Merge aosp-toolchain/gcc/gcc-4_9 changes.
[official-gcc.git] / gcc-4_9-mobile / gcc / mcf.c
blobf8997541a5aedf2074f45d39afc4fdcc183693d3
1 /* Routines to implement minimum-cost maximal flow algorithm used to smooth
2 basic block and edge frequency counts.
3 Copyright (C) 2008-2014 Free Software Foundation, Inc.
4 Contributed by Paul Yuan (yingbo.com@gmail.com) and
5 Vinodha Ramasamy (vinodha@google.com).
7 This file is part of GCC.
8 GCC is free software; you can redistribute it and/or modify it under
9 the terms of the GNU General Public License as published by the Free
10 Software Foundation; either version 3, or (at your option) any later
11 version.
13 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
14 WARRANTY; without even the implied warranty of MERCHANTABILITY or
15 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
16 for more details.
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING3. If not see
20 <http://www.gnu.org/licenses/>. */
22 /* References:
23 [1] "Feedback-directed Optimizations in GCC with Estimated Edge Profiles
24 from Hardware Event Sampling", Vinodha Ramasamy, Paul Yuan, Dehao Chen,
25 and Robert Hundt; GCC Summit 2008.
26 [2] "Complementing Missing and Inaccurate Profiling Using a Minimum Cost
27 Circulation Algorithm", Roy Levin, Ilan Newman and Gadi Haber;
28 HiPEAC '08.
30 Algorithm to smooth basic block and edge counts:
31 1. create_fixup_graph: Create fixup graph by translating function CFG into
32 a graph that satisfies MCF algorithm requirements.
33 2. find_max_flow: Find maximal flow.
34 3. compute_residual_flow: Form residual network.
35 4. Repeat:
36 cancel_negative_cycle: While G contains a negative cost cycle C, reverse
37 the flow on the found cycle by the minimum residual capacity in that
38 cycle.
39 5. Form the minimal cost flow
40 f(u,v) = rf(v, u).
41 6. adjust_cfg_counts: Update initial edge weights with corrected weights.
42 delta(u.v) = f(u,v) -f(v,u).
43 w*(u,v) = w(u,v) + delta(u,v). */
45 #include "config.h"
46 #include "system.h"
47 #include "coretypes.h"
48 #include "basic-block.h"
49 #include "gcov-io.h"
50 #include "params.h"
51 #include "diagnostic-core.h"
53 #include "tree.h"
54 #include "profile.h"
55 #include "dumpfile.h"
57 /* CAP_INFINITY: Constant to represent infinite capacity. */
58 #define CAP_INFINITY INTTYPE_MAXIMUM (HOST_WIDEST_INT)
60 /* COST FUNCTION. */
61 #define K_POS(b) ((b))
62 #define K_NEG(b) (50 * (b))
63 #define COST(k, w) ((k) / mcf_ln ((w) + 2))
64 /* Limit the number of iterations for cancel_negative_cycles() to ensure
65 reasonable compile time. */
66 #define MAX_ITER(n, e) (PARAM_VALUE (PARAM_MIN_MCF_CANCEL_ITERS) + \
67 (1000000 / ((n) * (e))))
69 typedef enum
71 INVALID_EDGE = 0,
72 VERTEX_SPLIT_EDGE, /* Edge to represent vertex with w(e) = w(v). */
73 REDIRECT_EDGE, /* Edge after vertex transformation. */
74 REVERSE_EDGE,
75 SOURCE_CONNECT_EDGE, /* Single edge connecting to single source. */
76 SINK_CONNECT_EDGE, /* Single edge connecting to single sink. */
77 SINK_SOURCE_EDGE, /* Single edge connecting sink to source. */
78 BALANCE_EDGE, /* Edge connecting with source/sink: cp(e) = 0. */
79 REDIRECT_NORMALIZED_EDGE, /* Normalized edge for a redirect edge. */
80 REVERSE_NORMALIZED_EDGE /* Normalized edge for a reverse edge. */
81 } edge_type;
83 /* Structure to represent an edge in the fixup graph. */
84 typedef struct fixup_edge_d
86 int src;
87 int dest;
88 /* Flag denoting type of edge and attributes for the flow field. */
89 edge_type type;
90 bool is_rflow_valid;
91 /* Index to the normalization vertex added for this edge. */
92 int norm_vertex_index;
93 /* Flow for this edge. */
94 gcov_type flow;
95 /* Residual flow for this edge - used during negative cycle canceling. */
96 gcov_type rflow;
97 gcov_type weight;
98 gcov_type cost;
99 gcov_type max_capacity;
100 } fixup_edge_type;
102 typedef fixup_edge_type *fixup_edge_p;
105 /* Structure to represent a vertex in the fixup graph. */
106 typedef struct fixup_vertex_d
108 vec<fixup_edge_p> succ_edges;
109 } fixup_vertex_type;
111 typedef fixup_vertex_type *fixup_vertex_p;
113 /* Fixup graph used in the MCF algorithm. */
114 typedef struct fixup_graph_d
116 /* Current number of vertices for the graph. */
117 int num_vertices;
118 /* Current number of edges for the graph. */
119 int num_edges;
120 /* Index of new entry vertex. */
121 int new_entry_index;
122 /* Index of new exit vertex. */
123 int new_exit_index;
124 /* Fixup vertex list. Adjacency list for fixup graph. */
125 fixup_vertex_p vertex_list;
126 /* Fixup edge list. */
127 fixup_edge_p edge_list;
128 } fixup_graph_type;
130 typedef struct queue_d
132 int *queue;
133 int head;
134 int tail;
135 int size;
136 } queue_type;
138 /* Structure used in the maximal flow routines to find augmenting path. */
139 typedef struct augmenting_path_d
141 /* Queue used to hold vertex indices. */
142 queue_type queue_list;
143 /* Vector to hold chain of pred vertex indices in augmenting path. */
144 int *bb_pred;
145 /* Vector that indicates if basic block i has been visited. */
146 int *is_visited;
147 } augmenting_path_type;
150 /* Function definitions. */
152 /* Dump routines to aid debugging. */
154 /* Print basic block with index N for FIXUP_GRAPH in n' and n'' format. */
156 static void
157 print_basic_block (FILE *file, fixup_graph_type *fixup_graph, int n)
159 if (n == ENTRY_BLOCK)
160 fputs ("ENTRY", file);
161 else if (n == ENTRY_BLOCK + 1)
162 fputs ("ENTRY''", file);
163 else if (n == 2 * EXIT_BLOCK)
164 fputs ("EXIT", file);
165 else if (n == 2 * EXIT_BLOCK + 1)
166 fputs ("EXIT''", file);
167 else if (n == fixup_graph->new_exit_index)
168 fputs ("NEW_EXIT", file);
169 else if (n == fixup_graph->new_entry_index)
170 fputs ("NEW_ENTRY", file);
171 else
173 fprintf (file, "%d", n / 2);
174 if (n % 2)
175 fputs ("''", file);
176 else
177 fputs ("'", file);
182 /* Print edge S->D for given fixup_graph with n' and n'' format.
183 PARAMETERS:
184 S is the index of the source vertex of the edge (input) and
185 D is the index of the destination vertex of the edge (input) for the given
186 fixup_graph (input). */
188 static void
189 print_edge (FILE *file, fixup_graph_type *fixup_graph, int s, int d)
191 print_basic_block (file, fixup_graph, s);
192 fputs ("->", file);
193 print_basic_block (file, fixup_graph, d);
197 /* Dump out the attributes of a given edge FEDGE in the fixup_graph to a
198 file. */
199 static void
200 dump_fixup_edge (FILE *file, fixup_graph_type *fixup_graph, fixup_edge_p fedge)
202 if (!fedge)
204 fputs ("NULL fixup graph edge.\n", file);
205 return;
208 print_edge (file, fixup_graph, fedge->src, fedge->dest);
209 fputs (": ", file);
211 if (fedge->type)
213 fprintf (file, "flow/capacity=" HOST_WIDEST_INT_PRINT_DEC "/",
214 fedge->flow);
215 if (fedge->max_capacity == CAP_INFINITY)
216 fputs ("+oo,", file);
217 else
218 fprintf (file, "" HOST_WIDEST_INT_PRINT_DEC ",", fedge->max_capacity);
221 if (fedge->is_rflow_valid)
223 if (fedge->rflow == CAP_INFINITY)
224 fputs (" rflow=+oo.", file);
225 else
226 fprintf (file, " rflow=" HOST_WIDEST_INT_PRINT_DEC ",", fedge->rflow);
229 fprintf (file, " cost=" HOST_WIDEST_INT_PRINT_DEC ".", fedge->cost);
231 fprintf (file, "\t(%d->%d)", fedge->src, fedge->dest);
233 if (fedge->type)
235 switch (fedge->type)
237 case VERTEX_SPLIT_EDGE:
238 fputs (" @VERTEX_SPLIT_EDGE", file);
239 break;
241 case REDIRECT_EDGE:
242 fputs (" @REDIRECT_EDGE", file);
243 break;
245 case SOURCE_CONNECT_EDGE:
246 fputs (" @SOURCE_CONNECT_EDGE", file);
247 break;
249 case SINK_CONNECT_EDGE:
250 fputs (" @SINK_CONNECT_EDGE", file);
251 break;
253 case SINK_SOURCE_EDGE:
254 fputs (" @SINK_SOURCE_EDGE", file);
255 break;
257 case REVERSE_EDGE:
258 fputs (" @REVERSE_EDGE", file);
259 break;
261 case BALANCE_EDGE:
262 fputs (" @BALANCE_EDGE", file);
263 break;
265 case REDIRECT_NORMALIZED_EDGE:
266 case REVERSE_NORMALIZED_EDGE:
267 fputs (" @NORMALIZED_EDGE", file);
268 break;
270 default:
271 fputs (" @INVALID_EDGE", file);
272 break;
275 fputs ("\n", file);
279 /* Print out the edges and vertices of the given FIXUP_GRAPH, into the dump
280 file. The input string MSG is printed out as a heading. */
282 static void
283 dump_fixup_graph (FILE *file, fixup_graph_type *fixup_graph, const char *msg)
285 int i, j;
286 int fnum_vertices, fnum_edges;
288 fixup_vertex_p fvertex_list, pfvertex;
289 fixup_edge_p pfedge;
291 gcc_assert (fixup_graph);
292 fvertex_list = fixup_graph->vertex_list;
293 fnum_vertices = fixup_graph->num_vertices;
294 fnum_edges = fixup_graph->num_edges;
296 fprintf (file, "\nDump fixup graph for %s(): %s.\n",
297 current_function_name (), msg);
298 fprintf (file,
299 "There are %d vertices and %d edges. new_exit_index is %d.\n\n",
300 fnum_vertices, fnum_edges, fixup_graph->new_exit_index);
302 for (i = 0; i < fnum_vertices; i++)
304 pfvertex = fvertex_list + i;
305 fprintf (file, "vertex_list[%d]: %d succ fixup edges.\n",
306 i, pfvertex->succ_edges.length ());
308 for (j = 0; pfvertex->succ_edges.iterate (j, &pfedge);
309 j++)
311 /* Distinguish forward edges and backward edges in the residual flow
312 network. */
313 if (pfedge->type)
314 fputs ("(f) ", file);
315 else if (pfedge->is_rflow_valid)
316 fputs ("(b) ", file);
317 dump_fixup_edge (file, fixup_graph, pfedge);
321 fputs ("\n", file);
325 /* Utility routines. */
326 /* ln() implementation: approximate calculation. Returns ln of X. */
328 static double
329 mcf_ln (double x)
331 #define E 2.71828
332 int l = 1;
333 double m = E;
335 gcc_assert (x >= 0);
337 while (m < x)
339 m *= E;
340 l++;
343 return l;
347 /* sqrt() implementation: based on open source QUAKE3 code (magic sqrt
348 implementation) by John Carmack. Returns sqrt of X. */
350 static double
351 mcf_sqrt (double x)
353 #define MAGIC_CONST1 0x1fbcf800
354 #define MAGIC_CONST2 0x5f3759df
355 union {
356 int intPart;
357 float floatPart;
358 } convertor, convertor2;
360 gcc_assert (x >= 0);
362 convertor.floatPart = x;
363 convertor2.floatPart = x;
364 convertor.intPart = MAGIC_CONST1 + (convertor.intPart >> 1);
365 convertor2.intPart = MAGIC_CONST2 - (convertor2.intPart >> 1);
367 return 0.5f * (convertor.floatPart + (x * convertor2.floatPart));
371 /* Common code shared between add_fixup_edge and add_rfixup_edge. Adds an edge
372 (SRC->DEST) to the edge_list maintained in FIXUP_GRAPH with cost of the edge
373 added set to COST. */
375 static fixup_edge_p
376 add_edge (fixup_graph_type *fixup_graph, int src, int dest, gcov_type cost)
378 fixup_vertex_p curr_vertex = fixup_graph->vertex_list + src;
379 fixup_edge_p curr_edge = fixup_graph->edge_list + fixup_graph->num_edges;
380 curr_edge->src = src;
381 curr_edge->dest = dest;
382 curr_edge->cost = cost;
383 fixup_graph->num_edges++;
384 if (dump_file)
385 dump_fixup_edge (dump_file, fixup_graph, curr_edge);
386 curr_vertex->succ_edges.safe_push (curr_edge);
387 return curr_edge;
391 /* Add a fixup edge (src->dest) with attributes TYPE, WEIGHT, COST and
392 MAX_CAPACITY to the edge_list in the fixup graph. */
394 static void
395 add_fixup_edge (fixup_graph_type *fixup_graph, int src, int dest,
396 edge_type type, gcov_type weight, gcov_type cost,
397 gcov_type max_capacity)
399 fixup_edge_p curr_edge = add_edge (fixup_graph, src, dest, cost);
400 curr_edge->type = type;
401 curr_edge->weight = weight;
402 curr_edge->max_capacity = max_capacity;
406 /* Add a residual edge (SRC->DEST) with attributes RFLOW and COST
407 to the fixup graph. */
409 static void
410 add_rfixup_edge (fixup_graph_type *fixup_graph, int src, int dest,
411 gcov_type rflow, gcov_type cost)
413 fixup_edge_p curr_edge = add_edge (fixup_graph, src, dest, cost);
414 curr_edge->rflow = rflow;
415 curr_edge->is_rflow_valid = true;
416 /* This edge is not a valid edge - merely used to hold residual flow. */
417 curr_edge->type = INVALID_EDGE;
421 /* Return the pointer to fixup edge SRC->DEST or NULL if edge does not
422 exist in the FIXUP_GRAPH. */
424 static fixup_edge_p
425 find_fixup_edge (fixup_graph_type *fixup_graph, int src, int dest)
427 int j;
428 fixup_edge_p pfedge;
429 fixup_vertex_p pfvertex;
431 gcc_assert (src < fixup_graph->num_vertices);
433 pfvertex = fixup_graph->vertex_list + src;
435 for (j = 0; pfvertex->succ_edges.iterate (j, &pfedge);
436 j++)
437 if (pfedge->dest == dest)
438 return pfedge;
440 return NULL;
444 /* Cleanup routine to free structures in FIXUP_GRAPH. */
446 static void
447 delete_fixup_graph (fixup_graph_type *fixup_graph)
449 int i;
450 int fnum_vertices = fixup_graph->num_vertices;
451 fixup_vertex_p pfvertex = fixup_graph->vertex_list;
453 for (i = 0; i < fnum_vertices; i++, pfvertex++)
454 pfvertex->succ_edges.release ();
456 free (fixup_graph->vertex_list);
457 free (fixup_graph->edge_list);
461 /* Creates a fixup graph FIXUP_GRAPH from the function CFG. */
463 static void
464 create_fixup_graph (fixup_graph_type *fixup_graph)
466 double sqrt_avg_vertex_weight = 0;
467 double total_vertex_weight = 0;
468 double k_pos = 0;
469 double k_neg = 0;
470 /* Vector to hold D(v) = sum_out_edges(v) - sum_in_edges(v). */
471 gcov_type *diff_out_in = NULL;
472 gcov_type supply_value = 0, demand_value = 0;
473 gcov_type fcost = 0;
474 int new_entry_index = 0, new_exit_index = 0;
475 int i = 0, j = 0;
476 int new_index = 0;
477 basic_block bb;
478 edge e;
479 edge_iterator ei;
480 fixup_edge_p pfedge, r_pfedge;
481 fixup_edge_p fedge_list;
482 int fnum_edges;
484 /* Each basic_block will be split into 2 during vertex transformation. */
485 int fnum_vertices_after_transform = 2 * n_basic_blocks_for_fn (cfun);
486 int fnum_edges_after_transform =
487 n_edges_for_fn (cfun) + n_basic_blocks_for_fn (cfun);
489 /* Count the new SOURCE and EXIT vertices to be added. */
490 int fmax_num_vertices =
491 (fnum_vertices_after_transform + n_edges_for_fn (cfun)
492 + n_basic_blocks_for_fn (cfun) + 2);
494 /* In create_fixup_graph: Each basic block and edge can be split into 3
495 edges. Number of balance edges = n_basic_blocks - 1. And there is 1 edge
496 connecting new_entry and new_exit, and 2 edges connecting new_entry to
497 entry, and exit to new_exit. So after create_fixup_graph:
498 max_edges = 4 * n_basic_blocks + 3 * n_edges + 2
499 Accounting for residual flow edges
501 max_edges = 2 * (4 * n_basic_blocks + 3 * n_edges)
502 = 8 * n_basic_blocks + 6 * n_edges
503 < 8 * n_basic_blocks + 8 * n_edges. */
504 int fmax_num_edges = 8 * (n_basic_blocks_for_fn (cfun) +
505 n_edges_for_fn (cfun));
507 /* Initial num of vertices in the fixup graph. */
508 fixup_graph->num_vertices = n_basic_blocks_for_fn (cfun);
510 /* Fixup graph vertex list. */
511 fixup_graph->vertex_list =
512 (fixup_vertex_p) xcalloc (fmax_num_vertices, sizeof (fixup_vertex_type));
514 /* Fixup graph edge list. */
515 fixup_graph->edge_list =
516 (fixup_edge_p) xcalloc (fmax_num_edges, sizeof (fixup_edge_type));
518 diff_out_in =
519 (gcov_type *) xcalloc (1 + fnum_vertices_after_transform,
520 sizeof (gcov_type));
522 /* Compute constants b, k_pos, k_neg used in the cost function calculation.
523 b = sqrt(avg_vertex_weight(cfg)); k_pos = b; k_neg = 50b. */
524 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR_FOR_FN (cfun), NULL, next_bb)
525 total_vertex_weight += bb->count;
527 sqrt_avg_vertex_weight = mcf_sqrt (total_vertex_weight /
528 n_basic_blocks_for_fn (cfun));
530 k_pos = K_POS (sqrt_avg_vertex_weight);
531 k_neg = K_NEG (sqrt_avg_vertex_weight);
533 /* 1. Vertex Transformation: Split each vertex v into two vertices v' and v'',
534 connected by an edge e from v' to v''. w(e) = w(v). */
536 if (dump_file)
537 fprintf (dump_file, "\nVertex transformation:\n");
539 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR_FOR_FN (cfun), NULL, next_bb)
541 /* v'->v'': index1->(index1+1). */
542 i = 2 * bb->index;
544 fcost = (gcov_type) COST (k_pos, bb->count);
545 add_fixup_edge (fixup_graph, i, i + 1, VERTEX_SPLIT_EDGE, bb->count,
546 fcost, CAP_INFINITY);
547 fixup_graph->num_vertices++;
549 FOR_EACH_EDGE (e, ei, bb->succs)
551 /* Edges with ignore attribute set should be treated like they don't
552 exist. */
553 if (EDGE_INFO (e) && EDGE_INFO (e)->ignore)
554 continue;
555 j = 2 * e->dest->index;
556 fcost = (gcov_type) COST (k_pos, e->count);
557 add_fixup_edge (fixup_graph, i + 1, j, REDIRECT_EDGE, e->count, fcost,
558 CAP_INFINITY);
562 /* After vertex transformation. */
563 gcc_assert (fixup_graph->num_vertices == fnum_vertices_after_transform);
564 /* Redirect edges are not added for edges with ignore attribute. */
565 gcc_assert (fixup_graph->num_edges <= fnum_edges_after_transform);
567 fnum_edges_after_transform = fixup_graph->num_edges;
569 /* 2. Initialize D(v). */
570 for (i = 0; i < fnum_edges_after_transform; i++)
572 pfedge = fixup_graph->edge_list + i;
573 diff_out_in[pfedge->src] += pfedge->weight;
574 diff_out_in[pfedge->dest] -= pfedge->weight;
577 /* Entry block - vertex indices 0, 1; EXIT block - vertex indices 2, 3. */
578 for (i = 0; i <= 3; i++)
579 diff_out_in[i] = 0;
581 /* 3. Add reverse edges: needed to decrease counts during smoothing. */
582 if (dump_file)
583 fprintf (dump_file, "\nReverse edges:\n");
584 for (i = 0; i < fnum_edges_after_transform; i++)
586 pfedge = fixup_graph->edge_list + i;
587 if ((pfedge->src == 0) || (pfedge->src == 2))
588 continue;
589 r_pfedge = find_fixup_edge (fixup_graph, pfedge->dest, pfedge->src);
590 if (!r_pfedge && pfedge->weight)
592 /* Skip adding reverse edges for edges with w(e) = 0, as its maximum
593 capacity is 0. */
594 fcost = (gcov_type) COST (k_neg, pfedge->weight);
595 add_fixup_edge (fixup_graph, pfedge->dest, pfedge->src,
596 REVERSE_EDGE, 0, fcost, pfedge->weight);
600 /* 4. Create single source and sink. Connect new source vertex s' to function
601 entry block. Connect sink vertex t' to function exit. */
602 if (dump_file)
603 fprintf (dump_file, "\ns'->S, T->t':\n");
605 new_entry_index = fixup_graph->new_entry_index = fixup_graph->num_vertices;
606 fixup_graph->num_vertices++;
607 /* Set capacity to 0 initially, it will be updated after
608 supply_value is computed. */
609 add_fixup_edge (fixup_graph, new_entry_index, ENTRY_BLOCK,
610 SOURCE_CONNECT_EDGE, 0 /* supply_value */, 0,
611 0 /* supply_value */);
612 add_fixup_edge (fixup_graph, ENTRY_BLOCK, new_entry_index,
613 SOURCE_CONNECT_EDGE, 0 /* supply_value */, 0,
614 0 /* supply_value */);
617 /* Set capacity to 0 initially, it will be updated after
618 demand_value is computed. */
619 new_exit_index = fixup_graph->new_exit_index = fixup_graph->num_vertices;
620 fixup_graph->num_vertices++;
621 add_fixup_edge (fixup_graph, 2 * EXIT_BLOCK + 1, new_exit_index,
622 SINK_CONNECT_EDGE,
623 0 /* demand_value */, 0, 0 /* demand_value */);
624 add_fixup_edge (fixup_graph, new_exit_index, 2 * EXIT_BLOCK + 1,
625 SINK_CONNECT_EDGE,
626 0 /* demand_value */, 0, 0 /* demand_value */);
629 /* Create a back edge from the new_exit to the new_entry.
630 Initially, its capacity will be set to 0 so that it does not
631 affect max flow, but later its capacity will be changed to
632 infinity to cancel negative cycles. */
633 add_fixup_edge (fixup_graph, new_exit_index, new_entry_index,
634 SINK_SOURCE_EDGE, 0, 0, 0);
638 /* Connect vertices with unbalanced D(v) to source/sink. */
639 if (dump_file)
640 fprintf (dump_file, "\nD(v) balance:\n");
642 /* Skip vertices for ENTRY (0, 1) and EXIT (2,3) blocks, so start
643 with i = 4. diff_out_in[v''] should be 0, but may not be due to
644 rounding error. So here we consider all vertices. */
645 for (i = 4; i < new_entry_index; i += 1)
647 if (diff_out_in[i] > 0)
649 add_fixup_edge (fixup_graph, i, new_exit_index, BALANCE_EDGE, 0, 0,
650 diff_out_in[i]);
651 demand_value += diff_out_in[i];
653 else if (diff_out_in[i] < 0)
655 add_fixup_edge (fixup_graph, new_entry_index, i, BALANCE_EDGE, 0, 0,
656 -diff_out_in[i]);
657 supply_value -= diff_out_in[i];
661 /* Set supply = demand. */
662 if (dump_file)
664 fprintf (dump_file, "\nAdjust supply and demand:\n");
665 fprintf (dump_file, "supply_value=" HOST_WIDEST_INT_PRINT_DEC "\n",
666 supply_value);
667 fprintf (dump_file, "demand_value=" HOST_WIDEST_INT_PRINT_DEC "\n",
668 demand_value);
671 if (demand_value > supply_value)
673 pfedge = find_fixup_edge (fixup_graph, new_entry_index, ENTRY_BLOCK);
674 pfedge->max_capacity += (demand_value - supply_value);
676 else
678 pfedge = find_fixup_edge (fixup_graph, 2 * EXIT_BLOCK + 1, new_exit_index);
679 pfedge->max_capacity += (supply_value - demand_value);
682 /* 6. Normalize edges: remove anti-parallel edges. Anti-parallel edges are
683 created by the vertex transformation step from self-edges in the original
684 CFG and by the reverse edges added earlier. */
685 if (dump_file)
686 fprintf (dump_file, "\nNormalize edges:\n");
688 fnum_edges = fixup_graph->num_edges;
689 fedge_list = fixup_graph->edge_list;
691 for (i = 0; i < fnum_edges; i++)
693 pfedge = fedge_list + i;
694 r_pfedge = find_fixup_edge (fixup_graph, pfedge->dest, pfedge->src);
695 if (((pfedge->type == VERTEX_SPLIT_EDGE)
696 || (pfedge->type == REDIRECT_EDGE)) && r_pfedge)
698 new_index = fixup_graph->num_vertices;
699 fixup_graph->num_vertices++;
701 if (dump_file)
703 fprintf (dump_file, "\nAnti-parallel edge:\n");
704 dump_fixup_edge (dump_file, fixup_graph, pfedge);
705 dump_fixup_edge (dump_file, fixup_graph, r_pfedge);
706 fprintf (dump_file, "New vertex is %d.\n", new_index);
707 fprintf (dump_file, "------------------\n");
710 pfedge->norm_vertex_index = new_index;
711 if (dump_file)
713 fprintf (dump_file, "After normalization:\n");
714 dump_fixup_edge (dump_file, fixup_graph, pfedge);
717 /* Add a new fixup edge: new_index->src. */
718 add_fixup_edge (fixup_graph, new_index, pfedge->src,
719 REVERSE_NORMALIZED_EDGE, 0, 0,
720 r_pfedge->max_capacity);
721 gcc_assert (fixup_graph->num_vertices <= fmax_num_vertices);
723 /* Edge: r_pfedge->src -> r_pfedge->dest
724 ==> r_pfedge->src -> new_index. */
725 r_pfedge->dest = new_index;
726 r_pfedge->type = REVERSE_NORMALIZED_EDGE;
727 r_pfedge->max_capacity = pfedge->max_capacity;
728 if (dump_file)
729 dump_fixup_edge (dump_file, fixup_graph, r_pfedge);
733 if (dump_file)
734 dump_fixup_graph (dump_file, fixup_graph, "After create_fixup_graph()");
736 /* Cleanup. */
737 free (diff_out_in);
741 /* Allocates space for the structures in AUGMENTING_PATH. The space needed is
742 proportional to the number of nodes in the graph, which is given by
743 GRAPH_SIZE. */
745 static void
746 init_augmenting_path (augmenting_path_type *augmenting_path, int graph_size)
748 augmenting_path->queue_list.queue = (int *)
749 xcalloc (graph_size + 2, sizeof (int));
750 augmenting_path->queue_list.size = graph_size + 2;
751 augmenting_path->bb_pred = (int *) xcalloc (graph_size, sizeof (int));
752 augmenting_path->is_visited = (int *) xcalloc (graph_size, sizeof (int));
755 /* Free the structures in AUGMENTING_PATH. */
756 static void
757 free_augmenting_path (augmenting_path_type *augmenting_path)
759 free (augmenting_path->queue_list.queue);
760 free (augmenting_path->bb_pred);
761 free (augmenting_path->is_visited);
765 /* Queue routines. Assumes queue will never overflow. */
767 static void
768 init_queue (queue_type *queue_list)
770 gcc_assert (queue_list);
771 queue_list->head = 0;
772 queue_list->tail = 0;
775 /* Return true if QUEUE_LIST is empty. */
776 static bool
777 is_empty (queue_type *queue_list)
779 return (queue_list->head == queue_list->tail);
782 /* Insert element X into QUEUE_LIST. */
783 static void
784 enqueue (queue_type *queue_list, int x)
786 gcc_assert (queue_list->tail < queue_list->size);
787 queue_list->queue[queue_list->tail] = x;
788 (queue_list->tail)++;
791 /* Return the first element in QUEUE_LIST. */
792 static int
793 dequeue (queue_type *queue_list)
795 int x;
796 gcc_assert (queue_list->head >= 0);
797 x = queue_list->queue[queue_list->head];
798 (queue_list->head)++;
799 return x;
803 /* Finds a negative cycle in the residual network using
804 the Bellman-Ford algorithm. The flow on the found cycle is reversed by the
805 minimum residual capacity of that cycle. ENTRY and EXIT vertices are not
806 considered.
808 Parameters:
809 FIXUP_GRAPH - Residual graph (input/output)
810 The following are allocated/freed by the caller:
811 PI - Vector to hold predecessors in path (pi = pred index)
812 D - D[I] holds minimum cost of path from i to sink
813 CYCLE - Vector to hold the minimum cost cycle
815 Return:
816 true if a negative cycle was found, false otherwise. */
818 static bool
819 cancel_negative_cycle (fixup_graph_type *fixup_graph,
820 int *pi, gcov_type *d, int *cycle)
822 int i, j, k;
823 int fnum_vertices, fnum_edges;
824 fixup_edge_p fedge_list, pfedge, r_pfedge;
825 bool found_cycle = false;
826 int cycle_start = 0, cycle_end = 0;
827 gcov_type sum_cost = 0, cycle_flow = 0;
828 bool propagated = false;
830 gcc_assert (fixup_graph);
831 fnum_vertices = fixup_graph->num_vertices;
832 fnum_edges = fixup_graph->num_edges;
833 fedge_list = fixup_graph->edge_list;
835 /* Initialize. */
836 /* Skip ENTRY. */
837 for (i = 1; i < fnum_vertices; i++)
839 d[i] = CAP_INFINITY;
840 pi[i] = -1;
841 cycle[i] = -1;
843 d[ENTRY_BLOCK] = 0;
845 /* Relax. */
846 for (k = 1; k < fnum_vertices; k++)
848 propagated = false;
849 for (i = 0; i < fnum_edges; i++)
851 pfedge = fedge_list + i;
852 if (pfedge->is_rflow_valid && pfedge->rflow
853 && d[pfedge->src] != CAP_INFINITY
854 && (d[pfedge->dest] > d[pfedge->src] + pfedge->cost))
856 d[pfedge->dest] = d[pfedge->src] + pfedge->cost;
857 pi[pfedge->dest] = pfedge->src;
858 propagated = true;
861 if (!propagated)
862 break;
865 if (!propagated)
866 /* No negative cycles exist. */
867 return 0;
869 /* Detect. */
870 for (i = 0; i < fnum_edges; i++)
872 pfedge = fedge_list + i;
873 if (pfedge->is_rflow_valid && pfedge->rflow
874 && d[pfedge->src] != CAP_INFINITY
875 && (d[pfedge->dest] > d[pfedge->src] + pfedge->cost))
877 found_cycle = true;
878 break;
882 if (!found_cycle)
883 return 0;
885 /* Augment the cycle with the cycle's minimum residual capacity. */
886 found_cycle = false;
887 cycle[0] = pfedge->dest;
888 j = pfedge->dest;
890 for (i = 1; i < fnum_vertices; i++)
892 j = pi[j];
893 cycle[i] = j;
894 for (k = 0; k < i; k++)
896 if (cycle[k] == j)
898 /* cycle[k] -> ... -> cycle[i]. */
899 cycle_start = k;
900 cycle_end = i;
901 found_cycle = true;
902 break;
905 if (found_cycle)
906 break;
909 gcc_assert (cycle[cycle_start] == cycle[cycle_end]);
910 if (dump_file)
911 fprintf (dump_file, "\nNegative cycle length is %d:\n",
912 cycle_end - cycle_start);
914 sum_cost = 0;
915 cycle_flow = CAP_INFINITY;
916 for (k = cycle_start; k < cycle_end; k++)
918 pfedge = find_fixup_edge (fixup_graph, cycle[k + 1], cycle[k]);
919 cycle_flow = MIN (cycle_flow, pfedge->rflow);
920 sum_cost += pfedge->cost;
921 if (dump_file)
922 fprintf (dump_file, "%d ", cycle[k]);
925 if (dump_file)
927 fprintf (dump_file, "%d", cycle[k]);
928 fprintf (dump_file,
929 ": (" HOST_WIDEST_INT_PRINT_DEC ", " HOST_WIDEST_INT_PRINT_DEC
930 ")\n", sum_cost, cycle_flow);
931 fprintf (dump_file,
932 "Augment cycle with " HOST_WIDEST_INT_PRINT_DEC "\n",
933 cycle_flow);
936 for (k = cycle_start; k < cycle_end; k++)
938 pfedge = find_fixup_edge (fixup_graph, cycle[k + 1], cycle[k]);
939 r_pfedge = find_fixup_edge (fixup_graph, cycle[k], cycle[k + 1]);
940 if (pfedge->rflow != CAP_INFINITY)
941 pfedge->rflow -= cycle_flow;
942 if (pfedge->type)
943 pfedge->flow += cycle_flow;
944 if (r_pfedge->rflow != CAP_INFINITY)
945 r_pfedge->rflow += cycle_flow;
946 if (r_pfedge->type)
947 r_pfedge->flow -= cycle_flow;
950 return true;
954 /* Computes the residual flow for FIXUP_GRAPH by setting the rflow field of
955 the edges. ENTRY and EXIT vertices should not be considered. */
957 static void
958 compute_residual_flow (fixup_graph_type *fixup_graph)
960 int i;
961 int fnum_edges;
962 fixup_edge_p fedge_list, pfedge;
964 gcc_assert (fixup_graph);
966 if (dump_file)
967 fputs ("\ncompute_residual_flow():\n", dump_file);
969 fnum_edges = fixup_graph->num_edges;
970 fedge_list = fixup_graph->edge_list;
972 for (i = 0; i < fnum_edges; i++)
974 pfedge = fedge_list + i;
975 pfedge->rflow = pfedge->max_capacity == CAP_INFINITY ?
976 CAP_INFINITY : pfedge->max_capacity - pfedge->flow;
977 pfedge->is_rflow_valid = true;
978 add_rfixup_edge (fixup_graph, pfedge->dest, pfedge->src, pfedge->flow,
979 -pfedge->cost);
984 /* Uses Edmonds-Karp algorithm - BFS to find augmenting path from SOURCE to
985 SINK. The fields in the edge vector in the FIXUP_GRAPH are not modified by
986 this routine. The vector bb_pred in the AUGMENTING_PATH structure is updated
987 to reflect the path found.
988 Returns: 0 if no augmenting path is found, 1 otherwise. */
990 static int
991 find_augmenting_path (fixup_graph_type *fixup_graph,
992 augmenting_path_type *augmenting_path, int source,
993 int sink)
995 int u = 0;
996 int i;
997 fixup_vertex_p fvertex_list, pfvertex;
998 fixup_edge_p pfedge;
999 int *bb_pred, *is_visited;
1000 queue_type *queue_list;
1002 gcc_assert (augmenting_path);
1003 bb_pred = augmenting_path->bb_pred;
1004 gcc_assert (bb_pred);
1005 is_visited = augmenting_path->is_visited;
1006 gcc_assert (is_visited);
1007 queue_list = &(augmenting_path->queue_list);
1009 gcc_assert (fixup_graph);
1011 fvertex_list = fixup_graph->vertex_list;
1013 for (u = 0; u < fixup_graph->num_vertices; u++)
1014 is_visited[u] = 0;
1016 init_queue (queue_list);
1017 enqueue (queue_list, source);
1018 bb_pred[source] = -1;
1020 while (!is_empty (queue_list))
1022 u = dequeue (queue_list);
1023 is_visited[u] = 1;
1024 pfvertex = fvertex_list + u;
1025 for (i = 0; pfvertex->succ_edges.iterate (i, &pfedge);
1026 i++)
1028 int dest = pfedge->dest;
1029 if ((pfedge->rflow > 0) && (is_visited[dest] == 0))
1031 enqueue (queue_list, dest);
1032 bb_pred[dest] = u;
1033 is_visited[dest] = 1;
1034 if (dest == sink)
1035 return 1;
1040 return 0;
1044 /* Routine to find the maximal flow:
1045 Algorithm:
1046 1. Initialize flow to 0
1047 2. Find an augmenting path form source to sink.
1048 3. Send flow equal to the path's residual capacity along the edges of this path.
1049 4. Repeat steps 2 and 3 until no new augmenting path is found.
1051 Parameters:
1052 SOURCE: index of source vertex (input)
1053 SINK: index of sink vertex (input)
1054 FIXUP_GRAPH: adjacency matrix representing the graph. The flow of the edges will be
1055 set to have a valid maximal flow by this routine. (input)
1056 Return: Maximum flow possible. */
1058 static gcov_type
1059 find_max_flow (fixup_graph_type *fixup_graph, int source, int sink)
1061 int fnum_edges;
1062 augmenting_path_type augmenting_path;
1063 int *bb_pred;
1064 gcov_type max_flow = 0;
1065 int i, u;
1066 fixup_edge_p fedge_list, pfedge, r_pfedge;
1068 gcc_assert (fixup_graph);
1070 fnum_edges = fixup_graph->num_edges;
1071 fedge_list = fixup_graph->edge_list;
1073 /* Initialize flow to 0. */
1074 for (i = 0; i < fnum_edges; i++)
1076 pfedge = fedge_list + i;
1077 pfedge->flow = 0;
1080 compute_residual_flow (fixup_graph);
1082 init_augmenting_path (&augmenting_path, fixup_graph->num_vertices);
1084 bb_pred = augmenting_path.bb_pred;
1085 while (find_augmenting_path (fixup_graph, &augmenting_path, source, sink))
1087 /* Determine the amount by which we can increment the flow. */
1088 gcov_type increment = CAP_INFINITY;
1089 for (u = sink; u != source; u = bb_pred[u])
1091 pfedge = find_fixup_edge (fixup_graph, bb_pred[u], u);
1092 increment = MIN (increment, pfedge->rflow);
1094 max_flow += increment;
1096 /* Now increment the flow. EXIT vertex index is 1. */
1097 for (u = sink; u != source; u = bb_pred[u])
1099 pfedge = find_fixup_edge (fixup_graph, bb_pred[u], u);
1100 r_pfedge = find_fixup_edge (fixup_graph, u, bb_pred[u]);
1102 if (pfedge->rflow != CAP_INFINITY)
1103 pfedge->rflow -= increment;
1104 if (r_pfedge->rflow != CAP_INFINITY)
1105 r_pfedge->rflow += increment;
1107 if (pfedge->type)
1109 /* forward edge. */
1110 pfedge->flow += increment;
1112 else
1114 /* backward edge. */
1115 gcc_assert (r_pfedge->type);
1116 r_pfedge->flow -= increment;
1120 if (dump_file)
1122 fprintf (dump_file, "\nDump augmenting path:\n");
1123 for (u = sink; u != source; u = bb_pred[u])
1125 print_basic_block (dump_file, fixup_graph, u);
1126 fprintf (dump_file, "<-");
1128 fprintf (dump_file,
1129 "ENTRY (path_capacity=" HOST_WIDEST_INT_PRINT_DEC ")\n",
1130 increment);
1131 fprintf (dump_file,
1132 "Network flow is " HOST_WIDEST_INT_PRINT_DEC ".\n",
1133 max_flow);
1137 free_augmenting_path (&augmenting_path);
1138 if (dump_file)
1139 dump_fixup_graph (dump_file, fixup_graph, "After find_max_flow()");
1140 return max_flow;
1144 /* Computes the corrected edge and basic block weights using FIXUP_GRAPH
1145 after applying the find_minimum_cost_flow() routine. */
1147 static void
1148 adjust_cfg_counts (fixup_graph_type *fixup_graph)
1150 basic_block bb;
1151 edge e;
1152 edge_iterator ei;
1153 int i, j;
1154 fixup_edge_p pfedge, pfedge_n;
1156 gcc_assert (fixup_graph);
1158 if (dump_file)
1159 fprintf (dump_file, "\nadjust_cfg_counts():\n");
1161 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR_FOR_FN (cfun),
1162 EXIT_BLOCK_PTR_FOR_FN (cfun), next_bb)
1164 i = 2 * bb->index;
1166 /* Fixup BB. */
1167 if (dump_file)
1168 fprintf (dump_file,
1169 "BB%d: " HOST_WIDEST_INT_PRINT_DEC "", bb->index, bb->count);
1171 pfedge = find_fixup_edge (fixup_graph, i, i + 1);
1172 if (pfedge->flow)
1174 bb->count += pfedge->flow;
1175 if (dump_file)
1177 fprintf (dump_file, " + " HOST_WIDEST_INT_PRINT_DEC "(",
1178 pfedge->flow);
1179 print_edge (dump_file, fixup_graph, i, i + 1);
1180 fprintf (dump_file, ")");
1184 pfedge_n =
1185 find_fixup_edge (fixup_graph, i + 1, pfedge->norm_vertex_index);
1186 /* Deduct flow from normalized reverse edge. */
1187 if (pfedge->norm_vertex_index && pfedge_n->flow)
1189 bb->count -= pfedge_n->flow;
1190 if (dump_file)
1192 fprintf (dump_file, " - " HOST_WIDEST_INT_PRINT_DEC "(",
1193 pfedge_n->flow);
1194 print_edge (dump_file, fixup_graph, i + 1,
1195 pfedge->norm_vertex_index);
1196 fprintf (dump_file, ")");
1199 if (dump_file)
1200 fprintf (dump_file, " = " HOST_WIDEST_INT_PRINT_DEC "\n", bb->count);
1202 /* Fixup edge. */
1203 FOR_EACH_EDGE (e, ei, bb->succs)
1205 /* Treat edges with ignore attribute set as if they don't exist. */
1206 if (EDGE_INFO (e) && EDGE_INFO (e)->ignore)
1207 continue;
1209 j = 2 * e->dest->index;
1210 if (dump_file)
1211 fprintf (dump_file, "%d->%d: " HOST_WIDEST_INT_PRINT_DEC "",
1212 bb->index, e->dest->index, e->count);
1214 pfedge = find_fixup_edge (fixup_graph, i + 1, j);
1216 if (bb->index != e->dest->index)
1218 /* Non-self edge. */
1219 if (pfedge->flow)
1221 e->count += pfedge->flow;
1222 if (dump_file)
1224 fprintf (dump_file, " + " HOST_WIDEST_INT_PRINT_DEC "(",
1225 pfedge->flow);
1226 print_edge (dump_file, fixup_graph, i + 1, j);
1227 fprintf (dump_file, ")");
1231 pfedge_n =
1232 find_fixup_edge (fixup_graph, j, pfedge->norm_vertex_index);
1233 /* Deduct flow from normalized reverse edge. */
1234 if (pfedge->norm_vertex_index && pfedge_n->flow)
1236 e->count -= pfedge_n->flow;
1237 if (dump_file)
1239 fprintf (dump_file, " - " HOST_WIDEST_INT_PRINT_DEC "(",
1240 pfedge_n->flow);
1241 print_edge (dump_file, fixup_graph, j,
1242 pfedge->norm_vertex_index);
1243 fprintf (dump_file, ")");
1247 else
1249 /* Handle self edges. Self edge is split with a normalization
1250 vertex. Here i=j. */
1251 pfedge = find_fixup_edge (fixup_graph, j, i + 1);
1252 pfedge_n =
1253 find_fixup_edge (fixup_graph, i + 1, pfedge->norm_vertex_index);
1254 e->count += pfedge_n->flow;
1255 bb->count += pfedge_n->flow;
1256 if (dump_file)
1258 fprintf (dump_file, "(self edge)");
1259 fprintf (dump_file, " + " HOST_WIDEST_INT_PRINT_DEC "(",
1260 pfedge_n->flow);
1261 print_edge (dump_file, fixup_graph, i + 1,
1262 pfedge->norm_vertex_index);
1263 fprintf (dump_file, ")");
1267 if (bb->count)
1268 e->probability = REG_BR_PROB_BASE * e->count / bb->count;
1269 if (dump_file)
1270 fprintf (dump_file, " = " HOST_WIDEST_INT_PRINT_DEC "\t(%.1f%%)\n",
1271 e->count, e->probability * 100.0 / REG_BR_PROB_BASE);
1275 ENTRY_BLOCK_PTR_FOR_FN (cfun)->count =
1276 sum_edge_counts (ENTRY_BLOCK_PTR_FOR_FN (cfun)->succs);
1277 EXIT_BLOCK_PTR_FOR_FN (cfun)->count =
1278 sum_edge_counts (EXIT_BLOCK_PTR_FOR_FN (cfun)->preds);
1280 /* Compute edge probabilities. */
1281 FOR_ALL_BB_FN (bb, cfun)
1283 if (bb->count)
1285 FOR_EACH_EDGE (e, ei, bb->succs)
1286 e->probability = REG_BR_PROB_BASE * e->count / bb->count;
1288 else
1290 int total = 0;
1291 FOR_EACH_EDGE (e, ei, bb->succs)
1292 if (!(e->flags & (EDGE_COMPLEX | EDGE_FAKE)))
1293 total++;
1294 if (total)
1296 FOR_EACH_EDGE (e, ei, bb->succs)
1298 if (!(e->flags & (EDGE_COMPLEX | EDGE_FAKE)))
1299 e->probability = REG_BR_PROB_BASE / total;
1300 else
1301 e->probability = 0;
1304 else
1306 total += EDGE_COUNT (bb->succs);
1307 FOR_EACH_EDGE (e, ei, bb->succs)
1308 e->probability = REG_BR_PROB_BASE / total;
1313 if (dump_file)
1315 fprintf (dump_file, "\nCheck %s() CFG flow conservation:\n",
1316 current_function_name ());
1317 FOR_EACH_BB_FN (bb, cfun)
1319 if ((bb->count != sum_edge_counts (bb->preds))
1320 || (bb->count != sum_edge_counts (bb->succs)))
1322 fprintf (dump_file,
1323 "BB%d(" HOST_WIDEST_INT_PRINT_DEC ") **INVALID**: ",
1324 bb->index, bb->count);
1325 fprintf (stderr,
1326 "******** BB%d(" HOST_WIDEST_INT_PRINT_DEC
1327 ") **INVALID**: \n", bb->index, bb->count);
1328 fprintf (dump_file, "in_edges=" HOST_WIDEST_INT_PRINT_DEC " ",
1329 sum_edge_counts (bb->preds));
1330 fprintf (dump_file, "out_edges=" HOST_WIDEST_INT_PRINT_DEC "\n",
1331 sum_edge_counts (bb->succs));
1338 /* Called before negative_cycle_cancellation, to form a cycle between
1339 * new_exit to new_entry in FIXUP_GRAPH with capacity MAX_FLOW. We
1340 * don't want the flow in the BALANCE_EDGE to be modified, so we set
1341 * the residural flow of those edges to 0 */
1343 static void
1344 modify_sink_source_capacity (fixup_graph_type *fixup_graph, gcov_type max_flow)
1346 fixup_edge_p edge, r_edge;
1347 int i;
1348 int entry = ENTRY_BLOCK;
1349 int exit = 2 * EXIT_BLOCK + 1;
1350 int new_entry = fixup_graph->new_entry_index;
1351 int new_exit = fixup_graph->new_exit_index;
1353 edge = find_fixup_edge (fixup_graph, new_entry, entry);
1354 edge->max_capacity = CAP_INFINITY;
1355 edge->rflow = CAP_INFINITY;
1357 edge = find_fixup_edge (fixup_graph, entry, new_entry);
1358 edge->max_capacity = CAP_INFINITY;
1359 edge->rflow = CAP_INFINITY;
1361 edge = find_fixup_edge (fixup_graph, exit, new_exit);
1362 edge->max_capacity = CAP_INFINITY;
1363 edge->rflow = CAP_INFINITY;
1365 edge = find_fixup_edge (fixup_graph, new_exit, exit);
1366 edge->max_capacity = CAP_INFINITY;
1367 edge->rflow = CAP_INFINITY;
1369 edge = find_fixup_edge (fixup_graph, new_exit, new_entry);
1370 edge->max_capacity = CAP_INFINITY;
1371 edge->flow = max_flow;
1372 edge->rflow = CAP_INFINITY;
1374 r_edge = find_fixup_edge (fixup_graph, new_entry, new_exit);
1375 r_edge->rflow = max_flow;
1377 /* Find all the backwards residual edges corresponding to
1378 BALANCE_EDGEs and set their residual flow to 0 to enforce a
1379 minimum flow constraint on these edges. */
1380 for (i = 4; i < new_entry; i += 1)
1382 edge = find_fixup_edge (fixup_graph, i, new_entry);
1383 if (edge)
1384 edge->rflow = 0;
1385 edge = find_fixup_edge (fixup_graph, new_exit, i);
1386 if (edge)
1387 edge->rflow = 0;
1392 /* Implements the negative cycle canceling algorithm to compute a minimum cost
1393 flow.
1394 Algorithm:
1395 1. Find maximal flow.
1396 2. Form residual network
1397 3. Repeat:
1398 While G contains a negative cost cycle C, reverse the flow on the found cycle
1399 by the minimum residual capacity in that cycle.
1400 4. Form the minimal cost flow
1401 f(u,v) = rf(v, u)
1402 Input:
1403 FIXUP_GRAPH - Initial fixup graph.
1404 The flow field is modified to represent the minimum cost flow. */
1406 static void
1407 find_minimum_cost_flow (fixup_graph_type *fixup_graph)
1409 /* Holds the index of predecessor in path. */
1410 int *pred;
1411 /* Used to hold the minimum cost cycle. */
1412 int *cycle;
1413 /* Used to record the number of iterations of cancel_negative_cycle. */
1414 int iteration;
1415 /* Vector d[i] holds the minimum cost of path from i to sink. */
1416 gcov_type *d;
1417 int fnum_vertices;
1418 int new_exit_index;
1419 int new_entry_index;
1420 gcov_type max_flow;
1422 gcc_assert (fixup_graph);
1423 fnum_vertices = fixup_graph->num_vertices;
1424 new_exit_index = fixup_graph->new_exit_index;
1425 new_entry_index = fixup_graph->new_entry_index;
1427 max_flow = find_max_flow (fixup_graph, new_entry_index, new_exit_index);
1429 /* Adjust the fixup graph to translate things into a minimum cost
1430 circulation problem. */
1431 modify_sink_source_capacity (fixup_graph, max_flow);
1433 /* Initialize the structures for find_negative_cycle(). */
1434 pred = (int *) xcalloc (fnum_vertices, sizeof (int));
1435 d = (gcov_type *) xcalloc (fnum_vertices, sizeof (gcov_type));
1436 cycle = (int *) xcalloc (fnum_vertices, sizeof (int));
1438 /* Repeatedly find and cancel negative cost cycles, until
1439 no more negative cycles exist. This also updates the flow field
1440 to represent the minimum cost flow so far. */
1441 iteration = 0;
1442 while (cancel_negative_cycle (fixup_graph, pred, d, cycle))
1444 iteration++;
1445 if (iteration > MAX_ITER (fixup_graph->num_vertices,
1446 fixup_graph->num_edges))
1448 if (dump_enabled_p ())
1449 dump_printf_loc (MSG_NOTE,
1450 DECL_SOURCE_LOCATION (current_function_decl),
1451 "Exiting profile correction early to avoid "
1452 "excessive compile time");
1453 break;
1457 if (dump_file)
1458 dump_fixup_graph (dump_file, fixup_graph,
1459 "After find_minimum_cost_flow()");
1461 /* Cleanup structures. */
1462 free (pred);
1463 free (d);
1464 free (cycle);
1468 /* Compute the sum of the edge counts in TO_EDGES. */
1470 gcov_type
1471 sum_edge_counts (vec<edge, va_gc> *to_edges)
1473 gcov_type sum = 0;
1474 edge e;
1475 edge_iterator ei;
1477 FOR_EACH_EDGE (e, ei, to_edges)
1479 if (EDGE_INFO (e) && EDGE_INFO (e)->ignore)
1480 continue;
1481 sum += e->count;
1483 return sum;
1487 /* Main routine. Smoothes the initial assigned basic block and edge counts using
1488 a minimum cost flow algorithm, to ensure that the flow consistency rule is
1489 obeyed: sum of outgoing edges = sum of incoming edges for each basic
1490 block. */
1492 void
1493 mcf_smooth_cfg (void)
1495 fixup_graph_type fixup_graph;
1496 memset (&fixup_graph, 0, sizeof (fixup_graph));
1497 create_fixup_graph (&fixup_graph);
1498 find_minimum_cost_flow (&fixup_graph);
1499 adjust_cfg_counts (&fixup_graph);
1500 delete_fixup_graph (&fixup_graph);