1 /*
2 * Copyright 2005-2008 Sun Microsystems, Inc. All Rights Reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 * have any questions.
22 *
23 */
24
25 #include "incls/_precompiled.incl"
26 #include "incls/_macro.cpp.incl"
27
28
29 //
30 // Replace any references to "oldref" in inputs to "use" with "newref".
31 // Returns the number of replacements made.
32 //
33 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
34 int nreplacements = 0;
35 uint req = use->req();
36 for (uint j = 0; j < use->len(); j++) {
37 Node *uin = use->in(j);
38 if (uin == oldref) {
39 if (j < req)
40 use->set_req(j, newref);
41 else
42 use->set_prec(j, newref);
43 nreplacements++;
44 } else if (j >= req && uin == NULL) {
45 break;
46 }
47 }
48 return nreplacements;
49 }
50
51 void PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) {
52 // Copy debug information and adjust JVMState information
53 uint old_dbg_start = oldcall->tf()->domain()->cnt();
54 uint new_dbg_start = newcall->tf()->domain()->cnt();
55 int jvms_adj = new_dbg_start - old_dbg_start;
56 assert (new_dbg_start == newcall->req(), "argument count mismatch");
57
58 Dict* sosn_map = new Dict(cmpkey,hashkey);
59 for (uint i = old_dbg_start; i < oldcall->req(); i++) {
60 Node* old_in = oldcall->in(i);
61 // Clone old SafePointScalarObjectNodes, adjusting their field contents.
62 if (old_in != NULL && old_in->is_SafePointScalarObject()) {
63 SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject();
64 uint old_unique = C->unique();
65 Node* new_in = old_sosn->clone(jvms_adj, sosn_map);
66 if (old_unique != C->unique()) {
67 new_in = transform_later(new_in); // Register new node.
68 }
69 old_in = new_in;
70 }
71 newcall->add_req(old_in);
72 }
73
74 newcall->set_jvms(oldcall->jvms());
75 for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) {
76 jvms->set_map(newcall);
77 jvms->set_locoff(jvms->locoff()+jvms_adj);
78 jvms->set_stkoff(jvms->stkoff()+jvms_adj);
79 jvms->set_monoff(jvms->monoff()+jvms_adj);
80 jvms->set_scloff(jvms->scloff()+jvms_adj);
81 jvms->set_endoff(jvms->endoff()+jvms_adj);
82 }
83 }
84
85 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
86 Node* cmp;
87 if (mask != 0) {
88 Node* and_node = transform_later(new (C, 3) AndXNode(word, MakeConX(mask)));
89 cmp = transform_later(new (C, 3) CmpXNode(and_node, MakeConX(bits)));
90 } else {
91 cmp = word;
92 }
93 Node* bol = transform_later(new (C, 2) BoolNode(cmp, BoolTest::ne));
94 IfNode* iff = new (C, 2) IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
95 transform_later(iff);
96
97 // Fast path taken.
98 Node *fast_taken = transform_later( new (C, 1) IfFalseNode(iff) );
99
100 // Fast path not-taken, i.e. slow path
101 Node *slow_taken = transform_later( new (C, 1) IfTrueNode(iff) );
102
103 if (return_fast_path) {
104 region->init_req(edge, slow_taken); // Capture slow-control
105 return fast_taken;
106 } else {
107 region->init_req(edge, fast_taken); // Capture fast-control
108 return slow_taken;
109 }
110 }
111
112 //--------------------copy_predefined_input_for_runtime_call--------------------
113 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
114 // Set fixed predefined input arguments
115 call->init_req( TypeFunc::Control, ctrl );
116 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) );
117 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
118 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
119 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
120 }
121
122 //------------------------------make_slow_call---------------------------------
123 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type, address slow_call, const char* leaf_name, Node* slow_path, Node* parm0, Node* parm1) {
124
125 // Slow-path call
126 int size = slow_call_type->domain()->cnt();
127 CallNode *call = leaf_name
128 ? (CallNode*)new (C, size) CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
129 : (CallNode*)new (C, size) CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM );
130
131 // Slow path call has no side-effects, uses few values
132 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
133 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0);
134 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1);
135 copy_call_debug_info(oldcall, call);
136 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
137 _igvn.hash_delete(oldcall);
138 _igvn.subsume_node(oldcall, call);
139 transform_later(call);
140
141 return call;
142 }
143
144 void PhaseMacroExpand::extract_call_projections(CallNode *call) {
145 _fallthroughproj = NULL;
146 _fallthroughcatchproj = NULL;
147 _ioproj_fallthrough = NULL;
148 _ioproj_catchall = NULL;
149 _catchallcatchproj = NULL;
150 _memproj_fallthrough = NULL;
151 _memproj_catchall = NULL;
152 _resproj = NULL;
153 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) {
154 ProjNode *pn = call->fast_out(i)->as_Proj();
155 switch (pn->_con) {
156 case TypeFunc::Control:
157 {
158 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj
159 _fallthroughproj = pn;
160 DUIterator_Fast jmax, j = pn->fast_outs(jmax);
161 const Node *cn = pn->fast_out(j);
162 if (cn->is_Catch()) {
163 ProjNode *cpn = NULL;
164 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) {
165 cpn = cn->fast_out(k)->as_Proj();
166 assert(cpn->is_CatchProj(), "must be a CatchProjNode");
167 if (cpn->_con == CatchProjNode::fall_through_index)
168 _fallthroughcatchproj = cpn;
169 else {
170 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index.");
171 _catchallcatchproj = cpn;
172 }
173 }
174 }
175 break;
176 }
177 case TypeFunc::I_O:
178 if (pn->_is_io_use)
179 _ioproj_catchall = pn;
180 else
181 _ioproj_fallthrough = pn;
182 break;
183 case TypeFunc::Memory:
184 if (pn->_is_io_use)
185 _memproj_catchall = pn;
186 else
187 _memproj_fallthrough = pn;
188 break;
189 case TypeFunc::Parms:
190 _resproj = pn;
191 break;
192 default:
193 assert(false, "unexpected projection from allocation node.");
194 }
195 }
196
197 }
198
199 // Eliminate a card mark sequence. p2x is a ConvP2XNode
200 void PhaseMacroExpand::eliminate_card_mark(Node *p2x) {
201 assert(p2x->Opcode() == Op_CastP2X, "ConvP2XNode required");
202 Node *shift = p2x->unique_out();
203 Node *addp = shift->unique_out();
204 for (DUIterator_Last jmin, j = addp->last_outs(jmin); j >= jmin; --j) {
205 Node *st = addp->last_out(j);
206 assert(st->is_Store(), "store required");
207 _igvn.replace_node(st, st->in(MemNode::Memory));
208 }
209 }
210
211 // Search for a memory operation for the specified memory slice.
212 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
213 Node *orig_mem = mem;
214 Node *alloc_mem = alloc->in(TypeFunc::Memory);
215 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
216 while (true) {
217 if (mem == alloc_mem || mem == start_mem ) {
218 return mem; // hit one of our sentinals
219 } else if (mem->is_MergeMem()) {
220 mem = mem->as_MergeMem()->memory_at(alias_idx);
221 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
222 Node *in = mem->in(0);
223 // we can safely skip over safepoints, calls, locks and membars because we
224 // already know that the object is safe to eliminate.
225 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
226 return in;
227 } else if (in->is_Call()) {
228 CallNode *call = in->as_Call();
229 if (!call->may_modify(tinst, phase)) {
230 mem = call->in(TypeFunc::Memory);
231 }
232 mem = in->in(TypeFunc::Memory);
233 } else if (in->is_MemBar()) {
234 mem = in->in(TypeFunc::Memory);
235 } else {
236 assert(false, "unexpected projection");
237 }
238 } else if (mem->is_Store()) {
239 const TypePtr* atype = mem->as_Store()->adr_type();
240 int adr_idx = Compile::current()->get_alias_index(atype);
241 if (adr_idx == alias_idx) {
242 assert(atype->isa_oopptr(), "address type must be oopptr");
243 int adr_offset = atype->offset();
244 uint adr_iid = atype->is_oopptr()->instance_id();
245 // Array elements references have the same alias_idx
246 // but different offset and different instance_id.
247 if (adr_offset == offset && adr_iid == alloc->_idx)
248 return mem;
249 } else {
250 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
251 }
252 mem = mem->in(MemNode::Memory);
253 } else if (mem->Opcode() == Op_SCMemProj) {
254 assert(mem->in(0)->is_LoadStore(), "sanity");
255 const TypePtr* atype = mem->in(0)->in(MemNode::Address)->bottom_type()->is_ptr();
256 int adr_idx = Compile::current()->get_alias_index(atype);
257 if (adr_idx == alias_idx) {
258 assert(false, "Object is not scalar replaceable if a LoadStore node access its field");
259 return NULL;
260 }
261 mem = mem->in(0)->in(MemNode::Memory);
262 } else {
263 return mem;
264 }
265 assert(mem != orig_mem, "dead memory loop");
266 }
267 }
268
269 //
270 // Given a Memory Phi, compute a value Phi containing the values from stores
271 // on the input paths.
272 // Note: this function is recursive, its depth is limied by the "level" argument
273 // Returns the computed Phi, or NULL if it cannot compute it.
274 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) {
275 assert(mem->is_Phi(), "sanity");
276 int alias_idx = C->get_alias_index(adr_t);
277 int offset = adr_t->offset();
278 int instance_id = adr_t->instance_id();
279
280 // Check if an appropriate value phi already exists.
281 Node* region = mem->in(0);
282 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
283 Node* phi = region->fast_out(k);
284 if (phi->is_Phi() && phi != mem &&
285 phi->as_Phi()->is_same_inst_field(phi_type, instance_id, alias_idx, offset)) {
286 return phi;
287 }
288 }
289 // Check if an appropriate new value phi already exists.
290 Node* new_phi = NULL;
291 uint size = value_phis->size();
292 for (uint i=0; i < size; i++) {
293 if ( mem->_idx == value_phis->index_at(i) ) {
294 return value_phis->node_at(i);
295 }
296 }
297
298 if (level <= 0) {
299 return NULL; // Give up: phi tree too deep
300 }
301 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
302 Node *alloc_mem = alloc->in(TypeFunc::Memory);
303
304 uint length = mem->req();
305 GrowableArray <Node *> values(length, length, NULL);
306
307 // create a new Phi for the value
308 PhiNode *phi = new (C, length) PhiNode(mem->in(0), phi_type, NULL, instance_id, alias_idx, offset);
309 transform_later(phi);
310 value_phis->push(phi, mem->_idx);
311
312 for (uint j = 1; j < length; j++) {
313 Node *in = mem->in(j);
314 if (in == NULL || in->is_top()) {
315 values.at_put(j, in);
316 } else {
317 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
318 if (val == start_mem || val == alloc_mem) {
319 // hit a sentinel, return appropriate 0 value
320 values.at_put(j, _igvn.zerocon(ft));
321 continue;
322 }
323 if (val->is_Initialize()) {
324 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
325 }
326 if (val == NULL) {
327 return NULL; // can't find a value on this path
328 }
329 if (val == mem) {
330 values.at_put(j, mem);
331 } else if (val->is_Store()) {
332 values.at_put(j, val->in(MemNode::ValueIn));
333 } else if(val->is_Proj() && val->in(0) == alloc) {
334 values.at_put(j, _igvn.zerocon(ft));
335 } else if (val->is_Phi()) {
336 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
337 if (val == NULL) {
338 return NULL;
339 }
340 values.at_put(j, val);
341 } else if (val->Opcode() == Op_SCMemProj) {
342 assert(val->in(0)->is_LoadStore(), "sanity");
343 assert(false, "Object is not scalar replaceable if a LoadStore node access its field");
344 return NULL;
345 } else {
346 #ifdef ASSERT
347 val->dump();
348 assert(false, "unknown node on this path");
349 #endif
350 return NULL; // unknown node on this path
351 }
352 }
353 }
354 // Set Phi's inputs
355 for (uint j = 1; j < length; j++) {
356 if (values.at(j) == mem) {
357 phi->init_req(j, phi);
358 } else {
359 phi->init_req(j, values.at(j));
360 }
361 }
362 return phi;
363 }
364
365 // Search the last value stored into the object's field.
366 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, Node *alloc) {
367 assert(adr_t->is_known_instance_field(), "instance required");
368 int instance_id = adr_t->instance_id();
369 assert((uint)instance_id == alloc->_idx, "wrong allocation");
370
371 int alias_idx = C->get_alias_index(adr_t);
372 int offset = adr_t->offset();
373 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
374 Node *alloc_ctrl = alloc->in(TypeFunc::Control);
375 Node *alloc_mem = alloc->in(TypeFunc::Memory);
376 Arena *a = Thread::current()->resource_area();
377 VectorSet visited(a);
378
379
380 bool done = sfpt_mem == alloc_mem;
381 Node *mem = sfpt_mem;
382 while (!done) {
383 if (visited.test_set(mem->_idx)) {
384 return NULL; // found a loop, give up
385 }
386 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
387 if (mem == start_mem || mem == alloc_mem) {
388 done = true; // hit a sentinel, return appropriate 0 value
389 } else if (mem->is_Initialize()) {
390 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
391 if (mem == NULL) {
392 done = true; // Something go wrong.
393 } else if (mem->is_Store()) {
394 const TypePtr* atype = mem->as_Store()->adr_type();
395 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
396 done = true;
397 }
398 } else if (mem->is_Store()) {
399 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
400 assert(atype != NULL, "address type must be oopptr");
401 assert(C->get_alias_index(atype) == alias_idx &&
402 atype->is_known_instance_field() && atype->offset() == offset &&
403 atype->instance_id() == instance_id, "store is correct memory slice");
404 done = true;
405 } else if (mem->is_Phi()) {
406 // try to find a phi's unique input
407 Node *unique_input = NULL;
408 Node *top = C->top();
409 for (uint i = 1; i < mem->req(); i++) {
410 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
411 if (n == NULL || n == top || n == mem) {
412 continue;
413 } else if (unique_input == NULL) {
414 unique_input = n;
415 } else if (unique_input != n) {
416 unique_input = top;
417 break;
418 }
419 }
420 if (unique_input != NULL && unique_input != top) {
421 mem = unique_input;
422 } else {
423 done = true;
424 }
425 } else {
426 assert(false, "unexpected node");
427 }
428 }
429 if (mem != NULL) {
430 if (mem == start_mem || mem == alloc_mem) {
431 // hit a sentinel, return appropriate 0 value
432 return _igvn.zerocon(ft);
433 } else if (mem->is_Store()) {
434 return mem->in(MemNode::ValueIn);
435 } else if (mem->is_Phi()) {
436 // attempt to produce a Phi reflecting the values on the input paths of the Phi
437 Node_Stack value_phis(a, 8);
438 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
439 if (phi != NULL) {
440 return phi;
441 } else {
442 // Kill all new Phis
443 while(value_phis.is_nonempty()) {
444 Node* n = value_phis.node();
445 _igvn.hash_delete(n);
446 _igvn.subsume_node(n, C->top());
447 value_phis.pop();
448 }
449 }
450 }
451 }
452 // Something go wrong.
453 return NULL;
454 }
455
456 // Check the possibility of scalar replacement.
457 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
458 // Scan the uses of the allocation to check for anything that would
459 // prevent us from eliminating it.
460 NOT_PRODUCT( const char* fail_eliminate = NULL; )
461 DEBUG_ONLY( Node* disq_node = NULL; )
462 bool can_eliminate = true;
463
464 Node* res = alloc->result_cast();
465 const TypeOopPtr* res_type = NULL;
466 if (res == NULL) {
467 // All users were eliminated.
468 } else if (!res->is_CheckCastPP()) {
469 alloc->_is_scalar_replaceable = false; // don't try again
470 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
471 can_eliminate = false;
472 } else {
473 res_type = _igvn.type(res)->isa_oopptr();
474 if (res_type == NULL) {
475 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
476 can_eliminate = false;
477 } else if (res_type->isa_aryptr()) {
478 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
479 if (length < 0) {
480 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
481 can_eliminate = false;
482 }
483 }
484 }
485
486 if (can_eliminate && res != NULL) {
487 for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
488 j < jmax && can_eliminate; j++) {
489 Node* use = res->fast_out(j);
490
491 if (use->is_AddP()) {
492 const TypePtr* addp_type = _igvn.type(use)->is_ptr();
493 int offset = addp_type->offset();
494
495 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
496 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";)
497 can_eliminate = false;
498 break;
499 }
500 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
501 k < kmax && can_eliminate; k++) {
502 Node* n = use->fast_out(k);
503 if (!n->is_Store() && n->Opcode() != Op_CastP2X) {
504 DEBUG_ONLY(disq_node = n;)
505 if (n->is_Load() || n->is_LoadStore()) {
506 NOT_PRODUCT(fail_eliminate = "Field load";)
507 } else {
508 NOT_PRODUCT(fail_eliminate = "Not store field referrence";)
509 }
510 can_eliminate = false;
511 }
512 }
513 } else if (use->is_SafePoint()) {
514 SafePointNode* sfpt = use->as_SafePoint();
515 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
516 // Object is passed as argument.
517 DEBUG_ONLY(disq_node = use;)
518 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
519 can_eliminate = false;
520 }
521 Node* sfptMem = sfpt->memory();
522 if (sfptMem == NULL || sfptMem->is_top()) {
523 DEBUG_ONLY(disq_node = use;)
524 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";)
525 can_eliminate = false;
526 } else {
527 safepoints.append_if_missing(sfpt);
528 }
529 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
530 if (use->is_Phi()) {
531 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
532 NOT_PRODUCT(fail_eliminate = "Object is return value";)
533 } else {
534 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
535 }
536 DEBUG_ONLY(disq_node = use;)
537 } else {
538 if (use->Opcode() == Op_Return) {
539 NOT_PRODUCT(fail_eliminate = "Object is return value";)
540 }else {
541 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
542 }
543 DEBUG_ONLY(disq_node = use;)
544 }
545 can_eliminate = false;
546 }
547 }
548 }
549
550 #ifndef PRODUCT
551 if (PrintEliminateAllocations) {
552 if (can_eliminate) {
553 tty->print("Scalar ");
554 if (res == NULL)
555 alloc->dump();
556 else
557 res->dump();
558 } else {
559 tty->print("NotScalar (%s)", fail_eliminate);
560 if (res == NULL)
561 alloc->dump();
562 else
563 res->dump();
564 #ifdef ASSERT
565 if (disq_node != NULL) {
566 tty->print(" >>>> ");
567 disq_node->dump();
568 }
569 #endif /*ASSERT*/
570 }
571 }
572 #endif
573 return can_eliminate;
574 }
575
576 // Do scalar replacement.
577 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
578 GrowableArray <SafePointNode *> safepoints_done;
579
580 ciKlass* klass = NULL;
581 ciInstanceKlass* iklass = NULL;
582 int nfields = 0;
583 int array_base;
584 int element_size;
585 BasicType basic_elem_type;
586 ciType* elem_type;
587
588 Node* res = alloc->result_cast();
589 const TypeOopPtr* res_type = NULL;
590 if (res != NULL) { // Could be NULL when there are no users
591 res_type = _igvn.type(res)->isa_oopptr();
592 }
593
594 if (res != NULL) {
595 klass = res_type->klass();
596 if (res_type->isa_instptr()) {
597 // find the fields of the class which will be needed for safepoint debug information
598 assert(klass->is_instance_klass(), "must be an instance klass.");
599 iklass = klass->as_instance_klass();
600 nfields = iklass->nof_nonstatic_fields();
601 } else {
602 // find the array's elements which will be needed for safepoint debug information
603 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
604 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass.");
605 elem_type = klass->as_array_klass()->element_type();
606 basic_elem_type = elem_type->basic_type();
607 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
608 element_size = type2aelembytes(basic_elem_type);
609 }
610 }
611 //
612 // Process the safepoint uses
613 //
614 while (safepoints.length() > 0) {
615 SafePointNode* sfpt = safepoints.pop();
616 Node* mem = sfpt->memory();
617 uint first_ind = sfpt->req();
618 SafePointScalarObjectNode* sobj = new (C, 1) SafePointScalarObjectNode(res_type,
619 #ifdef ASSERT
620 alloc,
621 #endif
622 first_ind, nfields);
623 sobj->init_req(0, sfpt->in(TypeFunc::Control));
624 transform_later(sobj);
625
626 // Scan object's fields adding an input to the safepoint for each field.
627 for (int j = 0; j < nfields; j++) {
628 intptr_t offset;
629 ciField* field = NULL;
630 if (iklass != NULL) {
631 field = iklass->nonstatic_field_at(j);
632 offset = field->offset();
633 elem_type = field->type();
634 basic_elem_type = field->layout_type();
635 } else {
636 offset = array_base + j * (intptr_t)element_size;
637 }
638
639 const Type *field_type;
640 // The next code is taken from Parse::do_get_xxx().
641 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) {
642 if (!elem_type->is_loaded()) {
643 field_type = TypeInstPtr::BOTTOM;
644 } else if (field != NULL && field->is_constant()) {
645 // This can happen if the constant oop is non-perm.
646 ciObject* con = field->constant_value().as_object();
647 // Do not "join" in the previous type; it doesn't add value,
648 // and may yield a vacuous result if the field is of interface type.
649 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
650 assert(field_type != NULL, "field singleton type must be consistent");
651 } else {
652 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
653 }
654 if (UseCompressedOops) {
655 field_type = field_type->make_narrowoop();
656 basic_elem_type = T_NARROWOOP;
657 }
658 } else {
659 field_type = Type::get_const_basic_type(basic_elem_type);
660 }
661
662 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr();
663
664 Node *field_val = value_from_mem(mem, basic_elem_type, field_type, field_addr_type, alloc);
665 if (field_val == NULL) {
666 // we weren't able to find a value for this field,
667 // give up on eliminating this allocation
668 alloc->_is_scalar_replaceable = false; // don't try again
669 // remove any extra entries we added to the safepoint
670 uint last = sfpt->req() - 1;
671 for (int k = 0; k < j; k++) {
672 sfpt->del_req(last--);
673 }
674 // rollback processed safepoints
675 while (safepoints_done.length() > 0) {
676 SafePointNode* sfpt_done = safepoints_done.pop();
677 // remove any extra entries we added to the safepoint
678 last = sfpt_done->req() - 1;
679 for (int k = 0; k < nfields; k++) {
680 sfpt_done->del_req(last--);
681 }
682 JVMState *jvms = sfpt_done->jvms();
683 jvms->set_endoff(sfpt_done->req());
684 // Now make a pass over the debug information replacing any references
685 // to SafePointScalarObjectNode with the allocated object.
686 int start = jvms->debug_start();
687 int end = jvms->debug_end();
688 for (int i = start; i < end; i++) {
689 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
690 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
691 if (scobj->first_index() == sfpt_done->req() &&
692 scobj->n_fields() == (uint)nfields) {
693 assert(scobj->alloc() == alloc, "sanity");
694 sfpt_done->set_req(i, res);
695 }
696 }
697 }
698 }
699 #ifndef PRODUCT
700 if (PrintEliminateAllocations) {
701 if (field != NULL) {
702 tty->print("=== At SafePoint node %d can't find value of Field: ",
703 sfpt->_idx);
704 field->print();
705 int field_idx = C->get_alias_index(field_addr_type);
706 tty->print(" (alias_idx=%d)", field_idx);
707 } else { // Array's element
708 tty->print("=== At SafePoint node %d can't find value of array element [%d]",
709 sfpt->_idx, j);
710 }
711 tty->print(", which prevents elimination of: ");
712 if (res == NULL)
713 alloc->dump();
714 else
715 res->dump();
716 }
717 #endif
718 return false;
719 }
720 if (UseCompressedOops && field_type->isa_narrowoop()) {
721 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
722 // to be able scalar replace the allocation.
723 if (field_val->is_EncodeP()) {
724 field_val = field_val->in(1);
725 } else {
726 field_val = transform_later(new (C, 2) DecodeNNode(field_val, field_val->bottom_type()->make_ptr()));
727 }
728 }
729 sfpt->add_req(field_val);
730 }
731 JVMState *jvms = sfpt->jvms();
732 jvms->set_endoff(sfpt->req());
733 // Now make a pass over the debug information replacing any references
734 // to the allocated object with "sobj"
735 int start = jvms->debug_start();
736 int end = jvms->debug_end();
737 for (int i = start; i < end; i++) {
738 if (sfpt->in(i) == res) {
739 sfpt->set_req(i, sobj);
740 }
741 }
742 safepoints_done.append_if_missing(sfpt); // keep it for rollback
743 }
744 return true;
745 }
746
747 // Process users of eliminated allocation.
748 void PhaseMacroExpand::process_users_of_allocation(AllocateNode *alloc) {
749 Node* res = alloc->result_cast();
750 if (res != NULL) {
751 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
752 Node *use = res->last_out(j);
753 uint oc1 = res->outcnt();
754
755 if (use->is_AddP()) {
756 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
757 Node *n = use->last_out(k);
758 uint oc2 = use->outcnt();
759 if (n->is_Store()) {
760 _igvn.replace_node(n, n->in(MemNode::Memory));
761 } else {
762 assert( n->Opcode() == Op_CastP2X, "CastP2X required");
763 eliminate_card_mark(n);
764 }
765 k -= (oc2 - use->outcnt());
766 }
767 } else {
768 assert( !use->is_SafePoint(), "safepoint uses must have been already elimiated");
769 assert( use->Opcode() == Op_CastP2X, "CastP2X required");
770 eliminate_card_mark(use);
771 }
772 j -= (oc1 - res->outcnt());
773 }
774 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
775 _igvn.remove_dead_node(res);
776 }
777
778 //
779 // Process other users of allocation's projections
780 //
781 if (_resproj != NULL && _resproj->outcnt() != 0) {
782 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) {
783 Node *use = _resproj->last_out(j);
784 uint oc1 = _resproj->outcnt();
785 if (use->is_Initialize()) {
786 // Eliminate Initialize node.
787 InitializeNode *init = use->as_Initialize();
788 assert(init->outcnt() <= 2, "only a control and memory projection expected");
789 Node *ctrl_proj = init->proj_out(TypeFunc::Control);
790 if (ctrl_proj != NULL) {
791 assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection");
792 _igvn.replace_node(ctrl_proj, _fallthroughcatchproj);
793 }
794 Node *mem_proj = init->proj_out(TypeFunc::Memory);
795 if (mem_proj != NULL) {
796 Node *mem = init->in(TypeFunc::Memory);
797 #ifdef ASSERT
798 if (mem->is_MergeMem()) {
799 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection");
800 } else {
801 assert(mem == _memproj_fallthrough, "allocation memory projection");
802 }
803 #endif
804 _igvn.replace_node(mem_proj, mem);
805 }
806 } else if (use->is_AddP()) {
807 // raw memory addresses used only by the initialization
808 _igvn.hash_delete(use);
809 _igvn.subsume_node(use, C->top());
810 } else {
811 assert(false, "only Initialize or AddP expected");
812 }
813 j -= (oc1 - _resproj->outcnt());
814 }
815 }
816 if (_fallthroughcatchproj != NULL) {
817 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control));
818 }
819 if (_memproj_fallthrough != NULL) {
820 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory));
821 }
822 if (_memproj_catchall != NULL) {
823 _igvn.replace_node(_memproj_catchall, C->top());
824 }
825 if (_ioproj_fallthrough != NULL) {
826 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O));
827 }
828 if (_ioproj_catchall != NULL) {
829 _igvn.replace_node(_ioproj_catchall, C->top());
830 }
831 if (_catchallcatchproj != NULL) {
832 _igvn.replace_node(_catchallcatchproj, C->top());
833 }
834 }
835
836 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
837
838 if (!EliminateAllocations || !alloc->_is_scalar_replaceable) {
839 return false;
840 }
841
842 extract_call_projections(alloc);
843
844 GrowableArray <SafePointNode *> safepoints;
845 if (!can_eliminate_allocation(alloc, safepoints)) {
846 return false;
847 }
848
849 if (!scalar_replacement(alloc, safepoints)) {
850 return false;
851 }
852
853 process_users_of_allocation(alloc);
854
855 #ifndef PRODUCT
856 if (PrintEliminateAllocations) {
857 if (alloc->is_AllocateArray())
858 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
859 else
860 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
861 }
862 #endif
863
864 return true;
865 }
866
867
868 //---------------------------set_eden_pointers-------------------------
869 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) {
870 if (UseTLAB) { // Private allocation: load from TLS
871 Node* thread = transform_later(new (C, 1) ThreadLocalNode());
872 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset());
873 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset());
874 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset);
875 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset);
876 } else { // Shared allocation: load from globals
877 CollectedHeap* ch = Universe::heap();
878 address top_adr = (address)ch->top_addr();
879 address end_adr = (address)ch->end_addr();
880 eden_top_adr = makecon(TypeRawPtr::make(top_adr));
881 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr);
882 }
883 }
884
885
886 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
887 Node* adr = basic_plus_adr(base, offset);
888 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
889 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt);
890 transform_later(value);
891 return value;
892 }
893
894
895 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
896 Node* adr = basic_plus_adr(base, offset);
897 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt);
898 transform_later(mem);
899 return mem;
900 }
901
902 //=============================================================================
903 //
904 // A L L O C A T I O N
905 //
906 // Allocation attempts to be fast in the case of frequent small objects.
907 // It breaks down like this:
908 //
909 // 1) Size in doublewords is computed. This is a constant for objects and
910 // variable for most arrays. Doubleword units are used to avoid size
911 // overflow of huge doubleword arrays. We need doublewords in the end for
912 // rounding.
913 //
914 // 2) Size is checked for being 'too large'. Too-large allocations will go
915 // the slow path into the VM. The slow path can throw any required
916 // exceptions, and does all the special checks for very large arrays. The
917 // size test can constant-fold away for objects. For objects with
918 // finalizers it constant-folds the otherway: you always go slow with
919 // finalizers.
920 //
921 // 3) If NOT using TLABs, this is the contended loop-back point.
922 // Load-Locked the heap top. If using TLABs normal-load the heap top.
923 //
924 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
925 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
926 // "size*8" we always enter the VM, where "largish" is a constant picked small
927 // enough that there's always space between the eden max and 4Gig (old space is
928 // there so it's quite large) and large enough that the cost of entering the VM
929 // is dwarfed by the cost to initialize the space.
930 //
931 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
932 // down. If contended, repeat at step 3. If using TLABs normal-store
933 // adjusted heap top back down; there is no contention.
934 //
935 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
936 // fields.
937 //
938 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
939 // oop flavor.
940 //
941 //=============================================================================
942 // FastAllocateSizeLimit value is in DOUBLEWORDS.
943 // Allocations bigger than this always go the slow route.
944 // This value must be small enough that allocation attempts that need to
945 // trigger exceptions go the slow route. Also, it must be small enough so
946 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
947 //=============================================================================j//
948 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
949 // The allocator will coalesce int->oop copies away. See comment in
950 // coalesce.cpp about how this works. It depends critically on the exact
951 // code shape produced here, so if you are changing this code shape
952 // make sure the GC info for the heap-top is correct in and around the
953 // slow-path call.
954 //
955
956 void PhaseMacroExpand::expand_allocate_common(
957 AllocateNode* alloc, // allocation node to be expanded
958 Node* length, // array length for an array allocation
959 const TypeFunc* slow_call_type, // Type of slow call
960 address slow_call_address // Address of slow call
961 )
962 {
963
964 Node* ctrl = alloc->in(TypeFunc::Control);
965 Node* mem = alloc->in(TypeFunc::Memory);
966 Node* i_o = alloc->in(TypeFunc::I_O);
967 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
968 Node* klass_node = alloc->in(AllocateNode::KlassNode);
969 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
970
971 assert(ctrl != NULL, "must have control");
972 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
973 // they will not be used if "always_slow" is set
974 enum { slow_result_path = 1, fast_result_path = 2 };
975 Node *result_region;
976 Node *result_phi_rawmem;
977 Node *result_phi_rawoop;
978 Node *result_phi_i_o;
979
980 // The initial slow comparison is a size check, the comparison
981 // we want to do is a BoolTest::gt
982 bool always_slow = false;
983 int tv = _igvn.find_int_con(initial_slow_test, -1);
984 if (tv >= 0) {
985 always_slow = (tv == 1);
986 initial_slow_test = NULL;
987 } else {
988 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
989 }
990
991 if (DTraceAllocProbes ||
992 !UseTLAB && (!Universe::heap()->supports_inline_contig_alloc() ||
993 (UseConcMarkSweepGC && CMSIncrementalMode))) {
994 // Force slow-path allocation
995 always_slow = true;
996 initial_slow_test = NULL;
997 }
998
999
1000 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1001 Node *slow_region = NULL;
1002 Node *toobig_false = ctrl;
1003
1004 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent");
1005 // generate the initial test if necessary
1006 if (initial_slow_test != NULL ) {
1007 slow_region = new (C, 3) RegionNode(3);
1008
1009 // Now make the initial failure test. Usually a too-big test but
1010 // might be a TRUE for finalizers or a fancy class check for
1011 // newInstance0.
1012 IfNode *toobig_iff = new (C, 2) IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1013 transform_later(toobig_iff);
1014 // Plug the failing-too-big test into the slow-path region
1015 Node *toobig_true = new (C, 1) IfTrueNode( toobig_iff );
1016 transform_later(toobig_true);
1017 slow_region ->init_req( too_big_or_final_path, toobig_true );
1018 toobig_false = new (C, 1) IfFalseNode( toobig_iff );
1019 transform_later(toobig_false);
1020 } else { // No initial test, just fall into next case
1021 toobig_false = ctrl;
1022 debug_only(slow_region = NodeSentinel);
1023 }
1024
1025 Node *slow_mem = mem; // save the current memory state for slow path
1026 // generate the fast allocation code unless we know that the initial test will always go slow
1027 if (!always_slow) {
1028 // Fast path modifies only raw memory.
1029 if (mem->is_MergeMem()) {
1030 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1031 }
1032
1033 Node* eden_top_adr;
1034 Node* eden_end_adr;
1035
1036 set_eden_pointers(eden_top_adr, eden_end_adr);
1037
1038 // Load Eden::end. Loop invariant and hoisted.
1039 //
1040 // Note: We set the control input on "eden_end" and "old_eden_top" when using
1041 // a TLAB to work around a bug where these values were being moved across
1042 // a safepoint. These are not oops, so they cannot be include in the oop
1043 // map, but the can be changed by a GC. The proper way to fix this would
1044 // be to set the raw memory state when generating a SafepointNode. However
1045 // this will require extensive changes to the loop optimization in order to
1046 // prevent a degradation of the optimization.
1047 // See comment in memnode.hpp, around line 227 in class LoadPNode.
1048 Node *eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS);
1049
1050 // allocate the Region and Phi nodes for the result
1051 result_region = new (C, 3) RegionNode(3);
1052 result_phi_rawmem = new (C, 3) PhiNode( result_region, Type::MEMORY, TypeRawPtr::BOTTOM );
1053 result_phi_rawoop = new (C, 3) PhiNode( result_region, TypeRawPtr::BOTTOM );
1054 result_phi_i_o = new (C, 3) PhiNode( result_region, Type::ABIO ); // I/O is used for Prefetch
1055
1056 // We need a Region for the loop-back contended case.
1057 enum { fall_in_path = 1, contended_loopback_path = 2 };
1058 Node *contended_region;
1059 Node *contended_phi_rawmem;
1060 if( UseTLAB ) {
1061 contended_region = toobig_false;
1062 contended_phi_rawmem = mem;
1063 } else {
1064 contended_region = new (C, 3) RegionNode(3);
1065 contended_phi_rawmem = new (C, 3) PhiNode( contended_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1066 // Now handle the passing-too-big test. We fall into the contended
1067 // loop-back merge point.
1068 contended_region ->init_req( fall_in_path, toobig_false );
1069 contended_phi_rawmem->init_req( fall_in_path, mem );
1070 transform_later(contended_region);
1071 transform_later(contended_phi_rawmem);
1072 }
1073
1074 // Load(-locked) the heap top.
1075 // See note above concerning the control input when using a TLAB
1076 Node *old_eden_top = UseTLAB
1077 ? new (C, 3) LoadPNode ( ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM )
1078 : new (C, 3) LoadPLockedNode( contended_region, contended_phi_rawmem, eden_top_adr );
1079
1080 transform_later(old_eden_top);
1081 // Add to heap top to get a new heap top
1082 Node *new_eden_top = new (C, 4) AddPNode( top(), old_eden_top, size_in_bytes );
1083 transform_later(new_eden_top);
1084 // Check for needing a GC; compare against heap end
1085 Node *needgc_cmp = new (C, 3) CmpPNode( new_eden_top, eden_end );
1086 transform_later(needgc_cmp);
1087 Node *needgc_bol = new (C, 2) BoolNode( needgc_cmp, BoolTest::ge );
1088 transform_later(needgc_bol);
1089 IfNode *needgc_iff = new (C, 2) IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1090 transform_later(needgc_iff);
1091
1092 // Plug the failing-heap-space-need-gc test into the slow-path region
1093 Node *needgc_true = new (C, 1) IfTrueNode( needgc_iff );
1094 transform_later(needgc_true);
1095 if( initial_slow_test ) {
1096 slow_region ->init_req( need_gc_path, needgc_true );
1097 // This completes all paths into the slow merge point
1098 transform_later(slow_region);
1099 } else { // No initial slow path needed!
1100 // Just fall from the need-GC path straight into the VM call.
1101 slow_region = needgc_true;
1102 }
1103 // No need for a GC. Setup for the Store-Conditional
1104 Node *needgc_false = new (C, 1) IfFalseNode( needgc_iff );
1105 transform_later(needgc_false);
1106
1107 // Grab regular I/O before optional prefetch may change it.
1108 // Slow-path does no I/O so just set it to the original I/O.
1109 result_phi_i_o->init_req( slow_result_path, i_o );
1110
1111 i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem,
1112 old_eden_top, new_eden_top, length);
1113
1114 // Store (-conditional) the modified eden top back down.
1115 // StorePConditional produces flags for a test PLUS a modified raw
1116 // memory state.
1117 Node *store_eden_top;
1118 Node *fast_oop_ctrl;
1119 if( UseTLAB ) {
1120 store_eden_top = new (C, 4) StorePNode( needgc_false, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, new_eden_top );
1121 transform_later(store_eden_top);
1122 fast_oop_ctrl = needgc_false; // No contention, so this is the fast path
1123 } else {
1124 store_eden_top = new (C, 5) StorePConditionalNode( needgc_false, contended_phi_rawmem, eden_top_adr, new_eden_top, old_eden_top );
1125 transform_later(store_eden_top);
1126 Node *contention_check = new (C, 2) BoolNode( store_eden_top, BoolTest::ne );
1127 transform_later(contention_check);
1128 store_eden_top = new (C, 1) SCMemProjNode(store_eden_top);
1129 transform_later(store_eden_top);
1130
1131 // If not using TLABs, check to see if there was contention.
1132 IfNode *contention_iff = new (C, 2) IfNode ( needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN );
1133 transform_later(contention_iff);
1134 Node *contention_true = new (C, 1) IfTrueNode( contention_iff );
1135 transform_later(contention_true);
1136 // If contention, loopback and try again.
1137 contended_region->init_req( contended_loopback_path, contention_true );
1138 contended_phi_rawmem->init_req( contended_loopback_path, store_eden_top );
1139
1140 // Fast-path succeeded with no contention!
1141 Node *contention_false = new (C, 1) IfFalseNode( contention_iff );
1142 transform_later(contention_false);
1143 fast_oop_ctrl = contention_false;
1144 }
1145
1146 // Rename successful fast-path variables to make meaning more obvious
1147 Node* fast_oop = old_eden_top;
1148 Node* fast_oop_rawmem = store_eden_top;
1149 fast_oop_rawmem = initialize_object(alloc,
1150 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1151 klass_node, length, size_in_bytes);
1152
1153 if (ExtendedDTraceProbes) {
1154 // Slow-path call
1155 int size = TypeFunc::Parms + 2;
1156 CallLeafNode *call = new (C, size) CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1157 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base),
1158 "dtrace_object_alloc",
1159 TypeRawPtr::BOTTOM);
1160
1161 // Get base of thread-local storage area
1162 Node* thread = new (C, 1) ThreadLocalNode();
1163 transform_later(thread);
1164
1165 call->init_req(TypeFunc::Parms+0, thread);
1166 call->init_req(TypeFunc::Parms+1, fast_oop);
1167 call->init_req( TypeFunc::Control, fast_oop_ctrl );
1168 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o
1169 call->init_req( TypeFunc::Memory , fast_oop_rawmem );
1170 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) );
1171 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) );
1172 transform_later(call);
1173 fast_oop_ctrl = new (C, 1) ProjNode(call,TypeFunc::Control);
1174 transform_later(fast_oop_ctrl);
1175 fast_oop_rawmem = new (C, 1) ProjNode(call,TypeFunc::Memory);
1176 transform_later(fast_oop_rawmem);
1177 }
1178
1179 // Plug in the successful fast-path into the result merge point
1180 result_region ->init_req( fast_result_path, fast_oop_ctrl );
1181 result_phi_rawoop->init_req( fast_result_path, fast_oop );
1182 result_phi_i_o ->init_req( fast_result_path, i_o );
1183 result_phi_rawmem->init_req( fast_result_path, fast_oop_rawmem );
1184 } else {
1185 slow_region = ctrl;
1186 }
1187
1188 // Generate slow-path call
1189 CallNode *call = new (C, slow_call_type->domain()->cnt())
1190 CallStaticJavaNode(slow_call_type, slow_call_address,
1191 OptoRuntime::stub_name(slow_call_address),
1192 alloc->jvms()->bci(),
1193 TypePtr::BOTTOM);
1194 call->init_req( TypeFunc::Control, slow_region );
1195 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o
1196 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs
1197 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) );
1198 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) );
1199
1200 call->init_req(TypeFunc::Parms+0, klass_node);
1201 if (length != NULL) {
1202 call->init_req(TypeFunc::Parms+1, length);
1203 }
1204
1205 // Copy debug information and adjust JVMState information, then replace
1206 // allocate node with the call
1207 copy_call_debug_info((CallNode *) alloc, call);
1208 if (!always_slow) {
1209 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1210 }
1211 _igvn.hash_delete(alloc);
1212 _igvn.subsume_node(alloc, call);
1213 transform_later(call);
1214
1215 // Identify the output projections from the allocate node and
1216 // adjust any references to them.
1217 // The control and io projections look like:
1218 //
1219 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1220 // Allocate Catch
1221 // ^---Proj(io) <-------+ ^---CatchProj(io)
1222 //
1223 // We are interested in the CatchProj nodes.
1224 //
1225 extract_call_projections(call);
1226
1227 // An allocate node has separate memory projections for the uses on the control and i_o paths
1228 // Replace uses of the control memory projection with result_phi_rawmem (unless we are only generating a slow call)
1229 if (!always_slow && _memproj_fallthrough != NULL) {
1230 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) {
1231 Node *use = _memproj_fallthrough->fast_out(i);
1232 _igvn.hash_delete(use);
1233 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem);
1234 _igvn._worklist.push(use);
1235 // back up iterator
1236 --i;
1237 }
1238 }
1239 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete _memproj_catchall so
1240 // we end up with a call that has only 1 memory projection
1241 if (_memproj_catchall != NULL ) {
1242 if (_memproj_fallthrough == NULL) {
1243 _memproj_fallthrough = new (C, 1) ProjNode(call, TypeFunc::Memory);
1244 transform_later(_memproj_fallthrough);
1245 }
1246 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) {
1247 Node *use = _memproj_catchall->fast_out(i);
1248 _igvn.hash_delete(use);
1249 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough);
1250 _igvn._worklist.push(use);
1251 // back up iterator
1252 --i;
1253 }
1254 }
1255
1256 // An allocate node has separate i_o projections for the uses on the control and i_o paths
1257 // Replace uses of the control i_o projection with result_phi_i_o (unless we are only generating a slow call)
1258 if (_ioproj_fallthrough == NULL) {
1259 _ioproj_fallthrough = new (C, 1) ProjNode(call, TypeFunc::I_O);
1260 transform_later(_ioproj_fallthrough);
1261 } else if (!always_slow) {
1262 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) {
1263 Node *use = _ioproj_fallthrough->fast_out(i);
1264
1265 _igvn.hash_delete(use);
1266 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o);
1267 _igvn._worklist.push(use);
1268 // back up iterator
1269 --i;
1270 }
1271 }
1272 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete _ioproj_catchall so
1273 // we end up with a call that has only 1 control projection
1274 if (_ioproj_catchall != NULL ) {
1275 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) {
1276 Node *use = _ioproj_catchall->fast_out(i);
1277 _igvn.hash_delete(use);
1278 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough);
1279 _igvn._worklist.push(use);
1280 // back up iterator
1281 --i;
1282 }
1283 }
1284
1285 // if we generated only a slow call, we are done
1286 if (always_slow)
1287 return;
1288
1289
1290 if (_fallthroughcatchproj != NULL) {
1291 ctrl = _fallthroughcatchproj->clone();
1292 transform_later(ctrl);
1293 _igvn.hash_delete(_fallthroughcatchproj);
1294 _igvn.subsume_node(_fallthroughcatchproj, result_region);
1295 } else {
1296 ctrl = top();
1297 }
1298 Node *slow_result;
1299 if (_resproj == NULL) {
1300 // no uses of the allocation result
1301 slow_result = top();
1302 } else {
1303 slow_result = _resproj->clone();
1304 transform_later(slow_result);
1305 _igvn.hash_delete(_resproj);
1306 _igvn.subsume_node(_resproj, result_phi_rawoop);
1307 }
1308
1309 // Plug slow-path into result merge point
1310 result_region ->init_req( slow_result_path, ctrl );
1311 result_phi_rawoop->init_req( slow_result_path, slow_result);
1312 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough );
1313 transform_later(result_region);
1314 transform_later(result_phi_rawoop);
1315 transform_later(result_phi_rawmem);
1316 transform_later(result_phi_i_o);
1317 // This completes all paths into the result merge point
1318 }
1319
1320
1321 // Helper for PhaseMacroExpand::expand_allocate_common.
1322 // Initializes the newly-allocated storage.
1323 Node*
1324 PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1325 Node* control, Node* rawmem, Node* object,
1326 Node* klass_node, Node* length,
1327 Node* size_in_bytes) {
1328 InitializeNode* init = alloc->initialization();
1329 // Store the klass & mark bits
1330 Node* mark_node = NULL;
1331 // For now only enable fast locking for non-array types
1332 if (UseBiasedLocking && (length == NULL)) {
1333 mark_node = make_load(NULL, rawmem, klass_node, Klass::prototype_header_offset_in_bytes() + sizeof(oopDesc), TypeRawPtr::BOTTOM, T_ADDRESS);
1334 } else {
1335 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype()));
1336 }
1337 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
1338
1339 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_OBJECT);
1340 int header_size = alloc->minimum_header_size(); // conservatively small
1341
1342 // Array length
1343 if (length != NULL) { // Arrays need length field
1344 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1345 // conservatively small header size:
1346 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1347 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1348 if (k->is_array_klass()) // we know the exact header size in most cases:
1349 header_size = Klass::layout_helper_header_size(k->layout_helper());
1350 }
1351
1352 // Clear the object body, if necessary.
1353 if (init == NULL) {
1354 // The init has somehow disappeared; be cautious and clear everything.
1355 //
1356 // This can happen if a node is allocated but an uncommon trap occurs
1357 // immediately. In this case, the Initialize gets associated with the
1358 // trap, and may be placed in a different (outer) loop, if the Allocate
1359 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1360 // there can be two Allocates to one Initialize. The answer in all these
1361 // edge cases is safety first. It is always safe to clear immediately
1362 // within an Allocate, and then (maybe or maybe not) clear some more later.
1363 if (!ZeroTLAB)
1364 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1365 header_size, size_in_bytes,
1366 &_igvn);
1367 } else {
1368 if (!init->is_complete()) {
1369 // Try to win by zeroing only what the init does not store.
1370 // We can also try to do some peephole optimizations,
1371 // such as combining some adjacent subword stores.
1372 rawmem = init->complete_stores(control, rawmem, object,
1373 header_size, size_in_bytes, &_igvn);
1374 }
1375 // We have no more use for this link, since the AllocateNode goes away:
1376 init->set_req(InitializeNode::RawAddress, top());
1377 // (If we keep the link, it just confuses the register allocator,
1378 // who thinks he sees a real use of the address by the membar.)
1379 }
1380
1381 return rawmem;
1382 }
1383
1384 // Generate prefetch instructions for next allocations.
1385 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
1386 Node*& contended_phi_rawmem,
1387 Node* old_eden_top, Node* new_eden_top,
1388 Node* length) {
1389 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
1390 // Generate prefetch allocation with watermark check.
1391 // As an allocation hits the watermark, we will prefetch starting
1392 // at a "distance" away from watermark.
1393 enum { fall_in_path = 1, pf_path = 2 };
1394
1395 Node *pf_region = new (C, 3) RegionNode(3);
1396 Node *pf_phi_rawmem = new (C, 3) PhiNode( pf_region, Type::MEMORY,
1397 TypeRawPtr::BOTTOM );
1398 // I/O is used for Prefetch
1399 Node *pf_phi_abio = new (C, 3) PhiNode( pf_region, Type::ABIO );
1400
1401 Node *thread = new (C, 1) ThreadLocalNode();
1402 transform_later(thread);
1403
1404 Node *eden_pf_adr = new (C, 4) AddPNode( top()/*not oop*/, thread,
1405 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
1406 transform_later(eden_pf_adr);
1407
1408 Node *old_pf_wm = new (C, 3) LoadPNode( needgc_false,
1409 contended_phi_rawmem, eden_pf_adr,
1410 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM );
1411 transform_later(old_pf_wm);
1412
1413 // check against new_eden_top
1414 Node *need_pf_cmp = new (C, 3) CmpPNode( new_eden_top, old_pf_wm );
1415 transform_later(need_pf_cmp);
1416 Node *need_pf_bol = new (C, 2) BoolNode( need_pf_cmp, BoolTest::ge );
1417 transform_later(need_pf_bol);
1418 IfNode *need_pf_iff = new (C, 2) IfNode( needgc_false, need_pf_bol,
1419 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1420 transform_later(need_pf_iff);
1421
1422 // true node, add prefetchdistance
1423 Node *need_pf_true = new (C, 1) IfTrueNode( need_pf_iff );
1424 transform_later(need_pf_true);
1425
1426 Node *need_pf_false = new (C, 1) IfFalseNode( need_pf_iff );
1427 transform_later(need_pf_false);
1428
1429 Node *new_pf_wmt = new (C, 4) AddPNode( top(), old_pf_wm,
1430 _igvn.MakeConX(AllocatePrefetchDistance) );
1431 transform_later(new_pf_wmt );
1432 new_pf_wmt->set_req(0, need_pf_true);
1433
1434 Node *store_new_wmt = new (C, 4) StorePNode( need_pf_true,
1435 contended_phi_rawmem, eden_pf_adr,
1436 TypeRawPtr::BOTTOM, new_pf_wmt );
1437 transform_later(store_new_wmt);
1438
1439 // adding prefetches
1440 pf_phi_abio->init_req( fall_in_path, i_o );
1441
1442 Node *prefetch_adr;
1443 Node *prefetch;
1444 uint lines = AllocatePrefetchDistance / AllocatePrefetchStepSize;
1445 uint step_size = AllocatePrefetchStepSize;
1446 uint distance = 0;
1447
1448 for ( uint i = 0; i < lines; i++ ) {
1449 prefetch_adr = new (C, 4) AddPNode( old_pf_wm, new_pf_wmt,
1450 _igvn.MakeConX(distance) );
1451 transform_later(prefetch_adr);
1452 prefetch = new (C, 3) PrefetchWriteNode( i_o, prefetch_adr );
1453 transform_later(prefetch);
1454 distance += step_size;
1455 i_o = prefetch;
1456 }
1457 pf_phi_abio->set_req( pf_path, i_o );
1458
1459 pf_region->init_req( fall_in_path, need_pf_false );
1460 pf_region->init_req( pf_path, need_pf_true );
1461
1462 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
1463 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
1464
1465 transform_later(pf_region);
1466 transform_later(pf_phi_rawmem);
1467 transform_later(pf_phi_abio);
1468
1469 needgc_false = pf_region;
1470 contended_phi_rawmem = pf_phi_rawmem;
1471 i_o = pf_phi_abio;
1472 } else if( AllocatePrefetchStyle > 0 ) {
1473 // Insert a prefetch for each allocation only on the fast-path
1474 Node *prefetch_adr;
1475 Node *prefetch;
1476 // Generate several prefetch instructions only for arrays.
1477 uint lines = (length != NULL) ? AllocatePrefetchLines : 1;
1478 uint step_size = AllocatePrefetchStepSize;
1479 uint distance = AllocatePrefetchDistance;
1480 for ( uint i = 0; i < lines; i++ ) {
1481 prefetch_adr = new (C, 4) AddPNode( old_eden_top, new_eden_top,
1482 _igvn.MakeConX(distance) );
1483 transform_later(prefetch_adr);
1484 prefetch = new (C, 3) PrefetchWriteNode( i_o, prefetch_adr );
1485 // Do not let it float too high, since if eden_top == eden_end,
1486 // both might be null.
1487 if( i == 0 ) { // Set control for first prefetch, next follows it
1488 prefetch->init_req(0, needgc_false);
1489 }
1490 transform_later(prefetch);
1491 distance += step_size;
1492 i_o = prefetch;
1493 }
1494 }
1495 return i_o;
1496 }
1497
1498
1499 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
1500 expand_allocate_common(alloc, NULL,
1501 OptoRuntime::new_instance_Type(),
1502 OptoRuntime::new_instance_Java());
1503 }
1504
1505 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
1506 Node* length = alloc->in(AllocateNode::ALength);
1507 expand_allocate_common(alloc, length,
1508 OptoRuntime::new_array_Type(),
1509 OptoRuntime::new_array_Java());
1510 }
1511
1512
1513 // we have determined that this lock/unlock can be eliminated, we simply
1514 // eliminate the node without expanding it.
1515 //
1516 // Note: The membar's associated with the lock/unlock are currently not
1517 // eliminated. This should be investigated as a future enhancement.
1518 //
1519 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
1520
1521 if (!alock->is_eliminated()) {
1522 return false;
1523 }
1524 if (alock->is_Lock() && !alock->is_coarsened()) {
1525 // Create new "eliminated" BoxLock node and use it
1526 // in monitor debug info for the same object.
1527 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
1528 Node* obj = alock->obj_node();
1529 if (!oldbox->is_eliminated()) {
1530 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
1531 newbox->set_eliminated();
1532 transform_later(newbox);
1533 // Replace old box node with new box for all users
1534 // of the same object.
1535 for (uint i = 0; i < oldbox->outcnt();) {
1536
1537 bool next_edge = true;
1538 Node* u = oldbox->raw_out(i);
1539 if (u == alock) {
1540 i++;
1541 continue; // It will be removed below
1542 }
1543 if (u->is_Lock() &&
1544 u->as_Lock()->obj_node() == obj &&
1545 // oldbox could be referenced in debug info also
1546 u->as_Lock()->box_node() == oldbox) {
1547 assert(u->as_Lock()->is_eliminated(), "sanity");
1548 _igvn.hash_delete(u);
1549 u->set_req(TypeFunc::Parms + 1, newbox);
1550 next_edge = false;
1551 #ifdef ASSERT
1552 } else if (u->is_Unlock() && u->as_Unlock()->obj_node() == obj) {
1553 assert(u->as_Unlock()->is_eliminated(), "sanity");
1554 #endif
1555 }
1556 // Replace old box in monitor debug info.
1557 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
1558 SafePointNode* sfn = u->as_SafePoint();
1559 JVMState* youngest_jvms = sfn->jvms();
1560 int max_depth = youngest_jvms->depth();
1561 for (int depth = 1; depth <= max_depth; depth++) {
1562 JVMState* jvms = youngest_jvms->of_depth(depth);
1563 int num_mon = jvms->nof_monitors();
1564 // Loop over monitors
1565 for (int idx = 0; idx < num_mon; idx++) {
1566 Node* obj_node = sfn->monitor_obj(jvms, idx);
1567 Node* box_node = sfn->monitor_box(jvms, idx);
1568 if (box_node == oldbox && obj_node == obj) {
1569 int j = jvms->monitor_box_offset(idx);
1570 _igvn.hash_delete(u);
1571 u->set_req(j, newbox);
1572 next_edge = false;
1573 }
1574 } // for (int idx = 0;
1575 } // for (int depth = 1;
1576 } // if (u->is_SafePoint()
1577 if (next_edge) i++;
1578 } // for (uint i = 0; i < oldbox->outcnt();)
1579 } // if (!oldbox->is_eliminated())
1580 } // if (alock->is_Lock() && !lock->is_coarsened())
1581
1582 #ifndef PRODUCT
1583 if (PrintEliminateLocks) {
1584 if (alock->is_Lock()) {
1585 tty->print_cr("++++ Eliminating: %d Lock", alock->_idx);
1586 } else {
1587 tty->print_cr("++++ Eliminating: %d Unlock", alock->_idx);
1588 }
1589 }
1590 #endif
1591
1592 Node* mem = alock->in(TypeFunc::Memory);
1593 Node* ctrl = alock->in(TypeFunc::Control);
1594
1595 extract_call_projections(alock);
1596 // There are 2 projections from the lock. The lock node will
1597 // be deleted when its last use is subsumed below.
1598 assert(alock->outcnt() == 2 &&
1599 _fallthroughproj != NULL &&
1600 _memproj_fallthrough != NULL,
1601 "Unexpected projections from Lock/Unlock");
1602
1603 Node* fallthroughproj = _fallthroughproj;
1604 Node* memproj_fallthrough = _memproj_fallthrough;
1605
1606 // The memory projection from a lock/unlock is RawMem
1607 // The input to a Lock is merged memory, so extract its RawMem input
1608 // (unless the MergeMem has been optimized away.)
1609 if (alock->is_Lock()) {
1610 // Seach for MemBarAcquire node and delete it also.
1611 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
1612 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquire, "");
1613 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
1614 Node* memproj = membar->proj_out(TypeFunc::Memory);
1615 _igvn.hash_delete(ctrlproj);
1616 _igvn.subsume_node(ctrlproj, fallthroughproj);
1617 _igvn.hash_delete(memproj);
1618 _igvn.subsume_node(memproj, memproj_fallthrough);
1619
1620 // Delete FastLock node also if this Lock node is unique user
1621 // (a loop peeling may clone a Lock node).
1622 Node* flock = alock->as_Lock()->fastlock_node();
1623 if (flock->outcnt() == 1) {
1624 assert(flock->unique_out() == alock, "sanity");
1625 _igvn.hash_delete(flock);
1626 _igvn.subsume_node(flock, top());
1627 }
1628 }
1629
1630 // Seach for MemBarRelease node and delete it also.
1631 if (alock->is_Unlock() && ctrl != NULL && ctrl->is_Proj() &&
1632 ctrl->in(0)->is_MemBar()) {
1633 MemBarNode* membar = ctrl->in(0)->as_MemBar();
1634 assert(membar->Opcode() == Op_MemBarRelease &&
1635 mem->is_Proj() && membar == mem->in(0), "");
1636 _igvn.hash_delete(fallthroughproj);
1637 _igvn.subsume_node(fallthroughproj, ctrl);
1638 _igvn.hash_delete(memproj_fallthrough);
1639 _igvn.subsume_node(memproj_fallthrough, mem);
1640 fallthroughproj = ctrl;
1641 memproj_fallthrough = mem;
1642 ctrl = membar->in(TypeFunc::Control);
1643 mem = membar->in(TypeFunc::Memory);
1644 }
1645
1646 _igvn.hash_delete(fallthroughproj);
1647 _igvn.subsume_node(fallthroughproj, ctrl);
1648 _igvn.hash_delete(memproj_fallthrough);
1649 _igvn.subsume_node(memproj_fallthrough, mem);
1650 return true;
1651 }
1652
1653
1654 //------------------------------expand_lock_node----------------------
1655 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
1656
1657 Node* ctrl = lock->in(TypeFunc::Control);
1658 Node* mem = lock->in(TypeFunc::Memory);
1659 Node* obj = lock->obj_node();
1660 Node* box = lock->box_node();
1661 Node* flock = lock->fastlock_node();
1662
1663 // Make the merge point
1664 Node *region;
1665 Node *mem_phi;
1666 Node *slow_path;
1667
1668 if (UseOptoBiasInlining) {
1669 /*
1670 * See the full descrition in MacroAssembler::biased_locking_enter().
1671 *
1672 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) {
1673 * // The object is biased.
1674 * proto_node = klass->prototype_header;
1675 * o_node = thread | proto_node;
1676 * x_node = o_node ^ mark_word;
1677 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ?
1678 * // Done.
1679 * } else {
1680 * if( (x_node & biased_lock_mask) != 0 ) {
1681 * // The klass's prototype header is no longer biased.
1682 * cas(&mark_word, mark_word, proto_node)
1683 * goto cas_lock;
1684 * } else {
1685 * // The klass's prototype header is still biased.
1686 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch?
1687 * old = mark_word;
1688 * new = o_node;
1689 * } else {
1690 * // Different thread or anonymous biased.
1691 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask);
1692 * new = thread | old;
1693 * }
1694 * // Try to rebias.
1695 * if( cas(&mark_word, old, new) == 0 ) {
1696 * // Done.
1697 * } else {
1698 * goto slow_path; // Failed.
1699 * }
1700 * }
1701 * }
1702 * } else {
1703 * // The object is not biased.
1704 * cas_lock:
1705 * if( FastLock(obj) == 0 ) {
1706 * // Done.
1707 * } else {
1708 * slow_path:
1709 * OptoRuntime::complete_monitor_locking_Java(obj);
1710 * }
1711 * }
1712 */
1713
1714 region = new (C, 5) RegionNode(5);
1715 // create a Phi for the memory state
1716 mem_phi = new (C, 5) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
1717
1718 Node* fast_lock_region = new (C, 3) RegionNode(3);
1719 Node* fast_lock_mem_phi = new (C, 3) PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1720
1721 // First, check mark word for the biased lock pattern.
1722 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
1723
1724 // Get fast path - mark word has the biased lock pattern.
1725 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node,
1726 markOopDesc::biased_lock_mask_in_place,
1727 markOopDesc::biased_lock_pattern, true);
1728 // fast_lock_region->in(1) is set to slow path.
1729 fast_lock_mem_phi->init_req(1, mem);
1730
1731 // Now check that the lock is biased to the current thread and has
1732 // the same epoch and bias as Klass::_prototype_header.
1733
1734 // Special-case a fresh allocation to avoid building nodes:
1735 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn);
1736 if (klass_node == NULL) {
1737 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1738 klass_node = transform_later( LoadKlassNode::make(_igvn, mem, k_adr, _igvn.type(k_adr)->is_ptr()) );
1739 #ifdef _LP64
1740 if (UseCompressedOops && klass_node->is_DecodeN()) {
1741 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity");
1742 klass_node->in(1)->init_req(0, ctrl);
1743 } else
1744 #endif
1745 klass_node->init_req(0, ctrl);
1746 }
1747 Node *proto_node = make_load(ctrl, mem, klass_node, Klass::prototype_header_offset_in_bytes() + sizeof(oopDesc), TypeX_X, TypeX_X->basic_type());
1748
1749 Node* thread = transform_later(new (C, 1) ThreadLocalNode());
1750 Node* cast_thread = transform_later(new (C, 2) CastP2XNode(ctrl, thread));
1751 Node* o_node = transform_later(new (C, 3) OrXNode(cast_thread, proto_node));
1752 Node* x_node = transform_later(new (C, 3) XorXNode(o_node, mark_node));
1753
1754 // Get slow path - mark word does NOT match the value.
1755 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node,
1756 (~markOopDesc::age_mask_in_place), 0);
1757 // region->in(3) is set to fast path - the object is biased to the current thread.
1758 mem_phi->init_req(3, mem);
1759
1760
1761 // Mark word does NOT match the value (thread | Klass::_prototype_header).
1762
1763
1764 // First, check biased pattern.
1765 // Get fast path - _prototype_header has the same biased lock pattern.
1766 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node,
1767 markOopDesc::biased_lock_mask_in_place, 0, true);
1768
1769 not_biased_ctrl = fast_lock_region->in(2); // Slow path
1770 // fast_lock_region->in(2) - the prototype header is no longer biased
1771 // and we have to revoke the bias on this object.
1772 // We are going to try to reset the mark of this object to the prototype
1773 // value and fall through to the CAS-based locking scheme.
1774 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
1775 Node* cas = new (C, 5) StoreXConditionalNode(not_biased_ctrl, mem, adr,
1776 proto_node, mark_node);
1777 transform_later(cas);
1778 Node* proj = transform_later( new (C, 1) SCMemProjNode(cas));
1779 fast_lock_mem_phi->init_req(2, proj);
1780
1781
1782 // Second, check epoch bits.
1783 Node* rebiased_region = new (C, 3) RegionNode(3);
1784 Node* old_phi = new (C, 3) PhiNode( rebiased_region, TypeX_X);
1785 Node* new_phi = new (C, 3) PhiNode( rebiased_region, TypeX_X);
1786
1787 // Get slow path - mark word does NOT match epoch bits.
1788 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node,
1789 markOopDesc::epoch_mask_in_place, 0);
1790 // The epoch of the current bias is not valid, attempt to rebias the object
1791 // toward the current thread.
1792 rebiased_region->init_req(2, epoch_ctrl);
1793 old_phi->init_req(2, mark_node);
1794 new_phi->init_req(2, o_node);
1795
1796 // rebiased_region->in(1) is set to fast path.
1797 // The epoch of the current bias is still valid but we know
1798 // nothing about the owner; it might be set or it might be clear.
1799 Node* cmask = MakeConX(markOopDesc::biased_lock_mask_in_place |
1800 markOopDesc::age_mask_in_place |
1801 markOopDesc::epoch_mask_in_place);
1802 Node* old = transform_later(new (C, 3) AndXNode(mark_node, cmask));
1803 cast_thread = transform_later(new (C, 2) CastP2XNode(ctrl, thread));
1804 Node* new_mark = transform_later(new (C, 3) OrXNode(cast_thread, old));
1805 old_phi->init_req(1, old);
1806 new_phi->init_req(1, new_mark);
1807
1808 transform_later(rebiased_region);
1809 transform_later(old_phi);
1810 transform_later(new_phi);
1811
1812 // Try to acquire the bias of the object using an atomic operation.
1813 // If this fails we will go in to the runtime to revoke the object's bias.
1814 cas = new (C, 5) StoreXConditionalNode(rebiased_region, mem, adr,
1815 new_phi, old_phi);
1816 transform_later(cas);
1817 proj = transform_later( new (C, 1) SCMemProjNode(cas));
1818
1819 // Get slow path - Failed to CAS.
1820 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0);
1821 mem_phi->init_req(4, proj);
1822 // region->in(4) is set to fast path - the object is rebiased to the current thread.
1823
1824 // Failed to CAS.
1825 slow_path = new (C, 3) RegionNode(3);
1826 Node *slow_mem = new (C, 3) PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM);
1827
1828 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control
1829 slow_mem->init_req(1, proj);
1830
1831 // Call CAS-based locking scheme (FastLock node).
1832
1833 transform_later(fast_lock_region);
1834 transform_later(fast_lock_mem_phi);
1835
1836 // Get slow path - FastLock failed to lock the object.
1837 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0);
1838 mem_phi->init_req(2, fast_lock_mem_phi);
1839 // region->in(2) is set to fast path - the object is locked to the current thread.
1840
1841 slow_path->init_req(2, ctrl); // Capture slow-control
1842 slow_mem->init_req(2, fast_lock_mem_phi);
1843
1844 transform_later(slow_path);
1845 transform_later(slow_mem);
1846 // Reset lock's memory edge.
1847 lock->set_req(TypeFunc::Memory, slow_mem);
1848
1849 } else {
1850 region = new (C, 3) RegionNode(3);
1851 // create a Phi for the memory state
1852 mem_phi = new (C, 3) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
1853
1854 // Optimize test; set region slot 2
1855 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
1856 mem_phi->init_req(2, mem);
1857 }
1858
1859 // Make slow path call
1860 CallNode *call = make_slow_call( (CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(), OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path, obj, box );
1861
1862 extract_call_projections(call);
1863
1864 // Slow path can only throw asynchronous exceptions, which are always
1865 // de-opted. So the compiler thinks the slow-call can never throw an
1866 // exception. If it DOES throw an exception we would need the debug
1867 // info removed first (since if it throws there is no monitor).
1868 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
1869 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
1870
1871 // Capture slow path
1872 // disconnect fall-through projection from call and create a new one
1873 // hook up users of fall-through projection to region
1874 Node *slow_ctrl = _fallthroughproj->clone();
1875 transform_later(slow_ctrl);
1876 _igvn.hash_delete(_fallthroughproj);
1877 _fallthroughproj->disconnect_inputs(NULL);
1878 region->init_req(1, slow_ctrl);
1879 // region inputs are now complete
1880 transform_later(region);
1881 _igvn.subsume_node(_fallthroughproj, region);
1882
1883 Node *memproj = transform_later( new(C, 1) ProjNode(call, TypeFunc::Memory) );
1884 mem_phi->init_req(1, memproj );
1885 transform_later(mem_phi);
1886 _igvn.hash_delete(_memproj_fallthrough);
1887 _igvn.subsume_node(_memproj_fallthrough, mem_phi);
1888 }
1889
1890 //------------------------------expand_unlock_node----------------------
1891 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
1892
1893 Node* ctrl = unlock->in(TypeFunc::Control);
1894 Node* mem = unlock->in(TypeFunc::Memory);
1895 Node* obj = unlock->obj_node();
1896 Node* box = unlock->box_node();
1897
1898 // No need for a null check on unlock
1899
1900 // Make the merge point
1901 Node *region;
1902 Node *mem_phi;
1903
1904 if (UseOptoBiasInlining) {
1905 // Check for biased locking unlock case, which is a no-op.
1906 // See the full descrition in MacroAssembler::biased_locking_exit().
1907 region = new (C, 4) RegionNode(4);
1908 // create a Phi for the memory state
1909 mem_phi = new (C, 4) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
1910 mem_phi->init_req(3, mem);
1911
1912 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
1913 ctrl = opt_bits_test(ctrl, region, 3, mark_node,
1914 markOopDesc::biased_lock_mask_in_place,
1915 markOopDesc::biased_lock_pattern);
1916 } else {
1917 region = new (C, 3) RegionNode(3);
1918 // create a Phi for the memory state
1919 mem_phi = new (C, 3) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
1920 }
1921
1922 FastUnlockNode *funlock = new (C, 3) FastUnlockNode( ctrl, obj, box );
1923 funlock = transform_later( funlock )->as_FastUnlock();
1924 // Optimize test; set region slot 2
1925 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
1926
1927 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 );
1928
1929 extract_call_projections(call);
1930
1931 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
1932 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
1933
1934 // No exceptions for unlocking
1935 // Capture slow path
1936 // disconnect fall-through projection from call and create a new one
1937 // hook up users of fall-through projection to region
1938 Node *slow_ctrl = _fallthroughproj->clone();
1939 transform_later(slow_ctrl);
1940 _igvn.hash_delete(_fallthroughproj);
1941 _fallthroughproj->disconnect_inputs(NULL);
1942 region->init_req(1, slow_ctrl);
1943 // region inputs are now complete
1944 transform_later(region);
1945 _igvn.subsume_node(_fallthroughproj, region);
1946
1947 Node *memproj = transform_later( new(C, 1) ProjNode(call, TypeFunc::Memory) );
1948 mem_phi->init_req(1, memproj );
1949 mem_phi->init_req(2, mem);
1950 transform_later(mem_phi);
1951 _igvn.hash_delete(_memproj_fallthrough);
1952 _igvn.subsume_node(_memproj_fallthrough, mem_phi);
1953 }
1954
1955 //------------------------------expand_macro_nodes----------------------
1956 // Returns true if a failure occurred.
1957 bool PhaseMacroExpand::expand_macro_nodes() {
1958 if (C->macro_count() == 0)
1959 return false;
1960 // First, attempt to eliminate locks
1961 bool progress = true;
1962 while (progress) {
1963 progress = false;
1964 for (int i = C->macro_count(); i > 0; i--) {
1965 Node * n = C->macro_node(i-1);
1966 bool success = false;
1967 debug_only(int old_macro_count = C->macro_count(););
1968 if (n->is_AbstractLock()) {
1969 success = eliminate_locking_node(n->as_AbstractLock());
1970 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) {
1971 _igvn.add_users_to_worklist(n);
1972 _igvn.hash_delete(n);
1973 _igvn.subsume_node(n, n->in(1));
1974 success = true;
1975 }
1976 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
1977 progress = progress || success;
1978 }
1979 }
1980 // Next, attempt to eliminate allocations
1981 progress = true;
1982 while (progress) {
1983 progress = false;
1984 for (int i = C->macro_count(); i > 0; i--) {
1985 Node * n = C->macro_node(i-1);
1986 bool success = false;
1987 debug_only(int old_macro_count = C->macro_count(););
1988 switch (n->class_id()) {
1989 case Node::Class_Allocate:
1990 case Node::Class_AllocateArray:
1991 success = eliminate_allocate_node(n->as_Allocate());
1992 break;
1993 case Node::Class_Lock:
1994 case Node::Class_Unlock:
1995 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
1996 break;
1997 default:
1998 assert(false, "unknown node type in macro list");
1999 }
2000 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2001 progress = progress || success;
2002 }
2003 }
2004 // Make sure expansion will not cause node limit to be exceeded.
2005 // Worst case is a macro node gets expanded into about 50 nodes.
2006 // Allow 50% more for optimization.
2007 if (C->check_node_count(C->macro_count() * 75, "out of nodes before macro expansion" ) )
2008 return true;
2009
2010 // expand "macro" nodes
2011 // nodes are removed from the macro list as they are processed
2012 while (C->macro_count() > 0) {
2013 int macro_count = C->macro_count();
2014 Node * n = C->macro_node(macro_count-1);
2015 assert(n->is_macro(), "only macro nodes expected here");
2016 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) {
2017 // node is unreachable, so don't try to expand it
2018 C->remove_macro_node(n);
2019 continue;
2020 }
2021 switch (n->class_id()) {
2022 case Node::Class_Allocate:
2023 expand_allocate(n->as_Allocate());
2024 break;
2025 case Node::Class_AllocateArray:
2026 expand_allocate_array(n->as_AllocateArray());
2027 break;
2028 case Node::Class_Lock:
2029 expand_lock_node(n->as_Lock());
2030 break;
2031 case Node::Class_Unlock:
2032 expand_unlock_node(n->as_Unlock());
2033 break;
2034 default:
2035 assert(false, "unknown node type in macro list");
2036 }
2037 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2038 if (C->failing()) return true;
2039 }
2040
2041 _igvn.set_delay_transform(false);
2042 _igvn.optimize();
2043 return false;
2044 }