Mercurial > hg > truffle
comparison src/share/vm/opto/loopnode.cpp @ 0:a61af66fc99e jdk7-b24
Initial load
author | duke |
---|---|
date | Sat, 01 Dec 2007 00:00:00 +0000 |
parents | |
children | ff5961f4c095 |
comparison
equal
deleted
inserted
replaced
-1:000000000000 | 0:a61af66fc99e |
---|---|
1 /* | |
2 * Copyright 1998-2007 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 #include "incls/_precompiled.incl" | |
26 #include "incls/_loopnode.cpp.incl" | |
27 | |
28 //============================================================================= | |
29 //------------------------------is_loop_iv------------------------------------- | |
30 // Determine if a node is Counted loop induction variable. | |
31 // The method is declared in node.hpp. | |
32 const Node* Node::is_loop_iv() const { | |
33 if (this->is_Phi() && !this->as_Phi()->is_copy() && | |
34 this->as_Phi()->region()->is_CountedLoop() && | |
35 this->as_Phi()->region()->as_CountedLoop()->phi() == this) { | |
36 return this; | |
37 } else { | |
38 return NULL; | |
39 } | |
40 } | |
41 | |
42 //============================================================================= | |
43 //------------------------------dump_spec-------------------------------------- | |
44 // Dump special per-node info | |
45 #ifndef PRODUCT | |
46 void LoopNode::dump_spec(outputStream *st) const { | |
47 if( is_inner_loop () ) st->print( "inner " ); | |
48 if( is_partial_peel_loop () ) st->print( "partial_peel " ); | |
49 if( partial_peel_has_failed () ) st->print( "partial_peel_failed " ); | |
50 } | |
51 #endif | |
52 | |
53 //------------------------------get_early_ctrl--------------------------------- | |
54 // Compute earliest legal control | |
55 Node *PhaseIdealLoop::get_early_ctrl( Node *n ) { | |
56 assert( !n->is_Phi() && !n->is_CFG(), "this code only handles data nodes" ); | |
57 uint i; | |
58 Node *early; | |
59 if( n->in(0) ) { | |
60 early = n->in(0); | |
61 if( !early->is_CFG() ) // Might be a non-CFG multi-def | |
62 early = get_ctrl(early); // So treat input as a straight data input | |
63 i = 1; | |
64 } else { | |
65 early = get_ctrl(n->in(1)); | |
66 i = 2; | |
67 } | |
68 uint e_d = dom_depth(early); | |
69 assert( early, "" ); | |
70 for( ; i < n->req(); i++ ) { | |
71 Node *cin = get_ctrl(n->in(i)); | |
72 assert( cin, "" ); | |
73 // Keep deepest dominator depth | |
74 uint c_d = dom_depth(cin); | |
75 if( c_d > e_d ) { // Deeper guy? | |
76 early = cin; // Keep deepest found so far | |
77 e_d = c_d; | |
78 } else if( c_d == e_d && // Same depth? | |
79 early != cin ) { // If not equal, must use slower algorithm | |
80 // If same depth but not equal, one _must_ dominate the other | |
81 // and we want the deeper (i.e., dominated) guy. | |
82 Node *n1 = early; | |
83 Node *n2 = cin; | |
84 while( 1 ) { | |
85 n1 = idom(n1); // Walk up until break cycle | |
86 n2 = idom(n2); | |
87 if( n1 == cin || // Walked early up to cin | |
88 dom_depth(n2) < c_d ) | |
89 break; // early is deeper; keep him | |
90 if( n2 == early || // Walked cin up to early | |
91 dom_depth(n1) < c_d ) { | |
92 early = cin; // cin is deeper; keep him | |
93 break; | |
94 } | |
95 } | |
96 e_d = dom_depth(early); // Reset depth register cache | |
97 } | |
98 } | |
99 | |
100 // Return earliest legal location | |
101 assert(early == find_non_split_ctrl(early), "unexpected early control"); | |
102 | |
103 return early; | |
104 } | |
105 | |
106 //------------------------------set_early_ctrl--------------------------------- | |
107 // Set earliest legal control | |
108 void PhaseIdealLoop::set_early_ctrl( Node *n ) { | |
109 Node *early = get_early_ctrl(n); | |
110 | |
111 // Record earliest legal location | |
112 set_ctrl(n, early); | |
113 } | |
114 | |
115 //------------------------------set_subtree_ctrl------------------------------- | |
116 // set missing _ctrl entries on new nodes | |
117 void PhaseIdealLoop::set_subtree_ctrl( Node *n ) { | |
118 // Already set? Get out. | |
119 if( _nodes[n->_idx] ) return; | |
120 // Recursively set _nodes array to indicate where the Node goes | |
121 uint i; | |
122 for( i = 0; i < n->req(); ++i ) { | |
123 Node *m = n->in(i); | |
124 if( m && m != C->root() ) | |
125 set_subtree_ctrl( m ); | |
126 } | |
127 | |
128 // Fixup self | |
129 set_early_ctrl( n ); | |
130 } | |
131 | |
132 //------------------------------is_counted_loop-------------------------------- | |
133 Node *PhaseIdealLoop::is_counted_loop( Node *x, IdealLoopTree *loop ) { | |
134 PhaseGVN *gvn = &_igvn; | |
135 | |
136 // Counted loop head must be a good RegionNode with only 3 not NULL | |
137 // control input edges: Self, Entry, LoopBack. | |
138 if ( x->in(LoopNode::Self) == NULL || x->req() != 3 ) | |
139 return NULL; | |
140 | |
141 Node *init_control = x->in(LoopNode::EntryControl); | |
142 Node *back_control = x->in(LoopNode::LoopBackControl); | |
143 if( init_control == NULL || back_control == NULL ) // Partially dead | |
144 return NULL; | |
145 // Must also check for TOP when looking for a dead loop | |
146 if( init_control->is_top() || back_control->is_top() ) | |
147 return NULL; | |
148 | |
149 // Allow funny placement of Safepoint | |
150 if( back_control->Opcode() == Op_SafePoint ) | |
151 back_control = back_control->in(TypeFunc::Control); | |
152 | |
153 // Controlling test for loop | |
154 Node *iftrue = back_control; | |
155 uint iftrue_op = iftrue->Opcode(); | |
156 if( iftrue_op != Op_IfTrue && | |
157 iftrue_op != Op_IfFalse ) | |
158 // I have a weird back-control. Probably the loop-exit test is in | |
159 // the middle of the loop and I am looking at some trailing control-flow | |
160 // merge point. To fix this I would have to partially peel the loop. | |
161 return NULL; // Obscure back-control | |
162 | |
163 // Get boolean guarding loop-back test | |
164 Node *iff = iftrue->in(0); | |
165 if( get_loop(iff) != loop || !iff->in(1)->is_Bool() ) return NULL; | |
166 BoolNode *test = iff->in(1)->as_Bool(); | |
167 BoolTest::mask bt = test->_test._test; | |
168 float cl_prob = iff->as_If()->_prob; | |
169 if( iftrue_op == Op_IfFalse ) { | |
170 bt = BoolTest(bt).negate(); | |
171 cl_prob = 1.0 - cl_prob; | |
172 } | |
173 // Get backedge compare | |
174 Node *cmp = test->in(1); | |
175 int cmp_op = cmp->Opcode(); | |
176 if( cmp_op != Op_CmpI ) | |
177 return NULL; // Avoid pointer & float compares | |
178 | |
179 // Find the trip-counter increment & limit. Limit must be loop invariant. | |
180 Node *incr = cmp->in(1); | |
181 Node *limit = cmp->in(2); | |
182 | |
183 // --------- | |
184 // need 'loop()' test to tell if limit is loop invariant | |
185 // --------- | |
186 | |
187 if( !is_member( loop, get_ctrl(incr) ) ) { // Swapped trip counter and limit? | |
188 Node *tmp = incr; // Then reverse order into the CmpI | |
189 incr = limit; | |
190 limit = tmp; | |
191 bt = BoolTest(bt).commute(); // And commute the exit test | |
192 } | |
193 if( is_member( loop, get_ctrl(limit) ) ) // Limit must loop-invariant | |
194 return NULL; | |
195 | |
196 // Trip-counter increment must be commutative & associative. | |
197 uint incr_op = incr->Opcode(); | |
198 if( incr_op == Op_Phi && incr->req() == 3 ) { | |
199 incr = incr->in(2); // Assume incr is on backedge of Phi | |
200 incr_op = incr->Opcode(); | |
201 } | |
202 Node* trunc1 = NULL; | |
203 Node* trunc2 = NULL; | |
204 const TypeInt* iv_trunc_t = NULL; | |
205 if (!(incr = CountedLoopNode::match_incr_with_optional_truncation(incr, &trunc1, &trunc2, &iv_trunc_t))) { | |
206 return NULL; // Funny increment opcode | |
207 } | |
208 | |
209 // Get merge point | |
210 Node *xphi = incr->in(1); | |
211 Node *stride = incr->in(2); | |
212 if( !stride->is_Con() ) { // Oops, swap these | |
213 if( !xphi->is_Con() ) // Is the other guy a constant? | |
214 return NULL; // Nope, unknown stride, bail out | |
215 Node *tmp = xphi; // 'incr' is commutative, so ok to swap | |
216 xphi = stride; | |
217 stride = tmp; | |
218 } | |
219 //if( loop(xphi) != l) return NULL;// Merge point is in inner loop?? | |
220 if( !xphi->is_Phi() ) return NULL; // Too much math on the trip counter | |
221 PhiNode *phi = xphi->as_Phi(); | |
222 | |
223 // Stride must be constant | |
224 const Type *stride_t = stride->bottom_type(); | |
225 int stride_con = stride_t->is_int()->get_con(); | |
226 assert( stride_con, "missed some peephole opt" ); | |
227 | |
228 // Phi must be of loop header; backedge must wrap to increment | |
229 if( phi->region() != x ) return NULL; | |
230 if( trunc1 == NULL && phi->in(LoopNode::LoopBackControl) != incr || | |
231 trunc1 != NULL && phi->in(LoopNode::LoopBackControl) != trunc1 ) { | |
232 return NULL; | |
233 } | |
234 Node *init_trip = phi->in(LoopNode::EntryControl); | |
235 //if (!init_trip->is_Con()) return NULL; // avoid rolling over MAXINT/MININT | |
236 | |
237 // If iv trunc type is smaller than int, check for possible wrap. | |
238 if (!TypeInt::INT->higher_equal(iv_trunc_t)) { | |
239 assert(trunc1 != NULL, "must have found some truncation"); | |
240 | |
241 // Get a better type for the phi (filtered thru if's) | |
242 const TypeInt* phi_ft = filtered_type(phi); | |
243 | |
244 // Can iv take on a value that will wrap? | |
245 // | |
246 // Ensure iv's limit is not within "stride" of the wrap value. | |
247 // | |
248 // Example for "short" type | |
249 // Truncation ensures value is in the range -32768..32767 (iv_trunc_t) | |
250 // If the stride is +10, then the last value of the induction | |
251 // variable before the increment (phi_ft->_hi) must be | |
252 // <= 32767 - 10 and (phi_ft->_lo) must be >= -32768 to | |
253 // ensure no truncation occurs after the increment. | |
254 | |
255 if (stride_con > 0) { | |
256 if (iv_trunc_t->_hi - phi_ft->_hi < stride_con || | |
257 iv_trunc_t->_lo > phi_ft->_lo) { | |
258 return NULL; // truncation may occur | |
259 } | |
260 } else if (stride_con < 0) { | |
261 if (iv_trunc_t->_lo - phi_ft->_lo > stride_con || | |
262 iv_trunc_t->_hi < phi_ft->_hi) { | |
263 return NULL; // truncation may occur | |
264 } | |
265 } | |
266 // No possibility of wrap so truncation can be discarded | |
267 // Promote iv type to Int | |
268 } else { | |
269 assert(trunc1 == NULL && trunc2 == NULL, "no truncation for int"); | |
270 } | |
271 | |
272 // ================================================= | |
273 // ---- SUCCESS! Found A Trip-Counted Loop! ----- | |
274 // | |
275 // Canonicalize the condition on the test. If we can exactly determine | |
276 // the trip-counter exit value, then set limit to that value and use | |
277 // a '!=' test. Otherwise use conditon '<' for count-up loops and | |
278 // '>' for count-down loops. If the condition is inverted and we will | |
279 // be rolling through MININT to MAXINT, then bail out. | |
280 | |
281 C->print_method("Before CountedLoop", 3); | |
282 | |
283 // Check for SafePoint on backedge and remove | |
284 Node *sfpt = x->in(LoopNode::LoopBackControl); | |
285 if( sfpt->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt)) { | |
286 lazy_replace( sfpt, iftrue ); | |
287 loop->_tail = iftrue; | |
288 } | |
289 | |
290 | |
291 // If compare points to incr, we are ok. Otherwise the compare | |
292 // can directly point to the phi; in this case adjust the compare so that | |
293 // it points to the incr by adusting the limit. | |
294 if( cmp->in(1) == phi || cmp->in(2) == phi ) | |
295 limit = gvn->transform(new (C, 3) AddINode(limit,stride)); | |
296 | |
297 // trip-count for +-tive stride should be: (limit - init_trip + stride - 1)/stride. | |
298 // Final value for iterator should be: trip_count * stride + init_trip. | |
299 const Type *limit_t = limit->bottom_type(); | |
300 const Type *init_t = init_trip->bottom_type(); | |
301 Node *one_p = gvn->intcon( 1); | |
302 Node *one_m = gvn->intcon(-1); | |
303 | |
304 Node *trip_count = NULL; | |
305 Node *hook = new (C, 6) Node(6); | |
306 switch( bt ) { | |
307 case BoolTest::eq: | |
308 return NULL; // Bail out, but this loop trips at most twice! | |
309 case BoolTest::ne: // Ahh, the case we desire | |
310 if( stride_con == 1 ) | |
311 trip_count = gvn->transform(new (C, 3) SubINode(limit,init_trip)); | |
312 else if( stride_con == -1 ) | |
313 trip_count = gvn->transform(new (C, 3) SubINode(init_trip,limit)); | |
314 else | |
315 return NULL; // Odd stride; must prove we hit limit exactly | |
316 set_subtree_ctrl( trip_count ); | |
317 //_loop.map(trip_count->_idx,loop(limit)); | |
318 break; | |
319 case BoolTest::le: // Maybe convert to '<' case | |
320 limit = gvn->transform(new (C, 3) AddINode(limit,one_p)); | |
321 set_subtree_ctrl( limit ); | |
322 hook->init_req(4, limit); | |
323 | |
324 bt = BoolTest::lt; | |
325 // Make the new limit be in the same loop nest as the old limit | |
326 //_loop.map(limit->_idx,limit_loop); | |
327 // Fall into next case | |
328 case BoolTest::lt: { // Maybe convert to '!=' case | |
329 if( stride_con < 0 ) return NULL; // Count down loop rolls through MAXINT | |
330 Node *range = gvn->transform(new (C, 3) SubINode(limit,init_trip)); | |
331 set_subtree_ctrl( range ); | |
332 hook->init_req(0, range); | |
333 | |
334 Node *bias = gvn->transform(new (C, 3) AddINode(range,stride)); | |
335 set_subtree_ctrl( bias ); | |
336 hook->init_req(1, bias); | |
337 | |
338 Node *bias1 = gvn->transform(new (C, 3) AddINode(bias,one_m)); | |
339 set_subtree_ctrl( bias1 ); | |
340 hook->init_req(2, bias1); | |
341 | |
342 trip_count = gvn->transform(new (C, 3) DivINode(0,bias1,stride)); | |
343 set_subtree_ctrl( trip_count ); | |
344 hook->init_req(3, trip_count); | |
345 break; | |
346 } | |
347 | |
348 case BoolTest::ge: // Maybe convert to '>' case | |
349 limit = gvn->transform(new (C, 3) AddINode(limit,one_m)); | |
350 set_subtree_ctrl( limit ); | |
351 hook->init_req(4 ,limit); | |
352 | |
353 bt = BoolTest::gt; | |
354 // Make the new limit be in the same loop nest as the old limit | |
355 //_loop.map(limit->_idx,limit_loop); | |
356 // Fall into next case | |
357 case BoolTest::gt: { // Maybe convert to '!=' case | |
358 if( stride_con > 0 ) return NULL; // count up loop rolls through MININT | |
359 Node *range = gvn->transform(new (C, 3) SubINode(limit,init_trip)); | |
360 set_subtree_ctrl( range ); | |
361 hook->init_req(0, range); | |
362 | |
363 Node *bias = gvn->transform(new (C, 3) AddINode(range,stride)); | |
364 set_subtree_ctrl( bias ); | |
365 hook->init_req(1, bias); | |
366 | |
367 Node *bias1 = gvn->transform(new (C, 3) AddINode(bias,one_p)); | |
368 set_subtree_ctrl( bias1 ); | |
369 hook->init_req(2, bias1); | |
370 | |
371 trip_count = gvn->transform(new (C, 3) DivINode(0,bias1,stride)); | |
372 set_subtree_ctrl( trip_count ); | |
373 hook->init_req(3, trip_count); | |
374 break; | |
375 } | |
376 } | |
377 | |
378 Node *span = gvn->transform(new (C, 3) MulINode(trip_count,stride)); | |
379 set_subtree_ctrl( span ); | |
380 hook->init_req(5, span); | |
381 | |
382 limit = gvn->transform(new (C, 3) AddINode(span,init_trip)); | |
383 set_subtree_ctrl( limit ); | |
384 | |
385 // Build a canonical trip test. | |
386 // Clone code, as old values may be in use. | |
387 incr = incr->clone(); | |
388 incr->set_req(1,phi); | |
389 incr->set_req(2,stride); | |
390 incr = _igvn.register_new_node_with_optimizer(incr); | |
391 set_early_ctrl( incr ); | |
392 _igvn.hash_delete(phi); | |
393 phi->set_req_X( LoopNode::LoopBackControl, incr, &_igvn ); | |
394 | |
395 // If phi type is more restrictive than Int, raise to | |
396 // Int to prevent (almost) infinite recursion in igvn | |
397 // which can only handle integer types for constants or minint..maxint. | |
398 if (!TypeInt::INT->higher_equal(phi->bottom_type())) { | |
399 Node* nphi = PhiNode::make(phi->in(0), phi->in(LoopNode::EntryControl), TypeInt::INT); | |
400 nphi->set_req(LoopNode::LoopBackControl, phi->in(LoopNode::LoopBackControl)); | |
401 nphi = _igvn.register_new_node_with_optimizer(nphi); | |
402 set_ctrl(nphi, get_ctrl(phi)); | |
403 _igvn.subsume_node(phi, nphi); | |
404 phi = nphi->as_Phi(); | |
405 } | |
406 cmp = cmp->clone(); | |
407 cmp->set_req(1,incr); | |
408 cmp->set_req(2,limit); | |
409 cmp = _igvn.register_new_node_with_optimizer(cmp); | |
410 set_ctrl(cmp, iff->in(0)); | |
411 | |
412 Node *tmp = test->clone(); | |
413 assert( tmp->is_Bool(), "" ); | |
414 test = (BoolNode*)tmp; | |
415 (*(BoolTest*)&test->_test)._test = bt; //BoolTest::ne; | |
416 test->set_req(1,cmp); | |
417 _igvn.register_new_node_with_optimizer(test); | |
418 set_ctrl(test, iff->in(0)); | |
419 // If the exit test is dead, STOP! | |
420 if( test == NULL ) return NULL; | |
421 _igvn.hash_delete(iff); | |
422 iff->set_req_X( 1, test, &_igvn ); | |
423 | |
424 // Replace the old IfNode with a new LoopEndNode | |
425 Node *lex = _igvn.register_new_node_with_optimizer(new (C, 2) CountedLoopEndNode( iff->in(0), iff->in(1), cl_prob, iff->as_If()->_fcnt )); | |
426 IfNode *le = lex->as_If(); | |
427 uint dd = dom_depth(iff); | |
428 set_idom(le, le->in(0), dd); // Update dominance for loop exit | |
429 set_loop(le, loop); | |
430 | |
431 // Get the loop-exit control | |
432 Node *if_f = iff->as_If()->proj_out(!(iftrue_op == Op_IfTrue)); | |
433 | |
434 // Need to swap loop-exit and loop-back control? | |
435 if( iftrue_op == Op_IfFalse ) { | |
436 Node *ift2=_igvn.register_new_node_with_optimizer(new (C, 1) IfTrueNode (le)); | |
437 Node *iff2=_igvn.register_new_node_with_optimizer(new (C, 1) IfFalseNode(le)); | |
438 | |
439 loop->_tail = back_control = ift2; | |
440 set_loop(ift2, loop); | |
441 set_loop(iff2, get_loop(if_f)); | |
442 | |
443 // Lazy update of 'get_ctrl' mechanism. | |
444 lazy_replace_proj( if_f , iff2 ); | |
445 lazy_replace_proj( iftrue, ift2 ); | |
446 | |
447 // Swap names | |
448 if_f = iff2; | |
449 iftrue = ift2; | |
450 } else { | |
451 _igvn.hash_delete(if_f ); | |
452 _igvn.hash_delete(iftrue); | |
453 if_f ->set_req_X( 0, le, &_igvn ); | |
454 iftrue->set_req_X( 0, le, &_igvn ); | |
455 } | |
456 | |
457 set_idom(iftrue, le, dd+1); | |
458 set_idom(if_f, le, dd+1); | |
459 | |
460 // Now setup a new CountedLoopNode to replace the existing LoopNode | |
461 CountedLoopNode *l = new (C, 3) CountedLoopNode(init_control, back_control); | |
462 // The following assert is approximately true, and defines the intention | |
463 // of can_be_counted_loop. It fails, however, because phase->type | |
464 // is not yet initialized for this loop and its parts. | |
465 //assert(l->can_be_counted_loop(this), "sanity"); | |
466 _igvn.register_new_node_with_optimizer(l); | |
467 set_loop(l, loop); | |
468 loop->_head = l; | |
469 // Fix all data nodes placed at the old loop head. | |
470 // Uses the lazy-update mechanism of 'get_ctrl'. | |
471 lazy_replace( x, l ); | |
472 set_idom(l, init_control, dom_depth(x)); | |
473 | |
474 // Check for immediately preceeding SafePoint and remove | |
475 Node *sfpt2 = le->in(0); | |
476 if( sfpt2->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt2)) | |
477 lazy_replace( sfpt2, sfpt2->in(TypeFunc::Control)); | |
478 | |
479 // Free up intermediate goo | |
480 _igvn.remove_dead_node(hook); | |
481 | |
482 C->print_method("After CountedLoop", 3); | |
483 | |
484 // Return trip counter | |
485 return trip_count; | |
486 } | |
487 | |
488 | |
489 //------------------------------Ideal------------------------------------------ | |
490 // Return a node which is more "ideal" than the current node. | |
491 // Attempt to convert into a counted-loop. | |
492 Node *LoopNode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
493 if (!can_be_counted_loop(phase)) { | |
494 phase->C->set_major_progress(); | |
495 } | |
496 return RegionNode::Ideal(phase, can_reshape); | |
497 } | |
498 | |
499 | |
500 //============================================================================= | |
501 //------------------------------Ideal------------------------------------------ | |
502 // Return a node which is more "ideal" than the current node. | |
503 // Attempt to convert into a counted-loop. | |
504 Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
505 return RegionNode::Ideal(phase, can_reshape); | |
506 } | |
507 | |
508 //------------------------------dump_spec-------------------------------------- | |
509 // Dump special per-node info | |
510 #ifndef PRODUCT | |
511 void CountedLoopNode::dump_spec(outputStream *st) const { | |
512 LoopNode::dump_spec(st); | |
513 if( stride_is_con() ) { | |
514 st->print("stride: %d ",stride_con()); | |
515 } else { | |
516 st->print("stride: not constant "); | |
517 } | |
518 if( is_pre_loop () ) st->print("pre of N%d" , _main_idx ); | |
519 if( is_main_loop() ) st->print("main of N%d", _idx ); | |
520 if( is_post_loop() ) st->print("post of N%d", _main_idx ); | |
521 } | |
522 #endif | |
523 | |
524 //============================================================================= | |
525 int CountedLoopEndNode::stride_con() const { | |
526 return stride()->bottom_type()->is_int()->get_con(); | |
527 } | |
528 | |
529 | |
530 //----------------------match_incr_with_optional_truncation-------------------- | |
531 // Match increment with optional truncation: | |
532 // CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16 | |
533 // Return NULL for failure. Success returns the increment node. | |
534 Node* CountedLoopNode::match_incr_with_optional_truncation( | |
535 Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type) { | |
536 // Quick cutouts: | |
537 if (expr == NULL || expr->req() != 3) return false; | |
538 | |
539 Node *t1 = NULL; | |
540 Node *t2 = NULL; | |
541 const TypeInt* trunc_t = TypeInt::INT; | |
542 Node* n1 = expr; | |
543 int n1op = n1->Opcode(); | |
544 | |
545 // Try to strip (n1 & M) or (n1 << N >> N) from n1. | |
546 if (n1op == Op_AndI && | |
547 n1->in(2)->is_Con() && | |
548 n1->in(2)->bottom_type()->is_int()->get_con() == 0x7fff) { | |
549 // %%% This check should match any mask of 2**K-1. | |
550 t1 = n1; | |
551 n1 = t1->in(1); | |
552 n1op = n1->Opcode(); | |
553 trunc_t = TypeInt::CHAR; | |
554 } else if (n1op == Op_RShiftI && | |
555 n1->in(1) != NULL && | |
556 n1->in(1)->Opcode() == Op_LShiftI && | |
557 n1->in(2) == n1->in(1)->in(2) && | |
558 n1->in(2)->is_Con()) { | |
559 jint shift = n1->in(2)->bottom_type()->is_int()->get_con(); | |
560 // %%% This check should match any shift in [1..31]. | |
561 if (shift == 16 || shift == 8) { | |
562 t1 = n1; | |
563 t2 = t1->in(1); | |
564 n1 = t2->in(1); | |
565 n1op = n1->Opcode(); | |
566 if (shift == 16) { | |
567 trunc_t = TypeInt::SHORT; | |
568 } else if (shift == 8) { | |
569 trunc_t = TypeInt::BYTE; | |
570 } | |
571 } | |
572 } | |
573 | |
574 // If (maybe after stripping) it is an AddI, we won: | |
575 if (n1op == Op_AddI) { | |
576 *trunc1 = t1; | |
577 *trunc2 = t2; | |
578 *trunc_type = trunc_t; | |
579 return n1; | |
580 } | |
581 | |
582 // failed | |
583 return NULL; | |
584 } | |
585 | |
586 | |
587 //------------------------------filtered_type-------------------------------- | |
588 // Return a type based on condition control flow | |
589 // A successful return will be a type that is restricted due | |
590 // to a series of dominating if-tests, such as: | |
591 // if (i < 10) { | |
592 // if (i > 0) { | |
593 // here: "i" type is [1..10) | |
594 // } | |
595 // } | |
596 // or a control flow merge | |
597 // if (i < 10) { | |
598 // do { | |
599 // phi( , ) -- at top of loop type is [min_int..10) | |
600 // i = ? | |
601 // } while ( i < 10) | |
602 // | |
603 const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) { | |
604 assert(n && n->bottom_type()->is_int(), "must be int"); | |
605 const TypeInt* filtered_t = NULL; | |
606 if (!n->is_Phi()) { | |
607 assert(n_ctrl != NULL || n_ctrl == C->top(), "valid control"); | |
608 filtered_t = filtered_type_from_dominators(n, n_ctrl); | |
609 | |
610 } else { | |
611 Node* phi = n->as_Phi(); | |
612 Node* region = phi->in(0); | |
613 assert(n_ctrl == NULL || n_ctrl == region, "ctrl parameter must be region"); | |
614 if (region && region != C->top()) { | |
615 for (uint i = 1; i < phi->req(); i++) { | |
616 Node* val = phi->in(i); | |
617 Node* use_c = region->in(i); | |
618 const TypeInt* val_t = filtered_type_from_dominators(val, use_c); | |
619 if (val_t != NULL) { | |
620 if (filtered_t == NULL) { | |
621 filtered_t = val_t; | |
622 } else { | |
623 filtered_t = filtered_t->meet(val_t)->is_int(); | |
624 } | |
625 } | |
626 } | |
627 } | |
628 } | |
629 const TypeInt* n_t = _igvn.type(n)->is_int(); | |
630 if (filtered_t != NULL) { | |
631 n_t = n_t->join(filtered_t)->is_int(); | |
632 } | |
633 return n_t; | |
634 } | |
635 | |
636 | |
637 //------------------------------filtered_type_from_dominators-------------------------------- | |
638 // Return a possibly more restrictive type for val based on condition control flow of dominators | |
639 const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) { | |
640 if (val->is_Con()) { | |
641 return val->bottom_type()->is_int(); | |
642 } | |
643 uint if_limit = 10; // Max number of dominating if's visited | |
644 const TypeInt* rtn_t = NULL; | |
645 | |
646 if (use_ctrl && use_ctrl != C->top()) { | |
647 Node* val_ctrl = get_ctrl(val); | |
648 uint val_dom_depth = dom_depth(val_ctrl); | |
649 Node* pred = use_ctrl; | |
650 uint if_cnt = 0; | |
651 while (if_cnt < if_limit) { | |
652 if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) { | |
653 if_cnt++; | |
654 const TypeInt* if_t = filtered_type_at_if(val, pred); | |
655 if (if_t != NULL) { | |
656 if (rtn_t == NULL) { | |
657 rtn_t = if_t; | |
658 } else { | |
659 rtn_t = rtn_t->join(if_t)->is_int(); | |
660 } | |
661 } | |
662 } | |
663 pred = idom(pred); | |
664 if (pred == NULL || pred == C->top()) { | |
665 break; | |
666 } | |
667 // Stop if going beyond definition block of val | |
668 if (dom_depth(pred) < val_dom_depth) { | |
669 break; | |
670 } | |
671 } | |
672 } | |
673 return rtn_t; | |
674 } | |
675 | |
676 | |
677 //------------------------------filtered_type_at_if-------------------------------- | |
678 // Return a possibly more restrictive type for val based on condition control flow for an if | |
679 const TypeInt* PhaseIdealLoop::filtered_type_at_if( Node* val, Node *if_proj) { | |
680 assert(if_proj && | |
681 (if_proj->Opcode() == Op_IfTrue || if_proj->Opcode() == Op_IfFalse), "expecting an if projection"); | |
682 if (if_proj->in(0) && if_proj->in(0)->is_If()) { | |
683 IfNode* iff = if_proj->in(0)->as_If(); | |
684 if (iff->in(1) && iff->in(1)->is_Bool()) { | |
685 BoolNode* bol = iff->in(1)->as_Bool(); | |
686 if (bol->in(1) && bol->in(1)->is_Cmp()) { | |
687 const CmpNode* cmp = bol->in(1)->as_Cmp(); | |
688 if (cmp->in(1) == val) { | |
689 const TypeInt* cmp2_t = _igvn.type(cmp->in(2))->isa_int(); | |
690 if (cmp2_t != NULL) { | |
691 jint lo = cmp2_t->_lo; | |
692 jint hi = cmp2_t->_hi; | |
693 BoolTest::mask msk = if_proj->Opcode() == Op_IfTrue ? bol->_test._test : bol->_test.negate(); | |
694 switch (msk) { | |
695 case BoolTest::ne: | |
696 // Can't refine type | |
697 return NULL; | |
698 case BoolTest::eq: | |
699 return cmp2_t; | |
700 case BoolTest::lt: | |
701 lo = TypeInt::INT->_lo; | |
702 if (hi - 1 < hi) { | |
703 hi = hi - 1; | |
704 } | |
705 break; | |
706 case BoolTest::le: | |
707 lo = TypeInt::INT->_lo; | |
708 break; | |
709 case BoolTest::gt: | |
710 if (lo + 1 > lo) { | |
711 lo = lo + 1; | |
712 } | |
713 hi = TypeInt::INT->_hi; | |
714 break; | |
715 case BoolTest::ge: | |
716 // lo unchanged | |
717 hi = TypeInt::INT->_hi; | |
718 break; | |
719 } | |
720 const TypeInt* rtn_t = TypeInt::make(lo, hi, cmp2_t->_widen); | |
721 return rtn_t; | |
722 } | |
723 } | |
724 } | |
725 } | |
726 } | |
727 return NULL; | |
728 } | |
729 | |
730 //------------------------------dump_spec-------------------------------------- | |
731 // Dump special per-node info | |
732 #ifndef PRODUCT | |
733 void CountedLoopEndNode::dump_spec(outputStream *st) const { | |
734 if( in(TestValue)->is_Bool() ) { | |
735 BoolTest bt( test_trip()); // Added this for g++. | |
736 | |
737 st->print("["); | |
738 bt.dump_on(st); | |
739 st->print("]"); | |
740 } | |
741 st->print(" "); | |
742 IfNode::dump_spec(st); | |
743 } | |
744 #endif | |
745 | |
746 //============================================================================= | |
747 //------------------------------is_member-------------------------------------- | |
748 // Is 'l' a member of 'this'? | |
749 int IdealLoopTree::is_member( const IdealLoopTree *l ) const { | |
750 while( l->_nest > _nest ) l = l->_parent; | |
751 return l == this; | |
752 } | |
753 | |
754 //------------------------------set_nest--------------------------------------- | |
755 // Set loop tree nesting depth. Accumulate _has_call bits. | |
756 int IdealLoopTree::set_nest( uint depth ) { | |
757 _nest = depth; | |
758 int bits = _has_call; | |
759 if( _child ) bits |= _child->set_nest(depth+1); | |
760 if( bits ) _has_call = 1; | |
761 if( _next ) bits |= _next ->set_nest(depth ); | |
762 return bits; | |
763 } | |
764 | |
765 //------------------------------split_fall_in---------------------------------- | |
766 // Split out multiple fall-in edges from the loop header. Move them to a | |
767 // private RegionNode before the loop. This becomes the loop landing pad. | |
768 void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) { | |
769 PhaseIterGVN &igvn = phase->_igvn; | |
770 uint i; | |
771 | |
772 // Make a new RegionNode to be the landing pad. | |
773 Node *landing_pad = new (phase->C, fall_in_cnt+1) RegionNode( fall_in_cnt+1 ); | |
774 phase->set_loop(landing_pad,_parent); | |
775 // Gather all the fall-in control paths into the landing pad | |
776 uint icnt = fall_in_cnt; | |
777 uint oreq = _head->req(); | |
778 for( i = oreq-1; i>0; i-- ) | |
779 if( !phase->is_member( this, _head->in(i) ) ) | |
780 landing_pad->set_req(icnt--,_head->in(i)); | |
781 | |
782 // Peel off PhiNode edges as well | |
783 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) { | |
784 Node *oj = _head->fast_out(j); | |
785 if( oj->is_Phi() ) { | |
786 PhiNode* old_phi = oj->as_Phi(); | |
787 assert( old_phi->region() == _head, "" ); | |
788 igvn.hash_delete(old_phi); // Yank from hash before hacking edges | |
789 Node *p = PhiNode::make_blank(landing_pad, old_phi); | |
790 uint icnt = fall_in_cnt; | |
791 for( i = oreq-1; i>0; i-- ) { | |
792 if( !phase->is_member( this, _head->in(i) ) ) { | |
793 p->init_req(icnt--, old_phi->in(i)); | |
794 // Go ahead and clean out old edges from old phi | |
795 old_phi->del_req(i); | |
796 } | |
797 } | |
798 // Search for CSE's here, because ZKM.jar does a lot of | |
799 // loop hackery and we need to be a little incremental | |
800 // with the CSE to avoid O(N^2) node blow-up. | |
801 Node *p2 = igvn.hash_find_insert(p); // Look for a CSE | |
802 if( p2 ) { // Found CSE | |
803 p->destruct(); // Recover useless new node | |
804 p = p2; // Use old node | |
805 } else { | |
806 igvn.register_new_node_with_optimizer(p, old_phi); | |
807 } | |
808 // Make old Phi refer to new Phi. | |
809 old_phi->add_req(p); | |
810 // Check for the special case of making the old phi useless and | |
811 // disappear it. In JavaGrande I have a case where this useless | |
812 // Phi is the loop limit and prevents recognizing a CountedLoop | |
813 // which in turn prevents removing an empty loop. | |
814 Node *id_old_phi = old_phi->Identity( &igvn ); | |
815 if( id_old_phi != old_phi ) { // Found a simple identity? | |
816 // Note that I cannot call 'subsume_node' here, because | |
817 // that will yank the edge from old_phi to the Region and | |
818 // I'm mid-iteration over the Region's uses. | |
819 for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) { | |
820 Node* use = old_phi->last_out(i); | |
821 igvn.hash_delete(use); | |
822 igvn._worklist.push(use); | |
823 uint uses_found = 0; | |
824 for (uint j = 0; j < use->len(); j++) { | |
825 if (use->in(j) == old_phi) { | |
826 if (j < use->req()) use->set_req (j, id_old_phi); | |
827 else use->set_prec(j, id_old_phi); | |
828 uses_found++; | |
829 } | |
830 } | |
831 i -= uses_found; // we deleted 1 or more copies of this edge | |
832 } | |
833 } | |
834 igvn._worklist.push(old_phi); | |
835 } | |
836 } | |
837 // Finally clean out the fall-in edges from the RegionNode | |
838 for( i = oreq-1; i>0; i-- ) { | |
839 if( !phase->is_member( this, _head->in(i) ) ) { | |
840 _head->del_req(i); | |
841 } | |
842 } | |
843 // Transform landing pad | |
844 igvn.register_new_node_with_optimizer(landing_pad, _head); | |
845 // Insert landing pad into the header | |
846 _head->add_req(landing_pad); | |
847 } | |
848 | |
849 //------------------------------split_outer_loop------------------------------- | |
850 // Split out the outermost loop from this shared header. | |
851 void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) { | |
852 PhaseIterGVN &igvn = phase->_igvn; | |
853 | |
854 // Find index of outermost loop; it should also be my tail. | |
855 uint outer_idx = 1; | |
856 while( _head->in(outer_idx) != _tail ) outer_idx++; | |
857 | |
858 // Make a LoopNode for the outermost loop. | |
859 Node *ctl = _head->in(LoopNode::EntryControl); | |
860 Node *outer = new (phase->C, 3) LoopNode( ctl, _head->in(outer_idx) ); | |
861 outer = igvn.register_new_node_with_optimizer(outer, _head); | |
862 phase->set_created_loop_node(); | |
863 // Outermost loop falls into '_head' loop | |
864 _head->set_req(LoopNode::EntryControl, outer); | |
865 _head->del_req(outer_idx); | |
866 // Split all the Phis up between '_head' loop and 'outer' loop. | |
867 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) { | |
868 Node *out = _head->fast_out(j); | |
869 if( out->is_Phi() ) { | |
870 PhiNode *old_phi = out->as_Phi(); | |
871 assert( old_phi->region() == _head, "" ); | |
872 Node *phi = PhiNode::make_blank(outer, old_phi); | |
873 phi->init_req(LoopNode::EntryControl, old_phi->in(LoopNode::EntryControl)); | |
874 phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx)); | |
875 phi = igvn.register_new_node_with_optimizer(phi, old_phi); | |
876 // Make old Phi point to new Phi on the fall-in path | |
877 igvn.hash_delete(old_phi); | |
878 old_phi->set_req(LoopNode::EntryControl, phi); | |
879 old_phi->del_req(outer_idx); | |
880 igvn._worklist.push(old_phi); | |
881 } | |
882 } | |
883 | |
884 // Use the new loop head instead of the old shared one | |
885 _head = outer; | |
886 phase->set_loop(_head, this); | |
887 } | |
888 | |
889 //------------------------------fix_parent------------------------------------- | |
890 static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) { | |
891 loop->_parent = parent; | |
892 if( loop->_child ) fix_parent( loop->_child, loop ); | |
893 if( loop->_next ) fix_parent( loop->_next , parent ); | |
894 } | |
895 | |
896 //------------------------------estimate_path_freq----------------------------- | |
897 static float estimate_path_freq( Node *n ) { | |
898 // Try to extract some path frequency info | |
899 IfNode *iff; | |
900 for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests | |
901 uint nop = n->Opcode(); | |
902 if( nop == Op_SafePoint ) { // Skip any safepoint | |
903 n = n->in(0); | |
904 continue; | |
905 } | |
906 if( nop == Op_CatchProj ) { // Get count from a prior call | |
907 // Assume call does not always throw exceptions: means the call-site | |
908 // count is also the frequency of the fall-through path. | |
909 assert( n->is_CatchProj(), "" ); | |
910 if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index ) | |
911 return 0.0f; // Assume call exception path is rare | |
912 Node *call = n->in(0)->in(0)->in(0); | |
913 assert( call->is_Call(), "expect a call here" ); | |
914 const JVMState *jvms = ((CallNode*)call)->jvms(); | |
915 ciMethodData* methodData = jvms->method()->method_data(); | |
916 if (!methodData->is_mature()) return 0.0f; // No call-site data | |
917 ciProfileData* data = methodData->bci_to_data(jvms->bci()); | |
918 if ((data == NULL) || !data->is_CounterData()) { | |
919 // no call profile available, try call's control input | |
920 n = n->in(0); | |
921 continue; | |
922 } | |
923 return data->as_CounterData()->count()/FreqCountInvocations; | |
924 } | |
925 // See if there's a gating IF test | |
926 Node *n_c = n->in(0); | |
927 if( !n_c->is_If() ) break; // No estimate available | |
928 iff = n_c->as_If(); | |
929 if( iff->_fcnt != COUNT_UNKNOWN ) // Have a valid count? | |
930 // Compute how much count comes on this path | |
931 return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt; | |
932 // Have no count info. Skip dull uncommon-trap like branches. | |
933 if( (nop == Op_IfTrue && iff->_prob < PROB_LIKELY_MAG(5)) || | |
934 (nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) ) | |
935 break; | |
936 // Skip through never-taken branch; look for a real loop exit. | |
937 n = iff->in(0); | |
938 } | |
939 return 0.0f; // No estimate available | |
940 } | |
941 | |
942 //------------------------------merge_many_backedges--------------------------- | |
943 // Merge all the backedges from the shared header into a private Region. | |
944 // Feed that region as the one backedge to this loop. | |
945 void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) { | |
946 uint i; | |
947 | |
948 // Scan for the top 2 hottest backedges | |
949 float hotcnt = 0.0f; | |
950 float warmcnt = 0.0f; | |
951 uint hot_idx = 0; | |
952 // Loop starts at 2 because slot 1 is the fall-in path | |
953 for( i = 2; i < _head->req(); i++ ) { | |
954 float cnt = estimate_path_freq(_head->in(i)); | |
955 if( cnt > hotcnt ) { // Grab hottest path | |
956 warmcnt = hotcnt; | |
957 hotcnt = cnt; | |
958 hot_idx = i; | |
959 } else if( cnt > warmcnt ) { // And 2nd hottest path | |
960 warmcnt = cnt; | |
961 } | |
962 } | |
963 | |
964 // See if the hottest backedge is worthy of being an inner loop | |
965 // by being much hotter than the next hottest backedge. | |
966 if( hotcnt <= 0.0001 || | |
967 hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge | |
968 | |
969 // Peel out the backedges into a private merge point; peel | |
970 // them all except optionally hot_idx. | |
971 PhaseIterGVN &igvn = phase->_igvn; | |
972 | |
973 Node *hot_tail = NULL; | |
974 // Make a Region for the merge point | |
975 Node *r = new (phase->C, 1) RegionNode(1); | |
976 for( i = 2; i < _head->req(); i++ ) { | |
977 if( i != hot_idx ) | |
978 r->add_req( _head->in(i) ); | |
979 else hot_tail = _head->in(i); | |
980 } | |
981 igvn.register_new_node_with_optimizer(r, _head); | |
982 // Plug region into end of loop _head, followed by hot_tail | |
983 while( _head->req() > 3 ) _head->del_req( _head->req()-1 ); | |
984 _head->set_req(2, r); | |
985 if( hot_idx ) _head->add_req(hot_tail); | |
986 | |
987 // Split all the Phis up between '_head' loop and the Region 'r' | |
988 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) { | |
989 Node *out = _head->fast_out(j); | |
990 if( out->is_Phi() ) { | |
991 PhiNode* n = out->as_Phi(); | |
992 igvn.hash_delete(n); // Delete from hash before hacking edges | |
993 Node *hot_phi = NULL; | |
994 Node *phi = new (phase->C, r->req()) PhiNode(r, n->type(), n->adr_type()); | |
995 // Check all inputs for the ones to peel out | |
996 uint j = 1; | |
997 for( uint i = 2; i < n->req(); i++ ) { | |
998 if( i != hot_idx ) | |
999 phi->set_req( j++, n->in(i) ); | |
1000 else hot_phi = n->in(i); | |
1001 } | |
1002 // Register the phi but do not transform until whole place transforms | |
1003 igvn.register_new_node_with_optimizer(phi, n); | |
1004 // Add the merge phi to the old Phi | |
1005 while( n->req() > 3 ) n->del_req( n->req()-1 ); | |
1006 n->set_req(2, phi); | |
1007 if( hot_idx ) n->add_req(hot_phi); | |
1008 } | |
1009 } | |
1010 | |
1011 | |
1012 // Insert a new IdealLoopTree inserted below me. Turn it into a clone | |
1013 // of self loop tree. Turn self into a loop headed by _head and with | |
1014 // tail being the new merge point. | |
1015 IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail ); | |
1016 phase->set_loop(_tail,ilt); // Adjust tail | |
1017 _tail = r; // Self's tail is new merge point | |
1018 phase->set_loop(r,this); | |
1019 ilt->_child = _child; // New guy has my children | |
1020 _child = ilt; // Self has new guy as only child | |
1021 ilt->_parent = this; // new guy has self for parent | |
1022 ilt->_nest = _nest; // Same nesting depth (for now) | |
1023 | |
1024 // Starting with 'ilt', look for child loop trees using the same shared | |
1025 // header. Flatten these out; they will no longer be loops in the end. | |
1026 IdealLoopTree **pilt = &_child; | |
1027 while( ilt ) { | |
1028 if( ilt->_head == _head ) { | |
1029 uint i; | |
1030 for( i = 2; i < _head->req(); i++ ) | |
1031 if( _head->in(i) == ilt->_tail ) | |
1032 break; // Still a loop | |
1033 if( i == _head->req() ) { // No longer a loop | |
1034 // Flatten ilt. Hang ilt's "_next" list from the end of | |
1035 // ilt's '_child' list. Move the ilt's _child up to replace ilt. | |
1036 IdealLoopTree **cp = &ilt->_child; | |
1037 while( *cp ) cp = &(*cp)->_next; // Find end of child list | |
1038 *cp = ilt->_next; // Hang next list at end of child list | |
1039 *pilt = ilt->_child; // Move child up to replace ilt | |
1040 ilt->_head = NULL; // Flag as a loop UNIONED into parent | |
1041 ilt = ilt->_child; // Repeat using new ilt | |
1042 continue; // do not advance over ilt->_child | |
1043 } | |
1044 assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" ); | |
1045 phase->set_loop(_head,ilt); | |
1046 } | |
1047 pilt = &ilt->_child; // Advance to next | |
1048 ilt = *pilt; | |
1049 } | |
1050 | |
1051 if( _child ) fix_parent( _child, this ); | |
1052 } | |
1053 | |
1054 //------------------------------beautify_loops--------------------------------- | |
1055 // Split shared headers and insert loop landing pads. | |
1056 // Insert a LoopNode to replace the RegionNode. | |
1057 // Return TRUE if loop tree is structurally changed. | |
1058 bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) { | |
1059 bool result = false; | |
1060 // Cache parts in locals for easy | |
1061 PhaseIterGVN &igvn = phase->_igvn; | |
1062 | |
1063 phase->C->print_method("Before beautify loops", 3); | |
1064 | |
1065 igvn.hash_delete(_head); // Yank from hash before hacking edges | |
1066 | |
1067 // Check for multiple fall-in paths. Peel off a landing pad if need be. | |
1068 int fall_in_cnt = 0; | |
1069 for( uint i = 1; i < _head->req(); i++ ) | |
1070 if( !phase->is_member( this, _head->in(i) ) ) | |
1071 fall_in_cnt++; | |
1072 assert( fall_in_cnt, "at least 1 fall-in path" ); | |
1073 if( fall_in_cnt > 1 ) // Need a loop landing pad to merge fall-ins | |
1074 split_fall_in( phase, fall_in_cnt ); | |
1075 | |
1076 // Swap inputs to the _head and all Phis to move the fall-in edge to | |
1077 // the left. | |
1078 fall_in_cnt = 1; | |
1079 while( phase->is_member( this, _head->in(fall_in_cnt) ) ) | |
1080 fall_in_cnt++; | |
1081 if( fall_in_cnt > 1 ) { | |
1082 // Since I am just swapping inputs I do not need to update def-use info | |
1083 Node *tmp = _head->in(1); | |
1084 _head->set_req( 1, _head->in(fall_in_cnt) ); | |
1085 _head->set_req( fall_in_cnt, tmp ); | |
1086 // Swap also all Phis | |
1087 for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) { | |
1088 Node* phi = _head->fast_out(i); | |
1089 if( phi->is_Phi() ) { | |
1090 igvn.hash_delete(phi); // Yank from hash before hacking edges | |
1091 tmp = phi->in(1); | |
1092 phi->set_req( 1, phi->in(fall_in_cnt) ); | |
1093 phi->set_req( fall_in_cnt, tmp ); | |
1094 } | |
1095 } | |
1096 } | |
1097 assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" ); | |
1098 assert( phase->is_member( this, _head->in(2) ), "right edge is loop" ); | |
1099 | |
1100 // If I am a shared header (multiple backedges), peel off the many | |
1101 // backedges into a private merge point and use the merge point as | |
1102 // the one true backedge. | |
1103 if( _head->req() > 3 ) { | |
1104 // Merge the many backedges into a single backedge. | |
1105 merge_many_backedges( phase ); | |
1106 result = true; | |
1107 } | |
1108 | |
1109 // If I am a shared header (multiple backedges), peel off myself loop. | |
1110 // I better be the outermost loop. | |
1111 if( _head->req() > 3 ) { | |
1112 split_outer_loop( phase ); | |
1113 result = true; | |
1114 | |
1115 } else if( !_head->is_Loop() && !_irreducible ) { | |
1116 // Make a new LoopNode to replace the old loop head | |
1117 Node *l = new (phase->C, 3) LoopNode( _head->in(1), _head->in(2) ); | |
1118 l = igvn.register_new_node_with_optimizer(l, _head); | |
1119 phase->set_created_loop_node(); | |
1120 // Go ahead and replace _head | |
1121 phase->_igvn.subsume_node( _head, l ); | |
1122 _head = l; | |
1123 phase->set_loop(_head, this); | |
1124 for (DUIterator_Fast imax, i = l->fast_outs(imax); i < imax; i++) | |
1125 phase->_igvn.add_users_to_worklist(l->fast_out(i)); | |
1126 } | |
1127 | |
1128 phase->C->print_method("After beautify loops", 3); | |
1129 | |
1130 // Now recursively beautify nested loops | |
1131 if( _child ) result |= _child->beautify_loops( phase ); | |
1132 if( _next ) result |= _next ->beautify_loops( phase ); | |
1133 return result; | |
1134 } | |
1135 | |
1136 //------------------------------allpaths_check_safepts---------------------------- | |
1137 // Allpaths backwards scan from loop tail, terminating each path at first safepoint | |
1138 // encountered. Helper for check_safepts. | |
1139 void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) { | |
1140 assert(stack.size() == 0, "empty stack"); | |
1141 stack.push(_tail); | |
1142 visited.Clear(); | |
1143 visited.set(_tail->_idx); | |
1144 while (stack.size() > 0) { | |
1145 Node* n = stack.pop(); | |
1146 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) { | |
1147 // Terminate this path | |
1148 } else if (n->Opcode() == Op_SafePoint) { | |
1149 if (_phase->get_loop(n) != this) { | |
1150 if (_required_safept == NULL) _required_safept = new Node_List(); | |
1151 _required_safept->push(n); // save the one closest to the tail | |
1152 } | |
1153 // Terminate this path | |
1154 } else { | |
1155 uint start = n->is_Region() ? 1 : 0; | |
1156 uint end = n->is_Region() && !n->is_Loop() ? n->req() : start + 1; | |
1157 for (uint i = start; i < end; i++) { | |
1158 Node* in = n->in(i); | |
1159 assert(in->is_CFG(), "must be"); | |
1160 if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) { | |
1161 stack.push(in); | |
1162 } | |
1163 } | |
1164 } | |
1165 } | |
1166 } | |
1167 | |
1168 //------------------------------check_safepts---------------------------- | |
1169 // Given dominators, try to find loops with calls that must always be | |
1170 // executed (call dominates loop tail). These loops do not need non-call | |
1171 // safepoints (ncsfpt). | |
1172 // | |
1173 // A complication is that a safepoint in a inner loop may be needed | |
1174 // by an outer loop. In the following, the inner loop sees it has a | |
1175 // call (block 3) on every path from the head (block 2) to the | |
1176 // backedge (arc 3->2). So it deletes the ncsfpt (non-call safepoint) | |
1177 // in block 2, _but_ this leaves the outer loop without a safepoint. | |
1178 // | |
1179 // entry 0 | |
1180 // | | |
1181 // v | |
1182 // outer 1,2 +->1 | |
1183 // | | | |
1184 // | v | |
1185 // | 2<---+ ncsfpt in 2 | |
1186 // |_/|\ | | |
1187 // | v | | |
1188 // inner 2,3 / 3 | call in 3 | |
1189 // / | | | |
1190 // v +--+ | |
1191 // exit 4 | |
1192 // | |
1193 // | |
1194 // This method creates a list (_required_safept) of ncsfpt nodes that must | |
1195 // be protected is created for each loop. When a ncsfpt maybe deleted, it | |
1196 // is first looked for in the lists for the outer loops of the current loop. | |
1197 // | |
1198 // The insights into the problem: | |
1199 // A) counted loops are okay | |
1200 // B) innermost loops are okay (only an inner loop can delete | |
1201 // a ncsfpt needed by an outer loop) | |
1202 // C) a loop is immune from an inner loop deleting a safepoint | |
1203 // if the loop has a call on the idom-path | |
1204 // D) a loop is also immune if it has a ncsfpt (non-call safepoint) on the | |
1205 // idom-path that is not in a nested loop | |
1206 // E) otherwise, an ncsfpt on the idom-path that is nested in an inner | |
1207 // loop needs to be prevented from deletion by an inner loop | |
1208 // | |
1209 // There are two analyses: | |
1210 // 1) The first, and cheaper one, scans the loop body from | |
1211 // tail to head following the idom (immediate dominator) | |
1212 // chain, looking for the cases (C,D,E) above. | |
1213 // Since inner loops are scanned before outer loops, there is summary | |
1214 // information about inner loops. Inner loops can be skipped over | |
1215 // when the tail of an inner loop is encountered. | |
1216 // | |
1217 // 2) The second, invoked if the first fails to find a call or ncsfpt on | |
1218 // the idom path (which is rare), scans all predecessor control paths | |
1219 // from the tail to the head, terminating a path when a call or sfpt | |
1220 // is encountered, to find the ncsfpt's that are closest to the tail. | |
1221 // | |
1222 void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) { | |
1223 // Bottom up traversal | |
1224 IdealLoopTree* ch = _child; | |
1225 while (ch != NULL) { | |
1226 ch->check_safepts(visited, stack); | |
1227 ch = ch->_next; | |
1228 } | |
1229 | |
1230 if (!_head->is_CountedLoop() && !_has_sfpt && _parent != NULL && !_irreducible) { | |
1231 bool has_call = false; // call on dom-path | |
1232 bool has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth | |
1233 Node* nonlocal_ncsfpt = NULL; // ncsfpt on dom-path at a deeper depth | |
1234 // Scan the dom-path nodes from tail to head | |
1235 for (Node* n = tail(); n != _head; n = _phase->idom(n)) { | |
1236 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) { | |
1237 has_call = true; | |
1238 _has_sfpt = 1; // Then no need for a safept! | |
1239 break; | |
1240 } else if (n->Opcode() == Op_SafePoint) { | |
1241 if (_phase->get_loop(n) == this) { | |
1242 has_local_ncsfpt = true; | |
1243 break; | |
1244 } | |
1245 if (nonlocal_ncsfpt == NULL) { | |
1246 nonlocal_ncsfpt = n; // save the one closest to the tail | |
1247 } | |
1248 } else { | |
1249 IdealLoopTree* nlpt = _phase->get_loop(n); | |
1250 if (this != nlpt) { | |
1251 // If at an inner loop tail, see if the inner loop has already | |
1252 // recorded seeing a call on the dom-path (and stop.) If not, | |
1253 // jump to the head of the inner loop. | |
1254 assert(is_member(nlpt), "nested loop"); | |
1255 Node* tail = nlpt->_tail; | |
1256 if (tail->in(0)->is_If()) tail = tail->in(0); | |
1257 if (n == tail) { | |
1258 // If inner loop has call on dom-path, so does outer loop | |
1259 if (nlpt->_has_sfpt) { | |
1260 has_call = true; | |
1261 _has_sfpt = 1; | |
1262 break; | |
1263 } | |
1264 // Skip to head of inner loop | |
1265 assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head"); | |
1266 n = nlpt->_head; | |
1267 } | |
1268 } | |
1269 } | |
1270 } | |
1271 // Record safept's that this loop needs preserved when an | |
1272 // inner loop attempts to delete it's safepoints. | |
1273 if (_child != NULL && !has_call && !has_local_ncsfpt) { | |
1274 if (nonlocal_ncsfpt != NULL) { | |
1275 if (_required_safept == NULL) _required_safept = new Node_List(); | |
1276 _required_safept->push(nonlocal_ncsfpt); | |
1277 } else { | |
1278 // Failed to find a suitable safept on the dom-path. Now use | |
1279 // an all paths walk from tail to head, looking for safepoints to preserve. | |
1280 allpaths_check_safepts(visited, stack); | |
1281 } | |
1282 } | |
1283 } | |
1284 } | |
1285 | |
1286 //---------------------------is_deleteable_safept---------------------------- | |
1287 // Is safept not required by an outer loop? | |
1288 bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) { | |
1289 assert(sfpt->Opcode() == Op_SafePoint, ""); | |
1290 IdealLoopTree* lp = get_loop(sfpt)->_parent; | |
1291 while (lp != NULL) { | |
1292 Node_List* sfpts = lp->_required_safept; | |
1293 if (sfpts != NULL) { | |
1294 for (uint i = 0; i < sfpts->size(); i++) { | |
1295 if (sfpt == sfpts->at(i)) | |
1296 return false; | |
1297 } | |
1298 } | |
1299 lp = lp->_parent; | |
1300 } | |
1301 return true; | |
1302 } | |
1303 | |
1304 //------------------------------counted_loop----------------------------------- | |
1305 // Convert to counted loops where possible | |
1306 void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) { | |
1307 | |
1308 // For grins, set the inner-loop flag here | |
1309 if( !_child ) { | |
1310 if( _head->is_Loop() ) _head->as_Loop()->set_inner_loop(); | |
1311 } | |
1312 | |
1313 if( _head->is_CountedLoop() || | |
1314 phase->is_counted_loop( _head, this ) ) { | |
1315 _has_sfpt = 1; // Indicate we do not need a safepoint here | |
1316 | |
1317 // Look for a safepoint to remove | |
1318 for (Node* n = tail(); n != _head; n = phase->idom(n)) | |
1319 if (n->Opcode() == Op_SafePoint && phase->get_loop(n) == this && | |
1320 phase->is_deleteable_safept(n)) | |
1321 phase->lazy_replace(n,n->in(TypeFunc::Control)); | |
1322 | |
1323 CountedLoopNode *cl = _head->as_CountedLoop(); | |
1324 Node *incr = cl->incr(); | |
1325 if( !incr ) return; // Dead loop? | |
1326 Node *init = cl->init_trip(); | |
1327 Node *phi = cl->phi(); | |
1328 // protect against stride not being a constant | |
1329 if( !cl->stride_is_con() ) return; | |
1330 int stride_con = cl->stride_con(); | |
1331 | |
1332 // Look for induction variables | |
1333 | |
1334 // Visit all children, looking for Phis | |
1335 for (DUIterator i = cl->outs(); cl->has_out(i); i++) { | |
1336 Node *out = cl->out(i); | |
1337 if (!out->is_Phi()) continue; // Looking for phis | |
1338 PhiNode* phi2 = out->as_Phi(); | |
1339 Node *incr2 = phi2->in( LoopNode::LoopBackControl ); | |
1340 // Look for induction variables of the form: X += constant | |
1341 if( phi2->region() != _head || | |
1342 incr2->req() != 3 || | |
1343 incr2->in(1) != phi2 || | |
1344 incr2 == incr || | |
1345 incr2->Opcode() != Op_AddI || | |
1346 !incr2->in(2)->is_Con() ) | |
1347 continue; | |
1348 | |
1349 // Check for parallel induction variable (parallel to trip counter) | |
1350 // via an affine function. In particular, count-down loops with | |
1351 // count-up array indices are common. We only RCE references off | |
1352 // the trip-counter, so we need to convert all these to trip-counter | |
1353 // expressions. | |
1354 Node *init2 = phi2->in( LoopNode::EntryControl ); | |
1355 int stride_con2 = incr2->in(2)->get_int(); | |
1356 | |
1357 // The general case here gets a little tricky. We want to find the | |
1358 // GCD of all possible parallel IV's and make a new IV using this | |
1359 // GCD for the loop. Then all possible IVs are simple multiples of | |
1360 // the GCD. In practice, this will cover very few extra loops. | |
1361 // Instead we require 'stride_con2' to be a multiple of 'stride_con', | |
1362 // where +/-1 is the common case, but other integer multiples are | |
1363 // also easy to handle. | |
1364 int ratio_con = stride_con2/stride_con; | |
1365 | |
1366 if( ratio_con * stride_con == stride_con2 ) { // Check for exact | |
1367 // Convert to using the trip counter. The parallel induction | |
1368 // variable differs from the trip counter by a loop-invariant | |
1369 // amount, the difference between their respective initial values. | |
1370 // It is scaled by the 'ratio_con'. | |
1371 Compile* C = phase->C; | |
1372 Node* ratio = phase->_igvn.intcon(ratio_con); | |
1373 phase->set_ctrl(ratio, C->root()); | |
1374 Node* ratio_init = new (C, 3) MulINode(init, ratio); | |
1375 phase->_igvn.register_new_node_with_optimizer(ratio_init, init); | |
1376 phase->set_early_ctrl(ratio_init); | |
1377 Node* diff = new (C, 3) SubINode(init2, ratio_init); | |
1378 phase->_igvn.register_new_node_with_optimizer(diff, init2); | |
1379 phase->set_early_ctrl(diff); | |
1380 Node* ratio_idx = new (C, 3) MulINode(phi, ratio); | |
1381 phase->_igvn.register_new_node_with_optimizer(ratio_idx, phi); | |
1382 phase->set_ctrl(ratio_idx, cl); | |
1383 Node* add = new (C, 3) AddINode(ratio_idx, diff); | |
1384 phase->_igvn.register_new_node_with_optimizer(add); | |
1385 phase->set_ctrl(add, cl); | |
1386 phase->_igvn.hash_delete( phi2 ); | |
1387 phase->_igvn.subsume_node( phi2, add ); | |
1388 // Sometimes an induction variable is unused | |
1389 if (add->outcnt() == 0) { | |
1390 phase->_igvn.remove_dead_node(add); | |
1391 } | |
1392 --i; // deleted this phi; rescan starting with next position | |
1393 continue; | |
1394 } | |
1395 } | |
1396 } else if (_parent != NULL && !_irreducible) { | |
1397 // Not a counted loop. | |
1398 // Look for a safepoint on the idom-path to remove, preserving the first one | |
1399 bool found = false; | |
1400 Node* n = tail(); | |
1401 for (; n != _head && !found; n = phase->idom(n)) { | |
1402 if (n->Opcode() == Op_SafePoint && phase->get_loop(n) == this) | |
1403 found = true; // Found one | |
1404 } | |
1405 // Skip past it and delete the others | |
1406 for (; n != _head; n = phase->idom(n)) { | |
1407 if (n->Opcode() == Op_SafePoint && phase->get_loop(n) == this && | |
1408 phase->is_deleteable_safept(n)) | |
1409 phase->lazy_replace(n,n->in(TypeFunc::Control)); | |
1410 } | |
1411 } | |
1412 | |
1413 // Recursively | |
1414 if( _child ) _child->counted_loop( phase ); | |
1415 if( _next ) _next ->counted_loop( phase ); | |
1416 } | |
1417 | |
1418 #ifndef PRODUCT | |
1419 //------------------------------dump_head-------------------------------------- | |
1420 // Dump 1 liner for loop header info | |
1421 void IdealLoopTree::dump_head( ) const { | |
1422 for( uint i=0; i<_nest; i++ ) | |
1423 tty->print(" "); | |
1424 tty->print("Loop: N%d/N%d ",_head->_idx,_tail->_idx); | |
1425 if( _irreducible ) tty->print(" IRREDUCIBLE"); | |
1426 if( _head->is_CountedLoop() ) { | |
1427 CountedLoopNode *cl = _head->as_CountedLoop(); | |
1428 tty->print(" counted"); | |
1429 if( cl->is_pre_loop () ) tty->print(" pre" ); | |
1430 if( cl->is_main_loop() ) tty->print(" main"); | |
1431 if( cl->is_post_loop() ) tty->print(" post"); | |
1432 } | |
1433 tty->cr(); | |
1434 } | |
1435 | |
1436 //------------------------------dump------------------------------------------- | |
1437 // Dump loops by loop tree | |
1438 void IdealLoopTree::dump( ) const { | |
1439 dump_head(); | |
1440 if( _child ) _child->dump(); | |
1441 if( _next ) _next ->dump(); | |
1442 } | |
1443 | |
1444 #endif | |
1445 | |
1446 //============================================================================= | |
1447 //------------------------------PhaseIdealLoop--------------------------------- | |
1448 // Create a PhaseLoop. Build the ideal Loop tree. Map each Ideal Node to | |
1449 // its corresponding LoopNode. If 'optimize' is true, do some loop cleanups. | |
1450 PhaseIdealLoop::PhaseIdealLoop( PhaseIterGVN &igvn, const PhaseIdealLoop *verify_me, bool do_split_ifs ) | |
1451 : PhaseTransform(Ideal_Loop), | |
1452 _igvn(igvn), | |
1453 _dom_lca_tags(C->comp_arena()) { | |
1454 // Reset major-progress flag for the driver's heuristics | |
1455 C->clear_major_progress(); | |
1456 | |
1457 #ifndef PRODUCT | |
1458 // Capture for later assert | |
1459 uint unique = C->unique(); | |
1460 _loop_invokes++; | |
1461 _loop_work += unique; | |
1462 #endif | |
1463 | |
1464 // True if the method has at least 1 irreducible loop | |
1465 _has_irreducible_loops = false; | |
1466 | |
1467 _created_loop_node = false; | |
1468 | |
1469 Arena *a = Thread::current()->resource_area(); | |
1470 VectorSet visited(a); | |
1471 // Pre-grow the mapping from Nodes to IdealLoopTrees. | |
1472 _nodes.map(C->unique(), NULL); | |
1473 memset(_nodes.adr(), 0, wordSize * C->unique()); | |
1474 | |
1475 // Pre-build the top-level outermost loop tree entry | |
1476 _ltree_root = new IdealLoopTree( this, C->root(), C->root() ); | |
1477 // Do not need a safepoint at the top level | |
1478 _ltree_root->_has_sfpt = 1; | |
1479 | |
1480 // Empty pre-order array | |
1481 allocate_preorders(); | |
1482 | |
1483 // Build a loop tree on the fly. Build a mapping from CFG nodes to | |
1484 // IdealLoopTree entries. Data nodes are NOT walked. | |
1485 build_loop_tree(); | |
1486 // Check for bailout, and return | |
1487 if (C->failing()) { | |
1488 return; | |
1489 } | |
1490 | |
1491 // No loops after all | |
1492 if( !_ltree_root->_child ) C->set_has_loops(false); | |
1493 | |
1494 // There should always be an outer loop containing the Root and Return nodes. | |
1495 // If not, we have a degenerate empty program. Bail out in this case. | |
1496 if (!has_node(C->root())) { | |
1497 C->clear_major_progress(); | |
1498 C->record_method_not_compilable("empty program detected during loop optimization"); | |
1499 return; | |
1500 } | |
1501 | |
1502 // Nothing to do, so get out | |
1503 if( !C->has_loops() && !do_split_ifs && !verify_me) { | |
1504 _igvn.optimize(); // Cleanup NeverBranches | |
1505 return; | |
1506 } | |
1507 | |
1508 // Set loop nesting depth | |
1509 _ltree_root->set_nest( 0 ); | |
1510 | |
1511 // Split shared headers and insert loop landing pads. | |
1512 // Do not bother doing this on the Root loop of course. | |
1513 if( !verify_me && _ltree_root->_child ) { | |
1514 if( _ltree_root->_child->beautify_loops( this ) ) { | |
1515 // Re-build loop tree! | |
1516 _ltree_root->_child = NULL; | |
1517 _nodes.clear(); | |
1518 reallocate_preorders(); | |
1519 build_loop_tree(); | |
1520 // Check for bailout, and return | |
1521 if (C->failing()) { | |
1522 return; | |
1523 } | |
1524 // Reset loop nesting depth | |
1525 _ltree_root->set_nest( 0 ); | |
1526 } | |
1527 } | |
1528 | |
1529 // Build Dominators for elision of NULL checks & loop finding. | |
1530 // Since nodes do not have a slot for immediate dominator, make | |
1531 // a persistant side array for that info indexed on node->_idx. | |
1532 _idom_size = C->unique(); | |
1533 _idom = NEW_RESOURCE_ARRAY( Node*, _idom_size ); | |
1534 _dom_depth = NEW_RESOURCE_ARRAY( uint, _idom_size ); | |
1535 _dom_stk = NULL; // Allocated on demand in recompute_dom_depth | |
1536 memset( _dom_depth, 0, _idom_size * sizeof(uint) ); | |
1537 | |
1538 Dominators(); | |
1539 | |
1540 // As a side effect, Dominators removed any unreachable CFG paths | |
1541 // into RegionNodes. It doesn't do this test against Root, so | |
1542 // we do it here. | |
1543 for( uint i = 1; i < C->root()->req(); i++ ) { | |
1544 if( !_nodes[C->root()->in(i)->_idx] ) { // Dead path into Root? | |
1545 _igvn.hash_delete(C->root()); | |
1546 C->root()->del_req(i); | |
1547 _igvn._worklist.push(C->root()); | |
1548 i--; // Rerun same iteration on compressed edges | |
1549 } | |
1550 } | |
1551 | |
1552 // Given dominators, try to find inner loops with calls that must | |
1553 // always be executed (call dominates loop tail). These loops do | |
1554 // not need a seperate safepoint. | |
1555 Node_List cisstack(a); | |
1556 _ltree_root->check_safepts(visited, cisstack); | |
1557 | |
1558 // Walk the DATA nodes and place into loops. Find earliest control | |
1559 // node. For CFG nodes, the _nodes array starts out and remains | |
1560 // holding the associated IdealLoopTree pointer. For DATA nodes, the | |
1561 // _nodes array holds the earliest legal controlling CFG node. | |
1562 | |
1563 // Allocate stack with enough space to avoid frequent realloc | |
1564 int stack_size = (C->unique() >> 1) + 16; // (unique>>1)+16 from Java2D stats | |
1565 Node_Stack nstack( a, stack_size ); | |
1566 | |
1567 visited.Clear(); | |
1568 Node_List worklist(a); | |
1569 // Don't need C->root() on worklist since | |
1570 // it will be processed among C->top() inputs | |
1571 worklist.push( C->top() ); | |
1572 visited.set( C->top()->_idx ); // Set C->top() as visited now | |
1573 build_loop_early( visited, worklist, nstack, verify_me ); | |
1574 | |
1575 // Given early legal placement, try finding counted loops. This placement | |
1576 // is good enough to discover most loop invariants. | |
1577 if( !verify_me ) | |
1578 _ltree_root->counted_loop( this ); | |
1579 | |
1580 // Find latest loop placement. Find ideal loop placement. | |
1581 visited.Clear(); | |
1582 init_dom_lca_tags(); | |
1583 // Need C->root() on worklist when processing outs | |
1584 worklist.push( C->root() ); | |
1585 NOT_PRODUCT( C->verify_graph_edges(); ) | |
1586 worklist.push( C->top() ); | |
1587 build_loop_late( visited, worklist, nstack, verify_me ); | |
1588 | |
1589 // clear out the dead code | |
1590 while(_deadlist.size()) { | |
1591 igvn.remove_globally_dead_node(_deadlist.pop()); | |
1592 } | |
1593 | |
1594 #ifndef PRODUCT | |
1595 C->verify_graph_edges(); | |
1596 if( verify_me ) { // Nested verify pass? | |
1597 // Check to see if the verify mode is broken | |
1598 assert(C->unique() == unique, "non-optimize mode made Nodes? ? ?"); | |
1599 return; | |
1600 } | |
1601 if( VerifyLoopOptimizations ) verify(); | |
1602 #endif | |
1603 | |
1604 if (ReassociateInvariants) { | |
1605 // Reassociate invariants and prep for split_thru_phi | |
1606 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { | |
1607 IdealLoopTree* lpt = iter.current(); | |
1608 if (!lpt->is_counted() || !lpt->is_inner()) continue; | |
1609 | |
1610 lpt->reassociate_invariants(this); | |
1611 | |
1612 // Because RCE opportunities can be masked by split_thru_phi, | |
1613 // look for RCE candidates and inhibit split_thru_phi | |
1614 // on just their loop-phi's for this pass of loop opts | |
1615 if( SplitIfBlocks && do_split_ifs ) { | |
1616 if (lpt->policy_range_check(this)) { | |
1617 lpt->_rce_candidate = true; | |
1618 } | |
1619 } | |
1620 } | |
1621 } | |
1622 | |
1623 // Check for aggressive application of split-if and other transforms | |
1624 // that require basic-block info (like cloning through Phi's) | |
1625 if( SplitIfBlocks && do_split_ifs ) { | |
1626 visited.Clear(); | |
1627 split_if_with_blocks( visited, nstack ); | |
1628 NOT_PRODUCT( if( VerifyLoopOptimizations ) verify(); ); | |
1629 } | |
1630 | |
1631 // Perform iteration-splitting on inner loops. Split iterations to avoid | |
1632 // range checks or one-shot null checks. | |
1633 | |
1634 // If split-if's didn't hack the graph too bad (no CFG changes) | |
1635 // then do loop opts. | |
1636 if( C->has_loops() && !C->major_progress() ) { | |
1637 memset( worklist.adr(), 0, worklist.Size()*sizeof(Node*) ); | |
1638 _ltree_root->_child->iteration_split( this, worklist ); | |
1639 // No verify after peeling! GCM has hoisted code out of the loop. | |
1640 // After peeling, the hoisted code could sink inside the peeled area. | |
1641 // The peeling code does not try to recompute the best location for | |
1642 // all the code before the peeled area, so the verify pass will always | |
1643 // complain about it. | |
1644 } | |
1645 // Do verify graph edges in any case | |
1646 NOT_PRODUCT( C->verify_graph_edges(); ); | |
1647 | |
1648 if( !do_split_ifs ) { | |
1649 // We saw major progress in Split-If to get here. We forced a | |
1650 // pass with unrolling and not split-if, however more split-if's | |
1651 // might make progress. If the unrolling didn't make progress | |
1652 // then the major-progress flag got cleared and we won't try | |
1653 // another round of Split-If. In particular the ever-common | |
1654 // instance-of/check-cast pattern requires at least 2 rounds of | |
1655 // Split-If to clear out. | |
1656 C->set_major_progress(); | |
1657 } | |
1658 | |
1659 // Repeat loop optimizations if new loops were seen | |
1660 if (created_loop_node()) { | |
1661 C->set_major_progress(); | |
1662 } | |
1663 | |
1664 // Convert scalar to superword operations | |
1665 | |
1666 if (UseSuperWord && C->has_loops() && !C->major_progress()) { | |
1667 // SuperWord transform | |
1668 SuperWord sw(this); | |
1669 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { | |
1670 IdealLoopTree* lpt = iter.current(); | |
1671 if (lpt->is_counted()) { | |
1672 sw.transform_loop(lpt); | |
1673 } | |
1674 } | |
1675 } | |
1676 | |
1677 // Cleanup any modified bits | |
1678 _igvn.optimize(); | |
1679 | |
1680 // Do not repeat loop optimizations if irreducible loops are present | |
1681 // by claiming no-progress. | |
1682 if( _has_irreducible_loops ) | |
1683 C->clear_major_progress(); | |
1684 } | |
1685 | |
1686 #ifndef PRODUCT | |
1687 //------------------------------print_statistics------------------------------- | |
1688 int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes | |
1689 int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique | |
1690 void PhaseIdealLoop::print_statistics() { | |
1691 tty->print_cr("PhaseIdealLoop=%d, sum _unique=%d", _loop_invokes, _loop_work); | |
1692 } | |
1693 | |
1694 //------------------------------verify----------------------------------------- | |
1695 // Build a verify-only PhaseIdealLoop, and see that it agrees with me. | |
1696 static int fail; // debug only, so its multi-thread dont care | |
1697 void PhaseIdealLoop::verify() const { | |
1698 int old_progress = C->major_progress(); | |
1699 ResourceMark rm; | |
1700 PhaseIdealLoop loop_verify( _igvn, this, false ); | |
1701 VectorSet visited(Thread::current()->resource_area()); | |
1702 | |
1703 fail = 0; | |
1704 verify_compare( C->root(), &loop_verify, visited ); | |
1705 assert( fail == 0, "verify loops failed" ); | |
1706 // Verify loop structure is the same | |
1707 _ltree_root->verify_tree(loop_verify._ltree_root, NULL); | |
1708 // Reset major-progress. It was cleared by creating a verify version of | |
1709 // PhaseIdealLoop. | |
1710 for( int i=0; i<old_progress; i++ ) | |
1711 C->set_major_progress(); | |
1712 } | |
1713 | |
1714 //------------------------------verify_compare--------------------------------- | |
1715 // Make sure me and the given PhaseIdealLoop agree on key data structures | |
1716 void PhaseIdealLoop::verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const { | |
1717 if( !n ) return; | |
1718 if( visited.test_set( n->_idx ) ) return; | |
1719 if( !_nodes[n->_idx] ) { // Unreachable | |
1720 assert( !loop_verify->_nodes[n->_idx], "both should be unreachable" ); | |
1721 return; | |
1722 } | |
1723 | |
1724 uint i; | |
1725 for( i = 0; i < n->req(); i++ ) | |
1726 verify_compare( n->in(i), loop_verify, visited ); | |
1727 | |
1728 // Check the '_nodes' block/loop structure | |
1729 i = n->_idx; | |
1730 if( has_ctrl(n) ) { // We have control; verify has loop or ctrl | |
1731 if( _nodes[i] != loop_verify->_nodes[i] && | |
1732 get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) { | |
1733 tty->print("Mismatched control setting for: "); | |
1734 n->dump(); | |
1735 if( fail++ > 10 ) return; | |
1736 Node *c = get_ctrl_no_update(n); | |
1737 tty->print("We have it as: "); | |
1738 if( c->in(0) ) c->dump(); | |
1739 else tty->print_cr("N%d",c->_idx); | |
1740 tty->print("Verify thinks: "); | |
1741 if( loop_verify->has_ctrl(n) ) | |
1742 loop_verify->get_ctrl_no_update(n)->dump(); | |
1743 else | |
1744 loop_verify->get_loop_idx(n)->dump(); | |
1745 tty->cr(); | |
1746 } | |
1747 } else { // We have a loop | |
1748 IdealLoopTree *us = get_loop_idx(n); | |
1749 if( loop_verify->has_ctrl(n) ) { | |
1750 tty->print("Mismatched loop setting for: "); | |
1751 n->dump(); | |
1752 if( fail++ > 10 ) return; | |
1753 tty->print("We have it as: "); | |
1754 us->dump(); | |
1755 tty->print("Verify thinks: "); | |
1756 loop_verify->get_ctrl_no_update(n)->dump(); | |
1757 tty->cr(); | |
1758 } else if (!C->major_progress()) { | |
1759 // Loop selection can be messed up if we did a major progress | |
1760 // operation, like split-if. Do not verify in that case. | |
1761 IdealLoopTree *them = loop_verify->get_loop_idx(n); | |
1762 if( us->_head != them->_head || us->_tail != them->_tail ) { | |
1763 tty->print("Unequals loops for: "); | |
1764 n->dump(); | |
1765 if( fail++ > 10 ) return; | |
1766 tty->print("We have it as: "); | |
1767 us->dump(); | |
1768 tty->print("Verify thinks: "); | |
1769 them->dump(); | |
1770 tty->cr(); | |
1771 } | |
1772 } | |
1773 } | |
1774 | |
1775 // Check for immediate dominators being equal | |
1776 if( i >= _idom_size ) { | |
1777 if( !n->is_CFG() ) return; | |
1778 tty->print("CFG Node with no idom: "); | |
1779 n->dump(); | |
1780 return; | |
1781 } | |
1782 if( !n->is_CFG() ) return; | |
1783 if( n == C->root() ) return; // No IDOM here | |
1784 | |
1785 assert(n->_idx == i, "sanity"); | |
1786 Node *id = idom_no_update(n); | |
1787 if( id != loop_verify->idom_no_update(n) ) { | |
1788 tty->print("Unequals idoms for: "); | |
1789 n->dump(); | |
1790 if( fail++ > 10 ) return; | |
1791 tty->print("We have it as: "); | |
1792 id->dump(); | |
1793 tty->print("Verify thinks: "); | |
1794 loop_verify->idom_no_update(n)->dump(); | |
1795 tty->cr(); | |
1796 } | |
1797 | |
1798 } | |
1799 | |
1800 //------------------------------verify_tree------------------------------------ | |
1801 // Verify that tree structures match. Because the CFG can change, siblings | |
1802 // within the loop tree can be reordered. We attempt to deal with that by | |
1803 // reordering the verify's loop tree if possible. | |
1804 void IdealLoopTree::verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const { | |
1805 assert( _parent == parent, "Badly formed loop tree" ); | |
1806 | |
1807 // Siblings not in same order? Attempt to re-order. | |
1808 if( _head != loop->_head ) { | |
1809 // Find _next pointer to update | |
1810 IdealLoopTree **pp = &loop->_parent->_child; | |
1811 while( *pp != loop ) | |
1812 pp = &((*pp)->_next); | |
1813 // Find proper sibling to be next | |
1814 IdealLoopTree **nn = &loop->_next; | |
1815 while( (*nn) && (*nn)->_head != _head ) | |
1816 nn = &((*nn)->_next); | |
1817 | |
1818 // Check for no match. | |
1819 if( !(*nn) ) { | |
1820 // Annoyingly, irreducible loops can pick different headers | |
1821 // after a major_progress operation, so the rest of the loop | |
1822 // tree cannot be matched. | |
1823 if (_irreducible && Compile::current()->major_progress()) return; | |
1824 assert( 0, "failed to match loop tree" ); | |
1825 } | |
1826 | |
1827 // Move (*nn) to (*pp) | |
1828 IdealLoopTree *hit = *nn; | |
1829 *nn = hit->_next; | |
1830 hit->_next = loop; | |
1831 *pp = loop; | |
1832 loop = hit; | |
1833 // Now try again to verify | |
1834 } | |
1835 | |
1836 assert( _head == loop->_head , "mismatched loop head" ); | |
1837 Node *tail = _tail; // Inline a non-updating version of | |
1838 while( !tail->in(0) ) // the 'tail()' call. | |
1839 tail = tail->in(1); | |
1840 assert( tail == loop->_tail, "mismatched loop tail" ); | |
1841 | |
1842 // Counted loops that are guarded should be able to find their guards | |
1843 if( _head->is_CountedLoop() && _head->as_CountedLoop()->is_main_loop() ) { | |
1844 CountedLoopNode *cl = _head->as_CountedLoop(); | |
1845 Node *init = cl->init_trip(); | |
1846 Node *ctrl = cl->in(LoopNode::EntryControl); | |
1847 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" ); | |
1848 Node *iff = ctrl->in(0); | |
1849 assert( iff->Opcode() == Op_If, "" ); | |
1850 Node *bol = iff->in(1); | |
1851 assert( bol->Opcode() == Op_Bool, "" ); | |
1852 Node *cmp = bol->in(1); | |
1853 assert( cmp->Opcode() == Op_CmpI, "" ); | |
1854 Node *add = cmp->in(1); | |
1855 Node *opaq; | |
1856 if( add->Opcode() == Op_Opaque1 ) { | |
1857 opaq = add; | |
1858 } else { | |
1859 assert( add->Opcode() == Op_AddI || add->Opcode() == Op_ConI , "" ); | |
1860 assert( add == init, "" ); | |
1861 opaq = cmp->in(2); | |
1862 } | |
1863 assert( opaq->Opcode() == Op_Opaque1, "" ); | |
1864 | |
1865 } | |
1866 | |
1867 if (_child != NULL) _child->verify_tree(loop->_child, this); | |
1868 if (_next != NULL) _next ->verify_tree(loop->_next, parent); | |
1869 // Innermost loops need to verify loop bodies, | |
1870 // but only if no 'major_progress' | |
1871 int fail = 0; | |
1872 if (!Compile::current()->major_progress() && _child == NULL) { | |
1873 for( uint i = 0; i < _body.size(); i++ ) { | |
1874 Node *n = _body.at(i); | |
1875 if (n->outcnt() == 0) continue; // Ignore dead | |
1876 uint j; | |
1877 for( j = 0; j < loop->_body.size(); j++ ) | |
1878 if( loop->_body.at(j) == n ) | |
1879 break; | |
1880 if( j == loop->_body.size() ) { // Not found in loop body | |
1881 // Last ditch effort to avoid assertion: Its possible that we | |
1882 // have some users (so outcnt not zero) but are still dead. | |
1883 // Try to find from root. | |
1884 if (Compile::current()->root()->find(n->_idx)) { | |
1885 fail++; | |
1886 tty->print("We have that verify does not: "); | |
1887 n->dump(); | |
1888 } | |
1889 } | |
1890 } | |
1891 for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) { | |
1892 Node *n = loop->_body.at(i2); | |
1893 if (n->outcnt() == 0) continue; // Ignore dead | |
1894 uint j; | |
1895 for( j = 0; j < _body.size(); j++ ) | |
1896 if( _body.at(j) == n ) | |
1897 break; | |
1898 if( j == _body.size() ) { // Not found in loop body | |
1899 // Last ditch effort to avoid assertion: Its possible that we | |
1900 // have some users (so outcnt not zero) but are still dead. | |
1901 // Try to find from root. | |
1902 if (Compile::current()->root()->find(n->_idx)) { | |
1903 fail++; | |
1904 tty->print("Verify has that we do not: "); | |
1905 n->dump(); | |
1906 } | |
1907 } | |
1908 } | |
1909 assert( !fail, "loop body mismatch" ); | |
1910 } | |
1911 } | |
1912 | |
1913 #endif | |
1914 | |
1915 //------------------------------set_idom--------------------------------------- | |
1916 void PhaseIdealLoop::set_idom(Node* d, Node* n, uint dom_depth) { | |
1917 uint idx = d->_idx; | |
1918 if (idx >= _idom_size) { | |
1919 uint newsize = _idom_size<<1; | |
1920 while( idx >= newsize ) { | |
1921 newsize <<= 1; | |
1922 } | |
1923 _idom = REALLOC_RESOURCE_ARRAY( Node*, _idom,_idom_size,newsize); | |
1924 _dom_depth = REALLOC_RESOURCE_ARRAY( uint, _dom_depth,_idom_size,newsize); | |
1925 memset( _dom_depth + _idom_size, 0, (newsize - _idom_size) * sizeof(uint) ); | |
1926 _idom_size = newsize; | |
1927 } | |
1928 _idom[idx] = n; | |
1929 _dom_depth[idx] = dom_depth; | |
1930 } | |
1931 | |
1932 //------------------------------recompute_dom_depth--------------------------------------- | |
1933 // The dominator tree is constructed with only parent pointers. | |
1934 // This recomputes the depth in the tree by first tagging all | |
1935 // nodes as "no depth yet" marker. The next pass then runs up | |
1936 // the dom tree from each node marked "no depth yet", and computes | |
1937 // the depth on the way back down. | |
1938 void PhaseIdealLoop::recompute_dom_depth() { | |
1939 uint no_depth_marker = C->unique(); | |
1940 uint i; | |
1941 // Initialize depth to "no depth yet" | |
1942 for (i = 0; i < _idom_size; i++) { | |
1943 if (_dom_depth[i] > 0 && _idom[i] != NULL) { | |
1944 _dom_depth[i] = no_depth_marker; | |
1945 } | |
1946 } | |
1947 if (_dom_stk == NULL) { | |
1948 uint init_size = C->unique() / 100; // Guess that 1/100 is a reasonable initial size. | |
1949 if (init_size < 10) init_size = 10; | |
1950 _dom_stk = new (C->node_arena()) GrowableArray<uint>(C->node_arena(), init_size, 0, 0); | |
1951 } | |
1952 // Compute new depth for each node. | |
1953 for (i = 0; i < _idom_size; i++) { | |
1954 uint j = i; | |
1955 // Run up the dom tree to find a node with a depth | |
1956 while (_dom_depth[j] == no_depth_marker) { | |
1957 _dom_stk->push(j); | |
1958 j = _idom[j]->_idx; | |
1959 } | |
1960 // Compute the depth on the way back down this tree branch | |
1961 uint dd = _dom_depth[j] + 1; | |
1962 while (_dom_stk->length() > 0) { | |
1963 uint j = _dom_stk->pop(); | |
1964 _dom_depth[j] = dd; | |
1965 dd++; | |
1966 } | |
1967 } | |
1968 } | |
1969 | |
1970 //------------------------------sort------------------------------------------- | |
1971 // Insert 'loop' into the existing loop tree. 'innermost' is a leaf of the | |
1972 // loop tree, not the root. | |
1973 IdealLoopTree *PhaseIdealLoop::sort( IdealLoopTree *loop, IdealLoopTree *innermost ) { | |
1974 if( !innermost ) return loop; // New innermost loop | |
1975 | |
1976 int loop_preorder = get_preorder(loop->_head); // Cache pre-order number | |
1977 assert( loop_preorder, "not yet post-walked loop" ); | |
1978 IdealLoopTree **pp = &innermost; // Pointer to previous next-pointer | |
1979 IdealLoopTree *l = *pp; // Do I go before or after 'l'? | |
1980 | |
1981 // Insert at start of list | |
1982 while( l ) { // Insertion sort based on pre-order | |
1983 if( l == loop ) return innermost; // Already on list! | |
1984 int l_preorder = get_preorder(l->_head); // Cache pre-order number | |
1985 assert( l_preorder, "not yet post-walked l" ); | |
1986 // Check header pre-order number to figure proper nesting | |
1987 if( loop_preorder > l_preorder ) | |
1988 break; // End of insertion | |
1989 // If headers tie (e.g., shared headers) check tail pre-order numbers. | |
1990 // Since I split shared headers, you'd think this could not happen. | |
1991 // BUT: I must first do the preorder numbering before I can discover I | |
1992 // have shared headers, so the split headers all get the same preorder | |
1993 // number as the RegionNode they split from. | |
1994 if( loop_preorder == l_preorder && | |
1995 get_preorder(loop->_tail) < get_preorder(l->_tail) ) | |
1996 break; // Also check for shared headers (same pre#) | |
1997 pp = &l->_parent; // Chain up list | |
1998 l = *pp; | |
1999 } | |
2000 // Link into list | |
2001 // Point predecessor to me | |
2002 *pp = loop; | |
2003 // Point me to successor | |
2004 IdealLoopTree *p = loop->_parent; | |
2005 loop->_parent = l; // Point me to successor | |
2006 if( p ) sort( p, innermost ); // Insert my parents into list as well | |
2007 return innermost; | |
2008 } | |
2009 | |
2010 //------------------------------build_loop_tree-------------------------------- | |
2011 // I use a modified Vick/Tarjan algorithm. I need pre- and a post- visit | |
2012 // bits. The _nodes[] array is mapped by Node index and holds a NULL for | |
2013 // not-yet-pre-walked, pre-order # for pre-but-not-post-walked and holds the | |
2014 // tightest enclosing IdealLoopTree for post-walked. | |
2015 // | |
2016 // During my forward walk I do a short 1-layer lookahead to see if I can find | |
2017 // a loop backedge with that doesn't have any work on the backedge. This | |
2018 // helps me construct nested loops with shared headers better. | |
2019 // | |
2020 // Once I've done the forward recursion, I do the post-work. For each child | |
2021 // I check to see if there is a backedge. Backedges define a loop! I | |
2022 // insert an IdealLoopTree at the target of the backedge. | |
2023 // | |
2024 // During the post-work I also check to see if I have several children | |
2025 // belonging to different loops. If so, then this Node is a decision point | |
2026 // where control flow can choose to change loop nests. It is at this | |
2027 // decision point where I can figure out how loops are nested. At this | |
2028 // time I can properly order the different loop nests from my children. | |
2029 // Note that there may not be any backedges at the decision point! | |
2030 // | |
2031 // Since the decision point can be far removed from the backedges, I can't | |
2032 // order my loops at the time I discover them. Thus at the decision point | |
2033 // I need to inspect loop header pre-order numbers to properly nest my | |
2034 // loops. This means I need to sort my childrens' loops by pre-order. | |
2035 // The sort is of size number-of-control-children, which generally limits | |
2036 // it to size 2 (i.e., I just choose between my 2 target loops). | |
2037 void PhaseIdealLoop::build_loop_tree() { | |
2038 // Allocate stack of size C->unique()/2 to avoid frequent realloc | |
2039 GrowableArray <Node *> bltstack(C->unique() >> 1); | |
2040 Node *n = C->root(); | |
2041 bltstack.push(n); | |
2042 int pre_order = 1; | |
2043 int stack_size; | |
2044 | |
2045 while ( ( stack_size = bltstack.length() ) != 0 ) { | |
2046 n = bltstack.top(); // Leave node on stack | |
2047 if ( !is_visited(n) ) { | |
2048 // ---- Pre-pass Work ---- | |
2049 // Pre-walked but not post-walked nodes need a pre_order number. | |
2050 | |
2051 set_preorder_visited( n, pre_order ); // set as visited | |
2052 | |
2053 // ---- Scan over children ---- | |
2054 // Scan first over control projections that lead to loop headers. | |
2055 // This helps us find inner-to-outer loops with shared headers better. | |
2056 | |
2057 // Scan children's children for loop headers. | |
2058 for ( int i = n->outcnt() - 1; i >= 0; --i ) { | |
2059 Node* m = n->raw_out(i); // Child | |
2060 if( m->is_CFG() && !is_visited(m) ) { // Only for CFG children | |
2061 // Scan over children's children to find loop | |
2062 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { | |
2063 Node* l = m->fast_out(j); | |
2064 if( is_visited(l) && // Been visited? | |
2065 !is_postvisited(l) && // But not post-visited | |
2066 get_preorder(l) < pre_order ) { // And smaller pre-order | |
2067 // Found! Scan the DFS down this path before doing other paths | |
2068 bltstack.push(m); | |
2069 break; | |
2070 } | |
2071 } | |
2072 } | |
2073 } | |
2074 pre_order++; | |
2075 } | |
2076 else if ( !is_postvisited(n) ) { | |
2077 // Note: build_loop_tree_impl() adds out edges on rare occasions, | |
2078 // such as com.sun.rsasign.am::a. | |
2079 // For non-recursive version, first, process current children. | |
2080 // On next iteration, check if additional children were added. | |
2081 for ( int k = n->outcnt() - 1; k >= 0; --k ) { | |
2082 Node* u = n->raw_out(k); | |
2083 if ( u->is_CFG() && !is_visited(u) ) { | |
2084 bltstack.push(u); | |
2085 } | |
2086 } | |
2087 if ( bltstack.length() == stack_size ) { | |
2088 // There were no additional children, post visit node now | |
2089 (void)bltstack.pop(); // Remove node from stack | |
2090 pre_order = build_loop_tree_impl( n, pre_order ); | |
2091 // Check for bailout | |
2092 if (C->failing()) { | |
2093 return; | |
2094 } | |
2095 // Check to grow _preorders[] array for the case when | |
2096 // build_loop_tree_impl() adds new nodes. | |
2097 check_grow_preorders(); | |
2098 } | |
2099 } | |
2100 else { | |
2101 (void)bltstack.pop(); // Remove post-visited node from stack | |
2102 } | |
2103 } | |
2104 } | |
2105 | |
2106 //------------------------------build_loop_tree_impl--------------------------- | |
2107 int PhaseIdealLoop::build_loop_tree_impl( Node *n, int pre_order ) { | |
2108 // ---- Post-pass Work ---- | |
2109 // Pre-walked but not post-walked nodes need a pre_order number. | |
2110 | |
2111 // Tightest enclosing loop for this Node | |
2112 IdealLoopTree *innermost = NULL; | |
2113 | |
2114 // For all children, see if any edge is a backedge. If so, make a loop | |
2115 // for it. Then find the tightest enclosing loop for the self Node. | |
2116 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { | |
2117 Node* m = n->fast_out(i); // Child | |
2118 if( n == m ) continue; // Ignore control self-cycles | |
2119 if( !m->is_CFG() ) continue;// Ignore non-CFG edges | |
2120 | |
2121 IdealLoopTree *l; // Child's loop | |
2122 if( !is_postvisited(m) ) { // Child visited but not post-visited? | |
2123 // Found a backedge | |
2124 assert( get_preorder(m) < pre_order, "should be backedge" ); | |
2125 // Check for the RootNode, which is already a LoopNode and is allowed | |
2126 // to have multiple "backedges". | |
2127 if( m == C->root()) { // Found the root? | |
2128 l = _ltree_root; // Root is the outermost LoopNode | |
2129 } else { // Else found a nested loop | |
2130 // Insert a LoopNode to mark this loop. | |
2131 l = new IdealLoopTree(this, m, n); | |
2132 } // End of Else found a nested loop | |
2133 if( !has_loop(m) ) // If 'm' does not already have a loop set | |
2134 set_loop(m, l); // Set loop header to loop now | |
2135 | |
2136 } else { // Else not a nested loop | |
2137 if( !_nodes[m->_idx] ) continue; // Dead code has no loop | |
2138 l = get_loop(m); // Get previously determined loop | |
2139 // If successor is header of a loop (nest), move up-loop till it | |
2140 // is a member of some outer enclosing loop. Since there are no | |
2141 // shared headers (I've split them already) I only need to go up | |
2142 // at most 1 level. | |
2143 while( l && l->_head == m ) // Successor heads loop? | |
2144 l = l->_parent; // Move up 1 for me | |
2145 // If this loop is not properly parented, then this loop | |
2146 // has no exit path out, i.e. its an infinite loop. | |
2147 if( !l ) { | |
2148 // Make loop "reachable" from root so the CFG is reachable. Basically | |
2149 // insert a bogus loop exit that is never taken. 'm', the loop head, | |
2150 // points to 'n', one (of possibly many) fall-in paths. There may be | |
2151 // many backedges as well. | |
2152 | |
2153 // Here I set the loop to be the root loop. I could have, after | |
2154 // inserting a bogus loop exit, restarted the recursion and found my | |
2155 // new loop exit. This would make the infinite loop a first-class | |
2156 // loop and it would then get properly optimized. What's the use of | |
2157 // optimizing an infinite loop? | |
2158 l = _ltree_root; // Oops, found infinite loop | |
2159 | |
2160 // Insert the NeverBranch between 'm' and it's control user. | |
2161 NeverBranchNode *iff = new (C, 1) NeverBranchNode( m ); | |
2162 _igvn.register_new_node_with_optimizer(iff); | |
2163 set_loop(iff, l); | |
2164 Node *if_t = new (C, 1) CProjNode( iff, 0 ); | |
2165 _igvn.register_new_node_with_optimizer(if_t); | |
2166 set_loop(if_t, l); | |
2167 | |
2168 Node* cfg = NULL; // Find the One True Control User of m | |
2169 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { | |
2170 Node* x = m->fast_out(j); | |
2171 if (x->is_CFG() && x != m && x != iff) | |
2172 { cfg = x; break; } | |
2173 } | |
2174 assert(cfg != NULL, "must find the control user of m"); | |
2175 uint k = 0; // Probably cfg->in(0) | |
2176 while( cfg->in(k) != m ) k++; // But check incase cfg is a Region | |
2177 cfg->set_req( k, if_t ); // Now point to NeverBranch | |
2178 | |
2179 // Now create the never-taken loop exit | |
2180 Node *if_f = new (C, 1) CProjNode( iff, 1 ); | |
2181 _igvn.register_new_node_with_optimizer(if_f); | |
2182 set_loop(if_f, l); | |
2183 // Find frame ptr for Halt. Relies on the optimizer | |
2184 // V-N'ing. Easier and quicker than searching through | |
2185 // the program structure. | |
2186 Node *frame = new (C, 1) ParmNode( C->start(), TypeFunc::FramePtr ); | |
2187 _igvn.register_new_node_with_optimizer(frame); | |
2188 // Halt & Catch Fire | |
2189 Node *halt = new (C, TypeFunc::Parms) HaltNode( if_f, frame ); | |
2190 _igvn.register_new_node_with_optimizer(halt); | |
2191 set_loop(halt, l); | |
2192 C->root()->add_req(halt); | |
2193 set_loop(C->root(), _ltree_root); | |
2194 } | |
2195 } | |
2196 // Weeny check for irreducible. This child was already visited (this | |
2197 // IS the post-work phase). Is this child's loop header post-visited | |
2198 // as well? If so, then I found another entry into the loop. | |
2199 while( is_postvisited(l->_head) ) { | |
2200 // found irreducible | |
2201 l->_irreducible = true; | |
2202 l = l->_parent; | |
2203 _has_irreducible_loops = true; | |
2204 // Check for bad CFG here to prevent crash, and bailout of compile | |
2205 if (l == NULL) { | |
2206 C->record_method_not_compilable("unhandled CFG detected during loop optimization"); | |
2207 return pre_order; | |
2208 } | |
2209 } | |
2210 | |
2211 // This Node might be a decision point for loops. It is only if | |
2212 // it's children belong to several different loops. The sort call | |
2213 // does a trivial amount of work if there is only 1 child or all | |
2214 // children belong to the same loop. If however, the children | |
2215 // belong to different loops, the sort call will properly set the | |
2216 // _parent pointers to show how the loops nest. | |
2217 // | |
2218 // In any case, it returns the tightest enclosing loop. | |
2219 innermost = sort( l, innermost ); | |
2220 } | |
2221 | |
2222 // Def-use info will have some dead stuff; dead stuff will have no | |
2223 // loop decided on. | |
2224 | |
2225 // Am I a loop header? If so fix up my parent's child and next ptrs. | |
2226 if( innermost && innermost->_head == n ) { | |
2227 assert( get_loop(n) == innermost, "" ); | |
2228 IdealLoopTree *p = innermost->_parent; | |
2229 IdealLoopTree *l = innermost; | |
2230 while( p && l->_head == n ) { | |
2231 l->_next = p->_child; // Put self on parents 'next child' | |
2232 p->_child = l; // Make self as first child of parent | |
2233 l = p; // Now walk up the parent chain | |
2234 p = l->_parent; | |
2235 } | |
2236 } else { | |
2237 // Note that it is possible for a LoopNode to reach here, if the | |
2238 // backedge has been made unreachable (hence the LoopNode no longer | |
2239 // denotes a Loop, and will eventually be removed). | |
2240 | |
2241 // Record tightest enclosing loop for self. Mark as post-visited. | |
2242 set_loop(n, innermost); | |
2243 // Also record has_call flag early on | |
2244 if( innermost ) { | |
2245 if( n->is_Call() && !n->is_CallLeaf() && !n->is_macro() ) { | |
2246 // Do not count uncommon calls | |
2247 if( !n->is_CallStaticJava() || !n->as_CallStaticJava()->_name ) { | |
2248 Node *iff = n->in(0)->in(0); | |
2249 if( !iff->is_If() || | |
2250 (n->in(0)->Opcode() == Op_IfFalse && | |
2251 (1.0 - iff->as_If()->_prob) >= 0.01) || | |
2252 (iff->as_If()->_prob >= 0.01) ) | |
2253 innermost->_has_call = 1; | |
2254 } | |
2255 } | |
2256 } | |
2257 } | |
2258 | |
2259 // Flag as post-visited now | |
2260 set_postvisited(n); | |
2261 return pre_order; | |
2262 } | |
2263 | |
2264 | |
2265 //------------------------------build_loop_early------------------------------- | |
2266 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping. | |
2267 // First pass computes the earliest controlling node possible. This is the | |
2268 // controlling input with the deepest dominating depth. | |
2269 void PhaseIdealLoop::build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack, const PhaseIdealLoop *verify_me ) { | |
2270 while (worklist.size() != 0) { | |
2271 // Use local variables nstack_top_n & nstack_top_i to cache values | |
2272 // on nstack's top. | |
2273 Node *nstack_top_n = worklist.pop(); | |
2274 uint nstack_top_i = 0; | |
2275 //while_nstack_nonempty: | |
2276 while (true) { | |
2277 // Get parent node and next input's index from stack's top. | |
2278 Node *n = nstack_top_n; | |
2279 uint i = nstack_top_i; | |
2280 uint cnt = n->req(); // Count of inputs | |
2281 if (i == 0) { // Pre-process the node. | |
2282 if( has_node(n) && // Have either loop or control already? | |
2283 !has_ctrl(n) ) { // Have loop picked out already? | |
2284 // During "merge_many_backedges" we fold up several nested loops | |
2285 // into a single loop. This makes the members of the original | |
2286 // loop bodies pointing to dead loops; they need to move up | |
2287 // to the new UNION'd larger loop. I set the _head field of these | |
2288 // dead loops to NULL and the _parent field points to the owning | |
2289 // loop. Shades of UNION-FIND algorithm. | |
2290 IdealLoopTree *ilt; | |
2291 while( !(ilt = get_loop(n))->_head ) { | |
2292 // Normally I would use a set_loop here. But in this one special | |
2293 // case, it is legal (and expected) to change what loop a Node | |
2294 // belongs to. | |
2295 _nodes.map(n->_idx, (Node*)(ilt->_parent) ); | |
2296 } | |
2297 // Remove safepoints ONLY if I've already seen I don't need one. | |
2298 // (the old code here would yank a 2nd safepoint after seeing a | |
2299 // first one, even though the 1st did not dominate in the loop body | |
2300 // and thus could be avoided indefinitely) | |
2301 if( !verify_me && ilt->_has_sfpt && n->Opcode() == Op_SafePoint && | |
2302 is_deleteable_safept(n)) { | |
2303 Node *in = n->in(TypeFunc::Control); | |
2304 lazy_replace(n,in); // Pull safepoint now | |
2305 // Carry on with the recursion "as if" we are walking | |
2306 // only the control input | |
2307 if( !visited.test_set( in->_idx ) ) { | |
2308 worklist.push(in); // Visit this guy later, using worklist | |
2309 } | |
2310 // Get next node from nstack: | |
2311 // - skip n's inputs processing by setting i > cnt; | |
2312 // - we also will not call set_early_ctrl(n) since | |
2313 // has_node(n) == true (see the condition above). | |
2314 i = cnt + 1; | |
2315 } | |
2316 } | |
2317 } // if (i == 0) | |
2318 | |
2319 // Visit all inputs | |
2320 bool done = true; // Assume all n's inputs will be processed | |
2321 while (i < cnt) { | |
2322 Node *in = n->in(i); | |
2323 ++i; | |
2324 if (in == NULL) continue; | |
2325 if (in->pinned() && !in->is_CFG()) | |
2326 set_ctrl(in, in->in(0)); | |
2327 int is_visited = visited.test_set( in->_idx ); | |
2328 if (!has_node(in)) { // No controlling input yet? | |
2329 assert( !in->is_CFG(), "CFG Node with no controlling input?" ); | |
2330 assert( !is_visited, "visit only once" ); | |
2331 nstack.push(n, i); // Save parent node and next input's index. | |
2332 nstack_top_n = in; // Process current input now. | |
2333 nstack_top_i = 0; | |
2334 done = false; // Not all n's inputs processed. | |
2335 break; // continue while_nstack_nonempty; | |
2336 } else if (!is_visited) { | |
2337 // This guy has a location picked out for him, but has not yet | |
2338 // been visited. Happens to all CFG nodes, for instance. | |
2339 // Visit him using the worklist instead of recursion, to break | |
2340 // cycles. Since he has a location already we do not need to | |
2341 // find his location before proceeding with the current Node. | |
2342 worklist.push(in); // Visit this guy later, using worklist | |
2343 } | |
2344 } | |
2345 if (done) { | |
2346 // All of n's inputs have been processed, complete post-processing. | |
2347 | |
2348 // Compute earilest point this Node can go. | |
2349 // CFG, Phi, pinned nodes already know their controlling input. | |
2350 if (!has_node(n)) { | |
2351 // Record earliest legal location | |
2352 set_early_ctrl( n ); | |
2353 } | |
2354 if (nstack.is_empty()) { | |
2355 // Finished all nodes on stack. | |
2356 // Process next node on the worklist. | |
2357 break; | |
2358 } | |
2359 // Get saved parent node and next input's index. | |
2360 nstack_top_n = nstack.node(); | |
2361 nstack_top_i = nstack.index(); | |
2362 nstack.pop(); | |
2363 } | |
2364 } // while (true) | |
2365 } | |
2366 } | |
2367 | |
2368 //------------------------------dom_lca_internal-------------------------------- | |
2369 // Pair-wise LCA | |
2370 Node *PhaseIdealLoop::dom_lca_internal( Node *n1, Node *n2 ) const { | |
2371 if( !n1 ) return n2; // Handle NULL original LCA | |
2372 assert( n1->is_CFG(), "" ); | |
2373 assert( n2->is_CFG(), "" ); | |
2374 // find LCA of all uses | |
2375 uint d1 = dom_depth(n1); | |
2376 uint d2 = dom_depth(n2); | |
2377 while (n1 != n2) { | |
2378 if (d1 > d2) { | |
2379 n1 = idom(n1); | |
2380 d1 = dom_depth(n1); | |
2381 } else if (d1 < d2) { | |
2382 n2 = idom(n2); | |
2383 d2 = dom_depth(n2); | |
2384 } else { | |
2385 // Here d1 == d2. Due to edits of the dominator-tree, sections | |
2386 // of the tree might have the same depth. These sections have | |
2387 // to be searched more carefully. | |
2388 | |
2389 // Scan up all the n1's with equal depth, looking for n2. | |
2390 Node *t1 = idom(n1); | |
2391 while (dom_depth(t1) == d1) { | |
2392 if (t1 == n2) return n2; | |
2393 t1 = idom(t1); | |
2394 } | |
2395 // Scan up all the n2's with equal depth, looking for n1. | |
2396 Node *t2 = idom(n2); | |
2397 while (dom_depth(t2) == d2) { | |
2398 if (t2 == n1) return n1; | |
2399 t2 = idom(t2); | |
2400 } | |
2401 // Move up to a new dominator-depth value as well as up the dom-tree. | |
2402 n1 = t1; | |
2403 n2 = t2; | |
2404 d1 = dom_depth(n1); | |
2405 d2 = dom_depth(n2); | |
2406 } | |
2407 } | |
2408 return n1; | |
2409 } | |
2410 | |
2411 //------------------------------compute_idom----------------------------------- | |
2412 // Locally compute IDOM using dom_lca call. Correct only if the incoming | |
2413 // IDOMs are correct. | |
2414 Node *PhaseIdealLoop::compute_idom( Node *region ) const { | |
2415 assert( region->is_Region(), "" ); | |
2416 Node *LCA = NULL; | |
2417 for( uint i = 1; i < region->req(); i++ ) { | |
2418 if( region->in(i) != C->top() ) | |
2419 LCA = dom_lca( LCA, region->in(i) ); | |
2420 } | |
2421 return LCA; | |
2422 } | |
2423 | |
2424 //------------------------------get_late_ctrl---------------------------------- | |
2425 // Compute latest legal control. | |
2426 Node *PhaseIdealLoop::get_late_ctrl( Node *n, Node *early ) { | |
2427 assert(early != NULL, "early control should not be NULL"); | |
2428 | |
2429 // Compute LCA over list of uses | |
2430 Node *LCA = NULL; | |
2431 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && LCA != early; i++) { | |
2432 Node* c = n->fast_out(i); | |
2433 if (_nodes[c->_idx] == NULL) | |
2434 continue; // Skip the occasional dead node | |
2435 if( c->is_Phi() ) { // For Phis, we must land above on the path | |
2436 for( uint j=1; j<c->req(); j++ ) {// For all inputs | |
2437 if( c->in(j) == n ) { // Found matching input? | |
2438 Node *use = c->in(0)->in(j); | |
2439 LCA = dom_lca_for_get_late_ctrl( LCA, use, n ); | |
2440 } | |
2441 } | |
2442 } else { | |
2443 // For CFG data-users, use is in the block just prior | |
2444 Node *use = has_ctrl(c) ? get_ctrl(c) : c->in(0); | |
2445 LCA = dom_lca_for_get_late_ctrl( LCA, use, n ); | |
2446 } | |
2447 } | |
2448 | |
2449 // if this is a load, check for anti-dependent stores | |
2450 // We use a conservative algorithm to identify potential interfering | |
2451 // instructions and for rescheduling the load. The users of the memory | |
2452 // input of this load are examined. Any use which is not a load and is | |
2453 // dominated by early is considered a potentially interfering store. | |
2454 // This can produce false positives. | |
2455 if (n->is_Load() && LCA != early) { | |
2456 Node_List worklist; | |
2457 | |
2458 Node *mem = n->in(MemNode::Memory); | |
2459 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) { | |
2460 Node* s = mem->fast_out(i); | |
2461 worklist.push(s); | |
2462 } | |
2463 while(worklist.size() != 0 && LCA != early) { | |
2464 Node* s = worklist.pop(); | |
2465 if (s->is_Load()) { | |
2466 continue; | |
2467 } else if (s->is_MergeMem()) { | |
2468 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) { | |
2469 Node* s1 = s->fast_out(i); | |
2470 worklist.push(s1); | |
2471 } | |
2472 } else { | |
2473 Node *sctrl = has_ctrl(s) ? get_ctrl(s) : s->in(0); | |
2474 assert(sctrl != NULL || s->outcnt() == 0, "must have control"); | |
2475 if (sctrl != NULL && !sctrl->is_top() && is_dominator(early, sctrl)) { | |
2476 LCA = dom_lca_for_get_late_ctrl(LCA, sctrl, n); | |
2477 } | |
2478 } | |
2479 } | |
2480 } | |
2481 | |
2482 assert(LCA == find_non_split_ctrl(LCA), "unexpected late control"); | |
2483 return LCA; | |
2484 } | |
2485 | |
2486 // true if CFG node d dominates CFG node n | |
2487 bool PhaseIdealLoop::is_dominator(Node *d, Node *n) { | |
2488 if (d == n) | |
2489 return true; | |
2490 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes"); | |
2491 uint dd = dom_depth(d); | |
2492 while (dom_depth(n) >= dd) { | |
2493 if (n == d) | |
2494 return true; | |
2495 n = idom(n); | |
2496 } | |
2497 return false; | |
2498 } | |
2499 | |
2500 //------------------------------dom_lca_for_get_late_ctrl_internal------------- | |
2501 // Pair-wise LCA with tags. | |
2502 // Tag each index with the node 'tag' currently being processed | |
2503 // before advancing up the dominator chain using idom(). | |
2504 // Later calls that find a match to 'tag' know that this path has already | |
2505 // been considered in the current LCA (which is input 'n1' by convention). | |
2506 // Since get_late_ctrl() is only called once for each node, the tag array | |
2507 // does not need to be cleared between calls to get_late_ctrl(). | |
2508 // Algorithm trades a larger constant factor for better asymptotic behavior | |
2509 // | |
2510 Node *PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal( Node *n1, Node *n2, Node *tag ) { | |
2511 uint d1 = dom_depth(n1); | |
2512 uint d2 = dom_depth(n2); | |
2513 | |
2514 do { | |
2515 if (d1 > d2) { | |
2516 // current lca is deeper than n2 | |
2517 _dom_lca_tags.map(n1->_idx, tag); | |
2518 n1 = idom(n1); | |
2519 d1 = dom_depth(n1); | |
2520 } else if (d1 < d2) { | |
2521 // n2 is deeper than current lca | |
2522 Node *memo = _dom_lca_tags[n2->_idx]; | |
2523 if( memo == tag ) { | |
2524 return n1; // Return the current LCA | |
2525 } | |
2526 _dom_lca_tags.map(n2->_idx, tag); | |
2527 n2 = idom(n2); | |
2528 d2 = dom_depth(n2); | |
2529 } else { | |
2530 // Here d1 == d2. Due to edits of the dominator-tree, sections | |
2531 // of the tree might have the same depth. These sections have | |
2532 // to be searched more carefully. | |
2533 | |
2534 // Scan up all the n1's with equal depth, looking for n2. | |
2535 _dom_lca_tags.map(n1->_idx, tag); | |
2536 Node *t1 = idom(n1); | |
2537 while (dom_depth(t1) == d1) { | |
2538 if (t1 == n2) return n2; | |
2539 _dom_lca_tags.map(t1->_idx, tag); | |
2540 t1 = idom(t1); | |
2541 } | |
2542 // Scan up all the n2's with equal depth, looking for n1. | |
2543 _dom_lca_tags.map(n2->_idx, tag); | |
2544 Node *t2 = idom(n2); | |
2545 while (dom_depth(t2) == d2) { | |
2546 if (t2 == n1) return n1; | |
2547 _dom_lca_tags.map(t2->_idx, tag); | |
2548 t2 = idom(t2); | |
2549 } | |
2550 // Move up to a new dominator-depth value as well as up the dom-tree. | |
2551 n1 = t1; | |
2552 n2 = t2; | |
2553 d1 = dom_depth(n1); | |
2554 d2 = dom_depth(n2); | |
2555 } | |
2556 } while (n1 != n2); | |
2557 return n1; | |
2558 } | |
2559 | |
2560 //------------------------------init_dom_lca_tags------------------------------ | |
2561 // Tag could be a node's integer index, 32bits instead of 64bits in some cases | |
2562 // Intended use does not involve any growth for the array, so it could | |
2563 // be of fixed size. | |
2564 void PhaseIdealLoop::init_dom_lca_tags() { | |
2565 uint limit = C->unique() + 1; | |
2566 _dom_lca_tags.map( limit, NULL ); | |
2567 #ifdef ASSERT | |
2568 for( uint i = 0; i < limit; ++i ) { | |
2569 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer"); | |
2570 } | |
2571 #endif // ASSERT | |
2572 } | |
2573 | |
2574 //------------------------------clear_dom_lca_tags------------------------------ | |
2575 // Tag could be a node's integer index, 32bits instead of 64bits in some cases | |
2576 // Intended use does not involve any growth for the array, so it could | |
2577 // be of fixed size. | |
2578 void PhaseIdealLoop::clear_dom_lca_tags() { | |
2579 uint limit = C->unique() + 1; | |
2580 _dom_lca_tags.map( limit, NULL ); | |
2581 _dom_lca_tags.clear(); | |
2582 #ifdef ASSERT | |
2583 for( uint i = 0; i < limit; ++i ) { | |
2584 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer"); | |
2585 } | |
2586 #endif // ASSERT | |
2587 } | |
2588 | |
2589 //------------------------------build_loop_late-------------------------------- | |
2590 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping. | |
2591 // Second pass finds latest legal placement, and ideal loop placement. | |
2592 void PhaseIdealLoop::build_loop_late( VectorSet &visited, Node_List &worklist, Node_Stack &nstack, const PhaseIdealLoop *verify_me ) { | |
2593 while (worklist.size() != 0) { | |
2594 Node *n = worklist.pop(); | |
2595 // Only visit once | |
2596 if (visited.test_set(n->_idx)) continue; | |
2597 uint cnt = n->outcnt(); | |
2598 uint i = 0; | |
2599 while (true) { | |
2600 assert( _nodes[n->_idx], "no dead nodes" ); | |
2601 // Visit all children | |
2602 if (i < cnt) { | |
2603 Node* use = n->raw_out(i); | |
2604 ++i; | |
2605 // Check for dead uses. Aggressively prune such junk. It might be | |
2606 // dead in the global sense, but still have local uses so I cannot | |
2607 // easily call 'remove_dead_node'. | |
2608 if( _nodes[use->_idx] != NULL || use->is_top() ) { // Not dead? | |
2609 // Due to cycles, we might not hit the same fixed point in the verify | |
2610 // pass as we do in the regular pass. Instead, visit such phis as | |
2611 // simple uses of the loop head. | |
2612 if( use->in(0) && (use->is_CFG() || use->is_Phi()) ) { | |
2613 if( !visited.test(use->_idx) ) | |
2614 worklist.push(use); | |
2615 } else if( !visited.test_set(use->_idx) ) { | |
2616 nstack.push(n, i); // Save parent and next use's index. | |
2617 n = use; // Process all children of current use. | |
2618 cnt = use->outcnt(); | |
2619 i = 0; | |
2620 } | |
2621 } else { | |
2622 // Do not visit around the backedge of loops via data edges. | |
2623 // push dead code onto a worklist | |
2624 _deadlist.push(use); | |
2625 } | |
2626 } else { | |
2627 // All of n's children have been processed, complete post-processing. | |
2628 build_loop_late_post(n, verify_me); | |
2629 if (nstack.is_empty()) { | |
2630 // Finished all nodes on stack. | |
2631 // Process next node on the worklist. | |
2632 break; | |
2633 } | |
2634 // Get saved parent node and next use's index. Visit the rest of uses. | |
2635 n = nstack.node(); | |
2636 cnt = n->outcnt(); | |
2637 i = nstack.index(); | |
2638 nstack.pop(); | |
2639 } | |
2640 } | |
2641 } | |
2642 } | |
2643 | |
2644 //------------------------------build_loop_late_post--------------------------- | |
2645 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping. | |
2646 // Second pass finds latest legal placement, and ideal loop placement. | |
2647 void PhaseIdealLoop::build_loop_late_post( Node *n, const PhaseIdealLoop *verify_me ) { | |
2648 | |
2649 if (n->req() == 2 && n->Opcode() == Op_ConvI2L && !C->major_progress()) { | |
2650 _igvn._worklist.push(n); // Maybe we'll normalize it, if no more loops. | |
2651 } | |
2652 | |
2653 // CFG and pinned nodes already handled | |
2654 if( n->in(0) ) { | |
2655 if( n->in(0)->is_top() ) return; // Dead? | |
2656 | |
2657 // We'd like +VerifyLoopOptimizations to not believe that Mod's/Loads | |
2658 // _must_ be pinned (they have to observe their control edge of course). | |
2659 // Unlike Stores (which modify an unallocable resource, the memory | |
2660 // state), Mods/Loads can float around. So free them up. | |
2661 bool pinned = true; | |
2662 switch( n->Opcode() ) { | |
2663 case Op_DivI: | |
2664 case Op_DivF: | |
2665 case Op_DivD: | |
2666 case Op_ModI: | |
2667 case Op_ModF: | |
2668 case Op_ModD: | |
2669 case Op_LoadB: // Same with Loads; they can sink | |
2670 case Op_LoadC: // during loop optimizations. | |
2671 case Op_LoadD: | |
2672 case Op_LoadF: | |
2673 case Op_LoadI: | |
2674 case Op_LoadKlass: | |
2675 case Op_LoadL: | |
2676 case Op_LoadS: | |
2677 case Op_LoadP: | |
2678 case Op_LoadRange: | |
2679 case Op_LoadD_unaligned: | |
2680 case Op_LoadL_unaligned: | |
2681 case Op_StrComp: // Does a bunch of load-like effects | |
2682 pinned = false; | |
2683 } | |
2684 if( pinned ) { | |
2685 IdealLoopTree *choosen_loop = get_loop(n->is_CFG() ? n : get_ctrl(n)); | |
2686 if( !choosen_loop->_child ) // Inner loop? | |
2687 choosen_loop->_body.push(n); // Collect inner loops | |
2688 return; | |
2689 } | |
2690 } else { // No slot zero | |
2691 if( n->is_CFG() ) { // CFG with no slot 0 is dead | |
2692 _nodes.map(n->_idx,0); // No block setting, it's globally dead | |
2693 return; | |
2694 } | |
2695 assert(!n->is_CFG() || n->outcnt() == 0, ""); | |
2696 } | |
2697 | |
2698 // Do I have a "safe range" I can select over? | |
2699 Node *early = get_ctrl(n);// Early location already computed | |
2700 | |
2701 // Compute latest point this Node can go | |
2702 Node *LCA = get_late_ctrl( n, early ); | |
2703 // LCA is NULL due to uses being dead | |
2704 if( LCA == NULL ) { | |
2705 #ifdef ASSERT | |
2706 for (DUIterator i1 = n->outs(); n->has_out(i1); i1++) { | |
2707 assert( _nodes[n->out(i1)->_idx] == NULL, "all uses must also be dead"); | |
2708 } | |
2709 #endif | |
2710 _nodes.map(n->_idx, 0); // This node is useless | |
2711 _deadlist.push(n); | |
2712 return; | |
2713 } | |
2714 assert(LCA != NULL && !LCA->is_top(), "no dead nodes"); | |
2715 | |
2716 Node *legal = LCA; // Walk 'legal' up the IDOM chain | |
2717 Node *least = legal; // Best legal position so far | |
2718 while( early != legal ) { // While not at earliest legal | |
2719 // Find least loop nesting depth | |
2720 legal = idom(legal); // Bump up the IDOM tree | |
2721 // Check for lower nesting depth | |
2722 if( get_loop(legal)->_nest < get_loop(least)->_nest ) | |
2723 least = legal; | |
2724 } | |
2725 | |
2726 // Try not to place code on a loop entry projection | |
2727 // which can inhibit range check elimination. | |
2728 if (least != early) { | |
2729 Node* ctrl_out = least->unique_ctrl_out(); | |
2730 if (ctrl_out && ctrl_out->is_CountedLoop() && | |
2731 least == ctrl_out->in(LoopNode::EntryControl)) { | |
2732 Node* least_dom = idom(least); | |
2733 if (get_loop(least_dom)->is_member(get_loop(least))) { | |
2734 least = least_dom; | |
2735 } | |
2736 } | |
2737 } | |
2738 | |
2739 #ifdef ASSERT | |
2740 // If verifying, verify that 'verify_me' has a legal location | |
2741 // and choose it as our location. | |
2742 if( verify_me ) { | |
2743 Node *v_ctrl = verify_me->get_ctrl_no_update(n); | |
2744 Node *legal = LCA; | |
2745 while( early != legal ) { // While not at earliest legal | |
2746 if( legal == v_ctrl ) break; // Check for prior good location | |
2747 legal = idom(legal) ;// Bump up the IDOM tree | |
2748 } | |
2749 // Check for prior good location | |
2750 if( legal == v_ctrl ) least = legal; // Keep prior if found | |
2751 } | |
2752 #endif | |
2753 | |
2754 // Assign discovered "here or above" point | |
2755 least = find_non_split_ctrl(least); | |
2756 set_ctrl(n, least); | |
2757 | |
2758 // Collect inner loop bodies | |
2759 IdealLoopTree *choosen_loop = get_loop(least); | |
2760 if( !choosen_loop->_child ) // Inner loop? | |
2761 choosen_loop->_body.push(n);// Collect inner loops | |
2762 } | |
2763 | |
2764 #ifndef PRODUCT | |
2765 //------------------------------dump------------------------------------------- | |
2766 void PhaseIdealLoop::dump( ) const { | |
2767 ResourceMark rm; | |
2768 Arena* arena = Thread::current()->resource_area(); | |
2769 Node_Stack stack(arena, C->unique() >> 2); | |
2770 Node_List rpo_list; | |
2771 VectorSet visited(arena); | |
2772 visited.set(C->top()->_idx); | |
2773 rpo( C->root(), stack, visited, rpo_list ); | |
2774 // Dump root loop indexed by last element in PO order | |
2775 dump( _ltree_root, rpo_list.size(), rpo_list ); | |
2776 } | |
2777 | |
2778 void PhaseIdealLoop::dump( IdealLoopTree *loop, uint idx, Node_List &rpo_list ) const { | |
2779 | |
2780 // Indent by loop nesting depth | |
2781 for( uint x = 0; x < loop->_nest; x++ ) | |
2782 tty->print(" "); | |
2783 tty->print_cr("---- Loop N%d-N%d ----", loop->_head->_idx,loop->_tail->_idx); | |
2784 | |
2785 // Now scan for CFG nodes in the same loop | |
2786 for( uint j=idx; j > 0; j-- ) { | |
2787 Node *n = rpo_list[j-1]; | |
2788 if( !_nodes[n->_idx] ) // Skip dead nodes | |
2789 continue; | |
2790 if( get_loop(n) != loop ) { // Wrong loop nest | |
2791 if( get_loop(n)->_head == n && // Found nested loop? | |
2792 get_loop(n)->_parent == loop ) | |
2793 dump(get_loop(n),rpo_list.size(),rpo_list); // Print it nested-ly | |
2794 continue; | |
2795 } | |
2796 | |
2797 // Dump controlling node | |
2798 for( uint x = 0; x < loop->_nest; x++ ) | |
2799 tty->print(" "); | |
2800 tty->print("C"); | |
2801 if( n == C->root() ) { | |
2802 n->dump(); | |
2803 } else { | |
2804 Node* cached_idom = idom_no_update(n); | |
2805 Node *computed_idom = n->in(0); | |
2806 if( n->is_Region() ) { | |
2807 computed_idom = compute_idom(n); | |
2808 // computed_idom() will return n->in(0) when idom(n) is an IfNode (or | |
2809 // any MultiBranch ctrl node), so apply a similar transform to | |
2810 // the cached idom returned from idom_no_update. | |
2811 cached_idom = find_non_split_ctrl(cached_idom); | |
2812 } | |
2813 tty->print(" ID:%d",computed_idom->_idx); | |
2814 n->dump(); | |
2815 if( cached_idom != computed_idom ) { | |
2816 tty->print_cr("*** BROKEN IDOM! Computed as: %d, cached as: %d", | |
2817 computed_idom->_idx, cached_idom->_idx); | |
2818 } | |
2819 } | |
2820 // Dump nodes it controls | |
2821 for( uint k = 0; k < _nodes.Size(); k++ ) { | |
2822 // (k < C->unique() && get_ctrl(find(k)) == n) | |
2823 if (k < C->unique() && _nodes[k] == (Node*)((intptr_t)n + 1)) { | |
2824 Node *m = C->root()->find(k); | |
2825 if( m && m->outcnt() > 0 ) { | |
2826 if (!(has_ctrl(m) && get_ctrl_no_update(m) == n)) { | |
2827 tty->print_cr("*** BROKEN CTRL ACCESSOR! _nodes[k] is %p, ctrl is %p", | |
2828 _nodes[k], has_ctrl(m) ? get_ctrl_no_update(m) : NULL); | |
2829 } | |
2830 for( uint j = 0; j < loop->_nest; j++ ) | |
2831 tty->print(" "); | |
2832 tty->print(" "); | |
2833 m->dump(); | |
2834 } | |
2835 } | |
2836 } | |
2837 } | |
2838 } | |
2839 | |
2840 // Collect a R-P-O for the whole CFG. | |
2841 // Result list is in post-order (scan backwards for RPO) | |
2842 void PhaseIdealLoop::rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const { | |
2843 stk.push(start, 0); | |
2844 visited.set(start->_idx); | |
2845 | |
2846 while (stk.is_nonempty()) { | |
2847 Node* m = stk.node(); | |
2848 uint idx = stk.index(); | |
2849 if (idx < m->outcnt()) { | |
2850 stk.set_index(idx + 1); | |
2851 Node* n = m->raw_out(idx); | |
2852 if (n->is_CFG() && !visited.test_set(n->_idx)) { | |
2853 stk.push(n, 0); | |
2854 } | |
2855 } else { | |
2856 rpo_list.push(m); | |
2857 stk.pop(); | |
2858 } | |
2859 } | |
2860 } | |
2861 #endif | |
2862 | |
2863 | |
2864 //============================================================================= | |
2865 //------------------------------LoopTreeIterator----------------------------------- | |
2866 | |
2867 // Advance to next loop tree using a preorder, left-to-right traversal. | |
2868 void LoopTreeIterator::next() { | |
2869 assert(!done(), "must not be done."); | |
2870 if (_curnt->_child != NULL) { | |
2871 _curnt = _curnt->_child; | |
2872 } else if (_curnt->_next != NULL) { | |
2873 _curnt = _curnt->_next; | |
2874 } else { | |
2875 while (_curnt != _root && _curnt->_next == NULL) { | |
2876 _curnt = _curnt->_parent; | |
2877 } | |
2878 if (_curnt == _root) { | |
2879 _curnt = NULL; | |
2880 assert(done(), "must be done."); | |
2881 } else { | |
2882 assert(_curnt->_next != NULL, "must be more to do"); | |
2883 _curnt = _curnt->_next; | |
2884 } | |
2885 } | |
2886 } |