2 * Copyright (c) 1983 Regents of the University of California.
5 * Redistribution and use in source and binary forms are permitted
6 * provided that: (1) source distributions retain this entire copyright
7 * notice and comment, and (2) distributions including binaries display
8 * the following acknowledgement: ``This product includes software
9 * developed by the University of California, Berkeley and its contributors''
10 * in the documentation or other materials provided with the distribution
11 * and in all advertising materials mentioning features or use of this
12 * software. Neither the name of the University nor the names of its
13 * contributors may be used to endorse or promote products derived
14 * from this software without specific prior written permission.
15 * THIS SOFTWARE IS PROVIDED ``AS IS'' AND WITHOUT ANY EXPRESS OR
16 * IMPLIED WARRANTIES, INCLUDING, WITHOUT LIMITATION, THE IMPLIED
17 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE.
19 #include "libiberty.h"
21 #include "call_graph.h"
34 * Return TRUE iff PARENT has an arc to covers the address
35 * range covered by CHILD.
38 DEFUN (arc_lookup, (parent, child), Sym * parent AND Sym * child)
42 if (!parent || !child)
44 printf ("[arc_lookup] parent == 0 || child == 0\n");
47 DBG (LOOKUPDEBUG, printf ("[arc_lookup] parent %s child %s\n",
48 parent->name, child->name));
49 for (arc = parent->cg.children; arc; arc = arc->next_child)
51 DBG (LOOKUPDEBUG, printf ("[arc_lookup]\t parent %s child %s\n",
52 arc->parent->name, arc->child->name));
53 if (child->addr >= arc->child->addr
54 && child->end_addr <= arc->child->end_addr)
64 * Add (or just increment) an arc:
67 DEFUN (arc_add, (parent, child, count),
68 Sym * parent AND Sym * child AND int count)
70 static int maxarcs = 0;
73 DBG (TALLYDEBUG, printf ("[arc_add] %d arcs from %s to %s\n",
74 count, parent->name, child->name));
75 arc = arc_lookup (parent, child);
79 * A hit: just increment the count.
81 DBG (TALLYDEBUG, printf ("[tally] hit %d += %d\n",
86 arc = (Arc *) xmalloc (sizeof (*arc));
91 /* If this isn't an arc for a recursive call to parent, then add it
92 to the array of arcs. */
95 /* If we've exhausted space in our current array, get a new one
96 and copy the contents. We might want to throttle the doubling
98 if (numarcs == maxarcs)
100 /* Determine how much space we want to allocate. */
105 /* Allocate the new array. */
106 newarcs = (Arc **)xmalloc(sizeof (Arc *) * maxarcs);
108 /* Copy the old array's contents into the new array. */
109 memcpy (newarcs, arcs, numarcs * sizeof (Arc *));
111 /* Free up the old array. */
114 /* And make the new array be the current array. */
118 /* Place this arc in the arc array. */
119 arcs[numarcs++] = arc;
122 /* prepend this child to the children of this parent: */
123 arc->next_child = parent->cg.children;
124 parent->cg.children = arc;
126 /* prepend this parent to the parents of this child: */
127 arc->next_parent = child->cg.parents;
128 child->cg.parents = arc;
133 DEFUN (cmp_topo, (lp, rp), const PTR lp AND const PTR rp)
135 const Sym *left = *(const Sym **) lp;
136 const Sym *right = *(const Sym **) rp;
138 return left->cg.top_order - right->cg.top_order;
143 DEFUN (propagate_time, (parent), Sym * parent)
147 double share, prop_share;
149 if (parent->cg.prop.fract == 0.0)
154 /* gather time from children of this parent: */
156 for (arc = parent->cg.children; arc; arc = arc->next_child)
159 if (arc->count == 0 || child == parent || child->cg.prop.fract == 0)
163 if (child->cg.cyc.head != child)
165 if (parent->cg.cyc.num == child->cg.cyc.num)
169 if (parent->cg.top_order <= child->cg.top_order)
171 fprintf (stderr, "[propagate] toporder botches\n");
173 child = child->cg.cyc.head;
177 if (parent->cg.top_order <= child->cg.top_order)
179 fprintf (stderr, "[propagate] toporder botches\n");
183 if (child->ncalls == 0)
188 /* distribute time for this arc: */
189 arc->time = child->hist.time * (((double) arc->count)
190 / ((double) child->ncalls));
191 arc->child_time = child->cg.child_time
192 * (((double) arc->count) / ((double) child->ncalls));
193 share = arc->time + arc->child_time;
194 parent->cg.child_time += share;
196 /* (1 - cg.prop.fract) gets lost along the way: */
197 prop_share = parent->cg.prop.fract * share;
199 /* fix things for printing: */
200 parent->cg.prop.child += prop_share;
201 arc->time *= parent->cg.prop.fract;
202 arc->child_time *= parent->cg.prop.fract;
204 /* add this share to the parent's cycle header, if any: */
205 if (parent->cg.cyc.head != parent)
207 parent->cg.cyc.head->cg.child_time += share;
208 parent->cg.cyc.head->cg.prop.child += prop_share;
211 printf ("[prop_time] child \t");
213 printf (" with %f %f %d/%d\n", child->hist.time,
214 child->cg.child_time, arc->count, child->ncalls);
215 printf ("[prop_time] parent\t");
217 printf ("\n[prop_time] share %f\n", share));
223 * Compute the time of a cycle as the sum of the times of all
227 DEFUN_VOID (cycle_time)
231 for (cyc = &cycle_header[1]; cyc <= &cycle_header[num_cycles]; ++cyc)
233 for (member = cyc->cg.cyc.next; member; member = member->cg.cyc.next)
235 if (member->cg.prop.fract == 0.0)
238 * All members have the same propfraction except those
239 * that were excluded with -E.
243 cyc->hist.time += member->hist.time;
245 cyc->cg.prop.self = cyc->cg.prop.fract * cyc->hist.time;
251 DEFUN_VOID (cycle_link)
253 Sym *sym, *cyc, *member;
257 /* count the number of cycles, and initialize the cycle lists: */
260 for (sym = symtab.base; sym < symtab.limit; ++sym)
262 /* this is how you find unattached cycles: */
263 if (sym->cg.cyc.head == sym && sym->cg.cyc.next)
270 * cycle_header is indexed by cycle number: i.e. it is origin 1,
273 cycle_header = (Sym *) xmalloc ((num_cycles + 1) * sizeof (Sym));
276 * Now link cycles to true cycle-heads, number them, accumulate
277 * the data for the cycle.
281 for (sym = symtab.base; sym < symtab.limit; ++sym)
283 if (!(sym->cg.cyc.head == sym && sym->cg.cyc.next != 0))
290 cyc->cg.print_flag = TRUE; /* should this be printed? */
291 cyc->cg.top_order = DFN_NAN; /* graph call chain top-sort order */
292 cyc->cg.cyc.num = num; /* internal number of cycle on */
293 cyc->cg.cyc.head = cyc; /* pointer to head of cycle */
294 cyc->cg.cyc.next = sym; /* pointer to next member of cycle */
295 DBG (CYCLEDEBUG, printf ("[cycle_link] ");
297 printf (" is the head of cycle %d\n", num));
299 /* link members to cycle header: */
300 for (member = sym; member; member = member->cg.cyc.next)
302 member->cg.cyc.num = num;
303 member->cg.cyc.head = cyc;
307 * Count calls from outside the cycle and those among cycle
310 for (member = sym; member; member = member->cg.cyc.next)
312 for (arc = member->cg.parents; arc; arc = arc->next_parent)
314 if (arc->parent == member)
318 if (arc->parent->cg.cyc.num == num)
320 cyc->cg.self_calls += arc->count;
324 cyc->ncalls += arc->count;
333 * Check if any parent of this child (or outside parents of this
334 * cycle) have their print flags on and set the print flag of the
335 * child (cycle) appropriately. Similarly, deal with propagation
336 * fractions from parents.
339 DEFUN (inherit_flags, (child), Sym * child)
341 Sym *head, *parent, *member;
344 head = child->cg.cyc.head;
347 /* just a regular child, check its parents: */
348 child->cg.print_flag = FALSE;
349 child->cg.prop.fract = 0.0;
350 for (arc = child->cg.parents; arc; arc = arc->next_parent)
352 parent = arc->parent;
357 child->cg.print_flag |= parent->cg.print_flag;
359 * If the child was never actually called (e.g., this arc
360 * is static (and all others are, too)) no time propagates
365 child->cg.prop.fract += parent->cg.prop.fract
366 * (((double) arc->count) / ((double) child->ncalls));
373 * Its a member of a cycle, look at all parents from outside
376 head->cg.print_flag = FALSE;
377 head->cg.prop.fract = 0.0;
378 for (member = head->cg.cyc.next; member; member = member->cg.cyc.next)
380 for (arc = member->cg.parents; arc; arc = arc->next_parent)
382 if (arc->parent->cg.cyc.head == head)
386 parent = arc->parent;
387 head->cg.print_flag |= parent->cg.print_flag;
389 * If the cycle was never actually called (e.g. this
390 * arc is static (and all others are, too)) no time
391 * propagates along this arc.
395 head->cg.prop.fract += parent->cg.prop.fract
396 * (((double) arc->count) / ((double) head->ncalls));
400 for (member = head; member; member = member->cg.cyc.next)
402 member->cg.print_flag = head->cg.print_flag;
403 member->cg.prop.fract = head->cg.prop.fract;
410 * In one top-to-bottom pass over the topologically sorted symbols
412 * cg.print_flag as the union of parents' print_flags
413 * propfraction as the sum of fractional parents' propfractions
414 * and while we're here, sum time for functions.
417 DEFUN (propagate_flags, (symbols), Sym ** symbols)
420 Sym *old_head, *child;
423 for (index = symtab.len - 1; index >= 0; --index)
425 child = symbols[index];
427 * If we haven't done this function or cycle, inherit things
428 * from parent. This way, we are linear in the number of arcs
429 * since we do all members of a cycle (and the cycle itself)
430 * as we hit the first member of the cycle.
432 if (child->cg.cyc.head != old_head)
434 old_head = child->cg.cyc.head;
435 inherit_flags (child);
438 printf ("[prop_flags] ");
440 printf ("inherits print-flag %d and prop-fract %f\n",
441 child->cg.print_flag, child->cg.prop.fract));
442 if (!child->cg.print_flag)
445 * Printflag is off. It gets turned on by being in the
446 * INCL_GRAPH table, or there being an empty INCL_GRAPH
447 * table and not being in the EXCL_GRAPH table.
449 if (sym_lookup (&syms[INCL_GRAPH], child->addr)
450 || (syms[INCL_GRAPH].len == 0
451 && !sym_lookup (&syms[EXCL_GRAPH], child->addr)))
453 child->cg.print_flag = TRUE;
459 * This function has printing parents: maybe someone wants
460 * to shut it up by putting it in the EXCL_GRAPH table.
461 * (But favor INCL_GRAPH over EXCL_GRAPH.)
463 if (!sym_lookup (&syms[INCL_GRAPH], child->addr)
464 && sym_lookup (&syms[EXCL_GRAPH], child->addr))
466 child->cg.print_flag = FALSE;
469 if (child->cg.prop.fract == 0.0)
472 * No parents to pass time to. Collect time from children
473 * if its in the INCL_TIME table, or there is an empty
474 * INCL_TIME table and its not in the EXCL_TIME table.
476 if (sym_lookup (&syms[INCL_TIME], child->addr)
477 || (syms[INCL_TIME].len == 0
478 && !sym_lookup (&syms[EXCL_TIME], child->addr)))
480 child->cg.prop.fract = 1.0;
486 * It has parents to pass time to, but maybe someone wants
487 * to shut it up by puttting it in the EXCL_TIME table.
488 * (But favor being in INCL_TIME tabe over being in
491 if (!sym_lookup (&syms[INCL_TIME], child->addr)
492 && sym_lookup (&syms[EXCL_TIME], child->addr))
494 child->cg.prop.fract = 0.0;
497 child->cg.prop.self = child->hist.time * child->cg.prop.fract;
498 print_time += child->cg.prop.self;
500 printf ("[prop_flags] ");
502 printf (" ends up with printflag %d and prop-fract %f\n",
503 child->cg.print_flag, child->cg.prop.fract);
504 printf ("[prop_flags] time %f propself %f print_time %f\n",
505 child->hist.time, child->cg.prop.self, print_time));
511 * Compare by decreasing propagated time. If times are equal, but one
512 * is a cycle header, say that's first (e.g. less, i.e. -1). If one's
513 * name doesn't have an underscore and the other does, say that one is
514 * first. All else being equal, compare by names.
517 DEFUN (cmp_total, (lp, rp), const PTR lp AND const PTR rp)
519 const Sym *left = *(const Sym **) lp;
520 const Sym *right = *(const Sym **) rp;
523 diff = (left->cg.prop.self + left->cg.prop.child)
524 - (right->cg.prop.self + right->cg.prop.child);
533 if (!left->name && left->cg.cyc.num != 0)
537 if (!right->name && right->cg.cyc.num != 0)
549 if (left->name[0] != '_' && right->name[0] == '_')
553 if (left->name[0] == '_' && right->name[0] != '_')
557 if (left->ncalls > right->ncalls)
561 if (left->ncalls < right->ncalls)
565 return strcmp (left->name, right->name);
570 * Topologically sort the graph (collapsing cycles), and propagates
571 * time bottom up and flags top down.
574 DEFUN_VOID (cg_assemble)
576 Sym *parent, **time_sorted_syms, **top_sorted_syms;
579 extern void find_call PARAMS ((Sym * parent,
580 bfd_vma p_lowpc, bfd_vma p_highpc));
582 * initialize various things:
583 * zero out child times.
584 * count self-recursive calls.
585 * indicate that nothing is on cycles.
587 for (parent = symtab.base; parent < symtab.limit; parent++)
589 parent->cg.child_time = 0.0;
590 arc = arc_lookup (parent, parent);
591 if (arc && parent == arc->child)
593 parent->ncalls -= arc->count;
594 parent->cg.self_calls = arc->count;
598 parent->cg.self_calls = 0;
600 parent->cg.prop.fract = 0.0;
601 parent->cg.prop.self = 0.0;
602 parent->cg.prop.child = 0.0;
603 parent->cg.print_flag = FALSE;
604 parent->cg.top_order = DFN_NAN;
605 parent->cg.cyc.num = 0;
606 parent->cg.cyc.head = parent;
607 parent->cg.cyc.next = 0;
608 if (ignore_direct_calls)
610 find_call (parent, parent->addr, (parent + 1)->addr);
614 * Topologically order things. If any node is unnumbered, number
615 * it and any of its descendents.
617 for (parent = symtab.base; parent < symtab.limit; parent++)
619 if (parent->cg.top_order == DFN_NAN)
625 /* link together nodes on the same cycle: */
628 /* sort the symbol table in reverse topological order: */
629 top_sorted_syms = (Sym **) xmalloc (symtab.len * sizeof (Sym *));
630 for (index = 0; index < symtab.len; ++index)
632 top_sorted_syms[index] = &symtab.base[index];
634 qsort (top_sorted_syms, symtab.len, sizeof (Sym *), cmp_topo);
636 printf ("[cg_assemble] topological sort listing\n");
637 for (index = 0; index < symtab.len; ++index)
639 printf ("[cg_assemble] ");
640 printf ("%d:", top_sorted_syms[index]->cg.top_order);
641 print_name (top_sorted_syms[index]);
646 * Starting from the topological top, propagate print flags to
647 * children. also, calculate propagation fractions. this happens
648 * before time propagation since time propagation uses the
651 propagate_flags (top_sorted_syms);
654 * Starting from the topological bottom, propogate children times
658 for (index = 0; index < symtab.len; ++index)
660 propagate_time (top_sorted_syms[index]);
663 free (top_sorted_syms);
666 * Now, sort by CG.PROP.SELF + CG.PROP.CHILD. Sorting both the regular
667 * function names and cycle headers.
669 time_sorted_syms = (Sym **) xmalloc ((symtab.len + num_cycles) * sizeof (Sym *));
670 for (index = 0; index < symtab.len; index++)
672 time_sorted_syms[index] = &symtab.base[index];
674 for (index = 1; index <= num_cycles; index++)
676 time_sorted_syms[symtab.len + index - 1] = &cycle_header[index];
678 qsort (time_sorted_syms, symtab.len + num_cycles, sizeof (Sym *),
680 for (index = 0; index < symtab.len + num_cycles; index++)
682 time_sorted_syms[index]->cg.index = index + 1;
684 return time_sorted_syms;