amd64: small changes w.r.t. stack alignment.
[libfirm] / ir / be / bespill.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       Spill module selection; Preparation steps
23  * @author      Matthias Braun
24  * @date        29.09.2005
25  * @version     $Id$
26  */
27 #include "config.h"
28
29 #include "irtools.h"
30 #include "debug.h"
31 #include "iredges_t.h"
32 #include "raw_bitset.h"
33 #include "statev.h"
34 #include "irgwalk.h"
35
36 #include "bespill.h"
37 #include "bemodule.h"
38 #include "be.h"
39 #include "belive_t.h"
40 #include "beirg.h"
41 #include "bearch.h"
42 #include "benode.h"
43 #include "besched.h"
44 #include "bera.h"
45 #include "beintlive_t.h"
46
47 #include "lc_opts.h"
48 #include "lc_opts_enum.h"
49
50 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
51
52 typedef struct be_pre_spill_env_t {
53         be_irg_t                    *birg;
54         const arch_register_class_t *cls;
55 } be_pre_spill_env_t;
56
57 static void prepare_constr_insn(be_pre_spill_env_t *env, ir_node *node)
58 {
59         const arch_register_class_t *cls = env->cls;
60         ir_node  *block      = get_nodes_block(node);
61         const be_irg_t *birg = env->birg;
62         be_lv_t *lv          = birg->lv;
63         unsigned *tmp        = NULL;
64         unsigned *def_constr = NULL;
65         int       arity      = get_irn_arity(node);
66
67         int i, i2;
68
69         /* Insert a copy for constraint inputs attached to a value which can't
70          * fullfil the constraint
71          * (typical example: stack pointer as input to copyb)
72          * TODO: This really just checks precolored registers at the moment and
73          *       ignore the general case of not matching in/out constraints
74          */
75         for (i = 0; i < arity; ++i) {
76                 ir_node *op = get_irn_n(node, i);
77                 ir_node *copy;
78                 const arch_register_t *reg;
79                 const arch_register_req_t *req;
80
81                 req = arch_get_register_req(node, i);
82                 if (req->cls != cls)
83                         continue;
84                 reg = arch_get_irn_register(op);
85                 if (reg == NULL)
86                         continue;
87
88                 /* precolored with an ignore register (which is not a joker like
89                    unknown/noreg) */
90                 if (arch_register_type_is(reg, joker)
91                                 || !arch_register_type_is(reg, ignore))
92                         continue;
93
94                 if (! (req->type & arch_register_req_type_limited))
95                         continue;
96                 if (rbitset_is_set(req->limited, reg->index))
97                         continue;
98
99                 copy = be_new_Copy(cls, block, op);
100                 stat_ev_int("constr_copy", 1);
101                 sched_add_before(node, copy);
102                 set_irn_n(node, i, copy);
103                 DBG((dbg, LEVEL_3, "inserting ignore arg copy %+F for %+F pos %d\n", copy, node, i));
104         }
105
106         /* insert copies for nodes that occur constrained more than once. */
107         for (i = 0; i < arity; ++i) {
108                 ir_node                   *in;
109                 ir_node                   *copy;
110                 const arch_register_req_t *req;
111
112                 req = arch_get_register_req(node, i);
113                 if (req->cls != cls)
114                         continue;
115
116                 if (! (req->type & arch_register_req_type_limited))
117                         continue;
118
119                 in = get_irn_n(node, i);
120                 if (!arch_irn_consider_in_reg_alloc(cls, in))
121                         continue;
122
123                 for (i2 = i + 1; i2 < arity; ++i2) {
124                         ir_node *in2;
125                         const arch_register_req_t *req2;
126
127                         req2 = arch_get_register_req(node, i2);
128                         if (req2->cls != cls)
129                                 continue;
130                         if (! (req2->type & arch_register_req_type_limited))
131                                 continue;
132
133                         in2 = get_irn_n(node, i2);
134                         if (in2 != in)
135                                 continue;
136
137                         /* if the constraint is the same, no copy is necessary
138                          * TODO generalise unequal but overlapping constraints */
139                         if (rbitsets_equal(req->limited, req2->limited, cls->n_regs))
140                                 continue;
141
142                         copy = be_new_Copy(cls, block, in);
143                         stat_ev_int("constr_copy", 1);
144
145                         sched_add_before(node, copy);
146                         set_irn_n(node, i2, copy);
147                         DBG((dbg, LEVEL_3,
148                              "inserting multiple constr copy %+F for %+F pos %d\n",
149                              copy, node, i2));
150                 }
151         }
152
153         /* collect all registers occurring in out constraints. */
154         if (get_irn_mode(node) == mode_T) {
155                 const ir_edge_t *edge;
156
157                 foreach_out_edge(node, edge) {
158                         ir_node                   *proj = get_edge_src_irn(edge);
159                         const arch_register_req_t *req  = arch_get_register_req_out(proj);
160                         if (! (req->type & arch_register_req_type_limited))
161                                 continue;
162
163                         if (def_constr == NULL) {
164                                 rbitset_alloca(def_constr, cls->n_regs);
165                         }
166                         rbitset_or(def_constr, req->limited, cls->n_regs);
167                 }
168         } else {
169                 const arch_register_req_t *req = arch_get_register_req_out(node);
170                 if (req->type & arch_register_req_type_limited) {
171                         rbitset_alloca(def_constr, cls->n_regs);
172                         rbitset_or(def_constr, req->limited, cls->n_regs);
173                 }
174         }
175
176         /* no output constraints => we're good */
177         if (def_constr == NULL) {
178                 return;
179         }
180
181         /*
182          * insert copies for all constrained arguments living through the node
183          * and being constrained to a register which also occurs in out constraints.
184          */
185         rbitset_alloca(tmp, cls->n_regs);
186         for (i = 0; i < arity; ++i) {
187                 const arch_register_req_t *req;
188                 ir_node                   *in;
189                 ir_node                   *copy;
190
191                 /*
192                  * Check, if
193                  * 1) the operand is constrained.
194                  * 2) lives through the node.
195                  * 3) is constrained to a register occurring in out constraints.
196                  */
197                 req = arch_get_register_req(node, i);
198                 if (req->cls != cls)
199                         continue;
200                 if (!(req->type & arch_register_req_type_limited))
201                         continue;
202
203                 in = get_irn_n(node, i);
204                 if (!arch_irn_consider_in_reg_alloc(cls, in))
205                         continue;
206                 if (!be_values_interfere(lv, node, in))
207                         continue;
208
209                 rbitset_copy(tmp, req->limited, cls->n_regs);
210                 rbitset_and(tmp, def_constr, cls->n_regs);
211
212                 if (rbitset_is_empty(tmp, cls->n_regs))
213                         continue;
214
215                 /*
216                  * only create the copy if the operand is no copy.
217                  * this is necessary since the assure constraints phase inserts
218                  * Copies and Keeps for operands which must be different from the
219                  * results. Additional copies here would destroy this.
220                  */
221                 if (be_is_Copy(in))
222                         continue;
223
224                 copy = be_new_Copy(cls, block, in);
225                 sched_add_before(node, copy);
226                 set_irn_n(node, i, copy);
227                 DBG((dbg, LEVEL_3, "inserting constr copy %+F for %+F pos %d\n",
228                      copy, node, i));
229                 be_liveness_update(lv, in);
230         }
231 }
232
233 static void pre_spill_prepare_constr_walker(ir_node *block, void *data)
234 {
235         be_pre_spill_env_t *env = data;
236         ir_node *node;
237         sched_foreach(block, node) {
238                 prepare_constr_insn(env, node);
239         }
240 }
241
242 void be_pre_spill_prepare_constr(be_irg_t *birg,
243                                  const arch_register_class_t *cls)
244 {
245         ir_graph *irg = birg->irg;
246         be_pre_spill_env_t env;
247         memset(&env, 0, sizeof(env));
248         env.birg = birg;
249         env.cls  = cls;
250
251         be_assure_liveness(birg);
252
253         irg_block_walk_graph(irg, pre_spill_prepare_constr_walker, NULL, &env);
254 }
255
256
257
258 int be_coalesce_spill_slots = 1;
259 int be_do_remats = 1;
260
261 static const lc_opt_table_entry_t be_spill_options[] = {
262         LC_OPT_ENT_BOOL ("coalesce_slots", "coalesce the spill slots", &be_coalesce_spill_slots),
263         LC_OPT_ENT_BOOL ("remat", "try to rematerialize values instead of reloading", &be_do_remats),
264         LC_OPT_LAST
265 };
266
267 static be_module_list_entry_t *spillers = NULL;
268 static const be_spiller_t *selected_spiller = NULL;
269
270 void be_register_spiller(const char *name, be_spiller_t *spiller)
271 {
272         if (selected_spiller == NULL)
273                 selected_spiller = spiller;
274         be_add_module_to_list(&spillers, name, spiller);
275 }
276
277 void be_do_spill(be_irg_t *birg, const arch_register_class_t *cls)
278 {
279         assert(selected_spiller != NULL);
280
281         selected_spiller->spill(birg, cls);
282 }
283
284 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spilloptions);
285 void be_init_spilloptions(void)
286 {
287         lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
288         lc_opt_entry_t *spill_grp = lc_opt_get_grp(be_grp, "spill");
289
290         lc_opt_add_table(spill_grp, be_spill_options);
291         be_add_module_list_opt(spill_grp, "spiller", "spill algorithm",
292                                &spillers, (void**) &selected_spiller);
293
294         FIRM_DBG_REGISTER(dbg, "firm.be.spillprepare");
295 }