1 /* Thread edges through blocks and update the control flow and SSA graphs.
2 Copyright (C) 2004, 2005 Free Software Foundation, Inc.
4 This file is part of GCC.
6 GCC is free software; you can redistribute it and/or modify
7 it under the terms of the GNU General Public License as published by
8 the Free Software Foundation; either version 2, or (at your option)
11 GCC is distributed in the hope that it will be useful,
12 but WITHOUT ANY WARRANTY; without even the implied warranty of
13 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 GNU General Public License for more details.
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING. If not, write to
18 the Free Software Foundation, 59 Temple Place - Suite 330,
19 Boston, MA 02111-1307, USA. */
23 #include "coretypes.h"
30 #include "basic-block.h"
34 #include "diagnostic.h"
35 #include "tree-flow.h"
36 #include "tree-dump.h"
37 #include "tree-pass.h"
40 /* Given a block B, update the CFG and SSA graph to reflect redirecting
41 one or more in-edges to B to instead reach the destination of an
42 out-edge from B while preserving any side effects in B.
44 i.e., given A->B and B->C, change A->B to be A->C yet still preserve the
45 side effects of executing B.
47 1. Make a copy of B (including its outgoing edges and statements). Call
48 the copy B'. Note B' has no incoming edges or PHIs at this time.
50 2. Remove the control statement at the end of B' and all outgoing edges
53 3. Add a new argument to each PHI in C with the same value as the existing
54 argument associated with edge B->C. Associate the new PHI arguments
57 4. For each PHI in B, find or create a PHI in B' with an identical
58 PHI_RESULT. Add an argument to the PHI in B' which has the same
59 value as the PHI in B associated with the edge A->B. Associate
60 the new argument in the PHI in B' with the edge A->B.
62 5. Change the edge A->B to A->B'.
64 5a. This automatically deletes any PHI arguments associated with the
67 5b. This automatically associates each new argument added in step 4
70 6. Repeat for other incoming edges into B.
72 7. Put the duplicated resources in B and all the B' blocks into SSA form.
74 Note that block duplication can be minimized by first collecting the
75 the set of unique destination blocks that the incoming edges should
76 be threaded to. Block duplication can be further minimized by using
77 B instead of creating B' for one destination if all edges into B are
78 going to be threaded to a successor of B.
80 We further reduce the number of edges and statements we create by
81 not copying all the outgoing edges and the control statement in
82 step #1. We instead create a template block without the outgoing
83 edges and duplicate the template. */
86 /* Steps #5 and #6 of the above algorithm are best implemented by walking
87 all the incoming edges which thread to the same destination edge at
88 the same time. That avoids lots of table lookups to get information
89 for the destination edge.
91 To realize that implementation we create a list of incoming edges
92 which thread to the same outgoing edge. Thus to implement steps
93 #5 and #6 we traverse our hash table of outgoing edge information.
94 For each entry we walk the list of incoming edges which thread to
95 the current outgoing edge. */
103 /* Main data structure recording information regarding B's duplicate
106 /* We need to efficiently record the unique thread destinations of this
107 block and specific information associated with those destinations. We
108 may have many incoming edges threaded to the same outgoing edge. This
109 can be naturally implemented with a hash table. */
111 struct redirection_data
113 /* A duplicate of B with the trailing control statement removed and which
114 targets a single successor of B. */
115 basic_block dup_block;
117 /* An outgoing edge from B. DUP_BLOCK will have OUTGOING_EDGE->dest as
118 its single successor. */
121 /* A list of incoming edges which we want to thread to
122 OUTGOING_EDGE->dest. */
123 struct el *incoming_edges;
125 /* Flag indicating whether or not we should create a duplicate block
126 for this thread destination. This is only true if we are threading
127 all incoming edges and thus are using BB itself as a duplicate block. */
128 bool do_not_duplicate;
131 /* Main data structure to hold information for duplicates of BB. */
132 static htab_t redirection_data;
134 bool rediscover_loops_after_threading;
136 /* Data structure of information to pass to hash table traversal routines. */
139 /* The current block we are working on. */
142 /* A template copy of BB with no outgoing edges or control statement that
143 we use for creating copies. */
144 basic_block template_block;
146 /* TRUE if we thread one or more jumps, FALSE otherwise. */
150 /* Remove the last statement in block BB if it is a control statement
151 Also remove all outgoing edges except the edge which reaches DEST_BB.
152 If DEST_BB is NULL, then remove all outgoing edges. */
155 remove_ctrl_stmt_and_useless_edges (basic_block bb, basic_block dest_bb)
157 block_stmt_iterator bsi;
163 /* If the duplicate ends with a control statement, then remove it.
165 Note that if we are duplicating the template block rather than the
166 original basic block, then the duplicate might not have any real
170 && (TREE_CODE (bsi_stmt (bsi)) == COND_EXPR
171 || TREE_CODE (bsi_stmt (bsi)) == GOTO_EXPR
172 || TREE_CODE (bsi_stmt (bsi)) == SWITCH_EXPR))
175 for (ei = ei_start (bb->succs); (e = ei_safe_edge (ei)); )
177 if (e->dest != dest_bb)
184 /* Create a duplicate of BB which only reaches the destination of the edge
185 stored in RD. Record the duplicate block in RD. */
188 create_block_for_threading (basic_block bb, struct redirection_data *rd)
190 /* We can use the generic block duplication code and simply remove
191 the stuff we do not need. */
192 rd->dup_block = duplicate_block (bb, NULL);
194 /* Zero out the profile, since the block is unreachable for now. */
195 rd->dup_block->frequency = 0;
196 rd->dup_block->count = 0;
198 /* The call to duplicate_block will copy everything, including the
199 useless COND_EXPR or SWITCH_EXPR at the end of BB. We just remove
200 the useless COND_EXPR or SWITCH_EXPR here rather than having a
201 specialized block copier. We also remove all outgoing edges
202 from the duplicate block. The appropriate edge will be created
204 remove_ctrl_stmt_and_useless_edges (rd->dup_block, NULL);
207 /* Hashing and equality routines for our hash table. */
209 redirection_data_hash (const void *p)
211 edge e = ((struct redirection_data *)p)->outgoing_edge;
212 return e->dest->index;
216 redirection_data_eq (const void *p1, const void *p2)
218 edge e1 = ((struct redirection_data *)p1)->outgoing_edge;
219 edge e2 = ((struct redirection_data *)p2)->outgoing_edge;
224 /* Given an outgoing edge E lookup and return its entry in our hash table.
226 If INSERT is true, then we insert the entry into the hash table if
227 it is not already present. INCOMING_EDGE is added to the list of incoming
228 edges associated with E in the hash table. */
230 static struct redirection_data *
231 lookup_redirection_data (edge e, edge incoming_edge, enum insert_option insert)
234 struct redirection_data *elt;
236 /* Build a hash table element so we can see if E is already
238 elt = xmalloc (sizeof (struct redirection_data));
239 elt->outgoing_edge = e;
240 elt->dup_block = NULL;
241 elt->do_not_duplicate = false;
242 elt->incoming_edges = NULL;
244 slot = htab_find_slot (redirection_data, elt, insert);
246 /* This will only happen if INSERT is false and the entry is not
247 in the hash table. */
254 /* This will only happen if E was not in the hash table and
259 elt->incoming_edges = xmalloc (sizeof (struct el));
260 elt->incoming_edges->e = incoming_edge;
261 elt->incoming_edges->next = NULL;
264 /* E was in the hash table. */
267 /* Free ELT as we do not need it anymore, we will extract the
268 relevant entry from the hash table itself. */
271 /* Get the entry stored in the hash table. */
272 elt = (struct redirection_data *) *slot;
274 /* If insertion was requested, then we need to add INCOMING_EDGE
275 to the list of incoming edges associated with E. */
278 struct el *el = xmalloc (sizeof (struct el));
279 el->next = elt->incoming_edges;
280 el->e = incoming_edge;
281 elt->incoming_edges = el;
288 /* Given a duplicate block and its single destination (both stored
289 in RD). Create an edge between the duplicate and its single
292 Add an additional argument to any PHI nodes at the single
296 create_edge_and_update_destination_phis (struct redirection_data *rd)
298 edge e = make_edge (rd->dup_block, rd->outgoing_edge->dest, EDGE_FALLTHRU);
301 e->probability = REG_BR_PROB_BASE;
302 e->count = rd->dup_block->count;
304 /* If there are any PHI nodes at the destination of the outgoing edge
305 from the duplicate block, then we will need to add a new argument
306 to them. The argument should have the same value as the argument
307 associated with the outgoing edge stored in RD. */
308 for (phi = phi_nodes (e->dest); phi; phi = PHI_CHAIN (phi))
310 int indx = rd->outgoing_edge->dest_idx;
311 add_phi_arg (phi, PHI_ARG_DEF (phi, indx), e);
315 /* Hash table traversal callback routine to create duplicate blocks. */
318 create_duplicates (void **slot, void *data)
320 struct redirection_data *rd = (struct redirection_data *) *slot;
321 struct local_info *local_info = (struct local_info *)data;
323 /* If this entry should not have a duplicate created, then there's
325 if (rd->do_not_duplicate)
328 /* Create a template block if we have not done so already. Otherwise
329 use the template to create a new block. */
330 if (local_info->template_block == NULL)
332 create_block_for_threading (local_info->bb, rd);
333 local_info->template_block = rd->dup_block;
335 /* We do not create any outgoing edges for the template. We will
336 take care of that in a later traversal. That way we do not
337 create edges that are going to just be deleted. */
341 create_block_for_threading (local_info->template_block, rd);
343 /* Go ahead and wire up outgoing edges and update PHIs for the duplicate
345 create_edge_and_update_destination_phis (rd);
348 /* Keep walking the hash table. */
352 /* We did not create any outgoing edges for the template block during
353 block creation. This hash table traversal callback creates the
354 outgoing edge for the template block. */
357 fixup_template_block (void **slot, void *data)
359 struct redirection_data *rd = (struct redirection_data *) *slot;
360 struct local_info *local_info = (struct local_info *)data;
362 /* If this is the template block, then create its outgoing edges
363 and halt the hash table traversal. */
364 if (rd->dup_block && rd->dup_block == local_info->template_block)
366 create_edge_and_update_destination_phis (rd);
373 /* Not all jump threading requests are useful. In particular some
374 jump threading requests can create irreducible regions which are
377 This routine will examine the BB's incoming edges for jump threading
378 requests which, if acted upon, would create irreducible regions. Any
379 such jump threading requests found will be pruned away. */
382 prune_undesirable_thread_requests (basic_block bb)
386 bool may_create_irreducible_region = false;
387 unsigned int num_outgoing_edges_into_loop = 0;
389 /* For the heuristics below, we need to know if BB has more than
390 one outgoing edge into a loop. */
391 FOR_EACH_EDGE (e, ei, bb->succs)
392 num_outgoing_edges_into_loop += ((e->flags & EDGE_LOOP_EXIT) == 0);
394 if (num_outgoing_edges_into_loop > 1)
396 edge backedge = NULL;
398 /* Consider the effect of threading the edge (0, 1) to 2 on the left
399 CFG to produce the right CFG:
413 Threading the (0, 1) edge to 2 effectively creates two loops
414 (2, 4, 1) and (4, 1, 3) which are neither disjoint nor nested.
417 However, we do need to be able to thread (0, 1) to 2 or 3
418 in the left CFG below (which creates the middle and right
419 CFGs with nested loops).
423 1<--+ 2<----+ 3<-+<-+
427 3---+ 1--+--+ 2-----+
430 A safe heuristic appears to be to only allow threading if BB
431 has a single incoming backedge from one of its direct successors. */
433 FOR_EACH_EDGE (e, ei, bb->preds)
435 if (e->flags & EDGE_DFS_BACK)
449 if (backedge && find_edge (bb, backedge->src))
452 may_create_irreducible_region = true;
458 /* If we thread across the loop entry block (BB) into the
459 loop and BB is still reached from outside the loop, then
460 we would create an irreducible CFG. Consider the effect
461 of threading the edge (1, 4) to 5 on the left CFG to produce
473 Threading the (1, 4) edge to 5 creates two entry points
474 into the loop (4, 5) (one from block 1, the other from
475 block 2). A classic irreducible region.
477 So look at all of BB's incoming edges which are not
478 backedges and which are not threaded to the loop exit.
479 If that subset of incoming edges do not all thread
480 to the same block, then threading any of them will create
481 an irreducible region. */
483 FOR_EACH_EDGE (e, ei, bb->preds)
487 /* We ignore back edges for now. This may need refinement
488 as threading a backedge creates an inner loop which
489 we would need to verify has a single entry point.
491 If all backedges thread to new locations, then this
492 block will no longer have incoming backedges and we
493 need not worry about creating irreducible regions
494 by threading through BB. I don't think this happens
495 enough in practice to worry about it. */
496 if (e->flags & EDGE_DFS_BACK)
499 /* If the incoming edge threads to the loop exit, then it
502 if (e2 && (e2->flags & EDGE_LOOP_EXIT))
505 /* E enters the loop header and is not threaded. We can
506 not allow any other incoming edges to thread into
507 the loop as that would create an irreducible region. */
510 may_create_irreducible_region = true;
514 /* We know that this incoming edge threads to a block inside
515 the loop. This edge must thread to the same target in
516 the loop as any previously seen threaded edges. Otherwise
517 we will create an irreducible region. */
522 may_create_irreducible_region = true;
528 /* If we might create an irreducible region, then cancel any of
529 the jump threading requests for incoming edges which are
530 not backedges and which do not thread to the exit block. */
531 if (may_create_irreducible_region)
533 FOR_EACH_EDGE (e, ei, bb->preds)
537 /* Ignore back edges. */
538 if (e->flags & EDGE_DFS_BACK)
543 /* If this incoming edge was not threaded, then there is
548 /* If this incoming edge threaded to the loop exit,
549 then it can be ignored as it is safe. */
550 if (e2->flags & EDGE_LOOP_EXIT)
555 /* This edge threaded into the loop and the jump thread
556 request must be cancelled. */
557 if (dump_file && (dump_flags & TDF_DETAILS))
558 fprintf (dump_file, " Not threading jump %d --> %d to %d\n",
559 e->src->index, e->dest->index, e2->dest->index);
566 /* Hash table traversal callback to redirect each incoming edge
567 associated with this hash table element to its new destination. */
570 redirect_edges (void **slot, void *data)
572 struct redirection_data *rd = (struct redirection_data *) *slot;
573 struct local_info *local_info = (struct local_info *)data;
574 struct el *next, *el;
576 /* Walk over all the incoming edges associated associated with this
578 for (el = rd->incoming_edges; el; el = next)
582 /* Go ahead and free this element from the list. Doing this now
583 avoids the need for another list walk when we destroy the hash
588 /* Go ahead and clear E->aux. It's not needed anymore and failure
589 to clear it will cause all kinds of unpleasant problems later. */
596 if (dump_file && (dump_flags & TDF_DETAILS))
597 fprintf (dump_file, " Threaded jump %d --> %d to %d\n",
598 e->src->index, e->dest->index, rd->dup_block->index);
600 /* Redirect the incoming edge to the appropriate duplicate
602 e2 = redirect_edge_and_branch (e, rd->dup_block);
603 flush_pending_stmts (e2);
605 if ((dump_file && (dump_flags & TDF_DETAILS))
606 && e->src != e2->src)
607 fprintf (dump_file, " basic block %d created\n", e2->src->index);
611 if (dump_file && (dump_flags & TDF_DETAILS))
612 fprintf (dump_file, " Threaded jump %d --> %d to %d\n",
613 e->src->index, e->dest->index, local_info->bb->index);
615 /* We are using BB as the duplicate. Remove the unnecessary
616 outgoing edges and statements from BB. */
617 remove_ctrl_stmt_and_useless_edges (local_info->bb,
618 rd->outgoing_edge->dest);
620 /* And fixup the flags on the single remaining edge. */
621 single_succ_edge (local_info->bb)->flags
622 &= ~(EDGE_TRUE_VALUE | EDGE_FALSE_VALUE | EDGE_ABNORMAL);
623 single_succ_edge (local_info->bb)->flags |= EDGE_FALLTHRU;
627 /* Indicate that we actually threaded one or more jumps. */
628 if (rd->incoming_edges)
629 local_info->jumps_threaded = true;
634 /* BB is a block which ends with a COND_EXPR or SWITCH_EXPR and when BB
635 is reached via one or more specific incoming edges, we know which
636 outgoing edge from BB will be traversed.
638 We want to redirect those incoming edges to the target of the
639 appropriate outgoing edge. Doing so avoids a conditional branch
640 and may expose new optimization opportunities. Note that we have
641 to update dominator tree and SSA graph after such changes.
643 The key to keeping the SSA graph update manageable is to duplicate
644 the side effects occurring in BB so that those side effects still
645 occur on the paths which bypass BB after redirecting edges.
647 We accomplish this by creating duplicates of BB and arranging for
648 the duplicates to unconditionally pass control to one specific
649 successor of BB. We then revector the incoming edges into BB to
650 the appropriate duplicate of BB.
652 BB and its duplicates will have assignments to the same set of
653 SSA_NAMEs. Right now, we just call into update_ssa to update the
654 SSA graph for those names.
656 We are also going to experiment with a true incremental update
657 scheme for the duplicated resources. One of the interesting
658 properties we can exploit here is that all the resources set
659 in BB will have the same IDFS, so we have one IDFS computation
660 per block with incoming threaded edges, which can lower the
661 cost of the true incremental update algorithm. */
664 thread_block (basic_block bb)
666 /* E is an incoming edge into BB that we may or may not want to
667 redirect to a duplicate of BB. */
670 struct local_info local_info;
672 /* FOUND_BACKEDGE indicates that we found an incoming backedge
673 into BB, in which case we may ignore certain jump threads
674 to avoid creating irreducible regions. */
675 bool found_backedge = false;
677 /* ALL indicates whether or not all incoming edges into BB should
678 be threaded to a duplicate of BB. */
681 /* To avoid scanning a linear array for the element we need we instead
682 use a hash table. For normal code there should be no noticeable
683 difference. However, if we have a block with a large number of
684 incoming and outgoing edges such linear searches can get expensive. */
685 redirection_data = htab_create (EDGE_COUNT (bb->succs),
686 redirection_data_hash,
690 FOR_EACH_EDGE (e, ei, bb->preds)
691 found_backedge |= ((e->flags & EDGE_DFS_BACK) != 0);
693 /* If BB has incoming backedges, then threading across BB might
694 introduce an irreducible region, which would be undesirable
695 as that inhibits various optimizations later. Prune away
696 any jump threading requests which we know will result in
697 an irreducible region. */
699 prune_undesirable_thread_requests (bb);
701 /* Record each unique threaded destination into a hash table for
702 efficient lookups. */
703 FOR_EACH_EDGE (e, ei, bb->preds)
713 /* If we thread to a loop exit edge, then we will need to
714 rediscover the loop exit edges. While it may seem that
715 the new edge is a loop exit edge, that is not the case.
716 Consider threading the edge (5,6) to E in the CFG on the
717 left which creates the CFG on the right:
732 After threading, the edge (0, 1) is the loop exit edge and
733 the nodes 0, 2, 6 are the only nodes in the loop. */
734 if (e2->flags & EDGE_LOOP_EXIT)
735 rediscover_loops_after_threading = true;
737 /* Insert the outgoing edge into the hash table if it is not
738 already in the hash table. */
739 lookup_redirection_data (e2, e, INSERT);
743 /* If we are going to thread all incoming edges to an outgoing edge, then
744 BB will become unreachable. Rather than just throwing it away, use
745 it for one of the duplicates. Mark the first incoming edge with the
746 DO_NOT_DUPLICATE attribute. */
749 edge e = EDGE_PRED (bb, 0)->aux;
750 lookup_redirection_data (e, NULL, NO_INSERT)->do_not_duplicate = true;
753 /* Now create duplicates of BB.
755 Note that for a block with a high outgoing degree we can waste
756 a lot of time and memory creating and destroying useless edges.
758 So we first duplicate BB and remove the control structure at the
759 tail of the duplicate as well as all outgoing edges from the
760 duplicate. We then use that duplicate block as a template for
761 the rest of the duplicates. */
762 local_info.template_block = NULL;
764 local_info.jumps_threaded = false;
765 htab_traverse (redirection_data, create_duplicates, &local_info);
767 /* The template does not have an outgoing edge. Create that outgoing
768 edge and update PHI nodes as the edge's target as necessary.
770 We do this after creating all the duplicates to avoid creating
771 unnecessary edges. */
772 htab_traverse (redirection_data, fixup_template_block, &local_info);
774 /* The hash table traversals above created the duplicate blocks (and the
775 statements within the duplicate blocks). This loop creates PHI nodes for
776 the duplicated blocks and redirects the incoming edges into BB to reach
777 the duplicates of BB. */
778 htab_traverse (redirection_data, redirect_edges, &local_info);
780 /* Done with this block. Clear REDIRECTION_DATA. */
781 htab_delete (redirection_data);
782 redirection_data = NULL;
784 /* Indicate to our caller whether or not any jumps were threaded. */
785 return local_info.jumps_threaded;
788 /* Walk through all blocks and thread incoming edges to the block's
789 destinations as requested. This is the only entry point into this
792 Blocks which have one or more incoming edges have INCOMING_EDGE_THREADED
793 set in the block's annotation.
795 Each edge that should be threaded has the new destination edge stored in
796 the original edge's AUX field.
798 This routine (or one of its callees) will clear INCOMING_EDGE_THREADED
799 in the block annotations and the AUX field in the edges.
801 It is the caller's responsibility to fix the dominance information
802 and rewrite duplicated SSA_NAMEs back into SSA form.
804 Returns true if one or more edges were threaded, false otherwise. */
807 thread_through_all_blocks (bitmap threaded_blocks)
813 rediscover_loops_after_threading = false;
815 EXECUTE_IF_SET_IN_BITMAP (threaded_blocks, 0, i, bi)
817 basic_block bb = BASIC_BLOCK (i);
819 if (EDGE_COUNT (bb->preds) > 0)
820 retval |= thread_block (bb);