1 | /* |
2 | * Copyright (c) 1997, 2018, Oracle and/or its affiliates. All rights reserved. |
3 | * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
4 | * |
5 | * This code is free software; you can redistribute it and/or modify it |
6 | * under the terms of the GNU General Public License version 2 only, as |
7 | * published by the Free Software Foundation. |
8 | * |
9 | * This code is distributed in the hope that it will be useful, but WITHOUT |
10 | * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or |
11 | * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
12 | * version 2 for more details (a copy is included in the LICENSE file that |
13 | * accompanied this code). |
14 | * |
15 | * You should have received a copy of the GNU General Public License version |
16 | * 2 along with this work; if not, write to the Free Software Foundation, |
17 | * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. |
18 | * |
19 | * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
20 | * or visit www.oracle.com if you need additional information or have any |
21 | * questions. |
22 | * |
23 | */ |
24 | |
25 | #include "precompiled.hpp" |
26 | #include "classfile/systemDictionary.hpp" |
27 | #include "gc/shared/barrierSet.hpp" |
28 | #include "gc/shared/c2/barrierSetC2.hpp" |
29 | #include "memory/allocation.inline.hpp" |
30 | #include "memory/resourceArea.hpp" |
31 | #include "oops/objArrayKlass.hpp" |
32 | #include "opto/addnode.hpp" |
33 | #include "opto/castnode.hpp" |
34 | #include "opto/cfgnode.hpp" |
35 | #include "opto/connode.hpp" |
36 | #include "opto/convertnode.hpp" |
37 | #include "opto/loopnode.hpp" |
38 | #include "opto/machnode.hpp" |
39 | #include "opto/movenode.hpp" |
40 | #include "opto/narrowptrnode.hpp" |
41 | #include "opto/mulnode.hpp" |
42 | #include "opto/phaseX.hpp" |
43 | #include "opto/regmask.hpp" |
44 | #include "opto/runtime.hpp" |
45 | #include "opto/subnode.hpp" |
46 | #include "utilities/vmError.hpp" |
47 | |
48 | // Portions of code courtesy of Clifford Click |
49 | |
50 | // Optimization - Graph Style |
51 | |
52 | //============================================================================= |
53 | //------------------------------Value------------------------------------------ |
54 | // Compute the type of the RegionNode. |
55 | const Type* RegionNode::Value(PhaseGVN* phase) const { |
56 | for( uint i=1; i<req(); ++i ) { // For all paths in |
57 | Node *n = in(i); // Get Control source |
58 | if( !n ) continue; // Missing inputs are TOP |
59 | if( phase->type(n) == Type::CONTROL ) |
60 | return Type::CONTROL; |
61 | } |
62 | return Type::TOP; // All paths dead? Then so are we |
63 | } |
64 | |
65 | //------------------------------Identity--------------------------------------- |
66 | // Check for Region being Identity. |
67 | Node* RegionNode::Identity(PhaseGVN* phase) { |
68 | // Cannot have Region be an identity, even if it has only 1 input. |
69 | // Phi users cannot have their Region input folded away for them, |
70 | // since they need to select the proper data input |
71 | return this; |
72 | } |
73 | |
74 | //------------------------------merge_region----------------------------------- |
75 | // If a Region flows into a Region, merge into one big happy merge. This is |
76 | // hard to do if there is stuff that has to happen |
77 | static Node *merge_region(RegionNode *region, PhaseGVN *phase) { |
78 | if( region->Opcode() != Op_Region ) // Do not do to LoopNodes |
79 | return NULL; |
80 | Node *progress = NULL; // Progress flag |
81 | PhaseIterGVN *igvn = phase->is_IterGVN(); |
82 | |
83 | uint rreq = region->req(); |
84 | for( uint i = 1; i < rreq; i++ ) { |
85 | Node *r = region->in(i); |
86 | if( r && r->Opcode() == Op_Region && // Found a region? |
87 | r->in(0) == r && // Not already collapsed? |
88 | r != region && // Avoid stupid situations |
89 | r->outcnt() == 2 ) { // Self user and 'region' user only? |
90 | assert(!r->as_Region()->has_phi(), "no phi users" ); |
91 | if( !progress ) { // No progress |
92 | if (region->has_phi()) { |
93 | return NULL; // Only flatten if no Phi users |
94 | // igvn->hash_delete( phi ); |
95 | } |
96 | igvn->hash_delete( region ); |
97 | progress = region; // Making progress |
98 | } |
99 | igvn->hash_delete( r ); |
100 | |
101 | // Append inputs to 'r' onto 'region' |
102 | for( uint j = 1; j < r->req(); j++ ) { |
103 | // Move an input from 'r' to 'region' |
104 | region->add_req(r->in(j)); |
105 | r->set_req(j, phase->C->top()); |
106 | // Update phis of 'region' |
107 | //for( uint k = 0; k < max; k++ ) { |
108 | // Node *phi = region->out(k); |
109 | // if( phi->is_Phi() ) { |
110 | // phi->add_req(phi->in(i)); |
111 | // } |
112 | //} |
113 | |
114 | rreq++; // One more input to Region |
115 | } // Found a region to merge into Region |
116 | igvn->_worklist.push(r); |
117 | // Clobber pointer to the now dead 'r' |
118 | region->set_req(i, phase->C->top()); |
119 | } |
120 | } |
121 | |
122 | return progress; |
123 | } |
124 | |
125 | |
126 | |
127 | //--------------------------------has_phi-------------------------------------- |
128 | // Helper function: Return any PhiNode that uses this region or NULL |
129 | PhiNode* RegionNode::has_phi() const { |
130 | for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) { |
131 | Node* phi = fast_out(i); |
132 | if (phi->is_Phi()) { // Check for Phi users |
133 | assert(phi->in(0) == (Node*)this, "phi uses region only via in(0)" ); |
134 | return phi->as_Phi(); // this one is good enough |
135 | } |
136 | } |
137 | |
138 | return NULL; |
139 | } |
140 | |
141 | |
142 | //-----------------------------has_unique_phi---------------------------------- |
143 | // Helper function: Return the only PhiNode that uses this region or NULL |
144 | PhiNode* RegionNode::has_unique_phi() const { |
145 | // Check that only one use is a Phi |
146 | PhiNode* only_phi = NULL; |
147 | for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) { |
148 | Node* phi = fast_out(i); |
149 | if (phi->is_Phi()) { // Check for Phi users |
150 | assert(phi->in(0) == (Node*)this, "phi uses region only via in(0)" ); |
151 | if (only_phi == NULL) { |
152 | only_phi = phi->as_Phi(); |
153 | } else { |
154 | return NULL; // multiple phis |
155 | } |
156 | } |
157 | } |
158 | |
159 | return only_phi; |
160 | } |
161 | |
162 | |
163 | //------------------------------check_phi_clipping----------------------------- |
164 | // Helper function for RegionNode's identification of FP clipping |
165 | // Check inputs to the Phi |
166 | static bool check_phi_clipping( PhiNode *phi, ConNode * &min, uint &min_idx, ConNode * &max, uint &max_idx, Node * &val, uint &val_idx ) { |
167 | min = NULL; |
168 | max = NULL; |
169 | val = NULL; |
170 | min_idx = 0; |
171 | max_idx = 0; |
172 | val_idx = 0; |
173 | uint phi_max = phi->req(); |
174 | if( phi_max == 4 ) { |
175 | for( uint j = 1; j < phi_max; ++j ) { |
176 | Node *n = phi->in(j); |
177 | int opcode = n->Opcode(); |
178 | switch( opcode ) { |
179 | case Op_ConI: |
180 | { |
181 | if( min == NULL ) { |
182 | min = n->Opcode() == Op_ConI ? (ConNode*)n : NULL; |
183 | min_idx = j; |
184 | } else { |
185 | max = n->Opcode() == Op_ConI ? (ConNode*)n : NULL; |
186 | max_idx = j; |
187 | if( min->get_int() > max->get_int() ) { |
188 | // Swap min and max |
189 | ConNode *temp; |
190 | uint temp_idx; |
191 | temp = min; min = max; max = temp; |
192 | temp_idx = min_idx; min_idx = max_idx; max_idx = temp_idx; |
193 | } |
194 | } |
195 | } |
196 | break; |
197 | default: |
198 | { |
199 | val = n; |
200 | val_idx = j; |
201 | } |
202 | break; |
203 | } |
204 | } |
205 | } |
206 | return ( min && max && val && (min->get_int() <= 0) && (max->get_int() >=0) ); |
207 | } |
208 | |
209 | |
210 | //------------------------------check_if_clipping------------------------------ |
211 | // Helper function for RegionNode's identification of FP clipping |
212 | // Check that inputs to Region come from two IfNodes, |
213 | // |
214 | // If |
215 | // False True |
216 | // If | |
217 | // False True | |
218 | // | | | |
219 | // RegionNode_inputs |
220 | // |
221 | static bool check_if_clipping( const RegionNode *region, IfNode * &bot_if, IfNode * &top_if ) { |
222 | top_if = NULL; |
223 | bot_if = NULL; |
224 | |
225 | // Check control structure above RegionNode for (if ( if ) ) |
226 | Node *in1 = region->in(1); |
227 | Node *in2 = region->in(2); |
228 | Node *in3 = region->in(3); |
229 | // Check that all inputs are projections |
230 | if( in1->is_Proj() && in2->is_Proj() && in3->is_Proj() ) { |
231 | Node *in10 = in1->in(0); |
232 | Node *in20 = in2->in(0); |
233 | Node *in30 = in3->in(0); |
234 | // Check that #1 and #2 are ifTrue and ifFalse from same If |
235 | if( in10 != NULL && in10->is_If() && |
236 | in20 != NULL && in20->is_If() && |
237 | in30 != NULL && in30->is_If() && in10 == in20 && |
238 | (in1->Opcode() != in2->Opcode()) ) { |
239 | Node *in100 = in10->in(0); |
240 | Node *in1000 = (in100 != NULL && in100->is_Proj()) ? in100->in(0) : NULL; |
241 | // Check that control for in10 comes from other branch of IF from in3 |
242 | if( in1000 != NULL && in1000->is_If() && |
243 | in30 == in1000 && (in3->Opcode() != in100->Opcode()) ) { |
244 | // Control pattern checks |
245 | top_if = (IfNode*)in1000; |
246 | bot_if = (IfNode*)in10; |
247 | } |
248 | } |
249 | } |
250 | |
251 | return (top_if != NULL); |
252 | } |
253 | |
254 | |
255 | //------------------------------check_convf2i_clipping------------------------- |
256 | // Helper function for RegionNode's identification of FP clipping |
257 | // Verify that the value input to the phi comes from "ConvF2I; LShift; RShift" |
258 | static bool check_convf2i_clipping( PhiNode *phi, uint idx, ConvF2INode * &convf2i, Node *min, Node *max) { |
259 | convf2i = NULL; |
260 | |
261 | // Check for the RShiftNode |
262 | Node *rshift = phi->in(idx); |
263 | assert( rshift, "Previous checks ensure phi input is present" ); |
264 | if( rshift->Opcode() != Op_RShiftI ) { return false; } |
265 | |
266 | // Check for the LShiftNode |
267 | Node *lshift = rshift->in(1); |
268 | assert( lshift, "Previous checks ensure phi input is present" ); |
269 | if( lshift->Opcode() != Op_LShiftI ) { return false; } |
270 | |
271 | // Check for the ConvF2INode |
272 | Node *conv = lshift->in(1); |
273 | if( conv->Opcode() != Op_ConvF2I ) { return false; } |
274 | |
275 | // Check that shift amounts are only to get sign bits set after F2I |
276 | jint max_cutoff = max->get_int(); |
277 | jint min_cutoff = min->get_int(); |
278 | jint left_shift = lshift->in(2)->get_int(); |
279 | jint right_shift = rshift->in(2)->get_int(); |
280 | jint max_post_shift = nth_bit(BitsPerJavaInteger - left_shift - 1); |
281 | if( left_shift != right_shift || |
282 | 0 > left_shift || left_shift >= BitsPerJavaInteger || |
283 | max_post_shift < max_cutoff || |
284 | max_post_shift < -min_cutoff ) { |
285 | // Shifts are necessary but current transformation eliminates them |
286 | return false; |
287 | } |
288 | |
289 | // OK to return the result of ConvF2I without shifting |
290 | convf2i = (ConvF2INode*)conv; |
291 | return true; |
292 | } |
293 | |
294 | |
295 | //------------------------------check_compare_clipping------------------------- |
296 | // Helper function for RegionNode's identification of FP clipping |
297 | static bool check_compare_clipping( bool less_than, IfNode *iff, ConNode *limit, Node * & input ) { |
298 | Node *i1 = iff->in(1); |
299 | if ( !i1->is_Bool() ) { return false; } |
300 | BoolNode *bool1 = i1->as_Bool(); |
301 | if( less_than && bool1->_test._test != BoolTest::le ) { return false; } |
302 | else if( !less_than && bool1->_test._test != BoolTest::lt ) { return false; } |
303 | const Node *cmpF = bool1->in(1); |
304 | if( cmpF->Opcode() != Op_CmpF ) { return false; } |
305 | // Test that the float value being compared against |
306 | // is equivalent to the int value used as a limit |
307 | Node *nodef = cmpF->in(2); |
308 | if( nodef->Opcode() != Op_ConF ) { return false; } |
309 | jfloat conf = nodef->getf(); |
310 | jint coni = limit->get_int(); |
311 | if( ((int)conf) != coni ) { return false; } |
312 | input = cmpF->in(1); |
313 | return true; |
314 | } |
315 | |
316 | //------------------------------is_unreachable_region-------------------------- |
317 | // Find if the Region node is reachable from the root. |
318 | bool RegionNode::is_unreachable_region(PhaseGVN *phase) const { |
319 | assert(req() == 2, "" ); |
320 | |
321 | // First, cut the simple case of fallthrough region when NONE of |
322 | // region's phis references itself directly or through a data node. |
323 | uint max = outcnt(); |
324 | uint i; |
325 | for (i = 0; i < max; i++) { |
326 | Node* phi = raw_out(i); |
327 | if (phi != NULL && phi->is_Phi()) { |
328 | assert(phase->eqv(phi->in(0), this) && phi->req() == 2, "" ); |
329 | if (phi->outcnt() == 0) |
330 | continue; // Safe case - no loops |
331 | if (phi->outcnt() == 1) { |
332 | Node* u = phi->raw_out(0); |
333 | // Skip if only one use is an other Phi or Call or Uncommon trap. |
334 | // It is safe to consider this case as fallthrough. |
335 | if (u != NULL && (u->is_Phi() || u->is_CFG())) |
336 | continue; |
337 | } |
338 | // Check when phi references itself directly or through an other node. |
339 | if (phi->as_Phi()->simple_data_loop_check(phi->in(1)) >= PhiNode::Unsafe) |
340 | break; // Found possible unsafe data loop. |
341 | } |
342 | } |
343 | if (i >= max) |
344 | return false; // An unsafe case was NOT found - don't need graph walk. |
345 | |
346 | // Unsafe case - check if the Region node is reachable from root. |
347 | ResourceMark rm; |
348 | |
349 | Arena *a = Thread::current()->resource_area(); |
350 | Node_List nstack(a); |
351 | VectorSet visited(a); |
352 | |
353 | // Mark all control nodes reachable from root outputs |
354 | Node *n = (Node*)phase->C->root(); |
355 | nstack.push(n); |
356 | visited.set(n->_idx); |
357 | while (nstack.size() != 0) { |
358 | n = nstack.pop(); |
359 | uint max = n->outcnt(); |
360 | for (uint i = 0; i < max; i++) { |
361 | Node* m = n->raw_out(i); |
362 | if (m != NULL && m->is_CFG()) { |
363 | if (phase->eqv(m, this)) { |
364 | return false; // We reached the Region node - it is not dead. |
365 | } |
366 | if (!visited.test_set(m->_idx)) |
367 | nstack.push(m); |
368 | } |
369 | } |
370 | } |
371 | |
372 | return true; // The Region node is unreachable - it is dead. |
373 | } |
374 | |
375 | bool RegionNode::try_clean_mem_phi(PhaseGVN *phase) { |
376 | // Incremental inlining + PhaseStringOpts sometimes produce: |
377 | // |
378 | // cmpP with 1 top input |
379 | // | |
380 | // If |
381 | // / \ |
382 | // IfFalse IfTrue /- Some Node |
383 | // \ / / / |
384 | // Region / /-MergeMem |
385 | // \---Phi |
386 | // |
387 | // |
388 | // It's expected by PhaseStringOpts that the Region goes away and is |
389 | // replaced by If's control input but because there's still a Phi, |
390 | // the Region stays in the graph. The top input from the cmpP is |
391 | // propagated forward and a subgraph that is useful goes away. The |
392 | // code below replaces the Phi with the MergeMem so that the Region |
393 | // is simplified. |
394 | |
395 | PhiNode* phi = has_unique_phi(); |
396 | if (phi && phi->type() == Type::MEMORY && req() == 3 && phi->is_diamond_phi(true)) { |
397 | MergeMemNode* m = NULL; |
398 | assert(phi->req() == 3, "same as region" ); |
399 | for (uint i = 1; i < 3; ++i) { |
400 | Node *mem = phi->in(i); |
401 | if (mem && mem->is_MergeMem() && in(i)->outcnt() == 1) { |
402 | // Nothing is control-dependent on path #i except the region itself. |
403 | m = mem->as_MergeMem(); |
404 | uint j = 3 - i; |
405 | Node* other = phi->in(j); |
406 | if (other && other == m->base_memory()) { |
407 | // m is a successor memory to other, and is not pinned inside the diamond, so push it out. |
408 | // This will allow the diamond to collapse completely. |
409 | phase->is_IterGVN()->replace_node(phi, m); |
410 | return true; |
411 | } |
412 | } |
413 | } |
414 | } |
415 | return false; |
416 | } |
417 | |
418 | //------------------------------Ideal------------------------------------------ |
419 | // Return a node which is more "ideal" than the current node. Must preserve |
420 | // the CFG, but we can still strip out dead paths. |
421 | Node *RegionNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
422 | if( !can_reshape && !in(0) ) return NULL; // Already degraded to a Copy |
423 | assert(!in(0) || !in(0)->is_Root(), "not a specially hidden merge" ); |
424 | |
425 | // Check for RegionNode with no Phi users and both inputs come from either |
426 | // arm of the same IF. If found, then the control-flow split is useless. |
427 | bool has_phis = false; |
428 | if (can_reshape) { // Need DU info to check for Phi users |
429 | has_phis = (has_phi() != NULL); // Cache result |
430 | if (has_phis && try_clean_mem_phi(phase)) { |
431 | has_phis = false; |
432 | } |
433 | |
434 | if (!has_phis) { // No Phi users? Nothing merging? |
435 | for (uint i = 1; i < req()-1; i++) { |
436 | Node *if1 = in(i); |
437 | if( !if1 ) continue; |
438 | Node *iff = if1->in(0); |
439 | if( !iff || !iff->is_If() ) continue; |
440 | for( uint j=i+1; j<req(); j++ ) { |
441 | if( in(j) && in(j)->in(0) == iff && |
442 | if1->Opcode() != in(j)->Opcode() ) { |
443 | // Add the IF Projections to the worklist. They (and the IF itself) |
444 | // will be eliminated if dead. |
445 | phase->is_IterGVN()->add_users_to_worklist(iff); |
446 | set_req(i, iff->in(0));// Skip around the useless IF diamond |
447 | set_req(j, NULL); |
448 | return this; // Record progress |
449 | } |
450 | } |
451 | } |
452 | } |
453 | } |
454 | |
455 | // Remove TOP or NULL input paths. If only 1 input path remains, this Region |
456 | // degrades to a copy. |
457 | bool add_to_worklist = false; |
458 | bool modified = false; |
459 | int cnt = 0; // Count of values merging |
460 | DEBUG_ONLY( int cnt_orig = req(); ) // Save original inputs count |
461 | int del_it = 0; // The last input path we delete |
462 | // For all inputs... |
463 | for( uint i=1; i<req(); ++i ){// For all paths in |
464 | Node *n = in(i); // Get the input |
465 | if( n != NULL ) { |
466 | // Remove useless control copy inputs |
467 | if( n->is_Region() && n->as_Region()->is_copy() ) { |
468 | set_req(i, n->nonnull_req()); |
469 | modified = true; |
470 | i--; |
471 | continue; |
472 | } |
473 | if( n->is_Proj() ) { // Remove useless rethrows |
474 | Node *call = n->in(0); |
475 | if (call->is_Call() && call->as_Call()->entry_point() == OptoRuntime::rethrow_stub()) { |
476 | set_req(i, call->in(0)); |
477 | modified = true; |
478 | i--; |
479 | continue; |
480 | } |
481 | } |
482 | if( phase->type(n) == Type::TOP ) { |
483 | set_req(i, NULL); // Ignore TOP inputs |
484 | modified = true; |
485 | i--; |
486 | continue; |
487 | } |
488 | cnt++; // One more value merging |
489 | |
490 | } else if (can_reshape) { // Else found dead path with DU info |
491 | PhaseIterGVN *igvn = phase->is_IterGVN(); |
492 | del_req(i); // Yank path from self |
493 | del_it = i; |
494 | uint max = outcnt(); |
495 | DUIterator j; |
496 | bool progress = true; |
497 | while(progress) { // Need to establish property over all users |
498 | progress = false; |
499 | for (j = outs(); has_out(j); j++) { |
500 | Node *n = out(j); |
501 | if( n->req() != req() && n->is_Phi() ) { |
502 | assert( n->in(0) == this, "" ); |
503 | igvn->hash_delete(n); // Yank from hash before hacking edges |
504 | n->set_req_X(i,NULL,igvn);// Correct DU info |
505 | n->del_req(i); // Yank path from Phis |
506 | if( max != outcnt() ) { |
507 | progress = true; |
508 | j = refresh_out_pos(j); |
509 | max = outcnt(); |
510 | } |
511 | } |
512 | } |
513 | } |
514 | add_to_worklist = true; |
515 | i--; |
516 | } |
517 | } |
518 | |
519 | if (can_reshape && cnt == 1) { |
520 | // Is it dead loop? |
521 | // If it is LoopNopde it had 2 (+1 itself) inputs and |
522 | // one of them was cut. The loop is dead if it was EntryContol. |
523 | // Loop node may have only one input because entry path |
524 | // is removed in PhaseIdealLoop::Dominators(). |
525 | assert(!this->is_Loop() || cnt_orig <= 3, "Loop node should have 3 or less inputs" ); |
526 | if ((this->is_Loop() && (del_it == LoopNode::EntryControl || |
527 | (del_it == 0 && is_unreachable_region(phase)))) || |
528 | (!this->is_Loop() && has_phis && is_unreachable_region(phase))) { |
529 | // Yes, the region will be removed during the next step below. |
530 | // Cut the backedge input and remove phis since no data paths left. |
531 | // We don't cut outputs to other nodes here since we need to put them |
532 | // on the worklist. |
533 | PhaseIterGVN *igvn = phase->is_IterGVN(); |
534 | if (in(1)->outcnt() == 1) { |
535 | igvn->_worklist.push(in(1)); |
536 | } |
537 | del_req(1); |
538 | cnt = 0; |
539 | assert( req() == 1, "no more inputs expected" ); |
540 | uint max = outcnt(); |
541 | bool progress = true; |
542 | Node *top = phase->C->top(); |
543 | DUIterator j; |
544 | while(progress) { |
545 | progress = false; |
546 | for (j = outs(); has_out(j); j++) { |
547 | Node *n = out(j); |
548 | if( n->is_Phi() ) { |
549 | assert( igvn->eqv(n->in(0), this), "" ); |
550 | assert( n->req() == 2 && n->in(1) != NULL, "Only one data input expected" ); |
551 | // Break dead loop data path. |
552 | // Eagerly replace phis with top to avoid phis copies generation. |
553 | igvn->replace_node(n, top); |
554 | if( max != outcnt() ) { |
555 | progress = true; |
556 | j = refresh_out_pos(j); |
557 | max = outcnt(); |
558 | } |
559 | } |
560 | } |
561 | } |
562 | add_to_worklist = true; |
563 | } |
564 | } |
565 | if (add_to_worklist) { |
566 | phase->is_IterGVN()->add_users_to_worklist(this); // Revisit collapsed Phis |
567 | } |
568 | |
569 | if( cnt <= 1 ) { // Only 1 path in? |
570 | set_req(0, NULL); // Null control input for region copy |
571 | if( cnt == 0 && !can_reshape) { // Parse phase - leave the node as it is. |
572 | // No inputs or all inputs are NULL. |
573 | return NULL; |
574 | } else if (can_reshape) { // Optimization phase - remove the node |
575 | PhaseIterGVN *igvn = phase->is_IterGVN(); |
576 | // Strip mined (inner) loop is going away, remove outer loop. |
577 | if (is_CountedLoop() && |
578 | as_Loop()->is_strip_mined()) { |
579 | Node* outer_sfpt = as_CountedLoop()->outer_safepoint(); |
580 | Node* outer_out = as_CountedLoop()->outer_loop_exit(); |
581 | if (outer_sfpt != NULL && outer_out != NULL) { |
582 | Node* in = outer_sfpt->in(0); |
583 | igvn->replace_node(outer_out, in); |
584 | LoopNode* outer = as_CountedLoop()->outer_loop(); |
585 | igvn->replace_input_of(outer, LoopNode::LoopBackControl, igvn->C->top()); |
586 | } |
587 | } |
588 | Node *parent_ctrl; |
589 | if( cnt == 0 ) { |
590 | assert( req() == 1, "no inputs expected" ); |
591 | // During IGVN phase such region will be subsumed by TOP node |
592 | // so region's phis will have TOP as control node. |
593 | // Kill phis here to avoid it. PhiNode::is_copy() will be always false. |
594 | // Also set other user's input to top. |
595 | parent_ctrl = phase->C->top(); |
596 | } else { |
597 | // The fallthrough case since we already checked dead loops above. |
598 | parent_ctrl = in(1); |
599 | assert(parent_ctrl != NULL, "Region is a copy of some non-null control" ); |
600 | assert(!igvn->eqv(parent_ctrl, this), "Close dead loop" ); |
601 | } |
602 | if (!add_to_worklist) |
603 | igvn->add_users_to_worklist(this); // Check for further allowed opts |
604 | for (DUIterator_Last imin, i = last_outs(imin); i >= imin; --i) { |
605 | Node* n = last_out(i); |
606 | igvn->hash_delete(n); // Remove from worklist before modifying edges |
607 | if( n->is_Phi() ) { // Collapse all Phis |
608 | // Eagerly replace phis to avoid copies generation. |
609 | Node* in; |
610 | if( cnt == 0 ) { |
611 | assert( n->req() == 1, "No data inputs expected" ); |
612 | in = parent_ctrl; // replaced by top |
613 | } else { |
614 | assert( n->req() == 2 && n->in(1) != NULL, "Only one data input expected" ); |
615 | in = n->in(1); // replaced by unique input |
616 | if( n->as_Phi()->is_unsafe_data_reference(in) ) |
617 | in = phase->C->top(); // replaced by top |
618 | } |
619 | igvn->replace_node(n, in); |
620 | } |
621 | else if( n->is_Region() ) { // Update all incoming edges |
622 | assert( !igvn->eqv(n, this), "Must be removed from DefUse edges" ); |
623 | uint uses_found = 0; |
624 | for( uint k=1; k < n->req(); k++ ) { |
625 | if( n->in(k) == this ) { |
626 | n->set_req(k, parent_ctrl); |
627 | uses_found++; |
628 | } |
629 | } |
630 | if( uses_found > 1 ) { // (--i) done at the end of the loop. |
631 | i -= (uses_found - 1); |
632 | } |
633 | } |
634 | else { |
635 | assert( igvn->eqv(n->in(0), this), "Expect RegionNode to be control parent" ); |
636 | n->set_req(0, parent_ctrl); |
637 | } |
638 | #ifdef ASSERT |
639 | for( uint k=0; k < n->req(); k++ ) { |
640 | assert( !igvn->eqv(n->in(k), this), "All uses of RegionNode should be gone" ); |
641 | } |
642 | #endif |
643 | } |
644 | // Remove the RegionNode itself from DefUse info |
645 | igvn->remove_dead_node(this); |
646 | return NULL; |
647 | } |
648 | return this; // Record progress |
649 | } |
650 | |
651 | |
652 | // If a Region flows into a Region, merge into one big happy merge. |
653 | if (can_reshape) { |
654 | Node *m = merge_region(this, phase); |
655 | if (m != NULL) return m; |
656 | } |
657 | |
658 | // Check if this region is the root of a clipping idiom on floats |
659 | if( ConvertFloat2IntClipping && can_reshape && req() == 4 ) { |
660 | // Check that only one use is a Phi and that it simplifies to two constants + |
661 | PhiNode* phi = has_unique_phi(); |
662 | if (phi != NULL) { // One Phi user |
663 | // Check inputs to the Phi |
664 | ConNode *min; |
665 | ConNode *max; |
666 | Node *val; |
667 | uint min_idx; |
668 | uint max_idx; |
669 | uint val_idx; |
670 | if( check_phi_clipping( phi, min, min_idx, max, max_idx, val, val_idx ) ) { |
671 | IfNode *top_if; |
672 | IfNode *bot_if; |
673 | if( check_if_clipping( this, bot_if, top_if ) ) { |
674 | // Control pattern checks, now verify compares |
675 | Node *top_in = NULL; // value being compared against |
676 | Node *bot_in = NULL; |
677 | if( check_compare_clipping( true, bot_if, min, bot_in ) && |
678 | check_compare_clipping( false, top_if, max, top_in ) ) { |
679 | if( bot_in == top_in ) { |
680 | PhaseIterGVN *gvn = phase->is_IterGVN(); |
681 | assert( gvn != NULL, "Only had DefUse info in IterGVN" ); |
682 | // Only remaining check is that bot_in == top_in == (Phi's val + mods) |
683 | |
684 | // Check for the ConvF2INode |
685 | ConvF2INode *convf2i; |
686 | if( check_convf2i_clipping( phi, val_idx, convf2i, min, max ) && |
687 | convf2i->in(1) == bot_in ) { |
688 | // Matched pattern, including LShiftI; RShiftI, replace with integer compares |
689 | // max test |
690 | Node *cmp = gvn->register_new_node_with_optimizer(new CmpINode( convf2i, min )); |
691 | Node *boo = gvn->register_new_node_with_optimizer(new BoolNode( cmp, BoolTest::lt )); |
692 | IfNode *iff = (IfNode*)gvn->register_new_node_with_optimizer(new IfNode( top_if->in(0), boo, PROB_UNLIKELY_MAG(5), top_if->_fcnt )); |
693 | Node *if_min= gvn->register_new_node_with_optimizer(new IfTrueNode (iff)); |
694 | Node *ifF = gvn->register_new_node_with_optimizer(new IfFalseNode(iff)); |
695 | // min test |
696 | cmp = gvn->register_new_node_with_optimizer(new CmpINode( convf2i, max )); |
697 | boo = gvn->register_new_node_with_optimizer(new BoolNode( cmp, BoolTest::gt )); |
698 | iff = (IfNode*)gvn->register_new_node_with_optimizer(new IfNode( ifF, boo, PROB_UNLIKELY_MAG(5), bot_if->_fcnt )); |
699 | Node *if_max= gvn->register_new_node_with_optimizer(new IfTrueNode (iff)); |
700 | ifF = gvn->register_new_node_with_optimizer(new IfFalseNode(iff)); |
701 | // update input edges to region node |
702 | set_req_X( min_idx, if_min, gvn ); |
703 | set_req_X( max_idx, if_max, gvn ); |
704 | set_req_X( val_idx, ifF, gvn ); |
705 | // remove unnecessary 'LShiftI; RShiftI' idiom |
706 | gvn->hash_delete(phi); |
707 | phi->set_req_X( val_idx, convf2i, gvn ); |
708 | gvn->hash_find_insert(phi); |
709 | // Return transformed region node |
710 | return this; |
711 | } |
712 | } |
713 | } |
714 | } |
715 | } |
716 | } |
717 | } |
718 | |
719 | if (can_reshape) { |
720 | modified |= optimize_trichotomy(phase->is_IterGVN()); |
721 | } |
722 | |
723 | return modified ? this : NULL; |
724 | } |
725 | |
726 | //------------------------------optimize_trichotomy-------------------------- |
727 | // Optimize nested comparisons of the following kind: |
728 | // |
729 | // int compare(int a, int b) { |
730 | // return (a < b) ? -1 : (a == b) ? 0 : 1; |
731 | // } |
732 | // |
733 | // Shape 1: |
734 | // if (compare(a, b) == 1) { ... } -> if (a > b) { ... } |
735 | // |
736 | // Shape 2: |
737 | // if (compare(a, b) == 0) { ... } -> if (a == b) { ... } |
738 | // |
739 | // Above code leads to the following IR shapes where both Ifs compare the |
740 | // same value and two out of three region inputs idx1 and idx2 map to |
741 | // the same value and control flow. |
742 | // |
743 | // (1) If (2) If |
744 | // / \ / \ |
745 | // Proj Proj Proj Proj |
746 | // | \ | \ |
747 | // | If | If If |
748 | // | / \ | / \ / \ |
749 | // | Proj Proj | Proj Proj ==> Proj Proj |
750 | // | / / \ | / | / |
751 | // Region / \ | / | / |
752 | // \ / \ | / | / |
753 | // Region Region Region |
754 | // |
755 | // The method returns true if 'this' is modified and false otherwise. |
756 | bool RegionNode::optimize_trichotomy(PhaseIterGVN* igvn) { |
757 | int idx1 = 1, idx2 = 2; |
758 | Node* region = NULL; |
759 | if (req() == 3 && in(1) != NULL && in(2) != NULL) { |
760 | // Shape 1: Check if one of the inputs is a region that merges two control |
761 | // inputs and has no other users (especially no Phi users). |
762 | region = in(1)->isa_Region() ? in(1) : in(2)->isa_Region(); |
763 | if (region == NULL || region->outcnt() != 2 || region->req() != 3) { |
764 | return false; // No suitable region input found |
765 | } |
766 | } else if (req() == 4) { |
767 | // Shape 2: Check if two control inputs map to the same value of the unique phi |
768 | // user and treat these as if they would come from another region (shape (1)). |
769 | PhiNode* phi = has_unique_phi(); |
770 | if (phi == NULL) { |
771 | return false; // No unique phi user |
772 | } |
773 | if (phi->in(idx1) != phi->in(idx2)) { |
774 | idx2 = 3; |
775 | if (phi->in(idx1) != phi->in(idx2)) { |
776 | idx1 = 2; |
777 | if (phi->in(idx1) != phi->in(idx2)) { |
778 | return false; // No equal phi inputs found |
779 | } |
780 | } |
781 | } |
782 | assert(phi->in(idx1) == phi->in(idx2), "must be" ); // Region is merging same value |
783 | region = this; |
784 | } |
785 | if (region == NULL || region->in(idx1) == NULL || region->in(idx2) == NULL) { |
786 | return false; // Region does not merge two control inputs |
787 | } |
788 | // At this point we know that region->in(idx1) and region->(idx2) map to the same |
789 | // value and control flow. Now search for ifs that feed into these region inputs. |
790 | ProjNode* proj1 = region->in(idx1)->isa_Proj(); |
791 | ProjNode* proj2 = region->in(idx2)->isa_Proj(); |
792 | if (proj1 == NULL || proj1->outcnt() != 1 || |
793 | proj2 == NULL || proj2->outcnt() != 1) { |
794 | return false; // No projection inputs with region as unique user found |
795 | } |
796 | assert(proj1 != proj2, "should be different projections" ); |
797 | IfNode* iff1 = proj1->in(0)->isa_If(); |
798 | IfNode* iff2 = proj2->in(0)->isa_If(); |
799 | if (iff1 == NULL || iff1->outcnt() != 2 || |
800 | iff2 == NULL || iff2->outcnt() != 2) { |
801 | return false; // No ifs found |
802 | } |
803 | if (iff1 == iff2) { |
804 | igvn->add_users_to_worklist(iff1); // Make sure dead if is eliminated |
805 | igvn->replace_input_of(region, idx1, iff1->in(0)); |
806 | igvn->replace_input_of(region, idx2, igvn->C->top()); |
807 | return (region == this); // Remove useless if (both projections map to the same control/value) |
808 | } |
809 | BoolNode* bol1 = iff1->in(1)->isa_Bool(); |
810 | BoolNode* bol2 = iff2->in(1)->isa_Bool(); |
811 | if (bol1 == NULL || bol2 == NULL) { |
812 | return false; // No bool inputs found |
813 | } |
814 | Node* cmp1 = bol1->in(1); |
815 | Node* cmp2 = bol2->in(1); |
816 | bool commute = false; |
817 | if (!cmp1->is_Cmp() || !cmp2->is_Cmp()) { |
818 | return false; // No comparison |
819 | } else if (cmp1->Opcode() == Op_CmpF || cmp1->Opcode() == Op_CmpD || |
820 | cmp2->Opcode() == Op_CmpF || cmp2->Opcode() == Op_CmpD || |
821 | cmp1->Opcode() == Op_CmpP || cmp1->Opcode() == Op_CmpN || |
822 | cmp2->Opcode() == Op_CmpP || cmp2->Opcode() == Op_CmpN) { |
823 | // Floats and pointers don't exactly obey trichotomy. To be on the safe side, don't transform their tests. |
824 | return false; |
825 | } else if (cmp1 != cmp2) { |
826 | if (cmp1->in(1) == cmp2->in(2) && |
827 | cmp1->in(2) == cmp2->in(1)) { |
828 | commute = true; // Same but swapped inputs, commute the test |
829 | } else { |
830 | return false; // Ifs are not comparing the same values |
831 | } |
832 | } |
833 | proj1 = proj1->other_if_proj(); |
834 | proj2 = proj2->other_if_proj(); |
835 | if (!((proj1->unique_ctrl_out() == iff2 && |
836 | proj2->unique_ctrl_out() == this) || |
837 | (proj2->unique_ctrl_out() == iff1 && |
838 | proj1->unique_ctrl_out() == this))) { |
839 | return false; // Ifs are not connected through other projs |
840 | } |
841 | // Found 'iff -> proj -> iff -> proj -> this' shape where all other projs are merged |
842 | // through 'region' and map to the same value. Merge the boolean tests and replace |
843 | // the ifs by a single comparison. |
844 | BoolTest test1 = (proj1->_con == 1) ? bol1->_test : bol1->_test.negate(); |
845 | BoolTest test2 = (proj2->_con == 1) ? bol2->_test : bol2->_test.negate(); |
846 | test1 = commute ? test1.commute() : test1; |
847 | // After possibly commuting test1, if we can merge test1 & test2, then proj2/iff2/bol2 are the nodes to refine. |
848 | BoolTest::mask res = test1.merge(test2); |
849 | if (res == BoolTest::illegal) { |
850 | return false; // Unable to merge tests |
851 | } |
852 | // Adjust iff1 to always pass (only iff2 will remain) |
853 | igvn->replace_input_of(iff1, 1, igvn->intcon(proj1->_con)); |
854 | if (res == BoolTest::never) { |
855 | // Merged test is always false, adjust iff2 to always fail |
856 | igvn->replace_input_of(iff2, 1, igvn->intcon(1 - proj2->_con)); |
857 | } else { |
858 | // Replace bool input of iff2 with merged test |
859 | BoolNode* new_bol = new BoolNode(bol2->in(1), res); |
860 | igvn->replace_input_of(iff2, 1, igvn->transform((proj2->_con == 1) ? new_bol : new_bol->negate(igvn))); |
861 | } |
862 | return false; |
863 | } |
864 | |
865 | const RegMask &RegionNode::out_RegMask() const { |
866 | return RegMask::Empty; |
867 | } |
868 | |
869 | // Find the one non-null required input. RegionNode only |
870 | Node *Node::nonnull_req() const { |
871 | assert( is_Region(), "" ); |
872 | for( uint i = 1; i < _cnt; i++ ) |
873 | if( in(i) ) |
874 | return in(i); |
875 | ShouldNotReachHere(); |
876 | return NULL; |
877 | } |
878 | |
879 | |
880 | //============================================================================= |
881 | // note that these functions assume that the _adr_type field is flattened |
882 | uint PhiNode::hash() const { |
883 | const Type* at = _adr_type; |
884 | return TypeNode::hash() + (at ? at->hash() : 0); |
885 | } |
886 | bool PhiNode::cmp( const Node &n ) const { |
887 | return TypeNode::cmp(n) && _adr_type == ((PhiNode&)n)._adr_type; |
888 | } |
889 | static inline |
890 | const TypePtr* flatten_phi_adr_type(const TypePtr* at) { |
891 | if (at == NULL || at == TypePtr::BOTTOM) return at; |
892 | return Compile::current()->alias_type(at)->adr_type(); |
893 | } |
894 | |
895 | //----------------------------make--------------------------------------------- |
896 | // create a new phi with edges matching r and set (initially) to x |
897 | PhiNode* PhiNode::make(Node* r, Node* x, const Type *t, const TypePtr* at) { |
898 | uint preds = r->req(); // Number of predecessor paths |
899 | assert(t != Type::MEMORY || at == flatten_phi_adr_type(at), "flatten at" ); |
900 | PhiNode* p = new PhiNode(r, t, at); |
901 | for (uint j = 1; j < preds; j++) { |
902 | // Fill in all inputs, except those which the region does not yet have |
903 | if (r->in(j) != NULL) |
904 | p->init_req(j, x); |
905 | } |
906 | return p; |
907 | } |
908 | PhiNode* PhiNode::make(Node* r, Node* x) { |
909 | const Type* t = x->bottom_type(); |
910 | const TypePtr* at = NULL; |
911 | if (t == Type::MEMORY) at = flatten_phi_adr_type(x->adr_type()); |
912 | return make(r, x, t, at); |
913 | } |
914 | PhiNode* PhiNode::make_blank(Node* r, Node* x) { |
915 | const Type* t = x->bottom_type(); |
916 | const TypePtr* at = NULL; |
917 | if (t == Type::MEMORY) at = flatten_phi_adr_type(x->adr_type()); |
918 | return new PhiNode(r, t, at); |
919 | } |
920 | |
921 | |
922 | //------------------------slice_memory----------------------------------------- |
923 | // create a new phi with narrowed memory type |
924 | PhiNode* PhiNode::slice_memory(const TypePtr* adr_type) const { |
925 | PhiNode* mem = (PhiNode*) clone(); |
926 | *(const TypePtr**)&mem->_adr_type = adr_type; |
927 | // convert self-loops, or else we get a bad graph |
928 | for (uint i = 1; i < req(); i++) { |
929 | if ((const Node*)in(i) == this) mem->set_req(i, mem); |
930 | } |
931 | mem->verify_adr_type(); |
932 | return mem; |
933 | } |
934 | |
935 | //------------------------split_out_instance----------------------------------- |
936 | // Split out an instance type from a bottom phi. |
937 | PhiNode* PhiNode::split_out_instance(const TypePtr* at, PhaseIterGVN *igvn) const { |
938 | const TypeOopPtr *t_oop = at->isa_oopptr(); |
939 | assert(t_oop != NULL && t_oop->is_known_instance(), "expecting instance oopptr" ); |
940 | const TypePtr *t = adr_type(); |
941 | assert(type() == Type::MEMORY && |
942 | (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM || |
943 | t->isa_oopptr() && !t->is_oopptr()->is_known_instance() && |
944 | t->is_oopptr()->cast_to_exactness(true) |
945 | ->is_oopptr()->cast_to_ptr_type(t_oop->ptr()) |
946 | ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop), |
947 | "bottom or raw memory required" ); |
948 | |
949 | // Check if an appropriate node already exists. |
950 | Node *region = in(0); |
951 | for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) { |
952 | Node* use = region->fast_out(k); |
953 | if( use->is_Phi()) { |
954 | PhiNode *phi2 = use->as_Phi(); |
955 | if (phi2->type() == Type::MEMORY && phi2->adr_type() == at) { |
956 | return phi2; |
957 | } |
958 | } |
959 | } |
960 | Compile *C = igvn->C; |
961 | Arena *a = Thread::current()->resource_area(); |
962 | Node_Array node_map = new Node_Array(a); |
963 | Node_Stack stack(a, C->live_nodes() >> 4); |
964 | PhiNode *nphi = slice_memory(at); |
965 | igvn->register_new_node_with_optimizer( nphi ); |
966 | node_map.map(_idx, nphi); |
967 | stack.push((Node *)this, 1); |
968 | while(!stack.is_empty()) { |
969 | PhiNode *ophi = stack.node()->as_Phi(); |
970 | uint i = stack.index(); |
971 | assert(i >= 1, "not control edge" ); |
972 | stack.pop(); |
973 | nphi = node_map[ophi->_idx]->as_Phi(); |
974 | for (; i < ophi->req(); i++) { |
975 | Node *in = ophi->in(i); |
976 | if (in == NULL || igvn->type(in) == Type::TOP) |
977 | continue; |
978 | Node *opt = MemNode::optimize_simple_memory_chain(in, t_oop, NULL, igvn); |
979 | PhiNode *optphi = opt->is_Phi() ? opt->as_Phi() : NULL; |
980 | if (optphi != NULL && optphi->adr_type() == TypePtr::BOTTOM) { |
981 | opt = node_map[optphi->_idx]; |
982 | if (opt == NULL) { |
983 | stack.push(ophi, i); |
984 | nphi = optphi->slice_memory(at); |
985 | igvn->register_new_node_with_optimizer( nphi ); |
986 | node_map.map(optphi->_idx, nphi); |
987 | ophi = optphi; |
988 | i = 0; // will get incremented at top of loop |
989 | continue; |
990 | } |
991 | } |
992 | nphi->set_req(i, opt); |
993 | } |
994 | } |
995 | return nphi; |
996 | } |
997 | |
998 | //------------------------verify_adr_type-------------------------------------- |
999 | #ifdef ASSERT |
1000 | void PhiNode::verify_adr_type(VectorSet& visited, const TypePtr* at) const { |
1001 | if (visited.test_set(_idx)) return; //already visited |
1002 | |
1003 | // recheck constructor invariants: |
1004 | verify_adr_type(false); |
1005 | |
1006 | // recheck local phi/phi consistency: |
1007 | assert(_adr_type == at || _adr_type == TypePtr::BOTTOM, |
1008 | "adr_type must be consistent across phi nest" ); |
1009 | |
1010 | // walk around |
1011 | for (uint i = 1; i < req(); i++) { |
1012 | Node* n = in(i); |
1013 | if (n == NULL) continue; |
1014 | const Node* np = in(i); |
1015 | if (np->is_Phi()) { |
1016 | np->as_Phi()->verify_adr_type(visited, at); |
1017 | } else if (n->bottom_type() == Type::TOP |
1018 | || (n->is_Mem() && n->in(MemNode::Address)->bottom_type() == Type::TOP)) { |
1019 | // ignore top inputs |
1020 | } else { |
1021 | const TypePtr* nat = flatten_phi_adr_type(n->adr_type()); |
1022 | // recheck phi/non-phi consistency at leaves: |
1023 | assert((nat != NULL) == (at != NULL), "" ); |
1024 | assert(nat == at || nat == TypePtr::BOTTOM, |
1025 | "adr_type must be consistent at leaves of phi nest" ); |
1026 | } |
1027 | } |
1028 | } |
1029 | |
1030 | // Verify a whole nest of phis rooted at this one. |
1031 | void PhiNode::verify_adr_type(bool recursive) const { |
1032 | if (VMError::is_error_reported()) return; // muzzle asserts when debugging an error |
1033 | if (Node::in_dump()) return; // muzzle asserts when printing |
1034 | |
1035 | assert((_type == Type::MEMORY) == (_adr_type != NULL), "adr_type for memory phis only" ); |
1036 | |
1037 | if (!VerifyAliases) return; // verify thoroughly only if requested |
1038 | |
1039 | assert(_adr_type == flatten_phi_adr_type(_adr_type), |
1040 | "Phi::adr_type must be pre-normalized" ); |
1041 | |
1042 | if (recursive) { |
1043 | VectorSet visited(Thread::current()->resource_area()); |
1044 | verify_adr_type(visited, _adr_type); |
1045 | } |
1046 | } |
1047 | #endif |
1048 | |
1049 | |
1050 | //------------------------------Value------------------------------------------ |
1051 | // Compute the type of the PhiNode |
1052 | const Type* PhiNode::Value(PhaseGVN* phase) const { |
1053 | Node *r = in(0); // RegionNode |
1054 | if( !r ) // Copy or dead |
1055 | return in(1) ? phase->type(in(1)) : Type::TOP; |
1056 | |
1057 | // Note: During parsing, phis are often transformed before their regions. |
1058 | // This means we have to use type_or_null to defend against untyped regions. |
1059 | if( phase->type_or_null(r) == Type::TOP ) // Dead code? |
1060 | return Type::TOP; |
1061 | |
1062 | // Check for trip-counted loop. If so, be smarter. |
1063 | CountedLoopNode* l = r->is_CountedLoop() ? r->as_CountedLoop() : NULL; |
1064 | if (l && ((const Node*)l->phi() == this)) { // Trip counted loop! |
1065 | // protect against init_trip() or limit() returning NULL |
1066 | if (l->can_be_counted_loop(phase)) { |
1067 | const Node *init = l->init_trip(); |
1068 | const Node *limit = l->limit(); |
1069 | const Node* stride = l->stride(); |
1070 | if (init != NULL && limit != NULL && stride != NULL) { |
1071 | const TypeInt* lo = phase->type(init)->isa_int(); |
1072 | const TypeInt* hi = phase->type(limit)->isa_int(); |
1073 | const TypeInt* stride_t = phase->type(stride)->isa_int(); |
1074 | if (lo != NULL && hi != NULL && stride_t != NULL) { // Dying loops might have TOP here |
1075 | assert(stride_t->_hi >= stride_t->_lo, "bad stride type" ); |
1076 | BoolTest::mask bt = l->loopexit()->test_trip(); |
1077 | // If the loop exit condition is "not equal", the condition |
1078 | // would not trigger if init > limit (if stride > 0) or if |
1079 | // init < limit if (stride > 0) so we can't deduce bounds |
1080 | // for the iv from the exit condition. |
1081 | if (bt != BoolTest::ne) { |
1082 | if (stride_t->_hi < 0) { // Down-counter loop |
1083 | swap(lo, hi); |
1084 | return TypeInt::make(MIN2(lo->_lo, hi->_lo) , hi->_hi, 3); |
1085 | } else if (stride_t->_lo >= 0) { |
1086 | return TypeInt::make(lo->_lo, MAX2(lo->_hi, hi->_hi), 3); |
1087 | } |
1088 | } |
1089 | } |
1090 | } |
1091 | } else if (l->in(LoopNode::LoopBackControl) != NULL && |
1092 | in(LoopNode::EntryControl) != NULL && |
1093 | phase->type(l->in(LoopNode::LoopBackControl)) == Type::TOP) { |
1094 | // During CCP, if we saturate the type of a counted loop's Phi |
1095 | // before the special code for counted loop above has a chance |
1096 | // to run (that is as long as the type of the backedge's control |
1097 | // is top), we might end up with non monotonic types |
1098 | return phase->type(in(LoopNode::EntryControl))->filter_speculative(_type); |
1099 | } |
1100 | } |
1101 | |
1102 | // Until we have harmony between classes and interfaces in the type |
1103 | // lattice, we must tread carefully around phis which implicitly |
1104 | // convert the one to the other. |
1105 | const TypePtr* ttp = _type->make_ptr(); |
1106 | const TypeInstPtr* ttip = (ttp != NULL) ? ttp->isa_instptr() : NULL; |
1107 | const TypeKlassPtr* ttkp = (ttp != NULL) ? ttp->isa_klassptr() : NULL; |
1108 | bool is_intf = false; |
1109 | if (ttip != NULL) { |
1110 | ciKlass* k = ttip->klass(); |
1111 | if (k->is_loaded() && k->is_interface()) |
1112 | is_intf = true; |
1113 | } |
1114 | if (ttkp != NULL) { |
1115 | ciKlass* k = ttkp->klass(); |
1116 | if (k->is_loaded() && k->is_interface()) |
1117 | is_intf = true; |
1118 | } |
1119 | |
1120 | // Default case: merge all inputs |
1121 | const Type *t = Type::TOP; // Merged type starting value |
1122 | for (uint i = 1; i < req(); ++i) {// For all paths in |
1123 | // Reachable control path? |
1124 | if (r->in(i) && phase->type(r->in(i)) == Type::CONTROL) { |
1125 | const Type* ti = phase->type(in(i)); |
1126 | // We assume that each input of an interface-valued Phi is a true |
1127 | // subtype of that interface. This might not be true of the meet |
1128 | // of all the input types. The lattice is not distributive in |
1129 | // such cases. Ward off asserts in type.cpp by refusing to do |
1130 | // meets between interfaces and proper classes. |
1131 | const TypePtr* tip = ti->make_ptr(); |
1132 | const TypeInstPtr* tiip = (tip != NULL) ? tip->isa_instptr() : NULL; |
1133 | if (tiip) { |
1134 | bool ti_is_intf = false; |
1135 | ciKlass* k = tiip->klass(); |
1136 | if (k->is_loaded() && k->is_interface()) |
1137 | ti_is_intf = true; |
1138 | if (is_intf != ti_is_intf) |
1139 | { t = _type; break; } |
1140 | } |
1141 | t = t->meet_speculative(ti); |
1142 | } |
1143 | } |
1144 | |
1145 | // The worst-case type (from ciTypeFlow) should be consistent with "t". |
1146 | // That is, we expect that "t->higher_equal(_type)" holds true. |
1147 | // There are various exceptions: |
1148 | // - Inputs which are phis might in fact be widened unnecessarily. |
1149 | // For example, an input might be a widened int while the phi is a short. |
1150 | // - Inputs might be BotPtrs but this phi is dependent on a null check, |
1151 | // and postCCP has removed the cast which encodes the result of the check. |
1152 | // - The type of this phi is an interface, and the inputs are classes. |
1153 | // - Value calls on inputs might produce fuzzy results. |
1154 | // (Occurrences of this case suggest improvements to Value methods.) |
1155 | // |
1156 | // It is not possible to see Type::BOTTOM values as phi inputs, |
1157 | // because the ciTypeFlow pre-pass produces verifier-quality types. |
1158 | const Type* ft = t->filter_speculative(_type); // Worst case type |
1159 | |
1160 | #ifdef ASSERT |
1161 | // The following logic has been moved into TypeOopPtr::filter. |
1162 | const Type* jt = t->join_speculative(_type); |
1163 | if (jt->empty()) { // Emptied out??? |
1164 | |
1165 | // Check for evil case of 't' being a class and '_type' expecting an |
1166 | // interface. This can happen because the bytecodes do not contain |
1167 | // enough type info to distinguish a Java-level interface variable |
1168 | // from a Java-level object variable. If we meet 2 classes which |
1169 | // both implement interface I, but their meet is at 'j/l/O' which |
1170 | // doesn't implement I, we have no way to tell if the result should |
1171 | // be 'I' or 'j/l/O'. Thus we'll pick 'j/l/O'. If this then flows |
1172 | // into a Phi which "knows" it's an Interface type we'll have to |
1173 | // uplift the type. |
1174 | if (!t->empty() && ttip && ttip->is_loaded() && ttip->klass()->is_interface()) { |
1175 | assert(ft == _type, "" ); // Uplift to interface |
1176 | } else if (!t->empty() && ttkp && ttkp->is_loaded() && ttkp->klass()->is_interface()) { |
1177 | assert(ft == _type, "" ); // Uplift to interface |
1178 | } else { |
1179 | // We also have to handle 'evil cases' of interface- vs. class-arrays |
1180 | Type::get_arrays_base_elements(jt, _type, NULL, &ttip); |
1181 | if (!t->empty() && ttip != NULL && ttip->is_loaded() && ttip->klass()->is_interface()) { |
1182 | assert(ft == _type, "" ); // Uplift to array of interface |
1183 | } else { |
1184 | // Otherwise it's something stupid like non-overlapping int ranges |
1185 | // found on dying counted loops. |
1186 | assert(ft == Type::TOP, "" ); // Canonical empty value |
1187 | } |
1188 | } |
1189 | } |
1190 | |
1191 | else { |
1192 | |
1193 | // If we have an interface-typed Phi and we narrow to a class type, the join |
1194 | // should report back the class. However, if we have a J/L/Object |
1195 | // class-typed Phi and an interface flows in, it's possible that the meet & |
1196 | // join report an interface back out. This isn't possible but happens |
1197 | // because the type system doesn't interact well with interfaces. |
1198 | const TypePtr *jtp = jt->make_ptr(); |
1199 | const TypeInstPtr *jtip = (jtp != NULL) ? jtp->isa_instptr() : NULL; |
1200 | const TypeKlassPtr *jtkp = (jtp != NULL) ? jtp->isa_klassptr() : NULL; |
1201 | if( jtip && ttip ) { |
1202 | if( jtip->is_loaded() && jtip->klass()->is_interface() && |
1203 | ttip->is_loaded() && !ttip->klass()->is_interface() ) { |
1204 | assert(ft == ttip->cast_to_ptr_type(jtip->ptr()) || |
1205 | ft->isa_narrowoop() && ft->make_ptr() == ttip->cast_to_ptr_type(jtip->ptr()), "" ); |
1206 | jt = ft; |
1207 | } |
1208 | } |
1209 | if( jtkp && ttkp ) { |
1210 | if( jtkp->is_loaded() && jtkp->klass()->is_interface() && |
1211 | !jtkp->klass_is_exact() && // Keep exact interface klass (6894807) |
1212 | ttkp->is_loaded() && !ttkp->klass()->is_interface() ) { |
1213 | assert(ft == ttkp->cast_to_ptr_type(jtkp->ptr()) || |
1214 | ft->isa_narrowklass() && ft->make_ptr() == ttkp->cast_to_ptr_type(jtkp->ptr()), "" ); |
1215 | jt = ft; |
1216 | } |
1217 | } |
1218 | if (jt != ft && jt->base() == ft->base()) { |
1219 | if (jt->isa_int() && |
1220 | jt->is_int()->_lo == ft->is_int()->_lo && |
1221 | jt->is_int()->_hi == ft->is_int()->_hi) |
1222 | jt = ft; |
1223 | if (jt->isa_long() && |
1224 | jt->is_long()->_lo == ft->is_long()->_lo && |
1225 | jt->is_long()->_hi == ft->is_long()->_hi) |
1226 | jt = ft; |
1227 | } |
1228 | if (jt != ft) { |
1229 | tty->print("merge type: " ); t->dump(); tty->cr(); |
1230 | tty->print("kill type: " ); _type->dump(); tty->cr(); |
1231 | tty->print("join type: " ); jt->dump(); tty->cr(); |
1232 | tty->print("filter type: " ); ft->dump(); tty->cr(); |
1233 | } |
1234 | assert(jt == ft, "" ); |
1235 | } |
1236 | #endif //ASSERT |
1237 | |
1238 | // Deal with conversion problems found in data loops. |
1239 | ft = phase->saturate(ft, phase->type_or_null(this), _type); |
1240 | |
1241 | return ft; |
1242 | } |
1243 | |
1244 | |
1245 | //------------------------------is_diamond_phi--------------------------------- |
1246 | // Does this Phi represent a simple well-shaped diamond merge? Return the |
1247 | // index of the true path or 0 otherwise. |
1248 | // If check_control_only is true, do not inspect the If node at the |
1249 | // top, and return -1 (not an edge number) on success. |
1250 | int PhiNode::is_diamond_phi(bool check_control_only) const { |
1251 | // Check for a 2-path merge |
1252 | Node *region = in(0); |
1253 | if( !region ) return 0; |
1254 | if( region->req() != 3 ) return 0; |
1255 | if( req() != 3 ) return 0; |
1256 | // Check that both paths come from the same If |
1257 | Node *ifp1 = region->in(1); |
1258 | Node *ifp2 = region->in(2); |
1259 | if( !ifp1 || !ifp2 ) return 0; |
1260 | Node *iff = ifp1->in(0); |
1261 | if( !iff || !iff->is_If() ) return 0; |
1262 | if( iff != ifp2->in(0) ) return 0; |
1263 | if (check_control_only) return -1; |
1264 | // Check for a proper bool/cmp |
1265 | const Node *b = iff->in(1); |
1266 | if( !b->is_Bool() ) return 0; |
1267 | const Node *cmp = b->in(1); |
1268 | if( !cmp->is_Cmp() ) return 0; |
1269 | |
1270 | // Check for branching opposite expected |
1271 | if( ifp2->Opcode() == Op_IfTrue ) { |
1272 | assert( ifp1->Opcode() == Op_IfFalse, "" ); |
1273 | return 2; |
1274 | } else { |
1275 | assert( ifp1->Opcode() == Op_IfTrue, "" ); |
1276 | return 1; |
1277 | } |
1278 | } |
1279 | |
1280 | //----------------------------check_cmove_id----------------------------------- |
1281 | // Check for CMove'ing a constant after comparing against the constant. |
1282 | // Happens all the time now, since if we compare equality vs a constant in |
1283 | // the parser, we "know" the variable is constant on one path and we force |
1284 | // it. Thus code like "if( x==0 ) {/*EMPTY*/}" ends up inserting a |
1285 | // conditional move: "x = (x==0)?0:x;". Yucko. This fix is slightly more |
1286 | // general in that we don't need constants. Since CMove's are only inserted |
1287 | // in very special circumstances, we do it here on generic Phi's. |
1288 | Node* PhiNode::is_cmove_id(PhaseTransform* phase, int true_path) { |
1289 | assert(true_path !=0, "only diamond shape graph expected" ); |
1290 | |
1291 | // is_diamond_phi() has guaranteed the correctness of the nodes sequence: |
1292 | // phi->region->if_proj->ifnode->bool->cmp |
1293 | Node* region = in(0); |
1294 | Node* iff = region->in(1)->in(0); |
1295 | BoolNode* b = iff->in(1)->as_Bool(); |
1296 | Node* cmp = b->in(1); |
1297 | Node* tval = in(true_path); |
1298 | Node* fval = in(3-true_path); |
1299 | Node* id = CMoveNode::is_cmove_id(phase, cmp, tval, fval, b); |
1300 | if (id == NULL) |
1301 | return NULL; |
1302 | |
1303 | // Either value might be a cast that depends on a branch of 'iff'. |
1304 | // Since the 'id' value will float free of the diamond, either |
1305 | // decast or return failure. |
1306 | Node* ctl = id->in(0); |
1307 | if (ctl != NULL && ctl->in(0) == iff) { |
1308 | if (id->is_ConstraintCast()) { |
1309 | return id->in(1); |
1310 | } else { |
1311 | // Don't know how to disentangle this value. |
1312 | return NULL; |
1313 | } |
1314 | } |
1315 | |
1316 | return id; |
1317 | } |
1318 | |
1319 | //------------------------------Identity--------------------------------------- |
1320 | // Check for Region being Identity. |
1321 | Node* PhiNode::Identity(PhaseGVN* phase) { |
1322 | // Check for no merging going on |
1323 | // (There used to be special-case code here when this->region->is_Loop. |
1324 | // It would check for a tributary phi on the backedge that the main phi |
1325 | // trivially, perhaps with a single cast. The unique_input method |
1326 | // does all this and more, by reducing such tributaries to 'this'.) |
1327 | Node* uin = unique_input(phase, false); |
1328 | if (uin != NULL) { |
1329 | return uin; |
1330 | } |
1331 | |
1332 | int true_path = is_diamond_phi(); |
1333 | if (true_path != 0) { |
1334 | Node* id = is_cmove_id(phase, true_path); |
1335 | if (id != NULL) return id; |
1336 | } |
1337 | |
1338 | return this; // No identity |
1339 | } |
1340 | |
1341 | //-----------------------------unique_input------------------------------------ |
1342 | // Find the unique value, discounting top, self-loops, and casts. |
1343 | // Return top if there are no inputs, and self if there are multiple. |
1344 | Node* PhiNode::unique_input(PhaseTransform* phase, bool uncast) { |
1345 | // 1) One unique direct input, |
1346 | // or if uncast is true: |
1347 | // 2) some of the inputs have an intervening ConstraintCast |
1348 | // 3) an input is a self loop |
1349 | // |
1350 | // 1) input or 2) input or 3) input __ |
1351 | // / \ / \ \ / \ |
1352 | // \ / | cast phi cast |
1353 | // phi \ / / \ / |
1354 | // phi / -- |
1355 | |
1356 | Node* r = in(0); // RegionNode |
1357 | if (r == NULL) return in(1); // Already degraded to a Copy |
1358 | Node* input = NULL; // The unique direct input (maybe uncasted = ConstraintCasts removed) |
1359 | |
1360 | for (uint i = 1, cnt = req(); i < cnt; ++i) { |
1361 | Node* rc = r->in(i); |
1362 | if (rc == NULL || phase->type(rc) == Type::TOP) |
1363 | continue; // ignore unreachable control path |
1364 | Node* n = in(i); |
1365 | if (n == NULL) |
1366 | continue; |
1367 | Node* un = n; |
1368 | if (uncast) { |
1369 | #ifdef ASSERT |
1370 | Node* m = un->uncast(); |
1371 | #endif |
1372 | while (un != NULL && un->req() == 2 && un->is_ConstraintCast()) { |
1373 | Node* next = un->in(1); |
1374 | if (phase->type(next)->isa_rawptr() && phase->type(un)->isa_oopptr()) { |
1375 | // risk exposing raw ptr at safepoint |
1376 | break; |
1377 | } |
1378 | un = next; |
1379 | } |
1380 | assert(m == un || un->in(1) == m, "Only expected at CheckCastPP from allocation" ); |
1381 | } |
1382 | if (un == NULL || un == this || phase->type(un) == Type::TOP) { |
1383 | continue; // ignore if top, or in(i) and "this" are in a data cycle |
1384 | } |
1385 | // Check for a unique input (maybe uncasted) |
1386 | if (input == NULL) { |
1387 | input = un; |
1388 | } else if (input != un) { |
1389 | input = NodeSentinel; // no unique input |
1390 | } |
1391 | } |
1392 | if (input == NULL) { |
1393 | return phase->C->top(); // no inputs |
1394 | } |
1395 | |
1396 | if (input != NodeSentinel) { |
1397 | return input; // one unique direct input |
1398 | } |
1399 | |
1400 | // Nothing. |
1401 | return NULL; |
1402 | } |
1403 | |
1404 | //------------------------------is_x2logic------------------------------------- |
1405 | // Check for simple convert-to-boolean pattern |
1406 | // If:(C Bool) Region:(IfF IfT) Phi:(Region 0 1) |
1407 | // Convert Phi to an ConvIB. |
1408 | static Node *is_x2logic( PhaseGVN *phase, PhiNode *phi, int true_path ) { |
1409 | assert(true_path !=0, "only diamond shape graph expected" ); |
1410 | // Convert the true/false index into an expected 0/1 return. |
1411 | // Map 2->0 and 1->1. |
1412 | int flipped = 2-true_path; |
1413 | |
1414 | // is_diamond_phi() has guaranteed the correctness of the nodes sequence: |
1415 | // phi->region->if_proj->ifnode->bool->cmp |
1416 | Node *region = phi->in(0); |
1417 | Node *iff = region->in(1)->in(0); |
1418 | BoolNode *b = (BoolNode*)iff->in(1); |
1419 | const CmpNode *cmp = (CmpNode*)b->in(1); |
1420 | |
1421 | Node *zero = phi->in(1); |
1422 | Node *one = phi->in(2); |
1423 | const Type *tzero = phase->type( zero ); |
1424 | const Type *tone = phase->type( one ); |
1425 | |
1426 | // Check for compare vs 0 |
1427 | const Type *tcmp = phase->type(cmp->in(2)); |
1428 | if( tcmp != TypeInt::ZERO && tcmp != TypePtr::NULL_PTR ) { |
1429 | // Allow cmp-vs-1 if the other input is bounded by 0-1 |
1430 | if( !(tcmp == TypeInt::ONE && phase->type(cmp->in(1)) == TypeInt::BOOL) ) |
1431 | return NULL; |
1432 | flipped = 1-flipped; // Test is vs 1 instead of 0! |
1433 | } |
1434 | |
1435 | // Check for setting zero/one opposite expected |
1436 | if( tzero == TypeInt::ZERO ) { |
1437 | if( tone == TypeInt::ONE ) { |
1438 | } else return NULL; |
1439 | } else if( tzero == TypeInt::ONE ) { |
1440 | if( tone == TypeInt::ZERO ) { |
1441 | flipped = 1-flipped; |
1442 | } else return NULL; |
1443 | } else return NULL; |
1444 | |
1445 | // Check for boolean test backwards |
1446 | if( b->_test._test == BoolTest::ne ) { |
1447 | } else if( b->_test._test == BoolTest::eq ) { |
1448 | flipped = 1-flipped; |
1449 | } else return NULL; |
1450 | |
1451 | // Build int->bool conversion |
1452 | Node *in1 = cmp->in(1); |
1453 | BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); |
1454 | in1 = bs->step_over_gc_barrier(in1); |
1455 | Node *n = new Conv2BNode(in1); |
1456 | if( flipped ) |
1457 | n = new XorINode( phase->transform(n), phase->intcon(1) ); |
1458 | |
1459 | return n; |
1460 | } |
1461 | |
1462 | //------------------------------is_cond_add------------------------------------ |
1463 | // Check for simple conditional add pattern: "(P < Q) ? X+Y : X;" |
1464 | // To be profitable the control flow has to disappear; there can be no other |
1465 | // values merging here. We replace the test-and-branch with: |
1466 | // "(sgn(P-Q))&Y) + X". Basically, convert "(P < Q)" into 0 or -1 by |
1467 | // moving the carry bit from (P-Q) into a register with 'sbb EAX,EAX'. |
1468 | // Then convert Y to 0-or-Y and finally add. |
1469 | // This is a key transform for SpecJava _201_compress. |
1470 | static Node* is_cond_add(PhaseGVN *phase, PhiNode *phi, int true_path) { |
1471 | assert(true_path !=0, "only diamond shape graph expected" ); |
1472 | |
1473 | // is_diamond_phi() has guaranteed the correctness of the nodes sequence: |
1474 | // phi->region->if_proj->ifnode->bool->cmp |
1475 | RegionNode *region = (RegionNode*)phi->in(0); |
1476 | Node *iff = region->in(1)->in(0); |
1477 | BoolNode* b = iff->in(1)->as_Bool(); |
1478 | const CmpNode *cmp = (CmpNode*)b->in(1); |
1479 | |
1480 | // Make sure only merging this one phi here |
1481 | if (region->has_unique_phi() != phi) return NULL; |
1482 | |
1483 | // Make sure each arm of the diamond has exactly one output, which we assume |
1484 | // is the region. Otherwise, the control flow won't disappear. |
1485 | if (region->in(1)->outcnt() != 1) return NULL; |
1486 | if (region->in(2)->outcnt() != 1) return NULL; |
1487 | |
1488 | // Check for "(P < Q)" of type signed int |
1489 | if (b->_test._test != BoolTest::lt) return NULL; |
1490 | if (cmp->Opcode() != Op_CmpI) return NULL; |
1491 | |
1492 | Node *p = cmp->in(1); |
1493 | Node *q = cmp->in(2); |
1494 | Node *n1 = phi->in( true_path); |
1495 | Node *n2 = phi->in(3-true_path); |
1496 | |
1497 | int op = n1->Opcode(); |
1498 | if( op != Op_AddI // Need zero as additive identity |
1499 | /*&&op != Op_SubI && |
1500 | op != Op_AddP && |
1501 | op != Op_XorI && |
1502 | op != Op_OrI*/ ) |
1503 | return NULL; |
1504 | |
1505 | Node *x = n2; |
1506 | Node *y = NULL; |
1507 | if( x == n1->in(1) ) { |
1508 | y = n1->in(2); |
1509 | } else if( x == n1->in(2) ) { |
1510 | y = n1->in(1); |
1511 | } else return NULL; |
1512 | |
1513 | // Not so profitable if compare and add are constants |
1514 | if( q->is_Con() && phase->type(q) != TypeInt::ZERO && y->is_Con() ) |
1515 | return NULL; |
1516 | |
1517 | Node *cmplt = phase->transform( new CmpLTMaskNode(p,q) ); |
1518 | Node *j_and = phase->transform( new AndINode(cmplt,y) ); |
1519 | return new AddINode(j_and,x); |
1520 | } |
1521 | |
1522 | //------------------------------is_absolute------------------------------------ |
1523 | // Check for absolute value. |
1524 | static Node* is_absolute( PhaseGVN *phase, PhiNode *phi_root, int true_path) { |
1525 | assert(true_path !=0, "only diamond shape graph expected" ); |
1526 | |
1527 | int cmp_zero_idx = 0; // Index of compare input where to look for zero |
1528 | int phi_x_idx = 0; // Index of phi input where to find naked x |
1529 | |
1530 | // ABS ends with the merge of 2 control flow paths. |
1531 | // Find the false path from the true path. With only 2 inputs, 3 - x works nicely. |
1532 | int false_path = 3 - true_path; |
1533 | |
1534 | // is_diamond_phi() has guaranteed the correctness of the nodes sequence: |
1535 | // phi->region->if_proj->ifnode->bool->cmp |
1536 | BoolNode *bol = phi_root->in(0)->in(1)->in(0)->in(1)->as_Bool(); |
1537 | |
1538 | // Check bool sense |
1539 | switch( bol->_test._test ) { |
1540 | case BoolTest::lt: cmp_zero_idx = 1; phi_x_idx = true_path; break; |
1541 | case BoolTest::le: cmp_zero_idx = 2; phi_x_idx = false_path; break; |
1542 | case BoolTest::gt: cmp_zero_idx = 2; phi_x_idx = true_path; break; |
1543 | case BoolTest::ge: cmp_zero_idx = 1; phi_x_idx = false_path; break; |
1544 | default: return NULL; break; |
1545 | } |
1546 | |
1547 | // Test is next |
1548 | Node *cmp = bol->in(1); |
1549 | const Type *tzero = NULL; |
1550 | switch( cmp->Opcode() ) { |
1551 | case Op_CmpF: tzero = TypeF::ZERO; break; // Float ABS |
1552 | case Op_CmpD: tzero = TypeD::ZERO; break; // Double ABS |
1553 | default: return NULL; |
1554 | } |
1555 | |
1556 | // Find zero input of compare; the other input is being abs'd |
1557 | Node *x = NULL; |
1558 | bool flip = false; |
1559 | if( phase->type(cmp->in(cmp_zero_idx)) == tzero ) { |
1560 | x = cmp->in(3 - cmp_zero_idx); |
1561 | } else if( phase->type(cmp->in(3 - cmp_zero_idx)) == tzero ) { |
1562 | // The test is inverted, we should invert the result... |
1563 | x = cmp->in(cmp_zero_idx); |
1564 | flip = true; |
1565 | } else { |
1566 | return NULL; |
1567 | } |
1568 | |
1569 | // Next get the 2 pieces being selected, one is the original value |
1570 | // and the other is the negated value. |
1571 | if( phi_root->in(phi_x_idx) != x ) return NULL; |
1572 | |
1573 | // Check other phi input for subtract node |
1574 | Node *sub = phi_root->in(3 - phi_x_idx); |
1575 | |
1576 | // Allow only Sub(0,X) and fail out for all others; Neg is not OK |
1577 | if( tzero == TypeF::ZERO ) { |
1578 | if( sub->Opcode() != Op_SubF || |
1579 | sub->in(2) != x || |
1580 | phase->type(sub->in(1)) != tzero ) return NULL; |
1581 | x = new AbsFNode(x); |
1582 | if (flip) { |
1583 | x = new SubFNode(sub->in(1), phase->transform(x)); |
1584 | } |
1585 | } else { |
1586 | if( sub->Opcode() != Op_SubD || |
1587 | sub->in(2) != x || |
1588 | phase->type(sub->in(1)) != tzero ) return NULL; |
1589 | x = new AbsDNode(x); |
1590 | if (flip) { |
1591 | x = new SubDNode(sub->in(1), phase->transform(x)); |
1592 | } |
1593 | } |
1594 | |
1595 | return x; |
1596 | } |
1597 | |
1598 | //------------------------------split_once------------------------------------- |
1599 | // Helper for split_flow_path |
1600 | static void split_once(PhaseIterGVN *igvn, Node *phi, Node *val, Node *n, Node *newn) { |
1601 | igvn->hash_delete(n); // Remove from hash before hacking edges |
1602 | |
1603 | uint j = 1; |
1604 | for (uint i = phi->req()-1; i > 0; i--) { |
1605 | if (phi->in(i) == val) { // Found a path with val? |
1606 | // Add to NEW Region/Phi, no DU info |
1607 | newn->set_req( j++, n->in(i) ); |
1608 | // Remove from OLD Region/Phi |
1609 | n->del_req(i); |
1610 | } |
1611 | } |
1612 | |
1613 | // Register the new node but do not transform it. Cannot transform until the |
1614 | // entire Region/Phi conglomerate has been hacked as a single huge transform. |
1615 | igvn->register_new_node_with_optimizer( newn ); |
1616 | |
1617 | // Now I can point to the new node. |
1618 | n->add_req(newn); |
1619 | igvn->_worklist.push(n); |
1620 | } |
1621 | |
1622 | //------------------------------split_flow_path-------------------------------- |
1623 | // Check for merging identical values and split flow paths |
1624 | static Node* split_flow_path(PhaseGVN *phase, PhiNode *phi) { |
1625 | BasicType bt = phi->type()->basic_type(); |
1626 | if( bt == T_ILLEGAL || type2size[bt] <= 0 ) |
1627 | return NULL; // Bail out on funny non-value stuff |
1628 | if( phi->req() <= 3 ) // Need at least 2 matched inputs and a |
1629 | return NULL; // third unequal input to be worth doing |
1630 | |
1631 | // Scan for a constant |
1632 | uint i; |
1633 | for( i = 1; i < phi->req()-1; i++ ) { |
1634 | Node *n = phi->in(i); |
1635 | if( !n ) return NULL; |
1636 | if( phase->type(n) == Type::TOP ) return NULL; |
1637 | if( n->Opcode() == Op_ConP || n->Opcode() == Op_ConN || n->Opcode() == Op_ConNKlass ) |
1638 | break; |
1639 | } |
1640 | if( i >= phi->req() ) // Only split for constants |
1641 | return NULL; |
1642 | |
1643 | Node *val = phi->in(i); // Constant to split for |
1644 | uint hit = 0; // Number of times it occurs |
1645 | Node *r = phi->region(); |
1646 | |
1647 | for( ; i < phi->req(); i++ ){ // Count occurrences of constant |
1648 | Node *n = phi->in(i); |
1649 | if( !n ) return NULL; |
1650 | if( phase->type(n) == Type::TOP ) return NULL; |
1651 | if( phi->in(i) == val ) { |
1652 | hit++; |
1653 | if (PhaseIdealLoop::find_predicate(r->in(i)) != NULL) { |
1654 | return NULL; // don't split loop entry path |
1655 | } |
1656 | } |
1657 | } |
1658 | |
1659 | if( hit <= 1 || // Make sure we find 2 or more |
1660 | hit == phi->req()-1 ) // and not ALL the same value |
1661 | return NULL; |
1662 | |
1663 | // Now start splitting out the flow paths that merge the same value. |
1664 | // Split first the RegionNode. |
1665 | PhaseIterGVN *igvn = phase->is_IterGVN(); |
1666 | RegionNode *newr = new RegionNode(hit+1); |
1667 | split_once(igvn, phi, val, r, newr); |
1668 | |
1669 | // Now split all other Phis than this one |
1670 | for (DUIterator_Fast kmax, k = r->fast_outs(kmax); k < kmax; k++) { |
1671 | Node* phi2 = r->fast_out(k); |
1672 | if( phi2->is_Phi() && phi2->as_Phi() != phi ) { |
1673 | PhiNode *newphi = PhiNode::make_blank(newr, phi2); |
1674 | split_once(igvn, phi, val, phi2, newphi); |
1675 | } |
1676 | } |
1677 | |
1678 | // Clean up this guy |
1679 | igvn->hash_delete(phi); |
1680 | for( i = phi->req()-1; i > 0; i-- ) { |
1681 | if( phi->in(i) == val ) { |
1682 | phi->del_req(i); |
1683 | } |
1684 | } |
1685 | phi->add_req(val); |
1686 | |
1687 | return phi; |
1688 | } |
1689 | |
1690 | //============================================================================= |
1691 | //------------------------------simple_data_loop_check------------------------- |
1692 | // Try to determining if the phi node in a simple safe/unsafe data loop. |
1693 | // Returns: |
1694 | // enum LoopSafety { Safe = 0, Unsafe, UnsafeLoop }; |
1695 | // Safe - safe case when the phi and it's inputs reference only safe data |
1696 | // nodes; |
1697 | // Unsafe - the phi and it's inputs reference unsafe data nodes but there |
1698 | // is no reference back to the phi - need a graph walk |
1699 | // to determine if it is in a loop; |
1700 | // UnsafeLoop - unsafe case when the phi references itself directly or through |
1701 | // unsafe data node. |
1702 | // Note: a safe data node is a node which could/never reference itself during |
1703 | // GVN transformations. For now it is Con, Proj, Phi, CastPP, CheckCastPP. |
1704 | // I mark Phi nodes as safe node not only because they can reference itself |
1705 | // but also to prevent mistaking the fallthrough case inside an outer loop |
1706 | // as dead loop when the phi references itselfs through an other phi. |
1707 | PhiNode::LoopSafety PhiNode::simple_data_loop_check(Node *in) const { |
1708 | // It is unsafe loop if the phi node references itself directly. |
1709 | if (in == (Node*)this) |
1710 | return UnsafeLoop; // Unsafe loop |
1711 | // Unsafe loop if the phi node references itself through an unsafe data node. |
1712 | // Exclude cases with null inputs or data nodes which could reference |
1713 | // itself (safe for dead loops). |
1714 | if (in != NULL && !in->is_dead_loop_safe()) { |
1715 | // Check inputs of phi's inputs also. |
1716 | // It is much less expensive then full graph walk. |
1717 | uint cnt = in->req(); |
1718 | uint i = (in->is_Proj() && !in->is_CFG()) ? 0 : 1; |
1719 | for (; i < cnt; ++i) { |
1720 | Node* m = in->in(i); |
1721 | if (m == (Node*)this) |
1722 | return UnsafeLoop; // Unsafe loop |
1723 | if (m != NULL && !m->is_dead_loop_safe()) { |
1724 | // Check the most common case (about 30% of all cases): |
1725 | // phi->Load/Store->AddP->(ConP ConP Con)/(Parm Parm Con). |
1726 | Node *m1 = (m->is_AddP() && m->req() > 3) ? m->in(1) : NULL; |
1727 | if (m1 == (Node*)this) |
1728 | return UnsafeLoop; // Unsafe loop |
1729 | if (m1 != NULL && m1 == m->in(2) && |
1730 | m1->is_dead_loop_safe() && m->in(3)->is_Con()) { |
1731 | continue; // Safe case |
1732 | } |
1733 | // The phi references an unsafe node - need full analysis. |
1734 | return Unsafe; |
1735 | } |
1736 | } |
1737 | } |
1738 | return Safe; // Safe case - we can optimize the phi node. |
1739 | } |
1740 | |
1741 | //------------------------------is_unsafe_data_reference----------------------- |
1742 | // If phi can be reached through the data input - it is data loop. |
1743 | bool PhiNode::is_unsafe_data_reference(Node *in) const { |
1744 | assert(req() > 1, "" ); |
1745 | // First, check simple cases when phi references itself directly or |
1746 | // through an other node. |
1747 | LoopSafety safety = simple_data_loop_check(in); |
1748 | if (safety == UnsafeLoop) |
1749 | return true; // phi references itself - unsafe loop |
1750 | else if (safety == Safe) |
1751 | return false; // Safe case - phi could be replaced with the unique input. |
1752 | |
1753 | // Unsafe case when we should go through data graph to determine |
1754 | // if the phi references itself. |
1755 | |
1756 | ResourceMark rm; |
1757 | |
1758 | Arena *a = Thread::current()->resource_area(); |
1759 | Node_List nstack(a); |
1760 | VectorSet visited(a); |
1761 | |
1762 | nstack.push(in); // Start with unique input. |
1763 | visited.set(in->_idx); |
1764 | while (nstack.size() != 0) { |
1765 | Node* n = nstack.pop(); |
1766 | uint cnt = n->req(); |
1767 | uint i = (n->is_Proj() && !n->is_CFG()) ? 0 : 1; |
1768 | for (; i < cnt; i++) { |
1769 | Node* m = n->in(i); |
1770 | if (m == (Node*)this) { |
1771 | return true; // Data loop |
1772 | } |
1773 | if (m != NULL && !m->is_dead_loop_safe()) { // Only look for unsafe cases. |
1774 | if (!visited.test_set(m->_idx)) |
1775 | nstack.push(m); |
1776 | } |
1777 | } |
1778 | } |
1779 | return false; // The phi is not reachable from its inputs |
1780 | } |
1781 | |
1782 | |
1783 | //------------------------------Ideal------------------------------------------ |
1784 | // Return a node which is more "ideal" than the current node. Must preserve |
1785 | // the CFG, but we can still strip out dead paths. |
1786 | Node *PhiNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
1787 | // The next should never happen after 6297035 fix. |
1788 | if( is_copy() ) // Already degraded to a Copy ? |
1789 | return NULL; // No change |
1790 | |
1791 | Node *r = in(0); // RegionNode |
1792 | assert(r->in(0) == NULL || !r->in(0)->is_Root(), "not a specially hidden merge" ); |
1793 | |
1794 | // Note: During parsing, phis are often transformed before their regions. |
1795 | // This means we have to use type_or_null to defend against untyped regions. |
1796 | if( phase->type_or_null(r) == Type::TOP ) // Dead code? |
1797 | return NULL; // No change |
1798 | |
1799 | Node *top = phase->C->top(); |
1800 | bool new_phi = (outcnt() == 0); // transforming new Phi |
1801 | // No change for igvn if new phi is not hooked |
1802 | if (new_phi && can_reshape) |
1803 | return NULL; |
1804 | |
1805 | // The are 2 situations when only one valid phi's input is left |
1806 | // (in addition to Region input). |
1807 | // One: region is not loop - replace phi with this input. |
1808 | // Two: region is loop - replace phi with top since this data path is dead |
1809 | // and we need to break the dead data loop. |
1810 | Node* progress = NULL; // Record if any progress made |
1811 | for( uint j = 1; j < req(); ++j ){ // For all paths in |
1812 | // Check unreachable control paths |
1813 | Node* rc = r->in(j); |
1814 | Node* n = in(j); // Get the input |
1815 | if (rc == NULL || phase->type(rc) == Type::TOP) { |
1816 | if (n != top) { // Not already top? |
1817 | PhaseIterGVN *igvn = phase->is_IterGVN(); |
1818 | if (can_reshape && igvn != NULL) { |
1819 | igvn->_worklist.push(r); |
1820 | } |
1821 | // Nuke it down |
1822 | if (can_reshape) { |
1823 | set_req_X(j, top, igvn); |
1824 | } else { |
1825 | set_req(j, top); |
1826 | } |
1827 | progress = this; // Record progress |
1828 | } |
1829 | } |
1830 | } |
1831 | |
1832 | if (can_reshape && outcnt() == 0) { |
1833 | // set_req() above may kill outputs if Phi is referenced |
1834 | // only by itself on the dead (top) control path. |
1835 | return top; |
1836 | } |
1837 | |
1838 | bool uncasted = false; |
1839 | Node* uin = unique_input(phase, false); |
1840 | if (uin == NULL && can_reshape) { |
1841 | uncasted = true; |
1842 | uin = unique_input(phase, true); |
1843 | } |
1844 | if (uin == top) { // Simplest case: no alive inputs. |
1845 | if (can_reshape) // IGVN transformation |
1846 | return top; |
1847 | else |
1848 | return NULL; // Identity will return TOP |
1849 | } else if (uin != NULL) { |
1850 | // Only one not-NULL unique input path is left. |
1851 | // Determine if this input is backedge of a loop. |
1852 | // (Skip new phis which have no uses and dead regions). |
1853 | if (outcnt() > 0 && r->in(0) != NULL) { |
1854 | // First, take the short cut when we know it is a loop and |
1855 | // the EntryControl data path is dead. |
1856 | // Loop node may have only one input because entry path |
1857 | // is removed in PhaseIdealLoop::Dominators(). |
1858 | assert(!r->is_Loop() || r->req() <= 3, "Loop node should have 3 or less inputs" ); |
1859 | bool is_loop = (r->is_Loop() && r->req() == 3); |
1860 | // Then, check if there is a data loop when phi references itself directly |
1861 | // or through other data nodes. |
1862 | if ((is_loop && !uin->eqv_uncast(in(LoopNode::EntryControl))) || |
1863 | (!is_loop && is_unsafe_data_reference(uin))) { |
1864 | // Break this data loop to avoid creation of a dead loop. |
1865 | if (can_reshape) { |
1866 | return top; |
1867 | } else { |
1868 | // We can't return top if we are in Parse phase - cut inputs only |
1869 | // let Identity to handle the case. |
1870 | replace_edge(uin, top); |
1871 | return NULL; |
1872 | } |
1873 | } |
1874 | } |
1875 | |
1876 | if (uncasted) { |
1877 | // Add cast nodes between the phi to be removed and its unique input. |
1878 | // Wait until after parsing for the type information to propagate from the casts. |
1879 | assert(can_reshape, "Invalid during parsing" ); |
1880 | const Type* phi_type = bottom_type(); |
1881 | assert(phi_type->isa_int() || phi_type->isa_ptr(), "bad phi type" ); |
1882 | // Add casts to carry the control dependency of the Phi that is |
1883 | // going away |
1884 | Node* cast = NULL; |
1885 | if (phi_type->isa_int()) { |
1886 | cast = ConstraintCastNode::make_cast(Op_CastII, r, uin, phi_type, true); |
1887 | } else { |
1888 | const Type* uin_type = phase->type(uin); |
1889 | if (!phi_type->isa_oopptr() && !uin_type->isa_oopptr()) { |
1890 | cast = ConstraintCastNode::make_cast(Op_CastPP, r, uin, phi_type, true); |
1891 | } else { |
1892 | // Use a CastPP for a cast to not null and a CheckCastPP for |
1893 | // a cast to a new klass (and both if both null-ness and |
1894 | // klass change). |
1895 | |
1896 | // If the type of phi is not null but the type of uin may be |
1897 | // null, uin's type must be casted to not null |
1898 | if (phi_type->join(TypePtr::NOTNULL) == phi_type->remove_speculative() && |
1899 | uin_type->join(TypePtr::NOTNULL) != uin_type->remove_speculative()) { |
1900 | cast = ConstraintCastNode::make_cast(Op_CastPP, r, uin, TypePtr::NOTNULL, true); |
1901 | } |
1902 | |
1903 | // If the type of phi and uin, both casted to not null, |
1904 | // differ the klass of uin must be (check)cast'ed to match |
1905 | // that of phi |
1906 | if (phi_type->join_speculative(TypePtr::NOTNULL) != uin_type->join_speculative(TypePtr::NOTNULL)) { |
1907 | Node* n = uin; |
1908 | if (cast != NULL) { |
1909 | cast = phase->transform(cast); |
1910 | n = cast; |
1911 | } |
1912 | cast = ConstraintCastNode::make_cast(Op_CheckCastPP, r, n, phi_type, true); |
1913 | } |
1914 | if (cast == NULL) { |
1915 | cast = ConstraintCastNode::make_cast(Op_CastPP, r, uin, phi_type, true); |
1916 | } |
1917 | } |
1918 | } |
1919 | assert(cast != NULL, "cast should be set" ); |
1920 | cast = phase->transform(cast); |
1921 | // set all inputs to the new cast(s) so the Phi is removed by Identity |
1922 | PhaseIterGVN* igvn = phase->is_IterGVN(); |
1923 | for (uint i = 1; i < req(); i++) { |
1924 | set_req_X(i, cast, igvn); |
1925 | } |
1926 | uin = cast; |
1927 | } |
1928 | |
1929 | // One unique input. |
1930 | debug_only(Node* ident = Identity(phase)); |
1931 | // The unique input must eventually be detected by the Identity call. |
1932 | #ifdef ASSERT |
1933 | if (ident != uin && !ident->is_top()) { |
1934 | // print this output before failing assert |
1935 | r->dump(3); |
1936 | this->dump(3); |
1937 | ident->dump(); |
1938 | uin->dump(); |
1939 | } |
1940 | #endif |
1941 | assert(ident == uin || ident->is_top(), "Identity must clean this up" ); |
1942 | return NULL; |
1943 | } |
1944 | |
1945 | Node* opt = NULL; |
1946 | int true_path = is_diamond_phi(); |
1947 | if( true_path != 0 ) { |
1948 | // Check for CMove'ing identity. If it would be unsafe, |
1949 | // handle it here. In the safe case, let Identity handle it. |
1950 | Node* unsafe_id = is_cmove_id(phase, true_path); |
1951 | if( unsafe_id != NULL && is_unsafe_data_reference(unsafe_id) ) |
1952 | opt = unsafe_id; |
1953 | |
1954 | // Check for simple convert-to-boolean pattern |
1955 | if( opt == NULL ) |
1956 | opt = is_x2logic(phase, this, true_path); |
1957 | |
1958 | // Check for absolute value |
1959 | if( opt == NULL ) |
1960 | opt = is_absolute(phase, this, true_path); |
1961 | |
1962 | // Check for conditional add |
1963 | if( opt == NULL && can_reshape ) |
1964 | opt = is_cond_add(phase, this, true_path); |
1965 | |
1966 | // These 4 optimizations could subsume the phi: |
1967 | // have to check for a dead data loop creation. |
1968 | if( opt != NULL ) { |
1969 | if( opt == unsafe_id || is_unsafe_data_reference(opt) ) { |
1970 | // Found dead loop. |
1971 | if( can_reshape ) |
1972 | return top; |
1973 | // We can't return top if we are in Parse phase - cut inputs only |
1974 | // to stop further optimizations for this phi. Identity will return TOP. |
1975 | assert(req() == 3, "only diamond merge phi here" ); |
1976 | set_req(1, top); |
1977 | set_req(2, top); |
1978 | return NULL; |
1979 | } else { |
1980 | return opt; |
1981 | } |
1982 | } |
1983 | } |
1984 | |
1985 | // Check for merging identical values and split flow paths |
1986 | if (can_reshape) { |
1987 | opt = split_flow_path(phase, this); |
1988 | // This optimization only modifies phi - don't need to check for dead loop. |
1989 | assert(opt == NULL || phase->eqv(opt, this), "do not elide phi" ); |
1990 | if (opt != NULL) return opt; |
1991 | } |
1992 | |
1993 | if (in(1) != NULL && in(1)->Opcode() == Op_AddP && can_reshape) { |
1994 | // Try to undo Phi of AddP: |
1995 | // (Phi (AddP base base y) (AddP base2 base2 y)) |
1996 | // becomes: |
1997 | // newbase := (Phi base base2) |
1998 | // (AddP newbase newbase y) |
1999 | // |
2000 | // This occurs as a result of unsuccessful split_thru_phi and |
2001 | // interferes with taking advantage of addressing modes. See the |
2002 | // clone_shift_expressions code in matcher.cpp |
2003 | Node* addp = in(1); |
2004 | const Type* type = addp->in(AddPNode::Base)->bottom_type(); |
2005 | Node* y = addp->in(AddPNode::Offset); |
2006 | if (y != NULL && addp->in(AddPNode::Base) == addp->in(AddPNode::Address)) { |
2007 | // make sure that all the inputs are similar to the first one, |
2008 | // i.e. AddP with base == address and same offset as first AddP |
2009 | bool doit = true; |
2010 | for (uint i = 2; i < req(); i++) { |
2011 | if (in(i) == NULL || |
2012 | in(i)->Opcode() != Op_AddP || |
2013 | in(i)->in(AddPNode::Base) != in(i)->in(AddPNode::Address) || |
2014 | in(i)->in(AddPNode::Offset) != y) { |
2015 | doit = false; |
2016 | break; |
2017 | } |
2018 | // Accumulate type for resulting Phi |
2019 | type = type->meet_speculative(in(i)->in(AddPNode::Base)->bottom_type()); |
2020 | } |
2021 | Node* base = NULL; |
2022 | if (doit) { |
2023 | // Check for neighboring AddP nodes in a tree. |
2024 | // If they have a base, use that it. |
2025 | for (DUIterator_Fast kmax, k = this->fast_outs(kmax); k < kmax; k++) { |
2026 | Node* u = this->fast_out(k); |
2027 | if (u->is_AddP()) { |
2028 | Node* base2 = u->in(AddPNode::Base); |
2029 | if (base2 != NULL && !base2->is_top()) { |
2030 | if (base == NULL) |
2031 | base = base2; |
2032 | else if (base != base2) |
2033 | { doit = false; break; } |
2034 | } |
2035 | } |
2036 | } |
2037 | } |
2038 | if (doit) { |
2039 | if (base == NULL) { |
2040 | base = new PhiNode(in(0), type, NULL); |
2041 | for (uint i = 1; i < req(); i++) { |
2042 | base->init_req(i, in(i)->in(AddPNode::Base)); |
2043 | } |
2044 | phase->is_IterGVN()->register_new_node_with_optimizer(base); |
2045 | } |
2046 | return new AddPNode(base, base, y); |
2047 | } |
2048 | } |
2049 | } |
2050 | |
2051 | // Split phis through memory merges, so that the memory merges will go away. |
2052 | // Piggy-back this transformation on the search for a unique input.... |
2053 | // It will be as if the merged memory is the unique value of the phi. |
2054 | // (Do not attempt this optimization unless parsing is complete. |
2055 | // It would make the parser's memory-merge logic sick.) |
2056 | // (MergeMemNode is not dead_loop_safe - need to check for dead loop.) |
2057 | if (progress == NULL && can_reshape && type() == Type::MEMORY) { |
2058 | // see if this phi should be sliced |
2059 | uint merge_width = 0; |
2060 | bool saw_self = false; |
2061 | for( uint i=1; i<req(); ++i ) {// For all paths in |
2062 | Node *ii = in(i); |
2063 | // TOP inputs should not be counted as safe inputs because if the |
2064 | // Phi references itself through all other inputs then splitting the |
2065 | // Phi through memory merges would create dead loop at later stage. |
2066 | if (ii == top) { |
2067 | return NULL; // Delay optimization until graph is cleaned. |
2068 | } |
2069 | if (ii->is_MergeMem()) { |
2070 | MergeMemNode* n = ii->as_MergeMem(); |
2071 | merge_width = MAX2(merge_width, n->req()); |
2072 | saw_self = saw_self || phase->eqv(n->base_memory(), this); |
2073 | } |
2074 | } |
2075 | |
2076 | // This restriction is temporarily necessary to ensure termination: |
2077 | if (!saw_self && adr_type() == TypePtr::BOTTOM) merge_width = 0; |
2078 | |
2079 | if (merge_width > Compile::AliasIdxRaw) { |
2080 | // found at least one non-empty MergeMem |
2081 | const TypePtr* at = adr_type(); |
2082 | if (at != TypePtr::BOTTOM) { |
2083 | // Patch the existing phi to select an input from the merge: |
2084 | // Phi:AT1(...MergeMem(m0, m1, m2)...) into |
2085 | // Phi:AT1(...m1...) |
2086 | int alias_idx = phase->C->get_alias_index(at); |
2087 | for (uint i=1; i<req(); ++i) { |
2088 | Node *ii = in(i); |
2089 | if (ii->is_MergeMem()) { |
2090 | MergeMemNode* n = ii->as_MergeMem(); |
2091 | // compress paths and change unreachable cycles to TOP |
2092 | // If not, we can update the input infinitely along a MergeMem cycle |
2093 | // Equivalent code is in MemNode::Ideal_common |
2094 | Node *m = phase->transform(n); |
2095 | if (outcnt() == 0) { // Above transform() may kill us! |
2096 | return top; |
2097 | } |
2098 | // If transformed to a MergeMem, get the desired slice |
2099 | // Otherwise the returned node represents memory for every slice |
2100 | Node *new_mem = (m->is_MergeMem()) ? |
2101 | m->as_MergeMem()->memory_at(alias_idx) : m; |
2102 | // Update input if it is progress over what we have now |
2103 | if (new_mem != ii) { |
2104 | set_req(i, new_mem); |
2105 | progress = this; |
2106 | } |
2107 | } |
2108 | } |
2109 | } else { |
2110 | // We know that at least one MergeMem->base_memory() == this |
2111 | // (saw_self == true). If all other inputs also references this phi |
2112 | // (directly or through data nodes) - it is dead loop. |
2113 | bool saw_safe_input = false; |
2114 | for (uint j = 1; j < req(); ++j) { |
2115 | Node *n = in(j); |
2116 | if (n->is_MergeMem() && n->as_MergeMem()->base_memory() == this) |
2117 | continue; // skip known cases |
2118 | if (!is_unsafe_data_reference(n)) { |
2119 | saw_safe_input = true; // found safe input |
2120 | break; |
2121 | } |
2122 | } |
2123 | if (!saw_safe_input) |
2124 | return top; // all inputs reference back to this phi - dead loop |
2125 | |
2126 | // Phi(...MergeMem(m0, m1:AT1, m2:AT2)...) into |
2127 | // MergeMem(Phi(...m0...), Phi:AT1(...m1...), Phi:AT2(...m2...)) |
2128 | PhaseIterGVN *igvn = phase->is_IterGVN(); |
2129 | Node* hook = new Node(1); |
2130 | PhiNode* new_base = (PhiNode*) clone(); |
2131 | // Must eagerly register phis, since they participate in loops. |
2132 | if (igvn) { |
2133 | igvn->register_new_node_with_optimizer(new_base); |
2134 | hook->add_req(new_base); |
2135 | } |
2136 | MergeMemNode* result = MergeMemNode::make(new_base); |
2137 | for (uint i = 1; i < req(); ++i) { |
2138 | Node *ii = in(i); |
2139 | if (ii->is_MergeMem()) { |
2140 | MergeMemNode* n = ii->as_MergeMem(); |
2141 | for (MergeMemStream mms(result, n); mms.next_non_empty2(); ) { |
2142 | // If we have not seen this slice yet, make a phi for it. |
2143 | bool made_new_phi = false; |
2144 | if (mms.is_empty()) { |
2145 | Node* new_phi = new_base->slice_memory(mms.adr_type(phase->C)); |
2146 | made_new_phi = true; |
2147 | if (igvn) { |
2148 | igvn->register_new_node_with_optimizer(new_phi); |
2149 | hook->add_req(new_phi); |
2150 | } |
2151 | mms.set_memory(new_phi); |
2152 | } |
2153 | Node* phi = mms.memory(); |
2154 | assert(made_new_phi || phi->in(i) == n, "replace the i-th merge by a slice" ); |
2155 | phi->set_req(i, mms.memory2()); |
2156 | } |
2157 | } |
2158 | } |
2159 | // Distribute all self-loops. |
2160 | { // (Extra braces to hide mms.) |
2161 | for (MergeMemStream mms(result); mms.next_non_empty(); ) { |
2162 | Node* phi = mms.memory(); |
2163 | for (uint i = 1; i < req(); ++i) { |
2164 | if (phi->in(i) == this) phi->set_req(i, phi); |
2165 | } |
2166 | } |
2167 | } |
2168 | // now transform the new nodes, and return the mergemem |
2169 | for (MergeMemStream mms(result); mms.next_non_empty(); ) { |
2170 | Node* phi = mms.memory(); |
2171 | mms.set_memory(phase->transform(phi)); |
2172 | } |
2173 | if (igvn) { // Unhook. |
2174 | igvn->hash_delete(hook); |
2175 | for (uint i = 1; i < hook->req(); i++) { |
2176 | hook->set_req(i, NULL); |
2177 | } |
2178 | } |
2179 | // Replace self with the result. |
2180 | return result; |
2181 | } |
2182 | } |
2183 | // |
2184 | // Other optimizations on the memory chain |
2185 | // |
2186 | const TypePtr* at = adr_type(); |
2187 | for( uint i=1; i<req(); ++i ) {// For all paths in |
2188 | Node *ii = in(i); |
2189 | Node *new_in = MemNode::optimize_memory_chain(ii, at, NULL, phase); |
2190 | if (ii != new_in ) { |
2191 | set_req(i, new_in); |
2192 | progress = this; |
2193 | } |
2194 | } |
2195 | } |
2196 | |
2197 | #ifdef _LP64 |
2198 | // Push DecodeN/DecodeNKlass down through phi. |
2199 | // The rest of phi graph will transform by split EncodeP node though phis up. |
2200 | if ((UseCompressedOops || UseCompressedClassPointers) && can_reshape && progress == NULL) { |
2201 | bool may_push = true; |
2202 | bool has_decodeN = false; |
2203 | bool is_decodeN = false; |
2204 | for (uint i=1; i<req(); ++i) {// For all paths in |
2205 | Node *ii = in(i); |
2206 | if (ii->is_DecodeNarrowPtr() && ii->bottom_type() == bottom_type()) { |
2207 | // Do optimization if a non dead path exist. |
2208 | if (ii->in(1)->bottom_type() != Type::TOP) { |
2209 | has_decodeN = true; |
2210 | is_decodeN = ii->is_DecodeN(); |
2211 | } |
2212 | } else if (!ii->is_Phi()) { |
2213 | may_push = false; |
2214 | } |
2215 | } |
2216 | |
2217 | if (has_decodeN && may_push) { |
2218 | PhaseIterGVN *igvn = phase->is_IterGVN(); |
2219 | // Make narrow type for new phi. |
2220 | const Type* narrow_t; |
2221 | if (is_decodeN) { |
2222 | narrow_t = TypeNarrowOop::make(this->bottom_type()->is_ptr()); |
2223 | } else { |
2224 | narrow_t = TypeNarrowKlass::make(this->bottom_type()->is_ptr()); |
2225 | } |
2226 | PhiNode* new_phi = new PhiNode(r, narrow_t); |
2227 | uint orig_cnt = req(); |
2228 | for (uint i=1; i<req(); ++i) {// For all paths in |
2229 | Node *ii = in(i); |
2230 | Node* new_ii = NULL; |
2231 | if (ii->is_DecodeNarrowPtr()) { |
2232 | assert(ii->bottom_type() == bottom_type(), "sanity" ); |
2233 | new_ii = ii->in(1); |
2234 | } else { |
2235 | assert(ii->is_Phi(), "sanity" ); |
2236 | if (ii->as_Phi() == this) { |
2237 | new_ii = new_phi; |
2238 | } else { |
2239 | if (is_decodeN) { |
2240 | new_ii = new EncodePNode(ii, narrow_t); |
2241 | } else { |
2242 | new_ii = new EncodePKlassNode(ii, narrow_t); |
2243 | } |
2244 | igvn->register_new_node_with_optimizer(new_ii); |
2245 | } |
2246 | } |
2247 | new_phi->set_req(i, new_ii); |
2248 | } |
2249 | igvn->register_new_node_with_optimizer(new_phi, this); |
2250 | if (is_decodeN) { |
2251 | progress = new DecodeNNode(new_phi, bottom_type()); |
2252 | } else { |
2253 | progress = new DecodeNKlassNode(new_phi, bottom_type()); |
2254 | } |
2255 | } |
2256 | } |
2257 | #endif |
2258 | |
2259 | return progress; // Return any progress |
2260 | } |
2261 | |
2262 | //------------------------------is_tripcount----------------------------------- |
2263 | bool PhiNode::is_tripcount() const { |
2264 | return (in(0) != NULL && in(0)->is_CountedLoop() && |
2265 | in(0)->as_CountedLoop()->phi() == this); |
2266 | } |
2267 | |
2268 | //------------------------------out_RegMask------------------------------------ |
2269 | const RegMask &PhiNode::in_RegMask(uint i) const { |
2270 | return i ? out_RegMask() : RegMask::Empty; |
2271 | } |
2272 | |
2273 | const RegMask &PhiNode::out_RegMask() const { |
2274 | uint ideal_reg = _type->ideal_reg(); |
2275 | assert( ideal_reg != Node::NotAMachineReg, "invalid type at Phi" ); |
2276 | if( ideal_reg == 0 ) return RegMask::Empty; |
2277 | assert(ideal_reg != Op_RegFlags, "flags register is not spillable" ); |
2278 | return *(Compile::current()->matcher()->idealreg2spillmask[ideal_reg]); |
2279 | } |
2280 | |
2281 | #ifndef PRODUCT |
2282 | void PhiNode::related(GrowableArray<Node*> *in_rel, GrowableArray<Node*> *out_rel, bool compact) const { |
2283 | // For a PhiNode, the set of related nodes includes all inputs till level 2, |
2284 | // and all outputs till level 1. In compact mode, inputs till level 1 are |
2285 | // collected. |
2286 | this->collect_nodes(in_rel, compact ? 1 : 2, false, false); |
2287 | this->collect_nodes(out_rel, -1, false, false); |
2288 | } |
2289 | |
2290 | void PhiNode::dump_spec(outputStream *st) const { |
2291 | TypeNode::dump_spec(st); |
2292 | if (is_tripcount()) { |
2293 | st->print(" #tripcount" ); |
2294 | } |
2295 | } |
2296 | #endif |
2297 | |
2298 | |
2299 | //============================================================================= |
2300 | const Type* GotoNode::Value(PhaseGVN* phase) const { |
2301 | // If the input is reachable, then we are executed. |
2302 | // If the input is not reachable, then we are not executed. |
2303 | return phase->type(in(0)); |
2304 | } |
2305 | |
2306 | Node* GotoNode::Identity(PhaseGVN* phase) { |
2307 | return in(0); // Simple copy of incoming control |
2308 | } |
2309 | |
2310 | const RegMask &GotoNode::out_RegMask() const { |
2311 | return RegMask::Empty; |
2312 | } |
2313 | |
2314 | #ifndef PRODUCT |
2315 | //-----------------------------related----------------------------------------- |
2316 | // The related nodes of a GotoNode are all inputs at level 1, as well as the |
2317 | // outputs at level 1. This is regardless of compact mode. |
2318 | void GotoNode::related(GrowableArray<Node*> *in_rel, GrowableArray<Node*> *out_rel, bool compact) const { |
2319 | this->collect_nodes(in_rel, 1, false, false); |
2320 | this->collect_nodes(out_rel, -1, false, false); |
2321 | } |
2322 | #endif |
2323 | |
2324 | |
2325 | //============================================================================= |
2326 | const RegMask &JumpNode::out_RegMask() const { |
2327 | return RegMask::Empty; |
2328 | } |
2329 | |
2330 | #ifndef PRODUCT |
2331 | //-----------------------------related----------------------------------------- |
2332 | // The related nodes of a JumpNode are all inputs at level 1, as well as the |
2333 | // outputs at level 2 (to include actual jump targets beyond projection nodes). |
2334 | // This is regardless of compact mode. |
2335 | void JumpNode::related(GrowableArray<Node*> *in_rel, GrowableArray<Node*> *out_rel, bool compact) const { |
2336 | this->collect_nodes(in_rel, 1, false, false); |
2337 | this->collect_nodes(out_rel, -2, false, false); |
2338 | } |
2339 | #endif |
2340 | |
2341 | //============================================================================= |
2342 | const RegMask &JProjNode::out_RegMask() const { |
2343 | return RegMask::Empty; |
2344 | } |
2345 | |
2346 | //============================================================================= |
2347 | const RegMask &CProjNode::out_RegMask() const { |
2348 | return RegMask::Empty; |
2349 | } |
2350 | |
2351 | |
2352 | |
2353 | //============================================================================= |
2354 | |
2355 | uint PCTableNode::hash() const { return Node::hash() + _size; } |
2356 | bool PCTableNode::cmp( const Node &n ) const |
2357 | { return _size == ((PCTableNode&)n)._size; } |
2358 | |
2359 | const Type *PCTableNode::bottom_type() const { |
2360 | const Type** f = TypeTuple::fields(_size); |
2361 | for( uint i = 0; i < _size; i++ ) f[i] = Type::CONTROL; |
2362 | return TypeTuple::make(_size, f); |
2363 | } |
2364 | |
2365 | //------------------------------Value------------------------------------------ |
2366 | // Compute the type of the PCTableNode. If reachable it is a tuple of |
2367 | // Control, otherwise the table targets are not reachable |
2368 | const Type* PCTableNode::Value(PhaseGVN* phase) const { |
2369 | if( phase->type(in(0)) == Type::CONTROL ) |
2370 | return bottom_type(); |
2371 | return Type::TOP; // All paths dead? Then so are we |
2372 | } |
2373 | |
2374 | //------------------------------Ideal------------------------------------------ |
2375 | // Return a node which is more "ideal" than the current node. Strip out |
2376 | // control copies |
2377 | Node *PCTableNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
2378 | return remove_dead_region(phase, can_reshape) ? this : NULL; |
2379 | } |
2380 | |
2381 | //============================================================================= |
2382 | uint JumpProjNode::hash() const { |
2383 | return Node::hash() + _dest_bci; |
2384 | } |
2385 | |
2386 | bool JumpProjNode::cmp( const Node &n ) const { |
2387 | return ProjNode::cmp(n) && |
2388 | _dest_bci == ((JumpProjNode&)n)._dest_bci; |
2389 | } |
2390 | |
2391 | #ifndef PRODUCT |
2392 | void JumpProjNode::dump_spec(outputStream *st) const { |
2393 | ProjNode::dump_spec(st); |
2394 | st->print("@bci %d " ,_dest_bci); |
2395 | } |
2396 | |
2397 | void JumpProjNode::dump_compact_spec(outputStream *st) const { |
2398 | ProjNode::dump_compact_spec(st); |
2399 | st->print("(%d)%d@%d" , _switch_val, _proj_no, _dest_bci); |
2400 | } |
2401 | |
2402 | void JumpProjNode::related(GrowableArray<Node*> *in_rel, GrowableArray<Node*> *out_rel, bool compact) const { |
2403 | // The related nodes of a JumpProjNode are its inputs and outputs at level 1. |
2404 | this->collect_nodes(in_rel, 1, false, false); |
2405 | this->collect_nodes(out_rel, -1, false, false); |
2406 | } |
2407 | #endif |
2408 | |
2409 | //============================================================================= |
2410 | //------------------------------Value------------------------------------------ |
2411 | // Check for being unreachable, or for coming from a Rethrow. Rethrow's cannot |
2412 | // have the default "fall_through_index" path. |
2413 | const Type* CatchNode::Value(PhaseGVN* phase) const { |
2414 | // Unreachable? Then so are all paths from here. |
2415 | if( phase->type(in(0)) == Type::TOP ) return Type::TOP; |
2416 | // First assume all paths are reachable |
2417 | const Type** f = TypeTuple::fields(_size); |
2418 | for( uint i = 0; i < _size; i++ ) f[i] = Type::CONTROL; |
2419 | // Identify cases that will always throw an exception |
2420 | // () rethrow call |
2421 | // () virtual or interface call with NULL receiver |
2422 | // () call is a check cast with incompatible arguments |
2423 | if( in(1)->is_Proj() ) { |
2424 | Node *i10 = in(1)->in(0); |
2425 | if( i10->is_Call() ) { |
2426 | CallNode *call = i10->as_Call(); |
2427 | // Rethrows always throw exceptions, never return |
2428 | if (call->entry_point() == OptoRuntime::rethrow_stub()) { |
2429 | f[CatchProjNode::fall_through_index] = Type::TOP; |
2430 | } else if( call->req() > TypeFunc::Parms ) { |
2431 | const Type *arg0 = phase->type( call->in(TypeFunc::Parms) ); |
2432 | // Check for null receiver to virtual or interface calls |
2433 | if( call->is_CallDynamicJava() && |
2434 | arg0->higher_equal(TypePtr::NULL_PTR) ) { |
2435 | f[CatchProjNode::fall_through_index] = Type::TOP; |
2436 | } |
2437 | } // End of if not a runtime stub |
2438 | } // End of if have call above me |
2439 | } // End of slot 1 is not a projection |
2440 | return TypeTuple::make(_size, f); |
2441 | } |
2442 | |
2443 | //============================================================================= |
2444 | uint CatchProjNode::hash() const { |
2445 | return Node::hash() + _handler_bci; |
2446 | } |
2447 | |
2448 | |
2449 | bool CatchProjNode::cmp( const Node &n ) const { |
2450 | return ProjNode::cmp(n) && |
2451 | _handler_bci == ((CatchProjNode&)n)._handler_bci; |
2452 | } |
2453 | |
2454 | |
2455 | //------------------------------Identity--------------------------------------- |
2456 | // If only 1 target is possible, choose it if it is the main control |
2457 | Node* CatchProjNode::Identity(PhaseGVN* phase) { |
2458 | // If my value is control and no other value is, then treat as ID |
2459 | const TypeTuple *t = phase->type(in(0))->is_tuple(); |
2460 | if (t->field_at(_con) != Type::CONTROL) return this; |
2461 | // If we remove the last CatchProj and elide the Catch/CatchProj, then we |
2462 | // also remove any exception table entry. Thus we must know the call |
2463 | // feeding the Catch will not really throw an exception. This is ok for |
2464 | // the main fall-thru control (happens when we know a call can never throw |
2465 | // an exception) or for "rethrow", because a further optimization will |
2466 | // yank the rethrow (happens when we inline a function that can throw an |
2467 | // exception and the caller has no handler). Not legal, e.g., for passing |
2468 | // a NULL receiver to a v-call, or passing bad types to a slow-check-cast. |
2469 | // These cases MUST throw an exception via the runtime system, so the VM |
2470 | // will be looking for a table entry. |
2471 | Node *proj = in(0)->in(1); // Expect a proj feeding CatchNode |
2472 | CallNode *call; |
2473 | if (_con != TypeFunc::Control && // Bail out if not the main control. |
2474 | !(proj->is_Proj() && // AND NOT a rethrow |
2475 | proj->in(0)->is_Call() && |
2476 | (call = proj->in(0)->as_Call()) && |
2477 | call->entry_point() == OptoRuntime::rethrow_stub())) |
2478 | return this; |
2479 | |
2480 | // Search for any other path being control |
2481 | for (uint i = 0; i < t->cnt(); i++) { |
2482 | if (i != _con && t->field_at(i) == Type::CONTROL) |
2483 | return this; |
2484 | } |
2485 | // Only my path is possible; I am identity on control to the jump |
2486 | return in(0)->in(0); |
2487 | } |
2488 | |
2489 | |
2490 | #ifndef PRODUCT |
2491 | void CatchProjNode::dump_spec(outputStream *st) const { |
2492 | ProjNode::dump_spec(st); |
2493 | st->print("@bci %d " ,_handler_bci); |
2494 | } |
2495 | #endif |
2496 | |
2497 | //============================================================================= |
2498 | //------------------------------Identity--------------------------------------- |
2499 | // Check for CreateEx being Identity. |
2500 | Node* CreateExNode::Identity(PhaseGVN* phase) { |
2501 | if( phase->type(in(1)) == Type::TOP ) return in(1); |
2502 | if( phase->type(in(0)) == Type::TOP ) return in(0); |
2503 | // We only come from CatchProj, unless the CatchProj goes away. |
2504 | // If the CatchProj is optimized away, then we just carry the |
2505 | // exception oop through. |
2506 | CallNode *call = in(1)->in(0)->as_Call(); |
2507 | |
2508 | return ( in(0)->is_CatchProj() && in(0)->in(0)->in(1) == in(1) ) |
2509 | ? this |
2510 | : call->in(TypeFunc::Parms); |
2511 | } |
2512 | |
2513 | //============================================================================= |
2514 | //------------------------------Value------------------------------------------ |
2515 | // Check for being unreachable. |
2516 | const Type* NeverBranchNode::Value(PhaseGVN* phase) const { |
2517 | if (!in(0) || in(0)->is_top()) return Type::TOP; |
2518 | return bottom_type(); |
2519 | } |
2520 | |
2521 | //------------------------------Ideal------------------------------------------ |
2522 | // Check for no longer being part of a loop |
2523 | Node *NeverBranchNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
2524 | if (can_reshape && !in(0)->is_Loop()) { |
2525 | // Dead code elimination can sometimes delete this projection so |
2526 | // if it's not there, there's nothing to do. |
2527 | Node* fallthru = proj_out_or_null(0); |
2528 | if (fallthru != NULL) { |
2529 | phase->is_IterGVN()->replace_node(fallthru, in(0)); |
2530 | } |
2531 | return phase->C->top(); |
2532 | } |
2533 | return NULL; |
2534 | } |
2535 | |
2536 | #ifndef PRODUCT |
2537 | void NeverBranchNode::format( PhaseRegAlloc *ra_, outputStream *st) const { |
2538 | st->print("%s" , Name()); |
2539 | } |
2540 | #endif |
2541 | |