1 /*
2 * Copyright (c) 2005, 2018, 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/shared/collectedHeap.inline.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/graphKit.hpp"
37 #include "opto/locknode.hpp"
38 #include "opto/loopnode.hpp"
39 #include "opto/macro.hpp"
40 #include "opto/memnode.hpp"
41 #include "opto/narrowptrnode.hpp"
42 #include "opto/node.hpp"
43 #include "opto/opaquenode.hpp"
44 #include "opto/phaseX.hpp"
45 #include "opto/rootnode.hpp"
46 #include "opto/runtime.hpp"
47 #include "opto/subnode.hpp"
48 #include "opto/type.hpp"
49 #include "runtime/sharedRuntime.hpp"
50 #if INCLUDE_ALL_GCS
51 #include "gc/g1/g1ThreadLocalData.hpp"
52 #endif // INCLUDE_ALL_GCS
53
54
55 //
56 // Replace any references to "oldref" in inputs to "use" with "newref".
57 // Returns the number of replacements made.
58 //
59 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
60 int nreplacements = 0;
61 uint req = use->req();
62 for (uint j = 0; j < use->len(); j++) {
63 Node *uin = use->in(j);
64 if (uin == oldref) {
65 if (j < req)
66 use->set_req(j, newref);
67 else
68 use->set_prec(j, newref);
69 nreplacements++;
70 } else if (j >= req && uin == NULL) {
71 break;
72 }
73 }
74 return nreplacements;
75 }
76
77 void PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) {
78 // Copy debug information and adjust JVMState information
79 uint old_dbg_start = oldcall->tf()->domain()->cnt();
80 uint new_dbg_start = newcall->tf()->domain()->cnt();
81 int jvms_adj = new_dbg_start - old_dbg_start;
82 assert (new_dbg_start == newcall->req(), "argument count mismatch");
83
84 // SafePointScalarObject node could be referenced several times in debug info.
85 // Use Dict to record cloned nodes.
86 Dict* sosn_map = new Dict(cmpkey,hashkey);
87 for (uint i = old_dbg_start; i < oldcall->req(); i++) {
88 Node* old_in = oldcall->in(i);
89 // Clone old SafePointScalarObjectNodes, adjusting their field contents.
90 if (old_in != NULL && old_in->is_SafePointScalarObject()) {
91 SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject();
92 uint old_unique = C->unique();
93 Node* new_in = old_sosn->clone(sosn_map);
94 if (old_unique != C->unique()) { // New node?
95 new_in->set_req(0, C->root()); // reset control edge
96 new_in = transform_later(new_in); // Register new node.
97 }
98 old_in = new_in;
99 }
100 newcall->add_req(old_in);
101 }
102
103 // JVMS may be shared so clone it before we modify it
104 newcall->set_jvms(oldcall->jvms() != NULL ? oldcall->jvms()->clone_deep(C) : NULL);
105 for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) {
106 jvms->set_map(newcall);
107 jvms->set_locoff(jvms->locoff()+jvms_adj);
108 jvms->set_stkoff(jvms->stkoff()+jvms_adj);
109 jvms->set_monoff(jvms->monoff()+jvms_adj);
110 jvms->set_scloff(jvms->scloff()+jvms_adj);
111 jvms->set_endoff(jvms->endoff()+jvms_adj);
112 }
113 }
114
115 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
116 Node* cmp;
117 if (mask != 0) {
118 Node* and_node = transform_later(new AndXNode(word, MakeConX(mask)));
119 cmp = transform_later(new CmpXNode(and_node, MakeConX(bits)));
120 } else {
121 cmp = word;
122 }
123 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne));
124 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
125 transform_later(iff);
126
127 // Fast path taken.
128 Node *fast_taken = transform_later(new IfFalseNode(iff));
129
130 // Fast path not-taken, i.e. slow path
131 Node *slow_taken = transform_later(new IfTrueNode(iff));
132
133 if (return_fast_path) {
134 region->init_req(edge, slow_taken); // Capture slow-control
135 return fast_taken;
136 } else {
137 region->init_req(edge, fast_taken); // Capture fast-control
138 return slow_taken;
139 }
140 }
141
142 //--------------------copy_predefined_input_for_runtime_call--------------------
143 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
144 // Set fixed predefined input arguments
145 call->init_req( TypeFunc::Control, ctrl );
146 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) );
147 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
148 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
149 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
150 }
151
152 //------------------------------make_slow_call---------------------------------
153 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type,
154 address slow_call, const char* leaf_name, Node* slow_path,
155 Node* parm0, Node* parm1, Node* parm2) {
156
157 // Slow-path call
158 CallNode *call = leaf_name
159 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
160 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM );
161
162 // Slow path call has no side-effects, uses few values
163 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
164 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0);
165 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1);
166 if (parm2 != NULL) call->init_req(TypeFunc::Parms+2, parm2);
167 copy_call_debug_info(oldcall, call);
168 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
169 _igvn.replace_node(oldcall, call);
170 transform_later(call);
171
172 return call;
173 }
174
175 void PhaseMacroExpand::extract_call_projections(CallNode *call) {
176 _fallthroughproj = NULL;
177 _fallthroughcatchproj = NULL;
178 _ioproj_fallthrough = NULL;
179 _ioproj_catchall = NULL;
180 _catchallcatchproj = NULL;
181 _memproj_fallthrough = NULL;
182 _memproj_catchall = NULL;
183 _resproj = NULL;
184 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) {
185 ProjNode *pn = call->fast_out(i)->as_Proj();
186 switch (pn->_con) {
187 case TypeFunc::Control:
188 {
189 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj
190 _fallthroughproj = pn;
191 DUIterator_Fast jmax, j = pn->fast_outs(jmax);
192 const Node *cn = pn->fast_out(j);
193 if (cn->is_Catch()) {
194 ProjNode *cpn = NULL;
195 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) {
196 cpn = cn->fast_out(k)->as_Proj();
197 assert(cpn->is_CatchProj(), "must be a CatchProjNode");
198 if (cpn->_con == CatchProjNode::fall_through_index)
199 _fallthroughcatchproj = cpn;
200 else {
201 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index.");
202 _catchallcatchproj = cpn;
203 }
204 }
205 }
206 break;
207 }
208 case TypeFunc::I_O:
209 if (pn->_is_io_use)
210 _ioproj_catchall = pn;
211 else
212 _ioproj_fallthrough = pn;
213 break;
214 case TypeFunc::Memory:
215 if (pn->_is_io_use)
216 _memproj_catchall = pn;
217 else
218 _memproj_fallthrough = pn;
219 break;
220 case TypeFunc::Parms:
221 _resproj = pn;
222 break;
223 default:
224 assert(false, "unexpected projection from allocation node.");
225 }
226 }
227
228 }
229
230 void PhaseMacroExpand::eliminate_gc_barrier(Node* p2x) {
231 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
232 bs->eliminate_gc_barrier(this, p2x);
233 }
234
235 // Search for a memory operation for the specified memory slice.
236 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
237 Node *orig_mem = mem;
238 Node *alloc_mem = alloc->in(TypeFunc::Memory);
239 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
240 while (true) {
241 if (mem == alloc_mem || mem == start_mem ) {
242 return mem; // hit one of our sentinels
243 } else if (mem->is_MergeMem()) {
244 mem = mem->as_MergeMem()->memory_at(alias_idx);
245 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
246 Node *in = mem->in(0);
247 // we can safely skip over safepoints, calls, locks and membars because we
248 // already know that the object is safe to eliminate.
249 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
250 return in;
251 } else if (in->is_Call()) {
252 CallNode *call = in->as_Call();
253 if (call->may_modify(tinst, phase)) {
254 assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape");
255 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) {
256 return in;
257 }
258 }
259 mem = in->in(TypeFunc::Memory);
260 } else if (in->is_MemBar()) {
261 ArrayCopyNode* ac = NULL;
262 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) {
263 assert(ac != NULL && ac->is_clonebasic(), "Only basic clone is a non escaping clone");
264 return ac;
265 }
266 mem = in->in(TypeFunc::Memory);
267 } else {
268 assert(false, "unexpected projection");
269 }
270 } else if (mem->is_Store()) {
271 const TypePtr* atype = mem->as_Store()->adr_type();
272 int adr_idx = phase->C->get_alias_index(atype);
273 if (adr_idx == alias_idx) {
274 assert(atype->isa_oopptr(), "address type must be oopptr");
275 int adr_offset = atype->offset();
276 uint adr_iid = atype->is_oopptr()->instance_id();
277 // Array elements references have the same alias_idx
278 // but different offset and different instance_id.
279 if (adr_offset == offset && adr_iid == alloc->_idx)
280 return mem;
281 } else {
282 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
283 }
284 mem = mem->in(MemNode::Memory);
285 } else if (mem->is_ClearArray()) {
286 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
287 // Can not bypass initialization of the instance
288 // we are looking.
289 debug_only(intptr_t offset;)
290 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
291 InitializeNode* init = alloc->as_Allocate()->initialization();
292 // We are looking for stored value, return Initialize node
293 // or memory edge from Allocate node.
294 if (init != NULL)
295 return init;
296 else
297 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers).
298 }
299 // Otherwise skip it (the call updated 'mem' value).
300 } else if (mem->Opcode() == Op_SCMemProj) {
301 mem = mem->in(0);
302 Node* adr = NULL;
303 if (mem->is_LoadStore()) {
304 adr = mem->in(MemNode::Address);
305 } else {
306 assert(mem->Opcode() == Op_EncodeISOArray ||
307 mem->Opcode() == Op_StrCompressedCopy, "sanity");
308 adr = mem->in(3); // Destination array
309 }
310 const TypePtr* atype = adr->bottom_type()->is_ptr();
311 int adr_idx = phase->C->get_alias_index(atype);
312 if (adr_idx == alias_idx) {
313 DEBUG_ONLY(mem->dump();)
314 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
315 return NULL;
316 }
317 mem = mem->in(MemNode::Memory);
318 } else if (mem->Opcode() == Op_StrInflatedCopy) {
319 Node* adr = mem->in(3); // Destination array
320 const TypePtr* atype = adr->bottom_type()->is_ptr();
321 int adr_idx = phase->C->get_alias_index(atype);
322 if (adr_idx == alias_idx) {
323 DEBUG_ONLY(mem->dump();)
324 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
325 return NULL;
326 }
327 mem = mem->in(MemNode::Memory);
328 } else {
329 return mem;
330 }
331 assert(mem != orig_mem, "dead memory loop");
332 }
333 }
334
335 // Generate loads from source of the arraycopy for fields of
336 // destination needed at a deoptimization point
337 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) {
338 BasicType bt = ft;
339 const Type *type = ftype;
340 if (ft == T_NARROWOOP) {
341 bt = T_OBJECT;
342 type = ftype->make_oopptr();
343 }
344 Node* res = NULL;
345 if (ac->is_clonebasic()) {
346 Node* base = ac->in(ArrayCopyNode::Src)->in(AddPNode::Base);
347 Node* adr = _igvn.transform(new AddPNode(base, base, MakeConX(offset)));
348 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
349 res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned);
350 } else {
351 if (ac->modifies(offset, offset, &_igvn, true)) {
352 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
353 uint shift = exact_log2(type2aelembytes(bt));
354 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
355 #ifdef _LP64
356 diff = _igvn.transform(new ConvI2LNode(diff));
357 #endif
358 diff = _igvn.transform(new LShiftXNode(diff, intcon(shift)));
359
360 Node* off = _igvn.transform(new AddXNode(MakeConX(offset), diff));
361 Node* base = ac->in(ArrayCopyNode::Src);
362 Node* adr = _igvn.transform(new AddPNode(base, base, off));
363 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
364 res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned);
365 }
366 }
367 if (res != NULL) {
368 res = _igvn.transform(res);
369 if (ftype->isa_narrowoop()) {
370 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
371 res = _igvn.transform(new EncodePNode(res, ftype));
372 }
373 return res;
374 }
375 return NULL;
376 }
377
378 //
379 // Given a Memory Phi, compute a value Phi containing the values from stores
380 // on the input paths.
381 // Note: this function is recursive, its depth is limited by the "level" argument
382 // Returns the computed Phi, or NULL if it cannot compute it.
383 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) {
384 assert(mem->is_Phi(), "sanity");
385 int alias_idx = C->get_alias_index(adr_t);
386 int offset = adr_t->offset();
387 int instance_id = adr_t->instance_id();
388
389 // Check if an appropriate value phi already exists.
390 Node* region = mem->in(0);
391 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
392 Node* phi = region->fast_out(k);
393 if (phi->is_Phi() && phi != mem &&
394 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) {
395 return phi;
396 }
397 }
398 // Check if an appropriate new value phi already exists.
399 Node* new_phi = value_phis->find(mem->_idx);
400 if (new_phi != NULL)
401 return new_phi;
402
403 if (level <= 0) {
404 return NULL; // Give up: phi tree too deep
405 }
406 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
407 Node *alloc_mem = alloc->in(TypeFunc::Memory);
408
409 uint length = mem->req();
410 GrowableArray <Node *> values(length, length, NULL, false);
411
412 // create a new Phi for the value
413 PhiNode *phi = new PhiNode(mem->in(0), phi_type, NULL, mem->_idx, instance_id, alias_idx, offset);
414 transform_later(phi);
415 value_phis->push(phi, mem->_idx);
416
417 for (uint j = 1; j < length; j++) {
418 Node *in = mem->in(j);
419 if (in == NULL || in->is_top()) {
420 values.at_put(j, in);
421 } else {
422 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
423 if (val == start_mem || val == alloc_mem) {
424 // hit a sentinel, return appropriate 0 value
425 values.at_put(j, _igvn.zerocon(ft));
426 continue;
427 }
428 if (val->is_Initialize()) {
429 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
430 }
431 if (val == NULL) {
432 return NULL; // can't find a value on this path
433 }
434 if (val == mem) {
435 values.at_put(j, mem);
436 } else if (val->is_Store()) {
437 values.at_put(j, val->in(MemNode::ValueIn));
438 } else if(val->is_Proj() && val->in(0) == alloc) {
439 values.at_put(j, _igvn.zerocon(ft));
440 } else if (val->is_Phi()) {
441 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
442 if (val == NULL) {
443 return NULL;
444 }
445 values.at_put(j, val);
446 } else if (val->Opcode() == Op_SCMemProj) {
447 assert(val->in(0)->is_LoadStore() ||
448 val->in(0)->Opcode() == Op_EncodeISOArray ||
449 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity");
450 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
451 return NULL;
452 } else if (val->is_ArrayCopy()) {
453 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc);
454 if (res == NULL) {
455 return NULL;
456 }
457 values.at_put(j, res);
458 } else {
459 #ifdef ASSERT
460 val->dump();
461 assert(false, "unknown node on this path");
462 #endif
463 return NULL; // unknown node on this path
464 }
465 }
466 }
467 // Set Phi's inputs
468 for (uint j = 1; j < length; j++) {
469 if (values.at(j) == mem) {
470 phi->init_req(j, phi);
471 } else {
472 phi->init_req(j, values.at(j));
473 }
474 }
475 return phi;
476 }
477
478 // Search the last value stored into the object's field.
479 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) {
480 assert(adr_t->is_known_instance_field(), "instance required");
481 int instance_id = adr_t->instance_id();
482 assert((uint)instance_id == alloc->_idx, "wrong allocation");
483
484 int alias_idx = C->get_alias_index(adr_t);
485 int offset = adr_t->offset();
486 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
487 Node *alloc_ctrl = alloc->in(TypeFunc::Control);
488 Node *alloc_mem = alloc->in(TypeFunc::Memory);
489 Arena *a = Thread::current()->resource_area();
490 VectorSet visited(a);
491
492
493 bool done = sfpt_mem == alloc_mem;
494 Node *mem = sfpt_mem;
495 while (!done) {
496 if (visited.test_set(mem->_idx)) {
497 return NULL; // found a loop, give up
498 }
499 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
500 if (mem == start_mem || mem == alloc_mem) {
501 done = true; // hit a sentinel, return appropriate 0 value
502 } else if (mem->is_Initialize()) {
503 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
504 if (mem == NULL) {
505 done = true; // Something go wrong.
506 } else if (mem->is_Store()) {
507 const TypePtr* atype = mem->as_Store()->adr_type();
508 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
509 done = true;
510 }
511 } else if (mem->is_Store()) {
512 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
513 assert(atype != NULL, "address type must be oopptr");
514 assert(C->get_alias_index(atype) == alias_idx &&
515 atype->is_known_instance_field() && atype->offset() == offset &&
516 atype->instance_id() == instance_id, "store is correct memory slice");
517 done = true;
518 } else if (mem->is_Phi()) {
519 // try to find a phi's unique input
520 Node *unique_input = NULL;
521 Node *top = C->top();
522 for (uint i = 1; i < mem->req(); i++) {
523 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
524 if (n == NULL || n == top || n == mem) {
525 continue;
526 } else if (unique_input == NULL) {
527 unique_input = n;
528 } else if (unique_input != n) {
529 unique_input = top;
530 break;
531 }
532 }
533 if (unique_input != NULL && unique_input != top) {
534 mem = unique_input;
535 } else {
536 done = true;
537 }
538 } else if (mem->is_ArrayCopy()) {
539 done = true;
540 } else {
541 assert(false, "unexpected node");
542 }
543 }
544 if (mem != NULL) {
545 if (mem == start_mem || mem == alloc_mem) {
546 // hit a sentinel, return appropriate 0 value
547 return _igvn.zerocon(ft);
548 } else if (mem->is_Store()) {
549 return mem->in(MemNode::ValueIn);
550 } else if (mem->is_Phi()) {
551 // attempt to produce a Phi reflecting the values on the input paths of the Phi
552 Node_Stack value_phis(a, 8);
553 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
554 if (phi != NULL) {
555 return phi;
556 } else {
557 // Kill all new Phis
558 while(value_phis.is_nonempty()) {
559 Node* n = value_phis.node();
560 _igvn.replace_node(n, C->top());
561 value_phis.pop();
562 }
563 }
564 } else if (mem->is_ArrayCopy()) {
565 Node* ctl = mem->in(0);
566 Node* m = mem->in(TypeFunc::Memory);
567 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) {
568 // pin the loads in the uncommon trap path
569 ctl = sfpt_ctl;
570 m = sfpt_mem;
571 }
572 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc);
573 }
574 }
575 // Something go wrong.
576 return NULL;
577 }
578
579 // Check the possibility of scalar replacement.
580 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
581 // Scan the uses of the allocation to check for anything that would
582 // prevent us from eliminating it.
583 NOT_PRODUCT( const char* fail_eliminate = NULL; )
584 DEBUG_ONLY( Node* disq_node = NULL; )
585 bool can_eliminate = true;
586
587 Node* res = alloc->result_cast();
588 const TypeOopPtr* res_type = NULL;
589 if (res == NULL) {
590 // All users were eliminated.
591 } else if (!res->is_CheckCastPP()) {
592 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
593 can_eliminate = false;
594 } else {
595 res_type = _igvn.type(res)->isa_oopptr();
596 if (res_type == NULL) {
597 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
598 can_eliminate = false;
599 } else if (res_type->isa_aryptr()) {
600 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
601 if (length < 0) {
602 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
603 can_eliminate = false;
604 }
605 }
606 }
607
608 if (can_eliminate && res != NULL) {
609 for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
610 j < jmax && can_eliminate; j++) {
611 Node* use = res->fast_out(j);
612
613 if (use->is_AddP()) {
614 const TypePtr* addp_type = _igvn.type(use)->is_ptr();
615 int offset = addp_type->offset();
616
617 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
618 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";)
619 can_eliminate = false;
620 break;
621 }
622 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
623 k < kmax && can_eliminate; k++) {
624 Node* n = use->fast_out(k);
625 if (!n->is_Store() && n->Opcode() != Op_CastP2X &&
626 !(n->is_ArrayCopy() &&
627 n->as_ArrayCopy()->is_clonebasic() &&
628 n->in(ArrayCopyNode::Dest) == use)) {
629 DEBUG_ONLY(disq_node = n;)
630 if (n->is_Load() || n->is_LoadStore()) {
631 NOT_PRODUCT(fail_eliminate = "Field load";)
632 } else {
633 NOT_PRODUCT(fail_eliminate = "Not store field referrence";)
634 }
635 can_eliminate = false;
636 }
637 }
638 } else if (use->is_ArrayCopy() &&
639 (use->as_ArrayCopy()->is_arraycopy_validated() ||
640 use->as_ArrayCopy()->is_copyof_validated() ||
641 use->as_ArrayCopy()->is_copyofrange_validated()) &&
642 use->in(ArrayCopyNode::Dest) == res) {
643 // ok to eliminate
644 } else if (use->is_SafePoint()) {
645 SafePointNode* sfpt = use->as_SafePoint();
646 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
647 // Object is passed as argument.
648 DEBUG_ONLY(disq_node = use;)
649 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
650 can_eliminate = false;
651 }
652 Node* sfptMem = sfpt->memory();
653 if (sfptMem == NULL || sfptMem->is_top()) {
654 DEBUG_ONLY(disq_node = use;)
655 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";)
656 can_eliminate = false;
657 } else {
658 safepoints.append_if_missing(sfpt);
659 }
660 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
661 if (use->is_Phi()) {
662 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
663 NOT_PRODUCT(fail_eliminate = "Object is return value";)
664 } else {
665 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
666 }
667 DEBUG_ONLY(disq_node = use;)
668 } else {
669 if (use->Opcode() == Op_Return) {
670 NOT_PRODUCT(fail_eliminate = "Object is return value";)
671 }else {
672 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
673 }
674 DEBUG_ONLY(disq_node = use;)
675 }
676 can_eliminate = false;
677 }
678 }
679 }
680
681 #ifndef PRODUCT
682 if (PrintEliminateAllocations) {
683 if (can_eliminate) {
684 tty->print("Scalar ");
685 if (res == NULL)
686 alloc->dump();
687 else
688 res->dump();
689 } else if (alloc->_is_scalar_replaceable) {
690 tty->print("NotScalar (%s)", fail_eliminate);
691 if (res == NULL)
692 alloc->dump();
693 else
694 res->dump();
695 #ifdef ASSERT
696 if (disq_node != NULL) {
697 tty->print(" >>>> ");
698 disq_node->dump();
699 }
700 #endif /*ASSERT*/
701 }
702 }
703 #endif
704 return can_eliminate;
705 }
706
707 // Do scalar replacement.
708 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
709 GrowableArray <SafePointNode *> safepoints_done;
710
711 ciKlass* klass = NULL;
712 ciInstanceKlass* iklass = NULL;
713 int nfields = 0;
714 int array_base = 0;
715 int element_size = 0;
716 BasicType basic_elem_type = T_ILLEGAL;
717 ciType* elem_type = NULL;
718
719 Node* res = alloc->result_cast();
720 assert(res == NULL || res->is_CheckCastPP(), "unexpected AllocateNode result");
721 const TypeOopPtr* res_type = NULL;
722 if (res != NULL) { // Could be NULL when there are no users
723 res_type = _igvn.type(res)->isa_oopptr();
724 }
725
726 if (res != NULL) {
727 klass = res_type->klass();
728 if (res_type->isa_instptr()) {
729 // find the fields of the class which will be needed for safepoint debug information
730 assert(klass->is_instance_klass(), "must be an instance klass.");
731 iklass = klass->as_instance_klass();
732 nfields = iklass->nof_nonstatic_fields();
733 } else {
734 // find the array's elements which will be needed for safepoint debug information
735 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
736 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass.");
737 elem_type = klass->as_array_klass()->element_type();
738 basic_elem_type = elem_type->basic_type();
739 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
740 element_size = type2aelembytes(basic_elem_type);
741 }
742 }
743 //
744 // Process the safepoint uses
745 //
746 while (safepoints.length() > 0) {
747 SafePointNode* sfpt = safepoints.pop();
748 Node* mem = sfpt->memory();
749 Node* ctl = sfpt->control();
750 assert(sfpt->jvms() != NULL, "missed JVMS");
751 // Fields of scalar objs are referenced only at the end
752 // of regular debuginfo at the last (youngest) JVMS.
753 // Record relative start index.
754 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
755 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type,
756 #ifdef ASSERT
757 alloc,
758 #endif
759 first_ind, nfields);
760 sobj->init_req(0, C->root());
761 transform_later(sobj);
762
763 // Scan object's fields adding an input to the safepoint for each field.
764 for (int j = 0; j < nfields; j++) {
765 intptr_t offset;
766 ciField* field = NULL;
767 if (iklass != NULL) {
768 field = iklass->nonstatic_field_at(j);
769 offset = field->offset();
770 elem_type = field->type();
771 basic_elem_type = field->layout_type();
772 } else {
773 offset = array_base + j * (intptr_t)element_size;
774 }
775
776 const Type *field_type;
777 // The next code is taken from Parse::do_get_xxx().
778 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) {
779 if (!elem_type->is_loaded()) {
780 field_type = TypeInstPtr::BOTTOM;
781 } else if (field != NULL && field->is_static_constant()) {
782 // This can happen if the constant oop is non-perm.
783 ciObject* con = field->constant_value().as_object();
784 // Do not "join" in the previous type; it doesn't add value,
785 // and may yield a vacuous result if the field is of interface type.
786 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
787 assert(field_type != NULL, "field singleton type must be consistent");
788 } else {
789 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
790 }
791 if (UseCompressedOops) {
792 field_type = field_type->make_narrowoop();
793 basic_elem_type = T_NARROWOOP;
794 }
795 } else {
796 field_type = Type::get_const_basic_type(basic_elem_type);
797 }
798
799 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr();
800
801 Node *field_val = value_from_mem(mem, ctl, basic_elem_type, field_type, field_addr_type, alloc);
802 if (field_val == NULL) {
803 // We weren't able to find a value for this field,
804 // give up on eliminating this allocation.
805
806 // Remove any extra entries we added to the safepoint.
807 uint last = sfpt->req() - 1;
808 for (int k = 0; k < j; k++) {
809 sfpt->del_req(last--);
810 }
811 _igvn._worklist.push(sfpt);
812 // rollback processed safepoints
813 while (safepoints_done.length() > 0) {
814 SafePointNode* sfpt_done = safepoints_done.pop();
815 // remove any extra entries we added to the safepoint
816 last = sfpt_done->req() - 1;
817 for (int k = 0; k < nfields; k++) {
818 sfpt_done->del_req(last--);
819 }
820 JVMState *jvms = sfpt_done->jvms();
821 jvms->set_endoff(sfpt_done->req());
822 // Now make a pass over the debug information replacing any references
823 // to SafePointScalarObjectNode with the allocated object.
824 int start = jvms->debug_start();
825 int end = jvms->debug_end();
826 for (int i = start; i < end; i++) {
827 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
828 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
829 if (scobj->first_index(jvms) == sfpt_done->req() &&
830 scobj->n_fields() == (uint)nfields) {
831 assert(scobj->alloc() == alloc, "sanity");
832 sfpt_done->set_req(i, res);
833 }
834 }
835 }
836 _igvn._worklist.push(sfpt_done);
837 }
838 #ifndef PRODUCT
839 if (PrintEliminateAllocations) {
840 if (field != NULL) {
841 tty->print("=== At SafePoint node %d can't find value of Field: ",
842 sfpt->_idx);
843 field->print();
844 int field_idx = C->get_alias_index(field_addr_type);
845 tty->print(" (alias_idx=%d)", field_idx);
846 } else { // Array's element
847 tty->print("=== At SafePoint node %d can't find value of array element [%d]",
848 sfpt->_idx, j);
849 }
850 tty->print(", which prevents elimination of: ");
851 if (res == NULL)
852 alloc->dump();
853 else
854 res->dump();
855 }
856 #endif
857 return false;
858 }
859 if (UseCompressedOops && field_type->isa_narrowoop()) {
860 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
861 // to be able scalar replace the allocation.
862 if (field_val->is_EncodeP()) {
863 field_val = field_val->in(1);
864 } else {
865 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
866 }
867 }
868 sfpt->add_req(field_val);
869 }
870 JVMState *jvms = sfpt->jvms();
871 jvms->set_endoff(sfpt->req());
872 // Now make a pass over the debug information replacing any references
873 // to the allocated object with "sobj"
874 int start = jvms->debug_start();
875 int end = jvms->debug_end();
876 sfpt->replace_edges_in_range(res, sobj, start, end);
877 _igvn._worklist.push(sfpt);
878 safepoints_done.append_if_missing(sfpt); // keep it for rollback
879 }
880 return true;
881 }
882
883 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
884 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control);
885 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory);
886 if (ctl_proj != NULL) {
887 igvn.replace_node(ctl_proj, n->in(0));
888 }
889 if (mem_proj != NULL) {
890 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
891 }
892 }
893
894 // Process users of eliminated allocation.
895 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) {
896 Node* res = alloc->result_cast();
897 if (res != NULL) {
898 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
899 Node *use = res->last_out(j);
900 uint oc1 = res->outcnt();
901
902 if (use->is_AddP()) {
903 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
904 Node *n = use->last_out(k);
905 uint oc2 = use->outcnt();
906 if (n->is_Store()) {
907 #ifdef ASSERT
908 // Verify that there is no dependent MemBarVolatile nodes,
909 // they should be removed during IGVN, see MemBarNode::Ideal().
910 for (DUIterator_Fast pmax, p = n->fast_outs(pmax);
911 p < pmax; p++) {
912 Node* mb = n->fast_out(p);
913 assert(mb->is_Initialize() || !mb->is_MemBar() ||
914 mb->req() <= MemBarNode::Precedent ||
915 mb->in(MemBarNode::Precedent) != n,
916 "MemBarVolatile should be eliminated for non-escaping object");
917 }
918 #endif
919 _igvn.replace_node(n, n->in(MemNode::Memory));
920 } else if (n->is_ArrayCopy()) {
921 // Disconnect ArrayCopy node
922 ArrayCopyNode* ac = n->as_ArrayCopy();
923 assert(ac->is_clonebasic(), "unexpected array copy kind");
924 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
925 disconnect_projections(ac, _igvn);
926 assert(alloc->in(0)->is_Proj() && alloc->in(0)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
927 Node* membar_before = alloc->in(0)->in(0);
928 disconnect_projections(membar_before->as_MemBar(), _igvn);
929 if (membar_after->is_MemBar()) {
930 disconnect_projections(membar_after->as_MemBar(), _igvn);
931 }
932 } else {
933 eliminate_gc_barrier(n);
934 }
935 k -= (oc2 - use->outcnt());
936 }
937 } else if (use->is_ArrayCopy()) {
938 // Disconnect ArrayCopy node
939 ArrayCopyNode* ac = use->as_ArrayCopy();
940 assert(ac->is_arraycopy_validated() ||
941 ac->is_copyof_validated() ||
942 ac->is_copyofrange_validated(), "unsupported");
943 CallProjections callprojs;
944 ac->extract_projections(&callprojs, true);
945
946 _igvn.replace_node(callprojs.fallthrough_ioproj, ac->in(TypeFunc::I_O));
947 _igvn.replace_node(callprojs.fallthrough_memproj, ac->in(TypeFunc::Memory));
948 _igvn.replace_node(callprojs.fallthrough_catchproj, ac->in(TypeFunc::Control));
949
950 // Set control to top. IGVN will remove the remaining projections
951 ac->set_req(0, top());
952 ac->replace_edge(res, top());
953
954 // Disconnect src right away: it can help find new
955 // opportunities for allocation elimination
956 Node* src = ac->in(ArrayCopyNode::Src);
957 ac->replace_edge(src, top());
958 // src can be top at this point if src and dest of the
959 // arraycopy were the same
960 if (src->outcnt() == 0 && !src->is_top()) {
961 _igvn.remove_dead_node(src);
962 }
963
964 _igvn._worklist.push(ac);
965 } else {
966 eliminate_gc_barrier(use);
967 }
968 j -= (oc1 - res->outcnt());
969 }
970 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
971 _igvn.remove_dead_node(res);
972 }
973
974 //
975 // Process other users of allocation's projections
976 //
977 if (_resproj != NULL && _resproj->outcnt() != 0) {
978 // First disconnect stores captured by Initialize node.
979 // If Initialize node is eliminated first in the following code,
980 // it will kill such stores and DUIterator_Last will assert.
981 for (DUIterator_Fast jmax, j = _resproj->fast_outs(jmax); j < jmax; j++) {
982 Node *use = _resproj->fast_out(j);
983 if (use->is_AddP()) {
984 // raw memory addresses used only by the initialization
985 _igvn.replace_node(use, C->top());
986 --j; --jmax;
987 }
988 }
989 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) {
990 Node *use = _resproj->last_out(j);
991 uint oc1 = _resproj->outcnt();
992 if (use->is_Initialize()) {
993 // Eliminate Initialize node.
994 InitializeNode *init = use->as_Initialize();
995 assert(init->outcnt() <= 2, "only a control and memory projection expected");
996 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control);
997 if (ctrl_proj != NULL) {
998 assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection");
999 _igvn.replace_node(ctrl_proj, _fallthroughcatchproj);
1000 }
1001 Node *mem_proj = init->proj_out_or_null(TypeFunc::Memory);
1002 if (mem_proj != NULL) {
1003 Node *mem = init->in(TypeFunc::Memory);
1004 #ifdef ASSERT
1005 if (mem->is_MergeMem()) {
1006 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection");
1007 } else {
1008 assert(mem == _memproj_fallthrough, "allocation memory projection");
1009 }
1010 #endif
1011 _igvn.replace_node(mem_proj, mem);
1012 }
1013 } else {
1014 assert(false, "only Initialize or AddP expected");
1015 }
1016 j -= (oc1 - _resproj->outcnt());
1017 }
1018 }
1019 if (_fallthroughcatchproj != NULL) {
1020 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control));
1021 }
1022 if (_memproj_fallthrough != NULL) {
1023 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory));
1024 }
1025 if (_memproj_catchall != NULL) {
1026 _igvn.replace_node(_memproj_catchall, C->top());
1027 }
1028 if (_ioproj_fallthrough != NULL) {
1029 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O));
1030 }
1031 if (_ioproj_catchall != NULL) {
1032 _igvn.replace_node(_ioproj_catchall, C->top());
1033 }
1034 if (_catchallcatchproj != NULL) {
1035 _igvn.replace_node(_catchallcatchproj, C->top());
1036 }
1037 }
1038
1039 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1040 // Don't do scalar replacement if the frame can be popped by JVMTI:
1041 // if reallocation fails during deoptimization we'll pop all
1042 // interpreter frames for this compiled frame and that won't play
1043 // nice with JVMTI popframe.
1044 if (!EliminateAllocations || JvmtiExport::can_pop_frame() || !alloc->_is_non_escaping) {
1045 return false;
1046 }
1047 Node* klass = alloc->in(AllocateNode::KlassNode);
1048 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1049 Node* res = alloc->result_cast();
1050 // Eliminate boxing allocations which are not used
1051 // regardless scalar replacable status.
1052 bool boxing_alloc = C->eliminate_boxing() &&
1053 tklass->klass()->is_instance_klass() &&
1054 tklass->klass()->as_instance_klass()->is_box_klass();
1055 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != NULL))) {
1056 return false;
1057 }
1058
1059 extract_call_projections(alloc);
1060
1061 GrowableArray <SafePointNode *> safepoints;
1062 if (!can_eliminate_allocation(alloc, safepoints)) {
1063 return false;
1064 }
1065
1066 if (!alloc->_is_scalar_replaceable) {
1067 assert(res == NULL, "sanity");
1068 // We can only eliminate allocation if all debug info references
1069 // are already replaced with SafePointScalarObject because
1070 // we can't search for a fields value without instance_id.
1071 if (safepoints.length() > 0) {
1072 return false;
1073 }
1074 }
1075
1076 if (!scalar_replacement(alloc, safepoints)) {
1077 return false;
1078 }
1079
1080 CompileLog* log = C->log();
1081 if (log != NULL) {
1082 log->head("eliminate_allocation type='%d'",
1083 log->identify(tklass->klass()));
1084 JVMState* p = alloc->jvms();
1085 while (p != NULL) {
1086 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1087 p = p->caller();
1088 }
1089 log->tail("eliminate_allocation");
1090 }
1091
1092 process_users_of_allocation(alloc);
1093
1094 #ifndef PRODUCT
1095 if (PrintEliminateAllocations) {
1096 if (alloc->is_AllocateArray())
1097 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1098 else
1099 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1100 }
1101 #endif
1102
1103 return true;
1104 }
1105
1106 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1107 // EA should remove all uses of non-escaping boxing node.
1108 if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != NULL) {
1109 return false;
1110 }
1111
1112 assert(boxing->result_cast() == NULL, "unexpected boxing node result");
1113
1114 extract_call_projections(boxing);
1115
1116 const TypeTuple* r = boxing->tf()->range();
1117 assert(r->cnt() > TypeFunc::Parms, "sanity");
1118 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1119 assert(t != NULL, "sanity");
1120
1121 CompileLog* log = C->log();
1122 if (log != NULL) {
1123 log->head("eliminate_boxing type='%d'",
1124 log->identify(t->klass()));
1125 JVMState* p = boxing->jvms();
1126 while (p != NULL) {
1127 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1128 p = p->caller();
1129 }
1130 log->tail("eliminate_boxing");
1131 }
1132
1133 process_users_of_allocation(boxing);
1134
1135 #ifndef PRODUCT
1136 if (PrintEliminateAllocations) {
1137 tty->print("++++ Eliminated: %d ", boxing->_idx);
1138 boxing->method()->print_short_name(tty);
1139 tty->cr();
1140 }
1141 #endif
1142
1143 return true;
1144 }
1145
1146 //---------------------------set_eden_pointers-------------------------
1147 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) {
1148 if (UseTLAB) { // Private allocation: load from TLS
1149 Node* thread = transform_later(new ThreadLocalNode());
1150 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset());
1151 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset());
1152 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset);
1153 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset);
1154 } else { // Shared allocation: load from globals
1155 CollectedHeap* ch = Universe::heap();
1156 address top_adr = (address)ch->top_addr();
1157 address end_adr = (address)ch->end_addr();
1158 eden_top_adr = makecon(TypeRawPtr::make(top_adr));
1159 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr);
1160 }
1161 }
1162
1163
1164 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
1165 Node* adr = basic_plus_adr(base, offset);
1166 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
1167 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered);
1168 transform_later(value);
1169 return value;
1170 }
1171
1172
1173 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
1174 Node* adr = basic_plus_adr(base, offset);
1175 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt, MemNode::unordered);
1176 transform_later(mem);
1177 return mem;
1178 }
1179
1180 //=============================================================================
1181 //
1182 // A L L O C A T I O N
1183 //
1184 // Allocation attempts to be fast in the case of frequent small objects.
1185 // It breaks down like this:
1186 //
1187 // 1) Size in doublewords is computed. This is a constant for objects and
1188 // variable for most arrays. Doubleword units are used to avoid size
1189 // overflow of huge doubleword arrays. We need doublewords in the end for
1190 // rounding.
1191 //
1192 // 2) Size is checked for being 'too large'. Too-large allocations will go
1193 // the slow path into the VM. The slow path can throw any required
1194 // exceptions, and does all the special checks for very large arrays. The
1195 // size test can constant-fold away for objects. For objects with
1196 // finalizers it constant-folds the otherway: you always go slow with
1197 // finalizers.
1198 //
1199 // 3) If NOT using TLABs, this is the contended loop-back point.
1200 // Load-Locked the heap top. If using TLABs normal-load the heap top.
1201 //
1202 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
1203 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
1204 // "size*8" we always enter the VM, where "largish" is a constant picked small
1205 // enough that there's always space between the eden max and 4Gig (old space is
1206 // there so it's quite large) and large enough that the cost of entering the VM
1207 // is dwarfed by the cost to initialize the space.
1208 //
1209 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1210 // down. If contended, repeat at step 3. If using TLABs normal-store
1211 // adjusted heap top back down; there is no contention.
1212 //
1213 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
1214 // fields.
1215 //
1216 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1217 // oop flavor.
1218 //
1219 //=============================================================================
1220 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1221 // Allocations bigger than this always go the slow route.
1222 // This value must be small enough that allocation attempts that need to
1223 // trigger exceptions go the slow route. Also, it must be small enough so
1224 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1225 //=============================================================================j//
1226 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1227 // The allocator will coalesce int->oop copies away. See comment in
1228 // coalesce.cpp about how this works. It depends critically on the exact
1229 // code shape produced here, so if you are changing this code shape
1230 // make sure the GC info for the heap-top is correct in and around the
1231 // slow-path call.
1232 //
1233
1234 void PhaseMacroExpand::expand_allocate_common(
1235 AllocateNode* alloc, // allocation node to be expanded
1236 Node* length, // array length for an array allocation
1237 const TypeFunc* slow_call_type, // Type of slow call
1238 address slow_call_address // Address of slow call
1239 )
1240 {
1241
1242 Node* ctrl = alloc->in(TypeFunc::Control);
1243 Node* mem = alloc->in(TypeFunc::Memory);
1244 Node* i_o = alloc->in(TypeFunc::I_O);
1245 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1246 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1247 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1248
1249 assert(ctrl != NULL, "must have control");
1250 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1251 // they will not be used if "always_slow" is set
1252 enum { slow_result_path = 1, fast_result_path = 2 };
1253 Node *result_region = NULL;
1254 Node *result_phi_rawmem = NULL;
1255 Node *result_phi_rawoop = NULL;
1256 Node *result_phi_i_o = NULL;
1257
1258 // The initial slow comparison is a size check, the comparison
1259 // we want to do is a BoolTest::gt
1260 bool always_slow = false;
1261 int tv = _igvn.find_int_con(initial_slow_test, -1);
1262 if (tv >= 0) {
1263 always_slow = (tv == 1);
1264 initial_slow_test = NULL;
1265 } else {
1266 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1267 }
1268
1269 if (C->env()->dtrace_alloc_probes() ||
1270 (!UseTLAB && !Universe::heap()->supports_inline_contig_alloc())) {
1271 // Force slow-path allocation
1272 always_slow = true;
1273 initial_slow_test = NULL;
1274 }
1275
1276
1277 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1278 Node *slow_region = NULL;
1279 Node *toobig_false = ctrl;
1280
1281 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent");
1282 // generate the initial test if necessary
1283 if (initial_slow_test != NULL ) {
1284 slow_region = new RegionNode(3);
1285
1286 // Now make the initial failure test. Usually a too-big test but
1287 // might be a TRUE for finalizers or a fancy class check for
1288 // newInstance0.
1289 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1290 transform_later(toobig_iff);
1291 // Plug the failing-too-big test into the slow-path region
1292 Node *toobig_true = new IfTrueNode( toobig_iff );
1293 transform_later(toobig_true);
1294 slow_region ->init_req( too_big_or_final_path, toobig_true );
1295 toobig_false = new IfFalseNode( toobig_iff );
1296 transform_later(toobig_false);
1297 } else { // No initial test, just fall into next case
1298 toobig_false = ctrl;
1299 debug_only(slow_region = NodeSentinel);
1300 }
1301
1302 Node *slow_mem = mem; // save the current memory state for slow path
1303 // generate the fast allocation code unless we know that the initial test will always go slow
1304 if (!always_slow) {
1305 // Fast path modifies only raw memory.
1306 if (mem->is_MergeMem()) {
1307 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1308 }
1309
1310 Node* eden_top_adr;
1311 Node* eden_end_adr;
1312
1313 set_eden_pointers(eden_top_adr, eden_end_adr);
1314
1315 // Load Eden::end. Loop invariant and hoisted.
1316 //
1317 // Note: We set the control input on "eden_end" and "old_eden_top" when using
1318 // a TLAB to work around a bug where these values were being moved across
1319 // a safepoint. These are not oops, so they cannot be include in the oop
1320 // map, but they can be changed by a GC. The proper way to fix this would
1321 // be to set the raw memory state when generating a SafepointNode. However
1322 // this will require extensive changes to the loop optimization in order to
1323 // prevent a degradation of the optimization.
1324 // See comment in memnode.hpp, around line 227 in class LoadPNode.
1325 Node *eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS);
1326
1327 // allocate the Region and Phi nodes for the result
1328 result_region = new RegionNode(3);
1329 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1330 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1331 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1332
1333 // We need a Region for the loop-back contended case.
1334 enum { fall_in_path = 1, contended_loopback_path = 2 };
1335 Node *contended_region;
1336 Node *contended_phi_rawmem;
1337 if (UseTLAB) {
1338 contended_region = toobig_false;
1339 contended_phi_rawmem = mem;
1340 } else {
1341 contended_region = new RegionNode(3);
1342 contended_phi_rawmem = new PhiNode(contended_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1343 // Now handle the passing-too-big test. We fall into the contended
1344 // loop-back merge point.
1345 contended_region ->init_req(fall_in_path, toobig_false);
1346 contended_phi_rawmem->init_req(fall_in_path, mem);
1347 transform_later(contended_region);
1348 transform_later(contended_phi_rawmem);
1349 }
1350
1351 // Load(-locked) the heap top.
1352 // See note above concerning the control input when using a TLAB
1353 Node *old_eden_top = UseTLAB
1354 ? new LoadPNode (ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered)
1355 : new LoadPLockedNode(contended_region, contended_phi_rawmem, eden_top_adr, MemNode::acquire);
1356
1357 transform_later(old_eden_top);
1358 // Add to heap top to get a new heap top
1359 Node *new_eden_top = new AddPNode(top(), old_eden_top, size_in_bytes);
1360 transform_later(new_eden_top);
1361 // Check for needing a GC; compare against heap end
1362 Node *needgc_cmp = new CmpPNode(new_eden_top, eden_end);
1363 transform_later(needgc_cmp);
1364 Node *needgc_bol = new BoolNode(needgc_cmp, BoolTest::ge);
1365 transform_later(needgc_bol);
1366 IfNode *needgc_iff = new IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN);
1367 transform_later(needgc_iff);
1368
1369 // Plug the failing-heap-space-need-gc test into the slow-path region
1370 Node *needgc_true = new IfTrueNode(needgc_iff);
1371 transform_later(needgc_true);
1372 if (initial_slow_test) {
1373 slow_region->init_req(need_gc_path, needgc_true);
1374 // This completes all paths into the slow merge point
1375 transform_later(slow_region);
1376 } else { // No initial slow path needed!
1377 // Just fall from the need-GC path straight into the VM call.
1378 slow_region = needgc_true;
1379 }
1380 // No need for a GC. Setup for the Store-Conditional
1381 Node *needgc_false = new IfFalseNode(needgc_iff);
1382 transform_later(needgc_false);
1383
1384 // Grab regular I/O before optional prefetch may change it.
1385 // Slow-path does no I/O so just set it to the original I/O.
1386 result_phi_i_o->init_req(slow_result_path, i_o);
1387
1388 i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem,
1389 old_eden_top, new_eden_top, length);
1390
1391 // Name successful fast-path variables
1392 Node* fast_oop = old_eden_top;
1393 Node* fast_oop_ctrl;
1394 Node* fast_oop_rawmem;
1395
1396 // Store (-conditional) the modified eden top back down.
1397 // StorePConditional produces flags for a test PLUS a modified raw
1398 // memory state.
1399 if (UseTLAB) {
1400 Node* store_eden_top =
1401 new StorePNode(needgc_false, contended_phi_rawmem, eden_top_adr,
1402 TypeRawPtr::BOTTOM, new_eden_top, MemNode::unordered);
1403 transform_later(store_eden_top);
1404 fast_oop_ctrl = needgc_false; // No contention, so this is the fast path
1405 fast_oop_rawmem = store_eden_top;
1406 } else {
1407 Node* store_eden_top =
1408 new StorePConditionalNode(needgc_false, contended_phi_rawmem, eden_top_adr,
1409 new_eden_top, fast_oop/*old_eden_top*/);
1410 transform_later(store_eden_top);
1411 Node *contention_check = new BoolNode(store_eden_top, BoolTest::ne);
1412 transform_later(contention_check);
1413 store_eden_top = new SCMemProjNode(store_eden_top);
1414 transform_later(store_eden_top);
1415
1416 // If not using TLABs, check to see if there was contention.
1417 IfNode *contention_iff = new IfNode (needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN);
1418 transform_later(contention_iff);
1419 Node *contention_true = new IfTrueNode(contention_iff);
1420 transform_later(contention_true);
1421 // If contention, loopback and try again.
1422 contended_region->init_req(contended_loopback_path, contention_true);
1423 contended_phi_rawmem->init_req(contended_loopback_path, store_eden_top);
1424
1425 // Fast-path succeeded with no contention!
1426 Node *contention_false = new IfFalseNode(contention_iff);
1427 transform_later(contention_false);
1428 fast_oop_ctrl = contention_false;
1429
1430 // Bump total allocated bytes for this thread
1431 Node* thread = new ThreadLocalNode();
1432 transform_later(thread);
1433 Node* alloc_bytes_adr = basic_plus_adr(top()/*not oop*/, thread,
1434 in_bytes(JavaThread::allocated_bytes_offset()));
1435 Node* alloc_bytes = make_load(fast_oop_ctrl, store_eden_top, alloc_bytes_adr,
1436 0, TypeLong::LONG, T_LONG);
1437 #ifdef _LP64
1438 Node* alloc_size = size_in_bytes;
1439 #else
1440 Node* alloc_size = new ConvI2LNode(size_in_bytes);
1441 transform_later(alloc_size);
1442 #endif
1443 Node* new_alloc_bytes = new AddLNode(alloc_bytes, alloc_size);
1444 transform_later(new_alloc_bytes);
1445 fast_oop_rawmem = make_store(fast_oop_ctrl, store_eden_top, alloc_bytes_adr,
1446 0, new_alloc_bytes, T_LONG);
1447 }
1448
1449 InitializeNode* init = alloc->initialization();
1450 fast_oop_rawmem = initialize_object(alloc,
1451 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1452 klass_node, length, size_in_bytes);
1453
1454 // If initialization is performed by an array copy, any required
1455 // MemBarStoreStore was already added. If the object does not
1456 // escape no need for a MemBarStoreStore. If the object does not
1457 // escape in its initializer and memory barrier (MemBarStoreStore or
1458 // stronger) is already added at exit of initializer, also no need
1459 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore
1460 // so that stores that initialize this object can't be reordered
1461 // with a subsequent store that makes this object accessible by
1462 // other threads.
1463 // Other threads include java threads and JVM internal threads
1464 // (for example concurrent GC threads). Current concurrent GC
1465 // implementation: CMS and G1 will not scan newly created object,
1466 // so it's safe to skip storestore barrier when allocation does
1467 // not escape.
1468 if (!alloc->does_not_escape_thread() &&
1469 !alloc->is_allocation_MemBar_redundant() &&
1470 (init == NULL || !init->is_complete_with_arraycopy())) {
1471 if (init == NULL || init->req() < InitializeNode::RawStores) {
1472 // No InitializeNode or no stores captured by zeroing
1473 // elimination. Simply add the MemBarStoreStore after object
1474 // initialization.
1475 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1476 transform_later(mb);
1477
1478 mb->init_req(TypeFunc::Memory, fast_oop_rawmem);
1479 mb->init_req(TypeFunc::Control, fast_oop_ctrl);
1480 fast_oop_ctrl = new ProjNode(mb,TypeFunc::Control);
1481 transform_later(fast_oop_ctrl);
1482 fast_oop_rawmem = new ProjNode(mb,TypeFunc::Memory);
1483 transform_later(fast_oop_rawmem);
1484 } else {
1485 // Add the MemBarStoreStore after the InitializeNode so that
1486 // all stores performing the initialization that were moved
1487 // before the InitializeNode happen before the storestore
1488 // barrier.
1489
1490 Node* init_ctrl = init->proj_out_or_null(TypeFunc::Control);
1491 Node* init_mem = init->proj_out_or_null(TypeFunc::Memory);
1492
1493 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1494 transform_later(mb);
1495
1496 Node* ctrl = new ProjNode(init,TypeFunc::Control);
1497 transform_later(ctrl);
1498 Node* mem = new ProjNode(init,TypeFunc::Memory);
1499 transform_later(mem);
1500
1501 // The MemBarStoreStore depends on control and memory coming
1502 // from the InitializeNode
1503 mb->init_req(TypeFunc::Memory, mem);
1504 mb->init_req(TypeFunc::Control, ctrl);
1505
1506 ctrl = new ProjNode(mb,TypeFunc::Control);
1507 transform_later(ctrl);
1508 mem = new ProjNode(mb,TypeFunc::Memory);
1509 transform_later(mem);
1510
1511 // All nodes that depended on the InitializeNode for control
1512 // and memory must now depend on the MemBarNode that itself
1513 // depends on the InitializeNode
1514 if (init_ctrl != NULL) {
1515 _igvn.replace_node(init_ctrl, ctrl);
1516 }
1517 if (init_mem != NULL) {
1518 _igvn.replace_node(init_mem, mem);
1519 }
1520 }
1521 }
1522
1523 if (C->env()->dtrace_extended_probes()) {
1524 // Slow-path call
1525 int size = TypeFunc::Parms + 2;
1526 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1527 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base),
1528 "dtrace_object_alloc",
1529 TypeRawPtr::BOTTOM);
1530
1531 // Get base of thread-local storage area
1532 Node* thread = new ThreadLocalNode();
1533 transform_later(thread);
1534
1535 call->init_req(TypeFunc::Parms+0, thread);
1536 call->init_req(TypeFunc::Parms+1, fast_oop);
1537 call->init_req(TypeFunc::Control, fast_oop_ctrl);
1538 call->init_req(TypeFunc::I_O , top()); // does no i/o
1539 call->init_req(TypeFunc::Memory , fast_oop_rawmem);
1540 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1541 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1542 transform_later(call);
1543 fast_oop_ctrl = new ProjNode(call,TypeFunc::Control);
1544 transform_later(fast_oop_ctrl);
1545 fast_oop_rawmem = new ProjNode(call,TypeFunc::Memory);
1546 transform_later(fast_oop_rawmem);
1547 }
1548
1549 // Plug in the successful fast-path into the result merge point
1550 result_region ->init_req(fast_result_path, fast_oop_ctrl);
1551 result_phi_rawoop->init_req(fast_result_path, fast_oop);
1552 result_phi_i_o ->init_req(fast_result_path, i_o);
1553 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1554 } else {
1555 slow_region = ctrl;
1556 result_phi_i_o = i_o; // Rename it to use in the following code.
1557 }
1558
1559 // Generate slow-path call
1560 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1561 OptoRuntime::stub_name(slow_call_address),
1562 alloc->jvms()->bci(),
1563 TypePtr::BOTTOM);
1564 call->init_req( TypeFunc::Control, slow_region );
1565 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o
1566 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs
1567 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) );
1568 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) );
1569
1570 call->init_req(TypeFunc::Parms+0, klass_node);
1571 if (length != NULL) {
1572 call->init_req(TypeFunc::Parms+1, length);
1573 }
1574
1575 // Copy debug information and adjust JVMState information, then replace
1576 // allocate node with the call
1577 copy_call_debug_info((CallNode *) alloc, call);
1578 if (!always_slow) {
1579 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1580 } else {
1581 // Hook i_o projection to avoid its elimination during allocation
1582 // replacement (when only a slow call is generated).
1583 call->set_req(TypeFunc::I_O, result_phi_i_o);
1584 }
1585 _igvn.replace_node(alloc, call);
1586 transform_later(call);
1587
1588 // Identify the output projections from the allocate node and
1589 // adjust any references to them.
1590 // The control and io projections look like:
1591 //
1592 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1593 // Allocate Catch
1594 // ^---Proj(io) <-------+ ^---CatchProj(io)
1595 //
1596 // We are interested in the CatchProj nodes.
1597 //
1598 extract_call_projections(call);
1599
1600 // An allocate node has separate memory projections for the uses on
1601 // the control and i_o paths. Replace the control memory projection with
1602 // result_phi_rawmem (unless we are only generating a slow call when
1603 // both memory projections are combined)
1604 if (!always_slow && _memproj_fallthrough != NULL) {
1605 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) {
1606 Node *use = _memproj_fallthrough->fast_out(i);
1607 _igvn.rehash_node_delayed(use);
1608 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem);
1609 // back up iterator
1610 --i;
1611 }
1612 }
1613 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete
1614 // _memproj_catchall so we end up with a call that has only 1 memory projection.
1615 if (_memproj_catchall != NULL ) {
1616 if (_memproj_fallthrough == NULL) {
1617 _memproj_fallthrough = new ProjNode(call, TypeFunc::Memory);
1618 transform_later(_memproj_fallthrough);
1619 }
1620 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) {
1621 Node *use = _memproj_catchall->fast_out(i);
1622 _igvn.rehash_node_delayed(use);
1623 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough);
1624 // back up iterator
1625 --i;
1626 }
1627 assert(_memproj_catchall->outcnt() == 0, "all uses must be deleted");
1628 _igvn.remove_dead_node(_memproj_catchall);
1629 }
1630
1631 // An allocate node has separate i_o projections for the uses on the control
1632 // and i_o paths. Always replace the control i_o projection with result i_o
1633 // otherwise incoming i_o become dead when only a slow call is generated
1634 // (it is different from memory projections where both projections are
1635 // combined in such case).
1636 if (_ioproj_fallthrough != NULL) {
1637 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) {
1638 Node *use = _ioproj_fallthrough->fast_out(i);
1639 _igvn.rehash_node_delayed(use);
1640 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o);
1641 // back up iterator
1642 --i;
1643 }
1644 }
1645 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete
1646 // _ioproj_catchall so we end up with a call that has only 1 i_o projection.
1647 if (_ioproj_catchall != NULL ) {
1648 if (_ioproj_fallthrough == NULL) {
1649 _ioproj_fallthrough = new ProjNode(call, TypeFunc::I_O);
1650 transform_later(_ioproj_fallthrough);
1651 }
1652 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) {
1653 Node *use = _ioproj_catchall->fast_out(i);
1654 _igvn.rehash_node_delayed(use);
1655 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough);
1656 // back up iterator
1657 --i;
1658 }
1659 assert(_ioproj_catchall->outcnt() == 0, "all uses must be deleted");
1660 _igvn.remove_dead_node(_ioproj_catchall);
1661 }
1662
1663 // if we generated only a slow call, we are done
1664 if (always_slow) {
1665 // Now we can unhook i_o.
1666 if (result_phi_i_o->outcnt() > 1) {
1667 call->set_req(TypeFunc::I_O, top());
1668 } else {
1669 assert(result_phi_i_o->unique_ctrl_out() == call, "");
1670 // Case of new array with negative size known during compilation.
1671 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1672 // following code since call to runtime will throw exception.
1673 // As result there will be no users of i_o after the call.
1674 // Leave i_o attached to this call to avoid problems in preceding graph.
1675 }
1676 return;
1677 }
1678
1679
1680 if (_fallthroughcatchproj != NULL) {
1681 ctrl = _fallthroughcatchproj->clone();
1682 transform_later(ctrl);
1683 _igvn.replace_node(_fallthroughcatchproj, result_region);
1684 } else {
1685 ctrl = top();
1686 }
1687 Node *slow_result;
1688 if (_resproj == NULL) {
1689 // no uses of the allocation result
1690 slow_result = top();
1691 } else {
1692 slow_result = _resproj->clone();
1693 transform_later(slow_result);
1694 _igvn.replace_node(_resproj, result_phi_rawoop);
1695 }
1696
1697 // Plug slow-path into result merge point
1698 result_region ->init_req( slow_result_path, ctrl );
1699 result_phi_rawoop->init_req( slow_result_path, slow_result);
1700 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough );
1701 transform_later(result_region);
1702 transform_later(result_phi_rawoop);
1703 transform_later(result_phi_rawmem);
1704 transform_later(result_phi_i_o);
1705 // This completes all paths into the result merge point
1706 }
1707
1708
1709 // Helper for PhaseMacroExpand::expand_allocate_common.
1710 // Initializes the newly-allocated storage.
1711 Node*
1712 PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1713 Node* control, Node* rawmem, Node* object,
1714 Node* klass_node, Node* length,
1715 Node* size_in_bytes) {
1716 InitializeNode* init = alloc->initialization();
1717 // Store the klass & mark bits
1718 Node* mark_node = NULL;
1719 // For now only enable fast locking for non-array types
1720 if (UseBiasedLocking && (length == NULL)) {
1721 mark_node = make_load(control, rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
1722 } else {
1723 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype()));
1724 }
1725 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
1726
1727 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1728 int header_size = alloc->minimum_header_size(); // conservatively small
1729
1730 // Array length
1731 if (length != NULL) { // Arrays need length field
1732 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1733 // conservatively small header size:
1734 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1735 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1736 if (k->is_array_klass()) // we know the exact header size in most cases:
1737 header_size = Klass::layout_helper_header_size(k->layout_helper());
1738 }
1739
1740 // Clear the object body, if necessary.
1741 if (init == NULL) {
1742 // The init has somehow disappeared; be cautious and clear everything.
1743 //
1744 // This can happen if a node is allocated but an uncommon trap occurs
1745 // immediately. In this case, the Initialize gets associated with the
1746 // trap, and may be placed in a different (outer) loop, if the Allocate
1747 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1748 // there can be two Allocates to one Initialize. The answer in all these
1749 // edge cases is safety first. It is always safe to clear immediately
1750 // within an Allocate, and then (maybe or maybe not) clear some more later.
1751 if (!(UseTLAB && ZeroTLAB)) {
1752 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1753 header_size, size_in_bytes,
1754 &_igvn);
1755 }
1756 } else {
1757 if (!init->is_complete()) {
1758 // Try to win by zeroing only what the init does not store.
1759 // We can also try to do some peephole optimizations,
1760 // such as combining some adjacent subword stores.
1761 rawmem = init->complete_stores(control, rawmem, object,
1762 header_size, size_in_bytes, &_igvn);
1763 }
1764 // We have no more use for this link, since the AllocateNode goes away:
1765 init->set_req(InitializeNode::RawAddress, top());
1766 // (If we keep the link, it just confuses the register allocator,
1767 // who thinks he sees a real use of the address by the membar.)
1768 }
1769
1770 return rawmem;
1771 }
1772
1773 // Generate prefetch instructions for next allocations.
1774 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
1775 Node*& contended_phi_rawmem,
1776 Node* old_eden_top, Node* new_eden_top,
1777 Node* length) {
1778 enum { fall_in_path = 1, pf_path = 2 };
1779 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
1780 // Generate prefetch allocation with watermark check.
1781 // As an allocation hits the watermark, we will prefetch starting
1782 // at a "distance" away from watermark.
1783
1784 Node *pf_region = new RegionNode(3);
1785 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY,
1786 TypeRawPtr::BOTTOM );
1787 // I/O is used for Prefetch
1788 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO );
1789
1790 Node *thread = new ThreadLocalNode();
1791 transform_later(thread);
1792
1793 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread,
1794 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
1795 transform_later(eden_pf_adr);
1796
1797 Node *old_pf_wm = new LoadPNode(needgc_false,
1798 contended_phi_rawmem, eden_pf_adr,
1799 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM,
1800 MemNode::unordered);
1801 transform_later(old_pf_wm);
1802
1803 // check against new_eden_top
1804 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm );
1805 transform_later(need_pf_cmp);
1806 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge );
1807 transform_later(need_pf_bol);
1808 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol,
1809 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1810 transform_later(need_pf_iff);
1811
1812 // true node, add prefetchdistance
1813 Node *need_pf_true = new IfTrueNode( need_pf_iff );
1814 transform_later(need_pf_true);
1815
1816 Node *need_pf_false = new IfFalseNode( need_pf_iff );
1817 transform_later(need_pf_false);
1818
1819 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm,
1820 _igvn.MakeConX(AllocatePrefetchDistance) );
1821 transform_later(new_pf_wmt );
1822 new_pf_wmt->set_req(0, need_pf_true);
1823
1824 Node *store_new_wmt = new StorePNode(need_pf_true,
1825 contended_phi_rawmem, eden_pf_adr,
1826 TypeRawPtr::BOTTOM, new_pf_wmt,
1827 MemNode::unordered);
1828 transform_later(store_new_wmt);
1829
1830 // adding prefetches
1831 pf_phi_abio->init_req( fall_in_path, i_o );
1832
1833 Node *prefetch_adr;
1834 Node *prefetch;
1835 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1836 uint step_size = AllocatePrefetchStepSize;
1837 uint distance = 0;
1838
1839 for ( uint i = 0; i < lines; i++ ) {
1840 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt,
1841 _igvn.MakeConX(distance) );
1842 transform_later(prefetch_adr);
1843 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1844 transform_later(prefetch);
1845 distance += step_size;
1846 i_o = prefetch;
1847 }
1848 pf_phi_abio->set_req( pf_path, i_o );
1849
1850 pf_region->init_req( fall_in_path, need_pf_false );
1851 pf_region->init_req( pf_path, need_pf_true );
1852
1853 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
1854 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
1855
1856 transform_later(pf_region);
1857 transform_later(pf_phi_rawmem);
1858 transform_later(pf_phi_abio);
1859
1860 needgc_false = pf_region;
1861 contended_phi_rawmem = pf_phi_rawmem;
1862 i_o = pf_phi_abio;
1863 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) {
1864 // Insert a prefetch instruction for each allocation.
1865 // This code is used to generate 1 prefetch instruction per cache line.
1866
1867 // Generate several prefetch instructions.
1868 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1869 uint step_size = AllocatePrefetchStepSize;
1870 uint distance = AllocatePrefetchDistance;
1871
1872 // Next cache address.
1873 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top,
1874 _igvn.MakeConX(step_size + distance));
1875 transform_later(cache_adr);
1876 cache_adr = new CastP2XNode(needgc_false, cache_adr);
1877 transform_later(cache_adr);
1878 // Address is aligned to execute prefetch to the beginning of cache line size
1879 // (it is important when BIS instruction is used on SPARC as prefetch).
1880 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1));
1881 cache_adr = new AndXNode(cache_adr, mask);
1882 transform_later(cache_adr);
1883 cache_adr = new CastX2PNode(cache_adr);
1884 transform_later(cache_adr);
1885
1886 // Prefetch
1887 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr );
1888 prefetch->set_req(0, needgc_false);
1889 transform_later(prefetch);
1890 contended_phi_rawmem = prefetch;
1891 Node *prefetch_adr;
1892 distance = step_size;
1893 for ( uint i = 1; i < lines; i++ ) {
1894 prefetch_adr = new AddPNode( cache_adr, cache_adr,
1895 _igvn.MakeConX(distance) );
1896 transform_later(prefetch_adr);
1897 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr );
1898 transform_later(prefetch);
1899 distance += step_size;
1900 contended_phi_rawmem = prefetch;
1901 }
1902 } else if( AllocatePrefetchStyle > 0 ) {
1903 // Insert a prefetch for each allocation only on the fast-path
1904 Node *prefetch_adr;
1905 Node *prefetch;
1906 // Generate several prefetch instructions.
1907 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1908 uint step_size = AllocatePrefetchStepSize;
1909 uint distance = AllocatePrefetchDistance;
1910 for ( uint i = 0; i < lines; i++ ) {
1911 prefetch_adr = new AddPNode( old_eden_top, new_eden_top,
1912 _igvn.MakeConX(distance) );
1913 transform_later(prefetch_adr);
1914 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1915 // Do not let it float too high, since if eden_top == eden_end,
1916 // both might be null.
1917 if( i == 0 ) { // Set control for first prefetch, next follows it
1918 prefetch->init_req(0, needgc_false);
1919 }
1920 transform_later(prefetch);
1921 distance += step_size;
1922 i_o = prefetch;
1923 }
1924 }
1925 return i_o;
1926 }
1927
1928
1929 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
1930 expand_allocate_common(alloc, NULL,
1931 OptoRuntime::new_instance_Type(),
1932 OptoRuntime::new_instance_Java());
1933 }
1934
1935 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
1936 Node* length = alloc->in(AllocateNode::ALength);
1937 InitializeNode* init = alloc->initialization();
1938 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1939 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1940 address slow_call_address; // Address of slow call
1941 if (init != NULL && init->is_complete_with_arraycopy() &&
1942 k->is_type_array_klass()) {
1943 // Don't zero type array during slow allocation in VM since
1944 // it will be initialized later by arraycopy in compiled code.
1945 slow_call_address = OptoRuntime::new_array_nozero_Java();
1946 } else {
1947 slow_call_address = OptoRuntime::new_array_Java();
1948 }
1949 expand_allocate_common(alloc, length,
1950 OptoRuntime::new_array_Type(),
1951 slow_call_address);
1952 }
1953
1954 //-------------------mark_eliminated_box----------------------------------
1955 //
1956 // During EA obj may point to several objects but after few ideal graph
1957 // transformations (CCP) it may point to only one non escaping object
1958 // (but still using phi), corresponding locks and unlocks will be marked
1959 // for elimination. Later obj could be replaced with a new node (new phi)
1960 // and which does not have escape information. And later after some graph
1961 // reshape other locks and unlocks (which were not marked for elimination
1962 // before) are connected to this new obj (phi) but they still will not be
1963 // marked for elimination since new obj has no escape information.
1964 // Mark all associated (same box and obj) lock and unlock nodes for
1965 // elimination if some of them marked already.
1966 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) {
1967 if (oldbox->as_BoxLock()->is_eliminated())
1968 return; // This BoxLock node was processed already.
1969
1970 // New implementation (EliminateNestedLocks) has separate BoxLock
1971 // node for each locked region so mark all associated locks/unlocks as
1972 // eliminated even if different objects are referenced in one locked region
1973 // (for example, OSR compilation of nested loop inside locked scope).
1974 if (EliminateNestedLocks ||
1975 oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj)) {
1976 // Box is used only in one lock region. Mark this box as eliminated.
1977 _igvn.hash_delete(oldbox);
1978 oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value
1979 _igvn.hash_insert(oldbox);
1980
1981 for (uint i = 0; i < oldbox->outcnt(); i++) {
1982 Node* u = oldbox->raw_out(i);
1983 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) {
1984 AbstractLockNode* alock = u->as_AbstractLock();
1985 // Check lock's box since box could be referenced by Lock's debug info.
1986 if (alock->box_node() == oldbox) {
1987 // Mark eliminated all related locks and unlocks.
1988 #ifdef ASSERT
1989 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4");
1990 #endif
1991 alock->set_non_esc_obj();
1992 }
1993 }
1994 }
1995 return;
1996 }
1997
1998 // Create new "eliminated" BoxLock node and use it in monitor debug info
1999 // instead of oldbox for the same object.
2000 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
2001
2002 // Note: BoxLock node is marked eliminated only here and it is used
2003 // to indicate that all associated lock and unlock nodes are marked
2004 // for elimination.
2005 newbox->set_eliminated();
2006 transform_later(newbox);
2007
2008 // Replace old box node with new box for all users of the same object.
2009 for (uint i = 0; i < oldbox->outcnt();) {
2010 bool next_edge = true;
2011
2012 Node* u = oldbox->raw_out(i);
2013 if (u->is_AbstractLock()) {
2014 AbstractLockNode* alock = u->as_AbstractLock();
2015 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) {
2016 // Replace Box and mark eliminated all related locks and unlocks.
2017 #ifdef ASSERT
2018 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5");
2019 #endif
2020 alock->set_non_esc_obj();
2021 _igvn.rehash_node_delayed(alock);
2022 alock->set_box_node(newbox);
2023 next_edge = false;
2024 }
2025 }
2026 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) {
2027 FastLockNode* flock = u->as_FastLock();
2028 assert(flock->box_node() == oldbox, "sanity");
2029 _igvn.rehash_node_delayed(flock);
2030 flock->set_box_node(newbox);
2031 next_edge = false;
2032 }
2033
2034 // Replace old box in monitor debug info.
2035 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
2036 SafePointNode* sfn = u->as_SafePoint();
2037 JVMState* youngest_jvms = sfn->jvms();
2038 int max_depth = youngest_jvms->depth();
2039 for (int depth = 1; depth <= max_depth; depth++) {
2040 JVMState* jvms = youngest_jvms->of_depth(depth);
2041 int num_mon = jvms->nof_monitors();
2042 // Loop over monitors
2043 for (int idx = 0; idx < num_mon; idx++) {
2044 Node* obj_node = sfn->monitor_obj(jvms, idx);
2045 Node* box_node = sfn->monitor_box(jvms, idx);
2046 if (box_node == oldbox && obj_node->eqv_uncast(obj)) {
2047 int j = jvms->monitor_box_offset(idx);
2048 _igvn.replace_input_of(u, j, newbox);
2049 next_edge = false;
2050 }
2051 }
2052 }
2053 }
2054 if (next_edge) i++;
2055 }
2056 }
2057
2058 //-----------------------mark_eliminated_locking_nodes-----------------------
2059 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) {
2060 if (EliminateNestedLocks) {
2061 if (alock->is_nested()) {
2062 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity");
2063 return;
2064 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened
2065 // Only Lock node has JVMState needed here.
2066 // Not that preceding claim is documented anywhere else.
2067 if (alock->jvms() != NULL) {
2068 if (alock->as_Lock()->is_nested_lock_region()) {
2069 // Mark eliminated related nested locks and unlocks.
2070 Node* obj = alock->obj_node();
2071 BoxLockNode* box_node = alock->box_node()->as_BoxLock();
2072 assert(!box_node->is_eliminated(), "should not be marked yet");
2073 // Note: BoxLock node is marked eliminated only here
2074 // and it is used to indicate that all associated lock
2075 // and unlock nodes are marked for elimination.
2076 box_node->set_eliminated(); // Box's hash is always NO_HASH here
2077 for (uint i = 0; i < box_node->outcnt(); i++) {
2078 Node* u = box_node->raw_out(i);
2079 if (u->is_AbstractLock()) {
2080 alock = u->as_AbstractLock();
2081 if (alock->box_node() == box_node) {
2082 // Verify that this Box is referenced only by related locks.
2083 assert(alock->obj_node()->eqv_uncast(obj), "");
2084 // Mark all related locks and unlocks.
2085 #ifdef ASSERT
2086 alock->log_lock_optimization(C, "eliminate_lock_set_nested");
2087 #endif
2088 alock->set_nested();
2089 }
2090 }
2091 }
2092 } else {
2093 #ifdef ASSERT
2094 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region");
2095 if (C->log() != NULL)
2096 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output
2097 #endif
2098 }
2099 }
2100 return;
2101 }
2102 // Process locks for non escaping object
2103 assert(alock->is_non_esc_obj(), "");
2104 } // EliminateNestedLocks
2105
2106 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2107 // Look for all locks of this object and mark them and
2108 // corresponding BoxLock nodes as eliminated.
2109 Node* obj = alock->obj_node();
2110 for (uint j = 0; j < obj->outcnt(); j++) {
2111 Node* o = obj->raw_out(j);
2112 if (o->is_AbstractLock() &&
2113 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2114 alock = o->as_AbstractLock();
2115 Node* box = alock->box_node();
2116 // Replace old box node with new eliminated box for all users
2117 // of the same object and mark related locks as eliminated.
2118 mark_eliminated_box(box, obj);
2119 }
2120 }
2121 }
2122 }
2123
2124 // we have determined that this lock/unlock can be eliminated, we simply
2125 // eliminate the node without expanding it.
2126 //
2127 // Note: The membar's associated with the lock/unlock are currently not
2128 // eliminated. This should be investigated as a future enhancement.
2129 //
2130 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2131
2132 if (!alock->is_eliminated()) {
2133 return false;
2134 }
2135 #ifdef ASSERT
2136 if (!alock->is_coarsened()) {
2137 // Check that new "eliminated" BoxLock node is created.
2138 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2139 assert(oldbox->is_eliminated(), "should be done already");
2140 }
2141 #endif
2142
2143 alock->log_lock_optimization(C, "eliminate_lock");
2144
2145 #ifndef PRODUCT
2146 if (PrintEliminateLocks) {
2147 if (alock->is_Lock()) {
2148 tty->print_cr("++++ Eliminated: %d Lock", alock->_idx);
2149 } else {
2150 tty->print_cr("++++ Eliminated: %d Unlock", alock->_idx);
2151 }
2152 }
2153 #endif
2154
2155 Node* mem = alock->in(TypeFunc::Memory);
2156 Node* ctrl = alock->in(TypeFunc::Control);
2157
2158 extract_call_projections(alock);
2159 // There are 2 projections from the lock. The lock node will
2160 // be deleted when its last use is subsumed below.
2161 assert(alock->outcnt() == 2 &&
2162 _fallthroughproj != NULL &&
2163 _memproj_fallthrough != NULL,
2164 "Unexpected projections from Lock/Unlock");
2165
2166 Node* fallthroughproj = _fallthroughproj;
2167 Node* memproj_fallthrough = _memproj_fallthrough;
2168
2169 // The memory projection from a lock/unlock is RawMem
2170 // The input to a Lock is merged memory, so extract its RawMem input
2171 // (unless the MergeMem has been optimized away.)
2172 if (alock->is_Lock()) {
2173 // Seach for MemBarAcquireLock node and delete it also.
2174 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2175 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, "");
2176 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2177 Node* memproj = membar->proj_out(TypeFunc::Memory);
2178 _igvn.replace_node(ctrlproj, fallthroughproj);
2179 _igvn.replace_node(memproj, memproj_fallthrough);
2180
2181 // Delete FastLock node also if this Lock node is unique user
2182 // (a loop peeling may clone a Lock node).
2183 Node* flock = alock->as_Lock()->fastlock_node();
2184 if (flock->outcnt() == 1) {
2185 assert(flock->unique_out() == alock, "sanity");
2186 _igvn.replace_node(flock, top());
2187 }
2188 }
2189
2190 // Seach for MemBarReleaseLock node and delete it also.
2191 if (alock->is_Unlock() && ctrl != NULL && ctrl->is_Proj() &&
2192 ctrl->in(0)->is_MemBar()) {
2193 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2194 assert(membar->Opcode() == Op_MemBarReleaseLock &&
2195 mem->is_Proj() && membar == mem->in(0), "");
2196 _igvn.replace_node(fallthroughproj, ctrl);
2197 _igvn.replace_node(memproj_fallthrough, mem);
2198 fallthroughproj = ctrl;
2199 memproj_fallthrough = mem;
2200 ctrl = membar->in(TypeFunc::Control);
2201 mem = membar->in(TypeFunc::Memory);
2202 }
2203
2204 _igvn.replace_node(fallthroughproj, ctrl);
2205 _igvn.replace_node(memproj_fallthrough, mem);
2206 return true;
2207 }
2208
2209
2210 //------------------------------expand_lock_node----------------------
2211 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
2212
2213 Node* ctrl = lock->in(TypeFunc::Control);
2214 Node* mem = lock->in(TypeFunc::Memory);
2215 Node* obj = lock->obj_node();
2216 Node* box = lock->box_node();
2217 Node* flock = lock->fastlock_node();
2218
2219 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2220
2221 // Make the merge point
2222 Node *region;
2223 Node *mem_phi;
2224 Node *slow_path;
2225
2226 if (UseOptoBiasInlining) {
2227 /*
2228 * See the full description in MacroAssembler::biased_locking_enter().
2229 *
2230 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) {
2231 * // The object is biased.
2232 * proto_node = klass->prototype_header;
2233 * o_node = thread | proto_node;
2234 * x_node = o_node ^ mark_word;
2235 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ?
2236 * // Done.
2237 * } else {
2238 * if( (x_node & biased_lock_mask) != 0 ) {
2239 * // The klass's prototype header is no longer biased.
2240 * cas(&mark_word, mark_word, proto_node)
2241 * goto cas_lock;
2242 * } else {
2243 * // The klass's prototype header is still biased.
2244 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch?
2245 * old = mark_word;
2246 * new = o_node;
2247 * } else {
2248 * // Different thread or anonymous biased.
2249 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask);
2250 * new = thread | old;
2251 * }
2252 * // Try to rebias.
2253 * if( cas(&mark_word, old, new) == 0 ) {
2254 * // Done.
2255 * } else {
2256 * goto slow_path; // Failed.
2257 * }
2258 * }
2259 * }
2260 * } else {
2261 * // The object is not biased.
2262 * cas_lock:
2263 * if( FastLock(obj) == 0 ) {
2264 * // Done.
2265 * } else {
2266 * slow_path:
2267 * OptoRuntime::complete_monitor_locking_Java(obj);
2268 * }
2269 * }
2270 */
2271
2272 region = new RegionNode(5);
2273 // create a Phi for the memory state
2274 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2275
2276 Node* fast_lock_region = new RegionNode(3);
2277 Node* fast_lock_mem_phi = new PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM);
2278
2279 // First, check mark word for the biased lock pattern.
2280 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2281
2282 // Get fast path - mark word has the biased lock pattern.
2283 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node,
2284 markOopDesc::biased_lock_mask_in_place,
2285 markOopDesc::biased_lock_pattern, true);
2286 // fast_lock_region->in(1) is set to slow path.
2287 fast_lock_mem_phi->init_req(1, mem);
2288
2289 // Now check that the lock is biased to the current thread and has
2290 // the same epoch and bias as Klass::_prototype_header.
2291
2292 // Special-case a fresh allocation to avoid building nodes:
2293 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn);
2294 if (klass_node == NULL) {
2295 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
2296 klass_node = transform_later(LoadKlassNode::make(_igvn, NULL, mem, k_adr, _igvn.type(k_adr)->is_ptr()));
2297 #ifdef _LP64
2298 if (UseCompressedClassPointers && klass_node->is_DecodeNKlass()) {
2299 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity");
2300 klass_node->in(1)->init_req(0, ctrl);
2301 } else
2302 #endif
2303 klass_node->init_req(0, ctrl);
2304 }
2305 Node *proto_node = make_load(ctrl, mem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeX_X, TypeX_X->basic_type());
2306
2307 Node* thread = transform_later(new ThreadLocalNode());
2308 Node* cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2309 Node* o_node = transform_later(new OrXNode(cast_thread, proto_node));
2310 Node* x_node = transform_later(new XorXNode(o_node, mark_node));
2311
2312 // Get slow path - mark word does NOT match the value.
2313 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node,
2314 (~markOopDesc::age_mask_in_place), 0);
2315 // region->in(3) is set to fast path - the object is biased to the current thread.
2316 mem_phi->init_req(3, mem);
2317
2318
2319 // Mark word does NOT match the value (thread | Klass::_prototype_header).
2320
2321
2322 // First, check biased pattern.
2323 // Get fast path - _prototype_header has the same biased lock pattern.
2324 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node,
2325 markOopDesc::biased_lock_mask_in_place, 0, true);
2326
2327 not_biased_ctrl = fast_lock_region->in(2); // Slow path
2328 // fast_lock_region->in(2) - the prototype header is no longer biased
2329 // and we have to revoke the bias on this object.
2330 // We are going to try to reset the mark of this object to the prototype
2331 // value and fall through to the CAS-based locking scheme.
2332 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
2333 Node* cas = new StoreXConditionalNode(not_biased_ctrl, mem, adr,
2334 proto_node, mark_node);
2335 transform_later(cas);
2336 Node* proj = transform_later(new SCMemProjNode(cas));
2337 fast_lock_mem_phi->init_req(2, proj);
2338
2339
2340 // Second, check epoch bits.
2341 Node* rebiased_region = new RegionNode(3);
2342 Node* old_phi = new PhiNode( rebiased_region, TypeX_X);
2343 Node* new_phi = new PhiNode( rebiased_region, TypeX_X);
2344
2345 // Get slow path - mark word does NOT match epoch bits.
2346 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node,
2347 markOopDesc::epoch_mask_in_place, 0);
2348 // The epoch of the current bias is not valid, attempt to rebias the object
2349 // toward the current thread.
2350 rebiased_region->init_req(2, epoch_ctrl);
2351 old_phi->init_req(2, mark_node);
2352 new_phi->init_req(2, o_node);
2353
2354 // rebiased_region->in(1) is set to fast path.
2355 // The epoch of the current bias is still valid but we know
2356 // nothing about the owner; it might be set or it might be clear.
2357 Node* cmask = MakeConX(markOopDesc::biased_lock_mask_in_place |
2358 markOopDesc::age_mask_in_place |
2359 markOopDesc::epoch_mask_in_place);
2360 Node* old = transform_later(new AndXNode(mark_node, cmask));
2361 cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2362 Node* new_mark = transform_later(new OrXNode(cast_thread, old));
2363 old_phi->init_req(1, old);
2364 new_phi->init_req(1, new_mark);
2365
2366 transform_later(rebiased_region);
2367 transform_later(old_phi);
2368 transform_later(new_phi);
2369
2370 // Try to acquire the bias of the object using an atomic operation.
2371 // If this fails we will go in to the runtime to revoke the object's bias.
2372 cas = new StoreXConditionalNode(rebiased_region, mem, adr, new_phi, old_phi);
2373 transform_later(cas);
2374 proj = transform_later(new SCMemProjNode(cas));
2375
2376 // Get slow path - Failed to CAS.
2377 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0);
2378 mem_phi->init_req(4, proj);
2379 // region->in(4) is set to fast path - the object is rebiased to the current thread.
2380
2381 // Failed to CAS.
2382 slow_path = new RegionNode(3);
2383 Node *slow_mem = new PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM);
2384
2385 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control
2386 slow_mem->init_req(1, proj);
2387
2388 // Call CAS-based locking scheme (FastLock node).
2389
2390 transform_later(fast_lock_region);
2391 transform_later(fast_lock_mem_phi);
2392
2393 // Get slow path - FastLock failed to lock the object.
2394 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0);
2395 mem_phi->init_req(2, fast_lock_mem_phi);
2396 // region->in(2) is set to fast path - the object is locked to the current thread.
2397
2398 slow_path->init_req(2, ctrl); // Capture slow-control
2399 slow_mem->init_req(2, fast_lock_mem_phi);
2400
2401 transform_later(slow_path);
2402 transform_later(slow_mem);
2403 // Reset lock's memory edge.
2404 lock->set_req(TypeFunc::Memory, slow_mem);
2405
2406 } else {
2407 region = new RegionNode(3);
2408 // create a Phi for the memory state
2409 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2410
2411 // Optimize test; set region slot 2
2412 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2413 mem_phi->init_req(2, mem);
2414 }
2415
2416 // Make slow path call
2417 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(),
2418 OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path,
2419 obj, box, NULL);
2420
2421 extract_call_projections(call);
2422
2423 // Slow path can only throw asynchronous exceptions, which are always
2424 // de-opted. So the compiler thinks the slow-call can never throw an
2425 // exception. If it DOES throw an exception we would need the debug
2426 // info removed first (since if it throws there is no monitor).
2427 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2428 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2429
2430 // Capture slow path
2431 // disconnect fall-through projection from call and create a new one
2432 // hook up users of fall-through projection to region
2433 Node *slow_ctrl = _fallthroughproj->clone();
2434 transform_later(slow_ctrl);
2435 _igvn.hash_delete(_fallthroughproj);
2436 _fallthroughproj->disconnect_inputs(NULL, C);
2437 region->init_req(1, slow_ctrl);
2438 // region inputs are now complete
2439 transform_later(region);
2440 _igvn.replace_node(_fallthroughproj, region);
2441
2442 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2443 mem_phi->init_req(1, memproj );
2444 transform_later(mem_phi);
2445 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2446 }
2447
2448 //------------------------------expand_unlock_node----------------------
2449 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2450
2451 Node* ctrl = unlock->in(TypeFunc::Control);
2452 Node* mem = unlock->in(TypeFunc::Memory);
2453 Node* obj = unlock->obj_node();
2454 Node* box = unlock->box_node();
2455
2456 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2457
2458 // No need for a null check on unlock
2459
2460 // Make the merge point
2461 Node *region;
2462 Node *mem_phi;
2463
2464 if (UseOptoBiasInlining) {
2465 // Check for biased locking unlock case, which is a no-op.
2466 // See the full description in MacroAssembler::biased_locking_exit().
2467 region = new RegionNode(4);
2468 // create a Phi for the memory state
2469 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2470 mem_phi->init_req(3, mem);
2471
2472 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2473 ctrl = opt_bits_test(ctrl, region, 3, mark_node,
2474 markOopDesc::biased_lock_mask_in_place,
2475 markOopDesc::biased_lock_pattern);
2476 } else {
2477 region = new RegionNode(3);
2478 // create a Phi for the memory state
2479 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2480 }
2481
2482 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2483 funlock = transform_later( funlock )->as_FastUnlock();
2484 // Optimize test; set region slot 2
2485 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2486 Node *thread = transform_later(new ThreadLocalNode());
2487
2488 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2489 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2490 "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2491
2492 extract_call_projections(call);
2493
2494 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2495 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2496
2497 // No exceptions for unlocking
2498 // Capture slow path
2499 // disconnect fall-through projection from call and create a new one
2500 // hook up users of fall-through projection to region
2501 Node *slow_ctrl = _fallthroughproj->clone();
2502 transform_later(slow_ctrl);
2503 _igvn.hash_delete(_fallthroughproj);
2504 _fallthroughproj->disconnect_inputs(NULL, C);
2505 region->init_req(1, slow_ctrl);
2506 // region inputs are now complete
2507 transform_later(region);
2508 _igvn.replace_node(_fallthroughproj, region);
2509
2510 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2511 mem_phi->init_req(1, memproj );
2512 mem_phi->init_req(2, mem);
2513 transform_later(mem_phi);
2514 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2515 }
2516
2517 //---------------------------eliminate_macro_nodes----------------------
2518 // Eliminate scalar replaced allocations and associated locks.
2519 void PhaseMacroExpand::eliminate_macro_nodes() {
2520 if (C->macro_count() == 0)
2521 return;
2522
2523 // First, attempt to eliminate locks
2524 int cnt = C->macro_count();
2525 for (int i=0; i < cnt; i++) {
2526 Node *n = C->macro_node(i);
2527 if (n->is_AbstractLock()) { // Lock and Unlock nodes
2528 // Before elimination mark all associated (same box and obj)
2529 // lock and unlock nodes.
2530 mark_eliminated_locking_nodes(n->as_AbstractLock());
2531 }
2532 }
2533 bool progress = true;
2534 while (progress) {
2535 progress = false;
2536 for (int i = C->macro_count(); i > 0; i--) {
2537 Node * n = C->macro_node(i-1);
2538 bool success = false;
2539 debug_only(int old_macro_count = C->macro_count(););
2540 if (n->is_AbstractLock()) {
2541 success = eliminate_locking_node(n->as_AbstractLock());
2542 }
2543 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2544 progress = progress || success;
2545 }
2546 }
2547 // Next, attempt to eliminate allocations
2548 _has_locks = false;
2549 progress = true;
2550 while (progress) {
2551 progress = false;
2552 for (int i = C->macro_count(); i > 0; i--) {
2553 Node * n = C->macro_node(i-1);
2554 bool success = false;
2555 debug_only(int old_macro_count = C->macro_count(););
2556 switch (n->class_id()) {
2557 case Node::Class_Allocate:
2558 case Node::Class_AllocateArray:
2559 success = eliminate_allocate_node(n->as_Allocate());
2560 break;
2561 case Node::Class_CallStaticJava:
2562 success = eliminate_boxing_node(n->as_CallStaticJava());
2563 break;
2564 case Node::Class_Lock:
2565 case Node::Class_Unlock:
2566 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2567 _has_locks = true;
2568 break;
2569 case Node::Class_ArrayCopy:
2570 break;
2571 case Node::Class_OuterStripMinedLoop:
2572 break;
2573 default:
2574 assert(n->Opcode() == Op_LoopLimit ||
2575 n->Opcode() == Op_Opaque1 ||
2576 n->Opcode() == Op_Opaque2 ||
2577 n->Opcode() == Op_Opaque3, "unknown node type in macro list");
2578 }
2579 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2580 progress = progress || success;
2581 }
2582 }
2583 }
2584
2585 //------------------------------expand_macro_nodes----------------------
2586 // Returns true if a failure occurred.
2587 bool PhaseMacroExpand::expand_macro_nodes() {
2588 // Last attempt to eliminate macro nodes.
2589 eliminate_macro_nodes();
2590
2591 // Make sure expansion will not cause node limit to be exceeded.
2592 // Worst case is a macro node gets expanded into about 200 nodes.
2593 // Allow 50% more for optimization.
2594 if (C->check_node_count(C->macro_count() * 300, "out of nodes before macro expansion" ) )
2595 return true;
2596
2597 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2598 bool progress = true;
2599 while (progress) {
2600 progress = false;
2601 for (int i = C->macro_count(); i > 0; i--) {
2602 Node * n = C->macro_node(i-1);
2603 bool success = false;
2604 debug_only(int old_macro_count = C->macro_count(););
2605 if (n->Opcode() == Op_LoopLimit) {
2606 // Remove it from macro list and put on IGVN worklist to optimize.
2607 C->remove_macro_node(n);
2608 _igvn._worklist.push(n);
2609 success = true;
2610 } else if (n->Opcode() == Op_CallStaticJava) {
2611 // Remove it from macro list and put on IGVN worklist to optimize.
2612 C->remove_macro_node(n);
2613 _igvn._worklist.push(n);
2614 success = true;
2615 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) {
2616 _igvn.replace_node(n, n->in(1));
2617 success = true;
2618 #if INCLUDE_RTM_OPT
2619 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) {
2620 assert(C->profile_rtm(), "should be used only in rtm deoptimization code");
2621 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), "");
2622 Node* cmp = n->unique_out();
2623 #ifdef ASSERT
2624 // Validate graph.
2625 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), "");
2626 BoolNode* bol = cmp->unique_out()->as_Bool();
2627 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() &&
2628 (bol->_test._test == BoolTest::ne), "");
2629 IfNode* ifn = bol->unique_out()->as_If();
2630 assert((ifn->outcnt() == 2) &&
2631 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != NULL, "");
2632 #endif
2633 Node* repl = n->in(1);
2634 if (!_has_locks) {
2635 // Remove RTM state check if there are no locks in the code.
2636 // Replace input to compare the same value.
2637 repl = (cmp->in(1) == n) ? cmp->in(2) : cmp->in(1);
2638 }
2639 _igvn.replace_node(n, repl);
2640 success = true;
2641 #endif
2642 } else if (n->Opcode() == Op_OuterStripMinedLoop) {
2643 n->as_OuterStripMinedLoop()->adjust_strip_mined_loop(&_igvn);
2644 C->remove_macro_node(n);
2645 success = true;
2646 }
2647 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2648 progress = progress || success;
2649 }
2650 }
2651
2652 // expand arraycopy "macro" nodes first
2653 // For ReduceBulkZeroing, we must first process all arraycopy nodes
2654 // before the allocate nodes are expanded.
2655 int macro_idx = C->macro_count() - 1;
2656 while (macro_idx >= 0) {
2657 Node * n = C->macro_node(macro_idx);
2658 assert(n->is_macro(), "only macro nodes expected here");
2659 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) {
2660 // node is unreachable, so don't try to expand it
2661 C->remove_macro_node(n);
2662 } else if (n->is_ArrayCopy()){
2663 int macro_count = C->macro_count();
2664 expand_arraycopy_node(n->as_ArrayCopy());
2665 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2666 }
2667 if (C->failing()) return true;
2668 macro_idx --;
2669 }
2670
2671 // expand "macro" nodes
2672 // nodes are removed from the macro list as they are processed
2673 while (C->macro_count() > 0) {
2674 int macro_count = C->macro_count();
2675 Node * n = C->macro_node(macro_count-1);
2676 assert(n->is_macro(), "only macro nodes expected here");
2677 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) {
2678 // node is unreachable, so don't try to expand it
2679 C->remove_macro_node(n);
2680 continue;
2681 }
2682 switch (n->class_id()) {
2683 case Node::Class_Allocate:
2684 expand_allocate(n->as_Allocate());
2685 break;
2686 case Node::Class_AllocateArray:
2687 expand_allocate_array(n->as_AllocateArray());
2688 break;
2689 case Node::Class_Lock:
2690 expand_lock_node(n->as_Lock());
2691 break;
2692 case Node::Class_Unlock:
2693 expand_unlock_node(n->as_Unlock());
2694 break;
2695 default:
2696 assert(false, "unknown node type in macro list");
2697 }
2698 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2699 if (C->failing()) return true;
2700 }
2701
2702 _igvn.set_delay_transform(false);
2703 _igvn.optimize();
2704 if (C->failing()) return true;
2705 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2706 return bs->expand_macro_nodes(this);
2707 }