bdiff: early pruning of common prefix before doing expensive computations
authorMads Kiilerich <madski@unity3d.com>
Wed, 16 Nov 2016 19:45:35 +0100
changeset 30566 7c0c722d568d
parent 30565 783016005122
child 30567 b3a9ef3d30e8
bdiff: early pruning of common prefix before doing expensive computations It seems quite common that files don't change completely. New lines are often pretty much appended, and modifications will often only change a small section of the file which on average will be in the middle. There can thus be a big win by pruning a common prefix before starting the more expensive search for longest common substrings. Worst case, it will scan through a long sequence of similar bytes without encountering a newline. Splitlines will then have to do the same again ... twice for each side. If similar lines are found, splitlines will save the double iteration and hashing of the lines ... plus there will be less lines to find common substrings in. This change might in some cases make the algorith pick shorter or less optimal common substrings. We can't have the cake and eat it. This make hg --time bundle --base null -r 4.0 go from 14.5 to 15 s - a 3% increase. On mozilla-unified: perfbdiff -m 3041e4d59df2 ! wall 0.053088 comb 0.060000 user 0.060000 sys 0.000000 (best of 100) to ! wall 0.024618 comb 0.020000 user 0.020000 sys 0.000000 (best of 116) perfbdiff 0e9928989e9c --alldata --count 10 ! wall 0.702075 comb 0.700000 user 0.700000 sys 0.000000 (best of 15) to ! wall 0.579235 comb 0.580000 user 0.580000 sys 0.000000 (best of 18)
mercurial/bdiff_module.c
tests/test-bdiff.py.out
--- a/mercurial/bdiff_module.c	Sat Oct 22 15:02:11 2016 +0900
+++ b/mercurial/bdiff_module.c	Wed Nov 16 19:45:35 2016 +0100
@@ -61,12 +61,12 @@
 
 static PyObject *bdiff(PyObject *self, PyObject *args)
 {
-	char *sa, *sb, *rb;
+	char *sa, *sb, *rb, *ia, *ib;
 	PyObject *result = NULL;
 	struct bdiff_line *al, *bl;
 	struct bdiff_hunk l, *h;
 	int an, bn, count;
-	Py_ssize_t len = 0, la, lb;
+	Py_ssize_t len = 0, la, lb, li = 0, lcommon = 0, lmax;
 	PyThreadState *_save;
 
 	l.next = NULL;
@@ -80,8 +80,17 @@
 	}
 
 	_save = PyEval_SaveThread();
-	an = bdiff_splitlines(sa, la, &al);
-	bn = bdiff_splitlines(sb, lb, &bl);
+
+	lmax = la > lb ? lb : la;
+	for (ia = sa, ib = sb;
+	     li < lmax && *ia == *ib;
+	     ++li, ++ia, ++ib)
+		if (*ia == '\n')
+			lcommon = li + 1;
+	/* we can almost add: if (li == lmax) lcommon = li; */
+
+	an = bdiff_splitlines(sa + lcommon, la - lcommon, &al);
+	bn = bdiff_splitlines(sb + lcommon, lb - lcommon, &bl);
 	if (!al || !bl)
 		goto nomem;
 
@@ -112,8 +121,8 @@
 	for (h = l.next; h; h = h->next) {
 		if (h->a1 != la || h->b1 != lb) {
 			len = bl[h->b1].l - bl[lb].l;
-			putbe32((uint32_t)(al[la].l - al->l), rb);
-			putbe32((uint32_t)(al[h->a1].l - al->l), rb + 4);
+			putbe32((uint32_t)(al[la].l + lcommon - al->l), rb);
+			putbe32((uint32_t)(al[h->a1].l + lcommon - al->l), rb + 4);
 			putbe32((uint32_t)len, rb + 8);
 			memcpy(rb + 12, bl[lb].l, len);
 			rb += 12 + len;
--- a/tests/test-bdiff.py.out	Sat Oct 22 15:02:11 2016 +0900
+++ b/tests/test-bdiff.py.out	Wed Nov 16 19:45:35 2016 +0100
@@ -34,10 +34,9 @@
 showdiff(
   'a\nb\nb\nb\nc\n.\nd\ne\n.\nf\n',
   'a\nb\nb\na\nb\nb\nb\nc\n.\nb\nc\n.\nd\ne\nf\n'):
- 0 0 '' -> 'a\nb\nb\n'
- 'a\nb\nb\nb\nc\n.\n'
- 12 12 '' -> 'b\nc\n.\n'
- 'd\ne\n'
+ 'a\nb\nb\n'
+ 6 6 '' -> 'a\nb\nb\nb\nc\n.\n'
+ 'b\nc\n.\nd\ne\n'
  16 18 '.\n' -> ''
  'f\n'
 done