match it can remove both a source and a destination whereas the relevant_sources filter can only remove a source. * We need to filter out the source pairs in a preliminary pass instead of adding a strset_contains(relevant_sources, one->path) check within the nested matrix loop. The reason for that is if we have 30k renames, doing 30k * 30k = 900M strset_contains() calls becomes extraordinarily expensive and defeats the performance gains from this change; we only want to do 30k such calls instead. If callers pass NULL for relevant_sources, that is special cases to treat all sources as relevant. Since all callers currently pass NULL, this optimization does not yet have any effect. Subsequent commits will have merge-ort compute a set of relevant_sources to restrict which sources we detect renames for, and have merge-ort pass that set of relevant_sources to diffcore_rename_extended(). A note about filtering order: Some may be curious why we don't filter out irrelevant sources at the same time we filter out exact renames. While that technically could be done at this point, there are two reasons to defer it: First, was to reinforce a lesson that was too easy to forget. As I mentioned above, in the past I filtered irrelevant sources out before exact rename checking, and then discovered that exact renames' ability to remove both sources and destinations was an important consideration and thus doing the filtering after exact rename checking would speed things up. Then at some point I realized that basename matching could also remove both sources and destinations, and decided to put irrelevant source filtering after basename filtering. That slowed things down a lot. But, despite learning about this important ordering, in later restructuring I forgot and made the same mistake of putting the filtering after basename guided rename detection again. So, I have this series of patches structured to do the irrelevant filtering last to start to show how much extra it costs, and then add relevant filtering in to find_basename_matches() to show how much it speeds things up. Basically, it's a way to reinforce something that apparently was too easy to forget, and make sure the commit messages record this lesson. Second, the items in the "relevant_sources" in this patch series will include all sources that *might be* relevant. It has to be conservative and catch anything that might need a rename, but in the patch series after this one we'll find ways to weed out more of the *might be* relevant ones. Unfortunately, merge-ort does not have sufficient information to weed those ones out, and there isn't enough information at the time of filtering exact renames out to remove the extra ones either. It has to be deferred. So the deferral is in part to simplify some later additions. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-03-11 00:38:24 +00:00
if (break_idx && relevant_sources)
BUG("break detection incompatible with source specification");
if (!minimum_score)
minimum_score = DEFAULT_RENAME_SCORE;
for (i = 0; i < q->nr; i++) {
struct diff_filepair *p = q->queue[i];
if (!DIFF_FILE_VALID(p->one)) {
if (!DIFF_FILE_VALID(p->two))
continue; /* unmerged */
else if (options->single_follow &&
strcmp(options->single_follow, p->two->path))
continue; /* not interested */
diff: make struct diff_flags members lowercase Now that the flags stored in struct diff_flags are being accessed directly and not through macros, change all struct members from being uppercase to lowercase. This conversion is done using the following semantic patch: @@ expression E; @@ - E.RECURSIVE + E.recursive @@ expression E; @@ - E.TREE_IN_RECURSIVE + E.tree_in_recursive @@ expression E; @@ - E.BINARY + E.binary @@ expression E; @@ - E.TEXT + E.text @@ expression E; @@ - E.FULL_INDEX + E.full_index @@ expression E; @@ - E.SILENT_ON_REMOVE + E.silent_on_remove @@ expression E; @@ - E.FIND_COPIES_HARDER + E.find_copies_harder @@ expression E; @@ - E.FOLLOW_RENAMES + E.follow_renames @@ expression E; @@ - E.RENAME_EMPTY + E.rename_empty @@ expression E; @@ - E.HAS_CHANGES + E.has_changes @@ expression E; @@ - E.QUICK + E.quick @@ expression E; @@ - E.NO_INDEX + E.no_index @@ expression E; @@ - E.ALLOW_EXTERNAL + E.allow_external @@ expression E; @@ - E.EXIT_WITH_STATUS + E.exit_with_status @@ expression E; @@ - E.REVERSE_DIFF + E.reverse_diff @@ expression E; @@ - E.CHECK_FAILED + E.check_failed @@ expression E; @@ - E.RELATIVE_NAME + E.relative_name @@ expression E; @@ - E.IGNORE_SUBMODULES + E.ignore_submodules @@ expression E; @@ - E.DIRSTAT_CUMULATIVE + E.dirstat_cumulative @@ expression E; @@ - E.DIRSTAT_BY_FILE + E.dirstat_by_file @@ expression E; @@ - E.ALLOW_TEXTCONV + E.allow_textconv @@ expression E; @@ - E.TEXTCONV_SET_VIA_CMDLINE + E.textconv_set_via_cmdline @@ expression E; @@ - E.DIFF_FROM_CONTENTS + E.diff_from_contents @@ expression E; @@ - E.DIRTY_SUBMODULES + E.dirty_submodules @@ expression E; @@ - E.IGNORE_UNTRACKED_IN_SUBMODULES + E.ignore_untracked_in_submodules @@ expression E; @@ - E.IGNORE_DIRTY_SUBMODULES + E.ignore_dirty_submodules @@ expression E; @@ - E.OVERRIDE_SUBMODULE_CONFIG + E.override_submodule_config @@ expression E; @@ - E.DIRSTAT_BY_LINE + E.dirstat_by_line @@ expression E; @@ - E.FUNCCONTEXT + E.funccontext @@ expression E; @@ - E.PICKAXE_IGNORE_CASE + E.pickaxe_ignore_case @@ expression E; @@ - E.DEFAULT_FOLLOW_RENAMES + E.default_follow_renames Signed-off-by: Brandon Williams <bmwill@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-10-31 18:19:11 +00:00
else if (!options->flags.rename_empty &&
is_empty_blob_oid(&p->two->oid))
teach diffcore-rename to optionally ignore empty content Our rename detection is a heuristic, matching pairs of removed and added files with similar or identical content. It's unlikely to be wrong when there is actual content to compare, and we already take care not to do inexact rename detection when there is not enough content to produce good results. However, we always do exact rename detection, even when the blob is tiny or empty. It's easy to get false positives with an empty blob, simply because it is an obvious content to use as a boilerplate (e.g., when telling git that an empty directory is worth tracking via an empty .gitignore). This patch lets callers specify whether or not they are interested in using empty files as rename sources and destinations. The default is "yes", keeping the original behavior. It works by detecting the empty-blob sha1 for rename sources and destinations. One more flexible alternative would be to allow the caller to specify a minimum size for a blob to be "interesting" for rename detection. But that would catch small boilerplate files, not large ones (e.g., if you had the GPL COPYING file in many directories). A better alternative would be to allow a "-rename" gitattribute to allow boilerplate files to be marked as such. I'll leave the complexity of that solution until such time as somebody actually wants it. The complaints we've seen so far revolve around empty files, so let's start with the simple thing. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2012-03-22 22:52:13 +00:00
continue;
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 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. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-11 09:08:47 +00:00
else if (add_rename_dst(p) < 0) {
warning("skipping rename detection, detected"
" duplicate destination '%s'",
p->two->path);
goto cleanup;
}
}
diff: make struct diff_flags members lowercase Now that the flags stored in struct diff_flags are being accessed directly and not through macros, change all struct members from being uppercase to lowercase. This conversion is done using the following semantic patch: @@ expression E; @@ - E.RECURSIVE + E.recursive @@ expression E; @@ - E.TREE_IN_RECURSIVE + E.tree_in_recursive @@ expression E; @@ - E.BINARY + E.binary @@ expression E; @@ - E.TEXT + E.text @@ expression E; @@ - E.FULL_INDEX + E.full_index @@ expression E; @@ - E.SILENT_ON_REMOVE + E.silent_on_remove @@ expression E; @@ - E.FIND_COPIES_HARDER + E.find_copies_harder @@ expression E; @@ - E.FOLLOW_RENAMES + E.follow_renames @@ expression E; @@ - E.RENAME_EMPTY + E.rename_empty @@ expression E; @@ - E.HAS_CHANGES + E.has_changes @@ expression E; @@ - E.QUICK + E.quick @@ expression E; @@ - E.NO_INDEX + E.no_index @@ expression E; @@ - E.ALLOW_EXTERNAL + E.allow_external @@ expression E; @@ - E.EXIT_WITH_STATUS + E.exit_with_status @@ expression E; @@ - E.REVERSE_DIFF + E.reverse_diff @@ expression E; @@ - E.CHECK_FAILED + E.check_failed @@ expression E; @@ - E.RELATIVE_NAME + E.relative_name @@ expression E; @@ - E.IGNORE_SUBMODULES + E.ignore_submodules @@ expression E; @@ - E.DIRSTAT_CUMULATIVE + E.dirstat_cumulative @@ expression E; @@ - E.DIRSTAT_BY_FILE + E.dirstat_by_file @@ expression E; @@ - E.ALLOW_TEXTCONV + E.allow_textconv @@ expression E; @@ - E.TEXTCONV_SET_VIA_CMDLINE + E.textconv_set_via_cmdline @@ expression E; @@ - E.DIFF_FROM_CONTENTS + E.diff_from_contents @@ expression E; @@ - E.DIRTY_SUBMODULES + E.dirty_submodules @@ expression E; @@ - E.IGNORE_UNTRACKED_IN_SUBMODULES + E.ignore_untracked_in_submodules @@ expression E; @@ - E.IGNORE_DIRTY_SUBMODULES + E.ignore_dirty_submodules @@ expression E; @@ - E.OVERRIDE_SUBMODULE_CONFIG + E.override_submodule_config @@ expression E; @@ - E.DIRSTAT_BY_LINE + E.dirstat_by_line @@ expression E; @@ - E.FUNCCONTEXT + E.funccontext @@ expression E; @@ - E.PICKAXE_IGNORE_CASE + E.pickaxe_ignore_case @@ expression E; @@ - E.DEFAULT_FOLLOW_RENAMES + E.default_follow_renames Signed-off-by: Brandon Williams <bmwill@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-10-31 18:19:11 +00:00
else if (!options->flags.rename_empty &&
is_empty_blob_oid(&p->one->oid))
teach diffcore-rename to optionally ignore empty content Our rename detection is a heuristic, matching pairs of removed and added files with similar or identical content. It's unlikely to be wrong when there is actual content to compare, and we already take care not to do inexact rename detection when there is not enough content to produce good results. However, we always do exact rename detection, even when the blob is tiny or empty. It's easy to get false positives with an empty blob, simply because it is an obvious content to use as a boilerplate (e.g., when telling git that an empty directory is worth tracking via an empty .gitignore). This patch lets callers specify whether or not they are interested in using empty files as rename sources and destinations. The default is "yes", keeping the original behavior. It works by detecting the empty-blob sha1 for rename sources and destinations. One more flexible alternative would be to allow the caller to specify a minimum size for a blob to be "interesting" for rename detection. But that would catch small boilerplate files, not large ones (e.g., if you had the GPL COPYING file in many directories). A better alternative would be to allow a "-rename" gitattribute to allow boilerplate files to be marked as such. I'll leave the complexity of that solution until such time as somebody actually wants it. The complaints we've seen so far revolve around empty files, so let's start with the simple thing. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2012-03-22 22:52:13 +00:00
continue;
diffcore-rename: don't consider unmerged path as source Since e9c8409 (diff-index --cached --raw: show tree entry on the LHS for unmerged entries., 2007-01-05), an unmerged entry should be detected by using DIFF_PAIR_UNMERGED(p), not by noticing both one and two sides of the filepair records mode=0 entries. However, it forgot to update some parts of the rename detection logic. This only makes difference in the "diff --cached" codepath where an unmerged filepair carries information on the entries that came from the tree. It probably hasn't been noticed for a long time because nobody would run "diff -M" during a conflict resolution, but "git status" uses rename detection when it internally runs "diff-index" and "diff-files" and gives nonsense results. In an unmerged pair, "one" side can have a valid filespec to record the tree entry (e.g. what's in HEAD) when running "diff --cached". This can be used as a rename source to other paths in the index that are not unmerged. The path that is unmerged by definition does not have the final content yet (i.e. "two" side cannot have a valid filespec), so it can never be a rename destination. Use the DIFF_PAIR_UNMERGED() to detect unmerged filepair correctly, and allow the valid "one" side of an unmerged filepair to be considered a potential rename source, but never to be considered a rename destination. Commit message and first two test cases by Junio, the rest by Martin. Signed-off-by: Martin von Zweigbergk <martin.von.zweigbergk@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2011-03-24 02:41:01 +00:00
else if (!DIFF_PAIR_UNMERGED(p) && !DIFF_FILE_VALID(p->two)) {
/*
* If the source is a broken "delete", and
* they did not really want to get broken,
* that means the source actually stays.
