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)) {
296 /* Count innerloop branches */
297 foreach_out_edge_kind(node, edge, EDGE_KIND_BLOCK) {
298 ir_node *succ = get_edge_src_irn(edge);
299 if (is_Block(succ) && is_in_loop(succ))
303 ++loop_info.branches;
305 if (node_in_loop && !pred_in_loop && loop_head_valid) {
306 ir_node *cfgpred = get_Block_cfgpred(node, i);
308 if (!is_in_loop(cfgpred)) {
309 DB((dbg, LEVEL_5, "potential head %+F because inloop and pred %+F not inloop\n",
311 /* another head? We do not touch this. */
312 if (loop_head && loop_head != node) {
313 loop_head_valid = false;
323 /* Finds all edges with users outside of the loop
324 * and definition inside the loop. */
325 static void get_loop_entries(ir_node *node, void *env)
327 unsigned node_in_loop, pred_in_loop;
331 arity = get_irn_arity(node);
332 for (i = 0; i < arity; ++i) {
333 ir_node *pred = get_irn_n(node, i);
335 pred_in_loop = is_in_loop(pred);
336 node_in_loop = is_in_loop(node);
338 if (pred_in_loop && !node_in_loop) {
343 ARR_APP1(entry_edge, loop_entries, entry);
349 static ir_node *ssa_second_def;
350 static ir_node *ssa_second_def_block;
353 * Walks the graph bottom up, searching for definitions and creates phis.
355 static ir_node *search_def_and_create_phis(ir_node *block, ir_mode *mode, int first)
359 ir_graph *irg = get_irn_irg(block);
363 DB((dbg, LEVEL_5, "ssa search_def_and_create_phis: block %N\n", block));
365 /* Prevents creation of phi that would be bad anyway.
366 * Dead and bad blocks. */
367 if (get_irn_arity(block) < 1 || is_Bad(block)) {
368 DB((dbg, LEVEL_5, "ssa bad %N\n", block));
369 return new_r_Bad(irg, mode);
372 if (block == ssa_second_def_block && !first) {
373 DB((dbg, LEVEL_5, "ssa found second definition: use second def %N\n", ssa_second_def));
374 return ssa_second_def;
377 /* already processed this block? */
378 if (irn_visited(block)) {
379 ir_node *value = (ir_node *) get_irn_link(block);
380 DB((dbg, LEVEL_5, "ssa already visited: use linked %N\n", value));
384 assert(block != get_irg_start_block(irg));
386 /* a Block with only 1 predecessor needs no Phi */
387 n_cfgpreds = get_Block_n_cfgpreds(block);
388 if (n_cfgpreds == 1) {
389 ir_node *pred_block = get_Block_cfgpred_block(block, 0);
392 DB((dbg, LEVEL_5, "ssa 1 pred: walk pred %N\n", pred_block));
394 value = search_def_and_create_phis(pred_block, mode, 0);
395 set_irn_link(block, value);
396 mark_irn_visited(block);
401 /* create a new Phi */
402 NEW_ARR_A(ir_node*, in, n_cfgpreds);
403 for (i = 0; i < n_cfgpreds; ++i)
404 in[i] = new_r_Dummy(irg, mode);
406 phi = new_r_Phi(block, n_cfgpreds, in, mode);
407 /* Important: always keep block phi list up to date. */
408 add_Block_phi(block, phi);
409 DB((dbg, LEVEL_5, "ssa phi creation: link new phi %N to block %N\n", phi, block));
410 set_irn_link(block, phi);
411 mark_irn_visited(block);
413 /* set Phi predecessors */
414 for (i = 0; i < n_cfgpreds; ++i) {
416 ir_node *pred_block = get_Block_cfgpred_block(block, i);
417 assert(pred_block != NULL);
418 pred_val = search_def_and_create_phis(pred_block, mode, 0);
420 assert(pred_val != NULL);
422 DB((dbg, LEVEL_5, "ssa phi pred:phi %N, pred %N\n", phi, pred_val));
423 set_irn_n(phi, i, pred_val);
431 * Given a set of values this function constructs SSA-form for the users of the
432 * first value (the users are determined through the out-edges of the value).
433 * Works without using the dominance tree.
435 static void construct_ssa(ir_node *orig_block, ir_node *orig_val,
436 ir_node *second_block, ir_node *second_val)
441 assert(orig_block && orig_val && second_block && second_val &&
442 "no parameter of construct_ssa may be NULL");
444 if (orig_val == second_val)
447 irg = get_irn_irg(orig_val);
449 ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED);
450 inc_irg_visited(irg);
452 mode = get_irn_mode(orig_val);
453 set_irn_link(orig_block, orig_val);
454 mark_irn_visited(orig_block);
456 ssa_second_def_block = second_block;
457 ssa_second_def = second_val;
459 /* Only fix the users of the first, i.e. the original node */
460 foreach_out_edge_safe(orig_val, edge) {
461 ir_node *user = get_edge_src_irn(edge);
462 int j = get_edge_src_pos(edge);
463 ir_node *user_block = get_nodes_block(user);
470 DB((dbg, LEVEL_5, "original user %N\n", user));
473 ir_node *pred_block = get_Block_cfgpred_block(user_block, j);
474 newval = search_def_and_create_phis(pred_block, mode, 1);
476 newval = search_def_and_create_phis(user_block, mode, 1);
478 if (newval != user && !is_Bad(newval))
479 set_irn_n(user, j, newval);
482 ir_free_resources(irg, IR_RESOURCE_IRN_VISITED);
486 /***** Unrolling Helper Functions *****/
488 /* Assign the copy with index nr to node n */
489 static void set_unroll_copy(ir_node *n, int nr, ir_node *cp)
491 unrolling_node_info *info;
492 assert(nr != 0 && "0 reserved");
494 info = ir_nodemap_get(unrolling_node_info, &map, n);
496 ir_node **const arr = NEW_ARR_DZ(ir_node*, &obst, unroll_nr);
498 info = OALLOCZ(&obst, unrolling_node_info);
500 ir_nodemap_insert(&map, n, info);
505 info->copies[nr] = cp;
508 /* Returns a nodes copy if it exists, else NULL. */
509 static ir_node *get_unroll_copy(ir_node *n, int nr)
512 unrolling_node_info *info = ir_nodemap_get(unrolling_node_info, &map, n);
516 cp = info->copies[nr];
521 /***** Inversion Helper Functions *****/
523 /* Sets copy cp of node n. */
524 static void set_inversion_copy(ir_node *n, ir_node *cp)
526 ir_nodemap_insert(&map, n, cp);
529 /* Getter of copy of n for inversion */
530 static ir_node *get_inversion_copy(ir_node *n)
532 ir_node *cp = ir_nodemap_get(ir_node, &map, n);
536 /* Resets block mark for given node. For use with walker */
537 static void reset_block_mark(ir_node *node, void * env)
542 set_Block_mark(node, 0);
545 /* Returns mark of node, or its block if node is not a block.
546 * Used in this context to determine if node is in the condition chain. */
547 static bool is_nodes_block_marked(const ir_node* node)
549 return get_Block_mark(get_block_const(node));
552 /* Extends a nodes ins by node new.
553 * NOTE: This is slow if a node n needs to be extended more than once. */
554 static void extend_irn(ir_node *n, ir_node *newnode, bool new_is_backedge)
557 int arity = get_irn_arity(n);
558 int new_arity = arity + 1;
559 ir_node **ins = XMALLOCN(ir_node*, new_arity);
560 bool *bes = XMALLOCN(bool, new_arity);
563 /* Bes are important!
564 * Another way would be recreating the looptree,
565 * but after that we cannot distinguish already processed loops
566 * from not yet processed ones. */
568 for(i = 0; i < arity; ++i) {
569 bes[i] = is_backedge(n, i);
571 bes[i] = new_is_backedge;
574 for(i = 0; i < arity; ++i) {
575 ins[i] = get_irn_n(n, i);
579 set_irn_in(n, new_arity, ins);
583 for(i = 0; i < new_arity; ++i) {
592 /* Extends a block by a copy of its pred at pos,
593 * fixing also the phis in the same way. */
594 static void extend_ins_by_copy(ir_node *block, int pos)
598 assert(is_Block(block));
600 /* Extend block by copy of definition at pos */
601 ir_node *const pred = get_Block_cfgpred(block, pos);
602 new_in = get_inversion_copy(pred);
603 DB((dbg, LEVEL_5, "Extend block %N by %N cp of %N\n", block, new_in, pred));
604 extend_irn(block, new_in, false);
606 /* Extend block phis by copy of definition at pos */
607 for_each_phi(block, phi) {
610 pred = get_irn_n(phi, pos);
611 cp = get_inversion_copy(pred);
612 /* If the phis in is not in the condition chain (eg. a constant),
613 * there is no copy. */
619 DB((dbg, LEVEL_5, "Extend phi %N by %N cp of %N\n", phi, new_in, pred));
620 extend_irn(phi, new_in, false);
624 /* Returns the number of blocks backedges. With or without alien bes. */
625 static int get_backedge_n(ir_node *block, bool with_alien)
628 int const arity = get_Block_n_cfgpreds(block);
629 for (int i = 0; i < arity; ++i) {
630 ir_node *const pred = get_Block_cfgpred(block, i);
631 if (is_backedge(block, i) && (with_alien || is_in_loop(pred)))
637 /* Returns a raw copy of the given node.
638 * Attributes are kept/set according to the needs of loop inversion. */
639 static ir_node *copy_node(ir_node *node)
644 cp = exact_copy(node);
645 arity = get_irn_arity(node);
647 /* Keep backedge info */
648 for (i = 0; i < arity; ++i) {
649 if (is_backedge(node, i))
654 set_Block_mark(cp, 0);
662 * This walker copies all walked nodes.
663 * If the walk_condition is true for a node, it is copied.
664 * All nodes node_info->copy have to be NULL prior to every walk.
665 * Order of ins is important for later usage.
667 static void copy_walk(ir_node *node, walker_condition *walk_condition,
676 * break condition and cycle resolver, creating temporary node copies
678 if (irn_visited(node)) {
679 /* Here we rely on nodestate's copy being initialized with NULL */
680 DB((dbg, LEVEL_5, "copy_walk: We have already visited %N\n", node));
681 if (get_inversion_copy(node) == NULL) {
682 cp = copy_node(node);
683 set_inversion_copy(node, cp);
685 DB((dbg, LEVEL_5, "The TEMP copy of %N is created %N\n", node, cp));
691 mark_irn_visited(node);
693 if (!is_Block(node)) {
694 ir_node *pred = get_nodes_block(node);
695 if (walk_condition(pred))
696 DB((dbg, LEVEL_5, "walk block %N\n", pred));
697 copy_walk(pred, walk_condition, set_loop);
700 arity = get_irn_arity(node);
702 NEW_ARR_A(ir_node *, cpin, arity);
704 for (i = 0; i < arity; ++i) {
705 ir_node *pred = get_irn_n(node, i);
707 if (walk_condition(pred)) {
708 DB((dbg, LEVEL_5, "walk node %N\n", pred));
709 copy_walk(pred, walk_condition, set_loop);
710 cpin[i] = get_inversion_copy(pred);
711 DB((dbg, LEVEL_5, "copy of %N gets new in %N which is copy of %N\n",
712 node, get_inversion_copy(pred), pred));
718 /* copy node / finalize temp node */
719 if (get_inversion_copy(node) == NULL) {
720 /* No temporary copy existent */
721 cp = copy_node(node);
722 set_inversion_copy(node, cp);
723 DB((dbg, LEVEL_5, "The FINAL copy of %N is CREATED %N\n", node, cp));
725 /* temporary copy is existent but without correct ins */
726 cp = get_inversion_copy(node);
727 DB((dbg, LEVEL_5, "The FINAL copy of %N is EXISTENT %N\n", node, cp));
730 if (!is_Block(node)) {
731 ir_node *cpblock = get_inversion_copy(get_nodes_block(node));
733 set_nodes_block(cp, cpblock );
735 add_Block_phi(cpblock, cp);
738 /* Keeps phi list of temporary node. */
739 set_irn_in(cp, ARR_LEN(cpin), cpin);
743 * This walker copies all walked nodes.
744 * If the walk_condition is true for a node, it is copied.
745 * All nodes node_info->copy have to be NULL prior to every walk.
746 * Order of ins is important for later usage.
747 * Takes copy_index, to phase-link copy at specific index.
749 static void copy_walk_n(ir_node *node, walker_condition *walk_condition,
758 * break condition and cycle resolver, creating temporary node copies
760 if (irn_visited(node)) {
761 /* Here we rely on nodestate's copy being initialized with NULL */
762 DB((dbg, LEVEL_5, "copy_walk: We have already visited %N\n", node));
763 if (get_unroll_copy(node, copy_index) == NULL) {
766 set_unroll_copy(node, copy_index, u);
767 DB((dbg, LEVEL_5, "The TEMP unknown of %N is created %N\n", node, u));
773 mark_irn_visited(node);
775 if (!is_Block(node)) {
776 ir_node *block = get_nodes_block(node);
777 if (walk_condition(block))
778 DB((dbg, LEVEL_5, "walk block %N\n", block));
779 copy_walk_n(block, walk_condition, copy_index);
782 arity = get_irn_arity(node);
783 NEW_ARR_A(ir_node *, cpin, arity);
785 for (i = 0; i < arity; ++i) {
786 ir_node *pred = get_irn_n(node, i);
788 if (walk_condition(pred)) {
789 DB((dbg, LEVEL_5, "walk node %N\n", pred));
790 copy_walk_n(pred, walk_condition, copy_index);
791 cpin[i] = get_unroll_copy(pred, copy_index);
797 /* copy node / finalize temp node */
798 cp = get_unroll_copy(node, copy_index);
799 if (cp == NULL || is_Unknown(cp)) {
800 cp = copy_node(node);
801 set_unroll_copy(node, copy_index, cp);
802 DB((dbg, LEVEL_5, "The FINAL copy of %N is CREATED %N\n", node, cp));
804 /* temporary copy is existent but without correct ins */
805 cp = get_unroll_copy(node, copy_index);
806 DB((dbg, LEVEL_5, "The FINAL copy of %N is EXISTENT %N\n", node, cp));
809 if (!is_Block(node)) {
810 ir_node *cpblock = get_unroll_copy(get_nodes_block(node), copy_index);
812 set_nodes_block(cp, cpblock );
814 add_Block_phi(cpblock, cp);
817 /* Keeps phi list of temporary node. */
818 set_irn_in(cp, ARR_LEN(cpin), cpin);
821 /* Removes alle Blocks with non marked predecessors from the condition chain. */
822 static void unmark_not_allowed_cc_blocks(void)
824 size_t blocks = ARR_LEN(cc_blocks);
827 for(i = 0; i < blocks; ++i) {
828 ir_node *block = cc_blocks[i];
830 /* Head is an exception. */
831 if (block == loop_head)
834 int const arity = get_Block_n_cfgpreds(block);
835 for (int a = 0; a < arity; ++a) {
836 if (!is_nodes_block_marked(get_Block_cfgpred(block, a))) {
837 set_Block_mark(block, 0);
838 --inversion_blocks_in_cc;
839 DB((dbg, LEVEL_5, "Removed %N from cc (blocks in cc %d)\n",
840 block, inversion_blocks_in_cc));
848 /* Unmarks all cc blocks using cc_blocks except head.
849 * TODO: invert head for unrolling? */
850 static void unmark_cc_blocks(void)
852 size_t blocks = ARR_LEN(cc_blocks);
855 for(i = 0; i < blocks; ++i) {
856 ir_node *block = cc_blocks[i];
858 /* TODO Head is an exception. */
859 /*if (block != loop_head)*/
860 set_Block_mark(block, 0);
862 /*inversion_blocks_in_cc = 1;*/
863 inversion_blocks_in_cc = 0;
866 loop_info.cc_size = 0;
870 * Populates head_entries with (node, pred_pos) tuple
871 * whereas the node's pred at pred_pos is in the cc but not the node itself.
872 * Also finds df loops inside the cc.
873 * Head and condition chain blocks have been marked previously.
875 static void get_head_outs(ir_node *node, void *env)
878 int arity = get_irn_arity(node);
881 for (i = 0; i < arity; ++i) {
882 if (!is_nodes_block_marked(node) && is_nodes_block_marked(get_irn_n(node, i))) {
886 /* Saving also predecessor seems redundant, but becomes
887 * necessary when changing position of it, before
888 * dereferencing it.*/
889 entry.pred = get_irn_n(node, i);
890 ARR_APP1(entry_edge, cur_head_outs, entry);
894 arity = get_irn_arity(loop_head);
896 /* Find df loops inside the cc */
897 if (is_Phi(node) && get_nodes_block(node) == loop_head) {
898 for (i = 0; i < arity; ++i) {
899 if (is_own_backedge(loop_head, i)) {
900 if (is_nodes_block_marked(get_irn_n(node, i))) {
904 entry.pred = get_irn_n(node, i);
905 ARR_APP1(entry_edge, head_df_loop, entry);
906 DB((dbg, LEVEL_5, "Found incc assignment node %N @%d is pred %N, graph %N %N\n",
907 node, i, entry.pred, current_ir_graph, get_irg_start_block(current_ir_graph)));
915 * Find condition chains, and add them to be inverted.
916 * A block belongs to the chain if a condition branches out of the loop.
917 * (Some blocks need to be removed once again.)
918 * Returns 1 if the given block belongs to the condition chain.
920 static void find_condition_chain(ir_node *block)
924 bool jmp_only = true;
925 unsigned nodes_n = 0;
927 mark_irn_visited(block);
929 DB((dbg, LEVEL_5, "condition_chains for block %N\n", block));
932 foreach_out_edge_kind(block, edge, EDGE_KIND_NORMAL) {
936 /* Check if node count would exceed maximum cc size.
938 * This is not optimal, as we search depth-first and break here,
939 * continuing with another subtree. */
940 if (loop_info.cc_size + nodes_n > opt_params.max_cc_size) {
941 set_Block_mark(block, 0);
945 /* Check if block only has a jmp instruction. */
946 foreach_out_edge(block, edge) {
947 ir_node *src = get_edge_src_irn(edge);
949 if (!is_Block(src) && !is_Jmp(src)) {
954 /* Check cf outs if one is leaving the loop,
955 * or if this node has a backedge. */
956 foreach_block_succ(block, edge) {
957 ir_node *src = get_edge_src_irn(edge);
958 int pos = get_edge_src_pos(edge);
960 if (!is_in_loop(src))
963 /* Inverting blocks with backedge outs leads to a cf edge
964 * from the inverted head, into the inverted head (skipping the body).
965 * As the body becomes the new loop head,
966 * this would introduce another loop in the existing loop.
967 * This loop inversion cannot cope with this case. */
968 if (is_backedge(src, pos)) {
974 /* We need all predecessors to already belong to the condition chain.
975 * Example of wrong case: * == in cc
986 /* Collect blocks containing only a Jmp.
987 * Do not collect blocks with backedge outs. */
988 if ((jmp_only || mark) && !has_be) {
989 set_Block_mark(block, 1);
990 ++inversion_blocks_in_cc;
991 loop_info.cc_size += nodes_n;
992 DB((dbg, LEVEL_5, "block %N is part of condition chain\n", block));
993 ARR_APP1(ir_node *, cc_blocks, block);
995 set_Block_mark(block, 0);
998 foreach_block_succ(block, edge) {
999 ir_node *src = get_edge_src_irn( edge );
1001 if (is_in_loop(src) && ! irn_visited(src))
1002 find_condition_chain(src);
1007 * Rewires the copied condition chain. Removes backedges
1008 * as this condition chain is prior to the loop.
1009 * Copy of loop_head must have phi list and old (unfixed) backedge info of the loop head.
1010 * (loop_head is already fixed, we cannot rely on it.)
1012 static void fix_copy_inversion(void)
1017 ir_node *phi, *next;
1018 ir_node *head_cp = get_inversion_copy(loop_head);
1019 ir_graph *irg = get_irn_irg(head_cp);
1020 int arity = get_irn_arity(head_cp);
1021 int backedges = get_backedge_n(head_cp, false);
1022 int new_arity = arity - backedges;
1026 NEW_ARR_A(ir_node *, ins, new_arity);
1029 /* Remove block backedges */
1030 for(i = 0; i < arity; ++i) {
1031 if (!is_backedge(head_cp, i))
1032 ins[pos++] = get_irn_n(head_cp, i);
1035 new_head = new_r_Block(irg, new_arity, ins);
1037 phis = NEW_ARR_F(ir_node *, 0);
1039 for_each_phi_safe(get_Block_phis(head_cp), phi, next) {
1041 NEW_ARR_A(ir_node *, ins, new_arity);
1043 for(i = 0; i < arity; ++i) {
1044 if (!is_backedge(head_cp, i))
1045 ins[pos++] = get_irn_n(phi, i);
1047 new_phi = new_rd_Phi(get_irn_dbg_info(phi),
1048 new_head, new_arity, ins,
1050 ARR_APP1(ir_node *, phis, new_phi);
1054 for_each_phi_safe(get_Block_phis(head_cp), phi, next) {
1055 exchange(phi, phis[pos++]);
1058 exchange(head_cp, new_head);
1064 /* Puts the original condition chain at the end of the loop,
1065 * subsequently to the body.
1066 * Relies on block phi list and correct backedges.
1068 static void fix_head_inversion(void)
1072 ir_node *phi, *next;
1074 ir_graph *irg = get_irn_irg(loop_head);
1075 int arity = get_irn_arity(loop_head);
1076 int backedges = get_backedge_n(loop_head, false);
1077 int new_arity = backedges;
1081 NEW_ARR_A(ir_node *, ins, new_arity);
1084 /* Keep only backedges */
1085 for(i = 0; i < arity; ++i) {
1086 if (is_own_backedge(loop_head, i))
1087 ins[pos++] = get_irn_n(loop_head, i);
1090 new_head = new_r_Block(irg, new_arity, ins);
1092 phis = NEW_ARR_F(ir_node *, 0);
1094 for_each_phi(loop_head, phi) {
1096 DB((dbg, LEVEL_5, "Fixing phi %N of loop head\n", phi));
1098 NEW_ARR_A(ir_node *, ins, new_arity);
1101 for (i = 0; i < arity; ++i) {
1102 ir_node *pred = get_irn_n(phi, i);
1104 if (is_own_backedge(loop_head, i)) {
1105 /* If assignment is in the condition chain,
1106 * we need to create a phi in the new loop head.
1107 * This can only happen for df, not cf. See find_condition_chains. */
1108 /*if (is_nodes_block_marked(pred)) {
1116 new_phi = new_rd_Phi(get_irn_dbg_info(phi),
1117 new_head, new_arity, ins,
1120 ARR_APP1(ir_node *, phis, new_phi);
1122 DB((dbg, LEVEL_5, "fix inverted head should exch %N by %N (pos %d)\n", phi, new_phi, pos ));
1126 for_each_phi_safe(get_Block_phis(loop_head), phi, next) {
1127 DB((dbg, LEVEL_5, "fix inverted exch phi %N by %N\n", phi, phis[pos]));
1128 if (phis[pos] != phi)
1129 exchange(phi, phis[pos++]);
1134 DB((dbg, LEVEL_5, "fix inverted head exch head block %N by %N\n", loop_head, new_head));
1135 exchange(loop_head, new_head);
1138 /* Does the loop inversion. */
1139 static void inversion_walk(ir_graph *irg, entry_edge *head_entries)
1144 * The order of rewiring bottom-up is crucial.
1145 * Any change of the order leads to lost information that would be needed later.
1148 ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED);
1150 /* 1. clone condition chain */
1151 inc_irg_visited(irg);
1153 for (i = 0; i < ARR_LEN(head_entries); ++i) {
1154 entry_edge entry = head_entries[i];
1155 ir_node *pred = get_irn_n(entry.node, entry.pos);
1157 DB((dbg, LEVEL_5, "\nInit walk block %N\n", pred));
1159 copy_walk(pred, is_nodes_block_marked, cur_loop);
1162 ir_free_resources(irg, IR_RESOURCE_IRN_VISITED);
1164 /* 2. Extends the head control flow successors ins
1165 * with the definitions of the copied head node. */
1166 for (i = 0; i < ARR_LEN(head_entries); ++i) {
1167 entry_edge head_out = head_entries[i];
1169 if (is_Block(head_out.node))
1170 extend_ins_by_copy(head_out.node, head_out.pos);
1173 /* 3. construct_ssa for users of definitions in the condition chain,
1174 * as there is now a second definition. */
1175 for (i = 0; i < ARR_LEN(head_entries); ++i) {
1176 entry_edge head_out = head_entries[i];
1178 /* Ignore keepalives */
1179 if (is_End(head_out.node))
1182 /* Construct ssa for assignments in the condition chain. */
1183 if (!is_Block(head_out.node)) {
1184 ir_node *pred, *cppred, *block, *cpblock;
1186 pred = head_out.pred;
1187 cppred = get_inversion_copy(pred);
1188 block = get_nodes_block(pred);
1189 cpblock = get_nodes_block(cppred);
1190 construct_ssa(block, pred, cpblock, cppred);
1195 * If there is an assignment in the condition chain
1196 * with a user also in the condition chain,
1197 * the dominance frontier is in the new loop head.
1198 * The dataflow loop is completely in the condition chain.
1203 * Phi_cp | | copied condition chain
1207 * Phi* | | new loop head with newly created phi.
1209 * Phi | | original, inverted condition chain
1214 for (i = 0; i < ARR_LEN(head_df_loop); ++i) {
1215 entry_edge head_out = head_df_loop[i];
1217 /* Construct ssa for assignments in the condition chain. */
1218 ir_node *pred, *cppred, *block, *cpblock;
1220 pred = head_out.pred;
1221 cppred = get_inversion_copy(pred);
1222 assert(cppred && pred);
1223 block = get_nodes_block(pred);
1224 cpblock = get_nodes_block(cppred);
1225 construct_ssa(block, pred, cpblock, cppred);
1228 /* 4. Remove the ins which are no backedges from the original condition chain
1229 * as the cc is now subsequent to the body. */
1230 fix_head_inversion();
1232 /* 5. Remove the backedges of the copied condition chain,
1233 * because it is going to be the new 'head' in advance to the loop. */
1234 fix_copy_inversion();
1238 /* Performs loop inversion of cur_loop if possible and reasonable. */
1239 static void loop_inversion(ir_graph *irg)
1242 unsigned max_loop_nodes = opt_params.max_loop_size;
1243 unsigned max_loop_nodes_adapted;
1244 int depth_adaption = opt_params.depth_adaption;
1246 bool do_inversion = true;
1248 /* Depth of 0 is the procedure and 1 a topmost loop. */
1249 loop_depth = get_loop_depth(cur_loop) - 1;
1251 /* Calculating in per mil. */
1252 max_loop_nodes_adapted = get_max_nodes_adapted(loop_depth);
1254 DB((dbg, LEVEL_1, "max_nodes: %d\nmax_nodes_adapted %d at depth of %d (adaption %d)\n",
1255 max_loop_nodes, max_loop_nodes_adapted, loop_depth, depth_adaption));
1257 if (loop_info.nodes == 0)
1260 if (loop_info.nodes > max_loop_nodes) {
1261 /* Only for stats */
1262 DB((dbg, LEVEL_1, "Nodes %d > allowed nodes %d\n",
1263 loop_info.nodes, loop_depth, max_loop_nodes));
1266 /* Adaption might change it */
1269 /* Limit processing to loops smaller than given parameter. */
1270 if (loop_info.nodes > max_loop_nodes_adapted) {
1271 DB((dbg, LEVEL_1, "Nodes %d > allowed nodes (depth %d adapted) %d\n",
1272 loop_info.nodes, loop_depth, max_loop_nodes_adapted));
1273 ++stats.too_large_adapted;
1277 if (loop_info.calls > opt_params.allowed_calls) {
1278 DB((dbg, LEVEL_1, "Calls %d > allowed calls %d\n",
1279 loop_info.calls, opt_params.allowed_calls));
1280 ++stats.calls_limit;
1284 /*inversion_head_node_limit = INT_MAX;*/
1285 ir_reserve_resources(irg, IR_RESOURCE_BLOCK_MARK);
1287 /* Reset block marks.
1288 * We use block marks to flag blocks of the original condition chain. */
1289 irg_walk_graph(irg, reset_block_mark, NULL, NULL);
1291 /*loop_info.blocks = get_loop_n_blocks(cur_loop);*/
1292 cond_chain_entries = NEW_ARR_F(entry_edge, 0);
1293 head_df_loop = NEW_ARR_F(entry_edge, 0);
1295 /*head_inversion_node_count = 0;*/
1296 inversion_blocks_in_cc = 0;
1298 /* Use phase to keep copy of nodes from the condition chain. */
1299 ir_nodemap_init(&map, irg);
1300 obstack_init(&obst);
1302 /* Search for condition chains and temporarily save the blocks in an array. */
1303 cc_blocks = NEW_ARR_F(ir_node *, 0);
1304 inc_irg_visited(irg);
1305 find_condition_chain(loop_head);
1307 unmark_not_allowed_cc_blocks();
1308 DEL_ARR_F(cc_blocks);
1310 /* Condition chain too large.
1311 * Loop should better be small enough to fit into the cache. */
1312 /* TODO Of course, we should take a small enough cc in the first place,
1313 * which is not that simple. (bin packing) */
1314 if (loop_info.cc_size > opt_params.max_cc_size) {
1315 ++stats.cc_limit_reached;
1317 do_inversion = false;
1319 /* Unmark cc blocks except the head.
1320 * Invert head only for possible unrolling. */
1324 /* We also catch endless loops here,
1325 * because they do not have a condition chain. */
1326 if (inversion_blocks_in_cc < 1) {
1327 do_inversion = false;
1329 "Loop contains %d (less than 1) invertible blocks => No Inversion done.\n",
1330 inversion_blocks_in_cc));
1334 cur_head_outs = NEW_ARR_F(entry_edge, 0);
1336 /* Get all edges pointing into the condition chain. */
1337 irg_walk_graph(irg, get_head_outs, NULL, NULL);
1339 /* Do the inversion */
1340 inversion_walk(irg, cur_head_outs);
1342 DEL_ARR_F(cur_head_outs);
1344 /* Duplicated blocks changed doms */
1345 clear_irg_properties(irg, IR_GRAPH_PROPERTY_CONSISTENT_DOMINANCE
1346 | IR_GRAPH_PROPERTY_CONSISTENT_LOOPINFO);
1352 obstack_free(&obst, NULL);
1353 ir_nodemap_destroy(&map);
1354 DEL_ARR_F(cond_chain_entries);
1355 DEL_ARR_F(head_df_loop);
1357 ir_free_resources(irg, IR_RESOURCE_BLOCK_MARK);
1360 /* Fix the original loop_heads ins for invariant unrolling case. */
1361 static void unrolling_fix_loop_head_inv(void)
1365 ir_node *proj = new_Proj(loop_info.duff_cond, mode_X, 0);
1366 ir_node *head_pred = get_irn_n(loop_head, loop_info.be_src_pos);
1367 ir_node *loop_condition = get_unroll_copy(head_pred, unroll_nr - 1);
1369 /* Original loop_heads ins are:
1370 * duff block and the own backedge */
1372 ins[0] = loop_condition;
1374 set_irn_in(loop_head, 2, ins);
1375 DB((dbg, LEVEL_4, "Rewire ins of block loophead %N to pred %N and duffs entry %N \n" , loop_head, ins[0], ins[1]));
1377 for_each_phi(loop_head, phi) {
1378 ir_node *pred = get_irn_n(phi, loop_info.be_src_pos);
1379 /* TODO we think it is a phi, but for Mergesort it is not the case.*/
1381 ir_node *last_pred = get_unroll_copy(pred, unroll_nr - 1);
1384 ins[1] = (ir_node*)get_irn_link(phi);
1385 set_irn_in(phi, 2, ins);
1386 DB((dbg, LEVEL_4, "Rewire ins of loophead phi %N to pred %N and duffs entry %N \n" , phi, ins[0], ins[1]));
1390 /* Removes previously created phis with only 1 in. */
1391 static void correct_phis(ir_node *node, void *env)
1395 if (is_Phi(node) && get_irn_arity(node) == 1) {
1399 in[0] = get_irn_n(node, 0);
1401 exch = new_rd_Phi(get_irn_dbg_info(node),
1402 get_nodes_block(node), 1, in,
1403 get_irn_mode(node));
1405 exchange(node, exch);
1409 /* Unrolling: Rewire floating copies. */
1410 static void place_copies(int copies)
1412 ir_node *loophead = loop_head;
1415 int be_src_pos = loop_info.be_src_pos;
1417 /* Serialize loops by fixing their head ins.
1418 * Processed are the copies.
1419 * The original loop is done after that, to keep backedge infos. */
1420 for (c = 0; c < copies; ++c) {
1421 ir_node *upper = get_unroll_copy(loophead, c);
1422 ir_node *lower = get_unroll_copy(loophead, c + 1);
1424 ir_node *topmost_be_block = get_nodes_block(get_irn_n(loophead, be_src_pos));
1426 /* Important: get the preds first and then their copy. */
1427 ir_node *upper_be_block = get_unroll_copy(topmost_be_block, c);
1428 ir_node *new_jmp = new_r_Jmp(upper_be_block);
1429 DB((dbg, LEVEL_5, " place_copies upper %N lower %N\n", upper, lower));
1431 DB((dbg, LEVEL_5, "topmost be block %N \n", topmost_be_block));
1433 if (loop_info.unroll_kind == constant) {
1436 set_irn_in(lower, 1, ins);
1438 for_each_phi(loophead, phi) {
1439 ir_node *topmost_def = get_irn_n(phi, be_src_pos);
1440 ir_node *upper_def = get_unroll_copy(topmost_def, c);
1441 ir_node *lower_phi = get_unroll_copy(phi, c + 1);
1443 /* It is possible, that the value used
1444 * in the OWN backedge path is NOT defined in this loop. */
1445 if (is_in_loop(topmost_def))
1448 ins[0] = topmost_def;
1450 set_irn_in(lower_phi, 1, ins);
1451 /* Need to replace phis with 1 in later. */
1454 /* Invariant case */
1455 /* Every node has 2 ins. One from the duff blocks
1456 * and one from the previously unrolled loop. */
1458 /* Calculate corresponding projection of mod result for this copy c */
1459 ir_node *proj = new_Proj(loop_info.duff_cond, mode_X, unroll_nr - c - 1);
1460 DB((dbg, LEVEL_4, "New duff proj %N\n" , proj));
1464 set_irn_in(lower, 2, ins);
1465 DB((dbg, LEVEL_4, "Rewire ins of Block %N to pred %N and duffs entry %N \n" , lower, ins[0], ins[1]));
1467 for_each_phi(loophead, phi) {
1468 ir_node *topmost_phi_pred = get_irn_n(phi, be_src_pos);
1469 ir_node *upper_phi_pred;
1473 lower_phi = get_unroll_copy(phi, c + 1);
1474 duff_phi = (ir_node*)get_irn_link(phi);
1475 DB((dbg, LEVEL_4, "DD Link of %N is %N\n" , phi, duff_phi));
1478 if (is_in_loop(topmost_phi_pred)) {
1479 upper_phi_pred = get_unroll_copy(topmost_phi_pred, c);
1481 upper_phi_pred = topmost_phi_pred;
1484 ins[0] = upper_phi_pred;
1486 set_irn_in(lower_phi, 2, ins);
1487 DB((dbg, LEVEL_4, "Rewire ins of %N to pred %N and duffs entry %N \n" , lower_phi, ins[0], ins[1]));
1492 /* Reconnect last copy. */
1493 for (i = 0; i < ARR_LEN(loop_entries); ++i) {
1494 entry_edge edge = loop_entries[i];
1495 /* Last copy is at the bottom */
1496 ir_node *new_pred = get_unroll_copy(edge.pred, copies);
1497 set_irn_n(edge.node, edge.pos, new_pred);
1500 /* Fix original loops head.
1501 * Done in the end, as ins and be info were needed before. */
1502 if (loop_info.unroll_kind == constant) {
1504 ir_node *head_pred = get_irn_n(loop_head, be_src_pos);
1505 ir_node *loop_condition = get_unroll_copy(head_pred, unroll_nr - 1);
1507 set_irn_n(loop_head, loop_info.be_src_pos, loop_condition);
1509 for_each_phi(loop_head, phi) {
1510 ir_node *pred = get_irn_n(phi, be_src_pos);
1513 /* It is possible, that the value used
1514 * in the OWN backedge path is NOT assigned in this loop. */
1515 if (is_in_loop(pred))
1516 last_pred = get_unroll_copy(pred, copies);
1519 set_irn_n(phi, be_src_pos, last_pred);
1523 unrolling_fix_loop_head_inv();
1527 /* Copies the cur_loop several times. */
1528 static void copy_loop(entry_edge *cur_loop_outs, int copies)
1532 ir_reserve_resources(current_ir_graph, IR_RESOURCE_IRN_VISITED);
1534 for (c = 0; c < copies; ++c) {
1537 inc_irg_visited(current_ir_graph);
1539 DB((dbg, LEVEL_5, " ### Copy_loop copy nr: %d ###\n", c));
1540 for (i = 0; i < ARR_LEN(cur_loop_outs); ++i) {
1541 entry_edge entry = cur_loop_outs[i];
1542 ir_node *pred = get_irn_n(entry.node, entry.pos);
1544 copy_walk_n(pred, is_in_loop, c + 1);
1548 ir_free_resources(current_ir_graph, IR_RESOURCE_IRN_VISITED);
1552 /* Creates a new phi from the given phi node omitting own bes,
1553 * using be_block as supplier of backedge informations. */
1554 static ir_node *clone_phis_sans_bes(ir_node *phi, ir_node *be_block, ir_node *dest_block)
1557 int arity = get_irn_arity(phi);
1561 assert(get_irn_arity(phi) == get_irn_arity(be_block));
1562 assert(is_Phi(phi));
1564 ins = NEW_ARR_F(ir_node *, arity);
1565 for (i = 0; i < arity; ++i) {
1566 if (! is_own_backedge(be_block, i)) {
1567 ins[c] = get_irn_n(phi, i);
1572 newphi = new_r_Phi(dest_block, c, ins, get_irn_mode(phi));
1574 set_irn_link(phi, newphi);
1575 DB((dbg, LEVEL_4, "Linking for duffs device %N to %N\n", phi, newphi));
1580 /* Creates a new block from the given block node omitting own bes,
1581 * using be_block as supplier of backedge informations. */
1582 static ir_node *clone_block_sans_bes(ir_node *node, ir_node *be_block)
1584 int arity = get_irn_arity(node);
1588 assert(get_irn_arity(node) == get_irn_arity(be_block));
1589 assert(is_Block(node));
1591 NEW_ARR_A(ir_node *, ins, arity);
1592 for (i = 0; i < arity; ++i) {
1593 if (! is_own_backedge(be_block, i)) {
1594 ins[c] = get_irn_n(node, i);
1599 return new_Block(c, ins);
1602 /* Creates a structure to calculate absolute value of node op.
1603 * Returns mux node with absolute value. */
1604 static ir_node *new_Abs(ir_node *op, ir_mode *mode)
1606 ir_graph *irg = get_irn_irg(op);
1607 ir_node *block = get_nodes_block(op);
1608 ir_node *zero = new_r_Const(irg, get_mode_null(mode));
1609 ir_node *cmp = new_r_Cmp(block, op, zero, ir_relation_less);
1610 ir_node *minus_op = new_r_Minus(block, op, mode);
1611 ir_node *mux = new_r_Mux(block, cmp, op, minus_op, mode);
1617 /* Creates blocks for duffs device, using previously obtained
1618 * informations about the iv.
1620 static void create_duffs_block(void)
1624 ir_node *block1, *count_block, *duff_block;
1625 ir_node *ems, *ems_mod, *ems_div, *ems_mod_proj, *cmp_null,
1626 *ems_mode_cond, *x_true, *x_false, *const_null;
1627 ir_node *true_val, *false_val;
1630 ir_node *duff_mod, *proj, *cond;
1632 ir_node *count, *correction, *unroll_c;
1633 ir_node *cmp_bad_count, *good_count, *bad_count, *count_phi, *bad_count_neg;
1636 mode = get_irn_mode(loop_info.end_val);
1637 const_null = new_Const(get_mode_null(mode));
1640 * 1. Calculate first approach to count.
1641 * Condition: (end - start) % step == 0 */
1642 block1 = clone_block_sans_bes(loop_head, loop_head);
1643 DB((dbg, LEVEL_4, "Duff block 1 %N\n", block1));
1645 /* Create loop entry phis in first duff block
1646 * as it becomes the loops preheader */
1647 for_each_phi(loop_head, phi) {
1648 /* Returns phis pred if phi would have arity 1*/
1649 ir_node *new_phi = clone_phis_sans_bes(phi, loop_head, block1);
1651 DB((dbg, LEVEL_4, "HEAD %N phi %N\n", loop_head, phi));
1652 DB((dbg, LEVEL_4, "BLOCK1 %N phi %N\n", block1, new_phi));
1655 ems = new_r_Sub(block1, loop_info.end_val, loop_info.start_val,
1656 get_irn_mode(loop_info.end_val));
1657 DB((dbg, LEVEL_4, "BLOCK1 sub %N\n", ems));
1660 ems = new_Sub(loop_info.end_val, loop_info.start_val,
1661 get_irn_mode(loop_info.end_val));
1663 DB((dbg, LEVEL_4, "mod ins %N %N\n", ems, loop_info.step));
1664 ems_mod = new_r_Mod(block1,
1669 op_pin_state_pinned);
1670 ems_div = new_r_Div(block1,
1675 op_pin_state_pinned);
1677 DB((dbg, LEVEL_4, "New module node %N\n", ems_mod));
1679 ems_mod_proj = new_r_Proj(ems_mod, mode_Iu, pn_Mod_res);
1680 cmp_null = new_r_Cmp(block1, ems_mod_proj, const_null, ir_relation_less);
1681 ems_mode_cond = new_r_Cond(block1, cmp_null);
1683 /* ems % step == 0 */
1684 x_true = new_r_Proj(ems_mode_cond, mode_X, pn_Cond_true);
1685 /* ems % step != 0 */
1686 x_false = new_r_Proj(ems_mode_cond, mode_X, pn_Cond_false);
1689 * Assures, duffs device receives a valid count.
1691 * decreasing: count < 0
1692 * increasing: count > 0
1697 count_block = new_Block(2, ins);
1698 DB((dbg, LEVEL_4, "Duff block 2 %N\n", count_block));
1701 /* Increase loop-taken-count depending on the loop condition
1702 * uses the latest iv to compare to. */
1703 if (loop_info.latest_value == 1) {
1704 /* ems % step == 0 : +0 */
1705 true_val = new_Const(get_mode_null(mode));
1706 /* ems % step != 0 : +1 */
1707 false_val = new_Const(get_mode_one(mode));
1709 ir_tarval *tv_two = new_tarval_from_long(2, mode);
1710 /* ems % step == 0 : +1 */
1711 true_val = new_Const(get_mode_one(mode));
1712 /* ems % step != 0 : +2 */
1713 false_val = new_Const(tv_two);
1719 correction = new_r_Phi(count_block, 2, ins, mode);
1721 count = new_r_Proj(ems_div, mode, pn_Div_res);
1723 /* (end - start) / step + correction */
1724 count = new_Add(count, correction, mode);
1726 /* We preconditioned the loop to be tail-controlled.
1727 * So, if count is something 'wrong' like 0,
1728 * negative/positive (depending on step direction),
1729 * we may take the loop once (tail-contr.) and leave it
1730 * to the existing condition, to break; */
1732 /* Depending on step direction, we have to check for > or < 0 */
1733 if (loop_info.decreasing == 1) {
1734 cmp_bad_count = new_r_Cmp(count_block, count, const_null,
1737 cmp_bad_count = new_r_Cmp(count_block, count, const_null,
1738 ir_relation_greater);
1741 bad_count_neg = new_r_Cond(count_block, cmp_bad_count);
1742 good_count = new_Proj(bad_count_neg, mode_X, pn_Cond_true);
1743 bad_count = new_Proj(ems_mode_cond, mode_X, pn_Cond_false);
1746 * Contains module to decide which loop to start from. */
1748 ins[0] = good_count;
1750 duff_block = new_Block(2, ins);
1751 DB((dbg, LEVEL_4, "Duff block 3 %N\n", duff_block));
1753 /* Get absolute value */
1754 ins[0] = new_Abs(count, mode);
1755 /* Manually feed the aforementioned count = 1 (bad case)*/
1756 ins[1] = new_Const(get_mode_one(mode));
1757 count_phi = new_r_Phi(duff_block, 2, ins, mode);
1759 unroll_c = new_Const(new_tarval_from_long((long)unroll_nr, mode));
1761 /* count % unroll_nr */
1762 duff_mod = new_r_Mod(duff_block,
1767 op_pin_state_pinned);
1770 proj = new_Proj(duff_mod, mode, pn_Mod_res);
1771 /* condition does NOT create itself in the block of the proj! */
1772 cond = new_r_Cond(duff_block, proj);
1774 loop_info.duff_cond = cond;
1777 /* Returns 1 if given node is not in loop,
1778 * or if it is a phi of the loop head with only loop invariant defs.
1780 static unsigned is_loop_invariant_def(ir_node *node)
1784 if (! is_in_loop(node)) {
1785 DB((dbg, LEVEL_4, "Not in loop %N\n", node));
1786 /* || is_Const(node) || is_SymConst(node)) {*/
1790 /* If this is a phi of the loophead shared by more than 1 loop,
1791 * we need to check if all defs are not in the loop. */
1794 block = get_nodes_block(node);
1796 /* To prevent unexpected situations. */
1797 if (block != loop_head) {
1801 for (i = 0; i < get_irn_arity(node); ++i) {
1802 /* Check if all bes are just loopbacks. */
1803 if (is_own_backedge(block, i) && get_irn_n(node, i) != node)
1806 DB((dbg, LEVEL_4, "invar %N\n", node));
1809 DB((dbg, LEVEL_4, "Not invar %N\n", node));
1814 /* Returns 1 if one pred of node is invariant and the other is not.
1815 * invar_pred and other are set analogously. */
1816 static unsigned get_invariant_pred(ir_node *node, ir_node **invar_pred, ir_node **other)
1818 ir_node *pred0 = get_irn_n(node, 0);
1819 ir_node *pred1 = get_irn_n(node, 1);
1824 if (is_loop_invariant_def(pred0)) {
1825 DB((dbg, LEVEL_4, "pred0 invar %N\n", pred0));
1826 *invar_pred = pred0;
1830 if (is_loop_invariant_def(pred1)) {
1831 DB((dbg, LEVEL_4, "pred1 invar %N\n", pred1));
1833 if (*invar_pred != NULL) {
1834 /* RETURN. We do not want both preds to be invariant. */
1839 *invar_pred = pred1;
1842 DB((dbg, LEVEL_4, "pred1 not invar %N\n", pred1));
1844 if (*invar_pred != NULL)
1851 /* Starts from a phi that may belong to an iv.
1852 * If an add forms a loop with iteration_phi,
1853 * and add uses a constant, 1 is returned
1854 * and 'start' as well as 'add' are sane. */
1855 static unsigned get_start_and_add(ir_node *iteration_phi, unrolling_kind_flag role)
1858 ir_node *found_add = loop_info.add;
1859 int arity = get_irn_arity(iteration_phi);
1861 DB((dbg, LEVEL_4, "Find start and add from %N\n", iteration_phi));
1863 for (i = 0; i < arity; ++i) {
1865 /* Find start_val which needs to be pred of the iteration_phi.
1866 * If start_val already known, sanity check. */
1867 if (!is_backedge(get_nodes_block(loop_info.iteration_phi), i)) {
1868 ir_node *found_start_val = get_irn_n(loop_info.iteration_phi, i);
1870 DB((dbg, LEVEL_4, "found_start_val %N\n", found_start_val));
1872 /* We already found a start_val it has to be always the same. */
1873 if (loop_info.start_val && found_start_val != loop_info.start_val)
1876 if ((role == constant) && !(is_SymConst(found_start_val) || is_Const(found_start_val)))
1878 else if((role == constant) && !(is_loop_invariant_def(found_start_val)))
1881 loop_info.start_val = found_start_val;
1884 /* The phi has to be in the loop head.
1885 * Follow all own backedges. Every value supplied from these preds of the phi
1886 * needs to origin from the same add. */
1887 if (is_own_backedge(get_nodes_block(loop_info.iteration_phi), i)) {
1888 ir_node *new_found = get_irn_n(loop_info.iteration_phi,i);
1890 DB((dbg, LEVEL_4, "is add? %N\n", new_found));
1892 if (! (is_Add(new_found) || is_Sub(new_found)) || (found_add && found_add != new_found))
1895 found_add = new_found;
1899 loop_info.add = found_add;
1905 /* Returns 1 if one pred of node is a const value and the other is not.
1906 * const_pred and other are set analogously. */
1907 static unsigned get_const_pred(ir_node *node, ir_node **const_pred, ir_node **other)
1909 ir_node *pred0 = get_irn_n(node, 0);
1910 ir_node *pred1 = get_irn_n(node, 1);
1912 DB((dbg, LEVEL_4, "Checking for constant pred of %N\n", node));
1917 /*DB((dbg, LEVEL_4, "is %N const\n", pred0));*/
1918 if (is_Const(pred0) || is_SymConst(pred0)) {
1919 *const_pred = pred0;
1923 /*DB((dbg, LEVEL_4, "is %N const\n", pred1));*/
1924 if (is_Const(pred1) || is_SymConst(pred1)) {
1925 if (*const_pred != NULL) {
1926 /* RETURN. We do not want both preds to be constant. */
1931 *const_pred = pred1;
1934 if (*const_pred == NULL)
1940 /* Returns 1 if loop exits within 2 steps of the iv.
1941 * Norm_proj means we do not exit the loop.*/
1942 static unsigned simulate_next(ir_tarval **count_tar,
1943 ir_tarval *stepped, ir_tarval *step_tar, ir_tarval *end_tar,
1944 ir_relation norm_proj)
1948 DB((dbg, LEVEL_4, "Loop taken if (stepped)%ld %s (end)%ld ",
1949 get_tarval_long(stepped),
1950 get_relation_string((norm_proj)),
1951 get_tarval_long(end_tar)));
1952 DB((dbg, LEVEL_4, "comparing latest value %d\n", loop_info.latest_value));
1954 /* If current iv does not stay in the loop,
1955 * this run satisfied the exit condition. */
1956 if (! (tarval_cmp(stepped, end_tar) & norm_proj))
1959 DB((dbg, LEVEL_4, "Result: (stepped)%ld IS %s (end)%ld\n",
1960 get_tarval_long(stepped),
1961 get_relation_string(tarval_cmp(stepped, end_tar)),
1962 get_tarval_long(end_tar)));
1965 if (is_Add(loop_info.add))
1966 next = tarval_add(stepped, step_tar);
1969 next = tarval_sub(stepped, step_tar, get_irn_mode(loop_info.end_val));
1971 DB((dbg, LEVEL_4, "Loop taken if %ld %s %ld ",
1972 get_tarval_long(next),
1973 get_relation_string(norm_proj),
1974 get_tarval_long(end_tar)));
1975 DB((dbg, LEVEL_4, "comparing latest value %d\n", loop_info.latest_value));
1977 /* Increase steps. */
1978 *count_tar = tarval_add(*count_tar, get_tarval_one(get_tarval_mode(*count_tar)));
1980 /* Next has to fail the loop condition, or we will never exit. */
1981 if (! (tarval_cmp(next, end_tar) & norm_proj))
1987 /* Check if loop meets requirements for a 'simple loop':
1988 * - Exactly one cf out
1990 * - Max nodes after unrolling
1994 * Returns Projection of cmp node or NULL; */
1995 static ir_node *is_simple_loop(void)
1998 ir_node *loop_block, *exit_block, *projx, *cond, *cmp;
2000 /* Maximum of one condition, and no endless loops. */
2001 if (loop_info.cf_outs != 1)
2004 DB((dbg, LEVEL_4, "1 loop exit\n"));
2006 /* Calculate maximum unroll_nr keeping node count below limit. */
2007 loop_info.max_unroll = (int)((double)opt_params.max_unrolled_loop_size / (double)loop_info.nodes);
2008 if (loop_info.max_unroll < 2) {
2013 DB((dbg, LEVEL_4, "maximum unroll factor %u, to not exceed node limit \n",
2014 opt_params.max_unrolled_loop_size));
2016 arity = get_irn_arity(loop_head);
2017 /* RETURN if we have more than 1 be. */
2018 /* Get my backedges without alien bes. */
2020 for (i = 0; i < arity; ++i) {
2021 ir_node *pred = get_irn_n(loop_head, i);
2022 if (is_own_backedge(loop_head, i)) {
2024 /* Our simple loops may have only one backedge. */
2027 loop_block = get_nodes_block(pred);
2028 loop_info.be_src_pos = i;
2033 DB((dbg, LEVEL_4, "loop has 1 own backedge.\n"));
2035 exit_block = get_nodes_block(loop_info.cf_out.pred);
2036 /* The loop has to be tail-controlled.
2037 * This can be changed/improved,
2038 * but we would need a duff iv. */
2039 if (exit_block != loop_block)
2042 DB((dbg, LEVEL_4, "tail-controlled loop.\n"));
2044 /* find value on which loop exit depends */
2045 projx = loop_info.cf_out.pred;
2046 cond = get_irn_n(projx, 0);
2047 cmp = get_irn_n(cond, 0);
2052 DB((dbg, LEVEL_5, "projection is %s\n", get_relation_string(get_Cmp_relation(cmp))));
2054 switch(get_Proj_proj(projx)) {
2056 loop_info.exit_cond = 0;
2059 loop_info.exit_cond = 1;
2062 panic("Cond Proj_proj other than true/false");
2065 DB((dbg, LEVEL_4, "Valid Cmp.\n"));
2069 /* Returns 1 if all nodes are mode_Iu or mode_Is. */
2070 static unsigned are_mode_I(ir_node *n1, ir_node* n2, ir_node *n3)
2072 ir_mode *m1 = get_irn_mode(n1);
2073 ir_mode *m2 = get_irn_mode(n2);
2074 ir_mode *m3 = get_irn_mode(n3);
2076 if ((m1 == mode_Iu && m2 == mode_Iu && m3 == mode_Iu) ||
2077 (m1 == mode_Is && m2 == mode_Is && m3 == mode_Is))
2083 /* Checks if cur_loop is a simple tail-controlled counting loop
2084 * with start and end value loop invariant, step constant. */
2085 static unsigned get_unroll_decision_invariant(void)
2088 ir_node *projres, *loop_condition, *iteration_path;
2090 ir_tarval *step_tar;
2094 /* RETURN if loop is not 'simple' */
2095 projres = is_simple_loop();
2096 if (projres == NULL)
2099 /* Use a minimal size for the invariant unrolled loop,
2100 * as duffs device produces overhead */
2101 if (loop_info.nodes < opt_params.invar_unrolling_min_size)
2104 loop_condition = get_irn_n(projres, 0);
2106 success = get_invariant_pred(loop_condition, &loop_info.end_val, &iteration_path);
2107 DB((dbg, LEVEL_4, "pred invar %d\n", success));
2112 DB((dbg, LEVEL_4, "Invariant End_val %N, other %N\n", loop_info.end_val, iteration_path));
2114 /* We may find the add or the phi first.
2115 * Until now we only have end_val. */
2116 if (is_Add(iteration_path) || is_Sub(iteration_path)) {
2118 loop_info.add = iteration_path;
2119 DB((dbg, LEVEL_4, "Case 1: Got add %N (maybe not sane)\n", loop_info.add));
2121 /* Preds of the add should be step and the iteration_phi */
2122 success = get_const_pred(loop_info.add, &loop_info.step, &loop_info.iteration_phi);
2126 DB((dbg, LEVEL_4, "Got step %N\n", loop_info.step));
2128 if (! is_Phi(loop_info.iteration_phi))
2131 DB((dbg, LEVEL_4, "Got phi %N\n", loop_info.iteration_phi));
2134 * Does necessary sanity check of add, if it is already set. */
2135 success = get_start_and_add(loop_info.iteration_phi, invariant);
2139 DB((dbg, LEVEL_4, "Got start A %N\n", loop_info.start_val));
2141 } else if (is_Phi(iteration_path)) {
2142 ir_node *new_iteration_phi;
2144 loop_info.iteration_phi = iteration_path;
2145 DB((dbg, LEVEL_4, "Case 2: Got phi %N\n", loop_info.iteration_phi));
2147 /* Find start_val and add-node.
2148 * Does necessary sanity check of add, if it is already set. */
2149 success = get_start_and_add(loop_info.iteration_phi, invariant);
2153 DB((dbg, LEVEL_4, "Got start B %N\n", loop_info.start_val));
2154 DB((dbg, LEVEL_4, "Got add or sub %N\n", loop_info.add));
2156 success = get_const_pred(loop_info.add, &loop_info.step, &new_iteration_phi);
2160 DB((dbg, LEVEL_4, "Got step (B) %N\n", loop_info.step));
2162 if (loop_info.iteration_phi != new_iteration_phi)
2169 mode = get_irn_mode(loop_info.end_val);
2171 DB((dbg, LEVEL_4, "start %N, end %N, step %N\n",
2172 loop_info.start_val, loop_info.end_val, loop_info.step));
2174 if (mode != mode_Is && mode != mode_Iu)
2177 /* TODO necessary? */
2178 if (!are_mode_I(loop_info.start_val, loop_info.step, loop_info.end_val))
2181 DB((dbg, LEVEL_4, "mode integer\n"));
2183 step_tar = get_Const_tarval(loop_info.step);
2185 if (tarval_is_null(step_tar)) {
2186 /* TODO Might be worth a warning. */
2190 DB((dbg, LEVEL_4, "step is not 0\n"));
2192 create_duffs_block();
2194 return loop_info.max_unroll;
2197 /* Returns unroll factor,
2198 * given maximum unroll factor and number of loop passes. */
2199 static unsigned get_preferred_factor_constant(ir_tarval *count_tar)
2201 ir_tarval *tar_6, *tar_5, *tar_4, *tar_3, *tar_2;
2203 ir_mode *mode = get_irn_mode(loop_info.end_val);
2205 tar_6 = new_tarval_from_long(6, mode);
2206 tar_5 = new_tarval_from_long(5, mode);
2207 tar_4 = new_tarval_from_long(4, mode);
2208 tar_3 = new_tarval_from_long(3, mode);
2209 tar_2 = new_tarval_from_long(2, mode);
2211 /* loop passes % {6, 5, 4, 3, 2} == 0 */
2212 if (tarval_is_null(tarval_mod(count_tar, tar_6)))
2214 else if (tarval_is_null(tarval_mod(count_tar, tar_5)))
2216 else if (tarval_is_null(tarval_mod(count_tar, tar_4)))
2218 else if (tarval_is_null(tarval_mod(count_tar, tar_3)))
2220 else if (tarval_is_null(tarval_mod(count_tar, tar_2)))
2223 /* gcd(max_unroll, count_tar) */
2224 int a = loop_info.max_unroll;
2225 int b = (int)get_tarval_long(count_tar);
2228 DB((dbg, LEVEL_4, "gcd of max_unroll %d and count_tar %d: ", a, b));
2235 DB((dbg, LEVEL_4, "%d\n", a));
2239 DB((dbg, LEVEL_4, "preferred unroll factor %d\n", prefer));
2242 * If our preference is greater than the allowed unroll factor
2243 * we either might reduce the preferred factor and prevent a duffs device block,
2244 * or create a duffs device block, from which in this case (constants only)
2245 * we know the startloop at compiletime.
2246 * The latter yields the following graphs.
2247 * but for code generation we would want to use graph A.
2248 * The graphs are equivalent. So, we can only reduce the preferred factor.
2258 if (prefer <= loop_info.max_unroll)
2263 if (loop_info.max_unroll >= 3)
2265 else if (loop_info.max_unroll >= 2)
2271 if (loop_info.max_unroll >= 2)
2282 /* Check if cur_loop is a simple counting loop.
2283 * Start, step and end are constants.
2284 * TODO The whole constant case should use procedures similar to
2285 * the invariant case, as they are more versatile. */
2287 static unsigned get_unroll_decision_constant(void)
2289 ir_node *cmp, *iteration_path;
2290 unsigned success, is_latest_val;
2291 ir_tarval *start_tar, *end_tar, *step_tar, *diff_tar, *count_tar;
2293 ir_relation proj_proj, norm_proj;
2296 /* RETURN if loop is not 'simple' */
2297 cmp = is_simple_loop();
2301 /* One in of the loop condition needs to be loop invariant. => end_val
2302 * The other in is assigned by an add. => add
2303 * The add uses a loop invariant value => step
2304 * and a phi with a loop invariant start_val and the add node as ins.
2316 success = get_const_pred(cmp, &loop_info.end_val, &iteration_path);
2320 DB((dbg, LEVEL_4, "End_val %N, other %N\n", loop_info.end_val, iteration_path));
2322 /* We may find the add or the phi first.
2323 * Until now we only have end_val. */
2324 if (is_Add(iteration_path) || is_Sub(iteration_path)) {
2326 /* We test against the latest value of the iv. */
2329 loop_info.add = iteration_path;
2330 DB((dbg, LEVEL_4, "Case 2: Got add %N (maybe not sane)\n", loop_info.add));
2332 /* Preds of the add should be step and the iteration_phi */
2333 success = get_const_pred(loop_info.add, &loop_info.step, &loop_info.iteration_phi);
2337 DB((dbg, LEVEL_4, "Got step %N\n", loop_info.step));
2339 if (! is_Phi(loop_info.iteration_phi))
2342 DB((dbg, LEVEL_4, "Got phi %N\n", loop_info.iteration_phi));
2345 * Does necessary sanity check of add, if it is already set. */
2346 success = get_start_and_add(loop_info.iteration_phi, constant);
2350 DB((dbg, LEVEL_4, "Got start %N\n", loop_info.start_val));
2352 } else if (is_Phi(iteration_path)) {
2353 ir_node *new_iteration_phi;
2355 /* We compare with the value the iv had entering this run. */
2358 loop_info.iteration_phi = iteration_path;
2359 DB((dbg, LEVEL_4, "Case 1: Got phi %N \n", loop_info.iteration_phi));
2361 /* Find start_val and add-node.
2362 * Does necessary sanity check of add, if it is already set. */
2363 success = get_start_and_add(loop_info.iteration_phi, constant);
2367 DB((dbg, LEVEL_4, "Got start %N\n", loop_info.start_val));
2368 DB((dbg, LEVEL_4, "Got add or sub %N\n", loop_info.add));
2370 success = get_const_pred(loop_info.add, &loop_info.step, &new_iteration_phi);
2374 DB((dbg, LEVEL_4, "Got step %N\n", loop_info.step));
2376 if (loop_info.iteration_phi != new_iteration_phi)
2384 mode = get_irn_mode(loop_info.end_val);
2386 DB((dbg, LEVEL_4, "start %N, end %N, step %N\n",
2387 loop_info.start_val, loop_info.end_val, loop_info.step));
2389 if (mode != mode_Is && mode != mode_Iu)
2392 /* TODO necessary? */
2393 if (!are_mode_I(loop_info.start_val, loop_info.step, loop_info.end_val))
2396 DB((dbg, LEVEL_4, "mode integer\n"));
2398 end_tar = get_Const_tarval(loop_info.end_val);
2399 start_tar = get_Const_tarval(loop_info.start_val);
2400 step_tar = get_Const_tarval(loop_info.step);
2402 if (tarval_is_null(step_tar))
2403 /* TODO Might be worth a warning. */
2406 DB((dbg, LEVEL_4, "step is not 0\n"));
2408 if ((!tarval_is_negative(step_tar)) ^ (!is_Sub(loop_info.add)))
2409 loop_info.decreasing = 1;
2411 diff_tar = tarval_sub(end_tar, start_tar, mode);
2413 /* We need at least count_tar steps to be close to end_val, maybe more.
2414 * No way, that we have gone too many steps.
2415 * This represents the 'latest value'.
2416 * (If condition checks against latest value, is checked later) */
2417 count_tar = tarval_div(diff_tar, step_tar);
2419 /* Iv will not pass end_val (except overflows).
2420 * Nothing done, as it would yield to no advantage. */
2421 if (tarval_is_negative(count_tar)) {
2422 DB((dbg, LEVEL_4, "Loop is endless or never taken."));
2423 /* TODO Might be worth a warning. */
2427 ++stats.u_simple_counting_loop;
2429 loop_info.latest_value = is_latest_val;
2432 if (! is_simple_counting_loop(&count_tar))
2436 /* stepped can be negative, if step < 0 */
2437 stepped = tarval_mul(count_tar, step_tar);
2439 /* step as close to end_val as possible, */
2440 /* |stepped| <= |end_tar|, and dist(stepped, end_tar) is smaller than a step. */
2441 if (is_Sub(loop_info.add))
2442 stepped = tarval_sub(start_tar, stepped, mode_Is);
2444 stepped = tarval_add(start_tar, stepped);
2446 DB((dbg, LEVEL_4, "stepped to %ld\n", get_tarval_long(stepped)));
2448 proj_proj = get_Cmp_relation(cmp);
2449 /* Assure that norm_proj is the stay-in-loop case. */
2450 if (loop_info.exit_cond == 1)
2451 norm_proj = get_negated_relation(proj_proj);
2453 norm_proj = proj_proj;
2455 DB((dbg, LEVEL_4, "normalized projection %s\n", get_relation_string(norm_proj)));
2456 /* Executed at most once (stay in counting loop if a Eq b) */
2457 if (norm_proj == ir_relation_equal)
2458 /* TODO Might be worth a warning. */
2461 /* calculates next values and increases count_tar according to it */
2462 success = simulate_next(&count_tar, stepped, step_tar, end_tar, norm_proj);
2466 /* We run loop once more, if we compare to the
2467 * not yet in-/decreased iv. */
2468 if (is_latest_val == 0) {
2469 DB((dbg, LEVEL_4, "condition uses not latest iv value\n"));
2470 count_tar = tarval_add(count_tar, get_tarval_one(mode));
2473 DB((dbg, LEVEL_4, "loop taken %ld times\n", get_tarval_long(count_tar)));
2475 /* Assure the loop is taken at least 1 time. */
2476 if (tarval_is_null(count_tar)) {
2477 /* TODO Might be worth a warning. */
2481 loop_info.count_tar = count_tar;
2482 return get_preferred_factor_constant(count_tar);
2488 static void unroll_loop(void)
2491 if (! (loop_info.nodes > 0))
2494 if (loop_info.nodes > opt_params.max_unrolled_loop_size) {
2495 DB((dbg, LEVEL_2, "Nodes %d > allowed nodes %d\n",
2496 loop_info.nodes, opt_params.max_unrolled_loop_size));
2501 if (loop_info.calls > 0) {
2502 DB((dbg, LEVEL_2, "Calls %d > allowed calls 0\n",
2504 ++stats.calls_limit;
2510 /* get_unroll_decision_constant and invariant are completely
2511 * independent for flexibility.
2512 * Some checks may be performed twice. */
2514 /* constant case? */
2515 if (opt_params.allow_const_unrolling)
2516 unroll_nr = get_unroll_decision_constant();
2517 if (unroll_nr > 1) {
2518 loop_info.unroll_kind = constant;
2521 /* invariant case? */
2522 if (opt_params.allow_invar_unrolling)
2523 unroll_nr = get_unroll_decision_invariant();
2525 loop_info.unroll_kind = invariant;
2528 DB((dbg, LEVEL_2, " *** Unrolling %d times ***\n", unroll_nr));
2530 if (unroll_nr > 1) {
2531 loop_entries = NEW_ARR_F(entry_edge, 0);
2534 irg_walk_graph(current_ir_graph, get_loop_entries, NULL, NULL);
2536 if (loop_info.unroll_kind == constant) {
2537 if ((int)get_tarval_long(loop_info.count_tar) == unroll_nr)
2538 loop_info.needs_backedge = 0;
2540 loop_info.needs_backedge = 1;
2542 loop_info.needs_backedge = 1;
2545 /* Use phase to keep copy of nodes from the condition chain. */
2546 ir_nodemap_init(&map, current_ir_graph);
2547 obstack_init(&obst);
2549 /* Copies the loop */
2550 copy_loop(loop_entries, unroll_nr - 1);
2552 /* Line up the floating copies. */
2553 place_copies(unroll_nr - 1);
2555 /* Remove phis with 1 in
2556 * If there were no nested phis, this would not be necessary.
2557 * Avoiding the creation in the first place
2558 * leads to complex special cases. */
2559 irg_walk_graph(current_ir_graph, correct_phis, NULL, NULL);
2561 if (loop_info.unroll_kind == constant)
2562 ++stats.constant_unroll;
2564 ++stats.invariant_unroll;
2566 clear_irg_properties(current_ir_graph, IR_GRAPH_PROPERTY_CONSISTENT_DOMINANCE);
2568 DEL_ARR_F(loop_entries);
2569 obstack_free(&obst, NULL);
2570 ir_nodemap_destroy(&map);
2575 /* Analyzes the loop, and checks if size is within allowed range.
2576 * Decides if loop will be processed. */
2577 static void init_analyze(ir_graph *irg, ir_loop *loop)
2582 loop_head_valid = true;
2584 /* Reset loop info */
2585 memset(&loop_info, 0, sizeof(loop_info_t));
2587 DB((dbg, LEVEL_1, " >>>> current loop %ld <<<\n",
2588 get_loop_loop_nr(loop)));
2590 /* Collect loop informations: head, node counts. */
2591 irg_walk_graph(irg, get_loop_info, NULL, NULL);
2593 /* RETURN if there is no valid head */
2594 if (!loop_head || !loop_head_valid) {
2595 DB((dbg, LEVEL_1, "No valid loop head. Nothing done.\n"));
2598 DB((dbg, LEVEL_1, "Loophead: %N\n", loop_head));
2601 if (loop_info.branches > opt_params.max_branches) {
2602 DB((dbg, LEVEL_1, "Branches %d > allowed branches %d\n",
2603 loop_info.branches, opt_params.max_branches));
2604 ++stats.calls_limit;
2609 case loop_op_inversion:
2610 loop_inversion(irg);
2613 case loop_op_unrolling:
2618 panic("Loop optimization not implemented.");
2620 DB((dbg, LEVEL_1, " <<<< end of loop with node %ld >>>>\n",
2621 get_loop_loop_nr(loop)));
2624 /* Find innermost loops and add them to loops. */
2625 static void find_innermost_loop(ir_loop *loop)
2627 bool had_sons = false;
2628 size_t n_elements = get_loop_n_elements(loop);
2631 for (e = 0; e < n_elements; ++e) {
2632 loop_element element = get_loop_element(loop, e);
2633 if (*element.kind == k_ir_loop) {
2634 find_innermost_loop(element.son);
2640 ARR_APP1(ir_loop*, loops, loop);
2644 static void set_loop_params(void)
2646 opt_params.max_loop_size = 100;
2647 opt_params.depth_adaption = -50;
2648 opt_params.count_phi = true;
2649 opt_params.count_proj = false;
2650 opt_params.allowed_calls = 0;
2652 opt_params.max_cc_size = 5;
2655 opt_params.allow_const_unrolling = true;
2656 opt_params.allow_invar_unrolling = false;
2658 opt_params.invar_unrolling_min_size = 20;
2659 opt_params.max_unrolled_loop_size = 400;
2660 opt_params.max_branches = 9999;
2663 /* Assure preconditions are met and go through all loops. */
2664 void loop_optimization(ir_graph *irg)
2670 assure_irg_properties(irg,
2671 IR_GRAPH_PROPERTY_CONSISTENT_OUT_EDGES
2672 | IR_GRAPH_PROPERTY_CONSISTENT_OUTS
2673 | IR_GRAPH_PROPERTY_CONSISTENT_LOOPINFO);
2677 /* Reset stats for this procedure */
2681 set_current_ir_graph(irg);
2683 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_PHI_LIST);
2684 collect_phiprojs(irg);
2686 loop = get_irg_loop(irg);
2688 loops = NEW_ARR_F(ir_loop *, 0);
2689 /* List all inner loops */
2690 n_elements = get_loop_n_elements(loop);
2691 for (i = 0; i < n_elements; ++i) {
2692 loop_element element = get_loop_element(loop, i);
2693 if (*element.kind != k_ir_loop)
2695 find_innermost_loop(element.son);
2698 /* Set all links to NULL */
2699 irg_walk_graph(irg, reset_link, NULL, NULL);
2701 for (i = 0; i < ARR_LEN(loops); ++i) {
2702 ir_loop *loop = loops[i];
2706 /* Analyze and handle loop */
2707 init_analyze(irg, loop);
2709 /* Copied blocks do not have their phi list yet */
2710 collect_phiprojs(irg);
2712 /* Set links to NULL
2713 * TODO Still necessary? */
2714 irg_walk_graph(irg, reset_link, NULL, NULL);
2720 ir_free_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_PHI_LIST);
2722 confirm_irg_properties(irg, IR_GRAPH_PROPERTIES_NONE);
2725 void do_loop_unrolling(ir_graph *irg)
2727 loop_op = loop_op_unrolling;
2728 loop_optimization(irg);
2731 void do_loop_inversion(ir_graph *irg)
2733 loop_op = loop_op_inversion;
2734 loop_optimization(irg);
2737 void do_loop_peeling(ir_graph *irg)
2739 loop_op = loop_op_peeling;
2740 loop_optimization(irg);
2743 ir_graph_pass_t *loop_inversion_pass(const char *name)
2745 return def_graph_pass(name ? name : "loop_inversion", do_loop_inversion);
2748 ir_graph_pass_t *loop_unroll_pass(const char *name)
2750 return def_graph_pass(name ? name : "loop_unroll", do_loop_unrolling);
2753 ir_graph_pass_t *loop_peeling_pass(const char *name)
2755 return def_graph_pass(name ? name : "loop_peeling", do_loop_peeling);
2758 void firm_init_loop_opt(void)
2760 FIRM_DBG_REGISTER(dbg, "firm.opt.loop");