Split RN into 2 phases.
[libfirm] / heuristical_co.c
1 /*
2  * Copyright (C) 1995-2008 University of Karlsruhe.  All right reserved.
3  *
4  * This file is part of libFirm.
5  *
6  * This file may be distributed and/or modified under the terms of the
7  * GNU General Public License version 2 as published by the Free Software
8  * Foundation and appearing in the file LICENSE.GPL included in the
9  * packaging of this file.
10  *
11  * Licensees holding valid libFirm Professional Edition licenses may use
12  * this file in accordance with the libFirm Commercial License.
13  * Agreement provided with the Software.
14  *
15  * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16  * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
17  * PURPOSE.
18  */
19
20 /**
21  * @file
22  * @brief   Heuristic PBQP solver for SSA-based register allocation.
23  * @date    18.09.2009
24  * @author  Thomas Bersch
25  * @version $Id$
26  */
27 #include "config.h"
28
29 #include "adt/array.h"
30 #include "assert.h"
31 #include "error.h"
32
33 #include "bucket.h"
34 #include "heuristical_co.h"
35 #include "optimal.h"
36 #if     KAPS_DUMP
37 #include "html_dumper.h"
38 #endif
39 #include "kaps.h"
40 #include "matrix.h"
41 #include "pbqp_edge.h"
42 #include "pbqp_edge_t.h"
43 #include "pbqp_node.h"
44 #include "pbqp_node_t.h"
45 #include "vector.h"
46
47 #include "plist.h"
48 #include "timing.h"
49
50 static void merge_into_RN_node(pbqp *pbqp, plist_t *rpeo)
51 {
52         pbqp_node   *node         = NULL;
53
54         assert(pbqp);
55
56         /* We want to reduce the first node in reverse perfect elimination order. */
57         do {
58                 /* get first element from reverse perfect elimination order */
59                 node = plist_first(rpeo)->data;
60                 /* remove element from reverse perfect elimination order */
61                 plist_erase(rpeo, plist_first(rpeo));
62                 /* insert node at the end of rpeo so the rpeo already exits after pbqp solving */
63                 plist_insert_back(rpeo, node);
64         } while(node_is_reduced(node));
65
66         assert(node);
67         assert(pbqp_node_get_degree(node) > 2);
68
69 #if     KAPS_DUMP
70         if (pbqp->dump_file) {
71                 char     txt[100];
72                 sprintf(txt, "RN-Reduction of Node n%d", node->index);
73                 dump_section(pbqp->dump_file, 2, txt);
74                 pbqp_dump_graph(pbqp);
75         }
76 #endif
77         /* Check whether we can merge a neighbor into the current node. */
78         apply_RM(pbqp, node);
79 }
80
81 static void apply_RN_co(pbqp *pbqp)
82 {
83         pbqp_node   *node      = NULL;
84         unsigned     min_index = 0;
85
86         assert(pbqp);
87
88         node = node_bucket_pop(&rn_bucket);
89         assert(node);
90
91         if (node_is_reduced(node))
92                 return;
93
94         min_index = get_local_minimal_alternative(pbqp, node);
95
96 #if     KAPS_DUMP
97         if (pbqp->dump_file) {
98                 fprintf(pbqp->dump_file, "node n%d is set to %d<br><br>\n",
99                                         node->index, min_index);
100         }
101 #endif
102
103 #if KAPS_STATISTIC
104                 FILE *fh = fopen("solutions.pb", "a");
105                 fprintf(fh, "[%u]", min_index);
106                 fclose(fh);
107                 pbqp->num_rn++;
108 #endif
109
110         /* Now that we found the local minimum set all other costs to infinity. */
111         select_alternative(node, min_index);
112 }
113
114 static void apply_heuristic_reductions_co(pbqp *pbqp, plist_t *rpeo)
115 {
116         #if KAPS_TIMING
117                 /* create timers */
118                 ir_timer_t *t_edge = ir_timer_new();
119                 ir_timer_t *t_r1   = ir_timer_new();
120                 ir_timer_t *t_r2   = ir_timer_new();
121                 ir_timer_t *t_rn   = ir_timer_new();
122         #endif
123
124         for (;;) {
125                 if (edge_bucket_get_length(edge_bucket) > 0) {
126                         #if KAPS_TIMING
127                                 ir_timer_start(t_edge);
128                         #endif
129
130                         apply_edge(pbqp);
131
132                         #if KAPS_TIMING
133                                 ir_timer_stop(t_edge);
134                         #endif
135                 } else if (node_bucket_get_length(node_buckets[1]) > 0) {
136                         #if KAPS_TIMING
137                                 ir_timer_start(t_r1);
138                         #endif
139
140                         apply_RI(pbqp);
141
142                         #if KAPS_TIMING
143                                 ir_timer_stop(t_r1);
144                         #endif
145                 } else if (node_bucket_get_length(node_buckets[2]) > 0) {
146                         #if KAPS_TIMING
147                                 ir_timer_start(t_r2);
148                         #endif
149
150                         apply_RII(pbqp);
151
152                         #if KAPS_TIMING
153                                 ir_timer_stop(t_r2);
154                         #endif
155                 } else if (node_bucket_get_length(rn_bucket) > 0) {
156                         #if KAPS_TIMING
157                                 ir_timer_start(t_rn);
158                         #endif
159
160                         apply_RN_co(pbqp);
161
162                         #if KAPS_TIMING
163                                 ir_timer_stop(t_rn);
164                         #endif
165                 } else if (node_bucket_get_length(node_buckets[3]) > 0) {
166                         #if KAPS_TIMING
167                                 ir_timer_start(t_rn);
168                         #endif
169
170                         merge_into_RN_node(pbqp, rpeo);
171
172                         #if KAPS_TIMING
173                                 ir_timer_stop(t_rn);
174                         #endif
175                 } else {
176                         #if KAPS_TIMING
177                                 printf("PBQP RE reductions:           %10.3lf msec\n", (double)ir_timer_elapsed_usec(t_edge) / 1000.0);
178                                 printf("PBQP R1 reductions:           %10.3lf msec\n", (double)ir_timer_elapsed_usec(t_r1) / 1000.0);
179                                 printf("PBQP R2 reductions:           %10.3lf msec\n", (double)ir_timer_elapsed_usec(t_r2) / 1000.0);
180                                 printf("PBQP RN reductions:           %10.3lf msec\n", (double)ir_timer_elapsed_usec(t_rn) / 1000.0);
181                         #endif
182
183                         return;
184                 }
185         }
186 }
187
188 void solve_pbqp_heuristical_co(pbqp *pbqp, plist_t *rpeo)
189 {
190         /* Reduce nodes degree ... */
191         initial_simplify_edges(pbqp);
192
193         /* ... and put node into bucket representing their degree. */
194         fill_node_buckets(pbqp);
195
196         #if KAPS_STATISTIC
197                 FILE *fh = fopen("solutions.pb", "a");
198                 fprintf(fh, "Solution");
199                 fclose(fh);
200         #endif
201
202         apply_heuristic_reductions_co(pbqp, rpeo);
203
204         pbqp->solution = determine_solution(pbqp);
205
206         #if KAPS_STATISTIC
207                 fh = fopen("solutions.pb", "a");
208                 #if KAPS_USE_UNSIGNED
209                         fprintf(fh, ": %u RE:%u R0:%u R1:%u R2:%u RM:%u RN/BF:%u\n", pbqp->solution,
210                                         pbqp->num_edges, pbqp->num_r0, pbqp->num_r1, pbqp->num_r2,
211                                         pbqp->num_rm, pbqp->num_rn);
212                 #else
213                         fprintf(fh, ": %lld RE:%u R0:%u R1:%u R2:%u RM:%u RN/BF:%u\n", pbqp->solution,
214                                         pbqp->num_edges, pbqp->num_r0, pbqp->num_r1, pbqp->num_r2,
215                 #endif
216                 fclose(fh);
217         #endif
218
219         /* Solve reduced nodes. */
220         back_propagate(pbqp);
221
222         free_buckets();
223 }