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"
42 #include "besched_t.h"
45 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
47 static const arch_env_t *arch_env;
49 static ir_node *current_node;
50 static ir_node *prev_node;
51 ir_node ***register_values;
53 static void clear_reg_value(ir_node *node)
55 const arch_register_t *reg;
56 const arch_register_class_t *cls;
60 if(!mode_is_data(get_irn_mode(node)))
63 reg = arch_get_irn_register(arch_env, node);
65 panic("No register assigned at %+F\n", node);
67 if(arch_register_type_is(reg, virtual))
69 cls = arch_register_get_class(reg);
70 reg_idx = arch_register_get_index(reg);
71 cls_idx = arch_register_class_index(cls);
73 //assert(register_values[cls_idx][reg_idx] != NULL);
74 DBG((dbg, LEVEL_1, "Clear Register %s\n", reg->name));
75 register_values[cls_idx][reg_idx] = NULL;
78 static void set_reg_value(ir_node *node)
80 const arch_register_t *reg;
81 const arch_register_class_t *cls;
85 if(!mode_is_data(get_irn_mode(node)))
88 reg = arch_get_irn_register(arch_env, node);
90 panic("No register assigned at %+F\n", node);
92 if(arch_register_type_is(reg, virtual))
94 cls = arch_register_get_class(reg);
95 reg_idx = arch_register_get_index(reg);
96 cls_idx = arch_register_class_index(cls);
98 DBG((dbg, LEVEL_1, "Set Register %s: %+F\n", reg->name, node));
99 register_values[cls_idx][reg_idx] = node;
102 static void clear_defs(ir_node *node)
104 /* clear values defined */
105 if(get_irn_mode(node) == mode_T) {
106 const ir_edge_t *edge;
107 foreach_out_edge(node, edge) {
108 ir_node *proj = get_edge_src_irn(edge);
109 clear_reg_value(proj);
112 clear_reg_value(node);
116 static void set_uses(ir_node *node)
120 /* set values used */
121 arity = get_irn_arity(node);
122 for(i = 0; i < arity; ++i) {
123 ir_node *in = get_irn_n(node, i);
128 void be_peephole_new_node(ir_node *const nw)
130 be_liveness_introduce(lv, nw);
134 * must be called from peephole optimisations before a node will be killed
135 * and its users will be redirected to new_node.
136 * so bepeephole can update it's internal state.
138 * Note: killing a node and rewiring os only allowed if new_node produces
139 * the same registers as old_node.
141 void be_peephole_before_exchange(const ir_node *old_node, ir_node *new_node)
143 const arch_register_t *reg;
144 const arch_register_class_t *cls;
148 DBG((dbg, LEVEL_1, "About to exchange and kill %+F with %+F\n", old_node, new_node));
150 if (current_node == old_node) {
151 /* next node to be processed will be killed. Its scheduling predecessor
152 * must be processed next. */
153 prev_node = sched_prev(current_node);
156 if (!mode_is_data(get_irn_mode(old_node)))
159 reg = arch_get_irn_register(arch_env, old_node);
161 panic("No register assigned at %+F\n", old_node);
163 assert(reg == arch_get_irn_register(arch_env, new_node) &&
164 "KILLING a node and replacing by different register is not allowed");
166 cls = arch_register_get_class(reg);
167 reg_idx = arch_register_get_index(reg);
168 cls_idx = arch_register_class_index(cls);
170 if (register_values[cls_idx][reg_idx] == old_node) {
171 register_values[cls_idx][reg_idx] = new_node;
174 be_liveness_remove(lv, old_node);
177 void be_peephole_exchange(ir_node *const old, ir_node *const nw)
179 be_peephole_before_exchange(old, nw);
182 be_peephole_new_node(nw);
186 * block-walker: run peephole optimization on the given block.
188 static void process_block(ir_node *block, void *data)
195 /* construct initial register assignment */
196 n_classes = arch_env_get_n_reg_class(arch_env);
197 for(i = 0; i < n_classes; ++i) {
198 const arch_register_class_t *cls = arch_env_get_reg_class(arch_env, i);
199 unsigned n_regs = arch_register_class_n_regs(cls);
200 memset(register_values[i], 0, sizeof(ir_node*) * n_regs);
203 assert(lv->nodes && "live sets must be computed");
204 DBG((dbg, LEVEL_1, "\nProcessing block %+F (from end)\n", block));
205 be_lv_foreach(lv, block, be_lv_state_end, l) {
206 ir_node *node = be_lv_get_irn(lv, block, l);
209 DBG((dbg, LEVEL_1, "\nstart processing\n"));
211 /* walk the block from last insn to the first */
212 current_node = sched_last(block);
213 for( ; !sched_is_begin(current_node);
214 current_node = prev_node != NULL ? prev_node : sched_prev(current_node)) {
217 peephole_opt_func peephole_node;
219 assert(!is_Bad(current_node));
221 if (is_Phi(current_node))
224 clear_defs(current_node);
225 set_uses(current_node);
227 op = get_irn_op(current_node);
228 peephole_node = (peephole_opt_func)op->ops.generic;
229 if (peephole_node == NULL)
233 peephole_node(current_node);
234 /* was the current node replaced? */
235 if (current_node != last)
236 set_uses(current_node);
241 * Walk through the block schedule and skip all barrier nodes.
