1 /*
2 * Copyright (c) 1997, 2019, 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 #include "utilities/powerOfTwo.hpp"
40
grow(uint i)41 void Block_Array::grow( uint i ) {
42 assert(i >= Max(), "must be an overflow");
43 debug_only(_limit = i+1);
44 if( i < _size ) return;
45 if( !_size ) {
46 _size = 1;
47 _blocks = (Block**)_arena->Amalloc( _size * sizeof(Block*) );
48 _blocks[0] = NULL;
49 }
50 uint old = _size;
51 _size = next_power_of_2(i);
52 _blocks = (Block**)_arena->Arealloc( _blocks, old*sizeof(Block*),_size*sizeof(Block*));
53 Copy::zero_to_bytes( &_blocks[old], (_size-old)*sizeof(Block*) );
54 }
55
remove(uint i)56 void Block_List::remove(uint i) {
57 assert(i < _cnt, "index out of bounds");
58 Copy::conjoint_words_to_lower((HeapWord*)&_blocks[i+1], (HeapWord*)&_blocks[i], ((_cnt-i-1)*sizeof(Block*)));
59 pop(); // shrink list by one block
60 }
61
insert(uint i,Block * b)62 void Block_List::insert(uint i, Block *b) {
63 push(b); // grow list by one block
64 Copy::conjoint_words_to_higher((HeapWord*)&_blocks[i], (HeapWord*)&_blocks[i+1], ((_cnt-i-1)*sizeof(Block*)));
65 _blocks[i] = b;
66 }
67
68 #ifndef PRODUCT
print()69 void Block_List::print() {
70 for (uint i=0; i < size(); i++) {
71 tty->print("B%d ", _blocks[i]->_pre_order);
72 }
73 tty->print("size = %d\n", size());
74 }
75 #endif
76
code_alignment() const77 uint Block::code_alignment() const {
78 // Check for Root block
79 if (_pre_order == 0) return CodeEntryAlignment;
80 // Check for Start block
81 if (_pre_order == 1) return InteriorEntryAlignment;
82 // Check for loop alignment
83 if (has_loop_alignment()) return loop_alignment();
84
85 return relocInfo::addr_unit(); // no particular alignment
86 }
87
compute_loop_alignment()88 uint Block::compute_loop_alignment() {
89 Node *h = head();
90 int unit_sz = relocInfo::addr_unit();
91 if (h->is_Loop() && h->as_Loop()->is_inner_loop()) {
92 // Pre- and post-loops have low trip count so do not bother with
93 // NOPs for align loop head. The constants are hidden from tuning
94 // but only because my "divide by 4" heuristic surely gets nearly
95 // all possible gain (a "do not align at all" heuristic has a
96 // chance of getting a really tiny gain).
97 if (h->is_CountedLoop() && (h->as_CountedLoop()->is_pre_loop() ||
98 h->as_CountedLoop()->is_post_loop())) {
99 return (OptoLoopAlignment > 4*unit_sz) ? (OptoLoopAlignment>>2) : unit_sz;
100 }
101 // Loops with low backedge frequency should not be aligned.
102 Node *n = h->in(LoopNode::LoopBackControl)->in(0);
103 if (n->is_MachIf() && n->as_MachIf()->_prob < 0.01) {
104 return unit_sz; // Loop does not loop, more often than not!
105 }
106 return OptoLoopAlignment; // Otherwise align loop head
107 }
108
109 return unit_sz; // no particular alignment
110 }
111
112 // Compute the size of first 'inst_cnt' instructions in this block.
113 // Return the number of instructions left to compute if the block has
114 // less then 'inst_cnt' instructions. Stop, and return 0 if sum_size
115 // exceeds OptoLoopAlignment.
compute_first_inst_size(uint & sum_size,uint inst_cnt,PhaseRegAlloc * ra)116 uint Block::compute_first_inst_size(uint& sum_size, uint inst_cnt,
117 PhaseRegAlloc* ra) {
118 uint last_inst = number_of_nodes();
119 for( uint j = 0; j < last_inst && inst_cnt > 0; j++ ) {
120 uint inst_size = get_node(j)->size(ra);
121 if( inst_size > 0 ) {
122 inst_cnt--;
123 uint sz = sum_size + inst_size;
124 if( sz <= (uint)OptoLoopAlignment ) {
125 // Compute size of instructions which fit into fetch buffer only
126 // since all inst_cnt instructions will not fit even if we align them.
127 sum_size = sz;
128 } else {
129 return 0;
130 }
131 }
132 }
133 return inst_cnt;
134 }
135
find_node(const Node * n) const136 uint Block::find_node( const Node *n ) const {
137 for( uint i = 0; i < number_of_nodes(); i++ ) {
138 if( get_node(i) == n )
139 return i;
140 }
141 ShouldNotReachHere();
142 return 0;
143 }
144
145 // Find and remove n from block list
find_remove(const Node * n)146 void Block::find_remove( const Node *n ) {
147 remove_node(find_node(n));
148 }
149
contains(const Node * n) const150 bool Block::contains(const Node *n) const {
151 return _nodes.contains(n);
152 }
153
154 // Return empty status of a block. Empty blocks contain only the head, other
155 // ideal nodes, and an optional trailing goto.
is_Empty() const156 int Block::is_Empty() const {
157
158 // Root or start block is not considered empty
159 if (head()->is_Root() || head()->is_Start()) {
160 return not_empty;
161 }
162
163 int success_result = completely_empty;
164 int end_idx = number_of_nodes() - 1;
165
166 // Check for ending goto
167 if ((end_idx > 0) && (get_node(end_idx)->is_MachGoto())) {
168 success_result = empty_with_goto;
169 end_idx--;
170 }
171
172 // Unreachable blocks are considered empty
173 if (num_preds() <= 1) {
174 return success_result;
175 }
176
177 // Ideal nodes are allowable in empty blocks: skip them Only MachNodes
178 // turn directly into code, because only MachNodes have non-trivial
179 // emit() functions.
180 while ((end_idx > 0) && !get_node(end_idx)->is_Mach()) {
181 end_idx--;
182 }
183
184 // No room for any interesting instructions?
185 if (end_idx == 0) {
186 return success_result;
187 }
188
189 return not_empty;
190 }
191
192 // Return true if the block's code implies that it is likely to be
193 // executed infrequently. Check to see if the block ends in a Halt or
194 // a low probability call.
has_uncommon_code() const195 bool Block::has_uncommon_code() const {
196 Node* en = end();
197
198 if (en->is_MachGoto())
199 en = en->in(0);
200 if (en->is_Catch())
201 en = en->in(0);
202 if (en->is_MachProj() && en->in(0)->is_MachCall()) {
203 MachCallNode* call = en->in(0)->as_MachCall();
204 if (call->cnt() != COUNT_UNKNOWN && call->cnt() <= PROB_UNLIKELY_MAG(4)) {
205 // This is true for slow-path stubs like new_{instance,array},
206 // slow_arraycopy, complete_monitor_locking, uncommon_trap.
207 // The magic number corresponds to the probability of an uncommon_trap,
208 // even though it is a count not a probability.
209 return true;
210 }
211 }
212
213 int op = en->is_Mach() ? en->as_Mach()->ideal_Opcode() : en->Opcode();
214 return op == Op_Halt;
215 }
216
217 // True if block is low enough frequency or guarded by a test which
218 // mostly does not go here.
is_uncommon(const Block * block)219 bool PhaseCFG::is_uncommon(const Block* block) {
220 // Initial blocks must never be moved, so are never uncommon.
221 if (block->head()->is_Root() || block->head()->is_Start()) return false;
222
223 // Check for way-low freq
224 if(block->_freq < BLOCK_FREQUENCY(0.00001f) ) return true;
225
226 // Look for code shape indicating uncommon_trap or slow path
227 if (block->has_uncommon_code()) return true;
228
229 const float epsilon = 0.05f;
230 const float guard_factor = PROB_UNLIKELY_MAG(4) / (1.f - epsilon);
231 uint uncommon_preds = 0;
232 uint freq_preds = 0;
233 uint uncommon_for_freq_preds = 0;
234
235 for( uint i=1; i< block->num_preds(); i++ ) {
236 Block* guard = get_block_for_node(block->pred(i));
237 // Check to see if this block follows its guard 1 time out of 10000
238 // or less.
239 //
240 // See list of magnitude-4 unlikely probabilities in cfgnode.hpp which
241 // we intend to be "uncommon", such as slow-path TLE allocation,
242 // predicted call failure, and uncommon trap triggers.
243 //
244 // Use an epsilon value of 5% to allow for variability in frequency
245 // predictions and floating point calculations. The net effect is
246 // that guard_factor is set to 9500.
247 //
248 // Ignore low-frequency blocks.
249 // The next check is (guard->_freq < 1.e-5 * 9500.).
250 if(guard->_freq*BLOCK_FREQUENCY(guard_factor) < BLOCK_FREQUENCY(0.00001f)) {
251 uncommon_preds++;
252 } else {
253 freq_preds++;
254 if(block->_freq < guard->_freq * guard_factor ) {
255 uncommon_for_freq_preds++;
256 }
257 }
258 }
259 if( block->num_preds() > 1 &&
260 // The block is uncommon if all preds are uncommon or
261 (uncommon_preds == (block->num_preds()-1) ||
262 // it is uncommon for all frequent preds.
263 uncommon_for_freq_preds == freq_preds) ) {
264 return true;
265 }
266 return false;
267 }
268
269 #ifndef PRODUCT
dump_bidx(const Block * orig,outputStream * st) const270 void Block::dump_bidx(const Block* orig, outputStream* st) const {
271 if (_pre_order) st->print("B%d", _pre_order);
272 else st->print("N%d", head()->_idx);
273
274 if (Verbose && orig != this) {
275 // Dump the original block's idx
276 st->print(" (");
277 orig->dump_bidx(orig, st);
278 st->print(")");
279 }
280 }
281
dump_pred(const PhaseCFG * cfg,Block * orig,outputStream * st) const282 void Block::dump_pred(const PhaseCFG* cfg, Block* orig, outputStream* st) const {
283 if (is_connector()) {
284 for (uint i=1; i<num_preds(); i++) {
285 Block *p = cfg->get_block_for_node(pred(i));
286 p->dump_pred(cfg, orig, st);
287 }
288 } else {
289 dump_bidx(orig, st);
290 st->print(" ");
291 }
292 }
293
dump_head(const PhaseCFG * cfg,outputStream * st) const294 void Block::dump_head(const PhaseCFG* cfg, outputStream* st) const {
295 // Print the basic block.
296 dump_bidx(this, st);
297 st->print(": ");
298
299 // Print the outgoing CFG edges.
300 st->print("#\tout( ");
301 for( uint i=0; i<_num_succs; i++ ) {
302 non_connector_successor(i)->dump_bidx(_succs[i], st);
303 st->print(" ");
304 }
305
306 // Print the incoming CFG edges.
307 st->print(") <- ");
308 if( head()->is_block_start() ) {
309 st->print("in( ");
310 for (uint i=1; i<num_preds(); i++) {
311 Node *s = pred(i);
312 if (cfg != NULL) {
313 Block *p = cfg->get_block_for_node(s);
314 p->dump_pred(cfg, p, st);
315 } else {
316 while (!s->is_block_start()) {
317 s = s->in(0);
318 }
319 st->print("N%d ", s->_idx );
320 }
321 }
322 st->print(") ");
323 } else {
324 st->print("BLOCK HEAD IS JUNK ");
325 }
326
327 // Print loop, if any
328 const Block *bhead = this; // Head of self-loop
329 Node *bh = bhead->head();
330
331 if ((cfg != NULL) && bh->is_Loop() && !head()->is_Root()) {
332 LoopNode *loop = bh->as_Loop();
333 const Block *bx = cfg->get_block_for_node(loop->in(LoopNode::LoopBackControl));
334 while (bx->is_connector()) {
335 bx = cfg->get_block_for_node(bx->pred(1));
336 }
337 st->print("Loop( B%d-B%d ", bhead->_pre_order, bx->_pre_order);
338 // Dump any loop-specific bits, especially for CountedLoops.
339 loop->dump_spec(st);
340 st->print(")");
341 } else if (has_loop_alignment()) {
342 st->print("top-of-loop");
343 }
344
345 // Print frequency and other optimization-relevant information
346 st->print(" Freq: %g",_freq);
347 if( Verbose || WizardMode ) {
348 st->print(" IDom: %d/#%d", _idom ? _idom->_pre_order : 0, _dom_depth);
349 st->print(" RegPressure: %d",_reg_pressure);
350 st->print(" IHRP Index: %d",_ihrp_index);
351 st->print(" FRegPressure: %d",_freg_pressure);
352 st->print(" FHRP Index: %d",_fhrp_index);
353 }
354 st->cr();
355 }
356
dump() const357 void Block::dump() const {
358 dump(NULL);
359 }
360
dump(const PhaseCFG * cfg) const361 void Block::dump(const PhaseCFG* cfg) const {
362 dump_head(cfg);
363 for (uint i=0; i< number_of_nodes(); i++) {
364 get_node(i)->dump();
365 }
366 tty->print("\n");
367 }
368 #endif
369
PhaseCFG(Arena * arena,RootNode * root,Matcher & matcher)370 PhaseCFG::PhaseCFG(Arena* arena, RootNode* root, Matcher& matcher)
371 : Phase(CFG)
372 , _root(root)
373 , _block_arena(arena)
374 , _regalloc(NULL)
375 , _scheduling_for_pressure(false)
376 , _matcher(matcher)
377 , _node_to_block_mapping(arena)
378 , _node_latency(NULL)
379 #ifndef PRODUCT
380 , _trace_opto_pipelining(C->directive()->TraceOptoPipeliningOption)
381 #endif
382 #ifdef ASSERT
383 , _raw_oops(arena)
384 #endif
385 {
386 ResourceMark rm;
387 // I'll need a few machine-specific GotoNodes. Make an Ideal GotoNode,
388 // then Match it into a machine-specific Node. Then clone the machine
389 // Node on demand.
390 Node *x = new GotoNode(NULL);
391 x->init_req(0, x);
392 _goto = matcher.match_tree(x);
393 assert(_goto != NULL, "");
394 _goto->set_req(0,_goto);
395
396 // Build the CFG in Reverse Post Order
397 _number_of_blocks = build_cfg();
398 _root_block = get_block_for_node(_root);
399 }
400
401 // Build a proper looking CFG. Make every block begin with either a StartNode
402 // or a RegionNode. Make every block end with either a Goto, If or Return.
403 // The RootNode both starts and ends it's own block. Do this with a recursive
404 // backwards walk over the control edges.
build_cfg()405 uint PhaseCFG::build_cfg() {
406 Arena *a = Thread::current()->resource_area();
407 VectorSet visited(a);
408
409 // Allocate stack with enough space to avoid frequent realloc
410 Node_Stack nstack(a, C->live_nodes() >> 1);
411 nstack.push(_root, 0);
412 uint sum = 0; // Counter for blocks
413
414 while (nstack.is_nonempty()) {
415 // node and in's index from stack's top
416 // 'np' is _root (see above) or RegionNode, StartNode: we push on stack
417 // only nodes which point to the start of basic block (see below).
418 Node *np = nstack.node();
419 // idx > 0, except for the first node (_root) pushed on stack
420 // at the beginning when idx == 0.
421 // We will use the condition (idx == 0) later to end the build.
422 uint idx = nstack.index();
423 Node *proj = np->in(idx);
424 const Node *x = proj->is_block_proj();
425 // Does the block end with a proper block-ending Node? One of Return,
426 // If or Goto? (This check should be done for visited nodes also).
427 if (x == NULL) { // Does not end right...
428 Node *g = _goto->clone(); // Force it to end in a Goto
429 g->set_req(0, proj);
430 np->set_req(idx, g);
431 x = proj = g;
432 }
433 if (!visited.test_set(x->_idx)) { // Visit this block once
434 // Skip any control-pinned middle'in stuff
435 Node *p = proj;
436 do {
437 proj = p; // Update pointer to last Control
438 p = p->in(0); // Move control forward
439 } while( !p->is_block_proj() &&
440 !p->is_block_start() );
441 // Make the block begin with one of Region or StartNode.
442 if( !p->is_block_start() ) {
443 RegionNode *r = new RegionNode( 2 );
444 r->init_req(1, p); // Insert RegionNode in the way
445 proj->set_req(0, r); // Insert RegionNode in the way
446 p = r;
447 }
448 // 'p' now points to the start of this basic block
449
450 // Put self in array of basic blocks
451 Block *bb = new (_block_arena) Block(_block_arena, p);
452 map_node_to_block(p, bb);
453 map_node_to_block(x, bb);
454 if( x != p ) { // Only for root is x == p
455 bb->push_node((Node*)x);
456 }
457 // Now handle predecessors
458 ++sum; // Count 1 for self block
459 uint cnt = bb->num_preds();
460 for (int i = (cnt - 1); i > 0; i-- ) { // For all predecessors
461 Node *prevproj = p->in(i); // Get prior input
462 assert( !prevproj->is_Con(), "dead input not removed" );
463 // Check to see if p->in(i) is a "control-dependent" CFG edge -
464 // i.e., it splits at the source (via an IF or SWITCH) and merges
465 // at the destination (via a many-input Region).
466 // This breaks critical edges. The RegionNode to start the block
467 // will be added when <p,i> is pulled off the node stack
468 if ( cnt > 2 ) { // Merging many things?
469 assert( prevproj== bb->pred(i),"");
470 if(prevproj->is_block_proj() != prevproj) { // Control-dependent edge?
471 // Force a block on the control-dependent edge
472 Node *g = _goto->clone(); // Force it to end in a Goto
473 g->set_req(0,prevproj);
474 p->set_req(i,g);
475 }
476 }
477 nstack.push(p, i); // 'p' is RegionNode or StartNode
478 }
479 } else { // Post-processing visited nodes
480 nstack.pop(); // remove node from stack
481 // Check if it the fist node pushed on stack at the beginning.
482 if (idx == 0) break; // end of the build
483 // Find predecessor basic block
484 Block *pb = get_block_for_node(x);
485 // Insert into nodes array, if not already there
486 if (!has_block(proj)) {
487 assert( x != proj, "" );
488 // Map basic block of projection
489 map_node_to_block(proj, pb);
490 pb->push_node(proj);
491 }
492 // Insert self as a child of my predecessor block
493 pb->_succs.map(pb->_num_succs++, get_block_for_node(np));
494 assert( pb->get_node(pb->number_of_nodes() - pb->_num_succs)->is_block_proj(),
495 "too many control users, not a CFG?" );
496 }
497 }
498 // Return number of basic blocks for all children and self
499 return sum;
500 }
501
502 // Inserts a goto & corresponding basic block between
503 // block[block_no] and its succ_no'th successor block
insert_goto_at(uint block_no,uint succ_no)504 void PhaseCFG::insert_goto_at(uint block_no, uint succ_no) {
505 // get block with block_no
506 assert(block_no < number_of_blocks(), "illegal block number");
507 Block* in = get_block(block_no);
508 // get successor block succ_no
509 assert(succ_no < in->_num_succs, "illegal successor number");
510 Block* out = in->_succs[succ_no];
511 // Compute frequency of the new block. Do this before inserting
512 // new block in case succ_prob() needs to infer the probability from
513 // surrounding blocks.
514 float freq = in->_freq * in->succ_prob(succ_no);
515 // get ProjNode corresponding to the succ_no'th successor of the in block
516 ProjNode* proj = in->get_node(in->number_of_nodes() - in->_num_succs + succ_no)->as_Proj();
517 // create region for basic block
518 RegionNode* region = new RegionNode(2);
519 region->init_req(1, proj);
520 // setup corresponding basic block
521 Block* block = new (_block_arena) Block(_block_arena, region);
522 map_node_to_block(region, block);
523 C->regalloc()->set_bad(region->_idx);
524 // add a goto node
525 Node* gto = _goto->clone(); // get a new goto node
526 gto->set_req(0, region);
527 // add it to the basic block
528 block->push_node(gto);
529 map_node_to_block(gto, block);
530 C->regalloc()->set_bad(gto->_idx);
531 // hook up successor block
532 block->_succs.map(block->_num_succs++, out);
533 // remap successor's predecessors if necessary
534 for (uint i = 1; i < out->num_preds(); i++) {
535 if (out->pred(i) == proj) out->head()->set_req(i, gto);
536 }
537 // remap predecessor's successor to new block
538 in->_succs.map(succ_no, block);
539 // Set the frequency of the new block
540 block->_freq = freq;
541 // add new basic block to basic block list
542 add_block_at(block_no + 1, block);
543 }
544
545 // Does this block end in a multiway branch that cannot have the default case
546 // flipped for another case?
no_flip_branch(Block * b)547 static bool no_flip_branch(Block *b) {
548 int branch_idx = b->number_of_nodes() - b->_num_succs-1;
549 if (branch_idx < 1) {
550 return false;
551 }
552 Node *branch = b->get_node(branch_idx);
553 if (branch->is_Catch()) {
554 return true;
555 }
556 if (branch->is_Mach()) {
557 if (branch->is_MachNullCheck()) {
558 return true;
559 }
560 int iop = branch->as_Mach()->ideal_Opcode();
561 if (iop == Op_FastLock || iop == Op_FastUnlock) {
562 return true;
563 }
564 // Don't flip if branch has an implicit check.
565 if (branch->as_Mach()->is_TrapBasedCheckNode()) {
566 return true;
567 }
568 }
569 return false;
570 }
571
572 // Check for NeverBranch at block end. This needs to become a GOTO to the
573 // true target. NeverBranch are treated as a conditional branch that always
574 // goes the same direction for most of the optimizer and are used to give a
575 // fake exit path to infinite loops. At this late stage they need to turn
576 // into Goto's so that when you enter the infinite loop you indeed hang.
convert_NeverBranch_to_Goto(Block * b)577 void PhaseCFG::convert_NeverBranch_to_Goto(Block *b) {
578 // Find true target
579 int end_idx = b->end_idx();
580 int idx = b->get_node(end_idx+1)->as_Proj()->_con;
581 Block *succ = b->_succs[idx];
582 Node* gto = _goto->clone(); // get a new goto node
583 gto->set_req(0, b->head());
584 Node *bp = b->get_node(end_idx);
585 b->map_node(gto, end_idx); // Slam over NeverBranch
586 map_node_to_block(gto, b);
587 C->regalloc()->set_bad(gto->_idx);
588 b->pop_node(); // Yank projections
589 b->pop_node(); // Yank projections
590 b->_succs.map(0,succ); // Map only successor
591 b->_num_succs = 1;
592 // remap successor's predecessors if necessary
593 uint j;
594 for( j = 1; j < succ->num_preds(); j++)
595 if( succ->pred(j)->in(0) == bp )
596 succ->head()->set_req(j, gto);
597 // Kill alternate exit path
598 Block *dead = b->_succs[1-idx];
599 for( j = 1; j < dead->num_preds(); j++)
600 if( dead->pred(j)->in(0) == bp )
601 break;
602 // Scan through block, yanking dead path from
603 // all regions and phis.
604 dead->head()->del_req(j);
605 for( int k = 1; dead->get_node(k)->is_Phi(); k++ )
606 dead->get_node(k)->del_req(j);
607 }
608
609 // Helper function to move block bx to the slot following b_index. Return
610 // true if the move is successful, otherwise false
move_to_next(Block * bx,uint b_index)611 bool PhaseCFG::move_to_next(Block* bx, uint b_index) {
612 if (bx == NULL) return false;
613
614 // Return false if bx is already scheduled.
615 uint bx_index = bx->_pre_order;
616 if ((bx_index <= b_index) && (get_block(bx_index) == bx)) {
617 return false;
618 }
619
620 // Find the current index of block bx on the block list
621 bx_index = b_index + 1;
622 while (bx_index < number_of_blocks() && get_block(bx_index) != bx) {
623 bx_index++;
624 }
625 assert(get_block(bx_index) == bx, "block not found");
626
627 // If the previous block conditionally falls into bx, return false,
628 // because moving bx will create an extra jump.
629 for(uint k = 1; k < bx->num_preds(); k++ ) {
630 Block* pred = get_block_for_node(bx->pred(k));
631 if (pred == get_block(bx_index - 1)) {
632 if (pred->_num_succs != 1) {
633 return false;
634 }
635 }
636 }
637
638 // Reinsert bx just past block 'b'
639 _blocks.remove(bx_index);
640 _blocks.insert(b_index + 1, bx);
641 return true;
642 }
643
644 // Move empty and uncommon blocks to the end.
move_to_end(Block * b,uint i)645 void PhaseCFG::move_to_end(Block *b, uint i) {
646 int e = b->is_Empty();
647 if (e != Block::not_empty) {
648 if (e == Block::empty_with_goto) {
649 // Remove the goto, but leave the block.
650 b->pop_node();
651 }
652 // Mark this block as a connector block, which will cause it to be
653 // ignored in certain functions such as non_connector_successor().
654 b->set_connector();
655 }
656 // Move the empty block to the end, and don't recheck.
657 _blocks.remove(i);
658 _blocks.push(b);
659 }
660
661 // Set loop alignment for every block
set_loop_alignment()662 void PhaseCFG::set_loop_alignment() {
663 uint last = number_of_blocks();
664 assert(get_block(0) == get_root_block(), "");
665
666 for (uint i = 1; i < last; i++) {
667 Block* block = get_block(i);
668 if (block->head()->is_Loop()) {
669 block->set_loop_alignment(block);
670 }
671 }
672 }
673
674 // Make empty basic blocks to be "connector" blocks, Move uncommon blocks
675 // to the end.
remove_empty_blocks()676 void PhaseCFG::remove_empty_blocks() {
677 // Move uncommon blocks to the end
678 uint last = number_of_blocks();
679 assert(get_block(0) == get_root_block(), "");
680
681 for (uint i = 1; i < last; i++) {
682 Block* block = get_block(i);
683 if (block->is_connector()) {
684 break;
685 }
686
687 // Check for NeverBranch at block end. This needs to become a GOTO to the
688 // true target. NeverBranch are treated as a conditional branch that
689 // always goes the same direction for most of the optimizer and are used
690 // to give a fake exit path to infinite loops. At this late stage they
691 // need to turn into Goto's so that when you enter the infinite loop you
692 // indeed hang.
693 if (block->get_node(block->end_idx())->Opcode() == Op_NeverBranch) {
694 convert_NeverBranch_to_Goto(block);
695 }
696
697 // Look for uncommon blocks and move to end.
698 if (!C->do_freq_based_layout()) {
699 if (is_uncommon(block)) {
700 move_to_end(block, i);
701 last--; // No longer check for being uncommon!
702 if (no_flip_branch(block)) { // Fall-thru case must follow?
703 // Find the fall-thru block
704 block = get_block(i);
705 move_to_end(block, i);
706 last--;
707 }
708 // backup block counter post-increment
709 i--;
710 }
711 }
712 }
713
714 // Move empty blocks to the end
715 last = number_of_blocks();
716 for (uint i = 1; i < last; i++) {
717 Block* block = get_block(i);
718 if (block->is_Empty() != Block::not_empty) {
719 move_to_end(block, i);
720 last--;
721 i--;
722 }
723 } // End of for all blocks
724 }
725
fixup_trap_based_check(Node * branch,Block * block,int block_pos,Block * bnext)726 Block *PhaseCFG::fixup_trap_based_check(Node *branch, Block *block, int block_pos, Block *bnext) {
727 // Trap based checks must fall through to the successor with
728 // PROB_ALWAYS.
729 // They should be an If with 2 successors.
730 assert(branch->is_MachIf(), "must be If");
731 assert(block->_num_succs == 2, "must have 2 successors");
732
733 // Get the If node and the projection for the first successor.
734 MachIfNode *iff = block->get_node(block->number_of_nodes()-3)->as_MachIf();
735 ProjNode *proj0 = block->get_node(block->number_of_nodes()-2)->as_Proj();
736 ProjNode *proj1 = block->get_node(block->number_of_nodes()-1)->as_Proj();
737 ProjNode *projt = (proj0->Opcode() == Op_IfTrue) ? proj0 : proj1;
738 ProjNode *projf = (proj0->Opcode() == Op_IfFalse) ? proj0 : proj1;
739
740 // Assert that proj0 and succs[0] match up. Similarly for proj1 and succs[1].
741 assert(proj0->raw_out(0) == block->_succs[0]->head(), "Mismatch successor 0");
742 assert(proj1->raw_out(0) == block->_succs[1]->head(), "Mismatch successor 1");
743
744 ProjNode *proj_always;
745 ProjNode *proj_never;
746 // We must negate the branch if the implicit check doesn't follow
747 // the branch's TRUE path. Then, the new TRUE branch target will
748 // be the old FALSE branch target.
749 if (iff->_prob <= 2*PROB_NEVER) { // There are small rounding errors.
750 proj_never = projt;
751 proj_always = projf;
752 } else {
753 // We must negate the branch if the trap doesn't follow the
754 // branch's TRUE path. Then, the new TRUE branch target will
755 // be the old FALSE branch target.
756 proj_never = projf;
757 proj_always = projt;
758 iff->negate();
759 }
760 assert(iff->_prob <= 2*PROB_NEVER, "Trap based checks are expected to trap never!");
761 // Map the successors properly
762 block->_succs.map(0, get_block_for_node(proj_never ->raw_out(0))); // The target of the trap.
763 block->_succs.map(1, get_block_for_node(proj_always->raw_out(0))); // The fall through target.
764
765 if (block->get_node(block->number_of_nodes() - block->_num_succs + 1) != proj_always) {
766 block->map_node(proj_never, block->number_of_nodes() - block->_num_succs + 0);
767 block->map_node(proj_always, block->number_of_nodes() - block->_num_succs + 1);
768 }
769
770 // Place the fall through block after this block.
771 Block *bs1 = block->non_connector_successor(1);
772 if (bs1 != bnext && move_to_next(bs1, block_pos)) {
773 bnext = bs1;
774 }
775 // If the fall through block still is not the next block, insert a goto.
776 if (bs1 != bnext) {
777 insert_goto_at(block_pos, 1);
778 }
779 return bnext;
780 }
781
782 // Fix up the final control flow for basic blocks.
fixup_flow()783 void PhaseCFG::fixup_flow() {
784 // Fixup final control flow for the blocks. Remove jump-to-next
785 // block. If neither arm of an IF follows the conditional branch, we
786 // have to add a second jump after the conditional. We place the
787 // TRUE branch target in succs[0] for both GOTOs and IFs.
788 for (uint i = 0; i < number_of_blocks(); i++) {
789 Block* block = get_block(i);
790 block->_pre_order = i; // turn pre-order into block-index
791
792 // Connector blocks need no further processing.
793 if (block->is_connector()) {
794 assert((i+1) == number_of_blocks() || get_block(i + 1)->is_connector(), "All connector blocks should sink to the end");
795 continue;
796 }
797 assert(block->is_Empty() != Block::completely_empty, "Empty blocks should be connectors");
798
799 Block* bnext = (i < number_of_blocks() - 1) ? get_block(i + 1) : NULL;
800 Block* bs0 = block->non_connector_successor(0);
801
802 // Check for multi-way branches where I cannot negate the test to
803 // exchange the true and false targets.
804 if (no_flip_branch(block)) {
805 // Find fall through case - if must fall into its target.
806 // Get the index of the branch's first successor.
807 int branch_idx = block->number_of_nodes() - block->_num_succs;
808
809 // The branch is 1 before the branch's first successor.
810 Node *branch = block->get_node(branch_idx-1);
811
812 // Handle no-flip branches which have implicit checks and which require
813 // special block ordering and individual semantics of the 'fall through
814 // case'.
815 if ((TrapBasedNullChecks || TrapBasedRangeChecks) &&
816 branch->is_Mach() && branch->as_Mach()->is_TrapBasedCheckNode()) {
817 bnext = fixup_trap_based_check(branch, block, i, bnext);
818 } else {
819 // Else, default handling for no-flip branches
820 for (uint j2 = 0; j2 < block->_num_succs; j2++) {
821 const ProjNode* p = block->get_node(branch_idx + j2)->as_Proj();
822 if (p->_con == 0) {
823 // successor j2 is fall through case
824 if (block->non_connector_successor(j2) != bnext) {
825 // but it is not the next block => insert a goto
826 insert_goto_at(i, j2);
827 }
828 // Put taken branch in slot 0
829 if (j2 == 0 && block->_num_succs == 2) {
830 // Flip targets in succs map
831 Block *tbs0 = block->_succs[0];
832 Block *tbs1 = block->_succs[1];
833 block->_succs.map(0, tbs1);
834 block->_succs.map(1, tbs0);
835 }
836 break;
837 }
838 }
839 }
840
841 // Remove all CatchProjs
842 for (uint j = 0; j < block->_num_succs; j++) {
843 block->pop_node();
844 }
845
846 } else if (block->_num_succs == 1) {
847 // Block ends in a Goto?
848 if (bnext == bs0) {
849 // We fall into next block; remove the Goto
850 block->pop_node();
851 }
852
853 } else if(block->_num_succs == 2) { // Block ends in a If?
854 // Get opcode of 1st projection (matches _succs[0])
855 // Note: Since this basic block has 2 exits, the last 2 nodes must
856 // be projections (in any order), the 3rd last node must be
857 // the IfNode (we have excluded other 2-way exits such as
858 // CatchNodes already).
859 MachNode* iff = block->get_node(block->number_of_nodes() - 3)->as_Mach();
860 ProjNode* proj0 = block->get_node(block->number_of_nodes() - 2)->as_Proj();
861 ProjNode* proj1 = block->get_node(block->number_of_nodes() - 1)->as_Proj();
862
863 // Assert that proj0 and succs[0] match up. Similarly for proj1 and succs[1].
864 assert(proj0->raw_out(0) == block->_succs[0]->head(), "Mismatch successor 0");
865 assert(proj1->raw_out(0) == block->_succs[1]->head(), "Mismatch successor 1");
866
867 Block* bs1 = block->non_connector_successor(1);
868
869 // Check for neither successor block following the current
870 // block ending in a conditional. If so, move one of the
871 // successors after the current one, provided that the
872 // successor was previously unscheduled, but moveable
873 // (i.e., all paths to it involve a branch).
874 if (!C->do_freq_based_layout() && bnext != bs0 && bnext != bs1) {
875 // Choose the more common successor based on the probability
876 // of the conditional branch.
877 Block* bx = bs0;
878 Block* by = bs1;
879
880 // _prob is the probability of taking the true path. Make
881 // p the probability of taking successor #1.
882 float p = iff->as_MachIf()->_prob;
883 if (proj0->Opcode() == Op_IfTrue) {
884 p = 1.0 - p;
885 }
886
887 // Prefer successor #1 if p > 0.5
888 if (p > PROB_FAIR) {
889 bx = bs1;
890 by = bs0;
891 }
892
893 // Attempt the more common successor first
894 if (move_to_next(bx, i)) {
895 bnext = bx;
896 } else if (move_to_next(by, i)) {
897 bnext = by;
898 }
899 }
900
901 // Check for conditional branching the wrong way. Negate
902 // conditional, if needed, so it falls into the following block
903 // and branches to the not-following block.
904
905 // Check for the next block being in succs[0]. We are going to branch
906 // to succs[0], so we want the fall-thru case as the next block in
907 // succs[1].
908 if (bnext == bs0) {
909 // Fall-thru case in succs[0], so flip targets in succs map
910 Block* tbs0 = block->_succs[0];
911 Block* tbs1 = block->_succs[1];
912 block->_succs.map(0, tbs1);
913 block->_succs.map(1, tbs0);
914 // Flip projection for each target
915 ProjNode* tmp = proj0;
916 proj0 = proj1;
917 proj1 = tmp;
918
919 } else if(bnext != bs1) {
920 // Need a double-branch
921 // The existing conditional branch need not change.
922 // Add a unconditional branch to the false target.
923 // Alas, it must appear in its own block and adding a
924 // block this late in the game is complicated. Sigh.
925 insert_goto_at(i, 1);
926 }
927
928 // Make sure we TRUE branch to the target
929 if (proj0->Opcode() == Op_IfFalse) {
930 iff->as_MachIf()->negate();
931 }
932
933 block->pop_node(); // Remove IfFalse & IfTrue projections
934 block->pop_node();
935
936 } else {
937 // Multi-exit block, e.g. a switch statement
938 // But we don't need to do anything here
939 }
940 } // End of for all blocks
941 }
942
943
944 // postalloc_expand: Expand nodes after register allocation.
945 //
946 // postalloc_expand has to be called after register allocation, just
947 // before output (i.e. scheduling). It only gets called if
948 // Matcher::require_postalloc_expand is true.
949 //
950 // Background:
951 //
952 // Nodes that are expandend (one compound node requiring several
953 // assembler instructions to be implemented split into two or more
954 // non-compound nodes) after register allocation are not as nice as
955 // the ones expanded before register allocation - they don't
956 // participate in optimizations as global code motion. But after
957 // register allocation we can expand nodes that use registers which
958 // are not spillable or registers that are not allocated, because the
959 // old compound node is simply replaced (in its location in the basic
960 // block) by a new subgraph which does not contain compound nodes any
961 // more. The scheduler called during output can later on process these
962 // non-compound nodes.
963 //
964 // Implementation:
965 //
966 // Nodes requiring postalloc expand are specified in the ad file by using
967 // a postalloc_expand statement instead of ins_encode. A postalloc_expand
968 // contains a single call to an encoding, as does an ins_encode
969 // statement. Instead of an emit() function a postalloc_expand() function
970 // is generated that doesn't emit assembler but creates a new
971 // subgraph. The code below calls this postalloc_expand function for each
972 // node with the appropriate attribute. This function returns the new
973 // nodes generated in an array passed in the call. The old node,
974 // potential MachTemps before and potential Projs after it then get
975 // disconnected and replaced by the new nodes. The instruction
976 // generating the result has to be the last one in the array. In
977 // general it is assumed that Projs after the node expanded are
978 // kills. These kills are not required any more after expanding as
979 // there are now explicitly visible def-use chains and the Projs are
980 // removed. This does not hold for calls: They do not only have
981 // kill-Projs but also Projs defining values. Therefore Projs after
982 // the node expanded are removed for all but for calls. If a node is
983 // to be reused, it must be added to the nodes list returned, and it
984 // will be added again.
985 //
986 // Implementing the postalloc_expand function for a node in an enc_class
987 // is rather tedious. It requires knowledge about many node details, as
988 // the nodes and the subgraph must be hand crafted. To simplify this,
989 // adlc generates some utility variables into the postalloc_expand function,
990 // e.g., holding the operands as specified by the postalloc_expand encoding
991 // specification, e.g.:
992 // * unsigned idx_<par_name> holding the index of the node in the ins
993 // * Node *n_<par_name> holding the node loaded from the ins
994 // * MachOpnd *op_<par_name> holding the corresponding operand
995 //
996 // The ordering of operands can not be determined by looking at a
997 // rule. Especially if a match rule matches several different trees,
998 // several nodes are generated from one instruct specification with
999 // different operand orderings. In this case the adlc generated
1000 // variables are the only way to access the ins and operands
1001 // deterministically.
1002 //
1003 // If assigning a register to a node that contains an oop, don't
1004 // forget to call ra_->set_oop() for the node.
postalloc_expand(PhaseRegAlloc * _ra)1005 void PhaseCFG::postalloc_expand(PhaseRegAlloc* _ra) {
1006 GrowableArray <Node *> new_nodes(32); // Array with new nodes filled by postalloc_expand function of node.
1007 GrowableArray <Node *> remove(32);
1008 GrowableArray <Node *> succs(32);
1009 unsigned int max_idx = C->unique(); // Remember to distinguish new from old nodes.
1010 DEBUG_ONLY(bool foundNode = false);
1011
1012 // for all blocks
1013 for (uint i = 0; i < number_of_blocks(); i++) {
1014 Block *b = _blocks[i];
1015 // For all instructions in the current block.
1016 for (uint j = 0; j < b->number_of_nodes(); j++) {
1017 Node *n = b->get_node(j);
1018 if (n->is_Mach() && n->as_Mach()->requires_postalloc_expand()) {
1019 #ifdef ASSERT
1020 if (TracePostallocExpand) {
1021 if (!foundNode) {
1022 foundNode = true;
1023 tty->print("POSTALLOC EXPANDING %d %s\n", C->compile_id(),
1024 C->method() ? C->method()->name()->as_utf8() : C->stub_name());
1025 }
1026 tty->print(" postalloc expanding "); n->dump();
1027 if (Verbose) {
1028 tty->print(" with ins:\n");
1029 for (uint k = 0; k < n->len(); ++k) {
1030 if (n->in(k)) { tty->print(" "); n->in(k)->dump(); }
1031 }
1032 }
1033 }
1034 #endif
1035 new_nodes.clear();
1036 // Collect nodes that have to be removed from the block later on.
1037 uint req = n->req();
1038 remove.clear();
1039 for (uint k = 0; k < req; ++k) {
1040 if (n->in(k) && n->in(k)->is_MachTemp()) {
1041 remove.push(n->in(k)); // MachTemps which are inputs to the old node have to be removed.
1042 n->in(k)->del_req(0);
1043 j--;
1044 }
1045 }
1046
1047 // Check whether we can allocate enough nodes. We set a fix limit for
1048 // the size of postalloc expands with this.
1049 uint unique_limit = C->unique() + 40;
1050 if (unique_limit >= _ra->node_regs_max_index()) {
1051 Compile::current()->record_failure("out of nodes in postalloc expand");
1052 return;
1053 }
1054
1055 // Emit (i.e. generate new nodes).
1056 n->as_Mach()->postalloc_expand(&new_nodes, _ra);
1057
1058 assert(C->unique() < unique_limit, "You allocated too many nodes in your postalloc expand.");
1059
1060 // Disconnect the inputs of the old node.
1061 //
1062 // We reuse MachSpillCopy nodes. If we need to expand them, there
1063 // are many, so reusing pays off. If reused, the node already
1064 // has the new ins. n must be the last node on new_nodes list.
1065 if (!n->is_MachSpillCopy()) {
1066 for (int k = req - 1; k >= 0; --k) {
1067 n->del_req(k);
1068 }
1069 }
1070
1071 #ifdef ASSERT
1072 // Check that all nodes have proper operands.
1073 for (int k = 0; k < new_nodes.length(); ++k) {
1074 if (new_nodes.at(k)->_idx < max_idx || !new_nodes.at(k)->is_Mach()) continue; // old node, Proj ...
1075 MachNode *m = new_nodes.at(k)->as_Mach();
1076 for (unsigned int l = 0; l < m->num_opnds(); ++l) {
1077 if (MachOper::notAnOper(m->_opnds[l])) {
1078 outputStream *os = tty;
1079 os->print("Node %s ", m->Name());
1080 os->print("has invalid opnd %d: %p\n", l, m->_opnds[l]);
1081 assert(0, "Invalid operands, see inline trace in hs_err_pid file.");
1082 }
1083 }
1084 }
1085 #endif
1086
1087 // Collect succs of old node in remove (for projections) and in succs (for
1088 // all other nodes) do _not_ collect projections in remove (but in succs)
1089 // in case the node is a call. We need the projections for calls as they are
1090 // associated with registes (i.e. they are defs).
1091 succs.clear();
1092 for (DUIterator k = n->outs(); n->has_out(k); k++) {
1093 if (n->out(k)->is_Proj() && !n->is_MachCall() && !n->is_MachBranch()) {
1094 remove.push(n->out(k));
1095 } else {
1096 succs.push(n->out(k));
1097 }
1098 }
1099 // Replace old node n as input of its succs by last of the new nodes.
1100 for (int k = 0; k < succs.length(); ++k) {
1101 Node *succ = succs.at(k);
1102 for (uint l = 0; l < succ->req(); ++l) {
1103 if (succ->in(l) == n) {
1104 succ->set_req(l, new_nodes.at(new_nodes.length() - 1));
1105 }
1106 }
1107 for (uint l = succ->req(); l < succ->len(); ++l) {
1108 if (succ->in(l) == n) {
1109 succ->set_prec(l, new_nodes.at(new_nodes.length() - 1));
1110 }
1111 }
1112 }
1113
1114 // Index of old node in block.
1115 uint index = b->find_node(n);
1116 // Insert new nodes into block and map them in nodes->blocks array
1117 // and remember last node in n2.
1118 Node *n2 = NULL;
1119 for (int k = 0; k < new_nodes.length(); ++k) {
1120 n2 = new_nodes.at(k);
1121 b->insert_node(n2, ++index);
1122 map_node_to_block(n2, b);
1123 }
1124
1125 // Add old node n to remove and remove them all from block.
1126 remove.push(n);
1127 j--;
1128 #ifdef ASSERT
1129 if (TracePostallocExpand && Verbose) {
1130 tty->print(" removing:\n");
1131 for (int k = 0; k < remove.length(); ++k) {
1132 tty->print(" "); remove.at(k)->dump();
1133 }
1134 tty->print(" inserting:\n");
1135 for (int k = 0; k < new_nodes.length(); ++k) {
1136 tty->print(" "); new_nodes.at(k)->dump();
1137 }
1138 }
1139 #endif
1140 for (int k = 0; k < remove.length(); ++k) {
1141 if (b->contains(remove.at(k))) {
1142 b->find_remove(remove.at(k));
1143 } else {
1144 assert(remove.at(k)->is_Proj() && (remove.at(k)->in(0)->is_MachBranch()), "");
1145 }
1146 }
1147 // If anything has been inserted (n2 != NULL), continue after last node inserted.
1148 // This does not always work. Some postalloc expands don't insert any nodes, if they
1149 // do optimizations (e.g., max(x,x)). In this case we decrement j accordingly.
1150 j = n2 ? b->find_node(n2) : j;
1151 }
1152 }
1153 }
1154
1155 #ifdef ASSERT
1156 if (foundNode) {
1157 tty->print("FINISHED %d %s\n", C->compile_id(),
1158 C->method() ? C->method()->name()->as_utf8() : C->stub_name());
1159 tty->flush();
1160 }
1161 #endif
1162 }
1163
1164
1165 //------------------------------dump-------------------------------------------
1166 #ifndef PRODUCT
_dump_cfg(const Node * end,VectorSet & visited) const1167 void PhaseCFG::_dump_cfg( const Node *end, VectorSet &visited ) const {
1168 const Node *x = end->is_block_proj();
1169 assert( x, "not a CFG" );
1170
1171 // Do not visit this block again
1172 if( visited.test_set(x->_idx) ) return;
1173
1174 // Skip through this block
1175 const Node *p = x;
1176 do {
1177 p = p->in(0); // Move control forward
1178 assert( !p->is_block_proj() || p->is_Root(), "not a CFG" );
1179 } while( !p->is_block_start() );
1180
1181 // Recursively visit
1182 for (uint i = 1; i < p->req(); i++) {
1183 _dump_cfg(p->in(i), visited);
1184 }
1185
1186 // Dump the block
1187 get_block_for_node(p)->dump(this);
1188 }
1189
dump() const1190 void PhaseCFG::dump( ) const {
1191 tty->print("\n--- CFG --- %d BBs\n", number_of_blocks());
1192 if (_blocks.size()) { // Did we do basic-block layout?
1193 for (uint i = 0; i < number_of_blocks(); i++) {
1194 const Block* block = get_block(i);
1195 block->dump(this);
1196 }
1197 } else { // Else do it with a DFS
1198 VectorSet visited(_block_arena);
1199 _dump_cfg(_root,visited);
1200 }
1201 }
1202
dump_headers()1203 void PhaseCFG::dump_headers() {
1204 for (uint i = 0; i < number_of_blocks(); i++) {
1205 Block* block = get_block(i);
1206 if (block != NULL) {
1207 block->dump_head(this);
1208 }
1209 }
1210 }
1211
verify() const1212 void PhaseCFG::verify() const {
1213 #ifdef ASSERT
1214 // Verify sane CFG
1215 for (uint i = 0; i < number_of_blocks(); i++) {
1216 Block* block = get_block(i);
1217 uint cnt = block->number_of_nodes();
1218 uint j;
1219 for (j = 0; j < cnt; j++) {
1220 Node *n = block->get_node(j);
1221 assert(get_block_for_node(n) == block, "");
1222 if (j >= 1 && n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_CreateEx) {
1223 assert(j == 1 || block->get_node(j-1)->is_Phi(), "CreateEx must be first instruction in block");
1224 }
1225 if (n->needs_anti_dependence_check()) {
1226 verify_anti_dependences(block, n);
1227 }
1228 for (uint k = 0; k < n->req(); k++) {
1229 Node *def = n->in(k);
1230 if (def && def != n) {
1231 assert(get_block_for_node(def) || def->is_Con(), "must have block; constants for debug info ok");
1232 // Verify that instructions in the block is in correct order.
1233 // Uses must follow their definition if they are at the same block.
1234 // Mostly done to check that MachSpillCopy nodes are placed correctly
1235 // when CreateEx node is moved in build_ifg_physical().
1236 if (get_block_for_node(def) == block && !(block->head()->is_Loop() && n->is_Phi()) &&
1237 // See (+++) comment in reg_split.cpp
1238 !(n->jvms() != NULL && n->jvms()->is_monitor_use(k))) {
1239 bool is_loop = false;
1240 if (n->is_Phi()) {
1241 for (uint l = 1; l < def->req(); l++) {
1242 if (n == def->in(l)) {
1243 is_loop = true;
1244 break; // Some kind of loop
1245 }
1246 }
1247 }
1248 assert(is_loop || block->find_node(def) < j, "uses must follow definitions");
1249 }
1250 }
1251 }
1252 }
1253
1254 j = block->end_idx();
1255 Node* bp = (Node*)block->get_node(block->number_of_nodes() - 1)->is_block_proj();
1256 assert(bp, "last instruction must be a block proj");
1257 assert(bp == block->get_node(j), "wrong number of successors for this block");
1258 if (bp->is_Catch()) {
1259 while (block->get_node(--j)->is_MachProj()) {
1260 ;
1261 }
1262 assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call");
1263 } else if (bp->is_Mach() && bp->as_Mach()->ideal_Opcode() == Op_If) {
1264 assert(block->_num_succs == 2, "Conditional branch must have two targets");
1265 }
1266 }
1267 #endif
1268 }
1269 #endif
1270
UnionFind(uint max)1271 UnionFind::UnionFind( uint max ) : _cnt(max), _max(max), _indices(NEW_RESOURCE_ARRAY(uint,max)) {
1272 Copy::zero_to_bytes( _indices, sizeof(uint)*max );
1273 }
1274
extend(uint from_idx,uint to_idx)1275 void UnionFind::extend( uint from_idx, uint to_idx ) {
1276 _nesting.check();
1277 if( from_idx >= _max ) {
1278 uint size = 16;
1279 while( size <= from_idx ) size <<=1;
1280 _indices = REALLOC_RESOURCE_ARRAY( uint, _indices, _max, size );
1281 _max = size;
1282 }
1283 while( _cnt <= from_idx ) _indices[_cnt++] = 0;
1284 _indices[from_idx] = to_idx;
1285 }
1286
reset(uint max)1287 void UnionFind::reset( uint max ) {
1288 // Force the Union-Find mapping to be at least this large
1289 extend(max,0);
1290 // Initialize to be the ID mapping.
1291 for( uint i=0; i<max; i++ ) map(i,i);
1292 }
1293
1294 // Straight out of Tarjan's union-find algorithm
Find_compress(uint idx)1295 uint UnionFind::Find_compress( uint idx ) {
1296 uint cur = idx;
1297 uint next = lookup(cur);
1298 while( next != cur ) { // Scan chain of equivalences
1299 assert( next < cur, "always union smaller" );
1300 cur = next; // until find a fixed-point
1301 next = lookup(cur);
1302 }
1303 // Core of union-find algorithm: update chain of
1304 // equivalences to be equal to the root.
1305 while( idx != next ) {
1306 uint tmp = lookup(idx);
1307 map(idx, next);
1308 idx = tmp;
1309 }
1310 return idx;
1311 }
1312
1313 // Like Find above, but no path compress, so bad asymptotic behavior
Find_const(uint idx) const1314 uint UnionFind::Find_const( uint idx ) const {
1315 if( idx == 0 ) return idx; // Ignore the zero idx
1316 // Off the end? This can happen during debugging dumps
1317 // when data structures have not finished being updated.
1318 if( idx >= _max ) return idx;
1319 uint next = lookup(idx);
1320 while( next != idx ) { // Scan chain of equivalences
1321 idx = next; // until find a fixed-point
1322 next = lookup(idx);
1323 }
1324 return next;
1325 }
1326
1327 // union 2 sets together.
Union(uint idx1,uint idx2)1328 void UnionFind::Union( uint idx1, uint idx2 ) {
1329 uint src = Find(idx1);
1330 uint dst = Find(idx2);
1331 assert( src, "" );
1332 assert( dst, "" );
1333 assert( src < _max, "oob" );
1334 assert( dst < _max, "oob" );
1335 assert( src < dst, "always union smaller" );
1336 map(dst,src);
1337 }
1338
1339 #ifndef PRODUCT
dump() const1340 void Trace::dump( ) const {
1341 tty->print_cr("Trace (freq %f)", first_block()->_freq);
1342 for (Block *b = first_block(); b != NULL; b = next(b)) {
1343 tty->print(" B%d", b->_pre_order);
1344 if (b->head()->is_Loop()) {
1345 tty->print(" (L%d)", b->compute_loop_alignment());
1346 }
1347 if (b->has_loop_alignment()) {
1348 tty->print(" (T%d)", b->code_alignment());
1349 }
1350 }
1351 tty->cr();
1352 }
1353
dump() const1354 void CFGEdge::dump( ) const {
1355 tty->print(" B%d --> B%d Freq: %f out:%3d%% in:%3d%% State: ",
1356 from()->_pre_order, to()->_pre_order, freq(), _from_pct, _to_pct);
1357 switch(state()) {
1358 case connected:
1359 tty->print("connected");
1360 break;
1361 case open:
1362 tty->print("open");
1363 break;
1364 case interior:
1365 tty->print("interior");
1366 break;
1367 }
1368 if (infrequent()) {
1369 tty->print(" infrequent");
1370 }
1371 tty->cr();
1372 }
1373 #endif
1374
1375 // Comparison function for edges
edge_order(CFGEdge ** e0,CFGEdge ** e1)1376 static int edge_order(CFGEdge **e0, CFGEdge **e1) {
1377 float freq0 = (*e0)->freq();
1378 float freq1 = (*e1)->freq();
1379 if (freq0 != freq1) {
1380 return freq0 > freq1 ? -1 : 1;
1381 }
1382
1383 int dist0 = (*e0)->to()->_rpo - (*e0)->from()->_rpo;
1384 int dist1 = (*e1)->to()->_rpo - (*e1)->from()->_rpo;
1385
1386 return dist1 - dist0;
1387 }
1388
1389 // Comparison function for edges
trace_frequency_order(const void * p0,const void * p1)1390 extern "C" int trace_frequency_order(const void *p0, const void *p1) {
1391 Trace *tr0 = *(Trace **) p0;
1392 Trace *tr1 = *(Trace **) p1;
1393 Block *b0 = tr0->first_block();
1394 Block *b1 = tr1->first_block();
1395
1396 // The trace of connector blocks goes at the end;
1397 // we only expect one such trace
1398 if (b0->is_connector() != b1->is_connector()) {
1399 return b1->is_connector() ? -1 : 1;
1400 }
1401
1402 // Pull more frequently executed blocks to the beginning
1403 float freq0 = b0->_freq;
1404 float freq1 = b1->_freq;
1405 if (freq0 != freq1) {
1406 return freq0 > freq1 ? -1 : 1;
1407 }
1408
1409 int diff = tr0->first_block()->_rpo - tr1->first_block()->_rpo;
1410
1411 return diff;
1412 }
1413
1414 // Find edges of interest, i.e, those which can fall through. Presumes that
1415 // edges which don't fall through are of low frequency and can be generally
1416 // ignored. Initialize the list of traces.
find_edges()1417 void PhaseBlockLayout::find_edges() {
1418 // Walk the blocks, creating edges and Traces
1419 uint i;
1420 Trace *tr = NULL;
1421 for (i = 0; i < _cfg.number_of_blocks(); i++) {
1422 Block* b = _cfg.get_block(i);
1423 tr = new Trace(b, next, prev);
1424 traces[tr->id()] = tr;
1425
1426 // All connector blocks should be at the end of the list
1427 if (b->is_connector()) break;
1428
1429 // If this block and the next one have a one-to-one successor
1430 // predecessor relationship, simply append the next block
1431 int nfallthru = b->num_fall_throughs();
1432 while (nfallthru == 1 &&
1433 b->succ_fall_through(0)) {
1434 Block *n = b->_succs[0];
1435
1436 // Skip over single-entry connector blocks, we don't want to
1437 // add them to the trace.
1438 while (n->is_connector() && n->num_preds() == 1) {
1439 n = n->_succs[0];
1440 }
1441
1442 // We see a merge point, so stop search for the next block
1443 if (n->num_preds() != 1) break;
1444
1445 i++;
1446 assert(n == _cfg.get_block(i), "expecting next block");
1447 tr->append(n);
1448 uf->map(n->_pre_order, tr->id());
1449 traces[n->_pre_order] = NULL;
1450 nfallthru = b->num_fall_throughs();
1451 b = n;
1452 }
1453
1454 if (nfallthru > 0) {
1455 // Create a CFGEdge for each outgoing
1456 // edge that could be a fall-through.
1457 for (uint j = 0; j < b->_num_succs; j++ ) {
1458 if (b->succ_fall_through(j)) {
1459 Block *target = b->non_connector_successor(j);
1460 float freq = b->_freq * b->succ_prob(j);
1461 int from_pct = (int) ((100 * freq) / b->_freq);
1462 int to_pct = (int) ((100 * freq) / target->_freq);
1463 edges->append(new CFGEdge(b, target, freq, from_pct, to_pct));
1464 }
1465 }
1466 }
1467 }
1468
1469 // Group connector blocks into one trace
1470 for (i++; i < _cfg.number_of_blocks(); i++) {
1471 Block *b = _cfg.get_block(i);
1472 assert(b->is_connector(), "connector blocks at the end");
1473 tr->append(b);
1474 uf->map(b->_pre_order, tr->id());
1475 traces[b->_pre_order] = NULL;
1476 }
1477 }
1478
1479 // Union two traces together in uf, and null out the trace in the list
union_traces(Trace * updated_trace,Trace * old_trace)1480 void PhaseBlockLayout::union_traces(Trace* updated_trace, Trace* old_trace) {
1481 uint old_id = old_trace->id();
1482 uint updated_id = updated_trace->id();
1483
1484 uint lo_id = updated_id;
1485 uint hi_id = old_id;
1486
1487 // If from is greater than to, swap values to meet
1488 // UnionFind guarantee.
1489 if (updated_id > old_id) {
1490 lo_id = old_id;
1491 hi_id = updated_id;
1492
1493 // Fix up the trace ids
1494 traces[lo_id] = traces[updated_id];
1495 updated_trace->set_id(lo_id);
1496 }
1497
1498 // Union the lower with the higher and remove the pointer
1499 // to the higher.
1500 uf->Union(lo_id, hi_id);
1501 traces[hi_id] = NULL;
1502 }
1503
1504 // Append traces together via the most frequently executed edges
grow_traces()1505 void PhaseBlockLayout::grow_traces() {
1506 // Order the edges, and drive the growth of Traces via the most
1507 // frequently executed edges.
1508 edges->sort(edge_order);
1509 for (int i = 0; i < edges->length(); i++) {
1510 CFGEdge *e = edges->at(i);
1511
1512 if (e->state() != CFGEdge::open) continue;
1513
1514 Block *src_block = e->from();
1515 Block *targ_block = e->to();
1516
1517 // Don't grow traces along backedges?
1518 if (!BlockLayoutRotateLoops) {
1519 if (targ_block->_rpo <= src_block->_rpo) {
1520 targ_block->set_loop_alignment(targ_block);
1521 continue;
1522 }
1523 }
1524
1525 Trace *src_trace = trace(src_block);
1526 Trace *targ_trace = trace(targ_block);
1527
1528 // If the edge in question can join two traces at their ends,
1529 // append one trace to the other.
1530 if (src_trace->last_block() == src_block) {
1531 if (src_trace == targ_trace) {
1532 e->set_state(CFGEdge::interior);
1533 if (targ_trace->backedge(e)) {
1534 // Reset i to catch any newly eligible edge
1535 // (Or we could remember the first "open" edge, and reset there)
1536 i = 0;
1537 }
1538 } else if (targ_trace->first_block() == targ_block) {
1539 e->set_state(CFGEdge::connected);
1540 src_trace->append(targ_trace);
1541 union_traces(src_trace, targ_trace);
1542 }
1543 }
1544 }
1545 }
1546
1547 // Embed one trace into another, if the fork or join points are sufficiently
1548 // balanced.
merge_traces(bool fall_thru_only)1549 void PhaseBlockLayout::merge_traces(bool fall_thru_only) {
1550 // Walk the edge list a another time, looking at unprocessed edges.
1551 // Fold in diamonds
1552 for (int i = 0; i < edges->length(); i++) {
1553 CFGEdge *e = edges->at(i);
1554
1555 if (e->state() != CFGEdge::open) continue;
1556 if (fall_thru_only) {
1557 if (e->infrequent()) continue;
1558 }
1559
1560 Block *src_block = e->from();
1561 Trace *src_trace = trace(src_block);
1562 bool src_at_tail = src_trace->last_block() == src_block;
1563
1564 Block *targ_block = e->to();
1565 Trace *targ_trace = trace(targ_block);
1566 bool targ_at_start = targ_trace->first_block() == targ_block;
1567
1568 if (src_trace == targ_trace) {
1569 // This may be a loop, but we can't do much about it.
1570 e->set_state(CFGEdge::interior);
1571 continue;
1572 }
1573
1574 if (fall_thru_only) {
1575 // If the edge links the middle of two traces, we can't do anything.
1576 // Mark the edge and continue.
1577 if (!src_at_tail & !targ_at_start) {
1578 continue;
1579 }
1580
1581 // Don't grow traces along backedges?
1582 if (!BlockLayoutRotateLoops && (targ_block->_rpo <= src_block->_rpo)) {
1583 continue;
1584 }
1585
1586 // If both ends of the edge are available, why didn't we handle it earlier?
1587 assert(src_at_tail ^ targ_at_start, "Should have caught this edge earlier.");
1588
1589 if (targ_at_start) {
1590 // Insert the "targ" trace in the "src" trace if the insertion point
1591 // is a two way branch.
1592 // Better profitability check possible, but may not be worth it.
1593 // Someday, see if the this "fork" has an associated "join";
1594 // then make a policy on merging this trace at the fork or join.
1595 // For example, other things being equal, it may be better to place this
1596 // trace at the join point if the "src" trace ends in a two-way, but
1597 // the insertion point is one-way.
1598 assert(src_block->num_fall_throughs() == 2, "unexpected diamond");
1599 e->set_state(CFGEdge::connected);
1600 src_trace->insert_after(src_block, targ_trace);
1601 union_traces(src_trace, targ_trace);
1602 } else if (src_at_tail) {
1603 if (src_trace != trace(_cfg.get_root_block())) {
1604 e->set_state(CFGEdge::connected);
1605 targ_trace->insert_before(targ_block, src_trace);
1606 union_traces(targ_trace, src_trace);
1607 }
1608 }
1609 } else if (e->state() == CFGEdge::open) {
1610 // Append traces, even without a fall-thru connection.
1611 // But leave root entry at the beginning of the block list.
1612 if (targ_trace != trace(_cfg.get_root_block())) {
1613 e->set_state(CFGEdge::connected);
1614 src_trace->append(targ_trace);
1615 union_traces(src_trace, targ_trace);
1616 }
1617 }
1618 }
1619 }
1620
1621 // Order the sequence of the traces in some desirable way, and fixup the
1622 // jumps at the end of each block.
reorder_traces(int count)1623 void PhaseBlockLayout::reorder_traces(int count) {
1624 ResourceArea *area = Thread::current()->resource_area();
1625 Trace ** new_traces = NEW_ARENA_ARRAY(area, Trace *, count);
1626 Block_List worklist;
1627 int new_count = 0;
1628
1629 // Compact the traces.
1630 for (int i = 0; i < count; i++) {
1631 Trace *tr = traces[i];
1632 if (tr != NULL) {
1633 new_traces[new_count++] = tr;
1634 }
1635 }
1636
1637 // The entry block should be first on the new trace list.
1638 Trace *tr = trace(_cfg.get_root_block());
1639 assert(tr == new_traces[0], "entry trace misplaced");
1640
1641 // Sort the new trace list by frequency
1642 qsort(new_traces + 1, new_count - 1, sizeof(new_traces[0]), trace_frequency_order);
1643
1644 // Patch up the successor blocks
1645 _cfg.clear_blocks();
1646 for (int i = 0; i < new_count; i++) {
1647 Trace *tr = new_traces[i];
1648 if (tr != NULL) {
1649 tr->fixup_blocks(_cfg);
1650 }
1651 }
1652 }
1653
1654 // Order basic blocks based on frequency
PhaseBlockLayout(PhaseCFG & cfg)1655 PhaseBlockLayout::PhaseBlockLayout(PhaseCFG &cfg)
1656 : Phase(BlockLayout)
1657 , _cfg(cfg) {
1658 ResourceMark rm;
1659 ResourceArea *area = Thread::current()->resource_area();
1660
1661 // List of traces
1662 int size = _cfg.number_of_blocks() + 1;
1663 traces = NEW_ARENA_ARRAY(area, Trace *, size);
1664 memset(traces, 0, size*sizeof(Trace*));
1665 next = NEW_ARENA_ARRAY(area, Block *, size);
1666 memset(next, 0, size*sizeof(Block *));
1667 prev = NEW_ARENA_ARRAY(area, Block *, size);
1668 memset(prev , 0, size*sizeof(Block *));
1669
1670 // List of edges
1671 edges = new GrowableArray<CFGEdge*>;
1672
1673 // Mapping block index --> block_trace
1674 uf = new UnionFind(size);
1675 uf->reset(size);
1676
1677 // Find edges and create traces.
1678 find_edges();
1679
1680 // Grow traces at their ends via most frequent edges.
1681 grow_traces();
1682
1683 // Merge one trace into another, but only at fall-through points.
1684 // This may make diamonds and other related shapes in a trace.
1685 merge_traces(true);
1686
1687 // Run merge again, allowing two traces to be catenated, even if
1688 // one does not fall through into the other. This appends loosely
1689 // related traces to be near each other.
1690 merge_traces(false);
1691
1692 // Re-order all the remaining traces by frequency
1693 reorder_traces(size);
1694
1695 assert(_cfg.number_of_blocks() >= (uint) (size - 1), "number of blocks can not shrink");
1696 }
1697
1698
1699 // Edge e completes a loop in a trace. If the target block is head of the
1700 // loop, rotate the loop block so that the loop ends in a conditional branch.
backedge(CFGEdge * e)1701 bool Trace::backedge(CFGEdge *e) {
1702 bool loop_rotated = false;
1703 Block *src_block = e->from();
1704 Block *targ_block = e->to();
1705
1706 assert(last_block() == src_block, "loop discovery at back branch");
1707 if (first_block() == targ_block) {
1708 if (BlockLayoutRotateLoops && last_block()->num_fall_throughs() < 2) {
1709 // Find the last block in the trace that has a conditional
1710 // branch.
1711 Block *b;
1712 for (b = last_block(); b != NULL; b = prev(b)) {
1713 if (b->num_fall_throughs() == 2) {
1714 break;
1715 }
1716 }
1717
1718 if (b != last_block() && b != NULL) {
1719 loop_rotated = true;
1720
1721 // Rotate the loop by doing two-part linked-list surgery.
1722 append(first_block());
1723 break_loop_after(b);
1724 }
1725 }
1726
1727 // Backbranch to the top of a trace
1728 // Scroll forward through the trace from the targ_block. If we find
1729 // a loop head before another loop top, use the the loop head alignment.
1730 for (Block *b = targ_block; b != NULL; b = next(b)) {
1731 if (b->has_loop_alignment()) {
1732 break;
1733 }
1734 if (b->head()->is_Loop()) {
1735 targ_block = b;
1736 break;
1737 }
1738 }
1739
1740 first_block()->set_loop_alignment(targ_block);
1741
1742 } else {
1743 // That loop may already have a loop top (we're reaching it again
1744 // through the backedge of an outer loop)
1745 Block* b = prev(targ_block);
1746 bool has_top = targ_block->head()->is_Loop() && b->has_loop_alignment() && !b->head()->is_Loop();
1747 if (!has_top) {
1748 // Backbranch into the middle of a trace
1749 targ_block->set_loop_alignment(targ_block);
1750 }
1751 }
1752
1753 return loop_rotated;
1754 }
1755
1756 // push blocks onto the CFG list
1757 // ensure that blocks have the correct two-way branch sense
fixup_blocks(PhaseCFG & cfg)1758 void Trace::fixup_blocks(PhaseCFG &cfg) {
1759 Block *last = last_block();
1760 for (Block *b = first_block(); b != NULL; b = next(b)) {
1761 cfg.add_block(b);
1762 if (!b->is_connector()) {
1763 int nfallthru = b->num_fall_throughs();
1764 if (b != last) {
1765 if (nfallthru == 2) {
1766 // Ensure that the sense of the branch is correct
1767 Block *bnext = next(b);
1768 Block *bs0 = b->non_connector_successor(0);
1769
1770 MachNode *iff = b->get_node(b->number_of_nodes() - 3)->as_Mach();
1771 ProjNode *proj0 = b->get_node(b->number_of_nodes() - 2)->as_Proj();
1772 ProjNode *proj1 = b->get_node(b->number_of_nodes() - 1)->as_Proj();
1773
1774 if (bnext == bs0) {
1775 // Fall-thru case in succs[0], should be in succs[1]
1776
1777 // Flip targets in _succs map
1778 Block *tbs0 = b->_succs[0];
1779 Block *tbs1 = b->_succs[1];
1780 b->_succs.map( 0, tbs1 );
1781 b->_succs.map( 1, tbs0 );
1782
1783 // Flip projections to match targets
1784 b->map_node(proj1, b->number_of_nodes() - 2);
1785 b->map_node(proj0, b->number_of_nodes() - 1);
1786 }
1787 }
1788 }
1789 }
1790 }
1791 }
1792