* So we increment the "rename_used" score
* by one, to indicate ourselves as a user
*/
if (p->broken_pair && !p->score)
p->one->rename_used++;
register_rename_src(p);
}
else if (want_copies) {
/*
* Increment the "rename_used" score by
* one, to indicate ourselves as a user.
*/
p->one->rename_used++;
register_rename_src(p);
}
}
merge-ort: begin performance work; instrument with trace2_region_* calls Add some timing instrumentation for both merge-ort and diffcore-rename; I used these to measure and optimize performance in both, and several future patch series will build on these to reduce the timings of some select testcases. === Setup === The primary testcase I used involved rebasing a random topic in the linux kernel (consisting of 35 patches) against an older version. I added two variants, one where I rename a toplevel directory, and another where I only rebase one patch instead of the whole topic. The setup is as follows: $ git clone git://git.kernel.org/pub/scm/linux/kernel/git/stable/linux-stable.git $ git branch hwmon-updates fd8bdb23b91876ac1e624337bb88dc1dcc21d67e $ git branch hwmon-just-one fd8bdb23b91876ac1e624337bb88dc1dcc21d67e~34 $ git branch base 4703d9119972bf586d2cca76ec6438f819ffa30e $ git switch -c 5.4-renames v5.4 $ git mv drivers pilots # Introduce over 26,000 renames $ git commit -m "Rename drivers/ to pilots/" $ git config merge.renameLimit 30000 $ git config merge.directoryRenames true === Testcases === Now with REBASE standing for either "git rebase [--merge]" (using merge-recursive) or "test-tool fast-rebase" (using merge-ort), the testcases are: Testcase #1: no-renames $ git checkout v5.4^0 $ REBASE --onto HEAD base hwmon-updates Note: technically the name is misleading; there are some renames, but very few. Rename detection only takes about half the overall time. Testcase #2: mega-renames $ git checkout 5.4-renames^0 $ REBASE --onto HEAD base hwmon-updates Testcase #3: just-one-mega $ git checkout 5.4-renames^0 $ REBASE --onto HEAD base hwmon-just-one === Timing results === Overall timings, using hyperfine (1 warmup run, 3 runs for mega-renames, 10 runs for the other two cases): merge-recursive merge-ort no-renames: 18.912 s ± 0.174 s 14.263 s ± 0.053 s mega-renames: 5964.031 s ± 10.459 s 5504.231 s ± 5.150 s just-one-mega: 149.583 s ± 0.751 s 158.534 s ± 0.498 s A single re-run of each with some breakdowns: --- no-renames --- merge-recursive merge-ort overall runtime: 19.302 s 14.257 s inexact rename detection: 7.603 s 7.906 s everything else: 11.699 s 6.351 s --- mega-renames --- merge-recursive merge-ort overall runtime: 5950.195 s 5499.672 s inexact rename detection: 5746.309 s 5487.120 s everything else: 203.886 s 17.552 s --- just-one-mega --- merge-recursive merge-ort overall runtime: 151.001 s 158.582 s inexact rename detection: 143.448 s 157.835 s everything else: 7.553 s 0.747 s === Timing observations === 0) Maximum speedup The "everything else" row represents the maximum speedup we could achieve if we were to somehow infinitely parallelize inexact rename detection, but leave everything else alone. The fact that this is so much smaller than the real runtime (even in the case with virtually no renames) makes it clear just how overwhelmingly large the time spent on rename detection can be. 1) no-renames 1a) merge-ort is faster than merge-recursive, which is nice. However, this still should not be considered good enough. Although the "merge" backend to rebase (merge-recursive) is sometimes faster than the "apply" backend, this is one of those cases where it is not. In fact, even merge-ort is slower. The "apply" backend can complete this testcase in 6.940 s ± 0.485 s which is about 2x faster than merge-ort and 3x faster than merge-recursive. One goal of the merge-ort performance work will be to make it faster than git-am on this (and similar) testcases. 2) mega-renames 2a) Obviously rename detection is a huge cost; it's where most the time is spent. We need to cut that down. If we could somehow infinitely parallelize it and drive its time to 0, the merge-recursive time would drop to about 204s, and the merge-ort time would drop to about 17s. I think this particular stat shows I've subtly baked a couple performance improvements into merge-ort and into fast-rebase already. 3) just-one-mega 3a) not much to say here, it just gives some flavor for how rebasing only one patch compares to rebasing 35. === Goals === This patch is obviously just the beginning. Here are some of my goals that this measurement will help us achieve: * Drive the cost of rename detection down considerably for merges * After the above has been achieved, see if there are other slowness factors (which would have previously been overshadowed by rename detection costs) which we can then focus on and also optimize. * Ensure our rebase testcase that requires little rename detection is noticeably faster with merge-ort than with apply-based rebase. Signed-off-by: Elijah Newren <newren@gmail.com> Acked-by: Taylor Blau <ttaylorr@github.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-01-24 06:01:12 +00:00
trace2_region_leave("diff", "setup", options->repo);
if (rename_dst_nr == 0 || rename_src_nr == 0)
goto cleanup; /* nothing to do */
merge-ort: begin performance work; instrument with trace2_region_* calls Add some timing instrumentation for both merge-ort and diffcore-rename; I used these to measure and optimize performance in both, and several future patch series will build on these to reduce the timings of some select testcases. === Setup === The primary testcase I used involved rebasing a random topic in the linux kernel (consisting of 35 patches) against an older version. I added two variants, one where I rename a toplevel directory, and another where I only rebase one patch instead of the whole topic. The setup is as follows: $ git clone git://git.kernel.org/pub/scm/linux/kernel/git/stable/linux-stable.git $ git branch hwmon-updates fd8bdb23b91876ac1e624337bb88dc1dcc21d67e $ git branch hwmon-just-one fd8bdb23b91876ac1e624337bb88dc1dcc21d67e~34 $ git branch base 4703d9119972bf586d2cca76ec6438f819ffa30e $ git switch -c 5.4-renames v5.4 $ git mv drivers pilots # Introduce over 26,000 renames $ git commit -m "Rename drivers/ to pilots/" $ git config merge.renameLimit 30000 $ git config merge.directoryRenames true === Testcases === Now with REBASE standing for either "git rebase [--merge]" (using merge-recursive) or "test-tool fast-rebase" (using merge-ort), the testcases are: Testcase #1: no-renames $ git checkout v5.4^0 $ REBASE --onto HEAD base hwmon-updates Note: technically the name is misleading; there are some renames, but very few. Rename detection only takes about half the overall time. Testcase #2: mega-renames $ git checkout 5.4-renames^0 $ REBASE --onto HEAD base hwmon-updates Testcase #3: just-one-mega $ git checkout 5.4-renames^0 $ REBASE --onto HEAD base hwmon-just-one === Timing results === Overall timings, using hyperfine (1 warmup run, 3 runs for mega-renames, 10 runs for the other two cases): merge-recursive merge-ort no-renames: 18.912 s ± 0.174 s 14.263 s ± 0.053 s mega-renames: 5964.031 s ± 10.459 s 5504.231 s ± 5.150 s just-one-mega: 149.583 s ± 0.751 s 158.534 s ± 0.498 s A single re-run of each with some breakdowns: --- no-renames --- merge-recursive merge-ort overall runtime: 19.302 s 14.257 s inexact rename detection: 7.603 s 7.906 s everything else: 11.699 s 6.351 s --- mega-renames --- merge-recursive merge-ort overall runtime: 5950.195 s 5499.672 s inexact rename detection: 5746.309 s 5487.120 s everything else: 203.886 s 17.552 s --- just-one-mega --- merge-recursive merge-ort overall runtime: 151.001 s 158.582 s inexact rename detection: 143.448 s 157.835 s everything else: 7.553 s 0.747 s === Timing observations === 0) Maximum speedup The "everything else" row represents the maximum speedup we could achieve if we were to somehow infinitely parallelize inexact rename detection, but leave everything else alone. The fact that this is so much smaller than the real runtime (even in the case with virtually no renames) makes it clear just how overwhelmingly large the time spent on rename detection can be. 1) no-renames 1a) merge-ort is faster than merge-recursive, which is nice. However, this still should not be considered good enough. Although the "merge" backend to rebase (merge-recursive) is sometimes faster than the "apply" backend, this is one of those cases where it is not. In fact, even merge-ort is slower. The "apply" backend can complete this testcase in 6.940 s ± 0.485 s which is about 2x faster than merge-ort and 3x faster than merge-recursive. One goal of the merge-ort performance work will be to make it faster than git-am on this (and similar) testcases. 2) mega-renames 2a) Obviously rename detection is a huge cost; it's where most the time is spent. We need to cut that down. If we could somehow infinitely parallelize it and drive its time to 0, the merge-recursive time would drop to about 204s, and the merge-ort time would drop to about 17s. I think this particular stat shows I've subtly baked a couple performance improvements into merge-ort and into fast-rebase already. 3) just-one-mega 3a) not much to say here, it just gives some flavor for how rebasing only one patch compares to rebasing 35. === Goals === This patch is obviously just the beginning. Here are some of my goals that this measurement will help us achieve: * Drive the cost of rename detection down considerably for merges * After the above has been achieved, see if there are other slowness factors (which would have previously been overshadowed by rename detection costs) which we can then focus on and also optimize. * Ensure our rebase testcase that requires little rename detection is noticeably faster with merge-ort than with apply-based rebase. Signed-off-by: Elijah Newren <newren@gmail.com> Acked-by: Taylor Blau <ttaylorr@github.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-01-24 06:01:12 +00:00
trace2_region_enter("diff", "exact renames", options->repo);
mem_pool_init(&local_pool, 32*1024);
Do exact rename detection regardless of rename limits Now that the exact rename detection is linear-time (with a very small constant factor to boot), there is no longer any reason to limit it by the number of files involved. In some trivial testing, I created a repository with a directory that had a hundred thousand files in it (all with different contents), and then moved that directory to show the effects of renaming 100,000 files. With the new code, that resulted in [torvalds@woody big-rename]$ time ~/git/git show -C | wc -l 400006 real 0m2.071s user 0m1.520s sys 0m0.576s ie the code can correctly detect the hundred thousand renames in about 2 seconds (the number "400006" comes from four lines for each rename: diff --git a/really-big-dir/file-1-1-1-1-1 b/moved-big-dir/file-1-1-1-1-1 similarity index 100% rename from really-big-dir/file-1-1-1-1-1 rename to moved-big-dir/file-1-1-1-1-1 and the extra six lines is from a one-liner commit message and all the commit information and spacing). Most of those two seconds weren't even really the rename detection, it's really all the other stuff needed to get there. With the old code, this wouldn't have been practically possible. Doing a pairwise check of the ten billion possible pairs would have been prohibitively expensive. In fact, even with the rename limiter in place, the old code would waste a lot of time just on the diff_filespec checks, and despite not even trying to find renames, it used to look like: [torvalds@woody big-rename]$ time git show -C | wc -l 1400006 real 0m12.337s user 0m12.285s sys 0m0.192s ie we used to take 12 seconds for this load and not even do any rename detection! (The number 1400006 comes from fourteen lines per file moved: seven lines each for the delete and the create of a one-liner file, and the same extra six lines of commit information). Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2007-10-25 18:24:47 +00:00
/*
* We really want to cull the candidates list early
* with cheap tests in order to avoid doing deltas.
*/
rename_count = find_exact_renames(options, &local_pool);
/*
* Discard local_pool immediately instead of at "cleanup:" in order
* to reduce maximum memory usage; inexact rename detection uses up
* a fair amount of memory, and mem_pools can too.
*/
mem_pool_discard(&local_pool, 0);
merge-ort: begin performance work; instrument with trace2_region_* calls Add some timing instrumentation for both merge-ort and diffcore-rename; I used these to measure and optimize performance in both, and several future patch series will build on these to reduce the timings of some select testcases. === Setup === The primary testcase I used involved rebasing a random topic in the linux kernel (consisting of 35 patches) against an older version. I added two variants, one where I rename a toplevel directory, and another where I only rebase one patch instead of the whole topic. The setup is as follows: $ git clone git://git.kernel.org/pub/scm/linux/kernel/git/stable/linux-stable.git $ git branch hwmon-updates fd8bdb23b91876ac1e624337bb88dc1dcc21d67e $ git branch hwmon-just-one fd8bdb23b91876ac1e624337bb88dc1dcc21d67e~34 $ git branch base 4703d9119972bf586d2cca76ec6438f819ffa30e $ git switch -c 5.4-renames v5.4 $ git mv drivers pilots # Introduce over 26,000 renames $ git commit -m "Rename drivers/ to pilots/" $ git config merge.renameLimit 30000 $ git config merge.directoryRenames true === Testcases === Now with REBASE standing for either "git rebase [--merge]" (using merge-recursive) or "test-tool fast-rebase" (using merge-ort), the testcases are: Testcase #1: no-renames $ git checkout v5.4^0 $ REBASE --onto HEAD base hwmon-updates Note: technically the name is misleading; there are some renames, but very few. Rename detection only takes about half the overall time. Testcase #2: mega-renames $ git checkout 5.4-renames^0 $ REBASE --onto HEAD base hwmon-updates Testcase #3: just-one-mega $ git checkout 5.4-renames^0 $ REBASE --onto HEAD base hwmon-just-one === Timing results === Overall timings, using hyperfine (1 warmup run, 3 runs for mega-renames, 10 runs for the other two cases): merge-recursive merge-ort no-renames: 18.912 s ± 0.174 s 14.263 s ± 0.053 s mega-renames: 5964.031 s ± 10.459 s 5504.231 s ± 5.150 s just-one-mega: 149.583 s ± 0.751 s 158.534 s ± 0.498 s A single re-run of each with some breakdowns: --- no-renames --- merge-recursive merge-ort overall runtime: 19.302 s 14.257 s inexact rename detection: 7.603 s 7.906 s everything else: 11.699 s 6.351 s --- mega-renames --- merge-recursive merge-ort overall runtime: 5950.195 s 5499.672 s inexact rename detection: 5746.309 s 5487.120 s everything else: 203.886 s 17.552 s --- just-one-mega --- merge-recursive merge-ort overall runtime: 151.001 s 158.582 s inexact rename detection: 143.448 s 157.835 s everything else: 7.553 s 0.747 s === Timing observations === 0) Maximum speedup The "everything else" row represents the maximum speedup we could achieve if we were to somehow infinitely parallelize inexact rename detection, but leave everything else alone. The fact that this is so much smaller than the real runtime (even in the case with virtually no renames) makes it clear just how overwhelmingly large the time spent on rename detection can be. 1) no-renames 1a) merge-ort is faster than merge-recursive, which is nice. However, this still should not be considered good enough. Although the "merge" backend to rebase (merge-recursive) is sometimes faster than the "apply" backend, this is one of those cases where it is not. In fact, even merge-ort is slower. The "apply" backend can complete this testcase in 6.940 s ± 0.485 s which is about 2x faster than merge-ort and 3x faster than merge-recursive. One goal of the merge-ort performance work will be to make it faster than git-am on this (and similar) testcases. 2) mega-renames 2a) Obviously rename detection is a huge cost; it's where most the time is spent. We need to cut that down. If we could somehow infinitely parallelize it and drive its time to 0, the merge-recursive time would drop to about 204s, and the merge-ort time would drop to about 17s. I think this particular stat shows I've subtly baked a couple performance improvements into merge-ort and into fast-rebase already. 3) just-one-mega 3a) not much to say here, it just gives some flavor for how rebasing only one patch compares to rebasing 35. === Goals === This patch is obviously just the beginning. Here are some of my goals that this measurement will help us achieve: * Drive the cost of rename detection down considerably for merges * After the above has been achieved, see if there are other slowness factors (which would have previously been overshadowed by rename detection costs) which we can then focus on and also optimize. * Ensure our rebase testcase that requires little rename detection is noticeably faster with merge-ort than with apply-based rebase. Signed-off-by: Elijah Newren <newren@gmail.com> Acked-by: Taylor Blau <ttaylorr@github.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-01-24 06:01:12 +00:00
trace2_region_leave("diff", "exact renames", options->repo);
Do exact rename detection regardless of rename limits Now that the exact rename detection is linear-time (with a very small constant factor to boot), there is no longer any reason to limit it by the number of files involved. In some trivial testing, I created a repository with a directory that had a hundred thousand files in it (all with different contents), and then moved that directory to show the effects of renaming 100,000 files. With the new code, that resulted in [torvalds@woody big-rename]$ time ~/git/git show -C | wc -l 400006 real 0m2.071s user 0m1.520s sys 0m0.576s ie the code can correctly detect the hundred thousand renames in about 2 seconds (the number "400006" comes from four lines for each rename: diff --git a/really-big-dir/file-1-1-1-1-1 b/moved-big-dir/file-1-1-1-1-1 similarity index 100% rename from really-big-dir/file-1-1-1-1-1 rename to moved-big-dir/file-1-1-1-1-1 and the extra six lines is from a one-liner commit message and all the commit information and spacing). Most of those two seconds weren't even really the rename detection, it's really all the other stuff needed to get there. With the old code, this wouldn't have been practically possible. Doing a pairwise check of the ten billion possible pairs would have been prohibitively expensive. In fact, even with the rename limiter in place, the old code would waste a lot of time just on the diff_filespec checks, and despite not even trying to find renames, it used to look like: [torvalds@woody big-rename]$ time git show -C | wc -l 1400006 real 0m12.337s user 0m12.285s sys 0m0.192s ie we used to take 12 seconds for this load and not even do any rename detection! (The number 1400006 comes from fourteen lines per file moved: seven lines each for the delete and the create of a one-liner file, and the same extra six lines of commit information). Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2007-10-25 18:24:47 +00:00
/* Did we only want exact renames? */
if (minimum_score == MAX_SCORE)
goto cleanup;
num_sources = rename_src_nr;
diffcore-rename: guide inexact rename detection based on basenames Make use of the new find_basename_matches() function added in the last two patches, to find renames more rapidly in cases where we can match up files based on basenames. As a quick reminder (see the last two commit messages for more details), this means for example that docs/extensions.txt and docs/config/extensions.txt are considered likely renames if there are no remaining 'extensions.txt' files elsewhere among the added and deleted files, and if a similarity check confirms they are similar, then they are marked as a rename without looking for a better similarity match among other files. This is a behavioral change, as covered in more detail in the previous commit message. We do not use this heuristic together with either break or copy detection. The point of break detection is to say that filename similarity does not imply file content similarity, and we only want to know about file content similarity. The point of copy detection is to use more resources to check for additional similarities, while this is an optimization that uses far less resources but which might also result in finding slightly fewer similarities. So the idea behind this optimization goes against both of those features, and will be turned off for both. For the testcases mentioned in commit 557ac0350d ("merge-ort: begin performance work; instrument with trace2_region_* calls", 2020-10-28), this change improves the performance as follows: Before After no-renames: 13.815 s ± 0.062 s 13.294 s ± 0.103 s mega-renames: 1799.937 s ± 0.493 s 187.248 s ± 0.882 s just-one-mega: 51.289 s ± 0.019 s 5.557 s ± 0.017 s Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-02-14 07:51:49 +00:00
if (want_copies || break_idx) {
/*
* Cull sources:
* - remove ones corresponding to exact renames
diffcore-rename: enable filtering possible rename sources Add the ability to diffcore_rename_extended() to allow external callers to declare that they only need renames detected for a subset of source files, and use that information to skip detecting renames for them. There are two important pieces to this optimization that may not be obvious at first glance: * We do not require callers to just filter the filepairs out to remove the non-relevant sources, because exact rename detection is fast and when it finds a match it can remove both a source and a destination whereas the relevant_sources filter can only remove a source. * We need to filter out the source pairs in a preliminary pass instead of adding a strset_contains(relevant_sources, one->path) check within the nested matrix loop. The reason for that is if we have 30k renames, doing 30k * 30k = 900M strset_contains() calls becomes extraordinarily expensive and defeats the performance gains from this change; we only want to do 30k such calls instead. If callers pass NULL for relevant_sources, that is special cases to treat all sources as relevant. Since all callers currently pass NULL, this optimization does not yet have any effect. Subsequent commits will have merge-ort compute a set of relevant_sources to restrict which sources we detect renames for, and have merge-ort pass that set of relevant_sources to diffcore_rename_extended(). A note about filtering order: Some may be curious why we don't filter out irrelevant sources at the same time we filter out exact renames. While that technically could be done at this point, there are two reasons to defer it: First, was to reinforce a lesson that was too easy to forget. As I mentioned above, in the past I filtered irrelevant sources out before exact rename checking, and then discovered that exact renames' ability to remove both sources and destinations was an important consideration and thus doing the filtering after exact rename checking would speed things up. Then at some point I realized that basename matching could also remove both sources and destinations, and decided to put irrelevant source filtering after basename filtering. That slowed things down a lot. But, despite learning about this important ordering, in later restructuring I forgot and made the same mistake of putting the filtering after basename guided rename detection again. So, I have this series of patches structured to do the irrelevant filtering last to start to show how much extra it costs, and then add relevant filtering in to find_basename_matches() to show how much it speeds things up. Basically, it's a way to reinforce something that apparently was too easy to forget, and make sure the commit messages record this lesson. Second, the items in the "relevant_sources" in this patch series will include all sources that *might be* relevant. It has to be conservative and catch anything that might need a rename, but in the patch series after this one we'll find ways to weed out more of the *might be* relevant ones. Unfortunately, merge-ort does not have sufficient information to weed those ones out, and there isn't enough information at the time of filtering exact renames out to remove the extra ones either. It has to be deferred. So the deferral is in part to simplify some later additions. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-03-11 00:38:24 +00:00
* - remove ones not found in relevant_sources
diffcore-rename: guide inexact rename detection based on basenames Make use of the new find_basename_matches() function added in the last two patches, to find renames more rapidly in cases where we can match up files based on basenames. As a quick reminder (see the last two commit messages for more details), this means for example that docs/extensions.txt and docs/config/extensions.txt are considered likely renames if there are no remaining 'extensions.txt' files elsewhere among the added and deleted files, and if a similarity check confirms they are similar, then they are marked as a rename without looking for a better similarity match among other files. This is a behavioral change, as covered in more detail in the previous commit message. We do not use this heuristic together with either break or copy detection. The point of break detection is to say that filename similarity does not imply file content similarity, and we only want to know about file content similarity. The point of copy detection is to use more resources to check for additional similarities, while this is an optimization that uses far less resources but which might also result in finding slightly fewer similarities. So the idea behind this optimization goes against both of those features, and will be turned off for both. For the testcases mentioned in commit 557ac0350d ("merge-ort: begin performance work; instrument with trace2_region_* calls", 2020-10-28), this change improves the performance as follows: Before After no-renames: 13.815 s ± 0.062 s 13.294 s ± 0.103 s mega-renames: 1799.937 s ± 0.493 s 187.248 s ± 0.882 s just-one-mega: 51.289 s ± 0.019 s 5.557 s ± 0.017 s Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-02-14 07:51:49 +00:00
