2 * Copyright (C) 1995-2011 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 Naive spilling algorithm
23 * @author Matthias Braun
26 * This implements a naive spilling algorithm. It is designed to produce
27 * similar effects to the spill decisions produced by traditional graph
28 * coloring register allocators that spill while they are coloring the graph.
30 * This spiller walks over all blocks and looks for places with too high
31 * register pressure where it spills the values that are cheapest to spill.
32 * Spilling in this context means placing a spill instruction behind the
33 * definition of the value and a reload before each usage.
39 #include "irnodeset.h"
42 #include "iredges_t.h"
47 #include "bespillutil.h"
56 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
58 static spill_env_t *spill_env;
59 static unsigned n_regs;
60 static const arch_register_class_t *cls;
61 static const be_lv_t *lv;
62 static bitset_t *spilled_nodes;
64 typedef struct spill_candidate_t spill_candidate_t;
65 struct spill_candidate_t {
70 static int compare_spill_candidates_desc(const void *d1, const void *d2)
72 const spill_candidate_t *c1 = (const spill_candidate_t*)d1;
73 const spill_candidate_t *c2 = (const spill_candidate_t*)d2;
75 return (int) (c1->costs - c2->costs);
78 static double get_spill_costs(ir_node *node)
80 ir_node *spill_place = skip_Proj(node);
81 double costs = be_get_spill_costs(spill_env, node, spill_place);
83 foreach_out_edge(node, edge) {
84 ir_node *use = get_edge_src_irn(edge);
86 /* keeps should be directly below the node */
87 if (be_is_Keep(use)) {
92 int in = get_edge_src_pos(edge);
93 ir_node *block = get_nodes_block(use);
95 costs += be_get_reload_costs_on_edge(spill_env, node, block, in);
97 costs += be_get_reload_costs(spill_env, node, use);
105 * spills a node by placing a reload before each usage
107 static void spill_node(ir_node *node)
109 DBG((dbg, LEVEL_3, "\tspilling %+F\n", node));
111 foreach_out_edge(node, edge) {
112 ir_node *use = get_edge_src_irn(edge);
119 int in = get_edge_src_pos(edge);
120 ir_node *block = get_nodes_block(use);
122 be_add_reload_on_edge(spill_env, node, block, in, cls, 1);
124 be_add_reload(spill_env, node, use, cls, 1);
128 bitset_set(spilled_nodes, get_irn_idx(node));
131 static unsigned get_value_width(const ir_node *node)
133 const arch_register_req_t *req = arch_get_irn_register_req(node);
138 * spill @p n nodes from a nodeset. Removes the nodes from the nodeset and
139 * sets the spilled bits in spilled_nodes.
141 static void do_spilling(ir_nodeset_t *live_nodes, ir_node *node)
143 size_t n_live_nodes = ir_nodeset_size(live_nodes);
144 size_t values_defined = 0;
145 size_t free_regs_needed = 0;
146 spill_candidate_t *candidates;
147 ir_nodeset_iterator_t iter;
155 be_foreach_definition(node, cls, value,
156 assert(req_->width >= 1);
157 values_defined += req_->width;
160 /* we need registers for the non-live argument values */
161 arity = get_irn_arity(node);
162 for (i = 0; i < arity; ++i) {
163 ir_node *pred = get_irn_n(node, i);
164 if (arch_irn_consider_in_reg_alloc(cls, pred)
165 && !ir_nodeset_contains(live_nodes, pred)) {
166 free_regs_needed += get_value_width(pred);
170 /* we can reuse all reloaded values for the defined values, but we might
171 * need even more registers */
172 if (values_defined > free_regs_needed)
173 free_regs_needed = values_defined;
175 spills_needed = (n_live_nodes + free_regs_needed) - n_regs;
176 if (spills_needed <= 0)
178 DBG((dbg, LEVEL_2, "\tspills needed after %+F: %d\n", node, spills_needed));
180 candidates = ALLOCAN(spill_candidate_t, n_live_nodes);
182 /* construct array with spill candidates and calculate their costs */
184 foreach_ir_nodeset(live_nodes, n, iter) {
185 spill_candidate_t *candidate = & candidates[c];
187 assert(!bitset_is_set(spilled_nodes, get_irn_idx(n)));
190 candidate->costs = get_spill_costs(n);
193 assert(c == n_live_nodes);
195 /* sort spill candidates */
196 qsort(candidates, n_live_nodes, sizeof(candidates[0]),
197 compare_spill_candidates_desc);
199 /* spill cheapest ones */
201 while (spills_needed > 0) {
203 spill_candidate_t *candidate;
206 if (cand_idx >= n_live_nodes) {
207 panic("can't spill enough values for node %+F", node);
210 candidate = &candidates[cand_idx];
211 cand_node = candidate->node;
