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"
49 #include "irbackedge_t.h"
50 #include "irnodemap.h"
53 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
56 * Convenience macro for iterating over every phi node of the given block.
57 * Requires phi list per block.
59 #define for_each_phi(block, phi) \
60 for ((phi) = get_Block_phis( (block) ); (phi) ; (phi) = get_Phi_next((phi)))
62 #define for_each_phi_safe(head, phi, next) \
63 for ((phi) = (head), (next) = (head) ? get_Phi_next((head)) : NULL; \
64 (phi) ; (phi) = (next), (next) = (next) ? get_Phi_next((next)) : NULL)
66 /* Currently processed loop. */
67 static ir_loop *cur_loop;
69 /* Flag for kind of unrolling. */
73 } unrolling_kind_flag;
75 /* Condition for performing visiting a node during copy_walk. */
76 typedef bool walker_condition(const ir_node *);
78 /* Node and position of a predecessor. */
79 typedef struct entry_edge {
85 /* Node info for unrolling. */
86 typedef struct unrolling_node_info {
88 } unrolling_node_info;
90 /* Outs of the nodes head. */
91 static entry_edge *cur_head_outs;
93 /* Information about the loop head */
94 static ir_node *loop_head = NULL;
95 static bool loop_head_valid = true;
97 /* List of all inner loops, that are processed. */
98 static ir_loop **loops;
101 typedef struct loop_stats_t {
105 unsigned too_large_adapted;
106 unsigned cc_limit_reached;
107 unsigned calls_limit;
109 unsigned u_simple_counting_loop;
110 unsigned constant_unroll;
111 unsigned invariant_unroll;
116 static loop_stats_t stats;
118 /* Set stats to sero */
119 static void reset_stats(void)
121 memset(&stats, 0, sizeof(loop_stats_t));
125 static void print_stats(void)
127 DB((dbg, LEVEL_2, "---------------------------------------\n"));
128 DB((dbg, LEVEL_2, "loops : %d\n",stats.loops));
129 DB((dbg, LEVEL_2, "inverted : %d\n",stats.inverted));
130 DB((dbg, LEVEL_2, "too_large : %d\n",stats.too_large));
131 DB((dbg, LEVEL_2, "too_large_adapted : %d\n",stats.too_large_adapted));
132 DB((dbg, LEVEL_2, "cc_limit_reached : %d\n",stats.cc_limit_reached));
133 DB((dbg, LEVEL_2, "calls_limit : %d\n",stats.calls_limit));
134 DB((dbg, LEVEL_2, "u_simple_counting : %d\n",stats.u_simple_counting_loop));
135 DB((dbg, LEVEL_2, "constant_unroll : %d\n",stats.constant_unroll));
136 DB((dbg, LEVEL_2, "invariant_unroll : %d\n",stats.invariant_unroll));
137 DB((dbg, LEVEL_2, "=======================================\n"));
140 /* Commandline parameters */
141 typedef struct loop_opt_params_t {
142 unsigned max_loop_size; /* Maximum number of nodes [nodes]*/
143 int depth_adaption; /* Loop nest depth adaption [percent] */
144 unsigned allowed_calls; /* Number of calls allowed [number] */
145 bool count_phi; /* Count phi nodes */
146 bool count_proj; /* Count projections */
148 unsigned max_cc_size; /* Maximum condition chain size [nodes] */
149 unsigned max_branches;
151 unsigned max_unrolled_loop_size; /* [nodes] */
152 bool allow_const_unrolling;
153 bool allow_invar_unrolling;
154 unsigned invar_unrolling_min_size; /* [nodes] */
158 static loop_opt_params_t opt_params;
160 /* Loop analysis informations */
161 typedef struct loop_info_t {
162 unsigned nodes; /* node count */
163 unsigned ld_st; /* load and store nodes */
164 unsigned branches; /* number of conditions */
165 unsigned calls; /* number of calls */
166 unsigned cf_outs; /* number of cf edges which leave the loop */
167 entry_edge cf_out; /* single loop leaving cf edge */
168 int be_src_pos; /* position of the single own backedge in the head */
171 unsigned cc_size; /* nodes in the condition chain */
174 unsigned max_unroll; /* Number of unrolls satisfying max_loop_size */
175 unsigned exit_cond; /* 1 if condition==true exits the loop. */
176 unsigned latest_value:1; /* 1 if condition is checked against latest counter value */
177 unsigned needs_backedge:1; /* 0 if loop is completely unrolled */
178 unsigned decreasing:1; /* Step operation is_Sub, or step is<0 */
180 /* IV informations of a simple loop */
184 ir_node *iteration_phi;
187 ir_tarval *count_tar; /* Number of loop iterations */
189 ir_node *duff_cond; /* Duff mod */
190 unrolling_kind_flag unroll_kind; /* constant or invariant unrolling */
193 /* Information about the current loop */
194 static loop_info_t loop_info;
196 /* Outs of the condition chain (loop inversion). */
197 static ir_node **cc_blocks;
198 /* df/cf edges with def in the condition chain */
199 static entry_edge *cond_chain_entries;
200 /* Array of df loops found in the condition chain. */
201 static entry_edge *head_df_loop;
202 /* Number of blocks in cc */
203 static unsigned inversion_blocks_in_cc;
206 /* Cf/df edges leaving the loop.
207 * Called entries here, as they are used to enter the loop with walkers. */
208 static entry_edge *loop_entries;
209 /* Number of unrolls to perform */
210 static int unroll_nr;
211 /* Phase is used to keep copies of nodes. */
212 static ir_nodemap map;
213 static struct obstack obst;
215 /* Loop operations. */
216 typedef enum loop_op_t {
222 /* Saves which loop operation to do until after basic tests. */
223 static loop_op_t loop_op;
225 /* Returns the maximum nodes for the given nest depth */
226 static unsigned get_max_nodes_adapted(unsigned depth)
228 double perc = 100.0 + (double)opt_params.depth_adaption;
229 double factor = pow(perc / 100.0, depth);
231 return (int)((double)opt_params.max_loop_size * factor);
234 /* Reset nodes link. For use with a walker. */
235 static void reset_link(ir_node *node, void *env)
238 set_irn_link(node, NULL);
241 /* Returns 0 if the node or block is not in cur_loop. */
242 static bool is_in_loop(const ir_node *node)
244 return get_irn_loop(get_block_const(node)) == cur_loop;
247 /* Returns 0 if the given edge is not a backedge
248 * with its pred in the cur_loop. */
249 static bool is_own_backedge(const ir_node *n, int pos)
251 return is_backedge(n, pos) && is_in_loop(get_irn_n(n, pos));
254 /* Finds loop head and some loop_info as calls or else if necessary. */
255 static void get_loop_info(ir_node *node, void *env)
257 bool node_in_loop = is_in_loop(node);
261 /* collect some loop information */
263 if (is_Phi(node) && opt_params.count_phi)
265 else if (is_Proj(node) && opt_params.count_proj)
267 else if (!is_Confirm(node) && !is_Const(node) && !is_SymConst(node))
270 if (is_Load(node) || is_Store(node))
277 arity = get_irn_arity(node);
278 for (i = 0; i < arity; i++) {
279 ir_node *pred = get_irn_n(node, i);
280 bool pred_in_loop = is_in_loop(pred);
282 if (is_Block(node) && !node_in_loop && pred_in_loop) {
289 loop_info.cf_out = entry;
292 /* Find the loops head/the blocks with cfpred outside of the loop */
293 if (is_Block(node)) {
294 const ir_edge_t *edge;
297 /* Count innerloop branches */
298 foreach_out_edge_kind(node, edge, EDGE_KIND_BLOCK) {
299 ir_node *succ = get_edge_src_irn(edge);
300 if (is_Block(succ) && is_in_loop(succ))
304 ++loop_info.branches;
306 if (node_in_loop && !pred_in_loop && loop_head_valid) {
307 ir_node *cfgpred = get_Block_cfgpred(node, i);
309 if (!is_in_loop(cfgpred)) {
310 DB((dbg, LEVEL_5, "potential head %+F because inloop and pred %+F not inloop\n",
312 /* another head? We do not touch this. */
313 if (loop_head && loop_head != node) {
314 loop_head_valid = false;
324 /* Finds all edges with users outside of the loop
325 * and definition inside the loop. */
326 static void get_loop_entries(ir_node *node, void *env)
328 unsigned node_in_loop, pred_in_loop;
332 arity = get_irn_arity(node);
333 for (i = 0; i < arity; ++i) {
334 ir_node *pred = get_irn_n(node, i);
336 pred_in_loop = is_in_loop(pred);
337 node_in_loop = is_in_loop(node);
339 if (pred_in_loop && !node_in_loop) {
344 ARR_APP1(entry_edge, loop_entries, entry);
350 static ir_node *ssa_second_def;
351 static ir_node *ssa_second_def_block;
354 * Walks the graph bottom up, searching for definitions and creates phis.
356 static ir_node *search_def_and_create_phis(ir_node *block, ir_mode *mode, int first)
360 ir_graph *irg = get_irn_irg(block);
364 DB((dbg, LEVEL_5, "ssa search_def_and_create_phis: block %N\n", block));
366 /* Prevents creation of phi that would be bad anyway.
367 * Dead and bad blocks. */
368 if (get_irn_arity(block) < 1 || is_Bad(block)) {
369 DB((dbg, LEVEL_5, "ssa bad %N\n", block));
370 return new_r_Bad(irg, mode);
373 if (block == ssa_second_def_block && !first) {
374 DB((dbg, LEVEL_5, "ssa found second definition: use second def %N\n", ssa_second_def));
375 return ssa_second_def;
378 /* already processed this block? */
379 if (irn_visited(block)) {
380 ir_node *value = (ir_node *) get_irn_link(block);
381 DB((dbg, LEVEL_5, "ssa already visited: use linked %N\n", value));
385 assert(block != get_irg_start_block(irg));
387 /* a Block with only 1 predecessor needs no Phi */
388 n_cfgpreds = get_Block_n_cfgpreds(block);
389 if (n_cfgpreds == 1) {
390 ir_node *pred_block = get_Block_cfgpred_block(block, 0);
393 DB((dbg, LEVEL_5, "ssa 1 pred: walk pred %N\n", pred_block));
395 value = search_def_and_create_phis(pred_block, mode, 0);
396 set_irn_link(block, value);
397 mark_irn_visited(block);
402 /* create a new Phi */
403 NEW_ARR_A(ir_node*, in, n_cfgpreds);
404 for (i = 0; i < n_cfgpreds; ++i)
405 in[i] = new_r_Dummy(irg, mode);
407 phi = new_r_Phi(block, n_cfgpreds, in, mode);
408 /* Important: always keep block phi list up to date. */
409 add_Block_phi(block, phi);
410 DB((dbg, LEVEL_5, "ssa phi creation: link new phi %N to block %N\n", phi, block));
411 set_irn_link(block, phi);
412 mark_irn_visited(block);
414 /* set Phi predecessors */
415 for (i = 0; i < n_cfgpreds; ++i) {
417 ir_node *pred_block = get_Block_cfgpred_block(block, i);
418 assert(pred_block != NULL);
419 pred_val = search_def_and_create_phis(pred_block, mode, 0);
421 assert(pred_val != NULL);
423 DB((dbg, LEVEL_5, "ssa phi pred:phi %N, pred %N\n", phi, pred_val));
424 set_irn_n(phi, i, pred_val);
432 * Given a set of values this function constructs SSA-form for the users of the
433 * first value (the users are determined through the out-edges of the value).
434 * Works without using the dominance tree.
436 static void construct_ssa(ir_node *orig_block, ir_node *orig_val,
437 ir_node *second_block, ir_node *second_val)
441 const ir_edge_t *edge;
442 const ir_edge_t *next;
444 assert(orig_block && orig_val && second_block && second_val &&
445 "no parameter of construct_ssa may be NULL");
447 if (orig_val == second_val)
450 irg = get_irn_irg(orig_val);
452 ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED);
453 inc_irg_visited(irg);
455 mode = get_irn_mode(orig_val);
456 set_irn_link(orig_block, orig_val);
457 mark_irn_visited(orig_block);
459 ssa_second_def_block = second_block;
460 ssa_second_def = second_val;
462 /* Only fix the users of the first, i.e. the original node */
463 foreach_out_edge_safe(orig_val, edge, next) {
464 ir_node *user = get_edge_src_irn(edge);
465 int j = get_edge_src_pos(edge);
466 ir_node *user_block = get_nodes_block(user);
473 DB((dbg, LEVEL_5, "original user %N\n", user));
476 ir_node *pred_block = get_Block_cfgpred_block(user_block, j);
477 newval = search_def_and_create_phis(pred_block, mode, 1);
479 newval = search_def_and_create_phis(user_block, mode, 1);
481 if (newval != user && !is_Bad(newval))
482 set_irn_n(user, j, newval);
485 ir_free_resources(irg, IR_RESOURCE_IRN_VISITED);
489 /***** Unrolling Helper Functions *****/
491 /* Assign the copy with index nr to node n */
492 static void set_unroll_copy(ir_node *n, int nr, ir_node *cp)
494 unrolling_node_info *info;
495 assert(nr != 0 && "0 reserved");
497 info = (unrolling_node_info*)ir_nodemap_get(&map, n);
499 ir_node **arr = NEW_ARR_D(ir_node*, &obst, unroll_nr);
500 memset(arr, 0, unroll_nr * sizeof(ir_node*));
502 info = OALLOCZ(&obst, unrolling_node_info);
504 ir_nodemap_insert(&map, n, info);
509 info->copies[nr] = cp;
512 /* Returns a nodes copy if it exists, else NULL. */
513 static ir_node *get_unroll_copy(ir_node *n, int nr)
516 unrolling_node_info *info = (unrolling_node_info *)ir_nodemap_get(&map, n);
520 cp = info->copies[nr];
525 /***** Inversion Helper Functions *****/
527 /* Sets copy cp of node n. */
528 static void set_inversion_copy(ir_node *n, ir_node *cp)
530 ir_nodemap_insert(&map, n, cp);
533 /* Getter of copy of n for inversion */
534 static ir_node *get_inversion_copy(ir_node *n)
536 ir_node *cp = (ir_node *)ir_nodemap_get(&map, n);
540 /* Resets block mark for given node. For use with walker */
541 static void reset_block_mark(ir_node *node, void * env)
546 set_Block_mark(node, 0);
549 /* Returns mark of node, or its block if node is not a block.
550 * Used in this context to determine if node is in the condition chain. */
551 static bool is_nodes_block_marked(const ir_node* node)
553 return get_Block_mark(get_block_const(node));
556 /* Extends a nodes ins by node new.
557 * NOTE: This is slow if a node n needs to be extended more than once. */
558 static void extend_irn(ir_node *n, ir_node *newnode, bool new_is_backedge)
561 int arity = get_irn_arity(n);
562 int new_arity = arity + 1;
563 ir_node **ins = XMALLOCN(ir_node*, new_arity);
564 bool *bes = XMALLOCN(bool, new_arity);
567 /* Bes are important!
568 * Another way would be recreating the looptree,
569 * but after that we cannot distinguish already processed loops
570 * from not yet processed ones. */
572 for(i = 0; i < arity; ++i) {
573 bes[i] = is_backedge(n, i);
575 bes[i] = new_is_backedge;
578 for(i = 0; i < arity; ++i) {
579 ins[i] = get_irn_n(n, i);
583 set_irn_in(n, new_arity, ins);
587 for(i = 0; i < new_arity; ++i) {
596 /* Extends a block by a copy of its pred at pos,
597 * fixing also the phis in the same way. */
598 static void extend_ins_by_copy(ir_node *block, int pos)
603 assert(is_Block(block));
605 /* Extend block by copy of definition at pos */
606 pred = get_irn_n(block, pos);
607 new_in = get_inversion_copy(pred);
608 DB((dbg, LEVEL_5, "Extend block %N by %N cp of %N\n", block, new_in, pred));
609 extend_irn(block, new_in, false);
611 /* Extend block phis by copy of definition at pos */
612 for_each_phi(block, phi) {
615 pred = get_irn_n(phi, pos);
616 cp = get_inversion_copy(pred);
617 /* If the phis in is not in the condition chain (eg. a constant),
618 * there is no copy. */
624 DB((dbg, LEVEL_5, "Extend phi %N by %N cp of %N\n", phi, new_in, pred));
625 extend_irn(phi, new_in, false);
629 /* Returns the number of blocks backedges. With or without alien bes. */
630 static int get_backedge_n(ir_node *block, bool with_alien)
634 int arity = get_irn_arity(block);
636 assert(is_Block(block));
638 for (i = 0; i < arity; ++i) {
639 ir_node *pred = get_irn_n(block, i);
640 if (is_backedge(block, i) && (with_alien || is_in_loop(pred)))
646 /* Returns a raw copy of the given node.
647 * Attributes are kept/set according to the needs of loop inversion. */
648 static ir_node *copy_node(ir_node *node)
653 cp = exact_copy(node);
654 arity = get_irn_arity(node);
656 /* Keep backedge info */
657 for (i = 0; i < arity; ++i) {
658 if (is_backedge(node, i))
663 set_Block_mark(cp, 0);
671 * This walker copies all walked nodes.
672 * If the walk_condition is true for a node, it is copied.
673 * All nodes node_info->copy have to be NULL prior to every walk.
674 * Order of ins is important for later usage.
676 static void copy_walk(ir_node *node, walker_condition *walk_condition,
683 ir_graph *irg = current_ir_graph;
686 * break condition and cycle resolver, creating temporary node copies
688 if (get_irn_visited(node) >= get_irg_visited(irg)) {
689 /* Here we rely on nodestate's copy being initialized with NULL */
690 DB((dbg, LEVEL_5, "copy_walk: We have already visited %N\n", node));
691 if (get_inversion_copy(node) == NULL) {
692 cp = copy_node(node);
693 set_inversion_copy(node, cp);
695 DB((dbg, LEVEL_5, "The TEMP copy of %N is created %N\n", node, cp));
701 mark_irn_visited(node);
703 if (!is_Block(node)) {
704 ir_node *pred = get_nodes_block(node);
705 if (walk_condition(pred))
706 DB((dbg, LEVEL_5, "walk block %N\n", pred));
707 copy_walk(pred, walk_condition, set_loop);
710 arity = get_irn_arity(node);
712 NEW_ARR_A(ir_node *, cpin, arity);
714 for (i = 0; i < arity; ++i) {
715 ir_node *pred = get_irn_n(node, i);
717 if (walk_condition(pred)) {
718 DB((dbg, LEVEL_5, "walk node %N\n", pred));
719 copy_walk(pred, walk_condition, set_loop);
720 cpin[i] = get_inversion_copy(pred);
721 DB((dbg, LEVEL_5, "copy of %N gets new in %N which is copy of %N\n",
722 node, get_inversion_copy(pred), pred));
728 /* copy node / finalize temp node */
729 if (get_inversion_copy(node) == NULL) {
730 /* No temporary copy existent */
731 cp = copy_node(node);
732 set_inversion_copy(node, cp);
733 DB((dbg, LEVEL_5, "The FINAL copy of %N is CREATED %N\n", node, cp));
735 /* temporary copy is existent but without correct ins */
736 cp = get_inversion_copy(node);
737 DB((dbg, LEVEL_5, "The FINAL copy of %N is EXISTENT %N\n", node, cp));
740 if (!is_Block(node)) {
741 ir_node *cpblock = get_inversion_copy(get_nodes_block(node));
743 set_nodes_block(cp, cpblock );
745 add_Block_phi(cpblock, cp);
748 /* Keeps phi list of temporary node. */
749 set_irn_in(cp, ARR_LEN(cpin), cpin);
753 * This walker copies all walked nodes.
754 * If the walk_condition is true for a node, it is copied.
755 * All nodes node_info->copy have to be NULL prior to every walk.
756 * Order of ins is important for later usage.
757 * Takes copy_index, to phase-link copy at specific index.
