Merge branch 'ew/fd-cloexec-fix' into maint
[gitweb.git] / xdiff / xdiffi.c
index b3c684887502c3c473a614114e4acdfbc57ea094..93a65680a18284041ce57f61191d20dcc763e022 100644 (file)
@@ -400,138 +400,544 @@ static xdchange_t *xdl_add_change(xdchange_t *xscr, long i1, long i2, long chg1,
 }
 
 
-static int is_blank_line(xrecord_t **recs, long ix, long flags)
+static int recs_match(xrecord_t *rec1, xrecord_t *rec2, long flags)
 {
-       return xdl_blankline(recs[ix]->ptr, recs[ix]->size, flags);
+       return (rec1->ha == rec2->ha &&
+               xdl_recmatch(rec1->ptr, rec1->size,
+                            rec2->ptr, rec2->size,
+                            flags));
 }
 
-static int recs_match(xrecord_t **recs, long ixs, long ix, long flags)
+/*
+ * If a line is indented more than this, get_indent() just returns this value.
+ * This avoids having to do absurd amounts of work for data that are not
+ * human-readable text, and also ensures that the output of get_indent fits within
+ * an int.
+ */
+#define MAX_INDENT 200
+
+/*
+ * Return the amount of indentation of the specified line, treating TAB as 8
+ * columns. Return -1 if line is empty or contains only whitespace. Clamp the
+ * output value at MAX_INDENT.
+ */
+static int get_indent(xrecord_t *rec)
 {
-       return (recs[ixs]->ha == recs[ix]->ha &&
-               xdl_recmatch(recs[ixs]->ptr, recs[ixs]->size,
-                            recs[ix]->ptr, recs[ix]->size,
-                            flags));
+       long i;
+       int ret = 0;
+
+       for (i = 0; i < rec->size; i++) {
+               char c = rec->ptr[i];
+
+               if (!XDL_ISSPACE(c))
+                       return ret;
+               else if (c == ' ')
+                       ret += 1;
+               else if (c == '\t')
+                       ret += 8 - ret % 8;
+               /* ignore other whitespace characters */
+
+               if (ret >= MAX_INDENT)
+                       return MAX_INDENT;
+       }
+
+       /* The line contains only whitespace. */
+       return -1;
 }
 
-int xdl_change_compact(xdfile_t *xdf, xdfile_t *xdfo, long flags) {
-       long ix, ixo, ixs, ixref, grpsiz, nrec = xdf->nrec;
-       char *rchg = xdf->rchg, *rchgo = xdfo->rchg;
-       unsigned int blank_lines;
-       xrecord_t **recs = xdf->recs;
+/*
+ * If more than this number of consecutive blank rows are found, just return this
+ * value. This avoids requiring O(N^2) work for pathological cases, and also
+ * ensures that the output of score_split fits in an int.
+ */
+#define MAX_BLANKS 20
 
+/* Characteristics measured about a hypothetical split position. */
+struct split_measurement {
        /*
-        * This is the same of what GNU diff does. Move back and forward
-        * change groups for a consistent and pretty diff output. This also
-        * helps in finding joinable change groups and reduce the diff size.
+        * Is the split at the end of the file (aside from any blank lines)?
         */
-       for (ix = ixo = 0;;) {
-               /*
-                * Find the first changed line in the to-be-compacted file.
-                * We need to keep track of both indexes, so if we find a
-                * changed lines group on the other file, while scanning the
-                * to-be-compacted file, we need to skip it properly. Note
-                * that loops that are testing for changed lines on rchg* do
-                * not need index bounding since the array is prepared with
-                * a zero at position -1 and N.
-                */
-               for (; ix < nrec && !rchg[ix]; ix++)
-                       while (rchgo[ixo++]);
-               if (ix == nrec)
+       int end_of_file;
+
+       /*
+        * How much is the line immediately following the split indented (or -1 if
+        * the line is blank):
+        */
+       int indent;
+
+       /*
+        * How many consecutive lines above the split are blank?
+        */
+       int pre_blank;
+
+       /*
+        * How much is the nearest non-blank line above the split indented (or -1
+        * if there is no such line)?
+        */
+       int pre_indent;
+
+       /*
+        * How many lines after the line following the split are blank?
+        */
+       int post_blank;
+
+       /*
+        * How much is the nearest non-blank line after the line following the
+        * split indented (or -1 if there is no such line)?
+        */
+       int post_indent;
+};
+
+struct split_score {
+       /* The effective indent of this split (smaller is preferred). */
+       int effective_indent;
+
+       /* Penalty for this split (smaller is preferred). */
+       int penalty;
+};
+
+/*
+ * Fill m with information about a hypothetical split of xdf above line split.
+ */
+static void measure_split(const xdfile_t *xdf, long split,
+                         struct split_measurement *m)
+{
+       long i;
+
+       if (split >= xdf->nrec) {
+               m->end_of_file = 1;
+               m->indent = -1;
+       } else {
+               m->end_of_file = 0;
+               m->indent = get_indent(xdf->recs[split]);
+       }
+
+       m->pre_blank = 0;
+       m->pre_indent = -1;
+       for (i = split - 1; i >= 0; i--) {
+               m->pre_indent = get_indent(xdf->recs[i]);
+               if (m->pre_indent != -1)
                        break;
+               m->pre_blank += 1;
+               if (m->pre_blank == MAX_BLANKS) {
+                       m->pre_indent = 0;
+                       break;
+               }
+       }
+
+       m->post_blank = 0;
+       m->post_indent = -1;
+       for (i = split + 1; i < xdf->nrec; i++) {
+               m->post_indent = get_indent(xdf->recs[i]);
+               if (m->post_indent != -1)
+                       break;
+               m->post_blank += 1;
+               if (m->post_blank == MAX_BLANKS) {
+                       m->post_indent = 0;
+                       break;
+               }
+       }
+}
+
+/*
+ * The empirically-determined weight factors used by score_split() below.
+ * Larger values means that the position is a less favorable place to split.
+ *
+ * Note that scores are only ever compared against each other, so multiplying
+ * all of these weight/penalty values by the same factor wouldn't change the
+ * heuristic's behavior. Still, we need to set that arbitrary scale *somehow*.
+ * In practice, these numbers are chosen to be large enough that they can be
+ * adjusted relative to each other with sufficient precision despite using
+ * integer math.
+ */
+
+/* Penalty if there are no non-blank lines before the split */
+#define START_OF_FILE_PENALTY 1
+
+/* Penalty if there are no non-blank lines after the split */
+#define END_OF_FILE_PENALTY 21
+
+/* Multiplier for the number of blank lines around the split */
+#define TOTAL_BLANK_WEIGHT (-30)
+
+/* Multiplier for the number of blank lines after the split */
+#define POST_BLANK_WEIGHT 6
+
+/*
+ * Penalties applied if the line is indented more than its predecessor
+ */
+#define RELATIVE_INDENT_PENALTY (-4)
+#define RELATIVE_INDENT_WITH_BLANK_PENALTY 10
 
+/*
+ * Penalties applied if the line is indented less than both its predecessor and
+ * its successor
+ */
+#define RELATIVE_OUTDENT_PENALTY 24
+#define RELATIVE_OUTDENT_WITH_BLANK_PENALTY 17
+
+/*
+ * Penalties applied if the line is indented less than its predecessor but not
+ * less than its successor
+ */
+#define RELATIVE_DEDENT_PENALTY 23
+#define RELATIVE_DEDENT_WITH_BLANK_PENALTY 17
+
+/*
+ * We only consider whether the sum of the effective indents for splits are
+ * less than (-1), equal to (0), or greater than (+1) each other. The resulting
+ * value is multiplied by the following weight and combined with the penalty to
+ * determine the better of two scores.
+ */
+#define INDENT_WEIGHT 60
+
+/*
+ * Compute a badness score for the hypothetical split whose measurements are
+ * stored in m. The weight factors were determined empirically using the tools and
+ * corpus described in
+ *
+ *     https://github.com/mhagger/diff-slider-tools
+ *
+ * Also see that project if you want to improve the weights based on, for example,
+ * a larger or more diverse corpus.
+ */
+static void score_add_split(const struct split_measurement *m, struct split_score *s)
+{
+       /*
+        * A place to accumulate penalty factors (positive makes this index more
+        * favored):
+        */
+       int post_blank, total_blank, indent, any_blanks;
+
+       if (m->pre_indent == -1 && m->pre_blank == 0)
+               s->penalty += START_OF_FILE_PENALTY;
+
+       if (m->end_of_file)
+               s->penalty += END_OF_FILE_PENALTY;
+
+       /*
+        * Set post_blank to the number of blank lines following the split,
+        * including the line immediately after the split:
+        */
+       post_blank = (m->indent == -1) ? 1 + m->post_blank : 0;
+       total_blank = m->pre_blank + post_blank;
+
+       /* Penalties based on nearby blank lines: */
+       s->penalty += TOTAL_BLANK_WEIGHT * total_blank;
+       s->penalty += POST_BLANK_WEIGHT * post_blank;
+
+       if (m->indent != -1)
+               indent = m->indent;
+       else
+               indent = m->post_indent;
+
+       any_blanks = (total_blank != 0);
+
+       /* Note that the effective indent is -1 at the end of the file: */
+       s->effective_indent += indent;
+
+       if (indent == -1) {
+               /* No additional adjustments needed. */
+       } else if (m->pre_indent == -1) {
+               /* No additional adjustments needed. */
+       } else if (indent > m->pre_indent) {
+               /*
+                * The line is indented more than its predecessor.
+                */
+               s->penalty += any_blanks ?
+                       RELATIVE_INDENT_WITH_BLANK_PENALTY :
+                       RELATIVE_INDENT_PENALTY;
+       } else if (indent == m->pre_indent) {
                /*
-                * Record the start of a changed-group in the to-be-compacted file
-                * and find the end of it, on both to-be-compacted and other file
-                * indexes (ix and ixo).
+                * The line has the same indentation level as its predecessor.
+                * No additional adjustments needed.
                 */
-               ixs = ix;
-               for (ix++; rchg[ix]; ix++);
-               for (; rchgo[ixo]; ixo++);
+       } else {
+               /*
+                * The line is indented less than its predecessor. It could be
+                * the block terminator of the previous block, but it could
+                * also be the start of a new block (e.g., an "else" block, or
+                * maybe the previous block didn't have a block terminator).
+                * Try to distinguish those cases based on what comes next:
+                */
+               if (m->post_indent != -1 && m->post_indent > indent) {
+                       /*
+                        * The following line is indented more. So it is likely
+                        * that this line is the start of a block.
+                        */
+                       s->penalty += any_blanks ?
+                               RELATIVE_OUTDENT_WITH_BLANK_PENALTY :
+                               RELATIVE_OUTDENT_PENALTY;
+               } else {
+                       /*
+                        * That was probably the end of a block.
+                        */
+                       s->penalty += any_blanks ?
+                               RELATIVE_DEDENT_WITH_BLANK_PENALTY :
+                               RELATIVE_DEDENT_PENALTY;
+               }
+       }
+}
+
+static int score_cmp(struct split_score *s1, struct split_score *s2)
+{
+       /* -1 if s1.effective_indent < s2->effective_indent, etc. */
+       int cmp_indents = ((s1->effective_indent > s2->effective_indent) -
+                          (s1->effective_indent < s2->effective_indent));
+
+       return INDENT_WEIGHT * cmp_indents + (s1->penalty - s2->penalty);
+}
+
+/*
+ * Represent a group of changed lines in an xdfile_t (i.e., a contiguous group
+ * of lines that was inserted or deleted from the corresponding version of the
+ * file). We consider there to be such a group at the beginning of the file, at
+ * the end of the file, and between any two unchanged lines, though most such
+ * groups will usually be empty.
+ *
+ * If the first line in a group is equal to the line following the group, then
+ * the group can be slid down. Similarly, if the last line in a group is equal
+ * to the line preceding the group, then the group can be slid up. See
+ * group_slide_down() and group_slide_up().
+ *
+ * Note that loops that are testing for changed lines in xdf->rchg do not need
+ * index bounding since the array is prepared with a zero at position -1 and N.
+ */
+struct xdlgroup {
+       /*
+        * The index of the first changed line in the group, or the index of
+        * the unchanged line above which the (empty) group is located.
+        */
+       long start;
+
+       /*
+        * The index of the first unchanged line after the group. For an empty
+        * group, end is equal to start.
+        */
+       long end;
+};
+
+/*
+ * Initialize g to point at the first group in xdf.
+ */
+static void group_init(xdfile_t *xdf, struct xdlgroup *g)
+{
+       g->start = g->end = 0;
+       while (xdf->rchg[g->end])
+               g->end++;
+}
 
