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 Move nodes to a block where they will be executed the least
24 * @author Christian Schaefer, Goetz Lindenmaier, Sebastian Felis,
28 * The idea here is to push nodes as deep into the dominance tree as their
29 * dependencies allow. After pushing them back up out of as many loops as
36 #include "iroptimize.h"
42 #include "opt_manage.h"
44 static bool is_block_reachable(ir_node *block)
46 return get_Block_dom_depth(block) >= 0;
50 * Find the earliest correct block for node n. --- Place n into the
51 * same Block as its dominance-deepest Input.
53 * move_out_of_loops() expects that place_floats_early() have placed
54 * all "living" nodes into a living block. That's why we must
55 * move nodes in dead block with "live" successors into a valid
57 * We move them just into the same block as its successor (or
58 * in case of a Phi into the effective use block). For Phi successors,
59 * this may still be a dead block, but then there is no real use, as
60 * the control flow will be dead later.
62 static void place_floats_early(ir_node *n, waitq *worklist)
72 /* we must not run into an infinite loop */
73 if (irn_visited_else_mark(n))
76 /* The algorithm relies on the fact that all predecessors of a block are
77 * moved up after a call to place_float_early of the predecessors
78 * (see the loop below).
79 * However we have to break cycles somewhere. Relying on the visited flag
80 * will result in nodes not being moved up despite their place_floats_early
82 * Instead we break cycles at pinned nodes which won't move anyway:
83 * This works because in firm each cycle contains a Phi or Block node
86 if (get_irn_pinned(n) != op_pin_state_floats) {
87 /* we can't move pinned nodes */
88 arity = get_irn_arity(n);
89 for (i = 0; i < arity; ++i) {
90 ir_node *pred = get_irn_n(n, i);
91 pdeq_putr(worklist, pred);
94 pdeq_putr(worklist, get_nodes_block(n));
98 block = get_nodes_block(n);
100 /* first move predecessors up */
101 arity = get_irn_arity(n);
102 place_floats_early(block, worklist);
103 for (i = 0; i < arity; ++i) {
104 ir_node *pred = get_irn_n(n, i);
105 place_floats_early(pred, worklist);
108 /* determine earliest point */
111 for (i = 0; i < arity; ++i) {
112 ir_node *pred = get_irn_n(n, i);
113 ir_node *pred_block = get_nodes_block(pred);
114 int pred_depth = get_Block_dom_depth(pred_block);
115 if (pred_depth > new_depth) {
116 new_depth = pred_depth;
117 new_block = pred_block;
121 /* avoid moving nodes into the start block if we are not in the backend */
122 irg = get_irn_irg(n);
123 start_block = get_irg_start_block(irg);
124 if (new_block == start_block && block != start_block &&
125 get_irg_phase_state(irg) != phase_backend) {
126 assert(get_Block_n_cfg_outs(start_block) == 1);
127 new_block = get_Block_cfg_out(start_block, 0);
130 /* Set the new block */
131 if (new_block != NULL)
132 set_nodes_block(n, new_block);
136 * Floating nodes form subgraphs that begin at nodes as Const, Load,
137 * Start, Call and that end at op_pin_state_pinned nodes as Store, Call.
138 * Place_early places all floating nodes reachable from its argument through
139 * floating nodes and adds all beginnings at op_pin_state_pinned nodes to the
142 * @param worklist a worklist, used for the algorithm, empty on in/output
144 static void place_early(ir_graph *irg, waitq *worklist)
147 inc_irg_visited(irg);
149 /* this inits the worklist */
150 place_floats_early(get_irg_end(irg), worklist);
152 /* Work the content of the worklist. */
153 while (!waitq_empty(worklist)) {
154 ir_node *n = (ir_node*)waitq_get(worklist);
156 place_floats_early(n, worklist);
158 set_irg_pinned(irg, op_pin_state_pinned);
162 * Compute the deepest common dominator tree ancestor of block and dca.
164 * @param dca the deepest common dominator tree ancestor so far,
166 * @param block a block
168 * @return the deepest common dominator tree ancestor of block and dca
170 static ir_node *calc_dom_dca(ir_node *dca, ir_node *block)
172 assert(block != NULL);
174 /* We found a first legal placement. */
178 /* Find a placement that is dominates both, dca and block. */
179 while (get_Block_dom_depth(block) > get_Block_dom_depth(dca))
180 block = get_Block_idom(block);
182 while (get_Block_dom_depth(dca) > get_Block_dom_depth(block)) {
183 dca = get_Block_idom(dca);
186 while (block != dca) {
187 block = get_Block_idom(block); dca = get_Block_idom(dca);
193 * Deepest common dominance ancestor of DCA and CONSUMER of PRODUCER.
194 * I.e., DCA is the block where we might place PRODUCER.
195 * A data flow edge points from producer to consumer.
197 static ir_node *consumer_dom_dca(ir_node *dca, ir_node *consumer,
200 /* Compute the last block into which we can place a node so that it is
202 if (is_Phi(consumer)) {
203 /* our consumer is a Phi-node, the effective use is in all those
204 blocks through which the Phi-node reaches producer */
205 ir_node *phi_block = get_nodes_block(consumer);
206 int arity = get_irn_arity(consumer);
209 for (i = 0; i < arity; i++) {
210 if (get_Phi_pred(consumer, i) == producer) {
211 ir_node *new_block = get_Block_cfgpred_block(phi_block, i);
212 if (is_Bad(new_block))
215 assert(is_block_reachable(new_block));
216 dca = calc_dom_dca(dca, new_block);
220 dca = calc_dom_dca(dca, get_nodes_block(consumer));
225 static inline int get_block_loop_depth(ir_node *block)
227 return get_loop_depth(get_irn_loop(block));
231 * Move n to a block with less loop depth than its current block. The
232 * new block must be dominated by early.
234 * @param n the node that should be moved
235 * @param early the earliest block we can n move to
237 static void move_out_of_loops(ir_node *n, ir_node *early)
239 ir_node *block = get_nodes_block(n);
240 ir_node *best = block;
241 int best_depth = get_block_loop_depth(best);
243 /* Find the region deepest in the dominator tree dominating
244 dca with the least loop nesting depth, but still dominated
245 by our early placement. */
246 while (block != early) {
247 ir_node *idom = get_Block_idom(block);
248 int idom_depth = get_block_loop_depth(idom);
249 if (idom_depth < best_depth) {
251 best_depth = idom_depth;
255 if (best != get_nodes_block(n))
256 set_nodes_block(n, best);
260 * Calculate the deepest common ancestor in the dominator tree of all nodes'
261 * blocks depending on node; our final placement has to dominate DCA.
263 * @param node the definition node
264 * @param dca the deepest common ancestor block so far, initially
267 * @return the deepest common dominator ancestor of all blocks of node's users
269 static ir_node *get_deepest_common_dom_ancestor(ir_node *node, ir_node *dca)
273 for (i = get_irn_n_outs(node) - 1; i >= 0; --i) {
274 ir_node *succ = get_irn_out(node, i);
276 /* keepalive edges are special and don't respect the dominance */
281 /* Proj nodes are in the same block as node, so
282 * the users of Proj are our users. */
283 dca = get_deepest_common_dom_ancestor(succ, dca);
285 assert(is_block_reachable(get_nodes_block(succ)));
286 dca = consumer_dom_dca(dca, succ, node);
293 * Put all the Proj nodes of a node into a given block.
295 * @param node the mode_T node
296 * @param block the block to put the Proj nodes to
298 static void set_projs_block(ir_node *node, ir_node *block)
302 for (i = get_irn_n_outs(node) - 1; i >= 0; --i) {
303 ir_node *succ = get_irn_out(node, i);
305 assert(is_Proj(succ));
307 if (get_irn_mode(succ) == mode_T) {
308 set_projs_block(succ, block);
310 set_nodes_block(succ, block);
315 * Find the latest legal block for N and place N into the
316 * `optimal' Block between the latest and earliest legal block.
317 * The `optimal' block is the dominance-deepest block of those
318 * with the least loop-nesting-depth. This places N out of as many
319 * loops as possible and then makes it as control dependent as
322 static void place_floats_late(ir_node *n, pdeq *worklist)
329 if (irn_visited_else_mark(n))
332 n_outs = get_irn_n_outs(n);
333 /* break cycles at pinned nodes (see place place_floats_early) as to why */
334 if (get_irn_pinned(n) != op_pin_state_floats) {
335 for (i = 0; i < n_outs; ++i) {
336 ir_node *succ = get_irn_out(n, i);
337 pdeq_putr(worklist, succ);
342 /* place our users */
343 for (i = 0; i < n_outs; ++i) {
344 ir_node *succ = get_irn_out(n, i);
345 place_floats_late(succ, worklist);
348 /* no point in moving Projs around, they are moved with their predecessor */
351 /* some nodes should simply stay in the startblock */
352 if (is_irn_start_block_placed(n)) {
354 ir_graph *irg = get_irn_irg(n);
355 ir_node *start_block = get_irg_start_block(irg);
356 assert(get_nodes_block(n) == start_block);
361 block = get_nodes_block(n);
362 assert(is_block_reachable(block));
364 /* deepest common ancestor in the dominator tree of all nodes'
365 blocks depending on us; our final placement has to dominate
367 dca = get_deepest_common_dom_ancestor(n, NULL);
369 set_nodes_block(n, dca);
370 move_out_of_loops(n, block);
371 if (get_irn_mode(n) == mode_T) {
372 set_projs_block(n, get_nodes_block(n));
378 * Place floating nodes on the given worklist as late as possible using
379 * the dominance tree.
381 * @param worklist the worklist containing the nodes to place
383 static void place_late(ir_graph *irg, waitq *worklist)
386 inc_irg_visited(irg);
388 /* This fills the worklist initially. */
389 place_floats_late(get_irg_start_block(irg), worklist);
391 /* And now empty the worklist again... */
392 while (!waitq_empty(worklist)) {
393 ir_node *n = (ir_node*)waitq_get(worklist);
395 place_floats_late(n, worklist);
399 /* Code Placement. */
400 static ir_graph_state_t do_codeplacement(ir_graph *irg)
404 /* Handle graph state */
405 assert(get_irg_phase_state(irg) != phase_building);
407 /* Place all floating nodes as early as possible. This guarantees
408 a legal code placement. */
409 worklist = new_waitq();
410 place_early(irg, worklist);
412 /* While GCSE might place nodes in unreachable blocks,
413 * these are now placed in reachable blocks. */
415 /* Note: place_early changes only blocks, no data edges. So, the
416 * data out edges are still valid, no need to recalculate them here. */
418 /* Now move the nodes down in the dominator tree. This reduces the
419 unnecessary executions of the node. */
420 place_late(irg, worklist);
426 static optdesc_t opt_codeplacement = {
428 IR_GRAPH_STATE_NO_CRITICAL_EDGES |
429 IR_GRAPH_STATE_CONSISTENT_OUTS |
430 IR_GRAPH_STATE_CONSISTENT_DOMINANCE |
431 IR_GRAPH_STATE_CONSISTENT_LOOPINFO,
435 void place_code(ir_graph *irg)
437 perform_irg_optimization(irg, &opt_codeplacement);
441 * Wrapper for place_code() inside the place_code pass.
443 static void place_code_wrapper(ir_graph *irg)
445 set_opt_global_cse(1);
446 optimize_graph_df(irg);
448 set_opt_global_cse(0);
452 static ir_graph_state_t do_gcse(ir_graph *irg)
454 set_opt_global_cse(1);
455 optimize_graph_df(irg);
456 do_codeplacement(irg);
457 set_opt_global_cse(0);
461 static optdesc_t opt_gcse = {
463 IR_GRAPH_STATE_NO_CRITICAL_EDGES |
464 IR_GRAPH_STATE_CONSISTENT_OUTS |
465 IR_GRAPH_STATE_CONSISTENT_DOMINANCE |
466 IR_GRAPH_STATE_CONSISTENT_LOOPINFO,
471 ir_graph_pass_t *place_code_pass(const char *name)
473 return def_graph_pass(name ? name : "place", place_code_wrapper);