1 /*
2 * Copyright (c) 2005, 2015, 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 "gc/shenandoah/brooksPointer.hpp"
28 #include "libadt/vectset.hpp"
29 #include "opto/addnode.hpp"
30 #include "opto/arraycopynode.hpp"
31 #include "opto/callnode.hpp"
32 #include "opto/castnode.hpp"
33 #include "opto/cfgnode.hpp"
34 #include "opto/compile.hpp"
35 #include "opto/convertnode.hpp"
36 #include "opto/locknode.hpp"
37 #include "opto/loopnode.hpp"
38 #include "opto/macro.hpp"
39 #include "opto/memnode.hpp"
40 #include "opto/narrowptrnode.hpp"
41 #include "opto/node.hpp"
42 #include "opto/opaquenode.hpp"
43 #include "opto/phaseX.hpp"
44 #include "opto/rootnode.hpp"
45 #include "opto/runtime.hpp"
46 #include "opto/shenandoahSupport.hpp"
47 #include "opto/subnode.hpp"
48 #include "opto/type.hpp"
49 #include "runtime/sharedRuntime.hpp"
50
51
52 //
53 // Replace any references to "oldref" in inputs to "use" with "newref".
54 // Returns the number of replacements made.
55 //
56 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
57 int nreplacements = 0;
58 uint req = use->req();
59 for (uint j = 0; j < use->len(); j++) {
60 Node *uin = use->in(j);
61 if (uin == oldref) {
62 if (j < req)
63 use->set_req(j, newref);
64 else
65 use->set_prec(j, newref);
66 nreplacements++;
67 } else if (j >= req && uin == NULL) {
68 break;
69 }
70 }
71 return nreplacements;
72 }
73
74 void PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) {
75 // Copy debug information and adjust JVMState information
76 uint old_dbg_start = oldcall->tf()->domain()->cnt();
77 uint new_dbg_start = newcall->tf()->domain()->cnt();
78 int jvms_adj = new_dbg_start - old_dbg_start;
79 assert (new_dbg_start == newcall->req(), "argument count mismatch");
80
81 // SafePointScalarObject node could be referenced several times in debug info.
82 // Use Dict to record cloned nodes.
83 Dict* sosn_map = new Dict(cmpkey,hashkey);
84 for (uint i = old_dbg_start; i < oldcall->req(); i++) {
85 Node* old_in = oldcall->in(i);
86 // Clone old SafePointScalarObjectNodes, adjusting their field contents.
87 if (old_in != NULL && old_in->is_SafePointScalarObject()) {
88 SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject();
89 uint old_unique = C->unique();
90 Node* new_in = old_sosn->clone(sosn_map);
91 if (old_unique != C->unique()) { // New node?
92 new_in->set_req(0, C->root()); // reset control edge
93 new_in = transform_later(new_in); // Register new node.
94 }
95 old_in = new_in;
96 }
97 newcall->add_req(old_in);
98 }
99
100 // JVMS may be shared so clone it before we modify it
101 newcall->set_jvms(oldcall->jvms() != NULL ? oldcall->jvms()->clone_deep(C) : NULL);
102 for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) {
103 jvms->set_map(newcall);
104 jvms->set_locoff(jvms->locoff()+jvms_adj);
105 jvms->set_stkoff(jvms->stkoff()+jvms_adj);
106 jvms->set_monoff(jvms->monoff()+jvms_adj);
107 jvms->set_scloff(jvms->scloff()+jvms_adj);
108 jvms->set_endoff(jvms->endoff()+jvms_adj);
109 }
110 }
111
112 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
113 Node* cmp;
114 if (mask != 0) {
115 Node* and_node = transform_later(new AndXNode(word, MakeConX(mask)));
116 cmp = transform_later(new CmpXNode(and_node, MakeConX(bits)));
117 } else {
118 cmp = word;
119 }
120 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne));
121 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
122 transform_later(iff);
123
124 // Fast path taken.
125 Node *fast_taken = transform_later(new IfFalseNode(iff));
126
127 // Fast path not-taken, i.e. slow path
128 Node *slow_taken = transform_later(new IfTrueNode(iff));
129
130 if (return_fast_path) {
131 region->init_req(edge, slow_taken); // Capture slow-control
132 return fast_taken;
133 } else {
134 region->init_req(edge, fast_taken); // Capture fast-control
135 return slow_taken;
136 }
137 }
138
139 //--------------------copy_predefined_input_for_runtime_call--------------------
140 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
141 // Set fixed predefined input arguments
142 call->init_req( TypeFunc::Control, ctrl );
143 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) );
144 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
145 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
146 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
147 }
148
149 //------------------------------make_slow_call---------------------------------
150 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type,
151 address slow_call, const char* leaf_name, Node* slow_path,
152 Node* parm0, Node* parm1, Node* parm2) {
153
154 // Slow-path call
155 CallNode *call = leaf_name
156 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
157 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM );
158
159 // Slow path call has no side-effects, uses few values
160 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
161 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0);
162 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1);
163 if (parm2 != NULL) call->init_req(TypeFunc::Parms+2, parm2);
164 copy_call_debug_info(oldcall, call);
165 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
166 _igvn.replace_node(oldcall, call);
167 transform_later(call);
168
169 return call;
170 }
171
172 void PhaseMacroExpand::extract_call_projections(CallNode *call) {
173 _fallthroughproj = NULL;
174 _fallthroughcatchproj = NULL;
175 _ioproj_fallthrough = NULL;
176 _ioproj_catchall = NULL;
177 _catchallcatchproj = NULL;
178 _memproj_fallthrough = NULL;
179 _memproj_catchall = NULL;
180 _resproj = NULL;
181 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) {
182 ProjNode *pn = call->fast_out(i)->as_Proj();
183 switch (pn->_con) {
184 case TypeFunc::Control:
185 {
186 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj
187 _fallthroughproj = pn;
188 DUIterator_Fast jmax, j = pn->fast_outs(jmax);
189 const Node *cn = pn->fast_out(j);
190 if (cn->is_Catch()) {
191 ProjNode *cpn = NULL;
192 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) {
193 cpn = cn->fast_out(k)->as_Proj();
194 assert(cpn->is_CatchProj(), "must be a CatchProjNode");
195 if (cpn->_con == CatchProjNode::fall_through_index)
196 _fallthroughcatchproj = cpn;
197 else {
198 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index.");
199 _catchallcatchproj = cpn;
200 }
201 }
202 }
203 break;
204 }
205 case TypeFunc::I_O:
206 if (pn->_is_io_use)
207 _ioproj_catchall = pn;
208 else
209 _ioproj_fallthrough = pn;
210 break;
211 case TypeFunc::Memory:
212 if (pn->_is_io_use)
213 _memproj_catchall = pn;
214 else
215 _memproj_fallthrough = pn;
216 break;
217 case TypeFunc::Parms:
218 _resproj = pn;
219 break;
220 default:
221 assert(false, "unexpected projection from allocation node.");
222 }
223 }
224
225 }
226
227 // Eliminate a card mark sequence. p2x is a ConvP2XNode
228 void PhaseMacroExpand::eliminate_card_mark(Node* p2x) {
229 assert(p2x->Opcode() == Op_CastP2X, "ConvP2XNode required");
230 if (!UseG1GC) {
231 // vanilla/CMS post barrier
232 Node *shift = p2x->unique_out();
233 Node *addp = shift->unique_out();
234 for (DUIterator_Last jmin, j = addp->last_outs(jmin); j >= jmin; --j) {
235 Node *mem = addp->last_out(j);
236 if (UseCondCardMark && mem->is_Load()) {
237 assert(mem->Opcode() == Op_LoadB, "unexpected code shape");
238 // The load is checking if the card has been written so
239 // replace it with zero to fold the test.
240 _igvn.replace_node(mem, intcon(0));
241 continue;
242 }
243 assert(mem->is_Store(), "store required");
244 _igvn.replace_node(mem, mem->in(MemNode::Memory));
245 }
246 } else {
247 // G1 pre/post barriers
248 assert(p2x->outcnt() <= 2, "expects 1 or 2 users: Xor and URShift nodes");
249 // It could be only one user, URShift node, in Object.clone() instrinsic
250 // but the new allocation is passed to arraycopy stub and it could not
251 // be scalar replaced. So we don't check the case.
252
253 // An other case of only one user (Xor) is when the value check for NULL
254 // in G1 post barrier is folded after CCP so the code which used URShift
255 // is removed.
256
257 // Take Region node before eliminating post barrier since it also
258 // eliminates CastP2X node when it has only one user.
259 Node* this_region = p2x->in(0);
260 assert(this_region != NULL, "");
261
262 // Remove G1 post barrier.
263
264 // Search for CastP2X->Xor->URShift->Cmp path which
265 // checks if the store done to a different from the value's region.
266 // And replace Cmp with #0 (false) to collapse G1 post barrier.
267 Node* xorx = p2x->find_out_with(Op_XorX);
268 assert(xorx != NULL, "missing G1 post barrier");
269 Node* shift = xorx->unique_out();
270 Node* cmpx = shift->unique_out();
271 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() &&
272 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne,
273 "missing region check in G1 post barrier");
274 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ));
275
276 // Remove G1 pre barrier.
277
278 // Search "if (marking != 0)" check and set it to "false".
279 // There is no G1 pre barrier if previous stored value is NULL
280 // (for example, after initialization).
281 if (this_region->is_Region() && this_region->req() == 3) {
282 int ind = 1;
283 if (!this_region->in(ind)->is_IfFalse()) {
284 ind = 2;
285 }
286 if (this_region->in(ind)->is_IfFalse()) {
287 Node* bol = this_region->in(ind)->in(0)->in(1);
288 assert(bol->is_Bool(), "");
289 cmpx = bol->in(1);
290 if (bol->as_Bool()->_test._test == BoolTest::ne &&
291 cmpx->is_Cmp() && cmpx->in(2) == intcon(0) &&
292 cmpx->in(1)->is_Load()) {
293 Node* adr = cmpx->in(1)->as_Load()->in(MemNode::Address);
294 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() +
295 PtrQueue::byte_offset_of_active());
296 if (adr->is_AddP() && adr->in(AddPNode::Base) == top() &&
297 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
298 adr->in(AddPNode::Offset) == MakeConX(marking_offset)) {
299 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ));
300 }
301 }
302 }
303 }
304 // Now CastP2X can be removed since it is used only on dead path
305 // which currently still alive until igvn optimize it.
306 assert(p2x->outcnt() == 0 || p2x->unique_out()->Opcode() == Op_URShiftX, "");
307 _igvn.replace_node(p2x, top());
308 }
309 }
310
311 // Search for a memory operation for the specified memory slice.
312 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
313 Node *orig_mem = mem;
314 Node *alloc_mem = alloc->in(TypeFunc::Memory);
315 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
316 while (true) {
317 if (mem == alloc_mem || mem == start_mem ) {
318 return mem; // hit one of our sentinels
319 } else if (mem->is_MergeMem()) {
320 mem = mem->as_MergeMem()->memory_at(alias_idx);
321 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
322 Node *in = mem->in(0);
323 // we can safely skip over safepoints, calls, locks and membars because we
324 // already know that the object is safe to eliminate.
325 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
326 return in;
327 } else if (in->is_Call()) {
328 CallNode *call = in->as_Call();
329 if (call->may_modify(tinst, phase)) {
330 assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape");
331
332 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) {
333 return in;
334 }
335 }
336 mem = in->in(TypeFunc::Memory);
337 } else if (in->is_MemBar()) {
338 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase)) {
339 assert(in->in(0)->is_Proj() && in->in(0)->in(0)->is_ArrayCopy(), "should be arraycopy");
340 ArrayCopyNode* ac = in->in(0)->in(0)->as_ArrayCopy();
341 assert(ac->is_clonebasic(), "Only basic clone is a non escaping clone");
342 return ac;
343 }
344 mem = in->in(TypeFunc::Memory);
345 } else {
346 assert(false, "unexpected projection");
347 }
348 } else if (mem->is_Store()) {
349 const TypePtr* atype = mem->as_Store()->adr_type();
350 int adr_idx = phase->C->get_alias_index(atype);
351 if (adr_idx == alias_idx) {
352 assert(atype->isa_oopptr(), "address type must be oopptr");
353 int adr_offset = atype->offset();
354 uint adr_iid = atype->is_oopptr()->instance_id();
355 // Array elements references have the same alias_idx
356 // but different offset and different instance_id.
357 if (adr_offset == offset && adr_iid == alloc->_idx)
358 return mem;
359 } else {
360 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
361 }
362 mem = mem->in(MemNode::Memory);
363 } else if (mem->is_ClearArray()) {
364 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
365 // Can not bypass initialization of the instance
366 // we are looking.
367 debug_only(intptr_t offset;)
368 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
369 InitializeNode* init = alloc->as_Allocate()->initialization();
370 // We are looking for stored value, return Initialize node
371 // or memory edge from Allocate node.
372 if (init != NULL)
373 return init;
374 else
375 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers).
376 }
377 // Otherwise skip it (the call updated 'mem' value).
378 } else if (mem->Opcode() == Op_SCMemProj) {
379 mem = mem->in(0);
380 Node* adr = NULL;
381 if (mem->is_LoadStore()) {
382 adr = mem->in(MemNode::Address);
383 } else {
384 assert(mem->Opcode() == Op_EncodeISOArray, "sanity");
385 adr = mem->in(3); // Destination array
386 }
387 const TypePtr* atype = adr->bottom_type()->is_ptr();
388 int adr_idx = phase->C->get_alias_index(atype);
389 if (adr_idx == alias_idx) {
390 assert(false, "Object is not scalar replaceable if a LoadStore node access its field");
391 return NULL;
392 }
393 mem = mem->in(MemNode::Memory);
394 } else {
395 return mem;
396 }
397 assert(mem != orig_mem, "dead memory loop");
398 }
399 }
400
401 // Generate loads from source of the arraycopy for fields of
402 // destination needed at a deoptimization point
403 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, BasicType ft, const Type *ftype, AllocateNode *alloc) {
404 BasicType bt = ft;
405 const Type *type = ftype;
406 if (ft == T_NARROWOOP) {
407 bt = T_OBJECT;
408 type = ftype->make_oopptr();
409 }
410 Node* res = NULL;
411 if (ac->is_clonebasic()) {
412 Node* base = ac->in(ArrayCopyNode::Src)->in(AddPNode::Base);
413 Node* adr = _igvn.transform(new AddPNode(base, base, MakeConX(offset)));
414 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
415 Node* m = ac->in(TypeFunc::Memory);
416 while (m->is_MergeMem()) {
417 m = m->as_MergeMem()->memory_at(C->get_alias_index(adr_type));
418 if (m->is_Proj() && m->in(0)->is_MemBar()) {
419 m = m->in(0)->in(TypeFunc::Memory);
420 }
421 }
422 res = LoadNode::make(_igvn, ctl, m, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned);
423 } else {
424 if (ac->modifies(offset, offset, &_igvn, true)) {
425 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
426 uint shift = exact_log2(type2aelembytes(bt));
427 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
428 #ifdef _LP64
429 diff = _igvn.transform(new ConvI2LNode(diff));
430 #endif
431 diff = _igvn.transform(new LShiftXNode(diff, intcon(shift)));
432
433 Node* off = _igvn.transform(new AddXNode(MakeConX(offset), diff));
434 Node* base = ac->in(ArrayCopyNode::Src);
435 Node* adr = _igvn.transform(new AddPNode(base, base, off));
436 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
437 Node* m = ac->in(TypeFunc::Memory);
438 res = LoadNode::make(_igvn, ctl, m, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned);
439 }
440 }
441 if (res != NULL) {
442 res = _igvn.transform(res);
443 if (ftype->isa_narrowoop()) {
444 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
445 res = _igvn.transform(new EncodePNode(res, ftype));
446 }
447 return res;
448 }
449 return NULL;
450 }
451
452 //
453 // Given a Memory Phi, compute a value Phi containing the values from stores
454 // on the input paths.
455 // Note: this function is recursive, its depth is limited by the "level" argument
456 // Returns the computed Phi, or NULL if it cannot compute it.
457 Node *PhaseMacroExpand::value_from_mem_phi(Node *mem, BasicType ft, const Type *phi_type, const TypeOopPtr *adr_t, AllocateNode *alloc, Node_Stack *value_phis, int level) {
458 assert(mem->is_Phi(), "sanity");
459 int alias_idx = C->get_alias_index(adr_t);
460 int offset = adr_t->offset();
461 int instance_id = adr_t->instance_id();
462
463 // Check if an appropriate value phi already exists.
464 Node* region = mem->in(0);
465 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
466 Node* phi = region->fast_out(k);
467 if (phi->is_Phi() && phi != mem &&
468 phi->as_Phi()->is_same_inst_field(phi_type, instance_id, alias_idx, offset)) {
469 return phi;
470 }
471 }
472 // Check if an appropriate new value phi already exists.
473 Node* new_phi = value_phis->find(mem->_idx);
474 if (new_phi != NULL)
475 return new_phi;
476
477 if (level <= 0) {
478 return NULL; // Give up: phi tree too deep
479 }
480 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
481 Node *alloc_mem = alloc->in(TypeFunc::Memory);
482
483 uint length = mem->req();
484 GrowableArray <Node *> values(length, length, NULL, false);
485
486 // create a new Phi for the value
487 PhiNode *phi = new PhiNode(mem->in(0), phi_type, NULL, instance_id, alias_idx, offset);
488 transform_later(phi);
489 value_phis->push(phi, mem->_idx);
490
491 for (uint j = 1; j < length; j++) {
492 Node *in = mem->in(j);
493 if (in == NULL || in->is_top()) {
494 values.at_put(j, in);
495 } else {
496 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
497 if (val == start_mem || val == alloc_mem) {
498 // hit a sentinel, return appropriate 0 value
499 values.at_put(j, _igvn.zerocon(ft));
500 continue;
501 }
502 if (val->is_Initialize()) {
503 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
504 }
505 if (val == NULL) {
506 return NULL; // can't find a value on this path
507 }
508 if (val == mem) {
509 values.at_put(j, mem);
510 } else if (val->is_Store()) {
511 values.at_put(j, val->in(MemNode::ValueIn));
512 } else if(val->is_Proj() && val->in(0) == alloc) {
513 values.at_put(j, _igvn.zerocon(ft));
514 } else if (val->is_Phi()) {
515 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
516 if (val == NULL) {
517 return NULL;
518 }
519 values.at_put(j, val);
520 } else if (val->Opcode() == Op_SCMemProj) {
521 assert(val->in(0)->is_LoadStore() || val->in(0)->Opcode() == Op_EncodeISOArray, "sanity");
522 assert(false, "Object is not scalar replaceable if a LoadStore node access its field");
523 return NULL;
524 } else if (val->is_ArrayCopy()) {
525 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), ft, phi_type, alloc);
526 if (res == NULL) {
527 return NULL;
528 }
529 values.at_put(j, res);
530 } else {
531 #ifdef ASSERT
532 val->dump();
533 assert(false, "unknown node on this path");
534 #endif
535 return NULL; // unknown node on this path
536 }
537 }
538 }
539 // Set Phi's inputs
540 for (uint j = 1; j < length; j++) {
541 if (values.at(j) == mem) {
542 phi->init_req(j, phi);
543 } else {
544 phi->init_req(j, values.at(j));
545 }
546 }
547 return phi;
548 }
549
550 // Search the last value stored into the object's field.
551 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) {
552 assert(adr_t->is_known_instance_field(), "instance required");
553 int instance_id = adr_t->instance_id();
554 assert((uint)instance_id == alloc->_idx, "wrong allocation");
555
556 int alias_idx = C->get_alias_index(adr_t);
557 int offset = adr_t->offset();
558 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
559 Node *alloc_ctrl = alloc->in(TypeFunc::Control);
560 Node *alloc_mem = alloc->in(TypeFunc::Memory);
561 Arena *a = Thread::current()->resource_area();
562 VectorSet visited(a);
563
564
565 bool done = sfpt_mem == alloc_mem;
566 Node *mem = sfpt_mem;
567 while (!done) {
568 if (visited.test_set(mem->_idx)) {
569 return NULL; // found a loop, give up
570 }
571 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
572 if (mem == start_mem || mem == alloc_mem) {
573 done = true; // hit a sentinel, return appropriate 0 value
574 } else if (mem->is_Initialize()) {
575 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
576 if (mem == NULL) {
577 done = true; // Something go wrong.
578 } else if (mem->is_Store()) {
579 const TypePtr* atype = mem->as_Store()->adr_type();
580 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
581 done = true;
582 }
583 } else if (mem->is_Store()) {
584 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
585 assert(atype != NULL, "address type must be oopptr");
586 assert(C->get_alias_index(atype) == alias_idx &&
587 atype->is_known_instance_field() && atype->offset() == offset &&
588 atype->instance_id() == instance_id, "store is correct memory slice");
589 done = true;
590 } else if (mem->is_Phi()) {
591 // try to find a phi's unique input
592 Node *unique_input = NULL;
593 Node *top = C->top();
594 for (uint i = 1; i < mem->req(); i++) {
595 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
596 if (n == NULL || n == top || n == mem) {
597 continue;
598 } else if (unique_input == NULL) {
599 unique_input = n;
600 } else if (unique_input != n) {
601 unique_input = top;
602 break;
603 }
604 }
605 if (unique_input != NULL && unique_input != top) {
606 mem = unique_input;
607 } else {
608 done = true;
609 }
610 } else if (mem->is_ArrayCopy()) {
611 done = true;
612 } else {
613 assert(false, "unexpected node");
614 }
615 }
616 if (mem != NULL) {
617 if (mem == start_mem || mem == alloc_mem) {
618 // hit a sentinel, return appropriate 0 value
619 return _igvn.zerocon(ft);
620 } else if (mem->is_Store()) {
621 return mem->in(MemNode::ValueIn);
622 } else if (mem->is_Phi()) {
623 // attempt to produce a Phi reflecting the values on the input paths of the Phi
624 Node_Stack value_phis(a, 8);
625 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
626 if (phi != NULL) {
627 return phi;
628 } else {
629 // Kill all new Phis
630 while(value_phis.is_nonempty()) {
631 Node* n = value_phis.node();
632 _igvn.replace_node(n, C->top());
633 value_phis.pop();
634 }
635 }
636 } else if (mem->is_ArrayCopy()) {
637 Node* ctl = mem->in(0);
638 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) {
639 // pin the loads in the uncommon trap path
640 ctl = sfpt_ctl;
641 }
642 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, ft, ftype, alloc);
643 }
644 }
645 // Something go wrong.
646 return NULL;
647 }
648
649 // Check the possibility of scalar replacement.
650 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
651 // Scan the uses of the allocation to check for anything that would
652 // prevent us from eliminating it.
653 NOT_PRODUCT( const char* fail_eliminate = NULL; )
654 DEBUG_ONLY( Node* disq_node = NULL; )
655 bool can_eliminate = true;
656
657 Node* res = alloc->result_cast();
658 const TypeOopPtr* res_type = NULL;
659 if (res == NULL) {
660 // All users were eliminated.
661 } else if (!res->is_CheckCastPP()) {
662 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
663 can_eliminate = false;
664 } else {
665 res_type = _igvn.type(res)->isa_oopptr();
666 if (res_type == NULL) {
667 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
668 can_eliminate = false;
669 } else if (res_type->isa_aryptr()) {
670 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
671 if (length < 0) {
672 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
673 can_eliminate = false;
674 }
675 }
676 }
677
678 if (can_eliminate && res != NULL) {
679 for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
680 j < jmax && can_eliminate; j++) {
681 Node* use = res->fast_out(j);
682
683 if (use->is_AddP()) {
684 const TypePtr* addp_type = _igvn.type(use)->is_ptr();
685 int offset = addp_type->offset();
686
687 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
688 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";)
689 can_eliminate = false;
690 break;
691 }
692 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
693 k < kmax && can_eliminate; k++) {
694 Node* n = use->fast_out(k);
695 if (!n->is_Store() && n->Opcode() != Op_CastP2X &&
696 !(n->is_ArrayCopy() &&
697 n->as_ArrayCopy()->is_clonebasic() &&
698 n->in(ArrayCopyNode::Dest) == use)) {
699 DEBUG_ONLY(disq_node = n;)
700 if (n->is_Load() || n->is_LoadStore()) {
701 NOT_PRODUCT(fail_eliminate = "Field load";)
702 } else {
703 NOT_PRODUCT(fail_eliminate = "Not store field referrence";)
704 }
705 can_eliminate = false;
706 }
707 }
708 } else if (use->is_ArrayCopy() &&
709 (use->as_ArrayCopy()->is_arraycopy_validated() ||
710 use->as_ArrayCopy()->is_copyof_validated() ||
711 use->as_ArrayCopy()->is_copyofrange_validated()) &&
712 use->in(ArrayCopyNode::Dest) == res) {
713 // ok to eliminate
714 } else if (use->is_SafePoint()) {
715 SafePointNode* sfpt = use->as_SafePoint();
716 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
717 // Object is passed as argument.
718 DEBUG_ONLY(disq_node = use;)
719 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
720 can_eliminate = false;
721 }
722 Node* sfptMem = sfpt->memory();
723 if (sfptMem == NULL || sfptMem->is_top()) {
724 DEBUG_ONLY(disq_node = use;)
725 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";)
726 can_eliminate = false;
727 } else {
728 safepoints.append_if_missing(sfpt);
729 }
730 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
731 if (use->is_Phi()) {
732 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
733 NOT_PRODUCT(fail_eliminate = "Object is return value";)
734 } else {
735 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
736 }
737 DEBUG_ONLY(disq_node = use;)
738 } else {
739 if (use->Opcode() == Op_Return) {
740 NOT_PRODUCT(fail_eliminate = "Object is return value";)
741 }else {
742 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
743 }
744 DEBUG_ONLY(disq_node = use;)
745 }
746 can_eliminate = false;
747 }
748 }
749 }
750
751 #ifndef PRODUCT
752 if (PrintEliminateAllocations) {
753 if (can_eliminate) {
754 tty->print("Scalar ");
755 if (res == NULL)
756 alloc->dump();
757 else
758 res->dump();
759 } else if (alloc->_is_scalar_replaceable) {
760 tty->print("NotScalar (%s)", fail_eliminate);
761 if (res == NULL)
762 alloc->dump();
763 else
764 res->dump();
765 #ifdef ASSERT
766 if (disq_node != NULL) {
767 tty->print(" >>>> ");
768 disq_node->dump();
769 }
770 #endif /*ASSERT*/
771 }
772 }
773 #endif
774 return can_eliminate;
775 }
776
777 // Do scalar replacement.
778 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
779 GrowableArray <SafePointNode *> safepoints_done;
780
781 ciKlass* klass = NULL;
782 ciInstanceKlass* iklass = NULL;
783 int nfields = 0;
784 int array_base;
785 int element_size;
786 BasicType basic_elem_type;
787 ciType* elem_type;
788
789 Node* res = alloc->result_cast();
790 assert(res == NULL || res->is_CheckCastPP(), "unexpected AllocateNode result");
791 const TypeOopPtr* res_type = NULL;
792 if (res != NULL) { // Could be NULL when there are no users
793 res_type = _igvn.type(res)->isa_oopptr();
794 }
795
796 if (res != NULL) {
797 klass = res_type->klass();
798 if (res_type->isa_instptr()) {
799 // find the fields of the class which will be needed for safepoint debug information
800 assert(klass->is_instance_klass(), "must be an instance klass.");
801 iklass = klass->as_instance_klass();
802 nfields = iklass->nof_nonstatic_fields();
803 } else {
804 // find the array's elements which will be needed for safepoint debug information
805 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
806 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass.");
807 elem_type = klass->as_array_klass()->element_type();
808 basic_elem_type = elem_type->basic_type();
809 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
810 element_size = type2aelembytes(basic_elem_type);
811 }
812 }
813 //
814 // Process the safepoint uses
815 //
816 while (safepoints.length() > 0) {
817 SafePointNode* sfpt = safepoints.pop();
818 Node* mem = sfpt->memory();
819 Node* ctl = sfpt->control();
820 assert(sfpt->jvms() != NULL, "missed JVMS");
821 // Fields of scalar objs are referenced only at the end
822 // of regular debuginfo at the last (youngest) JVMS.
823 // Record relative start index.
824 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
825 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type,
826 #ifdef ASSERT
827 alloc,
828 #endif
829 first_ind, nfields);
830 sobj->init_req(0, C->root());
831 transform_later(sobj);
832
833 // Scan object's fields adding an input to the safepoint for each field.
834 for (int j = 0; j < nfields; j++) {
835 intptr_t offset;
836 ciField* field = NULL;
837 if (iklass != NULL) {
838 field = iklass->nonstatic_field_at(j);
839 offset = field->offset();
840 elem_type = field->type();
841 basic_elem_type = field->layout_type();
842 } else {
843 offset = array_base + j * (intptr_t)element_size;
844 }
845
846 const Type *field_type;
847 // The next code is taken from Parse::do_get_xxx().
848 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) {
849 if (!elem_type->is_loaded()) {
850 field_type = TypeInstPtr::BOTTOM;
851 } else if (field != NULL && field->is_constant() && field->is_static()) {
852 // This can happen if the constant oop is non-perm.
853 ciObject* con = field->constant_value().as_object();
854 // Do not "join" in the previous type; it doesn't add value,
855 // and may yield a vacuous result if the field is of interface type.
856 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
857 assert(field_type != NULL, "field singleton type must be consistent");
858 } else {
859 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
860 }
861 if (UseCompressedOops) {
862 field_type = field_type->make_narrowoop();
863 basic_elem_type = T_NARROWOOP;
864 }
865 } else {
866 field_type = Type::get_const_basic_type(basic_elem_type);
867 }
868
869 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr();
870
871 Node *field_val = value_from_mem(mem, ctl, basic_elem_type, field_type, field_addr_type, alloc);
872 if (field_val == NULL) {
873 // We weren't able to find a value for this field,
874 // give up on eliminating this allocation.
875
876 // Remove any extra entries we added to the safepoint.
877 uint last = sfpt->req() - 1;
878 for (int k = 0; k < j; k++) {
879 sfpt->del_req(last--);
880 }
881 _igvn._worklist.push(sfpt);
882 // rollback processed safepoints
883 while (safepoints_done.length() > 0) {
884 SafePointNode* sfpt_done = safepoints_done.pop();
885 // remove any extra entries we added to the safepoint
886 last = sfpt_done->req() - 1;
887 for (int k = 0; k < nfields; k++) {
888 sfpt_done->del_req(last--);
889 }
890 JVMState *jvms = sfpt_done->jvms();
891 jvms->set_endoff(sfpt_done->req());
892 // Now make a pass over the debug information replacing any references
893 // to SafePointScalarObjectNode with the allocated object.
894 int start = jvms->debug_start();
895 int end = jvms->debug_end();
896 for (int i = start; i < end; i++) {
897 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
898 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
899 if (scobj->first_index(jvms) == sfpt_done->req() &&
900 scobj->n_fields() == (uint)nfields) {
901 assert(scobj->alloc() == alloc, "sanity");
902 sfpt_done->set_req(i, res);
903 }
904 }
905 }
906 _igvn._worklist.push(sfpt_done);
907 }
908 #ifndef PRODUCT
909 if (PrintEliminateAllocations) {
910 if (field != NULL) {
911 tty->print("=== At SafePoint node %d can't find value of Field: ",
912 sfpt->_idx);
913 field->print();
914 int field_idx = C->get_alias_index(field_addr_type);
915 tty->print(" (alias_idx=%d)", field_idx);
916 } else { // Array's element
917 tty->print("=== At SafePoint node %d can't find value of array element [%d]",
918 sfpt->_idx, j);
919 }
920 tty->print(", which prevents elimination of: ");
921 if (res == NULL)
922 alloc->dump();
923 else
924 res->dump();
925 }
926 #endif
927 return false;
928 }
929 if (UseCompressedOops && field_type->isa_narrowoop()) {
930 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
931 // to be able scalar replace the allocation.
932 if (field_val->is_EncodeP()) {
933 field_val = field_val->in(1);
934 } else {
935 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
936 }
937 }
938 if (field_val->isa_ShenandoahBarrier()) {
939 field_val = field_val->in(ShenandoahBarrierNode::ValueIn);
940 }
941 sfpt->add_req(field_val);
942 }
943 JVMState *jvms = sfpt->jvms();
944 jvms->set_endoff(sfpt->req());
945 // Now make a pass over the debug information replacing any references
946 // to the allocated object with "sobj"
947 int start = jvms->debug_start();
948 int end = jvms->debug_end();
949 sfpt->replace_edges_in_range(res, sobj, start, end);
950 _igvn._worklist.push(sfpt);
951 safepoints_done.append_if_missing(sfpt); // keep it for rollback
952 }
953 return true;
954 }
955
956 // Process users of eliminated allocation.
957 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) {
958 Node* res = alloc->result_cast();
959 if (res != NULL) {
960 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
961 Node *use = res->last_out(j);
962 uint oc1 = res->outcnt();
963
964 if (use->is_AddP()) {
965 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
966 Node *n = use->last_out(k);
967 uint oc2 = use->outcnt();
968 if (n->is_Store()) {
969 #ifdef ASSERT
970 // Verify that there is no dependent MemBarVolatile nodes,
971 // they should be removed during IGVN, see MemBarNode::Ideal().
972 for (DUIterator_Fast pmax, p = n->fast_outs(pmax);
973 p < pmax; p++) {
974 Node* mb = n->fast_out(p);
975 assert(mb->is_Initialize() || !mb->is_MemBar() ||
976 mb->req() <= MemBarNode::Precedent ||
977 mb->in(MemBarNode::Precedent) != n,
978 "MemBarVolatile should be eliminated for non-escaping object");
979 }
980 #endif
981 _igvn.replace_node(n, n->in(MemNode::Memory));
982 } else if (n->is_ArrayCopy()) {
983 // Disconnect ArrayCopy node
984 ArrayCopyNode* ac = n->as_ArrayCopy();
985 assert(ac->is_clonebasic(), "unexpected array copy kind");
986 Node* ctl_proj = ac->proj_out(TypeFunc::Control);
987 Node* mem_proj = ac->proj_out(TypeFunc::Memory);
988 if (ctl_proj != NULL) {
989 _igvn.replace_node(ctl_proj, n->in(0));
990 }
991 if (mem_proj != NULL) {
992 _igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
993 }
994 } else {
995 eliminate_card_mark(n);
996 }
997 k -= (oc2 - use->outcnt());
998 }
999 } else if (use->is_ArrayCopy()) {
1000 // Disconnect ArrayCopy node
1001 ArrayCopyNode* ac = use->as_ArrayCopy();
1002 assert(ac->is_arraycopy_validated() ||
1003 ac->is_copyof_validated() ||
1004 ac->is_copyofrange_validated(), "unsupported");
1005 CallProjections callprojs;
1006 ac->extract_projections(&callprojs, true);
1007
1008 _igvn.replace_node(callprojs.fallthrough_ioproj, ac->in(TypeFunc::I_O));
1009 _igvn.replace_node(callprojs.fallthrough_memproj, ac->in(TypeFunc::Memory));
1010 _igvn.replace_node(callprojs.fallthrough_catchproj, ac->in(TypeFunc::Control));
1011
1012 // Set control to top. IGVN will remove the remaining projections
1013 ac->set_req(0, top());
1014 ac->replace_edge(res, top());
1015
1016 // Disconnect src right away: it can help find new
1017 // opportunities for allocation elimination
1018 Node* src = ac->in(ArrayCopyNode::Src);
1019 ac->replace_edge(src, top());
1020 if (src->outcnt() == 0) {
1021 _igvn.remove_dead_node(src);
1022 }
1023
1024 _igvn._worklist.push(ac);
1025 } else {
1026 eliminate_card_mark(use);
1027 }
1028 j -= (oc1 - res->outcnt());
1029 }
1030 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
1031 _igvn.remove_dead_node(res);
1032 }
1033
1034 //
1035 // Process other users of allocation's projections
1036 //
1037 if (_resproj != NULL && _resproj->outcnt() != 0) {
1038 // First disconnect stores captured by Initialize node.
1039 // If Initialize node is eliminated first in the following code,
1040 // it will kill such stores and DUIterator_Last will assert.
1041 for (DUIterator_Fast jmax, j = _resproj->fast_outs(jmax); j < jmax; j++) {
1042 Node *use = _resproj->fast_out(j);
1043 if (use->is_AddP()) {
1044 // raw memory addresses used only by the initialization
1045 _igvn.replace_node(use, C->top());
1046 --j; --jmax;
1047 }
1048 }
1049 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) {
1050 Node *use = _resproj->last_out(j);
1051 uint oc1 = _resproj->outcnt();
1052 if (use->is_Initialize()) {
1053 // Eliminate Initialize node.
1054 InitializeNode *init = use->as_Initialize();
1055 assert(init->outcnt() <= 2, "only a control and memory projection expected");
1056 Node *ctrl_proj = init->proj_out(TypeFunc::Control);
1057 if (ctrl_proj != NULL) {
1058 assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection");
1059 _igvn.replace_node(ctrl_proj, _fallthroughcatchproj);
1060 }
1061 Node *mem_proj = init->proj_out(TypeFunc::Memory);
1062 if (mem_proj != NULL) {
1063 Node *mem = init->in(TypeFunc::Memory);
1064 #ifdef ASSERT
1065 if (mem->is_MergeMem()) {
1066 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection");
1067 } else {
1068 assert(mem == _memproj_fallthrough, "allocation memory projection");
1069 }
1070 #endif
1071 _igvn.replace_node(mem_proj, mem);
1072 }
1073 } else {
1074 assert(false, "only Initialize or AddP expected");
1075 }
1076 j -= (oc1 - _resproj->outcnt());
1077 }
1078 }
1079 if (_fallthroughcatchproj != NULL) {
1080 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control));
1081 }
1082 if (_memproj_fallthrough != NULL) {
1083 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory));
1084 }
1085 if (_memproj_catchall != NULL) {
1086 _igvn.replace_node(_memproj_catchall, C->top());
1087 }
1088 if (_ioproj_fallthrough != NULL) {
1089 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O));
1090 }
1091 if (_ioproj_catchall != NULL) {
1092 _igvn.replace_node(_ioproj_catchall, C->top());
1093 }
1094 if (_catchallcatchproj != NULL) {
1095 _igvn.replace_node(_catchallcatchproj, C->top());
1096 }
1097 }
1098
1099 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1100 // Don't do scalar replacement if the frame can be popped by JVMTI:
1101 // if reallocation fails during deoptimization we'll pop all
1102 // interpreter frames for this compiled frame and that won't play
1103 // nice with JVMTI popframe.
1104 if (!EliminateAllocations || JvmtiExport::can_pop_frame() || !alloc->_is_non_escaping) {
1105 return false;
1106 }
1107 Node* klass = alloc->in(AllocateNode::KlassNode);
1108 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1109 Node* res = alloc->result_cast();
1110 // Eliminate boxing allocations which are not used
1111 // regardless scalar replacable status.
1112 bool boxing_alloc = C->eliminate_boxing() &&
1113 tklass->klass()->is_instance_klass() &&
1114 tklass->klass()->as_instance_klass()->is_box_klass();
1115 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != NULL))) {
1116 return false;
1117 }
1118
1119 extract_call_projections(alloc);
1120
1121 GrowableArray <SafePointNode *> safepoints;
1122 if (!can_eliminate_allocation(alloc, safepoints)) {
1123 return false;
1124 }
1125
1126 if (!alloc->_is_scalar_replaceable) {
1127 assert(res == NULL, "sanity");
1128 // We can only eliminate allocation if all debug info references
1129 // are already replaced with SafePointScalarObject because
1130 // we can't search for a fields value without instance_id.
1131 if (safepoints.length() > 0) {
1132 return false;
1133 }
1134 }
1135
1136 if (!scalar_replacement(alloc, safepoints)) {
1137 return false;
1138 }
1139
1140 CompileLog* log = C->log();
1141 if (log != NULL) {
1142 log->head("eliminate_allocation type='%d'",
1143 log->identify(tklass->klass()));
1144 JVMState* p = alloc->jvms();
1145 while (p != NULL) {
1146 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1147 p = p->caller();
1148 }
1149 log->tail("eliminate_allocation");
1150 }
1151
1152 process_users_of_allocation(alloc);
1153
1154 #ifndef PRODUCT
1155 if (PrintEliminateAllocations) {
1156 if (alloc->is_AllocateArray())
1157 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1158 else
1159 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1160 }
1161 #endif
1162
1163 return true;
1164 }
1165
1166 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1167 // EA should remove all uses of non-escaping boxing node.
1168 if (!C->eliminate_boxing() || boxing->proj_out(TypeFunc::Parms) != NULL) {
1169 return false;
1170 }
1171
1172 assert(boxing->result_cast() == NULL, "unexpected boxing node result");
1173
1174 extract_call_projections(boxing);
1175
1176 const TypeTuple* r = boxing->tf()->range();
1177 assert(r->cnt() > TypeFunc::Parms, "sanity");
1178 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1179 assert(t != NULL, "sanity");
1180
1181 CompileLog* log = C->log();
1182 if (log != NULL) {
1183 log->head("eliminate_boxing type='%d'",
1184 log->identify(t->klass()));
1185 JVMState* p = boxing->jvms();
1186 while (p != NULL) {
1187 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1188 p = p->caller();
1189 }
1190 log->tail("eliminate_boxing");
1191 }
1192
1193 process_users_of_allocation(boxing);
1194
1195 #ifndef PRODUCT
1196 if (PrintEliminateAllocations) {
1197 tty->print("++++ Eliminated: %d ", boxing->_idx);
1198 boxing->method()->print_short_name(tty);
1199 tty->cr();
1200 }
1201 #endif
1202
1203 return true;
1204 }
1205
1206 //---------------------------set_eden_pointers-------------------------
1207 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) {
1208 if (UseTLAB) { // Private allocation: load from TLS
1209 Node* thread = transform_later(new ThreadLocalNode());
1210 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset());
1211 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset());
1212 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset);
1213 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset);
1214 } else { // Shared allocation: load from globals
1215 CollectedHeap* ch = Universe::heap();
1216 address top_adr = (address)ch->top_addr();
1217 address end_adr = (address)ch->end_addr();
1218 eden_top_adr = makecon(TypeRawPtr::make(top_adr));
1219 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr);
1220 }
1221 }
1222
1223
1224 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
1225 Node* adr = basic_plus_adr(base, offset);
1226 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
1227 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered);
1228 transform_later(value);
1229 return value;
1230 }
1231
1232
1233 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
1234 Node* adr = basic_plus_adr(base, offset);
1235 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt, MemNode::unordered);
1236 transform_later(mem);
1237 return mem;
1238 }
1239
1240 //=============================================================================
1241 //
1242 // A L L O C A T I O N
1243 //
1244 // Allocation attempts to be fast in the case of frequent small objects.
1245 // It breaks down like this:
1246 //
1247 // 1) Size in doublewords is computed. This is a constant for objects and
1248 // variable for most arrays. Doubleword units are used to avoid size
1249 // overflow of huge doubleword arrays. We need doublewords in the end for
1250 // rounding.
1251 //
1252 // 2) Size is checked for being 'too large'. Too-large allocations will go
1253 // the slow path into the VM. The slow path can throw any required
1254 // exceptions, and does all the special checks for very large arrays. The
1255 // size test can constant-fold away for objects. For objects with
1256 // finalizers it constant-folds the otherway: you always go slow with
1257 // finalizers.
1258 //
1259 // 3) If NOT using TLABs, this is the contended loop-back point.
1260 // Load-Locked the heap top. If using TLABs normal-load the heap top.
1261 //
1262 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
1263 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
1264 // "size*8" we always enter the VM, where "largish" is a constant picked small
1265 // enough that there's always space between the eden max and 4Gig (old space is
1266 // there so it's quite large) and large enough that the cost of entering the VM
1267 // is dwarfed by the cost to initialize the space.
1268 //
1269 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1270 // down. If contended, repeat at step 3. If using TLABs normal-store
1271 // adjusted heap top back down; there is no contention.
1272 //
1273 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
1274 // fields.
1275 //
1276 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1277 // oop flavor.
1278 //
1279 //=============================================================================
1280 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1281 // Allocations bigger than this always go the slow route.
1282 // This value must be small enough that allocation attempts that need to
1283 // trigger exceptions go the slow route. Also, it must be small enough so
1284 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1285 //=============================================================================j//
1286 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1287 // The allocator will coalesce int->oop copies away. See comment in
1288 // coalesce.cpp about how this works. It depends critically on the exact
1289 // code shape produced here, so if you are changing this code shape
1290 // make sure the GC info for the heap-top is correct in and around the
1291 // slow-path call.
1292 //
1293
1294 void PhaseMacroExpand::expand_allocate_common(
1295 AllocateNode* alloc, // allocation node to be expanded
1296 Node* length, // array length for an array allocation
1297 const TypeFunc* slow_call_type, // Type of slow call
1298 address slow_call_address // Address of slow call
1299 )
1300 {
1301
1302 Node* ctrl = alloc->in(TypeFunc::Control);
1303 Node* mem = alloc->in(TypeFunc::Memory);
1304 Node* i_o = alloc->in(TypeFunc::I_O);
1305 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1306 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1307 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1308
1309 assert(ctrl != NULL, "must have control");
1310 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1311 // they will not be used if "always_slow" is set
1312 enum { slow_result_path = 1, fast_result_path = 2 };
1313 Node *result_region;
1314 Node *result_phi_rawmem;
1315 Node *result_phi_rawoop;
1316 Node *result_phi_i_o;
1317
1318 // The initial slow comparison is a size check, the comparison
1319 // we want to do is a BoolTest::gt
1320 bool always_slow = false;
1321 int tv = _igvn.find_int_con(initial_slow_test, -1);
1322 if (tv >= 0) {
1323 always_slow = (tv == 1);
1324 initial_slow_test = NULL;
1325 } else {
1326 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1327 }
1328
1329 if (C->env()->dtrace_alloc_probes() ||
1330 !UseTLAB && (!Universe::heap()->supports_inline_contig_alloc())) {
1331 // Force slow-path allocation
1332 always_slow = true;
1333 initial_slow_test = NULL;
1334 }
1335
1336
1337 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1338 Node *slow_region = NULL;
1339 Node *toobig_false = ctrl;
1340
1341 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent");
1342 // generate the initial test if necessary
1343 if (initial_slow_test != NULL ) {
1344 slow_region = new RegionNode(3);
1345
1346 // Now make the initial failure test. Usually a too-big test but
1347 // might be a TRUE for finalizers or a fancy class check for
1348 // newInstance0.
1349 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1350 transform_later(toobig_iff);
1351 // Plug the failing-too-big test into the slow-path region
1352 Node *toobig_true = new IfTrueNode( toobig_iff );
1353 transform_later(toobig_true);
1354 slow_region ->init_req( too_big_or_final_path, toobig_true );
1355 toobig_false = new IfFalseNode( toobig_iff );
1356 transform_later(toobig_false);
1357 } else { // No initial test, just fall into next case
1358 toobig_false = ctrl;
1359 debug_only(slow_region = NodeSentinel);
1360 }
1361
1362 Node *slow_mem = mem; // save the current memory state for slow path
1363 // generate the fast allocation code unless we know that the initial test will always go slow
1364 if (!always_slow) {
1365 // Fast path modifies only raw memory.
1366 if (mem->is_MergeMem()) {
1367 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1368 }
1369
1370 Node* eden_top_adr;
1371 Node* eden_end_adr;
1372
1373 set_eden_pointers(eden_top_adr, eden_end_adr);
1374
1375 // Load Eden::end. Loop invariant and hoisted.
1376 //
1377 // Note: We set the control input on "eden_end" and "old_eden_top" when using
1378 // a TLAB to work around a bug where these values were being moved across
1379 // a safepoint. These are not oops, so they cannot be include in the oop
1380 // map, but they can be changed by a GC. The proper way to fix this would
1381 // be to set the raw memory state when generating a SafepointNode. However
1382 // this will require extensive changes to the loop optimization in order to
1383 // prevent a degradation of the optimization.
1384 // See comment in memnode.hpp, around line 227 in class LoadPNode.
1385 Node *eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS);
1386
1387 // allocate the Region and Phi nodes for the result
1388 result_region = new RegionNode(3);
1389 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1390 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1391 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1392
1393 // We need a Region for the loop-back contended case.
1394 enum { fall_in_path = 1, contended_loopback_path = 2 };
1395 Node *contended_region;
1396 Node *contended_phi_rawmem;
1397 if (UseTLAB) {
1398 contended_region = toobig_false;
1399 contended_phi_rawmem = mem;
1400 } else {
1401 contended_region = new RegionNode(3);
1402 contended_phi_rawmem = new PhiNode(contended_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1403 // Now handle the passing-too-big test. We fall into the contended
1404 // loop-back merge point.
1405 contended_region ->init_req(fall_in_path, toobig_false);
1406 contended_phi_rawmem->init_req(fall_in_path, mem);
1407 transform_later(contended_region);
1408 transform_later(contended_phi_rawmem);
1409 }
1410
1411 // Load(-locked) the heap top.
1412 // See note above concerning the control input when using a TLAB
1413 Node *old_eden_top = UseTLAB
1414 ? new LoadPNode (ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered)
1415 : new LoadPLockedNode(contended_region, contended_phi_rawmem, eden_top_adr, MemNode::acquire);
1416
1417 transform_later(old_eden_top);
1418 // Add to heap top to get a new heap top
1419
1420 if (UseShenandoahGC) {
1421 // Allocate one word more for the Shenandoah brooks pointer.
1422 size_in_bytes = new AddLNode(size_in_bytes, _igvn.MakeConX(8));
1423 transform_later(size_in_bytes);
1424 }
1425
1426 Node *new_eden_top = new AddPNode(top(), old_eden_top, size_in_bytes);
1427 transform_later(new_eden_top);
1428 // Check for needing a GC; compare against heap end
1429 Node *needgc_cmp = new CmpPNode(new_eden_top, eden_end);
1430 transform_later(needgc_cmp);
1431 Node *needgc_bol = new BoolNode(needgc_cmp, BoolTest::ge);
1432 transform_later(needgc_bol);
1433 IfNode *needgc_iff = new IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN);
1434 transform_later(needgc_iff);
1435
1436 // Plug the failing-heap-space-need-gc test into the slow-path region
1437 Node *needgc_true = new IfTrueNode(needgc_iff);
1438 transform_later(needgc_true);
1439 if (initial_slow_test) {
1440 slow_region->init_req(need_gc_path, needgc_true);
1441 // This completes all paths into the slow merge point
1442 transform_later(slow_region);
1443 } else { // No initial slow path needed!
1444 // Just fall from the need-GC path straight into the VM call.
1445 slow_region = needgc_true;
1446 }
1447 // No need for a GC. Setup for the Store-Conditional
1448 Node *needgc_false = new IfFalseNode(needgc_iff);
1449 transform_later(needgc_false);
1450
1451 // Grab regular I/O before optional prefetch may change it.
1452 // Slow-path does no I/O so just set it to the original I/O.
1453 result_phi_i_o->init_req(slow_result_path, i_o);
1454
1455 i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem,
1456 old_eden_top, new_eden_top, length);
1457
1458 // Name successful fast-path variables
1459 Node* fast_oop = old_eden_top;
1460 Node* fast_oop_ctrl;
1461 Node* fast_oop_rawmem;
1462
1463 // Store (-conditional) the modified eden top back down.
1464 // StorePConditional produces flags for a test PLUS a modified raw
1465 // memory state.
1466 if (UseTLAB) {
1467 Node* store_eden_top =
1468 new StorePNode(needgc_false, contended_phi_rawmem, eden_top_adr,
1469 TypeRawPtr::BOTTOM, new_eden_top, MemNode::unordered);
1470 transform_later(store_eden_top);
1471 fast_oop_ctrl = needgc_false; // No contention, so this is the fast path
1472 fast_oop_rawmem = store_eden_top;
1473 } else {
1474 Node* store_eden_top =
1475 new StorePConditionalNode(needgc_false, contended_phi_rawmem, eden_top_adr,
1476 new_eden_top, fast_oop/*old_eden_top*/);
1477 transform_later(store_eden_top);
1478 Node *contention_check = new BoolNode(store_eden_top, BoolTest::ne);
1479 transform_later(contention_check);
1480 store_eden_top = new SCMemProjNode(store_eden_top);
1481 transform_later(store_eden_top);
1482
1483 // If not using TLABs, check to see if there was contention.
1484 IfNode *contention_iff = new IfNode (needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN);
1485 transform_later(contention_iff);
1486 Node *contention_true = new IfTrueNode(contention_iff);
1487 transform_later(contention_true);
1488 // If contention, loopback and try again.
1489 contended_region->init_req(contended_loopback_path, contention_true);
1490 contended_phi_rawmem->init_req(contended_loopback_path, store_eden_top);
1491
1492 // Fast-path succeeded with no contention!
1493 Node *contention_false = new IfFalseNode(contention_iff);
1494 transform_later(contention_false);
1495 fast_oop_ctrl = contention_false;
1496
1497 // Bump total allocated bytes for this thread
1498 Node* thread = new ThreadLocalNode();
1499 transform_later(thread);
1500 Node* alloc_bytes_adr = basic_plus_adr(top()/*not oop*/, thread,
1501 in_bytes(JavaThread::allocated_bytes_offset()));
1502 Node* alloc_bytes = make_load(fast_oop_ctrl, store_eden_top, alloc_bytes_adr,
1503 0, TypeLong::LONG, T_LONG);
1504 #ifdef _LP64
1505 Node* alloc_size = size_in_bytes;
1506 #else
1507 Node* alloc_size = new ConvI2LNode(size_in_bytes);
1508 transform_later(alloc_size);
1509 #endif
1510 Node* new_alloc_bytes = new AddLNode(alloc_bytes, alloc_size);
1511 transform_later(new_alloc_bytes);
1512 fast_oop_rawmem = make_store(fast_oop_ctrl, store_eden_top, alloc_bytes_adr,
1513 0, new_alloc_bytes, T_LONG);
1514 }
1515
1516 if (UseShenandoahGC) {
1517 // Bump up object by one word. The preceding word is used for
1518 // the Shenandoah brooks pointer.
1519 fast_oop = new AddPNode(top(), fast_oop, _igvn.MakeConX(8));
1520 transform_later(fast_oop);
1521 }
1522
1523 InitializeNode* init = alloc->initialization();
1524 fast_oop_rawmem = initialize_object(alloc,
1525 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1526 klass_node, length, size_in_bytes);
1527
1528 // If initialization is performed by an array copy, any required
1529 // MemBarStoreStore was already added. If the object does not
1530 // escape no need for a MemBarStoreStore. Otherwise we need a
1531 // MemBarStoreStore so that stores that initialize this object
1532 // can't be reordered with a subsequent store that makes this
1533 // object accessible by other threads.
1534 if (init == NULL || (!init->is_complete_with_arraycopy() && !init->does_not_escape())) {
1535 if (init == NULL || init->req() < InitializeNode::RawStores) {
1536 // No InitializeNode or no stores captured by zeroing
1537 // elimination. Simply add the MemBarStoreStore after object
1538 // initialization.
1539 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1540 transform_later(mb);
1541
1542 mb->init_req(TypeFunc::Memory, fast_oop_rawmem);
1543 mb->init_req(TypeFunc::Control, fast_oop_ctrl);
1544 fast_oop_ctrl = new ProjNode(mb,TypeFunc::Control);
1545 transform_later(fast_oop_ctrl);
1546 fast_oop_rawmem = new ProjNode(mb,TypeFunc::Memory);
1547 transform_later(fast_oop_rawmem);
1548 } else {
1549 // Add the MemBarStoreStore after the InitializeNode so that
1550 // all stores performing the initialization that were moved
1551 // before the InitializeNode happen before the storestore
1552 // barrier.
1553
1554 Node* init_ctrl = init->proj_out(TypeFunc::Control);
1555 Node* init_mem = init->proj_out(TypeFunc::Memory);
1556
1557 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1558 transform_later(mb);
1559
1560 Node* ctrl = new ProjNode(init,TypeFunc::Control);
1561 transform_later(ctrl);
1562 Node* mem = new ProjNode(init,TypeFunc::Memory);
1563 transform_later(mem);
1564
1565 // The MemBarStoreStore depends on control and memory coming
1566 // from the InitializeNode
1567 mb->init_req(TypeFunc::Memory, mem);
1568 mb->init_req(TypeFunc::Control, ctrl);
1569
1570 ctrl = new ProjNode(mb,TypeFunc::Control);
1571 transform_later(ctrl);
1572 mem = new ProjNode(mb,TypeFunc::Memory);
1573 transform_later(mem);
1574
1575 // All nodes that depended on the InitializeNode for control
1576 // and memory must now depend on the MemBarNode that itself
1577 // depends on the InitializeNode
1578 _igvn.replace_node(init_ctrl, ctrl);
1579 _igvn.replace_node(init_mem, mem);
1580 }
1581 }
1582
1583 if (C->env()->dtrace_extended_probes()) {
1584 // Slow-path call
1585 int size = TypeFunc::Parms + 2;
1586 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1587 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base),
1588 "dtrace_object_alloc",
1589 TypeRawPtr::BOTTOM);
1590
1591 // Get base of thread-local storage area
1592 Node* thread = new ThreadLocalNode();
1593 transform_later(thread);
1594
1595 call->init_req(TypeFunc::Parms+0, thread);
1596 call->init_req(TypeFunc::Parms+1, fast_oop);
1597 call->init_req(TypeFunc::Control, fast_oop_ctrl);
1598 call->init_req(TypeFunc::I_O , top()); // does no i/o
1599 call->init_req(TypeFunc::Memory , fast_oop_rawmem);
1600 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1601 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1602 transform_later(call);
1603 fast_oop_ctrl = new ProjNode(call,TypeFunc::Control);
1604 transform_later(fast_oop_ctrl);
1605 fast_oop_rawmem = new ProjNode(call,TypeFunc::Memory);
1606 transform_later(fast_oop_rawmem);
1607 }
1608
1609 // Plug in the successful fast-path into the result merge point
1610 result_region ->init_req(fast_result_path, fast_oop_ctrl);
1611 result_phi_rawoop->init_req(fast_result_path, fast_oop);
1612 result_phi_i_o ->init_req(fast_result_path, i_o);
1613 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1614 } else {
1615 slow_region = ctrl;
1616 result_phi_i_o = i_o; // Rename it to use in the following code.
1617 }
1618
1619 // Generate slow-path call
1620 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1621 OptoRuntime::stub_name(slow_call_address),
1622 alloc->jvms()->bci(),
1623 TypePtr::BOTTOM);
1624 call->init_req( TypeFunc::Control, slow_region );
1625 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o
1626 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs
1627 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) );
1628 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) );
1629
1630 call->init_req(TypeFunc::Parms+0, klass_node);
1631 if (length != NULL) {
1632 call->init_req(TypeFunc::Parms+1, length);
1633 }
1634
1635 // Copy debug information and adjust JVMState information, then replace
1636 // allocate node with the call
1637 copy_call_debug_info((CallNode *) alloc, call);
1638 if (!always_slow) {
1639 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1640 } else {
1641 // Hook i_o projection to avoid its elimination during allocation
1642 // replacement (when only a slow call is generated).
1643 call->set_req(TypeFunc::I_O, result_phi_i_o);
1644 }
1645 _igvn.replace_node(alloc, call);
1646 transform_later(call);
1647
1648 // Identify the output projections from the allocate node and
1649 // adjust any references to them.
1650 // The control and io projections look like:
1651 //
1652 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1653 // Allocate Catch
1654 // ^---Proj(io) <-------+ ^---CatchProj(io)
1655 //
1656 // We are interested in the CatchProj nodes.
1657 //
1658 extract_call_projections(call);
1659
1660 // An allocate node has separate memory projections for the uses on
1661 // the control and i_o paths. Replace the control memory projection with
1662 // result_phi_rawmem (unless we are only generating a slow call when
1663 // both memory projections are combined)
1664 if (!always_slow && _memproj_fallthrough != NULL) {
1665 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) {
1666 Node *use = _memproj_fallthrough->fast_out(i);
1667 _igvn.rehash_node_delayed(use);
1668 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem);
1669 // back up iterator
1670 --i;
1671 }
1672 }
1673 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete
1674 // _memproj_catchall so we end up with a call that has only 1 memory projection.
1675 if (_memproj_catchall != NULL ) {
1676 if (_memproj_fallthrough == NULL) {
1677 _memproj_fallthrough = new ProjNode(call, TypeFunc::Memory);
1678 transform_later(_memproj_fallthrough);
1679 }
1680 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) {
1681 Node *use = _memproj_catchall->fast_out(i);
1682 _igvn.rehash_node_delayed(use);
1683 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough);
1684 // back up iterator
1685 --i;
1686 }
1687 assert(_memproj_catchall->outcnt() == 0, "all uses must be deleted");
1688 _igvn.remove_dead_node(_memproj_catchall);
1689 }
1690
1691 // An allocate node has separate i_o projections for the uses on the control
1692 // and i_o paths. Always replace the control i_o projection with result i_o
1693 // otherwise incoming i_o become dead when only a slow call is generated
1694 // (it is different from memory projections where both projections are
1695 // combined in such case).
1696 if (_ioproj_fallthrough != NULL) {
1697 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) {
1698 Node *use = _ioproj_fallthrough->fast_out(i);
1699 _igvn.rehash_node_delayed(use);
1700 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o);
1701 // back up iterator
1702 --i;
1703 }
1704 }
1705 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete
1706 // _ioproj_catchall so we end up with a call that has only 1 i_o projection.
1707 if (_ioproj_catchall != NULL ) {
1708 if (_ioproj_fallthrough == NULL) {
1709 _ioproj_fallthrough = new ProjNode(call, TypeFunc::I_O);
1710 transform_later(_ioproj_fallthrough);
1711 }
1712 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) {
1713 Node *use = _ioproj_catchall->fast_out(i);
1714 _igvn.rehash_node_delayed(use);
1715 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough);
1716 // back up iterator
1717 --i;
1718 }
1719 assert(_ioproj_catchall->outcnt() == 0, "all uses must be deleted");
1720 _igvn.remove_dead_node(_ioproj_catchall);
1721 }
1722
1723 // if we generated only a slow call, we are done
1724 if (always_slow) {
1725 // Now we can unhook i_o.
1726 if (result_phi_i_o->outcnt() > 1) {
1727 call->set_req(TypeFunc::I_O, top());
1728 } else {
1729 assert(result_phi_i_o->unique_ctrl_out() == call, "");
1730 // Case of new array with negative size known during compilation.
1731 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1732 // following code since call to runtime will throw exception.
1733 // As result there will be no users of i_o after the call.
1734 // Leave i_o attached to this call to avoid problems in preceding graph.
1735 }
1736 return;
1737 }
1738
1739
1740 if (_fallthroughcatchproj != NULL) {
1741 ctrl = _fallthroughcatchproj->clone();
1742 transform_later(ctrl);
1743 _igvn.replace_node(_fallthroughcatchproj, result_region);
1744 } else {
1745 ctrl = top();
1746 }
1747 Node *slow_result;
1748 if (_resproj == NULL) {
1749 // no uses of the allocation result
1750 slow_result = top();
1751 } else {
1752 slow_result = _resproj->clone();
1753 transform_later(slow_result);
1754 _igvn.replace_node(_resproj, result_phi_rawoop);
1755 }
1756
1757 // Plug slow-path into result merge point
1758 result_region ->init_req( slow_result_path, ctrl );
1759 result_phi_rawoop->init_req( slow_result_path, slow_result);
1760 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough );
1761 transform_later(result_region);
1762 transform_later(result_phi_rawoop);
1763 transform_later(result_phi_rawmem);
1764 transform_later(result_phi_i_o);
1765 // This completes all paths into the result merge point
1766 }
1767
1768
1769 // Helper for PhaseMacroExpand::expand_allocate_common.
1770 // Initializes the newly-allocated storage.
1771 Node*
1772 PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1773 Node* control, Node* rawmem, Node* object,
1774 Node* klass_node, Node* length,
1775 Node* size_in_bytes) {
1776 InitializeNode* init = alloc->initialization();
1777 // Store the klass & mark bits
1778 Node* mark_node = NULL;
1779 // For now only enable fast locking for non-array types
1780 if (UseBiasedLocking && (length == NULL)) {
1781 mark_node = make_load(control, rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
1782 } else {
1783 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype()));
1784 }
1785 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
1786
1787 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1788 int header_size = alloc->minimum_header_size(); // conservatively small
1789
1790 // Array length
1791 if (length != NULL) { // Arrays need length field
1792 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1793 // conservatively small header size:
1794 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1795 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1796 if (k->is_array_klass()) // we know the exact header size in most cases:
1797 header_size = Klass::layout_helper_header_size(k->layout_helper());
1798 }
1799
1800 if (UseShenandoahGC) {
1801 // Initialize Shenandoah brooks pointer to point to the object itself.
1802 rawmem = make_store(control, rawmem, object, BrooksPointer::BYTE_OFFSET, object, T_OBJECT);
1803 }
1804
1805 // Clear the object body, if necessary.
1806 if (init == NULL) {
1807 // The init has somehow disappeared; be cautious and clear everything.
1808 //
1809 // This can happen if a node is allocated but an uncommon trap occurs
1810 // immediately. In this case, the Initialize gets associated with the
1811 // trap, and may be placed in a different (outer) loop, if the Allocate
1812 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1813 // there can be two Allocates to one Initialize. The answer in all these
1814 // edge cases is safety first. It is always safe to clear immediately
1815 // within an Allocate, and then (maybe or maybe not) clear some more later.
1816 if (!ZeroTLAB)
1817 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1818 header_size, size_in_bytes,
1819 &_igvn);
1820 } else {
1821 if (!init->is_complete()) {
1822 // Try to win by zeroing only what the init does not store.
1823 // We can also try to do some peephole optimizations,
1824 // such as combining some adjacent subword stores.
1825 rawmem = init->complete_stores(control, rawmem, object,
1826 header_size, size_in_bytes, &_igvn);
1827 }
1828 // We have no more use for this link, since the AllocateNode goes away:
1829 init->set_req(InitializeNode::RawAddress, top());
1830 // (If we keep the link, it just confuses the register allocator,
1831 // who thinks he sees a real use of the address by the membar.)
1832 }
1833
1834 return rawmem;
1835 }
1836
1837 // Generate prefetch instructions for next allocations.
1838 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
1839 Node*& contended_phi_rawmem,
1840 Node* old_eden_top, Node* new_eden_top,
1841 Node* length) {
1842 enum { fall_in_path = 1, pf_path = 2 };
1843 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
1844 // Generate prefetch allocation with watermark check.
1845 // As an allocation hits the watermark, we will prefetch starting
1846 // at a "distance" away from watermark.
1847
1848 Node *pf_region = new RegionNode(3);
1849 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY,
1850 TypeRawPtr::BOTTOM );
1851 // I/O is used for Prefetch
1852 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO );
1853
1854 Node *thread = new ThreadLocalNode();
1855 transform_later(thread);
1856
1857 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread,
1858 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
1859 transform_later(eden_pf_adr);
1860
1861 Node *old_pf_wm = new LoadPNode(needgc_false,
1862 contended_phi_rawmem, eden_pf_adr,
1863 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM,
1864 MemNode::unordered);
1865 transform_later(old_pf_wm);
1866
1867 // check against new_eden_top
1868 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm );
1869 transform_later(need_pf_cmp);
1870 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge );
1871 transform_later(need_pf_bol);
1872 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol,
1873 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1874 transform_later(need_pf_iff);
1875
1876 // true node, add prefetchdistance
1877 Node *need_pf_true = new IfTrueNode( need_pf_iff );
1878 transform_later(need_pf_true);
1879
1880 Node *need_pf_false = new IfFalseNode( need_pf_iff );
1881 transform_later(need_pf_false);
1882
1883 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm,
1884 _igvn.MakeConX(AllocatePrefetchDistance) );
1885 transform_later(new_pf_wmt );
1886 new_pf_wmt->set_req(0, need_pf_true);
1887
1888 Node *store_new_wmt = new StorePNode(need_pf_true,
1889 contended_phi_rawmem, eden_pf_adr,
1890 TypeRawPtr::BOTTOM, new_pf_wmt,
1891 MemNode::unordered);
1892 transform_later(store_new_wmt);
1893
1894 // adding prefetches
1895 pf_phi_abio->init_req( fall_in_path, i_o );
1896
1897 Node *prefetch_adr;
1898 Node *prefetch;
1899 uint lines = AllocatePrefetchDistance / AllocatePrefetchStepSize;
1900 uint step_size = AllocatePrefetchStepSize;
1901 uint distance = 0;
1902
1903 for ( uint i = 0; i < lines; i++ ) {
1904 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt,
1905 _igvn.MakeConX(distance) );
1906 transform_later(prefetch_adr);
1907 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1908 transform_later(prefetch);
1909 distance += step_size;
1910 i_o = prefetch;
1911 }
1912 pf_phi_abio->set_req( pf_path, i_o );
1913
1914 pf_region->init_req( fall_in_path, need_pf_false );
1915 pf_region->init_req( pf_path, need_pf_true );
1916
1917 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
1918 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
1919
1920 transform_later(pf_region);
1921 transform_later(pf_phi_rawmem);
1922 transform_later(pf_phi_abio);
1923
1924 needgc_false = pf_region;
1925 contended_phi_rawmem = pf_phi_rawmem;
1926 i_o = pf_phi_abio;
1927 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) {
1928 // Insert a prefetch for each allocation.
1929 // This code is used for Sparc with BIS.
1930 Node *pf_region = new RegionNode(3);
1931 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY,
1932 TypeRawPtr::BOTTOM );
1933 transform_later(pf_region);
1934
1935 // Generate several prefetch instructions.
1936 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1937 uint step_size = AllocatePrefetchStepSize;
1938 uint distance = AllocatePrefetchDistance;
1939
1940 // Next cache address.
1941 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top,
1942 _igvn.MakeConX(distance));
1943 transform_later(cache_adr);
1944 cache_adr = new CastP2XNode(needgc_false, cache_adr);
1945 transform_later(cache_adr);
1946 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1));
1947 cache_adr = new AndXNode(cache_adr, mask);
1948 transform_later(cache_adr);
1949 cache_adr = new CastX2PNode(cache_adr);
1950 transform_later(cache_adr);
1951
1952 // Prefetch
1953 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr );
1954 prefetch->set_req(0, needgc_false);
1955 transform_later(prefetch);
1956 contended_phi_rawmem = prefetch;
1957 Node *prefetch_adr;
1958 distance = step_size;
1959 for ( uint i = 1; i < lines; i++ ) {
1960 prefetch_adr = new AddPNode( cache_adr, cache_adr,
1961 _igvn.MakeConX(distance) );
1962 transform_later(prefetch_adr);
1963 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr );
1964 transform_later(prefetch);
1965 distance += step_size;
1966 contended_phi_rawmem = prefetch;
1967 }
1968 } else if( AllocatePrefetchStyle > 0 ) {
1969 // Insert a prefetch for each allocation only on the fast-path
1970 Node *prefetch_adr;
1971 Node *prefetch;
1972 // Generate several prefetch instructions.
1973 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1974 uint step_size = AllocatePrefetchStepSize;
1975 uint distance = AllocatePrefetchDistance;
1976 for ( uint i = 0; i < lines; i++ ) {
1977 prefetch_adr = new AddPNode( old_eden_top, new_eden_top,
1978 _igvn.MakeConX(distance) );
1979 transform_later(prefetch_adr);
1980 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1981 // Do not let it float too high, since if eden_top == eden_end,
1982 // both might be null.
1983 if( i == 0 ) { // Set control for first prefetch, next follows it
1984 prefetch->init_req(0, needgc_false);
1985 }
1986 transform_later(prefetch);
1987 distance += step_size;
1988 i_o = prefetch;
1989 }
1990 }
1991 return i_o;
1992 }
1993
1994
1995 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
1996 expand_allocate_common(alloc, NULL,
1997 OptoRuntime::new_instance_Type(),
1998 OptoRuntime::new_instance_Java());
1999 }
2000
2001 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
2002 Node* length = alloc->in(AllocateNode::ALength);
2003 InitializeNode* init = alloc->initialization();
2004 Node* klass_node = alloc->in(AllocateNode::KlassNode);
2005 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
2006 address slow_call_address; // Address of slow call
2007 if (init != NULL && init->is_complete_with_arraycopy() &&
2008 k->is_type_array_klass()) {
2009 // Don't zero type array during slow allocation in VM since
2010 // it will be initialized later by arraycopy in compiled code.
2011 slow_call_address = OptoRuntime::new_array_nozero_Java();
2012 } else {
2013 slow_call_address = OptoRuntime::new_array_Java();
2014 }
2015 expand_allocate_common(alloc, length,
2016 OptoRuntime::new_array_Type(),
2017 slow_call_address);
2018 }
2019
2020 //-------------------mark_eliminated_box----------------------------------
2021 //
2022 // During EA obj may point to several objects but after few ideal graph
2023 // transformations (CCP) it may point to only one non escaping object
2024 // (but still using phi), corresponding locks and unlocks will be marked
2025 // for elimination. Later obj could be replaced with a new node (new phi)
2026 // and which does not have escape information. And later after some graph
2027 // reshape other locks and unlocks (which were not marked for elimination
2028 // before) are connected to this new obj (phi) but they still will not be
2029 // marked for elimination since new obj has no escape information.
2030 // Mark all associated (same box and obj) lock and unlock nodes for
2031 // elimination if some of them marked already.
2032 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) {
2033 if (oldbox->as_BoxLock()->is_eliminated())
2034 return; // This BoxLock node was processed already.
2035
2036 // New implementation (EliminateNestedLocks) has separate BoxLock
2037 // node for each locked region so mark all associated locks/unlocks as
2038 // eliminated even if different objects are referenced in one locked region
2039 // (for example, OSR compilation of nested loop inside locked scope).
2040 if (EliminateNestedLocks ||
2041 oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj)) {
2042 // Box is used only in one lock region. Mark this box as eliminated.
2043 _igvn.hash_delete(oldbox);
2044 oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value
2045 _igvn.hash_insert(oldbox);
2046
2047 for (uint i = 0; i < oldbox->outcnt(); i++) {
2048 Node* u = oldbox->raw_out(i);
2049 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) {
2050 AbstractLockNode* alock = u->as_AbstractLock();
2051 // Check lock's box since box could be referenced by Lock's debug info.
2052 if (alock->box_node() == oldbox) {
2053 // Mark eliminated all related locks and unlocks.
2054 #ifdef ASSERT
2055 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4");
2056 #endif
2057 alock->set_non_esc_obj();
2058 }
2059 }
2060 }
2061 return;
2062 }
2063
2064 // Create new "eliminated" BoxLock node and use it in monitor debug info
2065 // instead of oldbox for the same object.
2066 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
2067
2068 // Note: BoxLock node is marked eliminated only here and it is used
2069 // to indicate that all associated lock and unlock nodes are marked
2070 // for elimination.
2071 newbox->set_eliminated();
2072 transform_later(newbox);
2073
2074 // Replace old box node with new box for all users of the same object.
2075 for (uint i = 0; i < oldbox->outcnt();) {
2076 bool next_edge = true;
2077
2078 Node* u = oldbox->raw_out(i);
2079 if (u->is_AbstractLock()) {
2080 AbstractLockNode* alock = u->as_AbstractLock();
2081 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) {
2082 // Replace Box and mark eliminated all related locks and unlocks.
2083 #ifdef ASSERT
2084 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5");
2085 #endif
2086 alock->set_non_esc_obj();
2087 _igvn.rehash_node_delayed(alock);
2088 alock->set_box_node(newbox);
2089 next_edge = false;
2090 }
2091 }
2092 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) {
2093 FastLockNode* flock = u->as_FastLock();
2094 assert(flock->box_node() == oldbox, "sanity");
2095 _igvn.rehash_node_delayed(flock);
2096 flock->set_box_node(newbox);
2097 next_edge = false;
2098 }
2099
2100 // Replace old box in monitor debug info.
2101 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
2102 SafePointNode* sfn = u->as_SafePoint();
2103 JVMState* youngest_jvms = sfn->jvms();
2104 int max_depth = youngest_jvms->depth();
2105 for (int depth = 1; depth <= max_depth; depth++) {
2106 JVMState* jvms = youngest_jvms->of_depth(depth);
2107 int num_mon = jvms->nof_monitors();
2108 // Loop over monitors
2109 for (int idx = 0; idx < num_mon; idx++) {
2110 Node* obj_node = sfn->monitor_obj(jvms, idx);
2111 Node* box_node = sfn->monitor_box(jvms, idx);
2112 if (box_node == oldbox && obj_node->eqv_uncast(obj)) {
2113 int j = jvms->monitor_box_offset(idx);
2114 _igvn.replace_input_of(u, j, newbox);
2115 next_edge = false;
2116 }
2117 }
2118 }
2119 }
2120 if (next_edge) i++;
2121 }
2122 }
2123
2124 //-----------------------mark_eliminated_locking_nodes-----------------------
2125 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) {
2126 if (EliminateNestedLocks) {
2127 if (alock->is_nested()) {
2128 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity");
2129 return;
2130 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened
2131 // Only Lock node has JVMState needed here.
2132 // Not that preceding claim is documented anywhere else.
2133 if (alock->jvms() != NULL) {
2134 if (alock->as_Lock()->is_nested_lock_region()) {
2135 // Mark eliminated related nested locks and unlocks.
2136 Node* obj = alock->obj_node();
2137 BoxLockNode* box_node = alock->box_node()->as_BoxLock();
2138 assert(!box_node->is_eliminated(), "should not be marked yet");
2139 // Note: BoxLock node is marked eliminated only here
2140 // and it is used to indicate that all associated lock
2141 // and unlock nodes are marked for elimination.
2142 box_node->set_eliminated(); // Box's hash is always NO_HASH here
2143 for (uint i = 0; i < box_node->outcnt(); i++) {
2144 Node* u = box_node->raw_out(i);
2145 if (u->is_AbstractLock()) {
2146 alock = u->as_AbstractLock();
2147 if (alock->box_node() == box_node) {
2148 // Verify that this Box is referenced only by related locks.
2149 assert(alock->obj_node()->eqv_uncast(obj), "");
2150 // Mark all related locks and unlocks.
2151 #ifdef ASSERT
2152 alock->log_lock_optimization(C, "eliminate_lock_set_nested");
2153 #endif
2154 alock->set_nested();
2155 }
2156 }
2157 }
2158 } else {
2159 #ifdef ASSERT
2160 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region");
2161 if (C->log() != NULL)
2162 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output
2163 #endif
2164 }
2165 }
2166 return;
2167 }
2168 // Process locks for non escaping object
2169 assert(alock->is_non_esc_obj(), "");
2170 } // EliminateNestedLocks
2171
2172 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2173 // Look for all locks of this object and mark them and
2174 // corresponding BoxLock nodes as eliminated.
2175 Node* obj = alock->obj_node();
2176 for (uint j = 0; j < obj->outcnt(); j++) {
2177 Node* o = obj->raw_out(j);
2178 if (o->is_AbstractLock() &&
2179 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2180 alock = o->as_AbstractLock();
2181 Node* box = alock->box_node();
2182 // Replace old box node with new eliminated box for all users
2183 // of the same object and mark related locks as eliminated.
2184 mark_eliminated_box(box, obj);
2185 }
2186 }
2187 }
2188 }
2189
2190 // we have determined that this lock/unlock can be eliminated, we simply
2191 // eliminate the node without expanding it.
2192 //
2193 // Note: The membar's associated with the lock/unlock are currently not
2194 // eliminated. This should be investigated as a future enhancement.
2195 //
2196 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2197
2198 if (!alock->is_eliminated()) {
2199 return false;
2200 }
2201 #ifdef ASSERT
2202 if (!alock->is_coarsened()) {
2203 // Check that new "eliminated" BoxLock node is created.
2204 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2205 assert(oldbox->is_eliminated(), "should be done already");
2206 }
2207 #endif
2208
2209 alock->log_lock_optimization(C, "eliminate_lock");
2210
2211 #ifndef PRODUCT
2212 if (PrintEliminateLocks) {
2213 if (alock->is_Lock()) {
2214 tty->print_cr("++++ Eliminated: %d Lock", alock->_idx);
2215 } else {
2216 tty->print_cr("++++ Eliminated: %d Unlock", alock->_idx);
2217 }
2218 }
2219 #endif
2220
2221 Node* mem = alock->in(TypeFunc::Memory);
2222 Node* ctrl = alock->in(TypeFunc::Control);
2223
2224 extract_call_projections(alock);
2225 // There are 2 projections from the lock. The lock node will
2226 // be deleted when its last use is subsumed below.
2227 assert(alock->outcnt() == 2 &&
2228 _fallthroughproj != NULL &&
2229 _memproj_fallthrough != NULL,
2230 "Unexpected projections from Lock/Unlock");
2231
2232 Node* fallthroughproj = _fallthroughproj;
2233 Node* memproj_fallthrough = _memproj_fallthrough;
2234
2235 // The memory projection from a lock/unlock is RawMem
2236 // The input to a Lock is merged memory, so extract its RawMem input
2237 // (unless the MergeMem has been optimized away.)
2238 if (alock->is_Lock()) {
2239 // Seach for MemBarAcquireLock node and delete it also.
2240 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2241 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, "");
2242 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2243 Node* memproj = membar->proj_out(TypeFunc::Memory);
2244 _igvn.replace_node(ctrlproj, fallthroughproj);
2245 _igvn.replace_node(memproj, memproj_fallthrough);
2246
2247 // Delete FastLock node also if this Lock node is unique user
2248 // (a loop peeling may clone a Lock node).
2249 Node* flock = alock->as_Lock()->fastlock_node();
2250 if (flock->outcnt() == 1) {
2251 assert(flock->unique_out() == alock, "sanity");
2252 _igvn.replace_node(flock, top());
2253 }
2254 }
2255
2256 // Seach for MemBarReleaseLock node and delete it also.
2257 if (alock->is_Unlock() && ctrl != NULL && ctrl->is_Proj() &&
2258 ctrl->in(0)->is_MemBar()) {
2259 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2260 assert(membar->Opcode() == Op_MemBarReleaseLock &&
2261 mem->is_Proj() && membar == mem->in(0), "");
2262 _igvn.replace_node(fallthroughproj, ctrl);
2263 _igvn.replace_node(memproj_fallthrough, mem);
2264 fallthroughproj = ctrl;
2265 memproj_fallthrough = mem;
2266 ctrl = membar->in(TypeFunc::Control);
2267 mem = membar->in(TypeFunc::Memory);
2268 }
2269
2270 _igvn.replace_node(fallthroughproj, ctrl);
2271 _igvn.replace_node(memproj_fallthrough, mem);
2272 return true;
2273 }
2274
2275
2276 //------------------------------expand_lock_node----------------------
2277 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
2278
2279 Node* ctrl = lock->in(TypeFunc::Control);
2280 Node* mem = lock->in(TypeFunc::Memory);
2281 Node* obj = lock->obj_node();
2282 Node* box = lock->box_node();
2283 Node* flock = lock->fastlock_node();
2284
2285 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2286
2287 // Make the merge point
2288 Node *region;
2289 Node *mem_phi;
2290 Node *slow_path;
2291
2292 if (UseOptoBiasInlining) {
2293 /*
2294 * See the full description in MacroAssembler::biased_locking_enter().
2295 *
2296 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) {
2297 * // The object is biased.
2298 * proto_node = klass->prototype_header;
2299 * o_node = thread | proto_node;
2300 * x_node = o_node ^ mark_word;
2301 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ?
2302 * // Done.
2303 * } else {
2304 * if( (x_node & biased_lock_mask) != 0 ) {
2305 * // The klass's prototype header is no longer biased.
2306 * cas(&mark_word, mark_word, proto_node)
2307 * goto cas_lock;
2308 * } else {
2309 * // The klass's prototype header is still biased.
2310 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch?
2311 * old = mark_word;
2312 * new = o_node;
2313 * } else {
2314 * // Different thread or anonymous biased.
2315 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask);
2316 * new = thread | old;
2317 * }
2318 * // Try to rebias.
2319 * if( cas(&mark_word, old, new) == 0 ) {
2320 * // Done.
2321 * } else {
2322 * goto slow_path; // Failed.
2323 * }
2324 * }
2325 * }
2326 * } else {
2327 * // The object is not biased.
2328 * cas_lock:
2329 * if( FastLock(obj) == 0 ) {
2330 * // Done.
2331 * } else {
2332 * slow_path:
2333 * OptoRuntime::complete_monitor_locking_Java(obj);
2334 * }
2335 * }
2336 */
2337
2338 region = new RegionNode(5);
2339 // create a Phi for the memory state
2340 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2341
2342 Node* fast_lock_region = new RegionNode(3);
2343 Node* fast_lock_mem_phi = new PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM);
2344
2345 // First, check mark word for the biased lock pattern.
2346 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2347
2348 // Get fast path - mark word has the biased lock pattern.
2349 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node,
2350 markOopDesc::biased_lock_mask_in_place,
2351 markOopDesc::biased_lock_pattern, true);
2352 // fast_lock_region->in(1) is set to slow path.
2353 fast_lock_mem_phi->init_req(1, mem);
2354
2355 // Now check that the lock is biased to the current thread and has
2356 // the same epoch and bias as Klass::_prototype_header.
2357
2358 // Special-case a fresh allocation to avoid building nodes:
2359 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn);
2360 if (klass_node == NULL) {
2361 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
2362 klass_node = transform_later(LoadKlassNode::make(_igvn, NULL, mem, k_adr, _igvn.type(k_adr)->is_ptr()));
2363 #ifdef _LP64
2364 if (UseCompressedClassPointers && klass_node->is_DecodeNKlass()) {
2365 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity");
2366 klass_node->in(1)->init_req(0, ctrl);
2367 } else
2368 #endif
2369 klass_node->init_req(0, ctrl);
2370 }
2371 Node *proto_node = make_load(ctrl, mem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeX_X, TypeX_X->basic_type());
2372
2373 Node* thread = transform_later(new ThreadLocalNode());
2374 Node* cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2375 Node* o_node = transform_later(new OrXNode(cast_thread, proto_node));
2376 Node* x_node = transform_later(new XorXNode(o_node, mark_node));
2377
2378 // Get slow path - mark word does NOT match the value.
2379 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node,
2380 (~markOopDesc::age_mask_in_place), 0);
2381 // region->in(3) is set to fast path - the object is biased to the current thread.
2382 mem_phi->init_req(3, mem);
2383
2384
2385 // Mark word does NOT match the value (thread | Klass::_prototype_header).
2386
2387
2388 // First, check biased pattern.
2389 // Get fast path - _prototype_header has the same biased lock pattern.
2390 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node,
2391 markOopDesc::biased_lock_mask_in_place, 0, true);
2392
2393 not_biased_ctrl = fast_lock_region->in(2); // Slow path
2394 // fast_lock_region->in(2) - the prototype header is no longer biased
2395 // and we have to revoke the bias on this object.
2396 // We are going to try to reset the mark of this object to the prototype
2397 // value and fall through to the CAS-based locking scheme.
2398 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
2399 Node* cas = new StoreXConditionalNode(not_biased_ctrl, mem, adr,
2400 proto_node, mark_node);
2401 transform_later(cas);
2402 Node* proj = transform_later(new SCMemProjNode(cas));
2403 fast_lock_mem_phi->init_req(2, proj);
2404
2405
2406 // Second, check epoch bits.
2407 Node* rebiased_region = new RegionNode(3);
2408 Node* old_phi = new PhiNode( rebiased_region, TypeX_X);
2409 Node* new_phi = new PhiNode( rebiased_region, TypeX_X);
2410
2411 // Get slow path - mark word does NOT match epoch bits.
2412 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node,
2413 markOopDesc::epoch_mask_in_place, 0);
2414 // The epoch of the current bias is not valid, attempt to rebias the object
2415 // toward the current thread.
2416 rebiased_region->init_req(2, epoch_ctrl);
2417 old_phi->init_req(2, mark_node);
2418 new_phi->init_req(2, o_node);
2419
2420 // rebiased_region->in(1) is set to fast path.
2421 // The epoch of the current bias is still valid but we know
2422 // nothing about the owner; it might be set or it might be clear.
2423 Node* cmask = MakeConX(markOopDesc::biased_lock_mask_in_place |
2424 markOopDesc::age_mask_in_place |
2425 markOopDesc::epoch_mask_in_place);
2426 Node* old = transform_later(new AndXNode(mark_node, cmask));
2427 cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2428 Node* new_mark = transform_later(new OrXNode(cast_thread, old));
2429 old_phi->init_req(1, old);
2430 new_phi->init_req(1, new_mark);
2431
2432 transform_later(rebiased_region);
2433 transform_later(old_phi);
2434 transform_later(new_phi);
2435
2436 // Try to acquire the bias of the object using an atomic operation.
2437 // If this fails we will go in to the runtime to revoke the object's bias.
2438 cas = new StoreXConditionalNode(rebiased_region, mem, adr, new_phi, old_phi);
2439 transform_later(cas);
2440 proj = transform_later(new SCMemProjNode(cas));
2441
2442 // Get slow path - Failed to CAS.
2443 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0);
2444 mem_phi->init_req(4, proj);
2445 // region->in(4) is set to fast path - the object is rebiased to the current thread.
2446
2447 // Failed to CAS.
2448 slow_path = new RegionNode(3);
2449 Node *slow_mem = new PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM);
2450
2451 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control
2452 slow_mem->init_req(1, proj);
2453
2454 // Call CAS-based locking scheme (FastLock node).
2455
2456 transform_later(fast_lock_region);
2457 transform_later(fast_lock_mem_phi);
2458
2459 // Get slow path - FastLock failed to lock the object.
2460 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0);
2461 mem_phi->init_req(2, fast_lock_mem_phi);
2462 // region->in(2) is set to fast path - the object is locked to the current thread.
2463
2464 slow_path->init_req(2, ctrl); // Capture slow-control
2465 slow_mem->init_req(2, fast_lock_mem_phi);
2466
2467 transform_later(slow_path);
2468 transform_later(slow_mem);
2469 // Reset lock's memory edge.
2470 lock->set_req(TypeFunc::Memory, slow_mem);
2471
2472 } else {
2473 region = new RegionNode(3);
2474 // create a Phi for the memory state
2475 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2476
2477 // Optimize test; set region slot 2
2478 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2479 mem_phi->init_req(2, mem);
2480 }
2481
2482 // Make slow path call
2483 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(),
2484 OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path,
2485 obj, box, NULL);
2486
2487 extract_call_projections(call);
2488
2489 // Slow path can only throw asynchronous exceptions, which are always
2490 // de-opted. So the compiler thinks the slow-call can never throw an
2491 // exception. If it DOES throw an exception we would need the debug
2492 // info removed first (since if it throws there is no monitor).
2493 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2494 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2495
2496 // Capture slow path
2497 // disconnect fall-through projection from call and create a new one
2498 // hook up users of fall-through projection to region
2499 Node *slow_ctrl = _fallthroughproj->clone();
2500 transform_later(slow_ctrl);
2501 _igvn.hash_delete(_fallthroughproj);
2502 _fallthroughproj->disconnect_inputs(NULL, C);
2503 region->init_req(1, slow_ctrl);
2504 // region inputs are now complete
2505 transform_later(region);
2506 _igvn.replace_node(_fallthroughproj, region);
2507
2508 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2509 mem_phi->init_req(1, memproj );
2510 transform_later(mem_phi);
2511 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2512 }
2513
2514 //------------------------------expand_unlock_node----------------------
2515 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2516
2517 Node* ctrl = unlock->in(TypeFunc::Control);
2518 Node* mem = unlock->in(TypeFunc::Memory);
2519 Node* obj = unlock->obj_node();
2520 Node* box = unlock->box_node();
2521
2522 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2523
2524 // No need for a null check on unlock
2525
2526 // Make the merge point
2527 Node *region;
2528 Node *mem_phi;
2529
2530 if (UseOptoBiasInlining) {
2531 // Check for biased locking unlock case, which is a no-op.
2532 // See the full description in MacroAssembler::biased_locking_exit().
2533 region = new RegionNode(4);
2534 // create a Phi for the memory state
2535 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2536 mem_phi->init_req(3, mem);
2537
2538 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2539 ctrl = opt_bits_test(ctrl, region, 3, mark_node,
2540 markOopDesc::biased_lock_mask_in_place,
2541 markOopDesc::biased_lock_pattern);
2542 } else {
2543 region = new RegionNode(3);
2544 // create a Phi for the memory state
2545 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2546 }
2547
2548 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2549 funlock = transform_later( funlock )->as_FastUnlock();
2550 // Optimize test; set region slot 2
2551 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2552 Node *thread = transform_later(new ThreadLocalNode());
2553
2554 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2555 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2556 "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2557
2558 extract_call_projections(call);
2559
2560 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2561 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2562
2563 // No exceptions for unlocking
2564 // Capture slow path
2565 // disconnect fall-through projection from call and create a new one
2566 // hook up users of fall-through projection to region
2567 Node *slow_ctrl = _fallthroughproj->clone();
2568 transform_later(slow_ctrl);
2569 _igvn.hash_delete(_fallthroughproj);
2570 _fallthroughproj->disconnect_inputs(NULL, C);
2571 region->init_req(1, slow_ctrl);
2572 // region inputs are now complete
2573 transform_later(region);
2574 _igvn.replace_node(_fallthroughproj, region);
2575
2576 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2577 mem_phi->init_req(1, memproj );
2578 mem_phi->init_req(2, mem);
2579 transform_later(mem_phi);
2580 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2581 }
2582
2583 //---------------------------eliminate_macro_nodes----------------------
2584 // Eliminate scalar replaced allocations and associated locks.
2585 void PhaseMacroExpand::eliminate_macro_nodes() {
2586 if (C->macro_count() == 0)
2587 return;
2588
2589 // First, attempt to eliminate locks
2590 int cnt = C->macro_count();
2591 for (int i=0; i < cnt; i++) {
2592 Node *n = C->macro_node(i);
2593 if (n->is_AbstractLock()) { // Lock and Unlock nodes
2594 // Before elimination mark all associated (same box and obj)
2595 // lock and unlock nodes.
2596 mark_eliminated_locking_nodes(n->as_AbstractLock());
2597 }
2598 }
2599 bool progress = true;
2600 while (progress) {
2601 progress = false;
2602 for (int i = C->macro_count(); i > 0; i--) {
2603 Node * n = C->macro_node(i-1);
2604 bool success = false;
2605 debug_only(int old_macro_count = C->macro_count(););
2606 if (n->is_AbstractLock()) {
2607 success = eliminate_locking_node(n->as_AbstractLock());
2608 }
2609 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2610 progress = progress || success;
2611 }
2612 }
2613 // Next, attempt to eliminate allocations
2614 _has_locks = false;
2615 progress = true;
2616 while (progress) {
2617 progress = false;
2618 for (int i = C->macro_count(); i > 0; i--) {
2619 Node * n = C->macro_node(i-1);
2620 bool success = false;
2621 debug_only(int old_macro_count = C->macro_count(););
2622 switch (n->class_id()) {
2623 case Node::Class_Allocate:
2624 case Node::Class_AllocateArray:
2625 success = eliminate_allocate_node(n->as_Allocate());
2626 break;
2627 case Node::Class_CallStaticJava:
2628 success = eliminate_boxing_node(n->as_CallStaticJava());
2629 break;
2630 case Node::Class_Lock:
2631 case Node::Class_Unlock:
2632 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2633 _has_locks = true;
2634 break;
2635 case Node::Class_ArrayCopy:
2636 break;
2637 default:
2638 assert(n->Opcode() == Op_LoopLimit ||
2639 n->Opcode() == Op_Opaque1 ||
2640 n->Opcode() == Op_Opaque2 ||
2641 n->Opcode() == Op_Opaque3, "unknown node type in macro list");
2642 }
2643 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2644 progress = progress || success;
2645 }
2646 }
2647 }
2648
2649 //------------------------------expand_macro_nodes----------------------
2650 // Returns true if a failure occurred.
2651 bool PhaseMacroExpand::expand_macro_nodes() {
2652 // Last attempt to eliminate macro nodes.
2653 eliminate_macro_nodes();
2654
2655 // Make sure expansion will not cause node limit to be exceeded.
2656 // Worst case is a macro node gets expanded into about 50 nodes.
2657 // Allow 50% more for optimization.
2658 if (C->check_node_count(C->macro_count() * 75, "out of nodes before macro expansion" ) )
2659 return true;
2660
2661 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2662 bool progress = true;
2663 while (progress) {
2664 progress = false;
2665 for (int i = C->macro_count(); i > 0; i--) {
2666 Node * n = C->macro_node(i-1);
2667 bool success = false;
2668 debug_only(int old_macro_count = C->macro_count(););
2669 if (n->Opcode() == Op_LoopLimit) {
2670 // Remove it from macro list and put on IGVN worklist to optimize.
2671 C->remove_macro_node(n);
2672 _igvn._worklist.push(n);
2673 success = true;
2674 } else if (n->Opcode() == Op_CallStaticJava) {
2675 // Remove it from macro list and put on IGVN worklist to optimize.
2676 C->remove_macro_node(n);
2677 _igvn._worklist.push(n);
2678 success = true;
2679 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) {
2680 _igvn.replace_node(n, n->in(1));
2681 success = true;
2682 #if INCLUDE_RTM_OPT
2683 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) {
2684 assert(C->profile_rtm(), "should be used only in rtm deoptimization code");
2685 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), "");
2686 Node* cmp = n->unique_out();
2687 #ifdef ASSERT
2688 // Validate graph.
2689 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), "");
2690 BoolNode* bol = cmp->unique_out()->as_Bool();
2691 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() &&
2692 (bol->_test._test == BoolTest::ne), "");
2693 IfNode* ifn = bol->unique_out()->as_If();
2694 assert((ifn->outcnt() == 2) &&
2695 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != NULL, "");
2696 #endif
2697 Node* repl = n->in(1);
2698 if (!_has_locks) {
2699 // Remove RTM state check if there are no locks in the code.
2700 // Replace input to compare the same value.
2701 repl = (cmp->in(1) == n) ? cmp->in(2) : cmp->in(1);
2702 }
2703 _igvn.replace_node(n, repl);
2704 success = true;
2705 #endif
2706 }
2707 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2708 progress = progress || success;
2709 }
2710 }
2711
2712 // expand arraycopy "macro" nodes first
2713 // For ReduceBulkZeroing, we must first process all arraycopy nodes
2714 // before the allocate nodes are expanded.
2715 int macro_idx = C->macro_count() - 1;
2716 while (macro_idx >= 0) {
2717 Node * n = C->macro_node(macro_idx);
2718 assert(n->is_macro(), "only macro nodes expected here");
2719 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) {
2720 // node is unreachable, so don't try to expand it
2721 C->remove_macro_node(n);
2722 } else if (n->is_ArrayCopy()){
2723 int macro_count = C->macro_count();
2724 expand_arraycopy_node(n->as_ArrayCopy());
2725 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2726 }
2727 if (C->failing()) return true;
2728 macro_idx --;
2729 }
2730
2731 // expand "macro" nodes
2732 // nodes are removed from the macro list as they are processed
2733 while (C->macro_count() > 0) {
2734 int macro_count = C->macro_count();
2735 Node * n = C->macro_node(macro_count-1);
2736 assert(n->is_macro(), "only macro nodes expected here");
2737 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) {
2738 // node is unreachable, so don't try to expand it
2739 C->remove_macro_node(n);
2740 continue;
2741 }
2742 switch (n->class_id()) {
2743 case Node::Class_Allocate:
2744 expand_allocate(n->as_Allocate());
2745 break;
2746 case Node::Class_AllocateArray:
2747 expand_allocate_array(n->as_AllocateArray());
2748 break;
2749 case Node::Class_Lock:
2750 expand_lock_node(n->as_Lock());
2751 break;
2752 case Node::Class_Unlock:
2753 expand_unlock_node(n->as_Unlock());
2754 break;
2755 default:
2756 assert(false, "unknown node type in macro list");
2757 }
2758 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2759 if (C->failing()) return true;
2760 }
2761
2762 _igvn.set_delay_transform(false);
2763 _igvn.optimize();
2764 if (C->failing()) return true;
2765 return false;
2766 }