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
29 #include "iroptimize.h"
48 #include "irbackedge_t.h"
49 #include "irphase_t.h"
53 DEBUG_ONLY(static firm_dbg_module_t *dbg);
55 /* DBG print stats for every procedure. */
56 #define LOOP_OPT_STATS 1
58 /* DBG: Ignore node limits and process every possible loop. */
59 #define LOOP_IGNORE_NODE_LIMITS 0
62 * Convenience macro for iterating over every phi node of the given block.
63 * Requires phi list per block.
65 #define for_each_phi(block, phi) \
66 for ((phi) = get_Block_phis( (block) ); (phi) ; (phi) = get_Phi_next((phi)))
68 #define for_each_phi_safe(head, phi, next) \
69 for ((phi) = (head), (next) = (head) ? get_Phi_next((head)) : NULL; \
70 (phi) ; (phi) = (next), (next) = (next) ? get_Phi_next((next)) : NULL)
72 /* Currently processed loop. */
73 static ir_loop *cur_loop;
75 /* Flag for kind of unrolling. */
79 } unrolling_kind_flag;
81 /* Condition for performing visiting a node during copy_walk. */
82 typedef unsigned walker_condition(ir_node *);
84 /* Node and position of a predecessor. */
85 typedef struct entry_edge {
91 /* Node info for unrolling. */
92 typedef struct unrolling_node_info {
94 } unrolling_node_info;
96 /* Outs of the nodes head. */
97 static entry_edge *cur_head_outs;
99 /* Information about the loop head */
100 static ir_node *loop_head = NULL;
101 static unsigned loop_head_valid = 1;
103 /* List of all inner loops, that are processed. */
104 static ir_loop **loops;
108 #define count_stats(val) (++val)
109 #define print_stats() (do_print_stats())
110 #define reset_stats() (do_reset_stats())
113 typedef struct loop_stats_t {
117 unsigned too_large_adapted;
118 unsigned cc_limit_reached;
119 unsigned calls_limit;
121 unsigned u_simple_counting_loop;
122 unsigned constant_unroll;
123 unsigned invariant_unroll;
128 static loop_stats_t stats;
130 /* Set stats to sero */
131 static void do_reset_stats(void)
133 memset(&stats, 0, sizeof(loop_stats_t));
137 static void do_print_stats(void)
139 DB((dbg, LEVEL_2, "---------------------------------------\n"));
140 DB((dbg, LEVEL_2, "loops : %d\n",stats.loops));
141 DB((dbg, LEVEL_2, "inverted : %d\n",stats.inverted));
142 DB((dbg, LEVEL_2, "too_large : %d\n",stats.too_large));
143 DB((dbg, LEVEL_2, "too_large_adapted : %d\n",stats.too_large_adapted));
144 DB((dbg, LEVEL_2, "cc_limit_reached : %d\n",stats.cc_limit_reached));
145 DB((dbg, LEVEL_2, "calls_limit : %d\n",stats.calls_limit));
146 DB((dbg, LEVEL_2, "u_simple_counting : %d\n",stats.u_simple_counting_loop));
147 DB((dbg, LEVEL_2, "constant_unroll : %d\n",stats.constant_unroll));
148 DB((dbg, LEVEL_2, "invariant_unroll : %d\n",stats.invariant_unroll));
149 DB((dbg, LEVEL_2, "=======================================\n"));
153 #define count_stats(val) ((void)0)
154 #define print_stats() ((void)0)
155 #define reset_stats() ((void)0)
159 /* Commandline parameters */
160 typedef struct loop_opt_params_t {
161 unsigned max_loop_size; /* Maximum number of nodes [nodes]*/
162 int depth_adaption; /* Loop nest depth adaption [percent] */
163 unsigned allowed_calls; /* Number of calls allowed [number] */
164 unsigned count_phi:1; /* Count phi nodes */
165 unsigned count_proj:1; /* Count projections */
167 unsigned max_cc_size; /* Maximum condition chain size [nodes] */
168 unsigned max_branches;
170 unsigned max_unrolled_loop_size; /* [nodes] */
171 unsigned allow_const_unrolling:1;
172 unsigned allow_invar_unrolling:1;
173 unsigned invar_unrolling_min_size; /* [nodes] */
177 static loop_opt_params_t opt_params;
179 /* Loop analysis informations */
180 typedef struct loop_info_t {
181 unsigned nodes; /* node count */
182 unsigned ld_st; /* load and store nodes */
183 unsigned branches; /* number of conditions */
184 unsigned calls; /* number of calls */
185 unsigned cf_outs; /* number of cf edges which leave the loop */
186 entry_edge cf_out; /* single loop leaving cf edge */
187 int be_src_pos; /* position of the single own backedge in the head */
190 unsigned cc_size; /* nodes in the condition chain */
193 unsigned max_unroll; /* Number of unrolls satisfying max_loop_size */
194 unsigned exit_cond; /* 1 if condition==true exits the loop. */
195 unsigned latest_value:1; /* 1 if condition is checked against latest counter value */
196 unsigned needs_backedge:1; /* 0 if loop is completely unrolled */
197 unsigned decreasing:1; /* Step operation is_Sub, or step is<0 */
199 /* IV informations of a simple loop */
203 ir_node *iteration_phi;
206 ir_tarval *count_tar; /* Number of loop iterations */
208 ir_node *duff_cond; /* Duff mod */
209 unrolling_kind_flag unroll_kind; /* constant or invariant unrolling */
212 /* Information about the current loop */
213 static loop_info_t loop_info;
215 /* Outs of the condition chain (loop inversion). */
216 static ir_node **cc_blocks;
217 /* df/cf edges with def in the condition chain */
218 static entry_edge *cond_chain_entries;
219 /* Array of df loops found in the condition chain. */
220 static entry_edge *head_df_loop;
221 /* Number of blocks in cc */
222 static unsigned inversion_blocks_in_cc;
225 /* Cf/df edges leaving the loop.
226 * Called entries here, as they are used to enter the loop with walkers. */
227 static entry_edge *loop_entries;
228 /* Number of unrolls to perform */
229 static int unroll_nr;
230 /* Phase is used to keep copies of nodes. */
231 static ir_phase *phase;
233 /* Loop operations. */
234 typedef enum loop_op_t {
240 /* Saves which loop operation to do until after basic tests. */
241 static loop_op_t loop_op;
243 /* Returns the maximum nodes for the given nest depth */
244 static unsigned get_max_nodes_adapted(unsigned depth)
246 double perc = 100.0 + (double)opt_params.depth_adaption;
247 double factor = pow(perc / 100.0, depth);
249 return (int)((double)opt_params.max_loop_size * factor);
252 /* Reset nodes link. For use with a walker. */
253 static void reset_link(ir_node *node, void *env)
256 set_irn_link(node, NULL);
259 /* Returns 0 if the node or block is not in cur_loop. */
260 static unsigned is_in_loop(ir_node *node)
262 return (get_irn_loop(get_block(node)) == cur_loop);
265 /* Returns 0 if the given edge is not a backedge
266 * with its pred in the cur_loop. */
267 static unsigned is_own_backedge(ir_node *n, int pos)
269 return (is_backedge(n, pos) && is_in_loop(get_irn_n(n, pos)));
272 /* Finds loop head and some loop_info as calls or else if necessary. */
273 static void get_loop_info(ir_node *node, void *env)
275 unsigned node_in_loop, pred_in_loop;
279 arity = get_irn_arity(node);
280 for (i = 0; i < arity; i++) {
281 ir_node *pred = get_irn_n(node, i);
283 pred_in_loop = is_in_loop(pred);
284 node_in_loop = is_in_loop(node);
286 if (!node_in_loop && pred_in_loop && is_Block(node))
294 loop_info.cf_out = entry;
297 /* collect some loop information */
299 if (is_Phi(node) && opt_params.count_phi)
301 else if (is_Proj(node) && opt_params.count_proj)
303 else if (!is_Confirm(node) && !is_Const(node) && !is_SymConst(node))
306 if (is_Load(node) || is_Store(node))
314 /* Find the loops head/the blocks with cfpred outside of the loop */
315 if (is_Block(node)) {
316 const ir_edge_t *edge;
319 /* Count innerloop branches */
320 foreach_out_edge_kind(node, edge, EDGE_KIND_NORMAL) {
321 if (is_Block(get_edge_src_irn(edge)) && is_in_loop(get_edge_src_irn(edge)))
325 ++loop_info.branches;
327 if (node_in_loop && !pred_in_loop && loop_head_valid) {
328 ir_node *cfgpred = get_Block_cfgpred(node, i);
330 if (!is_in_loop(cfgpred)) {
331 DB((dbg, LEVEL_5, "potential head %+F because inloop and pred %+F not inloop\n",
333 /* another head? We do not touch this. */
334 if (loop_head && loop_head != node) {
345 /* Finds all edges with users outside of the loop
346 * and definition inside the loop. */
347 static void get_loop_entries(ir_node *node, void *env)
349 unsigned node_in_loop, pred_in_loop;
353 arity = get_irn_arity(node);
354 for (i = 0; i < arity; ++i) {
355 ir_node *pred = get_irn_n(node, i);
357 pred_in_loop = is_in_loop(pred);
358 node_in_loop = is_in_loop(node);
360 if (pred_in_loop && !node_in_loop) {
365 ARR_APP1(entry_edge, loop_entries, entry);
371 static ir_node *ssa_second_def;
372 static ir_node *ssa_second_def_block;
375 * Walks the graph bottom up, searching for definitions and creates phis.
377 static ir_node *search_def_and_create_phis(ir_node *block, ir_mode *mode, int first)
385 DB((dbg, LEVEL_5, "ssa search_def_and_create_phis: block %N\n", block));
387 /* Prevents creation of phi that would be bad anyway.
388 * Dead and bad blocks. */
389 if (get_irn_arity(block) < 1 || is_Bad(block)) {
390 DB((dbg, LEVEL_5, "ssa bad %N\n", block));
394 if (block == ssa_second_def_block && !first) {
395 DB((dbg, LEVEL_5, "ssa found second definition: use second def %N\n", ssa_second_def));
396 return ssa_second_def;
399 /* already processed this block? */
400 if (irn_visited(block)) {
401 ir_node *value = (ir_node *) get_irn_link(block);
402 DB((dbg, LEVEL_5, "ssa already visited: use linked %N\n", value));
406 irg = get_irn_irg(block);
407 assert(block != get_irg_start_block(irg));
409 /* a Block with only 1 predecessor needs no Phi */
410 n_cfgpreds = get_Block_n_cfgpreds(block);
411 if (n_cfgpreds == 1) {
412 ir_node *pred_block = get_Block_cfgpred_block(block, 0);
415 DB((dbg, LEVEL_5, "ssa 1 pred: walk pred %N\n", pred_block));
417 value = search_def_and_create_phis(pred_block, mode, 0);
418 set_irn_link(block, value);
419 mark_irn_visited(block);
424 /* create a new Phi */
425 NEW_ARR_A(ir_node*, in, n_cfgpreds);
426 for (i = 0; i < n_cfgpreds; ++i)
427 in[i] = new_Unknown(mode);
429 phi = new_r_Phi(block, n_cfgpreds, in, mode);
430 /* Important: always keep block phi list up to date. */
431 add_Block_phi(block, phi);
432 DB((dbg, LEVEL_5, "ssa phi creation: link new phi %N to block %N\n", phi, block));
433 set_irn_link(block, phi);
434 mark_irn_visited(block);
436 /* set Phi predecessors */
437 for (i = 0; i < n_cfgpreds; ++i) {
439 ir_node *pred_block = get_Block_cfgpred_block(block, i);
440 assert(pred_block != NULL);
441 pred_val = search_def_and_create_phis(pred_block, mode, 0);
443 assert(pred_val != NULL);
445 DB((dbg, LEVEL_5, "ssa phi pred:phi %N, pred %N\n", phi, pred_val));
446 set_irn_n(phi, i, pred_val);
454 * Given a set of values this function constructs SSA-form for the users of the
455 * first value (the users are determined through the out-edges of the value).
456 * Works without using the dominance tree.
458 static void construct_ssa(ir_node *orig_block, ir_node *orig_val,
459 ir_node *second_block, ir_node *second_val)
463 const ir_edge_t *edge;
464 const ir_edge_t *next;
466 assert(orig_block && orig_val && second_block && second_val &&
467 "no parameter of construct_ssa may be NULL");
469 if (orig_val == second_val)
472 irg = get_irn_irg(orig_val);
474 ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED);
475 inc_irg_visited(irg);
477 mode = get_irn_mode(orig_val);
478 set_irn_link(orig_block, orig_val);
479 mark_irn_visited(orig_block);
481 ssa_second_def_block = second_block;
482 ssa_second_def = second_val;
484 /* Only fix the users of the first, i.e. the original node */
485 foreach_out_edge_safe(orig_val, edge, next) {
486 ir_node *user = get_edge_src_irn(edge);
487 int j = get_edge_src_pos(edge);
488 ir_node *user_block = get_nodes_block(user);
495 DB((dbg, LEVEL_5, "original user %N\n", user));
498 ir_node *pred_block = get_Block_cfgpred_block(user_block, j);
499 newval = search_def_and_create_phis(pred_block, mode, 1);
501 newval = search_def_and_create_phis(user_block, mode, 1);
503 if (newval != user && !is_Bad(newval))
504 set_irn_n(user, j, newval);
507 ir_free_resources(irg, IR_RESOURCE_IRN_VISITED);
511 /***** Unrolling Helper Functions *****/
513 /* Assign the copy with index nr to node n */
514 static void set_unroll_copy(ir_node *n, int nr, ir_node *cp)
516 unrolling_node_info *info;
517 assert(nr != 0 && "0 reserved");
519 info = (unrolling_node_info *)phase_get_irn_data(phase, n);
523 info = XMALLOCZ(unrolling_node_info);
524 arr = NEW_ARR_F(ir_node *, unroll_nr);
526 memset(info->copies, 0, (unroll_nr) * sizeof(ir_node *));
528 phase_set_irn_data(phase, n, info);
533 info->copies[nr] = cp;
536 /* Returns a nodes copy if it exists, else NULL. */
537 static ir_node *get_unroll_copy(ir_node *n, int nr)
540 unrolling_node_info *info = (unrolling_node_info *)phase_get_irn_data(phase, n);
544 cp = info->copies[nr];
549 /***** Inversion Helper Functions *****/
551 /* Sets copy cp of node n. */
552 static void set_inversion_copy(ir_node *n, ir_node *cp)
554 phase_set_irn_data(phase, n, cp);
557 /* Getter of copy of n for inversion */
558 static ir_node *get_inversion_copy(ir_node *n)
560 ir_node *cp = (ir_node *)phase_get_irn_data(phase, n);
564 /* Resets block mark for given node. For use with walker */
565 static void reset_block_mark(ir_node *node, void * env)
570 set_Block_mark(node, 0);
573 /* Returns mark of node, or its block if node is not a block.
574 * Used in this context to determine if node is in the condition chain. */
575 static unsigned is_nodes_block_marked(ir_node* node)
578 return get_Block_mark(node);
580 return get_Block_mark(get_block(node));
583 /* Extends a nodes ins by node new.
584 * NOTE: This is slow if a node n needs to be extended more than once. */
585 static void extend_irn(ir_node *n, ir_node *newnode, int new_is_backedge)
589 int arity = get_irn_arity(n);
590 int new_arity = arity + 1;
593 NEW_ARR_A(int, bes, new_arity);
594 NEW_ARR_A(ir_node *, ins, new_arity);
597 /* Bes are important!
598 * Another way would be recreating the looptree,
599 * but after that we cannot distinguish already processed loops
600 * from not yet processed ones. */
602 for(i = 0; i < arity; ++i) {
603 bes[i] = is_backedge(n, i);
605 bes[i] = new_is_backedge;
608 for(i = 0; i < arity; ++i) {
609 ins[i] = get_irn_n(n, i);
613 set_irn_in(n, new_arity, ins);
617 for(i = 0; i < new_arity; ++i) {
624 /* Extends a block by a copy of its pred at pos,
625 * fixing also the phis in the same way. */
626 static void extend_ins_by_copy(ir_node *block, int pos)
631 assert(is_Block(block));
633 /* Extend block by copy of definition at pos */
634 pred = get_irn_n(block, pos);
635 new_in = get_inversion_copy(pred);
636 DB((dbg, LEVEL_5, "Extend block %N by %N cp of %N\n", block, new_in, pred));
637 extend_irn(block, new_in, 0);
639 /* Extend block phis by copy of definition at pos */
640 for_each_phi(block, phi) {
643 pred = get_irn_n(phi, pos);
644 cp = get_inversion_copy(pred);
645 /* If the phis in is not in the condition chain (eg. a constant),
646 * there is no copy. */
652 DB((dbg, LEVEL_5, "Extend phi %N by %N cp of %N\n", phi, new_in, pred));
653 extend_irn(phi, new_in, 0);
657 /* Returns the number of blocks backedges. With or without alien bes. */
658 static int get_backedge_n(ir_node *block, unsigned with_alien)
662 int arity = get_irn_arity(block);
664 assert(is_Block(block));
666 for (i = 0; i < arity; ++i) {
667 ir_node *pred = get_irn_n(block, i);
668 if (is_backedge(block, i) && (with_alien || is_in_loop(pred)))
674 /* Returns a raw copy of the given node.
675 * Attributes are kept/set according to the needs of loop inversion. */
676 static ir_node *copy_node(ir_node *node)
681 cp = exact_copy(node);
682 arity = get_irn_arity(node);
684 /* Keep backedge info */
685 for (i = 0; i < arity; ++i) {
686 if (is_backedge(node, i))
691 set_Block_mark(cp, 0);
699 * This walker copies all walked nodes.
700 * If the walk_condition is true for a node, it is copied.
701 * All nodes node_info->copy have to be NULL prior to every walk.
702 * Order of ins is important for later usage.
704 static void copy_walk(ir_node *node, walker_condition *walk_condition,
711 ir_graph *irg = current_ir_graph;
714 * break condition and cycle resolver, creating temporary node copies
716 if (get_irn_visited(node) >= get_irg_visited(irg)) {
717 /* Here we rely on nodestate's copy being initialized with NULL */
718 DB((dbg, LEVEL_5, "copy_walk: We have already visited %N\n", node));
719 if (get_inversion_copy(node) == NULL) {
720 cp = copy_node(node);
721 set_inversion_copy(node, cp);
723 DB((dbg, LEVEL_5, "The TEMP copy of %N is created %N\n", node, cp));
729 mark_irn_visited(node);
731 if (!is_Block(node)) {
732 ir_node *pred = get_nodes_block(node);
733 if (walk_condition(pred))
734 DB((dbg, LEVEL_5, "walk block %N\n", pred));
735 copy_walk(pred, walk_condition, set_loop);
738 arity = get_irn_arity(node);
740 NEW_ARR_A(ir_node *, cpin, arity);
742 for (i = 0; i < arity; ++i) {
743 ir_node *pred = get_irn_n(node, i);
745 if (walk_condition(pred)) {
746 DB((dbg, LEVEL_5, "walk node %N\n", pred));
747 copy_walk(pred, walk_condition, set_loop);
748 cpin[i] = get_inversion_copy(pred);
749 DB((dbg, LEVEL_5, "copy of %N gets new in %N which is copy of %N\n",
750 node, get_inversion_copy(pred), pred));
756 /* copy node / finalize temp node */
757 if (get_inversion_copy(node) == NULL) {
758 /* No temporary copy existent */
759 cp = copy_node(node);
760 set_inversion_copy(node, cp);
761 DB((dbg, LEVEL_5, "The FINAL copy of %N is CREATED %N\n", node, cp));
763 /* temporary copy is existent but without correct ins */
764 cp = get_inversion_copy(node);
765 DB((dbg, LEVEL_5, "The FINAL copy of %N is EXISTENT %N\n", node, cp));
768 if (!is_Block(node)) {
769 ir_node *cpblock = get_inversion_copy(get_nodes_block(node));
771 set_nodes_block(cp, cpblock );
773 add_Block_phi(cpblock, cp);
776 /* Keeps phi list of temporary node. */
777 set_irn_in(cp, ARR_LEN(cpin), cpin);
781 * This walker copies all walked nodes.
782 * If the walk_condition is true for a node, it is copied.
783 * All nodes node_info->copy have to be NULL prior to every walk.
784 * Order of ins is important for later usage.
785 * Takes copy_index, to phase-link copy at specific index.
787 static void copy_walk_n(ir_node *node, walker_condition *walk_condition,
796 * break condition and cycle resolver, creating temporary node copies
798 if (irn_visited(node)) {
799 /* Here we rely on nodestate's copy being initialized with NULL */
800 DB((dbg, LEVEL_5, "copy_walk: We have already visited %N\n", node));
801 if (get_unroll_copy(node, copy_index) == NULL) {
804 set_unroll_copy(node, copy_index, u);
805 DB((dbg, LEVEL_5, "The TEMP unknown of %N is created %N\n", node, u));
811 mark_irn_visited(node);
813 if (!is_Block(node)) {
814 ir_node *block = get_nodes_block(node);
815 if (walk_condition(block))
816 DB((dbg, LEVEL_5, "walk block %N\n", block));
817 copy_walk_n(block, walk_condition, copy_index);
820 arity = get_irn_arity(node);
821 NEW_ARR_A(ir_node *, cpin, arity);
823 for (i = 0; i < arity; ++i) {
824 ir_node *pred = get_irn_n(node, i);
826 if (walk_condition(pred)) {
827 DB((dbg, LEVEL_5, "walk node %N\n", pred));
828 copy_walk_n(pred, walk_condition, copy_index);
829 cpin[i] = get_unroll_copy(pred, copy_index);
835 /* copy node / finalize temp node */
836 cp = get_unroll_copy(node, copy_index);
837 if (cp == NULL || is_Unknown(cp)) {
838 cp = copy_node(node);
839 set_unroll_copy(node, copy_index, cp);
840 DB((dbg, LEVEL_5, "The FINAL copy of %N is CREATED %N\n", node, cp));
842 /* temporary copy is existent but without correct ins */
843 cp = get_unroll_copy(node, copy_index);
844 DB((dbg, LEVEL_5, "The FINAL copy of %N is EXISTENT %N\n", node, cp));
847 if (!is_Block(node)) {
848 ir_node *cpblock = get_unroll_copy(get_nodes_block(node), copy_index);
850 set_nodes_block(cp, cpblock );
852 add_Block_phi(cpblock, cp);
855 /* Keeps phi list of temporary node. */
856 set_irn_in(cp, ARR_LEN(cpin), cpin);
859 /* Removes alle Blocks with non marked predecessors from the condition chain. */
860 static void unmark_not_allowed_cc_blocks(void)
862 size_t blocks = ARR_LEN(cc_blocks);
865 for(i = 0; i < blocks; ++i) {
866 ir_node *block = cc_blocks[i];
868 int arity = get_irn_arity(block);
870 /* Head is an exception. */
871 if (block == loop_head)
874 for(a = 0; a < arity; ++a) {
875 if (! is_nodes_block_marked(get_irn_n(block, a))) {
876 set_Block_mark(block, 0);
877 --inversion_blocks_in_cc;
878 DB((dbg, LEVEL_5, "Removed %N from cc (blocks in cc %d)\n",
879 block, inversion_blocks_in_cc));
887 /* Unmarks all cc blocks using cc_blocks except head.
888 * TODO: invert head for unrolling? */
889 static void unmark_cc_blocks(void)
891 size_t blocks = ARR_LEN(cc_blocks);
894 for(i = 0; i < blocks; ++i) {
895 ir_node *block = cc_blocks[i];
897 /* TODO Head is an exception. */
898 /*if (block != loop_head)*/
899 set_Block_mark(block, 0);
901 /*inversion_blocks_in_cc = 1;*/
902 inversion_blocks_in_cc = 0;
905 loop_info.cc_size = 0;
909 * Populates head_entries with (node, pred_pos) tuple
910 * whereas the node's pred at pred_pos is in the cc but not the node itself.
911 * Also finds df loops inside the cc.
912 * Head and condition chain blocks have been marked previously.
914 static void get_head_outs(ir_node *node, void *env)
917 int arity = get_irn_arity(node);
920 for (i = 0; i < arity; ++i) {
921 if (!is_nodes_block_marked(node) && is_nodes_block_marked(get_irn_n(node, i))) {
925 /* Saving also predecessor seems redundant, but becomes
926 * necessary when changing position of it, before
927 * dereferencing it.*/
928 entry.pred = get_irn_n(node, i);
929 ARR_APP1(entry_edge, cur_head_outs, entry);
933 arity = get_irn_arity(loop_head);
935 /* Find df loops inside the cc */
936 if (is_Phi(node) && get_nodes_block(node) == loop_head) {
937 for (i = 0; i < arity; ++i) {
938 if (is_own_backedge(loop_head, i)) {
939 if (is_nodes_block_marked(get_irn_n(node, i))) {
943 entry.pred = get_irn_n(node, i);
944 ARR_APP1(entry_edge, head_df_loop, entry);
945 DB((dbg, LEVEL_5, "Found incc assignment node %N @%d is pred %N, graph %N %N\n",
946 node, i, entry.pred, current_ir_graph, get_irg_start_block(current_ir_graph)));
954 * Find condition chains, and add them to be inverted.
955 * A block belongs to the chain if a condition branches out of the loop.
956 * (Some blocks need to be removed once again.)
957 * Returns 1 if the given block belongs to the condition chain.
959 static unsigned find_condition_chain(ir_node *block)
961 const ir_edge_t *edge;
965 unsigned nodes_n = 0;
967 mark_irn_visited(block);
969 DB((dbg, LEVEL_5, "condition_chains for block %N\n", block));
972 foreach_out_edge_kind(block, edge, EDGE_KIND_NORMAL) {
976 /* Check if node count would exceed maximum cc size.
978 * This is not optimal, as we search depth-first and break here,
979 * continuing with another subtree. */
980 if (loop_info.cc_size + nodes_n > opt_params.max_cc_size) {
981 set_Block_mark(block, 0);
985 /* Check if block only has a jmp instruction. */
987 foreach_out_edge(block, edge) {
988 ir_node *src = get_edge_src_irn(edge);
990 if (! is_Block(src) && ! is_Jmp(src)) {
995 /* Check cf outs if one is leaving the loop,
996 * or if this node has a backedge. */
997 foreach_block_succ(block, edge) {
998 ir_node *src = get_edge_src_irn(edge);
999 int pos = get_edge_src_pos(edge);
1001 if (! is_in_loop(src))
1004 /* Inverting blocks with backedge outs leads to a cf edge
1005 * from the inverted head, into the inverted head (skipping the body).
1006 * As the body becomes the new loop head,
1007 * this would introduce another loop in the existing loop.
1008 * This loop inversion cannot cope with this case. */
1009 if (is_backedge(src, pos)) {
1015 /* We need all predecessors to already belong to the condition chain.
1016 * Example of wrong case: * == in cc
1027 /* Collect blocks containing only a Jmp.
1028 * Do not collect blocks with backedge outs. */
1029 if ((jmp_only == 1 || mark == 1) && has_be == 0) {
1030 set_Block_mark(block, 1);
1031 ++inversion_blocks_in_cc;
1032 loop_info.cc_size += nodes_n;
1033 DB((dbg, LEVEL_5, "block %N is part of condition chain\n", block));
1034 ARR_APP1(ir_node *, cc_blocks, block);
1036 set_Block_mark(block, 0);
1039 foreach_block_succ(block, edge) {
1040 ir_node *src = get_edge_src_irn( edge );
1042 if (is_in_loop(src) && ! irn_visited(src))
1043 find_condition_chain(src);
1050 * Rewires the copied condition chain. Removes backedges
1051 * as this condition chain is prior to the loop.
1052 * Copy of loop_head must have phi list and old (unfixed) backedge info of the loop head.
1053 * (loop_head is already fixed, we cannot rely on it.)
1055 static void fix_copy_inversion(void)
1060 ir_node *phi, *next;
1061 ir_node *head_cp = get_inversion_copy(loop_head);
1062 int arity = get_irn_arity(head_cp);
1063 int backedges = get_backedge_n(head_cp, 0);
1064 int new_arity = arity - backedges;
1068 NEW_ARR_A(ir_node *, ins, new_arity);
1071 /* Remove block backedges */
1072 for(i = 0; i < arity; ++i) {
1073 if (!is_backedge(head_cp, i))
1074 ins[pos++] = get_irn_n(head_cp, i);
1077 new_head = new_Block(new_arity, ins);
1079 phis = NEW_ARR_F(ir_node *, 0);
1081 for_each_phi_safe(get_Block_phis(head_cp), phi, next) {
1083 NEW_ARR_A(ir_node *, ins, new_arity);
1085 for(i = 0; i < arity; ++i) {
1086 if (!is_backedge(head_cp, i))
1087 ins[pos++] = get_irn_n(phi, i);
1089 new_phi = new_rd_Phi(get_irn_dbg_info(phi),
1090 new_head, new_arity, ins,
1092 ARR_APP1(ir_node *, phis, new_phi);
1096 for_each_phi_safe(get_Block_phis(head_cp), phi, next) {
1097 exchange(phi, phis[pos++]);
1100 exchange(head_cp, new_head);
1106 /* Puts the original condition chain at the end of the loop,
1107 * subsequently to the body.
1108 * Relies on block phi list and correct backedges.
1110 static void fix_head_inversion(void)
1114 ir_node *phi, *next;
1116 int arity = get_irn_arity(loop_head);
1117 int backedges = get_backedge_n(loop_head, 0);
1118 int new_arity = backedges;
1122 NEW_ARR_A(ir_node *, ins, new_arity);
1125 /* Keep only backedges */
1126 for(i = 0; i < arity; ++i) {
1127 if (is_own_backedge(loop_head, i))
1128 ins[pos++] = get_irn_n(loop_head, i);
1131 new_head = new_Block(new_arity, ins);
1133 phis = NEW_ARR_F(ir_node *, 0);
1135 for_each_phi(loop_head, phi) {
1137 DB((dbg, LEVEL_5, "Fixing phi %N of loop head\n", phi));
1139 NEW_ARR_A(ir_node *, ins, new_arity);
1142 for (i = 0; i < arity; ++i) {
1143 ir_node *pred = get_irn_n(phi, i);
1145 if (is_own_backedge(loop_head, i)) {
1146 /* If assignment is in the condition chain,
1147 * we need to create a phi in the new loop head.
1148 * This can only happen for df, not cf. See find_condition_chains. */
1149 /*if (is_nodes_block_marked(pred)) {
1157 new_phi = new_rd_Phi(get_irn_dbg_info(phi),
1158 new_head, new_arity, ins,
1161 ARR_APP1(ir_node *, phis, new_phi);
1163 DB((dbg, LEVEL_5, "fix inverted head should exch %N by %N (pos %d)\n", phi, new_phi, pos ));
1167 for_each_phi_safe(get_Block_phis(loop_head), phi, next) {
1168 DB((dbg, LEVEL_5, "fix inverted exch phi %N by %N\n", phi, phis[pos]));
1169 if (phis[pos] != phi)
1170 exchange(phi, phis[pos++]);
1175 DB((dbg, LEVEL_5, "fix inverted head exch head block %N by %N\n", loop_head, new_head));
1176 exchange(loop_head, new_head);
1179 /* Does the loop inversion. */
1180 static void inversion_walk(entry_edge *head_entries)
1185 * The order of rewiring bottom-up is crucial.
1186 * Any change of the order leads to lost information that would be needed later.
1189 ir_reserve_resources(current_ir_graph, IR_RESOURCE_IRN_VISITED);
1191 /* 1. clone condition chain */
1192 inc_irg_visited(current_ir_graph);
1194 for (i = 0; i < ARR_LEN(head_entries); ++i) {
1195 entry_edge entry = head_entries[i];
1196 ir_node *pred = get_irn_n(entry.node, entry.pos);
1198 DB((dbg, LEVEL_5, "\nInit walk block %N\n", pred));
1200 copy_walk(pred, is_nodes_block_marked, cur_loop);
1203 ir_free_resources(current_ir_graph, IR_RESOURCE_IRN_VISITED);
1205 /* 2. Extends the head control flow successors ins
1206 * with the definitions of the copied head node. */
1207 for (i = 0; i < ARR_LEN(head_entries); ++i) {
1208 entry_edge head_out = head_entries[i];
1210 if (is_Block(head_out.node))
1211 extend_ins_by_copy(head_out.node, head_out.pos);
1214 /* 3. construct_ssa for users of definitions in the condition chain,
1215 * as there is now a second definition. */
1216 for (i = 0; i < ARR_LEN(head_entries); ++i) {
1217 entry_edge head_out = head_entries[i];
1219 /* Ignore keepalives */
1220 if (is_End(head_out.node))
1223 /* Construct ssa for assignments in the condition chain. */
1224 if (!is_Block(head_out.node)) {
1225 ir_node *pred, *cppred, *block, *cpblock;
1227 pred = head_out.pred;
1228 cppred = get_inversion_copy(pred);
1229 block = get_nodes_block(pred);
1230 cpblock = get_nodes_block(cppred);
1231 construct_ssa(block, pred, cpblock, cppred);
1236 * If there is an assignment in the condition chain
1237 * with a user also in the condition chain,
1238 * the dominance frontier is in the new loop head.
1239 * The dataflow loop is completely in the condition chain.
1244 * Phi_cp | | copied condition chain
1248 * Phi* | | new loop head with newly created phi.
1250 * Phi | | original, inverted condition chain
1255 for (i = 0; i < ARR_LEN(head_df_loop); ++i) {
1256 entry_edge head_out = head_df_loop[i];
1258 /* Construct ssa for assignments in the condition chain. */
1259 ir_node *pred, *cppred, *block, *cpblock;
1261 pred = head_out.pred;
1262 cppred = get_inversion_copy(pred);
1263 assert(cppred && pred);
1264 block = get_nodes_block(pred);
1265 cpblock = get_nodes_block(cppred);
1266 construct_ssa(block, pred, cpblock, cppred);
1269 /* 4. Remove the ins which are no backedges from the original condition chain
1270 * as the cc is now subsequent to the body. */
1271 fix_head_inversion();
1273 /* 5. Remove the backedges of the copied condition chain,
1274 * because it is going to be the new 'head' in advance to the loop. */
1275 fix_copy_inversion();
1279 /* Performs loop inversion of cur_loop if possible and reasonable. */
1280 static void loop_inversion(void)
1283 unsigned max_loop_nodes = opt_params.max_loop_size;
1284 unsigned max_loop_nodes_adapted;
1285 int depth_adaption = opt_params.depth_adaption;
1287 unsigned do_inversion = 1;
1288 unsigned has_cc = 0;
1290 /* Depth of 0 is the procedure and 1 a topmost loop. */
1291 loop_depth = get_loop_depth(cur_loop) - 1;
1293 /* Calculating in per mil. */
1294 max_loop_nodes_adapted = get_max_nodes_adapted(loop_depth);
1296 DB((dbg, LEVEL_1, "max_nodes: %d\nmax_nodes_adapted %d at depth of %d (adaption %d)\n",
1297 max_loop_nodes, max_loop_nodes_adapted, loop_depth, depth_adaption));
1299 if (! (loop_info.nodes > 0))
1302 #if LOOP_IGNORE_NODE_LIMITS
1303 DB((dbg, LEVEL_1, "WARNING: Loop node limitations ignored."));
1305 if (loop_info.nodes > max_loop_nodes) {
1306 /* Only for stats */
1307 DB((dbg, LEVEL_1, "Nodes %d > allowed nodes %d\n",
1308 loop_info.nodes, loop_depth, max_loop_nodes));
1309 count_stats(stats.too_large);
1311 /* Adaption might change it */
1314 /* Limit processing to loops smaller than given parameter. */
1315 if (loop_info.nodes > max_loop_nodes_adapted) {
1316 DB((dbg, LEVEL_1, "Nodes %d > allowed nodes (depth %d adapted) %d\n",
1317 loop_info.nodes, loop_depth, max_loop_nodes_adapted));
1318 count_stats(stats.too_large_adapted);
1322 if (loop_info.calls > opt_params.allowed_calls) {
1323 DB((dbg, LEVEL_1, "Calls %d > allowed calls %d\n",
1324 loop_info.calls, opt_params.allowed_calls));
1325 count_stats(stats.calls_limit);
1330 /*inversion_head_node_limit = INT_MAX;*/
1331 ir_reserve_resources(current_ir_graph, IR_RESOURCE_BLOCK_MARK);
1333 /* Reset block marks.
1334 * We use block marks to flag blocks of the original condition chain. */
1335 irg_walk_graph(current_ir_graph, reset_block_mark, NULL, NULL);
1337 /*loop_info.blocks = get_loop_n_blocks(cur_loop);*/
1338 cond_chain_entries = NEW_ARR_F(entry_edge, 0);
1339 head_df_loop = NEW_ARR_F(entry_edge, 0);
1341 /*head_inversion_node_count = 0;*/
1342 inversion_blocks_in_cc = 0;
1344 /* Use phase to keep copy of nodes from the condition chain. */
1345 phase = new_phase(current_ir_graph, phase_irn_init_default);
1347 /* Search for condition chains and temporarily save the blocks in an array. */
1348 cc_blocks = NEW_ARR_F(ir_node *, 0);
1349 inc_irg_visited(current_ir_graph);
1350 has_cc = find_condition_chain(loop_head);
1352 unmark_not_allowed_cc_blocks();
1353 DEL_ARR_F(cc_blocks);
1355 #if LOOP_IGNORE_NODE_LIMITS
1356 (void) unmark_cc_blocks;
1358 /* Condition chain too large.
1359 * Loop should better be small enough to fit into the cache. */
1360 /* TODO Of course, we should take a small enough cc in the first place,
1361 * which is not that simple. (bin packing) */
1362 if (loop_info.cc_size > opt_params.max_cc_size) {
1363 count_stats(stats.cc_limit_reached);
1367 /* Unmark cc blocks except the head.
1368 * Invert head only for possible unrolling. */
1374 /* We also catch endless loops here,
1375 * because they do not have a condition chain. */
1376 if (inversion_blocks_in_cc < 1) {
1379 "Loop contains %d (less than 1) invertible blocks => No Inversion done.\n",
1380 inversion_blocks_in_cc));
1384 cur_head_outs = NEW_ARR_F(entry_edge, 0);
1386 /* Get all edges pointing into the condition chain. */
1387 irg_walk_graph(current_ir_graph, get_head_outs, NULL, NULL);
1389 /* Do the inversion */
1390 inversion_walk(cur_head_outs);
1392 DEL_ARR_F(cur_head_outs);
1394 /* Duplicated blocks changed doms */
1395 set_irg_doms_inconsistent(current_ir_graph);
1396 /* Loop content changed */
1397 set_irg_loopinfo_inconsistent(current_ir_graph);
1398 /* TODO are they? Depends on set_irn_in and set_irn_n exchange and new_node. */
1399 set_irg_outs_inconsistent(current_ir_graph);
1401 count_stats(stats.inverted);
1406 DEL_ARR_F(cond_chain_entries);
1407 DEL_ARR_F(head_df_loop);
1409 ir_free_resources(current_ir_graph, IR_RESOURCE_BLOCK_MARK);
1412 /* Fix the original loop_heads ins for invariant unrolling case. */
1413 static void unrolling_fix_loop_head_inv(void)
1417 ir_node *proj = new_Proj(loop_info.duff_cond, mode_X, 0);
1418 ir_node *head_pred = get_irn_n(loop_head, loop_info.be_src_pos);
1419 ir_node *loop_condition = get_unroll_copy(head_pred, unroll_nr - 1);
1421 /* Original loop_heads ins are:
1422 * duff block and the own backedge */
1424 ins[0] = loop_condition;
1426 set_irn_in(loop_head, 2, ins);
1427 DB((dbg, LEVEL_4, "Rewire ins of block loophead %N to pred %N and duffs entry %N \n" , loop_head, ins[0], ins[1]));
1429 for_each_phi(loop_head, phi) {
1430 ir_node *pred = get_irn_n(phi, loop_info.be_src_pos);
1431 /* TODO we think it is a phi, but for Mergesort it is not the case.*/
1433 ir_node *last_pred = get_unroll_copy(pred, unroll_nr - 1);
1436 ins[1] = (ir_node*)get_irn_link(phi);
1437 set_irn_in(phi, 2, ins);
1438 DB((dbg, LEVEL_4, "Rewire ins of loophead phi %N to pred %N and duffs entry %N \n" , phi, ins[0], ins[1]));
1442 /* Removes previously created phis with only 1 in. */
1443 static void correct_phis(ir_node *node, void *env)
1447 if (is_Phi(node) && get_irn_arity(node) == 1) {
1451 in[0] = get_irn_n(node, 0);
1453 exch = new_rd_Phi(get_irn_dbg_info(node),
1454 get_nodes_block(node), 1, in,
1455 get_irn_mode(node));
1457 exchange(node, exch);
1461 /* Unrolling: Rewire floating copies. */
1462 static void place_copies(int copies)
1464 ir_node *loophead = loop_head;
1467 int be_src_pos = loop_info.be_src_pos;
1469 /* Serialize loops by fixing their head ins.
1470 * Processed are the copies.
1471 * The original loop is done after that, to keep backedge infos. */
1472 for (c = 0; c < copies; ++c) {
1473 ir_node *upper = get_unroll_copy(loophead, c);
1474 ir_node *lower = get_unroll_copy(loophead, c + 1);
1476 ir_node *topmost_be_block = get_nodes_block(get_irn_n(loophead, be_src_pos));
1478 /* Important: get the preds first and then their copy. */
1479 ir_node *upper_be_block = get_unroll_copy(topmost_be_block, c);
1480 ir_node *new_jmp = new_r_Jmp(upper_be_block);
1481 DB((dbg, LEVEL_5, " place_copies upper %N lower %N\n", upper, lower));
1483 DB((dbg, LEVEL_5, "topmost be block %N \n", topmost_be_block));
1485 if (loop_info.unroll_kind == constant) {
1488 set_irn_in(lower, 1, ins);
1490 for_each_phi(loophead, phi) {
1491 ir_node *topmost_def = get_irn_n(phi, be_src_pos);
1492 ir_node *upper_def = get_unroll_copy(topmost_def, c);
1493 ir_node *lower_phi = get_unroll_copy(phi, c + 1);
1495 /* It is possible, that the value used
1496 * in the OWN backedge path is NOT defined in this loop. */
1497 if (is_in_loop(topmost_def))
1500 ins[0] = topmost_def;
1502 set_irn_in(lower_phi, 1, ins);
1503 /* Need to replace phis with 1 in later. */
1506 /* Invariant case */
1507 /* Every node has 2 ins. One from the duff blocks
1508 * and one from the previously unrolled loop. */
1510 /* Calculate corresponding projection of mod result for this copy c */
1511 ir_node *proj = new_Proj(loop_info.duff_cond, mode_X, unroll_nr - c - 1);
1512 DB((dbg, LEVEL_4, "New duff proj %N\n" , proj));
1516 set_irn_in(lower, 2, ins);
1517 DB((dbg, LEVEL_4, "Rewire ins of Block %N to pred %N and duffs entry %N \n" , lower, ins[0], ins[1]));
1519 for_each_phi(loophead, phi) {
1520 ir_node *topmost_phi_pred = get_irn_n(phi, be_src_pos);
1521 ir_node *upper_phi_pred;
1525 lower_phi = get_unroll_copy(phi, c + 1);
1526 duff_phi = (ir_node*)get_irn_link(phi);
1527 DB((dbg, LEVEL_4, "DD Link of %N is %N\n" , phi, duff_phi));
1530 if (is_in_loop(topmost_phi_pred)) {
1531 upper_phi_pred = get_unroll_copy(topmost_phi_pred, c);
1533 upper_phi_pred = topmost_phi_pred;
1536 ins[0] = upper_phi_pred;
1538 set_irn_in(lower_phi, 2, ins);
1539 DB((dbg, LEVEL_4, "Rewire ins of %N to pred %N and duffs entry %N \n" , lower_phi, ins[0], ins[1]));
1544 /* Reconnect last copy. */
1545 for (i = 0; i < ARR_LEN(loop_entries); ++i) {
1546 entry_edge edge = loop_entries[i];
1547 /* Last copy is at the bottom */
1548 ir_node *new_pred = get_unroll_copy(edge.pred, copies);
1549 set_irn_n(edge.node, edge.pos, new_pred);
1552 /* Fix original loops head.
1553 * Done in the end, as ins and be info were needed before. */
1554 if (loop_info.unroll_kind == constant) {
1556 ir_node *head_pred = get_irn_n(loop_head, be_src_pos);
1557 ir_node *loop_condition = get_unroll_copy(head_pred, unroll_nr - 1);
1559 set_irn_n(loop_head, loop_info.be_src_pos, loop_condition);
1561 for_each_phi(loop_head, phi) {
1562 ir_node *pred = get_irn_n(phi, be_src_pos);
1565 /* It is possible, that the value used
1566 * in the OWN backedge path is NOT assigned in this loop. */
1567 if (is_in_loop(pred))
1568 last_pred = get_unroll_copy(pred, copies);
1571 set_irn_n(phi, be_src_pos, last_pred);
1575 unrolling_fix_loop_head_inv();
1579 /* Copies the cur_loop several times. */
1580 static void copy_loop(entry_edge *cur_loop_outs, int copies)
1584 ir_reserve_resources(current_ir_graph, IR_RESOURCE_IRN_VISITED);
1586 for (c = 0; c < copies; ++c) {
1589 inc_irg_visited(current_ir_graph);
1591 DB((dbg, LEVEL_5, " ### Copy_loop copy nr: %d ###\n", c));
1592 for (i = 0; i < ARR_LEN(cur_loop_outs); ++i) {
1593 entry_edge entry = cur_loop_outs[i];
1594 ir_node *pred = get_irn_n(entry.node, entry.pos);
1596 copy_walk_n(pred, is_in_loop, c + 1);
1600 ir_free_resources(current_ir_graph, IR_RESOURCE_IRN_VISITED);
1604 /* Creates a new phi from the given phi node omitting own bes,
1605 * using be_block as supplier of backedge informations. */
1606 static ir_node *clone_phis_sans_bes(ir_node *phi, ir_node *be_block, ir_node *dest_block)
1609 int arity = get_irn_arity(phi);
1613 assert(get_irn_arity(phi) == get_irn_arity(be_block));
1614 assert(is_Phi(phi));
1616 ins = NEW_ARR_F(ir_node *, arity);
1617 for (i = 0; i < arity; ++i) {
1618 if (! is_own_backedge(be_block, i)) {
1619 ins[c] = get_irn_n(phi, i);
1624 newphi = new_r_Phi(dest_block, c, ins, get_irn_mode(phi));
1626 set_irn_link(phi, newphi);
1627 DB((dbg, LEVEL_4, "Linking for duffs device %N to %N\n", phi, newphi));
1632 /* Creates a new block from the given block node omitting own bes,
1633 * using be_block as supplier of backedge informations. */
1634 static ir_node *clone_block_sans_bes(ir_node *node, ir_node *be_block)
1637 int arity = get_irn_arity(node);
1640 assert(get_irn_arity(node) == get_irn_arity(be_block));
1641 assert(is_Block(node));
1643 NEW_ARR_A(ir_node *, ins, arity);
1644 for (i = 0; i < arity; ++i) {
1645 if (! is_own_backedge(be_block, i)) {
1646 ins[c] = get_irn_n(node, i);
1651 return new_Block(c, ins);
1654 /* Creates a structure to calculate absolute value of node op.
1655 * Returns mux node with absolute value. */
1656 static ir_node *new_Abs(ir_node *op, ir_mode *mode)
1658 ir_graph *irg = get_irn_irg(op);
1659 ir_node *block = get_nodes_block(op);
1660 ir_node *zero = new_r_Const(irg, get_mode_null(mode));
1661 ir_node *cmp = new_r_Cmp(block, op, zero);
1662 ir_node *cond = new_r_Proj(cmp, mode_b, pn_Cmp_Lt);
1663 ir_node *minus_op = new_r_Minus(block, op, mode);
1664 ir_node *mux = new_r_Mux(block, cond, op, minus_op, mode);
1670 /* Creates blocks for duffs device, using previously obtained
1671 * informations about the iv.
1673 static void create_duffs_block(void)
1677 ir_node *block1, *count_block, *duff_block;
1678 ir_node *ems, *ems_divmod, *ems_mod_proj, *cmp_null,
1679 *cmp_proj, *ems_mode_cond, *x_true, *x_false, *const_null;
1680 ir_node *true_val, *false_val;
1683 ir_node *duff_mod, *proj, *cond;
1685 ir_node *count, *correction, *unroll_c;
1686 ir_node *cmp_bad_count, *good_count, *bad_count, *count_phi, *bad_count_neg;
1689 mode = get_irn_mode(loop_info.end_val);
1690 const_null = new_Const(get_mode_null(mode));
1693 * 1. Calculate first approach to count.
1694 * Condition: (end - start) % step == 0 */
1695 block1 = clone_block_sans_bes(loop_head, loop_head);
1696 DB((dbg, LEVEL_4, "Duff block 1 %N\n", block1));
1698 /* Create loop entry phis in first duff block
1699 * as it becomes the loops preheader */
1700 for_each_phi(loop_head, phi) {
1701 /* Returns phis pred if phi would have arity 1*/
1702 ir_node *new_phi = clone_phis_sans_bes(phi, loop_head, block1);
1704 DB((dbg, LEVEL_4, "HEAD %N phi %N\n", loop_head, phi));
1705 DB((dbg, LEVEL_4, "BLOCK1 %N phi %N\n", block1, new_phi));
1708 ems = new_r_Sub(block1, loop_info.end_val, loop_info.start_val,
1709 get_irn_mode(loop_info.end_val));
1710 DB((dbg, LEVEL_4, "BLOCK1 sub %N\n", ems));
1713 ems = new_Sub(loop_info.end_val, loop_info.start_val,
1714 get_irn_mode(loop_info.end_val));
1716 DB((dbg, LEVEL_4, "divmod ins %N %N\n", ems, loop_info.step));
1717 ems_divmod = new_r_DivMod(block1,
1722 op_pin_state_pinned);
1724 DB((dbg, LEVEL_4, "New module node %N\n", ems_divmod));
1726 ems_mod_proj = new_r_Proj(ems_divmod, mode_Iu, pn_DivMod_res_mod);
1727 cmp_null = new_r_Cmp(block1, ems_mod_proj, const_null);
1728 cmp_proj = new_r_Proj(cmp_null, mode_b, pn_Cmp_Eq);
1729 ems_mode_cond = new_r_Cond(block1, cmp_proj);
1732 /* ems % step == 0 */
1733 x_true = new_r_Proj(ems_mode_cond, mode_X, pn_Cond_true);
1734 /* ems % step != 0 */
1735 x_false = new_r_Proj(ems_mode_cond, mode_X, pn_Cond_false);
1738 * Assures, duffs device receives a valid count.
1740 * decreasing: count < 0
1741 * increasing: count > 0
1746 count_block = new_Block(2, ins);
1747 DB((dbg, LEVEL_4, "Duff block 2 %N\n", count_block));
1750 /* Increase loop-taken-count depending on the loop condition
1751 * uses the latest iv to compare to. */
1752 if (loop_info.latest_value == 1) {
1753 /* ems % step == 0 : +0 */
1754 true_val = new_Const(get_mode_null(mode));
1755 /* ems % step != 0 : +1 */
1756 false_val = new_Const(get_mode_one(mode));
1758 ir_tarval *tv_two = new_tarval_from_long(2, mode);
1759 /* ems % step == 0 : +1 */
1760 true_val = new_Const(get_mode_one(mode));
1761 /* ems % step != 0 : +2 */
1762 false_val = new_Const(tv_two);
1768 correction = new_r_Phi(count_block, 2, ins, mode);
1770 count = new_r_Proj(ems_divmod, mode, pn_DivMod_res_div);
1772 /* (end - start) / step + correction */
1773 count = new_Add(count, correction, mode);
1775 cmp_bad_count = new_r_Cmp(count_block, count, const_null);
1777 /* We preconditioned the loop to be tail-controlled.
1778 * So, if count is something 'wrong' like 0,
1779 * negative/positive (depending on step direction),
1780 * we may take the loop once (tail-contr.) and leave it
1781 * to the existing condition, to break; */
1783 /* Depending on step direction, we have to check for > or < 0 */
1784 if (loop_info.decreasing == 1) {
1785 bad_count_neg = new_r_Proj(cmp_bad_count, mode_b, pn_Cmp_Lt);
1787 bad_count_neg = new_r_Proj(cmp_bad_count, mode_b, pn_Cmp_Gt);
1790 bad_count_neg = new_r_Cond(count_block, bad_count_neg);
1791 good_count = new_Proj(bad_count_neg, mode_X, pn_Cond_true);
1792 bad_count = new_Proj(ems_mode_cond, mode_X, pn_Cond_false);
1795 * Contains module to decide which loop to start from. */
1797 ins[0] = good_count;
1799 duff_block = new_Block(2, ins);
1800 DB((dbg, LEVEL_4, "Duff block 3 %N\n", duff_block));
1802 /* Get absolute value */
1803 ins[0] = new_Abs(count, mode);
1804 /* Manually feed the aforementioned count = 1 (bad case)*/
1805 ins[1] = new_Const(get_mode_one(mode));
1806 count_phi = new_r_Phi(duff_block, 2, ins, mode);
1808 unroll_c = new_Const(new_tarval_from_long((long)unroll_nr, mode));
1810 /* count % unroll_nr */
1811 duff_mod = new_r_Mod(duff_block,
1816 op_pin_state_pinned);
1819 proj = new_Proj(duff_mod, mode, pn_Mod_res);
1820 /* condition does NOT create itself in the block of the proj! */
1821 cond = new_r_Cond(duff_block, proj);
1823 loop_info.duff_cond = cond;
1826 /* Returns 1 if given node is not in loop,
1827 * or if it is a phi of the loop head with only loop invariant defs.
1829 static unsigned is_loop_invariant_def(ir_node *node)
1833 if (! is_in_loop(node)) {
1834 DB((dbg, LEVEL_4, "Not in loop %N\n", node));
1835 /* || is_Const(node) || is_SymConst(node)) {*/
1839 /* If this is a phi of the loophead shared by more than 1 loop,
1840 * we need to check if all defs are not in the loop. */
1843 block = get_nodes_block(node);
1845 /* To prevent unexpected situations. */
1846 if (block != loop_head) {
1850 for (i = 0; i < get_irn_arity(node); ++i) {
1851 /* Check if all bes are just loopbacks. */
1852 if (is_own_backedge(block, i) && get_irn_n(node, i) != node)
1855 DB((dbg, LEVEL_4, "invar %N\n", node));
1858 DB((dbg, LEVEL_4, "Not invar %N\n", node));
1863 /* Returns 1 if one pred of node is invariant and the other is not.
1864 * invar_pred and other are set analogously. */
1865 static unsigned get_invariant_pred(ir_node *node, ir_node **invar_pred, ir_node **other)
1867 ir_node *pred0 = get_irn_n(node, 0);
1868 ir_node *pred1 = get_irn_n(node, 1);
1873 if (is_loop_invariant_def(pred0)) {
1874 DB((dbg, LEVEL_4, "pred0 invar %N\n", pred0));
1875 *invar_pred = pred0;
1879 if (is_loop_invariant_def(pred1)) {
1880 DB((dbg, LEVEL_4, "pred1 invar %N\n", pred1));
1882 if (*invar_pred != NULL) {
1883 /* RETURN. We do not want both preds to be invariant. */
1888 *invar_pred = pred1;
1891 DB((dbg, LEVEL_4, "pred1 not invar %N\n", pred1));
1893 if (*invar_pred != NULL)
1900 /* Starts from a phi that may belong to an iv.
1901 * If an add forms a loop with iteration_phi,
1902 * and add uses a constant, 1 is returned
1903 * and 'start' as well as 'add' are sane. */
1904 static unsigned get_start_and_add(ir_node *iteration_phi, unrolling_kind_flag role)
1907 ir_node *found_add = loop_info.add;
1908 int arity = get_irn_arity(iteration_phi);
1910 DB((dbg, LEVEL_4, "Find start and add from %N\n", iteration_phi));
1912 for (i = 0; i < arity; ++i) {
1914 /* Find start_val which needs to be pred of the iteration_phi.
1915 * If start_val already known, sanity check. */
1916 if (!is_backedge(get_nodes_block(loop_info.iteration_phi), i)) {
1917 ir_node *found_start_val = get_irn_n(loop_info.iteration_phi, i);
1919 DB((dbg, LEVEL_4, "found_start_val %N\n", found_start_val));
1921 /* We already found a start_val it has to be always the same. */
1922 if (loop_info.start_val && found_start_val != loop_info.start_val)
1925 if ((role == constant) && !(is_SymConst(found_start_val) || is_Const(found_start_val)))
1927 else if((role == constant) && !(is_loop_invariant_def(found_start_val)))
1930 loop_info.start_val = found_start_val;
1933 /* The phi has to be in the loop head.
1934 * Follow all own backedges. Every value supplied from these preds of the phi
1935 * needs to origin from the same add. */
1936 if (is_own_backedge(get_nodes_block(loop_info.iteration_phi), i)) {
1937 ir_node *new_found = get_irn_n(loop_info.iteration_phi,i);
1939 DB((dbg, LEVEL_4, "is add? %N\n", new_found));
1941 if (! (is_Add(new_found) || is_Sub(new_found)) || (found_add && found_add != new_found))
1944 found_add = new_found;
1948 loop_info.add = found_add;
1954 /* Returns 1 if one pred of node is a const value and the other is not.
1955 * const_pred and other are set analogously. */
1956 static unsigned get_const_pred(ir_node *node, ir_node **const_pred, ir_node **other)
1958 ir_node *pred0 = get_irn_n(node, 0);
1959 ir_node *pred1 = get_irn_n(node, 1);
1961 DB((dbg, LEVEL_4, "Checking for constant pred of %N\n", node));
1966 /*DB((dbg, LEVEL_4, "is %N const\n", pred0));*/
1967 if (is_Const(pred0) || is_SymConst(pred0)) {
1968 *const_pred = pred0;
1972 /*DB((dbg, LEVEL_4, "is %N const\n", pred1));*/
1973 if (is_Const(pred1) || is_SymConst(pred1)) {
1974 if (*const_pred != NULL) {
1975 /* RETURN. We do not want both preds to be constant. */
1980 *const_pred = pred1;
1983 if (*const_pred == NULL)
1989 /* Returns the mathematically inverted pn_Cmp. */
1990 static pn_Cmp get_math_inverted_case(pn_Cmp proj)
2006 panic("Unhandled pn_Cmp.");
2010 /* Returns 1 if loop exits within 2 steps of the iv.
2011 * Norm_proj means we do not exit the loop.*/
2012 static unsigned simulate_next(ir_tarval **count_tar,
2013 ir_tarval *stepped, ir_tarval *step_tar, ir_tarval *end_tar,
2018 DB((dbg, LEVEL_4, "Loop taken if (stepped)%ld %s (end)%ld ",
2019 get_tarval_long(stepped),
2020 get_pnc_string((norm_proj)),
2021 get_tarval_long(end_tar)));
2022 DB((dbg, LEVEL_4, "comparing latest value %d\n", loop_info.latest_value));
2024 /* If current iv does not stay in the loop,
2025 * this run satisfied the exit condition. */
2026 if (! (tarval_cmp(stepped, end_tar) & norm_proj))
2029 DB((dbg, LEVEL_4, "Result: (stepped)%ld IS %s (end)%ld\n",
2030 get_tarval_long(stepped),
2031 get_pnc_string(tarval_cmp(stepped, end_tar)),
2032 get_tarval_long(end_tar)));
2035 if (is_Add(loop_info.add))
2036 next = tarval_add(stepped, step_tar);
2039 next = tarval_sub(stepped, step_tar, get_irn_mode(loop_info.end_val));
2041 DB((dbg, LEVEL_4, "Loop taken if %ld %s %ld ",
2042 get_tarval_long(next),
2043 get_pnc_string(norm_proj),
2044 get_tarval_long(end_tar)));
2045 DB((dbg, LEVEL_4, "comparing latest value %d\n", loop_info.latest_value));
2047 /* Increase steps. */
2048 *count_tar = tarval_add(*count_tar, get_tarval_one(get_tarval_mode(*count_tar)));
2050 /* Next has to fail the loop condition, or we will never exit. */
2051 if (! (tarval_cmp(next, end_tar) & norm_proj))
2057 /* Check if loop meets requirements for a 'simple loop':
2058 * - Exactly one cf out
2060 * - Max nodes after unrolling
2064 * Returns Projection of cmp node or NULL; */
2065 static ir_node *is_simple_loop(void)
2068 ir_node *loop_block, *exit_block, *projx, *cond, *projres, *loop_condition;
2070 /* Maximum of one condition, and no endless loops. */
2071 if (loop_info.cf_outs != 1)
2074 DB((dbg, LEVEL_4, "1 loop exit\n"));
2076 #if LOOP_IGNORE_NODE_LIMITS
2077 /* Ignore loop size. Probably not wise in other than testcases. */
2078 loop_info.max_unroll = 40;
2080 /* Calculate maximum unroll_nr keeping node count below limit. */
2081 loop_info.max_unroll = (int)((double)opt_params.max_unrolled_loop_size / (double)loop_info.nodes);
2082 if (loop_info.max_unroll < 2) {
2083 count_stats(stats.too_large);
2089 DB((dbg, LEVEL_4, "maximum unroll factor %u, to not exceed node limit \n",
2090 opt_params.max_unrolled_loop_size));
2092 arity = get_irn_arity(loop_head);
2093 /* RETURN if we have more than 1 be. */
2094 /* Get my backedges without alien bes. */
2096 for (i = 0; i < arity; ++i) {
2097 ir_node *pred = get_irn_n(loop_head, i);
2098 if (is_own_backedge(loop_head, i)) {
2100 /* Our simple loops may have only one backedge. */
2103 loop_block = get_nodes_block(pred);
2104 loop_info.be_src_pos = i;
2109 DB((dbg, LEVEL_4, "loop has 1 own backedge.\n"));
2111 exit_block = get_nodes_block(loop_info.cf_out.pred);
2112 /* The loop has to be tail-controlled.
2113 * This can be changed/improved,
2114 * but we would need a duff iv. */
2115 if (exit_block != loop_block)
2118 DB((dbg, LEVEL_4, "tail-controlled loop.\n"));
2120 /* find value on which loop exit depends */
2121 projx = loop_info.cf_out.pred;
2122 cond = get_irn_n(projx, 0);
2123 projres = get_irn_n(cond, 0);
2124 loop_condition = get_irn_n(projres, 0);
2126 if (!is_Cmp(loop_condition))
2129 DB((dbg, LEVEL_5, "projection is %s\n", get_pnc_string(get_Proj_proj(projx))));
2131 switch(get_Proj_proj(projx)) {
2133 loop_info.exit_cond = 0;
2136 loop_info.exit_cond = 1;
2139 panic("Cond Proj_proj other than true/false");
2142 DB((dbg, LEVEL_4, "Valid Cmp.\n"));
2147 /* Returns 1 if all nodes are mode_Iu or mode_Is. */
2148 static unsigned are_mode_I(ir_node *n1, ir_node* n2, ir_node *n3)
2150 ir_mode *m1 = get_irn_mode(n1);
2151 ir_mode *m2 = get_irn_mode(n2);
2152 ir_mode *m3 = get_irn_mode(n3);
2154 if ((m1 == mode_Iu && m2 == mode_Iu && m3 == mode_Iu) ||
2155 (m1 == mode_Is && m2 == mode_Is && m3 == mode_Is))
2161 /* Checks if cur_loop is a simple tail-controlled counting loop
2162 * with start and end value loop invariant, step constant. */
2163 static unsigned get_unroll_decision_invariant(void)
2166 ir_node *projres, *loop_condition, *iteration_path;
2167 unsigned success, is_latest_val;
2168 ir_tarval *step_tar;
2172 /* RETURN if loop is not 'simple' */
2173 projres = is_simple_loop();
2174 if (projres == NULL)
2177 /* Use a minimal size for the invariant unrolled loop,
2178 * as duffs device produces overhead */
2179 if (loop_info.nodes < opt_params.invar_unrolling_min_size)
2182 loop_condition = get_irn_n(projres, 0);
2184 success = get_invariant_pred(loop_condition, &loop_info.end_val, &iteration_path);
2185 DB((dbg, LEVEL_4, "pred invar %d\n", success));
2190 DB((dbg, LEVEL_4, "Invariant End_val %N, other %N\n", loop_info.end_val, iteration_path));
2192 /* We may find the add or the phi first.
2193 * Until now we only have end_val. */
2194 if (is_Add(iteration_path) || is_Sub(iteration_path)) {
2196 /* We test against the latest value of the iv. */
2199 loop_info.add = iteration_path;
2200 DB((dbg, LEVEL_4, "Case 1: Got add %N (maybe not sane)\n", loop_info.add));
2202 /* Preds of the add should be step and the iteration_phi */
2203 success = get_const_pred(loop_info.add, &loop_info.step, &loop_info.iteration_phi);
2207 DB((dbg, LEVEL_4, "Got step %N\n", loop_info.step));
2209 if (! is_Phi(loop_info.iteration_phi))
2212 DB((dbg, LEVEL_4, "Got phi %N\n", loop_info.iteration_phi));
2215 * Does necessary sanity check of add, if it is already set. */
2216 success = get_start_and_add(loop_info.iteration_phi, invariant);
2220 DB((dbg, LEVEL_4, "Got start A %N\n", loop_info.start_val));
2222 } else if (is_Phi(iteration_path)) {
2223 ir_node *new_iteration_phi;
2225 /* We compare with the value the iv had entering this run. */
2228 loop_info.iteration_phi = iteration_path;
2229 DB((dbg, LEVEL_4, "Case 2: Got phi %N\n", loop_info.iteration_phi));
2231 /* Find start_val and add-node.
2232 * Does necessary sanity check of add, if it is already set. */
2233 success = get_start_and_add(loop_info.iteration_phi, invariant);
2237 DB((dbg, LEVEL_4, "Got start B %N\n", loop_info.start_val));
2238 DB((dbg, LEVEL_4, "Got add or sub %N\n", loop_info.add));
2240 success = get_const_pred(loop_info.add, &loop_info.step, &new_iteration_phi);
2244 DB((dbg, LEVEL_4, "Got step (B) %N\n", loop_info.step));
2246 if (loop_info.iteration_phi != new_iteration_phi)
2253 mode = get_irn_mode(loop_info.end_val);
2255 DB((dbg, LEVEL_4, "start %N, end %N, step %N\n",
2256 loop_info.start_val, loop_info.end_val, loop_info.step));
2258 if (mode != mode_Is && mode != mode_Iu)
2261 /* TODO necessary? */
2262 if (!are_mode_I(loop_info.start_val, loop_info.step, loop_info.end_val))
2265 DB((dbg, LEVEL_4, "mode integer\n"));
2267 step_tar = get_Const_tarval(loop_info.step);
2269 if (tarval_is_null(step_tar)) {
2270 /* TODO Might be worth a warning. */
2274 DB((dbg, LEVEL_4, "step is not 0\n"));
2276 create_duffs_block();
2278 return loop_info.max_unroll;
2281 /* Returns unroll factor,
2282 * given maximum unroll factor and number of loop passes. */
2283 static unsigned get_preferred_factor_constant(ir_tarval *count_tar)
2285 ir_tarval *tar_6, *tar_5, *tar_4, *tar_3, *tar_2;
2287 ir_mode *mode = get_irn_mode(loop_info.end_val);
2289 tar_6 = new_tarval_from_long(6, mode);
2290 tar_5 = new_tarval_from_long(5, mode);
2291 tar_4 = new_tarval_from_long(4, mode);
2292 tar_3 = new_tarval_from_long(3, mode);
2293 tar_2 = new_tarval_from_long(2, mode);
2295 /* loop passes % {6, 5, 4, 3, 2} == 0 */
2296 if (tarval_is_null(tarval_mod(count_tar, tar_6)))
2298 else if (tarval_is_null(tarval_mod(count_tar, tar_5)))
2300 else if (tarval_is_null(tarval_mod(count_tar, tar_4)))
2302 else if (tarval_is_null(tarval_mod(count_tar, tar_3)))
2304 else if (tarval_is_null(tarval_mod(count_tar, tar_2)))
2307 /* gcd(max_unroll, count_tar) */
2308 int a = loop_info.max_unroll;
2309 int b = (int)get_tarval_long(count_tar);
2312 DB((dbg, LEVEL_4, "gcd of max_unroll %d and count_tar %d: ", a, b));
2319 DB((dbg, LEVEL_4, "%d\n", a));
2323 DB((dbg, LEVEL_4, "preferred unroll factor %d\n", prefer));
2326 * If our preference is greater than the allowed unroll factor
2327 * we either might reduce the preferred factor and prevent a duffs device block,
2328 * or create a duffs device block, from which in this case (constants only)
2329 * we know the startloop at compiletime.
2330 * The latter yields the following graphs.
2331 * but for code generation we would want to use graph A.
2332 * The graphs are equivalent. So, we can only reduce the preferred factor.
2342 if (prefer <= loop_info.max_unroll)
2347 if (loop_info.max_unroll >= 3)
2349 else if (loop_info.max_unroll >= 2)
2355 if (loop_info.max_unroll >= 2)
2366 /* Check if cur_loop is a simple counting loop.
2367 * Start, step and end are constants.
2368 * TODO The whole constant case should use procedures similar to
2369 * the invariant case, as they are more versatile. */
2371 static unsigned get_unroll_decision_constant(void)
2373 ir_node *projres, *loop_condition, *iteration_path;
2374 unsigned success, is_latest_val;
2375 ir_tarval *start_tar, *end_tar, *step_tar, *diff_tar, *count_tar, *stepped;
2376 pn_Cmp proj_proj, norm_proj;
2379 /* RETURN if loop is not 'simple' */
2380 projres = is_simple_loop();
2381 if (projres == NULL)
2384 /* One in of the loop condition needs to be loop invariant. => end_val
2385 * The other in is assigned by an add. => add
2386 * The add uses a loop invariant value => step
2387 * and a phi with a loop invariant start_val and the add node as ins.
2399 loop_condition = get_irn_n(projres, 0);
2401 success = get_const_pred(loop_condition, &loop_info.end_val, &iteration_path);
2405 DB((dbg, LEVEL_4, "End_val %N, other %N\n", loop_info.end_val, iteration_path));
2407 /* We may find the add or the phi first.
2408 * Until now we only have end_val. */
2409 if (is_Add(iteration_path) || is_Sub(iteration_path)) {
2411 /* We test against the latest value of the iv. */
2414 loop_info.add = iteration_path;
2415 DB((dbg, LEVEL_4, "Case 2: Got add %N (maybe not sane)\n", loop_info.add));
2417 /* Preds of the add should be step and the iteration_phi */
2418 success = get_const_pred(loop_info.add, &loop_info.step, &loop_info.iteration_phi);
2422 DB((dbg, LEVEL_4, "Got step %N\n", loop_info.step));
2424 if (! is_Phi(loop_info.iteration_phi))
2427 DB((dbg, LEVEL_4, "Got phi %N\n", loop_info.iteration_phi));
2430 * Does necessary sanity check of add, if it is already set. */
2431 success = get_start_and_add(loop_info.iteration_phi, constant);
2435 DB((dbg, LEVEL_4, "Got start %N\n", loop_info.start_val));
2437 } else if (is_Phi(iteration_path)) {
2438 ir_node *new_iteration_phi;
2440 /* We compare with the value the iv had entering this run. */
2443 loop_info.iteration_phi = iteration_path;
2444 DB((dbg, LEVEL_4, "Case 1: Got phi %N \n", loop_info.iteration_phi));
2446 /* Find start_val and add-node.
2447 * Does necessary sanity check of add, if it is already set. */
2448 success = get_start_and_add(loop_info.iteration_phi, constant);
2452 DB((dbg, LEVEL_4, "Got start %N\n", loop_info.start_val));
2453 DB((dbg, LEVEL_4, "Got add or sub %N\n", loop_info.add));
2455 success = get_const_pred(loop_info.add, &loop_info.step, &new_iteration_phi);
2459 DB((dbg, LEVEL_4, "Got step %N\n", loop_info.step));
2461 if (loop_info.iteration_phi != new_iteration_phi)
2469 mode = get_irn_mode(loop_info.end_val);
2471 DB((dbg, LEVEL_4, "start %N, end %N, step %N\n",
2472 loop_info.start_val, loop_info.end_val, loop_info.step));
2474 if (mode != mode_Is && mode != mode_Iu)
2477 /* TODO necessary? */
2478 if (!are_mode_I(loop_info.start_val, loop_info.step, loop_info.end_val))
2481 DB((dbg, LEVEL_4, "mode integer\n"));
2483 end_tar = get_Const_tarval(loop_info.end_val);
2484 start_tar = get_Const_tarval(loop_info.start_val);
2485 step_tar = get_Const_tarval(loop_info.step);
2487 if (tarval_is_null(step_tar))
2488 /* TODO Might be worth a warning. */
2491 DB((dbg, LEVEL_4, "step is not 0\n"));
2493 if ((!tarval_is_negative(step_tar)) ^ (!is_Sub(loop_info.add)))
2494 loop_info.decreasing = 1;
2496 diff_tar = tarval_sub(end_tar, start_tar, mode);
2498 /* We need at least count_tar steps to be close to end_val, maybe more.
2499 * No way, that we have gone too many steps.
2500 * This represents the 'latest value'.
2501 * (If condition checks against latest value, is checked later) */
2502 count_tar = tarval_div(diff_tar, step_tar);
2504 /* Iv will not pass end_val (except overflows).
2505 * Nothing done, as it would yield to no advantage. */
2506 if (tarval_is_negative(count_tar)) {
2507 DB((dbg, LEVEL_4, "Loop is endless or never taken."));
2508 /* TODO Might be worth a warning. */
2512 count_stats(stats.u_simple_counting_loop);
2514 loop_info.latest_value = is_latest_val;
2517 if (! is_simple_counting_loop(&count_tar))
2521 /* stepped can be negative, if step < 0 */
2522 stepped = tarval_mul(count_tar, step_tar);
2524 /* step as close to end_val as possible, */
2525 /* |stepped| <= |end_tar|, and dist(stepped, end_tar) is smaller than a step. */
2526 if (is_Sub(loop_info.add))
2527 stepped = tarval_sub(start_tar, stepped, mode_Is);
2529 stepped = tarval_add(start_tar, stepped);
2531 DB((dbg, LEVEL_4, "stepped to %ld\n", get_tarval_long(stepped)));
2533 proj_proj = get_Proj_pn_cmp(projres);
2534 /* Assure that norm_proj is the stay-in-loop case. */
2535 if (loop_info.exit_cond == 1)
2536 norm_proj = get_math_inverted_case(proj_proj);
2538 norm_proj = proj_proj;
2540 DB((dbg, LEVEL_4, "normalized projection %s\n", get_pnc_string(norm_proj)));
2542 /* Executed at most once (stay in counting loop if a Eq b) */
2543 if (norm_proj == pn_Cmp_Eq)
2544 /* TODO Might be worth a warning. */
2547 /* calculates next values and increases count_tar according to it */
2548 success = simulate_next(&count_tar, stepped, step_tar, end_tar, norm_proj);
2552 /* We run loop once more, if we compare to the
2553 * not yet in-/decreased iv. */
2554 if (is_latest_val == 0) {
2555 DB((dbg, LEVEL_4, "condition uses not latest iv value\n"));
2556 count_tar = tarval_add(count_tar, get_tarval_one(mode));
2559 DB((dbg, LEVEL_4, "loop taken %ld times\n", get_tarval_long(count_tar)));
2561 /* Assure the loop is taken at least 1 time. */
2562 if (tarval_is_null(count_tar)) {
2563 /* TODO Might be worth a warning. */
2567 loop_info.count_tar = count_tar;
2568 return get_preferred_factor_constant(count_tar);
2574 static void unroll_loop(void)
2577 if (! (loop_info.nodes > 0))
2580 #if LOOP_IGNORE_NODE_LIMITS
2581 DB((dbg, LEVEL_1, "WARNING: Loop node limitations ignored."));
2583 if (loop_info.nodes > opt_params.max_unrolled_loop_size) {
2584 DB((dbg, LEVEL_2, "Nodes %d > allowed nodes %d\n",
2585 loop_info.nodes, opt_params.max_unrolled_loop_size));
2586 count_stats(stats.too_large);
2590 if (loop_info.calls > 0) {
2591 DB((dbg, LEVEL_2, "Calls %d > allowed calls 0\n",
2593 count_stats(stats.calls_limit);
2600 /* get_unroll_decision_constant and invariant are completely
2601 * independent for flexibility.
2602 * Some checks may be performed twice. */
2604 /* constant case? */
2605 if (opt_params.allow_const_unrolling)
2606 unroll_nr = get_unroll_decision_constant();
2607 if (unroll_nr > 1) {
2608 loop_info.unroll_kind = constant;
2611 /* invariant case? */
2612 if (opt_params.allow_invar_unrolling)
2613 unroll_nr = get_unroll_decision_invariant();
2615 loop_info.unroll_kind = invariant;
2618 DB((dbg, LEVEL_2, " *** Unrolling %d times ***\n", unroll_nr));
2620 if (unroll_nr > 1) {
2621 loop_entries = NEW_ARR_F(entry_edge, 0);
2624 irg_walk_graph(current_ir_graph, get_loop_entries, NULL, NULL);
2626 if (loop_info.unroll_kind == constant) {
2627 if ((int)get_tarval_long(loop_info.count_tar) == unroll_nr)
2628 loop_info.needs_backedge = 0;
2630 loop_info.needs_backedge = 1;
2632 loop_info.needs_backedge = 1;
2635 /* Use phase to keep copy of nodes from the condition chain. */
2636 phase = new_phase(current_ir_graph, phase_irn_init_default);
2638 /* Copies the loop */
2639 copy_loop(loop_entries, unroll_nr - 1);
2641 /* Line up the floating copies. */
2642 place_copies(unroll_nr - 1);
2644 /* Remove phis with 1 in
2645 * If there were no nested phis, this would not be necessary.
2646 * Avoiding the creation in the first place
2647 * leads to complex special cases. */
2648 irg_walk_graph(current_ir_graph, correct_phis, NULL, NULL);
2650 if (loop_info.unroll_kind == constant)
2651 count_stats(stats.constant_unroll);
2653 count_stats(stats.invariant_unroll);
2655 set_irg_doms_inconsistent(current_ir_graph);
2656 set_irg_loopinfo_inconsistent(current_ir_graph);
2658 set_irg_outs_inconsistent(current_ir_graph);
2660 DEL_ARR_F(loop_entries);
2665 /* Analyzes the loop, and checks if size is within allowed range.
2666 * Decides if loop will be processed. */
2667 static void init_analyze(ir_loop *loop)
2672 loop_head_valid = 1;
2674 /* Reset loop info */
2675 memset(&loop_info, 0, sizeof(loop_info_t));
2677 DB((dbg, LEVEL_1, " >>>> current loop includes node %N <<<\n",
2678 get_loop_node(loop, 0)));
2680 /* Collect loop informations: head, node counts. */
2681 irg_walk_graph(current_ir_graph, get_loop_info, NULL, NULL);
2683 /* RETURN if there is no valid head */
2684 if (!loop_head || !loop_head_valid) {
2685 DB((dbg, LEVEL_1, "No valid loop head. Nothing done.\n"));
2688 DB((dbg, LEVEL_1, "Loophead: %N\n", loop_head));
2691 if (loop_info.branches > opt_params.max_branches) {
2692 DB((dbg, LEVEL_1, "Branches %d > allowed branches %d\n",
2693 loop_info.branches, opt_params.max_branches));
2694 count_stats(stats.calls_limit);
2699 case loop_op_inversion:
2703 case loop_op_unrolling:
2708 panic("Loop optimization not implemented.");
2710 DB((dbg, LEVEL_1, " <<<< end of loop with node %N >>>>\n",
2711 get_loop_node(loop, 0)));
2714 /* Find innermost loops and add them to loops. */
2715 static void find_innermost_loop(ir_loop *loop)
2717 /* descend into sons */
2718 int sons = get_loop_n_sons(loop);
2721 ARR_APP1(ir_loop *, loops, loop);
2724 for (s=0; s<sons; s++) {
2725 find_innermost_loop(get_loop_son(loop, s));
2730 static void set_loop_params(void)
2732 opt_params.max_loop_size = 100;
2733 opt_params.depth_adaption = -50;
2734 opt_params.count_phi = 1;
2735 opt_params.count_proj = 0;
2736 opt_params.allowed_calls = 0;
2738 opt_params.max_cc_size = 5;
2741 opt_params.allow_const_unrolling = 1;
2742 opt_params.allow_invar_unrolling = 0;
2744 opt_params.invar_unrolling_min_size = 20;
2745 opt_params.max_unrolled_loop_size = 400;
2746 opt_params.max_branches = 9999;
2749 /* Assure preconditions are met and go through all loops. */
2750 void loop_optimization(ir_graph *irg)
2758 /* Reset stats for this procedure */
2762 set_current_ir_graph(irg);
2765 assure_irg_outs(irg);
2767 /* NOTE: sets only the loop attribute of blocks, not nodes */
2768 /* NOTE: Kills links */
2769 assure_cf_loop(irg);
2771 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_PHI_LIST);
2772 collect_phiprojs(irg);
2773 ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
2775 loop = get_irg_loop(irg);
2776 sons = get_loop_n_sons(loop);
2778 loops = NEW_ARR_F(ir_loop *, 0);
2779 /* List all inner loops */
2780 for (nr = 0; nr < sons; ++nr) {
2781 find_innermost_loop(get_loop_son(loop, nr));
2784 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
2785 /* Set all links to NULL */
2786 irg_walk_graph(current_ir_graph, reset_link, NULL, NULL);
2788 for (i = 0; i < ARR_LEN(loops); ++i) {
2789 ir_loop *loop = loops[i];
2791 count_stats(stats.loops);
2793 /* Analyze and handle loop */
2796 /* Copied blocks do not have their phi list yet */
2797 collect_phiprojs(irg);
2799 /* Set links to NULL
2800 * TODO Still necessary? */
2801 irg_walk_graph(current_ir_graph, reset_link, NULL, NULL);
2807 ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
2808 ir_free_resources(irg, IR_RESOURCE_PHI_LIST);
2811 void do_loop_unrolling(ir_graph *irg)
2813 loop_op = loop_op_unrolling;
2815 DB((dbg, LEVEL_1, " >>> unrolling (Startnode %N) <<<\n",
2816 get_irg_start(irg)));
2818 loop_optimization(irg);
2820 DB((dbg, LEVEL_1, " >>> unrolling done (Startnode %N) <<<\n",
2821 get_irg_start(irg)));
2824 void do_loop_inversion(ir_graph *irg)
2826 loop_op = loop_op_inversion;
2828 DB((dbg, LEVEL_1, " >>> inversion (Startnode %N) <<<\n",
2829 get_irg_start(irg)));
2831 loop_optimization(irg);
2833 assure_cf_loop(irg);
2835 DB((dbg, LEVEL_1, " >>> inversion done (Startnode %N) <<<\n",
2836 get_irg_start(irg)));
2839 void do_loop_peeling(ir_graph *irg)
2841 loop_op = loop_op_peeling;
2843 DB((dbg, LEVEL_1, " >>> peeling (Startnode %N) <<<\n",
2844 get_irg_start(irg)));
2846 loop_optimization(irg);
2848 DB((dbg, LEVEL_1, " >>> peeling done (Startnode %N) <<<\n",
2849 get_irg_start(irg)));
2853 ir_graph_pass_t *loop_inversion_pass(const char *name)
2855 return def_graph_pass(name ? name : "loop_inversion", do_loop_inversion);
2858 ir_graph_pass_t *loop_unroll_pass(const char *name)
2860 return def_graph_pass(name ? name : "loop_unroll", do_loop_unrolling);
2863 ir_graph_pass_t *loop_peeling_pass(const char *name)
2865 return def_graph_pass(name ? name : "loop_peeling", do_loop_peeling);
2868 void firm_init_loop_opt(void)
2870 FIRM_DBG_REGISTER(dbg, "firm.opt.loop");