*/
trace2_region_enter("diff", "cull after exact", options->repo);
diffcore-rename: enable filtering possible rename sources Add the ability to diffcore_rename_extended() to allow external callers to declare that they only need renames detected for a subset of source files, and use that information to skip detecting renames for them. There are two important pieces to this optimization that may not be obvious at first glance: * We do not require callers to just filter the filepairs out to remove the non-relevant sources, because exact rename detection is fast and when it finds a match it can remove both a source and a destination whereas the relevant_sources filter can only remove a source. * We need to filter out the source pairs in a preliminary pass instead of adding a strset_contains(relevant_sources, one->path) check within the nested matrix loop. The reason for that is if we have 30k renames, doing 30k * 30k = 900M strset_contains() calls becomes extraordinarily expensive and defeats the performance gains from this change; we only want to do 30k such calls instead. If callers pass NULL for relevant_sources, that is special cases to treat all sources as relevant. Since all callers currently pass NULL, this optimization does not yet have any effect. Subsequent commits will have merge-ort compute a set of relevant_sources to restrict which sources we detect renames for, and have merge-ort pass that set of relevant_sources to diffcore_rename_extended(). A note about filtering order: Some may be curious why we don't filter out irrelevant sources at the same time we filter out exact renames. While that technically could be done at this point, there are two reasons to defer it: First, was to reinforce a lesson that was too easy to forget. As I mentioned above, in the past I filtered irrelevant sources out before exact rename checking, and then discovered that exact renames' ability to remove both sources and destinations was an important consideration and thus doing the filtering after exact rename checking would speed things up. Then at some point I realized that basename matching could also remove both sources and destinations, and decided to put irrelevant source filtering after basename filtering. That slowed things down a lot. But, despite learning about this important ordering, in later restructuring I forgot and made the same mistake of putting the filtering after basename guided rename detection again. So, I have this series of patches structured to do the irrelevant filtering last to start to show how much extra it costs, and then add relevant filtering in to find_basename_matches() to show how much it speeds things up. Basically, it's a way to reinforce something that apparently was too easy to forget, and make sure the commit messages record this lesson. Second, the items in the "relevant_sources" in this patch series will include all sources that *might be* relevant. It has to be conservative and catch anything that might need a rename, but in the patch series after this one we'll find ways to weed out more of the *might be* relevant ones. Unfortunately, merge-ort does not have sufficient information to weed those ones out, and there isn't enough information at the time of filtering exact renames out to remove the extra ones either. It has to be deferred. So the deferral is in part to simplify some later additions. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-03-11 00:38:24 +00:00
remove_unneeded_paths_from_src(want_copies, relevant_sources);
diffcore-rename: guide inexact rename detection based on basenames Make use of the new find_basename_matches() function added in the last two patches, to find renames more rapidly in cases where we can match up files based on basenames. As a quick reminder (see the last two commit messages for more details), this means for example that docs/extensions.txt and docs/config/extensions.txt are considered likely renames if there are no remaining 'extensions.txt' files elsewhere among the added and deleted files, and if a similarity check confirms they are similar, then they are marked as a rename without looking for a better similarity match among other files. This is a behavioral change, as covered in more detail in the previous commit message. We do not use this heuristic together with either break or copy detection. The point of break detection is to say that filename similarity does not imply file content similarity, and we only want to know about file content similarity. The point of copy detection is to use more resources to check for additional similarities, while this is an optimization that uses far less resources but which might also result in finding slightly fewer similarities. So the idea behind this optimization goes against both of those features, and will be turned off for both. For the testcases mentioned in commit 557ac0350d ("merge-ort: begin performance work; instrument with trace2_region_* calls", 2020-10-28), this change improves the performance as follows: Before After no-renames: 13.815 s ± 0.062 s 13.294 s ± 0.103 s mega-renames: 1799.937 s ± 0.493 s 187.248 s ± 0.882 s just-one-mega: 51.289 s ± 0.019 s 5.557 s ± 0.017 s Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-02-14 07:51:49 +00:00
trace2_region_leave("diff", "cull after exact", options->repo);
} else {
/* Determine minimum score to match basenames */
double factor = 0.5;
char *basename_factor = getenv("GIT_BASENAME_FACTOR");
int min_basename_score;
if (basename_factor)
factor = strtol(basename_factor, NULL, 10)/100.0;
assert(factor >= 0.0 && factor <= 1.0);
min_basename_score = minimum_score +
(int)(factor * (MAX_SCORE - minimum_score));
/*
* Cull sources:
* - remove ones involved in renames (found via exact match)
*/
trace2_region_enter("diff", "cull after exact", options->repo);
diffcore-rename: enable filtering possible rename sources Add the ability to diffcore_rename_extended() to allow external callers to declare that they only need renames detected for a subset of source files, and use that information to skip detecting renames for them. There are two important pieces to this optimization that may not be obvious at first glance: * We do not require callers to just filter the filepairs out to remove the non-relevant sources, because exact rename detection is fast and when it finds a match it can remove both a source and a destination whereas the relevant_sources filter can only remove a source. * We need to filter out the source pairs in a preliminary pass instead of adding a strset_contains(relevant_sources, one->path) check within the nested matrix loop. The reason for that is if we have 30k renames, doing 30k * 30k = 900M strset_contains() calls becomes extraordinarily expensive and defeats the performance gains from this change; we only want to do 30k such calls instead. If callers pass NULL for relevant_sources, that is special cases to treat all sources as relevant. Since all callers currently pass NULL, this optimization does not yet have any effect. Subsequent commits will have merge-ort compute a set of relevant_sources to restrict which sources we detect renames for, and have merge-ort pass that set of relevant_sources to diffcore_rename_extended(). A note about filtering order: Some may be curious why we don't filter out irrelevant sources at the same time we filter out exact renames. While that technically could be done at this point, there are two reasons to defer it: First, was to reinforce a lesson that was too easy to forget. As I mentioned above, in the past I filtered irrelevant sources out before exact rename checking, and then discovered that exact renames' ability to remove both sources and destinations was an important consideration and thus doing the filtering after exact rename checking would speed things up. Then at some point I realized that basename matching could also remove both sources and destinations, and decided to put irrelevant source filtering after basename filtering. That slowed things down a lot. But, despite learning about this important ordering, in later restructuring I forgot and made the same mistake of putting the filtering after basename guided rename detection again. So, I have this series of patches structured to do the irrelevant filtering last to start to show how much extra it costs, and then add relevant filtering in to find_basename_matches() to show how much it speeds things up. Basically, it's a way to reinforce something that apparently was too easy to forget, and make sure the commit messages record this lesson. Second, the items in the "relevant_sources" in this patch series will include all sources that *might be* relevant. It has to be conservative and catch anything that might need a rename, but in the patch series after this one we'll find ways to weed out more of the *might be* relevant ones. Unfortunately, merge-ort does not have sufficient information to weed those ones out, and there isn't enough information at the time of filtering exact renames out to remove the extra ones either. It has to be deferred. So the deferral is in part to simplify some later additions. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-03-11 00:38:24 +00:00
remove_unneeded_paths_from_src(want_copies, NULL);
diffcore-rename: guide inexact rename detection based on basenames Make use of the new find_basename_matches() function added in the last two patches, to find renames more rapidly in cases where we can match up files based on basenames. As a quick reminder (see the last two commit messages for more details), this means for example that docs/extensions.txt and docs/config/extensions.txt are considered likely renames if there are no remaining 'extensions.txt' files elsewhere among the added and deleted files, and if a similarity check confirms they are similar, then they are marked as a rename without looking for a better similarity match among other files. This is a behavioral change, as covered in more detail in the previous commit message. We do not use this heuristic together with either break or copy detection. The point of break detection is to say that filename similarity does not imply file content similarity, and we only want to know about file content similarity. The point of copy detection is to use more resources to check for additional similarities, while this is an optimization that uses far less resources but which might also result in finding slightly fewer similarities. So the idea behind this optimization goes against both of those features, and will be turned off for both. For the testcases mentioned in commit 557ac0350d ("merge-ort: begin performance work; instrument with trace2_region_* calls", 2020-10-28), this change improves the performance as follows: Before After no-renames: 13.815 s ± 0.062 s 13.294 s ± 0.103 s mega-renames: 1799.937 s ± 0.493 s 187.248 s ± 0.882 s just-one-mega: 51.289 s ± 0.019 s 5.557 s ± 0.017 s Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-02-14 07:51:49 +00:00
trace2_region_leave("diff", "cull after exact", options->repo);
/* Preparation for basename-driven matching. */
trace2_region_enter("diff", "dir rename setup", options->repo);
initialize_dir_rename_info(&info, relevant_sources,
merge-ort, diffcore-rename: employ cached renames when possible When there are many renames between the old base of a series of commits and the new base, the way sequencer.c, merge-recursive.c, and diffcore-rename.c have traditionally split the work resulted in redetecting the same renames with each and every commit being transplanted. To address this, the last several commits have been creating a cache of rename detection results, determining when it was safe to use such a cache in subsequent merge operations, adding helper functions, and so on. See the previous half dozen commit messages for additional discussion of this optimization, particularly the message a few commits ago entitled "add code to check for whether cached renames can be reused". This commit finally ties all of that work together, modifying the merge algorithm to make use of these cached renames. For the testcases mentioned in commit 557ac0350d ("merge-ort: begin performance work; instrument with trace2_region_* calls", 2020-10-28), this change improves the performance as follows: Before After no-renames: 5.665 s ± 0.129 s 5.622 s ± 0.059 s mega-renames: 11.435 s ± 0.158 s 10.127 s ± 0.073 s just-one-mega: 494.2 ms ± 6.1 ms 500.3 ms ± 3.8 ms That's a fairly small improvement, but mostly because the previous optimizations were so effective for these particular testcases; this optimization only kicks in when the others don't. If we undid the basename-guided rename detection and skip-irrelevant-renames optimizations, then we'd see that this series by itself improved performance as follows: Before Basename Series After Just This Series no-renames: 13.815 s ± 0.062 s 5.697 s ± 0.080 s mega-renames: 1799.937 s ± 0.493 s 205.709 s ± 0.457 s Since this optimization kicks in to help accelerate cases where the previous optimizations do not apply, this last comparison shows that this cached-renames optimization has the potential to help signficantly in cases that don't meet the requirements for the other optimizations to be effective. The changes made in this optimization also lay some important groundwork for a future optimization around having collect_merge_info() avoid recursing into subtrees in more cases. However, for this optimization to be effective, merge_switch_to_result() should only be called when the rebase or cherry-pick operation has either completed or hit a case where the user needs to resolve a conflict or edit the result. If it is called after every commit, as sequencer.c does, then the working tree and index are needlessly updated with every commit and the cached metadata is tossed, defeating this optimization. Refactoring sequencer.c to only call merge_switch_to_result() at the end of the operation is a bigger undertaking, and the practical benefits of this optimization will not be realized until that work is performed. Since `test-tool fast-rebase` only updates at the end of the operation, it was used to obtain the timings above. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-05-20 06:09:41 +00:00
dirs_removed, dir_rename_count,
cached_pairs);
trace2_region_leave("diff", "dir rename setup", options->repo);
diffcore-rename: guide inexact rename detection based on basenames Make use of the new find_basename_matches() function added in the last two patches, to find renames more rapidly in cases where we can match up files based on basenames. As a quick reminder (see the last two commit messages for more details), this means for example that docs/extensions.txt and docs/config/extensions.txt are considered likely renames if there are no remaining 'extensions.txt' files elsewhere among the added and deleted files, and if a similarity check confirms they are similar, then they are marked as a rename without looking for a better similarity match among other files. This is a behavioral change, as covered in more detail in the previous commit message. We do not use this heuristic together with either break or copy detection. The point of break detection is to say that filename similarity does not imply file content similarity, and we only want to know about file content similarity. The point of copy detection is to use more resources to check for additional similarities, while this is an optimization that uses far less resources but which might also result in finding slightly fewer similarities. So the idea behind this optimization goes against both of those features, and will be turned off for both. For the testcases mentioned in commit 557ac0350d ("merge-ort: begin performance work; instrument with trace2_region_* calls", 2020-10-28), this change improves the performance as follows: Before After no-renames: 13.815 s ± 0.062 s 13.294 s ± 0.103 s mega-renames: 1799.937 s ± 0.493 s 187.248 s ± 0.882 s just-one-mega: 51.289 s ± 0.019 s 5.557 s ± 0.017 s Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-02-14 07:51:49 +00:00
/* Utilize file basenames to quickly find renames. */
trace2_region_enter("diff", "basename matches", options->repo);
rename_count += find_basename_matches(options,
min_basename_score,
&info,
relevant_sources,
dirs_removed);
diffcore-rename: guide inexact rename detection based on basenames Make use of the new find_basename_matches() function added in the last two patches, to find renames more rapidly in cases where we can match up files based on basenames. As a quick reminder (see the last two commit messages for more details), this means for example that docs/extensions.txt and docs/config/extensions.txt are considered likely renames if there are no remaining 'extensions.txt' files elsewhere among the added and deleted files, and if a similarity check confirms they are similar, then they are marked as a rename without looking for a better similarity match among other files. This is a behavioral change, as covered in more detail in the previous commit message. We do not use this heuristic together with either break or copy detection. The point of break detection is to say that filename similarity does not imply file content similarity, and we only want to know about file content similarity. The point of copy detection is to use more resources to check for additional similarities, while this is an optimization that uses far less resources but which might also result in finding slightly fewer similarities. So the idea behind this optimization goes against both of those features, and will be turned off for both. For the testcases mentioned in commit 557ac0350d ("merge-ort: begin performance work; instrument with trace2_region_* calls", 2020-10-28), this change improves the performance as follows: Before After no-renames: 13.815 s ± 0.062 s 13.294 s ± 0.103 s mega-renames: 1799.937 s ± 0.493 s 187.248 s ± 0.882 s just-one-mega: 51.289 s ± 0.019 s 5.557 s ± 0.017 s Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-02-14 07:51:49 +00:00
trace2_region_leave("diff", "basename matches", options->repo);
/*
* Cull sources, again:
* - remove ones involved in renames (found via basenames)
diffcore-rename: enable filtering possible rename sources Add the ability to diffcore_rename_extended() to allow external callers to declare that they only need renames detected for a subset of source files, and use that information to skip detecting renames for them. There are two important pieces to this optimization that may not be obvious at first glance: * We do not require callers to just filter the filepairs out to remove the non-relevant sources, because exact rename detection is fast and when it finds a match it can remove both a source and a destination whereas the relevant_sources filter can only remove a source. * We need to filter out the source pairs in a preliminary pass instead of adding a strset_contains(relevant_sources, one->path) check within the nested matrix loop. The reason for that is if we have 30k renames, doing 30k * 30k = 900M strset_contains() calls becomes extraordinarily expensive and defeats the performance gains from this change; we only want to do 30k such calls instead. If callers pass NULL for relevant_sources, that is special cases to treat all sources as relevant. Since all callers currently pass NULL, this optimization does not yet have any effect. Subsequent commits will have merge-ort compute a set of relevant_sources to restrict which sources we detect renames for, and have merge-ort pass that set of relevant_sources to diffcore_rename_extended(). A note about filtering order: Some may be curious why we don't filter out irrelevant sources at the same time we filter out exact renames. While that technically could be done at this point, there are two reasons to defer it: First, was to reinforce a lesson that was too easy to forget. As I mentioned above, in the past I filtered irrelevant sources out before exact rename checking, and then discovered that exact renames' ability to remove both sources and destinations was an important consideration and thus doing the filtering after exact rename checking would speed things up. Then at some point I realized that basename matching could also remove both sources and destinations, and decided to put irrelevant source filtering after basename filtering. That slowed things down a lot. But, despite learning about this important ordering, in later restructuring I forgot and made the same mistake of putting the filtering after basename guided rename detection again. So, I have this series of patches structured to do the irrelevant filtering last to start to show how much extra it costs, and then add relevant filtering in to find_basename_matches() to show how much it speeds things up. Basically, it's a way to reinforce something that apparently was too easy to forget, and make sure the commit messages record this lesson. Second, the items in the "relevant_sources" in this patch series will include all sources that *might be* relevant. It has to be conservative and catch anything that might need a rename, but in the patch series after this one we'll find ways to weed out more of the *might be* relevant ones. Unfortunately, merge-ort does not have sufficient information to weed those ones out, and there isn't enough information at the time of filtering exact renames out to remove the extra ones either. It has to be deferred. So the deferral is in part to simplify some later additions. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-03-11 00:38:24 +00:00
* - remove ones not found in relevant_sources
diffcore-rename: take advantage of "majority rules" to skip more renames In directory rename detection (when a directory is removed on one side of history and the other side adds new files to that directory), we work to find where the greatest number of files within that directory were renamed to so that the new files can be moved with the majority of the files. Naively, we can just do this by detecting renames for *all* files within the removed/renamed directory, looking at all the destination directories where files within that directory were moved, and if there is more than one such directory then taking the one with the greatest number of files as the directory where the old directory was renamed to. However, sometimes there are enough renames from exact rename detection or basename-guided rename detection that we have enough information to determine the majority winner already. Add a function meant to compute whether particular renames are still needed based on this majority rules check. The next several commits will then add the necessary infrastructure to get the information we need to compute which additional rename sources we can skip. An important side note for future further optimization: There is a possible improvement to this optimization that I have not yet attempted and will not be included in this series of patches: we could first check whether exact renames provide enough information for us to determine directory renames, and avoid doing basename-guided rename detection on some or all of the RELEVANT_LOCATION files within those directories. In effect, this variant would mean doing the handle_early_known_dir_renames() both after exact rename detection and again after basename-guided rename detection, though it would also mean decrementing the number of "unknown" renames for each rename we found from basename-guided rename detection. Adding this additional check for skippable renames right after exact rename detection might turn out to be valuable, especially for partial clones where it might allow us to download certain source files entirely. However, this particular optimization was actually the last one I did in original implementation order, and by the time I implemented this idea, every testcase I had was sufficiently fast that further optimization was unwarranted. If future testcases arise that tax rename detection more heavily (or perhaps partial clones can benefit from avoiding loading more objects), it may be worth implementing this more involved variant. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-03-13 22:22:01 +00:00
* and
* - remove ones in relevant_sources which are needed only
* for directory renames IF no ancestory directory
* actually needs to know any more individual path
* renames under them
diffcore-rename: guide inexact rename detection based on basenames Make use of the new find_basename_matches() function added in the last two patches, to find renames more rapidly in cases where we can match up files based on basenames. As a quick reminder (see the last two commit messages for more details), this means for example that docs/extensions.txt and docs/config/extensions.txt are considered likely renames if there are no remaining 'extensions.txt' files elsewhere among the added and deleted files, and if a similarity check confirms they are similar, then they are marked as a rename without looking for a better similarity match among other files. This is a behavioral change, as covered in more detail in the previous commit message. We do not use this heuristic together with either break or copy detection. The point of break detection is to say that filename similarity does not imply file content similarity, and we only want to know about file content similarity. The point of copy detection is to use more resources to check for additional similarities, while this is an optimization that uses far less resources but which might also result in finding slightly fewer similarities. So the idea behind this optimization goes against both of those features, and will be turned off for both. For the testcases mentioned in commit 557ac0350d ("merge-ort: begin performance work; instrument with trace2_region_* calls", 2020-10-28), this change improves the performance as follows: Before After no-renames: 13.815 s ± 0.062 s 13.294 s ± 0.103 s mega-renames: 1799.937 s ± 0.493 s 187.248 s ± 0.882 s just-one-mega: 51.289 s ± 0.019 s 5.557 s ± 0.017 s Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-02-14 07:51:49 +00:00
*/
trace2_region_enter("diff", "cull basename", options->repo);
diffcore-rename: enable filtering possible rename sources Add the ability to diffcore_rename_extended() to allow external callers to declare that they only need renames detected for a subset of source files, and use that information to skip detecting renames for them. There are two important pieces to this optimization that may not be obvious at first glance: * We do not require callers to just filter the filepairs out to remove the non-relevant sources, because exact rename detection is fast and when it finds a match it can remove both a source and a destination whereas the relevant_sources filter can only remove a source. * We need to filter out the source pairs in a preliminary pass instead of adding a strset_contains(relevant_sources, one->path) check within the nested matrix loop. The reason for that is if we have 30k renames, doing 30k * 30k = 900M strset_contains() calls becomes extraordinarily expensive and defeats the performance gains from this change; we only want to do 30k such calls instead. If callers pass NULL for relevant_sources, that is special cases to treat all sources as relevant. Since all callers currently pass NULL, this optimization does not yet have any effect. Subsequent commits will have merge-ort compute a set of relevant_sources to restrict which sources we detect renames for, and have merge-ort pass that set of relevant_sources to diffcore_rename_extended(). A note about filtering order: Some may be curious why we don't filter out irrelevant sources at the same time we filter out exact renames. While that technically could be done at this point, there are two reasons to defer it: First, was to reinforce a lesson that was too easy to forget. As I mentioned above, in the past I filtered irrelevant sources out before exact rename checking, and then discovered that exact renames' ability to remove both sources and destinations was an important consideration and thus doing the filtering after exact rename checking would speed things up. Then at some point I realized that basename matching could also remove both sources and destinations, and decided to put irrelevant source filtering after basename filtering. That slowed things down a lot. But, despite learning about this important ordering, in later restructuring I forgot and made the same mistake of putting the filtering after basename guided rename detection again. So, I have this series of patches structured to do the irrelevant filtering last to start to show how much extra it costs, and then add relevant filtering in to find_basename_matches() to show how much it speeds things up. Basically, it's a way to reinforce something that apparently was too easy to forget, and make sure the commit messages record this lesson. Second, the items in the "relevant_sources" in this patch series will include all sources that *might be* relevant. It has to be conservative and catch anything that might need a rename, but in the patch series after this one we'll find ways to weed out more of the *might be* relevant ones. Unfortunately, merge-ort does not have sufficient information to weed those ones out, and there isn't enough information at the time of filtering exact renames out to remove the extra ones either. It has to be deferred. So the deferral is in part to simplify some later additions. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-03-11 00:38:24 +00:00
remove_unneeded_paths_from_src(want_copies, relevant_sources);
diffcore-rename: take advantage of "majority rules" to skip more renames In directory rename detection (when a directory is removed on one side of history and the other side adds new files to that directory), we work to find where the greatest number of files within that directory were renamed to so that the new files can be moved with the majority of the files. Naively, we can just do this by detecting renames for *all* files within the removed/renamed directory, looking at all the destination directories where files within that directory were moved, and if there is more than one such directory then taking the one with the greatest number of files as the directory where the old directory was renamed to. However, sometimes there are enough renames from exact rename detection or basename-guided rename detection that we have enough information to determine the majority winner already. Add a function meant to compute whether particular renames are still needed based on this majority rules check. The next several commits will then add the necessary infrastructure to get the information we need to compute which additional rename sources we can skip. An important side note for future further optimization: There is a possible improvement to this optimization that I have not yet attempted and will not be included in this series of patches: we could first check whether exact renames provide enough information for us to determine directory renames, and avoid doing basename-guided rename detection on some or all of the RELEVANT_LOCATION files within those directories. In effect, this variant would mean doing the handle_early_known_dir_renames() both after exact rename detection and again after basename-guided rename detection, though it would also mean decrementing the number of "unknown" renames for each rename we found from basename-guided rename detection. Adding this additional check for skippable renames right after exact rename detection might turn out to be valuable, especially for partial clones where it might allow us to download certain source files entirely. However, this particular optimization was actually the last one I did in original implementation order, and by the time I implemented this idea, every testcase I had was sufficiently fast that further optimization was unwarranted. If future testcases arise that tax rename detection more heavily (or perhaps partial clones can benefit from avoiding loading more objects), it may be worth implementing this more involved variant. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-03-13 22:22:01 +00:00
handle_early_known_dir_renames(&info, relevant_sources,
dirs_removed);
diffcore-rename: guide inexact rename detection based on basenames Make use of the new find_basename_matches() function added in the last two patches, to find renames more rapidly in cases where we can match up files based on basenames. As a quick reminder (see the last two commit messages for more details), this means for example that docs/extensions.txt and docs/config/extensions.txt are considered likely renames if there are no remaining 'extensions.txt' files elsewhere among the added and deleted files, and if a similarity check confirms they are similar, then they are marked as a rename without looking for a better similarity match among other files. This is a behavioral change, as covered in more detail in the previous commit message. We do not use this heuristic together with either break or copy detection. The point of break detection is to say that filename similarity does not imply file content similarity, and we only want to know about file content similarity. The point of copy detection is to use more resources to check for additional similarities, while this is an optimization that uses far less resources but which might also result in finding slightly fewer similarities. So the idea behind this optimization goes against both of those features, and will be turned off for both. For the testcases mentioned in commit 557ac0350d ("merge-ort: begin performance work; instrument with trace2_region_* calls", 2020-10-28), this change improves the performance as follows: Before After no-renames: 13.815 s ± 0.062 s 13.294 s ± 0.103 s mega-renames: 1799.937 s ± 0.493 s 187.248 s ± 0.882 s just-one-mega: 51.289 s ± 0.019 s 5.557 s ± 0.017 s Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-02-14 07:51:49 +00:00
trace2_region_leave("diff", "cull basename", options->repo);
}
/* Calculate how many rename destinations are left */
num_destinations = (rename_dst_nr - rename_count);
num_sources = rename_src_nr; /* rename_src_nr reflects lower number */
/* All done? */
if (!num_destinations || !num_sources)
goto cleanup;
switch (too_many_rename_candidates(num_destinations, num_sources,
options)) {
case 1:
goto cleanup;
case 2:
options->degraded_cc_to_c = 1;
skip_unmodified = 1;
break;
default:
break;
}
merge-ort: begin performance work; instrument with trace2_region_* calls Add some timing instrumentation for both merge-ort and diffcore-rename; I used these to measure and optimize performance in both, and several future patch series will build on these to reduce the timings of some select testcases. === Setup === The primary testcase I used involved rebasing a random topic in the linux kernel (consisting of 35 patches) against an older version. I added two variants, one where I rename a toplevel directory, and another where I only rebase one patch instead of the whole topic. The setup is as follows: $ git clone git://git.kernel.org/pub/scm/linux/kernel/git/stable/linux-stable.git $ git branch hwmon-updates fd8bdb23b91876ac1e624337bb88dc1dcc21d67e $ git branch hwmon-just-one fd8bdb23b91876ac1e624337bb88dc1dcc21d67e~34 $ git branch base 4703d9119972bf586d2cca76ec6438f819ffa30e $ git switch -c 5.4-renames v5.4 $ git mv drivers pilots # Introduce over 26,000 renames $ git commit -m "Rename drivers/ to pilots/" $ git config merge.renameLimit 30000 $ git config merge.directoryRenames true === Testcases === Now with REBASE standing for either "git rebase [--merge]" (using merge-recursive) or "test-tool fast-rebase" (using merge-ort), the testcases are: Testcase #1: no-renames $ git checkout v5.4^0 $ REBASE --onto HEAD base hwmon-updates Note: technically the name is misleading; there are some renames, but very few. Rename detection only takes about half the overall time. Testcase #2: mega-renames $ git checkout 5.4-renames^0 $ REBASE --onto HEAD base hwmon-updates Testcase #3: just-one-mega $ git checkout 5.4-renames^0 $ REBASE --onto HEAD base hwmon-just-one === Timing results === Overall timings, using hyperfine (1 warmup run, 3 runs for mega-renames, 10 runs for the other two cases): merge-recursive merge-ort no-renames: 18.912 s ± 0.174 s 14.263 s ± 0.053 s mega-renames: 5964.031 s ± 10.459 s 5504.231 s ± 5.150 s just-one-mega: 149.583 s ± 0.751 s 158.534 s ± 0.498 s A single re-run of each with some breakdowns: --- no-renames --- merge-recursive merge-ort overall runtime: 19.302 s 14.257 s inexact rename detection: 7.603 s 7.906 s everything else: 11.699 s 6.351 s --- mega-renames --- merge-recursive merge-ort overall runtime: 5950.195 s 5499.672 s inexact rename detection: 5746.309 s 5487.120 s everything else: 203.886 s 17.552 s --- just-one-mega --- merge-recursive merge-ort overall runtime: 151.001 s 158.582 s inexact rename detection: 143.448 s 157.835 s everything else: 7.553 s 0.747 s === Timing observations === 0) Maximum speedup The "everything else" row represents the maximum speedup we could achieve if we were to somehow infinitely parallelize inexact rename detection, but leave everything else alone. The fact that this is so much smaller than the real runtime (even in the case with virtually no renames) makes it clear just how overwhelmingly large the time spent on rename detection can be. 1) no-renames 1a) merge-ort is faster than merge-recursive, which is nice. However, this still should not be considered good enough. Although the "merge" backend to rebase (merge-recursive) is sometimes faster than the "apply" backend, this is one of those cases where it is not. In fact, even merge-ort is slower. The "apply" backend can complete this testcase in 6.940 s ± 0.485 s which is about 2x faster than merge-ort and 3x faster than merge-recursive. One goal of the merge-ort performance work will be to make it faster than git-am on this (and similar) testcases. 2) mega-renames 2a) Obviously rename detection is a huge cost; it's where most the time is spent. We need to cut that down. If we could somehow infinitely parallelize it and drive its time to 0, the merge-recursive time would drop to about 204s, and the merge-ort time would drop to about 17s. I think this particular stat shows I've subtly baked a couple performance improvements into merge-ort and into fast-rebase already. 3) just-one-mega 3a) not much to say here, it just gives some flavor for how rebasing only one patch compares to rebasing 35. === Goals === This patch is obviously just the beginning. Here are some of my goals that this measurement will help us achieve: * Drive the cost of rename detection down considerably for merges * After the above has been achieved, see if there are other slowness factors (which would have previously been overshadowed by rename detection costs) which we can then focus on and also optimize. * Ensure our rebase testcase that requires little rename detection is noticeably faster with merge-ort than with apply-based rebase. Signed-off-by: Elijah Newren <newren@gmail.com> Acked-by: Taylor Blau <ttaylorr@github.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-01-24 06:01:12 +00:00
trace2_region_enter("diff", "inexact renames", options->repo);
if (options->show_rename_progress) {
progress: simplify "delayed" progress API We used to expose the full power of the delayed progress API to the callers, so that they can specify, not just the message to show and expected total amount of work that is used to compute the percentage of work performed so far, the percent-threshold parameter P and the delay-seconds parameter N. The progress meter starts to show at N seconds into the operation only if we have not yet completed P per-cent of the total work. Most callers used either (0%, 2s) or (50%, 1s) as (P, N), but there are oddballs that chose more random-looking values like 95%. For a smoother workload, (50%, 1s) would allow us to start showing the progress meter earlier than (0%, 2s), while keeping the chance of not showing progress meter for long running operation the same as the latter. For a task that would take 2s or more to complete, it is likely that less than half of it would complete within the first second, if the workload is smooth. But for a spiky workload whose earlier part is easier, such a setting is likely to fail to show the progress meter entirely and (0%, 2s) is more appropriate. But that is merely a theory. Realistically, it is of dubious value to ask each codepath to carefully consider smoothness of their workload and specify their own setting by passing two extra parameters. Let's simplify the API by dropping both parameters and have everybody use (0%, 2s). Oh, by the way, the percent-threshold parameter and the structure member were consistently misspelled, which also is now fixed ;-) Helped-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-08-19 17:39:41 +00:00
progress = start_delayed_progress(
_("Performing inexact rename detection"),
(uint64_t)num_destinations * (uint64_t)num_sources);
}
/* Finish setting up dpf_options */
prefetch_options.skip_unmodified = skip_unmodified;
if (options->repo == the_repository && repo_has_promisor_remote(the_repository)) {
diffcore-rename: use a different prefetch for basename comparisons merge-ort was designed to minimize the amount of data needed and used, and several changes were made to diffcore-rename to take advantage of extra metadata to enable this data minimization (particularly the relevant_sources variable for skipping "irrelevant" renames). This effort obviously succeeded in drastically reducing computation times, but should also theoretically allow partial clones to download much less information. Previously, though, the "prefetch" command used in diffcore-rename had never been modified and downloaded many blobs that were unnecessary for merge-ort. This commit corrects that. When doing basename comparisons, we want to fetch only the objects that will be used for basename comparisons. If after basename fetching this leaves us with no more relevant sources (or no more destinations), then we won't need to do the full inexact rename detection and can skip downloading additional source and destination files. Even if we have to do that later full inexact rename detection, irrelevant sources are culled after basename matching and before the full inexact rename detection, so we can still avoid downloading the blobs for irrelevant sources. Rename prefetch() to inexact_prefetch(), and introduce a new basename_prefetch() to take advantage of this. If we modify the testcase from commit 557ac0350d ("merge-ort: begin performance work; instrument with trace2_region_* calls", 2021-01-23) to pass --sparse --filter=blob:none to the clone command, and use the new trace2 "fetch_count" output from a few commits ago to track both the number of fetch subcommands invoked and the number of objects fetched across all those fetches, then for the mega-renames testcase we observe the following: BEFORE this commit, rebasing 35 patches: strategy # of fetches total # of objects fetched --------- ------------ -------------------------- recursive 62 11423 ort 30 11391 AFTER this commit, rebasing the same 35 patches: ort 32 63 This means that the new code only needs to download less than 2 blobs per patch being rebased. That is especially interesting given that the repository at the start only had approximately half a dozen TOTAL blobs downloaded to start with (because the default sparse-checkout of just the toplevel directory was in use). So, for this particular linux kernel testcase that involved ~26,000 renames on the upstream side (drivers/ -> pilots/) across which 35 patches were being rebased, this change reduces the number of blobs that need to be downloaded by a factor of ~180. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-06-22 08:04:40 +00:00
dpf_options.missing_object_cb = inexact_prefetch;
dpf_options.missing_object_data = &prefetch_options;
}
CALLOC_ARRAY(mx, st_mult(NUM_CANDIDATE_PER_DST, num_destinations));
for (dst_cnt = i = 0; i < rename_dst_nr; i++) {
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 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. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-11 09:08:47 +00:00
struct diff_filespec *two = rename_dst[i].p->two;
struct diff_score *m;
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 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. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-11 09:08:47 +00:00
if (rename_dst[i].is_rename)
diffcore-rename: guide inexact rename detection based on basenames Make use of the new find_basename_matches() function added in the last two patches, to find renames more rapidly in cases where we can match up files based on basenames. As a quick reminder (see the last two commit messages for more details), this means for example that docs/extensions.txt and docs/config/extensions.txt are considered likely renames if there are no remaining 'extensions.txt' files elsewhere among the added and deleted files, and if a similarity check confirms they are similar, then they are marked as a rename without looking for a better similarity match among other files. This is a behavioral change, as covered in more detail in the previous commit message. We do not use this heuristic together with either break or copy detection. The point of break detection is to say that filename similarity does not imply file content similarity, and we only want to know about file content similarity. The point of copy detection is to use more resources to check for additional similarities, while this is an optimization that uses far less resources but which might also result in finding slightly fewer similarities. So the idea behind this optimization goes against both of those features, and will be turned off for both. For the testcases mentioned in commit 557ac0350d ("merge-ort: begin performance work; instrument with trace2_region_* calls", 2020-10-28), this change improves the performance as follows: Before After no-renames: 13.815 s ± 0.062 s 13.294 s ± 0.103 s mega-renames: 1799.937 s ± 0.493 s 187.248 s ± 0.882 s just-one-mega: 51.289 s ± 0.019 s 5.557 s ± 0.017 s Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-02-14 07:51:49 +00:00
continue; /* exact or basename match already handled */
m = &mx[dst_cnt * NUM_CANDIDATE_PER_DST];
for (j = 0; j < NUM_CANDIDATE_PER_DST; j++)
m[j].dst = -1;
for (j = 0; j < rename_src_nr; j++) {
struct diff_filespec *one = rename_src[j].p->one;
struct diff_score this_src;
assert(!one->rename_used || want_copies || break_idx);
if (skip_unmodified &&
diff_unmodified_pair(rename_src[j].p))
continue;
this_src.score = estimate_similarity(options->repo,
one, two,
minimum_score,
&dpf_options);
this_src.name_score = basename_same(one, two);
this_src.dst = i;
this_src.src = j;
record_if_better(m, &this_src);
/*
* Once we run estimate_similarity,
* We do not need the text anymore.
*/
diff_free_filespec_blob(one);
diff_free_filespec_blob(two);
}
dst_cnt++;
display_progress(progress,
(uint64_t)dst_cnt * (uint64_t)num_sources);
}
stop_progress(&progress);
/* cost matrix sorted by most to least similar pair */
STABLE_QSORT(mx, dst_cnt * NUM_CANDIDATE_PER_DST, score_compare);
rename_count += find_renames(mx, dst_cnt, minimum_score, 0,
&info, dirs_removed);
if (want_copies)
rename_count += find_renames(mx, dst_cnt, minimum_score, 1,
&info, dirs_removed);
free(mx);
merge-ort: begin performance work; instrument with trace2_region_* calls Add some timing instrumentation for both merge-ort and diffcore-rename; I used these to measure and optimize performance in both, and several future patch series will build on these to reduce the timings of some select testcases. === Setup === The primary testcase I used involved rebasing a random topic in the linux kernel (consisting of 35 patches) against an older version. I added two variants, one where I rename a toplevel directory, and another where I only rebase one patch instead of the whole topic. The setup is as follows: $ git clone git://git.kernel.org/pub/scm/linux/kernel/git/stable/linux-stable.git $ git branch hwmon-updates fd8bdb23b91876ac1e624337bb88dc1dcc21d67e $ git branch hwmon-just-one fd8bdb23b91876ac1e624337bb88dc1dcc21d67e~34 $ git branch base 4703d9119972bf586d2cca76ec6438f819ffa30e $ git switch -c 5.4-renames v5.4 $ git mv drivers pilots # Introduce over 26,000 renames $ git commit -m "Rename drivers/ to pilots/" $ git config merge.renameLimit 30000 $ git config merge.directoryRenames true === Testcases === Now with REBASE standing for either "git rebase [--merge]" (using merge-recursive) or "test-tool fast-rebase" (using merge-ort), the testcases are: Testcase #1: no-renames $ git checkout v5.4^0 $ REBASE --onto HEAD base hwmon-updates Note: technically the name is misleading; there are some renames, but very few. Rename detection only takes about half the overall time. Testcase #2: mega-renames $ git checkout 5.4-renames^0 $ REBASE --onto HEAD base hwmon-updates Testcase #3: just-one-mega $ git checkout 5.4-renames^0 $ REBASE --onto HEAD base hwmon-just-one === Timing results === Overall timings, using hyperfine (1 warmup run, 3 runs for mega-renames, 10 runs for the other two cases): merge-recursive merge-ort no-renames: 18.912 s ± 0.174 s 14.263 s ± 0.053 s mega-renames: 5964.031 s ± 10.459 s 5504.231 s ± 5.150 s just-one-mega: 149.583 s ± 0.751 s 158.534 s ± 0.498 s A single re-run of each with some breakdowns: --- no-renames --- merge-recursive merge-ort overall runtime: 19.302 s 14.257 s inexact rename detection: 7.603 s 7.906 s everything else: 11.699 s 6.351 s --- mega-renames --- merge-recursive merge-ort overall runtime: 5950.195 s 5499.672 s inexact rename detection: 5746.309 s 5487.120 s everything else: 203.886 s 17.552 s --- just-one-mega --- merge-recursive merge-ort overall runtime: 151.001 s 158.582 s inexact rename detection: 143.448 s 157.835 s everything else: 7.553 s 0.747 s === Timing observations === 0) Maximum speedup The "everything else" row represents the maximum speedup we could achieve if we were to somehow infinitely parallelize inexact rename detection, but leave everything else alone. The fact that this is so much smaller than the real runtime (even in the case with virtually no renames) makes it clear just how overwhelmingly large the time spent on rename detection can be. 1) no-renames 1a) merge-ort is faster than merge-recursive, which is nice. However, this still should not be considered good enough. Although the "merge" backend to rebase (merge-recursive) is sometimes faster than the "apply" backend, this is one of those cases where it is not. In fact, even merge-ort is slower. The "apply" backend can complete this testcase in 6.940 s ± 0.485 s which is about 2x faster than merge-ort and 3x faster than merge-recursive. One goal of the merge-ort performance work will be to make it faster than git-am on this (and similar) testcases. 2) mega-renames 2a) Obviously rename detection is a huge cost; it's where most the time is spent. We need to cut that down. If we could somehow infinitely parallelize it and drive its time to 0, the merge-recursive time would drop to about 204s, and the merge-ort time would drop to about 17s. I think this particular stat shows I've subtly baked a couple performance improvements into merge-ort and into fast-rebase already. 3) just-one-mega 3a) not much to say here, it just gives some flavor for how rebasing only one patch compares to rebasing 35. === Goals === This patch is obviously just the beginning. Here are some of my goals that this measurement will help us achieve: * Drive the cost of rename detection down considerably for merges * After the above has been achieved, see if there are other slowness factors (which would have previously been overshadowed by rename detection costs) which we can then focus on and also optimize. * Ensure our rebase testcase that requires little rename detection is noticeably faster with merge-ort than with apply-based rebase. Signed-off-by: Elijah Newren <newren@gmail.com> Acked-by: Taylor Blau <ttaylorr@github.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-01-24 06:01:12 +00:00
trace2_region_leave("diff", "inexact renames", options->repo);
cleanup:
/* At this point, we have found some renames and copies and they
* are recorded in rename_dst. The original list is still in *q.
*/
merge-ort: begin performance work; instrument with trace2_region_* calls Add some timing instrumentation for both merge-ort and diffcore-rename; I used these to measure and optimize performance in both, and several future patch series will build on these to reduce the timings of some select testcases. === Setup === The primary testcase I used involved rebasing a random topic in the linux kernel (consisting of 35 patches) against an older version. I added two variants, one where I rename a toplevel directory, and another where I only rebase one patch instead of the whole topic. The setup is as follows: $ git clone git://git.kernel.org/pub/scm/linux/kernel/git/stable/linux-stable.git $ git branch hwmon-updates fd8bdb23b91876ac1e624337bb88dc1dcc21d67e $ git branch hwmon-just-one fd8bdb23b91876ac1e624337bb88dc1dcc21d67e~34 $ git branch base 4703d9119972bf586d2cca76ec6438f819ffa30e $ git switch -c 5.4-renames v5.4 $ git mv drivers pilots # Introduce over 26,000 renames $ git commit -m "Rename drivers/ to pilots/" $ git config merge.renameLimit 30000 $ git config merge.directoryRenames true === Testcases === Now with REBASE standing for either "git rebase [--merge]" (using merge-recursive) or "test-tool fast-rebase" (using merge-ort), the testcases are: Testcase #1: no-renames $ git checkout v5.4^0 $ REBASE --onto HEAD base hwmon-updates Note: technically the name is misleading; there are some renames, but very few. Rename detection only takes about half the overall time. Testcase #2: mega-renames $ git checkout 5.4-renames^0 $ REBASE --onto HEAD base hwmon-updates Testcase #3: just-one-mega $ git checkout 5.4-renames^0 $ REBASE --onto HEAD base hwmon-just-one === Timing results === Overall timings, using hyperfine (1 warmup run, 3 runs for mega-renames, 10 runs for the other two cases): merge-recursive merge-ort no-renames: 18.912 s ± 0.174 s 14.263 s ± 0.053 s mega-renames: 5964.031 s ± 10.459 s 5504.231 s ± 5.150 s just-one-mega: 149.583 s ± 0.751 s 158.534 s ± 0.498 s A single re-run of each with some breakdowns: --- no-renames --- merge-recursive merge-ort overall runtime: 19.302 s 14.257 s inexact rename detection: 7.603 s 7.906 s everything else: 11.699 s 6.351 s --- mega-renames --- merge-recursive merge-ort overall runtime: 5950.195 s 5499.672 s inexact rename detection: 5746.309 s 5487.120 s everything else: 203.886 s 17.552 s --- just-one-mega --- merge-recursive merge-ort overall runtime: 151.001 s 158.582 s inexact rename detection: 143.448 s 157.835 s everything else: 7.553 s 0.747 s === Timing observations === 0) Maximum speedup The "everything else" row represents the maximum speedup we could achieve if we were to somehow infinitely parallelize inexact rename detection, but leave everything else alone. The fact that this is so much smaller than the real runtime (even in the case with virtually no renames) makes it clear just how overwhelmingly large the time spent on rename detection can be. 1) no-renames 1a) merge-ort is faster than merge-recursive, which is nice. However, this still should not be considered good enough. Although the "merge" backend to rebase (merge-recursive) is sometimes faster than the "apply" backend, this is one of those cases where it is not. In fact, even merge-ort is slower. The "apply" backend can complete this testcase in 6.940 s ± 0.485 s which is about 2x faster than merge-ort and 3x faster than merge-recursive. One goal of the merge-ort performance work will be to make it faster than git-am on this (and similar) testcases. 2) mega-renames 2a) Obviously rename detection is a huge cost; it's where most the time is spent. We need to cut that down. If we could somehow infinitely parallelize it and drive its time to 0, the merge-recursive time would drop to about 204s, and the merge-ort time would drop to about 17s. I think this particular stat shows I've subtly baked a couple performance improvements into merge-ort and into fast-rebase already. 3) just-one-mega 3a) not much to say here, it just gives some flavor for how rebasing only one patch compares to rebasing 35. === Goals === This patch is obviously just the beginning. Here are some of my goals that this measurement will help us achieve: * Drive the cost of rename detection down considerably for merges * After the above has been achieved, see if there are other slowness factors (which would have previously been overshadowed by rename detection costs) which we can then focus on and also optimize. * Ensure our rebase testcase that requires little rename detection is noticeably faster with merge-ort than with apply-based rebase. Signed-off-by: Elijah Newren <newren@gmail.com> Acked-by: Taylor Blau <ttaylorr@github.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-01-24 06:01:12 +00:00
trace2_region_enter("diff", "write back to queue", options->repo);
DIFF_QUEUE_CLEAR(&outq);
for (i = 0; i < q->nr; i++) {
struct diff_filepair *p = q->queue[i];
struct diff_filepair *pair_to_free = NULL;
diffcore-rename: don't consider unmerged path as source Since e9c8409 (diff-index --cached --raw: show tree entry on the LHS for unmerged entries., 2007-01-05), an unmerged entry should be detected by using DIFF_PAIR_UNMERGED(p), not by noticing both one and two sides of the filepair records mode=0 entries. However, it forgot to update some parts of the rename detection logic. This only makes difference in the "diff --cached" codepath where an unmerged filepair carries information on the entries that came from the tree. It probably hasn't been noticed for a long time because nobody would run "diff -M" during a conflict resolution, but "git status" uses rename detection when it internally runs "diff-index" and "diff-files" and gives nonsense results. In an unmerged pair, "one" side can have a valid filespec to record the tree entry (e.g. what's in HEAD) when running "diff --cached". This can be used as a rename source to other paths in the index that are not unmerged. The path that is unmerged by definition does not have the final content yet (i.e. "two" side cannot have a valid filespec), so it can never be a rename destination. Use the DIFF_PAIR_UNMERGED() to detect unmerged filepair correctly, and allow the valid "one" side of an unmerged filepair to be considered a potential rename source, but never to be considered a rename destination. Commit message and first two test cases by Junio, the rest by Martin. Signed-off-by: Martin von Zweigbergk <martin.von.zweigbergk@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2011-03-24 02:41:01 +00:00
if (DIFF_PAIR_UNMERGED(p)) {
diff_q(&outq, p);
}
else if (!DIFF_FILE_VALID(p->one) && DIFF_FILE_VALID(p->two)) {
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 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. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-11 09:08:47 +00:00
/* Creation */
diff_q(&outq, p);
}
else if (DIFF_FILE_VALID(p->one) && !DIFF_FILE_VALID(p->two)) {
/*
* Deletion
*
* We would output this delete record if:
*
* (1) this is a broken delete and the counterpart
* broken create remains in the output; or
* (2) this is not a broken delete, and rename_dst
* does not have a rename/copy to move p->one->path
* out of existence.
*
* Otherwise, the counterpart broken create
* has been turned into a rename-edit; or
* delete did not have a matching create to
* begin with.
*/
if (DIFF_PAIR_BROKEN(p)) {
/* broken delete */
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 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. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-11 09:08:47 +00:00
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;
}
else {
if (p->one->rename_used)
/* this path remains */
pair_to_free = p;
}
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 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. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-11 09:08:47 +00:00
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 */
pair_to_free = p;
if (pair_to_free)
pool_diff_free_filepair(pool, pair_to_free);
}
diff_debug_queue("done copying original", &outq);
free(q->queue);
*q = outq;
diff_debug_queue("done collapsing", q);
for (i = 0; i < rename_dst_nr; i++)
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 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. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-11 09:08:47 +00:00
if (rename_dst[i].filespec_to_free)
pool_free_filespec(pool, rename_dst[i].filespec_to_free);
cleanup_dir_rename_info(&info, dirs_removed, dir_rename_count != NULL);
FREE_AND_NULL(rename_dst);
rename_dst_nr = rename_dst_alloc = 0;
FREE_AND_NULL(rename_src);
rename_src_nr = rename_src_alloc = 0;
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 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. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-11 09:08:47 +00:00
if (break_idx) {
strintmap_clear(break_idx);
FREE_AND_NULL(break_idx);
}
merge-ort: begin performance work; instrument with trace2_region_* calls Add some timing instrumentation for both merge-ort and diffcore-rename; I used these to measure and optimize performance in both, and several future patch series will build on these to reduce the timings of some select testcases. === Setup === The primary testcase I used involved rebasing a random topic in the linux kernel (consisting of 35 patches) against an older version. I added two variants, one where I rename a toplevel directory, and another where I only rebase one patch instead of the whole topic. The setup is as follows: $ git clone git://git.kernel.org/pub/scm/linux/kernel/git/stable/linux-stable.git $ git branch hwmon-updates fd8bdb23b91876ac1e624337bb88dc1dcc21d67e $ git branch hwmon-just-one fd8bdb23b91876ac1e624337bb88dc1dcc21d67e~34 $ git branch base 4703d9119972bf586d2cca76ec6438f819ffa30e $ git switch -c 5.4-renames v5.4 $ git mv drivers pilots # Introduce over 26,000 renames $ git commit -m "Rename drivers/ to pilots/" $ git config merge.renameLimit 30000 $ git config merge.directoryRenames true === Testcases === Now with REBASE standing for either "git rebase [--merge]" (using merge-recursive) or "test-tool fast-rebase" (using merge-ort), the testcases are: Testcase #1: no-renames $ git checkout v5.4^0 $ REBASE --onto HEAD base hwmon-updates Note: technically the name is misleading; there are some renames, but very few. Rename detection only takes about half the overall time. Testcase #2: mega-renames $ git checkout 5.4-renames^0 $ REBASE --onto HEAD base hwmon-updates Testcase #3: just-one-mega $ git checkout 5.4-renames^0 $ REBASE --onto HEAD base hwmon-just-one === Timing results === Overall timings, using hyperfine (1 warmup run, 3 runs for mega-renames, 10 runs for the other two cases): merge-recursive merge-ort no-renames: 18.912 s ± 0.174 s 14.263 s ± 0.053 s mega-renames: 5964.031 s ± 10.459 s 5504.231 s ± 5.150 s just-one-mega: 149.583 s ± 0.751 s 158.534 s ± 0.498 s A single re-run of each with some breakdowns: --- no-renames --- merge-recursive merge-ort overall runtime: 19.302 s 14.257 s inexact rename detection: 7.603 s 7.906 s everything else: 11.699 s 6.351 s --- mega-renames --- merge-recursive merge-ort overall runtime: 5950.195 s 5499.672 s inexact rename detection: 5746.309 s 5487.120 s everything else: 203.886 s 17.552 s --- just-one-mega --- merge-recursive merge-ort overall runtime: 151.001 s 158.582 s inexact rename detection: 143.448 s 157.835 s everything else: 7.553 s 0.747 s === Timing observations === 0) Maximum speedup The "everything else" row represents the maximum speedup we could achieve if we were to somehow infinitely parallelize inexact rename detection, but leave everything else alone. The fact that this is so much smaller than the real runtime (even in the case with virtually no renames) makes it clear just how overwhelmingly large the time spent on rename detection can be. 1) no-renames 1a) merge-ort is faster than merge-recursive, which is nice. However, this still should not be considered good enough. Although the "merge" backend to rebase (merge-recursive) is sometimes faster than the "apply" backend, this is one of those cases where it is not. In fact, even merge-ort is slower. The "apply" backend can complete this testcase in 6.940 s ± 0.485 s which is about 2x faster than merge-ort and 3x faster than merge-recursive. One goal of the merge-ort performance work will be to make it faster than git-am on this (and similar) testcases. 2) mega-renames 2a) Obviously rename detection is a huge cost; it's where most the time is spent. We need to cut that down. If we could somehow infinitely parallelize it and drive its time to 0, the merge-recursive time would drop to about 204s, and the merge-ort time would drop to about 17s. I think this particular stat shows I've subtly baked a couple performance improvements into merge-ort and into fast-rebase already. 3) just-one-mega 3a) not much to say here, it just gives some flavor for how rebasing only one patch compares to rebasing 35. === Goals === This patch is obviously just the beginning. Here are some of my goals that this measurement will help us achieve: * Drive the cost of rename detection down considerably for merges * After the above has been achieved, see if there are other slowness factors (which would have previously been overshadowed by rename detection costs) which we can then focus on and also optimize. * Ensure our rebase testcase that requires little rename detection is noticeably faster with merge-ort than with apply-based rebase. Signed-off-by: Elijah Newren <newren@gmail.com> Acked-by: Taylor Blau <ttaylorr@github.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-01-24 06:01:12 +00:00
trace2_region_leave("diff", "write back to queue", options->repo);
return;
}
void diffcore_rename(struct diff_options *options)
{
diffcore_rename_extended(options, NULL, NULL, NULL, NULL, NULL);
}