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,
27 * The idea here is to push nodes as deep into the dominance tree as their
28 * dependencies allow. After pushing them back up out of as many loops as
35 #include "iroptimize.h"
38 #include "iredges_t.h"
42 static bool is_block_reachable(ir_node *block)
44 return get_Block_dom_depth(block) >= 0;
48 * Find the earliest correct block for node n. --- Place n into the
49 * same Block as its dominance-deepest Input.
51 * move_out_of_loops() expects that place_floats_early() have placed
52 * all "living" nodes into a living block. That's why we must
53 * move nodes in dead block with "live" successors into a valid
55 * We move them just into the same block as its successor (or
56 * in case of a Phi into the effective use block). For Phi successors,
57 * this may still be a dead block, but then there is no real use, as
58 * the control flow will be dead later.
60 static void place_floats_early(ir_node *n, waitq *worklist)
70 /* we must not run into an infinite loop */
71 if (irn_visited_else_mark(n))
74 /* The algorithm relies on the fact that all predecessors of a block are
75 * moved up after a call to place_float_early of the predecessors
76 * (see the loop below).
77 * However we have to break cycles somewhere. Relying on the visited flag
78 * will result in nodes not being moved up despite their place_floats_early
80 * Instead we break cycles at pinned nodes which won't move anyway:
81 * This works because in firm each cycle contains a Phi or Block node
84 if (get_irn_pinned(n) != op_pin_state_floats || only_used_by_keepalive(n)) {
85 /* we can't move pinned nodes */
86 arity = get_irn_arity(n);
87 for (i = 0; i < arity; ++i) {
88 ir_node *pred = get_irn_n(n, i);
89 pdeq_putr(worklist, pred);
92 pdeq_putr(worklist, get_nodes_block(n));
96 block = get_nodes_block(n);
98 /* first move predecessors up */
99 arity = get_irn_arity(n);
100 place_floats_early(block, worklist);
101 for (i = 0; i < arity; ++i) {
102 ir_node *pred = get_irn_n(n, i);
103 place_floats_early(pred, worklist);
106 /* determine earliest point */
109 for (i = 0; i < arity; ++i) {
110 ir_node *pred = get_irn_n(n, i);
111 ir_node *pred_block = get_nodes_block(pred);
112 int pred_depth = get_Block_dom_depth(pred_block);
113 if (pred_depth > new_depth) {
114 new_depth = pred_depth;
115 new_block = pred_block;
119 /* avoid moving nodes into the start block if we are not in the backend */
120 irg = get_irn_irg(n);
121 start_block = get_irg_start_block(irg);
122 if (new_block == start_block && block != start_block &&
123 get_irg_phase_state(irg) != phase_backend) {
124 assert(get_irn_n_edges_kind(start_block, EDGE_KIND_BLOCK) == 1);
125 const ir_edge_t *edge = get_block_succ_first(start_block);
126 new_block = get_edge_src_irn(edge);
129 /* Set the new block */
130 if (new_block != NULL)
131 set_nodes_block(n, new_block);
135 * Floating nodes form subgraphs that begin at nodes as Const, Load,
136 * Start, Call and that end at op_pin_state_pinned nodes as Store, Call.
137 * Place_early places all floating nodes reachable from its argument through
138 * floating nodes and adds all beginnings at op_pin_state_pinned nodes to the
141 * @param worklist a worklist, used for the algorithm, empty on in/output
143 static void place_early(ir_graph *irg, waitq *worklist)
146 inc_irg_visited(irg);
148 /* this inits the worklist */
149 place_floats_early(get_irg_end(irg), worklist);
151 /* Work the content of the worklist. */
152 while (!waitq_empty(worklist)) {
153 ir_node *n = (ir_node*)waitq_get(worklist);
155 place_floats_early(n, worklist);
157 set_irg_pinned(irg, op_pin_state_pinned);
161 * Compute the deepest common dominator tree ancestor of block and dca.
163 * @param dca the deepest common dominator tree ancestor so far,
165 * @param block a block
167 * @return the deepest common dominator tree ancestor of block and dca
169 static ir_node *calc_dom_dca(ir_node *dca, ir_node *block)
171 assert(block != NULL);
173 /* We found a first legal placement. */
177 /* Find a placement that is dominates both, dca and block. */
178 while (get_Block_dom_depth(block) > get_Block_dom_depth(dca))
179 block = get_Block_idom(block);
181 while (get_Block_dom_depth(dca) > get_Block_dom_depth(block)) {
182 dca = get_Block_idom(dca);
185 while (block != dca) {
186 block = get_Block_idom(block); dca = get_Block_idom(dca);
192 * Deepest common dominance ancestor of DCA and CONSUMER of PRODUCER.
193 * I.e., DCA is the block where we might place PRODUCER.
194 * A data flow edge points from producer to consumer.
196 static ir_node *consumer_dom_dca(ir_node *dca, ir_node *consumer,
199 /* Compute the last block into which we can place a node so that it is
201 if (is_Phi(consumer)) {
202 /* our consumer is a Phi-node, the effective use is in all those
203 blocks through which the Phi-node reaches producer */
204 ir_node *phi_block = get_nodes_block(consumer);
205 int arity = get_irn_arity(consumer);
208 for (i = 0; i < arity; i++) {
209 if (get_Phi_pred(consumer, i) == producer) {
210 ir_node *new_block = get_Block_cfgpred_block(phi_block, i);
211 if (is_Bad(new_block))
214 assert(is_block_reachable(new_block));
215 dca = calc_dom_dca(dca, new_block);
219 dca = calc_dom_dca(dca, get_nodes_block(consumer));
224 static inline int get_block_loop_depth(ir_node *block)
226 return get_loop_depth(get_irn_loop(block));
230 * Move n to a block with less loop depth than its current block. The
231 * new block must be dominated by early.
233 * @param n the node that should be moved
234 * @param early the earliest block we can n move to
236 static void move_out_of_loops(ir_node *n, ir_node *early)
238 ir_node *block = get_nodes_block(n);
239 ir_node *best = block;
240 int best_depth = get_block_loop_depth(best);
242 /* Find the region deepest in the dominator tree dominating
243 dca with the least loop nesting depth, but still dominated
244 by our early placement. */
245 while (block != early) {
246 ir_node *idom = get_Block_idom(block);
247 int idom_depth = get_block_loop_depth(idom);
248 if (idom_depth < best_depth) {
250 best_depth = idom_depth;
254 if (best != get_nodes_block(n))
255 set_nodes_block(n, best);
259 * Calculate the deepest common ancestor in the dominator tree of all nodes'
260 * blocks depending on node; our final placement has to dominate DCA.
262 * @param node the definition node
263 * @param dca the deepest common ancestor block so far, initially
266 * @return the deepest common dominator ancestor of all blocks of node's users
268 static ir_node *get_deepest_common_dom_ancestor(ir_node *node, ir_node *dca)
270 foreach_out_edge(node, edge) {
271 ir_node *succ = get_edge_src_irn(edge);
273 /* keepalive edges are special and don't respect the dominance */
278 /* Proj nodes are in the same block as node, so
279 * the users of Proj are our users. */
280 dca = get_deepest_common_dom_ancestor(succ, dca);
282 assert(is_block_reachable(get_nodes_block(succ)));
283 dca = consumer_dom_dca(dca, succ, node);
286 /* respect the keepalive rule: if our only user is a keepalive, then we must
287 * not move the node any further */
289 assert(only_used_by_keepalive(node));
290 return get_nodes_block(node);
293 foreach_out_edge_kind(node, edge, EDGE_KIND_DEP) {
294 ir_node *succ = get_edge_src_irn(edge);
295 assert(is_block_reachable(get_nodes_block(succ)));
296 dca = consumer_dom_dca(dca, succ, node);
302 * Put all the Proj nodes of a node into a given block.
304 * @param node the mode_T node
305 * @param block the block to put the Proj nodes to
307 static void set_projs_block(ir_node *node, ir_node *block)
309 foreach_out_edge(node, edge) {
310 ir_node *succ = get_edge_src_irn(edge);
315 set_nodes_block(succ, block);
316 if (get_irn_mode(succ) == mode_T) {
317 set_projs_block(succ, block);
323 * Find the latest legal block for N and place N into the
324 * `optimal' Block between the latest and earliest legal block.
325 * The `optimal' block is the dominance-deepest block of those
326 * with the least loop-nesting-depth. This places N out of as many
327 * loops as possible and then makes it as control dependent as
330 static void place_floats_late(ir_node *n, pdeq *worklist)
335 if (irn_visited_else_mark(n))
338 /* break cycles at pinned nodes (see place place_floats_early) as to why */
339 if (get_irn_pinned(n) != op_pin_state_floats) {
340 foreach_out_edge(n, edge) {
341 ir_node *succ = get_edge_src_irn(edge);
342 pdeq_putr(worklist, succ);
347 /* place our users */
348 foreach_out_edge(n, edge) {
349 ir_node *succ = get_edge_src_irn(edge);
350 place_floats_late(succ, worklist);
353 /* no point in moving Projs around, they are moved with their predecessor */
356 /* some nodes should simply stay in the startblock */
357 if (is_irn_start_block_placed(n)) {
358 assert(get_nodes_block(n) == get_irg_start_block(get_irn_irg(n)));
362 block = get_nodes_block(n);
363 assert(is_block_reachable(block));
365 /* deepest common ancestor in the dominator tree of all nodes'
366 blocks depending on us; our final placement has to dominate
368 dca = get_deepest_common_dom_ancestor(n, NULL);
370 set_nodes_block(n, dca);
371 move_out_of_loops(n, block);
372 if (get_irn_mode(n) == mode_T) {
373 set_projs_block(n, get_nodes_block(n));
379 * Place floating nodes on the given worklist as late as possible using
380 * the dominance tree.
382 * @param worklist the worklist containing the nodes to place
384 static void place_late(ir_graph *irg, waitq *worklist)
387 inc_irg_visited(irg);
389 /* This fills the worklist initially. */
390 place_floats_late(get_irg_start_block(irg), worklist);
392 /* And now empty the worklist again... */
393 while (!waitq_empty(worklist)) {
394 ir_node *n = (ir_node*)waitq_get(worklist);
396 place_floats_late(n, worklist);
400 /* Code Placement. */
401 void place_code(ir_graph *irg)
405 /* Handle graph state */
406 assert(get_irg_phase_state(irg) != phase_building);
407 assure_irg_properties(irg,
408 IR_GRAPH_PROPERTY_NO_CRITICAL_EDGES |
409 IR_GRAPH_PROPERTY_NO_UNREACHABLE_CODE |
410 IR_GRAPH_PROPERTY_CONSISTENT_OUT_EDGES |
411 IR_GRAPH_PROPERTY_CONSISTENT_DOMINANCE |
412 IR_GRAPH_PROPERTY_CONSISTENT_LOOPINFO);
414 /* Place all floating nodes as early as possible. This guarantees
415 a legal code placement. */
416 worklist = new_waitq();
417 place_early(irg, worklist);
419 /* While GCSE might place nodes in unreachable blocks,
420 * these are now placed in reachable blocks. */
422 /* Note: place_early changes only blocks, no data edges. So, the
423 * data out edges are still valid, no need to recalculate them here. */
425 /* Now move the nodes down in the dominator tree. This reduces the
426 unnecessary executions of the node. */
427 place_late(irg, worklist);
430 confirm_irg_properties(irg, IR_GRAPH_PROPERTIES_CONTROL_FLOW);
434 * Wrapper for place_code() inside the place_code pass.
436 static void place_code_wrapper(ir_graph *irg)
438 set_opt_global_cse(1);
439 optimize_graph_df(irg);
441 set_opt_global_cse(0);
444 ir_graph_pass_t *place_code_pass(const char *name)
446 return def_graph_pass(name ? name : "place", place_code_wrapper);