Print this page
rev 6883 : 8057622: java/util/stream/test/org/openjdk/tests/java/util/stream/InfiniteStreamWithLimitOpTest: SEGV inside compiled code (sparc)
Summary: In Parse::array_store_check(), add control edge FROM IfTrue branch of runtime type check of the destination array TO loading _element_klass from destination array.
Reviewed-by: kvn, roland, anoll
Contributed-by: Zoltan Majo <zoltan.majo@oracle.com>
Split |
Split |
Close |
Expand all |
Collapse all |
--- old/hotspot/src/share/vm/opto/memnode.cpp
+++ new/hotspot/src/share/vm/opto/memnode.cpp
1 1 /*
2 2 * Copyright (c) 1997, 2014, 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 TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) {
107 107 assert((t_oop != NULL), "sanity");
108 108 bool is_instance = t_oop->is_known_instance_field();
109 109 bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() &&
110 110 (load != NULL) && load->is_Load() &&
111 111 (phase->is_IterGVN() != NULL);
112 112 if (!(is_instance || is_boxed_value_load))
113 113 return mchain; // don't try to optimize non-instance types
114 114 uint instance_id = t_oop->instance_id();
115 115 Node *start_mem = phase->C->start()->proj_out(TypeFunc::Memory);
116 116 Node *prev = NULL;
117 117 Node *result = mchain;
118 118 while (prev != result) {
119 119 prev = result;
120 120 if (result == start_mem)
121 121 break; // hit one of our sentinels
122 122 // skip over a call which does not affect this memory slice
123 123 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
124 124 Node *proj_in = result->in(0);
125 125 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
126 126 break; // hit one of our sentinels
127 127 } else if (proj_in->is_Call()) {
128 128 CallNode *call = proj_in->as_Call();
129 129 if (!call->may_modify(t_oop, phase)) { // returns false for instances
130 130 result = call->in(TypeFunc::Memory);
131 131 }
132 132 } else if (proj_in->is_Initialize()) {
133 133 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
134 134 // Stop if this is the initialization for the object instance which
135 135 // which contains this memory slice, otherwise skip over it.
136 136 if ((alloc == NULL) || (alloc->_idx == instance_id)) {
137 137 break;
138 138 }
139 139 if (is_instance) {
140 140 result = proj_in->in(TypeFunc::Memory);
141 141 } else if (is_boxed_value_load) {
142 142 Node* klass = alloc->in(AllocateNode::KlassNode);
143 143 const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr();
144 144 if (tklass->klass_is_exact() && !tklass->klass()->equals(t_oop->klass())) {
145 145 result = proj_in->in(TypeFunc::Memory); // not related allocation
146 146 }
147 147 }
148 148 } else if (proj_in->is_MemBar()) {
149 149 result = proj_in->in(TypeFunc::Memory);
150 150 } else {
151 151 assert(false, "unexpected projection");
152 152 }
153 153 } else if (result->is_ClearArray()) {
154 154 if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) {
155 155 // Can not bypass initialization of the instance
156 156 // we are looking for.
157 157 break;
158 158 }
159 159 // Otherwise skip it (the call updated 'result' value).
160 160 } else if (result->is_MergeMem()) {
161 161 result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, NULL, tty);
162 162 }
163 163 }
164 164 return result;
165 165 }
166 166
167 167 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) {
168 168 const TypeOopPtr* t_oop = t_adr->isa_oopptr();
169 169 if (t_oop == NULL)
170 170 return mchain; // don't try to optimize non-oop types
171 171 Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase);
172 172 bool is_instance = t_oop->is_known_instance_field();
173 173 PhaseIterGVN *igvn = phase->is_IterGVN();
174 174 if (is_instance && igvn != NULL && result->is_Phi()) {
175 175 PhiNode *mphi = result->as_Phi();
176 176 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
177 177 const TypePtr *t = mphi->adr_type();
178 178 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ||
179 179 t->isa_oopptr() && !t->is_oopptr()->is_known_instance() &&
180 180 t->is_oopptr()->cast_to_exactness(true)
181 181 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
182 182 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop) {
183 183 // clone the Phi with our address type
184 184 result = mphi->split_out_instance(t_adr, igvn);
185 185 } else {
186 186 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
187 187 }
188 188 }
189 189 return result;
190 190 }
191 191
192 192 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
193 193 uint alias_idx = phase->C->get_alias_index(tp);
194 194 Node *mem = mmem;
195 195 #ifdef ASSERT
196 196 {
197 197 // Check that current type is consistent with the alias index used during graph construction
198 198 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
199 199 bool consistent = adr_check == NULL || adr_check->empty() ||
200 200 phase->C->must_alias(adr_check, alias_idx );
201 201 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
202 202 if( !consistent && adr_check != NULL && !adr_check->empty() &&
203 203 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
204 204 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
205 205 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
206 206 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
207 207 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
208 208 // don't assert if it is dead code.
209 209 consistent = true;
210 210 }
211 211 if( !consistent ) {
212 212 st->print("alias_idx==%d, adr_check==", alias_idx);
213 213 if( adr_check == NULL ) {
214 214 st->print("NULL");
215 215 } else {
216 216 adr_check->dump();
217 217 }
218 218 st->cr();
219 219 print_alias_types();
220 220 assert(consistent, "adr_check must match alias idx");
221 221 }
222 222 }
223 223 #endif
224 224 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
225 225 // means an array I have not precisely typed yet. Do not do any
226 226 // alias stuff with it any time soon.
227 227 const TypeOopPtr *toop = tp->isa_oopptr();
228 228 if( tp->base() != Type::AnyPtr &&
229 229 !(toop &&
230 230 toop->klass() != NULL &&
231 231 toop->klass()->is_java_lang_Object() &&
232 232 toop->offset() == Type::OffsetBot) ) {
233 233 // compress paths and change unreachable cycles to TOP
234 234 // If not, we can update the input infinitely along a MergeMem cycle
235 235 // Equivalent code in PhiNode::Ideal
236 236 Node* m = phase->transform(mmem);
237 237 // If transformed to a MergeMem, get the desired slice
238 238 // Otherwise the returned node represents memory for every slice
239 239 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
240 240 // Update input if it is progress over what we have now
241 241 }
242 242 return mem;
243 243 }
244 244
245 245 //--------------------------Ideal_common---------------------------------------
246 246 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
247 247 // Unhook non-raw memories from complete (macro-expanded) initializations.
248 248 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
249 249 // If our control input is a dead region, kill all below the region
250 250 Node *ctl = in(MemNode::Control);
251 251 if (ctl && remove_dead_region(phase, can_reshape))
252 252 return this;
253 253 ctl = in(MemNode::Control);
254 254 // Don't bother trying to transform a dead node
255 255 if (ctl && ctl->is_top()) return NodeSentinel;
256 256
257 257 PhaseIterGVN *igvn = phase->is_IterGVN();
258 258 // Wait if control on the worklist.
259 259 if (ctl && can_reshape && igvn != NULL) {
260 260 Node* bol = NULL;
261 261 Node* cmp = NULL;
262 262 if (ctl->in(0)->is_If()) {
263 263 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
264 264 bol = ctl->in(0)->in(1);
265 265 if (bol->is_Bool())
266 266 cmp = ctl->in(0)->in(1)->in(1);
267 267 }
268 268 if (igvn->_worklist.member(ctl) ||
269 269 (bol != NULL && igvn->_worklist.member(bol)) ||
270 270 (cmp != NULL && igvn->_worklist.member(cmp)) ) {
271 271 // This control path may be dead.
272 272 // Delay this memory node transformation until the control is processed.
273 273 phase->is_IterGVN()->_worklist.push(this);
274 274 return NodeSentinel; // caller will return NULL
275 275 }
276 276 }
277 277 // Ignore if memory is dead, or self-loop
278 278 Node *mem = in(MemNode::Memory);
279 279 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return NULL
280 280 assert(mem != this, "dead loop in MemNode::Ideal");
281 281
282 282 if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) {
283 283 // This memory slice may be dead.
284 284 // Delay this mem node transformation until the memory is processed.
285 285 phase->is_IterGVN()->_worklist.push(this);
286 286 return NodeSentinel; // caller will return NULL
287 287 }
288 288
289 289 Node *address = in(MemNode::Address);
290 290 const Type *t_adr = phase->type(address);
291 291 if (t_adr == Type::TOP) return NodeSentinel; // caller will return NULL
292 292
293 293 if (can_reshape && igvn != NULL &&
294 294 (igvn->_worklist.member(address) ||
295 295 igvn->_worklist.size() > 0 && (t_adr != adr_type())) ) {
296 296 // The address's base and type may change when the address is processed.
297 297 // Delay this mem node transformation until the address is processed.
298 298 phase->is_IterGVN()->_worklist.push(this);
299 299 return NodeSentinel; // caller will return NULL
300 300 }
301 301
302 302 // Do NOT remove or optimize the next lines: ensure a new alias index
303 303 // is allocated for an oop pointer type before Escape Analysis.
304 304 // Note: C++ will not remove it since the call has side effect.
305 305 if (t_adr->isa_oopptr()) {
306 306 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
307 307 }
308 308
309 309 Node* base = NULL;
310 310 if (address->is_AddP()) {
311 311 base = address->in(AddPNode::Base);
312 312 }
313 313 if (base != NULL && phase->type(base)->higher_equal(TypePtr::NULL_PTR) &&
314 314 !t_adr->isa_rawptr()) {
315 315 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true.
316 316 // Skip this node optimization if its address has TOP base.
317 317 return NodeSentinel; // caller will return NULL
318 318 }
319 319
320 320 // Avoid independent memory operations
321 321 Node* old_mem = mem;
322 322
323 323 // The code which unhooks non-raw memories from complete (macro-expanded)
324 324 // initializations was removed. After macro-expansion all stores catched
325 325 // by Initialize node became raw stores and there is no information
326 326 // which memory slices they modify. So it is unsafe to move any memory
327 327 // operation above these stores. Also in most cases hooked non-raw memories
328 328 // were already unhooked by using information from detect_ptr_independence()
329 329 // and find_previous_store().
330 330
331 331 if (mem->is_MergeMem()) {
332 332 MergeMemNode* mmem = mem->as_MergeMem();
333 333 const TypePtr *tp = t_adr->is_ptr();
334 334
335 335 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
336 336 }
337 337
338 338 if (mem != old_mem) {
339 339 set_req(MemNode::Memory, mem);
340 340 if (can_reshape && old_mem->outcnt() == 0) {
341 341 igvn->_worklist.push(old_mem);
342 342 }
343 343 if (phase->type( mem ) == Type::TOP) return NodeSentinel;
344 344 return this;
345 345 }
346 346
347 347 // let the subclass continue analyzing...
348 348 return NULL;
349 349 }
350 350
351 351 // Helper function for proving some simple control dominations.
352 352 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
353 353 // Already assumes that 'dom' is available at 'sub', and that 'sub'
354 354 // is not a constant (dominated by the method's StartNode).
355 355 // Used by MemNode::find_previous_store to prove that the
356 356 // control input of a memory operation predates (dominates)
357 357 // an allocation it wants to look past.
358 358 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
359 359 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
360 360 return false; // Conservative answer for dead code
361 361
362 362 // Check 'dom'. Skip Proj and CatchProj nodes.
363 363 dom = dom->find_exact_control(dom);
364 364 if (dom == NULL || dom->is_top())
365 365 return false; // Conservative answer for dead code
366 366
367 367 if (dom == sub) {
368 368 // For the case when, for example, 'sub' is Initialize and the original
369 369 // 'dom' is Proj node of the 'sub'.
370 370 return false;
371 371 }
372 372
373 373 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)
374 374 return true;
375 375
376 376 // 'dom' dominates 'sub' if its control edge and control edges
377 377 // of all its inputs dominate or equal to sub's control edge.
378 378
379 379 // Currently 'sub' is either Allocate, Initialize or Start nodes.
380 380 // Or Region for the check in LoadNode::Ideal();
381 381 // 'sub' should have sub->in(0) != NULL.
382 382 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
383 383 sub->is_Region() || sub->is_Call(), "expecting only these nodes");
384 384
385 385 // Get control edge of 'sub'.
386 386 Node* orig_sub = sub;
387 387 sub = sub->find_exact_control(sub->in(0));
388 388 if (sub == NULL || sub->is_top())
389 389 return false; // Conservative answer for dead code
390 390
391 391 assert(sub->is_CFG(), "expecting control");
392 392
393 393 if (sub == dom)
394 394 return true;
395 395
396 396 if (sub->is_Start() || sub->is_Root())
397 397 return false;
398 398
399 399 {
400 400 // Check all control edges of 'dom'.
401 401
402 402 ResourceMark rm;
403 403 Arena* arena = Thread::current()->resource_area();
404 404 Node_List nlist(arena);
405 405 Unique_Node_List dom_list(arena);
406 406
407 407 dom_list.push(dom);
408 408 bool only_dominating_controls = false;
409 409
410 410 for (uint next = 0; next < dom_list.size(); next++) {
411 411 Node* n = dom_list.at(next);
412 412 if (n == orig_sub)
413 413 return false; // One of dom's inputs dominated by sub.
414 414 if (!n->is_CFG() && n->pinned()) {
415 415 // Check only own control edge for pinned non-control nodes.
416 416 n = n->find_exact_control(n->in(0));
417 417 if (n == NULL || n->is_top())
418 418 return false; // Conservative answer for dead code
419 419 assert(n->is_CFG(), "expecting control");
420 420 dom_list.push(n);
421 421 } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
422 422 only_dominating_controls = true;
423 423 } else if (n->is_CFG()) {
424 424 if (n->dominates(sub, nlist))
425 425 only_dominating_controls = true;
426 426 else
427 427 return false;
428 428 } else {
429 429 // First, own control edge.
430 430 Node* m = n->find_exact_control(n->in(0));
431 431 if (m != NULL) {
432 432 if (m->is_top())
433 433 return false; // Conservative answer for dead code
434 434 dom_list.push(m);
435 435 }
436 436 // Now, the rest of edges.
437 437 uint cnt = n->req();
438 438 for (uint i = 1; i < cnt; i++) {
439 439 m = n->find_exact_control(n->in(i));
440 440 if (m == NULL || m->is_top())
441 441 continue;
442 442 dom_list.push(m);
443 443 }
444 444 }
445 445 }
446 446 return only_dominating_controls;
447 447 }
448 448 }
449 449
450 450 //---------------------detect_ptr_independence---------------------------------
451 451 // Used by MemNode::find_previous_store to prove that two base
452 452 // pointers are never equal.
453 453 // The pointers are accompanied by their associated allocations,
454 454 // if any, which have been previously discovered by the caller.
455 455 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
456 456 Node* p2, AllocateNode* a2,
457 457 PhaseTransform* phase) {
458 458 // Attempt to prove that these two pointers cannot be aliased.
459 459 // They may both manifestly be allocations, and they should differ.
460 460 // Or, if they are not both allocations, they can be distinct constants.
461 461 // Otherwise, one is an allocation and the other a pre-existing value.
462 462 if (a1 == NULL && a2 == NULL) { // neither an allocation
463 463 return (p1 != p2) && p1->is_Con() && p2->is_Con();
464 464 } else if (a1 != NULL && a2 != NULL) { // both allocations
465 465 return (a1 != a2);
466 466 } else if (a1 != NULL) { // one allocation a1
467 467 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
468 468 return all_controls_dominate(p2, a1);
469 469 } else { //(a2 != NULL) // one allocation a2
470 470 return all_controls_dominate(p1, a2);
471 471 }
472 472 return false;
473 473 }
474 474
475 475
476 476 // The logic for reordering loads and stores uses four steps:
477 477 // (a) Walk carefully past stores and initializations which we
478 478 // can prove are independent of this load.
479 479 // (b) Observe that the next memory state makes an exact match
480 480 // with self (load or store), and locate the relevant store.
481 481 // (c) Ensure that, if we were to wire self directly to the store,
482 482 // the optimizer would fold it up somehow.
483 483 // (d) Do the rewiring, and return, depending on some other part of
484 484 // the optimizer to fold up the load.
485 485 // This routine handles steps (a) and (b). Steps (c) and (d) are
486 486 // specific to loads and stores, so they are handled by the callers.
487 487 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
488 488 //
489 489 Node* MemNode::find_previous_store(PhaseTransform* phase) {
490 490 Node* ctrl = in(MemNode::Control);
491 491 Node* adr = in(MemNode::Address);
492 492 intptr_t offset = 0;
493 493 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
494 494 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
495 495
496 496 if (offset == Type::OffsetBot)
497 497 return NULL; // cannot unalias unless there are precise offsets
498 498
499 499 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
500 500
501 501 intptr_t size_in_bytes = memory_size();
502 502
503 503 Node* mem = in(MemNode::Memory); // start searching here...
504 504
505 505 int cnt = 50; // Cycle limiter
506 506 for (;;) { // While we can dance past unrelated stores...
507 507 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
508 508
509 509 if (mem->is_Store()) {
510 510 Node* st_adr = mem->in(MemNode::Address);
511 511 intptr_t st_offset = 0;
512 512 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
513 513 if (st_base == NULL)
514 514 break; // inscrutable pointer
515 515 if (st_offset != offset && st_offset != Type::OffsetBot) {
516 516 const int MAX_STORE = BytesPerLong;
517 517 if (st_offset >= offset + size_in_bytes ||
518 518 st_offset <= offset - MAX_STORE ||
519 519 st_offset <= offset - mem->as_Store()->memory_size()) {
520 520 // Success: The offsets are provably independent.
521 521 // (You may ask, why not just test st_offset != offset and be done?
522 522 // The answer is that stores of different sizes can co-exist
523 523 // in the same sequence of RawMem effects. We sometimes initialize
524 524 // a whole 'tile' of array elements with a single jint or jlong.)
525 525 mem = mem->in(MemNode::Memory);
526 526 continue; // (a) advance through independent store memory
527 527 }
528 528 }
529 529 if (st_base != base &&
530 530 detect_ptr_independence(base, alloc,
531 531 st_base,
532 532 AllocateNode::Ideal_allocation(st_base, phase),
533 533 phase)) {
534 534 // Success: The bases are provably independent.
535 535 mem = mem->in(MemNode::Memory);
536 536 continue; // (a) advance through independent store memory
537 537 }
538 538
539 539 // (b) At this point, if the bases or offsets do not agree, we lose,
540 540 // since we have not managed to prove 'this' and 'mem' independent.
541 541 if (st_base == base && st_offset == offset) {
542 542 return mem; // let caller handle steps (c), (d)
543 543 }
544 544
545 545 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
546 546 InitializeNode* st_init = mem->in(0)->as_Initialize();
547 547 AllocateNode* st_alloc = st_init->allocation();
548 548 if (st_alloc == NULL)
549 549 break; // something degenerated
550 550 bool known_identical = false;
551 551 bool known_independent = false;
552 552 if (alloc == st_alloc)
553 553 known_identical = true;
554 554 else if (alloc != NULL)
555 555 known_independent = true;
556 556 else if (all_controls_dominate(this, st_alloc))
557 557 known_independent = true;
558 558
559 559 if (known_independent) {
560 560 // The bases are provably independent: Either they are
561 561 // manifestly distinct allocations, or else the control
562 562 // of this load dominates the store's allocation.
563 563 int alias_idx = phase->C->get_alias_index(adr_type());
564 564 if (alias_idx == Compile::AliasIdxRaw) {
565 565 mem = st_alloc->in(TypeFunc::Memory);
566 566 } else {
567 567 mem = st_init->memory(alias_idx);
568 568 }
569 569 continue; // (a) advance through independent store memory
570 570 }
571 571
572 572 // (b) at this point, if we are not looking at a store initializing
573 573 // the same allocation we are loading from, we lose.
574 574 if (known_identical) {
575 575 // From caller, can_see_stored_value will consult find_captured_store.
576 576 return mem; // let caller handle steps (c), (d)
577 577 }
578 578
579 579 } else if (addr_t != NULL && addr_t->is_known_instance_field()) {
580 580 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
581 581 if (mem->is_Proj() && mem->in(0)->is_Call()) {
582 582 CallNode *call = mem->in(0)->as_Call();
583 583 if (!call->may_modify(addr_t, phase)) {
584 584 mem = call->in(TypeFunc::Memory);
585 585 continue; // (a) advance through independent call memory
586 586 }
587 587 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
588 588 mem = mem->in(0)->in(TypeFunc::Memory);
589 589 continue; // (a) advance through independent MemBar memory
590 590 } else if (mem->is_ClearArray()) {
591 591 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) {
592 592 // (the call updated 'mem' value)
593 593 continue; // (a) advance through independent allocation memory
594 594 } else {
595 595 // Can not bypass initialization of the instance
596 596 // we are looking for.
597 597 return mem;
598 598 }
599 599 } else if (mem->is_MergeMem()) {
600 600 int alias_idx = phase->C->get_alias_index(adr_type());
601 601 mem = mem->as_MergeMem()->memory_at(alias_idx);
602 602 continue; // (a) advance through independent MergeMem memory
603 603 }
604 604 }
605 605
606 606 // Unless there is an explicit 'continue', we must bail out here,
607 607 // because 'mem' is an inscrutable memory state (e.g., a call).
608 608 break;
609 609 }
610 610
611 611 return NULL; // bail out
612 612 }
613 613
614 614 //----------------------calculate_adr_type-------------------------------------
615 615 // Helper function. Notices when the given type of address hits top or bottom.
616 616 // Also, asserts a cross-check of the type against the expected address type.
617 617 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
618 618 if (t == Type::TOP) return NULL; // does not touch memory any more?
619 619 #ifdef PRODUCT
620 620 cross_check = NULL;
621 621 #else
622 622 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL;
623 623 #endif
624 624 const TypePtr* tp = t->isa_ptr();
625 625 if (tp == NULL) {
626 626 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
627 627 return TypePtr::BOTTOM; // touches lots of memory
628 628 } else {
629 629 #ifdef ASSERT
630 630 // %%%% [phh] We don't check the alias index if cross_check is
631 631 // TypeRawPtr::BOTTOM. Needs to be investigated.
632 632 if (cross_check != NULL &&
633 633 cross_check != TypePtr::BOTTOM &&
634 634 cross_check != TypeRawPtr::BOTTOM) {
635 635 // Recheck the alias index, to see if it has changed (due to a bug).
636 636 Compile* C = Compile::current();
637 637 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
638 638 "must stay in the original alias category");
639 639 // The type of the address must be contained in the adr_type,
640 640 // disregarding "null"-ness.
641 641 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
642 642 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
643 643 assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(),
644 644 "real address must not escape from expected memory type");
645 645 }
646 646 #endif
647 647 return tp;
648 648 }
649 649 }
650 650
651 651 //------------------------adr_phi_is_loop_invariant----------------------------
652 652 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted
653 653 // loop is loop invariant. Make a quick traversal of Phi and associated
654 654 // CastPP nodes, looking to see if they are a closed group within the loop.
655 655 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) {
656 656 // The idea is that the phi-nest must boil down to only CastPP nodes
657 657 // with the same data. This implies that any path into the loop already
658 658 // includes such a CastPP, and so the original cast, whatever its input,
659 659 // must be covered by an equivalent cast, with an earlier control input.
660 660 ResourceMark rm;
661 661
662 662 // The loop entry input of the phi should be the unique dominating
663 663 // node for every Phi/CastPP in the loop.
664 664 Unique_Node_List closure;
665 665 closure.push(adr_phi->in(LoopNode::EntryControl));
666 666
667 667 // Add the phi node and the cast to the worklist.
668 668 Unique_Node_List worklist;
669 669 worklist.push(adr_phi);
670 670 if( cast != NULL ){
671 671 if( !cast->is_ConstraintCast() ) return false;
672 672 worklist.push(cast);
673 673 }
674 674
675 675 // Begin recursive walk of phi nodes.
676 676 while( worklist.size() ){
677 677 // Take a node off the worklist
678 678 Node *n = worklist.pop();
679 679 if( !closure.member(n) ){
680 680 // Add it to the closure.
681 681 closure.push(n);
682 682 // Make a sanity check to ensure we don't waste too much time here.
683 683 if( closure.size() > 20) return false;
684 684 // This node is OK if:
685 685 // - it is a cast of an identical value
686 686 // - or it is a phi node (then we add its inputs to the worklist)
687 687 // Otherwise, the node is not OK, and we presume the cast is not invariant
688 688 if( n->is_ConstraintCast() ){
689 689 worklist.push(n->in(1));
690 690 } else if( n->is_Phi() ) {
691 691 for( uint i = 1; i < n->req(); i++ ) {
692 692 worklist.push(n->in(i));
693 693 }
694 694 } else {
695 695 return false;
696 696 }
697 697 }
698 698 }
699 699
700 700 // Quit when the worklist is empty, and we've found no offending nodes.
701 701 return true;
702 702 }
703 703
704 704 //------------------------------Ideal_DU_postCCP-------------------------------
705 705 // Find any cast-away of null-ness and keep its control. Null cast-aways are
706 706 // going away in this pass and we need to make this memory op depend on the
707 707 // gating null check.
708 708 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) {
709 709 return Ideal_common_DU_postCCP(ccp, this, in(MemNode::Address));
710 710 }
711 711
712 712 // I tried to leave the CastPP's in. This makes the graph more accurate in
713 713 // some sense; we get to keep around the knowledge that an oop is not-null
714 714 // after some test. Alas, the CastPP's interfere with GVN (some values are
715 715 // the regular oop, some are the CastPP of the oop, all merge at Phi's which
716 716 // cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed
717 717 // some of the more trivial cases in the optimizer. Removing more useless
718 718 // Phi's started allowing Loads to illegally float above null checks. I gave
719 719 // up on this approach. CNC 10/20/2000
720 720 // This static method may be called not from MemNode (EncodePNode calls it).
721 721 // Only the control edge of the node 'n' might be updated.
722 722 Node *MemNode::Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr ) {
723 723 Node *skipped_cast = NULL;
724 724 // Need a null check? Regular static accesses do not because they are
725 725 // from constant addresses. Array ops are gated by the range check (which
726 726 // always includes a NULL check). Just check field ops.
727 727 if( n->in(MemNode::Control) == NULL ) {
728 728 // Scan upwards for the highest location we can place this memory op.
729 729 while( true ) {
730 730 switch( adr->Opcode() ) {
731 731
732 732 case Op_AddP: // No change to NULL-ness, so peek thru AddP's
733 733 adr = adr->in(AddPNode::Base);
734 734 continue;
735 735
736 736 case Op_DecodeN: // No change to NULL-ness, so peek thru
737 737 case Op_DecodeNKlass:
738 738 adr = adr->in(1);
739 739 continue;
740 740
741 741 case Op_EncodeP:
742 742 case Op_EncodePKlass:
743 743 // EncodeP node's control edge could be set by this method
744 744 // when EncodeP node depends on CastPP node.
745 745 //
746 746 // Use its control edge for memory op because EncodeP may go away
747 747 // later when it is folded with following or preceding DecodeN node.
748 748 if (adr->in(0) == NULL) {
749 749 // Keep looking for cast nodes.
750 750 adr = adr->in(1);
751 751 continue;
752 752 }
753 753 ccp->hash_delete(n);
754 754 n->set_req(MemNode::Control, adr->in(0));
755 755 ccp->hash_insert(n);
756 756 return n;
757 757
758 758 case Op_CastPP:
759 759 // If the CastPP is useless, just peek on through it.
760 760 if( ccp->type(adr) == ccp->type(adr->in(1)) ) {
761 761 // Remember the cast that we've peeked though. If we peek
762 762 // through more than one, then we end up remembering the highest
763 763 // one, that is, if in a loop, the one closest to the top.
764 764 skipped_cast = adr;
765 765 adr = adr->in(1);
766 766 continue;
767 767 }
768 768 // CastPP is going away in this pass! We need this memory op to be
769 769 // control-dependent on the test that is guarding the CastPP.
770 770 ccp->hash_delete(n);
771 771 n->set_req(MemNode::Control, adr->in(0));
772 772 ccp->hash_insert(n);
773 773 return n;
774 774
775 775 case Op_Phi:
776 776 // Attempt to float above a Phi to some dominating point.
777 777 if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) {
778 778 // If we've already peeked through a Cast (which could have set the
779 779 // control), we can't float above a Phi, because the skipped Cast
780 780 // may not be loop invariant.
781 781 if (adr_phi_is_loop_invariant(adr, skipped_cast)) {
782 782 adr = adr->in(1);
783 783 continue;
784 784 }
785 785 }
786 786
787 787 // Intentional fallthrough!
788 788
789 789 // No obvious dominating point. The mem op is pinned below the Phi
790 790 // by the Phi itself. If the Phi goes away (no true value is merged)
791 791 // then the mem op can float, but not indefinitely. It must be pinned
792 792 // behind the controls leading to the Phi.
793 793 case Op_CheckCastPP:
794 794 // These usually stick around to change address type, however a
795 795 // useless one can be elided and we still need to pick up a control edge
796 796 if (adr->in(0) == NULL) {
797 797 // This CheckCastPP node has NO control and is likely useless. But we
798 798 // need check further up the ancestor chain for a control input to keep
799 799 // the node in place. 4959717.
800 800 skipped_cast = adr;
801 801 adr = adr->in(1);
802 802 continue;
803 803 }
804 804 ccp->hash_delete(n);
805 805 n->set_req(MemNode::Control, adr->in(0));
806 806 ccp->hash_insert(n);
807 807 return n;
808 808
809 809 // List of "safe" opcodes; those that implicitly block the memory
810 810 // op below any null check.
811 811 case Op_CastX2P: // no null checks on native pointers
812 812 case Op_Parm: // 'this' pointer is not null
813 813 case Op_LoadP: // Loading from within a klass
814 814 case Op_LoadN: // Loading from within a klass
815 815 case Op_LoadKlass: // Loading from within a klass
816 816 case Op_LoadNKlass: // Loading from within a klass
817 817 case Op_ConP: // Loading from a klass
818 818 case Op_ConN: // Loading from a klass
819 819 case Op_ConNKlass: // Loading from a klass
820 820 case Op_CreateEx: // Sucking up the guts of an exception oop
821 821 case Op_Con: // Reading from TLS
822 822 case Op_CMoveP: // CMoveP is pinned
823 823 case Op_CMoveN: // CMoveN is pinned
824 824 break; // No progress
825 825
826 826 case Op_Proj: // Direct call to an allocation routine
827 827 case Op_SCMemProj: // Memory state from store conditional ops
828 828 #ifdef ASSERT
829 829 {
830 830 assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value");
831 831 const Node* call = adr->in(0);
832 832 if (call->is_CallJava()) {
833 833 const CallJavaNode* call_java = call->as_CallJava();
834 834 const TypeTuple *r = call_java->tf()->range();
835 835 assert(r->cnt() > TypeFunc::Parms, "must return value");
836 836 const Type* ret_type = r->field_at(TypeFunc::Parms);
837 837 assert(ret_type && ret_type->isa_ptr(), "must return pointer");
838 838 // We further presume that this is one of
839 839 // new_instance_Java, new_array_Java, or
840 840 // the like, but do not assert for this.
841 841 } else if (call->is_Allocate()) {
842 842 // similar case to new_instance_Java, etc.
843 843 } else if (!call->is_CallLeaf()) {
844 844 // Projections from fetch_oop (OSR) are allowed as well.
845 845 ShouldNotReachHere();
846 846 }
847 847 }
848 848 #endif
849 849 break;
850 850 default:
851 851 ShouldNotReachHere();
↓ open down ↓ |
851 lines elided |
↑ open up ↑ |
852 852 }
853 853 break;
854 854 }
855 855 }
856 856
857 857 return NULL; // No progress
858 858 }
859 859
860 860
861 861 //=============================================================================
862 +// Should LoadNode::Ideal() attempt to remove control edges?
863 +bool LoadNode::can_remove_control() const {
864 + return true;
865 +}
862 866 uint LoadNode::size_of() const { return sizeof(*this); }
863 867 uint LoadNode::cmp( const Node &n ) const
864 868 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
865 869 const Type *LoadNode::bottom_type() const { return _type; }
866 870 uint LoadNode::ideal_reg() const {
867 871 return _type->ideal_reg();
868 872 }
869 873
870 874 #ifndef PRODUCT
871 875 void LoadNode::dump_spec(outputStream *st) const {
872 876 MemNode::dump_spec(st);
873 877 if( !Verbose && !WizardMode ) {
874 878 // standard dump does this in Verbose and WizardMode
875 879 st->print(" #"); _type->dump_on(st);
876 880 }
877 881 }
878 882 #endif
879 883
880 884 #ifdef ASSERT
881 885 //----------------------------is_immutable_value-------------------------------
882 886 // Helper function to allow a raw load without control edge for some cases
883 887 bool LoadNode::is_immutable_value(Node* adr) {
884 888 return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
885 889 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
886 890 (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) ==
887 891 in_bytes(JavaThread::osthread_offset())));
888 892 }
889 893 #endif
890 894
891 895 //----------------------------LoadNode::make-----------------------------------
892 896 // Polymorphic factory method:
893 897 Node *LoadNode::make(PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt, MemOrd mo) {
894 898 Compile* C = gvn.C;
895 899
896 900 // sanity check the alias category against the created node type
897 901 assert(!(adr_type->isa_oopptr() &&
898 902 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
899 903 "use LoadKlassNode instead");
900 904 assert(!(adr_type->isa_aryptr() &&
901 905 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
902 906 "use LoadRangeNode instead");
903 907 // Check control edge of raw loads
904 908 assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
905 909 // oop will be recorded in oop map if load crosses safepoint
906 910 rt->isa_oopptr() || is_immutable_value(adr),
907 911 "raw memory operations should have control edge");
908 912 switch (bt) {
909 913 case T_BOOLEAN: return new (C) LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo);
910 914 case T_BYTE: return new (C) LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo);
911 915 case T_INT: return new (C) LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo);
912 916 case T_CHAR: return new (C) LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo);
913 917 case T_SHORT: return new (C) LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo);
914 918 case T_LONG: return new (C) LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo);
915 919 case T_FLOAT: return new (C) LoadFNode (ctl, mem, adr, adr_type, rt, mo);
916 920 case T_DOUBLE: return new (C) LoadDNode (ctl, mem, adr, adr_type, rt, mo);
917 921 case T_ADDRESS: return new (C) LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo);
918 922 case T_OBJECT:
919 923 #ifdef _LP64
920 924 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
921 925 Node* load = gvn.transform(new (C) LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo));
922 926 return new (C) DecodeNNode(load, load->bottom_type()->make_ptr());
923 927 } else
924 928 #endif
925 929 {
926 930 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop");
927 931 return new (C) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr(), mo);
928 932 }
929 933 }
930 934 ShouldNotReachHere();
931 935 return (LoadNode*)NULL;
932 936 }
933 937
934 938 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo) {
935 939 bool require_atomic = true;
936 940 return new (C) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), mo, require_atomic);
937 941 }
938 942
939 943
940 944
941 945
942 946 //------------------------------hash-------------------------------------------
943 947 uint LoadNode::hash() const {
944 948 // unroll addition of interesting fields
945 949 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
946 950 }
947 951
948 952 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) {
949 953 if ((atp != NULL) && (atp->index() >= Compile::AliasIdxRaw)) {
950 954 bool non_volatile = (atp->field() != NULL) && !atp->field()->is_volatile();
951 955 bool is_stable_ary = FoldStableValues &&
952 956 (tp != NULL) && (tp->isa_aryptr() != NULL) &&
953 957 tp->isa_aryptr()->is_stable();
954 958
955 959 return (eliminate_boxing && non_volatile) || is_stable_ary;
956 960 }
957 961
958 962 return false;
959 963 }
960 964
961 965 //---------------------------can_see_stored_value------------------------------
962 966 // This routine exists to make sure this set of tests is done the same
963 967 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
964 968 // will change the graph shape in a way which makes memory alive twice at the
965 969 // same time (uses the Oracle model of aliasing), then some
966 970 // LoadXNode::Identity will fold things back to the equivalence-class model
967 971 // of aliasing.
968 972 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
969 973 Node* ld_adr = in(MemNode::Address);
970 974 intptr_t ld_off = 0;
971 975 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
972 976 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
973 977 Compile::AliasType* atp = (tp != NULL) ? phase->C->alias_type(tp) : NULL;
974 978 // This is more general than load from boxing objects.
975 979 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) {
976 980 uint alias_idx = atp->index();
977 981 bool final = !atp->is_rewritable();
978 982 Node* result = NULL;
979 983 Node* current = st;
980 984 // Skip through chains of MemBarNodes checking the MergeMems for
981 985 // new states for the slice of this load. Stop once any other
982 986 // kind of node is encountered. Loads from final memory can skip
983 987 // through any kind of MemBar but normal loads shouldn't skip
984 988 // through MemBarAcquire since the could allow them to move out of
985 989 // a synchronized region.
986 990 while (current->is_Proj()) {
987 991 int opc = current->in(0)->Opcode();
988 992 if ((final && (opc == Op_MemBarAcquire ||
989 993 opc == Op_MemBarAcquireLock ||
990 994 opc == Op_LoadFence)) ||
991 995 opc == Op_MemBarRelease ||
992 996 opc == Op_StoreFence ||
993 997 opc == Op_MemBarReleaseLock ||
994 998 opc == Op_MemBarCPUOrder) {
995 999 Node* mem = current->in(0)->in(TypeFunc::Memory);
996 1000 if (mem->is_MergeMem()) {
997 1001 MergeMemNode* merge = mem->as_MergeMem();
998 1002 Node* new_st = merge->memory_at(alias_idx);
999 1003 if (new_st == merge->base_memory()) {
1000 1004 // Keep searching
1001 1005 current = new_st;
1002 1006 continue;
1003 1007 }
1004 1008 // Save the new memory state for the slice and fall through
1005 1009 // to exit.
1006 1010 result = new_st;
1007 1011 }
1008 1012 }
1009 1013 break;
1010 1014 }
1011 1015 if (result != NULL) {
1012 1016 st = result;
1013 1017 }
1014 1018 }
1015 1019
1016 1020 // Loop around twice in the case Load -> Initialize -> Store.
1017 1021 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
1018 1022 for (int trip = 0; trip <= 1; trip++) {
1019 1023
1020 1024 if (st->is_Store()) {
1021 1025 Node* st_adr = st->in(MemNode::Address);
1022 1026 if (!phase->eqv(st_adr, ld_adr)) {
1023 1027 // Try harder before giving up... Match raw and non-raw pointers.
1024 1028 intptr_t st_off = 0;
1025 1029 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off);
1026 1030 if (alloc == NULL) return NULL;
1027 1031 if (alloc != ld_alloc) return NULL;
1028 1032 if (ld_off != st_off) return NULL;
1029 1033 // At this point we have proven something like this setup:
1030 1034 // A = Allocate(...)
1031 1035 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off))
1032 1036 // S = StoreQ(, AddP(, A.Parm , #Off), V)
1033 1037 // (Actually, we haven't yet proven the Q's are the same.)
1034 1038 // In other words, we are loading from a casted version of
1035 1039 // the same pointer-and-offset that we stored to.
1036 1040 // Thus, we are able to replace L by V.
1037 1041 }
1038 1042 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
1039 1043 if (store_Opcode() != st->Opcode())
1040 1044 return NULL;
1041 1045 return st->in(MemNode::ValueIn);
1042 1046 }
1043 1047
1044 1048 // A load from a freshly-created object always returns zero.
1045 1049 // (This can happen after LoadNode::Ideal resets the load's memory input
1046 1050 // to find_captured_store, which returned InitializeNode::zero_memory.)
1047 1051 if (st->is_Proj() && st->in(0)->is_Allocate() &&
1048 1052 (st->in(0) == ld_alloc) &&
1049 1053 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) {
1050 1054 // return a zero value for the load's basic type
1051 1055 // (This is one of the few places where a generic PhaseTransform
1052 1056 // can create new nodes. Think of it as lazily manifesting
1053 1057 // virtually pre-existing constants.)
1054 1058 return phase->zerocon(memory_type());
1055 1059 }
1056 1060
1057 1061 // A load from an initialization barrier can match a captured store.
1058 1062 if (st->is_Proj() && st->in(0)->is_Initialize()) {
1059 1063 InitializeNode* init = st->in(0)->as_Initialize();
1060 1064 AllocateNode* alloc = init->allocation();
1061 1065 if ((alloc != NULL) && (alloc == ld_alloc)) {
1062 1066 // examine a captured store value
1063 1067 st = init->find_captured_store(ld_off, memory_size(), phase);
1064 1068 if (st != NULL)
1065 1069 continue; // take one more trip around
1066 1070 }
1067 1071 }
1068 1072
1069 1073 // Load boxed value from result of valueOf() call is input parameter.
1070 1074 if (this->is_Load() && ld_adr->is_AddP() &&
1071 1075 (tp != NULL) && tp->is_ptr_to_boxed_value()) {
1072 1076 intptr_t ignore = 0;
1073 1077 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore);
1074 1078 if (base != NULL && base->is_Proj() &&
1075 1079 base->as_Proj()->_con == TypeFunc::Parms &&
1076 1080 base->in(0)->is_CallStaticJava() &&
1077 1081 base->in(0)->as_CallStaticJava()->is_boxing_method()) {
1078 1082 return base->in(0)->in(TypeFunc::Parms);
1079 1083 }
1080 1084 }
1081 1085
1082 1086 break;
1083 1087 }
1084 1088
1085 1089 return NULL;
1086 1090 }
1087 1091
1088 1092 //----------------------is_instance_field_load_with_local_phi------------------
1089 1093 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1090 1094 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl &&
1091 1095 in(Address)->is_AddP() ) {
1092 1096 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr();
1093 1097 // Only instances and boxed values.
1094 1098 if( t_oop != NULL &&
1095 1099 (t_oop->is_ptr_to_boxed_value() ||
1096 1100 t_oop->is_known_instance_field()) &&
1097 1101 t_oop->offset() != Type::OffsetBot &&
1098 1102 t_oop->offset() != Type::OffsetTop) {
1099 1103 return true;
1100 1104 }
1101 1105 }
1102 1106 return false;
1103 1107 }
1104 1108
1105 1109 //------------------------------Identity---------------------------------------
1106 1110 // Loads are identity if previous store is to same address
1107 1111 Node *LoadNode::Identity( PhaseTransform *phase ) {
1108 1112 // If the previous store-maker is the right kind of Store, and the store is
1109 1113 // to the same address, then we are equal to the value stored.
1110 1114 Node* mem = in(Memory);
1111 1115 Node* value = can_see_stored_value(mem, phase);
1112 1116 if( value ) {
1113 1117 // byte, short & char stores truncate naturally.
1114 1118 // A load has to load the truncated value which requires
1115 1119 // some sort of masking operation and that requires an
1116 1120 // Ideal call instead of an Identity call.
1117 1121 if (memory_size() < BytesPerInt) {
1118 1122 // If the input to the store does not fit with the load's result type,
1119 1123 // it must be truncated via an Ideal call.
1120 1124 if (!phase->type(value)->higher_equal(phase->type(this)))
1121 1125 return this;
1122 1126 }
1123 1127 // (This works even when value is a Con, but LoadNode::Value
1124 1128 // usually runs first, producing the singleton type of the Con.)
1125 1129 return value;
1126 1130 }
1127 1131
1128 1132 // Search for an existing data phi which was generated before for the same
1129 1133 // instance's field to avoid infinite generation of phis in a loop.
1130 1134 Node *region = mem->in(0);
1131 1135 if (is_instance_field_load_with_local_phi(region)) {
1132 1136 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr();
1133 1137 int this_index = phase->C->get_alias_index(addr_t);
1134 1138 int this_offset = addr_t->offset();
1135 1139 int this_iid = addr_t->instance_id();
1136 1140 if (!addr_t->is_known_instance() &&
1137 1141 addr_t->is_ptr_to_boxed_value()) {
1138 1142 // Use _idx of address base (could be Phi node) for boxed values.
1139 1143 intptr_t ignore = 0;
1140 1144 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1141 1145 this_iid = base->_idx;
1142 1146 }
1143 1147 const Type* this_type = bottom_type();
1144 1148 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1145 1149 Node* phi = region->fast_out(i);
1146 1150 if (phi->is_Phi() && phi != mem &&
1147 1151 phi->as_Phi()->is_same_inst_field(this_type, this_iid, this_index, this_offset)) {
1148 1152 return phi;
1149 1153 }
1150 1154 }
1151 1155 }
1152 1156
1153 1157 return this;
1154 1158 }
1155 1159
1156 1160 // We're loading from an object which has autobox behaviour.
1157 1161 // If this object is result of a valueOf call we'll have a phi
1158 1162 // merging a newly allocated object and a load from the cache.
1159 1163 // We want to replace this load with the original incoming
1160 1164 // argument to the valueOf call.
1161 1165 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
1162 1166 assert(phase->C->eliminate_boxing(), "sanity");
1163 1167 intptr_t ignore = 0;
1164 1168 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1165 1169 if ((base == NULL) || base->is_Phi()) {
1166 1170 // Push the loads from the phi that comes from valueOf up
1167 1171 // through it to allow elimination of the loads and the recovery
1168 1172 // of the original value. It is done in split_through_phi().
1169 1173 return NULL;
1170 1174 } else if (base->is_Load() ||
1171 1175 base->is_DecodeN() && base->in(1)->is_Load()) {
1172 1176 // Eliminate the load of boxed value for integer types from the cache
1173 1177 // array by deriving the value from the index into the array.
1174 1178 // Capture the offset of the load and then reverse the computation.
1175 1179
1176 1180 // Get LoadN node which loads a boxing object from 'cache' array.
1177 1181 if (base->is_DecodeN()) {
1178 1182 base = base->in(1);
1179 1183 }
1180 1184 if (!base->in(Address)->is_AddP()) {
1181 1185 return NULL; // Complex address
1182 1186 }
1183 1187 AddPNode* address = base->in(Address)->as_AddP();
1184 1188 Node* cache_base = address->in(AddPNode::Base);
1185 1189 if ((cache_base != NULL) && cache_base->is_DecodeN()) {
1186 1190 // Get ConP node which is static 'cache' field.
1187 1191 cache_base = cache_base->in(1);
1188 1192 }
1189 1193 if ((cache_base != NULL) && cache_base->is_Con()) {
1190 1194 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr();
1191 1195 if ((base_type != NULL) && base_type->is_autobox_cache()) {
1192 1196 Node* elements[4];
1193 1197 int shift = exact_log2(type2aelembytes(T_OBJECT));
1194 1198 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements));
1195 1199 if ((count > 0) && elements[0]->is_Con() &&
1196 1200 ((count == 1) ||
1197 1201 (count == 2) && elements[1]->Opcode() == Op_LShiftX &&
1198 1202 elements[1]->in(2) == phase->intcon(shift))) {
1199 1203 ciObjArray* array = base_type->const_oop()->as_obj_array();
1200 1204 // Fetch the box object cache[0] at the base of the array and get its value
1201 1205 ciInstance* box = array->obj_at(0)->as_instance();
1202 1206 ciInstanceKlass* ik = box->klass()->as_instance_klass();
1203 1207 assert(ik->is_box_klass(), "sanity");
1204 1208 assert(ik->nof_nonstatic_fields() == 1, "change following code");
1205 1209 if (ik->nof_nonstatic_fields() == 1) {
1206 1210 // This should be true nonstatic_field_at requires calling
1207 1211 // nof_nonstatic_fields so check it anyway
1208 1212 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1209 1213 BasicType bt = c.basic_type();
1210 1214 // Only integer types have boxing cache.
1211 1215 assert(bt == T_BOOLEAN || bt == T_CHAR ||
1212 1216 bt == T_BYTE || bt == T_SHORT ||
1213 1217 bt == T_INT || bt == T_LONG, err_msg_res("wrong type = %s", type2name(bt)));
1214 1218 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int();
1215 1219 if (cache_low != (int)cache_low) {
1216 1220 return NULL; // should not happen since cache is array indexed by value
1217 1221 }
1218 1222 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift);
1219 1223 if (offset != (int)offset) {
1220 1224 return NULL; // should not happen since cache is array indexed by value
1221 1225 }
1222 1226 // Add up all the offsets making of the address of the load
1223 1227 Node* result = elements[0];
1224 1228 for (int i = 1; i < count; i++) {
1225 1229 result = phase->transform(new (phase->C) AddXNode(result, elements[i]));
1226 1230 }
1227 1231 // Remove the constant offset from the address and then
1228 1232 result = phase->transform(new (phase->C) AddXNode(result, phase->MakeConX(-(int)offset)));
1229 1233 // remove the scaling of the offset to recover the original index.
1230 1234 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
1231 1235 // Peel the shift off directly but wrap it in a dummy node
1232 1236 // since Ideal can't return existing nodes
1233 1237 result = new (phase->C) RShiftXNode(result->in(1), phase->intcon(0));
1234 1238 } else if (result->is_Add() && result->in(2)->is_Con() &&
1235 1239 result->in(1)->Opcode() == Op_LShiftX &&
1236 1240 result->in(1)->in(2) == phase->intcon(shift)) {
1237 1241 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z)
1238 1242 // but for boxing cache access we know that X<<Z will not overflow
1239 1243 // (there is range check) so we do this optimizatrion by hand here.
1240 1244 Node* add_con = new (phase->C) RShiftXNode(result->in(2), phase->intcon(shift));
1241 1245 result = new (phase->C) AddXNode(result->in(1)->in(1), phase->transform(add_con));
1242 1246 } else {
1243 1247 result = new (phase->C) RShiftXNode(result, phase->intcon(shift));
1244 1248 }
1245 1249 #ifdef _LP64
1246 1250 if (bt != T_LONG) {
1247 1251 result = new (phase->C) ConvL2INode(phase->transform(result));
1248 1252 }
1249 1253 #else
1250 1254 if (bt == T_LONG) {
1251 1255 result = new (phase->C) ConvI2LNode(phase->transform(result));
1252 1256 }
1253 1257 #endif
1254 1258 return result;
1255 1259 }
1256 1260 }
1257 1261 }
1258 1262 }
1259 1263 }
1260 1264 return NULL;
1261 1265 }
1262 1266
1263 1267 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) {
1264 1268 Node* region = phi->in(0);
1265 1269 if (region == NULL) {
1266 1270 return false; // Wait stable graph
1267 1271 }
1268 1272 uint cnt = phi->req();
1269 1273 for (uint i = 1; i < cnt; i++) {
1270 1274 Node* rc = region->in(i);
1271 1275 if (rc == NULL || phase->type(rc) == Type::TOP)
1272 1276 return false; // Wait stable graph
1273 1277 Node* in = phi->in(i);
1274 1278 if (in == NULL || phase->type(in) == Type::TOP)
1275 1279 return false; // Wait stable graph
1276 1280 }
1277 1281 return true;
1278 1282 }
1279 1283 //------------------------------split_through_phi------------------------------
1280 1284 // Split instance or boxed field load through Phi.
1281 1285 Node *LoadNode::split_through_phi(PhaseGVN *phase) {
1282 1286 Node* mem = in(Memory);
1283 1287 Node* address = in(Address);
1284 1288 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr();
1285 1289
1286 1290 assert((t_oop != NULL) &&
1287 1291 (t_oop->is_known_instance_field() ||
1288 1292 t_oop->is_ptr_to_boxed_value()), "invalide conditions");
1289 1293
1290 1294 Compile* C = phase->C;
1291 1295 intptr_t ignore = 0;
1292 1296 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1293 1297 bool base_is_phi = (base != NULL) && base->is_Phi();
1294 1298 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() &&
1295 1299 (base != NULL) && (base == address->in(AddPNode::Base)) &&
1296 1300 phase->type(base)->higher_equal(TypePtr::NOTNULL);
1297 1301
1298 1302 if (!((mem->is_Phi() || base_is_phi) &&
1299 1303 (load_boxed_values || t_oop->is_known_instance_field()))) {
1300 1304 return NULL; // memory is not Phi
1301 1305 }
1302 1306
1303 1307 if (mem->is_Phi()) {
1304 1308 if (!stable_phi(mem->as_Phi(), phase)) {
1305 1309 return NULL; // Wait stable graph
1306 1310 }
1307 1311 uint cnt = mem->req();
1308 1312 // Check for loop invariant memory.
1309 1313 if (cnt == 3) {
1310 1314 for (uint i = 1; i < cnt; i++) {
1311 1315 Node* in = mem->in(i);
1312 1316 Node* m = optimize_memory_chain(in, t_oop, this, phase);
1313 1317 if (m == mem) {
1314 1318 set_req(Memory, mem->in(cnt - i));
1315 1319 return this; // made change
1316 1320 }
1317 1321 }
1318 1322 }
1319 1323 }
1320 1324 if (base_is_phi) {
1321 1325 if (!stable_phi(base->as_Phi(), phase)) {
1322 1326 return NULL; // Wait stable graph
1323 1327 }
1324 1328 uint cnt = base->req();
1325 1329 // Check for loop invariant memory.
1326 1330 if (cnt == 3) {
1327 1331 for (uint i = 1; i < cnt; i++) {
1328 1332 if (base->in(i) == base) {
1329 1333 return NULL; // Wait stable graph
1330 1334 }
1331 1335 }
1332 1336 }
1333 1337 }
1334 1338
1335 1339 bool load_boxed_phi = load_boxed_values && base_is_phi && (base->in(0) == mem->in(0));
1336 1340
1337 1341 // Split through Phi (see original code in loopopts.cpp).
1338 1342 assert(C->have_alias_type(t_oop), "instance should have alias type");
1339 1343
1340 1344 // Do nothing here if Identity will find a value
1341 1345 // (to avoid infinite chain of value phis generation).
1342 1346 if (!phase->eqv(this, this->Identity(phase)))
1343 1347 return NULL;
1344 1348
1345 1349 // Select Region to split through.
1346 1350 Node* region;
1347 1351 if (!base_is_phi) {
1348 1352 assert(mem->is_Phi(), "sanity");
1349 1353 region = mem->in(0);
1350 1354 // Skip if the region dominates some control edge of the address.
1351 1355 if (!MemNode::all_controls_dominate(address, region))
1352 1356 return NULL;
1353 1357 } else if (!mem->is_Phi()) {
1354 1358 assert(base_is_phi, "sanity");
1355 1359 region = base->in(0);
1356 1360 // Skip if the region dominates some control edge of the memory.
1357 1361 if (!MemNode::all_controls_dominate(mem, region))
1358 1362 return NULL;
1359 1363 } else if (base->in(0) != mem->in(0)) {
1360 1364 assert(base_is_phi && mem->is_Phi(), "sanity");
1361 1365 if (MemNode::all_controls_dominate(mem, base->in(0))) {
1362 1366 region = base->in(0);
1363 1367 } else if (MemNode::all_controls_dominate(address, mem->in(0))) {
1364 1368 region = mem->in(0);
1365 1369 } else {
1366 1370 return NULL; // complex graph
1367 1371 }
1368 1372 } else {
1369 1373 assert(base->in(0) == mem->in(0), "sanity");
1370 1374 region = mem->in(0);
1371 1375 }
1372 1376
1373 1377 const Type* this_type = this->bottom_type();
1374 1378 int this_index = C->get_alias_index(t_oop);
1375 1379 int this_offset = t_oop->offset();
1376 1380 int this_iid = t_oop->instance_id();
1377 1381 if (!t_oop->is_known_instance() && load_boxed_values) {
1378 1382 // Use _idx of address base for boxed values.
1379 1383 this_iid = base->_idx;
1380 1384 }
1381 1385 PhaseIterGVN* igvn = phase->is_IterGVN();
1382 1386 Node* phi = new (C) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset);
1383 1387 for (uint i = 1; i < region->req(); i++) {
1384 1388 Node* x;
1385 1389 Node* the_clone = NULL;
1386 1390 if (region->in(i) == C->top()) {
1387 1391 x = C->top(); // Dead path? Use a dead data op
1388 1392 } else {
1389 1393 x = this->clone(); // Else clone up the data op
1390 1394 the_clone = x; // Remember for possible deletion.
1391 1395 // Alter data node to use pre-phi inputs
1392 1396 if (this->in(0) == region) {
1393 1397 x->set_req(0, region->in(i));
1394 1398 } else {
1395 1399 x->set_req(0, NULL);
1396 1400 }
1397 1401 if (mem->is_Phi() && (mem->in(0) == region)) {
1398 1402 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone.
1399 1403 }
1400 1404 if (address->is_Phi() && address->in(0) == region) {
1401 1405 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone
1402 1406 }
1403 1407 if (base_is_phi && (base->in(0) == region)) {
1404 1408 Node* base_x = base->in(i); // Clone address for loads from boxed objects.
1405 1409 Node* adr_x = phase->transform(new (C) AddPNode(base_x,base_x,address->in(AddPNode::Offset)));
1406 1410 x->set_req(Address, adr_x);
1407 1411 }
1408 1412 }
1409 1413 // Check for a 'win' on some paths
1410 1414 const Type *t = x->Value(igvn);
1411 1415
1412 1416 bool singleton = t->singleton();
1413 1417
1414 1418 // See comments in PhaseIdealLoop::split_thru_phi().
1415 1419 if (singleton && t == Type::TOP) {
1416 1420 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1417 1421 }
1418 1422
1419 1423 if (singleton) {
1420 1424 x = igvn->makecon(t);
1421 1425 } else {
1422 1426 // We now call Identity to try to simplify the cloned node.
1423 1427 // Note that some Identity methods call phase->type(this).
1424 1428 // Make sure that the type array is big enough for
1425 1429 // our new node, even though we may throw the node away.
1426 1430 // (This tweaking with igvn only works because x is a new node.)
1427 1431 igvn->set_type(x, t);
1428 1432 // If x is a TypeNode, capture any more-precise type permanently into Node
1429 1433 // otherwise it will be not updated during igvn->transform since
1430 1434 // igvn->type(x) is set to x->Value() already.
1431 1435 x->raise_bottom_type(t);
1432 1436 Node *y = x->Identity(igvn);
1433 1437 if (y != x) {
1434 1438 x = y;
1435 1439 } else {
1436 1440 y = igvn->hash_find_insert(x);
1437 1441 if (y) {
1438 1442 x = y;
1439 1443 } else {
1440 1444 // Else x is a new node we are keeping
1441 1445 // We do not need register_new_node_with_optimizer
1442 1446 // because set_type has already been called.
1443 1447 igvn->_worklist.push(x);
1444 1448 }
1445 1449 }
1446 1450 }
1447 1451 if (x != the_clone && the_clone != NULL) {
↓ open down ↓ |
576 lines elided |
↑ open up ↑ |
1448 1452 igvn->remove_dead_node(the_clone);
1449 1453 }
1450 1454 phi->set_req(i, x);
1451 1455 }
1452 1456 // Record Phi
1453 1457 igvn->register_new_node_with_optimizer(phi);
1454 1458 return phi;
1455 1459 }
1456 1460
1457 1461 //------------------------------Ideal------------------------------------------
1458 -// If the load is from Field memory and the pointer is non-null, we can
1462 +// If the load is from Field memory and the pointer is non-null, it might be possible to
1459 1463 // zero out the control input.
1460 1464 // If the offset is constant and the base is an object allocation,
1461 1465 // try to hook me up to the exact initializing store.
1462 1466 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1463 1467 Node* p = MemNode::Ideal_common(phase, can_reshape);
1464 1468 if (p) return (p == NodeSentinel) ? NULL : p;
1465 1469
1466 1470 Node* ctrl = in(MemNode::Control);
1467 1471 Node* address = in(MemNode::Address);
1468 1472
1469 1473 // Skip up past a SafePoint control. Cannot do this for Stores because
1470 1474 // pointer stores & cardmarks must stay on the same side of a SafePoint.
1471 1475 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
1472 1476 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) {
↓ open down ↓ |
4 lines elided |
↑ open up ↑ |
1473 1477 ctrl = ctrl->in(0);
1474 1478 set_req(MemNode::Control,ctrl);
1475 1479 }
1476 1480
1477 1481 intptr_t ignore = 0;
1478 1482 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1479 1483 if (base != NULL
1480 1484 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
1481 1485 // Check for useless control edge in some common special cases
1482 1486 if (in(MemNode::Control) != NULL
1487 + && can_remove_control()
1483 1488 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1484 1489 && all_controls_dominate(base, phase->C->start())) {
1485 1490 // A method-invariant, non-null address (constant or 'this' argument).
1486 1491 set_req(MemNode::Control, NULL);
1487 1492 }
1488 1493 }
1489 1494
1490 1495 Node* mem = in(MemNode::Memory);
1491 1496 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1492 1497
1493 1498 if (can_reshape && (addr_t != NULL)) {
1494 1499 // try to optimize our memory input
1495 1500 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase);
1496 1501 if (opt_mem != mem) {
1497 1502 set_req(MemNode::Memory, opt_mem);
1498 1503 if (phase->type( opt_mem ) == Type::TOP) return NULL;
1499 1504 return this;
1500 1505 }
1501 1506 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1502 1507 if ((t_oop != NULL) &&
1503 1508 (t_oop->is_known_instance_field() ||
1504 1509 t_oop->is_ptr_to_boxed_value())) {
1505 1510 PhaseIterGVN *igvn = phase->is_IterGVN();
1506 1511 if (igvn != NULL && igvn->_worklist.member(opt_mem)) {
1507 1512 // Delay this transformation until memory Phi is processed.
1508 1513 phase->is_IterGVN()->_worklist.push(this);
1509 1514 return NULL;
1510 1515 }
1511 1516 // Split instance field load through Phi.
1512 1517 Node* result = split_through_phi(phase);
1513 1518 if (result != NULL) return result;
1514 1519
1515 1520 if (t_oop->is_ptr_to_boxed_value()) {
1516 1521 Node* result = eliminate_autobox(phase);
1517 1522 if (result != NULL) return result;
1518 1523 }
1519 1524 }
1520 1525 }
1521 1526
1522 1527 // Check for prior store with a different base or offset; make Load
1523 1528 // independent. Skip through any number of them. Bail out if the stores
1524 1529 // are in an endless dead cycle and report no progress. This is a key
1525 1530 // transform for Reflection. However, if after skipping through the Stores
1526 1531 // we can't then fold up against a prior store do NOT do the transform as
1527 1532 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
1528 1533 // array memory alive twice: once for the hoisted Load and again after the
1529 1534 // bypassed Store. This situation only works if EVERYBODY who does
1530 1535 // anti-dependence work knows how to bypass. I.e. we need all
1531 1536 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
1532 1537 // the alias index stuff. So instead, peek through Stores and IFF we can
1533 1538 // fold up, do so.
1534 1539 Node* prev_mem = find_previous_store(phase);
1535 1540 // Steps (a), (b): Walk past independent stores to find an exact match.
1536 1541 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
1537 1542 // (c) See if we can fold up on the spot, but don't fold up here.
1538 1543 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
1539 1544 // just return a prior value, which is done by Identity calls.
1540 1545 if (can_see_stored_value(prev_mem, phase)) {
1541 1546 // Make ready for step (d):
1542 1547 set_req(MemNode::Memory, prev_mem);
1543 1548 return this;
1544 1549 }
1545 1550 }
1546 1551
1547 1552 return NULL; // No further progress
1548 1553 }
1549 1554
1550 1555 // Helper to recognize certain Klass fields which are invariant across
1551 1556 // some group of array types (e.g., int[] or all T[] where T < Object).
1552 1557 const Type*
1553 1558 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1554 1559 ciKlass* klass) const {
1555 1560 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) {
1556 1561 // The field is Klass::_modifier_flags. Return its (constant) value.
1557 1562 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1558 1563 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1559 1564 return TypeInt::make(klass->modifier_flags());
1560 1565 }
1561 1566 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) {
1562 1567 // The field is Klass::_access_flags. Return its (constant) value.
1563 1568 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1564 1569 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1565 1570 return TypeInt::make(klass->access_flags());
1566 1571 }
1567 1572 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
1568 1573 // The field is Klass::_layout_helper. Return its constant value if known.
1569 1574 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1570 1575 return TypeInt::make(klass->layout_helper());
1571 1576 }
1572 1577
1573 1578 // No match.
1574 1579 return NULL;
1575 1580 }
1576 1581
1577 1582 // Try to constant-fold a stable array element.
1578 1583 static const Type* fold_stable_ary_elem(const TypeAryPtr* ary, int off, BasicType loadbt) {
1579 1584 assert(ary->const_oop(), "array should be constant");
1580 1585 assert(ary->is_stable(), "array should be stable");
1581 1586
1582 1587 // Decode the results of GraphKit::array_element_address.
1583 1588 ciArray* aobj = ary->const_oop()->as_array();
1584 1589 ciConstant con = aobj->element_value_by_offset(off);
1585 1590
1586 1591 if (con.basic_type() != T_ILLEGAL && !con.is_null_or_zero()) {
1587 1592 const Type* con_type = Type::make_from_constant(con);
1588 1593 if (con_type != NULL) {
1589 1594 if (con_type->isa_aryptr()) {
1590 1595 // Join with the array element type, in case it is also stable.
1591 1596 int dim = ary->stable_dimension();
1592 1597 con_type = con_type->is_aryptr()->cast_to_stable(true, dim-1);
1593 1598 }
1594 1599 if (loadbt == T_NARROWOOP && con_type->isa_oopptr()) {
1595 1600 con_type = con_type->make_narrowoop();
1596 1601 }
1597 1602 #ifndef PRODUCT
1598 1603 if (TraceIterativeGVN) {
1599 1604 tty->print("FoldStableValues: array element [off=%d]: con_type=", off);
1600 1605 con_type->dump(); tty->cr();
1601 1606 }
1602 1607 #endif //PRODUCT
1603 1608 return con_type;
1604 1609 }
1605 1610 }
1606 1611 return NULL;
1607 1612 }
1608 1613
1609 1614 //------------------------------Value-----------------------------------------
1610 1615 const Type *LoadNode::Value( PhaseTransform *phase ) const {
1611 1616 // Either input is TOP ==> the result is TOP
1612 1617 Node* mem = in(MemNode::Memory);
1613 1618 const Type *t1 = phase->type(mem);
1614 1619 if (t1 == Type::TOP) return Type::TOP;
1615 1620 Node* adr = in(MemNode::Address);
1616 1621 const TypePtr* tp = phase->type(adr)->isa_ptr();
1617 1622 if (tp == NULL || tp->empty()) return Type::TOP;
1618 1623 int off = tp->offset();
1619 1624 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1620 1625 Compile* C = phase->C;
1621 1626
1622 1627 // Try to guess loaded type from pointer type
1623 1628 if (tp->isa_aryptr()) {
1624 1629 const TypeAryPtr* ary = tp->is_aryptr();
1625 1630 const Type* t = ary->elem();
1626 1631
1627 1632 // Determine whether the reference is beyond the header or not, by comparing
1628 1633 // the offset against the offset of the start of the array's data.
1629 1634 // Different array types begin at slightly different offsets (12 vs. 16).
1630 1635 // We choose T_BYTE as an example base type that is least restrictive
1631 1636 // as to alignment, which will therefore produce the smallest
1632 1637 // possible base offset.
1633 1638 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1634 1639 const bool off_beyond_header = ((uint)off >= (uint)min_base_off);
1635 1640
1636 1641 // Try to constant-fold a stable array element.
1637 1642 if (FoldStableValues && ary->is_stable() && ary->const_oop() != NULL) {
1638 1643 // Make sure the reference is not into the header and the offset is constant
1639 1644 if (off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) {
1640 1645 const Type* con_type = fold_stable_ary_elem(ary, off, memory_type());
1641 1646 if (con_type != NULL) {
1642 1647 return con_type;
1643 1648 }
1644 1649 }
1645 1650 }
1646 1651
1647 1652 // Don't do this for integer types. There is only potential profit if
1648 1653 // the element type t is lower than _type; that is, for int types, if _type is
1649 1654 // more restrictive than t. This only happens here if one is short and the other
1650 1655 // char (both 16 bits), and in those cases we've made an intentional decision
1651 1656 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
1652 1657 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
1653 1658 //
1654 1659 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
1655 1660 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
1656 1661 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
1657 1662 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
1658 1663 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
1659 1664 // In fact, that could have been the original type of p1, and p1 could have
1660 1665 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
1661 1666 // expression (LShiftL quux 3) independently optimized to the constant 8.
1662 1667 if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
1663 1668 && (_type->isa_vect() == NULL)
1664 1669 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
1665 1670 // t might actually be lower than _type, if _type is a unique
1666 1671 // concrete subclass of abstract class t.
1667 1672 if (off_beyond_header) { // is the offset beyond the header?
1668 1673 const Type* jt = t->join_speculative(_type);
1669 1674 // In any case, do not allow the join, per se, to empty out the type.
1670 1675 if (jt->empty() && !t->empty()) {
1671 1676 // This can happen if a interface-typed array narrows to a class type.
1672 1677 jt = _type;
1673 1678 }
1674 1679 #ifdef ASSERT
1675 1680 if (phase->C->eliminate_boxing() && adr->is_AddP()) {
1676 1681 // The pointers in the autobox arrays are always non-null
1677 1682 Node* base = adr->in(AddPNode::Base);
1678 1683 if ((base != NULL) && base->is_DecodeN()) {
1679 1684 // Get LoadN node which loads IntegerCache.cache field
1680 1685 base = base->in(1);
1681 1686 }
1682 1687 if ((base != NULL) && base->is_Con()) {
1683 1688 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
1684 1689 if ((base_type != NULL) && base_type->is_autobox_cache()) {
1685 1690 // It could be narrow oop
1686 1691 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
1687 1692 }
1688 1693 }
1689 1694 }
1690 1695 #endif
1691 1696 return jt;
1692 1697 }
1693 1698 }
1694 1699 } else if (tp->base() == Type::InstPtr) {
1695 1700 ciEnv* env = C->env();
1696 1701 const TypeInstPtr* tinst = tp->is_instptr();
1697 1702 ciKlass* klass = tinst->klass();
1698 1703 assert( off != Type::OffsetBot ||
1699 1704 // arrays can be cast to Objects
1700 1705 tp->is_oopptr()->klass()->is_java_lang_Object() ||
1701 1706 // unsafe field access may not have a constant offset
1702 1707 C->has_unsafe_access(),
1703 1708 "Field accesses must be precise" );
1704 1709 // For oop loads, we expect the _type to be precise
1705 1710 if (klass == env->String_klass() &&
1706 1711 adr->is_AddP() && off != Type::OffsetBot) {
1707 1712 // For constant Strings treat the final fields as compile time constants.
1708 1713 Node* base = adr->in(AddPNode::Base);
1709 1714 const TypeOopPtr* t = phase->type(base)->isa_oopptr();
1710 1715 if (t != NULL && t->singleton()) {
1711 1716 ciField* field = env->String_klass()->get_field_by_offset(off, false);
1712 1717 if (field != NULL && field->is_final()) {
1713 1718 ciObject* string = t->const_oop();
1714 1719 ciConstant constant = string->as_instance()->field_value(field);
1715 1720 if (constant.basic_type() == T_INT) {
1716 1721 return TypeInt::make(constant.as_int());
1717 1722 } else if (constant.basic_type() == T_ARRAY) {
1718 1723 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
1719 1724 return TypeNarrowOop::make_from_constant(constant.as_object(), true);
1720 1725 } else {
1721 1726 return TypeOopPtr::make_from_constant(constant.as_object(), true);
1722 1727 }
1723 1728 }
1724 1729 }
1725 1730 }
1726 1731 }
1727 1732 // Optimizations for constant objects
1728 1733 ciObject* const_oop = tinst->const_oop();
1729 1734 if (const_oop != NULL) {
1730 1735 // For constant Boxed value treat the target field as a compile time constant.
1731 1736 if (tinst->is_ptr_to_boxed_value()) {
1732 1737 return tinst->get_const_boxed_value();
1733 1738 } else
1734 1739 // For constant CallSites treat the target field as a compile time constant.
1735 1740 if (const_oop->is_call_site()) {
1736 1741 ciCallSite* call_site = const_oop->as_call_site();
1737 1742 ciField* field = call_site->klass()->as_instance_klass()->get_field_by_offset(off, /*is_static=*/ false);
1738 1743 if (field != NULL && field->is_call_site_target()) {
1739 1744 ciMethodHandle* target = call_site->get_target();
1740 1745 if (target != NULL) { // just in case
1741 1746 ciConstant constant(T_OBJECT, target);
1742 1747 const Type* t;
1743 1748 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
1744 1749 t = TypeNarrowOop::make_from_constant(constant.as_object(), true);
1745 1750 } else {
1746 1751 t = TypeOopPtr::make_from_constant(constant.as_object(), true);
1747 1752 }
1748 1753 // Add a dependence for invalidation of the optimization.
1749 1754 if (!call_site->is_constant_call_site()) {
1750 1755 C->dependencies()->assert_call_site_target_value(call_site, target);
1751 1756 }
1752 1757 return t;
1753 1758 }
1754 1759 }
1755 1760 }
1756 1761 }
1757 1762 } else if (tp->base() == Type::KlassPtr) {
1758 1763 assert( off != Type::OffsetBot ||
1759 1764 // arrays can be cast to Objects
1760 1765 tp->is_klassptr()->klass()->is_java_lang_Object() ||
1761 1766 // also allow array-loading from the primary supertype
1762 1767 // array during subtype checks
1763 1768 Opcode() == Op_LoadKlass,
1764 1769 "Field accesses must be precise" );
1765 1770 // For klass/static loads, we expect the _type to be precise
1766 1771 }
1767 1772
1768 1773 const TypeKlassPtr *tkls = tp->isa_klassptr();
1769 1774 if (tkls != NULL && !StressReflectiveCode) {
1770 1775 ciKlass* klass = tkls->klass();
1771 1776 if (klass->is_loaded() && tkls->klass_is_exact()) {
1772 1777 // We are loading a field from a Klass metaobject whose identity
1773 1778 // is known at compile time (the type is "exact" or "precise").
1774 1779 // Check for fields we know are maintained as constants by the VM.
1775 1780 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
1776 1781 // The field is Klass::_super_check_offset. Return its (constant) value.
1777 1782 // (Folds up type checking code.)
1778 1783 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
1779 1784 return TypeInt::make(klass->super_check_offset());
1780 1785 }
1781 1786 // Compute index into primary_supers array
1782 1787 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1783 1788 // Check for overflowing; use unsigned compare to handle the negative case.
1784 1789 if( depth < ciKlass::primary_super_limit() ) {
1785 1790 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1786 1791 // (Folds up type checking code.)
1787 1792 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1788 1793 ciKlass *ss = klass->super_of_depth(depth);
1789 1794 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1790 1795 }
1791 1796 const Type* aift = load_array_final_field(tkls, klass);
1792 1797 if (aift != NULL) return aift;
1793 1798 if (tkls->offset() == in_bytes(ArrayKlass::component_mirror_offset())
1794 1799 && klass->is_array_klass()) {
1795 1800 // The field is ArrayKlass::_component_mirror. Return its (constant) value.
1796 1801 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.)
1797 1802 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror");
1798 1803 return TypeInstPtr::make(klass->as_array_klass()->component_mirror());
1799 1804 }
1800 1805 if (tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
1801 1806 // The field is Klass::_java_mirror. Return its (constant) value.
1802 1807 // (Folds up the 2nd indirection in anObjConstant.getClass().)
1803 1808 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1804 1809 return TypeInstPtr::make(klass->java_mirror());
1805 1810 }
1806 1811 }
1807 1812
1808 1813 // We can still check if we are loading from the primary_supers array at a
1809 1814 // shallow enough depth. Even though the klass is not exact, entries less
1810 1815 // than or equal to its super depth are correct.
1811 1816 if (klass->is_loaded() ) {
1812 1817 ciType *inner = klass;
1813 1818 while( inner->is_obj_array_klass() )
1814 1819 inner = inner->as_obj_array_klass()->base_element_type();
1815 1820 if( inner->is_instance_klass() &&
1816 1821 !inner->as_instance_klass()->flags().is_interface() ) {
1817 1822 // Compute index into primary_supers array
1818 1823 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1819 1824 // Check for overflowing; use unsigned compare to handle the negative case.
1820 1825 if( depth < ciKlass::primary_super_limit() &&
1821 1826 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
1822 1827 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1823 1828 // (Folds up type checking code.)
1824 1829 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1825 1830 ciKlass *ss = klass->super_of_depth(depth);
1826 1831 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1827 1832 }
1828 1833 }
1829 1834 }
1830 1835
1831 1836 // If the type is enough to determine that the thing is not an array,
1832 1837 // we can give the layout_helper a positive interval type.
1833 1838 // This will help short-circuit some reflective code.
1834 1839 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())
1835 1840 && !klass->is_array_klass() // not directly typed as an array
1836 1841 && !klass->is_interface() // specifically not Serializable & Cloneable
1837 1842 && !klass->is_java_lang_Object() // not the supertype of all T[]
1838 1843 ) {
1839 1844 // Note: When interfaces are reliable, we can narrow the interface
1840 1845 // test to (klass != Serializable && klass != Cloneable).
1841 1846 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
1842 1847 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
1843 1848 // The key property of this type is that it folds up tests
1844 1849 // for array-ness, since it proves that the layout_helper is positive.
1845 1850 // Thus, a generic value like the basic object layout helper works fine.
1846 1851 return TypeInt::make(min_size, max_jint, Type::WidenMin);
1847 1852 }
1848 1853 }
1849 1854
1850 1855 // If we are loading from a freshly-allocated object, produce a zero,
1851 1856 // if the load is provably beyond the header of the object.
1852 1857 // (Also allow a variable load from a fresh array to produce zero.)
1853 1858 const TypeOopPtr *tinst = tp->isa_oopptr();
1854 1859 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field();
1855 1860 bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value();
1856 1861 if (ReduceFieldZeroing || is_instance || is_boxed_value) {
1857 1862 Node* value = can_see_stored_value(mem,phase);
1858 1863 if (value != NULL && value->is_Con()) {
1859 1864 assert(value->bottom_type()->higher_equal(_type),"sanity");
1860 1865 return value->bottom_type();
1861 1866 }
1862 1867 }
1863 1868
1864 1869 if (is_instance) {
1865 1870 // If we have an instance type and our memory input is the
1866 1871 // programs's initial memory state, there is no matching store,
1867 1872 // so just return a zero of the appropriate type
1868 1873 Node *mem = in(MemNode::Memory);
1869 1874 if (mem->is_Parm() && mem->in(0)->is_Start()) {
1870 1875 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
1871 1876 return Type::get_zero_type(_type->basic_type());
1872 1877 }
1873 1878 }
1874 1879 return _type;
1875 1880 }
1876 1881
1877 1882 //------------------------------match_edge-------------------------------------
1878 1883 // Do we Match on this edge index or not? Match only the address.
1879 1884 uint LoadNode::match_edge(uint idx) const {
1880 1885 return idx == MemNode::Address;
1881 1886 }
1882 1887
1883 1888 //--------------------------LoadBNode::Ideal--------------------------------------
1884 1889 //
1885 1890 // If the previous store is to the same address as this load,
1886 1891 // and the value stored was larger than a byte, replace this load
1887 1892 // with the value stored truncated to a byte. If no truncation is
1888 1893 // needed, the replacement is done in LoadNode::Identity().
1889 1894 //
1890 1895 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1891 1896 Node* mem = in(MemNode::Memory);
1892 1897 Node* value = can_see_stored_value(mem,phase);
1893 1898 if( value && !phase->type(value)->higher_equal( _type ) ) {
1894 1899 Node *result = phase->transform( new (phase->C) LShiftINode(value, phase->intcon(24)) );
1895 1900 return new (phase->C) RShiftINode(result, phase->intcon(24));
1896 1901 }
1897 1902 // Identity call will handle the case where truncation is not needed.
1898 1903 return LoadNode::Ideal(phase, can_reshape);
1899 1904 }
1900 1905
1901 1906 const Type* LoadBNode::Value(PhaseTransform *phase) const {
1902 1907 Node* mem = in(MemNode::Memory);
1903 1908 Node* value = can_see_stored_value(mem,phase);
1904 1909 if (value != NULL && value->is_Con() &&
1905 1910 !value->bottom_type()->higher_equal(_type)) {
1906 1911 // If the input to the store does not fit with the load's result type,
1907 1912 // it must be truncated. We can't delay until Ideal call since
1908 1913 // a singleton Value is needed for split_thru_phi optimization.
1909 1914 int con = value->get_int();
1910 1915 return TypeInt::make((con << 24) >> 24);
1911 1916 }
1912 1917 return LoadNode::Value(phase);
1913 1918 }
1914 1919
1915 1920 //--------------------------LoadUBNode::Ideal-------------------------------------
1916 1921 //
1917 1922 // If the previous store is to the same address as this load,
1918 1923 // and the value stored was larger than a byte, replace this load
1919 1924 // with the value stored truncated to a byte. If no truncation is
1920 1925 // needed, the replacement is done in LoadNode::Identity().
1921 1926 //
1922 1927 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1923 1928 Node* mem = in(MemNode::Memory);
1924 1929 Node* value = can_see_stored_value(mem, phase);
1925 1930 if (value && !phase->type(value)->higher_equal(_type))
1926 1931 return new (phase->C) AndINode(value, phase->intcon(0xFF));
1927 1932 // Identity call will handle the case where truncation is not needed.
1928 1933 return LoadNode::Ideal(phase, can_reshape);
1929 1934 }
1930 1935
1931 1936 const Type* LoadUBNode::Value(PhaseTransform *phase) const {
1932 1937 Node* mem = in(MemNode::Memory);
1933 1938 Node* value = can_see_stored_value(mem,phase);
1934 1939 if (value != NULL && value->is_Con() &&
1935 1940 !value->bottom_type()->higher_equal(_type)) {
1936 1941 // If the input to the store does not fit with the load's result type,
1937 1942 // it must be truncated. We can't delay until Ideal call since
1938 1943 // a singleton Value is needed for split_thru_phi optimization.
1939 1944 int con = value->get_int();
1940 1945 return TypeInt::make(con & 0xFF);
1941 1946 }
1942 1947 return LoadNode::Value(phase);
1943 1948 }
1944 1949
1945 1950 //--------------------------LoadUSNode::Ideal-------------------------------------
1946 1951 //
1947 1952 // If the previous store is to the same address as this load,
1948 1953 // and the value stored was larger than a char, replace this load
1949 1954 // with the value stored truncated to a char. If no truncation is
1950 1955 // needed, the replacement is done in LoadNode::Identity().
1951 1956 //
1952 1957 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1953 1958 Node* mem = in(MemNode::Memory);
1954 1959 Node* value = can_see_stored_value(mem,phase);
1955 1960 if( value && !phase->type(value)->higher_equal( _type ) )
1956 1961 return new (phase->C) AndINode(value,phase->intcon(0xFFFF));
1957 1962 // Identity call will handle the case where truncation is not needed.
1958 1963 return LoadNode::Ideal(phase, can_reshape);
1959 1964 }
1960 1965
1961 1966 const Type* LoadUSNode::Value(PhaseTransform *phase) const {
1962 1967 Node* mem = in(MemNode::Memory);
1963 1968 Node* value = can_see_stored_value(mem,phase);
1964 1969 if (value != NULL && value->is_Con() &&
1965 1970 !value->bottom_type()->higher_equal(_type)) {
1966 1971 // If the input to the store does not fit with the load's result type,
1967 1972 // it must be truncated. We can't delay until Ideal call since
1968 1973 // a singleton Value is needed for split_thru_phi optimization.
1969 1974 int con = value->get_int();
1970 1975 return TypeInt::make(con & 0xFFFF);
1971 1976 }
1972 1977 return LoadNode::Value(phase);
1973 1978 }
1974 1979
1975 1980 //--------------------------LoadSNode::Ideal--------------------------------------
1976 1981 //
1977 1982 // If the previous store is to the same address as this load,
1978 1983 // and the value stored was larger than a short, replace this load
1979 1984 // with the value stored truncated to a short. If no truncation is
1980 1985 // needed, the replacement is done in LoadNode::Identity().
1981 1986 //
1982 1987 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1983 1988 Node* mem = in(MemNode::Memory);
1984 1989 Node* value = can_see_stored_value(mem,phase);
1985 1990 if( value && !phase->type(value)->higher_equal( _type ) ) {
1986 1991 Node *result = phase->transform( new (phase->C) LShiftINode(value, phase->intcon(16)) );
1987 1992 return new (phase->C) RShiftINode(result, phase->intcon(16));
1988 1993 }
1989 1994 // Identity call will handle the case where truncation is not needed.
1990 1995 return LoadNode::Ideal(phase, can_reshape);
1991 1996 }
1992 1997
1993 1998 const Type* LoadSNode::Value(PhaseTransform *phase) const {
1994 1999 Node* mem = in(MemNode::Memory);
1995 2000 Node* value = can_see_stored_value(mem,phase);
1996 2001 if (value != NULL && value->is_Con() &&
1997 2002 !value->bottom_type()->higher_equal(_type)) {
1998 2003 // If the input to the store does not fit with the load's result type,
1999 2004 // it must be truncated. We can't delay until Ideal call since
↓ open down ↓ |
507 lines elided |
↑ open up ↑ |
2000 2005 // a singleton Value is needed for split_thru_phi optimization.
2001 2006 int con = value->get_int();
2002 2007 return TypeInt::make((con << 16) >> 16);
2003 2008 }
2004 2009 return LoadNode::Value(phase);
2005 2010 }
2006 2011
2007 2012 //=============================================================================
2008 2013 //----------------------------LoadKlassNode::make------------------------------
2009 2014 // Polymorphic factory method:
2010 -Node *LoadKlassNode::make( PhaseGVN& gvn, Node *mem, Node *adr, const TypePtr* at, const TypeKlassPtr *tk ) {
2015 +Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node *mem, Node *adr, const TypePtr* at, const TypeKlassPtr *tk) {
2011 2016 Compile* C = gvn.C;
2012 - Node *ctl = NULL;
2013 2017 // sanity check the alias category against the created node type
2014 2018 const TypePtr *adr_type = adr->bottom_type()->isa_ptr();
2015 2019 assert(adr_type != NULL, "expecting TypeKlassPtr");
2016 2020 #ifdef _LP64
2017 2021 if (adr_type->is_ptr_to_narrowklass()) {
2018 2022 assert(UseCompressedClassPointers, "no compressed klasses");
2019 2023 Node* load_klass = gvn.transform(new (C) LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered));
2020 2024 return new (C) DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
2021 2025 }
2022 2026 #endif
2023 2027 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
2024 2028 return new (C) LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered);
2025 2029 }
2026 2030
2027 2031 //------------------------------Value------------------------------------------
2028 2032 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const {
2029 2033 return klass_value_common(phase);
2030 2034 }
2031 2035
2036 +// In most cases, LoadKlassNode does not have the control input set. If the control
2037 +// input is set, it must not be removed (by LoadNode::Ideal()).
2038 +bool LoadKlassNode::can_remove_control() const {
2039 + return false;
2040 +}
2041 +
2032 2042 const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const {
2033 2043 // Either input is TOP ==> the result is TOP
2034 2044 const Type *t1 = phase->type( in(MemNode::Memory) );
2035 2045 if (t1 == Type::TOP) return Type::TOP;
2036 2046 Node *adr = in(MemNode::Address);
2037 2047 const Type *t2 = phase->type( adr );
2038 2048 if (t2 == Type::TOP) return Type::TOP;
2039 2049 const TypePtr *tp = t2->is_ptr();
2040 2050 if (TypePtr::above_centerline(tp->ptr()) ||
2041 2051 tp->ptr() == TypePtr::Null) return Type::TOP;
2042 2052
2043 2053 // Return a more precise klass, if possible
2044 2054 const TypeInstPtr *tinst = tp->isa_instptr();
2045 2055 if (tinst != NULL) {
2046 2056 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
2047 2057 int offset = tinst->offset();
2048 2058 if (ik == phase->C->env()->Class_klass()
2049 2059 && (offset == java_lang_Class::klass_offset_in_bytes() ||
2050 2060 offset == java_lang_Class::array_klass_offset_in_bytes())) {
2051 2061 // We are loading a special hidden field from a Class mirror object,
2052 2062 // the field which points to the VM's Klass metaobject.
2053 2063 ciType* t = tinst->java_mirror_type();
2054 2064 // java_mirror_type returns non-null for compile-time Class constants.
2055 2065 if (t != NULL) {
2056 2066 // constant oop => constant klass
2057 2067 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
2058 2068 if (t->is_void()) {
2059 2069 // We cannot create a void array. Since void is a primitive type return null
2060 2070 // klass. Users of this result need to do a null check on the returned klass.
2061 2071 return TypePtr::NULL_PTR;
2062 2072 }
2063 2073 return TypeKlassPtr::make(ciArrayKlass::make(t));
2064 2074 }
2065 2075 if (!t->is_klass()) {
2066 2076 // a primitive Class (e.g., int.class) has NULL for a klass field
2067 2077 return TypePtr::NULL_PTR;
2068 2078 }
2069 2079 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
2070 2080 return TypeKlassPtr::make(t->as_klass());
2071 2081 }
2072 2082 // non-constant mirror, so we can't tell what's going on
2073 2083 }
2074 2084 if( !ik->is_loaded() )
2075 2085 return _type; // Bail out if not loaded
2076 2086 if (offset == oopDesc::klass_offset_in_bytes()) {
2077 2087 if (tinst->klass_is_exact()) {
2078 2088 return TypeKlassPtr::make(ik);
2079 2089 }
2080 2090 // See if we can become precise: no subklasses and no interface
2081 2091 // (Note: We need to support verified interfaces.)
2082 2092 if (!ik->is_interface() && !ik->has_subklass()) {
2083 2093 //assert(!UseExactTypes, "this code should be useless with exact types");
2084 2094 // Add a dependence; if any subclass added we need to recompile
2085 2095 if (!ik->is_final()) {
2086 2096 // %%% should use stronger assert_unique_concrete_subtype instead
2087 2097 phase->C->dependencies()->assert_leaf_type(ik);
2088 2098 }
2089 2099 // Return precise klass
2090 2100 return TypeKlassPtr::make(ik);
2091 2101 }
2092 2102
2093 2103 // Return root of possible klass
2094 2104 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
2095 2105 }
2096 2106 }
2097 2107
2098 2108 // Check for loading klass from an array
2099 2109 const TypeAryPtr *tary = tp->isa_aryptr();
2100 2110 if( tary != NULL ) {
2101 2111 ciKlass *tary_klass = tary->klass();
2102 2112 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP
2103 2113 && tary->offset() == oopDesc::klass_offset_in_bytes()) {
2104 2114 if (tary->klass_is_exact()) {
2105 2115 return TypeKlassPtr::make(tary_klass);
2106 2116 }
2107 2117 ciArrayKlass *ak = tary->klass()->as_array_klass();
2108 2118 // If the klass is an object array, we defer the question to the
2109 2119 // array component klass.
2110 2120 if( ak->is_obj_array_klass() ) {
2111 2121 assert( ak->is_loaded(), "" );
2112 2122 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
2113 2123 if( base_k->is_loaded() && base_k->is_instance_klass() ) {
2114 2124 ciInstanceKlass* ik = base_k->as_instance_klass();
2115 2125 // See if we can become precise: no subklasses and no interface
2116 2126 if (!ik->is_interface() && !ik->has_subklass()) {
2117 2127 //assert(!UseExactTypes, "this code should be useless with exact types");
2118 2128 // Add a dependence; if any subclass added we need to recompile
2119 2129 if (!ik->is_final()) {
2120 2130 phase->C->dependencies()->assert_leaf_type(ik);
2121 2131 }
2122 2132 // Return precise array klass
2123 2133 return TypeKlassPtr::make(ak);
2124 2134 }
2125 2135 }
2126 2136 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
2127 2137 } else { // Found a type-array?
2128 2138 //assert(!UseExactTypes, "this code should be useless with exact types");
2129 2139 assert( ak->is_type_array_klass(), "" );
2130 2140 return TypeKlassPtr::make(ak); // These are always precise
2131 2141 }
2132 2142 }
2133 2143 }
2134 2144
2135 2145 // Check for loading klass from an array klass
2136 2146 const TypeKlassPtr *tkls = tp->isa_klassptr();
2137 2147 if (tkls != NULL && !StressReflectiveCode) {
2138 2148 ciKlass* klass = tkls->klass();
2139 2149 if( !klass->is_loaded() )
2140 2150 return _type; // Bail out if not loaded
2141 2151 if( klass->is_obj_array_klass() &&
2142 2152 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2143 2153 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
2144 2154 // // Always returning precise element type is incorrect,
2145 2155 // // e.g., element type could be object and array may contain strings
2146 2156 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2147 2157
2148 2158 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2149 2159 // according to the element type's subclassing.
2150 2160 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
2151 2161 }
2152 2162 if( klass->is_instance_klass() && tkls->klass_is_exact() &&
2153 2163 tkls->offset() == in_bytes(Klass::super_offset())) {
2154 2164 ciKlass* sup = klass->as_instance_klass()->super();
2155 2165 // The field is Klass::_super. Return its (constant) value.
2156 2166 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
2157 2167 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
2158 2168 }
2159 2169 }
2160 2170
2161 2171 // Bailout case
2162 2172 return LoadNode::Value(phase);
2163 2173 }
2164 2174
2165 2175 //------------------------------Identity---------------------------------------
2166 2176 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
2167 2177 // Also feed through the klass in Allocate(...klass...)._klass.
2168 2178 Node* LoadKlassNode::Identity( PhaseTransform *phase ) {
2169 2179 return klass_identity_common(phase);
2170 2180 }
2171 2181
2172 2182 Node* LoadNode::klass_identity_common(PhaseTransform *phase ) {
2173 2183 Node* x = LoadNode::Identity(phase);
2174 2184 if (x != this) return x;
2175 2185
2176 2186 // Take apart the address into an oop and and offset.
2177 2187 // Return 'this' if we cannot.
2178 2188 Node* adr = in(MemNode::Address);
2179 2189 intptr_t offset = 0;
2180 2190 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2181 2191 if (base == NULL) return this;
2182 2192 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
2183 2193 if (toop == NULL) return this;
2184 2194
2185 2195 // We can fetch the klass directly through an AllocateNode.
2186 2196 // This works even if the klass is not constant (clone or newArray).
2187 2197 if (offset == oopDesc::klass_offset_in_bytes()) {
2188 2198 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
2189 2199 if (allocated_klass != NULL) {
2190 2200 return allocated_klass;
2191 2201 }
2192 2202 }
2193 2203
2194 2204 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
2195 2205 // Simplify ak.component_mirror.array_klass to plain ak, ak an ArrayKlass.
2196 2206 // See inline_native_Class_query for occurrences of these patterns.
2197 2207 // Java Example: x.getClass().isAssignableFrom(y)
2198 2208 // Java Example: Array.newInstance(x.getClass().getComponentType(), n)
2199 2209 //
2200 2210 // This improves reflective code, often making the Class
2201 2211 // mirror go completely dead. (Current exception: Class
2202 2212 // mirrors may appear in debug info, but we could clean them out by
2203 2213 // introducing a new debug info operator for Klass*.java_mirror).
2204 2214 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
2205 2215 && (offset == java_lang_Class::klass_offset_in_bytes() ||
2206 2216 offset == java_lang_Class::array_klass_offset_in_bytes())) {
2207 2217 // We are loading a special hidden field from a Class mirror,
2208 2218 // the field which points to its Klass or ArrayKlass metaobject.
2209 2219 if (base->is_Load()) {
2210 2220 Node* adr2 = base->in(MemNode::Address);
2211 2221 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2212 2222 if (tkls != NULL && !tkls->empty()
2213 2223 && (tkls->klass()->is_instance_klass() ||
2214 2224 tkls->klass()->is_array_klass())
2215 2225 && adr2->is_AddP()
2216 2226 ) {
2217 2227 int mirror_field = in_bytes(Klass::java_mirror_offset());
2218 2228 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
2219 2229 mirror_field = in_bytes(ArrayKlass::component_mirror_offset());
2220 2230 }
2221 2231 if (tkls->offset() == mirror_field) {
2222 2232 return adr2->in(AddPNode::Base);
2223 2233 }
2224 2234 }
2225 2235 }
2226 2236 }
2227 2237
2228 2238 return this;
2229 2239 }
2230 2240
2231 2241
2232 2242 //------------------------------Value------------------------------------------
2233 2243 const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const {
2234 2244 const Type *t = klass_value_common(phase);
2235 2245 if (t == Type::TOP)
2236 2246 return t;
2237 2247
2238 2248 return t->make_narrowklass();
2239 2249 }
2240 2250
2241 2251 //------------------------------Identity---------------------------------------
2242 2252 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
2243 2253 // Also feed through the klass in Allocate(...klass...)._klass.
2244 2254 Node* LoadNKlassNode::Identity( PhaseTransform *phase ) {
2245 2255 Node *x = klass_identity_common(phase);
2246 2256
2247 2257 const Type *t = phase->type( x );
2248 2258 if( t == Type::TOP ) return x;
2249 2259 if( t->isa_narrowklass()) return x;
2250 2260 assert (!t->isa_narrowoop(), "no narrow oop here");
2251 2261
2252 2262 return phase->transform(new (phase->C) EncodePKlassNode(x, t->make_narrowklass()));
2253 2263 }
2254 2264
2255 2265 //------------------------------Value-----------------------------------------
2256 2266 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const {
2257 2267 // Either input is TOP ==> the result is TOP
2258 2268 const Type *t1 = phase->type( in(MemNode::Memory) );
2259 2269 if( t1 == Type::TOP ) return Type::TOP;
2260 2270 Node *adr = in(MemNode::Address);
2261 2271 const Type *t2 = phase->type( adr );
2262 2272 if( t2 == Type::TOP ) return Type::TOP;
2263 2273 const TypePtr *tp = t2->is_ptr();
2264 2274 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
2265 2275 const TypeAryPtr *tap = tp->isa_aryptr();
2266 2276 if( !tap ) return _type;
2267 2277 return tap->size();
2268 2278 }
2269 2279
2270 2280 //-------------------------------Ideal---------------------------------------
2271 2281 // Feed through the length in AllocateArray(...length...)._length.
2272 2282 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2273 2283 Node* p = MemNode::Ideal_common(phase, can_reshape);
2274 2284 if (p) return (p == NodeSentinel) ? NULL : p;
2275 2285
2276 2286 // Take apart the address into an oop and and offset.
2277 2287 // Return 'this' if we cannot.
2278 2288 Node* adr = in(MemNode::Address);
2279 2289 intptr_t offset = 0;
2280 2290 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2281 2291 if (base == NULL) return NULL;
2282 2292 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2283 2293 if (tary == NULL) return NULL;
2284 2294
2285 2295 // We can fetch the length directly through an AllocateArrayNode.
2286 2296 // This works even if the length is not constant (clone or newArray).
2287 2297 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2288 2298 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2289 2299 if (alloc != NULL) {
2290 2300 Node* allocated_length = alloc->Ideal_length();
2291 2301 Node* len = alloc->make_ideal_length(tary, phase);
2292 2302 if (allocated_length != len) {
2293 2303 // New CastII improves on this.
2294 2304 return len;
2295 2305 }
2296 2306 }
2297 2307 }
2298 2308
2299 2309 return NULL;
2300 2310 }
2301 2311
2302 2312 //------------------------------Identity---------------------------------------
2303 2313 // Feed through the length in AllocateArray(...length...)._length.
2304 2314 Node* LoadRangeNode::Identity( PhaseTransform *phase ) {
2305 2315 Node* x = LoadINode::Identity(phase);
2306 2316 if (x != this) return x;
2307 2317
2308 2318 // Take apart the address into an oop and and offset.
2309 2319 // Return 'this' if we cannot.
2310 2320 Node* adr = in(MemNode::Address);
2311 2321 intptr_t offset = 0;
2312 2322 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2313 2323 if (base == NULL) return this;
2314 2324 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2315 2325 if (tary == NULL) return this;
2316 2326
2317 2327 // We can fetch the length directly through an AllocateArrayNode.
2318 2328 // This works even if the length is not constant (clone or newArray).
2319 2329 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2320 2330 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2321 2331 if (alloc != NULL) {
2322 2332 Node* allocated_length = alloc->Ideal_length();
2323 2333 // Do not allow make_ideal_length to allocate a CastII node.
2324 2334 Node* len = alloc->make_ideal_length(tary, phase, false);
2325 2335 if (allocated_length == len) {
2326 2336 // Return allocated_length only if it would not be improved by a CastII.
2327 2337 return allocated_length;
2328 2338 }
2329 2339 }
2330 2340 }
2331 2341
2332 2342 return this;
2333 2343
2334 2344 }
2335 2345
2336 2346 //=============================================================================
2337 2347 //---------------------------StoreNode::make-----------------------------------
2338 2348 // Polymorphic factory method:
2339 2349 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo) {
2340 2350 assert((mo == unordered || mo == release), "unexpected");
2341 2351 Compile* C = gvn.C;
2342 2352 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
2343 2353 ctl != NULL, "raw memory operations should have control edge");
2344 2354
2345 2355 switch (bt) {
2346 2356 case T_BOOLEAN:
2347 2357 case T_BYTE: return new (C) StoreBNode(ctl, mem, adr, adr_type, val, mo);
2348 2358 case T_INT: return new (C) StoreINode(ctl, mem, adr, adr_type, val, mo);
2349 2359 case T_CHAR:
2350 2360 case T_SHORT: return new (C) StoreCNode(ctl, mem, adr, adr_type, val, mo);
2351 2361 case T_LONG: return new (C) StoreLNode(ctl, mem, adr, adr_type, val, mo);
2352 2362 case T_FLOAT: return new (C) StoreFNode(ctl, mem, adr, adr_type, val, mo);
2353 2363 case T_DOUBLE: return new (C) StoreDNode(ctl, mem, adr, adr_type, val, mo);
2354 2364 case T_METADATA:
2355 2365 case T_ADDRESS:
2356 2366 case T_OBJECT:
2357 2367 #ifdef _LP64
2358 2368 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2359 2369 val = gvn.transform(new (C) EncodePNode(val, val->bottom_type()->make_narrowoop()));
2360 2370 return new (C) StoreNNode(ctl, mem, adr, adr_type, val, mo);
2361 2371 } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
2362 2372 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() &&
2363 2373 adr->bottom_type()->isa_rawptr())) {
2364 2374 val = gvn.transform(new (C) EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
2365 2375 return new (C) StoreNKlassNode(ctl, mem, adr, adr_type, val, mo);
2366 2376 }
2367 2377 #endif
2368 2378 {
2369 2379 return new (C) StorePNode(ctl, mem, adr, adr_type, val, mo);
2370 2380 }
2371 2381 }
2372 2382 ShouldNotReachHere();
2373 2383 return (StoreNode*)NULL;
2374 2384 }
2375 2385
2376 2386 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
2377 2387 bool require_atomic = true;
2378 2388 return new (C) StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
2379 2389 }
2380 2390
2381 2391
2382 2392 //--------------------------bottom_type----------------------------------------
2383 2393 const Type *StoreNode::bottom_type() const {
2384 2394 return Type::MEMORY;
2385 2395 }
2386 2396
2387 2397 //------------------------------hash-------------------------------------------
2388 2398 uint StoreNode::hash() const {
2389 2399 // unroll addition of interesting fields
2390 2400 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2391 2401
2392 2402 // Since they are not commoned, do not hash them:
2393 2403 return NO_HASH;
2394 2404 }
2395 2405
2396 2406 //------------------------------Ideal------------------------------------------
2397 2407 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
2398 2408 // When a store immediately follows a relevant allocation/initialization,
2399 2409 // try to capture it into the initialization, or hoist it above.
2400 2410 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2401 2411 Node* p = MemNode::Ideal_common(phase, can_reshape);
2402 2412 if (p) return (p == NodeSentinel) ? NULL : p;
2403 2413
2404 2414 Node* mem = in(MemNode::Memory);
2405 2415 Node* address = in(MemNode::Address);
2406 2416
2407 2417 // Back-to-back stores to same address? Fold em up. Generally
2408 2418 // unsafe if I have intervening uses... Also disallowed for StoreCM
2409 2419 // since they must follow each StoreP operation. Redundant StoreCMs
2410 2420 // are eliminated just before matching in final_graph_reshape.
2411 2421 if (mem->is_Store() && mem->in(MemNode::Address)->eqv_uncast(address) &&
2412 2422 mem->Opcode() != Op_StoreCM) {
2413 2423 // Looking at a dead closed cycle of memory?
2414 2424 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
2415 2425
2416 2426 assert(Opcode() == mem->Opcode() ||
2417 2427 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw,
2418 2428 "no mismatched stores, except on raw memory");
2419 2429
2420 2430 if (mem->outcnt() == 1 && // check for intervening uses
2421 2431 mem->as_Store()->memory_size() <= this->memory_size()) {
2422 2432 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away.
2423 2433 // For example, 'mem' might be the final state at a conditional return.
2424 2434 // Or, 'mem' might be used by some node which is live at the same time
2425 2435 // 'this' is live, which might be unschedulable. So, require exactly
2426 2436 // ONE user, the 'this' store, until such time as we clone 'mem' for
2427 2437 // each of 'mem's uses (thus making the exactly-1-user-rule hold true).
2428 2438 if (can_reshape) { // (%%% is this an anachronism?)
2429 2439 set_req_X(MemNode::Memory, mem->in(MemNode::Memory),
2430 2440 phase->is_IterGVN());
2431 2441 } else {
2432 2442 // It's OK to do this in the parser, since DU info is always accurate,
2433 2443 // and the parser always refers to nodes via SafePointNode maps.
2434 2444 set_req(MemNode::Memory, mem->in(MemNode::Memory));
2435 2445 }
2436 2446 return this;
2437 2447 }
2438 2448 }
2439 2449
2440 2450 // Capture an unaliased, unconditional, simple store into an initializer.
2441 2451 // Or, if it is independent of the allocation, hoist it above the allocation.
2442 2452 if (ReduceFieldZeroing && /*can_reshape &&*/
2443 2453 mem->is_Proj() && mem->in(0)->is_Initialize()) {
2444 2454 InitializeNode* init = mem->in(0)->as_Initialize();
2445 2455 intptr_t offset = init->can_capture_store(this, phase, can_reshape);
2446 2456 if (offset > 0) {
2447 2457 Node* moved = init->capture_store(this, offset, phase, can_reshape);
2448 2458 // If the InitializeNode captured me, it made a raw copy of me,
2449 2459 // and I need to disappear.
2450 2460 if (moved != NULL) {
2451 2461 // %%% hack to ensure that Ideal returns a new node:
2452 2462 mem = MergeMemNode::make(phase->C, mem);
2453 2463 return mem; // fold me away
2454 2464 }
2455 2465 }
2456 2466 }
2457 2467
2458 2468 return NULL; // No further progress
2459 2469 }
2460 2470
2461 2471 //------------------------------Value-----------------------------------------
2462 2472 const Type *StoreNode::Value( PhaseTransform *phase ) const {
2463 2473 // Either input is TOP ==> the result is TOP
2464 2474 const Type *t1 = phase->type( in(MemNode::Memory) );
2465 2475 if( t1 == Type::TOP ) return Type::TOP;
2466 2476 const Type *t2 = phase->type( in(MemNode::Address) );
2467 2477 if( t2 == Type::TOP ) return Type::TOP;
2468 2478 const Type *t3 = phase->type( in(MemNode::ValueIn) );
2469 2479 if( t3 == Type::TOP ) return Type::TOP;
2470 2480 return Type::MEMORY;
2471 2481 }
2472 2482
2473 2483 //------------------------------Identity---------------------------------------
2474 2484 // Remove redundant stores:
2475 2485 // Store(m, p, Load(m, p)) changes to m.
2476 2486 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
2477 2487 Node *StoreNode::Identity( PhaseTransform *phase ) {
2478 2488 Node* mem = in(MemNode::Memory);
2479 2489 Node* adr = in(MemNode::Address);
2480 2490 Node* val = in(MemNode::ValueIn);
2481 2491
2482 2492 // Load then Store? Then the Store is useless
2483 2493 if (val->is_Load() &&
2484 2494 val->in(MemNode::Address)->eqv_uncast(adr) &&
2485 2495 val->in(MemNode::Memory )->eqv_uncast(mem) &&
2486 2496 val->as_Load()->store_Opcode() == Opcode()) {
2487 2497 return mem;
2488 2498 }
2489 2499
2490 2500 // Two stores in a row of the same value?
2491 2501 if (mem->is_Store() &&
2492 2502 mem->in(MemNode::Address)->eqv_uncast(adr) &&
2493 2503 mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
2494 2504 mem->Opcode() == Opcode()) {
2495 2505 return mem;
2496 2506 }
2497 2507
2498 2508 // Store of zero anywhere into a freshly-allocated object?
2499 2509 // Then the store is useless.
2500 2510 // (It must already have been captured by the InitializeNode.)
2501 2511 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
2502 2512 // a newly allocated object is already all-zeroes everywhere
2503 2513 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
2504 2514 return mem;
2505 2515 }
2506 2516
2507 2517 // the store may also apply to zero-bits in an earlier object
2508 2518 Node* prev_mem = find_previous_store(phase);
2509 2519 // Steps (a), (b): Walk past independent stores to find an exact match.
2510 2520 if (prev_mem != NULL) {
2511 2521 Node* prev_val = can_see_stored_value(prev_mem, phase);
2512 2522 if (prev_val != NULL && phase->eqv(prev_val, val)) {
2513 2523 // prev_val and val might differ by a cast; it would be good
2514 2524 // to keep the more informative of the two.
2515 2525 return mem;
2516 2526 }
2517 2527 }
2518 2528 }
2519 2529
2520 2530 return this;
2521 2531 }
2522 2532
2523 2533 //------------------------------match_edge-------------------------------------
2524 2534 // Do we Match on this edge index or not? Match only memory & value
2525 2535 uint StoreNode::match_edge(uint idx) const {
2526 2536 return idx == MemNode::Address || idx == MemNode::ValueIn;
2527 2537 }
2528 2538
2529 2539 //------------------------------cmp--------------------------------------------
2530 2540 // Do not common stores up together. They generally have to be split
2531 2541 // back up anyways, so do not bother.
2532 2542 uint StoreNode::cmp( const Node &n ) const {
2533 2543 return (&n == this); // Always fail except on self
2534 2544 }
2535 2545
2536 2546 //------------------------------Ideal_masked_input-----------------------------
2537 2547 // Check for a useless mask before a partial-word store
2538 2548 // (StoreB ... (AndI valIn conIa) )
2539 2549 // If (conIa & mask == mask) this simplifies to
2540 2550 // (StoreB ... (valIn) )
2541 2551 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
2542 2552 Node *val = in(MemNode::ValueIn);
2543 2553 if( val->Opcode() == Op_AndI ) {
2544 2554 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2545 2555 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
2546 2556 set_req(MemNode::ValueIn, val->in(1));
2547 2557 return this;
2548 2558 }
2549 2559 }
2550 2560 return NULL;
2551 2561 }
2552 2562
2553 2563
2554 2564 //------------------------------Ideal_sign_extended_input----------------------
2555 2565 // Check for useless sign-extension before a partial-word store
2556 2566 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
2557 2567 // If (conIL == conIR && conIR <= num_bits) this simplifies to
2558 2568 // (StoreB ... (valIn) )
2559 2569 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
2560 2570 Node *val = in(MemNode::ValueIn);
2561 2571 if( val->Opcode() == Op_RShiftI ) {
2562 2572 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2563 2573 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
2564 2574 Node *shl = val->in(1);
2565 2575 if( shl->Opcode() == Op_LShiftI ) {
2566 2576 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
2567 2577 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
2568 2578 set_req(MemNode::ValueIn, shl->in(1));
2569 2579 return this;
2570 2580 }
2571 2581 }
2572 2582 }
2573 2583 }
2574 2584 return NULL;
2575 2585 }
2576 2586
2577 2587 //------------------------------value_never_loaded-----------------------------------
2578 2588 // Determine whether there are any possible loads of the value stored.
2579 2589 // For simplicity, we actually check if there are any loads from the
2580 2590 // address stored to, not just for loads of the value stored by this node.
2581 2591 //
2582 2592 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
2583 2593 Node *adr = in(Address);
2584 2594 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
2585 2595 if (adr_oop == NULL)
2586 2596 return false;
2587 2597 if (!adr_oop->is_known_instance_field())
2588 2598 return false; // if not a distinct instance, there may be aliases of the address
2589 2599 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
2590 2600 Node *use = adr->fast_out(i);
2591 2601 int opc = use->Opcode();
2592 2602 if (use->is_Load() || use->is_LoadStore()) {
2593 2603 return false;
2594 2604 }
2595 2605 }
2596 2606 return true;
2597 2607 }
2598 2608
2599 2609 //=============================================================================
2600 2610 //------------------------------Ideal------------------------------------------
2601 2611 // If the store is from an AND mask that leaves the low bits untouched, then
2602 2612 // we can skip the AND operation. If the store is from a sign-extension
2603 2613 // (a left shift, then right shift) we can skip both.
2604 2614 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
2605 2615 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
2606 2616 if( progress != NULL ) return progress;
2607 2617
2608 2618 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
2609 2619 if( progress != NULL ) return progress;
2610 2620
2611 2621 // Finally check the default case
2612 2622 return StoreNode::Ideal(phase, can_reshape);
2613 2623 }
2614 2624
2615 2625 //=============================================================================
2616 2626 //------------------------------Ideal------------------------------------------
2617 2627 // If the store is from an AND mask that leaves the low bits untouched, then
2618 2628 // we can skip the AND operation
2619 2629 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
2620 2630 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
2621 2631 if( progress != NULL ) return progress;
2622 2632
2623 2633 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
2624 2634 if( progress != NULL ) return progress;
2625 2635
2626 2636 // Finally check the default case
2627 2637 return StoreNode::Ideal(phase, can_reshape);
2628 2638 }
2629 2639
2630 2640 //=============================================================================
2631 2641 //------------------------------Identity---------------------------------------
2632 2642 Node *StoreCMNode::Identity( PhaseTransform *phase ) {
2633 2643 // No need to card mark when storing a null ptr
2634 2644 Node* my_store = in(MemNode::OopStore);
2635 2645 if (my_store->is_Store()) {
2636 2646 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
2637 2647 if( t1 == TypePtr::NULL_PTR ) {
2638 2648 return in(MemNode::Memory);
2639 2649 }
2640 2650 }
2641 2651 return this;
2642 2652 }
2643 2653
2644 2654 //=============================================================================
2645 2655 //------------------------------Ideal---------------------------------------
2646 2656 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){
2647 2657 Node* progress = StoreNode::Ideal(phase, can_reshape);
2648 2658 if (progress != NULL) return progress;
2649 2659
2650 2660 Node* my_store = in(MemNode::OopStore);
2651 2661 if (my_store->is_MergeMem()) {
2652 2662 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx());
2653 2663 set_req(MemNode::OopStore, mem);
2654 2664 return this;
2655 2665 }
2656 2666
2657 2667 return NULL;
2658 2668 }
2659 2669
2660 2670 //------------------------------Value-----------------------------------------
2661 2671 const Type *StoreCMNode::Value( PhaseTransform *phase ) const {
2662 2672 // Either input is TOP ==> the result is TOP
2663 2673 const Type *t = phase->type( in(MemNode::Memory) );
2664 2674 if( t == Type::TOP ) return Type::TOP;
2665 2675 t = phase->type( in(MemNode::Address) );
2666 2676 if( t == Type::TOP ) return Type::TOP;
2667 2677 t = phase->type( in(MemNode::ValueIn) );
2668 2678 if( t == Type::TOP ) return Type::TOP;
2669 2679 // If extra input is TOP ==> the result is TOP
2670 2680 t = phase->type( in(MemNode::OopStore) );
2671 2681 if( t == Type::TOP ) return Type::TOP;
2672 2682
2673 2683 return StoreNode::Value( phase );
2674 2684 }
2675 2685
2676 2686
2677 2687 //=============================================================================
2678 2688 //----------------------------------SCMemProjNode------------------------------
2679 2689 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const
2680 2690 {
2681 2691 return bottom_type();
2682 2692 }
2683 2693
2684 2694 //=============================================================================
2685 2695 //----------------------------------LoadStoreNode------------------------------
2686 2696 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
2687 2697 : Node(required),
2688 2698 _type(rt),
2689 2699 _adr_type(at)
2690 2700 {
2691 2701 init_req(MemNode::Control, c );
2692 2702 init_req(MemNode::Memory , mem);
2693 2703 init_req(MemNode::Address, adr);
2694 2704 init_req(MemNode::ValueIn, val);
2695 2705 init_class_id(Class_LoadStore);
2696 2706 }
2697 2707
2698 2708 uint LoadStoreNode::ideal_reg() const {
2699 2709 return _type->ideal_reg();
2700 2710 }
2701 2711
2702 2712 bool LoadStoreNode::result_not_used() const {
2703 2713 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) {
2704 2714 Node *x = fast_out(i);
2705 2715 if (x->Opcode() == Op_SCMemProj) continue;
2706 2716 return false;
2707 2717 }
2708 2718 return true;
2709 2719 }
2710 2720
2711 2721 uint LoadStoreNode::size_of() const { return sizeof(*this); }
2712 2722
2713 2723 //=============================================================================
2714 2724 //----------------------------------LoadStoreConditionalNode--------------------
2715 2725 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) {
2716 2726 init_req(ExpectedIn, ex );
2717 2727 }
2718 2728
2719 2729 //=============================================================================
2720 2730 //-------------------------------adr_type--------------------------------------
2721 2731 // Do we Match on this edge index or not? Do not match memory
2722 2732 const TypePtr* ClearArrayNode::adr_type() const {
2723 2733 Node *adr = in(3);
2724 2734 return MemNode::calculate_adr_type(adr->bottom_type());
2725 2735 }
2726 2736
2727 2737 //------------------------------match_edge-------------------------------------
2728 2738 // Do we Match on this edge index or not? Do not match memory
2729 2739 uint ClearArrayNode::match_edge(uint idx) const {
2730 2740 return idx > 1;
2731 2741 }
2732 2742
2733 2743 //------------------------------Identity---------------------------------------
2734 2744 // Clearing a zero length array does nothing
2735 2745 Node *ClearArrayNode::Identity( PhaseTransform *phase ) {
2736 2746 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
2737 2747 }
2738 2748
2739 2749 //------------------------------Idealize---------------------------------------
2740 2750 // Clearing a short array is faster with stores
2741 2751 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){
2742 2752 const int unit = BytesPerLong;
2743 2753 const TypeX* t = phase->type(in(2))->isa_intptr_t();
2744 2754 if (!t) return NULL;
2745 2755 if (!t->is_con()) return NULL;
2746 2756 intptr_t raw_count = t->get_con();
2747 2757 intptr_t size = raw_count;
2748 2758 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
2749 2759 // Clearing nothing uses the Identity call.
2750 2760 // Negative clears are possible on dead ClearArrays
2751 2761 // (see jck test stmt114.stmt11402.val).
2752 2762 if (size <= 0 || size % unit != 0) return NULL;
2753 2763 intptr_t count = size / unit;
2754 2764 // Length too long; use fast hardware clear
2755 2765 if (size > Matcher::init_array_short_size) return NULL;
2756 2766 Node *mem = in(1);
2757 2767 if( phase->type(mem)==Type::TOP ) return NULL;
2758 2768 Node *adr = in(3);
2759 2769 const Type* at = phase->type(adr);
2760 2770 if( at==Type::TOP ) return NULL;
2761 2771 const TypePtr* atp = at->isa_ptr();
2762 2772 // adjust atp to be the correct array element address type
2763 2773 if (atp == NULL) atp = TypePtr::BOTTOM;
2764 2774 else atp = atp->add_offset(Type::OffsetBot);
2765 2775 // Get base for derived pointer purposes
2766 2776 if( adr->Opcode() != Op_AddP ) Unimplemented();
2767 2777 Node *base = adr->in(1);
2768 2778
2769 2779 Node *zero = phase->makecon(TypeLong::ZERO);
2770 2780 Node *off = phase->MakeConX(BytesPerLong);
2771 2781 mem = new (phase->C) StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
2772 2782 count--;
2773 2783 while( count-- ) {
2774 2784 mem = phase->transform(mem);
2775 2785 adr = phase->transform(new (phase->C) AddPNode(base,adr,off));
2776 2786 mem = new (phase->C) StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
2777 2787 }
2778 2788 return mem;
2779 2789 }
2780 2790
2781 2791 //----------------------------step_through----------------------------------
2782 2792 // Return allocation input memory edge if it is different instance
2783 2793 // or itself if it is the one we are looking for.
2784 2794 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) {
2785 2795 Node* n = *np;
2786 2796 assert(n->is_ClearArray(), "sanity");
2787 2797 intptr_t offset;
2788 2798 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
2789 2799 // This method is called only before Allocate nodes are expanded during
2790 2800 // macro nodes expansion. Before that ClearArray nodes are only generated
2791 2801 // in LibraryCallKit::generate_arraycopy() which follows allocations.
2792 2802 assert(alloc != NULL, "should have allocation");
2793 2803 if (alloc->_idx == instance_id) {
2794 2804 // Can not bypass initialization of the instance we are looking for.
2795 2805 return false;
2796 2806 }
2797 2807 // Otherwise skip it.
2798 2808 InitializeNode* init = alloc->initialization();
2799 2809 if (init != NULL)
2800 2810 *np = init->in(TypeFunc::Memory);
2801 2811 else
2802 2812 *np = alloc->in(TypeFunc::Memory);
2803 2813 return true;
2804 2814 }
2805 2815
2806 2816 //----------------------------clear_memory-------------------------------------
2807 2817 // Generate code to initialize object storage to zero.
2808 2818 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2809 2819 intptr_t start_offset,
2810 2820 Node* end_offset,
2811 2821 PhaseGVN* phase) {
2812 2822 Compile* C = phase->C;
2813 2823 intptr_t offset = start_offset;
2814 2824
2815 2825 int unit = BytesPerLong;
2816 2826 if ((offset % unit) != 0) {
2817 2827 Node* adr = new (C) AddPNode(dest, dest, phase->MakeConX(offset));
2818 2828 adr = phase->transform(adr);
2819 2829 const TypePtr* atp = TypeRawPtr::BOTTOM;
2820 2830 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
2821 2831 mem = phase->transform(mem);
2822 2832 offset += BytesPerInt;
2823 2833 }
2824 2834 assert((offset % unit) == 0, "");
2825 2835
2826 2836 // Initialize the remaining stuff, if any, with a ClearArray.
2827 2837 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
2828 2838 }
2829 2839
2830 2840 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2831 2841 Node* start_offset,
2832 2842 Node* end_offset,
2833 2843 PhaseGVN* phase) {
2834 2844 if (start_offset == end_offset) {
2835 2845 // nothing to do
2836 2846 return mem;
2837 2847 }
2838 2848
2839 2849 Compile* C = phase->C;
2840 2850 int unit = BytesPerLong;
2841 2851 Node* zbase = start_offset;
2842 2852 Node* zend = end_offset;
2843 2853
2844 2854 // Scale to the unit required by the CPU:
2845 2855 if (!Matcher::init_array_count_is_in_bytes) {
2846 2856 Node* shift = phase->intcon(exact_log2(unit));
2847 2857 zbase = phase->transform( new(C) URShiftXNode(zbase, shift) );
2848 2858 zend = phase->transform( new(C) URShiftXNode(zend, shift) );
2849 2859 }
2850 2860
2851 2861 // Bulk clear double-words
2852 2862 Node* zsize = phase->transform( new(C) SubXNode(zend, zbase) );
2853 2863 Node* adr = phase->transform( new(C) AddPNode(dest, dest, start_offset) );
2854 2864 mem = new (C) ClearArrayNode(ctl, mem, zsize, adr);
2855 2865 return phase->transform(mem);
2856 2866 }
2857 2867
2858 2868 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2859 2869 intptr_t start_offset,
2860 2870 intptr_t end_offset,
2861 2871 PhaseGVN* phase) {
2862 2872 if (start_offset == end_offset) {
2863 2873 // nothing to do
2864 2874 return mem;
2865 2875 }
2866 2876
2867 2877 Compile* C = phase->C;
2868 2878 assert((end_offset % BytesPerInt) == 0, "odd end offset");
2869 2879 intptr_t done_offset = end_offset;
2870 2880 if ((done_offset % BytesPerLong) != 0) {
2871 2881 done_offset -= BytesPerInt;
2872 2882 }
2873 2883 if (done_offset > start_offset) {
2874 2884 mem = clear_memory(ctl, mem, dest,
2875 2885 start_offset, phase->MakeConX(done_offset), phase);
2876 2886 }
2877 2887 if (done_offset < end_offset) { // emit the final 32-bit store
2878 2888 Node* adr = new (C) AddPNode(dest, dest, phase->MakeConX(done_offset));
2879 2889 adr = phase->transform(adr);
2880 2890 const TypePtr* atp = TypeRawPtr::BOTTOM;
2881 2891 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
2882 2892 mem = phase->transform(mem);
2883 2893 done_offset += BytesPerInt;
2884 2894 }
2885 2895 assert(done_offset == end_offset, "");
2886 2896 return mem;
2887 2897 }
2888 2898
2889 2899 //=============================================================================
2890 2900 // Do not match memory edge.
2891 2901 uint StrIntrinsicNode::match_edge(uint idx) const {
2892 2902 return idx == 2 || idx == 3;
2893 2903 }
2894 2904
2895 2905 //------------------------------Ideal------------------------------------------
2896 2906 // Return a node which is more "ideal" than the current node. Strip out
2897 2907 // control copies
2898 2908 Node *StrIntrinsicNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2899 2909 if (remove_dead_region(phase, can_reshape)) return this;
2900 2910 // Don't bother trying to transform a dead node
2901 2911 if (in(0) && in(0)->is_top()) return NULL;
2902 2912
2903 2913 if (can_reshape) {
2904 2914 Node* mem = phase->transform(in(MemNode::Memory));
2905 2915 // If transformed to a MergeMem, get the desired slice
2906 2916 uint alias_idx = phase->C->get_alias_index(adr_type());
2907 2917 mem = mem->is_MergeMem() ? mem->as_MergeMem()->memory_at(alias_idx) : mem;
2908 2918 if (mem != in(MemNode::Memory)) {
2909 2919 set_req(MemNode::Memory, mem);
2910 2920 return this;
2911 2921 }
2912 2922 }
2913 2923 return NULL;
2914 2924 }
2915 2925
2916 2926 //------------------------------Value------------------------------------------
2917 2927 const Type *StrIntrinsicNode::Value( PhaseTransform *phase ) const {
2918 2928 if (in(0) && phase->type(in(0)) == Type::TOP) return Type::TOP;
2919 2929 return bottom_type();
2920 2930 }
2921 2931
2922 2932 //=============================================================================
2923 2933 //------------------------------match_edge-------------------------------------
2924 2934 // Do not match memory edge
2925 2935 uint EncodeISOArrayNode::match_edge(uint idx) const {
2926 2936 return idx == 2 || idx == 3; // EncodeISOArray src (Binary dst len)
2927 2937 }
2928 2938
2929 2939 //------------------------------Ideal------------------------------------------
2930 2940 // Return a node which is more "ideal" than the current node. Strip out
2931 2941 // control copies
2932 2942 Node *EncodeISOArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2933 2943 return remove_dead_region(phase, can_reshape) ? this : NULL;
2934 2944 }
2935 2945
2936 2946 //------------------------------Value------------------------------------------
2937 2947 const Type *EncodeISOArrayNode::Value(PhaseTransform *phase) const {
2938 2948 if (in(0) && phase->type(in(0)) == Type::TOP) return Type::TOP;
2939 2949 return bottom_type();
2940 2950 }
2941 2951
2942 2952 //=============================================================================
2943 2953 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
2944 2954 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
2945 2955 _adr_type(C->get_adr_type(alias_idx))
2946 2956 {
2947 2957 init_class_id(Class_MemBar);
2948 2958 Node* top = C->top();
2949 2959 init_req(TypeFunc::I_O,top);
2950 2960 init_req(TypeFunc::FramePtr,top);
2951 2961 init_req(TypeFunc::ReturnAdr,top);
2952 2962 if (precedent != NULL)
2953 2963 init_req(TypeFunc::Parms, precedent);
2954 2964 }
2955 2965
2956 2966 //------------------------------cmp--------------------------------------------
2957 2967 uint MemBarNode::hash() const { return NO_HASH; }
2958 2968 uint MemBarNode::cmp( const Node &n ) const {
2959 2969 return (&n == this); // Always fail except on self
2960 2970 }
2961 2971
2962 2972 //------------------------------make-------------------------------------------
2963 2973 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
2964 2974 switch (opcode) {
2965 2975 case Op_MemBarAcquire: return new(C) MemBarAcquireNode(C, atp, pn);
2966 2976 case Op_LoadFence: return new(C) LoadFenceNode(C, atp, pn);
2967 2977 case Op_MemBarRelease: return new(C) MemBarReleaseNode(C, atp, pn);
2968 2978 case Op_StoreFence: return new(C) StoreFenceNode(C, atp, pn);
2969 2979 case Op_MemBarAcquireLock: return new(C) MemBarAcquireLockNode(C, atp, pn);
2970 2980 case Op_MemBarReleaseLock: return new(C) MemBarReleaseLockNode(C, atp, pn);
2971 2981 case Op_MemBarVolatile: return new(C) MemBarVolatileNode(C, atp, pn);
2972 2982 case Op_MemBarCPUOrder: return new(C) MemBarCPUOrderNode(C, atp, pn);
2973 2983 case Op_Initialize: return new(C) InitializeNode(C, atp, pn);
2974 2984 case Op_MemBarStoreStore: return new(C) MemBarStoreStoreNode(C, atp, pn);
2975 2985 default: ShouldNotReachHere(); return NULL;
2976 2986 }
2977 2987 }
2978 2988
2979 2989 //------------------------------Ideal------------------------------------------
2980 2990 // Return a node which is more "ideal" than the current node. Strip out
2981 2991 // control copies
2982 2992 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2983 2993 if (remove_dead_region(phase, can_reshape)) return this;
2984 2994 // Don't bother trying to transform a dead node
2985 2995 if (in(0) && in(0)->is_top()) {
2986 2996 return NULL;
2987 2997 }
2988 2998
2989 2999 // Eliminate volatile MemBars for scalar replaced objects.
2990 3000 if (can_reshape && req() == (Precedent+1)) {
2991 3001 bool eliminate = false;
2992 3002 int opc = Opcode();
2993 3003 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) {
2994 3004 // Volatile field loads and stores.
2995 3005 Node* my_mem = in(MemBarNode::Precedent);
2996 3006 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge
2997 3007 if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) {
2998 3008 // if the Precedent is a decodeN and its input (a Load) is used at more than one place,
2999 3009 // replace this Precedent (decodeN) with the Load instead.
3000 3010 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) {
3001 3011 Node* load_node = my_mem->in(1);
3002 3012 set_req(MemBarNode::Precedent, load_node);
3003 3013 phase->is_IterGVN()->_worklist.push(my_mem);
3004 3014 my_mem = load_node;
3005 3015 } else {
3006 3016 assert(my_mem->unique_out() == this, "sanity");
3007 3017 del_req(Precedent);
3008 3018 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later
3009 3019 my_mem = NULL;
3010 3020 }
3011 3021 }
3012 3022 if (my_mem != NULL && my_mem->is_Mem()) {
3013 3023 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
3014 3024 // Check for scalar replaced object reference.
3015 3025 if( t_oop != NULL && t_oop->is_known_instance_field() &&
3016 3026 t_oop->offset() != Type::OffsetBot &&
3017 3027 t_oop->offset() != Type::OffsetTop) {
3018 3028 eliminate = true;
3019 3029 }
3020 3030 }
3021 3031 } else if (opc == Op_MemBarRelease) {
3022 3032 // Final field stores.
3023 3033 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase);
3024 3034 if ((alloc != NULL) && alloc->is_Allocate() &&
3025 3035 alloc->as_Allocate()->_is_non_escaping) {
3026 3036 // The allocated object does not escape.
3027 3037 eliminate = true;
3028 3038 }
3029 3039 }
3030 3040 if (eliminate) {
3031 3041 // Replace MemBar projections by its inputs.
3032 3042 PhaseIterGVN* igvn = phase->is_IterGVN();
3033 3043 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
3034 3044 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
3035 3045 // Must return either the original node (now dead) or a new node
3036 3046 // (Do not return a top here, since that would break the uniqueness of top.)
3037 3047 return new (phase->C) ConINode(TypeInt::ZERO);
3038 3048 }
3039 3049 }
3040 3050 return NULL;
3041 3051 }
3042 3052
3043 3053 //------------------------------Value------------------------------------------
3044 3054 const Type *MemBarNode::Value( PhaseTransform *phase ) const {
3045 3055 if( !in(0) ) return Type::TOP;
3046 3056 if( phase->type(in(0)) == Type::TOP )
3047 3057 return Type::TOP;
3048 3058 return TypeTuple::MEMBAR;
3049 3059 }
3050 3060
3051 3061 //------------------------------match------------------------------------------
3052 3062 // Construct projections for memory.
3053 3063 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
3054 3064 switch (proj->_con) {
3055 3065 case TypeFunc::Control:
3056 3066 case TypeFunc::Memory:
3057 3067 return new (m->C) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
3058 3068 }
3059 3069 ShouldNotReachHere();
3060 3070 return NULL;
3061 3071 }
3062 3072
3063 3073 //===========================InitializeNode====================================
3064 3074 // SUMMARY:
3065 3075 // This node acts as a memory barrier on raw memory, after some raw stores.
3066 3076 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
3067 3077 // The Initialize can 'capture' suitably constrained stores as raw inits.
3068 3078 // It can coalesce related raw stores into larger units (called 'tiles').
3069 3079 // It can avoid zeroing new storage for memory units which have raw inits.
3070 3080 // At macro-expansion, it is marked 'complete', and does not optimize further.
3071 3081 //
3072 3082 // EXAMPLE:
3073 3083 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
3074 3084 // ctl = incoming control; mem* = incoming memory
3075 3085 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
3076 3086 // First allocate uninitialized memory and fill in the header:
3077 3087 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
3078 3088 // ctl := alloc.Control; mem* := alloc.Memory*
3079 3089 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
3080 3090 // Then initialize to zero the non-header parts of the raw memory block:
3081 3091 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
3082 3092 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
3083 3093 // After the initialize node executes, the object is ready for service:
3084 3094 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
3085 3095 // Suppose its body is immediately initialized as {1,2}:
3086 3096 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3087 3097 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3088 3098 // mem.SLICE(#short[*]) := store2
3089 3099 //
3090 3100 // DETAILS:
3091 3101 // An InitializeNode collects and isolates object initialization after
3092 3102 // an AllocateNode and before the next possible safepoint. As a
3093 3103 // memory barrier (MemBarNode), it keeps critical stores from drifting
3094 3104 // down past any safepoint or any publication of the allocation.
3095 3105 // Before this barrier, a newly-allocated object may have uninitialized bits.
3096 3106 // After this barrier, it may be treated as a real oop, and GC is allowed.
3097 3107 //
3098 3108 // The semantics of the InitializeNode include an implicit zeroing of
3099 3109 // the new object from object header to the end of the object.
3100 3110 // (The object header and end are determined by the AllocateNode.)
3101 3111 //
3102 3112 // Certain stores may be added as direct inputs to the InitializeNode.
3103 3113 // These stores must update raw memory, and they must be to addresses
3104 3114 // derived from the raw address produced by AllocateNode, and with
3105 3115 // a constant offset. They must be ordered by increasing offset.
3106 3116 // The first one is at in(RawStores), the last at in(req()-1).
3107 3117 // Unlike most memory operations, they are not linked in a chain,
3108 3118 // but are displayed in parallel as users of the rawmem output of
3109 3119 // the allocation.
3110 3120 //
3111 3121 // (See comments in InitializeNode::capture_store, which continue
3112 3122 // the example given above.)
3113 3123 //
3114 3124 // When the associated Allocate is macro-expanded, the InitializeNode
3115 3125 // may be rewritten to optimize collected stores. A ClearArrayNode
3116 3126 // may also be created at that point to represent any required zeroing.
3117 3127 // The InitializeNode is then marked 'complete', prohibiting further
3118 3128 // capturing of nearby memory operations.
3119 3129 //
3120 3130 // During macro-expansion, all captured initializations which store
3121 3131 // constant values of 32 bits or smaller are coalesced (if advantageous)
3122 3132 // into larger 'tiles' 32 or 64 bits. This allows an object to be
3123 3133 // initialized in fewer memory operations. Memory words which are
3124 3134 // covered by neither tiles nor non-constant stores are pre-zeroed
3125 3135 // by explicit stores of zero. (The code shape happens to do all
3126 3136 // zeroing first, then all other stores, with both sequences occurring
3127 3137 // in order of ascending offsets.)
3128 3138 //
3129 3139 // Alternatively, code may be inserted between an AllocateNode and its
3130 3140 // InitializeNode, to perform arbitrary initialization of the new object.
3131 3141 // E.g., the object copying intrinsics insert complex data transfers here.
3132 3142 // The initialization must then be marked as 'complete' disable the
3133 3143 // built-in zeroing semantics and the collection of initializing stores.
3134 3144 //
3135 3145 // While an InitializeNode is incomplete, reads from the memory state
3136 3146 // produced by it are optimizable if they match the control edge and
3137 3147 // new oop address associated with the allocation/initialization.
3138 3148 // They return a stored value (if the offset matches) or else zero.
3139 3149 // A write to the memory state, if it matches control and address,
3140 3150 // and if it is to a constant offset, may be 'captured' by the
3141 3151 // InitializeNode. It is cloned as a raw memory operation and rewired
3142 3152 // inside the initialization, to the raw oop produced by the allocation.
3143 3153 // Operations on addresses which are provably distinct (e.g., to
3144 3154 // other AllocateNodes) are allowed to bypass the initialization.
3145 3155 //
3146 3156 // The effect of all this is to consolidate object initialization
3147 3157 // (both arrays and non-arrays, both piecewise and bulk) into a
3148 3158 // single location, where it can be optimized as a unit.
3149 3159 //
3150 3160 // Only stores with an offset less than TrackedInitializationLimit words
3151 3161 // will be considered for capture by an InitializeNode. This puts a
3152 3162 // reasonable limit on the complexity of optimized initializations.
3153 3163
3154 3164 //---------------------------InitializeNode------------------------------------
3155 3165 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
3156 3166 : _is_complete(Incomplete), _does_not_escape(false),
3157 3167 MemBarNode(C, adr_type, rawoop)
3158 3168 {
3159 3169 init_class_id(Class_Initialize);
3160 3170
3161 3171 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
3162 3172 assert(in(RawAddress) == rawoop, "proper init");
3163 3173 // Note: allocation() can be NULL, for secondary initialization barriers
3164 3174 }
3165 3175
3166 3176 // Since this node is not matched, it will be processed by the
3167 3177 // register allocator. Declare that there are no constraints
3168 3178 // on the allocation of the RawAddress edge.
3169 3179 const RegMask &InitializeNode::in_RegMask(uint idx) const {
3170 3180 // This edge should be set to top, by the set_complete. But be conservative.
3171 3181 if (idx == InitializeNode::RawAddress)
3172 3182 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
3173 3183 return RegMask::Empty;
3174 3184 }
3175 3185
3176 3186 Node* InitializeNode::memory(uint alias_idx) {
3177 3187 Node* mem = in(Memory);
3178 3188 if (mem->is_MergeMem()) {
3179 3189 return mem->as_MergeMem()->memory_at(alias_idx);
3180 3190 } else {
3181 3191 // incoming raw memory is not split
3182 3192 return mem;
3183 3193 }
3184 3194 }
3185 3195
3186 3196 bool InitializeNode::is_non_zero() {
3187 3197 if (is_complete()) return false;
3188 3198 remove_extra_zeroes();
3189 3199 return (req() > RawStores);
3190 3200 }
3191 3201
3192 3202 void InitializeNode::set_complete(PhaseGVN* phase) {
3193 3203 assert(!is_complete(), "caller responsibility");
3194 3204 _is_complete = Complete;
3195 3205
3196 3206 // After this node is complete, it contains a bunch of
3197 3207 // raw-memory initializations. There is no need for
3198 3208 // it to have anything to do with non-raw memory effects.
3199 3209 // Therefore, tell all non-raw users to re-optimize themselves,
3200 3210 // after skipping the memory effects of this initialization.
3201 3211 PhaseIterGVN* igvn = phase->is_IterGVN();
3202 3212 if (igvn) igvn->add_users_to_worklist(this);
3203 3213 }
3204 3214
3205 3215 // convenience function
3206 3216 // return false if the init contains any stores already
3207 3217 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
3208 3218 InitializeNode* init = initialization();
3209 3219 if (init == NULL || init->is_complete()) return false;
3210 3220 init->remove_extra_zeroes();
3211 3221 // for now, if this allocation has already collected any inits, bail:
3212 3222 if (init->is_non_zero()) return false;
3213 3223 init->set_complete(phase);
3214 3224 return true;
3215 3225 }
3216 3226
3217 3227 void InitializeNode::remove_extra_zeroes() {
3218 3228 if (req() == RawStores) return;
3219 3229 Node* zmem = zero_memory();
3220 3230 uint fill = RawStores;
3221 3231 for (uint i = fill; i < req(); i++) {
3222 3232 Node* n = in(i);
3223 3233 if (n->is_top() || n == zmem) continue; // skip
3224 3234 if (fill < i) set_req(fill, n); // compact
3225 3235 ++fill;
3226 3236 }
3227 3237 // delete any empty spaces created:
3228 3238 while (fill < req()) {
3229 3239 del_req(fill);
3230 3240 }
3231 3241 }
3232 3242
3233 3243 // Helper for remembering which stores go with which offsets.
3234 3244 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
3235 3245 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
3236 3246 intptr_t offset = -1;
3237 3247 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
3238 3248 phase, offset);
3239 3249 if (base == NULL) return -1; // something is dead,
3240 3250 if (offset < 0) return -1; // dead, dead
3241 3251 return offset;
3242 3252 }
3243 3253
3244 3254 // Helper for proving that an initialization expression is
3245 3255 // "simple enough" to be folded into an object initialization.
3246 3256 // Attempts to prove that a store's initial value 'n' can be captured
3247 3257 // within the initialization without creating a vicious cycle, such as:
3248 3258 // { Foo p = new Foo(); p.next = p; }
3249 3259 // True for constants and parameters and small combinations thereof.
3250 3260 bool InitializeNode::detect_init_independence(Node* n, int& count) {
3251 3261 if (n == NULL) return true; // (can this really happen?)
3252 3262 if (n->is_Proj()) n = n->in(0);
3253 3263 if (n == this) return false; // found a cycle
3254 3264 if (n->is_Con()) return true;
3255 3265 if (n->is_Start()) return true; // params, etc., are OK
3256 3266 if (n->is_Root()) return true; // even better
3257 3267
3258 3268 Node* ctl = n->in(0);
3259 3269 if (ctl != NULL && !ctl->is_top()) {
3260 3270 if (ctl->is_Proj()) ctl = ctl->in(0);
3261 3271 if (ctl == this) return false;
3262 3272
3263 3273 // If we already know that the enclosing memory op is pinned right after
3264 3274 // the init, then any control flow that the store has picked up
3265 3275 // must have preceded the init, or else be equal to the init.
3266 3276 // Even after loop optimizations (which might change control edges)
3267 3277 // a store is never pinned *before* the availability of its inputs.
3268 3278 if (!MemNode::all_controls_dominate(n, this))
3269 3279 return false; // failed to prove a good control
3270 3280 }
3271 3281
3272 3282 // Check data edges for possible dependencies on 'this'.
3273 3283 if ((count += 1) > 20) return false; // complexity limit
3274 3284 for (uint i = 1; i < n->req(); i++) {
3275 3285 Node* m = n->in(i);
3276 3286 if (m == NULL || m == n || m->is_top()) continue;
3277 3287 uint first_i = n->find_edge(m);
3278 3288 if (i != first_i) continue; // process duplicate edge just once
3279 3289 if (!detect_init_independence(m, count)) {
3280 3290 return false;
3281 3291 }
3282 3292 }
3283 3293
3284 3294 return true;
3285 3295 }
3286 3296
3287 3297 // Here are all the checks a Store must pass before it can be moved into
3288 3298 // an initialization. Returns zero if a check fails.
3289 3299 // On success, returns the (constant) offset to which the store applies,
3290 3300 // within the initialized memory.
3291 3301 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase, bool can_reshape) {
3292 3302 const int FAIL = 0;
3293 3303 if (st->req() != MemNode::ValueIn + 1)
3294 3304 return FAIL; // an inscrutable StoreNode (card mark?)
3295 3305 Node* ctl = st->in(MemNode::Control);
3296 3306 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
3297 3307 return FAIL; // must be unconditional after the initialization
3298 3308 Node* mem = st->in(MemNode::Memory);
3299 3309 if (!(mem->is_Proj() && mem->in(0) == this))
3300 3310 return FAIL; // must not be preceded by other stores
3301 3311 Node* adr = st->in(MemNode::Address);
3302 3312 intptr_t offset;
3303 3313 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
3304 3314 if (alloc == NULL)
3305 3315 return FAIL; // inscrutable address
3306 3316 if (alloc != allocation())
3307 3317 return FAIL; // wrong allocation! (store needs to float up)
3308 3318 Node* val = st->in(MemNode::ValueIn);
3309 3319 int complexity_count = 0;
3310 3320 if (!detect_init_independence(val, complexity_count))
3311 3321 return FAIL; // stored value must be 'simple enough'
3312 3322
3313 3323 // The Store can be captured only if nothing after the allocation
3314 3324 // and before the Store is using the memory location that the store
3315 3325 // overwrites.
3316 3326 bool failed = false;
3317 3327 // If is_complete_with_arraycopy() is true the shape of the graph is
3318 3328 // well defined and is safe so no need for extra checks.
3319 3329 if (!is_complete_with_arraycopy()) {
3320 3330 // We are going to look at each use of the memory state following
3321 3331 // the allocation to make sure nothing reads the memory that the
3322 3332 // Store writes.
3323 3333 const TypePtr* t_adr = phase->type(adr)->isa_ptr();
3324 3334 int alias_idx = phase->C->get_alias_index(t_adr);
3325 3335 ResourceMark rm;
3326 3336 Unique_Node_List mems;
3327 3337 mems.push(mem);
3328 3338 Node* unique_merge = NULL;
3329 3339 for (uint next = 0; next < mems.size(); ++next) {
3330 3340 Node *m = mems.at(next);
3331 3341 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
3332 3342 Node *n = m->fast_out(j);
3333 3343 if (n->outcnt() == 0) {
3334 3344 continue;
3335 3345 }
3336 3346 if (n == st) {
3337 3347 continue;
3338 3348 } else if (n->in(0) != NULL && n->in(0) != ctl) {
3339 3349 // If the control of this use is different from the control
3340 3350 // of the Store which is right after the InitializeNode then
3341 3351 // this node cannot be between the InitializeNode and the
3342 3352 // Store.
3343 3353 continue;
3344 3354 } else if (n->is_MergeMem()) {
3345 3355 if (n->as_MergeMem()->memory_at(alias_idx) == m) {
3346 3356 // We can hit a MergeMemNode (that will likely go away
3347 3357 // later) that is a direct use of the memory state
3348 3358 // following the InitializeNode on the same slice as the
3349 3359 // store node that we'd like to capture. We need to check
3350 3360 // the uses of the MergeMemNode.
3351 3361 mems.push(n);
3352 3362 }
3353 3363 } else if (n->is_Mem()) {
3354 3364 Node* other_adr = n->in(MemNode::Address);
3355 3365 if (other_adr == adr) {
3356 3366 failed = true;
3357 3367 break;
3358 3368 } else {
3359 3369 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
3360 3370 if (other_t_adr != NULL) {
3361 3371 int other_alias_idx = phase->C->get_alias_index(other_t_adr);
3362 3372 if (other_alias_idx == alias_idx) {
3363 3373 // A load from the same memory slice as the store right
3364 3374 // after the InitializeNode. We check the control of the
3365 3375 // object/array that is loaded from. If it's the same as
3366 3376 // the store control then we cannot capture the store.
3367 3377 assert(!n->is_Store(), "2 stores to same slice on same control?");
3368 3378 Node* base = other_adr;
3369 3379 assert(base->is_AddP(), err_msg_res("should be addp but is %s", base->Name()));
3370 3380 base = base->in(AddPNode::Base);
3371 3381 if (base != NULL) {
3372 3382 base = base->uncast();
3373 3383 if (base->is_Proj() && base->in(0) == alloc) {
3374 3384 failed = true;
3375 3385 break;
3376 3386 }
3377 3387 }
3378 3388 }
3379 3389 }
3380 3390 }
3381 3391 } else {
3382 3392 failed = true;
3383 3393 break;
3384 3394 }
3385 3395 }
3386 3396 }
3387 3397 }
3388 3398 if (failed) {
3389 3399 if (!can_reshape) {
3390 3400 // We decided we couldn't capture the store during parsing. We
3391 3401 // should try again during the next IGVN once the graph is
3392 3402 // cleaner.
3393 3403 phase->C->record_for_igvn(st);
3394 3404 }
3395 3405 return FAIL;
3396 3406 }
3397 3407
3398 3408 return offset; // success
3399 3409 }
3400 3410
3401 3411 // Find the captured store in(i) which corresponds to the range
3402 3412 // [start..start+size) in the initialized object.
3403 3413 // If there is one, return its index i. If there isn't, return the
3404 3414 // negative of the index where it should be inserted.
3405 3415 // Return 0 if the queried range overlaps an initialization boundary
3406 3416 // or if dead code is encountered.
3407 3417 // If size_in_bytes is zero, do not bother with overlap checks.
3408 3418 int InitializeNode::captured_store_insertion_point(intptr_t start,
3409 3419 int size_in_bytes,
3410 3420 PhaseTransform* phase) {
3411 3421 const int FAIL = 0, MAX_STORE = BytesPerLong;
3412 3422
3413 3423 if (is_complete())
3414 3424 return FAIL; // arraycopy got here first; punt
3415 3425
3416 3426 assert(allocation() != NULL, "must be present");
3417 3427
3418 3428 // no negatives, no header fields:
3419 3429 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
3420 3430
3421 3431 // after a certain size, we bail out on tracking all the stores:
3422 3432 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3423 3433 if (start >= ti_limit) return FAIL;
3424 3434
3425 3435 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
3426 3436 if (i >= limit) return -(int)i; // not found; here is where to put it
3427 3437
3428 3438 Node* st = in(i);
3429 3439 intptr_t st_off = get_store_offset(st, phase);
3430 3440 if (st_off < 0) {
3431 3441 if (st != zero_memory()) {
3432 3442 return FAIL; // bail out if there is dead garbage
3433 3443 }
3434 3444 } else if (st_off > start) {
3435 3445 // ...we are done, since stores are ordered
3436 3446 if (st_off < start + size_in_bytes) {
3437 3447 return FAIL; // the next store overlaps
3438 3448 }
3439 3449 return -(int)i; // not found; here is where to put it
3440 3450 } else if (st_off < start) {
3441 3451 if (size_in_bytes != 0 &&
3442 3452 start < st_off + MAX_STORE &&
3443 3453 start < st_off + st->as_Store()->memory_size()) {
3444 3454 return FAIL; // the previous store overlaps
3445 3455 }
3446 3456 } else {
3447 3457 if (size_in_bytes != 0 &&
3448 3458 st->as_Store()->memory_size() != size_in_bytes) {
3449 3459 return FAIL; // mismatched store size
3450 3460 }
3451 3461 return i;
3452 3462 }
3453 3463
3454 3464 ++i;
3455 3465 }
3456 3466 }
3457 3467
3458 3468 // Look for a captured store which initializes at the offset 'start'
3459 3469 // with the given size. If there is no such store, and no other
3460 3470 // initialization interferes, then return zero_memory (the memory
3461 3471 // projection of the AllocateNode).
3462 3472 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
3463 3473 PhaseTransform* phase) {
3464 3474 assert(stores_are_sane(phase), "");
3465 3475 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3466 3476 if (i == 0) {
3467 3477 return NULL; // something is dead
3468 3478 } else if (i < 0) {
3469 3479 return zero_memory(); // just primordial zero bits here
3470 3480 } else {
3471 3481 Node* st = in(i); // here is the store at this position
3472 3482 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
3473 3483 return st;
3474 3484 }
3475 3485 }
3476 3486
3477 3487 // Create, as a raw pointer, an address within my new object at 'offset'.
3478 3488 Node* InitializeNode::make_raw_address(intptr_t offset,
3479 3489 PhaseTransform* phase) {
3480 3490 Node* addr = in(RawAddress);
3481 3491 if (offset != 0) {
3482 3492 Compile* C = phase->C;
3483 3493 addr = phase->transform( new (C) AddPNode(C->top(), addr,
3484 3494 phase->MakeConX(offset)) );
3485 3495 }
3486 3496 return addr;
3487 3497 }
3488 3498
3489 3499 // Clone the given store, converting it into a raw store
3490 3500 // initializing a field or element of my new object.
3491 3501 // Caller is responsible for retiring the original store,
3492 3502 // with subsume_node or the like.
3493 3503 //
3494 3504 // From the example above InitializeNode::InitializeNode,
3495 3505 // here are the old stores to be captured:
3496 3506 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3497 3507 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3498 3508 //
3499 3509 // Here is the changed code; note the extra edges on init:
3500 3510 // alloc = (Allocate ...)
3501 3511 // rawoop = alloc.RawAddress
3502 3512 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
3503 3513 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
3504 3514 // init = (Initialize alloc.Control alloc.Memory rawoop
3505 3515 // rawstore1 rawstore2)
3506 3516 //
3507 3517 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
3508 3518 PhaseTransform* phase, bool can_reshape) {
3509 3519 assert(stores_are_sane(phase), "");
3510 3520
3511 3521 if (start < 0) return NULL;
3512 3522 assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
3513 3523
3514 3524 Compile* C = phase->C;
3515 3525 int size_in_bytes = st->memory_size();
3516 3526 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3517 3527 if (i == 0) return NULL; // bail out
3518 3528 Node* prev_mem = NULL; // raw memory for the captured store
3519 3529 if (i > 0) {
3520 3530 prev_mem = in(i); // there is a pre-existing store under this one
3521 3531 set_req(i, C->top()); // temporarily disconnect it
3522 3532 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
3523 3533 } else {
3524 3534 i = -i; // no pre-existing store
3525 3535 prev_mem = zero_memory(); // a slice of the newly allocated object
3526 3536 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
3527 3537 set_req(--i, C->top()); // reuse this edge; it has been folded away
3528 3538 else
3529 3539 ins_req(i, C->top()); // build a new edge
3530 3540 }
3531 3541 Node* new_st = st->clone();
3532 3542 new_st->set_req(MemNode::Control, in(Control));
3533 3543 new_st->set_req(MemNode::Memory, prev_mem);
3534 3544 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
3535 3545 new_st = phase->transform(new_st);
3536 3546
3537 3547 // At this point, new_st might have swallowed a pre-existing store
3538 3548 // at the same offset, or perhaps new_st might have disappeared,
3539 3549 // if it redundantly stored the same value (or zero to fresh memory).
3540 3550
3541 3551 // In any case, wire it in:
3542 3552 set_req(i, new_st);
3543 3553
3544 3554 // The caller may now kill the old guy.
3545 3555 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
3546 3556 assert(check_st == new_st || check_st == NULL, "must be findable");
3547 3557 assert(!is_complete(), "");
3548 3558 return new_st;
3549 3559 }
3550 3560
3551 3561 static bool store_constant(jlong* tiles, int num_tiles,
3552 3562 intptr_t st_off, int st_size,
3553 3563 jlong con) {
3554 3564 if ((st_off & (st_size-1)) != 0)
3555 3565 return false; // strange store offset (assume size==2**N)
3556 3566 address addr = (address)tiles + st_off;
3557 3567 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
3558 3568 switch (st_size) {
3559 3569 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
3560 3570 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
3561 3571 case sizeof(jint): *(jint*) addr = (jint) con; break;
3562 3572 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
3563 3573 default: return false; // strange store size (detect size!=2**N here)
3564 3574 }
3565 3575 return true; // return success to caller
3566 3576 }
3567 3577
3568 3578 // Coalesce subword constants into int constants and possibly
3569 3579 // into long constants. The goal, if the CPU permits,
3570 3580 // is to initialize the object with a small number of 64-bit tiles.
3571 3581 // Also, convert floating-point constants to bit patterns.
3572 3582 // Non-constants are not relevant to this pass.
3573 3583 //
3574 3584 // In terms of the running example on InitializeNode::InitializeNode
3575 3585 // and InitializeNode::capture_store, here is the transformation
3576 3586 // of rawstore1 and rawstore2 into rawstore12:
3577 3587 // alloc = (Allocate ...)
3578 3588 // rawoop = alloc.RawAddress
3579 3589 // tile12 = 0x00010002
3580 3590 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
3581 3591 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
3582 3592 //
3583 3593 void
3584 3594 InitializeNode::coalesce_subword_stores(intptr_t header_size,
3585 3595 Node* size_in_bytes,
3586 3596 PhaseGVN* phase) {
3587 3597 Compile* C = phase->C;
3588 3598
3589 3599 assert(stores_are_sane(phase), "");
3590 3600 // Note: After this pass, they are not completely sane,
3591 3601 // since there may be some overlaps.
3592 3602
3593 3603 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
3594 3604
3595 3605 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3596 3606 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
3597 3607 size_limit = MIN2(size_limit, ti_limit);
3598 3608 size_limit = align_size_up(size_limit, BytesPerLong);
3599 3609 int num_tiles = size_limit / BytesPerLong;
3600 3610
3601 3611 // allocate space for the tile map:
3602 3612 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
3603 3613 jlong tiles_buf[small_len];
3604 3614 Node* nodes_buf[small_len];
3605 3615 jlong inits_buf[small_len];
3606 3616 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
3607 3617 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3608 3618 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
3609 3619 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
3610 3620 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
3611 3621 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3612 3622 // tiles: exact bitwise model of all primitive constants
3613 3623 // nodes: last constant-storing node subsumed into the tiles model
3614 3624 // inits: which bytes (in each tile) are touched by any initializations
3615 3625
3616 3626 //// Pass A: Fill in the tile model with any relevant stores.
3617 3627
3618 3628 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
3619 3629 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
3620 3630 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
3621 3631 Node* zmem = zero_memory(); // initially zero memory state
3622 3632 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3623 3633 Node* st = in(i);
3624 3634 intptr_t st_off = get_store_offset(st, phase);
3625 3635
3626 3636 // Figure out the store's offset and constant value:
3627 3637 if (st_off < header_size) continue; //skip (ignore header)
3628 3638 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
3629 3639 int st_size = st->as_Store()->memory_size();
3630 3640 if (st_off + st_size > size_limit) break;
3631 3641
3632 3642 // Record which bytes are touched, whether by constant or not.
3633 3643 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
3634 3644 continue; // skip (strange store size)
3635 3645
3636 3646 const Type* val = phase->type(st->in(MemNode::ValueIn));
3637 3647 if (!val->singleton()) continue; //skip (non-con store)
3638 3648 BasicType type = val->basic_type();
3639 3649
3640 3650 jlong con = 0;
3641 3651 switch (type) {
3642 3652 case T_INT: con = val->is_int()->get_con(); break;
3643 3653 case T_LONG: con = val->is_long()->get_con(); break;
3644 3654 case T_FLOAT: con = jint_cast(val->getf()); break;
3645 3655 case T_DOUBLE: con = jlong_cast(val->getd()); break;
3646 3656 default: continue; //skip (odd store type)
3647 3657 }
3648 3658
3649 3659 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
3650 3660 st->Opcode() == Op_StoreL) {
3651 3661 continue; // This StoreL is already optimal.
3652 3662 }
3653 3663
3654 3664 // Store down the constant.
3655 3665 store_constant(tiles, num_tiles, st_off, st_size, con);
3656 3666
3657 3667 intptr_t j = st_off >> LogBytesPerLong;
3658 3668
3659 3669 if (type == T_INT && st_size == BytesPerInt
3660 3670 && (st_off & BytesPerInt) == BytesPerInt) {
3661 3671 jlong lcon = tiles[j];
3662 3672 if (!Matcher::isSimpleConstant64(lcon) &&
3663 3673 st->Opcode() == Op_StoreI) {
3664 3674 // This StoreI is already optimal by itself.
3665 3675 jint* intcon = (jint*) &tiles[j];
3666 3676 intcon[1] = 0; // undo the store_constant()
3667 3677
3668 3678 // If the previous store is also optimal by itself, back up and
3669 3679 // undo the action of the previous loop iteration... if we can.
3670 3680 // But if we can't, just let the previous half take care of itself.
3671 3681 st = nodes[j];
3672 3682 st_off -= BytesPerInt;
3673 3683 con = intcon[0];
3674 3684 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
3675 3685 assert(st_off >= header_size, "still ignoring header");
3676 3686 assert(get_store_offset(st, phase) == st_off, "must be");
3677 3687 assert(in(i-1) == zmem, "must be");
3678 3688 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
3679 3689 assert(con == tcon->is_int()->get_con(), "must be");
3680 3690 // Undo the effects of the previous loop trip, which swallowed st:
3681 3691 intcon[0] = 0; // undo store_constant()
3682 3692 set_req(i-1, st); // undo set_req(i, zmem)
3683 3693 nodes[j] = NULL; // undo nodes[j] = st
3684 3694 --old_subword; // undo ++old_subword
3685 3695 }
3686 3696 continue; // This StoreI is already optimal.
3687 3697 }
3688 3698 }
3689 3699
3690 3700 // This store is not needed.
3691 3701 set_req(i, zmem);
3692 3702 nodes[j] = st; // record for the moment
3693 3703 if (st_size < BytesPerLong) // something has changed
3694 3704 ++old_subword; // includes int/float, but who's counting...
3695 3705 else ++old_long;
3696 3706 }
3697 3707
3698 3708 if ((old_subword + old_long) == 0)
3699 3709 return; // nothing more to do
3700 3710
3701 3711 //// Pass B: Convert any non-zero tiles into optimal constant stores.
3702 3712 // Be sure to insert them before overlapping non-constant stores.
3703 3713 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
3704 3714 for (int j = 0; j < num_tiles; j++) {
3705 3715 jlong con = tiles[j];
3706 3716 jlong init = inits[j];
3707 3717 if (con == 0) continue;
3708 3718 jint con0, con1; // split the constant, address-wise
3709 3719 jint init0, init1; // split the init map, address-wise
3710 3720 { union { jlong con; jint intcon[2]; } u;
3711 3721 u.con = con;
3712 3722 con0 = u.intcon[0];
3713 3723 con1 = u.intcon[1];
3714 3724 u.con = init;
3715 3725 init0 = u.intcon[0];
3716 3726 init1 = u.intcon[1];
3717 3727 }
3718 3728
3719 3729 Node* old = nodes[j];
3720 3730 assert(old != NULL, "need the prior store");
3721 3731 intptr_t offset = (j * BytesPerLong);
3722 3732
3723 3733 bool split = !Matcher::isSimpleConstant64(con);
3724 3734
3725 3735 if (offset < header_size) {
3726 3736 assert(offset + BytesPerInt >= header_size, "second int counts");
3727 3737 assert(*(jint*)&tiles[j] == 0, "junk in header");
3728 3738 split = true; // only the second word counts
3729 3739 // Example: int a[] = { 42 ... }
3730 3740 } else if (con0 == 0 && init0 == -1) {
3731 3741 split = true; // first word is covered by full inits
3732 3742 // Example: int a[] = { ... foo(), 42 ... }
3733 3743 } else if (con1 == 0 && init1 == -1) {
3734 3744 split = true; // second word is covered by full inits
3735 3745 // Example: int a[] = { ... 42, foo() ... }
3736 3746 }
3737 3747
3738 3748 // Here's a case where init0 is neither 0 nor -1:
3739 3749 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
3740 3750 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
3741 3751 // In this case the tile is not split; it is (jlong)42.
3742 3752 // The big tile is stored down, and then the foo() value is inserted.
3743 3753 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
3744 3754
3745 3755 Node* ctl = old->in(MemNode::Control);
3746 3756 Node* adr = make_raw_address(offset, phase);
3747 3757 const TypePtr* atp = TypeRawPtr::BOTTOM;
3748 3758
3749 3759 // One or two coalesced stores to plop down.
3750 3760 Node* st[2];
3751 3761 intptr_t off[2];
3752 3762 int nst = 0;
3753 3763 if (!split) {
3754 3764 ++new_long;
3755 3765 off[nst] = offset;
3756 3766 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3757 3767 phase->longcon(con), T_LONG, MemNode::unordered);
3758 3768 } else {
3759 3769 // Omit either if it is a zero.
3760 3770 if (con0 != 0) {
3761 3771 ++new_int;
3762 3772 off[nst] = offset;
3763 3773 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3764 3774 phase->intcon(con0), T_INT, MemNode::unordered);
3765 3775 }
3766 3776 if (con1 != 0) {
3767 3777 ++new_int;
3768 3778 offset += BytesPerInt;
3769 3779 adr = make_raw_address(offset, phase);
3770 3780 off[nst] = offset;
3771 3781 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3772 3782 phase->intcon(con1), T_INT, MemNode::unordered);
3773 3783 }
3774 3784 }
3775 3785
3776 3786 // Insert second store first, then the first before the second.
3777 3787 // Insert each one just before any overlapping non-constant stores.
3778 3788 while (nst > 0) {
3779 3789 Node* st1 = st[--nst];
3780 3790 C->copy_node_notes_to(st1, old);
3781 3791 st1 = phase->transform(st1);
3782 3792 offset = off[nst];
3783 3793 assert(offset >= header_size, "do not smash header");
3784 3794 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
3785 3795 guarantee(ins_idx != 0, "must re-insert constant store");
3786 3796 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
3787 3797 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
3788 3798 set_req(--ins_idx, st1);
3789 3799 else
3790 3800 ins_req(ins_idx, st1);
3791 3801 }
3792 3802 }
3793 3803
3794 3804 if (PrintCompilation && WizardMode)
3795 3805 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
3796 3806 old_subword, old_long, new_int, new_long);
3797 3807 if (C->log() != NULL)
3798 3808 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
3799 3809 old_subword, old_long, new_int, new_long);
3800 3810
3801 3811 // Clean up any remaining occurrences of zmem:
3802 3812 remove_extra_zeroes();
3803 3813 }
3804 3814
3805 3815 // Explore forward from in(start) to find the first fully initialized
3806 3816 // word, and return its offset. Skip groups of subword stores which
3807 3817 // together initialize full words. If in(start) is itself part of a
3808 3818 // fully initialized word, return the offset of in(start). If there
3809 3819 // are no following full-word stores, or if something is fishy, return
3810 3820 // a negative value.
3811 3821 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
3812 3822 int int_map = 0;
3813 3823 intptr_t int_map_off = 0;
3814 3824 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
3815 3825
3816 3826 for (uint i = start, limit = req(); i < limit; i++) {
3817 3827 Node* st = in(i);
3818 3828
3819 3829 intptr_t st_off = get_store_offset(st, phase);
3820 3830 if (st_off < 0) break; // return conservative answer
3821 3831
3822 3832 int st_size = st->as_Store()->memory_size();
3823 3833 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
3824 3834 return st_off; // we found a complete word init
3825 3835 }
3826 3836
3827 3837 // update the map:
3828 3838
3829 3839 intptr_t this_int_off = align_size_down(st_off, BytesPerInt);
3830 3840 if (this_int_off != int_map_off) {
3831 3841 // reset the map:
3832 3842 int_map = 0;
3833 3843 int_map_off = this_int_off;
3834 3844 }
3835 3845
3836 3846 int subword_off = st_off - this_int_off;
3837 3847 int_map |= right_n_bits(st_size) << subword_off;
3838 3848 if ((int_map & FULL_MAP) == FULL_MAP) {
3839 3849 return this_int_off; // we found a complete word init
3840 3850 }
3841 3851
3842 3852 // Did this store hit or cross the word boundary?
3843 3853 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt);
3844 3854 if (next_int_off == this_int_off + BytesPerInt) {
3845 3855 // We passed the current int, without fully initializing it.
3846 3856 int_map_off = next_int_off;
3847 3857 int_map >>= BytesPerInt;
3848 3858 } else if (next_int_off > this_int_off + BytesPerInt) {
3849 3859 // We passed the current and next int.
3850 3860 return this_int_off + BytesPerInt;
3851 3861 }
3852 3862 }
3853 3863
3854 3864 return -1;
3855 3865 }
3856 3866
3857 3867
3858 3868 // Called when the associated AllocateNode is expanded into CFG.
3859 3869 // At this point, we may perform additional optimizations.
3860 3870 // Linearize the stores by ascending offset, to make memory
3861 3871 // activity as coherent as possible.
3862 3872 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
3863 3873 intptr_t header_size,
3864 3874 Node* size_in_bytes,
3865 3875 PhaseGVN* phase) {
3866 3876 assert(!is_complete(), "not already complete");
3867 3877 assert(stores_are_sane(phase), "");
3868 3878 assert(allocation() != NULL, "must be present");
3869 3879
3870 3880 remove_extra_zeroes();
3871 3881
3872 3882 if (ReduceFieldZeroing || ReduceBulkZeroing)
3873 3883 // reduce instruction count for common initialization patterns
3874 3884 coalesce_subword_stores(header_size, size_in_bytes, phase);
3875 3885
3876 3886 Node* zmem = zero_memory(); // initially zero memory state
3877 3887 Node* inits = zmem; // accumulating a linearized chain of inits
3878 3888 #ifdef ASSERT
3879 3889 intptr_t first_offset = allocation()->minimum_header_size();
3880 3890 intptr_t last_init_off = first_offset; // previous init offset
3881 3891 intptr_t last_init_end = first_offset; // previous init offset+size
3882 3892 intptr_t last_tile_end = first_offset; // previous tile offset+size
3883 3893 #endif
3884 3894 intptr_t zeroes_done = header_size;
3885 3895
3886 3896 bool do_zeroing = true; // we might give up if inits are very sparse
3887 3897 int big_init_gaps = 0; // how many large gaps have we seen?
3888 3898
3889 3899 if (ZeroTLAB) do_zeroing = false;
3890 3900 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
3891 3901
3892 3902 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3893 3903 Node* st = in(i);
3894 3904 intptr_t st_off = get_store_offset(st, phase);
3895 3905 if (st_off < 0)
3896 3906 break; // unknown junk in the inits
3897 3907 if (st->in(MemNode::Memory) != zmem)
3898 3908 break; // complicated store chains somehow in list
3899 3909
3900 3910 int st_size = st->as_Store()->memory_size();
3901 3911 intptr_t next_init_off = st_off + st_size;
3902 3912
3903 3913 if (do_zeroing && zeroes_done < next_init_off) {
3904 3914 // See if this store needs a zero before it or under it.
3905 3915 intptr_t zeroes_needed = st_off;
3906 3916
3907 3917 if (st_size < BytesPerInt) {
3908 3918 // Look for subword stores which only partially initialize words.
3909 3919 // If we find some, we must lay down some word-level zeroes first,
3910 3920 // underneath the subword stores.
3911 3921 //
3912 3922 // Examples:
3913 3923 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
3914 3924 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
3915 3925 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
3916 3926 //
3917 3927 // Note: coalesce_subword_stores may have already done this,
3918 3928 // if it was prompted by constant non-zero subword initializers.
3919 3929 // But this case can still arise with non-constant stores.
3920 3930
3921 3931 intptr_t next_full_store = find_next_fullword_store(i, phase);
3922 3932
3923 3933 // In the examples above:
3924 3934 // in(i) p q r s x y z
3925 3935 // st_off 12 13 14 15 12 13 14
3926 3936 // st_size 1 1 1 1 1 1 1
3927 3937 // next_full_s. 12 16 16 16 16 16 16
3928 3938 // z's_done 12 16 16 16 12 16 12
3929 3939 // z's_needed 12 16 16 16 16 16 16
3930 3940 // zsize 0 0 0 0 4 0 4
3931 3941 if (next_full_store < 0) {
3932 3942 // Conservative tack: Zero to end of current word.
3933 3943 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt);
3934 3944 } else {
3935 3945 // Zero to beginning of next fully initialized word.
3936 3946 // Or, don't zero at all, if we are already in that word.
3937 3947 assert(next_full_store >= zeroes_needed, "must go forward");
3938 3948 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
3939 3949 zeroes_needed = next_full_store;
3940 3950 }
3941 3951 }
3942 3952
3943 3953 if (zeroes_needed > zeroes_done) {
3944 3954 intptr_t zsize = zeroes_needed - zeroes_done;
3945 3955 // Do some incremental zeroing on rawmem, in parallel with inits.
3946 3956 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3947 3957 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3948 3958 zeroes_done, zeroes_needed,
3949 3959 phase);
3950 3960 zeroes_done = zeroes_needed;
3951 3961 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2)
3952 3962 do_zeroing = false; // leave the hole, next time
3953 3963 }
3954 3964 }
3955 3965
3956 3966 // Collect the store and move on:
3957 3967 st->set_req(MemNode::Memory, inits);
3958 3968 inits = st; // put it on the linearized chain
3959 3969 set_req(i, zmem); // unhook from previous position
3960 3970
3961 3971 if (zeroes_done == st_off)
3962 3972 zeroes_done = next_init_off;
3963 3973
3964 3974 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
3965 3975
3966 3976 #ifdef ASSERT
3967 3977 // Various order invariants. Weaker than stores_are_sane because
3968 3978 // a large constant tile can be filled in by smaller non-constant stores.
3969 3979 assert(st_off >= last_init_off, "inits do not reverse");
3970 3980 last_init_off = st_off;
3971 3981 const Type* val = NULL;
3972 3982 if (st_size >= BytesPerInt &&
3973 3983 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
3974 3984 (int)val->basic_type() < (int)T_OBJECT) {
3975 3985 assert(st_off >= last_tile_end, "tiles do not overlap");
3976 3986 assert(st_off >= last_init_end, "tiles do not overwrite inits");
3977 3987 last_tile_end = MAX2(last_tile_end, next_init_off);
3978 3988 } else {
3979 3989 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong);
3980 3990 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
3981 3991 assert(st_off >= last_init_end, "inits do not overlap");
3982 3992 last_init_end = next_init_off; // it's a non-tile
3983 3993 }
3984 3994 #endif //ASSERT
3985 3995 }
3986 3996
3987 3997 remove_extra_zeroes(); // clear out all the zmems left over
3988 3998 add_req(inits);
3989 3999
3990 4000 if (!ZeroTLAB) {
3991 4001 // If anything remains to be zeroed, zero it all now.
3992 4002 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3993 4003 // if it is the last unused 4 bytes of an instance, forget about it
3994 4004 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
3995 4005 if (zeroes_done + BytesPerLong >= size_limit) {
3996 4006 assert(allocation() != NULL, "");
3997 4007 if (allocation()->Opcode() == Op_Allocate) {
3998 4008 Node* klass_node = allocation()->in(AllocateNode::KlassNode);
3999 4009 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
4000 4010 if (zeroes_done == k->layout_helper())
4001 4011 zeroes_done = size_limit;
4002 4012 }
4003 4013 }
4004 4014 if (zeroes_done < size_limit) {
4005 4015 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
4006 4016 zeroes_done, size_in_bytes, phase);
4007 4017 }
4008 4018 }
4009 4019
4010 4020 set_complete(phase);
4011 4021 return rawmem;
4012 4022 }
4013 4023
4014 4024
4015 4025 #ifdef ASSERT
4016 4026 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
4017 4027 if (is_complete())
4018 4028 return true; // stores could be anything at this point
4019 4029 assert(allocation() != NULL, "must be present");
4020 4030 intptr_t last_off = allocation()->minimum_header_size();
4021 4031 for (uint i = InitializeNode::RawStores; i < req(); i++) {
4022 4032 Node* st = in(i);
4023 4033 intptr_t st_off = get_store_offset(st, phase);
4024 4034 if (st_off < 0) continue; // ignore dead garbage
4025 4035 if (last_off > st_off) {
4026 4036 tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off);
4027 4037 this->dump(2);
4028 4038 assert(false, "ascending store offsets");
4029 4039 return false;
4030 4040 }
4031 4041 last_off = st_off + st->as_Store()->memory_size();
4032 4042 }
4033 4043 return true;
4034 4044 }
4035 4045 #endif //ASSERT
4036 4046
4037 4047
4038 4048
4039 4049
4040 4050 //============================MergeMemNode=====================================
4041 4051 //
4042 4052 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
4043 4053 // contributing store or call operations. Each contributor provides the memory
4044 4054 // state for a particular "alias type" (see Compile::alias_type). For example,
4045 4055 // if a MergeMem has an input X for alias category #6, then any memory reference
4046 4056 // to alias category #6 may use X as its memory state input, as an exact equivalent
4047 4057 // to using the MergeMem as a whole.
4048 4058 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
4049 4059 //
4050 4060 // (Here, the <N> notation gives the index of the relevant adr_type.)
4051 4061 //
4052 4062 // In one special case (and more cases in the future), alias categories overlap.
4053 4063 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
4054 4064 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
4055 4065 // it is exactly equivalent to that state W:
4056 4066 // MergeMem(<Bot>: W) <==> W
4057 4067 //
4058 4068 // Usually, the merge has more than one input. In that case, where inputs
4059 4069 // overlap (i.e., one is Bot), the narrower alias type determines the memory
4060 4070 // state for that type, and the wider alias type (Bot) fills in everywhere else:
4061 4071 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
4062 4072 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
4063 4073 //
4064 4074 // A merge can take a "wide" memory state as one of its narrow inputs.
4065 4075 // This simply means that the merge observes out only the relevant parts of
4066 4076 // the wide input. That is, wide memory states arriving at narrow merge inputs
4067 4077 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
4068 4078 //
4069 4079 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
4070 4080 // and that memory slices "leak through":
4071 4081 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
4072 4082 //
4073 4083 // But, in such a cascade, repeated memory slices can "block the leak":
4074 4084 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
4075 4085 //
4076 4086 // In the last example, Y is not part of the combined memory state of the
4077 4087 // outermost MergeMem. The system must, of course, prevent unschedulable
4078 4088 // memory states from arising, so you can be sure that the state Y is somehow
4079 4089 // a precursor to state Y'.
4080 4090 //
4081 4091 //
4082 4092 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
4083 4093 // of each MergeMemNode array are exactly the numerical alias indexes, including
4084 4094 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
4085 4095 // Compile::alias_type (and kin) produce and manage these indexes.
4086 4096 //
4087 4097 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
4088 4098 // (Note that this provides quick access to the top node inside MergeMem methods,
4089 4099 // without the need to reach out via TLS to Compile::current.)
4090 4100 //
4091 4101 // As a consequence of what was just described, a MergeMem that represents a full
4092 4102 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
4093 4103 // containing all alias categories.
4094 4104 //
4095 4105 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
4096 4106 //
4097 4107 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
4098 4108 // a memory state for the alias type <N>, or else the top node, meaning that
4099 4109 // there is no particular input for that alias type. Note that the length of
4100 4110 // a MergeMem is variable, and may be extended at any time to accommodate new
4101 4111 // memory states at larger alias indexes. When merges grow, they are of course
4102 4112 // filled with "top" in the unused in() positions.
4103 4113 //
4104 4114 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
4105 4115 // (Top was chosen because it works smoothly with passes like GCM.)
4106 4116 //
4107 4117 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
4108 4118 // the type of random VM bits like TLS references.) Since it is always the
4109 4119 // first non-Bot memory slice, some low-level loops use it to initialize an
4110 4120 // index variable: for (i = AliasIdxRaw; i < req(); i++).
4111 4121 //
4112 4122 //
4113 4123 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
4114 4124 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
4115 4125 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
4116 4126 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
4117 4127 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
4118 4128 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
4119 4129 //
4120 4130 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
4121 4131 // really that different from the other memory inputs. An abbreviation called
4122 4132 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
4123 4133 //
4124 4134 //
4125 4135 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
4126 4136 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
4127 4137 // that "emerges though" the base memory will be marked as excluding the alias types
4128 4138 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
4129 4139 //
4130 4140 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
4131 4141 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
4132 4142 //
4133 4143 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
4134 4144 // (It is currently unimplemented.) As you can see, the resulting merge is
4135 4145 // actually a disjoint union of memory states, rather than an overlay.
4136 4146 //
4137 4147
4138 4148 //------------------------------MergeMemNode-----------------------------------
4139 4149 Node* MergeMemNode::make_empty_memory() {
4140 4150 Node* empty_memory = (Node*) Compile::current()->top();
4141 4151 assert(empty_memory->is_top(), "correct sentinel identity");
4142 4152 return empty_memory;
4143 4153 }
4144 4154
4145 4155 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
4146 4156 init_class_id(Class_MergeMem);
4147 4157 // all inputs are nullified in Node::Node(int)
4148 4158 // set_input(0, NULL); // no control input
4149 4159
4150 4160 // Initialize the edges uniformly to top, for starters.
4151 4161 Node* empty_mem = make_empty_memory();
4152 4162 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
4153 4163 init_req(i,empty_mem);
4154 4164 }
4155 4165 assert(empty_memory() == empty_mem, "");
4156 4166
4157 4167 if( new_base != NULL && new_base->is_MergeMem() ) {
4158 4168 MergeMemNode* mdef = new_base->as_MergeMem();
4159 4169 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
4160 4170 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
4161 4171 mms.set_memory(mms.memory2());
4162 4172 }
4163 4173 assert(base_memory() == mdef->base_memory(), "");
4164 4174 } else {
4165 4175 set_base_memory(new_base);
4166 4176 }
4167 4177 }
4168 4178
4169 4179 // Make a new, untransformed MergeMem with the same base as 'mem'.
4170 4180 // If mem is itself a MergeMem, populate the result with the same edges.
4171 4181 MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) {
4172 4182 return new(C) MergeMemNode(mem);
4173 4183 }
4174 4184
4175 4185 //------------------------------cmp--------------------------------------------
4176 4186 uint MergeMemNode::hash() const { return NO_HASH; }
4177 4187 uint MergeMemNode::cmp( const Node &n ) const {
4178 4188 return (&n == this); // Always fail except on self
4179 4189 }
4180 4190
4181 4191 //------------------------------Identity---------------------------------------
4182 4192 Node* MergeMemNode::Identity(PhaseTransform *phase) {
4183 4193 // Identity if this merge point does not record any interesting memory
4184 4194 // disambiguations.
4185 4195 Node* base_mem = base_memory();
4186 4196 Node* empty_mem = empty_memory();
4187 4197 if (base_mem != empty_mem) { // Memory path is not dead?
4188 4198 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4189 4199 Node* mem = in(i);
4190 4200 if (mem != empty_mem && mem != base_mem) {
4191 4201 return this; // Many memory splits; no change
4192 4202 }
4193 4203 }
4194 4204 }
4195 4205 return base_mem; // No memory splits; ID on the one true input
4196 4206 }
4197 4207
4198 4208 //------------------------------Ideal------------------------------------------
4199 4209 // This method is invoked recursively on chains of MergeMem nodes
4200 4210 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4201 4211 // Remove chain'd MergeMems
4202 4212 //
4203 4213 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
4204 4214 // relative to the "in(Bot)". Since we are patching both at the same time,
4205 4215 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
4206 4216 // but rewrite each "in(i)" relative to the new "in(Bot)".
4207 4217 Node *progress = NULL;
4208 4218
4209 4219
4210 4220 Node* old_base = base_memory();
4211 4221 Node* empty_mem = empty_memory();
4212 4222 if (old_base == empty_mem)
4213 4223 return NULL; // Dead memory path.
4214 4224
4215 4225 MergeMemNode* old_mbase;
4216 4226 if (old_base != NULL && old_base->is_MergeMem())
4217 4227 old_mbase = old_base->as_MergeMem();
4218 4228 else
4219 4229 old_mbase = NULL;
4220 4230 Node* new_base = old_base;
4221 4231
4222 4232 // simplify stacked MergeMems in base memory
4223 4233 if (old_mbase) new_base = old_mbase->base_memory();
4224 4234
4225 4235 // the base memory might contribute new slices beyond my req()
4226 4236 if (old_mbase) grow_to_match(old_mbase);
4227 4237
4228 4238 // Look carefully at the base node if it is a phi.
4229 4239 PhiNode* phi_base;
4230 4240 if (new_base != NULL && new_base->is_Phi())
4231 4241 phi_base = new_base->as_Phi();
4232 4242 else
4233 4243 phi_base = NULL;
4234 4244
4235 4245 Node* phi_reg = NULL;
4236 4246 uint phi_len = (uint)-1;
4237 4247 if (phi_base != NULL && !phi_base->is_copy()) {
4238 4248 // do not examine phi if degraded to a copy
4239 4249 phi_reg = phi_base->region();
4240 4250 phi_len = phi_base->req();
4241 4251 // see if the phi is unfinished
4242 4252 for (uint i = 1; i < phi_len; i++) {
4243 4253 if (phi_base->in(i) == NULL) {
4244 4254 // incomplete phi; do not look at it yet!
4245 4255 phi_reg = NULL;
4246 4256 phi_len = (uint)-1;
4247 4257 break;
4248 4258 }
4249 4259 }
4250 4260 }
4251 4261
4252 4262 // Note: We do not call verify_sparse on entry, because inputs
4253 4263 // can normalize to the base_memory via subsume_node or similar
4254 4264 // mechanisms. This method repairs that damage.
4255 4265
4256 4266 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
4257 4267
4258 4268 // Look at each slice.
4259 4269 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4260 4270 Node* old_in = in(i);
4261 4271 // calculate the old memory value
4262 4272 Node* old_mem = old_in;
4263 4273 if (old_mem == empty_mem) old_mem = old_base;
4264 4274 assert(old_mem == memory_at(i), "");
4265 4275
4266 4276 // maybe update (reslice) the old memory value
4267 4277
4268 4278 // simplify stacked MergeMems
4269 4279 Node* new_mem = old_mem;
4270 4280 MergeMemNode* old_mmem;
4271 4281 if (old_mem != NULL && old_mem->is_MergeMem())
4272 4282 old_mmem = old_mem->as_MergeMem();
4273 4283 else
4274 4284 old_mmem = NULL;
4275 4285 if (old_mmem == this) {
4276 4286 // This can happen if loops break up and safepoints disappear.
4277 4287 // A merge of BotPtr (default) with a RawPtr memory derived from a
4278 4288 // safepoint can be rewritten to a merge of the same BotPtr with
4279 4289 // the BotPtr phi coming into the loop. If that phi disappears
4280 4290 // also, we can end up with a self-loop of the mergemem.
4281 4291 // In general, if loops degenerate and memory effects disappear,
4282 4292 // a mergemem can be left looking at itself. This simply means
4283 4293 // that the mergemem's default should be used, since there is
4284 4294 // no longer any apparent effect on this slice.
4285 4295 // Note: If a memory slice is a MergeMem cycle, it is unreachable
4286 4296 // from start. Update the input to TOP.
4287 4297 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
4288 4298 }
4289 4299 else if (old_mmem != NULL) {
4290 4300 new_mem = old_mmem->memory_at(i);
4291 4301 }
4292 4302 // else preceding memory was not a MergeMem
4293 4303
4294 4304 // replace equivalent phis (unfortunately, they do not GVN together)
4295 4305 if (new_mem != NULL && new_mem != new_base &&
4296 4306 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
4297 4307 if (new_mem->is_Phi()) {
4298 4308 PhiNode* phi_mem = new_mem->as_Phi();
4299 4309 for (uint i = 1; i < phi_len; i++) {
4300 4310 if (phi_base->in(i) != phi_mem->in(i)) {
4301 4311 phi_mem = NULL;
4302 4312 break;
4303 4313 }
4304 4314 }
4305 4315 if (phi_mem != NULL) {
4306 4316 // equivalent phi nodes; revert to the def
4307 4317 new_mem = new_base;
4308 4318 }
4309 4319 }
4310 4320 }
4311 4321
4312 4322 // maybe store down a new value
4313 4323 Node* new_in = new_mem;
4314 4324 if (new_in == new_base) new_in = empty_mem;
4315 4325
4316 4326 if (new_in != old_in) {
4317 4327 // Warning: Do not combine this "if" with the previous "if"
4318 4328 // A memory slice might have be be rewritten even if it is semantically
4319 4329 // unchanged, if the base_memory value has changed.
4320 4330 set_req(i, new_in);
4321 4331 progress = this; // Report progress
4322 4332 }
4323 4333 }
4324 4334
4325 4335 if (new_base != old_base) {
4326 4336 set_req(Compile::AliasIdxBot, new_base);
4327 4337 // Don't use set_base_memory(new_base), because we need to update du.
4328 4338 assert(base_memory() == new_base, "");
4329 4339 progress = this;
4330 4340 }
4331 4341
4332 4342 if( base_memory() == this ) {
4333 4343 // a self cycle indicates this memory path is dead
4334 4344 set_req(Compile::AliasIdxBot, empty_mem);
4335 4345 }
4336 4346
4337 4347 // Resolve external cycles by calling Ideal on a MergeMem base_memory
4338 4348 // Recursion must occur after the self cycle check above
4339 4349 if( base_memory()->is_MergeMem() ) {
4340 4350 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
4341 4351 Node *m = phase->transform(new_mbase); // Rollup any cycles
4342 4352 if( m != NULL && (m->is_top() ||
4343 4353 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) {
4344 4354 // propagate rollup of dead cycle to self
4345 4355 set_req(Compile::AliasIdxBot, empty_mem);
4346 4356 }
4347 4357 }
4348 4358
4349 4359 if( base_memory() == empty_mem ) {
4350 4360 progress = this;
4351 4361 // Cut inputs during Parse phase only.
4352 4362 // During Optimize phase a dead MergeMem node will be subsumed by Top.
4353 4363 if( !can_reshape ) {
4354 4364 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4355 4365 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
4356 4366 }
4357 4367 }
4358 4368 }
4359 4369
4360 4370 if( !progress && base_memory()->is_Phi() && can_reshape ) {
4361 4371 // Check if PhiNode::Ideal's "Split phis through memory merges"
4362 4372 // transform should be attempted. Look for this->phi->this cycle.
4363 4373 uint merge_width = req();
4364 4374 if (merge_width > Compile::AliasIdxRaw) {
4365 4375 PhiNode* phi = base_memory()->as_Phi();
4366 4376 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
4367 4377 if (phi->in(i) == this) {
4368 4378 phase->is_IterGVN()->_worklist.push(phi);
4369 4379 break;
4370 4380 }
4371 4381 }
4372 4382 }
4373 4383 }
4374 4384
4375 4385 assert(progress || verify_sparse(), "please, no dups of base");
4376 4386 return progress;
4377 4387 }
4378 4388
4379 4389 //-------------------------set_base_memory-------------------------------------
4380 4390 void MergeMemNode::set_base_memory(Node *new_base) {
4381 4391 Node* empty_mem = empty_memory();
4382 4392 set_req(Compile::AliasIdxBot, new_base);
4383 4393 assert(memory_at(req()) == new_base, "must set default memory");
4384 4394 // Clear out other occurrences of new_base:
4385 4395 if (new_base != empty_mem) {
4386 4396 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4387 4397 if (in(i) == new_base) set_req(i, empty_mem);
4388 4398 }
4389 4399 }
4390 4400 }
4391 4401
4392 4402 //------------------------------out_RegMask------------------------------------
4393 4403 const RegMask &MergeMemNode::out_RegMask() const {
4394 4404 return RegMask::Empty;
4395 4405 }
4396 4406
4397 4407 //------------------------------dump_spec--------------------------------------
4398 4408 #ifndef PRODUCT
4399 4409 void MergeMemNode::dump_spec(outputStream *st) const {
4400 4410 st->print(" {");
4401 4411 Node* base_mem = base_memory();
4402 4412 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
4403 4413 Node* mem = memory_at(i);
4404 4414 if (mem == base_mem) { st->print(" -"); continue; }
4405 4415 st->print( " N%d:", mem->_idx );
4406 4416 Compile::current()->get_adr_type(i)->dump_on(st);
4407 4417 }
4408 4418 st->print(" }");
4409 4419 }
4410 4420 #endif // !PRODUCT
4411 4421
4412 4422
4413 4423 #ifdef ASSERT
4414 4424 static bool might_be_same(Node* a, Node* b) {
4415 4425 if (a == b) return true;
4416 4426 if (!(a->is_Phi() || b->is_Phi())) return false;
4417 4427 // phis shift around during optimization
4418 4428 return true; // pretty stupid...
4419 4429 }
4420 4430
4421 4431 // verify a narrow slice (either incoming or outgoing)
4422 4432 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
4423 4433 if (!VerifyAliases) return; // don't bother to verify unless requested
4424 4434 if (is_error_reported()) return; // muzzle asserts when debugging an error
4425 4435 if (Node::in_dump()) return; // muzzle asserts when printing
4426 4436 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
4427 4437 assert(n != NULL, "");
4428 4438 // Elide intervening MergeMem's
4429 4439 while (n->is_MergeMem()) {
4430 4440 n = n->as_MergeMem()->memory_at(alias_idx);
4431 4441 }
4432 4442 Compile* C = Compile::current();
4433 4443 const TypePtr* n_adr_type = n->adr_type();
4434 4444 if (n == m->empty_memory()) {
4435 4445 // Implicit copy of base_memory()
4436 4446 } else if (n_adr_type != TypePtr::BOTTOM) {
4437 4447 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
4438 4448 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
4439 4449 } else {
4440 4450 // A few places like make_runtime_call "know" that VM calls are narrow,
4441 4451 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
4442 4452 bool expected_wide_mem = false;
4443 4453 if (n == m->base_memory()) {
4444 4454 expected_wide_mem = true;
4445 4455 } else if (alias_idx == Compile::AliasIdxRaw ||
4446 4456 n == m->memory_at(Compile::AliasIdxRaw)) {
4447 4457 expected_wide_mem = true;
4448 4458 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
4449 4459 // memory can "leak through" calls on channels that
4450 4460 // are write-once. Allow this also.
4451 4461 expected_wide_mem = true;
4452 4462 }
4453 4463 assert(expected_wide_mem, "expected narrow slice replacement");
4454 4464 }
4455 4465 }
4456 4466 #else // !ASSERT
4457 4467 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op
4458 4468 #endif
4459 4469
4460 4470
4461 4471 //-----------------------------memory_at---------------------------------------
4462 4472 Node* MergeMemNode::memory_at(uint alias_idx) const {
4463 4473 assert(alias_idx >= Compile::AliasIdxRaw ||
4464 4474 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
4465 4475 "must avoid base_memory and AliasIdxTop");
4466 4476
4467 4477 // Otherwise, it is a narrow slice.
4468 4478 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
4469 4479 Compile *C = Compile::current();
4470 4480 if (is_empty_memory(n)) {
4471 4481 // the array is sparse; empty slots are the "top" node
4472 4482 n = base_memory();
4473 4483 assert(Node::in_dump()
4474 4484 || n == NULL || n->bottom_type() == Type::TOP
4475 4485 || n->adr_type() == NULL // address is TOP
4476 4486 || n->adr_type() == TypePtr::BOTTOM
4477 4487 || n->adr_type() == TypeRawPtr::BOTTOM
4478 4488 || Compile::current()->AliasLevel() == 0,
4479 4489 "must be a wide memory");
4480 4490 // AliasLevel == 0 if we are organizing the memory states manually.
4481 4491 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
4482 4492 } else {
4483 4493 // make sure the stored slice is sane
4484 4494 #ifdef ASSERT
4485 4495 if (is_error_reported() || Node::in_dump()) {
4486 4496 } else if (might_be_same(n, base_memory())) {
4487 4497 // Give it a pass: It is a mostly harmless repetition of the base.
4488 4498 // This can arise normally from node subsumption during optimization.
4489 4499 } else {
4490 4500 verify_memory_slice(this, alias_idx, n);
4491 4501 }
4492 4502 #endif
4493 4503 }
4494 4504 return n;
4495 4505 }
4496 4506
4497 4507 //---------------------------set_memory_at-------------------------------------
4498 4508 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
4499 4509 verify_memory_slice(this, alias_idx, n);
4500 4510 Node* empty_mem = empty_memory();
4501 4511 if (n == base_memory()) n = empty_mem; // collapse default
4502 4512 uint need_req = alias_idx+1;
4503 4513 if (req() < need_req) {
4504 4514 if (n == empty_mem) return; // already the default, so do not grow me
4505 4515 // grow the sparse array
4506 4516 do {
4507 4517 add_req(empty_mem);
4508 4518 } while (req() < need_req);
4509 4519 }
4510 4520 set_req( alias_idx, n );
4511 4521 }
4512 4522
4513 4523
4514 4524
4515 4525 //--------------------------iteration_setup------------------------------------
4516 4526 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
4517 4527 if (other != NULL) {
4518 4528 grow_to_match(other);
4519 4529 // invariant: the finite support of mm2 is within mm->req()
4520 4530 #ifdef ASSERT
4521 4531 for (uint i = req(); i < other->req(); i++) {
4522 4532 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
4523 4533 }
4524 4534 #endif
4525 4535 }
4526 4536 // Replace spurious copies of base_memory by top.
4527 4537 Node* base_mem = base_memory();
4528 4538 if (base_mem != NULL && !base_mem->is_top()) {
4529 4539 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
4530 4540 if (in(i) == base_mem)
4531 4541 set_req(i, empty_memory());
4532 4542 }
4533 4543 }
4534 4544 }
4535 4545
4536 4546 //---------------------------grow_to_match-------------------------------------
4537 4547 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
4538 4548 Node* empty_mem = empty_memory();
4539 4549 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
4540 4550 // look for the finite support of the other memory
4541 4551 for (uint i = other->req(); --i >= req(); ) {
4542 4552 if (other->in(i) != empty_mem) {
4543 4553 uint new_len = i+1;
4544 4554 while (req() < new_len) add_req(empty_mem);
4545 4555 break;
4546 4556 }
4547 4557 }
4548 4558 }
4549 4559
4550 4560 //---------------------------verify_sparse-------------------------------------
4551 4561 #ifndef PRODUCT
4552 4562 bool MergeMemNode::verify_sparse() const {
4553 4563 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
4554 4564 Node* base_mem = base_memory();
4555 4565 // The following can happen in degenerate cases, since empty==top.
4556 4566 if (is_empty_memory(base_mem)) return true;
4557 4567 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4558 4568 assert(in(i) != NULL, "sane slice");
4559 4569 if (in(i) == base_mem) return false; // should have been the sentinel value!
4560 4570 }
4561 4571 return true;
4562 4572 }
4563 4573
4564 4574 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
4565 4575 Node* n;
4566 4576 n = mm->in(idx);
4567 4577 if (mem == n) return true; // might be empty_memory()
4568 4578 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
4569 4579 if (mem == n) return true;
4570 4580 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) {
4571 4581 if (mem == n) return true;
4572 4582 if (n == NULL) break;
4573 4583 }
4574 4584 return false;
4575 4585 }
4576 4586 #endif // !PRODUCT
↓ open down ↓ |
2535 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX