1 /*
2 * Copyright (c) 1997, 2025, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "libadt/vectset.hpp"
26 #include "memory/allocation.inline.hpp"
27 #include "memory/resourceArea.hpp"
28 #include "compiler/compilerDirectives.hpp"
29 #include "opto/block.hpp"
30 #include "opto/cfgnode.hpp"
31 #include "opto/chaitin.hpp"
32 #include "opto/loopnode.hpp"
33 #include "opto/machnode.hpp"
34 #include "opto/matcher.hpp"
35 #include "opto/opcodes.hpp"
36 #include "opto/rootnode.hpp"
37 #include "utilities/copy.hpp"
38 #include "utilities/powerOfTwo.hpp"
39
40 void Block_Array::grow(uint i) {
41 assert(i >= Max(), "Should have been checked before, use maybe_grow?");
42 DEBUG_ONLY(_limit = i+1);
43 if( i < _size ) return;
44 if( !_size ) {
45 _size = 1;
46 _blocks = (Block**)_arena->Amalloc( _size * sizeof(Block*) );
47 _blocks[0] = nullptr;
48 }
49 uint old = _size;
50 _size = next_power_of_2(i);
51 _blocks = (Block**)_arena->Arealloc( _blocks, old*sizeof(Block*),_size*sizeof(Block*));
52 Copy::zero_to_bytes( &_blocks[old], (_size-old)*sizeof(Block*) );
53 }
54
55 void Block_List::remove(uint i) {
56 assert(i < _cnt, "index out of bounds");
57 Copy::conjoint_words_to_lower((HeapWord*)&_blocks[i+1], (HeapWord*)&_blocks[i], ((_cnt-i-1)*sizeof(Block*)));
58 pop(); // shrink list by one block
59 }
60
61 void Block_List::insert(uint i, Block *b) {
62 push(b); // grow list by one block
63 Copy::conjoint_words_to_higher((HeapWord*)&_blocks[i], (HeapWord*)&_blocks[i+1], ((_cnt-i-1)*sizeof(Block*)));
64 _blocks[i] = b;
65 }
66
67 #ifndef PRODUCT
68 void Block_List::print() {
69 for (uint i=0; i < size(); i++) {
70 tty->print("B%d ", _blocks[i]->_pre_order);
71 }
72 tty->print("size = %d\n", size());
73 }
74 #endif
75
76 uint Block::code_alignment() const {
77 // Check for Root block
78 if (_pre_order == 0) return CodeEntryAlignment;
79 // Check for Start block
80 if (_pre_order == 1) return InteriorEntryAlignment;
81 // Check for loop alignment
82 if (has_loop_alignment()) return loop_alignment();
83
84 return relocInfo::addr_unit(); // no particular alignment
85 }
86
87 uint Block::compute_loop_alignment() {
88 Node *h = head();
89 int unit_sz = relocInfo::addr_unit();
90 if (h->is_Loop() && h->as_Loop()->is_inner_loop()) {
91 // Pre- and post-loops have low trip count so do not bother with
92 // NOPs for align loop head. The constants are hidden from tuning
93 // but only because my "divide by 4" heuristic surely gets nearly
94 // all possible gain (a "do not align at all" heuristic has a
95 // chance of getting a really tiny gain).
96 if (h->is_CountedLoop() && (h->as_CountedLoop()->is_pre_loop() ||
97 h->as_CountedLoop()->is_post_loop())) {
98 return (OptoLoopAlignment > 4*unit_sz) ? (OptoLoopAlignment>>2) : unit_sz;
99 }
100 // Loops with low backedge frequency should not be aligned.
101 Node *n = h->in(LoopNode::LoopBackControl)->in(0);
102 if (n->is_MachIf() && n->as_MachIf()->_prob < 0.01) {
103 return unit_sz; // Loop does not loop, more often than not!
104 }
105 return OptoLoopAlignment; // Otherwise align loop head
106 }
107
108 return unit_sz; // no particular alignment
109 }
110
111 // Compute the size of first 'inst_cnt' instructions in this block.
112 // Return the number of instructions left to compute if the block has
113 // less then 'inst_cnt' instructions. Stop, and return 0 if sum_size
114 // exceeds OptoLoopAlignment.
115 uint Block::compute_first_inst_size(uint& sum_size, uint inst_cnt,
116 PhaseRegAlloc* ra) {
117 uint last_inst = number_of_nodes();
118 for( uint j = 0; j < last_inst && inst_cnt > 0; j++ ) {
119 uint inst_size = get_node(j)->size(ra);
120 if( inst_size > 0 ) {
121 inst_cnt--;
122 uint sz = sum_size + inst_size;
123 if( sz <= (uint)OptoLoopAlignment ) {
124 // Compute size of instructions which fit into fetch buffer only
125 // since all inst_cnt instructions will not fit even if we align them.
126 sum_size = sz;
127 } else {
128 return 0;
129 }
130 }
131 }
132 return inst_cnt;
133 }
134
135 uint Block::find_node( const Node *n ) const {
136 for( uint i = 0; i < number_of_nodes(); i++ ) {
137 if( get_node(i) == n )
138 return i;
139 }
140 ShouldNotReachHere();
141 return 0;
142 }
143
144 // Find and remove n from block list
145 void Block::find_remove( const Node *n ) {
146 remove_node(find_node(n));
147 }
148
149 bool Block::contains(const Node *n) const {
150 return _nodes.contains(n);
151 }
152
153 bool Block::is_trivially_unreachable() const {
154 return num_preds() <= 1 && !head()->is_Root() && !head()->is_Start();
155 }
156
157 // Return empty status of a block. Empty blocks contain only the head, other
158 // ideal nodes, and an optional trailing goto.
159 int Block::is_Empty() const {
160
161 // Root or start block is not considered empty
162 if (head()->is_Root() || head()->is_Start()) {
163 return not_empty;
164 }
165
166 int success_result = completely_empty;
167 int end_idx = number_of_nodes() - 1;
168
169 // Check for ending goto
170 if ((end_idx > 0) && (get_node(end_idx)->is_MachGoto())) {
171 success_result = empty_with_goto;
172 end_idx--;
173 }
174
175 // Unreachable blocks are considered empty
176 if (is_trivially_unreachable()) {
177 return success_result;
178 }
179
180 // Ideal nodes (except BoxLock) are allowable in empty blocks: skip them. Only
181 // Mach and BoxLock nodes turn directly into code via emit().
182 while ((end_idx > 0) &&
183 !get_node(end_idx)->is_Mach() &&
184 !get_node(end_idx)->is_BoxLock()) {
185 end_idx--;
186 }
187
188 // No room for any interesting instructions?
189 if (end_idx == 0) {
190 return success_result;
191 }
192
193 return not_empty;
194 }
195
196 // Return true if the block's code implies that it is likely to be
197 // executed infrequently. Check to see if the block ends in a Halt or
198 // a low probability call.
199 bool Block::has_uncommon_code() const {
200 Node* en = end();
201
202 if (en->is_MachGoto())
203 en = en->in(0);
204 if (en->is_Catch())
205 en = en->in(0);
206 if (en->is_MachProj() && en->in(0)->is_MachCall()) {
207 MachCallNode* call = en->in(0)->as_MachCall();
208 if (call->cnt() != COUNT_UNKNOWN && call->cnt() <= PROB_UNLIKELY_MAG(4)) {
209 // This is true for slow-path stubs like new_{instance,array},
210 // slow_arraycopy, complete_monitor_locking, uncommon_trap.
211 // The magic number corresponds to the probability of an uncommon_trap,
212 // even though it is a count not a probability.
213 return true;
214 }
215 }
216
217 int op = en->is_Mach() ? en->as_Mach()->ideal_Opcode() : en->Opcode();
218 return op == Op_Halt;
219 }
220
221 // True if block is low enough frequency or guarded by a test which
222 // mostly does not go here.
223 bool PhaseCFG::is_uncommon(const Block* block) {
224 // Initial blocks must never be moved, so are never uncommon.
225 if (block->head()->is_Root() || block->head()->is_Start()) return false;
226
227 // Check for way-low freq
228 if(block->_freq < BLOCK_FREQUENCY(0.00001f) ) return true;
229
230 // Look for code shape indicating uncommon_trap or slow path
231 if (block->has_uncommon_code()) return true;
232
233 const float epsilon = 0.05f;
234 const float guard_factor = PROB_UNLIKELY_MAG(4) / (1.f - epsilon);
235 uint uncommon_preds = 0;
236 uint freq_preds = 0;
237 uint uncommon_for_freq_preds = 0;
238
239 for( uint i=1; i< block->num_preds(); i++ ) {
240 Block* guard = get_block_for_node(block->pred(i));
241 // Check to see if this block follows its guard 1 time out of 10000
242 // or less.
243 //
244 // See list of magnitude-4 unlikely probabilities in cfgnode.hpp which
245 // we intend to be "uncommon", such as slow-path TLE allocation,
246 // predicted call failure, and uncommon trap triggers.
247 //
248 // Use an epsilon value of 5% to allow for variability in frequency
249 // predictions and floating point calculations. The net effect is
250 // that guard_factor is set to 9500.
251 //
252 // Ignore low-frequency blocks.
253 // The next check is (guard->_freq < 1.e-5 * 9500.).
254 if(guard->_freq*BLOCK_FREQUENCY(guard_factor) < BLOCK_FREQUENCY(0.00001f)) {
255 uncommon_preds++;
256 } else {
257 freq_preds++;
258 if(block->_freq < guard->_freq * guard_factor ) {
259 uncommon_for_freq_preds++;
260 }
261 }
262 }
263 if( block->num_preds() > 1 &&
264 // The block is uncommon if all preds are uncommon or
265 (uncommon_preds == (block->num_preds()-1) ||
266 // it is uncommon for all frequent preds.
