1 /*
2 * Copyright (c) 1997, 2014, 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 "opto/block.hpp"
29 #include "opto/cfgnode.hpp"
30 #include "opto/chaitin.hpp"
31 #include "opto/loopnode.hpp"
32 #include "opto/machnode.hpp"
33 #include "opto/matcher.hpp"
34 #include "opto/opcodes.hpp"
35 #include "opto/rootnode.hpp"
36 #include "utilities/copy.hpp"
37
38 void Block_Array::grow( uint i ) {
39 assert(i >= Max(), "must be an overflow");
40 debug_only(_limit = i+1);
41 if( i < _size ) return;
42 if( !_size ) {
43 _size = 1;
44 _blocks = (Block**)_arena->Amalloc( _size * sizeof(Block*) );
45 _blocks[0] = NULL;
46 }
47 uint old = _size;
48 while( i >= _size ) _size <<= 1; // Double to fit
49 _blocks = (Block**)_arena->Arealloc( _blocks, old*sizeof(Block*),_size*sizeof(Block*));
50 Copy::zero_to_bytes( &_blocks[old], (_size-old)*sizeof(Block*) );
51 }
52
53 void Block_List::remove(uint i) {
54 assert(i < _cnt, "index out of bounds");
55 Copy::conjoint_words_to_lower((HeapWord*)&_blocks[i+1], (HeapWord*)&_blocks[i], ((_cnt-i-1)*sizeof(Block*)));
56 pop(); // shrink list by one block
57 }
58
59 void Block_List::insert(uint i, Block *b) {
60 push(b); // grow list by one block
61 Copy::conjoint_words_to_higher((HeapWord*)&_blocks[i], (HeapWord*)&_blocks[i+1], ((_cnt-i-1)*sizeof(Block*)));
62 _blocks[i] = b;
63 }
64
65 #ifndef PRODUCT
66 void Block_List::print() {
67 for (uint i=0; i < size(); i++) {
68 tty->print("B%d ", _blocks[i]->_pre_order);
69 }
70 tty->print("size = %d\n", size());
71 }
72 #endif
73
74 uint Block::code_alignment() {
75 // Check for Root block
76 if (_pre_order == 0) return CodeEntryAlignment;
77 // Check for Start block
78 if (_pre_order == 1) return InteriorEntryAlignment;
79 // Check for loop alignment
80 if (has_loop_alignment()) return loop_alignment();
81
82 return relocInfo::addr_unit(); // no particular alignment
83 }
84
85 uint Block::compute_loop_alignment() {
86 Node *h = head();
87 int unit_sz = relocInfo::addr_unit();
88 if (h->is_Loop() && h->as_Loop()->is_inner_loop()) {
89 // Pre- and post-loops have low trip count so do not bother with
90 // NOPs for align loop head. The constants are hidden from tuning
91 // but only because my "divide by 4" heuristic surely gets nearly
92 // all possible gain (a "do not align at all" heuristic has a
93 // chance of getting a really tiny gain).
94 if (h->is_CountedLoop() && (h->as_CountedLoop()->is_pre_loop() ||
95 h->as_CountedLoop()->is_post_loop())) {
96 return (OptoLoopAlignment > 4*unit_sz) ? (OptoLoopAlignment>>2) : unit_sz;
97 }
98 // Loops with low backedge frequency should not be aligned.
99 Node *n = h->in(LoopNode::LoopBackControl)->in(0);
100 if (n->is_MachIf() && n->as_MachIf()->_prob < 0.01) {
101 return unit_sz; // Loop does not loop, more often than not!
102 }
103 return OptoLoopAlignment; // Otherwise align loop head
104 }
105
106 return unit_sz; // no particular alignment
107 }
108
109 // Compute the size of first 'inst_cnt' instructions in this block.
110 // Return the number of instructions left to compute if the block has
111 // less then 'inst_cnt' instructions. Stop, and return 0 if sum_size
112 // exceeds OptoLoopAlignment.
113 uint Block::compute_first_inst_size(uint& sum_size, uint inst_cnt,
114 PhaseRegAlloc* ra) {
115 uint last_inst = number_of_nodes();
116 for( uint j = 0; j < last_inst && inst_cnt > 0; j++ ) {
117 uint inst_size = get_node(j)->size(ra);
118 if( inst_size > 0 ) {
119 inst_cnt--;
120 uint sz = sum_size + inst_size;
121 if( sz <= (uint)OptoLoopAlignment ) {
122 // Compute size of instructions which fit into fetch buffer only
123 // since all inst_cnt instructions will not fit even if we align them.
124 sum_size = sz;
125 } else {
126 return 0;
127 }
128 }
129 }
130 return inst_cnt;
131 }
132
133 uint Block::find_node( const Node *n ) const {
134 for( uint i = 0; i < number_of_nodes(); i++ ) {
135 if( get_node(i) == n )
136 return i;
137 }
138 ShouldNotReachHere();
139 return 0;
140 }
141
142 // Find and remove n from block list
143 void Block::find_remove( const Node *n ) {
144 remove_node(find_node(n));
145 }
146
147 bool Block::contains(const Node *n) const {
148 return _nodes.contains(n);
149 }
150
151 // Return empty status of a block. Empty blocks contain only the head, other
152 // ideal nodes, and an optional trailing goto.
153 int Block::is_Empty() const {
154
155 // Root or start block is not considered empty
156 if (head()->is_Root() || head()->is_Start()) {
157 return not_empty;
158 }
159
160 int success_result = completely_empty;
161 int end_idx = number_of_nodes() - 1;
162
163 // Check for ending goto
164 if ((end_idx > 0) && (get_node(end_idx)->is_MachGoto())) {
165 success_result = empty_with_goto;
166 end_idx--;
167 }
168
169 // Unreachable blocks are considered empty
170 if (num_preds() <= 1) {
171 return success_result;
172 }
173
174 // Ideal nodes are allowable in empty blocks: skip them Only MachNodes
175 // turn directly into code, because only MachNodes have non-trivial
176 // emit() functions.
177 while ((end_idx > 0) && !get_node(end_idx)->is_Mach()) {
178 end_idx--;
179 }
180
181 // No room for any interesting instructions?
182 if (end_idx == 0) {
183 return success_result;
184 }
185
186 return not_empty;
187 }
188
189 // Return true if the block's code implies that it is likely to be
190 // executed infrequently. Check to see if the block ends in a Halt or
191 // a low probability call.
192 bool Block::has_uncommon_code() const {
193 Node* en = end();
194
195 if (en->is_MachGoto())
196 en = en->in(0);
197 if (en->is_Catch())
198 en = en->in(0);
199 if (en->is_MachProj() && en->in(0)->is_MachCall()) {
200 MachCallNode* call = en->in(0)->as_MachCall();
201 if (call->cnt() != COUNT_UNKNOWN && call->cnt() <= PROB_UNLIKELY_MAG(4)) {
202 // This is true for slow-path stubs like new_{instance,array},
203 // slow_arraycopy, complete_monitor_locking, uncommon_trap.
204 // The magic number corresponds to the probability of an uncommon_trap,
205 // even though it is a count not a probability.
206 return true;
207 }
208 }
209
210 int op = en->is_Mach() ? en->as_Mach()->ideal_Opcode() : en->Opcode();
211 return op == Op_Halt;
212 }
213
214 // True if block is low enough frequency or guarded by a test which
215 // mostly does not go here.
216 bool PhaseCFG::is_uncommon(const Block* block) {
217 // Initial blocks must never be moved, so are never uncommon.
218 if (block->head()->is_Root() || block->head()->is_Start()) return false;
219
220 // Check for way-low freq
221 if(block->_freq < BLOCK_FREQUENCY(0.00001f) ) return true;
222
223 // Look for code shape indicating uncommon_trap or slow path
224 if (block->has_uncommon_code()) return true;
225
226 const float epsilon = 0.05f;
227 const float guard_factor = PROB_UNLIKELY_MAG(4) / (1.f - epsilon);
228 uint uncommon_preds = 0;
229 uint freq_preds = 0;
230 uint uncommon_for_freq_preds = 0;
231
232 for( uint i=1; i< block->num_preds(); i++ ) {
233 Block* guard = get_block_for_node(block->pred(i));
234 // Check to see if this block follows its guard 1 time out of 10000
235 // or less.
236 //
237 // See list of magnitude-4 unlikely probabilities in cfgnode.hpp which
238 // we intend to be "uncommon", such as slow-path TLE allocation,
239 // predicted call failure, and uncommon trap triggers.
240 //
241 // Use an epsilon value of 5% to allow for variability in frequency
242 // predictions and floating point calculations. The net effect is
243 // that guard_factor is set to 9500.
244 //
245 // Ignore low-frequency blocks.
246 // The next check is (guard->_freq < 1.e-5 * 9500.).
247 if(guard->_freq*BLOCK_FREQUENCY(guard_factor) < BLOCK_FREQUENCY(0.00001f)) {
248 uncommon_preds++;
249 } else {
250 freq_preds++;
251 if(block->_freq < guard->_freq * guard_factor ) {
252 uncommon_for_freq_preds++;
253 }
254 }
255 }
256 if( block->num_preds() > 1 &&
257 // The block is uncommon if all preds are uncommon or
258 (uncommon_preds == (block->num_preds()-1) ||
259 // it is uncommon for all frequent preds.
260 uncommon_for_freq_preds == freq_preds) ) {
261 return true;
262 }
263 return false;
264 }
265
266 #ifndef PRODUCT
267 void Block::dump_bidx(const Block* orig, outputStream* st) const {
268 if (_pre_order) st->print("B%d",_pre_order);
269 else st->print("N%d", head()->_idx);
270
271 if (Verbose && orig != this) {
272 // Dump the original block's idx
273 st->print(" (");
274 orig->dump_bidx(orig, st);
275 st->print(")");
276 }
277 }
278
279 void Block::dump_pred(const PhaseCFG* cfg, Block* orig, outputStream* st) const {
280 if (is_connector()) {
281 for (uint i=1; i<num_preds(); i++) {
282 Block *p = cfg->get_block_for_node(pred(i));
283 p->dump_pred(cfg, orig, st);
284 }
285 } else {
286 dump_bidx(orig, st);
287 st->print(" ");
288 }
289 }
290
291 void Block::dump_head(const PhaseCFG* cfg, outputStream* st) const {
292 // Print the basic block
293 dump_bidx(this, st);
294 st->print(": #\t");
295
296 // Print the incoming CFG edges and the outgoing CFG edges
297 for( uint i=0; i<_num_succs; i++ ) {
298 non_connector_successor(i)->dump_bidx(_succs[i], st);
299 st->print(" ");
300 }
301 st->print("<- ");
302 if( head()->is_block_start() ) {
303 for (uint i=1; i<num_preds(); i++) {
304 Node *s = pred(i);
305 if (cfg != NULL) {
306 Block *p = cfg->get_block_for_node(s);
307 p->dump_pred(cfg, p, st);
308 } else {
309 while (!s->is_block_start())
310 s = s->in(0);
311 st->print("N%d ", s->_idx );
312 }
313 }
314 } else {
315 st->print("BLOCK HEAD IS JUNK ");
316 }
317
318 // Print loop, if any
319 const Block *bhead = this; // Head of self-loop
320 Node *bh = bhead->head();
321
322 if ((cfg != NULL) && bh->is_Loop() && !head()->is_Root()) {
323 LoopNode *loop = bh->as_Loop();
324 const Block *bx = cfg->get_block_for_node(loop->in(LoopNode::LoopBackControl));
325 while (bx->is_connector()) {
326 bx = cfg->get_block_for_node(bx->pred(1));
327 }
328 st->print("\tLoop: B%d-B%d ", bhead->_pre_order, bx->_pre_order);
329 // Dump any loop-specific bits, especially for CountedLoops.
330 loop->dump_spec(st);
331 } else if (has_loop_alignment()) {
332 st->print(" top-of-loop");
333 }
334 st->print(" Freq: %g",_freq);
335 if( Verbose || WizardMode ) {
336 st->print(" IDom: %d/#%d", _idom ? _idom->_pre_order : 0, _dom_depth);
337 st->print(" RegPressure: %d",_reg_pressure);
338 st->print(" IHRP Index: %d",_ihrp_index);
339 st->print(" FRegPressure: %d",_freg_pressure);
340 st->print(" FHRP Index: %d",_fhrp_index);
341 }
342 st->cr();
343 }
344
345 void Block::dump() const {
346 dump(NULL);
347 }
348
349 void Block::dump(const PhaseCFG* cfg) const {
350 dump_head(cfg);
351 for (uint i=0; i< number_of_nodes(); i++) {
352 get_node(i)->dump();
353 }
354 tty->print("\n");
355 }
356 #endif
357
358 PhaseCFG::PhaseCFG(Arena* arena, RootNode* root, Matcher& matcher)
359 : Phase(CFG)
360 , _block_arena(arena)
361 , _root(root)
362 , _matcher(matcher)
363 , _node_to_block_mapping(arena)
364 , _node_latency(NULL)
365 #ifndef PRODUCT
366 , _trace_opto_pipelining(TraceOptoPipelining || C->method_has_option("TraceOptoPipelining"))
367 #endif
368 #ifdef ASSERT
369 , _raw_oops(arena)
370 #endif
371 {
372 ResourceMark rm;
373 // I'll need a few machine-specific GotoNodes. Make an Ideal GotoNode,
374 // then Match it into a machine-specific Node. Then clone the machine
375 // Node on demand.
376 Node *x = new GotoNode(NULL);
377 x->init_req(0, x);
378 _goto = matcher.match_tree(x);
379 assert(_goto != NULL, "");
380 _goto->set_req(0,_goto);
381
382 // Build the CFG in Reverse Post Order
383 _number_of_blocks = build_cfg();
384 _root_block = get_block_for_node(_root);
385 }
386
387 // Build a proper looking CFG. Make every block begin with either a StartNode
388 // or a RegionNode. Make every block end with either a Goto, If or Return.
389 // The RootNode both starts and ends it's own block. Do this with a recursive
390 // backwards walk over the control edges.
391 uint PhaseCFG::build_cfg() {
392 Arena *a = Thread::current()->resource_area();
393 VectorSet visited(a);
394
395 // Allocate stack with enough space to avoid frequent realloc
396 Node_Stack nstack(a, C->live_nodes() >> 1);
397 nstack.push(_root, 0);
398 uint sum = 0; // Counter for blocks
399
400 while (nstack.is_nonempty()) {
401 // node and in's index from stack's top
402 // 'np' is _root (see above) or RegionNode, StartNode: we push on stack
403 // only nodes which point to the start of basic block (see below).
404 Node *np = nstack.node();
405 // idx > 0, except for the first node (_root) pushed on stack
406 // at the beginning when idx == 0.
407 // We will use the condition (idx == 0) later to end the build.
408 uint idx = nstack.index();
409 Node *proj = np->in(idx);
410 const Node *x = proj->is_block_proj();
411 // Does the block end with a proper block-ending Node? One of Return,
412 // If or Goto? (This check should be done for visited nodes also).
413 if (x == NULL) { // Does not end right...
414 Node *g = _goto->clone(); // Force it to end in a Goto
415 g->set_req(0, proj);
416 np->set_req(idx, g);
417 x = proj = g;
418 }
419 if (!visited.test_set(x->_idx)) { // Visit this block once
420 // Skip any control-pinned middle'in stuff
421 Node *p = proj;
422 do {
423 proj = p; // Update pointer to last Control
424 p = p->in(0); // Move control forward
425 } while( !p->is_block_proj() &&
426 !p->is_block_start() );
427 // Make the block begin with one of Region or StartNode.
428 if( !p->is_block_start() ) {
429 RegionNode *r = new RegionNode( 2 );
430 r->init_req(1, p); // Insert RegionNode in the way
431 proj->set_req(0, r); // Insert RegionNode in the way
432 p = r;
433 }
434 // 'p' now points to the start of this basic block
435
436 // Put self in array of basic blocks
437 Block *bb = new (_block_arena) Block(_block_arena, p);
438 map_node_to_block(p, bb);
439 map_node_to_block(x, bb);
440 if( x != p ) { // Only for root is x == p
441 bb->push_node((Node*)x);
442 }
443 // Now handle predecessors
444 ++sum; // Count 1 for self block
445 uint cnt = bb->num_preds();
446 for (int i = (cnt - 1); i > 0; i-- ) { // For all predecessors
447 Node *prevproj = p->in(i); // Get prior input
448 assert( !prevproj->is_Con(), "dead input not removed" );
449 // Check to see if p->in(i) is a "control-dependent" CFG edge -
450 // i.e., it splits at the source (via an IF or SWITCH) and merges
451 // at the destination (via a many-input Region).
452 // This breaks critical edges. The RegionNode to start the block
453 // will be added when <p,i> is pulled off the node stack
454 if ( cnt > 2 ) { // Merging many things?
455 assert( prevproj== bb->pred(i),"");
456 if(prevproj->is_block_proj() != prevproj) { // Control-dependent edge?
457 // Force a block on the control-dependent edge
458 Node *g = _goto->clone(); // Force it to end in a Goto
459 g->set_req(0,prevproj);
460 p->set_req(i,g);
461 }
462 }
463 nstack.push(p, i); // 'p' is RegionNode or StartNode
464 }
465 } else { // Post-processing visited nodes
466 nstack.pop(); // remove node from stack
467 // Check if it the fist node pushed on stack at the beginning.
468 if (idx == 0) break; // end of the build
469 // Find predecessor basic block
470 Block *pb = get_block_for_node(x);
471 // Insert into nodes array, if not already there
472 if (!has_block(proj)) {
473 assert( x != proj, "" );
474 // Map basic block of projection
475 map_node_to_block(proj, pb);
476 pb->push_node(proj);
477 }
478 // Insert self as a child of my predecessor block
479 pb->_succs.map(pb->_num_succs++, get_block_for_node(np));
480 assert( pb->get_node(pb->number_of_nodes() - pb->_num_succs)->is_block_proj(),
481 "too many control users, not a CFG?" );
482 }
483 }
484 // Return number of basic blocks for all children and self
485 return sum;
486 }
487
488 // Inserts a goto & corresponding basic block between
489 // block[block_no] and its succ_no'th successor block
490 void PhaseCFG::insert_goto_at(uint block_no, uint succ_no) {
491 // get block with block_no
492 assert(block_no < number_of_blocks(), "illegal block number");
493 Block* in = get_block(block_no);
494 // get successor block succ_no
495 assert(succ_no < in->_num_succs, "illegal successor number");
496 Block* out = in->_succs[succ_no];
497 // Compute frequency of the new block. Do this before inserting
498 // new block in case succ_prob() needs to infer the probability from
499 // surrounding blocks.
500 float freq = in->_freq * in->succ_prob(succ_no);
501 // get ProjNode corresponding to the succ_no'th successor of the in block
502 ProjNode* proj = in->get_node(in->number_of_nodes() - in->_num_succs + succ_no)->as_Proj();
503 // create region for basic block
504 RegionNode* region = new RegionNode(2);
505 region->init_req(1, proj);
506 // setup corresponding basic block
507 Block* block = new (_block_arena) Block(_block_arena, region);
508 map_node_to_block(region, block);
509 C->regalloc()->set_bad(region->_idx);
510 // add a goto node
511 Node* gto = _goto->clone(); // get a new goto node
512 gto->set_req(0, region);
513 // add it to the basic block
514 block->push_node(gto);
515 map_node_to_block(gto, block);
516 C->regalloc()->set_bad(gto->_idx);
517 // hook up successor block
518 block->_succs.map(block->_num_succs++, out);
519 // remap successor's predecessors if necessary
520 for (uint i = 1; i < out->num_preds(); i++) {
521 if (out->pred(i) == proj) out->head()->set_req(i, gto);
522 }
523 // remap predecessor's successor to new block
524 in->_succs.map(succ_no, block);
525 // Set the frequency of the new block
526 block->_freq = freq;
527 // add new basic block to basic block list
528 add_block_at(block_no + 1, block);
529 }
530
531 // Does this block end in a multiway branch that cannot have the default case
532 // flipped for another case?
533 static bool no_flip_branch(Block *b) {
534 int branch_idx = b->number_of_nodes() - b->_num_succs-1;
535 if (branch_idx < 1) {
536 return false;
537 }
538 Node *branch = b->get_node(branch_idx);
539 if (branch->is_Catch()) {
540 return true;
541 }
542 if (branch->is_Mach()) {
543 if (branch->is_MachNullCheck()) {
544 return true;
545 }
546 int iop = branch->as_Mach()->ideal_Opcode();
547 if (iop == Op_FastLock || iop == Op_FastUnlock) {
548 return true;
549 }
550 // Don't flip if branch has an implicit check.
551 if (branch->as_Mach()->is_TrapBasedCheckNode()) {
552 return true;
553 }
554 }
555 return false;
556 }
557
558 // Check for NeverBranch at block end. This needs to become a GOTO to the
559 // true target. NeverBranch are treated as a conditional branch that always
560 // goes the same direction for most of the optimizer and are used to give a
561 // fake exit path to infinite loops. At this late stage they need to turn
562 // into Goto's so that when you enter the infinite loop you indeed hang.
563 void PhaseCFG::convert_NeverBranch_to_Goto(Block *b) {
564 // Find true target
565 int end_idx = b->end_idx();
566 int idx = b->get_node(end_idx+1)->as_Proj()->_con;
567 Block *succ = b->_succs[idx];
568 Node* gto = _goto->clone(); // get a new goto node
569 gto->set_req(0, b->head());
570 Node *bp = b->get_node(end_idx);
571 b->map_node(gto, end_idx); // Slam over NeverBranch
572 map_node_to_block(gto, b);
573 C->regalloc()->set_bad(gto->_idx);
574 b->pop_node(); // Yank projections
575 b->pop_node(); // Yank projections
576 b->_succs.map(0,succ); // Map only successor
577 b->_num_succs = 1;
578 // remap successor's predecessors if necessary
579 uint j;
580 for( j = 1; j < succ->num_preds(); j++)
581 if( succ->pred(j)->in(0) == bp )
582 succ->head()->set_req(j, gto);
583 // Kill alternate exit path
584 Block *dead = b->_succs[1-idx];
585 for( j = 1; j < dead->num_preds(); j++)
586 if( dead->pred(j)->in(0) == bp )
587 break;
588 // Scan through block, yanking dead path from
589 // all regions and phis.
590 dead->head()->del_req(j);
591 for( int k = 1; dead->get_node(k)->is_Phi(); k++ )
592 dead->get_node(k)->del_req(j);
593 }
594
595 // Helper function to move block bx to the slot following b_index. Return
596 // true if the move is successful, otherwise false
597 bool PhaseCFG::move_to_next(Block* bx, uint b_index) {
598 if (bx == NULL) return false;
599
600 // Return false if bx is already scheduled.
601 uint bx_index = bx->_pre_order;
602 if ((bx_index <= b_index) && (get_block(bx_index) == bx)) {
603 return false;
604 }
605
606 // Find the current index of block bx on the block list
607 bx_index = b_index + 1;
608 while (bx_index < number_of_blocks() && get_block(bx_index) != bx) {
609 bx_index++;
610 }
611 assert(get_block(bx_index) == bx, "block not found");
612
613 // If the previous block conditionally falls into bx, return false,
614 // because moving bx will create an extra jump.
615 for(uint k = 1; k < bx->num_preds(); k++ ) {
616 Block* pred = get_block_for_node(bx->pred(k));
617 if (pred == get_block(bx_index - 1)) {
618 if (pred->_num_succs != 1) {
619 return false;
620 }
621 }
622 }
623
624 // Reinsert bx just past block 'b'
625 _blocks.remove(bx_index);
626 _blocks.insert(b_index + 1, bx);
627 return true;
628 }
629
630 // Move empty and uncommon blocks to the end.
631 void PhaseCFG::move_to_end(Block *b, uint i) {
632 int e = b->is_Empty();
633 if (e != Block::not_empty) {
634 if (e == Block::empty_with_goto) {
635 // Remove the goto, but leave the block.
636 b->pop_node();
637 }
638 // Mark this block as a connector block, which will cause it to be
639 // ignored in certain functions such as non_connector_successor().
640 b->set_connector();
641 }
642 // Move the empty block to the end, and don't recheck.
643 _blocks.remove(i);
644 _blocks.push(b);
645 }
646
647 // Set loop alignment for every block
648 void PhaseCFG::set_loop_alignment() {
649 uint last = number_of_blocks();
650 assert(get_block(0) == get_root_block(), "");
651
652 for (uint i = 1; i < last; i++) {
653 Block* block = get_block(i);
654 if (block->head()->is_Loop()) {
655 block->set_loop_alignment(block);
656 }
657 }
658 }
659
660 // Make empty basic blocks to be "connector" blocks, Move uncommon blocks
661 // to the end.
662 void PhaseCFG::remove_empty_blocks() {
663 // Move uncommon blocks to the end
664 uint last = number_of_blocks();
665 assert(get_block(0) == get_root_block(), "");
666
667 for (uint i = 1; i < last; i++) {
668 Block* block = get_block(i);
669 if (block->is_connector()) {
670 break;
671 }
672
673 // Check for NeverBranch at block end. This needs to become a GOTO to the
674 // true target. NeverBranch are treated as a conditional branch that
675 // always goes the same direction for most of the optimizer and are used
676 // to give a fake exit path to infinite loops. At this late stage they
677 // need to turn into Goto's so that when you enter the infinite loop you
678 // indeed hang.
679 if (block->get_node(block->end_idx())->Opcode() == Op_NeverBranch) {
680 convert_NeverBranch_to_Goto(block);
681 }
682
683 // Look for uncommon blocks and move to end.
684 if (!C->do_freq_based_layout()) {
685 if (is_uncommon(block)) {
686 move_to_end(block, i);
687 last--; // No longer check for being uncommon!
688 if (no_flip_branch(block)) { // Fall-thru case must follow?
689 // Find the fall-thru block
690 block = get_block(i);
691 move_to_end(block, i);
692 last--;
693 }
694 // backup block counter post-increment
695 i--;
696 }
697 }
698 }
699
700 // Move empty blocks to the end
701 last = number_of_blocks();
702 for (uint i = 1; i < last; i++) {
703 Block* block = get_block(i);
704 if (block->is_Empty() != Block::not_empty) {
705 move_to_end(block, i);
706 last--;
707 i--;
708 }
709 } // End of for all blocks
710 }
711
712 Block *PhaseCFG::fixup_trap_based_check(Node *branch, Block *block, int block_pos, Block *bnext) {
713 // Trap based checks must fall through to the successor with
714 // PROB_ALWAYS.
715 // They should be an If with 2 successors.
716 assert(branch->is_MachIf(), "must be If");
717 assert(block->_num_succs == 2, "must have 2 successors");
718
719 // Get the If node and the projection for the first successor.
720 MachIfNode *iff = block->get_node(block->number_of_nodes()-3)->as_MachIf();
721 ProjNode *proj0 = block->get_node(block->number_of_nodes()-2)->as_Proj();
722 ProjNode *proj1 = block->get_node(block->number_of_nodes()-1)->as_Proj();
723 ProjNode *projt = (proj0->Opcode() == Op_IfTrue) ? proj0 : proj1;
724 ProjNode *projf = (proj0->Opcode() == Op_IfFalse) ? proj0 : proj1;
725
726 // Assert that proj0 and succs[0] match up. Similarly for proj1 and succs[1].
727 assert(proj0->raw_out(0) == block->_succs[0]->head(), "Mismatch successor 0");
728 assert(proj1->raw_out(0) == block->_succs[1]->head(), "Mismatch successor 1");
729
730 ProjNode *proj_always;
731 ProjNode *proj_never;
732 // We must negate the branch if the implicit check doesn't follow
733 // the branch's TRUE path. Then, the new TRUE branch target will
734 // be the old FALSE branch target.
735 if (iff->_prob <= 2*PROB_NEVER) { // There are small rounding errors.
736 proj_never = projt;
737 proj_always = projf;
738 } else {
739 // We must negate the branch if the trap doesn't follow the
740 // branch's TRUE path. Then, the new TRUE branch target will
741 // be the old FALSE branch target.
742 proj_never = projf;
743 proj_always = projt;
744 iff->negate();
745 }
746 assert(iff->_prob <= 2*PROB_NEVER, "Trap based checks are expected to trap never!");
747 // Map the successors properly
748 block->_succs.map(0, get_block_for_node(proj_never ->raw_out(0))); // The target of the trap.
749 block->_succs.map(1, get_block_for_node(proj_always->raw_out(0))); // The fall through target.
750
751 if (block->get_node(block->number_of_nodes() - block->_num_succs + 1) != proj_always) {
752 block->map_node(proj_never, block->number_of_nodes() - block->_num_succs + 0);
753 block->map_node(proj_always, block->number_of_nodes() - block->_num_succs + 1);
754 }
755
756 // Place the fall through block after this block.
757 Block *bs1 = block->non_connector_successor(1);
758 if (bs1 != bnext && move_to_next(bs1, block_pos)) {
759 bnext = bs1;
760 }
761 // If the fall through block still is not the next block, insert a goto.
762 if (bs1 != bnext) {
763 insert_goto_at(block_pos, 1);
764 }
765 return bnext;
766 }
767
768 // Fix up the final control flow for basic blocks.
769 void PhaseCFG::fixup_flow() {
770 // Fixup final control flow for the blocks. Remove jump-to-next
771 // block. If neither arm of an IF follows the conditional branch, we
772 // have to add a second jump after the conditional. We place the
773 // TRUE branch target in succs[0] for both GOTOs and IFs.
774 for (uint i = 0; i < number_of_blocks(); i++) {
775 Block* block = get_block(i);
776 block->_pre_order = i; // turn pre-order into block-index
777
778 // Connector blocks need no further processing.
779 if (block->is_connector()) {
780 assert((i+1) == number_of_blocks() || get_block(i + 1)->is_connector(), "All connector blocks should sink to the end");
781 continue;
782 }
783 assert(block->is_Empty() != Block::completely_empty, "Empty blocks should be connectors");
784
785 Block* bnext = (i < number_of_blocks() - 1) ? get_block(i + 1) : NULL;
786 Block* bs0 = block->non_connector_successor(0);
787
788 // Check for multi-way branches where I cannot negate the test to
789 // exchange the true and false targets.
790 if (no_flip_branch(block)) {
791 // Find fall through case - if must fall into its target.
792 // Get the index of the branch's first successor.
793 int branch_idx = block->number_of_nodes() - block->_num_succs;
794
795 // The branch is 1 before the branch's first successor.
796 Node *branch = block->get_node(branch_idx-1);
797
798 // Handle no-flip branches which have implicit checks and which require
799 // special block ordering and individual semantics of the 'fall through
800 // case'.
801 if ((TrapBasedNullChecks || TrapBasedRangeChecks) &&
802 branch->is_Mach() && branch->as_Mach()->is_TrapBasedCheckNode()) {
803 bnext = fixup_trap_based_check(branch, block, i, bnext);
804 } else {
805 // Else, default handling for no-flip branches
806 for (uint j2 = 0; j2 < block->_num_succs; j2++) {
807 const ProjNode* p = block->get_node(branch_idx + j2)->as_Proj();
808 if (p->_con == 0) {
809 // successor j2 is fall through case
810 if (block->non_connector_successor(j2) != bnext) {
811 // but it is not the next block => insert a goto
812 insert_goto_at(i, j2);
813 }
814 // Put taken branch in slot 0
815 if (j2 == 0 && block->_num_succs == 2) {
816 // Flip targets in succs map
817 Block *tbs0 = block->_succs[0];
818 Block *tbs1 = block->_succs[1];
819 block->_succs.map(0, tbs1);
820 block->_succs.map(1, tbs0);
821 }
822 break;
823 }
824 }
825 }
826
827 // Remove all CatchProjs
828 for (uint j = 0; j < block->_num_succs; j++) {
829 block->pop_node();
830 }
831
832 } else if (block->_num_succs == 1) {
833 // Block ends in a Goto?
834 if (bnext == bs0) {
835 // We fall into next block; remove the Goto
836 block->pop_node();
837 }
838
839 } else if(block->_num_succs == 2) { // Block ends in a If?
840 // Get opcode of 1st projection (matches _succs[0])
841 // Note: Since this basic block has 2 exits, the last 2 nodes must
842 // be projections (in any order), the 3rd last node must be
843 // the IfNode (we have excluded other 2-way exits such as
844 // CatchNodes already).
845 MachNode* iff = block->get_node(block->number_of_nodes() - 3)->as_Mach();
846 ProjNode* proj0 = block->get_node(block->number_of_nodes() - 2)->as_Proj();
847 ProjNode* proj1 = block->get_node(block->number_of_nodes() - 1)->as_Proj();
848
849 // Assert that proj0 and succs[0] match up. Similarly for proj1 and succs[1].
850 assert(proj0->raw_out(0) == block->_succs[0]->head(), "Mismatch successor 0");
851 assert(proj1->raw_out(0) == block->_succs[1]->head(), "Mismatch successor 1");
852
853 Block* bs1 = block->non_connector_successor(1);
854
855 // Check for neither successor block following the current
856 // block ending in a conditional. If so, move one of the
857 // successors after the current one, provided that the
858 // successor was previously unscheduled, but moveable
859 // (i.e., all paths to it involve a branch).
860 if (!C->do_freq_based_layout() && bnext != bs0 && bnext != bs1) {
861 // Choose the more common successor based on the probability
862 // of the conditional branch.
863 Block* bx = bs0;
864 Block* by = bs1;
865
866 // _prob is the probability of taking the true path. Make
867 // p the probability of taking successor #1.
868 float p = iff->as_MachIf()->_prob;
869 if (proj0->Opcode() == Op_IfTrue) {
870 p = 1.0 - p;
871 }
872
873 // Prefer successor #1 if p > 0.5
874 if (p > PROB_FAIR) {
875 bx = bs1;
876 by = bs0;
877 }
878
879 // Attempt the more common successor first
880 if (move_to_next(bx, i)) {
881 bnext = bx;
882 } else if (move_to_next(by, i)) {
883 bnext = by;
884 }
885 }
886
887 // Check for conditional branching the wrong way. Negate
888 // conditional, if needed, so it falls into the following block
889 // and branches to the not-following block.
890
891 // Check for the next block being in succs[0]. We are going to branch
892 // to succs[0], so we want the fall-thru case as the next block in
893 // succs[1].
894 if (bnext == bs0) {
895 // Fall-thru case in succs[0], so flip targets in succs map
896 Block* tbs0 = block->_succs[0];
897 Block* tbs1 = block->_succs[1];
898 block->_succs.map(0, tbs1);
899 block->_succs.map(1, tbs0);
900 // Flip projection for each target
901 ProjNode* tmp = proj0;
902 proj0 = proj1;
903 proj1 = tmp;
904
905 } else if(bnext != bs1) {
906 // Need a double-branch
907 // The existing conditional branch need not change.
908 // Add a unconditional branch to the false target.
909 // Alas, it must appear in its own block and adding a
910 // block this late in the game is complicated. Sigh.
911 insert_goto_at(i, 1);
912 }
913
914 // Make sure we TRUE branch to the target
915 if (proj0->Opcode() == Op_IfFalse) {
916 iff->as_MachIf()->negate();
917 }
918
919 block->pop_node(); // Remove IfFalse & IfTrue projections
920 block->pop_node();
921
922 } else {
923 // Multi-exit block, e.g. a switch statement
924 // But we don't need to do anything here
925 }
926 } // End of for all blocks
927 }
928
929
930 // postalloc_expand: Expand nodes after register allocation.
931 //
932 // postalloc_expand has to be called after register allocation, just
933 // before output (i.e. scheduling). It only gets called if
934 // Matcher::require_postalloc_expand is true.
935 //
936 // Background:
937 //
938 // Nodes that are expandend (one compound node requiring several
939 // assembler instructions to be implemented split into two or more
940 // non-compound nodes) after register allocation are not as nice as
941 // the ones expanded before register allocation - they don't
942 // participate in optimizations as global code motion. But after
943 // register allocation we can expand nodes that use registers which
944 // are not spillable or registers that are not allocated, because the
945 // old compound node is simply replaced (in its location in the basic
946 // block) by a new subgraph which does not contain compound nodes any
947 // more. The scheduler called during output can later on process these
948 // non-compound nodes.
949 //
950 // Implementation:
951 //
952 // Nodes requiring postalloc expand are specified in the ad file by using
953 // a postalloc_expand statement instead of ins_encode. A postalloc_expand
954 // contains a single call to an encoding, as does an ins_encode
955 // statement. Instead of an emit() function a postalloc_expand() function
956 // is generated that doesn't emit assembler but creates a new
957 // subgraph. The code below calls this postalloc_expand function for each
958 // node with the appropriate attribute. This function returns the new
959 // nodes generated in an array passed in the call. The old node,
960 // potential MachTemps before and potential Projs after it then get
961 // disconnected and replaced by the new nodes. The instruction
962 // generating the result has to be the last one in the array. In
963 // general it is assumed that Projs after the node expanded are
964 // kills. These kills are not required any more after expanding as
965 // there are now explicitly visible def-use chains and the Projs are
966 // removed. This does not hold for calls: They do not only have
967 // kill-Projs but also Projs defining values. Therefore Projs after
968 // the node expanded are removed for all but for calls. If a node is
969 // to be reused, it must be added to the nodes list returned, and it
970 // will be added again.
971 //
972 // Implementing the postalloc_expand function for a node in an enc_class
973 // is rather tedious. It requires knowledge about many node details, as
974 // the nodes and the subgraph must be hand crafted. To simplify this,
975 // adlc generates some utility variables into the postalloc_expand function,
976 // e.g., holding the operands as specified by the postalloc_expand encoding
977 // specification, e.g.:
978 // * unsigned idx_<par_name> holding the index of the node in the ins
979 // * Node *n_<par_name> holding the node loaded from the ins
980 // * MachOpnd *op_<par_name> holding the corresponding operand
981 //
982 // The ordering of operands can not be determined by looking at a
983 // rule. Especially if a match rule matches several different trees,
984 // several nodes are generated from one instruct specification with
985 // different operand orderings. In this case the adlc generated
986 // variables are the only way to access the ins and operands
987 // deterministically.
988 //
989 // If assigning a register to a node that contains an oop, don't
990 // forget to call ra_->set_oop() for the node.
991 void PhaseCFG::postalloc_expand(PhaseRegAlloc* _ra) {
992 GrowableArray <Node *> new_nodes(32); // Array with new nodes filled by postalloc_expand function of node.
993 GrowableArray <Node *> remove(32);
994 GrowableArray <Node *> succs(32);
995 unsigned int max_idx = C->unique(); // Remember to distinguish new from old nodes.
996 DEBUG_ONLY(bool foundNode = false);
997
998 // for all blocks
999 for (uint i = 0; i < number_of_blocks(); i++) {
1000 Block *b = _blocks[i];
1001 // For all instructions in the current block.
1002 for (uint j = 0; j < b->number_of_nodes(); j++) {
1003 Node *n = b->get_node(j);
1004 if (n->is_Mach() && n->as_Mach()->requires_postalloc_expand()) {
1005 #ifdef ASSERT
1006 if (TracePostallocExpand) {
1007 if (!foundNode) {
1008 foundNode = true;
1009 tty->print("POSTALLOC EXPANDING %d %s\n", C->compile_id(),
1010 C->method() ? C->method()->name()->as_utf8() : C->stub_name());
1011 }
1012 tty->print(" postalloc expanding "); n->dump();
1013 if (Verbose) {
1014 tty->print(" with ins:\n");
1015 for (uint k = 0; k < n->len(); ++k) {
1016 if (n->in(k)) { tty->print(" "); n->in(k)->dump(); }
1017 }
1018 }
1019 }
1020 #endif
1021 new_nodes.clear();
1022 // Collect nodes that have to be removed from the block later on.
1023 uint req = n->req();
1024 remove.clear();
1025 for (uint k = 0; k < req; ++k) {
1026 if (n->in(k) && n->in(k)->is_MachTemp()) {
1027 remove.push(n->in(k)); // MachTemps which are inputs to the old node have to be removed.
1028 n->in(k)->del_req(0);
1029 j--;
1030 }
1031 }
1032
1033 // Check whether we can allocate enough nodes. We set a fix limit for
1034 // the size of postalloc expands with this.
1035 uint unique_limit = C->unique() + 40;
1036 if (unique_limit >= _ra->node_regs_max_index()) {
1037 Compile::current()->record_failure("out of nodes in postalloc expand");
1038 return;
1039 }
1040
1041 // Emit (i.e. generate new nodes).
1042 n->as_Mach()->postalloc_expand(&new_nodes, _ra);
1043
1044 assert(C->unique() < unique_limit, "You allocated too many nodes in your postalloc expand.");
1045
1046 // Disconnect the inputs of the old node.
1047 //
1048 // We reuse MachSpillCopy nodes. If we need to expand them, there
1049 // are many, so reusing pays off. If reused, the node already
1050 // has the new ins. n must be the last node on new_nodes list.
1051 if (!n->is_MachSpillCopy()) {
1052 for (int k = req - 1; k >= 0; --k) {
1053 n->del_req(k);
1054 }
1055 }
1056
1057 #ifdef ASSERT
1058 // Check that all nodes have proper operands.
1059 for (int k = 0; k < new_nodes.length(); ++k) {
1060 if (new_nodes.at(k)->_idx < max_idx || !new_nodes.at(k)->is_Mach()) continue; // old node, Proj ...
1061 MachNode *m = new_nodes.at(k)->as_Mach();
1062 for (unsigned int l = 0; l < m->num_opnds(); ++l) {
1063 if (MachOper::notAnOper(m->_opnds[l])) {
1064 outputStream *os = tty;
1065 os->print("Node %s ", m->Name());
1066 os->print("has invalid opnd %d: %p\n", l, m->_opnds[l]);
1067 assert(0, "Invalid operands, see inline trace in hs_err_pid file.");
1068 }
1069 }
1070 }
1071 #endif
1072
1073 // Collect succs of old node in remove (for projections) and in succs (for
1074 // all other nodes) do _not_ collect projections in remove (but in succs)
1075 // in case the node is a call. We need the projections for calls as they are
1076 // associated with registes (i.e. they are defs).
1077 succs.clear();
1078 for (DUIterator k = n->outs(); n->has_out(k); k++) {
1079 if (n->out(k)->is_Proj() && !n->is_MachCall() && !n->is_MachBranch()) {
1080 remove.push(n->out(k));
1081 } else {
1082 succs.push(n->out(k));
1083 }
1084 }
1085 // Replace old node n as input of its succs by last of the new nodes.
1086 for (int k = 0; k < succs.length(); ++k) {
1087 Node *succ = succs.at(k);
1088 for (uint l = 0; l < succ->req(); ++l) {
1089 if (succ->in(l) == n) {
1090 succ->set_req(l, new_nodes.at(new_nodes.length() - 1));
1091 }
1092 }
1093 for (uint l = succ->req(); l < succ->len(); ++l) {
1094 if (succ->in(l) == n) {
1095 succ->set_prec(l, new_nodes.at(new_nodes.length() - 1));
1096 }
1097 }
1098 }
1099
1100 // Index of old node in block.
1101 uint index = b->find_node(n);
1102 // Insert new nodes into block and map them in nodes->blocks array
1103 // and remember last node in n2.
1104 Node *n2 = NULL;
1105 for (int k = 0; k < new_nodes.length(); ++k) {
1106 n2 = new_nodes.at(k);
1107 b->insert_node(n2, ++index);
1108 map_node_to_block(n2, b);
1109 }
1110
1111 // Add old node n to remove and remove them all from block.
1112 remove.push(n);
1113 j--;
1114 #ifdef ASSERT
1115 if (TracePostallocExpand && Verbose) {
1116 tty->print(" removing:\n");
1117 for (int k = 0; k < remove.length(); ++k) {
1118 tty->print(" "); remove.at(k)->dump();
1119 }
1120 tty->print(" inserting:\n");
1121 for (int k = 0; k < new_nodes.length(); ++k) {
1122 tty->print(" "); new_nodes.at(k)->dump();
1123 }
1124 }
1125 #endif
1126 for (int k = 0; k < remove.length(); ++k) {
1127 if (b->contains(remove.at(k))) {
1128 b->find_remove(remove.at(k));
1129 } else {
1130 assert(remove.at(k)->is_Proj() && (remove.at(k)->in(0)->is_MachBranch()), "");
1131 }
1132 }
1133 // If anything has been inserted (n2 != NULL), continue after last node inserted.
1134 // This does not always work. Some postalloc expands don't insert any nodes, if they
1135 // do optimizations (e.g., max(x,x)). In this case we decrement j accordingly.
1136 j = n2 ? b->find_node(n2) : j;
1137 }
1138 }
1139 }
1140
1141 #ifdef ASSERT
1142 if (foundNode) {
1143 tty->print("FINISHED %d %s\n", C->compile_id(),
1144 C->method() ? C->method()->name()->as_utf8() : C->stub_name());
1145 tty->flush();
1146 }
1147 #endif
1148 }
1149
1150
1151 //------------------------------dump-------------------------------------------
1152 #ifndef PRODUCT
1153 void PhaseCFG::_dump_cfg( const Node *end, VectorSet &visited ) const {
1154 const Node *x = end->is_block_proj();
1155 assert( x, "not a CFG" );
1156
1157 // Do not visit this block again
1158 if( visited.test_set(x->_idx) ) return;
1159
1160 // Skip through this block
1161 const Node *p = x;
1162 do {
1163 p = p->in(0); // Move control forward
1164 assert( !p->is_block_proj() || p->is_Root(), "not a CFG" );
1165 } while( !p->is_block_start() );
1166
1167 // Recursively visit
1168 for (uint i = 1; i < p->req(); i++) {
1169 _dump_cfg(p->in(i), visited);
1170 }
1171
1172 // Dump the block
1173 get_block_for_node(p)->dump(this);
1174 }
1175
1176 void PhaseCFG::dump( ) const {
1177 tty->print("\n--- CFG --- %d BBs\n", number_of_blocks());
1178 if (_blocks.size()) { // Did we do basic-block layout?
1179 for (uint i = 0; i < number_of_blocks(); i++) {
1180 const Block* block = get_block(i);
1181 block->dump(this);
1182 }
1183 } else { // Else do it with a DFS
1184 VectorSet visited(_block_arena);
1185 _dump_cfg(_root,visited);
1186 }
1187 }
1188
1189 void PhaseCFG::dump_headers() {
1190 for (uint i = 0; i < number_of_blocks(); i++) {
1191 Block* block = get_block(i);
1192 if (block != NULL) {
1193 block->dump_head(this);
1194 }
1195 }
1196 }
1197
1198 void PhaseCFG::verify() const {
1199 #ifdef ASSERT
1200 // Verify sane CFG
1201 for (uint i = 0; i < number_of_blocks(); i++) {
1202 Block* block = get_block(i);
1203 uint cnt = block->number_of_nodes();
1204 uint j;
1205 for (j = 0; j < cnt; j++) {
1206 Node *n = block->get_node(j);
1207 assert(get_block_for_node(n) == block, "");
1208 if (j >= 1 && n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_CreateEx) {
1209 assert(j == 1 || block->get_node(j-1)->is_Phi(), "CreateEx must be first instruction in block");
1210 }
1211 for (uint k = 0; k < n->req(); k++) {
1212 Node *def = n->in(k);
1213 if (def && def != n) {
1214 assert(get_block_for_node(def) || def->is_Con(), "must have block; constants for debug info ok");
1215 // Verify that instructions in the block is in correct order.
1216 // Uses must follow their definition if they are at the same block.
1217 // Mostly done to check that MachSpillCopy nodes are placed correctly
1218 // when CreateEx node is moved in build_ifg_physical().
1219 if (get_block_for_node(def) == block && !(block->head()->is_Loop() && n->is_Phi()) &&
1220 // See (+++) comment in reg_split.cpp
1221 !(n->jvms() != NULL && n->jvms()->is_monitor_use(k))) {
1222 bool is_loop = false;
1223 if (n->is_Phi()) {
1224 for (uint l = 1; l < def->req(); l++) {
1225 if (n == def->in(l)) {
1226 is_loop = true;
1227 break; // Some kind of loop
1228 }
1229 }
1230 }
1231 assert(is_loop || block->find_node(def) < j, "uses must follow definitions");
1232 }
1233 }
1234 }
1235 }
1236
1237 j = block->end_idx();
1238 Node* bp = (Node*)block->get_node(block->number_of_nodes() - 1)->is_block_proj();
1239 assert(bp, "last instruction must be a block proj");
1240 assert(bp == block->get_node(j), "wrong number of successors for this block");
1241 if (bp->is_Catch()) {
1242 while (block->get_node(--j)->is_MachProj()) {
1243 ;
1244 }
1245 assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call");
1246 } else if (bp->is_Mach() && bp->as_Mach()->ideal_Opcode() == Op_If) {
1247 assert(block->_num_succs == 2, "Conditional branch must have two targets");
1248 }
1249 }
1250 #endif
1251 }
1252 #endif
1253
1254 UnionFind::UnionFind( uint max ) : _cnt(max), _max(max), _indices(NEW_RESOURCE_ARRAY(uint,max)) {
1255 Copy::zero_to_bytes( _indices, sizeof(uint)*max );
1256 }
1257
1258 void UnionFind::extend( uint from_idx, uint to_idx ) {
1259 _nesting.check();
1260 if( from_idx >= _max ) {
1261 uint size = 16;
1262 while( size <= from_idx ) size <<=1;
1263 _indices = REALLOC_RESOURCE_ARRAY( uint, _indices, _max, size );
1264 _max = size;
1265 }
1266 while( _cnt <= from_idx ) _indices[_cnt++] = 0;
1267 _indices[from_idx] = to_idx;
1268 }
1269
1270 void UnionFind::reset( uint max ) {
1271 // Force the Union-Find mapping to be at least this large
1272 extend(max,0);
1273 // Initialize to be the ID mapping.
1274 for( uint i=0; i<max; i++ ) map(i,i);
1275 }
1276
1277 // Straight out of Tarjan's union-find algorithm
1278 uint UnionFind::Find_compress( uint idx ) {
1279 uint cur = idx;
1280 uint next = lookup(cur);
1281 while( next != cur ) { // Scan chain of equivalences
1282 assert( next < cur, "always union smaller" );
1283 cur = next; // until find a fixed-point
1284 next = lookup(cur);
1285 }
1286 // Core of union-find algorithm: update chain of
1287 // equivalences to be equal to the root.
1288 while( idx != next ) {
1289 uint tmp = lookup(idx);
1290 map(idx, next);
1291 idx = tmp;
1292 }
1293 return idx;
1294 }
1295
1296 // Like Find above, but no path compress, so bad asymptotic behavior
1297 uint UnionFind::Find_const( uint idx ) const {
1298 if( idx == 0 ) return idx; // Ignore the zero idx
1299 // Off the end? This can happen during debugging dumps
1300 // when data structures have not finished being updated.
1301 if( idx >= _max ) return idx;
1302 uint next = lookup(idx);
1303 while( next != idx ) { // Scan chain of equivalences
1304 idx = next; // until find a fixed-point
1305 next = lookup(idx);
1306 }
1307 return next;
1308 }
1309
1310 // union 2 sets together.
1311 void UnionFind::Union( uint idx1, uint idx2 ) {
1312 uint src = Find(idx1);
1313 uint dst = Find(idx2);
1314 assert( src, "" );
1315 assert( dst, "" );
1316 assert( src < _max, "oob" );
1317 assert( dst < _max, "oob" );
1318 assert( src < dst, "always union smaller" );
1319 map(dst,src);
1320 }
1321
1322 #ifndef PRODUCT
1323 void Trace::dump( ) const {
1324 tty->print_cr("Trace (freq %f)", first_block()->_freq);
1325 for (Block *b = first_block(); b != NULL; b = next(b)) {
1326 tty->print(" B%d", b->_pre_order);
1327 if (b->head()->is_Loop()) {
1328 tty->print(" (L%d)", b->compute_loop_alignment());
1329 }
1330 if (b->has_loop_alignment()) {
1331 tty->print(" (T%d)", b->code_alignment());
1332 }
1333 }
1334 tty->cr();
1335 }
1336
1337 void CFGEdge::dump( ) const {
1338 tty->print(" B%d --> B%d Freq: %f out:%3d%% in:%3d%% State: ",
1339 from()->_pre_order, to()->_pre_order, freq(), _from_pct, _to_pct);
1340 switch(state()) {
1341 case connected:
1342 tty->print("connected");
1343 break;
1344 case open:
1345 tty->print("open");
1346 break;
1347 case interior:
1348 tty->print("interior");
1349 break;
1350 }
1351 if (infrequent()) {
1352 tty->print(" infrequent");
1353 }
1354 tty->cr();
1355 }
1356 #endif
1357
1358 // Comparison function for edges
1359 static int edge_order(CFGEdge **e0, CFGEdge **e1) {
1360 float freq0 = (*e0)->freq();
1361 float freq1 = (*e1)->freq();
1362 if (freq0 != freq1) {
1363 return freq0 > freq1 ? -1 : 1;
1364 }
1365
1366 int dist0 = (*e0)->to()->_rpo - (*e0)->from()->_rpo;
1367 int dist1 = (*e1)->to()->_rpo - (*e1)->from()->_rpo;
1368
1369 return dist1 - dist0;
1370 }
1371
1372 // Comparison function for edges
1373 extern "C" int trace_frequency_order(const void *p0, const void *p1) {
1374 Trace *tr0 = *(Trace **) p0;
1375 Trace *tr1 = *(Trace **) p1;
1376 Block *b0 = tr0->first_block();
1377 Block *b1 = tr1->first_block();
1378
1379 // The trace of connector blocks goes at the end;
1380 // we only expect one such trace
1381 if (b0->is_connector() != b1->is_connector()) {
1382 return b1->is_connector() ? -1 : 1;
1383 }
1384
1385 // Pull more frequently executed blocks to the beginning
1386 float freq0 = b0->_freq;
1387 float freq1 = b1->_freq;
1388 if (freq0 != freq1) {
1389 return freq0 > freq1 ? -1 : 1;
1390 }
1391
1392 int diff = tr0->first_block()->_rpo - tr1->first_block()->_rpo;
1393
1394 return diff;
1395 }
1396
1397 // Find edges of interest, i.e, those which can fall through. Presumes that
1398 // edges which don't fall through are of low frequency and can be generally
1399 // ignored. Initialize the list of traces.
1400 void PhaseBlockLayout::find_edges() {
1401 // Walk the blocks, creating edges and Traces
1402 uint i;
1403 Trace *tr = NULL;
1404 for (i = 0; i < _cfg.number_of_blocks(); i++) {
1405 Block* b = _cfg.get_block(i);
1406 tr = new Trace(b, next, prev);
1407 traces[tr->id()] = tr;
1408
1409 // All connector blocks should be at the end of the list
1410 if (b->is_connector()) break;
1411
1412 // If this block and the next one have a one-to-one successor
1413 // predecessor relationship, simply append the next block
1414 int nfallthru = b->num_fall_throughs();
1415 while (nfallthru == 1 &&
1416 b->succ_fall_through(0)) {
1417 Block *n = b->_succs[0];
1418
1419 // Skip over single-entry connector blocks, we don't want to
1420 // add them to the trace.
1421 while (n->is_connector() && n->num_preds() == 1) {
1422 n = n->_succs[0];
1423 }
1424
1425 // We see a merge point, so stop search for the next block
1426 if (n->num_preds() != 1) break;
1427
1428 i++;
1429 assert(n = _cfg.get_block(i), "expecting next block");
1430 tr->append(n);
1431 uf->map(n->_pre_order, tr->id());
1432 traces[n->_pre_order] = NULL;
1433 nfallthru = b->num_fall_throughs();
1434 b = n;
1435 }
1436
1437 if (nfallthru > 0) {
1438 // Create a CFGEdge for each outgoing
1439 // edge that could be a fall-through.
1440 for (uint j = 0; j < b->_num_succs; j++ ) {
1441 if (b->succ_fall_through(j)) {
1442 Block *target = b->non_connector_successor(j);
1443 float freq = b->_freq * b->succ_prob(j);
1444 int from_pct = (int) ((100 * freq) / b->_freq);
1445 int to_pct = (int) ((100 * freq) / target->_freq);
1446 edges->append(new CFGEdge(b, target, freq, from_pct, to_pct));
1447 }
1448 }
1449 }
1450 }
1451
1452 // Group connector blocks into one trace
1453 for (i++; i < _cfg.number_of_blocks(); i++) {
1454 Block *b = _cfg.get_block(i);
1455 assert(b->is_connector(), "connector blocks at the end");
1456 tr->append(b);
1457 uf->map(b->_pre_order, tr->id());
1458 traces[b->_pre_order] = NULL;
1459 }
1460 }
1461
1462 // Union two traces together in uf, and null out the trace in the list
1463 void PhaseBlockLayout::union_traces(Trace* updated_trace, Trace* old_trace) {
1464 uint old_id = old_trace->id();
1465 uint updated_id = updated_trace->id();
1466
1467 uint lo_id = updated_id;
1468 uint hi_id = old_id;
1469
1470 // If from is greater than to, swap values to meet
1471 // UnionFind guarantee.
1472 if (updated_id > old_id) {
1473 lo_id = old_id;
1474 hi_id = updated_id;
1475
1476 // Fix up the trace ids
1477 traces[lo_id] = traces[updated_id];
1478 updated_trace->set_id(lo_id);
1479 }
1480
1481 // Union the lower with the higher and remove the pointer
1482 // to the higher.
1483 uf->Union(lo_id, hi_id);
1484 traces[hi_id] = NULL;
1485 }
1486
1487 // Append traces together via the most frequently executed edges
1488 void PhaseBlockLayout::grow_traces() {
1489 // Order the edges, and drive the growth of Traces via the most
1490 // frequently executed edges.
1491 edges->sort(edge_order);
1492 for (int i = 0; i < edges->length(); i++) {
1493 CFGEdge *e = edges->at(i);
1494
1495 if (e->state() != CFGEdge::open) continue;
1496
1497 Block *src_block = e->from();
1498 Block *targ_block = e->to();
1499
1500 // Don't grow traces along backedges?
1501 if (!BlockLayoutRotateLoops) {
1502 if (targ_block->_rpo <= src_block->_rpo) {
1503 targ_block->set_loop_alignment(targ_block);
1504 continue;
1505 }
1506 }
1507
1508 Trace *src_trace = trace(src_block);
1509 Trace *targ_trace = trace(targ_block);
1510
1511 // If the edge in question can join two traces at their ends,
1512 // append one trace to the other.
1513 if (src_trace->last_block() == src_block) {
1514 if (src_trace == targ_trace) {
1515 e->set_state(CFGEdge::interior);
1516 if (targ_trace->backedge(e)) {
1517 // Reset i to catch any newly eligible edge
1518 // (Or we could remember the first "open" edge, and reset there)
1519 i = 0;
1520 }
1521 } else if (targ_trace->first_block() == targ_block) {
1522 e->set_state(CFGEdge::connected);
1523 src_trace->append(targ_trace);
1524 union_traces(src_trace, targ_trace);
1525 }
1526 }
1527 }
1528 }
1529
1530 // Embed one trace into another, if the fork or join points are sufficiently
1531 // balanced.
1532 void PhaseBlockLayout::merge_traces(bool fall_thru_only) {
1533 // Walk the edge list a another time, looking at unprocessed edges.
1534 // Fold in diamonds
1535 for (int i = 0; i < edges->length(); i++) {
1536 CFGEdge *e = edges->at(i);
1537
1538 if (e->state() != CFGEdge::open) continue;
1539 if (fall_thru_only) {
1540 if (e->infrequent()) continue;
1541 }
1542
1543 Block *src_block = e->from();
1544 Trace *src_trace = trace(src_block);
1545 bool src_at_tail = src_trace->last_block() == src_block;
1546
1547 Block *targ_block = e->to();
1548 Trace *targ_trace = trace(targ_block);
1549 bool targ_at_start = targ_trace->first_block() == targ_block;
1550
1551 if (src_trace == targ_trace) {
1552 // This may be a loop, but we can't do much about it.
1553 e->set_state(CFGEdge::interior);
1554 continue;
1555 }
1556
1557 if (fall_thru_only) {
1558 // If the edge links the middle of two traces, we can't do anything.
1559 // Mark the edge and continue.
1560 if (!src_at_tail & !targ_at_start) {
1561 continue;
1562 }
1563
1564 // Don't grow traces along backedges?
1565 if (!BlockLayoutRotateLoops && (targ_block->_rpo <= src_block->_rpo)) {
1566 continue;
1567 }
1568
1569 // If both ends of the edge are available, why didn't we handle it earlier?
1570 assert(src_at_tail ^ targ_at_start, "Should have caught this edge earlier.");
1571
1572 if (targ_at_start) {
1573 // Insert the "targ" trace in the "src" trace if the insertion point
1574 // is a two way branch.
1575 // Better profitability check possible, but may not be worth it.
1576 // Someday, see if the this "fork" has an associated "join";
1577 // then make a policy on merging this trace at the fork or join.
1578 // For example, other things being equal, it may be better to place this
1579 // trace at the join point if the "src" trace ends in a two-way, but
1580 // the insertion point is one-way.
1581 assert(src_block->num_fall_throughs() == 2, "unexpected diamond");
1582 e->set_state(CFGEdge::connected);
1583 src_trace->insert_after(src_block, targ_trace);
1584 union_traces(src_trace, targ_trace);
1585 } else if (src_at_tail) {
1586 if (src_trace != trace(_cfg.get_root_block())) {
1587 e->set_state(CFGEdge::connected);
1588 targ_trace->insert_before(targ_block, src_trace);
1589 union_traces(targ_trace, src_trace);
1590 }
1591 }
1592 } else if (e->state() == CFGEdge::open) {
1593 // Append traces, even without a fall-thru connection.
1594 // But leave root entry at the beginning of the block list.
1595 if (targ_trace != trace(_cfg.get_root_block())) {
1596 e->set_state(CFGEdge::connected);
1597 src_trace->append(targ_trace);
1598 union_traces(src_trace, targ_trace);
1599 }
1600 }
1601 }
1602 }
1603
1604 // Order the sequence of the traces in some desirable way, and fixup the
1605 // jumps at the end of each block.
1606 void PhaseBlockLayout::reorder_traces(int count) {
1607 ResourceArea *area = Thread::current()->resource_area();
1608 Trace ** new_traces = NEW_ARENA_ARRAY(area, Trace *, count);
1609 Block_List worklist;
1610 int new_count = 0;
1611
1612 // Compact the traces.
1613 for (int i = 0; i < count; i++) {
1614 Trace *tr = traces[i];
1615 if (tr != NULL) {
1616 new_traces[new_count++] = tr;
1617 }
1618 }
1619
1620 // The entry block should be first on the new trace list.
1621 Trace *tr = trace(_cfg.get_root_block());
1622 assert(tr == new_traces[0], "entry trace misplaced");
1623
1624 // Sort the new trace list by frequency
1625 qsort(new_traces + 1, new_count - 1, sizeof(new_traces[0]), trace_frequency_order);
1626
1627 // Patch up the successor blocks
1628 _cfg.clear_blocks();
1629 for (int i = 0; i < new_count; i++) {
1630 Trace *tr = new_traces[i];
1631 if (tr != NULL) {
1632 tr->fixup_blocks(_cfg);
1633 }
1634 }
1635 }
1636
1637 // Order basic blocks based on frequency
1638 PhaseBlockLayout::PhaseBlockLayout(PhaseCFG &cfg)
1639 : Phase(BlockLayout)
1640 , _cfg(cfg) {
1641 ResourceMark rm;
1642 ResourceArea *area = Thread::current()->resource_area();
1643
1644 // List of traces
1645 int size = _cfg.number_of_blocks() + 1;
1646 traces = NEW_ARENA_ARRAY(area, Trace *, size);
1647 memset(traces, 0, size*sizeof(Trace*));
1648 next = NEW_ARENA_ARRAY(area, Block *, size);
1649 memset(next, 0, size*sizeof(Block *));
1650 prev = NEW_ARENA_ARRAY(area, Block *, size);
1651 memset(prev , 0, size*sizeof(Block *));
1652
1653 // List of edges
1654 edges = new GrowableArray<CFGEdge*>;
1655
1656 // Mapping block index --> block_trace
1657 uf = new UnionFind(size);
1658 uf->reset(size);
1659
1660 // Find edges and create traces.
1661 find_edges();
1662
1663 // Grow traces at their ends via most frequent edges.
1664 grow_traces();
1665
1666 // Merge one trace into another, but only at fall-through points.
1667 // This may make diamonds and other related shapes in a trace.
1668 merge_traces(true);
1669
1670 // Run merge again, allowing two traces to be catenated, even if
1671 // one does not fall through into the other. This appends loosely
1672 // related traces to be near each other.
1673 merge_traces(false);
1674
1675 // Re-order all the remaining traces by frequency
1676 reorder_traces(size);
1677
1678 assert(_cfg.number_of_blocks() >= (uint) (size - 1), "number of blocks can not shrink");
1679 }
1680
1681
1682 // Edge e completes a loop in a trace. If the target block is head of the
1683 // loop, rotate the loop block so that the loop ends in a conditional branch.
1684 bool Trace::backedge(CFGEdge *e) {
1685 bool loop_rotated = false;
1686 Block *src_block = e->from();
1687 Block *targ_block = e->to();
1688
1689 assert(last_block() == src_block, "loop discovery at back branch");
1690 if (first_block() == targ_block) {
1691 if (BlockLayoutRotateLoops && last_block()->num_fall_throughs() < 2) {
1692 // Find the last block in the trace that has a conditional
1693 // branch.
1694 Block *b;
1695 for (b = last_block(); b != NULL; b = prev(b)) {
1696 if (b->num_fall_throughs() == 2) {
1697 break;
1698 }
1699 }
1700
1701 if (b != last_block() && b != NULL) {
1702 loop_rotated = true;
1703
1704 // Rotate the loop by doing two-part linked-list surgery.
1705 append(first_block());
1706 break_loop_after(b);
1707 }
1708 }
1709
1710 // Backbranch to the top of a trace
1711 // Scroll forward through the trace from the targ_block. If we find
1712 // a loop head before another loop top, use the the loop head alignment.
1713 for (Block *b = targ_block; b != NULL; b = next(b)) {
1714 if (b->has_loop_alignment()) {
1715 break;
1716 }
1717 if (b->head()->is_Loop()) {
1718 targ_block = b;
1719 break;
1720 }
1721 }
1722
1723 first_block()->set_loop_alignment(targ_block);
1724
1725 } else {
1726 // Backbranch into the middle of a trace
1727 targ_block->set_loop_alignment(targ_block);
1728 }
1729
1730 return loop_rotated;
1731 }
1732
1733 // push blocks onto the CFG list
1734 // ensure that blocks have the correct two-way branch sense
1735 void Trace::fixup_blocks(PhaseCFG &cfg) {
1736 Block *last = last_block();
1737 for (Block *b = first_block(); b != NULL; b = next(b)) {
1738 cfg.add_block(b);
1739 if (!b->is_connector()) {
1740 int nfallthru = b->num_fall_throughs();
1741 if (b != last) {
1742 if (nfallthru == 2) {
1743 // Ensure that the sense of the branch is correct
1744 Block *bnext = next(b);
1745 Block *bs0 = b->non_connector_successor(0);
1746
1747 MachNode *iff = b->get_node(b->number_of_nodes() - 3)->as_Mach();
1748 ProjNode *proj0 = b->get_node(b->number_of_nodes() - 2)->as_Proj();
1749 ProjNode *proj1 = b->get_node(b->number_of_nodes() - 1)->as_Proj();
1750
1751 if (bnext == bs0) {
1752 // Fall-thru case in succs[0], should be in succs[1]
1753
1754 // Flip targets in _succs map
1755 Block *tbs0 = b->_succs[0];
1756 Block *tbs1 = b->_succs[1];
1757 b->_succs.map( 0, tbs1 );
1758 b->_succs.map( 1, tbs0 );
1759
1760 // Flip projections to match targets
1761 b->map_node(proj1, b->number_of_nodes() - 2);
1762 b->map_node(proj0, b->number_of_nodes() - 1);
1763 }
1764 }
1765 }
1766 }
1767 }
1768 }