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)
381 ir_graph *irg = get_irn_irg(block);
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));
391 return new_r_Bad(irg, mode);
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 assert(block != get_irg_start_block(irg));
408 /* a Block with only 1 predecessor needs no Phi */
409 n_cfgpreds = get_Block_n_cfgpreds(block);
410 if (n_cfgpreds == 1) {
411 ir_node *pred_block = get_Block_cfgpred_block(block, 0);
414 DB((dbg, LEVEL_5, "ssa 1 pred: walk pred %N\n", pred_block));
416 value = search_def_and_create_phis(pred_block, mode, 0);
417 set_irn_link(block, value);
418 mark_irn_visited(block);
423 /* create a new Phi */
424 NEW_ARR_A(ir_node*, in, n_cfgpreds);
425 for (i = 0; i < n_cfgpreds; ++i)
426 in[i] = new_r_Dummy(irg, mode);
428 phi = new_r_Phi(block, n_cfgpreds, in, mode);
429 /* Important: always keep block phi list up to date. */
430 add_Block_phi(block, phi);
431 DB((dbg, LEVEL_5, "ssa phi creation: link new phi %N to block %N\n", phi, block));
432 set_irn_link(block, phi);
433 mark_irn_visited(block);
435 /* set Phi predecessors */
436 for (i = 0; i < n_cfgpreds; ++i) {
438 ir_node *pred_block = get_Block_cfgpred_block(block, i);
439 assert(pred_block != NULL);
440 pred_val = search_def_and_create_phis(pred_block, mode, 0);
442 assert(pred_val != NULL);
444 DB((dbg, LEVEL_5, "ssa phi pred:phi %N, pred %N\n", phi, pred_val));
445 set_irn_n(phi, i, pred_val);
453 * Given a set of values this function constructs SSA-form for the users of the
454 * first value (the users are determined through the out-edges of the value).
455 * Works without using the dominance tree.
457 static void construct_ssa(ir_node *orig_block, ir_node *orig_val,
458 ir_node *second_block, ir_node *second_val)
462 const ir_edge_t *edge;
463 const ir_edge_t *next;
465 assert(orig_block && orig_val && second_block && second_val &&
466 "no parameter of construct_ssa may be NULL");
468 if (orig_val == second_val)
471 irg = get_irn_irg(orig_val);
473 ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED);
474 inc_irg_visited(irg);
476 mode = get_irn_mode(orig_val);
477 set_irn_link(orig_block, orig_val);
478 mark_irn_visited(orig_block);
480 ssa_second_def_block = second_block;
481 ssa_second_def = second_val;
483 /* Only fix the users of the first, i.e. the original node */
484 foreach_out_edge_safe(orig_val, edge, next) {
485 ir_node *user = get_edge_src_irn(edge);
486 int j = get_edge_src_pos(edge);
487 ir_node *user_block = get_nodes_block(user);
494 DB((dbg, LEVEL_5, "original user %N\n", user));
497 ir_node *pred_block = get_Block_cfgpred_block(user_block, j);
498 newval = search_def_and_create_phis(pred_block, mode, 1);
500 newval = search_def_and_create_phis(user_block, mode, 1);
502 if (newval != user && !is_Bad(newval))
503 set_irn_n(user, j, newval);
506 ir_free_resources(irg, IR_RESOURCE_IRN_VISITED);
510 /***** Unrolling Helper Functions *****/
512 /* Assign the copy with index nr to node n */
513 static void set_unroll_copy(ir_node *n, int nr, ir_node *cp)
515 unrolling_node_info *info;
516 assert(nr != 0 && "0 reserved");
518 info = (unrolling_node_info *)phase_get_irn_data(phase, n);
522 info = XMALLOCZ(unrolling_node_info);
523 arr = NEW_ARR_F(ir_node *, unroll_nr);
525 memset(info->copies, 0, (unroll_nr) * sizeof(ir_node *));
527 phase_set_irn_data(phase, n, info);
532 info->copies[nr] = cp;
535 /* Returns a nodes copy if it exists, else NULL. */
536 static ir_node *get_unroll_copy(ir_node *n, int nr)
539 unrolling_node_info *info = (unrolling_node_info *)phase_get_irn_data(phase, n);
543 cp = info->copies[nr];
548 /***** Inversion Helper Functions *****/
550 /* Sets copy cp of node n. */
551 static void set_inversion_copy(ir_node *n, ir_node *cp)
553 phase_set_irn_data(phase, n, cp);
556 /* Getter of copy of n for inversion */
557 static ir_node *get_inversion_copy(ir_node *n)
559 ir_node *cp = (ir_node *)phase_get_irn_data(phase, n);
563 /* Resets block mark for given node. For use with walker */
564 static void reset_block_mark(ir_node *node, void * env)
569 set_Block_mark(node, 0);
572 /* Returns mark of node, or its block if node is not a block.
573 * Used in this context to determine if node is in the condition chain. */
574 static unsigned is_nodes_block_marked(ir_node* node)
577 return get_Block_mark(node);
579 return get_Block_mark(get_block(node));
582 /* Extends a nodes ins by node new.
583 * NOTE: This is slow if a node n needs to be extended more than once. */
584 static void extend_irn(ir_node *n, ir_node *newnode, int new_is_backedge)
588 int arity = get_irn_arity(n);
589 int new_arity = arity + 1;
592 NEW_ARR_A(int, bes, new_arity);
593 NEW_ARR_A(ir_node *, ins, new_arity);
596 /* Bes are important!
597 * Another way would be recreating the looptree,
598 * but after that we cannot distinguish already processed loops
599 * from not yet processed ones. */
601 for(i = 0; i < arity; ++i) {
602 bes[i] = is_backedge(n, i);
604 bes[i] = new_is_backedge;
607 for(i = 0; i < arity; ++i) {
608 ins[i] = get_irn_n(n, i);
612 set_irn_in(n, new_arity, ins);
616 for(i = 0; i < new_arity; ++i) {
623 /* Extends a block by a copy of its pred at pos,
624 * fixing also the phis in the same way. */
625 static void extend_ins_by_copy(ir_node *block, int pos)
630 assert(is_Block(block));
632 /* Extend block by copy of definition at pos */
633 pred = get_irn_n(block, pos);
634 new_in = get_inversion_copy(pred);
635 DB((dbg, LEVEL_5, "Extend block %N by %N cp of %N\n", block, new_in, pred));
636 extend_irn(block, new_in, 0);
638 /* Extend block phis by copy of definition at pos */
639 for_each_phi(block, phi) {
642 pred = get_irn_n(phi, pos);
643 cp = get_inversion_copy(pred);
644 /* If the phis in is not in the condition chain (eg. a constant),
645 * there is no copy. */
651 DB((dbg, LEVEL_5, "Extend phi %N by %N cp of %N\n", phi, new_in, pred));
652 extend_irn(phi, new_in, 0);
656 /* Returns the number of blocks backedges. With or without alien bes. */
657 static int get_backedge_n(ir_node *block, unsigned with_alien)
661 int arity = get_irn_arity(block);
663 assert(is_Block(block));
665 for (i = 0; i < arity; ++i) {
666 ir_node *pred = get_irn_n(block, i);
667 if (is_backedge(block, i) && (with_alien || is_in_loop(pred)))
673 /* Returns a raw copy of the given node.
674 * Attributes are kept/set according to the needs of loop inversion. */
675 static ir_node *copy_node(ir_node *node)
680 cp = exact_copy(node);
681 arity = get_irn_arity(node);
683 /* Keep backedge info */
684 for (i = 0; i < arity; ++i) {
685 if (is_backedge(node, i))
690 set_Block_mark(cp, 0);
698 * This walker copies all walked nodes.
699 * If the walk_condition is true for a node, it is copied.
700 * All nodes node_info->copy have to be NULL prior to every walk.
701 * Order of ins is important for later usage.
703 static void copy_walk(ir_node *node, walker_condition *walk_condition,
710 ir_graph *irg = current_ir_graph;
713 * break condition and cycle resolver, creating temporary node copies
715 if (get_irn_visited(node) >= get_irg_visited(irg)) {
716 /* Here we rely on nodestate's copy being initialized with NULL */
717 DB((dbg, LEVEL_5, "copy_walk: We have already visited %N\n", node));
718 if (get_inversion_copy(node) == NULL) {
719 cp = copy_node(node);
720 set_inversion_copy(node, cp);
722 DB((dbg, LEVEL_5, "The TEMP copy of %N is created %N\n", node, cp));
728 mark_irn_visited(node);
730 if (!is_Block(node)) {
731 ir_node *pred = get_nodes_block(node);
732 if (walk_condition(pred))
733 DB((dbg, LEVEL_5, "walk block %N\n", pred));
734 copy_walk(pred, walk_condition, set_loop);
737 arity = get_irn_arity(node);
739 NEW_ARR_A(ir_node *, cpin, arity);
741 for (i = 0; i < arity; ++i) {
742 ir_node *pred = get_irn_n(node, i);
744 if (walk_condition(pred)) {
745 DB((dbg, LEVEL_5, "walk node %N\n", pred));
746 copy_walk(pred, walk_condition, set_loop);
747 cpin[i] = get_inversion_copy(pred);
748 DB((dbg, LEVEL_5, "copy of %N gets new in %N which is copy of %N\n",
749 node, get_inversion_copy(pred), pred));
755 /* copy node / finalize temp node */
756 if (get_inversion_copy(node) == NULL) {
757 /* No temporary copy existent */
758 cp = copy_node(node);
759 set_inversion_copy(node, cp);
760 DB((dbg, LEVEL_5, "The FINAL copy of %N is CREATED %N\n", node, cp));
762 /* temporary copy is existent but without correct ins */
763 cp = get_inversion_copy(node);
764 DB((dbg, LEVEL_5, "The FINAL copy of %N is EXISTENT %N\n", node, cp));
767 if (!is_Block(node)) {
768 ir_node *cpblock = get_inversion_copy(get_nodes_block(node));
770 set_nodes_block(cp, cpblock );
772 add_Block_phi(cpblock, cp);
775 /* Keeps phi list of temporary node. */
776 set_irn_in(cp, ARR_LEN(cpin), cpin);
780 * This walker copies all walked nodes.
781 * If the walk_condition is true for a node, it is copied.
782 * All nodes node_info->copy have to be NULL prior to every walk.
783 * Order of ins is important for later usage.
784 * Takes copy_index, to phase-link copy at specific index.
786 static void copy_walk_n(ir_node *node, walker_condition *walk_condition,
795 * break condition and cycle resolver, creating temporary node copies
797 if (irn_visited(node)) {
798 /* Here we rely on nodestate's copy being initialized with NULL */
799 DB((dbg, LEVEL_5, "copy_walk: We have already visited %N\n", node));
800 if (get_unroll_copy(node, copy_index) == NULL) {
803 set_unroll_copy(node, copy_index, u);
804 DB((dbg, LEVEL_5, "The TEMP unknown of %N is created %N\n", node, u));
810 mark_irn_visited(node);
812 if (!is_Block(node)) {
813 ir_node *block = get_nodes_block(node);
814 if (walk_condition(block))
815 DB((dbg, LEVEL_5, "walk block %N\n", block));
816 copy_walk_n(block, walk_condition, copy_index);
819 arity = get_irn_arity(node);
820 NEW_ARR_A(ir_node *, cpin, arity);
822 for (i = 0; i < arity; ++i) {
823 ir_node *pred = get_irn_n(node, i);
825 if (walk_condition(pred)) {
826 DB((dbg, LEVEL_5, "walk node %N\n", pred));
827 copy_walk_n(pred, walk_condition, copy_index);
828 cpin[i] = get_unroll_copy(pred, copy_index);
834 /* copy node / finalize temp node */
835 cp = get_unroll_copy(node, copy_index);
836 if (cp == NULL || is_Unknown(cp)) {
837 cp = copy_node(node);
838 set_unroll_copy(node, copy_index, cp);
839 DB((dbg, LEVEL_5, "The FINAL copy of %N is CREATED %N\n", node, cp));
841 /* temporary copy is existent but without correct ins */
842 cp = get_unroll_copy(node, copy_index);
843 DB((dbg, LEVEL_5, "The FINAL copy of %N is EXISTENT %N\n", node, cp));
846 if (!is_Block(node)) {
847 ir_node *cpblock = get_unroll_copy(get_nodes_block(node), copy_index);
849 set_nodes_block(cp, cpblock );
851 add_Block_phi(cpblock, cp);
854 /* Keeps phi list of temporary node. */
855 set_irn_in(cp, ARR_LEN(cpin), cpin);
858 /* Removes alle Blocks with non marked predecessors from the condition chain. */
859 static void unmark_not_allowed_cc_blocks(void)
861 size_t blocks = ARR_LEN(cc_blocks);
864 for(i = 0; i < blocks; ++i) {
865 ir_node *block = cc_blocks[i];
867 int arity = get_irn_arity(block);
869 /* Head is an exception. */
870 if (block == loop_head)
873 for(a = 0; a < arity; ++a) {
874 if (! is_nodes_block_marked(get_irn_n(block, a))) {
875 set_Block_mark(block, 0);
876 --inversion_blocks_in_cc;
877 DB((dbg, LEVEL_5, "Removed %N from cc (blocks in cc %d)\n",
878 block, inversion_blocks_in_cc));
886 /* Unmarks all cc blocks using cc_blocks except head.
887 * TODO: invert head for unrolling? */
888 static void unmark_cc_blocks(void)
890 size_t blocks = ARR_LEN(cc_blocks);
893 for(i = 0; i < blocks; ++i) {
894 ir_node *block = cc_blocks[i];
896 /* TODO Head is an exception. */
897 /*if (block != loop_head)*/
898 set_Block_mark(block, 0);
900 /*inversion_blocks_in_cc = 1;*/
901 inversion_blocks_in_cc = 0;
904 loop_info.cc_size = 0;
908 * Populates head_entries with (node, pred_pos) tuple
909 * whereas the node's pred at pred_pos is in the cc but not the node itself.
910 * Also finds df loops inside the cc.
911 * Head and condition chain blocks have been marked previously.
913 static void get_head_outs(ir_node *node, void *env)
916 int arity = get_irn_arity(node);
919 for (i = 0; i < arity; ++i) {
920 if (!is_nodes_block_marked(node) && is_nodes_block_marked(get_irn_n(node, i))) {
924 /* Saving also predecessor seems redundant, but becomes
925 * necessary when changing position of it, before
926 * dereferencing it.*/
927 entry.pred = get_irn_n(node, i);
928 ARR_APP1(entry_edge, cur_head_outs, entry);
932 arity = get_irn_arity(loop_head);
934 /* Find df loops inside the cc */
935 if (is_Phi(node) && get_nodes_block(node) == loop_head) {
936 for (i = 0; i < arity; ++i) {
937 if (is_own_backedge(loop_head, i)) {
938 if (is_nodes_block_marked(get_irn_n(node, i))) {
942 entry.pred = get_irn_n(node, i);
943 ARR_APP1(entry_edge, head_df_loop, entry);
944 DB((dbg, LEVEL_5, "Found incc assignment node %N @%d is pred %N, graph %N %N\n",
945 node, i, entry.pred, current_ir_graph, get_irg_start_block(current_ir_graph)));
953 * Find condition chains, and add them to be inverted.
954 * A block belongs to the chain if a condition branches out of the loop.
955 * (Some blocks need to be removed once again.)
956 * Returns 1 if the given block belongs to the condition chain.
958 static unsigned find_condition_chain(ir_node *block)
960 const ir_edge_t *edge;
964 unsigned nodes_n = 0;
966 mark_irn_visited(block);
968 DB((dbg, LEVEL_5, "condition_chains for block %N\n", block));
971 foreach_out_edge_kind(block, edge, EDGE_KIND_NORMAL) {
975 /* Check if node count would exceed maximum cc size.
977 * This is not optimal, as we search depth-first and break here,
978 * continuing with another subtree. */
979 if (loop_info.cc_size + nodes_n > opt_params.max_cc_size) {
980 set_Block_mark(block, 0);
984 /* Check if block only has a jmp instruction. */
986 foreach_out_edge(block, edge) {
987 ir_node *src = get_edge_src_irn(edge);
989 if (! is_Block(src) && ! is_Jmp(src)) {
994 /* Check cf outs if one is leaving the loop,
995 * or if this node has a backedge. */
996 foreach_block_succ(block, edge) {
997 ir_node *src = get_edge_src_irn(edge);
998 int pos = get_edge_src_pos(edge);
1000 if (! is_in_loop(src))
1003 /* Inverting blocks with backedge outs leads to a cf edge
1004 * from the inverted head, into the inverted head (skipping the body).
1005 * As the body becomes the new loop head,
1006 * this would introduce another loop in the existing loop.
1007 * This loop inversion cannot cope with this case. */
1008 if (is_backedge(src, pos)) {
1014 /* We need all predecessors to already belong to the condition chain.
1015 * Example of wrong case: * == in cc
1026 /* Collect blocks containing only a Jmp.
1027 * Do not collect blocks with backedge outs. */
1028 if ((jmp_only == 1 || mark == 1) && has_be == 0) {
1029 set_Block_mark(block, 1);
1030 ++inversion_blocks_in_cc;
1031 loop_info.cc_size += nodes_n;
1032 DB((dbg, LEVEL_5, "block %N is part of condition chain\n", block));
1033 ARR_APP1(ir_node *, cc_blocks, block);
1035 set_Block_mark(block, 0);
1038 foreach_block_succ(block, edge) {
1039 ir_node *src = get_edge_src_irn( edge );
1041 if (is_in_loop(src) && ! irn_visited(src))
1042 find_condition_chain(src);
1049 * Rewires the copied condition chain. Removes backedges
1050 * as this condition chain is prior to the loop.
1051 * Copy of loop_head must have phi list and old (unfixed) backedge info of the loop head.
1052 * (loop_head is already fixed, we cannot rely on it.)
1054 static void fix_copy_inversion(void)
1059 ir_node *phi, *next;
1060 ir_node *head_cp = get_inversion_copy(loop_head);
1061 ir_graph *irg = get_irn_irg(head_cp);
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_r_Block(irg, 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 ir_graph *irg = get_irn_irg(loop_head);
1117 int arity = get_irn_arity(loop_head);
1118 int backedges = get_backedge_n(loop_head, 0);
1119 int new_arity = backedges;
1123 NEW_ARR_A(ir_node *, ins, new_arity);
1126 /* Keep only backedges */
1127 for(i = 0; i < arity; ++i) {
1128 if (is_own_backedge(loop_head, i))
1129 ins[pos++] = get_irn_n(loop_head, i);
1132 new_head = new_r_Block(irg, new_arity, ins);
1134 phis = NEW_ARR_F(ir_node *, 0);
1136 for_each_phi(loop_head, phi) {
1138 DB((dbg, LEVEL_5, "Fixing phi %N of loop head\n", phi));
1140 NEW_ARR_A(ir_node *, ins, new_arity);
1143 for (i = 0; i < arity; ++i) {
1144 ir_node *pred = get_irn_n(phi, i);
1146 if (is_own_backedge(loop_head, i)) {
1147 /* If assignment is in the condition chain,
1148 * we need to create a phi in the new loop head.
1149 * This can only happen for df, not cf. See find_condition_chains. */
1150 /*if (is_nodes_block_marked(pred)) {
1158 new_phi = new_rd_Phi(get_irn_dbg_info(phi),
1159 new_head, new_arity, ins,
1162 ARR_APP1(ir_node *, phis, new_phi);
1164 DB((dbg, LEVEL_5, "fix inverted head should exch %N by %N (pos %d)\n", phi, new_phi, pos ));
1168 for_each_phi_safe(get_Block_phis(loop_head), phi, next) {
1169 DB((dbg, LEVEL_5, "fix inverted exch phi %N by %N\n", phi, phis[pos]));
1170 if (phis[pos] != phi)
1171 exchange(phi, phis[pos++]);
1176 DB((dbg, LEVEL_5, "fix inverted head exch head block %N by %N\n", loop_head, new_head));
1177 exchange(loop_head, new_head);
1180 /* Does the loop inversion. */
1181 static void inversion_walk(entry_edge *head_entries)
1186 * The order of rewiring bottom-up is crucial.
1187 * Any change of the order leads to lost information that would be needed later.
1190 ir_reserve_resources(current_ir_graph, IR_RESOURCE_IRN_VISITED);
1192 /* 1. clone condition chain */
1193 inc_irg_visited(current_ir_graph);
1195 for (i = 0; i < ARR_LEN(head_entries); ++i) {
1196 entry_edge entry = head_entries[i];
1197 ir_node *pred = get_irn_n(entry.node, entry.pos);
1199 DB((dbg, LEVEL_5, "\nInit walk block %N\n", pred));
1201 copy_walk(pred, is_nodes_block_marked, cur_loop);
1204 ir_free_resources(current_ir_graph, IR_RESOURCE_IRN_VISITED);
1206 /* 2. Extends the head control flow successors ins
1207 * with the definitions of the copied head node. */
1208 for (i = 0; i < ARR_LEN(head_entries); ++i) {
1209 entry_edge head_out = head_entries[i];
1211 if (is_Block(head_out.node))
1212 extend_ins_by_copy(head_out.node, head_out.pos);
1215 /* 3. construct_ssa for users of definitions in the condition chain,
1216 * as there is now a second definition. */
1217 for (i = 0; i < ARR_LEN(head_entries); ++i) {
1218 entry_edge head_out = head_entries[i];
1220 /* Ignore keepalives */
1221 if (is_End(head_out.node))
1224 /* Construct ssa for assignments in the condition chain. */
1225 if (!is_Block(head_out.node)) {
1226 ir_node *pred, *cppred, *block, *cpblock;
1228 pred = head_out.pred;
1229 cppred = get_inversion_copy(pred);
1230 block = get_nodes_block(pred);
1231 cpblock = get_nodes_block(cppred);
1232 construct_ssa(block, pred, cpblock, cppred);
1237 * If there is an assignment in the condition chain
1238 * with a user also in the condition chain,
1239 * the dominance frontier is in the new loop head.
1240 * The dataflow loop is completely in the condition chain.
1245 * Phi_cp | | copied condition chain
1249 * Phi* | | new loop head with newly created phi.
1251 * Phi | | original, inverted condition chain
1256 for (i = 0; i < ARR_LEN(head_df_loop); ++i) {
1257 entry_edge head_out = head_df_loop[i];
1259 /* Construct ssa for assignments in the condition chain. */
1260 ir_node *pred, *cppred, *block, *cpblock;
1262 pred = head_out.pred;
1263 cppred = get_inversion_copy(pred);
1264 assert(cppred && pred);
1265 block = get_nodes_block(pred);
1266 cpblock = get_nodes_block(cppred);
1267 construct_ssa(block, pred, cpblock, cppred);
1270 /* 4. Remove the ins which are no backedges from the original condition chain
1271 * as the cc is now subsequent to the body. */
1272 fix_head_inversion();
1274 /* 5. Remove the backedges of the copied condition chain,
1275 * because it is going to be the new 'head' in advance to the loop. */
1276 fix_copy_inversion();
1280 /* Performs loop inversion of cur_loop if possible and reasonable. */
1281 static void loop_inversion(void)
1284 unsigned max_loop_nodes = opt_params.max_loop_size;
1285 unsigned max_loop_nodes_adapted;
1286 int depth_adaption = opt_params.depth_adaption;
1288 unsigned do_inversion = 1;
1289 unsigned has_cc = 0;
1291 /* Depth of 0 is the procedure and 1 a topmost loop. */
1292 loop_depth = get_loop_depth(cur_loop) - 1;
1294 /* Calculating in per mil. */
1295 max_loop_nodes_adapted = get_max_nodes_adapted(loop_depth);
1297 DB((dbg, LEVEL_1, "max_nodes: %d\nmax_nodes_adapted %d at depth of %d (adaption %d)\n",
1298 max_loop_nodes, max_loop_nodes_adapted, loop_depth, depth_adaption));
1300 if (! (loop_info.nodes > 0))
1303 #if LOOP_IGNORE_NODE_LIMITS
1304 DB((dbg, LEVEL_1, "WARNING: Loop node limitations ignored."));
1306 if (loop_info.nodes > max_loop_nodes) {
1307 /* Only for stats */
1308 DB((dbg, LEVEL_1, "Nodes %d > allowed nodes %d\n",
1309 loop_info.nodes, loop_depth, max_loop_nodes));
1310 count_stats(stats.too_large);
1312 /* Adaption might change it */
1315 /* Limit processing to loops smaller than given parameter. */
1316 if (loop_info.nodes > max_loop_nodes_adapted) {
1317 DB((dbg, LEVEL_1, "Nodes %d > allowed nodes (depth %d adapted) %d\n",
1318 loop_info.nodes, loop_depth, max_loop_nodes_adapted));
1319 count_stats(stats.too_large_adapted);
1323 if (loop_info.calls > opt_params.allowed_calls) {
1324 DB((dbg, LEVEL_1, "Calls %d > allowed calls %d\n",
1325 loop_info.calls, opt_params.allowed_calls));
1326 count_stats(stats.calls_limit);
1331 /*inversion_head_node_limit = INT_MAX;*/
1332 ir_reserve_resources(current_ir_graph, IR_RESOURCE_BLOCK_MARK);
1334 /* Reset block marks.
1335 * We use block marks to flag blocks of the original condition chain. */
1336 irg_walk_graph(current_ir_graph, reset_block_mark, NULL, NULL);
1338 /*loop_info.blocks = get_loop_n_blocks(cur_loop);*/
1339 cond_chain_entries = NEW_ARR_F(entry_edge, 0);
1340 head_df_loop = NEW_ARR_F(entry_edge, 0);
1342 /*head_inversion_node_count = 0;*/
1343 inversion_blocks_in_cc = 0;
1345 /* Use phase to keep copy of nodes from the condition chain. */
1346 phase = new_phase(current_ir_graph, phase_irn_init_default);
1348 /* Search for condition chains and temporarily save the blocks in an array. */
1349 cc_blocks = NEW_ARR_F(ir_node *, 0);
1350 inc_irg_visited(current_ir_graph);
1351 has_cc = find_condition_chain(loop_head);
1353 unmark_not_allowed_cc_blocks();
1354 DEL_ARR_F(cc_blocks);
1356 #if LOOP_IGNORE_NODE_LIMITS
1357 (void) unmark_cc_blocks;
1359 /* Condition chain too large.
1360 * Loop should better be small enough to fit into the cache. */
1361 /* TODO Of course, we should take a small enough cc in the first place,
1362 * which is not that simple. (bin packing) */
1363 if (loop_info.cc_size > opt_params.max_cc_size) {
1364 count_stats(stats.cc_limit_reached);
1368 /* Unmark cc blocks except the head.
1369 * Invert head only for possible unrolling. */
1375 /* We also catch endless loops here,
1376 * because they do not have a condition chain. */
1377 if (inversion_blocks_in_cc < 1) {
1380 "Loop contains %d (less than 1) invertible blocks => No Inversion done.\n",
1381 inversion_blocks_in_cc));
1385 cur_head_outs = NEW_ARR_F(entry_edge, 0);
1387 /* Get all edges pointing into the condition chain. */
1388 irg_walk_graph(current_ir_graph, get_head_outs, NULL, NULL);
1390 /* Do the inversion */
1391 inversion_walk(cur_head_outs);
1393 DEL_ARR_F(cur_head_outs);
1395 /* Duplicated blocks changed doms */
1396 set_irg_doms_inconsistent(current_ir_graph);
1397 /* Loop content changed */
1398 set_irg_loopinfo_inconsistent(current_ir_graph);
1400 count_stats(stats.inverted);
1405 DEL_ARR_F(cond_chain_entries);
1406 DEL_ARR_F(head_df_loop);
1408 ir_free_resources(current_ir_graph, IR_RESOURCE_BLOCK_MARK);
1411 /* Fix the original loop_heads ins for invariant unrolling case. */
1412 static void unrolling_fix_loop_head_inv(void)
1416 ir_node *proj = new_Proj(loop_info.duff_cond, mode_X, 0);
1417 ir_node *head_pred = get_irn_n(loop_head, loop_info.be_src_pos);
1418 ir_node *loop_condition = get_unroll_copy(head_pred, unroll_nr - 1);
1420 /* Original loop_heads ins are:
1421 * duff block and the own backedge */
1423 ins[0] = loop_condition;
1425 set_irn_in(loop_head, 2, ins);
1426 DB((dbg, LEVEL_4, "Rewire ins of block loophead %N to pred %N and duffs entry %N \n" , loop_head, ins[0], ins[1]));
1428 for_each_phi(loop_head, phi) {
1429 ir_node *pred = get_irn_n(phi, loop_info.be_src_pos);
1430 /* TODO we think it is a phi, but for Mergesort it is not the case.*/
1432 ir_node *last_pred = get_unroll_copy(pred, unroll_nr - 1);
1435 ins[1] = (ir_node*)get_irn_link(phi);
1436 set_irn_in(phi, 2, ins);
1437 DB((dbg, LEVEL_4, "Rewire ins of loophead phi %N to pred %N and duffs entry %N \n" , phi, ins[0], ins[1]));
1441 /* Removes previously created phis with only 1 in. */
1442 static void correct_phis(ir_node *node, void *env)
1446 if (is_Phi(node) && get_irn_arity(node) == 1) {
1450 in[0] = get_irn_n(node, 0);
1452 exch = new_rd_Phi(get_irn_dbg_info(node),
1453 get_nodes_block(node), 1, in,
1454 get_irn_mode(node));
1456 exchange(node, exch);
1460 /* Unrolling: Rewire floating copies. */
1461 static void place_copies(int copies)
1463 ir_node *loophead = loop_head;
1466 int be_src_pos = loop_info.be_src_pos;
1468 /* Serialize loops by fixing their head ins.
1469 * Processed are the copies.
1470 * The original loop is done after that, to keep backedge infos. */
1471 for (c = 0; c < copies; ++c) {
1472 ir_node *upper = get_unroll_copy(loophead, c);
1473 ir_node *lower = get_unroll_copy(loophead, c + 1);
1475 ir_node *topmost_be_block = get_nodes_block(get_irn_n(loophead, be_src_pos));
1477 /* Important: get the preds first and then their copy. */
1478 ir_node *upper_be_block = get_unroll_copy(topmost_be_block, c);
1479 ir_node *new_jmp = new_r_Jmp(upper_be_block);
1480 DB((dbg, LEVEL_5, " place_copies upper %N lower %N\n", upper, lower));
1482 DB((dbg, LEVEL_5, "topmost be block %N \n", topmost_be_block));
1484 if (loop_info.unroll_kind == constant) {
1487 set_irn_in(lower, 1, ins);
1489 for_each_phi(loophead, phi) {
1490 ir_node *topmost_def = get_irn_n(phi, be_src_pos);
1491 ir_node *upper_def = get_unroll_copy(topmost_def, c);
1492 ir_node *lower_phi = get_unroll_copy(phi, c + 1);
1494 /* It is possible, that the value used
1495 * in the OWN backedge path is NOT defined in this loop. */
1496 if (is_in_loop(topmost_def))
1499 ins[0] = topmost_def;
1501 set_irn_in(lower_phi, 1, ins);
1502 /* Need to replace phis with 1 in later. */
1505 /* Invariant case */
1506 /* Every node has 2 ins. One from the duff blocks
1507 * and one from the previously unrolled loop. */
1509 /* Calculate corresponding projection of mod result for this copy c */
1510 ir_node *proj = new_Proj(loop_info.duff_cond, mode_X, unroll_nr - c - 1);
1511 DB((dbg, LEVEL_4, "New duff proj %N\n" , proj));
1515 set_irn_in(lower, 2, ins);
1516 DB((dbg, LEVEL_4, "Rewire ins of Block %N to pred %N and duffs entry %N \n" , lower, ins[0], ins[1]));
1518 for_each_phi(loophead, phi) {
1519 ir_node *topmost_phi_pred = get_irn_n(phi, be_src_pos);
1520 ir_node *upper_phi_pred;
1524 lower_phi = get_unroll_copy(phi, c + 1);
1525 duff_phi = (ir_node*)get_irn_link(phi);
1526 DB((dbg, LEVEL_4, "DD Link of %N is %N\n" , phi, duff_phi));
1529 if (is_in_loop(topmost_phi_pred)) {
1530 upper_phi_pred = get_unroll_copy(topmost_phi_pred, c);
1532 upper_phi_pred = topmost_phi_pred;
1535 ins[0] = upper_phi_pred;
1537 set_irn_in(lower_phi, 2, ins);
1538 DB((dbg, LEVEL_4, "Rewire ins of %N to pred %N and duffs entry %N \n" , lower_phi, ins[0], ins[1]));
1543 /* Reconnect last copy. */
1544 for (i = 0; i < ARR_LEN(loop_entries); ++i) {
1545 entry_edge edge = loop_entries[i];
1546 /* Last copy is at the bottom */
1547 ir_node *new_pred = get_unroll_copy(edge.pred, copies);
1548 set_irn_n(edge.node, edge.pos, new_pred);
1551 /* Fix original loops head.
1552 * Done in the end, as ins and be info were needed before. */
1553 if (loop_info.unroll_kind == constant) {
1555 ir_node *head_pred = get_irn_n(loop_head, be_src_pos);
1556 ir_node *loop_condition = get_unroll_copy(head_pred, unroll_nr - 1);
1558 set_irn_n(loop_head, loop_info.be_src_pos, loop_condition);
1560 for_each_phi(loop_head, phi) {
1561 ir_node *pred = get_irn_n(phi, be_src_pos);
1564 /* It is possible, that the value used
1565 * in the OWN backedge path is NOT assigned in this loop. */
1566 if (is_in_loop(pred))
1567 last_pred = get_unroll_copy(pred, copies);
1570 set_irn_n(phi, be_src_pos, last_pred);
1574 unrolling_fix_loop_head_inv();
1578 /* Copies the cur_loop several times. */
1579 static void copy_loop(entry_edge *cur_loop_outs, int copies)
1583 ir_reserve_resources(current_ir_graph, IR_RESOURCE_IRN_VISITED);
1585 for (c = 0; c < copies; ++c) {
1588 inc_irg_visited(current_ir_graph);
1590 DB((dbg, LEVEL_5, " ### Copy_loop copy nr: %d ###\n", c));
1591 for (i = 0; i < ARR_LEN(cur_loop_outs); ++i) {
1592 entry_edge entry = cur_loop_outs[i];
1593 ir_node *pred = get_irn_n(entry.node, entry.pos);
1595 copy_walk_n(pred, is_in_loop, c + 1);
1599 ir_free_resources(current_ir_graph, IR_RESOURCE_IRN_VISITED);
1603 /* Creates a new phi from the given phi node omitting own bes,
1604 * using be_block as supplier of backedge informations. */
1605 static ir_node *clone_phis_sans_bes(ir_node *phi, ir_node *be_block, ir_node *dest_block)
1608 int arity = get_irn_arity(phi);
1612 assert(get_irn_arity(phi) == get_irn_arity(be_block));
1613 assert(is_Phi(phi));
1615 ins = NEW_ARR_F(ir_node *, arity);
1616 for (i = 0; i < arity; ++i) {
1617 if (! is_own_backedge(be_block, i)) {
1618 ins[c] = get_irn_n(phi, i);
1623 newphi = new_r_Phi(dest_block, c, ins, get_irn_mode(phi));
1625 set_irn_link(phi, newphi);
1626 DB((dbg, LEVEL_4, "Linking for duffs device %N to %N\n", phi, newphi));
1631 /* Creates a new block from the given block node omitting own bes,
1632 * using be_block as supplier of backedge informations. */
1633 static ir_node *clone_block_sans_bes(ir_node *node, ir_node *be_block)
1636 int arity = get_irn_arity(node);
1639 assert(get_irn_arity(node) == get_irn_arity(be_block));
1640 assert(is_Block(node));
1642 NEW_ARR_A(ir_node *, ins, arity);
1643 for (i = 0; i < arity; ++i) {
1644 if (! is_own_backedge(be_block, i)) {
1645 ins[c] = get_irn_n(node, i);
1650 return new_Block(c, ins);
1653 /* Creates a structure to calculate absolute value of node op.
1654 * Returns mux node with absolute value. */
1655 static ir_node *new_Abs(ir_node *op, ir_mode *mode)
1657 ir_graph *irg = get_irn_irg(op);
1658 ir_node *block = get_nodes_block(op);
1659 ir_node *zero = new_r_Const(irg, get_mode_null(mode));
1660 ir_node *cmp = new_r_Cmp(block, op, zero, ir_relation_less);
1661 ir_node *minus_op = new_r_Minus(block, op, mode);
1662 ir_node *mux = new_r_Mux(block, cmp, op, minus_op, mode);
1668 /* Creates blocks for duffs device, using previously obtained
1669 * informations about the iv.
1671 static void create_duffs_block(void)
1675 ir_node *block1, *count_block, *duff_block;
1676 ir_node *ems, *ems_mod, *ems_div, *ems_mod_proj, *cmp_null,
1677 *ems_mode_cond, *x_true, *x_false, *const_null;
1678 ir_node *true_val, *false_val;
1681 ir_node *duff_mod, *proj, *cond;
1683 ir_node *count, *correction, *unroll_c;
1684 ir_node *cmp_bad_count, *good_count, *bad_count, *count_phi, *bad_count_neg;
1687 mode = get_irn_mode(loop_info.end_val);
1688 const_null = new_Const(get_mode_null(mode));
1691 * 1. Calculate first approach to count.
1692 * Condition: (end - start) % step == 0 */
1693 block1 = clone_block_sans_bes(loop_head, loop_head);
1694 DB((dbg, LEVEL_4, "Duff block 1 %N\n", block1));
1696 /* Create loop entry phis in first duff block
1697 * as it becomes the loops preheader */
1698 for_each_phi(loop_head, phi) {
1699 /* Returns phis pred if phi would have arity 1*/
1700 ir_node *new_phi = clone_phis_sans_bes(phi, loop_head, block1);
1702 DB((dbg, LEVEL_4, "HEAD %N phi %N\n", loop_head, phi));
1703 DB((dbg, LEVEL_4, "BLOCK1 %N phi %N\n", block1, new_phi));
1706 ems = new_r_Sub(block1, loop_info.end_val, loop_info.start_val,
1707 get_irn_mode(loop_info.end_val));
1708 DB((dbg, LEVEL_4, "BLOCK1 sub %N\n", ems));
1711 ems = new_Sub(loop_info.end_val, loop_info.start_val,
1712 get_irn_mode(loop_info.end_val));
1714 DB((dbg, LEVEL_4, "mod ins %N %N\n", ems, loop_info.step));
1715 ems_mod = new_r_Mod(block1,
1720 op_pin_state_pinned);
1721 ems_div = new_r_Div(block1,
1726 op_pin_state_pinned);
1728 DB((dbg, LEVEL_4, "New module node %N\n", ems_mod));
1730 ems_mod_proj = new_r_Proj(ems_mod, mode_Iu, pn_Mod_res);
1731 cmp_null = new_r_Cmp(block1, ems_mod_proj, const_null, ir_relation_less);
1732 ems_mode_cond = new_r_Cond(block1, cmp_null);
1734 /* ems % step == 0 */
1735 x_true = new_r_Proj(ems_mode_cond, mode_X, pn_Cond_true);
1736 /* ems % step != 0 */
1737 x_false = new_r_Proj(ems_mode_cond, mode_X, pn_Cond_false);
1740 * Assures, duffs device receives a valid count.
1742 * decreasing: count < 0
1743 * increasing: count > 0
1748 count_block = new_Block(2, ins);
1749 DB((dbg, LEVEL_4, "Duff block 2 %N\n", count_block));
1752 /* Increase loop-taken-count depending on the loop condition
1753 * uses the latest iv to compare to. */
1754 if (loop_info.latest_value == 1) {
1755 /* ems % step == 0 : +0 */
1756 true_val = new_Const(get_mode_null(mode));
1757 /* ems % step != 0 : +1 */
1758 false_val = new_Const(get_mode_one(mode));
1760 ir_tarval *tv_two = new_tarval_from_long(2, mode);
1761 /* ems % step == 0 : +1 */
1762 true_val = new_Const(get_mode_one(mode));
1763 /* ems % step != 0 : +2 */
1764 false_val = new_Const(tv_two);
1770 correction = new_r_Phi(count_block, 2, ins, mode);
1772 count = new_r_Proj(ems_div, mode, pn_Div_res);
1774 /* (end - start) / step + correction */
1775 count = new_Add(count, correction, mode);
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 cmp_bad_count = new_r_Cmp(count_block, count, const_null,
1788 cmp_bad_count = new_r_Cmp(count_block, count, const_null,
1789 ir_relation_greater);
1792 bad_count_neg = new_r_Cond(count_block, cmp_bad_count);
1793 good_count = new_Proj(bad_count_neg, mode_X, pn_Cond_true);
1794 bad_count = new_Proj(ems_mode_cond, mode_X, pn_Cond_false);
1797 * Contains module to decide which loop to start from. */
1799 ins[0] = good_count;
1801 duff_block = new_Block(2, ins);
1802 DB((dbg, LEVEL_4, "Duff block 3 %N\n", duff_block));
1804 /* Get absolute value */
1805 ins[0] = new_Abs(count, mode);
1806 /* Manually feed the aforementioned count = 1 (bad case)*/
1807 ins[1] = new_Const(get_mode_one(mode));
1808 count_phi = new_r_Phi(duff_block, 2, ins, mode);
1810 unroll_c = new_Const(new_tarval_from_long((long)unroll_nr, mode));
1812 /* count % unroll_nr */
1813 duff_mod = new_r_Mod(duff_block,
1818 op_pin_state_pinned);
1821 proj = new_Proj(duff_mod, mode, pn_Mod_res);
1822 /* condition does NOT create itself in the block of the proj! */
1823 cond = new_r_Cond(duff_block, proj);
1825 loop_info.duff_cond = cond;
1828 /* Returns 1 if given node is not in loop,
1829 * or if it is a phi of the loop head with only loop invariant defs.
1831 static unsigned is_loop_invariant_def(ir_node *node)
1835 if (! is_in_loop(node)) {
1836 DB((dbg, LEVEL_4, "Not in loop %N\n", node));
1837 /* || is_Const(node) || is_SymConst(node)) {*/
1841 /* If this is a phi of the loophead shared by more than 1 loop,
1842 * we need to check if all defs are not in the loop. */
1845 block = get_nodes_block(node);
1847 /* To prevent unexpected situations. */
1848 if (block != loop_head) {
1852 for (i = 0; i < get_irn_arity(node); ++i) {
1853 /* Check if all bes are just loopbacks. */
1854 if (is_own_backedge(block, i) && get_irn_n(node, i) != node)
1857 DB((dbg, LEVEL_4, "invar %N\n", node));
1860 DB((dbg, LEVEL_4, "Not invar %N\n", node));
1865 /* Returns 1 if one pred of node is invariant and the other is not.
1866 * invar_pred and other are set analogously. */
1867 static unsigned get_invariant_pred(ir_node *node, ir_node **invar_pred, ir_node **other)
1869 ir_node *pred0 = get_irn_n(node, 0);
1870 ir_node *pred1 = get_irn_n(node, 1);
1875 if (is_loop_invariant_def(pred0)) {
1876 DB((dbg, LEVEL_4, "pred0 invar %N\n", pred0));
1877 *invar_pred = pred0;
1881 if (is_loop_invariant_def(pred1)) {
1882 DB((dbg, LEVEL_4, "pred1 invar %N\n", pred1));
1884 if (*invar_pred != NULL) {
1885 /* RETURN. We do not want both preds to be invariant. */
1890 *invar_pred = pred1;
1893 DB((dbg, LEVEL_4, "pred1 not invar %N\n", pred1));
1895 if (*invar_pred != NULL)
1902 /* Starts from a phi that may belong to an iv.
1903 * If an add forms a loop with iteration_phi,
1904 * and add uses a constant, 1 is returned
1905 * and 'start' as well as 'add' are sane. */
1906 static unsigned get_start_and_add(ir_node *iteration_phi, unrolling_kind_flag role)
1909 ir_node *found_add = loop_info.add;
1910 int arity = get_irn_arity(iteration_phi);
1912 DB((dbg, LEVEL_4, "Find start and add from %N\n", iteration_phi));
1914 for (i = 0; i < arity; ++i) {
1916 /* Find start_val which needs to be pred of the iteration_phi.
1917 * If start_val already known, sanity check. */
1918 if (!is_backedge(get_nodes_block(loop_info.iteration_phi), i)) {
1919 ir_node *found_start_val = get_irn_n(loop_info.iteration_phi, i);
1921 DB((dbg, LEVEL_4, "found_start_val %N\n", found_start_val));
1923 /* We already found a start_val it has to be always the same. */
1924 if (loop_info.start_val && found_start_val != loop_info.start_val)
1927 if ((role == constant) && !(is_SymConst(found_start_val) || is_Const(found_start_val)))
1929 else if((role == constant) && !(is_loop_invariant_def(found_start_val)))
1932 loop_info.start_val = found_start_val;
1935 /* The phi has to be in the loop head.
1936 * Follow all own backedges. Every value supplied from these preds of the phi
1937 * needs to origin from the same add. */
1938 if (is_own_backedge(get_nodes_block(loop_info.iteration_phi), i)) {
1939 ir_node *new_found = get_irn_n(loop_info.iteration_phi,i);
1941 DB((dbg, LEVEL_4, "is add? %N\n", new_found));
1943 if (! (is_Add(new_found) || is_Sub(new_found)) || (found_add && found_add != new_found))
1946 found_add = new_found;
1950 loop_info.add = found_add;
1956 /* Returns 1 if one pred of node is a const value and the other is not.
1957 * const_pred and other are set analogously. */
1958 static unsigned get_const_pred(ir_node *node, ir_node **const_pred, ir_node **other)
1960 ir_node *pred0 = get_irn_n(node, 0);
1961 ir_node *pred1 = get_irn_n(node, 1);
1963 DB((dbg, LEVEL_4, "Checking for constant pred of %N\n", node));
1968 /*DB((dbg, LEVEL_4, "is %N const\n", pred0));*/
1969 if (is_Const(pred0) || is_SymConst(pred0)) {
1970 *const_pred = pred0;
1974 /*DB((dbg, LEVEL_4, "is %N const\n", pred1));*/
1975 if (is_Const(pred1) || is_SymConst(pred1)) {
1976 if (*const_pred != NULL) {
1977 /* RETURN. We do not want both preds to be constant. */
1982 *const_pred = pred1;
1985 if (*const_pred == NULL)
1991 /* Returns 1 if loop exits within 2 steps of the iv.
1992 * Norm_proj means we do not exit the loop.*/
1993 static unsigned simulate_next(ir_tarval **count_tar,
1994 ir_tarval *stepped, ir_tarval *step_tar, ir_tarval *end_tar,
1995 ir_relation norm_proj)
1999 DB((dbg, LEVEL_4, "Loop taken if (stepped)%ld %s (end)%ld ",
2000 get_tarval_long(stepped),
2001 get_relation_string((norm_proj)),
2002 get_tarval_long(end_tar)));
2003 DB((dbg, LEVEL_4, "comparing latest value %d\n", loop_info.latest_value));
2005 /* If current iv does not stay in the loop,
2006 * this run satisfied the exit condition. */
2007 if (! (tarval_cmp(stepped, end_tar) & norm_proj))
2010 DB((dbg, LEVEL_4, "Result: (stepped)%ld IS %s (end)%ld\n",
2011 get_tarval_long(stepped),
2012 get_relation_string(tarval_cmp(stepped, end_tar)),
2013 get_tarval_long(end_tar)));
2016 if (is_Add(loop_info.add))
2017 next = tarval_add(stepped, step_tar);
2020 next = tarval_sub(stepped, step_tar, get_irn_mode(loop_info.end_val));
2022 DB((dbg, LEVEL_4, "Loop taken if %ld %s %ld ",
2023 get_tarval_long(next),
2024 get_relation_string(norm_proj),
2025 get_tarval_long(end_tar)));
2026 DB((dbg, LEVEL_4, "comparing latest value %d\n", loop_info.latest_value));
2028 /* Increase steps. */
2029 *count_tar = tarval_add(*count_tar, get_tarval_one(get_tarval_mode(*count_tar)));
2031 /* Next has to fail the loop condition, or we will never exit. */
2032 if (! (tarval_cmp(next, end_tar) & norm_proj))
2038 /* Check if loop meets requirements for a 'simple loop':
2039 * - Exactly one cf out
2041 * - Max nodes after unrolling
2045 * Returns Projection of cmp node or NULL; */
2046 static ir_node *is_simple_loop(void)
2049 ir_node *loop_block, *exit_block, *projx, *cond, *cmp;
2051 /* Maximum of one condition, and no endless loops. */
2052 if (loop_info.cf_outs != 1)
2055 DB((dbg, LEVEL_4, "1 loop exit\n"));
2057 #if LOOP_IGNORE_NODE_LIMITS
2058 /* Ignore loop size. Probably not wise in other than testcases. */
2059 loop_info.max_unroll = 40;
2061 /* Calculate maximum unroll_nr keeping node count below limit. */
2062 loop_info.max_unroll = (int)((double)opt_params.max_unrolled_loop_size / (double)loop_info.nodes);
2063 if (loop_info.max_unroll < 2) {
2064 count_stats(stats.too_large);
2070 DB((dbg, LEVEL_4, "maximum unroll factor %u, to not exceed node limit \n",
2071 opt_params.max_unrolled_loop_size));
2073 arity = get_irn_arity(loop_head);
2074 /* RETURN if we have more than 1 be. */
2075 /* Get my backedges without alien bes. */
2077 for (i = 0; i < arity; ++i) {
2078 ir_node *pred = get_irn_n(loop_head, i);
2079 if (is_own_backedge(loop_head, i)) {
2081 /* Our simple loops may have only one backedge. */
2084 loop_block = get_nodes_block(pred);
2085 loop_info.be_src_pos = i;
2090 DB((dbg, LEVEL_4, "loop has 1 own backedge.\n"));
2092 exit_block = get_nodes_block(loop_info.cf_out.pred);
2093 /* The loop has to be tail-controlled.
2094 * This can be changed/improved,
2095 * but we would need a duff iv. */
2096 if (exit_block != loop_block)
2099 DB((dbg, LEVEL_4, "tail-controlled loop.\n"));
2101 /* find value on which loop exit depends */
2102 projx = loop_info.cf_out.pred;
2103 cond = get_irn_n(projx, 0);
2104 cmp = get_irn_n(cond, 0);
2109 DB((dbg, LEVEL_5, "projection is %s\n", get_relation_string(get_Proj_proj(projx))));
2111 switch(get_Proj_proj(projx)) {
2113 loop_info.exit_cond = 0;
2116 loop_info.exit_cond = 1;
2119 panic("Cond Proj_proj other than true/false");
2122 DB((dbg, LEVEL_4, "Valid Cmp.\n"));
2126 /* Returns 1 if all nodes are mode_Iu or mode_Is. */
2127 static unsigned are_mode_I(ir_node *n1, ir_node* n2, ir_node *n3)
2129 ir_mode *m1 = get_irn_mode(n1);
2130 ir_mode *m2 = get_irn_mode(n2);
2131 ir_mode *m3 = get_irn_mode(n3);
2133 if ((m1 == mode_Iu && m2 == mode_Iu && m3 == mode_Iu) ||
2134 (m1 == mode_Is && m2 == mode_Is && m3 == mode_Is))
2140 /* Checks if cur_loop is a simple tail-controlled counting loop
2141 * with start and end value loop invariant, step constant. */
2142 static unsigned get_unroll_decision_invariant(void)
2145 ir_node *projres, *loop_condition, *iteration_path;
2146 unsigned success, is_latest_val;
2147 ir_tarval *step_tar;
2151 /* RETURN if loop is not 'simple' */
2152 projres = is_simple_loop();
2153 if (projres == NULL)
2156 /* Use a minimal size for the invariant unrolled loop,
2157 * as duffs device produces overhead */
2158 if (loop_info.nodes < opt_params.invar_unrolling_min_size)
2161 loop_condition = get_irn_n(projres, 0);
2163 success = get_invariant_pred(loop_condition, &loop_info.end_val, &iteration_path);
2164 DB((dbg, LEVEL_4, "pred invar %d\n", success));
2169 DB((dbg, LEVEL_4, "Invariant End_val %N, other %N\n", loop_info.end_val, iteration_path));
2171 /* We may find the add or the phi first.
2172 * Until now we only have end_val. */
2173 if (is_Add(iteration_path) || is_Sub(iteration_path)) {
2175 /* We test against the latest value of the iv. */
2178 loop_info.add = iteration_path;
2179 DB((dbg, LEVEL_4, "Case 1: Got add %N (maybe not sane)\n", loop_info.add));
2181 /* Preds of the add should be step and the iteration_phi */
2182 success = get_const_pred(loop_info.add, &loop_info.step, &loop_info.iteration_phi);
2186 DB((dbg, LEVEL_4, "Got step %N\n", loop_info.step));
2188 if (! is_Phi(loop_info.iteration_phi))
2191 DB((dbg, LEVEL_4, "Got phi %N\n", loop_info.iteration_phi));
2194 * Does necessary sanity check of add, if it is already set. */
2195 success = get_start_and_add(loop_info.iteration_phi, invariant);
2199 DB((dbg, LEVEL_4, "Got start A %N\n", loop_info.start_val));
2201 } else if (is_Phi(iteration_path)) {
2202 ir_node *new_iteration_phi;
2204 /* We compare with the value the iv had entering this run. */
2207 loop_info.iteration_phi = iteration_path;
2208 DB((dbg, LEVEL_4, "Case 2: Got phi %N\n", loop_info.iteration_phi));
2210 /* Find start_val and add-node.
2211 * Does necessary sanity check of add, if it is already set. */
2212 success = get_start_and_add(loop_info.iteration_phi, invariant);
2216 DB((dbg, LEVEL_4, "Got start B %N\n", loop_info.start_val));
2217 DB((dbg, LEVEL_4, "Got add or sub %N\n", loop_info.add));
2219 success = get_const_pred(loop_info.add, &loop_info.step, &new_iteration_phi);
2223 DB((dbg, LEVEL_4, "Got step (B) %N\n", loop_info.step));
2225 if (loop_info.iteration_phi != new_iteration_phi)
2232 mode = get_irn_mode(loop_info.end_val);
2234 DB((dbg, LEVEL_4, "start %N, end %N, step %N\n",
2235 loop_info.start_val, loop_info.end_val, loop_info.step));
2237 if (mode != mode_Is && mode != mode_Iu)
2240 /* TODO necessary? */
2241 if (!are_mode_I(loop_info.start_val, loop_info.step, loop_info.end_val))
2244 DB((dbg, LEVEL_4, "mode integer\n"));
2246 step_tar = get_Const_tarval(loop_info.step);
2248 if (tarval_is_null(step_tar)) {
2249 /* TODO Might be worth a warning. */
2253 DB((dbg, LEVEL_4, "step is not 0\n"));
2255 create_duffs_block();
2257 return loop_info.max_unroll;
2260 /* Returns unroll factor,
2261 * given maximum unroll factor and number of loop passes. */
2262 static unsigned get_preferred_factor_constant(ir_tarval *count_tar)
2264 ir_tarval *tar_6, *tar_5, *tar_4, *tar_3, *tar_2;
2266 ir_mode *mode = get_irn_mode(loop_info.end_val);
2268 tar_6 = new_tarval_from_long(6, mode);
2269 tar_5 = new_tarval_from_long(5, mode);
2270 tar_4 = new_tarval_from_long(4, mode);
2271 tar_3 = new_tarval_from_long(3, mode);
2272 tar_2 = new_tarval_from_long(2, mode);
2274 /* loop passes % {6, 5, 4, 3, 2} == 0 */
2275 if (tarval_is_null(tarval_mod(count_tar, tar_6)))
2277 else if (tarval_is_null(tarval_mod(count_tar, tar_5)))
2279 else if (tarval_is_null(tarval_mod(count_tar, tar_4)))
2281 else if (tarval_is_null(tarval_mod(count_tar, tar_3)))
2283 else if (tarval_is_null(tarval_mod(count_tar, tar_2)))
2286 /* gcd(max_unroll, count_tar) */
2287 int a = loop_info.max_unroll;
2288 int b = (int)get_tarval_long(count_tar);
2291 DB((dbg, LEVEL_4, "gcd of max_unroll %d and count_tar %d: ", a, b));
2298 DB((dbg, LEVEL_4, "%d\n", a));
2302 DB((dbg, LEVEL_4, "preferred unroll factor %d\n", prefer));
2305 * If our preference is greater than the allowed unroll factor
2306 * we either might reduce the preferred factor and prevent a duffs device block,
2307 * or create a duffs device block, from which in this case (constants only)
2308 * we know the startloop at compiletime.
2309 * The latter yields the following graphs.
2310 * but for code generation we would want to use graph A.
2311 * The graphs are equivalent. So, we can only reduce the preferred factor.
2321 if (prefer <= loop_info.max_unroll)
2326 if (loop_info.max_unroll >= 3)
2328 else if (loop_info.max_unroll >= 2)
2334 if (loop_info.max_unroll >= 2)
2345 /* Check if cur_loop is a simple counting loop.
2346 * Start, step and end are constants.
2347 * TODO The whole constant case should use procedures similar to
2348 * the invariant case, as they are more versatile. */
2350 static unsigned get_unroll_decision_constant(void)
2352 ir_node *cmp, *iteration_path;
2353 unsigned success, is_latest_val;
2354 ir_tarval *start_tar, *end_tar, *step_tar, *diff_tar, *count_tar;
2356 ir_relation proj_proj, norm_proj;
2359 /* RETURN if loop is not 'simple' */
2360 cmp = is_simple_loop();
2364 /* One in of the loop condition needs to be loop invariant. => end_val
2365 * The other in is assigned by an add. => add
2366 * The add uses a loop invariant value => step
2367 * and a phi with a loop invariant start_val and the add node as ins.
2379 success = get_const_pred(cmp, &loop_info.end_val, &iteration_path);
2383 DB((dbg, LEVEL_4, "End_val %N, other %N\n", loop_info.end_val, iteration_path));
2385 /* We may find the add or the phi first.
2386 * Until now we only have end_val. */
2387 if (is_Add(iteration_path) || is_Sub(iteration_path)) {
2389 /* We test against the latest value of the iv. */
2392 loop_info.add = iteration_path;
2393 DB((dbg, LEVEL_4, "Case 2: Got add %N (maybe not sane)\n", loop_info.add));
2395 /* Preds of the add should be step and the iteration_phi */
2396 success = get_const_pred(loop_info.add, &loop_info.step, &loop_info.iteration_phi);
2400 DB((dbg, LEVEL_4, "Got step %N\n", loop_info.step));
2402 if (! is_Phi(loop_info.iteration_phi))
2405 DB((dbg, LEVEL_4, "Got phi %N\n", loop_info.iteration_phi));
2408 * Does necessary sanity check of add, if it is already set. */
2409 success = get_start_and_add(loop_info.iteration_phi, constant);
2413 DB((dbg, LEVEL_4, "Got start %N\n", loop_info.start_val));
2415 } else if (is_Phi(iteration_path)) {
2416 ir_node *new_iteration_phi;
2418 /* We compare with the value the iv had entering this run. */
2421 loop_info.iteration_phi = iteration_path;
2422 DB((dbg, LEVEL_4, "Case 1: Got phi %N \n", loop_info.iteration_phi));
2424 /* Find start_val and add-node.
2425 * Does necessary sanity check of add, if it is already set. */
2426 success = get_start_and_add(loop_info.iteration_phi, constant);
2430 DB((dbg, LEVEL_4, "Got start %N\n", loop_info.start_val));
2431 DB((dbg, LEVEL_4, "Got add or sub %N\n", loop_info.add));
2433 success = get_const_pred(loop_info.add, &loop_info.step, &new_iteration_phi);
2437 DB((dbg, LEVEL_4, "Got step %N\n", loop_info.step));
2439 if (loop_info.iteration_phi != new_iteration_phi)
2447 mode = get_irn_mode(loop_info.end_val);
2449 DB((dbg, LEVEL_4, "start %N, end %N, step %N\n",
2450 loop_info.start_val, loop_info.end_val, loop_info.step));
2452 if (mode != mode_Is && mode != mode_Iu)
2455 /* TODO necessary? */
2456 if (!are_mode_I(loop_info.start_val, loop_info.step, loop_info.end_val))
2459 DB((dbg, LEVEL_4, "mode integer\n"));
2461 end_tar = get_Const_tarval(loop_info.end_val);
2462 start_tar = get_Const_tarval(loop_info.start_val);
2463 step_tar = get_Const_tarval(loop_info.step);
2465 if (tarval_is_null(step_tar))
2466 /* TODO Might be worth a warning. */
2469 DB((dbg, LEVEL_4, "step is not 0\n"));
2471 if ((!tarval_is_negative(step_tar)) ^ (!is_Sub(loop_info.add)))
2472 loop_info.decreasing = 1;
2474 diff_tar = tarval_sub(end_tar, start_tar, mode);
2476 /* We need at least count_tar steps to be close to end_val, maybe more.
2477 * No way, that we have gone too many steps.
2478 * This represents the 'latest value'.
2479 * (If condition checks against latest value, is checked later) */
2480 count_tar = tarval_div(diff_tar, step_tar);
2482 /* Iv will not pass end_val (except overflows).
2483 * Nothing done, as it would yield to no advantage. */
2484 if (tarval_is_negative(count_tar)) {
2485 DB((dbg, LEVEL_4, "Loop is endless or never taken."));
2486 /* TODO Might be worth a warning. */
2490 count_stats(stats.u_simple_counting_loop);
2492 loop_info.latest_value = is_latest_val;
2495 if (! is_simple_counting_loop(&count_tar))
2499 /* stepped can be negative, if step < 0 */
2500 stepped = tarval_mul(count_tar, step_tar);
2502 /* step as close to end_val as possible, */
2503 /* |stepped| <= |end_tar|, and dist(stepped, end_tar) is smaller than a step. */
2504 if (is_Sub(loop_info.add))
2505 stepped = tarval_sub(start_tar, stepped, mode_Is);
2507 stepped = tarval_add(start_tar, stepped);
2509 DB((dbg, LEVEL_4, "stepped to %ld\n", get_tarval_long(stepped)));
2511 proj_proj = get_Cmp_relation(cmp);
2512 /* Assure that norm_proj is the stay-in-loop case. */
2513 if (loop_info.exit_cond == 1)
2514 norm_proj = get_negated_relation(proj_proj);
2516 norm_proj = proj_proj;
2518 DB((dbg, LEVEL_4, "normalized projection %s\n", get_relation_string(norm_proj)));
2519 /* Executed at most once (stay in counting loop if a Eq b) */
2520 if (norm_proj == ir_relation_equal)
2521 /* TODO Might be worth a warning. */
2524 /* calculates next values and increases count_tar according to it */
2525 success = simulate_next(&count_tar, stepped, step_tar, end_tar, norm_proj);
2529 /* We run loop once more, if we compare to the
2530 * not yet in-/decreased iv. */
2531 if (is_latest_val == 0) {
2532 DB((dbg, LEVEL_4, "condition uses not latest iv value\n"));
2533 count_tar = tarval_add(count_tar, get_tarval_one(mode));
2536 DB((dbg, LEVEL_4, "loop taken %ld times\n", get_tarval_long(count_tar)));
2538 /* Assure the loop is taken at least 1 time. */
2539 if (tarval_is_null(count_tar)) {
2540 /* TODO Might be worth a warning. */
2544 loop_info.count_tar = count_tar;
2545 return get_preferred_factor_constant(count_tar);
2551 static void unroll_loop(void)
2554 if (! (loop_info.nodes > 0))
2557 #if LOOP_IGNORE_NODE_LIMITS
2558 DB((dbg, LEVEL_1, "WARNING: Loop node limitations ignored."));
2560 if (loop_info.nodes > opt_params.max_unrolled_loop_size) {
2561 DB((dbg, LEVEL_2, "Nodes %d > allowed nodes %d\n",
2562 loop_info.nodes, opt_params.max_unrolled_loop_size));
2563 count_stats(stats.too_large);
2567 if (loop_info.calls > 0) {
2568 DB((dbg, LEVEL_2, "Calls %d > allowed calls 0\n",
2570 count_stats(stats.calls_limit);
2577 /* get_unroll_decision_constant and invariant are completely
2578 * independent for flexibility.
2579 * Some checks may be performed twice. */
2581 /* constant case? */
2582 if (opt_params.allow_const_unrolling)
2583 unroll_nr = get_unroll_decision_constant();
2584 if (unroll_nr > 1) {
2585 loop_info.unroll_kind = constant;
2588 /* invariant case? */
2589 if (opt_params.allow_invar_unrolling)
2590 unroll_nr = get_unroll_decision_invariant();
2592 loop_info.unroll_kind = invariant;
2595 DB((dbg, LEVEL_2, " *** Unrolling %d times ***\n", unroll_nr));
2597 if (unroll_nr > 1) {
2598 loop_entries = NEW_ARR_F(entry_edge, 0);
2601 irg_walk_graph(current_ir_graph, get_loop_entries, NULL, NULL);
2603 if (loop_info.unroll_kind == constant) {
2604 if ((int)get_tarval_long(loop_info.count_tar) == unroll_nr)
2605 loop_info.needs_backedge = 0;
2607 loop_info.needs_backedge = 1;
2609 loop_info.needs_backedge = 1;
2612 /* Use phase to keep copy of nodes from the condition chain. */
2613 phase = new_phase(current_ir_graph, phase_irn_init_default);
2615 /* Copies the loop */
2616 copy_loop(loop_entries, unroll_nr - 1);
2618 /* Line up the floating copies. */
2619 place_copies(unroll_nr - 1);
2621 /* Remove phis with 1 in
2622 * If there were no nested phis, this would not be necessary.
2623 * Avoiding the creation in the first place
2624 * leads to complex special cases. */
2625 irg_walk_graph(current_ir_graph, correct_phis, NULL, NULL);
2627 if (loop_info.unroll_kind == constant)
2628 count_stats(stats.constant_unroll);
2630 count_stats(stats.invariant_unroll);
2632 set_irg_doms_inconsistent(current_ir_graph);
2633 set_irg_loopinfo_inconsistent(current_ir_graph);
2635 DEL_ARR_F(loop_entries);
2640 /* Analyzes the loop, and checks if size is within allowed range.
2641 * Decides if loop will be processed. */
2642 static void init_analyze(ir_loop *loop)
2647 loop_head_valid = 1;
2649 /* Reset loop info */
2650 memset(&loop_info, 0, sizeof(loop_info_t));
2652 DB((dbg, LEVEL_1, " >>>> current loop includes node %N <<<\n",
2653 get_loop_node(loop, 0)));
2655 /* Collect loop informations: head, node counts. */
2656 irg_walk_graph(current_ir_graph, get_loop_info, NULL, NULL);
2658 /* RETURN if there is no valid head */
2659 if (!loop_head || !loop_head_valid) {
2660 DB((dbg, LEVEL_1, "No valid loop head. Nothing done.\n"));
2663 DB((dbg, LEVEL_1, "Loophead: %N\n", loop_head));
2666 if (loop_info.branches > opt_params.max_branches) {
2667 DB((dbg, LEVEL_1, "Branches %d > allowed branches %d\n",
2668 loop_info.branches, opt_params.max_branches));
2669 count_stats(stats.calls_limit);
2674 case loop_op_inversion:
2678 case loop_op_unrolling:
2683 panic("Loop optimization not implemented.");
2685 DB((dbg, LEVEL_1, " <<<< end of loop with node %N >>>>\n",
2686 get_loop_node(loop, 0)));
2689 /* Find innermost loops and add them to loops. */
2690 static void find_innermost_loop(ir_loop *loop)
2692 /* descend into sons */
2693 size_t sons = get_loop_n_sons(loop);
2696 ARR_APP1(ir_loop *, loops, loop);
2699 for (s = 0; s < sons; ++s) {
2700 find_innermost_loop(get_loop_son(loop, s));
2705 static void set_loop_params(void)
2707 opt_params.max_loop_size = 100;
2708 opt_params.depth_adaption = -50;
2709 opt_params.count_phi = 1;
2710 opt_params.count_proj = 0;
2711 opt_params.allowed_calls = 0;
2713 opt_params.max_cc_size = 5;
2716 opt_params.allow_const_unrolling = 1;
2717 opt_params.allow_invar_unrolling = 0;
2719 opt_params.invar_unrolling_min_size = 20;
2720 opt_params.max_unrolled_loop_size = 400;
2721 opt_params.max_branches = 9999;
2724 /* Assure preconditions are met and go through all loops. */
2725 void loop_optimization(ir_graph *irg)
2733 /* Reset stats for this procedure */
2737 set_current_ir_graph(irg);
2740 assure_irg_outs(irg);
2742 /* NOTE: sets only the loop attribute of blocks, not nodes */
2743 /* NOTE: Kills links */
2744 assure_cf_loop(irg);
2746 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_PHI_LIST);
2747 collect_phiprojs(irg);
2748 ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
2750 loop = get_irg_loop(irg);
2751 sons = get_loop_n_sons(loop);
2753 loops = NEW_ARR_F(ir_loop *, 0);
2754 /* List all inner loops */
2755 for (nr = 0; nr < sons; ++nr) {
2756 find_innermost_loop(get_loop_son(loop, nr));
2759 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
2760 /* Set all links to NULL */
2761 irg_walk_graph(current_ir_graph, reset_link, NULL, NULL);
2763 for (i = 0; i < ARR_LEN(loops); ++i) {
2764 ir_loop *loop = loops[i];
2766 count_stats(stats.loops);
2768 /* Analyze and handle loop */
2771 /* Copied blocks do not have their phi list yet */
2772 collect_phiprojs(irg);
2774 /* Set links to NULL
2775 * TODO Still necessary? */
2776 irg_walk_graph(current_ir_graph, reset_link, NULL, NULL);
2782 ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
2783 ir_free_resources(irg, IR_RESOURCE_PHI_LIST);
2786 void do_loop_unrolling(ir_graph *irg)
2788 loop_op = loop_op_unrolling;
2790 DB((dbg, LEVEL_1, " >>> unrolling (Startnode %N) <<<\n",
2791 get_irg_start(irg)));
2793 loop_optimization(irg);
2795 DB((dbg, LEVEL_1, " >>> unrolling done (Startnode %N) <<<\n",
2796 get_irg_start(irg)));
2799 void do_loop_inversion(ir_graph *irg)
2801 loop_op = loop_op_inversion;
2803 DB((dbg, LEVEL_1, " >>> inversion (Startnode %N) <<<\n",
2804 get_irg_start(irg)));
2806 loop_optimization(irg);
2808 assure_cf_loop(irg);
2810 DB((dbg, LEVEL_1, " >>> inversion done (Startnode %N) <<<\n",
2811 get_irg_start(irg)));
2814 void do_loop_peeling(ir_graph *irg)
2816 loop_op = loop_op_peeling;
2818 DB((dbg, LEVEL_1, " >>> peeling (Startnode %N) <<<\n",
2819 get_irg_start(irg)));
2821 loop_optimization(irg);
2823 DB((dbg, LEVEL_1, " >>> peeling done (Startnode %N) <<<\n",
2824 get_irg_start(irg)));
2828 ir_graph_pass_t *loop_inversion_pass(const char *name)
2830 return def_graph_pass(name ? name : "loop_inversion", do_loop_inversion);
2833 ir_graph_pass_t *loop_unroll_pass(const char *name)
2835 return def_graph_pass(name ? name : "loop_unroll", do_loop_unrolling);
2838 ir_graph_pass_t *loop_peeling_pass(const char *name)
2840 return def_graph_pass(name ? name : "loop_peeling", do_loop_peeling);
2843 void firm_init_loop_opt(void)
2845 FIRM_DBG_REGISTER(dbg, "firm.opt.loop");