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) {
595 /* Extends a block by a copy of its pred at pos,
596 * fixing also the phis in the same way. */
597 static void extend_ins_by_copy(ir_node *block, int pos)
602 assert(is_Block(block));
604 /* Extend block by copy of definition at pos */
605 pred = get_irn_n(block, pos);
606 new_in = get_inversion_copy(pred);
607 DB((dbg, LEVEL_5, "Extend block %N by %N cp of %N\n", block, new_in, pred));
608 extend_irn(block, new_in, false);
610 /* Extend block phis by copy of definition at pos */
611 for_each_phi(block, phi) {
614 pred = get_irn_n(phi, pos);
615 cp = get_inversion_copy(pred);
616 /* If the phis in is not in the condition chain (eg. a constant),
617 * there is no copy. */
623 DB((dbg, LEVEL_5, "Extend phi %N by %N cp of %N\n", phi, new_in, pred));
624 extend_irn(phi, new_in, false);
628 /* Returns the number of blocks backedges. With or without alien bes. */
629 static int get_backedge_n(ir_node *block, bool with_alien)
633 int arity = get_irn_arity(block);
635 assert(is_Block(block));
637 for (i = 0; i < arity; ++i) {
638 ir_node *pred = get_irn_n(block, i);
639 if (is_backedge(block, i) && (with_alien || is_in_loop(pred)))
645 /* Returns a raw copy of the given node.
646 * Attributes are kept/set according to the needs of loop inversion. */
647 static ir_node *copy_node(ir_node *node)
652 cp = exact_copy(node);
653 arity = get_irn_arity(node);
655 /* Keep backedge info */
656 for (i = 0; i < arity; ++i) {
657 if (is_backedge(node, i))
662 set_Block_mark(cp, 0);
670 * This walker copies all walked nodes.
671 * If the walk_condition is true for a node, it is copied.
672 * All nodes node_info->copy have to be NULL prior to every walk.
673 * Order of ins is important for later usage.
675 static void copy_walk(ir_node *node, walker_condition *walk_condition,
682 ir_graph *irg = current_ir_graph;
685 * break condition and cycle resolver, creating temporary node copies
687 if (get_irn_visited(node) >= get_irg_visited(irg)) {
688 /* Here we rely on nodestate's copy being initialized with NULL */
689 DB((dbg, LEVEL_5, "copy_walk: We have already visited %N\n", node));
690 if (get_inversion_copy(node) == NULL) {
691 cp = copy_node(node);
692 set_inversion_copy(node, cp);
694 DB((dbg, LEVEL_5, "The TEMP copy of %N is created %N\n", node, cp));
700 mark_irn_visited(node);
702 if (!is_Block(node)) {
703 ir_node *pred = get_nodes_block(node);
704 if (walk_condition(pred))
705 DB((dbg, LEVEL_5, "walk block %N\n", pred));
706 copy_walk(pred, walk_condition, set_loop);
709 arity = get_irn_arity(node);
711 NEW_ARR_A(ir_node *, cpin, arity);
713 for (i = 0; i < arity; ++i) {
714 ir_node *pred = get_irn_n(node, i);
716 if (walk_condition(pred)) {
717 DB((dbg, LEVEL_5, "walk node %N\n", pred));
718 copy_walk(pred, walk_condition, set_loop);
719 cpin[i] = get_inversion_copy(pred);
720 DB((dbg, LEVEL_5, "copy of %N gets new in %N which is copy of %N\n",
721 node, get_inversion_copy(pred), pred));
727 /* copy node / finalize temp node */
728 if (get_inversion_copy(node) == NULL) {
729 /* No temporary copy existent */
730 cp = copy_node(node);
731 set_inversion_copy(node, cp);
732 DB((dbg, LEVEL_5, "The FINAL copy of %N is CREATED %N\n", node, cp));
734 /* temporary copy is existent but without correct ins */
735 cp = get_inversion_copy(node);
736 DB((dbg, LEVEL_5, "The FINAL copy of %N is EXISTENT %N\n", node, cp));
739 if (!is_Block(node)) {
740 ir_node *cpblock = get_inversion_copy(get_nodes_block(node));
742 set_nodes_block(cp, cpblock );
744 add_Block_phi(cpblock, cp);
747 /* Keeps phi list of temporary node. */
748 set_irn_in(cp, ARR_LEN(cpin), cpin);
752 * This walker copies all walked nodes.
753 * If the walk_condition is true for a node, it is copied.
754 * All nodes node_info->copy have to be NULL prior to every walk.
755 * Order of ins is important for later usage.
756 * Takes copy_index, to phase-link copy at specific index.
758 static void copy_walk_n(ir_node *node, walker_condition *walk_condition,
767 * break condition and cycle resolver, creating temporary node copies
769 if (irn_visited(node)) {
770 /* Here we rely on nodestate's copy being initialized with NULL */
771 DB((dbg, LEVEL_5, "copy_walk: We have already visited %N\n", node));
772 if (get_unroll_copy(node, copy_index) == NULL) {
775 set_unroll_copy(node, copy_index, u);
776 DB((dbg, LEVEL_5, "The TEMP unknown of %N is created %N\n", node, u));
782 mark_irn_visited(node);
784 if (!is_Block(node)) {
785 ir_node *block = get_nodes_block(node);
786 if (walk_condition(block))
787 DB((dbg, LEVEL_5, "walk block %N\n", block));
788 copy_walk_n(block, walk_condition, copy_index);
791 arity = get_irn_arity(node);
792 NEW_ARR_A(ir_node *, cpin, arity);
794 for (i = 0; i < arity; ++i) {
795 ir_node *pred = get_irn_n(node, i);
797 if (walk_condition(pred)) {
798 DB((dbg, LEVEL_5, "walk node %N\n", pred));
799 copy_walk_n(pred, walk_condition, copy_index);
800 cpin[i] = get_unroll_copy(pred, copy_index);
806 /* copy node / finalize temp node */
807 cp = get_unroll_copy(node, copy_index);
808 if (cp == NULL || is_Unknown(cp)) {
809 cp = copy_node(node);
810 set_unroll_copy(node, copy_index, cp);
811 DB((dbg, LEVEL_5, "The FINAL copy of %N is CREATED %N\n", node, cp));
813 /* temporary copy is existent but without correct ins */
814 cp = get_unroll_copy(node, copy_index);
815 DB((dbg, LEVEL_5, "The FINAL copy of %N is EXISTENT %N\n", node, cp));
818 if (!is_Block(node)) {
819 ir_node *cpblock = get_unroll_copy(get_nodes_block(node), copy_index);
821 set_nodes_block(cp, cpblock );
823 add_Block_phi(cpblock, cp);
826 /* Keeps phi list of temporary node. */
827 set_irn_in(cp, ARR_LEN(cpin), cpin);
830 /* Removes alle Blocks with non marked predecessors from the condition chain. */
831 static void unmark_not_allowed_cc_blocks(void)
833 size_t blocks = ARR_LEN(cc_blocks);
836 for(i = 0; i < blocks; ++i) {
837 ir_node *block = cc_blocks[i];
839 int arity = get_irn_arity(block);
841 /* Head is an exception. */
842 if (block == loop_head)
845 for(a = 0; a < arity; ++a) {
846 if (! is_nodes_block_marked(get_irn_n(block, a))) {
847 set_Block_mark(block, 0);
848 --inversion_blocks_in_cc;
849 DB((dbg, LEVEL_5, "Removed %N from cc (blocks in cc %d)\n",
850 block, inversion_blocks_in_cc));
858 /* Unmarks all cc blocks using cc_blocks except head.
859 * TODO: invert head for unrolling? */
860 static void unmark_cc_blocks(void)
862 size_t blocks = ARR_LEN(cc_blocks);
865 for(i = 0; i < blocks; ++i) {
866 ir_node *block = cc_blocks[i];
868 /* TODO Head is an exception. */
869 /*if (block != loop_head)*/
870 set_Block_mark(block, 0);
872 /*inversion_blocks_in_cc = 1;*/
873 inversion_blocks_in_cc = 0;
876 loop_info.cc_size = 0;
880 * Populates head_entries with (node, pred_pos) tuple
881 * whereas the node's pred at pred_pos is in the cc but not the node itself.
882 * Also finds df loops inside the cc.
883 * Head and condition chain blocks have been marked previously.
