1 | /* |
2 | * Copyright (c) 2000, 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 "ci/ciTypeFlow.hpp" |
27 | #include "memory/allocation.inline.hpp" |
28 | #include "memory/resourceArea.hpp" |
29 | #include "opto/addnode.hpp" |
30 | #include "opto/castnode.hpp" |
31 | #include "opto/cfgnode.hpp" |
32 | #include "opto/connode.hpp" |
33 | #include "opto/loopnode.hpp" |
34 | #include "opto/phaseX.hpp" |
35 | #include "opto/runtime.hpp" |
36 | #include "opto/rootnode.hpp" |
37 | #include "opto/subnode.hpp" |
38 | |
39 | // Portions of code courtesy of Clifford Click |
40 | |
41 | // Optimization - Graph Style |
42 | |
43 | |
44 | #ifndef PRODUCT |
45 | extern int explicit_null_checks_elided; |
46 | #endif |
47 | |
48 | //============================================================================= |
49 | //------------------------------Value------------------------------------------ |
50 | // Return a tuple for whichever arm of the IF is reachable |
51 | const Type* IfNode::Value(PhaseGVN* phase) const { |
52 | if( !in(0) ) return Type::TOP; |
53 | if( phase->type(in(0)) == Type::TOP ) |
54 | return Type::TOP; |
55 | const Type *t = phase->type(in(1)); |
56 | if( t == Type::TOP ) // data is undefined |
57 | return TypeTuple::IFNEITHER; // unreachable altogether |
58 | if( t == TypeInt::ZERO ) // zero, or false |
59 | return TypeTuple::IFFALSE; // only false branch is reachable |
60 | if( t == TypeInt::ONE ) // 1, or true |
61 | return TypeTuple::IFTRUE; // only true branch is reachable |
62 | assert( t == TypeInt::BOOL, "expected boolean type" ); |
63 | |
64 | return TypeTuple::IFBOTH; // No progress |
65 | } |
66 | |
67 | const RegMask &IfNode::out_RegMask() const { |
68 | return RegMask::Empty; |
69 | } |
70 | |
71 | //------------------------------split_if--------------------------------------- |
72 | // Look for places where we merge constants, then test on the merged value. |
73 | // If the IF test will be constant folded on the path with the constant, we |
74 | // win by splitting the IF to before the merge point. |
75 | static Node* split_if(IfNode *iff, PhaseIterGVN *igvn) { |
76 | // I could be a lot more general here, but I'm trying to squeeze this |
77 | // in before the Christmas '98 break so I'm gonna be kinda restrictive |
78 | // on the patterns I accept. CNC |
79 | |
80 | // Look for a compare of a constant and a merged value |
81 | Node *i1 = iff->in(1); |
82 | if( !i1->is_Bool() ) return NULL; |
83 | BoolNode *b = i1->as_Bool(); |
84 | Node *cmp = b->in(1); |
85 | if( !cmp->is_Cmp() ) return NULL; |
86 | i1 = cmp->in(1); |
87 | if( i1 == NULL || !i1->is_Phi() ) return NULL; |
88 | PhiNode *phi = i1->as_Phi(); |
89 | if( phi->is_copy() ) return NULL; |
90 | Node *con2 = cmp->in(2); |
91 | if( !con2->is_Con() ) return NULL; |
92 | // See that the merge point contains some constants |
93 | Node *con1=NULL; |
94 | uint i4; |
95 | for( i4 = 1; i4 < phi->req(); i4++ ) { |
96 | con1 = phi->in(i4); |
97 | if( !con1 ) return NULL; // Do not optimize partially collapsed merges |
98 | if( con1->is_Con() ) break; // Found a constant |
99 | // Also allow null-vs-not-null checks |
100 | const TypePtr *tp = igvn->type(con1)->isa_ptr(); |
101 | if( tp && tp->_ptr == TypePtr::NotNull ) |
102 | break; |
103 | } |
104 | if( i4 >= phi->req() ) return NULL; // Found no constants |
105 | |
106 | igvn->C->set_has_split_ifs(true); // Has chance for split-if |
107 | |
108 | // Make sure that the compare can be constant folded away |
109 | Node *cmp2 = cmp->clone(); |
110 | cmp2->set_req(1,con1); |
111 | cmp2->set_req(2,con2); |
112 | const Type *t = cmp2->Value(igvn); |
113 | // This compare is dead, so whack it! |
114 | igvn->remove_dead_node(cmp2); |
115 | if( !t->singleton() ) return NULL; |
116 | |
117 | // No intervening control, like a simple Call |
118 | Node *r = iff->in(0); |
119 | if( !r->is_Region() ) return NULL; |
120 | if (r->is_Loop() && r->in(LoopNode::LoopBackControl)->is_top()) return NULL; // going away anyway |
121 | if( phi->region() != r ) return NULL; |
122 | // No other users of the cmp/bool |
123 | if (b->outcnt() != 1 || cmp->outcnt() != 1) { |
124 | //tty->print_cr("many users of cmp/bool"); |
125 | return NULL; |
126 | } |
127 | |
128 | // Make sure we can determine where all the uses of merged values go |
129 | for (DUIterator_Fast jmax, j = r->fast_outs(jmax); j < jmax; j++) { |
130 | Node* u = r->fast_out(j); |
131 | if( u == r ) continue; |
132 | if( u == iff ) continue; |
133 | if( u->outcnt() == 0 ) continue; // use is dead & ignorable |
134 | if( !u->is_Phi() ) { |
135 | /* |
136 | if( u->is_Start() ) { |
137 | tty->print_cr("Region has inlined start use"); |
138 | } else { |
139 | tty->print_cr("Region has odd use"); |
140 | u->dump(2); |
141 | }*/ |
142 | return NULL; |
143 | } |
144 | if( u != phi ) { |
145 | // CNC - do not allow any other merged value |
146 | //tty->print_cr("Merging another value"); |
147 | //u->dump(2); |
148 | return NULL; |
149 | } |
150 | // Make sure we can account for all Phi uses |
151 | for (DUIterator_Fast kmax, k = u->fast_outs(kmax); k < kmax; k++) { |
152 | Node* v = u->fast_out(k); // User of the phi |
153 | // CNC - Allow only really simple patterns. |
154 | // In particular I disallow AddP of the Phi, a fairly common pattern |
155 | if (v == cmp) continue; // The compare is OK |
156 | if (v->is_ConstraintCast()) { |
157 | // If the cast is derived from data flow edges, it may not have a control edge. |
158 | // If so, it should be safe to split. But follow-up code can not deal with |
159 | // this (l. 359). So skip. |
160 | if (v->in(0) == NULL) { |
161 | return NULL; |
162 | } |
163 | if (v->in(0)->in(0) == iff) { |
164 | continue; // CastPP/II of the IfNode is OK |
165 | } |
166 | } |
167 | // Disabled following code because I cannot tell if exactly one |
168 | // path dominates without a real dominator check. CNC 9/9/1999 |
169 | //uint vop = v->Opcode(); |
170 | //if( vop == Op_Phi ) { // Phi from another merge point might be OK |
171 | // Node *r = v->in(0); // Get controlling point |
172 | // if( !r ) return NULL; // Degraded to a copy |
173 | // // Find exactly one path in (either True or False doms, but not IFF) |
174 | // int cnt = 0; |
175 | // for( uint i = 1; i < r->req(); i++ ) |
176 | // if( r->in(i) && r->in(i)->in(0) == iff ) |
177 | // cnt++; |
178 | // if( cnt == 1 ) continue; // Exactly one of True or False guards Phi |
179 | //} |
180 | if( !v->is_Call() ) { |
181 | /* |
182 | if( v->Opcode() == Op_AddP ) { |
183 | tty->print_cr("Phi has AddP use"); |
184 | } else if( v->Opcode() == Op_CastPP ) { |
185 | tty->print_cr("Phi has CastPP use"); |
186 | } else if( v->Opcode() == Op_CastII ) { |
187 | tty->print_cr("Phi has CastII use"); |
188 | } else { |
189 | tty->print_cr("Phi has use I cant be bothered with"); |
190 | } |
191 | */ |
192 | } |
193 | return NULL; |
194 | |
195 | /* CNC - Cut out all the fancy acceptance tests |
196 | // Can we clone this use when doing the transformation? |
197 | // If all uses are from Phis at this merge or constants, then YES. |
198 | if( !v->in(0) && v != cmp ) { |
199 | tty->print_cr("Phi has free-floating use"); |
200 | v->dump(2); |
201 | return NULL; |
202 | } |
203 | for( uint l = 1; l < v->req(); l++ ) { |
204 | if( (!v->in(l)->is_Phi() || v->in(l)->in(0) != r) && |
205 | !v->in(l)->is_Con() ) { |
206 | tty->print_cr("Phi has use"); |
207 | v->dump(2); |
208 | return NULL; |
209 | } // End of if Phi-use input is neither Phi nor Constant |
210 | } // End of for all inputs to Phi-use |
211 | */ |
212 | } // End of for all uses of Phi |
213 | } // End of for all uses of Region |
214 | |
215 | // Only do this if the IF node is in a sane state |
216 | if (iff->outcnt() != 2) |
217 | return NULL; |
218 | |
219 | // Got a hit! Do the Mondo Hack! |
220 | // |
221 | //ABC a1c def ghi B 1 e h A C a c d f g i |
222 | // R - Phi - Phi - Phi Rc - Phi - Phi - Phi Rx - Phi - Phi - Phi |
223 | // cmp - 2 cmp - 2 cmp - 2 |
224 | // bool bool_c bool_x |
225 | // if if_c if_x |
226 | // T F T F T F |
227 | // ..s.. ..t .. ..s.. ..t.. ..s.. ..t.. |
228 | // |
229 | // Split the paths coming into the merge point into 2 separate groups of |
230 | // merges. On the left will be all the paths feeding constants into the |
231 | // Cmp's Phi. On the right will be the remaining paths. The Cmp's Phi |
232 | // will fold up into a constant; this will let the Cmp fold up as well as |
233 | // all the control flow. Below the original IF we have 2 control |
234 | // dependent regions, 's' and 't'. Now we will merge the two paths |
235 | // just prior to 's' and 't' from the two IFs. At least 1 path (and quite |
236 | // likely 2 or more) will promptly constant fold away. |
237 | PhaseGVN *phase = igvn; |
238 | |
239 | // Make a region merging constants and a region merging the rest |
240 | uint req_c = 0; |
241 | Node* predicate_proj = NULL; |
242 | int nb_predicate_proj = 0; |
243 | for (uint ii = 1; ii < r->req(); ii++) { |
244 | if (phi->in(ii) == con1) { |
245 | req_c++; |
246 | } |
247 | Node* proj = PhaseIdealLoop::find_predicate(r->in(ii)); |
248 | if (proj != NULL) { |
249 | nb_predicate_proj++; |
250 | predicate_proj = proj; |
251 | } |
252 | } |
253 | |
254 | // If all the defs of the phi are the same constant, we already have the desired end state. |
255 | // Skip the split that would create empty phi and region nodes. |
256 | if((r->req() - req_c) == 1) { |
257 | return NULL; |
258 | } |
259 | |
260 | if (nb_predicate_proj > 1) { |
261 | // Can happen in case of loop unswitching and when the loop is |
262 | // optimized out: it's not a loop anymore so we don't care about |
263 | // predicates. |
264 | assert(!r->is_Loop(), "this must not be a loop anymore" ); |
265 | predicate_proj = NULL; |
266 | } |
267 | Node* predicate_c = NULL; |
268 | Node* predicate_x = NULL; |
269 | bool counted_loop = r->is_CountedLoop(); |
270 | if (counted_loop) { |
271 | // Ignore counted loops for now because the split-if logic does not work |
272 | // in all the cases (for example, with strip mined loops). Also, above |
273 | // checks only pass for already degraded loops without a tripcount phi |
274 | // and these are essentially dead and will go away during igvn. |
275 | return NULL; |
276 | } |
277 | |
278 | Node *region_c = new RegionNode(req_c + 1); |
279 | Node *phi_c = con1; |
280 | uint len = r->req(); |
281 | Node *region_x = new RegionNode(len - req_c); |
282 | Node *phi_x = PhiNode::make_blank(region_x, phi); |
283 | for (uint i = 1, i_c = 1, i_x = 1; i < len; i++) { |
284 | if (phi->in(i) == con1) { |
285 | region_c->init_req( i_c++, r ->in(i) ); |
286 | if (r->in(i) == predicate_proj) |
287 | predicate_c = predicate_proj; |
288 | } else { |
289 | region_x->init_req( i_x, r ->in(i) ); |
290 | phi_x ->init_req( i_x++, phi->in(i) ); |
291 | if (r->in(i) == predicate_proj) |
292 | predicate_x = predicate_proj; |
293 | } |
294 | } |
295 | if (predicate_c != NULL && (req_c > 1)) { |
296 | assert(predicate_x == NULL, "only one predicate entry expected" ); |
297 | predicate_c = NULL; // Do not clone predicate below merge point |
298 | } |
299 | if (predicate_x != NULL && ((len - req_c) > 2)) { |
300 | assert(predicate_c == NULL, "only one predicate entry expected" ); |
301 | predicate_x = NULL; // Do not clone predicate below merge point |
302 | } |
303 | |
304 | // Register the new RegionNodes but do not transform them. Cannot |
305 | // transform until the entire Region/Phi conglomerate has been hacked |
306 | // as a single huge transform. |
307 | igvn->register_new_node_with_optimizer( region_c ); |
308 | igvn->register_new_node_with_optimizer( region_x ); |
309 | // Prevent the untimely death of phi_x. Currently he has no uses. He is |
310 | // about to get one. If this only use goes away, then phi_x will look dead. |
311 | // However, he will be picking up some more uses down below. |
312 | Node *hook = new Node(4); |
313 | hook->init_req(0, phi_x); |
314 | hook->init_req(1, phi_c); |
315 | phi_x = phase->transform( phi_x ); |
316 | |
317 | // Make the compare |
318 | Node *cmp_c = phase->makecon(t); |
319 | Node *cmp_x = cmp->clone(); |
320 | cmp_x->set_req(1,phi_x); |
321 | cmp_x->set_req(2,con2); |
322 | cmp_x = phase->transform(cmp_x); |
323 | // Make the bool |
324 | Node *b_c = phase->transform(new BoolNode(cmp_c,b->_test._test)); |
325 | Node *b_x = phase->transform(new BoolNode(cmp_x,b->_test._test)); |
326 | // Make the IfNode |
327 | IfNode* iff_c = iff->clone()->as_If(); |
328 | iff_c->set_req(0, region_c); |
329 | iff_c->set_req(1, b_c); |
330 | igvn->set_type_bottom(iff_c); |
331 | igvn->_worklist.push(iff_c); |
332 | hook->init_req(2, iff_c); |
333 | |
334 | IfNode* iff_x = iff->clone()->as_If(); |
335 | iff_x->set_req(0, region_x); |
336 | iff_x->set_req(1, b_x); |
337 | igvn->set_type_bottom(iff_x); |
338 | igvn->_worklist.push(iff_x); |
339 | hook->init_req(3, iff_x); |
340 | |
341 | // Make the true/false arms |
342 | Node *iff_c_t = phase->transform(new IfTrueNode (iff_c)); |
343 | Node *iff_c_f = phase->transform(new IfFalseNode(iff_c)); |
344 | if (predicate_c != NULL) { |
345 | assert(predicate_x == NULL, "only one predicate entry expected" ); |
346 | // Clone loop predicates to each path |
347 | iff_c_t = igvn->clone_loop_predicates(predicate_c, iff_c_t, !counted_loop); |
348 | iff_c_f = igvn->clone_loop_predicates(predicate_c, iff_c_f, !counted_loop); |
349 | } |
350 | Node *iff_x_t = phase->transform(new IfTrueNode (iff_x)); |
351 | Node *iff_x_f = phase->transform(new IfFalseNode(iff_x)); |
352 | if (predicate_x != NULL) { |
353 | assert(predicate_c == NULL, "only one predicate entry expected" ); |
354 | // Clone loop predicates to each path |
355 | iff_x_t = igvn->clone_loop_predicates(predicate_x, iff_x_t, !counted_loop); |
356 | iff_x_f = igvn->clone_loop_predicates(predicate_x, iff_x_f, !counted_loop); |
357 | } |
358 | |
359 | // Merge the TRUE paths |
360 | Node *region_s = new RegionNode(3); |
361 | igvn->_worklist.push(region_s); |
362 | region_s->init_req(1, iff_c_t); |
363 | region_s->init_req(2, iff_x_t); |
364 | igvn->register_new_node_with_optimizer( region_s ); |
365 | |
366 | // Merge the FALSE paths |
367 | Node *region_f = new RegionNode(3); |
368 | igvn->_worklist.push(region_f); |
369 | region_f->init_req(1, iff_c_f); |
370 | region_f->init_req(2, iff_x_f); |
371 | igvn->register_new_node_with_optimizer( region_f ); |
372 | |
373 | igvn->hash_delete(cmp);// Remove soon-to-be-dead node from hash table. |
374 | cmp->set_req(1,NULL); // Whack the inputs to cmp because it will be dead |
375 | cmp->set_req(2,NULL); |
376 | // Check for all uses of the Phi and give them a new home. |
377 | // The 'cmp' got cloned, but CastPP/IIs need to be moved. |
378 | Node *phi_s = NULL; // do not construct unless needed |
379 | Node *phi_f = NULL; // do not construct unless needed |
380 | for (DUIterator_Last i2min, i2 = phi->last_outs(i2min); i2 >= i2min; --i2) { |
381 | Node* v = phi->last_out(i2);// User of the phi |
382 | igvn->rehash_node_delayed(v); // Have to fixup other Phi users |
383 | uint vop = v->Opcode(); |
384 | Node *proj = NULL; |
385 | if( vop == Op_Phi ) { // Remote merge point |
386 | Node *r = v->in(0); |
387 | for (uint i3 = 1; i3 < r->req(); i3++) |
388 | if (r->in(i3) && r->in(i3)->in(0) == iff) { |
389 | proj = r->in(i3); |
390 | break; |
391 | } |
392 | } else if( v->is_ConstraintCast() ) { |
393 | proj = v->in(0); // Controlling projection |
394 | } else { |
395 | assert( 0, "do not know how to handle this guy" ); |
396 | } |
397 | guarantee(proj != NULL, "sanity" ); |
398 | |
399 | Node *proj_path_data, *proj_path_ctrl; |
400 | if( proj->Opcode() == Op_IfTrue ) { |
401 | if( phi_s == NULL ) { |
402 | // Only construct phi_s if needed, otherwise provides |
403 | // interfering use. |
404 | phi_s = PhiNode::make_blank(region_s,phi); |
405 | phi_s->init_req( 1, phi_c ); |
406 | phi_s->init_req( 2, phi_x ); |
407 | hook->add_req(phi_s); |
408 | phi_s = phase->transform(phi_s); |
409 | } |
410 | proj_path_data = phi_s; |
411 | proj_path_ctrl = region_s; |
412 | } else { |
413 | if( phi_f == NULL ) { |
414 | // Only construct phi_f if needed, otherwise provides |
415 | // interfering use. |
416 | phi_f = PhiNode::make_blank(region_f,phi); |
417 | phi_f->init_req( 1, phi_c ); |
418 | phi_f->init_req( 2, phi_x ); |
419 | hook->add_req(phi_f); |
420 | phi_f = phase->transform(phi_f); |
421 | } |
422 | proj_path_data = phi_f; |
423 | proj_path_ctrl = region_f; |
424 | } |
425 | |
426 | // Fixup 'v' for for the split |
427 | if( vop == Op_Phi ) { // Remote merge point |
428 | uint i; |
429 | for( i = 1; i < v->req(); i++ ) |
430 | if( v->in(i) == phi ) |
431 | break; |
432 | v->set_req(i, proj_path_data ); |
433 | } else if( v->is_ConstraintCast() ) { |
434 | v->set_req(0, proj_path_ctrl ); |
435 | v->set_req(1, proj_path_data ); |
436 | } else |
437 | ShouldNotReachHere(); |
438 | } |
439 | |
440 | // Now replace the original iff's True/False with region_s/region_t. |
441 | // This makes the original iff go dead. |
442 | for (DUIterator_Last i3min, i3 = iff->last_outs(i3min); i3 >= i3min; --i3) { |
443 | Node* p = iff->last_out(i3); |
444 | assert( p->Opcode() == Op_IfTrue || p->Opcode() == Op_IfFalse, "" ); |
445 | Node *u = (p->Opcode() == Op_IfTrue) ? region_s : region_f; |
446 | // Replace p with u |
447 | igvn->add_users_to_worklist(p); |
448 | for (DUIterator_Last lmin, l = p->last_outs(lmin); l >= lmin;) { |
449 | Node* x = p->last_out(l); |
450 | igvn->hash_delete(x); |
451 | uint uses_found = 0; |
452 | for( uint j = 0; j < x->req(); j++ ) { |
453 | if( x->in(j) == p ) { |
454 | x->set_req(j, u); |
455 | uses_found++; |
456 | } |
457 | } |
458 | l -= uses_found; // we deleted 1 or more copies of this edge |
459 | } |
460 | igvn->remove_dead_node(p); |
461 | } |
462 | |
463 | // Force the original merge dead |
464 | igvn->hash_delete(r); |
465 | // First, remove region's dead users. |
466 | for (DUIterator_Last lmin, l = r->last_outs(lmin); l >= lmin;) { |
467 | Node* u = r->last_out(l); |
468 | if( u == r ) { |
469 | r->set_req(0, NULL); |
470 | } else { |
471 | assert(u->outcnt() == 0, "only dead users" ); |
472 | igvn->remove_dead_node(u); |
473 | } |
474 | l -= 1; |
475 | } |
476 | igvn->remove_dead_node(r); |
477 | |
478 | // Now remove the bogus extra edges used to keep things alive |
479 | igvn->remove_dead_node( hook ); |
480 | |
481 | // Must return either the original node (now dead) or a new node |
482 | // (Do not return a top here, since that would break the uniqueness of top.) |
483 | return new ConINode(TypeInt::ZERO); |
484 | } |
485 | |
486 | // if this IfNode follows a range check pattern return the projection |
487 | // for the failed path |
488 | ProjNode* IfNode::range_check_trap_proj(int& flip_test, Node*& l, Node*& r) { |
489 | if (outcnt() != 2) { |
490 | return NULL; |
491 | } |
492 | Node* b = in(1); |
493 | if (b == NULL || !b->is_Bool()) return NULL; |
494 | BoolNode* bn = b->as_Bool(); |
495 | Node* cmp = bn->in(1); |