214 if (arch_irn_is(skip_Proj_const(cand_node), dont_spill))
217 /* make sure the node is not an argument of the instruction */
218 for (i = 0; i < arity; ++i) {
219 ir_node *in = get_irn_n(node, i);
220 if (in == cand_node) {
228 spill_node(cand_node);
229 ir_nodeset_remove(live_nodes, cand_node);
230 spills_needed -= get_value_width(cand_node);
235 * removes all values from the nodeset that are defined by node
237 static void remove_defs(ir_node *node, ir_nodeset_t *nodeset)
240 /* You must break out of your loop when hitting the first phi function. */
241 assert(!is_Phi(node));
243 be_foreach_definition(node, cls, value,
244 ir_nodeset_remove(nodeset, value);
248 static void add_uses(ir_node *node, ir_nodeset_t *nodeset)
252 arity = get_irn_arity(node);
253 for (i = 0; i < arity; ++i) {
254 ir_node *op = get_irn_n(node, i);
256 if (arch_irn_consider_in_reg_alloc(cls, op) &&
257 !bitset_is_set(spilled_nodes, get_irn_idx(op))) {
258 ir_nodeset_insert(nodeset, op);
263 static __attribute__((unused))
264 void print_nodeset(ir_nodeset_t *nodeset)
266 ir_nodeset_iterator_t iter;
269 foreach_ir_nodeset(nodeset, node, iter) {
270 ir_fprintf(stderr, "%+F ", node);
272 fprintf(stderr, "\n");
276 * make sure register pressure in a block is always equal or below the number
277 * of available registers
279 static void spill_block(ir_node *block, void *data)
281 ir_nodeset_t live_nodes;
282 ir_nodeset_iterator_t iter;
283 int n_phi_values_spilled;
285 int live_nodes_pressure;
286 int phi_spills_needed;
289 DBG((dbg, LEVEL_1, "spilling block %+F\n", block));
291 /* construct set of live nodes at end of block */
292 ir_nodeset_init(&live_nodes);
293 be_liveness_end_of_block(lv, cls, block, &live_nodes);
295 /* remove already spilled nodes from liveset */
298 foreach_ir_nodeset(&live_nodes, node, iter) {
299 DBG((dbg, LEVEL_2, "\t%+F is live-end... ", node));
300 if (bitset_is_set(spilled_nodes, get_irn_idx(node))) {
301 DBG((dbg, LEVEL_2, "but spilled; removing.\n"));
302 ir_nodeset_remove_iterator(&live_nodes, &iter);
304 DBG((dbg, LEVEL_2, "keeping.\n"));
309 /* walk schedule backwards and spill until register pressure is fine at
311 sched_foreach_reverse(block, node) {
315 remove_defs(node, &live_nodes);
316 do_spilling(&live_nodes, node);
317 add_uses(node, &live_nodes);
320 /* until now only the values of some phis have been spilled the phis itself
321 * are still there and occupy registers, so we need to count them and might
322 * have to spill some of them. */
323 n_phi_values_spilled = 0;
324 sched_foreach(block, node) {
328 if (bitset_is_set(spilled_nodes, get_irn_idx(node))) {
329 n_phi_values_spilled += get_value_width(node);
333 live_nodes_pressure = 0;
336 foreach_ir_nodeset(&live_nodes, node, iter) {
337 live_nodes_pressure += get_value_width(node);
341 /* calculate how many of the phis need to be spilled */
342 regpressure = live_nodes_pressure + n_phi_values_spilled;
343 phi_spills_needed = regpressure - n_regs;
344 DBG((dbg, LEVEL_3, "Regpressure before phis: %d phispills: %d\n",
345 regpressure, phi_spills_needed));
347 /* spill as many phis as needed */
348 /* TODO: we should really estimate costs of the phi spill as well...
349 * and preferably spill phis with lower costs... */
350 sched_foreach(block, node) {
353 if (phi_spills_needed <= 0)
356 if (!bitset_is_set(spilled_nodes, get_irn_idx(node)))
359 be_spill_phi(spill_env, node);
360 phi_spills_needed -= get_value_width(node);
362 assert(phi_spills_needed <= 0);
364 ir_nodeset_destroy(&live_nodes);
367 static void be_spill_daemel(ir_graph *irg, const arch_register_class_t *new_cls)
369 n_regs = be_get_n_allocatable_regs(irg, new_cls);
373 be_assure_live_sets(irg);
375 spill_env = be_new_spill_env(irg);
377 lv = be_get_irg_liveness(irg);
378 spilled_nodes = bitset_malloc(get_irg_last_idx(irg));
380 DBG((dbg, LEVEL_1, "*** RegClass %s\n", cls->name));
382 irg_block_walk_graph(irg, spill_block, NULL, NULL);
384 bitset_free(spilled_nodes);
386 be_insert_spills_reloads(spill_env);
387 be_delete_spill_env(spill_env);
390 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_daemelspill)
391 void be_init_daemelspill(void)
393 static be_spiller_t daemel_spiller = {
397 be_register_spiller("daemel", &daemel_spiller);
398 FIRM_DBG_REGISTER(dbg, "firm.be.spilldaemel");