1 /*
2 * Copyright 2005-2007 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->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_iff(Node* region, Node* iff) {
86 IfNode *opt_iff = transform_later(iff)->as_If();
87
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
91
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 }
96
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 }
106
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) {
109
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 );
115
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);
125
126 return call;
127 }
128
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 }
181
182 }
183
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 }
195
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) {
198 Node *orig_mem = mem;
199 Node *alloc_mem = alloc->in(TypeFunc::Memory);
200 while (true) {
201 if (mem == alloc_mem || mem == start_mem ) {
202 return mem; // hit one of our sentinals
203 } else if (mem->is_MergeMem()) {
204 mem = mem->as_MergeMem()->memory_at(alias_idx);
205 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
206 Node *in = mem->in(0);
207 // we can safely skip over safepoints, calls, locks and membars because we
208 // already know that the object is safe to eliminate.
209 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
210 return in;
211 } else if (in->is_Call() || in->is_MemBar()) {
212 mem = in->in(TypeFunc::Memory);
213 } else {
214 assert(false, "unexpected projection");
215 }
216 } else if (mem->is_Store()) {
217 const TypePtr* atype = mem->as_Store()->adr_type();
218 int adr_idx = Compile::current()->get_alias_index(atype);
219 if (adr_idx == alias_idx) {
220 assert(atype->isa_oopptr(), "address type must be oopptr");
221 int adr_offset = atype->offset();
222 uint adr_iid = atype->is_oopptr()->instance_id();
223 // Array elements references have the same alias_idx
224 // but different offset and different instance_id.
225 if (adr_offset == offset && adr_iid == alloc->_idx)
226 return mem;
227 } else {
228 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
229 }
230 mem = mem->in(MemNode::Memory);
231 } else {
232 return mem;
233 }
234 if (mem == orig_mem)
235 return mem;
236 }
237 }
238
239 //
240 // Given a Memory Phi, compute a value Phi containing the values from stores
241 // on the input paths.
242 // Note: this function is recursive, its depth is limied by the "level" argument
243 // Returns the computed Phi, or NULL if it cannot compute it.
244 Node *PhaseMacroExpand::value_from_mem_phi(Node *mem, BasicType ft, const Type *phi_type, const TypeOopPtr *adr_t, Node *alloc, int level) {
245
246 if (level <= 0) {
247 return NULL;
248 }
249 int alias_idx = C->get_alias_index(adr_t);
250 int offset = adr_t->offset();
251 int instance_id = adr_t->instance_id();
252
253 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
254 Node *alloc_mem = alloc->in(TypeFunc::Memory);
255
256 uint length = mem->req();
257 GrowableArray <Node *> values(length, length, NULL);
258
259 for (uint j = 1; j < length; j++) {
260 Node *in = mem->in(j);
261 if (in == NULL || in->is_top()) {
262 values.at_put(j, in);
263 } else {
264 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc);
265 if (val == start_mem || val == alloc_mem) {
266 // hit a sentinel, return appropriate 0 value
267 values.at_put(j, _igvn.zerocon(ft));
268 continue;
269 }
270 if (val->is_Initialize()) {
271 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
272 }
273 if (val == NULL) {
274 return NULL; // can't find a value on this path
275 }
276 if (val == mem) {
277 values.at_put(j, mem);
278 } else if (val->is_Store()) {
279 values.at_put(j, val->in(MemNode::ValueIn));
280 } else if(val->is_Proj() && val->in(0) == alloc) {
281 values.at_put(j, _igvn.zerocon(ft));
282 } else if (val->is_Phi()) {
283 // Check if an appropriate node already exists.
284 Node* region = val->in(0);
285 Node* old_phi = NULL;
286 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
287 Node* phi = region->fast_out(k);
288 if (phi->is_Phi() && phi != val &&
289 phi->as_Phi()->is_same_inst_field(phi_type, instance_id, alias_idx, offset)) {
290 old_phi = phi;
291 break;
292 }
293 }
294 if (old_phi == NULL) {
295 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, level-1);
296 if (val == NULL) {
297 return NULL;
298 }
299 values.at_put(j, val);
300 } else {
301 values.at_put(j, old_phi);
302 }
303 } else {
304 return NULL; // unknown node on this path
305 }
306 }
307 }
308 // create a new Phi for the value
309 PhiNode *phi = new (C, length) PhiNode(mem->in(0), phi_type, NULL, instance_id, alias_idx, offset);
310 for (uint j = 1; j < length; j++) {
311 if (values.at(j) == mem) {
312 phi->init_req(j, phi);
313 } else {
314 phi->init_req(j, values.at(j));
315 }
316 }
317 transform_later(phi);
318 return phi;
319 }
320
321 // Search the last value stored into the object's field.
322 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, Node *alloc) {
323 assert(adr_t->is_instance_field(), "instance required");
324 uint instance_id = adr_t->instance_id();
325 assert(instance_id == alloc->_idx, "wrong allocation");
326
327 int alias_idx = C->get_alias_index(adr_t);
328 int offset = adr_t->offset();
329 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
330 Node *alloc_ctrl = alloc->in(TypeFunc::Control);
331 Node *alloc_mem = alloc->in(TypeFunc::Memory);
332 VectorSet visited(Thread::current()->resource_area());
333
334
335 bool done = sfpt_mem == alloc_mem;
336 Node *mem = sfpt_mem;
337 while (!done) {
338 if (visited.test_set(mem->_idx)) {
339 return NULL; // found a loop, give up
340 }
341 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc);
342 if (mem == start_mem || mem == alloc_mem) {
343 done = true; // hit a sentinel, return appropriate 0 value
344 } else if (mem->is_Initialize()) {
345 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
346 if (mem == NULL) {
347 done = true; // Something go wrong.
348 } else if (mem->is_Store()) {
349 const TypePtr* atype = mem->as_Store()->adr_type();
350 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
351 done = true;
352 }
353 } else if (mem->is_Store()) {
354 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
355 assert(atype != NULL, "address type must be oopptr");
356 assert(C->get_alias_index(atype) == alias_idx &&
357 atype->is_instance_field() && atype->offset() == offset &&
358 atype->instance_id() == instance_id, "store is correct memory slice");
359 done = true;
360 } else if (mem->is_Phi()) {
361 // try to find a phi's unique input
362 Node *unique_input = NULL;
363 Node *top = C->top();
364 for (uint i = 1; i < mem->req(); i++) {
365 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc);
366 if (n == NULL || n == top || n == mem) {
367 continue;
368 } else if (unique_input == NULL) {
369 unique_input = n;
370 } else if (unique_input != n) {
371 unique_input = top;
372 break;
373 }
374 }
375 if (unique_input != NULL && unique_input != top) {
376 mem = unique_input;
377 } else {
378 done = true;
379 }
380 } else {
381 assert(false, "unexpected node");
382 }
383 }
384 if (mem != NULL) {
385 if (mem == start_mem || mem == alloc_mem) {
386 // hit a sentinel, return appropriate 0 value
387 return _igvn.zerocon(ft);
388 } else if (mem->is_Store()) {
389 return mem->in(MemNode::ValueIn);
390 } else if (mem->is_Phi()) {
391 // attempt to produce a Phi reflecting the values on the input paths of the Phi
392 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, 8);
393 if (phi != NULL) {
394 return phi;
395 }
396 }
397 }
398 // Something go wrong.
399 return NULL;
400 }
401
402 // Check the possibility of scalar replacement.
403 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
404 // Scan the uses of the allocation to check for anything that would
405 // prevent us from eliminating it.
406 NOT_PRODUCT( const char* fail_eliminate = NULL; )
407 DEBUG_ONLY( Node* disq_node = NULL; )
408 bool can_eliminate = true;
409
410 Node* res = alloc->result_cast();
411 const TypeOopPtr* res_type = NULL;
412 if (res == NULL) {
413 // All users were eliminated.
414 } else if (!res->is_CheckCastPP()) {
415 alloc->_is_scalar_replaceable = false; // don't try again
416 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
417 can_eliminate = false;
418 } else {
419 res_type = _igvn.type(res)->isa_oopptr();
420 if (res_type == NULL) {
421 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
422 can_eliminate = false;
423 } else if (res_type->isa_aryptr()) {
424 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
425 if (length < 0) {
426 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
427 can_eliminate = false;
428 }
429 }
430 }
431
432 if (can_eliminate && res != NULL) {
433 for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
434 j < jmax && can_eliminate; j++) {
435 Node* use = res->fast_out(j);
436
437 if (use->is_AddP()) {
438 const TypePtr* addp_type = _igvn.type(use)->is_ptr();
439 int offset = addp_type->offset();
440
441 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
442 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";)
443 can_eliminate = false;
444 break;
445 }
446 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
447 k < kmax && can_eliminate; k++) {
448 Node* n = use->fast_out(k);
449 if (!n->is_Store() && n->Opcode() != Op_CastP2X) {
450 DEBUG_ONLY(disq_node = n;)
451 if (n->is_Load()) {
452 NOT_PRODUCT(fail_eliminate = "Field load";)
453 } else {
454 NOT_PRODUCT(fail_eliminate = "Not store field referrence";)
455 }
456 can_eliminate = false;
457 }
458 }
459 } else if (use->is_SafePoint()) {
460 SafePointNode* sfpt = use->as_SafePoint();
461 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
462 // Object is passed as argument.
463 DEBUG_ONLY(disq_node = use;)
464 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
465 can_eliminate = false;
466 }
467 Node* sfptMem = sfpt->memory();
468 if (sfptMem == NULL || sfptMem->is_top()) {
469 DEBUG_ONLY(disq_node = use;)
470 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";)
471 can_eliminate = false;
472 } else {
473 safepoints.append_if_missing(sfpt);
474 }
475 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
476 if (use->is_Phi()) {
477 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
478 NOT_PRODUCT(fail_eliminate = "Object is return value";)
479 } else {
480 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
481 }
482 DEBUG_ONLY(disq_node = use;)
483 } else {
484 if (use->Opcode() == Op_Return) {
485 NOT_PRODUCT(fail_eliminate = "Object is return value";)
486 }else {
487 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
488 }
489 DEBUG_ONLY(disq_node = use;)
490 }
491 can_eliminate = false;
492 }
493 }
494 }
495
496 #ifndef PRODUCT
497 if (PrintEliminateAllocations) {
498 if (can_eliminate) {
499 tty->print("Scalar ");
500 if (res == NULL)
501 alloc->dump();
502 else
503 res->dump();
504 } else {
505 tty->print("NotScalar (%s)", fail_eliminate);
506 if (res == NULL)
507 alloc->dump();
508 else
509 res->dump();
510 #ifdef ASSERT
511 if (disq_node != NULL) {
512 tty->print(" >>>> ");
513 disq_node->dump();
514 }
515 #endif /*ASSERT*/
516 }
517 }
518 #endif
519 return can_eliminate;
520 }
521
522 // Do scalar replacement.
523 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
524 GrowableArray <SafePointNode *> safepoints_done;
525
526 ciKlass* klass = NULL;
527 ciInstanceKlass* iklass = NULL;
528 int nfields = 0;
529 int array_base;
530 int element_size;
531 BasicType basic_elem_type;
532 ciType* elem_type;
533
534 Node* res = alloc->result_cast();
535 const TypeOopPtr* res_type = NULL;
536 if (res != NULL) { // Could be NULL when there are no users
537 res_type = _igvn.type(res)->isa_oopptr();
538 }
539
540 if (res != NULL) {
541 klass = res_type->klass();
542 if (res_type->isa_instptr()) {
543 // find the fields of the class which will be needed for safepoint debug information
544 assert(klass->is_instance_klass(), "must be an instance klass.");
545 iklass = klass->as_instance_klass();
546 nfields = iklass->nof_nonstatic_fields();
547 } else {
548 // find the array's elements which will be needed for safepoint debug information
549 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
550 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass.");
551 elem_type = klass->as_array_klass()->element_type();
552 basic_elem_type = elem_type->basic_type();
553 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
554 element_size = type2aelembytes(basic_elem_type);
555 }
556 }
557 //
558 // Process the safepoint uses
559 //
560 while (safepoints.length() > 0) {
561 SafePointNode* sfpt = safepoints.pop();
562 Node* mem = sfpt->memory();
563 uint first_ind = sfpt->req();
564 SafePointScalarObjectNode* sobj = new (C, 1) SafePointScalarObjectNode(res_type,
565 #ifdef ASSERT
566 alloc,
567 #endif
568 first_ind, nfields);
569 sobj->init_req(0, sfpt->in(TypeFunc::Control));
570 transform_later(sobj);
571
572 // Scan object's fields adding an input to the safepoint for each field.
573 for (int j = 0; j < nfields; j++) {
574 int offset;
575 ciField* field = NULL;
576 if (iklass != NULL) {
577 field = iklass->nonstatic_field_at(j);
578 offset = field->offset();
579 elem_type = field->type();
580 basic_elem_type = field->layout_type();
581 } else {
582 offset = array_base + j * element_size;
583 }
584
585 const Type *field_type;
586 // The next code is taken from Parse::do_get_xxx().
587 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) {
588 if (!elem_type->is_loaded()) {
589 field_type = TypeInstPtr::BOTTOM;
590 } else if (field != NULL && field->is_constant()) {
591 // This can happen if the constant oop is non-perm.
592 ciObject* con = field->constant_value().as_object();
593 // Do not "join" in the previous type; it doesn't add value,
594 // and may yield a vacuous result if the field is of interface type.
595 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
596 assert(field_type != NULL, "field singleton type must be consistent");
597 } else {
598 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
599 }
600 if (UseCompressedOops) {
601 field_type = field_type->is_oopptr()->make_narrowoop();
602 basic_elem_type = T_NARROWOOP;
603 }
604 } else {
605 field_type = Type::get_const_basic_type(basic_elem_type);
606 }
607
608 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr();
609
610 Node *field_val = value_from_mem(mem, basic_elem_type, field_type, field_addr_type, alloc);
611 if (field_val == NULL) {
612 // we weren't able to find a value for this field,
613 // give up on eliminating this allocation
614 alloc->_is_scalar_replaceable = false; // don't try again
615 // remove any extra entries we added to the safepoint
616 uint last = sfpt->req() - 1;
617 for (int k = 0; k < j; k++) {
618 sfpt->del_req(last--);
619 }
620 // rollback processed safepoints
621 while (safepoints_done.length() > 0) {
622 SafePointNode* sfpt_done = safepoints_done.pop();
623 // remove any extra entries we added to the safepoint
624 last = sfpt_done->req() - 1;
625 for (int k = 0; k < nfields; k++) {
626 sfpt_done->del_req(last--);
627 }
628 JVMState *jvms = sfpt_done->jvms();
629 jvms->set_endoff(sfpt_done->req());
630 // Now make a pass over the debug information replacing any references
631 // to SafePointScalarObjectNode with the allocated object.
632 int start = jvms->debug_start();
633 int end = jvms->debug_end();
634 for (int i = start; i < end; i++) {
635 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
636 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
637 if (scobj->first_index() == sfpt_done->req() &&
638 scobj->n_fields() == (uint)nfields) {
639 assert(scobj->alloc() == alloc, "sanity");
640 sfpt_done->set_req(i, res);
641 }
642 }
643 }
644 }
645 #ifndef PRODUCT
646 if (PrintEliminateAllocations) {
647 if (field != NULL) {
648 tty->print("=== At SafePoint node %d can't find value of Field: ",
649 sfpt->_idx);
650 field->print();
651 int field_idx = C->get_alias_index(field_addr_type);
652 tty->print(" (alias_idx=%d)", field_idx);
653 } else { // Array's element
654 tty->print("=== At SafePoint node %d can't find value of array element [%d]",
655 sfpt->_idx, j);
656 }
657 tty->print(", which prevents elimination of: ");
658 if (res == NULL)
659 alloc->dump();
660 else
661 res->dump();
662 }
663 #endif
664 return false;
665 }
666 if (UseCompressedOops && field_type->isa_narrowoop()) {
667 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
668 // to be able scalar replace the allocation.
669 _igvn.set_delay_transform(false);
670 field_val = DecodeNNode::decode(&_igvn, field_val);
671 _igvn.set_delay_transform(true);
672 }
673 sfpt->add_req(field_val);
674 }
675 JVMState *jvms = sfpt->jvms();
676 jvms->set_endoff(sfpt->req());
677 // Now make a pass over the debug information replacing any references
678 // to the allocated object with "sobj"
679 int start = jvms->debug_start();
680 int end = jvms->debug_end();
681 for (int i = start; i < end; i++) {
682 if (sfpt->in(i) == res) {
683 sfpt->set_req(i, sobj);
684 }
685 }
686 safepoints_done.append_if_missing(sfpt); // keep it for rollback
687 }
688 return true;
689 }
690
691 // Process users of eliminated allocation.
692 void PhaseMacroExpand::process_users_of_allocation(AllocateNode *alloc) {
693 Node* res = alloc->result_cast();
694 if (res != NULL) {
695 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
696 Node *use = res->last_out(j);
697 uint oc1 = res->outcnt();
698
699 if (use->is_AddP()) {
700 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
701 Node *n = use->last_out(k);
702 uint oc2 = use->outcnt();
703 if (n->is_Store()) {
704 _igvn.replace_node(n, n->in(MemNode::Memory));
705 } else {
706 assert( n->Opcode() == Op_CastP2X, "CastP2X required");
707 eliminate_card_mark(n);
708 }
709 k -= (oc2 - use->outcnt());
710 }
711 } else {
712 assert( !use->is_SafePoint(), "safepoint uses must have been already elimiated");
713 assert( use->Opcode() == Op_CastP2X, "CastP2X required");
714 eliminate_card_mark(use);
715 }
716 j -= (oc1 - res->outcnt());
717 }
718 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
719 _igvn.remove_dead_node(res);
720 }
721
722 //
723 // Process other users of allocation's projections
724 //
725 if (_resproj != NULL && _resproj->outcnt() != 0) {
726 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) {
727 Node *use = _resproj->last_out(j);
728 uint oc1 = _resproj->outcnt();
729 if (use->is_Initialize()) {
730 // Eliminate Initialize node.
731 InitializeNode *init = use->as_Initialize();
732 assert(init->outcnt() <= 2, "only a control and memory projection expected");
733 Node *ctrl_proj = init->proj_out(TypeFunc::Control);
734 if (ctrl_proj != NULL) {
735 assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection");
736 _igvn.replace_node(ctrl_proj, _fallthroughcatchproj);
737 }
738 Node *mem_proj = init->proj_out(TypeFunc::Memory);
739 if (mem_proj != NULL) {
740 Node *mem = init->in(TypeFunc::Memory);
741 #ifdef ASSERT
742 if (mem->is_MergeMem()) {
743 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection");
744 } else {
745 assert(mem == _memproj_fallthrough, "allocation memory projection");
746 }
747 #endif
748 _igvn.replace_node(mem_proj, mem);
749 }
750 } else if (use->is_AddP()) {
751 // raw memory addresses used only by the initialization
752 _igvn.hash_delete(use);
753 _igvn.subsume_node(use, C->top());
754 } else {
755 assert(false, "only Initialize or AddP expected");
756 }
757 j -= (oc1 - _resproj->outcnt());
758 }
759 }
760 if (_fallthroughcatchproj != NULL) {
761 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control));
762 }
763 if (_memproj_fallthrough != NULL) {
764 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory));
765 }
766 if (_memproj_catchall != NULL) {
767 _igvn.replace_node(_memproj_catchall, C->top());
768 }
769 if (_ioproj_fallthrough != NULL) {
770 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O));
771 }
772 if (_ioproj_catchall != NULL) {
773 _igvn.replace_node(_ioproj_catchall, C->top());
774 }
775 if (_catchallcatchproj != NULL) {
776 _igvn.replace_node(_catchallcatchproj, C->top());
777 }
778 }
779
780 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
781
782 if (!EliminateAllocations || !alloc->_is_scalar_replaceable) {
783 return false;
784 }
785
786 extract_call_projections(alloc);
787
788 GrowableArray <SafePointNode *> safepoints;
789 if (!can_eliminate_allocation(alloc, safepoints)) {
790 return false;
791 }
792
793 if (!scalar_replacement(alloc, safepoints)) {
794 return false;
795 }
796
797 process_users_of_allocation(alloc);
798
799 #ifndef PRODUCT
800 if (PrintEliminateAllocations) {
801 if (alloc->is_AllocateArray())
802 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
803 else
804 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
805 }
806 #endif
807
808 return true;
809 }
810
811
812 //---------------------------set_eden_pointers-------------------------
813 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) {
814 if (UseTLAB) { // Private allocation: load from TLS
815 Node* thread = transform_later(new (C, 1) ThreadLocalNode());
816 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset());
817 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset());
818 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset);
819 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset);
820 } else { // Shared allocation: load from globals
821 CollectedHeap* ch = Universe::heap();
822 address top_adr = (address)ch->top_addr();
823 address end_adr = (address)ch->end_addr();
824 eden_top_adr = makecon(TypeRawPtr::make(top_adr));
825 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr);
826 }
827 }
828
829
830 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
831 Node* adr = basic_plus_adr(base, offset);
832 const TypePtr* adr_type = TypeRawPtr::BOTTOM;
833 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt);
834 transform_later(value);
835 return value;
836 }
837
838
839 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
840 Node* adr = basic_plus_adr(base, offset);
841 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt);
842 transform_later(mem);
843 return mem;
844 }
845
846 //=============================================================================
847 //
848 // A L L O C A T I O N
849 //
850 // Allocation attempts to be fast in the case of frequent small objects.
851 // It breaks down like this:
852 //
853 // 1) Size in doublewords is computed. This is a constant for objects and
854 // variable for most arrays. Doubleword units are used to avoid size
855 // overflow of huge doubleword arrays. We need doublewords in the end for
856 // rounding.
857 //
858 // 2) Size is checked for being 'too large'. Too-large allocations will go
859 // the slow path into the VM. The slow path can throw any required
860 // exceptions, and does all the special checks for very large arrays. The
861 // size test can constant-fold away for objects. For objects with
862 // finalizers it constant-folds the otherway: you always go slow with
863 // finalizers.
864 //
865 // 3) If NOT using TLABs, this is the contended loop-back point.
866 // Load-Locked the heap top. If using TLABs normal-load the heap top.
867 //
868 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
869 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
870 // "size*8" we always enter the VM, where "largish" is a constant picked small
871 // enough that there's always space between the eden max and 4Gig (old space is
872 // there so it's quite large) and large enough that the cost of entering the VM
873 // is dwarfed by the cost to initialize the space.
874 //
875 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
876 // down. If contended, repeat at step 3. If using TLABs normal-store
877 // adjusted heap top back down; there is no contention.
878 //
879 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
880 // fields.
881 //
882 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
883 // oop flavor.
884 //
885 //=============================================================================
886 // FastAllocateSizeLimit value is in DOUBLEWORDS.
887 // Allocations bigger than this always go the slow route.
888 // This value must be small enough that allocation attempts that need to
889 // trigger exceptions go the slow route. Also, it must be small enough so
890 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
891 //=============================================================================j//
892 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
893 // The allocator will coalesce int->oop copies away. See comment in
894 // coalesce.cpp about how this works. It depends critically on the exact
895 // code shape produced here, so if you are changing this code shape
896 // make sure the GC info for the heap-top is correct in and around the
897 // slow-path call.
898 //
899
900 void PhaseMacroExpand::expand_allocate_common(
901 AllocateNode* alloc, // allocation node to be expanded
902 Node* length, // array length for an array allocation
903 const TypeFunc* slow_call_type, // Type of slow call
904 address slow_call_address // Address of slow call
905 )
906 {
907
908 Node* ctrl = alloc->in(TypeFunc::Control);
909 Node* mem = alloc->in(TypeFunc::Memory);
910 Node* i_o = alloc->in(TypeFunc::I_O);
911 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
912 Node* klass_node = alloc->in(AllocateNode::KlassNode);
913 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
914
915 // With escape analysis, the entire memory state was needed to be able to
916 // eliminate the allocation. Since the allocations cannot be eliminated,
917 // optimize it to the raw slice.
918 if (mem->is_MergeMem()) {
919 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
920 }
921
922 Node* eden_top_adr;
923 Node* eden_end_adr;
924 set_eden_pointers(eden_top_adr, eden_end_adr);
925
926 uint raw_idx = C->get_alias_index(TypeRawPtr::BOTTOM);
927 assert(ctrl != NULL, "must have control");
928
929 // Load Eden::end. Loop invariant and hoisted.
930 //
931 // Note: We set the control input on "eden_end" and "old_eden_top" when using
932 // a TLAB to work around a bug where these values were being moved across
933 // a safepoint. These are not oops, so they cannot be include in the oop
934 // map, but the can be changed by a GC. The proper way to fix this would
935 // be to set the raw memory state when generating a SafepointNode. However
936 // this will require extensive changes to the loop optimization in order to
937 // prevent a degradation of the optimization.
938 // See comment in memnode.hpp, around line 227 in class LoadPNode.
939 Node* eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS);
940
941 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
942 // they will not be used if "always_slow" is set
943 enum { slow_result_path = 1, fast_result_path = 2 };
944 Node *result_region;
945 Node *result_phi_rawmem;
946 Node *result_phi_rawoop;
947 Node *result_phi_i_o;
948
949 // The initial slow comparison is a size check, the comparison
950 // we want to do is a BoolTest::gt
951 bool always_slow = false;
952 int tv = _igvn.find_int_con(initial_slow_test, -1);
953 if (tv >= 0) {
954 always_slow = (tv == 1);
955 initial_slow_test = NULL;
956 } else {
957 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
958 }
959
960 if (DTraceAllocProbes) {
961 // Force slow-path allocation
962 always_slow = true;
963 initial_slow_test = NULL;
964 }
965
966 enum { too_big_or_final_path = 1, need_gc_path = 2 };
967 Node *slow_region = NULL;
968 Node *toobig_false = ctrl;
969
970 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent");
971 // generate the initial test if necessary
972 if (initial_slow_test != NULL ) {
973 slow_region = new (C, 3) RegionNode(3);
974
975 // Now make the initial failure test. Usually a too-big test but
976 // might be a TRUE for finalizers or a fancy class check for
977 // newInstance0.
978 IfNode *toobig_iff = new (C, 2) IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
979 transform_later(toobig_iff);
980 // Plug the failing-too-big test into the slow-path region
981 Node *toobig_true = new (C, 1) IfTrueNode( toobig_iff );
982 transform_later(toobig_true);
983 slow_region ->init_req( too_big_or_final_path, toobig_true );
984 toobig_false = new (C, 1) IfFalseNode( toobig_iff );
985 transform_later(toobig_false);
986 } else { // No initial test, just fall into next case
987 toobig_false = ctrl;
988 debug_only(slow_region = NodeSentinel);
989 }
990
991 Node *slow_mem = mem; // save the current memory state for slow path
992 // generate the fast allocation code unless we know that the initial test will always go slow
993 if (!always_slow) {
994 // allocate the Region and Phi nodes for the result
995 result_region = new (C, 3) RegionNode(3);
996 result_phi_rawmem = new (C, 3) PhiNode( result_region, Type::MEMORY, TypeRawPtr::BOTTOM );
997 result_phi_rawoop = new (C, 3) PhiNode( result_region, TypeRawPtr::BOTTOM );
998 result_phi_i_o = new (C, 3) PhiNode( result_region, Type::ABIO ); // I/O is used for Prefetch
999
1000 // We need a Region for the loop-back contended case.
1001 enum { fall_in_path = 1, contended_loopback_path = 2 };
1002 Node *contended_region;
1003 Node *contended_phi_rawmem;
1004 if( UseTLAB ) {
1005 contended_region = toobig_false;
1006 contended_phi_rawmem = mem;
1007 } else {
1008 contended_region = new (C, 3) RegionNode(3);
1009 contended_phi_rawmem = new (C, 3) PhiNode( contended_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1010 // Now handle the passing-too-big test. We fall into the contended
1011 // loop-back merge point.
1012 contended_region ->init_req( fall_in_path, toobig_false );
1013 contended_phi_rawmem->init_req( fall_in_path, mem );
1014 transform_later(contended_region);
1015 transform_later(contended_phi_rawmem);
1016 }
1017
1018 // Load(-locked) the heap top.
1019 // See note above concerning the control input when using a TLAB
1020 Node *old_eden_top = UseTLAB
1021 ? new (C, 3) LoadPNode ( ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM )
1022 : new (C, 3) LoadPLockedNode( contended_region, contended_phi_rawmem, eden_top_adr );
1023
1024 transform_later(old_eden_top);
1025 // Add to heap top to get a new heap top
1026 Node *new_eden_top = new (C, 4) AddPNode( top(), old_eden_top, size_in_bytes );
1027 transform_later(new_eden_top);
1028 // Check for needing a GC; compare against heap end
1029 Node *needgc_cmp = new (C, 3) CmpPNode( new_eden_top, eden_end );
1030 transform_later(needgc_cmp);
1031 Node *needgc_bol = new (C, 2) BoolNode( needgc_cmp, BoolTest::ge );
1032 transform_later(needgc_bol);
1033 IfNode *needgc_iff = new (C, 2) IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1034 transform_later(needgc_iff);
1035
1036 // Plug the failing-heap-space-need-gc test into the slow-path region
1037 Node *needgc_true = new (C, 1) IfTrueNode( needgc_iff );
1038 transform_later(needgc_true);
1039 if( initial_slow_test ) {
1040 slow_region ->init_req( need_gc_path, needgc_true );
1041 // This completes all paths into the slow merge point
1042 transform_later(slow_region);
1043 } else { // No initial slow path needed!
1044 // Just fall from the need-GC path straight into the VM call.
1045 slow_region = needgc_true;
1046 }
1047 // No need for a GC. Setup for the Store-Conditional
1048 Node *needgc_false = new (C, 1) IfFalseNode( needgc_iff );
1049 transform_later(needgc_false);
1050
1051 // Grab regular I/O before optional prefetch may change it.
1052 // Slow-path does no I/O so just set it to the original I/O.
1053 result_phi_i_o->init_req( slow_result_path, i_o );
1054
1055 i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem,
1056 old_eden_top, new_eden_top, length);
1057
1058 // Store (-conditional) the modified eden top back down.
1059 // StorePConditional produces flags for a test PLUS a modified raw
1060 // memory state.
1061 Node *store_eden_top;
1062 Node *fast_oop_ctrl;
1063 if( UseTLAB ) {
1064 store_eden_top = new (C, 4) StorePNode( needgc_false, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, new_eden_top );
1065 transform_later(store_eden_top);
1066 fast_oop_ctrl = needgc_false; // No contention, so this is the fast path
1067 } else {
1068 store_eden_top = new (C, 5) StorePConditionalNode( needgc_false, contended_phi_rawmem, eden_top_adr, new_eden_top, old_eden_top );
1069 transform_later(store_eden_top);
1070 Node *contention_check = new (C, 2) BoolNode( store_eden_top, BoolTest::ne );
1071 transform_later(contention_check);
1072 store_eden_top = new (C, 1) SCMemProjNode(store_eden_top);
1073 transform_later(store_eden_top);
1074
1075 // If not using TLABs, check to see if there was contention.
1076 IfNode *contention_iff = new (C, 2) IfNode ( needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN );
1077 transform_later(contention_iff);
1078 Node *contention_true = new (C, 1) IfTrueNode( contention_iff );
1079 transform_later(contention_true);
1080 // If contention, loopback and try again.
1081 contended_region->init_req( contended_loopback_path, contention_true );
1082 contended_phi_rawmem->init_req( contended_loopback_path, store_eden_top );
1083
1084 // Fast-path succeeded with no contention!
1085 Node *contention_false = new (C, 1) IfFalseNode( contention_iff );
1086 transform_later(contention_false);
1087 fast_oop_ctrl = contention_false;
1088 }
1089
1090 // Rename successful fast-path variables to make meaning more obvious
1091 Node* fast_oop = old_eden_top;
1092 Node* fast_oop_rawmem = store_eden_top;
1093 fast_oop_rawmem = initialize_object(alloc,
1094 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1095 klass_node, length, size_in_bytes);
1096
1097 if (ExtendedDTraceProbes) {
1098 // Slow-path call
1099 int size = TypeFunc::Parms + 2;
1100 CallLeafNode *call = new (C, size) CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1101 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base),
1102 "dtrace_object_alloc",
1103 TypeRawPtr::BOTTOM);
1104
1105 // Get base of thread-local storage area
1106 Node* thread = new (C, 1) ThreadLocalNode();
1107 transform_later(thread);
1108
1109 call->init_req(TypeFunc::Parms+0, thread);
1110 call->init_req(TypeFunc::Parms+1, fast_oop);
1111 call->init_req( TypeFunc::Control, fast_oop_ctrl );
1112 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o
1113 call->init_req( TypeFunc::Memory , fast_oop_rawmem );
1114 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) );
1115 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) );
1116 transform_later(call);
1117 fast_oop_ctrl = new (C, 1) ProjNode(call,TypeFunc::Control);
1118 transform_later(fast_oop_ctrl);
1119 fast_oop_rawmem = new (C, 1) ProjNode(call,TypeFunc::Memory);
1120 transform_later(fast_oop_rawmem);
1121 }
1122
1123 // Plug in the successful fast-path into the result merge point
1124 result_region ->init_req( fast_result_path, fast_oop_ctrl );
1125 result_phi_rawoop->init_req( fast_result_path, fast_oop );
1126 result_phi_i_o ->init_req( fast_result_path, i_o );
1127 result_phi_rawmem->init_req( fast_result_path, fast_oop_rawmem );
1128 } else {
1129 slow_region = ctrl;
1130 }
1131
1132 // Generate slow-path call
1133 CallNode *call = new (C, slow_call_type->domain()->cnt())
1134 CallStaticJavaNode(slow_call_type, slow_call_address,
1135 OptoRuntime::stub_name(slow_call_address),
1136 alloc->jvms()->bci(),
1137 TypePtr::BOTTOM);
1138 call->init_req( TypeFunc::Control, slow_region );
1139 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o
1140 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs
1141 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) );
1142 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) );
1143
1144 call->init_req(TypeFunc::Parms+0, klass_node);
1145 if (length != NULL) {
1146 call->init_req(TypeFunc::Parms+1, length);
1147 }
1148
1149 // Copy debug information and adjust JVMState information, then replace
1150 // allocate node with the call
1151 copy_call_debug_info((CallNode *) alloc, call);
1152 if (!always_slow) {
1153 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1154 }
1155 _igvn.hash_delete(alloc);
1156 _igvn.subsume_node(alloc, call);
1157 transform_later(call);
1158
1159 // Identify the output projections from the allocate node and
1160 // adjust any references to them.
1161 // The control and io projections look like:
1162 //
1163 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1164 // Allocate Catch
1165 // ^---Proj(io) <-------+ ^---CatchProj(io)
1166 //
1167 // We are interested in the CatchProj nodes.
1168 //
1169 extract_call_projections(call);
1170
1171 // An allocate node has separate memory projections for the uses on the control and i_o paths
1172 // Replace uses of the control memory projection with result_phi_rawmem (unless we are only generating a slow call)
1173 if (!always_slow && _memproj_fallthrough != NULL) {
1174 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) {
1175 Node *use = _memproj_fallthrough->fast_out(i);
1176 _igvn.hash_delete(use);
1177 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem);
1178 _igvn._worklist.push(use);
1179 // back up iterator
1180 --i;
1181 }
1182 }
1183 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete _memproj_catchall so
1184 // we end up with a call that has only 1 memory projection
1185 if (_memproj_catchall != NULL ) {
1186 if (_memproj_fallthrough == NULL) {
1187 _memproj_fallthrough = new (C, 1) ProjNode(call, TypeFunc::Memory);
1188 transform_later(_memproj_fallthrough);
1189 }
1190 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) {
1191 Node *use = _memproj_catchall->fast_out(i);
1192 _igvn.hash_delete(use);
1193 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough);
1194 _igvn._worklist.push(use);
1195 // back up iterator
1196 --i;
1197 }
1198 }
1199
1200 mem = result_phi_rawmem;
1201
1202 // An allocate node has separate i_o projections for the uses on the control and i_o paths
1203 // Replace uses of the control i_o projection with result_phi_i_o (unless we are only generating a slow call)
1204 if (_ioproj_fallthrough == NULL) {
1205 _ioproj_fallthrough = new (C, 1) ProjNode(call, TypeFunc::I_O);
1206 transform_later(_ioproj_fallthrough);
1207 } else if (!always_slow) {
1208 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) {
1209 Node *use = _ioproj_fallthrough->fast_out(i);
1210
1211 _igvn.hash_delete(use);
1212 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o);
1213 _igvn._worklist.push(use);
1214 // back up iterator
1215 --i;
1216 }
1217 }
1218 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete _ioproj_catchall so
1219 // we end up with a call that has only 1 control projection
1220 if (_ioproj_catchall != NULL ) {
1221 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) {
1222 Node *use = _ioproj_catchall->fast_out(i);
1223 _igvn.hash_delete(use);
1224 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough);
1225 _igvn._worklist.push(use);
1226 // back up iterator
1227 --i;
1228 }
1229 }
1230
1231 // if we generated only a slow call, we are done
1232 if (always_slow)
1233 return;
1234
1235
1236 if (_fallthroughcatchproj != NULL) {
1237 ctrl = _fallthroughcatchproj->clone();
1238 transform_later(ctrl);
1239 _igvn.hash_delete(_fallthroughcatchproj);
1240 _igvn.subsume_node(_fallthroughcatchproj, result_region);
1241 } else {
1242 ctrl = top();
1243 }
1244 Node *slow_result;
1245 if (_resproj == NULL) {
1246 // no uses of the allocation result
1247 slow_result = top();
1248 } else {
1249 slow_result = _resproj->clone();
1250 transform_later(slow_result);
1251 _igvn.hash_delete(_resproj);
1252 _igvn.subsume_node(_resproj, result_phi_rawoop);
1253 }
1254
1255 // Plug slow-path into result merge point
1256 result_region ->init_req( slow_result_path, ctrl );
1257 result_phi_rawoop->init_req( slow_result_path, slow_result);
1258 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough );
1259 transform_later(result_region);
1260 transform_later(result_phi_rawoop);
1261 transform_later(result_phi_rawmem);
1262 transform_later(result_phi_i_o);
1263 // This completes all paths into the result merge point
1264 }
1265
1266
1267 // Helper for PhaseMacroExpand::expand_allocate_common.
1268 // Initializes the newly-allocated storage.
1269 Node*
1270 PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1271 Node* control, Node* rawmem, Node* object,
1272 Node* klass_node, Node* length,
1273 Node* size_in_bytes) {
1274 InitializeNode* init = alloc->initialization();
1275 // Store the klass & mark bits
1276 Node* mark_node = NULL;
1277 // For now only enable fast locking for non-array types
1278 if (UseBiasedLocking && (length == NULL)) {
1279 mark_node = make_load(NULL, rawmem, klass_node, Klass::prototype_header_offset_in_bytes() + sizeof(oopDesc), TypeRawPtr::BOTTOM, T_ADDRESS);
1280 } else {
1281 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype()));
1282 }
1283 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
1284
1285 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_OBJECT);
1286 int header_size = alloc->minimum_header_size(); // conservatively small
1287
1288 // Array length
1289 if (length != NULL) { // Arrays need length field
1290 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1291 // conservatively small header size:
1292 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1293 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1294 if (k->is_array_klass()) // we know the exact header size in most cases:
1295 header_size = Klass::layout_helper_header_size(k->layout_helper());
1296 }
1297
1298 // Clear the object body, if necessary.
1299 if (init == NULL) {
1300 // The init has somehow disappeared; be cautious and clear everything.
1301 //
1302 // This can happen if a node is allocated but an uncommon trap occurs
1303 // immediately. In this case, the Initialize gets associated with the
1304 // trap, and may be placed in a different (outer) loop, if the Allocate
1305 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1306 // there can be two Allocates to one Initialize. The answer in all these
1307 // edge cases is safety first. It is always safe to clear immediately
1308 // within an Allocate, and then (maybe or maybe not) clear some more later.
1309 if (!ZeroTLAB)
1310 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1311 header_size, size_in_bytes,
1312 &_igvn);
1313 } else {
1314 if (!init->is_complete()) {
1315 // Try to win by zeroing only what the init does not store.
1316 // We can also try to do some peephole optimizations,
1317 // such as combining some adjacent subword stores.
1318 rawmem = init->complete_stores(control, rawmem, object,
1319 header_size, size_in_bytes, &_igvn);
1320 }
1321 // We have no more use for this link, since the AllocateNode goes away:
1322 init->set_req(InitializeNode::RawAddress, top());
1323 // (If we keep the link, it just confuses the register allocator,
1324 // who thinks he sees a real use of the address by the membar.)
1325 }
1326
1327 return rawmem;
1328 }
1329
1330 // Generate prefetch instructions for next allocations.
1331 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
1332 Node*& contended_phi_rawmem,
1333 Node* old_eden_top, Node* new_eden_top,
1334 Node* length) {
1335 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
1336 // Generate prefetch allocation with watermark check.
1337 // As an allocation hits the watermark, we will prefetch starting
1338 // at a "distance" away from watermark.
1339 enum { fall_in_path = 1, pf_path = 2 };
1340
1341 Node *pf_region = new (C, 3) RegionNode(3);
1342 Node *pf_phi_rawmem = new (C, 3) PhiNode( pf_region, Type::MEMORY,
1343 TypeRawPtr::BOTTOM );
1344 // I/O is used for Prefetch
1345 Node *pf_phi_abio = new (C, 3) PhiNode( pf_region, Type::ABIO );
1346
1347 Node *thread = new (C, 1) ThreadLocalNode();
1348 transform_later(thread);
1349
1350 Node *eden_pf_adr = new (C, 4) AddPNode( top()/*not oop*/, thread,
1351 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
1352 transform_later(eden_pf_adr);
1353
1354 Node *old_pf_wm = new (C, 3) LoadPNode( needgc_false,
1355 contended_phi_rawmem, eden_pf_adr,
1356 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM );
1357 transform_later(old_pf_wm);
1358
1359 // check against new_eden_top
1360 Node *need_pf_cmp = new (C, 3) CmpPNode( new_eden_top, old_pf_wm );
1361 transform_later(need_pf_cmp);
1362 Node *need_pf_bol = new (C, 2) BoolNode( need_pf_cmp, BoolTest::ge );
1363 transform_later(need_pf_bol);
1364 IfNode *need_pf_iff = new (C, 2) IfNode( needgc_false, need_pf_bol,
1365 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1366 transform_later(need_pf_iff);
1367
1368 // true node, add prefetchdistance
1369 Node *need_pf_true = new (C, 1) IfTrueNode( need_pf_iff );
1370 transform_later(need_pf_true);
1371
1372 Node *need_pf_false = new (C, 1) IfFalseNode( need_pf_iff );
1373 transform_later(need_pf_false);
1374
1375 Node *new_pf_wmt = new (C, 4) AddPNode( top(), old_pf_wm,
1376 _igvn.MakeConX(AllocatePrefetchDistance) );
1377 transform_later(new_pf_wmt );
1378 new_pf_wmt->set_req(0, need_pf_true);
1379
1380 Node *store_new_wmt = new (C, 4) StorePNode( need_pf_true,
1381 contended_phi_rawmem, eden_pf_adr,
1382 TypeRawPtr::BOTTOM, new_pf_wmt );
1383 transform_later(store_new_wmt);
1384
1385 // adding prefetches
1386 pf_phi_abio->init_req( fall_in_path, i_o );
1387
1388 Node *prefetch_adr;
1389 Node *prefetch;
1390 uint lines = AllocatePrefetchDistance / AllocatePrefetchStepSize;
1391 uint step_size = AllocatePrefetchStepSize;
1392 uint distance = 0;
1393
1394 for ( uint i = 0; i < lines; i++ ) {
1395 prefetch_adr = new (C, 4) AddPNode( old_pf_wm, new_pf_wmt,
1396 _igvn.MakeConX(distance) );
1397 transform_later(prefetch_adr);
1398 prefetch = new (C, 3) PrefetchWriteNode( i_o, prefetch_adr );
1399 transform_later(prefetch);
1400 distance += step_size;
1401 i_o = prefetch;
1402 }
1403 pf_phi_abio->set_req( pf_path, i_o );
1404
1405 pf_region->init_req( fall_in_path, need_pf_false );
1406 pf_region->init_req( pf_path, need_pf_true );
1407
1408 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
1409 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
1410
1411 transform_later(pf_region);
1412 transform_later(pf_phi_rawmem);
1413 transform_later(pf_phi_abio);
1414
1415 needgc_false = pf_region;
1416 contended_phi_rawmem = pf_phi_rawmem;
1417 i_o = pf_phi_abio;
1418 } else if( AllocatePrefetchStyle > 0 ) {
1419 // Insert a prefetch for each allocation only on the fast-path
1420 Node *prefetch_adr;
1421 Node *prefetch;
1422 // Generate several prefetch instructions only for arrays.
1423 uint lines = (length != NULL) ? AllocatePrefetchLines : 1;
1424 uint step_size = AllocatePrefetchStepSize;
1425 uint distance = AllocatePrefetchDistance;
1426 for ( uint i = 0; i < lines; i++ ) {
1427 prefetch_adr = new (C, 4) AddPNode( old_eden_top, new_eden_top,
1428 _igvn.MakeConX(distance) );
1429 transform_later(prefetch_adr);
1430 prefetch = new (C, 3) PrefetchWriteNode( i_o, prefetch_adr );
1431 // Do not let it float too high, since if eden_top == eden_end,
1432 // both might be null.
1433 if( i == 0 ) { // Set control for first prefetch, next follows it
1434 prefetch->init_req(0, needgc_false);
1435 }
1436 transform_later(prefetch);
1437 distance += step_size;
1438 i_o = prefetch;
1439 }
1440 }
1441 return i_o;
1442 }
1443
1444
1445 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
1446 expand_allocate_common(alloc, NULL,
1447 OptoRuntime::new_instance_Type(),
1448 OptoRuntime::new_instance_Java());
1449 }
1450
1451 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
1452 Node* length = alloc->in(AllocateNode::ALength);
1453 expand_allocate_common(alloc, length,
1454 OptoRuntime::new_array_Type(),
1455 OptoRuntime::new_array_Java());
1456 }
1457
1458
1459 // we have determined that this lock/unlock can be eliminated, we simply
1460 // eliminate the node without expanding it.
1461 //
1462 // Note: The membar's associated with the lock/unlock are currently not
1463 // eliminated. This should be investigated as a future enhancement.
1464 //
1465 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
1466
1467 if (!alock->is_eliminated()) {
1468 return false;
1469 }
1470 // Mark the box lock as eliminated if all correspondent locks are eliminated
1471 // to construct correct debug info.
1472 BoxLockNode* box = alock->box_node()->as_BoxLock();
1473 if (!box->is_eliminated()) {
1474 bool eliminate = true;
1475 for (DUIterator_Fast imax, i = box->fast_outs(imax); i < imax; i++) {
1476 Node *lck = box->fast_out(i);
1477 if (lck->is_Lock() && !lck->as_AbstractLock()->is_eliminated()) {
1478 eliminate = false;
1479 break;
1480 }
1481 }
1482 if (eliminate)
1483 box->set_eliminated();
1484 }
1485
1486 #ifndef PRODUCT
1487 if (PrintEliminateLocks) {
1488 if (alock->is_Lock()) {
1489 tty->print_cr("++++ Eliminating: %d Lock", alock->_idx);
1490 } else {
1491 tty->print_cr("++++ Eliminating: %d Unlock", alock->_idx);
1492 }
1493 }
1494 #endif
1495
1496 Node* mem = alock->in(TypeFunc::Memory);
1497 Node* ctrl = alock->in(TypeFunc::Control);
1498
1499 extract_call_projections(alock);
1500 // There are 2 projections from the lock. The lock node will
1501 // be deleted when its last use is subsumed below.
1502 assert(alock->outcnt() == 2 &&
1503 _fallthroughproj != NULL &&
1504 _memproj_fallthrough != NULL,
1505 "Unexpected projections from Lock/Unlock");
1506
1507 Node* fallthroughproj = _fallthroughproj;
1508 Node* memproj_fallthrough = _memproj_fallthrough;
1509
1510 // The memory projection from a lock/unlock is RawMem
1511 // The input to a Lock is merged memory, so extract its RawMem input
1512 // (unless the MergeMem has been optimized away.)
1513 if (alock->is_Lock()) {
1514 // Seach for MemBarAcquire node and delete it also.
1515 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
1516 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquire, "");
1517 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
1518 Node* memproj = membar->proj_out(TypeFunc::Memory);
1519 _igvn.hash_delete(ctrlproj);
1520 _igvn.subsume_node(ctrlproj, fallthroughproj);
1521 _igvn.hash_delete(memproj);
1522 _igvn.subsume_node(memproj, memproj_fallthrough);
1523 }
1524
1525 // Seach for MemBarRelease node and delete it also.
1526 if (alock->is_Unlock() && ctrl != NULL && ctrl->is_Proj() &&
1527 ctrl->in(0)->is_MemBar()) {
1528 MemBarNode* membar = ctrl->in(0)->as_MemBar();
1529 assert(membar->Opcode() == Op_MemBarRelease &&
1530 mem->is_Proj() && membar == mem->in(0), "");
1531 _igvn.hash_delete(fallthroughproj);
1532 _igvn.subsume_node(fallthroughproj, ctrl);
1533 _igvn.hash_delete(memproj_fallthrough);
1534 _igvn.subsume_node(memproj_fallthrough, mem);
1535 fallthroughproj = ctrl;
1536 memproj_fallthrough = mem;
1537 ctrl = membar->in(TypeFunc::Control);
1538 mem = membar->in(TypeFunc::Memory);
1539 }
1540
1541 _igvn.hash_delete(fallthroughproj);
1542 _igvn.subsume_node(fallthroughproj, ctrl);
1543 _igvn.hash_delete(memproj_fallthrough);
1544 _igvn.subsume_node(memproj_fallthrough, mem);
1545 return true;
1546 }
1547
1548
1549 //------------------------------expand_lock_node----------------------
1550 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
1551
1552 Node* ctrl = lock->in(TypeFunc::Control);
1553 Node* mem = lock->in(TypeFunc::Memory);
1554 Node* obj = lock->obj_node();
1555 Node* box = lock->box_node();
1556 Node* flock = lock->fastlock_node();
1557
1558 // Make the merge point
1559 Node *region = new (C, 3) RegionNode(3);
1560
1561 Node *bol = transform_later(new (C, 2) BoolNode(flock,BoolTest::ne));
1562 Node *iff = new (C, 2) IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
1563 // Optimize test; set region slot 2
1564 Node *slow_path = opt_iff(region,iff);
1565
1566 // Make slow path call
1567 CallNode *call = make_slow_call( (CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(), OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path, obj, box );
1568
1569 extract_call_projections(call);
1570
1571 // Slow path can only throw asynchronous exceptions, which are always
1572 // de-opted. So the compiler thinks the slow-call can never throw an
1573 // exception. If it DOES throw an exception we would need the debug
1574 // info removed first (since if it throws there is no monitor).
1575 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
1576 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
1577
1578 // Capture slow path
1579 // disconnect fall-through projection from call and create a new one
1580 // hook up users of fall-through projection to region
1581 Node *slow_ctrl = _fallthroughproj->clone();
1582 transform_later(slow_ctrl);
1583 _igvn.hash_delete(_fallthroughproj);
1584 _fallthroughproj->disconnect_inputs(NULL);
1585 region->init_req(1, slow_ctrl);
1586 // region inputs are now complete
1587 transform_later(region);
1588 _igvn.subsume_node(_fallthroughproj, region);
1589
1590 // create a Phi for the memory state
1591 Node *mem_phi = new (C, 3) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
1592 Node *memproj = transform_later( new (C, 1) ProjNode(call, TypeFunc::Memory) );
1593 mem_phi->init_req(1, memproj );
1594 mem_phi->init_req(2, mem);
1595 transform_later(mem_phi);
1596 _igvn.hash_delete(_memproj_fallthrough);
1597 _igvn.subsume_node(_memproj_fallthrough, mem_phi);
1598
1599
1600 }
1601
1602 //------------------------------expand_unlock_node----------------------
1603 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
1604
1605 Node* ctrl = unlock->in(TypeFunc::Control);
1606 Node* mem = unlock->in(TypeFunc::Memory);
1607 Node* obj = unlock->obj_node();
1608 Node* box = unlock->box_node();
1609
1610 // No need for a null check on unlock
1611
1612 // Make the merge point
1613 RegionNode *region = new (C, 3) RegionNode(3);
1614
1615 FastUnlockNode *funlock = new (C, 3) FastUnlockNode( ctrl, obj, box );
1616 funlock = transform_later( funlock )->as_FastUnlock();
1617 Node *bol = transform_later(new (C, 2) BoolNode(funlock,BoolTest::ne));
1618 Node *iff = new (C, 2) IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
1619 // Optimize test; set region slot 2
1620 Node *slow_path = opt_iff(region,iff);
1621
1622 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 );
1623
1624 extract_call_projections(call);
1625
1626 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
1627 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
1628
1629 // No exceptions for unlocking
1630 // Capture slow path
1631 // disconnect fall-through projection from call and create a new one
1632 // hook up users of fall-through projection to region
1633 Node *slow_ctrl = _fallthroughproj->clone();
1634 transform_later(slow_ctrl);
1635 _igvn.hash_delete(_fallthroughproj);
1636 _fallthroughproj->disconnect_inputs(NULL);
1637 region->init_req(1, slow_ctrl);
1638 // region inputs are now complete
1639 transform_later(region);
1640 _igvn.subsume_node(_fallthroughproj, region);
1641
1642 // create a Phi for the memory state
1643 Node *mem_phi = new (C, 3) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
1644 Node *memproj = transform_later( new(C, 1) ProjNode(call, TypeFunc::Memory) );
1645 mem_phi->init_req(1, memproj );
1646 mem_phi->init_req(2, mem);
1647 transform_later(mem_phi);
1648 _igvn.hash_delete(_memproj_fallthrough);
1649 _igvn.subsume_node(_memproj_fallthrough, mem_phi);
1650
1651
1652 }
1653
1654 //------------------------------expand_macro_nodes----------------------
1655 // Returns true if a failure occurred.
1656 bool PhaseMacroExpand::expand_macro_nodes() {
1657 if (C->macro_count() == 0)
1658 return false;
1659 // attempt to eliminate allocations
1660 bool progress = true;
1661 while (progress) {
1662 progress = false;
1663 for (int i = C->macro_count(); i > 0; i--) {
1664 Node * n = C->macro_node(i-1);
1665 bool success = false;
1666 debug_only(int old_macro_count = C->macro_count(););
1667 switch (n->class_id()) {
1668 case Node::Class_Allocate:
1669 case Node::Class_AllocateArray:
1670 success = eliminate_allocate_node(n->as_Allocate());
1671 break;
1672 case Node::Class_Lock:
1673 case Node::Class_Unlock:
1674 success = eliminate_locking_node(n->as_AbstractLock());
1675 break;
1676 default:
1677 if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) {
1678 _igvn.add_users_to_worklist(n);
1679 _igvn.hash_delete(n);
1680 _igvn.subsume_node(n, n->in(1));
1681 success = true;
1682 } else {
1683 assert(false, "unknown node type in macro list");
1684 }
1685 }
1686 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
1687 progress = progress || success;
1688 }
1689 }
1690 // Make sure expansion will not cause node limit to be exceeded.
1691 // Worst case is a macro node gets expanded into about 50 nodes.
1692 // Allow 50% more for optimization.
1693 if (C->check_node_count(C->macro_count() * 75, "out of nodes before macro expansion" ) )
1694 return true;
1695
1696 // expand "macro" nodes
1697 // nodes are removed from the macro list as they are processed
1698 while (C->macro_count() > 0) {
1699 int macro_count = C->macro_count();
1700 Node * n = C->macro_node(macro_count-1);
1701 assert(n->is_macro(), "only macro nodes expected here");
1702 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) {
1703 // node is unreachable, so don't try to expand it
1704 C->remove_macro_node(n);
1705 continue;
1706 }
1707 switch (n->class_id()) {
1708 case Node::Class_Allocate:
1709 expand_allocate(n->as_Allocate());
1710 break;
1711 case Node::Class_AllocateArray:
1712 expand_allocate_array(n->as_AllocateArray());
1713 break;
1714 case Node::Class_Lock:
1715 expand_lock_node(n->as_Lock());
1716 break;
1717 case Node::Class_Unlock:
1718 expand_unlock_node(n->as_Unlock());
1719 break;
1720 default:
1721 assert(false, "unknown node type in macro list");
1722 }
1723 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
1724 if (C->failing()) return true;
1725 }
1726
1727 _igvn.set_delay_transform(false);
1728 _igvn.optimize();
1729 return false;
1730 }