1 /*
2 * Copyright 2005-2008 Sun Microsystems, Inc. All Rights Reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 * have any questions.
22 *
23 */
24
25 #include "incls/_precompiled.incl"
26 #include "incls/_macro.cpp.incl"
27
28
29 //
30 // Replace any references to "oldref" in inputs to "use" with "newref".
31 // Returns the number of replacements made.
32 //
33 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
34 int nreplacements = 0;
35 uint req = use->req();
36 for (uint j = 0; j < use->len(); j++) {
37 Node *uin = use->in(j);
38 if (uin == oldref) {
39 if (j < req)
40 use->set_req(j, newref);
41 else
42 use->set_prec(j, newref);
43 nreplacements++;
44 } else if (j >= req && uin == NULL) {
45 break;
46 }
47 }
48 return nreplacements;
49 }
50
51 void PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) {
52 // Copy debug information and adjust JVMState information
53 uint old_dbg_start = oldcall->tf()->domain()->cnt();
54 uint new_dbg_start = newcall->tf()->domain()->cnt();
55 int jvms_adj = new_dbg_start - old_dbg_start;
56 assert (new_dbg_start == newcall->req(), "argument count mismatch");
57
58 Dict* sosn_map = new Dict(cmpkey,hashkey);
59 for (uint i = old_dbg_start; i < oldcall->req(); i++) {
60 Node* old_in = oldcall->in(i);
61 // Clone old SafePointScalarObjectNodes, adjusting their field contents.
62 if (old_in != NULL && old_in->is_SafePointScalarObject()) {
63 SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject();
64 uint old_unique = C->unique();
65 Node* new_in = old_sosn->clone(jvms_adj, sosn_map);
66 if (old_unique != C->unique()) {
67 new_in = transform_later(new_in); // Register new node.
68 }
69 old_in = new_in;
70 }
71 newcall->add_req(old_in);
72 }
73
74 newcall->set_jvms(oldcall->jvms());
75 for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) {
76 jvms->set_map(newcall);
77 jvms->set_locoff(jvms->locoff()+jvms_adj);
78 jvms->set_stkoff(jvms->stkoff()+jvms_adj);
79 jvms->set_monoff(jvms->monoff()+jvms_adj);
80 jvms->set_scloff(jvms->scloff()+jvms_adj);
81 jvms->set_endoff(jvms->endoff()+jvms_adj);
82 }
83 }
84
85 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
86 Node* cmp;
87 if (mask != 0) {
88 Node* and_node = transform_later(new (C, 3) AndXNode(word, MakeConX(mask)));
89 cmp = transform_later(new (C, 3) CmpXNode(and_node, MakeConX(bits)));
90 } else {
91 cmp = word;
92 }
93 Node* bol = transform_later(new (C, 2) BoolNode(cmp, BoolTest::ne));
94 IfNode* iff = new (C, 2) IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
95 transform_later(iff);
96
97 // Fast path taken.
98 Node *fast_taken = transform_later( new (C, 1) IfFalseNode(iff) );
99
100 // Fast path not-taken, i.e. slow path
101 Node *slow_taken = transform_later( new (C, 1) IfTrueNode(iff) );
102
103 if (return_fast_path) {
104 region->init_req(edge, slow_taken); // Capture slow-control
105 return fast_taken;
106 } else {
107 region->init_req(edge, fast_taken); // Capture fast-control
108 return slow_taken;
109 }
110 }
111
112 //--------------------copy_predefined_input_for_runtime_call--------------------
113 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
114 // Set fixed predefined input arguments
115 call->init_req( TypeFunc::Control, ctrl );
116 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) );
117 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
118 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
119 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
120 }
121
122 //------------------------------make_slow_call---------------------------------
123 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type, address slow_call, const char* leaf_name, Node* slow_path, Node* parm0, Node* parm1) {
124
125 // Slow-path call
126 int size = slow_call_type->domain()->cnt();
127 CallNode *call = leaf_name
128 ? (CallNode*)new (C, size) CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
129 : (CallNode*)new (C, size) CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM );
130
131 // Slow path call has no side-effects, uses few values
132 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
133 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0);
134 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1);
135 copy_call_debug_info(oldcall, call);
136 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
137 _igvn.hash_delete(oldcall);
138 _igvn.subsume_node(oldcall, call);
139 transform_later(call);
140
141 return call;
142 }
143
144 void PhaseMacroExpand::extract_call_projections(CallNode *call) {
145 _fallthroughproj = NULL;
146 _fallthroughcatchproj = NULL;
147 _ioproj_fallthrough = NULL;
148 _ioproj_catchall = NULL;
149 _catchallcatchproj = NULL;
150 _memproj_fallthrough = NULL;
151 _memproj_catchall = NULL;
152 _resproj = NULL;
153 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) {
154 ProjNode *pn = call->fast_out(i)->as_Proj();
155 switch (pn->_con) {
156 case TypeFunc::Control:
157 {
158 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj
159 _fallthroughproj = pn;
160 DUIterator_Fast jmax, j = pn->fast_outs(jmax);
161 const Node *cn = pn->fast_out(j);
162 if (cn->is_Catch()) {
163 ProjNode *cpn = NULL;
164 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) {
165 cpn = cn->fast_out(k)->as_Proj();
166 assert(cpn->is_CatchProj(), "must be a CatchProjNode");
167 if (cpn->_con == CatchProjNode::fall_through_index)
168 _fallthroughcatchproj = cpn;
169 else {
170 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index.");
171 _catchallcatchproj = cpn;
172 }
173 }
174 }
175 break;
176 }
177 case TypeFunc::I_O:
178 if (pn->_is_io_use)
179 _ioproj_catchall = pn;
180 else
181 _ioproj_fallthrough = pn;
182 break;
183 case TypeFunc::Memory:
184 if (pn->_is_io_use)
185 _memproj_catchall = pn;
186 else
187 _memproj_fallthrough = pn;
188 break;
189 case TypeFunc::Parms:
190 _resproj = pn;
191 break;
192 default:
193 assert(false, "unexpected projection from allocation node.");
194 }
195 }
196
197 }
198
199 // Eliminate a card mark sequence. p2x is a ConvP2XNode
200 void PhaseMacroExpand::eliminate_card_mark(Node *p2x) {
201 assert(p2x->Opcode() == Op_CastP2X, "ConvP2XNode required");
202 Node *shift = p2x->unique_out();
203 Node *addp = shift->unique_out();
204 for (DUIterator_Last jmin, j = addp->last_outs(jmin); j >= jmin; --j) {
205 Node *st = addp->last_out(j);
206 assert(st->is_Store(), "store required");
207 _igvn.replace_node(st, st->in(MemNode::Memory));
208 }
209 }
210
211 // Search for a memory operation for the specified memory slice.
212 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
213 Node *orig_mem = mem;
214 Node *alloc_mem = alloc->in(TypeFunc::Memory);
215 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
216 while (true) {
217 if (mem == alloc_mem || mem == start_mem ) {
218 return mem; // hit one of our sentinals
219 } else if (mem->is_MergeMem()) {
220 mem = mem->as_MergeMem()->memory_at(alias_idx);
221 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
222 Node *in = mem->in(0);
223 // we can safely skip over safepoints, calls, locks and membars because we
224 // already know that the object is safe to eliminate.
225 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
226 return in;
227 } else if (in->is_Call()) {
228 CallNode *call = in->as_Call();
229 if (!call->may_modify(tinst, phase)) {
230 mem = call->in(TypeFunc::Memory);
231 }
232 mem = in->in(TypeFunc::Memory);
233 } else if (in->is_MemBar()) {
234 mem = in->in(TypeFunc::Memory);
235 } else {
236 assert(false, "unexpected projection");
237 }
238 } else if (mem->is_Store()) {
239 const TypePtr* atype = mem->as_Store()->adr_type();
240 int adr_idx = Compile::current()->get_alias_index(atype);
241 if (adr_idx == alias_idx) {
242 assert(atype->isa_oopptr(), "address type must be oopptr");
243 int adr_offset = atype->offset();
244 uint adr_iid = atype->is_oopptr()->instance_id();
245 // Array elements references have the same alias_idx
246 // but different offset and different instance_id.
247 if (adr_offset == offset && adr_iid == alloc->_idx)
248 return mem;
249 } else {
250 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
251 }
252 mem = mem->in(MemNode::Memory);
253 } else {
254 return mem;
255 }
256 assert(mem != orig_mem, "dead memory loop");
257 }
258 }
259
260 //
261 // Given a Memory Phi, compute a value Phi containing the values from stores
262 // on the input paths.
263 // Note: this function is recursive, its depth is limied by the "level" argument
264 // Returns the computed Phi, or NULL if it cannot compute it.
265 Node *PhaseMacroExpand::value_from_mem_phi(Node *mem, BasicType ft, const Type *phi_type, const TypeOopPtr *adr_t, Node *alloc, Node_Stack *value_phis, int level) {
266 assert(mem->is_Phi(), "sanity");
267 int alias_idx = C->get_alias_index(adr_t);
268 int offset = adr_t->offset();
269 int instance_id = adr_t->instance_id();
270
271 // Check if an appropriate value phi already exists.
272 Node* region = mem->in(0);
273 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
274 Node* phi = region->fast_out(k);
275 if (phi->is_Phi() && phi != mem &&
276 phi->as_Phi()->is_same_inst_field(phi_type, instance_id, alias_idx, offset)) {
277 return phi;
278 }
279 }
280 // Check if an appropriate new value phi already exists.
281 Node* new_phi = NULL;
282 uint size = value_phis->size();
283 for (uint i=0; i < size; i++) {
284 if ( mem->_idx == value_phis->index_at(i) ) {
285 return value_phis->node_at(i);
286 }
287 }
288
289 if (level <= 0) {
290 return NULL; // Give up: phi tree too deep
291 }
292 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
293 Node *alloc_mem = alloc->in(TypeFunc::Memory);
294
295 uint length = mem->req();
296 GrowableArray <Node *> values(length, length, NULL);
297
298 // create a new Phi for the value
299 PhiNode *phi = new (C, length) PhiNode(mem->in(0), phi_type, NULL, instance_id, alias_idx, offset);
300 transform_later(phi);
301 value_phis->push(phi, mem->_idx);
302
303 for (uint j = 1; j < length; j++) {
304 Node *in = mem->in(j);
305 if (in == NULL || in->is_top()) {
306 values.at_put(j, in);
307 } else {
308 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
309 if (val == start_mem || val == alloc_mem) {
310 // hit a sentinel, return appropriate 0 value
311 values.at_put(j, _igvn.zerocon(ft));
312 continue;
313 }
314 if (val->is_Initialize()) {
315 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
316 }
317 if (val == NULL) {
318 return NULL; // can't find a value on this path
319 }
320 if (val == mem) {
321 values.at_put(j, mem);
322 } else if (val->is_Store()) {
323 values.at_put(j, val->in(MemNode::ValueIn));
324 } else if(val->is_Proj() && val->in(0) == alloc) {
325 values.at_put(j, _igvn.zerocon(ft));
326 } else if (val->is_Phi()) {
327 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
328 if (val == NULL) {
329 return NULL;
330 }
331 values.at_put(j, val);
332 } else {
333 assert(false, "unknown node on this path");
334 return NULL; // unknown node on this path
335 }
336 }
337 }
338 // Set Phi's inputs
339 for (uint j = 1; j < length; j++) {
340 if (values.at(j) == mem) {
341 phi->init_req(j, phi);
342 } else {
343 phi->init_req(j, values.at(j));
344 }
345 }
346 return phi;
347 }
348
349 // Search the last value stored into the object's field.
350 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, Node *alloc) {
351 assert(adr_t->is_known_instance_field(), "instance required");
352 int instance_id = adr_t->instance_id();
353 assert((uint)instance_id == alloc->_idx, "wrong allocation");
354
355 int alias_idx = C->get_alias_index(adr_t);
356 int offset = adr_t->offset();
357 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
358 Node *alloc_ctrl = alloc->in(TypeFunc::Control);
359 Node *alloc_mem = alloc->in(TypeFunc::Memory);
360 Arena *a = Thread::current()->resource_area();
361 VectorSet visited(a);
362
363
364 bool done = sfpt_mem == alloc_mem;
365 Node *mem = sfpt_mem;
366 while (!done) {
367 if (visited.test_set(mem->_idx)) {
368 return NULL; // found a loop, give up
369 }
370 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
371 if (mem == start_mem || mem == alloc_mem) {
372 done = true; // hit a sentinel, return appropriate 0 value
373 } else if (mem->is_Initialize()) {
374 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
375 if (mem == NULL) {
376 done = true; // Something go wrong.
377 } else if (mem->is_Store()) {
378 const TypePtr* atype = mem->as_Store()->adr_type();
379 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
380 done = true;
381 }
382 } else if (mem->is_Store()) {
383 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
384 assert(atype != NULL, "address type must be oopptr");
385 assert(C->get_alias_index(atype) == alias_idx &&
386 atype->is_known_instance_field() && atype->offset() == offset &&
387 atype->instance_id() == instance_id, "store is correct memory slice");
388 done = true;
389 } else if (mem->is_Phi()) {
390 // try to find a phi's unique input
391 Node *unique_input = NULL;
392 Node *top = C->top();
393 for (uint i = 1; i < mem->req(); i++) {
394 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
395 if (n == NULL || n == top || n == mem) {
396 continue;
397 } else if (unique_input == NULL) {
398 unique_input = n;
399 } else if (unique_input != n) {
400 unique_input = top;
401 break;
402 }
403 }
404 if (unique_input != NULL && unique_input != top) {
405 mem = unique_input;
406 } else {
407 done = true;
408 }
409 } else {
410 assert(false, "unexpected node");
411 }
412 }
413 if (mem != NULL) {
414 if (mem == start_mem || mem == alloc_mem) {
415 // hit a sentinel, return appropriate 0 value
416 return _igvn.zerocon(ft);
417 } else if (mem->is_Store()) {
418 return mem->in(MemNode::ValueIn);
419 } else if (mem->is_Phi()) {
420 // attempt to produce a Phi reflecting the values on the input paths of the Phi
421 Node_Stack value_phis(a, 8);
422 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
423 if (phi != NULL) {
424 return phi;
425 } else {
426 // Kill all new Phis
427 while(value_phis.is_nonempty()) {
428 Node* n = value_phis.node();
429 _igvn.hash_delete(n);
430 _igvn.subsume_node(n, C->top());
431 value_phis.pop();
432 }
433 }
434 }
435 }
436 // Something go wrong.
437 return NULL;
438 }
439
440 // Check the possibility of scalar replacement.
441 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
442 // Scan the uses of the allocation to check for anything that would
443 // prevent us from eliminating it.
444 NOT_PRODUCT( const char* fail_eliminate = NULL; )
445 DEBUG_ONLY( Node* disq_node = NULL; )
446 bool can_eliminate = true;
447
448 Node* res = alloc->result_cast();
449 const TypeOopPtr* res_type = NULL;
450 if (res == NULL) {
451 // All users were eliminated.
452 } else if (!res->is_CheckCastPP()) {
453 alloc->_is_scalar_replaceable = false; // don't try again
454 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
455 can_eliminate = false;
456 } else {
457 res_type = _igvn.type(res)->isa_oopptr();
458 if (res_type == NULL) {
459 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
460 can_eliminate = false;
461 } else if (res_type->isa_aryptr()) {
462 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
463 if (length < 0) {
464 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
465 can_eliminate = false;
466 }
467 }
468 }
469
470 if (can_eliminate && res != NULL) {
471 for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
472 j < jmax && can_eliminate; j++) {
473 Node* use = res->fast_out(j);
474
475 if (use->is_AddP()) {
476 const TypePtr* addp_type = _igvn.type(use)->is_ptr();
477 int offset = addp_type->offset();
478
479 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
480 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";)
481 can_eliminate = false;
482 break;
483 }
484 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
485 k < kmax && can_eliminate; k++) {
486 Node* n = use->fast_out(k);
487 if (!n->is_Store() && n->Opcode() != Op_CastP2X) {
488 DEBUG_ONLY(disq_node = n;)
489 if (n->is_Load() || n->is_LoadStore()) {
490 NOT_PRODUCT(fail_eliminate = "Field load";)
491 } else {
492 NOT_PRODUCT(fail_eliminate = "Not store field referrence";)
493 }
494 can_eliminate = false;
495 }
496 }
497 } else if (use->is_SafePoint()) {
498 SafePointNode* sfpt = use->as_SafePoint();
499 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
500 // Object is passed as argument.
501 DEBUG_ONLY(disq_node = use;)
502 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
503 can_eliminate = false;
504 }
505 Node* sfptMem = sfpt->memory();
506 if (sfptMem == NULL || sfptMem->is_top()) {
507 DEBUG_ONLY(disq_node = use;)
508 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";)
509 can_eliminate = false;
510 } else {
511 safepoints.append_if_missing(sfpt);
512 }
513 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
514 if (use->is_Phi()) {
515 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
516 NOT_PRODUCT(fail_eliminate = "Object is return value";)
517 } else {
518 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
519 }
520 DEBUG_ONLY(disq_node = use;)
521 } else {
522 if (use->Opcode() == Op_Return) {
523 NOT_PRODUCT(fail_eliminate = "Object is return value";)
524 }else {
525 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
526 }
527 DEBUG_ONLY(disq_node = use;)
528 }
529 can_eliminate = false;
530 }
531 }
532 }
533
534 #ifndef PRODUCT
535 if (PrintEliminateAllocations) {
536 if (can_eliminate) {
537 tty->print("Scalar ");
538 if (res == NULL)
539 alloc->dump();
540 else
541 res->dump();
542 } else {
543 tty->print("NotScalar (%s)", fail_eliminate);
544 if (res == NULL)
545 alloc->dump();
546 else
547 res->dump();
548 #ifdef ASSERT
549 if (disq_node != NULL) {
550 tty->print(" >>>> ");
551 disq_node->dump();
552 }
553 #endif /*ASSERT*/
554 }
555 }
556 #endif
557 return can_eliminate;
558 }
559
560 // Do scalar replacement.
561 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
562 GrowableArray <SafePointNode *> safepoints_done;
563
564 ciKlass* klass = NULL;
565 ciInstanceKlass* iklass = NULL;
566 int nfields = 0;
567 int array_base;
568 int element_size;
569 BasicType basic_elem_type;
570 ciType* elem_type;
571
572 Node* res = alloc->result_cast();
573 const TypeOopPtr* res_type = NULL;
574 if (res != NULL) { // Could be NULL when there are no users
575 res_type = _igvn.type(res)->isa_oopptr();
576 }
577
578 if (res != NULL) {
579 klass = res_type->klass();
580 if (res_type->isa_instptr()) {
581 // find the fields of the class which will be needed for safepoint debug information
582 assert(klass->is_instance_klass(), "must be an instance klass.");
583 iklass = klass->as_instance_klass();
584 nfields = iklass->nof_nonstatic_fields();
585 } else {
586 // find the array's elements which will be needed for safepoint debug information
587 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
588 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass.");
589 elem_type = klass->as_array_klass()->element_type();
590 basic_elem_type = elem_type->basic_type();
591 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
592 element_size = type2aelembytes(basic_elem_type);
593 }
594 }
595 //
596 // Process the safepoint uses
597 //
598 while (safepoints.length() > 0) {
599 SafePointNode* sfpt = safepoints.pop();
600 Node* mem = sfpt->memory();
601 uint first_ind = sfpt->req();
602 SafePointScalarObjectNode* sobj = new (C, 1) SafePointScalarObjectNode(res_type,
603 #ifdef ASSERT
604 alloc,
605 #endif
606 first_ind, nfields);
607 sobj->init_req(0, sfpt->in(TypeFunc::Control));
608 transform_later(sobj);
609
610 // Scan object's fields adding an input to the safepoint for each field.
611 for (int j = 0; j < nfields; j++) {
612 intptr_t offset;
613 ciField* field = NULL;
614 if (iklass != NULL) {
615 field = iklass->nonstatic_field_at(j);
616 offset = field->offset();
617 elem_type = field->type();
618 basic_elem_type = field->layout_type();
619 } else {
620 offset = array_base + j * (intptr_t)element_size;
621 }
622
623 const Type *field_type;
624 // The next code is taken from Parse::do_get_xxx().
625 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) {
626 if (!elem_type->is_loaded()) {
627 field_type = TypeInstPtr::BOTTOM;
628 } else if (field != NULL && field->is_constant()) {
629 // This can happen if the constant oop is non-perm.
630 ciObject* con = field->constant_value().as_object();
631 // Do not "join" in the previous type; it doesn't add value,
632 // and may yield a vacuous result if the field is of interface type.
633 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
634 assert(field_type != NULL, "field singleton type must be consistent");
635 } else {
636 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
637 }
638 if (UseCompressedOops) {
639 field_type = field_type->make_narrowoop();
640 basic_elem_type = T_NARROWOOP;
641 }
642 } else {
643 field_type = Type::get_const_basic_type(basic_elem_type);
644 }
645
646 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr();
647
648 Node *field_val = value_from_mem(mem, basic_elem_type, field_type, field_addr_type, alloc);
649 if (field_val == NULL) {
650 // we weren't able to find a value for this field,
651 // give up on eliminating this allocation
652 alloc->_is_scalar_replaceable = false; // don't try again
653 // remove any extra entries we added to the safepoint
654 uint last = sfpt->req() - 1;
655 for (int k = 0; k < j; k++) {
656 sfpt->del_req(last--);
657 }
658 // rollback processed safepoints
659 while (safepoints_done.length() > 0) {
660 SafePointNode* sfpt_done = safepoints_done.pop();
661 // remove any extra entries we added to the safepoint
662 last = sfpt_done->req() - 1;
663 for (int k = 0; k < nfields; k++) {
664 sfpt_done->del_req(last--);
665 }
666 JVMState *jvms = sfpt_done->jvms();
667 jvms->set_endoff(sfpt_done->req());
668 // Now make a pass over the debug information replacing any references
669 // to SafePointScalarObjectNode with the allocated object.
670 int start = jvms->debug_start();
671 int end = jvms->debug_end();
672 for (int i = start; i < end; i++) {
673 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
674 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
675 if (scobj->first_index() == sfpt_done->req() &&
676 scobj->n_fields() == (uint)nfields) {
677 assert(scobj->alloc() == alloc, "sanity");
678 sfpt_done->set_req(i, res);
679 }
680 }
681 }
682 }
683 #ifndef PRODUCT
684 if (PrintEliminateAllocations) {
685 if (field != NULL) {
686 tty->print("=== At SafePoint node %d can't find value of Field: ",
687 sfpt->_idx);
688 field->print();
689 int field_idx = C->get_alias_index(field_addr_type);
690 tty->print(" (alias_idx=%d)", field_idx);
691 } else { // Array's element
692 tty->print("=== At SafePoint node %d can't find value of array element [%d]",
693 sfpt->_idx, j);
694 }
695 tty->print(", which prevents elimination of: ");
696 if (res == NULL)
697 alloc->dump();
698 else
699 res->dump();
700 }
701 #endif
702 return false;
703 }
704 if (UseCompressedOops && field_type->isa_narrowoop()) {
705 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
706 // to be able scalar replace the allocation.
707 if (field_val->is_EncodeP()) {
708 field_val = field_val->in(1);
709 } else {
710 field_val = transform_later(new (C, 2) DecodeNNode(field_val, field_val->bottom_type()->make_ptr()));
711 }
712 }
713 sfpt->add_req(field_val);
714 }
715 JVMState *jvms = sfpt->jvms();
716 jvms->set_endoff(sfpt->req());
717 // Now make a pass over the debug information replacing any references
718 // to the allocated object with "sobj"
719 int start = jvms->debug_start();
720 int end = jvms->debug_end();
721 for (int i = start; i < end; i++) {
722 if (sfpt->in(i) == res) {
723 sfpt->set_req(i, sobj);
724 }
725 }
726 safepoints_done.append_if_missing(sfpt); // keep it for rollback
727 }
728 return true;
729 }
730
731 // Process users of eliminated allocation.
732 void PhaseMacroExpand::process_users_of_allocation(AllocateNode *alloc) {
733 Node* res = alloc->result_cast();
734 if (res != NULL) {
735 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
736 Node *use = res->last_out(j);
737 uint oc1 = res->outcnt();
738
739 if (use->is_AddP()) {
740 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
741 Node *n = use->last_out(k);
742 uint oc2 = use->outcnt();
743 if (n->is_Store()) {
744 _igvn.replace_node(n, n->in(MemNode::Memory));
745 } else {
746 assert( n->Opcode() == Op_CastP2X, "CastP2X required");
747 eliminate_card_mark(n);
748 }
749 k -= (oc2 - use->outcnt());
750 }
751 } else {
752 assert( !use->is_SafePoint(), "safepoint uses must have been already elimiated");
753 assert( use->Opcode() == Op_CastP2X, "CastP2X required");
754 eliminate_card_mark(use);
755 }
756 j -= (oc1 - res->outcnt());
757 }
758 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
759 _igvn.remove_dead_node(res);
760 }
761
762 //
763 // Process other users of allocation's projections
764 //
765 if (_resproj != NULL && _resproj->outcnt() != 0) {
766 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) {
767 Node *use = _resproj->last_out(j);
768 uint oc1 = _resproj->outcnt();
769 if (use->is_Initialize()) {
770 // Eliminate Initialize node.
771 InitializeNode *init = use->as_Initialize();
772 assert(init->outcnt() <= 2, "only a control and memory projection expected");
773 Node *ctrl_proj = init->proj_out(TypeFunc::Control);
774 if (ctrl_proj != NULL) {
775 assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection");
776 _igvn.replace_node(ctrl_proj, _fallthroughcatchproj);
777 }
778 Node *mem_proj = init->proj_out(TypeFunc::Memory);
779 if (mem_proj != NULL) {
780 Node *mem = init->in(TypeFunc::Memory);
781 #ifdef ASSERT
782 if (mem->is_MergeMem()) {
783 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection");
784 } else {
785 assert(mem == _memproj_fallthrough, "allocation memory projection");
786 }
787 #endif
788 _igvn.replace_node(mem_proj, mem);
789 }
790 } else if (use->is_AddP()) {
791 // raw memory addresses used only by the initialization
792 _igvn.hash_delete(use);
793 _igvn.subsume_node(use, C->top());
794 } else {
795 assert(false, "only Initialize or AddP expected");
796 }
797 j -= (oc1 - _resproj->outcnt());
798 }
799 }
800 if (_fallthroughcatchproj != NULL) {
801 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control));
802 }
803 if (_memproj_fallthrough != NULL) {
804 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory));
805 }
806 if (_memproj_catchall != NULL) {
807 _igvn.replace_node(_memproj_catchall, C->top());
808 }
809 if (_ioproj_fallthrough != NULL) {
810 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O));
811 }
812 if (_ioproj_catchall != NULL) {
813 _igvn.replace_node(_ioproj_catchall, C->top());
814 }
815 if (_catchallcatchproj != NULL) {
816 _igvn.replace_node(_catchallcatchproj, C->top());
817 }
818 }
819
820 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
821
822 if (!EliminateAllocations || !alloc->_is_scalar_replaceable) {
823 return false;
824 }
825
826 extract_call_projections(alloc);
827
828 GrowableArray <SafePointNode *> safepoints;
829 if (!can_eliminate_allocation(alloc, safepoints)) {
830 return false;
831 }
832
833 if (!scalar_replacement(alloc, safepoints)) {
834 return false;
835 }
836
837 process_users_of_allocation(alloc);
838
839 #ifndef PRODUCT
840 if (PrintEliminateAllocations) {
841 if (alloc->is_AllocateArray())
842 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
843 else
844 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
845 }
846 #endif
847
848 return true;
849 }
850
851
852 //---------------------------set_eden_pointers-------------------------
853 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) {
854 if (UseTLAB) { // Private allocation: load from TLS
855 Node* thread = transform_later(new (C, 1) ThreadLocalNode());
856 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset());
857 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset());
858 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset);
859 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset);
860 } else { // Shared allocation: load from globals
861 CollectedHeap* ch = Universe::heap();
862 address top_adr = (address)ch->top_addr();
863 address end_adr = (address)ch->end_addr();
864 eden_top_adr = makecon(TypeRawPtr::make(top_adr));
865 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr);
866 }
867 }
868
869
870 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
871 Node* adr = basic_plus_adr(base, offset);
872 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
873 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt);
874 transform_later(value);
875 return value;
876 }
877
878
879 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
880 Node* adr = basic_plus_adr(base, offset);
881 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt);
882 transform_later(mem);
883 return mem;
884 }
885
886 //=============================================================================
887 //
888 // A L L O C A T I O N
889 //
890 // Allocation attempts to be fast in the case of frequent small objects.
891 // It breaks down like this:
892 //
893 // 1) Size in doublewords is computed. This is a constant for objects and
894 // variable for most arrays. Doubleword units are used to avoid size
895 // overflow of huge doubleword arrays. We need doublewords in the end for
896 // rounding.
897 //
898 // 2) Size is checked for being 'too large'. Too-large allocations will go
899 // the slow path into the VM. The slow path can throw any required
900 // exceptions, and does all the special checks for very large arrays. The
901 // size test can constant-fold away for objects. For objects with
902 // finalizers it constant-folds the otherway: you always go slow with
903 // finalizers.
904 //
905 // 3) If NOT using TLABs, this is the contended loop-back point.
906 // Load-Locked the heap top. If using TLABs normal-load the heap top.
907 //
908 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
909 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
910 // "size*8" we always enter the VM, where "largish" is a constant picked small
911 // enough that there's always space between the eden max and 4Gig (old space is
912 // there so it's quite large) and large enough that the cost of entering the VM
913 // is dwarfed by the cost to initialize the space.
914 //
915 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
916 // down. If contended, repeat at step 3. If using TLABs normal-store
917 // adjusted heap top back down; there is no contention.
918 //
919 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
920 // fields.
921 //
922 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
923 // oop flavor.
924 //
925 //=============================================================================
926 // FastAllocateSizeLimit value is in DOUBLEWORDS.
927 // Allocations bigger than this always go the slow route.
928 // This value must be small enough that allocation attempts that need to
929 // trigger exceptions go the slow route. Also, it must be small enough so
930 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
931 //=============================================================================j//
932 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
933 // The allocator will coalesce int->oop copies away. See comment in
934 // coalesce.cpp about how this works. It depends critically on the exact
935 // code shape produced here, so if you are changing this code shape
936 // make sure the GC info for the heap-top is correct in and around the
937 // slow-path call.
938 //
939
940 void PhaseMacroExpand::expand_allocate_common(
941 AllocateNode* alloc, // allocation node to be expanded
942 Node* length, // array length for an array allocation
943 const TypeFunc* slow_call_type, // Type of slow call
944 address slow_call_address // Address of slow call
945 )
946 {
947
948 Node* ctrl = alloc->in(TypeFunc::Control);
949 Node* mem = alloc->in(TypeFunc::Memory);
950 Node* i_o = alloc->in(TypeFunc::I_O);
951 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
952 Node* klass_node = alloc->in(AllocateNode::KlassNode);
953 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
954
955 // With escape analysis, the entire memory state was needed to be able to
956 // eliminate the allocation. Since the allocations cannot be eliminated,
957 // optimize it to the raw slice.
958 if (mem->is_MergeMem()) {
959 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
960 }
961
962 assert(ctrl != NULL, "must have control");
963 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
964 // they will not be used if "always_slow" is set
965 enum { slow_result_path = 1, fast_result_path = 2 };
966 Node *result_region;
967 Node *result_phi_rawmem;
968 Node *result_phi_rawoop;
969 Node *result_phi_i_o;
970
971 // The initial slow comparison is a size check, the comparison
972 // we want to do is a BoolTest::gt
973 bool always_slow = false;
974 int tv = _igvn.find_int_con(initial_slow_test, -1);
975 if (tv >= 0) {
976 always_slow = (tv == 1);
977 initial_slow_test = NULL;
978 } else {
979 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
980 }
981
982 if (DTraceAllocProbes ||
983 !UseTLAB && (!Universe::heap()->supports_inline_contig_alloc() ||
984 (UseConcMarkSweepGC && CMSIncrementalMode))) {
985 // Force slow-path allocation
986 always_slow = true;
987 initial_slow_test = NULL;
988 }
989
990
991 enum { too_big_or_final_path = 1, need_gc_path = 2 };
992 Node *slow_region = NULL;
993 Node *toobig_false = ctrl;
994
995 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent");
996 // generate the initial test if necessary
997 if (initial_slow_test != NULL ) {
998 slow_region = new (C, 3) RegionNode(3);
999
1000 // Now make the initial failure test. Usually a too-big test but
1001 // might be a TRUE for finalizers or a fancy class check for
1002 // newInstance0.
1003 IfNode *toobig_iff = new (C, 2) IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1004 transform_later(toobig_iff);
1005 // Plug the failing-too-big test into the slow-path region
1006 Node *toobig_true = new (C, 1) IfTrueNode( toobig_iff );
1007 transform_later(toobig_true);
1008 slow_region ->init_req( too_big_or_final_path, toobig_true );
1009 toobig_false = new (C, 1) IfFalseNode( toobig_iff );
1010 transform_later(toobig_false);
1011 } else { // No initial test, just fall into next case
1012 toobig_false = ctrl;
1013 debug_only(slow_region = NodeSentinel);
1014 }
1015
1016 Node *slow_mem = mem; // save the current memory state for slow path
1017 // generate the fast allocation code unless we know that the initial test will always go slow
1018 if (!always_slow) {
1019 Node* eden_top_adr;
1020 Node* eden_end_adr;
1021
1022 set_eden_pointers(eden_top_adr, eden_end_adr);
1023
1024 // Load Eden::end. Loop invariant and hoisted.
1025 //
1026 // Note: We set the control input on "eden_end" and "old_eden_top" when using
1027 // a TLAB to work around a bug where these values were being moved across
1028 // a safepoint. These are not oops, so they cannot be include in the oop
1029 // map, but the can be changed by a GC. The proper way to fix this would
1030 // be to set the raw memory state when generating a SafepointNode. However
1031 // this will require extensive changes to the loop optimization in order to
1032 // prevent a degradation of the optimization.
1033 // See comment in memnode.hpp, around line 227 in class LoadPNode.
1034 Node *eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS);
1035
1036 // allocate the Region and Phi nodes for the result
1037 result_region = new (C, 3) RegionNode(3);
1038 result_phi_rawmem = new (C, 3) PhiNode( result_region, Type::MEMORY, TypeRawPtr::BOTTOM );
1039 result_phi_rawoop = new (C, 3) PhiNode( result_region, TypeRawPtr::BOTTOM );
1040 result_phi_i_o = new (C, 3) PhiNode( result_region, Type::ABIO ); // I/O is used for Prefetch
1041
1042 // We need a Region for the loop-back contended case.
1043 enum { fall_in_path = 1, contended_loopback_path = 2 };
1044 Node *contended_region;
1045 Node *contended_phi_rawmem;
1046 if( UseTLAB ) {
1047 contended_region = toobig_false;
1048 contended_phi_rawmem = mem;
1049 } else {
1050 contended_region = new (C, 3) RegionNode(3);
1051 contended_phi_rawmem = new (C, 3) PhiNode( contended_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1052 // Now handle the passing-too-big test. We fall into the contended
1053 // loop-back merge point.
1054 contended_region ->init_req( fall_in_path, toobig_false );
1055 contended_phi_rawmem->init_req( fall_in_path, mem );
1056 transform_later(contended_region);
1057 transform_later(contended_phi_rawmem);
1058 }
1059
1060 // Load(-locked) the heap top.
1061 // See note above concerning the control input when using a TLAB
1062 Node *old_eden_top = UseTLAB
1063 ? new (C, 3) LoadPNode ( ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM )
1064 : new (C, 3) LoadPLockedNode( contended_region, contended_phi_rawmem, eden_top_adr );
1065
1066 transform_later(old_eden_top);
1067 // Add to heap top to get a new heap top
1068 Node *new_eden_top = new (C, 4) AddPNode( top(), old_eden_top, size_in_bytes );
1069 transform_later(new_eden_top);
1070 // Check for needing a GC; compare against heap end
1071 Node *needgc_cmp = new (C, 3) CmpPNode( new_eden_top, eden_end );
1072 transform_later(needgc_cmp);
1073 Node *needgc_bol = new (C, 2) BoolNode( needgc_cmp, BoolTest::ge );
1074 transform_later(needgc_bol);
1075 IfNode *needgc_iff = new (C, 2) IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1076 transform_later(needgc_iff);
1077
1078 // Plug the failing-heap-space-need-gc test into the slow-path region
1079 Node *needgc_true = new (C, 1) IfTrueNode( needgc_iff );
1080 transform_later(needgc_true);
1081 if( initial_slow_test ) {
1082 slow_region ->init_req( need_gc_path, needgc_true );
1083 // This completes all paths into the slow merge point
1084 transform_later(slow_region);
1085 } else { // No initial slow path needed!
1086 // Just fall from the need-GC path straight into the VM call.
1087 slow_region = needgc_true;
1088 }
1089 // No need for a GC. Setup for the Store-Conditional
1090 Node *needgc_false = new (C, 1) IfFalseNode( needgc_iff );
1091 transform_later(needgc_false);
1092
1093 // Grab regular I/O before optional prefetch may change it.
1094 // Slow-path does no I/O so just set it to the original I/O.
1095 result_phi_i_o->init_req( slow_result_path, i_o );
1096
1097 i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem,
1098 old_eden_top, new_eden_top, length);
1099
1100 // Store (-conditional) the modified eden top back down.
1101 // StorePConditional produces flags for a test PLUS a modified raw
1102 // memory state.
1103 Node *store_eden_top;
1104 Node *fast_oop_ctrl;
1105 if( UseTLAB ) {
1106 store_eden_top = new (C, 4) StorePNode( needgc_false, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, new_eden_top );
1107 transform_later(store_eden_top);
1108 fast_oop_ctrl = needgc_false; // No contention, so this is the fast path
1109 } else {
1110 store_eden_top = new (C, 5) StorePConditionalNode( needgc_false, contended_phi_rawmem, eden_top_adr, new_eden_top, old_eden_top );
1111 transform_later(store_eden_top);
1112 Node *contention_check = new (C, 2) BoolNode( store_eden_top, BoolTest::ne );
1113 transform_later(contention_check);
1114 store_eden_top = new (C, 1) SCMemProjNode(store_eden_top);
1115 transform_later(store_eden_top);
1116
1117 // If not using TLABs, check to see if there was contention.
1118 IfNode *contention_iff = new (C, 2) IfNode ( needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN );
1119 transform_later(contention_iff);
1120 Node *contention_true = new (C, 1) IfTrueNode( contention_iff );
1121 transform_later(contention_true);
1122 // If contention, loopback and try again.
1123 contended_region->init_req( contended_loopback_path, contention_true );
1124 contended_phi_rawmem->init_req( contended_loopback_path, store_eden_top );
1125
1126 // Fast-path succeeded with no contention!
1127 Node *contention_false = new (C, 1) IfFalseNode( contention_iff );
1128 transform_later(contention_false);
1129 fast_oop_ctrl = contention_false;
1130 }
1131
1132 // Rename successful fast-path variables to make meaning more obvious
1133 Node* fast_oop = old_eden_top;
1134 Node* fast_oop_rawmem = store_eden_top;
1135 fast_oop_rawmem = initialize_object(alloc,
1136 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1137 klass_node, length, size_in_bytes);
1138
1139 if (ExtendedDTraceProbes) {
1140 // Slow-path call
1141 int size = TypeFunc::Parms + 2;
1142 CallLeafNode *call = new (C, size) CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1143 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base),
1144 "dtrace_object_alloc",
1145 TypeRawPtr::BOTTOM);
1146
1147 // Get base of thread-local storage area
1148 Node* thread = new (C, 1) ThreadLocalNode();
1149 transform_later(thread);
1150
1151 call->init_req(TypeFunc::Parms+0, thread);
1152 call->init_req(TypeFunc::Parms+1, fast_oop);
1153 call->init_req( TypeFunc::Control, fast_oop_ctrl );
1154 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o
1155 call->init_req( TypeFunc::Memory , fast_oop_rawmem );
1156 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) );
1157 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) );
1158 transform_later(call);
1159 fast_oop_ctrl = new (C, 1) ProjNode(call,TypeFunc::Control);
1160 transform_later(fast_oop_ctrl);
1161 fast_oop_rawmem = new (C, 1) ProjNode(call,TypeFunc::Memory);
1162 transform_later(fast_oop_rawmem);
1163 }
1164
1165 // Plug in the successful fast-path into the result merge point
1166 result_region ->init_req( fast_result_path, fast_oop_ctrl );
1167 result_phi_rawoop->init_req( fast_result_path, fast_oop );
1168 result_phi_i_o ->init_req( fast_result_path, i_o );
1169 result_phi_rawmem->init_req( fast_result_path, fast_oop_rawmem );
1170 } else {
1171 slow_region = ctrl;
1172 }
1173
1174 // Generate slow-path call
1175 CallNode *call = new (C, slow_call_type->domain()->cnt())
1176 CallStaticJavaNode(slow_call_type, slow_call_address,
1177 OptoRuntime::stub_name(slow_call_address),
1178 alloc->jvms()->bci(),
1179 TypePtr::BOTTOM);
1180 call->init_req( TypeFunc::Control, slow_region );
1181 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o
1182 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs
1183 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) );
1184 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) );
1185
1186 call->init_req(TypeFunc::Parms+0, klass_node);
1187 if (length != NULL) {
1188 call->init_req(TypeFunc::Parms+1, length);
1189 }
1190
1191 // Copy debug information and adjust JVMState information, then replace
1192 // allocate node with the call
1193 copy_call_debug_info((CallNode *) alloc, call);
1194 if (!always_slow) {
1195 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1196 }
1197 _igvn.hash_delete(alloc);
1198 _igvn.subsume_node(alloc, call);
1199 transform_later(call);
1200
1201 // Identify the output projections from the allocate node and
1202 // adjust any references to them.
1203 // The control and io projections look like:
1204 //
1205 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1206 // Allocate Catch
1207 // ^---Proj(io) <-------+ ^---CatchProj(io)
1208 //
1209 // We are interested in the CatchProj nodes.
1210 //
1211 extract_call_projections(call);
1212
1213 // An allocate node has separate memory projections for the uses on the control and i_o paths
1214 // Replace uses of the control memory projection with result_phi_rawmem (unless we are only generating a slow call)
1215 if (!always_slow && _memproj_fallthrough != NULL) {
1216 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) {
1217 Node *use = _memproj_fallthrough->fast_out(i);
1218 _igvn.hash_delete(use);
1219 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem);
1220 _igvn._worklist.push(use);
1221 // back up iterator
1222 --i;
1223 }
1224 }
1225 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete _memproj_catchall so
1226 // we end up with a call that has only 1 memory projection
1227 if (_memproj_catchall != NULL ) {
1228 if (_memproj_fallthrough == NULL) {
1229 _memproj_fallthrough = new (C, 1) ProjNode(call, TypeFunc::Memory);
1230 transform_later(_memproj_fallthrough);
1231 }
1232 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) {
1233 Node *use = _memproj_catchall->fast_out(i);
1234 _igvn.hash_delete(use);
1235 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough);
1236 _igvn._worklist.push(use);
1237 // back up iterator
1238 --i;
1239 }
1240 }
1241
1242 mem = result_phi_rawmem;
1243
1244 // An allocate node has separate i_o projections for the uses on the control and i_o paths
1245 // Replace uses of the control i_o projection with result_phi_i_o (unless we are only generating a slow call)
1246 if (_ioproj_fallthrough == NULL) {
1247 _ioproj_fallthrough = new (C, 1) ProjNode(call, TypeFunc::I_O);
1248 transform_later(_ioproj_fallthrough);
1249 } else if (!always_slow) {
1250 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) {
1251 Node *use = _ioproj_fallthrough->fast_out(i);
1252
1253 _igvn.hash_delete(use);
1254 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o);
1255 _igvn._worklist.push(use);
1256 // back up iterator
1257 --i;
1258 }
1259 }
1260 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete _ioproj_catchall so
1261 // we end up with a call that has only 1 control projection
1262 if (_ioproj_catchall != NULL ) {
1263 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) {
1264 Node *use = _ioproj_catchall->fast_out(i);
1265 _igvn.hash_delete(use);
1266 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough);
1267 _igvn._worklist.push(use);
1268 // back up iterator
1269 --i;
1270 }
1271 }
1272
1273 // if we generated only a slow call, we are done
1274 if (always_slow)
1275 return;
1276
1277
1278 if (_fallthroughcatchproj != NULL) {
1279 ctrl = _fallthroughcatchproj->clone();
1280 transform_later(ctrl);
1281 _igvn.hash_delete(_fallthroughcatchproj);
1282 _igvn.subsume_node(_fallthroughcatchproj, result_region);
1283 } else {
1284 ctrl = top();
1285 }
1286 Node *slow_result;
1287 if (_resproj == NULL) {
1288 // no uses of the allocation result
1289 slow_result = top();
1290 } else {
1291 slow_result = _resproj->clone();
1292 transform_later(slow_result);
1293 _igvn.hash_delete(_resproj);
1294 _igvn.subsume_node(_resproj, result_phi_rawoop);
1295 }
1296
1297 // Plug slow-path into result merge point
1298 result_region ->init_req( slow_result_path, ctrl );
1299 result_phi_rawoop->init_req( slow_result_path, slow_result);
1300 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough );
1301 transform_later(result_region);
1302 transform_later(result_phi_rawoop);
1303 transform_later(result_phi_rawmem);
1304 transform_later(result_phi_i_o);
1305 // This completes all paths into the result merge point
1306 }
1307
1308
1309 // Helper for PhaseMacroExpand::expand_allocate_common.
1310 // Initializes the newly-allocated storage.
1311 Node*
1312 PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1313 Node* control, Node* rawmem, Node* object,
1314 Node* klass_node, Node* length,
1315 Node* size_in_bytes) {
1316 InitializeNode* init = alloc->initialization();
1317 // Store the klass & mark bits
1318 Node* mark_node = NULL;
1319 // For now only enable fast locking for non-array types
1320 if (UseBiasedLocking && (length == NULL)) {
1321 mark_node = make_load(NULL, rawmem, klass_node, Klass::prototype_header_offset_in_bytes() + sizeof(oopDesc), TypeRawPtr::BOTTOM, T_ADDRESS);
1322 } else {
1323 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype()));
1324 }
1325 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
1326
1327 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_OBJECT);
1328 int header_size = alloc->minimum_header_size(); // conservatively small
1329
1330 // Array length
1331 if (length != NULL) { // Arrays need length field
1332 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1333 // conservatively small header size:
1334 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1335 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1336 if (k->is_array_klass()) // we know the exact header size in most cases:
1337 header_size = Klass::layout_helper_header_size(k->layout_helper());
1338 }
1339
1340 // Clear the object body, if necessary.
1341 if (init == NULL) {
1342 // The init has somehow disappeared; be cautious and clear everything.
1343 //
1344 // This can happen if a node is allocated but an uncommon trap occurs
1345 // immediately. In this case, the Initialize gets associated with the
1346 // trap, and may be placed in a different (outer) loop, if the Allocate
1347 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1348 // there can be two Allocates to one Initialize. The answer in all these
1349 // edge cases is safety first. It is always safe to clear immediately
1350 // within an Allocate, and then (maybe or maybe not) clear some more later.
1351 if (!ZeroTLAB)
1352 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1353 header_size, size_in_bytes,
1354 &_igvn);
1355 } else {
1356 if (!init->is_complete()) {
1357 // Try to win by zeroing only what the init does not store.
1358 // We can also try to do some peephole optimizations,
1359 // such as combining some adjacent subword stores.
1360 rawmem = init->complete_stores(control, rawmem, object,
1361 header_size, size_in_bytes, &_igvn);
1362 }
1363 // We have no more use for this link, since the AllocateNode goes away:
1364 init->set_req(InitializeNode::RawAddress, top());
1365 // (If we keep the link, it just confuses the register allocator,
1366 // who thinks he sees a real use of the address by the membar.)
1367 }
1368
1369 return rawmem;
1370 }
1371
1372 // Generate prefetch instructions for next allocations.
1373 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
1374 Node*& contended_phi_rawmem,
1375 Node* old_eden_top, Node* new_eden_top,
1376 Node* length) {
1377 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
1378 // Generate prefetch allocation with watermark check.
1379 // As an allocation hits the watermark, we will prefetch starting
1380 // at a "distance" away from watermark.
1381 enum { fall_in_path = 1, pf_path = 2 };
1382
1383 Node *pf_region = new (C, 3) RegionNode(3);
1384 Node *pf_phi_rawmem = new (C, 3) PhiNode( pf_region, Type::MEMORY,
1385 TypeRawPtr::BOTTOM );
1386 // I/O is used for Prefetch
1387 Node *pf_phi_abio = new (C, 3) PhiNode( pf_region, Type::ABIO );
1388
1389 Node *thread = new (C, 1) ThreadLocalNode();
1390 transform_later(thread);
1391
1392 Node *eden_pf_adr = new (C, 4) AddPNode( top()/*not oop*/, thread,
1393 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
1394 transform_later(eden_pf_adr);
1395
1396 Node *old_pf_wm = new (C, 3) LoadPNode( needgc_false,
1397 contended_phi_rawmem, eden_pf_adr,
1398 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM );
1399 transform_later(old_pf_wm);
1400
1401 // check against new_eden_top
1402 Node *need_pf_cmp = new (C, 3) CmpPNode( new_eden_top, old_pf_wm );
1403 transform_later(need_pf_cmp);
1404 Node *need_pf_bol = new (C, 2) BoolNode( need_pf_cmp, BoolTest::ge );
1405 transform_later(need_pf_bol);
1406 IfNode *need_pf_iff = new (C, 2) IfNode( needgc_false, need_pf_bol,
1407 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1408 transform_later(need_pf_iff);
1409
1410 // true node, add prefetchdistance
1411 Node *need_pf_true = new (C, 1) IfTrueNode( need_pf_iff );
1412 transform_later(need_pf_true);
1413
1414 Node *need_pf_false = new (C, 1) IfFalseNode( need_pf_iff );
1415 transform_later(need_pf_false);
1416
1417 Node *new_pf_wmt = new (C, 4) AddPNode( top(), old_pf_wm,
1418 _igvn.MakeConX(AllocatePrefetchDistance) );
1419 transform_later(new_pf_wmt );
1420 new_pf_wmt->set_req(0, need_pf_true);
1421
1422 Node *store_new_wmt = new (C, 4) StorePNode( need_pf_true,
1423 contended_phi_rawmem, eden_pf_adr,
1424 TypeRawPtr::BOTTOM, new_pf_wmt );
1425 transform_later(store_new_wmt);
1426
1427 // adding prefetches
1428 pf_phi_abio->init_req( fall_in_path, i_o );
1429
1430 Node *prefetch_adr;
1431 Node *prefetch;
1432 uint lines = AllocatePrefetchDistance / AllocatePrefetchStepSize;
1433 uint step_size = AllocatePrefetchStepSize;
1434 uint distance = 0;
1435
1436 for ( uint i = 0; i < lines; i++ ) {
1437 prefetch_adr = new (C, 4) AddPNode( old_pf_wm, new_pf_wmt,
1438 _igvn.MakeConX(distance) );
1439 transform_later(prefetch_adr);
1440 prefetch = new (C, 3) PrefetchWriteNode( i_o, prefetch_adr );
1441 transform_later(prefetch);
1442 distance += step_size;
1443 i_o = prefetch;
1444 }
1445 pf_phi_abio->set_req( pf_path, i_o );
1446
1447 pf_region->init_req( fall_in_path, need_pf_false );
1448 pf_region->init_req( pf_path, need_pf_true );
1449
1450 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
1451 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
1452
1453 transform_later(pf_region);
1454 transform_later(pf_phi_rawmem);
1455 transform_later(pf_phi_abio);
1456
1457 needgc_false = pf_region;
1458 contended_phi_rawmem = pf_phi_rawmem;
1459 i_o = pf_phi_abio;
1460 } else if( AllocatePrefetchStyle > 0 ) {
1461 // Insert a prefetch for each allocation only on the fast-path
1462 Node *prefetch_adr;
1463 Node *prefetch;
1464 // Generate several prefetch instructions only for arrays.
1465 uint lines = (length != NULL) ? AllocatePrefetchLines : 1;
1466 uint step_size = AllocatePrefetchStepSize;
1467 uint distance = AllocatePrefetchDistance;
1468 for ( uint i = 0; i < lines; i++ ) {
1469 prefetch_adr = new (C, 4) AddPNode( old_eden_top, new_eden_top,
1470 _igvn.MakeConX(distance) );
1471 transform_later(prefetch_adr);
1472 prefetch = new (C, 3) PrefetchWriteNode( i_o, prefetch_adr );
1473 // Do not let it float too high, since if eden_top == eden_end,
1474 // both might be null.
1475 if( i == 0 ) { // Set control for first prefetch, next follows it
1476 prefetch->init_req(0, needgc_false);
1477 }
1478 transform_later(prefetch);
1479 distance += step_size;
1480 i_o = prefetch;
1481 }
1482 }
1483 return i_o;
1484 }
1485
1486
1487 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
1488 expand_allocate_common(alloc, NULL,
1489 OptoRuntime::new_instance_Type(),
1490 OptoRuntime::new_instance_Java());
1491 }
1492
1493 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
1494 Node* length = alloc->in(AllocateNode::ALength);
1495 expand_allocate_common(alloc, length,
1496 OptoRuntime::new_array_Type(),
1497 OptoRuntime::new_array_Java());
1498 }
1499
1500
1501 // we have determined that this lock/unlock can be eliminated, we simply
1502 // eliminate the node without expanding it.
1503 //
1504 // Note: The membar's associated with the lock/unlock are currently not
1505 // eliminated. This should be investigated as a future enhancement.
1506 //
1507 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
1508
1509 if (!alock->is_eliminated()) {
1510 return false;
1511 }
1512 if (alock->is_Lock() && !alock->is_coarsened()) {
1513 // Create new "eliminated" BoxLock node and use it
1514 // in monitor debug info for the same object.
1515 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
1516 Node* obj = alock->obj_node();
1517 if (!oldbox->is_eliminated()) {
1518 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
1519 newbox->set_eliminated();
1520 transform_later(newbox);
1521 // Replace old box node with new box for all users
1522 // of the same object.
1523 for (uint i = 0; i < oldbox->outcnt();) {
1524
1525 bool next_edge = true;
1526 Node* u = oldbox->raw_out(i);
1527 if (u == alock) {
1528 i++;
1529 continue; // It will be removed below
1530 }
1531 if (u->is_Lock() &&
1532 u->as_Lock()->obj_node() == obj &&
1533 // oldbox could be referenced in debug info also
1534 u->as_Lock()->box_node() == oldbox) {
1535 assert(u->as_Lock()->is_eliminated(), "sanity");
1536 _igvn.hash_delete(u);
1537 u->set_req(TypeFunc::Parms + 1, newbox);
1538 next_edge = false;
1539 #ifdef ASSERT
1540 } else if (u->is_Unlock() && u->as_Unlock()->obj_node() == obj) {
1541 assert(u->as_Unlock()->is_eliminated(), "sanity");
1542 #endif
1543 }
1544 // Replace old box in monitor debug info.
1545 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
1546 SafePointNode* sfn = u->as_SafePoint();
1547 JVMState* youngest_jvms = sfn->jvms();
1548 int max_depth = youngest_jvms->depth();
1549 for (int depth = 1; depth <= max_depth; depth++) {
1550 JVMState* jvms = youngest_jvms->of_depth(depth);
1551 int num_mon = jvms->nof_monitors();
1552 // Loop over monitors
1553 for (int idx = 0; idx < num_mon; idx++) {
1554 Node* obj_node = sfn->monitor_obj(jvms, idx);
1555 Node* box_node = sfn->monitor_box(jvms, idx);
1556 if (box_node == oldbox && obj_node == obj) {
1557 int j = jvms->monitor_box_offset(idx);
1558 _igvn.hash_delete(u);
1559 u->set_req(j, newbox);
1560 next_edge = false;
1561 }
1562 } // for (int idx = 0;
1563 } // for (int depth = 1;
1564 } // if (u->is_SafePoint()
1565 if (next_edge) i++;
1566 } // for (uint i = 0; i < oldbox->outcnt();)
1567 } // if (!oldbox->is_eliminated())
1568 } // if (alock->is_Lock() && !lock->is_coarsened())
1569
1570 #ifndef PRODUCT
1571 if (PrintEliminateLocks) {
1572 if (alock->is_Lock()) {
1573 tty->print_cr("++++ Eliminating: %d Lock", alock->_idx);
1574 } else {
1575 tty->print_cr("++++ Eliminating: %d Unlock", alock->_idx);
1576 }
1577 }
1578 #endif
1579
1580 Node* mem = alock->in(TypeFunc::Memory);
1581 Node* ctrl = alock->in(TypeFunc::Control);
1582
1583 extract_call_projections(alock);
1584 // There are 2 projections from the lock. The lock node will
1585 // be deleted when its last use is subsumed below.
1586 assert(alock->outcnt() == 2 &&
1587 _fallthroughproj != NULL &&
1588 _memproj_fallthrough != NULL,
1589 "Unexpected projections from Lock/Unlock");
1590
1591 Node* fallthroughproj = _fallthroughproj;
1592 Node* memproj_fallthrough = _memproj_fallthrough;
1593
1594 // The memory projection from a lock/unlock is RawMem
1595 // The input to a Lock is merged memory, so extract its RawMem input
1596 // (unless the MergeMem has been optimized away.)
1597 if (alock->is_Lock()) {
1598 // Seach for MemBarAcquire node and delete it also.
1599 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
1600 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquire, "");
1601 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
1602 Node* memproj = membar->proj_out(TypeFunc::Memory);
1603 _igvn.hash_delete(ctrlproj);
1604 _igvn.subsume_node(ctrlproj, fallthroughproj);
1605 _igvn.hash_delete(memproj);
1606 _igvn.subsume_node(memproj, memproj_fallthrough);
1607
1608 // Delete FastLock node also if this Lock node is unique user
1609 // (a loop peeling may clone a Lock node).
1610 Node* flock = alock->as_Lock()->fastlock_node();
1611 if (flock->outcnt() == 1) {
1612 assert(flock->unique_out() == alock, "sanity");
1613 _igvn.hash_delete(flock);
1614 _igvn.subsume_node(flock, top());
1615 }
1616 }
1617
1618 // Seach for MemBarRelease node and delete it also.
1619 if (alock->is_Unlock() && ctrl != NULL && ctrl->is_Proj() &&
1620 ctrl->in(0)->is_MemBar()) {
1621 MemBarNode* membar = ctrl->in(0)->as_MemBar();
1622 assert(membar->Opcode() == Op_MemBarRelease &&
1623 mem->is_Proj() && membar == mem->in(0), "");
1624 _igvn.hash_delete(fallthroughproj);
1625 _igvn.subsume_node(fallthroughproj, ctrl);
1626 _igvn.hash_delete(memproj_fallthrough);
1627 _igvn.subsume_node(memproj_fallthrough, mem);
1628 fallthroughproj = ctrl;
1629 memproj_fallthrough = mem;
1630 ctrl = membar->in(TypeFunc::Control);
1631 mem = membar->in(TypeFunc::Memory);
1632 }
1633
1634 _igvn.hash_delete(fallthroughproj);
1635 _igvn.subsume_node(fallthroughproj, ctrl);
1636 _igvn.hash_delete(memproj_fallthrough);
1637 _igvn.subsume_node(memproj_fallthrough, mem);
1638 return true;
1639 }
1640
1641
1642 //------------------------------expand_lock_node----------------------
1643 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
1644
1645 Node* ctrl = lock->in(TypeFunc::Control);
1646 Node* mem = lock->in(TypeFunc::Memory);
1647 Node* obj = lock->obj_node();
1648 Node* box = lock->box_node();
1649 Node* flock = lock->fastlock_node();
1650
1651 // Make the merge point
1652 Node *region;
1653 Node *mem_phi;
1654 Node *slow_path;
1655
1656 if (UseOptoBiasInlining) {
1657 /*
1658 * See the full descrition in MacroAssembler::biased_locking_enter().
1659 *
1660 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) {
1661 * // The object is biased.
1662 * proto_node = klass->prototype_header;
1663 * o_node = thread | proto_node;
1664 * x_node = o_node ^ mark_word;
1665 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ?
1666 * // Done.
1667 * } else {
1668 * if( (x_node & biased_lock_mask) != 0 ) {
1669 * // The klass's prototype header is no longer biased.
1670 * cas(&mark_word, mark_word, proto_node)
1671 * goto cas_lock;
1672 * } else {
1673 * // The klass's prototype header is still biased.
1674 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch?
1675 * old = mark_word;
1676 * new = o_node;
1677 * } else {
1678 * // Different thread or anonymous biased.
1679 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask);
1680 * new = thread | old;
1681 * }
1682 * // Try to rebias.
1683 * if( cas(&mark_word, old, new) == 0 ) {
1684 * // Done.
1685 * } else {
1686 * goto slow_path; // Failed.
1687 * }
1688 * }
1689 * }
1690 * } else {
1691 * // The object is not biased.
1692 * cas_lock:
1693 * if( FastLock(obj) == 0 ) {
1694 * // Done.
1695 * } else {
1696 * slow_path:
1697 * OptoRuntime::complete_monitor_locking_Java(obj);
1698 * }
1699 * }
1700 */
1701
1702 region = new (C, 5) RegionNode(5);
1703 // create a Phi for the memory state
1704 mem_phi = new (C, 5) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
1705
1706 Node* fast_lock_region = new (C, 3) RegionNode(3);
1707 Node* fast_lock_mem_phi = new (C, 3) PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1708
1709 // First, check mark word for the biased lock pattern.
1710 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
1711
1712 // Get fast path - mark word has the biased lock pattern.
1713 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node,
1714 markOopDesc::biased_lock_mask_in_place,
1715 markOopDesc::biased_lock_pattern, true);
1716 // fast_lock_region->in(1) is set to slow path.
1717 fast_lock_mem_phi->init_req(1, mem);
1718
1719 // Now check that the lock is biased to the current thread and has
1720 // the same epoch and bias as Klass::_prototype_header.
1721
1722 // Special-case a fresh allocation to avoid building nodes:
1723 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn);
1724 if (klass_node == NULL) {
1725 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1726 klass_node = transform_later( LoadKlassNode::make(_igvn, mem, k_adr, _igvn.type(k_adr)->is_ptr()) );
1727 }
1728 Node *proto_node = make_load(ctrl, mem, klass_node, Klass::prototype_header_offset_in_bytes() + sizeof(oopDesc), TypeX_X, TypeX_X->basic_type());
1729
1730 Node* thread = transform_later(new (C, 1) ThreadLocalNode());
1731 Node* cast_thread = transform_later(new (C, 2) CastP2XNode(ctrl, thread));
1732 Node* o_node = transform_later(new (C, 3) OrXNode(cast_thread, proto_node));
1733 Node* x_node = transform_later(new (C, 3) XorXNode(o_node, mark_node));
1734
1735 // Get slow path - mark word does NOT match the value.
1736 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node,
1737 (~markOopDesc::age_mask_in_place), 0);
1738 // region->in(3) is set to fast path - the object is biased to the current thread.
1739 mem_phi->init_req(3, mem);
1740
1741
1742 // Mark word does NOT match the value (thread | Klass::_prototype_header).
1743
1744
1745 // First, check biased pattern.
1746 // Get fast path - _prototype_header has the same biased lock pattern.
1747 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node,
1748 markOopDesc::biased_lock_mask_in_place, 0, true);
1749
1750 not_biased_ctrl = fast_lock_region->in(2); // Slow path
1751 // fast_lock_region->in(2) - the prototype header is no longer biased
1752 // and we have to revoke the bias on this object.
1753 // We are going to try to reset the mark of this object to the prototype
1754 // value and fall through to the CAS-based locking scheme.
1755 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
1756 Node* cas = new (C, 5) StoreXConditionalNode(not_biased_ctrl, mem, adr,
1757 proto_node, mark_node);
1758 transform_later(cas);
1759 Node* proj = transform_later( new (C, 1) SCMemProjNode(cas));
1760 fast_lock_mem_phi->init_req(2, proj);
1761
1762
1763 // Second, check epoch bits.
1764 Node* rebiased_region = new (C, 3) RegionNode(3);
1765 Node* old_phi = new (C, 3) PhiNode( rebiased_region, TypeX_X);
1766 Node* new_phi = new (C, 3) PhiNode( rebiased_region, TypeX_X);
1767
1768 // Get slow path - mark word does NOT match epoch bits.
1769 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node,
1770 markOopDesc::epoch_mask_in_place, 0);
1771 // The epoch of the current bias is not valid, attempt to rebias the object
1772 // toward the current thread.
1773 rebiased_region->init_req(2, epoch_ctrl);
1774 old_phi->init_req(2, mark_node);
1775 new_phi->init_req(2, o_node);
1776
1777 // rebiased_region->in(1) is set to fast path.
1778 // The epoch of the current bias is still valid but we know
1779 // nothing about the owner; it might be set or it might be clear.
1780 Node* cmask = MakeConX(markOopDesc::biased_lock_mask_in_place |
1781 markOopDesc::age_mask_in_place |
1782 markOopDesc::epoch_mask_in_place);
1783 Node* old = transform_later(new (C, 3) AndXNode(mark_node, cmask));
1784 cast_thread = transform_later(new (C, 2) CastP2XNode(ctrl, thread));
1785 Node* new_mark = transform_later(new (C, 3) OrXNode(cast_thread, old));
1786 old_phi->init_req(1, old);
1787 new_phi->init_req(1, new_mark);
1788
1789 transform_later(rebiased_region);
1790 transform_later(old_phi);
1791 transform_later(new_phi);
1792
1793 // Try to acquire the bias of the object using an atomic operation.
1794 // If this fails we will go in to the runtime to revoke the object's bias.
1795 cas = new (C, 5) StoreXConditionalNode(rebiased_region, mem, adr,
1796 new_phi, old_phi);
1797 transform_later(cas);
1798 proj = transform_later( new (C, 1) SCMemProjNode(cas));
1799
1800 // Get slow path - Failed to CAS.
1801 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0);
1802 mem_phi->init_req(4, proj);
1803 // region->in(4) is set to fast path - the object is rebiased to the current thread.
1804
1805 // Failed to CAS.
1806 slow_path = new (C, 3) RegionNode(3);
1807 Node *slow_mem = new (C, 3) PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM);
1808
1809 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control
1810 slow_mem->init_req(1, proj);
1811
1812 // Call CAS-based locking scheme (FastLock node).
1813
1814 transform_later(fast_lock_region);
1815 transform_later(fast_lock_mem_phi);
1816
1817 // Get slow path - FastLock failed to lock the object.
1818 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0);
1819 mem_phi->init_req(2, fast_lock_mem_phi);
1820 // region->in(2) is set to fast path - the object is locked to the current thread.
1821
1822 slow_path->init_req(2, ctrl); // Capture slow-control
1823 slow_mem->init_req(2, fast_lock_mem_phi);
1824
1825 transform_later(slow_path);
1826 transform_later(slow_mem);
1827 // Reset lock's memory edge.
1828 lock->set_req(TypeFunc::Memory, slow_mem);
1829
1830 } else {
1831 region = new (C, 3) RegionNode(3);
1832 // create a Phi for the memory state
1833 mem_phi = new (C, 3) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
1834
1835 // Optimize test; set region slot 2
1836 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
1837 mem_phi->init_req(2, mem);
1838 }
1839
1840 // Make slow path call
1841 CallNode *call = make_slow_call( (CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(), OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path, obj, box );
1842
1843 extract_call_projections(call);
1844
1845 // Slow path can only throw asynchronous exceptions, which are always
1846 // de-opted. So the compiler thinks the slow-call can never throw an
1847 // exception. If it DOES throw an exception we would need the debug
1848 // info removed first (since if it throws there is no monitor).
1849 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
1850 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
1851
1852 // Capture slow path
1853 // disconnect fall-through projection from call and create a new one
1854 // hook up users of fall-through projection to region
1855 Node *slow_ctrl = _fallthroughproj->clone();
1856 transform_later(slow_ctrl);
1857 _igvn.hash_delete(_fallthroughproj);
1858 _fallthroughproj->disconnect_inputs(NULL);
1859 region->init_req(1, slow_ctrl);
1860 // region inputs are now complete
1861 transform_later(region);
1862 _igvn.subsume_node(_fallthroughproj, region);
1863
1864 Node *memproj = transform_later( new(C, 1) ProjNode(call, TypeFunc::Memory) );
1865 mem_phi->init_req(1, memproj );
1866 transform_later(mem_phi);
1867 _igvn.hash_delete(_memproj_fallthrough);
1868 _igvn.subsume_node(_memproj_fallthrough, mem_phi);
1869 }
1870
1871 //------------------------------expand_unlock_node----------------------
1872 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
1873
1874 Node* ctrl = unlock->in(TypeFunc::Control);
1875 Node* mem = unlock->in(TypeFunc::Memory);
1876 Node* obj = unlock->obj_node();
1877 Node* box = unlock->box_node();
1878
1879 // No need for a null check on unlock
1880
1881 // Make the merge point
1882 Node *region;
1883 Node *mem_phi;
1884
1885 if (UseOptoBiasInlining) {
1886 // Check for biased locking unlock case, which is a no-op.
1887 // See the full descrition in MacroAssembler::biased_locking_exit().
1888 region = new (C, 4) RegionNode(4);
1889 // create a Phi for the memory state
1890 mem_phi = new (C, 4) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
1891 mem_phi->init_req(3, mem);
1892
1893 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
1894 ctrl = opt_bits_test(ctrl, region, 3, mark_node,
1895 markOopDesc::biased_lock_mask_in_place,
1896 markOopDesc::biased_lock_pattern);
1897 } else {
1898 region = new (C, 3) RegionNode(3);
1899 // create a Phi for the memory state
1900 mem_phi = new (C, 3) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
1901 }
1902
1903 FastUnlockNode *funlock = new (C, 3) FastUnlockNode( ctrl, obj, box );
1904 funlock = transform_later( funlock )->as_FastUnlock();
1905 // Optimize test; set region slot 2
1906 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
1907
1908 CallNode *call = make_slow_call( (CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C), "complete_monitor_unlocking_C", slow_path, obj, box );
1909
1910 extract_call_projections(call);
1911
1912 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
1913 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
1914
1915 // No exceptions for unlocking
1916 // Capture slow path
1917 // disconnect fall-through projection from call and create a new one
1918 // hook up users of fall-through projection to region
1919 Node *slow_ctrl = _fallthroughproj->clone();
1920 transform_later(slow_ctrl);
1921 _igvn.hash_delete(_fallthroughproj);
1922 _fallthroughproj->disconnect_inputs(NULL);
1923 region->init_req(1, slow_ctrl);
1924 // region inputs are now complete
1925 transform_later(region);
1926 _igvn.subsume_node(_fallthroughproj, region);
1927
1928 Node *memproj = transform_later( new(C, 1) ProjNode(call, TypeFunc::Memory) );
1929 mem_phi->init_req(1, memproj );
1930 mem_phi->init_req(2, mem);
1931 transform_later(mem_phi);
1932 _igvn.hash_delete(_memproj_fallthrough);
1933 _igvn.subsume_node(_memproj_fallthrough, mem_phi);
1934 }
1935
1936 //------------------------------expand_macro_nodes----------------------
1937 // Returns true if a failure occurred.
1938 bool PhaseMacroExpand::expand_macro_nodes() {
1939 if (C->macro_count() == 0)
1940 return false;
1941 // First, attempt to eliminate locks
1942 bool progress = true;
1943 while (progress) {
1944 progress = false;
1945 for (int i = C->macro_count(); i > 0; i--) {
1946 Node * n = C->macro_node(i-1);
1947 bool success = false;
1948 debug_only(int old_macro_count = C->macro_count(););
1949 if (n->is_AbstractLock()) {
1950 success = eliminate_locking_node(n->as_AbstractLock());
1951 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) {
1952 _igvn.add_users_to_worklist(n);
1953 _igvn.hash_delete(n);
1954 _igvn.subsume_node(n, n->in(1));
1955 success = true;
1956 }
1957 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
1958 progress = progress || success;
1959 }
1960 }
1961 // Next, attempt to eliminate allocations
1962 progress = true;
1963 while (progress) {
1964 progress = false;
1965 for (int i = C->macro_count(); i > 0; i--) {
1966 Node * n = C->macro_node(i-1);
1967 bool success = false;
1968 debug_only(int old_macro_count = C->macro_count(););
1969 switch (n->class_id()) {
1970 case Node::Class_Allocate:
1971 case Node::Class_AllocateArray:
1972 success = eliminate_allocate_node(n->as_Allocate());
1973 break;
1974 case Node::Class_Lock:
1975 case Node::Class_Unlock:
1976 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
1977 break;
1978 default:
1979 assert(false, "unknown node type in macro list");
1980 }
1981 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
1982 progress = progress || success;
1983 }
1984 }
1985 // Make sure expansion will not cause node limit to be exceeded.
1986 // Worst case is a macro node gets expanded into about 50 nodes.
1987 // Allow 50% more for optimization.
1988 if (C->check_node_count(C->macro_count() * 75, "out of nodes before macro expansion" ) )
1989 return true;
1990
1991 // expand "macro" nodes
1992 // nodes are removed from the macro list as they are processed
1993 while (C->macro_count() > 0) {
1994 int macro_count = C->macro_count();
1995 Node * n = C->macro_node(macro_count-1);
1996 assert(n->is_macro(), "only macro nodes expected here");
1997 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) {
1998 // node is unreachable, so don't try to expand it
1999 C->remove_macro_node(n);
2000 continue;
2001 }
2002 switch (n->class_id()) {
2003 case Node::Class_Allocate:
2004 expand_allocate(n->as_Allocate());
2005 break;
2006 case Node::Class_AllocateArray:
2007 expand_allocate_array(n->as_AllocateArray());
2008 break;
2009 case Node::Class_Lock:
2010 expand_lock_node(n->as_Lock());
2011 break;
2012 case Node::Class_Unlock:
2013 expand_unlock_node(n->as_Unlock());
2014 break;
2015 default:
2016 assert(false, "unknown node type in macro list");
2017 }
2018 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2019 if (C->failing()) return true;
2020 }
2021
2022 _igvn.set_delay_transform(false);
2023 _igvn.optimize();
2024 return false;
2025 }