243 static void skip_barrier(ir_node *ret_blk, ir_graph *irg) {
246 sched_foreach_reverse(ret_blk, irn) {
247 if (be_is_Barrier(irn)) {
248 const ir_edge_t *edge, *next;
250 foreach_out_edge_safe(irn, edge, next) {
251 ir_node *proj = get_edge_src_irn(edge);
252 int pn = (int)get_Proj_proj(proj);
253 ir_node *pred = get_irn_n(irn, pn);
255 edges_reroute_kind(proj, pred, EDGE_KIND_NORMAL, irg);
256 edges_reroute_kind(proj, pred, EDGE_KIND_DEP, irg);
266 * Kill the Barrier nodes for better peephole optimization.
268 static void kill_barriers(ir_graph *irg) {
269 ir_node *end_blk = get_irg_end_block(irg);
273 /* skip the barrier on all return blocks */
274 for (i = get_Block_n_cfgpreds(end_blk) - 1; i >= 0; --i) {
275 ir_node *be_ret = get_Block_cfgpred(end_blk, i);
276 ir_node *ret_blk = get_nodes_block(be_ret);
278 skip_barrier(ret_blk, irg);
281 /* skip the barrier on the start block */
282 start_blk = get_irg_start_block(irg);
283 skip_barrier(start_blk, irg);
287 * Tries to optimize a beIncSp node with it's previous IncSP node.
288 * Must be run from a be_peephole_opt() context.
290 ir_node *be_peephole_IncSP_IncSP(ir_node *node)
295 ir_node *pred = be_get_IncSP_pred(node);
297 if (!be_is_IncSP(pred))
300 if (get_irn_n_edges(pred) > 1)
303 pred_offs = be_get_IncSP_offset(pred);
304 curr_offs = be_get_IncSP_offset(node);
306 if (pred_offs == BE_STACK_FRAME_SIZE_EXPAND) {
307 if (curr_offs != BE_STACK_FRAME_SIZE_SHRINK) {
311 } else if (pred_offs == BE_STACK_FRAME_SIZE_SHRINK) {
312 if (curr_offs != BE_STACK_FRAME_SIZE_EXPAND) {
316 } else if (curr_offs == BE_STACK_FRAME_SIZE_EXPAND ||
317 curr_offs == BE_STACK_FRAME_SIZE_SHRINK) {
320 offs = curr_offs + pred_offs;
323 /* add node offset to pred and remove our IncSP */
324 be_set_IncSP_offset(pred, offs);
326 be_peephole_exchange(node, pred);
330 void be_peephole_opt(be_irg_t *birg)
332 ir_graph *irg = be_get_birg_irg(birg);
336 /* barrier nodes are used for register allocations. They hinders
337 * peephole optimizations, so remove them here. */
340 /* we sometimes find BadE nodes in float apps like optest_float.c or
341 * kahansum.c for example... */
342 be_liveness_invalidate(birg->lv);
343 be_liveness_assure_sets(be_assure_liveness(birg));
345 arch_env = be_get_birg_arch_env(birg);
346 lv = be_get_birg_liveness(birg);
348 n_classes = arch_env_get_n_reg_class(arch_env);
349 register_values = alloca(sizeof(register_values[0]) * n_classes);
350 for(i = 0; i < n_classes; ++i) {
351 const arch_register_class_t *cls = arch_env_get_reg_class(arch_env, i);
352 unsigned n_regs = arch_register_class_n_regs(cls);
353 register_values[i] = alloca(sizeof(ir_node*) * n_regs);
356 irg_block_walk_graph(irg, process_block, NULL, NULL);
359 void be_peephole_init(void)
361 clear_irp_opcodes_generic_func();
364 void be_init_peephole(void)
366 FIRM_DBG_REGISTER(dbg, "firm.be.peephole");
369 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillbelady);