1 /*
2 * Copyright 2005-2009 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->set_req(0, newcall->in(0)); // reset control edge
68 new_in = transform_later(new_in); // Register new node.
69 }
70 old_in = new_in;
71 }
72 newcall->add_req(old_in);
73 }
74
75 newcall->set_jvms(oldcall->jvms());
76 for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) {
77 jvms->set_map(newcall);
78 jvms->set_locoff(jvms->locoff()+jvms_adj);
79 jvms->set_stkoff(jvms->stkoff()+jvms_adj);
80 jvms->set_monoff(jvms->monoff()+jvms_adj);
81 jvms->set_scloff(jvms->scloff()+jvms_adj);
82 jvms->set_endoff(jvms->endoff()+jvms_adj);
83 }
84 }
85
86 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
87 Node* cmp;
88 if (mask != 0) {
89 Node* and_node = transform_later(new (C, 3) AndXNode(word, MakeConX(mask)));
90 cmp = transform_later(new (C, 3) CmpXNode(and_node, MakeConX(bits)));
91 } else {
92 cmp = word;
93 }
94 Node* bol = transform_later(new (C, 2) BoolNode(cmp, BoolTest::ne));
95 IfNode* iff = new (C, 2) IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
96 transform_later(iff);
97
98 // Fast path taken.
99 Node *fast_taken = transform_later( new (C, 1) IfFalseNode(iff) );
100
101 // Fast path not-taken, i.e. slow path
102 Node *slow_taken = transform_later( new (C, 1) IfTrueNode(iff) );
103
104 if (return_fast_path) {
105 region->init_req(edge, slow_taken); // Capture slow-control
106 return fast_taken;
107 } else {
108 region->init_req(edge, fast_taken); // Capture fast-control
109 return slow_taken;
110 }
111 }
112
113 //--------------------copy_predefined_input_for_runtime_call--------------------
114 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
115 // Set fixed predefined input arguments
116 call->init_req( TypeFunc::Control, ctrl );
117 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) );
118 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
119 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
120 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
121 }
122
123 //------------------------------make_slow_call---------------------------------
124 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) {
125
126 // Slow-path call
127 int size = slow_call_type->domain()->cnt();
128 CallNode *call = leaf_name
129 ? (CallNode*)new (C, size) CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
130 : (CallNode*)new (C, size) CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM );
131
132 // Slow path call has no side-effects, uses few values
133 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
134 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0);
135 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1);
136 copy_call_debug_info(oldcall, call);
137 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
138 _igvn.hash_delete(oldcall);
139 _igvn.subsume_node(oldcall, call);
140 transform_later(call);
141
142 return call;
143 }
144
145 void PhaseMacroExpand::extract_call_projections(CallNode *call) {
146 _fallthroughproj = NULL;
147 _fallthroughcatchproj = NULL;
148 _ioproj_fallthrough = NULL;
149 _ioproj_catchall = NULL;
150 _catchallcatchproj = NULL;
151 _memproj_fallthrough = NULL;
152 _memproj_catchall = NULL;
153 _resproj = NULL;
154 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) {
155 ProjNode *pn = call->fast_out(i)->as_Proj();
156 switch (pn->_con) {
157 case TypeFunc::Control:
158 {
159 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj
160 _fallthroughproj = pn;
161 DUIterator_Fast jmax, j = pn->fast_outs(jmax);
162 const Node *cn = pn->fast_out(j);
163 if (cn->is_Catch()) {
164 ProjNode *cpn = NULL;
165 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) {
166 cpn = cn->fast_out(k)->as_Proj();
167 assert(cpn->is_CatchProj(), "must be a CatchProjNode");
168 if (cpn->_con == CatchProjNode::fall_through_index)
169 _fallthroughcatchproj = cpn;
170 else {
171 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index.");
172 _catchallcatchproj = cpn;
173 }
174 }
175 }
176 break;
177 }
178 case TypeFunc::I_O:
179 if (pn->_is_io_use)
180 _ioproj_catchall = pn;
181 else
182 _ioproj_fallthrough = pn;
183 break;
184 case TypeFunc::Memory:
185 if (pn->_is_io_use)
186 _memproj_catchall = pn;
187 else
188 _memproj_fallthrough = pn;
189 break;
190 case TypeFunc::Parms:
191 _resproj = pn;
192 break;
193 default:
194 assert(false, "unexpected projection from allocation node.");
195 }
196 }
197
198 }
199
200 // Eliminate a card mark sequence. p2x is a ConvP2XNode
201 void PhaseMacroExpand::eliminate_card_mark(Node *p2x) {
202 assert(p2x->Opcode() == Op_CastP2X, "ConvP2XNode required");
203 Node *shift = p2x->unique_out();
204 Node *addp = shift->unique_out();
205 for (DUIterator_Last jmin, j = addp->last_outs(jmin); j >= jmin; --j) {
206 Node *st = addp->last_out(j);
207 assert(st->is_Store(), "store required");
208 _igvn.replace_node(st, st->in(MemNode::Memory));
209 }
210 }
211
212 // Search for a memory operation for the specified memory slice.
213 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
214 Node *orig_mem = mem;
215 Node *alloc_mem = alloc->in(TypeFunc::Memory);
216 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
217 while (true) {
218 if (mem == alloc_mem || mem == start_mem ) {
219 return mem; // hit one of our sentinels
220 } else if (mem->is_MergeMem()) {
221 mem = mem->as_MergeMem()->memory_at(alias_idx);
222 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
223 Node *in = mem->in(0);
224 // we can safely skip over safepoints, calls, locks and membars because we
225 // already know that the object is safe to eliminate.
226 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
227 return in;
228 } else if (in->is_Call()) {
229 CallNode *call = in->as_Call();
230 if (!call->may_modify(tinst, phase)) {
231 mem = call->in(TypeFunc::Memory);
232 }
233 mem = in->in(TypeFunc::Memory);
234 } else if (in->is_MemBar()) {
235 mem = in->in(TypeFunc::Memory);
236 } else {
237 assert(false, "unexpected projection");
238 }
239 } else if (mem->is_Store()) {
240 const TypePtr* atype = mem->as_Store()->adr_type();
241 int adr_idx = Compile::current()->get_alias_index(atype);
242 if (adr_idx == alias_idx) {
243 assert(atype->isa_oopptr(), "address type must be oopptr");
244 int adr_offset = atype->offset();
245 uint adr_iid = atype->is_oopptr()->instance_id();
246 // Array elements references have the same alias_idx
247 // but different offset and different instance_id.
