Print this page
rev 6883 : 8057622: java/util/stream/test/org/openjdk/tests/java/util/stream/InfiniteStreamWithLimitOpTest: SEGV inside compiled code (sparc)
Summary: In Parse::array_store_check(), add control edge FROM IfTrue branch of runtime type check of the destination array TO loading _element_klass from destination array.
Reviewed-by: kvn, roland, anoll
Contributed-by: Zoltan Majo <zoltan.majo@oracle.com>
Split |
Split |
Close |
Expand all |
Collapse all |
--- old/hotspot/src/share/vm/opto/graphKit.cpp
+++ new/hotspot/src/share/vm/opto/graphKit.cpp
1 1 /*
2 2 * Copyright (c) 2001, 2013, Oracle and/or its affiliates. All rights reserved.
3 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 4 *
5 5 * This code is free software; you can redistribute it and/or modify it
6 6 * under the terms of the GNU General Public License version 2 only, as
7 7 * published by the Free Software Foundation.
8 8 *
9 9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 12 * version 2 for more details (a copy is included in the LICENSE file that
13 13 * accompanied this code).
14 14 *
15 15 * You should have received a copy of the GNU General Public License version
16 16 * 2 along with this work; if not, write to the Free Software Foundation,
17 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 18 *
19 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 20 * or visit www.oracle.com if you need additional information or have any
21 21 * questions.
22 22 *
23 23 */
24 24
25 25 #include "precompiled.hpp"
26 26 #include "compiler/compileLog.hpp"
27 27 #include "gc_implementation/g1/g1SATBCardTableModRefBS.hpp"
28 28 #include "gc_implementation/g1/heapRegion.hpp"
29 29 #include "gc_interface/collectedHeap.hpp"
30 30 #include "memory/barrierSet.hpp"
31 31 #include "memory/cardTableModRefBS.hpp"
32 32 #include "opto/addnode.hpp"
33 33 #include "opto/graphKit.hpp"
34 34 #include "opto/idealKit.hpp"
35 35 #include "opto/locknode.hpp"
36 36 #include "opto/machnode.hpp"
37 37 #include "opto/parse.hpp"
38 38 #include "opto/rootnode.hpp"
39 39 #include "opto/runtime.hpp"
40 40 #include "runtime/deoptimization.hpp"
41 41 #include "runtime/sharedRuntime.hpp"
42 42
43 43 //----------------------------GraphKit-----------------------------------------
44 44 // Main utility constructor.
45 45 GraphKit::GraphKit(JVMState* jvms)
46 46 : Phase(Phase::Parser),
47 47 _env(C->env()),
48 48 _gvn(*C->initial_gvn())
49 49 {
50 50 _exceptions = jvms->map()->next_exception();
51 51 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL);
52 52 set_jvms(jvms);
53 53 }
54 54
55 55 // Private constructor for parser.
56 56 GraphKit::GraphKit()
57 57 : Phase(Phase::Parser),
58 58 _env(C->env()),
59 59 _gvn(*C->initial_gvn())
60 60 {
61 61 _exceptions = NULL;
62 62 set_map(NULL);
63 63 debug_only(_sp = -99);
64 64 debug_only(set_bci(-99));
65 65 }
66 66
67 67
68 68
69 69 //---------------------------clean_stack---------------------------------------
70 70 // Clear away rubbish from the stack area of the JVM state.
71 71 // This destroys any arguments that may be waiting on the stack.
72 72 void GraphKit::clean_stack(int from_sp) {
73 73 SafePointNode* map = this->map();
74 74 JVMState* jvms = this->jvms();
75 75 int stk_size = jvms->stk_size();
76 76 int stkoff = jvms->stkoff();
77 77 Node* top = this->top();
78 78 for (int i = from_sp; i < stk_size; i++) {
79 79 if (map->in(stkoff + i) != top) {
80 80 map->set_req(stkoff + i, top);
81 81 }
82 82 }
83 83 }
84 84
85 85
86 86 //--------------------------------sync_jvms-----------------------------------
87 87 // Make sure our current jvms agrees with our parse state.
88 88 JVMState* GraphKit::sync_jvms() const {
89 89 JVMState* jvms = this->jvms();
90 90 jvms->set_bci(bci()); // Record the new bci in the JVMState
91 91 jvms->set_sp(sp()); // Record the new sp in the JVMState
92 92 assert(jvms_in_sync(), "jvms is now in sync");
93 93 return jvms;
94 94 }
95 95
96 96 //--------------------------------sync_jvms_for_reexecute---------------------
97 97 // Make sure our current jvms agrees with our parse state. This version
98 98 // uses the reexecute_sp for reexecuting bytecodes.
99 99 JVMState* GraphKit::sync_jvms_for_reexecute() {
100 100 JVMState* jvms = this->jvms();
101 101 jvms->set_bci(bci()); // Record the new bci in the JVMState
102 102 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState
103 103 return jvms;
104 104 }
105 105
106 106 #ifdef ASSERT
107 107 bool GraphKit::jvms_in_sync() const {
108 108 Parse* parse = is_Parse();
109 109 if (parse == NULL) {
110 110 if (bci() != jvms()->bci()) return false;
111 111 if (sp() != (int)jvms()->sp()) return false;
112 112 return true;
113 113 }
114 114 if (jvms()->method() != parse->method()) return false;
115 115 if (jvms()->bci() != parse->bci()) return false;
116 116 int jvms_sp = jvms()->sp();
117 117 if (jvms_sp != parse->sp()) return false;
118 118 int jvms_depth = jvms()->depth();
119 119 if (jvms_depth != parse->depth()) return false;
120 120 return true;
121 121 }
122 122
123 123 // Local helper checks for special internal merge points
124 124 // used to accumulate and merge exception states.
125 125 // They are marked by the region's in(0) edge being the map itself.
126 126 // Such merge points must never "escape" into the parser at large,
127 127 // until they have been handed to gvn.transform.
128 128 static bool is_hidden_merge(Node* reg) {
129 129 if (reg == NULL) return false;
130 130 if (reg->is_Phi()) {
131 131 reg = reg->in(0);
132 132 if (reg == NULL) return false;
133 133 }
134 134 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root();
135 135 }
136 136
137 137 void GraphKit::verify_map() const {
138 138 if (map() == NULL) return; // null map is OK
139 139 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
140 140 assert(!map()->has_exceptions(), "call add_exception_states_from 1st");
141 141 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
142 142 }
143 143
144 144 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
145 145 assert(ex_map->next_exception() == NULL, "not already part of a chain");
146 146 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
147 147 }
148 148 #endif
149 149
150 150 //---------------------------stop_and_kill_map---------------------------------
151 151 // Set _map to NULL, signalling a stop to further bytecode execution.
152 152 // First smash the current map's control to a constant, to mark it dead.
153 153 void GraphKit::stop_and_kill_map() {
154 154 SafePointNode* dead_map = stop();
155 155 if (dead_map != NULL) {
156 156 dead_map->disconnect_inputs(NULL, C); // Mark the map as killed.
157 157 assert(dead_map->is_killed(), "must be so marked");
158 158 }
159 159 }
160 160
161 161
162 162 //--------------------------------stopped--------------------------------------
163 163 // Tell if _map is NULL, or control is top.
164 164 bool GraphKit::stopped() {
165 165 if (map() == NULL) return true;
166 166 else if (control() == top()) return true;
167 167 else return false;
168 168 }
169 169
170 170
171 171 //-----------------------------has_ex_handler----------------------------------
172 172 // Tell if this method or any caller method has exception handlers.
173 173 bool GraphKit::has_ex_handler() {
174 174 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) {
175 175 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
176 176 return true;
177 177 }
178 178 }
179 179 return false;
180 180 }
181 181
182 182 //------------------------------save_ex_oop------------------------------------
183 183 // Save an exception without blowing stack contents or other JVM state.
184 184 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
185 185 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
186 186 ex_map->add_req(ex_oop);
187 187 debug_only(verify_exception_state(ex_map));
188 188 }
189 189
190 190 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
191 191 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
192 192 Node* ex_oop = ex_map->in(ex_map->req()-1);
193 193 if (clear_it) ex_map->del_req(ex_map->req()-1);
194 194 return ex_oop;
195 195 }
196 196
197 197 //-----------------------------saved_ex_oop------------------------------------
198 198 // Recover a saved exception from its map.
199 199 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
200 200 return common_saved_ex_oop(ex_map, false);
201 201 }
202 202
203 203 //--------------------------clear_saved_ex_oop---------------------------------
204 204 // Erase a previously saved exception from its map.
205 205 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
206 206 return common_saved_ex_oop(ex_map, true);
207 207 }
208 208
209 209 #ifdef ASSERT
210 210 //---------------------------has_saved_ex_oop----------------------------------
211 211 // Erase a previously saved exception from its map.
212 212 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
213 213 return ex_map->req() == ex_map->jvms()->endoff()+1;
214 214 }
215 215 #endif
216 216
217 217 //-------------------------make_exception_state--------------------------------
218 218 // Turn the current JVM state into an exception state, appending the ex_oop.
219 219 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
220 220 sync_jvms();
221 221 SafePointNode* ex_map = stop(); // do not manipulate this map any more
222 222 set_saved_ex_oop(ex_map, ex_oop);
223 223 return ex_map;
224 224 }
225 225
226 226
227 227 //--------------------------add_exception_state--------------------------------
228 228 // Add an exception to my list of exceptions.
229 229 void GraphKit::add_exception_state(SafePointNode* ex_map) {
230 230 if (ex_map == NULL || ex_map->control() == top()) {
231 231 return;
232 232 }
233 233 #ifdef ASSERT
234 234 verify_exception_state(ex_map);
235 235 if (has_exceptions()) {
236 236 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
237 237 }
238 238 #endif
239 239
240 240 // If there is already an exception of exactly this type, merge with it.
241 241 // In particular, null-checks and other low-level exceptions common up here.
242 242 Node* ex_oop = saved_ex_oop(ex_map);
243 243 const Type* ex_type = _gvn.type(ex_oop);
244 244 if (ex_oop == top()) {
245 245 // No action needed.
246 246 return;
247 247 }
248 248 assert(ex_type->isa_instptr(), "exception must be an instance");
249 249 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) {
250 250 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
251 251 // We check sp also because call bytecodes can generate exceptions
252 252 // both before and after arguments are popped!
253 253 if (ex_type2 == ex_type
254 254 && e2->_jvms->sp() == ex_map->_jvms->sp()) {
255 255 combine_exception_states(ex_map, e2);
256 256 return;
257 257 }
258 258 }
259 259
260 260 // No pre-existing exception of the same type. Chain it on the list.
261 261 push_exception_state(ex_map);
262 262 }
263 263
264 264 //-----------------------add_exception_states_from-----------------------------
265 265 void GraphKit::add_exception_states_from(JVMState* jvms) {
266 266 SafePointNode* ex_map = jvms->map()->next_exception();
267 267 if (ex_map != NULL) {
268 268 jvms->map()->set_next_exception(NULL);
269 269 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) {
270 270 next_map = ex_map->next_exception();
271 271 ex_map->set_next_exception(NULL);
272 272 add_exception_state(ex_map);
273 273 }
274 274 }
275 275 }
276 276
277 277 //-----------------------transfer_exceptions_into_jvms-------------------------
278 278 JVMState* GraphKit::transfer_exceptions_into_jvms() {
279 279 if (map() == NULL) {
280 280 // We need a JVMS to carry the exceptions, but the map has gone away.
281 281 // Create a scratch JVMS, cloned from any of the exception states...
282 282 if (has_exceptions()) {
283 283 _map = _exceptions;
284 284 _map = clone_map();
285 285 _map->set_next_exception(NULL);
286 286 clear_saved_ex_oop(_map);
287 287 debug_only(verify_map());
288 288 } else {
289 289 // ...or created from scratch
290 290 JVMState* jvms = new (C) JVMState(_method, NULL);
291 291 jvms->set_bci(_bci);
292 292 jvms->set_sp(_sp);
293 293 jvms->set_map(new (C) SafePointNode(TypeFunc::Parms, jvms));
294 294 set_jvms(jvms);
295 295 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top());
296 296 set_all_memory(top());
297 297 while (map()->req() < jvms->endoff()) map()->add_req(top());
298 298 }
299 299 // (This is a kludge, in case you didn't notice.)
300 300 set_control(top());
301 301 }
302 302 JVMState* jvms = sync_jvms();
303 303 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
304 304 jvms->map()->set_next_exception(_exceptions);
305 305 _exceptions = NULL; // done with this set of exceptions
306 306 return jvms;
307 307 }
308 308
309 309 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
310 310 assert(is_hidden_merge(dstphi), "must be a special merge node");
311 311 assert(is_hidden_merge(srcphi), "must be a special merge node");
312 312 uint limit = srcphi->req();
313 313 for (uint i = PhiNode::Input; i < limit; i++) {
314 314 dstphi->add_req(srcphi->in(i));
315 315 }
316 316 }
317 317 static inline void add_one_req(Node* dstphi, Node* src) {
318 318 assert(is_hidden_merge(dstphi), "must be a special merge node");
319 319 assert(!is_hidden_merge(src), "must not be a special merge node");
320 320 dstphi->add_req(src);
321 321 }
322 322
323 323 //-----------------------combine_exception_states------------------------------
324 324 // This helper function combines exception states by building phis on a
325 325 // specially marked state-merging region. These regions and phis are
326 326 // untransformed, and can build up gradually. The region is marked by
327 327 // having a control input of its exception map, rather than NULL. Such
328 328 // regions do not appear except in this function, and in use_exception_state.
329 329 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
330 330 if (failing()) return; // dying anyway...
331 331 JVMState* ex_jvms = ex_map->_jvms;
332 332 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
333 333 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
334 334 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
335 335 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
336 336 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
337 337 assert(ex_map->req() == phi_map->req(), "matching maps");
338 338 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
339 339 Node* hidden_merge_mark = root();
340 340 Node* region = phi_map->control();
341 341 MergeMemNode* phi_mem = phi_map->merged_memory();
342 342 MergeMemNode* ex_mem = ex_map->merged_memory();
343 343 if (region->in(0) != hidden_merge_mark) {
344 344 // The control input is not (yet) a specially-marked region in phi_map.
345 345 // Make it so, and build some phis.
346 346 region = new (C) RegionNode(2);
347 347 _gvn.set_type(region, Type::CONTROL);
348 348 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
349 349 region->init_req(1, phi_map->control());
350 350 phi_map->set_control(region);
351 351 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
352 352 record_for_igvn(io_phi);
353 353 _gvn.set_type(io_phi, Type::ABIO);
354 354 phi_map->set_i_o(io_phi);
355 355 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
356 356 Node* m = mms.memory();
357 357 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
358 358 record_for_igvn(m_phi);
359 359 _gvn.set_type(m_phi, Type::MEMORY);
360 360 mms.set_memory(m_phi);
361 361 }
362 362 }
363 363
364 364 // Either or both of phi_map and ex_map might already be converted into phis.
365 365 Node* ex_control = ex_map->control();
366 366 // if there is special marking on ex_map also, we add multiple edges from src
367 367 bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
368 368 // how wide was the destination phi_map, originally?
369 369 uint orig_width = region->req();
370 370
371 371 if (add_multiple) {
372 372 add_n_reqs(region, ex_control);
373 373 add_n_reqs(phi_map->i_o(), ex_map->i_o());
374 374 } else {
375 375 // ex_map has no merges, so we just add single edges everywhere
376 376 add_one_req(region, ex_control);
377 377 add_one_req(phi_map->i_o(), ex_map->i_o());
378 378 }
379 379 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
380 380 if (mms.is_empty()) {
381 381 // get a copy of the base memory, and patch some inputs into it
382 382 const TypePtr* adr_type = mms.adr_type(C);
383 383 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
384 384 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
385 385 mms.set_memory(phi);
386 386 // Prepare to append interesting stuff onto the newly sliced phi:
387 387 while (phi->req() > orig_width) phi->del_req(phi->req()-1);
388 388 }
389 389 // Append stuff from ex_map:
390 390 if (add_multiple) {
391 391 add_n_reqs(mms.memory(), mms.memory2());
392 392 } else {
393 393 add_one_req(mms.memory(), mms.memory2());
394 394 }
395 395 }
396 396 uint limit = ex_map->req();
397 397 for (uint i = TypeFunc::Parms; i < limit; i++) {
398 398 // Skip everything in the JVMS after tos. (The ex_oop follows.)
399 399 if (i == tos) i = ex_jvms->monoff();
400 400 Node* src = ex_map->in(i);
401 401 Node* dst = phi_map->in(i);
402 402 if (src != dst) {
403 403 PhiNode* phi;
404 404 if (dst->in(0) != region) {
405 405 dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
406 406 record_for_igvn(phi);
407 407 _gvn.set_type(phi, phi->type());
408 408 phi_map->set_req(i, dst);
409 409 // Prepare to append interesting stuff onto the new phi:
410 410 while (dst->req() > orig_width) dst->del_req(dst->req()-1);
411 411 } else {
412 412 assert(dst->is_Phi(), "nobody else uses a hidden region");
413 413 phi = dst->as_Phi();
414 414 }
415 415 if (add_multiple && src->in(0) == ex_control) {
416 416 // Both are phis.
417 417 add_n_reqs(dst, src);
418 418 } else {
419 419 while (dst->req() < region->req()) add_one_req(dst, src);
420 420 }
421 421 const Type* srctype = _gvn.type(src);
422 422 if (phi->type() != srctype) {
423 423 const Type* dsttype = phi->type()->meet_speculative(srctype);
424 424 if (phi->type() != dsttype) {
425 425 phi->set_type(dsttype);
426 426 _gvn.set_type(phi, dsttype);
427 427 }
428 428 }
429 429 }
430 430 }
431 431 phi_map->merge_replaced_nodes_with(ex_map);
432 432 }
433 433
434 434 //--------------------------use_exception_state--------------------------------
435 435 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
436 436 if (failing()) { stop(); return top(); }
437 437 Node* region = phi_map->control();
438 438 Node* hidden_merge_mark = root();
439 439 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
440 440 Node* ex_oop = clear_saved_ex_oop(phi_map);
441 441 if (region->in(0) == hidden_merge_mark) {
442 442 // Special marking for internal ex-states. Process the phis now.
443 443 region->set_req(0, region); // now it's an ordinary region
444 444 set_jvms(phi_map->jvms()); // ...so now we can use it as a map
445 445 // Note: Setting the jvms also sets the bci and sp.
446 446 set_control(_gvn.transform(region));
447 447 uint tos = jvms()->stkoff() + sp();
448 448 for (uint i = 1; i < tos; i++) {
449 449 Node* x = phi_map->in(i);
450 450 if (x->in(0) == region) {
451 451 assert(x->is_Phi(), "expected a special phi");
452 452 phi_map->set_req(i, _gvn.transform(x));
453 453 }
454 454 }
455 455 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
456 456 Node* x = mms.memory();
457 457 if (x->in(0) == region) {
458 458 assert(x->is_Phi(), "nobody else uses a hidden region");
459 459 mms.set_memory(_gvn.transform(x));
460 460 }
461 461 }
462 462 if (ex_oop->in(0) == region) {
463 463 assert(ex_oop->is_Phi(), "expected a special phi");
464 464 ex_oop = _gvn.transform(ex_oop);
465 465 }
466 466 } else {
467 467 set_jvms(phi_map->jvms());
468 468 }
469 469
470 470 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
471 471 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
472 472 return ex_oop;
473 473 }
474 474
475 475 //---------------------------------java_bc-------------------------------------
476 476 Bytecodes::Code GraphKit::java_bc() const {
477 477 ciMethod* method = this->method();
478 478 int bci = this->bci();
479 479 if (method != NULL && bci != InvocationEntryBci)
480 480 return method->java_code_at_bci(bci);
481 481 else
482 482 return Bytecodes::_illegal;
483 483 }
484 484
485 485 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
486 486 bool must_throw) {
487 487 // if the exception capability is set, then we will generate code
488 488 // to check the JavaThread.should_post_on_exceptions flag to see
489 489 // if we actually need to report exception events (for this
490 490 // thread). If we don't need to report exception events, we will
491 491 // take the normal fast path provided by add_exception_events. If
492 492 // exception event reporting is enabled for this thread, we will
493 493 // take the uncommon_trap in the BuildCutout below.
494 494
495 495 // first must access the should_post_on_exceptions_flag in this thread's JavaThread
496 496 Node* jthread = _gvn.transform(new (C) ThreadLocalNode());
497 497 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
498 498 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered);
499 499
500 500 // Test the should_post_on_exceptions_flag vs. 0
501 501 Node* chk = _gvn.transform( new (C) CmpINode(should_post_flag, intcon(0)) );
502 502 Node* tst = _gvn.transform( new (C) BoolNode(chk, BoolTest::eq) );
503 503
504 504 // Branch to slow_path if should_post_on_exceptions_flag was true
505 505 { BuildCutout unless(this, tst, PROB_MAX);
506 506 // Do not try anything fancy if we're notifying the VM on every throw.
507 507 // Cf. case Bytecodes::_athrow in parse2.cpp.
508 508 uncommon_trap(reason, Deoptimization::Action_none,
509 509 (ciKlass*)NULL, (char*)NULL, must_throw);
510 510 }
511 511
512 512 }
513 513
514 514 //------------------------------builtin_throw----------------------------------
515 515 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
516 516 bool must_throw = true;
517 517
518 518 if (env()->jvmti_can_post_on_exceptions()) {
519 519 // check if we must post exception events, take uncommon trap if so
520 520 uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
521 521 // here if should_post_on_exceptions is false
522 522 // continue on with the normal codegen
523 523 }
524 524
525 525 // If this particular condition has not yet happened at this
526 526 // bytecode, then use the uncommon trap mechanism, and allow for
527 527 // a future recompilation if several traps occur here.
528 528 // If the throw is hot, try to use a more complicated inline mechanism
529 529 // which keeps execution inside the compiled code.
530 530 bool treat_throw_as_hot = false;
531 531 ciMethodData* md = method()->method_data();
532 532
533 533 if (ProfileTraps) {
534 534 if (too_many_traps(reason)) {
535 535 treat_throw_as_hot = true;
536 536 }
537 537 // (If there is no MDO at all, assume it is early in
538 538 // execution, and that any deopts are part of the
539 539 // startup transient, and don't need to be remembered.)
540 540
541 541 // Also, if there is a local exception handler, treat all throws
542 542 // as hot if there has been at least one in this method.
543 543 if (C->trap_count(reason) != 0
544 544 && method()->method_data()->trap_count(reason) != 0
545 545 && has_ex_handler()) {
546 546 treat_throw_as_hot = true;
547 547 }
548 548 }
549 549
550 550 // If this throw happens frequently, an uncommon trap might cause
551 551 // a performance pothole. If there is a local exception handler,
552 552 // and if this particular bytecode appears to be deoptimizing often,
553 553 // let us handle the throw inline, with a preconstructed instance.
554 554 // Note: If the deopt count has blown up, the uncommon trap
555 555 // runtime is going to flush this nmethod, not matter what.
556 556 if (treat_throw_as_hot
557 557 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
558 558 // If the throw is local, we use a pre-existing instance and
559 559 // punt on the backtrace. This would lead to a missing backtrace
560 560 // (a repeat of 4292742) if the backtrace object is ever asked
561 561 // for its backtrace.
562 562 // Fixing this remaining case of 4292742 requires some flavor of
563 563 // escape analysis. Leave that for the future.
564 564 ciInstance* ex_obj = NULL;
565 565 switch (reason) {
566 566 case Deoptimization::Reason_null_check:
567 567 ex_obj = env()->NullPointerException_instance();
568 568 break;
569 569 case Deoptimization::Reason_div0_check:
570 570 ex_obj = env()->ArithmeticException_instance();
571 571 break;
572 572 case Deoptimization::Reason_range_check:
573 573 ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
574 574 break;
575 575 case Deoptimization::Reason_class_check:
576 576 if (java_bc() == Bytecodes::_aastore) {
577 577 ex_obj = env()->ArrayStoreException_instance();
578 578 } else {
579 579 ex_obj = env()->ClassCastException_instance();
580 580 }
581 581 break;
582 582 }
583 583 if (failing()) { stop(); return; } // exception allocation might fail
584 584 if (ex_obj != NULL) {
585 585 // Cheat with a preallocated exception object.
586 586 if (C->log() != NULL)
587 587 C->log()->elem("hot_throw preallocated='1' reason='%s'",
588 588 Deoptimization::trap_reason_name(reason));
589 589 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj);
590 590 Node* ex_node = _gvn.transform( ConNode::make(C, ex_con) );
591 591
592 592 // Clear the detail message of the preallocated exception object.
593 593 // Weblogic sometimes mutates the detail message of exceptions
594 594 // using reflection.
595 595 int offset = java_lang_Throwable::get_detailMessage_offset();
596 596 const TypePtr* adr_typ = ex_con->add_offset(offset);
597 597
598 598 Node *adr = basic_plus_adr(ex_node, ex_node, offset);
599 599 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
600 600 // Conservatively release stores of object references.
601 601 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, MemNode::release);
602 602
603 603 add_exception_state(make_exception_state(ex_node));
604 604 return;
605 605 }
606 606 }
607 607
608 608 // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
609 609 // It won't be much cheaper than bailing to the interp., since we'll
610 610 // have to pass up all the debug-info, and the runtime will have to
611 611 // create the stack trace.
612 612
613 613 // Usual case: Bail to interpreter.
614 614 // Reserve the right to recompile if we haven't seen anything yet.
615 615
616 616 assert(!Deoptimization::reason_is_speculate(reason), "unsupported");
617 617 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
618 618 if (treat_throw_as_hot
619 619 && (method()->method_data()->trap_recompiled_at(bci(), NULL)
620 620 || C->too_many_traps(reason))) {
621 621 // We cannot afford to take more traps here. Suffer in the interpreter.
622 622 if (C->log() != NULL)
623 623 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
624 624 Deoptimization::trap_reason_name(reason),
625 625 C->trap_count(reason));
626 626 action = Deoptimization::Action_none;
627 627 }
628 628
629 629 // "must_throw" prunes the JVM state to include only the stack, if there
630 630 // are no local exception handlers. This should cut down on register
631 631 // allocation time and code size, by drastically reducing the number
632 632 // of in-edges on the call to the uncommon trap.
633 633
634 634 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
635 635 }
636 636
637 637
638 638 //----------------------------PreserveJVMState---------------------------------
639 639 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
640 640 debug_only(kit->verify_map());
641 641 _kit = kit;
642 642 _map = kit->map(); // preserve the map
643 643 _sp = kit->sp();
644 644 kit->set_map(clone_map ? kit->clone_map() : NULL);
645 645 #ifdef ASSERT
646 646 _bci = kit->bci();
647 647 Parse* parser = kit->is_Parse();
648 648 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
649 649 _block = block;
650 650 #endif
651 651 }
652 652 PreserveJVMState::~PreserveJVMState() {
653 653 GraphKit* kit = _kit;
654 654 #ifdef ASSERT
655 655 assert(kit->bci() == _bci, "bci must not shift");
656 656 Parse* parser = kit->is_Parse();
657 657 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
658 658 assert(block == _block, "block must not shift");
659 659 #endif
660 660 kit->set_map(_map);
661 661 kit->set_sp(_sp);
662 662 }
663 663
664 664
665 665 //-----------------------------BuildCutout-------------------------------------
666 666 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
667 667 : PreserveJVMState(kit)
668 668 {
669 669 assert(p->is_Con() || p->is_Bool(), "test must be a bool");
670 670 SafePointNode* outer_map = _map; // preserved map is caller's
671 671 SafePointNode* inner_map = kit->map();
672 672 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
673 673 outer_map->set_control(kit->gvn().transform( new (kit->C) IfTrueNode(iff) ));
674 674 inner_map->set_control(kit->gvn().transform( new (kit->C) IfFalseNode(iff) ));
675 675 }
676 676 BuildCutout::~BuildCutout() {
677 677 GraphKit* kit = _kit;
678 678 assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
679 679 }
680 680
681 681 //---------------------------PreserveReexecuteState----------------------------
682 682 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
683 683 assert(!kit->stopped(), "must call stopped() before");
684 684 _kit = kit;
685 685 _sp = kit->sp();
686 686 _reexecute = kit->jvms()->_reexecute;
687 687 }
688 688 PreserveReexecuteState::~PreserveReexecuteState() {
689 689 if (_kit->stopped()) return;
690 690 _kit->jvms()->_reexecute = _reexecute;
691 691 _kit->set_sp(_sp);
692 692 }
693 693
694 694 //------------------------------clone_map--------------------------------------
695 695 // Implementation of PreserveJVMState
696 696 //
697 697 // Only clone_map(...) here. If this function is only used in the
698 698 // PreserveJVMState class we may want to get rid of this extra
699 699 // function eventually and do it all there.
700 700
701 701 SafePointNode* GraphKit::clone_map() {
702 702 if (map() == NULL) return NULL;
703 703
704 704 // Clone the memory edge first
705 705 Node* mem = MergeMemNode::make(C, map()->memory());
706 706 gvn().set_type_bottom(mem);
707 707
708 708 SafePointNode *clonemap = (SafePointNode*)map()->clone();
709 709 JVMState* jvms = this->jvms();
710 710 JVMState* clonejvms = jvms->clone_shallow(C);
711 711 clonemap->set_memory(mem);
712 712 clonemap->set_jvms(clonejvms);
713 713 clonejvms->set_map(clonemap);
714 714 record_for_igvn(clonemap);
715 715 gvn().set_type_bottom(clonemap);
716 716 return clonemap;
717 717 }
718 718
719 719
720 720 //-----------------------------set_map_clone-----------------------------------
721 721 void GraphKit::set_map_clone(SafePointNode* m) {
722 722 _map = m;
723 723 _map = clone_map();
724 724 _map->set_next_exception(NULL);
725 725 debug_only(verify_map());
726 726 }
727 727
728 728
729 729 //----------------------------kill_dead_locals---------------------------------
730 730 // Detect any locals which are known to be dead, and force them to top.
731 731 void GraphKit::kill_dead_locals() {
732 732 // Consult the liveness information for the locals. If any
733 733 // of them are unused, then they can be replaced by top(). This
734 734 // should help register allocation time and cut down on the size
735 735 // of the deoptimization information.
736 736
737 737 // This call is made from many of the bytecode handling
738 738 // subroutines called from the Big Switch in do_one_bytecode.
739 739 // Every bytecode which might include a slow path is responsible
740 740 // for killing its dead locals. The more consistent we
741 741 // are about killing deads, the fewer useless phis will be
742 742 // constructed for them at various merge points.
743 743
744 744 // bci can be -1 (InvocationEntryBci). We return the entry
745 745 // liveness for the method.
746 746
747 747 if (method() == NULL || method()->code_size() == 0) {
748 748 // We are building a graph for a call to a native method.
749 749 // All locals are live.
750 750 return;
751 751 }
752 752
753 753 ResourceMark rm;
754 754
755 755 // Consult the liveness information for the locals. If any
756 756 // of them are unused, then they can be replaced by top(). This
757 757 // should help register allocation time and cut down on the size
758 758 // of the deoptimization information.
759 759 MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
760 760
761 761 int len = (int)live_locals.size();
762 762 assert(len <= jvms()->loc_size(), "too many live locals");
763 763 for (int local = 0; local < len; local++) {
764 764 if (!live_locals.at(local)) {
765 765 set_local(local, top());
766 766 }
767 767 }
768 768 }
769 769
770 770 #ifdef ASSERT
771 771 //-------------------------dead_locals_are_killed------------------------------
772 772 // Return true if all dead locals are set to top in the map.
773 773 // Used to assert "clean" debug info at various points.
774 774 bool GraphKit::dead_locals_are_killed() {
775 775 if (method() == NULL || method()->code_size() == 0) {
776 776 // No locals need to be dead, so all is as it should be.
777 777 return true;
778 778 }
779 779
780 780 // Make sure somebody called kill_dead_locals upstream.
781 781 ResourceMark rm;
782 782 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
783 783 if (jvms->loc_size() == 0) continue; // no locals to consult
784 784 SafePointNode* map = jvms->map();
785 785 ciMethod* method = jvms->method();
786 786 int bci = jvms->bci();
787 787 if (jvms == this->jvms()) {
788 788 bci = this->bci(); // it might not yet be synched
789 789 }
790 790 MethodLivenessResult live_locals = method->liveness_at_bci(bci);
791 791 int len = (int)live_locals.size();
792 792 if (!live_locals.is_valid() || len == 0)
793 793 // This method is trivial, or is poisoned by a breakpoint.
794 794 return true;
795 795 assert(len == jvms->loc_size(), "live map consistent with locals map");
796 796 for (int local = 0; local < len; local++) {
797 797 if (!live_locals.at(local) && map->local(jvms, local) != top()) {
798 798 if (PrintMiscellaneous && (Verbose || WizardMode)) {
799 799 tty->print_cr("Zombie local %d: ", local);
800 800 jvms->dump();
801 801 }
802 802 return false;
803 803 }
804 804 }
805 805 }
806 806 return true;
807 807 }
808 808
809 809 #endif //ASSERT
810 810
811 811 // Helper function for enforcing certain bytecodes to reexecute if
812 812 // deoptimization happens
813 813 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
814 814 ciMethod* cur_method = jvms->method();
815 815 int cur_bci = jvms->bci();
816 816 if (cur_method != NULL && cur_bci != InvocationEntryBci) {
817 817 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
818 818 return Interpreter::bytecode_should_reexecute(code) ||
819 819 is_anewarray && code == Bytecodes::_multianewarray;
820 820 // Reexecute _multianewarray bytecode which was replaced with
821 821 // sequence of [a]newarray. See Parse::do_multianewarray().
822 822 //
823 823 // Note: interpreter should not have it set since this optimization
824 824 // is limited by dimensions and guarded by flag so in some cases
825 825 // multianewarray() runtime calls will be generated and
826 826 // the bytecode should not be reexecutes (stack will not be reset).
827 827 } else
828 828 return false;
829 829 }
830 830
831 831 // Helper function for adding JVMState and debug information to node
832 832 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
833 833 // Add the safepoint edges to the call (or other safepoint).
834 834
835 835 // Make sure dead locals are set to top. This
836 836 // should help register allocation time and cut down on the size
837 837 // of the deoptimization information.
838 838 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
839 839
840 840 // Walk the inline list to fill in the correct set of JVMState's
841 841 // Also fill in the associated edges for each JVMState.
842 842
843 843 // If the bytecode needs to be reexecuted we need to put
844 844 // the arguments back on the stack.
845 845 const bool should_reexecute = jvms()->should_reexecute();
846 846 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
847 847
848 848 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
849 849 // undefined if the bci is different. This is normal for Parse but it
850 850 // should not happen for LibraryCallKit because only one bci is processed.
851 851 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
852 852 "in LibraryCallKit the reexecute bit should not change");
853 853
854 854 // If we are guaranteed to throw, we can prune everything but the
855 855 // input to the current bytecode.
856 856 bool can_prune_locals = false;
857 857 uint stack_slots_not_pruned = 0;
858 858 int inputs = 0, depth = 0;
859 859 if (must_throw) {
860 860 assert(method() == youngest_jvms->method(), "sanity");
861 861 if (compute_stack_effects(inputs, depth)) {
862 862 can_prune_locals = true;
863 863 stack_slots_not_pruned = inputs;
864 864 }
865 865 }
866 866
867 867 if (env()->jvmti_can_access_local_variables()) {
868 868 // At any safepoint, this method can get breakpointed, which would
869 869 // then require an immediate deoptimization.
870 870 can_prune_locals = false; // do not prune locals
871 871 stack_slots_not_pruned = 0;
872 872 }
873 873
874 874 // do not scribble on the input jvms
875 875 JVMState* out_jvms = youngest_jvms->clone_deep(C);
876 876 call->set_jvms(out_jvms); // Start jvms list for call node
877 877
878 878 // For a known set of bytecodes, the interpreter should reexecute them if
879 879 // deoptimization happens. We set the reexecute state for them here
880 880 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
881 881 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
882 882 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
883 883 }
884 884
885 885 // Presize the call:
886 886 DEBUG_ONLY(uint non_debug_edges = call->req());
887 887 call->add_req_batch(top(), youngest_jvms->debug_depth());
888 888 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
889 889
890 890 // Set up edges so that the call looks like this:
891 891 // Call [state:] ctl io mem fptr retadr
892 892 // [parms:] parm0 ... parmN
893 893 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
894 894 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
895 895 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
896 896 // Note that caller debug info precedes callee debug info.
897 897
898 898 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
899 899 uint debug_ptr = call->req();
900 900
901 901 // Loop over the map input edges associated with jvms, add them
902 902 // to the call node, & reset all offsets to match call node array.
903 903 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
904 904 uint debug_end = debug_ptr;
905 905 uint debug_start = debug_ptr - in_jvms->debug_size();
906 906 debug_ptr = debug_start; // back up the ptr
907 907
908 908 uint p = debug_start; // walks forward in [debug_start, debug_end)
909 909 uint j, k, l;
910 910 SafePointNode* in_map = in_jvms->map();
911 911 out_jvms->set_map(call);
912 912
913 913 if (can_prune_locals) {
914 914 assert(in_jvms->method() == out_jvms->method(), "sanity");
915 915 // If the current throw can reach an exception handler in this JVMS,
916 916 // then we must keep everything live that can reach that handler.
917 917 // As a quick and dirty approximation, we look for any handlers at all.
918 918 if (in_jvms->method()->has_exception_handlers()) {
919 919 can_prune_locals = false;
920 920 }
921 921 }
922 922
923 923 // Add the Locals
924 924 k = in_jvms->locoff();
925 925 l = in_jvms->loc_size();
926 926 out_jvms->set_locoff(p);
927 927 if (!can_prune_locals) {
928 928 for (j = 0; j < l; j++)
929 929 call->set_req(p++, in_map->in(k+j));
930 930 } else {
931 931 p += l; // already set to top above by add_req_batch
932 932 }
933 933
934 934 // Add the Expression Stack
935 935 k = in_jvms->stkoff();
936 936 l = in_jvms->sp();
937 937 out_jvms->set_stkoff(p);
938 938 if (!can_prune_locals) {
939 939 for (j = 0; j < l; j++)
940 940 call->set_req(p++, in_map->in(k+j));
941 941 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
942 942 // Divide stack into {S0,...,S1}, where S0 is set to top.
943 943 uint s1 = stack_slots_not_pruned;
944 944 stack_slots_not_pruned = 0; // for next iteration
945 945 if (s1 > l) s1 = l;
946 946 uint s0 = l - s1;
947 947 p += s0; // skip the tops preinstalled by add_req_batch
948 948 for (j = s0; j < l; j++)
949 949 call->set_req(p++, in_map->in(k+j));
950 950 } else {
951 951 p += l; // already set to top above by add_req_batch
952 952 }
953 953
954 954 // Add the Monitors
955 955 k = in_jvms->monoff();
956 956 l = in_jvms->mon_size();
957 957 out_jvms->set_monoff(p);
958 958 for (j = 0; j < l; j++)
959 959 call->set_req(p++, in_map->in(k+j));
960 960
961 961 // Copy any scalar object fields.
962 962 k = in_jvms->scloff();
963 963 l = in_jvms->scl_size();
964 964 out_jvms->set_scloff(p);
965 965 for (j = 0; j < l; j++)
966 966 call->set_req(p++, in_map->in(k+j));
967 967
968 968 // Finish the new jvms.
969 969 out_jvms->set_endoff(p);
970 970
971 971 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
972 972 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
973 973 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
974 974 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
975 975 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
976 976 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
977 977
978 978 // Update the two tail pointers in parallel.
979 979 out_jvms = out_jvms->caller();
980 980 in_jvms = in_jvms->caller();
981 981 }
982 982
983 983 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
984 984
985 985 // Test the correctness of JVMState::debug_xxx accessors:
986 986 assert(call->jvms()->debug_start() == non_debug_edges, "");
987 987 assert(call->jvms()->debug_end() == call->req(), "");
988 988 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
989 989 }
990 990
991 991 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
992 992 Bytecodes::Code code = java_bc();
993 993 if (code == Bytecodes::_wide) {
994 994 code = method()->java_code_at_bci(bci() + 1);
995 995 }
996 996
997 997 BasicType rtype = T_ILLEGAL;
998 998 int rsize = 0;
999 999
1000 1000 if (code != Bytecodes::_illegal) {
1001 1001 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1002 1002 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1003 1003 if (rtype < T_CONFLICT)
1004 1004 rsize = type2size[rtype];
1005 1005 }
1006 1006
1007 1007 switch (code) {
1008 1008 case Bytecodes::_illegal:
1009 1009 return false;
1010 1010
1011 1011 case Bytecodes::_ldc:
1012 1012 case Bytecodes::_ldc_w:
1013 1013 case Bytecodes::_ldc2_w:
1014 1014 inputs = 0;
1015 1015 break;
1016 1016
1017 1017 case Bytecodes::_dup: inputs = 1; break;
1018 1018 case Bytecodes::_dup_x1: inputs = 2; break;
1019 1019 case Bytecodes::_dup_x2: inputs = 3; break;
1020 1020 case Bytecodes::_dup2: inputs = 2; break;
1021 1021 case Bytecodes::_dup2_x1: inputs = 3; break;
1022 1022 case Bytecodes::_dup2_x2: inputs = 4; break;
1023 1023 case Bytecodes::_swap: inputs = 2; break;
1024 1024 case Bytecodes::_arraylength: inputs = 1; break;
1025 1025
1026 1026 case Bytecodes::_getstatic:
1027 1027 case Bytecodes::_putstatic:
1028 1028 case Bytecodes::_getfield:
1029 1029 case Bytecodes::_putfield:
1030 1030 {
1031 1031 bool ignored_will_link;
1032 1032 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1033 1033 int size = field->type()->size();
1034 1034 bool is_get = (depth >= 0), is_static = (depth & 1);
1035 1035 inputs = (is_static ? 0 : 1);
1036 1036 if (is_get) {
1037 1037 depth = size - inputs;
1038 1038 } else {
1039 1039 inputs += size; // putxxx pops the value from the stack
1040 1040 depth = - inputs;
1041 1041 }
1042 1042 }
1043 1043 break;
1044 1044
1045 1045 case Bytecodes::_invokevirtual:
1046 1046 case Bytecodes::_invokespecial:
1047 1047 case Bytecodes::_invokestatic:
1048 1048 case Bytecodes::_invokedynamic:
1049 1049 case Bytecodes::_invokeinterface:
1050 1050 {
1051 1051 bool ignored_will_link;
1052 1052 ciSignature* declared_signature = NULL;
1053 1053 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1054 1054 assert(declared_signature != NULL, "cannot be null");
1055 1055 inputs = declared_signature->arg_size_for_bc(code);
1056 1056 int size = declared_signature->return_type()->size();
1057 1057 depth = size - inputs;
1058 1058 }
1059 1059 break;
1060 1060
1061 1061 case Bytecodes::_multianewarray:
1062 1062 {
1063 1063 ciBytecodeStream iter(method());
1064 1064 iter.reset_to_bci(bci());
1065 1065 iter.next();
1066 1066 inputs = iter.get_dimensions();
1067 1067 assert(rsize == 1, "");
1068 1068 depth = rsize - inputs;
1069 1069 }
1070 1070 break;
1071 1071
1072 1072 case Bytecodes::_ireturn:
1073 1073 case Bytecodes::_lreturn:
1074 1074 case Bytecodes::_freturn:
1075 1075 case Bytecodes::_dreturn:
1076 1076 case Bytecodes::_areturn:
1077 1077 assert(rsize = -depth, "");
1078 1078 inputs = rsize;
1079 1079 break;
1080 1080
1081 1081 case Bytecodes::_jsr:
1082 1082 case Bytecodes::_jsr_w:
1083 1083 inputs = 0;
1084 1084 depth = 1; // S.B. depth=1, not zero
1085 1085 break;
1086 1086
1087 1087 default:
1088 1088 // bytecode produces a typed result
1089 1089 inputs = rsize - depth;
1090 1090 assert(inputs >= 0, "");
1091 1091 break;
1092 1092 }
1093 1093
1094 1094 #ifdef ASSERT
1095 1095 // spot check
1096 1096 int outputs = depth + inputs;
1097 1097 assert(outputs >= 0, "sanity");
1098 1098 switch (code) {
1099 1099 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1100 1100 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1101 1101 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1102 1102 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1103 1103 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1104 1104 }
1105 1105 #endif //ASSERT
1106 1106
1107 1107 return true;
1108 1108 }
1109 1109
1110 1110
1111 1111
1112 1112 //------------------------------basic_plus_adr---------------------------------
1113 1113 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1114 1114 // short-circuit a common case
1115 1115 if (offset == intcon(0)) return ptr;
1116 1116 return _gvn.transform( new (C) AddPNode(base, ptr, offset) );
1117 1117 }
1118 1118
1119 1119 Node* GraphKit::ConvI2L(Node* offset) {
1120 1120 // short-circuit a common case
1121 1121 jint offset_con = find_int_con(offset, Type::OffsetBot);
1122 1122 if (offset_con != Type::OffsetBot) {
1123 1123 return longcon((jlong) offset_con);
1124 1124 }
1125 1125 return _gvn.transform( new (C) ConvI2LNode(offset));
1126 1126 }
1127 1127
1128 1128 Node* GraphKit::ConvI2UL(Node* offset) {
1129 1129 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1130 1130 if (offset_con != (juint) Type::OffsetBot) {
1131 1131 return longcon((julong) offset_con);
1132 1132 }
1133 1133 Node* conv = _gvn.transform( new (C) ConvI2LNode(offset));
1134 1134 Node* mask = _gvn.transform( ConLNode::make(C, (julong) max_juint) );
1135 1135 return _gvn.transform( new (C) AndLNode(conv, mask) );
1136 1136 }
1137 1137
1138 1138 Node* GraphKit::ConvL2I(Node* offset) {
1139 1139 // short-circuit a common case
1140 1140 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1141 1141 if (offset_con != (jlong)Type::OffsetBot) {
1142 1142 return intcon((int) offset_con);
↓ open down ↓ |
1142 lines elided |
↑ open up ↑ |
1143 1143 }
1144 1144 return _gvn.transform( new (C) ConvL2INode(offset));
1145 1145 }
1146 1146
1147 1147 //-------------------------load_object_klass-----------------------------------
1148 1148 Node* GraphKit::load_object_klass(Node* obj) {
1149 1149 // Special-case a fresh allocation to avoid building nodes:
1150 1150 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1151 1151 if (akls != NULL) return akls;
1152 1152 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1153 - return _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS) );
1153 + return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS));
1154 1154 }
1155 1155
1156 1156 //-------------------------load_array_length-----------------------------------
1157 1157 Node* GraphKit::load_array_length(Node* array) {
1158 1158 // Special-case a fresh allocation to avoid building nodes:
1159 1159 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1160 1160 Node *alen;
1161 1161 if (alloc == NULL) {
1162 1162 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1163 1163 alen = _gvn.transform( new (C) LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1164 1164 } else {
1165 1165 alen = alloc->Ideal_length();
1166 1166 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1167 1167 if (ccast != alen) {
1168 1168 alen = _gvn.transform(ccast);
1169 1169 }
1170 1170 }
1171 1171 return alen;
1172 1172 }
1173 1173
1174 1174 //------------------------------do_null_check----------------------------------
1175 1175 // Helper function to do a NULL pointer check. Returned value is
1176 1176 // the incoming address with NULL casted away. You are allowed to use the
1177 1177 // not-null value only if you are control dependent on the test.
1178 1178 extern int explicit_null_checks_inserted,
1179 1179 explicit_null_checks_elided;
1180 1180 Node* GraphKit::null_check_common(Node* value, BasicType type,
1181 1181 // optional arguments for variations:
1182 1182 bool assert_null,
1183 1183 Node* *null_control) {
1184 1184 assert(!assert_null || null_control == NULL, "not both at once");
1185 1185 if (stopped()) return top();
1186 1186 if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) {
1187 1187 // For some performance testing, we may wish to suppress null checking.
1188 1188 value = cast_not_null(value); // Make it appear to be non-null (4962416).
1189 1189 return value;
1190 1190 }
1191 1191 explicit_null_checks_inserted++;
1192 1192
1193 1193 // Construct NULL check
1194 1194 Node *chk = NULL;
1195 1195 switch(type) {
1196 1196 case T_LONG : chk = new (C) CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1197 1197 case T_INT : chk = new (C) CmpINode(value, _gvn.intcon(0)); break;
1198 1198 case T_ARRAY : // fall through
1199 1199 type = T_OBJECT; // simplify further tests
1200 1200 case T_OBJECT : {
1201 1201 const Type *t = _gvn.type( value );
1202 1202
1203 1203 const TypeOopPtr* tp = t->isa_oopptr();
1204 1204 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1205 1205 // Only for do_null_check, not any of its siblings:
1206 1206 && !assert_null && null_control == NULL) {
1207 1207 // Usually, any field access or invocation on an unloaded oop type
1208 1208 // will simply fail to link, since the statically linked class is
1209 1209 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1210 1210 // the static class is loaded but the sharper oop type is not.
1211 1211 // Rather than checking for this obscure case in lots of places,
1212 1212 // we simply observe that a null check on an unloaded class
1213 1213 // will always be followed by a nonsense operation, so we
1214 1214 // can just issue the uncommon trap here.
1215 1215 // Our access to the unloaded class will only be correct
1216 1216 // after it has been loaded and initialized, which requires
1217 1217 // a trip through the interpreter.
1218 1218 #ifndef PRODUCT
1219 1219 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1220 1220 #endif
1221 1221 uncommon_trap(Deoptimization::Reason_unloaded,
1222 1222 Deoptimization::Action_reinterpret,
1223 1223 tp->klass(), "!loaded");
1224 1224 return top();
1225 1225 }
1226 1226
1227 1227 if (assert_null) {
1228 1228 // See if the type is contained in NULL_PTR.
1229 1229 // If so, then the value is already null.
1230 1230 if (t->higher_equal(TypePtr::NULL_PTR)) {
1231 1231 explicit_null_checks_elided++;
1232 1232 return value; // Elided null assert quickly!
1233 1233 }
1234 1234 } else {
1235 1235 // See if mixing in the NULL pointer changes type.
1236 1236 // If so, then the NULL pointer was not allowed in the original
1237 1237 // type. In other words, "value" was not-null.
1238 1238 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1239 1239 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1240 1240 explicit_null_checks_elided++;
1241 1241 return value; // Elided null check quickly!
1242 1242 }
1243 1243 }
1244 1244 chk = new (C) CmpPNode( value, null() );
1245 1245 break;
1246 1246 }
1247 1247
1248 1248 default:
1249 1249 fatal(err_msg_res("unexpected type: %s", type2name(type)));
1250 1250 }
1251 1251 assert(chk != NULL, "sanity check");
1252 1252 chk = _gvn.transform(chk);
1253 1253
1254 1254 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1255 1255 BoolNode *btst = new (C) BoolNode( chk, btest);
1256 1256 Node *tst = _gvn.transform( btst );
1257 1257
1258 1258 //-----------
1259 1259 // if peephole optimizations occurred, a prior test existed.
1260 1260 // If a prior test existed, maybe it dominates as we can avoid this test.
1261 1261 if (tst != btst && type == T_OBJECT) {
1262 1262 // At this point we want to scan up the CFG to see if we can
1263 1263 // find an identical test (and so avoid this test altogether).
1264 1264 Node *cfg = control();
1265 1265 int depth = 0;
1266 1266 while( depth < 16 ) { // Limit search depth for speed
1267 1267 if( cfg->Opcode() == Op_IfTrue &&
1268 1268 cfg->in(0)->in(1) == tst ) {
1269 1269 // Found prior test. Use "cast_not_null" to construct an identical
1270 1270 // CastPP (and hence hash to) as already exists for the prior test.
1271 1271 // Return that casted value.
1272 1272 if (assert_null) {
1273 1273 replace_in_map(value, null());
1274 1274 return null(); // do not issue the redundant test
1275 1275 }
1276 1276 Node *oldcontrol = control();
1277 1277 set_control(cfg);
1278 1278 Node *res = cast_not_null(value);
1279 1279 set_control(oldcontrol);
1280 1280 explicit_null_checks_elided++;
1281 1281 return res;
1282 1282 }
1283 1283 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1284 1284 if (cfg == NULL) break; // Quit at region nodes
1285 1285 depth++;
1286 1286 }
1287 1287 }
1288 1288
1289 1289 //-----------
1290 1290 // Branch to failure if null
1291 1291 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1292 1292 Deoptimization::DeoptReason reason;
1293 1293 if (assert_null)
1294 1294 reason = Deoptimization::Reason_null_assert;
1295 1295 else if (type == T_OBJECT)
1296 1296 reason = Deoptimization::Reason_null_check;
1297 1297 else
1298 1298 reason = Deoptimization::Reason_div0_check;
1299 1299
1300 1300 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1301 1301 // ciMethodData::has_trap_at will return a conservative -1 if any
1302 1302 // must-be-null assertion has failed. This could cause performance
1303 1303 // problems for a method after its first do_null_assert failure.
1304 1304 // Consider using 'Reason_class_check' instead?
1305 1305
1306 1306 // To cause an implicit null check, we set the not-null probability
1307 1307 // to the maximum (PROB_MAX). For an explicit check the probability
1308 1308 // is set to a smaller value.
1309 1309 if (null_control != NULL || too_many_traps(reason)) {
1310 1310 // probability is less likely
1311 1311 ok_prob = PROB_LIKELY_MAG(3);
1312 1312 } else if (!assert_null &&
1313 1313 (ImplicitNullCheckThreshold > 0) &&
1314 1314 method() != NULL &&
1315 1315 (method()->method_data()->trap_count(reason)
1316 1316 >= (uint)ImplicitNullCheckThreshold)) {
1317 1317 ok_prob = PROB_LIKELY_MAG(3);
1318 1318 }
1319 1319
1320 1320 if (null_control != NULL) {
1321 1321 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1322 1322 Node* null_true = _gvn.transform( new (C) IfFalseNode(iff));
1323 1323 set_control( _gvn.transform( new (C) IfTrueNode(iff)));
1324 1324 if (null_true == top())
1325 1325 explicit_null_checks_elided++;
1326 1326 (*null_control) = null_true;
1327 1327 } else {
1328 1328 BuildCutout unless(this, tst, ok_prob);
1329 1329 // Check for optimizer eliding test at parse time
1330 1330 if (stopped()) {
1331 1331 // Failure not possible; do not bother making uncommon trap.
1332 1332 explicit_null_checks_elided++;
1333 1333 } else if (assert_null) {
1334 1334 uncommon_trap(reason,
1335 1335 Deoptimization::Action_make_not_entrant,
1336 1336 NULL, "assert_null");
1337 1337 } else {
1338 1338 replace_in_map(value, zerocon(type));
1339 1339 builtin_throw(reason);
1340 1340 }
1341 1341 }
1342 1342
1343 1343 // Must throw exception, fall-thru not possible?
1344 1344 if (stopped()) {
1345 1345 return top(); // No result
1346 1346 }
1347 1347
1348 1348 if (assert_null) {
1349 1349 // Cast obj to null on this path.
1350 1350 replace_in_map(value, zerocon(type));
1351 1351 return zerocon(type);
1352 1352 }
1353 1353
1354 1354 // Cast obj to not-null on this path, if there is no null_control.
1355 1355 // (If there is a null_control, a non-null value may come back to haunt us.)
1356 1356 if (type == T_OBJECT) {
1357 1357 Node* cast = cast_not_null(value, false);
1358 1358 if (null_control == NULL || (*null_control) == top())
1359 1359 replace_in_map(value, cast);
1360 1360 value = cast;
1361 1361 }
1362 1362
1363 1363 return value;
1364 1364 }
1365 1365
1366 1366
1367 1367 //------------------------------cast_not_null----------------------------------
1368 1368 // Cast obj to not-null on this path
1369 1369 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1370 1370 const Type *t = _gvn.type(obj);
1371 1371 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1372 1372 // Object is already not-null?
1373 1373 if( t == t_not_null ) return obj;
1374 1374
1375 1375 Node *cast = new (C) CastPPNode(obj,t_not_null);
1376 1376 cast->init_req(0, control());
1377 1377 cast = _gvn.transform( cast );
1378 1378
1379 1379 // Scan for instances of 'obj' in the current JVM mapping.
1380 1380 // These instances are known to be not-null after the test.
1381 1381 if (do_replace_in_map)
1382 1382 replace_in_map(obj, cast);
1383 1383
1384 1384 return cast; // Return casted value
1385 1385 }
1386 1386
1387 1387
1388 1388 //--------------------------replace_in_map-------------------------------------
1389 1389 void GraphKit::replace_in_map(Node* old, Node* neww) {
1390 1390 if (old == neww) {
1391 1391 return;
1392 1392 }
1393 1393
1394 1394 map()->replace_edge(old, neww);
1395 1395
1396 1396 // Note: This operation potentially replaces any edge
1397 1397 // on the map. This includes locals, stack, and monitors
1398 1398 // of the current (innermost) JVM state.
1399 1399
1400 1400 // don't let inconsistent types from profiling escape this
1401 1401 // method
1402 1402
1403 1403 const Type* told = _gvn.type(old);
1404 1404 const Type* tnew = _gvn.type(neww);
1405 1405
1406 1406 if (!tnew->higher_equal(told)) {
1407 1407 return;
1408 1408 }
1409 1409
1410 1410 map()->record_replaced_node(old, neww);
1411 1411 }
1412 1412
1413 1413
1414 1414 //=============================================================================
1415 1415 //--------------------------------memory---------------------------------------
1416 1416 Node* GraphKit::memory(uint alias_idx) {
1417 1417 MergeMemNode* mem = merged_memory();
1418 1418 Node* p = mem->memory_at(alias_idx);
1419 1419 _gvn.set_type(p, Type::MEMORY); // must be mapped
1420 1420 return p;
1421 1421 }
1422 1422
1423 1423 //-----------------------------reset_memory------------------------------------
1424 1424 Node* GraphKit::reset_memory() {
1425 1425 Node* mem = map()->memory();
1426 1426 // do not use this node for any more parsing!
1427 1427 debug_only( map()->set_memory((Node*)NULL) );
1428 1428 return _gvn.transform( mem );
1429 1429 }
1430 1430
1431 1431 //------------------------------set_all_memory---------------------------------
1432 1432 void GraphKit::set_all_memory(Node* newmem) {
1433 1433 Node* mergemem = MergeMemNode::make(C, newmem);
1434 1434 gvn().set_type_bottom(mergemem);
1435 1435 map()->set_memory(mergemem);
1436 1436 }
1437 1437
1438 1438 //------------------------------set_all_memory_call----------------------------
1439 1439 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1440 1440 Node* newmem = _gvn.transform( new (C) ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1441 1441 set_all_memory(newmem);
1442 1442 }
1443 1443
1444 1444 //=============================================================================
1445 1445 //
1446 1446 // parser factory methods for MemNodes
1447 1447 //
1448 1448 // These are layered on top of the factory methods in LoadNode and StoreNode,
1449 1449 // and integrate with the parser's memory state and _gvn engine.
1450 1450 //
1451 1451
1452 1452 // factory methods in "int adr_idx"
1453 1453 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1454 1454 int adr_idx,
1455 1455 MemNode::MemOrd mo, bool require_atomic_access) {
1456 1456 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1457 1457 const TypePtr* adr_type = NULL; // debug-mode-only argument
1458 1458 debug_only(adr_type = C->get_adr_type(adr_idx));
1459 1459 Node* mem = memory(adr_idx);
1460 1460 Node* ld;
1461 1461 if (require_atomic_access && bt == T_LONG) {
1462 1462 ld = LoadLNode::make_atomic(C, ctl, mem, adr, adr_type, t, mo);
1463 1463 } else {
1464 1464 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo);
1465 1465 }
1466 1466 ld = _gvn.transform(ld);
1467 1467 if ((bt == T_OBJECT) && C->do_escape_analysis() || C->eliminate_boxing()) {
1468 1468 // Improve graph before escape analysis and boxing elimination.
1469 1469 record_for_igvn(ld);
1470 1470 }
1471 1471 return ld;
1472 1472 }
1473 1473
1474 1474 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1475 1475 int adr_idx,
1476 1476 MemNode::MemOrd mo,
1477 1477 bool require_atomic_access) {
1478 1478 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1479 1479 const TypePtr* adr_type = NULL;
1480 1480 debug_only(adr_type = C->get_adr_type(adr_idx));
1481 1481 Node *mem = memory(adr_idx);
1482 1482 Node* st;
1483 1483 if (require_atomic_access && bt == T_LONG) {
1484 1484 st = StoreLNode::make_atomic(C, ctl, mem, adr, adr_type, val, mo);
1485 1485 } else {
1486 1486 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo);
1487 1487 }
1488 1488 st = _gvn.transform(st);
1489 1489 set_memory(st, adr_idx);
1490 1490 // Back-to-back stores can only remove intermediate store with DU info
1491 1491 // so push on worklist for optimizer.
1492 1492 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1493 1493 record_for_igvn(st);
1494 1494
1495 1495 return st;
1496 1496 }
1497 1497
1498 1498
1499 1499 void GraphKit::pre_barrier(bool do_load,
1500 1500 Node* ctl,
1501 1501 Node* obj,
1502 1502 Node* adr,
1503 1503 uint adr_idx,
1504 1504 Node* val,
1505 1505 const TypeOopPtr* val_type,
1506 1506 Node* pre_val,
1507 1507 BasicType bt) {
1508 1508
1509 1509 BarrierSet* bs = Universe::heap()->barrier_set();
1510 1510 set_control(ctl);
1511 1511 switch (bs->kind()) {
1512 1512 case BarrierSet::G1SATBCT:
1513 1513 case BarrierSet::G1SATBCTLogging:
1514 1514 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt);
1515 1515 break;
1516 1516
1517 1517 case BarrierSet::CardTableModRef:
1518 1518 case BarrierSet::CardTableExtension:
1519 1519 case BarrierSet::ModRef:
1520 1520 break;
1521 1521
1522 1522 case BarrierSet::Other:
1523 1523 default :
1524 1524 ShouldNotReachHere();
1525 1525
1526 1526 }
1527 1527 }
1528 1528
1529 1529 bool GraphKit::can_move_pre_barrier() const {
1530 1530 BarrierSet* bs = Universe::heap()->barrier_set();
1531 1531 switch (bs->kind()) {
1532 1532 case BarrierSet::G1SATBCT:
1533 1533 case BarrierSet::G1SATBCTLogging:
1534 1534 return true; // Can move it if no safepoint
1535 1535
1536 1536 case BarrierSet::CardTableModRef:
1537 1537 case BarrierSet::CardTableExtension:
1538 1538 case BarrierSet::ModRef:
1539 1539 return true; // There is no pre-barrier
1540 1540
1541 1541 case BarrierSet::Other:
1542 1542 default :
1543 1543 ShouldNotReachHere();
1544 1544 }
1545 1545 return false;
1546 1546 }
1547 1547
1548 1548 void GraphKit::post_barrier(Node* ctl,
1549 1549 Node* store,
1550 1550 Node* obj,
1551 1551 Node* adr,
1552 1552 uint adr_idx,
1553 1553 Node* val,
1554 1554 BasicType bt,
1555 1555 bool use_precise) {
1556 1556 BarrierSet* bs = Universe::heap()->barrier_set();
1557 1557 set_control(ctl);
1558 1558 switch (bs->kind()) {
1559 1559 case BarrierSet::G1SATBCT:
1560 1560 case BarrierSet::G1SATBCTLogging:
1561 1561 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise);
1562 1562 break;
1563 1563
1564 1564 case BarrierSet::CardTableModRef:
1565 1565 case BarrierSet::CardTableExtension:
1566 1566 write_barrier_post(store, obj, adr, adr_idx, val, use_precise);
1567 1567 break;
1568 1568
1569 1569 case BarrierSet::ModRef:
1570 1570 break;
1571 1571
1572 1572 case BarrierSet::Other:
1573 1573 default :
1574 1574 ShouldNotReachHere();
1575 1575
1576 1576 }
1577 1577 }
1578 1578
1579 1579 Node* GraphKit::store_oop(Node* ctl,
1580 1580 Node* obj,
1581 1581 Node* adr,
1582 1582 const TypePtr* adr_type,
1583 1583 Node* val,
1584 1584 const TypeOopPtr* val_type,
1585 1585 BasicType bt,
1586 1586 bool use_precise,
1587 1587 MemNode::MemOrd mo) {
1588 1588 // Transformation of a value which could be NULL pointer (CastPP #NULL)
1589 1589 // could be delayed during Parse (for example, in adjust_map_after_if()).
1590 1590 // Execute transformation here to avoid barrier generation in such case.
1591 1591 if (_gvn.type(val) == TypePtr::NULL_PTR)
1592 1592 val = _gvn.makecon(TypePtr::NULL_PTR);
1593 1593
1594 1594 set_control(ctl);
1595 1595 if (stopped()) return top(); // Dead path ?
1596 1596
1597 1597 assert(bt == T_OBJECT, "sanity");
1598 1598 assert(val != NULL, "not dead path");
1599 1599 uint adr_idx = C->get_alias_index(adr_type);
1600 1600 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1601 1601
1602 1602 pre_barrier(true /* do_load */,
1603 1603 control(), obj, adr, adr_idx, val, val_type,
1604 1604 NULL /* pre_val */,
1605 1605 bt);
1606 1606
1607 1607 Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo);
1608 1608 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise);
1609 1609 return store;
1610 1610 }
1611 1611
1612 1612 // Could be an array or object we don't know at compile time (unsafe ref.)
1613 1613 Node* GraphKit::store_oop_to_unknown(Node* ctl,
1614 1614 Node* obj, // containing obj
1615 1615 Node* adr, // actual adress to store val at
1616 1616 const TypePtr* adr_type,
1617 1617 Node* val,
1618 1618 BasicType bt,
1619 1619 MemNode::MemOrd mo) {
1620 1620 Compile::AliasType* at = C->alias_type(adr_type);
1621 1621 const TypeOopPtr* val_type = NULL;
1622 1622 if (adr_type->isa_instptr()) {
1623 1623 if (at->field() != NULL) {
1624 1624 // known field. This code is a copy of the do_put_xxx logic.
1625 1625 ciField* field = at->field();
1626 1626 if (!field->type()->is_loaded()) {
1627 1627 val_type = TypeInstPtr::BOTTOM;
1628 1628 } else {
1629 1629 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass());
1630 1630 }
1631 1631 }
1632 1632 } else if (adr_type->isa_aryptr()) {
1633 1633 val_type = adr_type->is_aryptr()->elem()->make_oopptr();
1634 1634 }
1635 1635 if (val_type == NULL) {
1636 1636 val_type = TypeInstPtr::BOTTOM;
1637 1637 }
1638 1638 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo);
1639 1639 }
1640 1640
1641 1641
1642 1642 //-------------------------array_element_address-------------------------
1643 1643 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1644 1644 const TypeInt* sizetype) {
1645 1645 uint shift = exact_log2(type2aelembytes(elembt));
1646 1646 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1647 1647
1648 1648 // short-circuit a common case (saves lots of confusing waste motion)
1649 1649 jint idx_con = find_int_con(idx, -1);
1650 1650 if (idx_con >= 0) {
1651 1651 intptr_t offset = header + ((intptr_t)idx_con << shift);
1652 1652 return basic_plus_adr(ary, offset);
1653 1653 }
1654 1654
1655 1655 // must be correct type for alignment purposes
1656 1656 Node* base = basic_plus_adr(ary, header);
1657 1657 #ifdef _LP64
1658 1658 // The scaled index operand to AddP must be a clean 64-bit value.
1659 1659 // Java allows a 32-bit int to be incremented to a negative
1660 1660 // value, which appears in a 64-bit register as a large
1661 1661 // positive number. Using that large positive number as an
1662 1662 // operand in pointer arithmetic has bad consequences.
1663 1663 // On the other hand, 32-bit overflow is rare, and the possibility
1664 1664 // can often be excluded, if we annotate the ConvI2L node with
1665 1665 // a type assertion that its value is known to be a small positive
1666 1666 // number. (The prior range check has ensured this.)
1667 1667 // This assertion is used by ConvI2LNode::Ideal.
1668 1668 int index_max = max_jint - 1; // array size is max_jint, index is one less
1669 1669 if (sizetype != NULL) index_max = sizetype->_hi - 1;
1670 1670 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax);
1671 1671 idx = _gvn.transform( new (C) ConvI2LNode(idx, lidxtype) );
1672 1672 #endif
1673 1673 Node* scale = _gvn.transform( new (C) LShiftXNode(idx, intcon(shift)) );
1674 1674 return basic_plus_adr(ary, base, scale);
1675 1675 }
1676 1676
1677 1677 //-------------------------load_array_element-------------------------
1678 1678 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1679 1679 const Type* elemtype = arytype->elem();
1680 1680 BasicType elembt = elemtype->array_element_basic_type();
1681 1681 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1682 1682 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered);
1683 1683 return ld;
1684 1684 }
1685 1685
1686 1686 //-------------------------set_arguments_for_java_call-------------------------
1687 1687 // Arguments (pre-popped from the stack) are taken from the JVMS.
1688 1688 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1689 1689 // Add the call arguments:
1690 1690 uint nargs = call->method()->arg_size();
1691 1691 for (uint i = 0; i < nargs; i++) {
1692 1692 Node* arg = argument(i);
1693 1693 call->init_req(i + TypeFunc::Parms, arg);
1694 1694 }
1695 1695 }
1696 1696
1697 1697 //---------------------------set_edges_for_java_call---------------------------
1698 1698 // Connect a newly created call into the current JVMS.
1699 1699 // A return value node (if any) is returned from set_edges_for_java_call.
1700 1700 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1701 1701
1702 1702 // Add the predefined inputs:
1703 1703 call->init_req( TypeFunc::Control, control() );
1704 1704 call->init_req( TypeFunc::I_O , i_o() );
1705 1705 call->init_req( TypeFunc::Memory , reset_memory() );
1706 1706 call->init_req( TypeFunc::FramePtr, frameptr() );
1707 1707 call->init_req( TypeFunc::ReturnAdr, top() );
1708 1708
1709 1709 add_safepoint_edges(call, must_throw);
1710 1710
1711 1711 Node* xcall = _gvn.transform(call);
1712 1712
1713 1713 if (xcall == top()) {
1714 1714 set_control(top());
1715 1715 return;
1716 1716 }
1717 1717 assert(xcall == call, "call identity is stable");
1718 1718
1719 1719 // Re-use the current map to produce the result.
1720 1720
1721 1721 set_control(_gvn.transform(new (C) ProjNode(call, TypeFunc::Control)));
1722 1722 set_i_o( _gvn.transform(new (C) ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1723 1723 set_all_memory_call(xcall, separate_io_proj);
1724 1724
1725 1725 //return xcall; // no need, caller already has it
1726 1726 }
1727 1727
1728 1728 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) {
1729 1729 if (stopped()) return top(); // maybe the call folded up?
1730 1730
1731 1731 // Capture the return value, if any.
1732 1732 Node* ret;
1733 1733 if (call->method() == NULL ||
1734 1734 call->method()->return_type()->basic_type() == T_VOID)
1735 1735 ret = top();
1736 1736 else ret = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
1737 1737
1738 1738 // Note: Since any out-of-line call can produce an exception,
1739 1739 // we always insert an I_O projection from the call into the result.
1740 1740
1741 1741 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj);
1742 1742
1743 1743 if (separate_io_proj) {
1744 1744 // The caller requested separate projections be used by the fall
1745 1745 // through and exceptional paths, so replace the projections for
1746 1746 // the fall through path.
1747 1747 set_i_o(_gvn.transform( new (C) ProjNode(call, TypeFunc::I_O) ));
1748 1748 set_all_memory(_gvn.transform( new (C) ProjNode(call, TypeFunc::Memory) ));
1749 1749 }
1750 1750 return ret;
1751 1751 }
1752 1752
1753 1753 //--------------------set_predefined_input_for_runtime_call--------------------
1754 1754 // Reading and setting the memory state is way conservative here.
1755 1755 // The real problem is that I am not doing real Type analysis on memory,
1756 1756 // so I cannot distinguish card mark stores from other stores. Across a GC
1757 1757 // point the Store Barrier and the card mark memory has to agree. I cannot
1758 1758 // have a card mark store and its barrier split across the GC point from
1759 1759 // either above or below. Here I get that to happen by reading ALL of memory.
1760 1760 // A better answer would be to separate out card marks from other memory.
1761 1761 // For now, return the input memory state, so that it can be reused
1762 1762 // after the call, if this call has restricted memory effects.
1763 1763 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) {
1764 1764 // Set fixed predefined input arguments
1765 1765 Node* memory = reset_memory();
1766 1766 call->init_req( TypeFunc::Control, control() );
1767 1767 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1768 1768 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs
1769 1769 call->init_req( TypeFunc::FramePtr, frameptr() );
1770 1770 call->init_req( TypeFunc::ReturnAdr, top() );
1771 1771 return memory;
1772 1772 }
1773 1773
1774 1774 //-------------------set_predefined_output_for_runtime_call--------------------
1775 1775 // Set control and memory (not i_o) from the call.
1776 1776 // If keep_mem is not NULL, use it for the output state,
1777 1777 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1778 1778 // If hook_mem is NULL, this call produces no memory effects at all.
1779 1779 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1780 1780 // then only that memory slice is taken from the call.
1781 1781 // In the last case, we must put an appropriate memory barrier before
1782 1782 // the call, so as to create the correct anti-dependencies on loads
1783 1783 // preceding the call.
1784 1784 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1785 1785 Node* keep_mem,
1786 1786 const TypePtr* hook_mem) {
1787 1787 // no i/o
1788 1788 set_control(_gvn.transform( new (C) ProjNode(call,TypeFunc::Control) ));
1789 1789 if (keep_mem) {
1790 1790 // First clone the existing memory state
1791 1791 set_all_memory(keep_mem);
1792 1792 if (hook_mem != NULL) {
1793 1793 // Make memory for the call
1794 1794 Node* mem = _gvn.transform( new (C) ProjNode(call, TypeFunc::Memory) );
1795 1795 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
1796 1796 // We also use hook_mem to extract specific effects from arraycopy stubs.
1797 1797 set_memory(mem, hook_mem);
1798 1798 }
1799 1799 // ...else the call has NO memory effects.
1800 1800
1801 1801 // Make sure the call advertises its memory effects precisely.
1802 1802 // This lets us build accurate anti-dependences in gcm.cpp.
1803 1803 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1804 1804 "call node must be constructed correctly");
1805 1805 } else {
1806 1806 assert(hook_mem == NULL, "");
1807 1807 // This is not a "slow path" call; all memory comes from the call.
1808 1808 set_all_memory_call(call);
1809 1809 }
1810 1810 }
1811 1811
1812 1812
1813 1813 // Replace the call with the current state of the kit.
1814 1814 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) {
1815 1815 JVMState* ejvms = NULL;
1816 1816 if (has_exceptions()) {
1817 1817 ejvms = transfer_exceptions_into_jvms();
1818 1818 }
1819 1819
1820 1820 ReplacedNodes replaced_nodes = map()->replaced_nodes();
1821 1821 ReplacedNodes replaced_nodes_exception;
1822 1822 Node* ex_ctl = top();
1823 1823
1824 1824 SafePointNode* final_state = stop();
1825 1825
1826 1826 // Find all the needed outputs of this call
1827 1827 CallProjections callprojs;
1828 1828 call->extract_projections(&callprojs, true);
1829 1829
1830 1830 Node* init_mem = call->in(TypeFunc::Memory);
1831 1831 Node* final_mem = final_state->in(TypeFunc::Memory);
1832 1832 Node* final_ctl = final_state->in(TypeFunc::Control);
1833 1833 Node* final_io = final_state->in(TypeFunc::I_O);
1834 1834
1835 1835 // Replace all the old call edges with the edges from the inlining result
1836 1836 if (callprojs.fallthrough_catchproj != NULL) {
1837 1837 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
1838 1838 }
1839 1839 if (callprojs.fallthrough_memproj != NULL) {
1840 1840 if (final_mem->is_MergeMem()) {
1841 1841 // Parser's exits MergeMem was not transformed but may be optimized
1842 1842 final_mem = _gvn.transform(final_mem);
1843 1843 }
1844 1844 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem);
1845 1845 }
1846 1846 if (callprojs.fallthrough_ioproj != NULL) {
1847 1847 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io);
1848 1848 }
1849 1849
1850 1850 // Replace the result with the new result if it exists and is used
1851 1851 if (callprojs.resproj != NULL && result != NULL) {
1852 1852 C->gvn_replace_by(callprojs.resproj, result);
1853 1853 }
1854 1854
1855 1855 if (ejvms == NULL) {
1856 1856 // No exception edges to simply kill off those paths
1857 1857 if (callprojs.catchall_catchproj != NULL) {
1858 1858 C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1859 1859 }
1860 1860 if (callprojs.catchall_memproj != NULL) {
1861 1861 C->gvn_replace_by(callprojs.catchall_memproj, C->top());
1862 1862 }
1863 1863 if (callprojs.catchall_ioproj != NULL) {
1864 1864 C->gvn_replace_by(callprojs.catchall_ioproj, C->top());
1865 1865 }
1866 1866 // Replace the old exception object with top
1867 1867 if (callprojs.exobj != NULL) {
1868 1868 C->gvn_replace_by(callprojs.exobj, C->top());
1869 1869 }
1870 1870 } else {
1871 1871 GraphKit ekit(ejvms);
1872 1872
1873 1873 // Load my combined exception state into the kit, with all phis transformed:
1874 1874 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1875 1875 replaced_nodes_exception = ex_map->replaced_nodes();
1876 1876
1877 1877 Node* ex_oop = ekit.use_exception_state(ex_map);
1878 1878
1879 1879 if (callprojs.catchall_catchproj != NULL) {
1880 1880 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
1881 1881 ex_ctl = ekit.control();
1882 1882 }
1883 1883 if (callprojs.catchall_memproj != NULL) {
1884 1884 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory());
1885 1885 }
1886 1886 if (callprojs.catchall_ioproj != NULL) {
1887 1887 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o());
1888 1888 }
1889 1889
1890 1890 // Replace the old exception object with the newly created one
1891 1891 if (callprojs.exobj != NULL) {
1892 1892 C->gvn_replace_by(callprojs.exobj, ex_oop);
1893 1893 }
1894 1894 }
1895 1895
1896 1896 // Disconnect the call from the graph
1897 1897 call->disconnect_inputs(NULL, C);
1898 1898 C->gvn_replace_by(call, C->top());
1899 1899
1900 1900 // Clean up any MergeMems that feed other MergeMems since the
1901 1901 // optimizer doesn't like that.
1902 1902 if (final_mem->is_MergeMem()) {
1903 1903 Node_List wl;
1904 1904 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) {
1905 1905 Node* m = i.get();
1906 1906 if (m->is_MergeMem() && !wl.contains(m)) {
1907 1907 wl.push(m);
1908 1908 }
1909 1909 }
1910 1910 while (wl.size() > 0) {
1911 1911 _gvn.transform(wl.pop());
1912 1912 }
1913 1913 }
1914 1914
1915 1915 if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) {
1916 1916 replaced_nodes.apply(C, final_ctl);
1917 1917 }
1918 1918 if (!ex_ctl->is_top() && do_replaced_nodes) {
1919 1919 replaced_nodes_exception.apply(C, ex_ctl);
1920 1920 }
1921 1921 }
1922 1922
1923 1923
1924 1924 //------------------------------increment_counter------------------------------
1925 1925 // for statistics: increment a VM counter by 1
1926 1926
1927 1927 void GraphKit::increment_counter(address counter_addr) {
1928 1928 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
1929 1929 increment_counter(adr1);
1930 1930 }
1931 1931
1932 1932 void GraphKit::increment_counter(Node* counter_addr) {
1933 1933 int adr_type = Compile::AliasIdxRaw;
1934 1934 Node* ctrl = control();
1935 1935 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
1936 1936 Node* incr = _gvn.transform(new (C) AddINode(cnt, _gvn.intcon(1)));
1937 1937 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered);
1938 1938 }
1939 1939
1940 1940
1941 1941 //------------------------------uncommon_trap----------------------------------
1942 1942 // Bail out to the interpreter in mid-method. Implemented by calling the
1943 1943 // uncommon_trap blob. This helper function inserts a runtime call with the
1944 1944 // right debug info.
1945 1945 void GraphKit::uncommon_trap(int trap_request,
1946 1946 ciKlass* klass, const char* comment,
1947 1947 bool must_throw,
1948 1948 bool keep_exact_action) {
1949 1949 if (failing()) stop();
1950 1950 if (stopped()) return; // trap reachable?
1951 1951
1952 1952 // Note: If ProfileTraps is true, and if a deopt. actually
1953 1953 // occurs here, the runtime will make sure an MDO exists. There is
1954 1954 // no need to call method()->ensure_method_data() at this point.
1955 1955
1956 1956 // Set the stack pointer to the right value for reexecution:
1957 1957 set_sp(reexecute_sp());
1958 1958
1959 1959 #ifdef ASSERT
1960 1960 if (!must_throw) {
1961 1961 // Make sure the stack has at least enough depth to execute
1962 1962 // the current bytecode.
1963 1963 int inputs, ignored_depth;
1964 1964 if (compute_stack_effects(inputs, ignored_depth)) {
1965 1965 assert(sp() >= inputs, err_msg_res("must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
1966 1966 Bytecodes::name(java_bc()), sp(), inputs));
1967 1967 }
1968 1968 }
1969 1969 #endif
1970 1970
1971 1971 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
1972 1972 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
1973 1973
1974 1974 switch (action) {
1975 1975 case Deoptimization::Action_maybe_recompile:
1976 1976 case Deoptimization::Action_reinterpret:
1977 1977 // Temporary fix for 6529811 to allow virtual calls to be sure they
1978 1978 // get the chance to go from mono->bi->mega
1979 1979 if (!keep_exact_action &&
1980 1980 Deoptimization::trap_request_index(trap_request) < 0 &&
1981 1981 too_many_recompiles(reason)) {
1982 1982 // This BCI is causing too many recompilations.
1983 1983 action = Deoptimization::Action_none;
1984 1984 trap_request = Deoptimization::make_trap_request(reason, action);
1985 1985 } else {
1986 1986 C->set_trap_can_recompile(true);
1987 1987 }
1988 1988 break;
1989 1989 case Deoptimization::Action_make_not_entrant:
1990 1990 C->set_trap_can_recompile(true);
1991 1991 break;
1992 1992 #ifdef ASSERT
1993 1993 case Deoptimization::Action_none:
1994 1994 case Deoptimization::Action_make_not_compilable:
1995 1995 break;
1996 1996 default:
1997 1997 fatal(err_msg_res("unknown action %d: %s", action, Deoptimization::trap_action_name(action)));
1998 1998 break;
1999 1999 #endif
2000 2000 }
2001 2001
2002 2002 if (TraceOptoParse) {
2003 2003 char buf[100];
2004 2004 tty->print_cr("Uncommon trap %s at bci:%d",
2005 2005 Deoptimization::format_trap_request(buf, sizeof(buf),
2006 2006 trap_request), bci());
2007 2007 }
2008 2008
2009 2009 CompileLog* log = C->log();
2010 2010 if (log != NULL) {
2011 2011 int kid = (klass == NULL)? -1: log->identify(klass);
2012 2012 log->begin_elem("uncommon_trap bci='%d'", bci());
2013 2013 char buf[100];
2014 2014 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
2015 2015 trap_request));
2016 2016 if (kid >= 0) log->print(" klass='%d'", kid);
2017 2017 if (comment != NULL) log->print(" comment='%s'", comment);
2018 2018 log->end_elem();
2019 2019 }
2020 2020
2021 2021 // Make sure any guarding test views this path as very unlikely
2022 2022 Node *i0 = control()->in(0);
2023 2023 if (i0 != NULL && i0->is_If()) { // Found a guarding if test?
2024 2024 IfNode *iff = i0->as_If();
2025 2025 float f = iff->_prob; // Get prob
2026 2026 if (control()->Opcode() == Op_IfTrue) {
2027 2027 if (f > PROB_UNLIKELY_MAG(4))
2028 2028 iff->_prob = PROB_MIN;
2029 2029 } else {
2030 2030 if (f < PROB_LIKELY_MAG(4))
2031 2031 iff->_prob = PROB_MAX;
2032 2032 }
2033 2033 }
2034 2034
2035 2035 // Clear out dead values from the debug info.
2036 2036 kill_dead_locals();
2037 2037
2038 2038 // Now insert the uncommon trap subroutine call
2039 2039 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
2040 2040 const TypePtr* no_memory_effects = NULL;
2041 2041 // Pass the index of the class to be loaded
2042 2042 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2043 2043 (must_throw ? RC_MUST_THROW : 0),
2044 2044 OptoRuntime::uncommon_trap_Type(),
2045 2045 call_addr, "uncommon_trap", no_memory_effects,
2046 2046 intcon(trap_request));
2047 2047 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2048 2048 "must extract request correctly from the graph");
2049 2049 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2050 2050
2051 2051 call->set_req(TypeFunc::ReturnAdr, returnadr());
2052 2052 // The debug info is the only real input to this call.
2053 2053
2054 2054 // Halt-and-catch fire here. The above call should never return!
2055 2055 HaltNode* halt = new(C) HaltNode(control(), frameptr());
2056 2056 _gvn.set_type_bottom(halt);
2057 2057 root()->add_req(halt);
2058 2058
2059 2059 stop_and_kill_map();
2060 2060 }
2061 2061
2062 2062
2063 2063 //--------------------------just_allocated_object------------------------------
2064 2064 // Report the object that was just allocated.
2065 2065 // It must be the case that there are no intervening safepoints.
2066 2066 // We use this to determine if an object is so "fresh" that
2067 2067 // it does not require card marks.
2068 2068 Node* GraphKit::just_allocated_object(Node* current_control) {
2069 2069 if (C->recent_alloc_ctl() == current_control)
2070 2070 return C->recent_alloc_obj();
2071 2071 return NULL;
2072 2072 }
2073 2073
2074 2074
2075 2075 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2076 2076 // (Note: TypeFunc::make has a cache that makes this fast.)
2077 2077 const TypeFunc* tf = TypeFunc::make(dest_method);
2078 2078 int nargs = tf->_domain->_cnt - TypeFunc::Parms;
2079 2079 for (int j = 0; j < nargs; j++) {
2080 2080 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms);
2081 2081 if( targ->basic_type() == T_DOUBLE ) {
2082 2082 // If any parameters are doubles, they must be rounded before
2083 2083 // the call, dstore_rounding does gvn.transform
2084 2084 Node *arg = argument(j);
2085 2085 arg = dstore_rounding(arg);
2086 2086 set_argument(j, arg);
2087 2087 }
2088 2088 }
2089 2089 }
2090 2090
2091 2091 /**
2092 2092 * Record profiling data exact_kls for Node n with the type system so
2093 2093 * that it can propagate it (speculation)
2094 2094 *
2095 2095 * @param n node that the type applies to
2096 2096 * @param exact_kls type from profiling
2097 2097 *
2098 2098 * @return node with improved type
2099 2099 */
2100 2100 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls) {
2101 2101 const Type* current_type = _gvn.type(n);
2102 2102 assert(UseTypeSpeculation, "type speculation must be on");
2103 2103
2104 2104 const TypeOopPtr* speculative = current_type->speculative();
2105 2105
2106 2106 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2107 2107 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls);
2108 2108 const TypeOopPtr* xtype = tklass->as_instance_type();
2109 2109 assert(xtype->klass_is_exact(), "Should be exact");
2110 2110 // record the new speculative type's depth
2111 2111 speculative = xtype->with_inline_depth(jvms()->depth());
2112 2112 }
2113 2113
2114 2114 if (speculative != current_type->speculative()) {
2115 2115 // Build a type with a speculative type (what we think we know
2116 2116 // about the type but will need a guard when we use it)
2117 2117 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative);
2118 2118 // We're changing the type, we need a new CheckCast node to carry
2119 2119 // the new type. The new type depends on the control: what
2120 2120 // profiling tells us is only valid from here as far as we can
2121 2121 // tell.
2122 2122 Node* cast = new(C) CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2123 2123 cast = _gvn.transform(cast);
2124 2124 replace_in_map(n, cast);
2125 2125 n = cast;
2126 2126 }
2127 2127
2128 2128 return n;
2129 2129 }
2130 2130
2131 2131 /**
2132 2132 * Record profiling data from receiver profiling at an invoke with the
2133 2133 * type system so that it can propagate it (speculation)
2134 2134 *
2135 2135 * @param n receiver node
2136 2136 *
2137 2137 * @return node with improved type
2138 2138 */
2139 2139 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2140 2140 if (!UseTypeSpeculation) {
2141 2141 return n;
2142 2142 }
2143 2143 ciKlass* exact_kls = profile_has_unique_klass();
2144 2144 return record_profile_for_speculation(n, exact_kls);
2145 2145 }
2146 2146
2147 2147 /**
2148 2148 * Record profiling data from argument profiling at an invoke with the
2149 2149 * type system so that it can propagate it (speculation)
2150 2150 *
2151 2151 * @param dest_method target method for the call
2152 2152 * @param bc what invoke bytecode is this?
2153 2153 */
2154 2154 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2155 2155 if (!UseTypeSpeculation) {
2156 2156 return;
2157 2157 }
2158 2158 const TypeFunc* tf = TypeFunc::make(dest_method);
2159 2159 int nargs = tf->_domain->_cnt - TypeFunc::Parms;
2160 2160 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2161 2161 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2162 2162 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms);
2163 2163 if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) {
2164 2164 ciKlass* better_type = method()->argument_profiled_type(bci(), i);
2165 2165 if (better_type != NULL) {
2166 2166 record_profile_for_speculation(argument(j), better_type);
2167 2167 }
2168 2168 i++;
2169 2169 }
2170 2170 }
2171 2171 }
2172 2172
2173 2173 /**
2174 2174 * Record profiling data from parameter profiling at an invoke with
2175 2175 * the type system so that it can propagate it (speculation)
2176 2176 */
2177 2177 void GraphKit::record_profiled_parameters_for_speculation() {
2178 2178 if (!UseTypeSpeculation) {
2179 2179 return;
2180 2180 }
2181 2181 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2182 2182 if (_gvn.type(local(i))->isa_oopptr()) {
2183 2183 ciKlass* better_type = method()->parameter_profiled_type(j);
2184 2184 if (better_type != NULL) {
2185 2185 record_profile_for_speculation(local(i), better_type);
2186 2186 }
2187 2187 j++;
2188 2188 }
2189 2189 }
2190 2190 }
2191 2191
2192 2192 void GraphKit::round_double_result(ciMethod* dest_method) {
2193 2193 // A non-strict method may return a double value which has an extended
2194 2194 // exponent, but this must not be visible in a caller which is 'strict'
2195 2195 // If a strict caller invokes a non-strict callee, round a double result
2196 2196
2197 2197 BasicType result_type = dest_method->return_type()->basic_type();
2198 2198 assert( method() != NULL, "must have caller context");
2199 2199 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
2200 2200 // Destination method's return value is on top of stack
2201 2201 // dstore_rounding() does gvn.transform
2202 2202 Node *result = pop_pair();
2203 2203 result = dstore_rounding(result);
2204 2204 push_pair(result);
2205 2205 }
2206 2206 }
2207 2207
2208 2208 // rounding for strict float precision conformance
2209 2209 Node* GraphKit::precision_rounding(Node* n) {
2210 2210 return UseStrictFP && _method->flags().is_strict()
2211 2211 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
2212 2212 ? _gvn.transform( new (C) RoundFloatNode(0, n) )
2213 2213 : n;
2214 2214 }
2215 2215
2216 2216 // rounding for strict double precision conformance
2217 2217 Node* GraphKit::dprecision_rounding(Node *n) {
2218 2218 return UseStrictFP && _method->flags().is_strict()
2219 2219 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
2220 2220 ? _gvn.transform( new (C) RoundDoubleNode(0, n) )
2221 2221 : n;
2222 2222 }
2223 2223
2224 2224 // rounding for non-strict double stores
2225 2225 Node* GraphKit::dstore_rounding(Node* n) {
2226 2226 return Matcher::strict_fp_requires_explicit_rounding
2227 2227 && UseSSE <= 1
2228 2228 ? _gvn.transform( new (C) RoundDoubleNode(0, n) )
2229 2229 : n;
2230 2230 }
2231 2231
2232 2232 //=============================================================================
2233 2233 // Generate a fast path/slow path idiom. Graph looks like:
2234 2234 // [foo] indicates that 'foo' is a parameter
2235 2235 //
2236 2236 // [in] NULL
2237 2237 // \ /
2238 2238 // CmpP
2239 2239 // Bool ne
2240 2240 // If
2241 2241 // / \
2242 2242 // True False-<2>
2243 2243 // / |
2244 2244 // / cast_not_null
2245 2245 // Load | | ^
2246 2246 // [fast_test] | |
2247 2247 // gvn to opt_test | |
2248 2248 // / \ | <1>
2249 2249 // True False |
2250 2250 // | \\ |
2251 2251 // [slow_call] \[fast_result]
2252 2252 // Ctl Val \ \
2253 2253 // | \ \
2254 2254 // Catch <1> \ \
2255 2255 // / \ ^ \ \
2256 2256 // Ex No_Ex | \ \
2257 2257 // | \ \ | \ <2> \
2258 2258 // ... \ [slow_res] | | \ [null_result]
2259 2259 // \ \--+--+--- | |
2260 2260 // \ | / \ | /
2261 2261 // --------Region Phi
2262 2262 //
2263 2263 //=============================================================================
2264 2264 // Code is structured as a series of driver functions all called 'do_XXX' that
2265 2265 // call a set of helper functions. Helper functions first, then drivers.
2266 2266
2267 2267 //------------------------------null_check_oop---------------------------------
2268 2268 // Null check oop. Set null-path control into Region in slot 3.
2269 2269 // Make a cast-not-nullness use the other not-null control. Return cast.
2270 2270 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2271 2271 bool never_see_null, bool safe_for_replace) {
2272 2272 // Initial NULL check taken path
2273 2273 (*null_control) = top();
2274 2274 Node* cast = null_check_common(value, T_OBJECT, false, null_control);
2275 2275
2276 2276 // Generate uncommon_trap:
2277 2277 if (never_see_null && (*null_control) != top()) {
2278 2278 // If we see an unexpected null at a check-cast we record it and force a
2279 2279 // recompile; the offending check-cast will be compiled to handle NULLs.
2280 2280 // If we see more than one offending BCI, then all checkcasts in the
2281 2281 // method will be compiled to handle NULLs.
2282 2282 PreserveJVMState pjvms(this);
2283 2283 set_control(*null_control);
2284 2284 replace_in_map(value, null());
2285 2285 uncommon_trap(Deoptimization::Reason_null_check,
2286 2286 Deoptimization::Action_make_not_entrant);
2287 2287 (*null_control) = top(); // NULL path is dead
2288 2288 }
2289 2289 if ((*null_control) == top() && safe_for_replace) {
2290 2290 replace_in_map(value, cast);
2291 2291 }
2292 2292
2293 2293 // Cast away null-ness on the result
2294 2294 return cast;
2295 2295 }
2296 2296
2297 2297 //------------------------------opt_iff----------------------------------------
2298 2298 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2299 2299 // Return slow-path control.
2300 2300 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2301 2301 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2302 2302
2303 2303 // Fast path taken; set region slot 2
2304 2304 Node *fast_taken = _gvn.transform( new (C) IfFalseNode(opt_iff) );
2305 2305 region->init_req(2,fast_taken); // Capture fast-control
2306 2306
2307 2307 // Fast path not-taken, i.e. slow path
2308 2308 Node *slow_taken = _gvn.transform( new (C) IfTrueNode(opt_iff) );
2309 2309 return slow_taken;
2310 2310 }
2311 2311
2312 2312 //-----------------------------make_runtime_call-------------------------------
2313 2313 Node* GraphKit::make_runtime_call(int flags,
2314 2314 const TypeFunc* call_type, address call_addr,
2315 2315 const char* call_name,
2316 2316 const TypePtr* adr_type,
2317 2317 // The following parms are all optional.
2318 2318 // The first NULL ends the list.
2319 2319 Node* parm0, Node* parm1,
2320 2320 Node* parm2, Node* parm3,
2321 2321 Node* parm4, Node* parm5,
2322 2322 Node* parm6, Node* parm7) {
2323 2323 // Slow-path call
2324 2324 bool is_leaf = !(flags & RC_NO_LEAF);
2325 2325 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2326 2326 if (call_name == NULL) {
2327 2327 assert(!is_leaf, "must supply name for leaf");
2328 2328 call_name = OptoRuntime::stub_name(call_addr);
2329 2329 }
2330 2330 CallNode* call;
2331 2331 if (!is_leaf) {
2332 2332 call = new(C) CallStaticJavaNode(call_type, call_addr, call_name,
2333 2333 bci(), adr_type);
2334 2334 } else if (flags & RC_NO_FP) {
2335 2335 call = new(C) CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2336 2336 } else {
2337 2337 call = new(C) CallLeafNode(call_type, call_addr, call_name, adr_type);
2338 2338 }
2339 2339
2340 2340 // The following is similar to set_edges_for_java_call,
2341 2341 // except that the memory effects of the call are restricted to AliasIdxRaw.
2342 2342
2343 2343 // Slow path call has no side-effects, uses few values
2344 2344 bool wide_in = !(flags & RC_NARROW_MEM);
2345 2345 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2346 2346
2347 2347 Node* prev_mem = NULL;
2348 2348 if (wide_in) {
2349 2349 prev_mem = set_predefined_input_for_runtime_call(call);
2350 2350 } else {
2351 2351 assert(!wide_out, "narrow in => narrow out");
2352 2352 Node* narrow_mem = memory(adr_type);
2353 2353 prev_mem = reset_memory();
2354 2354 map()->set_memory(narrow_mem);
2355 2355 set_predefined_input_for_runtime_call(call);
2356 2356 }
2357 2357
2358 2358 // Hook each parm in order. Stop looking at the first NULL.
2359 2359 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2360 2360 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2361 2361 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2362 2362 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2363 2363 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2364 2364 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2365 2365 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2366 2366 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2367 2367 /* close each nested if ===> */ } } } } } } } }
2368 2368 assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2369 2369
2370 2370 if (!is_leaf) {
2371 2371 // Non-leaves can block and take safepoints:
2372 2372 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2373 2373 }
2374 2374 // Non-leaves can throw exceptions:
2375 2375 if (has_io) {
2376 2376 call->set_req(TypeFunc::I_O, i_o());
2377 2377 }
2378 2378
2379 2379 if (flags & RC_UNCOMMON) {
2380 2380 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2381 2381 // (An "if" probability corresponds roughly to an unconditional count.
2382 2382 // Sort of.)
2383 2383 call->set_cnt(PROB_UNLIKELY_MAG(4));
2384 2384 }
2385 2385
2386 2386 Node* c = _gvn.transform(call);
2387 2387 assert(c == call, "cannot disappear");
2388 2388
2389 2389 if (wide_out) {
2390 2390 // Slow path call has full side-effects.
2391 2391 set_predefined_output_for_runtime_call(call);
2392 2392 } else {
2393 2393 // Slow path call has few side-effects, and/or sets few values.
2394 2394 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2395 2395 }
2396 2396
2397 2397 if (has_io) {
2398 2398 set_i_o(_gvn.transform(new (C) ProjNode(call, TypeFunc::I_O)));
2399 2399 }
2400 2400 return call;
2401 2401
2402 2402 }
2403 2403
2404 2404 //------------------------------merge_memory-----------------------------------
2405 2405 // Merge memory from one path into the current memory state.
2406 2406 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2407 2407 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2408 2408 Node* old_slice = mms.force_memory();
2409 2409 Node* new_slice = mms.memory2();
2410 2410 if (old_slice != new_slice) {
2411 2411 PhiNode* phi;
2412 2412 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2413 2413 if (mms.is_empty()) {
2414 2414 // clone base memory Phi's inputs for this memory slice
2415 2415 assert(old_slice == mms.base_memory(), "sanity");
2416 2416 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C));
2417 2417 _gvn.set_type(phi, Type::MEMORY);
2418 2418 for (uint i = 1; i < phi->req(); i++) {
2419 2419 phi->init_req(i, old_slice->in(i));
2420 2420 }
2421 2421 } else {
2422 2422 phi = old_slice->as_Phi(); // Phi was generated already
2423 2423 }
2424 2424 } else {
2425 2425 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2426 2426 _gvn.set_type(phi, Type::MEMORY);
2427 2427 }
2428 2428 phi->set_req(new_path, new_slice);
2429 2429 mms.set_memory(phi);
2430 2430 }
2431 2431 }
2432 2432 }
2433 2433
2434 2434 //------------------------------make_slow_call_ex------------------------------
2435 2435 // Make the exception handler hookups for the slow call
2436 2436 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) {
2437 2437 if (stopped()) return;
2438 2438
2439 2439 // Make a catch node with just two handlers: fall-through and catch-all
2440 2440 Node* i_o = _gvn.transform( new (C) ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2441 2441 Node* catc = _gvn.transform( new (C) CatchNode(control(), i_o, 2) );
2442 2442 Node* norm = _gvn.transform( new (C) CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2443 2443 Node* excp = _gvn.transform( new (C) CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2444 2444
2445 2445 { PreserveJVMState pjvms(this);
2446 2446 set_control(excp);
2447 2447 set_i_o(i_o);
2448 2448
2449 2449 if (excp != top()) {
2450 2450 if (deoptimize) {
2451 2451 // Deoptimize if an exception is caught. Don't construct exception state in this case.
2452 2452 uncommon_trap(Deoptimization::Reason_unhandled,
2453 2453 Deoptimization::Action_none);
2454 2454 } else {
2455 2455 // Create an exception state also.
2456 2456 // Use an exact type if the caller has specified a specific exception.
2457 2457 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2458 2458 Node* ex_oop = new (C) CreateExNode(ex_type, control(), i_o);
2459 2459 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2460 2460 }
2461 2461 }
2462 2462 }
2463 2463
2464 2464 // Get the no-exception control from the CatchNode.
2465 2465 set_control(norm);
2466 2466 }
2467 2467
2468 2468
2469 2469 //-------------------------------gen_subtype_check-----------------------------
2470 2470 // Generate a subtyping check. Takes as input the subtype and supertype.
2471 2471 // Returns 2 values: sets the default control() to the true path and returns
2472 2472 // the false path. Only reads invariant memory; sets no (visible) memory.
2473 2473 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2474 2474 // but that's not exposed to the optimizer. This call also doesn't take in an
2475 2475 // Object; if you wish to check an Object you need to load the Object's class
2476 2476 // prior to coming here.
2477 2477 Node* GraphKit::gen_subtype_check(Node* subklass, Node* superklass) {
2478 2478 // Fast check for identical types, perhaps identical constants.
2479 2479 // The types can even be identical non-constants, in cases
2480 2480 // involving Array.newInstance, Object.clone, etc.
2481 2481 if (subklass == superklass)
2482 2482 return top(); // false path is dead; no test needed.
2483 2483
2484 2484 if (_gvn.type(superklass)->singleton()) {
2485 2485 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
2486 2486 ciKlass* subk = _gvn.type(subklass)->is_klassptr()->klass();
2487 2487
2488 2488 // In the common case of an exact superklass, try to fold up the
2489 2489 // test before generating code. You may ask, why not just generate
2490 2490 // the code and then let it fold up? The answer is that the generated
2491 2491 // code will necessarily include null checks, which do not always
2492 2492 // completely fold away. If they are also needless, then they turn
2493 2493 // into a performance loss. Example:
2494 2494 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2495 2495 // Here, the type of 'fa' is often exact, so the store check
2496 2496 // of fa[1]=x will fold up, without testing the nullness of x.
2497 2497 switch (static_subtype_check(superk, subk)) {
2498 2498 case SSC_always_false:
2499 2499 {
2500 2500 Node* always_fail = control();
2501 2501 set_control(top());
2502 2502 return always_fail;
2503 2503 }
2504 2504 case SSC_always_true:
2505 2505 return top();
2506 2506 case SSC_easy_test:
2507 2507 {
2508 2508 // Just do a direct pointer compare and be done.
2509 2509 Node* cmp = _gvn.transform( new(C) CmpPNode(subklass, superklass) );
2510 2510 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) );
2511 2511 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
2512 2512 set_control( _gvn.transform( new(C) IfTrueNode (iff) ) );
2513 2513 return _gvn.transform( new(C) IfFalseNode(iff) );
2514 2514 }
2515 2515 case SSC_full_test:
2516 2516 break;
2517 2517 default:
2518 2518 ShouldNotReachHere();
2519 2519 }
2520 2520 }
2521 2521
2522 2522 // %%% Possible further optimization: Even if the superklass is not exact,
2523 2523 // if the subklass is the unique subtype of the superklass, the check
2524 2524 // will always succeed. We could leave a dependency behind to ensure this.
2525 2525
2526 2526 // First load the super-klass's check-offset
2527 2527 Node *p1 = basic_plus_adr( superklass, superklass, in_bytes(Klass::super_check_offset_offset()) );
2528 2528 Node *chk_off = _gvn.transform(new (C) LoadINode(NULL, memory(p1), p1, _gvn.type(p1)->is_ptr(),
2529 2529 TypeInt::INT, MemNode::unordered));
2530 2530 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2531 2531 bool might_be_cache = (find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2532 2532
2533 2533 // Load from the sub-klass's super-class display list, or a 1-word cache of
2534 2534 // the secondary superclass list, or a failing value with a sentinel offset
↓ open down ↓ |
1371 lines elided |
↑ open up ↑ |
2535 2535 // if the super-klass is an interface or exceptionally deep in the Java
2536 2536 // hierarchy and we have to scan the secondary superclass list the hard way.
2537 2537 // Worst-case type is a little odd: NULL is allowed as a result (usually
2538 2538 // klass loads can never produce a NULL).
2539 2539 Node *chk_off_X = ConvI2X(chk_off);
2540 2540 Node *p2 = _gvn.transform( new (C) AddPNode(subklass,subklass,chk_off_X) );
2541 2541 // For some types like interfaces the following loadKlass is from a 1-word
2542 2542 // cache which is mutable so can't use immutable memory. Other
2543 2543 // types load from the super-class display table which is immutable.
2544 2544 Node *kmem = might_be_cache ? memory(p2) : immutable_memory();
2545 - Node *nkls = _gvn.transform( LoadKlassNode::make( _gvn, kmem, p2, _gvn.type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL ) );
2545 + Node* nkls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, kmem, p2, _gvn.type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL));
2546 2546
2547 2547 // Compile speed common case: ARE a subtype and we canNOT fail
2548 2548 if( superklass == nkls )
2549 2549 return top(); // false path is dead; no test needed.
2550 2550
2551 2551 // See if we get an immediate positive hit. Happens roughly 83% of the
2552 2552 // time. Test to see if the value loaded just previously from the subklass
2553 2553 // is exactly the superklass.
2554 2554 Node *cmp1 = _gvn.transform( new (C) CmpPNode( superklass, nkls ) );
2555 2555 Node *bol1 = _gvn.transform( new (C) BoolNode( cmp1, BoolTest::eq ) );
2556 2556 IfNode *iff1 = create_and_xform_if( control(), bol1, PROB_LIKELY(0.83f), COUNT_UNKNOWN );
2557 2557 Node *iftrue1 = _gvn.transform( new (C) IfTrueNode ( iff1 ) );
2558 2558 set_control( _gvn.transform( new (C) IfFalseNode( iff1 ) ) );
2559 2559
2560 2560 // Compile speed common case: Check for being deterministic right now. If
2561 2561 // chk_off is a constant and not equal to cacheoff then we are NOT a
2562 2562 // subklass. In this case we need exactly the 1 test above and we can
2563 2563 // return those results immediately.
2564 2564 if (!might_be_cache) {
2565 2565 Node* not_subtype_ctrl = control();
2566 2566 set_control(iftrue1); // We need exactly the 1 test above
2567 2567 return not_subtype_ctrl;
2568 2568 }
2569 2569
2570 2570 // Gather the various success & failures here
2571 2571 RegionNode *r_ok_subtype = new (C) RegionNode(4);
2572 2572 record_for_igvn(r_ok_subtype);
2573 2573 RegionNode *r_not_subtype = new (C) RegionNode(3);
2574 2574 record_for_igvn(r_not_subtype);
2575 2575
2576 2576 r_ok_subtype->init_req(1, iftrue1);
2577 2577
2578 2578 // Check for immediate negative hit. Happens roughly 11% of the time (which
2579 2579 // is roughly 63% of the remaining cases). Test to see if the loaded
2580 2580 // check-offset points into the subklass display list or the 1-element
2581 2581 // cache. If it points to the display (and NOT the cache) and the display
2582 2582 // missed then it's not a subtype.
2583 2583 Node *cacheoff = _gvn.intcon(cacheoff_con);
2584 2584 Node *cmp2 = _gvn.transform( new (C) CmpINode( chk_off, cacheoff ) );
2585 2585 Node *bol2 = _gvn.transform( new (C) BoolNode( cmp2, BoolTest::ne ) );
2586 2586 IfNode *iff2 = create_and_xform_if( control(), bol2, PROB_LIKELY(0.63f), COUNT_UNKNOWN );
2587 2587 r_not_subtype->init_req(1, _gvn.transform( new (C) IfTrueNode (iff2) ) );
2588 2588 set_control( _gvn.transform( new (C) IfFalseNode(iff2) ) );
2589 2589
2590 2590 // Check for self. Very rare to get here, but it is taken 1/3 the time.
2591 2591 // No performance impact (too rare) but allows sharing of secondary arrays
2592 2592 // which has some footprint reduction.
2593 2593 Node *cmp3 = _gvn.transform( new (C) CmpPNode( subklass, superklass ) );
2594 2594 Node *bol3 = _gvn.transform( new (C) BoolNode( cmp3, BoolTest::eq ) );
2595 2595 IfNode *iff3 = create_and_xform_if( control(), bol3, PROB_LIKELY(0.36f), COUNT_UNKNOWN );
2596 2596 r_ok_subtype->init_req(2, _gvn.transform( new (C) IfTrueNode ( iff3 ) ) );
2597 2597 set_control( _gvn.transform( new (C) IfFalseNode( iff3 ) ) );
2598 2598
2599 2599 // -- Roads not taken here: --
2600 2600 // We could also have chosen to perform the self-check at the beginning
2601 2601 // of this code sequence, as the assembler does. This would not pay off
2602 2602 // the same way, since the optimizer, unlike the assembler, can perform
2603 2603 // static type analysis to fold away many successful self-checks.
2604 2604 // Non-foldable self checks work better here in second position, because
2605 2605 // the initial primary superclass check subsumes a self-check for most
2606 2606 // types. An exception would be a secondary type like array-of-interface,
2607 2607 // which does not appear in its own primary supertype display.
2608 2608 // Finally, we could have chosen to move the self-check into the
2609 2609 // PartialSubtypeCheckNode, and from there out-of-line in a platform
2610 2610 // dependent manner. But it is worthwhile to have the check here,
2611 2611 // where it can be perhaps be optimized. The cost in code space is
2612 2612 // small (register compare, branch).
2613 2613
2614 2614 // Now do a linear scan of the secondary super-klass array. Again, no real
2615 2615 // performance impact (too rare) but it's gotta be done.
2616 2616 // Since the code is rarely used, there is no penalty for moving it
2617 2617 // out of line, and it can only improve I-cache density.
2618 2618 // The decision to inline or out-of-line this final check is platform
2619 2619 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2620 2620 Node* psc = _gvn.transform(
2621 2621 new (C) PartialSubtypeCheckNode(control(), subklass, superklass) );
2622 2622
2623 2623 Node *cmp4 = _gvn.transform( new (C) CmpPNode( psc, null() ) );
2624 2624 Node *bol4 = _gvn.transform( new (C) BoolNode( cmp4, BoolTest::ne ) );
2625 2625 IfNode *iff4 = create_and_xform_if( control(), bol4, PROB_FAIR, COUNT_UNKNOWN );
2626 2626 r_not_subtype->init_req(2, _gvn.transform( new (C) IfTrueNode (iff4) ) );
2627 2627 r_ok_subtype ->init_req(3, _gvn.transform( new (C) IfFalseNode(iff4) ) );
2628 2628
2629 2629 // Return false path; set default control to true path.
2630 2630 set_control( _gvn.transform(r_ok_subtype) );
2631 2631 return _gvn.transform(r_not_subtype);
2632 2632 }
2633 2633
2634 2634 //----------------------------static_subtype_check-----------------------------
2635 2635 // Shortcut important common cases when superklass is exact:
2636 2636 // (0) superklass is java.lang.Object (can occur in reflective code)
2637 2637 // (1) subklass is already limited to a subtype of superklass => always ok
2638 2638 // (2) subklass does not overlap with superklass => always fail
2639 2639 // (3) superklass has NO subtypes and we can check with a simple compare.
2640 2640 int GraphKit::static_subtype_check(ciKlass* superk, ciKlass* subk) {
2641 2641 if (StressReflectiveCode) {
2642 2642 return SSC_full_test; // Let caller generate the general case.
2643 2643 }
2644 2644
2645 2645 if (superk == env()->Object_klass()) {
2646 2646 return SSC_always_true; // (0) this test cannot fail
2647 2647 }
2648 2648
2649 2649 ciType* superelem = superk;
2650 2650 if (superelem->is_array_klass())
2651 2651 superelem = superelem->as_array_klass()->base_element_type();
2652 2652
2653 2653 if (!subk->is_interface()) { // cannot trust static interface types yet
2654 2654 if (subk->is_subtype_of(superk)) {
2655 2655 return SSC_always_true; // (1) false path dead; no dynamic test needed
2656 2656 }
2657 2657 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
2658 2658 !superk->is_subtype_of(subk)) {
2659 2659 return SSC_always_false;
2660 2660 }
2661 2661 }
2662 2662
2663 2663 // If casting to an instance klass, it must have no subtypes
2664 2664 if (superk->is_interface()) {
2665 2665 // Cannot trust interfaces yet.
2666 2666 // %%% S.B. superk->nof_implementors() == 1
2667 2667 } else if (superelem->is_instance_klass()) {
2668 2668 ciInstanceKlass* ik = superelem->as_instance_klass();
2669 2669 if (!ik->has_subklass() && !ik->is_interface()) {
2670 2670 if (!ik->is_final()) {
2671 2671 // Add a dependency if there is a chance of a later subclass.
2672 2672 C->dependencies()->assert_leaf_type(ik);
2673 2673 }
2674 2674 return SSC_easy_test; // (3) caller can do a simple ptr comparison
2675 2675 }
2676 2676 } else {
2677 2677 // A primitive array type has no subtypes.
2678 2678 return SSC_easy_test; // (3) caller can do a simple ptr comparison
2679 2679 }
2680 2680
2681 2681 return SSC_full_test;
2682 2682 }
2683 2683
2684 2684 // Profile-driven exact type check:
2685 2685 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2686 2686 float prob,
2687 2687 Node* *casted_receiver) {
2688 2688 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2689 2689 Node* recv_klass = load_object_klass(receiver);
2690 2690 Node* want_klass = makecon(tklass);
2691 2691 Node* cmp = _gvn.transform( new(C) CmpPNode(recv_klass, want_klass) );
2692 2692 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) );
2693 2693 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2694 2694 set_control( _gvn.transform( new(C) IfTrueNode (iff) ));
2695 2695 Node* fail = _gvn.transform( new(C) IfFalseNode(iff) );
2696 2696
2697 2697 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2698 2698 assert(recv_xtype->klass_is_exact(), "");
2699 2699
2700 2700 // Subsume downstream occurrences of receiver with a cast to
2701 2701 // recv_xtype, since now we know what the type will be.
2702 2702 Node* cast = new(C) CheckCastPPNode(control(), receiver, recv_xtype);
2703 2703 (*casted_receiver) = _gvn.transform(cast);
2704 2704 // (User must make the replace_in_map call.)
2705 2705
2706 2706 return fail;
2707 2707 }
2708 2708
2709 2709
2710 2710 //------------------------------seems_never_null-------------------------------
2711 2711 // Use null_seen information if it is available from the profile.
2712 2712 // If we see an unexpected null at a type check we record it and force a
2713 2713 // recompile; the offending check will be recompiled to handle NULLs.
2714 2714 // If we see several offending BCIs, then all checks in the
2715 2715 // method will be recompiled.
2716 2716 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data) {
2717 2717 if (UncommonNullCast // Cutout for this technique
2718 2718 && obj != null() // And not the -Xcomp stupid case?
2719 2719 && !too_many_traps(Deoptimization::Reason_null_check)
2720 2720 ) {
2721 2721 if (data == NULL)
2722 2722 // Edge case: no mature data. Be optimistic here.
2723 2723 return true;
2724 2724 // If the profile has not seen a null, assume it won't happen.
2725 2725 assert(java_bc() == Bytecodes::_checkcast ||
2726 2726 java_bc() == Bytecodes::_instanceof ||
2727 2727 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
2728 2728 return !data->as_BitData()->null_seen();
2729 2729 }
2730 2730 return false;
2731 2731 }
2732 2732
2733 2733 //------------------------maybe_cast_profiled_receiver-------------------------
2734 2734 // If the profile has seen exactly one type, narrow to exactly that type.
2735 2735 // Subsequent type checks will always fold up.
2736 2736 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
2737 2737 ciKlass* require_klass,
2738 2738 ciKlass* spec_klass,
2739 2739 bool safe_for_replace) {
2740 2740 if (!UseTypeProfile || !TypeProfileCasts) return NULL;
2741 2741
2742 2742 Deoptimization::DeoptReason reason = spec_klass == NULL ? Deoptimization::Reason_class_check : Deoptimization::Reason_speculate_class_check;
2743 2743
2744 2744 // Make sure we haven't already deoptimized from this tactic.
2745 2745 if (too_many_traps(reason))
2746 2746 return NULL;
2747 2747
2748 2748 // (No, this isn't a call, but it's enough like a virtual call
2749 2749 // to use the same ciMethod accessor to get the profile info...)
2750 2750 // If we have a speculative type use it instead of profiling (which
2751 2751 // may not help us)
2752 2752 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass;
2753 2753 if (exact_kls != NULL) {// no cast failures here
2754 2754 if (require_klass == NULL ||
2755 2755 static_subtype_check(require_klass, exact_kls) == SSC_always_true) {
2756 2756 // If we narrow the type to match what the type profile sees or
2757 2757 // the speculative type, we can then remove the rest of the
2758 2758 // cast.
2759 2759 // This is a win, even if the exact_kls is very specific,
2760 2760 // because downstream operations, such as method calls,
2761 2761 // will often benefit from the sharper type.
2762 2762 Node* exact_obj = not_null_obj; // will get updated in place...
2763 2763 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2764 2764 &exact_obj);
2765 2765 { PreserveJVMState pjvms(this);
2766 2766 set_control(slow_ctl);
2767 2767 uncommon_trap(reason,
2768 2768 Deoptimization::Action_maybe_recompile);
2769 2769 }
2770 2770 if (safe_for_replace) {
2771 2771 replace_in_map(not_null_obj, exact_obj);
2772 2772 }
2773 2773 return exact_obj;
2774 2774 }
2775 2775 // assert(ssc == SSC_always_true)... except maybe the profile lied to us.
2776 2776 }
2777 2777
2778 2778 return NULL;
2779 2779 }
2780 2780
2781 2781 /**
2782 2782 * Cast obj to type and emit guard unless we had too many traps here
2783 2783 * already
2784 2784 *
2785 2785 * @param obj node being casted
2786 2786 * @param type type to cast the node to
2787 2787 * @param not_null true if we know node cannot be null
2788 2788 */
2789 2789 Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
2790 2790 ciKlass* type,
2791 2791 bool not_null) {
2792 2792 // type == NULL if profiling tells us this object is always null
2793 2793 if (type != NULL) {
2794 2794 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
2795 2795 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_null_check;
2796 2796 if (!too_many_traps(null_reason) &&
2797 2797 !too_many_traps(class_reason)) {
2798 2798 Node* not_null_obj = NULL;
2799 2799 // not_null is true if we know the object is not null and
2800 2800 // there's no need for a null check
2801 2801 if (!not_null) {
2802 2802 Node* null_ctl = top();
2803 2803 not_null_obj = null_check_oop(obj, &null_ctl, true, true);
2804 2804 assert(null_ctl->is_top(), "no null control here");
2805 2805 } else {
2806 2806 not_null_obj = obj;
2807 2807 }
2808 2808
2809 2809 Node* exact_obj = not_null_obj;
2810 2810 ciKlass* exact_kls = type;
2811 2811 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2812 2812 &exact_obj);
2813 2813 {
2814 2814 PreserveJVMState pjvms(this);
2815 2815 set_control(slow_ctl);
2816 2816 uncommon_trap(class_reason,
2817 2817 Deoptimization::Action_maybe_recompile);
2818 2818 }
2819 2819 replace_in_map(not_null_obj, exact_obj);
2820 2820 obj = exact_obj;
2821 2821 }
2822 2822 } else {
2823 2823 if (!too_many_traps(Deoptimization::Reason_null_assert)) {
2824 2824 Node* exact_obj = null_assert(obj);
2825 2825 replace_in_map(obj, exact_obj);
2826 2826 obj = exact_obj;
2827 2827 }
2828 2828 }
2829 2829 return obj;
2830 2830 }
2831 2831
2832 2832 //-------------------------------gen_instanceof--------------------------------
2833 2833 // Generate an instance-of idiom. Used by both the instance-of bytecode
2834 2834 // and the reflective instance-of call.
2835 2835 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
2836 2836 kill_dead_locals(); // Benefit all the uncommon traps
2837 2837 assert( !stopped(), "dead parse path should be checked in callers" );
2838 2838 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
2839 2839 "must check for not-null not-dead klass in callers");
2840 2840
2841 2841 // Make the merge point
2842 2842 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
2843 2843 RegionNode* region = new(C) RegionNode(PATH_LIMIT);
2844 2844 Node* phi = new(C) PhiNode(region, TypeInt::BOOL);
2845 2845 C->set_has_split_ifs(true); // Has chance for split-if optimization
2846 2846
2847 2847 ciProfileData* data = NULL;
2848 2848 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
2849 2849 data = method()->method_data()->bci_to_data(bci());
2850 2850 }
2851 2851 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
2852 2852 && seems_never_null(obj, data));
2853 2853
2854 2854 // Null check; get casted pointer; set region slot 3
2855 2855 Node* null_ctl = top();
2856 2856 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace);
2857 2857
2858 2858 // If not_null_obj is dead, only null-path is taken
2859 2859 if (stopped()) { // Doing instance-of on a NULL?
2860 2860 set_control(null_ctl);
2861 2861 return intcon(0);
2862 2862 }
2863 2863 region->init_req(_null_path, null_ctl);
2864 2864 phi ->init_req(_null_path, intcon(0)); // Set null path value
2865 2865 if (null_ctl == top()) {
2866 2866 // Do this eagerly, so that pattern matches like is_diamond_phi
2867 2867 // will work even during parsing.
2868 2868 assert(_null_path == PATH_LIMIT-1, "delete last");
2869 2869 region->del_req(_null_path);
2870 2870 phi ->del_req(_null_path);
2871 2871 }
2872 2872
2873 2873 // Do we know the type check always succeed?
2874 2874 bool known_statically = false;
2875 2875 if (_gvn.type(superklass)->singleton()) {
2876 2876 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
2877 2877 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass();
2878 2878 if (subk != NULL && subk->is_loaded()) {
2879 2879 int static_res = static_subtype_check(superk, subk);
2880 2880 known_statically = (static_res == SSC_always_true || static_res == SSC_always_false);
2881 2881 }
2882 2882 }
2883 2883
2884 2884 if (known_statically && UseTypeSpeculation) {
2885 2885 // If we know the type check always succeeds then we don't use the
2886 2886 // profiling data at this bytecode. Don't lose it, feed it to the
2887 2887 // type system as a speculative type.
2888 2888 not_null_obj = record_profiled_receiver_for_speculation(not_null_obj);
2889 2889 } else {
2890 2890 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
2891 2891 // We may not have profiling here or it may not help us. If we
2892 2892 // have a speculative type use it to perform an exact cast.
2893 2893 ciKlass* spec_obj_type = obj_type->speculative_type();
2894 2894 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) {
2895 2895 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace);
2896 2896 if (stopped()) { // Profile disagrees with this path.
2897 2897 set_control(null_ctl); // Null is the only remaining possibility.
2898 2898 return intcon(0);
2899 2899 }
2900 2900 if (cast_obj != NULL) {
2901 2901 not_null_obj = cast_obj;
2902 2902 }
2903 2903 }
2904 2904 }
2905 2905
2906 2906 // Load the object's klass
2907 2907 Node* obj_klass = load_object_klass(not_null_obj);
2908 2908
2909 2909 // Generate the subtype check
2910 2910 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
2911 2911
2912 2912 // Plug in the success path to the general merge in slot 1.
2913 2913 region->init_req(_obj_path, control());
2914 2914 phi ->init_req(_obj_path, intcon(1));
2915 2915
2916 2916 // Plug in the failing path to the general merge in slot 2.
2917 2917 region->init_req(_fail_path, not_subtype_ctrl);
2918 2918 phi ->init_req(_fail_path, intcon(0));
2919 2919
2920 2920 // Return final merged results
2921 2921 set_control( _gvn.transform(region) );
2922 2922 record_for_igvn(region);
2923 2923 return _gvn.transform(phi);
2924 2924 }
2925 2925
2926 2926 //-------------------------------gen_checkcast---------------------------------
2927 2927 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
2928 2928 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
2929 2929 // uncommon-trap paths work. Adjust stack after this call.
2930 2930 // If failure_control is supplied and not null, it is filled in with
2931 2931 // the control edge for the cast failure. Otherwise, an appropriate
2932 2932 // uncommon trap or exception is thrown.
2933 2933 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
2934 2934 Node* *failure_control) {
2935 2935 kill_dead_locals(); // Benefit all the uncommon traps
2936 2936 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
2937 2937 const Type *toop = TypeOopPtr::make_from_klass(tk->klass());
2938 2938
2939 2939 // Fast cutout: Check the case that the cast is vacuously true.
2940 2940 // This detects the common cases where the test will short-circuit
2941 2941 // away completely. We do this before we perform the null check,
2942 2942 // because if the test is going to turn into zero code, we don't
2943 2943 // want a residual null check left around. (Causes a slowdown,
2944 2944 // for example, in some objArray manipulations, such as a[i]=a[j].)
2945 2945 if (tk->singleton()) {
2946 2946 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
2947 2947 if (objtp != NULL && objtp->klass() != NULL) {
2948 2948 switch (static_subtype_check(tk->klass(), objtp->klass())) {
2949 2949 case SSC_always_true:
2950 2950 // If we know the type check always succeed then we don't use
2951 2951 // the profiling data at this bytecode. Don't lose it, feed it
2952 2952 // to the type system as a speculative type.
2953 2953 return record_profiled_receiver_for_speculation(obj);
2954 2954 case SSC_always_false:
2955 2955 // It needs a null check because a null will *pass* the cast check.
2956 2956 // A non-null value will always produce an exception.
2957 2957 return null_assert(obj);
2958 2958 }
2959 2959 }
2960 2960 }
2961 2961
2962 2962 ciProfileData* data = NULL;
2963 2963 bool safe_for_replace = false;
2964 2964 if (failure_control == NULL) { // use MDO in regular case only
2965 2965 assert(java_bc() == Bytecodes::_aastore ||
2966 2966 java_bc() == Bytecodes::_checkcast,
2967 2967 "interpreter profiles type checks only for these BCs");
2968 2968 data = method()->method_data()->bci_to_data(bci());
2969 2969 safe_for_replace = true;
2970 2970 }
2971 2971
2972 2972 // Make the merge point
2973 2973 enum { _obj_path = 1, _null_path, PATH_LIMIT };
2974 2974 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
2975 2975 Node* phi = new (C) PhiNode(region, toop);
2976 2976 C->set_has_split_ifs(true); // Has chance for split-if optimization
2977 2977
2978 2978 // Use null-cast information if it is available
2979 2979 bool never_see_null = ((failure_control == NULL) // regular case only
2980 2980 && seems_never_null(obj, data));
2981 2981
2982 2982 // Null check; get casted pointer; set region slot 3
2983 2983 Node* null_ctl = top();
2984 2984 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace);
2985 2985
2986 2986 // If not_null_obj is dead, only null-path is taken
2987 2987 if (stopped()) { // Doing instance-of on a NULL?
2988 2988 set_control(null_ctl);
2989 2989 return null();
2990 2990 }
2991 2991 region->init_req(_null_path, null_ctl);
2992 2992 phi ->init_req(_null_path, null()); // Set null path value
2993 2993 if (null_ctl == top()) {
2994 2994 // Do this eagerly, so that pattern matches like is_diamond_phi
2995 2995 // will work even during parsing.
2996 2996 assert(_null_path == PATH_LIMIT-1, "delete last");
2997 2997 region->del_req(_null_path);
2998 2998 phi ->del_req(_null_path);
2999 2999 }
3000 3000
3001 3001 Node* cast_obj = NULL;
3002 3002 if (tk->klass_is_exact()) {
3003 3003 // The following optimization tries to statically cast the speculative type of the object
3004 3004 // (for example obtained during profiling) to the type of the superklass and then do a
3005 3005 // dynamic check that the type of the object is what we expect. To work correctly
3006 3006 // for checkcast and aastore the type of superklass should be exact.
3007 3007 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3008 3008 // We may not have profiling here or it may not help us. If we have
3009 3009 // a speculative type use it to perform an exact cast.
3010 3010 ciKlass* spec_obj_type = obj_type->speculative_type();
3011 3011 if (spec_obj_type != NULL ||
3012 3012 (data != NULL &&
3013 3013 // Counter has never been decremented (due to cast failure).
3014 3014 // ...This is a reasonable thing to expect. It is true of
3015 3015 // all casts inserted by javac to implement generic types.
3016 3016 data->as_CounterData()->count() >= 0)) {
3017 3017 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace);
3018 3018 if (cast_obj != NULL) {
3019 3019 if (failure_control != NULL) // failure is now impossible
3020 3020 (*failure_control) = top();
3021 3021 // adjust the type of the phi to the exact klass:
3022 3022 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3023 3023 }
3024 3024 }
3025 3025 }
3026 3026
3027 3027 if (cast_obj == NULL) {
3028 3028 // Load the object's klass
3029 3029 Node* obj_klass = load_object_klass(not_null_obj);
3030 3030
3031 3031 // Generate the subtype check
3032 3032 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
3033 3033
3034 3034 // Plug in success path into the merge
3035 3035 cast_obj = _gvn.transform(new (C) CheckCastPPNode(control(),
3036 3036 not_null_obj, toop));
3037 3037 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3038 3038 if (failure_control == NULL) {
3039 3039 if (not_subtype_ctrl != top()) { // If failure is possible
3040 3040 PreserveJVMState pjvms(this);
3041 3041 set_control(not_subtype_ctrl);
3042 3042 builtin_throw(Deoptimization::Reason_class_check, obj_klass);
3043 3043 }
3044 3044 } else {
3045 3045 (*failure_control) = not_subtype_ctrl;
3046 3046 }
3047 3047 }
3048 3048
3049 3049 region->init_req(_obj_path, control());
3050 3050 phi ->init_req(_obj_path, cast_obj);
3051 3051
3052 3052 // A merge of NULL or Casted-NotNull obj
3053 3053 Node* res = _gvn.transform(phi);
3054 3054
3055 3055 // Note I do NOT always 'replace_in_map(obj,result)' here.
3056 3056 // if( tk->klass()->can_be_primary_super() )
3057 3057 // This means that if I successfully store an Object into an array-of-String
3058 3058 // I 'forget' that the Object is really now known to be a String. I have to
3059 3059 // do this because we don't have true union types for interfaces - if I store
3060 3060 // a Baz into an array-of-Interface and then tell the optimizer it's an
3061 3061 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3062 3062 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3063 3063 // replace_in_map( obj, res );
3064 3064
3065 3065 // Return final merged results
3066 3066 set_control( _gvn.transform(region) );
3067 3067 record_for_igvn(region);
3068 3068 return res;
3069 3069 }
3070 3070
3071 3071 //------------------------------next_monitor-----------------------------------
3072 3072 // What number should be given to the next monitor?
3073 3073 int GraphKit::next_monitor() {
3074 3074 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3075 3075 int next = current + C->sync_stack_slots();
3076 3076 // Keep the toplevel high water mark current:
3077 3077 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3078 3078 return current;
3079 3079 }
3080 3080
3081 3081 //------------------------------insert_mem_bar---------------------------------
3082 3082 // Memory barrier to avoid floating things around
3083 3083 // The membar serves as a pinch point between both control and all memory slices.
3084 3084 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3085 3085 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3086 3086 mb->init_req(TypeFunc::Control, control());
3087 3087 mb->init_req(TypeFunc::Memory, reset_memory());
3088 3088 Node* membar = _gvn.transform(mb);
3089 3089 set_control(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Control)));
3090 3090 set_all_memory_call(membar);
3091 3091 return membar;
3092 3092 }
3093 3093
3094 3094 //-------------------------insert_mem_bar_volatile----------------------------
3095 3095 // Memory barrier to avoid floating things around
3096 3096 // The membar serves as a pinch point between both control and memory(alias_idx).
3097 3097 // If you want to make a pinch point on all memory slices, do not use this
3098 3098 // function (even with AliasIdxBot); use insert_mem_bar() instead.
3099 3099 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
3100 3100 // When Parse::do_put_xxx updates a volatile field, it appends a series
3101 3101 // of MemBarVolatile nodes, one for *each* volatile field alias category.
3102 3102 // The first membar is on the same memory slice as the field store opcode.
3103 3103 // This forces the membar to follow the store. (Bug 6500685 broke this.)
3104 3104 // All the other membars (for other volatile slices, including AliasIdxBot,
3105 3105 // which stands for all unknown volatile slices) are control-dependent
3106 3106 // on the first membar. This prevents later volatile loads or stores
3107 3107 // from sliding up past the just-emitted store.
3108 3108
3109 3109 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
3110 3110 mb->set_req(TypeFunc::Control,control());
3111 3111 if (alias_idx == Compile::AliasIdxBot) {
3112 3112 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
3113 3113 } else {
3114 3114 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
3115 3115 mb->set_req(TypeFunc::Memory, memory(alias_idx));
3116 3116 }
3117 3117 Node* membar = _gvn.transform(mb);
3118 3118 set_control(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Control)));
3119 3119 if (alias_idx == Compile::AliasIdxBot) {
3120 3120 merged_memory()->set_base_memory(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Memory)));
3121 3121 } else {
3122 3122 set_memory(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Memory)),alias_idx);
3123 3123 }
3124 3124 return membar;
3125 3125 }
3126 3126
3127 3127 //------------------------------shared_lock------------------------------------
3128 3128 // Emit locking code.
3129 3129 FastLockNode* GraphKit::shared_lock(Node* obj) {
3130 3130 // bci is either a monitorenter bc or InvocationEntryBci
3131 3131 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3132 3132 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3133 3133
3134 3134 if( !GenerateSynchronizationCode )
3135 3135 return NULL; // Not locking things?
3136 3136 if (stopped()) // Dead monitor?
3137 3137 return NULL;
3138 3138
3139 3139 assert(dead_locals_are_killed(), "should kill locals before sync. point");
3140 3140
3141 3141 // Box the stack location
3142 3142 Node* box = _gvn.transform(new (C) BoxLockNode(next_monitor()));
3143 3143 Node* mem = reset_memory();
3144 3144
3145 3145 FastLockNode * flock = _gvn.transform(new (C) FastLockNode(0, obj, box) )->as_FastLock();
3146 3146 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) {
3147 3147 // Create the counters for this fast lock.
3148 3148 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3149 3149 }
3150 3150
3151 3151 // Create the rtm counters for this fast lock if needed.
3152 3152 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3153 3153
3154 3154 // Add monitor to debug info for the slow path. If we block inside the
3155 3155 // slow path and de-opt, we need the monitor hanging around
3156 3156 map()->push_monitor( flock );
3157 3157
3158 3158 const TypeFunc *tf = LockNode::lock_type();
3159 3159 LockNode *lock = new (C) LockNode(C, tf);
3160 3160
3161 3161 lock->init_req( TypeFunc::Control, control() );
3162 3162 lock->init_req( TypeFunc::Memory , mem );
3163 3163 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3164 3164 lock->init_req( TypeFunc::FramePtr, frameptr() );
3165 3165 lock->init_req( TypeFunc::ReturnAdr, top() );
3166 3166
3167 3167 lock->init_req(TypeFunc::Parms + 0, obj);
3168 3168 lock->init_req(TypeFunc::Parms + 1, box);
3169 3169 lock->init_req(TypeFunc::Parms + 2, flock);
3170 3170 add_safepoint_edges(lock);
3171 3171
3172 3172 lock = _gvn.transform( lock )->as_Lock();
3173 3173
3174 3174 // lock has no side-effects, sets few values
3175 3175 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
3176 3176
3177 3177 insert_mem_bar(Op_MemBarAcquireLock);
3178 3178
3179 3179 // Add this to the worklist so that the lock can be eliminated
3180 3180 record_for_igvn(lock);
3181 3181
3182 3182 #ifndef PRODUCT
3183 3183 if (PrintLockStatistics) {
3184 3184 // Update the counter for this lock. Don't bother using an atomic
3185 3185 // operation since we don't require absolute accuracy.
3186 3186 lock->create_lock_counter(map()->jvms());
3187 3187 increment_counter(lock->counter()->addr());
3188 3188 }
3189 3189 #endif
3190 3190
3191 3191 return flock;
3192 3192 }
3193 3193
3194 3194
3195 3195 //------------------------------shared_unlock----------------------------------
3196 3196 // Emit unlocking code.
3197 3197 void GraphKit::shared_unlock(Node* box, Node* obj) {
3198 3198 // bci is either a monitorenter bc or InvocationEntryBci
3199 3199 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3200 3200 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3201 3201
3202 3202 if( !GenerateSynchronizationCode )
3203 3203 return;
3204 3204 if (stopped()) { // Dead monitor?
3205 3205 map()->pop_monitor(); // Kill monitor from debug info
3206 3206 return;
3207 3207 }
3208 3208
3209 3209 // Memory barrier to avoid floating things down past the locked region
3210 3210 insert_mem_bar(Op_MemBarReleaseLock);
3211 3211
3212 3212 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3213 3213 UnlockNode *unlock = new (C) UnlockNode(C, tf);
3214 3214 uint raw_idx = Compile::AliasIdxRaw;
3215 3215 unlock->init_req( TypeFunc::Control, control() );
3216 3216 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3217 3217 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3218 3218 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3219 3219 unlock->init_req( TypeFunc::ReturnAdr, top() );
3220 3220
3221 3221 unlock->init_req(TypeFunc::Parms + 0, obj);
3222 3222 unlock->init_req(TypeFunc::Parms + 1, box);
3223 3223 unlock = _gvn.transform(unlock)->as_Unlock();
3224 3224
3225 3225 Node* mem = reset_memory();
3226 3226
3227 3227 // unlock has no side-effects, sets few values
3228 3228 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3229 3229
3230 3230 // Kill monitor from debug info
3231 3231 map()->pop_monitor( );
3232 3232 }
3233 3233
3234 3234 //-------------------------------get_layout_helper-----------------------------
3235 3235 // If the given klass is a constant or known to be an array,
3236 3236 // fetch the constant layout helper value into constant_value
3237 3237 // and return (Node*)NULL. Otherwise, load the non-constant
3238 3238 // layout helper value, and return the node which represents it.
3239 3239 // This two-faced routine is useful because allocation sites
3240 3240 // almost always feature constant types.
3241 3241 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3242 3242 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
3243 3243 if (!StressReflectiveCode && inst_klass != NULL) {
3244 3244 ciKlass* klass = inst_klass->klass();
3245 3245 bool xklass = inst_klass->klass_is_exact();
3246 3246 if (xklass || klass->is_array_klass()) {
3247 3247 jint lhelper = klass->layout_helper();
3248 3248 if (lhelper != Klass::_lh_neutral_value) {
3249 3249 constant_value = lhelper;
3250 3250 return (Node*) NULL;
3251 3251 }
3252 3252 }
3253 3253 }
3254 3254 constant_value = Klass::_lh_neutral_value; // put in a known value
3255 3255 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3256 3256 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3257 3257 }
3258 3258
3259 3259 // We just put in an allocate/initialize with a big raw-memory effect.
3260 3260 // Hook selected additional alias categories on the initialization.
3261 3261 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3262 3262 MergeMemNode* init_in_merge,
3263 3263 Node* init_out_raw) {
3264 3264 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3265 3265 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3266 3266
3267 3267 Node* prevmem = kit.memory(alias_idx);
3268 3268 init_in_merge->set_memory_at(alias_idx, prevmem);
3269 3269 kit.set_memory(init_out_raw, alias_idx);
3270 3270 }
3271 3271
3272 3272 //---------------------------set_output_for_allocation-------------------------
3273 3273 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3274 3274 const TypeOopPtr* oop_type,
3275 3275 bool deoptimize_on_exception) {
3276 3276 int rawidx = Compile::AliasIdxRaw;
3277 3277 alloc->set_req( TypeFunc::FramePtr, frameptr() );
3278 3278 add_safepoint_edges(alloc);
3279 3279 Node* allocx = _gvn.transform(alloc);
3280 3280 set_control( _gvn.transform(new (C) ProjNode(allocx, TypeFunc::Control) ) );
3281 3281 // create memory projection for i_o
3282 3282 set_memory ( _gvn.transform( new (C) ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3283 3283 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
3284 3284
3285 3285 // create a memory projection as for the normal control path
3286 3286 Node* malloc = _gvn.transform(new (C) ProjNode(allocx, TypeFunc::Memory));
3287 3287 set_memory(malloc, rawidx);
3288 3288
3289 3289 // a normal slow-call doesn't change i_o, but an allocation does
3290 3290 // we create a separate i_o projection for the normal control path
3291 3291 set_i_o(_gvn.transform( new (C) ProjNode(allocx, TypeFunc::I_O, false) ) );
3292 3292 Node* rawoop = _gvn.transform( new (C) ProjNode(allocx, TypeFunc::Parms) );
3293 3293
3294 3294 // put in an initialization barrier
3295 3295 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3296 3296 rawoop)->as_Initialize();
3297 3297 assert(alloc->initialization() == init, "2-way macro link must work");
3298 3298 assert(init ->allocation() == alloc, "2-way macro link must work");
3299 3299 {
3300 3300 // Extract memory strands which may participate in the new object's
3301 3301 // initialization, and source them from the new InitializeNode.
3302 3302 // This will allow us to observe initializations when they occur,
3303 3303 // and link them properly (as a group) to the InitializeNode.
3304 3304 assert(init->in(InitializeNode::Memory) == malloc, "");
3305 3305 MergeMemNode* minit_in = MergeMemNode::make(C, malloc);
3306 3306 init->set_req(InitializeNode::Memory, minit_in);
3307 3307 record_for_igvn(minit_in); // fold it up later, if possible
3308 3308 Node* minit_out = memory(rawidx);
3309 3309 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3310 3310 if (oop_type->isa_aryptr()) {
3311 3311 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3312 3312 int elemidx = C->get_alias_index(telemref);
3313 3313 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3314 3314 } else if (oop_type->isa_instptr()) {
3315 3315 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
3316 3316 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3317 3317 ciField* field = ik->nonstatic_field_at(i);
3318 3318 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3319 3319 continue; // do not bother to track really large numbers of fields
3320 3320 // Find (or create) the alias category for this field:
3321 3321 int fieldidx = C->alias_type(field)->index();
3322 3322 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3323 3323 }
3324 3324 }
3325 3325 }
3326 3326
3327 3327 // Cast raw oop to the real thing...
3328 3328 Node* javaoop = new (C) CheckCastPPNode(control(), rawoop, oop_type);
3329 3329 javaoop = _gvn.transform(javaoop);
3330 3330 C->set_recent_alloc(control(), javaoop);
3331 3331 assert(just_allocated_object(control()) == javaoop, "just allocated");
3332 3332
3333 3333 #ifdef ASSERT
3334 3334 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3335 3335 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
3336 3336 "Ideal_allocation works");
3337 3337 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
3338 3338 "Ideal_allocation works");
3339 3339 if (alloc->is_AllocateArray()) {
3340 3340 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
3341 3341 "Ideal_allocation works");
3342 3342 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
3343 3343 "Ideal_allocation works");
3344 3344 } else {
3345 3345 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3346 3346 }
3347 3347 }
3348 3348 #endif //ASSERT
3349 3349
3350 3350 return javaoop;
3351 3351 }
3352 3352
3353 3353 //---------------------------new_instance--------------------------------------
3354 3354 // This routine takes a klass_node which may be constant (for a static type)
3355 3355 // or may be non-constant (for reflective code). It will work equally well
3356 3356 // for either, and the graph will fold nicely if the optimizer later reduces
3357 3357 // the type to a constant.
3358 3358 // The optional arguments are for specialized use by intrinsics:
3359 3359 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3360 3360 // - If 'return_size_val', report the the total object size to the caller.
3361 3361 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3362 3362 Node* GraphKit::new_instance(Node* klass_node,
3363 3363 Node* extra_slow_test,
3364 3364 Node* *return_size_val,
3365 3365 bool deoptimize_on_exception) {
3366 3366 // Compute size in doublewords
3367 3367 // The size is always an integral number of doublewords, represented
3368 3368 // as a positive bytewise size stored in the klass's layout_helper.
3369 3369 // The layout_helper also encodes (in a low bit) the need for a slow path.
3370 3370 jint layout_con = Klass::_lh_neutral_value;
3371 3371 Node* layout_val = get_layout_helper(klass_node, layout_con);
3372 3372 int layout_is_con = (layout_val == NULL);
3373 3373
3374 3374 if (extra_slow_test == NULL) extra_slow_test = intcon(0);
3375 3375 // Generate the initial go-slow test. It's either ALWAYS (return a
3376 3376 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3377 3377 // case) a computed value derived from the layout_helper.
3378 3378 Node* initial_slow_test = NULL;
3379 3379 if (layout_is_con) {
3380 3380 assert(!StressReflectiveCode, "stress mode does not use these paths");
3381 3381 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3382 3382 initial_slow_test = must_go_slow? intcon(1): extra_slow_test;
3383 3383
3384 3384 } else { // reflective case
3385 3385 // This reflective path is used by Unsafe.allocateInstance.
3386 3386 // (It may be stress-tested by specifying StressReflectiveCode.)
3387 3387 // Basically, we want to get into the VM is there's an illegal argument.
3388 3388 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3389 3389 initial_slow_test = _gvn.transform( new (C) AndINode(layout_val, bit) );
3390 3390 if (extra_slow_test != intcon(0)) {
3391 3391 initial_slow_test = _gvn.transform( new (C) OrINode(initial_slow_test, extra_slow_test) );
3392 3392 }
3393 3393 // (Macro-expander will further convert this to a Bool, if necessary.)
3394 3394 }
3395 3395
3396 3396 // Find the size in bytes. This is easy; it's the layout_helper.
3397 3397 // The size value must be valid even if the slow path is taken.
3398 3398 Node* size = NULL;
3399 3399 if (layout_is_con) {
3400 3400 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
3401 3401 } else { // reflective case
3402 3402 // This reflective path is used by clone and Unsafe.allocateInstance.
3403 3403 size = ConvI2X(layout_val);
3404 3404
3405 3405 // Clear the low bits to extract layout_helper_size_in_bytes:
3406 3406 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3407 3407 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3408 3408 size = _gvn.transform( new (C) AndXNode(size, mask) );
3409 3409 }
3410 3410 if (return_size_val != NULL) {
3411 3411 (*return_size_val) = size;
3412 3412 }
3413 3413
3414 3414 // This is a precise notnull oop of the klass.
3415 3415 // (Actually, it need not be precise if this is a reflective allocation.)
3416 3416 // It's what we cast the result to.
3417 3417 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3418 3418 if (!tklass) tklass = TypeKlassPtr::OBJECT;
3419 3419 const TypeOopPtr* oop_type = tklass->as_instance_type();
3420 3420
3421 3421 // Now generate allocation code
3422 3422
3423 3423 // The entire memory state is needed for slow path of the allocation
3424 3424 // since GC and deoptimization can happened.
3425 3425 Node *mem = reset_memory();
3426 3426 set_all_memory(mem); // Create new memory state
3427 3427
3428 3428 AllocateNode* alloc
3429 3429 = new (C) AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3430 3430 control(), mem, i_o(),
3431 3431 size, klass_node,
3432 3432 initial_slow_test);
3433 3433
3434 3434 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
3435 3435 }
3436 3436
3437 3437 //-------------------------------new_array-------------------------------------
3438 3438 // helper for both newarray and anewarray
3439 3439 // The 'length' parameter is (obviously) the length of the array.
3440 3440 // See comments on new_instance for the meaning of the other arguments.
3441 3441 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
3442 3442 Node* length, // number of array elements
3443 3443 int nargs, // number of arguments to push back for uncommon trap
3444 3444 Node* *return_size_val,
3445 3445 bool deoptimize_on_exception) {
3446 3446 jint layout_con = Klass::_lh_neutral_value;
3447 3447 Node* layout_val = get_layout_helper(klass_node, layout_con);
3448 3448 int layout_is_con = (layout_val == NULL);
3449 3449
3450 3450 if (!layout_is_con && !StressReflectiveCode &&
3451 3451 !too_many_traps(Deoptimization::Reason_class_check)) {
3452 3452 // This is a reflective array creation site.
3453 3453 // Optimistically assume that it is a subtype of Object[],
3454 3454 // so that we can fold up all the address arithmetic.
3455 3455 layout_con = Klass::array_layout_helper(T_OBJECT);
3456 3456 Node* cmp_lh = _gvn.transform( new(C) CmpINode(layout_val, intcon(layout_con)) );
3457 3457 Node* bol_lh = _gvn.transform( new(C) BoolNode(cmp_lh, BoolTest::eq) );
3458 3458 { BuildCutout unless(this, bol_lh, PROB_MAX);
3459 3459 inc_sp(nargs);
3460 3460 uncommon_trap(Deoptimization::Reason_class_check,
3461 3461 Deoptimization::Action_maybe_recompile);
3462 3462 }
3463 3463 layout_val = NULL;
3464 3464 layout_is_con = true;
3465 3465 }
3466 3466
3467 3467 // Generate the initial go-slow test. Make sure we do not overflow
3468 3468 // if length is huge (near 2Gig) or negative! We do not need
3469 3469 // exact double-words here, just a close approximation of needed
3470 3470 // double-words. We can't add any offset or rounding bits, lest we
3471 3471 // take a size -1 of bytes and make it positive. Use an unsigned
3472 3472 // compare, so negative sizes look hugely positive.
3473 3473 int fast_size_limit = FastAllocateSizeLimit;
3474 3474 if (layout_is_con) {
3475 3475 assert(!StressReflectiveCode, "stress mode does not use these paths");
3476 3476 // Increase the size limit if we have exact knowledge of array type.
3477 3477 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3478 3478 fast_size_limit <<= (LogBytesPerLong - log2_esize);
3479 3479 }
3480 3480
3481 3481 Node* initial_slow_cmp = _gvn.transform( new (C) CmpUNode( length, intcon( fast_size_limit ) ) );
3482 3482 Node* initial_slow_test = _gvn.transform( new (C) BoolNode( initial_slow_cmp, BoolTest::gt ) );
3483 3483 if (initial_slow_test->is_Bool()) {
3484 3484 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3485 3485 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3486 3486 }
3487 3487
3488 3488 // --- Size Computation ---
3489 3489 // array_size = round_to_heap(array_header + (length << elem_shift));
3490 3490 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes)
3491 3491 // and round_to(x, y) == ((x + y-1) & ~(y-1))
3492 3492 // The rounding mask is strength-reduced, if possible.
3493 3493 int round_mask = MinObjAlignmentInBytes - 1;
3494 3494 Node* header_size = NULL;
3495 3495 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3496 3496 // (T_BYTE has the weakest alignment and size restrictions...)
3497 3497 if (layout_is_con) {
3498 3498 int hsize = Klass::layout_helper_header_size(layout_con);
3499 3499 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3500 3500 BasicType etype = Klass::layout_helper_element_type(layout_con);
3501 3501 if ((round_mask & ~right_n_bits(eshift)) == 0)
3502 3502 round_mask = 0; // strength-reduce it if it goes away completely
3503 3503 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3504 3504 assert(header_size_min <= hsize, "generic minimum is smallest");
3505 3505 header_size_min = hsize;
3506 3506 header_size = intcon(hsize + round_mask);
3507 3507 } else {
3508 3508 Node* hss = intcon(Klass::_lh_header_size_shift);
3509 3509 Node* hsm = intcon(Klass::_lh_header_size_mask);
3510 3510 Node* hsize = _gvn.transform( new(C) URShiftINode(layout_val, hss) );
3511 3511 hsize = _gvn.transform( new(C) AndINode(hsize, hsm) );
3512 3512 Node* mask = intcon(round_mask);
3513 3513 header_size = _gvn.transform( new(C) AddINode(hsize, mask) );
3514 3514 }
3515 3515
3516 3516 Node* elem_shift = NULL;
3517 3517 if (layout_is_con) {
3518 3518 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3519 3519 if (eshift != 0)
3520 3520 elem_shift = intcon(eshift);
3521 3521 } else {
3522 3522 // There is no need to mask or shift this value.
3523 3523 // The semantics of LShiftINode include an implicit mask to 0x1F.
3524 3524 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3525 3525 elem_shift = layout_val;
3526 3526 }
3527 3527
3528 3528 // Transition to native address size for all offset calculations:
3529 3529 Node* lengthx = ConvI2X(length);
3530 3530 Node* headerx = ConvI2X(header_size);
3531 3531 #ifdef _LP64
3532 3532 { const TypeLong* tllen = _gvn.find_long_type(lengthx);
3533 3533 if (tllen != NULL && tllen->_lo < 0) {
3534 3534 // Add a manual constraint to a positive range. Cf. array_element_address.
3535 3535 jlong size_max = arrayOopDesc::max_array_length(T_BYTE);
3536 3536 if (size_max > tllen->_hi) size_max = tllen->_hi;
3537 3537 const TypeLong* tlcon = TypeLong::make(CONST64(0), size_max, Type::WidenMin);
3538 3538 lengthx = _gvn.transform( new (C) ConvI2LNode(length, tlcon));
3539 3539 }
3540 3540 }
3541 3541 #endif
3542 3542
3543 3543 // Combine header size (plus rounding) and body size. Then round down.
3544 3544 // This computation cannot overflow, because it is used only in two
3545 3545 // places, one where the length is sharply limited, and the other
3546 3546 // after a successful allocation.
3547 3547 Node* abody = lengthx;
3548 3548 if (elem_shift != NULL)
3549 3549 abody = _gvn.transform( new(C) LShiftXNode(lengthx, elem_shift) );
3550 3550 Node* size = _gvn.transform( new(C) AddXNode(headerx, abody) );
3551 3551 if (round_mask != 0) {
3552 3552 Node* mask = MakeConX(~round_mask);
3553 3553 size = _gvn.transform( new(C) AndXNode(size, mask) );
3554 3554 }
3555 3555 // else if round_mask == 0, the size computation is self-rounding
3556 3556
3557 3557 if (return_size_val != NULL) {
3558 3558 // This is the size
3559 3559 (*return_size_val) = size;
3560 3560 }
3561 3561
3562 3562 // Now generate allocation code
3563 3563
3564 3564 // The entire memory state is needed for slow path of the allocation
3565 3565 // since GC and deoptimization can happened.
3566 3566 Node *mem = reset_memory();
3567 3567 set_all_memory(mem); // Create new memory state
3568 3568
3569 3569 // Create the AllocateArrayNode and its result projections
3570 3570 AllocateArrayNode* alloc
3571 3571 = new (C) AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
3572 3572 control(), mem, i_o(),
3573 3573 size, klass_node,
3574 3574 initial_slow_test,
3575 3575 length);
3576 3576
3577 3577 // Cast to correct type. Note that the klass_node may be constant or not,
3578 3578 // and in the latter case the actual array type will be inexact also.
3579 3579 // (This happens via a non-constant argument to inline_native_newArray.)
3580 3580 // In any case, the value of klass_node provides the desired array type.
3581 3581 const TypeInt* length_type = _gvn.find_int_type(length);
3582 3582 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3583 3583 if (ary_type->isa_aryptr() && length_type != NULL) {
3584 3584 // Try to get a better type than POS for the size
3585 3585 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3586 3586 }
3587 3587
3588 3588 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
3589 3589
3590 3590 // Cast length on remaining path to be as narrow as possible
3591 3591 if (map()->find_edge(length) >= 0) {
3592 3592 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
3593 3593 if (ccast != length) {
3594 3594 _gvn.set_type_bottom(ccast);
3595 3595 record_for_igvn(ccast);
3596 3596 replace_in_map(length, ccast);
3597 3597 }
3598 3598 }
3599 3599
3600 3600 return javaoop;
3601 3601 }
3602 3602
3603 3603 // The following "Ideal_foo" functions are placed here because they recognize
3604 3604 // the graph shapes created by the functions immediately above.
3605 3605
3606 3606 //---------------------------Ideal_allocation----------------------------------
3607 3607 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
3608 3608 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
3609 3609 if (ptr == NULL) { // reduce dumb test in callers
3610 3610 return NULL;
3611 3611 }
3612 3612 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
3613 3613 ptr = ptr->in(1);
3614 3614 if (ptr == NULL) return NULL;
3615 3615 }
3616 3616 // Return NULL for allocations with several casts:
3617 3617 // j.l.reflect.Array.newInstance(jobject, jint)
3618 3618 // Object.clone()
3619 3619 // to keep more precise type from last cast.
3620 3620 if (ptr->is_Proj()) {
3621 3621 Node* allo = ptr->in(0);
3622 3622 if (allo != NULL && allo->is_Allocate()) {
3623 3623 return allo->as_Allocate();
3624 3624 }
3625 3625 }
3626 3626 // Report failure to match.
3627 3627 return NULL;
3628 3628 }
3629 3629
3630 3630 // Fancy version which also strips off an offset (and reports it to caller).
3631 3631 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
3632 3632 intptr_t& offset) {
3633 3633 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
3634 3634 if (base == NULL) return NULL;
3635 3635 return Ideal_allocation(base, phase);
3636 3636 }
3637 3637
3638 3638 // Trace Initialize <- Proj[Parm] <- Allocate
3639 3639 AllocateNode* InitializeNode::allocation() {
3640 3640 Node* rawoop = in(InitializeNode::RawAddress);
3641 3641 if (rawoop->is_Proj()) {
3642 3642 Node* alloc = rawoop->in(0);
3643 3643 if (alloc->is_Allocate()) {
3644 3644 return alloc->as_Allocate();
3645 3645 }
3646 3646 }
3647 3647 return NULL;
3648 3648 }
3649 3649
3650 3650 // Trace Allocate -> Proj[Parm] -> Initialize
3651 3651 InitializeNode* AllocateNode::initialization() {
3652 3652 ProjNode* rawoop = proj_out(AllocateNode::RawAddress);
3653 3653 if (rawoop == NULL) return NULL;
3654 3654 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
3655 3655 Node* init = rawoop->fast_out(i);
3656 3656 if (init->is_Initialize()) {
3657 3657 assert(init->as_Initialize()->allocation() == this, "2-way link");
3658 3658 return init->as_Initialize();
3659 3659 }
3660 3660 }
3661 3661 return NULL;
3662 3662 }
3663 3663
3664 3664 //----------------------------- loop predicates ---------------------------
3665 3665
3666 3666 //------------------------------add_predicate_impl----------------------------
3667 3667 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) {
3668 3668 // Too many traps seen?
3669 3669 if (too_many_traps(reason)) {
3670 3670 #ifdef ASSERT
3671 3671 if (TraceLoopPredicate) {
3672 3672 int tc = C->trap_count(reason);
3673 3673 tty->print("too many traps=%s tcount=%d in ",
3674 3674 Deoptimization::trap_reason_name(reason), tc);
3675 3675 method()->print(); // which method has too many predicate traps
3676 3676 tty->cr();
3677 3677 }
3678 3678 #endif
3679 3679 // We cannot afford to take more traps here,
3680 3680 // do not generate predicate.
3681 3681 return;
3682 3682 }
3683 3683
3684 3684 Node *cont = _gvn.intcon(1);
3685 3685 Node* opq = _gvn.transform(new (C) Opaque1Node(C, cont));
3686 3686 Node *bol = _gvn.transform(new (C) Conv2BNode(opq));
3687 3687 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
3688 3688 Node* iffalse = _gvn.transform(new (C) IfFalseNode(iff));
3689 3689 C->add_predicate_opaq(opq);
3690 3690 {
3691 3691 PreserveJVMState pjvms(this);
3692 3692 set_control(iffalse);
3693 3693 inc_sp(nargs);
3694 3694 uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
3695 3695 }
3696 3696 Node* iftrue = _gvn.transform(new (C) IfTrueNode(iff));
3697 3697 set_control(iftrue);
3698 3698 }
3699 3699
3700 3700 //------------------------------add_predicate---------------------------------
3701 3701 void GraphKit::add_predicate(int nargs) {
3702 3702 if (UseLoopPredicate) {
3703 3703 add_predicate_impl(Deoptimization::Reason_predicate, nargs);
3704 3704 }
3705 3705 // loop's limit check predicate should be near the loop.
3706 3706 if (LoopLimitCheck) {
3707 3707 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs);
3708 3708 }
3709 3709 }
3710 3710
3711 3711 //----------------------------- store barriers ----------------------------
3712 3712 #define __ ideal.
3713 3713
3714 3714 void GraphKit::sync_kit(IdealKit& ideal) {
3715 3715 set_all_memory(__ merged_memory());
3716 3716 set_i_o(__ i_o());
3717 3717 set_control(__ ctrl());
3718 3718 }
3719 3719
3720 3720 void GraphKit::final_sync(IdealKit& ideal) {
3721 3721 // Final sync IdealKit and graphKit.
3722 3722 sync_kit(ideal);
3723 3723 }
3724 3724
3725 3725 // vanilla/CMS post barrier
3726 3726 // Insert a write-barrier store. This is to let generational GC work; we have
3727 3727 // to flag all oop-stores before the next GC point.
3728 3728 void GraphKit::write_barrier_post(Node* oop_store,
3729 3729 Node* obj,
3730 3730 Node* adr,
3731 3731 uint adr_idx,
3732 3732 Node* val,
3733 3733 bool use_precise) {
3734 3734 // No store check needed if we're storing a NULL or an old object
3735 3735 // (latter case is probably a string constant). The concurrent
3736 3736 // mark sweep garbage collector, however, needs to have all nonNull
3737 3737 // oop updates flagged via card-marks.
3738 3738 if (val != NULL && val->is_Con()) {
3739 3739 // must be either an oop or NULL
3740 3740 const Type* t = val->bottom_type();
3741 3741 if (t == TypePtr::NULL_PTR || t == Type::TOP)
3742 3742 // stores of null never (?) need barriers
3743 3743 return;
3744 3744 }
3745 3745
3746 3746 if (use_ReduceInitialCardMarks()
3747 3747 && obj == just_allocated_object(control())) {
3748 3748 // We can skip marks on a freshly-allocated object in Eden.
3749 3749 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp.
3750 3750 // That routine informs GC to take appropriate compensating steps,
3751 3751 // upon a slow-path allocation, so as to make this card-mark
3752 3752 // elision safe.
3753 3753 return;
3754 3754 }
3755 3755
3756 3756 if (!use_precise) {
3757 3757 // All card marks for a (non-array) instance are in one place:
3758 3758 adr = obj;
3759 3759 }
3760 3760 // (Else it's an array (or unknown), and we want more precise card marks.)
3761 3761 assert(adr != NULL, "");
3762 3762
3763 3763 IdealKit ideal(this, true);
3764 3764
3765 3765 // Convert the pointer to an int prior to doing math on it
3766 3766 Node* cast = __ CastPX(__ ctrl(), adr);
3767 3767
3768 3768 // Divide by card size
3769 3769 assert(Universe::heap()->barrier_set()->kind() == BarrierSet::CardTableModRef,
3770 3770 "Only one we handle so far.");
3771 3771 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
3772 3772
3773 3773 // Combine card table base and card offset
3774 3774 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset );
3775 3775
3776 3776 // Get the alias_index for raw card-mark memory
3777 3777 int adr_type = Compile::AliasIdxRaw;
3778 3778 Node* zero = __ ConI(0); // Dirty card value
3779 3779 BasicType bt = T_BYTE;
3780 3780
3781 3781 if (UseCondCardMark) {
3782 3782 // The classic GC reference write barrier is typically implemented
3783 3783 // as a store into the global card mark table. Unfortunately
3784 3784 // unconditional stores can result in false sharing and excessive
3785 3785 // coherence traffic as well as false transactional aborts.
3786 3786 // UseCondCardMark enables MP "polite" conditional card mark
3787 3787 // stores. In theory we could relax the load from ctrl() to
3788 3788 // no_ctrl, but that doesn't buy much latitude.
3789 3789 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type);
3790 3790 __ if_then(card_val, BoolTest::ne, zero);
3791 3791 }
3792 3792
3793 3793 // Smash zero into card
3794 3794 if( !UseConcMarkSweepGC ) {
3795 3795 __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::release);
3796 3796 } else {
3797 3797 // Specialized path for CM store barrier
3798 3798 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type);
3799 3799 }
3800 3800
3801 3801 if (UseCondCardMark) {
3802 3802 __ end_if();
3803 3803 }
3804 3804
3805 3805 // Final sync IdealKit and GraphKit.
3806 3806 final_sync(ideal);
3807 3807 }
3808 3808
3809 3809 // G1 pre/post barriers
3810 3810 void GraphKit::g1_write_barrier_pre(bool do_load,
3811 3811 Node* obj,
3812 3812 Node* adr,
3813 3813 uint alias_idx,
3814 3814 Node* val,
3815 3815 const TypeOopPtr* val_type,
3816 3816 Node* pre_val,
3817 3817 BasicType bt) {
3818 3818
3819 3819 // Some sanity checks
3820 3820 // Note: val is unused in this routine.
3821 3821
3822 3822 if (do_load) {
3823 3823 // We need to generate the load of the previous value
3824 3824 assert(obj != NULL, "must have a base");
3825 3825 assert(adr != NULL, "where are loading from?");
3826 3826 assert(pre_val == NULL, "loaded already?");
3827 3827 assert(val_type != NULL, "need a type");
3828 3828 } else {
3829 3829 // In this case both val_type and alias_idx are unused.
3830 3830 assert(pre_val != NULL, "must be loaded already");
3831 3831 // Nothing to be done if pre_val is null.
3832 3832 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return;
3833 3833 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
3834 3834 }
3835 3835 assert(bt == T_OBJECT, "or we shouldn't be here");
3836 3836
3837 3837 IdealKit ideal(this, true);
3838 3838
3839 3839 Node* tls = __ thread(); // ThreadLocalStorage
3840 3840
3841 3841 Node* no_ctrl = NULL;
3842 3842 Node* no_base = __ top();
3843 3843 Node* zero = __ ConI(0);
3844 3844 Node* zeroX = __ ConX(0);
3845 3845
3846 3846 float likely = PROB_LIKELY(0.999);
3847 3847 float unlikely = PROB_UNLIKELY(0.999);
3848 3848
3849 3849 BasicType active_type = in_bytes(PtrQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE;
3850 3850 assert(in_bytes(PtrQueue::byte_width_of_active()) == 4 || in_bytes(PtrQueue::byte_width_of_active()) == 1, "flag width");
3851 3851
3852 3852 // Offsets into the thread
3853 3853 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648
3854 3854 PtrQueue::byte_offset_of_active());
3855 3855 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656
3856 3856 PtrQueue::byte_offset_of_index());
3857 3857 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652
3858 3858 PtrQueue::byte_offset_of_buf());
3859 3859
3860 3860 // Now the actual pointers into the thread
3861 3861 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset));
3862 3862 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
3863 3863 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
3864 3864
3865 3865 // Now some of the values
3866 3866 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw);
3867 3867
3868 3868 // if (!marking)
3869 3869 __ if_then(marking, BoolTest::ne, zero, unlikely); {
3870 3870 BasicType index_bt = TypeX_X->basic_type();
3871 3871 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 PtrQueue::_index with wrong size.");
3872 3872 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
3873 3873
3874 3874 if (do_load) {
3875 3875 // load original value
3876 3876 // alias_idx correct??
3877 3877 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx);
3878 3878 }
3879 3879
3880 3880 // if (pre_val != NULL)
3881 3881 __ if_then(pre_val, BoolTest::ne, null()); {
3882 3882 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
3883 3883
3884 3884 // is the queue for this thread full?
3885 3885 __ if_then(index, BoolTest::ne, zeroX, likely); {
3886 3886
3887 3887 // decrement the index
3888 3888 Node* next_index = _gvn.transform(new (C) SubXNode(index, __ ConX(sizeof(intptr_t))));
3889 3889
3890 3890 // Now get the buffer location we will log the previous value into and store it
3891 3891 Node *log_addr = __ AddP(no_base, buffer, next_index);
3892 3892 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered);
3893 3893 // update the index
3894 3894 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered);
3895 3895
3896 3896 } __ else_(); {
3897 3897
3898 3898 // logging buffer is full, call the runtime
3899 3899 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type();
3900 3900 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls);
3901 3901 } __ end_if(); // (!index)
3902 3902 } __ end_if(); // (pre_val != NULL)
3903 3903 } __ end_if(); // (!marking)
3904 3904
3905 3905 // Final sync IdealKit and GraphKit.
3906 3906 final_sync(ideal);
3907 3907 }
3908 3908
3909 3909 //
3910 3910 // Update the card table and add card address to the queue
3911 3911 //
3912 3912 void GraphKit::g1_mark_card(IdealKit& ideal,
3913 3913 Node* card_adr,
3914 3914 Node* oop_store,
3915 3915 uint oop_alias_idx,
3916 3916 Node* index,
3917 3917 Node* index_adr,
3918 3918 Node* buffer,
3919 3919 const TypeFunc* tf) {
3920 3920
3921 3921 Node* zero = __ ConI(0);
3922 3922 Node* zeroX = __ ConX(0);
3923 3923 Node* no_base = __ top();
3924 3924 BasicType card_bt = T_BYTE;
3925 3925 // Smash zero into card. MUST BE ORDERED WRT TO STORE
3926 3926 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw);
3927 3927
3928 3928 // Now do the queue work
3929 3929 __ if_then(index, BoolTest::ne, zeroX); {
3930 3930
3931 3931 Node* next_index = _gvn.transform(new (C) SubXNode(index, __ ConX(sizeof(intptr_t))));
3932 3932 Node* log_addr = __ AddP(no_base, buffer, next_index);
3933 3933
3934 3934 // Order, see storeCM.
3935 3935 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered);
3936 3936 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered);
3937 3937
3938 3938 } __ else_(); {
3939 3939 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread());
3940 3940 } __ end_if();
3941 3941
3942 3942 }
3943 3943
3944 3944 void GraphKit::g1_write_barrier_post(Node* oop_store,
3945 3945 Node* obj,
3946 3946 Node* adr,
3947 3947 uint alias_idx,
3948 3948 Node* val,
3949 3949 BasicType bt,
3950 3950 bool use_precise) {
3951 3951 // If we are writing a NULL then we need no post barrier
3952 3952
3953 3953 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) {
3954 3954 // Must be NULL
3955 3955 const Type* t = val->bottom_type();
3956 3956 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL");
3957 3957 // No post barrier if writing NULLx
3958 3958 return;
3959 3959 }
3960 3960
3961 3961 if (!use_precise) {
3962 3962 // All card marks for a (non-array) instance are in one place:
3963 3963 adr = obj;
3964 3964 }
3965 3965 // (Else it's an array (or unknown), and we want more precise card marks.)
3966 3966 assert(adr != NULL, "");
3967 3967
3968 3968 IdealKit ideal(this, true);
3969 3969
3970 3970 Node* tls = __ thread(); // ThreadLocalStorage
3971 3971
3972 3972 Node* no_base = __ top();
3973 3973 float likely = PROB_LIKELY(0.999);
3974 3974 float unlikely = PROB_UNLIKELY(0.999);
3975 3975 Node* young_card = __ ConI((jint)G1SATBCardTableModRefBS::g1_young_card_val());
3976 3976 Node* dirty_card = __ ConI((jint)CardTableModRefBS::dirty_card_val());
3977 3977 Node* zeroX = __ ConX(0);
3978 3978
3979 3979 // Get the alias_index for raw card-mark memory
3980 3980 const TypePtr* card_type = TypeRawPtr::BOTTOM;
3981 3981
3982 3982 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type();
3983 3983
3984 3984 // Offsets into the thread
3985 3985 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
3986 3986 PtrQueue::byte_offset_of_index());
3987 3987 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
3988 3988 PtrQueue::byte_offset_of_buf());
3989 3989
3990 3990 // Pointers into the thread
3991 3991
3992 3992 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
3993 3993 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
3994 3994
3995 3995 // Now some values
3996 3996 // Use ctrl to avoid hoisting these values past a safepoint, which could
3997 3997 // potentially reset these fields in the JavaThread.
3998 3998 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw);
3999 3999 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
4000 4000
4001 4001 // Convert the store obj pointer to an int prior to doing math on it
4002 4002 // Must use ctrl to prevent "integerized oop" existing across safepoint
4003 4003 Node* cast = __ CastPX(__ ctrl(), adr);
4004 4004
4005 4005 // Divide pointer by card size
4006 4006 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
4007 4007
4008 4008 // Combine card table base and card offset
4009 4009 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset );
4010 4010
4011 4011 // If we know the value being stored does it cross regions?
4012 4012
4013 4013 if (val != NULL) {
4014 4014 // Does the store cause us to cross regions?
4015 4015
4016 4016 // Should be able to do an unsigned compare of region_size instead of
4017 4017 // and extra shift. Do we have an unsigned compare??
4018 4018 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes);
4019 4019 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes));
4020 4020
4021 4021 // if (xor_res == 0) same region so skip
4022 4022 __ if_then(xor_res, BoolTest::ne, zeroX); {
4023 4023
4024 4024 // No barrier if we are storing a NULL
4025 4025 __ if_then(val, BoolTest::ne, null(), unlikely); {
4026 4026
4027 4027 // Ok must mark the card if not already dirty
4028 4028
4029 4029 // load the original value of the card
4030 4030 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4031 4031
4032 4032 __ if_then(card_val, BoolTest::ne, young_card); {
4033 4033 sync_kit(ideal);
4034 4034 // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier.
4035 4035 insert_mem_bar(Op_MemBarVolatile, oop_store);
4036 4036 __ sync_kit(this);
4037 4037
4038 4038 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4039 4039 __ if_then(card_val_reload, BoolTest::ne, dirty_card); {
4040 4040 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4041 4041 } __ end_if();
4042 4042 } __ end_if();
4043 4043 } __ end_if();
4044 4044 } __ end_if();
4045 4045 } else {
4046 4046 // Object.clone() instrinsic uses this path.
4047 4047 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4048 4048 }
4049 4049
4050 4050 // Final sync IdealKit and GraphKit.
4051 4051 final_sync(ideal);
4052 4052 }
4053 4053 #undef __
4054 4054
4055 4055
4056 4056
4057 4057 Node* GraphKit::load_String_offset(Node* ctrl, Node* str) {
4058 4058 if (java_lang_String::has_offset_field()) {
4059 4059 int offset_offset = java_lang_String::offset_offset_in_bytes();
4060 4060 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4061 4061 false, NULL, 0);
4062 4062 const TypePtr* offset_field_type = string_type->add_offset(offset_offset);
4063 4063 int offset_field_idx = C->get_alias_index(offset_field_type);
4064 4064 return make_load(ctrl,
4065 4065 basic_plus_adr(str, str, offset_offset),
4066 4066 TypeInt::INT, T_INT, offset_field_idx, MemNode::unordered);
4067 4067 } else {
4068 4068 return intcon(0);
4069 4069 }
4070 4070 }
4071 4071
4072 4072 Node* GraphKit::load_String_length(Node* ctrl, Node* str) {
4073 4073 if (java_lang_String::has_count_field()) {
4074 4074 int count_offset = java_lang_String::count_offset_in_bytes();
4075 4075 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4076 4076 false, NULL, 0);
4077 4077 const TypePtr* count_field_type = string_type->add_offset(count_offset);
4078 4078 int count_field_idx = C->get_alias_index(count_field_type);
4079 4079 return make_load(ctrl,
4080 4080 basic_plus_adr(str, str, count_offset),
4081 4081 TypeInt::INT, T_INT, count_field_idx, MemNode::unordered);
4082 4082 } else {
4083 4083 return load_array_length(load_String_value(ctrl, str));
4084 4084 }
4085 4085 }
4086 4086
4087 4087 Node* GraphKit::load_String_value(Node* ctrl, Node* str) {
4088 4088 int value_offset = java_lang_String::value_offset_in_bytes();
4089 4089 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4090 4090 false, NULL, 0);
4091 4091 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4092 4092 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4093 4093 TypeAry::make(TypeInt::CHAR,TypeInt::POS),
4094 4094 ciTypeArrayKlass::make(T_CHAR), true, 0);
4095 4095 int value_field_idx = C->get_alias_index(value_field_type);
4096 4096 Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset),
4097 4097 value_type, T_OBJECT, value_field_idx, MemNode::unordered);
4098 4098 // String.value field is known to be @Stable.
4099 4099 if (UseImplicitStableValues) {
4100 4100 load = cast_array_to_stable(load, value_type);
4101 4101 }
4102 4102 return load;
4103 4103 }
4104 4104
4105 4105 void GraphKit::store_String_offset(Node* ctrl, Node* str, Node* value) {
4106 4106 int offset_offset = java_lang_String::offset_offset_in_bytes();
4107 4107 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4108 4108 false, NULL, 0);
4109 4109 const TypePtr* offset_field_type = string_type->add_offset(offset_offset);
4110 4110 int offset_field_idx = C->get_alias_index(offset_field_type);
4111 4111 store_to_memory(ctrl, basic_plus_adr(str, offset_offset),
4112 4112 value, T_INT, offset_field_idx, MemNode::unordered);
4113 4113 }
4114 4114
4115 4115 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) {
4116 4116 int value_offset = java_lang_String::value_offset_in_bytes();
4117 4117 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4118 4118 false, NULL, 0);
4119 4119 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4120 4120
4121 4121 store_oop_to_object(ctrl, str, basic_plus_adr(str, value_offset), value_field_type,
4122 4122 value, TypeAryPtr::CHARS, T_OBJECT, MemNode::unordered);
4123 4123 }
4124 4124
4125 4125 void GraphKit::store_String_length(Node* ctrl, Node* str, Node* value) {
4126 4126 int count_offset = java_lang_String::count_offset_in_bytes();
4127 4127 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4128 4128 false, NULL, 0);
4129 4129 const TypePtr* count_field_type = string_type->add_offset(count_offset);
4130 4130 int count_field_idx = C->get_alias_index(count_field_type);
4131 4131 store_to_memory(ctrl, basic_plus_adr(str, count_offset),
4132 4132 value, T_INT, count_field_idx, MemNode::unordered);
4133 4133 }
4134 4134
4135 4135 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) {
4136 4136 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity
4137 4137 // assumption of CCP analysis.
4138 4138 return _gvn.transform(new(C) CastPPNode(ary, ary_type->cast_to_stable(true)));
4139 4139 }
↓ open down ↓ |
1584 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX