1 /*
2 * Copyright (c) 1997, 2009, 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 // Optimization - Graph Style
26
27 #include "incls/_precompiled.incl"
28 #include "incls/_block.cpp.incl"
29
30
31 //-----------------------------------------------------------------------------
32 void Block_Array::grow( uint i ) {
33 assert(i >= Max(), "must be an overflow");
34 debug_only(_limit = i+1);
35 if( i < _size ) return;
36 if( !_size ) {
37 _size = 1;
38 _blocks = (Block**)_arena->Amalloc( _size * sizeof(Block*) );
39 _blocks[0] = NULL;
40 }
41 uint old = _size;
42 while( i >= _size ) _size <<= 1; // Double to fit
43 _blocks = (Block**)_arena->Arealloc( _blocks, old*sizeof(Block*),_size*sizeof(Block*));
44 Copy::zero_to_bytes( &_blocks[old], (_size-old)*sizeof(Block*) );
45 }
46
47 //=============================================================================
48 void Block_List::remove(uint i) {
49 assert(i < _cnt, "index out of bounds");
50 Copy::conjoint_words_to_lower((HeapWord*)&_blocks[i+1], (HeapWord*)&_blocks[i], ((_cnt-i-1)*sizeof(Block*)));
51 pop(); // shrink list by one block
52 }
53
54 void Block_List::insert(uint i, Block *b) {
55 push(b); // grow list by one block
56 Copy::conjoint_words_to_higher((HeapWord*)&_blocks[i], (HeapWord*)&_blocks[i+1], ((_cnt-i-1)*sizeof(Block*)));
57 _blocks[i] = b;
58 }
59
60 #ifndef PRODUCT
61 void Block_List::print() {
62 for (uint i=0; i < size(); i++) {
63 tty->print("B%d ", _blocks[i]->_pre_order);
64 }
65 tty->print("size = %d\n", size());
66 }
67 #endif
68
69 //=============================================================================
70
71 uint Block::code_alignment() {
72 // Check for Root block
73 if( _pre_order == 0 ) return CodeEntryAlignment;
74 // Check for Start block
75 if( _pre_order == 1 ) return InteriorEntryAlignment;
76 // Check for loop alignment
77 if (has_loop_alignment()) return loop_alignment();
78
79 return 1; // no particular alignment
80 }
81
82 uint Block::compute_loop_alignment() {
83 Node *h = head();
84 if( h->is_Loop() && h->as_Loop()->is_inner_loop() ) {
85 // Pre- and post-loops have low trip count so do not bother with
86 // NOPs for align loop head. The constants are hidden from tuning
87 // but only because my "divide by 4" heuristic surely gets nearly
88 // all possible gain (a "do not align at all" heuristic has a
89 // chance of getting a really tiny gain).
90 if( h->is_CountedLoop() && (h->as_CountedLoop()->is_pre_loop() ||
91 h->as_CountedLoop()->is_post_loop()) )
92 return (OptoLoopAlignment > 4) ? (OptoLoopAlignment>>2) : 1;
93 // Loops with low backedge frequency should not be aligned.
94 Node *n = h->in(LoopNode::LoopBackControl)->in(0);
95 if( n->is_MachIf() && n->as_MachIf()->_prob < 0.01 ) {
96 return 1; // Loop does not loop, more often than not!
97 }
98 return OptoLoopAlignment; // Otherwise align loop head
99 }
100
101 return 1; // no particular alignment
102 }
103
104 //-----------------------------------------------------------------------------
105 // Compute the size of first 'inst_cnt' instructions in this block.
106 // Return the number of instructions left to compute if the block has
107 // less then 'inst_cnt' instructions. Stop, and return 0 if sum_size
108 // exceeds OptoLoopAlignment.
109 uint Block::compute_first_inst_size(uint& sum_size, uint inst_cnt,
110 PhaseRegAlloc* ra) {
111 uint last_inst = _nodes.size();
112 for( uint j = 0; j < last_inst && inst_cnt > 0; j++ ) {
113 uint inst_size = _nodes[j]->size(ra);
114 if( inst_size > 0 ) {
115 inst_cnt--;
116 uint sz = sum_size + inst_size;
117 if( sz <= (uint)OptoLoopAlignment ) {
118 // Compute size of instructions which fit into fetch buffer only
119 // since all inst_cnt instructions will not fit even if we align them.
120 sum_size = sz;
121 } else {
122 return 0;
123 }
124 }
125 }
126 return inst_cnt;
127 }
128
129 //-----------------------------------------------------------------------------
130 uint Block::find_node( const Node *n ) const {
131 for( uint i = 0; i < _nodes.size(); i++ ) {
132 if( _nodes[i] == n )
133 return i;
134 }
135 ShouldNotReachHere();
136 return 0;
137 }
138
139 // Find and remove n from block list
140 void Block::find_remove( const Node *n ) {
141 _nodes.remove(find_node(n));
142 }
143
144 //------------------------------is_Empty---------------------------------------
145 // Return empty status of a block. Empty blocks contain only the head, other
146 // ideal nodes, and an optional trailing goto.
147 int Block::is_Empty() const {
148
149 // Root or start block is not considered empty
150 if (head()->is_Root() || head()->is_Start()) {
151 return not_empty;
152 }
153
154 int success_result = completely_empty;
155 int end_idx = _nodes.size()-1;
156
157 // Check for ending goto
158 if ((end_idx > 0) && (_nodes[end_idx]->is_Goto())) {
159 success_result = empty_with_goto;
160 end_idx--;
161 }
162
163 // Unreachable blocks are considered empty
164 if (num_preds() <= 1) {
165 return success_result;
166 }
167
168 // Ideal nodes are allowable in empty blocks: skip them Only MachNodes
169 // turn directly into code, because only MachNodes have non-trivial
170 // emit() functions.
171 while ((end_idx > 0) && !_nodes[end_idx]->is_Mach()) {
172 end_idx--;
173 }
174
175 // No room for any interesting instructions?
176 if (end_idx == 0) {
177 return success_result;
178 }
179
180 return not_empty;
181 }
182
183 //------------------------------has_uncommon_code------------------------------
184 // Return true if the block's code implies that it is likely to be
185 // executed infrequently. Check to see if the block ends in a Halt or
186 // a low probability call.
187 bool Block::has_uncommon_code() const {
188 Node* en = end();
189
190 if (en->is_Goto())
191 en = en->in(0);
192 if (en->is_Catch())
193 en = en->in(0);
194 if (en->is_Proj() && en->in(0)->is_MachCall()) {
195 MachCallNode* call = en->in(0)->as_MachCall();
196 if (call->cnt() != COUNT_UNKNOWN && call->cnt() <= PROB_UNLIKELY_MAG(4)) {
197 // This is true for slow-path stubs like new_{instance,array},
198 // slow_arraycopy, complete_monitor_locking, uncommon_trap.
199 // The magic number corresponds to the probability of an uncommon_trap,
200 // even though it is a count not a probability.
201 return true;
202 }
203 }
204
205 int op = en->is_Mach() ? en->as_Mach()->ideal_Opcode() : en->Opcode();
206 return op == Op_Halt;
207 }
208
209 //------------------------------is_uncommon------------------------------------
210 // True if block is low enough frequency or guarded by a test which
211 // mostly does not go here.
212 bool Block::is_uncommon( Block_Array &bbs ) const {
213 // Initial blocks must never be moved, so are never uncommon.
