Mercurial > hg
view mercurial/bdiff.c @ 39627:ee7ee0c516ca
narrow: intersect provided matcher with narrowmatcher in `hg diff`
This provides significant speedups when running diff, and no change in behavior
that I'm aware of (or that the tests found). I tested with a repo that I started
using narrow in after it was created and attempted to run `hg diff -c .` and
similar commands in it on a commit that had files not in the narrowspec.
Timing numbers below, using a similar setup as my previous commits.
before=9db85644, m-u is mozilla-unified at eb39298e432d (flatmanifest) and
0553b7f29eaf (treemanifest). l-d-r is a repo simulating a situation I've
encountered where there's one directory with 30k+ subdirectories. N means
narrow, T means treemanifest. The narrowspec is pretty small when in use, and
importantly the narrowspec is applied *after* doing the initial checkout
(without narrowing), so all of these files exist in the filesystem, which is not
normally the case if someone has been using narrow for the entire life of the
clone.
Anything less than a 5% difference in performance is most likely noise.
diff --git:
repo | N | T | before (mean +- stdev) | after (mean +- stdev) | % of before
------+---+---+------------------------+-----------------------+------------
m-u | | | 1.292 s +- 0.009 s | 1.295 s +- 0.010 s | 100.2%
m-u | | x | 1.296 s +- 0.042 s | 1.299 s +- 0.026 s | 100.2%
m-u | x | | 1.292 s +- 0.010 s | 1.297 s +- 0.021 s | 100.4%
m-u | x | x | 84.2 ms +- 1.2 ms | 83.6 ms +- 0.2 ms | 99.3%
l-d-r | | | 188.7 ms +- 2.7 ms | 188.8 ms +- 2.0 ms | 100.1%
l-d-r | | x | 189.9 ms +- 1.5 ms | 189.4 ms +- 1.2 ms | 99.7%
l-d-r | x | | 97.1 ms +- 1.0 ms | 87.1 ms +- 1.0 ms | 89.7% <--
l-d-r | x | x | 96.9 ms +- 0.8 ms | 87.2 ms +- 0.7 ms | 90.0% <--
diff -c . --git:
repo | N | T | before (mean +- stdev) | after (mean +- stdev) | % of before
------+---+---+------------------------+-----------------------+------------
m-u | | | 231.6 ms +- 3.1 ms | 228.9 ms +- 1.6 ms | 98.8%
m-u | | x | 150.5 ms +- 1.7 ms | 150.7 ms +- 1.4 ms | 100.1%
m-u | x | | 233.7 ms +- 2.4 ms | 232.2 ms +- 1.9 ms | 99.4%
m-u | x | x | 126.1 ms +- 1.2 ms | 126.8 ms +- 1.2 ms | 100.6%
l-d-r | | | 82.1 ms +- 2.0 ms | 81.8 ms +- 1.4 ms | 99.6%
l-d-r | | x | 3.732 s +- 0.020 s | 3.746 s +- 0.027 s | 100.4%
l-d-r | x | | 83.1 ms +- 0.8 ms | 107.6 ms +- 2.4 ms | 129.5% <--
l-d-r | x | x | 758.2 ms +- 38.8 ms | 188.5 ms +- 1.8 ms | 24.9% <--
rebase -r . --keep -d .^^:
repo | N | T | before (mean +- stdev) | after (mean +- stdev) | % of before
------+---+---+------------------------+-----------------------+------------
m-u | | | 5.532 s +- 0.087 s | 5.496 s +- 0.016 s | 99.3%
m-u | | x | 5.554 s +- 0.061 s | 5.532 s +- 0.013 s | 99.6%
m-u | x | | 5.602 s +- 0.134 s | 5.508 s +- 0.035 s | 98.3%
m-u | x | x | 582.2 ms +- 15.2 ms | 572.9 ms +- 12.0 ms | 98.4%
l-d-r | | | 629.5 ms +- 12.3 ms | 622.5 ms +- 7.3 ms | 98.9%
l-d-r | | x | 6.173 s +- 0.062 s | 6.185 s +- 0.076 s | 100.2%
l-d-r | x | | 274.5 ms +- 10.0 ms | 272.1 ms +- 6.2 ms | 99.1%
l-d-r | x | x | 4.835 s +- 0.056 s | 4.826 s +- 0.034 s | 99.8%
status --change . --copies:
repo | N | T | before (mean +- stdev) | after (mean +- stdev) | % of before
------+---+---+------------------------+-----------------------+------------
m-u | | | 214.4 ms +- 1.4 ms | 212.2 ms +- 1.7 ms | 99.0%
m-u | | x | 130.9 ms +- 1.2 ms | 131.7 ms +- 1.1 ms | 100.6%
m-u | x | | 215.0 ms +- 2.1 ms | 214.9 ms +- 2.7 ms | 100.0%
m-u | x | x | 109.5 ms +- 2.3 ms | 107.8 ms +- 0.9 ms | 98.4%
l-d-r | | | 79.6 ms +- 0.9 ms | 79.8 ms +- 1.6 ms | 100.3%
l-d-r | | x | 3.799 s +- 0.037 s | 3.928 s +- 0.021 s | 103.4% <--?
l-d-r | x | | 82.7 ms +- 0.7 ms | 83.2 ms +- 1.0 ms | 100.6%
l-d-r | x | x | 746.8 ms +- 6.1 ms | 739.0 ms +- 4.2 ms | 99.0%
status --copies:
repo | N | T | before (mean +- stdev) | after (mean +- stdev) | % of before
------+---+---+------------------------+-----------------------+------------
m-u | | | 1.884 s +- 0.012 s | 1.885 s +- 0.013 s | 100.1%
m-u | | x | 1.897 s +- 0.027 s | 1.909 s +- 0.077 s | 100.6%
m-u | x | | 1.886 s +- 0.021 s | 1.891 s +- 0.030 s | 100.3%
m-u | x | x | 92.0 ms +- 0.7 ms | 92.4 ms +- 0.4 ms | 100.4%
l-d-r | | | 570.3 ms +- 18.7 ms | 552.2 ms +- 4.5 ms | 96.8%
l-d-r | | x | 568.9 ms +- 16.1 ms | 567.2 ms +- 11.9 ms | 99.7%
l-d-r | x | | 171.1 ms +- 2.5 ms | 170.4 ms +- 1.2 ms | 99.6%
l-d-r | x | x | 171.6 ms +- 3.4 ms | 171.5 ms +- 1.7 ms | 99.9%
update $rev^; ~/src/hg/hg{hg}/hg update $rev:
repo | N | T | before (mean +- stdev) | after (mean +- stdev) | % of before
------+---+---+------------------------+-----------------------+------------
m-u | | | 3.107 s +- 0.017 s | 3.116 s +- 0.012 s | 100.3%
m-u | | x | 2.943 s +- 0.010 s | 2.945 s +- 0.019 s | 100.1%
m-u | x | | 3.116 s +- 0.033 s | 3.118 s +- 0.027 s | 100.1%
m-u | x | x | 318.5 ms +- 2.7 ms | 320.8 ms +- 4.8 ms | 100.7%
l-d-r | | | 428.9 ms +- 4.4 ms | 429.5 ms +- 4.0 ms | 100.1%
l-d-r | | x | 9.593 s +- 0.081 s | 9.869 s +- 0.043 s | 102.9%
l-d-r | x | | 253.2 ms +- 3.6 ms | 254.0 ms +- 2.8 ms | 100.3%
l-d-r | x | x | 1.613 s +- 0.009 s | 1.630 s +- 0.017 s | 101.1%
Differential Revision: https://phab.mercurial-scm.org/D4587
author | spectral <spectral@google.com> |
---|---|
date | Thu, 13 Sep 2018 22:48:27 -0700 |
parents | 068e774ae29e |
children | 763b45bc4483 |
line wrap: on
line source
/* bdiff.c - efficient binary diff extension for Mercurial Copyright 2005, 2006 Matt Mackall <mpm@selenic.com> This software may be used and distributed according to the terms of the GNU General Public License, incorporated herein by reference. Based roughly on Python difflib */ #include <limits.h> #include <stdlib.h> #include <string.h> #include "bdiff.h" #include "bitmanipulation.h" #include "compat.h" /* Hash implementation from diffutils */ #define ROL(v, n) ((v) << (n) | (v) >> (sizeof(v) * CHAR_BIT - (n))) #define HASH(h, c) ((c) + ROL(h, 7)) struct pos { int pos, len; }; int bdiff_splitlines(const char *a, ssize_t len, struct bdiff_line **lr) { unsigned hash; int i; const char *p, *b = a; const char *const plast = a + len - 1; struct bdiff_line *l; /* count the lines */ i = 1; /* extra line for sentinel */ for (p = a; p < plast; p++) if (*p == '\n') i++; if (p == plast) i++; *lr = l = (struct bdiff_line *)calloc(i, sizeof(struct bdiff_line)); if (!l) return -1; /* build the line array and calculate hashes */ hash = 0; for (p = a; p < plast; p++) { hash = HASH(hash, *p); if (*p == '\n') { l->hash = hash; hash = 0; l->len = p - b + 1; l->l = b; l->n = INT_MAX; l++; b = p + 1; } } if (p == plast) { hash = HASH(hash, *p); l->hash = hash; l->len = p - b + 1; l->l = b; l->n = INT_MAX; l++; } /* set up a sentinel */ l->hash = 0; l->len = 0; l->l = a + len; return i - 1; } static inline int cmp(struct bdiff_line *a, struct bdiff_line *b) { return a->hash != b->hash || a->len != b->len || memcmp(a->l, b->l, a->len); } static int equatelines(struct bdiff_line *a, int an, struct bdiff_line *b, int bn) { int i, j, buckets = 1, t, scale; struct pos *h = NULL; /* build a hash table of the next highest power of 2 */ while (buckets < bn + 1) buckets *= 2; /* try to allocate a large hash table to avoid collisions */ for (scale = 4; scale; scale /= 2) { h = (struct pos *)calloc(buckets, scale * sizeof(struct pos)); if (h) break; } if (!h) return 0; buckets = buckets * scale - 1; /* clear the hash table */ for (i = 0; i <= buckets; i++) { h[i].pos = -1; h[i].len = 0; } /* add lines to the hash table chains */ for (i = 0; i < bn; i++) { /* find the equivalence class */ for (j = b[i].hash & buckets; h[j].pos != -1; j = (j + 1) & buckets) if (!cmp(b + i, b + h[j].pos)) break; /* add to the head of the equivalence class */ b[i].n = h[j].pos; b[i].e = j; h[j].pos = i; h[j].len++; /* keep track of popularity */ } /* compute popularity threshold */ t = (bn >= 31000) ? bn / 1000 : 1000000 / (bn + 1); /* match items in a to their equivalence class in b */ for (i = 0; i < an; i++) { /* find the equivalence class */ for (j = a[i].hash & buckets; h[j].pos != -1; j = (j + 1) & buckets) if (!cmp(a + i, b + h[j].pos)) break; a[i].e = j; /* use equivalence class for quick compare */ if (h[j].len <= t) a[i].n = h[j].pos; /* point to head of match list */ else a[i].n = -1; /* too popular */ } /* discard hash tables */ free(h); return 1; } static int longest_match(struct bdiff_line *a, struct bdiff_line *b, struct pos *pos, int a1, int a2, int b1, int b2, int *omi, int *omj) { int mi = a1, mj = b1, mk = 0, i, j, k, half, bhalf; /* window our search on large regions to better bound worst-case performance. by choosing a window at the end, we reduce skipping overhead on the b chains. */ if (a2 - a1 > 30000) a1 = a2 - 30000; half = (a1 + a2 - 1) / 2; bhalf = (b1 + b2 - 1) / 2; for (i = a1; i < a2; i++) { /* skip all lines in b after the current block */ for (j = a[i].n; j >= b2; j = b[j].n) ; /* loop through all lines match a[i] in b */ for (; j >= b1; j = b[j].n) { /* does this extend an earlier match? */ for (k = 1; j - k >= b1 && i - k >= a1; k++) { /* reached an earlier match? */ if (pos[j - k].pos == i - k) { k += pos[j - k].len; break; } /* previous line mismatch? */ if (a[i - k].e != b[j - k].e) break; } pos[j].pos = i; pos[j].len = k; /* best match so far? we prefer matches closer to the middle to balance recursion */ if (k > mk) { /* a longer match */ mi = i; mj = j; mk = k; } else if (k == mk) { if (i > mi && i <= half && j > b1) { /* same match but closer to half */ mi = i; mj = j; } else if (i == mi && (mj > bhalf || i == a1)) { /* same i but best earlier j */ mj = j; } } } } if (mk) { mi = mi - mk + 1; mj = mj - mk + 1; } /* expand match to include subsequent popular lines */ while (mi + mk < a2 && mj + mk < b2 && a[mi + mk].e == b[mj + mk].e) mk++; *omi = mi; *omj = mj; return mk; } static struct bdiff_hunk *recurse(struct bdiff_line *a, struct bdiff_line *b, struct pos *pos, int a1, int a2, int b1, int b2, struct bdiff_hunk *l) { int i, j, k; while (1) { /* find the longest match in this chunk */ k = longest_match(a, b, pos, a1, a2, b1, b2, &i, &j); if (!k) return l; /* and recurse on the remaining chunks on either side */ l = recurse(a, b, pos, a1, i, b1, j, l); if (!l) return NULL; l->next = (struct bdiff_hunk *)malloc(sizeof(struct bdiff_hunk)); if (!l->next) return NULL; l = l->next; l->a1 = i; l->a2 = i + k; l->b1 = j; l->b2 = j + k; l->next = NULL; /* tail-recursion didn't happen, so do equivalent iteration */ a1 = i + k; b1 = j + k; } } int bdiff_diff(struct bdiff_line *a, int an, struct bdiff_line *b, int bn, struct bdiff_hunk *base) { struct bdiff_hunk *curr; struct pos *pos; int t, count = 0; /* allocate and fill arrays */ t = equatelines(a, an, b, bn); pos = (struct pos *)calloc(bn ? bn : 1, sizeof(struct pos)); if (pos && t) { /* generate the matching block list */ curr = recurse(a, b, pos, 0, an, 0, bn, base); if (!curr) return -1; /* sentinel end hunk */ curr->next = (struct bdiff_hunk *)malloc(sizeof(struct bdiff_hunk)); if (!curr->next) return -1; curr = curr->next; curr->a1 = curr->a2 = an; curr->b1 = curr->b2 = bn; curr->next = NULL; } free(pos); /* normalize the hunk list, try to push each hunk towards the end */ for (curr = base->next; curr; curr = curr->next) { struct bdiff_hunk *next = curr->next; if (!next) break; if (curr->a2 == next->a1 || curr->b2 == next->b1) while (curr->a2 < an && curr->b2 < bn && next->a1 < next->a2 && next->b1 < next->b2 && !cmp(a + curr->a2, b + curr->b2)) { curr->a2++; next->a1++; curr->b2++; next->b1++; } } for (curr = base->next; curr; curr = curr->next) count++; return count; } /* deallocate list of hunks; l may be NULL */ void bdiff_freehunks(struct bdiff_hunk *l) { struct bdiff_hunk *n; for (; l; l = n) { n = l->next; free(l); } }