Print this page
| Split |
Split |
Close |
| Expand all |
| Collapse all |
--- old/src/share/vm/opto/graphKit.cpp
+++ new/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 }
432 432
433 433 //--------------------------use_exception_state--------------------------------
434 434 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
435 435 if (failing()) { stop(); return top(); }
436 436 Node* region = phi_map->control();
437 437 Node* hidden_merge_mark = root();
438 438 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
439 439 Node* ex_oop = clear_saved_ex_oop(phi_map);
440 440 if (region->in(0) == hidden_merge_mark) {
441 441 // Special marking for internal ex-states. Process the phis now.
442 442 region->set_req(0, region); // now it's an ordinary region
443 443 set_jvms(phi_map->jvms()); // ...so now we can use it as a map
444 444 // Note: Setting the jvms also sets the bci and sp.
445 445 set_control(_gvn.transform(region));
446 446 uint tos = jvms()->stkoff() + sp();
447 447 for (uint i = 1; i < tos; i++) {
448 448 Node* x = phi_map->in(i);
449 449 if (x->in(0) == region) {
450 450 assert(x->is_Phi(), "expected a special phi");
451 451 phi_map->set_req(i, _gvn.transform(x));
452 452 }
453 453 }
454 454 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
455 455 Node* x = mms.memory();
456 456 if (x->in(0) == region) {
457 457 assert(x->is_Phi(), "nobody else uses a hidden region");
458 458 mms.set_memory(_gvn.transform(x));
459 459 }
460 460 }
461 461 if (ex_oop->in(0) == region) {
462 462 assert(ex_oop->is_Phi(), "expected a special phi");
463 463 ex_oop = _gvn.transform(ex_oop);
464 464 }
465 465 } else {
466 466 set_jvms(phi_map->jvms());
467 467 }
468 468
469 469 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
470 470 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
471 471 return ex_oop;
472 472 }
473 473
474 474 //---------------------------------java_bc-------------------------------------
475 475 Bytecodes::Code GraphKit::java_bc() const {
476 476 ciMethod* method = this->method();
477 477 int bci = this->bci();
478 478 if (method != NULL && bci != InvocationEntryBci)
479 479 return method->java_code_at_bci(bci);
480 480 else
481 481 return Bytecodes::_illegal;
482 482 }
483 483
484 484 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
485 485 bool must_throw) {
486 486 // if the exception capability is set, then we will generate code
487 487 // to check the JavaThread.should_post_on_exceptions flag to see
488 488 // if we actually need to report exception events (for this
489 489 // thread). If we don't need to report exception events, we will
490 490 // take the normal fast path provided by add_exception_events. If
491 491 // exception event reporting is enabled for this thread, we will
492 492 // take the uncommon_trap in the BuildCutout below.
493 493
494 494 // first must access the should_post_on_exceptions_flag in this thread's JavaThread
495 495 Node* jthread = _gvn.transform(new (C) ThreadLocalNode());
496 496 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
497 497 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered);
498 498
499 499 // Test the should_post_on_exceptions_flag vs. 0
500 500 Node* chk = _gvn.transform( new (C) CmpINode(should_post_flag, intcon(0)) );
501 501 Node* tst = _gvn.transform( new (C) BoolNode(chk, BoolTest::eq) );
502 502
503 503 // Branch to slow_path if should_post_on_exceptions_flag was true
504 504 { BuildCutout unless(this, tst, PROB_MAX);
505 505 // Do not try anything fancy if we're notifying the VM on every throw.
506 506 // Cf. case Bytecodes::_athrow in parse2.cpp.
507 507 uncommon_trap(reason, Deoptimization::Action_none,
508 508 (ciKlass*)NULL, (char*)NULL, must_throw);
509 509 }
510 510
511 511 }
512 512
513 513 //------------------------------builtin_throw----------------------------------
514 514 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
515 515 bool must_throw = true;
516 516
517 517 if (env()->jvmti_can_post_on_exceptions()) {
518 518 // check if we must post exception events, take uncommon trap if so
519 519 uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
520 520 // here if should_post_on_exceptions is false
521 521 // continue on with the normal codegen
522 522 }
523 523
524 524 // If this particular condition has not yet happened at this
525 525 // bytecode, then use the uncommon trap mechanism, and allow for
526 526 // a future recompilation if several traps occur here.
527 527 // If the throw is hot, try to use a more complicated inline mechanism
528 528 // which keeps execution inside the compiled code.
529 529 bool treat_throw_as_hot = false;
530 530 ciMethodData* md = method()->method_data();
531 531
532 532 if (ProfileTraps) {
533 533 if (too_many_traps(reason)) {
534 534 treat_throw_as_hot = true;
535 535 }
536 536 // (If there is no MDO at all, assume it is early in
537 537 // execution, and that any deopts are part of the
538 538 // startup transient, and don't need to be remembered.)
539 539
540 540 // Also, if there is a local exception handler, treat all throws
541 541 // as hot if there has been at least one in this method.
542 542 if (C->trap_count(reason) != 0
543 543 && method()->method_data()->trap_count(reason) != 0
544 544 && has_ex_handler()) {
545 545 treat_throw_as_hot = true;
546 546 }
547 547 }
548 548
549 549 // If this throw happens frequently, an uncommon trap might cause
550 550 // a performance pothole. If there is a local exception handler,
551 551 // and if this particular bytecode appears to be deoptimizing often,
552 552 // let us handle the throw inline, with a preconstructed instance.
553 553 // Note: If the deopt count has blown up, the uncommon trap
554 554 // runtime is going to flush this nmethod, not matter what.
555 555 if (treat_throw_as_hot
556 556 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
557 557 // If the throw is local, we use a pre-existing instance and
558 558 // punt on the backtrace. This would lead to a missing backtrace
559 559 // (a repeat of 4292742) if the backtrace object is ever asked
560 560 // for its backtrace.
561 561 // Fixing this remaining case of 4292742 requires some flavor of
562 562 // escape analysis. Leave that for the future.
563 563 ciInstance* ex_obj = NULL;
564 564 switch (reason) {
565 565 case Deoptimization::Reason_null_check:
566 566 ex_obj = env()->NullPointerException_instance();
567 567 break;
568 568 case Deoptimization::Reason_div0_check:
569 569 ex_obj = env()->ArithmeticException_instance();
570 570 break;
571 571 case Deoptimization::Reason_range_check:
572 572 ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
573 573 break;
574 574 case Deoptimization::Reason_class_check:
575 575 if (java_bc() == Bytecodes::_aastore) {
576 576 ex_obj = env()->ArrayStoreException_instance();
577 577 } else {
578 578 ex_obj = env()->ClassCastException_instance();
579 579 }
580 580 break;
581 581 }
582 582 if (failing()) { stop(); return; } // exception allocation might fail
583 583 if (ex_obj != NULL) {
584 584 // Cheat with a preallocated exception object.
585 585 if (C->log() != NULL)
586 586 C->log()->elem("hot_throw preallocated='1' reason='%s'",
587 587 Deoptimization::trap_reason_name(reason));
588 588 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj);
589 589 Node* ex_node = _gvn.transform( ConNode::make(C, ex_con) );
590 590
591 591 // Clear the detail message of the preallocated exception object.
592 592 // Weblogic sometimes mutates the detail message of exceptions
593 593 // using reflection.
594 594 int offset = java_lang_Throwable::get_detailMessage_offset();
595 595 const TypePtr* adr_typ = ex_con->add_offset(offset);
596 596
597 597 Node *adr = basic_plus_adr(ex_node, ex_node, offset);
598 598 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
599 599 // Conservatively release stores of object references.
600 600 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, MemNode::release);
601 601
602 602 add_exception_state(make_exception_state(ex_node));
603 603 return;
604 604 }
605 605 }
606 606
607 607 // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
608 608 // It won't be much cheaper than bailing to the interp., since we'll
609 609 // have to pass up all the debug-info, and the runtime will have to
610 610 // create the stack trace.
611 611
612 612 // Usual case: Bail to interpreter.
613 613 // Reserve the right to recompile if we haven't seen anything yet.
614 614
615 615 assert(!Deoptimization::reason_is_speculate(reason), "unsupported");
616 616 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
617 617 if (treat_throw_as_hot
618 618 && (method()->method_data()->trap_recompiled_at(bci(), NULL)
619 619 || C->too_many_traps(reason))) {
620 620 // We cannot afford to take more traps here. Suffer in the interpreter.
621 621 if (C->log() != NULL)
622 622 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
623 623 Deoptimization::trap_reason_name(reason),
624 624 C->trap_count(reason));
625 625 action = Deoptimization::Action_none;
626 626 }
627 627
628 628 // "must_throw" prunes the JVM state to include only the stack, if there
629 629 // are no local exception handlers. This should cut down on register
630 630 // allocation time and code size, by drastically reducing the number
631 631 // of in-edges on the call to the uncommon trap.
632 632
633 633 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
634 634 }
635 635
636 636
637 637 //----------------------------PreserveJVMState---------------------------------
638 638 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
639 639 debug_only(kit->verify_map());
640 640 _kit = kit;
641 641 _map = kit->map(); // preserve the map
642 642 _sp = kit->sp();
643 643 kit->set_map(clone_map ? kit->clone_map() : NULL);
644 644 Compile::current()->inc_preserve_jvm_state();
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 Compile::current()->dec_preserve_jvm_state();
663 663 }
664 664
665 665
666 666 //-----------------------------BuildCutout-------------------------------------
667 667 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
668 668 : PreserveJVMState(kit)
669 669 {
670 670 assert(p->is_Con() || p->is_Bool(), "test must be a bool");
671 671 SafePointNode* outer_map = _map; // preserved map is caller's
672 672 SafePointNode* inner_map = kit->map();
673 673 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
674 674 outer_map->set_control(kit->gvn().transform( new (kit->C) IfTrueNode(iff) ));
675 675 inner_map->set_control(kit->gvn().transform( new (kit->C) IfFalseNode(iff) ));
676 676 }
677 677 BuildCutout::~BuildCutout() {
678 678 GraphKit* kit = _kit;
679 679 assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
680 680 }
681 681
682 682 //---------------------------PreserveReexecuteState----------------------------
683 683 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
684 684 assert(!kit->stopped(), "must call stopped() before");
685 685 _kit = kit;
686 686 _sp = kit->sp();
687 687 _reexecute = kit->jvms()->_reexecute;
688 688 }
689 689 PreserveReexecuteState::~PreserveReexecuteState() {
690 690 if (_kit->stopped()) return;
691 691 _kit->jvms()->_reexecute = _reexecute;
692 692 _kit->set_sp(_sp);
693 693 }
694 694
695 695 //------------------------------clone_map--------------------------------------
696 696 // Implementation of PreserveJVMState
697 697 //
698 698 // Only clone_map(...) here. If this function is only used in the
699 699 // PreserveJVMState class we may want to get rid of this extra
700 700 // function eventually and do it all there.
701 701
702 702 SafePointNode* GraphKit::clone_map() {
703 703 if (map() == NULL) return NULL;
704 704
705 705 // Clone the memory edge first
706 706 Node* mem = MergeMemNode::make(C, map()->memory());
707 707 gvn().set_type_bottom(mem);
708 708
709 709 SafePointNode *clonemap = (SafePointNode*)map()->clone();
710 710 JVMState* jvms = this->jvms();
711 711 JVMState* clonejvms = jvms->clone_shallow(C);
712 712 clonemap->set_memory(mem);
713 713 clonemap->set_jvms(clonejvms);
714 714 clonejvms->set_map(clonemap);
715 715 record_for_igvn(clonemap);
716 716 gvn().set_type_bottom(clonemap);
717 717 return clonemap;
718 718 }
719 719
720 720
721 721 //-----------------------------set_map_clone-----------------------------------
722 722 void GraphKit::set_map_clone(SafePointNode* m) {
723 723 _map = m;
724 724 _map = clone_map();
725 725 _map->set_next_exception(NULL);
726 726 debug_only(verify_map());
727 727 }
728 728
729 729
730 730 //----------------------------kill_dead_locals---------------------------------
731 731 // Detect any locals which are known to be dead, and force them to top.
732 732 void GraphKit::kill_dead_locals() {
733 733 // Consult the liveness information for the locals. If any
734 734 // of them are unused, then they can be replaced by top(). This
735 735 // should help register allocation time and cut down on the size
736 736 // of the deoptimization information.
737 737
738 738 // This call is made from many of the bytecode handling
739 739 // subroutines called from the Big Switch in do_one_bytecode.
740 740 // Every bytecode which might include a slow path is responsible
741 741 // for killing its dead locals. The more consistent we
742 742 // are about killing deads, the fewer useless phis will be
743 743 // constructed for them at various merge points.
744 744
745 745 // bci can be -1 (InvocationEntryBci). We return the entry
746 746 // liveness for the method.
747 747
748 748 if (method() == NULL || method()->code_size() == 0) {
749 749 // We are building a graph for a call to a native method.
750 750 // All locals are live.
751 751 return;
752 752 }
753 753
754 754 ResourceMark rm;
755 755
756 756 // Consult the liveness information for the locals. If any
757 757 // of them are unused, then they can be replaced by top(). This
758 758 // should help register allocation time and cut down on the size
759 759 // of the deoptimization information.
760 760 MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
761 761
762 762 int len = (int)live_locals.size();
763 763 assert(len <= jvms()->loc_size(), "too many live locals");
764 764 for (int local = 0; local < len; local++) {
765 765 if (!live_locals.at(local)) {
766 766 set_local(local, top());
767 767 }
768 768 }
769 769 }
770 770
771 771 #ifdef ASSERT
772 772 //-------------------------dead_locals_are_killed------------------------------
773 773 // Return true if all dead locals are set to top in the map.
774 774 // Used to assert "clean" debug info at various points.
775 775 bool GraphKit::dead_locals_are_killed() {
776 776 if (method() == NULL || method()->code_size() == 0) {
777 777 // No locals need to be dead, so all is as it should be.
778 778 return true;
779 779 }
780 780
781 781 // Make sure somebody called kill_dead_locals upstream.
782 782 ResourceMark rm;
783 783 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
784 784 if (jvms->loc_size() == 0) continue; // no locals to consult
785 785 SafePointNode* map = jvms->map();
786 786 ciMethod* method = jvms->method();
787 787 int bci = jvms->bci();
788 788 if (jvms == this->jvms()) {
789 789 bci = this->bci(); // it might not yet be synched
790 790 }
791 791 MethodLivenessResult live_locals = method->liveness_at_bci(bci);
792 792 int len = (int)live_locals.size();
793 793 if (!live_locals.is_valid() || len == 0)
794 794 // This method is trivial, or is poisoned by a breakpoint.
795 795 return true;
796 796 assert(len == jvms->loc_size(), "live map consistent with locals map");
797 797 for (int local = 0; local < len; local++) {
798 798 if (!live_locals.at(local) && map->local(jvms, local) != top()) {
799 799 if (PrintMiscellaneous && (Verbose || WizardMode)) {
800 800 tty->print_cr("Zombie local %d: ", local);
801 801 jvms->dump();
802 802 }
803 803 return false;
804 804 }
805 805 }
806 806 }
807 807 return true;
808 808 }
809 809
810 810 #endif //ASSERT
811 811
812 812 // Helper function for enforcing certain bytecodes to reexecute if
813 813 // deoptimization happens
814 814 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
815 815 ciMethod* cur_method = jvms->method();
816 816 int cur_bci = jvms->bci();
817 817 if (cur_method != NULL && cur_bci != InvocationEntryBci) {
818 818 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
819 819 return Interpreter::bytecode_should_reexecute(code) ||
820 820 is_anewarray && code == Bytecodes::_multianewarray;
821 821 // Reexecute _multianewarray bytecode which was replaced with
822 822 // sequence of [a]newarray. See Parse::do_multianewarray().
823 823 //
824 824 // Note: interpreter should not have it set since this optimization
825 825 // is limited by dimensions and guarded by flag so in some cases
826 826 // multianewarray() runtime calls will be generated and
827 827 // the bytecode should not be reexecutes (stack will not be reset).
828 828 } else
829 829 return false;
830 830 }
831 831
832 832 // Helper function for adding JVMState and debug information to node
833 833 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
834 834 // Add the safepoint edges to the call (or other safepoint).
835 835
836 836 // Make sure dead locals are set to top. This
837 837 // should help register allocation time and cut down on the size
838 838 // of the deoptimization information.
839 839 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
840 840
841 841 // Walk the inline list to fill in the correct set of JVMState's
842 842 // Also fill in the associated edges for each JVMState.
843 843
844 844 // If the bytecode needs to be reexecuted we need to put
845 845 // the arguments back on the stack.
846 846 const bool should_reexecute = jvms()->should_reexecute();
847 847 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
848 848
849 849 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
850 850 // undefined if the bci is different. This is normal for Parse but it
851 851 // should not happen for LibraryCallKit because only one bci is processed.
852 852 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
853 853 "in LibraryCallKit the reexecute bit should not change");
854 854
855 855 // If we are guaranteed to throw, we can prune everything but the
856 856 // input to the current bytecode.
857 857 bool can_prune_locals = false;
858 858 uint stack_slots_not_pruned = 0;
859 859 int inputs = 0, depth = 0;
860 860 if (must_throw) {
861 861 assert(method() == youngest_jvms->method(), "sanity");
862 862 if (compute_stack_effects(inputs, depth)) {
863 863 can_prune_locals = true;
864 864 stack_slots_not_pruned = inputs;
865 865 }
866 866 }
867 867
868 868 if (env()->should_retain_local_variables()) {
869 869 // At any safepoint, this method can get breakpointed, which would
870 870 // then require an immediate deoptimization.
871 871 can_prune_locals = false; // do not prune locals
872 872 stack_slots_not_pruned = 0;
873 873 }
874 874
875 875 // do not scribble on the input jvms
876 876 JVMState* out_jvms = youngest_jvms->clone_deep(C);
877 877 call->set_jvms(out_jvms); // Start jvms list for call node
878 878
879 879 // For a known set of bytecodes, the interpreter should reexecute them if
880 880 // deoptimization happens. We set the reexecute state for them here
881 881 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
882 882 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
883 883 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
884 884 }
885 885
886 886 // Presize the call:
887 887 DEBUG_ONLY(uint non_debug_edges = call->req());
888 888 call->add_req_batch(top(), youngest_jvms->debug_depth());
889 889 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
890 890
891 891 // Set up edges so that the call looks like this:
892 892 // Call [state:] ctl io mem fptr retadr
893 893 // [parms:] parm0 ... parmN
894 894 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
895 895 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
896 896 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
897 897 // Note that caller debug info precedes callee debug info.
898 898
899 899 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
900 900 uint debug_ptr = call->req();
901 901
902 902 // Loop over the map input edges associated with jvms, add them
903 903 // to the call node, & reset all offsets to match call node array.
904 904 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
905 905 uint debug_end = debug_ptr;
906 906 uint debug_start = debug_ptr - in_jvms->debug_size();
907 907 debug_ptr = debug_start; // back up the ptr
908 908
909 909 uint p = debug_start; // walks forward in [debug_start, debug_end)
910 910 uint j, k, l;
911 911 SafePointNode* in_map = in_jvms->map();
912 912 out_jvms->set_map(call);
913 913
914 914 if (can_prune_locals) {
915 915 assert(in_jvms->method() == out_jvms->method(), "sanity");
916 916 // If the current throw can reach an exception handler in this JVMS,
917 917 // then we must keep everything live that can reach that handler.
918 918 // As a quick and dirty approximation, we look for any handlers at all.
919 919 if (in_jvms->method()->has_exception_handlers()) {
920 920 can_prune_locals = false;
921 921 }
922 922 }
923 923
924 924 // Add the Locals
925 925 k = in_jvms->locoff();
926 926 l = in_jvms->loc_size();
927 927 out_jvms->set_locoff(p);
928 928 if (!can_prune_locals) {
929 929 for (j = 0; j < l; j++)
930 930 call->set_req(p++, in_map->in(k+j));
931 931 } else {
932 932 p += l; // already set to top above by add_req_batch
933 933 }
934 934
935 935 // Add the Expression Stack
936 936 k = in_jvms->stkoff();
937 937 l = in_jvms->sp();
938 938 out_jvms->set_stkoff(p);
939 939 if (!can_prune_locals) {
940 940 for (j = 0; j < l; j++)
941 941 call->set_req(p++, in_map->in(k+j));
942 942 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
943 943 // Divide stack into {S0,...,S1}, where S0 is set to top.
944 944 uint s1 = stack_slots_not_pruned;
945 945 stack_slots_not_pruned = 0; // for next iteration
946 946 if (s1 > l) s1 = l;
947 947 uint s0 = l - s1;
948 948 p += s0; // skip the tops preinstalled by add_req_batch
949 949 for (j = s0; j < l; j++)
950 950 call->set_req(p++, in_map->in(k+j));
951 951 } else {
952 952 p += l; // already set to top above by add_req_batch
953 953 }
954 954
955 955 // Add the Monitors
956 956 k = in_jvms->monoff();
957 957 l = in_jvms->mon_size();
958 958 out_jvms->set_monoff(p);
959 959 for (j = 0; j < l; j++)
960 960 call->set_req(p++, in_map->in(k+j));
961 961
962 962 // Copy any scalar object fields.
963 963 k = in_jvms->scloff();
964 964 l = in_jvms->scl_size();
965 965 out_jvms->set_scloff(p);
966 966 for (j = 0; j < l; j++)
967 967 call->set_req(p++, in_map->in(k+j));
968 968
969 969 // Finish the new jvms.
970 970 out_jvms->set_endoff(p);
971 971
972 972 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
973 973 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
974 974 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
975 975 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
976 976 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
977 977 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
978 978
979 979 // Update the two tail pointers in parallel.
980 980 out_jvms = out_jvms->caller();
981 981 in_jvms = in_jvms->caller();
982 982 }
983 983
984 984 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
985 985
986 986 // Test the correctness of JVMState::debug_xxx accessors:
987 987 assert(call->jvms()->debug_start() == non_debug_edges, "");
988 988 assert(call->jvms()->debug_end() == call->req(), "");
989 989 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
990 990 }
991 991
992 992 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
993 993 Bytecodes::Code code = java_bc();
994 994 if (code == Bytecodes::_wide) {
995 995 code = method()->java_code_at_bci(bci() + 1);
996 996 }
997 997
998 998 BasicType rtype = T_ILLEGAL;
999 999 int rsize = 0;
1000 1000
1001 1001 if (code != Bytecodes::_illegal) {
1002 1002 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1003 1003 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1004 1004 if (rtype < T_CONFLICT)
1005 1005 rsize = type2size[rtype];
1006 1006 }
1007 1007
1008 1008 switch (code) {
1009 1009 case Bytecodes::_illegal:
1010 1010 return false;
1011 1011
1012 1012 case Bytecodes::_ldc:
1013 1013 case Bytecodes::_ldc_w:
1014 1014 case Bytecodes::_ldc2_w:
1015 1015 inputs = 0;
1016 1016 break;
1017 1017
1018 1018 case Bytecodes::_dup: inputs = 1; break;
1019 1019 case Bytecodes::_dup_x1: inputs = 2; break;
1020 1020 case Bytecodes::_dup_x2: inputs = 3; break;
1021 1021 case Bytecodes::_dup2: inputs = 2; break;
1022 1022 case Bytecodes::_dup2_x1: inputs = 3; break;
1023 1023 case Bytecodes::_dup2_x2: inputs = 4; break;
1024 1024 case Bytecodes::_swap: inputs = 2; break;
1025 1025 case Bytecodes::_arraylength: inputs = 1; break;
1026 1026
1027 1027 case Bytecodes::_getstatic:
1028 1028 case Bytecodes::_putstatic:
1029 1029 case Bytecodes::_getfield:
1030 1030 case Bytecodes::_putfield:
1031 1031 {
1032 1032 bool ignored_will_link;
1033 1033 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1034 1034 int size = field->type()->size();
1035 1035 bool is_get = (depth >= 0), is_static = (depth & 1);
1036 1036 inputs = (is_static ? 0 : 1);
1037 1037 if (is_get) {
1038 1038 depth = size - inputs;
1039 1039 } else {
1040 1040 inputs += size; // putxxx pops the value from the stack
1041 1041 depth = - inputs;
1042 1042 }
1043 1043 }
1044 1044 break;
1045 1045
1046 1046 case Bytecodes::_invokevirtual:
1047 1047 case Bytecodes::_invokespecial:
1048 1048 case Bytecodes::_invokestatic:
1049 1049 case Bytecodes::_invokedynamic:
1050 1050 case Bytecodes::_invokeinterface:
1051 1051 {
1052 1052 bool ignored_will_link;
1053 1053 ciSignature* declared_signature = NULL;
1054 1054 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1055 1055 assert(declared_signature != NULL, "cannot be null");
1056 1056 inputs = declared_signature->arg_size_for_bc(code);
1057 1057 int size = declared_signature->return_type()->size();
1058 1058 depth = size - inputs;
1059 1059 }
1060 1060 break;
1061 1061
1062 1062 case Bytecodes::_multianewarray:
1063 1063 {
1064 1064 ciBytecodeStream iter(method());
1065 1065 iter.reset_to_bci(bci());
1066 1066 iter.next();
1067 1067 inputs = iter.get_dimensions();
1068 1068 assert(rsize == 1, "");
1069 1069 depth = rsize - inputs;
1070 1070 }
1071 1071 break;
1072 1072
1073 1073 case Bytecodes::_ireturn:
1074 1074 case Bytecodes::_lreturn:
1075 1075 case Bytecodes::_freturn:
1076 1076 case Bytecodes::_dreturn:
1077 1077 case Bytecodes::_areturn:
1078 1078 assert(rsize = -depth, "");
1079 1079 inputs = rsize;
1080 1080 break;
1081 1081
1082 1082 case Bytecodes::_jsr:
1083 1083 case Bytecodes::_jsr_w:
1084 1084 inputs = 0;
1085 1085 depth = 1; // S.B. depth=1, not zero
1086 1086 break;
1087 1087
1088 1088 default:
1089 1089 // bytecode produces a typed result
1090 1090 inputs = rsize - depth;
1091 1091 assert(inputs >= 0, "");
1092 1092 break;
1093 1093 }
1094 1094
1095 1095 #ifdef ASSERT
1096 1096 // spot check
1097 1097 int outputs = depth + inputs;
1098 1098 assert(outputs >= 0, "sanity");
1099 1099 switch (code) {
1100 1100 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1101 1101 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1102 1102 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1103 1103 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1104 1104 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1105 1105 }
1106 1106 #endif //ASSERT
1107 1107
1108 1108 return true;
1109 1109 }
1110 1110
1111 1111
1112 1112
1113 1113 //------------------------------basic_plus_adr---------------------------------
1114 1114 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1115 1115 // short-circuit a common case
1116 1116 if (offset == intcon(0)) return ptr;
1117 1117 return _gvn.transform( new (C) AddPNode(base, ptr, offset) );
|
↓ open down ↓ |
1117 lines elided |
↑ open up ↑ |
1118 1118 }
1119 1119
1120 1120 Node* GraphKit::ConvI2L(Node* offset) {
1121 1121 // short-circuit a common case
1122 1122 jint offset_con = find_int_con(offset, Type::OffsetBot);
1123 1123 if (offset_con != Type::OffsetBot) {
1124 1124 return longcon((jlong) offset_con);
1125 1125 }
1126 1126 return _gvn.transform( new (C) ConvI2LNode(offset));
1127 1127 }
1128 +
1129 +Node* GraphKit::ConvI2UL(Node* offset) {
1130 + juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1131 + if (offset_con != (juint) Type::OffsetBot) {
1132 + return longcon((julong) offset_con);
1133 + }
1134 + Node* conv = _gvn.transform( new (C) ConvI2LNode(offset));
1135 + Node* mask = _gvn.transform( ConLNode::make(C, (julong) max_juint) );
1136 + return _gvn.transform( new (C) AndLNode(conv, mask) );
1137 +}
1138 +
1128 1139 Node* GraphKit::ConvL2I(Node* offset) {
1129 1140 // short-circuit a common case
1130 1141 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1131 1142 if (offset_con != (jlong)Type::OffsetBot) {
1132 1143 return intcon((int) offset_con);
1133 1144 }
1134 1145 return _gvn.transform( new (C) ConvL2INode(offset));
1135 1146 }
1136 1147
1137 1148 //-------------------------load_object_klass-----------------------------------
1138 1149 Node* GraphKit::load_object_klass(Node* obj) {
1139 1150 // Special-case a fresh allocation to avoid building nodes:
1140 1151 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1141 1152 if (akls != NULL) return akls;
1142 1153 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1143 1154 return _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS) );
1144 1155 }
1145 1156
1146 1157 //-------------------------load_array_length-----------------------------------
1147 1158 Node* GraphKit::load_array_length(Node* array) {
1148 1159 // Special-case a fresh allocation to avoid building nodes:
1149 1160 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1150 1161 Node *alen;
1151 1162 if (alloc == NULL) {
1152 1163 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1153 1164 alen = _gvn.transform( new (C) LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1154 1165 } else {
1155 1166 alen = alloc->Ideal_length();
1156 1167 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1157 1168 if (ccast != alen) {
1158 1169 alen = _gvn.transform(ccast);
1159 1170 }
1160 1171 }
1161 1172 return alen;
1162 1173 }
1163 1174
1164 1175 //------------------------------do_null_check----------------------------------
1165 1176 // Helper function to do a NULL pointer check. Returned value is
1166 1177 // the incoming address with NULL casted away. You are allowed to use the
1167 1178 // not-null value only if you are control dependent on the test.
1168 1179 extern int explicit_null_checks_inserted,
1169 1180 explicit_null_checks_elided;
1170 1181 Node* GraphKit::null_check_common(Node* value, BasicType type,
1171 1182 // optional arguments for variations:
1172 1183 bool assert_null,
1173 1184 Node* *null_control) {
1174 1185 assert(!assert_null || null_control == NULL, "not both at once");
1175 1186 if (stopped()) return top();
1176 1187 if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) {
1177 1188 // For some performance testing, we may wish to suppress null checking.
1178 1189 value = cast_not_null(value); // Make it appear to be non-null (4962416).
1179 1190 return value;
1180 1191 }
1181 1192 explicit_null_checks_inserted++;
1182 1193
1183 1194 // Construct NULL check
1184 1195 Node *chk = NULL;
1185 1196 switch(type) {
1186 1197 case T_LONG : chk = new (C) CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1187 1198 case T_INT : chk = new (C) CmpINode(value, _gvn.intcon(0)); break;
1188 1199 case T_ARRAY : // fall through
1189 1200 type = T_OBJECT; // simplify further tests
1190 1201 case T_OBJECT : {
1191 1202 const Type *t = _gvn.type( value );
1192 1203
1193 1204 const TypeOopPtr* tp = t->isa_oopptr();
1194 1205 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1195 1206 // Only for do_null_check, not any of its siblings:
1196 1207 && !assert_null && null_control == NULL) {
1197 1208 // Usually, any field access or invocation on an unloaded oop type
1198 1209 // will simply fail to link, since the statically linked class is
1199 1210 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1200 1211 // the static class is loaded but the sharper oop type is not.
1201 1212 // Rather than checking for this obscure case in lots of places,
1202 1213 // we simply observe that a null check on an unloaded class
1203 1214 // will always be followed by a nonsense operation, so we
1204 1215 // can just issue the uncommon trap here.
1205 1216 // Our access to the unloaded class will only be correct
1206 1217 // after it has been loaded and initialized, which requires
1207 1218 // a trip through the interpreter.
1208 1219 #ifndef PRODUCT
1209 1220 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1210 1221 #endif
1211 1222 uncommon_trap(Deoptimization::Reason_unloaded,
1212 1223 Deoptimization::Action_reinterpret,
1213 1224 tp->klass(), "!loaded");
1214 1225 return top();
1215 1226 }
1216 1227
1217 1228 if (assert_null) {
1218 1229 // See if the type is contained in NULL_PTR.
1219 1230 // If so, then the value is already null.
1220 1231 if (t->higher_equal(TypePtr::NULL_PTR)) {
1221 1232 explicit_null_checks_elided++;
1222 1233 return value; // Elided null assert quickly!
1223 1234 }
1224 1235 } else {
1225 1236 // See if mixing in the NULL pointer changes type.
1226 1237 // If so, then the NULL pointer was not allowed in the original
1227 1238 // type. In other words, "value" was not-null.
1228 1239 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1229 1240 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1230 1241 explicit_null_checks_elided++;
1231 1242 return value; // Elided null check quickly!
1232 1243 }
1233 1244 }
1234 1245 chk = new (C) CmpPNode( value, null() );
1235 1246 break;
1236 1247 }
1237 1248
1238 1249 default:
1239 1250 fatal(err_msg_res("unexpected type: %s", type2name(type)));
1240 1251 }
1241 1252 assert(chk != NULL, "sanity check");
1242 1253 chk = _gvn.transform(chk);
1243 1254
1244 1255 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1245 1256 BoolNode *btst = new (C) BoolNode( chk, btest);
1246 1257 Node *tst = _gvn.transform( btst );
1247 1258
1248 1259 //-----------
1249 1260 // if peephole optimizations occurred, a prior test existed.
1250 1261 // If a prior test existed, maybe it dominates as we can avoid this test.
1251 1262 if (tst != btst && type == T_OBJECT) {
1252 1263 // At this point we want to scan up the CFG to see if we can
1253 1264 // find an identical test (and so avoid this test altogether).
1254 1265 Node *cfg = control();
1255 1266 int depth = 0;
1256 1267 while( depth < 16 ) { // Limit search depth for speed
1257 1268 if( cfg->Opcode() == Op_IfTrue &&
1258 1269 cfg->in(0)->in(1) == tst ) {
1259 1270 // Found prior test. Use "cast_not_null" to construct an identical
1260 1271 // CastPP (and hence hash to) as already exists for the prior test.
1261 1272 // Return that casted value.
1262 1273 if (assert_null) {
1263 1274 replace_in_map(value, null());
1264 1275 return null(); // do not issue the redundant test
1265 1276 }
1266 1277 Node *oldcontrol = control();
1267 1278 set_control(cfg);
1268 1279 Node *res = cast_not_null(value);
1269 1280 set_control(oldcontrol);
1270 1281 explicit_null_checks_elided++;
1271 1282 return res;
1272 1283 }
1273 1284 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1274 1285 if (cfg == NULL) break; // Quit at region nodes
1275 1286 depth++;
1276 1287 }
1277 1288 }
1278 1289
1279 1290 //-----------
1280 1291 // Branch to failure if null
1281 1292 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1282 1293 Deoptimization::DeoptReason reason;
1283 1294 if (assert_null)
1284 1295 reason = Deoptimization::Reason_null_assert;
1285 1296 else if (type == T_OBJECT)
1286 1297 reason = Deoptimization::Reason_null_check;
1287 1298 else
1288 1299 reason = Deoptimization::Reason_div0_check;
1289 1300
1290 1301 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1291 1302 // ciMethodData::has_trap_at will return a conservative -1 if any
1292 1303 // must-be-null assertion has failed. This could cause performance
1293 1304 // problems for a method after its first do_null_assert failure.
1294 1305 // Consider using 'Reason_class_check' instead?
1295 1306
1296 1307 // To cause an implicit null check, we set the not-null probability
1297 1308 // to the maximum (PROB_MAX). For an explicit check the probability
1298 1309 // is set to a smaller value.
1299 1310 if (null_control != NULL || too_many_traps(reason)) {
1300 1311 // probability is less likely
1301 1312 ok_prob = PROB_LIKELY_MAG(3);
1302 1313 } else if (!assert_null &&
1303 1314 (ImplicitNullCheckThreshold > 0) &&
1304 1315 method() != NULL &&
1305 1316 (method()->method_data()->trap_count(reason)
1306 1317 >= (uint)ImplicitNullCheckThreshold)) {
1307 1318 ok_prob = PROB_LIKELY_MAG(3);
1308 1319 }
1309 1320
1310 1321 if (null_control != NULL) {
1311 1322 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1312 1323 Node* null_true = _gvn.transform( new (C) IfFalseNode(iff));
1313 1324 set_control( _gvn.transform( new (C) IfTrueNode(iff)));
1314 1325 if (null_true == top())
1315 1326 explicit_null_checks_elided++;
1316 1327 (*null_control) = null_true;
1317 1328 } else {
1318 1329 BuildCutout unless(this, tst, ok_prob);
1319 1330 // Check for optimizer eliding test at parse time
1320 1331 if (stopped()) {
1321 1332 // Failure not possible; do not bother making uncommon trap.
1322 1333 explicit_null_checks_elided++;
1323 1334 } else if (assert_null) {
1324 1335 uncommon_trap(reason,
1325 1336 Deoptimization::Action_make_not_entrant,
1326 1337 NULL, "assert_null");
1327 1338 } else {
1328 1339 replace_in_map(value, zerocon(type));
1329 1340 builtin_throw(reason);
1330 1341 }
1331 1342 }
1332 1343
1333 1344 // Must throw exception, fall-thru not possible?
1334 1345 if (stopped()) {
1335 1346 return top(); // No result
1336 1347 }
1337 1348
1338 1349 if (assert_null) {
1339 1350 // Cast obj to null on this path.
1340 1351 replace_in_map(value, zerocon(type));
1341 1352 return zerocon(type);
1342 1353 }
1343 1354
1344 1355 // Cast obj to not-null on this path, if there is no null_control.
1345 1356 // (If there is a null_control, a non-null value may come back to haunt us.)
1346 1357 if (type == T_OBJECT) {
1347 1358 Node* cast = cast_not_null(value, false);
1348 1359 if (null_control == NULL || (*null_control) == top())
1349 1360 replace_in_map(value, cast);
1350 1361 value = cast;
1351 1362 }
1352 1363
1353 1364 return value;
1354 1365 }
1355 1366
1356 1367
1357 1368 //------------------------------cast_not_null----------------------------------
1358 1369 // Cast obj to not-null on this path
1359 1370 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1360 1371 const Type *t = _gvn.type(obj);
1361 1372 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1362 1373 // Object is already not-null?
1363 1374 if( t == t_not_null ) return obj;
1364 1375
1365 1376 Node *cast = new (C) CastPPNode(obj,t_not_null);
1366 1377 cast->init_req(0, control());
1367 1378 cast = _gvn.transform( cast );
1368 1379
1369 1380 // Scan for instances of 'obj' in the current JVM mapping.
1370 1381 // These instances are known to be not-null after the test.
1371 1382 if (do_replace_in_map)
1372 1383 replace_in_map(obj, cast);
1373 1384
1374 1385 return cast; // Return casted value
1375 1386 }
1376 1387
1377 1388
1378 1389 //--------------------------replace_in_map-------------------------------------
1379 1390 void GraphKit::replace_in_map(Node* old, Node* neww) {
1380 1391 if (old == neww) {
1381 1392 return;
1382 1393 }
1383 1394
1384 1395 map()->replace_edge(old, neww);
1385 1396
1386 1397 // Note: This operation potentially replaces any edge
1387 1398 // on the map. This includes locals, stack, and monitors
1388 1399 // of the current (innermost) JVM state.
1389 1400
1390 1401 if (!ReplaceInParentMaps) {
1391 1402 return;
1392 1403 }
1393 1404
1394 1405 // PreserveJVMState doesn't do a deep copy so we can't modify
1395 1406 // parents
1396 1407 if (Compile::current()->has_preserve_jvm_state()) {
1397 1408 return;
1398 1409 }
1399 1410
1400 1411 Parse* parser = is_Parse();
1401 1412 bool progress = true;
1402 1413 Node* ctrl = map()->in(0);
1403 1414 // Follow the chain of parsers and see whether the update can be
1404 1415 // done in the map of callers. We can do the replace for a caller if
1405 1416 // the current control post dominates the control of a caller.
1406 1417 while (parser != NULL && parser->caller() != NULL && progress) {
1407 1418 progress = false;
1408 1419 Node* parent_map = parser->caller()->map();
1409 1420 assert(parser->exits().map()->jvms()->depth() == parser->caller()->depth(), "map mismatch");
1410 1421
1411 1422 Node* parent_ctrl = parent_map->in(0);
1412 1423
1413 1424 while (parent_ctrl->is_Region()) {
1414 1425 Node* n = parent_ctrl->as_Region()->is_copy();
1415 1426 if (n == NULL) {
1416 1427 break;
1417 1428 }
1418 1429 parent_ctrl = n;
1419 1430 }
1420 1431
1421 1432 for (;;) {
1422 1433 if (ctrl == parent_ctrl) {
1423 1434 // update the map of the exits which is the one that will be
1424 1435 // used when compilation resume after inlining
1425 1436 parser->exits().map()->replace_edge(old, neww);
1426 1437 progress = true;
1427 1438 break;
1428 1439 }
1429 1440 if (ctrl->is_Proj() && ctrl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) {
1430 1441 ctrl = ctrl->in(0)->in(0);
1431 1442 } else if (ctrl->is_Region()) {
1432 1443 Node* n = ctrl->as_Region()->is_copy();
1433 1444 if (n == NULL) {
1434 1445 break;
1435 1446 }
1436 1447 ctrl = n;
1437 1448 } else {
1438 1449 break;
1439 1450 }
1440 1451 }
1441 1452
1442 1453 parser = parser->parent_parser();
1443 1454 }
1444 1455 }
1445 1456
1446 1457
1447 1458 //=============================================================================
1448 1459 //--------------------------------memory---------------------------------------
1449 1460 Node* GraphKit::memory(uint alias_idx) {
1450 1461 MergeMemNode* mem = merged_memory();
1451 1462 Node* p = mem->memory_at(alias_idx);
1452 1463 _gvn.set_type(p, Type::MEMORY); // must be mapped
1453 1464 return p;
1454 1465 }
1455 1466
1456 1467 //-----------------------------reset_memory------------------------------------
1457 1468 Node* GraphKit::reset_memory() {
1458 1469 Node* mem = map()->memory();
1459 1470 // do not use this node for any more parsing!
1460 1471 debug_only( map()->set_memory((Node*)NULL) );
1461 1472 return _gvn.transform( mem );
1462 1473 }
1463 1474
1464 1475 //------------------------------set_all_memory---------------------------------
1465 1476 void GraphKit::set_all_memory(Node* newmem) {
1466 1477 Node* mergemem = MergeMemNode::make(C, newmem);
1467 1478 gvn().set_type_bottom(mergemem);
1468 1479 map()->set_memory(mergemem);
1469 1480 }
1470 1481
1471 1482 //------------------------------set_all_memory_call----------------------------
1472 1483 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1473 1484 Node* newmem = _gvn.transform( new (C) ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1474 1485 set_all_memory(newmem);
1475 1486 }
1476 1487
1477 1488 //=============================================================================
1478 1489 //
1479 1490 // parser factory methods for MemNodes
1480 1491 //
1481 1492 // These are layered on top of the factory methods in LoadNode and StoreNode,
1482 1493 // and integrate with the parser's memory state and _gvn engine.
1483 1494 //
1484 1495
1485 1496 // factory methods in "int adr_idx"
1486 1497 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1487 1498 int adr_idx,
1488 1499 MemNode::MemOrd mo, bool require_atomic_access) {
1489 1500 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1490 1501 const TypePtr* adr_type = NULL; // debug-mode-only argument
1491 1502 debug_only(adr_type = C->get_adr_type(adr_idx));
1492 1503 Node* mem = memory(adr_idx);
1493 1504 Node* ld;
1494 1505 if (require_atomic_access && bt == T_LONG) {
1495 1506 ld = LoadLNode::make_atomic(C, ctl, mem, adr, adr_type, t, mo);
1496 1507 } else {
1497 1508 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo);
1498 1509 }
1499 1510 ld = _gvn.transform(ld);
1500 1511 if ((bt == T_OBJECT) && C->do_escape_analysis() || C->eliminate_boxing()) {
1501 1512 // Improve graph before escape analysis and boxing elimination.
1502 1513 record_for_igvn(ld);
1503 1514 }
1504 1515 return ld;
1505 1516 }
1506 1517
1507 1518 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1508 1519 int adr_idx,
1509 1520 MemNode::MemOrd mo,
1510 1521 bool require_atomic_access) {
1511 1522 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1512 1523 const TypePtr* adr_type = NULL;
1513 1524 debug_only(adr_type = C->get_adr_type(adr_idx));
1514 1525 Node *mem = memory(adr_idx);
1515 1526 Node* st;
1516 1527 if (require_atomic_access && bt == T_LONG) {
1517 1528 st = StoreLNode::make_atomic(C, ctl, mem, adr, adr_type, val, mo);
1518 1529 } else {
1519 1530 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo);
1520 1531 }
1521 1532 st = _gvn.transform(st);
1522 1533 set_memory(st, adr_idx);
1523 1534 // Back-to-back stores can only remove intermediate store with DU info
1524 1535 // so push on worklist for optimizer.
1525 1536 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1526 1537 record_for_igvn(st);
1527 1538
1528 1539 return st;
1529 1540 }
1530 1541
1531 1542
1532 1543 void GraphKit::pre_barrier(bool do_load,
1533 1544 Node* ctl,
1534 1545 Node* obj,
1535 1546 Node* adr,
1536 1547 uint adr_idx,
1537 1548 Node* val,
1538 1549 const TypeOopPtr* val_type,
1539 1550 Node* pre_val,
1540 1551 BasicType bt) {
1541 1552
1542 1553 BarrierSet* bs = Universe::heap()->barrier_set();
1543 1554 set_control(ctl);
1544 1555 switch (bs->kind()) {
1545 1556 case BarrierSet::G1SATBCT:
1546 1557 case BarrierSet::G1SATBCTLogging:
1547 1558 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt);
1548 1559 break;
1549 1560
1550 1561 case BarrierSet::CardTableModRef:
1551 1562 case BarrierSet::CardTableExtension:
1552 1563 case BarrierSet::ModRef:
1553 1564 break;
1554 1565
1555 1566 case BarrierSet::Other:
1556 1567 default :
1557 1568 ShouldNotReachHere();
1558 1569
1559 1570 }
1560 1571 }
1561 1572
1562 1573 bool GraphKit::can_move_pre_barrier() const {
1563 1574 BarrierSet* bs = Universe::heap()->barrier_set();
1564 1575 switch (bs->kind()) {
1565 1576 case BarrierSet::G1SATBCT:
1566 1577 case BarrierSet::G1SATBCTLogging:
1567 1578 return true; // Can move it if no safepoint
1568 1579
1569 1580 case BarrierSet::CardTableModRef:
1570 1581 case BarrierSet::CardTableExtension:
1571 1582 case BarrierSet::ModRef:
1572 1583 return true; // There is no pre-barrier
1573 1584
1574 1585 case BarrierSet::Other:
1575 1586 default :
1576 1587 ShouldNotReachHere();
1577 1588 }
1578 1589 return false;
1579 1590 }
1580 1591
1581 1592 void GraphKit::post_barrier(Node* ctl,
1582 1593 Node* store,
1583 1594 Node* obj,
1584 1595 Node* adr,
1585 1596 uint adr_idx,
1586 1597 Node* val,
1587 1598 BasicType bt,
1588 1599 bool use_precise) {
1589 1600 BarrierSet* bs = Universe::heap()->barrier_set();
1590 1601 set_control(ctl);
1591 1602 switch (bs->kind()) {
1592 1603 case BarrierSet::G1SATBCT:
1593 1604 case BarrierSet::G1SATBCTLogging:
1594 1605 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise);
1595 1606 break;
1596 1607
1597 1608 case BarrierSet::CardTableModRef:
1598 1609 case BarrierSet::CardTableExtension:
1599 1610 write_barrier_post(store, obj, adr, adr_idx, val, use_precise);
1600 1611 break;
1601 1612
1602 1613 case BarrierSet::ModRef:
1603 1614 break;
1604 1615
1605 1616 case BarrierSet::Other:
1606 1617 default :
1607 1618 ShouldNotReachHere();
1608 1619
1609 1620 }
1610 1621 }
1611 1622
1612 1623 Node* GraphKit::store_oop(Node* ctl,
1613 1624 Node* obj,
1614 1625 Node* adr,
1615 1626 const TypePtr* adr_type,
1616 1627 Node* val,
1617 1628 const TypeOopPtr* val_type,
1618 1629 BasicType bt,
1619 1630 bool use_precise,
1620 1631 MemNode::MemOrd mo) {
1621 1632 // Transformation of a value which could be NULL pointer (CastPP #NULL)
1622 1633 // could be delayed during Parse (for example, in adjust_map_after_if()).
1623 1634 // Execute transformation here to avoid barrier generation in such case.
1624 1635 if (_gvn.type(val) == TypePtr::NULL_PTR)
1625 1636 val = _gvn.makecon(TypePtr::NULL_PTR);
1626 1637
1627 1638 set_control(ctl);
1628 1639 if (stopped()) return top(); // Dead path ?
1629 1640
1630 1641 assert(bt == T_OBJECT, "sanity");
1631 1642 assert(val != NULL, "not dead path");
1632 1643 uint adr_idx = C->get_alias_index(adr_type);
1633 1644 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1634 1645
1635 1646 pre_barrier(true /* do_load */,
1636 1647 control(), obj, adr, adr_idx, val, val_type,
1637 1648 NULL /* pre_val */,
1638 1649 bt);
1639 1650
1640 1651 Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo);
1641 1652 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise);
1642 1653 return store;
1643 1654 }
1644 1655
1645 1656 // Could be an array or object we don't know at compile time (unsafe ref.)
1646 1657 Node* GraphKit::store_oop_to_unknown(Node* ctl,
1647 1658 Node* obj, // containing obj
1648 1659 Node* adr, // actual adress to store val at
1649 1660 const TypePtr* adr_type,
1650 1661 Node* val,
1651 1662 BasicType bt,
1652 1663 MemNode::MemOrd mo) {
1653 1664 Compile::AliasType* at = C->alias_type(adr_type);
1654 1665 const TypeOopPtr* val_type = NULL;
1655 1666 if (adr_type->isa_instptr()) {
1656 1667 if (at->field() != NULL) {
1657 1668 // known field. This code is a copy of the do_put_xxx logic.
1658 1669 ciField* field = at->field();
1659 1670 if (!field->type()->is_loaded()) {
1660 1671 val_type = TypeInstPtr::BOTTOM;
1661 1672 } else {
1662 1673 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass());
1663 1674 }
1664 1675 }
1665 1676 } else if (adr_type->isa_aryptr()) {
1666 1677 val_type = adr_type->is_aryptr()->elem()->make_oopptr();
1667 1678 }
1668 1679 if (val_type == NULL) {
1669 1680 val_type = TypeInstPtr::BOTTOM;
1670 1681 }
1671 1682 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo);
1672 1683 }
1673 1684
1674 1685
1675 1686 //-------------------------array_element_address-------------------------
1676 1687 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1677 1688 const TypeInt* sizetype) {
1678 1689 uint shift = exact_log2(type2aelembytes(elembt));
1679 1690 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1680 1691
1681 1692 // short-circuit a common case (saves lots of confusing waste motion)
1682 1693 jint idx_con = find_int_con(idx, -1);
1683 1694 if (idx_con >= 0) {
1684 1695 intptr_t offset = header + ((intptr_t)idx_con << shift);
1685 1696 return basic_plus_adr(ary, offset);
1686 1697 }
1687 1698
1688 1699 // must be correct type for alignment purposes
1689 1700 Node* base = basic_plus_adr(ary, header);
1690 1701 #ifdef _LP64
1691 1702 // The scaled index operand to AddP must be a clean 64-bit value.
1692 1703 // Java allows a 32-bit int to be incremented to a negative
1693 1704 // value, which appears in a 64-bit register as a large
1694 1705 // positive number. Using that large positive number as an
1695 1706 // operand in pointer arithmetic has bad consequences.
1696 1707 // On the other hand, 32-bit overflow is rare, and the possibility
1697 1708 // can often be excluded, if we annotate the ConvI2L node with
1698 1709 // a type assertion that its value is known to be a small positive
1699 1710 // number. (The prior range check has ensured this.)
1700 1711 // This assertion is used by ConvI2LNode::Ideal.
1701 1712 int index_max = max_jint - 1; // array size is max_jint, index is one less
1702 1713 if (sizetype != NULL) index_max = sizetype->_hi - 1;
1703 1714 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax);
1704 1715 idx = _gvn.transform( new (C) ConvI2LNode(idx, lidxtype) );
1705 1716 #endif
1706 1717 Node* scale = _gvn.transform( new (C) LShiftXNode(idx, intcon(shift)) );
1707 1718 return basic_plus_adr(ary, base, scale);
1708 1719 }
1709 1720
1710 1721 //-------------------------load_array_element-------------------------
1711 1722 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1712 1723 const Type* elemtype = arytype->elem();
1713 1724 BasicType elembt = elemtype->array_element_basic_type();
1714 1725 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1715 1726 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered);
1716 1727 return ld;
1717 1728 }
1718 1729
1719 1730 //-------------------------set_arguments_for_java_call-------------------------
1720 1731 // Arguments (pre-popped from the stack) are taken from the JVMS.
1721 1732 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1722 1733 // Add the call arguments:
1723 1734 uint nargs = call->method()->arg_size();
1724 1735 for (uint i = 0; i < nargs; i++) {
1725 1736 Node* arg = argument(i);
1726 1737 call->init_req(i + TypeFunc::Parms, arg);
1727 1738 }
1728 1739 }
1729 1740
1730 1741 //---------------------------set_edges_for_java_call---------------------------
1731 1742 // Connect a newly created call into the current JVMS.
1732 1743 // A return value node (if any) is returned from set_edges_for_java_call.
1733 1744 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1734 1745
1735 1746 // Add the predefined inputs:
1736 1747 call->init_req( TypeFunc::Control, control() );
1737 1748 call->init_req( TypeFunc::I_O , i_o() );
1738 1749 call->init_req( TypeFunc::Memory , reset_memory() );
1739 1750 call->init_req( TypeFunc::FramePtr, frameptr() );
1740 1751 call->init_req( TypeFunc::ReturnAdr, top() );
1741 1752
1742 1753 add_safepoint_edges(call, must_throw);
1743 1754
1744 1755 Node* xcall = _gvn.transform(call);
1745 1756
1746 1757 if (xcall == top()) {
1747 1758 set_control(top());
1748 1759 return;
1749 1760 }
1750 1761 assert(xcall == call, "call identity is stable");
1751 1762
1752 1763 // Re-use the current map to produce the result.
1753 1764
1754 1765 set_control(_gvn.transform(new (C) ProjNode(call, TypeFunc::Control)));
1755 1766 set_i_o( _gvn.transform(new (C) ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1756 1767 set_all_memory_call(xcall, separate_io_proj);
1757 1768
1758 1769 //return xcall; // no need, caller already has it
1759 1770 }
1760 1771
1761 1772 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) {
1762 1773 if (stopped()) return top(); // maybe the call folded up?
1763 1774
1764 1775 // Capture the return value, if any.
1765 1776 Node* ret;
1766 1777 if (call->method() == NULL ||
1767 1778 call->method()->return_type()->basic_type() == T_VOID)
1768 1779 ret = top();
1769 1780 else ret = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
1770 1781
1771 1782 // Note: Since any out-of-line call can produce an exception,
1772 1783 // we always insert an I_O projection from the call into the result.
1773 1784
1774 1785 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj);
1775 1786
1776 1787 if (separate_io_proj) {
1777 1788 // The caller requested separate projections be used by the fall
1778 1789 // through and exceptional paths, so replace the projections for
1779 1790 // the fall through path.
1780 1791 set_i_o(_gvn.transform( new (C) ProjNode(call, TypeFunc::I_O) ));
1781 1792 set_all_memory(_gvn.transform( new (C) ProjNode(call, TypeFunc::Memory) ));
1782 1793 }
1783 1794 return ret;
1784 1795 }
1785 1796
1786 1797 //--------------------set_predefined_input_for_runtime_call--------------------
1787 1798 // Reading and setting the memory state is way conservative here.
1788 1799 // The real problem is that I am not doing real Type analysis on memory,
1789 1800 // so I cannot distinguish card mark stores from other stores. Across a GC
1790 1801 // point the Store Barrier and the card mark memory has to agree. I cannot
1791 1802 // have a card mark store and its barrier split across the GC point from
1792 1803 // either above or below. Here I get that to happen by reading ALL of memory.
1793 1804 // A better answer would be to separate out card marks from other memory.
1794 1805 // For now, return the input memory state, so that it can be reused
1795 1806 // after the call, if this call has restricted memory effects.
1796 1807 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) {
1797 1808 // Set fixed predefined input arguments
1798 1809 Node* memory = reset_memory();
1799 1810 call->init_req( TypeFunc::Control, control() );
1800 1811 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1801 1812 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs
1802 1813 call->init_req( TypeFunc::FramePtr, frameptr() );
1803 1814 call->init_req( TypeFunc::ReturnAdr, top() );
1804 1815 return memory;
1805 1816 }
1806 1817
1807 1818 //-------------------set_predefined_output_for_runtime_call--------------------
1808 1819 // Set control and memory (not i_o) from the call.
1809 1820 // If keep_mem is not NULL, use it for the output state,
1810 1821 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1811 1822 // If hook_mem is NULL, this call produces no memory effects at all.
1812 1823 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1813 1824 // then only that memory slice is taken from the call.
1814 1825 // In the last case, we must put an appropriate memory barrier before
1815 1826 // the call, so as to create the correct anti-dependencies on loads
1816 1827 // preceding the call.
1817 1828 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1818 1829 Node* keep_mem,
1819 1830 const TypePtr* hook_mem) {
1820 1831 // no i/o
1821 1832 set_control(_gvn.transform( new (C) ProjNode(call,TypeFunc::Control) ));
1822 1833 if (keep_mem) {
1823 1834 // First clone the existing memory state
1824 1835 set_all_memory(keep_mem);
1825 1836 if (hook_mem != NULL) {
1826 1837 // Make memory for the call
1827 1838 Node* mem = _gvn.transform( new (C) ProjNode(call, TypeFunc::Memory) );
1828 1839 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
1829 1840 // We also use hook_mem to extract specific effects from arraycopy stubs.
1830 1841 set_memory(mem, hook_mem);
1831 1842 }
1832 1843 // ...else the call has NO memory effects.
1833 1844
1834 1845 // Make sure the call advertises its memory effects precisely.
1835 1846 // This lets us build accurate anti-dependences in gcm.cpp.
1836 1847 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1837 1848 "call node must be constructed correctly");
1838 1849 } else {
1839 1850 assert(hook_mem == NULL, "");
1840 1851 // This is not a "slow path" call; all memory comes from the call.
1841 1852 set_all_memory_call(call);
1842 1853 }
1843 1854 }
1844 1855
1845 1856
1846 1857 // Replace the call with the current state of the kit.
1847 1858 void GraphKit::replace_call(CallNode* call, Node* result) {
1848 1859 JVMState* ejvms = NULL;
1849 1860 if (has_exceptions()) {
1850 1861 ejvms = transfer_exceptions_into_jvms();
1851 1862 }
1852 1863
1853 1864 SafePointNode* final_state = stop();
1854 1865
1855 1866 // Find all the needed outputs of this call
1856 1867 CallProjections callprojs;
1857 1868 call->extract_projections(&callprojs, true);
1858 1869
1859 1870 Node* init_mem = call->in(TypeFunc::Memory);
1860 1871 Node* final_mem = final_state->in(TypeFunc::Memory);
1861 1872 Node* final_ctl = final_state->in(TypeFunc::Control);
1862 1873 Node* final_io = final_state->in(TypeFunc::I_O);
1863 1874
1864 1875 // Replace all the old call edges with the edges from the inlining result
1865 1876 if (callprojs.fallthrough_catchproj != NULL) {
1866 1877 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
1867 1878 }
1868 1879 if (callprojs.fallthrough_memproj != NULL) {
1869 1880 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem);
1870 1881 }
1871 1882 if (callprojs.fallthrough_ioproj != NULL) {
1872 1883 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io);
1873 1884 }
1874 1885
1875 1886 // Replace the result with the new result if it exists and is used
1876 1887 if (callprojs.resproj != NULL && result != NULL) {
1877 1888 C->gvn_replace_by(callprojs.resproj, result);
1878 1889 }
1879 1890
1880 1891 if (ejvms == NULL) {
1881 1892 // No exception edges to simply kill off those paths
1882 1893 if (callprojs.catchall_catchproj != NULL) {
1883 1894 C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1884 1895 }
1885 1896 if (callprojs.catchall_memproj != NULL) {
1886 1897 C->gvn_replace_by(callprojs.catchall_memproj, C->top());
1887 1898 }
1888 1899 if (callprojs.catchall_ioproj != NULL) {
1889 1900 C->gvn_replace_by(callprojs.catchall_ioproj, C->top());
1890 1901 }
1891 1902 // Replace the old exception object with top
1892 1903 if (callprojs.exobj != NULL) {
1893 1904 C->gvn_replace_by(callprojs.exobj, C->top());
1894 1905 }
1895 1906 } else {
1896 1907 GraphKit ekit(ejvms);
1897 1908
1898 1909 // Load my combined exception state into the kit, with all phis transformed:
1899 1910 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1900 1911
1901 1912 Node* ex_oop = ekit.use_exception_state(ex_map);
1902 1913 if (callprojs.catchall_catchproj != NULL) {
1903 1914 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
1904 1915 }
1905 1916 if (callprojs.catchall_memproj != NULL) {
1906 1917 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory());
1907 1918 }
1908 1919 if (callprojs.catchall_ioproj != NULL) {
1909 1920 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o());
1910 1921 }
1911 1922
1912 1923 // Replace the old exception object with the newly created one
1913 1924 if (callprojs.exobj != NULL) {
1914 1925 C->gvn_replace_by(callprojs.exobj, ex_oop);
1915 1926 }
1916 1927 }
1917 1928
1918 1929 // Disconnect the call from the graph
1919 1930 call->disconnect_inputs(NULL, C);
1920 1931 C->gvn_replace_by(call, C->top());
1921 1932
1922 1933 // Clean up any MergeMems that feed other MergeMems since the
1923 1934 // optimizer doesn't like that.
1924 1935 if (final_mem->is_MergeMem()) {
1925 1936 Node_List wl;
1926 1937 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) {
1927 1938 Node* m = i.get();
1928 1939 if (m->is_MergeMem() && !wl.contains(m)) {
1929 1940 wl.push(m);
1930 1941 }
1931 1942 }
1932 1943 while (wl.size() > 0) {
1933 1944 _gvn.transform(wl.pop());
1934 1945 }
1935 1946 }
1936 1947 }
1937 1948
1938 1949
1939 1950 //------------------------------increment_counter------------------------------
1940 1951 // for statistics: increment a VM counter by 1
1941 1952
1942 1953 void GraphKit::increment_counter(address counter_addr) {
1943 1954 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
1944 1955 increment_counter(adr1);
1945 1956 }
1946 1957
1947 1958 void GraphKit::increment_counter(Node* counter_addr) {
1948 1959 int adr_type = Compile::AliasIdxRaw;
1949 1960 Node* ctrl = control();
1950 1961 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
1951 1962 Node* incr = _gvn.transform(new (C) AddINode(cnt, _gvn.intcon(1)));
1952 1963 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered);
1953 1964 }
1954 1965
1955 1966
1956 1967 //------------------------------uncommon_trap----------------------------------
1957 1968 // Bail out to the interpreter in mid-method. Implemented by calling the
1958 1969 // uncommon_trap blob. This helper function inserts a runtime call with the
1959 1970 // right debug info.
1960 1971 void GraphKit::uncommon_trap(int trap_request,
1961 1972 ciKlass* klass, const char* comment,
1962 1973 bool must_throw,
1963 1974 bool keep_exact_action) {
1964 1975 if (failing()) stop();
1965 1976 if (stopped()) return; // trap reachable?
1966 1977
1967 1978 // Note: If ProfileTraps is true, and if a deopt. actually
1968 1979 // occurs here, the runtime will make sure an MDO exists. There is
1969 1980 // no need to call method()->ensure_method_data() at this point.
1970 1981
1971 1982 // Set the stack pointer to the right value for reexecution:
1972 1983 set_sp(reexecute_sp());
1973 1984
1974 1985 #ifdef ASSERT
1975 1986 if (!must_throw) {
1976 1987 // Make sure the stack has at least enough depth to execute
1977 1988 // the current bytecode.
1978 1989 int inputs, ignored_depth;
1979 1990 if (compute_stack_effects(inputs, ignored_depth)) {
1980 1991 assert(sp() >= inputs, err_msg_res("must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
1981 1992 Bytecodes::name(java_bc()), sp(), inputs));
1982 1993 }
1983 1994 }
1984 1995 #endif
1985 1996
1986 1997 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
1987 1998 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
1988 1999
1989 2000 switch (action) {
1990 2001 case Deoptimization::Action_maybe_recompile:
1991 2002 case Deoptimization::Action_reinterpret:
1992 2003 // Temporary fix for 6529811 to allow virtual calls to be sure they
1993 2004 // get the chance to go from mono->bi->mega
1994 2005 if (!keep_exact_action &&
1995 2006 Deoptimization::trap_request_index(trap_request) < 0 &&
1996 2007 too_many_recompiles(reason)) {
1997 2008 // This BCI is causing too many recompilations.
1998 2009 action = Deoptimization::Action_none;
1999 2010 trap_request = Deoptimization::make_trap_request(reason, action);
2000 2011 } else {
2001 2012 C->set_trap_can_recompile(true);
2002 2013 }
2003 2014 break;
2004 2015 case Deoptimization::Action_make_not_entrant:
2005 2016 C->set_trap_can_recompile(true);
2006 2017 break;
2007 2018 #ifdef ASSERT
2008 2019 case Deoptimization::Action_none:
2009 2020 case Deoptimization::Action_make_not_compilable:
2010 2021 break;
2011 2022 default:
2012 2023 fatal(err_msg_res("unknown action %d: %s", action, Deoptimization::trap_action_name(action)));
2013 2024 break;
2014 2025 #endif
2015 2026 }
2016 2027
2017 2028 if (TraceOptoParse) {
2018 2029 char buf[100];
2019 2030 tty->print_cr("Uncommon trap %s at bci:%d",
2020 2031 Deoptimization::format_trap_request(buf, sizeof(buf),
2021 2032 trap_request), bci());
2022 2033 }
2023 2034
2024 2035 CompileLog* log = C->log();
2025 2036 if (log != NULL) {
2026 2037 int kid = (klass == NULL)? -1: log->identify(klass);
2027 2038 log->begin_elem("uncommon_trap bci='%d'", bci());
2028 2039 char buf[100];
2029 2040 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
2030 2041 trap_request));
2031 2042 if (kid >= 0) log->print(" klass='%d'", kid);
2032 2043 if (comment != NULL) log->print(" comment='%s'", comment);
2033 2044 log->end_elem();
2034 2045 }
2035 2046
2036 2047 // Make sure any guarding test views this path as very unlikely
2037 2048 Node *i0 = control()->in(0);
2038 2049 if (i0 != NULL && i0->is_If()) { // Found a guarding if test?
2039 2050 IfNode *iff = i0->as_If();
2040 2051 float f = iff->_prob; // Get prob
2041 2052 if (control()->Opcode() == Op_IfTrue) {
2042 2053 if (f > PROB_UNLIKELY_MAG(4))
2043 2054 iff->_prob = PROB_MIN;
2044 2055 } else {
2045 2056 if (f < PROB_LIKELY_MAG(4))
2046 2057 iff->_prob = PROB_MAX;
2047 2058 }
2048 2059 }
2049 2060
2050 2061 // Clear out dead values from the debug info.
2051 2062 kill_dead_locals();
2052 2063
2053 2064 // Now insert the uncommon trap subroutine call
2054 2065 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
2055 2066 const TypePtr* no_memory_effects = NULL;
2056 2067 // Pass the index of the class to be loaded
2057 2068 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2058 2069 (must_throw ? RC_MUST_THROW : 0),
2059 2070 OptoRuntime::uncommon_trap_Type(),
2060 2071 call_addr, "uncommon_trap", no_memory_effects,
2061 2072 intcon(trap_request));
2062 2073 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2063 2074 "must extract request correctly from the graph");
2064 2075 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2065 2076
2066 2077 call->set_req(TypeFunc::ReturnAdr, returnadr());
2067 2078 // The debug info is the only real input to this call.
2068 2079
2069 2080 // Halt-and-catch fire here. The above call should never return!
2070 2081 HaltNode* halt = new(C) HaltNode(control(), frameptr());
2071 2082 _gvn.set_type_bottom(halt);
2072 2083 root()->add_req(halt);
2073 2084
2074 2085 stop_and_kill_map();
2075 2086 }
2076 2087
2077 2088
2078 2089 //--------------------------just_allocated_object------------------------------
2079 2090 // Report the object that was just allocated.
2080 2091 // It must be the case that there are no intervening safepoints.
2081 2092 // We use this to determine if an object is so "fresh" that
2082 2093 // it does not require card marks.
2083 2094 Node* GraphKit::just_allocated_object(Node* current_control) {
2084 2095 if (C->recent_alloc_ctl() == current_control)
2085 2096 return C->recent_alloc_obj();
2086 2097 return NULL;
2087 2098 }
2088 2099
2089 2100
2090 2101 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2091 2102 // (Note: TypeFunc::make has a cache that makes this fast.)
2092 2103 const TypeFunc* tf = TypeFunc::make(dest_method);
2093 2104 int nargs = tf->_domain->_cnt - TypeFunc::Parms;
2094 2105 for (int j = 0; j < nargs; j++) {
2095 2106 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms);
2096 2107 if( targ->basic_type() == T_DOUBLE ) {
2097 2108 // If any parameters are doubles, they must be rounded before
2098 2109 // the call, dstore_rounding does gvn.transform
2099 2110 Node *arg = argument(j);
2100 2111 arg = dstore_rounding(arg);
2101 2112 set_argument(j, arg);
2102 2113 }
2103 2114 }
2104 2115 }
2105 2116
2106 2117 /**
2107 2118 * Record profiling data exact_kls for Node n with the type system so
2108 2119 * that it can propagate it (speculation)
2109 2120 *
2110 2121 * @param n node that the type applies to
2111 2122 * @param exact_kls type from profiling
2112 2123 *
2113 2124 * @return node with improved type
2114 2125 */
2115 2126 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls) {
2116 2127 const Type* current_type = _gvn.type(n);
2117 2128 assert(UseTypeSpeculation, "type speculation must be on");
2118 2129
2119 2130 const TypeOopPtr* speculative = current_type->speculative();
2120 2131
2121 2132 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2122 2133 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls);
2123 2134 const TypeOopPtr* xtype = tklass->as_instance_type();
2124 2135 assert(xtype->klass_is_exact(), "Should be exact");
2125 2136 // record the new speculative type's depth
2126 2137 speculative = xtype->with_inline_depth(jvms()->depth());
2127 2138 }
2128 2139
2129 2140 if (speculative != current_type->speculative()) {
2130 2141 // Build a type with a speculative type (what we think we know
2131 2142 // about the type but will need a guard when we use it)
2132 2143 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative);
2133 2144 // We're changing the type, we need a new CheckCast node to carry
2134 2145 // the new type. The new type depends on the control: what
2135 2146 // profiling tells us is only valid from here as far as we can
2136 2147 // tell.
2137 2148 Node* cast = new(C) CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2138 2149 cast = _gvn.transform(cast);
2139 2150 replace_in_map(n, cast);
2140 2151 n = cast;
2141 2152 }
2142 2153
2143 2154 return n;
2144 2155 }
2145 2156
2146 2157 /**
2147 2158 * Record profiling data from receiver profiling at an invoke with the
2148 2159 * type system so that it can propagate it (speculation)
2149 2160 *
2150 2161 * @param n receiver node
2151 2162 *
2152 2163 * @return node with improved type
2153 2164 */
2154 2165 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2155 2166 if (!UseTypeSpeculation) {
2156 2167 return n;
2157 2168 }
2158 2169 ciKlass* exact_kls = profile_has_unique_klass();
2159 2170 return record_profile_for_speculation(n, exact_kls);
2160 2171 }
2161 2172
2162 2173 /**
2163 2174 * Record profiling data from argument profiling at an invoke with the
2164 2175 * type system so that it can propagate it (speculation)
2165 2176 *
2166 2177 * @param dest_method target method for the call
2167 2178 * @param bc what invoke bytecode is this?
2168 2179 */
2169 2180 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2170 2181 if (!UseTypeSpeculation) {
2171 2182 return;
2172 2183 }
2173 2184 const TypeFunc* tf = TypeFunc::make(dest_method);
2174 2185 int nargs = tf->_domain->_cnt - TypeFunc::Parms;
2175 2186 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2176 2187 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2177 2188 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms);
2178 2189 if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) {
2179 2190 ciKlass* better_type = method()->argument_profiled_type(bci(), i);
2180 2191 if (better_type != NULL) {
2181 2192 record_profile_for_speculation(argument(j), better_type);
2182 2193 }
2183 2194 i++;
2184 2195 }
2185 2196 }
2186 2197 }
2187 2198
2188 2199 /**
2189 2200 * Record profiling data from parameter profiling at an invoke with
2190 2201 * the type system so that it can propagate it (speculation)
2191 2202 */
2192 2203 void GraphKit::record_profiled_parameters_for_speculation() {
2193 2204 if (!UseTypeSpeculation) {
2194 2205 return;
2195 2206 }
2196 2207 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2197 2208 if (_gvn.type(local(i))->isa_oopptr()) {
2198 2209 ciKlass* better_type = method()->parameter_profiled_type(j);
2199 2210 if (better_type != NULL) {
2200 2211 record_profile_for_speculation(local(i), better_type);
2201 2212 }
2202 2213 j++;
2203 2214 }
2204 2215 }
2205 2216 }
2206 2217
2207 2218 void GraphKit::round_double_result(ciMethod* dest_method) {
2208 2219 // A non-strict method may return a double value which has an extended
2209 2220 // exponent, but this must not be visible in a caller which is 'strict'
2210 2221 // If a strict caller invokes a non-strict callee, round a double result
2211 2222
2212 2223 BasicType result_type = dest_method->return_type()->basic_type();
2213 2224 assert( method() != NULL, "must have caller context");
2214 2225 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
2215 2226 // Destination method's return value is on top of stack
2216 2227 // dstore_rounding() does gvn.transform
2217 2228 Node *result = pop_pair();
2218 2229 result = dstore_rounding(result);
2219 2230 push_pair(result);
2220 2231 }
2221 2232 }
2222 2233
2223 2234 // rounding for strict float precision conformance
2224 2235 Node* GraphKit::precision_rounding(Node* n) {
2225 2236 return UseStrictFP && _method->flags().is_strict()
2226 2237 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
2227 2238 ? _gvn.transform( new (C) RoundFloatNode(0, n) )
2228 2239 : n;
2229 2240 }
2230 2241
2231 2242 // rounding for strict double precision conformance
2232 2243 Node* GraphKit::dprecision_rounding(Node *n) {
2233 2244 return UseStrictFP && _method->flags().is_strict()
2234 2245 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
2235 2246 ? _gvn.transform( new (C) RoundDoubleNode(0, n) )
2236 2247 : n;
2237 2248 }
2238 2249
2239 2250 // rounding for non-strict double stores
2240 2251 Node* GraphKit::dstore_rounding(Node* n) {
2241 2252 return Matcher::strict_fp_requires_explicit_rounding
2242 2253 && UseSSE <= 1
2243 2254 ? _gvn.transform( new (C) RoundDoubleNode(0, n) )
2244 2255 : n;
2245 2256 }
2246 2257
2247 2258 //=============================================================================
2248 2259 // Generate a fast path/slow path idiom. Graph looks like:
2249 2260 // [foo] indicates that 'foo' is a parameter
2250 2261 //
2251 2262 // [in] NULL
2252 2263 // \ /
2253 2264 // CmpP
2254 2265 // Bool ne
2255 2266 // If
2256 2267 // / \
2257 2268 // True False-<2>
2258 2269 // / |
2259 2270 // / cast_not_null
2260 2271 // Load | | ^
2261 2272 // [fast_test] | |
2262 2273 // gvn to opt_test | |
2263 2274 // / \ | <1>
2264 2275 // True False |
2265 2276 // | \\ |
2266 2277 // [slow_call] \[fast_result]
2267 2278 // Ctl Val \ \
2268 2279 // | \ \
2269 2280 // Catch <1> \ \
2270 2281 // / \ ^ \ \
2271 2282 // Ex No_Ex | \ \
2272 2283 // | \ \ | \ <2> \
2273 2284 // ... \ [slow_res] | | \ [null_result]
2274 2285 // \ \--+--+--- | |
2275 2286 // \ | / \ | /
2276 2287 // --------Region Phi
2277 2288 //
2278 2289 //=============================================================================
2279 2290 // Code is structured as a series of driver functions all called 'do_XXX' that
2280 2291 // call a set of helper functions. Helper functions first, then drivers.
2281 2292
2282 2293 //------------------------------null_check_oop---------------------------------
2283 2294 // Null check oop. Set null-path control into Region in slot 3.
2284 2295 // Make a cast-not-nullness use the other not-null control. Return cast.
2285 2296 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2286 2297 bool never_see_null, bool safe_for_replace) {
2287 2298 // Initial NULL check taken path
2288 2299 (*null_control) = top();
2289 2300 Node* cast = null_check_common(value, T_OBJECT, false, null_control);
2290 2301
2291 2302 // Generate uncommon_trap:
2292 2303 if (never_see_null && (*null_control) != top()) {
2293 2304 // If we see an unexpected null at a check-cast we record it and force a
2294 2305 // recompile; the offending check-cast will be compiled to handle NULLs.
2295 2306 // If we see more than one offending BCI, then all checkcasts in the
2296 2307 // method will be compiled to handle NULLs.
2297 2308 PreserveJVMState pjvms(this);
2298 2309 set_control(*null_control);
2299 2310 replace_in_map(value, null());
2300 2311 uncommon_trap(Deoptimization::Reason_null_check,
2301 2312 Deoptimization::Action_make_not_entrant);
2302 2313 (*null_control) = top(); // NULL path is dead
2303 2314 }
2304 2315 if ((*null_control) == top() && safe_for_replace) {
2305 2316 replace_in_map(value, cast);
2306 2317 }
2307 2318
2308 2319 // Cast away null-ness on the result
2309 2320 return cast;
2310 2321 }
2311 2322
2312 2323 //------------------------------opt_iff----------------------------------------
2313 2324 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2314 2325 // Return slow-path control.
2315 2326 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2316 2327 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2317 2328
2318 2329 // Fast path taken; set region slot 2
2319 2330 Node *fast_taken = _gvn.transform( new (C) IfFalseNode(opt_iff) );
2320 2331 region->init_req(2,fast_taken); // Capture fast-control
2321 2332
2322 2333 // Fast path not-taken, i.e. slow path
2323 2334 Node *slow_taken = _gvn.transform( new (C) IfTrueNode(opt_iff) );
2324 2335 return slow_taken;
2325 2336 }
2326 2337
2327 2338 //-----------------------------make_runtime_call-------------------------------
2328 2339 Node* GraphKit::make_runtime_call(int flags,
2329 2340 const TypeFunc* call_type, address call_addr,
2330 2341 const char* call_name,
2331 2342 const TypePtr* adr_type,
2332 2343 // The following parms are all optional.
2333 2344 // The first NULL ends the list.
2334 2345 Node* parm0, Node* parm1,
2335 2346 Node* parm2, Node* parm3,
2336 2347 Node* parm4, Node* parm5,
2337 2348 Node* parm6, Node* parm7) {
2338 2349 // Slow-path call
2339 2350 bool is_leaf = !(flags & RC_NO_LEAF);
2340 2351 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2341 2352 if (call_name == NULL) {
2342 2353 assert(!is_leaf, "must supply name for leaf");
2343 2354 call_name = OptoRuntime::stub_name(call_addr);
2344 2355 }
2345 2356 CallNode* call;
2346 2357 if (!is_leaf) {
2347 2358 call = new(C) CallStaticJavaNode(call_type, call_addr, call_name,
2348 2359 bci(), adr_type);
2349 2360 } else if (flags & RC_NO_FP) {
2350 2361 call = new(C) CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2351 2362 } else {
2352 2363 call = new(C) CallLeafNode(call_type, call_addr, call_name, adr_type);
2353 2364 }
2354 2365
2355 2366 // The following is similar to set_edges_for_java_call,
2356 2367 // except that the memory effects of the call are restricted to AliasIdxRaw.
2357 2368
2358 2369 // Slow path call has no side-effects, uses few values
2359 2370 bool wide_in = !(flags & RC_NARROW_MEM);
2360 2371 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2361 2372
2362 2373 Node* prev_mem = NULL;
2363 2374 if (wide_in) {
2364 2375 prev_mem = set_predefined_input_for_runtime_call(call);
2365 2376 } else {
2366 2377 assert(!wide_out, "narrow in => narrow out");
2367 2378 Node* narrow_mem = memory(adr_type);
2368 2379 prev_mem = reset_memory();
2369 2380 map()->set_memory(narrow_mem);
2370 2381 set_predefined_input_for_runtime_call(call);
2371 2382 }
2372 2383
2373 2384 // Hook each parm in order. Stop looking at the first NULL.
2374 2385 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2375 2386 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2376 2387 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2377 2388 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2378 2389 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2379 2390 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2380 2391 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2381 2392 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2382 2393 /* close each nested if ===> */ } } } } } } } }
2383 2394 assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2384 2395
2385 2396 if (!is_leaf) {
2386 2397 // Non-leaves can block and take safepoints:
2387 2398 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2388 2399 }
2389 2400 // Non-leaves can throw exceptions:
2390 2401 if (has_io) {
2391 2402 call->set_req(TypeFunc::I_O, i_o());
2392 2403 }
2393 2404
2394 2405 if (flags & RC_UNCOMMON) {
2395 2406 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2396 2407 // (An "if" probability corresponds roughly to an unconditional count.
2397 2408 // Sort of.)
2398 2409 call->set_cnt(PROB_UNLIKELY_MAG(4));
2399 2410 }
2400 2411
2401 2412 Node* c = _gvn.transform(call);
2402 2413 assert(c == call, "cannot disappear");
2403 2414
2404 2415 if (wide_out) {
2405 2416 // Slow path call has full side-effects.
2406 2417 set_predefined_output_for_runtime_call(call);
2407 2418 } else {
2408 2419 // Slow path call has few side-effects, and/or sets few values.
2409 2420 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2410 2421 }
2411 2422
2412 2423 if (has_io) {
2413 2424 set_i_o(_gvn.transform(new (C) ProjNode(call, TypeFunc::I_O)));
2414 2425 }
2415 2426 return call;
2416 2427
2417 2428 }
2418 2429
2419 2430 //------------------------------merge_memory-----------------------------------
2420 2431 // Merge memory from one path into the current memory state.
2421 2432 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2422 2433 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2423 2434 Node* old_slice = mms.force_memory();
2424 2435 Node* new_slice = mms.memory2();
2425 2436 if (old_slice != new_slice) {
2426 2437 PhiNode* phi;
2427 2438 if (new_slice->is_Phi() && new_slice->as_Phi()->region() == region) {
2428 2439 phi = new_slice->as_Phi();
2429 2440 #ifdef ASSERT
2430 2441 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region)
2431 2442 old_slice = old_slice->in(new_path);
2432 2443 // Caller is responsible for ensuring that any pre-existing
2433 2444 // phis are already aware of old memory.
2434 2445 int old_path = (new_path > 1) ? 1 : 2; // choose old_path != new_path
2435 2446 assert(phi->in(old_path) == old_slice, "pre-existing phis OK");
2436 2447 #endif
2437 2448 mms.set_memory(phi);
2438 2449 } else {
2439 2450 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2440 2451 _gvn.set_type(phi, Type::MEMORY);
2441 2452 phi->set_req(new_path, new_slice);
2442 2453 mms.set_memory(_gvn.transform(phi)); // assume it is complete
2443 2454 }
2444 2455 }
2445 2456 }
2446 2457 }
2447 2458
2448 2459 //------------------------------make_slow_call_ex------------------------------
2449 2460 // Make the exception handler hookups for the slow call
2450 2461 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj) {
2451 2462 if (stopped()) return;
2452 2463
2453 2464 // Make a catch node with just two handlers: fall-through and catch-all
2454 2465 Node* i_o = _gvn.transform( new (C) ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2455 2466 Node* catc = _gvn.transform( new (C) CatchNode(control(), i_o, 2) );
2456 2467 Node* norm = _gvn.transform( new (C) CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2457 2468 Node* excp = _gvn.transform( new (C) CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2458 2469
2459 2470 { PreserveJVMState pjvms(this);
2460 2471 set_control(excp);
2461 2472 set_i_o(i_o);
2462 2473
2463 2474 if (excp != top()) {
2464 2475 // Create an exception state also.
2465 2476 // Use an exact type if the caller has specified a specific exception.
2466 2477 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2467 2478 Node* ex_oop = new (C) CreateExNode(ex_type, control(), i_o);
2468 2479 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2469 2480 }
2470 2481 }
2471 2482
2472 2483 // Get the no-exception control from the CatchNode.
2473 2484 set_control(norm);
2474 2485 }
2475 2486
2476 2487
2477 2488 //-------------------------------gen_subtype_check-----------------------------
2478 2489 // Generate a subtyping check. Takes as input the subtype and supertype.
2479 2490 // Returns 2 values: sets the default control() to the true path and returns
2480 2491 // the false path. Only reads invariant memory; sets no (visible) memory.
2481 2492 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2482 2493 // but that's not exposed to the optimizer. This call also doesn't take in an
2483 2494 // Object; if you wish to check an Object you need to load the Object's class
2484 2495 // prior to coming here.
2485 2496 Node* GraphKit::gen_subtype_check(Node* subklass, Node* superklass) {
2486 2497 // Fast check for identical types, perhaps identical constants.
2487 2498 // The types can even be identical non-constants, in cases
2488 2499 // involving Array.newInstance, Object.clone, etc.
2489 2500 if (subklass == superklass)
2490 2501 return top(); // false path is dead; no test needed.
2491 2502
2492 2503 if (_gvn.type(superklass)->singleton()) {
2493 2504 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
2494 2505 ciKlass* subk = _gvn.type(subklass)->is_klassptr()->klass();
2495 2506
2496 2507 // In the common case of an exact superklass, try to fold up the
2497 2508 // test before generating code. You may ask, why not just generate
2498 2509 // the code and then let it fold up? The answer is that the generated
2499 2510 // code will necessarily include null checks, which do not always
2500 2511 // completely fold away. If they are also needless, then they turn
2501 2512 // into a performance loss. Example:
2502 2513 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2503 2514 // Here, the type of 'fa' is often exact, so the store check
2504 2515 // of fa[1]=x will fold up, without testing the nullness of x.
2505 2516 switch (static_subtype_check(superk, subk)) {
2506 2517 case SSC_always_false:
2507 2518 {
2508 2519 Node* always_fail = control();
2509 2520 set_control(top());
2510 2521 return always_fail;
2511 2522 }
2512 2523 case SSC_always_true:
2513 2524 return top();
2514 2525 case SSC_easy_test:
2515 2526 {
2516 2527 // Just do a direct pointer compare and be done.
2517 2528 Node* cmp = _gvn.transform( new(C) CmpPNode(subklass, superklass) );
2518 2529 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) );
2519 2530 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
2520 2531 set_control( _gvn.transform( new(C) IfTrueNode (iff) ) );
2521 2532 return _gvn.transform( new(C) IfFalseNode(iff) );
2522 2533 }
2523 2534 case SSC_full_test:
2524 2535 break;
2525 2536 default:
2526 2537 ShouldNotReachHere();
2527 2538 }
2528 2539 }
2529 2540
2530 2541 // %%% Possible further optimization: Even if the superklass is not exact,
2531 2542 // if the subklass is the unique subtype of the superklass, the check
2532 2543 // will always succeed. We could leave a dependency behind to ensure this.
2533 2544
2534 2545 // First load the super-klass's check-offset
2535 2546 Node *p1 = basic_plus_adr( superklass, superklass, in_bytes(Klass::super_check_offset_offset()) );
2536 2547 Node *chk_off = _gvn.transform(new (C) LoadINode(NULL, memory(p1), p1, _gvn.type(p1)->is_ptr(),
2537 2548 TypeInt::INT, MemNode::unordered));
2538 2549 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2539 2550 bool might_be_cache = (find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2540 2551
2541 2552 // Load from the sub-klass's super-class display list, or a 1-word cache of
2542 2553 // the secondary superclass list, or a failing value with a sentinel offset
2543 2554 // if the super-klass is an interface or exceptionally deep in the Java
2544 2555 // hierarchy and we have to scan the secondary superclass list the hard way.
2545 2556 // Worst-case type is a little odd: NULL is allowed as a result (usually
2546 2557 // klass loads can never produce a NULL).
2547 2558 Node *chk_off_X = ConvI2X(chk_off);
2548 2559 Node *p2 = _gvn.transform( new (C) AddPNode(subklass,subklass,chk_off_X) );
2549 2560 // For some types like interfaces the following loadKlass is from a 1-word
2550 2561 // cache which is mutable so can't use immutable memory. Other
2551 2562 // types load from the super-class display table which is immutable.
2552 2563 Node *kmem = might_be_cache ? memory(p2) : immutable_memory();
2553 2564 Node *nkls = _gvn.transform( LoadKlassNode::make( _gvn, kmem, p2, _gvn.type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL ) );
2554 2565
2555 2566 // Compile speed common case: ARE a subtype and we canNOT fail
2556 2567 if( superklass == nkls )
2557 2568 return top(); // false path is dead; no test needed.
2558 2569
2559 2570 // See if we get an immediate positive hit. Happens roughly 83% of the
2560 2571 // time. Test to see if the value loaded just previously from the subklass
2561 2572 // is exactly the superklass.
2562 2573 Node *cmp1 = _gvn.transform( new (C) CmpPNode( superklass, nkls ) );
2563 2574 Node *bol1 = _gvn.transform( new (C) BoolNode( cmp1, BoolTest::eq ) );
2564 2575 IfNode *iff1 = create_and_xform_if( control(), bol1, PROB_LIKELY(0.83f), COUNT_UNKNOWN );
2565 2576 Node *iftrue1 = _gvn.transform( new (C) IfTrueNode ( iff1 ) );
2566 2577 set_control( _gvn.transform( new (C) IfFalseNode( iff1 ) ) );
2567 2578
2568 2579 // Compile speed common case: Check for being deterministic right now. If
2569 2580 // chk_off is a constant and not equal to cacheoff then we are NOT a
2570 2581 // subklass. In this case we need exactly the 1 test above and we can
2571 2582 // return those results immediately.
2572 2583 if (!might_be_cache) {
2573 2584 Node* not_subtype_ctrl = control();
2574 2585 set_control(iftrue1); // We need exactly the 1 test above
2575 2586 return not_subtype_ctrl;
2576 2587 }
2577 2588
2578 2589 // Gather the various success & failures here
2579 2590 RegionNode *r_ok_subtype = new (C) RegionNode(4);
2580 2591 record_for_igvn(r_ok_subtype);
2581 2592 RegionNode *r_not_subtype = new (C) RegionNode(3);
2582 2593 record_for_igvn(r_not_subtype);
2583 2594
2584 2595 r_ok_subtype->init_req(1, iftrue1);
2585 2596
2586 2597 // Check for immediate negative hit. Happens roughly 11% of the time (which
2587 2598 // is roughly 63% of the remaining cases). Test to see if the loaded
2588 2599 // check-offset points into the subklass display list or the 1-element
2589 2600 // cache. If it points to the display (and NOT the cache) and the display
2590 2601 // missed then it's not a subtype.
2591 2602 Node *cacheoff = _gvn.intcon(cacheoff_con);
2592 2603 Node *cmp2 = _gvn.transform( new (C) CmpINode( chk_off, cacheoff ) );
2593 2604 Node *bol2 = _gvn.transform( new (C) BoolNode( cmp2, BoolTest::ne ) );
2594 2605 IfNode *iff2 = create_and_xform_if( control(), bol2, PROB_LIKELY(0.63f), COUNT_UNKNOWN );
2595 2606 r_not_subtype->init_req(1, _gvn.transform( new (C) IfTrueNode (iff2) ) );
2596 2607 set_control( _gvn.transform( new (C) IfFalseNode(iff2) ) );
2597 2608
2598 2609 // Check for self. Very rare to get here, but it is taken 1/3 the time.
2599 2610 // No performance impact (too rare) but allows sharing of secondary arrays
2600 2611 // which has some footprint reduction.
2601 2612 Node *cmp3 = _gvn.transform( new (C) CmpPNode( subklass, superklass ) );
2602 2613 Node *bol3 = _gvn.transform( new (C) BoolNode( cmp3, BoolTest::eq ) );
2603 2614 IfNode *iff3 = create_and_xform_if( control(), bol3, PROB_LIKELY(0.36f), COUNT_UNKNOWN );
2604 2615 r_ok_subtype->init_req(2, _gvn.transform( new (C) IfTrueNode ( iff3 ) ) );
2605 2616 set_control( _gvn.transform( new (C) IfFalseNode( iff3 ) ) );
2606 2617
2607 2618 // -- Roads not taken here: --
2608 2619 // We could also have chosen to perform the self-check at the beginning
2609 2620 // of this code sequence, as the assembler does. This would not pay off
2610 2621 // the same way, since the optimizer, unlike the assembler, can perform
2611 2622 // static type analysis to fold away many successful self-checks.
2612 2623 // Non-foldable self checks work better here in second position, because
2613 2624 // the initial primary superclass check subsumes a self-check for most
2614 2625 // types. An exception would be a secondary type like array-of-interface,
2615 2626 // which does not appear in its own primary supertype display.
2616 2627 // Finally, we could have chosen to move the self-check into the
2617 2628 // PartialSubtypeCheckNode, and from there out-of-line in a platform
2618 2629 // dependent manner. But it is worthwhile to have the check here,
2619 2630 // where it can be perhaps be optimized. The cost in code space is
2620 2631 // small (register compare, branch).
2621 2632
2622 2633 // Now do a linear scan of the secondary super-klass array. Again, no real
2623 2634 // performance impact (too rare) but it's gotta be done.
2624 2635 // Since the code is rarely used, there is no penalty for moving it
2625 2636 // out of line, and it can only improve I-cache density.
2626 2637 // The decision to inline or out-of-line this final check is platform
2627 2638 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2628 2639 Node* psc = _gvn.transform(
2629 2640 new (C) PartialSubtypeCheckNode(control(), subklass, superklass) );
2630 2641
2631 2642 Node *cmp4 = _gvn.transform( new (C) CmpPNode( psc, null() ) );
2632 2643 Node *bol4 = _gvn.transform( new (C) BoolNode( cmp4, BoolTest::ne ) );
2633 2644 IfNode *iff4 = create_and_xform_if( control(), bol4, PROB_FAIR, COUNT_UNKNOWN );
2634 2645 r_not_subtype->init_req(2, _gvn.transform( new (C) IfTrueNode (iff4) ) );
2635 2646 r_ok_subtype ->init_req(3, _gvn.transform( new (C) IfFalseNode(iff4) ) );
2636 2647
2637 2648 // Return false path; set default control to true path.
2638 2649 set_control( _gvn.transform(r_ok_subtype) );
2639 2650 return _gvn.transform(r_not_subtype);
2640 2651 }
2641 2652
2642 2653 //----------------------------static_subtype_check-----------------------------
2643 2654 // Shortcut important common cases when superklass is exact:
2644 2655 // (0) superklass is java.lang.Object (can occur in reflective code)
2645 2656 // (1) subklass is already limited to a subtype of superklass => always ok
2646 2657 // (2) subklass does not overlap with superklass => always fail
2647 2658 // (3) superklass has NO subtypes and we can check with a simple compare.
2648 2659 int GraphKit::static_subtype_check(ciKlass* superk, ciKlass* subk) {
2649 2660 if (StressReflectiveCode) {
2650 2661 return SSC_full_test; // Let caller generate the general case.
2651 2662 }
2652 2663
2653 2664 if (superk == env()->Object_klass()) {
2654 2665 return SSC_always_true; // (0) this test cannot fail
2655 2666 }
2656 2667
2657 2668 ciType* superelem = superk;
2658 2669 if (superelem->is_array_klass())
2659 2670 superelem = superelem->as_array_klass()->base_element_type();
2660 2671
2661 2672 if (!subk->is_interface()) { // cannot trust static interface types yet
2662 2673 if (subk->is_subtype_of(superk)) {
2663 2674 return SSC_always_true; // (1) false path dead; no dynamic test needed
2664 2675 }
2665 2676 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
2666 2677 !superk->is_subtype_of(subk)) {
2667 2678 return SSC_always_false;
2668 2679 }
2669 2680 }
2670 2681
2671 2682 // If casting to an instance klass, it must have no subtypes
2672 2683 if (superk->is_interface()) {
2673 2684 // Cannot trust interfaces yet.
2674 2685 // %%% S.B. superk->nof_implementors() == 1
2675 2686 } else if (superelem->is_instance_klass()) {
2676 2687 ciInstanceKlass* ik = superelem->as_instance_klass();
2677 2688 if (!ik->has_subklass() && !ik->is_interface()) {
2678 2689 if (!ik->is_final()) {
2679 2690 // Add a dependency if there is a chance of a later subclass.
2680 2691 C->dependencies()->assert_leaf_type(ik);
2681 2692 }
2682 2693 return SSC_easy_test; // (3) caller can do a simple ptr comparison
2683 2694 }
2684 2695 } else {
2685 2696 // A primitive array type has no subtypes.
2686 2697 return SSC_easy_test; // (3) caller can do a simple ptr comparison
2687 2698 }
2688 2699
2689 2700 return SSC_full_test;
2690 2701 }
2691 2702
2692 2703 // Profile-driven exact type check:
2693 2704 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2694 2705 float prob,
2695 2706 Node* *casted_receiver) {
2696 2707 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2697 2708 Node* recv_klass = load_object_klass(receiver);
2698 2709 Node* want_klass = makecon(tklass);
2699 2710 Node* cmp = _gvn.transform( new(C) CmpPNode(recv_klass, want_klass) );
2700 2711 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) );
2701 2712 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2702 2713 set_control( _gvn.transform( new(C) IfTrueNode (iff) ));
2703 2714 Node* fail = _gvn.transform( new(C) IfFalseNode(iff) );
2704 2715
2705 2716 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2706 2717 assert(recv_xtype->klass_is_exact(), "");
2707 2718
2708 2719 // Subsume downstream occurrences of receiver with a cast to
2709 2720 // recv_xtype, since now we know what the type will be.
2710 2721 Node* cast = new(C) CheckCastPPNode(control(), receiver, recv_xtype);
2711 2722 (*casted_receiver) = _gvn.transform(cast);
2712 2723 // (User must make the replace_in_map call.)
2713 2724
2714 2725 return fail;
2715 2726 }
2716 2727
2717 2728
2718 2729 //------------------------------seems_never_null-------------------------------
2719 2730 // Use null_seen information if it is available from the profile.
2720 2731 // If we see an unexpected null at a type check we record it and force a
2721 2732 // recompile; the offending check will be recompiled to handle NULLs.
2722 2733 // If we see several offending BCIs, then all checks in the
2723 2734 // method will be recompiled.
2724 2735 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data) {
2725 2736 if (UncommonNullCast // Cutout for this technique
2726 2737 && obj != null() // And not the -Xcomp stupid case?
2727 2738 && !too_many_traps(Deoptimization::Reason_null_check)
2728 2739 ) {
2729 2740 if (data == NULL)
2730 2741 // Edge case: no mature data. Be optimistic here.
2731 2742 return true;
2732 2743 // If the profile has not seen a null, assume it won't happen.
2733 2744 assert(java_bc() == Bytecodes::_checkcast ||
2734 2745 java_bc() == Bytecodes::_instanceof ||
2735 2746 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
2736 2747 return !data->as_BitData()->null_seen();
2737 2748 }
2738 2749 return false;
2739 2750 }
2740 2751
2741 2752 //------------------------maybe_cast_profiled_receiver-------------------------
2742 2753 // If the profile has seen exactly one type, narrow to exactly that type.
2743 2754 // Subsequent type checks will always fold up.
2744 2755 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
2745 2756 ciKlass* require_klass,
2746 2757 ciKlass* spec_klass,
2747 2758 bool safe_for_replace) {
2748 2759 if (!UseTypeProfile || !TypeProfileCasts) return NULL;
2749 2760
2750 2761 Deoptimization::DeoptReason reason = spec_klass == NULL ? Deoptimization::Reason_class_check : Deoptimization::Reason_speculate_class_check;
2751 2762
2752 2763 // Make sure we haven't already deoptimized from this tactic.
2753 2764 if (too_many_traps(reason))
2754 2765 return NULL;
2755 2766
2756 2767 // (No, this isn't a call, but it's enough like a virtual call
2757 2768 // to use the same ciMethod accessor to get the profile info...)
2758 2769 // If we have a speculative type use it instead of profiling (which
2759 2770 // may not help us)
2760 2771 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass;
2761 2772 if (exact_kls != NULL) {// no cast failures here
2762 2773 if (require_klass == NULL ||
2763 2774 static_subtype_check(require_klass, exact_kls) == SSC_always_true) {
2764 2775 // If we narrow the type to match what the type profile sees or
2765 2776 // the speculative type, we can then remove the rest of the
2766 2777 // cast.
2767 2778 // This is a win, even if the exact_kls is very specific,
2768 2779 // because downstream operations, such as method calls,
2769 2780 // will often benefit from the sharper type.
2770 2781 Node* exact_obj = not_null_obj; // will get updated in place...
2771 2782 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2772 2783 &exact_obj);
2773 2784 { PreserveJVMState pjvms(this);
2774 2785 set_control(slow_ctl);
2775 2786 uncommon_trap(reason,
2776 2787 Deoptimization::Action_maybe_recompile);
2777 2788 }
2778 2789 if (safe_for_replace) {
2779 2790 replace_in_map(not_null_obj, exact_obj);
2780 2791 }
2781 2792 return exact_obj;
2782 2793 }
2783 2794 // assert(ssc == SSC_always_true)... except maybe the profile lied to us.
2784 2795 }
2785 2796
2786 2797 return NULL;
2787 2798 }
2788 2799
2789 2800 /**
2790 2801 * Cast obj to type and emit guard unless we had too many traps here
2791 2802 * already
2792 2803 *
2793 2804 * @param obj node being casted
2794 2805 * @param type type to cast the node to
2795 2806 * @param not_null true if we know node cannot be null
2796 2807 */
2797 2808 Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
2798 2809 ciKlass* type,
2799 2810 bool not_null) {
2800 2811 // type == NULL if profiling tells us this object is always null
2801 2812 if (type != NULL) {
2802 2813 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
2803 2814 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_null_check;
2804 2815 if (!too_many_traps(null_reason) &&
2805 2816 !too_many_traps(class_reason)) {
2806 2817 Node* not_null_obj = NULL;
2807 2818 // not_null is true if we know the object is not null and
2808 2819 // there's no need for a null check
2809 2820 if (!not_null) {
2810 2821 Node* null_ctl = top();
2811 2822 not_null_obj = null_check_oop(obj, &null_ctl, true, true);
2812 2823 assert(null_ctl->is_top(), "no null control here");
2813 2824 } else {
2814 2825 not_null_obj = obj;
2815 2826 }
2816 2827
2817 2828 Node* exact_obj = not_null_obj;
2818 2829 ciKlass* exact_kls = type;
2819 2830 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2820 2831 &exact_obj);
2821 2832 {
2822 2833 PreserveJVMState pjvms(this);
2823 2834 set_control(slow_ctl);
2824 2835 uncommon_trap(class_reason,
2825 2836 Deoptimization::Action_maybe_recompile);
2826 2837 }
2827 2838 replace_in_map(not_null_obj, exact_obj);
2828 2839 obj = exact_obj;
2829 2840 }
2830 2841 } else {
2831 2842 if (!too_many_traps(Deoptimization::Reason_null_assert)) {
2832 2843 Node* exact_obj = null_assert(obj);
2833 2844 replace_in_map(obj, exact_obj);
2834 2845 obj = exact_obj;
2835 2846 }
2836 2847 }
2837 2848 return obj;
2838 2849 }
2839 2850
2840 2851 //-------------------------------gen_instanceof--------------------------------
2841 2852 // Generate an instance-of idiom. Used by both the instance-of bytecode
2842 2853 // and the reflective instance-of call.
2843 2854 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
2844 2855 kill_dead_locals(); // Benefit all the uncommon traps
2845 2856 assert( !stopped(), "dead parse path should be checked in callers" );
2846 2857 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
2847 2858 "must check for not-null not-dead klass in callers");
2848 2859
2849 2860 // Make the merge point
2850 2861 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
2851 2862 RegionNode* region = new(C) RegionNode(PATH_LIMIT);
2852 2863 Node* phi = new(C) PhiNode(region, TypeInt::BOOL);
2853 2864 C->set_has_split_ifs(true); // Has chance for split-if optimization
2854 2865
2855 2866 ciProfileData* data = NULL;
2856 2867 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
2857 2868 data = method()->method_data()->bci_to_data(bci());
2858 2869 }
2859 2870 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
2860 2871 && seems_never_null(obj, data));
2861 2872
2862 2873 // Null check; get casted pointer; set region slot 3
2863 2874 Node* null_ctl = top();
2864 2875 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace);
2865 2876
2866 2877 // If not_null_obj is dead, only null-path is taken
2867 2878 if (stopped()) { // Doing instance-of on a NULL?
2868 2879 set_control(null_ctl);
2869 2880 return intcon(0);
2870 2881 }
2871 2882 region->init_req(_null_path, null_ctl);
2872 2883 phi ->init_req(_null_path, intcon(0)); // Set null path value
2873 2884 if (null_ctl == top()) {
2874 2885 // Do this eagerly, so that pattern matches like is_diamond_phi
2875 2886 // will work even during parsing.
2876 2887 assert(_null_path == PATH_LIMIT-1, "delete last");
2877 2888 region->del_req(_null_path);
2878 2889 phi ->del_req(_null_path);
2879 2890 }
2880 2891
2881 2892 // Do we know the type check always succeed?
2882 2893 bool known_statically = false;
2883 2894 if (_gvn.type(superklass)->singleton()) {
2884 2895 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
2885 2896 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass();
2886 2897 if (subk != NULL && subk->is_loaded()) {
2887 2898 int static_res = static_subtype_check(superk, subk);
2888 2899 known_statically = (static_res == SSC_always_true || static_res == SSC_always_false);
2889 2900 }
2890 2901 }
2891 2902
2892 2903 if (known_statically && UseTypeSpeculation) {
2893 2904 // If we know the type check always succeeds then we don't use the
2894 2905 // profiling data at this bytecode. Don't lose it, feed it to the
2895 2906 // type system as a speculative type.
2896 2907 not_null_obj = record_profiled_receiver_for_speculation(not_null_obj);
2897 2908 } else {
2898 2909 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
2899 2910 // We may not have profiling here or it may not help us. If we
2900 2911 // have a speculative type use it to perform an exact cast.
2901 2912 ciKlass* spec_obj_type = obj_type->speculative_type();
2902 2913 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) {
2903 2914 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace);
2904 2915 if (stopped()) { // Profile disagrees with this path.
2905 2916 set_control(null_ctl); // Null is the only remaining possibility.
2906 2917 return intcon(0);
2907 2918 }
2908 2919 if (cast_obj != NULL) {
2909 2920 not_null_obj = cast_obj;
2910 2921 }
2911 2922 }
2912 2923 }
2913 2924
2914 2925 // Load the object's klass
2915 2926 Node* obj_klass = load_object_klass(not_null_obj);
2916 2927
2917 2928 // Generate the subtype check
2918 2929 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
2919 2930
2920 2931 // Plug in the success path to the general merge in slot 1.
2921 2932 region->init_req(_obj_path, control());
2922 2933 phi ->init_req(_obj_path, intcon(1));
2923 2934
2924 2935 // Plug in the failing path to the general merge in slot 2.
2925 2936 region->init_req(_fail_path, not_subtype_ctrl);
2926 2937 phi ->init_req(_fail_path, intcon(0));
2927 2938
2928 2939 // Return final merged results
2929 2940 set_control( _gvn.transform(region) );
2930 2941 record_for_igvn(region);
2931 2942 return _gvn.transform(phi);
2932 2943 }
2933 2944
2934 2945 //-------------------------------gen_checkcast---------------------------------
2935 2946 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
2936 2947 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
2937 2948 // uncommon-trap paths work. Adjust stack after this call.
2938 2949 // If failure_control is supplied and not null, it is filled in with
2939 2950 // the control edge for the cast failure. Otherwise, an appropriate
2940 2951 // uncommon trap or exception is thrown.
2941 2952 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
2942 2953 Node* *failure_control) {
2943 2954 kill_dead_locals(); // Benefit all the uncommon traps
2944 2955 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
2945 2956 const Type *toop = TypeOopPtr::make_from_klass(tk->klass());
2946 2957
2947 2958 // Fast cutout: Check the case that the cast is vacuously true.
2948 2959 // This detects the common cases where the test will short-circuit
2949 2960 // away completely. We do this before we perform the null check,
2950 2961 // because if the test is going to turn into zero code, we don't
2951 2962 // want a residual null check left around. (Causes a slowdown,
2952 2963 // for example, in some objArray manipulations, such as a[i]=a[j].)
2953 2964 if (tk->singleton()) {
2954 2965 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
2955 2966 if (objtp != NULL && objtp->klass() != NULL) {
2956 2967 switch (static_subtype_check(tk->klass(), objtp->klass())) {
2957 2968 case SSC_always_true:
2958 2969 // If we know the type check always succeed then we don't use
2959 2970 // the profiling data at this bytecode. Don't lose it, feed it
2960 2971 // to the type system as a speculative type.
2961 2972 return record_profiled_receiver_for_speculation(obj);
2962 2973 case SSC_always_false:
2963 2974 // It needs a null check because a null will *pass* the cast check.
2964 2975 // A non-null value will always produce an exception.
2965 2976 return null_assert(obj);
2966 2977 }
2967 2978 }
2968 2979 }
2969 2980
2970 2981 ciProfileData* data = NULL;
2971 2982 bool safe_for_replace = false;
2972 2983 if (failure_control == NULL) { // use MDO in regular case only
2973 2984 assert(java_bc() == Bytecodes::_aastore ||
2974 2985 java_bc() == Bytecodes::_checkcast,
2975 2986 "interpreter profiles type checks only for these BCs");
2976 2987 data = method()->method_data()->bci_to_data(bci());
2977 2988 safe_for_replace = true;
2978 2989 }
2979 2990
2980 2991 // Make the merge point
2981 2992 enum { _obj_path = 1, _null_path, PATH_LIMIT };
2982 2993 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
2983 2994 Node* phi = new (C) PhiNode(region, toop);
2984 2995 C->set_has_split_ifs(true); // Has chance for split-if optimization
2985 2996
2986 2997 // Use null-cast information if it is available
2987 2998 bool never_see_null = ((failure_control == NULL) // regular case only
2988 2999 && seems_never_null(obj, data));
2989 3000
2990 3001 // Null check; get casted pointer; set region slot 3
2991 3002 Node* null_ctl = top();
2992 3003 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace);
2993 3004
2994 3005 // If not_null_obj is dead, only null-path is taken
2995 3006 if (stopped()) { // Doing instance-of on a NULL?
2996 3007 set_control(null_ctl);
2997 3008 return null();
2998 3009 }
2999 3010 region->init_req(_null_path, null_ctl);
3000 3011 phi ->init_req(_null_path, null()); // Set null path value
3001 3012 if (null_ctl == top()) {
3002 3013 // Do this eagerly, so that pattern matches like is_diamond_phi
3003 3014 // will work even during parsing.
3004 3015 assert(_null_path == PATH_LIMIT-1, "delete last");
3005 3016 region->del_req(_null_path);
3006 3017 phi ->del_req(_null_path);
3007 3018 }
3008 3019
3009 3020 Node* cast_obj = NULL;
3010 3021 if (tk->klass_is_exact()) {
3011 3022 // The following optimization tries to statically cast the speculative type of the object
3012 3023 // (for example obtained during profiling) to the type of the superklass and then do a
3013 3024 // dynamic check that the type of the object is what we expect. To work correctly
3014 3025 // for checkcast and aastore the type of superklass should be exact.
3015 3026 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3016 3027 // We may not have profiling here or it may not help us. If we have
3017 3028 // a speculative type use it to perform an exact cast.
3018 3029 ciKlass* spec_obj_type = obj_type->speculative_type();
3019 3030 if (spec_obj_type != NULL ||
3020 3031 (data != NULL &&
3021 3032 // Counter has never been decremented (due to cast failure).
3022 3033 // ...This is a reasonable thing to expect. It is true of
3023 3034 // all casts inserted by javac to implement generic types.
3024 3035 data->as_CounterData()->count() >= 0)) {
3025 3036 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace);
3026 3037 if (cast_obj != NULL) {
3027 3038 if (failure_control != NULL) // failure is now impossible
3028 3039 (*failure_control) = top();
3029 3040 // adjust the type of the phi to the exact klass:
3030 3041 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3031 3042 }
3032 3043 }
3033 3044 }
3034 3045
3035 3046 if (cast_obj == NULL) {
3036 3047 // Load the object's klass
3037 3048 Node* obj_klass = load_object_klass(not_null_obj);
3038 3049
3039 3050 // Generate the subtype check
3040 3051 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
3041 3052
3042 3053 // Plug in success path into the merge
3043 3054 cast_obj = _gvn.transform(new (C) CheckCastPPNode(control(),
3044 3055 not_null_obj, toop));
3045 3056 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3046 3057 if (failure_control == NULL) {
3047 3058 if (not_subtype_ctrl != top()) { // If failure is possible
3048 3059 PreserveJVMState pjvms(this);
3049 3060 set_control(not_subtype_ctrl);
3050 3061 builtin_throw(Deoptimization::Reason_class_check, obj_klass);
3051 3062 }
3052 3063 } else {
3053 3064 (*failure_control) = not_subtype_ctrl;
3054 3065 }
3055 3066 }
3056 3067
3057 3068 region->init_req(_obj_path, control());
3058 3069 phi ->init_req(_obj_path, cast_obj);
3059 3070
3060 3071 // A merge of NULL or Casted-NotNull obj
3061 3072 Node* res = _gvn.transform(phi);
3062 3073
3063 3074 // Note I do NOT always 'replace_in_map(obj,result)' here.
3064 3075 // if( tk->klass()->can_be_primary_super() )
3065 3076 // This means that if I successfully store an Object into an array-of-String
3066 3077 // I 'forget' that the Object is really now known to be a String. I have to
3067 3078 // do this because we don't have true union types for interfaces - if I store
3068 3079 // a Baz into an array-of-Interface and then tell the optimizer it's an
3069 3080 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3070 3081 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3071 3082 // replace_in_map( obj, res );
3072 3083
3073 3084 // Return final merged results
3074 3085 set_control( _gvn.transform(region) );
3075 3086 record_for_igvn(region);
3076 3087 return res;
3077 3088 }
3078 3089
3079 3090 //------------------------------next_monitor-----------------------------------
3080 3091 // What number should be given to the next monitor?
3081 3092 int GraphKit::next_monitor() {
3082 3093 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3083 3094 int next = current + C->sync_stack_slots();
3084 3095 // Keep the toplevel high water mark current:
3085 3096 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3086 3097 return current;
3087 3098 }
3088 3099
3089 3100 //------------------------------insert_mem_bar---------------------------------
3090 3101 // Memory barrier to avoid floating things around
3091 3102 // The membar serves as a pinch point between both control and all memory slices.
3092 3103 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3093 3104 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3094 3105 mb->init_req(TypeFunc::Control, control());
3095 3106 mb->init_req(TypeFunc::Memory, reset_memory());
3096 3107 Node* membar = _gvn.transform(mb);
3097 3108 set_control(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Control)));
3098 3109 set_all_memory_call(membar);
3099 3110 return membar;
3100 3111 }
3101 3112
3102 3113 //-------------------------insert_mem_bar_volatile----------------------------
3103 3114 // Memory barrier to avoid floating things around
3104 3115 // The membar serves as a pinch point between both control and memory(alias_idx).
3105 3116 // If you want to make a pinch point on all memory slices, do not use this
3106 3117 // function (even with AliasIdxBot); use insert_mem_bar() instead.
3107 3118 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
3108 3119 // When Parse::do_put_xxx updates a volatile field, it appends a series
3109 3120 // of MemBarVolatile nodes, one for *each* volatile field alias category.
3110 3121 // The first membar is on the same memory slice as the field store opcode.
3111 3122 // This forces the membar to follow the store. (Bug 6500685 broke this.)
3112 3123 // All the other membars (for other volatile slices, including AliasIdxBot,
3113 3124 // which stands for all unknown volatile slices) are control-dependent
3114 3125 // on the first membar. This prevents later volatile loads or stores
3115 3126 // from sliding up past the just-emitted store.
3116 3127
3117 3128 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
3118 3129 mb->set_req(TypeFunc::Control,control());
3119 3130 if (alias_idx == Compile::AliasIdxBot) {
3120 3131 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
3121 3132 } else {
3122 3133 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
3123 3134 mb->set_req(TypeFunc::Memory, memory(alias_idx));
3124 3135 }
3125 3136 Node* membar = _gvn.transform(mb);
3126 3137 set_control(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Control)));
3127 3138 if (alias_idx == Compile::AliasIdxBot) {
3128 3139 merged_memory()->set_base_memory(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Memory)));
3129 3140 } else {
3130 3141 set_memory(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Memory)),alias_idx);
3131 3142 }
3132 3143 return membar;
3133 3144 }
3134 3145
3135 3146 //------------------------------shared_lock------------------------------------
3136 3147 // Emit locking code.
3137 3148 FastLockNode* GraphKit::shared_lock(Node* obj) {
3138 3149 // bci is either a monitorenter bc or InvocationEntryBci
3139 3150 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3140 3151 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3141 3152
3142 3153 if( !GenerateSynchronizationCode )
3143 3154 return NULL; // Not locking things?
3144 3155 if (stopped()) // Dead monitor?
3145 3156 return NULL;
3146 3157
3147 3158 assert(dead_locals_are_killed(), "should kill locals before sync. point");
3148 3159
3149 3160 // Box the stack location
3150 3161 Node* box = _gvn.transform(new (C) BoxLockNode(next_monitor()));
3151 3162 Node* mem = reset_memory();
3152 3163
3153 3164 FastLockNode * flock = _gvn.transform(new (C) FastLockNode(0, obj, box) )->as_FastLock();
3154 3165 if (PrintPreciseBiasedLockingStatistics) {
3155 3166 // Create the counters for this fast lock.
3156 3167 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3157 3168 }
3158 3169 // Add monitor to debug info for the slow path. If we block inside the
3159 3170 // slow path and de-opt, we need the monitor hanging around
3160 3171 map()->push_monitor( flock );
3161 3172
3162 3173 const TypeFunc *tf = LockNode::lock_type();
3163 3174 LockNode *lock = new (C) LockNode(C, tf);
3164 3175
3165 3176 lock->init_req( TypeFunc::Control, control() );
3166 3177 lock->init_req( TypeFunc::Memory , mem );
3167 3178 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3168 3179 lock->init_req( TypeFunc::FramePtr, frameptr() );
3169 3180 lock->init_req( TypeFunc::ReturnAdr, top() );
3170 3181
3171 3182 lock->init_req(TypeFunc::Parms + 0, obj);
3172 3183 lock->init_req(TypeFunc::Parms + 1, box);
3173 3184 lock->init_req(TypeFunc::Parms + 2, flock);
3174 3185 add_safepoint_edges(lock);
3175 3186
3176 3187 lock = _gvn.transform( lock )->as_Lock();
3177 3188
3178 3189 // lock has no side-effects, sets few values
3179 3190 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
3180 3191
3181 3192 insert_mem_bar(Op_MemBarAcquireLock);
3182 3193
3183 3194 // Add this to the worklist so that the lock can be eliminated
3184 3195 record_for_igvn(lock);
3185 3196
3186 3197 #ifndef PRODUCT
3187 3198 if (PrintLockStatistics) {
3188 3199 // Update the counter for this lock. Don't bother using an atomic
3189 3200 // operation since we don't require absolute accuracy.
3190 3201 lock->create_lock_counter(map()->jvms());
3191 3202 increment_counter(lock->counter()->addr());
3192 3203 }
3193 3204 #endif
3194 3205
3195 3206 return flock;
3196 3207 }
3197 3208
3198 3209
3199 3210 //------------------------------shared_unlock----------------------------------
3200 3211 // Emit unlocking code.
3201 3212 void GraphKit::shared_unlock(Node* box, Node* obj) {
3202 3213 // bci is either a monitorenter bc or InvocationEntryBci
3203 3214 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3204 3215 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3205 3216
3206 3217 if( !GenerateSynchronizationCode )
3207 3218 return;
3208 3219 if (stopped()) { // Dead monitor?
3209 3220 map()->pop_monitor(); // Kill monitor from debug info
3210 3221 return;
3211 3222 }
3212 3223
3213 3224 // Memory barrier to avoid floating things down past the locked region
3214 3225 insert_mem_bar(Op_MemBarReleaseLock);
3215 3226
3216 3227 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3217 3228 UnlockNode *unlock = new (C) UnlockNode(C, tf);
3218 3229 uint raw_idx = Compile::AliasIdxRaw;
3219 3230 unlock->init_req( TypeFunc::Control, control() );
3220 3231 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3221 3232 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3222 3233 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3223 3234 unlock->init_req( TypeFunc::ReturnAdr, top() );
3224 3235
3225 3236 unlock->init_req(TypeFunc::Parms + 0, obj);
3226 3237 unlock->init_req(TypeFunc::Parms + 1, box);
3227 3238 unlock = _gvn.transform(unlock)->as_Unlock();
3228 3239
3229 3240 Node* mem = reset_memory();
3230 3241
3231 3242 // unlock has no side-effects, sets few values
3232 3243 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3233 3244
3234 3245 // Kill monitor from debug info
3235 3246 map()->pop_monitor( );
3236 3247 }
3237 3248
3238 3249 //-------------------------------get_layout_helper-----------------------------
3239 3250 // If the given klass is a constant or known to be an array,
3240 3251 // fetch the constant layout helper value into constant_value
3241 3252 // and return (Node*)NULL. Otherwise, load the non-constant
3242 3253 // layout helper value, and return the node which represents it.
3243 3254 // This two-faced routine is useful because allocation sites
3244 3255 // almost always feature constant types.
3245 3256 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3246 3257 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
3247 3258 if (!StressReflectiveCode && inst_klass != NULL) {
3248 3259 ciKlass* klass = inst_klass->klass();
3249 3260 bool xklass = inst_klass->klass_is_exact();
3250 3261 if (xklass || klass->is_array_klass()) {
3251 3262 jint lhelper = klass->layout_helper();
3252 3263 if (lhelper != Klass::_lh_neutral_value) {
3253 3264 constant_value = lhelper;
3254 3265 return (Node*) NULL;
3255 3266 }
3256 3267 }
3257 3268 }
3258 3269 constant_value = Klass::_lh_neutral_value; // put in a known value
3259 3270 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3260 3271 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3261 3272 }
3262 3273
3263 3274 // We just put in an allocate/initialize with a big raw-memory effect.
3264 3275 // Hook selected additional alias categories on the initialization.
3265 3276 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3266 3277 MergeMemNode* init_in_merge,
3267 3278 Node* init_out_raw) {
3268 3279 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3269 3280 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3270 3281
3271 3282 Node* prevmem = kit.memory(alias_idx);
3272 3283 init_in_merge->set_memory_at(alias_idx, prevmem);
3273 3284 kit.set_memory(init_out_raw, alias_idx);
3274 3285 }
3275 3286
3276 3287 //---------------------------set_output_for_allocation-------------------------
3277 3288 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3278 3289 const TypeOopPtr* oop_type) {
3279 3290 int rawidx = Compile::AliasIdxRaw;
3280 3291 alloc->set_req( TypeFunc::FramePtr, frameptr() );
3281 3292 add_safepoint_edges(alloc);
3282 3293 Node* allocx = _gvn.transform(alloc);
3283 3294 set_control( _gvn.transform(new (C) ProjNode(allocx, TypeFunc::Control) ) );
3284 3295 // create memory projection for i_o
3285 3296 set_memory ( _gvn.transform( new (C) ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3286 3297 make_slow_call_ex(allocx, env()->Throwable_klass(), true);
3287 3298
3288 3299 // create a memory projection as for the normal control path
3289 3300 Node* malloc = _gvn.transform(new (C) ProjNode(allocx, TypeFunc::Memory));
3290 3301 set_memory(malloc, rawidx);
3291 3302
3292 3303 // a normal slow-call doesn't change i_o, but an allocation does
3293 3304 // we create a separate i_o projection for the normal control path
3294 3305 set_i_o(_gvn.transform( new (C) ProjNode(allocx, TypeFunc::I_O, false) ) );
3295 3306 Node* rawoop = _gvn.transform( new (C) ProjNode(allocx, TypeFunc::Parms) );
3296 3307
3297 3308 // put in an initialization barrier
3298 3309 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3299 3310 rawoop)->as_Initialize();
3300 3311 assert(alloc->initialization() == init, "2-way macro link must work");
3301 3312 assert(init ->allocation() == alloc, "2-way macro link must work");
3302 3313 {
3303 3314 // Extract memory strands which may participate in the new object's
3304 3315 // initialization, and source them from the new InitializeNode.
3305 3316 // This will allow us to observe initializations when they occur,
3306 3317 // and link them properly (as a group) to the InitializeNode.
3307 3318 assert(init->in(InitializeNode::Memory) == malloc, "");
3308 3319 MergeMemNode* minit_in = MergeMemNode::make(C, malloc);
3309 3320 init->set_req(InitializeNode::Memory, minit_in);
3310 3321 record_for_igvn(minit_in); // fold it up later, if possible
3311 3322 Node* minit_out = memory(rawidx);
3312 3323 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3313 3324 if (oop_type->isa_aryptr()) {
3314 3325 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3315 3326 int elemidx = C->get_alias_index(telemref);
3316 3327 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3317 3328 } else if (oop_type->isa_instptr()) {
3318 3329 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
3319 3330 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3320 3331 ciField* field = ik->nonstatic_field_at(i);
3321 3332 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3322 3333 continue; // do not bother to track really large numbers of fields
3323 3334 // Find (or create) the alias category for this field:
3324 3335 int fieldidx = C->alias_type(field)->index();
3325 3336 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3326 3337 }
3327 3338 }
3328 3339 }
3329 3340
3330 3341 // Cast raw oop to the real thing...
3331 3342 Node* javaoop = new (C) CheckCastPPNode(control(), rawoop, oop_type);
3332 3343 javaoop = _gvn.transform(javaoop);
3333 3344 C->set_recent_alloc(control(), javaoop);
3334 3345 assert(just_allocated_object(control()) == javaoop, "just allocated");
3335 3346
3336 3347 #ifdef ASSERT
3337 3348 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3338 3349 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
3339 3350 "Ideal_allocation works");
3340 3351 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
3341 3352 "Ideal_allocation works");
3342 3353 if (alloc->is_AllocateArray()) {
3343 3354 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
3344 3355 "Ideal_allocation works");
3345 3356 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
3346 3357 "Ideal_allocation works");
3347 3358 } else {
3348 3359 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3349 3360 }
3350 3361 }
3351 3362 #endif //ASSERT
3352 3363
3353 3364 return javaoop;
3354 3365 }
3355 3366
3356 3367 //---------------------------new_instance--------------------------------------
3357 3368 // This routine takes a klass_node which may be constant (for a static type)
3358 3369 // or may be non-constant (for reflective code). It will work equally well
3359 3370 // for either, and the graph will fold nicely if the optimizer later reduces
3360 3371 // the type to a constant.
3361 3372 // The optional arguments are for specialized use by intrinsics:
3362 3373 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3363 3374 // - If 'return_size_val', report the the total object size to the caller.
3364 3375 Node* GraphKit::new_instance(Node* klass_node,
3365 3376 Node* extra_slow_test,
3366 3377 Node* *return_size_val) {
3367 3378 // Compute size in doublewords
3368 3379 // The size is always an integral number of doublewords, represented
3369 3380 // as a positive bytewise size stored in the klass's layout_helper.
3370 3381 // The layout_helper also encodes (in a low bit) the need for a slow path.
3371 3382 jint layout_con = Klass::_lh_neutral_value;
3372 3383 Node* layout_val = get_layout_helper(klass_node, layout_con);
3373 3384 int layout_is_con = (layout_val == NULL);
3374 3385
3375 3386 if (extra_slow_test == NULL) extra_slow_test = intcon(0);
3376 3387 // Generate the initial go-slow test. It's either ALWAYS (return a
3377 3388 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3378 3389 // case) a computed value derived from the layout_helper.
3379 3390 Node* initial_slow_test = NULL;
3380 3391 if (layout_is_con) {
3381 3392 assert(!StressReflectiveCode, "stress mode does not use these paths");
3382 3393 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3383 3394 initial_slow_test = must_go_slow? intcon(1): extra_slow_test;
3384 3395
3385 3396 } else { // reflective case
3386 3397 // This reflective path is used by Unsafe.allocateInstance.
3387 3398 // (It may be stress-tested by specifying StressReflectiveCode.)
3388 3399 // Basically, we want to get into the VM is there's an illegal argument.
3389 3400 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3390 3401 initial_slow_test = _gvn.transform( new (C) AndINode(layout_val, bit) );
3391 3402 if (extra_slow_test != intcon(0)) {
3392 3403 initial_slow_test = _gvn.transform( new (C) OrINode(initial_slow_test, extra_slow_test) );
3393 3404 }
3394 3405 // (Macro-expander will further convert this to a Bool, if necessary.)
3395 3406 }
3396 3407
3397 3408 // Find the size in bytes. This is easy; it's the layout_helper.
3398 3409 // The size value must be valid even if the slow path is taken.
3399 3410 Node* size = NULL;
3400 3411 if (layout_is_con) {
3401 3412 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
3402 3413 } else { // reflective case
3403 3414 // This reflective path is used by clone and Unsafe.allocateInstance.
3404 3415 size = ConvI2X(layout_val);
3405 3416
3406 3417 // Clear the low bits to extract layout_helper_size_in_bytes:
3407 3418 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3408 3419 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3409 3420 size = _gvn.transform( new (C) AndXNode(size, mask) );
3410 3421 }
3411 3422 if (return_size_val != NULL) {
3412 3423 (*return_size_val) = size;
3413 3424 }
3414 3425
3415 3426 // This is a precise notnull oop of the klass.
3416 3427 // (Actually, it need not be precise if this is a reflective allocation.)
3417 3428 // It's what we cast the result to.
3418 3429 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3419 3430 if (!tklass) tklass = TypeKlassPtr::OBJECT;
3420 3431 const TypeOopPtr* oop_type = tklass->as_instance_type();
3421 3432
3422 3433 // Now generate allocation code
3423 3434
3424 3435 // The entire memory state is needed for slow path of the allocation
3425 3436 // since GC and deoptimization can happened.
3426 3437 Node *mem = reset_memory();
3427 3438 set_all_memory(mem); // Create new memory state
3428 3439
3429 3440 AllocateNode* alloc
3430 3441 = new (C) AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3431 3442 control(), mem, i_o(),
3432 3443 size, klass_node,
3433 3444 initial_slow_test);
3434 3445
3435 3446 return set_output_for_allocation(alloc, oop_type);
3436 3447 }
3437 3448
3438 3449 //-------------------------------new_array-------------------------------------
3439 3450 // helper for both newarray and anewarray
3440 3451 // The 'length' parameter is (obviously) the length of the array.
3441 3452 // See comments on new_instance for the meaning of the other arguments.
3442 3453 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
3443 3454 Node* length, // number of array elements
3444 3455 int nargs, // number of arguments to push back for uncommon trap
3445 3456 Node* *return_size_val) {
3446 3457 jint layout_con = Klass::_lh_neutral_value;
3447 3458 Node* layout_val = get_layout_helper(klass_node, layout_con);
3448 3459 int layout_is_con = (layout_val == NULL);
3449 3460
3450 3461 if (!layout_is_con && !StressReflectiveCode &&
3451 3462 !too_many_traps(Deoptimization::Reason_class_check)) {
3452 3463 // This is a reflective array creation site.
3453 3464 // Optimistically assume that it is a subtype of Object[],
3454 3465 // so that we can fold up all the address arithmetic.
3455 3466 layout_con = Klass::array_layout_helper(T_OBJECT);
3456 3467 Node* cmp_lh = _gvn.transform( new(C) CmpINode(layout_val, intcon(layout_con)) );
3457 3468 Node* bol_lh = _gvn.transform( new(C) BoolNode(cmp_lh, BoolTest::eq) );
3458 3469 { BuildCutout unless(this, bol_lh, PROB_MAX);
3459 3470 inc_sp(nargs);
3460 3471 uncommon_trap(Deoptimization::Reason_class_check,
3461 3472 Deoptimization::Action_maybe_recompile);
3462 3473 }
3463 3474 layout_val = NULL;
3464 3475 layout_is_con = true;
3465 3476 }
3466 3477
3467 3478 // Generate the initial go-slow test. Make sure we do not overflow
3468 3479 // if length is huge (near 2Gig) or negative! We do not need
3469 3480 // exact double-words here, just a close approximation of needed
3470 3481 // double-words. We can't add any offset or rounding bits, lest we
3471 3482 // take a size -1 of bytes and make it positive. Use an unsigned
3472 3483 // compare, so negative sizes look hugely positive.
3473 3484 int fast_size_limit = FastAllocateSizeLimit;
3474 3485 if (layout_is_con) {
3475 3486 assert(!StressReflectiveCode, "stress mode does not use these paths");
3476 3487 // Increase the size limit if we have exact knowledge of array type.
3477 3488 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3478 3489 fast_size_limit <<= (LogBytesPerLong - log2_esize);
3479 3490 }
3480 3491
3481 3492 Node* initial_slow_cmp = _gvn.transform( new (C) CmpUNode( length, intcon( fast_size_limit ) ) );
3482 3493 Node* initial_slow_test = _gvn.transform( new (C) BoolNode( initial_slow_cmp, BoolTest::gt ) );
3483 3494 if (initial_slow_test->is_Bool()) {
3484 3495 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3485 3496 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3486 3497 }
3487 3498
3488 3499 // --- Size Computation ---
3489 3500 // array_size = round_to_heap(array_header + (length << elem_shift));
3490 3501 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes)
3491 3502 // and round_to(x, y) == ((x + y-1) & ~(y-1))
3492 3503 // The rounding mask is strength-reduced, if possible.
3493 3504 int round_mask = MinObjAlignmentInBytes - 1;
3494 3505 Node* header_size = NULL;
3495 3506 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3496 3507 // (T_BYTE has the weakest alignment and size restrictions...)
3497 3508 if (layout_is_con) {
3498 3509 int hsize = Klass::layout_helper_header_size(layout_con);
3499 3510 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3500 3511 BasicType etype = Klass::layout_helper_element_type(layout_con);
3501 3512 if ((round_mask & ~right_n_bits(eshift)) == 0)
3502 3513 round_mask = 0; // strength-reduce it if it goes away completely
3503 3514 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3504 3515 assert(header_size_min <= hsize, "generic minimum is smallest");
3505 3516 header_size_min = hsize;
3506 3517 header_size = intcon(hsize + round_mask);
3507 3518 } else {
3508 3519 Node* hss = intcon(Klass::_lh_header_size_shift);
3509 3520 Node* hsm = intcon(Klass::_lh_header_size_mask);
3510 3521 Node* hsize = _gvn.transform( new(C) URShiftINode(layout_val, hss) );
3511 3522 hsize = _gvn.transform( new(C) AndINode(hsize, hsm) );
3512 3523 Node* mask = intcon(round_mask);
3513 3524 header_size = _gvn.transform( new(C) AddINode(hsize, mask) );
3514 3525 }
3515 3526
3516 3527 Node* elem_shift = NULL;
3517 3528 if (layout_is_con) {
3518 3529 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3519 3530 if (eshift != 0)
3520 3531 elem_shift = intcon(eshift);
3521 3532 } else {
3522 3533 // There is no need to mask or shift this value.
3523 3534 // The semantics of LShiftINode include an implicit mask to 0x1F.
3524 3535 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3525 3536 elem_shift = layout_val;
3526 3537 }
3527 3538
3528 3539 // Transition to native address size for all offset calculations:
3529 3540 Node* lengthx = ConvI2X(length);
3530 3541 Node* headerx = ConvI2X(header_size);
3531 3542 #ifdef _LP64
3532 3543 { const TypeLong* tllen = _gvn.find_long_type(lengthx);
3533 3544 if (tllen != NULL && tllen->_lo < 0) {
3534 3545 // Add a manual constraint to a positive range. Cf. array_element_address.
3535 3546 jlong size_max = arrayOopDesc::max_array_length(T_BYTE);
3536 3547 if (size_max > tllen->_hi) size_max = tllen->_hi;
3537 3548 const TypeLong* tlcon = TypeLong::make(CONST64(0), size_max, Type::WidenMin);
3538 3549 lengthx = _gvn.transform( new (C) ConvI2LNode(length, tlcon));
3539 3550 }
3540 3551 }
3541 3552 #endif
3542 3553
3543 3554 // Combine header size (plus rounding) and body size. Then round down.
3544 3555 // This computation cannot overflow, because it is used only in two
3545 3556 // places, one where the length is sharply limited, and the other
3546 3557 // after a successful allocation.
3547 3558 Node* abody = lengthx;
3548 3559 if (elem_shift != NULL)
3549 3560 abody = _gvn.transform( new(C) LShiftXNode(lengthx, elem_shift) );
3550 3561 Node* size = _gvn.transform( new(C) AddXNode(headerx, abody) );
3551 3562 if (round_mask != 0) {
3552 3563 Node* mask = MakeConX(~round_mask);
3553 3564 size = _gvn.transform( new(C) AndXNode(size, mask) );
3554 3565 }
3555 3566 // else if round_mask == 0, the size computation is self-rounding
3556 3567
3557 3568 if (return_size_val != NULL) {
3558 3569 // This is the size
3559 3570 (*return_size_val) = size;
3560 3571 }
3561 3572
3562 3573 // Now generate allocation code
3563 3574
3564 3575 // The entire memory state is needed for slow path of the allocation
3565 3576 // since GC and deoptimization can happened.
3566 3577 Node *mem = reset_memory();
3567 3578 set_all_memory(mem); // Create new memory state
3568 3579
3569 3580 // Create the AllocateArrayNode and its result projections
3570 3581 AllocateArrayNode* alloc
3571 3582 = new (C) AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
3572 3583 control(), mem, i_o(),
3573 3584 size, klass_node,
3574 3585 initial_slow_test,
3575 3586 length);
3576 3587
3577 3588 // Cast to correct type. Note that the klass_node may be constant or not,
3578 3589 // and in the latter case the actual array type will be inexact also.
3579 3590 // (This happens via a non-constant argument to inline_native_newArray.)
3580 3591 // In any case, the value of klass_node provides the desired array type.
3581 3592 const TypeInt* length_type = _gvn.find_int_type(length);
3582 3593 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3583 3594 if (ary_type->isa_aryptr() && length_type != NULL) {
3584 3595 // Try to get a better type than POS for the size
3585 3596 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3586 3597 }
3587 3598
3588 3599 Node* javaoop = set_output_for_allocation(alloc, ary_type);
3589 3600
3590 3601 // Cast length on remaining path to be as narrow as possible
3591 3602 if (map()->find_edge(length) >= 0) {
3592 3603 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
3593 3604 if (ccast != length) {
3594 3605 _gvn.set_type_bottom(ccast);
3595 3606 record_for_igvn(ccast);
3596 3607 replace_in_map(length, ccast);
3597 3608 }
3598 3609 }
3599 3610
3600 3611 return javaoop;
3601 3612 }
3602 3613
3603 3614 // The following "Ideal_foo" functions are placed here because they recognize
3604 3615 // the graph shapes created by the functions immediately above.
3605 3616
3606 3617 //---------------------------Ideal_allocation----------------------------------
3607 3618 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
3608 3619 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
3609 3620 if (ptr == NULL) { // reduce dumb test in callers
3610 3621 return NULL;
3611 3622 }
3612 3623 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
3613 3624 ptr = ptr->in(1);
3614 3625 if (ptr == NULL) return NULL;
3615 3626 }
3616 3627 // Return NULL for allocations with several casts:
3617 3628 // j.l.reflect.Array.newInstance(jobject, jint)
3618 3629 // Object.clone()
3619 3630 // to keep more precise type from last cast.
3620 3631 if (ptr->is_Proj()) {
3621 3632 Node* allo = ptr->in(0);
3622 3633 if (allo != NULL && allo->is_Allocate()) {
3623 3634 return allo->as_Allocate();
3624 3635 }
3625 3636 }
3626 3637 // Report failure to match.
3627 3638 return NULL;
3628 3639 }
3629 3640
3630 3641 // Fancy version which also strips off an offset (and reports it to caller).
3631 3642 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
3632 3643 intptr_t& offset) {
3633 3644 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
3634 3645 if (base == NULL) return NULL;
3635 3646 return Ideal_allocation(base, phase);
3636 3647 }
3637 3648
3638 3649 // Trace Initialize <- Proj[Parm] <- Allocate
3639 3650 AllocateNode* InitializeNode::allocation() {
3640 3651 Node* rawoop = in(InitializeNode::RawAddress);
3641 3652 if (rawoop->is_Proj()) {
3642 3653 Node* alloc = rawoop->in(0);
3643 3654 if (alloc->is_Allocate()) {
3644 3655 return alloc->as_Allocate();
3645 3656 }
3646 3657 }
3647 3658 return NULL;
3648 3659 }
3649 3660
3650 3661 // Trace Allocate -> Proj[Parm] -> Initialize
3651 3662 InitializeNode* AllocateNode::initialization() {
3652 3663 ProjNode* rawoop = proj_out(AllocateNode::RawAddress);
3653 3664 if (rawoop == NULL) return NULL;
3654 3665 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
3655 3666 Node* init = rawoop->fast_out(i);
3656 3667 if (init->is_Initialize()) {
3657 3668 assert(init->as_Initialize()->allocation() == this, "2-way link");
3658 3669 return init->as_Initialize();
3659 3670 }
3660 3671 }
3661 3672 return NULL;
3662 3673 }
3663 3674
3664 3675 //----------------------------- loop predicates ---------------------------
3665 3676
3666 3677 //------------------------------add_predicate_impl----------------------------
3667 3678 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) {
3668 3679 // Too many traps seen?
3669 3680 if (too_many_traps(reason)) {
3670 3681 #ifdef ASSERT
3671 3682 if (TraceLoopPredicate) {
3672 3683 int tc = C->trap_count(reason);
3673 3684 tty->print("too many traps=%s tcount=%d in ",
3674 3685 Deoptimization::trap_reason_name(reason), tc);
3675 3686 method()->print(); // which method has too many predicate traps
3676 3687 tty->cr();
3677 3688 }
3678 3689 #endif
3679 3690 // We cannot afford to take more traps here,
3680 3691 // do not generate predicate.
3681 3692 return;
3682 3693 }
3683 3694
3684 3695 Node *cont = _gvn.intcon(1);
3685 3696 Node* opq = _gvn.transform(new (C) Opaque1Node(C, cont));
3686 3697 Node *bol = _gvn.transform(new (C) Conv2BNode(opq));
3687 3698 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
3688 3699 Node* iffalse = _gvn.transform(new (C) IfFalseNode(iff));
3689 3700 C->add_predicate_opaq(opq);
3690 3701 {
3691 3702 PreserveJVMState pjvms(this);
3692 3703 set_control(iffalse);
3693 3704 inc_sp(nargs);
3694 3705 uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
3695 3706 }
3696 3707 Node* iftrue = _gvn.transform(new (C) IfTrueNode(iff));
3697 3708 set_control(iftrue);
3698 3709 }
3699 3710
3700 3711 //------------------------------add_predicate---------------------------------
3701 3712 void GraphKit::add_predicate(int nargs) {
3702 3713 if (UseLoopPredicate) {
3703 3714 add_predicate_impl(Deoptimization::Reason_predicate, nargs);
3704 3715 }
3705 3716 // loop's limit check predicate should be near the loop.
3706 3717 if (LoopLimitCheck) {
3707 3718 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs);
3708 3719 }
3709 3720 }
3710 3721
3711 3722 //----------------------------- store barriers ----------------------------
3712 3723 #define __ ideal.
3713 3724
3714 3725 void GraphKit::sync_kit(IdealKit& ideal) {
3715 3726 set_all_memory(__ merged_memory());
3716 3727 set_i_o(__ i_o());
3717 3728 set_control(__ ctrl());
3718 3729 }
3719 3730
3720 3731 void GraphKit::final_sync(IdealKit& ideal) {
3721 3732 // Final sync IdealKit and graphKit.
3722 3733 sync_kit(ideal);
3723 3734 }
3724 3735
3725 3736 // vanilla/CMS post barrier
3726 3737 // Insert a write-barrier store. This is to let generational GC work; we have
3727 3738 // to flag all oop-stores before the next GC point.
3728 3739 void GraphKit::write_barrier_post(Node* oop_store,
3729 3740 Node* obj,
3730 3741 Node* adr,
3731 3742 uint adr_idx,
3732 3743 Node* val,
3733 3744 bool use_precise) {
3734 3745 // No store check needed if we're storing a NULL or an old object
3735 3746 // (latter case is probably a string constant). The concurrent
3736 3747 // mark sweep garbage collector, however, needs to have all nonNull
3737 3748 // oop updates flagged via card-marks.
3738 3749 if (val != NULL && val->is_Con()) {
3739 3750 // must be either an oop or NULL
3740 3751 const Type* t = val->bottom_type();
3741 3752 if (t == TypePtr::NULL_PTR || t == Type::TOP)
3742 3753 // stores of null never (?) need barriers
3743 3754 return;
3744 3755 }
3745 3756
3746 3757 if (use_ReduceInitialCardMarks()
3747 3758 && obj == just_allocated_object(control())) {
3748 3759 // We can skip marks on a freshly-allocated object in Eden.
3749 3760 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp.
3750 3761 // That routine informs GC to take appropriate compensating steps,
3751 3762 // upon a slow-path allocation, so as to make this card-mark
3752 3763 // elision safe.
3753 3764 return;
3754 3765 }
3755 3766
3756 3767 if (!use_precise) {
3757 3768 // All card marks for a (non-array) instance are in one place:
3758 3769 adr = obj;
3759 3770 }
3760 3771 // (Else it's an array (or unknown), and we want more precise card marks.)
3761 3772 assert(adr != NULL, "");
3762 3773
3763 3774 IdealKit ideal(this, true);
3764 3775
3765 3776 // Convert the pointer to an int prior to doing math on it
3766 3777 Node* cast = __ CastPX(__ ctrl(), adr);
3767 3778
3768 3779 // Divide by card size
3769 3780 assert(Universe::heap()->barrier_set()->kind() == BarrierSet::CardTableModRef,
3770 3781 "Only one we handle so far.");
3771 3782 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
3772 3783
3773 3784 // Combine card table base and card offset
3774 3785 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset );
3775 3786
3776 3787 // Get the alias_index for raw card-mark memory
3777 3788 int adr_type = Compile::AliasIdxRaw;
3778 3789 Node* zero = __ ConI(0); // Dirty card value
3779 3790 BasicType bt = T_BYTE;
3780 3791
3781 3792 if (UseCondCardMark) {
3782 3793 // The classic GC reference write barrier is typically implemented
3783 3794 // as a store into the global card mark table. Unfortunately
3784 3795 // unconditional stores can result in false sharing and excessive
3785 3796 // coherence traffic as well as false transactional aborts.
3786 3797 // UseCondCardMark enables MP "polite" conditional card mark
3787 3798 // stores. In theory we could relax the load from ctrl() to
3788 3799 // no_ctrl, but that doesn't buy much latitude.
3789 3800 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type);
3790 3801 __ if_then(card_val, BoolTest::ne, zero);
3791 3802 }
3792 3803
3793 3804 // Smash zero into card
3794 3805 if( !UseConcMarkSweepGC ) {
3795 3806 __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::release);
3796 3807 } else {
3797 3808 // Specialized path for CM store barrier
3798 3809 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type);
3799 3810 }
3800 3811
3801 3812 if (UseCondCardMark) {
3802 3813 __ end_if();
3803 3814 }
3804 3815
3805 3816 // Final sync IdealKit and GraphKit.
3806 3817 final_sync(ideal);
3807 3818 }
3808 3819
3809 3820 // G1 pre/post barriers
3810 3821 void GraphKit::g1_write_barrier_pre(bool do_load,
3811 3822 Node* obj,
3812 3823 Node* adr,
3813 3824 uint alias_idx,
3814 3825 Node* val,
3815 3826 const TypeOopPtr* val_type,
3816 3827 Node* pre_val,
3817 3828 BasicType bt) {
3818 3829
3819 3830 // Some sanity checks
3820 3831 // Note: val is unused in this routine.
3821 3832
3822 3833 if (do_load) {
3823 3834 // We need to generate the load of the previous value
3824 3835 assert(obj != NULL, "must have a base");
3825 3836 assert(adr != NULL, "where are loading from?");
3826 3837 assert(pre_val == NULL, "loaded already?");
3827 3838 assert(val_type != NULL, "need a type");
3828 3839 } else {
3829 3840 // In this case both val_type and alias_idx are unused.
3830 3841 assert(pre_val != NULL, "must be loaded already");
3831 3842 // Nothing to be done if pre_val is null.
3832 3843 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return;
3833 3844 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
3834 3845 }
3835 3846 assert(bt == T_OBJECT, "or we shouldn't be here");
3836 3847
3837 3848 IdealKit ideal(this, true);
3838 3849
3839 3850 Node* tls = __ thread(); // ThreadLocalStorage
3840 3851
3841 3852 Node* no_ctrl = NULL;
3842 3853 Node* no_base = __ top();
3843 3854 Node* zero = __ ConI(0);
3844 3855 Node* zeroX = __ ConX(0);
3845 3856
3846 3857 float likely = PROB_LIKELY(0.999);
3847 3858 float unlikely = PROB_UNLIKELY(0.999);
3848 3859
3849 3860 BasicType active_type = in_bytes(PtrQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE;
3850 3861 assert(in_bytes(PtrQueue::byte_width_of_active()) == 4 || in_bytes(PtrQueue::byte_width_of_active()) == 1, "flag width");
3851 3862
3852 3863 // Offsets into the thread
3853 3864 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648
3854 3865 PtrQueue::byte_offset_of_active());
3855 3866 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656
3856 3867 PtrQueue::byte_offset_of_index());
3857 3868 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652
3858 3869 PtrQueue::byte_offset_of_buf());
3859 3870
3860 3871 // Now the actual pointers into the thread
3861 3872 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset));
3862 3873 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
3863 3874 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
3864 3875
3865 3876 // Now some of the values
3866 3877 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw);
3867 3878
3868 3879 // if (!marking)
3869 3880 __ if_then(marking, BoolTest::ne, zero, unlikely); {
3870 3881 BasicType index_bt = TypeX_X->basic_type();
3871 3882 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 PtrQueue::_index with wrong size.");
3872 3883 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
3873 3884
3874 3885 if (do_load) {
3875 3886 // load original value
3876 3887 // alias_idx correct??
3877 3888 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx);
3878 3889 }
3879 3890
3880 3891 // if (pre_val != NULL)
3881 3892 __ if_then(pre_val, BoolTest::ne, null()); {
3882 3893 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
3883 3894
3884 3895 // is the queue for this thread full?
3885 3896 __ if_then(index, BoolTest::ne, zeroX, likely); {
3886 3897
3887 3898 // decrement the index
3888 3899 Node* next_index = _gvn.transform(new (C) SubXNode(index, __ ConX(sizeof(intptr_t))));
3889 3900
3890 3901 // Now get the buffer location we will log the previous value into and store it
3891 3902 Node *log_addr = __ AddP(no_base, buffer, next_index);
3892 3903 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered);
3893 3904 // update the index
3894 3905 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered);
3895 3906
3896 3907 } __ else_(); {
3897 3908
3898 3909 // logging buffer is full, call the runtime
3899 3910 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type();
3900 3911 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls);
3901 3912 } __ end_if(); // (!index)
3902 3913 } __ end_if(); // (pre_val != NULL)
3903 3914 } __ end_if(); // (!marking)
3904 3915
3905 3916 // Final sync IdealKit and GraphKit.
3906 3917 final_sync(ideal);
3907 3918 }
3908 3919
3909 3920 //
3910 3921 // Update the card table and add card address to the queue
3911 3922 //
3912 3923 void GraphKit::g1_mark_card(IdealKit& ideal,
3913 3924 Node* card_adr,
3914 3925 Node* oop_store,
3915 3926 uint oop_alias_idx,
3916 3927 Node* index,
3917 3928 Node* index_adr,
3918 3929 Node* buffer,
3919 3930 const TypeFunc* tf) {
3920 3931
3921 3932 Node* zero = __ ConI(0);
3922 3933 Node* zeroX = __ ConX(0);
3923 3934 Node* no_base = __ top();
3924 3935 BasicType card_bt = T_BYTE;
3925 3936 // Smash zero into card. MUST BE ORDERED WRT TO STORE
3926 3937 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw);
3927 3938
3928 3939 // Now do the queue work
3929 3940 __ if_then(index, BoolTest::ne, zeroX); {
3930 3941
3931 3942 Node* next_index = _gvn.transform(new (C) SubXNode(index, __ ConX(sizeof(intptr_t))));
3932 3943 Node* log_addr = __ AddP(no_base, buffer, next_index);
3933 3944
3934 3945 // Order, see storeCM.
3935 3946 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered);
3936 3947 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered);
3937 3948
3938 3949 } __ else_(); {
3939 3950 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread());
3940 3951 } __ end_if();
3941 3952
3942 3953 }
3943 3954
3944 3955 void GraphKit::g1_write_barrier_post(Node* oop_store,
3945 3956 Node* obj,
3946 3957 Node* adr,
3947 3958 uint alias_idx,
3948 3959 Node* val,
3949 3960 BasicType bt,
3950 3961 bool use_precise) {
3951 3962 // If we are writing a NULL then we need no post barrier
3952 3963
3953 3964 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) {
3954 3965 // Must be NULL
3955 3966 const Type* t = val->bottom_type();
3956 3967 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL");
3957 3968 // No post barrier if writing NULLx
3958 3969 return;
3959 3970 }
3960 3971
3961 3972 if (!use_precise) {
3962 3973 // All card marks for a (non-array) instance are in one place:
3963 3974 adr = obj;
3964 3975 }
3965 3976 // (Else it's an array (or unknown), and we want more precise card marks.)
3966 3977 assert(adr != NULL, "");
3967 3978
3968 3979 IdealKit ideal(this, true);
3969 3980
3970 3981 Node* tls = __ thread(); // ThreadLocalStorage
3971 3982
3972 3983 Node* no_base = __ top();
3973 3984 float likely = PROB_LIKELY(0.999);
3974 3985 float unlikely = PROB_UNLIKELY(0.999);
3975 3986 Node* young_card = __ ConI((jint)G1SATBCardTableModRefBS::g1_young_card_val());
3976 3987 Node* dirty_card = __ ConI((jint)CardTableModRefBS::dirty_card_val());
3977 3988 Node* zeroX = __ ConX(0);
3978 3989
3979 3990 // Get the alias_index for raw card-mark memory
3980 3991 const TypePtr* card_type = TypeRawPtr::BOTTOM;
3981 3992
3982 3993 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type();
3983 3994
3984 3995 // Offsets into the thread
3985 3996 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
3986 3997 PtrQueue::byte_offset_of_index());
3987 3998 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
3988 3999 PtrQueue::byte_offset_of_buf());
3989 4000
3990 4001 // Pointers into the thread
3991 4002
3992 4003 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
3993 4004 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
3994 4005
3995 4006 // Now some values
3996 4007 // Use ctrl to avoid hoisting these values past a safepoint, which could
3997 4008 // potentially reset these fields in the JavaThread.
3998 4009 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw);
3999 4010 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
4000 4011
4001 4012 // Convert the store obj pointer to an int prior to doing math on it
4002 4013 // Must use ctrl to prevent "integerized oop" existing across safepoint
4003 4014 Node* cast = __ CastPX(__ ctrl(), adr);
4004 4015
4005 4016 // Divide pointer by card size
4006 4017 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
4007 4018
4008 4019 // Combine card table base and card offset
4009 4020 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset );
4010 4021
4011 4022 // If we know the value being stored does it cross regions?
4012 4023
4013 4024 if (val != NULL) {
4014 4025 // Does the store cause us to cross regions?
4015 4026
4016 4027 // Should be able to do an unsigned compare of region_size instead of
4017 4028 // and extra shift. Do we have an unsigned compare??
4018 4029 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes);
4019 4030 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes));
4020 4031
4021 4032 // if (xor_res == 0) same region so skip
4022 4033 __ if_then(xor_res, BoolTest::ne, zeroX); {
4023 4034
4024 4035 // No barrier if we are storing a NULL
4025 4036 __ if_then(val, BoolTest::ne, null(), unlikely); {
4026 4037
4027 4038 // Ok must mark the card if not already dirty
4028 4039
4029 4040 // load the original value of the card
4030 4041 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4031 4042
4032 4043 __ if_then(card_val, BoolTest::ne, young_card); {
4033 4044 sync_kit(ideal);
4034 4045 // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier.
4035 4046 insert_mem_bar(Op_MemBarVolatile, oop_store);
4036 4047 __ sync_kit(this);
4037 4048
4038 4049 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4039 4050 __ if_then(card_val_reload, BoolTest::ne, dirty_card); {
4040 4051 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4041 4052 } __ end_if();
4042 4053 } __ end_if();
4043 4054 } __ end_if();
4044 4055 } __ end_if();
4045 4056 } else {
4046 4057 // Object.clone() instrinsic uses this path.
4047 4058 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4048 4059 }
4049 4060
4050 4061 // Final sync IdealKit and GraphKit.
4051 4062 final_sync(ideal);
4052 4063 }
4053 4064 #undef __
4054 4065
4055 4066
4056 4067
4057 4068 Node* GraphKit::load_String_offset(Node* ctrl, Node* str) {
4058 4069 if (java_lang_String::has_offset_field()) {
4059 4070 int offset_offset = java_lang_String::offset_offset_in_bytes();
4060 4071 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4061 4072 false, NULL, 0);
4062 4073 const TypePtr* offset_field_type = string_type->add_offset(offset_offset);
4063 4074 int offset_field_idx = C->get_alias_index(offset_field_type);
4064 4075 return make_load(ctrl,
4065 4076 basic_plus_adr(str, str, offset_offset),
4066 4077 TypeInt::INT, T_INT, offset_field_idx, MemNode::unordered);
4067 4078 } else {
4068 4079 return intcon(0);
4069 4080 }
4070 4081 }
4071 4082
4072 4083 Node* GraphKit::load_String_length(Node* ctrl, Node* str) {
4073 4084 if (java_lang_String::has_count_field()) {
4074 4085 int count_offset = java_lang_String::count_offset_in_bytes();
4075 4086 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4076 4087 false, NULL, 0);
4077 4088 const TypePtr* count_field_type = string_type->add_offset(count_offset);
4078 4089 int count_field_idx = C->get_alias_index(count_field_type);
4079 4090 return make_load(ctrl,
4080 4091 basic_plus_adr(str, str, count_offset),
4081 4092 TypeInt::INT, T_INT, count_field_idx, MemNode::unordered);
4082 4093 } else {
4083 4094 return load_array_length(load_String_value(ctrl, str));
4084 4095 }
4085 4096 }
4086 4097
4087 4098 Node* GraphKit::load_String_value(Node* ctrl, Node* str) {
4088 4099 int value_offset = java_lang_String::value_offset_in_bytes();
4089 4100 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4090 4101 false, NULL, 0);
4091 4102 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4092 4103 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4093 4104 TypeAry::make(TypeInt::CHAR,TypeInt::POS),
4094 4105 ciTypeArrayKlass::make(T_CHAR), true, 0);
4095 4106 int value_field_idx = C->get_alias_index(value_field_type);
4096 4107 Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset),
4097 4108 value_type, T_OBJECT, value_field_idx, MemNode::unordered);
4098 4109 // String.value field is known to be @Stable.
4099 4110 if (UseImplicitStableValues) {
4100 4111 load = cast_array_to_stable(load, value_type);
4101 4112 }
4102 4113 return load;
4103 4114 }
4104 4115
4105 4116 void GraphKit::store_String_offset(Node* ctrl, Node* str, Node* value) {
4106 4117 int offset_offset = java_lang_String::offset_offset_in_bytes();
4107 4118 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4108 4119 false, NULL, 0);
4109 4120 const TypePtr* offset_field_type = string_type->add_offset(offset_offset);
4110 4121 int offset_field_idx = C->get_alias_index(offset_field_type);
4111 4122 store_to_memory(ctrl, basic_plus_adr(str, offset_offset),
4112 4123 value, T_INT, offset_field_idx, MemNode::unordered);
4113 4124 }
4114 4125
4115 4126 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) {
4116 4127 int value_offset = java_lang_String::value_offset_in_bytes();
4117 4128 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4118 4129 false, NULL, 0);
4119 4130 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4120 4131
4121 4132 store_oop_to_object(ctrl, str, basic_plus_adr(str, value_offset), value_field_type,
4122 4133 value, TypeAryPtr::CHARS, T_OBJECT, MemNode::unordered);
4123 4134 }
4124 4135
4125 4136 void GraphKit::store_String_length(Node* ctrl, Node* str, Node* value) {
4126 4137 int count_offset = java_lang_String::count_offset_in_bytes();
4127 4138 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4128 4139 false, NULL, 0);
4129 4140 const TypePtr* count_field_type = string_type->add_offset(count_offset);
4130 4141 int count_field_idx = C->get_alias_index(count_field_type);
4131 4142 store_to_memory(ctrl, basic_plus_adr(str, count_offset),
4132 4143 value, T_INT, count_field_idx, MemNode::unordered);
4133 4144 }
4134 4145
4135 4146 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) {
4136 4147 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity
4137 4148 // assumption of CCP analysis.
4138 4149 return _gvn.transform(new(C) CastPPNode(ary, ary_type->cast_to_stable(true)));
4139 4150 }
|
↓ open down ↓ |
3002 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX