1 /* Basic block reordering routines for the GNU compiler.
2 Copyright (C) 2000, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2010
3 Free Software Foundation, Inc.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it
8 under the terms of the GNU General Public License as published by
9 the Free Software Foundation; either version 3, or (at your option)
12 GCC is distributed in the hope that it will be useful, but WITHOUT
13 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
14 or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public
15 License for more details.
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
21 /* This (greedy) algorithm constructs traces in several rounds.
22 The construction starts from "seeds". The seed for the first round
23 is the entry point of function. When there are more than one seed
24 that one is selected first that has the lowest key in the heap
25 (see function bb_to_key). Then the algorithm repeatedly adds the most
26 probable successor to the end of a trace. Finally it connects the traces.
28 There are two parameters: Branch Threshold and Exec Threshold.
29 If the edge to a successor of the actual basic block is lower than
30 Branch Threshold or the frequency of the successor is lower than
31 Exec Threshold the successor will be the seed in one of the next rounds.
32 Each round has these parameters lower than the previous one.
33 The last round has to have these parameters set to zero
34 so that the remaining blocks are picked up.
36 The algorithm selects the most probable successor from all unvisited
37 successors and successors that have been added to this trace.
38 The other successors (that has not been "sent" to the next round) will be
39 other seeds for this round and the secondary traces will start in them.
40 If the successor has not been visited in this trace it is added to the trace
41 (however, there is some heuristic for simple branches).
42 If the successor has been visited in this trace the loop has been found.
43 If the loop has many iterations the loop is rotated so that the
44 source block of the most probable edge going out from the loop
45 is the last block of the trace.
46 If the loop has few iterations and there is no edge from the last block of
47 the loop going out from loop the loop header is duplicated.
48 Finally, the construction of the trace is terminated.
50 When connecting traces it first checks whether there is an edge from the
51 last block of one trace to the first block of another trace.
52 When there are still some unconnected traces it checks whether there exists
53 a basic block BB such that BB is a successor of the last bb of one trace
54 and BB is a predecessor of the first block of another trace. In this case,
55 BB is duplicated and the traces are connected through this duplicate.
56 The rest of traces are simply connected so there will be a jump to the
57 beginning of the rest of trace.
62 "Software Trace Cache"
63 A. Ramirez, J. Larriba-Pey, C. Navarro, J. Torrellas and M. Valero; 1999
64 http://citeseer.nj.nec.com/15361.html
70 #include "coretypes.h"
77 #include "cfglayout.h"
85 #include "diagnostic-core.h"
86 #include "toplev.h" /* user_defined_section_attribute */
87 #include "tree-pass.h"
90 /* The number of rounds. In most cases there will only be 4 rounds, but
91 when partitioning hot and cold basic blocks into separate sections of
92 the .o file there will be an extra round.*/
95 /* Stubs in case we don't have a return insn.
96 We have to check at runtime too, not only compiletime. */
100 #define gen_return() NULL_RTX
104 /* Branch thresholds in thousandths (per mille) of the REG_BR_PROB_BASE. */
105 static int branch_threshold[N_ROUNDS] = {400, 200, 100, 0, 0};
107 /* Exec thresholds in thousandths (per mille) of the frequency of bb 0. */
108 static int exec_threshold[N_ROUNDS] = {500, 200, 50, 0, 0};
110 /* If edge frequency is lower than DUPLICATION_THRESHOLD per mille of entry
111 block the edge destination is not duplicated while connecting traces. */
112 #define DUPLICATION_THRESHOLD 100
114 /* Length of unconditional jump instruction. */
115 static int uncond_jump_length;
117 /* Structure to hold needed information for each basic block. */
118 typedef struct bbro_basic_block_data_def
120 /* Which trace is the bb start of (-1 means it is not a start of a trace). */
123 /* Which trace is the bb end of (-1 means it is not an end of a trace). */
126 /* Which trace is the bb in? */
129 /* Which heap is BB in (if any)? */
132 /* Which heap node is BB in (if any)? */
134 } bbro_basic_block_data;
136 /* The current size of the following dynamic array. */
137 static int array_size;
139 /* The array which holds needed information for basic blocks. */
140 static bbro_basic_block_data *bbd;
142 /* To avoid frequent reallocation the size of arrays is greater than needed,
143 the number of elements is (not less than) 1.25 * size_wanted. */
144 #define GET_ARRAY_SIZE(X) ((((X) / 4) + 1) * 5)
146 /* Free the memory and set the pointer to NULL. */
147 #define FREE(P) (gcc_assert (P), free (P), P = 0)
149 /* Structure for holding information about a trace. */
152 /* First and last basic block of the trace. */
153 basic_block first, last;
155 /* The round of the STC creation which this trace was found in. */
158 /* The length (i.e. the number of basic blocks) of the trace. */
162 /* Maximum frequency and count of one of the entry blocks. */
163 static int max_entry_frequency;
164 static gcov_type max_entry_count;
166 /* Local function prototypes. */
167 static void find_traces (int *, struct trace *);
168 static basic_block rotate_loop (edge, struct trace *, int);
169 static void mark_bb_visited (basic_block, int);
170 static void find_traces_1_round (int, int, gcov_type, struct trace *, int *,
171 int, fibheap_t *, int);
172 static basic_block copy_bb (basic_block, edge, basic_block, int);
173 static fibheapkey_t bb_to_key (basic_block);
174 static bool better_edge_p (const_basic_block, const_edge, int, int, int, int, const_edge);
175 static void connect_traces (int, struct trace *);
176 static bool copy_bb_p (const_basic_block, int);
177 static int get_uncond_jump_length (void);
178 static bool push_to_next_round_p (const_basic_block, int, int, int, gcov_type);
179 static void find_rarely_executed_basic_blocks_and_crossing_edges (edge **,
182 static void add_labels_and_missing_jumps (edge *, int);
183 static void add_reg_crossing_jump_notes (void);
184 static void fix_up_fall_thru_edges (void);
185 static void fix_edges_for_rarely_executed_code (edge *, int);
186 static void fix_crossing_conditional_branches (void);
187 static void fix_crossing_unconditional_branches (void);
189 /* Check to see if bb should be pushed into the next round of trace
190 collections or not. Reasons for pushing the block forward are 1).
191 If the block is cold, we are doing partitioning, and there will be
192 another round (cold partition blocks are not supposed to be
193 collected into traces until the very last round); or 2). There will
194 be another round, and the basic block is not "hot enough" for the
195 current round of trace collection. */
198 push_to_next_round_p (const_basic_block bb, int round, int number_of_rounds,
199 int exec_th, gcov_type count_th)
201 bool there_exists_another_round;
202 bool block_not_hot_enough;
204 there_exists_another_round = round < number_of_rounds - 1;
206 block_not_hot_enough = (bb->frequency < exec_th
207 || bb->count < count_th
208 || probably_never_executed_bb_p (bb));
210 if (there_exists_another_round
211 && block_not_hot_enough)
217 /* Find the traces for Software Trace Cache. Chain each trace through
218 RBI()->next. Store the number of traces to N_TRACES and description of
222 find_traces (int *n_traces, struct trace *traces)
225 int number_of_rounds;
230 /* Add one extra round of trace collection when partitioning hot/cold
231 basic blocks into separate sections. The last round is for all the
232 cold blocks (and ONLY the cold blocks). */
234 number_of_rounds = N_ROUNDS - 1;
236 /* Insert entry points of function into heap. */
237 heap = fibheap_new ();
238 max_entry_frequency = 0;
240 FOR_EACH_EDGE (e, ei, ENTRY_BLOCK_PTR->succs)
242 bbd[e->dest->index].heap = heap;
243 bbd[e->dest->index].node = fibheap_insert (heap, bb_to_key (e->dest),
245 if (e->dest->frequency > max_entry_frequency)
246 max_entry_frequency = e->dest->frequency;
247 if (e->dest->count > max_entry_count)
248 max_entry_count = e->dest->count;
251 /* Find the traces. */
252 for (i = 0; i < number_of_rounds; i++)
254 gcov_type count_threshold;
257 fprintf (dump_file, "STC - round %d\n", i + 1);
259 if (max_entry_count < INT_MAX / 1000)
260 count_threshold = max_entry_count * exec_threshold[i] / 1000;
262 count_threshold = max_entry_count / 1000 * exec_threshold[i];
264 find_traces_1_round (REG_BR_PROB_BASE * branch_threshold[i] / 1000,
265 max_entry_frequency * exec_threshold[i] / 1000,
266 count_threshold, traces, n_traces, i, &heap,
269 fibheap_delete (heap);
273 for (i = 0; i < *n_traces; i++)
276 fprintf (dump_file, "Trace %d (round %d): ", i + 1,
277 traces[i].round + 1);
278 for (bb = traces[i].first; bb != traces[i].last; bb = (basic_block) bb->aux)
279 fprintf (dump_file, "%d [%d] ", bb->index, bb->frequency);
280 fprintf (dump_file, "%d [%d]\n", bb->index, bb->frequency);
286 /* Rotate loop whose back edge is BACK_EDGE in the tail of trace TRACE
287 (with sequential number TRACE_N). */
290 rotate_loop (edge back_edge, struct trace *trace, int trace_n)
294 /* Information about the best end (end after rotation) of the loop. */
295 basic_block best_bb = NULL;
296 edge best_edge = NULL;
298 gcov_type best_count = -1;
299 /* The best edge is preferred when its destination is not visited yet
300 or is a start block of some trace. */
301 bool is_preferred = false;
303 /* Find the most frequent edge that goes out from current trace. */
304 bb = back_edge->dest;
310 FOR_EACH_EDGE (e, ei, bb->succs)
311 if (e->dest != EXIT_BLOCK_PTR
312 && e->dest->il.rtl->visited != trace_n
313 && (e->flags & EDGE_CAN_FALLTHRU)
314 && !(e->flags & EDGE_COMPLEX))
318 /* The best edge is preferred. */
319 if (!e->dest->il.rtl->visited
320 || bbd[e->dest->index].start_of_trace >= 0)
322 /* The current edge E is also preferred. */
323 int freq = EDGE_FREQUENCY (e);
324 if (freq > best_freq || e->count > best_count)
327 best_count = e->count;
335 if (!e->dest->il.rtl->visited
336 || bbd[e->dest->index].start_of_trace >= 0)
338 /* The current edge E is preferred. */
340 best_freq = EDGE_FREQUENCY (e);
341 best_count = e->count;
347 int freq = EDGE_FREQUENCY (e);
348 if (!best_edge || freq > best_freq || e->count > best_count)
351 best_count = e->count;
358 bb = (basic_block) bb->aux;
360 while (bb != back_edge->dest);
364 /* Rotate the loop so that the BEST_EDGE goes out from the last block of
366 if (back_edge->dest == trace->first)
368 trace->first = (basic_block) best_bb->aux;
374 for (prev_bb = trace->first;
375 prev_bb->aux != back_edge->dest;
376 prev_bb = (basic_block) prev_bb->aux)
378 prev_bb->aux = best_bb->aux;
380 /* Try to get rid of uncond jump to cond jump. */
381 if (single_succ_p (prev_bb))
383 basic_block header = single_succ (prev_bb);
385 /* Duplicate HEADER if it is a small block containing cond jump
387 if (any_condjump_p (BB_END (header)) && copy_bb_p (header, 0)
388 && !find_reg_note (BB_END (header), REG_CROSSING_JUMP,
390 copy_bb (header, single_succ_edge (prev_bb), prev_bb, trace_n);
396 /* We have not found suitable loop tail so do no rotation. */
397 best_bb = back_edge->src;
403 /* This function marks BB that it was visited in trace number TRACE. */
406 mark_bb_visited (basic_block bb, int trace)
408 bb->il.rtl->visited = trace;
409 if (bbd[bb->index].heap)
411 fibheap_delete_node (bbd[bb->index].heap, bbd[bb->index].node);
412 bbd[bb->index].heap = NULL;
413 bbd[bb->index].node = NULL;
417 /* One round of finding traces. Find traces for BRANCH_TH and EXEC_TH i.e. do
418 not include basic blocks their probability is lower than BRANCH_TH or their
419 frequency is lower than EXEC_TH into traces (or count is lower than
420 COUNT_TH). It stores the new traces into TRACES and modifies the number of
421 traces *N_TRACES. Sets the round (which the trace belongs to) to ROUND. It
422 expects that starting basic blocks are in *HEAP and at the end it deletes
423 *HEAP and stores starting points for the next round into new *HEAP. */
426 find_traces_1_round (int branch_th, int exec_th, gcov_type count_th,
427 struct trace *traces, int *n_traces, int round,
428 fibheap_t *heap, int number_of_rounds)
430 /* Heap for discarded basic blocks which are possible starting points for
432 fibheap_t new_heap = fibheap_new ();
434 while (!fibheap_empty (*heap))
442 bb = (basic_block) fibheap_extract_min (*heap);
443 bbd[bb->index].heap = NULL;
444 bbd[bb->index].node = NULL;
447 fprintf (dump_file, "Getting bb %d\n", bb->index);
449 /* If the BB's frequency is too low send BB to the next round. When
450 partitioning hot/cold blocks into separate sections, make sure all
451 the cold blocks (and ONLY the cold blocks) go into the (extra) final
454 if (push_to_next_round_p (bb, round, number_of_rounds, exec_th,
457 int key = bb_to_key (bb);
458 bbd[bb->index].heap = new_heap;
459 bbd[bb->index].node = fibheap_insert (new_heap, key, bb);
463 " Possible start point of next round: %d (key: %d)\n",
468 trace = traces + *n_traces;
470 trace->round = round;
472 bbd[bb->index].in_trace = *n_traces;
480 /* The probability and frequency of the best edge. */
481 int best_prob = INT_MIN / 2;
482 int best_freq = INT_MIN / 2;
485 mark_bb_visited (bb, *n_traces);
489 fprintf (dump_file, "Basic block %d was visited in trace %d\n",
490 bb->index, *n_traces - 1);
492 ends_in_call = block_ends_with_call_p (bb);
494 /* Select the successor that will be placed after BB. */
495 FOR_EACH_EDGE (e, ei, bb->succs)
497 gcc_assert (!(e->flags & EDGE_FAKE));
499 if (e->dest == EXIT_BLOCK_PTR)
502 if (e->dest->il.rtl->visited
503 && e->dest->il.rtl->visited != *n_traces)
506 if (BB_PARTITION (e->dest) != BB_PARTITION (bb))
509 prob = e->probability;
510 freq = e->dest->frequency;
512 /* The only sensible preference for a call instruction is the
513 fallthru edge. Don't bother selecting anything else. */
516 if (e->flags & EDGE_CAN_FALLTHRU)
525 /* Edge that cannot be fallthru or improbable or infrequent
526 successor (i.e. it is unsuitable successor). */
527 if (!(e->flags & EDGE_CAN_FALLTHRU) || (e->flags & EDGE_COMPLEX)
528 || prob < branch_th || EDGE_FREQUENCY (e) < exec_th
529 || e->count < count_th)
532 /* If partitioning hot/cold basic blocks, don't consider edges
533 that cross section boundaries. */
535 if (better_edge_p (bb, e, prob, freq, best_prob, best_freq,
544 /* If the best destination has multiple predecessors, and can be
545 duplicated cheaper than a jump, don't allow it to be added
546 to a trace. We'll duplicate it when connecting traces. */
547 if (best_edge && EDGE_COUNT (best_edge->dest->preds) >= 2
548 && copy_bb_p (best_edge->dest, 0))
551 /* Add all non-selected successors to the heaps. */
552 FOR_EACH_EDGE (e, ei, bb->succs)
555 || e->dest == EXIT_BLOCK_PTR
556 || e->dest->il.rtl->visited)
559 key = bb_to_key (e->dest);
561 if (bbd[e->dest->index].heap)
563 /* E->DEST is already in some heap. */
564 if (key != bbd[e->dest->index].node->key)
569 "Changing key for bb %d from %ld to %ld.\n",
571 (long) bbd[e->dest->index].node->key,
574 fibheap_replace_key (bbd[e->dest->index].heap,
575 bbd[e->dest->index].node, key);
580 fibheap_t which_heap = *heap;
582 prob = e->probability;
583 freq = EDGE_FREQUENCY (e);
585 if (!(e->flags & EDGE_CAN_FALLTHRU)
586 || (e->flags & EDGE_COMPLEX)
587 || prob < branch_th || freq < exec_th
588 || e->count < count_th)
590 /* When partitioning hot/cold basic blocks, make sure
591 the cold blocks (and only the cold blocks) all get
592 pushed to the last round of trace collection. */
594 if (push_to_next_round_p (e->dest, round,
597 which_heap = new_heap;
600 bbd[e->dest->index].heap = which_heap;
601 bbd[e->dest->index].node = fibheap_insert (which_heap,
607 " Possible start of %s round: %d (key: %ld)\n",
608 (which_heap == new_heap) ? "next" : "this",
609 e->dest->index, (long) key);
615 if (best_edge) /* Suitable successor was found. */
617 if (best_edge->dest->il.rtl->visited == *n_traces)
619 /* We do nothing with one basic block loops. */
620 if (best_edge->dest != bb)
622 if (EDGE_FREQUENCY (best_edge)
623 > 4 * best_edge->dest->frequency / 5)
625 /* The loop has at least 4 iterations. If the loop
626 header is not the first block of the function
627 we can rotate the loop. */
629 if (best_edge->dest != ENTRY_BLOCK_PTR->next_bb)
634 "Rotating loop %d - %d\n",
635 best_edge->dest->index, bb->index);
637 bb->aux = best_edge->dest;
638 bbd[best_edge->dest->index].in_trace =
640 bb = rotate_loop (best_edge, trace, *n_traces);
645 /* The loop has less than 4 iterations. */
647 if (single_succ_p (bb)
648 && copy_bb_p (best_edge->dest,
649 optimize_edge_for_speed_p (best_edge)))
651 bb = copy_bb (best_edge->dest, best_edge, bb,
658 /* Terminate the trace. */
663 /* Check for a situation
672 EDGE_FREQUENCY (AB) + EDGE_FREQUENCY (BC)
673 >= EDGE_FREQUENCY (AC).
674 (i.e. 2 * B->frequency >= EDGE_FREQUENCY (AC) )
675 Best ordering is then A B C.
677 This situation is created for example by:
684 FOR_EACH_EDGE (e, ei, bb->succs)
686 && (e->flags & EDGE_CAN_FALLTHRU)
687 && !(e->flags & EDGE_COMPLEX)
688 && !e->dest->il.rtl->visited
689 && single_pred_p (e->dest)
690 && !(e->flags & EDGE_CROSSING)
691 && single_succ_p (e->dest)
692 && (single_succ_edge (e->dest)->flags
694 && !(single_succ_edge (e->dest)->flags & EDGE_COMPLEX)
695 && single_succ (e->dest) == best_edge->dest
696 && 2 * e->dest->frequency >= EDGE_FREQUENCY (best_edge))
700 fprintf (dump_file, "Selecting BB %d\n",
701 best_edge->dest->index);
705 bb->aux = best_edge->dest;
706 bbd[best_edge->dest->index].in_trace = (*n_traces) - 1;
707 bb = best_edge->dest;
713 bbd[trace->first->index].start_of_trace = *n_traces - 1;
714 bbd[trace->last->index].end_of_trace = *n_traces - 1;
716 /* The trace is terminated so we have to recount the keys in heap
717 (some block can have a lower key because now one of its predecessors
718 is an end of the trace). */
719 FOR_EACH_EDGE (e, ei, bb->succs)
721 if (e->dest == EXIT_BLOCK_PTR
722 || e->dest->il.rtl->visited)
725 if (bbd[e->dest->index].heap)
727 key = bb_to_key (e->dest);
728 if (key != bbd[e->dest->index].node->key)
733 "Changing key for bb %d from %ld to %ld.\n",
735 (long) bbd[e->dest->index].node->key, key);
737 fibheap_replace_key (bbd[e->dest->index].heap,
738 bbd[e->dest->index].node,
745 fibheap_delete (*heap);
747 /* "Return" the new heap. */
751 /* Create a duplicate of the basic block OLD_BB and redirect edge E to it, add
752 it to trace after BB, mark OLD_BB visited and update pass' data structures
753 (TRACE is a number of trace which OLD_BB is duplicated to). */
756 copy_bb (basic_block old_bb, edge e, basic_block bb, int trace)
760 new_bb = duplicate_block (old_bb, e, bb);
761 BB_COPY_PARTITION (new_bb, old_bb);
763 gcc_assert (e->dest == new_bb);
764 gcc_assert (!e->dest->il.rtl->visited);
768 "Duplicated bb %d (created bb %d)\n",
769 old_bb->index, new_bb->index);
770 new_bb->il.rtl->visited = trace;
771 new_bb->aux = bb->aux;
774 if (new_bb->index >= array_size || last_basic_block > array_size)
779 new_size = MAX (last_basic_block, new_bb->index + 1);
780 new_size = GET_ARRAY_SIZE (new_size);
781 bbd = XRESIZEVEC (bbro_basic_block_data, bbd, new_size);
782 for (i = array_size; i < new_size; i++)
784 bbd[i].start_of_trace = -1;
785 bbd[i].in_trace = -1;
786 bbd[i].end_of_trace = -1;
790 array_size = new_size;
795 "Growing the dynamic array to %d elements.\n",
800 bbd[new_bb->index].in_trace = trace;
805 /* Compute and return the key (for the heap) of the basic block BB. */
808 bb_to_key (basic_block bb)
814 /* Do not start in probably never executed blocks. */
816 if (BB_PARTITION (bb) == BB_COLD_PARTITION
817 || probably_never_executed_bb_p (bb))
820 /* Prefer blocks whose predecessor is an end of some trace
821 or whose predecessor edge is EDGE_DFS_BACK. */
822 FOR_EACH_EDGE (e, ei, bb->preds)
824 if ((e->src != ENTRY_BLOCK_PTR && bbd[e->src->index].end_of_trace >= 0)
825 || (e->flags & EDGE_DFS_BACK))
827 int edge_freq = EDGE_FREQUENCY (e);
829 if (edge_freq > priority)
830 priority = edge_freq;
835 /* The block with priority should have significantly lower key. */
836 return -(100 * BB_FREQ_MAX + 100 * priority + bb->frequency);
837 return -bb->frequency;
840 /* Return true when the edge E from basic block BB is better than the temporary
841 best edge (details are in function). The probability of edge E is PROB. The
842 frequency of the successor is FREQ. The current best probability is
843 BEST_PROB, the best frequency is BEST_FREQ.
844 The edge is considered to be equivalent when PROB does not differ much from
845 BEST_PROB; similarly for frequency. */
848 better_edge_p (const_basic_block bb, const_edge e, int prob, int freq, int best_prob,
849 int best_freq, const_edge cur_best_edge)
853 /* The BEST_* values do not have to be best, but can be a bit smaller than
855 int diff_prob = best_prob / 10;
856 int diff_freq = best_freq / 10;
858 if (prob > best_prob + diff_prob)
859 /* The edge has higher probability than the temporary best edge. */
860 is_better_edge = true;
861 else if (prob < best_prob - diff_prob)
862 /* The edge has lower probability than the temporary best edge. */
863 is_better_edge = false;
864 else if (freq < best_freq - diff_freq)
865 /* The edge and the temporary best edge have almost equivalent
866 probabilities. The higher frequency of a successor now means
867 that there is another edge going into that successor.
868 This successor has lower frequency so it is better. */
869 is_better_edge = true;
870 else if (freq > best_freq + diff_freq)
871 /* This successor has higher frequency so it is worse. */
872 is_better_edge = false;
873 else if (e->dest->prev_bb == bb)
874 /* The edges have equivalent probabilities and the successors
875 have equivalent frequencies. Select the previous successor. */
876 is_better_edge = true;
878 is_better_edge = false;
880 /* If we are doing hot/cold partitioning, make sure that we always favor
881 non-crossing edges over crossing edges. */
884 && flag_reorder_blocks_and_partition
886 && (cur_best_edge->flags & EDGE_CROSSING)
887 && !(e->flags & EDGE_CROSSING))
888 is_better_edge = true;
890 return is_better_edge;
893 /* Connect traces in array TRACES, N_TRACES is the count of traces. */
896 connect_traces (int n_traces, struct trace *traces)
903 int current_partition;
905 gcov_type count_threshold;
907 freq_threshold = max_entry_frequency * DUPLICATION_THRESHOLD / 1000;
908 if (max_entry_count < INT_MAX / 1000)
909 count_threshold = max_entry_count * DUPLICATION_THRESHOLD / 1000;
911 count_threshold = max_entry_count / 1000 * DUPLICATION_THRESHOLD;
913 connected = XCNEWVEC (bool, n_traces);
916 current_partition = BB_PARTITION (traces[0].first);
919 if (flag_reorder_blocks_and_partition)
920 for (i = 0; i < n_traces && !two_passes; i++)
921 if (BB_PARTITION (traces[0].first)
922 != BB_PARTITION (traces[i].first))
925 for (i = 0; i < n_traces || (two_passes && current_pass == 1) ; i++)
934 gcc_assert (two_passes && current_pass == 1);
938 if (current_partition == BB_HOT_PARTITION)
939 current_partition = BB_COLD_PARTITION;
941 current_partition = BB_HOT_PARTITION;
948 && BB_PARTITION (traces[t].first) != current_partition)
953 /* Find the predecessor traces. */
954 for (t2 = t; t2 > 0;)
959 FOR_EACH_EDGE (e, ei, traces[t2].first->preds)
961 int si = e->src->index;
963 if (e->src != ENTRY_BLOCK_PTR
964 && (e->flags & EDGE_CAN_FALLTHRU)
965 && !(e->flags & EDGE_COMPLEX)
966 && bbd[si].end_of_trace >= 0
967 && !connected[bbd[si].end_of_trace]
968 && (BB_PARTITION (e->src) == current_partition)
970 || e->probability > best->probability
971 || (e->probability == best->probability
972 && traces[bbd[si].end_of_trace].length > best_len)))
975 best_len = traces[bbd[si].end_of_trace].length;
980 best->src->aux = best->dest;
981 t2 = bbd[best->src->index].end_of_trace;
982 connected[t2] = true;
986 fprintf (dump_file, "Connection: %d %d\n",
987 best->src->index, best->dest->index);
995 traces[last_trace].last->aux = traces[t2].first;
998 /* Find the successor traces. */
1001 /* Find the continuation of the chain. */
1005 FOR_EACH_EDGE (e, ei, traces[t].last->succs)
1007 int di = e->dest->index;
1009 if (e->dest != EXIT_BLOCK_PTR
1010 && (e->flags & EDGE_CAN_FALLTHRU)
1011 && !(e->flags & EDGE_COMPLEX)
1012 && bbd[di].start_of_trace >= 0
1013 && !connected[bbd[di].start_of_trace]
1014 && (BB_PARTITION (e->dest) == current_partition)
1016 || e->probability > best->probability
1017 || (e->probability == best->probability
1018 && traces[bbd[di].start_of_trace].length > best_len)))
1021 best_len = traces[bbd[di].start_of_trace].length;
1029 fprintf (dump_file, "Connection: %d %d\n",
1030 best->src->index, best->dest->index);
1032 t = bbd[best->dest->index].start_of_trace;
1033 traces[last_trace].last->aux = traces[t].first;
1034 connected[t] = true;
1039 /* Try to connect the traces by duplication of 1 block. */
1041 basic_block next_bb = NULL;
1042 bool try_copy = false;
1044 FOR_EACH_EDGE (e, ei, traces[t].last->succs)
1045 if (e->dest != EXIT_BLOCK_PTR
1046 && (e->flags & EDGE_CAN_FALLTHRU)
1047 && !(e->flags & EDGE_COMPLEX)
1048 && (!best || e->probability > best->probability))
1054 /* If the destination is a start of a trace which is only
1055 one block long, then no need to search the successor
1056 blocks of the trace. Accept it. */
1057 if (bbd[e->dest->index].start_of_trace >= 0
1058 && traces[bbd[e->dest->index].start_of_trace].length
1066 FOR_EACH_EDGE (e2, ei, e->dest->succs)
1068 int di = e2->dest->index;
1070 if (e2->dest == EXIT_BLOCK_PTR
1071 || ((e2->flags & EDGE_CAN_FALLTHRU)
1072 && !(e2->flags & EDGE_COMPLEX)
1073 && bbd[di].start_of_trace >= 0
1074 && !connected[bbd[di].start_of_trace]
1075 && (BB_PARTITION (e2->dest) == current_partition)
1076 && (EDGE_FREQUENCY (e2) >= freq_threshold)
1077 && (e2->count >= count_threshold)
1079 || e2->probability > best2->probability
1080 || (e2->probability == best2->probability
1081 && traces[bbd[di].start_of_trace].length
1086 if (e2->dest != EXIT_BLOCK_PTR)
1087 best2_len = traces[bbd[di].start_of_trace].length;
1089 best2_len = INT_MAX;
1096 if (flag_reorder_blocks_and_partition)
1099 /* Copy tiny blocks always; copy larger blocks only when the
1100 edge is traversed frequently enough. */
1102 && copy_bb_p (best->dest,
1103 optimize_edge_for_speed_p (best)
1104 && EDGE_FREQUENCY (best) >= freq_threshold
1105 && best->count >= count_threshold))
1111 fprintf (dump_file, "Connection: %d %d ",
1112 traces[t].last->index, best->dest->index);
1114 fputc ('\n', dump_file);
1115 else if (next_bb == EXIT_BLOCK_PTR)
1116 fprintf (dump_file, "exit\n");
1118 fprintf (dump_file, "%d\n", next_bb->index);
1121 new_bb = copy_bb (best->dest, best, traces[t].last, t);
1122 traces[t].last = new_bb;
1123 if (next_bb && next_bb != EXIT_BLOCK_PTR)
1125 t = bbd[next_bb->index].start_of_trace;
1126 traces[last_trace].last->aux = traces[t].first;
1127 connected[t] = true;
1131 break; /* Stop finding the successor traces. */
1134 break; /* Stop finding the successor traces. */
1143 fprintf (dump_file, "Final order:\n");
1144 for (bb = traces[0].first; bb; bb = (basic_block) bb->aux)
1145 fprintf (dump_file, "%d ", bb->index);
1146 fprintf (dump_file, "\n");
1153 /* Return true when BB can and should be copied. CODE_MAY_GROW is true
1154 when code size is allowed to grow by duplication. */
1157 copy_bb_p (const_basic_block bb, int code_may_grow)
1160 int max_size = uncond_jump_length;
1165 if (EDGE_COUNT (bb->preds) < 2)
1167 if (!can_duplicate_block_p (bb))
1170 /* Avoid duplicating blocks which have many successors (PR/13430). */
1171 if (EDGE_COUNT (bb->succs) > 8)
1174 if (code_may_grow && optimize_bb_for_speed_p (bb))
1175 max_size *= PARAM_VALUE (PARAM_MAX_GROW_COPY_BB_INSNS);
1177 FOR_BB_INSNS (bb, insn)
1180 size += get_attr_min_length (insn);
1183 if (size <= max_size)
1189 "Block %d can't be copied because its size = %d.\n",
1196 /* Return the length of unconditional jump instruction. */
1199 get_uncond_jump_length (void)
1204 label = emit_label_before (gen_label_rtx (), get_insns ());
1205 jump = emit_jump_insn (gen_jump (label));
1207 length = get_attr_min_length (jump);
1210 delete_insn (label);
1214 /* Find the basic blocks that are rarely executed and need to be moved to
1215 a separate section of the .o file (to cut down on paging and improve
1219 find_rarely_executed_basic_blocks_and_crossing_edges (edge **crossing_edges,
1220 int *n_crossing_edges,
1228 /* Mark which partition (hot/cold) each basic block belongs in. */
1232 if (probably_never_executed_bb_p (bb))
1233 BB_SET_PARTITION (bb, BB_COLD_PARTITION);
1235 BB_SET_PARTITION (bb, BB_HOT_PARTITION);
1238 /* Mark every edge that crosses between sections. */
1242 FOR_EACH_EDGE (e, ei, bb->succs)
1244 if (e->src != ENTRY_BLOCK_PTR
1245 && e->dest != EXIT_BLOCK_PTR
1246 && BB_PARTITION (e->src) != BB_PARTITION (e->dest))
1248 e->flags |= EDGE_CROSSING;
1252 *crossing_edges = XRESIZEVEC (edge, *crossing_edges, *max_idx);
1254 (*crossing_edges)[i++] = e;
1257 e->flags &= ~EDGE_CROSSING;
1259 *n_crossing_edges = i;
1262 /* If any destination of a crossing edge does not have a label, add label;
1263 Convert any fall-through crossing edges (for blocks that do not contain
1264 a jump) to unconditional jumps. */
1267 add_labels_and_missing_jumps (edge *crossing_edges, int n_crossing_edges)
1276 for (i=0; i < n_crossing_edges; i++)
1278 if (crossing_edges[i])
1280 src = crossing_edges[i]->src;
1281 dest = crossing_edges[i]->dest;
1283 /* Make sure dest has a label. */
1285 if (dest && (dest != EXIT_BLOCK_PTR))
1287 label = block_label (dest);
1289 /* Make sure source block ends with a jump. If the
1290 source block does not end with a jump it might end
1291 with a call_insn; this case will be handled in
1292 fix_up_fall_thru_edges function. */
1294 if (src && (src != ENTRY_BLOCK_PTR))
1296 if (!JUMP_P (BB_END (src)) && !block_ends_with_call_p (src))
1297 /* bb just falls through. */
1299 /* make sure there's only one successor */
1300 gcc_assert (single_succ_p (src));
1302 /* Find label in dest block. */
1303 label = block_label (dest);
1305 new_jump = emit_jump_insn_after (gen_jump (label),
1307 barrier = emit_barrier_after (new_jump);
1308 JUMP_LABEL (new_jump) = label;
1309 LABEL_NUSES (label) += 1;
1310 src->il.rtl->footer = unlink_insn_chain (barrier, barrier);
1311 /* Mark edge as non-fallthru. */
1312 crossing_edges[i]->flags &= ~EDGE_FALLTHRU;
1313 } /* end: 'if (GET_CODE ... ' */
1314 } /* end: 'if (src && src->index...' */
1315 } /* end: 'if (dest && dest->index...' */
1316 } /* end: 'if (crossing_edges[i]...' */
1317 } /* end for loop */
1320 /* Find any bb's where the fall-through edge is a crossing edge (note that
1321 these bb's must also contain a conditional jump or end with a call
1322 instruction; we've already dealt with fall-through edges for blocks
1323 that didn't have a conditional jump or didn't end with call instruction
1324 in the call to add_labels_and_missing_jumps). Convert the fall-through
1325 edge to non-crossing edge by inserting a new bb to fall-through into.
1326 The new bb will contain an unconditional jump (crossing edge) to the
1327 original fall through destination. */
1330 fix_up_fall_thru_edges (void)
1337 edge cond_jump = NULL;
1339 bool cond_jump_crosses;
1342 rtx fall_thru_label;
1345 FOR_EACH_BB (cur_bb)
1348 if (EDGE_COUNT (cur_bb->succs) > 0)
1349 succ1 = EDGE_SUCC (cur_bb, 0);
1353 if (EDGE_COUNT (cur_bb->succs) > 1)
1354 succ2 = EDGE_SUCC (cur_bb, 1);
1358 /* Find the fall-through edge. */
1361 && (succ1->flags & EDGE_FALLTHRU))
1367 && (succ2->flags & EDGE_FALLTHRU))
1372 else if (!fall_thru && succ1 && block_ends_with_call_p (cur_bb))
1377 /* Find EDGE_CAN_FALLTHRU edge. */
1378 FOR_EACH_EDGE (e, ei, cur_bb->succs)
1379 if (e->flags & EDGE_CAN_FALLTHRU)
1386 if (fall_thru && (fall_thru->dest != EXIT_BLOCK_PTR))
1388 /* Check to see if the fall-thru edge is a crossing edge. */
1390 if (fall_thru->flags & EDGE_CROSSING)
1392 /* The fall_thru edge crosses; now check the cond jump edge, if
1395 cond_jump_crosses = true;
1397 old_jump = BB_END (cur_bb);
1399 /* Find the jump instruction, if there is one. */
1403 if (!(cond_jump->flags & EDGE_CROSSING))
1404 cond_jump_crosses = false;
1406 /* We know the fall-thru edge crosses; if the cond
1407 jump edge does NOT cross, and its destination is the
1408 next block in the bb order, invert the jump
1409 (i.e. fix it so the fall thru does not cross and
1410 the cond jump does). */
1412 if (!cond_jump_crosses
1413 && cur_bb->aux == cond_jump->dest)
1415 /* Find label in fall_thru block. We've already added
1416 any missing labels, so there must be one. */
1418 fall_thru_label = block_label (fall_thru->dest);
1420 if (old_jump && JUMP_P (old_jump) && fall_thru_label)
1421 invert_worked = invert_jump (old_jump,
1425 fall_thru->flags &= ~EDGE_FALLTHRU;
1426 cond_jump->flags |= EDGE_FALLTHRU;
1427 update_br_prob_note (cur_bb);
1429 fall_thru = cond_jump;
1431 cond_jump->flags |= EDGE_CROSSING;
1432 fall_thru->flags &= ~EDGE_CROSSING;
1437 if (cond_jump_crosses || !invert_worked)
1439 /* This is the case where both edges out of the basic
1440 block are crossing edges. Here we will fix up the
1441 fall through edge. The jump edge will be taken care
1442 of later. The EDGE_CROSSING flag of fall_thru edge
1443 is unset before the call to force_nonfallthru
1444 function because if a new basic-block is created
1445 this edge remains in the current section boundary
1446 while the edge between new_bb and the fall_thru->dest
1447 becomes EDGE_CROSSING. */
1449 fall_thru->flags &= ~EDGE_CROSSING;
1450 new_bb = force_nonfallthru (fall_thru);
1454 new_bb->aux = cur_bb->aux;
1455 cur_bb->aux = new_bb;
1457 /* Make sure new fall-through bb is in same
1458 partition as bb it's falling through from. */
1460 BB_COPY_PARTITION (new_bb, cur_bb);
1461 single_succ_edge (new_bb)->flags |= EDGE_CROSSING;
1465 /* If a new basic-block was not created; restore
1466 the EDGE_CROSSING flag. */
1467 fall_thru->flags |= EDGE_CROSSING;
1470 /* Add barrier after new jump */
1474 barrier = emit_barrier_after (BB_END (new_bb));
1475 new_bb->il.rtl->footer = unlink_insn_chain (barrier,
1480 barrier = emit_barrier_after (BB_END (cur_bb));
1481 cur_bb->il.rtl->footer = unlink_insn_chain (barrier,
1490 /* This function checks the destination block of a "crossing jump" to
1491 see if it has any crossing predecessors that begin with a code label
1492 and end with an unconditional jump. If so, it returns that predecessor
1493 block. (This is to avoid creating lots of new basic blocks that all
1494 contain unconditional jumps to the same destination). */
1497 find_jump_block (basic_block jump_dest)
1499 basic_block source_bb = NULL;
1504 FOR_EACH_EDGE (e, ei, jump_dest->preds)
1505 if (e->flags & EDGE_CROSSING)
1507 basic_block src = e->src;
1509 /* Check each predecessor to see if it has a label, and contains
1510 only one executable instruction, which is an unconditional jump.
1511 If so, we can use it. */
1513 if (LABEL_P (BB_HEAD (src)))
1514 for (insn = BB_HEAD (src);
1515 !INSN_P (insn) && insn != NEXT_INSN (BB_END (src));
1516 insn = NEXT_INSN (insn))
1519 && insn == BB_END (src)
1521 && !any_condjump_p (insn))
1535 /* Find all BB's with conditional jumps that are crossing edges;
1536 insert a new bb and make the conditional jump branch to the new
1537 bb instead (make the new bb same color so conditional branch won't
1538 be a 'crossing' edge). Insert an unconditional jump from the
1539 new bb to the original destination of the conditional jump. */
1542 fix_crossing_conditional_branches (void)
1546 basic_block last_bb;
1554 rtx old_label = NULL_RTX;
1559 last_bb = EXIT_BLOCK_PTR->prev_bb;
1561 FOR_EACH_BB (cur_bb)
1563 crossing_edge = NULL;
1564 if (EDGE_COUNT (cur_bb->succs) > 0)
1565 succ1 = EDGE_SUCC (cur_bb, 0);
1569 if (EDGE_COUNT (cur_bb->succs) > 1)
1570 succ2 = EDGE_SUCC (cur_bb, 1);
1574 /* We already took care of fall-through edges, so only one successor
1575 can be a crossing edge. */
1577 if (succ1 && (succ1->flags & EDGE_CROSSING))
1578 crossing_edge = succ1;
1579 else if (succ2 && (succ2->flags & EDGE_CROSSING))
1580 crossing_edge = succ2;
1584 old_jump = BB_END (cur_bb);
1586 /* Check to make sure the jump instruction is a
1587 conditional jump. */
1591 if (any_condjump_p (old_jump))
1593 if (GET_CODE (PATTERN (old_jump)) == SET)
1594 set_src = SET_SRC (PATTERN (old_jump));
1595 else if (GET_CODE (PATTERN (old_jump)) == PARALLEL)
1597 set_src = XVECEXP (PATTERN (old_jump), 0,0);
1598 if (GET_CODE (set_src) == SET)
1599 set_src = SET_SRC (set_src);
1605 if (set_src && (GET_CODE (set_src) == IF_THEN_ELSE))
1607 if (GET_CODE (XEXP (set_src, 1)) == PC)
1608 old_label = XEXP (set_src, 2);
1609 else if (GET_CODE (XEXP (set_src, 2)) == PC)
1610 old_label = XEXP (set_src, 1);
1612 /* Check to see if new bb for jumping to that dest has
1613 already been created; if so, use it; if not, create
1616 new_bb = find_jump_block (crossing_edge->dest);
1619 new_label = block_label (new_bb);
1622 /* Create new basic block to be dest for
1623 conditional jump. */
1625 new_bb = create_basic_block (NULL, NULL, last_bb);
1626 new_bb->aux = last_bb->aux;
1627 last_bb->aux = new_bb;
1629 /* Put appropriate instructions in new bb. */
1631 new_label = gen_label_rtx ();
1632 emit_label_before (new_label, BB_HEAD (new_bb));
1633 BB_HEAD (new_bb) = new_label;
1635 if (GET_CODE (old_label) == LABEL_REF)
1637 old_label = JUMP_LABEL (old_jump);
1638 new_jump = emit_jump_insn_after (gen_jump
1644 gcc_assert (HAVE_return
1645 && GET_CODE (old_label) == RETURN);
1646 new_jump = emit_jump_insn_after (gen_return (),
1650 barrier = emit_barrier_after (new_jump);
1651 JUMP_LABEL (new_jump) = old_label;
1652 new_bb->il.rtl->footer = unlink_insn_chain (barrier,
1655 /* Make sure new bb is in same partition as source
1656 of conditional branch. */
1657 BB_COPY_PARTITION (new_bb, cur_bb);
1660 /* Make old jump branch to new bb. */
1662 redirect_jump (old_jump, new_label, 0);
1664 /* Remove crossing_edge as predecessor of 'dest'. */
1666 dest = crossing_edge->dest;
1668 redirect_edge_succ (crossing_edge, new_bb);
1670 /* Make a new edge from new_bb to old dest; new edge
1671 will be a successor for new_bb and a predecessor
1674 if (EDGE_COUNT (new_bb->succs) == 0)
1675 new_edge = make_edge (new_bb, dest, 0);
1677 new_edge = EDGE_SUCC (new_bb, 0);
1679 crossing_edge->flags &= ~EDGE_CROSSING;
1680 new_edge->flags |= EDGE_CROSSING;
1686 /* Find any unconditional branches that cross between hot and cold
1687 sections. Convert them into indirect jumps instead. */
1690 fix_crossing_unconditional_branches (void)
1696 rtx indirect_jump_sequence;
1697 rtx jump_insn = NULL_RTX;
1702 FOR_EACH_BB (cur_bb)
1704 last_insn = BB_END (cur_bb);
1706 if (EDGE_COUNT (cur_bb->succs) < 1)
1709 succ = EDGE_SUCC (cur_bb, 0);
1711 /* Check to see if bb ends in a crossing (unconditional) jump. At
1712 this point, no crossing jumps should be conditional. */
1714 if (JUMP_P (last_insn)
1715 && (succ->flags & EDGE_CROSSING))
1719 gcc_assert (!any_condjump_p (last_insn));
1721 /* Make sure the jump is not already an indirect or table jump. */
1723 if (!computed_jump_p (last_insn)
1724 && !tablejump_p (last_insn, &label2, &table))
1726 /* We have found a "crossing" unconditional branch. Now
1727 we must convert it to an indirect jump. First create
1728 reference of label, as target for jump. */
1730 label = JUMP_LABEL (last_insn);
1731 label_addr = gen_rtx_LABEL_REF (Pmode, label);
1732 LABEL_NUSES (label) += 1;
1734 /* Get a register to use for the indirect jump. */
1736 new_reg = gen_reg_rtx (Pmode);
1738 /* Generate indirect the jump sequence. */
1741 emit_move_insn (new_reg, label_addr);
1742 emit_indirect_jump (new_reg);
1743 indirect_jump_sequence = get_insns ();
1746 /* Make sure every instruction in the new jump sequence has
1747 its basic block set to be cur_bb. */
1749 for (cur_insn = indirect_jump_sequence; cur_insn;
1750 cur_insn = NEXT_INSN (cur_insn))
1752 if (!BARRIER_P (cur_insn))
1753 BLOCK_FOR_INSN (cur_insn) = cur_bb;
1754 if (JUMP_P (cur_insn))
1755 jump_insn = cur_insn;
1758 /* Insert the new (indirect) jump sequence immediately before
1759 the unconditional jump, then delete the unconditional jump. */
1761 emit_insn_before (indirect_jump_sequence, last_insn);
1762 delete_insn (last_insn);
1764 /* Make BB_END for cur_bb be the jump instruction (NOT the
1765 barrier instruction at the end of the sequence...). */
1767 BB_END (cur_bb) = jump_insn;
1773 /* Add REG_CROSSING_JUMP note to all crossing jump insns. */
1776 add_reg_crossing_jump_notes (void)
1783 FOR_EACH_EDGE (e, ei, bb->succs)
1784 if ((e->flags & EDGE_CROSSING)
1785 && JUMP_P (BB_END (e->src)))
1786 add_reg_note (BB_END (e->src), REG_CROSSING_JUMP, NULL_RTX);
1789 /* Hot and cold basic blocks are partitioned and put in separate
1790 sections of the .o file, to reduce paging and improve cache
1791 performance (hopefully). This can result in bits of code from the
1792 same function being widely separated in the .o file. However this
1793 is not obvious to the current bb structure. Therefore we must take
1794 care to ensure that: 1). There are no fall_thru edges that cross
1795 between sections; 2). For those architectures which have "short"
1796 conditional branches, all conditional branches that attempt to
1797 cross between sections are converted to unconditional branches;
1798 and, 3). For those architectures which have "short" unconditional
1799 branches, all unconditional branches that attempt to cross between
1800 sections are converted to indirect jumps.
1802 The code for fixing up fall_thru edges that cross between hot and
1803 cold basic blocks does so by creating new basic blocks containing
1804 unconditional branches to the appropriate label in the "other"
1805 section. The new basic block is then put in the same (hot or cold)
1806 section as the original conditional branch, and the fall_thru edge
1807 is modified to fall into the new basic block instead. By adding
1808 this level of indirection we end up with only unconditional branches
1809 crossing between hot and cold sections.
1811 Conditional branches are dealt with by adding a level of indirection.
1812 A new basic block is added in the same (hot/cold) section as the
1813 conditional branch, and the conditional branch is retargeted to the
1814 new basic block. The new basic block contains an unconditional branch
1815 to the original target of the conditional branch (in the other section).
1817 Unconditional branches are dealt with by converting them into
1821 fix_edges_for_rarely_executed_code (edge *crossing_edges,
1822 int n_crossing_edges)
1824 /* Make sure the source of any crossing edge ends in a jump and the
1825 destination of any crossing edge has a label. */
1827 add_labels_and_missing_jumps (crossing_edges, n_crossing_edges);
1829 /* Convert all crossing fall_thru edges to non-crossing fall
1830 thrus to unconditional jumps (that jump to the original fall
1833 fix_up_fall_thru_edges ();
1835 /* If the architecture does not have conditional branches that can
1836 span all of memory, convert crossing conditional branches into
1837 crossing unconditional branches. */
1839 if (!HAS_LONG_COND_BRANCH)
1840 fix_crossing_conditional_branches ();
1842 /* If the architecture does not have unconditional branches that
1843 can span all of memory, convert crossing unconditional branches
1844 into indirect jumps. Since adding an indirect jump also adds
1845 a new register usage, update the register usage information as
1848 if (!HAS_LONG_UNCOND_BRANCH)
1849 fix_crossing_unconditional_branches ();
1851 add_reg_crossing_jump_notes ();
1854 /* Verify, in the basic block chain, that there is at most one switch
1855 between hot/cold partitions. This is modelled on
1856 rtl_verify_flow_info_1, but it cannot go inside that function
1857 because this condition will not be true until after
1858 reorder_basic_blocks is called. */
1861 verify_hot_cold_block_grouping (void)
1865 bool switched_sections = false;
1866 int current_partition = 0;
1870 if (!current_partition)
1871 current_partition = BB_PARTITION (bb);
1872 if (BB_PARTITION (bb) != current_partition)
1874 if (switched_sections)
1876 error ("multiple hot/cold transitions found (bb %i)",
1882 switched_sections = true;
1883 current_partition = BB_PARTITION (bb);
1891 /* Reorder basic blocks. The main entry point to this file. FLAGS is
1892 the set of flags to pass to cfg_layout_initialize(). */
1895 reorder_basic_blocks (void)
1899 struct trace *traces;
1901 gcc_assert (current_ir_type () == IR_RTL_CFGLAYOUT);
1903 if (n_basic_blocks <= NUM_FIXED_BLOCKS + 1)
1906 set_edge_can_fallthru_flag ();
1907 mark_dfs_back_edges ();
1909 /* We are estimating the length of uncond jump insn only once since the code
1910 for getting the insn length always returns the minimal length now. */
1911 if (uncond_jump_length == 0)
1912 uncond_jump_length = get_uncond_jump_length ();
1914 /* We need to know some information for each basic block. */
1915 array_size = GET_ARRAY_SIZE (last_basic_block);
1916 bbd = XNEWVEC (bbro_basic_block_data, array_size);
1917 for (i = 0; i < array_size; i++)
1919 bbd[i].start_of_trace = -1;
1920 bbd[i].in_trace = -1;
1921 bbd[i].end_of_trace = -1;
1926 traces = XNEWVEC (struct trace, n_basic_blocks);
1928 find_traces (&n_traces, traces);
1929 connect_traces (n_traces, traces);
1933 relink_block_chain (/*stay_in_cfglayout_mode=*/true);
1936 dump_flow_info (dump_file, dump_flags);
1938 if (flag_reorder_blocks_and_partition)
1939 verify_hot_cold_block_grouping ();
1942 /* Determine which partition the first basic block in the function
1943 belongs to, then find the first basic block in the current function
1944 that belongs to a different section, and insert a
1945 NOTE_INSN_SWITCH_TEXT_SECTIONS note immediately before it in the
1946 instruction stream. When writing out the assembly code,
1947 encountering this note will make the compiler switch between the
1948 hot and cold text sections. */
1951 insert_section_boundary_note (void)
1955 int first_partition = 0;
1957 if (flag_reorder_blocks_and_partition)
1960 if (!first_partition)
1961 first_partition = BB_PARTITION (bb);
1962 if (BB_PARTITION (bb) != first_partition)
1964 new_note = emit_note_before (NOTE_INSN_SWITCH_TEXT_SECTIONS,
1966 /* ??? This kind of note always lives between basic blocks,
1967 but add_insn_before will set BLOCK_FOR_INSN anyway. */
1968 BLOCK_FOR_INSN (new_note) = NULL;
1974 /* Duplicate the blocks containing computed gotos. This basically unfactors
1975 computed gotos that were factored early on in the compilation process to
1976 speed up edge based data flow. We used to not unfactoring them again,
1977 which can seriously pessimize code with many computed jumps in the source
1978 code, such as interpreters. See e.g. PR15242. */
1981 gate_duplicate_computed_gotos (void)
1983 if (targetm.cannot_modify_jumps_p ())
1985 return (optimize > 0
1986 && flag_expensive_optimizations
1987 && ! optimize_function_for_size_p (cfun));
1992 duplicate_computed_gotos (void)
1994 basic_block bb, new_bb;
1998 if (n_basic_blocks <= NUM_FIXED_BLOCKS + 1)
2001 cfg_layout_initialize (0);
2003 /* We are estimating the length of uncond jump insn only once
2004 since the code for getting the insn length always returns
2005 the minimal length now. */
2006 if (uncond_jump_length == 0)
2007 uncond_jump_length = get_uncond_jump_length ();
2009 max_size = uncond_jump_length * PARAM_VALUE (PARAM_MAX_GOTO_DUPLICATION_INSNS);
2010 candidates = BITMAP_ALLOC (NULL);
2012 /* Look for blocks that end in a computed jump, and see if such blocks
2013 are suitable for unfactoring. If a block is a candidate for unfactoring,
2014 mark it in the candidates. */
2020 int size, all_flags;
2022 /* Build the reorder chain for the original order of blocks. */
2023 if (bb->next_bb != EXIT_BLOCK_PTR)
2024 bb->aux = bb->next_bb;
2026 /* Obviously the block has to end in a computed jump. */
2027 if (!computed_jump_p (BB_END (bb)))
2030 /* Only consider blocks that can be duplicated. */
2031 if (find_reg_note (BB_END (bb), REG_CROSSING_JUMP, NULL_RTX)
2032 || !can_duplicate_block_p (bb))
2035 /* Make sure that the block is small enough. */
2037 FOR_BB_INSNS (bb, insn)
2040 size += get_attr_min_length (insn);
2041 if (size > max_size)
2044 if (size > max_size)
2047 /* Final check: there must not be any incoming abnormal edges. */
2049 FOR_EACH_EDGE (e, ei, bb->preds)
2050 all_flags |= e->flags;
2051 if (all_flags & EDGE_COMPLEX)
2054 bitmap_set_bit (candidates, bb->index);
2057 /* Nothing to do if there is no computed jump here. */
2058 if (bitmap_empty_p (candidates))
2061 /* Duplicate computed gotos. */
2064 if (bb->il.rtl->visited)
2067 bb->il.rtl->visited = 1;
2069 /* BB must have one outgoing edge. That edge must not lead to
2070 the exit block or the next block.
2071 The destination must have more than one predecessor. */
2072 if (!single_succ_p (bb)
2073 || single_succ (bb) == EXIT_BLOCK_PTR
2074 || single_succ (bb) == bb->next_bb
2075 || single_pred_p (single_succ (bb)))
2078 /* The successor block has to be a duplication candidate. */
2079 if (!bitmap_bit_p (candidates, single_succ (bb)->index))
2082 new_bb = duplicate_block (single_succ (bb), single_succ_edge (bb), bb);
2083 new_bb->aux = bb->aux;
2085 new_bb->il.rtl->visited = 1;
2089 cfg_layout_finalize ();
2091 BITMAP_FREE (candidates);
2095 struct rtl_opt_pass pass_duplicate_computed_gotos =
2099 "compgotos", /* name */
2100 gate_duplicate_computed_gotos, /* gate */
2101 duplicate_computed_gotos, /* execute */
2104 0, /* static_pass_number */
2105 TV_REORDER_BLOCKS, /* tv_id */
2106 0, /* properties_required */
2107 0, /* properties_provided */
2108 0, /* properties_destroyed */
2109 0, /* todo_flags_start */
2110 TODO_dump_func | TODO_verify_rtl_sharing,/* todo_flags_finish */
2115 /* This function is the main 'entrance' for the optimization that
2116 partitions hot and cold basic blocks into separate sections of the
2117 .o file (to improve performance and cache locality). Ideally it
2118 would be called after all optimizations that rearrange the CFG have
2119 been called. However part of this optimization may introduce new
2120 register usage, so it must be called before register allocation has
2121 occurred. This means that this optimization is actually called
2122 well before the optimization that reorders basic blocks (see
2125 This optimization checks the feedback information to determine
2126 which basic blocks are hot/cold, updates flags on the basic blocks
2127 to indicate which section they belong in. This information is
2128 later used for writing out sections in the .o file. Because hot
2129 and cold sections can be arbitrarily large (within the bounds of
2130 memory), far beyond the size of a single function, it is necessary
2131 to fix up all edges that cross section boundaries, to make sure the
2132 instructions used can actually span the required distance. The
2133 fixes are described below.
2135 Fall-through edges must be changed into jumps; it is not safe or
2136 legal to fall through across a section boundary. Whenever a
2137 fall-through edge crossing a section boundary is encountered, a new
2138 basic block is inserted (in the same section as the fall-through
2139 source), and the fall through edge is redirected to the new basic
2140 block. The new basic block contains an unconditional jump to the
2141 original fall-through target. (If the unconditional jump is
2142 insufficient to cross section boundaries, that is dealt with a
2143 little later, see below).
2145 In order to deal with architectures that have short conditional
2146 branches (which cannot span all of memory) we take any conditional
2147 jump that attempts to cross a section boundary and add a level of
2148 indirection: it becomes a conditional jump to a new basic block, in
2149 the same section. The new basic block contains an unconditional
2150 jump to the original target, in the other section.
2152 For those architectures whose unconditional branch is also
2153 incapable of reaching all of memory, those unconditional jumps are
2154 converted into indirect jumps, through a register.
2156 IMPORTANT NOTE: This optimization causes some messy interactions
2157 with the cfg cleanup optimizations; those optimizations want to
2158 merge blocks wherever possible, and to collapse indirect jump
2159 sequences (change "A jumps to B jumps to C" directly into "A jumps
2160 to C"). Those optimizations can undo the jump fixes that
2161 partitioning is required to make (see above), in order to ensure
2162 that jumps attempting to cross section boundaries are really able
2163 to cover whatever distance the jump requires (on many architectures
2164 conditional or unconditional jumps are not able to reach all of
2165 memory). Therefore tests have to be inserted into each such
2166 optimization to make sure that it does not undo stuff necessary to
2167 cross partition boundaries. This would be much less of a problem
2168 if we could perform this optimization later in the compilation, but
2169 unfortunately the fact that we may need to create indirect jumps
2170 (through registers) requires that this optimization be performed
2171 before register allocation. */
2174 partition_hot_cold_basic_blocks (void)
2176 edge *crossing_edges;
2177 int n_crossing_edges;
2178 int max_edges = 2 * last_basic_block;
2180 if (n_basic_blocks <= NUM_FIXED_BLOCKS + 1)
2183 crossing_edges = XCNEWVEC (edge, max_edges);
2185 find_rarely_executed_basic_blocks_and_crossing_edges (&crossing_edges,
2189 if (n_crossing_edges > 0)
2190 fix_edges_for_rarely_executed_code (crossing_edges, n_crossing_edges);
2192 free (crossing_edges);
2196 gate_handle_reorder_blocks (void)
2198 if (targetm.cannot_modify_jumps_p ())
2200 return (optimize > 0);
2204 /* Reorder basic blocks. */
2206 rest_of_handle_reorder_blocks (void)
2210 /* Last attempt to optimize CFG, as scheduling, peepholing and insn
2211 splitting possibly introduced more crossjumping opportunities. */
2212 cfg_layout_initialize (CLEANUP_EXPENSIVE);
2214 if ((flag_reorder_blocks || flag_reorder_blocks_and_partition)
2215 /* Don't reorder blocks when optimizing for size because extra jump insns may
2216 be created; also barrier may create extra padding.
2218 More correctly we should have a block reordering mode that tried to
2219 minimize the combined size of all the jumps. This would more or less
2220 automatically remove extra jumps, but would also try to use more short
2221 jumps instead of long jumps. */
2222 && optimize_function_for_speed_p (cfun))
2224 reorder_basic_blocks ();
2225 cleanup_cfg (CLEANUP_EXPENSIVE);
2229 if (bb->next_bb != EXIT_BLOCK_PTR)
2230 bb->aux = bb->next_bb;
2231 cfg_layout_finalize ();
2233 /* Add NOTE_INSN_SWITCH_TEXT_SECTIONS notes. */
2234 insert_section_boundary_note ();
2238 struct rtl_opt_pass pass_reorder_blocks =
2243 gate_handle_reorder_blocks, /* gate */
2244 rest_of_handle_reorder_blocks, /* execute */
2247 0, /* static_pass_number */
2248 TV_REORDER_BLOCKS, /* tv_id */
2249 0, /* properties_required */
2250 0, /* properties_provided */
2251 0, /* properties_destroyed */
2252 0, /* todo_flags_start */
2253 TODO_dump_func | TODO_verify_rtl_sharing,/* todo_flags_finish */
2258 gate_handle_partition_blocks (void)
2260 /* The optimization to partition hot/cold basic blocks into separate
2261 sections of the .o file does not work well with linkonce or with
2262 user defined section attributes. Don't call it if either case
2265 return (flag_reorder_blocks_and_partition
2266 && !DECL_ONE_ONLY (current_function_decl)
2267 && !user_defined_section_attribute);
2270 /* Partition hot and cold basic blocks. */
2272 rest_of_handle_partition_blocks (void)
2274 partition_hot_cold_basic_blocks ();
2278 struct rtl_opt_pass pass_partition_blocks =
2282 "bbpart", /* name */
2283 gate_handle_partition_blocks, /* gate */
2284 rest_of_handle_partition_blocks, /* execute */
2287 0, /* static_pass_number */
2288 TV_REORDER_BLOCKS, /* tv_id */
2289 PROP_cfglayout, /* properties_required */
2290 0, /* properties_provided */
2291 0, /* properties_destroyed */
2292 0, /* todo_flags_start */
2293 TODO_dump_func | TODO_verify_rtl_sharing/* todo_flags_finish */