759 static void copy_walk_n(ir_node *node, walker_condition *walk_condition,
768 * break condition and cycle resolver, creating temporary node copies
770 if (irn_visited(node)) {
771 /* Here we rely on nodestate's copy being initialized with NULL */
772 DB((dbg, LEVEL_5, "copy_walk: We have already visited %N\n", node));
773 if (get_unroll_copy(node, copy_index) == NULL) {
776 set_unroll_copy(node, copy_index, u);
777 DB((dbg, LEVEL_5, "The TEMP unknown of %N is created %N\n", node, u));
783 mark_irn_visited(node);
785 if (!is_Block(node)) {
786 ir_node *block = get_nodes_block(node);
787 if (walk_condition(block))
788 DB((dbg, LEVEL_5, "walk block %N\n", block));
789 copy_walk_n(block, walk_condition, copy_index);
792 arity = get_irn_arity(node);
793 NEW_ARR_A(ir_node *, cpin, arity);
795 for (i = 0; i < arity; ++i) {
796 ir_node *pred = get_irn_n(node, i);
798 if (walk_condition(pred)) {
799 DB((dbg, LEVEL_5, "walk node %N\n", pred));
800 copy_walk_n(pred, walk_condition, copy_index);
801 cpin[i] = get_unroll_copy(pred, copy_index);
807 /* copy node / finalize temp node */
808 cp = get_unroll_copy(node, copy_index);
809 if (cp == NULL || is_Unknown(cp)) {
810 cp = copy_node(node);
811 set_unroll_copy(node, copy_index, cp);
812 DB((dbg, LEVEL_5, "The FINAL copy of %N is CREATED %N\n", node, cp));
814 /* temporary copy is existent but without correct ins */
815 cp = get_unroll_copy(node, copy_index);
816 DB((dbg, LEVEL_5, "The FINAL copy of %N is EXISTENT %N\n", node, cp));
819 if (!is_Block(node)) {
820 ir_node *cpblock = get_unroll_copy(get_nodes_block(node), copy_index);
822 set_nodes_block(cp, cpblock );
824 add_Block_phi(cpblock, cp);
827 /* Keeps phi list of temporary node. */
828 set_irn_in(cp, ARR_LEN(cpin), cpin);
831 /* Removes alle Blocks with non marked predecessors from the condition chain. */
832 static void unmark_not_allowed_cc_blocks(void)
834 size_t blocks = ARR_LEN(cc_blocks);
837 for(i = 0; i < blocks; ++i) {
838 ir_node *block = cc_blocks[i];
840 int arity = get_irn_arity(block);
842 /* Head is an exception. */
843 if (block == loop_head)
846 for(a = 0; a < arity; ++a) {
847 if (! is_nodes_block_marked(get_irn_n(block, a))) {
848 set_Block_mark(block, 0);
849 --inversion_blocks_in_cc;
850 DB((dbg, LEVEL_5, "Removed %N from cc (blocks in cc %d)\n",
851 block, inversion_blocks_in_cc));
859 /* Unmarks all cc blocks using cc_blocks except head.
860 * TODO: invert head for unrolling? */
861 static void unmark_cc_blocks(void)
863 size_t blocks = ARR_LEN(cc_blocks);
866 for(i = 0; i < blocks; ++i) {
867 ir_node *block = cc_blocks[i];
869 /* TODO Head is an exception. */
870 /*if (block != loop_head)*/
871 set_Block_mark(block, 0);
873 /*inversion_blocks_in_cc = 1;*/
874 inversion_blocks_in_cc = 0;
877 loop_info.cc_size = 0;
881 * Populates head_entries with (node, pred_pos) tuple
882 * whereas the node's pred at pred_pos is in the cc but not the node itself.
883 * Also finds df loops inside the cc.
884 * Head and condition chain blocks have been marked previously.
886 static void get_head_outs(ir_node *node, void *env)
889 int arity = get_irn_arity(node);
892 for (i = 0; i < arity; ++i) {
893 if (!is_nodes_block_marked(node) && is_nodes_block_marked(get_irn_n(node, i))) {
897 /* Saving also predecessor seems redundant, but becomes
898 * necessary when changing position of it, before
899 * dereferencing it.*/
900 entry.pred = get_irn_n(node, i);
901 ARR_APP1(entry_edge, cur_head_outs, entry);
905 arity = get_irn_arity(loop_head);
907 /* Find df loops inside the cc */
908 if (is_Phi(node) && get_nodes_block(node) == loop_head) {
909 for (i = 0; i < arity; ++i) {
910 if (is_own_backedge(loop_head, i)) {
911 if (is_nodes_block_marked(get_irn_n(node, i))) {
915 entry.pred = get_irn_n(node, i);
916 ARR_APP1(entry_edge, head_df_loop, entry);
917 DB((dbg, LEVEL_5, "Found incc assignment node %N @%d is pred %N, graph %N %N\n",
918 node, i, entry.pred, current_ir_graph, get_irg_start_block(current_ir_graph)));
926 * Find condition chains, and add them to be inverted.
927 * A block belongs to the chain if a condition branches out of the loop.
928 * (Some blocks need to be removed once again.)
929 * Returns 1 if the given block belongs to the condition chain.
931 static void find_condition_chain(ir_node *block)
933 const ir_edge_t *edge;
936 bool jmp_only = true;
937 unsigned nodes_n = 0;
939 mark_irn_visited(block);
941 DB((dbg, LEVEL_5, "condition_chains for block %N\n", block));
944 foreach_out_edge_kind(block, edge, EDGE_KIND_NORMAL) {
948 /* Check if node count would exceed maximum cc size.
950 * This is not optimal, as we search depth-first and break here,
951 * continuing with another subtree. */
952 if (loop_info.cc_size + nodes_n > opt_params.max_cc_size) {
953 set_Block_mark(block, 0);
957 /* Check if block only has a jmp instruction. */
958 foreach_out_edge(block, edge) {
959 ir_node *src = get_edge_src_irn(edge);
961 if (!is_Block(src) && !is_Jmp(src)) {
966 /* Check cf outs if one is leaving the loop,
967 * or if this node has a backedge. */
968 foreach_block_succ(block, edge) {
969 ir_node *src = get_edge_src_irn(edge);
970 int pos = get_edge_src_pos(edge);
972 if (!is_in_loop(src))
975 /* Inverting blocks with backedge outs leads to a cf edge
976 * from the inverted head, into the inverted head (skipping the body).
977 * As the body becomes the new loop head,
978 * this would introduce another loop in the existing loop.
979 * This loop inversion cannot cope with this case. */
980 if (is_backedge(src, pos)) {
986 /* We need all predecessors to already belong to the condition chain.
987 * Example of wrong case: * == in cc
998 /* Collect blocks containing only a Jmp.
999 * Do not collect blocks with backedge outs. */
1000 if ((jmp_only || mark) && !has_be) {
1001 set_Block_mark(block, 1);
1002 ++inversion_blocks_in_cc;
1003 loop_info.cc_size += nodes_n;
1004 DB((dbg, LEVEL_5, "block %N is part of condition chain\n", block));
1005 ARR_APP1(ir_node *, cc_blocks, block);
1007 set_Block_mark(block, 0);
1010 foreach_block_succ(block, edge) {
1011 ir_node *src = get_edge_src_irn( edge );
1013 if (is_in_loop(src) && ! irn_visited(src))
1014 find_condition_chain(src);
1019 * Rewires the copied condition chain. Removes backedges
1020 * as this condition chain is prior to the loop.
1021 * Copy of loop_head must have phi list and old (unfixed) backedge info of the loop head.
1022 * (loop_head is already fixed, we cannot rely on it.)
1024 static void fix_copy_inversion(void)
1029 ir_node *phi, *next;
1030 ir_node *head_cp = get_inversion_copy(loop_head);
1031 ir_graph *irg = get_irn_irg(head_cp);
1032 int arity = get_irn_arity(head_cp);
1033 int backedges = get_backedge_n(head_cp, false);
1034 int new_arity = arity - backedges;
1038 NEW_ARR_A(ir_node *, ins, new_arity);
1041 /* Remove block backedges */
1042 for(i = 0; i < arity; ++i) {
1043 if (!is_backedge(head_cp, i))
1044 ins[pos++] = get_irn_n(head_cp, i);
1047 new_head = new_r_Block(irg, new_arity, ins);
1049 phis = NEW_ARR_F(ir_node *, 0);
1051 for_each_phi_safe(get_Block_phis(head_cp), phi, next) {
1053 NEW_ARR_A(ir_node *, ins, new_arity);
1055 for(i = 0; i < arity; ++i) {
1056 if (!is_backedge(head_cp, i))
1057 ins[pos++] = get_irn_n(phi, i);
1059 new_phi = new_rd_Phi(get_irn_dbg_info(phi),
1060 new_head, new_arity, ins,
1062 ARR_APP1(ir_node *, phis, new_phi);
1066 for_each_phi_safe(get_Block_phis(head_cp), phi, next) {
1067 exchange(phi, phis[pos++]);
1070 exchange(head_cp, new_head);
1076 /* Puts the original condition chain at the end of the loop,
1077 * subsequently to the body.
1078 * Relies on block phi list and correct backedges.
1080 static void fix_head_inversion(void)
1084 ir_node *phi, *next;
1086 ir_graph *irg = get_irn_irg(loop_head);
1087 int arity = get_irn_arity(loop_head);
1088 int backedges = get_backedge_n(loop_head, false);
1089 int new_arity = backedges;
1093 NEW_ARR_A(ir_node *, ins, new_arity);
1096 /* Keep only backedges */
1097 for(i = 0; i < arity; ++i) {
1098 if (is_own_backedge(loop_head, i))
1099 ins[pos++] = get_irn_n(loop_head, i);
1102 new_head = new_r_Block(irg, new_arity, ins);
1104 phis = NEW_ARR_F(ir_node *, 0);
1106 for_each_phi(loop_head, phi) {
1108 DB((dbg, LEVEL_5, "Fixing phi %N of loop head\n", phi));
1110 NEW_ARR_A(ir_node *, ins, new_arity);
1113 for (i = 0; i < arity; ++i) {
1114 ir_node *pred = get_irn_n(phi, i);
1116 if (is_own_backedge(loop_head, i)) {
1117 /* If assignment is in the condition chain,
1118 * we need to create a phi in the new loop head.
1119 * This can only happen for df, not cf. See find_condition_chains. */
1120 /*if (is_nodes_block_marked(pred)) {
1128 new_phi = new_rd_Phi(get_irn_dbg_info(phi),
1129 new_head, new_arity, ins,
1132 ARR_APP1(ir_node *, phis, new_phi);
1134 DB((dbg, LEVEL_5, "fix inverted head should exch %N by %N (pos %d)\n", phi, new_phi, pos ));
1138 for_each_phi_safe(get_Block_phis(loop_head), phi, next) {
1139 DB((dbg, LEVEL_5, "fix inverted exch phi %N by %N\n", phi, phis[pos]));
1140 if (phis[pos] != phi)
1141 exchange(phi, phis[pos++]);
1146 DB((dbg, LEVEL_5, "fix inverted head exch head block %N by %N\n", loop_head, new_head));
1147 exchange(loop_head, new_head);
1150 /* Does the loop inversion. */
1151 static void inversion_walk(ir_graph *irg, entry_edge *head_entries)
1156 * The order of rewiring bottom-up is crucial.
1157 * Any change of the order leads to lost information that would be needed later.
1160 ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED);
1162 /* 1. clone condition chain */
1163 inc_irg_visited(irg);
1165 for (i = 0; i < ARR_LEN(head_entries); ++i) {
1166 entry_edge entry = head_entries[i];
1167 ir_node *pred = get_irn_n(entry.node, entry.pos);
1169 DB((dbg, LEVEL_5, "\nInit walk block %N\n", pred));
1171 copy_walk(pred, is_nodes_block_marked, cur_loop);
1174 ir_free_resources(irg, IR_RESOURCE_IRN_VISITED);
1176 /* 2. Extends the head control flow successors ins
1177 * with the definitions of the copied head node. */
1178 for (i = 0; i < ARR_LEN(head_entries); ++i) {
1179 entry_edge head_out = head_entries[i];
1181 if (is_Block(head_out.node))
1182 extend_ins_by_copy(head_out.node, head_out.pos);
1185 /* 3. construct_ssa for users of definitions in the condition chain,
1186 * as there is now a second definition. */
1187 for (i = 0; i < ARR_LEN(head_entries); ++i) {
1188 entry_edge head_out = head_entries[i];
1190 /* Ignore keepalives */
1191 if (is_End(head_out.node))
1194 /* Construct ssa for assignments in the condition chain. */
1195 if (!is_Block(head_out.node)) {
1196 ir_node *pred, *cppred, *block, *cpblock;
1198 pred = head_out.pred;
1199 cppred = get_inversion_copy(pred);
1200 block = get_nodes_block(pred);
1201 cpblock = get_nodes_block(cppred);
1202 construct_ssa(block, pred, cpblock, cppred);
1207 * If there is an assignment in the condition chain
1208 * with a user also in the condition chain,
1209 * the dominance frontier is in the new loop head.
1210 * The dataflow loop is completely in the condition chain.
1215 * Phi_cp | | copied condition chain
1219 * Phi* | | new loop head with newly created phi.
1221 * Phi | | original, inverted condition chain
1226 for (i = 0; i < ARR_LEN(head_df_loop); ++i) {
1227 entry_edge head_out = head_df_loop[i];
1229 /* Construct ssa for assignments in the condition chain. */
1230 ir_node *pred, *cppred, *block, *cpblock;
1232 pred = head_out.pred;
1233 cppred = get_inversion_copy(pred);
1234 assert(cppred && pred);
1235 block = get_nodes_block(pred);
1236 cpblock = get_nodes_block(cppred);
1237 construct_ssa(block, pred, cpblock, cppred);
1240 /* 4. Remove the ins which are no backedges from the original condition chain
1241 * as the cc is now subsequent to the body. */
1242 fix_head_inversion();
1244 /* 5. Remove the backedges of the copied condition chain,
1245 * because it is going to be the new 'head' in advance to the loop. */
1246 fix_copy_inversion();
1250 /* Performs loop inversion of cur_loop if possible and reasonable. */
1251 static void loop_inversion(ir_graph *irg)
1254 unsigned max_loop_nodes = opt_params.max_loop_size;
1255 unsigned max_loop_nodes_adapted;
1256 int depth_adaption = opt_params.depth_adaption;
1258 bool do_inversion = true;
1260 /* Depth of 0 is the procedure and 1 a topmost loop. */
1261 loop_depth = get_loop_depth(cur_loop) - 1;
1263 /* Calculating in per mil. */
1264 max_loop_nodes_adapted = get_max_nodes_adapted(loop_depth);
1266 DB((dbg, LEVEL_1, "max_nodes: %d\nmax_nodes_adapted %d at depth of %d (adaption %d)\n",
1267 max_loop_nodes, max_loop_nodes_adapted, loop_depth, depth_adaption));
1269 if (loop_info.nodes == 0)
1272 if (loop_info.nodes > max_loop_nodes) {
1273 /* Only for stats */
1274 DB((dbg, LEVEL_1, "Nodes %d > allowed nodes %d\n",
1275 loop_info.nodes, loop_depth, max_loop_nodes));
1278 /* Adaption might change it */
1281 /* Limit processing to loops smaller than given parameter. */
1282 if (loop_info.nodes > max_loop_nodes_adapted) {
1283 DB((dbg, LEVEL_1, "Nodes %d > allowed nodes (depth %d adapted) %d\n",
1284 loop_info.nodes, loop_depth, max_loop_nodes_adapted));
1285 ++stats.too_large_adapted;
1289 if (loop_info.calls > opt_params.allowed_calls) {
1290 DB((dbg, LEVEL_1, "Calls %d > allowed calls %d\n",
1291 loop_info.calls, opt_params.allowed_calls));
1292 ++stats.calls_limit;
1296 /*inversion_head_node_limit = INT_MAX;*/
1297 ir_reserve_resources(irg, IR_RESOURCE_BLOCK_MARK);
1299 /* Reset block marks.
1300 * We use block marks to flag blocks of the original condition chain. */
1301 irg_walk_graph(irg, reset_block_mark, NULL, NULL);
1303 /*loop_info.blocks = get_loop_n_blocks(cur_loop);*/
1304 cond_chain_entries = NEW_ARR_F(entry_edge, 0);
1305 head_df_loop = NEW_ARR_F(entry_edge, 0);
1307 /*head_inversion_node_count = 0;*/
1308 inversion_blocks_in_cc = 0;
1310 /* Use phase to keep copy of nodes from the condition chain. */
1311 ir_nodemap_init(&map, irg);
1312 obstack_init(&obst);
1314 /* Search for condition chains and temporarily save the blocks in an array. */
1315 cc_blocks = NEW_ARR_F(ir_node *, 0);
1316 inc_irg_visited(irg);
1317 find_condition_chain(loop_head);
1319 unmark_not_allowed_cc_blocks();
1320 DEL_ARR_F(cc_blocks);
1322 /* Condition chain too large.
1323 * Loop should better be small enough to fit into the cache. */
1324 /* TODO Of course, we should take a small enough cc in the first place,
1325 * which is not that simple. (bin packing) */
1326 if (loop_info.cc_size > opt_params.max_cc_size) {
1327 ++stats.cc_limit_reached;
1329 do_inversion = false;
1331 /* Unmark cc blocks except the head.
1332 * Invert head only for possible unrolling. */
1336 /* We also catch endless loops here,
1337 * because they do not have a condition chain. */
1338 if (inversion_blocks_in_cc < 1) {
1339 do_inversion = false;
1341 "Loop contains %d (less than 1) invertible blocks => No Inversion done.\n",
1342 inversion_blocks_in_cc));
1346 cur_head_outs = NEW_ARR_F(entry_edge, 0);
1348 /* Get all edges pointing into the condition chain. */
1349 irg_walk_graph(irg, get_head_outs, NULL, NULL);
1351 /* Do the inversion */
1352 inversion_walk(irg, cur_head_outs);
1354 DEL_ARR_F(cur_head_outs);
1356 /* Duplicated blocks changed doms */
1357 clear_irg_properties(irg, IR_GRAPH_PROPERTY_CONSISTENT_DOMINANCE
1358 | IR_GRAPH_PROPERTY_CONSISTENT_LOOPINFO);
1364 obstack_free(&obst, NULL);
1365 ir_nodemap_destroy(&map);
1366 DEL_ARR_F(cond_chain_entries);
1367 DEL_ARR_F(head_df_loop);
1369 ir_free_resources(irg, IR_RESOURCE_BLOCK_MARK);
1372 /* Fix the original loop_heads ins for invariant unrolling case. */
1373 static void unrolling_fix_loop_head_inv(void)
1377 ir_node *proj = new_Proj(loop_info.duff_cond, mode_X, 0);
1378 ir_node *head_pred = get_irn_n(loop_head, loop_info.be_src_pos);
1379 ir_node *loop_condition = get_unroll_copy(head_pred, unroll_nr - 1);
1381 /* Original loop_heads ins are:
1382 * duff block and the own backedge */
1384 ins[0] = loop_condition;
1386 set_irn_in(loop_head, 2, ins);
1387 DB((dbg, LEVEL_4, "Rewire ins of block loophead %N to pred %N and duffs entry %N \n" , loop_head, ins[0], ins[1]));
1389 for_each_phi(loop_head, phi) {
1390 ir_node *pred = get_irn_n(phi, loop_info.be_src_pos);
1391 /* TODO we think it is a phi, but for Mergesort it is not the case.*/
1393 ir_node *last_pred = get_unroll_copy(pred, unroll_nr - 1);
1396 ins[1] = (ir_node*)get_irn_link(phi);
1397 set_irn_in(phi, 2, ins);
1398 DB((dbg, LEVEL_4, "Rewire ins of loophead phi %N to pred %N and duffs entry %N \n" , phi, ins[0], ins[1]));
1402 /* Removes previously created phis with only 1 in. */
1403 static void correct_phis(ir_node *node, void *env)
1407 if (is_Phi(node) && get_irn_arity(node) == 1) {
1411 in[0] = get_irn_n(node, 0);
1413 exch = new_rd_Phi(get_irn_dbg_info(node),
1414 get_nodes_block(node), 1, in,
1415 get_irn_mode(node));
1417 exchange(node, exch);
1421 /* Unrolling: Rewire floating copies. */
1422 static void place_copies(int copies)
1424 ir_node *loophead = loop_head;
1427 int be_src_pos = loop_info.be_src_pos;
1429 /* Serialize loops by fixing their head ins.
1430 * Processed are the copies.
1431 * The original loop is done after that, to keep backedge infos. */
1432 for (c = 0; c < copies; ++c) {
1433 ir_node *upper = get_unroll_copy(loophead, c);
1434 ir_node *lower = get_unroll_copy(loophead, c + 1);
1436 ir_node *topmost_be_block = get_nodes_block(get_irn_n(loophead, be_src_pos));
1438 /* Important: get the preds first and then their copy. */
1439 ir_node *upper_be_block = get_unroll_copy(topmost_be_block, c);
1440 ir_node *new_jmp = new_r_Jmp(upper_be_block);
1441 DB((dbg, LEVEL_5, " place_copies upper %N lower %N\n", upper, lower));
1443 DB((dbg, LEVEL_5, "topmost be block %N \n", topmost_be_block));
1445 if (loop_info.unroll_kind == constant) {
1448 set_irn_in(lower, 1, ins);
1450 for_each_phi(loophead, phi) {
1451 ir_node *topmost_def = get_irn_n(phi, be_src_pos);
1452 ir_node *upper_def = get_unroll_copy(topmost_def, c);
1453 ir_node *lower_phi = get_unroll_copy(phi, c + 1);
1455 /* It is possible, that the value used
1456 * in the OWN backedge path is NOT defined in this loop. */
1457 if (is_in_loop(topmost_def))
1460 ins[0] = topmost_def;
1462 set_irn_in(lower_phi, 1, ins);
1463 /* Need to replace phis with 1 in later. */
1466 /* Invariant case */
1467 /* Every node has 2 ins. One from the duff blocks
1468 * and one from the previously unrolled loop. */
1470 /* Calculate corresponding projection of mod result for this copy c */
1471 ir_node *proj = new_Proj(loop_info.duff_cond, mode_X, unroll_nr - c - 1);
1472 DB((dbg, LEVEL_4, "New duff proj %N\n" , proj));
1476 set_irn_in(lower, 2, ins);
1477 DB((dbg, LEVEL_4, "Rewire ins of Block %N to pred %N and duffs entry %N \n" , lower, ins[0], ins[1]));
1479 for_each_phi(loophead, phi) {
1480 ir_node *topmost_phi_pred = get_irn_n(phi, be_src_pos);
1481 ir_node *upper_phi_pred;
1485 lower_phi = get_unroll_copy(phi, c + 1);
1486 duff_phi = (ir_node*)get_irn_link(phi);
1487 DB((dbg, LEVEL_4, "DD Link of %N is %N\n" , phi, duff_phi));
1490 if (is_in_loop(topmost_phi_pred)) {
1491 upper_phi_pred = get_unroll_copy(topmost_phi_pred, c);
1493 upper_phi_pred = topmost_phi_pred;
1496 ins[0] = upper_phi_pred;
1498 set_irn_in(lower_phi, 2, ins);
1499 DB((dbg, LEVEL_4, "Rewire ins of %N to pred %N and duffs entry %N \n" , lower_phi, ins[0], ins[1]));
1504 /* Reconnect last copy. */
1505 for (i = 0; i < ARR_LEN(loop_entries); ++i) {
1506 entry_edge edge = loop_entries[i];
1507 /* Last copy is at the bottom */
1508 ir_node *new_pred = get_unroll_copy(edge.pred, copies);
1509 set_irn_n(edge.node, edge.pos, new_pred);
1512 /* Fix original loops head.
1513 * Done in the end, as ins and be info were needed before. */
1514 if (loop_info.unroll_kind == constant) {
1516 ir_node *head_pred = get_irn_n(loop_head, be_src_pos);
1517 ir_node *loop_condition = get_unroll_copy(head_pred, unroll_nr - 1);
1519 set_irn_n(loop_head, loop_info.be_src_pos, loop_condition);
1521 for_each_phi(loop_head, phi) {
1522 ir_node *pred = get_irn_n(phi, be_src_pos);
1525 /* It is possible, that the value used
1526 * in the OWN backedge path is NOT assigned in this loop. */
1527 if (is_in_loop(pred))
1528 last_pred = get_unroll_copy(pred, copies);
1531 set_irn_n(phi, be_src_pos, last_pred);
1535 unrolling_fix_loop_head_inv();
1539 /* Copies the cur_loop several times. */
1540 static void copy_loop(entry_edge *cur_loop_outs, int copies)
1544 ir_reserve_resources(current_ir_graph, IR_RESOURCE_IRN_VISITED);
1546 for (c = 0; c < copies; ++c) {
1549 inc_irg_visited(current_ir_graph);
1551 DB((dbg, LEVEL_5, " ### Copy_loop copy nr: %d ###\n", c));
1552 for (i = 0; i < ARR_LEN(cur_loop_outs); ++i) {
1553 entry_edge entry = cur_loop_outs[i];
1554 ir_node *pred = get_irn_n(entry.node, entry.pos);
1556 copy_walk_n(pred, is_in_loop, c + 1);
1560 ir_free_resources(current_ir_graph, IR_RESOURCE_IRN_VISITED);
1564 /* Creates a new phi from the given phi node omitting own bes,
1565 * using be_block as supplier of backedge informations. */
1566 static ir_node *clone_phis_sans_bes(ir_node *phi, ir_node *be_block, ir_node *dest_block)
1569 int arity = get_irn_arity(phi);
1573 assert(get_irn_arity(phi) == get_irn_arity(be_block));
1574 assert(is_Phi(phi));
1576 ins = NEW_ARR_F(ir_node *, arity);
1577 for (i = 0; i < arity; ++i) {
1578 if (! is_own_backedge(be_block, i)) {
1579 ins[c] = get_irn_n(phi, i);
1584 newphi = new_r_Phi(dest_block, c, ins, get_irn_mode(phi));
1586 set_irn_link(phi, newphi);
1587 DB((dbg, LEVEL_4, "Linking for duffs device %N to %N\n", phi, newphi));
1592 /* Creates a new block from the given block node omitting own bes,
1593 * using be_block as supplier of backedge informations. */
1594 static ir_node *clone_block_sans_bes(ir_node *node, ir_node *be_block)
1596 int arity = get_irn_arity(node);
1600 assert(get_irn_arity(node) == get_irn_arity(be_block));
1601 assert(is_Block(node));
1603 NEW_ARR_A(ir_node *, ins, arity);
1604 for (i = 0; i < arity; ++i) {
1605 if (! is_own_backedge(be_block, i)) {
1606 ins[c] = get_irn_n(node, i);
1611 return new_Block(c, ins);
1614 /* Creates a structure to calculate absolute value of node op.
1615 * Returns mux node with absolute value. */
1616 static ir_node *new_Abs(ir_node *op, ir_mode *mode)
1618 ir_graph *irg = get_irn_irg(op);
1619 ir_node *block = get_nodes_block(op);
1620 ir_node *zero = new_r_Const(irg, get_mode_null(mode));
1621 ir_node *cmp = new_r_Cmp(block, op, zero, ir_relation_less);
1622 ir_node *minus_op = new_r_Minus(block, op, mode);
1623 ir_node *mux = new_r_Mux(block, cmp, op, minus_op, mode);
1629 /* Creates blocks for duffs device, using previously obtained
1630 * informations about the iv.
1632 static void create_duffs_block(void)
1636 ir_node *block1, *count_block, *duff_block;
1637 ir_node *ems, *ems_mod, *ems_div, *ems_mod_proj, *cmp_null,
1638 *ems_mode_cond, *x_true, *x_false, *const_null;
1639 ir_node *true_val, *false_val;
1642 ir_node *duff_mod, *proj, *cond;
1644 ir_node *count, *correction, *unroll_c;
1645 ir_node *cmp_bad_count, *good_count, *bad_count, *count_phi, *bad_count_neg;
1648 mode = get_irn_mode(loop_info.end_val);
1649 const_null = new_Const(get_mode_null(mode));
1652 * 1. Calculate first approach to count.
1653 * Condition: (end - start) % step == 0 */
1654 block1 = clone_block_sans_bes(loop_head, loop_head);
1655 DB((dbg, LEVEL_4, "Duff block 1 %N\n", block1));
1657 /* Create loop entry phis in first duff block
1658 * as it becomes the loops preheader */
1659 for_each_phi(loop_head, phi) {
1660 /* Returns phis pred if phi would have arity 1*/
1661 ir_node *new_phi = clone_phis_sans_bes(phi, loop_head, block1);
1663 DB((dbg, LEVEL_4, "HEAD %N phi %N\n", loop_head, phi));
1664 DB((dbg, LEVEL_4, "BLOCK1 %N phi %N\n", block1, new_phi));
1667 ems = new_r_Sub(block1, loop_info.end_val, loop_info.start_val,
1668 get_irn_mode(loop_info.end_val));
1669 DB((dbg, LEVEL_4, "BLOCK1 sub %N\n", ems));
1672 ems = new_Sub(loop_info.end_val, loop_info.start_val,
1673 get_irn_mode(loop_info.end_val));
1675 DB((dbg, LEVEL_4, "mod ins %N %N\n", ems, loop_info.step));
1676 ems_mod = new_r_Mod(block1,
1681 op_pin_state_pinned);
1682 ems_div = new_r_Div(block1,
1687 op_pin_state_pinned);
1689 DB((dbg, LEVEL_4, "New module node %N\n", ems_mod));
1691 ems_mod_proj = new_r_Proj(ems_mod, mode_Iu, pn_Mod_res);
1692 cmp_null = new_r_Cmp(block1, ems_mod_proj, const_null, ir_relation_less);
1693 ems_mode_cond = new_r_Cond(block1, cmp_null);
1695 /* ems % step == 0 */
1696 x_true = new_r_Proj(ems_mode_cond, mode_X, pn_Cond_true);
1697 /* ems % step != 0 */
1698 x_false = new_r_Proj(ems_mode_cond, mode_X, pn_Cond_false);
1701 * Assures, duffs device receives a valid count.
1703 * decreasing: count < 0
1704 * increasing: count > 0
1709 count_block = new_Block(2, ins);
1710 DB((dbg, LEVEL_4, "Duff block 2 %N\n", count_block));
1713 /* Increase loop-taken-count depending on the loop condition
1714 * uses the latest iv to compare to. */
1715 if (loop_info.latest_value == 1) {
1716 /* ems % step == 0 : +0 */
1717 true_val = new_Const(get_mode_null(mode));
1718 /* ems % step != 0 : +1 */
1719 false_val = new_Const(get_mode_one(mode));
1721 ir_tarval *tv_two = new_tarval_from_long(2, mode);
1722 /* ems % step == 0 : +1 */
1723 true_val = new_Const(get_mode_one(mode));
1724 /* ems % step != 0 : +2 */
1725 false_val = new_Const(tv_two);
1731 correction = new_r_Phi(count_block, 2, ins, mode);
1733 count = new_r_Proj(ems_div, mode, pn_Div_res);
1735 /* (end - start) / step + correction */
1736 count = new_Add(count, correction, mode);
1738 /* We preconditioned the loop to be tail-controlled.
1739 * So, if count is something 'wrong' like 0,
1740 * negative/positive (depending on step direction),
1741 * we may take the loop once (tail-contr.) and leave it
1742 * to the existing condition, to break; */
1744 /* Depending on step direction, we have to check for > or < 0 */
1745 if (loop_info.decreasing == 1) {
1746 cmp_bad_count = new_r_Cmp(count_block, count, const_null,
1749 cmp_bad_count = new_r_Cmp(count_block, count, const_null,
1750 ir_relation_greater);
1753 bad_count_neg = new_r_Cond(count_block, cmp_bad_count);
1754 good_count = new_Proj(bad_count_neg, mode_X, pn_Cond_true);
1755 bad_count = new_Proj(ems_mode_cond, mode_X, pn_Cond_false);
1758 * Contains module to decide which loop to start from. */
1760 ins[0] = good_count;
1762 duff_block = new_Block(2, ins);
1763 DB((dbg, LEVEL_4, "Duff block 3 %N\n", duff_block));
1765 /* Get absolute value */
1766 ins[0] = new_Abs(count, mode);
1767 /* Manually feed the aforementioned count = 1 (bad case)*/
1768 ins[1] = new_Const(get_mode_one(mode));
1769 count_phi = new_r_Phi(duff_block, 2, ins, mode);
1771 unroll_c = new_Const(new_tarval_from_long((long)unroll_nr, mode));
1773 /* count % unroll_nr */
1774 duff_mod = new_r_Mod(duff_block,
1779 op_pin_state_pinned);
1782 proj = new_Proj(duff_mod, mode, pn_Mod_res);
1783 /* condition does NOT create itself in the block of the proj! */
1784 cond = new_r_Cond(duff_block, proj);
1786 loop_info.duff_cond = cond;
1789 /* Returns 1 if given node is not in loop,
1790 * or if it is a phi of the loop head with only loop invariant defs.
1792 static unsigned is_loop_invariant_def(ir_node *node)
1796 if (! is_in_loop(node)) {
1797 DB((dbg, LEVEL_4, "Not in loop %N\n", node));
1798 /* || is_Const(node) || is_SymConst(node)) {*/
1802 /* If this is a phi of the loophead shared by more than 1 loop,
1803 * we need to check if all defs are not in the loop. */
1806 block = get_nodes_block(node);
1808 /* To prevent unexpected situations. */
1809 if (block != loop_head) {
1813 for (i = 0; i < get_irn_arity(node); ++i) {
1814 /* Check if all bes are just loopbacks. */
1815 if (is_own_backedge(block, i) && get_irn_n(node, i) != node)
1818 DB((dbg, LEVEL_4, "invar %N\n", node));
1821 DB((dbg, LEVEL_4, "Not invar %N\n", node));
1826 /* Returns 1 if one pred of node is invariant and the other is not.
1827 * invar_pred and other are set analogously. */
1828 static unsigned get_invariant_pred(ir_node *node, ir_node **invar_pred, ir_node **other)
1830 ir_node *pred0 = get_irn_n(node, 0);
1831 ir_node *pred1 = get_irn_n(node, 1);
1836 if (is_loop_invariant_def(pred0)) {
1837 DB((dbg, LEVEL_4, "pred0 invar %N\n", pred0));
1838 *invar_pred = pred0;
1842 if (is_loop_invariant_def(pred1)) {
1843 DB((dbg, LEVEL_4, "pred1 invar %N\n", pred1));
1845 if (*invar_pred != NULL) {
1846 /* RETURN. We do not want both preds to be invariant. */
1851 *invar_pred = pred1;
1854 DB((dbg, LEVEL_4, "pred1 not invar %N\n", pred1));
1856 if (*invar_pred != NULL)
1863 /* Starts from a phi that may belong to an iv.
1864 * If an add forms a loop with iteration_phi,
1865 * and add uses a constant, 1 is returned
1866 * and 'start' as well as 'add' are sane. */
1867 static unsigned get_start_and_add(ir_node *iteration_phi, unrolling_kind_flag role)
1870 ir_node *found_add = loop_info.add;
1871 int arity = get_irn_arity(iteration_phi);
1873 DB((dbg, LEVEL_4, "Find start and add from %N\n", iteration_phi));
1875 for (i = 0; i < arity; ++i) {
1877 /* Find start_val which needs to be pred of the iteration_phi.
1878 * If start_val already known, sanity check. */
1879 if (!is_backedge(get_nodes_block(loop_info.iteration_phi), i)) {
1880 ir_node *found_start_val = get_irn_n(loop_info.iteration_phi, i);
1882 DB((dbg, LEVEL_4, "found_start_val %N\n", found_start_val));
1884 /* We already found a start_val it has to be always the same. */
1885 if (loop_info.start_val && found_start_val != loop_info.start_val)
1888 if ((role == constant) && !(is_SymConst(found_start_val) || is_Const(found_start_val)))
1890 else if((role == constant) && !(is_loop_invariant_def(found_start_val)))
1893 loop_info.start_val = found_start_val;
1896 /* The phi has to be in the loop head.
1897 * Follow all own backedges. Every value supplied from these preds of the phi
1898 * needs to origin from the same add. */
1899 if (is_own_backedge(get_nodes_block(loop_info.iteration_phi), i)) {
1900 ir_node *new_found = get_irn_n(loop_info.iteration_phi,i);
1902 DB((dbg, LEVEL_4, "is add? %N\n", new_found));
1904 if (! (is_Add(new_found) || is_Sub(new_found)) || (found_add && found_add != new_found))
1907 found_add = new_found;
1911 loop_info.add = found_add;
1917 /* Returns 1 if one pred of node is a const value and the other is not.
1918 * const_pred and other are set analogously. */
1919 static unsigned get_const_pred(ir_node *node, ir_node **const_pred, ir_node **other)
1921 ir_node *pred0 = get_irn_n(node, 0);
1922 ir_node *pred1 = get_irn_n(node, 1);
1924 DB((dbg, LEVEL_4, "Checking for constant pred of %N\n", node));
1929 /*DB((dbg, LEVEL_4, "is %N const\n", pred0));*/
1930 if (is_Const(pred0) || is_SymConst(pred0)) {
1931 *const_pred = pred0;
1935 /*DB((dbg, LEVEL_4, "is %N const\n", pred1));*/
1936 if (is_Const(pred1) || is_SymConst(pred1)) {
1937 if (*const_pred != NULL) {
1938 /* RETURN. We do not want both preds to be constant. */
1943 *const_pred = pred1;
1946 if (*const_pred == NULL)
1952 /* Returns 1 if loop exits within 2 steps of the iv.
1953 * Norm_proj means we do not exit the loop.*/
1954 static unsigned simulate_next(ir_tarval **count_tar,
1955 ir_tarval *stepped, ir_tarval *step_tar, ir_tarval *end_tar,
1956 ir_relation norm_proj)
1960 DB((dbg, LEVEL_4, "Loop taken if (stepped)%ld %s (end)%ld ",
1961 get_tarval_long(stepped),
1962 get_relation_string((norm_proj)),
1963 get_tarval_long(end_tar)));
1964 DB((dbg, LEVEL_4, "comparing latest value %d\n", loop_info.latest_value));
1966 /* If current iv does not stay in the loop,
1967 * this run satisfied the exit condition. */
1968 if (! (tarval_cmp(stepped, end_tar) & norm_proj))
1971 DB((dbg, LEVEL_4, "Result: (stepped)%ld IS %s (end)%ld\n",
1972 get_tarval_long(stepped),
1973 get_relation_string(tarval_cmp(stepped, end_tar)),
1974 get_tarval_long(end_tar)));
1977 if (is_Add(loop_info.add))
1978 next = tarval_add(stepped, step_tar);
1981 next = tarval_sub(stepped, step_tar, get_irn_mode(loop_info.end_val));
1983 DB((dbg, LEVEL_4, "Loop taken if %ld %s %ld ",
1984 get_tarval_long(next),
1985 get_relation_string(norm_proj),
1986 get_tarval_long(end_tar)));
1987 DB((dbg, LEVEL_4, "comparing latest value %d\n", loop_info.latest_value));
1989 /* Increase steps. */
1990 *count_tar = tarval_add(*count_tar, get_tarval_one(get_tarval_mode(*count_tar)));
1992 /* Next has to fail the loop condition, or we will never exit. */
1993 if (! (tarval_cmp(next, end_tar) & norm_proj))
1999 /* Check if loop meets requirements for a 'simple loop':
2000 * - Exactly one cf out
2002 * - Max nodes after unrolling
2006 * Returns Projection of cmp node or NULL; */
2007 static ir_node *is_simple_loop(void)
2010 ir_node *loop_block, *exit_block, *projx, *cond, *cmp;
2012 /* Maximum of one condition, and no endless loops. */
2013 if (loop_info.cf_outs != 1)
2016 DB((dbg, LEVEL_4, "1 loop exit\n"));
2018 /* Calculate maximum unroll_nr keeping node count below limit. */
2019 loop_info.max_unroll = (int)((double)opt_params.max_unrolled_loop_size / (double)loop_info.nodes);
2020 if (loop_info.max_unroll < 2) {
2025 DB((dbg, LEVEL_4, "maximum unroll factor %u, to not exceed node limit \n",
2026 opt_params.max_unrolled_loop_size));
2028 arity = get_irn_arity(loop_head);
2029 /* RETURN if we have more than 1 be. */
2030 /* Get my backedges without alien bes. */
2032 for (i = 0; i < arity; ++i) {
2033 ir_node *pred = get_irn_n(loop_head, i);
2034 if (is_own_backedge(loop_head, i)) {
2036 /* Our simple loops may have only one backedge. */
2039 loop_block = get_nodes_block(pred);
2040 loop_info.be_src_pos = i;
2045 DB((dbg, LEVEL_4, "loop has 1 own backedge.\n"));
2047 exit_block = get_nodes_block(loop_info.cf_out.pred);
2048 /* The loop has to be tail-controlled.
2049 * This can be changed/improved,
2050 * but we would need a duff iv. */
2051 if (exit_block != loop_block)
2054 DB((dbg, LEVEL_4, "tail-controlled loop.\n"));
2056 /* find value on which loop exit depends */
2057 projx = loop_info.cf_out.pred;
2058 cond = get_irn_n(projx, 0);
2059 cmp = get_irn_n(cond, 0);
2064 DB((dbg, LEVEL_5, "projection is %s\n", get_relation_string(get_Proj_proj(projx))));
2066 switch(get_Proj_proj(projx)) {
2068 loop_info.exit_cond = 0;
2071 loop_info.exit_cond = 1;
2074 panic("Cond Proj_proj other than true/false");
2077 DB((dbg, LEVEL_4, "Valid Cmp.\n"));
2081 /* Returns 1 if all nodes are mode_Iu or mode_Is. */
2082 static unsigned are_mode_I(ir_node *n1, ir_node* n2, ir_node *n3)
2084 ir_mode *m1 = get_irn_mode(n1);
2085 ir_mode *m2 = get_irn_mode(n2);
2086 ir_mode *m3 = get_irn_mode(n3);
2088 if ((m1 == mode_Iu && m2 == mode_Iu && m3 == mode_Iu) ||
2089 (m1 == mode_Is && m2 == mode_Is && m3 == mode_Is))
2095 /* Checks if cur_loop is a simple tail-controlled counting loop
2096 * with start and end value loop invariant, step constant. */
2097 static unsigned get_unroll_decision_invariant(void)
2100 ir_node *projres, *loop_condition, *iteration_path;
2102 ir_tarval *step_tar;
2106 /* RETURN if loop is not 'simple' */
2107 projres = is_simple_loop();
2108 if (projres == NULL)
2111 /* Use a minimal size for the invariant unrolled loop,
2112 * as duffs device produces overhead */
2113 if (loop_info.nodes < opt_params.invar_unrolling_min_size)
2116 loop_condition = get_irn_n(projres, 0);
2118 success = get_invariant_pred(loop_condition, &loop_info.end_val, &iteration_path);
2119 DB((dbg, LEVEL_4, "pred invar %d\n", success));
2124 DB((dbg, LEVEL_4, "Invariant End_val %N, other %N\n", loop_info.end_val, iteration_path));
2126 /* We may find the add or the phi first.
2127 * Until now we only have end_val. */
2128 if (is_Add(iteration_path) || is_Sub(iteration_path)) {
2130 loop_info.add = iteration_path;
2131 DB((dbg, LEVEL_4, "Case 1: Got add %N (maybe not sane)\n", loop_info.add));
2133 /* Preds of the add should be step and the iteration_phi */
2134 success = get_const_pred(loop_info.add, &loop_info.step, &loop_info.iteration_phi);
2138 DB((dbg, LEVEL_4, "Got step %N\n", loop_info.step));
2140 if (! is_Phi(loop_info.iteration_phi))
2143 DB((dbg, LEVEL_4, "Got phi %N\n", loop_info.iteration_phi));
2146 * Does necessary sanity check of add, if it is already set. */
2147 success = get_start_and_add(loop_info.iteration_phi, invariant);
2151 DB((dbg, LEVEL_4, "Got start A %N\n", loop_info.start_val));
2153 } else if (is_Phi(iteration_path)) {
2154 ir_node *new_iteration_phi;
2156 loop_info.iteration_phi = iteration_path;
2157 DB((dbg, LEVEL_4, "Case 2: Got phi %N\n", loop_info.iteration_phi));
2159 /* Find start_val and add-node.
2160 * Does necessary sanity check of add, if it is already set. */
2161 success = get_start_and_add(loop_info.iteration_phi, invariant);
2165 DB((dbg, LEVEL_4, "Got start B %N\n", loop_info.start_val));
2166 DB((dbg, LEVEL_4, "Got add or sub %N\n", loop_info.add));
2168 success = get_const_pred(loop_info.add, &loop_info.step, &new_iteration_phi);
2172 DB((dbg, LEVEL_4, "Got step (B) %N\n", loop_info.step));
2174 if (loop_info.iteration_phi != new_iteration_phi)
2181 mode = get_irn_mode(loop_info.end_val);
2183 DB((dbg, LEVEL_4, "start %N, end %N, step %N\n",
2184 loop_info.start_val, loop_info.end_val, loop_info.step));
2186 if (mode != mode_Is && mode != mode_Iu)
2189 /* TODO necessary? */
2190 if (!are_mode_I(loop_info.start_val, loop_info.step, loop_info.end_val))
2193 DB((dbg, LEVEL_4, "mode integer\n"));
2195 step_tar = get_Const_tarval(loop_info.step);
2197 if (tarval_is_null(step_tar)) {
2198 /* TODO Might be worth a warning. */
2202 DB((dbg, LEVEL_4, "step is not 0\n"));
2204 create_duffs_block();
2206 return loop_info.max_unroll;
2209 /* Returns unroll factor,
2210 * given maximum unroll factor and number of loop passes. */
2211 static unsigned get_preferred_factor_constant(ir_tarval *count_tar)
2213 ir_tarval *tar_6, *tar_5, *tar_4, *tar_3, *tar_2;
2215 ir_mode *mode = get_irn_mode(loop_info.end_val);
2217 tar_6 = new_tarval_from_long(6, mode);
2218 tar_5 = new_tarval_from_long(5, mode);
2219 tar_4 = new_tarval_from_long(4, mode);
2220 tar_3 = new_tarval_from_long(3, mode);
2221 tar_2 = new_tarval_from_long(2, mode);
2223 /* loop passes % {6, 5, 4, 3, 2} == 0 */
2224 if (tarval_is_null(tarval_mod(count_tar, tar_6)))
2226 else if (tarval_is_null(tarval_mod(count_tar, tar_5)))
2228 else if (tarval_is_null(tarval_mod(count_tar, tar_4)))
2230 else if (tarval_is_null(tarval_mod(count_tar, tar_3)))
2232 else if (tarval_is_null(tarval_mod(count_tar, tar_2)))
2235 /* gcd(max_unroll, count_tar) */
2236 int a = loop_info.max_unroll;
2237 int b = (int)get_tarval_long(count_tar);
2240 DB((dbg, LEVEL_4, "gcd of max_unroll %d and count_tar %d: ", a, b));
2247 DB((dbg, LEVEL_4, "%d\n", a));
2251 DB((dbg, LEVEL_4, "preferred unroll factor %d\n", prefer));
2254 * If our preference is greater than the allowed unroll factor
2255 * we either might reduce the preferred factor and prevent a duffs device block,
2256 * or create a duffs device block, from which in this case (constants only)
2257 * we know the startloop at compiletime.
2258 * The latter yields the following graphs.
2259 * but for code generation we would want to use graph A.
2260 * The graphs are equivalent. So, we can only reduce the preferred factor.
2270 if (prefer <= loop_info.max_unroll)
2275 if (loop_info.max_unroll >= 3)
2277 else if (loop_info.max_unroll >= 2)
2283 if (loop_info.max_unroll >= 2)
2294 /* Check if cur_loop is a simple counting loop.
2295 * Start, step and end are constants.
2296 * TODO The whole constant case should use procedures similar to
2297 * the invariant case, as they are more versatile. */
2299 static unsigned get_unroll_decision_constant(void)
2301 ir_node *cmp, *iteration_path;
2302 unsigned success, is_latest_val;
2303 ir_tarval *start_tar, *end_tar, *step_tar, *diff_tar, *count_tar;
2305 ir_relation proj_proj, norm_proj;
2308 /* RETURN if loop is not 'simple' */
2309 cmp = is_simple_loop();
2313 /* One in of the loop condition needs to be loop invariant. => end_val
2314 * The other in is assigned by an add. => add
2315 * The add uses a loop invariant value => step
2316 * and a phi with a loop invariant start_val and the add node as ins.
2328 success = get_const_pred(cmp, &loop_info.end_val, &iteration_path);
2332 DB((dbg, LEVEL_4, "End_val %N, other %N\n", loop_info.end_val, iteration_path));
2334 /* We may find the add or the phi first.
2335 * Until now we only have end_val. */
2336 if (is_Add(iteration_path) || is_Sub(iteration_path)) {
2338 /* We test against the latest value of the iv. */
2341 loop_info.add = iteration_path;
2342 DB((dbg, LEVEL_4, "Case 2: Got add %N (maybe not sane)\n", loop_info.add));
2344 /* Preds of the add should be step and the iteration_phi */
2345 success = get_const_pred(loop_info.add, &loop_info.step, &loop_info.iteration_phi);
2349 DB((dbg, LEVEL_4, "Got step %N\n", loop_info.step));
2351 if (! is_Phi(loop_info.iteration_phi))
2354 DB((dbg, LEVEL_4, "Got phi %N\n", loop_info.iteration_phi));
2357 * Does necessary sanity check of add, if it is already set. */
2358 success = get_start_and_add(loop_info.iteration_phi, constant);
2362 DB((dbg, LEVEL_4, "Got start %N\n", loop_info.start_val));
2364 } else if (is_Phi(iteration_path)) {
2365 ir_node *new_iteration_phi;
2367 /* We compare with the value the iv had entering this run. */
2370 loop_info.iteration_phi = iteration_path;
2371 DB((dbg, LEVEL_4, "Case 1: Got phi %N \n", loop_info.iteration_phi));
2373 /* Find start_val and add-node.
2374 * Does necessary sanity check of add, if it is already set. */
2375 success = get_start_and_add(loop_info.iteration_phi, constant);
2379 DB((dbg, LEVEL_4, "Got start %N\n", loop_info.start_val));
2380 DB((dbg, LEVEL_4, "Got add or sub %N\n", loop_info.add));
2382 success = get_const_pred(loop_info.add, &loop_info.step, &new_iteration_phi);
2386 DB((dbg, LEVEL_4, "Got step %N\n", loop_info.step));
2388 if (loop_info.iteration_phi != new_iteration_phi)
2396 mode = get_irn_mode(loop_info.end_val);
2398 DB((dbg, LEVEL_4, "start %N, end %N, step %N\n",
2399 loop_info.start_val, loop_info.end_val, loop_info.step));
2401 if (mode != mode_Is && mode != mode_Iu)
2404 /* TODO necessary? */
2405 if (!are_mode_I(loop_info.start_val, loop_info.step, loop_info.end_val))
2408 DB((dbg, LEVEL_4, "mode integer\n"));
2410 end_tar = get_Const_tarval(loop_info.end_val);
2411 start_tar = get_Const_tarval(loop_info.start_val);
2412 step_tar = get_Const_tarval(loop_info.step);
2414 if (tarval_is_null(step_tar))
2415 /* TODO Might be worth a warning. */
2418 DB((dbg, LEVEL_4, "step is not 0\n"));
2420 if ((!tarval_is_negative(step_tar)) ^ (!is_Sub(loop_info.add)))
2421 loop_info.decreasing = 1;
2423 diff_tar = tarval_sub(end_tar, start_tar, mode);
2425 /* We need at least count_tar steps to be close to end_val, maybe more.
2426 * No way, that we have gone too many steps.
2427 * This represents the 'latest value'.
2428 * (If condition checks against latest value, is checked later) */
2429 count_tar = tarval_div(diff_tar, step_tar);
2431 /* Iv will not pass end_val (except overflows).
2432 * Nothing done, as it would yield to no advantage. */
2433 if (tarval_is_negative(count_tar)) {
2434 DB((dbg, LEVEL_4, "Loop is endless or never taken."));
2435 /* TODO Might be worth a warning. */
2439 ++stats.u_simple_counting_loop;
2441 loop_info.latest_value = is_latest_val;
2444 if (! is_simple_counting_loop(&count_tar))
2448 /* stepped can be negative, if step < 0 */
2449 stepped = tarval_mul(count_tar, step_tar);
2451 /* step as close to end_val as possible, */
2452 /* |stepped| <= |end_tar|, and dist(stepped, end_tar) is smaller than a step. */
2453 if (is_Sub(loop_info.add))
2454 stepped = tarval_sub(start_tar, stepped, mode_Is);
2456 stepped = tarval_add(start_tar, stepped);
2458 DB((dbg, LEVEL_4, "stepped to %ld\n", get_tarval_long(stepped)));
2460 proj_proj = get_Cmp_relation(cmp);
2461 /* Assure that norm_proj is the stay-in-loop case. */
2462 if (loop_info.exit_cond == 1)
2463 norm_proj = get_negated_relation(proj_proj);
2465 norm_proj = proj_proj;
2467 DB((dbg, LEVEL_4, "normalized projection %s\n", get_relation_string(norm_proj)));
2468 /* Executed at most once (stay in counting loop if a Eq b) */
2469 if (norm_proj == ir_relation_equal)
2470 /* TODO Might be worth a warning. */
2473 /* calculates next values and increases count_tar according to it */
2474 success = simulate_next(&count_tar, stepped, step_tar, end_tar, norm_proj);
2478 /* We run loop once more, if we compare to the
2479 * not yet in-/decreased iv. */
2480 if (is_latest_val == 0) {
2481 DB((dbg, LEVEL_4, "condition uses not latest iv value\n"));
2482 count_tar = tarval_add(count_tar, get_tarval_one(mode));
2485 DB((dbg, LEVEL_4, "loop taken %ld times\n", get_tarval_long(count_tar)));
2487 /* Assure the loop is taken at least 1 time. */
2488 if (tarval_is_null(count_tar)) {
2489 /* TODO Might be worth a warning. */
2493 loop_info.count_tar = count_tar;
2494 return get_preferred_factor_constant(count_tar);
2500 static void unroll_loop(void)
2503 if (! (loop_info.nodes > 0))
2506 if (loop_info.nodes > opt_params.max_unrolled_loop_size) {
2507 DB((dbg, LEVEL_2, "Nodes %d > allowed nodes %d\n",
2508 loop_info.nodes, opt_params.max_unrolled_loop_size));
2513 if (loop_info.calls > 0) {
2514 DB((dbg, LEVEL_2, "Calls %d > allowed calls 0\n",
2516 ++stats.calls_limit;
2522 /* get_unroll_decision_constant and invariant are completely
2523 * independent for flexibility.
2524 * Some checks may be performed twice. */
2526 /* constant case? */
2527 if (opt_params.allow_const_unrolling)
2528 unroll_nr = get_unroll_decision_constant();
2529 if (unroll_nr > 1) {
2530 loop_info.unroll_kind = constant;
2533 /* invariant case? */
2534 if (opt_params.allow_invar_unrolling)
2535 unroll_nr = get_unroll_decision_invariant();
2537 loop_info.unroll_kind = invariant;
2540 DB((dbg, LEVEL_2, " *** Unrolling %d times ***\n", unroll_nr));
2542 if (unroll_nr > 1) {
2543 loop_entries = NEW_ARR_F(entry_edge, 0);
2546 irg_walk_graph(current_ir_graph, get_loop_entries, NULL, NULL);
2548 if (loop_info.unroll_kind == constant) {
2549 if ((int)get_tarval_long(loop_info.count_tar) == unroll_nr)
2550 loop_info.needs_backedge = 0;
2552 loop_info.needs_backedge = 1;
2554 loop_info.needs_backedge = 1;
2557 /* Use phase to keep copy of nodes from the condition chain. */
2558 ir_nodemap_init(&map, current_ir_graph);
2559 obstack_init(&obst);
2561 /* Copies the loop */
2562 copy_loop(loop_entries, unroll_nr - 1);
2564 /* Line up the floating copies. */
2565 place_copies(unroll_nr - 1);
2567 /* Remove phis with 1 in
2568 * If there were no nested phis, this would not be necessary.
2569 * Avoiding the creation in the first place
2570 * leads to complex special cases. */
2571 irg_walk_graph(current_ir_graph, correct_phis, NULL, NULL);
2573 if (loop_info.unroll_kind == constant)
2574 ++stats.constant_unroll;
2576 ++stats.invariant_unroll;
2578 clear_irg_properties(current_ir_graph, IR_GRAPH_PROPERTY_CONSISTENT_DOMINANCE);
2580 DEL_ARR_F(loop_entries);
2581 obstack_free(&obst, NULL);
2582 ir_nodemap_destroy(&map);
2587 /* Analyzes the loop, and checks if size is within allowed range.
2588 * Decides if loop will be processed. */
2589 static void init_analyze(ir_graph *irg, ir_loop *loop)
2594 loop_head_valid = true;
2596 /* Reset loop info */
2597 memset(&loop_info, 0, sizeof(loop_info_t));
2599 DB((dbg, LEVEL_1, " >>>> current loop %ld <<<\n",
2600 get_loop_loop_nr(loop)));
2602 /* Collect loop informations: head, node counts. */
2603 irg_walk_graph(irg, get_loop_info, NULL, NULL);
2605 /* RETURN if there is no valid head */
2606 if (!loop_head || !loop_head_valid) {
2607 DB((dbg, LEVEL_1, "No valid loop head. Nothing done.\n"));
2610 DB((dbg, LEVEL_1, "Loophead: %N\n", loop_head));
2613 if (loop_info.branches > opt_params.max_branches) {
2614 DB((dbg, LEVEL_1, "Branches %d > allowed branches %d\n",
2615 loop_info.branches, opt_params.max_branches));
2616 ++stats.calls_limit;
2621 case loop_op_inversion:
2622 loop_inversion(irg);
2625 case loop_op_unrolling:
2630 panic("Loop optimization not implemented.");
2632 DB((dbg, LEVEL_1, " <<<< end of loop with node %ld >>>>\n",
2633 get_loop_loop_nr(loop)));
2636 /* Find innermost loops and add them to loops. */
2637 static void find_innermost_loop(ir_loop *loop)
2639 bool had_sons = false;
2640 size_t n_elements = get_loop_n_elements(loop);
2643 for (e = 0; e < n_elements; ++e) {
2644 loop_element element = get_loop_element(loop, e);
2645 if (*element.kind == k_ir_loop) {
2646 find_innermost_loop(element.son);
2652 ARR_APP1(ir_loop*, loops, loop);
2656 static void set_loop_params(void)
2658 opt_params.max_loop_size = 100;
2659 opt_params.depth_adaption = -50;
2660 opt_params.count_phi = true;
2661 opt_params.count_proj = false;
2662 opt_params.allowed_calls = 0;
2664 opt_params.max_cc_size = 5;
2667 opt_params.allow_const_unrolling = true;
2668 opt_params.allow_invar_unrolling = false;
2670 opt_params.invar_unrolling_min_size = 20;
2671 opt_params.max_unrolled_loop_size = 400;
2672 opt_params.max_branches = 9999;
2675 /* Assure preconditions are met and go through all loops. */
2676 void loop_optimization(ir_graph *irg)
2682 assure_irg_properties(irg,
2683 IR_GRAPH_PROPERTY_CONSISTENT_OUT_EDGES
2684 | IR_GRAPH_PROPERTY_CONSISTENT_OUTS
2685 | IR_GRAPH_PROPERTY_CONSISTENT_LOOPINFO);
2689 /* Reset stats for this procedure */
2693 set_current_ir_graph(irg);
2695 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_PHI_LIST);
2696 collect_phiprojs(irg);
2698 loop = get_irg_loop(irg);
2700 loops = NEW_ARR_F(ir_loop *, 0);
2701 /* List all inner loops */
2702 n_elements = get_loop_n_elements(loop);
2703 for (i = 0; i < n_elements; ++i) {
2704 loop_element element = get_loop_element(loop, i);
2705 if (*element.kind != k_ir_loop)
2707 find_innermost_loop(element.son);
2710 /* Set all links to NULL */
2711 irg_walk_graph(irg, reset_link, NULL, NULL);
2713 for (i = 0; i < ARR_LEN(loops); ++i) {
2714 ir_loop *loop = loops[i];
2718 /* Analyze and handle loop */
2719 init_analyze(irg, loop);
2721 /* Copied blocks do not have their phi list yet */
2722 collect_phiprojs(irg);
2724 /* Set links to NULL
2725 * TODO Still necessary? */
2726 irg_walk_graph(irg, reset_link, NULL, NULL);
2732 ir_free_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_PHI_LIST);
2734 confirm_irg_properties(irg, IR_GRAPH_PROPERTIES_NONE);
2737 void do_loop_unrolling(ir_graph *irg)
2739 loop_op = loop_op_unrolling;
2740 loop_optimization(irg);
2743 void do_loop_inversion(ir_graph *irg)
2745 loop_op = loop_op_inversion;
2746 loop_optimization(irg);
2749 void do_loop_peeling(ir_graph *irg)
2751 loop_op = loop_op_peeling;
2752 loop_optimization(irg);
2755 ir_graph_pass_t *loop_inversion_pass(const char *name)
2757 return def_graph_pass(name ? name : "loop_inversion", do_loop_inversion);
2760 ir_graph_pass_t *loop_unroll_pass(const char *name)
2762 return def_graph_pass(name ? name : "loop_unroll", do_loop_unrolling);
2765 ir_graph_pass_t *loop_peeling_pass(const char *name)
2767 return def_graph_pass(name ? name : "loop_peeling", do_loop_peeling);
2770 void firm_init_loop_opt(void)
2772 FIRM_DBG_REGISTER(dbg, "firm.opt.loop");