1 /*
2 * Copyright (c) 1997, 2016, 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 "precompiled.hpp"
26 #include "libadt/vectset.hpp"
27 #include "memory/allocation.inline.hpp"
28 #include "memory/resourceArea.hpp"
29 #include "compiler/compilerDirectives.hpp"
30 #include "opto/block.hpp"
31 #include "opto/cfgnode.hpp"
32 #include "opto/chaitin.hpp"
33 #include "opto/loopnode.hpp"
34 #include "opto/machnode.hpp"
35 #include "opto/matcher.hpp"
36 #include "opto/opcodes.hpp"
37 #include "opto/rootnode.hpp"
38 #include "utilities/copy.hpp"
39
40 void Block_Array::grow( uint i ) {
41 assert(i >= Max(), "must be an overflow");
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] = NULL;
48 }
49 uint old = _size;
50 while( i >= _size ) _size <<= 1; // Double to fit
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 // Return empty status of a block. Empty blocks contain only the head, other
154 // ideal nodes, and an optional trailing goto.
155 int Block::is_Empty() const {
156
157 // Root or start block is not considered empty
158 if (head()->is_Root() || head()->is_Start()) {
159 return not_empty;
160 }
161
162 int success_result = completely_empty;
163 int end_idx = number_of_nodes() - 1;
164
165 // Check for ending goto
166 if ((end_idx > 0) && (get_node(end_idx)->is_MachGoto())) {
167 success_result = empty_with_goto;
168 end_idx--;
169 }
170
171 // Unreachable blocks are considered empty
172 if (num_preds() <= 1) {
173 return success_result;
174 }
175
176 // Ideal nodes are allowable in empty blocks: skip them Only MachNodes
177 // turn directly into code, because only MachNodes have non-trivial
178 // emit() functions.
179 while ((end_idx > 0) && !get_node(end_idx)->is_Mach()) {
180 end_idx--;
181 }
182
183 // No room for any interesting instructions?
184 if (end_idx == 0) {
185 return success_result;
186 }
187
188 return not_empty;
189 }
190
191 // Return true if the block's code implies that it is likely to be
192 // executed infrequently. Check to see if the block ends in a Halt or
193 // a low probability call.
194 bool Block::has_uncommon_code() const {
195 Node* en = end();
196
197 if (en->is_MachGoto())
198 en = en->in(0);
199 if (en->is_Catch())
200 en = en->in(0);
201 if (en->is_MachProj() && en->in(0)->is_MachCall()) {
202 MachCallNode* call = en->in(0)->as_MachCall();
203 if (call->cnt() != COUNT_UNKNOWN && call->cnt() <= PROB_UNLIKELY_MAG(4)) {
204 // This is true for slow-path stubs like new_{instance,array},
205 // slow_arraycopy, complete_monitor_locking, uncommon_trap.
206 // The magic number corresponds to the probability of an uncommon_trap,
207 // even though it is a count not a probability.
208 return true;
209 }
210 }
211
212 int op = en->is_Mach() ? en->as_Mach()->ideal_Opcode() : en->Opcode();
213 return op == Op_Halt;
214 }
215
216 // True if block is low enough frequency or guarded by a test which
217 // mostly does not go here.
218 bool PhaseCFG::is_uncommon(const Block* block) {
219 // Initial blocks must never be moved, so are never uncommon.
220 if (block->head()->is_Root() || block->head()->is_Start()) return false;
221
222 // Check for way-low freq
223 if(block->_freq < BLOCK_FREQUENCY(0.00001f) ) return true;
224
225 // Look for code shape indicating uncommon_trap or slow path
226 if (block->has_uncommon_code()) return true;
227
228 const float epsilon = 0.05f;
229 const float guard_factor = PROB_UNLIKELY_MAG(4) / (1.f - epsilon);
230 uint uncommon_preds = 0;
231 uint freq_preds = 0;
232 uint uncommon_for_freq_preds = 0;
233
234 for( uint i=1; i< block->num_preds(); i++ ) {
235 Block* guard = get_block_for_node(block->pred(i));
236 // Check to see if this block follows its guard 1 time out of 10000
237 // or less.
238 //
239 // See list of magnitude-4 unlikely probabilities in cfgnode.hpp which
240 // we intend to be "uncommon", such as slow-path TLE allocation,
241 // predicted call failure, and uncommon trap triggers.
242 //
243 // Use an epsilon value of 5% to allow for variability in frequency
244 // predictions and floating point calculations. The net effect is
245 // that guard_factor is set to 9500.
246 //
247 // Ignore low-frequency blocks.
248 // The next check is (guard->_freq < 1.e-5 * 9500.).
249 if(guard->_freq*BLOCK_FREQUENCY(guard_factor) < BLOCK_FREQUENCY(0.00001f)) {
250 uncommon_preds++;
251 } else {
252 freq_preds++;
253 if(block->_freq < guard->_freq * guard_factor ) {
254 uncommon_for_freq_preds++;
255 }
256 }
257 }
258 if( block->num_preds() > 1 &&
259 // The block is uncommon if all preds are uncommon or
260 (uncommon_preds == (block->num_preds()-1) ||
261 // it is uncommon for all frequent preds.
262 uncommon_for_freq_preds == freq_preds) ) {
263 return true;
264 }
265 return false;
266 }
267
268 #ifndef PRODUCT
269 void Block::dump_bidx(const Block* orig, outputStream* st) const {
270 if (_pre_order) st->print("B%d",_pre_order);
271 else st->print("N%d", head()->_idx);
272
273 if (Verbose && orig != this) {
274 // Dump the original block's idx
275 st->print(" (");
276 orig->dump_bidx(orig, st);
277 st->print(")");
278 }
279 }
280
281 void Block::dump_pred(const PhaseCFG* cfg, Block* orig, outputStream* st) const {
282 if (is_connector()) {
283 for (uint i=1; i<num_preds(); i++) {
284 Block *p = cfg->get_block_for_node(pred(i));
285 p->dump_pred(cfg, orig, st);
286 }
287 } else {
288 dump_bidx(orig, st);
289 st->print(" ");
290 }
291 }
292
293 void Block::dump_head(const PhaseCFG* cfg, outputStream* st) const {
294 // Print the basic block
295 dump_bidx(this, st);
296 st->print(": #\t");
297
298 // Print the incoming CFG edges and the outgoing CFG edges
299 for( uint i=0; i<_num_succs; i++ ) {
300 non_connector_successor(i)->dump_bidx(_succs[i], st);
301 st->print(" ");
302 }
303 st->print("<- ");
304 if( head()->is_block_start() ) {
305 for (uint i=1; i<num_preds(); i++) {
306 Node *s = pred(i);
307 if (cfg != NULL) {
308 Block *p = cfg->get_block_for_node(s);
309 p->dump_pred(cfg, p, st);
310 } else {
311 while (!s->is_block_start())
312 s = s->in(0);
313 st->print("N%d ", s->_idx );
314 }
315 }
316 } else {
317 st->print("BLOCK HEAD IS JUNK ");
318 }
319
320 // Print loop, if any
321 const Block *bhead = this; // Head of self-loop
322 Node *bh = bhead->head();
323
324 if ((cfg != NULL) && bh->is_Loop() && !head()->is_Root()) {
325 LoopNode *loop = bh->as_Loop();
326 const Block *bx = cfg->get_block_for_node(loop->in(LoopNode::LoopBackControl));
327 while (bx->is_connector()) {
328 bx = cfg->get_block_for_node(bx->pred(1));
329 }
330 st->print("\tLoop: B%d-B%d ", bhead->_pre_order, bx->_pre_order);
331 // Dump any loop-specific bits, especially for CountedLoops.
332 loop->dump_spec(st);
333 } else if (has_loop_alignment()) {
334 st->print(" top-of-loop");
335 }
336 st->print(" Freq: %g",_freq);
337 if( Verbose || WizardMode ) {
338 st->print(" IDom: %d/#%d", _idom ? _idom->_pre_order : 0, _dom_depth);
339 st->print(" RegPressure: %d",_reg_pressure);
340 st->print(" IHRP Index: %d",_ihrp_index);
341 st->print(" FRegPressure: %d",_freg_pressure);
342 st->print(" FHRP Index: %d",_fhrp_index);
343 }
344 st->cr();
345 }
346
347 void Block::dump() const {
348 dump(NULL);
349 }
350
351 void Block::dump(const PhaseCFG* cfg) const {
352 dump_head(cfg);
353 for (uint i=0; i< number_of_nodes(); i++) {
354 get_node(i)->dump();
355 }
356 tty->print("\n");
357 }
358 #endif
359
360 PhaseCFG::PhaseCFG(Arena* arena, RootNode* root, Matcher& matcher)
361 : Phase(CFG)
362 , _root(root)
363 , _block_arena(arena)
364 , _regalloc(NULL)
365 , _scheduling_for_pressure(false)
366 , _matcher(matcher)
367 , _node_to_block_mapping(arena)
368 , _node_latency(NULL)
369 #ifndef PRODUCT
370 , _trace_opto_pipelining(C->directive()->TraceOptoPipeliningOption)
371 #endif
372 #ifdef ASSERT
373 , _raw_oops(arena)
374 #endif
375 {
376 ResourceMark rm;
377 // I'll need a few machine-specific GotoNodes. Make an Ideal GotoNode,
378 // then Match it into a machine-specific Node. Then clone the machine
379 // Node on demand.
380 Node *x = new GotoNode(NULL);
381 x->init_req(0, x);
382 _goto = matcher.match_tree(x);
383 assert(_goto != NULL, "");
384 _goto->set_req(0,_goto);
385
386 // Build the CFG in Reverse Post Order
387 _number_of_blocks = build_cfg();
388 _root_block = get_block_for_node(_root);
389 }
390
391 // Build a proper looking CFG. Make every block begin with either a StartNode
392 // or a RegionNode. Make every block end with either a Goto, If or Return.
393 // The RootNode both starts and ends it's own block. Do this with a recursive
394 // backwards walk over the control edges.
395 uint PhaseCFG::build_cfg() {
396 Arena *a = Thread::current()->resource_area();
397 VectorSet visited(a);
398
399 // Allocate stack with enough space to avoid frequent realloc
400 Node_Stack nstack(a, C->live_nodes() >> 1);
401 nstack.push(_root, 0);
402 uint sum = 0; // Counter for blocks
403
404 while (nstack.is_nonempty()) {
405 // node and in's index from stack's top
406 // 'np' is _root (see above) or RegionNode, StartNode: we push on stack
407 // only nodes which point to the start of basic block (see below).
408 Node *np = nstack.node();
409 // idx > 0, except for the first node (_root) pushed on stack
410 // at the beginning when idx == 0.
411 // We will use the condition (idx == 0) later to end the build.
412 uint idx = nstack.index();
413 Node *proj = np->in(idx);
414 const Node *x = proj->is_block_proj();
415 // Does the block end with a proper block-ending Node? One of Return,
416 // If or Goto? (This check should be done for visited nodes also).
417 if (x == NULL) { // Does not end right...
418 Node *g = _goto->clone(); // Force it to end in a Goto
419 g->set_req(0, proj);
420 np->set_req(idx, g);
421 x = proj = g;
422 }
423 if (!visited.test_set(x->_idx)) { // Visit this block once
424 // Skip any control-pinned middle'in stuff
425 Node *p = proj;
426 do {
427 proj = p; // Update pointer to last Control
428 p = p->in(0); // Move control forward
429 } while( !p->is_block_proj() &&
430 !p->is_block_start() );
431 // Make the block begin with one of Region or StartNode.
432 if( !p->is_block_start() ) {
433 RegionNode *r = new RegionNode( 2 );
434 r->init_req(1, p); // Insert RegionNode in the way
435 proj->set_req(0, r); // Insert RegionNode in the way
436 p = r;
437 }
438 // 'p' now points to the start of this basic block
439
440 // Put self in array of basic blocks
441 Block *bb = new (_block_arena) Block(_block_arena, p);
442 map_node_to_block(p, bb);
443 map_node_to_block(x, bb);
444 if( x != p ) { // Only for root is x == p
445 bb->push_node((Node*)x);
446 }
447 // Now handle predecessors
448 ++sum; // Count 1 for self block
449 uint cnt = bb->num_preds();
450 for (int i = (cnt - 1); i > 0; i-- ) { // For all predecessors
451 Node *prevproj = p->in(i); // Get prior input
452 assert( !prevproj->is_Con(), "dead input not removed" );
453 // Check to see if p->in(i) is a "control-dependent" CFG edge -
454 // i.e., it splits at the source (via an IF or SWITCH) and merges
455 // at the destination (via a many-input Region).
456 // This breaks critical edges. The RegionNode to start the block
457 // will be added when <p,i> is pulled off the node stack
458 if ( cnt > 2 ) { // Merging many things?
459 assert( prevproj== bb->pred(i),"");
460 if(prevproj->is_block_proj() != prevproj) { // Control-dependent edge?
461 // Force a block on the control-dependent edge
462 Node *g = _goto->clone(); // Force it to end in a Goto
463 g->set_req(0,prevproj);
464 p->set_req(i,g);
465 }
466 }
467 nstack.push(p, i); // 'p' is RegionNode or StartNode
468 }
469 } else { // Post-processing visited nodes
470 nstack.pop(); // remove node from stack
471 // Check if it the fist node pushed on stack at the beginning.
472 if (idx == 0) break; // end of the build
473 // Find predecessor basic block
474 Block *pb = get_block_for_node(x);
475 // Insert into nodes array, if not already there
476 if (!has_block(proj)) {
477 assert( x != proj, "" );
478 // Map basic block of projection
479 map_node_to_block(proj, pb);
480 pb->push_node(proj);
481 }
482 // Insert self as a child of my predecessor block
483 pb->_succs.map(pb->_num_succs++, get_block_for_node(np));
484 assert( pb->get_node(pb->number_of_nodes() - pb->_num_succs)->is_block_proj(),
485 "too many control users, not a CFG?" );
486 }
487 }
488 // Return number of basic blocks for all children and self
489 return sum;
490 }
491
492 // Inserts a goto & corresponding basic block between
493 // block[block_no] and its succ_no'th successor block
494 void PhaseCFG::insert_goto_at(uint block_no, uint succ_no) {
495 // get block with block_no
496 assert(block_no < number_of_blocks(), "illegal block number");
497 Block* in = get_block(block_no);
498 // get successor block succ_no
499 assert(succ_no < in->_num_succs, "illegal successor number");
500 Block* out = in->_succs[succ_no];
501 // Compute frequency of the new block. Do this before inserting
502 // new block in case succ_prob() needs to infer the probability from
503 // surrounding blocks.
504 float freq = in->_freq * in->succ_prob(succ_no);
505 // get ProjNode corresponding to the succ_no'th successor of the in block
506 ProjNode* proj = in->get_node(in->number_of_nodes() - in->_num_succs + succ_no)->as_Proj();
507 // create region for basic block
508 RegionNode* region = new RegionNode(2);
509 region->init_req(1, proj);
510 // setup corresponding basic block
511 Block* block = new (_block_arena) Block(_block_arena, region);
512 map_node_to_block(region, block);
513 C->regalloc()->set_bad(region->_idx);
514 // add a goto node
515 Node* gto = _goto->clone(); // get a new goto node
516 gto->set_req(0, region);
517 // add it to the basic block
518 block->push_node(gto);
519 map_node_to_block(gto, block);
520 C->regalloc()->set_bad(gto->_idx);
521 // hook up successor block
522 block->_succs.map(block->_num_succs++, out);
523 // remap successor's predecessors if necessary
524 for (uint i = 1; i < out->num_preds(); i++) {
525 if (out->pred(i) == proj) out->head()->set_req(i, gto);
526 }
527 // remap predecessor's successor to new block
528 in->_succs.map(succ_no, block);
529 // Set the frequency of the new block
530 block->_freq = freq;
531 // add new basic block to basic block list
532 add_block_at(block_no + 1, block);
533 }
534
535 // Does this block end in a multiway branch that cannot have the default case
536 // flipped for another case?
537 static bool no_flip_branch(Block *b) {
538 int branch_idx = b->number_of_nodes() - b->_num_succs-1;
539 if (branch_idx < 1) {
540 return false;
541 }
542 Node *branch = b->get_node(branch_idx);
543 if (branch->is_Catch()) {
544 return true;
545 }
546 if (branch->is_Mach()) {
547 if (branch->is_MachNullCheck()) {
548 return true;
549 }
550 int iop = branch->as_Mach()->ideal_Opcode();
551 if (iop == Op_FastLock || iop == Op_FastUnlock) {
552 return true;
553 }
554 // Don't flip if branch has an implicit check.
555 if (branch->as_Mach()->is_TrapBasedCheckNode()) {
556 return true;
557 }
558 }
559 return false;
560 }
561
562 // Check for NeverBranch at block end. This needs to become a GOTO to the
563 // true target. NeverBranch are treated as a conditional branch that always
564 // goes the same direction for most of the optimizer and are used to give a
565 // fake exit path to infinite loops. At this late stage they need to turn
566 // into Goto's so that when you enter the infinite loop you indeed hang.
567 void PhaseCFG::convert_NeverBranch_to_Goto(Block *b) {
568 // Find true target
569 int end_idx = b->end_idx();
570 int idx = b->get_node(end_idx+1)->as_Proj()->_con;
571 Block *succ = b->_succs[idx];
572 Node* gto = _goto->clone(); // get a new goto node
573 gto->set_req(0, b->head());
574 Node *bp = b->get_node(end_idx);
575 b->map_node(gto, end_idx); // Slam over NeverBranch
576 map_node_to_block(gto, b);
577 C->regalloc()->set_bad(gto->_idx);
578 b->pop_node(); // Yank projections
579 b->pop_node(); // Yank projections
580 b->_succs.map(0,succ); // Map only successor
581 b->_num_succs = 1;
582 // remap successor's predecessors if necessary
583 uint j;
584 for( j = 1; j < succ->num_preds(); j++)
585 if( succ->pred(j)->in(0) == bp )
586 succ->head()->set_req(j, gto);
587 // Kill alternate exit path
588 Block *dead = b->_succs[1-idx];
589 for( j = 1; j < dead->num_preds(); j++)
590 if( dead->pred(j)->in(0) == bp )
591 break;
592 // Scan through block, yanking dead path from
593 // all regions and phis.
594 dead->head()->del_req(j);
595 for( int k = 1; dead->get_node(k)->is_Phi(); k++ )
596 dead->get_node(k)->del_req(j);
597 }
598
599 // Helper function to move block bx to the slot following b_index. Return
600 // true if the move is successful, otherwise false
601 bool PhaseCFG::move_to_next(Block* bx, uint b_index) {
602 if (bx == NULL) return false;
603
604 // Return false if bx is already scheduled.
605 uint bx_index = bx->_pre_order;
606 if ((bx_index <= b_index) && (get_block(bx_index) == bx)) {
607 return false;
608 }
609
610 // Find the current index of block bx on the block list
611 bx_index = b_index + 1;
612 while (bx_index < number_of_blocks() && get_block(bx_index) != bx) {
613 bx_index++;
614 }
615 assert(get_block(bx_index) == bx, "block not found");
616
617 // If the previous block conditionally falls into bx, return false,
618 // because moving bx will create an extra jump.
619 for(uint k = 1; k < bx->num_preds(); k++ ) {
620 Block* pred = get_block_for_node(bx->pred(k));
621 if (pred == get_block(bx_index - 1)) {
622 if (pred->_num_succs != 1) {
623 return false;
624 }
625 }
626 }
627
628 // Reinsert bx just past block 'b'
629 _blocks.remove(bx_index);
630 _blocks.insert(b_index + 1, bx);
631 return true;
632 }
633
634 // Move empty and uncommon blocks to the end.
635 void PhaseCFG::move_to_end(Block *b, uint i) {
636 int e = b->is_Empty();
637 if (e != Block::not_empty) {
638 if (e == Block::empty_with_goto) {
639 // Remove the goto, but leave the block.
640 b->pop_node();
641 }
642 // Mark this block as a connector block, which will cause it to be
643 // ignored in certain functions such as non_connector_successor().
644 b->set_connector();
645 }
646 // Move the empty block to the end, and don't recheck.
647 _blocks.remove(i);
648 _blocks.push(b);
649 }
650
651 // Set loop alignment for every block
652 void PhaseCFG::set_loop_alignment() {
653 uint last = number_of_blocks();
654 assert(get_block(0) == get_root_block(), "");
655
656 for (uint i = 1; i < last; i++) {
657 Block* block = get_block(i);
658 if (block->head()->is_Loop()) {
659 block->set_loop_alignment(block);
660 }
661 }
662 }
663
664 // Make empty basic blocks to be "connector" blocks, Move uncommon blocks
665 // to the end.
666 void PhaseCFG::remove_empty_blocks() {
667 // Move uncommon blocks to the end
668 uint last = number_of_blocks();
669 assert(get_block(0) == get_root_block(), "");
670
671 for (uint i = 1; i < last; i++) {
672 Block* block = get_block(i);
673 if (block->is_connector()) {
674 break;
675 }
676
677 // Check for NeverBranch at block end. This needs to become a GOTO to the
678 // true target. NeverBranch are treated as a conditional branch that
679 // always goes the same direction for most of the optimizer and are used
680 // to give a fake exit path to infinite loops. At this late stage they
681 // need to turn into Goto's so that when you enter the infinite loop you
682 // indeed hang.
683 if (block->get_node(block->end_idx())->Opcode() == Op_NeverBranch) {
684 convert_NeverBranch_to_Goto(block);
685 }
686
687 // Look for uncommon blocks and move to end.
688 if (!C->do_freq_based_layout()) {
689 if (is_uncommon(block)) {
690 move_to_end(block, i);
691 last--; // No longer check for being uncommon!
692 if (no_flip_branch(block)) { // Fall-thru case must follow?
693 // Find the fall-thru block
694 block = get_block(i);
695 move_to_end(block, i);
696 last--;
697 }
698 // backup block counter post-increment
699 i--;
700 }
701 }
702 }
703
704 // Move empty blocks to the end
705 last = number_of_blocks();
706 for (uint i = 1; i < last; i++) {
707 Block* block = get_block(i);
708 if (block->is_Empty() != Block::not_empty) {
709 move_to_end(block, i);
710 last--;
711 i--;
712 }
713 } // End of for all blocks
714 }
715
716 Block *PhaseCFG::fixup_trap_based_check(Node *branch, Block *block, int block_pos, Block *bnext) {
717 // Trap based checks must fall through to the successor with
718 // PROB_ALWAYS.
719 // They should be an If with 2 successors.
720 assert(branch->is_MachIf(), "must be If");
721 assert(block->_num_succs == 2, "must have 2 successors");
722
723 // Get the If node and the projection for the first successor.
724 MachIfNode *iff = block->get_node(block->number_of_nodes()-3)->as_MachIf();
725 ProjNode *proj0 = block->get_node(block->number_of_nodes()-2)->as_Proj();
726 ProjNode *proj1 = block->get_node(block->number_of_nodes()-1)->as_Proj();
727 ProjNode *projt = (proj0->Opcode() == Op_IfTrue) ? proj0 : proj1;
728 ProjNode *projf = (proj0->Opcode() == Op_IfFalse) ? proj0 : proj1;
729
730 // Assert that proj0 and succs[0] match up. Similarly for proj1 and succs[1].
731 assert(proj0->raw_out(0) == block->_succs[0]->head(), "Mismatch successor 0");
732 assert(proj1->raw_out(0) == block->_succs[1]->head(), "Mismatch successor 1");
733
734 ProjNode *proj_always;
735 ProjNode *proj_never;
736 // We must negate the branch if the implicit check doesn't follow
737 // the branch's TRUE path. Then, the new TRUE branch target will
738 // be the old FALSE branch target.
739 if (iff->_prob <= 2*PROB_NEVER) { // There are small rounding errors.
740 proj_never = projt;
741 proj_always = projf;
742 } else {
743 // We must negate the branch if the trap doesn't follow the
744 // branch's TRUE path. Then, the new TRUE branch target will
745 // be the old FALSE branch target.
746 proj_never = projf;
747 proj_always = projt;
748 iff->negate();
749 }
750 assert(iff->_prob <= 2*PROB_NEVER, "Trap based checks are expected to trap never!");
751 // Map the successors properly
752 block->_succs.map(0, get_block_for_node(proj_never ->raw_out(0))); // The target of the trap.
753 block->_succs.map(1, get_block_for_node(proj_always->raw_out(0))); // The fall through target.
754
755 if (block->get_node(block->number_of_nodes() - block->_num_succs + 1) != proj_always) {
756 block->map_node(proj_never, block->number_of_nodes() - block->_num_succs + 0);
757 block->map_node(proj_always, block->number_of_nodes() - block->_num_succs + 1);
758 }
759
760 // Place the fall through block after this block.
761 Block *bs1 = block->non_connector_successor(1);
762 if (bs1 != bnext && move_to_next(bs1, block_pos)) {
763 bnext = bs1;
764 }
765 // If the fall through block still is not the next block, insert a goto.
766 if (bs1 != bnext) {
767 insert_goto_at(block_pos, 1);
768 }
769 return bnext;
770 }
771
772 // Fix up the final control flow for basic blocks.
773 void PhaseCFG::fixup_flow() {
774 // Fixup final control flow for the blocks. Remove jump-to-next
775 // block. If neither arm of an IF follows the conditional branch, we
776 // have to add a second jump after the conditional. We place the
777 // TRUE branch target in succs[0] for both GOTOs and IFs.
778 for (uint i = 0; i < number_of_blocks(); i++) {
779 Block* block = get_block(i);
780 block->_pre_order = i; // turn pre-order into block-index
781
782 // Connector blocks need no further processing.
783 if (block->is_connector()) {
784 assert((i+1) == number_of_blocks() || get_block(i + 1)->is_connector(), "All connector blocks should sink to the end");
785 continue;
786 }
787 assert(block->is_Empty() != Block::completely_empty, "Empty blocks should be connectors");
788
789 Block* bnext = (i < number_of_blocks() - 1) ? get_block(i + 1) : NULL;
790 Block* bs0 = block->non_connector_successor(0);
791
792 // Check for multi-way branches where I cannot negate the test to
793 // exchange the true and false targets.
794 if (no_flip_branch(block)) {
795 // Find fall through case - if must fall into its target.
796 // Get the index of the branch's first successor.
797 int branch_idx = block->number_of_nodes() - block->_num_succs;
798
799 // The branch is 1 before the branch's first successor.
800 Node *branch = block->get_node(branch_idx-1);
801
802 // Handle no-flip branches which have implicit checks and which require
803 // special block ordering and individual semantics of the 'fall through
804 // case'.
805 if ((TrapBasedNullChecks || TrapBasedRangeChecks) &&
806 branch->is_Mach() && branch->as_Mach()->is_TrapBasedCheckNode()) {
807 bnext = fixup_trap_based_check(branch, block, i, bnext);
808 } else {
809 // Else, default handling for no-flip branches
810 for (uint j2 = 0; j2 < block->_num_succs; j2++) {
811 const ProjNode* p = block->get_node(branch_idx + j2)->as_Proj();
812 if (p->_con == 0) {
813 // successor j2 is fall through case
814 if (block->non_connector_successor(j2) != bnext) {
815 // but it is not the next block => insert a goto
816 insert_goto_at(i, j2);
817 }
818 // Put taken branch in slot 0
819 if (j2 == 0 && block->_num_succs == 2) {
820 // Flip targets in succs map
821 Block *tbs0 = block->_succs[0];
822 Block *tbs1 = block->_succs[1];
823 block->_succs.map(0, tbs1);
824 block->_succs.map(1, tbs0);
825 }
826 break;
827 }
828 }
829 }
830
831 // Remove all CatchProjs
832 for (uint j = 0; j < block->_num_succs; j++) {
833 block->pop_node();
834 }
835
836 } else if (block->_num_succs == 1) {
837 // Block ends in a Goto?
838 if (bnext == bs0) {
839 // We fall into next block; remove the Goto
840 block->pop_node();
841 }
842
843 } else if(block->_num_succs == 2) { // Block ends in a If?
844 // Get opcode of 1st projection (matches _succs[0])
845 // Note: Since this basic block has 2 exits, the last 2 nodes must
846 // be projections (in any order), the 3rd last node must be
847 // the IfNode (we have excluded other 2-way exits such as
848 // CatchNodes already).
849 MachNode* iff = block->get_node(block->number_of_nodes() - 3)->as_Mach();
850 ProjNode* proj0 = block->get_node(block->number_of_nodes() - 2)->as_Proj();
851 ProjNode* proj1 = block->get_node(block->number_of_nodes() - 1)->as_Proj();
852
853 // Assert that proj0 and succs[0] match up. Similarly for proj1 and succs[1].
854 assert(proj0->raw_out(0) == block->_succs[0]->head(), "Mismatch successor 0");
855 assert(proj1->raw_out(0) == block->_succs[1]->head(), "Mismatch successor 1");
856
857 Block* bs1 = block->non_connector_successor(1);
858
859 // Check for neither successor block following the current
860 // block ending in a conditional. If so, move one of the
861 // successors after the current one, provided that the
862 // successor was previously unscheduled, but moveable
863 // (i.e., all paths to it involve a branch).
864 if (!C->do_freq_based_layout() && bnext != bs0 && bnext != bs1) {
865 // Choose the more common successor based on the probability
866 // of the conditional branch.
867 Block* bx = bs0;
868 Block* by = bs1;
869
870 // _prob is the probability of taking the true path. Make
871 // p the probability of taking successor #1.
872 float p = iff->as_MachIf()->_prob;
873 if (proj0->Opcode() == Op_IfTrue) {
874 p = 1.0 - p;
875 }
876
877 // Prefer successor #1 if p > 0.5
878 if (p > PROB_FAIR) {
879 bx = bs1;
880 by = bs0;
881 }
882
883 // Attempt the more common successor first
884 if (move_to_next(bx, i)) {
885 bnext = bx;
886 } else if (move_to_next(by, i)) {
887 bnext = by;
888 }
889 }
890
891 // Check for conditional branching the wrong way. Negate
892 // conditional, if needed, so it falls into the following block
893 // and branches to the not-following block.
894
895 // Check for the next block being in succs[0]. We are going to branch
896 // to succs[0], so we want the fall-thru case as the next block in
897 // succs[1].
898 if (bnext == bs0) {
899 // Fall-thru case in succs[0], so flip targets in succs map
900 Block* tbs0 = block->_succs[0];
901 Block* tbs1 = block->_succs[1];
902 block->_succs.map(0, tbs1);
903 block->_succs.map(1, tbs0);
904 // Flip projection for each target
905 ProjNode* tmp = proj0;
906 proj0 = proj1;
907 proj1 = tmp;
908
909 } else if(bnext != bs1) {
910 // Need a double-branch
911 // The existing conditional branch need not change.
912 // Add a unconditional branch to the false target.
913 // Alas, it must appear in its own block and adding a
914 // block this late in the game is complicated. Sigh.
915 insert_goto_at(i, 1);
916 }
917
918 // Make sure we TRUE branch to the target
919 if (proj0->Opcode() == Op_IfFalse) {
920 iff->as_MachIf()->negate();
921 }
922
923 block->pop_node(); // Remove IfFalse & IfTrue projections
924 block->pop_node();
925
926 } else {
927 // Multi-exit block, e.g. a switch statement
928 // But we don't need to do anything here
929 }
930 } // End of for all blocks
931 }
932
933
934 // postalloc_expand: Expand nodes after register allocation.
935 //
936 // postalloc_expand has to be called after register allocation, just
937 // before output (i.e. scheduling). It only gets called if
938 // Matcher::require_postalloc_expand is true.
939 //
940 // Background:
941 //
942 // Nodes that are expandend (one compound node requiring several
943 // assembler instructions to be implemented split into two or more
944 // non-compound nodes) after register allocation are not as nice as
945 // the ones expanded before register allocation - they don't
946 // participate in optimizations as global code motion. But after
947 // register allocation we can expand nodes that use registers which
948 // are not spillable or registers that are not allocated, because the
949 // old compound node is simply replaced (in its location in the basic
950 // block) by a new subgraph which does not contain compound nodes any
951 // more. The scheduler called during output can later on process these
952 // non-compound nodes.
953 //
954 // Implementation:
955 //
956 // Nodes requiring postalloc expand are specified in the ad file by using
957 // a postalloc_expand statement instead of ins_encode. A postalloc_expand
958 // contains a single call to an encoding, as does an ins_encode
959 // statement. Instead of an emit() function a postalloc_expand() function
960 // is generated that doesn't emit assembler but creates a new
961 // subgraph. The code below calls this postalloc_expand function for each
962 // node with the appropriate attribute. This function returns the new
963 // nodes generated in an array passed in the call. The old node,
964 // potential MachTemps before and potential Projs after it then get
965 // disconnected and replaced by the new nodes. The instruction
966 // generating the result has to be the last one in the array. In
967 // general it is assumed that Projs after the node expanded are
968 // kills. These kills are not required any more after expanding as
969 // there are now explicitly visible def-use chains and the Projs are
970 // removed. This does not hold for calls: They do not only have
971 // kill-Projs but also Projs defining values. Therefore Projs after
972 // the node expanded are removed for all but for calls. If a node is
973 // to be reused, it must be added to the nodes list returned, and it
974 // will be added again.
975 //
976 // Implementing the postalloc_expand function for a node in an enc_class
977 // is rather tedious. It requires knowledge about many node details, as
978 // the nodes and the subgraph must be hand crafted. To simplify this,
979 // adlc generates some utility variables into the postalloc_expand function,
980 // e.g., holding the operands as specified by the postalloc_expand encoding
981 // specification, e.g.:
982 // * unsigned idx_<par_name> holding the index of the node in the ins
983 // * Node *n_<par_name> holding the node loaded from the ins
984 // * MachOpnd *op_<par_name> holding the corresponding operand
985 //
986 // The ordering of operands can not be determined by looking at a
987 // rule. Especially if a match rule matches several different trees,
988 // several nodes are generated from one instruct specification with
989 // different operand orderings. In this case the adlc generated
990 // variables are the only way to access the ins and operands
991 // deterministically.
992 //
993 // If assigning a register to a node that contains an oop, don't
994 // forget to call ra_->set_oop() for the node.
995 void PhaseCFG::postalloc_expand(PhaseRegAlloc* _ra) {
996 GrowableArray <Node *> new_nodes(32); // Array with new nodes filled by postalloc_expand function of node.
997 GrowableArray <Node *> remove(32);
998 GrowableArray <Node *> succs(32);
999 unsigned int max_idx = C->unique(); // Remember to distinguish new from old nodes.
1000 DEBUG_ONLY(bool foundNode = false);
1001
1002 // for all blocks
1003 for (uint i = 0; i < number_of_blocks(); i++) {
1004 Block *b = _blocks[i];
1005 // For all instructions in the current block.
1006 for (uint j = 0; j < b->number_of_nodes(); j++) {
1007 Node *n = b->get_node(j);
1008 if (n->is_Mach() && n->as_Mach()->requires_postalloc_expand()) {
1009 #ifdef ASSERT
1010 if (TracePostallocExpand) {
1011 if (!foundNode) {
1012 foundNode = true;
1013 tty->print("POSTALLOC EXPANDING %d %s\n", C->compile_id(),
1014 C->method() ? C->method()->name()->as_utf8() : C->stub_name());
1015 }
1016 tty->print(" postalloc expanding "); n->dump();
1017 if (Verbose) {
1018 tty->print(" with ins:\n");
1019 for (uint k = 0; k < n->len(); ++k) {
1020 if (n->in(k)) { tty->print(" "); n->in(k)->dump(); }
1021 }
1022 }
1023 }
1024 #endif
1025 new_nodes.clear();
1026 // Collect nodes that have to be removed from the block later on.
1027 uint req = n->req();
1028 remove.clear();
1029 for (uint k = 0; k < req; ++k) {
1030 if (n->in(k) && n->in(k)->is_MachTemp()) {
1031 remove.push(n->in(k)); // MachTemps which are inputs to the old node have to be removed.
1032 n->in(k)->del_req(0);
1033 j--;
1034 }
1035 }
1036
1037 // Check whether we can allocate enough nodes. We set a fix limit for
1038 // the size of postalloc expands with this.
1039 uint unique_limit = C->unique() + 40;
1040 if (unique_limit >= _ra->node_regs_max_index()) {
1041 Compile::current()->record_failure("out of nodes in postalloc expand");
1042 return;
1043 }
1044
1045 // Emit (i.e. generate new nodes).
1046 n->as_Mach()->postalloc_expand(&new_nodes, _ra);
1047
1048 assert(C->unique() < unique_limit, "You allocated too many nodes in your postalloc expand.");
1049
1050 // Disconnect the inputs of the old node.
1051 //
1052 // We reuse MachSpillCopy nodes. If we need to expand them, there
1053 // are many, so reusing pays off. If reused, the node already
1054 // has the new ins. n must be the last node on new_nodes list.
1055 if (!n->is_MachSpillCopy()) {
1056 for (int k = req - 1; k >= 0; --k) {
1057 n->del_req(k);
1058 }
1059 }
1060
1061 #ifdef ASSERT
1062 // Check that all nodes have proper operands.
1063 for (int k = 0; k < new_nodes.length(); ++k) {
1064 if (new_nodes.at(k)->_idx < max_idx || !new_nodes.at(k)->is_Mach()) continue; // old node, Proj ...
1065 MachNode *m = new_nodes.at(k)->as_Mach();
1066 for (unsigned int l = 0; l < m->num_opnds(); ++l) {
1067 if (MachOper::notAnOper(m->_opnds[l])) {
1068 outputStream *os = tty;
1069 os->print("Node %s ", m->Name());
1070 os->print("has invalid opnd %d: %p\n", l, m->_opnds[l]);
1071 assert(0, "Invalid operands, see inline trace in hs_err_pid file.");
1072 }
1073 }
1074 }
1075 #endif
1076
1077 // Collect succs of old node in remove (for projections) and in succs (for
1078 // all other nodes) do _not_ collect projections in remove (but in succs)
1079 // in case the node is a call. We need the projections for calls as they are
1080 // associated with registes (i.e. they are defs).
1081 succs.clear();
1082 for (DUIterator k = n->outs(); n->has_out(k); k++) {
1083 if (n->out(k)->is_Proj() && !n->is_MachCall() && !n->is_MachBranch()) {
1084 remove.push(n->out(k));
1085 } else {
1086 succs.push(n->out(k));
1087 }
1088 }
1089 // Replace old node n as input of its succs by last of the new nodes.
1090 for (int k = 0; k < succs.length(); ++k) {
1091 Node *succ = succs.at(k);
1092 for (uint l = 0; l < succ->req(); ++l) {
1093 if (succ->in(l) == n) {
1094 succ->set_req(l, new_nodes.at(new_nodes.length() - 1));
1095 }
1096 }
1097 for (uint l = succ->req(); l < succ->len(); ++l) {
1098 if (succ->in(l) == n) {
1099 succ->set_prec(l, new_nodes.at(new_nodes.length() - 1));
1100 }
1101 }
1102 }
1103
1104 // Index of old node in block.
1105 uint index = b->find_node(n);
1106 // Insert new nodes into block and map them in nodes->blocks array
1107 // and remember last node in n2.
1108 Node *n2 = NULL;
1109 for (int k = 0; k < new_nodes.length(); ++k) {
1110 n2 = new_nodes.at(k);
1111 b->insert_node(n2, ++index);
1112 map_node_to_block(n2, b);
1113 }
1114
1115 // Add old node n to remove and remove them all from block.
1116 remove.push(n);
1117 j--;
1118 #ifdef ASSERT
1119 if (TracePostallocExpand && Verbose) {
1120 tty->print(" removing:\n");
1121 for (int k = 0; k < remove.length(); ++k) {
1122 tty->print(" "); remove.at(k)->dump();
1123 }
1124 tty->print(" inserting:\n");
1125 for (int k = 0; k < new_nodes.length(); ++k) {
1126 tty->print(" "); new_nodes.at(k)->dump();
1127 }
1128 }
1129 #endif
1130 for (int k = 0; k < remove.length(); ++k) {
1131 if (b->contains(remove.at(k))) {
1132 b->find_remove(remove.at(k));
1133 } else {
1134 assert(remove.at(k)->is_Proj() && (remove.at(k)->in(0)->is_MachBranch()), "");
1135 }
1136 }
1137 // If anything has been inserted (n2 != NULL), continue after last node inserted.
1138 // This does not always work. Some postalloc expands don't insert any nodes, if they
1139 // do optimizations (e.g., max(x,x)). In this case we decrement j accordingly.
1140 j = n2 ? b->find_node(n2) : j;
1141 }
1142 }
1143 }
1144
1145 #ifdef ASSERT
1146 if (foundNode) {
1147 tty->print("FINISHED %d %s\n", C->compile_id(),
1148 C->method() ? C->method()->name()->as_utf8() : C->stub_name());
1149 tty->flush();
1150 }
1151 #endif
1152 }
1153
1154
1155 //------------------------------dump-------------------------------------------
1156 #ifndef PRODUCT
1157 void PhaseCFG::_dump_cfg( const Node *end, VectorSet &visited ) const {
1158 const Node *x = end->is_block_proj();
1159 assert( x, "not a CFG" );
1160
1161 // Do not visit this block again
1162 if( visited.test_set(x->_idx) ) return;
1163
1164 // Skip through this block
1165 const Node *p = x;
1166 do {
1167 p = p->in(0); // Move control forward
1168 assert( !p->is_block_proj() || p->is_Root(), "not a CFG" );
1169 } while( !p->is_block_start() );
1170
1171 // Recursively visit
1172 for (uint i = 1; i < p->req(); i++) {
1173 _dump_cfg(p->in(i), visited);
1174 }
1175
1176 // Dump the block
1177 get_block_for_node(p)->dump(this);
1178 }
1179
1180 void PhaseCFG::dump( ) const {
1181 tty->print("\n--- CFG --- %d BBs\n", number_of_blocks());
1182 if (_blocks.size()) { // Did we do basic-block layout?
1183 for (uint i = 0; i < number_of_blocks(); i++) {
1184 const Block* block = get_block(i);
1185 block->dump(this);
1186 }
1187 } else { // Else do it with a DFS
1188 VectorSet visited(_block_arena);
1189 _dump_cfg(_root,visited);
1190 }
1191 }
1192
1193 void PhaseCFG::dump_headers() {
1194 for (uint i = 0; i < number_of_blocks(); i++) {
1195 Block* block = get_block(i);
1196 if (block != NULL) {
1197 block->dump_head(this);
1198 }
1199 }
1200 }
1201
1202 void PhaseCFG::verify() const {
1203 #ifdef ASSERT
1204 // Verify sane CFG
1205 for (uint i = 0; i < number_of_blocks(); i++) {
1206 Block* block = get_block(i);
1207 uint cnt = block->number_of_nodes();
1208 uint j;
1209 for (j = 0; j < cnt; j++) {
1210 Node *n = block->get_node(j);
1211 assert(get_block_for_node(n) == block, "");
1212 if (j >= 1 && n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_CreateEx) {
1213 assert(j == 1 || block->get_node(j-1)->is_Phi(), "CreateEx must be first instruction in block");
1214 }
1215 if (n->needs_anti_dependence_check()) {
1216 verify_anti_dependences(block, n);
1217 }
1218 for (uint k = 0; k < n->req(); k++) {
1219 Node *def = n->in(k);
1220 if (def && def != n) {
1221 assert(get_block_for_node(def) || def->is_Con(), "must have block; constants for debug info ok");
1222 // Verify that instructions in the block is in correct order.
1223 // Uses must follow their definition if they are at the same block.
1224 // Mostly done to check that MachSpillCopy nodes are placed correctly
1225 // when CreateEx node is moved in build_ifg_physical().
1226 if (get_block_for_node(def) == block && !(block->head()->is_Loop() && n->is_Phi()) &&
1227 // See (+++) comment in reg_split.cpp
1228 !(n->jvms() != NULL && n->jvms()->is_monitor_use(k))) {
1229 bool is_loop = false;
1230 if (n->is_Phi()) {
1231 for (uint l = 1; l < def->req(); l++) {
1232 if (n == def->in(l)) {
1233 is_loop = true;
1234 break; // Some kind of loop
1235 }
1236 }
1237 }
1238 assert(is_loop || block->find_node(def) < j, "uses must follow definitions");
1239 }
1240 }
1241 }
1242 }
1243
1244 j = block->end_idx();
1245 Node* bp = (Node*)block->get_node(block->number_of_nodes() - 1)->is_block_proj();
1246 assert(bp, "last instruction must be a block proj");
1247 assert(bp == block->get_node(j), "wrong number of successors for this block");
1248 if (bp->is_Catch()) {
1249 while (block->get_node(--j)->is_MachProj()) {
1250 ;
1251 }
1252 assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call");
1253 } else if (bp->is_Mach() && bp->as_Mach()->ideal_Opcode() == Op_If) {
1254 assert(block->_num_succs == 2, "Conditional branch must have two targets");
1255 }
1256 }
1257 #endif
1258 }
1259 #endif
1260
1261 UnionFind::UnionFind( uint max ) : _cnt(max), _max(max), _indices(NEW_RESOURCE_ARRAY(uint,max)) {
1262 Copy::zero_to_bytes( _indices, sizeof(uint)*max );
1263 }
1264
1265 void UnionFind::extend( uint from_idx, uint to_idx ) {
1266 _nesting.check();
1267 if( from_idx >= _max ) {
1268 uint size = 16;
1269 while( size <= from_idx ) size <<=1;
1270 _indices = REALLOC_RESOURCE_ARRAY( uint, _indices, _max, size );
1271 _max = size;
1272 }
1273 while( _cnt <= from_idx ) _indices[_cnt++] = 0;
1274 _indices[from_idx] = to_idx;
1275 }
1276
1277 void UnionFind::reset( uint max ) {
1278 // Force the Union-Find mapping to be at least this large
1279 extend(max,0);
1280 // Initialize to be the ID mapping.
1281 for( uint i=0; i<max; i++ ) map(i,i);
1282 }
1283
1284 // Straight out of Tarjan's union-find algorithm
1285 uint UnionFind::Find_compress( uint idx ) {
1286 uint cur = idx;
1287 uint next = lookup(cur);
1288 while( next != cur ) { // Scan chain of equivalences
1289 assert( next < cur, "always union smaller" );
1290 cur = next; // until find a fixed-point
1291 next = lookup(cur);
1292 }
1293 // Core of union-find algorithm: update chain of
1294 // equivalences to be equal to the root.
1295 while( idx != next ) {
1296 uint tmp = lookup(idx);
1297 map(idx, next);
1298 idx = tmp;
1299 }
1300 return idx;
1301 }
1302
1303 // Like Find above, but no path compress, so bad asymptotic behavior
1304 uint UnionFind::Find_const( uint idx ) const {
1305 if( idx == 0 ) return idx; // Ignore the zero idx
1306 // Off the end? This can happen during debugging dumps
1307 // when data structures have not finished being updated.
1308 if( idx >= _max ) return idx;
1309 uint next = lookup(idx);
1310 while( next != idx ) { // Scan chain of equivalences
1311 idx = next; // until find a fixed-point
1312 next = lookup(idx);
1313 }
1314 return next;
1315 }
1316
1317 // union 2 sets together.
1318 void UnionFind::Union( uint idx1, uint idx2 ) {
1319 uint src = Find(idx1);
1320 uint dst = Find(idx2);
1321 assert( src, "" );
1322 assert( dst, "" );
1323 assert( src < _max, "oob" );
1324 assert( dst < _max, "oob" );
1325 assert( src < dst, "always union smaller" );
1326 map(dst,src);
1327 }
1328
1329 #ifndef PRODUCT
1330 void Trace::dump( ) const {
1331 tty->print_cr("Trace (freq %f)", first_block()->_freq);
1332 for (Block *b = first_block(); b != NULL; b = next(b)) {
1333 tty->print(" B%d", b->_pre_order);
1334 if (b->head()->is_Loop()) {
1335 tty->print(" (L%d)", b->compute_loop_alignment());
1336 }
1337 if (b->has_loop_alignment()) {
1338 tty->print(" (T%d)", b->code_alignment());
1339 }
1340 }
1341 tty->cr();
1342 }
1343
1344 void CFGEdge::dump( ) const {
1345 tty->print(" B%d --> B%d Freq: %f out:%3d%% in:%3d%% State: ",
1346 from()->_pre_order, to()->_pre_order, freq(), _from_pct, _to_pct);
1347 switch(state()) {
1348 case connected:
1349 tty->print("connected");
1350 break;
1351 case open:
1352 tty->print("open");
1353 break;
1354 case interior:
1355 tty->print("interior");
1356 break;
1357 }
1358 if (infrequent()) {
1359 tty->print(" infrequent");
1360 }
1361 tty->cr();
1362 }
1363 #endif
1364
1365 // Comparison function for edges
1366 static int edge_order(CFGEdge **e0, CFGEdge **e1) {
1367 float freq0 = (*e0)->freq();
1368 float freq1 = (*e1)->freq();
1369 if (freq0 != freq1) {
1370 return freq0 > freq1 ? -1 : 1;
1371 }
1372
1373 int dist0 = (*e0)->to()->_rpo - (*e0)->from()->_rpo;
1374 int dist1 = (*e1)->to()->_rpo - (*e1)->from()->_rpo;
1375
1376 return dist1 - dist0;
1377 }
1378
1379 // Comparison function for edges
1380 extern "C" int trace_frequency_order(const void *p0, const void *p1) {
1381 Trace *tr0 = *(Trace **) p0;
1382 Trace *tr1 = *(Trace **) p1;
1383 Block *b0 = tr0->first_block();
1384 Block *b1 = tr1->first_block();
1385
1386 // The trace of connector blocks goes at the end;
1387 // we only expect one such trace
1388 if (b0->is_connector() != b1->is_connector()) {
1389 return b1->is_connector() ? -1 : 1;
1390 }
1391
1392 // Pull more frequently executed blocks to the beginning
1393 float freq0 = b0->_freq;
1394 float freq1 = b1->_freq;
1395 if (freq0 != freq1) {
1396 return freq0 > freq1 ? -1 : 1;
1397 }
1398
1399 int diff = tr0->first_block()->_rpo - tr1->first_block()->_rpo;
1400
1401 return diff;
1402 }
1403
1404 // Find edges of interest, i.e, those which can fall through. Presumes that
1405 // edges which don't fall through are of low frequency and can be generally
1406 // ignored. Initialize the list of traces.
1407 void PhaseBlockLayout::find_edges() {
1408 // Walk the blocks, creating edges and Traces
1409 uint i;
1410 Trace *tr = NULL;
1411 for (i = 0; i < _cfg.number_of_blocks(); i++) {
1412 Block* b = _cfg.get_block(i);
1413 tr = new Trace(b, next, prev);
1414 traces[tr->id()] = tr;
1415
1416 // All connector blocks should be at the end of the list
1417 if (b->is_connector()) break;
1418
1419 // If this block and the next one have a one-to-one successor
1420 // predecessor relationship, simply append the next block
1421 int nfallthru = b->num_fall_throughs();
1422 while (nfallthru == 1 &&
1423 b->succ_fall_through(0)) {
1424 Block *n = b->_succs[0];
1425
1426 // Skip over single-entry connector blocks, we don't want to
1427 // add them to the trace.
1428 while (n->is_connector() && n->num_preds() == 1) {
1429 n = n->_succs[0];
1430 }
1431
1432 // We see a merge point, so stop search for the next block
1433 if (n->num_preds() != 1) break;
1434
1435 i++;
1436 assert(n == _cfg.get_block(i), "expecting next block");
1437 tr->append(n);
1438 uf->map(n->_pre_order, tr->id());
1439 traces[n->_pre_order] = NULL;
1440 nfallthru = b->num_fall_throughs();
1441 b = n;
1442 }
1443
1444 if (nfallthru > 0) {
1445 // Create a CFGEdge for each outgoing
1446 // edge that could be a fall-through.
1447 for (uint j = 0; j < b->_num_succs; j++ ) {
1448 if (b->succ_fall_through(j)) {
1449 Block *target = b->non_connector_successor(j);
1450 float freq = b->_freq * b->succ_prob(j);
1451 int from_pct = (int) ((100 * freq) / b->_freq);
1452 int to_pct = (int) ((100 * freq) / target->_freq);
1453 edges->append(new CFGEdge(b, target, freq, from_pct, to_pct));
1454 }
1455 }
1456 }
1457 }
1458
1459 // Group connector blocks into one trace
1460 for (i++; i < _cfg.number_of_blocks(); i++) {
1461 Block *b = _cfg.get_block(i);
1462 assert(b->is_connector(), "connector blocks at the end");
1463 tr->append(b);
1464 uf->map(b->_pre_order, tr->id());
1465 traces[b->_pre_order] = NULL;
1466 }
1467 }
1468
1469 // Union two traces together in uf, and null out the trace in the list
1470 void PhaseBlockLayout::union_traces(Trace* updated_trace, Trace* old_trace) {
1471 uint old_id = old_trace->id();
1472 uint updated_id = updated_trace->id();
1473
1474 uint lo_id = updated_id;
1475 uint hi_id = old_id;
1476
1477 // If from is greater than to, swap values to meet
1478 // UnionFind guarantee.
1479 if (updated_id > old_id) {
1480 lo_id = old_id;
1481 hi_id = updated_id;
1482
1483 // Fix up the trace ids
1484 traces[lo_id] = traces[updated_id];
1485 updated_trace->set_id(lo_id);
1486 }
1487
1488 // Union the lower with the higher and remove the pointer
1489 // to the higher.
1490 uf->Union(lo_id, hi_id);
1491 traces[hi_id] = NULL;
1492 }
1493
1494 // Append traces together via the most frequently executed edges
1495 void PhaseBlockLayout::grow_traces() {
1496 // Order the edges, and drive the growth of Traces via the most
1497 // frequently executed edges.
1498 edges->sort(edge_order);
1499 for (int i = 0; i < edges->length(); i++) {
1500 CFGEdge *e = edges->at(i);
1501
1502 if (e->state() != CFGEdge::open) continue;
1503
1504 Block *src_block = e->from();
1505 Block *targ_block = e->to();
1506
1507 // Don't grow traces along backedges?
1508 if (!BlockLayoutRotateLoops) {
1509 if (targ_block->_rpo <= src_block->_rpo) {
1510 targ_block->set_loop_alignment(targ_block);
1511 continue;
1512 }
1513 }
1514
1515 Trace *src_trace = trace(src_block);
1516 Trace *targ_trace = trace(targ_block);
1517
1518 // If the edge in question can join two traces at their ends,
1519 // append one trace to the other.
1520 if (src_trace->last_block() == src_block) {
1521 if (src_trace == targ_trace) {
1522 e->set_state(CFGEdge::interior);
1523 if (targ_trace->backedge(e)) {
1524 // Reset i to catch any newly eligible edge
1525 // (Or we could remember the first "open" edge, and reset there)
1526 i = 0;
1527 }
1528 } else if (targ_trace->first_block() == targ_block) {
1529 e->set_state(CFGEdge::connected);
1530 src_trace->append(targ_trace);
1531 union_traces(src_trace, targ_trace);
1532 }
1533 }
1534 }
1535 }
1536
1537 // Embed one trace into another, if the fork or join points are sufficiently
1538 // balanced.
1539 void PhaseBlockLayout::merge_traces(bool fall_thru_only) {
1540 // Walk the edge list a another time, looking at unprocessed edges.
1541 // Fold in diamonds
1542 for (int i = 0; i < edges->length(); i++) {
1543 CFGEdge *e = edges->at(i);
1544
1545 if (e->state() != CFGEdge::open) continue;
1546 if (fall_thru_only) {
1547 if (e->infrequent()) continue;
1548 }
1549
1550 Block *src_block = e->from();
1551 Trace *src_trace = trace(src_block);
1552 bool src_at_tail = src_trace->last_block() == src_block;
1553
1554 Block *targ_block = e->to();
1555 Trace *targ_trace = trace(targ_block);
1556 bool targ_at_start = targ_trace->first_block() == targ_block;
1557
1558 if (src_trace == targ_trace) {
1559 // This may be a loop, but we can't do much about it.
1560 e->set_state(CFGEdge::interior);
1561 continue;
1562 }
1563
1564 if (fall_thru_only) {
1565 // If the edge links the middle of two traces, we can't do anything.
1566 // Mark the edge and continue.
1567 if (!src_at_tail & !targ_at_start) {
1568 continue;
1569 }
1570
1571 // Don't grow traces along backedges?
1572 if (!BlockLayoutRotateLoops && (targ_block->_rpo <= src_block->_rpo)) {
1573 continue;
1574 }
1575
1576 // If both ends of the edge are available, why didn't we handle it earlier?
1577 assert(src_at_tail ^ targ_at_start, "Should have caught this edge earlier.");
1578
1579 if (targ_at_start) {
1580 // Insert the "targ" trace in the "src" trace if the insertion point
1581 // is a two way branch.
1582 // Better profitability check possible, but may not be worth it.
1583 // Someday, see if the this "fork" has an associated "join";
1584 // then make a policy on merging this trace at the fork or join.
1585 // For example, other things being equal, it may be better to place this
1586 // trace at the join point if the "src" trace ends in a two-way, but
1587 // the insertion point is one-way.
1588 assert(src_block->num_fall_throughs() == 2, "unexpected diamond");
1589 e->set_state(CFGEdge::connected);
1590 src_trace->insert_after(src_block, targ_trace);
1591 union_traces(src_trace, targ_trace);
1592 } else if (src_at_tail) {
1593 if (src_trace != trace(_cfg.get_root_block())) {
1594 e->set_state(CFGEdge::connected);
1595 targ_trace->insert_before(targ_block, src_trace);
1596 union_traces(targ_trace, src_trace);
1597 }
1598 }
1599 } else if (e->state() == CFGEdge::open) {
1600 // Append traces, even without a fall-thru connection.
1601 // But leave root entry at the beginning of the block list.
1602 if (targ_trace != trace(_cfg.get_root_block())) {
1603 e->set_state(CFGEdge::connected);
1604 src_trace->append(targ_trace);
1605 union_traces(src_trace, targ_trace);
1606 }
1607 }
1608 }
1609 }
1610
1611 // Order the sequence of the traces in some desirable way, and fixup the
1612 // jumps at the end of each block.
1613 void PhaseBlockLayout::reorder_traces(int count) {
1614 ResourceArea *area = Thread::current()->resource_area();
1615 Trace ** new_traces = NEW_ARENA_ARRAY(area, Trace *, count);
1616 Block_List worklist;
1617 int new_count = 0;
1618
1619 // Compact the traces.
1620 for (int i = 0; i < count; i++) {
1621 Trace *tr = traces[i];
1622 if (tr != NULL) {
1623 new_traces[new_count++] = tr;
1624 }
1625 }
1626
1627 // The entry block should be first on the new trace list.
1628 Trace *tr = trace(_cfg.get_root_block());
1629 assert(tr == new_traces[0], "entry trace misplaced");
1630
1631 // Sort the new trace list by frequency
1632 qsort(new_traces + 1, new_count - 1, sizeof(new_traces[0]), trace_frequency_order);
1633
1634 // Patch up the successor blocks
1635 _cfg.clear_blocks();
1636 for (int i = 0; i < new_count; i++) {
1637 Trace *tr = new_traces[i];
1638 if (tr != NULL) {
1639 tr->fixup_blocks(_cfg);
1640 }
1641 }
1642 }
1643
1644 // Order basic blocks based on frequency
1645 PhaseBlockLayout::PhaseBlockLayout(PhaseCFG &cfg)
1646 : Phase(BlockLayout)
1647 , _cfg(cfg) {
1648 ResourceMark rm;
1649 ResourceArea *area = Thread::current()->resource_area();
1650
1651 // List of traces
1652 int size = _cfg.number_of_blocks() + 1;
1653 traces = NEW_ARENA_ARRAY(area, Trace *, size);
1654 memset(traces, 0, size*sizeof(Trace*));
1655 next = NEW_ARENA_ARRAY(area, Block *, size);
1656 memset(next, 0, size*sizeof(Block *));
1657 prev = NEW_ARENA_ARRAY(area, Block *, size);
1658 memset(prev , 0, size*sizeof(Block *));
1659
1660 // List of edges
1661 edges = new GrowableArray<CFGEdge*>;
1662
1663 // Mapping block index --> block_trace
1664 uf = new UnionFind(size);
1665 uf->reset(size);
1666
1667 // Find edges and create traces.
1668 find_edges();
1669
1670 // Grow traces at their ends via most frequent edges.
1671 grow_traces();
1672
1673 // Merge one trace into another, but only at fall-through points.
1674 // This may make diamonds and other related shapes in a trace.
1675 merge_traces(true);
1676
1677 // Run merge again, allowing two traces to be catenated, even if
1678 // one does not fall through into the other. This appends loosely
1679 // related traces to be near each other.
1680 merge_traces(false);
1681
1682 // Re-order all the remaining traces by frequency
1683 reorder_traces(size);
1684
1685 assert(_cfg.number_of_blocks() >= (uint) (size - 1), "number of blocks can not shrink");
1686 }
1687
1688
1689 // Edge e completes a loop in a trace. If the target block is head of the
1690 // loop, rotate the loop block so that the loop ends in a conditional branch.
1691 bool Trace::backedge(CFGEdge *e) {
1692 bool loop_rotated = false;
1693 Block *src_block = e->from();
1694 Block *targ_block = e->to();
1695
1696 assert(last_block() == src_block, "loop discovery at back branch");
1697 if (first_block() == targ_block) {
1698 if (BlockLayoutRotateLoops && last_block()->num_fall_throughs() < 2) {
1699 // Find the last block in the trace that has a conditional
1700 // branch.
1701 Block *b;
1702 for (b = last_block(); b != NULL; b = prev(b)) {
1703 if (b->num_fall_throughs() == 2) {
1704 break;
1705 }
1706 }
1707
1708 if (b != last_block() && b != NULL) {
1709 loop_rotated = true;
1710
1711 // Rotate the loop by doing two-part linked-list surgery.
1712 append(first_block());
1713 break_loop_after(b);
1714 }
1715 }
1716
1717 // Backbranch to the top of a trace
1718 // Scroll forward through the trace from the targ_block. If we find
1719 // a loop head before another loop top, use the the loop head alignment.
1720 for (Block *b = targ_block; b != NULL; b = next(b)) {
1721 if (b->has_loop_alignment()) {
1722 break;
1723 }
1724 if (b->head()->is_Loop()) {
1725 targ_block = b;
1726 break;
1727 }
1728 }
1729
1730 first_block()->set_loop_alignment(targ_block);
1731
1732 } else {
1733 // That loop may already have a loop top (we're reaching it again
1734 // through the backedge of an outer loop)
1735 Block* b = prev(targ_block);
1736 bool has_top = targ_block->head()->is_Loop() && b->has_loop_alignment() && !b->head()->is_Loop();
1737 if (!has_top) {
1738 // Backbranch into the middle of a trace
1739 targ_block->set_loop_alignment(targ_block);
1740 }
1741 }
1742
1743 return loop_rotated;
1744 }
1745
1746 // push blocks onto the CFG list
1747 // ensure that blocks have the correct two-way branch sense
1748 void Trace::fixup_blocks(PhaseCFG &cfg) {
1749 Block *last = last_block();
1750 for (Block *b = first_block(); b != NULL; b = next(b)) {
1751 cfg.add_block(b);
1752 if (!b->is_connector()) {
1753 int nfallthru = b->num_fall_throughs();
1754 if (b != last) {
1755 if (nfallthru == 2) {
1756 // Ensure that the sense of the branch is correct
1757 Block *bnext = next(b);
1758 Block *bs0 = b->non_connector_successor(0);
1759
1760 MachNode *iff = b->get_node(b->number_of_nodes() - 3)->as_Mach();
1761 ProjNode *proj0 = b->get_node(b->number_of_nodes() - 2)->as_Proj();
1762 ProjNode *proj1 = b->get_node(b->number_of_nodes() - 1)->as_Proj();
1763
1764 if (bnext == bs0) {
1765 // Fall-thru case in succs[0], should be in succs[1]
1766
1767 // Flip targets in _succs map
1768 Block *tbs0 = b->_succs[0];
1769 Block *tbs1 = b->_succs[1];
1770 b->_succs.map( 0, tbs1 );
1771 b->_succs.map( 1, tbs0 );
1772
1773 // Flip projections to match targets
1774 b->map_node(proj1, b->number_of_nodes() - 2);
1775 b->map_node(proj0, b->number_of_nodes() - 1);
1776 }
1777 }
1778 }
1779 }
1780 }
1781 }