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