885 static void get_head_outs(ir_node *node, void *env)
888 int arity = get_irn_arity(node);
891 for (i = 0; i < arity; ++i) {
892 if (!is_nodes_block_marked(node) && is_nodes_block_marked(get_irn_n(node, i))) {
896 /* Saving also predecessor seems redundant, but becomes
897 * necessary when changing position of it, before
898 * dereferencing it.*/
899 entry.pred = get_irn_n(node, i);
900 ARR_APP1(entry_edge, cur_head_outs, entry);
904 arity = get_irn_arity(loop_head);
906 /* Find df loops inside the cc */
907 if (is_Phi(node) && get_nodes_block(node) == loop_head) {
908 for (i = 0; i < arity; ++i) {
909 if (is_own_backedge(loop_head, i)) {
910 if (is_nodes_block_marked(get_irn_n(node, i))) {
914 entry.pred = get_irn_n(node, i);
915 ARR_APP1(entry_edge, head_df_loop, entry);
916 DB((dbg, LEVEL_5, "Found incc assignment node %N @%d is pred %N, graph %N %N\n",
917 node, i, entry.pred, current_ir_graph, get_irg_start_block(current_ir_graph)));
925 * Find condition chains, and add them to be inverted.
926 * A block belongs to the chain if a condition branches out of the loop.
927 * (Some blocks need to be removed once again.)
928 * Returns 1 if the given block belongs to the condition chain.
930 static void find_condition_chain(ir_node *block)
932 const ir_edge_t *edge;
935 bool jmp_only = true;
936 unsigned nodes_n = 0;
938 mark_irn_visited(block);
940 DB((dbg, LEVEL_5, "condition_chains for block %N\n", block));
943 foreach_out_edge_kind(block, edge, EDGE_KIND_NORMAL) {
947 /* Check if node count would exceed maximum cc size.
949 * This is not optimal, as we search depth-first and break here,
950 * continuing with another subtree. */
951 if (loop_info.cc_size + nodes_n > opt_params.max_cc_size) {
952 set_Block_mark(block, 0);
956 /* Check if block only has a jmp instruction. */
957 foreach_out_edge(block, edge) {
958 ir_node *src = get_edge_src_irn(edge);
960 if (!is_Block(src) && !is_Jmp(src)) {
965 /* Check cf outs if one is leaving the loop,
966 * or if this node has a backedge. */
967 foreach_block_succ(block, edge) {
968 ir_node *src = get_edge_src_irn(edge);
969 int pos = get_edge_src_pos(edge);
971 if (!is_in_loop(src))
974 /* Inverting blocks with backedge outs leads to a cf edge
975 * from the inverted head, into the inverted head (skipping the body).
976 * As the body becomes the new loop head,
977 * this would introduce another loop in the existing loop.
978 * This loop inversion cannot cope with this case. */
979 if (is_backedge(src, pos)) {
985 /* We need all predecessors to already belong to the condition chain.
986 * Example of wrong case: * == in cc
997 /* Collect blocks containing only a Jmp.
998 * Do not collect blocks with backedge outs. */
999 if ((jmp_only || mark) && !has_be) {
1000 set_Block_mark(block, 1);
1001 ++inversion_blocks_in_cc;
1002 loop_info.cc_size += nodes_n;
1003 DB((dbg, LEVEL_5, "block %N is part of condition chain\n", block));
1004 ARR_APP1(ir_node *, cc_blocks, block);
1006 set_Block_mark(block, 0);
1009 foreach_block_succ(block, edge) {
1010 ir_node *src = get_edge_src_irn( edge );
1012 if (is_in_loop(src) && ! irn_visited(src))
1013 find_condition_chain(src);
1018 * Rewires the copied condition chain. Removes backedges
1019 * as this condition chain is prior to the loop.
1020 * Copy of loop_head must have phi list and old (unfixed) backedge info of the loop head.
1021 * (loop_head is already fixed, we cannot rely on it.)
1023 static void fix_copy_inversion(void)
1028 ir_node *phi, *next;
1029 ir_node *head_cp = get_inversion_copy(loop_head);
1030 ir_graph *irg = get_irn_irg(head_cp);
1031 int arity = get_irn_arity(head_cp);
1032 int backedges = get_backedge_n(head_cp, false);
1033 int new_arity = arity - backedges;
1037 NEW_ARR_A(ir_node *, ins, new_arity);
1040 /* Remove block backedges */
1041 for(i = 0; i < arity; ++i) {
1042 if (!is_backedge(head_cp, i))
1043 ins[pos++] = get_irn_n(head_cp, i);
1046 new_head = new_r_Block(irg, new_arity, ins);
1048 phis = NEW_ARR_F(ir_node *, 0);
1050 for_each_phi_safe(get_Block_phis(head_cp), phi, next) {
1052 NEW_ARR_A(ir_node *, ins, new_arity);
1054 for(i = 0; i < arity; ++i) {
1055 if (!is_backedge(head_cp, i))
1056 ins[pos++] = get_irn_n(phi, i);
1058 new_phi = new_rd_Phi(get_irn_dbg_info(phi),
1059 new_head, new_arity, ins,
1061 ARR_APP1(ir_node *, phis, new_phi);
1065 for_each_phi_safe(get_Block_phis(head_cp), phi, next) {
1066 exchange(phi, phis[pos++]);
1069 exchange(head_cp, new_head);
1075 /* Puts the original condition chain at the end of the loop,
1076 * subsequently to the body.
1077 * Relies on block phi list and correct backedges.
1079 static void fix_head_inversion(void)
1083 ir_node *phi, *next;
1085 ir_graph *irg = get_irn_irg(loop_head);
1086 int arity = get_irn_arity(loop_head);
1087 int backedges = get_backedge_n(loop_head, false);
1088 int new_arity = backedges;
1092 NEW_ARR_A(ir_node *, ins, new_arity);
1095 /* Keep only backedges */
1096 for(i = 0; i < arity; ++i) {
1097 if (is_own_backedge(loop_head, i))
1098 ins[pos++] = get_irn_n(loop_head, i);
1101 new_head = new_r_Block(irg, new_arity, ins);
1103 phis = NEW_ARR_F(ir_node *, 0);
1105 for_each_phi(loop_head, phi) {
1107 DB((dbg, LEVEL_5, "Fixing phi %N of loop head\n", phi));
1109 NEW_ARR_A(ir_node *, ins, new_arity);
1112 for (i = 0; i < arity; ++i) {
1113 ir_node *pred = get_irn_n(phi, i);
1115 if (is_own_backedge(loop_head, i)) {
1116 /* If assignment is in the condition chain,
1117 * we need to create a phi in the new loop head.
1118 * This can only happen for df, not cf. See find_condition_chains. */
1119 /*if (is_nodes_block_marked(pred)) {
1127 new_phi = new_rd_Phi(get_irn_dbg_info(phi),
1128 new_head, new_arity, ins,
1131 ARR_APP1(ir_node *, phis, new_phi);
1133 DB((dbg, LEVEL_5, "fix inverted head should exch %N by %N (pos %d)\n", phi, new_phi, pos ));
1137 for_each_phi_safe(get_Block_phis(loop_head), phi, next) {
1138 DB((dbg, LEVEL_5, "fix inverted exch phi %N by %N\n", phi, phis[pos]));
1139 if (phis[pos] != phi)
1140 exchange(phi, phis[pos++]);
1145 DB((dbg, LEVEL_5, "fix inverted head exch head block %N by %N\n", loop_head, new_head));
1146 exchange(loop_head, new_head);
1149 /* Does the loop inversion. */
1150 static void inversion_walk(ir_graph *irg, entry_edge *head_entries)
1155 * The order of rewiring bottom-up is crucial.
1156 * Any change of the order leads to lost information that would be needed later.
1159 ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED);
1161 /* 1. clone condition chain */
1162 inc_irg_visited(irg);
1164 for (i = 0; i < ARR_LEN(head_entries); ++i) {
1165 entry_edge entry = head_entries[i];
1166 ir_node *pred = get_irn_n(entry.node, entry.pos);
1168 DB((dbg, LEVEL_5, "\nInit walk block %N\n", pred));
1170 copy_walk(pred, is_nodes_block_marked, cur_loop);
1173 ir_free_resources(irg, IR_RESOURCE_IRN_VISITED);
1175 /* 2. Extends the head control flow successors ins
1176 * with the definitions of the copied head node. */
1177 for (i = 0; i < ARR_LEN(head_entries); ++i) {
1178 entry_edge head_out = head_entries[i];
1180 if (is_Block(head_out.node))
1181 extend_ins_by_copy(head_out.node, head_out.pos);
1184 /* 3. construct_ssa for users of definitions in the condition chain,
1185 * as there is now a second definition. */
1186 for (i = 0; i < ARR_LEN(head_entries); ++i) {
1187 entry_edge head_out = head_entries[i];
1189 /* Ignore keepalives */
1190 if (is_End(head_out.node))
1193 /* Construct ssa for assignments in the condition chain. */
1194 if (!is_Block(head_out.node)) {
1195 ir_node *pred, *cppred, *block, *cpblock;
1197 pred = head_out.pred;
1198 cppred = get_inversion_copy(pred);
1199 block = get_nodes_block(pred);
1200 cpblock = get_nodes_block(cppred);
1201 construct_ssa(block, pred, cpblock, cppred);
1206 * If there is an assignment in the condition chain
1207 * with a user also in the condition chain,
1208 * the dominance frontier is in the new loop head.
1209 * The dataflow loop is completely in the condition chain.
1214 * Phi_cp | | copied condition chain
1218 * Phi* | | new loop head with newly created phi.
1220 * Phi | | original, inverted condition chain
1225 for (i = 0; i < ARR_LEN(head_df_loop); ++i) {
1226 entry_edge head_out = head_df_loop[i];
1228 /* Construct ssa for assignments in the condition chain. */
1229 ir_node *pred, *cppred, *block, *cpblock;
1231 pred = head_out.pred;
1232 cppred = get_inversion_copy(pred);
1233 assert(cppred && pred);
1234 block = get_nodes_block(pred);
1235 cpblock = get_nodes_block(cppred);
1236 construct_ssa(block, pred, cpblock, cppred);
1239 /* 4. Remove the ins which are no backedges from the original condition chain
1240 * as the cc is now subsequent to the body. */
1241 fix_head_inversion();
1243 /* 5. Remove the backedges of the copied condition chain,
1244 * because it is going to be the new 'head' in advance to the loop. */
1245 fix_copy_inversion();
1249 /* Performs loop inversion of cur_loop if possible and reasonable. */
1250 static void loop_inversion(ir_graph *irg)
1253 unsigned max_loop_nodes = opt_params.max_loop_size;
1254 unsigned max_loop_nodes_adapted;
1255 int depth_adaption = opt_params.depth_adaption;
1257 bool do_inversion = true;
1259 /* Depth of 0 is the procedure and 1 a topmost loop. */
1260 loop_depth = get_loop_depth(cur_loop) - 1;
1262 /* Calculating in per mil. */
1263 max_loop_nodes_adapted = get_max_nodes_adapted(loop_depth);
1265 DB((dbg, LEVEL_1, "max_nodes: %d\nmax_nodes_adapted %d at depth of %d (adaption %d)\n",
1266 max_loop_nodes, max_loop_nodes_adapted, loop_depth, depth_adaption));
1268 if (loop_info.nodes == 0)
1271 if (loop_info.nodes > max_loop_nodes) {
1272 /* Only for stats */
1273 DB((dbg, LEVEL_1, "Nodes %d > allowed nodes %d\n",
1274 loop_info.nodes, loop_depth, max_loop_nodes));
1277 /* Adaption might change it */
1280 /* Limit processing to loops smaller than given parameter. */
1281 if (loop_info.nodes > max_loop_nodes_adapted) {
1282 DB((dbg, LEVEL_1, "Nodes %d > allowed nodes (depth %d adapted) %d\n",
1283 loop_info.nodes, loop_depth, max_loop_nodes_adapted));
1284 ++stats.too_large_adapted;
1288 if (loop_info.calls > opt_params.allowed_calls) {
1289 DB((dbg, LEVEL_1, "Calls %d > allowed calls %d\n",
1290 loop_info.calls, opt_params.allowed_calls));
1291 ++stats.calls_limit;
1295 /*inversion_head_node_limit = INT_MAX;*/
1296 ir_reserve_resources(irg, IR_RESOURCE_BLOCK_MARK);
1298 /* Reset block marks.
1299 * We use block marks to flag blocks of the original condition chain. */
1300 irg_walk_graph(irg, reset_block_mark, NULL, NULL);
1302 /*loop_info.blocks = get_loop_n_blocks(cur_loop);*/
1303 cond_chain_entries = NEW_ARR_F(entry_edge, 0);
1304 head_df_loop = NEW_ARR_F(entry_edge, 0);
1306 /*head_inversion_node_count = 0;*/
1307 inversion_blocks_in_cc = 0;
1309 /* Use phase to keep copy of nodes from the condition chain. */
1310 ir_nodemap_init(&map, irg);
1311 obstack_init(&obst);
1313 /* Search for condition chains and temporarily save the blocks in an array. */
1314 cc_blocks = NEW_ARR_F(ir_node *, 0);
1315 inc_irg_visited(irg);
1316 find_condition_chain(loop_head);
1318 unmark_not_allowed_cc_blocks();
1319 DEL_ARR_F(cc_blocks);
1321 /* Condition chain too large.
1322 * Loop should better be small enough to fit into the cache. */
1323 /* TODO Of course, we should take a small enough cc in the first place,
1324 * which is not that simple. (bin packing) */
1325 if (loop_info.cc_size > opt_params.max_cc_size) {
1326 ++stats.cc_limit_reached;
1328 do_inversion = false;
1330 /* Unmark cc blocks except the head.
1331 * Invert head only for possible unrolling. */
1335 /* We also catch endless loops here,
1336 * because they do not have a condition chain. */
1337 if (inversion_blocks_in_cc < 1) {
1338 do_inversion = false;
1340 "Loop contains %d (less than 1) invertible blocks => No Inversion done.\n",
1341 inversion_blocks_in_cc));
1345 cur_head_outs = NEW_ARR_F(entry_edge, 0);
1347 /* Get all edges pointing into the condition chain. */
1348 irg_walk_graph(irg, get_head_outs, NULL, NULL);
1350 /* Do the inversion */
1351 inversion_walk(irg, cur_head_outs);
1353 DEL_ARR_F(cur_head_outs);
1355 /* Duplicated blocks changed doms */
1356 clear_irg_state(irg, IR_GRAPH_STATE_CONSISTENT_DOMINANCE
1357 | IR_GRAPH_STATE_CONSISTENT_LOOPINFO);
1363 obstack_free(&obst, NULL);
1364 ir_nodemap_destroy(&map);
1365 DEL_ARR_F(cond_chain_entries);
1366 DEL_ARR_F(head_df_loop);
1368 ir_free_resources(irg, IR_RESOURCE_BLOCK_MARK);
1371 /* Fix the original loop_heads ins for invariant unrolling case. */
1372 static void unrolling_fix_loop_head_inv(void)
1376 ir_node *proj = new_Proj(loop_info.duff_cond, mode_X, 0);
1377 ir_node *head_pred = get_irn_n(loop_head, loop_info.be_src_pos);
1378 ir_node *loop_condition = get_unroll_copy(head_pred, unroll_nr - 1);
1380 /* Original loop_heads ins are:
1381 * duff block and the own backedge */
1383 ins[0] = loop_condition;
1385 set_irn_in(loop_head, 2, ins);
1386 DB((dbg, LEVEL_4, "Rewire ins of block loophead %N to pred %N and duffs entry %N \n" , loop_head, ins[0], ins[1]));
1388 for_each_phi(loop_head, phi) {
1389 ir_node *pred = get_irn_n(phi, loop_info.be_src_pos);
1390 /* TODO we think it is a phi, but for Mergesort it is not the case.*/
1392 ir_node *last_pred = get_unroll_copy(pred, unroll_nr - 1);
1395 ins[1] = (ir_node*)get_irn_link(phi);
1396 set_irn_in(phi, 2, ins);
1397 DB((dbg, LEVEL_4, "Rewire ins of loophead phi %N to pred %N and duffs entry %N \n" , phi, ins[0], ins[1]));
1401 /* Removes previously created phis with only 1 in. */
1402 static void correct_phis(ir_node *node, void *env)
1406 if (is_Phi(node) && get_irn_arity(node) == 1) {
1410 in[0] = get_irn_n(node, 0);
1412 exch = new_rd_Phi(get_irn_dbg_info(node),
1413 get_nodes_block(node), 1, in,
1414 get_irn_mode(node));
1416 exchange(node, exch);
1420 /* Unrolling: Rewire floating copies. */
1421 static void place_copies(int copies)
1423 ir_node *loophead = loop_head;
1426 int be_src_pos = loop_info.be_src_pos;
1428 /* Serialize loops by fixing their head ins.
1429 * Processed are the copies.
1430 * The original loop is done after that, to keep backedge infos. */
1431 for (c = 0; c < copies; ++c) {
1432 ir_node *upper = get_unroll_copy(loophead, c);
1433 ir_node *lower = get_unroll_copy(loophead, c + 1);
1435 ir_node *topmost_be_block = get_nodes_block(get_irn_n(loophead, be_src_pos));
1437 /* Important: get the preds first and then their copy. */
1438 ir_node *upper_be_block = get_unroll_copy(topmost_be_block, c);
1439 ir_node *new_jmp = new_r_Jmp(upper_be_block);
1440 DB((dbg, LEVEL_5, " place_copies upper %N lower %N\n", upper, lower));
1442 DB((dbg, LEVEL_5, "topmost be block %N \n", topmost_be_block));
1444 if (loop_info.unroll_kind == constant) {
1447 set_irn_in(lower, 1, ins);
1449 for_each_phi(loophead, phi) {
1450 ir_node *topmost_def = get_irn_n(phi, be_src_pos);
1451 ir_node *upper_def = get_unroll_copy(topmost_def, c);
1452 ir_node *lower_phi = get_unroll_copy(phi, c + 1);
1454 /* It is possible, that the value used
1455 * in the OWN backedge path is NOT defined in this loop. */
1456 if (is_in_loop(topmost_def))
1459 ins[0] = topmost_def;
1461 set_irn_in(lower_phi, 1, ins);
1462 /* Need to replace phis with 1 in later. */
1465 /* Invariant case */
1466 /* Every node has 2 ins. One from the duff blocks
1467 * and one from the previously unrolled loop. */
1469 /* Calculate corresponding projection of mod result for this copy c */
1470 ir_node *proj = new_Proj(loop_info.duff_cond, mode_X, unroll_nr - c - 1);
1471 DB((dbg, LEVEL_4, "New duff proj %N\n" , proj));
1475 set_irn_in(lower, 2, ins);
1476 DB((dbg, LEVEL_4, "Rewire ins of Block %N to pred %N and duffs entry %N \n" , lower, ins[0], ins[1]));
1478 for_each_phi(loophead, phi) {
1479 ir_node *topmost_phi_pred = get_irn_n(phi, be_src_pos);
1480 ir_node *upper_phi_pred;
1484 lower_phi = get_unroll_copy(phi, c + 1);
1485 duff_phi = (ir_node*)get_irn_link(phi);
1486 DB((dbg, LEVEL_4, "DD Link of %N is %N\n" , phi, duff_phi));
1489 if (is_in_loop(topmost_phi_pred)) {
1490 upper_phi_pred = get_unroll_copy(topmost_phi_pred, c);
1492 upper_phi_pred = topmost_phi_pred;
1495 ins[0] = upper_phi_pred;
1497 set_irn_in(lower_phi, 2, ins);
1498 DB((dbg, LEVEL_4, "Rewire ins of %N to pred %N and duffs entry %N \n" , lower_phi, ins[0], ins[1]));
1503 /* Reconnect last copy. */
1504 for (i = 0; i < ARR_LEN(loop_entries); ++i) {
1505 entry_edge edge = loop_entries[i];
1506 /* Last copy is at the bottom */
1507 ir_node *new_pred = get_unroll_copy(edge.pred, copies);
1508 set_irn_n(edge.node, edge.pos, new_pred);
1511 /* Fix original loops head.
1512 * Done in the end, as ins and be info were needed before. */
1513 if (loop_info.unroll_kind == constant) {
1515 ir_node *head_pred = get_irn_n(loop_head, be_src_pos);
1516 ir_node *loop_condition = get_unroll_copy(head_pred, unroll_nr - 1);
1518 set_irn_n(loop_head, loop_info.be_src_pos, loop_condition);
1520 for_each_phi(loop_head, phi) {
1521 ir_node *pred = get_irn_n(phi, be_src_pos);
1524 /* It is possible, that the value used
1525 * in the OWN backedge path is NOT assigned in this loop. */
1526 if (is_in_loop(pred))
1527 last_pred = get_unroll_copy(pred, copies);
1530 set_irn_n(phi, be_src_pos, last_pred);
1534 unrolling_fix_loop_head_inv();
1538 /* Copies the cur_loop several times. */
1539 static void copy_loop(entry_edge *cur_loop_outs, int copies)
1543 ir_reserve_resources(current_ir_graph, IR_RESOURCE_IRN_VISITED);
1545 for (c = 0; c < copies; ++c) {
1548 inc_irg_visited(current_ir_graph);
1550 DB((dbg, LEVEL_5, " ### Copy_loop copy nr: %d ###\n", c));
1551 for (i = 0; i < ARR_LEN(cur_loop_outs); ++i) {
1552 entry_edge entry = cur_loop_outs[i];
1553 ir_node *pred = get_irn_n(entry.node, entry.pos);
1555 copy_walk_n(pred, is_in_loop, c + 1);
1559 ir_free_resources(current_ir_graph, IR_RESOURCE_IRN_VISITED);
1563 /* Creates a new phi from the given phi node omitting own bes,
1564 * using be_block as supplier of backedge informations. */
1565 static ir_node *clone_phis_sans_bes(ir_node *phi, ir_node *be_block, ir_node *dest_block)
1568 int arity = get_irn_arity(phi);
1572 assert(get_irn_arity(phi) == get_irn_arity(be_block));
1573 assert(is_Phi(phi));
1575 ins = NEW_ARR_F(ir_node *, arity);
1576 for (i = 0; i < arity; ++i) {
1577 if (! is_own_backedge(be_block, i)) {
1578 ins[c] = get_irn_n(phi, i);
1583 newphi = new_r_Phi(dest_block, c, ins, get_irn_mode(phi));
1585 set_irn_link(phi, newphi);
1586 DB((dbg, LEVEL_4, "Linking for duffs device %N to %N\n", phi, newphi));
1591 /* Creates a new block from the given block node omitting own bes,
1592 * using be_block as supplier of backedge informations. */
1593 static ir_node *clone_block_sans_bes(ir_node *node, ir_node *be_block)
1595 int arity = get_irn_arity(node);
1599 assert(get_irn_arity(node) == get_irn_arity(be_block));
1600 assert(is_Block(node));
1602 NEW_ARR_A(ir_node *, ins, arity);
1603 for (i = 0; i < arity; ++i) {
1604 if (! is_own_backedge(be_block, i)) {
1605 ins[c] = get_irn_n(node, i);
1610 return new_Block(c, ins);
1613 /* Creates a structure to calculate absolute value of node op.
1614 * Returns mux node with absolute value. */
1615 static ir_node *new_Abs(ir_node *op, ir_mode *mode)
1617 ir_graph *irg = get_irn_irg(op);
1618 ir_node *block = get_nodes_block(op);
1619 ir_node *zero = new_r_Const(irg, get_mode_null(mode));
1620 ir_node *cmp = new_r_Cmp(block, op, zero, ir_relation_less);
1621 ir_node *minus_op = new_r_Minus(block, op, mode);
1622 ir_node *mux = new_r_Mux(block, cmp, op, minus_op, mode);
1628 /* Creates blocks for duffs device, using previously obtained
1629 * informations about the iv.
1631 static void create_duffs_block(void)
1635 ir_node *block1, *count_block, *duff_block;
1636 ir_node *ems, *ems_mod, *ems_div, *ems_mod_proj, *cmp_null,
1637 *ems_mode_cond, *x_true, *x_false, *const_null;
1638 ir_node *true_val, *false_val;
1641 ir_node *duff_mod, *proj, *cond;
1643 ir_node *count, *correction, *unroll_c;
1644 ir_node *cmp_bad_count, *good_count, *bad_count, *count_phi, *bad_count_neg;
1647 mode = get_irn_mode(loop_info.end_val);
1648 const_null = new_Const(get_mode_null(mode));
1651 * 1. Calculate first approach to count.
1652 * Condition: (end - start) % step == 0 */
1653 block1 = clone_block_sans_bes(loop_head, loop_head);
1654 DB((dbg, LEVEL_4, "Duff block 1 %N\n", block1));
1656 /* Create loop entry phis in first duff block
1657 * as it becomes the loops preheader */
1658 for_each_phi(loop_head, phi) {
1659 /* Returns phis pred if phi would have arity 1*/
1660 ir_node *new_phi = clone_phis_sans_bes(phi, loop_head, block1);
1662 DB((dbg, LEVEL_4, "HEAD %N phi %N\n", loop_head, phi));
1663 DB((dbg, LEVEL_4, "BLOCK1 %N phi %N\n", block1, new_phi));
1666 ems = new_r_Sub(block1, loop_info.end_val, loop_info.start_val,
1667 get_irn_mode(loop_info.end_val));
1668 DB((dbg, LEVEL_4, "BLOCK1 sub %N\n", ems));
1671 ems = new_Sub(loop_info.end_val, loop_info.start_val,
1672 get_irn_mode(loop_info.end_val));
1674 DB((dbg, LEVEL_4, "mod ins %N %N\n", ems, loop_info.step));
1675 ems_mod = new_r_Mod(block1,
1680 op_pin_state_pinned);
1681 ems_div = new_r_Div(block1,
1686 op_pin_state_pinned);
1688 DB((dbg, LEVEL_4, "New module node %N\n", ems_mod));
1690 ems_mod_proj = new_r_Proj(ems_mod, mode_Iu, pn_Mod_res);
1691 cmp_null = new_r_Cmp(block1, ems_mod_proj, const_null, ir_relation_less);
1692 ems_mode_cond = new_r_Cond(block1, cmp_null);
1694 /* ems % step == 0 */
1695 x_true = new_r_Proj(ems_mode_cond, mode_X, pn_Cond_true);
1696 /* ems % step != 0 */
1697 x_false = new_r_Proj(ems_mode_cond, mode_X, pn_Cond_false);
1700 * Assures, duffs device receives a valid count.
