2 * Copyright (C) 1995-2011 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 * @author Christian Helmer
23 * @brief loop inversion and loop unrolling
30 #include "iroptimize.h"
47 #include "opt_manage.h"
50 #include "irbackedge_t.h"
51 #include "irnodemap.h"
54 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
57 * Convenience macro for iterating over every phi node of the given block.
58 * Requires phi list per block.
60 #define for_each_phi(block, phi) \
61 for ((phi) = get_Block_phis( (block) ); (phi) ; (phi) = get_Phi_next((phi)))
63 #define for_each_phi_safe(head, phi, next) \
64 for ((phi) = (head), (next) = (head) ? get_Phi_next((head)) : NULL; \
65 (phi) ; (phi) = (next), (next) = (next) ? get_Phi_next((next)) : NULL)
67 /* Currently processed loop. */
68 static ir_loop *cur_loop;
70 /* Flag for kind of unrolling. */
74 } unrolling_kind_flag;
76 /* Condition for performing visiting a node during copy_walk. */
77 typedef bool walker_condition(const ir_node *);
79 /* Node and position of a predecessor. */
80 typedef struct entry_edge {
86 /* Node info for unrolling. */
87 typedef struct unrolling_node_info {
89 } unrolling_node_info;
91 /* Outs of the nodes head. */
92 static entry_edge *cur_head_outs;
94 /* Information about the loop head */
95 static ir_node *loop_head = NULL;
96 static bool loop_head_valid = true;
98 /* List of all inner loops, that are processed. */
99 static ir_loop **loops;
102 typedef struct loop_stats_t {
106 unsigned too_large_adapted;
107 unsigned cc_limit_reached;
108 unsigned calls_limit;
110 unsigned u_simple_counting_loop;
111 unsigned constant_unroll;
112 unsigned invariant_unroll;
117 static loop_stats_t stats;
119 /* Set stats to sero */
120 static void reset_stats(void)
122 memset(&stats, 0, sizeof(loop_stats_t));
126 static void print_stats(void)
128 DB((dbg, LEVEL_2, "---------------------------------------\n"));
129 DB((dbg, LEVEL_2, "loops : %d\n",stats.loops));
130 DB((dbg, LEVEL_2, "inverted : %d\n",stats.inverted));
131 DB((dbg, LEVEL_2, "too_large : %d\n",stats.too_large));
132 DB((dbg, LEVEL_2, "too_large_adapted : %d\n",stats.too_large_adapted));
133 DB((dbg, LEVEL_2, "cc_limit_reached : %d\n",stats.cc_limit_reached));
134 DB((dbg, LEVEL_2, "calls_limit : %d\n",stats.calls_limit));
135 DB((dbg, LEVEL_2, "u_simple_counting : %d\n",stats.u_simple_counting_loop));
136 DB((dbg, LEVEL_2, "constant_unroll : %d\n",stats.constant_unroll));
137 DB((dbg, LEVEL_2, "invariant_unroll : %d\n",stats.invariant_unroll));
138 DB((dbg, LEVEL_2, "=======================================\n"));
141 /* Commandline parameters */
142 typedef struct loop_opt_params_t {
143 unsigned max_loop_size; /* Maximum number of nodes [nodes]*/
144 int depth_adaption; /* Loop nest depth adaption [percent] */
145 unsigned allowed_calls; /* Number of calls allowed [number] */
146 bool count_phi; /* Count phi nodes */
147 bool count_proj; /* Count projections */
149 unsigned max_cc_size; /* Maximum condition chain size [nodes] */
150 unsigned max_branches;
152 unsigned max_unrolled_loop_size; /* [nodes] */
153 bool allow_const_unrolling;
154 bool allow_invar_unrolling;
155 unsigned invar_unrolling_min_size; /* [nodes] */
159 static loop_opt_params_t opt_params;
161 /* Loop analysis informations */
162 typedef struct loop_info_t {
163 unsigned nodes; /* node count */
164 unsigned ld_st; /* load and store nodes */
165 unsigned branches; /* number of conditions */
166 unsigned calls; /* number of calls */
167 unsigned cf_outs; /* number of cf edges which leave the loop */
168 entry_edge cf_out; /* single loop leaving cf edge */
169 int be_src_pos; /* position of the single own backedge in the head */
172 unsigned cc_size; /* nodes in the condition chain */
175 unsigned max_unroll; /* Number of unrolls satisfying max_loop_size */
176 unsigned exit_cond; /* 1 if condition==true exits the loop. */
177 unsigned latest_value:1; /* 1 if condition is checked against latest counter value */
178 unsigned needs_backedge:1; /* 0 if loop is completely unrolled */
179 unsigned decreasing:1; /* Step operation is_Sub, or step is<0 */
181 /* IV informations of a simple loop */
185 ir_node *iteration_phi;
188 ir_tarval *count_tar; /* Number of loop iterations */
190 ir_node *duff_cond; /* Duff mod */
191 unrolling_kind_flag unroll_kind; /* constant or invariant unrolling */
194 /* Information about the current loop */
195 static loop_info_t loop_info;
197 /* Outs of the condition chain (loop inversion). */
198 static ir_node **cc_blocks;
199 /* df/cf edges with def in the condition chain */
200 static entry_edge *cond_chain_entries;
201 /* Array of df loops found in the condition chain. */
202 static entry_edge *head_df_loop;
203 /* Number of blocks in cc */
204 static unsigned inversion_blocks_in_cc;
207 /* Cf/df edges leaving the loop.
208 * Called entries here, as they are used to enter the loop with walkers. */
209 static entry_edge *loop_entries;
210 /* Number of unrolls to perform */
211 static int unroll_nr;
212 /* Phase is used to keep copies of nodes. */
213 static ir_nodemap map;
214 static struct obstack obst;
216 /* Loop operations. */
217 typedef enum loop_op_t {
223 /* Saves which loop operation to do until after basic tests. */
224 static loop_op_t loop_op;
226 /* Returns the maximum nodes for the given nest depth */
227 static unsigned get_max_nodes_adapted(unsigned depth)
229 double perc = 100.0 + (double)opt_params.depth_adaption;
230 double factor = pow(perc / 100.0, depth);
232 return (int)((double)opt_params.max_loop_size * factor);
235 /* Reset nodes link. For use with a walker. */
236 static void reset_link(ir_node *node, void *env)
239 set_irn_link(node, NULL);
242 /* Returns 0 if the node or block is not in cur_loop. */
243 static bool is_in_loop(const ir_node *node)
245 return get_irn_loop(get_block_const(node)) == cur_loop;
248 /* Returns 0 if the given edge is not a backedge
249 * with its pred in the cur_loop. */
250 static bool is_own_backedge(const ir_node *n, int pos)
252 return is_backedge(n, pos) && is_in_loop(get_irn_n(n, pos));
255 /* Finds loop head and some loop_info as calls or else if necessary. */
256 static void get_loop_info(ir_node *node, void *env)
258 bool node_in_loop = is_in_loop(node);
262 /* collect some loop information */
264 if (is_Phi(node) && opt_params.count_phi)
266 else if (is_Proj(node) && opt_params.count_proj)
268 else if (!is_Confirm(node) && !is_Const(node) && !is_SymConst(node))
271 if (is_Load(node) || is_Store(node))
278 arity = get_irn_arity(node);
279 for (i = 0; i < arity; i++) {
280 ir_node *pred = get_irn_n(node, i);
281 bool pred_in_loop = is_in_loop(pred);
283 if (is_Block(node) && !node_in_loop && pred_in_loop) {
290 loop_info.cf_out = entry;
293 /* Find the loops head/the blocks with cfpred outside of the loop */
294 if (is_Block(node)) {
295 const ir_edge_t *edge;
298 /* Count innerloop branches */
299 foreach_out_edge_kind(node, edge, EDGE_KIND_BLOCK) {
300 ir_node *succ = get_edge_src_irn(edge);
301 if (is_Block(succ) && is_in_loop(succ))
305 ++loop_info.branches;
307 if (node_in_loop && !pred_in_loop && loop_head_valid) {
308 ir_node *cfgpred = get_Block_cfgpred(node, i);
310 if (!is_in_loop(cfgpred)) {
311 DB((dbg, LEVEL_5, "potential head %+F because inloop and pred %+F not inloop\n",
313 /* another head? We do not touch this. */
314 if (loop_head && loop_head != node) {
315 loop_head_valid = false;
325 /* Finds all edges with users outside of the loop
326 * and definition inside the loop. */
327 static void get_loop_entries(ir_node *node, void *env)
329 unsigned node_in_loop, pred_in_loop;
333 arity = get_irn_arity(node);
334 for (i = 0; i < arity; ++i) {
335 ir_node *pred = get_irn_n(node, i);
337 pred_in_loop = is_in_loop(pred);
338 node_in_loop = is_in_loop(node);
340 if (pred_in_loop && !node_in_loop) {
345 ARR_APP1(entry_edge, loop_entries, entry);
351 static ir_node *ssa_second_def;
352 static ir_node *ssa_second_def_block;
355 * Walks the graph bottom up, searching for definitions and creates phis.
357 static ir_node *search_def_and_create_phis(ir_node *block, ir_mode *mode, int first)
361 ir_graph *irg = get_irn_irg(block);
365 DB((dbg, LEVEL_5, "ssa search_def_and_create_phis: block %N\n", block));
367 /* Prevents creation of phi that would be bad anyway.
368 * Dead and bad blocks. */
369 if (get_irn_arity(block) < 1 || is_Bad(block)) {
370 DB((dbg, LEVEL_5, "ssa bad %N\n", block));
371 return new_r_Bad(irg, mode);
374 if (block == ssa_second_def_block && !first) {
375 DB((dbg, LEVEL_5, "ssa found second definition: use second def %N\n", ssa_second_def));
376 return ssa_second_def;
379 /* already processed this block? */
380 if (irn_visited(block)) {
381 ir_node *value = (ir_node *) get_irn_link(block);
382 DB((dbg, LEVEL_5, "ssa already visited: use linked %N\n", value));
386 assert(block != get_irg_start_block(irg));
388 /* a Block with only 1 predecessor needs no Phi */
389 n_cfgpreds = get_Block_n_cfgpreds(block);
390 if (n_cfgpreds == 1) {
391 ir_node *pred_block = get_Block_cfgpred_block(block, 0);
394 DB((dbg, LEVEL_5, "ssa 1 pred: walk pred %N\n", pred_block));
396 value = search_def_and_create_phis(pred_block, mode, 0);
397 set_irn_link(block, value);
398 mark_irn_visited(block);
403 /* create a new Phi */
404 NEW_ARR_A(ir_node*, in, n_cfgpreds);
405 for (i = 0; i < n_cfgpreds; ++i)
406 in[i] = new_r_Dummy(irg, mode);
408 phi = new_r_Phi(block, n_cfgpreds, in, mode);
409 /* Important: always keep block phi list up to date. */
410 add_Block_phi(block, phi);
411 DB((dbg, LEVEL_5, "ssa phi creation: link new phi %N to block %N\n", phi, block));
412 set_irn_link(block, phi);
413 mark_irn_visited(block);
415 /* set Phi predecessors */
416 for (i = 0; i < n_cfgpreds; ++i) {
418 ir_node *pred_block = get_Block_cfgpred_block(block, i);
419 assert(pred_block != NULL);
420 pred_val = search_def_and_create_phis(pred_block, mode, 0);
422 assert(pred_val != NULL);
424 DB((dbg, LEVEL_5, "ssa phi pred:phi %N, pred %N\n", phi, pred_val));
425 set_irn_n(phi, i, pred_val);
433 * Given a set of values this function constructs SSA-form for the users of the
434 * first value (the users are determined through the out-edges of the value).
435 * Works without using the dominance tree.
437 static void construct_ssa(ir_node *orig_block, ir_node *orig_val,
438 ir_node *second_block, ir_node *second_val)
442 const ir_edge_t *edge;
443 const ir_edge_t *next;
445 assert(orig_block && orig_val && second_block && second_val &&
446 "no parameter of construct_ssa may be NULL");
448 if (orig_val == second_val)
451 irg = get_irn_irg(orig_val);
453 ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED);
454 inc_irg_visited(irg);
456 mode = get_irn_mode(orig_val);
457 set_irn_link(orig_block, orig_val);
458 mark_irn_visited(orig_block);
460 ssa_second_def_block = second_block;
461 ssa_second_def = second_val;
463 /* Only fix the users of the first, i.e. the original node */
464 foreach_out_edge_safe(orig_val, edge, next) {
465 ir_node *user = get_edge_src_irn(edge);
466 int j = get_edge_src_pos(edge);
467 ir_node *user_block = get_nodes_block(user);
474 DB((dbg, LEVEL_5, "original user %N\n", user));
477 ir_node *pred_block = get_Block_cfgpred_block(user_block, j);
478 newval = search_def_and_create_phis(pred_block, mode, 1);
480 newval = search_def_and_create_phis(user_block, mode, 1);
482 if (newval != user && !is_Bad(newval))
483 set_irn_n(user, j, newval);
486 ir_free_resources(irg, IR_RESOURCE_IRN_VISITED);
490 /***** Unrolling Helper Functions *****/
492 /* Assign the copy with index nr to node n */
493 static void set_unroll_copy(ir_node *n, int nr, ir_node *cp)
495 unrolling_node_info *info;
496 assert(nr != 0 && "0 reserved");
498 info = (unrolling_node_info*)ir_nodemap_get(&map, n);
500 ir_node **arr = NEW_ARR_D(ir_node*, &obst, unroll_nr);
501 memset(arr, 0, unroll_nr * sizeof(ir_node*));
503 info = OALLOCZ(&obst, unrolling_node_info);
505 ir_nodemap_insert(&map, n, info);
510 info->copies[nr] = cp;
513 /* Returns a nodes copy if it exists, else NULL. */
514 static ir_node *get_unroll_copy(ir_node *n, int nr)
517 unrolling_node_info *info = (unrolling_node_info *)ir_nodemap_get(&map, n);
521 cp = info->copies[nr];
526 /***** Inversion Helper Functions *****/
528 /* Sets copy cp of node n. */
529 static void set_inversion_copy(ir_node *n, ir_node *cp)
531 ir_nodemap_insert(&map, n, cp);
534 /* Getter of copy of n for inversion */
535 static ir_node *get_inversion_copy(ir_node *n)
537 ir_node *cp = (ir_node *)ir_nodemap_get(&map, n);
541 /* Resets block mark for given node. For use with walker */
542 static void reset_block_mark(ir_node *node, void * env)
547 set_Block_mark(node, 0);
550 /* Returns mark of node, or its block if node is not a block.
551 * Used in this context to determine if node is in the condition chain. */
552 static bool is_nodes_block_marked(const ir_node* node)
554 return get_Block_mark(get_block_const(node));
557 /* Extends a nodes ins by node new.
558 * NOTE: This is slow if a node n needs to be extended more than once. */
559 static void extend_irn(ir_node *n, ir_node *newnode, bool new_is_backedge)
562 int arity = get_irn_arity(n);
563 int new_arity = arity + 1;
564 ir_node **ins = XMALLOCN(ir_node*, new_arity);
565 bool *bes = XMALLOCN(bool, new_arity);
568 /* Bes are important!
569 * Another way would be recreating the looptree,
570 * but after that we cannot distinguish already processed loops
571 * from not yet processed ones. */
573 for(i = 0; i < arity; ++i) {
574 bes[i] = is_backedge(n, i);
576 bes[i] = new_is_backedge;
579 for(i = 0; i < arity; ++i) {
580 ins[i] = get_irn_n(n, i);
584 set_irn_in(n, new_arity, ins);
588 for(i = 0; i < new_arity; ++i) {
597 /* Extends a block by a copy of its pred at pos,
598 * fixing also the phis in the same way. */
599 static void extend_ins_by_copy(ir_node *block, int pos)
604 assert(is_Block(block));
606 /* Extend block by copy of definition at pos */
607 pred = get_irn_n(block, pos);
608 new_in = get_inversion_copy(pred);
609 DB((dbg, LEVEL_5, "Extend block %N by %N cp of %N\n", block, new_in, pred));
610 extend_irn(block, new_in, false);
612 /* Extend block phis by copy of definition at pos */
613 for_each_phi(block, phi) {
616 pred = get_irn_n(phi, pos);
617 cp = get_inversion_copy(pred);
618 /* If the phis in is not in the condition chain (eg. a constant),
619 * there is no copy. */
625 DB((dbg, LEVEL_5, "Extend phi %N by %N cp of %N\n", phi, new_in, pred));
626 extend_irn(phi, new_in, false);
630 /* Returns the number of blocks backedges. With or without alien bes. */
631 static int get_backedge_n(ir_node *block, bool with_alien)
635 int arity = get_irn_arity(block);
637 assert(is_Block(block));
639 for (i = 0; i < arity; ++i) {
640 ir_node *pred = get_irn_n(block, i);
641 if (is_backedge(block, i) && (with_alien || is_in_loop(pred)))
647 /* Returns a raw copy of the given node.
648 * Attributes are kept/set according to the needs of loop inversion. */
649 static ir_node *copy_node(ir_node *node)
654 cp = exact_copy(node);
655 arity = get_irn_arity(node);
657 /* Keep backedge info */
658 for (i = 0; i < arity; ++i) {
659 if (is_backedge(node, i))
664 set_Block_mark(cp, 0);
672 * This walker copies all walked nodes.
673 * If the walk_condition is true for a node, it is copied.
674 * All nodes node_info->copy have to be NULL prior to every walk.
675 * Order of ins is important for later usage.
677 static void copy_walk(ir_node *node, walker_condition *walk_condition,
684 ir_graph *irg = current_ir_graph;
687 * break condition and cycle resolver, creating temporary node copies
689 if (get_irn_visited(node) >= get_irg_visited(irg)) {
690 /* Here we rely on nodestate's copy being initialized with NULL */
691 DB((dbg, LEVEL_5, "copy_walk: We have already visited %N\n", node));
692 if (get_inversion_copy(node) == NULL) {
693 cp = copy_node(node);
694 set_inversion_copy(node, cp);
696 DB((dbg, LEVEL_5, "The TEMP copy of %N is created %N\n", node, cp));
702 mark_irn_visited(node);
704 if (!is_Block(node)) {
705 ir_node *pred = get_nodes_block(node);
706 if (walk_condition(pred))
707 DB((dbg, LEVEL_5, "walk block %N\n", pred));
708 copy_walk(pred, walk_condition, set_loop);
711 arity = get_irn_arity(node);
713 NEW_ARR_A(ir_node *, cpin, arity);
715 for (i = 0; i < arity; ++i) {
716 ir_node *pred = get_irn_n(node, i);
718 if (walk_condition(pred)) {
719 DB((dbg, LEVEL_5, "walk node %N\n", pred));
720 copy_walk(pred, walk_condition, set_loop);
721 cpin[i] = get_inversion_copy(pred);
722 DB((dbg, LEVEL_5, "copy of %N gets new in %N which is copy of %N\n",
723 node, get_inversion_copy(pred), pred));
729 /* copy node / finalize temp node */
730 if (get_inversion_copy(node) == NULL) {
731 /* No temporary copy existent */
732 cp = copy_node(node);
733 set_inversion_copy(node, cp);
734 DB((dbg, LEVEL_5, "The FINAL copy of %N is CREATED %N\n", node, cp));
736 /* temporary copy is existent but without correct ins */
737 cp = get_inversion_copy(node);
738 DB((dbg, LEVEL_5, "The FINAL copy of %N is EXISTENT %N\n", node, cp));
741 if (!is_Block(node)) {
742 ir_node *cpblock = get_inversion_copy(get_nodes_block(node));
744 set_nodes_block(cp, cpblock );
746 add_Block_phi(cpblock, cp);
749 /* Keeps phi list of temporary node. */
750 set_irn_in(cp, ARR_LEN(cpin), cpin);
754 * This walker copies all walked nodes.
755 * If the walk_condition is true for a node, it is copied.
756 * All nodes node_info->copy have to be NULL prior to every walk.
757 * Order of ins is important for later usage.
758 * Takes copy_index, to phase-link copy at specific index.
760 static void copy_walk_n(ir_node *node, walker_condition *walk_condition,
769 * break condition and cycle resolver, creating temporary node copies
771 if (irn_visited(node)) {
772 /* Here we rely on nodestate's copy being initialized with NULL */
773 DB((dbg, LEVEL_5, "copy_walk: We have already visited %N\n", node));
774 if (get_unroll_copy(node, copy_index) == NULL) {
777 set_unroll_copy(node, copy_index, u);
778 DB((dbg, LEVEL_5, "The TEMP unknown of %N is created %N\n", node, u));
784 mark_irn_visited(node);
786 if (!is_Block(node)) {
787 ir_node *block = get_nodes_block(node);
788 if (walk_condition(block))
789 DB((dbg, LEVEL_5, "walk block %N\n", block));
790 copy_walk_n(block, walk_condition, copy_index);
793 arity = get_irn_arity(node);
794 NEW_ARR_A(ir_node *, cpin, arity);
796 for (i = 0; i < arity; ++i) {
797 ir_node *pred = get_irn_n(node, i);
799 if (walk_condition(pred)) {
800 DB((dbg, LEVEL_5, "walk node %N\n", pred));
801 copy_walk_n(pred, walk_condition, copy_index);
802 cpin[i] = get_unroll_copy(pred, copy_index);
808 /* copy node / finalize temp node */
809 cp = get_unroll_copy(node, copy_index);
810 if (cp == NULL || is_Unknown(cp)) {
811 cp = copy_node(node);
812 set_unroll_copy(node, copy_index, cp);
813 DB((dbg, LEVEL_5, "The FINAL copy of %N is CREATED %N\n", node, cp));
815 /* temporary copy is existent but without correct ins */
816 cp = get_unroll_copy(node, copy_index);
817 DB((dbg, LEVEL_5, "The FINAL copy of %N is EXISTENT %N\n", node, cp));
820 if (!is_Block(node)) {
821 ir_node *cpblock = get_unroll_copy(get_nodes_block(node), copy_index);
823 set_nodes_block(cp, cpblock );
825 add_Block_phi(cpblock, cp);
828 /* Keeps phi list of temporary node. */
829 set_irn_in(cp, ARR_LEN(cpin), cpin);
832 /* Removes alle Blocks with non marked predecessors from the condition chain. */
833 static void unmark_not_allowed_cc_blocks(void)
835 size_t blocks = ARR_LEN(cc_blocks);
838 for(i = 0; i < blocks; ++i) {
839 ir_node *block = cc_blocks[i];
841 int arity = get_irn_arity(block);
843 /* Head is an exception. */
844 if (block == loop_head)
847 for(a = 0; a < arity; ++a) {
848 if (! is_nodes_block_marked(get_irn_n(block, a))) {
849 set_Block_mark(block, 0);
850 --inversion_blocks_in_cc;
851 DB((dbg, LEVEL_5, "Removed %N from cc (blocks in cc %d)\n",
852 block, inversion_blocks_in_cc));
860 /* Unmarks all cc blocks using cc_blocks except head.
861 * TODO: invert head for unrolling? */
862 static void unmark_cc_blocks(void)
864 size_t blocks = ARR_LEN(cc_blocks);
867 for(i = 0; i < blocks; ++i) {
868 ir_node *block = cc_blocks[i];
870 /* TODO Head is an exception. */
871 /*if (block != loop_head)*/
872 set_Block_mark(block, 0);
874 /*inversion_blocks_in_cc = 1;*/
875 inversion_blocks_in_cc = 0;
878 loop_info.cc_size = 0;
882 * Populates head_entries with (node, pred_pos) tuple
883 * whereas the node's pred at pred_pos is in the cc but not the node itself.
884 * Also finds df loops inside the cc.
885 * Head and condition chain blocks have been marked previously.
887 static void get_head_outs(ir_node *node, void *env)
890 int arity = get_irn_arity(node);
893 for (i = 0; i < arity; ++i) {
894 if (!is_nodes_block_marked(node) && is_nodes_block_marked(get_irn_n(node, i))) {
898 /* Saving also predecessor seems redundant, but becomes
899 * necessary when changing position of it, before
900 * dereferencing it.*/
901 entry.pred = get_irn_n(node, i);
902 ARR_APP1(entry_edge, cur_head_outs, entry);
906 arity = get_irn_arity(loop_head);
908 /* Find df loops inside the cc */
909 if (is_Phi(node) && get_nodes_block(node) == loop_head) {
910 for (i = 0; i < arity; ++i) {
911 if (is_own_backedge(loop_head, i)) {
912 if (is_nodes_block_marked(get_irn_n(node, i))) {
916 entry.pred = get_irn_n(node, i);
917 ARR_APP1(entry_edge, head_df_loop, entry);
918 DB((dbg, LEVEL_5, "Found incc assignment node %N @%d is pred %N, graph %N %N\n",
919 node, i, entry.pred, current_ir_graph, get_irg_start_block(current_ir_graph)));
927 * Find condition chains, and add them to be inverted.
928 * A block belongs to the chain if a condition branches out of the loop.
929 * (Some blocks need to be removed once again.)
930 * Returns 1 if the given block belongs to the condition chain.
932 static void find_condition_chain(ir_node *block)
934 const ir_edge_t *edge;
937 bool jmp_only = true;
938 unsigned nodes_n = 0;
940 mark_irn_visited(block);
942 DB((dbg, LEVEL_5, "condition_chains for block %N\n", block));
945 foreach_out_edge_kind(block, edge, EDGE_KIND_NORMAL) {
949 /* Check if node count would exceed maximum cc size.
951 * This is not optimal, as we search depth-first and break here,
952 * continuing with another subtree. */
953 if (loop_info.cc_size + nodes_n > opt_params.max_cc_size) {
954 set_Block_mark(block, 0);
958 /* Check if block only has a jmp instruction. */
959 foreach_out_edge(block, edge) {
960 ir_node *src = get_edge_src_irn(edge);
962 if (!is_Block(src) && !is_Jmp(src)) {
967 /* Check cf outs if one is leaving the loop,
968 * or if this node has a backedge. */
969 foreach_block_succ(block, edge) {
970 ir_node *src = get_edge_src_irn(edge);
971 int pos = get_edge_src_pos(edge);
973 if (!is_in_loop(src))
976 /* Inverting blocks with backedge outs leads to a cf edge
977 * from the inverted head, into the inverted head (skipping the body).
978 * As the body becomes the new loop head,
979 * this would introduce another loop in the existing loop.
980 * This loop inversion cannot cope with this case. */
981 if (is_backedge(src, pos)) {
987 /* We need all predecessors to already belong to the condition chain.
988 * Example of wrong case: * == in cc
999 /* Collect blocks containing only a Jmp.
1000 * Do not collect blocks with backedge outs. */
1001 if ((jmp_only || mark) && !has_be) {
1002 set_Block_mark(block, 1);
1003 ++inversion_blocks_in_cc;
1004 loop_info.cc_size += nodes_n;
1005 DB((dbg, LEVEL_5, "block %N is part of condition chain\n", block));
1006 ARR_APP1(ir_node *, cc_blocks, block);
1008 set_Block_mark(block, 0);
1011 foreach_block_succ(block, edge) {
1012 ir_node *src = get_edge_src_irn( edge );
1014 if (is_in_loop(src) && ! irn_visited(src))
1015 find_condition_chain(src);
1020 * Rewires the copied condition chain. Removes backedges
1021 * as this condition chain is prior to the loop.
1022 * Copy of loop_head must have phi list and old (unfixed) backedge info of the loop head.
1023 * (loop_head is already fixed, we cannot rely on it.)
1025 static void fix_copy_inversion(void)
1030 ir_node *phi, *next;
1031 ir_node *head_cp = get_inversion_copy(loop_head);
1032 ir_graph *irg = get_irn_irg(head_cp);
1033 int arity = get_irn_arity(head_cp);
1034 int backedges = get_backedge_n(head_cp, false);
1035 int new_arity = arity - backedges;
1039 NEW_ARR_A(ir_node *, ins, new_arity);
1042 /* Remove block backedges */
1043 for(i = 0; i < arity; ++i) {
1044 if (!is_backedge(head_cp, i))
1045 ins[pos++] = get_irn_n(head_cp, i);
1048 new_head = new_r_Block(irg, new_arity, ins);
1050 phis = NEW_ARR_F(ir_node *, 0);
1052 for_each_phi_safe(get_Block_phis(head_cp), phi, next) {
1054 NEW_ARR_A(ir_node *, ins, new_arity);
1056 for(i = 0; i < arity; ++i) {
1057 if (!is_backedge(head_cp, i))
1058 ins[pos++] = get_irn_n(phi, i);
1060 new_phi = new_rd_Phi(get_irn_dbg_info(phi),
1061 new_head, new_arity, ins,
1063 ARR_APP1(ir_node *, phis, new_phi);
1067 for_each_phi_safe(get_Block_phis(head_cp), phi, next) {
1068 exchange(phi, phis[pos++]);
1071 exchange(head_cp, new_head);
1077 /* Puts the original condition chain at the end of the loop,
1078 * subsequently to the body.
1079 * Relies on block phi list and correct backedges.
1081 static void fix_head_inversion(void)
1085 ir_node *phi, *next;
1087 ir_graph *irg = get_irn_irg(loop_head);
1088 int arity = get_irn_arity(loop_head);
1089 int backedges = get_backedge_n(loop_head, false);
1090 int new_arity = backedges;
1094 NEW_ARR_A(ir_node *, ins, new_arity);
1097 /* Keep only backedges */
1098 for(i = 0; i < arity; ++i) {
1099 if (is_own_backedge(loop_head, i))
1100 ins[pos++] = get_irn_n(loop_head, i);
1103 new_head = new_r_Block(irg, new_arity, ins);
1105 phis = NEW_ARR_F(ir_node *, 0);
1107 for_each_phi(loop_head, phi) {
1109 DB((dbg, LEVEL_5, "Fixing phi %N of loop head\n", phi));
1111 NEW_ARR_A(ir_node *, ins, new_arity);
1114 for (i = 0; i < arity; ++i) {
1115 ir_node *pred = get_irn_n(phi, i);
1117 if (is_own_backedge(loop_head, i)) {
1118 /* If assignment is in the condition chain,
1119 * we need to create a phi in the new loop head.
1120 * This can only happen for df, not cf. See find_condition_chains. */
1121 /*if (is_nodes_block_marked(pred)) {
1129 new_phi = new_rd_Phi(get_irn_dbg_info(phi),
1130 new_head, new_arity, ins,
1133 ARR_APP1(ir_node *, phis, new_phi);
1135 DB((dbg, LEVEL_5, "fix inverted head should exch %N by %N (pos %d)\n", phi, new_phi, pos ));
1139 for_each_phi_safe(get_Block_phis(loop_head), phi, next) {
1140 DB((dbg, LEVEL_5, "fix inverted exch phi %N by %N\n", phi, phis[pos]));
1141 if (phis[pos] != phi)
1142 exchange(phi, phis[pos++]);
1147 DB((dbg, LEVEL_5, "fix inverted head exch head block %N by %N\n", loop_head, new_head));
1148 exchange(loop_head, new_head);
1151 /* Does the loop inversion. */
1152 static void inversion_walk(ir_graph *irg, entry_edge *head_entries)
1157 * The order of rewiring bottom-up is crucial.
1158 * Any change of the order leads to lost information that would be needed later.
1161 ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED);
1163 /* 1. clone condition chain */
1164 inc_irg_visited(irg);
1166 for (i = 0; i < ARR_LEN(head_entries); ++i) {
1167 entry_edge entry = head_entries[i];
1168 ir_node *pred = get_irn_n(entry.node, entry.pos);
1170 DB((dbg, LEVEL_5, "\nInit walk block %N\n", pred));
1172 copy_walk(pred, is_nodes_block_marked, cur_loop);
1175 ir_free_resources(irg, IR_RESOURCE_IRN_VISITED);
1177 /* 2. Extends the head control flow successors ins
1178 * with the definitions of the copied head node. */
1179 for (i = 0; i < ARR_LEN(head_entries); ++i) {
1180 entry_edge head_out = head_entries[i];
1182 if (is_Block(head_out.node))
1183 extend_ins_by_copy(head_out.node, head_out.pos);
1186 /* 3. construct_ssa for users of definitions in the condition chain,
1187 * as there is now a second definition. */
1188 for (i = 0; i < ARR_LEN(head_entries); ++i) {
1189 entry_edge head_out = head_entries[i];
1191 /* Ignore keepalives */
1192 if (is_End(head_out.node))
1195 /* Construct ssa for assignments in the condition chain. */
1196 if (!is_Block(head_out.node)) {
1197 ir_node *pred, *cppred, *block, *cpblock;
1199 pred = head_out.pred;
1200 cppred = get_inversion_copy(pred);
1201 block = get_nodes_block(pred);
1202 cpblock = get_nodes_block(cppred);
1203 construct_ssa(block, pred, cpblock, cppred);
1208 * If there is an assignment in the condition chain
1209 * with a user also in the condition chain,
1210 * the dominance frontier is in the new loop head.
1211 * The dataflow loop is completely in the condition chain.
1216 * Phi_cp | | copied condition chain
1220 * Phi* | | new loop head with newly created phi.
1222 * Phi | | original, inverted condition chain
1227 for (i = 0; i < ARR_LEN(head_df_loop); ++i) {
1228 entry_edge head_out = head_df_loop[i];
1230 /* Construct ssa for assignments in the condition chain. */
1231 ir_node *pred, *cppred, *block, *cpblock;
1233 pred = head_out.pred;
1234 cppred = get_inversion_copy(pred);
1235 assert(cppred && pred);
1236 block = get_nodes_block(pred);
1237 cpblock = get_nodes_block(cppred);
1238 construct_ssa(block, pred, cpblock, cppred);
1241 /* 4. Remove the ins which are no backedges from the original condition chain
1242 * as the cc is now subsequent to the body. */
1243 fix_head_inversion();
1245 /* 5. Remove the backedges of the copied condition chain,
1246 * because it is going to be the new 'head' in advance to the loop. */
1247 fix_copy_inversion();
1251 /* Performs loop inversion of cur_loop if possible and reasonable. */
1252 static void loop_inversion(ir_graph *irg)
1255 unsigned max_loop_nodes = opt_params.max_loop_size;
1256 unsigned max_loop_nodes_adapted;
1257 int depth_adaption = opt_params.depth_adaption;
1259 bool do_inversion = true;
1261 /* Depth of 0 is the procedure and 1 a topmost loop. */
1262 loop_depth = get_loop_depth(cur_loop) - 1;
1264 /* Calculating in per mil. */
1265 max_loop_nodes_adapted = get_max_nodes_adapted(loop_depth);
1267 DB((dbg, LEVEL_1, "max_nodes: %d\nmax_nodes_adapted %d at depth of %d (adaption %d)\n",
1268 max_loop_nodes, max_loop_nodes_adapted, loop_depth, depth_adaption));
1270 if (loop_info.nodes == 0)
1273 if (loop_info.nodes > max_loop_nodes) {
1274 /* Only for stats */
1275 DB((dbg, LEVEL_1, "Nodes %d > allowed nodes %d\n",
1276 loop_info.nodes, loop_depth, max_loop_nodes));
1279 /* Adaption might change it */
1282 /* Limit processing to loops smaller than given parameter. */
1283 if (loop_info.nodes > max_loop_nodes_adapted) {
1284 DB((dbg, LEVEL_1, "Nodes %d > allowed nodes (depth %d adapted) %d\n",
1285 loop_info.nodes, loop_depth, max_loop_nodes_adapted));
1286 ++stats.too_large_adapted;
1290 if (loop_info.calls > opt_params.allowed_calls) {
1291 DB((dbg, LEVEL_1, "Calls %d > allowed calls %d\n",
1292 loop_info.calls, opt_params.allowed_calls));
1293 ++stats.calls_limit;
1297 /*inversion_head_node_limit = INT_MAX;*/
1298 ir_reserve_resources(irg, IR_RESOURCE_BLOCK_MARK);
1300 /* Reset block marks.
1301 * We use block marks to flag blocks of the original condition chain. */
1302 irg_walk_graph(irg, reset_block_mark, NULL, NULL);
1304 /*loop_info.blocks = get_loop_n_blocks(cur_loop);*/
1305 cond_chain_entries = NEW_ARR_F(entry_edge, 0);
1306 head_df_loop = NEW_ARR_F(entry_edge, 0);
1308 /*head_inversion_node_count = 0;*/
1309 inversion_blocks_in_cc = 0;
1311 /* Use phase to keep copy of nodes from the condition chain. */
1312 ir_nodemap_init(&map, irg);
1313 obstack_init(&obst);
1315 /* Search for condition chains and temporarily save the blocks in an array. */
1316 cc_blocks = NEW_ARR_F(ir_node *, 0);
1317 inc_irg_visited(irg);
1318 find_condition_chain(loop_head);
1320 unmark_not_allowed_cc_blocks();
1321 DEL_ARR_F(cc_blocks);
1323 /* Condition chain too large.
1324 * Loop should better be small enough to fit into the cache. */
1325 /* TODO Of course, we should take a small enough cc in the first place,
1326 * which is not that simple. (bin packing) */
1327 if (loop_info.cc_size > opt_params.max_cc_size) {
1328 ++stats.cc_limit_reached;
1330 do_inversion = false;
1332 /* Unmark cc blocks except the head.
1333 * Invert head only for possible unrolling. */
1337 /* We also catch endless loops here,
1338 * because they do not have a condition chain. */
1339 if (inversion_blocks_in_cc < 1) {
1340 do_inversion = false;
1342 "Loop contains %d (less than 1) invertible blocks => No Inversion done.\n",
1343 inversion_blocks_in_cc));
1347 cur_head_outs = NEW_ARR_F(entry_edge, 0);
1349 /* Get all edges pointing into the condition chain. */
1350 irg_walk_graph(irg, get_head_outs, NULL, NULL);
1352 /* Do the inversion */
1353 inversion_walk(irg, cur_head_outs);
1355 DEL_ARR_F(cur_head_outs);
1357 /* Duplicated blocks changed doms */
1358 clear_irg_state(irg, IR_GRAPH_STATE_CONSISTENT_DOMINANCE
1359 | IR_GRAPH_STATE_CONSISTENT_LOOPINFO);
1365 obstack_free(&obst, NULL);
1366 ir_nodemap_destroy(&map);
1367 DEL_ARR_F(cond_chain_entries);
1368 DEL_ARR_F(head_df_loop);
1370 ir_free_resources(irg, IR_RESOURCE_BLOCK_MARK);
1373 /* Fix the original loop_heads ins for invariant unrolling case. */
1374 static void unrolling_fix_loop_head_inv(void)
1378 ir_node *proj = new_Proj(loop_info.duff_cond, mode_X, 0);
1379 ir_node *head_pred = get_irn_n(loop_head, loop_info.be_src_pos);
1380 ir_node *loop_condition = get_unroll_copy(head_pred, unroll_nr - 1);
1382 /* Original loop_heads ins are:
1383 * duff block and the own backedge */
1385 ins[0] = loop_condition;
1387 set_irn_in(loop_head, 2, ins);
1388 DB((dbg, LEVEL_4, "Rewire ins of block loophead %N to pred %N and duffs entry %N \n" , loop_head, ins[0], ins[1]));
1390 for_each_phi(loop_head, phi) {
1391 ir_node *pred = get_irn_n(phi, loop_info.be_src_pos);
1392 /* TODO we think it is a phi, but for Mergesort it is not the case.*/
1394 ir_node *last_pred = get_unroll_copy(pred, unroll_nr - 1);
1397 ins[1] = (ir_node*)get_irn_link(phi);
1398 set_irn_in(phi, 2, ins);
1399 DB((dbg, LEVEL_4, "Rewire ins of loophead phi %N to pred %N and duffs entry %N \n" , phi, ins[0], ins[1]));
1403 /* Removes previously created phis with only 1 in. */
1404 static void correct_phis(ir_node *node, void *env)
1408 if (is_Phi(node) && get_irn_arity(node) == 1) {
1412 in[0] = get_irn_n(node, 0);
1414 exch = new_rd_Phi(get_irn_dbg_info(node),
1415 get_nodes_block(node), 1, in,
1416 get_irn_mode(node));
1418 exchange(node, exch);
1422 /* Unrolling: Rewire floating copies. */
1423 static void place_copies(int copies)
1425 ir_node *loophead = loop_head;
1428 int be_src_pos = loop_info.be_src_pos;
1430 /* Serialize loops by fixing their head ins.
1431 * Processed are the copies.
1432 * The original loop is done after that, to keep backedge infos. */
1433 for (c = 0; c < copies; ++c) {
1434 ir_node *upper = get_unroll_copy(loophead, c);
1435 ir_node *lower = get_unroll_copy(loophead, c + 1);
1437 ir_node *topmost_be_block = get_nodes_block(get_irn_n(loophead, be_src_pos));
1439 /* Important: get the preds first and then their copy. */
1440 ir_node *upper_be_block = get_unroll_copy(topmost_be_block, c);
1441 ir_node *new_jmp = new_r_Jmp(upper_be_block);
1442 DB((dbg, LEVEL_5, " place_copies upper %N lower %N\n", upper, lower));
1444 DB((dbg, LEVEL_5, "topmost be block %N \n", topmost_be_block));
1446 if (loop_info.unroll_kind == constant) {
1449 set_irn_in(lower, 1, ins);
1451 for_each_phi(loophead, phi) {
1452 ir_node *topmost_def = get_irn_n(phi, be_src_pos);
1453 ir_node *upper_def = get_unroll_copy(topmost_def, c);
1454 ir_node *lower_phi = get_unroll_copy(phi, c + 1);
1456 /* It is possible, that the value used
1457 * in the OWN backedge path is NOT defined in this loop. */
1458 if (is_in_loop(topmost_def))
1461 ins[0] = topmost_def;
1463 set_irn_in(lower_phi, 1, ins);
1464 /* Need to replace phis with 1 in later. */
1467 /* Invariant case */
1468 /* Every node has 2 ins. One from the duff blocks
1469 * and one from the previously unrolled loop. */
1471 /* Calculate corresponding projection of mod result for this copy c */
1472 ir_node *proj = new_Proj(loop_info.duff_cond, mode_X, unroll_nr - c - 1);
1473 DB((dbg, LEVEL_4, "New duff proj %N\n" , proj));
1477 set_irn_in(lower, 2, ins);
1478 DB((dbg, LEVEL_4, "Rewire ins of Block %N to pred %N and duffs entry %N \n" , lower, ins[0], ins[1]));
1480 for_each_phi(loophead, phi) {
1481 ir_node *topmost_phi_pred = get_irn_n(phi, be_src_pos);
1482 ir_node *upper_phi_pred;
1486 lower_phi = get_unroll_copy(phi, c + 1);
1487 duff_phi = (ir_node*)get_irn_link(phi);
1488 DB((dbg, LEVEL_4, "DD Link of %N is %N\n" , phi, duff_phi));
1491 if (is_in_loop(topmost_phi_pred)) {
1492 upper_phi_pred = get_unroll_copy(topmost_phi_pred, c);
1494 upper_phi_pred = topmost_phi_pred;
1497 ins[0] = upper_phi_pred;
1499 set_irn_in(lower_phi, 2, ins);
1500 DB((dbg, LEVEL_4, "Rewire ins of %N to pred %N and duffs entry %N \n" , lower_phi, ins[0], ins[1]));
1505 /* Reconnect last copy. */
1506 for (i = 0; i < ARR_LEN(loop_entries); ++i) {
1507 entry_edge edge = loop_entries[i];
1508 /* Last copy is at the bottom */
1509 ir_node *new_pred = get_unroll_copy(edge.pred, copies);
1510 set_irn_n(edge.node, edge.pos, new_pred);
1513 /* Fix original loops head.
1514 * Done in the end, as ins and be info were needed before. */
1515 if (loop_info.unroll_kind == constant) {
1517 ir_node *head_pred = get_irn_n(loop_head, be_src_pos);
1518 ir_node *loop_condition = get_unroll_copy(head_pred, unroll_nr - 1);
1520 set_irn_n(loop_head, loop_info.be_src_pos, loop_condition);
1522 for_each_phi(loop_head, phi) {
1523 ir_node *pred = get_irn_n(phi, be_src_pos);
1526 /* It is possible, that the value used
1527 * in the OWN backedge path is NOT assigned in this loop. */
1528 if (is_in_loop(pred))
1529 last_pred = get_unroll_copy(pred, copies);
1532 set_irn_n(phi, be_src_pos, last_pred);
1536 unrolling_fix_loop_head_inv();
1540 /* Copies the cur_loop several times. */
1541 static void copy_loop(entry_edge *cur_loop_outs, int copies)
1545 ir_reserve_resources(current_ir_graph, IR_RESOURCE_IRN_VISITED);
1547 for (c = 0; c < copies; ++c) {
1550 inc_irg_visited(current_ir_graph);
1552 DB((dbg, LEVEL_5, " ### Copy_loop copy nr: %d ###\n", c));
1553 for (i = 0; i < ARR_LEN(cur_loop_outs); ++i) {
1554 entry_edge entry = cur_loop_outs[i];
1555 ir_node *pred = get_irn_n(entry.node, entry.pos);
1557 copy_walk_n(pred, is_in_loop, c + 1);
1561 ir_free_resources(current_ir_graph, IR_RESOURCE_IRN_VISITED);
1565 /* Creates a new phi from the given phi node omitting own bes,
1566 * using be_block as supplier of backedge informations. */
1567 static ir_node *clone_phis_sans_bes(ir_node *phi, ir_node *be_block, ir_node *dest_block)
1570 int arity = get_irn_arity(phi);
1574 assert(get_irn_arity(phi) == get_irn_arity(be_block));
1575 assert(is_Phi(phi));
1577 ins = NEW_ARR_F(ir_node *, arity);
1578 for (i = 0; i < arity; ++i) {
1579 if (! is_own_backedge(be_block, i)) {
1580 ins[c] = get_irn_n(phi, i);
1585 newphi = new_r_Phi(dest_block, c, ins, get_irn_mode(phi));
1587 set_irn_link(phi, newphi);
1588 DB((dbg, LEVEL_4, "Linking for duffs device %N to %N\n", phi, newphi));
1593 /* Creates a new block from the given block node omitting own bes,
1594 * using be_block as supplier of backedge informations. */
1595 static ir_node *clone_block_sans_bes(ir_node *node, ir_node *be_block)
1597 int arity = get_irn_arity(node);
1601 assert(get_irn_arity(node) == get_irn_arity(be_block));
1602 assert(is_Block(node));
1604 NEW_ARR_A(ir_node *, ins, arity);
1605 for (i = 0; i < arity; ++i) {
1606 if (! is_own_backedge(be_block, i)) {
1607 ins[c] = get_irn_n(node, i);
1612 return new_Block(c, ins);
1615 /* Creates a structure to calculate absolute value of node op.
1616 * Returns mux node with absolute value. */
1617 static ir_node *new_Abs(ir_node *op, ir_mode *mode)
1619 ir_graph *irg = get_irn_irg(op);
1620 ir_node *block = get_nodes_block(op);
1621 ir_node *zero = new_r_Const(irg, get_mode_null(mode));
1622 ir_node *cmp = new_r_Cmp(block, op, zero, ir_relation_less);
1623 ir_node *minus_op = new_r_Minus(block, op, mode);
1624 ir_node *mux = new_r_Mux(block, cmp, op, minus_op, mode);
1630 /* Creates blocks for duffs device, using previously obtained
1631 * informations about the iv.
1633 static void create_duffs_block(void)
1637 ir_node *block1, *count_block, *duff_block;
1638 ir_node *ems, *ems_mod, *ems_div, *ems_mod_proj, *cmp_null,
1639 *ems_mode_cond, *x_true, *x_false, *const_null;
1640 ir_node *true_val, *false_val;
1643 ir_node *duff_mod, *proj, *cond;
1645 ir_node *count, *correction, *unroll_c;
1646 ir_node *cmp_bad_count, *good_count, *bad_count, *count_phi, *bad_count_neg;
1649 mode = get_irn_mode(loop_info.end_val);
1650 const_null = new_Const(get_mode_null(mode));
1653 * 1. Calculate first approach to count.
1654 * Condition: (end - start) % step == 0 */
1655 block1 = clone_block_sans_bes(loop_head, loop_head);
1656 DB((dbg, LEVEL_4, "Duff block 1 %N\n", block1));
1658 /* Create loop entry phis in first duff block
1659 * as it becomes the loops preheader */
1660 for_each_phi(loop_head, phi) {
1661 /* Returns phis pred if phi would have arity 1*/
1662 ir_node *new_phi = clone_phis_sans_bes(phi, loop_head, block1);
1664 DB((dbg, LEVEL_4, "HEAD %N phi %N\n", loop_head, phi));
1665 DB((dbg, LEVEL_4, "BLOCK1 %N phi %N\n", block1, new_phi));
1668 ems = new_r_Sub(block1, loop_info.end_val, loop_info.start_val,
1669 get_irn_mode(loop_info.end_val));
1670 DB((dbg, LEVEL_4, "BLOCK1 sub %N\n", ems));
1673 ems = new_Sub(loop_info.end_val, loop_info.start_val,
1674 get_irn_mode(loop_info.end_val));
1676 DB((dbg, LEVEL_4, "mod ins %N %N\n", ems, loop_info.step));
1677 ems_mod = new_r_Mod(block1,
1682 op_pin_state_pinned);
1683 ems_div = new_r_Div(block1,
1688 op_pin_state_pinned);
1690 DB((dbg, LEVEL_4, "New module node %N\n", ems_mod));
1692 ems_mod_proj = new_r_Proj(ems_mod, mode_Iu, pn_Mod_res);
1693 cmp_null = new_r_Cmp(block1, ems_mod_proj, const_null, ir_relation_less);
1694 ems_mode_cond = new_r_Cond(block1, cmp_null);
1696 /* ems % step == 0 */
1697 x_true = new_r_Proj(ems_mode_cond, mode_X, pn_Cond_true);
1698 /* ems % step != 0 */
1699 x_false = new_r_Proj(ems_mode_cond, mode_X, pn_Cond_false);
1702 * Assures, duffs device receives a valid count.
1704 * decreasing: count < 0
1705 * increasing: count > 0
1710 count_block = new_Block(2, ins);
1711 DB((dbg, LEVEL_4, "Duff block 2 %N\n", count_block));
1714 /* Increase loop-taken-count depending on the loop condition
1715 * uses the latest iv to compare to. */
1716 if (loop_info.latest_value == 1) {
1717 /* ems % step == 0 : +0 */
1718 true_val = new_Const(get_mode_null(mode));
1719 /* ems % step != 0 : +1 */
1720 false_val = new_Const(get_mode_one(mode));
1722 ir_tarval *tv_two = new_tarval_from_long(2, mode);
1723 /* ems % step == 0 : +1 */
1724 true_val = new_Const(get_mode_one(mode));
1725 /* ems % step != 0 : +2 */
1726 false_val = new_Const(tv_two);
1732 correction = new_r_Phi(count_block, 2, ins, mode);
1734 count = new_r_Proj(ems_div, mode, pn_Div_res);
1736 /* (end - start) / step + correction */
1737 count = new_Add(count, correction, mode);
1739 /* We preconditioned the loop to be tail-controlled.
1740 * So, if count is something 'wrong' like 0,
1741 * negative/positive (depending on step direction),
1742 * we may take the loop once (tail-contr.) and leave it
1743 * to the existing condition, to break; */
1745 /* Depending on step direction, we have to check for > or < 0 */
1746 if (loop_info.decreasing == 1) {
1747 cmp_bad_count = new_r_Cmp(count_block, count, const_null,
1750 cmp_bad_count = new_r_Cmp(count_block, count, const_null,
1751 ir_relation_greater);
1754 bad_count_neg = new_r_Cond(count_block, cmp_bad_count);
1755 good_count = new_Proj(bad_count_neg, mode_X, pn_Cond_true);
1756 bad_count = new_Proj(ems_mode_cond, mode_X, pn_Cond_false);
1759 * Contains module to decide which loop to start from. */
1761 ins[0] = good_count;
1763 duff_block = new_Block(2, ins);
1764 DB((dbg, LEVEL_4, "Duff block 3 %N\n", duff_block));
1766 /* Get absolute value */
1767 ins[0] = new_Abs(count, mode);
1768 /* Manually feed the aforementioned count = 1 (bad case)*/
1769 ins[1] = new_Const(get_mode_one(mode));
1770 count_phi = new_r_Phi(duff_block, 2, ins, mode);
1772 unroll_c = new_Const(new_tarval_from_long((long)unroll_nr, mode));
1774 /* count % unroll_nr */
1775 duff_mod = new_r_Mod(duff_block,
1780 op_pin_state_pinned);
1783 proj = new_Proj(duff_mod, mode, pn_Mod_res);
1784 /* condition does NOT create itself in the block of the proj! */
1785 cond = new_r_Cond(duff_block, proj);
1787 loop_info.duff_cond = cond;
1790 /* Returns 1 if given node is not in loop,
1791 * or if it is a phi of the loop head with only loop invariant defs.
1793 static unsigned is_loop_invariant_def(ir_node *node)
1797 if (! is_in_loop(node)) {
1798 DB((dbg, LEVEL_4, "Not in loop %N\n", node));
1799 /* || is_Const(node) || is_SymConst(node)) {*/
1803 /* If this is a phi of the loophead shared by more than 1 loop,
1804 * we need to check if all defs are not in the loop. */
1807 block = get_nodes_block(node);
1809 /* To prevent unexpected situations. */
1810 if (block != loop_head) {
1814 for (i = 0; i < get_irn_arity(node); ++i) {
1815 /* Check if all bes are just loopbacks. */
1816 if (is_own_backedge(block, i) && get_irn_n(node, i) != node)
1819 DB((dbg, LEVEL_4, "invar %N\n", node));
1822 DB((dbg, LEVEL_4, "Not invar %N\n", node));
1827 /* Returns 1 if one pred of node is invariant and the other is not.
1828 * invar_pred and other are set analogously. */
1829 static unsigned get_invariant_pred(ir_node *node, ir_node **invar_pred, ir_node **other)
1831 ir_node *pred0 = get_irn_n(node, 0);
1832 ir_node *pred1 = get_irn_n(node, 1);
1837 if (is_loop_invariant_def(pred0)) {
1838 DB((dbg, LEVEL_4, "pred0 invar %N\n", pred0));
1839 *invar_pred = pred0;
1843 if (is_loop_invariant_def(pred1)) {
1844 DB((dbg, LEVEL_4, "pred1 invar %N\n", pred1));
1846 if (*invar_pred != NULL) {
1847 /* RETURN. We do not want both preds to be invariant. */
1852 *invar_pred = pred1;
1855 DB((dbg, LEVEL_4, "pred1 not invar %N\n", pred1));
1857 if (*invar_pred != NULL)
1864 /* Starts from a phi that may belong to an iv.
1865 * If an add forms a loop with iteration_phi,
1866 * and add uses a constant, 1 is returned
1867 * and 'start' as well as 'add' are sane. */
1868 static unsigned get_start_and_add(ir_node *iteration_phi, unrolling_kind_flag role)
1871 ir_node *found_add = loop_info.add;
1872 int arity = get_irn_arity(iteration_phi);
1874 DB((dbg, LEVEL_4, "Find start and add from %N\n", iteration_phi));
1876 for (i = 0; i < arity; ++i) {
1878 /* Find start_val which needs to be pred of the iteration_phi.
1879 * If start_val already known, sanity check. */
1880 if (!is_backedge(get_nodes_block(loop_info.iteration_phi), i)) {
1881 ir_node *found_start_val = get_irn_n(loop_info.iteration_phi, i);
1883 DB((dbg, LEVEL_4, "found_start_val %N\n", found_start_val));
1885 /* We already found a start_val it has to be always the same. */
1886 if (loop_info.start_val && found_start_val != loop_info.start_val)
1889 if ((role == constant) && !(is_SymConst(found_start_val) || is_Const(found_start_val)))
1891 else if((role == constant) && !(is_loop_invariant_def(found_start_val)))
1894 loop_info.start_val = found_start_val;
1897 /* The phi has to be in the loop head.
1898 * Follow all own backedges. Every value supplied from these preds of the phi
1899 * needs to origin from the same add. */
1900 if (is_own_backedge(get_nodes_block(loop_info.iteration_phi), i)) {
1901 ir_node *new_found = get_irn_n(loop_info.iteration_phi,i);
1903 DB((dbg, LEVEL_4, "is add? %N\n", new_found));
1905 if (! (is_Add(new_found) || is_Sub(new_found)) || (found_add && found_add != new_found))
1908 found_add = new_found;
1912 loop_info.add = found_add;
1918 /* Returns 1 if one pred of node is a const value and the other is not.
1919 * const_pred and other are set analogously. */
1920 static unsigned get_const_pred(ir_node *node, ir_node **const_pred, ir_node **other)
1922 ir_node *pred0 = get_irn_n(node, 0);
1923 ir_node *pred1 = get_irn_n(node, 1);
1925 DB((dbg, LEVEL_4, "Checking for constant pred of %N\n", node));
1930 /*DB((dbg, LEVEL_4, "is %N const\n", pred0));*/
1931 if (is_Const(pred0) || is_SymConst(pred0)) {
1932 *const_pred = pred0;
1936 /*DB((dbg, LEVEL_4, "is %N const\n", pred1));*/
1937 if (is_Const(pred1) || is_SymConst(pred1)) {
1938 if (*const_pred != NULL) {
1939 /* RETURN. We do not want both preds to be constant. */
1944 *const_pred = pred1;
1947 if (*const_pred == NULL)
1953 /* Returns 1 if loop exits within 2 steps of the iv.
1954 * Norm_proj means we do not exit the loop.*/
1955 static unsigned simulate_next(ir_tarval **count_tar,
1956 ir_tarval *stepped, ir_tarval *step_tar, ir_tarval *end_tar,
1957 ir_relation norm_proj)
1961 DB((dbg, LEVEL_4, "Loop taken if (stepped)%ld %s (end)%ld ",
1962 get_tarval_long(stepped),
1963 get_relation_string((norm_proj)),
1964 get_tarval_long(end_tar)));
1965 DB((dbg, LEVEL_4, "comparing latest value %d\n", loop_info.latest_value));
1967 /* If current iv does not stay in the loop,
1968 * this run satisfied the exit condition. */
1969 if (! (tarval_cmp(stepped, end_tar) & norm_proj))
1972 DB((dbg, LEVEL_4, "Result: (stepped)%ld IS %s (end)%ld\n",
1973 get_tarval_long(stepped),
1974 get_relation_string(tarval_cmp(stepped, end_tar)),
1975 get_tarval_long(end_tar)));
1978 if (is_Add(loop_info.add))
1979 next = tarval_add(stepped, step_tar);
1982 next = tarval_sub(stepped, step_tar, get_irn_mode(loop_info.end_val));
1984 DB((dbg, LEVEL_4, "Loop taken if %ld %s %ld ",
1985 get_tarval_long(next),
1986 get_relation_string(norm_proj),
1987 get_tarval_long(end_tar)));
1988 DB((dbg, LEVEL_4, "comparing latest value %d\n", loop_info.latest_value));
1990 /* Increase steps. */
1991 *count_tar = tarval_add(*count_tar, get_tarval_one(get_tarval_mode(*count_tar)));
1993 /* Next has to fail the loop condition, or we will never exit. */
1994 if (! (tarval_cmp(next, end_tar) & norm_proj))
2000 /* Check if loop meets requirements for a 'simple loop':
2001 * - Exactly one cf out
2003 * - Max nodes after unrolling
2007 * Returns Projection of cmp node or NULL; */
2008 static ir_node *is_simple_loop(void)
2011 ir_node *loop_block, *exit_block, *projx, *cond, *cmp;
2013 /* Maximum of one condition, and no endless loops. */
2014 if (loop_info.cf_outs != 1)
2017 DB((dbg, LEVEL_4, "1 loop exit\n"));
2019 /* Calculate maximum unroll_nr keeping node count below limit. */
2020 loop_info.max_unroll = (int)((double)opt_params.max_unrolled_loop_size / (double)loop_info.nodes);
2021 if (loop_info.max_unroll < 2) {
2026 DB((dbg, LEVEL_4, "maximum unroll factor %u, to not exceed node limit \n",
2027 opt_params.max_unrolled_loop_size));
2029 arity = get_irn_arity(loop_head);
2030 /* RETURN if we have more than 1 be. */
2031 /* Get my backedges without alien bes. */
2033 for (i = 0; i < arity; ++i) {
2034 ir_node *pred = get_irn_n(loop_head, i);
2035 if (is_own_backedge(loop_head, i)) {
2037 /* Our simple loops may have only one backedge. */
2040 loop_block = get_nodes_block(pred);
2041 loop_info.be_src_pos = i;
2046 DB((dbg, LEVEL_4, "loop has 1 own backedge.\n"));
2048 exit_block = get_nodes_block(loop_info.cf_out.pred);
2049 /* The loop has to be tail-controlled.
2050 * This can be changed/improved,
2051 * but we would need a duff iv. */
2052 if (exit_block != loop_block)
2055 DB((dbg, LEVEL_4, "tail-controlled loop.\n"));
2057 /* find value on which loop exit depends */
2058 projx = loop_info.cf_out.pred;
2059 cond = get_irn_n(projx, 0);
2060 cmp = get_irn_n(cond, 0);
2065 DB((dbg, LEVEL_5, "projection is %s\n", get_relation_string(get_Proj_proj(projx))));
2067 switch(get_Proj_proj(projx)) {
2069 loop_info.exit_cond = 0;
2072 loop_info.exit_cond = 1;
2075 panic("Cond Proj_proj other than true/false");
2078 DB((dbg, LEVEL_4, "Valid Cmp.\n"));
2082 /* Returns 1 if all nodes are mode_Iu or mode_Is. */
2083 static unsigned are_mode_I(ir_node *n1, ir_node* n2, ir_node *n3)
2085 ir_mode *m1 = get_irn_mode(n1);
2086 ir_mode *m2 = get_irn_mode(n2);
2087 ir_mode *m3 = get_irn_mode(n3);
2089 if ((m1 == mode_Iu && m2 == mode_Iu && m3 == mode_Iu) ||
2090 (m1 == mode_Is && m2 == mode_Is && m3 == mode_Is))
2096 /* Checks if cur_loop is a simple tail-controlled counting loop
2097 * with start and end value loop invariant, step constant. */
2098 static unsigned get_unroll_decision_invariant(void)
2101 ir_node *projres, *loop_condition, *iteration_path;
2103 ir_tarval *step_tar;
2107 /* RETURN if loop is not 'simple' */
2108 projres = is_simple_loop();
2109 if (projres == NULL)
2112 /* Use a minimal size for the invariant unrolled loop,
2113 * as duffs device produces overhead */
2114 if (loop_info.nodes < opt_params.invar_unrolling_min_size)
2117 loop_condition = get_irn_n(projres, 0);
2119 success = get_invariant_pred(loop_condition, &loop_info.end_val, &iteration_path);
2120 DB((dbg, LEVEL_4, "pred invar %d\n", success));
2125 DB((dbg, LEVEL_4, "Invariant End_val %N, other %N\n", loop_info.end_val, iteration_path));
2127 /* We may find the add or the phi first.
2128 * Until now we only have end_val. */
2129 if (is_Add(iteration_path) || is_Sub(iteration_path)) {
2131 loop_info.add = iteration_path;
2132 DB((dbg, LEVEL_4, "Case 1: Got add %N (maybe not sane)\n", loop_info.add));
2134 /* Preds of the add should be step and the iteration_phi */
2135 success = get_const_pred(loop_info.add, &loop_info.step, &loop_info.iteration_phi);
2139 DB((dbg, LEVEL_4, "Got step %N\n", loop_info.step));
2141 if (! is_Phi(loop_info.iteration_phi))
2144 DB((dbg, LEVEL_4, "Got phi %N\n", loop_info.iteration_phi));
2147 * Does necessary sanity check of add, if it is already set. */
2148 success = get_start_and_add(loop_info.iteration_phi, invariant);
2152 DB((dbg, LEVEL_4, "Got start A %N\n", loop_info.start_val));
2154 } else if (is_Phi(iteration_path)) {
2155 ir_node *new_iteration_phi;
2157 loop_info.iteration_phi = iteration_path;
2158 DB((dbg, LEVEL_4, "Case 2: Got phi %N\n", loop_info.iteration_phi));
2160 /* Find start_val and add-node.
2161 * Does necessary sanity check of add, if it is already set. */
2162 success = get_start_and_add(loop_info.iteration_phi, invariant);
2166 DB((dbg, LEVEL_4, "Got start B %N\n", loop_info.start_val));
2167 DB((dbg, LEVEL_4, "Got add or sub %N\n", loop_info.add));
2169 success = get_const_pred(loop_info.add, &loop_info.step, &new_iteration_phi);
2173 DB((dbg, LEVEL_4, "Got step (B) %N\n", loop_info.step));
2175 if (loop_info.iteration_phi != new_iteration_phi)
2182 mode = get_irn_mode(loop_info.end_val);
2184 DB((dbg, LEVEL_4, "start %N, end %N, step %N\n",
2185 loop_info.start_val, loop_info.end_val, loop_info.step));
2187 if (mode != mode_Is && mode != mode_Iu)
2190 /* TODO necessary? */
2191 if (!are_mode_I(loop_info.start_val, loop_info.step, loop_info.end_val))
2194 DB((dbg, LEVEL_4, "mode integer\n"));
2196 step_tar = get_Const_tarval(loop_info.step);
2198 if (tarval_is_null(step_tar)) {
2199 /* TODO Might be worth a warning. */
2203 DB((dbg, LEVEL_4, "step is not 0\n"));
2205 create_duffs_block();
2207 return loop_info.max_unroll;
2210 /* Returns unroll factor,
2211 * given maximum unroll factor and number of loop passes. */
2212 static unsigned get_preferred_factor_constant(ir_tarval *count_tar)
2214 ir_tarval *tar_6, *tar_5, *tar_4, *tar_3, *tar_2;
2216 ir_mode *mode = get_irn_mode(loop_info.end_val);
2218 tar_6 = new_tarval_from_long(6, mode);
2219 tar_5 = new_tarval_from_long(5, mode);
2220 tar_4 = new_tarval_from_long(4, mode);
2221 tar_3 = new_tarval_from_long(3, mode);
2222 tar_2 = new_tarval_from_long(2, mode);
2224 /* loop passes % {6, 5, 4, 3, 2} == 0 */
2225 if (tarval_is_null(tarval_mod(count_tar, tar_6)))
2227 else if (tarval_is_null(tarval_mod(count_tar, tar_5)))
2229 else if (tarval_is_null(tarval_mod(count_tar, tar_4)))
2231 else if (tarval_is_null(tarval_mod(count_tar, tar_3)))
2233 else if (tarval_is_null(tarval_mod(count_tar, tar_2)))
2236 /* gcd(max_unroll, count_tar) */
2237 int a = loop_info.max_unroll;
2238 int b = (int)get_tarval_long(count_tar);
2241 DB((dbg, LEVEL_4, "gcd of max_unroll %d and count_tar %d: ", a, b));
2248 DB((dbg, LEVEL_4, "%d\n", a));
2252 DB((dbg, LEVEL_4, "preferred unroll factor %d\n", prefer));
2255 * If our preference is greater than the allowed unroll factor
2256 * we either might reduce the preferred factor and prevent a duffs device block,
2257 * or create a duffs device block, from which in this case (constants only)
2258 * we know the startloop at compiletime.
2259 * The latter yields the following graphs.
2260 * but for code generation we would want to use graph A.
2261 * The graphs are equivalent. So, we can only reduce the preferred factor.
2271 if (prefer <= loop_info.max_unroll)
2276 if (loop_info.max_unroll >= 3)
2278 else if (loop_info.max_unroll >= 2)
2284 if (loop_info.max_unroll >= 2)
2295 /* Check if cur_loop is a simple counting loop.
2296 * Start, step and end are constants.
2297 * TODO The whole constant case should use procedures similar to
2298 * the invariant case, as they are more versatile. */
2300 static unsigned get_unroll_decision_constant(void)
2302 ir_node *cmp, *iteration_path;
2303 unsigned success, is_latest_val;
2304 ir_tarval *start_tar, *end_tar, *step_tar, *diff_tar, *count_tar;
2306 ir_relation proj_proj, norm_proj;
2309 /* RETURN if loop is not 'simple' */
2310 cmp = is_simple_loop();
2314 /* One in of the loop condition needs to be loop invariant. => end_val
2315 * The other in is assigned by an add. => add
2316 * The add uses a loop invariant value => step
2317 * and a phi with a loop invariant start_val and the add node as ins.
2329 success = get_const_pred(cmp, &loop_info.end_val, &iteration_path);
2333 DB((dbg, LEVEL_4, "End_val %N, other %N\n", loop_info.end_val, iteration_path));
2335 /* We may find the add or the phi first.
2336 * Until now we only have end_val. */
2337 if (is_Add(iteration_path) || is_Sub(iteration_path)) {
2339 /* We test against the latest value of the iv. */
2342 loop_info.add = iteration_path;
2343 DB((dbg, LEVEL_4, "Case 2: Got add %N (maybe not sane)\n", loop_info.add));
2345 /* Preds of the add should be step and the iteration_phi */
2346 success = get_const_pred(loop_info.add, &loop_info.step, &loop_info.iteration_phi);
2350 DB((dbg, LEVEL_4, "Got step %N\n", loop_info.step));
2352 if (! is_Phi(loop_info.iteration_phi))
2355 DB((dbg, LEVEL_4, "Got phi %N\n", loop_info.iteration_phi));
2358 * Does necessary sanity check of add, if it is already set. */
2359 success = get_start_and_add(loop_info.iteration_phi, constant);
2363 DB((dbg, LEVEL_4, "Got start %N\n", loop_info.start_val));
2365 } else if (is_Phi(iteration_path)) {
2366 ir_node *new_iteration_phi;
2368 /* We compare with the value the iv had entering this run. */
2371 loop_info.iteration_phi = iteration_path;
2372 DB((dbg, LEVEL_4, "Case 1: Got phi %N \n", loop_info.iteration_phi));
2374 /* Find start_val and add-node.
2375 * Does necessary sanity check of add, if it is already set. */
2376 success = get_start_and_add(loop_info.iteration_phi, constant);
2380 DB((dbg, LEVEL_4, "Got start %N\n", loop_info.start_val));
2381 DB((dbg, LEVEL_4, "Got add or sub %N\n", loop_info.add));
2383 success = get_const_pred(loop_info.add, &loop_info.step, &new_iteration_phi);
2387 DB((dbg, LEVEL_4, "Got step %N\n", loop_info.step));
2389 if (loop_info.iteration_phi != new_iteration_phi)
2397 mode = get_irn_mode(loop_info.end_val);
2399 DB((dbg, LEVEL_4, "start %N, end %N, step %N\n",
2400 loop_info.start_val, loop_info.end_val, loop_info.step));
2402 if (mode != mode_Is && mode != mode_Iu)
2405 /* TODO necessary? */
2406 if (!are_mode_I(loop_info.start_val, loop_info.step, loop_info.end_val))
2409 DB((dbg, LEVEL_4, "mode integer\n"));
2411 end_tar = get_Const_tarval(loop_info.end_val);
2412 start_tar = get_Const_tarval(loop_info.start_val);
2413 step_tar = get_Const_tarval(loop_info.step);
2415 if (tarval_is_null(step_tar))
2416 /* TODO Might be worth a warning. */
2419 DB((dbg, LEVEL_4, "step is not 0\n"));
2421 if ((!tarval_is_negative(step_tar)) ^ (!is_Sub(loop_info.add)))
2422 loop_info.decreasing = 1;
2424 diff_tar = tarval_sub(end_tar, start_tar, mode);
2426 /* We need at least count_tar steps to be close to end_val, maybe more.
2427 * No way, that we have gone too many steps.
2428 * This represents the 'latest value'.
2429 * (If condition checks against latest value, is checked later) */
2430 count_tar = tarval_div(diff_tar, step_tar);
2432 /* Iv will not pass end_val (except overflows).
2433 * Nothing done, as it would yield to no advantage. */
2434 if (tarval_is_negative(count_tar)) {
2435 DB((dbg, LEVEL_4, "Loop is endless or never taken."));
2436 /* TODO Might be worth a warning. */
2440 ++stats.u_simple_counting_loop;
2442 loop_info.latest_value = is_latest_val;
2445 if (! is_simple_counting_loop(&count_tar))
2449 /* stepped can be negative, if step < 0 */
2450 stepped = tarval_mul(count_tar, step_tar);
2452 /* step as close to end_val as possible, */
2453 /* |stepped| <= |end_tar|, and dist(stepped, end_tar) is smaller than a step. */
2454 if (is_Sub(loop_info.add))
2455 stepped = tarval_sub(start_tar, stepped, mode_Is);
2457 stepped = tarval_add(start_tar, stepped);
2459 DB((dbg, LEVEL_4, "stepped to %ld\n", get_tarval_long(stepped)));
2461 proj_proj = get_Cmp_relation(cmp);
2462 /* Assure that norm_proj is the stay-in-loop case. */
2463 if (loop_info.exit_cond == 1)
2464 norm_proj = get_negated_relation(proj_proj);
2466 norm_proj = proj_proj;
2468 DB((dbg, LEVEL_4, "normalized projection %s\n", get_relation_string(norm_proj)));
2469 /* Executed at most once (stay in counting loop if a Eq b) */
2470 if (norm_proj == ir_relation_equal)
2471 /* TODO Might be worth a warning. */
2474 /* calculates next values and increases count_tar according to it */
2475 success = simulate_next(&count_tar, stepped, step_tar, end_tar, norm_proj);
2479 /* We run loop once more, if we compare to the
2480 * not yet in-/decreased iv. */
2481 if (is_latest_val == 0) {
2482 DB((dbg, LEVEL_4, "condition uses not latest iv value\n"));
2483 count_tar = tarval_add(count_tar, get_tarval_one(mode));
2486 DB((dbg, LEVEL_4, "loop taken %ld times\n", get_tarval_long(count_tar)));
2488 /* Assure the loop is taken at least 1 time. */
2489 if (tarval_is_null(count_tar)) {
2490 /* TODO Might be worth a warning. */
2494 loop_info.count_tar = count_tar;
2495 return get_preferred_factor_constant(count_tar);
2501 static void unroll_loop(void)
2504 if (! (loop_info.nodes > 0))
2507 if (loop_info.nodes > opt_params.max_unrolled_loop_size) {
2508 DB((dbg, LEVEL_2, "Nodes %d > allowed nodes %d\n",
2509 loop_info.nodes, opt_params.max_unrolled_loop_size));
2514 if (loop_info.calls > 0) {
2515 DB((dbg, LEVEL_2, "Calls %d > allowed calls 0\n",
2517 ++stats.calls_limit;
2523 /* get_unroll_decision_constant and invariant are completely
2524 * independent for flexibility.
2525 * Some checks may be performed twice. */
2527 /* constant case? */
2528 if (opt_params.allow_const_unrolling)
2529 unroll_nr = get_unroll_decision_constant();
2530 if (unroll_nr > 1) {
2531 loop_info.unroll_kind = constant;
2534 /* invariant case? */
2535 if (opt_params.allow_invar_unrolling)
2536 unroll_nr = get_unroll_decision_invariant();
2538 loop_info.unroll_kind = invariant;
2541 DB((dbg, LEVEL_2, " *** Unrolling %d times ***\n", unroll_nr));
2543 if (unroll_nr > 1) {
2544 loop_entries = NEW_ARR_F(entry_edge, 0);
2547 irg_walk_graph(current_ir_graph, get_loop_entries, NULL, NULL);
2549 if (loop_info.unroll_kind == constant) {
2550 if ((int)get_tarval_long(loop_info.count_tar) == unroll_nr)
2551 loop_info.needs_backedge = 0;
2553 loop_info.needs_backedge = 1;
2555 loop_info.needs_backedge = 1;
2558 /* Use phase to keep copy of nodes from the condition chain. */
2559 ir_nodemap_init(&map, current_ir_graph);
2560 obstack_init(&obst);
2562 /* Copies the loop */
2563 copy_loop(loop_entries, unroll_nr - 1);
2565 /* Line up the floating copies. */
2566 place_copies(unroll_nr - 1);
2568 /* Remove phis with 1 in
2569 * If there were no nested phis, this would not be necessary.
2570 * Avoiding the creation in the first place
2571 * leads to complex special cases. */
2572 irg_walk_graph(current_ir_graph, correct_phis, NULL, NULL);
2574 if (loop_info.unroll_kind == constant)
2575 ++stats.constant_unroll;
2577 ++stats.invariant_unroll;
2579 clear_irg_state(current_ir_graph, IR_GRAPH_STATE_CONSISTENT_DOMINANCE);
2581 DEL_ARR_F(loop_entries);
2582 obstack_free(&obst, NULL);
2583 ir_nodemap_destroy(&map);
2588 /* Analyzes the loop, and checks if size is within allowed range.
2589 * Decides if loop will be processed. */
2590 static void init_analyze(ir_graph *irg, ir_loop *loop)
2595 loop_head_valid = true;
2597 /* Reset loop info */
2598 memset(&loop_info, 0, sizeof(loop_info_t));
2600 DB((dbg, LEVEL_1, " >>>> current loop %ld <<<\n",
2601 get_loop_loop_nr(loop)));
2603 /* Collect loop informations: head, node counts. */
2604 irg_walk_graph(irg, get_loop_info, NULL, NULL);
2606 /* RETURN if there is no valid head */
2607 if (!loop_head || !loop_head_valid) {
2608 DB((dbg, LEVEL_1, "No valid loop head. Nothing done.\n"));
2611 DB((dbg, LEVEL_1, "Loophead: %N\n", loop_head));
2614 if (loop_info.branches > opt_params.max_branches) {
2615 DB((dbg, LEVEL_1, "Branches %d > allowed branches %d\n",
2616 loop_info.branches, opt_params.max_branches));
2617 ++stats.calls_limit;
2622 case loop_op_inversion:
2623 loop_inversion(irg);
2626 case loop_op_unrolling:
2631 panic("Loop optimization not implemented.");
2633 DB((dbg, LEVEL_1, " <<<< end of loop with node %ld >>>>\n",
2634 get_loop_loop_nr(loop)));
2637 /* Find innermost loops and add them to loops. */
2638 static void find_innermost_loop(ir_loop *loop)
2640 bool had_sons = false;
2641 size_t n_elements = get_loop_n_elements(loop);
2644 for (e = 0; e < n_elements; ++e) {
2645 loop_element element = get_loop_element(loop, e);
2646 if (*element.kind == k_ir_loop) {
2647 find_innermost_loop(element.son);
2653 ARR_APP1(ir_loop*, loops, loop);
2657 static void set_loop_params(void)
2659 opt_params.max_loop_size = 100;
2660 opt_params.depth_adaption = -50;
2661 opt_params.count_phi = true;
2662 opt_params.count_proj = false;
2663 opt_params.allowed_calls = 0;
2665 opt_params.max_cc_size = 5;
2668 opt_params.allow_const_unrolling = true;
2669 opt_params.allow_invar_unrolling = false;
2671 opt_params.invar_unrolling_min_size = 20;
2672 opt_params.max_unrolled_loop_size = 400;
2673 opt_params.max_branches = 9999;
2676 /* Assure preconditions are met and go through all loops. */
2677 void loop_optimization(ir_graph *irg)
2685 /* Reset stats for this procedure */
2689 set_current_ir_graph(irg);
2691 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_PHI_LIST);
2692 collect_phiprojs(irg);
2694 loop = get_irg_loop(irg);
2696 loops = NEW_ARR_F(ir_loop *, 0);
2697 /* List all inner loops */
2698 n_elements = get_loop_n_elements(loop);
2699 for (i = 0; i < n_elements; ++i) {
2700 loop_element element = get_loop_element(loop, i);
2701 if (*element.kind != k_ir_loop)
2703 find_innermost_loop(element.son);
2706 /* Set all links to NULL */
2707 irg_walk_graph(irg, reset_link, NULL, NULL);
2709 for (i = 0; i < ARR_LEN(loops); ++i) {
2710 ir_loop *loop = loops[i];
2714 /* Analyze and handle loop */
2715 init_analyze(irg, loop);
2717 /* Copied blocks do not have their phi list yet */
2718 collect_phiprojs(irg);
2720 /* Set links to NULL
2721 * TODO Still necessary? */
2722 irg_walk_graph(irg, reset_link, NULL, NULL);
2728 ir_free_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_PHI_LIST);
2731 static ir_graph_state_t perform_loop_unrolling(ir_graph *irg)
2733 loop_op = loop_op_unrolling;
2734 loop_optimization(irg);
2738 static ir_graph_state_t perform_loop_inversion(ir_graph *irg)
2740 loop_op = loop_op_inversion;
2741 loop_optimization(irg);
2745 static ir_graph_state_t perform_loop_peeling(ir_graph *irg)
2747 loop_op = loop_op_peeling;
2748 loop_optimization(irg);
2752 static optdesc_t opt_unroll_loops = {
2754 IR_GRAPH_STATE_CONSISTENT_OUT_EDGES | IR_GRAPH_STATE_CONSISTENT_OUTS | IR_GRAPH_STATE_CONSISTENT_LOOPINFO,
2755 perform_loop_unrolling,
2758 static optdesc_t opt_invert_loops = {
2760 IR_GRAPH_STATE_CONSISTENT_OUT_EDGES | IR_GRAPH_STATE_CONSISTENT_OUTS | IR_GRAPH_STATE_CONSISTENT_LOOPINFO,
2761 perform_loop_inversion,
2764 static optdesc_t opt_peel_loops = {
2766 IR_GRAPH_STATE_CONSISTENT_OUT_EDGES | IR_GRAPH_STATE_CONSISTENT_OUTS | IR_GRAPH_STATE_CONSISTENT_LOOPINFO,
2767 perform_loop_peeling,
2770 void do_loop_unrolling(ir_graph *irg)
2771 { perform_irg_optimization(irg, &opt_unroll_loops); }
2773 void do_loop_inversion(ir_graph *irg)
2774 { perform_irg_optimization(irg, &opt_invert_loops); }
2776 void do_loop_peeling(ir_graph *irg)
2777 { perform_irg_optimization(irg, &opt_peel_loops); }
2779 ir_graph_pass_t *loop_inversion_pass(const char *name)
2781 return def_graph_pass(name ? name : "loop_inversion", do_loop_inversion);
2784 ir_graph_pass_t *loop_unroll_pass(const char *name)
2786 return def_graph_pass(name ? name : "loop_unroll", do_loop_unrolling);
2789 ir_graph_pass_t *loop_peeling_pass(const char *name)
2791 return def_graph_pass(name ? name : "loop_peeling", do_loop_peeling);
2794 void firm_init_loop_opt(void)
2796 FIRM_DBG_REGISTER(dbg, "firm.opt.loop");