Thu, 21 Aug 2008 23:36:31 -0400
Merge
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 */
25 #include "incls/_precompiled.incl"
26 #include "incls/_macro.cpp.incl"
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 }
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");
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->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 }
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 }
85 Node* PhaseMacroExpand::opt_iff(Node* region, Node* iff) {
86 IfNode *opt_iff = transform_later(iff)->as_If();
88 // Fast path taken; set region slot 2
89 Node *fast_taken = transform_later( new (C, 1) IfFalseNode(opt_iff) );
90 region->init_req(2,fast_taken); // Capture fast-control
92 // Fast path not-taken, i.e. slow path
93 Node *slow_taken = transform_later( new (C, 1) IfTrueNode(opt_iff) );
94 return slow_taken;
95 }
97 //--------------------copy_predefined_input_for_runtime_call--------------------
98 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
99 // Set fixed predefined input arguments
100 call->init_req( TypeFunc::Control, ctrl );
101 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) );
102 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
103 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
104 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
105 }
107 //------------------------------make_slow_call---------------------------------
108 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) {
110 // Slow-path call
111 int size = slow_call_type->domain()->cnt();
112 CallNode *call = leaf_name
113 ? (CallNode*)new (C, size) CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
114 : (CallNode*)new (C, size) CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM );
116 // Slow path call has no side-effects, uses few values
117 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
118 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0);
119 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1);
120 copy_call_debug_info(oldcall, call);
121 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
122 _igvn.hash_delete(oldcall);
123 _igvn.subsume_node(oldcall, call);
124 transform_later(call);
126 return call;
127 }
129 void PhaseMacroExpand::extract_call_projections(CallNode *call) {
130 _fallthroughproj = NULL;
131 _fallthroughcatchproj = NULL;
132 _ioproj_fallthrough = NULL;
133 _ioproj_catchall = NULL;
134 _catchallcatchproj = NULL;
135 _memproj_fallthrough = NULL;
136 _memproj_catchall = NULL;
137 _resproj = NULL;
138 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) {
139 ProjNode *pn = call->fast_out(i)->as_Proj();
140 switch (pn->_con) {
141 case TypeFunc::Control:
142 {
143 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj
144 _fallthroughproj = pn;
145 DUIterator_Fast jmax, j = pn->fast_outs(jmax);
146 const Node *cn = pn->fast_out(j);
147 if (cn->is_Catch()) {
148 ProjNode *cpn = NULL;
149 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) {
150 cpn = cn->fast_out(k)->as_Proj();
151 assert(cpn->is_CatchProj(), "must be a CatchProjNode");
152 if (cpn->_con == CatchProjNode::fall_through_index)
153 _fallthroughcatchproj = cpn;
154 else {
155 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index.");
156 _catchallcatchproj = cpn;
157 }
158 }
159 }
160 break;
161 }
162 case TypeFunc::I_O:
163 if (pn->_is_io_use)
164 _ioproj_catchall = pn;
165 else
166 _ioproj_fallthrough = pn;
167 break;
168 case TypeFunc::Memory:
169 if (pn->_is_io_use)
170 _memproj_catchall = pn;
171 else
172 _memproj_fallthrough = pn;
173 break;
174 case TypeFunc::Parms:
175 _resproj = pn;
176 break;
177 default:
178 assert(false, "unexpected projection from allocation node.");
179 }
180 }
182 }
184 // Eliminate a card mark sequence. p2x is a ConvP2XNode
185 void PhaseMacroExpand::eliminate_card_mark(Node *p2x) {
186 assert(p2x->Opcode() == Op_CastP2X, "ConvP2XNode required");
187 Node *shift = p2x->unique_out();
188 Node *addp = shift->unique_out();
189 for (DUIterator_Last jmin, j = addp->last_outs(jmin); j >= jmin; --j) {
190 Node *st = addp->last_out(j);
191 assert(st->is_Store(), "store required");
192 _igvn.replace_node(st, st->in(MemNode::Memory));
193 }
194 }
196 // Search for a memory operation for the specified memory slice.
197 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
198 Node *orig_mem = mem;
199 Node *alloc_mem = alloc->in(TypeFunc::Memory);
200 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
201 while (true) {
202 if (mem == alloc_mem || mem == start_mem ) {
203 return mem; // hit one of our sentinals
204 } else if (mem->is_MergeMem()) {
205 mem = mem->as_MergeMem()->memory_at(alias_idx);
206 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
207 Node *in = mem->in(0);
208 // we can safely skip over safepoints, calls, locks and membars because we
209 // already know that the object is safe to eliminate.
210 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
211 return in;
212 } else if (in->is_Call()) {
213 CallNode *call = in->as_Call();
214 if (!call->may_modify(tinst, phase)) {
215 mem = call->in(TypeFunc::Memory);
216 }
217 mem = in->in(TypeFunc::Memory);
218 } else if (in->is_MemBar()) {
219 mem = in->in(TypeFunc::Memory);
220 } else {
221 assert(false, "unexpected projection");
222 }
223 } else if (mem->is_Store()) {
224 const TypePtr* atype = mem->as_Store()->adr_type();
225 int adr_idx = Compile::current()->get_alias_index(atype);
226 if (adr_idx == alias_idx) {
227 assert(atype->isa_oopptr(), "address type must be oopptr");
228 int adr_offset = atype->offset();
229 uint adr_iid = atype->is_oopptr()->instance_id();
230 // Array elements references have the same alias_idx
231 // but different offset and different instance_id.
232 if (adr_offset == offset && adr_iid == alloc->_idx)
233 return mem;
234 } else {
235 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
236 }
237 mem = mem->in(MemNode::Memory);
238 } else {
239 return mem;
240 }
241 assert(mem != orig_mem, "dead memory loop");
242 }
243 }
245 //
246 // Given a Memory Phi, compute a value Phi containing the values from stores
247 // on the input paths.
248 // Note: this function is recursive, its depth is limied by the "level" argument
249 // Returns the computed Phi, or NULL if it cannot compute it.
250 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) {
251 assert(mem->is_Phi(), "sanity");
252 int alias_idx = C->get_alias_index(adr_t);
253 int offset = adr_t->offset();
254 int instance_id = adr_t->instance_id();
256 // Check if an appropriate value phi already exists.
257 Node* region = mem->in(0);
258 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
259 Node* phi = region->fast_out(k);
260 if (phi->is_Phi() && phi != mem &&
261 phi->as_Phi()->is_same_inst_field(phi_type, instance_id, alias_idx, offset)) {
262 return phi;
263 }
264 }
265 // Check if an appropriate new value phi already exists.
266 Node* new_phi = NULL;
267 uint size = value_phis->size();
268 for (uint i=0; i < size; i++) {
269 if ( mem->_idx == value_phis->index_at(i) ) {
270 return value_phis->node_at(i);
271 }
272 }
274 if (level <= 0) {
275 return NULL; // Give up: phi tree too deep
276 }
277 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
278 Node *alloc_mem = alloc->in(TypeFunc::Memory);
280 uint length = mem->req();
281 GrowableArray <Node *> values(length, length, NULL);
283 // create a new Phi for the value
284 PhiNode *phi = new (C, length) PhiNode(mem->in(0), phi_type, NULL, instance_id, alias_idx, offset);
285 transform_later(phi);
286 value_phis->push(phi, mem->_idx);
288 for (uint j = 1; j < length; j++) {
289 Node *in = mem->in(j);
290 if (in == NULL || in->is_top()) {
291 values.at_put(j, in);
292 } else {
293 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
294 if (val == start_mem || val == alloc_mem) {
295 // hit a sentinel, return appropriate 0 value
296 values.at_put(j, _igvn.zerocon(ft));
297 continue;
298 }
299 if (val->is_Initialize()) {
300 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
301 }
302 if (val == NULL) {
303 return NULL; // can't find a value on this path
304 }
305 if (val == mem) {
306 values.at_put(j, mem);
307 } else if (val->is_Store()) {
308 values.at_put(j, val->in(MemNode::ValueIn));
309 } else if(val->is_Proj() && val->in(0) == alloc) {
310 values.at_put(j, _igvn.zerocon(ft));
311 } else if (val->is_Phi()) {
312 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
313 if (val == NULL) {
314 return NULL;
315 }
316 values.at_put(j, val);
317 } else {
318 assert(false, "unknown node on this path");
319 return NULL; // unknown node on this path
320 }
321 }
322 }
323 // Set Phi's inputs
324 for (uint j = 1; j < length; j++) {
325 if (values.at(j) == mem) {
326 phi->init_req(j, phi);
327 } else {
328 phi->init_req(j, values.at(j));
329 }
330 }
331 return phi;
332 }
334 // Search the last value stored into the object's field.
335 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, Node *alloc) {
336 assert(adr_t->is_known_instance_field(), "instance required");
337 int instance_id = adr_t->instance_id();
338 assert((uint)instance_id == alloc->_idx, "wrong allocation");
340 int alias_idx = C->get_alias_index(adr_t);
341 int offset = adr_t->offset();
342 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
343 Node *alloc_ctrl = alloc->in(TypeFunc::Control);
344 Node *alloc_mem = alloc->in(TypeFunc::Memory);
345 Arena *a = Thread::current()->resource_area();
346 VectorSet visited(a);
349 bool done = sfpt_mem == alloc_mem;
350 Node *mem = sfpt_mem;
351 while (!done) {
352 if (visited.test_set(mem->_idx)) {
353 return NULL; // found a loop, give up
354 }
355 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
356 if (mem == start_mem || mem == alloc_mem) {
357 done = true; // hit a sentinel, return appropriate 0 value
358 } else if (mem->is_Initialize()) {
359 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
360 if (mem == NULL) {
361 done = true; // Something go wrong.
362 } else if (mem->is_Store()) {
363 const TypePtr* atype = mem->as_Store()->adr_type();
364 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
365 done = true;
366 }
367 } else if (mem->is_Store()) {
368 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
369 assert(atype != NULL, "address type must be oopptr");
370 assert(C->get_alias_index(atype) == alias_idx &&
371 atype->is_known_instance_field() && atype->offset() == offset &&
372 atype->instance_id() == instance_id, "store is correct memory slice");
373 done = true;
374 } else if (mem->is_Phi()) {
375 // try to find a phi's unique input
376 Node *unique_input = NULL;
377 Node *top = C->top();
378 for (uint i = 1; i < mem->req(); i++) {
379 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
380 if (n == NULL || n == top || n == mem) {
381 continue;
382 } else if (unique_input == NULL) {
383 unique_input = n;
384 } else if (unique_input != n) {
385 unique_input = top;
386 break;
387 }
388 }
389 if (unique_input != NULL && unique_input != top) {
390 mem = unique_input;
391 } else {
392 done = true;
393 }
394 } else {
395 assert(false, "unexpected node");
396 }
397 }
398 if (mem != NULL) {
399 if (mem == start_mem || mem == alloc_mem) {
400 // hit a sentinel, return appropriate 0 value
401 return _igvn.zerocon(ft);
402 } else if (mem->is_Store()) {
403 return mem->in(MemNode::ValueIn);
404 } else if (mem->is_Phi()) {
405 // attempt to produce a Phi reflecting the values on the input paths of the Phi
406 Node_Stack value_phis(a, 8);
407 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
408 if (phi != NULL) {
409 return phi;
410 } else {
411 // Kill all new Phis
412 while(value_phis.is_nonempty()) {
413 Node* n = value_phis.node();
414 _igvn.hash_delete(n);
415 _igvn.subsume_node(n, C->top());
416 value_phis.pop();
417 }
418 }
419 }
420 }
421 // Something go wrong.
422 return NULL;
423 }
425 // Check the possibility of scalar replacement.
426 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
427 // Scan the uses of the allocation to check for anything that would
428 // prevent us from eliminating it.
429 NOT_PRODUCT( const char* fail_eliminate = NULL; )
430 DEBUG_ONLY( Node* disq_node = NULL; )
431 bool can_eliminate = true;
433 Node* res = alloc->result_cast();
434 const TypeOopPtr* res_type = NULL;
435 if (res == NULL) {
436 // All users were eliminated.
437 } else if (!res->is_CheckCastPP()) {
438 alloc->_is_scalar_replaceable = false; // don't try again
439 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
440 can_eliminate = false;
441 } else {
442 res_type = _igvn.type(res)->isa_oopptr();
443 if (res_type == NULL) {
444 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
445 can_eliminate = false;
446 } else if (res_type->isa_aryptr()) {
447 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
448 if (length < 0) {
449 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
450 can_eliminate = false;
451 }
452 }
453 }
455 if (can_eliminate && res != NULL) {
456 for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
457 j < jmax && can_eliminate; j++) {
458 Node* use = res->fast_out(j);
460 if (use->is_AddP()) {
461 const TypePtr* addp_type = _igvn.type(use)->is_ptr();
462 int offset = addp_type->offset();
464 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
465 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";)
466 can_eliminate = false;
467 break;
468 }
469 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
470 k < kmax && can_eliminate; k++) {
471 Node* n = use->fast_out(k);
472 if (!n->is_Store() && n->Opcode() != Op_CastP2X) {
473 DEBUG_ONLY(disq_node = n;)
474 if (n->is_Load() || n->is_LoadStore()) {
475 NOT_PRODUCT(fail_eliminate = "Field load";)
476 } else {
477 NOT_PRODUCT(fail_eliminate = "Not store field referrence";)
478 }
479 can_eliminate = false;
480 }
481 }
482 } else if (use->is_SafePoint()) {
483 SafePointNode* sfpt = use->as_SafePoint();
484 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
485 // Object is passed as argument.
486 DEBUG_ONLY(disq_node = use;)
487 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
488 can_eliminate = false;
489 }
490 Node* sfptMem = sfpt->memory();
491 if (sfptMem == NULL || sfptMem->is_top()) {
492 DEBUG_ONLY(disq_node = use;)
493 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";)
494 can_eliminate = false;
495 } else {
496 safepoints.append_if_missing(sfpt);
497 }
498 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
499 if (use->is_Phi()) {
500 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
501 NOT_PRODUCT(fail_eliminate = "Object is return value";)
502 } else {
503 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
504 }
505 DEBUG_ONLY(disq_node = use;)
506 } else {
507 if (use->Opcode() == Op_Return) {
508 NOT_PRODUCT(fail_eliminate = "Object is return value";)
509 }else {
510 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
511 }
512 DEBUG_ONLY(disq_node = use;)
513 }
514 can_eliminate = false;
515 }
516 }
517 }
519 #ifndef PRODUCT
520 if (PrintEliminateAllocations) {
521 if (can_eliminate) {
522 tty->print("Scalar ");
523 if (res == NULL)
524 alloc->dump();
525 else
526 res->dump();
527 } else {
528 tty->print("NotScalar (%s)", fail_eliminate);
529 if (res == NULL)
530 alloc->dump();
531 else
532 res->dump();
533 #ifdef ASSERT
534 if (disq_node != NULL) {
535 tty->print(" >>>> ");
536 disq_node->dump();
537 }
538 #endif /*ASSERT*/
539 }
540 }
541 #endif
542 return can_eliminate;
543 }
545 // Do scalar replacement.
546 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
547 GrowableArray <SafePointNode *> safepoints_done;
549 ciKlass* klass = NULL;
550 ciInstanceKlass* iklass = NULL;
551 int nfields = 0;
552 int array_base;
553 int element_size;
554 BasicType basic_elem_type;
555 ciType* elem_type;
557 Node* res = alloc->result_cast();
558 const TypeOopPtr* res_type = NULL;
559 if (res != NULL) { // Could be NULL when there are no users
560 res_type = _igvn.type(res)->isa_oopptr();
561 }
563 if (res != NULL) {
564 klass = res_type->klass();
565 if (res_type->isa_instptr()) {
566 // find the fields of the class which will be needed for safepoint debug information
567 assert(klass->is_instance_klass(), "must be an instance klass.");
568 iklass = klass->as_instance_klass();
569 nfields = iklass->nof_nonstatic_fields();
570 } else {
571 // find the array's elements which will be needed for safepoint debug information
572 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
573 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass.");
574 elem_type = klass->as_array_klass()->element_type();
575 basic_elem_type = elem_type->basic_type();
576 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
577 element_size = type2aelembytes(basic_elem_type);
578 }
579 }
580 //
581 // Process the safepoint uses
582 //
583 while (safepoints.length() > 0) {
584 SafePointNode* sfpt = safepoints.pop();
585 Node* mem = sfpt->memory();
586 uint first_ind = sfpt->req();
587 SafePointScalarObjectNode* sobj = new (C, 1) SafePointScalarObjectNode(res_type,
588 #ifdef ASSERT
589 alloc,
590 #endif
591 first_ind, nfields);
592 sobj->init_req(0, sfpt->in(TypeFunc::Control));
593 transform_later(sobj);
595 // Scan object's fields adding an input to the safepoint for each field.
596 for (int j = 0; j < nfields; j++) {
597 int offset;
598 ciField* field = NULL;
599 if (iklass != NULL) {
600 field = iklass->nonstatic_field_at(j);
601 offset = field->offset();
602 elem_type = field->type();
603 basic_elem_type = field->layout_type();
604 } else {
605 offset = array_base + j * element_size;
606 }
608 const Type *field_type;
609 // The next code is taken from Parse::do_get_xxx().
610 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) {
611 if (!elem_type->is_loaded()) {
612 field_type = TypeInstPtr::BOTTOM;
613 } else if (field != NULL && field->is_constant()) {
614 // This can happen if the constant oop is non-perm.
615 ciObject* con = field->constant_value().as_object();
616 // Do not "join" in the previous type; it doesn't add value,
617 // and may yield a vacuous result if the field is of interface type.
618 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
619 assert(field_type != NULL, "field singleton type must be consistent");
620 } else {
621 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
622 }
623 if (UseCompressedOops) {
624 field_type = field_type->make_narrowoop();
625 basic_elem_type = T_NARROWOOP;
626 }
627 } else {
628 field_type = Type::get_const_basic_type(basic_elem_type);
629 }
631 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr();
633 Node *field_val = value_from_mem(mem, basic_elem_type, field_type, field_addr_type, alloc);
634 if (field_val == NULL) {
635 // we weren't able to find a value for this field,
636 // give up on eliminating this allocation
637 alloc->_is_scalar_replaceable = false; // don't try again
638 // remove any extra entries we added to the safepoint
639 uint last = sfpt->req() - 1;
640 for (int k = 0; k < j; k++) {
641 sfpt->del_req(last--);
642 }
643 // rollback processed safepoints
644 while (safepoints_done.length() > 0) {
645 SafePointNode* sfpt_done = safepoints_done.pop();
646 // remove any extra entries we added to the safepoint
647 last = sfpt_done->req() - 1;
648 for (int k = 0; k < nfields; k++) {
649 sfpt_done->del_req(last--);
650 }
651 JVMState *jvms = sfpt_done->jvms();
652 jvms->set_endoff(sfpt_done->req());
653 // Now make a pass over the debug information replacing any references
654 // to SafePointScalarObjectNode with the allocated object.
655 int start = jvms->debug_start();
656 int end = jvms->debug_end();
657 for (int i = start; i < end; i++) {
658 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
659 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
660 if (scobj->first_index() == sfpt_done->req() &&
661 scobj->n_fields() == (uint)nfields) {
662 assert(scobj->alloc() == alloc, "sanity");
663 sfpt_done->set_req(i, res);
664 }
665 }
666 }
667 }
668 #ifndef PRODUCT
669 if (PrintEliminateAllocations) {
670 if (field != NULL) {
671 tty->print("=== At SafePoint node %d can't find value of Field: ",
672 sfpt->_idx);
673 field->print();
674 int field_idx = C->get_alias_index(field_addr_type);
675 tty->print(" (alias_idx=%d)", field_idx);
676 } else { // Array's element
677 tty->print("=== At SafePoint node %d can't find value of array element [%d]",
678 sfpt->_idx, j);
679 }
680 tty->print(", which prevents elimination of: ");
681 if (res == NULL)
682 alloc->dump();
683 else
684 res->dump();
685 }
686 #endif
687 return false;
688 }
689 if (UseCompressedOops && field_type->isa_narrowoop()) {
690 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
691 // to be able scalar replace the allocation.
692 if (field_val->is_EncodeP()) {
693 field_val = field_val->in(1);
694 } else {
695 field_val = transform_later(new (C, 2) DecodeNNode(field_val, field_val->bottom_type()->make_ptr()));
696 }
697 }
698 sfpt->add_req(field_val);
699 }
700 JVMState *jvms = sfpt->jvms();
701 jvms->set_endoff(sfpt->req());
702 // Now make a pass over the debug information replacing any references
703 // to the allocated object with "sobj"
704 int start = jvms->debug_start();
705 int end = jvms->debug_end();
706 for (int i = start; i < end; i++) {
707 if (sfpt->in(i) == res) {
708 sfpt->set_req(i, sobj);
709 }
710 }
711 safepoints_done.append_if_missing(sfpt); // keep it for rollback
712 }
713 return true;
714 }
716 // Process users of eliminated allocation.
717 void PhaseMacroExpand::process_users_of_allocation(AllocateNode *alloc) {
718 Node* res = alloc->result_cast();
719 if (res != NULL) {
720 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
721 Node *use = res->last_out(j);
722 uint oc1 = res->outcnt();
724 if (use->is_AddP()) {
725 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
726 Node *n = use->last_out(k);
727 uint oc2 = use->outcnt();
728 if (n->is_Store()) {
729 _igvn.replace_node(n, n->in(MemNode::Memory));
730 } else {
731 assert( n->Opcode() == Op_CastP2X, "CastP2X required");
732 eliminate_card_mark(n);
733 }
734 k -= (oc2 - use->outcnt());
735 }
736 } else {
737 assert( !use->is_SafePoint(), "safepoint uses must have been already elimiated");
738 assert( use->Opcode() == Op_CastP2X, "CastP2X required");
739 eliminate_card_mark(use);
740 }
741 j -= (oc1 - res->outcnt());
742 }
743 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
744 _igvn.remove_dead_node(res);
745 }
747 //
748 // Process other users of allocation's projections
749 //
750 if (_resproj != NULL && _resproj->outcnt() != 0) {
751 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) {
752 Node *use = _resproj->last_out(j);
753 uint oc1 = _resproj->outcnt();
754 if (use->is_Initialize()) {
755 // Eliminate Initialize node.
756 InitializeNode *init = use->as_Initialize();
757 assert(init->outcnt() <= 2, "only a control and memory projection expected");
758 Node *ctrl_proj = init->proj_out(TypeFunc::Control);
759 if (ctrl_proj != NULL) {
760 assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection");
761 _igvn.replace_node(ctrl_proj, _fallthroughcatchproj);
762 }
763 Node *mem_proj = init->proj_out(TypeFunc::Memory);
764 if (mem_proj != NULL) {
765 Node *mem = init->in(TypeFunc::Memory);
766 #ifdef ASSERT
767 if (mem->is_MergeMem()) {
768 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection");
769 } else {
770 assert(mem == _memproj_fallthrough, "allocation memory projection");
771 }
772 #endif
773 _igvn.replace_node(mem_proj, mem);
774 }
775 } else if (use->is_AddP()) {
776 // raw memory addresses used only by the initialization
777 _igvn.hash_delete(use);
778 _igvn.subsume_node(use, C->top());
779 } else {
780 assert(false, "only Initialize or AddP expected");
781 }
782 j -= (oc1 - _resproj->outcnt());
783 }
784 }
785 if (_fallthroughcatchproj != NULL) {
786 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control));
787 }
788 if (_memproj_fallthrough != NULL) {
789 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory));
790 }
791 if (_memproj_catchall != NULL) {
792 _igvn.replace_node(_memproj_catchall, C->top());
793 }
794 if (_ioproj_fallthrough != NULL) {
795 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O));
796 }
797 if (_ioproj_catchall != NULL) {
798 _igvn.replace_node(_ioproj_catchall, C->top());
799 }
800 if (_catchallcatchproj != NULL) {
801 _igvn.replace_node(_catchallcatchproj, C->top());
802 }
803 }
805 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
807 if (!EliminateAllocations || !alloc->_is_scalar_replaceable) {
808 return false;
809 }
811 extract_call_projections(alloc);
813 GrowableArray <SafePointNode *> safepoints;
814 if (!can_eliminate_allocation(alloc, safepoints)) {
815 return false;
816 }
818 if (!scalar_replacement(alloc, safepoints)) {
819 return false;
820 }
822 process_users_of_allocation(alloc);
824 #ifndef PRODUCT
825 if (PrintEliminateAllocations) {
826 if (alloc->is_AllocateArray())
827 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
828 else
829 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
830 }
831 #endif
833 return true;
834 }
837 //---------------------------set_eden_pointers-------------------------
838 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) {
839 if (UseTLAB) { // Private allocation: load from TLS
840 Node* thread = transform_later(new (C, 1) ThreadLocalNode());
841 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset());
842 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset());
843 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset);
844 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset);
845 } else { // Shared allocation: load from globals
846 CollectedHeap* ch = Universe::heap();
847 address top_adr = (address)ch->top_addr();
848 address end_adr = (address)ch->end_addr();
849 eden_top_adr = makecon(TypeRawPtr::make(top_adr));
850 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr);
851 }
852 }
855 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
856 Node* adr = basic_plus_adr(base, offset);
857 const TypePtr* adr_type = TypeRawPtr::BOTTOM;
858 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt);
859 transform_later(value);
860 return value;
861 }
864 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
865 Node* adr = basic_plus_adr(base, offset);
866 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt);
867 transform_later(mem);
868 return mem;
869 }
871 //=============================================================================
872 //
873 // A L L O C A T I O N
874 //
875 // Allocation attempts to be fast in the case of frequent small objects.
876 // It breaks down like this:
877 //
878 // 1) Size in doublewords is computed. This is a constant for objects and
879 // variable for most arrays. Doubleword units are used to avoid size
880 // overflow of huge doubleword arrays. We need doublewords in the end for
881 // rounding.
882 //
883 // 2) Size is checked for being 'too large'. Too-large allocations will go
884 // the slow path into the VM. The slow path can throw any required
885 // exceptions, and does all the special checks for very large arrays. The
886 // size test can constant-fold away for objects. For objects with
887 // finalizers it constant-folds the otherway: you always go slow with
888 // finalizers.
889 //
890 // 3) If NOT using TLABs, this is the contended loop-back point.
891 // Load-Locked the heap top. If using TLABs normal-load the heap top.
892 //
893 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
894 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
895 // "size*8" we always enter the VM, where "largish" is a constant picked small
896 // enough that there's always space between the eden max and 4Gig (old space is
897 // there so it's quite large) and large enough that the cost of entering the VM
898 // is dwarfed by the cost to initialize the space.
899 //
900 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
901 // down. If contended, repeat at step 3. If using TLABs normal-store
902 // adjusted heap top back down; there is no contention.
903 //
904 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
905 // fields.
906 //
907 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
908 // oop flavor.
909 //
910 //=============================================================================
911 // FastAllocateSizeLimit value is in DOUBLEWORDS.
912 // Allocations bigger than this always go the slow route.
913 // This value must be small enough that allocation attempts that need to
914 // trigger exceptions go the slow route. Also, it must be small enough so
915 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
916 //=============================================================================j//
917 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
918 // The allocator will coalesce int->oop copies away. See comment in
919 // coalesce.cpp about how this works. It depends critically on the exact
920 // code shape produced here, so if you are changing this code shape
921 // make sure the GC info for the heap-top is correct in and around the
922 // slow-path call.
923 //
925 void PhaseMacroExpand::expand_allocate_common(
926 AllocateNode* alloc, // allocation node to be expanded
927 Node* length, // array length for an array allocation
928 const TypeFunc* slow_call_type, // Type of slow call
929 address slow_call_address // Address of slow call
930 )
931 {
933 Node* ctrl = alloc->in(TypeFunc::Control);
934 Node* mem = alloc->in(TypeFunc::Memory);
935 Node* i_o = alloc->in(TypeFunc::I_O);
936 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
937 Node* klass_node = alloc->in(AllocateNode::KlassNode);
938 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
940 // With escape analysis, the entire memory state was needed to be able to
941 // eliminate the allocation. Since the allocations cannot be eliminated,
942 // optimize it to the raw slice.
943 if (mem->is_MergeMem()) {
944 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
945 }
947 assert(ctrl != NULL, "must have control");
948 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
949 // they will not be used if "always_slow" is set
950 enum { slow_result_path = 1, fast_result_path = 2 };
951 Node *result_region;
952 Node *result_phi_rawmem;
953 Node *result_phi_rawoop;
954 Node *result_phi_i_o;
956 // The initial slow comparison is a size check, the comparison
957 // we want to do is a BoolTest::gt
958 bool always_slow = false;
959 int tv = _igvn.find_int_con(initial_slow_test, -1);
960 if (tv >= 0) {
961 always_slow = (tv == 1);
962 initial_slow_test = NULL;
963 } else {
964 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
965 }
967 if (DTraceAllocProbes ||
968 !UseTLAB && (!Universe::heap()->supports_inline_contig_alloc() ||
969 (UseConcMarkSweepGC && CMSIncrementalMode))) {
970 // Force slow-path allocation
971 always_slow = true;
972 initial_slow_test = NULL;
973 }
976 enum { too_big_or_final_path = 1, need_gc_path = 2 };
977 Node *slow_region = NULL;
978 Node *toobig_false = ctrl;
980 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent");
981 // generate the initial test if necessary
982 if (initial_slow_test != NULL ) {
983 slow_region = new (C, 3) RegionNode(3);
985 // Now make the initial failure test. Usually a too-big test but
986 // might be a TRUE for finalizers or a fancy class check for
987 // newInstance0.
988 IfNode *toobig_iff = new (C, 2) IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
989 transform_later(toobig_iff);
990 // Plug the failing-too-big test into the slow-path region
991 Node *toobig_true = new (C, 1) IfTrueNode( toobig_iff );
992 transform_later(toobig_true);
993 slow_region ->init_req( too_big_or_final_path, toobig_true );
994 toobig_false = new (C, 1) IfFalseNode( toobig_iff );
995 transform_later(toobig_false);
996 } else { // No initial test, just fall into next case
997 toobig_false = ctrl;
998 debug_only(slow_region = NodeSentinel);
999 }
1001 Node *slow_mem = mem; // save the current memory state for slow path
1002 // generate the fast allocation code unless we know that the initial test will always go slow
1003 if (!always_slow) {
1004 Node* eden_top_adr;
1005 Node* eden_end_adr;
1007 set_eden_pointers(eden_top_adr, eden_end_adr);
1009 // Load Eden::end. Loop invariant and hoisted.
1010 //
1011 // Note: We set the control input on "eden_end" and "old_eden_top" when using
1012 // a TLAB to work around a bug where these values were being moved across
1013 // a safepoint. These are not oops, so they cannot be include in the oop
1014 // map, but the can be changed by a GC. The proper way to fix this would
1015 // be to set the raw memory state when generating a SafepointNode. However
1016 // this will require extensive changes to the loop optimization in order to
1017 // prevent a degradation of the optimization.
1018 // See comment in memnode.hpp, around line 227 in class LoadPNode.
1019 Node *eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS);
1021 // allocate the Region and Phi nodes for the result
1022 result_region = new (C, 3) RegionNode(3);
1023 result_phi_rawmem = new (C, 3) PhiNode( result_region, Type::MEMORY, TypeRawPtr::BOTTOM );
1024 result_phi_rawoop = new (C, 3) PhiNode( result_region, TypeRawPtr::BOTTOM );
1025 result_phi_i_o = new (C, 3) PhiNode( result_region, Type::ABIO ); // I/O is used for Prefetch
1027 // We need a Region for the loop-back contended case.
1028 enum { fall_in_path = 1, contended_loopback_path = 2 };
1029 Node *contended_region;
1030 Node *contended_phi_rawmem;
1031 if( UseTLAB ) {
1032 contended_region = toobig_false;
1033 contended_phi_rawmem = mem;
1034 } else {
1035 contended_region = new (C, 3) RegionNode(3);
1036 contended_phi_rawmem = new (C, 3) PhiNode( contended_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1037 // Now handle the passing-too-big test. We fall into the contended
1038 // loop-back merge point.
1039 contended_region ->init_req( fall_in_path, toobig_false );
1040 contended_phi_rawmem->init_req( fall_in_path, mem );
1041 transform_later(contended_region);
1042 transform_later(contended_phi_rawmem);
1043 }
1045 // Load(-locked) the heap top.
1046 // See note above concerning the control input when using a TLAB
1047 Node *old_eden_top = UseTLAB
1048 ? new (C, 3) LoadPNode ( ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM )
1049 : new (C, 3) LoadPLockedNode( contended_region, contended_phi_rawmem, eden_top_adr );
1051 transform_later(old_eden_top);
1052 // Add to heap top to get a new heap top
1053 Node *new_eden_top = new (C, 4) AddPNode( top(), old_eden_top, size_in_bytes );
1054 transform_later(new_eden_top);
1055 // Check for needing a GC; compare against heap end
1056 Node *needgc_cmp = new (C, 3) CmpPNode( new_eden_top, eden_end );
1057 transform_later(needgc_cmp);
1058 Node *needgc_bol = new (C, 2) BoolNode( needgc_cmp, BoolTest::ge );
1059 transform_later(needgc_bol);
1060 IfNode *needgc_iff = new (C, 2) IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1061 transform_later(needgc_iff);
1063 // Plug the failing-heap-space-need-gc test into the slow-path region
1064 Node *needgc_true = new (C, 1) IfTrueNode( needgc_iff );
1065 transform_later(needgc_true);
1066 if( initial_slow_test ) {
1067 slow_region ->init_req( need_gc_path, needgc_true );
1068 // This completes all paths into the slow merge point
1069 transform_later(slow_region);
1070 } else { // No initial slow path needed!
1071 // Just fall from the need-GC path straight into the VM call.
1072 slow_region = needgc_true;
1073 }
1074 // No need for a GC. Setup for the Store-Conditional
1075 Node *needgc_false = new (C, 1) IfFalseNode( needgc_iff );
1076 transform_later(needgc_false);
1078 // Grab regular I/O before optional prefetch may change it.
1079 // Slow-path does no I/O so just set it to the original I/O.
1080 result_phi_i_o->init_req( slow_result_path, i_o );
1082 i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem,
1083 old_eden_top, new_eden_top, length);
1085 // Store (-conditional) the modified eden top back down.
1086 // StorePConditional produces flags for a test PLUS a modified raw
1087 // memory state.
1088 Node *store_eden_top;
1089 Node *fast_oop_ctrl;
1090 if( UseTLAB ) {
1091 store_eden_top = new (C, 4) StorePNode( needgc_false, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, new_eden_top );
1092 transform_later(store_eden_top);
1093 fast_oop_ctrl = needgc_false; // No contention, so this is the fast path
1094 } else {
1095 store_eden_top = new (C, 5) StorePConditionalNode( needgc_false, contended_phi_rawmem, eden_top_adr, new_eden_top, old_eden_top );
1096 transform_later(store_eden_top);
1097 Node *contention_check = new (C, 2) BoolNode( store_eden_top, BoolTest::ne );
1098 transform_later(contention_check);
1099 store_eden_top = new (C, 1) SCMemProjNode(store_eden_top);
1100 transform_later(store_eden_top);
1102 // If not using TLABs, check to see if there was contention.
1103 IfNode *contention_iff = new (C, 2) IfNode ( needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN );
1104 transform_later(contention_iff);
1105 Node *contention_true = new (C, 1) IfTrueNode( contention_iff );
1106 transform_later(contention_true);
1107 // If contention, loopback and try again.
1108 contended_region->init_req( contended_loopback_path, contention_true );
1109 contended_phi_rawmem->init_req( contended_loopback_path, store_eden_top );
1111 // Fast-path succeeded with no contention!
1112 Node *contention_false = new (C, 1) IfFalseNode( contention_iff );
1113 transform_later(contention_false);
1114 fast_oop_ctrl = contention_false;
1115 }
1117 // Rename successful fast-path variables to make meaning more obvious
1118 Node* fast_oop = old_eden_top;
1119 Node* fast_oop_rawmem = store_eden_top;
1120 fast_oop_rawmem = initialize_object(alloc,
1121 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1122 klass_node, length, size_in_bytes);
1124 if (ExtendedDTraceProbes) {
1125 // Slow-path call
1126 int size = TypeFunc::Parms + 2;
1127 CallLeafNode *call = new (C, size) CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1128 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base),
1129 "dtrace_object_alloc",
1130 TypeRawPtr::BOTTOM);
1132 // Get base of thread-local storage area
1133 Node* thread = new (C, 1) ThreadLocalNode();
1134 transform_later(thread);
1136 call->init_req(TypeFunc::Parms+0, thread);
1137 call->init_req(TypeFunc::Parms+1, fast_oop);
1138 call->init_req( TypeFunc::Control, fast_oop_ctrl );
1139 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o
1140 call->init_req( TypeFunc::Memory , fast_oop_rawmem );
1141 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) );
1142 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) );
1143 transform_later(call);
1144 fast_oop_ctrl = new (C, 1) ProjNode(call,TypeFunc::Control);
1145 transform_later(fast_oop_ctrl);
1146 fast_oop_rawmem = new (C, 1) ProjNode(call,TypeFunc::Memory);
1147 transform_later(fast_oop_rawmem);
1148 }
1150 // Plug in the successful fast-path into the result merge point
1151 result_region ->init_req( fast_result_path, fast_oop_ctrl );
1152 result_phi_rawoop->init_req( fast_result_path, fast_oop );
1153 result_phi_i_o ->init_req( fast_result_path, i_o );
1154 result_phi_rawmem->init_req( fast_result_path, fast_oop_rawmem );
1155 } else {
1156 slow_region = ctrl;
1157 }
1159 // Generate slow-path call
1160 CallNode *call = new (C, slow_call_type->domain()->cnt())
1161 CallStaticJavaNode(slow_call_type, slow_call_address,
1162 OptoRuntime::stub_name(slow_call_address),
1163 alloc->jvms()->bci(),
1164 TypePtr::BOTTOM);
1165 call->init_req( TypeFunc::Control, slow_region );
1166 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o
1167 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs
1168 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) );
1169 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) );
1171 call->init_req(TypeFunc::Parms+0, klass_node);
1172 if (length != NULL) {
1173 call->init_req(TypeFunc::Parms+1, length);
1174 }
1176 // Copy debug information and adjust JVMState information, then replace
1177 // allocate node with the call
1178 copy_call_debug_info((CallNode *) alloc, call);
1179 if (!always_slow) {
1180 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1181 }
1182 _igvn.hash_delete(alloc);
1183 _igvn.subsume_node(alloc, call);
1184 transform_later(call);
1186 // Identify the output projections from the allocate node and
1187 // adjust any references to them.
1188 // The control and io projections look like:
1189 //
1190 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1191 // Allocate Catch
1192 // ^---Proj(io) <-------+ ^---CatchProj(io)
1193 //
1194 // We are interested in the CatchProj nodes.
1195 //
1196 extract_call_projections(call);
1198 // An allocate node has separate memory projections for the uses on the control and i_o paths
1199 // Replace uses of the control memory projection with result_phi_rawmem (unless we are only generating a slow call)
1200 if (!always_slow && _memproj_fallthrough != NULL) {
1201 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) {
1202 Node *use = _memproj_fallthrough->fast_out(i);
1203 _igvn.hash_delete(use);
1204 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem);
1205 _igvn._worklist.push(use);
1206 // back up iterator
1207 --i;
1208 }
1209 }
1210 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete _memproj_catchall so
1211 // we end up with a call that has only 1 memory projection
1212 if (_memproj_catchall != NULL ) {
1213 if (_memproj_fallthrough == NULL) {
1214 _memproj_fallthrough = new (C, 1) ProjNode(call, TypeFunc::Memory);
1215 transform_later(_memproj_fallthrough);
1216 }
1217 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) {
1218 Node *use = _memproj_catchall->fast_out(i);
1219 _igvn.hash_delete(use);
1220 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough);
1221 _igvn._worklist.push(use);
1222 // back up iterator
1223 --i;
1224 }
1225 }
1227 mem = result_phi_rawmem;
1229 // An allocate node has separate i_o projections for the uses on the control and i_o paths
1230 // Replace uses of the control i_o projection with result_phi_i_o (unless we are only generating a slow call)
1231 if (_ioproj_fallthrough == NULL) {
1232 _ioproj_fallthrough = new (C, 1) ProjNode(call, TypeFunc::I_O);
1233 transform_later(_ioproj_fallthrough);
1234 } else if (!always_slow) {
1235 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) {
1236 Node *use = _ioproj_fallthrough->fast_out(i);
1238 _igvn.hash_delete(use);
1239 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o);
1240 _igvn._worklist.push(use);
1241 // back up iterator
1242 --i;
1243 }
1244 }
1245 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete _ioproj_catchall so
1246 // we end up with a call that has only 1 control projection
1247 if (_ioproj_catchall != NULL ) {
1248 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) {
1249 Node *use = _ioproj_catchall->fast_out(i);
1250 _igvn.hash_delete(use);
1251 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough);
1252 _igvn._worklist.push(use);
1253 // back up iterator
1254 --i;
1255 }
1256 }
1258 // if we generated only a slow call, we are done
1259 if (always_slow)
1260 return;
1263 if (_fallthroughcatchproj != NULL) {
1264 ctrl = _fallthroughcatchproj->clone();
1265 transform_later(ctrl);
1266 _igvn.hash_delete(_fallthroughcatchproj);
1267 _igvn.subsume_node(_fallthroughcatchproj, result_region);
1268 } else {
1269 ctrl = top();
1270 }
1271 Node *slow_result;
1272 if (_resproj == NULL) {
1273 // no uses of the allocation result
1274 slow_result = top();
1275 } else {
1276 slow_result = _resproj->clone();
1277 transform_later(slow_result);
1278 _igvn.hash_delete(_resproj);
1279 _igvn.subsume_node(_resproj, result_phi_rawoop);
1280 }
1282 // Plug slow-path into result merge point
1283 result_region ->init_req( slow_result_path, ctrl );
1284 result_phi_rawoop->init_req( slow_result_path, slow_result);
1285 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough );
1286 transform_later(result_region);
1287 transform_later(result_phi_rawoop);
1288 transform_later(result_phi_rawmem);
1289 transform_later(result_phi_i_o);
1290 // This completes all paths into the result merge point
1291 }
1294 // Helper for PhaseMacroExpand::expand_allocate_common.
1295 // Initializes the newly-allocated storage.
1296 Node*
1297 PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1298 Node* control, Node* rawmem, Node* object,
1299 Node* klass_node, Node* length,
1300 Node* size_in_bytes) {
1301 InitializeNode* init = alloc->initialization();
1302 // Store the klass & mark bits
1303 Node* mark_node = NULL;
1304 // For now only enable fast locking for non-array types
1305 if (UseBiasedLocking && (length == NULL)) {
1306 mark_node = make_load(NULL, rawmem, klass_node, Klass::prototype_header_offset_in_bytes() + sizeof(oopDesc), TypeRawPtr::BOTTOM, T_ADDRESS);
1307 } else {
1308 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype()));
1309 }
1310 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
1312 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_OBJECT);
1313 int header_size = alloc->minimum_header_size(); // conservatively small
1315 // Array length
1316 if (length != NULL) { // Arrays need length field
1317 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1318 // conservatively small header size:
1319 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1320 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1321 if (k->is_array_klass()) // we know the exact header size in most cases:
1322 header_size = Klass::layout_helper_header_size(k->layout_helper());
1323 }
1325 // Clear the object body, if necessary.
1326 if (init == NULL) {
1327 // The init has somehow disappeared; be cautious and clear everything.
1328 //
1329 // This can happen if a node is allocated but an uncommon trap occurs
1330 // immediately. In this case, the Initialize gets associated with the
1331 // trap, and may be placed in a different (outer) loop, if the Allocate
1332 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1333 // there can be two Allocates to one Initialize. The answer in all these
1334 // edge cases is safety first. It is always safe to clear immediately
1335 // within an Allocate, and then (maybe or maybe not) clear some more later.
1336 if (!ZeroTLAB)
1337 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1338 header_size, size_in_bytes,
1339 &_igvn);
1340 } else {
1341 if (!init->is_complete()) {
1342 // Try to win by zeroing only what the init does not store.
1343 // We can also try to do some peephole optimizations,
1344 // such as combining some adjacent subword stores.
1345 rawmem = init->complete_stores(control, rawmem, object,
1346 header_size, size_in_bytes, &_igvn);
1347 }
1348 // We have no more use for this link, since the AllocateNode goes away:
1349 init->set_req(InitializeNode::RawAddress, top());
1350 // (If we keep the link, it just confuses the register allocator,
1351 // who thinks he sees a real use of the address by the membar.)
1352 }
1354 return rawmem;
1355 }
1357 // Generate prefetch instructions for next allocations.
1358 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
1359 Node*& contended_phi_rawmem,
1360 Node* old_eden_top, Node* new_eden_top,
1361 Node* length) {
1362 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
1363 // Generate prefetch allocation with watermark check.
1364 // As an allocation hits the watermark, we will prefetch starting
1365 // at a "distance" away from watermark.
1366 enum { fall_in_path = 1, pf_path = 2 };
1368 Node *pf_region = new (C, 3) RegionNode(3);
1369 Node *pf_phi_rawmem = new (C, 3) PhiNode( pf_region, Type::MEMORY,
1370 TypeRawPtr::BOTTOM );
1371 // I/O is used for Prefetch
1372 Node *pf_phi_abio = new (C, 3) PhiNode( pf_region, Type::ABIO );
1374 Node *thread = new (C, 1) ThreadLocalNode();
1375 transform_later(thread);
1377 Node *eden_pf_adr = new (C, 4) AddPNode( top()/*not oop*/, thread,
1378 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
1379 transform_later(eden_pf_adr);
1381 Node *old_pf_wm = new (C, 3) LoadPNode( needgc_false,
1382 contended_phi_rawmem, eden_pf_adr,
1383 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM );
1384 transform_later(old_pf_wm);
1386 // check against new_eden_top
1387 Node *need_pf_cmp = new (C, 3) CmpPNode( new_eden_top, old_pf_wm );
1388 transform_later(need_pf_cmp);
1389 Node *need_pf_bol = new (C, 2) BoolNode( need_pf_cmp, BoolTest::ge );
1390 transform_later(need_pf_bol);
1391 IfNode *need_pf_iff = new (C, 2) IfNode( needgc_false, need_pf_bol,
1392 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1393 transform_later(need_pf_iff);
1395 // true node, add prefetchdistance
1396 Node *need_pf_true = new (C, 1) IfTrueNode( need_pf_iff );
1397 transform_later(need_pf_true);
1399 Node *need_pf_false = new (C, 1) IfFalseNode( need_pf_iff );
1400 transform_later(need_pf_false);
1402 Node *new_pf_wmt = new (C, 4) AddPNode( top(), old_pf_wm,
1403 _igvn.MakeConX(AllocatePrefetchDistance) );
1404 transform_later(new_pf_wmt );
1405 new_pf_wmt->set_req(0, need_pf_true);
1407 Node *store_new_wmt = new (C, 4) StorePNode( need_pf_true,
1408 contended_phi_rawmem, eden_pf_adr,
1409 TypeRawPtr::BOTTOM, new_pf_wmt );
1410 transform_later(store_new_wmt);
1412 // adding prefetches
1413 pf_phi_abio->init_req( fall_in_path, i_o );
1415 Node *prefetch_adr;
1416 Node *prefetch;
1417 uint lines = AllocatePrefetchDistance / AllocatePrefetchStepSize;
1418 uint step_size = AllocatePrefetchStepSize;
1419 uint distance = 0;
1421 for ( uint i = 0; i < lines; i++ ) {
1422 prefetch_adr = new (C, 4) AddPNode( old_pf_wm, new_pf_wmt,
1423 _igvn.MakeConX(distance) );
1424 transform_later(prefetch_adr);
1425 prefetch = new (C, 3) PrefetchWriteNode( i_o, prefetch_adr );
1426 transform_later(prefetch);
1427 distance += step_size;
1428 i_o = prefetch;
1429 }
1430 pf_phi_abio->set_req( pf_path, i_o );
1432 pf_region->init_req( fall_in_path, need_pf_false );
1433 pf_region->init_req( pf_path, need_pf_true );
1435 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
1436 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
1438 transform_later(pf_region);
1439 transform_later(pf_phi_rawmem);
1440 transform_later(pf_phi_abio);
1442 needgc_false = pf_region;
1443 contended_phi_rawmem = pf_phi_rawmem;
1444 i_o = pf_phi_abio;
1445 } else if( AllocatePrefetchStyle > 0 ) {
1446 // Insert a prefetch for each allocation only on the fast-path
1447 Node *prefetch_adr;
1448 Node *prefetch;
1449 // Generate several prefetch instructions only for arrays.
1450 uint lines = (length != NULL) ? AllocatePrefetchLines : 1;
1451 uint step_size = AllocatePrefetchStepSize;
1452 uint distance = AllocatePrefetchDistance;
1453 for ( uint i = 0; i < lines; i++ ) {
1454 prefetch_adr = new (C, 4) AddPNode( old_eden_top, new_eden_top,
1455 _igvn.MakeConX(distance) );
1456 transform_later(prefetch_adr);
1457 prefetch = new (C, 3) PrefetchWriteNode( i_o, prefetch_adr );
1458 // Do not let it float too high, since if eden_top == eden_end,
1459 // both might be null.
1460 if( i == 0 ) { // Set control for first prefetch, next follows it
1461 prefetch->init_req(0, needgc_false);
1462 }
1463 transform_later(prefetch);
1464 distance += step_size;
1465 i_o = prefetch;
1466 }
1467 }
1468 return i_o;
1469 }
1472 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
1473 expand_allocate_common(alloc, NULL,
1474 OptoRuntime::new_instance_Type(),
1475 OptoRuntime::new_instance_Java());
1476 }
1478 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
1479 Node* length = alloc->in(AllocateNode::ALength);
1480 expand_allocate_common(alloc, length,
1481 OptoRuntime::new_array_Type(),
1482 OptoRuntime::new_array_Java());
1483 }
1486 // we have determined that this lock/unlock can be eliminated, we simply
1487 // eliminate the node without expanding it.
1488 //
1489 // Note: The membar's associated with the lock/unlock are currently not
1490 // eliminated. This should be investigated as a future enhancement.
1491 //
1492 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
1494 if (!alock->is_eliminated()) {
1495 return false;
1496 }
1497 // Mark the box lock as eliminated if all correspondent locks are eliminated
1498 // to construct correct debug info.
1499 BoxLockNode* box = alock->box_node()->as_BoxLock();
1500 if (!box->is_eliminated()) {
1501 bool eliminate = true;
1502 for (DUIterator_Fast imax, i = box->fast_outs(imax); i < imax; i++) {
1503 Node *lck = box->fast_out(i);
1504 if (lck->is_Lock() && !lck->as_AbstractLock()->is_eliminated()) {
1505 eliminate = false;
1506 break;
1507 }
1508 }
1509 if (eliminate)
1510 box->set_eliminated();
1511 }
1513 #ifndef PRODUCT
1514 if (PrintEliminateLocks) {
1515 if (alock->is_Lock()) {
1516 tty->print_cr("++++ Eliminating: %d Lock", alock->_idx);
1517 } else {
1518 tty->print_cr("++++ Eliminating: %d Unlock", alock->_idx);
1519 }
1520 }
1521 #endif
1523 Node* mem = alock->in(TypeFunc::Memory);
1524 Node* ctrl = alock->in(TypeFunc::Control);
1526 extract_call_projections(alock);
1527 // There are 2 projections from the lock. The lock node will
1528 // be deleted when its last use is subsumed below.
1529 assert(alock->outcnt() == 2 &&
1530 _fallthroughproj != NULL &&
1531 _memproj_fallthrough != NULL,
1532 "Unexpected projections from Lock/Unlock");
1534 Node* fallthroughproj = _fallthroughproj;
1535 Node* memproj_fallthrough = _memproj_fallthrough;
1537 // The memory projection from a lock/unlock is RawMem
1538 // The input to a Lock is merged memory, so extract its RawMem input
1539 // (unless the MergeMem has been optimized away.)
1540 if (alock->is_Lock()) {
1541 // Seach for MemBarAcquire node and delete it also.
1542 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
1543 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquire, "");
1544 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
1545 Node* memproj = membar->proj_out(TypeFunc::Memory);
1546 _igvn.hash_delete(ctrlproj);
1547 _igvn.subsume_node(ctrlproj, fallthroughproj);
1548 _igvn.hash_delete(memproj);
1549 _igvn.subsume_node(memproj, memproj_fallthrough);
1550 }
1552 // Seach for MemBarRelease node and delete it also.
1553 if (alock->is_Unlock() && ctrl != NULL && ctrl->is_Proj() &&
1554 ctrl->in(0)->is_MemBar()) {
1555 MemBarNode* membar = ctrl->in(0)->as_MemBar();
1556 assert(membar->Opcode() == Op_MemBarRelease &&
1557 mem->is_Proj() && membar == mem->in(0), "");
1558 _igvn.hash_delete(fallthroughproj);
1559 _igvn.subsume_node(fallthroughproj, ctrl);
1560 _igvn.hash_delete(memproj_fallthrough);
1561 _igvn.subsume_node(memproj_fallthrough, mem);
1562 fallthroughproj = ctrl;
1563 memproj_fallthrough = mem;
1564 ctrl = membar->in(TypeFunc::Control);
1565 mem = membar->in(TypeFunc::Memory);
1566 }
1568 _igvn.hash_delete(fallthroughproj);
1569 _igvn.subsume_node(fallthroughproj, ctrl);
1570 _igvn.hash_delete(memproj_fallthrough);
1571 _igvn.subsume_node(memproj_fallthrough, mem);
1572 return true;
1573 }
1576 //------------------------------expand_lock_node----------------------
1577 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
1579 Node* ctrl = lock->in(TypeFunc::Control);
1580 Node* mem = lock->in(TypeFunc::Memory);
1581 Node* obj = lock->obj_node();
1582 Node* box = lock->box_node();
1583 Node* flock = lock->fastlock_node();
1585 // Make the merge point
1586 Node *region = new (C, 3) RegionNode(3);
1588 Node *bol = transform_later(new (C, 2) BoolNode(flock,BoolTest::ne));
1589 Node *iff = new (C, 2) IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
1590 // Optimize test; set region slot 2
1591 Node *slow_path = opt_iff(region,iff);
1593 // Make slow path call
1594 CallNode *call = make_slow_call( (CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(), OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path, obj, box );
1596 extract_call_projections(call);
1598 // Slow path can only throw asynchronous exceptions, which are always
1599 // de-opted. So the compiler thinks the slow-call can never throw an
1600 // exception. If it DOES throw an exception we would need the debug
1601 // info removed first (since if it throws there is no monitor).
1602 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
1603 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
1605 // Capture slow path
1606 // disconnect fall-through projection from call and create a new one
1607 // hook up users of fall-through projection to region
1608 Node *slow_ctrl = _fallthroughproj->clone();
1609 transform_later(slow_ctrl);
1610 _igvn.hash_delete(_fallthroughproj);
1611 _fallthroughproj->disconnect_inputs(NULL);
1612 region->init_req(1, slow_ctrl);
1613 // region inputs are now complete
1614 transform_later(region);
1615 _igvn.subsume_node(_fallthroughproj, region);
1617 // create a Phi for the memory state
1618 Node *mem_phi = new (C, 3) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
1619 Node *memproj = transform_later( new (C, 1) ProjNode(call, TypeFunc::Memory) );
1620 mem_phi->init_req(1, memproj );
1621 mem_phi->init_req(2, mem);
1622 transform_later(mem_phi);
1623 _igvn.hash_delete(_memproj_fallthrough);
1624 _igvn.subsume_node(_memproj_fallthrough, mem_phi);
1627 }
1629 //------------------------------expand_unlock_node----------------------
1630 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
1632 Node* ctrl = unlock->in(TypeFunc::Control);
1633 Node* mem = unlock->in(TypeFunc::Memory);
1634 Node* obj = unlock->obj_node();
1635 Node* box = unlock->box_node();
1637 // No need for a null check on unlock
1639 // Make the merge point
1640 RegionNode *region = new (C, 3) RegionNode(3);
1642 FastUnlockNode *funlock = new (C, 3) FastUnlockNode( ctrl, obj, box );
1643 funlock = transform_later( funlock )->as_FastUnlock();
1644 Node *bol = transform_later(new (C, 2) BoolNode(funlock,BoolTest::ne));
1645 Node *iff = new (C, 2) IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
1646 // Optimize test; set region slot 2
1647 Node *slow_path = opt_iff(region,iff);
1649 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 );
1651 extract_call_projections(call);
1653 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
1654 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
1656 // No exceptions for unlocking
1657 // Capture slow path
1658 // disconnect fall-through projection from call and create a new one
1659 // hook up users of fall-through projection to region
1660 Node *slow_ctrl = _fallthroughproj->clone();
1661 transform_later(slow_ctrl);
1662 _igvn.hash_delete(_fallthroughproj);
1663 _fallthroughproj->disconnect_inputs(NULL);
1664 region->init_req(1, slow_ctrl);
1665 // region inputs are now complete
1666 transform_later(region);
1667 _igvn.subsume_node(_fallthroughproj, region);
1669 // create a Phi for the memory state
1670 Node *mem_phi = new (C, 3) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
1671 Node *memproj = transform_later( new(C, 1) ProjNode(call, TypeFunc::Memory) );
1672 mem_phi->init_req(1, memproj );
1673 mem_phi->init_req(2, mem);
1674 transform_later(mem_phi);
1675 _igvn.hash_delete(_memproj_fallthrough);
1676 _igvn.subsume_node(_memproj_fallthrough, mem_phi);
1679 }
1681 //------------------------------expand_macro_nodes----------------------
1682 // Returns true if a failure occurred.
1683 bool PhaseMacroExpand::expand_macro_nodes() {
1684 if (C->macro_count() == 0)
1685 return false;
1686 // attempt to eliminate allocations
1687 bool progress = true;
1688 while (progress) {
1689 progress = false;
1690 for (int i = C->macro_count(); i > 0; i--) {
1691 Node * n = C->macro_node(i-1);
1692 bool success = false;
1693 debug_only(int old_macro_count = C->macro_count(););
1694 switch (n->class_id()) {
1695 case Node::Class_Allocate:
1696 case Node::Class_AllocateArray:
1697 success = eliminate_allocate_node(n->as_Allocate());
1698 break;
1699 case Node::Class_Lock:
1700 case Node::Class_Unlock:
1701 success = eliminate_locking_node(n->as_AbstractLock());
1702 break;
1703 default:
1704 if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) {
1705 _igvn.add_users_to_worklist(n);
1706 _igvn.hash_delete(n);
1707 _igvn.subsume_node(n, n->in(1));
1708 success = true;
1709 } else {
1710 assert(false, "unknown node type in macro list");
1711 }
1712 }
1713 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
1714 progress = progress || success;
1715 }
1716 }
1717 // Make sure expansion will not cause node limit to be exceeded.
1718 // Worst case is a macro node gets expanded into about 50 nodes.
1719 // Allow 50% more for optimization.
1720 if (C->check_node_count(C->macro_count() * 75, "out of nodes before macro expansion" ) )
1721 return true;
1723 // expand "macro" nodes
1724 // nodes are removed from the macro list as they are processed
1725 while (C->macro_count() > 0) {
1726 int macro_count = C->macro_count();
1727 Node * n = C->macro_node(macro_count-1);
1728 assert(n->is_macro(), "only macro nodes expected here");
1729 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) {
1730 // node is unreachable, so don't try to expand it
1731 C->remove_macro_node(n);
1732 continue;
1733 }
1734 switch (n->class_id()) {
1735 case Node::Class_Allocate:
1736 expand_allocate(n->as_Allocate());
1737 break;
1738 case Node::Class_AllocateArray:
1739 expand_allocate_array(n->as_AllocateArray());
1740 break;
1741 case Node::Class_Lock:
1742 expand_lock_node(n->as_Lock());
1743 break;
1744 case Node::Class_Unlock:
1745 expand_unlock_node(n->as_Unlock());
1746 break;
1747 default:
1748 assert(false, "unknown node type in macro list");
1749 }
1750 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
1751 if (C->failing()) return true;
1752 }
1754 _igvn.set_delay_transform(false);
1755 _igvn.optimize();
1756 return false;
1757 }