1702 * decreasing: count < 0
1703 * increasing: count > 0
1708 count_block = new_Block(2, ins);
1709 DB((dbg, LEVEL_4, "Duff block 2 %N\n", count_block));
1712 /* Increase loop-taken-count depending on the loop condition
1713 * uses the latest iv to compare to. */
1714 if (loop_info.latest_value == 1) {
1715 /* ems % step == 0 : +0 */
1716 true_val = new_Const(get_mode_null(mode));
1717 /* ems % step != 0 : +1 */
1718 false_val = new_Const(get_mode_one(mode));
1720 ir_tarval *tv_two = new_tarval_from_long(2, mode);
1721 /* ems % step == 0 : +1 */
1722 true_val = new_Const(get_mode_one(mode));
1723 /* ems % step != 0 : +2 */
1724 false_val = new_Const(tv_two);
1730 correction = new_r_Phi(count_block, 2, ins, mode);
1732 count = new_r_Proj(ems_div, mode, pn_Div_res);
1734 /* (end - start) / step + correction */
1735 count = new_Add(count, correction, mode);
1737 /* We preconditioned the loop to be tail-controlled.
1738 * So, if count is something 'wrong' like 0,
1739 * negative/positive (depending on step direction),
1740 * we may take the loop once (tail-contr.) and leave it
1741 * to the existing condition, to break; */
1743 /* Depending on step direction, we have to check for > or < 0 */
1744 if (loop_info.decreasing == 1) {
1745 cmp_bad_count = new_r_Cmp(count_block, count, const_null,
1748 cmp_bad_count = new_r_Cmp(count_block, count, const_null,
1749 ir_relation_greater);
1752 bad_count_neg = new_r_Cond(count_block, cmp_bad_count);
1753 good_count = new_Proj(bad_count_neg, mode_X, pn_Cond_true);
1754 bad_count = new_Proj(ems_mode_cond, mode_X, pn_Cond_false);
1757 * Contains module to decide which loop to start from. */
1759 ins[0] = good_count;
1761 duff_block = new_Block(2, ins);
1762 DB((dbg, LEVEL_4, "Duff block 3 %N\n", duff_block));
1764 /* Get absolute value */
1765 ins[0] = new_Abs(count, mode);
1766 /* Manually feed the aforementioned count = 1 (bad case)*/
1767 ins[1] = new_Const(get_mode_one(mode));
1768 count_phi = new_r_Phi(duff_block, 2, ins, mode);
1770 unroll_c = new_Const(new_tarval_from_long((long)unroll_nr, mode));
1772 /* count % unroll_nr */
1773 duff_mod = new_r_Mod(duff_block,
1778 op_pin_state_pinned);
1781 proj = new_Proj(duff_mod, mode, pn_Mod_res);
1782 /* condition does NOT create itself in the block of the proj! */
1783 cond = new_r_Cond(duff_block, proj);
1785 loop_info.duff_cond = cond;
1788 /* Returns 1 if given node is not in loop,
1789 * or if it is a phi of the loop head with only loop invariant defs.
1791 static unsigned is_loop_invariant_def(ir_node *node)
1795 if (! is_in_loop(node)) {
1796 DB((dbg, LEVEL_4, "Not in loop %N\n", node));
1797 /* || is_Const(node) || is_SymConst(node)) {*/
1801 /* If this is a phi of the loophead shared by more than 1 loop,
1802 * we need to check if all defs are not in the loop. */
1805 block = get_nodes_block(node);
1807 /* To prevent unexpected situations. */
1808 if (block != loop_head) {
1812 for (i = 0; i < get_irn_arity(node); ++i) {
1813 /* Check if all bes are just loopbacks. */
1814 if (is_own_backedge(block, i) && get_irn_n(node, i) != node)
1817 DB((dbg, LEVEL_4, "invar %N\n", node));
1820 DB((dbg, LEVEL_4, "Not invar %N\n", node));
1825 /* Returns 1 if one pred of node is invariant and the other is not.
1826 * invar_pred and other are set analogously. */
1827 static unsigned get_invariant_pred(ir_node *node, ir_node **invar_pred, ir_node **other)
1829 ir_node *pred0 = get_irn_n(node, 0);
1830 ir_node *pred1 = get_irn_n(node, 1);
1835 if (is_loop_invariant_def(pred0)) {
1836 DB((dbg, LEVEL_4, "pred0 invar %N\n", pred0));
1837 *invar_pred = pred0;
1841 if (is_loop_invariant_def(pred1)) {
1842 DB((dbg, LEVEL_4, "pred1 invar %N\n", pred1));
1844 if (*invar_pred != NULL) {
1845 /* RETURN. We do not want both preds to be invariant. */
1850 *invar_pred = pred1;
1853 DB((dbg, LEVEL_4, "pred1 not invar %N\n", pred1));
1855 if (*invar_pred != NULL)
1862 /* Starts from a phi that may belong to an iv.
1863 * If an add forms a loop with iteration_phi,
1864 * and add uses a constant, 1 is returned
1865 * and 'start' as well as 'add' are sane. */
1866 static unsigned get_start_and_add(ir_node *iteration_phi, unrolling_kind_flag role)
1869 ir_node *found_add = loop_info.add;
1870 int arity = get_irn_arity(iteration_phi);
1872 DB((dbg, LEVEL_4, "Find start and add from %N\n", iteration_phi));
1874 for (i = 0; i < arity; ++i) {
1876 /* Find start_val which needs to be pred of the iteration_phi.
1877 * If start_val already known, sanity check. */
1878 if (!is_backedge(get_nodes_block(loop_info.iteration_phi), i)) {
1879 ir_node *found_start_val = get_irn_n(loop_info.iteration_phi, i);
1881 DB((dbg, LEVEL_4, "found_start_val %N\n", found_start_val));
1883 /* We already found a start_val it has to be always the same. */
1884 if (loop_info.start_val && found_start_val != loop_info.start_val)
1887 if ((role == constant) && !(is_SymConst(found_start_val) || is_Const(found_start_val)))
1889 else if((role == constant) && !(is_loop_invariant_def(found_start_val)))
1892 loop_info.start_val = found_start_val;
1895 /* The phi has to be in the loop head.
1896 * Follow all own backedges. Every value supplied from these preds of the phi
1897 * needs to origin from the same add. */
1898 if (is_own_backedge(get_nodes_block(loop_info.iteration_phi), i)) {
1899 ir_node *new_found = get_irn_n(loop_info.iteration_phi,i);
1901 DB((dbg, LEVEL_4, "is add? %N\n", new_found));
1903 if (! (is_Add(new_found) || is_Sub(new_found)) || (found_add && found_add != new_found))
1906 found_add = new_found;
1910 loop_info.add = found_add;
1916 /* Returns 1 if one pred of node is a const value and the other is not.
1917 * const_pred and other are set analogously. */
1918 static unsigned get_const_pred(ir_node *node, ir_node **const_pred, ir_node **other)
1920 ir_node *pred0 = get_irn_n(node, 0);
1921 ir_node *pred1 = get_irn_n(node, 1);
1923 DB((dbg, LEVEL_4, "Checking for constant pred of %N\n", node));
1928 /*DB((dbg, LEVEL_4, "is %N const\n", pred0));*/
1929 if (is_Const(pred0) || is_SymConst(pred0)) {
1930 *const_pred = pred0;
1934 /*DB((dbg, LEVEL_4, "is %N const\n", pred1));*/
1935 if (is_Const(pred1) || is_SymConst(pred1)) {
1936 if (*const_pred != NULL) {
1937 /* RETURN. We do not want both preds to be constant. */
1942 *const_pred = pred1;
1945 if (*const_pred == NULL)
1951 /* Returns 1 if loop exits within 2 steps of the iv.
1952 * Norm_proj means we do not exit the loop.*/
1953 static unsigned simulate_next(ir_tarval **count_tar,
1954 ir_tarval *stepped, ir_tarval *step_tar, ir_tarval *end_tar,
1955 ir_relation norm_proj)
1959 DB((dbg, LEVEL_4, "Loop taken if (stepped)%ld %s (end)%ld ",
1960 get_tarval_long(stepped),
1961 get_relation_string((norm_proj)),
1962 get_tarval_long(end_tar)));
1963 DB((dbg, LEVEL_4, "comparing latest value %d\n", loop_info.latest_value));
1965 /* If current iv does not stay in the loop,
1966 * this run satisfied the exit condition. */
1967 if (! (tarval_cmp(stepped, end_tar) & norm_proj))
1970 DB((dbg, LEVEL_4, "Result: (stepped)%ld IS %s (end)%ld\n",
1971 get_tarval_long(stepped),
1972 get_relation_string(tarval_cmp(stepped, end_tar)),
1973 get_tarval_long(end_tar)));
1976 if (is_Add(loop_info.add))
1977 next = tarval_add(stepped, step_tar);
1980 next = tarval_sub(stepped, step_tar, get_irn_mode(loop_info.end_val));
1982 DB((dbg, LEVEL_4, "Loop taken if %ld %s %ld ",
1983 get_tarval_long(next),
1984 get_relation_string(norm_proj),
1985 get_tarval_long(end_tar)));
1986 DB((dbg, LEVEL_4, "comparing latest value %d\n", loop_info.latest_value));
1988 /* Increase steps. */
1989 *count_tar = tarval_add(*count_tar, get_tarval_one(get_tarval_mode(*count_tar)));
1991 /* Next has to fail the loop condition, or we will never exit. */
1992 if (! (tarval_cmp(next, end_tar) & norm_proj))
1998 /* Check if loop meets requirements for a 'simple loop':
1999 * - Exactly one cf out
2001 * - Max nodes after unrolling
2005 * Returns Projection of cmp node or NULL; */
2006 static ir_node *is_simple_loop(void)
2009 ir_node *loop_block, *exit_block, *projx, *cond, *cmp;
2011 /* Maximum of one condition, and no endless loops. */
2012 if (loop_info.cf_outs != 1)
2015 DB((dbg, LEVEL_4, "1 loop exit\n"));
2017 /* Calculate maximum unroll_nr keeping node count below limit. */
2018 loop_info.max_unroll = (int)((double)opt_params.max_unrolled_loop_size / (double)loop_info.nodes);
2019 if (loop_info.max_unroll < 2) {
2024 DB((dbg, LEVEL_4, "maximum unroll factor %u, to not exceed node limit \n",
2025 opt_params.max_unrolled_loop_size));
2027 arity = get_irn_arity(loop_head);
2028 /* RETURN if we have more than 1 be. */
2029 /* Get my backedges without alien bes. */
2031 for (i = 0; i < arity; ++i) {
2032 ir_node *pred = get_irn_n(loop_head, i);
2033 if (is_own_backedge(loop_head, i)) {
2035 /* Our simple loops may have only one backedge. */
2038 loop_block = get_nodes_block(pred);
2039 loop_info.be_src_pos = i;
2044 DB((dbg, LEVEL_4, "loop has 1 own backedge.\n"));
2046 exit_block = get_nodes_block(loop_info.cf_out.pred);
2047 /* The loop has to be tail-controlled.
2048 * This can be changed/improved,
2049 * but we would need a duff iv. */
2050 if (exit_block != loop_block)
2053 DB((dbg, LEVEL_4, "tail-controlled loop.\n"));
2055 /* find value on which loop exit depends */
2056 projx = loop_info.cf_out.pred;
2057 cond = get_irn_n(projx, 0);
2058 cmp = get_irn_n(cond, 0);
2063 DB((dbg, LEVEL_5, "projection is %s\n", get_relation_string(get_Proj_proj(projx))));
2065 switch(get_Proj_proj(projx)) {
2067 loop_info.exit_cond = 0;
2070 loop_info.exit_cond = 1;
2073 panic("Cond Proj_proj other than true/false");
2076 DB((dbg, LEVEL_4, "Valid Cmp.\n"));
2080 /* Returns 1 if all nodes are mode_Iu or mode_Is. */
2081 static unsigned are_mode_I(ir_node *n1, ir_node* n2, ir_node *n3)
2083 ir_mode *m1 = get_irn_mode(n1);
2084 ir_mode *m2 = get_irn_mode(n2);
2085 ir_mode *m3 = get_irn_mode(n3);
2087 if ((m1 == mode_Iu && m2 == mode_Iu && m3 == mode_Iu) ||
2088 (m1 == mode_Is && m2 == mode_Is && m3 == mode_Is))
2094 /* Checks if cur_loop is a simple tail-controlled counting loop
2095 * with start and end value loop invariant, step constant. */
2096 static unsigned get_unroll_decision_invariant(void)
2099 ir_node *projres, *loop_condition, *iteration_path;
2101 ir_tarval *step_tar;
2105 /* RETURN if loop is not 'simple' */
2106 projres = is_simple_loop();
2107 if (projres == NULL)
2110 /* Use a minimal size for the invariant unrolled loop,
2111 * as duffs device produces overhead */
2112 if (loop_info.nodes < opt_params.invar_unrolling_min_size)
2115 loop_condition = get_irn_n(projres, 0);
2117 success = get_invariant_pred(loop_condition, &loop_info.end_val, &iteration_path);
2118 DB((dbg, LEVEL_4, "pred invar %d\n", success));
2123 DB((dbg, LEVEL_4, "Invariant End_val %N, other %N\n", loop_info.end_val, iteration_path));
2125 /* We may find the add or the phi first.
2126 * Until now we only have end_val. */
2127 if (is_Add(iteration_path) || is_Sub(iteration_path)) {
2129 loop_info.add = iteration_path;
2130 DB((dbg, LEVEL_4, "Case 1: Got add %N (maybe not sane)\n", loop_info.add));
2132 /* Preds of the add should be step and the iteration_phi */
2133 success = get_const_pred(loop_info.add, &loop_info.step, &loop_info.iteration_phi);
2137 DB((dbg, LEVEL_4, "Got step %N\n", loop_info.step));
2139 if (! is_Phi(loop_info.iteration_phi))
2142 DB((dbg, LEVEL_4, "Got phi %N\n", loop_info.iteration_phi));
2145 * Does necessary sanity check of add, if it is already set. */
2146 success = get_start_and_add(loop_info.iteration_phi, invariant);
2150 DB((dbg, LEVEL_4, "Got start A %N\n", loop_info.start_val));
2152 } else if (is_Phi(iteration_path)) {
2153 ir_node *new_iteration_phi;
2155 loop_info.iteration_phi = iteration_path;
2156 DB((dbg, LEVEL_4, "Case 2: Got phi %N\n", loop_info.iteration_phi));
2158 /* Find start_val and add-node.
2159 * Does necessary sanity check of add, if it is already set. */
2160 success = get_start_and_add(loop_info.iteration_phi, invariant);
2164 DB((dbg, LEVEL_4, "Got start B %N\n", loop_info.start_val));
2165 DB((dbg, LEVEL_4, "Got add or sub %N\n", loop_info.add));
2167 success = get_const_pred(loop_info.add, &loop_info.step, &new_iteration_phi);
2171 DB((dbg, LEVEL_4, "Got step (B) %N\n", loop_info.step));
2173 if (loop_info.iteration_phi != new_iteration_phi)
2180 mode = get_irn_mode(loop_info.end_val);
2182 DB((dbg, LEVEL_4, "start %N, end %N, step %N\n",
2183 loop_info.start_val, loop_info.end_val, loop_info.step));
2185 if (mode != mode_Is && mode != mode_Iu)
2188 /* TODO necessary? */
2189 if (!are_mode_I(loop_info.start_val, loop_info.step, loop_info.end_val))
2192 DB((dbg, LEVEL_4, "mode integer\n"));
2194 step_tar = get_Const_tarval(loop_info.step);
2196 if (tarval_is_null(step_tar)) {
2197 /* TODO Might be worth a warning. */
2201 DB((dbg, LEVEL_4, "step is not 0\n"));
2203 create_duffs_block();
2205 return loop_info.max_unroll;
2208 /* Returns unroll factor,
2209 * given maximum unroll factor and number of loop passes. */
2210 static unsigned get_preferred_factor_constant(ir_tarval *count_tar)
2212 ir_tarval *tar_6, *tar_5, *tar_4, *tar_3, *tar_2;
2214 ir_mode *mode = get_irn_mode(loop_info.end_val);
2216 tar_6 = new_tarval_from_long(6, mode);
2217 tar_5 = new_tarval_from_long(5, mode);
2218 tar_4 = new_tarval_from_long(4, mode);
2219 tar_3 = new_tarval_from_long(3, mode);
2220 tar_2 = new_tarval_from_long(2, mode);
2222 /* loop passes % {6, 5, 4, 3, 2} == 0 */
2223 if (tarval_is_null(tarval_mod(count_tar, tar_6)))
2225 else if (tarval_is_null(tarval_mod(count_tar, tar_5)))
2227 else if (tarval_is_null(tarval_mod(count_tar, tar_4)))
2229 else if (tarval_is_null(tarval_mod(count_tar, tar_3)))
2231 else if (tarval_is_null(tarval_mod(count_tar, tar_2)))
2234 /* gcd(max_unroll, count_tar) */
2235 int a = loop_info.max_unroll;
2236 int b = (int)get_tarval_long(count_tar);
2239 DB((dbg, LEVEL_4, "gcd of max_unroll %d and count_tar %d: ", a, b));
2246 DB((dbg, LEVEL_4, "%d\n", a));
2250 DB((dbg, LEVEL_4, "preferred unroll factor %d\n", prefer));
2253 * If our preference is greater than the allowed unroll factor
2254 * we either might reduce the preferred factor and prevent a duffs device block,
2255 * or create a duffs device block, from which in this case (constants only)
2256 * we know the startloop at compiletime.
2257 * The latter yields the following graphs.
2258 * but for code generation we would want to use graph A.
2259 * The graphs are equivalent. So, we can only reduce the preferred factor.
2269 if (prefer <= loop_info.max_unroll)
2274 if (loop_info.max_unroll >= 3)
2276 else if (loop_info.max_unroll >= 2)
2282 if (loop_info.max_unroll >= 2)
2293 /* Check if cur_loop is a simple counting loop.
2294 * Start, step and end are constants.
2295 * TODO The whole constant case should use procedures similar to
2296 * the invariant case, as they are more versatile. */
2298 static unsigned get_unroll_decision_constant(void)
2300 ir_node *cmp, *iteration_path;
2301 unsigned success, is_latest_val;
2302 ir_tarval *start_tar, *end_tar, *step_tar, *diff_tar, *count_tar;
2304 ir_relation proj_proj, norm_proj;
2307 /* RETURN if loop is not 'simple' */
2308 cmp = is_simple_loop();
2312 /* One in of the loop condition needs to be loop invariant. => end_val
2313 * The other in is assigned by an add. => add
2314 * The add uses a loop invariant value => step
2315 * and a phi with a loop invariant start_val and the add node as ins.
2327 success = get_const_pred(cmp, &loop_info.end_val, &iteration_path);
2331 DB((dbg, LEVEL_4, "End_val %N, other %N\n", loop_info.end_val, iteration_path));
2333 /* We may find the add or the phi first.
2334 * Until now we only have end_val. */
2335 if (is_Add(iteration_path) || is_Sub(iteration_path)) {
2337 /* We test against the latest value of the iv. */
2340 loop_info.add = iteration_path;
2341 DB((dbg, LEVEL_4, "Case 2: Got add %N (maybe not sane)\n", loop_info.add));
2343 /* Preds of the add should be step and the iteration_phi */
2344 success = get_const_pred(loop_info.add, &loop_info.step, &loop_info.iteration_phi);
2348 DB((dbg, LEVEL_4, "Got step %N\n", loop_info.step));
2350 if (! is_Phi(loop_info.iteration_phi))
2353 DB((dbg, LEVEL_4, "Got phi %N\n", loop_info.iteration_phi));
2356 * Does necessary sanity check of add, if it is already set. */
2357 success = get_start_and_add(loop_info.iteration_phi, constant);
2361 DB((dbg, LEVEL_4, "Got start %N\n", loop_info.start_val));
2363 } else if (is_Phi(iteration_path)) {
2364 ir_node *new_iteration_phi;
2366 /* We compare with the value the iv had entering this run. */
2369 loop_info.iteration_phi = iteration_path;
2370 DB((dbg, LEVEL_4, "Case 1: Got phi %N \n", loop_info.iteration_phi));
2372 /* Find start_val and add-node.
2373 * Does necessary sanity check of add, if it is already set. */
2374 success = get_start_and_add(loop_info.iteration_phi, constant);
2378 DB((dbg, LEVEL_4, "Got start %N\n", loop_info.start_val));
2379 DB((dbg, LEVEL_4, "Got add or sub %N\n", loop_info.add));
2381 success = get_const_pred(loop_info.add, &loop_info.step, &new_iteration_phi);
2385 DB((dbg, LEVEL_4, "Got step %N\n", loop_info.step));
2387 if (loop_info.iteration_phi != new_iteration_phi)
2395 mode = get_irn_mode(loop_info.end_val);
2397 DB((dbg, LEVEL_4, "start %N, end %N, step %N\n",
2398 loop_info.start_val, loop_info.end_val, loop_info.step));
2400 if (mode != mode_Is && mode != mode_Iu)
2403 /* TODO necessary? */
2404 if (!are_mode_I(loop_info.start_val, loop_info.step, loop_info.end_val))
2407 DB((dbg, LEVEL_4, "mode integer\n"));
2409 end_tar = get_Const_tarval(loop_info.end_val);
2410 start_tar = get_Const_tarval(loop_info.start_val);
2411 step_tar = get_Const_tarval(loop_info.step);
2413 if (tarval_is_null(step_tar))
2414 /* TODO Might be worth a warning. */
2417 DB((dbg, LEVEL_4, "step is not 0\n"));
2419 if ((!tarval_is_negative(step_tar)) ^ (!is_Sub(loop_info.add)))
2420 loop_info.decreasing = 1;
2422 diff_tar = tarval_sub(end_tar, start_tar, mode);
2424 /* We need at least count_tar steps to be close to end_val, maybe more.
2425 * No way, that we have gone too many steps.
2426 * This represents the 'latest value'.
2427 * (If condition checks against latest value, is checked later) */
2428 count_tar = tarval_div(diff_tar, step_tar);
2430 /* Iv will not pass end_val (except overflows).
2431 * Nothing done, as it would yield to no advantage. */
2432 if (tarval_is_negative(count_tar)) {
2433 DB((dbg, LEVEL_4, "Loop is endless or never taken."));
2434 /* TODO Might be worth a warning. */
2438 ++stats.u_simple_counting_loop;
2440 loop_info.latest_value = is_latest_val;
2443 if (! is_simple_counting_loop(&count_tar))
2447 /* stepped can be negative, if step < 0 */
2448 stepped = tarval_mul(count_tar, step_tar);
2450 /* step as close to end_val as possible, */
2451 /* |stepped| <= |end_tar|, and dist(stepped, end_tar) is smaller than a step. */
2452 if (is_Sub(loop_info.add))
2453 stepped = tarval_sub(start_tar, stepped, mode_Is);
2455 stepped = tarval_add(start_tar, stepped);
2457 DB((dbg, LEVEL_4, "stepped to %ld\n", get_tarval_long(stepped)));
2459 proj_proj = get_Cmp_relation(cmp);
2460 /* Assure that norm_proj is the stay-in-loop case. */
2461 if (loop_info.exit_cond == 1)
2462 norm_proj = get_negated_relation(proj_proj);
2464 norm_proj = proj_proj;
2466 DB((dbg, LEVEL_4, "normalized projection %s\n", get_relation_string(norm_proj)));
2467 /* Executed at most once (stay in counting loop if a Eq b) */
2468 if (norm_proj == ir_relation_equal)
2469 /* TODO Might be worth a warning. */
2472 /* calculates next values and increases count_tar according to it */
2473 success = simulate_next(&count_tar, stepped, step_tar, end_tar, norm_proj);
2477 /* We run loop once more, if we compare to the
2478 * not yet in-/decreased iv. */
2479 if (is_latest_val == 0) {
2480 DB((dbg, LEVEL_4, "condition uses not latest iv value\n"));
2481 count_tar = tarval_add(count_tar, get_tarval_one(mode));
2484 DB((dbg, LEVEL_4, "loop taken %ld times\n", get_tarval_long(count_tar)));
2486 /* Assure the loop is taken at least 1 time. */
2487 if (tarval_is_null(count_tar)) {
2488 /* TODO Might be worth a warning. */
2492 loop_info.count_tar = count_tar;
2493 return get_preferred_factor_constant(count_tar);
2499 static void unroll_loop(void)
2502 if (! (loop_info.nodes > 0))
2505 if (loop_info.nodes > opt_params.max_unrolled_loop_size) {
2506 DB((dbg, LEVEL_2, "Nodes %d > allowed nodes %d\n",
2507 loop_info.nodes, opt_params.max_unrolled_loop_size));
2512 if (loop_info.calls > 0) {
2513 DB((dbg, LEVEL_2, "Calls %d > allowed calls 0\n",
2515 ++stats.calls_limit;
2521 /* get_unroll_decision_constant and invariant are completely
2522 * independent for flexibility.
2523 * Some checks may be performed twice. */
2525 /* constant case? */
2526 if (opt_params.allow_const_unrolling)
2527 unroll_nr = get_unroll_decision_constant();
2528 if (unroll_nr > 1) {
2529 loop_info.unroll_kind = constant;
2532 /* invariant case? */
2533 if (opt_params.allow_invar_unrolling)
2534 unroll_nr = get_unroll_decision_invariant();
2536 loop_info.unroll_kind = invariant;
2539 DB((dbg, LEVEL_2, " *** Unrolling %d times ***\n", unroll_nr));
2541 if (unroll_nr > 1) {
2542 loop_entries = NEW_ARR_F(entry_edge, 0);
2545 irg_walk_graph(current_ir_graph, get_loop_entries, NULL, NULL);
2547 if (loop_info.unroll_kind == constant) {
2548 if ((int)get_tarval_long(loop_info.count_tar) == unroll_nr)
2549 loop_info.needs_backedge = 0;
2551 loop_info.needs_backedge = 1;
2553 loop_info.needs_backedge = 1;
2556 /* Use phase to keep copy of nodes from the condition chain. */
2557 ir_nodemap_init(&map, current_ir_graph);
2558 obstack_init(&obst);
2560 /* Copies the loop */
2561 copy_loop(loop_entries, unroll_nr - 1);
2563 /* Line up the floating copies. */
2564 place_copies(unroll_nr - 1);
2566 /* Remove phis with 1 in
2567 * If there were no nested phis, this would not be necessary.
2568 * Avoiding the creation in the first place
2569 * leads to complex special cases. */
2570 irg_walk_graph(current_ir_graph, correct_phis, NULL, NULL);
2572 if (loop_info.unroll_kind == constant)
2573 ++stats.constant_unroll;
2575 ++stats.invariant_unroll;
2577 clear_irg_state(current_ir_graph, IR_GRAPH_STATE_CONSISTENT_DOMINANCE);
2579 DEL_ARR_F(loop_entries);
2580 obstack_free(&obst, NULL);
2581 ir_nodemap_destroy(&map);
2586 /* Analyzes the loop, and checks if size is within allowed range.
2587 * Decides if loop will be processed. */
2588 static void init_analyze(ir_graph *irg, ir_loop *loop)
2593 loop_head_valid = true;
2595 /* Reset loop info */
2596 memset(&loop_info, 0, sizeof(loop_info_t));
2598 DB((dbg, LEVEL_1, " >>>> current loop %ld <<<\n",
2599 get_loop_loop_nr(loop)));
2601 /* Collect loop informations: head, node counts. */
2602 irg_walk_graph(irg, get_loop_info, NULL, NULL);
2604 /* RETURN if there is no valid head */
2605 if (!loop_head || !loop_head_valid) {
2606 DB((dbg, LEVEL_1, "No valid loop head. Nothing done.\n"));
2609 DB((dbg, LEVEL_1, "Loophead: %N\n", loop_head));
2612 if (loop_info.branches > opt_params.max_branches) {
2613 DB((dbg, LEVEL_1, "Branches %d > allowed branches %d\n",
2614 loop_info.branches, opt_params.max_branches));
2615 ++stats.calls_limit;
2620 case loop_op_inversion:
2621 loop_inversion(irg);
2624 case loop_op_unrolling:
2629 panic("Loop optimization not implemented.");
2631 DB((dbg, LEVEL_1, " <<<< end of loop with node %ld >>>>\n",
2632 get_loop_loop_nr(loop)));
2635 /* Find innermost loops and add them to loops. */
2636 static void find_innermost_loop(ir_loop *loop)
2638 bool had_sons = false;
2639 size_t n_elements = get_loop_n_elements(loop);
2642 for (e = 0; e < n_elements; ++e) {
2643 loop_element element = get_loop_element(loop, e);
2644 if (*element.kind == k_ir_loop) {
2645 find_innermost_loop(element.son);
2651 ARR_APP1(ir_loop*, loops, loop);
2655 static void set_loop_params(void)
2657 opt_params.max_loop_size = 100;
2658 opt_params.depth_adaption = -50;
2659 opt_params.count_phi = true;
2660 opt_params.count_proj = false;
2661 opt_params.allowed_calls = 0;
2663 opt_params.max_cc_size = 5;
2666 opt_params.allow_const_unrolling = true;
2667 opt_params.allow_invar_unrolling = false;
2669 opt_params.invar_unrolling_min_size = 20;
2670 opt_params.max_unrolled_loop_size = 400;
2671 opt_params.max_branches = 9999;
2674 /* Assure preconditions are met and go through all loops. */
2675 void loop_optimization(ir_graph *irg)
2683 /* Reset stats for this procedure */
2687 set_current_ir_graph(irg);
2689 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_PHI_LIST);
2690 collect_phiprojs(irg);
2692 loop = get_irg_loop(irg);
2694 loops = NEW_ARR_F(ir_loop *, 0);
2695 /* List all inner loops */
2696 n_elements = get_loop_n_elements(loop);
2697 for (i = 0; i < n_elements; ++i) {
2698 loop_element element = get_loop_element(loop, i);
2699 if (*element.kind != k_ir_loop)
2701 find_innermost_loop(element.son);
2704 /* Set all links to NULL */
2705 irg_walk_graph(irg, reset_link, NULL, NULL);
2707 for (i = 0; i < ARR_LEN(loops); ++i) {
2708 ir_loop *loop = loops[i];
2712 /* Analyze and handle loop */
2713 init_analyze(irg, loop);
2715 /* Copied blocks do not have their phi list yet */
2716 collect_phiprojs(irg);
2718 /* Set links to NULL
2719 * TODO Still necessary? */
2720 irg_walk_graph(irg, reset_link, NULL, NULL);
2726 ir_free_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_PHI_LIST);
2729 static ir_graph_state_t perform_loop_unrolling(ir_graph *irg)
2731 loop_op = loop_op_unrolling;
2732 loop_optimization(irg);
2736 static ir_graph_state_t perform_loop_inversion(ir_graph *irg)
2738 loop_op = loop_op_inversion;
2739 loop_optimization(irg);
2743 static ir_graph_state_t perform_loop_peeling(ir_graph *irg)
2745 loop_op = loop_op_peeling;
2746 loop_optimization(irg);
2750 static optdesc_t opt_unroll_loops = {
2752 IR_GRAPH_STATE_CONSISTENT_OUT_EDGES | IR_GRAPH_STATE_CONSISTENT_OUTS | IR_GRAPH_STATE_CONSISTENT_LOOPINFO,
2753 perform_loop_unrolling,
2756 static optdesc_t opt_invert_loops = {
2758 IR_GRAPH_STATE_CONSISTENT_OUT_EDGES | IR_GRAPH_STATE_CONSISTENT_OUTS | IR_GRAPH_STATE_CONSISTENT_LOOPINFO,
2759 perform_loop_inversion,
2762 static optdesc_t opt_peel_loops = {
2764 IR_GRAPH_STATE_CONSISTENT_OUT_EDGES | IR_GRAPH_STATE_CONSISTENT_OUTS | IR_GRAPH_STATE_CONSISTENT_LOOPINFO,
2765 perform_loop_peeling,
2768 void do_loop_unrolling(ir_graph *irg)
2769 { perform_irg_optimization(irg, &opt_unroll_loops); }
2771 void do_loop_inversion(ir_graph *irg)
2772 { perform_irg_optimization(irg, &opt_invert_loops); }
2774 void do_loop_peeling(ir_graph *irg)
2775 { perform_irg_optimization(irg, &opt_peel_loops); }
2777 ir_graph_pass_t *loop_inversion_pass(const char *name)
2779 return def_graph_pass(name ? name : "loop_inversion", do_loop_inversion);
2782 ir_graph_pass_t *loop_unroll_pass(const char *name)
2784 return def_graph_pass(name ? name : "loop_unroll", do_loop_unrolling);
2787 ir_graph_pass_t *loop_peeling_pass(const char *name)
2789 return def_graph_pass(name ? name : "loop_peeling", do_loop_peeling);
2792 void firm_init_loop_opt(void)
2794 FIRM_DBG_REGISTER(dbg, "firm.opt.loop");