2 * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
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.
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.
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
22 * @brief Naiv spilling algorithm
23 * @author Matthias Braun
25 * @version $Id: bespillbelady.c 13913 2007-05-18 12:48:56Z matze $
27 * This implements a naiv spilling algorithm. It is design to produce similar
28 * effects to the spill decisions produced by traditional graph coloring
29 * register allocators that spill while they are coloring the graph.
31 * This spiller walks over all blocks and looks for places with too high
32 * register pressure where it spills the values that are cheapest to spill.
33 * Spilling in this context means placing a spill instruction behind the
34 * definition of the value and a reload before each usage.
42 #include "irnodeset.h"
45 #include "iredges_t.h"
49 #include "bespilloptions.h"
58 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
60 static spill_env_t *spill_env;
62 static const arch_register_class_t *cls;
63 static const be_lv_t *lv;
64 static bitset_t *spilled_nodes;
66 typedef struct spill_candidate_t spill_candidate_t;
67 struct spill_candidate_t {
72 static int compare_spill_candidates_desc(const void *d1, const void *d2)
74 const spill_candidate_t *c1 = d1;
75 const spill_candidate_t *c2 = d2;
77 return (int) (c1->costs - c2->costs);
80 static double get_spill_costs(ir_node *node)
82 const ir_edge_t *edge;
83 ir_node *spill_place = skip_Proj(node);
84 double costs = be_get_spill_costs(spill_env, node,
87 foreach_out_edge(node, edge) {
88 ir_node *use = get_edge_src_irn(edge);
90 /* keeps should be directly below the node */
96 int in = get_edge_src_pos(edge);
97 ir_node *block = get_nodes_block(use);
99 costs += be_get_reload_costs_on_edge(spill_env, node, block, in);
101 costs += be_get_reload_costs(spill_env, node, use);
105 /* TODO cache costs? */
111 * spills a node by placing a reload before each usage
113 static void spill_node(ir_node *node)
115 const ir_edge_t *edge;
117 DBG((dbg, LEVEL_3, "\tspilling %+F\n", node));
119 foreach_out_edge(node, edge) {
120 ir_node *use = get_edge_src_irn(edge);
127 int in = get_edge_src_pos(edge);
128 ir_node *block = get_nodes_block(use);
130 be_add_reload_on_edge(spill_env, node, block, in, cls, 1);
132 be_add_reload(spill_env, node, use, cls, 1);
136 bitset_set(spilled_nodes, get_irn_idx(node));
140 * spill @p n nodes from a nodeset. Removes the nodes from the nodeset and
141 * sets the spilled bits in spilled_nodes.
143 static void do_spilling(ir_nodeset_t *live_nodes, ir_node *node)
145 size_t n_live_nodes = ir_nodeset_size(live_nodes);
146 size_t values_defined = 0;
147 size_t free_regs_needed = 0;
148 spill_candidate_t *candidates;
149 ir_nodeset_iterator_t iter;
155 /* mode_T nodes define several values at once. Count them */
156 if(get_irn_mode(node) == mode_T) {
157 const ir_edge_t *edge;
159 foreach_out_edge(node, edge) {
160 const ir_node *proj = get_edge_src_irn(edge);
162 if (arch_irn_consider_in_reg_alloc(cls, proj)) {
166 } else if (arch_irn_consider_in_reg_alloc(cls, node)) {
170 /* we need registers for the non-live argument values */
171 arity = get_irn_arity(node);
172 for(i = 0; i < arity; ++i) {
173 ir_node *pred = get_irn_n(node, i);
174 if (arch_irn_consider_in_reg_alloc(cls, pred)
175 && !ir_nodeset_contains(live_nodes, pred)) {
180 /* we can reuse all reloaded values for the defined values, but we might
181 need even more registers */
182 if(values_defined > free_regs_needed)
183 free_regs_needed = values_defined;
185 spills_needed = (n_live_nodes + free_regs_needed) - n_regs;
186 if(spills_needed <= 0)
188 DBG((dbg, LEVEL_2, "\tspills needed after %+F: %d\n", node, spills_needed));
190 candidates = alloca(n_live_nodes * sizeof(candidates[0]));
192 /* construct array with spill candidates and calculate their costs */
194 foreach_ir_nodeset(live_nodes, n, iter) {
195 spill_candidate_t *candidate = & candidates[i];
197 assert(!bitset_is_set(spilled_nodes, get_irn_idx(n)));
200 candidate->costs = get_spill_costs(n);
203 assert(i == n_live_nodes);
205 /* sort spill candidates */
206 qsort(candidates, n_live_nodes, sizeof(candidates[0]),
207 compare_spill_candidates_desc);
209 /* spill cheapest ones */
211 while(spills_needed > 0) {
212 spill_candidate_t *candidate;
216 if (cand_idx >= n_live_nodes) {
217 panic("can't spill enough values for node %+F", node);
221 candidate = &candidates[cand_idx];
222 cand_node = candidate->node;
225 if (arch_irn_is(cand_node, dont_spill))
228 /* make sure the node is not an argument of the instruction */
230 for (i = 0; i < arity; ++i) {
231 ir_node *in = get_irn_n(node, i);
232 if(in == cand_node) {
241 spill_node(cand_node);
242 ir_nodeset_remove(live_nodes, cand_node);
248 * removes all values from the nodeset that are defined by node
250 static void remove_defs(ir_node *node, ir_nodeset_t *nodeset)
252 /* You should better break out of your loop when hitting the first phi
254 assert(!is_Phi(node) && "liveness_transfer produces invalid results for phi nodes");
256 if (get_irn_mode(node) == mode_T) {
257 const ir_edge_t *edge;
259 foreach_out_edge(node, edge) {
260 const ir_node *proj = get_edge_src_irn(edge);
262 if (arch_irn_consider_in_reg_alloc(cls, proj)) {
263 ir_nodeset_remove(nodeset, proj);
268 if (arch_irn_consider_in_reg_alloc(cls, node)) {
269 ir_nodeset_remove(nodeset, node);
273 static void add_uses(ir_node *node, ir_nodeset_t *nodeset)
277 arity = get_irn_arity(node);
278 for(i = 0; i < arity; ++i) {
279 ir_node *op = get_irn_n(node, i);
281 if (arch_irn_consider_in_reg_alloc(cls, op) &&
282 !bitset_is_set(spilled_nodes, get_irn_idx(op))) {
283 ir_nodeset_insert(nodeset, op);
288 static __attribute__((unused))
289 void print_nodeset(ir_nodeset_t *nodeset)
291 ir_nodeset_iterator_t iter;
294 foreach_ir_nodeset(nodeset, node, iter) {
295 ir_fprintf(stderr, "%+F ", node);
297 fprintf(stderr, "\n");
301 * make sure register pressure in a block is always equal or below the number
302 * of available registers
304 static void spill_block(ir_node *block, void *data)
306 ir_nodeset_t live_nodes;
307 ir_nodeset_iterator_t iter;
309 int n_phi_values_spilled;
311 int phi_spills_needed;
314 DBG((dbg, LEVEL_1, "spilling block %+F\n", block));
316 /* construct set of live nodes at end of block */
317 ir_nodeset_init(&live_nodes);
318 be_liveness_end_of_block(lv, cls, block, &live_nodes);
320 /* remove already spilled nodes from liveset */
321 foreach_ir_nodeset(&live_nodes, node, iter) {
322 DBG((dbg, LEVEL_2, "\t%+F is live-end... ", node));
323 if(bitset_is_set(spilled_nodes, get_irn_idx(node))) {
324 DBG((dbg, LEVEL_2, "but spilled; removing.\n"));
325 ir_nodeset_remove_iterator(&live_nodes, &iter);
327 DBG((dbg, LEVEL_2, "keeping.\n"));
331 /* walk schedule backwards and spill until register pressure is fine at
333 sched_foreach_reverse(block, node) {
337 remove_defs(node, &live_nodes);
338 do_spilling(&live_nodes, node);
339 add_uses(node, &live_nodes);
342 /* until now only the values of some phis have been spilled the phis itself
343 * are still there and occupy registers, so we need to count them and might
344 * have to spill some of them.
346 n_phi_values_spilled = 0;
347 sched_foreach(block, node) {
351 if(bitset_is_set(spilled_nodes, get_irn_idx(node))) {
352 ++n_phi_values_spilled;
356 /* calculate how many of the phis need to be spilled */
357 regpressure = ir_nodeset_size(&live_nodes) + n_phi_values_spilled;
358 phi_spills_needed = regpressure - n_regs;
359 DBG((dbg, LEVEL_3, "Regpressure before phis: %d phispills: %d\n",
360 regpressure, phi_spills_needed));
362 /* spill as many phis as needed */
363 /* TODO: we should really estimate costs of the phi spill as well...
364 * and preferably spill phis with lower costs... */
365 sched_foreach(block, node) {
368 if(phi_spills_needed <= 0)
371 if(bitset_is_set(spilled_nodes, get_irn_idx(node))) {
372 be_spill_phi(spill_env, node);
376 assert(phi_spills_needed <= 0);
378 ir_nodeset_destroy(&live_nodes);
381 void be_spill_daemel(be_irg_t *birg, const arch_register_class_t *new_cls)
383 ir_graph *irg = be_get_birg_irg(birg);
384 n_regs = new_cls->n_regs - be_put_ignore_regs(birg, new_cls, NULL);
389 be_liveness_assure_sets(be_assure_liveness(birg));
391 spill_env = be_new_spill_env(birg);
393 lv = be_get_birg_liveness(birg);
394 spilled_nodes = bitset_malloc(get_irg_last_idx(irg));
396 DBG((dbg, LEVEL_1, "*** RegClass %s\n", cls->name));
398 irg_block_walk_graph(irg, spill_block, NULL, NULL);
400 bitset_free(spilled_nodes);
401 spilled_nodes = NULL;
403 be_insert_spills_reloads(spill_env);
405 be_delete_spill_env(spill_env);
409 void be_init_daemelspill(void)
411 static be_spiller_t daemel_spiller = {
415 be_register_spiller("daemel", &daemel_spiller);
416 FIRM_DBG_REGISTER(dbg, "firm.be.spilldaemel");
419 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_daemelspill);