Wed, 28 Aug 2013 11:22:43 +0200
8023597: Optimize G1 barriers code for unsafe load_store
Summary: Avoid loading old values in G1 pre-barriers for inlined unsafe load_store nodes.
Reviewed-by: kvn, tonyp
Contributed-by: Martin Doerr <martin.doerr@sap.com>
1 /*
2 * Copyright (c) 2005, 2012, Oracle and/or its affiliates. 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 Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #include "precompiled.hpp"
26 #include "compiler/compileLog.hpp"
27 #include "libadt/vectset.hpp"
28 #include "opto/addnode.hpp"
29 #include "opto/callnode.hpp"
30 #include "opto/cfgnode.hpp"
31 #include "opto/compile.hpp"
32 #include "opto/connode.hpp"
33 #include "opto/locknode.hpp"
34 #include "opto/loopnode.hpp"
35 #include "opto/macro.hpp"
36 #include "opto/memnode.hpp"
37 #include "opto/node.hpp"
38 #include "opto/phaseX.hpp"
39 #include "opto/rootnode.hpp"
40 #include "opto/runtime.hpp"
41 #include "opto/subnode.hpp"
42 #include "opto/type.hpp"
43 #include "runtime/sharedRuntime.hpp"
46 //
47 // Replace any references to "oldref" in inputs to "use" with "newref".
48 // Returns the number of replacements made.
49 //
50 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
51 int nreplacements = 0;
52 uint req = use->req();
53 for (uint j = 0; j < use->len(); j++) {
54 Node *uin = use->in(j);
55 if (uin == oldref) {
56 if (j < req)
57 use->set_req(j, newref);
58 else
59 use->set_prec(j, newref);
60 nreplacements++;
61 } else if (j >= req && uin == NULL) {
62 break;
63 }
64 }
65 return nreplacements;
66 }
68 void PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) {
69 // Copy debug information and adjust JVMState information
70 uint old_dbg_start = oldcall->tf()->domain()->cnt();
71 uint new_dbg_start = newcall->tf()->domain()->cnt();
72 int jvms_adj = new_dbg_start - old_dbg_start;
73 assert (new_dbg_start == newcall->req(), "argument count mismatch");
75 // SafePointScalarObject node could be referenced several times in debug info.
76 // Use Dict to record cloned nodes.
77 Dict* sosn_map = new Dict(cmpkey,hashkey);
78 for (uint i = old_dbg_start; i < oldcall->req(); i++) {
79 Node* old_in = oldcall->in(i);
80 // Clone old SafePointScalarObjectNodes, adjusting their field contents.
81 if (old_in != NULL && old_in->is_SafePointScalarObject()) {
82 SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject();
83 uint old_unique = C->unique();
84 Node* new_in = old_sosn->clone(sosn_map);
85 if (old_unique != C->unique()) { // New node?
86 new_in->set_req(0, C->root()); // reset control edge
87 new_in = transform_later(new_in); // Register new node.
88 }
89 old_in = new_in;
90 }
91 newcall->add_req(old_in);
92 }
94 newcall->set_jvms(oldcall->jvms());
95 for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) {
96 jvms->set_map(newcall);
97 jvms->set_locoff(jvms->locoff()+jvms_adj);
98 jvms->set_stkoff(jvms->stkoff()+jvms_adj);
99 jvms->set_monoff(jvms->monoff()+jvms_adj);
100 jvms->set_scloff(jvms->scloff()+jvms_adj);
101 jvms->set_endoff(jvms->endoff()+jvms_adj);
102 }
103 }
105 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
106 Node* cmp;
107 if (mask != 0) {
108 Node* and_node = transform_later(new (C) AndXNode(word, MakeConX(mask)));
109 cmp = transform_later(new (C) CmpXNode(and_node, MakeConX(bits)));
110 } else {
111 cmp = word;
112 }
113 Node* bol = transform_later(new (C) BoolNode(cmp, BoolTest::ne));
114 IfNode* iff = new (C) IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
115 transform_later(iff);
117 // Fast path taken.
118 Node *fast_taken = transform_later( new (C) IfFalseNode(iff) );
120 // Fast path not-taken, i.e. slow path
121 Node *slow_taken = transform_later( new (C) IfTrueNode(iff) );
123 if (return_fast_path) {
124 region->init_req(edge, slow_taken); // Capture slow-control
125 return fast_taken;
126 } else {
127 region->init_req(edge, fast_taken); // Capture fast-control
128 return slow_taken;
129 }
130 }
132 //--------------------copy_predefined_input_for_runtime_call--------------------
133 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
134 // Set fixed predefined input arguments
135 call->init_req( TypeFunc::Control, ctrl );
136 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) );
137 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
138 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
139 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
140 }
142 //------------------------------make_slow_call---------------------------------
143 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) {
145 // Slow-path call
146 CallNode *call = leaf_name
147 ? (CallNode*)new (C) CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
148 : (CallNode*)new (C) CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM );
150 // Slow path call has no side-effects, uses few values
151 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
152 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0);
153 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1);
154 copy_call_debug_info(oldcall, call);
155 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
156 _igvn.replace_node(oldcall, call);
157 transform_later(call);
159 return call;
160 }
162 void PhaseMacroExpand::extract_call_projections(CallNode *call) {
163 _fallthroughproj = NULL;
164 _fallthroughcatchproj = NULL;
165 _ioproj_fallthrough = NULL;
166 _ioproj_catchall = NULL;
167 _catchallcatchproj = NULL;
168 _memproj_fallthrough = NULL;
169 _memproj_catchall = NULL;
170 _resproj = NULL;
171 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) {
172 ProjNode *pn = call->fast_out(i)->as_Proj();
173 switch (pn->_con) {
174 case TypeFunc::Control:
175 {
176 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj
177 _fallthroughproj = pn;
178 DUIterator_Fast jmax, j = pn->fast_outs(jmax);
179 const Node *cn = pn->fast_out(j);
180 if (cn->is_Catch()) {
181 ProjNode *cpn = NULL;
182 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) {
183 cpn = cn->fast_out(k)->as_Proj();
184 assert(cpn->is_CatchProj(), "must be a CatchProjNode");
185 if (cpn->_con == CatchProjNode::fall_through_index)
186 _fallthroughcatchproj = cpn;
187 else {
188 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index.");
189 _catchallcatchproj = cpn;
190 }
191 }
192 }
193 break;
194 }
195 case TypeFunc::I_O:
196 if (pn->_is_io_use)
197 _ioproj_catchall = pn;
198 else
199 _ioproj_fallthrough = pn;
200 break;
201 case TypeFunc::Memory:
202 if (pn->_is_io_use)
203 _memproj_catchall = pn;
204 else
205 _memproj_fallthrough = pn;
206 break;
207 case TypeFunc::Parms:
208 _resproj = pn;
209 break;
210 default:
211 assert(false, "unexpected projection from allocation node.");
212 }
213 }
215 }
217 // Eliminate a card mark sequence. p2x is a ConvP2XNode
218 void PhaseMacroExpand::eliminate_card_mark(Node* p2x) {
219 assert(p2x->Opcode() == Op_CastP2X, "ConvP2XNode required");
220 if (!UseG1GC) {
221 // vanilla/CMS post barrier
222 Node *shift = p2x->unique_out();
223 Node *addp = shift->unique_out();
224 for (DUIterator_Last jmin, j = addp->last_outs(jmin); j >= jmin; --j) {
225 Node *mem = addp->last_out(j);
226 if (UseCondCardMark && mem->is_Load()) {
227 assert(mem->Opcode() == Op_LoadB, "unexpected code shape");
228 // The load is checking if the card has been written so
229 // replace it with zero to fold the test.
230 _igvn.replace_node(mem, intcon(0));
231 continue;
232 }
233 assert(mem->is_Store(), "store required");
234 _igvn.replace_node(mem, mem->in(MemNode::Memory));
235 }
236 } else {
237 // G1 pre/post barriers
238 assert(p2x->outcnt() <= 2, "expects 1 or 2 users: Xor and URShift nodes");
239 // It could be only one user, URShift node, in Object.clone() instrinsic
240 // but the new allocation is passed to arraycopy stub and it could not
241 // be scalar replaced. So we don't check the case.
243 // An other case of only one user (Xor) is when the value check for NULL
244 // in G1 post barrier is folded after CCP so the code which used URShift
245 // is removed.
247 // Take Region node before eliminating post barrier since it also
248 // eliminates CastP2X node when it has only one user.
249 Node* this_region = p2x->in(0);
250 assert(this_region != NULL, "");
252 // Remove G1 post barrier.
254 // Search for CastP2X->Xor->URShift->Cmp path which
255 // checks if the store done to a different from the value's region.
256 // And replace Cmp with #0 (false) to collapse G1 post barrier.
257 Node* xorx = NULL;
258 for (DUIterator_Fast imax, i = p2x->fast_outs(imax); i < imax; i++) {
259 Node* u = p2x->fast_out(i);
260 if (u->Opcode() == Op_XorX) {
261 xorx = u;
262 break;
263 }
264 }
265 assert(xorx != NULL, "missing G1 post barrier");
266 Node* shift = xorx->unique_out();
267 Node* cmpx = shift->unique_out();
268 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() &&
269 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne,
270 "missing region check in G1 post barrier");
271 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ));
273 // Remove G1 pre barrier.
275 // Search "if (marking != 0)" check and set it to "false".
276 // There is no G1 pre barrier if previous stored value is NULL
277 // (for example, after initialization).
278 if (this_region->is_Region() && this_region->req() == 3) {
279 int ind = 1;
280 if (!this_region->in(ind)->is_IfFalse()) {
281 ind = 2;
282 }
283 if (this_region->in(ind)->is_IfFalse()) {
284 Node* bol = this_region->in(ind)->in(0)->in(1);
285 assert(bol->is_Bool(), "");
286 cmpx = bol->in(1);
287 if (bol->as_Bool()->_test._test == BoolTest::ne &&
288 cmpx->is_Cmp() && cmpx->in(2) == intcon(0) &&
289 cmpx->in(1)->is_Load()) {
290 Node* adr = cmpx->in(1)->as_Load()->in(MemNode::Address);
291 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() +
292 PtrQueue::byte_offset_of_active());
293 if (adr->is_AddP() && adr->in(AddPNode::Base) == top() &&
294 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
295 adr->in(AddPNode::Offset) == MakeConX(marking_offset)) {
296 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ));
297 }
298 }
299 }
300 }
301 // Now CastP2X can be removed since it is used only on dead path
302 // which currently still alive until igvn optimize it.
303 assert(p2x->outcnt() == 0 || p2x->unique_out()->Opcode() == Op_URShiftX, "");
304 _igvn.replace_node(p2x, top());
305 }
306 }
308 // Search for a memory operation for the specified memory slice.
309 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
310 Node *orig_mem = mem;
311 Node *alloc_mem = alloc->in(TypeFunc::Memory);
312 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
313 while (true) {
314 if (mem == alloc_mem || mem == start_mem ) {
315 return mem; // hit one of our sentinels
316 } else if (mem->is_MergeMem()) {
317 mem = mem->as_MergeMem()->memory_at(alias_idx);
318 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
319 Node *in = mem->in(0);
320 // we can safely skip over safepoints, calls, locks and membars because we
321 // already know that the object is safe to eliminate.
322 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
323 return in;
324 } else if (in->is_Call()) {
325 CallNode *call = in->as_Call();
326 if (!call->may_modify(tinst, phase)) {
327 mem = call->in(TypeFunc::Memory);
328 }
329 mem = in->in(TypeFunc::Memory);
330 } else if (in->is_MemBar()) {
331 mem = in->in(TypeFunc::Memory);
332 } else {
333 assert(false, "unexpected projection");
334 }
335 } else if (mem->is_Store()) {
336 const TypePtr* atype = mem->as_Store()->adr_type();
337 int adr_idx = Compile::current()->get_alias_index(atype);
338 if (adr_idx == alias_idx) {
339 assert(atype->isa_oopptr(), "address type must be oopptr");
340 int adr_offset = atype->offset();
341 uint adr_iid = atype->is_oopptr()->instance_id();
342 // Array elements references have the same alias_idx
343 // but different offset and different instance_id.
344 if (adr_offset == offset && adr_iid == alloc->_idx)
345 return mem;
346 } else {
347 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
348 }
349 mem = mem->in(MemNode::Memory);
350 } else if (mem->is_ClearArray()) {
351 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
352 // Can not bypass initialization of the instance
353 // we are looking.
354 debug_only(intptr_t offset;)
355 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
356 InitializeNode* init = alloc->as_Allocate()->initialization();
357 // We are looking for stored value, return Initialize node
358 // or memory edge from Allocate node.
359 if (init != NULL)
360 return init;
361 else
362 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers).
363 }
364 // Otherwise skip it (the call updated 'mem' value).
365 } else if (mem->Opcode() == Op_SCMemProj) {
366 mem = mem->in(0);
367 Node* adr = NULL;
368 if (mem->is_LoadStore()) {
369 adr = mem->in(MemNode::Address);
370 } else {
371 assert(mem->Opcode() == Op_EncodeISOArray, "sanity");
372 adr = mem->in(3); // Destination array
373 }
374 const TypePtr* atype = adr->bottom_type()->is_ptr();
375 int adr_idx = Compile::current()->get_alias_index(atype);
376 if (adr_idx == alias_idx) {
377 assert(false, "Object is not scalar replaceable if a LoadStore node access its field");
378 return NULL;
379 }
380 mem = mem->in(MemNode::Memory);
381 } else {
382 return mem;
383 }
384 assert(mem != orig_mem, "dead memory loop");
385 }
386 }
388 //
389 // Given a Memory Phi, compute a value Phi containing the values from stores
390 // on the input paths.
391 // Note: this function is recursive, its depth is limied by the "level" argument
392 // Returns the computed Phi, or NULL if it cannot compute it.
393 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) {
394 assert(mem->is_Phi(), "sanity");
395 int alias_idx = C->get_alias_index(adr_t);
396 int offset = adr_t->offset();
397 int instance_id = adr_t->instance_id();
399 // Check if an appropriate value phi already exists.
400 Node* region = mem->in(0);
401 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
402 Node* phi = region->fast_out(k);
403 if (phi->is_Phi() && phi != mem &&
404 phi->as_Phi()->is_same_inst_field(phi_type, instance_id, alias_idx, offset)) {
405 return phi;
406 }
407 }
408 // Check if an appropriate new value phi already exists.
409 Node* new_phi = value_phis->find(mem->_idx);
410 if (new_phi != NULL)
411 return new_phi;
413 if (level <= 0) {
414 return NULL; // Give up: phi tree too deep
415 }
416 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
417 Node *alloc_mem = alloc->in(TypeFunc::Memory);
419 uint length = mem->req();
420 GrowableArray <Node *> values(length, length, NULL, false);
422 // create a new Phi for the value
423 PhiNode *phi = new (C) PhiNode(mem->in(0), phi_type, NULL, instance_id, alias_idx, offset);
424 transform_later(phi);
425 value_phis->push(phi, mem->_idx);
427 for (uint j = 1; j < length; j++) {
428 Node *in = mem->in(j);
429 if (in == NULL || in->is_top()) {
430 values.at_put(j, in);
431 } else {
432 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
433 if (val == start_mem || val == alloc_mem) {
434 // hit a sentinel, return appropriate 0 value
435 values.at_put(j, _igvn.zerocon(ft));
436 continue;
437 }
438 if (val->is_Initialize()) {
439 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
440 }
441 if (val == NULL) {
442 return NULL; // can't find a value on this path
443 }
444 if (val == mem) {
445 values.at_put(j, mem);
446 } else if (val->is_Store()) {
447 values.at_put(j, val->in(MemNode::ValueIn));
448 } else if(val->is_Proj() && val->in(0) == alloc) {
449 values.at_put(j, _igvn.zerocon(ft));
450 } else if (val->is_Phi()) {
451 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
452 if (val == NULL) {
453 return NULL;
454 }
455 values.at_put(j, val);
456 } else if (val->Opcode() == Op_SCMemProj) {
457 assert(val->in(0)->is_LoadStore() || val->in(0)->Opcode() == Op_EncodeISOArray, "sanity");
458 assert(false, "Object is not scalar replaceable if a LoadStore node access its field");
459 return NULL;
460 } else {
461 #ifdef ASSERT
462 val->dump();
463 assert(false, "unknown node on this path");
464 #endif
465 return NULL; // unknown node on this path
466 }
467 }
468 }
469 // Set Phi's inputs
470 for (uint j = 1; j < length; j++) {
471 if (values.at(j) == mem) {
472 phi->init_req(j, phi);
473 } else {
474 phi->init_req(j, values.at(j));
475 }
476 }
477 return phi;
478 }
480 // Search the last value stored into the object's field.
481 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, Node *alloc) {
482 assert(adr_t->is_known_instance_field(), "instance required");
483 int instance_id = adr_t->instance_id();
484 assert((uint)instance_id == alloc->_idx, "wrong allocation");
486 int alias_idx = C->get_alias_index(adr_t);
487 int offset = adr_t->offset();
488 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
489 Node *alloc_ctrl = alloc->in(TypeFunc::Control);
490 Node *alloc_mem = alloc->in(TypeFunc::Memory);
491 Arena *a = Thread::current()->resource_area();
492 VectorSet visited(a);
495 bool done = sfpt_mem == alloc_mem;
496 Node *mem = sfpt_mem;
497 while (!done) {
498 if (visited.test_set(mem->_idx)) {
499 return NULL; // found a loop, give up
500 }
501 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
502 if (mem == start_mem || mem == alloc_mem) {
503 done = true; // hit a sentinel, return appropriate 0 value
504 } else if (mem->is_Initialize()) {
505 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
506 if (mem == NULL) {
507 done = true; // Something go wrong.
508 } else if (mem->is_Store()) {
509 const TypePtr* atype = mem->as_Store()->adr_type();
510 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
511 done = true;
512 }
513 } else if (mem->is_Store()) {
514 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
515 assert(atype != NULL, "address type must be oopptr");
516 assert(C->get_alias_index(atype) == alias_idx &&
517 atype->is_known_instance_field() && atype->offset() == offset &&
518 atype->instance_id() == instance_id, "store is correct memory slice");
519 done = true;
520 } else if (mem->is_Phi()) {
521 // try to find a phi's unique input
522 Node *unique_input = NULL;
523 Node *top = C->top();
524 for (uint i = 1; i < mem->req(); i++) {
525 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
526 if (n == NULL || n == top || n == mem) {
527 continue;
528 } else if (unique_input == NULL) {
529 unique_input = n;
530 } else if (unique_input != n) {
531 unique_input = top;
532 break;
533 }
534 }
535 if (unique_input != NULL && unique_input != top) {
536 mem = unique_input;
537 } else {
538 done = true;
539 }
540 } else {
541 assert(false, "unexpected node");
542 }
543 }
544 if (mem != NULL) {
545 if (mem == start_mem || mem == alloc_mem) {
546 // hit a sentinel, return appropriate 0 value
547 return _igvn.zerocon(ft);
548 } else if (mem->is_Store()) {
549 return mem->in(MemNode::ValueIn);
550 } else if (mem->is_Phi()) {
551 // attempt to produce a Phi reflecting the values on the input paths of the Phi
552 Node_Stack value_phis(a, 8);
553 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
554 if (phi != NULL) {
555 return phi;
556 } else {
557 // Kill all new Phis
558 while(value_phis.is_nonempty()) {
559 Node* n = value_phis.node();
560 _igvn.replace_node(n, C->top());
561 value_phis.pop();
562 }
563 }
564 }
565 }
566 // Something go wrong.
567 return NULL;
568 }
570 // Check the possibility of scalar replacement.
571 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
572 // Scan the uses of the allocation to check for anything that would
573 // prevent us from eliminating it.
574 NOT_PRODUCT( const char* fail_eliminate = NULL; )
575 DEBUG_ONLY( Node* disq_node = NULL; )
576 bool can_eliminate = true;
578 Node* res = alloc->result_cast();
579 const TypeOopPtr* res_type = NULL;
580 if (res == NULL) {
581 // All users were eliminated.
582 } else if (!res->is_CheckCastPP()) {
583 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
584 can_eliminate = false;
585 } else {
586 res_type = _igvn.type(res)->isa_oopptr();
587 if (res_type == NULL) {
588 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
589 can_eliminate = false;
590 } else if (res_type->isa_aryptr()) {
591 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
592 if (length < 0) {
593 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
594 can_eliminate = false;
595 }
596 }
597 }
599 if (can_eliminate && res != NULL) {
600 for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
601 j < jmax && can_eliminate; j++) {
602 Node* use = res->fast_out(j);
604 if (use->is_AddP()) {
605 const TypePtr* addp_type = _igvn.type(use)->is_ptr();
606 int offset = addp_type->offset();
608 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
609 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";)
610 can_eliminate = false;
611 break;
612 }
613 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
614 k < kmax && can_eliminate; k++) {
615 Node* n = use->fast_out(k);
616 if (!n->is_Store() && n->Opcode() != Op_CastP2X) {
617 DEBUG_ONLY(disq_node = n;)
618 if (n->is_Load() || n->is_LoadStore()) {
619 NOT_PRODUCT(fail_eliminate = "Field load";)
620 } else {
621 NOT_PRODUCT(fail_eliminate = "Not store field referrence";)
622 }
623 can_eliminate = false;
624 }
625 }
626 } else if (use->is_SafePoint()) {
627 SafePointNode* sfpt = use->as_SafePoint();
628 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
629 // Object is passed as argument.
630 DEBUG_ONLY(disq_node = use;)
631 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
632 can_eliminate = false;
633 }
634 Node* sfptMem = sfpt->memory();
635 if (sfptMem == NULL || sfptMem->is_top()) {
636 DEBUG_ONLY(disq_node = use;)
637 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";)
638 can_eliminate = false;
639 } else {
640 safepoints.append_if_missing(sfpt);
641 }
642 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
643 if (use->is_Phi()) {
644 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
645 NOT_PRODUCT(fail_eliminate = "Object is return value";)
646 } else {
647 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
648 }
649 DEBUG_ONLY(disq_node = use;)
650 } else {
651 if (use->Opcode() == Op_Return) {
652 NOT_PRODUCT(fail_eliminate = "Object is return value";)
653 }else {
654 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
655 }
656 DEBUG_ONLY(disq_node = use;)
657 }
658 can_eliminate = false;
659 }
660 }
661 }
663 #ifndef PRODUCT
664 if (PrintEliminateAllocations) {
665 if (can_eliminate) {
666 tty->print("Scalar ");
667 if (res == NULL)
668 alloc->dump();
669 else
670 res->dump();
671 } else if (alloc->_is_scalar_replaceable) {
672 tty->print("NotScalar (%s)", fail_eliminate);
673 if (res == NULL)
674 alloc->dump();
675 else
676 res->dump();
677 #ifdef ASSERT
678 if (disq_node != NULL) {
679 tty->print(" >>>> ");
680 disq_node->dump();
681 }
682 #endif /*ASSERT*/
683 }
684 }
685 #endif
686 return can_eliminate;
687 }
689 // Do scalar replacement.
690 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
691 GrowableArray <SafePointNode *> safepoints_done;
693 ciKlass* klass = NULL;
694 ciInstanceKlass* iklass = NULL;
695 int nfields = 0;
696 int array_base;
697 int element_size;
698 BasicType basic_elem_type;
699 ciType* elem_type;
701 Node* res = alloc->result_cast();
702 const TypeOopPtr* res_type = NULL;
703 if (res != NULL) { // Could be NULL when there are no users
704 res_type = _igvn.type(res)->isa_oopptr();
705 }
707 if (res != NULL) {
708 klass = res_type->klass();
709 if (res_type->isa_instptr()) {
710 // find the fields of the class which will be needed for safepoint debug information
711 assert(klass->is_instance_klass(), "must be an instance klass.");
712 iklass = klass->as_instance_klass();
713 nfields = iklass->nof_nonstatic_fields();
714 } else {
715 // find the array's elements which will be needed for safepoint debug information
716 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
717 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass.");
718 elem_type = klass->as_array_klass()->element_type();
719 basic_elem_type = elem_type->basic_type();
720 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
721 element_size = type2aelembytes(basic_elem_type);
722 }
723 }
724 //
725 // Process the safepoint uses
726 //
727 while (safepoints.length() > 0) {
728 SafePointNode* sfpt = safepoints.pop();
729 Node* mem = sfpt->memory();
730 assert(sfpt->jvms() != NULL, "missed JVMS");
731 // Fields of scalar objs are referenced only at the end
732 // of regular debuginfo at the last (youngest) JVMS.
733 // Record relative start index.
734 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
735 SafePointScalarObjectNode* sobj = new (C) SafePointScalarObjectNode(res_type,
736 #ifdef ASSERT
737 alloc,
738 #endif
739 first_ind, nfields);
740 sobj->init_req(0, C->root());
741 transform_later(sobj);
743 // Scan object's fields adding an input to the safepoint for each field.
744 for (int j = 0; j < nfields; j++) {
745 intptr_t offset;
746 ciField* field = NULL;
747 if (iklass != NULL) {
748 field = iklass->nonstatic_field_at(j);
749 offset = field->offset();
750 elem_type = field->type();
751 basic_elem_type = field->layout_type();
752 } else {
753 offset = array_base + j * (intptr_t)element_size;
754 }
756 const Type *field_type;
757 // The next code is taken from Parse::do_get_xxx().
758 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) {
759 if (!elem_type->is_loaded()) {
760 field_type = TypeInstPtr::BOTTOM;
761 } else if (field != NULL && field->is_constant() && field->is_static()) {
762 // This can happen if the constant oop is non-perm.
763 ciObject* con = field->constant_value().as_object();
764 // Do not "join" in the previous type; it doesn't add value,
765 // and may yield a vacuous result if the field is of interface type.
766 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
767 assert(field_type != NULL, "field singleton type must be consistent");
768 } else {
769 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
770 }
771 if (UseCompressedOops) {
772 field_type = field_type->make_narrowoop();
773 basic_elem_type = T_NARROWOOP;
774 }
775 } else {
776 field_type = Type::get_const_basic_type(basic_elem_type);
777 }
779 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr();
781 Node *field_val = value_from_mem(mem, basic_elem_type, field_type, field_addr_type, alloc);
782 if (field_val == NULL) {
783 // We weren't able to find a value for this field,
784 // give up on eliminating this allocation.
786 // Remove any extra entries we added to the safepoint.
787 uint last = sfpt->req() - 1;
788 for (int k = 0; k < j; k++) {
789 sfpt->del_req(last--);
790 }
791 // rollback processed safepoints
792 while (safepoints_done.length() > 0) {
793 SafePointNode* sfpt_done = safepoints_done.pop();
794 // remove any extra entries we added to the safepoint
795 last = sfpt_done->req() - 1;
796 for (int k = 0; k < nfields; k++) {
797 sfpt_done->del_req(last--);
798 }
799 JVMState *jvms = sfpt_done->jvms();
800 jvms->set_endoff(sfpt_done->req());
801 // Now make a pass over the debug information replacing any references
802 // to SafePointScalarObjectNode with the allocated object.
803 int start = jvms->debug_start();
804 int end = jvms->debug_end();
805 for (int i = start; i < end; i++) {
806 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
807 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
808 if (scobj->first_index(jvms) == sfpt_done->req() &&
809 scobj->n_fields() == (uint)nfields) {
810 assert(scobj->alloc() == alloc, "sanity");
811 sfpt_done->set_req(i, res);
812 }
813 }
814 }
815 }
816 #ifndef PRODUCT
817 if (PrintEliminateAllocations) {
818 if (field != NULL) {
819 tty->print("=== At SafePoint node %d can't find value of Field: ",
820 sfpt->_idx);
821 field->print();
822 int field_idx = C->get_alias_index(field_addr_type);
823 tty->print(" (alias_idx=%d)", field_idx);
824 } else { // Array's element
825 tty->print("=== At SafePoint node %d can't find value of array element [%d]",
826 sfpt->_idx, j);
827 }
828 tty->print(", which prevents elimination of: ");
829 if (res == NULL)
830 alloc->dump();
831 else
832 res->dump();
833 }
834 #endif
835 return false;
836 }
837 if (UseCompressedOops && field_type->isa_narrowoop()) {
838 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
839 // to be able scalar replace the allocation.
840 if (field_val->is_EncodeP()) {
841 field_val = field_val->in(1);
842 } else {
843 field_val = transform_later(new (C) DecodeNNode(field_val, field_val->get_ptr_type()));
844 }
845 }
846 sfpt->add_req(field_val);
847 }
848 JVMState *jvms = sfpt->jvms();
849 jvms->set_endoff(sfpt->req());
850 // Now make a pass over the debug information replacing any references
851 // to the allocated object with "sobj"
852 int start = jvms->debug_start();
853 int end = jvms->debug_end();
854 sfpt->replace_edges_in_range(res, sobj, start, end);
855 safepoints_done.append_if_missing(sfpt); // keep it for rollback
856 }
857 return true;
858 }
860 // Process users of eliminated allocation.
861 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) {
862 Node* res = alloc->result_cast();
863 if (res != NULL) {
864 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
865 Node *use = res->last_out(j);
866 uint oc1 = res->outcnt();
868 if (use->is_AddP()) {
869 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
870 Node *n = use->last_out(k);
871 uint oc2 = use->outcnt();
872 if (n->is_Store()) {
873 #ifdef ASSERT
874 // Verify that there is no dependent MemBarVolatile nodes,
875 // they should be removed during IGVN, see MemBarNode::Ideal().
876 for (DUIterator_Fast pmax, p = n->fast_outs(pmax);
877 p < pmax; p++) {
878 Node* mb = n->fast_out(p);
879 assert(mb->is_Initialize() || !mb->is_MemBar() ||
880 mb->req() <= MemBarNode::Precedent ||
881 mb->in(MemBarNode::Precedent) != n,
882 "MemBarVolatile should be eliminated for non-escaping object");
883 }
884 #endif
885 _igvn.replace_node(n, n->in(MemNode::Memory));
886 } else {
887 eliminate_card_mark(n);
888 }
889 k -= (oc2 - use->outcnt());
890 }
891 } else {
892 eliminate_card_mark(use);
893 }
894 j -= (oc1 - res->outcnt());
895 }
896 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
897 _igvn.remove_dead_node(res);
898 }
900 //
901 // Process other users of allocation's projections
902 //
903 if (_resproj != NULL && _resproj->outcnt() != 0) {
904 // First disconnect stores captured by Initialize node.
905 // If Initialize node is eliminated first in the following code,
906 // it will kill such stores and DUIterator_Last will assert.
907 for (DUIterator_Fast jmax, j = _resproj->fast_outs(jmax); j < jmax; j++) {
908 Node *use = _resproj->fast_out(j);
909 if (use->is_AddP()) {
910 // raw memory addresses used only by the initialization
911 _igvn.replace_node(use, C->top());
912 --j; --jmax;
913 }
914 }
915 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) {
916 Node *use = _resproj->last_out(j);
917 uint oc1 = _resproj->outcnt();
918 if (use->is_Initialize()) {
919 // Eliminate Initialize node.
920 InitializeNode *init = use->as_Initialize();
921 assert(init->outcnt() <= 2, "only a control and memory projection expected");
922 Node *ctrl_proj = init->proj_out(TypeFunc::Control);
923 if (ctrl_proj != NULL) {
924 assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection");
925 _igvn.replace_node(ctrl_proj, _fallthroughcatchproj);
926 }
927 Node *mem_proj = init->proj_out(TypeFunc::Memory);
928 if (mem_proj != NULL) {
929 Node *mem = init->in(TypeFunc::Memory);
930 #ifdef ASSERT
931 if (mem->is_MergeMem()) {
932 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection");
933 } else {
934 assert(mem == _memproj_fallthrough, "allocation memory projection");
935 }
936 #endif
937 _igvn.replace_node(mem_proj, mem);
938 }
939 } else {
940 assert(false, "only Initialize or AddP expected");
941 }
942 j -= (oc1 - _resproj->outcnt());
943 }
944 }
945 if (_fallthroughcatchproj != NULL) {
946 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control));
947 }
948 if (_memproj_fallthrough != NULL) {
949 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory));
950 }
951 if (_memproj_catchall != NULL) {
952 _igvn.replace_node(_memproj_catchall, C->top());
953 }
954 if (_ioproj_fallthrough != NULL) {
955 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O));
956 }
957 if (_ioproj_catchall != NULL) {
958 _igvn.replace_node(_ioproj_catchall, C->top());
959 }
960 if (_catchallcatchproj != NULL) {
961 _igvn.replace_node(_catchallcatchproj, C->top());
962 }
963 }
965 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
966 if (!EliminateAllocations || !alloc->_is_non_escaping) {
967 return false;
968 }
969 Node* klass = alloc->in(AllocateNode::KlassNode);
970 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
971 Node* res = alloc->result_cast();
972 // Eliminate boxing allocations which are not used
973 // regardless scalar replacable status.
974 bool boxing_alloc = C->eliminate_boxing() &&
975 tklass->klass()->is_instance_klass() &&
976 tklass->klass()->as_instance_klass()->is_box_klass();
977 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != NULL))) {
978 return false;
979 }
981 extract_call_projections(alloc);
983 GrowableArray <SafePointNode *> safepoints;
984 if (!can_eliminate_allocation(alloc, safepoints)) {
985 return false;
986 }
988 if (!alloc->_is_scalar_replaceable) {
989 assert(res == NULL, "sanity");
990 // We can only eliminate allocation if all debug info references
991 // are already replaced with SafePointScalarObject because
992 // we can't search for a fields value without instance_id.
993 if (safepoints.length() > 0) {
994 return false;
995 }
996 }
998 if (!scalar_replacement(alloc, safepoints)) {
999 return false;
1000 }
1002 CompileLog* log = C->log();
1003 if (log != NULL) {
1004 log->head("eliminate_allocation type='%d'",
1005 log->identify(tklass->klass()));
1006 JVMState* p = alloc->jvms();
1007 while (p != NULL) {
1008 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1009 p = p->caller();
1010 }
1011 log->tail("eliminate_allocation");
1012 }
1014 process_users_of_allocation(alloc);
1016 #ifndef PRODUCT
1017 if (PrintEliminateAllocations) {
1018 if (alloc->is_AllocateArray())
1019 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1020 else
1021 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1022 }
1023 #endif
1025 return true;
1026 }
1028 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1029 // EA should remove all uses of non-escaping boxing node.
1030 if (!C->eliminate_boxing() || boxing->proj_out(TypeFunc::Parms) != NULL) {
1031 return false;
1032 }
1034 extract_call_projections(boxing);
1036 const TypeTuple* r = boxing->tf()->range();
1037 assert(r->cnt() > TypeFunc::Parms, "sanity");
1038 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1039 assert(t != NULL, "sanity");
1041 CompileLog* log = C->log();
1042 if (log != NULL) {
1043 log->head("eliminate_boxing type='%d'",
1044 log->identify(t->klass()));
1045 JVMState* p = boxing->jvms();
1046 while (p != NULL) {
1047 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1048 p = p->caller();
1049 }
1050 log->tail("eliminate_boxing");
1051 }
1053 process_users_of_allocation(boxing);
1055 #ifndef PRODUCT
1056 if (PrintEliminateAllocations) {
1057 tty->print("++++ Eliminated: %d ", boxing->_idx);
1058 boxing->method()->print_short_name(tty);
1059 tty->cr();
1060 }
1061 #endif
1063 return true;
1064 }
1066 //---------------------------set_eden_pointers-------------------------
1067 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) {
1068 if (UseTLAB) { // Private allocation: load from TLS
1069 Node* thread = transform_later(new (C) ThreadLocalNode());
1070 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset());
1071 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset());
1072 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset);
1073 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset);
1074 } else { // Shared allocation: load from globals
1075 CollectedHeap* ch = Universe::heap();
1076 address top_adr = (address)ch->top_addr();
1077 address end_adr = (address)ch->end_addr();
1078 eden_top_adr = makecon(TypeRawPtr::make(top_adr));
1079 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr);
1080 }
1081 }
1084 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
1085 Node* adr = basic_plus_adr(base, offset);
1086 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
1087 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt);
1088 transform_later(value);
1089 return value;
1090 }
1093 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
1094 Node* adr = basic_plus_adr(base, offset);
1095 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt);
1096 transform_later(mem);
1097 return mem;
1098 }
1100 //=============================================================================
1101 //
1102 // A L L O C A T I O N
1103 //
1104 // Allocation attempts to be fast in the case of frequent small objects.
1105 // It breaks down like this:
1106 //
1107 // 1) Size in doublewords is computed. This is a constant for objects and
1108 // variable for most arrays. Doubleword units are used to avoid size
1109 // overflow of huge doubleword arrays. We need doublewords in the end for
1110 // rounding.
1111 //
1112 // 2) Size is checked for being 'too large'. Too-large allocations will go
1113 // the slow path into the VM. The slow path can throw any required
1114 // exceptions, and does all the special checks for very large arrays. The
1115 // size test can constant-fold away for objects. For objects with
1116 // finalizers it constant-folds the otherway: you always go slow with
1117 // finalizers.
1118 //
1119 // 3) If NOT using TLABs, this is the contended loop-back point.
1120 // Load-Locked the heap top. If using TLABs normal-load the heap top.
1121 //
1122 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
1123 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
1124 // "size*8" we always enter the VM, where "largish" is a constant picked small
1125 // enough that there's always space between the eden max and 4Gig (old space is
1126 // there so it's quite large) and large enough that the cost of entering the VM
1127 // is dwarfed by the cost to initialize the space.
1128 //
1129 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1130 // down. If contended, repeat at step 3. If using TLABs normal-store
1131 // adjusted heap top back down; there is no contention.
1132 //
1133 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
1134 // fields.
1135 //
1136 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1137 // oop flavor.
1138 //
1139 //=============================================================================
1140 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1141 // Allocations bigger than this always go the slow route.
1142 // This value must be small enough that allocation attempts that need to
1143 // trigger exceptions go the slow route. Also, it must be small enough so
1144 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1145 //=============================================================================j//
1146 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1147 // The allocator will coalesce int->oop copies away. See comment in
1148 // coalesce.cpp about how this works. It depends critically on the exact
1149 // code shape produced here, so if you are changing this code shape
1150 // make sure the GC info for the heap-top is correct in and around the
1151 // slow-path call.
1152 //
1154 void PhaseMacroExpand::expand_allocate_common(
1155 AllocateNode* alloc, // allocation node to be expanded
1156 Node* length, // array length for an array allocation
1157 const TypeFunc* slow_call_type, // Type of slow call
1158 address slow_call_address // Address of slow call
1159 )
1160 {
1162 Node* ctrl = alloc->in(TypeFunc::Control);
1163 Node* mem = alloc->in(TypeFunc::Memory);
1164 Node* i_o = alloc->in(TypeFunc::I_O);
1165 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1166 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1167 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1169 assert(ctrl != NULL, "must have control");
1170 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1171 // they will not be used if "always_slow" is set
1172 enum { slow_result_path = 1, fast_result_path = 2 };
1173 Node *result_region;
1174 Node *result_phi_rawmem;
1175 Node *result_phi_rawoop;
1176 Node *result_phi_i_o;
1178 // The initial slow comparison is a size check, the comparison
1179 // we want to do is a BoolTest::gt
1180 bool always_slow = false;
1181 int tv = _igvn.find_int_con(initial_slow_test, -1);
1182 if (tv >= 0) {
1183 always_slow = (tv == 1);
1184 initial_slow_test = NULL;
1185 } else {
1186 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1187 }
1189 if (C->env()->dtrace_alloc_probes() ||
1190 !UseTLAB && (!Universe::heap()->supports_inline_contig_alloc() ||
1191 (UseConcMarkSweepGC && CMSIncrementalMode))) {
1192 // Force slow-path allocation
1193 always_slow = true;
1194 initial_slow_test = NULL;
1195 }
1198 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1199 Node *slow_region = NULL;
1200 Node *toobig_false = ctrl;
1202 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent");
1203 // generate the initial test if necessary
1204 if (initial_slow_test != NULL ) {
1205 slow_region = new (C) RegionNode(3);
1207 // Now make the initial failure test. Usually a too-big test but
1208 // might be a TRUE for finalizers or a fancy class check for
1209 // newInstance0.
1210 IfNode *toobig_iff = new (C) IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1211 transform_later(toobig_iff);
1212 // Plug the failing-too-big test into the slow-path region
1213 Node *toobig_true = new (C) IfTrueNode( toobig_iff );
1214 transform_later(toobig_true);
1215 slow_region ->init_req( too_big_or_final_path, toobig_true );
1216 toobig_false = new (C) IfFalseNode( toobig_iff );
1217 transform_later(toobig_false);
1218 } else { // No initial test, just fall into next case
1219 toobig_false = ctrl;
1220 debug_only(slow_region = NodeSentinel);
1221 }
1223 Node *slow_mem = mem; // save the current memory state for slow path
1224 // generate the fast allocation code unless we know that the initial test will always go slow
1225 if (!always_slow) {
1226 // Fast path modifies only raw memory.
1227 if (mem->is_MergeMem()) {
1228 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1229 }
1231 Node* eden_top_adr;
1232 Node* eden_end_adr;
1234 set_eden_pointers(eden_top_adr, eden_end_adr);
1236 // Load Eden::end. Loop invariant and hoisted.
1237 //
1238 // Note: We set the control input on "eden_end" and "old_eden_top" when using
1239 // a TLAB to work around a bug where these values were being moved across
1240 // a safepoint. These are not oops, so they cannot be include in the oop
1241 // map, but they can be changed by a GC. The proper way to fix this would
1242 // be to set the raw memory state when generating a SafepointNode. However
1243 // this will require extensive changes to the loop optimization in order to
1244 // prevent a degradation of the optimization.
1245 // See comment in memnode.hpp, around line 227 in class LoadPNode.
1246 Node *eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS);
1248 // allocate the Region and Phi nodes for the result
1249 result_region = new (C) RegionNode(3);
1250 result_phi_rawmem = new (C) PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1251 result_phi_rawoop = new (C) PhiNode(result_region, TypeRawPtr::BOTTOM);
1252 result_phi_i_o = new (C) PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1254 // We need a Region for the loop-back contended case.
1255 enum { fall_in_path = 1, contended_loopback_path = 2 };
1256 Node *contended_region;
1257 Node *contended_phi_rawmem;
1258 if (UseTLAB) {
1259 contended_region = toobig_false;
1260 contended_phi_rawmem = mem;
1261 } else {
1262 contended_region = new (C) RegionNode(3);
1263 contended_phi_rawmem = new (C) PhiNode(contended_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1264 // Now handle the passing-too-big test. We fall into the contended
1265 // loop-back merge point.
1266 contended_region ->init_req(fall_in_path, toobig_false);
1267 contended_phi_rawmem->init_req(fall_in_path, mem);
1268 transform_later(contended_region);
1269 transform_later(contended_phi_rawmem);
1270 }
1272 // Load(-locked) the heap top.
1273 // See note above concerning the control input when using a TLAB
1274 Node *old_eden_top = UseTLAB
1275 ? new (C) LoadPNode (ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM)
1276 : new (C) LoadPLockedNode(contended_region, contended_phi_rawmem, eden_top_adr);
1278 transform_later(old_eden_top);
1279 // Add to heap top to get a new heap top
1280 Node *new_eden_top = new (C) AddPNode(top(), old_eden_top, size_in_bytes);
1281 transform_later(new_eden_top);
1282 // Check for needing a GC; compare against heap end
1283 Node *needgc_cmp = new (C) CmpPNode(new_eden_top, eden_end);
1284 transform_later(needgc_cmp);
1285 Node *needgc_bol = new (C) BoolNode(needgc_cmp, BoolTest::ge);
1286 transform_later(needgc_bol);
1287 IfNode *needgc_iff = new (C) IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN);
1288 transform_later(needgc_iff);
1290 // Plug the failing-heap-space-need-gc test into the slow-path region
1291 Node *needgc_true = new (C) IfTrueNode(needgc_iff);
1292 transform_later(needgc_true);
1293 if (initial_slow_test) {
1294 slow_region->init_req(need_gc_path, needgc_true);
1295 // This completes all paths into the slow merge point
1296 transform_later(slow_region);
1297 } else { // No initial slow path needed!
1298 // Just fall from the need-GC path straight into the VM call.
1299 slow_region = needgc_true;
1300 }
1301 // No need for a GC. Setup for the Store-Conditional
1302 Node *needgc_false = new (C) IfFalseNode(needgc_iff);
1303 transform_later(needgc_false);
1305 // Grab regular I/O before optional prefetch may change it.
1306 // Slow-path does no I/O so just set it to the original I/O.
1307 result_phi_i_o->init_req(slow_result_path, i_o);
1309 i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem,
1310 old_eden_top, new_eden_top, length);
1312 // Name successful fast-path variables
1313 Node* fast_oop = old_eden_top;
1314 Node* fast_oop_ctrl;
1315 Node* fast_oop_rawmem;
1317 // Store (-conditional) the modified eden top back down.
1318 // StorePConditional produces flags for a test PLUS a modified raw
1319 // memory state.
1320 if (UseTLAB) {
1321 Node* store_eden_top =
1322 new (C) StorePNode(needgc_false, contended_phi_rawmem, eden_top_adr,
1323 TypeRawPtr::BOTTOM, new_eden_top);
1324 transform_later(store_eden_top);
1325 fast_oop_ctrl = needgc_false; // No contention, so this is the fast path
1326 fast_oop_rawmem = store_eden_top;
1327 } else {
1328 Node* store_eden_top =
1329 new (C) StorePConditionalNode(needgc_false, contended_phi_rawmem, eden_top_adr,
1330 new_eden_top, fast_oop/*old_eden_top*/);
1331 transform_later(store_eden_top);
1332 Node *contention_check = new (C) BoolNode(store_eden_top, BoolTest::ne);
1333 transform_later(contention_check);
1334 store_eden_top = new (C) SCMemProjNode(store_eden_top);
1335 transform_later(store_eden_top);
1337 // If not using TLABs, check to see if there was contention.
1338 IfNode *contention_iff = new (C) IfNode (needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN);
1339 transform_later(contention_iff);
1340 Node *contention_true = new (C) IfTrueNode(contention_iff);
1341 transform_later(contention_true);
1342 // If contention, loopback and try again.
1343 contended_region->init_req(contended_loopback_path, contention_true);
1344 contended_phi_rawmem->init_req(contended_loopback_path, store_eden_top);
1346 // Fast-path succeeded with no contention!
1347 Node *contention_false = new (C) IfFalseNode(contention_iff);
1348 transform_later(contention_false);
1349 fast_oop_ctrl = contention_false;
1351 // Bump total allocated bytes for this thread
1352 Node* thread = new (C) ThreadLocalNode();
1353 transform_later(thread);
1354 Node* alloc_bytes_adr = basic_plus_adr(top()/*not oop*/, thread,
1355 in_bytes(JavaThread::allocated_bytes_offset()));
1356 Node* alloc_bytes = make_load(fast_oop_ctrl, store_eden_top, alloc_bytes_adr,
1357 0, TypeLong::LONG, T_LONG);
1358 #ifdef _LP64
1359 Node* alloc_size = size_in_bytes;
1360 #else
1361 Node* alloc_size = new (C) ConvI2LNode(size_in_bytes);
1362 transform_later(alloc_size);
1363 #endif
1364 Node* new_alloc_bytes = new (C) AddLNode(alloc_bytes, alloc_size);
1365 transform_later(new_alloc_bytes);
1366 fast_oop_rawmem = make_store(fast_oop_ctrl, store_eden_top, alloc_bytes_adr,
1367 0, new_alloc_bytes, T_LONG);
1368 }
1370 InitializeNode* init = alloc->initialization();
1371 fast_oop_rawmem = initialize_object(alloc,
1372 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1373 klass_node, length, size_in_bytes);
1375 // If initialization is performed by an array copy, any required
1376 // MemBarStoreStore was already added. If the object does not
1377 // escape no need for a MemBarStoreStore. Otherwise we need a
1378 // MemBarStoreStore so that stores that initialize this object
1379 // can't be reordered with a subsequent store that makes this
1380 // object accessible by other threads.
1381 if (init == NULL || (!init->is_complete_with_arraycopy() && !init->does_not_escape())) {
1382 if (init == NULL || init->req() < InitializeNode::RawStores) {
1383 // No InitializeNode or no stores captured by zeroing
1384 // elimination. Simply add the MemBarStoreStore after object
1385 // initialization.
1386 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1387 transform_later(mb);
1389 mb->init_req(TypeFunc::Memory, fast_oop_rawmem);
1390 mb->init_req(TypeFunc::Control, fast_oop_ctrl);
1391 fast_oop_ctrl = new (C) ProjNode(mb,TypeFunc::Control);
1392 transform_later(fast_oop_ctrl);
1393 fast_oop_rawmem = new (C) ProjNode(mb,TypeFunc::Memory);
1394 transform_later(fast_oop_rawmem);
1395 } else {
1396 // Add the MemBarStoreStore after the InitializeNode so that
1397 // all stores performing the initialization that were moved
1398 // before the InitializeNode happen before the storestore
1399 // barrier.
1401 Node* init_ctrl = init->proj_out(TypeFunc::Control);
1402 Node* init_mem = init->proj_out(TypeFunc::Memory);
1404 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1405 transform_later(mb);
1407 Node* ctrl = new (C) ProjNode(init,TypeFunc::Control);
1408 transform_later(ctrl);
1409 Node* mem = new (C) ProjNode(init,TypeFunc::Memory);
1410 transform_later(mem);
1412 // The MemBarStoreStore depends on control and memory coming
1413 // from the InitializeNode
1414 mb->init_req(TypeFunc::Memory, mem);
1415 mb->init_req(TypeFunc::Control, ctrl);
1417 ctrl = new (C) ProjNode(mb,TypeFunc::Control);
1418 transform_later(ctrl);
1419 mem = new (C) ProjNode(mb,TypeFunc::Memory);
1420 transform_later(mem);
1422 // All nodes that depended on the InitializeNode for control
1423 // and memory must now depend on the MemBarNode that itself
1424 // depends on the InitializeNode
1425 _igvn.replace_node(init_ctrl, ctrl);
1426 _igvn.replace_node(init_mem, mem);
1427 }
1428 }
1430 if (C->env()->dtrace_extended_probes()) {
1431 // Slow-path call
1432 int size = TypeFunc::Parms + 2;
1433 CallLeafNode *call = new (C) CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1434 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base),
1435 "dtrace_object_alloc",
1436 TypeRawPtr::BOTTOM);
1438 // Get base of thread-local storage area
1439 Node* thread = new (C) ThreadLocalNode();
1440 transform_later(thread);
1442 call->init_req(TypeFunc::Parms+0, thread);
1443 call->init_req(TypeFunc::Parms+1, fast_oop);
1444 call->init_req(TypeFunc::Control, fast_oop_ctrl);
1445 call->init_req(TypeFunc::I_O , top()); // does no i/o
1446 call->init_req(TypeFunc::Memory , fast_oop_rawmem);
1447 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1448 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1449 transform_later(call);
1450 fast_oop_ctrl = new (C) ProjNode(call,TypeFunc::Control);
1451 transform_later(fast_oop_ctrl);
1452 fast_oop_rawmem = new (C) ProjNode(call,TypeFunc::Memory);
1453 transform_later(fast_oop_rawmem);
1454 }
1456 // Plug in the successful fast-path into the result merge point
1457 result_region ->init_req(fast_result_path, fast_oop_ctrl);
1458 result_phi_rawoop->init_req(fast_result_path, fast_oop);
1459 result_phi_i_o ->init_req(fast_result_path, i_o);
1460 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1461 } else {
1462 slow_region = ctrl;
1463 result_phi_i_o = i_o; // Rename it to use in the following code.
1464 }
1466 // Generate slow-path call
1467 CallNode *call = new (C) CallStaticJavaNode(slow_call_type, slow_call_address,
1468 OptoRuntime::stub_name(slow_call_address),
1469 alloc->jvms()->bci(),
1470 TypePtr::BOTTOM);
1471 call->init_req( TypeFunc::Control, slow_region );
1472 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o
1473 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs
1474 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) );
1475 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) );
1477 call->init_req(TypeFunc::Parms+0, klass_node);
1478 if (length != NULL) {
1479 call->init_req(TypeFunc::Parms+1, length);
1480 }
1482 // Copy debug information and adjust JVMState information, then replace
1483 // allocate node with the call
1484 copy_call_debug_info((CallNode *) alloc, call);
1485 if (!always_slow) {
1486 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1487 } else {
1488 // Hook i_o projection to avoid its elimination during allocation
1489 // replacement (when only a slow call is generated).
1490 call->set_req(TypeFunc::I_O, result_phi_i_o);
1491 }
1492 _igvn.replace_node(alloc, call);
1493 transform_later(call);
1495 // Identify the output projections from the allocate node and
1496 // adjust any references to them.
1497 // The control and io projections look like:
1498 //
1499 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1500 // Allocate Catch
1501 // ^---Proj(io) <-------+ ^---CatchProj(io)
1502 //
1503 // We are interested in the CatchProj nodes.
1504 //
1505 extract_call_projections(call);
1507 // An allocate node has separate memory projections for the uses on
1508 // the control and i_o paths. Replace the control memory projection with
1509 // result_phi_rawmem (unless we are only generating a slow call when
1510 // both memory projections are combined)
1511 if (!always_slow && _memproj_fallthrough != NULL) {
1512 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) {
1513 Node *use = _memproj_fallthrough->fast_out(i);
1514 _igvn.rehash_node_delayed(use);
1515 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem);
1516 // back up iterator
1517 --i;
1518 }
1519 }
1520 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete
1521 // _memproj_catchall so we end up with a call that has only 1 memory projection.
1522 if (_memproj_catchall != NULL ) {
1523 if (_memproj_fallthrough == NULL) {
1524 _memproj_fallthrough = new (C) ProjNode(call, TypeFunc::Memory);
1525 transform_later(_memproj_fallthrough);
1526 }
1527 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) {
1528 Node *use = _memproj_catchall->fast_out(i);
1529 _igvn.rehash_node_delayed(use);
1530 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough);
1531 // back up iterator
1532 --i;
1533 }
1534 assert(_memproj_catchall->outcnt() == 0, "all uses must be deleted");
1535 _igvn.remove_dead_node(_memproj_catchall);
1536 }
1538 // An allocate node has separate i_o projections for the uses on the control
1539 // and i_o paths. Always replace the control i_o projection with result i_o
1540 // otherwise incoming i_o become dead when only a slow call is generated
1541 // (it is different from memory projections where both projections are
1542 // combined in such case).
1543 if (_ioproj_fallthrough != NULL) {
1544 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) {
1545 Node *use = _ioproj_fallthrough->fast_out(i);
1546 _igvn.rehash_node_delayed(use);
1547 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o);
1548 // back up iterator
1549 --i;
1550 }
1551 }
1552 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete
1553 // _ioproj_catchall so we end up with a call that has only 1 i_o projection.
1554 if (_ioproj_catchall != NULL ) {
1555 if (_ioproj_fallthrough == NULL) {
1556 _ioproj_fallthrough = new (C) ProjNode(call, TypeFunc::I_O);
1557 transform_later(_ioproj_fallthrough);
1558 }
1559 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) {
1560 Node *use = _ioproj_catchall->fast_out(i);
1561 _igvn.rehash_node_delayed(use);
1562 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough);
1563 // back up iterator
1564 --i;
1565 }
1566 assert(_ioproj_catchall->outcnt() == 0, "all uses must be deleted");
1567 _igvn.remove_dead_node(_ioproj_catchall);
1568 }
1570 // if we generated only a slow call, we are done
1571 if (always_slow) {
1572 // Now we can unhook i_o.
1573 if (result_phi_i_o->outcnt() > 1) {
1574 call->set_req(TypeFunc::I_O, top());
1575 } else {
1576 assert(result_phi_i_o->unique_ctrl_out() == call, "");
1577 // Case of new array with negative size known during compilation.
1578 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1579 // following code since call to runtime will throw exception.
1580 // As result there will be no users of i_o after the call.
1581 // Leave i_o attached to this call to avoid problems in preceding graph.
1582 }
1583 return;
1584 }
1587 if (_fallthroughcatchproj != NULL) {
1588 ctrl = _fallthroughcatchproj->clone();
1589 transform_later(ctrl);
1590 _igvn.replace_node(_fallthroughcatchproj, result_region);
1591 } else {
1592 ctrl = top();
1593 }
1594 Node *slow_result;
1595 if (_resproj == NULL) {
1596 // no uses of the allocation result
1597 slow_result = top();
1598 } else {
1599 slow_result = _resproj->clone();
1600 transform_later(slow_result);
1601 _igvn.replace_node(_resproj, result_phi_rawoop);
1602 }
1604 // Plug slow-path into result merge point
1605 result_region ->init_req( slow_result_path, ctrl );
1606 result_phi_rawoop->init_req( slow_result_path, slow_result);
1607 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough );
1608 transform_later(result_region);
1609 transform_later(result_phi_rawoop);
1610 transform_later(result_phi_rawmem);
1611 transform_later(result_phi_i_o);
1612 // This completes all paths into the result merge point
1613 }
1616 // Helper for PhaseMacroExpand::expand_allocate_common.
1617 // Initializes the newly-allocated storage.
1618 Node*
1619 PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1620 Node* control, Node* rawmem, Node* object,
1621 Node* klass_node, Node* length,
1622 Node* size_in_bytes) {
1623 InitializeNode* init = alloc->initialization();
1624 // Store the klass & mark bits
1625 Node* mark_node = NULL;
1626 // For now only enable fast locking for non-array types
1627 if (UseBiasedLocking && (length == NULL)) {
1628 mark_node = make_load(control, rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
1629 } else {
1630 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype()));
1631 }
1632 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
1634 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1635 int header_size = alloc->minimum_header_size(); // conservatively small
1637 // Array length
1638 if (length != NULL) { // Arrays need length field
1639 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1640 // conservatively small header size:
1641 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1642 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1643 if (k->is_array_klass()) // we know the exact header size in most cases:
1644 header_size = Klass::layout_helper_header_size(k->layout_helper());
1645 }
1647 // Clear the object body, if necessary.
1648 if (init == NULL) {
1649 // The init has somehow disappeared; be cautious and clear everything.
1650 //
1651 // This can happen if a node is allocated but an uncommon trap occurs
1652 // immediately. In this case, the Initialize gets associated with the
1653 // trap, and may be placed in a different (outer) loop, if the Allocate
1654 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1655 // there can be two Allocates to one Initialize. The answer in all these
1656 // edge cases is safety first. It is always safe to clear immediately
1657 // within an Allocate, and then (maybe or maybe not) clear some more later.
1658 if (!ZeroTLAB)
1659 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1660 header_size, size_in_bytes,
1661 &_igvn);
1662 } else {
1663 if (!init->is_complete()) {
1664 // Try to win by zeroing only what the init does not store.
1665 // We can also try to do some peephole optimizations,
1666 // such as combining some adjacent subword stores.
1667 rawmem = init->complete_stores(control, rawmem, object,
1668 header_size, size_in_bytes, &_igvn);
1669 }
1670 // We have no more use for this link, since the AllocateNode goes away:
1671 init->set_req(InitializeNode::RawAddress, top());
1672 // (If we keep the link, it just confuses the register allocator,
1673 // who thinks he sees a real use of the address by the membar.)
1674 }
1676 return rawmem;
1677 }
1679 // Generate prefetch instructions for next allocations.
1680 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
1681 Node*& contended_phi_rawmem,
1682 Node* old_eden_top, Node* new_eden_top,
1683 Node* length) {
1684 enum { fall_in_path = 1, pf_path = 2 };
1685 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
1686 // Generate prefetch allocation with watermark check.
1687 // As an allocation hits the watermark, we will prefetch starting
1688 // at a "distance" away from watermark.
1690 Node *pf_region = new (C) RegionNode(3);
1691 Node *pf_phi_rawmem = new (C) PhiNode( pf_region, Type::MEMORY,
1692 TypeRawPtr::BOTTOM );
1693 // I/O is used for Prefetch
1694 Node *pf_phi_abio = new (C) PhiNode( pf_region, Type::ABIO );
1696 Node *thread = new (C) ThreadLocalNode();
1697 transform_later(thread);
1699 Node *eden_pf_adr = new (C) AddPNode( top()/*not oop*/, thread,
1700 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
1701 transform_later(eden_pf_adr);
1703 Node *old_pf_wm = new (C) LoadPNode( needgc_false,
1704 contended_phi_rawmem, eden_pf_adr,
1705 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM );
1706 transform_later(old_pf_wm);
1708 // check against new_eden_top
1709 Node *need_pf_cmp = new (C) CmpPNode( new_eden_top, old_pf_wm );
1710 transform_later(need_pf_cmp);
1711 Node *need_pf_bol = new (C) BoolNode( need_pf_cmp, BoolTest::ge );
1712 transform_later(need_pf_bol);
1713 IfNode *need_pf_iff = new (C) IfNode( needgc_false, need_pf_bol,
1714 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1715 transform_later(need_pf_iff);
1717 // true node, add prefetchdistance
1718 Node *need_pf_true = new (C) IfTrueNode( need_pf_iff );
1719 transform_later(need_pf_true);
1721 Node *need_pf_false = new (C) IfFalseNode( need_pf_iff );
1722 transform_later(need_pf_false);
1724 Node *new_pf_wmt = new (C) AddPNode( top(), old_pf_wm,
1725 _igvn.MakeConX(AllocatePrefetchDistance) );
1726 transform_later(new_pf_wmt );
1727 new_pf_wmt->set_req(0, need_pf_true);
1729 Node *store_new_wmt = new (C) StorePNode( need_pf_true,
1730 contended_phi_rawmem, eden_pf_adr,
1731 TypeRawPtr::BOTTOM, new_pf_wmt );
1732 transform_later(store_new_wmt);
1734 // adding prefetches
1735 pf_phi_abio->init_req( fall_in_path, i_o );
1737 Node *prefetch_adr;
1738 Node *prefetch;
1739 uint lines = AllocatePrefetchDistance / AllocatePrefetchStepSize;
1740 uint step_size = AllocatePrefetchStepSize;
1741 uint distance = 0;
1743 for ( uint i = 0; i < lines; i++ ) {
1744 prefetch_adr = new (C) AddPNode( old_pf_wm, new_pf_wmt,
1745 _igvn.MakeConX(distance) );
1746 transform_later(prefetch_adr);
1747 prefetch = new (C) PrefetchAllocationNode( i_o, prefetch_adr );
1748 transform_later(prefetch);
1749 distance += step_size;
1750 i_o = prefetch;
1751 }
1752 pf_phi_abio->set_req( pf_path, i_o );
1754 pf_region->init_req( fall_in_path, need_pf_false );
1755 pf_region->init_req( pf_path, need_pf_true );
1757 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
1758 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
1760 transform_later(pf_region);
1761 transform_later(pf_phi_rawmem);
1762 transform_later(pf_phi_abio);
1764 needgc_false = pf_region;
1765 contended_phi_rawmem = pf_phi_rawmem;
1766 i_o = pf_phi_abio;
1767 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) {
1768 // Insert a prefetch for each allocation.
1769 // This code is used for Sparc with BIS.
1770 Node *pf_region = new (C) RegionNode(3);
1771 Node *pf_phi_rawmem = new (C) PhiNode( pf_region, Type::MEMORY,
1772 TypeRawPtr::BOTTOM );
1774 // Generate several prefetch instructions.
1775 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1776 uint step_size = AllocatePrefetchStepSize;
1777 uint distance = AllocatePrefetchDistance;
1779 // Next cache address.
1780 Node *cache_adr = new (C) AddPNode(old_eden_top, old_eden_top,
1781 _igvn.MakeConX(distance));
1782 transform_later(cache_adr);
1783 cache_adr = new (C) CastP2XNode(needgc_false, cache_adr);
1784 transform_later(cache_adr);
1785 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1));
1786 cache_adr = new (C) AndXNode(cache_adr, mask);
1787 transform_later(cache_adr);
1788 cache_adr = new (C) CastX2PNode(cache_adr);
1789 transform_later(cache_adr);
1791 // Prefetch
1792 Node *prefetch = new (C) PrefetchAllocationNode( contended_phi_rawmem, cache_adr );
1793 prefetch->set_req(0, needgc_false);
1794 transform_later(prefetch);
1795 contended_phi_rawmem = prefetch;
1796 Node *prefetch_adr;
1797 distance = step_size;
1798 for ( uint i = 1; i < lines; i++ ) {
1799 prefetch_adr = new (C) AddPNode( cache_adr, cache_adr,
1800 _igvn.MakeConX(distance) );
1801 transform_later(prefetch_adr);
1802 prefetch = new (C) PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr );
1803 transform_later(prefetch);
1804 distance += step_size;
1805 contended_phi_rawmem = prefetch;
1806 }
1807 } else if( AllocatePrefetchStyle > 0 ) {
1808 // Insert a prefetch for each allocation only on the fast-path
1809 Node *prefetch_adr;
1810 Node *prefetch;
1811 // Generate several prefetch instructions.
1812 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1813 uint step_size = AllocatePrefetchStepSize;
1814 uint distance = AllocatePrefetchDistance;
1815 for ( uint i = 0; i < lines; i++ ) {
1816 prefetch_adr = new (C) AddPNode( old_eden_top, new_eden_top,
1817 _igvn.MakeConX(distance) );
1818 transform_later(prefetch_adr);
1819 prefetch = new (C) PrefetchAllocationNode( i_o, prefetch_adr );
1820 // Do not let it float too high, since if eden_top == eden_end,
1821 // both might be null.
1822 if( i == 0 ) { // Set control for first prefetch, next follows it
1823 prefetch->init_req(0, needgc_false);
1824 }
1825 transform_later(prefetch);
1826 distance += step_size;
1827 i_o = prefetch;
1828 }
1829 }
1830 return i_o;
1831 }
1834 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
1835 expand_allocate_common(alloc, NULL,
1836 OptoRuntime::new_instance_Type(),
1837 OptoRuntime::new_instance_Java());
1838 }
1840 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
1841 Node* length = alloc->in(AllocateNode::ALength);
1842 InitializeNode* init = alloc->initialization();
1843 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1844 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1845 address slow_call_address; // Address of slow call
1846 if (init != NULL && init->is_complete_with_arraycopy() &&
1847 k->is_type_array_klass()) {
1848 // Don't zero type array during slow allocation in VM since
1849 // it will be initialized later by arraycopy in compiled code.
1850 slow_call_address = OptoRuntime::new_array_nozero_Java();
1851 } else {
1852 slow_call_address = OptoRuntime::new_array_Java();
1853 }
1854 expand_allocate_common(alloc, length,
1855 OptoRuntime::new_array_Type(),
1856 slow_call_address);
1857 }
1859 //-------------------mark_eliminated_box----------------------------------
1860 //
1861 // During EA obj may point to several objects but after few ideal graph
1862 // transformations (CCP) it may point to only one non escaping object
1863 // (but still using phi), corresponding locks and unlocks will be marked
1864 // for elimination. Later obj could be replaced with a new node (new phi)
1865 // and which does not have escape information. And later after some graph
1866 // reshape other locks and unlocks (which were not marked for elimination
1867 // before) are connected to this new obj (phi) but they still will not be
1868 // marked for elimination since new obj has no escape information.
1869 // Mark all associated (same box and obj) lock and unlock nodes for
1870 // elimination if some of them marked already.
1871 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) {
1872 if (oldbox->as_BoxLock()->is_eliminated())
1873 return; // This BoxLock node was processed already.
1875 // New implementation (EliminateNestedLocks) has separate BoxLock
1876 // node for each locked region so mark all associated locks/unlocks as
1877 // eliminated even if different objects are referenced in one locked region
1878 // (for example, OSR compilation of nested loop inside locked scope).
1879 if (EliminateNestedLocks ||
1880 oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj)) {
1881 // Box is used only in one lock region. Mark this box as eliminated.
1882 _igvn.hash_delete(oldbox);
1883 oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value
1884 _igvn.hash_insert(oldbox);
1886 for (uint i = 0; i < oldbox->outcnt(); i++) {
1887 Node* u = oldbox->raw_out(i);
1888 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) {
1889 AbstractLockNode* alock = u->as_AbstractLock();
1890 // Check lock's box since box could be referenced by Lock's debug info.
1891 if (alock->box_node() == oldbox) {
1892 // Mark eliminated all related locks and unlocks.
1893 alock->set_non_esc_obj();
1894 }
1895 }
1896 }
1897 return;
1898 }
1900 // Create new "eliminated" BoxLock node and use it in monitor debug info
1901 // instead of oldbox for the same object.
1902 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
1904 // Note: BoxLock node is marked eliminated only here and it is used
1905 // to indicate that all associated lock and unlock nodes are marked
1906 // for elimination.
1907 newbox->set_eliminated();
1908 transform_later(newbox);
1910 // Replace old box node with new box for all users of the same object.
1911 for (uint i = 0; i < oldbox->outcnt();) {
1912 bool next_edge = true;
1914 Node* u = oldbox->raw_out(i);
1915 if (u->is_AbstractLock()) {
1916 AbstractLockNode* alock = u->as_AbstractLock();
1917 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) {
1918 // Replace Box and mark eliminated all related locks and unlocks.
1919 alock->set_non_esc_obj();
1920 _igvn.rehash_node_delayed(alock);
1921 alock->set_box_node(newbox);
1922 next_edge = false;
1923 }
1924 }
1925 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) {
1926 FastLockNode* flock = u->as_FastLock();
1927 assert(flock->box_node() == oldbox, "sanity");
1928 _igvn.rehash_node_delayed(flock);
1929 flock->set_box_node(newbox);
1930 next_edge = false;
1931 }
1933 // Replace old box in monitor debug info.
1934 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
1935 SafePointNode* sfn = u->as_SafePoint();
1936 JVMState* youngest_jvms = sfn->jvms();
1937 int max_depth = youngest_jvms->depth();
1938 for (int depth = 1; depth <= max_depth; depth++) {
1939 JVMState* jvms = youngest_jvms->of_depth(depth);
1940 int num_mon = jvms->nof_monitors();
1941 // Loop over monitors
1942 for (int idx = 0; idx < num_mon; idx++) {
1943 Node* obj_node = sfn->monitor_obj(jvms, idx);
1944 Node* box_node = sfn->monitor_box(jvms, idx);
1945 if (box_node == oldbox && obj_node->eqv_uncast(obj)) {
1946 int j = jvms->monitor_box_offset(idx);
1947 _igvn.replace_input_of(u, j, newbox);
1948 next_edge = false;
1949 }
1950 }
1951 }
1952 }
1953 if (next_edge) i++;
1954 }
1955 }
1957 //-----------------------mark_eliminated_locking_nodes-----------------------
1958 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) {
1959 if (EliminateNestedLocks) {
1960 if (alock->is_nested()) {
1961 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity");
1962 return;
1963 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened
1964 // Only Lock node has JVMState needed here.
1965 if (alock->jvms() != NULL && alock->as_Lock()->is_nested_lock_region()) {
1966 // Mark eliminated related nested locks and unlocks.
1967 Node* obj = alock->obj_node();
1968 BoxLockNode* box_node = alock->box_node()->as_BoxLock();
1969 assert(!box_node->is_eliminated(), "should not be marked yet");
1970 // Note: BoxLock node is marked eliminated only here
1971 // and it is used to indicate that all associated lock
1972 // and unlock nodes are marked for elimination.
1973 box_node->set_eliminated(); // Box's hash is always NO_HASH here
1974 for (uint i = 0; i < box_node->outcnt(); i++) {
1975 Node* u = box_node->raw_out(i);
1976 if (u->is_AbstractLock()) {
1977 alock = u->as_AbstractLock();
1978 if (alock->box_node() == box_node) {
1979 // Verify that this Box is referenced only by related locks.
1980 assert(alock->obj_node()->eqv_uncast(obj), "");
1981 // Mark all related locks and unlocks.
1982 alock->set_nested();
1983 }
1984 }
1985 }
1986 }
1987 return;
1988 }
1989 // Process locks for non escaping object
1990 assert(alock->is_non_esc_obj(), "");
1991 } // EliminateNestedLocks
1993 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
1994 // Look for all locks of this object and mark them and
1995 // corresponding BoxLock nodes as eliminated.
1996 Node* obj = alock->obj_node();
1997 for (uint j = 0; j < obj->outcnt(); j++) {
1998 Node* o = obj->raw_out(j);
1999 if (o->is_AbstractLock() &&
2000 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2001 alock = o->as_AbstractLock();
2002 Node* box = alock->box_node();
2003 // Replace old box node with new eliminated box for all users
2004 // of the same object and mark related locks as eliminated.
2005 mark_eliminated_box(box, obj);
2006 }
2007 }
2008 }
2009 }
2011 // we have determined that this lock/unlock can be eliminated, we simply
2012 // eliminate the node without expanding it.
2013 //
2014 // Note: The membar's associated with the lock/unlock are currently not
2015 // eliminated. This should be investigated as a future enhancement.
2016 //
2017 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2019 if (!alock->is_eliminated()) {
2020 return false;
2021 }
2022 #ifdef ASSERT
2023 if (!alock->is_coarsened()) {
2024 // Check that new "eliminated" BoxLock node is created.
2025 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2026 assert(oldbox->is_eliminated(), "should be done already");
2027 }
2028 #endif
2029 CompileLog* log = C->log();
2030 if (log != NULL) {
2031 log->head("eliminate_lock lock='%d'",
2032 alock->is_Lock());
2033 JVMState* p = alock->jvms();
2034 while (p != NULL) {
2035 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
2036 p = p->caller();
2037 }
2038 log->tail("eliminate_lock");
2039 }
2041 #ifndef PRODUCT
2042 if (PrintEliminateLocks) {
2043 if (alock->is_Lock()) {
2044 tty->print_cr("++++ Eliminated: %d Lock", alock->_idx);
2045 } else {
2046 tty->print_cr("++++ Eliminated: %d Unlock", alock->_idx);
2047 }
2048 }
2049 #endif
2051 Node* mem = alock->in(TypeFunc::Memory);
2052 Node* ctrl = alock->in(TypeFunc::Control);
2054 extract_call_projections(alock);
2055 // There are 2 projections from the lock. The lock node will
2056 // be deleted when its last use is subsumed below.
2057 assert(alock->outcnt() == 2 &&
2058 _fallthroughproj != NULL &&
2059 _memproj_fallthrough != NULL,
2060 "Unexpected projections from Lock/Unlock");
2062 Node* fallthroughproj = _fallthroughproj;
2063 Node* memproj_fallthrough = _memproj_fallthrough;
2065 // The memory projection from a lock/unlock is RawMem
2066 // The input to a Lock is merged memory, so extract its RawMem input
2067 // (unless the MergeMem has been optimized away.)
2068 if (alock->is_Lock()) {
2069 // Seach for MemBarAcquireLock node and delete it also.
2070 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2071 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, "");
2072 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2073 Node* memproj = membar->proj_out(TypeFunc::Memory);
2074 _igvn.replace_node(ctrlproj, fallthroughproj);
2075 _igvn.replace_node(memproj, memproj_fallthrough);
2077 // Delete FastLock node also if this Lock node is unique user
2078 // (a loop peeling may clone a Lock node).
2079 Node* flock = alock->as_Lock()->fastlock_node();
2080 if (flock->outcnt() == 1) {
2081 assert(flock->unique_out() == alock, "sanity");
2082 _igvn.replace_node(flock, top());
2083 }
2084 }
2086 // Seach for MemBarReleaseLock node and delete it also.
2087 if (alock->is_Unlock() && ctrl != NULL && ctrl->is_Proj() &&
2088 ctrl->in(0)->is_MemBar()) {
2089 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2090 assert(membar->Opcode() == Op_MemBarReleaseLock &&
2091 mem->is_Proj() && membar == mem->in(0), "");
2092 _igvn.replace_node(fallthroughproj, ctrl);
2093 _igvn.replace_node(memproj_fallthrough, mem);
2094 fallthroughproj = ctrl;
2095 memproj_fallthrough = mem;
2096 ctrl = membar->in(TypeFunc::Control);
2097 mem = membar->in(TypeFunc::Memory);
2098 }
2100 _igvn.replace_node(fallthroughproj, ctrl);
2101 _igvn.replace_node(memproj_fallthrough, mem);
2102 return true;
2103 }
2106 //------------------------------expand_lock_node----------------------
2107 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
2109 Node* ctrl = lock->in(TypeFunc::Control);
2110 Node* mem = lock->in(TypeFunc::Memory);
2111 Node* obj = lock->obj_node();
2112 Node* box = lock->box_node();
2113 Node* flock = lock->fastlock_node();
2115 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2117 // Make the merge point
2118 Node *region;
2119 Node *mem_phi;
2120 Node *slow_path;
2122 if (UseOptoBiasInlining) {
2123 /*
2124 * See the full description in MacroAssembler::biased_locking_enter().
2125 *
2126 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) {
2127 * // The object is biased.
2128 * proto_node = klass->prototype_header;
2129 * o_node = thread | proto_node;
2130 * x_node = o_node ^ mark_word;
2131 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ?
2132 * // Done.
2133 * } else {
2134 * if( (x_node & biased_lock_mask) != 0 ) {
2135 * // The klass's prototype header is no longer biased.
2136 * cas(&mark_word, mark_word, proto_node)
2137 * goto cas_lock;
2138 * } else {
2139 * // The klass's prototype header is still biased.
2140 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch?
2141 * old = mark_word;
2142 * new = o_node;
2143 * } else {
2144 * // Different thread or anonymous biased.
2145 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask);
2146 * new = thread | old;
2147 * }
2148 * // Try to rebias.
2149 * if( cas(&mark_word, old, new) == 0 ) {
2150 * // Done.
2151 * } else {
2152 * goto slow_path; // Failed.
2153 * }
2154 * }
2155 * }
2156 * } else {
2157 * // The object is not biased.
2158 * cas_lock:
2159 * if( FastLock(obj) == 0 ) {
2160 * // Done.
2161 * } else {
2162 * slow_path:
2163 * OptoRuntime::complete_monitor_locking_Java(obj);
2164 * }
2165 * }
2166 */
2168 region = new (C) RegionNode(5);
2169 // create a Phi for the memory state
2170 mem_phi = new (C) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2172 Node* fast_lock_region = new (C) RegionNode(3);
2173 Node* fast_lock_mem_phi = new (C) PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM);
2175 // First, check mark word for the biased lock pattern.
2176 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2178 // Get fast path - mark word has the biased lock pattern.
2179 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node,
2180 markOopDesc::biased_lock_mask_in_place,
2181 markOopDesc::biased_lock_pattern, true);
2182 // fast_lock_region->in(1) is set to slow path.
2183 fast_lock_mem_phi->init_req(1, mem);
2185 // Now check that the lock is biased to the current thread and has
2186 // the same epoch and bias as Klass::_prototype_header.
2188 // Special-case a fresh allocation to avoid building nodes:
2189 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn);
2190 if (klass_node == NULL) {
2191 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
2192 klass_node = transform_later( LoadKlassNode::make(_igvn, mem, k_adr, _igvn.type(k_adr)->is_ptr()) );
2193 #ifdef _LP64
2194 if (UseCompressedKlassPointers && klass_node->is_DecodeNKlass()) {
2195 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity");
2196 klass_node->in(1)->init_req(0, ctrl);
2197 } else
2198 #endif
2199 klass_node->init_req(0, ctrl);
2200 }
2201 Node *proto_node = make_load(ctrl, mem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeX_X, TypeX_X->basic_type());
2203 Node* thread = transform_later(new (C) ThreadLocalNode());
2204 Node* cast_thread = transform_later(new (C) CastP2XNode(ctrl, thread));
2205 Node* o_node = transform_later(new (C) OrXNode(cast_thread, proto_node));
2206 Node* x_node = transform_later(new (C) XorXNode(o_node, mark_node));
2208 // Get slow path - mark word does NOT match the value.
2209 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node,
2210 (~markOopDesc::age_mask_in_place), 0);
2211 // region->in(3) is set to fast path - the object is biased to the current thread.
2212 mem_phi->init_req(3, mem);
2215 // Mark word does NOT match the value (thread | Klass::_prototype_header).
2218 // First, check biased pattern.
2219 // Get fast path - _prototype_header has the same biased lock pattern.
2220 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node,
2221 markOopDesc::biased_lock_mask_in_place, 0, true);
2223 not_biased_ctrl = fast_lock_region->in(2); // Slow path
2224 // fast_lock_region->in(2) - the prototype header is no longer biased
2225 // and we have to revoke the bias on this object.
2226 // We are going to try to reset the mark of this object to the prototype
2227 // value and fall through to the CAS-based locking scheme.
2228 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
2229 Node* cas = new (C) StoreXConditionalNode(not_biased_ctrl, mem, adr,
2230 proto_node, mark_node);
2231 transform_later(cas);
2232 Node* proj = transform_later( new (C) SCMemProjNode(cas));
2233 fast_lock_mem_phi->init_req(2, proj);
2236 // Second, check epoch bits.
2237 Node* rebiased_region = new (C) RegionNode(3);
2238 Node* old_phi = new (C) PhiNode( rebiased_region, TypeX_X);
2239 Node* new_phi = new (C) PhiNode( rebiased_region, TypeX_X);
2241 // Get slow path - mark word does NOT match epoch bits.
2242 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node,
2243 markOopDesc::epoch_mask_in_place, 0);
2244 // The epoch of the current bias is not valid, attempt to rebias the object
2245 // toward the current thread.
2246 rebiased_region->init_req(2, epoch_ctrl);
2247 old_phi->init_req(2, mark_node);
2248 new_phi->init_req(2, o_node);
2250 // rebiased_region->in(1) is set to fast path.
2251 // The epoch of the current bias is still valid but we know
2252 // nothing about the owner; it might be set or it might be clear.
2253 Node* cmask = MakeConX(markOopDesc::biased_lock_mask_in_place |
2254 markOopDesc::age_mask_in_place |
2255 markOopDesc::epoch_mask_in_place);
2256 Node* old = transform_later(new (C) AndXNode(mark_node, cmask));
2257 cast_thread = transform_later(new (C) CastP2XNode(ctrl, thread));
2258 Node* new_mark = transform_later(new (C) OrXNode(cast_thread, old));
2259 old_phi->init_req(1, old);
2260 new_phi->init_req(1, new_mark);
2262 transform_later(rebiased_region);
2263 transform_later(old_phi);
2264 transform_later(new_phi);
2266 // Try to acquire the bias of the object using an atomic operation.
2267 // If this fails we will go in to the runtime to revoke the object's bias.
2268 cas = new (C) StoreXConditionalNode(rebiased_region, mem, adr,
2269 new_phi, old_phi);
2270 transform_later(cas);
2271 proj = transform_later( new (C) SCMemProjNode(cas));
2273 // Get slow path - Failed to CAS.
2274 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0);
2275 mem_phi->init_req(4, proj);
2276 // region->in(4) is set to fast path - the object is rebiased to the current thread.
2278 // Failed to CAS.
2279 slow_path = new (C) RegionNode(3);
2280 Node *slow_mem = new (C) PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM);
2282 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control
2283 slow_mem->init_req(1, proj);
2285 // Call CAS-based locking scheme (FastLock node).
2287 transform_later(fast_lock_region);
2288 transform_later(fast_lock_mem_phi);
2290 // Get slow path - FastLock failed to lock the object.
2291 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0);
2292 mem_phi->init_req(2, fast_lock_mem_phi);
2293 // region->in(2) is set to fast path - the object is locked to the current thread.
2295 slow_path->init_req(2, ctrl); // Capture slow-control
2296 slow_mem->init_req(2, fast_lock_mem_phi);
2298 transform_later(slow_path);
2299 transform_later(slow_mem);
2300 // Reset lock's memory edge.
2301 lock->set_req(TypeFunc::Memory, slow_mem);
2303 } else {
2304 region = new (C) RegionNode(3);
2305 // create a Phi for the memory state
2306 mem_phi = new (C) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2308 // Optimize test; set region slot 2
2309 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2310 mem_phi->init_req(2, mem);
2311 }
2313 // Make slow path call
2314 CallNode *call = make_slow_call( (CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(), OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path, obj, box );
2316 extract_call_projections(call);
2318 // Slow path can only throw asynchronous exceptions, which are always
2319 // de-opted. So the compiler thinks the slow-call can never throw an
2320 // exception. If it DOES throw an exception we would need the debug
2321 // info removed first (since if it throws there is no monitor).
2322 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2323 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2325 // Capture slow path
2326 // disconnect fall-through projection from call and create a new one
2327 // hook up users of fall-through projection to region
2328 Node *slow_ctrl = _fallthroughproj->clone();
2329 transform_later(slow_ctrl);
2330 _igvn.hash_delete(_fallthroughproj);
2331 _fallthroughproj->disconnect_inputs(NULL, C);
2332 region->init_req(1, slow_ctrl);
2333 // region inputs are now complete
2334 transform_later(region);
2335 _igvn.replace_node(_fallthroughproj, region);
2337 Node *memproj = transform_later( new(C) ProjNode(call, TypeFunc::Memory) );
2338 mem_phi->init_req(1, memproj );
2339 transform_later(mem_phi);
2340 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2341 }
2343 //------------------------------expand_unlock_node----------------------
2344 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2346 Node* ctrl = unlock->in(TypeFunc::Control);
2347 Node* mem = unlock->in(TypeFunc::Memory);
2348 Node* obj = unlock->obj_node();
2349 Node* box = unlock->box_node();
2351 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2353 // No need for a null check on unlock
2355 // Make the merge point
2356 Node *region;
2357 Node *mem_phi;
2359 if (UseOptoBiasInlining) {
2360 // Check for biased locking unlock case, which is a no-op.
2361 // See the full description in MacroAssembler::biased_locking_exit().
2362 region = new (C) RegionNode(4);
2363 // create a Phi for the memory state
2364 mem_phi = new (C) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2365 mem_phi->init_req(3, mem);
2367 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2368 ctrl = opt_bits_test(ctrl, region, 3, mark_node,
2369 markOopDesc::biased_lock_mask_in_place,
2370 markOopDesc::biased_lock_pattern);
2371 } else {
2372 region = new (C) RegionNode(3);
2373 // create a Phi for the memory state
2374 mem_phi = new (C) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2375 }
2377 FastUnlockNode *funlock = new (C) FastUnlockNode( ctrl, obj, box );
2378 funlock = transform_later( funlock )->as_FastUnlock();
2379 // Optimize test; set region slot 2
2380 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2382 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 );
2384 extract_call_projections(call);
2386 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2387 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2389 // No exceptions for unlocking
2390 // Capture slow path
2391 // disconnect fall-through projection from call and create a new one
2392 // hook up users of fall-through projection to region
2393 Node *slow_ctrl = _fallthroughproj->clone();
2394 transform_later(slow_ctrl);
2395 _igvn.hash_delete(_fallthroughproj);
2396 _fallthroughproj->disconnect_inputs(NULL, C);
2397 region->init_req(1, slow_ctrl);
2398 // region inputs are now complete
2399 transform_later(region);
2400 _igvn.replace_node(_fallthroughproj, region);
2402 Node *memproj = transform_later( new(C) ProjNode(call, TypeFunc::Memory) );
2403 mem_phi->init_req(1, memproj );
2404 mem_phi->init_req(2, mem);
2405 transform_later(mem_phi);
2406 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2407 }
2409 //---------------------------eliminate_macro_nodes----------------------
2410 // Eliminate scalar replaced allocations and associated locks.
2411 void PhaseMacroExpand::eliminate_macro_nodes() {
2412 if (C->macro_count() == 0)
2413 return;
2415 // First, attempt to eliminate locks
2416 int cnt = C->macro_count();
2417 for (int i=0; i < cnt; i++) {
2418 Node *n = C->macro_node(i);
2419 if (n->is_AbstractLock()) { // Lock and Unlock nodes
2420 // Before elimination mark all associated (same box and obj)
2421 // lock and unlock nodes.
2422 mark_eliminated_locking_nodes(n->as_AbstractLock());
2423 }
2424 }
2425 bool progress = true;
2426 while (progress) {
2427 progress = false;
2428 for (int i = C->macro_count(); i > 0; i--) {
2429 Node * n = C->macro_node(i-1);
2430 bool success = false;
2431 debug_only(int old_macro_count = C->macro_count(););
2432 if (n->is_AbstractLock()) {
2433 success = eliminate_locking_node(n->as_AbstractLock());
2434 }
2435 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2436 progress = progress || success;
2437 }
2438 }
2439 // Next, attempt to eliminate allocations
2440 progress = true;
2441 while (progress) {
2442 progress = false;
2443 for (int i = C->macro_count(); i > 0; i--) {
2444 Node * n = C->macro_node(i-1);
2445 bool success = false;
2446 debug_only(int old_macro_count = C->macro_count(););
2447 switch (n->class_id()) {
2448 case Node::Class_Allocate:
2449 case Node::Class_AllocateArray:
2450 success = eliminate_allocate_node(n->as_Allocate());
2451 break;
2452 case Node::Class_CallStaticJava:
2453 success = eliminate_boxing_node(n->as_CallStaticJava());
2454 break;
2455 case Node::Class_Lock:
2456 case Node::Class_Unlock:
2457 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2458 break;
2459 default:
2460 assert(n->Opcode() == Op_LoopLimit ||
2461 n->Opcode() == Op_Opaque1 ||
2462 n->Opcode() == Op_Opaque2, "unknown node type in macro list");
2463 }
2464 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2465 progress = progress || success;
2466 }
2467 }
2468 }
2470 //------------------------------expand_macro_nodes----------------------
2471 // Returns true if a failure occurred.
2472 bool PhaseMacroExpand::expand_macro_nodes() {
2473 // Last attempt to eliminate macro nodes.
2474 eliminate_macro_nodes();
2476 // Make sure expansion will not cause node limit to be exceeded.
2477 // Worst case is a macro node gets expanded into about 50 nodes.
2478 // Allow 50% more for optimization.
2479 if (C->check_node_count(C->macro_count() * 75, "out of nodes before macro expansion" ) )
2480 return true;
2482 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2483 bool progress = true;
2484 while (progress) {
2485 progress = false;
2486 for (int i = C->macro_count(); i > 0; i--) {
2487 Node * n = C->macro_node(i-1);
2488 bool success = false;
2489 debug_only(int old_macro_count = C->macro_count(););
2490 if (n->Opcode() == Op_LoopLimit) {
2491 // Remove it from macro list and put on IGVN worklist to optimize.
2492 C->remove_macro_node(n);
2493 _igvn._worklist.push(n);
2494 success = true;
2495 } else if (n->Opcode() == Op_CallStaticJava) {
2496 // Remove it from macro list and put on IGVN worklist to optimize.
2497 C->remove_macro_node(n);
2498 _igvn._worklist.push(n);
2499 success = true;
2500 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) {
2501 _igvn.replace_node(n, n->in(1));
2502 success = true;
2503 }
2504 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2505 progress = progress || success;
2506 }
2507 }
2509 // expand "macro" nodes
2510 // nodes are removed from the macro list as they are processed
2511 while (C->macro_count() > 0) {
2512 int macro_count = C->macro_count();
2513 Node * n = C->macro_node(macro_count-1);
2514 assert(n->is_macro(), "only macro nodes expected here");
2515 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) {
2516 // node is unreachable, so don't try to expand it
2517 C->remove_macro_node(n);
2518 continue;
2519 }
2520 switch (n->class_id()) {
2521 case Node::Class_Allocate:
2522 expand_allocate(n->as_Allocate());
2523 break;
2524 case Node::Class_AllocateArray:
2525 expand_allocate_array(n->as_AllocateArray());
2526 break;
2527 case Node::Class_Lock:
2528 expand_lock_node(n->as_Lock());
2529 break;
2530 case Node::Class_Unlock:
2531 expand_unlock_node(n->as_Unlock());
2532 break;
2533 default:
2534 assert(false, "unknown node type in macro list");
2535 }
2536 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2537 if (C->failing()) return true;
2538 }
2540 _igvn.set_delay_transform(false);
2541 _igvn.optimize();
2542 if (C->failing()) return true;
2543 return false;
2544 }