248 if (adr_offset == offset && adr_iid == alloc->_idx)
249 return mem;
250 } else {
251 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
252 }
253 mem = mem->in(MemNode::Memory);
254 } else if (mem->Opcode() == Op_SCMemProj) {
255 assert(mem->in(0)->is_LoadStore(), "sanity");
256 const TypePtr* atype = mem->in(0)->in(MemNode::Address)->bottom_type()->is_ptr();
257 int adr_idx = Compile::current()->get_alias_index(atype);
258 if (adr_idx == alias_idx) {
259 assert(false, "Object is not scalar replaceable if a LoadStore node access its field");
260 return NULL;
261 }
262 mem = mem->in(0)->in(MemNode::Memory);
263 } else {
264 return mem;
265 }
266 assert(mem != orig_mem, "dead memory loop");
267 }
268 }
269
270 //
271 // Given a Memory Phi, compute a value Phi containing the values from stores
272 // on the input paths.
273 // Note: this function is recursive, its depth is limied by the "level" argument
274 // Returns the computed Phi, or NULL if it cannot compute it.
275 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) {
276 assert(mem->is_Phi(), "sanity");
277 int alias_idx = C->get_alias_index(adr_t);
278 int offset = adr_t->offset();
279 int instance_id = adr_t->instance_id();
280
281 // Check if an appropriate value phi already exists.
282 Node* region = mem->in(0);
283 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
284 Node* phi = region->fast_out(k);
285 if (phi->is_Phi() && phi != mem &&
286 phi->as_Phi()->is_same_inst_field(phi_type, instance_id, alias_idx, offset)) {
287 return phi;
288 }
289 }
290 // Check if an appropriate new value phi already exists.
291 Node* new_phi = NULL;
292 uint size = value_phis->size();
293 for (uint i=0; i < size; i++) {
294 if ( mem->_idx == value_phis->index_at(i) ) {
295 return value_phis->node_at(i);
296 }
297 }
298
299 if (level <= 0) {
300 return NULL; // Give up: phi tree too deep
301 }
302 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
303 Node *alloc_mem = alloc->in(TypeFunc::Memory);
304
305 uint length = mem->req();
306 GrowableArray <Node *> values(length, length, NULL);
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 transform_later(phi);
311 value_phis->push(phi, mem->_idx);
312
313 for (uint j = 1; j < length; j++) {
314 Node *in = mem->in(j);
315 if (in == NULL || in->is_top()) {
316 values.at_put(j, in);
317 } else {
318 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
319 if (val == start_mem || val == alloc_mem) {
320 // hit a sentinel, return appropriate 0 value
321 values.at_put(j, _igvn.zerocon(ft));
322 continue;
323 }
324 if (val->is_Initialize()) {
325 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
326 }
327 if (val == NULL) {
328 return NULL; // can't find a value on this path
329 }
330 if (val == mem) {
331 values.at_put(j, mem);
332 } else if (val->is_Store()) {
333 values.at_put(j, val->in(MemNode::ValueIn));
334 } else if(val->is_Proj() && val->in(0) == alloc) {
335 values.at_put(j, _igvn.zerocon(ft));
336 } else if (val->is_Phi()) {
337 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
338 if (val == NULL) {
339 return NULL;
340 }
341 values.at_put(j, val);
342 } else if (val->Opcode() == Op_SCMemProj) {
343 assert(val->in(0)->is_LoadStore(), "sanity");
344 assert(false, "Object is not scalar replaceable if a LoadStore node access its field");
345 return NULL;
346 } else {
347 #ifdef ASSERT
348 val->dump();
349 assert(false, "unknown node on this path");
350 #endif
351 return NULL; // unknown node on this path
352 }
353 }
354 }
355 // Set Phi's inputs
356 for (uint j = 1; j < length; j++) {
357 if (values.at(j) == mem) {
358 phi->init_req(j, phi);
359 } else {
360 phi->init_req(j, values.at(j));
361 }
362 }
363 return phi;
364 }
365
366 // Search the last value stored into the object's field.
367 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, Node *alloc) {
368 assert(adr_t->is_known_instance_field(), "instance required");
369 int instance_id = adr_t->instance_id();
370 assert((uint)instance_id == alloc->_idx, "wrong allocation");
371
372 int alias_idx = C->get_alias_index(adr_t);
373 int offset = adr_t->offset();
374 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
375 Node *alloc_ctrl = alloc->in(TypeFunc::Control);
376 Node *alloc_mem = alloc->in(TypeFunc::Memory);
377 Arena *a = Thread::current()->resource_area();
378 VectorSet visited(a);
379
380
381 bool done = sfpt_mem == alloc_mem;
382 Node *mem = sfpt_mem;
383 while (!done) {
384 if (visited.test_set(mem->_idx)) {
385 return NULL; // found a loop, give up
386 }
387 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
388 if (mem == start_mem || mem == alloc_mem) {
389 done = true; // hit a sentinel, return appropriate 0 value
390 } else if (mem->is_Initialize()) {
391 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
392 if (mem == NULL) {
393 done = true; // Something go wrong.
394 } else if (mem->is_Store()) {
395 const TypePtr* atype = mem->as_Store()->adr_type();
396 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
397 done = true;
398 }
399 } else if (mem->is_Store()) {
400 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
401 assert(atype != NULL, "address type must be oopptr");
402 assert(C->get_alias_index(atype) == alias_idx &&
403 atype->is_known_instance_field() && atype->offset() == offset &&
404 atype->instance_id() == instance_id, "store is correct memory slice");
405 done = true;
406 } else if (mem->is_Phi()) {
407 // try to find a phi's unique input
408 Node *unique_input = NULL;
409 Node *top = C->top();
410 for (uint i = 1; i < mem->req(); i++) {
411 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
412 if (n == NULL || n == top || n == mem) {
413 continue;
414 } else if (unique_input == NULL) {
415 unique_input = n;
416 } else if (unique_input != n) {
417 unique_input = top;
418 break;
419 }
420 }
421 if (unique_input != NULL && unique_input != top) {
422 mem = unique_input;
423 } else {
424 done = true;
425 }
426 } else {
427 assert(false, "unexpected node");
428 }
429 }
430 if (mem != NULL) {
431 if (mem == start_mem || mem == alloc_mem) {
432 // hit a sentinel, return appropriate 0 value
433 return _igvn.zerocon(ft);
434 } else if (mem->is_Store()) {
435 return mem->in(MemNode::ValueIn);
436 } else if (mem->is_Phi()) {
437 // attempt to produce a Phi reflecting the values on the input paths of the Phi
438 Node_Stack value_phis(a, 8);
439 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
440 if (phi != NULL) {
441 return phi;
442 } else {
443 // Kill all new Phis
444 while(value_phis.is_nonempty()) {
445 Node* n = value_phis.node();
446 _igvn.hash_delete(n);
447 _igvn.subsume_node(n, C->top());
448 value_phis.pop();
449 }
450 }
451 }
452 }
453 // Something go wrong.
454 return NULL;
455 }
456
457 // Check the possibility of scalar replacement.
458 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
459 // Scan the uses of the allocation to check for anything that would
460 // prevent us from eliminating it.
461 NOT_PRODUCT( const char* fail_eliminate = NULL; )
462 DEBUG_ONLY( Node* disq_node = NULL; )
463 bool can_eliminate = true;
464
465 Node* res = alloc->result_cast();
466 const TypeOopPtr* res_type = NULL;
467 if (res == NULL) {
468 // All users were eliminated.
469 } else if (!res->is_CheckCastPP()) {
470 alloc->_is_scalar_replaceable = false; // don't try again
471 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
472 can_eliminate = false;
473 } else {
474 res_type = _igvn.type(res)->isa_oopptr();
475 if (res_type == NULL) {
476 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
477 can_eliminate = false;
478 } else if (res_type->isa_aryptr()) {
479 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
480 if (length < 0) {
481 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
482 can_eliminate = false;
483 }
484 }
485 }
486
487 if (can_eliminate && res != NULL) {
488 for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
489 j < jmax && can_eliminate; j++) {
490 Node* use = res->fast_out(j);
491
492 if (use->is_AddP()) {
493 const TypePtr* addp_type = _igvn.type(use)->is_ptr();
494 int offset = addp_type->offset();
495
496 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
497 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";)
498 can_eliminate = false;
499 break;
500 }
501 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
502 k < kmax && can_eliminate; k++) {
503 Node* n = use->fast_out(k);
504 if (!n->is_Store() && n->Opcode() != Op_CastP2X) {
505 DEBUG_ONLY(disq_node = n;)
506 if (n->is_Load() || n->is_LoadStore()) {
507 NOT_PRODUCT(fail_eliminate = "Field load";)
508 } else {
509 NOT_PRODUCT(fail_eliminate = "Not store field referrence";)
510 }
511 can_eliminate = false;
512 }
513 }
514 } else if (use->is_SafePoint()) {
515 SafePointNode* sfpt = use->as_SafePoint();
516 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
517 // Object is passed as argument.
518 DEBUG_ONLY(disq_node = use;)
519 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
520 can_eliminate = false;
521 }
522 Node* sfptMem = sfpt->memory();
523 if (sfptMem == NULL || sfptMem->is_top()) {
524 DEBUG_ONLY(disq_node = use;)
525 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";)
526 can_eliminate = false;
527 } else {
528 safepoints.append_if_missing(sfpt);
529 }
530 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
531 if (use->is_Phi()) {
532 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
533 NOT_PRODUCT(fail_eliminate = "Object is return value";)
534 } else {
535 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
536 }
537 DEBUG_ONLY(disq_node = use;)
538 } else {
539 if (use->Opcode() == Op_Return) {
540 NOT_PRODUCT(fail_eliminate = "Object is return value";)
541 }else {
542 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
543 }
544 DEBUG_ONLY(disq_node = use;)
545 }
546 can_eliminate = false;
547 }
548 }
549 }
550
551 #ifndef PRODUCT
552 if (PrintEliminateAllocations) {
553 if (can_eliminate) {
554 tty->print("Scalar ");
555 if (res == NULL)
556 alloc->dump();
557 else
558 res->dump();
559 } else {
560 tty->print("NotScalar (%s)", fail_eliminate);
561 if (res == NULL)
562 alloc->dump();
563 else
564 res->dump();
565 #ifdef ASSERT
566 if (disq_node != NULL) {
567 tty->print(" >>>> ");
568 disq_node->dump();
569 }
570 #endif /*ASSERT*/
571 }
572 }
573 #endif
574 return can_eliminate;
575 }
576
577 // Do scalar replacement.
578 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
579 GrowableArray <SafePointNode *> safepoints_done;
580
581 ciKlass* klass = NULL;
582 ciInstanceKlass* iklass = NULL;
583 int nfields = 0;
584 int array_base;
585 int element_size;
586 BasicType basic_elem_type;
587 ciType* elem_type;
588
589 Node* res = alloc->result_cast();
590 const TypeOopPtr* res_type = NULL;
591 if (res != NULL) { // Could be NULL when there are no users
592 res_type = _igvn.type(res)->isa_oopptr();
593 }
594
595 if (res != NULL) {
596 klass = res_type->klass();
597 if (res_type->isa_instptr()) {
598 // find the fields of the class which will be needed for safepoint debug information
599 assert(klass->is_instance_klass(), "must be an instance klass.");
600 iklass = klass->as_instance_klass();
601 nfields = iklass->nof_nonstatic_fields();
602 } else {
603 // find the array's elements which will be needed for safepoint debug information
604 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
605 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass.");
606 elem_type = klass->as_array_klass()->element_type();
607 basic_elem_type = elem_type->basic_type();
608 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
609 element_size = type2aelembytes(basic_elem_type);
610 }
611 }
612 //
613 // Process the safepoint uses
614 //
615 while (safepoints.length() > 0) {
616 SafePointNode* sfpt = safepoints.pop();
617 Node* mem = sfpt->memory();
618 uint first_ind = sfpt->req();
619 SafePointScalarObjectNode* sobj = new (C, 1) SafePointScalarObjectNode(res_type,
620 #ifdef ASSERT
621 alloc,
622 #endif
623 first_ind, nfields);
624 sobj->init_req(0, sfpt->in(TypeFunc::Control));
625 transform_later(sobj);
626
627 // Scan object's fields adding an input to the safepoint for each field.
628 for (int j = 0; j < nfields; j++) {
629 intptr_t offset;
630 ciField* field = NULL;
631 if (iklass != NULL) {
632 field = iklass->nonstatic_field_at(j);
633 offset = field->offset();
634 elem_type = field->type();
635 basic_elem_type = field->layout_type();
636 } else {
637 offset = array_base + j * (intptr_t)element_size;
638 }
639
640 const Type *field_type;
641 // The next code is taken from Parse::do_get_xxx().
642 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) {
643 if (!elem_type->is_loaded()) {
644 field_type = TypeInstPtr::BOTTOM;
645 } else if (field != NULL && field->is_constant()) {
646 // This can happen if the constant oop is non-perm.
647 ciObject* con = field->constant_value().as_object();
648 // Do not "join" in the previous type; it doesn't add value,
649 // and may yield a vacuous result if the field is of interface type.
650 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
651 assert(field_type != NULL, "field singleton type must be consistent");
652 } else {
653 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
654 }
655 if (UseCompressedOops) {
656 field_type = field_type->make_narrowoop();
657 basic_elem_type = T_NARROWOOP;
658 }
659 } else {
660 field_type = Type::get_const_basic_type(basic_elem_type);
661 }
662
663 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr();
664
665 Node *field_val = value_from_mem(mem, basic_elem_type, field_type, field_addr_type, alloc);
666 if (field_val == NULL) {
667 // we weren't able to find a value for this field,
668 // give up on eliminating this allocation
669 alloc->_is_scalar_replaceable = false; // don't try again
670 // remove any extra entries we added to the safepoint
671 uint last = sfpt->req() - 1;
672 for (int k = 0; k < j; k++) {
673 sfpt->del_req(last--);
674 }
675 // rollback processed safepoints
676 while (safepoints_done.length() > 0) {
677 SafePointNode* sfpt_done = safepoints_done.pop();
678 // remove any extra entries we added to the safepoint
679 last = sfpt_done->req() - 1;
680 for (int k = 0; k < nfields; k++) {
681 sfpt_done->del_req(last--);
682 }
683 JVMState *jvms = sfpt_done->jvms();
684 jvms->set_endoff(sfpt_done->req());
685 // Now make a pass over the debug information replacing any references
686 // to SafePointScalarObjectNode with the allocated object.
687 int start = jvms->debug_start();
688 int end = jvms->debug_end();
689 for (int i = start; i < end; i++) {
690 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
691 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
692 if (scobj->first_index() == sfpt_done->req() &&
693 scobj->n_fields() == (uint)nfields) {
694 assert(scobj->alloc() == alloc, "sanity");
695 sfpt_done->set_req(i, res);
696 }
697 }
698 }
699 }
700 #ifndef PRODUCT
701 if (PrintEliminateAllocations) {
702 if (field != NULL) {
703 tty->print("=== At SafePoint node %d can't find value of Field: ",
704 sfpt->_idx);
705 field->print();
706 int field_idx = C->get_alias_index(field_addr_type);
707 tty->print(" (alias_idx=%d)", field_idx);
708 } else { // Array's element
709 tty->print("=== At SafePoint node %d can't find value of array element [%d]",
710 sfpt->_idx, j);
711 }
712 tty->print(", which prevents elimination of: ");
713 if (res == NULL)
714 alloc->dump();
715 else
716 res->dump();
717 }
718 #endif
719 return false;
720 }
721 if (UseCompressedOops && field_type->isa_narrowoop()) {
722 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
723 // to be able scalar replace the allocation.
724 if (field_val->is_EncodeP()) {
725 field_val = field_val->in(1);
726 } else {
727 field_val = transform_later(new (C, 2) DecodeNNode(field_val, field_val->bottom_type()->make_ptr()));
728 }
729 }
730 sfpt->add_req(field_val);
731 }
732 JVMState *jvms = sfpt->jvms();
733 jvms->set_endoff(sfpt->req());
734 // Now make a pass over the debug information replacing any references
735 // to the allocated object with "sobj"
736 int start = jvms->debug_start();
737 int end = jvms->debug_end();
738 for (int i = start; i < end; i++) {
739 if (sfpt->in(i) == res) {
740 sfpt->set_req(i, sobj);
741 }
742 }
743 safepoints_done.append_if_missing(sfpt); // keep it for rollback
744 }
745 return true;
746 }
747
748 // Process users of eliminated allocation.
749 void PhaseMacroExpand::process_users_of_allocation(AllocateNode *alloc) {
750 Node* res = alloc->result_cast();
751 if (res != NULL) {
752 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
753 Node *use = res->last_out(j);
754 uint oc1 = res->outcnt();
755
756 if (use->is_AddP()) {
757 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
758 Node *n = use->last_out(k);
759 uint oc2 = use->outcnt();
760 if (n->is_Store()) {
761 _igvn.replace_node(n, n->in(MemNode::Memory));
762 } else {
763 assert( n->Opcode() == Op_CastP2X, "CastP2X required");
764 eliminate_card_mark(n);
765 }
766 k -= (oc2 - use->outcnt());
767 }
768 } else {
769 assert( !use->is_SafePoint(), "safepoint uses must have been already elimiated");
770 assert( use->Opcode() == Op_CastP2X, "CastP2X required");
771 eliminate_card_mark(use);
772 }
773 j -= (oc1 - res->outcnt());
774 }
775 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
776 _igvn.remove_dead_node(res);
777 }
778
779 //
780 // Process other users of allocation's projections
781 //
782 if (_resproj != NULL && _resproj->outcnt() != 0) {
783 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) {
784 Node *use = _resproj->last_out(j);
785 uint oc1 = _resproj->outcnt();
786 if (use->is_Initialize()) {
787 // Eliminate Initialize node.
788 InitializeNode *init = use->as_Initialize();
789 assert(init->outcnt() <= 2, "only a control and memory projection expected");
790 Node *ctrl_proj = init->proj_out(TypeFunc::Control);
791 if (ctrl_proj != NULL) {
792 assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection");
793 _igvn.replace_node(ctrl_proj, _fallthroughcatchproj);
794 }
795 Node *mem_proj = init->proj_out(TypeFunc::Memory);
796 if (mem_proj != NULL) {
797 Node *mem = init->in(TypeFunc::Memory);
798 #ifdef ASSERT
799 if (mem->is_MergeMem()) {
800 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection");
801 } else {
802 assert(mem == _memproj_fallthrough, "allocation memory projection");
803 }
804 #endif
805 _igvn.replace_node(mem_proj, mem);
806 }
807 } else if (use->is_AddP()) {
808 // raw memory addresses used only by the initialization
809 _igvn.replace_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.replace_node(_fallthroughcatchproj, result_region);
1294 } else {
1295 ctrl = top();
1296 }
1297 Node *slow_result;
1298 if (_resproj == NULL) {
1299 // no uses of the allocation result
1300 slow_result = top();
1301 } else {
1302 slow_result = _resproj->clone();
1303 transform_later(slow_result);
1304 _igvn.replace_node(_resproj, result_phi_rawoop);
1305 }
1306
1307 // Plug slow-path into result merge point
1308 result_region ->init_req( slow_result_path, ctrl );
1309 result_phi_rawoop->init_req( slow_result_path, slow_result);
1310 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough );
1311 transform_later(result_region);
1312 transform_later(result_phi_rawoop);
1313 transform_later(result_phi_rawmem);
1314 transform_later(result_phi_i_o);
1315 // This completes all paths into the result merge point
1316 }
1317
1318
1319 // Helper for PhaseMacroExpand::expand_allocate_common.
1320 // Initializes the newly-allocated storage.
1321 Node*
1322 PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1323 Node* control, Node* rawmem, Node* object,
1324 Node* klass_node, Node* length,
1325 Node* size_in_bytes) {
1326 InitializeNode* init = alloc->initialization();
1327 // Store the klass & mark bits
1328 Node* mark_node = NULL;
1329 // For now only enable fast locking for non-array types
1330 if (UseBiasedLocking && (length == NULL)) {
1331 mark_node = make_load(NULL, rawmem, klass_node, Klass::prototype_header_offset_in_bytes() + sizeof(oopDesc), TypeRawPtr::BOTTOM, T_ADDRESS);
1332 } else {
1333 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype()));
1334 }
1335 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
1336
1337 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_OBJECT);
1338 int header_size = alloc->minimum_header_size(); // conservatively small
1339
1340 // Array length
1341 if (length != NULL) { // Arrays need length field
1342 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1343 // conservatively small header size:
1344 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1345 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1346 if (k->is_array_klass()) // we know the exact header size in most cases:
1347 header_size = Klass::layout_helper_header_size(k->layout_helper());
1348 }
1349
1350 // Clear the object body, if necessary.
1351 if (init == NULL) {
1352 // The init has somehow disappeared; be cautious and clear everything.
1353 //
1354 // This can happen if a node is allocated but an uncommon trap occurs
1355 // immediately. In this case, the Initialize gets associated with the
1356 // trap, and may be placed in a different (outer) loop, if the Allocate
1357 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1358 // there can be two Allocates to one Initialize. The answer in all these
1359 // edge cases is safety first. It is always safe to clear immediately
1360 // within an Allocate, and then (maybe or maybe not) clear some more later.
1361 if (!ZeroTLAB)
1362 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1363 header_size, size_in_bytes,
1364 &_igvn);
1365 } else {
1366 if (!init->is_complete()) {
1367 // Try to win by zeroing only what the init does not store.
1368 // We can also try to do some peephole optimizations,
1369 // such as combining some adjacent subword stores.
1370 rawmem = init->complete_stores(control, rawmem, object,
1371 header_size, size_in_bytes, &_igvn);
1372 }
1373 // We have no more use for this link, since the AllocateNode goes away:
1374 init->set_req(InitializeNode::RawAddress, top());
1375 // (If we keep the link, it just confuses the register allocator,
1376 // who thinks he sees a real use of the address by the membar.)
1377 }
1378
1379 return rawmem;
1380 }
1381
1382 // Generate prefetch instructions for next allocations.
1383 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
1384 Node*& contended_phi_rawmem,
1385 Node* old_eden_top, Node* new_eden_top,
1386 Node* length) {
1387 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
1388 // Generate prefetch allocation with watermark check.
1389 // As an allocation hits the watermark, we will prefetch starting
1390 // at a "distance" away from watermark.
1391 enum { fall_in_path = 1, pf_path = 2 };
1392
1393 Node *pf_region = new (C, 3) RegionNode(3);
1394 Node *pf_phi_rawmem = new (C, 3) PhiNode( pf_region, Type::MEMORY,
1395 TypeRawPtr::BOTTOM );
1396 // I/O is used for Prefetch
1397 Node *pf_phi_abio = new (C, 3) PhiNode( pf_region, Type::ABIO );
1398
1399 Node *thread = new (C, 1) ThreadLocalNode();
1400 transform_later(thread);
1401
1402 Node *eden_pf_adr = new (C, 4) AddPNode( top()/*not oop*/, thread,
1403 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
1404 transform_later(eden_pf_adr);
1405
1406 Node *old_pf_wm = new (C, 3) LoadPNode( needgc_false,
1407 contended_phi_rawmem, eden_pf_adr,
1408 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM );
1409 transform_later(old_pf_wm);
1410
1411 // check against new_eden_top
1412 Node *need_pf_cmp = new (C, 3) CmpPNode( new_eden_top, old_pf_wm );
1413 transform_later(need_pf_cmp);
1414 Node *need_pf_bol = new (C, 2) BoolNode( need_pf_cmp, BoolTest::ge );
1415 transform_later(need_pf_bol);
1416 IfNode *need_pf_iff = new (C, 2) IfNode( needgc_false, need_pf_bol,
1417 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1418 transform_later(need_pf_iff);
1419
1420 // true node, add prefetchdistance
1421 Node *need_pf_true = new (C, 1) IfTrueNode( need_pf_iff );
1422 transform_later(need_pf_true);
1423
1424 Node *need_pf_false = new (C, 1) IfFalseNode( need_pf_iff );
1425 transform_later(need_pf_false);
1426
1427 Node *new_pf_wmt = new (C, 4) AddPNode( top(), old_pf_wm,
1428 _igvn.MakeConX(AllocatePrefetchDistance) );
1429 transform_later(new_pf_wmt );
1430 new_pf_wmt->set_req(0, need_pf_true);
1431
1432 Node *store_new_wmt = new (C, 4) StorePNode( need_pf_true,
1433 contended_phi_rawmem, eden_pf_adr,
1434 TypeRawPtr::BOTTOM, new_pf_wmt );
1435 transform_later(store_new_wmt);
1436
1437 // adding prefetches
1438 pf_phi_abio->init_req( fall_in_path, i_o );
1439
1440 Node *prefetch_adr;
1441 Node *prefetch;
1442 uint lines = AllocatePrefetchDistance / AllocatePrefetchStepSize;
1443 uint step_size = AllocatePrefetchStepSize;
1444 uint distance = 0;
1445
1446 for ( uint i = 0; i < lines; i++ ) {
1447 prefetch_adr = new (C, 4) AddPNode( old_pf_wm, new_pf_wmt,
1448 _igvn.MakeConX(distance) );
1449 transform_later(prefetch_adr);
1450 prefetch = new (C, 3) PrefetchWriteNode( i_o, prefetch_adr );
1451 transform_later(prefetch);
1452 distance += step_size;
1453 i_o = prefetch;
1454 }
1455 pf_phi_abio->set_req( pf_path, i_o );
1456
1457 pf_region->init_req( fall_in_path, need_pf_false );
1458 pf_region->init_req( pf_path, need_pf_true );
1459
1460 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
1461 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
1462
1463 transform_later(pf_region);
1464 transform_later(pf_phi_rawmem);
1465 transform_later(pf_phi_abio);
1466
1467 needgc_false = pf_region;
1468 contended_phi_rawmem = pf_phi_rawmem;
1469 i_o = pf_phi_abio;
1470 } else if( AllocatePrefetchStyle > 0 ) {
1471 // Insert a prefetch for each allocation only on the fast-path
1472 Node *prefetch_adr;
1473 Node *prefetch;
1474 // Generate several prefetch instructions only for arrays.
1475 uint lines = (length != NULL) ? AllocatePrefetchLines : 1;
1476 uint step_size = AllocatePrefetchStepSize;
1477 uint distance = AllocatePrefetchDistance;
1478 for ( uint i = 0; i < lines; i++ ) {
1479 prefetch_adr = new (C, 4) AddPNode( old_eden_top, new_eden_top,
1480 _igvn.MakeConX(distance) );
1481 transform_later(prefetch_adr);
1482 prefetch = new (C, 3) PrefetchWriteNode( i_o, prefetch_adr );
1483 // Do not let it float too high, since if eden_top == eden_end,
1484 // both might be null.
1485 if( i == 0 ) { // Set control for first prefetch, next follows it
1486 prefetch->init_req(0, needgc_false);
1487 }
1488 transform_later(prefetch);
1489 distance += step_size;
1490 i_o = prefetch;
1491 }
1492 }
1493 return i_o;
1494 }
1495
1496
1497 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
1498 expand_allocate_common(alloc, NULL,
1499 OptoRuntime::new_instance_Type(),
1500 OptoRuntime::new_instance_Java());
1501 }
1502
1503 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
1504 Node* length = alloc->in(AllocateNode::ALength);
1505 expand_allocate_common(alloc, length,
1506 OptoRuntime::new_array_Type(),
1507 OptoRuntime::new_array_Java());
1508 }
1509
1510
1511 // we have determined that this lock/unlock can be eliminated, we simply
1512 // eliminate the node without expanding it.
1513 //
1514 // Note: The membar's associated with the lock/unlock are currently not
1515 // eliminated. This should be investigated as a future enhancement.
1516 //
1517 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
1518
1519 if (!alock->is_eliminated()) {
1520 return false;
1521 }
1522 if (alock->is_Lock() && !alock->is_coarsened()) {
1523 // Create new "eliminated" BoxLock node and use it
1524 // in monitor debug info for the same object.
1525 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
1526 Node* obj = alock->obj_node();
1527 if (!oldbox->is_eliminated()) {
1528 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
1529 newbox->set_eliminated();
1530 transform_later(newbox);
1531 // Replace old box node with new box for all users
1532 // of the same object.
1533 for (uint i = 0; i < oldbox->outcnt();) {
1534
1535 bool next_edge = true;
1536 Node* u = oldbox->raw_out(i);
1537 if (u == alock) {
1538 i++;
1539 continue; // It will be removed below
1540 }
1541 if (u->is_Lock() &&
1542 u->as_Lock()->obj_node() == obj &&
1543 // oldbox could be referenced in debug info also
1544 u->as_Lock()->box_node() == oldbox) {
1545 assert(u->as_Lock()->is_eliminated(), "sanity");
1546 _igvn.hash_delete(u);
1547 u->set_req(TypeFunc::Parms + 1, newbox);
1548 next_edge = false;
1549 #ifdef ASSERT
1550 } else if (u->is_Unlock() && u->as_Unlock()->obj_node() == obj) {
1551 assert(u->as_Unlock()->is_eliminated(), "sanity");
1552 #endif
1553 }
1554 // Replace old box in monitor debug info.
1555 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
1556 SafePointNode* sfn = u->as_SafePoint();
1557 JVMState* youngest_jvms = sfn->jvms();
1558 int max_depth = youngest_jvms->depth();
1559 for (int depth = 1; depth <= max_depth; depth++) {
1560 JVMState* jvms = youngest_jvms->of_depth(depth);
1561 int num_mon = jvms->nof_monitors();
1562 // Loop over monitors
1563 for (int idx = 0; idx < num_mon; idx++) {
1564 Node* obj_node = sfn->monitor_obj(jvms, idx);
1565 Node* box_node = sfn->monitor_box(jvms, idx);
1566 if (box_node == oldbox && obj_node == obj) {
1567 int j = jvms->monitor_box_offset(idx);
1568 _igvn.hash_delete(u);
1569 u->set_req(j, newbox);
1570 next_edge = false;
1571 }
1572 } // for (int idx = 0;
1573 } // for (int depth = 1;
1574 } // if (u->is_SafePoint()
1575 if (next_edge) i++;
1576 } // for (uint i = 0; i < oldbox->outcnt();)
1577 } // if (!oldbox->is_eliminated())
1578 } // if (alock->is_Lock() && !lock->is_coarsened())
1579
1580 #ifndef PRODUCT
1581 if (PrintEliminateLocks) {
1582 if (alock->is_Lock()) {
1583 tty->print_cr("++++ Eliminating: %d Lock", alock->_idx);
1584 } else {
1585 tty->print_cr("++++ Eliminating: %d Unlock", alock->_idx);
1586 }
1587 }
1588 #endif
1589
1590 Node* mem = alock->in(TypeFunc::Memory);
1591 Node* ctrl = alock->in(TypeFunc::Control);
1592
1593 extract_call_projections(alock);
1594 // There are 2 projections from the lock. The lock node will
1595 // be deleted when its last use is subsumed below.
1596 assert(alock->outcnt() == 2 &&
1597 _fallthroughproj != NULL &&
1598 _memproj_fallthrough != NULL,
1599 "Unexpected projections from Lock/Unlock");
1600
1601 Node* fallthroughproj = _fallthroughproj;
1602 Node* memproj_fallthrough = _memproj_fallthrough;
1603
1604 // The memory projection from a lock/unlock is RawMem
1605 // The input to a Lock is merged memory, so extract its RawMem input
1606 // (unless the MergeMem has been optimized away.)
1607 if (alock->is_Lock()) {
1608 // Seach for MemBarAcquire node and delete it also.
1609 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
1610 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquire, "");
1611 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
1612 Node* memproj = membar->proj_out(TypeFunc::Memory);
1613 _igvn.replace_node(ctrlproj, fallthroughproj);
1614 _igvn.replace_node(memproj, memproj_fallthrough);
1615
1616 // Delete FastLock node also if this Lock node is unique user
1617 // (a loop peeling may clone a Lock node).
1618 Node* flock = alock->as_Lock()->fastlock_node();
1619 if (flock->outcnt() == 1) {
1620 assert(flock->unique_out() == alock, "sanity");
1621 _igvn.replace_node(flock, top());
1622 }
1623 }
1624
1625 // Seach for MemBarRelease node and delete it also.
1626 if (alock->is_Unlock() && ctrl != NULL && ctrl->is_Proj() &&
1627 ctrl->in(0)->is_MemBar()) {
1628 MemBarNode* membar = ctrl->in(0)->as_MemBar();
1629 assert(membar->Opcode() == Op_MemBarRelease &&
1630 mem->is_Proj() && membar == mem->in(0), "");
1631 _igvn.replace_node(fallthroughproj, ctrl);
1632 _igvn.replace_node(memproj_fallthrough, mem);
1633 fallthroughproj = ctrl;
1634 memproj_fallthrough = mem;
1635 ctrl = membar->in(TypeFunc::Control);
1636 mem = membar->in(TypeFunc::Memory);
1637 }
1638
1639 _igvn.replace_node(fallthroughproj, ctrl);
1640 _igvn.replace_node(memproj_fallthrough, mem);
1641 return true;
1642 }
1643
1644
1645 //------------------------------expand_lock_node----------------------
1646 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
1647
1648 Node* ctrl = lock->in(TypeFunc::Control);
1649 Node* mem = lock->in(TypeFunc::Memory);
1650 Node* obj = lock->obj_node();
1651 Node* box = lock->box_node();
1652 Node* flock = lock->fastlock_node();
1653
1654 // Make the merge point
1655 Node *region;
1656 Node *mem_phi;
1657 Node *slow_path;
1658
1659 if (UseOptoBiasInlining) {
1660 /*
1661 * See the full description in MacroAssembler::biased_locking_enter().
1662 *
1663 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) {
1664 * // The object is biased.
1665 * proto_node = klass->prototype_header;
1666 * o_node = thread | proto_node;
1667 * x_node = o_node ^ mark_word;
1668 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ?
1669 * // Done.
1670 * } else {
1671 * if( (x_node & biased_lock_mask) != 0 ) {
1672 * // The klass's prototype header is no longer biased.
1673 * cas(&mark_word, mark_word, proto_node)
1674 * goto cas_lock;
1675 * } else {
1676 * // The klass's prototype header is still biased.
1677 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch?
1678 * old = mark_word;
1679 * new = o_node;
1680 * } else {
1681 * // Different thread or anonymous biased.
1682 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask);
1683 * new = thread | old;
1684 * }
1685 * // Try to rebias.
1686 * if( cas(&mark_word, old, new) == 0 ) {
1687 * // Done.
1688 * } else {
1689 * goto slow_path; // Failed.
1690 * }
1691 * }
1692 * }
1693 * } else {
1694 * // The object is not biased.
1695 * cas_lock:
1696 * if( FastLock(obj) == 0 ) {
1697 * // Done.
1698 * } else {
1699 * slow_path:
1700 * OptoRuntime::complete_monitor_locking_Java(obj);
1701 * }
1702 * }
1703 */
1704
1705 region = new (C, 5) RegionNode(5);
1706 // create a Phi for the memory state
1707 mem_phi = new (C, 5) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
1708
1709 Node* fast_lock_region = new (C, 3) RegionNode(3);
1710 Node* fast_lock_mem_phi = new (C, 3) PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1711
1712 // First, check mark word for the biased lock pattern.
1713 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
1714
1715 // Get fast path - mark word has the biased lock pattern.
1716 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node,
1717 markOopDesc::biased_lock_mask_in_place,
1718 markOopDesc::biased_lock_pattern, true);
1719 // fast_lock_region->in(1) is set to slow path.
1720 fast_lock_mem_phi->init_req(1, mem);
1721
1722 // Now check that the lock is biased to the current thread and has
1723 // the same epoch and bias as Klass::_prototype_header.
1724
1725 // Special-case a fresh allocation to avoid building nodes:
1726 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn);
1727 if (klass_node == NULL) {
1728 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1729 klass_node = transform_later( LoadKlassNode::make(_igvn, mem, k_adr, _igvn.type(k_adr)->is_ptr()) );
1730 #ifdef _LP64
1731 if (UseCompressedOops && klass_node->is_DecodeN()) {
1732 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity");
1733 klass_node->in(1)->init_req(0, ctrl);
1734 } else
1735 #endif
1736 klass_node->init_req(0, ctrl);
1737 }
1738 Node *proto_node = make_load(ctrl, mem, klass_node, Klass::prototype_header_offset_in_bytes() + sizeof(oopDesc), TypeX_X, TypeX_X->basic_type());
1739
1740 Node* thread = transform_later(new (C, 1) ThreadLocalNode());
1741 Node* cast_thread = transform_later(new (C, 2) CastP2XNode(ctrl, thread));
1742 Node* o_node = transform_later(new (C, 3) OrXNode(cast_thread, proto_node));
1743 Node* x_node = transform_later(new (C, 3) XorXNode(o_node, mark_node));
1744
1745 // Get slow path - mark word does NOT match the value.
1746 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node,
1747 (~markOopDesc::age_mask_in_place), 0);
1748 // region->in(3) is set to fast path - the object is biased to the current thread.
1749 mem_phi->init_req(3, mem);
1750
1751
1752 // Mark word does NOT match the value (thread | Klass::_prototype_header).
1753
1754
1755 // First, check biased pattern.
1756 // Get fast path - _prototype_header has the same biased lock pattern.
1757 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node,
1758 markOopDesc::biased_lock_mask_in_place, 0, true);
1759
1760 not_biased_ctrl = fast_lock_region->in(2); // Slow path
1761 // fast_lock_region->in(2) - the prototype header is no longer biased
1762 // and we have to revoke the bias on this object.
1763 // We are going to try to reset the mark of this object to the prototype
1764 // value and fall through to the CAS-based locking scheme.
1765 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
1766 Node* cas = new (C, 5) StoreXConditionalNode(not_biased_ctrl, mem, adr,
1767 proto_node, mark_node);
1768 transform_later(cas);
1769 Node* proj = transform_later( new (C, 1) SCMemProjNode(cas));
1770 fast_lock_mem_phi->init_req(2, proj);
1771
1772
1773 // Second, check epoch bits.
1774 Node* rebiased_region = new (C, 3) RegionNode(3);
1775 Node* old_phi = new (C, 3) PhiNode( rebiased_region, TypeX_X);
1776 Node* new_phi = new (C, 3) PhiNode( rebiased_region, TypeX_X);
1777
1778 // Get slow path - mark word does NOT match epoch bits.
1779 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node,
1780 markOopDesc::epoch_mask_in_place, 0);
1781 // The epoch of the current bias is not valid, attempt to rebias the object
1782 // toward the current thread.
1783 rebiased_region->init_req(2, epoch_ctrl);
1784 old_phi->init_req(2, mark_node);
1785 new_phi->init_req(2, o_node);
1786
1787 // rebiased_region->in(1) is set to fast path.
1788 // The epoch of the current bias is still valid but we know
1789 // nothing about the owner; it might be set or it might be clear.
1790 Node* cmask = MakeConX(markOopDesc::biased_lock_mask_in_place |
1791 markOopDesc::age_mask_in_place |
1792 markOopDesc::epoch_mask_in_place);
1793 Node* old = transform_later(new (C, 3) AndXNode(mark_node, cmask));
1794 cast_thread = transform_later(new (C, 2) CastP2XNode(ctrl, thread));
1795 Node* new_mark = transform_later(new (C, 3) OrXNode(cast_thread, old));
1796 old_phi->init_req(1, old);
1797 new_phi->init_req(1, new_mark);
1798
1799 transform_later(rebiased_region);
1800 transform_later(old_phi);
1801 transform_later(new_phi);
1802
1803 // Try to acquire the bias of the object using an atomic operation.
1804 // If this fails we will go in to the runtime to revoke the object's bias.
1805 cas = new (C, 5) StoreXConditionalNode(rebiased_region, mem, adr,
1806 new_phi, old_phi);
1807 transform_later(cas);
1808 proj = transform_later( new (C, 1) SCMemProjNode(cas));
1809
1810 // Get slow path - Failed to CAS.
1811 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0);
1812 mem_phi->init_req(4, proj);
1813 // region->in(4) is set to fast path - the object is rebiased to the current thread.
1814
1815 // Failed to CAS.
1816 slow_path = new (C, 3) RegionNode(3);
1817 Node *slow_mem = new (C, 3) PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM);
1818
1819 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control
1820 slow_mem->init_req(1, proj);
1821
1822 // Call CAS-based locking scheme (FastLock node).
1823
1824 transform_later(fast_lock_region);
1825 transform_later(fast_lock_mem_phi);
1826
1827 // Get slow path - FastLock failed to lock the object.
1828 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0);
1829 mem_phi->init_req(2, fast_lock_mem_phi);
1830 // region->in(2) is set to fast path - the object is locked to the current thread.
1831
1832 slow_path->init_req(2, ctrl); // Capture slow-control
1833 slow_mem->init_req(2, fast_lock_mem_phi);
1834
1835 transform_later(slow_path);
1836 transform_later(slow_mem);
1837 // Reset lock's memory edge.
1838 lock->set_req(TypeFunc::Memory, slow_mem);
1839
1840 } else {
1841 region = new (C, 3) RegionNode(3);
1842 // create a Phi for the memory state
1843 mem_phi = new (C, 3) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
1844
1845 // Optimize test; set region slot 2
1846 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
1847 mem_phi->init_req(2, mem);
1848 }
1849
1850 // Make slow path call
1851 CallNode *call = make_slow_call( (CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(), OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path, obj, box );
1852
1853 extract_call_projections(call);
1854
1855 // Slow path can only throw asynchronous exceptions, which are always
1856 // de-opted. So the compiler thinks the slow-call can never throw an
1857 // exception. If it DOES throw an exception we would need the debug
1858 // info removed first (since if it throws there is no monitor).
1859 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
1860 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
1861
1862 // Capture slow path
1863 // disconnect fall-through projection from call and create a new one
1864 // hook up users of fall-through projection to region
1865 Node *slow_ctrl = _fallthroughproj->clone();
1866 transform_later(slow_ctrl);
1867 _igvn.hash_delete(_fallthroughproj);
1868 _fallthroughproj->disconnect_inputs(NULL);
1869 region->init_req(1, slow_ctrl);
1870 // region inputs are now complete
1871 transform_later(region);
1872 _igvn.replace_node(_fallthroughproj, region);
1873
1874 Node *memproj = transform_later( new(C, 1) ProjNode(call, TypeFunc::Memory) );
1875 mem_phi->init_req(1, memproj );
1876 transform_later(mem_phi);
1877 _igvn.replace_node(_memproj_fallthrough, mem_phi);
1878 }
1879
1880 //------------------------------expand_unlock_node----------------------
1881 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
1882
1883 Node* ctrl = unlock->in(TypeFunc::Control);
1884 Node* mem = unlock->in(TypeFunc::Memory);
1885 Node* obj = unlock->obj_node();
1886 Node* box = unlock->box_node();
1887
1888 // No need for a null check on unlock
1889
1890 // Make the merge point
1891 Node *region;
1892 Node *mem_phi;
1893
1894 if (UseOptoBiasInlining) {
1895 // Check for biased locking unlock case, which is a no-op.
1896 // See the full description in MacroAssembler::biased_locking_exit().
1897 region = new (C, 4) RegionNode(4);
1898 // create a Phi for the memory state
1899 mem_phi = new (C, 4) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
1900 mem_phi->init_req(3, mem);
1901
1902 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
1903 ctrl = opt_bits_test(ctrl, region, 3, mark_node,
1904 markOopDesc::biased_lock_mask_in_place,
1905 markOopDesc::biased_lock_pattern);
1906 } else {
1907 region = new (C, 3) RegionNode(3);
1908 // create a Phi for the memory state
1909 mem_phi = new (C, 3) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
1910 }
1911
1912 FastUnlockNode *funlock = new (C, 3) FastUnlockNode( ctrl, obj, box );
1913 funlock = transform_later( funlock )->as_FastUnlock();
1914 // Optimize test; set region slot 2
1915 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
1916
1917 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 );
1918
1919 extract_call_projections(call);
1920
1921 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
1922 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
1923
1924 // No exceptions for unlocking
1925 // Capture slow path
1926 // disconnect fall-through projection from call and create a new one
1927 // hook up users of fall-through projection to region
1928 Node *slow_ctrl = _fallthroughproj->clone();
1929 transform_later(slow_ctrl);
1930 _igvn.hash_delete(_fallthroughproj);
1931 _fallthroughproj->disconnect_inputs(NULL);
1932 region->init_req(1, slow_ctrl);
1933 // region inputs are now complete
1934 transform_later(region);
1935 _igvn.replace_node(_fallthroughproj, region);
1936
1937 Node *memproj = transform_later( new(C, 1) ProjNode(call, TypeFunc::Memory) );
1938 mem_phi->init_req(1, memproj );
1939 mem_phi->init_req(2, mem);
1940 transform_later(mem_phi);
1941 _igvn.replace_node(_memproj_fallthrough, mem_phi);
1942 }
1943
1944 //------------------------------expand_macro_nodes----------------------
1945 // Returns true if a failure occurred.
1946 bool PhaseMacroExpand::expand_macro_nodes() {
1947 if (C->macro_count() == 0)
1948 return false;
1949 // First, attempt to eliminate locks
1950 bool progress = true;
1951 while (progress) {
1952 progress = false;
1953 for (int i = C->macro_count(); i > 0; i--) {
1954 Node * n = C->macro_node(i-1);
1955 bool success = false;
1956 debug_only(int old_macro_count = C->macro_count(););
1957 if (n->is_AbstractLock()) {
1958 success = eliminate_locking_node(n->as_AbstractLock());
1959 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) {
1960 _igvn.replace_node(n, n->in(1));
1961 success = true;
1962 }
1963 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
1964 progress = progress || success;
1965 }
1966 }
1967 // Next, attempt to eliminate allocations
1968 progress = true;
1969 while (progress) {
1970 progress = false;
1971 for (int i = C->macro_count(); i > 0; i--) {
1972 Node * n = C->macro_node(i-1);
1973 bool success = false;
1974 debug_only(int old_macro_count = C->macro_count(););
1975 switch (n->class_id()) {
1976 case Node::Class_Allocate:
1977 case Node::Class_AllocateArray:
1978 success = eliminate_allocate_node(n->as_Allocate());
1979 break;
1980 case Node::Class_Lock:
1981 case Node::Class_Unlock:
1982 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
1983 break;
1984 default:
1985 assert(false, "unknown node type in macro list");
1986 }
1987 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
1988 progress = progress || success;
1989 }
1990 }
1991 // Make sure expansion will not cause node limit to be exceeded.
1992 // Worst case is a macro node gets expanded into about 50 nodes.
1993 // Allow 50% more for optimization.
1994 if (C->check_node_count(C->macro_count() * 75, "out of nodes before macro expansion" ) )
1995 return true;
1996
1997 // expand "macro" nodes
1998 // nodes are removed from the macro list as they are processed
1999 while (C->macro_count() > 0) {
2000 int macro_count = C->macro_count();
2001 Node * n = C->macro_node(macro_count-1);
2002 assert(n->is_macro(), "only macro nodes expected here");
2003 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) {
2004 // node is unreachable, so don't try to expand it
2005 C->remove_macro_node(n);
2006 continue;
2007 }
2008 switch (n->class_id()) {
2009 case Node::Class_Allocate:
2010 expand_allocate(n->as_Allocate());
2011 break;
2012 case Node::Class_AllocateArray:
2013 expand_allocate_array(n->as_AllocateArray());
2014 break;
2015 case Node::Class_Lock:
2016 expand_lock_node(n->as_Lock());
2017 break;
2018 case Node::Class_Unlock:
2019 expand_unlock_node(n->as_Unlock());
2020 break;
2021 default:
2022 assert(false, "unknown node type in macro list");
2023 }
2024 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2025 if (C->failing()) return true;
2026 }
2027
2028 _igvn.set_delay_transform(false);
2029 _igvn.optimize();
2030 return false;
2031 }