+/*
+ * Move g to describe the next (possibly empty) group in xdf and return 0. If g
+ * is already at the end of the file, do nothing and return -1.
+ */
+static inline int group_next(xdfile_t *xdf, struct xdlgroup *g)
+{
+       if (g->end == xdf->nrec)
+               return -1;
+
+       g->start = g->end + 1;
+       for (g->end = g->start; xdf->rchg[g->end]; g->end++)
+               ;
+
+       return 0;
+}
+
+/*
+ * Move g to describe the previous (possibly empty) group in xdf and return 0.
+ * If g is already at the beginning of the file, do nothing and return -1.
+ */
+static inline int group_previous(xdfile_t *xdf, struct xdlgroup *g)
+{
+       if (g->start == 0)
+               return -1;
+
+       g->end = g->start - 1;
+       for (g->start = g->end; xdf->rchg[g->start - 1]; g->start--)
+               ;
+
+       return 0;
+}
+
+/*
+ * If g can be slid toward the end of the file, do so, and if it bumps into a
+ * following group, expand this group to include it. Return 0 on success or -1
+ * if g cannot be slid down.
+ */
+static int group_slide_down(xdfile_t *xdf, struct xdlgroup *g, long flags)
+{
+       if (g->end < xdf->nrec &&
+           recs_match(xdf->recs[g->start], xdf->recs[g->end], flags)) {
+               xdf->rchg[g->start++] = 0;
+               xdf->rchg[g->end++] = 1;
+
+               while (xdf->rchg[g->end])
+                       g->end++;
+
+               return 0;
+       } else {
+               return -1;
+       }
+}
+
+/*
+ * If g can be slid toward the beginning of the file, do so, and if it bumps
+ * into a previous group, expand this group to include it. Return 0 on success
+ * or -1 if g cannot be slid up.
+ */
+static int group_slide_up(xdfile_t *xdf, struct xdlgroup *g, long flags)
+{
+       if (g->start > 0 &&
+           recs_match(xdf->recs[g->start - 1], xdf->recs[g->end - 1], flags)) {
+               xdf->rchg[--g->start] = 1;
+               xdf->rchg[--g->end] = 0;
+
+               while (xdf->rchg[g->start - 1])
+                       g->start--;
+
+               return 0;
+       } else {
+               return -1;
+       }
+}
+
+static void xdl_bug(const char *msg)
+{
+       fprintf(stderr, "BUG: %s\n", msg);
+       exit(1);
+}
+
+/*
+ * Move back and forward change groups for a consistent and pretty diff output.
+ * This also helps in finding joinable change groups and reducing the diff
+ * size.
+ */
+int xdl_change_compact(xdfile_t *xdf, xdfile_t *xdfo, long flags) {
+       struct xdlgroup g, go;
+       long earliest_end, end_matching_other;
+       long groupsize;
+
+       group_init(xdf, &g);
+       group_init(xdfo, &go);
+
+       while (1) {
+               /* If the group is empty in the to-be-compacted file, skip it: */
+               if (g.end == g.start)
+                       goto next;
+
+               /*
+                * Now shift the change up and then down as far as possible in
+                * each direction. If it bumps into any other changes, merge them.
+                */
                do {
-                       grpsiz = ix - ixs;
-                       blank_lines = 0;
+                       groupsize = g.end - g.start;
 
                        /*
-                        * If the line before the current change group, is equal to
-                        * the last line of the current change group, shift backward
-                        * the group.
+                        * Keep track of the last "end" index that causes this
+                        * group to align with a group of changed lines in the
+                        * other file. -1 indicates that we haven't found such
+                        * a match yet:
                         */
-                       while (ixs > 0 && recs_match(recs, ixs - 1, ix - 1, flags)) {
-                               rchg[--ixs] = 1;
-                               rchg[--ix] = 0;
-
-                               /*
-                                * This change might have joined two change groups,
-                                * so we try to take this scenario in account by moving
-                                * the start index accordingly (and so the other-file
-                                * end-of-group index).
-                                */
-                               for (; rchg[ixs - 1]; ixs--);
-                               while (rchgo[--ixo]);
-                       }
+                       end_matching_other = -1;
 
-                       /*
-                        * Record the end-of-group position in case we are matched
-                        * with a group of changes in the other file (that is, the
-                        * change record before the end-of-group index in the other
-                        * file is set).
-                        */
-                       ixref = rchgo[ixo - 1] ? ix: nrec;
+                       /* Shift the group backward as much as possible: */
+                       while (!group_slide_up(xdf, &g, flags))
+                               if (group_previous(xdfo, &go))
+                                       xdl_bug("group sync broken sliding up");
 
                        /*
-                        * If the first line of the current change group, is equal to
-                        * the line next of the current change group, shift forward
-                        * the group.
+                        * This is this highest that this group can be shifted.
+                        * Record its end index:
                         */
-                       while (ix < nrec && recs_match(recs, ixs, ix, flags)) {
-                               blank_lines += is_blank_line(recs, ix, flags);
-
-                               rchg[ixs++] = 0;
-                               rchg[ix++] = 1;
-
-                               /*
-                                * This change might have joined two change groups,
-                                * so we try to take this scenario in account by moving
-                                * the start index accordingly (and so the other-file
-                                * end-of-group index). Keep tracking the reference
-                                * index in case we are shifting together with a
-                                * corresponding group of changes in the other file.
-                                */
-                               for (; rchg[ix]; ix++);
-                               while (rchgo[++ixo])
-                                       ixref = ix;
-                       }
-               } while (grpsiz != ix - ixs);
+                       earliest_end = g.end;
 
-               /*
-                * Try to move back the possibly merged group of changes, to match
-                * the recorded position in the other file.
-                */
-               while (ixref < ix) {
-                       rchg[--ixs] = 1;
-                       rchg[--ix] = 0;
-                       while (rchgo[--ixo]);
-               }
+                       if (go.end > go.start)
+                               end_matching_other = g.end;
+
+                       /* Now shift the group forward as far as possible: */
+                       while (1) {
+                               if (group_slide_down(xdf, &g, flags))
+                                       break;
+                               if (group_next(xdfo, &go))
+                                       xdl_bug("group sync broken sliding down");
+
+                               if (go.end > go.start)
+                                       end_matching_other = g.end;
+                       }
+               } while (groupsize != g.end - g.start);
 
                /*
-                * If a group can be moved back and forth, see if there is a
-                * blank line in the moving space. If there is a blank line,
-                * make sure the last blank line is the end of the group.
+                * If the group can be shifted, then we can possibly use this
+                * freedom to produce a more intuitive diff.
                 *
-                * As we already shifted the group forward as far as possible
-                * in the earlier loop, we need to shift it back only if at all.
+                * The group is currently shifted as far down as possible, so the
+                * heuristics below only have to handle upwards shifts.
                 */
-               if ((flags & XDF_COMPACTION_HEURISTIC) && blank_lines) {
-                       while (ixs > 0 &&
-                              !is_blank_line(recs, ix - 1, flags) &&
-                              recs_match(recs, ixs - 1, ix - 1, flags)) {
-                               rchg[--ixs] = 1;
-                               rchg[--ix] = 0;
+
+               if (g.end == earliest_end) {
+                       /* no shifting was possible */
+               } else if (end_matching_other != -1) {
+                       /*
+                        * Move the possibly merged group of changes back to line
+                        * up with the last group of changes from the other file
+                        * that it can align with.
+                        */
+                       while (go.end == go.start) {
+                               if (group_slide_up(xdf, &g, flags))
+                                       xdl_bug("match disappeared");
+                               if (group_previous(xdfo, &go))
+                                       xdl_bug("group sync broken sliding to match");
+                       }
+               } else if (flags & XDF_INDENT_HEURISTIC) {
+                       /*
+                        * Indent heuristic: a group of pure add/delete lines
+                        * implies two splits, one between the end of the "before"
+                        * context and the start of the group, and another between
+                        * the end of the group and the beginning of the "after"
+                        * context. Some splits are aesthetically better and some
+                        * are worse. We compute a badness "score" for each split,
+                        * and add the scores for the two splits to define a
+                        * "score" for each position that the group can be shifted
+                        * to. Then we pick the shift with the lowest score.
+                        */
+                       long shift, best_shift = -1;
+                       struct split_score best_score;
+
+                       for (shift = earliest_end; shift <= g.end; shift++) {
+                               struct split_measurement m;
+                               struct split_score score = {0, 0};
+
+                               measure_split(xdf, shift, &m);
+                               score_add_split(&m, &score);
+                               measure_split(xdf, shift - groupsize, &m);
+                               score_add_split(&m, &score);
+                               if (best_shift == -1 ||
+                                   score_cmp(&score, &best_score) <= 0) {
+                                       best_score.effective_indent = score.effective_indent;
+                                       best_score.penalty = score.penalty;
+                                       best_shift = shift;
+                               }
+                       }
+
+                       while (g.end > best_shift) {
+                               if (group_slide_up(xdf, &g, flags))
+                                       xdl_bug("best shift unreached");
+                               if (group_previous(xdfo, &go))
+                                       xdl_bug("group sync broken sliding to blank line");
                        }
                }
+
+       next:
+               /* Move past the just-processed group: */
+               if (group_next(xdf, &g))
+                       break;
+               if (group_next(xdfo, &go))
+                       xdl_bug("group sync broken moving to next group");
        }
 
+       if (!group_next(xdfo, &go))
+               xdl_bug("group sync broken at end of file");
+
        return 0;
 }