3 * File name: ir/opt/cfopt.c
4 * Purpose: control flow optimizations
8 * Copyright: (c) 1998-2004 Universität Karlsruhe
9 * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE.
20 #include "irgraph_t.h"
33 #include "irbackedge_t.h"
39 #include "iropt_dbg.h"
41 /*------------------------------------------------------------------*/
42 /* Control flow optimization. */
44 /* Removes Bad control flow predecessors and empty blocks. A block */
45 /* is empty if it contains only a Jmp node. */
46 /* Blocks can only be removed if they are not needed for the */
47 /* semantics of Phi nodes. */
48 /*------------------------------------------------------------------*/
51 * Replace binary Conds that jumps twice into the same block
57 * ProjX True ProjX False ==> | /
62 * Such pattern are the result of if-conversion.
64 * Note that the simple case that Block has only these two
65 * predecessors are already handled in equivalent_node_Block().
67 static void remove_senseless_conds(ir_node *bl, void *data)
70 int n = get_Block_n_cfgpreds(bl);
74 for (i = 0; i < n; ++i) {
75 ir_node *pred_i = get_Block_cfgpred(bl, i);
76 ir_node *cond_i = skip_Proj(pred_i);
78 for (j = i + 1; j < n; ++j) {
79 ir_node *pred_j = get_Block_cfgpred(bl, j);
80 ir_node *cond_j = skip_Proj(pred_j);
83 && get_irn_op(cond_i) == op_Cond
84 && get_irn_mode(get_Cond_selector(cond_i)) == mode_b) {
86 ir_node *jmp = new_r_Jmp(current_ir_graph, get_nodes_block(cond_i));
87 set_irn_n(bl, i, jmp);
88 set_irn_n(bl, j, new_Bad());
90 DBG_OPT_IFSIM2(cond_i, jmp);
99 * Removes Tuples from Block control flow predecessors.
100 * Optimizes blocks with equivalent_node(). This is tricky,
101 * as we want to avoid nodes that have as block predecessor Bads.
102 * Therefore we also optimize at control flow operations, depending
103 * how we first reach the Block.
105 static void merge_blocks(ir_node *n, void *env) {
109 /* clear the link field for ALL nodes first */
110 set_irn_link(n, NULL);
112 if (get_irn_op(n) == op_Block) {
114 for (i = 0; i < get_Block_n_cfgpreds(n); i++) {
115 /* GL @@@ : is this possible? if (get_opt_normalize()) -- added, all tests go through.
116 A different order of optimizations might cause problems. */
117 if (get_opt_normalize())
118 set_Block_cfgpred(n, i, skip_Tuple(get_Block_cfgpred(n, i)));
122 new_block = equivalent_node(n);
123 if (new_block != n && ! is_Block_dead(new_block))
124 exchange (n, new_block);
126 } else if (get_opt_optimize() && (get_irn_mode(n) == mode_X)) {
127 /* We will soon visit a block. Optimize it before visiting! */
128 ir_node *b = get_nodes_block(skip_Proj(n));
130 if (!is_Block_dead(b)) {
131 new_block = equivalent_node(b);
133 while (irn_not_visited(b) && (!is_Block_dead(new_block)) && (new_block != b)) {
134 /* We would have to run gigo() if new is bad, so we
135 promote it directly below. Nevertheless, we sometimes reach a block
136 the first time through a dataflow node. In this case we optimized the
137 block as such and have to promote the Bad here. */
138 assert((get_opt_control_flow_straightening() ||
139 get_opt_control_flow_weak_simplification()) &&
140 ("strange flag setting"));
141 exchange (b, new_block);
143 new_block = equivalent_node(b);
146 /* normally, we would create a Bad block here, but this must be
147 * prevented, so just set it's cf to Bad.
149 if (is_Block_dead(new_block))
150 exchange(n, new_Bad());
156 * Remove cf from dead block by inspecting dominance info
157 * Do not replace blocks by Bad. This optimization shall
158 * ensure, that all Bad control flow predecessors are
159 * removed, and no new other Bads are introduced.
161 * Must be run in the post walker.
163 static void remove_dead_block_cf(ir_node *block, void *env)
167 /* check block predecessors and turn control flow into bad */
168 for (i = 0, n = get_Block_n_cfgpreds(block); i < n; ++i) {
169 ir_node *pred_X = get_Block_cfgpred(block, i);
171 if (! is_Bad(pred_X)) {
172 ir_node *pred_bl = get_nodes_block(skip_Proj(pred_X));
174 if (is_Block_dead(pred_bl) || (get_Block_dom_depth(pred_bl) < 0))
175 exchange (pred_X, new_Bad());
181 * Collects all Phi nodes in link list of Block.
182 * Marks all blocks "block_visited" if they contain a node other
184 * Replaces n by Bad if n is unreachable control flow. We do that
185 * in the post walker, so we catch all blocks.
187 static void collect_nodes(ir_node *n, void *env) {
188 if (is_no_Block(n)) {
189 ir_node *b = get_nodes_block(n);
191 if ((get_irn_op(n) == op_Phi)) {
192 /* Collect Phi nodes to compact ins along with block's ins. */
193 set_irn_link(n, get_irn_link(b));
196 else if ((get_irn_op(n) != op_Jmp) && !is_Bad(b)) { /* Check for non empty block. */
197 mark_Block_block_visited(b);
202 /** Returns true if pred is predecessor of block. */
203 static int is_pred_of(ir_node *pred, ir_node *b) {
206 for (i = 0, n = get_Block_n_cfgpreds(b); i < n; ++i) {
207 ir_node *b_pred = get_nodes_block(get_Block_cfgpred(b, i));
208 if (b_pred == pred) return 1;
214 /** Test whether we can optimize away pred block pos of b.
216 * @param b A block node.
217 * @param pos The position of the predecessor block to judge about.
219 * @returns The number of predecessors
221 * The test is rather tricky.
223 * The situation is something like the following:
231 * b merges the control flow of an if-then-else. We may not remove
232 * the 'then' _and_ the 'else' block of an 'if' if there is a Phi
233 * node in b, even if both are empty. The destruction of this Phi
234 * requires that a copy is added before the merge. We have to
235 * keep one of the case blocks to place the copies in.
237 * To perform the test for pos, we must regard predecessors before pos
238 * as already removed.
240 static int test_whether_dispensable(ir_node *b, int pos) {
241 int i, j, n_preds = 1;
243 ir_node *cfop = get_Block_cfgpred(b, pos);
244 ir_node *pred = get_nodes_block(cfop);
246 /* Bad blocks will be optimized away, so we don't need space for them */
250 if (get_Block_block_visited(pred) + 1
251 < get_irg_block_visited(current_ir_graph)) {
253 if (!get_opt_optimize() || !get_opt_control_flow_strong_simplification()) {
254 /* Mark block so that is will not be removed: optimization is turned off. */
255 set_Block_block_visited(pred, get_irg_block_visited(current_ir_graph)-1);
259 /* Seems to be empty. At least we detected this in collect_nodes. */
260 if (!get_irn_link(b)) {
261 /* There are no Phi nodes ==> all predecessors are dispensable. */
262 n_preds = get_Block_n_cfgpreds(pred);
264 /* b's pred blocks and pred's pred blocks must be pairwise disjunct.
265 Work preds < pos as if they were already removed. */
266 for (i = 0; i < pos; i++) {
267 ir_node *b_pred = get_nodes_block(get_Block_cfgpred(b, i));
268 if (get_Block_block_visited(b_pred) + 1
269 < get_irg_block_visited(current_ir_graph)) {
270 for (j = 0; j < get_Block_n_cfgpreds(b_pred); j++) {
271 ir_node *b_pred_pred = get_nodes_block(get_Block_cfgpred(b_pred, j));
272 if (is_pred_of(b_pred_pred, pred)) dispensable = 0;
275 if (is_pred_of(b_pred, pred)) dispensable = 0;
278 for (i = pos +1; i < get_Block_n_cfgpreds(b); i++) {
279 ir_node *b_pred = get_nodes_block(get_Block_cfgpred(b, i));
280 if (is_pred_of(b_pred, pred)) dispensable = 0;
283 set_Block_block_visited(pred, get_irg_block_visited(current_ir_graph)-1);
286 n_preds = get_Block_n_cfgpreds(pred);
296 * This method removed Bad cf predecessors from Blocks and Phis, and removes
297 * empty blocks. A block is empty if it only contains Phi and Jmp nodes.
299 * We first adapt Phi nodes, then Block nodes, as we need the old ins
300 * of the Block to adapt the Phi nodes. We do this by computing new
301 * in arrays, and then replacing the old ones. So far we compute new in arrays
302 * for all nodes, not regarding whether there is a possibility for optimization.
304 * For each predecessor p of a Block b there are three cases:
305 * 1. The predecessor p is a Bad node: just skip it. The in array of b shrinks by one.
306 * 2. The predecessor p is empty. Remove p. All predecessors of p are now
308 * 3. The predecessor p is a block containing useful code. Just keep p as is.
310 * For Phi nodes f we have to check the conditions at the Block of f.
311 * For cases 1 and 3 we proceed as for Blocks. For case 2 we can have two
313 * 2a: The old predecessor of the Phi f is a Phi pred_f IN THE BLOCK REMOVED. In this
314 * case we proceed as for blocks. We remove pred_f. All
315 * predecessors of pred_f now are predecessors of f.
316 * 2b: The old predecessor of f is NOT in the block removed. It might be a Phi, too.
317 * We have to replicate f for each predecessor of the removed block. Or, with
318 * other words, the removed predecessor block has exactly one predecessor.
320 * Further there is a special case for self referencing blocks:
322 * then_b else_b then_b else_b
328 * | | | === optimized to ===> \ | | |
334 * If there is a Phi in pred_b, but we remove pred_b, we have to generate a
335 * Phi in loop_b, that has the ins of the Phi in pred_b and a self referencing
337 * @@@ It is negotiable whether we should do this ... there might end up a copy
338 * from the Phi in the loop when removing the Phis.
340 static void optimize_blocks(ir_node *b, void *env) {
341 int i, j, k, n, max_preds, n_preds, p_preds;
345 /* Count the number of predecessor if this block is merged with pred blocks
348 for (i = 0, k = get_Block_n_cfgpreds(b); i < k; ++i) {
349 max_preds += test_whether_dispensable(b, i);
351 in = xmalloc(max_preds * sizeof(*in));
354 printf(" working on "); DDMN(b);
355 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
356 pred = get_nodes_block(get_Block_cfgpred(b, i));
357 if (is_Bad(get_Block_cfgpred(b, i))) {
358 printf(" removing Bad %i\n ", i);
359 } else if (get_Block_block_visited(pred) +1
360 < get_irg_block_visited(current_ir_graph)) {
361 printf(" removing pred %i ", i); DDMN(pred);
362 } else { printf(" Nothing to do for "); DDMN(pred); }
364 * end Debug output -*/
366 /*- Fix the Phi nodes of the current block -*/
367 for (phi = get_irn_link(b); phi; ) {
368 assert(get_irn_op(phi) == op_Phi);
370 /* Find the new predecessors for the Phi */
372 for (i = 0, n = get_Block_n_cfgpreds(b); i < n; ++i) {
373 pred = get_nodes_block(get_Block_cfgpred(b, i));
375 if (is_Bad(get_Block_cfgpred(b, i))) {
376 /* case Phi 1: Do nothing */
378 else if (get_Block_block_visited(pred) + 1
379 < get_irg_block_visited(current_ir_graph)) {
380 /* case Phi 2: It's an empty block and not yet visited. */
381 ir_node *phi_pred = get_Phi_pred(phi, i);
383 for (j = 0, k = get_Block_n_cfgpreds(pred); j < k; j++) {
384 /* because of breaking loops, not all predecessors are Bad-clean,
385 * so we must check this here again */
386 if (! is_Bad(get_Block_cfgpred(pred, j))) {
387 if (get_nodes_block(phi_pred) == pred) {
389 assert(get_irn_op(phi_pred) == op_Phi); /* Block is empty!! */
391 in[p_preds++] = get_Phi_pred(phi_pred, j);
394 in[p_preds++] = phi_pred;
399 /* The Phi_pred node is replaced now if it is a Phi.
401 Somehow the removed Phi node can be used legally in loops.
402 Therefore we replace the old phi by the new one.
404 Further we have to remove the old Phi node by replacing it
405 by Bad. Else it will remain in the keep alive array of End
406 and cause illegal situations. So if there is no loop, we should
409 if (get_nodes_block(phi_pred) == pred) {
410 /* remove the Phi as it might be kept alive. Further there
411 might be other users. */
412 exchange(phi_pred, phi); /* geht, ist aber doch semantisch falsch! Warum?? */
416 in[p_preds++] = get_Phi_pred(phi, i);
419 assert(p_preds <= max_preds);
423 /* By removal of Bad ins the Phi might be degenerated. */
424 exchange(phi, in[0]);
426 set_irn_in(phi, p_preds, in);
428 phi = get_irn_link(phi);
431 /*- This happens only if merge between loop backedge and single loop entry.
432 See special case above. -*/
433 for (k = 0, n = get_Block_n_cfgpreds(b); k < n; ++k) {
434 pred = get_nodes_block(get_Block_cfgpred(b, k));
436 if (get_Block_block_visited(pred) + 1 < get_irg_block_visited(current_ir_graph)) {
437 /* we found a predecessor block at position k that will be removed */
438 for (phi = get_irn_link(pred); phi;) {
440 * the previous phase may already changed the phi, and even
441 * removed it at all, so check here if this node is still a phi
443 if (get_irn_op(phi) == op_Phi) {
446 /* move this phi from the predecessor into the block b */
447 set_nodes_block(phi, b);
449 /* first, copy all 0..k-1 predecessors */
450 for (i = 0; i < k; i++) {
451 pred = get_nodes_block(get_Block_cfgpred(b, i));
453 if (is_Bad(get_Block_cfgpred(b, i))) {
455 } else if (get_Block_block_visited(pred) + 1
456 < get_irg_block_visited(current_ir_graph)) {
457 /* It's an empty block and not yet visited. */
458 for (j = 0; j < get_Block_n_cfgpreds(pred); j++) {
459 /* @@@ Hier brauche ich Schleifeninformation!!! Kontrollflusskante
460 muss Rueckwaertskante sein! (An allen vier in[q_preds] = phi
461 Anweisungen.) Trotzdem tuts bisher!! */
462 if (! is_Bad(get_Block_cfgpred(pred, j)))
470 /* now we are at k, copy the phi predecessors */
471 pred = get_nodes_block(get_Block_cfgpred(b, k));
472 for (i = 0; i < get_Phi_n_preds(phi); i++) {
473 if (! is_Bad(get_Block_cfgpred(pred, i)))
474 in[q_preds++] = get_Phi_pred(phi, i);
477 /* and now all the rest */
478 for (i = k+1; i < get_Block_n_cfgpreds(b); i++) {
479 pred = get_nodes_block(get_Block_cfgpred(b, i));
481 if (is_Bad(get_Block_cfgpred(b, i))) {
483 } else if (get_Block_block_visited(pred) +1
484 < get_irg_block_visited(current_ir_graph)) {
485 /* It's an empty block and not yet visited. */
486 for (j = 0; j < get_Block_n_cfgpreds(pred); j++) {
487 if (! is_Bad(get_Block_cfgpred(pred, j)))
497 exchange(phi, in[0]);
499 set_irn_in(phi, q_preds, in);
501 assert(q_preds <= max_preds);
502 // assert(p_preds == q_preds && "Wrong Phi Fix");
504 phi = get_irn_link(phi);
509 /*- Fix the block -*/
511 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
512 pred = get_nodes_block(get_Block_cfgpred(b, i));
514 if (is_Bad(get_Block_cfgpred(b, i))) {
515 /* case 1: Do nothing */
516 } else if (get_Block_block_visited(pred) +1
517 < get_irg_block_visited(current_ir_graph)) {
518 /* case 2: It's an empty block and not yet visited. */
519 assert(get_Block_n_cfgpreds(b) > 1);
520 /* Else it should be optimized by equivalent_node. */
521 for (j = 0; j < get_Block_n_cfgpreds(pred); j++) {
522 ir_node *pred_block = get_Block_cfgpred(pred, j);
524 /* because of breaking loops, not all predecessors are Bad-clean,
525 * so we must check this here again */
526 if (! is_Bad(pred_block))
527 in[n_preds++] = pred_block;
529 /* Remove block as it might be kept alive. */
530 exchange(pred, b/*new_Bad()*/);
533 in[n_preds++] = get_Block_cfgpred(b, i);
536 assert(n_preds <= max_preds);
538 set_irn_in(b, n_preds, in);
540 assert(get_irn_link(b) == NULL || (n_preds == p_preds && "Wrong Phi Fix"));
546 /* Optimizations of the control flow that also require changes of Phi nodes.
548 * This optimization performs two passes over the graph.
550 * The first pass collects all Phi nodes in a link list in the block
551 * nodes. Further it performs simple control flow optimizations.
552 * Finally it marks all blocks that do not contain useful
553 * computations, i.e., these blocks might be removed.
555 * The second pass performs the optimizations intended by this algorithm.
556 * It walks only over block nodes and adapts these and the Phi nodes in these blocks,
557 * which it finds in a linked list computed by the first pass.
559 * We use the block_visited flag to mark empty blocks in the first
561 * @@@ It would be better to add a struct in the link field
562 * that keeps the Phi list and the mark. Place it on an obstack, as
563 * we will lose blocks and thereby generate memory leaks.
565 void optimize_cf(ir_graph *irg) {
568 ir_node *end = get_irg_end(irg);
569 ir_graph *rem = current_ir_graph;
570 irg_dom_state dom_state = get_irg_dom_state(current_ir_graph);
571 current_ir_graph = irg;
573 /* if the graph is not pinned, we cannot determine empty blocks */
574 assert(get_irg_pinned(irg) != op_pin_state_floats &&
575 "Control flow optimization need a pinned graph");
577 /* Handle graph state */
578 assert(get_irg_phase_state(irg) != phase_building);
579 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
580 set_irg_outs_inconsistent(current_ir_graph);
581 if (get_irg_dom_state(current_ir_graph) == dom_consistent)
582 set_irg_dom_inconsistent(current_ir_graph);
584 if (dom_state == dom_consistent && get_opt_optimize() && get_opt_unreachable_code()) {
585 ir_node *end = get_irg_end(irg);
587 /* we have dominance info, we can kill dead block */
588 irg_block_walk_graph(irg, NULL, remove_dead_block_cf, NULL);
590 /* fix the keep-alives */
591 for (i = 0, n = get_End_n_keepalives(end); i < n; ++i) {
592 ir_node *ka = get_End_keepalive(end, i);
595 if (get_Block_dom_depth(ka) == -1)
596 set_End_keepalive(end, i, new_Bad());
598 else if (get_Block_dom_depth(get_nodes_block(ka)) == -1)
599 set_End_keepalive(end, i, new_Bad());
603 irg_block_walk_graph(current_ir_graph, NULL, remove_senseless_conds, NULL);
605 /* Use block visited flag to mark non-empty blocks. */
606 inc_irg_block_visited(irg);
607 irg_walk(end, merge_blocks, collect_nodes, NULL);
609 /* Optimize the standard code. */
610 irg_block_walk(get_irg_end_block(irg), optimize_blocks, NULL, NULL);
612 /* Walk all keep alives, optimize them if block, add to new in-array
613 for end if useful. */
614 in = NEW_ARR_F (ir_node *, 1);
615 in[0] = get_nodes_block(end);
616 inc_irg_visited(current_ir_graph);
618 for (i = 0; i < get_End_n_keepalives(end); i++) {
619 ir_node *ka = get_End_keepalive(end, i);
621 if (irn_not_visited(ka)) {
622 if ((get_irn_op(ka) == op_Block) && Block_not_block_visited(ka)) {
623 set_irg_block_visited(current_ir_graph, /* Don't walk all the way to Start. */
624 get_irg_block_visited(current_ir_graph)-1);
625 irg_block_walk(ka, optimize_blocks, NULL, NULL);
626 mark_irn_visited(ka);
627 ARR_APP1 (ir_node *, in, ka);
628 } else if (get_irn_op(ka) == op_Phi) {
629 mark_irn_visited(ka);
630 ARR_APP1 (ir_node *, in, ka);
631 } else if (get_irn_op(ka) == op_IJmp) {
632 mark_irn_visited(ka);
633 ARR_APP1 (ir_node *, in, ka);
637 /* DEL_ARR_F(end->in); GL @@@ tut nicht ! */
641 /* the verifier doesn't work yet with floating nodes */
642 if (get_irg_pinned(irg) == op_pin_state_pinned) {
643 /* after optimize_cf(), only Bad data flow may remain. */
644 if (irg_vrfy_bads(irg, BAD_DF | BAD_BLOCK | TUPLE)) {
645 dump_ir_block_graph(irg, "-vrfy-cf");
646 dump_ir_graph(irg, "-vrfy-cf");
647 fprintf(stderr, "VRFY_BAD in optimize_cf()\n");
651 current_ir_graph = rem;