Fri, 16 Oct 2009 02:05:46 -0700
6888898: CMS: ReduceInitialCardMarks unsafe in the presence of cms precleaning
6889757: G1: enable card mark elision for initializing writes from compiled code (ReduceInitialCardMarks)
Summary: Defer the (compiler-elided) card-mark upon a slow-path allocation until after the store and before the next subsequent safepoint; G1 now answers yes to can_elide_tlab_write_barriers().
Reviewed-by: jcoomes, kvn, never
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 */
25 #include "incls/_precompiled.incl"
26 #include "incls/_macro.cpp.incl"
29 //
30 // Replace any references to "oldref" in inputs to "use" with "newref".
31 // Returns the number of replacements made.
32 //
33 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
34 int nreplacements = 0;
35 uint req = use->req();
36 for (uint j = 0; j < use->len(); j++) {
37 Node *uin = use->in(j);
38 if (uin == oldref) {
39 if (j < req)
40 use->set_req(j, newref);
41 else
42 use->set_prec(j, newref);
43 nreplacements++;
44 } else if (j >= req && uin == NULL) {
45 break;
46 }
47 }
48 return nreplacements;
49 }
51 void PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) {
52 // Copy debug information and adjust JVMState information
53 uint old_dbg_start = oldcall->tf()->domain()->cnt();
54 uint new_dbg_start = newcall->tf()->domain()->cnt();
55 int jvms_adj = new_dbg_start - old_dbg_start;
56 assert (new_dbg_start == newcall->req(), "argument count mismatch");
58 Dict* sosn_map = new Dict(cmpkey,hashkey);
59 for (uint i = old_dbg_start; i < oldcall->req(); i++) {
60 Node* old_in = oldcall->in(i);
61 // Clone old SafePointScalarObjectNodes, adjusting their field contents.
62 if (old_in != 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 }
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 }
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);
98 // Fast path taken.
99 Node *fast_taken = transform_later( new (C, 1) IfFalseNode(iff) );
101 // Fast path not-taken, i.e. slow path
102 Node *slow_taken = transform_later( new (C, 1) IfTrueNode(iff) );
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 }
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 }
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) {
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 );
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);
142 return call;
143 }
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 }
198 }
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 if (!UseG1GC) {
204 // vanilla/CMS post barrier
205 Node *shift = p2x->unique_out();
206 Node *addp = shift->unique_out();
207 for (DUIterator_Last jmin, j = addp->last_outs(jmin); j >= jmin; --j) {
208 Node *st = addp->last_out(j);
209 assert(st->is_Store(), "store required");
210 _igvn.replace_node(st, st->in(MemNode::Memory));
211 }
212 } else {
213 // G1 pre/post barriers
214 assert(p2x->outcnt() == 2, "expects 2 users: Xor and URShift nodes");
215 // It could be only one user, URShift node, in Object.clone() instrinsic
216 // but the new allocation is passed to arraycopy stub and it could not
217 // be scalar replaced. So we don't check the case.
219 // Remove G1 post barrier.
221 // Search for CastP2X->Xor->URShift->Cmp path which
222 // checks if the store done to a different from the value's region.
223 // And replace Cmp with #0 (false) to collapse G1 post barrier.
224 Node* xorx = NULL;
225 for (DUIterator_Fast imax, i = p2x->fast_outs(imax); i < imax; i++) {
226 Node* u = p2x->fast_out(i);
227 if (u->Opcode() == Op_XorX) {
228 xorx = u;
229 break;
230 }
231 }
232 assert(xorx != NULL, "missing G1 post barrier");
233 Node* shift = xorx->unique_out();
234 Node* cmpx = shift->unique_out();
235 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() &&
236 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne,
237 "missing region check in G1 post barrier");
238 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ));
240 // Remove G1 pre barrier.
242 // Search "if (marking != 0)" check and set it to "false".
243 Node* this_region = p2x->in(0);
244 assert(this_region != NULL, "");
245 // There is no G1 pre barrier if previous stored value is NULL
246 // (for example, after initialization).
247 if (this_region->is_Region() && this_region->req() == 3) {
248 int ind = 1;
249 if (!this_region->in(ind)->is_IfFalse()) {
250 ind = 2;
251 }
252 if (this_region->in(ind)->is_IfFalse()) {
253 Node* bol = this_region->in(ind)->in(0)->in(1);
254 assert(bol->is_Bool(), "");
255 cmpx = bol->in(1);
256 if (bol->as_Bool()->_test._test == BoolTest::ne &&
257 cmpx->is_Cmp() && cmpx->in(2) == intcon(0) &&
258 cmpx->in(1)->is_Load()) {
259 Node* adr = cmpx->in(1)->as_Load()->in(MemNode::Address);
260 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() +
261 PtrQueue::byte_offset_of_active());
262 if (adr->is_AddP() && adr->in(AddPNode::Base) == top() &&
263 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
264 adr->in(AddPNode::Offset) == MakeConX(marking_offset)) {
265 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ));
266 }
267 }
268 }
269 }
270 // Now CastP2X can be removed since it is used only on dead path
271 // which currently still alive until igvn optimize it.
272 assert(p2x->unique_out()->Opcode() == Op_URShiftX, "");
273 _igvn.replace_node(p2x, top());
274 }
275 }
277 // Search for a memory operation for the specified memory slice.
278 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
279 Node *orig_mem = mem;
280 Node *alloc_mem = alloc->in(TypeFunc::Memory);
281 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
282 while (true) {
283 if (mem == alloc_mem || mem == start_mem ) {
284 return mem; // hit one of our sentinels
285 } else if (mem->is_MergeMem()) {
286 mem = mem->as_MergeMem()->memory_at(alias_idx);
287 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
288 Node *in = mem->in(0);
289 // we can safely skip over safepoints, calls, locks and membars because we
290 // already know that the object is safe to eliminate.
291 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
292 return in;
293 } else if (in->is_Call()) {
294 CallNode *call = in->as_Call();
295 if (!call->may_modify(tinst, phase)) {
296 mem = call->in(TypeFunc::Memory);
297 }
298 mem = in->in(TypeFunc::Memory);
299 } else if (in->is_MemBar()) {
300 mem = in->in(TypeFunc::Memory);
301 } else {
302 assert(false, "unexpected projection");
303 }
304 } else if (mem->is_Store()) {
305 const TypePtr* atype = mem->as_Store()->adr_type();
306 int adr_idx = Compile::current()->get_alias_index(atype);
307 if (adr_idx == alias_idx) {
308 assert(atype->isa_oopptr(), "address type must be oopptr");
309 int adr_offset = atype->offset();
310 uint adr_iid = atype->is_oopptr()->instance_id();
311 // Array elements references have the same alias_idx
312 // but different offset and different instance_id.
313 if (adr_offset == offset && adr_iid == alloc->_idx)
314 return mem;
315 } else {
316 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
317 }
318 mem = mem->in(MemNode::Memory);
319 } else if (mem->Opcode() == Op_SCMemProj) {
320 assert(mem->in(0)->is_LoadStore(), "sanity");
321 const TypePtr* atype = mem->in(0)->in(MemNode::Address)->bottom_type()->is_ptr();
322 int adr_idx = Compile::current()->get_alias_index(atype);
323 if (adr_idx == alias_idx) {
324 assert(false, "Object is not scalar replaceable if a LoadStore node access its field");
325 return NULL;
326 }
327 mem = mem->in(0)->in(MemNode::Memory);
328 } else {
329 return mem;
330 }
331 assert(mem != orig_mem, "dead memory loop");
332 }
333 }
335 //
336 // Given a Memory Phi, compute a value Phi containing the values from stores
337 // on the input paths.
338 // Note: this function is recursive, its depth is limied by the "level" argument
339 // Returns the computed Phi, or NULL if it cannot compute it.
340 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) {
341 assert(mem->is_Phi(), "sanity");
342 int alias_idx = C->get_alias_index(adr_t);
343 int offset = adr_t->offset();
344 int instance_id = adr_t->instance_id();
346 // Check if an appropriate value phi already exists.
347 Node* region = mem->in(0);
348 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
349 Node* phi = region->fast_out(k);
350 if (phi->is_Phi() && phi != mem &&
351 phi->as_Phi()->is_same_inst_field(phi_type, instance_id, alias_idx, offset)) {
352 return phi;
353 }
354 }
355 // Check if an appropriate new value phi already exists.
356 Node* new_phi = NULL;
357 uint size = value_phis->size();
358 for (uint i=0; i < size; i++) {
359 if ( mem->_idx == value_phis->index_at(i) ) {
360 return value_phis->node_at(i);
361 }
362 }
364 if (level <= 0) {
365 return NULL; // Give up: phi tree too deep
366 }
367 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
368 Node *alloc_mem = alloc->in(TypeFunc::Memory);
370 uint length = mem->req();
371 GrowableArray <Node *> values(length, length, NULL);
373 // create a new Phi for the value
374 PhiNode *phi = new (C, length) PhiNode(mem->in(0), phi_type, NULL, instance_id, alias_idx, offset);
375 transform_later(phi);
376 value_phis->push(phi, mem->_idx);
378 for (uint j = 1; j < length; j++) {
379 Node *in = mem->in(j);
380 if (in == NULL || in->is_top()) {
381 values.at_put(j, in);
382 } else {
383 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
384 if (val == start_mem || val == alloc_mem) {
385 // hit a sentinel, return appropriate 0 value
386 values.at_put(j, _igvn.zerocon(ft));
387 continue;
388 }
389 if (val->is_Initialize()) {
390 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
391 }
392 if (val == NULL) {
393 return NULL; // can't find a value on this path
394 }
395 if (val == mem) {
396 values.at_put(j, mem);
397 } else if (val->is_Store()) {
398 values.at_put(j, val->in(MemNode::ValueIn));
399 } else if(val->is_Proj() && val->in(0) == alloc) {
400 values.at_put(j, _igvn.zerocon(ft));
401 } else if (val->is_Phi()) {
402 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
403 if (val == NULL) {
404 return NULL;
405 }
406 values.at_put(j, val);
407 } else if (val->Opcode() == Op_SCMemProj) {
408 assert(val->in(0)->is_LoadStore(), "sanity");
409 assert(false, "Object is not scalar replaceable if a LoadStore node access its field");
410 return NULL;
411 } else {
412 #ifdef ASSERT
413 val->dump();
414 assert(false, "unknown node on this path");
415 #endif
416 return NULL; // unknown node on this path
417 }
418 }
419 }
420 // Set Phi's inputs
421 for (uint j = 1; j < length; j++) {
422 if (values.at(j) == mem) {
423 phi->init_req(j, phi);
424 } else {
425 phi->init_req(j, values.at(j));
426 }
427 }
428 return phi;
429 }
431 // Search the last value stored into the object's field.
432 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, Node *alloc) {
433 assert(adr_t->is_known_instance_field(), "instance required");
434 int instance_id = adr_t->instance_id();
435 assert((uint)instance_id == alloc->_idx, "wrong allocation");
437 int alias_idx = C->get_alias_index(adr_t);
438 int offset = adr_t->offset();
439 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
440 Node *alloc_ctrl = alloc->in(TypeFunc::Control);
441 Node *alloc_mem = alloc->in(TypeFunc::Memory);
442 Arena *a = Thread::current()->resource_area();
443 VectorSet visited(a);
446 bool done = sfpt_mem == alloc_mem;
447 Node *mem = sfpt_mem;
448 while (!done) {
449 if (visited.test_set(mem->_idx)) {
450 return NULL; // found a loop, give up
451 }
452 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
453 if (mem == start_mem || mem == alloc_mem) {
454 done = true; // hit a sentinel, return appropriate 0 value
455 } else if (mem->is_Initialize()) {
456 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
457 if (mem == NULL) {
458 done = true; // Something go wrong.
459 } else if (mem->is_Store()) {
460 const TypePtr* atype = mem->as_Store()->adr_type();
461 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
462 done = true;
463 }
464 } else if (mem->is_Store()) {
465 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
466 assert(atype != NULL, "address type must be oopptr");
467 assert(C->get_alias_index(atype) == alias_idx &&
468 atype->is_known_instance_field() && atype->offset() == offset &&
469 atype->instance_id() == instance_id, "store is correct memory slice");
470 done = true;
471 } else if (mem->is_Phi()) {
472 // try to find a phi's unique input
473 Node *unique_input = NULL;
474 Node *top = C->top();
475 for (uint i = 1; i < mem->req(); i++) {
476 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
477 if (n == NULL || n == top || n == mem) {
478 continue;
479 } else if (unique_input == NULL) {
480 unique_input = n;
481 } else if (unique_input != n) {
482 unique_input = top;
483 break;
484 }
485 }
486 if (unique_input != NULL && unique_input != top) {
487 mem = unique_input;
488 } else {
489 done = true;
490 }
491 } else {
492 assert(false, "unexpected node");
493 }
494 }
495 if (mem != NULL) {
496 if (mem == start_mem || mem == alloc_mem) {
497 // hit a sentinel, return appropriate 0 value
498 return _igvn.zerocon(ft);
499 } else if (mem->is_Store()) {
500 return mem->in(MemNode::ValueIn);
501 } else if (mem->is_Phi()) {
502 // attempt to produce a Phi reflecting the values on the input paths of the Phi
503 Node_Stack value_phis(a, 8);
504 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
505 if (phi != NULL) {
506 return phi;
507 } else {
508 // Kill all new Phis
509 while(value_phis.is_nonempty()) {
510 Node* n = value_phis.node();
511 _igvn.hash_delete(n);
512 _igvn.subsume_node(n, C->top());
513 value_phis.pop();
514 }
515 }
516 }
517 }
518 // Something go wrong.
519 return NULL;
520 }
522 // Check the possibility of scalar replacement.
523 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
524 // Scan the uses of the allocation to check for anything that would
525 // prevent us from eliminating it.
526 NOT_PRODUCT( const char* fail_eliminate = NULL; )
527 DEBUG_ONLY( Node* disq_node = NULL; )
528 bool can_eliminate = true;
530 Node* res = alloc->result_cast();
531 const TypeOopPtr* res_type = NULL;
532 if (res == NULL) {
533 // All users were eliminated.
534 } else if (!res->is_CheckCastPP()) {
535 alloc->_is_scalar_replaceable = false; // don't try again
536 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
537 can_eliminate = false;
538 } else {
539 res_type = _igvn.type(res)->isa_oopptr();
540 if (res_type == NULL) {
541 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
542 can_eliminate = false;
543 } else if (res_type->isa_aryptr()) {
544 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
545 if (length < 0) {
546 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
547 can_eliminate = false;
548 }
549 }
550 }
552 if (can_eliminate && res != NULL) {
553 for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
554 j < jmax && can_eliminate; j++) {
555 Node* use = res->fast_out(j);
557 if (use->is_AddP()) {
558 const TypePtr* addp_type = _igvn.type(use)->is_ptr();
559 int offset = addp_type->offset();
561 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
562 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";)
563 can_eliminate = false;
564 break;
565 }
566 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
567 k < kmax && can_eliminate; k++) {
568 Node* n = use->fast_out(k);
569 if (!n->is_Store() && n->Opcode() != Op_CastP2X) {
570 DEBUG_ONLY(disq_node = n;)
571 if (n->is_Load() || n->is_LoadStore()) {
572 NOT_PRODUCT(fail_eliminate = "Field load";)
573 } else {
574 NOT_PRODUCT(fail_eliminate = "Not store field referrence";)
575 }
576 can_eliminate = false;
577 }
578 }
579 } else if (use->is_SafePoint()) {
580 SafePointNode* sfpt = use->as_SafePoint();
581 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
582 // Object is passed as argument.
583 DEBUG_ONLY(disq_node = use;)
584 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
585 can_eliminate = false;
586 }
587 Node* sfptMem = sfpt->memory();
588 if (sfptMem == NULL || sfptMem->is_top()) {
589 DEBUG_ONLY(disq_node = use;)
590 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";)
591 can_eliminate = false;
592 } else {
593 safepoints.append_if_missing(sfpt);
594 }
595 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
596 if (use->is_Phi()) {
597 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
598 NOT_PRODUCT(fail_eliminate = "Object is return value";)
599 } else {
600 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
601 }
602 DEBUG_ONLY(disq_node = use;)
603 } else {
604 if (use->Opcode() == Op_Return) {
605 NOT_PRODUCT(fail_eliminate = "Object is return value";)
606 }else {
607 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
608 }
609 DEBUG_ONLY(disq_node = use;)
610 }
611 can_eliminate = false;
612 }
613 }
614 }
616 #ifndef PRODUCT
617 if (PrintEliminateAllocations) {
618 if (can_eliminate) {
619 tty->print("Scalar ");
620 if (res == NULL)
621 alloc->dump();
622 else
623 res->dump();
624 } else {
625 tty->print("NotScalar (%s)", fail_eliminate);
626 if (res == NULL)
627 alloc->dump();
628 else
629 res->dump();
630 #ifdef ASSERT
631 if (disq_node != NULL) {
632 tty->print(" >>>> ");
633 disq_node->dump();
634 }
635 #endif /*ASSERT*/
636 }
637 }
638 #endif
639 return can_eliminate;
640 }
642 // Do scalar replacement.
643 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
644 GrowableArray <SafePointNode *> safepoints_done;
646 ciKlass* klass = NULL;
647 ciInstanceKlass* iklass = NULL;
648 int nfields = 0;
649 int array_base;
650 int element_size;
651 BasicType basic_elem_type;
652 ciType* elem_type;
654 Node* res = alloc->result_cast();
655 const TypeOopPtr* res_type = NULL;
656 if (res != NULL) { // Could be NULL when there are no users
657 res_type = _igvn.type(res)->isa_oopptr();
658 }
660 if (res != NULL) {
661 klass = res_type->klass();
662 if (res_type->isa_instptr()) {
663 // find the fields of the class which will be needed for safepoint debug information
664 assert(klass->is_instance_klass(), "must be an instance klass.");
665 iklass = klass->as_instance_klass();
666 nfields = iklass->nof_nonstatic_fields();
667 } else {
668 // find the array's elements which will be needed for safepoint debug information
669 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
670 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass.");
671 elem_type = klass->as_array_klass()->element_type();
672 basic_elem_type = elem_type->basic_type();
673 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
674 element_size = type2aelembytes(basic_elem_type);
675 }
676 }
677 //
678 // Process the safepoint uses
679 //
680 while (safepoints.length() > 0) {
681 SafePointNode* sfpt = safepoints.pop();
682 Node* mem = sfpt->memory();
683 uint first_ind = sfpt->req();
684 SafePointScalarObjectNode* sobj = new (C, 1) SafePointScalarObjectNode(res_type,
685 #ifdef ASSERT
686 alloc,
687 #endif
688 first_ind, nfields);
689 sobj->init_req(0, sfpt->in(TypeFunc::Control));
690 transform_later(sobj);
692 // Scan object's fields adding an input to the safepoint for each field.
693 for (int j = 0; j < nfields; j++) {
694 intptr_t offset;
695 ciField* field = NULL;
696 if (iklass != NULL) {
697 field = iklass->nonstatic_field_at(j);
698 offset = field->offset();
699 elem_type = field->type();
700 basic_elem_type = field->layout_type();
701 } else {
702 offset = array_base + j * (intptr_t)element_size;
703 }
705 const Type *field_type;
706 // The next code is taken from Parse::do_get_xxx().
707 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) {
708 if (!elem_type->is_loaded()) {
709 field_type = TypeInstPtr::BOTTOM;
710 } else if (field != NULL && field->is_constant()) {
711 // This can happen if the constant oop is non-perm.
712 ciObject* con = field->constant_value().as_object();
713 // Do not "join" in the previous type; it doesn't add value,
714 // and may yield a vacuous result if the field is of interface type.
715 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
716 assert(field_type != NULL, "field singleton type must be consistent");
717 } else {
718 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
719 }
720 if (UseCompressedOops) {
721 field_type = field_type->make_narrowoop();
722 basic_elem_type = T_NARROWOOP;
723 }
724 } else {
725 field_type = Type::get_const_basic_type(basic_elem_type);
726 }
728 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr();
730 Node *field_val = value_from_mem(mem, basic_elem_type, field_type, field_addr_type, alloc);
731 if (field_val == NULL) {
732 // we weren't able to find a value for this field,
733 // give up on eliminating this allocation
734 alloc->_is_scalar_replaceable = false; // don't try again
735 // remove any extra entries we added to the safepoint
736 uint last = sfpt->req() - 1;
737 for (int k = 0; k < j; k++) {
738 sfpt->del_req(last--);
739 }
740 // rollback processed safepoints
741 while (safepoints_done.length() > 0) {
742 SafePointNode* sfpt_done = safepoints_done.pop();
743 // remove any extra entries we added to the safepoint
744 last = sfpt_done->req() - 1;
745 for (int k = 0; k < nfields; k++) {
746 sfpt_done->del_req(last--);
747 }
748 JVMState *jvms = sfpt_done->jvms();
749 jvms->set_endoff(sfpt_done->req());
750 // Now make a pass over the debug information replacing any references
751 // to SafePointScalarObjectNode with the allocated object.
752 int start = jvms->debug_start();
753 int end = jvms->debug_end();
754 for (int i = start; i < end; i++) {
755 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
756 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
757 if (scobj->first_index() == sfpt_done->req() &&
758 scobj->n_fields() == (uint)nfields) {
759 assert(scobj->alloc() == alloc, "sanity");
760 sfpt_done->set_req(i, res);
761 }
762 }
763 }
764 }
765 #ifndef PRODUCT
766 if (PrintEliminateAllocations) {
767 if (field != NULL) {
768 tty->print("=== At SafePoint node %d can't find value of Field: ",
769 sfpt->_idx);
770 field->print();
771 int field_idx = C->get_alias_index(field_addr_type);
772 tty->print(" (alias_idx=%d)", field_idx);
773 } else { // Array's element
774 tty->print("=== At SafePoint node %d can't find value of array element [%d]",
775 sfpt->_idx, j);
776 }
777 tty->print(", which prevents elimination of: ");
778 if (res == NULL)
779 alloc->dump();
780 else
781 res->dump();
782 }
783 #endif
784 return false;
785 }
786 if (UseCompressedOops && field_type->isa_narrowoop()) {
787 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
788 // to be able scalar replace the allocation.
789 if (field_val->is_EncodeP()) {
790 field_val = field_val->in(1);
791 } else {
792 field_val = transform_later(new (C, 2) DecodeNNode(field_val, field_val->bottom_type()->make_ptr()));
793 }
794 }
795 sfpt->add_req(field_val);
796 }
797 JVMState *jvms = sfpt->jvms();
798 jvms->set_endoff(sfpt->req());
799 // Now make a pass over the debug information replacing any references
800 // to the allocated object with "sobj"
801 int start = jvms->debug_start();
802 int end = jvms->debug_end();
803 for (int i = start; i < end; i++) {
804 if (sfpt->in(i) == res) {
805 sfpt->set_req(i, sobj);
806 }
807 }
808 safepoints_done.append_if_missing(sfpt); // keep it for rollback
809 }
810 return true;
811 }
813 // Process users of eliminated allocation.
814 void PhaseMacroExpand::process_users_of_allocation(AllocateNode *alloc) {
815 Node* res = alloc->result_cast();
816 if (res != NULL) {
817 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
818 Node *use = res->last_out(j);
819 uint oc1 = res->outcnt();
821 if (use->is_AddP()) {
822 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
823 Node *n = use->last_out(k);
824 uint oc2 = use->outcnt();
825 if (n->is_Store()) {
826 _igvn.replace_node(n, n->in(MemNode::Memory));
827 } else {
828 eliminate_card_mark(n);
829 }
830 k -= (oc2 - use->outcnt());
831 }
832 } else {
833 eliminate_card_mark(use);
834 }
835 j -= (oc1 - res->outcnt());
836 }
837 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
838 _igvn.remove_dead_node(res);
839 }
841 //
842 // Process other users of allocation's projections
843 //
844 if (_resproj != NULL && _resproj->outcnt() != 0) {
845 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) {
846 Node *use = _resproj->last_out(j);
847 uint oc1 = _resproj->outcnt();
848 if (use->is_Initialize()) {
849 // Eliminate Initialize node.
850 InitializeNode *init = use->as_Initialize();
851 assert(init->outcnt() <= 2, "only a control and memory projection expected");
852 Node *ctrl_proj = init->proj_out(TypeFunc::Control);
853 if (ctrl_proj != NULL) {
854 assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection");
855 _igvn.replace_node(ctrl_proj, _fallthroughcatchproj);
856 }
857 Node *mem_proj = init->proj_out(TypeFunc::Memory);
858 if (mem_proj != NULL) {
859 Node *mem = init->in(TypeFunc::Memory);
860 #ifdef ASSERT
861 if (mem->is_MergeMem()) {
862 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection");
863 } else {
864 assert(mem == _memproj_fallthrough, "allocation memory projection");
865 }
866 #endif
867 _igvn.replace_node(mem_proj, mem);
868 }
869 } else if (use->is_AddP()) {
870 // raw memory addresses used only by the initialization
871 _igvn.replace_node(use, C->top());
872 } else {
873 assert(false, "only Initialize or AddP expected");
874 }
875 j -= (oc1 - _resproj->outcnt());
876 }
877 }
878 if (_fallthroughcatchproj != NULL) {
879 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control));
880 }
881 if (_memproj_fallthrough != NULL) {
882 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory));
883 }
884 if (_memproj_catchall != NULL) {
885 _igvn.replace_node(_memproj_catchall, C->top());
886 }
887 if (_ioproj_fallthrough != NULL) {
888 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O));
889 }
890 if (_ioproj_catchall != NULL) {
891 _igvn.replace_node(_ioproj_catchall, C->top());
892 }
893 if (_catchallcatchproj != NULL) {
894 _igvn.replace_node(_catchallcatchproj, C->top());
895 }
896 }
898 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
900 if (!EliminateAllocations || !alloc->_is_scalar_replaceable) {
901 return false;
902 }
904 extract_call_projections(alloc);
906 GrowableArray <SafePointNode *> safepoints;
907 if (!can_eliminate_allocation(alloc, safepoints)) {
908 return false;
909 }
911 if (!scalar_replacement(alloc, safepoints)) {
912 return false;
913 }
915 process_users_of_allocation(alloc);
917 #ifndef PRODUCT
918 if (PrintEliminateAllocations) {
919 if (alloc->is_AllocateArray())
920 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
921 else
922 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
923 }
924 #endif
926 return true;
927 }
930 //---------------------------set_eden_pointers-------------------------
931 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) {
932 if (UseTLAB) { // Private allocation: load from TLS
933 Node* thread = transform_later(new (C, 1) ThreadLocalNode());
934 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset());
935 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset());
936 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset);
937 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset);
938 } else { // Shared allocation: load from globals
939 CollectedHeap* ch = Universe::heap();
940 address top_adr = (address)ch->top_addr();
941 address end_adr = (address)ch->end_addr();
942 eden_top_adr = makecon(TypeRawPtr::make(top_adr));
943 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr);
944 }
945 }
948 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
949 Node* adr = basic_plus_adr(base, offset);
950 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
951 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt);
952 transform_later(value);
953 return value;
954 }
957 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
958 Node* adr = basic_plus_adr(base, offset);
959 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt);
960 transform_later(mem);
961 return mem;
962 }
964 //=============================================================================
965 //
966 // A L L O C A T I O N
967 //
968 // Allocation attempts to be fast in the case of frequent small objects.
969 // It breaks down like this:
970 //
971 // 1) Size in doublewords is computed. This is a constant for objects and
972 // variable for most arrays. Doubleword units are used to avoid size
973 // overflow of huge doubleword arrays. We need doublewords in the end for
974 // rounding.
975 //
976 // 2) Size is checked for being 'too large'. Too-large allocations will go
977 // the slow path into the VM. The slow path can throw any required
978 // exceptions, and does all the special checks for very large arrays. The
979 // size test can constant-fold away for objects. For objects with
980 // finalizers it constant-folds the otherway: you always go slow with
981 // finalizers.
982 //
983 // 3) If NOT using TLABs, this is the contended loop-back point.
984 // Load-Locked the heap top. If using TLABs normal-load the heap top.
985 //
986 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
987 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
988 // "size*8" we always enter the VM, where "largish" is a constant picked small
989 // enough that there's always space between the eden max and 4Gig (old space is
990 // there so it's quite large) and large enough that the cost of entering the VM
991 // is dwarfed by the cost to initialize the space.
992 //
993 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
994 // down. If contended, repeat at step 3. If using TLABs normal-store
995 // adjusted heap top back down; there is no contention.
996 //
997 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
998 // fields.
999 //
1000 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1001 // oop flavor.
1002 //
1003 //=============================================================================
1004 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1005 // Allocations bigger than this always go the slow route.
1006 // This value must be small enough that allocation attempts that need to
1007 // trigger exceptions go the slow route. Also, it must be small enough so
1008 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1009 //=============================================================================j//
1010 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1011 // The allocator will coalesce int->oop copies away. See comment in
1012 // coalesce.cpp about how this works. It depends critically on the exact
1013 // code shape produced here, so if you are changing this code shape
1014 // make sure the GC info for the heap-top is correct in and around the
1015 // slow-path call.
1016 //
1018 void PhaseMacroExpand::expand_allocate_common(
1019 AllocateNode* alloc, // allocation node to be expanded
1020 Node* length, // array length for an array allocation
1021 const TypeFunc* slow_call_type, // Type of slow call
1022 address slow_call_address // Address of slow call
1023 )
1024 {
1026 Node* ctrl = alloc->in(TypeFunc::Control);
1027 Node* mem = alloc->in(TypeFunc::Memory);
1028 Node* i_o = alloc->in(TypeFunc::I_O);
1029 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1030 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1031 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1033 assert(ctrl != NULL, "must have control");
1034 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1035 // they will not be used if "always_slow" is set
1036 enum { slow_result_path = 1, fast_result_path = 2 };
1037 Node *result_region;
1038 Node *result_phi_rawmem;
1039 Node *result_phi_rawoop;
1040 Node *result_phi_i_o;
1042 // The initial slow comparison is a size check, the comparison
1043 // we want to do is a BoolTest::gt
1044 bool always_slow = false;
1045 int tv = _igvn.find_int_con(initial_slow_test, -1);
1046 if (tv >= 0) {
1047 always_slow = (tv == 1);
1048 initial_slow_test = NULL;
1049 } else {
1050 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1051 }
1053 if (C->env()->dtrace_alloc_probes() ||
1054 !UseTLAB && (!Universe::heap()->supports_inline_contig_alloc() ||
1055 (UseConcMarkSweepGC && CMSIncrementalMode))) {
1056 // Force slow-path allocation
1057 always_slow = true;
1058 initial_slow_test = NULL;
1059 }
1062 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1063 Node *slow_region = NULL;
1064 Node *toobig_false = ctrl;
1066 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent");
1067 // generate the initial test if necessary
1068 if (initial_slow_test != NULL ) {
1069 slow_region = new (C, 3) RegionNode(3);
1071 // Now make the initial failure test. Usually a too-big test but
1072 // might be a TRUE for finalizers or a fancy class check for
1073 // newInstance0.
1074 IfNode *toobig_iff = new (C, 2) IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1075 transform_later(toobig_iff);
1076 // Plug the failing-too-big test into the slow-path region
1077 Node *toobig_true = new (C, 1) IfTrueNode( toobig_iff );
1078 transform_later(toobig_true);
1079 slow_region ->init_req( too_big_or_final_path, toobig_true );
1080 toobig_false = new (C, 1) IfFalseNode( toobig_iff );
1081 transform_later(toobig_false);
1082 } else { // No initial test, just fall into next case
1083 toobig_false = ctrl;
1084 debug_only(slow_region = NodeSentinel);
1085 }
1087 Node *slow_mem = mem; // save the current memory state for slow path
1088 // generate the fast allocation code unless we know that the initial test will always go slow
1089 if (!always_slow) {
1090 // Fast path modifies only raw memory.
1091 if (mem->is_MergeMem()) {
1092 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1093 }
1095 Node* eden_top_adr;
1096 Node* eden_end_adr;
1098 set_eden_pointers(eden_top_adr, eden_end_adr);
1100 // Load Eden::end. Loop invariant and hoisted.
1101 //
1102 // Note: We set the control input on "eden_end" and "old_eden_top" when using
1103 // a TLAB to work around a bug where these values were being moved across
1104 // a safepoint. These are not oops, so they cannot be include in the oop
1105 // map, but the can be changed by a GC. The proper way to fix this would
1106 // be to set the raw memory state when generating a SafepointNode. However
1107 // this will require extensive changes to the loop optimization in order to
1108 // prevent a degradation of the optimization.
1109 // See comment in memnode.hpp, around line 227 in class LoadPNode.
1110 Node *eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS);
1112 // allocate the Region and Phi nodes for the result
1113 result_region = new (C, 3) RegionNode(3);
1114 result_phi_rawmem = new (C, 3) PhiNode( result_region, Type::MEMORY, TypeRawPtr::BOTTOM );
1115 result_phi_rawoop = new (C, 3) PhiNode( result_region, TypeRawPtr::BOTTOM );
1116 result_phi_i_o = new (C, 3) PhiNode( result_region, Type::ABIO ); // I/O is used for Prefetch
1118 // We need a Region for the loop-back contended case.
1119 enum { fall_in_path = 1, contended_loopback_path = 2 };
1120 Node *contended_region;
1121 Node *contended_phi_rawmem;
1122 if( UseTLAB ) {
1123 contended_region = toobig_false;
1124 contended_phi_rawmem = mem;
1125 } else {
1126 contended_region = new (C, 3) RegionNode(3);
1127 contended_phi_rawmem = new (C, 3) PhiNode( contended_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1128 // Now handle the passing-too-big test. We fall into the contended
1129 // loop-back merge point.
1130 contended_region ->init_req( fall_in_path, toobig_false );
1131 contended_phi_rawmem->init_req( fall_in_path, mem );
1132 transform_later(contended_region);
1133 transform_later(contended_phi_rawmem);
1134 }
1136 // Load(-locked) the heap top.
1137 // See note above concerning the control input when using a TLAB
1138 Node *old_eden_top = UseTLAB
1139 ? new (C, 3) LoadPNode ( ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM )
1140 : new (C, 3) LoadPLockedNode( contended_region, contended_phi_rawmem, eden_top_adr );
1142 transform_later(old_eden_top);
1143 // Add to heap top to get a new heap top
1144 Node *new_eden_top = new (C, 4) AddPNode( top(), old_eden_top, size_in_bytes );
1145 transform_later(new_eden_top);
1146 // Check for needing a GC; compare against heap end
1147 Node *needgc_cmp = new (C, 3) CmpPNode( new_eden_top, eden_end );
1148 transform_later(needgc_cmp);
1149 Node *needgc_bol = new (C, 2) BoolNode( needgc_cmp, BoolTest::ge );
1150 transform_later(needgc_bol);
1151 IfNode *needgc_iff = new (C, 2) IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1152 transform_later(needgc_iff);
1154 // Plug the failing-heap-space-need-gc test into the slow-path region
1155 Node *needgc_true = new (C, 1) IfTrueNode( needgc_iff );
1156 transform_later(needgc_true);
1157 if( initial_slow_test ) {
1158 slow_region ->init_req( need_gc_path, needgc_true );
1159 // This completes all paths into the slow merge point
1160 transform_later(slow_region);
1161 } else { // No initial slow path needed!
1162 // Just fall from the need-GC path straight into the VM call.
1163 slow_region = needgc_true;
1164 }
1165 // No need for a GC. Setup for the Store-Conditional
1166 Node *needgc_false = new (C, 1) IfFalseNode( needgc_iff );
1167 transform_later(needgc_false);
1169 // Grab regular I/O before optional prefetch may change it.
1170 // Slow-path does no I/O so just set it to the original I/O.
1171 result_phi_i_o->init_req( slow_result_path, i_o );
1173 i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem,
1174 old_eden_top, new_eden_top, length);
1176 // Store (-conditional) the modified eden top back down.
1177 // StorePConditional produces flags for a test PLUS a modified raw
1178 // memory state.
1179 Node *store_eden_top;
1180 Node *fast_oop_ctrl;
1181 if( UseTLAB ) {
1182 store_eden_top = new (C, 4) StorePNode( needgc_false, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, new_eden_top );
1183 transform_later(store_eden_top);
1184 fast_oop_ctrl = needgc_false; // No contention, so this is the fast path
1185 } else {
1186 store_eden_top = new (C, 5) StorePConditionalNode( needgc_false, contended_phi_rawmem, eden_top_adr, new_eden_top, old_eden_top );
1187 transform_later(store_eden_top);
1188 Node *contention_check = new (C, 2) BoolNode( store_eden_top, BoolTest::ne );
1189 transform_later(contention_check);
1190 store_eden_top = new (C, 1) SCMemProjNode(store_eden_top);
1191 transform_later(store_eden_top);
1193 // If not using TLABs, check to see if there was contention.
1194 IfNode *contention_iff = new (C, 2) IfNode ( needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN );
1195 transform_later(contention_iff);
1196 Node *contention_true = new (C, 1) IfTrueNode( contention_iff );
1197 transform_later(contention_true);
1198 // If contention, loopback and try again.
1199 contended_region->init_req( contended_loopback_path, contention_true );
1200 contended_phi_rawmem->init_req( contended_loopback_path, store_eden_top );
1202 // Fast-path succeeded with no contention!
1203 Node *contention_false = new (C, 1) IfFalseNode( contention_iff );
1204 transform_later(contention_false);
1205 fast_oop_ctrl = contention_false;
1206 }
1208 // Rename successful fast-path variables to make meaning more obvious
1209 Node* fast_oop = old_eden_top;
1210 Node* fast_oop_rawmem = store_eden_top;
1211 fast_oop_rawmem = initialize_object(alloc,
1212 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1213 klass_node, length, size_in_bytes);
1215 if (C->env()->dtrace_extended_probes()) {
1216 // Slow-path call
1217 int size = TypeFunc::Parms + 2;
1218 CallLeafNode *call = new (C, size) CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1219 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base),
1220 "dtrace_object_alloc",
1221 TypeRawPtr::BOTTOM);
1223 // Get base of thread-local storage area
1224 Node* thread = new (C, 1) ThreadLocalNode();
1225 transform_later(thread);
1227 call->init_req(TypeFunc::Parms+0, thread);
1228 call->init_req(TypeFunc::Parms+1, fast_oop);
1229 call->init_req( TypeFunc::Control, fast_oop_ctrl );
1230 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o
1231 call->init_req( TypeFunc::Memory , fast_oop_rawmem );
1232 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) );
1233 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) );
1234 transform_later(call);
1235 fast_oop_ctrl = new (C, 1) ProjNode(call,TypeFunc::Control);
1236 transform_later(fast_oop_ctrl);
1237 fast_oop_rawmem = new (C, 1) ProjNode(call,TypeFunc::Memory);
1238 transform_later(fast_oop_rawmem);
1239 }
1241 // Plug in the successful fast-path into the result merge point
1242 result_region ->init_req( fast_result_path, fast_oop_ctrl );
1243 result_phi_rawoop->init_req( fast_result_path, fast_oop );
1244 result_phi_i_o ->init_req( fast_result_path, i_o );
1245 result_phi_rawmem->init_req( fast_result_path, fast_oop_rawmem );
1246 } else {
1247 slow_region = ctrl;
1248 }
1250 // Generate slow-path call
1251 CallNode *call = new (C, slow_call_type->domain()->cnt())
1252 CallStaticJavaNode(slow_call_type, slow_call_address,
1253 OptoRuntime::stub_name(slow_call_address),
1254 alloc->jvms()->bci(),
1255 TypePtr::BOTTOM);
1256 call->init_req( TypeFunc::Control, slow_region );
1257 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o
1258 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs
1259 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) );
1260 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) );
1262 call->init_req(TypeFunc::Parms+0, klass_node);
1263 if (length != NULL) {
1264 call->init_req(TypeFunc::Parms+1, length);
1265 }
1267 // Copy debug information and adjust JVMState information, then replace
1268 // allocate node with the call
1269 copy_call_debug_info((CallNode *) alloc, call);
1270 if (!always_slow) {
1271 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1272 }
1273 _igvn.hash_delete(alloc);
1274 _igvn.subsume_node(alloc, call);
1275 transform_later(call);
1277 // Identify the output projections from the allocate node and
1278 // adjust any references to them.
1279 // The control and io projections look like:
1280 //
1281 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1282 // Allocate Catch
1283 // ^---Proj(io) <-------+ ^---CatchProj(io)
1284 //
1285 // We are interested in the CatchProj nodes.
1286 //
1287 extract_call_projections(call);
1289 // An allocate node has separate memory projections for the uses on the control and i_o paths
1290 // Replace uses of the control memory projection with result_phi_rawmem (unless we are only generating a slow call)
1291 if (!always_slow && _memproj_fallthrough != NULL) {
1292 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) {
1293 Node *use = _memproj_fallthrough->fast_out(i);
1294 _igvn.hash_delete(use);
1295 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem);
1296 _igvn._worklist.push(use);
1297 // back up iterator
1298 --i;
1299 }
1300 }
1301 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete _memproj_catchall so
1302 // we end up with a call that has only 1 memory projection
1303 if (_memproj_catchall != NULL ) {
1304 if (_memproj_fallthrough == NULL) {
1305 _memproj_fallthrough = new (C, 1) ProjNode(call, TypeFunc::Memory);
1306 transform_later(_memproj_fallthrough);
1307 }
1308 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) {
1309 Node *use = _memproj_catchall->fast_out(i);
1310 _igvn.hash_delete(use);
1311 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough);
1312 _igvn._worklist.push(use);
1313 // back up iterator
1314 --i;
1315 }
1316 }
1318 // An allocate node has separate i_o projections for the uses on the control and i_o paths
1319 // Replace uses of the control i_o projection with result_phi_i_o (unless we are only generating a slow call)
1320 if (_ioproj_fallthrough == NULL) {
1321 _ioproj_fallthrough = new (C, 1) ProjNode(call, TypeFunc::I_O);
1322 transform_later(_ioproj_fallthrough);
1323 } else if (!always_slow) {
1324 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) {
1325 Node *use = _ioproj_fallthrough->fast_out(i);
1327 _igvn.hash_delete(use);
1328 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o);
1329 _igvn._worklist.push(use);
1330 // back up iterator
1331 --i;
1332 }
1333 }
1334 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete _ioproj_catchall so
1335 // we end up with a call that has only 1 control projection
1336 if (_ioproj_catchall != NULL ) {
1337 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) {
1338 Node *use = _ioproj_catchall->fast_out(i);
1339 _igvn.hash_delete(use);
1340 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough);
1341 _igvn._worklist.push(use);
1342 // back up iterator
1343 --i;
1344 }
1345 }
1347 // if we generated only a slow call, we are done
1348 if (always_slow)
1349 return;
1352 if (_fallthroughcatchproj != NULL) {
1353 ctrl = _fallthroughcatchproj->clone();
1354 transform_later(ctrl);
1355 _igvn.replace_node(_fallthroughcatchproj, result_region);
1356 } else {
1357 ctrl = top();
1358 }
1359 Node *slow_result;
1360 if (_resproj == NULL) {
1361 // no uses of the allocation result
1362 slow_result = top();
1363 } else {
1364 slow_result = _resproj->clone();
1365 transform_later(slow_result);
1366 _igvn.replace_node(_resproj, result_phi_rawoop);
1367 }
1369 // Plug slow-path into result merge point
1370 result_region ->init_req( slow_result_path, ctrl );
1371 result_phi_rawoop->init_req( slow_result_path, slow_result);
1372 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough );
1373 transform_later(result_region);
1374 transform_later(result_phi_rawoop);
1375 transform_later(result_phi_rawmem);
1376 transform_later(result_phi_i_o);
1377 // This completes all paths into the result merge point
1378 }
1381 // Helper for PhaseMacroExpand::expand_allocate_common.
1382 // Initializes the newly-allocated storage.
1383 Node*
1384 PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1385 Node* control, Node* rawmem, Node* object,
1386 Node* klass_node, Node* length,
1387 Node* size_in_bytes) {
1388 InitializeNode* init = alloc->initialization();
1389 // Store the klass & mark bits
1390 Node* mark_node = NULL;
1391 // For now only enable fast locking for non-array types
1392 if (UseBiasedLocking && (length == NULL)) {
1393 mark_node = make_load(NULL, rawmem, klass_node, Klass::prototype_header_offset_in_bytes() + sizeof(oopDesc), TypeRawPtr::BOTTOM, T_ADDRESS);
1394 } else {
1395 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype()));
1396 }
1397 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
1399 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_OBJECT);
1400 int header_size = alloc->minimum_header_size(); // conservatively small
1402 // Array length
1403 if (length != NULL) { // Arrays need length field
1404 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1405 // conservatively small header size:
1406 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1407 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1408 if (k->is_array_klass()) // we know the exact header size in most cases:
1409 header_size = Klass::layout_helper_header_size(k->layout_helper());
1410 }
1412 // Clear the object body, if necessary.
1413 if (init == NULL) {
1414 // The init has somehow disappeared; be cautious and clear everything.
1415 //
1416 // This can happen if a node is allocated but an uncommon trap occurs
1417 // immediately. In this case, the Initialize gets associated with the
1418 // trap, and may be placed in a different (outer) loop, if the Allocate
1419 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1420 // there can be two Allocates to one Initialize. The answer in all these
1421 // edge cases is safety first. It is always safe to clear immediately
1422 // within an Allocate, and then (maybe or maybe not) clear some more later.
1423 if (!ZeroTLAB)
1424 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1425 header_size, size_in_bytes,
1426 &_igvn);
1427 } else {
1428 if (!init->is_complete()) {
1429 // Try to win by zeroing only what the init does not store.
1430 // We can also try to do some peephole optimizations,
1431 // such as combining some adjacent subword stores.
1432 rawmem = init->complete_stores(control, rawmem, object,
1433 header_size, size_in_bytes, &_igvn);
1434 }
1435 // We have no more use for this link, since the AllocateNode goes away:
1436 init->set_req(InitializeNode::RawAddress, top());
1437 // (If we keep the link, it just confuses the register allocator,
1438 // who thinks he sees a real use of the address by the membar.)
1439 }
1441 return rawmem;
1442 }
1444 // Generate prefetch instructions for next allocations.
1445 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
1446 Node*& contended_phi_rawmem,
1447 Node* old_eden_top, Node* new_eden_top,
1448 Node* length) {
1449 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
1450 // Generate prefetch allocation with watermark check.
1451 // As an allocation hits the watermark, we will prefetch starting
1452 // at a "distance" away from watermark.
1453 enum { fall_in_path = 1, pf_path = 2 };
1455 Node *pf_region = new (C, 3) RegionNode(3);
1456 Node *pf_phi_rawmem = new (C, 3) PhiNode( pf_region, Type::MEMORY,
1457 TypeRawPtr::BOTTOM );
1458 // I/O is used for Prefetch
1459 Node *pf_phi_abio = new (C, 3) PhiNode( pf_region, Type::ABIO );
1461 Node *thread = new (C, 1) ThreadLocalNode();
1462 transform_later(thread);
1464 Node *eden_pf_adr = new (C, 4) AddPNode( top()/*not oop*/, thread,
1465 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
1466 transform_later(eden_pf_adr);
1468 Node *old_pf_wm = new (C, 3) LoadPNode( needgc_false,
1469 contended_phi_rawmem, eden_pf_adr,
1470 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM );
1471 transform_later(old_pf_wm);
1473 // check against new_eden_top
1474 Node *need_pf_cmp = new (C, 3) CmpPNode( new_eden_top, old_pf_wm );
1475 transform_later(need_pf_cmp);
1476 Node *need_pf_bol = new (C, 2) BoolNode( need_pf_cmp, BoolTest::ge );
1477 transform_later(need_pf_bol);
1478 IfNode *need_pf_iff = new (C, 2) IfNode( needgc_false, need_pf_bol,
1479 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1480 transform_later(need_pf_iff);
1482 // true node, add prefetchdistance
1483 Node *need_pf_true = new (C, 1) IfTrueNode( need_pf_iff );
1484 transform_later(need_pf_true);
1486 Node *need_pf_false = new (C, 1) IfFalseNode( need_pf_iff );
1487 transform_later(need_pf_false);
1489 Node *new_pf_wmt = new (C, 4) AddPNode( top(), old_pf_wm,
1490 _igvn.MakeConX(AllocatePrefetchDistance) );
1491 transform_later(new_pf_wmt );
1492 new_pf_wmt->set_req(0, need_pf_true);
1494 Node *store_new_wmt = new (C, 4) StorePNode( need_pf_true,
1495 contended_phi_rawmem, eden_pf_adr,
1496 TypeRawPtr::BOTTOM, new_pf_wmt );
1497 transform_later(store_new_wmt);
1499 // adding prefetches
1500 pf_phi_abio->init_req( fall_in_path, i_o );
1502 Node *prefetch_adr;
1503 Node *prefetch;
1504 uint lines = AllocatePrefetchDistance / AllocatePrefetchStepSize;
1505 uint step_size = AllocatePrefetchStepSize;
1506 uint distance = 0;
1508 for ( uint i = 0; i < lines; i++ ) {
1509 prefetch_adr = new (C, 4) AddPNode( old_pf_wm, new_pf_wmt,
1510 _igvn.MakeConX(distance) );
1511 transform_later(prefetch_adr);
1512 prefetch = new (C, 3) PrefetchWriteNode( i_o, prefetch_adr );
1513 transform_later(prefetch);
1514 distance += step_size;
1515 i_o = prefetch;
1516 }
1517 pf_phi_abio->set_req( pf_path, i_o );
1519 pf_region->init_req( fall_in_path, need_pf_false );
1520 pf_region->init_req( pf_path, need_pf_true );
1522 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
1523 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
1525 transform_later(pf_region);
1526 transform_later(pf_phi_rawmem);
1527 transform_later(pf_phi_abio);
1529 needgc_false = pf_region;
1530 contended_phi_rawmem = pf_phi_rawmem;
1531 i_o = pf_phi_abio;
1532 } else if( AllocatePrefetchStyle > 0 ) {
1533 // Insert a prefetch for each allocation only on the fast-path
1534 Node *prefetch_adr;
1535 Node *prefetch;
1536 // Generate several prefetch instructions only for arrays.
1537 uint lines = (length != NULL) ? AllocatePrefetchLines : 1;
1538 uint step_size = AllocatePrefetchStepSize;
1539 uint distance = AllocatePrefetchDistance;
1540 for ( uint i = 0; i < lines; i++ ) {
1541 prefetch_adr = new (C, 4) AddPNode( old_eden_top, new_eden_top,
1542 _igvn.MakeConX(distance) );
1543 transform_later(prefetch_adr);
1544 prefetch = new (C, 3) PrefetchWriteNode( i_o, prefetch_adr );
1545 // Do not let it float too high, since if eden_top == eden_end,
1546 // both might be null.
1547 if( i == 0 ) { // Set control for first prefetch, next follows it
1548 prefetch->init_req(0, needgc_false);
1549 }
1550 transform_later(prefetch);
1551 distance += step_size;
1552 i_o = prefetch;
1553 }
1554 }
1555 return i_o;
1556 }
1559 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
1560 expand_allocate_common(alloc, NULL,
1561 OptoRuntime::new_instance_Type(),
1562 OptoRuntime::new_instance_Java());
1563 }
1565 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
1566 Node* length = alloc->in(AllocateNode::ALength);
1567 expand_allocate_common(alloc, length,
1568 OptoRuntime::new_array_Type(),
1569 OptoRuntime::new_array_Java());
1570 }
1573 // we have determined that this lock/unlock can be eliminated, we simply
1574 // eliminate the node without expanding it.
1575 //
1576 // Note: The membar's associated with the lock/unlock are currently not
1577 // eliminated. This should be investigated as a future enhancement.
1578 //
1579 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
1581 if (!alock->is_eliminated()) {
1582 return false;
1583 }
1584 if (alock->is_Lock() && !alock->is_coarsened()) {
1585 // Create new "eliminated" BoxLock node and use it
1586 // in monitor debug info for the same object.
1587 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
1588 Node* obj = alock->obj_node();
1589 if (!oldbox->is_eliminated()) {
1590 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
1591 newbox->set_eliminated();
1592 transform_later(newbox);
1593 // Replace old box node with new box for all users
1594 // of the same object.
1595 for (uint i = 0; i < oldbox->outcnt();) {
1597 bool next_edge = true;
1598 Node* u = oldbox->raw_out(i);
1599 if (u == alock) {
1600 i++;
1601 continue; // It will be removed below
1602 }
1603 if (u->is_Lock() &&
1604 u->as_Lock()->obj_node() == obj &&
1605 // oldbox could be referenced in debug info also
1606 u->as_Lock()->box_node() == oldbox) {
1607 assert(u->as_Lock()->is_eliminated(), "sanity");
1608 _igvn.hash_delete(u);
1609 u->set_req(TypeFunc::Parms + 1, newbox);
1610 next_edge = false;
1611 #ifdef ASSERT
1612 } else if (u->is_Unlock() && u->as_Unlock()->obj_node() == obj) {
1613 assert(u->as_Unlock()->is_eliminated(), "sanity");
1614 #endif
1615 }
1616 // Replace old box in monitor debug info.
1617 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
1618 SafePointNode* sfn = u->as_SafePoint();
1619 JVMState* youngest_jvms = sfn->jvms();
1620 int max_depth = youngest_jvms->depth();
1621 for (int depth = 1; depth <= max_depth; depth++) {
1622 JVMState* jvms = youngest_jvms->of_depth(depth);
1623 int num_mon = jvms->nof_monitors();
1624 // Loop over monitors
1625 for (int idx = 0; idx < num_mon; idx++) {
1626 Node* obj_node = sfn->monitor_obj(jvms, idx);
1627 Node* box_node = sfn->monitor_box(jvms, idx);
1628 if (box_node == oldbox && obj_node == obj) {
1629 int j = jvms->monitor_box_offset(idx);
1630 _igvn.hash_delete(u);
1631 u->set_req(j, newbox);
1632 next_edge = false;
1633 }
1634 } // for (int idx = 0;
1635 } // for (int depth = 1;
1636 } // if (u->is_SafePoint()
1637 if (next_edge) i++;
1638 } // for (uint i = 0; i < oldbox->outcnt();)
1639 } // if (!oldbox->is_eliminated())
1640 } // if (alock->is_Lock() && !lock->is_coarsened())
1642 #ifndef PRODUCT
1643 if (PrintEliminateLocks) {
1644 if (alock->is_Lock()) {
1645 tty->print_cr("++++ Eliminating: %d Lock", alock->_idx);
1646 } else {
1647 tty->print_cr("++++ Eliminating: %d Unlock", alock->_idx);
1648 }
1649 }
1650 #endif
1652 Node* mem = alock->in(TypeFunc::Memory);
1653 Node* ctrl = alock->in(TypeFunc::Control);
1655 extract_call_projections(alock);
1656 // There are 2 projections from the lock. The lock node will
1657 // be deleted when its last use is subsumed below.
1658 assert(alock->outcnt() == 2 &&
1659 _fallthroughproj != NULL &&
1660 _memproj_fallthrough != NULL,
1661 "Unexpected projections from Lock/Unlock");
1663 Node* fallthroughproj = _fallthroughproj;
1664 Node* memproj_fallthrough = _memproj_fallthrough;
1666 // The memory projection from a lock/unlock is RawMem
1667 // The input to a Lock is merged memory, so extract its RawMem input
1668 // (unless the MergeMem has been optimized away.)
1669 if (alock->is_Lock()) {
1670 // Seach for MemBarAcquire node and delete it also.
1671 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
1672 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquire, "");
1673 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
1674 Node* memproj = membar->proj_out(TypeFunc::Memory);
1675 _igvn.replace_node(ctrlproj, fallthroughproj);
1676 _igvn.replace_node(memproj, memproj_fallthrough);
1678 // Delete FastLock node also if this Lock node is unique user
1679 // (a loop peeling may clone a Lock node).
1680 Node* flock = alock->as_Lock()->fastlock_node();
1681 if (flock->outcnt() == 1) {
1682 assert(flock->unique_out() == alock, "sanity");
1683 _igvn.replace_node(flock, top());
1684 }
1685 }
1687 // Seach for MemBarRelease node and delete it also.
1688 if (alock->is_Unlock() && ctrl != NULL && ctrl->is_Proj() &&
1689 ctrl->in(0)->is_MemBar()) {
1690 MemBarNode* membar = ctrl->in(0)->as_MemBar();
1691 assert(membar->Opcode() == Op_MemBarRelease &&
1692 mem->is_Proj() && membar == mem->in(0), "");
1693 _igvn.replace_node(fallthroughproj, ctrl);
1694 _igvn.replace_node(memproj_fallthrough, mem);
1695 fallthroughproj = ctrl;
1696 memproj_fallthrough = mem;
1697 ctrl = membar->in(TypeFunc::Control);
1698 mem = membar->in(TypeFunc::Memory);
1699 }
1701 _igvn.replace_node(fallthroughproj, ctrl);
1702 _igvn.replace_node(memproj_fallthrough, mem);
1703 return true;
1704 }
1707 //------------------------------expand_lock_node----------------------
1708 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
1710 Node* ctrl = lock->in(TypeFunc::Control);
1711 Node* mem = lock->in(TypeFunc::Memory);
1712 Node* obj = lock->obj_node();
1713 Node* box = lock->box_node();
1714 Node* flock = lock->fastlock_node();
1716 // Make the merge point
1717 Node *region;
1718 Node *mem_phi;
1719 Node *slow_path;
1721 if (UseOptoBiasInlining) {
1722 /*
1723 * See the full description in MacroAssembler::biased_locking_enter().
1724 *
1725 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) {
1726 * // The object is biased.
1727 * proto_node = klass->prototype_header;
1728 * o_node = thread | proto_node;
1729 * x_node = o_node ^ mark_word;
1730 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ?
1731 * // Done.
1732 * } else {
1733 * if( (x_node & biased_lock_mask) != 0 ) {
1734 * // The klass's prototype header is no longer biased.
1735 * cas(&mark_word, mark_word, proto_node)
1736 * goto cas_lock;
1737 * } else {
1738 * // The klass's prototype header is still biased.
1739 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch?
1740 * old = mark_word;
1741 * new = o_node;
1742 * } else {
1743 * // Different thread or anonymous biased.
1744 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask);
1745 * new = thread | old;
1746 * }
1747 * // Try to rebias.
1748 * if( cas(&mark_word, old, new) == 0 ) {
1749 * // Done.
1750 * } else {
1751 * goto slow_path; // Failed.
1752 * }
1753 * }
1754 * }
1755 * } else {
1756 * // The object is not biased.
1757 * cas_lock:
1758 * if( FastLock(obj) == 0 ) {
1759 * // Done.
1760 * } else {
1761 * slow_path:
1762 * OptoRuntime::complete_monitor_locking_Java(obj);
1763 * }
1764 * }
1765 */
1767 region = new (C, 5) RegionNode(5);
1768 // create a Phi for the memory state
1769 mem_phi = new (C, 5) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
1771 Node* fast_lock_region = new (C, 3) RegionNode(3);
1772 Node* fast_lock_mem_phi = new (C, 3) PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1774 // First, check mark word for the biased lock pattern.
1775 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
1777 // Get fast path - mark word has the biased lock pattern.
1778 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node,
1779 markOopDesc::biased_lock_mask_in_place,
1780 markOopDesc::biased_lock_pattern, true);
1781 // fast_lock_region->in(1) is set to slow path.
1782 fast_lock_mem_phi->init_req(1, mem);
1784 // Now check that the lock is biased to the current thread and has
1785 // the same epoch and bias as Klass::_prototype_header.
1787 // Special-case a fresh allocation to avoid building nodes:
1788 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn);
1789 if (klass_node == NULL) {
1790 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1791 klass_node = transform_later( LoadKlassNode::make(_igvn, mem, k_adr, _igvn.type(k_adr)->is_ptr()) );
1792 #ifdef _LP64
1793 if (UseCompressedOops && klass_node->is_DecodeN()) {
1794 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity");
1795 klass_node->in(1)->init_req(0, ctrl);
1796 } else
1797 #endif
1798 klass_node->init_req(0, ctrl);
1799 }
1800 Node *proto_node = make_load(ctrl, mem, klass_node, Klass::prototype_header_offset_in_bytes() + sizeof(oopDesc), TypeX_X, TypeX_X->basic_type());
1802 Node* thread = transform_later(new (C, 1) ThreadLocalNode());
1803 Node* cast_thread = transform_later(new (C, 2) CastP2XNode(ctrl, thread));
1804 Node* o_node = transform_later(new (C, 3) OrXNode(cast_thread, proto_node));
1805 Node* x_node = transform_later(new (C, 3) XorXNode(o_node, mark_node));
1807 // Get slow path - mark word does NOT match the value.
1808 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node,
1809 (~markOopDesc::age_mask_in_place), 0);
1810 // region->in(3) is set to fast path - the object is biased to the current thread.
1811 mem_phi->init_req(3, mem);
1814 // Mark word does NOT match the value (thread | Klass::_prototype_header).
1817 // First, check biased pattern.
1818 // Get fast path - _prototype_header has the same biased lock pattern.
1819 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node,
1820 markOopDesc::biased_lock_mask_in_place, 0, true);
1822 not_biased_ctrl = fast_lock_region->in(2); // Slow path
1823 // fast_lock_region->in(2) - the prototype header is no longer biased
1824 // and we have to revoke the bias on this object.
1825 // We are going to try to reset the mark of this object to the prototype
1826 // value and fall through to the CAS-based locking scheme.
1827 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
1828 Node* cas = new (C, 5) StoreXConditionalNode(not_biased_ctrl, mem, adr,
1829 proto_node, mark_node);
1830 transform_later(cas);
1831 Node* proj = transform_later( new (C, 1) SCMemProjNode(cas));
1832 fast_lock_mem_phi->init_req(2, proj);
1835 // Second, check epoch bits.
1836 Node* rebiased_region = new (C, 3) RegionNode(3);
1837 Node* old_phi = new (C, 3) PhiNode( rebiased_region, TypeX_X);
1838 Node* new_phi = new (C, 3) PhiNode( rebiased_region, TypeX_X);
1840 // Get slow path - mark word does NOT match epoch bits.
1841 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node,
1842 markOopDesc::epoch_mask_in_place, 0);
1843 // The epoch of the current bias is not valid, attempt to rebias the object
1844 // toward the current thread.
1845 rebiased_region->init_req(2, epoch_ctrl);
1846 old_phi->init_req(2, mark_node);
1847 new_phi->init_req(2, o_node);
1849 // rebiased_region->in(1) is set to fast path.
1850 // The epoch of the current bias is still valid but we know
1851 // nothing about the owner; it might be set or it might be clear.
1852 Node* cmask = MakeConX(markOopDesc::biased_lock_mask_in_place |
1853 markOopDesc::age_mask_in_place |
1854 markOopDesc::epoch_mask_in_place);
1855 Node* old = transform_later(new (C, 3) AndXNode(mark_node, cmask));
1856 cast_thread = transform_later(new (C, 2) CastP2XNode(ctrl, thread));
1857 Node* new_mark = transform_later(new (C, 3) OrXNode(cast_thread, old));
1858 old_phi->init_req(1, old);
1859 new_phi->init_req(1, new_mark);
1861 transform_later(rebiased_region);
1862 transform_later(old_phi);
1863 transform_later(new_phi);
1865 // Try to acquire the bias of the object using an atomic operation.
1866 // If this fails we will go in to the runtime to revoke the object's bias.
1867 cas = new (C, 5) StoreXConditionalNode(rebiased_region, mem, adr,
1868 new_phi, old_phi);
1869 transform_later(cas);
1870 proj = transform_later( new (C, 1) SCMemProjNode(cas));
1872 // Get slow path - Failed to CAS.
1873 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0);
1874 mem_phi->init_req(4, proj);
1875 // region->in(4) is set to fast path - the object is rebiased to the current thread.
1877 // Failed to CAS.
1878 slow_path = new (C, 3) RegionNode(3);
1879 Node *slow_mem = new (C, 3) PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM);
1881 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control
1882 slow_mem->init_req(1, proj);
1884 // Call CAS-based locking scheme (FastLock node).
1886 transform_later(fast_lock_region);
1887 transform_later(fast_lock_mem_phi);
1889 // Get slow path - FastLock failed to lock the object.
1890 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0);
1891 mem_phi->init_req(2, fast_lock_mem_phi);
1892 // region->in(2) is set to fast path - the object is locked to the current thread.
1894 slow_path->init_req(2, ctrl); // Capture slow-control
1895 slow_mem->init_req(2, fast_lock_mem_phi);
1897 transform_later(slow_path);
1898 transform_later(slow_mem);
1899 // Reset lock's memory edge.
1900 lock->set_req(TypeFunc::Memory, slow_mem);
1902 } else {
1903 region = new (C, 3) RegionNode(3);
1904 // create a Phi for the memory state
1905 mem_phi = new (C, 3) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
1907 // Optimize test; set region slot 2
1908 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
1909 mem_phi->init_req(2, mem);
1910 }
1912 // Make slow path call
1913 CallNode *call = make_slow_call( (CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(), OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path, obj, box );
1915 extract_call_projections(call);
1917 // Slow path can only throw asynchronous exceptions, which are always
1918 // de-opted. So the compiler thinks the slow-call can never throw an
1919 // exception. If it DOES throw an exception we would need the debug
1920 // info removed first (since if it throws there is no monitor).
1921 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
1922 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
1924 // Capture slow path
1925 // disconnect fall-through projection from call and create a new one
1926 // hook up users of fall-through projection to region
1927 Node *slow_ctrl = _fallthroughproj->clone();
1928 transform_later(slow_ctrl);
1929 _igvn.hash_delete(_fallthroughproj);
1930 _fallthroughproj->disconnect_inputs(NULL);
1931 region->init_req(1, slow_ctrl);
1932 // region inputs are now complete
1933 transform_later(region);
1934 _igvn.replace_node(_fallthroughproj, region);
1936 Node *memproj = transform_later( new(C, 1) ProjNode(call, TypeFunc::Memory) );
1937 mem_phi->init_req(1, memproj );
1938 transform_later(mem_phi);
1939 _igvn.replace_node(_memproj_fallthrough, mem_phi);
1940 }
1942 //------------------------------expand_unlock_node----------------------
1943 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
1945 Node* ctrl = unlock->in(TypeFunc::Control);
1946 Node* mem = unlock->in(TypeFunc::Memory);
1947 Node* obj = unlock->obj_node();
1948 Node* box = unlock->box_node();
1950 // No need for a null check on unlock
1952 // Make the merge point
1953 Node *region;
1954 Node *mem_phi;
1956 if (UseOptoBiasInlining) {
1957 // Check for biased locking unlock case, which is a no-op.
1958 // See the full description in MacroAssembler::biased_locking_exit().
1959 region = new (C, 4) RegionNode(4);
1960 // create a Phi for the memory state
1961 mem_phi = new (C, 4) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
1962 mem_phi->init_req(3, mem);
1964 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
1965 ctrl = opt_bits_test(ctrl, region, 3, mark_node,
1966 markOopDesc::biased_lock_mask_in_place,
1967 markOopDesc::biased_lock_pattern);
1968 } else {
1969 region = new (C, 3) RegionNode(3);
1970 // create a Phi for the memory state
1971 mem_phi = new (C, 3) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
1972 }
1974 FastUnlockNode *funlock = new (C, 3) FastUnlockNode( ctrl, obj, box );
1975 funlock = transform_later( funlock )->as_FastUnlock();
1976 // Optimize test; set region slot 2
1977 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
1979 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 );
1981 extract_call_projections(call);
1983 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
1984 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
1986 // No exceptions for unlocking
1987 // Capture slow path
1988 // disconnect fall-through projection from call and create a new one
1989 // hook up users of fall-through projection to region
1990 Node *slow_ctrl = _fallthroughproj->clone();
1991 transform_later(slow_ctrl);
1992 _igvn.hash_delete(_fallthroughproj);
1993 _fallthroughproj->disconnect_inputs(NULL);
1994 region->init_req(1, slow_ctrl);
1995 // region inputs are now complete
1996 transform_later(region);
1997 _igvn.replace_node(_fallthroughproj, region);
1999 Node *memproj = transform_later( new(C, 1) ProjNode(call, TypeFunc::Memory) );
2000 mem_phi->init_req(1, memproj );
2001 mem_phi->init_req(2, mem);
2002 transform_later(mem_phi);
2003 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2004 }
2006 //------------------------------expand_macro_nodes----------------------
2007 // Returns true if a failure occurred.
2008 bool PhaseMacroExpand::expand_macro_nodes() {
2009 if (C->macro_count() == 0)
2010 return false;
2011 // First, attempt to eliminate locks
2012 bool progress = true;
2013 while (progress) {
2014 progress = false;
2015 for (int i = C->macro_count(); i > 0; i--) {
2016 Node * n = C->macro_node(i-1);
2017 bool success = false;
2018 debug_only(int old_macro_count = C->macro_count(););
2019 if (n->is_AbstractLock()) {
2020 success = eliminate_locking_node(n->as_AbstractLock());
2021 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) {
2022 _igvn.replace_node(n, n->in(1));
2023 success = true;
2024 }
2025 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2026 progress = progress || success;
2027 }
2028 }
2029 // Next, attempt to eliminate allocations
2030 progress = true;
2031 while (progress) {
2032 progress = false;
2033 for (int i = C->macro_count(); i > 0; i--) {
2034 Node * n = C->macro_node(i-1);
2035 bool success = false;
2036 debug_only(int old_macro_count = C->macro_count(););
2037 switch (n->class_id()) {
2038 case Node::Class_Allocate:
2039 case Node::Class_AllocateArray:
2040 success = eliminate_allocate_node(n->as_Allocate());
2041 break;
2042 case Node::Class_Lock:
2043 case Node::Class_Unlock:
2044 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2045 break;
2046 default:
2047 assert(false, "unknown node type in macro list");
2048 }
2049 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2050 progress = progress || success;
2051 }
2052 }
2053 // Make sure expansion will not cause node limit to be exceeded.
2054 // Worst case is a macro node gets expanded into about 50 nodes.
2055 // Allow 50% more for optimization.
2056 if (C->check_node_count(C->macro_count() * 75, "out of nodes before macro expansion" ) )
2057 return true;
2059 // expand "macro" nodes
2060 // nodes are removed from the macro list as they are processed
2061 while (C->macro_count() > 0) {
2062 int macro_count = C->macro_count();
2063 Node * n = C->macro_node(macro_count-1);
2064 assert(n->is_macro(), "only macro nodes expected here");
2065 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) {
2066 // node is unreachable, so don't try to expand it
2067 C->remove_macro_node(n);
2068 continue;
2069 }
2070 switch (n->class_id()) {
2071 case Node::Class_Allocate:
2072 expand_allocate(n->as_Allocate());
2073 break;
2074 case Node::Class_AllocateArray:
2075 expand_allocate_array(n->as_AllocateArray());
2076 break;
2077 case Node::Class_Lock:
2078 expand_lock_node(n->as_Lock());
2079 break;
2080 case Node::Class_Unlock:
2081 expand_unlock_node(n->as_Unlock());
2082 break;
2083 default:
2084 assert(false, "unknown node type in macro list");
2085 }
2086 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2087 if (C->failing()) return true;
2088 }
2090 _igvn.set_delay_transform(false);
2091 _igvn.optimize();
2092 return false;
2093 }