|
1 /* |
|
2 * Copyright 1997-2006 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, |
|
20 * CA 95054 USA or visit www.sun.com if you need additional information or |
|
21 * have any questions. |
|
22 * |
|
23 */ |
|
24 |
|
25 // Optimization - Graph Style |
|
26 |
|
27 #include "incls/_precompiled.incl" |
|
28 #include "incls/_block.cpp.incl" |
|
29 |
|
30 |
|
31 //----------------------------------------------------------------------------- |
|
32 void Block_Array::grow( uint i ) { |
|
33 assert(i >= Max(), "must be an overflow"); |
|
34 debug_only(_limit = i+1); |
|
35 if( i < _size ) return; |
|
36 if( !_size ) { |
|
37 _size = 1; |
|
38 _blocks = (Block**)_arena->Amalloc( _size * sizeof(Block*) ); |
|
39 _blocks[0] = NULL; |
|
40 } |
|
41 uint old = _size; |
|
42 while( i >= _size ) _size <<= 1; // Double to fit |
|
43 _blocks = (Block**)_arena->Arealloc( _blocks, old*sizeof(Block*),_size*sizeof(Block*)); |
|
44 Copy::zero_to_bytes( &_blocks[old], (_size-old)*sizeof(Block*) ); |
|
45 } |
|
46 |
|
47 //============================================================================= |
|
48 void Block_List::remove(uint i) { |
|
49 assert(i < _cnt, "index out of bounds"); |
|
50 Copy::conjoint_words_to_lower((HeapWord*)&_blocks[i+1], (HeapWord*)&_blocks[i], ((_cnt-i-1)*sizeof(Block*))); |
|
51 pop(); // shrink list by one block |
|
52 } |
|
53 |
|
54 void Block_List::insert(uint i, Block *b) { |
|
55 push(b); // grow list by one block |
|
56 Copy::conjoint_words_to_higher((HeapWord*)&_blocks[i], (HeapWord*)&_blocks[i+1], ((_cnt-i-1)*sizeof(Block*))); |
|
57 _blocks[i] = b; |
|
58 } |
|
59 |
|
60 |
|
61 //============================================================================= |
|
62 |
|
63 uint Block::code_alignment() { |
|
64 // Check for Root block |
|
65 if( _pre_order == 0 ) return CodeEntryAlignment; |
|
66 // Check for Start block |
|
67 if( _pre_order == 1 ) return InteriorEntryAlignment; |
|
68 // Check for loop alignment |
|
69 Node *h = head(); |
|
70 if( h->is_Loop() && h->as_Loop()->is_inner_loop() ) { |
|
71 // Pre- and post-loops have low trip count so do not bother with |
|
72 // NOPs for align loop head. The constants are hidden from tuning |
|
73 // but only because my "divide by 4" heuristic surely gets nearly |
|
74 // all possible gain (a "do not align at all" heuristic has a |
|
75 // chance of getting a really tiny gain). |
|
76 if( h->is_CountedLoop() && (h->as_CountedLoop()->is_pre_loop() || |
|
77 h->as_CountedLoop()->is_post_loop()) ) |
|
78 return (OptoLoopAlignment > 4) ? (OptoLoopAlignment>>2) : 1; |
|
79 // Loops with low backedge frequency should not be aligned. |
|
80 Node *n = h->in(LoopNode::LoopBackControl)->in(0); |
|
81 if( n->is_MachIf() && n->as_MachIf()->_prob < 0.01 ) { |
|
82 return 1; // Loop does not loop, more often than not! |
|
83 } |
|
84 return OptoLoopAlignment; // Otherwise align loop head |
|
85 } |
|
86 return 1; // no particular alignment |
|
87 } |
|
88 |
|
89 //----------------------------------------------------------------------------- |
|
90 // Compute the size of first 'inst_cnt' instructions in this block. |
|
91 // Return the number of instructions left to compute if the block has |
|
92 // less then 'inst_cnt' instructions. |
|
93 uint Block::compute_first_inst_size(uint& sum_size, uint inst_cnt, |
|
94 PhaseRegAlloc* ra) { |
|
95 uint last_inst = _nodes.size(); |
|
96 for( uint j = 0; j < last_inst && inst_cnt > 0; j++ ) { |
|
97 uint inst_size = _nodes[j]->size(ra); |
|
98 if( inst_size > 0 ) { |
|
99 inst_cnt--; |
|
100 uint sz = sum_size + inst_size; |
|
101 if( sz <= (uint)OptoLoopAlignment ) { |
|
102 // Compute size of instructions which fit into fetch buffer only |
|
103 // since all inst_cnt instructions will not fit even if we align them. |
|
104 sum_size = sz; |
|
105 } else { |
|
106 return 0; |
|
107 } |
|
108 } |
|
109 } |
|
110 return inst_cnt; |
|
111 } |
|
112 |
|
113 //----------------------------------------------------------------------------- |
|
114 uint Block::find_node( const Node *n ) const { |
|
115 for( uint i = 0; i < _nodes.size(); i++ ) { |
|
116 if( _nodes[i] == n ) |
|
117 return i; |
|
118 } |
|
119 ShouldNotReachHere(); |
|
120 return 0; |
|
121 } |
|
122 |
|
123 // Find and remove n from block list |
|
124 void Block::find_remove( const Node *n ) { |
|
125 _nodes.remove(find_node(n)); |
|
126 } |
|
127 |
|
128 //------------------------------is_Empty--------------------------------------- |
|
129 // Return empty status of a block. Empty blocks contain only the head, other |
|
130 // ideal nodes, and an optional trailing goto. |
|
131 int Block::is_Empty() const { |
|
132 |
|
133 // Root or start block is not considered empty |
|
134 if (head()->is_Root() || head()->is_Start()) { |
|
135 return not_empty; |
|
136 } |
|
137 |
|
138 int success_result = completely_empty; |
|
139 int end_idx = _nodes.size()-1; |
|
140 |
|
141 // Check for ending goto |
|
142 if ((end_idx > 0) && (_nodes[end_idx]->is_Goto())) { |
|
143 success_result = empty_with_goto; |
|
144 end_idx--; |
|
145 } |
|
146 |
|
147 // Unreachable blocks are considered empty |
|
148 if (num_preds() <= 1) { |
|
149 return success_result; |
|
150 } |
|
151 |
|
152 // Ideal nodes are allowable in empty blocks: skip them Only MachNodes |
|
153 // turn directly into code, because only MachNodes have non-trivial |
|
154 // emit() functions. |
|
155 while ((end_idx > 0) && !_nodes[end_idx]->is_Mach()) { |
|
156 end_idx--; |
|
157 } |
|
158 |
|
159 // No room for any interesting instructions? |
|
160 if (end_idx == 0) { |
|
161 return success_result; |
|
162 } |
|
163 |
|
164 return not_empty; |
|
165 } |
|
166 |
|
167 //------------------------------has_uncommon_code------------------------------ |
|
168 // Return true if the block's code implies that it is not likely to be |
|
169 // executed infrequently. Check to see if the block ends in a Halt or |
|
170 // a low probability call. |
|
171 bool Block::has_uncommon_code() const { |
|
172 Node* en = end(); |
|
173 |
|
174 if (en->is_Goto()) |
|
175 en = en->in(0); |
|
176 if (en->is_Catch()) |
|
177 en = en->in(0); |
|
178 if (en->is_Proj() && en->in(0)->is_MachCall()) { |
|
179 MachCallNode* call = en->in(0)->as_MachCall(); |
|
180 if (call->cnt() != COUNT_UNKNOWN && call->cnt() <= PROB_UNLIKELY_MAG(4)) { |
|
181 // This is true for slow-path stubs like new_{instance,array}, |
|
182 // slow_arraycopy, complete_monitor_locking, uncommon_trap. |
|
183 // The magic number corresponds to the probability of an uncommon_trap, |
|
184 // even though it is a count not a probability. |
|
185 return true; |
|
186 } |
|
187 } |
|
188 |
|
189 int op = en->is_Mach() ? en->as_Mach()->ideal_Opcode() : en->Opcode(); |
|
190 return op == Op_Halt; |
|
191 } |
|
192 |
|
193 //------------------------------is_uncommon------------------------------------ |
|
194 // True if block is low enough frequency or guarded by a test which |
|
195 // mostly does not go here. |
|
196 bool Block::is_uncommon( Block_Array &bbs ) const { |
|
197 // Initial blocks must never be moved, so are never uncommon. |
|
198 if (head()->is_Root() || head()->is_Start()) return false; |
|
199 |
|
200 // Check for way-low freq |
|
201 if( _freq < BLOCK_FREQUENCY(0.00001f) ) return true; |
|
202 |
|
203 // Look for code shape indicating uncommon_trap or slow path |
|
204 if (has_uncommon_code()) return true; |
|
205 |
|
206 const float epsilon = 0.05f; |
|
207 const float guard_factor = PROB_UNLIKELY_MAG(4) / (1.f - epsilon); |
|
208 uint uncommon_preds = 0; |
|
209 uint freq_preds = 0; |
|
210 uint uncommon_for_freq_preds = 0; |
|
211 |
|
212 for( uint i=1; i<num_preds(); i++ ) { |
|
213 Block* guard = bbs[pred(i)->_idx]; |
|
214 // Check to see if this block follows its guard 1 time out of 10000 |
|
215 // or less. |
|
216 // |
|
217 // See list of magnitude-4 unlikely probabilities in cfgnode.hpp which |
|
218 // we intend to be "uncommon", such as slow-path TLE allocation, |
|
219 // predicted call failure, and uncommon trap triggers. |
|
220 // |
|
221 // Use an epsilon value of 5% to allow for variability in frequency |
|
222 // predictions and floating point calculations. The net effect is |
|
223 // that guard_factor is set to 9500. |
|
224 // |
|
225 // Ignore low-frequency blocks. |
|
226 // The next check is (guard->_freq < 1.e-5 * 9500.). |
|
227 if(guard->_freq*BLOCK_FREQUENCY(guard_factor) < BLOCK_FREQUENCY(0.00001f)) { |
|
228 uncommon_preds++; |
|
229 } else { |
|
230 freq_preds++; |
|
231 if( _freq < guard->_freq * guard_factor ) { |
|
232 uncommon_for_freq_preds++; |
|
233 } |
|
234 } |
|
235 } |
|
236 if( num_preds() > 1 && |
|
237 // The block is uncommon if all preds are uncommon or |
|
238 (uncommon_preds == (num_preds()-1) || |
|
239 // it is uncommon for all frequent preds. |
|
240 uncommon_for_freq_preds == freq_preds) ) { |
|
241 return true; |
|
242 } |
|
243 return false; |
|
244 } |
|
245 |
|
246 //------------------------------dump------------------------------------------- |
|
247 #ifndef PRODUCT |
|
248 void Block::dump_bidx(const Block* orig) const { |
|
249 if (_pre_order) tty->print("B%d",_pre_order); |
|
250 else tty->print("N%d", head()->_idx); |
|
251 |
|
252 if (Verbose && orig != this) { |
|
253 // Dump the original block's idx |
|
254 tty->print(" ("); |
|
255 orig->dump_bidx(orig); |
|
256 tty->print(")"); |
|
257 } |
|
258 } |
|
259 |
|
260 void Block::dump_pred(const Block_Array *bbs, Block* orig) const { |
|
261 if (is_connector()) { |
|
262 for (uint i=1; i<num_preds(); i++) { |
|
263 Block *p = ((*bbs)[pred(i)->_idx]); |
|
264 p->dump_pred(bbs, orig); |
|
265 } |
|
266 } else { |
|
267 dump_bidx(orig); |
|
268 tty->print(" "); |
|
269 } |
|
270 } |
|
271 |
|
272 void Block::dump_head( const Block_Array *bbs ) const { |
|
273 // Print the basic block |
|
274 dump_bidx(this); |
|
275 tty->print(": #\t"); |
|
276 |
|
277 // Print the incoming CFG edges and the outgoing CFG edges |
|
278 for( uint i=0; i<_num_succs; i++ ) { |
|
279 non_connector_successor(i)->dump_bidx(_succs[i]); |
|
280 tty->print(" "); |
|
281 } |
|
282 tty->print("<- "); |
|
283 if( head()->is_block_start() ) { |
|
284 for (uint i=1; i<num_preds(); i++) { |
|
285 Node *s = pred(i); |
|
286 if (bbs) { |
|
287 Block *p = (*bbs)[s->_idx]; |
|
288 p->dump_pred(bbs, p); |
|
289 } else { |
|
290 while (!s->is_block_start()) |
|
291 s = s->in(0); |
|
292 tty->print("N%d ", s->_idx ); |
|
293 } |
|
294 } |
|
295 } else |
|
296 tty->print("BLOCK HEAD IS JUNK "); |
|
297 |
|
298 // Print loop, if any |
|
299 const Block *bhead = this; // Head of self-loop |
|
300 Node *bh = bhead->head(); |
|
301 if( bbs && bh->is_Loop() && !head()->is_Root() ) { |
|
302 LoopNode *loop = bh->as_Loop(); |
|
303 const Block *bx = (*bbs)[loop->in(LoopNode::LoopBackControl)->_idx]; |
|
304 while (bx->is_connector()) { |
|
305 bx = (*bbs)[bx->pred(1)->_idx]; |
|
306 } |
|
307 tty->print("\tLoop: B%d-B%d ", bhead->_pre_order, bx->_pre_order); |
|
308 // Dump any loop-specific bits, especially for CountedLoops. |
|
309 loop->dump_spec(tty); |
|
310 } |
|
311 tty->print(" Freq: %g",_freq); |
|
312 if( Verbose || WizardMode ) { |
|
313 tty->print(" IDom: %d/#%d", _idom ? _idom->_pre_order : 0, _dom_depth); |
|
314 tty->print(" RegPressure: %d",_reg_pressure); |
|
315 tty->print(" IHRP Index: %d",_ihrp_index); |
|
316 tty->print(" FRegPressure: %d",_freg_pressure); |
|
317 tty->print(" FHRP Index: %d",_fhrp_index); |
|
318 } |
|
319 tty->print_cr(""); |
|
320 } |
|
321 |
|
322 void Block::dump() const { dump(0); } |
|
323 |
|
324 void Block::dump( const Block_Array *bbs ) const { |
|
325 dump_head(bbs); |
|
326 uint cnt = _nodes.size(); |
|
327 for( uint i=0; i<cnt; i++ ) |
|
328 _nodes[i]->dump(); |
|
329 tty->print("\n"); |
|
330 } |
|
331 #endif |
|
332 |
|
333 //============================================================================= |
|
334 //------------------------------PhaseCFG--------------------------------------- |
|
335 PhaseCFG::PhaseCFG( Arena *a, RootNode *r, Matcher &m ) : |
|
336 Phase(CFG), |
|
337 _bbs(a), |
|
338 _root(r) |
|
339 #ifndef PRODUCT |
|
340 , _trace_opto_pipelining(TraceOptoPipelining || C->method_has_option("TraceOptoPipelining")) |
|
341 #endif |
|
342 { |
|
343 ResourceMark rm; |
|
344 // I'll need a few machine-specific GotoNodes. Make an Ideal GotoNode, |
|
345 // then Match it into a machine-specific Node. Then clone the machine |
|
346 // Node on demand. |
|
347 Node *x = new (C, 1) GotoNode(NULL); |
|
348 x->init_req(0, x); |
|
349 _goto = m.match_tree(x); |
|
350 assert(_goto != NULL, ""); |
|
351 _goto->set_req(0,_goto); |
|
352 |
|
353 // Build the CFG in Reverse Post Order |
|
354 _num_blocks = build_cfg(); |
|
355 _broot = _bbs[_root->_idx]; |
|
356 } |
|
357 |
|
358 //------------------------------build_cfg-------------------------------------- |
|
359 // Build a proper looking CFG. Make every block begin with either a StartNode |
|
360 // or a RegionNode. Make every block end with either a Goto, If or Return. |
|
361 // The RootNode both starts and ends it's own block. Do this with a recursive |
|
362 // backwards walk over the control edges. |
|
363 uint PhaseCFG::build_cfg() { |
|
364 Arena *a = Thread::current()->resource_area(); |
|
365 VectorSet visited(a); |
|
366 |
|
367 // Allocate stack with enough space to avoid frequent realloc |
|
368 Node_Stack nstack(a, C->unique() >> 1); |
|
369 nstack.push(_root, 0); |
|
370 uint sum = 0; // Counter for blocks |
|
371 |
|
372 while (nstack.is_nonempty()) { |
|
373 // node and in's index from stack's top |
|
374 // 'np' is _root (see above) or RegionNode, StartNode: we push on stack |
|
375 // only nodes which point to the start of basic block (see below). |
|
376 Node *np = nstack.node(); |
|
377 // idx > 0, except for the first node (_root) pushed on stack |
|
378 // at the beginning when idx == 0. |
|
379 // We will use the condition (idx == 0) later to end the build. |
|
380 uint idx = nstack.index(); |
|
381 Node *proj = np->in(idx); |
|
382 const Node *x = proj->is_block_proj(); |
|
383 // Does the block end with a proper block-ending Node? One of Return, |
|
384 // If or Goto? (This check should be done for visited nodes also). |
|
385 if (x == NULL) { // Does not end right... |
|
386 Node *g = _goto->clone(); // Force it to end in a Goto |
|
387 g->set_req(0, proj); |
|
388 np->set_req(idx, g); |
|
389 x = proj = g; |
|
390 } |
|
391 if (!visited.test_set(x->_idx)) { // Visit this block once |
|
392 // Skip any control-pinned middle'in stuff |
|
393 Node *p = proj; |
|
394 do { |
|
395 proj = p; // Update pointer to last Control |
|
396 p = p->in(0); // Move control forward |
|
397 } while( !p->is_block_proj() && |
|
398 !p->is_block_start() ); |
|
399 // Make the block begin with one of Region or StartNode. |
|
400 if( !p->is_block_start() ) { |
|
401 RegionNode *r = new (C, 2) RegionNode( 2 ); |
|
402 r->init_req(1, p); // Insert RegionNode in the way |
|
403 proj->set_req(0, r); // Insert RegionNode in the way |
|
404 p = r; |
|
405 } |
|
406 // 'p' now points to the start of this basic block |
|
407 |
|
408 // Put self in array of basic blocks |
|
409 Block *bb = new (_bbs._arena) Block(_bbs._arena,p); |
|
410 _bbs.map(p->_idx,bb); |
|
411 _bbs.map(x->_idx,bb); |
|
412 if( x != p ) // Only for root is x == p |
|
413 bb->_nodes.push((Node*)x); |
|
414 |
|
415 // Now handle predecessors |
|
416 ++sum; // Count 1 for self block |
|
417 uint cnt = bb->num_preds(); |
|
418 for (int i = (cnt - 1); i > 0; i-- ) { // For all predecessors |
|
419 Node *prevproj = p->in(i); // Get prior input |
|
420 assert( !prevproj->is_Con(), "dead input not removed" ); |
|
421 // Check to see if p->in(i) is a "control-dependent" CFG edge - |
|
422 // i.e., it splits at the source (via an IF or SWITCH) and merges |
|
423 // at the destination (via a many-input Region). |
|
424 // This breaks critical edges. The RegionNode to start the block |
|
425 // will be added when <p,i> is pulled off the node stack |
|
426 if ( cnt > 2 ) { // Merging many things? |
|
427 assert( prevproj== bb->pred(i),""); |
|
428 if(prevproj->is_block_proj() != prevproj) { // Control-dependent edge? |
|
429 // Force a block on the control-dependent edge |
|
430 Node *g = _goto->clone(); // Force it to end in a Goto |
|
431 g->set_req(0,prevproj); |
|
432 p->set_req(i,g); |
|
433 } |
|
434 } |
|
435 nstack.push(p, i); // 'p' is RegionNode or StartNode |
|
436 } |
|
437 } else { // Post-processing visited nodes |
|
438 nstack.pop(); // remove node from stack |
|
439 // Check if it the fist node pushed on stack at the beginning. |
|
440 if (idx == 0) break; // end of the build |
|
441 // Find predecessor basic block |
|
442 Block *pb = _bbs[x->_idx]; |
|
443 // Insert into nodes array, if not already there |
|
444 if( !_bbs.lookup(proj->_idx) ) { |
|
445 assert( x != proj, "" ); |
|
446 // Map basic block of projection |
|
447 _bbs.map(proj->_idx,pb); |
|
448 pb->_nodes.push(proj); |
|
449 } |
|
450 // Insert self as a child of my predecessor block |
|
451 pb->_succs.map(pb->_num_succs++, _bbs[np->_idx]); |
|
452 assert( pb->_nodes[ pb->_nodes.size() - pb->_num_succs ]->is_block_proj(), |
|
453 "too many control users, not a CFG?" ); |
|
454 } |
|
455 } |
|
456 // Return number of basic blocks for all children and self |
|
457 return sum; |
|
458 } |
|
459 |
|
460 //------------------------------insert_goto_at--------------------------------- |
|
461 // Inserts a goto & corresponding basic block between |
|
462 // block[block_no] and its succ_no'th successor block |
|
463 void PhaseCFG::insert_goto_at(uint block_no, uint succ_no) { |
|
464 // get block with block_no |
|
465 assert(block_no < _num_blocks, "illegal block number"); |
|
466 Block* in = _blocks[block_no]; |
|
467 // get successor block succ_no |
|
468 assert(succ_no < in->_num_succs, "illegal successor number"); |
|
469 Block* out = in->_succs[succ_no]; |
|
470 // get ProjNode corresponding to the succ_no'th successor of the in block |
|
471 ProjNode* proj = in->_nodes[in->_nodes.size() - in->_num_succs + succ_no]->as_Proj(); |
|
472 // create region for basic block |
|
473 RegionNode* region = new (C, 2) RegionNode(2); |
|
474 region->init_req(1, proj); |
|
475 // setup corresponding basic block |
|
476 Block* block = new (_bbs._arena) Block(_bbs._arena, region); |
|
477 _bbs.map(region->_idx, block); |
|
478 C->regalloc()->set_bad(region->_idx); |
|
479 // add a goto node |
|
480 Node* gto = _goto->clone(); // get a new goto node |
|
481 gto->set_req(0, region); |
|
482 // add it to the basic block |
|
483 block->_nodes.push(gto); |
|
484 _bbs.map(gto->_idx, block); |
|
485 C->regalloc()->set_bad(gto->_idx); |
|
486 // hook up successor block |
|
487 block->_succs.map(block->_num_succs++, out); |
|
488 // remap successor's predecessors if necessary |
|
489 for (uint i = 1; i < out->num_preds(); i++) { |
|
490 if (out->pred(i) == proj) out->head()->set_req(i, gto); |
|
491 } |
|
492 // remap predecessor's successor to new block |
|
493 in->_succs.map(succ_no, block); |
|
494 // add new basic block to basic block list |
|
495 _blocks.insert(block_no + 1, block); |
|
496 _num_blocks++; |
|
497 } |
|
498 |
|
499 //------------------------------no_flip_branch--------------------------------- |
|
500 // Does this block end in a multiway branch that cannot have the default case |
|
501 // flipped for another case? |
|
502 static bool no_flip_branch( Block *b ) { |
|
503 int branch_idx = b->_nodes.size() - b->_num_succs-1; |
|
504 if( branch_idx < 1 ) return false; |
|
505 Node *bra = b->_nodes[branch_idx]; |
|
506 if( bra->is_Catch() ) return true; |
|
507 if( bra->is_Mach() ) { |
|
508 if( bra->is_MachNullCheck() ) return true; |
|
509 int iop = bra->as_Mach()->ideal_Opcode(); |
|
510 if( iop == Op_FastLock || iop == Op_FastUnlock ) |
|
511 return true; |
|
512 } |
|
513 return false; |
|
514 } |
|
515 |
|
516 //------------------------------convert_NeverBranch_to_Goto-------------------- |
|
517 // Check for NeverBranch at block end. This needs to become a GOTO to the |
|
518 // true target. NeverBranch are treated as a conditional branch that always |
|
519 // goes the same direction for most of the optimizer and are used to give a |
|
520 // fake exit path to infinite loops. At this late stage they need to turn |
|
521 // into Goto's so that when you enter the infinite loop you indeed hang. |
|
522 void PhaseCFG::convert_NeverBranch_to_Goto(Block *b) { |
|
523 // Find true target |
|
524 int end_idx = b->end_idx(); |
|
525 int idx = b->_nodes[end_idx+1]->as_Proj()->_con; |
|
526 Block *succ = b->_succs[idx]; |
|
527 Node* gto = _goto->clone(); // get a new goto node |
|
528 gto->set_req(0, b->head()); |
|
529 Node *bp = b->_nodes[end_idx]; |
|
530 b->_nodes.map(end_idx,gto); // Slam over NeverBranch |
|
531 _bbs.map(gto->_idx, b); |
|
532 C->regalloc()->set_bad(gto->_idx); |
|
533 b->_nodes.pop(); // Yank projections |
|
534 b->_nodes.pop(); // Yank projections |
|
535 b->_succs.map(0,succ); // Map only successor |
|
536 b->_num_succs = 1; |
|
537 // remap successor's predecessors if necessary |
|
538 uint j; |
|
539 for( j = 1; j < succ->num_preds(); j++) |
|
540 if( succ->pred(j)->in(0) == bp ) |
|
541 succ->head()->set_req(j, gto); |
|
542 // Kill alternate exit path |
|
543 Block *dead = b->_succs[1-idx]; |
|
544 for( j = 1; j < dead->num_preds(); j++) |
|
545 if( dead->pred(j)->in(0) == bp ) |
|
546 break; |
|
547 // Scan through block, yanking dead path from |
|
548 // all regions and phis. |
|
549 dead->head()->del_req(j); |
|
550 for( int k = 1; dead->_nodes[k]->is_Phi(); k++ ) |
|
551 dead->_nodes[k]->del_req(j); |
|
552 } |
|
553 |
|
554 //------------------------------MoveToNext------------------------------------- |
|
555 // Helper function to move block bx to the slot following b_index. Return |
|
556 // true if the move is successful, otherwise false |
|
557 bool PhaseCFG::MoveToNext(Block* bx, uint b_index) { |
|
558 if (bx == NULL) return false; |
|
559 |
|
560 // Return false if bx is already scheduled. |
|
561 uint bx_index = bx->_pre_order; |
|
562 if ((bx_index <= b_index) && (_blocks[bx_index] == bx)) { |
|
563 return false; |
|
564 } |
|
565 |
|
566 // Find the current index of block bx on the block list |
|
567 bx_index = b_index + 1; |
|
568 while( bx_index < _num_blocks && _blocks[bx_index] != bx ) bx_index++; |
|
569 assert(_blocks[bx_index] == bx, "block not found"); |
|
570 |
|
571 // If the previous block conditionally falls into bx, return false, |
|
572 // because moving bx will create an extra jump. |
|
573 for(uint k = 1; k < bx->num_preds(); k++ ) { |
|
574 Block* pred = _bbs[bx->pred(k)->_idx]; |
|
575 if (pred == _blocks[bx_index-1]) { |
|
576 if (pred->_num_succs != 1) { |
|
577 return false; |
|
578 } |
|
579 } |
|
580 } |
|
581 |
|
582 // Reinsert bx just past block 'b' |
|
583 _blocks.remove(bx_index); |
|
584 _blocks.insert(b_index + 1, bx); |
|
585 return true; |
|
586 } |
|
587 |
|
588 //------------------------------MoveToEnd-------------------------------------- |
|
589 // Move empty and uncommon blocks to the end. |
|
590 void PhaseCFG::MoveToEnd(Block *b, uint i) { |
|
591 int e = b->is_Empty(); |
|
592 if (e != Block::not_empty) { |
|
593 if (e == Block::empty_with_goto) { |
|
594 // Remove the goto, but leave the block. |
|
595 b->_nodes.pop(); |
|
596 } |
|
597 // Mark this block as a connector block, which will cause it to be |
|
598 // ignored in certain functions such as non_connector_successor(). |
|
599 b->set_connector(); |
|
600 } |
|
601 // Move the empty block to the end, and don't recheck. |
|
602 _blocks.remove(i); |
|
603 _blocks.push(b); |
|
604 } |
|
605 |
|
606 //------------------------------RemoveEmpty------------------------------------ |
|
607 // Remove empty basic blocks and useless branches. |
|
608 void PhaseCFG::RemoveEmpty() { |
|
609 // Move uncommon blocks to the end |
|
610 uint last = _num_blocks; |
|
611 uint i; |
|
612 assert( _blocks[0] == _broot, "" ); |
|
613 for( i = 1; i < last; i++ ) { |
|
614 Block *b = _blocks[i]; |
|
615 |
|
616 // Check for NeverBranch at block end. This needs to become a GOTO to the |
|
617 // true target. NeverBranch are treated as a conditional branch that |
|
618 // always goes the same direction for most of the optimizer and are used |
|
619 // to give a fake exit path to infinite loops. At this late stage they |
|
620 // need to turn into Goto's so that when you enter the infinite loop you |
|
621 // indeed hang. |
|
622 if( b->_nodes[b->end_idx()]->Opcode() == Op_NeverBranch ) |
|
623 convert_NeverBranch_to_Goto(b); |
|
624 |
|
625 // Look for uncommon blocks and move to end. |
|
626 if( b->is_uncommon(_bbs) ) { |
|
627 MoveToEnd(b, i); |
|
628 last--; // No longer check for being uncommon! |
|
629 if( no_flip_branch(b) ) { // Fall-thru case must follow? |
|
630 b = _blocks[i]; // Find the fall-thru block |
|
631 MoveToEnd(b, i); |
|
632 last--; |
|
633 } |
|
634 i--; // backup block counter post-increment |
|
635 } |
|
636 } |
|
637 |
|
638 // Remove empty blocks |
|
639 uint j1; |
|
640 last = _num_blocks; |
|
641 for( i=0; i < last; i++ ) { |
|
642 Block *b = _blocks[i]; |
|
643 if (i > 0) { |
|
644 if (b->is_Empty() != Block::not_empty) { |
|
645 MoveToEnd(b, i); |
|
646 last--; |
|
647 i--; |
|
648 } |
|
649 } |
|
650 } // End of for all blocks |
|
651 |
|
652 // Fixup final control flow for the blocks. Remove jump-to-next |
|
653 // block. If neither arm of a IF follows the conditional branch, we |
|
654 // have to add a second jump after the conditional. We place the |
|
655 // TRUE branch target in succs[0] for both GOTOs and IFs. |
|
656 for( i=0; i < _num_blocks; i++ ) { |
|
657 Block *b = _blocks[i]; |
|
658 b->_pre_order = i; // turn pre-order into block-index |
|
659 |
|
660 // Connector blocks need no further processing. |
|
661 if (b->is_connector()) { |
|
662 assert((i+1) == _num_blocks || _blocks[i+1]->is_connector(), |
|
663 "All connector blocks should sink to the end"); |
|
664 continue; |
|
665 } |
|
666 assert(b->is_Empty() != Block::completely_empty, |
|
667 "Empty blocks should be connectors"); |
|
668 |
|
669 Block *bnext = (i < _num_blocks-1) ? _blocks[i+1] : NULL; |
|
670 Block *bs0 = b->non_connector_successor(0); |
|
671 |
|
672 // Check for multi-way branches where I cannot negate the test to |
|
673 // exchange the true and false targets. |
|
674 if( no_flip_branch( b ) ) { |
|
675 // Find fall through case - if must fall into its target |
|
676 int branch_idx = b->_nodes.size() - b->_num_succs; |
|
677 for (uint j2 = 0; j2 < b->_num_succs; j2++) { |
|
678 const ProjNode* p = b->_nodes[branch_idx + j2]->as_Proj(); |
|
679 if (p->_con == 0) { |
|
680 // successor j2 is fall through case |
|
681 if (b->non_connector_successor(j2) != bnext) { |
|
682 // but it is not the next block => insert a goto |
|
683 insert_goto_at(i, j2); |
|
684 } |
|
685 // Put taken branch in slot 0 |
|
686 if( j2 == 0 && b->_num_succs == 2) { |
|
687 // Flip targets in succs map |
|
688 Block *tbs0 = b->_succs[0]; |
|
689 Block *tbs1 = b->_succs[1]; |
|
690 b->_succs.map( 0, tbs1 ); |
|
691 b->_succs.map( 1, tbs0 ); |
|
692 } |
|
693 break; |
|
694 } |
|
695 } |
|
696 // Remove all CatchProjs |
|
697 for (j1 = 0; j1 < b->_num_succs; j1++) b->_nodes.pop(); |
|
698 |
|
699 } else if (b->_num_succs == 1) { |
|
700 // Block ends in a Goto? |
|
701 if (bnext == bs0) { |
|
702 // We fall into next block; remove the Goto |
|
703 b->_nodes.pop(); |
|
704 } |
|
705 |
|
706 } else if( b->_num_succs == 2 ) { // Block ends in a If? |
|
707 // Get opcode of 1st projection (matches _succs[0]) |
|
708 // Note: Since this basic block has 2 exits, the last 2 nodes must |
|
709 // be projections (in any order), the 3rd last node must be |
|
710 // the IfNode (we have excluded other 2-way exits such as |
|
711 // CatchNodes already). |
|
712 MachNode *iff = b->_nodes[b->_nodes.size()-3]->as_Mach(); |
|
713 ProjNode *proj0 = b->_nodes[b->_nodes.size()-2]->as_Proj(); |
|
714 ProjNode *proj1 = b->_nodes[b->_nodes.size()-1]->as_Proj(); |
|
715 |
|
716 // Assert that proj0 and succs[0] match up. Similarly for proj1 and succs[1]. |
|
717 assert(proj0->raw_out(0) == b->_succs[0]->head(), "Mismatch successor 0"); |
|
718 assert(proj1->raw_out(0) == b->_succs[1]->head(), "Mismatch successor 1"); |
|
719 |
|
720 Block *bs1 = b->non_connector_successor(1); |
|
721 |
|
722 // Check for neither successor block following the current |
|
723 // block ending in a conditional. If so, move one of the |
|
724 // successors after the current one, provided that the |
|
725 // successor was previously unscheduled, but moveable |
|
726 // (i.e., all paths to it involve a branch). |
|
727 if( bnext != bs0 && bnext != bs1 ) { |
|
728 |
|
729 // Choose the more common successor based on the probability |
|
730 // of the conditional branch. |
|
731 Block *bx = bs0; |
|
732 Block *by = bs1; |
|
733 |
|
734 // _prob is the probability of taking the true path. Make |
|
735 // p the probability of taking successor #1. |
|
736 float p = iff->as_MachIf()->_prob; |
|
737 if( proj0->Opcode() == Op_IfTrue ) { |
|
738 p = 1.0 - p; |
|
739 } |
|
740 |
|
741 // Prefer successor #1 if p > 0.5 |
|
742 if (p > PROB_FAIR) { |
|
743 bx = bs1; |
|
744 by = bs0; |
|
745 } |
|
746 |
|
747 // Attempt the more common successor first |
|
748 if (MoveToNext(bx, i)) { |
|
749 bnext = bx; |
|
750 } else if (MoveToNext(by, i)) { |
|
751 bnext = by; |
|
752 } |
|
753 } |
|
754 |
|
755 // Check for conditional branching the wrong way. Negate |
|
756 // conditional, if needed, so it falls into the following block |
|
757 // and branches to the not-following block. |
|
758 |
|
759 // Check for the next block being in succs[0]. We are going to branch |
|
760 // to succs[0], so we want the fall-thru case as the next block in |
|
761 // succs[1]. |
|
762 if (bnext == bs0) { |
|
763 // Fall-thru case in succs[0], so flip targets in succs map |
|
764 Block *tbs0 = b->_succs[0]; |
|
765 Block *tbs1 = b->_succs[1]; |
|
766 b->_succs.map( 0, tbs1 ); |
|
767 b->_succs.map( 1, tbs0 ); |
|
768 // Flip projection for each target |
|
769 { ProjNode *tmp = proj0; proj0 = proj1; proj1 = tmp; } |
|
770 |
|
771 } else if( bnext == bs1 ) { // Fall-thru is already in succs[1] |
|
772 |
|
773 } else { // Else need a double-branch |
|
774 |
|
775 // The existing conditional branch need not change. |
|
776 // Add a unconditional branch to the false target. |
|
777 // Alas, it must appear in its own block and adding a |
|
778 // block this late in the game is complicated. Sigh. |
|
779 insert_goto_at(i, 1); |
|
780 } |
|
781 |
|
782 // Make sure we TRUE branch to the target |
|
783 if( proj0->Opcode() == Op_IfFalse ) |
|
784 iff->negate(); |
|
785 |
|
786 b->_nodes.pop(); // Remove IfFalse & IfTrue projections |
|
787 b->_nodes.pop(); |
|
788 |
|
789 } else { |
|
790 // Multi-exit block, e.g. a switch statement |
|
791 // But we don't need to do anything here |
|
792 } |
|
793 |
|
794 } // End of for all blocks |
|
795 |
|
796 } |
|
797 |
|
798 |
|
799 //------------------------------dump------------------------------------------- |
|
800 #ifndef PRODUCT |
|
801 void PhaseCFG::_dump_cfg( const Node *end, VectorSet &visited ) const { |
|
802 const Node *x = end->is_block_proj(); |
|
803 assert( x, "not a CFG" ); |
|
804 |
|
805 // Do not visit this block again |
|
806 if( visited.test_set(x->_idx) ) return; |
|
807 |
|
808 // Skip through this block |
|
809 const Node *p = x; |
|
810 do { |
|
811 p = p->in(0); // Move control forward |
|
812 assert( !p->is_block_proj() || p->is_Root(), "not a CFG" ); |
|
813 } while( !p->is_block_start() ); |
|
814 |
|
815 // Recursively visit |
|
816 for( uint i=1; i<p->req(); i++ ) |
|
817 _dump_cfg(p->in(i),visited); |
|
818 |
|
819 // Dump the block |
|
820 _bbs[p->_idx]->dump(&_bbs); |
|
821 } |
|
822 |
|
823 void PhaseCFG::dump( ) const { |
|
824 tty->print("\n--- CFG --- %d BBs\n",_num_blocks); |
|
825 if( _blocks.size() ) { // Did we do basic-block layout? |
|
826 for( uint i=0; i<_num_blocks; i++ ) |
|
827 _blocks[i]->dump(&_bbs); |
|
828 } else { // Else do it with a DFS |
|
829 VectorSet visited(_bbs._arena); |
|
830 _dump_cfg(_root,visited); |
|
831 } |
|
832 } |
|
833 |
|
834 void PhaseCFG::dump_headers() { |
|
835 for( uint i = 0; i < _num_blocks; i++ ) { |
|
836 if( _blocks[i] == NULL ) continue; |
|
837 _blocks[i]->dump_head(&_bbs); |
|
838 } |
|
839 } |
|
840 |
|
841 void PhaseCFG::verify( ) const { |
|
842 // Verify sane CFG |
|
843 for( uint i = 0; i < _num_blocks; i++ ) { |
|
844 Block *b = _blocks[i]; |
|
845 uint cnt = b->_nodes.size(); |
|
846 uint j; |
|
847 for( j = 0; j < cnt; j++ ) { |
|
848 Node *n = b->_nodes[j]; |
|
849 assert( _bbs[n->_idx] == b, "" ); |
|
850 if( j >= 1 && n->is_Mach() && |
|
851 n->as_Mach()->ideal_Opcode() == Op_CreateEx ) { |
|
852 assert( j == 1 || b->_nodes[j-1]->is_Phi(), |
|
853 "CreateEx must be first instruction in block" ); |
|
854 } |
|
855 for( uint k = 0; k < n->req(); k++ ) { |
|
856 Node *use = n->in(k); |
|
857 if( use && use != n ) { |
|
858 assert( _bbs[use->_idx] || use->is_Con(), |
|
859 "must have block; constants for debug info ok" ); |
|
860 } |
|
861 } |
|
862 } |
|
863 |
|
864 j = b->end_idx(); |
|
865 Node *bp = (Node*)b->_nodes[b->_nodes.size()-1]->is_block_proj(); |
|
866 assert( bp, "last instruction must be a block proj" ); |
|
867 assert( bp == b->_nodes[j], "wrong number of successors for this block" ); |
|
868 if( bp->is_Catch() ) { |
|
869 while( b->_nodes[--j]->Opcode() == Op_MachProj ) ; |
|
870 assert( b->_nodes[j]->is_Call(), "CatchProj must follow call" ); |
|
871 } |
|
872 else if( bp->is_Mach() && bp->as_Mach()->ideal_Opcode() == Op_If ) { |
|
873 assert( b->_num_succs == 2, "Conditional branch must have two targets"); |
|
874 } |
|
875 } |
|
876 } |
|
877 #endif |
|
878 |
|
879 //============================================================================= |
|
880 //------------------------------UnionFind-------------------------------------- |
|
881 UnionFind::UnionFind( uint max ) : _cnt(max), _max(max), _indices(NEW_RESOURCE_ARRAY(uint,max)) { |
|
882 Copy::zero_to_bytes( _indices, sizeof(uint)*max ); |
|
883 } |
|
884 |
|
885 void UnionFind::extend( uint from_idx, uint to_idx ) { |
|
886 _nesting.check(); |
|
887 if( from_idx >= _max ) { |
|
888 uint size = 16; |
|
889 while( size <= from_idx ) size <<=1; |
|
890 _indices = REALLOC_RESOURCE_ARRAY( uint, _indices, _max, size ); |
|
891 _max = size; |
|
892 } |
|
893 while( _cnt <= from_idx ) _indices[_cnt++] = 0; |
|
894 _indices[from_idx] = to_idx; |
|
895 } |
|
896 |
|
897 void UnionFind::reset( uint max ) { |
|
898 assert( max <= max_uint, "Must fit within uint" ); |
|
899 // Force the Union-Find mapping to be at least this large |
|
900 extend(max,0); |
|
901 // Initialize to be the ID mapping. |
|
902 for( uint i=0; i<_max; i++ ) map(i,i); |
|
903 } |
|
904 |
|
905 //------------------------------Find_compress---------------------------------- |
|
906 // Straight out of Tarjan's union-find algorithm |
|
907 uint UnionFind::Find_compress( uint idx ) { |
|
908 uint cur = idx; |
|
909 uint next = lookup(cur); |
|
910 while( next != cur ) { // Scan chain of equivalences |
|
911 assert( next < cur, "always union smaller" ); |
|
912 cur = next; // until find a fixed-point |
|
913 next = lookup(cur); |
|
914 } |
|
915 // Core of union-find algorithm: update chain of |
|
916 // equivalences to be equal to the root. |
|
917 while( idx != next ) { |
|
918 uint tmp = lookup(idx); |
|
919 map(idx, next); |
|
920 idx = tmp; |
|
921 } |
|
922 return idx; |
|
923 } |
|
924 |
|
925 //------------------------------Find_const------------------------------------- |
|
926 // Like Find above, but no path compress, so bad asymptotic behavior |
|
927 uint UnionFind::Find_const( uint idx ) const { |
|
928 if( idx == 0 ) return idx; // Ignore the zero idx |
|
929 // Off the end? This can happen during debugging dumps |
|
930 // when data structures have not finished being updated. |
|
931 if( idx >= _max ) return idx; |
|
932 uint next = lookup(idx); |
|
933 while( next != idx ) { // Scan chain of equivalences |
|
934 assert( next < idx, "always union smaller" ); |
|
935 idx = next; // until find a fixed-point |
|
936 next = lookup(idx); |
|
937 } |
|
938 return next; |
|
939 } |
|
940 |
|
941 //------------------------------Union------------------------------------------ |
|
942 // union 2 sets together. |
|
943 void UnionFind::Union( uint idx1, uint idx2 ) { |
|
944 uint src = Find(idx1); |
|
945 uint dst = Find(idx2); |
|
946 assert( src, "" ); |
|
947 assert( dst, "" ); |
|
948 assert( src < _max, "oob" ); |
|
949 assert( dst < _max, "oob" ); |
|
950 assert( src < dst, "always union smaller" ); |
|
951 map(dst,src); |
|
952 } |