1 /*
2 * Copyright (c) 2005, 2011, 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 */
24
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"
44
45
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 }
67
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");
74
75 Dict* sosn_map = new Dict(cmpkey,hashkey);
76 for (uint i = old_dbg_start; i < oldcall->req(); i++) {
77 Node* old_in = oldcall->in(i);
78 // Clone old SafePointScalarObjectNodes, adjusting their field contents.
79 if (old_in != NULL && old_in->is_SafePointScalarObject()) {
80 SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject();
81 uint old_unique = C->unique();
82 Node* new_in = old_sosn->clone(jvms_adj, sosn_map);
83 if (old_unique != C->unique()) {
84 new_in->set_req(0, C->root()); // reset control edge
85 new_in = transform_later(new_in); // Register new node.
86 }
87 old_in = new_in;
88 }
89 newcall->add_req(old_in);
90 }
91
92 newcall->set_jvms(oldcall->jvms());
93 for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) {
94 jvms->set_map(newcall);
95 jvms->set_locoff(jvms->locoff()+jvms_adj);
96 jvms->set_stkoff(jvms->stkoff()+jvms_adj);
97 jvms->set_monoff(jvms->monoff()+jvms_adj);
98 jvms->set_scloff(jvms->scloff()+jvms_adj);
99 jvms->set_endoff(jvms->endoff()+jvms_adj);
100 }
101 }
102
103 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
104 Node* cmp;
105 if (mask != 0) {
106 Node* and_node = transform_later(new (C) AndXNode(word, MakeConX(mask)));
107 cmp = transform_later(new (C) CmpXNode(and_node, MakeConX(bits)));
108 } else {
109 cmp = word;
110 }
111 Node* bol = transform_later(new (C) BoolNode(cmp, BoolTest::ne));
112 IfNode* iff = new (C) IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
113 transform_later(iff);
114
115 // Fast path taken.
116 Node *fast_taken = transform_later( new (C) IfFalseNode(iff) );
117
118 // Fast path not-taken, i.e. slow path
119 Node *slow_taken = transform_later( new (C) IfTrueNode(iff) );
120
121 if (return_fast_path) {
122 region->init_req(edge, slow_taken); // Capture slow-control
123 return fast_taken;
124 } else {
125 region->init_req(edge, fast_taken); // Capture fast-control
126 return slow_taken;
127 }
128 }
129
130 //--------------------copy_predefined_input_for_runtime_call--------------------
131 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
132 // Set fixed predefined input arguments
133 call->init_req( TypeFunc::Control, ctrl );
134 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) );
135 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
136 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
137 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
138 }
139
140 //------------------------------make_slow_call---------------------------------
141 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) {
142
143 // Slow-path call
144 CallNode *call = leaf_name
145 ? (CallNode*)new (C) CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
146 : (CallNode*)new (C) CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM );
147
148 // Slow path call has no side-effects, uses few values
149 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
150 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0);
151 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1);
152 copy_call_debug_info(oldcall, call);
153 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
154 _igvn.replace_node(oldcall, call);
155 transform_later(call);
156
157 return call;
158 }
159
160 void PhaseMacroExpand::extract_call_projections(CallNode *call) {
161 _fallthroughproj = NULL;
162 _fallthroughcatchproj = NULL;
163 _ioproj_fallthrough = NULL;
164 _ioproj_catchall = NULL;
165 _catchallcatchproj = NULL;
166 _memproj_fallthrough = NULL;
167 _memproj_catchall = NULL;
168 _resproj = NULL;
169 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) {
170 ProjNode *pn = call->fast_out(i)->as_Proj();
171 switch (pn->_con) {
172 case TypeFunc::Control:
173 {
174 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj
175 _fallthroughproj = pn;
176 DUIterator_Fast jmax, j = pn->fast_outs(jmax);
177 const Node *cn = pn->fast_out(j);
178 if (cn->is_Catch()) {
179 ProjNode *cpn = NULL;
180 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) {
181 cpn = cn->fast_out(k)->as_Proj();
182 assert(cpn->is_CatchProj(), "must be a CatchProjNode");
183 if (cpn->_con == CatchProjNode::fall_through_index)
184 _fallthroughcatchproj = cpn;
185 else {
186 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index.");
187 _catchallcatchproj = cpn;
188 }
189 }
190 }
191 break;
192 }
193 case TypeFunc::I_O:
194 if (pn->_is_io_use)
195 _ioproj_catchall = pn;
196 else
197 _ioproj_fallthrough = pn;
198 break;
199 case TypeFunc::Memory:
200 if (pn->_is_io_use)
201 _memproj_catchall = pn;
202 else
203 _memproj_fallthrough = pn;
204 break;
205 case TypeFunc::Parms:
206 _resproj = pn;
207 break;
208 default:
209 assert(false, "unexpected projection from allocation node.");
210 }
211 }
212
213 }
214
215 // Eliminate a card mark sequence. p2x is a ConvP2XNode
216 void PhaseMacroExpand::eliminate_card_mark(Node* p2x) {
217 assert(p2x->Opcode() == Op_CastP2X, "ConvP2XNode required");
218 if (!UseG1GC) {
219 // vanilla/CMS post barrier
220 Node *shift = p2x->unique_out();
221 Node *addp = shift->unique_out();
222 for (DUIterator_Last jmin, j = addp->last_outs(jmin); j >= jmin; --j) {
223 Node *mem = addp->last_out(j);
224 if (UseCondCardMark && mem->is_Load()) {
225 assert(mem->Opcode() == Op_LoadB, "unexpected code shape");
226 // The load is checking if the card has been written so
227 // replace it with zero to fold the test.
228 _igvn.replace_node(mem, intcon(0));
229 continue;
230 }
231 assert(mem->is_Store(), "store required");
232 _igvn.replace_node(mem, mem->in(MemNode::Memory));
233 }
234 } else {
235 // G1 pre/post barriers
236 assert(p2x->outcnt() <= 2, "expects 1 or 2 users: Xor and URShift nodes");
237 // It could be only one user, URShift node, in Object.clone() instrinsic
238 // but the new allocation is passed to arraycopy stub and it could not
239 // be scalar replaced. So we don't check the case.
240
241 // An other case of only one user (Xor) is when the value check for NULL
242 // in G1 post barrier is folded after CCP so the code which used URShift
243 // is removed.
244
245 // Take Region node before eliminating post barrier since it also
246 // eliminates CastP2X node when it has only one user.
247 Node* this_region = p2x->in(0);
248 assert(this_region != NULL, "");
249
250 // Remove G1 post barrier.
251
252 // Search for CastP2X->Xor->URShift->Cmp path which
253 // checks if the store done to a different from the value's region.
254 // And replace Cmp with #0 (false) to collapse G1 post barrier.
255 Node* xorx = NULL;
256 for (DUIterator_Fast imax, i = p2x->fast_outs(imax); i < imax; i++) {
257 Node* u = p2x->fast_out(i);
258 if (u->Opcode() == Op_XorX) {
259 xorx = u;
260 break;
261 }
262 }
263 assert(xorx != NULL, "missing G1 post barrier");
264 Node* shift = xorx->unique_out();
265 Node* cmpx = shift->unique_out();
266 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() &&
267 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne,
268 "missing region check in G1 post barrier");
269 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ));
270
271 // Remove G1 pre barrier.
272
273 // Search "if (marking != 0)" check and set it to "false".
274 // There is no G1 pre barrier if previous stored value is NULL
275 // (for example, after initialization).
276 if (this_region->is_Region() && this_region->req() == 3) {
277 int ind = 1;
278 if (!this_region->in(ind)->is_IfFalse()) {
279 ind = 2;
280 }
281 if (this_region->in(ind)->is_IfFalse()) {
282 Node* bol = this_region->in(ind)->in(0)->in(1);
283 assert(bol->is_Bool(), "");
284 cmpx = bol->in(1);
285 if (bol->as_Bool()->_test._test == BoolTest::ne &&
286 cmpx->is_Cmp() && cmpx->in(2) == intcon(0) &&
287 cmpx->in(1)->is_Load()) {
288 Node* adr = cmpx->in(1)->as_Load()->in(MemNode::Address);
289 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() +
290 PtrQueue::byte_offset_of_active());
291 if (adr->is_AddP() && adr->in(AddPNode::Base) == top() &&
292 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
293 adr->in(AddPNode::Offset) == MakeConX(marking_offset)) {
294 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ));
295 }
296 }
297 }
298 }
299 // Now CastP2X can be removed since it is used only on dead path
300 // which currently still alive until igvn optimize it.
301 assert(p2x->outcnt() == 0 || p2x->unique_out()->Opcode() == Op_URShiftX, "");
302 _igvn.replace_node(p2x, top());
303 }
304 }
305
306 // Search for a memory operation for the specified memory slice.
307 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
308 Node *orig_mem = mem;
309 Node *alloc_mem = alloc->in(TypeFunc::Memory);
310 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
311 while (true) {
312 if (mem == alloc_mem || mem == start_mem ) {
313 return mem; // hit one of our sentinels
314 } else if (mem->is_MergeMem()) {
315 mem = mem->as_MergeMem()->memory_at(alias_idx);
316 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
317 Node *in = mem->in(0);
318 // we can safely skip over safepoints, calls, locks and membars because we
319 // already know that the object is safe to eliminate.
320 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
321 return in;
322 } else if (in->is_Call()) {
323 CallNode *call = in->as_Call();
324 if (!call->may_modify(tinst, phase)) {
325 mem = call->in(TypeFunc::Memory);
326 }
327 mem = in->in(TypeFunc::Memory);
328 } else if (in->is_MemBar()) {
329 mem = in->in(TypeFunc::Memory);
330 } else {
331 assert(false, "unexpected projection");
332 }
333 } else if (mem->is_Store()) {
334 const TypePtr* atype = mem->as_Store()->adr_type();
335 int adr_idx = Compile::current()->get_alias_index(atype);
336 if (adr_idx == alias_idx) {
337 assert(atype->isa_oopptr(), "address type must be oopptr");
338 int adr_offset = atype->offset();
339 uint adr_iid = atype->is_oopptr()->instance_id();
340 // Array elements references have the same alias_idx
341 // but different offset and different instance_id.
342 if (adr_offset == offset && adr_iid == alloc->_idx)
343 return mem;
344 } else {
345 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
346 }
347 mem = mem->in(MemNode::Memory);
348 } else if (mem->is_ClearArray()) {
349 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
350 // Can not bypass initialization of the instance
351 // we are looking.
352 debug_only(intptr_t offset;)
353 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
354 InitializeNode* init = alloc->as_Allocate()->initialization();
355 // We are looking for stored value, return Initialize node
356 // or memory edge from Allocate node.
357 if (init != NULL)
358 return init;
359 else
360 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers).
361 }
362 // Otherwise skip it (the call updated 'mem' value).
363 } else if (mem->Opcode() == Op_SCMemProj) {
364 assert(mem->in(0)->is_LoadStore(), "sanity");
365 const TypePtr* atype = mem->in(0)->in(MemNode::Address)->bottom_type()->is_ptr();
366 int adr_idx = Compile::current()->get_alias_index(atype);
367 if (adr_idx == alias_idx) {
368 assert(false, "Object is not scalar replaceable if a LoadStore node access its field");
369 return NULL;
370 }
371 mem = mem->in(0)->in(MemNode::Memory);
372 } else {
373 return mem;
374 }
375 assert(mem != orig_mem, "dead memory loop");
376 }
377 }
378
379 //
380 // Given a Memory Phi, compute a value Phi containing the values from stores
381 // on the input paths.
382 // Note: this function is recursive, its depth is limied by the "level" argument
383 // Returns the computed Phi, or NULL if it cannot compute it.
384 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) {
385 assert(mem->is_Phi(), "sanity");
386 int alias_idx = C->get_alias_index(adr_t);
387 int offset = adr_t->offset();
388 int instance_id = adr_t->instance_id();
389
390 // Check if an appropriate value phi already exists.
391 Node* region = mem->in(0);
392 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
393 Node* phi = region->fast_out(k);
394 if (phi->is_Phi() && phi != mem &&
395 phi->as_Phi()->is_same_inst_field(phi_type, instance_id, alias_idx, offset)) {
396 return phi;
397 }
398 }
399 // Check if an appropriate new value phi already exists.
400 Node* new_phi = value_phis->find(mem->_idx);
401 if (new_phi != NULL)
402 return new_phi;
403
404 if (level <= 0) {
405 return NULL; // Give up: phi tree too deep
406 }
407 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
408 Node *alloc_mem = alloc->in(TypeFunc::Memory);
409
410 uint length = mem->req();
411 GrowableArray <Node *> values(length, length, NULL, false);
412
413 // create a new Phi for the value
414 PhiNode *phi = new (C) PhiNode(mem->in(0), phi_type, NULL, instance_id, alias_idx, offset);
415 transform_later(phi);
416 value_phis->push(phi, mem->_idx);
417
418 for (uint j = 1; j < length; j++) {
419 Node *in = mem->in(j);
420 if (in == NULL || in->is_top()) {
421 values.at_put(j, in);
422 } else {
423 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
424 if (val == start_mem || val == alloc_mem) {
425 // hit a sentinel, return appropriate 0 value
426 values.at_put(j, _igvn.zerocon(ft));
427 continue;
428 }
429 if (val->is_Initialize()) {
430 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
431 }
432 if (val == NULL) {
433 return NULL; // can't find a value on this path
434 }
435 if (val == mem) {
436 values.at_put(j, mem);
437 } else if (val->is_Store()) {
438 values.at_put(j, val->in(MemNode::ValueIn));
439 } else if(val->is_Proj() && val->in(0) == alloc) {
440 values.at_put(j, _igvn.zerocon(ft));
441 } else if (val->is_Phi()) {
442 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
443 if (val == NULL) {
444 return NULL;
445 }
446 values.at_put(j, val);
447 } else if (val->Opcode() == Op_SCMemProj) {
448 assert(val->in(0)->is_LoadStore(), "sanity");
449 assert(false, "Object is not scalar replaceable if a LoadStore node access its field");
450 return NULL;
451 } else {
452 #ifdef ASSERT
453 val->dump();
454 assert(false, "unknown node on this path");
455 #endif
456 return NULL; // unknown node on this path
457 }
458 }
459 }
460 // Set Phi's inputs
461 for (uint j = 1; j < length; j++) {
462 if (values.at(j) == mem) {
463 phi->init_req(j, phi);
464 } else {
465 phi->init_req(j, values.at(j));
466 }
467 }
468 return phi;
469 }
470
471 // Search the last value stored into the object's field.
472 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, Node *alloc) {
473 assert(adr_t->is_known_instance_field(), "instance required");
474 int instance_id = adr_t->instance_id();
475 assert((uint)instance_id == alloc->_idx, "wrong allocation");
476
477 int alias_idx = C->get_alias_index(adr_t);
478 int offset = adr_t->offset();
479 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
480 Node *alloc_ctrl = alloc->in(TypeFunc::Control);
481 Node *alloc_mem = alloc->in(TypeFunc::Memory);
482 Arena *a = Thread::current()->resource_area();
483 VectorSet visited(a);
484
485
486 bool done = sfpt_mem == alloc_mem;
487 Node *mem = sfpt_mem;
488 while (!done) {
489 if (visited.test_set(mem->_idx)) {
490 return NULL; // found a loop, give up
491 }
492 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
493 if (mem == start_mem || mem == alloc_mem) {
494 done = true; // hit a sentinel, return appropriate 0 value
495 } else if (mem->is_Initialize()) {
496 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
497 if (mem == NULL) {
498 done = true; // Something go wrong.
499 } else if (mem->is_Store()) {
500 const TypePtr* atype = mem->as_Store()->adr_type();
501 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
502 done = true;
503 }
504 } else if (mem->is_Store()) {
505 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
506 assert(atype != NULL, "address type must be oopptr");
507 assert(C->get_alias_index(atype) == alias_idx &&
508 atype->is_known_instance_field() && atype->offset() == offset &&
509 atype->instance_id() == instance_id, "store is correct memory slice");
510 done = true;
511 } else if (mem->is_Phi()) {
512 // try to find a phi's unique input
513 Node *unique_input = NULL;
514 Node *top = C->top();
515 for (uint i = 1; i < mem->req(); i++) {
516 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
517 if (n == NULL || n == top || n == mem) {
518 continue;
519 } else if (unique_input == NULL) {
520 unique_input = n;
521 } else if (unique_input != n) {
522 unique_input = top;
523 break;
524 }
525 }
526 if (unique_input != NULL && unique_input != top) {
527 mem = unique_input;
528 } else {
529 done = true;
530 }
531 } else {
532 assert(false, "unexpected node");
533 }
534 }
535 if (mem != NULL) {
536 if (mem == start_mem || mem == alloc_mem) {
537 // hit a sentinel, return appropriate 0 value
538 return _igvn.zerocon(ft);
539 } else if (mem->is_Store()) {
540 return mem->in(MemNode::ValueIn);
541 } else if (mem->is_Phi()) {
542 // attempt to produce a Phi reflecting the values on the input paths of the Phi
543 Node_Stack value_phis(a, 8);
544 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
545 if (phi != NULL) {
546 return phi;
547 } else {
548 // Kill all new Phis
549 while(value_phis.is_nonempty()) {
550 Node* n = value_phis.node();
551 _igvn.replace_node(n, C->top());
552 value_phis.pop();
553 }
554 }
555 }
556 }
557 // Something go wrong.
558 return NULL;
559 }
560
561 // Check the possibility of scalar replacement.
562 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
563 // Scan the uses of the allocation to check for anything that would
564 // prevent us from eliminating it.
565 NOT_PRODUCT( const char* fail_eliminate = NULL; )
566 DEBUG_ONLY( Node* disq_node = NULL; )
567 bool can_eliminate = true;
568
569 Node* res = alloc->result_cast();
570 const TypeOopPtr* res_type = NULL;
571 if (res == NULL) {
572 // All users were eliminated.
573 } else if (!res->is_CheckCastPP()) {
574 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
575 can_eliminate = false;
576 } else {
577 res_type = _igvn.type(res)->isa_oopptr();
578 if (res_type == NULL) {
579 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
580 can_eliminate = false;
581 } else if (res_type->isa_aryptr()) {
582 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
583 if (length < 0) {
584 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
585 can_eliminate = false;
586 }
587 }
588 }
589
590 if (can_eliminate && res != NULL) {
591 for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
592 j < jmax && can_eliminate; j++) {
593 Node* use = res->fast_out(j);
594
595 if (use->is_AddP()) {
596 const TypePtr* addp_type = _igvn.type(use)->is_ptr();
597 int offset = addp_type->offset();
598
599 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
600 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";)
601 can_eliminate = false;
602 break;
603 }
604 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
605 k < kmax && can_eliminate; k++) {
606 Node* n = use->fast_out(k);
607 if (!n->is_Store() && n->Opcode() != Op_CastP2X) {
608 DEBUG_ONLY(disq_node = n;)
609 if (n->is_Load() || n->is_LoadStore()) {
610 NOT_PRODUCT(fail_eliminate = "Field load";)
611 } else {
612 NOT_PRODUCT(fail_eliminate = "Not store field referrence";)
613 }
614 can_eliminate = false;
615 }
616 }
617 } else if (use->is_SafePoint()) {
618 SafePointNode* sfpt = use->as_SafePoint();
619 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
620 // Object is passed as argument.
621 DEBUG_ONLY(disq_node = use;)
622 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
623 can_eliminate = false;
624 }
625 Node* sfptMem = sfpt->memory();
626 if (sfptMem == NULL || sfptMem->is_top()) {
627 DEBUG_ONLY(disq_node = use;)
628 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";)
629 can_eliminate = false;
630 } else {
631 safepoints.append_if_missing(sfpt);
632 }
633 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
634 if (use->is_Phi()) {
635 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
636 NOT_PRODUCT(fail_eliminate = "Object is return value";)
637 } else {
638 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
639 }
640 DEBUG_ONLY(disq_node = use;)
641 } else {
642 if (use->Opcode() == Op_Return) {
643 NOT_PRODUCT(fail_eliminate = "Object is return value";)
644 }else {
645 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
646 }
647 DEBUG_ONLY(disq_node = use;)
648 }
649 can_eliminate = false;
650 }
651 }
652 }
653
654 #ifndef PRODUCT
655 if (PrintEliminateAllocations) {
656 if (can_eliminate) {
657 tty->print("Scalar ");
658 if (res == NULL)
659 alloc->dump();
660 else
661 res->dump();
662 } else {
663 tty->print("NotScalar (%s)", fail_eliminate);
664 if (res == NULL)
665 alloc->dump();
666 else
667 res->dump();
668 #ifdef ASSERT
669 if (disq_node != NULL) {
670 tty->print(" >>>> ");
671 disq_node->dump();
672 }
673 #endif /*ASSERT*/
674 }
675 }
676 #endif
677 return can_eliminate;
678 }
679
680 // Do scalar replacement.
681 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
682 GrowableArray <SafePointNode *> safepoints_done;
683
684 ciKlass* klass = NULL;
685 ciInstanceKlass* iklass = NULL;
686 int nfields = 0;
687 int array_base;
688 int element_size;
689 BasicType basic_elem_type;
690 ciType* elem_type;
691
692 Node* res = alloc->result_cast();
693 const TypeOopPtr* res_type = NULL;
694 if (res != NULL) { // Could be NULL when there are no users
695 res_type = _igvn.type(res)->isa_oopptr();
696 }
697
698 if (res != NULL) {
699 klass = res_type->klass();
700 if (res_type->isa_instptr()) {
701 // find the fields of the class which will be needed for safepoint debug information
702 assert(klass->is_instance_klass(), "must be an instance klass.");
703 iklass = klass->as_instance_klass();
704 nfields = iklass->nof_nonstatic_fields();
705 } else {
706 // find the array's elements which will be needed for safepoint debug information
707 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
708 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass.");
709 elem_type = klass->as_array_klass()->element_type();
710 basic_elem_type = elem_type->basic_type();
711 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
712 element_size = type2aelembytes(basic_elem_type);
713 }
714 }
715 //
716 // Process the safepoint uses
717 //
718 while (safepoints.length() > 0) {
719 SafePointNode* sfpt = safepoints.pop();
720 Node* mem = sfpt->memory();
721 uint first_ind = sfpt->req();
722 SafePointScalarObjectNode* sobj = new (C) SafePointScalarObjectNode(res_type,
723 #ifdef ASSERT
724 alloc,
725 #endif
726 first_ind, nfields);
727 sobj->init_req(0, C->root());
728 transform_later(sobj);
729
730 // Scan object's fields adding an input to the safepoint for each field.
731 for (int j = 0; j < nfields; j++) {
732 intptr_t offset;
733 ciField* field = NULL;
734 if (iklass != NULL) {
735 field = iklass->nonstatic_field_at(j);
736 offset = field->offset();
737 elem_type = field->type();
738 basic_elem_type = field->layout_type();
739 } else {
740 offset = array_base + j * (intptr_t)element_size;
741 }
742
743 const Type *field_type;
744 // The next code is taken from Parse::do_get_xxx().
745 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) {
746 if (!elem_type->is_loaded()) {
747 field_type = TypeInstPtr::BOTTOM;
748 } else if (field != NULL && field->is_constant() && field->is_static()) {
749 // This can happen if the constant oop is non-perm.
750 ciObject* con = field->constant_value().as_object();
751 // Do not "join" in the previous type; it doesn't add value,
752 // and may yield a vacuous result if the field is of interface type.
753 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
754 assert(field_type != NULL, "field singleton type must be consistent");
755 } else {
756 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
757 }
758 if (UseCompressedOops) {
759 field_type = field_type->make_narrowoop();
760 basic_elem_type = T_NARROWOOP;
761 }
762 } else {
763 field_type = Type::get_const_basic_type(basic_elem_type);
764 }
765
766 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr();
767
768 Node *field_val = value_from_mem(mem, basic_elem_type, field_type, field_addr_type, alloc);
769 if (field_val == NULL) {
770 // We weren't able to find a value for this field,
771 // give up on eliminating this allocation.
772
773 // Remove any extra entries we added to the safepoint.
774 uint last = sfpt->req() - 1;
775 for (int k = 0; k < j; k++) {
776 sfpt->del_req(last--);
777 }
778 // rollback processed safepoints
779 while (safepoints_done.length() > 0) {
780 SafePointNode* sfpt_done = safepoints_done.pop();
781 // remove any extra entries we added to the safepoint
782 last = sfpt_done->req() - 1;
783 for (int k = 0; k < nfields; k++) {
784 sfpt_done->del_req(last--);
785 }
786 JVMState *jvms = sfpt_done->jvms();
787 jvms->set_endoff(sfpt_done->req());
788 // Now make a pass over the debug information replacing any references
789 // to SafePointScalarObjectNode with the allocated object.
790 int start = jvms->debug_start();
791 int end = jvms->debug_end();
792 for (int i = start; i < end; i++) {
793 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
794 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
795 if (scobj->first_index() == sfpt_done->req() &&
796 scobj->n_fields() == (uint)nfields) {
797 assert(scobj->alloc() == alloc, "sanity");
798 sfpt_done->set_req(i, res);
799 }
800 }
801 }
802 }
803 #ifndef PRODUCT
804 if (PrintEliminateAllocations) {
805 if (field != NULL) {
806 tty->print("=== At SafePoint node %d can't find value of Field: ",
807 sfpt->_idx);
808 field->print();
809 int field_idx = C->get_alias_index(field_addr_type);
810 tty->print(" (alias_idx=%d)", field_idx);
811 } else { // Array's element
812 tty->print("=== At SafePoint node %d can't find value of array element [%d]",
813 sfpt->_idx, j);
814 }
815 tty->print(", which prevents elimination of: ");
816 if (res == NULL)
817 alloc->dump();
818 else
819 res->dump();
820 }
821 #endif
822 return false;
823 }
824 if (UseCompressedOops && field_type->isa_narrowoop()) {
825 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
826 // to be able scalar replace the allocation.
827 if (field_val->is_EncodeP()) {
828 field_val = field_val->in(1);
829 } else {
830 field_val = transform_later(new (C) DecodeNNode(field_val, field_val->get_ptr_type()));
831 }
832 }
833 sfpt->add_req(field_val);
834 }
835 JVMState *jvms = sfpt->jvms();
836 jvms->set_endoff(sfpt->req());
837 // Now make a pass over the debug information replacing any references
838 // to the allocated object with "sobj"
839 int start = jvms->debug_start();
840 int end = jvms->debug_end();
841 for (int i = start; i < end; i++) {
842 if (sfpt->in(i) == res) {
843 sfpt->set_req(i, sobj);
844 }
845 }
846 safepoints_done.append_if_missing(sfpt); // keep it for rollback
847 }
848 return true;
849 }
850
851 // Process users of eliminated allocation.
852 void PhaseMacroExpand::process_users_of_allocation(AllocateNode *alloc) {
853 Node* res = alloc->result_cast();
854 if (res != NULL) {
855 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
856 Node *use = res->last_out(j);
857 uint oc1 = res->outcnt();
858
859 if (use->is_AddP()) {
860 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
861 Node *n = use->last_out(k);
862 uint oc2 = use->outcnt();
863 if (n->is_Store()) {
864 #ifdef ASSERT
865 // Verify that there is no dependent MemBarVolatile nodes,
866 // they should be removed during IGVN, see MemBarNode::Ideal().
867 for (DUIterator_Fast pmax, p = n->fast_outs(pmax);
868 p < pmax; p++) {
869 Node* mb = n->fast_out(p);
870 assert(mb->is_Initialize() || !mb->is_MemBar() ||
871 mb->req() <= MemBarNode::Precedent ||
872 mb->in(MemBarNode::Precedent) != n,
873 "MemBarVolatile should be eliminated for non-escaping object");
874 }
875 #endif
876 _igvn.replace_node(n, n->in(MemNode::Memory));
877 } else {
878 eliminate_card_mark(n);
879 }
880 k -= (oc2 - use->outcnt());
881 }
882 } else {
883 eliminate_card_mark(use);
884 }
885 j -= (oc1 - res->outcnt());
886 }
887 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
888 _igvn.remove_dead_node(res);
889 }
890
891 //
892 // Process other users of allocation's projections
893 //
894 if (_resproj != NULL && _resproj->outcnt() != 0) {
895 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) {
896 Node *use = _resproj->last_out(j);
897 uint oc1 = _resproj->outcnt();
898 if (use->is_Initialize()) {
899 // Eliminate Initialize node.
900 InitializeNode *init = use->as_Initialize();
901 assert(init->outcnt() <= 2, "only a control and memory projection expected");
902 Node *ctrl_proj = init->proj_out(TypeFunc::Control);
903 if (ctrl_proj != NULL) {
904 assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection");
905 _igvn.replace_node(ctrl_proj, _fallthroughcatchproj);
906 }
907 Node *mem_proj = init->proj_out(TypeFunc::Memory);
908 if (mem_proj != NULL) {
909 Node *mem = init->in(TypeFunc::Memory);
910 #ifdef ASSERT
911 if (mem->is_MergeMem()) {
912 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection");
913 } else {
914 assert(mem == _memproj_fallthrough, "allocation memory projection");
915 }
916 #endif
917 _igvn.replace_node(mem_proj, mem);
918 }
919 } else if (use->is_AddP()) {
920 // raw memory addresses used only by the initialization
921 _igvn.replace_node(use, C->top());
922 } else {
923 assert(false, "only Initialize or AddP expected");
924 }
925 j -= (oc1 - _resproj->outcnt());
926 }
927 }
928 if (_fallthroughcatchproj != NULL) {
929 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control));
930 }
931 if (_memproj_fallthrough != NULL) {
932 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory));
933 }
934 if (_memproj_catchall != NULL) {
935 _igvn.replace_node(_memproj_catchall, C->top());
936 }
937 if (_ioproj_fallthrough != NULL) {
938 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O));
939 }
940 if (_ioproj_catchall != NULL) {
941 _igvn.replace_node(_ioproj_catchall, C->top());
942 }
943 if (_catchallcatchproj != NULL) {
944 _igvn.replace_node(_catchallcatchproj, C->top());
945 }
946 }
947
948 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
949
950 if (!EliminateAllocations || !alloc->_is_scalar_replaceable) {
951 return false;
952 }
953
954 extract_call_projections(alloc);
955
956 GrowableArray <SafePointNode *> safepoints;
957 if (!can_eliminate_allocation(alloc, safepoints)) {
958 return false;
959 }
960
961 if (!scalar_replacement(alloc, safepoints)) {
962 return false;
963 }
964
965 CompileLog* log = C->log();
966 if (log != NULL) {
967 Node* klass = alloc->in(AllocateNode::KlassNode);
968 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
969 log->head("eliminate_allocation type='%d'",
970 log->identify(tklass->klass()));
971 JVMState* p = alloc->jvms();
972 while (p != NULL) {
973 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
974 p = p->caller();
975 }
976 log->tail("eliminate_allocation");
977 }
978
979 process_users_of_allocation(alloc);
980
981 #ifndef PRODUCT
982 if (PrintEliminateAllocations) {
983 if (alloc->is_AllocateArray())
984 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
985 else
986 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
987 }
988 #endif
989
990 return true;
991 }
992
993
994 //---------------------------set_eden_pointers-------------------------
995 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) {
996 if (UseTLAB) { // Private allocation: load from TLS
997 Node* thread = transform_later(new (C) ThreadLocalNode());
998 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset());
999 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset());
1000 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset);
1001 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset);
1002 } else { // Shared allocation: load from globals
1003 CollectedHeap* ch = Universe::heap();
1004 address top_adr = (address)ch->top_addr();
1005 address end_adr = (address)ch->end_addr();
1006 eden_top_adr = makecon(TypeRawPtr::make(top_adr));
1007 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr);
1008 }
1009 }
1010
1011
1012 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
1013 Node* adr = basic_plus_adr(base, offset);
1014 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
1015 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt);
1016 transform_later(value);
1017 return value;
1018 }
1019
1020
1021 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
1022 Node* adr = basic_plus_adr(base, offset);
1023 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt);
1024 transform_later(mem);
1025 return mem;
1026 }
1027
1028 //=============================================================================
1029 //
1030 // A L L O C A T I O N
1031 //
1032 // Allocation attempts to be fast in the case of frequent small objects.
1033 // It breaks down like this:
1034 //
1035 // 1) Size in doublewords is computed. This is a constant for objects and
1036 // variable for most arrays. Doubleword units are used to avoid size
1037 // overflow of huge doubleword arrays. We need doublewords in the end for
1038 // rounding.
1039 //
1040 // 2) Size is checked for being 'too large'. Too-large allocations will go
1041 // the slow path into the VM. The slow path can throw any required
1042 // exceptions, and does all the special checks for very large arrays. The
1043 // size test can constant-fold away for objects. For objects with
1044 // finalizers it constant-folds the otherway: you always go slow with
1045 // finalizers.
1046 //
1047 // 3) If NOT using TLABs, this is the contended loop-back point.
1048 // Load-Locked the heap top. If using TLABs normal-load the heap top.
1049 //
1050 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
1051 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
1052 // "size*8" we always enter the VM, where "largish" is a constant picked small
1053 // enough that there's always space between the eden max and 4Gig (old space is
1054 // there so it's quite large) and large enough that the cost of entering the VM
1055 // is dwarfed by the cost to initialize the space.
1056 //
1057 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1058 // down. If contended, repeat at step 3. If using TLABs normal-store
1059 // adjusted heap top back down; there is no contention.
1060 //
1061 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
1062 // fields.
1063 //
1064 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1065 // oop flavor.
1066 //
1067 //=============================================================================
1068 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1069 // Allocations bigger than this always go the slow route.
1070 // This value must be small enough that allocation attempts that need to
1071 // trigger exceptions go the slow route. Also, it must be small enough so
1072 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1073 //=============================================================================j//
1074 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1075 // The allocator will coalesce int->oop copies away. See comment in
1076 // coalesce.cpp about how this works. It depends critically on the exact
1077 // code shape produced here, so if you are changing this code shape
1078 // make sure the GC info for the heap-top is correct in and around the
1079 // slow-path call.
1080 //
1081
1082 void PhaseMacroExpand::expand_allocate_common(
1083 AllocateNode* alloc, // allocation node to be expanded
1084 Node* length, // array length for an array allocation
1085 const TypeFunc* slow_call_type, // Type of slow call
1086 address slow_call_address // Address of slow call
1087 )
1088 {
1089
1090 Node* ctrl = alloc->in(TypeFunc::Control);
1091 Node* mem = alloc->in(TypeFunc::Memory);
1092 Node* i_o = alloc->in(TypeFunc::I_O);
1093 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1094 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1095 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1096
1097 assert(ctrl != NULL, "must have control");
1098 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1099 // they will not be used if "always_slow" is set
1100 enum { slow_result_path = 1, fast_result_path = 2 };
1101 Node *result_region;
1102 Node *result_phi_rawmem;
1103 Node *result_phi_rawoop;
1104 Node *result_phi_i_o;
1105
1106 // The initial slow comparison is a size check, the comparison
1107 // we want to do is a BoolTest::gt
1108 bool always_slow = false;
1109 int tv = _igvn.find_int_con(initial_slow_test, -1);
1110 if (tv >= 0) {
1111 always_slow = (tv == 1);
1112 initial_slow_test = NULL;
1113 } else {
1114 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1115 }
1116
1117 if (C->env()->dtrace_alloc_probes() ||
1118 !UseTLAB && (!Universe::heap()->supports_inline_contig_alloc() ||
1119 (UseConcMarkSweepGC && CMSIncrementalMode))) {
1120 // Force slow-path allocation
1121 always_slow = true;
1122 initial_slow_test = NULL;
1123 }
1124
1125
1126 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1127 Node *slow_region = NULL;
1128 Node *toobig_false = ctrl;
1129
1130 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent");
1131 // generate the initial test if necessary
1132 if (initial_slow_test != NULL ) {
1133 slow_region = new (C) RegionNode(3);
1134
1135 // Now make the initial failure test. Usually a too-big test but
1136 // might be a TRUE for finalizers or a fancy class check for
1137 // newInstance0.
1138 IfNode *toobig_iff = new (C) IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1139 transform_later(toobig_iff);
1140 // Plug the failing-too-big test into the slow-path region
1141 Node *toobig_true = new (C) IfTrueNode( toobig_iff );
1142 transform_later(toobig_true);
1143 slow_region ->init_req( too_big_or_final_path, toobig_true );
1144 toobig_false = new (C) IfFalseNode( toobig_iff );
1145 transform_later(toobig_false);
1146 } else { // No initial test, just fall into next case
1147 toobig_false = ctrl;
1148 debug_only(slow_region = NodeSentinel);
1149 }
1150
1151 Node *slow_mem = mem; // save the current memory state for slow path
1152 // generate the fast allocation code unless we know that the initial test will always go slow
1153 if (!always_slow) {
1154 // Fast path modifies only raw memory.
1155 if (mem->is_MergeMem()) {
1156 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1157 }
1158
1159 Node* eden_top_adr;
1160 Node* eden_end_adr;
1161
1162 set_eden_pointers(eden_top_adr, eden_end_adr);
1163
1164 // Load Eden::end. Loop invariant and hoisted.
1165 //
1166 // Note: We set the control input on "eden_end" and "old_eden_top" when using
1167 // a TLAB to work around a bug where these values were being moved across
1168 // a safepoint. These are not oops, so they cannot be include in the oop
1169 // map, but they can be changed by a GC. The proper way to fix this would
1170 // be to set the raw memory state when generating a SafepointNode. However
1171 // this will require extensive changes to the loop optimization in order to
1172 // prevent a degradation of the optimization.
1173 // See comment in memnode.hpp, around line 227 in class LoadPNode.
1174 Node *eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS);
1175
1176 // allocate the Region and Phi nodes for the result
1177 result_region = new (C) RegionNode(3);
1178 result_phi_rawmem = new (C) PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1179 result_phi_rawoop = new (C) PhiNode(result_region, TypeRawPtr::BOTTOM);
1180 result_phi_i_o = new (C) PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1181
1182 // We need a Region for the loop-back contended case.
1183 enum { fall_in_path = 1, contended_loopback_path = 2 };
1184 Node *contended_region;
1185 Node *contended_phi_rawmem;
1186 if (UseTLAB) {
1187 contended_region = toobig_false;
1188 contended_phi_rawmem = mem;
1189 } else {
1190 contended_region = new (C) RegionNode(3);
1191 contended_phi_rawmem = new (C) PhiNode(contended_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1192 // Now handle the passing-too-big test. We fall into the contended
1193 // loop-back merge point.
1194 contended_region ->init_req(fall_in_path, toobig_false);
1195 contended_phi_rawmem->init_req(fall_in_path, mem);
1196 transform_later(contended_region);
1197 transform_later(contended_phi_rawmem);
1198 }
1199
1200 // Load(-locked) the heap top.
1201 // See note above concerning the control input when using a TLAB
1202 Node *old_eden_top = UseTLAB
1203 ? new (C) LoadPNode (ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM)
1204 : new (C) LoadPLockedNode(contended_region, contended_phi_rawmem, eden_top_adr);
1205
1206 transform_later(old_eden_top);
1207 // Add to heap top to get a new heap top
1208 Node *new_eden_top = new (C) AddPNode(top(), old_eden_top, size_in_bytes);
1209 transform_later(new_eden_top);
1210 // Check for needing a GC; compare against heap end
1211 Node *needgc_cmp = new (C) CmpPNode(new_eden_top, eden_end);
1212 transform_later(needgc_cmp);
1213 Node *needgc_bol = new (C) BoolNode(needgc_cmp, BoolTest::ge);
1214 transform_later(needgc_bol);
1215 IfNode *needgc_iff = new (C) IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN);
1216 transform_later(needgc_iff);
1217
1218 // Plug the failing-heap-space-need-gc test into the slow-path region
1219 Node *needgc_true = new (C) IfTrueNode(needgc_iff);
1220 transform_later(needgc_true);
1221 if (initial_slow_test) {
1222 slow_region->init_req(need_gc_path, needgc_true);
1223 // This completes all paths into the slow merge point
1224 transform_later(slow_region);
1225 } else { // No initial slow path needed!
1226 // Just fall from the need-GC path straight into the VM call.
1227 slow_region = needgc_true;
1228 }
1229 // No need for a GC. Setup for the Store-Conditional
1230 Node *needgc_false = new (C) IfFalseNode(needgc_iff);
1231 transform_later(needgc_false);
1232
1233 // Grab regular I/O before optional prefetch may change it.
1234 // Slow-path does no I/O so just set it to the original I/O.
1235 result_phi_i_o->init_req(slow_result_path, i_o);
1236
1237 i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem,
1238 old_eden_top, new_eden_top, length);
1239
1240 // Name successful fast-path variables
1241 Node* fast_oop = old_eden_top;
1242 Node* fast_oop_ctrl;
1243 Node* fast_oop_rawmem;
1244
1245 // Store (-conditional) the modified eden top back down.
1246 // StorePConditional produces flags for a test PLUS a modified raw
1247 // memory state.
1248 if (UseTLAB) {
1249 Node* store_eden_top =
1250 new (C) StorePNode(needgc_false, contended_phi_rawmem, eden_top_adr,
1251 TypeRawPtr::BOTTOM, new_eden_top);
1252 transform_later(store_eden_top);
1253 fast_oop_ctrl = needgc_false; // No contention, so this is the fast path
1254 fast_oop_rawmem = store_eden_top;
1255 } else {
1256 Node* store_eden_top =
1257 new (C) StorePConditionalNode(needgc_false, contended_phi_rawmem, eden_top_adr,
1258 new_eden_top, fast_oop/*old_eden_top*/);
1259 transform_later(store_eden_top);
1260 Node *contention_check = new (C) BoolNode(store_eden_top, BoolTest::ne);
1261 transform_later(contention_check);
1262 store_eden_top = new (C) SCMemProjNode(store_eden_top);
1263 transform_later(store_eden_top);
1264
1265 // If not using TLABs, check to see if there was contention.
1266 IfNode *contention_iff = new (C) IfNode (needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN);
1267 transform_later(contention_iff);
1268 Node *contention_true = new (C) IfTrueNode(contention_iff);
1269 transform_later(contention_true);
1270 // If contention, loopback and try again.
1271 contended_region->init_req(contended_loopback_path, contention_true);
1272 contended_phi_rawmem->init_req(contended_loopback_path, store_eden_top);
1273
1274 // Fast-path succeeded with no contention!
1275 Node *contention_false = new (C) IfFalseNode(contention_iff);
1276 transform_later(contention_false);
1277 fast_oop_ctrl = contention_false;
1278
1279 // Bump total allocated bytes for this thread
1280 Node* thread = new (C) ThreadLocalNode();
1281 transform_later(thread);
1282 Node* alloc_bytes_adr = basic_plus_adr(top()/*not oop*/, thread,
1283 in_bytes(JavaThread::allocated_bytes_offset()));
1284 Node* alloc_bytes = make_load(fast_oop_ctrl, store_eden_top, alloc_bytes_adr,
1285 0, TypeLong::LONG, T_LONG);
1286 #ifdef _LP64
1287 Node* alloc_size = size_in_bytes;
1288 #else
1289 Node* alloc_size = new (C) ConvI2LNode(size_in_bytes);
1290 transform_later(alloc_size);
1291 #endif
1292 Node* new_alloc_bytes = new (C) AddLNode(alloc_bytes, alloc_size);
1293 transform_later(new_alloc_bytes);
1294 fast_oop_rawmem = make_store(fast_oop_ctrl, store_eden_top, alloc_bytes_adr,
1295 0, new_alloc_bytes, T_LONG);
1296 }
1297
1298 InitializeNode* init = alloc->initialization();
1299 fast_oop_rawmem = initialize_object(alloc,
1300 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1301 klass_node, length, size_in_bytes);
1302
1303 // If initialization is performed by an array copy, any required
1304 // MemBarStoreStore was already added. If the object does not
1305 // escape no need for a MemBarStoreStore. Otherwise we need a
1306 // MemBarStoreStore so that stores that initialize this object
1307 // can't be reordered with a subsequent store that makes this
1308 // object accessible by other threads.
1309 if (init == NULL || (!init->is_complete_with_arraycopy() && !init->does_not_escape())) {
1310 if (init == NULL || init->req() < InitializeNode::RawStores) {
1311 // No InitializeNode or no stores captured by zeroing
1312 // elimination. Simply add the MemBarStoreStore after object
1313 // initialization.
1314 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1315 transform_later(mb);
1316
1317 mb->init_req(TypeFunc::Memory, fast_oop_rawmem);
1318 mb->init_req(TypeFunc::Control, fast_oop_ctrl);
1319 fast_oop_ctrl = new (C) ProjNode(mb,TypeFunc::Control);
1320 transform_later(fast_oop_ctrl);
1321 fast_oop_rawmem = new (C) ProjNode(mb,TypeFunc::Memory);
1322 transform_later(fast_oop_rawmem);
1323 } else {
1324 // Add the MemBarStoreStore after the InitializeNode so that
1325 // all stores performing the initialization that were moved
1326 // before the InitializeNode happen before the storestore
1327 // barrier.
1328
1329 Node* init_ctrl = init->proj_out(TypeFunc::Control);
1330 Node* init_mem = init->proj_out(TypeFunc::Memory);
1331
1332 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1333 transform_later(mb);
1334
1335 Node* ctrl = new (C) ProjNode(init,TypeFunc::Control);
1336 transform_later(ctrl);
1337 Node* mem = new (C) ProjNode(init,TypeFunc::Memory);
1338 transform_later(mem);
1339
1340 // The MemBarStoreStore depends on control and memory coming
1341 // from the InitializeNode
1342 mb->init_req(TypeFunc::Memory, mem);
1343 mb->init_req(TypeFunc::Control, ctrl);
1344
1345 ctrl = new (C) ProjNode(mb,TypeFunc::Control);
1346 transform_later(ctrl);
1347 mem = new (C) ProjNode(mb,TypeFunc::Memory);
1348 transform_later(mem);
1349
1350 // All nodes that depended on the InitializeNode for control
1351 // and memory must now depend on the MemBarNode that itself
1352 // depends on the InitializeNode
1353 _igvn.replace_node(init_ctrl, ctrl);
1354 _igvn.replace_node(init_mem, mem);
1355 }
1356 }
1357
1358 if (C->env()->dtrace_extended_probes()) {
1359 // Slow-path call
1360 int size = TypeFunc::Parms + 2;
1361 CallLeafNode *call = new (C) CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1362 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base),
1363 "dtrace_object_alloc",
1364 TypeRawPtr::BOTTOM);
1365
1366 // Get base of thread-local storage area
1367 Node* thread = new (C) ThreadLocalNode();
1368 transform_later(thread);
1369
1370 call->init_req(TypeFunc::Parms+0, thread);
1371 call->init_req(TypeFunc::Parms+1, fast_oop);
1372 call->init_req(TypeFunc::Control, fast_oop_ctrl);
1373 call->init_req(TypeFunc::I_O , top()); // does no i/o
1374 call->init_req(TypeFunc::Memory , fast_oop_rawmem);
1375 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1376 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1377 transform_later(call);
1378 fast_oop_ctrl = new (C) ProjNode(call,TypeFunc::Control);
1379 transform_later(fast_oop_ctrl);
1380 fast_oop_rawmem = new (C) ProjNode(call,TypeFunc::Memory);
1381 transform_later(fast_oop_rawmem);
1382 }
1383
1384 // Plug in the successful fast-path into the result merge point
1385 result_region ->init_req(fast_result_path, fast_oop_ctrl);
1386 result_phi_rawoop->init_req(fast_result_path, fast_oop);
1387 result_phi_i_o ->init_req(fast_result_path, i_o);
1388 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1389 } else {
1390 slow_region = ctrl;
1391 result_phi_i_o = i_o; // Rename it to use in the following code.
1392 }
1393
1394 // Generate slow-path call
1395 CallNode *call = new (C) CallStaticJavaNode(slow_call_type, slow_call_address,
1396 OptoRuntime::stub_name(slow_call_address),
1397 alloc->jvms()->bci(),
1398 TypePtr::BOTTOM);
1399 call->init_req( TypeFunc::Control, slow_region );
1400 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o
1401 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs
1402 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) );
1403 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) );
1404
1405 call->init_req(TypeFunc::Parms+0, klass_node);
1406 if (length != NULL) {
1407 call->init_req(TypeFunc::Parms+1, length);
1408 }
1409
1410 // Copy debug information and adjust JVMState information, then replace
1411 // allocate node with the call
1412 copy_call_debug_info((CallNode *) alloc, call);
1413 if (!always_slow) {
1414 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1415 } else {
1416 // Hook i_o projection to avoid its elimination during allocation
1417 // replacement (when only a slow call is generated).
1418 call->set_req(TypeFunc::I_O, result_phi_i_o);
1419 }
1420 _igvn.replace_node(alloc, call);
1421 transform_later(call);
1422
1423 // Identify the output projections from the allocate node and
1424 // adjust any references to them.
1425 // The control and io projections look like:
1426 //
1427 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1428 // Allocate Catch
1429 // ^---Proj(io) <-------+ ^---CatchProj(io)
1430 //
1431 // We are interested in the CatchProj nodes.
1432 //
1433 extract_call_projections(call);
1434
1435 // An allocate node has separate memory projections for the uses on
1436 // the control and i_o paths. Replace the control memory projection with
1437 // result_phi_rawmem (unless we are only generating a slow call when
1438 // both memory projections are combined)
1439 if (!always_slow && _memproj_fallthrough != NULL) {
1440 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) {
1441 Node *use = _memproj_fallthrough->fast_out(i);
1442 _igvn.rehash_node_delayed(use);
1443 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem);
1444 // back up iterator
1445 --i;
1446 }
1447 }
1448 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete
1449 // _memproj_catchall so we end up with a call that has only 1 memory projection.
1450 if (_memproj_catchall != NULL ) {
1451 if (_memproj_fallthrough == NULL) {
1452 _memproj_fallthrough = new (C) ProjNode(call, TypeFunc::Memory);
1453 transform_later(_memproj_fallthrough);
1454 }
1455 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) {
1456 Node *use = _memproj_catchall->fast_out(i);
1457 _igvn.rehash_node_delayed(use);
1458 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough);
1459 // back up iterator
1460 --i;
1461 }
1462 assert(_memproj_catchall->outcnt() == 0, "all uses must be deleted");
1463 _igvn.remove_dead_node(_memproj_catchall);
1464 }
1465
1466 // An allocate node has separate i_o projections for the uses on the control
1467 // and i_o paths. Always replace the control i_o projection with result i_o
1468 // otherwise incoming i_o become dead when only a slow call is generated
1469 // (it is different from memory projections where both projections are
1470 // combined in such case).
1471 if (_ioproj_fallthrough != NULL) {
1472 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) {
1473 Node *use = _ioproj_fallthrough->fast_out(i);
1474 _igvn.rehash_node_delayed(use);
1475 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o);
1476 // back up iterator
1477 --i;
1478 }
1479 }
1480 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete
1481 // _ioproj_catchall so we end up with a call that has only 1 i_o projection.
1482 if (_ioproj_catchall != NULL ) {
1483 if (_ioproj_fallthrough == NULL) {
1484 _ioproj_fallthrough = new (C) ProjNode(call, TypeFunc::I_O);
1485 transform_later(_ioproj_fallthrough);
1486 }
1487 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) {
1488 Node *use = _ioproj_catchall->fast_out(i);
1489 _igvn.rehash_node_delayed(use);
1490 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough);
1491 // back up iterator
1492 --i;
1493 }
1494 assert(_ioproj_catchall->outcnt() == 0, "all uses must be deleted");
1495 _igvn.remove_dead_node(_ioproj_catchall);
1496 }
1497
1498 // if we generated only a slow call, we are done
1499 if (always_slow) {
1500 // Now we can unhook i_o.
1501 if (result_phi_i_o->outcnt() > 1) {
1502 call->set_req(TypeFunc::I_O, top());
1503 } else {
1504 assert(result_phi_i_o->unique_ctrl_out() == call, "");
1505 // Case of new array with negative size known during compilation.
1506 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1507 // following code since call to runtime will throw exception.
1508 // As result there will be no users of i_o after the call.
1509 // Leave i_o attached to this call to avoid problems in preceding graph.
1510 }
1511 return;
1512 }
1513
1514
1515 if (_fallthroughcatchproj != NULL) {
1516 ctrl = _fallthroughcatchproj->clone();
1517 transform_later(ctrl);
1518 _igvn.replace_node(_fallthroughcatchproj, result_region);
1519 } else {
1520 ctrl = top();
1521 }
1522 Node *slow_result;
1523 if (_resproj == NULL) {
1524 // no uses of the allocation result
1525 slow_result = top();
1526 } else {
1527 slow_result = _resproj->clone();
1528 transform_later(slow_result);
1529 _igvn.replace_node(_resproj, result_phi_rawoop);
1530 }
1531
1532 // Plug slow-path into result merge point
1533 result_region ->init_req( slow_result_path, ctrl );
1534 result_phi_rawoop->init_req( slow_result_path, slow_result);
1535 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough );
1536 transform_later(result_region);
1537 transform_later(result_phi_rawoop);
1538 transform_later(result_phi_rawmem);
1539 transform_later(result_phi_i_o);
1540 // This completes all paths into the result merge point
1541 }
1542
1543
1544 // Helper for PhaseMacroExpand::expand_allocate_common.
1545 // Initializes the newly-allocated storage.
1546 Node*
1547 PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1548 Node* control, Node* rawmem, Node* object,
1549 Node* klass_node, Node* length,
1550 Node* size_in_bytes) {
1551 InitializeNode* init = alloc->initialization();
1552 // Store the klass & mark bits
1553 Node* mark_node = NULL;
1554 // For now only enable fast locking for non-array types
1555 if (UseBiasedLocking && (length == NULL)) {
1556 mark_node = make_load(control, rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
1557 } else {
1558 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype()));
1559 }
1560 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
1561
1562 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_OBJECT);
1563 int header_size = alloc->minimum_header_size(); // conservatively small
1564
1565 // Array length
1566 if (length != NULL) { // Arrays need length field
1567 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1568 // conservatively small header size:
1569 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1570 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1571 if (k->is_array_klass()) // we know the exact header size in most cases:
1572 header_size = Klass::layout_helper_header_size(k->layout_helper());
1573 }
1574
1575 // Clear the object body, if necessary.
1576 if (init == NULL) {
1577 // The init has somehow disappeared; be cautious and clear everything.
1578 //
1579 // This can happen if a node is allocated but an uncommon trap occurs
1580 // immediately. In this case, the Initialize gets associated with the
1581 // trap, and may be placed in a different (outer) loop, if the Allocate
1582 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1583 // there can be two Allocates to one Initialize. The answer in all these
1584 // edge cases is safety first. It is always safe to clear immediately
1585 // within an Allocate, and then (maybe or maybe not) clear some more later.
1586 if (!ZeroTLAB)
1587 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1588 header_size, size_in_bytes,
1589 &_igvn);
1590 } else {
1591 if (!init->is_complete()) {
1592 // Try to win by zeroing only what the init does not store.
1593 // We can also try to do some peephole optimizations,
1594 // such as combining some adjacent subword stores.
1595 rawmem = init->complete_stores(control, rawmem, object,
1596 header_size, size_in_bytes, &_igvn);
1597 }
1598 // We have no more use for this link, since the AllocateNode goes away:
1599 init->set_req(InitializeNode::RawAddress, top());
1600 // (If we keep the link, it just confuses the register allocator,
1601 // who thinks he sees a real use of the address by the membar.)
1602 }
1603
1604 return rawmem;
1605 }
1606
1607 // Generate prefetch instructions for next allocations.
1608 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
1609 Node*& contended_phi_rawmem,
1610 Node* old_eden_top, Node* new_eden_top,
1611 Node* length) {
1612 enum { fall_in_path = 1, pf_path = 2 };
1613 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
1614 // Generate prefetch allocation with watermark check.
1615 // As an allocation hits the watermark, we will prefetch starting
1616 // at a "distance" away from watermark.
1617
1618 Node *pf_region = new (C) RegionNode(3);
1619 Node *pf_phi_rawmem = new (C) PhiNode( pf_region, Type::MEMORY,
1620 TypeRawPtr::BOTTOM );
1621 // I/O is used for Prefetch
1622 Node *pf_phi_abio = new (C) PhiNode( pf_region, Type::ABIO );
1623
1624 Node *thread = new (C) ThreadLocalNode();
1625 transform_later(thread);
1626
1627 Node *eden_pf_adr = new (C) AddPNode( top()/*not oop*/, thread,
1628 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
1629 transform_later(eden_pf_adr);
1630
1631 Node *old_pf_wm = new (C) LoadPNode( needgc_false,
1632 contended_phi_rawmem, eden_pf_adr,
1633 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM );
1634 transform_later(old_pf_wm);
1635
1636 // check against new_eden_top
1637 Node *need_pf_cmp = new (C) CmpPNode( new_eden_top, old_pf_wm );
1638 transform_later(need_pf_cmp);
1639 Node *need_pf_bol = new (C) BoolNode( need_pf_cmp, BoolTest::ge );
1640 transform_later(need_pf_bol);
1641 IfNode *need_pf_iff = new (C) IfNode( needgc_false, need_pf_bol,
1642 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1643 transform_later(need_pf_iff);
1644
1645 // true node, add prefetchdistance
1646 Node *need_pf_true = new (C) IfTrueNode( need_pf_iff );
1647 transform_later(need_pf_true);
1648
1649 Node *need_pf_false = new (C) IfFalseNode( need_pf_iff );
1650 transform_later(need_pf_false);
1651
1652 Node *new_pf_wmt = new (C) AddPNode( top(), old_pf_wm,
1653 _igvn.MakeConX(AllocatePrefetchDistance) );
1654 transform_later(new_pf_wmt );
1655 new_pf_wmt->set_req(0, need_pf_true);
1656
1657 Node *store_new_wmt = new (C) StorePNode( need_pf_true,
1658 contended_phi_rawmem, eden_pf_adr,
1659 TypeRawPtr::BOTTOM, new_pf_wmt );
1660 transform_later(store_new_wmt);
1661
1662 // adding prefetches
1663 pf_phi_abio->init_req( fall_in_path, i_o );
1664
1665 Node *prefetch_adr;
1666 Node *prefetch;
1667 uint lines = AllocatePrefetchDistance / AllocatePrefetchStepSize;
1668 uint step_size = AllocatePrefetchStepSize;
1669 uint distance = 0;
1670
1671 for ( uint i = 0; i < lines; i++ ) {
1672 prefetch_adr = new (C) AddPNode( old_pf_wm, new_pf_wmt,
1673 _igvn.MakeConX(distance) );
1674 transform_later(prefetch_adr);
1675 prefetch = new (C) PrefetchAllocationNode( i_o, prefetch_adr );
1676 transform_later(prefetch);
1677 distance += step_size;
1678 i_o = prefetch;
1679 }
1680 pf_phi_abio->set_req( pf_path, i_o );
1681
1682 pf_region->init_req( fall_in_path, need_pf_false );
1683 pf_region->init_req( pf_path, need_pf_true );
1684
1685 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
1686 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
1687
1688 transform_later(pf_region);
1689 transform_later(pf_phi_rawmem);
1690 transform_later(pf_phi_abio);
1691
1692 needgc_false = pf_region;
1693 contended_phi_rawmem = pf_phi_rawmem;
1694 i_o = pf_phi_abio;
1695 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) {
1696 // Insert a prefetch for each allocation.
1697 // This code is used for Sparc with BIS.
1698 Node *pf_region = new (C) RegionNode(3);
1699 Node *pf_phi_rawmem = new (C) PhiNode( pf_region, Type::MEMORY,
1700 TypeRawPtr::BOTTOM );
1701
1702 // Generate several prefetch instructions.
1703 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1704 uint step_size = AllocatePrefetchStepSize;
1705 uint distance = AllocatePrefetchDistance;
1706
1707 // Next cache address.
1708 Node *cache_adr = new (C) AddPNode(old_eden_top, old_eden_top,
1709 _igvn.MakeConX(distance));
1710 transform_later(cache_adr);
1711 cache_adr = new (C) CastP2XNode(needgc_false, cache_adr);
1712 transform_later(cache_adr);
1713 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1));
1714 cache_adr = new (C) AndXNode(cache_adr, mask);
1715 transform_later(cache_adr);
1716 cache_adr = new (C) CastX2PNode(cache_adr);
1717 transform_later(cache_adr);
1718
1719 // Prefetch
1720 Node *prefetch = new (C) PrefetchAllocationNode( contended_phi_rawmem, cache_adr );
1721 prefetch->set_req(0, needgc_false);
1722 transform_later(prefetch);
1723 contended_phi_rawmem = prefetch;
1724 Node *prefetch_adr;
1725 distance = step_size;
1726 for ( uint i = 1; i < lines; i++ ) {
1727 prefetch_adr = new (C) AddPNode( cache_adr, cache_adr,
1728 _igvn.MakeConX(distance) );
1729 transform_later(prefetch_adr);
1730 prefetch = new (C) PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr );
1731 transform_later(prefetch);
1732 distance += step_size;
1733 contended_phi_rawmem = prefetch;
1734 }
1735 } else if( AllocatePrefetchStyle > 0 ) {
1736 // Insert a prefetch for each allocation only on the fast-path
1737 Node *prefetch_adr;
1738 Node *prefetch;
1739 // Generate several prefetch instructions.
1740 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1741 uint step_size = AllocatePrefetchStepSize;
1742 uint distance = AllocatePrefetchDistance;
1743 for ( uint i = 0; i < lines; i++ ) {
1744 prefetch_adr = new (C) AddPNode( old_eden_top, new_eden_top,
1745 _igvn.MakeConX(distance) );
1746 transform_later(prefetch_adr);
1747 prefetch = new (C) PrefetchAllocationNode( i_o, prefetch_adr );
1748 // Do not let it float too high, since if eden_top == eden_end,
1749 // both might be null.
1750 if( i == 0 ) { // Set control for first prefetch, next follows it
1751 prefetch->init_req(0, needgc_false);
1752 }
1753 transform_later(prefetch);
1754 distance += step_size;
1755 i_o = prefetch;
1756 }
1757 }
1758 return i_o;
1759 }
1760
1761
1762 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
1763 expand_allocate_common(alloc, NULL,
1764 OptoRuntime::new_instance_Type(),
1765 OptoRuntime::new_instance_Java());
1766 }
1767
1768 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
1769 Node* length = alloc->in(AllocateNode::ALength);
1770 InitializeNode* init = alloc->initialization();
1771 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1772 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1773 address slow_call_address; // Address of slow call
1774 if (init != NULL && init->is_complete_with_arraycopy() &&
1775 k->is_type_array_klass()) {
1776 // Don't zero type array during slow allocation in VM since
1777 // it will be initialized later by arraycopy in compiled code.
1778 slow_call_address = OptoRuntime::new_array_nozero_Java();
1779 } else {
1780 slow_call_address = OptoRuntime::new_array_Java();
1781 }
1782 expand_allocate_common(alloc, length,
1783 OptoRuntime::new_array_Type(),
1784 slow_call_address);
1785 }
1786
1787 //-------------------mark_eliminated_box----------------------------------
1788 //
1789 // During EA obj may point to several objects but after few ideal graph
1790 // transformations (CCP) it may point to only one non escaping object
1791 // (but still using phi), corresponding locks and unlocks will be marked
1792 // for elimination. Later obj could be replaced with a new node (new phi)
1793 // and which does not have escape information. And later after some graph
1794 // reshape other locks and unlocks (which were not marked for elimination
1795 // before) are connected to this new obj (phi) but they still will not be
1796 // marked for elimination since new obj has no escape information.
1797 // Mark all associated (same box and obj) lock and unlock nodes for
1798 // elimination if some of them marked already.
1799 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) {
1800 if (oldbox->as_BoxLock()->is_eliminated())
1801 return; // This BoxLock node was processed already.
1802
1803 // New implementation (EliminateNestedLocks) has separate BoxLock
1804 // node for each locked region so mark all associated locks/unlocks as
1805 // eliminated even if different objects are referenced in one locked region
1806 // (for example, OSR compilation of nested loop inside locked scope).
1807 if (EliminateNestedLocks ||
1808 oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj)) {
1809 // Box is used only in one lock region. Mark this box as eliminated.
1810 _igvn.hash_delete(oldbox);
1811 oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value
1812 _igvn.hash_insert(oldbox);
1813
1814 for (uint i = 0; i < oldbox->outcnt(); i++) {
1815 Node* u = oldbox->raw_out(i);
1816 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) {
1817 AbstractLockNode* alock = u->as_AbstractLock();
1818 // Check lock's box since box could be referenced by Lock's debug info.
1819 if (alock->box_node() == oldbox) {
1820 // Mark eliminated all related locks and unlocks.
1821 alock->set_non_esc_obj();
1822 }
1823 }
1824 }
1825 return;
1826 }
1827
1828 // Create new "eliminated" BoxLock node and use it in monitor debug info
1829 // instead of oldbox for the same object.
1830 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
1831
1832 // Note: BoxLock node is marked eliminated only here and it is used
1833 // to indicate that all associated lock and unlock nodes are marked
1834 // for elimination.
1835 newbox->set_eliminated();
1836 transform_later(newbox);
1837
1838 // Replace old box node with new box for all users of the same object.
1839 for (uint i = 0; i < oldbox->outcnt();) {
1840 bool next_edge = true;
1841
1842 Node* u = oldbox->raw_out(i);
1843 if (u->is_AbstractLock()) {
1844 AbstractLockNode* alock = u->as_AbstractLock();
1845 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) {
1846 // Replace Box and mark eliminated all related locks and unlocks.
1847 alock->set_non_esc_obj();
1848 _igvn.rehash_node_delayed(alock);
1849 alock->set_box_node(newbox);
1850 next_edge = false;
1851 }
1852 }
1853 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) {
1854 FastLockNode* flock = u->as_FastLock();
1855 assert(flock->box_node() == oldbox, "sanity");
1856 _igvn.rehash_node_delayed(flock);
1857 flock->set_box_node(newbox);
1858 next_edge = false;
1859 }
1860
1861 // Replace old box in monitor debug info.
1862 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
1863 SafePointNode* sfn = u->as_SafePoint();
1864 JVMState* youngest_jvms = sfn->jvms();
1865 int max_depth = youngest_jvms->depth();
1866 for (int depth = 1; depth <= max_depth; depth++) {
1867 JVMState* jvms = youngest_jvms->of_depth(depth);
1868 int num_mon = jvms->nof_monitors();
1869 // Loop over monitors
1870 for (int idx = 0; idx < num_mon; idx++) {
1871 Node* obj_node = sfn->monitor_obj(jvms, idx);
1872 Node* box_node = sfn->monitor_box(jvms, idx);
1873 if (box_node == oldbox && obj_node->eqv_uncast(obj)) {
1874 int j = jvms->monitor_box_offset(idx);
1875 _igvn.replace_input_of(u, j, newbox);
1876 next_edge = false;
1877 }
1878 }
1879 }
1880 }
1881 if (next_edge) i++;
1882 }
1883 }
1884
1885 //-----------------------mark_eliminated_locking_nodes-----------------------
1886 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) {
1887 if (EliminateNestedLocks) {
1888 if (alock->is_nested()) {
1889 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity");
1890 return;
1891 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened
1892 // Only Lock node has JVMState needed here.
1893 if (alock->jvms() != NULL && alock->as_Lock()->is_nested_lock_region()) {
1894 // Mark eliminated related nested locks and unlocks.
1895 Node* obj = alock->obj_node();
1896 BoxLockNode* box_node = alock->box_node()->as_BoxLock();
1897 assert(!box_node->is_eliminated(), "should not be marked yet");
1898 // Note: BoxLock node is marked eliminated only here
1899 // and it is used to indicate that all associated lock
1900 // and unlock nodes are marked for elimination.
1901 box_node->set_eliminated(); // Box's hash is always NO_HASH here
1902 for (uint i = 0; i < box_node->outcnt(); i++) {
1903 Node* u = box_node->raw_out(i);
1904 if (u->is_AbstractLock()) {
1905 alock = u->as_AbstractLock();
1906 if (alock->box_node() == box_node) {
1907 // Verify that this Box is referenced only by related locks.
1908 assert(alock->obj_node()->eqv_uncast(obj), "");
1909 // Mark all related locks and unlocks.
1910 alock->set_nested();
1911 }
1912 }
1913 }
1914 }
1915 return;
1916 }
1917 // Process locks for non escaping object
1918 assert(alock->is_non_esc_obj(), "");
1919 } // EliminateNestedLocks
1920
1921 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
1922 // Look for all locks of this object and mark them and
1923 // corresponding BoxLock nodes as eliminated.
1924 Node* obj = alock->obj_node();
1925 for (uint j = 0; j < obj->outcnt(); j++) {
1926 Node* o = obj->raw_out(j);
1927 if (o->is_AbstractLock() &&
1928 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
1929 alock = o->as_AbstractLock();
1930 Node* box = alock->box_node();
1931 // Replace old box node with new eliminated box for all users
1932 // of the same object and mark related locks as eliminated.
1933 mark_eliminated_box(box, obj);
1934 }
1935 }
1936 }
1937 }
1938
1939 // we have determined that this lock/unlock can be eliminated, we simply
1940 // eliminate the node without expanding it.
1941 //
1942 // Note: The membar's associated with the lock/unlock are currently not
1943 // eliminated. This should be investigated as a future enhancement.
1944 //
1945 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
1946
1947 if (!alock->is_eliminated()) {
1948 return false;
1949 }
1950 #ifdef ASSERT
1951 if (!alock->is_coarsened()) {
1952 // Check that new "eliminated" BoxLock node is created.
1953 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
1954 assert(oldbox->is_eliminated(), "should be done already");
1955 }
1956 #endif
1957 CompileLog* log = C->log();
1958 if (log != NULL) {
1959 log->head("eliminate_lock lock='%d'",
1960 alock->is_Lock());
1961 JVMState* p = alock->jvms();
1962 while (p != NULL) {
1963 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1964 p = p->caller();
1965 }
1966 log->tail("eliminate_lock");
1967 }
1968
1969 #ifndef PRODUCT
1970 if (PrintEliminateLocks) {
1971 if (alock->is_Lock()) {
1972 tty->print_cr("++++ Eliminated: %d Lock", alock->_idx);
1973 } else {
1974 tty->print_cr("++++ Eliminated: %d Unlock", alock->_idx);
1975 }
1976 }
1977 #endif
1978
1979 Node* mem = alock->in(TypeFunc::Memory);
1980 Node* ctrl = alock->in(TypeFunc::Control);
1981
1982 extract_call_projections(alock);
1983 // There are 2 projections from the lock. The lock node will
1984 // be deleted when its last use is subsumed below.
1985 assert(alock->outcnt() == 2 &&
1986 _fallthroughproj != NULL &&
1987 _memproj_fallthrough != NULL,
1988 "Unexpected projections from Lock/Unlock");
1989
1990 Node* fallthroughproj = _fallthroughproj;
1991 Node* memproj_fallthrough = _memproj_fallthrough;
1992
1993 // The memory projection from a lock/unlock is RawMem
1994 // The input to a Lock is merged memory, so extract its RawMem input
1995 // (unless the MergeMem has been optimized away.)
1996 if (alock->is_Lock()) {
1997 // Seach for MemBarAcquireLock node and delete it also.
1998 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
1999 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, "");
2000 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2001 Node* memproj = membar->proj_out(TypeFunc::Memory);
2002 _igvn.replace_node(ctrlproj, fallthroughproj);
2003 _igvn.replace_node(memproj, memproj_fallthrough);
2004
2005 // Delete FastLock node also if this Lock node is unique user
2006 // (a loop peeling may clone a Lock node).
2007 Node* flock = alock->as_Lock()->fastlock_node();
2008 if (flock->outcnt() == 1) {
2009 assert(flock->unique_out() == alock, "sanity");
2010 _igvn.replace_node(flock, top());
2011 }
2012 }
2013
2014 // Seach for MemBarReleaseLock node and delete it also.
2015 if (alock->is_Unlock() && ctrl != NULL && ctrl->is_Proj() &&
2016 ctrl->in(0)->is_MemBar()) {
2017 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2018 assert(membar->Opcode() == Op_MemBarReleaseLock &&
2019 mem->is_Proj() && membar == mem->in(0), "");
2020 _igvn.replace_node(fallthroughproj, ctrl);
2021 _igvn.replace_node(memproj_fallthrough, mem);
2022 fallthroughproj = ctrl;
2023 memproj_fallthrough = mem;
2024 ctrl = membar->in(TypeFunc::Control);
2025 mem = membar->in(TypeFunc::Memory);
2026 }
2027
2028 _igvn.replace_node(fallthroughproj, ctrl);
2029 _igvn.replace_node(memproj_fallthrough, mem);
2030 return true;
2031 }
2032
2033
2034 //------------------------------expand_lock_node----------------------
2035 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
2036
2037 Node* ctrl = lock->in(TypeFunc::Control);
2038 Node* mem = lock->in(TypeFunc::Memory);
2039 Node* obj = lock->obj_node();
2040 Node* box = lock->box_node();
2041 Node* flock = lock->fastlock_node();
2042
2043 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2044
2045 // Make the merge point
2046 Node *region;
2047 Node *mem_phi;
2048 Node *slow_path;
2049
2050 if (UseOptoBiasInlining) {
2051 /*
2052 * See the full description in MacroAssembler::biased_locking_enter().
2053 *
2054 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) {
2055 * // The object is biased.
2056 * proto_node = klass->prototype_header;
2057 * o_node = thread | proto_node;
2058 * x_node = o_node ^ mark_word;
2059 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ?
2060 * // Done.
2061 * } else {
2062 * if( (x_node & biased_lock_mask) != 0 ) {
2063 * // The klass's prototype header is no longer biased.
2064 * cas(&mark_word, mark_word, proto_node)
2065 * goto cas_lock;
2066 * } else {
2067 * // The klass's prototype header is still biased.
2068 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch?
2069 * old = mark_word;
2070 * new = o_node;
2071 * } else {
2072 * // Different thread or anonymous biased.
2073 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask);
2074 * new = thread | old;
2075 * }
2076 * // Try to rebias.
2077 * if( cas(&mark_word, old, new) == 0 ) {
2078 * // Done.
2079 * } else {
2080 * goto slow_path; // Failed.
2081 * }
2082 * }
2083 * }
2084 * } else {
2085 * // The object is not biased.
2086 * cas_lock:
2087 * if( FastLock(obj) == 0 ) {
2088 * // Done.
2089 * } else {
2090 * slow_path:
2091 * OptoRuntime::complete_monitor_locking_Java(obj);
2092 * }
2093 * }
2094 */
2095
2096 region = new (C) RegionNode(5);
2097 // create a Phi for the memory state
2098 mem_phi = new (C) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2099
2100 Node* fast_lock_region = new (C) RegionNode(3);
2101 Node* fast_lock_mem_phi = new (C) PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM);
2102
2103 // First, check mark word for the biased lock pattern.
2104 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2105
2106 // Get fast path - mark word has the biased lock pattern.
2107 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node,
2108 markOopDesc::biased_lock_mask_in_place,
2109 markOopDesc::biased_lock_pattern, true);
2110 // fast_lock_region->in(1) is set to slow path.
2111 fast_lock_mem_phi->init_req(1, mem);
2112
2113 // Now check that the lock is biased to the current thread and has
2114 // the same epoch and bias as Klass::_prototype_header.
2115
2116 // Special-case a fresh allocation to avoid building nodes:
2117 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn);
2118 if (klass_node == NULL) {
2119 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
2120 klass_node = transform_later( LoadKlassNode::make(_igvn, mem, k_adr, _igvn.type(k_adr)->is_ptr()) );
2121 #ifdef _LP64
2122 if (UseCompressedOops && klass_node->is_DecodeN()) {
2123 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity");
2124 klass_node->in(1)->init_req(0, ctrl);
2125 } else
2126 #endif
2127 klass_node->init_req(0, ctrl);
2128 }
2129 Node *proto_node = make_load(ctrl, mem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeX_X, TypeX_X->basic_type());
2130
2131 Node* thread = transform_later(new (C) ThreadLocalNode());
2132 Node* cast_thread = transform_later(new (C) CastP2XNode(ctrl, thread));
2133 Node* o_node = transform_later(new (C) OrXNode(cast_thread, proto_node));
2134 Node* x_node = transform_later(new (C) XorXNode(o_node, mark_node));
2135
2136 // Get slow path - mark word does NOT match the value.
2137 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node,
2138 (~markOopDesc::age_mask_in_place), 0);
2139 // region->in(3) is set to fast path - the object is biased to the current thread.
2140 mem_phi->init_req(3, mem);
2141
2142
2143 // Mark word does NOT match the value (thread | Klass::_prototype_header).
2144
2145
2146 // First, check biased pattern.
2147 // Get fast path - _prototype_header has the same biased lock pattern.
2148 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node,
2149 markOopDesc::biased_lock_mask_in_place, 0, true);
2150
2151 not_biased_ctrl = fast_lock_region->in(2); // Slow path
2152 // fast_lock_region->in(2) - the prototype header is no longer biased
2153 // and we have to revoke the bias on this object.
2154 // We are going to try to reset the mark of this object to the prototype
2155 // value and fall through to the CAS-based locking scheme.
2156 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
2157 Node* cas = new (C) StoreXConditionalNode(not_biased_ctrl, mem, adr,
2158 proto_node, mark_node);
2159 transform_later(cas);
2160 Node* proj = transform_later( new (C) SCMemProjNode(cas));
2161 fast_lock_mem_phi->init_req(2, proj);
2162
2163
2164 // Second, check epoch bits.
2165 Node* rebiased_region = new (C) RegionNode(3);
2166 Node* old_phi = new (C) PhiNode( rebiased_region, TypeX_X);
2167 Node* new_phi = new (C) PhiNode( rebiased_region, TypeX_X);
2168
2169 // Get slow path - mark word does NOT match epoch bits.
2170 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node,
2171 markOopDesc::epoch_mask_in_place, 0);
2172 // The epoch of the current bias is not valid, attempt to rebias the object
2173 // toward the current thread.
2174 rebiased_region->init_req(2, epoch_ctrl);
2175 old_phi->init_req(2, mark_node);
2176 new_phi->init_req(2, o_node);
2177
2178 // rebiased_region->in(1) is set to fast path.
2179 // The epoch of the current bias is still valid but we know
2180 // nothing about the owner; it might be set or it might be clear.
2181 Node* cmask = MakeConX(markOopDesc::biased_lock_mask_in_place |
2182 markOopDesc::age_mask_in_place |
2183 markOopDesc::epoch_mask_in_place);
2184 Node* old = transform_later(new (C) AndXNode(mark_node, cmask));
2185 cast_thread = transform_later(new (C) CastP2XNode(ctrl, thread));
2186 Node* new_mark = transform_later(new (C) OrXNode(cast_thread, old));
2187 old_phi->init_req(1, old);
2188 new_phi->init_req(1, new_mark);
2189
2190 transform_later(rebiased_region);
2191 transform_later(old_phi);
2192 transform_later(new_phi);
2193
2194 // Try to acquire the bias of the object using an atomic operation.
2195 // If this fails we will go in to the runtime to revoke the object's bias.
2196 cas = new (C) StoreXConditionalNode(rebiased_region, mem, adr,
2197 new_phi, old_phi);
2198 transform_later(cas);
2199 proj = transform_later( new (C) SCMemProjNode(cas));
2200
2201 // Get slow path - Failed to CAS.
2202 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0);
2203 mem_phi->init_req(4, proj);
2204 // region->in(4) is set to fast path - the object is rebiased to the current thread.
2205
2206 // Failed to CAS.
2207 slow_path = new (C) RegionNode(3);
2208 Node *slow_mem = new (C) PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM);
2209
2210 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control
2211 slow_mem->init_req(1, proj);
2212
2213 // Call CAS-based locking scheme (FastLock node).
2214
2215 transform_later(fast_lock_region);
2216 transform_later(fast_lock_mem_phi);
2217
2218 // Get slow path - FastLock failed to lock the object.
2219 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0);
2220 mem_phi->init_req(2, fast_lock_mem_phi);
2221 // region->in(2) is set to fast path - the object is locked to the current thread.
2222
2223 slow_path->init_req(2, ctrl); // Capture slow-control
2224 slow_mem->init_req(2, fast_lock_mem_phi);
2225
2226 transform_later(slow_path);
2227 transform_later(slow_mem);
2228 // Reset lock's memory edge.
2229 lock->set_req(TypeFunc::Memory, slow_mem);
2230
2231 } else {
2232 region = new (C) RegionNode(3);
2233 // create a Phi for the memory state
2234 mem_phi = new (C) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2235
2236 // Optimize test; set region slot 2
2237 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2238 mem_phi->init_req(2, mem);
2239 }
2240
2241 // Make slow path call
2242 CallNode *call = make_slow_call( (CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(), OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path, obj, box );
2243
2244 extract_call_projections(call);
2245
2246 // Slow path can only throw asynchronous exceptions, which are always
2247 // de-opted. So the compiler thinks the slow-call can never throw an
2248 // exception. If it DOES throw an exception we would need the debug
2249 // info removed first (since if it throws there is no monitor).
2250 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2251 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2252
2253 // Capture slow path
2254 // disconnect fall-through projection from call and create a new one
2255 // hook up users of fall-through projection to region
2256 Node *slow_ctrl = _fallthroughproj->clone();
2257 transform_later(slow_ctrl);
2258 _igvn.hash_delete(_fallthroughproj);
2259 _fallthroughproj->disconnect_inputs(NULL, C);
2260 region->init_req(1, slow_ctrl);
2261 // region inputs are now complete
2262 transform_later(region);
2263 _igvn.replace_node(_fallthroughproj, region);
2264
2265 Node *memproj = transform_later( new(C) ProjNode(call, TypeFunc::Memory) );
2266 mem_phi->init_req(1, memproj );
2267 transform_later(mem_phi);
2268 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2269 }
2270
2271 //------------------------------expand_unlock_node----------------------
2272 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2273
2274 Node* ctrl = unlock->in(TypeFunc::Control);
2275 Node* mem = unlock->in(TypeFunc::Memory);
2276 Node* obj = unlock->obj_node();
2277 Node* box = unlock->box_node();
2278
2279 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2280
2281 // No need for a null check on unlock
2282
2283 // Make the merge point
2284 Node *region;
2285 Node *mem_phi;
2286
2287 if (UseOptoBiasInlining) {
2288 // Check for biased locking unlock case, which is a no-op.
2289 // See the full description in MacroAssembler::biased_locking_exit().
2290 region = new (C) RegionNode(4);
2291 // create a Phi for the memory state
2292 mem_phi = new (C) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2293 mem_phi->init_req(3, mem);
2294
2295 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2296 ctrl = opt_bits_test(ctrl, region, 3, mark_node,
2297 markOopDesc::biased_lock_mask_in_place,
2298 markOopDesc::biased_lock_pattern);
2299 } else {
2300 region = new (C) RegionNode(3);
2301 // create a Phi for the memory state
2302 mem_phi = new (C) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2303 }
2304
2305 FastUnlockNode *funlock = new (C) FastUnlockNode( ctrl, obj, box );
2306 funlock = transform_later( funlock )->as_FastUnlock();
2307 // Optimize test; set region slot 2
2308 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2309
2310 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 );
2311
2312 extract_call_projections(call);
2313
2314 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2315 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2316
2317 // No exceptions for unlocking
2318 // Capture slow path
2319 // disconnect fall-through projection from call and create a new one
2320 // hook up users of fall-through projection to region
2321 Node *slow_ctrl = _fallthroughproj->clone();
2322 transform_later(slow_ctrl);
2323 _igvn.hash_delete(_fallthroughproj);
2324 _fallthroughproj->disconnect_inputs(NULL, C);
2325 region->init_req(1, slow_ctrl);
2326 // region inputs are now complete
2327 transform_later(region);
2328 _igvn.replace_node(_fallthroughproj, region);
2329
2330 Node *memproj = transform_later( new(C) ProjNode(call, TypeFunc::Memory) );
2331 mem_phi->init_req(1, memproj );
2332 mem_phi->init_req(2, mem);
2333 transform_later(mem_phi);
2334 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2335 }
2336
2337 //---------------------------eliminate_macro_nodes----------------------
2338 // Eliminate scalar replaced allocations and associated locks.
2339 void PhaseMacroExpand::eliminate_macro_nodes() {
2340 if (C->macro_count() == 0)
2341 return;
2342
2343 // First, attempt to eliminate locks
2344 int cnt = C->macro_count();
2345 for (int i=0; i < cnt; i++) {
2346 Node *n = C->macro_node(i);
2347 if (n->is_AbstractLock()) { // Lock and Unlock nodes
2348 // Before elimination mark all associated (same box and obj)
2349 // lock and unlock nodes.
2350 mark_eliminated_locking_nodes(n->as_AbstractLock());
2351 }
2352 }
2353 bool progress = true;
2354 while (progress) {
2355 progress = false;
2356 for (int i = C->macro_count(); i > 0; i--) {
2357 Node * n = C->macro_node(i-1);
2358 bool success = false;
2359 debug_only(int old_macro_count = C->macro_count(););
2360 if (n->is_AbstractLock()) {
2361 success = eliminate_locking_node(n->as_AbstractLock());
2362 }
2363 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2364 progress = progress || success;
2365 }
2366 }
2367 // Next, attempt to eliminate allocations
2368 progress = true;
2369 while (progress) {
2370 progress = false;
2371 for (int i = C->macro_count(); i > 0; i--) {
2372 Node * n = C->macro_node(i-1);
2373 bool success = false;
2374 debug_only(int old_macro_count = C->macro_count(););
2375 switch (n->class_id()) {
2376 case Node::Class_Allocate:
2377 case Node::Class_AllocateArray:
2378 success = eliminate_allocate_node(n->as_Allocate());
2379 break;
2380 case Node::Class_Lock:
2381 case Node::Class_Unlock:
2382 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2383 break;
2384 default:
2385 assert(n->Opcode() == Op_LoopLimit ||
2386 n->Opcode() == Op_Opaque1 ||
2387 n->Opcode() == Op_Opaque2, "unknown node type in macro list");
2388 }
2389 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2390 progress = progress || success;
2391 }
2392 }
2393 }
2394
2395 //------------------------------expand_macro_nodes----------------------
2396 // Returns true if a failure occurred.
2397 bool PhaseMacroExpand::expand_macro_nodes() {
2398 // Last attempt to eliminate macro nodes.
2399 eliminate_macro_nodes();
2400
2401 // Make sure expansion will not cause node limit to be exceeded.
2402 // Worst case is a macro node gets expanded into about 50 nodes.
2403 // Allow 50% more for optimization.
2404 if (C->check_node_count(C->macro_count() * 75, "out of nodes before macro expansion" ) )
2405 return true;
2406
2407 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2408 bool progress = true;
2409 while (progress) {
2410 progress = false;
2411 for (int i = C->macro_count(); i > 0; i--) {
2412 Node * n = C->macro_node(i-1);
2413 bool success = false;
2414 debug_only(int old_macro_count = C->macro_count(););
2415 if (n->Opcode() == Op_LoopLimit) {
2416 // Remove it from macro list and put on IGVN worklist to optimize.
2417 C->remove_macro_node(n);
2418 _igvn._worklist.push(n);
2419 success = true;
2420 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) {
2421 _igvn.replace_node(n, n->in(1));
2422 success = true;
2423 }
2424 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2425 progress = progress || success;
2426 }
2427 }
2428
2429 // expand "macro" nodes
2430 // nodes are removed from the macro list as they are processed
2431 while (C->macro_count() > 0) {
2432 int macro_count = C->macro_count();
2433 Node * n = C->macro_node(macro_count-1);
2434 assert(n->is_macro(), "only macro nodes expected here");
2435 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) {
2436 // node is unreachable, so don't try to expand it
2437 C->remove_macro_node(n);
2438 continue;
2439 }
2440 switch (n->class_id()) {
2441 case Node::Class_Allocate:
2442 expand_allocate(n->as_Allocate());
2443 break;
2444 case Node::Class_AllocateArray:
2445 expand_allocate_array(n->as_AllocateArray());
2446 break;
2447 case Node::Class_Lock:
2448 expand_lock_node(n->as_Lock());
2449 break;
2450 case Node::Class_Unlock:
2451 expand_unlock_node(n->as_Unlock());
2452 break;
2453 default:
2454 assert(false, "unknown node type in macro list");
2455 }
2456 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2457 if (C->failing()) return true;
2458 }
2459
2460 _igvn.set_delay_transform(false);
2461 _igvn.optimize();
2462 if (C->failing()) return true;
2463 return false;
2464 }