267 uncommon_for_freq_preds == freq_preds) ) {
268 return true;
269 }
270 return false;
271 }
272
273 #ifndef PRODUCT
274 void Block::dump_bidx(const Block* orig, outputStream* st) const {
275 if (_pre_order) st->print("B%d", _pre_order);
276 else st->print("N%d", head()->_idx);
277
278 if (Verbose && orig != this) {
279 // Dump the original block's idx
280 st->print(" (");
281 orig->dump_bidx(orig, st);
282 st->print(")");
283 }
284 }
285
286 void Block::dump_pred(const PhaseCFG* cfg, Block* orig, outputStream* st) const {
287 if (is_connector()) {
288 for (uint i=1; i<num_preds(); i++) {
289 Block *p = cfg->get_block_for_node(pred(i));
290 p->dump_pred(cfg, orig, st);
291 }
292 } else {
293 dump_bidx(orig, st);
294 st->print(" ");
295 }
296 }
297
298 void Block::dump_head(const PhaseCFG* cfg, outputStream* st) const {
299 // Print the basic block.
300 dump_bidx(this, st);
301 st->print(": ");
302
303 // Print the outgoing CFG edges.
304 st->print("#\tout( ");
305 for( uint i=0; i<_num_succs; i++ ) {
306 non_connector_successor(i)->dump_bidx(_succs[i], st);
307 st->print(" ");
308 }
309
310 // Print the incoming CFG edges.
311 st->print(") <- ");
312 if( head()->is_block_start() ) {
313 st->print("in( ");
314 for (uint i=1; i<num_preds(); i++) {
315 Node *s = pred(i);
316 if (cfg != nullptr) {
317 Block *p = cfg->get_block_for_node(s);
318 p->dump_pred(cfg, p, st);
319 } else {
320 while (!s->is_block_start()) {
321 s = s->in(0);
322 }
323 st->print("N%d ", s->_idx );
324 }
325 }
326 st->print(") ");
327 } else {
328 st->print("BLOCK HEAD IS JUNK ");
329 }
330
331 // Print loop, if any
332 const Block *bhead = this; // Head of self-loop
333 Node *bh = bhead->head();
334
335 if ((cfg != nullptr) && bh->is_Loop() && !head()->is_Root()) {
336 LoopNode *loop = bh->as_Loop();
337 const Block *bx = cfg->get_block_for_node(loop->in(LoopNode::LoopBackControl));
338 while (bx->is_connector()) {
339 bx = cfg->get_block_for_node(bx->pred(1));
340 }
341 st->print("Loop( B%d-B%d ", bhead->_pre_order, bx->_pre_order);
342 // Dump any loop-specific bits, especially for CountedLoops.
343 loop->dump_spec(st);
344 st->print(")");
345 } else if (has_loop_alignment()) {
346 st->print("top-of-loop");
347 }
348
349 // Print frequency and other optimization-relevant information
350 st->print(" Freq: %g",_freq);
351 if( Verbose || WizardMode ) {
352 st->print(" IDom: %d/#%d", _idom ? _idom->_pre_order : 0, _dom_depth);
353 st->print(" RegPressure: %d",_reg_pressure);
354 st->print(" IHRP Index: %d",_ihrp_index);
355 st->print(" FRegPressure: %d",_freg_pressure);
356 st->print(" FHRP Index: %d",_fhrp_index);
357 }
358 st->cr();
359 }
360
361 void Block::dump() const {
362 dump(nullptr);
363 }
364
365 void Block::dump(const PhaseCFG* cfg) const {
366 dump_head(cfg);
367 for (uint i=0; i< number_of_nodes(); i++) {
368 get_node(i)->dump();
369 }
370 tty->print("\n");
371 }
372 #endif
373
374 PhaseCFG::PhaseCFG(Arena* arena, RootNode* root, Matcher& matcher)
375 : Phase(CFG)
376 , _root(root)
377 , _blocks(arena)
378 , _block_arena(arena)
379 , _regalloc(nullptr)
380 , _scheduling_for_pressure(false)
381 , _matcher(matcher)
382 , _node_to_block_mapping(arena)
383 , _node_latency(nullptr)
384 #ifndef PRODUCT
385 , _trace_opto_pipelining(C->directive()->TraceOptoPipeliningOption)
386 #endif
387 #ifdef ASSERT
388 , _raw_oops(arena)
389 #endif
390 {
391 ResourceMark rm;
392 // I'll need a few machine-specific GotoNodes. Make an Ideal GotoNode,
393 // then Match it into a machine-specific Node. Then clone the machine
394 // Node on demand.
395 Node *x = new GotoNode(nullptr);
396 x->init_req(0, x);
397 _goto = matcher.match_tree(x);
398 assert(_goto != nullptr || C->failure_is_artificial(), "");
399 if (C->failing()) {
400 return;
401 }
402 _goto->set_req(0,_goto);
403
404 // Build the CFG in Reverse Post Order
405 _number_of_blocks = build_cfg();
406 _root_block = get_block_for_node(_root);
407 }
408
409 // Build a proper looking CFG. Make every block begin with either a StartNode
410 // or a RegionNode. Make every block end with either a Goto, If or Return.
411 // The RootNode both starts and ends it's own block. Do this with a recursive
412 // backwards walk over the control edges.
413 uint PhaseCFG::build_cfg() {
414 VectorSet visited;
415
416 // Allocate stack with enough space to avoid frequent realloc
417 Node_Stack nstack(C->live_nodes() >> 1);
418 nstack.push(_root, 0);
419 uint sum = 0; // Counter for blocks
420
421 while (nstack.is_nonempty()) {
422 // node and in's index from stack's top
423 // 'np' is _root (see above) or RegionNode, StartNode: we push on stack
424 // only nodes which point to the start of basic block (see below).
425 Node *np = nstack.node();
426 // idx > 0, except for the first node (_root) pushed on stack
427 // at the beginning when idx == 0.
428 // We will use the condition (idx == 0) later to end the build.
429 uint idx = nstack.index();
430 Node *proj = np->in(idx);
431 const Node *x = proj->is_block_proj();
432 // Does the block end with a proper block-ending Node? One of Return,
433 // If or Goto? (This check should be done for visited nodes also).
434 if (x == nullptr) { // Does not end right...
435 Node *g = _goto->clone(); // Force it to end in a Goto
436 g->set_req(0, proj);
437 np->set_req(idx, g);
438 x = proj = g;
439 }
440 if (!visited.test_set(x->_idx)) { // Visit this block once
441 // Skip any control-pinned middle'in stuff
442 Node *p = proj;
443 do {
444 proj = p; // Update pointer to last Control
445 p = p->in(0); // Move control forward
446 } while( !p->is_block_proj() &&
447 !p->is_block_start() );
448 // Make the block begin with one of Region or StartNode.
449 if( !p->is_block_start() ) {
450 RegionNode *r = new RegionNode( 2 );
451 r->init_req(1, p); // Insert RegionNode in the way
452 proj->set_req(0, r); // Insert RegionNode in the way
453 p = r;
454 }
455 // 'p' now points to the start of this basic block
456
457 // Put self in array of basic blocks
458 Block *bb = new (_block_arena) Block(_block_arena, p);
459 map_node_to_block(p, bb);
460 map_node_to_block(x, bb);
461 if( x != p ) { // Only for root is x == p
462 bb->push_node((Node*)x);
463 }
464 // Now handle predecessors
465 ++sum; // Count 1 for self block
466 uint cnt = bb->num_preds();
467 for (int i = (cnt - 1); i > 0; i-- ) { // For all predecessors
468 Node *prevproj = p->in(i); // Get prior input
469 assert( !prevproj->is_Con(), "dead input not removed" );
470 // Check to see if p->in(i) is a "control-dependent" CFG edge -
471 // i.e., it splits at the source (via an IF or SWITCH) and merges
472 // at the destination (via a many-input Region).
473 // This breaks critical edges. The RegionNode to start the block
474 // will be added when <p,i> is pulled off the node stack
475 if ( cnt > 2 ) { // Merging many things?
476 assert( prevproj== bb->pred(i),"");
477 if(prevproj->is_block_proj() != prevproj) { // Control-dependent edge?
478 // Force a block on the control-dependent edge
479 Node *g = _goto->clone(); // Force it to end in a Goto
480 g->set_req(0,prevproj);
481 p->set_req(i,g);
482 }
483 }
484 nstack.push(p, i); // 'p' is RegionNode or StartNode
485 }
486 } else { // Post-processing visited nodes
487 nstack.pop(); // remove node from stack
488 // Check if it the fist node pushed on stack at the beginning.
489 if (idx == 0) break; // end of the build
490 // Find predecessor basic block
491 Block *pb = get_block_for_node(x);
492 // Insert into nodes array, if not already there
493 if (!has_block(proj)) {
494 assert( x != proj, "" );
495 // Map basic block of projection
496 map_node_to_block(proj, pb);
497 pb->push_node(proj);
498 }
499 // Insert self as a child of my predecessor block
500 pb->_succs.map(pb->_num_succs++, get_block_for_node(np));
501 assert( pb->get_node(pb->number_of_nodes() - pb->_num_succs)->is_block_proj(),
502 "too many control users, not a CFG?" );
503 }
504 }
505 // Return number of basic blocks for all children and self
506 return sum;
507 }
508
509 // Inserts a goto & corresponding basic block between
510 // block[block_no] and its succ_no'th successor block
511 void PhaseCFG::insert_goto_at(uint block_no, uint succ_no) {
512 // get block with block_no
513 assert(block_no < number_of_blocks(), "illegal block number");
514 Block* in = get_block(block_no);
515 // get successor block succ_no
516 assert(succ_no < in->_num_succs, "illegal successor number");
517 Block* out = in->_succs[succ_no];
518 // Compute frequency of the new block. Do this before inserting
519 // new block in case succ_prob() needs to infer the probability from
520 // surrounding blocks.
521 float freq = in->_freq * in->succ_prob(succ_no);
522 // get ProjNode corresponding to the succ_no'th successor of the in block
523 ProjNode* proj = in->get_node(in->number_of_nodes() - in->_num_succs + succ_no)->as_Proj();
524 // create region for basic block
525 RegionNode* region = new RegionNode(2);
526 region->init_req(1, proj);
527 // setup corresponding basic block
528 Block* block = new (_block_arena) Block(_block_arena, region);
529 map_node_to_block(region, block);
530 C->regalloc()->set_bad(region->_idx);
531 // add a goto node
532 Node* gto = _goto->clone(); // get a new goto node
533 gto->set_req(0, region);
534 // add it to the basic block
535 block->push_node(gto);
536 map_node_to_block(gto, block);
537 C->regalloc()->set_bad(gto->_idx);
538 // hook up successor block
539 block->_succs.map(block->_num_succs++, out);
540 // remap successor's predecessors if necessary
541 for (uint i = 1; i < out->num_preds(); i++) {
542 if (out->pred(i) == proj) out->head()->set_req(i, gto);
543 }
544 // remap predecessor's successor to new block
545 in->_succs.map(succ_no, block);
546 // Set the frequency of the new block
547 block->_freq = freq;
548 // add new basic block to basic block list
549 add_block_at(block_no + 1, block);
550 // Update dominator tree information of the new goto block.
551 block->_idom = in;
552 block->_dom_depth = in->_dom_depth + 1;
553 if (out->_idom != in) {
554 // The successor block was not immediately dominated by the predecessor
555 // block, so there is no dominator subtree to update.
556 return;
557 }
558 // Update immediate dominator of the successor block.
559 out->_idom = block;
560 // Increment the dominator tree depth of the goto block's descendants. These
561 // are found by a depth-first search starting from the successor block. Two
562 // domination properties guarantee that only descendant blocks are visited:
563 // 1) all dominators of a block b must appear in any path from the root to b;
564 // 2) if a block b does not dominate another block b', b cannot dominate any
565 // block reachable from b' either.
566 // The exploration uses header indices as block identifiers, since
567 // Block::_pre_order might not be unique in the context of this function.
568 ResourceMark rm;
569 VectorSet descendants;
570 descendants.set(block->head()->_idx); // The goto block is a descendant of itself.
571 Block_List worklist;
572 worklist.push(out); // Start exploring from the successor block.
573 while (worklist.size() > 0) {
574 Block* b = worklist.pop();
575 // The immediate dominator of b is a descendant, hence b is also a
576 // descendant. Even though all predecessors of b might not have been visited
577 // yet, we know that all dominators of b have been already visited (since
578 // they must appear in any path from the goto block to b).
579 descendants.set(b->head()->_idx);
580 b->_dom_depth++;
581 for (uint i = 0; i < b->_num_succs; i++) {
582 Block* s = b->_succs[i];
583 if (s != get_root_block() &&
584 !descendants.test(s->head()->_idx) &&
585 // Do not search below non-descendant successors, since any block
586 // reachable from them cannot be descendant either.
587 descendants.test(s->_idom->head()->_idx)) {
588 worklist.push(s);
589 }
590 }
591 }
592 }
593
594 // Does this block end in a multiway branch that cannot have the default case
595 // flipped for another case?
596 static bool no_flip_branch(Block *b) {
597 int branch_idx = b->number_of_nodes() - b->_num_succs-1;
598 if (branch_idx < 1) {
599 return false;
600 }
601 Node *branch = b->get_node(branch_idx);
602 if (branch->is_Catch()) {
603 return true;
604 }
605 if (branch->is_Mach()) {
606 if (branch->is_MachNullCheck()) {
607 return true;
608 }
609 int iop = branch->as_Mach()->ideal_Opcode();
610 if (iop == Op_FastLock || iop == Op_FastUnlock) {
611 return true;
612 }
613 // Don't flip if branch has an implicit check.
614 if (branch->as_Mach()->is_TrapBasedCheckNode()) {
615 return true;
616 }
617 }
618 return false;
619 }
620
621 // Check for NeverBranch at block end. This needs to become a GOTO to the
622 // true target. NeverBranch are treated as a conditional branch that always
623 // goes the same direction for most of the optimizer and are used to give a
624 // fake exit path to infinite loops. At this late stage they need to turn
625 // into Goto's so that when you enter the infinite loop you indeed hang.
626 void PhaseCFG::convert_NeverBranch_to_Goto(Block *b) {
627 int end_idx = b->end_idx();
628 NeverBranchNode* never_branch = b->get_node(end_idx)->as_NeverBranch();
629 Block* succ = get_block_for_node(never_branch->proj_out(0)->unique_ctrl_out());
630 Block* dead = get_block_for_node(never_branch->proj_out(1)->unique_ctrl_out());
631 assert(succ == b->_succs[0] || succ == b->_succs[1], "succ is a successor");
632 assert(dead == b->_succs[0] || dead == b->_succs[1], "dead is a successor");
633
634 Node* gto = _goto->clone(); // get a new goto node
635 gto->set_req(0, b->head());
636 Node *bp = b->get_node(end_idx);
637 b->map_node(gto, end_idx); // Slam over NeverBranch
638 map_node_to_block(gto, b);
639 C->regalloc()->set_bad(gto->_idx);
640 b->pop_node(); // Yank projections
641 b->pop_node(); // Yank projections
642 b->_succs.map(0,succ); // Map only successor
643 b->_num_succs = 1;
644 // remap successor's predecessors if necessary
645 uint j;
646 for (j = 1; j < succ->num_preds(); j++) {
647 if (succ->pred(j)->in(0) == bp) {
648 succ->head()->set_req(j, gto);
649 }
650 }
651 // Kill alternate exit path
652 for (j = 1; j < dead->num_preds(); j++) {
653 if (dead->pred(j)->in(0) == bp) {
654 break;
655 }
656 }
657 // Scan through block, yanking dead path from
658 // all regions and phis.
659 dead->head()->del_req(j);
660 for (int k = 1; dead->get_node(k)->is_Phi(); k++) {
661 dead->get_node(k)->del_req(j);
662 }
663 }
664
665 // Helper function to move block bx to the slot following b_index. Return
666 // true if the move is successful, otherwise false
667 bool PhaseCFG::move_to_next(Block* bx, uint b_index) {
668 if (bx == nullptr) return false;
669
670 // Return false if bx is already scheduled.
671 uint bx_index = bx->_pre_order;
672 if ((bx_index <= b_index) && (get_block(bx_index) == bx)) {
673 return false;
674 }
675
676 // Find the current index of block bx on the block list
677 bx_index = b_index + 1;
678 while (bx_index < number_of_blocks() && get_block(bx_index) != bx) {
679 bx_index++;
680 }
681 assert(get_block(bx_index) == bx, "block not found");
682
683 // If the previous block conditionally falls into bx, return false,
684 // because moving bx will create an extra jump.
685 for(uint k = 1; k < bx->num_preds(); k++ ) {
686 Block* pred = get_block_for_node(bx->pred(k));
687 if (pred == get_block(bx_index - 1)) {
688 if (pred->_num_succs != 1) {
689 return false;
690 }
691 }
692 }
693
694 // Reinsert bx just past block 'b'
695 _blocks.remove(bx_index);
696 _blocks.insert(b_index + 1, bx);
697 return true;
698 }
699
700 // Move empty and uncommon blocks to the end.
701 void PhaseCFG::move_to_end(Block *b, uint i) {
702 int e = b->is_Empty();
703 if (e != Block::not_empty) {
704 if (e == Block::empty_with_goto) {
705 // Remove the goto, but leave the block.
706 b->pop_node();
707 }
708 // Mark this block as a connector block, which will cause it to be
709 // ignored in certain functions such as non_connector_successor().
710 b->set_connector();
711 }
712 // Move the empty block to the end, and don't recheck.
713 _blocks.remove(i);
714 _blocks.push(b);
715 }
716
717 // Set loop alignment for every block
718 void PhaseCFG::set_loop_alignment() {
719 uint last = number_of_blocks();
720 assert(get_block(0) == get_root_block(), "");
721
722 for (uint i = 1; i < last; i++) {
723 Block* block = get_block(i);
724 if (block->head()->is_Loop()) {
725 block->set_loop_alignment(block);
726 }
727 }
728 }
729
730 // Make empty basic blocks to be "connector" blocks, Move uncommon blocks
731 // to the end.
732 void PhaseCFG::remove_empty_blocks() {
733 // Move uncommon blocks to the end
734 uint last = number_of_blocks();
735 assert(get_block(0) == get_root_block(), "");
736
737 for (uint i = 1; i < last; i++) {
738 Block* block = get_block(i);
739 if (block->is_connector()) {
740 break;
741 }
742
743 // Check for NeverBranch at block end. This needs to become a GOTO to the
744 // true target. NeverBranch are treated as a conditional branch that
745 // always goes the same direction for most of the optimizer and are used
746 // to give a fake exit path to infinite loops. At this late stage they
747 // need to turn into Goto's so that when you enter the infinite loop you
748 // indeed hang.
749 if (block->get_node(block->end_idx())->is_NeverBranch()) {
750 convert_NeverBranch_to_Goto(block);
751 }
752
753 // Look for uncommon blocks and move to end.
754 if (!C->do_freq_based_layout()) {
755 if (is_uncommon(block)) {
756 move_to_end(block, i);
757 last--; // No longer check for being uncommon!
758 if (no_flip_branch(block)) { // Fall-thru case must follow?
759 // Find the fall-thru block
760 block = get_block(i);
761 move_to_end(block, i);
762 last--;
763 }
764 // backup block counter post-increment
765 i--;
766 }
767 }
768 }
769
770 // Move empty blocks to the end
771 last = number_of_blocks();
772 for (uint i = 1; i < last; i++) {
773 Block* block = get_block(i);
774 if (block->is_Empty() != Block::not_empty) {
775 move_to_end(block, i);
776 last--;
777 i--;
778 }
779 } // End of for all blocks
780 }
781
782 Block *PhaseCFG::fixup_trap_based_check(Node *branch, Block *block, int block_pos, Block *bnext) {
783 // Trap based checks must fall through to the successor with
784 // PROB_ALWAYS.
785 // They should be an If with 2 successors.
786 assert(branch->is_MachIf(), "must be If");
787 assert(block->_num_succs == 2, "must have 2 successors");
788
789 // Get the If node and the projection for the first successor.
790 MachIfNode *iff = block->get_node(block->number_of_nodes()-3)->as_MachIf();
791 ProjNode *proj0 = block->get_node(block->number_of_nodes()-2)->as_Proj();
792 ProjNode *proj1 = block->get_node(block->number_of_nodes()-1)->as_Proj();
793 ProjNode *projt = (proj0->Opcode() == Op_IfTrue) ? proj0 : proj1;
794 ProjNode *projf = (proj0->Opcode() == Op_IfFalse) ? proj0 : proj1;
795
796 // Assert that proj0 and succs[0] match up. Similarly for proj1 and succs[1].
797 assert(proj0->raw_out(0) == block->_succs[0]->head(), "Mismatch successor 0");
798 assert(proj1->raw_out(0) == block->_succs[1]->head(), "Mismatch successor 1");
799
800 ProjNode *proj_always;
801 ProjNode *proj_never;
802 // We must negate the branch if the implicit check doesn't follow
803 // the branch's TRUE path. Then, the new TRUE branch target will
804 // be the old FALSE branch target.
805 if (iff->_prob <= 2*PROB_NEVER) { // There are small rounding errors.
806 proj_never = projt;
807 proj_always = projf;
808 } else {
809 // We must negate the branch if the trap doesn't follow the
810 // branch's TRUE path. Then, the new TRUE branch target will
811 // be the old FALSE branch target.
812 proj_never = projf;
813 proj_always = projt;
814 iff->negate();
815 }
816 assert(iff->_prob <= 2*PROB_NEVER, "Trap based checks are expected to trap never!");
817 // Map the successors properly
818 block->_succs.map(0, get_block_for_node(proj_never ->raw_out(0))); // The target of the trap.
819 block->_succs.map(1, get_block_for_node(proj_always->raw_out(0))); // The fall through target.
820
821 if (block->get_node(block->number_of_nodes() - block->_num_succs + 1) != proj_always) {
822 block->map_node(proj_never, block->number_of_nodes() - block->_num_succs + 0);
823 block->map_node(proj_always, block->number_of_nodes() - block->_num_succs + 1);
824 }
825
826 // Place the fall through block after this block.
827 Block *bs1 = block->non_connector_successor(1);
828 if (bs1 != bnext && move_to_next(bs1, block_pos)) {
829 bnext = bs1;
830 }
831 // If the fall through block still is not the next block, insert a goto.
832 if (bs1 != bnext) {
833 insert_goto_at(block_pos, 1);
834 }
835 return bnext;
836 }
837
838 // Fix up the final control flow for basic blocks.
839 void PhaseCFG::fixup_flow() {
840 // Fixup final control flow for the blocks. Remove jump-to-next
841 // block. If neither arm of an IF follows the conditional branch, we
842 // have to add a second jump after the conditional. We place the
843 // TRUE branch target in succs[0] for both GOTOs and IFs.
844 for (uint i = 0; i < number_of_blocks(); i++) {
845 Block* block = get_block(i);
846 block->_pre_order = i; // turn pre-order into block-index
847
848 // Connector blocks need no further processing.
849 if (block->is_connector()) {
850 assert((i+1) == number_of_blocks() || get_block(i + 1)->is_connector(), "All connector blocks should sink to the end");
851 continue;
852 }
853 assert(block->is_Empty() != Block::completely_empty, "Empty blocks should be connectors");
854
855 Block* bnext = (i < number_of_blocks() - 1) ? get_block(i + 1) : nullptr;
856 Block* bs0 = block->non_connector_successor(0);
857
858 // Check for multi-way branches where I cannot negate the test to
859 // exchange the true and false targets.
860 if (no_flip_branch(block)) {
861 // Find fall through case - if must fall into its target.
862 // Get the index of the branch's first successor.
863 int branch_idx = block->number_of_nodes() - block->_num_succs;
864
865 // The branch is 1 before the branch's first successor.
866 Node *branch = block->get_node(branch_idx-1);
867
868 // Handle no-flip branches which have implicit checks and which require
869 // special block ordering and individual semantics of the 'fall through
870 // case'.
871 if ((TrapBasedNullChecks || TrapBasedRangeChecks) &&
872 branch->is_Mach() && branch->as_Mach()->is_TrapBasedCheckNode()) {
873 bnext = fixup_trap_based_check(branch, block, i, bnext);
874 } else {
875 // Else, default handling for no-flip branches
876 for (uint j2 = 0; j2 < block->_num_succs; j2++) {
877 const ProjNode* p = block->get_node(branch_idx + j2)->as_Proj();
878 if (p->_con == 0) {
879 // successor j2 is fall through case
880 if (block->non_connector_successor(j2) != bnext) {
881 // but it is not the next block => insert a goto
882 insert_goto_at(i, j2);
883 }
884 // Put taken branch in slot 0
885 if (j2 == 0 && block->_num_succs == 2) {
886 // Flip targets in succs map
887 Block *tbs0 = block->_succs[0];
888 Block *tbs1 = block->_succs[1];
889 block->_succs.map(0, tbs1);
890 block->_succs.map(1, tbs0);
891 }
892 break;
893 }
894 }
895 }
896
897 // Remove all CatchProjs
898 for (uint j = 0; j < block->_num_succs; j++) {
899 block->pop_node();
900 }
901
902 } else if (block->_num_succs == 1) {
903 // Block ends in a Goto?
904 if (bnext == bs0) {
905 // We fall into next block; remove the Goto
906 block->pop_node();
907 }
908
909 } else if(block->_num_succs == 2) { // Block ends in a If?
910 // Get opcode of 1st projection (matches _succs[0])
911 // Note: Since this basic block has 2 exits, the last 2 nodes must
912 // be projections (in any order), the 3rd last node must be
913 // the IfNode (we have excluded other 2-way exits such as
914 // CatchNodes already).
915 MachNode* iff = block->get_node(block->number_of_nodes() - 3)->as_Mach();
916 ProjNode* proj0 = block->get_node(block->number_of_nodes() - 2)->as_Proj();
917 ProjNode* proj1 = block->get_node(block->number_of_nodes() - 1)->as_Proj();
918
919 // Assert that proj0 and succs[0] match up. Similarly for proj1 and succs[1].
920 assert(proj0->raw_out(0) == block->_succs[0]->head(), "Mismatch successor 0");
921 assert(proj1->raw_out(0) == block->_succs[1]->head(), "Mismatch successor 1");
922
923 Block* bs1 = block->non_connector_successor(1);
924
925 // Check for neither successor block following the current
926 // block ending in a conditional. If so, move one of the
927 // successors after the current one, provided that the
928 // successor was previously unscheduled, but moveable
929 // (i.e., all paths to it involve a branch).
930 if (!C->do_freq_based_layout() && bnext != bs0 && bnext != bs1) {
931 // Choose the more common successor based on the probability
932 // of the conditional branch.
933 Block* bx = bs0;
934 Block* by = bs1;
935
936 // _prob is the probability of taking the true path. Make
937 // p the probability of taking successor #1.
938 float p = iff->as_MachIf()->_prob;
939 if (proj0->Opcode() == Op_IfTrue) {
940 p = 1.0 - p;
941 }
942
943 // Prefer successor #1 if p > 0.5
944 if (p > PROB_FAIR) {
945 bx = bs1;
946 by = bs0;
947 }
948
949 // Attempt the more common successor first
950 if (move_to_next(bx, i)) {
951 bnext = bx;
952 } else if (move_to_next(by, i)) {
953 bnext = by;
954 }
955 }
956
957 // Check for conditional branching the wrong way. Negate
958 // conditional, if needed, so it falls into the following block
959 // and branches to the not-following block.
960
961 // Check for the next block being in succs[0]. We are going to branch
962 // to succs[0], so we want the fall-thru case as the next block in
963 // succs[1].
964 if (bnext == bs0) {
965 // Fall-thru case in succs[0], should be in succs[1], so flip targets in _succs map
966 Block* tbs0 = block->_succs[0];
967 Block* tbs1 = block->_succs[1];
968 block->_succs.map(0, tbs1);
969 block->_succs.map(1, tbs0);
970 // Flip projection for each target
971 swap(proj0, proj1);
972 } else if(bnext != bs1) {
973 // Need a double-branch
974 // The existing conditional branch need not change.
975 // Add a unconditional branch to the false target.
976 // Alas, it must appear in its own block and adding a
977 // block this late in the game is complicated. Sigh.
978 insert_goto_at(i, 1);
979 }
980
981 // Make sure we TRUE branch to the target
982 if (proj0->Opcode() == Op_IfFalse) {
983 iff->as_MachIf()->negate();
984 }
985
986 block->pop_node(); // Remove IfFalse & IfTrue projections
987 block->pop_node();
988
989 } else {
990 // Multi-exit block, e.g. a switch statement
991 // But we don't need to do anything here
992 }
993 } // End of for all blocks
994 }
995
996 void PhaseCFG::remove_unreachable_blocks() {
997 ResourceMark rm;
998 Block_List unreachable;
999 // Initialize worklist of unreachable blocks to be removed.
1000 for (uint i = 0; i < number_of_blocks(); i++) {
1001 Block* block = get_block(i);
1002 assert(block->_pre_order == i, "Block::pre_order does not match block index");
1003 if (block->is_trivially_unreachable()) {
1004 unreachable.push(block);
1005 }
1006 }
1007 // Now remove all blocks that are transitively unreachable.
1008 while (unreachable.size() > 0) {
1009 Block* dead = unreachable.pop();
1010 // When this code runs (after PhaseCFG::fixup_flow()), Block::_pre_order
1011 // does not contain pre-order but block-list indices. Ensure they stay
1012 // contiguous by decrementing _pre_order for all elements after 'dead'.
1013 // Block::_rpo does not contain valid reverse post-order indices anymore
1014 // (they are invalidated by block insertions in PhaseCFG::fixup_flow()),
1015 // so there is no need to update them.
1016 for (uint i = dead->_pre_order + 1; i < number_of_blocks(); i++) {
1017 get_block(i)->_pre_order--;
1018 }
1019 _blocks.remove(dead->_pre_order);
1020 _number_of_blocks--;
1021 // Update the successors' predecessor list and push new unreachable blocks.
1022 for (uint i = 0; i < dead->_num_succs; i++) {
1023 Block* succ = dead->_succs[i];
1024 Node* head = succ->head();
1025 for (int j = head->req() - 1; j >= 1; j--) {
1026 if (get_block_for_node(head->in(j)) == dead) {
1027 head->del_req(j);
1028 }
1029 }
1030 if (succ->is_trivially_unreachable()) {
1031 unreachable.push(succ);
1032 }
1033 }
1034 }
1035 }
1036
1037 // postalloc_expand: Expand nodes after register allocation.
1038 //
1039 // postalloc_expand has to be called after register allocation, just
1040 // before output (i.e. scheduling). It only gets called if
1041 // Matcher::require_postalloc_expand is true.
1042 //
1043 // Background:
1044 //
1045 // Nodes that are expandend (one compound node requiring several
1046 // assembler instructions to be implemented split into two or more
1047 // non-compound nodes) after register allocation are not as nice as
1048 // the ones expanded before register allocation - they don't
1049 // participate in optimizations as global code motion. But after
1050 // register allocation we can expand nodes that use registers which
1051 // are not spillable or registers that are not allocated, because the
1052 // old compound node is simply replaced (in its location in the basic
1053 // block) by a new subgraph which does not contain compound nodes any
1054 // more. The scheduler called during output can later on process these
1055 // non-compound nodes.
1056 //
1057 // Implementation:
1058 //
1059 // Nodes requiring postalloc expand are specified in the ad file by using
1060 // a postalloc_expand statement instead of ins_encode. A postalloc_expand
1061 // contains a single call to an encoding, as does an ins_encode
1062 // statement. Instead of an emit() function a postalloc_expand() function
1063 // is generated that doesn't emit assembler but creates a new
1064 // subgraph. The code below calls this postalloc_expand function for each
1065 // node with the appropriate attribute. This function returns the new
1066 // nodes generated in an array passed in the call. The old node,
1067 // potential MachTemps before and potential Projs after it then get
1068 // disconnected and replaced by the new nodes. The instruction
1069 // generating the result has to be the last one in the array. In
1070 // general it is assumed that Projs after the node expanded are
1071 // kills. These kills are not required any more after expanding as
1072 // there are now explicitly visible def-use chains and the Projs are
1073 // removed. This does not hold for calls: They do not only have
1074 // kill-Projs but also Projs defining values. Therefore Projs after
1075 // the node expanded are removed for all but for calls. If a node is
1076 // to be reused, it must be added to the nodes list returned, and it
1077 // will be added again.
1078 //
1079 // Implementing the postalloc_expand function for a node in an enc_class
1080 // is rather tedious. It requires knowledge about many node details, as
1081 // the nodes and the subgraph must be hand crafted. To simplify this,
1082 // adlc generates some utility variables into the postalloc_expand function,
1083 // e.g., holding the operands as specified by the postalloc_expand encoding
1084 // specification, e.g.:
1085 // * unsigned idx_<par_name> holding the index of the node in the ins
1086 // * Node *n_<par_name> holding the node loaded from the ins
1087 // * MachOpnd *op_<par_name> holding the corresponding operand
1088 //
1089 // The ordering of operands can not be determined by looking at a
1090 // rule. Especially if a match rule matches several different trees,
1091 // several nodes are generated from one instruct specification with
1092 // different operand orderings. In this case the adlc generated
1093 // variables are the only way to access the ins and operands
1094 // deterministically.
1095 //
1096 // If assigning a register to a node that contains an oop, don't
1097 // forget to call ra_->set_oop() for the node.
1098 void PhaseCFG::postalloc_expand(PhaseRegAlloc* _ra) {
1099 GrowableArray <Node *> new_nodes(32); // Array with new nodes filled by postalloc_expand function of node.
1100 GrowableArray <Node *> remove(32);
1101 GrowableArray <Node *> succs(32);
1102 unsigned int max_idx = C->unique(); // Remember to distinguish new from old nodes.
1103 DEBUG_ONLY(bool foundNode = false);
1104
1105 // for all blocks
1106 for (uint i = 0; i < number_of_blocks(); i++) {
1107 Block *b = _blocks[i];
1108 // For all instructions in the current block.
1109 for (uint j = 0; j < b->number_of_nodes(); j++) {
1110 Node *n = b->get_node(j);
1111 if (n->is_Mach() && n->as_Mach()->requires_postalloc_expand()) {
1112 #ifdef ASSERT
1113 if (TracePostallocExpand) {
1114 if (!foundNode) {
1115 foundNode = true;
1116 tty->print("POSTALLOC EXPANDING %d %s\n", C->compile_id(),
1117 C->method() ? C->method()->name()->as_utf8() : C->stub_name());
1118 }
1119 tty->print(" postalloc expanding "); n->dump();
1120 if (Verbose) {
1121 tty->print(" with ins:\n");
1122 for (uint k = 0; k < n->len(); ++k) {
1123 if (n->in(k)) { tty->print(" "); n->in(k)->dump(); }
1124 }
1125 }
1126 }
1127 #endif
1128 new_nodes.clear();
1129 // Collect nodes that have to be removed from the block later on.
1130 uint req = n->req();
1131 remove.clear();
1132 for (uint k = 0; k < req; ++k) {
1133 if (n->in(k) && n->in(k)->is_MachTemp()) {
1134 remove.push(n->in(k)); // MachTemps which are inputs to the old node have to be removed.
1135 n->in(k)->del_req(0);
1136 j--;
1137 }
1138 }
1139
1140 // Check whether we can allocate enough nodes. We set a fix limit for
1141 // the size of postalloc expands with this.
1142 uint unique_limit = C->unique() + 40;
1143 if (unique_limit >= _ra->node_regs_max_index()) {
1144 Compile::current()->record_failure("out of nodes in postalloc expand");
1145 return;
1146 }
1147
1148 // Emit (i.e. generate new nodes).
1149 n->as_Mach()->postalloc_expand(&new_nodes, _ra);
1150
1151 assert(C->unique() < unique_limit, "You allocated too many nodes in your postalloc expand.");
1152
1153 // Disconnect the inputs of the old node.
1154 //
1155 // We reuse MachSpillCopy nodes. If we need to expand them, there
1156 // are many, so reusing pays off. If reused, the node already
1157 // has the new ins. n must be the last node on new_nodes list.
1158 if (!n->is_MachSpillCopy()) {
1159 for (int k = req - 1; k >= 0; --k) {
1160 n->del_req(k);
1161 }
1162 }
1163
1164 #ifdef ASSERT
1165 // Check that all nodes have proper operands.
1166 for (int k = 0; k < new_nodes.length(); ++k) {
1167 if (new_nodes.at(k)->_idx < max_idx || !new_nodes.at(k)->is_Mach()) continue; // old node, Proj ...
1168 MachNode *m = new_nodes.at(k)->as_Mach();
1169 for (unsigned int l = 0; l < m->num_opnds(); ++l) {
1170 if (MachOper::notAnOper(m->_opnds[l])) {
1171 outputStream *os = tty;
1172 os->print("Node %s ", m->Name());
1173 os->print("has invalid opnd %d: %p\n", l, m->_opnds[l]);
1174 assert(0, "Invalid operands, see inline trace in hs_err_pid file.");
1175 }
1176 }
1177 }
1178 #endif
1179
1180 // Collect succs of old node in remove (for projections) and in succs (for
1181 // all other nodes) do _not_ collect projections in remove (but in succs)
1182 // in case the node is a call. We need the projections for calls as they are
1183 // associated with registers (i.e. they are defs).
1184 succs.clear();
1185 for (DUIterator k = n->outs(); n->has_out(k); k++) {
1186 if (n->out(k)->is_Proj() && !n->is_MachCall() && !n->is_MachBranch()) {
1187 remove.push(n->out(k));
1188 } else {
1189 succs.push(n->out(k));
1190 }
1191 }
1192 // Replace old node n as input of its succs by last of the new nodes.
1193 for (int k = 0; k < succs.length(); ++k) {
1194 Node *succ = succs.at(k);
1195 for (uint l = 0; l < succ->req(); ++l) {
1196 if (succ->in(l) == n) {
1197 succ->set_req(l, new_nodes.at(new_nodes.length() - 1));
1198 }
1199 }
1200 for (uint l = succ->req(); l < succ->len(); ++l) {
1201 if (succ->in(l) == n) {
1202 succ->set_prec(l, new_nodes.at(new_nodes.length() - 1));
1203 }
1204 }
1205 }
1206
1207 // Index of old node in block.
1208 uint index = b->find_node(n);
1209 // Insert new nodes into block and map them in nodes->blocks array
1210 // and remember last node in n2.
1211 Node *n2 = nullptr;
1212 for (int k = 0; k < new_nodes.length(); ++k) {
1213 n2 = new_nodes.at(k);
1214 b->insert_node(n2, ++index);
1215 map_node_to_block(n2, b);
1216 }
1217
1218 // Add old node n to remove and remove them all from block.
1219 remove.push(n);
1220 j--;
1221 #ifdef ASSERT
1222 if (TracePostallocExpand && Verbose) {
1223 tty->print(" removing:\n");
1224 for (int k = 0; k < remove.length(); ++k) {
1225 tty->print(" "); remove.at(k)->dump();
1226 }
1227 tty->print(" inserting:\n");
1228 for (int k = 0; k < new_nodes.length(); ++k) {
1229 tty->print(" "); new_nodes.at(k)->dump();
1230 }
1231 }
1232 #endif
1233 for (int k = 0; k < remove.length(); ++k) {
1234 if (b->contains(remove.at(k))) {
1235 b->find_remove(remove.at(k));
1236 } else {
1237 assert(remove.at(k)->is_Proj() && (remove.at(k)->in(0)->is_MachBranch()), "");
1238 }
1239 }
1240 // If anything has been inserted (n2 != nullptr), continue after last node inserted.
1241 // This does not always work. Some postalloc expands don't insert any nodes, if they
1242 // do optimizations (e.g., max(x,x)). In this case we decrement j accordingly.
1243 j = n2 ? b->find_node(n2) : j;
1244 }
1245 }
1246 }
1247
1248 #ifdef ASSERT
1249 if (foundNode) {
1250 tty->print("FINISHED %d %s\n", C->compile_id(),
1251 C->method() ? C->method()->name()->as_utf8() : C->stub_name());
1252 tty->flush();
1253 }
1254 #endif
1255 }
1256
1257
1258 //------------------------------dump-------------------------------------------
1259 #ifndef PRODUCT
1260 void PhaseCFG::_dump_cfg( const Node *end, VectorSet &visited ) const {
1261 const Node *x = end->is_block_proj();
1262 assert( x, "not a CFG" );
1263
1264 // Do not visit this block again
1265 if( visited.test_set(x->_idx) ) return;
1266
1267 // Skip through this block
1268 const Node *p = x;
1269 do {
1270 p = p->in(0); // Move control forward
1271 assert( !p->is_block_proj() || p->is_Root(), "not a CFG" );
1272 } while( !p->is_block_start() );
1273
1274 // Recursively visit
1275 for (uint i = 1; i < p->req(); i++) {
1276 _dump_cfg(p->in(i), visited);
1277 }
1278
1279 // Dump the block
1280 get_block_for_node(p)->dump(this);
1281 }
1282
1283 void PhaseCFG::dump( ) const {
1284 tty->print("\n--- CFG --- %d BBs\n", number_of_blocks());
1285 if (_blocks.size()) { // Did we do basic-block layout?
1286 for (uint i = 0; i < number_of_blocks(); i++) {
1287 const Block* block = get_block(i);
1288 block->dump(this);
1289 }
1290 } else { // Else do it with a DFS
1291 VectorSet visited(_block_arena);
1292 _dump_cfg(_root,visited);
1293 }
1294 }
1295
1296 void PhaseCFG::dump_headers() {
1297 for (uint i = 0; i < number_of_blocks(); i++) {
1298 Block* block = get_block(i);
1299 if (block != nullptr) {
1300 block->dump_head(this);
1301 }
1302 }
1303 }
1304 #endif // !PRODUCT
1305
1306 #ifdef ASSERT
1307 void PhaseCFG::verify_memory_writer_placement(const Block* b, const Node* n) const {
1308 if (!n->is_memory_writer()) {
1309 return;
1310 }
1311 CFGLoop* home_or_ancestor = find_block_for_node(n->in(0))->_loop;
1312 bool found = false;
1313 do {
1314 if (b->_loop == home_or_ancestor) {
1315 found = true;
1316 break;
1317 }
1318 home_or_ancestor = home_or_ancestor->parent();
1319 } while (home_or_ancestor != nullptr);
1320 assert(found, "block b is not in n's home loop or an ancestor of it");
1321 }
1322
1323 void PhaseCFG::verify_dominator_tree() const {
1324 for (uint i = 0; i < number_of_blocks(); i++) {
1325 Block* block = get_block(i);
1326 assert(block->_dom_depth <= number_of_blocks(), "unexpected dominator tree depth");
1327 if (block == get_root_block()) {
1328 assert(block->_dom_depth == 1, "unexpected root dominator tree depth");
1329 // The root block does not have an immediate dominator, stop checking.
1330 continue;
1331 }
1332 assert(block->_idom != nullptr, "non-root blocks must have immediate dominators");
1333 assert(block->_dom_depth == block->_idom->_dom_depth + 1,
1334 "the dominator tree depth of a node must succeed that of its immediate dominator");
1335 assert(block->num_preds() > 2 || block->_idom == get_block_for_node(block->pred(1)),
1336 "the immediate dominator of a single-predecessor block must be the predecessor");
1337 }
1338 }
1339
1340 void PhaseCFG::verify() const {
1341 // Verify sane CFG
1342 for (uint i = 0; i < number_of_blocks(); i++) {
1343 Block* block = get_block(i);
1344 uint cnt = block->number_of_nodes();
1345 uint j;
1346 for (j = 0; j < cnt; j++) {
1347 Node *n = block->get_node(j);
1348 assert(get_block_for_node(n) == block, "");
1349 if (j >= 1 && n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_CreateEx) {
1350 assert(j == 1 || block->get_node(j-1)->is_Phi(), "CreateEx must be first instruction in block");
1351 }
1352 verify_memory_writer_placement(block, n);
1353 if (n->needs_anti_dependence_check()) {
1354 verify_anti_dependences(block, n);
1355 if (C->failing()) {
1356 return;
1357 }
1358 }
1359 for (uint k = 0; k < n->req(); k++) {
1360 Node *def = n->in(k);
1361 if (def && def != n) {
1362 Block* def_block = get_block_for_node(def);
1363 assert(def_block || def->is_Con(), "must have block; constants for debug info ok");
1364 // Verify that all definitions dominate their uses (except for virtual
1365 // instructions merging multiple definitions).
1366 assert(n->is_Root() || n->is_Region() || n->is_Phi() || n->is_MachMerge() ||
1367 def_block->dominates(block),
1368 "uses must be dominated by definitions");
1369 // Verify that instructions in the block are in correct order.
1370 // Uses must follow their definition if they are at the same block.
1371 // Mostly done to check that MachSpillCopy nodes are placed correctly
1372 // when CreateEx node is moved in build_ifg_physical().
1373 if (def_block == block && !(block->head()->is_Loop() && n->is_Phi()) &&
1374 // See (+++) comment in reg_split.cpp
1375 !(n->jvms() != nullptr && n->jvms()->is_monitor_use(k))) {
1376 bool is_loop = false;
1377 if (n->is_Phi()) {
1378 for (uint l = 1; l < def->req(); l++) {
1379 if (n == def->in(l)) {
1380 is_loop = true;
1381 break; // Some kind of loop
1382 }
1383 }
1384 }
1385 // Uses must be before definition, except if:
1386 // - We are in some kind of loop we already detected
1387 // - We are in infinite loop, where Region may not have been turned into LoopNode
1388 assert(block->find_node(def) < j ||
1389 is_loop ||
1390 (n->is_Phi() && block->head()->as_Region()->is_in_infinite_subgraph()),
1391 "uses must follow definitions (except in loops)");
1392 }
1393 }
1394 }
1395 if (n->is_Proj()) {
1396 assert(j >= 1, "a projection cannot be the first instruction in a block");
1397 Node* pred = block->get_node(j - 1);
1398 Node* parent = n->in(0);
1399 assert(parent != nullptr, "projections must have a parent");
1400 assert(pred == parent || (pred->is_Proj() && pred->in(0) == parent),
1401 "projections must follow their parents or other sibling projections");
1402 }
1403 }
1404
1405 j = block->end_idx();
1406 Node* bp = (Node*)block->get_node(block->number_of_nodes() - 1)->is_block_proj();
1407 assert(bp, "last instruction must be a block proj");
1408 assert(bp == block->get_node(j), "wrong number of successors for this block");
1409 if (bp->is_Catch()) {
1410 while (block->get_node(--j)->is_MachProj()) {
1411 ;
1412 }
1413 assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call");
1414 } else if (bp->is_Mach() && bp->as_Mach()->ideal_Opcode() == Op_If) {
1415 assert(block->_num_succs == 2, "Conditional branch must have two targets");
1416 }
1417 }
1418 verify_dominator_tree();
1419 }
1420 #endif // ASSERT
1421
1422 UnionFind::UnionFind( uint max ) : _cnt(max), _max(max), _indices(NEW_RESOURCE_ARRAY(uint,max)) {
1423 Copy::zero_to_bytes( _indices, sizeof(uint)*max );
1424 }
1425
1426 void UnionFind::extend( uint from_idx, uint to_idx ) {
1427 _nesting.check(); // Check if a potential reallocation in the resource arena is safe
1428 if( from_idx >= _max ) {
1429 uint size = 16;
1430 while( size <= from_idx ) size <<=1;
1431 _indices = REALLOC_RESOURCE_ARRAY( uint, _indices, _max, size );
1432 _max = size;
1433 }
1434 while( _cnt <= from_idx ) _indices[_cnt++] = 0;
1435 _indices[from_idx] = to_idx;
1436 }
1437
1438 void UnionFind::reset( uint max ) {
1439 // Force the Union-Find mapping to be at least this large
1440 extend(max,0);
1441 // Initialize to be the ID mapping.
1442 for( uint i=0; i<max; i++ ) map(i,i);
1443 }
1444
1445 // Straight out of Tarjan's union-find algorithm
1446 uint UnionFind::Find_compress( uint idx ) {
1447 uint cur = idx;
1448 uint next = lookup(cur);
1449 while( next != cur ) { // Scan chain of equivalences
1450 assert( next < cur, "always union smaller" );
1451 cur = next; // until find a fixed-point
1452 next = lookup(cur);
1453 }
1454 // Core of union-find algorithm: update chain of
1455 // equivalences to be equal to the root.
1456 while( idx != next ) {
1457 uint tmp = lookup(idx);
1458 map(idx, next);
1459 idx = tmp;
1460 }
1461 return idx;
1462 }
1463
1464 // Like Find above, but no path compress, so bad asymptotic behavior
1465 uint UnionFind::Find_const( uint idx ) const {
1466 if( idx == 0 ) return idx; // Ignore the zero idx
1467 // Off the end? This can happen during debugging dumps
1468 // when data structures have not finished being updated.
1469 if( idx >= _max ) return idx;
1470 uint next = lookup(idx);
1471 while( next != idx ) { // Scan chain of equivalences
1472 idx = next; // until find a fixed-point
1473 next = lookup(idx);
1474 }
1475 return next;
1476 }
1477
1478 // union 2 sets together.
1479 void UnionFind::Union( uint idx1, uint idx2 ) {
1480 uint src = Find(idx1);
1481 uint dst = Find(idx2);
1482 assert( src, "" );
1483 assert( dst, "" );
1484 assert( src < _max, "oob" );
1485 assert( dst < _max, "oob" );
1486 assert( src < dst, "always union smaller" );
1487 map(dst,src);
1488 }
1489
1490 #ifndef PRODUCT
1491 void Trace::dump( ) const {
1492 tty->print_cr("Trace (freq %f)", first_block()->_freq);
1493 for (Block *b = first_block(); b != nullptr; b = next(b)) {
1494 tty->print(" B%d", b->_pre_order);
1495 if (b->head()->is_Loop()) {
1496 tty->print(" (L%d)", b->compute_loop_alignment());
1497 }
1498 if (b->has_loop_alignment()) {
1499 tty->print(" (T%d)", b->code_alignment());
1500 }
1501 }
1502 tty->cr();
1503 }
1504
1505 void CFGEdge::dump( ) const {
1506 tty->print(" B%d --> B%d Freq: %f out:%3d%% in:%3d%% State: ",
1507 from()->_pre_order, to()->_pre_order, freq(), _from_pct, _to_pct);
1508 switch(state()) {
1509 case connected:
1510 tty->print("connected");
1511 break;
1512 case open:
1513 tty->print("open");
1514 break;
1515 case interior:
1516 tty->print("interior");
1517 break;
1518 }
1519 if (infrequent()) {
1520 tty->print(" infrequent");
1521 }
1522 tty->cr();
1523 }
1524 #endif
1525
1526 // Comparison function for edges
1527 static int edge_order(CFGEdge **e0, CFGEdge **e1) {
1528 float freq0 = (*e0)->freq();
1529 float freq1 = (*e1)->freq();
1530 if (freq0 != freq1) {
1531 return freq0 > freq1 ? -1 : 1;
1532 }
1533
1534 int dist0 = (*e0)->to()->_rpo - (*e0)->from()->_rpo;
1535 int dist1 = (*e1)->to()->_rpo - (*e1)->from()->_rpo;
1536
1537 return dist1 - dist0;
1538 }
1539
1540 // Comparison function for edges
1541 extern "C" int trace_frequency_order(const void *p0, const void *p1) {
1542 Trace *tr0 = *(Trace **) p0;
1543 Trace *tr1 = *(Trace **) p1;
1544 Block *b0 = tr0->first_block();
1545 Block *b1 = tr1->first_block();
1546
1547 // The trace of connector blocks goes at the end;
1548 // we only expect one such trace
1549 if (b0->is_connector() != b1->is_connector()) {
1550 return b1->is_connector() ? -1 : 1;
1551 }
1552
1553 // Pull more frequently executed blocks to the beginning
1554 float freq0 = b0->_freq;
1555 float freq1 = b1->_freq;
1556 if (freq0 != freq1) {
1557 return freq0 > freq1 ? -1 : 1;
1558 }
1559
1560 int diff = tr0->first_block()->_rpo - tr1->first_block()->_rpo;
1561
1562 return diff;
1563 }
1564
1565 // Find edges of interest, i.e, those which can fall through. Presumes that
1566 // edges which don't fall through are of low frequency and can be generally
1567 // ignored. Initialize the list of traces.
1568 void PhaseBlockLayout::find_edges() {
1569 // Walk the blocks, creating edges and Traces
1570 uint i;
1571 Trace *tr = nullptr;
1572 for (i = 0; i < _cfg.number_of_blocks(); i++) {
1573 Block* b = _cfg.get_block(i);
1574 tr = new Trace(b, next, prev);
1575 traces[tr->id()] = tr;
1576
1577 // All connector blocks should be at the end of the list
1578 if (b->is_connector()) break;
1579
1580 // If this block and the next one have a one-to-one successor
1581 // predecessor relationship, simply append the next block
1582 int nfallthru = b->num_fall_throughs();
1583 while (nfallthru == 1 &&
1584 b->succ_fall_through(0)) {
1585 Block *n = b->_succs[0];
1586
1587 // Skip over single-entry connector blocks, we don't want to
1588 // add them to the trace.
1589 while (n->is_connector() && n->num_preds() == 1) {
1590 n = n->_succs[0];
1591 }
1592
1593 // We see a merge point, so stop search for the next block
1594 if (n->num_preds() != 1) break;
1595
1596 i++;
1597 assert(n == _cfg.get_block(i), "expecting next block");
1598 tr->append(n);
1599 uf->map(n->_pre_order, tr->id());
1600 traces[n->_pre_order] = nullptr;
1601 nfallthru = b->num_fall_throughs();
1602 b = n;
1603 }
1604
1605 if (nfallthru > 0) {
1606 // Create a CFGEdge for each outgoing
1607 // edge that could be a fall-through.
1608 for (uint j = 0; j < b->_num_succs; j++ ) {
1609 if (b->succ_fall_through(j)) {
1610 Block *target = b->non_connector_successor(j);
1611 float freq = b->_freq * b->succ_prob(j);
1612 int from_pct = (int) ((100 * freq) / b->_freq);
1613 float f_to_pct = (100 * freq) / target->_freq;
1614 int to_pct = (f_to_pct < 100.0) ? (int)f_to_pct : 100;
1615 edges->append(new CFGEdge(b, target, freq, from_pct, to_pct));
1616 }
1617 }
1618 }
1619 }
1620
1621 // Group connector blocks into one trace
1622 for (i++; i < _cfg.number_of_blocks(); i++) {
1623 Block *b = _cfg.get_block(i);
1624 assert(b->is_connector(), "connector blocks at the end");
1625 tr->append(b);
1626 uf->map(b->_pre_order, tr->id());
1627 traces[b->_pre_order] = nullptr;
1628 }
1629 }
1630
1631 // Union two traces together in uf, and null out the trace in the list
1632 void PhaseBlockLayout::union_traces(Trace* updated_trace, Trace* old_trace) {
1633 uint old_id = old_trace->id();
1634 uint updated_id = updated_trace->id();
1635
1636 uint lo_id = updated_id;
1637 uint hi_id = old_id;
1638
1639 // If from is greater than to, swap values to meet
1640 // UnionFind guarantee.
1641 if (updated_id > old_id) {
1642 lo_id = old_id;
1643 hi_id = updated_id;
1644
1645 // Fix up the trace ids
1646 traces[lo_id] = traces[updated_id];
1647 updated_trace->set_id(lo_id);
1648 }
1649
1650 // Union the lower with the higher and remove the pointer
1651 // to the higher.
1652 uf->Union(lo_id, hi_id);
1653 traces[hi_id] = nullptr;
1654 }
1655
1656 // Append traces together via the most frequently executed edges
1657 void PhaseBlockLayout::grow_traces() {
1658 // Order the edges, and drive the growth of Traces via the most
1659 // frequently executed edges.