496 | if (cmp == NULL) return NULL; |
497 | if (cmp->Opcode() != Op_CmpU) return NULL; |
498 | |
499 | l = cmp->in(1); |
500 | r = cmp->in(2); |
501 | flip_test = 1; |
502 | if (bn->_test._test == BoolTest::le) { |
503 | l = cmp->in(2); |
504 | r = cmp->in(1); |
505 | flip_test = 2; |
506 | } else if (bn->_test._test != BoolTest::lt) { |
507 | return NULL; |
508 | } |
509 | if (l->is_top()) return NULL; // Top input means dead test |
510 | if (r->Opcode() != Op_LoadRange && !is_RangeCheck()) return NULL; |
511 | |
512 | // We have recognized one of these forms: |
513 | // Flip 1: If (Bool[<] CmpU(l, LoadRange)) ... |
514 | // Flip 2: If (Bool[<=] CmpU(LoadRange, l)) ... |
515 | |
516 | ProjNode* iftrap = proj_out_or_null(flip_test == 2 ? true : false); |
517 | return iftrap; |
518 | } |
519 | |
520 | |
521 | //------------------------------is_range_check--------------------------------- |
522 | // Return 0 if not a range check. Return 1 if a range check and set index and |
523 | // offset. Return 2 if we had to negate the test. Index is NULL if the check |
524 | // is versus a constant. |
525 | int RangeCheckNode::is_range_check(Node* &range, Node* &index, jint &offset) { |
526 | int flip_test = 0; |
527 | Node* l = NULL; |
528 | Node* r = NULL; |
529 | ProjNode* iftrap = range_check_trap_proj(flip_test, l, r); |
530 | |
531 | if (iftrap == NULL) { |
532 | return 0; |
533 | } |
534 | |
535 | // Make sure it's a real range check by requiring an uncommon trap |
536 | // along the OOB path. Otherwise, it's possible that the user wrote |
537 | // something which optimized to look like a range check but behaves |
538 | // in some other way. |
539 | if (iftrap->is_uncommon_trap_proj(Deoptimization::Reason_range_check) == NULL) { |
540 | return 0; |
541 | } |
542 | |
543 | // Look for index+offset form |
544 | Node* ind = l; |
545 | jint off = 0; |
546 | if (l->is_top()) { |
547 | return 0; |
548 | } else if (l->Opcode() == Op_AddI) { |
549 | if ((off = l->in(1)->find_int_con(0)) != 0) { |
550 | ind = l->in(2)->uncast(); |
551 | } else if ((off = l->in(2)->find_int_con(0)) != 0) { |
552 | ind = l->in(1)->uncast(); |
553 | } |
554 | } else if ((off = l->find_int_con(-1)) >= 0) { |
555 | // constant offset with no variable index |
556 | ind = NULL; |
557 | } else { |
558 | // variable index with no constant offset (or dead negative index) |
559 | off = 0; |
560 | } |
561 | |
562 | // Return all the values: |
563 | index = ind; |
564 | offset = off; |
565 | range = r; |
566 | return flip_test; |
567 | } |
568 | |
569 | //------------------------------adjust_check----------------------------------- |
570 | // Adjust (widen) a prior range check |
571 | static void adjust_check(Node* proj, Node* range, Node* index, |
572 | int flip, jint off_lo, PhaseIterGVN* igvn) { |
573 | PhaseGVN *gvn = igvn; |
574 | // Break apart the old check |
575 | Node *iff = proj->in(0); |
576 | Node *bol = iff->in(1); |
577 | if( bol->is_top() ) return; // In case a partially dead range check appears |
578 | // bail (or bomb[ASSERT/DEBUG]) if NOT projection-->IfNode-->BoolNode |
579 | DEBUG_ONLY( if( !bol->is_Bool() ) { proj->dump(3); fatal("Expect projection-->IfNode-->BoolNode" ); } ) |
580 | if( !bol->is_Bool() ) return; |
581 | |
582 | Node *cmp = bol->in(1); |
583 | // Compute a new check |
584 | Node *new_add = gvn->intcon(off_lo); |
585 | if( index ) { |
586 | new_add = off_lo ? gvn->transform(new AddINode( index, new_add )) : index; |
587 | } |
588 | Node *new_cmp = (flip == 1) |
589 | ? new CmpUNode( new_add, range ) |
590 | : new CmpUNode( range, new_add ); |
591 | new_cmp = gvn->transform(new_cmp); |
592 | // See if no need to adjust the existing check |
593 | if( new_cmp == cmp ) return; |
594 | // Else, adjust existing check |
595 | Node *new_bol = gvn->transform( new BoolNode( new_cmp, bol->as_Bool()->_test._test ) ); |
596 | igvn->rehash_node_delayed( iff ); |
597 | iff->set_req_X( 1, new_bol, igvn ); |
598 | } |
599 | |
600 | //------------------------------up_one_dom------------------------------------- |
601 | // Walk up the dominator tree one step. Return NULL at root or true |
602 | // complex merges. Skips through small diamonds. |
603 | Node* IfNode::up_one_dom(Node *curr, bool linear_only) { |
604 | Node *dom = curr->in(0); |
605 | if( !dom ) // Found a Region degraded to a copy? |
606 | return curr->nonnull_req(); // Skip thru it |
607 | |
608 | if( curr != dom ) // Normal walk up one step? |
609 | return dom; |
610 | |
611 | // Use linear_only if we are still parsing, since we cannot |
612 | // trust the regions to be fully filled in. |
613 | if (linear_only) |
614 | return NULL; |
615 | |
616 | if( dom->is_Root() ) |
617 | return NULL; |
618 | |
619 | // Else hit a Region. Check for a loop header |
620 | if( dom->is_Loop() ) |
621 | return dom->in(1); // Skip up thru loops |
622 | |
623 | // Check for small diamonds |
624 | Node *din1, *din2, *din3, *din4; |
625 | if( dom->req() == 3 && // 2-path merge point |
626 | (din1 = dom ->in(1)) && // Left path exists |
627 | (din2 = dom ->in(2)) && // Right path exists |
628 | (din3 = din1->in(0)) && // Left path up one |
629 | (din4 = din2->in(0)) ) { // Right path up one |
630 | if( din3->is_Call() && // Handle a slow-path call on either arm |
631 | (din3 = din3->in(0)) ) |
632 | din3 = din3->in(0); |
633 | if( din4->is_Call() && // Handle a slow-path call on either arm |
634 | (din4 = din4->in(0)) ) |
635 | din4 = din4->in(0); |
636 | if (din3 != NULL && din3 == din4 && din3->is_If()) // Regions not degraded to a copy |
637 | return din3; // Skip around diamonds |
638 | } |
639 | |
640 | // Give up the search at true merges |
641 | return NULL; // Dead loop? Or hit root? |
642 | } |
643 | |
644 | |
645 | //------------------------------filtered_int_type-------------------------------- |
646 | // Return a possibly more restrictive type for val based on condition control flow for an if |
647 | const TypeInt* IfNode::filtered_int_type(PhaseGVN* gvn, Node *val, Node* if_proj) { |
648 | assert(if_proj && |
649 | (if_proj->Opcode() == Op_IfTrue || if_proj->Opcode() == Op_IfFalse), "expecting an if projection" ); |
650 | if (if_proj->in(0) && if_proj->in(0)->is_If()) { |
651 | IfNode* iff = if_proj->in(0)->as_If(); |
652 | if (iff->in(1) && iff->in(1)->is_Bool()) { |
653 | BoolNode* bol = iff->in(1)->as_Bool(); |
654 | if (bol->in(1) && bol->in(1)->is_Cmp()) { |
655 | const CmpNode* cmp = bol->in(1)->as_Cmp(); |
656 | if (cmp->in(1) == val) { |
657 | const TypeInt* cmp2_t = gvn->type(cmp->in(2))->isa_int(); |
658 | if (cmp2_t != NULL) { |
659 | jint lo = cmp2_t->_lo; |
660 | jint hi = cmp2_t->_hi; |
661 | BoolTest::mask msk = if_proj->Opcode() == Op_IfTrue ? bol->_test._test : bol->_test.negate(); |
662 | switch (msk) { |
663 | case BoolTest::ne: |
664 | // Can't refine type |
665 | return NULL; |
666 | case BoolTest::eq: |
667 | return cmp2_t; |
668 | case BoolTest::lt: |
669 | lo = TypeInt::INT->_lo; |
670 | if (hi - 1 < hi) { |
671 | hi = hi - 1; |
672 | } |
673 | break; |
674 | case BoolTest::le: |
675 | lo = TypeInt::INT->_lo; |
676 | break; |
677 | case BoolTest::gt: |
678 | if (lo + 1 > lo) { |
679 | lo = lo + 1; |
680 | } |
681 | hi = TypeInt::INT->_hi; |
682 | break; |
683 | case BoolTest::ge: |
684 | // lo unchanged |
685 | hi = TypeInt::INT->_hi; |
686 | break; |
687 | default: |
688 | break; |
689 | } |
690 | const TypeInt* rtn_t = TypeInt::make(lo, hi, cmp2_t->_widen); |
691 | return rtn_t; |
692 | } |
693 | } |
694 | } |
695 | } |
696 | } |
697 | return NULL; |
698 | } |
699 | |
700 | //------------------------------fold_compares---------------------------- |
701 | // See if a pair of CmpIs can be converted into a CmpU. In some cases |
702 | // the direction of this if is determined by the preceding if so it |
703 | // can be eliminate entirely. |
704 | // |
705 | // Given an if testing (CmpI n v) check for an immediately control |
706 | // dependent if that is testing (CmpI n v2) and has one projection |
707 | // leading to this if and the other projection leading to a region |
708 | // that merges one of this ifs control projections. |
709 | // |
710 | // If |
711 | // / | |
712 | // / | |
713 | // / | |
714 | // If | |
715 | // /\ | |
716 | // / \ | |
717 | // / \ | |
718 | // / Region |
719 | // |
720 | // Or given an if testing (CmpI n v) check for a dominating if that is |
721 | // testing (CmpI n v2), both having one projection leading to an |
722 | // uncommon trap. Allow Another independent guard in between to cover |
723 | // an explicit range check: |
724 | // if (index < 0 || index >= array.length) { |
725 | // which may need a null check to guard the LoadRange |
726 | // |
727 | // If |
728 | // / \ |
729 | // / \ |
730 | // / \ |
731 | // If unc |
732 | // /\ |
733 | // / \ |
734 | // / \ |
735 | // / unc |
736 | // |
737 | |
738 | // Is the comparison for this If suitable for folding? |
739 | bool IfNode::cmpi_folds(PhaseIterGVN* igvn) { |
740 | return in(1) != NULL && |
741 | in(1)->is_Bool() && |
742 | in(1)->in(1) != NULL && |
743 | in(1)->in(1)->Opcode() == Op_CmpI && |
744 | in(1)->in(1)->in(2) != NULL && |
745 | in(1)->in(1)->in(2) != igvn->C->top() && |
746 | (in(1)->as_Bool()->_test.is_less() || |
747 | in(1)->as_Bool()->_test.is_greater()); |
748 | } |
749 | |
750 | // Is a dominating control suitable for folding with this if? |
751 | bool IfNode::is_ctrl_folds(Node* ctrl, PhaseIterGVN* igvn) { |
752 | return ctrl != NULL && |
753 | ctrl->is_Proj() && |
754 | ctrl->in(0) != NULL && |
755 | ctrl->in(0)->Opcode() == Op_If && |
756 | ctrl->in(0)->outcnt() == 2 && |
757 | ctrl->in(0)->as_If()->cmpi_folds(igvn) && |
758 | // Must compare same value |
759 | ctrl->in(0)->in(1)->in(1)->in(1) != NULL && |
760 | ctrl->in(0)->in(1)->in(1)->in(1) == in(1)->in(1)->in(1); |
761 | } |
762 | |
763 | // Do this If and the dominating If share a region? |
764 | bool IfNode::has_shared_region(ProjNode* proj, ProjNode*& success, ProjNode*& fail) { |
765 | ProjNode* otherproj = proj->other_if_proj(); |
766 | Node* otherproj_ctrl_use = otherproj->unique_ctrl_out(); |
767 | RegionNode* region = (otherproj_ctrl_use != NULL && otherproj_ctrl_use->is_Region()) ? otherproj_ctrl_use->as_Region() : NULL; |
768 | success = NULL; |
769 | fail = NULL; |
770 | |
771 | if (otherproj->outcnt() == 1 && region != NULL && !region->has_phi()) { |
772 | for (int i = 0; i < 2; i++) { |
773 | ProjNode* proj = proj_out(i); |
774 | if (success == NULL && proj->outcnt() == 1 && proj->unique_out() == region) { |
775 | success = proj; |
776 | } else if (fail == NULL) { |
777 | fail = proj; |
778 | } else { |
779 | success = fail = NULL; |
780 | } |
781 | } |
782 | } |
783 | return success != NULL && fail != NULL; |
784 | } |
785 | |
786 | bool IfNode::is_dominator_unc(CallStaticJavaNode* dom_unc, CallStaticJavaNode* unc) { |
787 | // Different methods and methods containing jsrs are not supported. |
788 | ciMethod* method = unc->jvms()->method(); |
789 | ciMethod* dom_method = dom_unc->jvms()->method(); |
790 | if (method != dom_method || method->has_jsrs()) { |
791 | return false; |
792 | } |
793 | // Check that both traps are in the same activation of the method (instead |
794 | // of two activations being inlined through different call sites) by verifying |
795 | // that the call stacks are equal for both JVMStates. |
796 | JVMState* dom_caller = dom_unc->jvms()->caller(); |
797 | JVMState* caller = unc->jvms()->caller(); |
798 | if ((dom_caller == NULL) != (caller == NULL)) { |
799 | // The current method must either be inlined into both dom_caller and |
800 | // caller or must not be inlined at all (top method). Bail out otherwise. |
801 | return false; |
802 | } else if (dom_caller != NULL && !dom_caller->same_calls_as(caller)) { |
803 | return false; |
804 | } |
805 | // Check that the bci of the dominating uncommon trap dominates the bci |
806 | // of the dominated uncommon trap. Otherwise we may not re-execute |
807 | // the dominated check after deoptimization from the merged uncommon trap. |
808 | ciTypeFlow* flow = dom_method->get_flow_analysis(); |
809 | int bci = unc->jvms()->bci(); |
810 | int dom_bci = dom_unc->jvms()->bci(); |
811 | if (!flow->is_dominated_by(bci, dom_bci)) { |
812 | return false; |
813 | } |
814 | |
815 | return true; |
816 | } |
817 | |
818 | // Return projection that leads to an uncommon trap if any |
819 | ProjNode* IfNode::uncommon_trap_proj(CallStaticJavaNode*& call) const { |
820 | for (int i = 0; i < 2; i++) { |
821 | call = proj_out(i)->is_uncommon_trap_proj(Deoptimization::Reason_none); |
822 | if (call != NULL) { |
823 | return proj_out(i); |
824 | } |
825 | } |
826 | return NULL; |
827 | } |
828 | |
829 | // Do this If and the dominating If both branch out to an uncommon trap |
830 | bool IfNode::has_only_uncommon_traps(ProjNode* proj, ProjNode*& success, ProjNode*& fail, PhaseIterGVN* igvn) { |
831 | ProjNode* otherproj = proj->other_if_proj(); |
832 | CallStaticJavaNode* dom_unc = otherproj->is_uncommon_trap_proj(Deoptimization::Reason_none); |
833 | |
834 | if (otherproj->outcnt() == 1 && dom_unc != NULL) { |
835 | // We need to re-execute the folded Ifs after deoptimization from the merged traps |
836 | if (!dom_unc->jvms()->should_reexecute()) { |
837 | return false; |
838 | } |
839 | |
840 | CallStaticJavaNode* unc = NULL; |
841 | ProjNode* unc_proj = uncommon_trap_proj(unc); |
842 | if (unc_proj != NULL && unc_proj->outcnt() == 1) { |
843 | if (dom_unc == unc) { |
844 | // Allow the uncommon trap to be shared through a region |
845 | RegionNode* r = unc->in(0)->as_Region(); |
846 | if (r->outcnt() != 2 || r->req() != 3 || r->find_edge(otherproj) == -1 || r->find_edge(unc_proj) == -1) { |
847 | return false; |
848 | } |
849 | assert(r->has_phi() == NULL, "simple region shouldn't have a phi" ); |
850 | } else if (dom_unc->in(0) != otherproj || unc->in(0) != unc_proj) { |
851 | return false; |
852 | } |
853 | |
854 | if (!is_dominator_unc(dom_unc, unc)) { |
855 | return false; |
856 | } |
857 | |
858 | // See merge_uncommon_traps: the reason of the uncommon trap |
859 | // will be changed and the state of the dominating If will be |
860 | // used. Checked that we didn't apply this transformation in a |
861 | // previous compilation and it didn't cause too many traps |
862 | ciMethod* dom_method = dom_unc->jvms()->method(); |
863 | int dom_bci = dom_unc->jvms()->bci(); |
864 | if (!igvn->C->too_many_traps(dom_method, dom_bci, Deoptimization::Reason_unstable_fused_if) && |
865 | !igvn->C->too_many_traps(dom_method, dom_bci, Deoptimization::Reason_range_check)) { |
866 | success = unc_proj; |
867 | fail = unc_proj->other_if_proj(); |
868 | return true; |
869 | } |
870 | } |
871 | } |
872 | return false; |
873 | } |
874 | |
875 | // Check that the 2 CmpI can be folded into as single CmpU and proceed with the folding |
876 | bool IfNode::fold_compares_helper(ProjNode* proj, ProjNode* success, ProjNode* fail, PhaseIterGVN* igvn) { |
877 | Node* this_cmp = in(1)->in(1); |
878 | BoolNode* this_bool = in(1)->as_Bool(); |
879 | IfNode* dom_iff = proj->in(0)->as_If(); |
880 | BoolNode* dom_bool = dom_iff->in(1)->as_Bool(); |
881 | Node* lo = dom_iff->in(1)->in(1)->in(2); |
882 | Node* hi = this_cmp->in(2); |
883 | Node* n = this_cmp->in(1); |
884 | ProjNode* otherproj = proj->other_if_proj(); |
885 | |
886 | const TypeInt* lo_type = IfNode::filtered_int_type(igvn, n, otherproj); |
887 | const TypeInt* hi_type = IfNode::filtered_int_type(igvn, n, success); |
888 | |
889 | BoolTest::mask lo_test = dom_bool->_test._test; |
890 | BoolTest::mask hi_test = this_bool->_test._test; |
891 | BoolTest::mask cond = hi_test; |
892 | |
893 | // convert: |
894 | // |
895 | // dom_bool = x {<,<=,>,>=} a |
896 | // / \ |
897 | // proj = {True,False} / \ otherproj = {False,True} |
898 | // / |
899 | // this_bool = x {<,<=} b |
900 | // / \ |
901 | // fail = {True,False} / \ success = {False,True} |
902 | // / |
903 | // |
904 | // (Second test guaranteed canonicalized, first one may not have |
905 | // been canonicalized yet) |
906 | // |
907 | // into: |
908 | // |
909 | // cond = (x - lo) {<u,<=u,>u,>=u} adjusted_lim |
910 | // / \ |
911 | // fail / \ success |
912 | // / |
913 | // |
914 | |
915 | // Figure out which of the two tests sets the upper bound and which |
916 | // sets the lower bound if any. |
917 | Node* adjusted_lim = NULL; |
918 | if (lo_type != NULL && hi_type != NULL && hi_type->_lo > lo_type->_hi && |
919 | hi_type->_hi == max_jint && lo_type->_lo == min_jint) { |
920 | assert((dom_bool->_test.is_less() && !proj->_con) || |
921 | (dom_bool->_test.is_greater() && proj->_con), "incorrect test" ); |
922 | // this test was canonicalized |
923 | assert(this_bool->_test.is_less() && fail->_con, "incorrect test" ); |
924 | |
925 | // this_bool = < |
926 | // dom_bool = >= (proj = True) or dom_bool = < (proj = False) |
927 | // x in [a, b[ on the fail (= True) projection, b > a-1 (because of hi_type->_lo > lo_type->_hi test above): |
928 | // lo = a, hi = b, adjusted_lim = b-a, cond = <u |
929 | // dom_bool = > (proj = True) or dom_bool = <= (proj = False) |
930 | // x in ]a, b[ on the fail (= True) projection, b > a: |
931 | // lo = a+1, hi = b, adjusted_lim = b-a-1, cond = <u |
932 | // this_bool = <= |
933 | // dom_bool = >= (proj = True) or dom_bool = < (proj = False) |
934 | // x in [a, b] on the fail (= True) projection, b+1 > a-1: |
935 | // lo = a, hi = b, adjusted_lim = b-a+1, cond = <u |
936 | // lo = a, hi = b, adjusted_lim = b-a, cond = <=u doesn't work because b = a - 1 is possible, then b-a = -1 |
937 | // dom_bool = > (proj = True) or dom_bool = <= (proj = False) |
938 | // x in ]a, b] on the fail (= True) projection b+1 > a: |
939 | // lo = a+1, hi = b, adjusted_lim = b-a, cond = <u |
940 | // lo = a+1, hi = b, adjusted_lim = b-a-1, cond = <=u doesn't work because a = b is possible, then b-a-1 = -1 |
941 | |
942 | if (hi_test == BoolTest::lt) { |
943 | if (lo_test == BoolTest::gt || lo_test == BoolTest::le) { |
944 | lo = igvn->transform(new AddINode(lo, igvn->intcon(1))); |
945 | } |
946 | } else { |
947 | assert(hi_test == BoolTest::le, "bad test" ); |
948 | if (lo_test == BoolTest::ge || lo_test == BoolTest::lt) { |
949 | adjusted_lim = igvn->transform(new SubINode(hi, lo)); |
950 | adjusted_lim = igvn->transform(new AddINode(adjusted_lim, igvn->intcon(1))); |
951 | cond = BoolTest::lt; |
952 | } else { |
953 | assert(lo_test == BoolTest::gt || lo_test == BoolTest::le, "bad test" ); |
954 | adjusted_lim = igvn->transform(new SubINode(hi, lo)); |
955 | lo = igvn->transform(new AddINode(lo, igvn->intcon(1))); |
956 | cond = BoolTest::lt; |
957 | } |
958 | } |
959 | } else if (lo_type != NULL && hi_type != NULL && lo_type->_lo > hi_type->_hi && |
960 | lo_type->_hi == max_jint && hi_type->_lo == min_jint) { |
961 | |
962 | // this_bool = < |
963 | // dom_bool = < (proj = True) or dom_bool = >= (proj = False) |
964 | // x in [b, a[ on the fail (= False) projection, a > b-1 (because of lo_type->_lo > hi_type->_hi above): |
965 | // lo = b, hi = a, adjusted_lim = a-b, cond = >=u |
966 | // dom_bool = <= (proj = True) or dom_bool = > (proj = False) |
967 | // x in [b, a] on the fail (= False) projection, a+1 > b-1: |
968 | // lo = b, hi = a, adjusted_lim = a-b+1, cond = >=u |
969 | // lo = b, hi = a, adjusted_lim = a-b, cond = >u doesn't work because a = b - 1 is possible, then b-a = -1 |
970 | // this_bool = <= |
971 | // dom_bool = < (proj = True) or dom_bool = >= (proj = False) |
972 | // x in ]b, a[ on the fail (= False) projection, a > b: |
973 | // lo = b+1, hi = a, adjusted_lim = a-b-1, cond = >=u |
974 | // dom_bool = <= (proj = True) or dom_bool = > (proj = False) |
975 | // x in ]b, a] on the fail (= False) projection, a+1 > b: |
976 | // lo = b+1, hi = a, adjusted_lim = a-b, cond = >=u |
977 | // lo = b+1, hi = a, adjusted_lim = a-b-1, cond = >u doesn't work because a = b is possible, then b-a-1 = -1 |
978 | |
979 | swap(lo, hi); |
980 | swap(lo_type, hi_type); |
981 | swap(lo_test, hi_test); |
982 | |
983 | assert((dom_bool->_test.is_less() && proj->_con) || |
984 | (dom_bool->_test.is_greater() && !proj->_con), "incorrect test" ); |
985 | // this test was canonicalized |
986 | assert(this_bool->_test.is_less() && !fail->_con, "incorrect test" ); |
987 | |
988 | cond = (hi_test == BoolTest::le || hi_test == BoolTest::gt) ? BoolTest::gt : BoolTest::ge; |
989 | |
990 | if (lo_test == BoolTest::lt) { |
991 | if (hi_test == BoolTest::lt || hi_test == BoolTest::ge) { |
992 | cond = BoolTest::ge; |
993 | } else { |
994 | assert(hi_test == BoolTest::le || hi_test == BoolTest::gt, "bad test" ); |
995 | adjusted_lim = igvn->transform(new SubINode(hi, lo)); |
996 | adjusted_lim = igvn->transform(new AddINode(adjusted_lim, igvn->intcon(1))); |
997 | cond = BoolTest::ge; |
998 | } |
999 | } else if (lo_test == BoolTest::le) { |
1000 | if (hi_test == BoolTest::lt || hi_test == BoolTest::ge) { |
1001 | lo = igvn->transform(new AddINode(lo, igvn->intcon(1))); |
1002 | cond = BoolTest::ge; |
1003 | } else { |
1004 | assert(hi_test == BoolTest::le || hi_test == BoolTest::gt, "bad test" ); |
1005 | adjusted_lim = igvn->transform(new SubINode(hi, lo)); |
1006 | lo = igvn->transform(new AddINode(lo, igvn->intcon(1))); |
1007 | cond = BoolTest::ge; |
1008 | } |
1009 | } |
1010 | } else { |
1011 | const TypeInt* failtype = filtered_int_type(igvn, n, proj); |
1012 | if (failtype != NULL) { |
1013 | const TypeInt* type2 = filtered_int_type(igvn, n, fail); |
1014 | if (type2 != NULL) { |
1015 | failtype = failtype->join(type2)->is_int(); |
1016 | if (failtype->_lo > failtype->_hi) { |
1017 | // previous if determines the result of this if so |
1018 | // replace Bool with constant |
1019 | igvn->_worklist.push(in(1)); |
1020 | igvn->replace_input_of(this, 1, igvn->intcon(success->_con)); |
1021 | return true; |
1022 | } |
1023 | } |
1024 | } |
1025 | lo = NULL; |
1026 | hi = NULL; |
1027 | } |
1028 | |
1029 | if (lo && hi) { |
1030 | // Merge the two compares into a single unsigned compare by building (CmpU (n - lo) (hi - lo)) |
1031 | Node* adjusted_val = igvn->transform(new SubINode(n, lo)); |
1032 | if (adjusted_lim == NULL) { |
1033 | adjusted_lim = igvn->transform(new SubINode(hi, lo)); |
1034 | } |
1035 | Node* newcmp = igvn->transform(new CmpUNode(adjusted_val, adjusted_lim)); |
1036 | Node* newbool = igvn->transform(new BoolNode(newcmp, cond)); |
1037 | |
1038 | igvn->replace_input_of(dom_iff, 1, igvn->intcon(proj->_con)); |
1039 | igvn->_worklist.push(in(1)); |
1040 | igvn->replace_input_of(this, 1, newbool); |
1041 | |
1042 | return true; |
1043 | } |
1044 | return false; |
1045 | } |
1046 | |
1047 | // Merge the branches that trap for this If and the dominating If into |
1048 | // a single region that branches to the uncommon trap for the |
1049 | // dominating If |
1050 | Node* IfNode::merge_uncommon_traps(ProjNode* proj, ProjNode* success, ProjNode* fail, PhaseIterGVN* igvn) { |
1051 | Node* res = this; |
1052 | assert(success->in(0) == this, "bad projection" ); |
1053 | |
1054 | ProjNode* otherproj = proj->other_if_proj(); |
1055 | |
1056 | CallStaticJavaNode* unc = success->is_uncommon_trap_proj(Deoptimization::Reason_none); |
1057 | CallStaticJavaNode* dom_unc = otherproj->is_uncommon_trap_proj(Deoptimization::Reason_none); |
1058 | |
1059 | if (unc != dom_unc) { |
1060 | Node* r = new RegionNode(3); |
1061 | |
1062 | r->set_req(1, otherproj); |
1063 | r->set_req(2, success); |
1064 | r = igvn->transform(r); |
1065 | assert(r->is_Region(), "can't go away" ); |
1066 | |
1067 | // Make both If trap at the state of the first If: once the CmpI |
1068 | // nodes are merged, if we trap we don't know which of the CmpI |
1069 | // nodes would have caused the trap so we have to restart |
1070 | // execution at the first one |
1071 | igvn->replace_input_of(dom_unc, 0, r); |
1072 | igvn->replace_input_of(unc, 0, igvn->C->top()); |
1073 | } |
1074 | int trap_request = dom_unc->uncommon_trap_request(); |
1075 | Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); |
1076 | Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); |
1077 | |
1078 | int flip_test = 0; |
1079 | Node* l = NULL; |
1080 | Node* r = NULL; |
1081 | |
1082 | if (success->in(0)->as_If()->range_check_trap_proj(flip_test, l, r) != NULL) { |
1083 | // If this looks like a range check, change the trap to |
1084 | // Reason_range_check so the compiler recognizes it as a range |
1085 | // check and applies the corresponding optimizations |
1086 | trap_request = Deoptimization::make_trap_request(Deoptimization::Reason_range_check, action); |
1087 | |
1088 | improve_address_types(l, r, fail, igvn); |
1089 | |
1090 | res = igvn->transform(new RangeCheckNode(in(0), in(1), _prob, _fcnt)); |
1091 | } else if (unc != dom_unc) { |
1092 | // If we trap we won't know what CmpI would have caused the trap |
1093 | // so use a special trap reason to mark this pair of CmpI nodes as |
1094 | // bad candidate for folding. On recompilation we won't fold them |
1095 | // and we may trap again but this time we'll know what branch |
1096 | // traps |
1097 | trap_request = Deoptimization::make_trap_request(Deoptimization::Reason_unstable_fused_if, action); |
1098 | } |
1099 | igvn->replace_input_of(dom_unc, TypeFunc::Parms, igvn->intcon(trap_request)); |
1100 | return res; |
1101 | } |
1102 | |
1103 | // If we are turning 2 CmpI nodes into a CmpU that follows the pattern |
1104 | // of a rangecheck on index i, on 64 bit the compares may be followed |
1105 | // by memory accesses using i as index. In that case, the CmpU tells |
1106 | // us something about the values taken by i that can help the compiler |
1107 | // (see Compile::conv_I2X_index()) |
1108 | void IfNode::improve_address_types(Node* l, Node* r, ProjNode* fail, PhaseIterGVN* igvn) { |
1109 | #ifdef _LP64 |
1110 | ResourceMark rm; |
1111 | Node_Stack stack(2); |
1112 | |
1113 | assert(r->Opcode() == Op_LoadRange, "unexpected range check" ); |
1114 | const TypeInt* array_size = igvn->type(r)->is_int(); |
1115 | |
1116 | stack.push(l, 0); |
1117 | |
1118 | while(stack.size() > 0) { |
1119 | Node* n = stack.node(); |
1120 | uint start = stack.index(); |
1121 | |
1122 | uint i = start; |
1123 | for (; i < n->outcnt(); i++) { |
1124 | Node* use = n->raw_out(i); |
1125 | if (stack.size() == 1) { |
1126 | if (use->Opcode() == Op_ConvI2L) { |
1127 | const TypeLong* bounds = use->as_Type()->type()->is_long(); |
1128 | if (bounds->_lo <= array_size->_lo && bounds->_hi >= array_size->_hi && |
1129 | (bounds->_lo != array_size->_lo || bounds->_hi != array_size->_hi)) { |
1130 | stack.set_index(i+1); |
1131 | stack.push(use, 0); |
1132 | break; |
1133 | } |
1134 | } |
1135 | } else if (use->is_Mem()) { |
1136 | Node* ctrl = use->in(0); |
1137 | for (int i = 0; i < 10 && ctrl != NULL && ctrl != fail; i++) { |
1138 | ctrl = up_one_dom(ctrl); |
1139 | } |
1140 | if (ctrl == fail) { |
1141 | Node* init_n = stack.node_at(1); |
1142 | assert(init_n->Opcode() == Op_ConvI2L, "unexpected first node" ); |
1143 | // Create a new narrow ConvI2L node that is dependent on the range check |
1144 | Node* new_n = igvn->C->conv_I2X_index(igvn, l, array_size, fail); |
1145 | |
1146 | // The type of the ConvI2L may be widen and so the new |
1147 | // ConvI2L may not be better than an existing ConvI2L |
1148 | if (new_n != init_n) { |
1149 | for (uint j = 2; j < stack.size(); j++) { |
1150 | Node* n = stack.node_at(j); |
1151 | Node* clone = n->clone(); |
1152 | int rep = clone->replace_edge(init_n, new_n); |
1153 | assert(rep > 0, "can't find expected node?" ); |
1154 | clone = igvn->transform(clone); |
1155 | init_n = n; |
1156 | new_n = clone; |
1157 | } |
1158 | igvn->hash_delete(use); |
1159 | int rep = use->replace_edge(init_n, new_n); |
1160 | assert(rep > 0, "can't find expected node?" ); |
1161 | igvn->transform(use); |
1162 | if (init_n->outcnt() == 0) { |
1163 | igvn->_worklist.push(init_n); |
1164 | } |
1165 | } |
1166 | } |
1167 | } else if (use->in(0) == NULL && (igvn->type(use)->isa_long() || |
1168 | igvn->type(use)->isa_ptr())) { |
1169 | stack.set_index(i+1); |
1170 | stack.push(use, 0); |
1171 | break; |
1172 | } |
1173 | } |
1174 | if (i == n->outcnt()) { |
1175 | stack.pop(); |
1176 | } |
1177 | } |
1178 | #endif |
1179 | } |
1180 | |
1181 | bool IfNode::is_cmp_with_loadrange(ProjNode* proj) { |
1182 | if (in(1) != NULL && |
1183 | in(1)->in(1) != NULL && |
1184 | in(1)->in(1)->in(2) != NULL) { |
1185 | Node* other = in(1)->in(1)->in(2); |
1186 | if (other->Opcode() == Op_LoadRange && |
1187 | ((other->in(0) != NULL && other->in(0) == proj) || |
1188 | (other->in(0) == NULL && |
1189 | other->in(2) != NULL && |
1190 | other->in(2)->is_AddP() && |
1191 | other->in(2)->in(1) != NULL && |
1192 | other->in(2)->in(1)->Opcode() == Op_CastPP && |
1193 | other->in(2)->in(1)->in(0) == proj))) { |
1194 | return true; |
1195 | } |
1196 | } |
1197 | return false; |
1198 | } |
1199 | |
1200 | bool IfNode::is_null_check(ProjNode* proj, PhaseIterGVN* igvn) { |
1201 | Node* other = in(1)->in(1)->in(2); |
1202 | if (other->in(MemNode::Address) != NULL && |
1203 | proj->in(0)->in(1) != NULL && |
1204 | proj->in(0)->in(1)->is_Bool() && |
1205 | proj->in(0)->in(1)->in(1) != NULL && |
1206 | proj->in(0)->in(1)->in(1)->Opcode() == Op_CmpP && |
1207 | proj->in(0)->in(1)->in(1)->in(2) != NULL && |
1208 | proj->in(0)->in(1)->in(1)->in(1) == other->in(MemNode::Address)->in(AddPNode::Address)->uncast() && |
1209 | igvn->type(proj->in(0)->in(1)->in(1)->in(2)) == TypePtr::NULL_PTR) { |
1210 | return true; |
1211 | } |
1212 | return false; |
1213 | } |
1214 | |
1215 | // Check that the If that is in between the 2 integer comparisons has |
1216 | // no side effect |
1217 | bool IfNode::is_side_effect_free_test(ProjNode* proj, PhaseIterGVN* igvn) { |
1218 | if (proj == NULL) { |
1219 | return false; |
1220 | } |
1221 | CallStaticJavaNode* unc = proj->is_uncommon_trap_if_pattern(Deoptimization::Reason_none); |
1222 | if (unc != NULL && proj->outcnt() <= 2) { |
1223 | if (proj->outcnt() == 1 || |
1224 | // Allow simple null check from LoadRange |
1225 | (is_cmp_with_loadrange(proj) && is_null_check(proj, igvn))) { |
1226 | CallStaticJavaNode* unc = proj->is_uncommon_trap_if_pattern(Deoptimization::Reason_none); |
1227 | CallStaticJavaNode* dom_unc = proj->in(0)->in(0)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none); |
1228 | assert(dom_unc != NULL, "is_uncommon_trap_if_pattern returned NULL" ); |
1229 | |
1230 | // reroute_side_effect_free_unc changes the state of this |
1231 | // uncommon trap to restart execution at the previous |
1232 | // CmpI. Check that this change in a previous compilation didn't |
1233 | // cause too many traps. |
1234 | int trap_request = unc->uncommon_trap_request(); |
1235 | Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); |
1236 | |
1237 | if (igvn->C->too_many_traps(dom_unc->jvms()->method(), dom_unc->jvms()->bci(), reason)) { |
1238 | return false; |
1239 | } |
1240 | |
1241 | if (!is_dominator_unc(dom_unc, unc)) { |
1242 | return false; |
1243 | } |
1244 | |
1245 | return true; |
1246 | } |
1247 | } |
1248 | return false; |
1249 | } |
1250 | |
1251 | // Make the If between the 2 integer comparisons trap at the state of |
1252 | // the first If: the last CmpI is the one replaced by a CmpU and the |
1253 | // first CmpI is eliminated, so the test between the 2 CmpI nodes |
1254 | // won't be guarded by the first CmpI anymore. It can trap in cases |
1255 | // where the first CmpI would have prevented it from executing: on a |
1256 | // trap, we need to restart execution at the state of the first CmpI |
1257 | void IfNode::reroute_side_effect_free_unc(ProjNode* proj, ProjNode* dom_proj, PhaseIterGVN* igvn) { |
1258 | CallStaticJavaNode* dom_unc = dom_proj->is_uncommon_trap_if_pattern(Deoptimization::Reason_none); |
1259 | ProjNode* otherproj = proj->other_if_proj(); |
1260 | CallStaticJavaNode* unc = proj->is_uncommon_trap_if_pattern(Deoptimization::Reason_none); |
1261 | Node* call_proj = dom_unc->unique_ctrl_out(); |
1262 | Node* halt = call_proj->unique_ctrl_out(); |
1263 | |
1264 | Node* new_unc = dom_unc->clone(); |
1265 | call_proj = call_proj->clone(); |
1266 | halt = halt->clone(); |
1267 | Node* c = otherproj->clone(); |
1268 | |
1269 | c = igvn->transform(c); |
1270 | new_unc->set_req(TypeFunc::Parms, unc->in(TypeFunc::Parms)); |
1271 | new_unc->set_req(0, c); |
1272 | new_unc = igvn->transform(new_unc); |
1273 | call_proj->set_req(0, new_unc); |
1274 | call_proj = igvn->transform(call_proj); |
1275 | halt->set_req(0, call_proj); |
1276 | halt = igvn->transform(halt); |
1277 | |
1278 | igvn->replace_node(otherproj, igvn->C->top()); |
1279 | igvn->C->root()->add_req(halt); |
1280 | } |
1281 | |
1282 | Node* IfNode::fold_compares(PhaseIterGVN* igvn) { |
1283 | if (Opcode() != Op_If) return NULL; |
1284 | |
1285 | if (cmpi_folds(igvn)) { |
1286 | Node* ctrl = in(0); |
1287 | if (is_ctrl_folds(ctrl, igvn) && |
1288 | ctrl->outcnt() == 1) { |
1289 | // A integer comparison immediately dominated by another integer |
1290 | // comparison |
1291 | ProjNode* success = NULL; |
1292 | ProjNode* fail = NULL; |
1293 | ProjNode* dom_cmp = ctrl->as_Proj(); |
1294 | if (has_shared_region(dom_cmp, success, fail) && |
1295 | // Next call modifies graph so must be last |
1296 | fold_compares_helper(dom_cmp, success, fail, igvn)) { |
1297 | return this; |
1298 | } |
1299 | if (has_only_uncommon_traps(dom_cmp, success, fail, igvn) && |
1300 | // Next call modifies graph so must be last |
1301 | fold_compares_helper(dom_cmp, success, fail, igvn)) { |
1302 | return merge_uncommon_traps(dom_cmp, success, fail, igvn); |
1303 | } |
1304 | return NULL; |
1305 | } else if (ctrl->in(0) != NULL && |
1306 | ctrl->in(0)->in(0) != NULL) { |
1307 | ProjNode* success = NULL; |
1308 | ProjNode* fail = NULL; |
1309 | Node* dom = ctrl->in(0)->in(0); |
1310 | ProjNode* dom_cmp = dom->isa_Proj(); |
1311 | ProjNode* other_cmp = ctrl->isa_Proj(); |
1312 | |
1313 | // Check if it's an integer comparison dominated by another |
1314 | // integer comparison with another test in between |
1315 | if (is_ctrl_folds(dom, igvn) && |
1316 | has_only_uncommon_traps(dom_cmp, success, fail, igvn) && |
1317 | is_side_effect_free_test(other_cmp, igvn) && |
1318 | // Next call modifies graph so must be last |
1319 | fold_compares_helper(dom_cmp, success, fail, igvn)) { |
1320 | reroute_side_effect_free_unc(other_cmp, dom_cmp, igvn); |
1321 | return merge_uncommon_traps(dom_cmp, success, fail, igvn); |
1322 | } |
1323 | } |
1324 | } |
1325 | return NULL; |
1326 | } |
1327 | |
1328 | //------------------------------remove_useless_bool---------------------------- |
1329 | // Check for people making a useless boolean: things like |
1330 | // if( (x < y ? true : false) ) { ... } |
1331 | // Replace with if( x < y ) { ... } |
1332 | static Node *remove_useless_bool(IfNode *iff, PhaseGVN *phase) { |
1333 | Node *i1 = iff->in(1); |
1334 | if( !i1->is_Bool() ) return NULL; |
1335 | BoolNode *bol = i1->as_Bool(); |
1336 | |
1337 | Node *cmp = bol->in(1); |
1338 | if( cmp->Opcode() != Op_CmpI ) return NULL; |
1339 | |
1340 | // Must be comparing against a bool |
1341 | const Type *cmp2_t = phase->type( cmp->in(2) ); |
1342 | if( cmp2_t != TypeInt::ZERO && |
1343 | cmp2_t != TypeInt::ONE ) |
1344 | return NULL; |
1345 | |
1346 | // Find a prior merge point merging the boolean |
1347 | i1 = cmp->in(1); |
1348 | if( !i1->is_Phi() ) return NULL; |
1349 | PhiNode *phi = i1->as_Phi(); |
1350 | if( phase->type( phi ) != TypeInt::BOOL ) |
1351 | return NULL; |
1352 | |
1353 | // Check for diamond pattern |
1354 | int true_path = phi->is_diamond_phi(); |
1355 | if( true_path == 0 ) return NULL; |
1356 | |
1357 | // Make sure that iff and the control of the phi are different. This |
1358 | // should really only happen for dead control flow since it requires |
1359 | // an illegal cycle. |
1360 | if (phi->in(0)->in(1)->in(0) == iff) return NULL; |
1361 | |
1362 | // phi->region->if_proj->ifnode->bool->cmp |
1363 | BoolNode *bol2 = phi->in(0)->in(1)->in(0)->in(1)->as_Bool(); |
1364 | |
1365 | // Now get the 'sense' of the test correct so we can plug in |
1366 | // either iff2->in(1) or its complement. |
1367 | int flip = 0; |
1368 | if( bol->_test._test == BoolTest::ne ) flip = 1-flip; |
1369 | else if( bol->_test._test != BoolTest::eq ) return NULL; |
1370 | if( cmp2_t == TypeInt::ZERO ) flip = 1-flip; |
1371 | |
1372 | const Type *phi1_t = phase->type( phi->in(1) ); |
1373 | const Type *phi2_t = phase->type( phi->in(2) ); |
1374 | // Check for Phi(0,1) and flip |
1375 | if( phi1_t == TypeInt::ZERO ) { |
1376 | if( phi2_t != TypeInt::ONE ) return NULL; |
1377 | flip = 1-flip; |
1378 | } else { |
1379 | // Check for Phi(1,0) |
1380 | if( phi1_t != TypeInt::ONE ) return NULL; |
1381 | if( phi2_t != TypeInt::ZERO ) return NULL; |
1382 | } |
1383 | if( true_path == 2 ) { |
1384 | flip = 1-flip; |
1385 | } |
1386 | |
1387 | Node* new_bol = (flip ? phase->transform( bol2->negate(phase) ) : bol2); |
1388 | assert(new_bol != iff->in(1), "must make progress" ); |
1389 | iff->set_req(1, new_bol); |
1390 | // Intervening diamond probably goes dead |
1391 | phase->C->set_major_progress(); |
1392 | return iff; |
1393 | } |
1394 | |
1395 | static IfNode* idealize_test(PhaseGVN* phase, IfNode* iff); |
1396 | |
1397 | struct RangeCheck { |
1398 | Node* ctl; |
1399 | jint off; |
1400 | }; |
1401 | |
1402 | Node* IfNode::Ideal_common(PhaseGVN *phase, bool can_reshape) { |
1403 | if (remove_dead_region(phase, can_reshape)) return this; |
1404 | // No Def-Use info? |
1405 | if (!can_reshape) return NULL; |
1406 | |
1407 | // Don't bother trying to transform a dead if |
1408 | if (in(0)->is_top()) return NULL; |
1409 | // Don't bother trying to transform an if with a dead test |
1410 | if (in(1)->is_top()) return NULL; |
1411 | // Another variation of a dead test |
1412 | if (in(1)->is_Con()) return NULL; |
1413 | // Another variation of a dead if |
1414 | if (outcnt() < 2) return NULL; |
1415 | |
1416 | // Canonicalize the test. |
1417 | Node* idt_if = idealize_test(phase, this); |
1418 | if (idt_if != NULL) return idt_if; |
1419 | |
1420 | // Try to split the IF |
1421 | PhaseIterGVN *igvn = phase->is_IterGVN(); |
1422 | Node *s = split_if(this, igvn); |
1423 | if (s != NULL) return s; |
1424 | |
1425 | return NodeSentinel; |
1426 | } |
1427 | |
1428 | //------------------------------Ideal------------------------------------------ |
1429 | // Return a node which is more "ideal" than the current node. Strip out |
1430 | // control copies |
1431 | Node* IfNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
1432 | Node* res = Ideal_common(phase, can_reshape); |
1433 | if (res != NodeSentinel) { |
1434 | return res; |
1435 | } |
1436 | |
1437 | // Check for people making a useless boolean: things like |
1438 | // if( (x < y ? true : false) ) { ... } |
1439 | // Replace with if( x < y ) { ... } |
1440 | Node *bol2 = remove_useless_bool(this, phase); |
1441 | if( bol2 ) return bol2; |
1442 | |
1443 | if (in(0) == NULL) return NULL; // Dead loop? |
1444 | |
1445 | PhaseIterGVN *igvn = phase->is_IterGVN(); |
1446 | Node* result = fold_compares(igvn); |
1447 | if (result != NULL) { |
1448 | return result; |
1449 | } |
1450 | |
1451 | // Scan for an equivalent test |
1452 | Node *cmp; |
1453 | int dist = 0; // Cutoff limit for search |
1454 | int op = Opcode(); |
1455 | if( op == Op_If && |
1456 | (cmp=in(1)->in(1))->Opcode() == Op_CmpP ) { |
1457 | if( cmp->in(2) != NULL && // make sure cmp is not already dead |
1458 | cmp->in(2)->bottom_type() == TypePtr::NULL_PTR ) { |
1459 | dist = 64; // Limit for null-pointer scans |
1460 | } else { |
1461 | dist = 4; // Do not bother for random pointer tests |
1462 | } |
1463 | } else { |
1464 | dist = 4; // Limit for random junky scans |
1465 | } |
1466 | |
1467 | Node* prev_dom = search_identical(dist); |
1468 | |
1469 | if (prev_dom == NULL) { |
1470 | return NULL; |
1471 | } |
1472 | |
1473 | // Replace dominated IfNode |
1474 | return dominated_by(prev_dom, igvn); |
1475 | } |
1476 | |
1477 | //------------------------------dominated_by----------------------------------- |
1478 | Node* IfNode::dominated_by(Node* prev_dom, PhaseIterGVN *igvn) { |
1479 | #ifndef PRODUCT |
1480 | if (TraceIterativeGVN) { |
1481 | tty->print(" Removing IfNode: " ); this->dump(); |
1482 | } |
1483 | if (VerifyOpto && !igvn->allow_progress()) { |
1484 | // Found an equivalent dominating test, |
1485 | // we can not guarantee reaching a fix-point for these during iterativeGVN |
1486 | // since intervening nodes may not change. |
1487 | return NULL; |
1488 | } |
1489 | #endif |
1490 | |
1491 | igvn->hash_delete(this); // Remove self to prevent spurious V-N |
1492 | Node *idom = in(0); |
1493 | // Need opcode to decide which way 'this' test goes |
1494 | int prev_op = prev_dom->Opcode(); |
1495 | Node *top = igvn->C->top(); // Shortcut to top |
1496 | |
1497 | // Loop predicates may have depending checks which should not |
1498 | // be skipped. For example, range check predicate has two checks |
1499 | // for lower and upper bounds. |
1500 | ProjNode* unc_proj = proj_out(1 - prev_dom->as_Proj()->_con)->as_Proj(); |
1501 | if (unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate) != NULL || |
1502 | unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_profile_predicate) != NULL) { |
1503 | prev_dom = idom; |
1504 | } |
1505 | |
1506 | // Now walk the current IfNode's projections. |
1507 | // Loop ends when 'this' has no more uses. |
1508 | for (DUIterator_Last imin, i = last_outs(imin); i >= imin; --i) { |
1509 | Node *ifp = last_out(i); // Get IfTrue/IfFalse |
1510 | igvn->add_users_to_worklist(ifp); |
1511 | // Check which projection it is and set target. |
1512 | // Data-target is either the dominating projection of the same type |
1513 | // or TOP if the dominating projection is of opposite type. |
1514 | // Data-target will be used as the new control edge for the non-CFG |
1515 | // nodes like Casts and Loads. |
1516 | Node *data_target = (ifp->Opcode() == prev_op) ? prev_dom : top; |
1517 | // Control-target is just the If's immediate dominator or TOP. |
1518 | Node *ctrl_target = (ifp->Opcode() == prev_op) ? idom : top; |
1519 | |
1520 | // For each child of an IfTrue/IfFalse projection, reroute. |
1521 | // Loop ends when projection has no more uses. |
1522 | for (DUIterator_Last jmin, j = ifp->last_outs(jmin); j >= jmin; --j) { |
1523 | Node* s = ifp->last_out(j); // Get child of IfTrue/IfFalse |
1524 | if( !s->depends_only_on_test() ) { |
1525 | // Find the control input matching this def-use edge. |
1526 | // For Regions it may not be in slot 0. |
1527 | uint l; |
1528 | for( l = 0; s->in(l) != ifp; l++ ) { } |
1529 | igvn->replace_input_of(s, l, ctrl_target); |
1530 | } else { // Else, for control producers, |
1531 | igvn->replace_input_of(s, 0, data_target); // Move child to data-target |
1532 | } |
1533 | } // End for each child of a projection |
1534 | |
1535 | igvn->remove_dead_node(ifp); |
1536 | } // End for each IfTrue/IfFalse child of If |
1537 | |
1538 | // Kill the IfNode |
1539 | igvn->remove_dead_node(this); |
1540 | |
1541 | // Must return either the original node (now dead) or a new node |
1542 | // (Do not return a top here, since that would break the uniqueness of top.) |
1543 | return new ConINode(TypeInt::ZERO); |
1544 | } |
1545 | |
1546 | Node* IfNode::search_identical(int dist) { |
1547 | // Setup to scan up the CFG looking for a dominating test |
1548 | Node* dom = in(0); |
1549 | Node* prev_dom = this; |
1550 | int op = Opcode(); |
1551 | // Search up the dominator tree for an If with an identical test |
1552 | while (dom->Opcode() != op || // Not same opcode? |
1553 | dom->in(1) != in(1) || // Not same input 1? |
1554 | prev_dom->in(0) != dom) { // One path of test does not dominate? |
1555 | if (dist < 0) return NULL; |
1556 | |
1557 | dist--; |
1558 | prev_dom = dom; |
1559 | dom = up_one_dom(dom); |
1560 | if (!dom) return NULL; |
1561 | } |
1562 | |
1563 | // Check that we did not follow a loop back to ourselves |
1564 | if (this == dom) { |
1565 | return NULL; |
1566 | } |
1567 | |
1568 | #ifndef PRODUCT |
1569 | if (dist > 2) { // Add to count of NULL checks elided |
1570 | explicit_null_checks_elided++; |
1571 | } |
1572 | #endif |
1573 | |
1574 | return prev_dom; |
1575 | } |
1576 | |
1577 | //------------------------------Identity--------------------------------------- |
1578 | // If the test is constant & we match, then we are the input Control |
1579 | Node* IfProjNode::Identity(PhaseGVN* phase) { |
1580 | // Can only optimize if cannot go the other way |
1581 | const TypeTuple *t = phase->type(in(0))->is_tuple(); |
1582 | if (t == TypeTuple::IFNEITHER || (always_taken(t) && |
1583 | // During parsing (GVN) we don't remove dead code aggressively. |
1584 | // Cut off dead branch and let PhaseRemoveUseless take care of it. |
1585 | (!phase->is_IterGVN() || |
1586 | // During IGVN, first wait for the dead branch to be killed. |
1587 | // Otherwise, the IfNode's control will have two control uses (the IfNode |
1588 | // that doesn't go away because it still has uses and this branch of the |
1589 | // If) which breaks other optimizations. Node::has_special_unique_user() |
1590 | // will cause this node to be reprocessed once the dead branch is killed. |
1591 | in(0)->outcnt() == 1))) { |
1592 | // IfNode control |
1593 | return in(0)->in(0); |
1594 | } |
1595 | // no progress |
1596 | return this; |
1597 | } |
1598 | |
1599 | #ifndef PRODUCT |
1600 | //-------------------------------related--------------------------------------- |
1601 | // An IfProjNode's related node set consists of its input (an IfNode) including |
1602 | // the IfNode's condition, plus all of its outputs at level 1. In compact mode, |
1603 | // the restrictions for IfNode apply (see IfNode::rel). |
1604 | void IfProjNode::related(GrowableArray<Node*> *in_rel, GrowableArray<Node*> *out_rel, bool compact) const { |
1605 | Node* ifNode = this->in(0); |
1606 | in_rel->append(ifNode); |
1607 | if (compact) { |
1608 | ifNode->collect_nodes(in_rel, 3, false, true); |
1609 | } else { |
1610 | ifNode->collect_nodes_in_all_data(in_rel, false); |
1611 | } |
1612 | this->collect_nodes(out_rel, -1, false, false); |
1613 | } |
1614 | |
1615 | //------------------------------dump_spec-------------------------------------- |
1616 | void IfNode::dump_spec(outputStream *st) const { |
1617 | st->print("P=%f, C=%f" ,_prob,_fcnt); |
1618 | } |
1619 | |
1620 | //-------------------------------related--------------------------------------- |
1621 | // For an IfNode, the set of related output nodes is just the output nodes till |
1622 | // depth 2, i.e, the IfTrue/IfFalse projection nodes plus the nodes they refer. |
1623 | // The related input nodes contain no control nodes, but all data nodes |
1624 | // pertaining to the condition. In compact mode, the input nodes are collected |
1625 | // up to a depth of 3. |
1626 | void IfNode::related(GrowableArray <Node *> *in_rel, GrowableArray <Node *> *out_rel, bool compact) const { |
1627 | if (compact) { |
1628 | this->collect_nodes(in_rel, 3, false, true); |
1629 | } else { |
1630 | this->collect_nodes_in_all_data(in_rel, false); |
1631 | } |
1632 | this->collect_nodes(out_rel, -2, false, false); |
1633 | } |
1634 | #endif |
1635 | |
1636 | //------------------------------idealize_test---------------------------------- |
1637 | // Try to canonicalize tests better. Peek at the Cmp/Bool/If sequence and |
1638 | // come up with a canonical sequence. Bools getting 'eq', 'gt' and 'ge' forms |
1639 | // converted to 'ne', 'le' and 'lt' forms. IfTrue/IfFalse get swapped as |
1640 | // needed. |
1641 | static IfNode* idealize_test(PhaseGVN* phase, IfNode* iff) { |
1642 | assert(iff->in(0) != NULL, "If must be live" ); |
1643 | |
1644 | if (iff->outcnt() != 2) return NULL; // Malformed projections. |
1645 | Node* old_if_f = iff->proj_out(false); |
1646 | Node* old_if_t = iff->proj_out(true); |
1647 | |
1648 | // CountedLoopEnds want the back-control test to be TRUE, irregardless of |
1649 | // whether they are testing a 'gt' or 'lt' condition. The 'gt' condition |
1650 | // happens in count-down loops |
1651 | if (iff->is_CountedLoopEnd()) return NULL; |
1652 | if (!iff->in(1)->is_Bool()) return NULL; // Happens for partially optimized IF tests |
1653 | BoolNode *b = iff->in(1)->as_Bool(); |
1654 | BoolTest bt = b->_test; |
1655 | // Test already in good order? |
1656 | if( bt.is_canonical() ) |
1657 | return NULL; |
1658 | |
1659 | // Flip test to be canonical. Requires flipping the IfFalse/IfTrue and |
1660 | // cloning the IfNode. |
1661 | Node* new_b = phase->transform( new BoolNode(b->in(1), bt.negate()) ); |
1662 | if( !new_b->is_Bool() ) return NULL; |
1663 | b = new_b->as_Bool(); |
1664 | |
1665 | PhaseIterGVN *igvn = phase->is_IterGVN(); |
1666 | assert( igvn, "Test is not canonical in parser?" ); |
1667 | |
1668 | // The IF node never really changes, but it needs to be cloned |
1669 | iff = iff->clone()->as_If(); |
1670 | iff->set_req(1, b); |
1671 | iff->_prob = 1.0-iff->_prob; |
1672 | |
1673 | Node *prior = igvn->hash_find_insert(iff); |
1674 | if( prior ) { |
1675 | igvn->remove_dead_node(iff); |
1676 | iff = (IfNode*)prior; |
1677 | } else { |
1678 | // Cannot call transform on it just yet |
1679 | igvn->set_type_bottom(iff); |
1680 | } |
1681 | igvn->_worklist.push(iff); |
1682 | |
1683 | // Now handle projections. Cloning not required. |
1684 | Node* new_if_f = (Node*)(new IfFalseNode( iff )); |
1685 | Node* new_if_t = (Node*)(new IfTrueNode ( iff )); |
1686 | |
1687 | igvn->register_new_node_with_optimizer(new_if_f); |
1688 | igvn->register_new_node_with_optimizer(new_if_t); |
1689 | // Flip test, so flip trailing control |
1690 | igvn->replace_node(old_if_f, new_if_t); |
1691 | igvn->replace_node(old_if_t, new_if_f); |
1692 | |
1693 | // Progress |
1694 | return iff; |
1695 | } |
1696 | |
1697 | Node* RangeCheckNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
1698 | Node* res = Ideal_common(phase, can_reshape); |
1699 | if (res != NodeSentinel) { |
1700 | return res; |
1701 | } |
1702 | |
1703 | PhaseIterGVN *igvn = phase->is_IterGVN(); |
1704 | // Setup to scan up the CFG looking for a dominating test |
1705 | Node* prev_dom = this; |
1706 | |
1707 | // Check for range-check vs other kinds of tests |
1708 | Node* index1; |
1709 | Node* range1; |
1710 | jint offset1; |
1711 | int flip1 = is_range_check(range1, index1, offset1); |
1712 | if (flip1) { |
1713 | Node* dom = in(0); |
1714 | // Try to remove extra range checks. All 'up_one_dom' gives up at merges |
1715 | // so all checks we inspect post-dominate the top-most check we find. |
1716 | // If we are going to fail the current check and we reach the top check |
1717 | // then we are guaranteed to fail, so just start interpreting there. |
1718 | // We 'expand' the top 3 range checks to include all post-dominating |
1719 | // checks. |
1720 | |
1721 | // The top 3 range checks seen |
1722 | const int NRC =3; |
1723 | RangeCheck prev_checks[NRC]; |
1724 | int nb_checks = 0; |
1725 | |
1726 | // Low and high offsets seen so far |
1727 | jint off_lo = offset1; |
1728 | jint off_hi = offset1; |
1729 | |
1730 | bool found_immediate_dominator = false; |
1731 | |
1732 | // Scan for the top checks and collect range of offsets |
1733 | for (int dist = 0; dist < 999; dist++) { // Range-Check scan limit |
1734 | if (dom->Opcode() == Op_RangeCheck && // Not same opcode? |
1735 | prev_dom->in(0) == dom) { // One path of test does dominate? |
1736 | if (dom == this) return NULL; // dead loop |
1737 | // See if this is a range check |
1738 | Node* index2; |
1739 | Node* range2; |
1740 | jint offset2; |
1741 | int flip2 = dom->as_RangeCheck()->is_range_check(range2, index2, offset2); |
1742 | // See if this is a _matching_ range check, checking against |
1743 | // the same array bounds. |
1744 | if (flip2 == flip1 && range2 == range1 && index2 == index1 && |
1745 | dom->outcnt() == 2) { |
1746 | if (nb_checks == 0 && dom->in(1) == in(1)) { |
1747 | // Found an immediately dominating test at the same offset. |
1748 | // This kind of back-to-back test can be eliminated locally, |
1749 | // and there is no need to search further for dominating tests. |
1750 | assert(offset2 == offset1, "Same test but different offsets" ); |
1751 | found_immediate_dominator = true; |
1752 | break; |
1753 | } |
1754 | // Gather expanded bounds |
1755 | off_lo = MIN2(off_lo,offset2); |
1756 | off_hi = MAX2(off_hi,offset2); |
1757 | // Record top NRC range checks |
1758 | prev_checks[nb_checks%NRC].ctl = prev_dom; |
1759 | prev_checks[nb_checks%NRC].off = offset2; |
1760 | nb_checks++; |
1761 | } |
1762 | } |
1763 | prev_dom = dom; |
1764 | dom = up_one_dom(dom); |
1765 | if (!dom) break; |
1766 | } |
1767 | |
1768 | if (!found_immediate_dominator) { |
1769 | // Attempt to widen the dominating range check to cover some later |
1770 | // ones. Since range checks "fail" by uncommon-trapping to the |
1771 | // interpreter, widening a check can make us speculatively enter |
1772 | // the interpreter. If we see range-check deopt's, do not widen! |
1773 | if (!phase->C->allow_range_check_smearing()) return NULL; |
1774 | |
1775 | // Didn't find prior covering check, so cannot remove anything. |
1776 | if (nb_checks == 0) { |
1777 | return NULL; |
1778 | } |
1779 | // Constant indices only need to check the upper bound. |
1780 | // Non-constant indices must check both low and high. |
1781 | int chk0 = (nb_checks - 1) % NRC; |
1782 | if (index1) { |
1783 | if (nb_checks == 1) { |
1784 | return NULL; |
1785 | } else { |
1786 | // If the top range check's constant is the min or max of |
1787 | // all constants we widen the next one to cover the whole |
1788 | // range of constants. |
1789 | RangeCheck rc0 = prev_checks[chk0]; |
1790 | int chk1 = (nb_checks - 2) % NRC; |
1791 | RangeCheck rc1 = prev_checks[chk1]; |
1792 | if (rc0.off == off_lo) { |
1793 | adjust_check(rc1.ctl, range1, index1, flip1, off_hi, igvn); |
1794 | prev_dom = rc1.ctl; |
1795 | } else if (rc0.off == off_hi) { |
1796 | adjust_check(rc1.ctl, range1, index1, flip1, off_lo, igvn); |
1797 | prev_dom = rc1.ctl; |
1798 | } else { |
1799 | // If the top test's constant is not the min or max of all |
1800 | // constants, we need 3 range checks. We must leave the |
1801 | // top test unchanged because widening it would allow the |
1802 | // accesses it protects to successfully read/write out of |
1803 | // bounds. |
1804 | if (nb_checks == 2) { |
1805 | return NULL; |
1806 | } |
1807 | int chk2 = (nb_checks - 3) % NRC; |
1808 | RangeCheck rc2 = prev_checks[chk2]; |
1809 | // The top range check a+i covers interval: -a <= i < length-a |
1810 | // The second range check b+i covers interval: -b <= i < length-b |
1811 | if (rc1.off <= rc0.off) { |
1812 | // if b <= a, we change the second range check to: |
1813 | // -min_of_all_constants <= i < length-min_of_all_constants |
1814 | // Together top and second range checks now cover: |
1815 | // -min_of_all_constants <= i < length-a |
1816 | // which is more restrictive than -b <= i < length-b: |
1817 | // -b <= -min_of_all_constants <= i < length-a <= length-b |
1818 | // The third check is then changed to: |
1819 | // -max_of_all_constants <= i < length-max_of_all_constants |
1820 | // so 2nd and 3rd checks restrict allowed values of i to: |
1821 | // -min_of_all_constants <= i < length-max_of_all_constants |
1822 | adjust_check(rc1.ctl, range1, index1, flip1, off_lo, igvn); |
1823 | adjust_check(rc2.ctl, range1, index1, flip1, off_hi, igvn); |
1824 | } else { |
1825 | // if b > a, we change the second range check to: |
1826 | // -max_of_all_constants <= i < length-max_of_all_constants |
1827 | // Together top and second range checks now cover: |
1828 | // -a <= i < length-max_of_all_constants |
1829 | // which is more restrictive than -b <= i < length-b: |
1830 | // -b < -a <= i < length-max_of_all_constants <= length-b |
1831 | // The third check is then changed to: |
1832 | // -max_of_all_constants <= i < length-max_of_all_constants |
1833 | // so 2nd and 3rd checks restrict allowed values of i to: |
1834 | // -min_of_all_constants <= i < length-max_of_all_constants |
1835 | adjust_check(rc1.ctl, range1, index1, flip1, off_hi, igvn); |
1836 | adjust_check(rc2.ctl, range1, index1, flip1, off_lo, igvn); |
1837 | } |
1838 | prev_dom = rc2.ctl; |
1839 | } |
1840 | } |
1841 | } else { |
1842 | RangeCheck rc0 = prev_checks[chk0]; |
1843 | // 'Widen' the offset of the 1st and only covering check |
1844 | adjust_check(rc0.ctl, range1, index1, flip1, off_hi, igvn); |
1845 | // Test is now covered by prior checks, dominate it out |
1846 | prev_dom = rc0.ctl; |
1847 | } |
1848 | } |
1849 | } else { |
1850 | prev_dom = search_identical(4); |
1851 | |
1852 | if (prev_dom == NULL) { |
1853 | return NULL; |
1854 | } |
1855 | } |
1856 | |
1857 | // Replace dominated IfNode |
1858 | return dominated_by(prev_dom, igvn); |
1859 | } |
1860 | |