214 if (head()->is_Root() || head()->is_Start()) return false;
215
216 // Check for way-low freq
217 if( _freq < BLOCK_FREQUENCY(0.00001f) ) return true;
218
219 // Look for code shape indicating uncommon_trap or slow path
220 if (has_uncommon_code()) return true;
221
222 const float epsilon = 0.05f;
223 const float guard_factor = PROB_UNLIKELY_MAG(4) / (1.f - epsilon);
224 uint uncommon_preds = 0;
225 uint freq_preds = 0;
226 uint uncommon_for_freq_preds = 0;
227
228 for( uint i=1; i<num_preds(); i++ ) {
229 Block* guard = bbs[pred(i)->_idx];
230 // Check to see if this block follows its guard 1 time out of 10000
231 // or less.
232 //
233 // See list of magnitude-4 unlikely probabilities in cfgnode.hpp which
234 // we intend to be "uncommon", such as slow-path TLE allocation,
235 // predicted call failure, and uncommon trap triggers.
236 //
237 // Use an epsilon value of 5% to allow for variability in frequency
238 // predictions and floating point calculations. The net effect is
239 // that guard_factor is set to 9500.
240 //
241 // Ignore low-frequency blocks.
242 // The next check is (guard->_freq < 1.e-5 * 9500.).
243 if(guard->_freq*BLOCK_FREQUENCY(guard_factor) < BLOCK_FREQUENCY(0.00001f)) {
244 uncommon_preds++;
245 } else {
246 freq_preds++;
247 if( _freq < guard->_freq * guard_factor ) {
248 uncommon_for_freq_preds++;
249 }
250 }
251 }
252 if( num_preds() > 1 &&
253 // The block is uncommon if all preds are uncommon or
254 (uncommon_preds == (num_preds()-1) ||
255 // it is uncommon for all frequent preds.
256 uncommon_for_freq_preds == freq_preds) ) {
257 return true;
258 }
259 return false;
260 }
261
262 //------------------------------dump-------------------------------------------
263 #ifndef PRODUCT
264 void Block::dump_bidx(const Block* orig) const {
265 if (_pre_order) tty->print("B%d",_pre_order);
266 else tty->print("N%d", head()->_idx);
267
268 if (Verbose && orig != this) {
269 // Dump the original block's idx
270 tty->print(" (");
271 orig->dump_bidx(orig);
272 tty->print(")");
273 }
274 }
275
276 void Block::dump_pred(const Block_Array *bbs, Block* orig) const {
277 if (is_connector()) {
278 for (uint i=1; i<num_preds(); i++) {
279 Block *p = ((*bbs)[pred(i)->_idx]);
280 p->dump_pred(bbs, orig);
281 }
282 } else {
283 dump_bidx(orig);
284 tty->print(" ");
285 }
286 }
287
288 void Block::dump_head( const Block_Array *bbs ) const {
289 // Print the basic block
290 dump_bidx(this);
291 tty->print(": #\t");
292
293 // Print the incoming CFG edges and the outgoing CFG edges
294 for( uint i=0; i<_num_succs; i++ ) {
295 non_connector_successor(i)->dump_bidx(_succs[i]);
296 tty->print(" ");
297 }
298 tty->print("<- ");
299 if( head()->is_block_start() ) {
300 for (uint i=1; i<num_preds(); i++) {
301 Node *s = pred(i);
302 if (bbs) {
303 Block *p = (*bbs)[s->_idx];
304 p->dump_pred(bbs, p);
305 } else {
306 while (!s->is_block_start())
307 s = s->in(0);
308 tty->print("N%d ", s->_idx );
309 }
310 }
311 } else
312 tty->print("BLOCK HEAD IS JUNK ");
313
314 // Print loop, if any
315 const Block *bhead = this; // Head of self-loop
316 Node *bh = bhead->head();
317 if( bbs && bh->is_Loop() && !head()->is_Root() ) {
318 LoopNode *loop = bh->as_Loop();
319 const Block *bx = (*bbs)[loop->in(LoopNode::LoopBackControl)->_idx];
320 while (bx->is_connector()) {
321 bx = (*bbs)[bx->pred(1)->_idx];
322 }
323 tty->print("\tLoop: B%d-B%d ", bhead->_pre_order, bx->_pre_order);
324 // Dump any loop-specific bits, especially for CountedLoops.
325 loop->dump_spec(tty);
326 } else if (has_loop_alignment()) {
327 tty->print(" top-of-loop");
328 }
329 tty->print(" Freq: %g",_freq);
330 if( Verbose || WizardMode ) {
331 tty->print(" IDom: %d/#%d", _idom ? _idom->_pre_order : 0, _dom_depth);
332 tty->print(" RegPressure: %d",_reg_pressure);
333 tty->print(" IHRP Index: %d",_ihrp_index);
334 tty->print(" FRegPressure: %d",_freg_pressure);
335 tty->print(" FHRP Index: %d",_fhrp_index);
336 }
337 tty->print_cr("");
338 }
339
340 void Block::dump() const { dump(0); }
341
342 void Block::dump( const Block_Array *bbs ) const {
343 dump_head(bbs);
344 uint cnt = _nodes.size();
345 for( uint i=0; i<cnt; i++ )
346 _nodes[i]->dump();
347 tty->print("\n");
348 }
349 #endif
350
351 //=============================================================================
352 //------------------------------PhaseCFG---------------------------------------
353 PhaseCFG::PhaseCFG( Arena *a, RootNode *r, Matcher &m ) :
354 Phase(CFG),
355 _bbs(a),
356 _root(r),
357 _node_latency(NULL)
358 #ifndef PRODUCT
359 , _trace_opto_pipelining(TraceOptoPipelining || C->method_has_option("TraceOptoPipelining"))
360 #endif
361 #ifdef ASSERT
362 , _raw_oops(a)
363 #endif
364 {
365 ResourceMark rm;
366 // I'll need a few machine-specific GotoNodes. Make an Ideal GotoNode,
367 // then Match it into a machine-specific Node. Then clone the machine
368 // Node on demand.
369 Node *x = new (C, 1) GotoNode(NULL);
370 x->init_req(0, x);
371 _goto = m.match_tree(x);
372 assert(_goto != NULL, "");
373 _goto->set_req(0,_goto);
374
375 // Build the CFG in Reverse Post Order
376 _num_blocks = build_cfg();
377 _broot = _bbs[_root->_idx];
378 }
379
380 //------------------------------build_cfg--------------------------------------
381 // Build a proper looking CFG. Make every block begin with either a StartNode
382 // or a RegionNode. Make every block end with either a Goto, If or Return.
383 // The RootNode both starts and ends it's own block. Do this with a recursive
384 // backwards walk over the control edges.
385 uint PhaseCFG::build_cfg() {
386 Arena *a = Thread::current()->resource_area();
387 VectorSet visited(a);
388
389 // Allocate stack with enough space to avoid frequent realloc
390 Node_Stack nstack(a, C->unique() >> 1);
391 nstack.push(_root, 0);
392 uint sum = 0; // Counter for blocks
393
394 while (nstack.is_nonempty()) {
395 // node and in's index from stack's top
396 // 'np' is _root (see above) or RegionNode, StartNode: we push on stack
397 // only nodes which point to the start of basic block (see below).
398 Node *np = nstack.node();
399 // idx > 0, except for the first node (_root) pushed on stack
400 // at the beginning when idx == 0.
401 // We will use the condition (idx == 0) later to end the build.
402 uint idx = nstack.index();
403 Node *proj = np->in(idx);
404 const Node *x = proj->is_block_proj();
405 // Does the block end with a proper block-ending Node? One of Return,
406 // If or Goto? (This check should be done for visited nodes also).
407 if (x == NULL) { // Does not end right...
408 Node *g = _goto->clone(); // Force it to end in a Goto
409 g->set_req(0, proj);
410 np->set_req(idx, g);
411 x = proj = g;
412 }
413 if (!visited.test_set(x->_idx)) { // Visit this block once
414 // Skip any control-pinned middle'in stuff
415 Node *p = proj;
416 do {
417 proj = p; // Update pointer to last Control
418 p = p->in(0); // Move control forward
419 } while( !p->is_block_proj() &&
420 !p->is_block_start() );
421 // Make the block begin with one of Region or StartNode.
422 if( !p->is_block_start() ) {
423 RegionNode *r = new (C, 2) RegionNode( 2 );
424 r->init_req(1, p); // Insert RegionNode in the way
425 proj->set_req(0, r); // Insert RegionNode in the way
426 p = r;
427 }
428 // 'p' now points to the start of this basic block
429
430 // Put self in array of basic blocks
431 Block *bb = new (_bbs._arena) Block(_bbs._arena,p);
432 _bbs.map(p->_idx,bb);
433 _bbs.map(x->_idx,bb);
434 if( x != p ) // Only for root is x == p
435 bb->_nodes.push((Node*)x);
436
437 // Now handle predecessors
438 ++sum; // Count 1 for self block
439 uint cnt = bb->num_preds();
440 for (int i = (cnt - 1); i > 0; i-- ) { // For all predecessors
441 Node *prevproj = p->in(i); // Get prior input
442 assert( !prevproj->is_Con(), "dead input not removed" );
443 // Check to see if p->in(i) is a "control-dependent" CFG edge -
444 // i.e., it splits at the source (via an IF or SWITCH) and merges
445 // at the destination (via a many-input Region).
446 // This breaks critical edges. The RegionNode to start the block
447 // will be added when <p,i> is pulled off the node stack
448 if ( cnt > 2 ) { // Merging many things?
449 assert( prevproj== bb->pred(i),"");
450 if(prevproj->is_block_proj() != prevproj) { // Control-dependent edge?
451 // Force a block on the control-dependent edge
452 Node *g = _goto->clone(); // Force it to end in a Goto
453 g->set_req(0,prevproj);
454 p->set_req(i,g);
455 }
456 }
457 nstack.push(p, i); // 'p' is RegionNode or StartNode
458 }
459 } else { // Post-processing visited nodes
460 nstack.pop(); // remove node from stack
461 // Check if it the fist node pushed on stack at the beginning.
462 if (idx == 0) break; // end of the build
463 // Find predecessor basic block
464 Block *pb = _bbs[x->_idx];
465 // Insert into nodes array, if not already there
466 if( !_bbs.lookup(proj->_idx) ) {
467 assert( x != proj, "" );
468 // Map basic block of projection
469 _bbs.map(proj->_idx,pb);
470 pb->_nodes.push(proj);
471 }
472 // Insert self as a child of my predecessor block
473 pb->_succs.map(pb->_num_succs++, _bbs[np->_idx]);
474 assert( pb->_nodes[ pb->_nodes.size() - pb->_num_succs ]->is_block_proj(),
475 "too many control users, not a CFG?" );
476 }
477 }
478 // Return number of basic blocks for all children and self
479 return sum;
480 }
481
482 //------------------------------insert_goto_at---------------------------------
483 // Inserts a goto & corresponding basic block between
484 // block[block_no] and its succ_no'th successor block
485 void PhaseCFG::insert_goto_at(uint block_no, uint succ_no) {
486 // get block with block_no
487 assert(block_no < _num_blocks, "illegal block number");
488 Block* in = _blocks[block_no];
489 // get successor block succ_no
490 assert(succ_no < in->_num_succs, "illegal successor number");
491 Block* out = in->_succs[succ_no];
492 // Compute frequency of the new block. Do this before inserting
493 // new block in case succ_prob() needs to infer the probability from
494 // surrounding blocks.
495 float freq = in->_freq * in->succ_prob(succ_no);
496 // get ProjNode corresponding to the succ_no'th successor of the in block
497 ProjNode* proj = in->_nodes[in->_nodes.size() - in->_num_succs + succ_no]->as_Proj();
498 // create region for basic block
499 RegionNode* region = new (C, 2) RegionNode(2);
500 region->init_req(1, proj);
501 // setup corresponding basic block
502 Block* block = new (_bbs._arena) Block(_bbs._arena, region);
503 _bbs.map(region->_idx, block);
504 C->regalloc()->set_bad(region->_idx);
505 // add a goto node
506 Node* gto = _goto->clone(); // get a new goto node
507 gto->set_req(0, region);
508 // add it to the basic block
509 block->_nodes.push(gto);
510 _bbs.map(gto->_idx, block);
511 C->regalloc()->set_bad(gto->_idx);
512 // hook up successor block
513 block->_succs.map(block->_num_succs++, out);
514 // remap successor's predecessors if necessary
515 for (uint i = 1; i < out->num_preds(); i++) {
516 if (out->pred(i) == proj) out->head()->set_req(i, gto);
517 }
518 // remap predecessor's successor to new block
519 in->_succs.map(succ_no, block);
520 // Set the frequency of the new block
521 block->_freq = freq;
522 // add new basic block to basic block list
523 _blocks.insert(block_no + 1, block);
524 _num_blocks++;
525 }
526
527 //------------------------------no_flip_branch---------------------------------
528 // Does this block end in a multiway branch that cannot have the default case
529 // flipped for another case?
530 static bool no_flip_branch( Block *b ) {
531 int branch_idx = b->_nodes.size() - b->_num_succs-1;
532 if( branch_idx < 1 ) return false;
533 Node *bra = b->_nodes[branch_idx];
534 if( bra->is_Catch() )
535 return true;
536 if( bra->is_Mach() ) {
537 if( bra->is_MachNullCheck() )
538 return true;
539 int iop = bra->as_Mach()->ideal_Opcode();
540 if( iop == Op_FastLock || iop == Op_FastUnlock )
541 return true;
542 }
543 return false;
544 }
545
546 //------------------------------convert_NeverBranch_to_Goto--------------------
547 // Check for NeverBranch at block end. This needs to become a GOTO to the
548 // true target. NeverBranch are treated as a conditional branch that always
549 // goes the same direction for most of the optimizer and are used to give a
550 // fake exit path to infinite loops. At this late stage they need to turn
551 // into Goto's so that when you enter the infinite loop you indeed hang.
552 void PhaseCFG::convert_NeverBranch_to_Goto(Block *b) {
553 // Find true target
554 int end_idx = b->end_idx();
555 int idx = b->_nodes[end_idx+1]->as_Proj()->_con;
556 Block *succ = b->_succs[idx];
557 Node* gto = _goto->clone(); // get a new goto node
558 gto->set_req(0, b->head());
559 Node *bp = b->_nodes[end_idx];
560 b->_nodes.map(end_idx,gto); // Slam over NeverBranch
561 _bbs.map(gto->_idx, b);
562 C->regalloc()->set_bad(gto->_idx);
563 b->_nodes.pop(); // Yank projections
564 b->_nodes.pop(); // Yank projections
565 b->_succs.map(0,succ); // Map only successor
566 b->_num_succs = 1;
567 // remap successor's predecessors if necessary
568 uint j;
569 for( j = 1; j < succ->num_preds(); j++)
570 if( succ->pred(j)->in(0) == bp )
571 succ->head()->set_req(j, gto);
572 // Kill alternate exit path
573 Block *dead = b->_succs[1-idx];
574 for( j = 1; j < dead->num_preds(); j++)
575 if( dead->pred(j)->in(0) == bp )
576 break;
577 // Scan through block, yanking dead path from
578 // all regions and phis.
579 dead->head()->del_req(j);
580 for( int k = 1; dead->_nodes[k]->is_Phi(); k++ )
581 dead->_nodes[k]->del_req(j);
582 }
583
584 //------------------------------move_to_next-----------------------------------
585 // Helper function to move block bx to the slot following b_index. Return
586 // true if the move is successful, otherwise false
587 bool PhaseCFG::move_to_next(Block* bx, uint b_index) {
588 if (bx == NULL) return false;
589
590 // Return false if bx is already scheduled.
591 uint bx_index = bx->_pre_order;
592 if ((bx_index <= b_index) && (_blocks[bx_index] == bx)) {
593 return false;
594 }
595
596 // Find the current index of block bx on the block list
597 bx_index = b_index + 1;
598 while( bx_index < _num_blocks && _blocks[bx_index] != bx ) bx_index++;
599 assert(_blocks[bx_index] == bx, "block not found");
600
601 // If the previous block conditionally falls into bx, return false,
602 // because moving bx will create an extra jump.
603 for(uint k = 1; k < bx->num_preds(); k++ ) {
604 Block* pred = _bbs[bx->pred(k)->_idx];
605 if (pred == _blocks[bx_index-1]) {
606 if (pred->_num_succs != 1) {
607 return false;
608 }
609 }
610 }
611
612 // Reinsert bx just past block 'b'
613 _blocks.remove(bx_index);
614 _blocks.insert(b_index + 1, bx);
615 return true;
616 }
617
618 //------------------------------move_to_end------------------------------------
619 // Move empty and uncommon blocks to the end.
620 void PhaseCFG::move_to_end(Block *b, uint i) {
621 int e = b->is_Empty();
622 if (e != Block::not_empty) {
623 if (e == Block::empty_with_goto) {
624 // Remove the goto, but leave the block.
625 b->_nodes.pop();
626 }
627 // Mark this block as a connector block, which will cause it to be
628 // ignored in certain functions such as non_connector_successor().
629 b->set_connector();
630 }
631 // Move the empty block to the end, and don't recheck.
632 _blocks.remove(i);
633 _blocks.push(b);
634 }
635
636 //---------------------------set_loop_alignment--------------------------------
637 // Set loop alignment for every block
638 void PhaseCFG::set_loop_alignment() {
639 uint last = _num_blocks;
640 assert( _blocks[0] == _broot, "" );
641
642 for (uint i = 1; i < last; i++ ) {
643 Block *b = _blocks[i];
644 if (b->head()->is_Loop()) {
645 b->set_loop_alignment(b);
646 }
647 }
648 }
649
650 //-----------------------------remove_empty------------------------------------
651 // Make empty basic blocks to be "connector" blocks, Move uncommon blocks
652 // to the end.
653 void PhaseCFG::remove_empty() {
654 // Move uncommon blocks to the end
655 uint last = _num_blocks;
656 assert( _blocks[0] == _broot, "" );
657
658 for (uint i = 1; i < last; i++) {
659 Block *b = _blocks[i];
660 if (b->is_connector()) break;
661
662 // Check for NeverBranch at block end. This needs to become a GOTO to the
663 // true target. NeverBranch are treated as a conditional branch that
664 // always goes the same direction for most of the optimizer and are used
665 // to give a fake exit path to infinite loops. At this late stage they
666 // need to turn into Goto's so that when you enter the infinite loop you
667 // indeed hang.
668 if( b->_nodes[b->end_idx()]->Opcode() == Op_NeverBranch )
669 convert_NeverBranch_to_Goto(b);
670
671 // Look for uncommon blocks and move to end.
672 if (!C->do_freq_based_layout()) {
673 if( b->is_uncommon(_bbs) ) {
674 move_to_end(b, i);
675 last--; // No longer check for being uncommon!
676 if( no_flip_branch(b) ) { // Fall-thru case must follow?
677 b = _blocks[i]; // Find the fall-thru block
678 move_to_end(b, i);
679 last--;
680 }
681 i--; // backup block counter post-increment
682 }
683 }
684 }
685
686 // Move empty blocks to the end
687 last = _num_blocks;
688 for (uint i = 1; i < last; i++) {
689 Block *b = _blocks[i];
690 if (b->is_Empty() != Block::not_empty) {
691 move_to_end(b, i);
692 last--;
693 i--;
694 }
695 } // End of for all blocks
696 }
697
698 //-----------------------------fixup_flow--------------------------------------
699 // Fix up the final control flow for basic blocks.
700 void PhaseCFG::fixup_flow() {
701 // Fixup final control flow for the blocks. Remove jump-to-next
702 // block. If neither arm of a IF follows the conditional branch, we
703 // have to add a second jump after the conditional. We place the
704 // TRUE branch target in succs[0] for both GOTOs and IFs.
705 for (uint i=0; i < _num_blocks; i++) {
706 Block *b = _blocks[i];
707 b->_pre_order = i; // turn pre-order into block-index
708
709 // Connector blocks need no further processing.
710 if (b->is_connector()) {
711 assert((i+1) == _num_blocks || _blocks[i+1]->is_connector(),
712 "All connector blocks should sink to the end");
713 continue;
714 }
715 assert(b->is_Empty() != Block::completely_empty,
716 "Empty blocks should be connectors");
717
718 Block *bnext = (i < _num_blocks-1) ? _blocks[i+1] : NULL;
719 Block *bs0 = b->non_connector_successor(0);
720
721 // Check for multi-way branches where I cannot negate the test to
722 // exchange the true and false targets.
723 if( no_flip_branch( b ) ) {
724 // Find fall through case - if must fall into its target
725 int branch_idx = b->_nodes.size() - b->_num_succs;
726 for (uint j2 = 0; j2 < b->_num_succs; j2++) {
727 const ProjNode* p = b->_nodes[branch_idx + j2]->as_Proj();
728 if (p->_con == 0) {
729 // successor j2 is fall through case
730 if (b->non_connector_successor(j2) != bnext) {
731 // but it is not the next block => insert a goto
732 insert_goto_at(i, j2);
733 }
734 // Put taken branch in slot 0
735 if( j2 == 0 && b->_num_succs == 2) {
736 // Flip targets in succs map
737 Block *tbs0 = b->_succs[0];
738 Block *tbs1 = b->_succs[1];
739 b->_succs.map( 0, tbs1 );
740 b->_succs.map( 1, tbs0 );
741 }
742 break;
743 }
744 }
745 // Remove all CatchProjs
746 for (uint j1 = 0; j1 < b->_num_succs; j1++) b->_nodes.pop();
747
748 } else if (b->_num_succs == 1) {
749 // Block ends in a Goto?
750 if (bnext == bs0) {
751 // We fall into next block; remove the Goto
752 b->_nodes.pop();
753 }
754
755 } else if( b->_num_succs == 2 ) { // Block ends in a If?
756 // Get opcode of 1st projection (matches _succs[0])
757 // Note: Since this basic block has 2 exits, the last 2 nodes must
758 // be projections (in any order), the 3rd last node must be
759 // the IfNode (we have excluded other 2-way exits such as
760 // CatchNodes already).
761 MachNode *iff = b->_nodes[b->_nodes.size()-3]->as_Mach();
762 ProjNode *proj0 = b->_nodes[b->_nodes.size()-2]->as_Proj();
763 ProjNode *proj1 = b->_nodes[b->_nodes.size()-1]->as_Proj();
764
765 // Assert that proj0 and succs[0] match up. Similarly for proj1 and succs[1].
766 assert(proj0->raw_out(0) == b->_succs[0]->head(), "Mismatch successor 0");
767 assert(proj1->raw_out(0) == b->_succs[1]->head(), "Mismatch successor 1");
768
769 Block *bs1 = b->non_connector_successor(1);
770
771 // Check for neither successor block following the current
772 // block ending in a conditional. If so, move one of the
773 // successors after the current one, provided that the
774 // successor was previously unscheduled, but moveable
775 // (i.e., all paths to it involve a branch).
776 if( !C->do_freq_based_layout() && bnext != bs0 && bnext != bs1 ) {
777 // Choose the more common successor based on the probability
778 // of the conditional branch.
779 Block *bx = bs0;
780 Block *by = bs1;
781
782 // _prob is the probability of taking the true path. Make
783 // p the probability of taking successor #1.
784 float p = iff->as_MachIf()->_prob;
785 if( proj0->Opcode() == Op_IfTrue ) {
786 p = 1.0 - p;
787 }
788
789 // Prefer successor #1 if p > 0.5
790 if (p > PROB_FAIR) {
791 bx = bs1;
792 by = bs0;
793 }
794
795 // Attempt the more common successor first
796 if (move_to_next(bx, i)) {
797 bnext = bx;
798 } else if (move_to_next(by, i)) {
799 bnext = by;
800 }
801 }
802
803 // Check for conditional branching the wrong way. Negate
804 // conditional, if needed, so it falls into the following block
805 // and branches to the not-following block.
806
807 // Check for the next block being in succs[0]. We are going to branch
808 // to succs[0], so we want the fall-thru case as the next block in
809 // succs[1].
810 if (bnext == bs0) {
811 // Fall-thru case in succs[0], so flip targets in succs map
812 Block *tbs0 = b->_succs[0];
813 Block *tbs1 = b->_succs[1];
814 b->_succs.map( 0, tbs1 );
815 b->_succs.map( 1, tbs0 );
816 // Flip projection for each target
817 { ProjNode *tmp = proj0; proj0 = proj1; proj1 = tmp; }
818
819 } else if( bnext != bs1 ) {
820 // Need a double-branch
821 // The existing conditional branch need not change.
822 // Add a unconditional branch to the false target.
823 // Alas, it must appear in its own block and adding a
824 // block this late in the game is complicated. Sigh.
825 insert_goto_at(i, 1);
826 }
827
828 // Make sure we TRUE branch to the target
829 if( proj0->Opcode() == Op_IfFalse ) {
830 iff->negate();
831 }
832
833 b->_nodes.pop(); // Remove IfFalse & IfTrue projections
834 b->_nodes.pop();
835
836 } else {
837 // Multi-exit block, e.g. a switch statement
838 // But we don't need to do anything here
839 }
840 } // End of for all blocks
841 }
842
843
844 //------------------------------dump-------------------------------------------
845 #ifndef PRODUCT
846 void PhaseCFG::_dump_cfg( const Node *end, VectorSet &visited ) const {
847 const Node *x = end->is_block_proj();
848 assert( x, "not a CFG" );
849
850 // Do not visit this block again
851 if( visited.test_set(x->_idx) ) return;
852
853 // Skip through this block
854 const Node *p = x;
855 do {
856 p = p->in(0); // Move control forward
857 assert( !p->is_block_proj() || p->is_Root(), "not a CFG" );
858 } while( !p->is_block_start() );
859
860 // Recursively visit
861 for( uint i=1; i<p->req(); i++ )
862 _dump_cfg(p->in(i),visited);
863
864 // Dump the block
865 _bbs[p->_idx]->dump(&_bbs);
866 }
867
868 void PhaseCFG::dump( ) const {
869 tty->print("\n--- CFG --- %d BBs\n",_num_blocks);
870 if( _blocks.size() ) { // Did we do basic-block layout?
871 for( uint i=0; i<_num_blocks; i++ )
872 _blocks[i]->dump(&_bbs);
873 } else { // Else do it with a DFS
874 VectorSet visited(_bbs._arena);
875 _dump_cfg(_root,visited);
876 }
877 }
878
879 void PhaseCFG::dump_headers() {
880 for( uint i = 0; i < _num_blocks; i++ ) {
881 if( _blocks[i] == NULL ) continue;
882 _blocks[i]->dump_head(&_bbs);
883 }
884 }
885
886 void PhaseCFG::verify( ) const {
887 #ifdef ASSERT
888 // Verify sane CFG
889 for( uint i = 0; i < _num_blocks; i++ ) {
890 Block *b = _blocks[i];
891 uint cnt = b->_nodes.size();
892 uint j;
893 for( j = 0; j < cnt; j++ ) {
894 Node *n = b->_nodes[j];
895 assert( _bbs[n->_idx] == b, "" );
896 if( j >= 1 && n->is_Mach() &&
897 n->as_Mach()->ideal_Opcode() == Op_CreateEx ) {
898 assert( j == 1 || b->_nodes[j-1]->is_Phi(),
899 "CreateEx must be first instruction in block" );
900 }
901 for( uint k = 0; k < n->req(); k++ ) {
902 Node *def = n->in(k);
903 if( def && def != n ) {
904 assert( _bbs[def->_idx] || def->is_Con(),
905 "must have block; constants for debug info ok" );
906 // Verify that instructions in the block is in correct order.
907 // Uses must follow their definition if they are at the same block.
908 // Mostly done to check that MachSpillCopy nodes are placed correctly
909 // when CreateEx node is moved in build_ifg_physical().
910 if( _bbs[def->_idx] == b &&
911 !(b->head()->is_Loop() && n->is_Phi()) &&
912 // See (+++) comment in reg_split.cpp
913 !(n->jvms() != NULL && n->jvms()->is_monitor_use(k)) ) {
914 bool is_loop = false;
915 if (n->is_Phi()) {
916 for( uint l = 1; l < def->req(); l++ ) {
917 if (n == def->in(l)) {
918 is_loop = true;
919 break; // Some kind of loop
920 }
921 }
922 }
923 assert( is_loop || b->find_node(def) < j, "uses must follow definitions" );
924 }
925 if( def->is_SafePointScalarObject() ) {
926 assert(_bbs[def->_idx] == b, "SafePointScalarObject Node should be at the same block as its SafePoint node");
927 assert(_bbs[def->_idx] == _bbs[def->in(0)->_idx], "SafePointScalarObject Node should be at the same block as its control edge");
928 }
929 }
930 }
931 }
932
933 j = b->end_idx();
934 Node *bp = (Node*)b->_nodes[b->_nodes.size()-1]->is_block_proj();
935 assert( bp, "last instruction must be a block proj" );
936 assert( bp == b->_nodes[j], "wrong number of successors for this block" );
937 if( bp->is_Catch() ) {
938 while( b->_nodes[--j]->Opcode() == Op_MachProj ) ;
939 assert( b->_nodes[j]->is_Call(), "CatchProj must follow call" );
940 }
941 else if( bp->is_Mach() && bp->as_Mach()->ideal_Opcode() == Op_If ) {
942 assert( b->_num_succs == 2, "Conditional branch must have two targets");
943 }
944 }
945 #endif
946 }
947 #endif
948
949 //=============================================================================
950 //------------------------------UnionFind--------------------------------------
951 UnionFind::UnionFind( uint max ) : _cnt(max), _max(max), _indices(NEW_RESOURCE_ARRAY(uint,max)) {
952 Copy::zero_to_bytes( _indices, sizeof(uint)*max );
953 }
954
955 void UnionFind::extend( uint from_idx, uint to_idx ) {
956 _nesting.check();
957 if( from_idx >= _max ) {
958 uint size = 16;
959 while( size <= from_idx ) size <<=1;
960 _indices = REALLOC_RESOURCE_ARRAY( uint, _indices, _max, size );
961 _max = size;
962 }
963 while( _cnt <= from_idx ) _indices[_cnt++] = 0;
964 _indices[from_idx] = to_idx;
965 }
966
967 void UnionFind::reset( uint max ) {
968 assert( max <= max_uint, "Must fit within uint" );
969 // Force the Union-Find mapping to be at least this large
970 extend(max,0);
971 // Initialize to be the ID mapping.
972 for( uint i=0; i<max; i++ ) map(i,i);
973 }
974
975 //------------------------------Find_compress----------------------------------
976 // Straight out of Tarjan's union-find algorithm
977 uint UnionFind::Find_compress( uint idx ) {
978 uint cur = idx;
979 uint next = lookup(cur);
980 while( next != cur ) { // Scan chain of equivalences
981 assert( next < cur, "always union smaller" );
982 cur = next; // until find a fixed-point
983 next = lookup(cur);
984 }
985 // Core of union-find algorithm: update chain of
986 // equivalences to be equal to the root.
987 while( idx != next ) {
988 uint tmp = lookup(idx);
989 map(idx, next);
990 idx = tmp;
991 }
992 return idx;
993 }
994
995 //------------------------------Find_const-------------------------------------
996 // Like Find above, but no path compress, so bad asymptotic behavior
997 uint UnionFind::Find_const( uint idx ) const {
998 if( idx == 0 ) return idx; // Ignore the zero idx
999 // Off the end? This can happen during debugging dumps
1000 // when data structures have not finished being updated.
1001 if( idx >= _max ) return idx;
1002 uint next = lookup(idx);
1003 while( next != idx ) { // Scan chain of equivalences
1004 idx = next; // until find a fixed-point
1005 next = lookup(idx);
1006 }
1007 return next;
1008 }
1009
1010 //------------------------------Union------------------------------------------
1011 // union 2 sets together.
1012 void UnionFind::Union( uint idx1, uint idx2 ) {
1013 uint src = Find(idx1);
1014 uint dst = Find(idx2);
1015 assert( src, "" );
1016 assert( dst, "" );
1017 assert( src < _max, "oob" );
1018 assert( dst < _max, "oob" );
1019 assert( src < dst, "always union smaller" );
1020 map(dst,src);
1021 }
1022
1023 #ifndef PRODUCT
1024 static void edge_dump(GrowableArray<CFGEdge *> *edges) {
1025 tty->print_cr("---- Edges ----");
1026 for (int i = 0; i < edges->length(); i++) {
1027 CFGEdge *e = edges->at(i);
1028 if (e != NULL) {
1029 edges->at(i)->dump();
1030 }
1031 }
1032 }
1033
1034 static void trace_dump(Trace *traces[], int count) {
1035 tty->print_cr("---- Traces ----");
1036 for (int i = 0; i < count; i++) {
1037 Trace *tr = traces[i];
1038 if (tr != NULL) {
1039 tr->dump();
1040 }
1041 }
1042 }
1043
1044 void Trace::dump( ) const {
1045 tty->print_cr("Trace (freq %f)", first_block()->_freq);
1046 for (Block *b = first_block(); b != NULL; b = next(b)) {
1047 tty->print(" B%d", b->_pre_order);
1048 if (b->head()->is_Loop()) {
1049 tty->print(" (L%d)", b->compute_loop_alignment());
1050 }
1051 if (b->has_loop_alignment()) {
1052 tty->print(" (T%d)", b->code_alignment());
1053 }
1054 }
1055 tty->cr();
1056 }
1057
1058 void CFGEdge::dump( ) const {
1059 tty->print(" B%d --> B%d Freq: %f out:%3d%% in:%3d%% State: ",
1060 from()->_pre_order, to()->_pre_order, freq(), _from_pct, _to_pct);
1061 switch(state()) {
1062 case connected:
1063 tty->print("connected");
1064 break;
1065 case open:
1066 tty->print("open");
1067 break;
1068 case interior:
1069 tty->print("interior");
1070 break;
1071 }
1072 if (infrequent()) {
1073 tty->print(" infrequent");
1074 }
1075 tty->cr();
1076 }
1077 #endif
1078
1079 //=============================================================================
1080
1081 //------------------------------edge_order-------------------------------------
1082 // Comparison function for edges
1083 static int edge_order(CFGEdge **e0, CFGEdge **e1) {
1084 float freq0 = (*e0)->freq();
1085 float freq1 = (*e1)->freq();
1086 if (freq0 != freq1) {
1087 return freq0 > freq1 ? -1 : 1;
1088 }
1089
1090 int dist0 = (*e0)->to()->_rpo - (*e0)->from()->_rpo;
1091 int dist1 = (*e1)->to()->_rpo - (*e1)->from()->_rpo;
1092
1093 return dist1 - dist0;
1094 }
1095
1096 //------------------------------trace_frequency_order--------------------------
1097 // Comparison function for edges
1098 static int trace_frequency_order(const void *p0, const void *p1) {
1099 Trace *tr0 = *(Trace **) p0;
1100 Trace *tr1 = *(Trace **) p1;
1101 Block *b0 = tr0->first_block();
1102 Block *b1 = tr1->first_block();
1103
1104 // The trace of connector blocks goes at the end;
1105 // we only expect one such trace
1106 if (b0->is_connector() != b1->is_connector()) {
1107 return b1->is_connector() ? -1 : 1;
1108 }
1109
1110 // Pull more frequently executed blocks to the beginning
1111 float freq0 = b0->_freq;
1112 float freq1 = b1->_freq;
1113 if (freq0 != freq1) {
1114 return freq0 > freq1 ? -1 : 1;
1115 }
1116
1117 int diff = tr0->first_block()->_rpo - tr1->first_block()->_rpo;
1118
1119 return diff;
1120 }
1121
1122 //------------------------------find_edges-------------------------------------
1123 // Find edges of interest, i.e, those which can fall through. Presumes that
1124 // edges which don't fall through are of low frequency and can be generally
1125 // ignored. Initialize the list of traces.
1126 void PhaseBlockLayout::find_edges()
1127 {
1128 // Walk the blocks, creating edges and Traces
1129 uint i;
1130 Trace *tr = NULL;
1131 for (i = 0; i < _cfg._num_blocks; i++) {
1132 Block *b = _cfg._blocks[i];
1133 tr = new Trace(b, next, prev);
1134 traces[tr->id()] = tr;
1135
1136 // All connector blocks should be at the end of the list
1137 if (b->is_connector()) break;
1138
1139 // If this block and the next one have a one-to-one successor
1140 // predecessor relationship, simply append the next block
1141 int nfallthru = b->num_fall_throughs();
1142 while (nfallthru == 1 &&
1143 b->succ_fall_through(0)) {
1144 Block *n = b->_succs[0];
1145
1146 // Skip over single-entry connector blocks, we don't want to
1147 // add them to the trace.
1148 while (n->is_connector() && n->num_preds() == 1) {
1149 n = n->_succs[0];
1150 }
1151
1152 // We see a merge point, so stop search for the next block
1153 if (n->num_preds() != 1) break;
1154
1155 i++;
1156 assert(n = _cfg._blocks[i], "expecting next block");
1157 tr->append(n);
1158 uf->map(n->_pre_order, tr->id());
1159 traces[n->_pre_order] = NULL;
1160 nfallthru = b->num_fall_throughs();
1161 b = n;
1162 }
1163
1164 if (nfallthru > 0) {
1165 // Create a CFGEdge for each outgoing
1166 // edge that could be a fall-through.
1167 for (uint j = 0; j < b->_num_succs; j++ ) {
1168 if (b->succ_fall_through(j)) {
1169 Block *target = b->non_connector_successor(j);
1170 float freq = b->_freq * b->succ_prob(j);
1171 int from_pct = (int) ((100 * freq) / b->_freq);
1172 int to_pct = (int) ((100 * freq) / target->_freq);
1173 edges->append(new CFGEdge(b, target, freq, from_pct, to_pct));
1174 }
1175 }
1176 }
1177 }
1178
1179 // Group connector blocks into one trace
1180 for (i++; i < _cfg._num_blocks; i++) {
1181 Block *b = _cfg._blocks[i];
1182 assert(b->is_connector(), "connector blocks at the end");
1183 tr->append(b);
1184 uf->map(b->_pre_order, tr->id());
1185 traces[b->_pre_order] = NULL;
1186 }
1187 }
1188
1189 //------------------------------union_traces----------------------------------
1190 // Union two traces together in uf, and null out the trace in the list
1191 void PhaseBlockLayout::union_traces(Trace* updated_trace, Trace* old_trace)
1192 {
1193 uint old_id = old_trace->id();
1194 uint updated_id = updated_trace->id();
1195
1196 uint lo_id = updated_id;
1197 uint hi_id = old_id;
1198
1199 // If from is greater than to, swap values to meet
1200 // UnionFind guarantee.
1201 if (updated_id > old_id) {
1202 lo_id = old_id;
1203 hi_id = updated_id;
1204
1205 // Fix up the trace ids
1206 traces[lo_id] = traces[updated_id];
1207 updated_trace->set_id(lo_id);
1208 }
1209
1210 // Union the lower with the higher and remove the pointer
1211 // to the higher.
1212 uf->Union(lo_id, hi_id);
1213 traces[hi_id] = NULL;
1214 }
1215
1216 //------------------------------grow_traces-------------------------------------
1217 // Append traces together via the most frequently executed edges
1218 void PhaseBlockLayout::grow_traces()
1219 {
1220 // Order the edges, and drive the growth of Traces via the most
1221 // frequently executed edges.
1222 edges->sort(edge_order);
1223 for (int i = 0; i < edges->length(); i++) {
1224 CFGEdge *e = edges->at(i);
1225
1226 if (e->state() != CFGEdge::open) continue;
1227
1228 Block *src_block = e->from();
1229 Block *targ_block = e->to();
1230
1231 // Don't grow traces along backedges?
1232 if (!BlockLayoutRotateLoops) {
1233 if (targ_block->_rpo <= src_block->_rpo) {
1234 targ_block->set_loop_alignment(targ_block);
1235 continue;
1236 }
1237 }
1238
1239 Trace *src_trace = trace(src_block);
1240 Trace *targ_trace = trace(targ_block);
1241
1242 // If the edge in question can join two traces at their ends,
1243 // append one trace to the other.
1244 if (src_trace->last_block() == src_block) {
1245 if (src_trace == targ_trace) {
1246 e->set_state(CFGEdge::interior);
1247 if (targ_trace->backedge(e)) {
1248 // Reset i to catch any newly eligible edge
1249 // (Or we could remember the first "open" edge, and reset there)
1250 i = 0;
1251 }
1252 } else if (targ_trace->first_block() == targ_block) {
1253 e->set_state(CFGEdge::connected);
1254 src_trace->append(targ_trace);
1255 union_traces(src_trace, targ_trace);
1256 }
1257 }
1258 }
1259 }
1260
1261 //------------------------------merge_traces-----------------------------------
1262 // Embed one trace into another, if the fork or join points are sufficiently
1263 // balanced.
1264 void PhaseBlockLayout::merge_traces(bool fall_thru_only)
1265 {
1266 // Walk the edge list a another time, looking at unprocessed edges.
1267 // Fold in diamonds
1268 for (int i = 0; i < edges->length(); i++) {
1269 CFGEdge *e = edges->at(i);
1270
1271 if (e->state() != CFGEdge::open) continue;
1272 if (fall_thru_only) {
1273 if (e->infrequent()) continue;
1274 }
1275
1276 Block *src_block = e->from();
1277 Trace *src_trace = trace(src_block);
1278 bool src_at_tail = src_trace->last_block() == src_block;
1279
1280 Block *targ_block = e->to();
1281 Trace *targ_trace = trace(targ_block);
1282 bool targ_at_start = targ_trace->first_block() == targ_block;
1283
1284 if (src_trace == targ_trace) {
1285 // This may be a loop, but we can't do much about it.
1286 e->set_state(CFGEdge::interior);
1287 continue;
1288 }
1289
1290 if (fall_thru_only) {
1291 // If the edge links the middle of two traces, we can't do anything.
1292 // Mark the edge and continue.
1293 if (!src_at_tail & !targ_at_start) {
1294 continue;
1295 }
1296
1297 // Don't grow traces along backedges?
1298 if (!BlockLayoutRotateLoops && (targ_block->_rpo <= src_block->_rpo)) {
1299 continue;
1300 }
1301
1302 // If both ends of the edge are available, why didn't we handle it earlier?
1303 assert(src_at_tail ^ targ_at_start, "Should have caught this edge earlier.");
1304
1305 if (targ_at_start) {
1306 // Insert the "targ" trace in the "src" trace if the insertion point
1307 // is a two way branch.
1308 // Better profitability check possible, but may not be worth it.
1309 // Someday, see if the this "fork" has an associated "join";
1310 // then make a policy on merging this trace at the fork or join.
1311 // For example, other things being equal, it may be better to place this
1312 // trace at the join point if the "src" trace ends in a two-way, but
1313 // the insertion point is one-way.
1314 assert(src_block->num_fall_throughs() == 2, "unexpected diamond");
1315 e->set_state(CFGEdge::connected);
1316 src_trace->insert_after(src_block, targ_trace);
1317 union_traces(src_trace, targ_trace);
1318 } else if (src_at_tail) {
1319 if (src_trace != trace(_cfg._broot)) {
1320 e->set_state(CFGEdge::connected);
1321 targ_trace->insert_before(targ_block, src_trace);
1322 union_traces(targ_trace, src_trace);
1323 }
1324 }
1325 } else if (e->state() == CFGEdge::open) {
1326 // Append traces, even without a fall-thru connection.
1327 // But leave root entry at the beginning of the block list.
1328 if (targ_trace != trace(_cfg._broot)) {
1329 e->set_state(CFGEdge::connected);
1330 src_trace->append(targ_trace);
1331 union_traces(src_trace, targ_trace);
1332 }
1333 }
1334 }
1335 }
1336
1337 //----------------------------reorder_traces-----------------------------------
1338 // Order the sequence of the traces in some desirable way, and fixup the
1339 // jumps at the end of each block.
1340 void PhaseBlockLayout::reorder_traces(int count)
1341 {
1342 ResourceArea *area = Thread::current()->resource_area();
1343 Trace ** new_traces = NEW_ARENA_ARRAY(area, Trace *, count);
1344 Block_List worklist;
1345 int new_count = 0;
1346
1347 // Compact the traces.
1348 for (int i = 0; i < count; i++) {
1349 Trace *tr = traces[i];
1350 if (tr != NULL) {
1351 new_traces[new_count++] = tr;
1352 }
1353 }
1354
1355 // The entry block should be first on the new trace list.
1356 Trace *tr = trace(_cfg._broot);
1357 assert(tr == new_traces[0], "entry trace misplaced");
1358
1359 // Sort the new trace list by frequency
1360 qsort(new_traces + 1, new_count - 1, sizeof(new_traces[0]), trace_frequency_order);
1361
1362 // Patch up the successor blocks
1363 _cfg._blocks.reset();
1364 _cfg._num_blocks = 0;
1365 for (int i = 0; i < new_count; i++) {
1366 Trace *tr = new_traces[i];
1367 if (tr != NULL) {
1368 tr->fixup_blocks(_cfg);
1369 }
1370 }
1371 }
1372
1373 //------------------------------PhaseBlockLayout-------------------------------
1374 // Order basic blocks based on frequency
1375 PhaseBlockLayout::PhaseBlockLayout(PhaseCFG &cfg) :
1376 Phase(BlockLayout),
1377 _cfg(cfg)
1378 {
1379 ResourceMark rm;
1380 ResourceArea *area = Thread::current()->resource_area();
1381
1382 // List of traces
1383 int size = _cfg._num_blocks + 1;
1384 traces = NEW_ARENA_ARRAY(area, Trace *, size);
1385 memset(traces, 0, size*sizeof(Trace*));
1386 next = NEW_ARENA_ARRAY(area, Block *, size);
1387 memset(next, 0, size*sizeof(Block *));
1388 prev = NEW_ARENA_ARRAY(area, Block *, size);
1389 memset(prev , 0, size*sizeof(Block *));
1390
1391 // List of edges
1392 edges = new GrowableArray<CFGEdge*>;
1393
1394 // Mapping block index --> block_trace
1395 uf = new UnionFind(size);
1396 uf->reset(size);
1397
1398 // Find edges and create traces.
1399 find_edges();
1400
1401 // Grow traces at their ends via most frequent edges.
1402 grow_traces();
1403
1404 // Merge one trace into another, but only at fall-through points.
1405 // This may make diamonds and other related shapes in a trace.
1406 merge_traces(true);
1407
1408 // Run merge again, allowing two traces to be catenated, even if
1409 // one does not fall through into the other. This appends loosely
1410 // related traces to be near each other.
1411 merge_traces(false);
1412
1413 // Re-order all the remaining traces by frequency
1414 reorder_traces(size);
1415
1416 assert(_cfg._num_blocks >= (uint) (size - 1), "number of blocks can not shrink");
1417 }
1418
1419
1420 //------------------------------backedge---------------------------------------
1421 // Edge e completes a loop in a trace. If the target block is head of the
1422 // loop, rotate the loop block so that the loop ends in a conditional branch.
1423 bool Trace::backedge(CFGEdge *e) {
1424 bool loop_rotated = false;
1425 Block *src_block = e->from();
1426 Block *targ_block = e->to();
1427
1428 assert(last_block() == src_block, "loop discovery at back branch");
1429 if (first_block() == targ_block) {
1430 if (BlockLayoutRotateLoops && last_block()->num_fall_throughs() < 2) {
1431 // Find the last block in the trace that has a conditional
1432 // branch.
1433 Block *b;
1434 for (b = last_block(); b != NULL; b = prev(b)) {
1435 if (b->num_fall_throughs() == 2) {
1436 break;
1437 }
1438 }
1439
1440 if (b != last_block() && b != NULL) {
1441 loop_rotated = true;
1442
1443 // Rotate the loop by doing two-part linked-list surgery.
1444 append(first_block());
1445 break_loop_after(b);
1446 }
1447 }
1448
1449 // Backbranch to the top of a trace
1450 // Scroll forward through the trace from the targ_block. If we find
1451 // a loop head before another loop top, use the the loop head alignment.
1452 for (Block *b = targ_block; b != NULL; b = next(b)) {
1453 if (b->has_loop_alignment()) {
1454 break;
1455 }
1456 if (b->head()->is_Loop()) {
1457 targ_block = b;
1458 break;
1459 }
1460 }
1461
1462 first_block()->set_loop_alignment(targ_block);
1463
1464 } else {
1465 // Backbranch into the middle of a trace
1466 targ_block->set_loop_alignment(targ_block);
1467 }
1468
1469 return loop_rotated;
1470 }
1471
1472 //------------------------------fixup_blocks-----------------------------------
1473 // push blocks onto the CFG list
1474 // ensure that blocks have the correct two-way branch sense
1475 void Trace::fixup_blocks(PhaseCFG &cfg) {
1476 Block *last = last_block();
1477 for (Block *b = first_block(); b != NULL; b = next(b)) {
1478 cfg._blocks.push(b);
1479 cfg._num_blocks++;
1480 if (!b->is_connector()) {
1481 int nfallthru = b->num_fall_throughs();
1482 if (b != last) {
1483 if (nfallthru == 2) {
1484 // Ensure that the sense of the branch is correct
1485 Block *bnext = next(b);
1486 Block *bs0 = b->non_connector_successor(0);
1487
1488 MachNode *iff = b->_nodes[b->_nodes.size()-3]->as_Mach();
1489 ProjNode *proj0 = b->_nodes[b->_nodes.size()-2]->as_Proj();
1490 ProjNode *proj1 = b->_nodes[b->_nodes.size()-1]->as_Proj();
1491
1492 if (bnext == bs0) {
1493 // Fall-thru case in succs[0], should be in succs[1]
1494
1495 // Flip targets in _succs map
1496 Block *tbs0 = b->_succs[0];
1497 Block *tbs1 = b->_succs[1];
1498 b->_succs.map( 0, tbs1 );
1499 b->_succs.map( 1, tbs0 );
1500
1501 // Flip projections to match targets
1502 b->_nodes.map(b->_nodes.size()-2, proj1);
1503 b->_nodes.map(b->_nodes.size()-1, proj0);
1504 }
1505 }
1506 }
1507 }
1508 }
1509 }