2 * This file is part of libFirm.
3 * Copyright (C) 2012 University of Karlsruhe.
8 * @brief Peephole optimisation framework keeps track of which registers contain which values
9 * @author Matthias Braun
14 #include "bepeephole.h"
16 #include "iredges_t.h"
27 #include "beintlive_t.h"
32 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
34 static const arch_env_t *arch_env;
36 static ir_node *current_node;
37 ir_node **register_values;
39 static void clear_reg_value(ir_node *node)
41 const arch_register_t *reg;
44 if (!mode_is_data(get_irn_mode(node)))
47 reg = arch_get_irn_register(node);
49 panic("No register assigned at %+F", node);
51 if (reg->type & arch_register_type_virtual)
53 reg_idx = reg->global_index;
55 DB((dbg, LEVEL_1, "Clear Register %s\n", reg->name));
56 register_values[reg_idx] = NULL;
59 static void set_reg_value(ir_node *node)
61 const arch_register_t *reg;
64 if (!mode_is_data(get_irn_mode(node)))
67 reg = arch_get_irn_register(node);
69 panic("No register assigned at %+F", node);
71 if (reg->type & arch_register_type_virtual)
73 reg_idx = reg->global_index;
75 DB((dbg, LEVEL_1, "Set Register %s: %+F\n", reg->name, node));
76 register_values[reg_idx] = node;
79 static void clear_defs(ir_node *node)
81 /* clear values defined */
82 be_foreach_value(node, value,
83 clear_reg_value(value);
87 static void set_uses(ir_node *node)
92 arity = get_irn_arity(node);
93 for (i = 0; i < arity; ++i) {
94 ir_node *in = get_irn_n(node, i);
100 * must be called from peephole optimisations before a node will be killed
101 * and its users will be redirected to new_node.
102 * so bepeephole can update its internal state.
104 * Note: killing a node and rewiring is only allowed if new_node produces
105 * the same registers as old_node.
107 static void be_peephole_before_exchange(const ir_node *old_node,
110 const arch_register_t *reg;
112 bool old_is_current = false;
114 DB((dbg, LEVEL_1, "About to exchange and kill %+F with %+F\n", old_node, new_node));
116 assert(sched_is_scheduled(skip_Proj_const(old_node)));
117 assert(sched_is_scheduled(skip_Proj(new_node)));
119 if (current_node == old_node) {
120 old_is_current = true;
122 /* next node to be processed will be killed. Its scheduling predecessor
123 * must be processed next. */
124 current_node = sched_next(current_node);
125 assert (!is_Bad(current_node));
127 /* we can't handle liveness updates correctly when exchange current node
128 * with something behind it */
129 assert(value_dominates(skip_Proj(new_node), skip_Proj_const(old_node)));
132 if (!mode_is_data(get_irn_mode(old_node)))
135 reg = arch_get_irn_register(old_node);
137 panic("No register assigned at %+F", old_node);
139 assert(reg == arch_get_irn_register(new_node) &&
140 "KILLING a node and replacing by different register is not allowed");
142 reg_idx = reg->global_index;
143 if (register_values[reg_idx] == old_node || old_is_current) {
144 register_values[reg_idx] = new_node;
147 be_liveness_remove(lv, old_node);
150 void be_peephole_exchange(ir_node *old, ir_node *nw)
152 be_peephole_before_exchange(old, nw);
155 be_liveness_introduce(lv, nw);
159 * block-walker: run peephole optimization on the given block.
161 static void process_block(ir_node *block, void *data)
165 /* construct initial register assignment */
166 memset(register_values, 0, sizeof(ir_node*) * arch_env->n_registers);
168 assert(lv->sets_valid && "live sets must be computed");
169 DB((dbg, LEVEL_1, "\nProcessing block %+F (from end)\n", block));
170 be_lv_foreach(lv, block, be_lv_state_end, node) {
173 DB((dbg, LEVEL_1, "\nstart processing\n"));
175 /* walk the block from last insn to the first */
176 current_node = sched_last(block);
177 for ( ; !sched_is_begin(current_node);
178 current_node = sched_prev(current_node)) {
180 peephole_opt_func peephole_node;
182 assert(!is_Bad(current_node));
183 if (is_Phi(current_node))
186 clear_defs(current_node);
187 set_uses(current_node);
189 op = get_irn_op(current_node);
190 peephole_node = (peephole_opt_func)op->ops.generic;
191 if (peephole_node == NULL)
194 DB((dbg, LEVEL_2, "optimize %+F\n", current_node));
195 peephole_node(current_node);
196 assert(!is_Bad(current_node));
201 * Check whether the node has only one user. Explicitly ignore the anchor.
203 bool be_has_only_one_user(ir_node *node)
205 int n = get_irn_n_edges(node);
212 foreach_out_edge(node, edge) {
213 ir_node *src = get_edge_src_irn(edge);
214 /* ignore anchor and keep-alive edges */
215 if (is_Anchor(src) || is_End(src))
223 static inline bool overlapping_regs(const arch_register_t *reg0,
224 const arch_register_req_t *req0, const arch_register_t *reg1,
225 const arch_register_req_t *req1)
227 if (reg0 == NULL || reg1 == NULL)
229 return reg0->global_index < (unsigned)reg1->global_index + req1->width
230 && reg1->global_index < (unsigned)reg0->global_index + req0->width;
233 bool be_can_move_down(ir_heights_t *heights, const ir_node *node,
234 const ir_node *before)
236 assert(get_nodes_block(node) == get_nodes_block(before));
237 assert(sched_get_time_step(node) < sched_get_time_step(before));
239 int node_arity = get_irn_arity(node);
240 ir_node *schedpoint = sched_next(node);
242 while (schedpoint != before) {
243 /* schedpoint must not use our computed values */
244 if (heights_reachable_in_block(heights, schedpoint, node))
247 /* schedpoint must not overwrite registers of our inputs */
248 for (int i = 0; i < node_arity; ++i) {
249 ir_node *in = get_irn_n(node, i);
250 const arch_register_t *reg = arch_get_irn_register(in);
253 const arch_register_req_t *in_req
254 = arch_get_irn_register_req_in(node, i);
255 be_foreach_out(schedpoint, o) {
256 const arch_register_t *outreg
257 = arch_get_irn_register_out(schedpoint, o);
258 const arch_register_req_t *outreq
259 = arch_get_irn_register_req_out(schedpoint, o);
260 if (overlapping_regs(reg, in_req, outreg, outreq))
265 schedpoint = sched_next(schedpoint);
270 bool be_can_move_up(ir_heights_t *heights, const ir_node *node,
271 const ir_node *after)
273 const ir_node *node_block = get_nodes_block(node);
274 const ir_node *after_block = get_block_const(after);
275 const ir_node *schedpoint;
276 if (node_block != after_block) {
277 /* currently we can move up exactly 1 block */
278 assert(get_Block_cfgpred_block(node_block, 0) == after_block);
279 ir_node *first = sched_first(node_block);
281 /* do not move nodes changing memory */
282 if (is_memop(node)) {
283 ir_node *meminput = get_memop_mem(node);
284 if (!is_NoMem(meminput))
288 /* make sure we can move to the beginning of the succ block */
289 if (node != first && !be_can_move_up(heights, node, sched_prev(first)))
292 /* check if node overrides any of live-in values of other successors */
293 ir_graph *irg = get_irn_irg(node);
294 be_lv_t *lv = be_get_irg_liveness(irg);
295 foreach_block_succ(after_block, edge) {
296 ir_node *succ = get_edge_src_irn(edge);
297 if (succ == node_block)
300 be_lv_foreach(lv, succ, be_lv_state_in, live_node) {
301 const arch_register_t *reg = arch_get_irn_register(live_node);
302 const arch_register_req_t *req = arch_get_irn_register_req(live_node);
303 be_foreach_out(node, o) {
304 const arch_register_t *outreg
305 = arch_get_irn_register_out(node, o);
306 const arch_register_req_t *outreq
307 = arch_get_irn_register_req_out(node, o);
308 if (overlapping_regs(outreg, outreq, reg, req))
312 sched_foreach(succ, phi) {
315 const arch_register_t *reg = arch_get_irn_register(phi);
316 const arch_register_req_t *req = arch_get_irn_register_req(phi);
317 be_foreach_out(node, o) {
318 const arch_register_t *outreg
319 = arch_get_irn_register_out(node, o);
320 const arch_register_req_t *outreq
321 = arch_get_irn_register_req_out(node, o);
322 if (overlapping_regs(outreg, outreq, reg, req))
327 schedpoint = sched_last(after_block);
329 schedpoint = sched_prev(node);
332 /* move schedule upwards until we hit the "after" node */
333 while (schedpoint != after) {
334 /* TODO: the following heights query only works for nodes in the same
335 * block, otherwise we have to be conservative here */
336 if (get_nodes_block(node) != get_nodes_block(schedpoint))
338 /* node must not depend on schedpoint */
339 if (heights_reachable_in_block(heights, node, schedpoint))
342 /* node must not overwrite registers used by schedpoint */
343 int arity = get_irn_arity(schedpoint);
344 for (int i = 0; i < arity; ++i) {
345 const arch_register_t *reg
346 = arch_get_irn_register_in(schedpoint, i);
349 const arch_register_req_t *in_req
350 = arch_get_irn_register_req_in(schedpoint, i);
351 be_foreach_out(node, o) {
352 const arch_register_t *outreg
353 = arch_get_irn_register_out(node, o);
354 const arch_register_req_t *outreq
355 = arch_get_irn_register_req_out(node, o);
356 if (overlapping_regs(outreg, outreq, reg, in_req))
361 schedpoint = sched_prev(schedpoint);
367 * Tries to optimize a beIncSP node with its previous IncSP node.
368 * Must be run from a be_peephole_opt() context.
370 ir_node *be_peephole_IncSP_IncSP(ir_node *node)
375 ir_node *pred = be_get_IncSP_pred(node);
377 if (!be_is_IncSP(pred))
380 if (!be_has_only_one_user(pred))
383 pred_offs = be_get_IncSP_offset(pred);
384 curr_offs = be_get_IncSP_offset(node);
385 offs = curr_offs + pred_offs;
387 /* add node offset to pred and remove our IncSP */
388 be_set_IncSP_offset(pred, offs);
390 be_peephole_exchange(node, pred);
394 void be_peephole_opt(ir_graph *irg)
396 be_assure_live_sets(irg);
398 arch_env = be_get_irg_arch_env(irg);
399 lv = be_get_irg_liveness(irg);
401 register_values = XMALLOCN(ir_node*, arch_env->n_registers);
403 irg_block_walk_graph(irg, process_block, NULL, NULL);
405 xfree(register_values);
408 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_peephole)
409 void be_init_peephole(void)
411 FIRM_DBG_REGISTER(dbg, "firm.be.peephole");