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 Peephole optimisation framework keeps track of which registers contain which values
23 * @author Matthias Braun
30 #include "bepeephole.h"
32 #include "iredges_t.h"
41 #include "besched_t.h"
44 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
46 static const arch_env_t *arch_env;
48 static ir_node *current_node;
49 ir_node ***register_values;
51 static void clear_reg_value(ir_node *node)
53 const arch_register_t *reg;
54 const arch_register_class_t *cls;
58 if(!mode_is_data(get_irn_mode(node)))
61 reg = arch_get_irn_register(arch_env, node);
63 panic("No register assigned at %+F\n", node);
65 if(arch_register_type_is(reg, virtual))
67 cls = arch_register_get_class(reg);
68 reg_idx = arch_register_get_index(reg);
69 cls_idx = arch_register_class_index(cls);
71 //assert(register_values[cls_idx][reg_idx] != NULL);
72 DBG((dbg, LEVEL_1, "Clear Register %s\n", reg->name));
73 register_values[cls_idx][reg_idx] = NULL;
76 static void set_reg_value(ir_node *node)
78 const arch_register_t *reg;
79 const arch_register_class_t *cls;
83 if(!mode_is_data(get_irn_mode(node)))
86 reg = arch_get_irn_register(arch_env, node);
88 panic("No register assigned at %+F\n", node);
90 if(arch_register_type_is(reg, virtual))
92 cls = arch_register_get_class(reg);
93 reg_idx = arch_register_get_index(reg);
94 cls_idx = arch_register_class_index(cls);
96 DBG((dbg, LEVEL_1, "Set Register %s: %+F\n", reg->name, node));
97 register_values[cls_idx][reg_idx] = node;
100 static void clear_defs(ir_node *node)
102 /* clear values defined */
103 if(get_irn_mode(node) == mode_T) {
104 const ir_edge_t *edge;
105 foreach_out_edge(node, edge) {
106 ir_node *proj = get_edge_src_irn(edge);
107 clear_reg_value(proj);
110 clear_reg_value(node);
114 static void set_uses(ir_node *node)
118 /* set values used */
119 arity = get_irn_arity(node);
120 for(i = 0; i < arity; ++i) {
121 ir_node *in = get_irn_n(node, i);
126 void be_peephole_before_exchange(const ir_node *old_node, ir_node *new_node)
128 const arch_register_t *reg;
129 const arch_register_class_t *cls;
133 DBG((dbg, LEVEL_1, "About to exchange %+F with %+F\n", old_node, new_node));
135 if(old_node == current_node) {
136 if(is_Proj(new_node)) {
137 current_node = get_Proj_pred(new_node);
139 current_node = new_node;
143 if(!mode_is_data(get_irn_mode(old_node)))
146 reg = arch_get_irn_register(arch_env, old_node);
148 panic("No register assigned at %+F\n", old_node);
150 cls = arch_register_get_class(reg);
151 reg_idx = arch_register_get_index(reg);
152 cls_idx = arch_register_class_index(cls);
154 if(register_values[cls_idx][reg_idx] == old_node) {
155 register_values[cls_idx][reg_idx] = new_node;
158 be_liveness_remove(lv, old_node);
161 void be_peephole_after_exchange(ir_node *new_node)
163 be_liveness_introduce(lv, new_node);
166 static void process_block(ir_node *block, void *data)
173 /* construct initial register assignment */
174 n_classes = arch_env_get_n_reg_class(arch_env);
175 for(i = 0; i < n_classes; ++i) {
176 const arch_register_class_t *cls = arch_env_get_reg_class(arch_env, i);
177 unsigned n_regs = arch_register_class_n_regs(cls);
178 memset(register_values[i], 0, sizeof(ir_node*) * n_regs);
181 assert(lv->nodes && "live sets must be computed");
182 DBG((dbg, LEVEL_1, "\nProcessing block %+F (from end)\n", block));
183 be_lv_foreach(lv, block, be_lv_state_end, l) {
184 ir_node *node = be_lv_get_irn(lv, block, l);
187 DBG((dbg, LEVEL_1, "\nstart processing\n"));
189 /* walk the block from last insn to the first */
190 current_node = sched_last(block);
191 for( ; !sched_is_begin(current_node);
192 current_node = sched_prev(current_node)) {
195 peephole_opt_func func;
197 if(is_Phi(current_node))
200 clear_defs(current_node);
201 set_uses(current_node);
203 op = get_irn_op(current_node);
204 func = (peephole_opt_func) op->ops.generic;
210 /* was the current node replaced? */
211 if(current_node != last) {
212 set_uses(current_node);
218 * Walk through the block schedule and skip all barrier nodes.
220 static void skip_barrier(ir_node *ret_blk) {
223 sched_foreach_reverse(ret_blk, irn) {
226 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
227 ir_node *proj = get_irn_n(irn, i);
230 ir_node *barrier = get_Proj_pred(proj);
232 if (be_is_Barrier(barrier)) {
233 int pn = (int)get_Proj_proj(proj);
234 ir_node *pred = get_irn_n(barrier, pn);
236 set_irn_n(irn, i, pred);
237 if (sched_is_scheduled(barrier))
238 sched_remove(barrier);
242 for (i = get_irn_deps(irn) - 1; i >= 0; --i) {
243 ir_node *proj = get_irn_dep(irn, i);
246 ir_node *barrier = get_Proj_pred(proj);
248 if (be_is_Barrier(barrier)) {
249 int pn = (int)get_Proj_proj(proj);
250 ir_node *pred = get_irn_n(barrier, pn);
252 set_irn_dep(irn, i, pred);
253 if (sched_is_scheduled(barrier))
254 sched_remove(barrier);
262 * Kill the Barrier nodes for better peephole optimization.
264 static void kill_barriers(ir_graph *irg) {
265 ir_node *end_blk = get_irg_end_block(irg);
269 /* skip the barrier on all return blocks */
270 for (i = get_Block_n_cfgpreds(end_blk) - 1; i >= 0; --i) {
271 ir_node *be_ret = get_Block_cfgpred(end_blk, i);
272 ir_node *ret_blk = get_nodes_block(be_ret);
274 skip_barrier(ret_blk);
277 /* skip the barrier on the start block */
278 start_blk = get_irg_start_block(irg);
279 skip_barrier(start_blk);
283 void be_peephole_opt(be_irg_t *birg)
285 ir_graph *irg = be_get_birg_irg(birg);
289 /* barrier nodes are used for register allocations. They hinders
290 * peephole optimizations, so remove them here. */
293 /* we sometimes find BadE nodes in float apps like optest_float.c or
294 * kahansum.c for example... */
295 be_liveness_invalidate(birg->lv);
296 be_liveness_assure_sets(be_assure_liveness(birg));
298 arch_env = be_get_birg_arch_env(birg);
299 lv = be_get_birg_liveness(birg);
301 n_classes = arch_env_get_n_reg_class(arch_env);
302 register_values = alloca(sizeof(register_values[0]) * n_classes);
303 for(i = 0; i < n_classes; ++i) {
304 const arch_register_class_t *cls = arch_env_get_reg_class(arch_env, i);
305 unsigned n_regs = arch_register_class_n_regs(cls);
306 register_values[i] = alloca(sizeof(ir_node*) * n_regs);
309 irg_block_walk_graph(irg, process_block, NULL, NULL);
312 void be_peephole_init(void)
314 clear_irp_opcodes_generic_func();
317 void be_init_peephole(void)
319 FIRM_DBG_REGISTER(dbg, "firm.be.peephole");
322 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillbelady);