Print this page
Split |
Close |
Expand all |
Collapse all |
--- old/src/share/vm/opto/memnode.cpp
+++ new/src/share/vm/opto/memnode.cpp
1 1 /*
2 2 * Copyright (c) 1997, 2011, 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 "classfile/systemDictionary.hpp"
27 27 #include "compiler/compileLog.hpp"
28 28 #include "memory/allocation.inline.hpp"
29 29 #include "oops/objArrayKlass.hpp"
30 30 #include "opto/addnode.hpp"
31 31 #include "opto/cfgnode.hpp"
32 32 #include "opto/compile.hpp"
33 33 #include "opto/connode.hpp"
34 34 #include "opto/loopnode.hpp"
35 35 #include "opto/machnode.hpp"
36 36 #include "opto/matcher.hpp"
37 37 #include "opto/memnode.hpp"
38 38 #include "opto/mulnode.hpp"
39 39 #include "opto/phaseX.hpp"
40 40 #include "opto/regmask.hpp"
41 41
42 42 // Portions of code courtesy of Clifford Click
43 43
44 44 // Optimization - Graph Style
45 45
46 46 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
47 47
48 48 //=============================================================================
49 49 uint MemNode::size_of() const { return sizeof(*this); }
50 50
51 51 const TypePtr *MemNode::adr_type() const {
52 52 Node* adr = in(Address);
53 53 const TypePtr* cross_check = NULL;
54 54 DEBUG_ONLY(cross_check = _adr_type);
55 55 return calculate_adr_type(adr->bottom_type(), cross_check);
56 56 }
57 57
58 58 #ifndef PRODUCT
59 59 void MemNode::dump_spec(outputStream *st) const {
60 60 if (in(Address) == NULL) return; // node is dead
61 61 #ifndef ASSERT
62 62 // fake the missing field
63 63 const TypePtr* _adr_type = NULL;
64 64 if (in(Address) != NULL)
65 65 _adr_type = in(Address)->bottom_type()->isa_ptr();
66 66 #endif
67 67 dump_adr_type(this, _adr_type, st);
68 68
69 69 Compile* C = Compile::current();
70 70 if( C->alias_type(_adr_type)->is_volatile() )
71 71 st->print(" Volatile!");
72 72 }
73 73
74 74 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
75 75 st->print(" @");
76 76 if (adr_type == NULL) {
77 77 st->print("NULL");
78 78 } else {
79 79 adr_type->dump_on(st);
80 80 Compile* C = Compile::current();
81 81 Compile::AliasType* atp = NULL;
82 82 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type);
83 83 if (atp == NULL)
84 84 st->print(", idx=?\?;");
85 85 else if (atp->index() == Compile::AliasIdxBot)
86 86 st->print(", idx=Bot;");
87 87 else if (atp->index() == Compile::AliasIdxTop)
88 88 st->print(", idx=Top;");
89 89 else if (atp->index() == Compile::AliasIdxRaw)
90 90 st->print(", idx=Raw;");
91 91 else {
92 92 ciField* field = atp->field();
93 93 if (field) {
94 94 st->print(", name=");
95 95 field->print_name_on(st);
96 96 }
97 97 st->print(", idx=%d;", atp->index());
98 98 }
99 99 }
100 100 }
101 101
102 102 extern void print_alias_types();
103 103
104 104 #endif
105 105
106 106 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) {
107 107 const TypeOopPtr *tinst = t_adr->isa_oopptr();
108 108 if (tinst == NULL || !tinst->is_known_instance_field())
109 109 return mchain; // don't try to optimize non-instance types
110 110 uint instance_id = tinst->instance_id();
111 111 Node *start_mem = phase->C->start()->proj_out(TypeFunc::Memory);
112 112 Node *prev = NULL;
113 113 Node *result = mchain;
114 114 while (prev != result) {
115 115 prev = result;
116 116 if (result == start_mem)
117 117 break; // hit one of our sentinels
118 118 // skip over a call which does not affect this memory slice
119 119 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
120 120 Node *proj_in = result->in(0);
121 121 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
122 122 break; // hit one of our sentinels
123 123 } else if (proj_in->is_Call()) {
124 124 CallNode *call = proj_in->as_Call();
125 125 if (!call->may_modify(t_adr, phase)) {
126 126 result = call->in(TypeFunc::Memory);
127 127 }
128 128 } else if (proj_in->is_Initialize()) {
129 129 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
130 130 // Stop if this is the initialization for the object instance which
131 131 // which contains this memory slice, otherwise skip over it.
132 132 if (alloc != NULL && alloc->_idx != instance_id) {
133 133 result = proj_in->in(TypeFunc::Memory);
134 134 }
135 135 } else if (proj_in->is_MemBar()) {
136 136 result = proj_in->in(TypeFunc::Memory);
137 137 } else {
138 138 assert(false, "unexpected projection");
139 139 }
140 140 } else if (result->is_ClearArray()) {
141 141 if (!ClearArrayNode::step_through(&result, instance_id, phase)) {
142 142 // Can not bypass initialization of the instance
143 143 // we are looking for.
144 144 break;
145 145 }
146 146 // Otherwise skip it (the call updated 'result' value).
147 147 } else if (result->is_MergeMem()) {
148 148 result = step_through_mergemem(phase, result->as_MergeMem(), t_adr, NULL, tty);
149 149 }
150 150 }
151 151 return result;
152 152 }
153 153
154 154 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) {
155 155 const TypeOopPtr *t_oop = t_adr->isa_oopptr();
156 156 bool is_instance = (t_oop != NULL) && t_oop->is_known_instance_field();
157 157 PhaseIterGVN *igvn = phase->is_IterGVN();
158 158 Node *result = mchain;
159 159 result = optimize_simple_memory_chain(result, t_adr, phase);
160 160 if (is_instance && igvn != NULL && result->is_Phi()) {
161 161 PhiNode *mphi = result->as_Phi();
162 162 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
163 163 const TypePtr *t = mphi->adr_type();
164 164 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ||
165 165 t->isa_oopptr() && !t->is_oopptr()->is_known_instance() &&
166 166 t->is_oopptr()->cast_to_exactness(true)
167 167 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
168 168 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop) {
169 169 // clone the Phi with our address type
170 170 result = mphi->split_out_instance(t_adr, igvn);
171 171 } else {
172 172 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
173 173 }
174 174 }
175 175 return result;
176 176 }
177 177
178 178 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
179 179 uint alias_idx = phase->C->get_alias_index(tp);
180 180 Node *mem = mmem;
181 181 #ifdef ASSERT
182 182 {
183 183 // Check that current type is consistent with the alias index used during graph construction
184 184 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
185 185 bool consistent = adr_check == NULL || adr_check->empty() ||
186 186 phase->C->must_alias(adr_check, alias_idx );
187 187 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
188 188 if( !consistent && adr_check != NULL && !adr_check->empty() &&
189 189 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
190 190 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
191 191 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
192 192 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
193 193 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
194 194 // don't assert if it is dead code.
195 195 consistent = true;
196 196 }
197 197 if( !consistent ) {
198 198 st->print("alias_idx==%d, adr_check==", alias_idx);
199 199 if( adr_check == NULL ) {
200 200 st->print("NULL");
201 201 } else {
202 202 adr_check->dump();
203 203 }
204 204 st->cr();
205 205 print_alias_types();
206 206 assert(consistent, "adr_check must match alias idx");
207 207 }
208 208 }
209 209 #endif
210 210 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
211 211 // means an array I have not precisely typed yet. Do not do any
212 212 // alias stuff with it any time soon.
213 213 const TypeOopPtr *toop = tp->isa_oopptr();
214 214 if( tp->base() != Type::AnyPtr &&
215 215 !(toop &&
216 216 toop->klass() != NULL &&
217 217 toop->klass()->is_java_lang_Object() &&
218 218 toop->offset() == Type::OffsetBot) ) {
219 219 // compress paths and change unreachable cycles to TOP
220 220 // If not, we can update the input infinitely along a MergeMem cycle
221 221 // Equivalent code in PhiNode::Ideal
222 222 Node* m = phase->transform(mmem);
223 223 // If transformed to a MergeMem, get the desired slice
224 224 // Otherwise the returned node represents memory for every slice
225 225 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
226 226 // Update input if it is progress over what we have now
227 227 }
228 228 return mem;
229 229 }
230 230
231 231 //--------------------------Ideal_common---------------------------------------
232 232 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
233 233 // Unhook non-raw memories from complete (macro-expanded) initializations.
234 234 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
235 235 // If our control input is a dead region, kill all below the region
236 236 Node *ctl = in(MemNode::Control);
237 237 if (ctl && remove_dead_region(phase, can_reshape))
238 238 return this;
239 239 ctl = in(MemNode::Control);
240 240 // Don't bother trying to transform a dead node
241 241 if( ctl && ctl->is_top() ) return NodeSentinel;
242 242
243 243 PhaseIterGVN *igvn = phase->is_IterGVN();
244 244 // Wait if control on the worklist.
245 245 if (ctl && can_reshape && igvn != NULL) {
246 246 Node* bol = NULL;
247 247 Node* cmp = NULL;
248 248 if (ctl->in(0)->is_If()) {
249 249 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
250 250 bol = ctl->in(0)->in(1);
251 251 if (bol->is_Bool())
252 252 cmp = ctl->in(0)->in(1)->in(1);
253 253 }
254 254 if (igvn->_worklist.member(ctl) ||
255 255 (bol != NULL && igvn->_worklist.member(bol)) ||
256 256 (cmp != NULL && igvn->_worklist.member(cmp)) ) {
257 257 // This control path may be dead.
258 258 // Delay this memory node transformation until the control is processed.
259 259 phase->is_IterGVN()->_worklist.push(this);
260 260 return NodeSentinel; // caller will return NULL
261 261 }
262 262 }
263 263 // Ignore if memory is dead, or self-loop
264 264 Node *mem = in(MemNode::Memory);
265 265 if( phase->type( mem ) == Type::TOP ) return NodeSentinel; // caller will return NULL
266 266 assert( mem != this, "dead loop in MemNode::Ideal" );
267 267
268 268 Node *address = in(MemNode::Address);
269 269 const Type *t_adr = phase->type( address );
270 270 if( t_adr == Type::TOP ) return NodeSentinel; // caller will return NULL
271 271
272 272 if( can_reshape && igvn != NULL &&
273 273 (igvn->_worklist.member(address) ||
274 274 igvn->_worklist.size() > 0 && (phase->type(address) != adr_type())) ) {
275 275 // The address's base and type may change when the address is processed.
276 276 // Delay this mem node transformation until the address is processed.
277 277 phase->is_IterGVN()->_worklist.push(this);
278 278 return NodeSentinel; // caller will return NULL
279 279 }
280 280
281 281 // Do NOT remove or optimize the next lines: ensure a new alias index
282 282 // is allocated for an oop pointer type before Escape Analysis.
283 283 // Note: C++ will not remove it since the call has side effect.
284 284 if ( t_adr->isa_oopptr() ) {
285 285 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
286 286 }
287 287
288 288 #ifdef ASSERT
289 289 Node* base = NULL;
290 290 if (address->is_AddP())
291 291 base = address->in(AddPNode::Base);
292 292 assert(base == NULL || t_adr->isa_rawptr() ||
293 293 !phase->type(base)->higher_equal(TypePtr::NULL_PTR), "NULL+offs not RAW address?");
294 294 #endif
295 295
296 296 // Avoid independent memory operations
297 297 Node* old_mem = mem;
298 298
299 299 // The code which unhooks non-raw memories from complete (macro-expanded)
300 300 // initializations was removed. After macro-expansion all stores catched
301 301 // by Initialize node became raw stores and there is no information
302 302 // which memory slices they modify. So it is unsafe to move any memory
303 303 // operation above these stores. Also in most cases hooked non-raw memories
304 304 // were already unhooked by using information from detect_ptr_independence()
305 305 // and find_previous_store().
306 306
307 307 if (mem->is_MergeMem()) {
308 308 MergeMemNode* mmem = mem->as_MergeMem();
309 309 const TypePtr *tp = t_adr->is_ptr();
310 310
311 311 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
312 312 }
313 313
314 314 if (mem != old_mem) {
315 315 set_req(MemNode::Memory, mem);
316 316 if (phase->type( mem ) == Type::TOP) return NodeSentinel;
317 317 return this;
318 318 }
319 319
320 320 // let the subclass continue analyzing...
321 321 return NULL;
322 322 }
323 323
324 324 // Helper function for proving some simple control dominations.
325 325 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
326 326 // Already assumes that 'dom' is available at 'sub', and that 'sub'
327 327 // is not a constant (dominated by the method's StartNode).
328 328 // Used by MemNode::find_previous_store to prove that the
329 329 // control input of a memory operation predates (dominates)
330 330 // an allocation it wants to look past.
331 331 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
332 332 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
333 333 return false; // Conservative answer for dead code
334 334
335 335 // Check 'dom'. Skip Proj and CatchProj nodes.
336 336 dom = dom->find_exact_control(dom);
337 337 if (dom == NULL || dom->is_top())
338 338 return false; // Conservative answer for dead code
339 339
340 340 if (dom == sub) {
341 341 // For the case when, for example, 'sub' is Initialize and the original
342 342 // 'dom' is Proj node of the 'sub'.
343 343 return false;
344 344 }
345 345
346 346 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)
347 347 return true;
348 348
349 349 // 'dom' dominates 'sub' if its control edge and control edges
350 350 // of all its inputs dominate or equal to sub's control edge.
351 351
352 352 // Currently 'sub' is either Allocate, Initialize or Start nodes.
353 353 // Or Region for the check in LoadNode::Ideal();
354 354 // 'sub' should have sub->in(0) != NULL.
355 355 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
356 356 sub->is_Region(), "expecting only these nodes");
357 357
358 358 // Get control edge of 'sub'.
359 359 Node* orig_sub = sub;
360 360 sub = sub->find_exact_control(sub->in(0));
361 361 if (sub == NULL || sub->is_top())
362 362 return false; // Conservative answer for dead code
363 363
364 364 assert(sub->is_CFG(), "expecting control");
365 365
366 366 if (sub == dom)
367 367 return true;
368 368
369 369 if (sub->is_Start() || sub->is_Root())
370 370 return false;
371 371
372 372 {
373 373 // Check all control edges of 'dom'.
374 374
375 375 ResourceMark rm;
376 376 Arena* arena = Thread::current()->resource_area();
377 377 Node_List nlist(arena);
378 378 Unique_Node_List dom_list(arena);
379 379
380 380 dom_list.push(dom);
381 381 bool only_dominating_controls = false;
382 382
383 383 for (uint next = 0; next < dom_list.size(); next++) {
384 384 Node* n = dom_list.at(next);
385 385 if (n == orig_sub)
386 386 return false; // One of dom's inputs dominated by sub.
387 387 if (!n->is_CFG() && n->pinned()) {
388 388 // Check only own control edge for pinned non-control nodes.
389 389 n = n->find_exact_control(n->in(0));
390 390 if (n == NULL || n->is_top())
391 391 return false; // Conservative answer for dead code
392 392 assert(n->is_CFG(), "expecting control");
393 393 dom_list.push(n);
394 394 } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
395 395 only_dominating_controls = true;
396 396 } else if (n->is_CFG()) {
397 397 if (n->dominates(sub, nlist))
398 398 only_dominating_controls = true;
399 399 else
400 400 return false;
401 401 } else {
402 402 // First, own control edge.
403 403 Node* m = n->find_exact_control(n->in(0));
404 404 if (m != NULL) {
405 405 if (m->is_top())
406 406 return false; // Conservative answer for dead code
407 407 dom_list.push(m);
408 408 }
409 409 // Now, the rest of edges.
410 410 uint cnt = n->req();
411 411 for (uint i = 1; i < cnt; i++) {
412 412 m = n->find_exact_control(n->in(i));
413 413 if (m == NULL || m->is_top())
414 414 continue;
415 415 dom_list.push(m);
416 416 }
417 417 }
418 418 }
419 419 return only_dominating_controls;
420 420 }
421 421 }
422 422
423 423 //---------------------detect_ptr_independence---------------------------------
424 424 // Used by MemNode::find_previous_store to prove that two base
425 425 // pointers are never equal.
426 426 // The pointers are accompanied by their associated allocations,
427 427 // if any, which have been previously discovered by the caller.
428 428 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
429 429 Node* p2, AllocateNode* a2,
430 430 PhaseTransform* phase) {
431 431 // Attempt to prove that these two pointers cannot be aliased.
432 432 // They may both manifestly be allocations, and they should differ.
433 433 // Or, if they are not both allocations, they can be distinct constants.
434 434 // Otherwise, one is an allocation and the other a pre-existing value.
435 435 if (a1 == NULL && a2 == NULL) { // neither an allocation
436 436 return (p1 != p2) && p1->is_Con() && p2->is_Con();
437 437 } else if (a1 != NULL && a2 != NULL) { // both allocations
438 438 return (a1 != a2);
439 439 } else if (a1 != NULL) { // one allocation a1
440 440 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
441 441 return all_controls_dominate(p2, a1);
442 442 } else { //(a2 != NULL) // one allocation a2
443 443 return all_controls_dominate(p1, a2);
444 444 }
445 445 return false;
446 446 }
447 447
448 448
449 449 // The logic for reordering loads and stores uses four steps:
450 450 // (a) Walk carefully past stores and initializations which we
451 451 // can prove are independent of this load.
452 452 // (b) Observe that the next memory state makes an exact match
453 453 // with self (load or store), and locate the relevant store.
454 454 // (c) Ensure that, if we were to wire self directly to the store,
455 455 // the optimizer would fold it up somehow.
456 456 // (d) Do the rewiring, and return, depending on some other part of
457 457 // the optimizer to fold up the load.
458 458 // This routine handles steps (a) and (b). Steps (c) and (d) are
459 459 // specific to loads and stores, so they are handled by the callers.
460 460 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
461 461 //
462 462 Node* MemNode::find_previous_store(PhaseTransform* phase) {
463 463 Node* ctrl = in(MemNode::Control);
464 464 Node* adr = in(MemNode::Address);
465 465 intptr_t offset = 0;
466 466 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
467 467 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
468 468
469 469 if (offset == Type::OffsetBot)
470 470 return NULL; // cannot unalias unless there are precise offsets
471 471
472 472 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
473 473
474 474 intptr_t size_in_bytes = memory_size();
475 475
476 476 Node* mem = in(MemNode::Memory); // start searching here...
477 477
478 478 int cnt = 50; // Cycle limiter
479 479 for (;;) { // While we can dance past unrelated stores...
480 480 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
481 481
482 482 if (mem->is_Store()) {
483 483 Node* st_adr = mem->in(MemNode::Address);
484 484 intptr_t st_offset = 0;
485 485 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
486 486 if (st_base == NULL)
487 487 break; // inscrutable pointer
488 488 if (st_offset != offset && st_offset != Type::OffsetBot) {
489 489 const int MAX_STORE = BytesPerLong;
490 490 if (st_offset >= offset + size_in_bytes ||
491 491 st_offset <= offset - MAX_STORE ||
492 492 st_offset <= offset - mem->as_Store()->memory_size()) {
493 493 // Success: The offsets are provably independent.
494 494 // (You may ask, why not just test st_offset != offset and be done?
495 495 // The answer is that stores of different sizes can co-exist
496 496 // in the same sequence of RawMem effects. We sometimes initialize
497 497 // a whole 'tile' of array elements with a single jint or jlong.)
498 498 mem = mem->in(MemNode::Memory);
499 499 continue; // (a) advance through independent store memory
500 500 }
501 501 }
502 502 if (st_base != base &&
503 503 detect_ptr_independence(base, alloc,
504 504 st_base,
505 505 AllocateNode::Ideal_allocation(st_base, phase),
506 506 phase)) {
507 507 // Success: The bases are provably independent.
508 508 mem = mem->in(MemNode::Memory);
509 509 continue; // (a) advance through independent store memory
510 510 }
511 511
512 512 // (b) At this point, if the bases or offsets do not agree, we lose,
513 513 // since we have not managed to prove 'this' and 'mem' independent.
514 514 if (st_base == base && st_offset == offset) {
515 515 return mem; // let caller handle steps (c), (d)
516 516 }
517 517
518 518 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
519 519 InitializeNode* st_init = mem->in(0)->as_Initialize();
520 520 AllocateNode* st_alloc = st_init->allocation();
521 521 if (st_alloc == NULL)
522 522 break; // something degenerated
523 523 bool known_identical = false;
524 524 bool known_independent = false;
525 525 if (alloc == st_alloc)
526 526 known_identical = true;
527 527 else if (alloc != NULL)
528 528 known_independent = true;
529 529 else if (all_controls_dominate(this, st_alloc))
530 530 known_independent = true;
531 531
532 532 if (known_independent) {
533 533 // The bases are provably independent: Either they are
534 534 // manifestly distinct allocations, or else the control
535 535 // of this load dominates the store's allocation.
536 536 int alias_idx = phase->C->get_alias_index(adr_type());
537 537 if (alias_idx == Compile::AliasIdxRaw) {
538 538 mem = st_alloc->in(TypeFunc::Memory);
539 539 } else {
540 540 mem = st_init->memory(alias_idx);
541 541 }
542 542 continue; // (a) advance through independent store memory
543 543 }
544 544
545 545 // (b) at this point, if we are not looking at a store initializing
546 546 // the same allocation we are loading from, we lose.
547 547 if (known_identical) {
548 548 // From caller, can_see_stored_value will consult find_captured_store.
549 549 return mem; // let caller handle steps (c), (d)
550 550 }
551 551
552 552 } else if (addr_t != NULL && addr_t->is_known_instance_field()) {
553 553 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
554 554 if (mem->is_Proj() && mem->in(0)->is_Call()) {
555 555 CallNode *call = mem->in(0)->as_Call();
556 556 if (!call->may_modify(addr_t, phase)) {
557 557 mem = call->in(TypeFunc::Memory);
558 558 continue; // (a) advance through independent call memory
559 559 }
560 560 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
561 561 mem = mem->in(0)->in(TypeFunc::Memory);
562 562 continue; // (a) advance through independent MemBar memory
563 563 } else if (mem->is_ClearArray()) {
564 564 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) {
565 565 // (the call updated 'mem' value)
566 566 continue; // (a) advance through independent allocation memory
567 567 } else {
568 568 // Can not bypass initialization of the instance
569 569 // we are looking for.
570 570 return mem;
571 571 }
572 572 } else if (mem->is_MergeMem()) {
573 573 int alias_idx = phase->C->get_alias_index(adr_type());
574 574 mem = mem->as_MergeMem()->memory_at(alias_idx);
575 575 continue; // (a) advance through independent MergeMem memory
576 576 }
577 577 }
578 578
579 579 // Unless there is an explicit 'continue', we must bail out here,
580 580 // because 'mem' is an inscrutable memory state (e.g., a call).
581 581 break;
582 582 }
583 583
584 584 return NULL; // bail out
585 585 }
586 586
587 587 //----------------------calculate_adr_type-------------------------------------
588 588 // Helper function. Notices when the given type of address hits top or bottom.
589 589 // Also, asserts a cross-check of the type against the expected address type.
590 590 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
591 591 if (t == Type::TOP) return NULL; // does not touch memory any more?
592 592 #ifdef PRODUCT
593 593 cross_check = NULL;
594 594 #else
595 595 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL;
596 596 #endif
597 597 const TypePtr* tp = t->isa_ptr();
598 598 if (tp == NULL) {
599 599 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
600 600 return TypePtr::BOTTOM; // touches lots of memory
601 601 } else {
602 602 #ifdef ASSERT
603 603 // %%%% [phh] We don't check the alias index if cross_check is
604 604 // TypeRawPtr::BOTTOM. Needs to be investigated.
605 605 if (cross_check != NULL &&
606 606 cross_check != TypePtr::BOTTOM &&
607 607 cross_check != TypeRawPtr::BOTTOM) {
608 608 // Recheck the alias index, to see if it has changed (due to a bug).
609 609 Compile* C = Compile::current();
610 610 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
611 611 "must stay in the original alias category");
612 612 // The type of the address must be contained in the adr_type,
613 613 // disregarding "null"-ness.
614 614 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
615 615 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
616 616 assert(cross_check->meet(tp_notnull) == cross_check,
617 617 "real address must not escape from expected memory type");
618 618 }
619 619 #endif
620 620 return tp;
621 621 }
622 622 }
623 623
624 624 //------------------------adr_phi_is_loop_invariant----------------------------
625 625 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted
626 626 // loop is loop invariant. Make a quick traversal of Phi and associated
627 627 // CastPP nodes, looking to see if they are a closed group within the loop.
628 628 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) {
629 629 // The idea is that the phi-nest must boil down to only CastPP nodes
630 630 // with the same data. This implies that any path into the loop already
631 631 // includes such a CastPP, and so the original cast, whatever its input,
632 632 // must be covered by an equivalent cast, with an earlier control input.
633 633 ResourceMark rm;
634 634
635 635 // The loop entry input of the phi should be the unique dominating
636 636 // node for every Phi/CastPP in the loop.
637 637 Unique_Node_List closure;
638 638 closure.push(adr_phi->in(LoopNode::EntryControl));
639 639
640 640 // Add the phi node and the cast to the worklist.
641 641 Unique_Node_List worklist;
642 642 worklist.push(adr_phi);
643 643 if( cast != NULL ){
644 644 if( !cast->is_ConstraintCast() ) return false;
645 645 worklist.push(cast);
646 646 }
647 647
648 648 // Begin recursive walk of phi nodes.
649 649 while( worklist.size() ){
650 650 // Take a node off the worklist
651 651 Node *n = worklist.pop();
652 652 if( !closure.member(n) ){
653 653 // Add it to the closure.
654 654 closure.push(n);
655 655 // Make a sanity check to ensure we don't waste too much time here.
656 656 if( closure.size() > 20) return false;
657 657 // This node is OK if:
658 658 // - it is a cast of an identical value
659 659 // - or it is a phi node (then we add its inputs to the worklist)
660 660 // Otherwise, the node is not OK, and we presume the cast is not invariant
661 661 if( n->is_ConstraintCast() ){
662 662 worklist.push(n->in(1));
663 663 } else if( n->is_Phi() ) {
664 664 for( uint i = 1; i < n->req(); i++ ) {
665 665 worklist.push(n->in(i));
666 666 }
667 667 } else {
668 668 return false;
669 669 }
670 670 }
671 671 }
672 672
673 673 // Quit when the worklist is empty, and we've found no offending nodes.
674 674 return true;
675 675 }
676 676
677 677 //------------------------------Ideal_DU_postCCP-------------------------------
678 678 // Find any cast-away of null-ness and keep its control. Null cast-aways are
679 679 // going away in this pass and we need to make this memory op depend on the
680 680 // gating null check.
681 681 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) {
682 682 return Ideal_common_DU_postCCP(ccp, this, in(MemNode::Address));
683 683 }
684 684
685 685 // I tried to leave the CastPP's in. This makes the graph more accurate in
686 686 // some sense; we get to keep around the knowledge that an oop is not-null
687 687 // after some test. Alas, the CastPP's interfere with GVN (some values are
688 688 // the regular oop, some are the CastPP of the oop, all merge at Phi's which
689 689 // cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed
690 690 // some of the more trivial cases in the optimizer. Removing more useless
691 691 // Phi's started allowing Loads to illegally float above null checks. I gave
692 692 // up on this approach. CNC 10/20/2000
693 693 // This static method may be called not from MemNode (EncodePNode calls it).
694 694 // Only the control edge of the node 'n' might be updated.
695 695 Node *MemNode::Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr ) {
696 696 Node *skipped_cast = NULL;
697 697 // Need a null check? Regular static accesses do not because they are
698 698 // from constant addresses. Array ops are gated by the range check (which
699 699 // always includes a NULL check). Just check field ops.
700 700 if( n->in(MemNode::Control) == NULL ) {
701 701 // Scan upwards for the highest location we can place this memory op.
702 702 while( true ) {
703 703 switch( adr->Opcode() ) {
704 704
705 705 case Op_AddP: // No change to NULL-ness, so peek thru AddP's
706 706 adr = adr->in(AddPNode::Base);
707 707 continue;
708 708
709 709 case Op_DecodeN: // No change to NULL-ness, so peek thru
710 710 adr = adr->in(1);
711 711 continue;
712 712
713 713 case Op_CastPP:
714 714 // If the CastPP is useless, just peek on through it.
715 715 if( ccp->type(adr) == ccp->type(adr->in(1)) ) {
716 716 // Remember the cast that we've peeked though. If we peek
717 717 // through more than one, then we end up remembering the highest
718 718 // one, that is, if in a loop, the one closest to the top.
719 719 skipped_cast = adr;
720 720 adr = adr->in(1);
721 721 continue;
722 722 }
723 723 // CastPP is going away in this pass! We need this memory op to be
724 724 // control-dependent on the test that is guarding the CastPP.
725 725 ccp->hash_delete(n);
726 726 n->set_req(MemNode::Control, adr->in(0));
727 727 ccp->hash_insert(n);
728 728 return n;
729 729
730 730 case Op_Phi:
731 731 // Attempt to float above a Phi to some dominating point.
732 732 if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) {
733 733 // If we've already peeked through a Cast (which could have set the
734 734 // control), we can't float above a Phi, because the skipped Cast
735 735 // may not be loop invariant.
736 736 if (adr_phi_is_loop_invariant(adr, skipped_cast)) {
737 737 adr = adr->in(1);
738 738 continue;
739 739 }
740 740 }
741 741
742 742 // Intentional fallthrough!
743 743
744 744 // No obvious dominating point. The mem op is pinned below the Phi
745 745 // by the Phi itself. If the Phi goes away (no true value is merged)
746 746 // then the mem op can float, but not indefinitely. It must be pinned
747 747 // behind the controls leading to the Phi.
748 748 case Op_CheckCastPP:
749 749 // These usually stick around to change address type, however a
750 750 // useless one can be elided and we still need to pick up a control edge
751 751 if (adr->in(0) == NULL) {
752 752 // This CheckCastPP node has NO control and is likely useless. But we
753 753 // need check further up the ancestor chain for a control input to keep
754 754 // the node in place. 4959717.
755 755 skipped_cast = adr;
756 756 adr = adr->in(1);
757 757 continue;
758 758 }
759 759 ccp->hash_delete(n);
760 760 n->set_req(MemNode::Control, adr->in(0));
761 761 ccp->hash_insert(n);
762 762 return n;
763 763
764 764 // List of "safe" opcodes; those that implicitly block the memory
765 765 // op below any null check.
766 766 case Op_CastX2P: // no null checks on native pointers
767 767 case Op_Parm: // 'this' pointer is not null
768 768 case Op_LoadP: // Loading from within a klass
769 769 case Op_LoadN: // Loading from within a klass
770 770 case Op_LoadKlass: // Loading from within a klass
771 771 case Op_LoadNKlass: // Loading from within a klass
772 772 case Op_ConP: // Loading from a klass
773 773 case Op_ConN: // Loading from a klass
774 774 case Op_CreateEx: // Sucking up the guts of an exception oop
775 775 case Op_Con: // Reading from TLS
776 776 case Op_CMoveP: // CMoveP is pinned
777 777 case Op_CMoveN: // CMoveN is pinned
778 778 break; // No progress
779 779
780 780 case Op_Proj: // Direct call to an allocation routine
781 781 case Op_SCMemProj: // Memory state from store conditional ops
782 782 #ifdef ASSERT
783 783 {
784 784 assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value");
785 785 const Node* call = adr->in(0);
786 786 if (call->is_CallJava()) {
787 787 const CallJavaNode* call_java = call->as_CallJava();
788 788 const TypeTuple *r = call_java->tf()->range();
789 789 assert(r->cnt() > TypeFunc::Parms, "must return value");
790 790 const Type* ret_type = r->field_at(TypeFunc::Parms);
791 791 assert(ret_type && ret_type->isa_ptr(), "must return pointer");
792 792 // We further presume that this is one of
793 793 // new_instance_Java, new_array_Java, or
794 794 // the like, but do not assert for this.
795 795 } else if (call->is_Allocate()) {
796 796 // similar case to new_instance_Java, etc.
797 797 } else if (!call->is_CallLeaf()) {
798 798 // Projections from fetch_oop (OSR) are allowed as well.
799 799 ShouldNotReachHere();
800 800 }
801 801 }
802 802 #endif
803 803 break;
804 804 default:
805 805 ShouldNotReachHere();
806 806 }
807 807 break;
808 808 }
809 809 }
810 810
811 811 return NULL; // No progress
812 812 }
813 813
814 814
815 815 //=============================================================================
816 816 uint LoadNode::size_of() const { return sizeof(*this); }
817 817 uint LoadNode::cmp( const Node &n ) const
818 818 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
819 819 const Type *LoadNode::bottom_type() const { return _type; }
820 820 uint LoadNode::ideal_reg() const {
821 821 return Matcher::base2reg[_type->base()];
822 822 }
823 823
824 824 #ifndef PRODUCT
825 825 void LoadNode::dump_spec(outputStream *st) const {
826 826 MemNode::dump_spec(st);
827 827 if( !Verbose && !WizardMode ) {
828 828 // standard dump does this in Verbose and WizardMode
829 829 st->print(" #"); _type->dump_on(st);
830 830 }
831 831 }
832 832 #endif
833 833
834 834 #ifdef ASSERT
835 835 //----------------------------is_immutable_value-------------------------------
836 836 // Helper function to allow a raw load without control edge for some cases
837 837 bool LoadNode::is_immutable_value(Node* adr) {
838 838 return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
839 839 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
840 840 (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) ==
841 841 in_bytes(JavaThread::osthread_offset())));
842 842 }
843 843 #endif
844 844
845 845 //----------------------------LoadNode::make-----------------------------------
846 846 // Polymorphic factory method:
847 847 Node *LoadNode::make( PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) {
848 848 Compile* C = gvn.C;
849 849
850 850 // sanity check the alias category against the created node type
851 851 assert(!(adr_type->isa_oopptr() &&
852 852 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
853 853 "use LoadKlassNode instead");
854 854 assert(!(adr_type->isa_aryptr() &&
855 855 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
856 856 "use LoadRangeNode instead");
857 857 // Check control edge of raw loads
858 858 assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
859 859 // oop will be recorded in oop map if load crosses safepoint
860 860 rt->isa_oopptr() || is_immutable_value(adr),
861 861 "raw memory operations should have control edge");
862 862 switch (bt) {
863 863 case T_BOOLEAN: return new (C, 3) LoadUBNode(ctl, mem, adr, adr_type, rt->is_int() );
864 864 case T_BYTE: return new (C, 3) LoadBNode (ctl, mem, adr, adr_type, rt->is_int() );
865 865 case T_INT: return new (C, 3) LoadINode (ctl, mem, adr, adr_type, rt->is_int() );
866 866 case T_CHAR: return new (C, 3) LoadUSNode(ctl, mem, adr, adr_type, rt->is_int() );
867 867 case T_SHORT: return new (C, 3) LoadSNode (ctl, mem, adr, adr_type, rt->is_int() );
868 868 case T_LONG: return new (C, 3) LoadLNode (ctl, mem, adr, adr_type, rt->is_long() );
869 869 case T_FLOAT: return new (C, 3) LoadFNode (ctl, mem, adr, adr_type, rt );
870 870 case T_DOUBLE: return new (C, 3) LoadDNode (ctl, mem, adr, adr_type, rt );
871 871 case T_ADDRESS: return new (C, 3) LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr() );
872 872 case T_OBJECT:
873 873 #ifdef _LP64
874 874 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
875 875 Node* load = gvn.transform(new (C, 3) LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop()));
876 876 return new (C, 2) DecodeNNode(load, load->bottom_type()->make_ptr());
877 877 } else
878 878 #endif
879 879 {
880 880 assert(!adr->bottom_type()->is_ptr_to_narrowoop(), "should have got back a narrow oop");
881 881 return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr());
882 882 }
883 883 }
884 884 ShouldNotReachHere();
885 885 return (LoadNode*)NULL;
886 886 }
887 887
888 888 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) {
889 889 bool require_atomic = true;
890 890 return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic);
891 891 }
892 892
893 893
894 894
895 895
896 896 //------------------------------hash-------------------------------------------
897 897 uint LoadNode::hash() const {
898 898 // unroll addition of interesting fields
899 899 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
900 900 }
901 901
902 902 //---------------------------can_see_stored_value------------------------------
903 903 // This routine exists to make sure this set of tests is done the same
904 904 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
905 905 // will change the graph shape in a way which makes memory alive twice at the
906 906 // same time (uses the Oracle model of aliasing), then some
907 907 // LoadXNode::Identity will fold things back to the equivalence-class model
908 908 // of aliasing.
909 909 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
910 910 Node* ld_adr = in(MemNode::Address);
911 911
912 912 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
913 913 Compile::AliasType* atp = tp != NULL ? phase->C->alias_type(tp) : NULL;
914 914 if (EliminateAutoBox && atp != NULL && atp->index() >= Compile::AliasIdxRaw &&
915 915 atp->field() != NULL && !atp->field()->is_volatile()) {
916 916 uint alias_idx = atp->index();
917 917 bool final = atp->field()->is_final();
918 918 Node* result = NULL;
919 919 Node* current = st;
920 920 // Skip through chains of MemBarNodes checking the MergeMems for
921 921 // new states for the slice of this load. Stop once any other
922 922 // kind of node is encountered. Loads from final memory can skip
923 923 // through any kind of MemBar but normal loads shouldn't skip
924 924 // through MemBarAcquire since the could allow them to move out of
925 925 // a synchronized region.
926 926 while (current->is_Proj()) {
927 927 int opc = current->in(0)->Opcode();
928 928 if ((final && (opc == Op_MemBarAcquire || opc == Op_MemBarAcquireLock)) ||
929 929 opc == Op_MemBarRelease || opc == Op_MemBarCPUOrder ||
930 930 opc == Op_MemBarReleaseLock) {
931 931 Node* mem = current->in(0)->in(TypeFunc::Memory);
932 932 if (mem->is_MergeMem()) {
933 933 MergeMemNode* merge = mem->as_MergeMem();
934 934 Node* new_st = merge->memory_at(alias_idx);
935 935 if (new_st == merge->base_memory()) {
936 936 // Keep searching
937 937 current = merge->base_memory();
938 938 continue;
939 939 }
940 940 // Save the new memory state for the slice and fall through
941 941 // to exit.
942 942 result = new_st;
943 943 }
944 944 }
945 945 break;
946 946 }
947 947 if (result != NULL) {
948 948 st = result;
949 949 }
950 950 }
951 951
952 952
953 953 // Loop around twice in the case Load -> Initialize -> Store.
954 954 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
955 955 for (int trip = 0; trip <= 1; trip++) {
956 956
957 957 if (st->is_Store()) {
958 958 Node* st_adr = st->in(MemNode::Address);
959 959 if (!phase->eqv(st_adr, ld_adr)) {
960 960 // Try harder before giving up... Match raw and non-raw pointers.
961 961 intptr_t st_off = 0;
962 962 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off);
963 963 if (alloc == NULL) return NULL;
964 964 intptr_t ld_off = 0;
965 965 AllocateNode* allo2 = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
966 966 if (alloc != allo2) return NULL;
967 967 if (ld_off != st_off) return NULL;
968 968 // At this point we have proven something like this setup:
969 969 // A = Allocate(...)
970 970 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off))
971 971 // S = StoreQ(, AddP(, A.Parm , #Off), V)
972 972 // (Actually, we haven't yet proven the Q's are the same.)
973 973 // In other words, we are loading from a casted version of
974 974 // the same pointer-and-offset that we stored to.
975 975 // Thus, we are able to replace L by V.
976 976 }
977 977 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
978 978 if (store_Opcode() != st->Opcode())
979 979 return NULL;
980 980 return st->in(MemNode::ValueIn);
981 981 }
982 982
983 983 intptr_t offset = 0; // scratch
984 984
985 985 // A load from a freshly-created object always returns zero.
986 986 // (This can happen after LoadNode::Ideal resets the load's memory input
987 987 // to find_captured_store, which returned InitializeNode::zero_memory.)
988 988 if (st->is_Proj() && st->in(0)->is_Allocate() &&
989 989 st->in(0) == AllocateNode::Ideal_allocation(ld_adr, phase, offset) &&
990 990 offset >= st->in(0)->as_Allocate()->minimum_header_size()) {
991 991 // return a zero value for the load's basic type
992 992 // (This is one of the few places where a generic PhaseTransform
993 993 // can create new nodes. Think of it as lazily manifesting
994 994 // virtually pre-existing constants.)
995 995 return phase->zerocon(memory_type());
996 996 }
997 997
998 998 // A load from an initialization barrier can match a captured store.
999 999 if (st->is_Proj() && st->in(0)->is_Initialize()) {
1000 1000 InitializeNode* init = st->in(0)->as_Initialize();
1001 1001 AllocateNode* alloc = init->allocation();
1002 1002 if (alloc != NULL &&
1003 1003 alloc == AllocateNode::Ideal_allocation(ld_adr, phase, offset)) {
1004 1004 // examine a captured store value
1005 1005 st = init->find_captured_store(offset, memory_size(), phase);
1006 1006 if (st != NULL)
1007 1007 continue; // take one more trip around
1008 1008 }
1009 1009 }
1010 1010
1011 1011 break;
1012 1012 }
1013 1013
1014 1014 return NULL;
1015 1015 }
1016 1016
1017 1017 //----------------------is_instance_field_load_with_local_phi------------------
1018 1018 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1019 1019 if( in(MemNode::Memory)->is_Phi() && in(MemNode::Memory)->in(0) == ctrl &&
1020 1020 in(MemNode::Address)->is_AddP() ) {
1021 1021 const TypeOopPtr* t_oop = in(MemNode::Address)->bottom_type()->isa_oopptr();
1022 1022 // Only instances.
1023 1023 if( t_oop != NULL && t_oop->is_known_instance_field() &&
1024 1024 t_oop->offset() != Type::OffsetBot &&
1025 1025 t_oop->offset() != Type::OffsetTop) {
1026 1026 return true;
1027 1027 }
1028 1028 }
1029 1029 return false;
1030 1030 }
1031 1031
1032 1032 //------------------------------Identity---------------------------------------
1033 1033 // Loads are identity if previous store is to same address
1034 1034 Node *LoadNode::Identity( PhaseTransform *phase ) {
1035 1035 // If the previous store-maker is the right kind of Store, and the store is
1036 1036 // to the same address, then we are equal to the value stored.
1037 1037 Node* mem = in(MemNode::Memory);
1038 1038 Node* value = can_see_stored_value(mem, phase);
1039 1039 if( value ) {
1040 1040 // byte, short & char stores truncate naturally.
1041 1041 // A load has to load the truncated value which requires
1042 1042 // some sort of masking operation and that requires an
1043 1043 // Ideal call instead of an Identity call.
1044 1044 if (memory_size() < BytesPerInt) {
1045 1045 // If the input to the store does not fit with the load's result type,
1046 1046 // it must be truncated via an Ideal call.
1047 1047 if (!phase->type(value)->higher_equal(phase->type(this)))
1048 1048 return this;
1049 1049 }
1050 1050 // (This works even when value is a Con, but LoadNode::Value
1051 1051 // usually runs first, producing the singleton type of the Con.)
1052 1052 return value;
1053 1053 }
1054 1054
1055 1055 // Search for an existing data phi which was generated before for the same
1056 1056 // instance's field to avoid infinite generation of phis in a loop.
1057 1057 Node *region = mem->in(0);
1058 1058 if (is_instance_field_load_with_local_phi(region)) {
1059 1059 const TypePtr *addr_t = in(MemNode::Address)->bottom_type()->isa_ptr();
1060 1060 int this_index = phase->C->get_alias_index(addr_t);
1061 1061 int this_offset = addr_t->offset();
1062 1062 int this_id = addr_t->is_oopptr()->instance_id();
1063 1063 const Type* this_type = bottom_type();
1064 1064 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1065 1065 Node* phi = region->fast_out(i);
1066 1066 if (phi->is_Phi() && phi != mem &&
1067 1067 phi->as_Phi()->is_same_inst_field(this_type, this_id, this_index, this_offset)) {
1068 1068 return phi;
1069 1069 }
1070 1070 }
1071 1071 }
1072 1072
1073 1073 return this;
1074 1074 }
1075 1075
1076 1076
1077 1077 // Returns true if the AliasType refers to the field that holds the
1078 1078 // cached box array. Currently only handles the IntegerCache case.
1079 1079 static bool is_autobox_cache(Compile::AliasType* atp) {
1080 1080 if (atp != NULL && atp->field() != NULL) {
1081 1081 ciField* field = atp->field();
1082 1082 ciSymbol* klass = field->holder()->name();
1083 1083 if (field->name() == ciSymbol::cache_field_name() &&
1084 1084 field->holder()->uses_default_loader() &&
1085 1085 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
1086 1086 return true;
1087 1087 }
1088 1088 }
1089 1089 return false;
1090 1090 }
1091 1091
1092 1092 // Fetch the base value in the autobox array
1093 1093 static bool fetch_autobox_base(Compile::AliasType* atp, int& cache_offset) {
1094 1094 if (atp != NULL && atp->field() != NULL) {
1095 1095 ciField* field = atp->field();
1096 1096 ciSymbol* klass = field->holder()->name();
1097 1097 if (field->name() == ciSymbol::cache_field_name() &&
1098 1098 field->holder()->uses_default_loader() &&
1099 1099 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
1100 1100 assert(field->is_constant(), "what?");
1101 1101 ciObjArray* array = field->constant_value().as_object()->as_obj_array();
1102 1102 // Fetch the box object at the base of the array and get its value
1103 1103 ciInstance* box = array->obj_at(0)->as_instance();
1104 1104 ciInstanceKlass* ik = box->klass()->as_instance_klass();
1105 1105 if (ik->nof_nonstatic_fields() == 1) {
1106 1106 // This should be true nonstatic_field_at requires calling
1107 1107 // nof_nonstatic_fields so check it anyway
1108 1108 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1109 1109 cache_offset = c.as_int();
1110 1110 }
1111 1111 return true;
1112 1112 }
1113 1113 }
1114 1114 return false;
1115 1115 }
1116 1116
1117 1117 // Returns true if the AliasType refers to the value field of an
1118 1118 // autobox object. Currently only handles Integer.
1119 1119 static bool is_autobox_object(Compile::AliasType* atp) {
1120 1120 if (atp != NULL && atp->field() != NULL) {
1121 1121 ciField* field = atp->field();
1122 1122 ciSymbol* klass = field->holder()->name();
1123 1123 if (field->name() == ciSymbol::value_name() &&
1124 1124 field->holder()->uses_default_loader() &&
1125 1125 klass == ciSymbol::java_lang_Integer()) {
1126 1126 return true;
1127 1127 }
1128 1128 }
1129 1129 return false;
1130 1130 }
1131 1131
1132 1132
1133 1133 // We're loading from an object which has autobox behaviour.
1134 1134 // If this object is result of a valueOf call we'll have a phi
1135 1135 // merging a newly allocated object and a load from the cache.
1136 1136 // We want to replace this load with the original incoming
1137 1137 // argument to the valueOf call.
1138 1138 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
1139 1139 Node* base = in(Address)->in(AddPNode::Base);
1140 1140 if (base->is_Phi() && base->req() == 3) {
1141 1141 AllocateNode* allocation = NULL;
1142 1142 int allocation_index = -1;
1143 1143 int load_index = -1;
1144 1144 for (uint i = 1; i < base->req(); i++) {
1145 1145 allocation = AllocateNode::Ideal_allocation(base->in(i), phase);
1146 1146 if (allocation != NULL) {
1147 1147 allocation_index = i;
1148 1148 load_index = 3 - allocation_index;
1149 1149 break;
1150 1150 }
1151 1151 }
1152 1152 bool has_load = ( allocation != NULL &&
1153 1153 (base->in(load_index)->is_Load() ||
1154 1154 base->in(load_index)->is_DecodeN() &&
1155 1155 base->in(load_index)->in(1)->is_Load()) );
1156 1156 if (has_load && in(Memory)->is_Phi() && in(Memory)->in(0) == base->in(0)) {
1157 1157 // Push the loads from the phi that comes from valueOf up
1158 1158 // through it to allow elimination of the loads and the recovery
1159 1159 // of the original value.
1160 1160 Node* mem_phi = in(Memory);
1161 1161 Node* offset = in(Address)->in(AddPNode::Offset);
1162 1162 Node* region = base->in(0);
1163 1163
1164 1164 Node* in1 = clone();
1165 1165 Node* in1_addr = in1->in(Address)->clone();
1166 1166 in1_addr->set_req(AddPNode::Base, base->in(allocation_index));
1167 1167 in1_addr->set_req(AddPNode::Address, base->in(allocation_index));
1168 1168 in1_addr->set_req(AddPNode::Offset, offset);
1169 1169 in1->set_req(0, region->in(allocation_index));
1170 1170 in1->set_req(Address, in1_addr);
1171 1171 in1->set_req(Memory, mem_phi->in(allocation_index));
1172 1172
1173 1173 Node* in2 = clone();
1174 1174 Node* in2_addr = in2->in(Address)->clone();
1175 1175 in2_addr->set_req(AddPNode::Base, base->in(load_index));
1176 1176 in2_addr->set_req(AddPNode::Address, base->in(load_index));
1177 1177 in2_addr->set_req(AddPNode::Offset, offset);
1178 1178 in2->set_req(0, region->in(load_index));
1179 1179 in2->set_req(Address, in2_addr);
1180 1180 in2->set_req(Memory, mem_phi->in(load_index));
1181 1181
1182 1182 in1_addr = phase->transform(in1_addr);
1183 1183 in1 = phase->transform(in1);
1184 1184 in2_addr = phase->transform(in2_addr);
1185 1185 in2 = phase->transform(in2);
1186 1186
1187 1187 PhiNode* result = PhiNode::make_blank(region, this);
1188 1188 result->set_req(allocation_index, in1);
1189 1189 result->set_req(load_index, in2);
1190 1190 return result;
1191 1191 }
1192 1192 } else if (base->is_Load() ||
1193 1193 base->is_DecodeN() && base->in(1)->is_Load()) {
1194 1194 if (base->is_DecodeN()) {
1195 1195 // Get LoadN node which loads cached Integer object
1196 1196 base = base->in(1);
1197 1197 }
1198 1198 // Eliminate the load of Integer.value for integers from the cache
1199 1199 // array by deriving the value from the index into the array.
1200 1200 // Capture the offset of the load and then reverse the computation.
1201 1201 Node* load_base = base->in(Address)->in(AddPNode::Base);
1202 1202 if (load_base->is_DecodeN()) {
1203 1203 // Get LoadN node which loads IntegerCache.cache field
1204 1204 load_base = load_base->in(1);
1205 1205 }
1206 1206 if (load_base != NULL) {
1207 1207 Compile::AliasType* atp = phase->C->alias_type(load_base->adr_type());
1208 1208 intptr_t cache_offset;
1209 1209 int shift = -1;
1210 1210 Node* cache = NULL;
1211 1211 if (is_autobox_cache(atp)) {
1212 1212 shift = exact_log2(type2aelembytes(T_OBJECT));
1213 1213 cache = AddPNode::Ideal_base_and_offset(load_base->in(Address), phase, cache_offset);
1214 1214 }
1215 1215 if (cache != NULL && base->in(Address)->is_AddP()) {
1216 1216 Node* elements[4];
1217 1217 int count = base->in(Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements));
1218 1218 int cache_low;
1219 1219 if (count > 0 && fetch_autobox_base(atp, cache_low)) {
1220 1220 int offset = arrayOopDesc::base_offset_in_bytes(memory_type()) - (cache_low << shift);
1221 1221 // Add up all the offsets making of the address of the load
1222 1222 Node* result = elements[0];
1223 1223 for (int i = 1; i < count; i++) {
1224 1224 result = phase->transform(new (phase->C, 3) AddXNode(result, elements[i]));
1225 1225 }
1226 1226 // Remove the constant offset from the address and then
1227 1227 // remove the scaling of the offset to recover the original index.
1228 1228 result = phase->transform(new (phase->C, 3) AddXNode(result, phase->MakeConX(-offset)));
1229 1229 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
1230 1230 // Peel the shift off directly but wrap it in a dummy node
1231 1231 // since Ideal can't return existing nodes
1232 1232 result = new (phase->C, 3) RShiftXNode(result->in(1), phase->intcon(0));
1233 1233 } else {
1234 1234 result = new (phase->C, 3) RShiftXNode(result, phase->intcon(shift));
1235 1235 }
1236 1236 #ifdef _LP64
1237 1237 result = new (phase->C, 2) ConvL2INode(phase->transform(result));
1238 1238 #endif
1239 1239 return result;
1240 1240 }
1241 1241 }
1242 1242 }
1243 1243 }
1244 1244 return NULL;
1245 1245 }
1246 1246
1247 1247 //------------------------------split_through_phi------------------------------
1248 1248 // Split instance field load through Phi.
1249 1249 Node *LoadNode::split_through_phi(PhaseGVN *phase) {
1250 1250 Node* mem = in(MemNode::Memory);
1251 1251 Node* address = in(MemNode::Address);
1252 1252 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1253 1253 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1254 1254
1255 1255 assert(mem->is_Phi() && (t_oop != NULL) &&
1256 1256 t_oop->is_known_instance_field(), "invalide conditions");
1257 1257
1258 1258 Node *region = mem->in(0);
1259 1259 if (region == NULL) {
1260 1260 return NULL; // Wait stable graph
1261 1261 }
1262 1262 uint cnt = mem->req();
1263 1263 for (uint i = 1; i < cnt; i++) {
1264 1264 Node* rc = region->in(i);
1265 1265 if (rc == NULL || phase->type(rc) == Type::TOP)
1266 1266 return NULL; // Wait stable graph
1267 1267 Node *in = mem->in(i);
1268 1268 if (in == NULL) {
1269 1269 return NULL; // Wait stable graph
1270 1270 }
1271 1271 }
1272 1272 // Check for loop invariant.
1273 1273 if (cnt == 3) {
1274 1274 for (uint i = 1; i < cnt; i++) {
1275 1275 Node *in = mem->in(i);
1276 1276 Node* m = MemNode::optimize_memory_chain(in, addr_t, phase);
1277 1277 if (m == mem) {
1278 1278 set_req(MemNode::Memory, mem->in(cnt - i)); // Skip this phi.
1279 1279 return this;
1280 1280 }
1281 1281 }
1282 1282 }
1283 1283 // Split through Phi (see original code in loopopts.cpp).
1284 1284 assert(phase->C->have_alias_type(addr_t), "instance should have alias type");
1285 1285
1286 1286 // Do nothing here if Identity will find a value
1287 1287 // (to avoid infinite chain of value phis generation).
1288 1288 if (!phase->eqv(this, this->Identity(phase)))
1289 1289 return NULL;
1290 1290
1291 1291 // Skip the split if the region dominates some control edge of the address.
1292 1292 if (!MemNode::all_controls_dominate(address, region))
1293 1293 return NULL;
1294 1294
1295 1295 const Type* this_type = this->bottom_type();
1296 1296 int this_index = phase->C->get_alias_index(addr_t);
1297 1297 int this_offset = addr_t->offset();
1298 1298 int this_iid = addr_t->is_oopptr()->instance_id();
1299 1299 PhaseIterGVN *igvn = phase->is_IterGVN();
1300 1300 Node *phi = new (igvn->C, region->req()) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset);
1301 1301 for (uint i = 1; i < region->req(); i++) {
1302 1302 Node *x;
1303 1303 Node* the_clone = NULL;
1304 1304 if (region->in(i) == phase->C->top()) {
1305 1305 x = phase->C->top(); // Dead path? Use a dead data op
1306 1306 } else {
1307 1307 x = this->clone(); // Else clone up the data op
1308 1308 the_clone = x; // Remember for possible deletion.
1309 1309 // Alter data node to use pre-phi inputs
1310 1310 if (this->in(0) == region) {
1311 1311 x->set_req(0, region->in(i));
1312 1312 } else {
1313 1313 x->set_req(0, NULL);
1314 1314 }
1315 1315 for (uint j = 1; j < this->req(); j++) {
1316 1316 Node *in = this->in(j);
1317 1317 if (in->is_Phi() && in->in(0) == region)
1318 1318 x->set_req(j, in->in(i)); // Use pre-Phi input for the clone
1319 1319 }
1320 1320 }
1321 1321 // Check for a 'win' on some paths
1322 1322 const Type *t = x->Value(igvn);
1323 1323
1324 1324 bool singleton = t->singleton();
1325 1325
1326 1326 // See comments in PhaseIdealLoop::split_thru_phi().
1327 1327 if (singleton && t == Type::TOP) {
1328 1328 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1329 1329 }
1330 1330
1331 1331 if (singleton) {
1332 1332 x = igvn->makecon(t);
1333 1333 } else {
1334 1334 // We now call Identity to try to simplify the cloned node.
1335 1335 // Note that some Identity methods call phase->type(this).
1336 1336 // Make sure that the type array is big enough for
1337 1337 // our new node, even though we may throw the node away.
1338 1338 // (This tweaking with igvn only works because x is a new node.)
1339 1339 igvn->set_type(x, t);
1340 1340 // If x is a TypeNode, capture any more-precise type permanently into Node
1341 1341 // otherwise it will be not updated during igvn->transform since
1342 1342 // igvn->type(x) is set to x->Value() already.
1343 1343 x->raise_bottom_type(t);
1344 1344 Node *y = x->Identity(igvn);
1345 1345 if (y != x) {
1346 1346 x = y;
1347 1347 } else {
1348 1348 y = igvn->hash_find(x);
1349 1349 if (y) {
1350 1350 x = y;
1351 1351 } else {
1352 1352 // Else x is a new node we are keeping
1353 1353 // We do not need register_new_node_with_optimizer
1354 1354 // because set_type has already been called.
1355 1355 igvn->_worklist.push(x);
1356 1356 }
1357 1357 }
1358 1358 }
1359 1359 if (x != the_clone && the_clone != NULL)
1360 1360 igvn->remove_dead_node(the_clone);
1361 1361 phi->set_req(i, x);
1362 1362 }
1363 1363 // Record Phi
1364 1364 igvn->register_new_node_with_optimizer(phi);
1365 1365 return phi;
1366 1366 }
1367 1367
1368 1368 //------------------------------Ideal------------------------------------------
1369 1369 // If the load is from Field memory and the pointer is non-null, we can
1370 1370 // zero out the control input.
1371 1371 // If the offset is constant and the base is an object allocation,
1372 1372 // try to hook me up to the exact initializing store.
1373 1373 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1374 1374 Node* p = MemNode::Ideal_common(phase, can_reshape);
1375 1375 if (p) return (p == NodeSentinel) ? NULL : p;
1376 1376
1377 1377 Node* ctrl = in(MemNode::Control);
1378 1378 Node* address = in(MemNode::Address);
1379 1379
1380 1380 // Skip up past a SafePoint control. Cannot do this for Stores because
1381 1381 // pointer stores & cardmarks must stay on the same side of a SafePoint.
1382 1382 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
1383 1383 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) {
1384 1384 ctrl = ctrl->in(0);
1385 1385 set_req(MemNode::Control,ctrl);
1386 1386 }
1387 1387
1388 1388 intptr_t ignore = 0;
1389 1389 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1390 1390 if (base != NULL
1391 1391 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
1392 1392 // Check for useless control edge in some common special cases
1393 1393 if (in(MemNode::Control) != NULL
1394 1394 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1395 1395 && all_controls_dominate(base, phase->C->start())) {
1396 1396 // A method-invariant, non-null address (constant or 'this' argument).
1397 1397 set_req(MemNode::Control, NULL);
1398 1398 }
1399 1399
1400 1400 if (EliminateAutoBox && can_reshape) {
1401 1401 assert(!phase->type(base)->higher_equal(TypePtr::NULL_PTR), "the autobox pointer should be non-null");
1402 1402 Compile::AliasType* atp = phase->C->alias_type(adr_type());
1403 1403 if (is_autobox_object(atp)) {
1404 1404 Node* result = eliminate_autobox(phase);
1405 1405 if (result != NULL) return result;
1406 1406 }
1407 1407 }
1408 1408 }
1409 1409
1410 1410 Node* mem = in(MemNode::Memory);
1411 1411 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1412 1412
1413 1413 if (addr_t != NULL) {
1414 1414 // try to optimize our memory input
1415 1415 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, phase);
1416 1416 if (opt_mem != mem) {
1417 1417 set_req(MemNode::Memory, opt_mem);
1418 1418 if (phase->type( opt_mem ) == Type::TOP) return NULL;
1419 1419 return this;
1420 1420 }
1421 1421 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1422 1422 if (can_reshape && opt_mem->is_Phi() &&
1423 1423 (t_oop != NULL) && t_oop->is_known_instance_field()) {
1424 1424 // Split instance field load through Phi.
1425 1425 Node* result = split_through_phi(phase);
1426 1426 if (result != NULL) return result;
1427 1427 }
1428 1428 }
1429 1429
1430 1430 // Check for prior store with a different base or offset; make Load
1431 1431 // independent. Skip through any number of them. Bail out if the stores
1432 1432 // are in an endless dead cycle and report no progress. This is a key
1433 1433 // transform for Reflection. However, if after skipping through the Stores
1434 1434 // we can't then fold up against a prior store do NOT do the transform as
1435 1435 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
1436 1436 // array memory alive twice: once for the hoisted Load and again after the
1437 1437 // bypassed Store. This situation only works if EVERYBODY who does
1438 1438 // anti-dependence work knows how to bypass. I.e. we need all
1439 1439 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
1440 1440 // the alias index stuff. So instead, peek through Stores and IFF we can
1441 1441 // fold up, do so.
1442 1442 Node* prev_mem = find_previous_store(phase);
1443 1443 // Steps (a), (b): Walk past independent stores to find an exact match.
1444 1444 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
1445 1445 // (c) See if we can fold up on the spot, but don't fold up here.
1446 1446 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
1447 1447 // just return a prior value, which is done by Identity calls.
1448 1448 if (can_see_stored_value(prev_mem, phase)) {
1449 1449 // Make ready for step (d):
1450 1450 set_req(MemNode::Memory, prev_mem);
1451 1451 return this;
1452 1452 }
1453 1453 }
1454 1454
1455 1455 return NULL; // No further progress
1456 1456 }
1457 1457
1458 1458 // Helper to recognize certain Klass fields which are invariant across
1459 1459 // some group of array types (e.g., int[] or all T[] where T < Object).
1460 1460 const Type*
1461 1461 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1462 1462 ciKlass* klass) const {
1463 1463 if (tkls->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
1464 1464 // The field is Klass::_modifier_flags. Return its (constant) value.
1465 1465 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1466 1466 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1467 1467 return TypeInt::make(klass->modifier_flags());
1468 1468 }
1469 1469 if (tkls->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
1470 1470 // The field is Klass::_access_flags. Return its (constant) value.
1471 1471 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1472 1472 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1473 1473 return TypeInt::make(klass->access_flags());
1474 1474 }
1475 1475 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)) {
1476 1476 // The field is Klass::_layout_helper. Return its constant value if known.
1477 1477 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1478 1478 return TypeInt::make(klass->layout_helper());
1479 1479 }
1480 1480
1481 1481 // No match.
1482 1482 return NULL;
1483 1483 }
1484 1484
1485 1485 //------------------------------Value-----------------------------------------
↓ open down ↓ |
1485 lines elided |
↑ open up ↑ |
1486 1486 const Type *LoadNode::Value( PhaseTransform *phase ) const {
1487 1487 // Either input is TOP ==> the result is TOP
1488 1488 Node* mem = in(MemNode::Memory);
1489 1489 const Type *t1 = phase->type(mem);
1490 1490 if (t1 == Type::TOP) return Type::TOP;
1491 1491 Node* adr = in(MemNode::Address);
1492 1492 const TypePtr* tp = phase->type(adr)->isa_ptr();
1493 1493 if (tp == NULL || tp->empty()) return Type::TOP;
1494 1494 int off = tp->offset();
1495 1495 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1496 + Compile* C = phase->C;
1496 1497
1497 1498 // Try to guess loaded type from pointer type
1498 1499 if (tp->base() == Type::AryPtr) {
1499 1500 const Type *t = tp->is_aryptr()->elem();
1500 1501 // Don't do this for integer types. There is only potential profit if
1501 1502 // the element type t is lower than _type; that is, for int types, if _type is
1502 1503 // more restrictive than t. This only happens here if one is short and the other
1503 1504 // char (both 16 bits), and in those cases we've made an intentional decision
1504 1505 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
1505 1506 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
1506 1507 //
1507 1508 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
1508 1509 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
1509 1510 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
1510 1511 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
1511 1512 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
1512 1513 // In fact, that could have been the original type of p1, and p1 could have
1513 1514 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
1514 1515 // expression (LShiftL quux 3) independently optimized to the constant 8.
1515 1516 if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
1516 1517 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
1517 1518 // t might actually be lower than _type, if _type is a unique
1518 1519 // concrete subclass of abstract class t.
1519 1520 // Make sure the reference is not into the header, by comparing
1520 1521 // the offset against the offset of the start of the array's data.
1521 1522 // Different array types begin at slightly different offsets (12 vs. 16).
1522 1523 // We choose T_BYTE as an example base type that is least restrictive
1523 1524 // as to alignment, which will therefore produce the smallest
1524 1525 // possible base offset.
1525 1526 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1526 1527 if ((uint)off >= (uint)min_base_off) { // is the offset beyond the header?
1527 1528 const Type* jt = t->join(_type);
1528 1529 // In any case, do not allow the join, per se, to empty out the type.
↓ open down ↓ |
23 lines elided |
↑ open up ↑ |
1529 1530 if (jt->empty() && !t->empty()) {
1530 1531 // This can happen if a interface-typed array narrows to a class type.
1531 1532 jt = _type;
1532 1533 }
1533 1534
1534 1535 if (EliminateAutoBox && adr->is_AddP()) {
1535 1536 // The pointers in the autobox arrays are always non-null
1536 1537 Node* base = adr->in(AddPNode::Base);
1537 1538 if (base != NULL &&
1538 1539 !phase->type(base)->higher_equal(TypePtr::NULL_PTR)) {
1539 - Compile::AliasType* atp = phase->C->alias_type(base->adr_type());
1540 + Compile::AliasType* atp = C->alias_type(base->adr_type());
1540 1541 if (is_autobox_cache(atp)) {
1541 1542 return jt->join(TypePtr::NOTNULL)->is_ptr();
1542 1543 }
1543 1544 }
1544 1545 }
1545 1546 return jt;
1546 1547 }
1547 1548 }
1548 1549 } else if (tp->base() == Type::InstPtr) {
1550 + ciEnv* env = C->env();
1549 1551 const TypeInstPtr* tinst = tp->is_instptr();
1550 1552 ciKlass* klass = tinst->klass();
1551 1553 assert( off != Type::OffsetBot ||
1552 1554 // arrays can be cast to Objects
1553 1555 tp->is_oopptr()->klass()->is_java_lang_Object() ||
1554 1556 // unsafe field access may not have a constant offset
1555 - phase->C->has_unsafe_access(),
1557 + C->has_unsafe_access(),
1556 1558 "Field accesses must be precise" );
1557 1559 // For oop loads, we expect the _type to be precise
1558 - if (klass == phase->C->env()->String_klass() &&
1560 + if (klass == env->String_klass() &&
1559 1561 adr->is_AddP() && off != Type::OffsetBot) {
1560 1562 // For constant Strings treat the final fields as compile time constants.
1561 1563 Node* base = adr->in(AddPNode::Base);
1562 1564 const TypeOopPtr* t = phase->type(base)->isa_oopptr();
1563 1565 if (t != NULL && t->singleton()) {
1564 - ciField* field = phase->C->env()->String_klass()->get_field_by_offset(off, false);
1566 + ciField* field = env->String_klass()->get_field_by_offset(off, false);
1565 1567 if (field != NULL && field->is_final()) {
1566 1568 ciObject* string = t->const_oop();
1567 1569 ciConstant constant = string->as_instance()->field_value(field);
1568 1570 if (constant.basic_type() == T_INT) {
1569 1571 return TypeInt::make(constant.as_int());
1570 1572 } else if (constant.basic_type() == T_ARRAY) {
1571 1573 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
1572 1574 return TypeNarrowOop::make_from_constant(constant.as_object(), true);
1573 1575 } else {
1574 1576 return TypeOopPtr::make_from_constant(constant.as_object(), true);
1575 1577 }
1576 1578 }
1577 1579 }
1578 1580 }
1579 1581 }
1582 + // Optimizations for constant objects
1583 + ciObject* const_oop = tinst->const_oop();
1584 + if (const_oop != NULL) {
1585 + // For constant CallSites treat the target field as a compile time constant.
1586 + if (const_oop->is_call_site()) {
1587 + ciCallSite* call_site = const_oop->as_call_site();
1588 + ciField* field = call_site->klass()->as_instance_klass()->get_field_by_offset(off, /*is_static=*/ false);
1589 + if (field != NULL && field->is_call_site_target()) {
1590 + ciMethodHandle* target = call_site->get_target();
1591 + if (target != NULL) { // just in case
1592 + ciConstant constant(T_OBJECT, target);
1593 + const Type* t;
1594 + if (adr->bottom_type()->is_ptr_to_narrowoop()) {
1595 + t = TypeNarrowOop::make_from_constant(constant.as_object(), true);
1596 + } else {
1597 + t = TypeOopPtr::make_from_constant(constant.as_object(), true);
1598 + }
1599 + // Add a dependence for invalidation of the optimization.
1600 + if (!call_site->is_constant_call_site()) {
1601 + C->dependencies()->assert_call_site_target_value(call_site, target);
1602 + }
1603 + return t;
1604 + }
1605 + }
1606 + }
1607 + }
1580 1608 } else if (tp->base() == Type::KlassPtr) {
1581 1609 assert( off != Type::OffsetBot ||
1582 1610 // arrays can be cast to Objects
1583 1611 tp->is_klassptr()->klass()->is_java_lang_Object() ||
1584 1612 // also allow array-loading from the primary supertype
1585 1613 // array during subtype checks
1586 1614 Opcode() == Op_LoadKlass,
1587 1615 "Field accesses must be precise" );
1588 1616 // For klass/static loads, we expect the _type to be precise
1589 1617 }
1590 1618
1591 1619 const TypeKlassPtr *tkls = tp->isa_klassptr();
1592 1620 if (tkls != NULL && !StressReflectiveCode) {
1593 1621 ciKlass* klass = tkls->klass();
1594 1622 if (klass->is_loaded() && tkls->klass_is_exact()) {
1595 1623 // We are loading a field from a Klass metaobject whose identity
1596 1624 // is known at compile time (the type is "exact" or "precise").
1597 1625 // Check for fields we know are maintained as constants by the VM.
1598 1626 if (tkls->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) {
1599 1627 // The field is Klass::_super_check_offset. Return its (constant) value.
1600 1628 // (Folds up type checking code.)
1601 1629 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
1602 1630 return TypeInt::make(klass->super_check_offset());
1603 1631 }
1604 1632 // Compute index into primary_supers array
1605 1633 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
1606 1634 // Check for overflowing; use unsigned compare to handle the negative case.
1607 1635 if( depth < ciKlass::primary_super_limit() ) {
1608 1636 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1609 1637 // (Folds up type checking code.)
1610 1638 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1611 1639 ciKlass *ss = klass->super_of_depth(depth);
1612 1640 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1613 1641 }
1614 1642 const Type* aift = load_array_final_field(tkls, klass);
1615 1643 if (aift != NULL) return aift;
1616 1644 if (tkls->offset() == in_bytes(arrayKlass::component_mirror_offset()) + (int)sizeof(oopDesc)
1617 1645 && klass->is_array_klass()) {
1618 1646 // The field is arrayKlass::_component_mirror. Return its (constant) value.
1619 1647 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.)
1620 1648 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror");
1621 1649 return TypeInstPtr::make(klass->as_array_klass()->component_mirror());
1622 1650 }
1623 1651 if (tkls->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) {
1624 1652 // The field is Klass::_java_mirror. Return its (constant) value.
1625 1653 // (Folds up the 2nd indirection in anObjConstant.getClass().)
1626 1654 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1627 1655 return TypeInstPtr::make(klass->java_mirror());
1628 1656 }
1629 1657 }
1630 1658
1631 1659 // We can still check if we are loading from the primary_supers array at a
1632 1660 // shallow enough depth. Even though the klass is not exact, entries less
1633 1661 // than or equal to its super depth are correct.
1634 1662 if (klass->is_loaded() ) {
1635 1663 ciType *inner = klass->klass();
1636 1664 while( inner->is_obj_array_klass() )
1637 1665 inner = inner->as_obj_array_klass()->base_element_type();
1638 1666 if( inner->is_instance_klass() &&
1639 1667 !inner->as_instance_klass()->flags().is_interface() ) {
1640 1668 // Compute index into primary_supers array
1641 1669 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
1642 1670 // Check for overflowing; use unsigned compare to handle the negative case.
1643 1671 if( depth < ciKlass::primary_super_limit() &&
1644 1672 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
1645 1673 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1646 1674 // (Folds up type checking code.)
1647 1675 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1648 1676 ciKlass *ss = klass->super_of_depth(depth);
1649 1677 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1650 1678 }
1651 1679 }
1652 1680 }
1653 1681
1654 1682 // If the type is enough to determine that the thing is not an array,
1655 1683 // we can give the layout_helper a positive interval type.
1656 1684 // This will help short-circuit some reflective code.
1657 1685 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)
1658 1686 && !klass->is_array_klass() // not directly typed as an array
1659 1687 && !klass->is_interface() // specifically not Serializable & Cloneable
1660 1688 && !klass->is_java_lang_Object() // not the supertype of all T[]
1661 1689 ) {
1662 1690 // Note: When interfaces are reliable, we can narrow the interface
1663 1691 // test to (klass != Serializable && klass != Cloneable).
1664 1692 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
1665 1693 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
1666 1694 // The key property of this type is that it folds up tests
1667 1695 // for array-ness, since it proves that the layout_helper is positive.
1668 1696 // Thus, a generic value like the basic object layout helper works fine.
1669 1697 return TypeInt::make(min_size, max_jint, Type::WidenMin);
1670 1698 }
1671 1699 }
1672 1700
1673 1701 // If we are loading from a freshly-allocated object, produce a zero,
1674 1702 // if the load is provably beyond the header of the object.
1675 1703 // (Also allow a variable load from a fresh array to produce zero.)
1676 1704 const TypeOopPtr *tinst = tp->isa_oopptr();
1677 1705 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field();
1678 1706 if (ReduceFieldZeroing || is_instance) {
1679 1707 Node* value = can_see_stored_value(mem,phase);
1680 1708 if (value != NULL && value->is_Con())
1681 1709 return value->bottom_type();
1682 1710 }
1683 1711
1684 1712 if (is_instance) {
1685 1713 // If we have an instance type and our memory input is the
1686 1714 // programs's initial memory state, there is no matching store,
1687 1715 // so just return a zero of the appropriate type
1688 1716 Node *mem = in(MemNode::Memory);
1689 1717 if (mem->is_Parm() && mem->in(0)->is_Start()) {
1690 1718 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
1691 1719 return Type::get_zero_type(_type->basic_type());
1692 1720 }
1693 1721 }
1694 1722 return _type;
1695 1723 }
1696 1724
1697 1725 //------------------------------match_edge-------------------------------------
1698 1726 // Do we Match on this edge index or not? Match only the address.
1699 1727 uint LoadNode::match_edge(uint idx) const {
1700 1728 return idx == MemNode::Address;
1701 1729 }
1702 1730
1703 1731 //--------------------------LoadBNode::Ideal--------------------------------------
1704 1732 //
1705 1733 // If the previous store is to the same address as this load,
1706 1734 // and the value stored was larger than a byte, replace this load
1707 1735 // with the value stored truncated to a byte. If no truncation is
1708 1736 // needed, the replacement is done in LoadNode::Identity().
1709 1737 //
1710 1738 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1711 1739 Node* mem = in(MemNode::Memory);
1712 1740 Node* value = can_see_stored_value(mem,phase);
1713 1741 if( value && !phase->type(value)->higher_equal( _type ) ) {
1714 1742 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(24)) );
1715 1743 return new (phase->C, 3) RShiftINode(result, phase->intcon(24));
1716 1744 }
1717 1745 // Identity call will handle the case where truncation is not needed.
1718 1746 return LoadNode::Ideal(phase, can_reshape);
1719 1747 }
1720 1748
1721 1749 //--------------------------LoadUBNode::Ideal-------------------------------------
1722 1750 //
1723 1751 // If the previous store is to the same address as this load,
1724 1752 // and the value stored was larger than a byte, replace this load
1725 1753 // with the value stored truncated to a byte. If no truncation is
1726 1754 // needed, the replacement is done in LoadNode::Identity().
1727 1755 //
1728 1756 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1729 1757 Node* mem = in(MemNode::Memory);
1730 1758 Node* value = can_see_stored_value(mem, phase);
1731 1759 if (value && !phase->type(value)->higher_equal(_type))
1732 1760 return new (phase->C, 3) AndINode(value, phase->intcon(0xFF));
1733 1761 // Identity call will handle the case where truncation is not needed.
1734 1762 return LoadNode::Ideal(phase, can_reshape);
1735 1763 }
1736 1764
1737 1765 //--------------------------LoadUSNode::Ideal-------------------------------------
1738 1766 //
1739 1767 // If the previous store is to the same address as this load,
1740 1768 // and the value stored was larger than a char, replace this load
1741 1769 // with the value stored truncated to a char. If no truncation is
1742 1770 // needed, the replacement is done in LoadNode::Identity().
1743 1771 //
1744 1772 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1745 1773 Node* mem = in(MemNode::Memory);
1746 1774 Node* value = can_see_stored_value(mem,phase);
1747 1775 if( value && !phase->type(value)->higher_equal( _type ) )
1748 1776 return new (phase->C, 3) AndINode(value,phase->intcon(0xFFFF));
1749 1777 // Identity call will handle the case where truncation is not needed.
1750 1778 return LoadNode::Ideal(phase, can_reshape);
1751 1779 }
1752 1780
1753 1781 //--------------------------LoadSNode::Ideal--------------------------------------
1754 1782 //
1755 1783 // If the previous store is to the same address as this load,
1756 1784 // and the value stored was larger than a short, replace this load
1757 1785 // with the value stored truncated to a short. If no truncation is
1758 1786 // needed, the replacement is done in LoadNode::Identity().
1759 1787 //
1760 1788 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1761 1789 Node* mem = in(MemNode::Memory);
1762 1790 Node* value = can_see_stored_value(mem,phase);
1763 1791 if( value && !phase->type(value)->higher_equal( _type ) ) {
1764 1792 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(16)) );
1765 1793 return new (phase->C, 3) RShiftINode(result, phase->intcon(16));
1766 1794 }
1767 1795 // Identity call will handle the case where truncation is not needed.
1768 1796 return LoadNode::Ideal(phase, can_reshape);
1769 1797 }
1770 1798
1771 1799 //=============================================================================
1772 1800 //----------------------------LoadKlassNode::make------------------------------
1773 1801 // Polymorphic factory method:
1774 1802 Node *LoadKlassNode::make( PhaseGVN& gvn, Node *mem, Node *adr, const TypePtr* at, const TypeKlassPtr *tk ) {
1775 1803 Compile* C = gvn.C;
1776 1804 Node *ctl = NULL;
1777 1805 // sanity check the alias category against the created node type
1778 1806 const TypeOopPtr *adr_type = adr->bottom_type()->isa_oopptr();
1779 1807 assert(adr_type != NULL, "expecting TypeOopPtr");
1780 1808 #ifdef _LP64
1781 1809 if (adr_type->is_ptr_to_narrowoop()) {
1782 1810 Node* load_klass = gvn.transform(new (C, 3) LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowoop()));
1783 1811 return new (C, 2) DecodeNNode(load_klass, load_klass->bottom_type()->make_ptr());
1784 1812 }
1785 1813 #endif
1786 1814 assert(!adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
1787 1815 return new (C, 3) LoadKlassNode(ctl, mem, adr, at, tk);
1788 1816 }
1789 1817
1790 1818 //------------------------------Value------------------------------------------
1791 1819 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const {
1792 1820 return klass_value_common(phase);
1793 1821 }
1794 1822
1795 1823 const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const {
1796 1824 // Either input is TOP ==> the result is TOP
1797 1825 const Type *t1 = phase->type( in(MemNode::Memory) );
1798 1826 if (t1 == Type::TOP) return Type::TOP;
1799 1827 Node *adr = in(MemNode::Address);
1800 1828 const Type *t2 = phase->type( adr );
1801 1829 if (t2 == Type::TOP) return Type::TOP;
1802 1830 const TypePtr *tp = t2->is_ptr();
1803 1831 if (TypePtr::above_centerline(tp->ptr()) ||
1804 1832 tp->ptr() == TypePtr::Null) return Type::TOP;
1805 1833
1806 1834 // Return a more precise klass, if possible
1807 1835 const TypeInstPtr *tinst = tp->isa_instptr();
1808 1836 if (tinst != NULL) {
1809 1837 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
1810 1838 int offset = tinst->offset();
1811 1839 if (ik == phase->C->env()->Class_klass()
1812 1840 && (offset == java_lang_Class::klass_offset_in_bytes() ||
1813 1841 offset == java_lang_Class::array_klass_offset_in_bytes())) {
1814 1842 // We are loading a special hidden field from a Class mirror object,
1815 1843 // the field which points to the VM's Klass metaobject.
1816 1844 ciType* t = tinst->java_mirror_type();
1817 1845 // java_mirror_type returns non-null for compile-time Class constants.
1818 1846 if (t != NULL) {
1819 1847 // constant oop => constant klass
1820 1848 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
1821 1849 return TypeKlassPtr::make(ciArrayKlass::make(t));
1822 1850 }
1823 1851 if (!t->is_klass()) {
1824 1852 // a primitive Class (e.g., int.class) has NULL for a klass field
1825 1853 return TypePtr::NULL_PTR;
1826 1854 }
1827 1855 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
1828 1856 return TypeKlassPtr::make(t->as_klass());
1829 1857 }
1830 1858 // non-constant mirror, so we can't tell what's going on
1831 1859 }
1832 1860 if( !ik->is_loaded() )
1833 1861 return _type; // Bail out if not loaded
1834 1862 if (offset == oopDesc::klass_offset_in_bytes()) {
1835 1863 if (tinst->klass_is_exact()) {
1836 1864 return TypeKlassPtr::make(ik);
1837 1865 }
1838 1866 // See if we can become precise: no subklasses and no interface
1839 1867 // (Note: We need to support verified interfaces.)
1840 1868 if (!ik->is_interface() && !ik->has_subklass()) {
1841 1869 //assert(!UseExactTypes, "this code should be useless with exact types");
1842 1870 // Add a dependence; if any subclass added we need to recompile
1843 1871 if (!ik->is_final()) {
1844 1872 // %%% should use stronger assert_unique_concrete_subtype instead
1845 1873 phase->C->dependencies()->assert_leaf_type(ik);
1846 1874 }
1847 1875 // Return precise klass
1848 1876 return TypeKlassPtr::make(ik);
1849 1877 }
1850 1878
1851 1879 // Return root of possible klass
1852 1880 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
1853 1881 }
1854 1882 }
1855 1883
1856 1884 // Check for loading klass from an array
1857 1885 const TypeAryPtr *tary = tp->isa_aryptr();
1858 1886 if( tary != NULL ) {
1859 1887 ciKlass *tary_klass = tary->klass();
1860 1888 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP
1861 1889 && tary->offset() == oopDesc::klass_offset_in_bytes()) {
1862 1890 if (tary->klass_is_exact()) {
1863 1891 return TypeKlassPtr::make(tary_klass);
1864 1892 }
1865 1893 ciArrayKlass *ak = tary->klass()->as_array_klass();
1866 1894 // If the klass is an object array, we defer the question to the
1867 1895 // array component klass.
1868 1896 if( ak->is_obj_array_klass() ) {
1869 1897 assert( ak->is_loaded(), "" );
1870 1898 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
1871 1899 if( base_k->is_loaded() && base_k->is_instance_klass() ) {
1872 1900 ciInstanceKlass* ik = base_k->as_instance_klass();
1873 1901 // See if we can become precise: no subklasses and no interface
1874 1902 if (!ik->is_interface() && !ik->has_subklass()) {
1875 1903 //assert(!UseExactTypes, "this code should be useless with exact types");
1876 1904 // Add a dependence; if any subclass added we need to recompile
1877 1905 if (!ik->is_final()) {
1878 1906 phase->C->dependencies()->assert_leaf_type(ik);
1879 1907 }
1880 1908 // Return precise array klass
1881 1909 return TypeKlassPtr::make(ak);
1882 1910 }
1883 1911 }
1884 1912 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
1885 1913 } else { // Found a type-array?
1886 1914 //assert(!UseExactTypes, "this code should be useless with exact types");
1887 1915 assert( ak->is_type_array_klass(), "" );
1888 1916 return TypeKlassPtr::make(ak); // These are always precise
1889 1917 }
1890 1918 }
1891 1919 }
1892 1920
1893 1921 // Check for loading klass from an array klass
1894 1922 const TypeKlassPtr *tkls = tp->isa_klassptr();
1895 1923 if (tkls != NULL && !StressReflectiveCode) {
1896 1924 ciKlass* klass = tkls->klass();
1897 1925 if( !klass->is_loaded() )
1898 1926 return _type; // Bail out if not loaded
1899 1927 if( klass->is_obj_array_klass() &&
1900 1928 (uint)tkls->offset() == objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc)) {
1901 1929 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
1902 1930 // // Always returning precise element type is incorrect,
1903 1931 // // e.g., element type could be object and array may contain strings
1904 1932 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
1905 1933
1906 1934 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
1907 1935 // according to the element type's subclassing.
1908 1936 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
1909 1937 }
1910 1938 if( klass->is_instance_klass() && tkls->klass_is_exact() &&
1911 1939 (uint)tkls->offset() == Klass::super_offset_in_bytes() + sizeof(oopDesc)) {
1912 1940 ciKlass* sup = klass->as_instance_klass()->super();
1913 1941 // The field is Klass::_super. Return its (constant) value.
1914 1942 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
1915 1943 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
1916 1944 }
1917 1945 }
1918 1946
1919 1947 // Bailout case
1920 1948 return LoadNode::Value(phase);
1921 1949 }
1922 1950
1923 1951 //------------------------------Identity---------------------------------------
1924 1952 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
1925 1953 // Also feed through the klass in Allocate(...klass...)._klass.
1926 1954 Node* LoadKlassNode::Identity( PhaseTransform *phase ) {
1927 1955 return klass_identity_common(phase);
1928 1956 }
1929 1957
1930 1958 Node* LoadNode::klass_identity_common(PhaseTransform *phase ) {
1931 1959 Node* x = LoadNode::Identity(phase);
1932 1960 if (x != this) return x;
1933 1961
1934 1962 // Take apart the address into an oop and and offset.
1935 1963 // Return 'this' if we cannot.
1936 1964 Node* adr = in(MemNode::Address);
1937 1965 intptr_t offset = 0;
1938 1966 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
1939 1967 if (base == NULL) return this;
1940 1968 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
1941 1969 if (toop == NULL) return this;
1942 1970
1943 1971 // We can fetch the klass directly through an AllocateNode.
1944 1972 // This works even if the klass is not constant (clone or newArray).
1945 1973 if (offset == oopDesc::klass_offset_in_bytes()) {
1946 1974 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
1947 1975 if (allocated_klass != NULL) {
1948 1976 return allocated_klass;
1949 1977 }
1950 1978 }
1951 1979
1952 1980 // Simplify k.java_mirror.as_klass to plain k, where k is a klassOop.
1953 1981 // Simplify ak.component_mirror.array_klass to plain ak, ak an arrayKlass.
1954 1982 // See inline_native_Class_query for occurrences of these patterns.
1955 1983 // Java Example: x.getClass().isAssignableFrom(y)
1956 1984 // Java Example: Array.newInstance(x.getClass().getComponentType(), n)
1957 1985 //
1958 1986 // This improves reflective code, often making the Class
1959 1987 // mirror go completely dead. (Current exception: Class
1960 1988 // mirrors may appear in debug info, but we could clean them out by
1961 1989 // introducing a new debug info operator for klassOop.java_mirror).
1962 1990 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
1963 1991 && (offset == java_lang_Class::klass_offset_in_bytes() ||
1964 1992 offset == java_lang_Class::array_klass_offset_in_bytes())) {
1965 1993 // We are loading a special hidden field from a Class mirror,
1966 1994 // the field which points to its Klass or arrayKlass metaobject.
1967 1995 if (base->is_Load()) {
1968 1996 Node* adr2 = base->in(MemNode::Address);
1969 1997 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
1970 1998 if (tkls != NULL && !tkls->empty()
1971 1999 && (tkls->klass()->is_instance_klass() ||
1972 2000 tkls->klass()->is_array_klass())
1973 2001 && adr2->is_AddP()
1974 2002 ) {
1975 2003 int mirror_field = Klass::java_mirror_offset_in_bytes();
1976 2004 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
1977 2005 mirror_field = in_bytes(arrayKlass::component_mirror_offset());
1978 2006 }
1979 2007 if (tkls->offset() == mirror_field + (int)sizeof(oopDesc)) {
1980 2008 return adr2->in(AddPNode::Base);
1981 2009 }
1982 2010 }
1983 2011 }
1984 2012 }
1985 2013
1986 2014 return this;
1987 2015 }
1988 2016
1989 2017
1990 2018 //------------------------------Value------------------------------------------
1991 2019 const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const {
1992 2020 const Type *t = klass_value_common(phase);
1993 2021 if (t == Type::TOP)
1994 2022 return t;
1995 2023
1996 2024 return t->make_narrowoop();
1997 2025 }
1998 2026
1999 2027 //------------------------------Identity---------------------------------------
2000 2028 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
2001 2029 // Also feed through the klass in Allocate(...klass...)._klass.
2002 2030 Node* LoadNKlassNode::Identity( PhaseTransform *phase ) {
2003 2031 Node *x = klass_identity_common(phase);
2004 2032
2005 2033 const Type *t = phase->type( x );
2006 2034 if( t == Type::TOP ) return x;
2007 2035 if( t->isa_narrowoop()) return x;
2008 2036
2009 2037 return phase->transform(new (phase->C, 2) EncodePNode(x, t->make_narrowoop()));
2010 2038 }
2011 2039
2012 2040 //------------------------------Value-----------------------------------------
2013 2041 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const {
2014 2042 // Either input is TOP ==> the result is TOP
2015 2043 const Type *t1 = phase->type( in(MemNode::Memory) );
2016 2044 if( t1 == Type::TOP ) return Type::TOP;
2017 2045 Node *adr = in(MemNode::Address);
2018 2046 const Type *t2 = phase->type( adr );
2019 2047 if( t2 == Type::TOP ) return Type::TOP;
2020 2048 const TypePtr *tp = t2->is_ptr();
2021 2049 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
2022 2050 const TypeAryPtr *tap = tp->isa_aryptr();
2023 2051 if( !tap ) return _type;
2024 2052 return tap->size();
2025 2053 }
2026 2054
2027 2055 //-------------------------------Ideal---------------------------------------
2028 2056 // Feed through the length in AllocateArray(...length...)._length.
2029 2057 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2030 2058 Node* p = MemNode::Ideal_common(phase, can_reshape);
2031 2059 if (p) return (p == NodeSentinel) ? NULL : p;
2032 2060
2033 2061 // Take apart the address into an oop and and offset.
2034 2062 // Return 'this' if we cannot.
2035 2063 Node* adr = in(MemNode::Address);
2036 2064 intptr_t offset = 0;
2037 2065 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2038 2066 if (base == NULL) return NULL;
2039 2067 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2040 2068 if (tary == NULL) return NULL;
2041 2069
2042 2070 // We can fetch the length directly through an AllocateArrayNode.
2043 2071 // This works even if the length is not constant (clone or newArray).
2044 2072 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2045 2073 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2046 2074 if (alloc != NULL) {
2047 2075 Node* allocated_length = alloc->Ideal_length();
2048 2076 Node* len = alloc->make_ideal_length(tary, phase);
2049 2077 if (allocated_length != len) {
2050 2078 // New CastII improves on this.
2051 2079 return len;
2052 2080 }
2053 2081 }
2054 2082 }
2055 2083
2056 2084 return NULL;
2057 2085 }
2058 2086
2059 2087 //------------------------------Identity---------------------------------------
2060 2088 // Feed through the length in AllocateArray(...length...)._length.
2061 2089 Node* LoadRangeNode::Identity( PhaseTransform *phase ) {
2062 2090 Node* x = LoadINode::Identity(phase);
2063 2091 if (x != this) return x;
2064 2092
2065 2093 // Take apart the address into an oop and and offset.
2066 2094 // Return 'this' if we cannot.
2067 2095 Node* adr = in(MemNode::Address);
2068 2096 intptr_t offset = 0;
2069 2097 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2070 2098 if (base == NULL) return this;
2071 2099 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2072 2100 if (tary == NULL) return this;
2073 2101
2074 2102 // We can fetch the length directly through an AllocateArrayNode.
2075 2103 // This works even if the length is not constant (clone or newArray).
2076 2104 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2077 2105 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2078 2106 if (alloc != NULL) {
2079 2107 Node* allocated_length = alloc->Ideal_length();
2080 2108 // Do not allow make_ideal_length to allocate a CastII node.
2081 2109 Node* len = alloc->make_ideal_length(tary, phase, false);
2082 2110 if (allocated_length == len) {
2083 2111 // Return allocated_length only if it would not be improved by a CastII.
2084 2112 return allocated_length;
2085 2113 }
2086 2114 }
2087 2115 }
2088 2116
2089 2117 return this;
2090 2118
2091 2119 }
2092 2120
2093 2121 //=============================================================================
2094 2122 //---------------------------StoreNode::make-----------------------------------
2095 2123 // Polymorphic factory method:
2096 2124 StoreNode* StoreNode::make( PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) {
2097 2125 Compile* C = gvn.C;
2098 2126 assert( C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
2099 2127 ctl != NULL, "raw memory operations should have control edge");
2100 2128
2101 2129 switch (bt) {
2102 2130 case T_BOOLEAN:
2103 2131 case T_BYTE: return new (C, 4) StoreBNode(ctl, mem, adr, adr_type, val);
2104 2132 case T_INT: return new (C, 4) StoreINode(ctl, mem, adr, adr_type, val);
2105 2133 case T_CHAR:
2106 2134 case T_SHORT: return new (C, 4) StoreCNode(ctl, mem, adr, adr_type, val);
2107 2135 case T_LONG: return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val);
2108 2136 case T_FLOAT: return new (C, 4) StoreFNode(ctl, mem, adr, adr_type, val);
2109 2137 case T_DOUBLE: return new (C, 4) StoreDNode(ctl, mem, adr, adr_type, val);
2110 2138 case T_ADDRESS:
2111 2139 case T_OBJECT:
2112 2140 #ifdef _LP64
2113 2141 if (adr->bottom_type()->is_ptr_to_narrowoop() ||
2114 2142 (UseCompressedOops && val->bottom_type()->isa_klassptr() &&
2115 2143 adr->bottom_type()->isa_rawptr())) {
2116 2144 val = gvn.transform(new (C, 2) EncodePNode(val, val->bottom_type()->make_narrowoop()));
2117 2145 return new (C, 4) StoreNNode(ctl, mem, adr, adr_type, val);
2118 2146 } else
2119 2147 #endif
2120 2148 {
2121 2149 return new (C, 4) StorePNode(ctl, mem, adr, adr_type, val);
2122 2150 }
2123 2151 }
2124 2152 ShouldNotReachHere();
2125 2153 return (StoreNode*)NULL;
2126 2154 }
2127 2155
2128 2156 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) {
2129 2157 bool require_atomic = true;
2130 2158 return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic);
2131 2159 }
2132 2160
2133 2161
2134 2162 //--------------------------bottom_type----------------------------------------
2135 2163 const Type *StoreNode::bottom_type() const {
2136 2164 return Type::MEMORY;
2137 2165 }
2138 2166
2139 2167 //------------------------------hash-------------------------------------------
2140 2168 uint StoreNode::hash() const {
2141 2169 // unroll addition of interesting fields
2142 2170 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2143 2171
2144 2172 // Since they are not commoned, do not hash them:
2145 2173 return NO_HASH;
2146 2174 }
2147 2175
2148 2176 //------------------------------Ideal------------------------------------------
2149 2177 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
2150 2178 // When a store immediately follows a relevant allocation/initialization,
2151 2179 // try to capture it into the initialization, or hoist it above.
2152 2180 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2153 2181 Node* p = MemNode::Ideal_common(phase, can_reshape);
2154 2182 if (p) return (p == NodeSentinel) ? NULL : p;
2155 2183
2156 2184 Node* mem = in(MemNode::Memory);
2157 2185 Node* address = in(MemNode::Address);
2158 2186
2159 2187 // Back-to-back stores to same address? Fold em up. Generally
2160 2188 // unsafe if I have intervening uses... Also disallowed for StoreCM
2161 2189 // since they must follow each StoreP operation. Redundant StoreCMs
2162 2190 // are eliminated just before matching in final_graph_reshape.
2163 2191 if (mem->is_Store() && phase->eqv_uncast(mem->in(MemNode::Address), address) &&
2164 2192 mem->Opcode() != Op_StoreCM) {
2165 2193 // Looking at a dead closed cycle of memory?
2166 2194 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
2167 2195
2168 2196 assert(Opcode() == mem->Opcode() ||
2169 2197 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw,
2170 2198 "no mismatched stores, except on raw memory");
2171 2199
2172 2200 if (mem->outcnt() == 1 && // check for intervening uses
2173 2201 mem->as_Store()->memory_size() <= this->memory_size()) {
2174 2202 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away.
2175 2203 // For example, 'mem' might be the final state at a conditional return.
2176 2204 // Or, 'mem' might be used by some node which is live at the same time
2177 2205 // 'this' is live, which might be unschedulable. So, require exactly
2178 2206 // ONE user, the 'this' store, until such time as we clone 'mem' for
2179 2207 // each of 'mem's uses (thus making the exactly-1-user-rule hold true).
2180 2208 if (can_reshape) { // (%%% is this an anachronism?)
2181 2209 set_req_X(MemNode::Memory, mem->in(MemNode::Memory),
2182 2210 phase->is_IterGVN());
2183 2211 } else {
2184 2212 // It's OK to do this in the parser, since DU info is always accurate,
2185 2213 // and the parser always refers to nodes via SafePointNode maps.
2186 2214 set_req(MemNode::Memory, mem->in(MemNode::Memory));
2187 2215 }
2188 2216 return this;
2189 2217 }
2190 2218 }
2191 2219
2192 2220 // Capture an unaliased, unconditional, simple store into an initializer.
2193 2221 // Or, if it is independent of the allocation, hoist it above the allocation.
2194 2222 if (ReduceFieldZeroing && /*can_reshape &&*/
2195 2223 mem->is_Proj() && mem->in(0)->is_Initialize()) {
2196 2224 InitializeNode* init = mem->in(0)->as_Initialize();
2197 2225 intptr_t offset = init->can_capture_store(this, phase);
2198 2226 if (offset > 0) {
2199 2227 Node* moved = init->capture_store(this, offset, phase);
2200 2228 // If the InitializeNode captured me, it made a raw copy of me,
2201 2229 // and I need to disappear.
2202 2230 if (moved != NULL) {
2203 2231 // %%% hack to ensure that Ideal returns a new node:
2204 2232 mem = MergeMemNode::make(phase->C, mem);
2205 2233 return mem; // fold me away
2206 2234 }
2207 2235 }
2208 2236 }
2209 2237
2210 2238 return NULL; // No further progress
2211 2239 }
2212 2240
2213 2241 //------------------------------Value-----------------------------------------
2214 2242 const Type *StoreNode::Value( PhaseTransform *phase ) const {
2215 2243 // Either input is TOP ==> the result is TOP
2216 2244 const Type *t1 = phase->type( in(MemNode::Memory) );
2217 2245 if( t1 == Type::TOP ) return Type::TOP;
2218 2246 const Type *t2 = phase->type( in(MemNode::Address) );
2219 2247 if( t2 == Type::TOP ) return Type::TOP;
2220 2248 const Type *t3 = phase->type( in(MemNode::ValueIn) );
2221 2249 if( t3 == Type::TOP ) return Type::TOP;
2222 2250 return Type::MEMORY;
2223 2251 }
2224 2252
2225 2253 //------------------------------Identity---------------------------------------
2226 2254 // Remove redundant stores:
2227 2255 // Store(m, p, Load(m, p)) changes to m.
2228 2256 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
2229 2257 Node *StoreNode::Identity( PhaseTransform *phase ) {
2230 2258 Node* mem = in(MemNode::Memory);
2231 2259 Node* adr = in(MemNode::Address);
2232 2260 Node* val = in(MemNode::ValueIn);
2233 2261
2234 2262 // Load then Store? Then the Store is useless
2235 2263 if (val->is_Load() &&
2236 2264 phase->eqv_uncast( val->in(MemNode::Address), adr ) &&
2237 2265 phase->eqv_uncast( val->in(MemNode::Memory ), mem ) &&
2238 2266 val->as_Load()->store_Opcode() == Opcode()) {
2239 2267 return mem;
2240 2268 }
2241 2269
2242 2270 // Two stores in a row of the same value?
2243 2271 if (mem->is_Store() &&
2244 2272 phase->eqv_uncast( mem->in(MemNode::Address), adr ) &&
2245 2273 phase->eqv_uncast( mem->in(MemNode::ValueIn), val ) &&
2246 2274 mem->Opcode() == Opcode()) {
2247 2275 return mem;
2248 2276 }
2249 2277
2250 2278 // Store of zero anywhere into a freshly-allocated object?
2251 2279 // Then the store is useless.
2252 2280 // (It must already have been captured by the InitializeNode.)
2253 2281 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
2254 2282 // a newly allocated object is already all-zeroes everywhere
2255 2283 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
2256 2284 return mem;
2257 2285 }
2258 2286
2259 2287 // the store may also apply to zero-bits in an earlier object
2260 2288 Node* prev_mem = find_previous_store(phase);
2261 2289 // Steps (a), (b): Walk past independent stores to find an exact match.
2262 2290 if (prev_mem != NULL) {
2263 2291 Node* prev_val = can_see_stored_value(prev_mem, phase);
2264 2292 if (prev_val != NULL && phase->eqv(prev_val, val)) {
2265 2293 // prev_val and val might differ by a cast; it would be good
2266 2294 // to keep the more informative of the two.
2267 2295 return mem;
2268 2296 }
2269 2297 }
2270 2298 }
2271 2299
2272 2300 return this;
2273 2301 }
2274 2302
2275 2303 //------------------------------match_edge-------------------------------------
2276 2304 // Do we Match on this edge index or not? Match only memory & value
2277 2305 uint StoreNode::match_edge(uint idx) const {
2278 2306 return idx == MemNode::Address || idx == MemNode::ValueIn;
2279 2307 }
2280 2308
2281 2309 //------------------------------cmp--------------------------------------------
2282 2310 // Do not common stores up together. They generally have to be split
2283 2311 // back up anyways, so do not bother.
2284 2312 uint StoreNode::cmp( const Node &n ) const {
2285 2313 return (&n == this); // Always fail except on self
2286 2314 }
2287 2315
2288 2316 //------------------------------Ideal_masked_input-----------------------------
2289 2317 // Check for a useless mask before a partial-word store
2290 2318 // (StoreB ... (AndI valIn conIa) )
2291 2319 // If (conIa & mask == mask) this simplifies to
2292 2320 // (StoreB ... (valIn) )
2293 2321 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
2294 2322 Node *val = in(MemNode::ValueIn);
2295 2323 if( val->Opcode() == Op_AndI ) {
2296 2324 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2297 2325 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
2298 2326 set_req(MemNode::ValueIn, val->in(1));
2299 2327 return this;
2300 2328 }
2301 2329 }
2302 2330 return NULL;
2303 2331 }
2304 2332
2305 2333
2306 2334 //------------------------------Ideal_sign_extended_input----------------------
2307 2335 // Check for useless sign-extension before a partial-word store
2308 2336 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
2309 2337 // If (conIL == conIR && conIR <= num_bits) this simplifies to
2310 2338 // (StoreB ... (valIn) )
2311 2339 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
2312 2340 Node *val = in(MemNode::ValueIn);
2313 2341 if( val->Opcode() == Op_RShiftI ) {
2314 2342 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2315 2343 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
2316 2344 Node *shl = val->in(1);
2317 2345 if( shl->Opcode() == Op_LShiftI ) {
2318 2346 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
2319 2347 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
2320 2348 set_req(MemNode::ValueIn, shl->in(1));
2321 2349 return this;
2322 2350 }
2323 2351 }
2324 2352 }
2325 2353 }
2326 2354 return NULL;
2327 2355 }
2328 2356
2329 2357 //------------------------------value_never_loaded-----------------------------------
2330 2358 // Determine whether there are any possible loads of the value stored.
2331 2359 // For simplicity, we actually check if there are any loads from the
2332 2360 // address stored to, not just for loads of the value stored by this node.
2333 2361 //
2334 2362 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
2335 2363 Node *adr = in(Address);
2336 2364 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
2337 2365 if (adr_oop == NULL)
2338 2366 return false;
2339 2367 if (!adr_oop->is_known_instance_field())
2340 2368 return false; // if not a distinct instance, there may be aliases of the address
2341 2369 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
2342 2370 Node *use = adr->fast_out(i);
2343 2371 int opc = use->Opcode();
2344 2372 if (use->is_Load() || use->is_LoadStore()) {
2345 2373 return false;
2346 2374 }
2347 2375 }
2348 2376 return true;
2349 2377 }
2350 2378
2351 2379 //=============================================================================
2352 2380 //------------------------------Ideal------------------------------------------
2353 2381 // If the store is from an AND mask that leaves the low bits untouched, then
2354 2382 // we can skip the AND operation. If the store is from a sign-extension
2355 2383 // (a left shift, then right shift) we can skip both.
2356 2384 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
2357 2385 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
2358 2386 if( progress != NULL ) return progress;
2359 2387
2360 2388 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
2361 2389 if( progress != NULL ) return progress;
2362 2390
2363 2391 // Finally check the default case
2364 2392 return StoreNode::Ideal(phase, can_reshape);
2365 2393 }
2366 2394
2367 2395 //=============================================================================
2368 2396 //------------------------------Ideal------------------------------------------
2369 2397 // If the store is from an AND mask that leaves the low bits untouched, then
2370 2398 // we can skip the AND operation
2371 2399 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
2372 2400 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
2373 2401 if( progress != NULL ) return progress;
2374 2402
2375 2403 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
2376 2404 if( progress != NULL ) return progress;
2377 2405
2378 2406 // Finally check the default case
2379 2407 return StoreNode::Ideal(phase, can_reshape);
2380 2408 }
2381 2409
2382 2410 //=============================================================================
2383 2411 //------------------------------Identity---------------------------------------
2384 2412 Node *StoreCMNode::Identity( PhaseTransform *phase ) {
2385 2413 // No need to card mark when storing a null ptr
2386 2414 Node* my_store = in(MemNode::OopStore);
2387 2415 if (my_store->is_Store()) {
2388 2416 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
2389 2417 if( t1 == TypePtr::NULL_PTR ) {
2390 2418 return in(MemNode::Memory);
2391 2419 }
2392 2420 }
2393 2421 return this;
2394 2422 }
2395 2423
2396 2424 //=============================================================================
2397 2425 //------------------------------Ideal---------------------------------------
2398 2426 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){
2399 2427 Node* progress = StoreNode::Ideal(phase, can_reshape);
2400 2428 if (progress != NULL) return progress;
2401 2429
2402 2430 Node* my_store = in(MemNode::OopStore);
2403 2431 if (my_store->is_MergeMem()) {
2404 2432 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx());
2405 2433 set_req(MemNode::OopStore, mem);
2406 2434 return this;
2407 2435 }
2408 2436
2409 2437 return NULL;
2410 2438 }
2411 2439
2412 2440 //------------------------------Value-----------------------------------------
2413 2441 const Type *StoreCMNode::Value( PhaseTransform *phase ) const {
2414 2442 // Either input is TOP ==> the result is TOP
2415 2443 const Type *t = phase->type( in(MemNode::Memory) );
2416 2444 if( t == Type::TOP ) return Type::TOP;
2417 2445 t = phase->type( in(MemNode::Address) );
2418 2446 if( t == Type::TOP ) return Type::TOP;
2419 2447 t = phase->type( in(MemNode::ValueIn) );
2420 2448 if( t == Type::TOP ) return Type::TOP;
2421 2449 // If extra input is TOP ==> the result is TOP
2422 2450 t = phase->type( in(MemNode::OopStore) );
2423 2451 if( t == Type::TOP ) return Type::TOP;
2424 2452
2425 2453 return StoreNode::Value( phase );
2426 2454 }
2427 2455
2428 2456
2429 2457 //=============================================================================
2430 2458 //----------------------------------SCMemProjNode------------------------------
2431 2459 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const
2432 2460 {
2433 2461 return bottom_type();
2434 2462 }
2435 2463
2436 2464 //=============================================================================
2437 2465 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : Node(5) {
2438 2466 init_req(MemNode::Control, c );
2439 2467 init_req(MemNode::Memory , mem);
2440 2468 init_req(MemNode::Address, adr);
2441 2469 init_req(MemNode::ValueIn, val);
2442 2470 init_req( ExpectedIn, ex );
2443 2471 init_class_id(Class_LoadStore);
2444 2472
2445 2473 }
2446 2474
2447 2475 //=============================================================================
2448 2476 //-------------------------------adr_type--------------------------------------
2449 2477 // Do we Match on this edge index or not? Do not match memory
2450 2478 const TypePtr* ClearArrayNode::adr_type() const {
2451 2479 Node *adr = in(3);
2452 2480 return MemNode::calculate_adr_type(adr->bottom_type());
2453 2481 }
2454 2482
2455 2483 //------------------------------match_edge-------------------------------------
2456 2484 // Do we Match on this edge index or not? Do not match memory
2457 2485 uint ClearArrayNode::match_edge(uint idx) const {
2458 2486 return idx > 1;
2459 2487 }
2460 2488
2461 2489 //------------------------------Identity---------------------------------------
2462 2490 // Clearing a zero length array does nothing
2463 2491 Node *ClearArrayNode::Identity( PhaseTransform *phase ) {
2464 2492 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
2465 2493 }
2466 2494
2467 2495 //------------------------------Idealize---------------------------------------
2468 2496 // Clearing a short array is faster with stores
2469 2497 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){
2470 2498 const int unit = BytesPerLong;
2471 2499 const TypeX* t = phase->type(in(2))->isa_intptr_t();
2472 2500 if (!t) return NULL;
2473 2501 if (!t->is_con()) return NULL;
2474 2502 intptr_t raw_count = t->get_con();
2475 2503 intptr_t size = raw_count;
2476 2504 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
2477 2505 // Clearing nothing uses the Identity call.
2478 2506 // Negative clears are possible on dead ClearArrays
2479 2507 // (see jck test stmt114.stmt11402.val).
2480 2508 if (size <= 0 || size % unit != 0) return NULL;
2481 2509 intptr_t count = size / unit;
2482 2510 // Length too long; use fast hardware clear
2483 2511 if (size > Matcher::init_array_short_size) return NULL;
2484 2512 Node *mem = in(1);
2485 2513 if( phase->type(mem)==Type::TOP ) return NULL;
2486 2514 Node *adr = in(3);
2487 2515 const Type* at = phase->type(adr);
2488 2516 if( at==Type::TOP ) return NULL;
2489 2517 const TypePtr* atp = at->isa_ptr();
2490 2518 // adjust atp to be the correct array element address type
2491 2519 if (atp == NULL) atp = TypePtr::BOTTOM;
2492 2520 else atp = atp->add_offset(Type::OffsetBot);
2493 2521 // Get base for derived pointer purposes
2494 2522 if( adr->Opcode() != Op_AddP ) Unimplemented();
2495 2523 Node *base = adr->in(1);
2496 2524
2497 2525 Node *zero = phase->makecon(TypeLong::ZERO);
2498 2526 Node *off = phase->MakeConX(BytesPerLong);
2499 2527 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
2500 2528 count--;
2501 2529 while( count-- ) {
2502 2530 mem = phase->transform(mem);
2503 2531 adr = phase->transform(new (phase->C, 4) AddPNode(base,adr,off));
2504 2532 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
2505 2533 }
2506 2534 return mem;
2507 2535 }
2508 2536
2509 2537 //----------------------------step_through----------------------------------
2510 2538 // Return allocation input memory edge if it is different instance
2511 2539 // or itself if it is the one we are looking for.
2512 2540 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) {
2513 2541 Node* n = *np;
2514 2542 assert(n->is_ClearArray(), "sanity");
2515 2543 intptr_t offset;
2516 2544 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
2517 2545 // This method is called only before Allocate nodes are expanded during
2518 2546 // macro nodes expansion. Before that ClearArray nodes are only generated
2519 2547 // in LibraryCallKit::generate_arraycopy() which follows allocations.
2520 2548 assert(alloc != NULL, "should have allocation");
2521 2549 if (alloc->_idx == instance_id) {
2522 2550 // Can not bypass initialization of the instance we are looking for.
2523 2551 return false;
2524 2552 }
2525 2553 // Otherwise skip it.
2526 2554 InitializeNode* init = alloc->initialization();
2527 2555 if (init != NULL)
2528 2556 *np = init->in(TypeFunc::Memory);
2529 2557 else
2530 2558 *np = alloc->in(TypeFunc::Memory);
2531 2559 return true;
2532 2560 }
2533 2561
2534 2562 //----------------------------clear_memory-------------------------------------
2535 2563 // Generate code to initialize object storage to zero.
2536 2564 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2537 2565 intptr_t start_offset,
2538 2566 Node* end_offset,
2539 2567 PhaseGVN* phase) {
2540 2568 Compile* C = phase->C;
2541 2569 intptr_t offset = start_offset;
2542 2570
2543 2571 int unit = BytesPerLong;
2544 2572 if ((offset % unit) != 0) {
2545 2573 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(offset));
2546 2574 adr = phase->transform(adr);
2547 2575 const TypePtr* atp = TypeRawPtr::BOTTOM;
2548 2576 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
2549 2577 mem = phase->transform(mem);
2550 2578 offset += BytesPerInt;
2551 2579 }
2552 2580 assert((offset % unit) == 0, "");
2553 2581
2554 2582 // Initialize the remaining stuff, if any, with a ClearArray.
2555 2583 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
2556 2584 }
2557 2585
2558 2586 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2559 2587 Node* start_offset,
2560 2588 Node* end_offset,
2561 2589 PhaseGVN* phase) {
2562 2590 if (start_offset == end_offset) {
2563 2591 // nothing to do
2564 2592 return mem;
2565 2593 }
2566 2594
2567 2595 Compile* C = phase->C;
2568 2596 int unit = BytesPerLong;
2569 2597 Node* zbase = start_offset;
2570 2598 Node* zend = end_offset;
2571 2599
2572 2600 // Scale to the unit required by the CPU:
2573 2601 if (!Matcher::init_array_count_is_in_bytes) {
2574 2602 Node* shift = phase->intcon(exact_log2(unit));
2575 2603 zbase = phase->transform( new(C,3) URShiftXNode(zbase, shift) );
2576 2604 zend = phase->transform( new(C,3) URShiftXNode(zend, shift) );
2577 2605 }
2578 2606
2579 2607 Node* zsize = phase->transform( new(C,3) SubXNode(zend, zbase) );
2580 2608 Node* zinit = phase->zerocon((unit == BytesPerLong) ? T_LONG : T_INT);
2581 2609
2582 2610 // Bulk clear double-words
2583 2611 Node* adr = phase->transform( new(C,4) AddPNode(dest, dest, start_offset) );
2584 2612 mem = new (C, 4) ClearArrayNode(ctl, mem, zsize, adr);
2585 2613 return phase->transform(mem);
2586 2614 }
2587 2615
2588 2616 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2589 2617 intptr_t start_offset,
2590 2618 intptr_t end_offset,
2591 2619 PhaseGVN* phase) {
2592 2620 if (start_offset == end_offset) {
2593 2621 // nothing to do
2594 2622 return mem;
2595 2623 }
2596 2624
2597 2625 Compile* C = phase->C;
2598 2626 assert((end_offset % BytesPerInt) == 0, "odd end offset");
2599 2627 intptr_t done_offset = end_offset;
2600 2628 if ((done_offset % BytesPerLong) != 0) {
2601 2629 done_offset -= BytesPerInt;
2602 2630 }
2603 2631 if (done_offset > start_offset) {
2604 2632 mem = clear_memory(ctl, mem, dest,
2605 2633 start_offset, phase->MakeConX(done_offset), phase);
2606 2634 }
2607 2635 if (done_offset < end_offset) { // emit the final 32-bit store
2608 2636 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(done_offset));
2609 2637 adr = phase->transform(adr);
2610 2638 const TypePtr* atp = TypeRawPtr::BOTTOM;
2611 2639 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
2612 2640 mem = phase->transform(mem);
2613 2641 done_offset += BytesPerInt;
2614 2642 }
2615 2643 assert(done_offset == end_offset, "");
2616 2644 return mem;
2617 2645 }
2618 2646
2619 2647 //=============================================================================
2620 2648 // Do not match memory edge.
2621 2649 uint StrIntrinsicNode::match_edge(uint idx) const {
2622 2650 return idx == 2 || idx == 3;
2623 2651 }
2624 2652
2625 2653 //------------------------------Ideal------------------------------------------
2626 2654 // Return a node which is more "ideal" than the current node. Strip out
2627 2655 // control copies
2628 2656 Node *StrIntrinsicNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2629 2657 if (remove_dead_region(phase, can_reshape)) return this;
2630 2658
2631 2659 if (can_reshape) {
2632 2660 Node* mem = phase->transform(in(MemNode::Memory));
2633 2661 // If transformed to a MergeMem, get the desired slice
2634 2662 uint alias_idx = phase->C->get_alias_index(adr_type());
2635 2663 mem = mem->is_MergeMem() ? mem->as_MergeMem()->memory_at(alias_idx) : mem;
2636 2664 if (mem != in(MemNode::Memory)) {
2637 2665 set_req(MemNode::Memory, mem);
2638 2666 return this;
2639 2667 }
2640 2668 }
2641 2669 return NULL;
2642 2670 }
2643 2671
2644 2672 //=============================================================================
2645 2673 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
2646 2674 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
2647 2675 _adr_type(C->get_adr_type(alias_idx))
2648 2676 {
2649 2677 init_class_id(Class_MemBar);
2650 2678 Node* top = C->top();
2651 2679 init_req(TypeFunc::I_O,top);
2652 2680 init_req(TypeFunc::FramePtr,top);
2653 2681 init_req(TypeFunc::ReturnAdr,top);
2654 2682 if (precedent != NULL)
2655 2683 init_req(TypeFunc::Parms, precedent);
2656 2684 }
2657 2685
2658 2686 //------------------------------cmp--------------------------------------------
2659 2687 uint MemBarNode::hash() const { return NO_HASH; }
2660 2688 uint MemBarNode::cmp( const Node &n ) const {
2661 2689 return (&n == this); // Always fail except on self
2662 2690 }
2663 2691
2664 2692 //------------------------------make-------------------------------------------
2665 2693 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
2666 2694 int len = Precedent + (pn == NULL? 0: 1);
2667 2695 switch (opcode) {
2668 2696 case Op_MemBarAcquire: return new(C, len) MemBarAcquireNode(C, atp, pn);
2669 2697 case Op_MemBarRelease: return new(C, len) MemBarReleaseNode(C, atp, pn);
2670 2698 case Op_MemBarAcquireLock: return new(C, len) MemBarAcquireLockNode(C, atp, pn);
2671 2699 case Op_MemBarReleaseLock: return new(C, len) MemBarReleaseLockNode(C, atp, pn);
2672 2700 case Op_MemBarVolatile: return new(C, len) MemBarVolatileNode(C, atp, pn);
2673 2701 case Op_MemBarCPUOrder: return new(C, len) MemBarCPUOrderNode(C, atp, pn);
2674 2702 case Op_Initialize: return new(C, len) InitializeNode(C, atp, pn);
2675 2703 default: ShouldNotReachHere(); return NULL;
2676 2704 }
2677 2705 }
2678 2706
2679 2707 //------------------------------Ideal------------------------------------------
2680 2708 // Return a node which is more "ideal" than the current node. Strip out
2681 2709 // control copies
2682 2710 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2683 2711 if (remove_dead_region(phase, can_reshape)) return this;
2684 2712
2685 2713 // Eliminate volatile MemBars for scalar replaced objects.
2686 2714 if (can_reshape && req() == (Precedent+1) &&
2687 2715 (Opcode() == Op_MemBarAcquire || Opcode() == Op_MemBarVolatile)) {
2688 2716 // Volatile field loads and stores.
2689 2717 Node* my_mem = in(MemBarNode::Precedent);
2690 2718 if (my_mem != NULL && my_mem->is_Mem()) {
2691 2719 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
2692 2720 // Check for scalar replaced object reference.
2693 2721 if( t_oop != NULL && t_oop->is_known_instance_field() &&
2694 2722 t_oop->offset() != Type::OffsetBot &&
2695 2723 t_oop->offset() != Type::OffsetTop) {
2696 2724 // Replace MemBar projections by its inputs.
2697 2725 PhaseIterGVN* igvn = phase->is_IterGVN();
2698 2726 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
2699 2727 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
2700 2728 // Must return either the original node (now dead) or a new node
2701 2729 // (Do not return a top here, since that would break the uniqueness of top.)
2702 2730 return new (phase->C, 1) ConINode(TypeInt::ZERO);
2703 2731 }
2704 2732 }
2705 2733 }
2706 2734 return NULL;
2707 2735 }
2708 2736
2709 2737 //------------------------------Value------------------------------------------
2710 2738 const Type *MemBarNode::Value( PhaseTransform *phase ) const {
2711 2739 if( !in(0) ) return Type::TOP;
2712 2740 if( phase->type(in(0)) == Type::TOP )
2713 2741 return Type::TOP;
2714 2742 return TypeTuple::MEMBAR;
2715 2743 }
2716 2744
2717 2745 //------------------------------match------------------------------------------
2718 2746 // Construct projections for memory.
2719 2747 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
2720 2748 switch (proj->_con) {
2721 2749 case TypeFunc::Control:
2722 2750 case TypeFunc::Memory:
2723 2751 return new (m->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
2724 2752 }
2725 2753 ShouldNotReachHere();
2726 2754 return NULL;
2727 2755 }
2728 2756
2729 2757 //===========================InitializeNode====================================
2730 2758 // SUMMARY:
2731 2759 // This node acts as a memory barrier on raw memory, after some raw stores.
2732 2760 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
2733 2761 // The Initialize can 'capture' suitably constrained stores as raw inits.
2734 2762 // It can coalesce related raw stores into larger units (called 'tiles').
2735 2763 // It can avoid zeroing new storage for memory units which have raw inits.
2736 2764 // At macro-expansion, it is marked 'complete', and does not optimize further.
2737 2765 //
2738 2766 // EXAMPLE:
2739 2767 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
2740 2768 // ctl = incoming control; mem* = incoming memory
2741 2769 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
2742 2770 // First allocate uninitialized memory and fill in the header:
2743 2771 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
2744 2772 // ctl := alloc.Control; mem* := alloc.Memory*
2745 2773 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
2746 2774 // Then initialize to zero the non-header parts of the raw memory block:
2747 2775 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
2748 2776 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
2749 2777 // After the initialize node executes, the object is ready for service:
2750 2778 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
2751 2779 // Suppose its body is immediately initialized as {1,2}:
2752 2780 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
2753 2781 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
2754 2782 // mem.SLICE(#short[*]) := store2
2755 2783 //
2756 2784 // DETAILS:
2757 2785 // An InitializeNode collects and isolates object initialization after
2758 2786 // an AllocateNode and before the next possible safepoint. As a
2759 2787 // memory barrier (MemBarNode), it keeps critical stores from drifting
2760 2788 // down past any safepoint or any publication of the allocation.
2761 2789 // Before this barrier, a newly-allocated object may have uninitialized bits.
2762 2790 // After this barrier, it may be treated as a real oop, and GC is allowed.
2763 2791 //
2764 2792 // The semantics of the InitializeNode include an implicit zeroing of
2765 2793 // the new object from object header to the end of the object.
2766 2794 // (The object header and end are determined by the AllocateNode.)
2767 2795 //
2768 2796 // Certain stores may be added as direct inputs to the InitializeNode.
2769 2797 // These stores must update raw memory, and they must be to addresses
2770 2798 // derived from the raw address produced by AllocateNode, and with
2771 2799 // a constant offset. They must be ordered by increasing offset.
2772 2800 // The first one is at in(RawStores), the last at in(req()-1).
2773 2801 // Unlike most memory operations, they are not linked in a chain,
2774 2802 // but are displayed in parallel as users of the rawmem output of
2775 2803 // the allocation.
2776 2804 //
2777 2805 // (See comments in InitializeNode::capture_store, which continue
2778 2806 // the example given above.)
2779 2807 //
2780 2808 // When the associated Allocate is macro-expanded, the InitializeNode
2781 2809 // may be rewritten to optimize collected stores. A ClearArrayNode
2782 2810 // may also be created at that point to represent any required zeroing.
2783 2811 // The InitializeNode is then marked 'complete', prohibiting further
2784 2812 // capturing of nearby memory operations.
2785 2813 //
2786 2814 // During macro-expansion, all captured initializations which store
2787 2815 // constant values of 32 bits or smaller are coalesced (if advantageous)
2788 2816 // into larger 'tiles' 32 or 64 bits. This allows an object to be
2789 2817 // initialized in fewer memory operations. Memory words which are
2790 2818 // covered by neither tiles nor non-constant stores are pre-zeroed
2791 2819 // by explicit stores of zero. (The code shape happens to do all
2792 2820 // zeroing first, then all other stores, with both sequences occurring
2793 2821 // in order of ascending offsets.)
2794 2822 //
2795 2823 // Alternatively, code may be inserted between an AllocateNode and its
2796 2824 // InitializeNode, to perform arbitrary initialization of the new object.
2797 2825 // E.g., the object copying intrinsics insert complex data transfers here.
2798 2826 // The initialization must then be marked as 'complete' disable the
2799 2827 // built-in zeroing semantics and the collection of initializing stores.
2800 2828 //
2801 2829 // While an InitializeNode is incomplete, reads from the memory state
2802 2830 // produced by it are optimizable if they match the control edge and
2803 2831 // new oop address associated with the allocation/initialization.
2804 2832 // They return a stored value (if the offset matches) or else zero.
2805 2833 // A write to the memory state, if it matches control and address,
2806 2834 // and if it is to a constant offset, may be 'captured' by the
2807 2835 // InitializeNode. It is cloned as a raw memory operation and rewired
2808 2836 // inside the initialization, to the raw oop produced by the allocation.
2809 2837 // Operations on addresses which are provably distinct (e.g., to
2810 2838 // other AllocateNodes) are allowed to bypass the initialization.
2811 2839 //
2812 2840 // The effect of all this is to consolidate object initialization
2813 2841 // (both arrays and non-arrays, both piecewise and bulk) into a
2814 2842 // single location, where it can be optimized as a unit.
2815 2843 //
2816 2844 // Only stores with an offset less than TrackedInitializationLimit words
2817 2845 // will be considered for capture by an InitializeNode. This puts a
2818 2846 // reasonable limit on the complexity of optimized initializations.
2819 2847
2820 2848 //---------------------------InitializeNode------------------------------------
2821 2849 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
2822 2850 : _is_complete(false),
2823 2851 MemBarNode(C, adr_type, rawoop)
2824 2852 {
2825 2853 init_class_id(Class_Initialize);
2826 2854
2827 2855 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
2828 2856 assert(in(RawAddress) == rawoop, "proper init");
2829 2857 // Note: allocation() can be NULL, for secondary initialization barriers
2830 2858 }
2831 2859
2832 2860 // Since this node is not matched, it will be processed by the
2833 2861 // register allocator. Declare that there are no constraints
2834 2862 // on the allocation of the RawAddress edge.
2835 2863 const RegMask &InitializeNode::in_RegMask(uint idx) const {
2836 2864 // This edge should be set to top, by the set_complete. But be conservative.
2837 2865 if (idx == InitializeNode::RawAddress)
2838 2866 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
2839 2867 return RegMask::Empty;
2840 2868 }
2841 2869
2842 2870 Node* InitializeNode::memory(uint alias_idx) {
2843 2871 Node* mem = in(Memory);
2844 2872 if (mem->is_MergeMem()) {
2845 2873 return mem->as_MergeMem()->memory_at(alias_idx);
2846 2874 } else {
2847 2875 // incoming raw memory is not split
2848 2876 return mem;
2849 2877 }
2850 2878 }
2851 2879
2852 2880 bool InitializeNode::is_non_zero() {
2853 2881 if (is_complete()) return false;
2854 2882 remove_extra_zeroes();
2855 2883 return (req() > RawStores);
2856 2884 }
2857 2885
2858 2886 void InitializeNode::set_complete(PhaseGVN* phase) {
2859 2887 assert(!is_complete(), "caller responsibility");
2860 2888 _is_complete = true;
2861 2889
2862 2890 // After this node is complete, it contains a bunch of
2863 2891 // raw-memory initializations. There is no need for
2864 2892 // it to have anything to do with non-raw memory effects.
2865 2893 // Therefore, tell all non-raw users to re-optimize themselves,
2866 2894 // after skipping the memory effects of this initialization.
2867 2895 PhaseIterGVN* igvn = phase->is_IterGVN();
2868 2896 if (igvn) igvn->add_users_to_worklist(this);
2869 2897 }
2870 2898
2871 2899 // convenience function
2872 2900 // return false if the init contains any stores already
2873 2901 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
2874 2902 InitializeNode* init = initialization();
2875 2903 if (init == NULL || init->is_complete()) return false;
2876 2904 init->remove_extra_zeroes();
2877 2905 // for now, if this allocation has already collected any inits, bail:
2878 2906 if (init->is_non_zero()) return false;
2879 2907 init->set_complete(phase);
2880 2908 return true;
2881 2909 }
2882 2910
2883 2911 void InitializeNode::remove_extra_zeroes() {
2884 2912 if (req() == RawStores) return;
2885 2913 Node* zmem = zero_memory();
2886 2914 uint fill = RawStores;
2887 2915 for (uint i = fill; i < req(); i++) {
2888 2916 Node* n = in(i);
2889 2917 if (n->is_top() || n == zmem) continue; // skip
2890 2918 if (fill < i) set_req(fill, n); // compact
2891 2919 ++fill;
2892 2920 }
2893 2921 // delete any empty spaces created:
2894 2922 while (fill < req()) {
2895 2923 del_req(fill);
2896 2924 }
2897 2925 }
2898 2926
2899 2927 // Helper for remembering which stores go with which offsets.
2900 2928 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
2901 2929 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
2902 2930 intptr_t offset = -1;
2903 2931 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
2904 2932 phase, offset);
2905 2933 if (base == NULL) return -1; // something is dead,
2906 2934 if (offset < 0) return -1; // dead, dead
2907 2935 return offset;
2908 2936 }
2909 2937
2910 2938 // Helper for proving that an initialization expression is
2911 2939 // "simple enough" to be folded into an object initialization.
2912 2940 // Attempts to prove that a store's initial value 'n' can be captured
2913 2941 // within the initialization without creating a vicious cycle, such as:
2914 2942 // { Foo p = new Foo(); p.next = p; }
2915 2943 // True for constants and parameters and small combinations thereof.
2916 2944 bool InitializeNode::detect_init_independence(Node* n,
2917 2945 bool st_is_pinned,
2918 2946 int& count) {
2919 2947 if (n == NULL) return true; // (can this really happen?)
2920 2948 if (n->is_Proj()) n = n->in(0);
2921 2949 if (n == this) return false; // found a cycle
2922 2950 if (n->is_Con()) return true;
2923 2951 if (n->is_Start()) return true; // params, etc., are OK
2924 2952 if (n->is_Root()) return true; // even better
2925 2953
2926 2954 Node* ctl = n->in(0);
2927 2955 if (ctl != NULL && !ctl->is_top()) {
2928 2956 if (ctl->is_Proj()) ctl = ctl->in(0);
2929 2957 if (ctl == this) return false;
2930 2958
2931 2959 // If we already know that the enclosing memory op is pinned right after
2932 2960 // the init, then any control flow that the store has picked up
2933 2961 // must have preceded the init, or else be equal to the init.
2934 2962 // Even after loop optimizations (which might change control edges)
2935 2963 // a store is never pinned *before* the availability of its inputs.
2936 2964 if (!MemNode::all_controls_dominate(n, this))
2937 2965 return false; // failed to prove a good control
2938 2966
2939 2967 }
2940 2968
2941 2969 // Check data edges for possible dependencies on 'this'.
2942 2970 if ((count += 1) > 20) return false; // complexity limit
2943 2971 for (uint i = 1; i < n->req(); i++) {
2944 2972 Node* m = n->in(i);
2945 2973 if (m == NULL || m == n || m->is_top()) continue;
2946 2974 uint first_i = n->find_edge(m);
2947 2975 if (i != first_i) continue; // process duplicate edge just once
2948 2976 if (!detect_init_independence(m, st_is_pinned, count)) {
2949 2977 return false;
2950 2978 }
2951 2979 }
2952 2980
2953 2981 return true;
2954 2982 }
2955 2983
2956 2984 // Here are all the checks a Store must pass before it can be moved into
2957 2985 // an initialization. Returns zero if a check fails.
2958 2986 // On success, returns the (constant) offset to which the store applies,
2959 2987 // within the initialized memory.
2960 2988 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase) {
2961 2989 const int FAIL = 0;
2962 2990 if (st->req() != MemNode::ValueIn + 1)
2963 2991 return FAIL; // an inscrutable StoreNode (card mark?)
2964 2992 Node* ctl = st->in(MemNode::Control);
2965 2993 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
2966 2994 return FAIL; // must be unconditional after the initialization
2967 2995 Node* mem = st->in(MemNode::Memory);
2968 2996 if (!(mem->is_Proj() && mem->in(0) == this))
2969 2997 return FAIL; // must not be preceded by other stores
2970 2998 Node* adr = st->in(MemNode::Address);
2971 2999 intptr_t offset;
2972 3000 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
2973 3001 if (alloc == NULL)
2974 3002 return FAIL; // inscrutable address
2975 3003 if (alloc != allocation())
2976 3004 return FAIL; // wrong allocation! (store needs to float up)
2977 3005 Node* val = st->in(MemNode::ValueIn);
2978 3006 int complexity_count = 0;
2979 3007 if (!detect_init_independence(val, true, complexity_count))
2980 3008 return FAIL; // stored value must be 'simple enough'
2981 3009
2982 3010 return offset; // success
2983 3011 }
2984 3012
2985 3013 // Find the captured store in(i) which corresponds to the range
2986 3014 // [start..start+size) in the initialized object.
2987 3015 // If there is one, return its index i. If there isn't, return the
2988 3016 // negative of the index where it should be inserted.
2989 3017 // Return 0 if the queried range overlaps an initialization boundary
2990 3018 // or if dead code is encountered.
2991 3019 // If size_in_bytes is zero, do not bother with overlap checks.
2992 3020 int InitializeNode::captured_store_insertion_point(intptr_t start,
2993 3021 int size_in_bytes,
2994 3022 PhaseTransform* phase) {
2995 3023 const int FAIL = 0, MAX_STORE = BytesPerLong;
2996 3024
2997 3025 if (is_complete())
2998 3026 return FAIL; // arraycopy got here first; punt
2999 3027
3000 3028 assert(allocation() != NULL, "must be present");
3001 3029
3002 3030 // no negatives, no header fields:
3003 3031 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
3004 3032
3005 3033 // after a certain size, we bail out on tracking all the stores:
3006 3034 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3007 3035 if (start >= ti_limit) return FAIL;
3008 3036
3009 3037 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
3010 3038 if (i >= limit) return -(int)i; // not found; here is where to put it
3011 3039
3012 3040 Node* st = in(i);
3013 3041 intptr_t st_off = get_store_offset(st, phase);
3014 3042 if (st_off < 0) {
3015 3043 if (st != zero_memory()) {
3016 3044 return FAIL; // bail out if there is dead garbage
3017 3045 }
3018 3046 } else if (st_off > start) {
3019 3047 // ...we are done, since stores are ordered
3020 3048 if (st_off < start + size_in_bytes) {
3021 3049 return FAIL; // the next store overlaps
3022 3050 }
3023 3051 return -(int)i; // not found; here is where to put it
3024 3052 } else if (st_off < start) {
3025 3053 if (size_in_bytes != 0 &&
3026 3054 start < st_off + MAX_STORE &&
3027 3055 start < st_off + st->as_Store()->memory_size()) {
3028 3056 return FAIL; // the previous store overlaps
3029 3057 }
3030 3058 } else {
3031 3059 if (size_in_bytes != 0 &&
3032 3060 st->as_Store()->memory_size() != size_in_bytes) {
3033 3061 return FAIL; // mismatched store size
3034 3062 }
3035 3063 return i;
3036 3064 }
3037 3065
3038 3066 ++i;
3039 3067 }
3040 3068 }
3041 3069
3042 3070 // Look for a captured store which initializes at the offset 'start'
3043 3071 // with the given size. If there is no such store, and no other
3044 3072 // initialization interferes, then return zero_memory (the memory
3045 3073 // projection of the AllocateNode).
3046 3074 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
3047 3075 PhaseTransform* phase) {
3048 3076 assert(stores_are_sane(phase), "");
3049 3077 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3050 3078 if (i == 0) {
3051 3079 return NULL; // something is dead
3052 3080 } else if (i < 0) {
3053 3081 return zero_memory(); // just primordial zero bits here
3054 3082 } else {
3055 3083 Node* st = in(i); // here is the store at this position
3056 3084 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
3057 3085 return st;
3058 3086 }
3059 3087 }
3060 3088
3061 3089 // Create, as a raw pointer, an address within my new object at 'offset'.
3062 3090 Node* InitializeNode::make_raw_address(intptr_t offset,
3063 3091 PhaseTransform* phase) {
3064 3092 Node* addr = in(RawAddress);
3065 3093 if (offset != 0) {
3066 3094 Compile* C = phase->C;
3067 3095 addr = phase->transform( new (C, 4) AddPNode(C->top(), addr,
3068 3096 phase->MakeConX(offset)) );
3069 3097 }
3070 3098 return addr;
3071 3099 }
3072 3100
3073 3101 // Clone the given store, converting it into a raw store
3074 3102 // initializing a field or element of my new object.
3075 3103 // Caller is responsible for retiring the original store,
3076 3104 // with subsume_node or the like.
3077 3105 //
3078 3106 // From the example above InitializeNode::InitializeNode,
3079 3107 // here are the old stores to be captured:
3080 3108 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3081 3109 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3082 3110 //
3083 3111 // Here is the changed code; note the extra edges on init:
3084 3112 // alloc = (Allocate ...)
3085 3113 // rawoop = alloc.RawAddress
3086 3114 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
3087 3115 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
3088 3116 // init = (Initialize alloc.Control alloc.Memory rawoop
3089 3117 // rawstore1 rawstore2)
3090 3118 //
3091 3119 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
3092 3120 PhaseTransform* phase) {
3093 3121 assert(stores_are_sane(phase), "");
3094 3122
3095 3123 if (start < 0) return NULL;
3096 3124 assert(can_capture_store(st, phase) == start, "sanity");
3097 3125
3098 3126 Compile* C = phase->C;
3099 3127 int size_in_bytes = st->memory_size();
3100 3128 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3101 3129 if (i == 0) return NULL; // bail out
3102 3130 Node* prev_mem = NULL; // raw memory for the captured store
3103 3131 if (i > 0) {
3104 3132 prev_mem = in(i); // there is a pre-existing store under this one
3105 3133 set_req(i, C->top()); // temporarily disconnect it
3106 3134 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
3107 3135 } else {
3108 3136 i = -i; // no pre-existing store
3109 3137 prev_mem = zero_memory(); // a slice of the newly allocated object
3110 3138 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
3111 3139 set_req(--i, C->top()); // reuse this edge; it has been folded away
3112 3140 else
3113 3141 ins_req(i, C->top()); // build a new edge
3114 3142 }
3115 3143 Node* new_st = st->clone();
3116 3144 new_st->set_req(MemNode::Control, in(Control));
3117 3145 new_st->set_req(MemNode::Memory, prev_mem);
3118 3146 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
3119 3147 new_st = phase->transform(new_st);
3120 3148
3121 3149 // At this point, new_st might have swallowed a pre-existing store
3122 3150 // at the same offset, or perhaps new_st might have disappeared,
3123 3151 // if it redundantly stored the same value (or zero to fresh memory).
3124 3152
3125 3153 // In any case, wire it in:
3126 3154 set_req(i, new_st);
3127 3155
3128 3156 // The caller may now kill the old guy.
3129 3157 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
3130 3158 assert(check_st == new_st || check_st == NULL, "must be findable");
3131 3159 assert(!is_complete(), "");
3132 3160 return new_st;
3133 3161 }
3134 3162
3135 3163 static bool store_constant(jlong* tiles, int num_tiles,
3136 3164 intptr_t st_off, int st_size,
3137 3165 jlong con) {
3138 3166 if ((st_off & (st_size-1)) != 0)
3139 3167 return false; // strange store offset (assume size==2**N)
3140 3168 address addr = (address)tiles + st_off;
3141 3169 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
3142 3170 switch (st_size) {
3143 3171 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
3144 3172 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
3145 3173 case sizeof(jint): *(jint*) addr = (jint) con; break;
3146 3174 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
3147 3175 default: return false; // strange store size (detect size!=2**N here)
3148 3176 }
3149 3177 return true; // return success to caller
3150 3178 }
3151 3179
3152 3180 // Coalesce subword constants into int constants and possibly
3153 3181 // into long constants. The goal, if the CPU permits,
3154 3182 // is to initialize the object with a small number of 64-bit tiles.
3155 3183 // Also, convert floating-point constants to bit patterns.
3156 3184 // Non-constants are not relevant to this pass.
3157 3185 //
3158 3186 // In terms of the running example on InitializeNode::InitializeNode
3159 3187 // and InitializeNode::capture_store, here is the transformation
3160 3188 // of rawstore1 and rawstore2 into rawstore12:
3161 3189 // alloc = (Allocate ...)
3162 3190 // rawoop = alloc.RawAddress
3163 3191 // tile12 = 0x00010002
3164 3192 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
3165 3193 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
3166 3194 //
3167 3195 void
3168 3196 InitializeNode::coalesce_subword_stores(intptr_t header_size,
3169 3197 Node* size_in_bytes,
3170 3198 PhaseGVN* phase) {
3171 3199 Compile* C = phase->C;
3172 3200
3173 3201 assert(stores_are_sane(phase), "");
3174 3202 // Note: After this pass, they are not completely sane,
3175 3203 // since there may be some overlaps.
3176 3204
3177 3205 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
3178 3206
3179 3207 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3180 3208 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
3181 3209 size_limit = MIN2(size_limit, ti_limit);
3182 3210 size_limit = align_size_up(size_limit, BytesPerLong);
3183 3211 int num_tiles = size_limit / BytesPerLong;
3184 3212
3185 3213 // allocate space for the tile map:
3186 3214 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
3187 3215 jlong tiles_buf[small_len];
3188 3216 Node* nodes_buf[small_len];
3189 3217 jlong inits_buf[small_len];
3190 3218 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
3191 3219 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3192 3220 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
3193 3221 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
3194 3222 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
3195 3223 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3196 3224 // tiles: exact bitwise model of all primitive constants
3197 3225 // nodes: last constant-storing node subsumed into the tiles model
3198 3226 // inits: which bytes (in each tile) are touched by any initializations
3199 3227
3200 3228 //// Pass A: Fill in the tile model with any relevant stores.
3201 3229
3202 3230 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
3203 3231 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
3204 3232 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
3205 3233 Node* zmem = zero_memory(); // initially zero memory state
3206 3234 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3207 3235 Node* st = in(i);
3208 3236 intptr_t st_off = get_store_offset(st, phase);
3209 3237
3210 3238 // Figure out the store's offset and constant value:
3211 3239 if (st_off < header_size) continue; //skip (ignore header)
3212 3240 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
3213 3241 int st_size = st->as_Store()->memory_size();
3214 3242 if (st_off + st_size > size_limit) break;
3215 3243
3216 3244 // Record which bytes are touched, whether by constant or not.
3217 3245 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
3218 3246 continue; // skip (strange store size)
3219 3247
3220 3248 const Type* val = phase->type(st->in(MemNode::ValueIn));
3221 3249 if (!val->singleton()) continue; //skip (non-con store)
3222 3250 BasicType type = val->basic_type();
3223 3251
3224 3252 jlong con = 0;
3225 3253 switch (type) {
3226 3254 case T_INT: con = val->is_int()->get_con(); break;
3227 3255 case T_LONG: con = val->is_long()->get_con(); break;
3228 3256 case T_FLOAT: con = jint_cast(val->getf()); break;
3229 3257 case T_DOUBLE: con = jlong_cast(val->getd()); break;
3230 3258 default: continue; //skip (odd store type)
3231 3259 }
3232 3260
3233 3261 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
3234 3262 st->Opcode() == Op_StoreL) {
3235 3263 continue; // This StoreL is already optimal.
3236 3264 }
3237 3265
3238 3266 // Store down the constant.
3239 3267 store_constant(tiles, num_tiles, st_off, st_size, con);
3240 3268
3241 3269 intptr_t j = st_off >> LogBytesPerLong;
3242 3270
3243 3271 if (type == T_INT && st_size == BytesPerInt
3244 3272 && (st_off & BytesPerInt) == BytesPerInt) {
3245 3273 jlong lcon = tiles[j];
3246 3274 if (!Matcher::isSimpleConstant64(lcon) &&
3247 3275 st->Opcode() == Op_StoreI) {
3248 3276 // This StoreI is already optimal by itself.
3249 3277 jint* intcon = (jint*) &tiles[j];
3250 3278 intcon[1] = 0; // undo the store_constant()
3251 3279
3252 3280 // If the previous store is also optimal by itself, back up and
3253 3281 // undo the action of the previous loop iteration... if we can.
3254 3282 // But if we can't, just let the previous half take care of itself.
3255 3283 st = nodes[j];
3256 3284 st_off -= BytesPerInt;
3257 3285 con = intcon[0];
3258 3286 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
3259 3287 assert(st_off >= header_size, "still ignoring header");
3260 3288 assert(get_store_offset(st, phase) == st_off, "must be");
3261 3289 assert(in(i-1) == zmem, "must be");
3262 3290 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
3263 3291 assert(con == tcon->is_int()->get_con(), "must be");
3264 3292 // Undo the effects of the previous loop trip, which swallowed st:
3265 3293 intcon[0] = 0; // undo store_constant()
3266 3294 set_req(i-1, st); // undo set_req(i, zmem)
3267 3295 nodes[j] = NULL; // undo nodes[j] = st
3268 3296 --old_subword; // undo ++old_subword
3269 3297 }
3270 3298 continue; // This StoreI is already optimal.
3271 3299 }
3272 3300 }
3273 3301
3274 3302 // This store is not needed.
3275 3303 set_req(i, zmem);
3276 3304 nodes[j] = st; // record for the moment
3277 3305 if (st_size < BytesPerLong) // something has changed
3278 3306 ++old_subword; // includes int/float, but who's counting...
3279 3307 else ++old_long;
3280 3308 }
3281 3309
3282 3310 if ((old_subword + old_long) == 0)
3283 3311 return; // nothing more to do
3284 3312
3285 3313 //// Pass B: Convert any non-zero tiles into optimal constant stores.
3286 3314 // Be sure to insert them before overlapping non-constant stores.
3287 3315 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
3288 3316 for (int j = 0; j < num_tiles; j++) {
3289 3317 jlong con = tiles[j];
3290 3318 jlong init = inits[j];
3291 3319 if (con == 0) continue;
3292 3320 jint con0, con1; // split the constant, address-wise
3293 3321 jint init0, init1; // split the init map, address-wise
3294 3322 { union { jlong con; jint intcon[2]; } u;
3295 3323 u.con = con;
3296 3324 con0 = u.intcon[0];
3297 3325 con1 = u.intcon[1];
3298 3326 u.con = init;
3299 3327 init0 = u.intcon[0];
3300 3328 init1 = u.intcon[1];
3301 3329 }
3302 3330
3303 3331 Node* old = nodes[j];
3304 3332 assert(old != NULL, "need the prior store");
3305 3333 intptr_t offset = (j * BytesPerLong);
3306 3334
3307 3335 bool split = !Matcher::isSimpleConstant64(con);
3308 3336
3309 3337 if (offset < header_size) {
3310 3338 assert(offset + BytesPerInt >= header_size, "second int counts");
3311 3339 assert(*(jint*)&tiles[j] == 0, "junk in header");
3312 3340 split = true; // only the second word counts
3313 3341 // Example: int a[] = { 42 ... }
3314 3342 } else if (con0 == 0 && init0 == -1) {
3315 3343 split = true; // first word is covered by full inits
3316 3344 // Example: int a[] = { ... foo(), 42 ... }
3317 3345 } else if (con1 == 0 && init1 == -1) {
3318 3346 split = true; // second word is covered by full inits
3319 3347 // Example: int a[] = { ... 42, foo() ... }
3320 3348 }
3321 3349
3322 3350 // Here's a case where init0 is neither 0 nor -1:
3323 3351 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
3324 3352 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
3325 3353 // In this case the tile is not split; it is (jlong)42.
3326 3354 // The big tile is stored down, and then the foo() value is inserted.
3327 3355 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
3328 3356
3329 3357 Node* ctl = old->in(MemNode::Control);
3330 3358 Node* adr = make_raw_address(offset, phase);
3331 3359 const TypePtr* atp = TypeRawPtr::BOTTOM;
3332 3360
3333 3361 // One or two coalesced stores to plop down.
3334 3362 Node* st[2];
3335 3363 intptr_t off[2];
3336 3364 int nst = 0;
3337 3365 if (!split) {
3338 3366 ++new_long;
3339 3367 off[nst] = offset;
3340 3368 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3341 3369 phase->longcon(con), T_LONG);
3342 3370 } else {
3343 3371 // Omit either if it is a zero.
3344 3372 if (con0 != 0) {
3345 3373 ++new_int;
3346 3374 off[nst] = offset;
3347 3375 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3348 3376 phase->intcon(con0), T_INT);
3349 3377 }
3350 3378 if (con1 != 0) {
3351 3379 ++new_int;
3352 3380 offset += BytesPerInt;
3353 3381 adr = make_raw_address(offset, phase);
3354 3382 off[nst] = offset;
3355 3383 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3356 3384 phase->intcon(con1), T_INT);
3357 3385 }
3358 3386 }
3359 3387
3360 3388 // Insert second store first, then the first before the second.
3361 3389 // Insert each one just before any overlapping non-constant stores.
3362 3390 while (nst > 0) {
3363 3391 Node* st1 = st[--nst];
3364 3392 C->copy_node_notes_to(st1, old);
3365 3393 st1 = phase->transform(st1);
3366 3394 offset = off[nst];
3367 3395 assert(offset >= header_size, "do not smash header");
3368 3396 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
3369 3397 guarantee(ins_idx != 0, "must re-insert constant store");
3370 3398 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
3371 3399 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
3372 3400 set_req(--ins_idx, st1);
3373 3401 else
3374 3402 ins_req(ins_idx, st1);
3375 3403 }
3376 3404 }
3377 3405
3378 3406 if (PrintCompilation && WizardMode)
3379 3407 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
3380 3408 old_subword, old_long, new_int, new_long);
3381 3409 if (C->log() != NULL)
3382 3410 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
3383 3411 old_subword, old_long, new_int, new_long);
3384 3412
3385 3413 // Clean up any remaining occurrences of zmem:
3386 3414 remove_extra_zeroes();
3387 3415 }
3388 3416
3389 3417 // Explore forward from in(start) to find the first fully initialized
3390 3418 // word, and return its offset. Skip groups of subword stores which
3391 3419 // together initialize full words. If in(start) is itself part of a
3392 3420 // fully initialized word, return the offset of in(start). If there
3393 3421 // are no following full-word stores, or if something is fishy, return
3394 3422 // a negative value.
3395 3423 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
3396 3424 int int_map = 0;
3397 3425 intptr_t int_map_off = 0;
3398 3426 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
3399 3427
3400 3428 for (uint i = start, limit = req(); i < limit; i++) {
3401 3429 Node* st = in(i);
3402 3430
3403 3431 intptr_t st_off = get_store_offset(st, phase);
3404 3432 if (st_off < 0) break; // return conservative answer
3405 3433
3406 3434 int st_size = st->as_Store()->memory_size();
3407 3435 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
3408 3436 return st_off; // we found a complete word init
3409 3437 }
3410 3438
3411 3439 // update the map:
3412 3440
3413 3441 intptr_t this_int_off = align_size_down(st_off, BytesPerInt);
3414 3442 if (this_int_off != int_map_off) {
3415 3443 // reset the map:
3416 3444 int_map = 0;
3417 3445 int_map_off = this_int_off;
3418 3446 }
3419 3447
3420 3448 int subword_off = st_off - this_int_off;
3421 3449 int_map |= right_n_bits(st_size) << subword_off;
3422 3450 if ((int_map & FULL_MAP) == FULL_MAP) {
3423 3451 return this_int_off; // we found a complete word init
3424 3452 }
3425 3453
3426 3454 // Did this store hit or cross the word boundary?
3427 3455 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt);
3428 3456 if (next_int_off == this_int_off + BytesPerInt) {
3429 3457 // We passed the current int, without fully initializing it.
3430 3458 int_map_off = next_int_off;
3431 3459 int_map >>= BytesPerInt;
3432 3460 } else if (next_int_off > this_int_off + BytesPerInt) {
3433 3461 // We passed the current and next int.
3434 3462 return this_int_off + BytesPerInt;
3435 3463 }
3436 3464 }
3437 3465
3438 3466 return -1;
3439 3467 }
3440 3468
3441 3469
3442 3470 // Called when the associated AllocateNode is expanded into CFG.
3443 3471 // At this point, we may perform additional optimizations.
3444 3472 // Linearize the stores by ascending offset, to make memory
3445 3473 // activity as coherent as possible.
3446 3474 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
3447 3475 intptr_t header_size,
3448 3476 Node* size_in_bytes,
3449 3477 PhaseGVN* phase) {
3450 3478 assert(!is_complete(), "not already complete");
3451 3479 assert(stores_are_sane(phase), "");
3452 3480 assert(allocation() != NULL, "must be present");
3453 3481
3454 3482 remove_extra_zeroes();
3455 3483
3456 3484 if (ReduceFieldZeroing || ReduceBulkZeroing)
3457 3485 // reduce instruction count for common initialization patterns
3458 3486 coalesce_subword_stores(header_size, size_in_bytes, phase);
3459 3487
3460 3488 Node* zmem = zero_memory(); // initially zero memory state
3461 3489 Node* inits = zmem; // accumulating a linearized chain of inits
3462 3490 #ifdef ASSERT
3463 3491 intptr_t first_offset = allocation()->minimum_header_size();
3464 3492 intptr_t last_init_off = first_offset; // previous init offset
3465 3493 intptr_t last_init_end = first_offset; // previous init offset+size
3466 3494 intptr_t last_tile_end = first_offset; // previous tile offset+size
3467 3495 #endif
3468 3496 intptr_t zeroes_done = header_size;
3469 3497
3470 3498 bool do_zeroing = true; // we might give up if inits are very sparse
3471 3499 int big_init_gaps = 0; // how many large gaps have we seen?
3472 3500
3473 3501 if (ZeroTLAB) do_zeroing = false;
3474 3502 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
3475 3503
3476 3504 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3477 3505 Node* st = in(i);
3478 3506 intptr_t st_off = get_store_offset(st, phase);
3479 3507 if (st_off < 0)
3480 3508 break; // unknown junk in the inits
3481 3509 if (st->in(MemNode::Memory) != zmem)
3482 3510 break; // complicated store chains somehow in list
3483 3511
3484 3512 int st_size = st->as_Store()->memory_size();
3485 3513 intptr_t next_init_off = st_off + st_size;
3486 3514
3487 3515 if (do_zeroing && zeroes_done < next_init_off) {
3488 3516 // See if this store needs a zero before it or under it.
3489 3517 intptr_t zeroes_needed = st_off;
3490 3518
3491 3519 if (st_size < BytesPerInt) {
3492 3520 // Look for subword stores which only partially initialize words.
3493 3521 // If we find some, we must lay down some word-level zeroes first,
3494 3522 // underneath the subword stores.
3495 3523 //
3496 3524 // Examples:
3497 3525 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
3498 3526 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
3499 3527 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
3500 3528 //
3501 3529 // Note: coalesce_subword_stores may have already done this,
3502 3530 // if it was prompted by constant non-zero subword initializers.
3503 3531 // But this case can still arise with non-constant stores.
3504 3532
3505 3533 intptr_t next_full_store = find_next_fullword_store(i, phase);
3506 3534
3507 3535 // In the examples above:
3508 3536 // in(i) p q r s x y z
3509 3537 // st_off 12 13 14 15 12 13 14
3510 3538 // st_size 1 1 1 1 1 1 1
3511 3539 // next_full_s. 12 16 16 16 16 16 16
3512 3540 // z's_done 12 16 16 16 12 16 12
3513 3541 // z's_needed 12 16 16 16 16 16 16
3514 3542 // zsize 0 0 0 0 4 0 4
3515 3543 if (next_full_store < 0) {
3516 3544 // Conservative tack: Zero to end of current word.
3517 3545 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt);
3518 3546 } else {
3519 3547 // Zero to beginning of next fully initialized word.
3520 3548 // Or, don't zero at all, if we are already in that word.
3521 3549 assert(next_full_store >= zeroes_needed, "must go forward");
3522 3550 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
3523 3551 zeroes_needed = next_full_store;
3524 3552 }
3525 3553 }
3526 3554
3527 3555 if (zeroes_needed > zeroes_done) {
3528 3556 intptr_t zsize = zeroes_needed - zeroes_done;
3529 3557 // Do some incremental zeroing on rawmem, in parallel with inits.
3530 3558 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3531 3559 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3532 3560 zeroes_done, zeroes_needed,
3533 3561 phase);
3534 3562 zeroes_done = zeroes_needed;
3535 3563 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2)
3536 3564 do_zeroing = false; // leave the hole, next time
3537 3565 }
3538 3566 }
3539 3567
3540 3568 // Collect the store and move on:
3541 3569 st->set_req(MemNode::Memory, inits);
3542 3570 inits = st; // put it on the linearized chain
3543 3571 set_req(i, zmem); // unhook from previous position
3544 3572
3545 3573 if (zeroes_done == st_off)
3546 3574 zeroes_done = next_init_off;
3547 3575
3548 3576 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
3549 3577
3550 3578 #ifdef ASSERT
3551 3579 // Various order invariants. Weaker than stores_are_sane because
3552 3580 // a large constant tile can be filled in by smaller non-constant stores.
3553 3581 assert(st_off >= last_init_off, "inits do not reverse");
3554 3582 last_init_off = st_off;
3555 3583 const Type* val = NULL;
3556 3584 if (st_size >= BytesPerInt &&
3557 3585 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
3558 3586 (int)val->basic_type() < (int)T_OBJECT) {
3559 3587 assert(st_off >= last_tile_end, "tiles do not overlap");
3560 3588 assert(st_off >= last_init_end, "tiles do not overwrite inits");
3561 3589 last_tile_end = MAX2(last_tile_end, next_init_off);
3562 3590 } else {
3563 3591 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong);
3564 3592 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
3565 3593 assert(st_off >= last_init_end, "inits do not overlap");
3566 3594 last_init_end = next_init_off; // it's a non-tile
3567 3595 }
3568 3596 #endif //ASSERT
3569 3597 }
3570 3598
3571 3599 remove_extra_zeroes(); // clear out all the zmems left over
3572 3600 add_req(inits);
3573 3601
3574 3602 if (!ZeroTLAB) {
3575 3603 // If anything remains to be zeroed, zero it all now.
3576 3604 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3577 3605 // if it is the last unused 4 bytes of an instance, forget about it
3578 3606 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
3579 3607 if (zeroes_done + BytesPerLong >= size_limit) {
3580 3608 assert(allocation() != NULL, "");
3581 3609 if (allocation()->Opcode() == Op_Allocate) {
3582 3610 Node* klass_node = allocation()->in(AllocateNode::KlassNode);
3583 3611 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
3584 3612 if (zeroes_done == k->layout_helper())
3585 3613 zeroes_done = size_limit;
3586 3614 }
3587 3615 }
3588 3616 if (zeroes_done < size_limit) {
3589 3617 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3590 3618 zeroes_done, size_in_bytes, phase);
3591 3619 }
3592 3620 }
3593 3621
3594 3622 set_complete(phase);
3595 3623 return rawmem;
3596 3624 }
3597 3625
3598 3626
3599 3627 #ifdef ASSERT
3600 3628 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
3601 3629 if (is_complete())
3602 3630 return true; // stores could be anything at this point
3603 3631 assert(allocation() != NULL, "must be present");
3604 3632 intptr_t last_off = allocation()->minimum_header_size();
3605 3633 for (uint i = InitializeNode::RawStores; i < req(); i++) {
3606 3634 Node* st = in(i);
3607 3635 intptr_t st_off = get_store_offset(st, phase);
3608 3636 if (st_off < 0) continue; // ignore dead garbage
3609 3637 if (last_off > st_off) {
3610 3638 tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off);
3611 3639 this->dump(2);
3612 3640 assert(false, "ascending store offsets");
3613 3641 return false;
3614 3642 }
3615 3643 last_off = st_off + st->as_Store()->memory_size();
3616 3644 }
3617 3645 return true;
3618 3646 }
3619 3647 #endif //ASSERT
3620 3648
3621 3649
3622 3650
3623 3651
3624 3652 //============================MergeMemNode=====================================
3625 3653 //
3626 3654 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
3627 3655 // contributing store or call operations. Each contributor provides the memory
3628 3656 // state for a particular "alias type" (see Compile::alias_type). For example,
3629 3657 // if a MergeMem has an input X for alias category #6, then any memory reference
3630 3658 // to alias category #6 may use X as its memory state input, as an exact equivalent
3631 3659 // to using the MergeMem as a whole.
3632 3660 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
3633 3661 //
3634 3662 // (Here, the <N> notation gives the index of the relevant adr_type.)
3635 3663 //
3636 3664 // In one special case (and more cases in the future), alias categories overlap.
3637 3665 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
3638 3666 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
3639 3667 // it is exactly equivalent to that state W:
3640 3668 // MergeMem(<Bot>: W) <==> W
3641 3669 //
3642 3670 // Usually, the merge has more than one input. In that case, where inputs
3643 3671 // overlap (i.e., one is Bot), the narrower alias type determines the memory
3644 3672 // state for that type, and the wider alias type (Bot) fills in everywhere else:
3645 3673 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
3646 3674 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
3647 3675 //
3648 3676 // A merge can take a "wide" memory state as one of its narrow inputs.
3649 3677 // This simply means that the merge observes out only the relevant parts of
3650 3678 // the wide input. That is, wide memory states arriving at narrow merge inputs
3651 3679 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
3652 3680 //
3653 3681 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
3654 3682 // and that memory slices "leak through":
3655 3683 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
3656 3684 //
3657 3685 // But, in such a cascade, repeated memory slices can "block the leak":
3658 3686 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
3659 3687 //
3660 3688 // In the last example, Y is not part of the combined memory state of the
3661 3689 // outermost MergeMem. The system must, of course, prevent unschedulable
3662 3690 // memory states from arising, so you can be sure that the state Y is somehow
3663 3691 // a precursor to state Y'.
3664 3692 //
3665 3693 //
3666 3694 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
3667 3695 // of each MergeMemNode array are exactly the numerical alias indexes, including
3668 3696 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
3669 3697 // Compile::alias_type (and kin) produce and manage these indexes.
3670 3698 //
3671 3699 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
3672 3700 // (Note that this provides quick access to the top node inside MergeMem methods,
3673 3701 // without the need to reach out via TLS to Compile::current.)
3674 3702 //
3675 3703 // As a consequence of what was just described, a MergeMem that represents a full
3676 3704 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
3677 3705 // containing all alias categories.
3678 3706 //
3679 3707 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
3680 3708 //
3681 3709 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
3682 3710 // a memory state for the alias type <N>, or else the top node, meaning that
3683 3711 // there is no particular input for that alias type. Note that the length of
3684 3712 // a MergeMem is variable, and may be extended at any time to accommodate new
3685 3713 // memory states at larger alias indexes. When merges grow, they are of course
3686 3714 // filled with "top" in the unused in() positions.
3687 3715 //
3688 3716 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
3689 3717 // (Top was chosen because it works smoothly with passes like GCM.)
3690 3718 //
3691 3719 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
3692 3720 // the type of random VM bits like TLS references.) Since it is always the
3693 3721 // first non-Bot memory slice, some low-level loops use it to initialize an
3694 3722 // index variable: for (i = AliasIdxRaw; i < req(); i++).
3695 3723 //
3696 3724 //
3697 3725 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
3698 3726 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
3699 3727 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
3700 3728 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
3701 3729 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
3702 3730 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
3703 3731 //
3704 3732 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
3705 3733 // really that different from the other memory inputs. An abbreviation called
3706 3734 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
3707 3735 //
3708 3736 //
3709 3737 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
3710 3738 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
3711 3739 // that "emerges though" the base memory will be marked as excluding the alias types
3712 3740 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
3713 3741 //
3714 3742 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
3715 3743 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
3716 3744 //
3717 3745 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
3718 3746 // (It is currently unimplemented.) As you can see, the resulting merge is
3719 3747 // actually a disjoint union of memory states, rather than an overlay.
3720 3748 //
3721 3749
3722 3750 //------------------------------MergeMemNode-----------------------------------
3723 3751 Node* MergeMemNode::make_empty_memory() {
3724 3752 Node* empty_memory = (Node*) Compile::current()->top();
3725 3753 assert(empty_memory->is_top(), "correct sentinel identity");
3726 3754 return empty_memory;
3727 3755 }
3728 3756
3729 3757 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
3730 3758 init_class_id(Class_MergeMem);
3731 3759 // all inputs are nullified in Node::Node(int)
3732 3760 // set_input(0, NULL); // no control input
3733 3761
3734 3762 // Initialize the edges uniformly to top, for starters.
3735 3763 Node* empty_mem = make_empty_memory();
3736 3764 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
3737 3765 init_req(i,empty_mem);
3738 3766 }
3739 3767 assert(empty_memory() == empty_mem, "");
3740 3768
3741 3769 if( new_base != NULL && new_base->is_MergeMem() ) {
3742 3770 MergeMemNode* mdef = new_base->as_MergeMem();
3743 3771 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
3744 3772 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
3745 3773 mms.set_memory(mms.memory2());
3746 3774 }
3747 3775 assert(base_memory() == mdef->base_memory(), "");
3748 3776 } else {
3749 3777 set_base_memory(new_base);
3750 3778 }
3751 3779 }
3752 3780
3753 3781 // Make a new, untransformed MergeMem with the same base as 'mem'.
3754 3782 // If mem is itself a MergeMem, populate the result with the same edges.
3755 3783 MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) {
3756 3784 return new(C, 1+Compile::AliasIdxRaw) MergeMemNode(mem);
3757 3785 }
3758 3786
3759 3787 //------------------------------cmp--------------------------------------------
3760 3788 uint MergeMemNode::hash() const { return NO_HASH; }
3761 3789 uint MergeMemNode::cmp( const Node &n ) const {
3762 3790 return (&n == this); // Always fail except on self
3763 3791 }
3764 3792
3765 3793 //------------------------------Identity---------------------------------------
3766 3794 Node* MergeMemNode::Identity(PhaseTransform *phase) {
3767 3795 // Identity if this merge point does not record any interesting memory
3768 3796 // disambiguations.
3769 3797 Node* base_mem = base_memory();
3770 3798 Node* empty_mem = empty_memory();
3771 3799 if (base_mem != empty_mem) { // Memory path is not dead?
3772 3800 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3773 3801 Node* mem = in(i);
3774 3802 if (mem != empty_mem && mem != base_mem) {
3775 3803 return this; // Many memory splits; no change
3776 3804 }
3777 3805 }
3778 3806 }
3779 3807 return base_mem; // No memory splits; ID on the one true input
3780 3808 }
3781 3809
3782 3810 //------------------------------Ideal------------------------------------------
3783 3811 // This method is invoked recursively on chains of MergeMem nodes
3784 3812 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3785 3813 // Remove chain'd MergeMems
3786 3814 //
3787 3815 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
3788 3816 // relative to the "in(Bot)". Since we are patching both at the same time,
3789 3817 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
3790 3818 // but rewrite each "in(i)" relative to the new "in(Bot)".
3791 3819 Node *progress = NULL;
3792 3820
3793 3821
3794 3822 Node* old_base = base_memory();
3795 3823 Node* empty_mem = empty_memory();
3796 3824 if (old_base == empty_mem)
3797 3825 return NULL; // Dead memory path.
3798 3826
3799 3827 MergeMemNode* old_mbase;
3800 3828 if (old_base != NULL && old_base->is_MergeMem())
3801 3829 old_mbase = old_base->as_MergeMem();
3802 3830 else
3803 3831 old_mbase = NULL;
3804 3832 Node* new_base = old_base;
3805 3833
3806 3834 // simplify stacked MergeMems in base memory
3807 3835 if (old_mbase) new_base = old_mbase->base_memory();
3808 3836
3809 3837 // the base memory might contribute new slices beyond my req()
3810 3838 if (old_mbase) grow_to_match(old_mbase);
3811 3839
3812 3840 // Look carefully at the base node if it is a phi.
3813 3841 PhiNode* phi_base;
3814 3842 if (new_base != NULL && new_base->is_Phi())
3815 3843 phi_base = new_base->as_Phi();
3816 3844 else
3817 3845 phi_base = NULL;
3818 3846
3819 3847 Node* phi_reg = NULL;
3820 3848 uint phi_len = (uint)-1;
3821 3849 if (phi_base != NULL && !phi_base->is_copy()) {
3822 3850 // do not examine phi if degraded to a copy
3823 3851 phi_reg = phi_base->region();
3824 3852 phi_len = phi_base->req();
3825 3853 // see if the phi is unfinished
3826 3854 for (uint i = 1; i < phi_len; i++) {
3827 3855 if (phi_base->in(i) == NULL) {
3828 3856 // incomplete phi; do not look at it yet!
3829 3857 phi_reg = NULL;
3830 3858 phi_len = (uint)-1;
3831 3859 break;
3832 3860 }
3833 3861 }
3834 3862 }
3835 3863
3836 3864 // Note: We do not call verify_sparse on entry, because inputs
3837 3865 // can normalize to the base_memory via subsume_node or similar
3838 3866 // mechanisms. This method repairs that damage.
3839 3867
3840 3868 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
3841 3869
3842 3870 // Look at each slice.
3843 3871 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3844 3872 Node* old_in = in(i);
3845 3873 // calculate the old memory value
3846 3874 Node* old_mem = old_in;
3847 3875 if (old_mem == empty_mem) old_mem = old_base;
3848 3876 assert(old_mem == memory_at(i), "");
3849 3877
3850 3878 // maybe update (reslice) the old memory value
3851 3879
3852 3880 // simplify stacked MergeMems
3853 3881 Node* new_mem = old_mem;
3854 3882 MergeMemNode* old_mmem;
3855 3883 if (old_mem != NULL && old_mem->is_MergeMem())
3856 3884 old_mmem = old_mem->as_MergeMem();
3857 3885 else
3858 3886 old_mmem = NULL;
3859 3887 if (old_mmem == this) {
3860 3888 // This can happen if loops break up and safepoints disappear.
3861 3889 // A merge of BotPtr (default) with a RawPtr memory derived from a
3862 3890 // safepoint can be rewritten to a merge of the same BotPtr with
3863 3891 // the BotPtr phi coming into the loop. If that phi disappears
3864 3892 // also, we can end up with a self-loop of the mergemem.
3865 3893 // In general, if loops degenerate and memory effects disappear,
3866 3894 // a mergemem can be left looking at itself. This simply means
3867 3895 // that the mergemem's default should be used, since there is
3868 3896 // no longer any apparent effect on this slice.
3869 3897 // Note: If a memory slice is a MergeMem cycle, it is unreachable
3870 3898 // from start. Update the input to TOP.
3871 3899 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
3872 3900 }
3873 3901 else if (old_mmem != NULL) {
3874 3902 new_mem = old_mmem->memory_at(i);
3875 3903 }
3876 3904 // else preceding memory was not a MergeMem
3877 3905
3878 3906 // replace equivalent phis (unfortunately, they do not GVN together)
3879 3907 if (new_mem != NULL && new_mem != new_base &&
3880 3908 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
3881 3909 if (new_mem->is_Phi()) {
3882 3910 PhiNode* phi_mem = new_mem->as_Phi();
3883 3911 for (uint i = 1; i < phi_len; i++) {
3884 3912 if (phi_base->in(i) != phi_mem->in(i)) {
3885 3913 phi_mem = NULL;
3886 3914 break;
3887 3915 }
3888 3916 }
3889 3917 if (phi_mem != NULL) {
3890 3918 // equivalent phi nodes; revert to the def
3891 3919 new_mem = new_base;
3892 3920 }
3893 3921 }
3894 3922 }
3895 3923
3896 3924 // maybe store down a new value
3897 3925 Node* new_in = new_mem;
3898 3926 if (new_in == new_base) new_in = empty_mem;
3899 3927
3900 3928 if (new_in != old_in) {
3901 3929 // Warning: Do not combine this "if" with the previous "if"
3902 3930 // A memory slice might have be be rewritten even if it is semantically
3903 3931 // unchanged, if the base_memory value has changed.
3904 3932 set_req(i, new_in);
3905 3933 progress = this; // Report progress
3906 3934 }
3907 3935 }
3908 3936
3909 3937 if (new_base != old_base) {
3910 3938 set_req(Compile::AliasIdxBot, new_base);
3911 3939 // Don't use set_base_memory(new_base), because we need to update du.
3912 3940 assert(base_memory() == new_base, "");
3913 3941 progress = this;
3914 3942 }
3915 3943
3916 3944 if( base_memory() == this ) {
3917 3945 // a self cycle indicates this memory path is dead
3918 3946 set_req(Compile::AliasIdxBot, empty_mem);
3919 3947 }
3920 3948
3921 3949 // Resolve external cycles by calling Ideal on a MergeMem base_memory
3922 3950 // Recursion must occur after the self cycle check above
3923 3951 if( base_memory()->is_MergeMem() ) {
3924 3952 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
3925 3953 Node *m = phase->transform(new_mbase); // Rollup any cycles
3926 3954 if( m != NULL && (m->is_top() ||
3927 3955 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) {
3928 3956 // propagate rollup of dead cycle to self
3929 3957 set_req(Compile::AliasIdxBot, empty_mem);
3930 3958 }
3931 3959 }
3932 3960
3933 3961 if( base_memory() == empty_mem ) {
3934 3962 progress = this;
3935 3963 // Cut inputs during Parse phase only.
3936 3964 // During Optimize phase a dead MergeMem node will be subsumed by Top.
3937 3965 if( !can_reshape ) {
3938 3966 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3939 3967 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
3940 3968 }
3941 3969 }
3942 3970 }
3943 3971
3944 3972 if( !progress && base_memory()->is_Phi() && can_reshape ) {
3945 3973 // Check if PhiNode::Ideal's "Split phis through memory merges"
3946 3974 // transform should be attempted. Look for this->phi->this cycle.
3947 3975 uint merge_width = req();
3948 3976 if (merge_width > Compile::AliasIdxRaw) {
3949 3977 PhiNode* phi = base_memory()->as_Phi();
3950 3978 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
3951 3979 if (phi->in(i) == this) {
3952 3980 phase->is_IterGVN()->_worklist.push(phi);
3953 3981 break;
3954 3982 }
3955 3983 }
3956 3984 }
3957 3985 }
3958 3986
3959 3987 assert(progress || verify_sparse(), "please, no dups of base");
3960 3988 return progress;
3961 3989 }
3962 3990
3963 3991 //-------------------------set_base_memory-------------------------------------
3964 3992 void MergeMemNode::set_base_memory(Node *new_base) {
3965 3993 Node* empty_mem = empty_memory();
3966 3994 set_req(Compile::AliasIdxBot, new_base);
3967 3995 assert(memory_at(req()) == new_base, "must set default memory");
3968 3996 // Clear out other occurrences of new_base:
3969 3997 if (new_base != empty_mem) {
3970 3998 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3971 3999 if (in(i) == new_base) set_req(i, empty_mem);
3972 4000 }
3973 4001 }
3974 4002 }
3975 4003
3976 4004 //------------------------------out_RegMask------------------------------------
3977 4005 const RegMask &MergeMemNode::out_RegMask() const {
3978 4006 return RegMask::Empty;
3979 4007 }
3980 4008
3981 4009 //------------------------------dump_spec--------------------------------------
3982 4010 #ifndef PRODUCT
3983 4011 void MergeMemNode::dump_spec(outputStream *st) const {
3984 4012 st->print(" {");
3985 4013 Node* base_mem = base_memory();
3986 4014 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
3987 4015 Node* mem = memory_at(i);
3988 4016 if (mem == base_mem) { st->print(" -"); continue; }
3989 4017 st->print( " N%d:", mem->_idx );
3990 4018 Compile::current()->get_adr_type(i)->dump_on(st);
3991 4019 }
3992 4020 st->print(" }");
3993 4021 }
3994 4022 #endif // !PRODUCT
3995 4023
3996 4024
3997 4025 #ifdef ASSERT
3998 4026 static bool might_be_same(Node* a, Node* b) {
3999 4027 if (a == b) return true;
4000 4028 if (!(a->is_Phi() || b->is_Phi())) return false;
4001 4029 // phis shift around during optimization
4002 4030 return true; // pretty stupid...
4003 4031 }
4004 4032
4005 4033 // verify a narrow slice (either incoming or outgoing)
4006 4034 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
4007 4035 if (!VerifyAliases) return; // don't bother to verify unless requested
4008 4036 if (is_error_reported()) return; // muzzle asserts when debugging an error
4009 4037 if (Node::in_dump()) return; // muzzle asserts when printing
4010 4038 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
4011 4039 assert(n != NULL, "");
4012 4040 // Elide intervening MergeMem's
4013 4041 while (n->is_MergeMem()) {
4014 4042 n = n->as_MergeMem()->memory_at(alias_idx);
4015 4043 }
4016 4044 Compile* C = Compile::current();
4017 4045 const TypePtr* n_adr_type = n->adr_type();
4018 4046 if (n == m->empty_memory()) {
4019 4047 // Implicit copy of base_memory()
4020 4048 } else if (n_adr_type != TypePtr::BOTTOM) {
4021 4049 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
4022 4050 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
4023 4051 } else {
4024 4052 // A few places like make_runtime_call "know" that VM calls are narrow,
4025 4053 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
4026 4054 bool expected_wide_mem = false;
4027 4055 if (n == m->base_memory()) {
4028 4056 expected_wide_mem = true;
4029 4057 } else if (alias_idx == Compile::AliasIdxRaw ||
4030 4058 n == m->memory_at(Compile::AliasIdxRaw)) {
4031 4059 expected_wide_mem = true;
4032 4060 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
4033 4061 // memory can "leak through" calls on channels that
4034 4062 // are write-once. Allow this also.
4035 4063 expected_wide_mem = true;
4036 4064 }
4037 4065 assert(expected_wide_mem, "expected narrow slice replacement");
4038 4066 }
4039 4067 }
4040 4068 #else // !ASSERT
4041 4069 #define verify_memory_slice(m,i,n) (0) // PRODUCT version is no-op
4042 4070 #endif
4043 4071
4044 4072
4045 4073 //-----------------------------memory_at---------------------------------------
4046 4074 Node* MergeMemNode::memory_at(uint alias_idx) const {
4047 4075 assert(alias_idx >= Compile::AliasIdxRaw ||
4048 4076 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
4049 4077 "must avoid base_memory and AliasIdxTop");
4050 4078
4051 4079 // Otherwise, it is a narrow slice.
4052 4080 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
4053 4081 Compile *C = Compile::current();
4054 4082 if (is_empty_memory(n)) {
4055 4083 // the array is sparse; empty slots are the "top" node
4056 4084 n = base_memory();
4057 4085 assert(Node::in_dump()
4058 4086 || n == NULL || n->bottom_type() == Type::TOP
4059 4087 || n->adr_type() == NULL // address is TOP
4060 4088 || n->adr_type() == TypePtr::BOTTOM
4061 4089 || n->adr_type() == TypeRawPtr::BOTTOM
4062 4090 || Compile::current()->AliasLevel() == 0,
4063 4091 "must be a wide memory");
4064 4092 // AliasLevel == 0 if we are organizing the memory states manually.
4065 4093 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
4066 4094 } else {
4067 4095 // make sure the stored slice is sane
4068 4096 #ifdef ASSERT
4069 4097 if (is_error_reported() || Node::in_dump()) {
4070 4098 } else if (might_be_same(n, base_memory())) {
4071 4099 // Give it a pass: It is a mostly harmless repetition of the base.
4072 4100 // This can arise normally from node subsumption during optimization.
4073 4101 } else {
4074 4102 verify_memory_slice(this, alias_idx, n);
4075 4103 }
4076 4104 #endif
4077 4105 }
4078 4106 return n;
4079 4107 }
4080 4108
4081 4109 //---------------------------set_memory_at-------------------------------------
4082 4110 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
4083 4111 verify_memory_slice(this, alias_idx, n);
4084 4112 Node* empty_mem = empty_memory();
4085 4113 if (n == base_memory()) n = empty_mem; // collapse default
4086 4114 uint need_req = alias_idx+1;
4087 4115 if (req() < need_req) {
4088 4116 if (n == empty_mem) return; // already the default, so do not grow me
4089 4117 // grow the sparse array
4090 4118 do {
4091 4119 add_req(empty_mem);
4092 4120 } while (req() < need_req);
4093 4121 }
4094 4122 set_req( alias_idx, n );
4095 4123 }
4096 4124
4097 4125
4098 4126
4099 4127 //--------------------------iteration_setup------------------------------------
4100 4128 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
4101 4129 if (other != NULL) {
4102 4130 grow_to_match(other);
4103 4131 // invariant: the finite support of mm2 is within mm->req()
4104 4132 #ifdef ASSERT
4105 4133 for (uint i = req(); i < other->req(); i++) {
4106 4134 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
4107 4135 }
4108 4136 #endif
4109 4137 }
4110 4138 // Replace spurious copies of base_memory by top.
4111 4139 Node* base_mem = base_memory();
4112 4140 if (base_mem != NULL && !base_mem->is_top()) {
4113 4141 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
4114 4142 if (in(i) == base_mem)
4115 4143 set_req(i, empty_memory());
4116 4144 }
4117 4145 }
4118 4146 }
4119 4147
4120 4148 //---------------------------grow_to_match-------------------------------------
4121 4149 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
4122 4150 Node* empty_mem = empty_memory();
4123 4151 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
4124 4152 // look for the finite support of the other memory
4125 4153 for (uint i = other->req(); --i >= req(); ) {
4126 4154 if (other->in(i) != empty_mem) {
4127 4155 uint new_len = i+1;
4128 4156 while (req() < new_len) add_req(empty_mem);
4129 4157 break;
4130 4158 }
4131 4159 }
4132 4160 }
4133 4161
4134 4162 //---------------------------verify_sparse-------------------------------------
4135 4163 #ifndef PRODUCT
4136 4164 bool MergeMemNode::verify_sparse() const {
4137 4165 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
4138 4166 Node* base_mem = base_memory();
4139 4167 // The following can happen in degenerate cases, since empty==top.
4140 4168 if (is_empty_memory(base_mem)) return true;
4141 4169 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4142 4170 assert(in(i) != NULL, "sane slice");
4143 4171 if (in(i) == base_mem) return false; // should have been the sentinel value!
4144 4172 }
4145 4173 return true;
4146 4174 }
4147 4175
4148 4176 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
4149 4177 Node* n;
4150 4178 n = mm->in(idx);
4151 4179 if (mem == n) return true; // might be empty_memory()
4152 4180 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
4153 4181 if (mem == n) return true;
4154 4182 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) {
4155 4183 if (mem == n) return true;
4156 4184 if (n == NULL) break;
4157 4185 }
4158 4186 return false;
4159 4187 }
4160 4188 #endif // !PRODUCT
↓ open down ↓ |
2571 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX