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
28 #include "bepeephole.h"
30 #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(node);
63 panic("No register assigned at %+F", 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(node);
88 panic("No register assigned at %+F", 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_new_node(ir_node * nw)
128 be_liveness_introduce(lv, nw);
132 * must be called from peephole optimisations before a node will be killed
133 * and its users will be redirected to new_node.
134 * so bepeephole can update it's internal state.
136 * Note: killing a node and rewiring os only allowed if new_node produces
137 * the same registers as old_node.
139 void be_peephole_before_exchange(const ir_node *old_node, ir_node *new_node)
141 const arch_register_t *reg;
142 const arch_register_class_t *cls;
146 DBG((dbg, LEVEL_1, "About to exchange and kill %+F with %+F\n", old_node, new_node));
148 if (current_node == old_node) {
149 /* next node to be processed will be killed. Its scheduling predecessor
150 * must be processed next. */
151 current_node = sched_next(current_node);
152 assert (!is_Bad(current_node));
155 if (!mode_is_data(get_irn_mode(old_node)))
158 reg = arch_get_irn_register(old_node);
160 panic("No register assigned at %+F", old_node);
162 assert(reg == arch_get_irn_register(new_node) &&
163 "KILLING a node and replacing by different register is not allowed");
165 cls = arch_register_get_class(reg);
166 reg_idx = arch_register_get_index(reg);
167 cls_idx = arch_register_class_index(cls);
169 if (register_values[cls_idx][reg_idx] == old_node) {
170 register_values[cls_idx][reg_idx] = new_node;
173 be_liveness_remove(lv, old_node);
176 void be_peephole_exchange(ir_node *old, ir_node *nw)
178 be_peephole_before_exchange(old, nw);
181 be_peephole_new_node(nw);
185 * block-walker: run peephole optimization on the given block.
187 static void process_block(ir_node *block, void *data)
194 /* construct initial register assignment */
195 n_classes = arch_env_get_n_reg_class(arch_env);
196 for(i = 0; i < n_classes; ++i) {
197 const arch_register_class_t *cls = arch_env_get_reg_class(arch_env, i);
198 unsigned n_regs = arch_register_class_n_regs(cls);
199 memset(register_values[i], 0, sizeof(ir_node*) * n_regs);
202 assert(lv->nodes && "live sets must be computed");
203 DBG((dbg, LEVEL_1, "\nProcessing block %+F (from end)\n", block));
204 be_lv_foreach(lv, block, be_lv_state_end, l) {
205 ir_node *node = be_lv_get_irn(lv, block, l);
208 DBG((dbg, LEVEL_1, "\nstart processing\n"));
210 /* walk the block from last insn to the first */
211 current_node = sched_last(block);
212 for( ; !sched_is_begin(current_node);
213 current_node = sched_prev(current_node)) {
215 peephole_opt_func peephole_node;
217 assert(!is_Bad(current_node));
218 if (is_Phi(current_node))
221 clear_defs(current_node);
222 set_uses(current_node);
224 op = get_irn_op(current_node);
225 peephole_node = (peephole_opt_func)op->ops.generic;
226 if (peephole_node == NULL)
229 peephole_node(current_node);
230 assert(!is_Bad(current_node));
234 static void kill_node_and_preds(ir_node *node)
238 arity = get_irn_arity(node);
239 for (i = 0; i < arity; ++i) {
240 ir_node *pred = get_irn_n(node, i);
242 set_irn_n(node, i, new_Bad());
243 if (get_irn_n_edges(pred) != 0)
246 kill_node_and_preds(pred);
255 * Walk through the block schedule and skip all barrier nodes.
257 static void skip_barrier(ir_node *ret_blk, ir_graph *irg) {
260 sched_foreach_reverse(ret_blk, irn) {
261 const ir_edge_t *edge, *next;
263 if (!be_is_Barrier(irn))
266 foreach_out_edge_safe(irn, edge, next) {
267 ir_node *proj = get_edge_src_irn(edge);
268 int pn = (int)get_Proj_proj(proj);
269 ir_node *pred = get_irn_n(irn, pn);
271 edges_reroute_kind(proj, pred, EDGE_KIND_NORMAL, irg);
272 edges_reroute_kind(proj, pred, EDGE_KIND_DEP, irg);
275 kill_node_and_preds(irn);
281 * Kill the Barrier nodes for better peephole optimization.
283 static void kill_barriers(ir_graph *irg) {
284 ir_node *end_blk = get_irg_end_block(irg);
288 /* skip the barrier on all return blocks */
289 for (i = get_Block_n_cfgpreds(end_blk) - 1; i >= 0; --i) {
290 ir_node *be_ret = get_Block_cfgpred(end_blk, i);
291 ir_node *ret_blk = get_nodes_block(be_ret);
293 skip_barrier(ret_blk, irg);
296 /* skip the barrier on the start block */
297 start_blk = get_irg_start_block(irg);
298 skip_barrier(start_blk, irg);
302 * Check whether the node has only one user. Explicitly ignore the anchor.
304 static int has_only_one_user(ir_node *node)
306 int n = get_irn_n_edges(node);
307 const ir_edge_t *edge;
315 foreach_out_edge(node, edge) {
316 ir_node *src = get_edge_src_irn(edge);
325 * Tries to optimize a beIncSP node with its previous IncSP node.
326 * Must be run from a be_peephole_opt() context.
328 ir_node *be_peephole_IncSP_IncSP(ir_node *node)
333 ir_node *pred = be_get_IncSP_pred(node);
335 if (!be_is_IncSP(pred))
338 if (!has_only_one_user(pred))
341 pred_offs = be_get_IncSP_offset(pred);
342 curr_offs = be_get_IncSP_offset(node);
344 if (pred_offs == BE_STACK_FRAME_SIZE_EXPAND) {
345 if (curr_offs != BE_STACK_FRAME_SIZE_SHRINK) {
349 } else if (pred_offs == BE_STACK_FRAME_SIZE_SHRINK) {
350 if (curr_offs != BE_STACK_FRAME_SIZE_EXPAND) {
354 } else if (curr_offs == BE_STACK_FRAME_SIZE_EXPAND ||
355 curr_offs == BE_STACK_FRAME_SIZE_SHRINK) {
358 offs = curr_offs + pred_offs;
361 /* add node offset to pred and remove our IncSP */
362 be_set_IncSP_offset(pred, offs);
364 be_peephole_exchange(node, pred);
368 void be_peephole_opt(be_irg_t *birg)
370 ir_graph *irg = be_get_birg_irg(birg);
374 /* barrier nodes are used for register allocations. They hinders
375 * peephole optimizations, so remove them here. */
378 /* we sometimes find BadE nodes in float apps like optest_float.c or
379 * kahansum.c for example... */
380 be_liveness_invalidate(birg->lv);
381 be_liveness_assure_sets(be_assure_liveness(birg));
383 arch_env = be_get_birg_arch_env(birg);
384 lv = be_get_birg_liveness(birg);
386 n_classes = arch_env_get_n_reg_class(arch_env);
387 register_values = alloca(sizeof(register_values[0]) * n_classes);
388 for(i = 0; i < n_classes; ++i) {
389 const arch_register_class_t *cls = arch_env_get_reg_class(arch_env, i);
390 unsigned n_regs = arch_register_class_n_regs(cls);
391 register_values[i] = alloca(sizeof(ir_node*) * n_regs);
394 irg_block_walk_graph(irg, process_block, NULL, NULL);
397 void be_peephole_init(void)
399 clear_irp_opcodes_generic_func();
402 void be_init_peephole(void)
404 FIRM_DBG_REGISTER(dbg, "firm.be.peephole");
407 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_peephole);