1660 edges->sort(edge_order);
1661 for (int i = 0; i < edges->length(); i++) {
1662 CFGEdge *e = edges->at(i);
1663
1664 if (e->state() != CFGEdge::open) continue;
1665
1666 Block *src_block = e->from();
1667 Block *targ_block = e->to();
1668
1669 // Don't grow traces along backedges?
1670 if (!BlockLayoutRotateLoops) {
1671 if (targ_block->_rpo <= src_block->_rpo) {
1672 targ_block->set_loop_alignment(targ_block);
1673 continue;
1674 }
1675 }
1676
1677 Trace *src_trace = trace(src_block);
1678 Trace *targ_trace = trace(targ_block);
1679
1680 // If the edge in question can join two traces at their ends,
1681 // append one trace to the other.
1682 if (src_trace->last_block() == src_block) {
1683 if (src_trace == targ_trace) {
1684 e->set_state(CFGEdge::interior);
1685 if (targ_trace->backedge(e)) {
1686 // Reset i to catch any newly eligible edge
1687 // (Or we could remember the first "open" edge, and reset there)
1688 i = 0;
1689 }
1690 } else if (targ_trace->first_block() == targ_block) {
1691 e->set_state(CFGEdge::connected);
1692 src_trace->append(targ_trace);
1693 union_traces(src_trace, targ_trace);
1694 }
1695 }
1696 }
1697 }
1698
1699 // Embed one trace into another, if the fork or join points are sufficiently
1700 // balanced.
1701 void PhaseBlockLayout::merge_traces(bool fall_thru_only) {
1702 // Walk the edge list a another time, looking at unprocessed edges.
1703 // Fold in diamonds
1704 for (int i = 0; i < edges->length(); i++) {
1705 CFGEdge *e = edges->at(i);
1706
1707 if (e->state() != CFGEdge::open) continue;
1708 if (fall_thru_only) {
1709 if (e->infrequent()) continue;
1710 }
1711
1712 Block *src_block = e->from();
1713 Trace *src_trace = trace(src_block);
1714 bool src_at_tail = src_trace->last_block() == src_block;
1715
1716 Block *targ_block = e->to();
1717 Trace *targ_trace = trace(targ_block);
1718 bool targ_at_start = targ_trace->first_block() == targ_block;
1719
1720 if (src_trace == targ_trace) {
1721 // This may be a loop, but we can't do much about it.
1722 e->set_state(CFGEdge::interior);
1723 continue;
1724 }
1725
1726 if (fall_thru_only) {
1727 // If the edge links the middle of two traces, we can't do anything.
1728 // Mark the edge and continue.
1729 if (!src_at_tail & !targ_at_start) {
1730 continue;
1731 }
1732
1733 // Don't grow traces along backedges?
1734 if (!BlockLayoutRotateLoops && (targ_block->_rpo <= src_block->_rpo)) {
1735 continue;
1736 }
1737
1738 // If both ends of the edge are available, why didn't we handle it earlier?
1739 assert(src_at_tail ^ targ_at_start, "Should have caught this edge earlier.");
1740
1741 if (targ_at_start) {
1742 // Insert the "targ" trace in the "src" trace if the insertion point
1743 // is a two way branch.
1744 // Better profitability check possible, but may not be worth it.
1745 // Someday, see if the this "fork" has an associated "join";
1746 // then make a policy on merging this trace at the fork or join.
1747 // For example, other things being equal, it may be better to place this
1748 // trace at the join point if the "src" trace ends in a two-way, but
1749 // the insertion point is one-way.
1750 assert(src_block->num_fall_throughs() == 2, "unexpected diamond");
1751 e->set_state(CFGEdge::connected);
1752 src_trace->insert_after(src_block, targ_trace);
1753 union_traces(src_trace, targ_trace);
1754 } else if (src_at_tail) {
1755 if (src_trace != trace(_cfg.get_root_block())) {
1756 e->set_state(CFGEdge::connected);
1757 targ_trace->insert_before(targ_block, src_trace);
1758 union_traces(targ_trace, src_trace);
1759 }
1760 }
1761 } else if (e->state() == CFGEdge::open) {
1762 // Append traces, even without a fall-thru connection.
1763 // But leave root entry at the beginning of the block list.
1764 if (targ_trace != trace(_cfg.get_root_block())) {
1765 e->set_state(CFGEdge::connected);
1766 src_trace->append(targ_trace);
1767 union_traces(src_trace, targ_trace);
1768 }
1769 }
1770 }
1771 }
1772
1773 // Order the sequence of the traces in some desirable way
1774 void PhaseBlockLayout::reorder_traces(int count) {
1775 Trace** new_traces = NEW_RESOURCE_ARRAY(Trace*, count);
1776 Block_List worklist;
1777 int new_count = 0;
1778
1779 // Compact the traces.
1780 for (int i = 0; i < count; i++) {
1781 Trace* tr = traces[i];
1782 if (tr != nullptr) {
1783 new_traces[new_count++] = tr;
1784 }
1785 }
1786
1787 // The entry block should be first on the new trace list.
1788 Trace* tr = trace(_cfg.get_root_block());
1789 assert(tr == new_traces[0], "entry trace misplaced");
1790
1791 // Sort the new trace list by frequency
1792 qsort(new_traces + 1, new_count - 1, sizeof(new_traces[0]), trace_frequency_order);
1793
1794 // Collect all blocks from existing Traces
1795 _cfg.clear_blocks();
1796 for (int i = 0; i < new_count; i++) {
1797 Trace* tr = new_traces[i];
1798 if (tr != nullptr) {
1799 // push blocks onto the CFG list
1800 for (Block* b = tr->first_block(); b != nullptr; b = tr->next(b)) {
1801 _cfg.add_block(b);
1802 }
1803 }
1804 }
1805 }
1806
1807 // Order basic blocks based on frequency
1808 PhaseBlockLayout::PhaseBlockLayout(PhaseCFG &cfg)
1809 : Phase(BlockLayout)
1810 , _cfg(cfg) {
1811 ResourceMark rm;
1812
1813 // List of traces
1814 int size = _cfg.number_of_blocks() + 1;
1815 traces = NEW_RESOURCE_ARRAY(Trace*, size);
1816 memset(traces, 0, size*sizeof(Trace*));
1817 next = NEW_RESOURCE_ARRAY(Block*, size);
1818 memset(next, 0, size*sizeof(Block*));
1819 prev = NEW_RESOURCE_ARRAY(Block*, size);
1820 memset(prev , 0, size*sizeof(Block*));
1821
1822 // List of edges
1823 edges = new GrowableArray<CFGEdge*>;
1824
1825 // Mapping block index --> block_trace
1826 uf = new UnionFind(size);
1827 uf->reset(size);
1828
1829 // Find edges and create traces.
1830 find_edges();
1831
1832 // Grow traces at their ends via most frequent edges.
1833 grow_traces();
1834
1835 // Merge one trace into another, but only at fall-through points.
1836 // This may make diamonds and other related shapes in a trace.
1837 merge_traces(true);
1838
1839 // Run merge again, allowing two traces to be catenated, even if
1840 // one does not fall through into the other. This appends loosely
1841 // related traces to be near each other.
1842 merge_traces(false);
1843
1844 // Re-order all the remaining traces by frequency
1845 reorder_traces(size);
1846
1847 assert(_cfg.number_of_blocks() >= (uint) (size - 1), "number of blocks can not shrink");
1848 }
1849
1850
1851 // Edge e completes a loop in a trace. If the target block is head of the
1852 // loop, rotate the loop block so that the loop ends in a conditional branch.
1853 bool Trace::backedge(CFGEdge *e) {
1854 bool loop_rotated = false;
1855 Block *src_block = e->from();
1856 Block *targ_block = e->to();
1857
1858 assert(last_block() == src_block, "loop discovery at back branch");
1859 if (first_block() == targ_block) {
1860 if (BlockLayoutRotateLoops && last_block()->num_fall_throughs() < 2) {
1861 // Find the last block in the trace that has a conditional
1862 // branch.
1863 Block *b;
1864 for (b = last_block(); b != nullptr; b = prev(b)) {
1865 if (b->num_fall_throughs() == 2) {
1866 break;
1867 }
1868 }
1869
1870 if (b != last_block() && b != nullptr) {
1871 loop_rotated = true;
1872
1873 // Rotate the loop by doing two-part linked-list surgery.
1874 append(first_block());
1875 break_loop_after(b);
1876 }
1877 }
1878
1879 // Backbranch to the top of a trace
1880 // Scroll forward through the trace from the targ_block. If we find
1881 // a loop head before another loop top, use the loop head alignment.
1882 for (Block *b = targ_block; b != nullptr; b = next(b)) {
1883 if (b->has_loop_alignment()) {
1884 break;
1885 }
1886 if (b->head()->is_Loop()) {
1887 targ_block = b;
1888 break;
1889 }
1890 }
1891
1892 first_block()->set_loop_alignment(targ_block);
1893
1894 } else {
1895 // That loop may already have a loop top (we're reaching it again
1896 // through the backedge of an outer loop)
1897 Block* b = prev(targ_block);
1898 bool has_top = targ_block->head()->is_Loop() && b->has_loop_alignment() && !b->head()->is_Loop();
1899 if (!has_top) {
1900 // Backbranch into the middle of a trace
1901 targ_block->set_loop_alignment(targ_block);
1902 }
1903 }
1904
1905 return loop_rotated;
1906 }