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 "opto/valuetypenode.hpp"
50 #include "runtime/sharedRuntime.hpp"
51 #if INCLUDE_G1GC
52 #include "gc/g1/g1ThreadLocalData.hpp"
53 #endif // INCLUDE_G1GC
54
55
56 //
57 // Replace any references to "oldref" in inputs to "use" with "newref".
58 // Returns the number of replacements made.
59 //
60 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
61 int nreplacements = 0;
62 uint req = use->req();
63 for (uint j = 0; j < use->len(); j++) {
64 Node *uin = use->in(j);
65 if (uin == oldref) {
66 if (j < req)
67 use->set_req(j, newref);
68 else
69 use->set_prec(j, newref);
70 nreplacements++;
71 } else if (j >= req && uin == NULL) {
72 break;
73 }
74 }
75 return nreplacements;
76 }
77
78 void PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) {
79 // Copy debug information and adjust JVMState information
80 uint old_dbg_start = oldcall->tf()->domain_sig()->cnt();
81 uint new_dbg_start = newcall->tf()->domain_sig()->cnt();
82 int jvms_adj = new_dbg_start - old_dbg_start;
83 assert (new_dbg_start == newcall->req(), "argument count mismatch");
84
85 // SafePointScalarObject node could be referenced several times in debug info.
86 // Use Dict to record cloned nodes.
87 Dict* sosn_map = new Dict(cmpkey,hashkey);
88 for (uint i = old_dbg_start; i < oldcall->req(); i++) {
89 Node* old_in = oldcall->in(i);
90 // Clone old SafePointScalarObjectNodes, adjusting their field contents.
91 if (old_in != NULL && old_in->is_SafePointScalarObject()) {
92 SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject();
93 uint old_unique = C->unique();
94 Node* new_in = old_sosn->clone(sosn_map);
95 if (old_unique != C->unique()) { // New node?
96 new_in->set_req(0, C->root()); // reset control edge
97 new_in = transform_later(new_in); // Register new node.
98 }
99 old_in = new_in;
100 }
101 newcall->add_req(old_in);
102 }
103
104 // JVMS may be shared so clone it before we modify it
105 newcall->set_jvms(oldcall->jvms() != NULL ? oldcall->jvms()->clone_deep(C) : NULL);
106 for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) {
107 jvms->set_map(newcall);
108 jvms->set_locoff(jvms->locoff()+jvms_adj);
109 jvms->set_stkoff(jvms->stkoff()+jvms_adj);
110 jvms->set_monoff(jvms->monoff()+jvms_adj);
111 jvms->set_scloff(jvms->scloff()+jvms_adj);
112 jvms->set_endoff(jvms->endoff()+jvms_adj);
113 }
114 }
115
116 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
117 Node* cmp;
118 if (mask != 0) {
119 Node* and_node = transform_later(new AndXNode(word, MakeConX(mask)));
120 cmp = transform_later(new CmpXNode(and_node, MakeConX(bits)));
121 } else {
122 cmp = word;
123 }
124 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne));
125 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
126 transform_later(iff);
127
128 // Fast path taken.
129 Node *fast_taken = transform_later(new IfFalseNode(iff));
130
131 // Fast path not-taken, i.e. slow path
132 Node *slow_taken = transform_later(new IfTrueNode(iff));
133
134 if (return_fast_path) {
135 region->init_req(edge, slow_taken); // Capture slow-control
136 return fast_taken;
137 } else {
138 region->init_req(edge, fast_taken); // Capture fast-control
139 return slow_taken;
140 }
141 }
142
143 //--------------------copy_predefined_input_for_runtime_call--------------------
144 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
145 // Set fixed predefined input arguments
146 call->init_req( TypeFunc::Control, ctrl );
147 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) );
148 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
149 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
150 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
151 }
152
153 //------------------------------make_slow_call---------------------------------
154 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type,
155 address slow_call, const char* leaf_name, Node* slow_path,
156 Node* parm0, Node* parm1, Node* parm2) {
157
158 // Slow-path call
159 CallNode *call = leaf_name
160 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
161 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM );
162
163 // Slow path call has no side-effects, uses few values
164 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
165 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0);
166 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1);
167 if (parm2 != NULL) call->init_req(TypeFunc::Parms+2, parm2);
168 copy_call_debug_info(oldcall, call);
169 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
170 _igvn.replace_node(oldcall, call);
171 transform_later(call);
172
173 return call;
174 }
175
176 void PhaseMacroExpand::extract_call_projections(CallNode *call) {
177 _fallthroughproj = NULL;
178 _fallthroughcatchproj = NULL;
179 _ioproj_fallthrough = NULL;
180 _ioproj_catchall = NULL;
181 _catchallcatchproj = NULL;
182 _memproj_fallthrough = NULL;
183 _memproj_catchall = NULL;
184 _resproj = NULL;
185 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) {
186 ProjNode *pn = call->fast_out(i)->as_Proj();
187 switch (pn->_con) {
188 case TypeFunc::Control:
189 {
190 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj
191 _fallthroughproj = pn;
192 DUIterator_Fast jmax, j = pn->fast_outs(jmax);
193 const Node *cn = pn->fast_out(j);
194 if (cn->is_Catch()) {
195 ProjNode *cpn = NULL;
196 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) {
197 cpn = cn->fast_out(k)->as_Proj();
198 assert(cpn->is_CatchProj(), "must be a CatchProjNode");
199 if (cpn->_con == CatchProjNode::fall_through_index)
200 _fallthroughcatchproj = cpn;
201 else {
202 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index.");
203 _catchallcatchproj = cpn;
204 }
205 }
206 }
207 break;
208 }
209 case TypeFunc::I_O:
210 if (pn->_is_io_use)
211 _ioproj_catchall = pn;
212 else
213 _ioproj_fallthrough = pn;
214 break;
215 case TypeFunc::Memory:
216 if (pn->_is_io_use)
217 _memproj_catchall = pn;
218 else
219 _memproj_fallthrough = pn;
220 break;
221 case TypeFunc::Parms:
222 _resproj = pn;
223 break;
224 default:
225 assert(false, "unexpected projection from allocation node.");
226 }
227 }
228
229 }
230
231 void PhaseMacroExpand::eliminate_gc_barrier(Node* p2x) {
232 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
233 bs->eliminate_gc_barrier(this, p2x);
234 }
235
236 // Search for a memory operation for the specified memory slice.
237 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
238 Node *orig_mem = mem;
239 Node *alloc_mem = alloc->in(TypeFunc::Memory);
240 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
241 while (true) {
242 if (mem == alloc_mem || mem == start_mem ) {
243 return mem; // hit one of our sentinels
244 } else if (mem->is_MergeMem()) {
245 mem = mem->as_MergeMem()->memory_at(alias_idx);
246 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
247 Node *in = mem->in(0);
248 // we can safely skip over safepoints, calls, locks and membars because we
249 // already know that the object is safe to eliminate.
250 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
251 return in;
252 } else if (in->is_Call()) {
253 CallNode *call = in->as_Call();
254 if (call->may_modify(tinst, phase)) {
255 assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape");
256 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) {
257 return in;
258 }
259 }
260 mem = in->in(TypeFunc::Memory);
261 } else if (in->is_MemBar()) {
262 ArrayCopyNode* ac = NULL;
263 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) {
264 assert(ac != NULL && ac->is_clonebasic(), "Only basic clone is a non escaping clone");
265 return ac;
266 }
267 mem = in->in(TypeFunc::Memory);
268 } else {
269 assert(false, "unexpected projection");
270 }
271 } else if (mem->is_Store()) {
272 const TypePtr* atype = mem->as_Store()->adr_type();
273 int adr_idx = phase->C->get_alias_index(atype);
274 if (adr_idx == alias_idx) {
275 assert(atype->isa_oopptr(), "address type must be oopptr");
276 int adr_offset = atype->flattened_offset();
277 uint adr_iid = atype->is_oopptr()->instance_id();
278 // Array elements references have the same alias_idx
279 // but different offset and different instance_id.
280 if (adr_offset == offset && adr_iid == alloc->_idx)
281 return mem;
282 } else {
283 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
284 }
285 mem = mem->in(MemNode::Memory);
286 } else if (mem->is_ClearArray()) {
287 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
288 // Can not bypass initialization of the instance
289 // we are looking.
290 debug_only(intptr_t offset;)
291 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
292 InitializeNode* init = alloc->as_Allocate()->initialization();
293 // We are looking for stored value, return Initialize node
294 // or memory edge from Allocate node.
295 if (init != NULL)
296 return init;
297 else
298 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers).
299 }
300 // Otherwise skip it (the call updated 'mem' value).
301 } else if (mem->Opcode() == Op_SCMemProj) {
302 mem = mem->in(0);
303 Node* adr = NULL;
304 if (mem->is_LoadStore()) {
305 adr = mem->in(MemNode::Address);
306 } else {
307 assert(mem->Opcode() == Op_EncodeISOArray ||
308 mem->Opcode() == Op_StrCompressedCopy, "sanity");
309 adr = mem->in(3); // Destination array
310 }
311 const TypePtr* atype = adr->bottom_type()->is_ptr();
312 int adr_idx = phase->C->get_alias_index(atype);
313 if (adr_idx == alias_idx) {
314 DEBUG_ONLY(mem->dump();)
315 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
316 return NULL;
317 }
318 mem = mem->in(MemNode::Memory);
319 } else if (mem->Opcode() == Op_StrInflatedCopy) {
320 Node* adr = mem->in(3); // Destination array
321 const TypePtr* atype = adr->bottom_type()->is_ptr();
322 int adr_idx = phase->C->get_alias_index(atype);
323 if (adr_idx == alias_idx) {
324 DEBUG_ONLY(mem->dump();)
325 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
326 return NULL;
327 }
328 mem = mem->in(MemNode::Memory);
329 } else {
330 return mem;
331 }
332 assert(mem != orig_mem, "dead memory loop");
333 }
334 }
335
336 // Generate loads from source of the arraycopy for fields of
337 // destination needed at a deoptimization point
338 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) {
339 BasicType bt = ft;
340 const Type *type = ftype;
341 if (ft == T_NARROWOOP) {
342 bt = T_OBJECT;
343 type = ftype->make_oopptr();
344 }
345 Node* res = NULL;
346 if (ac->is_clonebasic()) {
347 Node* base = ac->in(ArrayCopyNode::Src)->in(AddPNode::Base);
348 Node* adr = _igvn.transform(new AddPNode(base, base, MakeConX(offset)));
349 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
350 res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned);
351 } else {
352 if (ac->modifies(offset, offset, &_igvn, true)) {
353 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
354 uint shift = exact_log2(type2aelembytes(bt));
355 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
356 #ifdef _LP64
357 diff = _igvn.transform(new ConvI2LNode(diff));
358 #endif
359 diff = _igvn.transform(new LShiftXNode(diff, intcon(shift)));
360
361 Node* off = _igvn.transform(new AddXNode(MakeConX(offset), diff));
362 Node* base = ac->in(ArrayCopyNode::Src);
363 Node* adr = _igvn.transform(new AddPNode(base, base, off));
364 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
365 res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned);
366 }
367 }
368 if (res != NULL) {
369 res = _igvn.transform(res);
370 if (ftype->isa_narrowoop()) {
371 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
372 res = _igvn.transform(new EncodePNode(res, ftype));
373 }
374 return res;
375 }
376 return NULL;
377 }
378
379 //
380 // Given a Memory Phi, compute a value Phi containing the values from stores
381 // on the input paths.
382 // Note: this function is recursive, its depth is limited by the "level" argument
383 // Returns the computed Phi, or NULL if it cannot compute it.
384 Node *PhaseMacroExpand::value_from_mem_phi(Node *mem, BasicType ft, const Type *phi_type, const TypeOopPtr *adr_t, AllocateNode *alloc, Node_Stack *value_phis, int level) {
385 assert(mem->is_Phi(), "sanity");
386 int alias_idx = C->get_alias_index(adr_t);
387 int offset = adr_t->flattened_offset();
388 int instance_id = adr_t->instance_id();
389
390 // Check if an appropriate value phi already exists.
391 Node* region = mem->in(0);
392 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
393 Node* phi = region->fast_out(k);
394 if (phi->is_Phi() && phi != mem &&
395 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) {
396 return phi;
397 }
398 }
399 // Check if an appropriate new value phi already exists.
400 Node* new_phi = value_phis->find(mem->_idx);
401 if (new_phi != NULL)
402 return new_phi;
403
404 if (level <= 0) {
405 return NULL; // Give up: phi tree too deep
406 }
407 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
408 Node *alloc_mem = alloc->in(TypeFunc::Memory);
409
410 uint length = mem->req();
411 GrowableArray <Node *> values(length, length, NULL, false);
412
413 // create a new Phi for the value
414 PhiNode *phi = new PhiNode(mem->in(0), phi_type, NULL, mem->_idx, instance_id, alias_idx, offset);
415 transform_later(phi);
416 value_phis->push(phi, mem->_idx);
417
418 for (uint j = 1; j < length; j++) {
419 Node *in = mem->in(j);
420 if (in == NULL || in->is_top()) {
421 values.at_put(j, in);
422 } else {
423 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
424 if (val == start_mem || val == alloc_mem) {
425 // hit a sentinel, return appropriate 0 value
426 values.at_put(j, _igvn.zerocon(ft));
427 continue;
428 }
429 if (val->is_Initialize()) {
430 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
431 }
432 if (val == NULL) {
433 return NULL; // can't find a value on this path
434 }
435 if (val == mem) {
436 values.at_put(j, mem);
437 } else if (val->is_Store()) {
438 values.at_put(j, val->in(MemNode::ValueIn));
439 } else if(val->is_Proj() && val->in(0) == alloc) {
440 values.at_put(j, _igvn.zerocon(ft));
441 } else if (val->is_Phi()) {
442 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
443 if (val == NULL) {
444 return NULL;
445 }
446 values.at_put(j, val);
447 } else if (val->Opcode() == Op_SCMemProj) {
448 assert(val->in(0)->is_LoadStore() ||
449 val->in(0)->Opcode() == Op_EncodeISOArray ||
450 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity");
451 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
452 return NULL;
453 } else if (val->is_ArrayCopy()) {
454 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc);
455 if (res == NULL) {
456 return NULL;
457 }
458 values.at_put(j, res);
459 } else {
460 #ifdef ASSERT
461 val->dump();
462 assert(false, "unknown node on this path");
463 #endif
464 return NULL; // unknown node on this path
465 }
466 }
467 }
468 // Set Phi's inputs
469 for (uint j = 1; j < length; j++) {
470 if (values.at(j) == mem) {
471 phi->init_req(j, phi);
472 } else {
473 phi->init_req(j, values.at(j));
474 }
475 }
476 return phi;
477 }
478
479 // Search the last value stored into the object's field.
480 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) {
481 assert(adr_t->is_known_instance_field(), "instance required");
482 int instance_id = adr_t->instance_id();
483 assert((uint)instance_id == alloc->_idx, "wrong allocation");
484
485 int alias_idx = C->get_alias_index(adr_t);
486 int offset = adr_t->flattened_offset();
487 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
488 Node *alloc_ctrl = alloc->in(TypeFunc::Control);
489 Node *alloc_mem = alloc->in(TypeFunc::Memory);
490 Arena *a = Thread::current()->resource_area();
491 VectorSet visited(a);
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 went 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->flattened_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 went wrong.
576 return NULL;
577 }
578
579 // Search the last value stored into the value type's fields.
580 Node* PhaseMacroExpand::value_type_from_mem(Node* mem, Node* ctl, ciValueKlass* vk, const TypeAryPtr* adr_type, int offset, AllocateNode* alloc) {
581 // Subtract the offset of the first field to account for the missing oop header
582 offset -= vk->first_field_offset();
583 // Create a new ValueTypeNode and retrieve the field values from memory
584 ValueTypeNode* vt = ValueTypeNode::make_uninitialized(_igvn, vk)->as_ValueType();
585 for (int i = 0; i < vk->nof_declared_nonstatic_fields(); ++i) {
586 ciType* field_type = vt->field_type(i);
587 int field_offset = offset + vt->field_offset(i);
588 // Each value type field has its own memory slice
589 adr_type = adr_type->with_field_offset(field_offset);
590 Node* value = NULL;
591 if (vt->field_is_flattened(i)) {
592 value = value_type_from_mem(mem, ctl, field_type->as_value_klass(), adr_type, field_offset, alloc);
593 } else {
594 const Type* ft = Type::get_const_type(field_type);
595 BasicType bt = field_type->basic_type();
596 if (UseCompressedOops && !is_java_primitive(bt)) {
597 ft = ft->make_narrowoop();
598 bt = T_NARROWOOP;
599 }
600 value = value_from_mem(mem, ctl, bt, ft, adr_type, alloc);
601 if (value != NULL && ft->isa_narrowoop()) {
602 assert(UseCompressedOops, "unexpected narrow oop");
603 value = transform_later(new DecodeNNode(value, value->get_ptr_type()));
604 }
605 }
606 if (value != NULL) {
607 vt->set_field_value(i, value);
608 } else {
609 // We might have reached the TrackedInitializationLimit
610 return NULL;
611 }
612 }
613 return vt;
614 }
615
616 // Check the possibility of scalar replacement.
617 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
618 // Scan the uses of the allocation to check for anything that would
619 // prevent us from eliminating it.
620 NOT_PRODUCT( const char* fail_eliminate = NULL; )
621 DEBUG_ONLY( Node* disq_node = NULL; )
622 bool can_eliminate = true;
623
624 Node* res = alloc->result_cast();
625 const TypeOopPtr* res_type = NULL;
626 if (res == NULL) {
627 // All users were eliminated.
628 } else if (!res->is_CheckCastPP()) {
629 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
630 can_eliminate = false;
631 } else {
632 res_type = _igvn.type(res)->isa_oopptr();
633 if (res_type == NULL) {
634 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
635 can_eliminate = false;
636 } else if (res_type->isa_aryptr()) {
637 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
638 if (length < 0) {
639 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
640 can_eliminate = false;
641 }
642 }
643 }
644
645 if (can_eliminate && res != NULL) {
646 for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
647 j < jmax && can_eliminate; j++) {
648 Node* use = res->fast_out(j);
649
650 if (use->is_AddP()) {
651 const TypePtr* addp_type = _igvn.type(use)->is_ptr();
652 int offset = addp_type->offset();
653
654 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
655 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";)
656 can_eliminate = false;
657 break;
658 }
659 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
660 k < kmax && can_eliminate; k++) {
661 Node* n = use->fast_out(k);
662 if (!n->is_Store() && n->Opcode() != Op_CastP2X &&
663 !(n->is_ArrayCopy() &&
664 n->as_ArrayCopy()->is_clonebasic() &&
665 n->in(ArrayCopyNode::Dest) == use)) {
666 DEBUG_ONLY(disq_node = n;)
667 if (n->is_Load() || n->is_LoadStore()) {
668 NOT_PRODUCT(fail_eliminate = "Field load";)
669 } else {
670 NOT_PRODUCT(fail_eliminate = "Not store field reference";)
671 }
672 can_eliminate = false;
673 }
674 }
675 } else if (use->is_ArrayCopy() &&
676 (use->as_ArrayCopy()->is_arraycopy_validated() ||
677 use->as_ArrayCopy()->is_copyof_validated() ||
678 use->as_ArrayCopy()->is_copyofrange_validated()) &&
679 use->in(ArrayCopyNode::Dest) == res) {
680 // ok to eliminate
681 } else if (use->is_SafePoint()) {
682 SafePointNode* sfpt = use->as_SafePoint();
683 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
684 // Object is passed as argument.
685 DEBUG_ONLY(disq_node = use;)
686 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
687 can_eliminate = false;
688 }
689 Node* sfptMem = sfpt->memory();
690 if (sfptMem == NULL || sfptMem->is_top()) {
691 DEBUG_ONLY(disq_node = use;)
692 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";)
693 can_eliminate = false;
694 } else {
695 safepoints.append_if_missing(sfpt);
696 }
697 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
698 if (use->is_Phi()) {
699 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
700 NOT_PRODUCT(fail_eliminate = "Object is return value";)
701 } else {
702 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
703 }
704 DEBUG_ONLY(disq_node = use;)
705 } else {
706 if (use->Opcode() == Op_Return) {
707 NOT_PRODUCT(fail_eliminate = "Object is return value";)
708 } else {
709 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
710 }
711 DEBUG_ONLY(disq_node = use;)
712 }
713 can_eliminate = false;
714 } else {
715 assert(use->Opcode() == Op_CastP2X, "should be");
716 assert(!use->has_out_with(Op_OrL), "should have been removed because oop is never null");
717 }
718 }
719 }
720
721 #ifndef PRODUCT
722 if (PrintEliminateAllocations) {
723 if (can_eliminate) {
724 tty->print("Scalar ");
725 if (res == NULL)
726 alloc->dump();
727 else
728 res->dump();
729 } else if (alloc->_is_scalar_replaceable) {
730 tty->print("NotScalar (%s)", fail_eliminate);
731 if (res == NULL)
732 alloc->dump();
733 else
734 res->dump();
735 #ifdef ASSERT
736 if (disq_node != NULL) {
737 tty->print(" >>>> ");
738 disq_node->dump();
739 }
740 #endif /*ASSERT*/
741 }
742 }
743 #endif
744 return can_eliminate;
745 }
746
747 // Do scalar replacement.
748 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
749 GrowableArray <SafePointNode *> safepoints_done;
750
751 ciKlass* klass = NULL;
752 ciInstanceKlass* iklass = NULL;
753 int nfields = 0;
754 int array_base = 0;
755 int element_size = 0;
756 BasicType basic_elem_type = T_ILLEGAL;
757 ciType* elem_type = NULL;
758
759 Node* res = alloc->result_cast();
760 assert(res == NULL || res->is_CheckCastPP(), "unexpected AllocateNode result");
761 const TypeOopPtr* res_type = NULL;
762 if (res != NULL) { // Could be NULL when there are no users
763 res_type = _igvn.type(res)->isa_oopptr();
764 }
765
766 if (res != NULL) {
767 klass = res_type->klass();
768 if (res_type->isa_instptr()) {
769 // find the fields of the class which will be needed for safepoint debug information
770 assert(klass->is_instance_klass(), "must be an instance klass.");
771 iklass = klass->as_instance_klass();
772 nfields = iklass->nof_nonstatic_fields();
773 } else {
774 // find the array's elements which will be needed for safepoint debug information
775 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
776 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass.");
777 elem_type = klass->as_array_klass()->element_type();
778 basic_elem_type = elem_type->basic_type();
779 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
780 element_size = type2aelembytes(basic_elem_type);
781 if (klass->is_value_array_klass()) {
782 // Flattened value type array
783 element_size = klass->as_value_array_klass()->element_byte_size();
784 }
785 }
786 }
787 //
788 // Process the safepoint uses
789 //
790 Unique_Node_List value_worklist;
791 while (safepoints.length() > 0) {
792 SafePointNode* sfpt = safepoints.pop();
793 Node* mem = sfpt->memory();
794 Node* ctl = sfpt->control();
795 assert(sfpt->jvms() != NULL, "missed JVMS");
796 // Fields of scalar objs are referenced only at the end
797 // of regular debuginfo at the last (youngest) JVMS.
798 // Record relative start index.
799 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
800 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type,
801 #ifdef ASSERT
802 alloc,
803 #endif
804 first_ind, nfields);
805 sobj->init_req(0, C->root());
806 transform_later(sobj);
807
808 // Scan object's fields adding an input to the safepoint for each field.
809 for (int j = 0; j < nfields; j++) {
810 intptr_t offset;
811 ciField* field = NULL;
812 if (iklass != NULL) {
813 field = iklass->nonstatic_field_at(j);
814 offset = field->offset();
815 elem_type = field->type();
816 basic_elem_type = field->layout_type();
817 assert(!field->is_flattened(), "flattened value type fields should not have safepoint uses");
818 } else {
819 offset = array_base + j * (intptr_t)element_size;
820 }
821
822 const Type *field_type;
823 // The next code is taken from Parse::do_get_xxx().
824 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) {
825 if (!elem_type->is_loaded()) {
826 field_type = TypeInstPtr::BOTTOM;
827 } else if (field != NULL && field->is_static_constant()) {
828 // This can happen if the constant oop is non-perm.
829 ciObject* con = field->constant_value().as_object();
830 // Do not "join" in the previous type; it doesn't add value,
831 // and may yield a vacuous result if the field is of interface type.
832 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
833 assert(field_type != NULL, "field singleton type must be consistent");
834 } else {
835 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
836 }
837 if (UseCompressedOops) {
838 field_type = field_type->make_narrowoop();
839 basic_elem_type = T_NARROWOOP;
840 }
841 } else {
842 field_type = Type::get_const_basic_type(basic_elem_type);
843 }
844
845 Node* field_val = NULL;
846 const TypeOopPtr* field_addr_type = res_type->add_offset(offset)->isa_oopptr();
847 if (klass->is_value_array_klass()) {
848 ciValueKlass* vk = elem_type->as_value_klass();
849 assert(vk->flatten_array(), "must be flattened");
850 field_val = value_type_from_mem(mem, ctl, vk, field_addr_type->isa_aryptr(), 0, alloc);
851 } else {
852 field_val = value_from_mem(mem, ctl, basic_elem_type, field_type, field_addr_type, alloc);
853 }
854 if (field_val == NULL) {
855 // We weren't able to find a value for this field,
856 // give up on eliminating this allocation.
857
858 // Remove any extra entries we added to the safepoint.
859 uint last = sfpt->req() - 1;
860 for (int k = 0; k < j; k++) {
861 sfpt->del_req(last--);
862 }
863 _igvn._worklist.push(sfpt);
864 // rollback processed safepoints
865 while (safepoints_done.length() > 0) {
866 SafePointNode* sfpt_done = safepoints_done.pop();
867 // remove any extra entries we added to the safepoint
868 last = sfpt_done->req() - 1;
869 for (int k = 0; k < nfields; k++) {
870 sfpt_done->del_req(last--);
871 }
872 JVMState *jvms = sfpt_done->jvms();
873 jvms->set_endoff(sfpt_done->req());
874 // Now make a pass over the debug information replacing any references
875 // to SafePointScalarObjectNode with the allocated object.
876 int start = jvms->debug_start();
877 int end = jvms->debug_end();
878 for (int i = start; i < end; i++) {
879 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
880 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
881 if (scobj->first_index(jvms) == sfpt_done->req() &&
882 scobj->n_fields() == (uint)nfields) {
883 assert(scobj->alloc() == alloc, "sanity");
884 sfpt_done->set_req(i, res);
885 }
886 }
887 }
888 _igvn._worklist.push(sfpt_done);
889 }
890 #ifndef PRODUCT
891 if (PrintEliminateAllocations) {
892 if (field != NULL) {
893 tty->print("=== At SafePoint node %d can't find value of Field: ",
894 sfpt->_idx);
895 field->print();
896 int field_idx = C->get_alias_index(field_addr_type);
897 tty->print(" (alias_idx=%d)", field_idx);
898 } else { // Array's element
899 tty->print("=== At SafePoint node %d can't find value of array element [%d]",
900 sfpt->_idx, j);
901 }
902 tty->print(", which prevents elimination of: ");
903 if (res == NULL)
904 alloc->dump();
905 else
906 res->dump();
907 }
908 #endif
909 return false;
910 }
911 if (UseCompressedOops && field_type->isa_narrowoop()) {
912 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
913 // to be able scalar replace the allocation.
914 if (field_val->is_EncodeP()) {
915 field_val = field_val->in(1);
916 } else {
917 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
918 }
919 } else if (field_val->is_ValueType()) {
920 // Keep track of value types to scalarize them later
921 value_worklist.push(field_val);
922 }
923 sfpt->add_req(field_val);
924 }
925 JVMState *jvms = sfpt->jvms();
926 jvms->set_endoff(sfpt->req());
927 // Now make a pass over the debug information replacing any references
928 // to the allocated object with "sobj"
929 int start = jvms->debug_start();
930 int end = jvms->debug_end();
931 sfpt->replace_edges_in_range(res, sobj, start, end);
932 _igvn._worklist.push(sfpt);
933 safepoints_done.append_if_missing(sfpt); // keep it for rollback
934 }
935 // Scalarize value types that were added to the safepoint
936 for (uint i = 0; i < value_worklist.size(); ++i) {
937 Node* vt = value_worklist.at(i);
938 vt->as_ValueType()->make_scalar_in_safepoints(C->root(), &_igvn);
939 }
940 return true;
941 }
942
943 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
944 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control);
945 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory);
946 if (ctl_proj != NULL) {
947 igvn.replace_node(ctl_proj, n->in(0));
948 }
949 if (mem_proj != NULL) {
950 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
951 }
952 }
953
954 // Process users of eliminated allocation.
955 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) {
956 Node* res = alloc->result_cast();
957 if (res != NULL) {
958 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
959 Node *use = res->last_out(j);
960 uint oc1 = res->outcnt();
961
962 if (use->is_AddP()) {
963 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
964 Node *n = use->last_out(k);
965 uint oc2 = use->outcnt();
966 if (n->is_Store()) {
967 #ifdef ASSERT
968 // Verify that there is no dependent MemBarVolatile nodes,
969 // they should be removed during IGVN, see MemBarNode::Ideal().
970 for (DUIterator_Fast pmax, p = n->fast_outs(pmax);
971 p < pmax; p++) {
972 Node* mb = n->fast_out(p);
973 assert(mb->is_Initialize() || !mb->is_MemBar() ||
974 mb->req() <= MemBarNode::Precedent ||
975 mb->in(MemBarNode::Precedent) != n,
976 "MemBarVolatile should be eliminated for non-escaping object");
977 }
978 #endif
979 _igvn.replace_node(n, n->in(MemNode::Memory));
980 } else if (n->is_ArrayCopy()) {
981 // Disconnect ArrayCopy node
982 ArrayCopyNode* ac = n->as_ArrayCopy();
983 assert(ac->is_clonebasic(), "unexpected array copy kind");
984 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
985 disconnect_projections(ac, _igvn);
986 assert(alloc->in(0)->is_Proj() && alloc->in(0)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
987 Node* membar_before = alloc->in(0)->in(0);
988 disconnect_projections(membar_before->as_MemBar(), _igvn);
989 if (membar_after->is_MemBar()) {
990 disconnect_projections(membar_after->as_MemBar(), _igvn);
991 }
992 } else {
993 eliminate_gc_barrier(n);
994 }
995 k -= (oc2 - use->outcnt());
996 }
997 } else if (use->is_ArrayCopy()) {
998 // Disconnect ArrayCopy node
999 ArrayCopyNode* ac = use->as_ArrayCopy();
1000 assert(ac->is_arraycopy_validated() ||
1001 ac->is_copyof_validated() ||
1002 ac->is_copyofrange_validated(), "unsupported");
1003 CallProjections* callprojs = ac->extract_projections(true);
1004
1005 _igvn.replace_node(callprojs->fallthrough_ioproj, ac->in(TypeFunc::I_O));
1006 _igvn.replace_node(callprojs->fallthrough_memproj, ac->in(TypeFunc::Memory));
1007 _igvn.replace_node(callprojs->fallthrough_catchproj, ac->in(TypeFunc::Control));
1008
1009 // Set control to top. IGVN will remove the remaining projections
1010 ac->set_req(0, top());
1011 ac->replace_edge(res, top());
1012
1013 // Disconnect src right away: it can help find new
1014 // opportunities for allocation elimination
1015 Node* src = ac->in(ArrayCopyNode::Src);
1016 ac->replace_edge(src, top());
1017 // src can be top at this point if src and dest of the
1018 // arraycopy were the same
1019 if (src->outcnt() == 0 && !src->is_top()) {
1020 _igvn.remove_dead_node(src);
1021 }
1022
1023 _igvn._worklist.push(ac);
1024 } else {
1025 eliminate_gc_barrier(use);
1026 }
1027 j -= (oc1 - res->outcnt());
1028 }
1029 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
1030 _igvn.remove_dead_node(res);
1031 }
1032
1033 //
1034 // Process other users of allocation's projections
1035 //
1036 if (_resproj != NULL && _resproj->outcnt() != 0) {
1037 // First disconnect stores captured by Initialize node.
1038 // If Initialize node is eliminated first in the following code,
1039 // it will kill such stores and DUIterator_Last will assert.
1040 for (DUIterator_Fast jmax, j = _resproj->fast_outs(jmax); j < jmax; j++) {
1041 Node *use = _resproj->fast_out(j);
1042 if (use->is_AddP()) {
1043 // raw memory addresses used only by the initialization
1044 _igvn.replace_node(use, C->top());
1045 --j; --jmax;
1046 }
1047 }
1048 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) {
1049 Node *use = _resproj->last_out(j);
1050 uint oc1 = _resproj->outcnt();
1051 if (use->is_Initialize()) {
1052 // Eliminate Initialize node.
1053 InitializeNode *init = use->as_Initialize();
1054 assert(init->outcnt() <= 2, "only a control and memory projection expected");
1055 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control);
1056 if (ctrl_proj != NULL) {
1057 assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection");
1058 _igvn.replace_node(ctrl_proj, _fallthroughcatchproj);
1059 }
1060 Node *mem_proj = init->proj_out_or_null(TypeFunc::Memory);
1061 if (mem_proj != NULL) {
1062 Node *mem = init->in(TypeFunc::Memory);
1063 #ifdef ASSERT
1064 if (mem->is_MergeMem()) {
1065 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection");
1066 } else {
1067 assert(mem == _memproj_fallthrough, "allocation memory projection");
1068 }
1069 #endif
1070 _igvn.replace_node(mem_proj, mem);
1071 }
1072 } else {
1073 assert(false, "only Initialize or AddP expected");
1074 }
1075 j -= (oc1 - _resproj->outcnt());
1076 }
1077 }
1078 if (_fallthroughcatchproj != NULL) {
1079 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control));
1080 }
1081 if (_memproj_fallthrough != NULL) {
1082 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory));
1083 }
1084 if (_memproj_catchall != NULL) {
1085 _igvn.replace_node(_memproj_catchall, C->top());
1086 }
1087 if (_ioproj_fallthrough != NULL) {
1088 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O));
1089 }
1090 if (_ioproj_catchall != NULL) {
1091 _igvn.replace_node(_ioproj_catchall, C->top());
1092 }
1093 if (_catchallcatchproj != NULL) {
1094 _igvn.replace_node(_catchallcatchproj, C->top());
1095 }
1096 }
1097
1098 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1099 // Don't do scalar replacement if the frame can be popped by JVMTI:
1100 // if reallocation fails during deoptimization we'll pop all
1101 // interpreter frames for this compiled frame and that won't play
1102 // nice with JVMTI popframe.
1103 if (!EliminateAllocations || JvmtiExport::can_pop_frame() || !alloc->_is_non_escaping) {
1104 return false;
1105 }
1106 Node* klass = alloc->in(AllocateNode::KlassNode);
1107 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1108 Node* res = alloc->result_cast();
1109 // Eliminate boxing allocations which are not used
1110 // regardless scalar replacable status.
1111 bool boxing_alloc = C->eliminate_boxing() &&
1112 tklass->klass()->is_instance_klass() &&
1113 tklass->klass()->as_instance_klass()->is_box_klass();
1114 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != NULL))) {
1115 return false;
1116 }
1117
1118 extract_call_projections(alloc);
1119
1120 GrowableArray <SafePointNode *> safepoints;
1121 if (!can_eliminate_allocation(alloc, safepoints)) {
1122 return false;
1123 }
1124
1125 if (!alloc->_is_scalar_replaceable) {
1126 assert(res == NULL, "sanity");
1127 // We can only eliminate allocation if all debug info references
1128 // are already replaced with SafePointScalarObject because
1129 // we can't search for a fields value without instance_id.
1130 if (safepoints.length() > 0) {
1131 return false;
1132 }
1133 }
1134
1135 if (!scalar_replacement(alloc, safepoints)) {
1136 return false;
1137 }
1138
1139 CompileLog* log = C->log();
1140 if (log != NULL) {
1141 log->head("eliminate_allocation type='%d'",
1142 log->identify(tklass->klass()));
1143 JVMState* p = alloc->jvms();
1144 while (p != NULL) {
1145 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1146 p = p->caller();
1147 }
1148 log->tail("eliminate_allocation");
1149 }
1150
1151 process_users_of_allocation(alloc);
1152
1153 #ifndef PRODUCT
1154 if (PrintEliminateAllocations) {
1155 if (alloc->is_AllocateArray())
1156 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1157 else
1158 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1159 }
1160 #endif
1161
1162 return true;
1163 }
1164
1165 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1166 // EA should remove all uses of non-escaping boxing node.
1167 if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != NULL) {
1168 return false;
1169 }
1170
1171 assert(boxing->result_cast() == NULL, "unexpected boxing node result");
1172
1173 extract_call_projections(boxing);
1174
1175 const TypeTuple* r = boxing->tf()->range_sig();
1176 assert(r->cnt() > TypeFunc::Parms, "sanity");
1177 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1178 assert(t != NULL, "sanity");
1179
1180 CompileLog* log = C->log();
1181 if (log != NULL) {
1182 log->head("eliminate_boxing type='%d'",
1183 log->identify(t->klass()));
1184 JVMState* p = boxing->jvms();
1185 while (p != NULL) {
1186 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1187 p = p->caller();
1188 }
1189 log->tail("eliminate_boxing");
1190 }
1191
1192 process_users_of_allocation(boxing);
1193
1194 #ifndef PRODUCT
1195 if (PrintEliminateAllocations) {
1196 tty->print("++++ Eliminated: %d ", boxing->_idx);
1197 boxing->method()->print_short_name(tty);
1198 tty->cr();
1199 }
1200 #endif
1201
1202 return true;
1203 }
1204
1205 //---------------------------set_eden_pointers-------------------------
1206 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) {
1207 if (UseTLAB) { // Private allocation: load from TLS
1208 Node* thread = transform_later(new ThreadLocalNode());
1209 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset());
1210 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset());
1211 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset);
1212 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset);
1213 } else { // Shared allocation: load from globals
1214 CollectedHeap* ch = Universe::heap();
1215 address top_adr = (address)ch->top_addr();
1216 address end_adr = (address)ch->end_addr();
1217 eden_top_adr = makecon(TypeRawPtr::make(top_adr));
1218 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr);
1219 }
1220 }
1221
1222
1223 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
1224 Node* adr = basic_plus_adr(base, offset);
1225 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
1226 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered);
1227 transform_later(value);
1228 return value;
1229 }
1230
1231
1232 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
1233 Node* adr = basic_plus_adr(base, offset);
1234 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt, MemNode::unordered);
1235 transform_later(mem);
1236 return mem;
1237 }
1238
1239 //=============================================================================
1240 //
1241 // A L L O C A T I O N
1242 //
1243 // Allocation attempts to be fast in the case of frequent small objects.
1244 // It breaks down like this:
1245 //
1246 // 1) Size in doublewords is computed. This is a constant for objects and
1247 // variable for most arrays. Doubleword units are used to avoid size
1248 // overflow of huge doubleword arrays. We need doublewords in the end for
1249 // rounding.
1250 //
1251 // 2) Size is checked for being 'too large'. Too-large allocations will go
1252 // the slow path into the VM. The slow path can throw any required
1253 // exceptions, and does all the special checks for very large arrays. The
1254 // size test can constant-fold away for objects. For objects with
1255 // finalizers it constant-folds the otherway: you always go slow with
1256 // finalizers.
1257 //
1258 // 3) If NOT using TLABs, this is the contended loop-back point.
1259 // Load-Locked the heap top. If using TLABs normal-load the heap top.
1260 //
1261 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
1262 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
1263 // "size*8" we always enter the VM, where "largish" is a constant picked small
1264 // enough that there's always space between the eden max and 4Gig (old space is
1265 // there so it's quite large) and large enough that the cost of entering the VM
1266 // is dwarfed by the cost to initialize the space.
1267 //
1268 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1269 // down. If contended, repeat at step 3. If using TLABs normal-store
1270 // adjusted heap top back down; there is no contention.
1271 //
1272 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
1273 // fields.
1274 //
1275 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1276 // oop flavor.
1277 //
1278 //=============================================================================
1279 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1280 // Allocations bigger than this always go the slow route.
1281 // This value must be small enough that allocation attempts that need to
1282 // trigger exceptions go the slow route. Also, it must be small enough so
1283 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1284 //=============================================================================j//
1285 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1286 // The allocator will coalesce int->oop copies away. See comment in
1287 // coalesce.cpp about how this works. It depends critically on the exact
1288 // code shape produced here, so if you are changing this code shape
1289 // make sure the GC info for the heap-top is correct in and around the
1290 // slow-path call.
1291 //
1292
1293 void PhaseMacroExpand::expand_allocate_common(
1294 AllocateNode* alloc, // allocation node to be expanded
1295 Node* length, // array length for an array allocation
1296 const TypeFunc* slow_call_type, // Type of slow call
1297 address slow_call_address // Address of slow call
1298 )
1299 {
1300
1301 Node* ctrl = alloc->in(TypeFunc::Control);
1302 Node* mem = alloc->in(TypeFunc::Memory);
1303 Node* i_o = alloc->in(TypeFunc::I_O);
1304 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1305 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1306 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1307
1308 assert(ctrl != NULL, "must have control");
1309 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1310 // they will not be used if "always_slow" is set
1311 enum { slow_result_path = 1, fast_result_path = 2 };
1312 Node *result_region = NULL;
1313 Node *result_phi_rawmem = NULL;
1314 Node *result_phi_rawoop = NULL;
1315 Node *result_phi_i_o = NULL;
1316
1317 // The initial slow comparison is a size check, the comparison
1318 // we want to do is a BoolTest::gt
1319 bool always_slow = false;
1320 int tv = _igvn.find_int_con(initial_slow_test, -1);
1321 if (tv >= 0) {
1322 always_slow = (tv == 1);
1323 initial_slow_test = NULL;
1324 } else {
1325 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1326 }
1327
1328 if (C->env()->dtrace_alloc_probes() ||
1329 (!UseTLAB && !Universe::heap()->supports_inline_contig_alloc())) {
1330 // Force slow-path allocation
1331 always_slow = true;
1332 initial_slow_test = NULL;
1333 }
1334
1335 Node *slow_region = NULL;
1336 Node *toobig_false = ctrl;
1337
1338 if (!always_slow && (alloc->initialization() == NULL || alloc->initialization()->is_unknown_value())) {
1339 const TypeOopPtr* ary_type = _igvn.type(klass_node)->is_klassptr()->as_instance_type();
1340 const TypeAryPtr* ary_ptr = ary_type->isa_aryptr();
1341
1342 ciKlass* elem_klass = NULL;
1343 if (ary_ptr != NULL && ary_ptr->klass() != NULL) {
1344 elem_klass = ary_ptr->klass()->as_array_klass()->element_klass();
1345 }
1346
1347 if (elem_klass == NULL || (elem_klass->is_java_lang_Object() && !ary_ptr->klass_is_exact()) || elem_klass->is_valuetype()) {
1348 // If it's an array of values we must go to the slow path so it is
1349 // correctly initialized with default values.
1350 Node* fast_region = new RegionNode(3);
1351 Node* not_obj_array = ctrl;
1352 Node* obj_array = generate_object_array_guard(¬_obj_array, mem, klass_node, NULL);
1353
1354 fast_region->init_req(1, not_obj_array);
1355 slow_region = new RegionNode(1);
1356 Node* k_adr = basic_plus_adr(klass_node, klass_node, in_bytes(ArrayKlass::element_klass_offset()));
1357 Node* elem_klass = LoadKlassNode::make(_igvn, NULL, C->immutable_memory(), k_adr, TypeInstPtr::KLASS);
1358 transform_later(elem_klass);
1359 Node* flags = make_load(NULL, mem, elem_klass, in_bytes(Klass::access_flags_offset()), TypeInt::INT, T_INT);
1360 Node* is_value_elem = new AndINode(flags, intcon(JVM_ACC_VALUE));
1361 transform_later(is_value_elem);
1362 Node* cmp = new CmpINode(is_value_elem, _igvn.intcon(0));
1363 transform_later(cmp);
1364 Node* bol = new BoolNode(cmp, BoolTest::ne);
1365 transform_later(bol);
1366 IfNode* value_array_iff = new IfNode(obj_array, bol, PROB_MIN, COUNT_UNKNOWN);
1367 transform_later(value_array_iff);
1368 Node* value_array = new IfTrueNode(value_array_iff);
1369 transform_later(value_array);
1370 slow_region->add_req(value_array);
1371 Node* not_value_array = new IfFalseNode(value_array_iff);
1372 transform_later(not_value_array);
1373 fast_region->init_req(2, not_value_array);
1374 transform_later(fast_region);
1375 ctrl = fast_region;
1376 }
1377 InitializeNode* init = alloc->initialization();
1378 if (init != NULL) {
1379 init->clear_unknown_value();
1380 }
1381 }
1382
1383 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent");
1384 // generate the initial test if necessary
1385 if (initial_slow_test != NULL ) {
1386 if (slow_region == NULL) {
1387 slow_region = new RegionNode(1);
1388 }
1389 // Now make the initial failure test. Usually a too-big test but
1390 // might be a TRUE for finalizers or a fancy class check for
1391 // newInstance0.
1392 IfNode* toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1393 transform_later(toobig_iff);
1394 // Plug the failing-too-big test into the slow-path region
1395 Node* toobig_true = new IfTrueNode(toobig_iff);
1396 transform_later(toobig_true);
1397 slow_region ->add_req(toobig_true);
1398 toobig_false = new IfFalseNode(toobig_iff);
1399 transform_later(toobig_false);
1400 } else { // No initial test, just fall into next case
1401 toobig_false = ctrl;
1402 }
1403
1404 Node *slow_mem = mem; // save the current memory state for slow path
1405 // generate the fast allocation code unless we know that the initial test will always go slow
1406 if (!always_slow) {
1407 // Fast path modifies only raw memory.
1408 if (mem->is_MergeMem()) {
1409 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1410 }
1411
1412 Node* eden_top_adr;
1413 Node* eden_end_adr;
1414
1415 set_eden_pointers(eden_top_adr, eden_end_adr);
1416
1417 // Load Eden::end. Loop invariant and hoisted.
1418 //
1419 // Note: We set the control input on "eden_end" and "old_eden_top" when using
1420 // a TLAB to work around a bug where these values were being moved across
1421 // a safepoint. These are not oops, so they cannot be include in the oop
1422 // map, but they can be changed by a GC. The proper way to fix this would
1423 // be to set the raw memory state when generating a SafepointNode. However
1424 // this will require extensive changes to the loop optimization in order to
1425 // prevent a degradation of the optimization.
1426 // See comment in memnode.hpp, around line 227 in class LoadPNode.
1427 Node *eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS);
1428
1429 // allocate the Region and Phi nodes for the result
1430 result_region = new RegionNode(3);
1431 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1432 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1433 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1434
1435 // We need a Region for the loop-back contended case.
1436 enum { fall_in_path = 1, contended_loopback_path = 2 };
1437 Node *contended_region;
1438 Node *contended_phi_rawmem;
1439 if (UseTLAB) {
1440 contended_region = toobig_false;
1441 contended_phi_rawmem = mem;
1442 } else {
1443 contended_region = new RegionNode(3);
1444 contended_phi_rawmem = new PhiNode(contended_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1445 // Now handle the passing-too-big test. We fall into the contended
1446 // loop-back merge point.
1447 contended_region ->init_req(fall_in_path, toobig_false);
1448 contended_phi_rawmem->init_req(fall_in_path, mem);
1449 transform_later(contended_region);
1450 transform_later(contended_phi_rawmem);
1451 }
1452
1453 // Load(-locked) the heap top.
1454 // See note above concerning the control input when using a TLAB
1455 Node *old_eden_top = UseTLAB
1456 ? new LoadPNode (ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered)
1457 : new LoadPLockedNode(contended_region, contended_phi_rawmem, eden_top_adr, MemNode::acquire);
1458
1459 transform_later(old_eden_top);
1460 // Add to heap top to get a new heap top
1461 Node *new_eden_top = new AddPNode(top(), old_eden_top, size_in_bytes);
1462 transform_later(new_eden_top);
1463 // Check for needing a GC; compare against heap end
1464 Node *needgc_cmp = new CmpPNode(new_eden_top, eden_end);
1465 transform_later(needgc_cmp);
1466 Node *needgc_bol = new BoolNode(needgc_cmp, BoolTest::ge);
1467 transform_later(needgc_bol);
1468 IfNode *needgc_iff = new IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN);
1469 transform_later(needgc_iff);
1470
1471 // Plug the failing-heap-space-need-gc test into the slow-path region
1472 Node *needgc_true = new IfTrueNode(needgc_iff);
1473 transform_later(needgc_true);
1474 if (slow_region != NULL) {
1475 slow_region->add_req(needgc_true);
1476 // This completes all paths into the slow merge point
1477 transform_later(slow_region);
1478 } else {
1479 // Just fall from the need-GC path straight into the VM call.
1480 slow_region = needgc_true;
1481 }
1482 // No need for a GC. Setup for the Store-Conditional
1483 Node *needgc_false = new IfFalseNode(needgc_iff);
1484 transform_later(needgc_false);
1485
1486 // Grab regular I/O before optional prefetch may change it.
1487 // Slow-path does no I/O so just set it to the original I/O.
1488 result_phi_i_o->init_req(slow_result_path, i_o);
1489
1490 i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem,
1491 old_eden_top, new_eden_top, length);
1492
1493 // Name successful fast-path variables
1494 Node* fast_oop = old_eden_top;
1495 Node* fast_oop_ctrl;
1496 Node* fast_oop_rawmem;
1497
1498 // Store (-conditional) the modified eden top back down.
1499 // StorePConditional produces flags for a test PLUS a modified raw
1500 // memory state.
1501 if (UseTLAB) {
1502 Node* store_eden_top =
1503 new StorePNode(needgc_false, contended_phi_rawmem, eden_top_adr,
1504 TypeRawPtr::BOTTOM, new_eden_top, MemNode::unordered);
1505 transform_later(store_eden_top);
1506 fast_oop_ctrl = needgc_false; // No contention, so this is the fast path
1507 fast_oop_rawmem = store_eden_top;
1508 } else {
1509 Node* store_eden_top =
1510 new StorePConditionalNode(needgc_false, contended_phi_rawmem, eden_top_adr,
1511 new_eden_top, fast_oop/*old_eden_top*/);
1512 transform_later(store_eden_top);
1513 Node *contention_check = new BoolNode(store_eden_top, BoolTest::ne);
1514 transform_later(contention_check);
1515 store_eden_top = new SCMemProjNode(store_eden_top);
1516 transform_later(store_eden_top);
1517
1518 // If not using TLABs, check to see if there was contention.
1519 IfNode *contention_iff = new IfNode (needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN);
1520 transform_later(contention_iff);
1521 Node *contention_true = new IfTrueNode(contention_iff);
1522 transform_later(contention_true);
1523 // If contention, loopback and try again.
1524 contended_region->init_req(contended_loopback_path, contention_true);
1525 contended_phi_rawmem->init_req(contended_loopback_path, store_eden_top);
1526
1527 // Fast-path succeeded with no contention!
1528 Node *contention_false = new IfFalseNode(contention_iff);
1529 transform_later(contention_false);
1530 fast_oop_ctrl = contention_false;
1531
1532 // Bump total allocated bytes for this thread
1533 Node* thread = new ThreadLocalNode();
1534 transform_later(thread);
1535 Node* alloc_bytes_adr = basic_plus_adr(top()/*not oop*/, thread,
1536 in_bytes(JavaThread::allocated_bytes_offset()));
1537 Node* alloc_bytes = make_load(fast_oop_ctrl, store_eden_top, alloc_bytes_adr,
1538 0, TypeLong::LONG, T_LONG);
1539 #ifdef _LP64
1540 Node* alloc_size = size_in_bytes;
1541 #else
1542 Node* alloc_size = new ConvI2LNode(size_in_bytes);
1543 transform_later(alloc_size);
1544 #endif
1545 Node* new_alloc_bytes = new AddLNode(alloc_bytes, alloc_size);
1546 transform_later(new_alloc_bytes);
1547 fast_oop_rawmem = make_store(fast_oop_ctrl, store_eden_top, alloc_bytes_adr,
1548 0, new_alloc_bytes, T_LONG);
1549 }
1550
1551 InitializeNode* init = alloc->initialization();
1552 fast_oop_rawmem = initialize_object(alloc,
1553 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1554 klass_node, length, size_in_bytes);
1555
1556 // If initialization is performed by an array copy, any required
1557 // MemBarStoreStore was already added. If the object does not
1558 // escape no need for a MemBarStoreStore. If the object does not
1559 // escape in its initializer and memory barrier (MemBarStoreStore or
1560 // stronger) is already added at exit of initializer, also no need
1561 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore
1562 // so that stores that initialize this object can't be reordered
1563 // with a subsequent store that makes this object accessible by
1564 // other threads.
1565 // Other threads include java threads and JVM internal threads
1566 // (for example concurrent GC threads). Current concurrent GC
1567 // implementation: CMS and G1 will not scan newly created object,
1568 // so it's safe to skip storestore barrier when allocation does
1569 // not escape.
1570 if (!alloc->does_not_escape_thread() &&
1571 !alloc->is_allocation_MemBar_redundant() &&
1572 (init == NULL || !init->is_complete_with_arraycopy())) {
1573 if (init == NULL || init->req() < InitializeNode::RawStores) {
1574 // No InitializeNode or no stores captured by zeroing
1575 // elimination. Simply add the MemBarStoreStore after object
1576 // initialization.
1577 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1578 transform_later(mb);
1579
1580 mb->init_req(TypeFunc::Memory, fast_oop_rawmem);
1581 mb->init_req(TypeFunc::Control, fast_oop_ctrl);
1582 fast_oop_ctrl = new ProjNode(mb,TypeFunc::Control);
1583 transform_later(fast_oop_ctrl);
1584 fast_oop_rawmem = new ProjNode(mb,TypeFunc::Memory);
1585 transform_later(fast_oop_rawmem);
1586 } else {
1587 // Add the MemBarStoreStore after the InitializeNode so that
1588 // all stores performing the initialization that were moved
1589 // before the InitializeNode happen before the storestore
1590 // barrier.
1591
1592 Node* init_ctrl = init->proj_out_or_null(TypeFunc::Control);
1593 Node* init_mem = init->proj_out_or_null(TypeFunc::Memory);
1594
1595 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1596 transform_later(mb);
1597
1598 Node* ctrl = new ProjNode(init,TypeFunc::Control);
1599 transform_later(ctrl);
1600 Node* mem = new ProjNode(init,TypeFunc::Memory);
1601 transform_later(mem);
1602
1603 // The MemBarStoreStore depends on control and memory coming
1604 // from the InitializeNode
1605 mb->init_req(TypeFunc::Memory, mem);
1606 mb->init_req(TypeFunc::Control, ctrl);
1607
1608 ctrl = new ProjNode(mb,TypeFunc::Control);
1609 transform_later(ctrl);
1610 mem = new ProjNode(mb,TypeFunc::Memory);
1611 transform_later(mem);
1612
1613 // All nodes that depended on the InitializeNode for control
1614 // and memory must now depend on the MemBarNode that itself
1615 // depends on the InitializeNode
1616 if (init_ctrl != NULL) {
1617 _igvn.replace_node(init_ctrl, ctrl);
1618 }
1619 if (init_mem != NULL) {
1620 _igvn.replace_node(init_mem, mem);
1621 }
1622 }
1623 }
1624
1625 if (C->env()->dtrace_extended_probes()) {
1626 // Slow-path call
1627 int size = TypeFunc::Parms + 2;
1628 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1629 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base),
1630 "dtrace_object_alloc",
1631 TypeRawPtr::BOTTOM);
1632
1633 // Get base of thread-local storage area
1634 Node* thread = new ThreadLocalNode();
1635 transform_later(thread);
1636
1637 call->init_req(TypeFunc::Parms+0, thread);
1638 call->init_req(TypeFunc::Parms+1, fast_oop);
1639 call->init_req(TypeFunc::Control, fast_oop_ctrl);
1640 call->init_req(TypeFunc::I_O , top()); // does no i/o
1641 call->init_req(TypeFunc::Memory , fast_oop_rawmem);
1642 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1643 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1644 transform_later(call);
1645 fast_oop_ctrl = new ProjNode(call,TypeFunc::Control);
1646 transform_later(fast_oop_ctrl);
1647 fast_oop_rawmem = new ProjNode(call,TypeFunc::Memory);
1648 transform_later(fast_oop_rawmem);
1649 }
1650
1651 // Plug in the successful fast-path into the result merge point
1652 result_region ->init_req(fast_result_path, fast_oop_ctrl);
1653 result_phi_rawoop->init_req(fast_result_path, fast_oop);
1654 result_phi_i_o ->init_req(fast_result_path, i_o);
1655 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1656 } else {
1657 slow_region = ctrl;
1658 result_phi_i_o = i_o; // Rename it to use in the following code.
1659 }
1660
1661 // Generate slow-path call
1662 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1663 OptoRuntime::stub_name(slow_call_address),
1664 alloc->jvms()->bci(),
1665 TypePtr::BOTTOM);
1666 call->init_req( TypeFunc::Control, slow_region );
1667 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o
1668 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs
1669 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) );
1670 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) );
1671
1672 call->init_req(TypeFunc::Parms+0, klass_node);
1673 if (length != NULL) {
1674 call->init_req(TypeFunc::Parms+1, length);
1675 }
1676
1677 // Copy debug information and adjust JVMState information, then replace
1678 // allocate node with the call
1679 copy_call_debug_info((CallNode *) alloc, call);
1680 if (!always_slow) {
1681 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1682 } else {
1683 // Hook i_o projection to avoid its elimination during allocation
1684 // replacement (when only a slow call is generated).
1685 call->set_req(TypeFunc::I_O, result_phi_i_o);
1686 }
1687 _igvn.replace_node(alloc, call);
1688 transform_later(call);
1689
1690 // Identify the output projections from the allocate node and
1691 // adjust any references to them.
1692 // The control and io projections look like:
1693 //
1694 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1695 // Allocate Catch
1696 // ^---Proj(io) <-------+ ^---CatchProj(io)
1697 //
1698 // We are interested in the CatchProj nodes.
1699 //
1700 extract_call_projections(call);
1701
1702 // An allocate node has separate memory projections for the uses on
1703 // the control and i_o paths. Replace the control memory projection with
1704 // result_phi_rawmem (unless we are only generating a slow call when
1705 // both memory projections are combined)
1706 if (!always_slow && _memproj_fallthrough != NULL) {
1707 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) {
1708 Node *use = _memproj_fallthrough->fast_out(i);
1709 _igvn.rehash_node_delayed(use);
1710 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem);
1711 // back up iterator
1712 --i;
1713 }
1714 }
1715 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete
1716 // _memproj_catchall so we end up with a call that has only 1 memory projection.
1717 if (_memproj_catchall != NULL ) {
1718 if (_memproj_fallthrough == NULL) {
1719 _memproj_fallthrough = new ProjNode(call, TypeFunc::Memory);
1720 transform_later(_memproj_fallthrough);
1721 }
1722 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) {
1723 Node *use = _memproj_catchall->fast_out(i);
1724 _igvn.rehash_node_delayed(use);
1725 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough);
1726 // back up iterator
1727 --i;
1728 }
1729 assert(_memproj_catchall->outcnt() == 0, "all uses must be deleted");
1730 _igvn.remove_dead_node(_memproj_catchall);
1731 }
1732
1733 // An allocate node has separate i_o projections for the uses on the control
1734 // and i_o paths. Always replace the control i_o projection with result i_o
1735 // otherwise incoming i_o become dead when only a slow call is generated
1736 // (it is different from memory projections where both projections are
1737 // combined in such case).
1738 if (_ioproj_fallthrough != NULL) {
1739 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) {
1740 Node *use = _ioproj_fallthrough->fast_out(i);
1741 _igvn.rehash_node_delayed(use);
1742 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o);
1743 // back up iterator
1744 --i;
1745 }
1746 }
1747 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete
1748 // _ioproj_catchall so we end up with a call that has only 1 i_o projection.
1749 if (_ioproj_catchall != NULL ) {
1750 if (_ioproj_fallthrough == NULL) {
1751 _ioproj_fallthrough = new ProjNode(call, TypeFunc::I_O);
1752 transform_later(_ioproj_fallthrough);
1753 }
1754 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) {
1755 Node *use = _ioproj_catchall->fast_out(i);
1756 _igvn.rehash_node_delayed(use);
1757 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough);
1758 // back up iterator
1759 --i;
1760 }
1761 assert(_ioproj_catchall->outcnt() == 0, "all uses must be deleted");
1762 _igvn.remove_dead_node(_ioproj_catchall);
1763 }
1764
1765 // if we generated only a slow call, we are done
1766 if (always_slow) {
1767 // Now we can unhook i_o.
1768 if (result_phi_i_o->outcnt() > 1) {
1769 call->set_req(TypeFunc::I_O, top());
1770 } else {
1771 assert(result_phi_i_o->unique_ctrl_out() == call, "");
1772 // Case of new array with negative size known during compilation.
1773 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1774 // following code since call to runtime will throw exception.
1775 // As result there will be no users of i_o after the call.
1776 // Leave i_o attached to this call to avoid problems in preceding graph.
1777 }
1778 return;
1779 }
1780
1781
1782 if (_fallthroughcatchproj != NULL) {
1783 ctrl = _fallthroughcatchproj->clone();
1784 transform_later(ctrl);
1785 _igvn.replace_node(_fallthroughcatchproj, result_region);
1786 } else {
1787 ctrl = top();
1788 }
1789 Node *slow_result;
1790 if (_resproj == NULL) {
1791 // no uses of the allocation result
1792 slow_result = top();
1793 } else {
1794 slow_result = _resproj->clone();
1795 transform_later(slow_result);
1796 _igvn.replace_node(_resproj, result_phi_rawoop);
1797 }
1798
1799 // Plug slow-path into result merge point
1800 result_region ->init_req( slow_result_path, ctrl );
1801 result_phi_rawoop->init_req( slow_result_path, slow_result);
1802 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough );
1803 transform_later(result_region);
1804 transform_later(result_phi_rawoop);
1805 transform_later(result_phi_rawmem);
1806 transform_later(result_phi_i_o);
1807 // This completes all paths into the result merge point
1808 }
1809
1810
1811 // Helper for PhaseMacroExpand::expand_allocate_common.
1812 // Initializes the newly-allocated storage.
1813 Node* PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1814 Node* control, Node* rawmem, Node* object,
1815 Node* klass_node, Node* length,
1816 Node* size_in_bytes) {
1817 InitializeNode* init = alloc->initialization();
1818 // Store the klass & mark bits
1819 Node* mark_node = NULL;
1820 // For now only enable fast locking for non-array types
1821 if ((EnableValhalla || UseBiasedLocking) && length == NULL) {
1822 mark_node = make_load(control, rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
1823 } else {
1824 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype()));
1825 }
1826 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
1827
1828 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1829 int header_size = alloc->minimum_header_size(); // conservatively small
1830
1831 // Array length
1832 if (length != NULL) { // Arrays need length field
1833 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1834 // conservatively small header size:
1835 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1836 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1837 if (k->is_array_klass()) // we know the exact header size in most cases:
1838 header_size = Klass::layout_helper_header_size(k->layout_helper());
1839 }
1840
1841 // Clear the object body, if necessary.
1842 if (init == NULL) {
1843 // The init has somehow disappeared; be cautious and clear everything.
1844 //
1845 // This can happen if a node is allocated but an uncommon trap occurs
1846 // immediately. In this case, the Initialize gets associated with the
1847 // trap, and may be placed in a different (outer) loop, if the Allocate
1848 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1849 // there can be two Allocates to one Initialize. The answer in all these
1850 // edge cases is safety first. It is always safe to clear immediately
1851 // within an Allocate, and then (maybe or maybe not) clear some more later.
1852 if (!(UseTLAB && ZeroTLAB)) {
1853 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1854 header_size, size_in_bytes,
1855 &_igvn);
1856 }
1857 } else {
1858 if (!init->is_complete()) {
1859 // Try to win by zeroing only what the init does not store.
1860 // We can also try to do some peephole optimizations,
1861 // such as combining some adjacent subword stores.
1862 rawmem = init->complete_stores(control, rawmem, object,
1863 header_size, size_in_bytes, &_igvn);
1864 }
1865 // We have no more use for this link, since the AllocateNode goes away:
1866 init->set_req(InitializeNode::RawAddress, top());
1867 // (If we keep the link, it just confuses the register allocator,
1868 // who thinks he sees a real use of the address by the membar.)
1869 }
1870
1871 return rawmem;
1872 }
1873
1874 // Generate prefetch instructions for next allocations.
1875 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
1876 Node*& contended_phi_rawmem,
1877 Node* old_eden_top, Node* new_eden_top,
1878 Node* length) {
1879 enum { fall_in_path = 1, pf_path = 2 };
1880 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
1881 // Generate prefetch allocation with watermark check.
1882 // As an allocation hits the watermark, we will prefetch starting
1883 // at a "distance" away from watermark.
1884
1885 Node *pf_region = new RegionNode(3);
1886 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY,
1887 TypeRawPtr::BOTTOM );
1888 // I/O is used for Prefetch
1889 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO );
1890
1891 Node *thread = new ThreadLocalNode();
1892 transform_later(thread);
1893
1894 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread,
1895 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
1896 transform_later(eden_pf_adr);
1897
1898 Node *old_pf_wm = new LoadPNode(needgc_false,
1899 contended_phi_rawmem, eden_pf_adr,
1900 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM,
1901 MemNode::unordered);
1902 transform_later(old_pf_wm);
1903
1904 // check against new_eden_top
1905 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm );
1906 transform_later(need_pf_cmp);
1907 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge );
1908 transform_later(need_pf_bol);
1909 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol,
1910 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1911 transform_later(need_pf_iff);
1912
1913 // true node, add prefetchdistance
1914 Node *need_pf_true = new IfTrueNode( need_pf_iff );
1915 transform_later(need_pf_true);
1916
1917 Node *need_pf_false = new IfFalseNode( need_pf_iff );
1918 transform_later(need_pf_false);
1919
1920 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm,
1921 _igvn.MakeConX(AllocatePrefetchDistance) );
1922 transform_later(new_pf_wmt );
1923 new_pf_wmt->set_req(0, need_pf_true);
1924
1925 Node *store_new_wmt = new StorePNode(need_pf_true,
1926 contended_phi_rawmem, eden_pf_adr,
1927 TypeRawPtr::BOTTOM, new_pf_wmt,
1928 MemNode::unordered);
1929 transform_later(store_new_wmt);
1930
1931 // adding prefetches
1932 pf_phi_abio->init_req( fall_in_path, i_o );
1933
1934 Node *prefetch_adr;
1935 Node *prefetch;
1936 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1937 uint step_size = AllocatePrefetchStepSize;
1938 uint distance = 0;
1939
1940 for ( uint i = 0; i < lines; i++ ) {
1941 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt,
1942 _igvn.MakeConX(distance) );
1943 transform_later(prefetch_adr);
1944 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1945 transform_later(prefetch);
1946 distance += step_size;
1947 i_o = prefetch;
1948 }
1949 pf_phi_abio->set_req( pf_path, i_o );
1950
1951 pf_region->init_req( fall_in_path, need_pf_false );
1952 pf_region->init_req( pf_path, need_pf_true );
1953
1954 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
1955 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
1956
1957 transform_later(pf_region);
1958 transform_later(pf_phi_rawmem);
1959 transform_later(pf_phi_abio);
1960
1961 needgc_false = pf_region;
1962 contended_phi_rawmem = pf_phi_rawmem;
1963 i_o = pf_phi_abio;
1964 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) {
1965 // Insert a prefetch instruction for each allocation.
1966 // This code is used to generate 1 prefetch instruction per cache line.
1967
1968 // Generate several prefetch instructions.
1969 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1970 uint step_size = AllocatePrefetchStepSize;
1971 uint distance = AllocatePrefetchDistance;
1972
1973 // Next cache address.
1974 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top,
1975 _igvn.MakeConX(step_size + distance));
1976 transform_later(cache_adr);
1977 cache_adr = new CastP2XNode(needgc_false, cache_adr);
1978 transform_later(cache_adr);
1979 // Address is aligned to execute prefetch to the beginning of cache line size
1980 // (it is important when BIS instruction is used on SPARC as prefetch).
1981 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1));
1982 cache_adr = new AndXNode(cache_adr, mask);
1983 transform_later(cache_adr);
1984 cache_adr = new CastX2PNode(cache_adr);
1985 transform_later(cache_adr);
1986
1987 // Prefetch
1988 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr );
1989 prefetch->set_req(0, needgc_false);
1990 transform_later(prefetch);
1991 contended_phi_rawmem = prefetch;
1992 Node *prefetch_adr;
1993 distance = step_size;
1994 for ( uint i = 1; i < lines; i++ ) {
1995 prefetch_adr = new AddPNode( cache_adr, cache_adr,
1996 _igvn.MakeConX(distance) );
1997 transform_later(prefetch_adr);
1998 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr );
1999 transform_later(prefetch);
2000 distance += step_size;
2001 contended_phi_rawmem = prefetch;
2002 }
2003 } else if( AllocatePrefetchStyle > 0 ) {
2004 // Insert a prefetch for each allocation only on the fast-path
2005 Node *prefetch_adr;
2006 Node *prefetch;
2007 // Generate several prefetch instructions.
2008 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
2009 uint step_size = AllocatePrefetchStepSize;
2010 uint distance = AllocatePrefetchDistance;
2011 for ( uint i = 0; i < lines; i++ ) {
2012 prefetch_adr = new AddPNode( old_eden_top, new_eden_top,
2013 _igvn.MakeConX(distance) );
2014 transform_later(prefetch_adr);
2015 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
2016 // Do not let it float too high, since if eden_top == eden_end,
2017 // both might be null.
2018 if( i == 0 ) { // Set control for first prefetch, next follows it
2019 prefetch->init_req(0, needgc_false);
2020 }
2021 transform_later(prefetch);
2022 distance += step_size;
2023 i_o = prefetch;
2024 }
2025 }
2026 return i_o;
2027 }
2028
2029
2030 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
2031 expand_allocate_common(alloc, NULL,
2032 OptoRuntime::new_instance_Type(),
2033 OptoRuntime::new_instance_Java());
2034 }
2035
2036 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
2037 Node* length = alloc->in(AllocateNode::ALength);
2038 InitializeNode* init = alloc->initialization();
2039 Node* klass_node = alloc->in(AllocateNode::KlassNode);
2040 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
2041 address slow_call_address; // Address of slow call
2042 if (init != NULL && init->is_complete_with_arraycopy() &&
2043 k->is_type_array_klass()) {
2044 // Don't zero type array during slow allocation in VM since
2045 // it will be initialized later by arraycopy in compiled code.
2046 slow_call_address = OptoRuntime::new_array_nozero_Java();
2047 } else {
2048 slow_call_address = OptoRuntime::new_array_Java();
2049 }
2050 expand_allocate_common(alloc, length,
2051 OptoRuntime::new_array_Type(),
2052 slow_call_address);
2053 }
2054
2055 //-------------------mark_eliminated_box----------------------------------
2056 //
2057 // During EA obj may point to several objects but after few ideal graph
2058 // transformations (CCP) it may point to only one non escaping object
2059 // (but still using phi), corresponding locks and unlocks will be marked
2060 // for elimination. Later obj could be replaced with a new node (new phi)
2061 // and which does not have escape information. And later after some graph
2062 // reshape other locks and unlocks (which were not marked for elimination
2063 // before) are connected to this new obj (phi) but they still will not be
2064 // marked for elimination since new obj has no escape information.
2065 // Mark all associated (same box and obj) lock and unlock nodes for
2066 // elimination if some of them marked already.
2067 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) {
2068 if (oldbox->as_BoxLock()->is_eliminated())
2069 return; // This BoxLock node was processed already.
2070
2071 // New implementation (EliminateNestedLocks) has separate BoxLock
2072 // node for each locked region so mark all associated locks/unlocks as
2073 // eliminated even if different objects are referenced in one locked region
2074 // (for example, OSR compilation of nested loop inside locked scope).
2075 if (EliminateNestedLocks ||
2076 oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj)) {
2077 // Box is used only in one lock region. Mark this box as eliminated.
2078 _igvn.hash_delete(oldbox);
2079 oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value
2080 _igvn.hash_insert(oldbox);
2081
2082 for (uint i = 0; i < oldbox->outcnt(); i++) {
2083 Node* u = oldbox->raw_out(i);
2084 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) {
2085 AbstractLockNode* alock = u->as_AbstractLock();
2086 // Check lock's box since box could be referenced by Lock's debug info.
2087 if (alock->box_node() == oldbox) {
2088 // Mark eliminated all related locks and unlocks.
2089 #ifdef ASSERT
2090 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4");
2091 #endif
2092 alock->set_non_esc_obj();
2093 }
2094 }
2095 }
2096 return;
2097 }
2098
2099 // Create new "eliminated" BoxLock node and use it in monitor debug info
2100 // instead of oldbox for the same object.
2101 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
2102
2103 // Note: BoxLock node is marked eliminated only here and it is used
2104 // to indicate that all associated lock and unlock nodes are marked
2105 // for elimination.
2106 newbox->set_eliminated();
2107 transform_later(newbox);
2108
2109 // Replace old box node with new box for all users of the same object.
2110 for (uint i = 0; i < oldbox->outcnt();) {
2111 bool next_edge = true;
2112
2113 Node* u = oldbox->raw_out(i);
2114 if (u->is_AbstractLock()) {
2115 AbstractLockNode* alock = u->as_AbstractLock();
2116 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) {
2117 // Replace Box and mark eliminated all related locks and unlocks.
2118 #ifdef ASSERT
2119 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5");
2120 #endif
2121 alock->set_non_esc_obj();
2122 _igvn.rehash_node_delayed(alock);
2123 alock->set_box_node(newbox);
2124 next_edge = false;
2125 }
2126 }
2127 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) {
2128 FastLockNode* flock = u->as_FastLock();
2129 assert(flock->box_node() == oldbox, "sanity");
2130 _igvn.rehash_node_delayed(flock);
2131 flock->set_box_node(newbox);
2132 next_edge = false;
2133 }
2134
2135 // Replace old box in monitor debug info.
2136 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
2137 SafePointNode* sfn = u->as_SafePoint();
2138 JVMState* youngest_jvms = sfn->jvms();
2139 int max_depth = youngest_jvms->depth();
2140 for (int depth = 1; depth <= max_depth; depth++) {
2141 JVMState* jvms = youngest_jvms->of_depth(depth);
2142 int num_mon = jvms->nof_monitors();
2143 // Loop over monitors
2144 for (int idx = 0; idx < num_mon; idx++) {
2145 Node* obj_node = sfn->monitor_obj(jvms, idx);
2146 Node* box_node = sfn->monitor_box(jvms, idx);
2147 if (box_node == oldbox && obj_node->eqv_uncast(obj)) {
2148 int j = jvms->monitor_box_offset(idx);
2149 _igvn.replace_input_of(u, j, newbox);
2150 next_edge = false;
2151 }
2152 }
2153 }
2154 }
2155 if (next_edge) i++;
2156 }
2157 }
2158
2159 //-----------------------mark_eliminated_locking_nodes-----------------------
2160 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) {
2161 if (EliminateNestedLocks) {
2162 if (alock->is_nested()) {
2163 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity");
2164 return;
2165 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened
2166 // Only Lock node has JVMState needed here.
2167 // Not that preceding claim is documented anywhere else.
2168 if (alock->jvms() != NULL) {
2169 if (alock->as_Lock()->is_nested_lock_region()) {
2170 // Mark eliminated related nested locks and unlocks.
2171 Node* obj = alock->obj_node();
2172 BoxLockNode* box_node = alock->box_node()->as_BoxLock();
2173 assert(!box_node->is_eliminated(), "should not be marked yet");
2174 // Note: BoxLock node is marked eliminated only here
2175 // and it is used to indicate that all associated lock
2176 // and unlock nodes are marked for elimination.
2177 box_node->set_eliminated(); // Box's hash is always NO_HASH here
2178 for (uint i = 0; i < box_node->outcnt(); i++) {
2179 Node* u = box_node->raw_out(i);
2180 if (u->is_AbstractLock()) {
2181 alock = u->as_AbstractLock();
2182 if (alock->box_node() == box_node) {
2183 // Verify that this Box is referenced only by related locks.
2184 assert(alock->obj_node()->eqv_uncast(obj), "");
2185 // Mark all related locks and unlocks.
2186 #ifdef ASSERT
2187 alock->log_lock_optimization(C, "eliminate_lock_set_nested");
2188 #endif
2189 alock->set_nested();
2190 }
2191 }
2192 }
2193 } else {
2194 #ifdef ASSERT
2195 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region");
2196 if (C->log() != NULL)
2197 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output
2198 #endif
2199 }
2200 }
2201 return;
2202 }
2203 // Process locks for non escaping object
2204 assert(alock->is_non_esc_obj(), "");
2205 } // EliminateNestedLocks
2206
2207 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2208 // Look for all locks of this object and mark them and
2209 // corresponding BoxLock nodes as eliminated.
2210 Node* obj = alock->obj_node();
2211 for (uint j = 0; j < obj->outcnt(); j++) {
2212 Node* o = obj->raw_out(j);
2213 if (o->is_AbstractLock() &&
2214 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2215 alock = o->as_AbstractLock();
2216 Node* box = alock->box_node();
2217 // Replace old box node with new eliminated box for all users
2218 // of the same object and mark related locks as eliminated.
2219 mark_eliminated_box(box, obj);
2220 }
2221 }
2222 }
2223 }
2224
2225 // we have determined that this lock/unlock can be eliminated, we simply
2226 // eliminate the node without expanding it.
2227 //
2228 // Note: The membar's associated with the lock/unlock are currently not
2229 // eliminated. This should be investigated as a future enhancement.
2230 //
2231 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2232
2233 if (!alock->is_eliminated()) {
2234 return false;
2235 }
2236 #ifdef ASSERT
2237 if (!alock->is_coarsened()) {
2238 // Check that new "eliminated" BoxLock node is created.
2239 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2240 assert(oldbox->is_eliminated(), "should be done already");
2241 }
2242 #endif
2243
2244 alock->log_lock_optimization(C, "eliminate_lock");
2245
2246 #ifndef PRODUCT
2247 if (PrintEliminateLocks) {
2248 if (alock->is_Lock()) {
2249 tty->print_cr("++++ Eliminated: %d Lock", alock->_idx);
2250 } else {
2251 tty->print_cr("++++ Eliminated: %d Unlock", alock->_idx);
2252 }
2253 }
2254 #endif
2255
2256 Node* mem = alock->in(TypeFunc::Memory);
2257 Node* ctrl = alock->in(TypeFunc::Control);
2258 guarantee(ctrl != NULL, "missing control projection, cannot replace_node() with NULL");
2259
2260 extract_call_projections(alock);
2261 // There are 2 projections from the lock. The lock node will
2262 // be deleted when its last use is subsumed below.
2263 assert(alock->outcnt() == 2 &&
2264 _fallthroughproj != NULL &&
2265 _memproj_fallthrough != NULL,
2266 "Unexpected projections from Lock/Unlock");
2267
2268 Node* fallthroughproj = _fallthroughproj;
2269 Node* memproj_fallthrough = _memproj_fallthrough;
2270
2271 // The memory projection from a lock/unlock is RawMem
2272 // The input to a Lock is merged memory, so extract its RawMem input
2273 // (unless the MergeMem has been optimized away.)
2274 if (alock->is_Lock()) {
2275 // Seach for MemBarAcquireLock node and delete it also.
2276 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2277 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, "");
2278 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2279 Node* memproj = membar->proj_out(TypeFunc::Memory);
2280 _igvn.replace_node(ctrlproj, fallthroughproj);
2281 _igvn.replace_node(memproj, memproj_fallthrough);
2282
2283 // Delete FastLock node also if this Lock node is unique user
2284 // (a loop peeling may clone a Lock node).
2285 Node* flock = alock->as_Lock()->fastlock_node();
2286 if (flock->outcnt() == 1) {
2287 assert(flock->unique_out() == alock, "sanity");
2288 _igvn.replace_node(flock, top());
2289 }
2290 }
2291
2292 // Seach for MemBarReleaseLock node and delete it also.
2293 if (alock->is_Unlock() && ctrl->is_Proj() && ctrl->in(0)->is_MemBar()) {
2294 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2295 assert(membar->Opcode() == Op_MemBarReleaseLock &&
2296 mem->is_Proj() && membar == mem->in(0), "");
2297 _igvn.replace_node(fallthroughproj, ctrl);
2298 _igvn.replace_node(memproj_fallthrough, mem);
2299 fallthroughproj = ctrl;
2300 memproj_fallthrough = mem;
2301 ctrl = membar->in(TypeFunc::Control);
2302 mem = membar->in(TypeFunc::Memory);
2303 }
2304
2305 _igvn.replace_node(fallthroughproj, ctrl);
2306 _igvn.replace_node(memproj_fallthrough, mem);
2307 return true;
2308 }
2309
2310
2311 //------------------------------expand_lock_node----------------------
2312 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
2313
2314 Node* ctrl = lock->in(TypeFunc::Control);
2315 Node* mem = lock->in(TypeFunc::Memory);
2316 Node* obj = lock->obj_node();
2317 Node* box = lock->box_node();
2318 Node* flock = lock->fastlock_node();
2319
2320 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2321
2322 // Make the merge point
2323 Node *region;
2324 Node *mem_phi;
2325 Node *slow_path;
2326
2327 if (UseOptoBiasInlining) {
2328 /*
2329 * See the full description in MacroAssembler::biased_locking_enter().
2330 *
2331 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) {
2332 * // The object is biased.
2333 * proto_node = klass->prototype_header;
2334 * o_node = thread | proto_node;
2335 * x_node = o_node ^ mark_word;
2336 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ?
2337 * // Done.
2338 * } else {
2339 * if( (x_node & biased_lock_mask) != 0 ) {
2340 * // The klass's prototype header is no longer biased.
2341 * cas(&mark_word, mark_word, proto_node)
2342 * goto cas_lock;
2343 * } else {
2344 * // The klass's prototype header is still biased.
2345 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch?
2346 * old = mark_word;
2347 * new = o_node;
2348 * } else {
2349 * // Different thread or anonymous biased.
2350 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask);
2351 * new = thread | old;
2352 * }
2353 * // Try to rebias.
2354 * if( cas(&mark_word, old, new) == 0 ) {
2355 * // Done.
2356 * } else {
2357 * goto slow_path; // Failed.
2358 * }
2359 * }
2360 * }
2361 * } else {
2362 * // The object is not biased.
2363 * cas_lock:
2364 * if( FastLock(obj) == 0 ) {
2365 * // Done.
2366 * } else {
2367 * slow_path:
2368 * OptoRuntime::complete_monitor_locking_Java(obj);
2369 * }
2370 * }
2371 */
2372
2373 region = new RegionNode(5);
2374 // create a Phi for the memory state
2375 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2376
2377 Node* fast_lock_region = new RegionNode(3);
2378 Node* fast_lock_mem_phi = new PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM);
2379
2380 // First, check mark word for the biased lock pattern.
2381 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2382
2383 // Get fast path - mark word has the biased lock pattern.
2384 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node,
2385 markOopDesc::biased_lock_mask_in_place,
2386 markOopDesc::biased_lock_pattern, true);
2387 // fast_lock_region->in(1) is set to slow path.
2388 fast_lock_mem_phi->init_req(1, mem);
2389
2390 // Now check that the lock is biased to the current thread and has
2391 // the same epoch and bias as Klass::_prototype_header.
2392
2393 // Special-case a fresh allocation to avoid building nodes:
2394 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn);
2395 if (klass_node == NULL) {
2396 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
2397 klass_node = transform_later(LoadKlassNode::make(_igvn, NULL, mem, k_adr, _igvn.type(k_adr)->is_ptr()));
2398 #ifdef _LP64
2399 if (UseCompressedClassPointers && klass_node->is_DecodeNKlass()) {
2400 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity");
2401 klass_node->in(1)->init_req(0, ctrl);
2402 } else
2403 #endif
2404 klass_node->init_req(0, ctrl);
2405 }
2406 Node *proto_node = make_load(ctrl, mem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeX_X, TypeX_X->basic_type());
2407
2408 Node* thread = transform_later(new ThreadLocalNode());
2409 Node* cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2410 Node* o_node = transform_later(new OrXNode(cast_thread, proto_node));
2411 Node* x_node = transform_later(new XorXNode(o_node, mark_node));
2412
2413 // Get slow path - mark word does NOT match the value.
2414 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node,
2415 (~markOopDesc::age_mask_in_place), 0);
2416 // region->in(3) is set to fast path - the object is biased to the current thread.
2417 mem_phi->init_req(3, mem);
2418
2419
2420 // Mark word does NOT match the value (thread | Klass::_prototype_header).
2421
2422
2423 // First, check biased pattern.
2424 // Get fast path - _prototype_header has the same biased lock pattern.
2425 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node,
2426 markOopDesc::biased_lock_mask_in_place, 0, true);
2427
2428 not_biased_ctrl = fast_lock_region->in(2); // Slow path
2429 // fast_lock_region->in(2) - the prototype header is no longer biased
2430 // and we have to revoke the bias on this object.
2431 // We are going to try to reset the mark of this object to the prototype
2432 // value and fall through to the CAS-based locking scheme.
2433 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
2434 Node* cas = new StoreXConditionalNode(not_biased_ctrl, mem, adr,
2435 proto_node, mark_node);
2436 transform_later(cas);
2437 Node* proj = transform_later(new SCMemProjNode(cas));
2438 fast_lock_mem_phi->init_req(2, proj);
2439
2440
2441 // Second, check epoch bits.
2442 Node* rebiased_region = new RegionNode(3);
2443 Node* old_phi = new PhiNode( rebiased_region, TypeX_X);
2444 Node* new_phi = new PhiNode( rebiased_region, TypeX_X);
2445
2446 // Get slow path - mark word does NOT match epoch bits.
2447 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node,
2448 markOopDesc::epoch_mask_in_place, 0);
2449 // The epoch of the current bias is not valid, attempt to rebias the object
2450 // toward the current thread.
2451 rebiased_region->init_req(2, epoch_ctrl);
2452 old_phi->init_req(2, mark_node);
2453 new_phi->init_req(2, o_node);
2454
2455 // rebiased_region->in(1) is set to fast path.
2456 // The epoch of the current bias is still valid but we know
2457 // nothing about the owner; it might be set or it might be clear.
2458 Node* cmask = MakeConX(markOopDesc::biased_lock_mask_in_place |
2459 markOopDesc::age_mask_in_place |
2460 markOopDesc::epoch_mask_in_place);
2461 Node* old = transform_later(new AndXNode(mark_node, cmask));
2462 cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2463 Node* new_mark = transform_later(new OrXNode(cast_thread, old));
2464 old_phi->init_req(1, old);
2465 new_phi->init_req(1, new_mark);
2466
2467 transform_later(rebiased_region);
2468 transform_later(old_phi);
2469 transform_later(new_phi);
2470
2471 // Try to acquire the bias of the object using an atomic operation.
2472 // If this fails we will go in to the runtime to revoke the object's bias.
2473 cas = new StoreXConditionalNode(rebiased_region, mem, adr, new_phi, old_phi);
2474 transform_later(cas);
2475 proj = transform_later(new SCMemProjNode(cas));
2476
2477 // Get slow path - Failed to CAS.
2478 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0);
2479 mem_phi->init_req(4, proj);
2480 // region->in(4) is set to fast path - the object is rebiased to the current thread.
2481
2482 // Failed to CAS.
2483 slow_path = new RegionNode(3);
2484 Node *slow_mem = new PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM);
2485
2486 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control
2487 slow_mem->init_req(1, proj);
2488
2489 // Call CAS-based locking scheme (FastLock node).
2490
2491 transform_later(fast_lock_region);
2492 transform_later(fast_lock_mem_phi);
2493
2494 // Get slow path - FastLock failed to lock the object.
2495 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0);
2496 mem_phi->init_req(2, fast_lock_mem_phi);
2497 // region->in(2) is set to fast path - the object is locked to the current thread.
2498
2499 slow_path->init_req(2, ctrl); // Capture slow-control
2500 slow_mem->init_req(2, fast_lock_mem_phi);
2501
2502 transform_later(slow_path);
2503 transform_later(slow_mem);
2504 // Reset lock's memory edge.
2505 lock->set_req(TypeFunc::Memory, slow_mem);
2506
2507 } else {
2508 region = new RegionNode(3);
2509 // create a Phi for the memory state
2510 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2511
2512 // Optimize test; set region slot 2
2513 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2514 mem_phi->init_req(2, mem);
2515 }
2516
2517 // Make slow path call
2518 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(),
2519 OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path,
2520 obj, box, NULL);
2521
2522 extract_call_projections(call);
2523
2524 // Slow path can only throw asynchronous exceptions, which are always
2525 // de-opted. So the compiler thinks the slow-call can never throw an
2526 // exception. If it DOES throw an exception we would need the debug
2527 // info removed first (since if it throws there is no monitor).
2528 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2529 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2530
2531 // Capture slow path
2532 // disconnect fall-through projection from call and create a new one
2533 // hook up users of fall-through projection to region
2534 Node *slow_ctrl = _fallthroughproj->clone();
2535 transform_later(slow_ctrl);
2536 _igvn.hash_delete(_fallthroughproj);
2537 _fallthroughproj->disconnect_inputs(NULL, C);
2538 region->init_req(1, slow_ctrl);
2539 // region inputs are now complete
2540 transform_later(region);
2541 _igvn.replace_node(_fallthroughproj, region);
2542
2543 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2544 mem_phi->init_req(1, memproj );
2545 transform_later(mem_phi);
2546 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2547 }
2548
2549 //------------------------------expand_unlock_node----------------------
2550 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2551
2552 Node* ctrl = unlock->in(TypeFunc::Control);
2553 Node* mem = unlock->in(TypeFunc::Memory);
2554 Node* obj = unlock->obj_node();
2555 Node* box = unlock->box_node();
2556
2557 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2558
2559 // No need for a null check on unlock
2560
2561 // Make the merge point
2562 Node *region;
2563 Node *mem_phi;
2564
2565 if (UseOptoBiasInlining) {
2566 // Check for biased locking unlock case, which is a no-op.
2567 // See the full description in MacroAssembler::biased_locking_exit().
2568 region = new RegionNode(4);
2569 // create a Phi for the memory state
2570 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2571 mem_phi->init_req(3, mem);
2572
2573 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2574 ctrl = opt_bits_test(ctrl, region, 3, mark_node,
2575 markOopDesc::biased_lock_mask_in_place,
2576 markOopDesc::biased_lock_pattern);
2577 } else {
2578 region = new RegionNode(3);
2579 // create a Phi for the memory state
2580 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2581 }
2582
2583 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2584 funlock = transform_later( funlock )->as_FastUnlock();
2585 // Optimize test; set region slot 2
2586 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2587 Node *thread = transform_later(new ThreadLocalNode());
2588
2589 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2590 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2591 "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2592
2593 extract_call_projections(call);
2594
2595 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2596 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2597
2598 // No exceptions for unlocking
2599 // Capture slow path
2600 // disconnect fall-through projection from call and create a new one
2601 // hook up users of fall-through projection to region
2602 Node *slow_ctrl = _fallthroughproj->clone();
2603 transform_later(slow_ctrl);
2604 _igvn.hash_delete(_fallthroughproj);
2605 _fallthroughproj->disconnect_inputs(NULL, C);
2606 region->init_req(1, slow_ctrl);
2607 // region inputs are now complete
2608 transform_later(region);
2609 _igvn.replace_node(_fallthroughproj, region);
2610
2611 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2612 mem_phi->init_req(1, memproj );
2613 mem_phi->init_req(2, mem);
2614 transform_later(mem_phi);
2615 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2616 }
2617
2618 // A value type is returned from the call but we don't know its
2619 // type. Either we get a buffered value (and nothing needs to be done)
2620 // or one of the values being returned is the klass of the value type
2621 // and we need to allocate a value type instance of that type and
2622 // initialize it with other values being returned. In that case, we
2623 // first try a fast path allocation and initialize the value with the
2624 // value klass's pack handler or we fall back to a runtime call.
2625 void PhaseMacroExpand::expand_mh_intrinsic_return(CallStaticJavaNode* call) {
2626 Node* ret = call->proj_out(TypeFunc::Parms);
2627 if (ret == NULL) {
2628 return;
2629 }
2630 // TODO fix this with the calling convention changes
2631 //assert(ret->bottom_type()->is_valuetypeptr()->is__Value(), "unexpected return type from MH intrinsic");
2632 const TypeFunc* tf = call->_tf;
2633 const TypeTuple* domain = OptoRuntime::store_value_type_fields_Type()->domain_cc();
2634 const TypeFunc* new_tf = TypeFunc::make(tf->domain_sig(), tf->domain_cc(), tf->range_sig(), domain);
2635 call->_tf = new_tf;
2636 // Make sure the change of type is applied before projections are
2637 // processed by igvn
2638 _igvn.set_type(call, call->Value(&_igvn));
2639 _igvn.set_type(ret, ret->Value(&_igvn));
2640
2641 // Before any new projection is added:
2642 CallProjections* projs = call->extract_projections(true, true);
2643
2644 Node* ctl = new Node(1);
2645 Node* mem = new Node(1);
2646 Node* io = new Node(1);
2647 Node* ex_ctl = new Node(1);
2648 Node* ex_mem = new Node(1);
2649 Node* ex_io = new Node(1);
2650 Node* res = new Node(1);
2651
2652 Node* cast = transform_later(new CastP2XNode(ctl, res));
2653 Node* mask = MakeConX(0x1);
2654 Node* masked = transform_later(new AndXNode(cast, mask));
2655 Node* cmp = transform_later(new CmpXNode(masked, mask));
2656 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
2657 IfNode* allocation_iff = new IfNode(ctl, bol, PROB_MAX, COUNT_UNKNOWN);
2658 transform_later(allocation_iff);
2659 Node* allocation_ctl = transform_later(new IfTrueNode(allocation_iff));
2660 Node* no_allocation_ctl = transform_later(new IfFalseNode(allocation_iff));
2661
2662 Node* no_allocation_res = transform_later(new CheckCastPPNode(no_allocation_ctl, res, TypeInstPtr::NOTNULL));
2663
2664 Node* mask2 = MakeConX(-2);
2665 Node* masked2 = transform_later(new AndXNode(cast, mask2));
2666 Node* rawklassptr = transform_later(new CastX2PNode(masked2));
2667 Node* klass_node = transform_later(new CheckCastPPNode(allocation_ctl, rawklassptr, TypeKlassPtr::OBJECT_OR_NULL));
2668
2669 Node* slowpath_bol = NULL;
2670 Node* top_adr = NULL;
2671 Node* old_top = NULL;
2672 Node* new_top = NULL;
2673 if (UseTLAB) {
2674 Node* end_adr = NULL;
2675 set_eden_pointers(top_adr, end_adr);
2676 Node* end = make_load(ctl, mem, end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS);
2677 old_top = new LoadPNode(ctl, mem, top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered);
2678 transform_later(old_top);
2679 Node* layout_val = make_load(NULL, mem, klass_node, in_bytes(Klass::layout_helper_offset()), TypeInt::INT, T_INT);
2680 Node* size_in_bytes = ConvI2X(layout_val);
2681 new_top = new AddPNode(top(), old_top, size_in_bytes);
2682 transform_later(new_top);
2683 Node* slowpath_cmp = new CmpPNode(new_top, end);
2684 transform_later(slowpath_cmp);
2685 slowpath_bol = new BoolNode(slowpath_cmp, BoolTest::ge);
2686 transform_later(slowpath_bol);
2687 } else {
2688 slowpath_bol = intcon(1);
2689 top_adr = top();
2690 old_top = top();
2691 new_top = top();
2692 }
2693 IfNode* slowpath_iff = new IfNode(allocation_ctl, slowpath_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN);
2694 transform_later(slowpath_iff);
2695
2696 Node* slowpath_true = new IfTrueNode(slowpath_iff);
2697 transform_later(slowpath_true);
2698
2699 CallStaticJavaNode* slow_call = new CallStaticJavaNode(OptoRuntime::store_value_type_fields_Type(),
2700 StubRoutines::store_value_type_fields_to_buf(),
2701 "store_value_type_fields",
2702 call->jvms()->bci(),
2703 TypePtr::BOTTOM);
2704 slow_call->init_req(TypeFunc::Control, slowpath_true);
2705 slow_call->init_req(TypeFunc::Memory, mem);
2706 slow_call->init_req(TypeFunc::I_O, io);
2707 slow_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2708 slow_call->init_req(TypeFunc::ReturnAdr, call->in(TypeFunc::ReturnAdr));
2709 slow_call->init_req(TypeFunc::Parms, res);
2710
2711 Node* slow_ctl = transform_later(new ProjNode(slow_call, TypeFunc::Control));
2712 Node* slow_mem = transform_later(new ProjNode(slow_call, TypeFunc::Memory));
2713 Node* slow_io = transform_later(new ProjNode(slow_call, TypeFunc::I_O));
2714 Node* slow_res = transform_later(new ProjNode(slow_call, TypeFunc::Parms));
2715 Node* slow_catc = transform_later(new CatchNode(slow_ctl, slow_io, 2));
2716 Node* slow_norm = transform_later(new CatchProjNode(slow_catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci));
2717 Node* slow_excp = transform_later(new CatchProjNode(slow_catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci));
2718
2719 Node* ex_r = new RegionNode(3);
2720 Node* ex_mem_phi = new PhiNode(ex_r, Type::MEMORY, TypePtr::BOTTOM);
2721 Node* ex_io_phi = new PhiNode(ex_r, Type::ABIO);
2722 ex_r->init_req(1, slow_excp);
2723 ex_mem_phi->init_req(1, slow_mem);
2724 ex_io_phi->init_req(1, slow_io);
2725 ex_r->init_req(2, ex_ctl);
2726 ex_mem_phi->init_req(2, ex_mem);
2727 ex_io_phi->init_req(2, ex_io);
2728
2729 transform_later(ex_r);
2730 transform_later(ex_mem_phi);
2731 transform_later(ex_io_phi);
2732
2733 Node* slowpath_false = new IfFalseNode(slowpath_iff);
2734 transform_later(slowpath_false);
2735 Node* rawmem = new StorePNode(slowpath_false, mem, top_adr, TypeRawPtr::BOTTOM, new_top, MemNode::unordered);
2736 transform_later(rawmem);
2737 Node* mark_node = mark_node = makecon(TypeRawPtr::make((address)markOopDesc::always_locked_prototype()));
2738 rawmem = make_store(slowpath_false, rawmem, old_top, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
2739 rawmem = make_store(slowpath_false, rawmem, old_top, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
2740 if (UseCompressedClassPointers) {
2741 rawmem = make_store(slowpath_false, rawmem, old_top, oopDesc::klass_gap_offset_in_bytes(), intcon(0), T_INT);
2742 }
2743 Node* pack_handler = make_load(slowpath_false, rawmem, klass_node, in_bytes(ValueKlass::pack_handler_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
2744
2745 CallLeafNoFPNode* handler_call = new CallLeafNoFPNode(OptoRuntime::pack_value_type_Type(),
2746 NULL,
2747 "pack handler",
2748 TypeRawPtr::BOTTOM);
2749 handler_call->init_req(TypeFunc::Control, slowpath_false);
2750 handler_call->init_req(TypeFunc::Memory, rawmem);
2751 handler_call->init_req(TypeFunc::I_O, top());
2752 handler_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2753 handler_call->init_req(TypeFunc::ReturnAdr, top());
2754 handler_call->init_req(TypeFunc::Parms, pack_handler);
2755 handler_call->init_req(TypeFunc::Parms+1, old_top);
2756
2757 // We don't know how many values are returned. This assumes the
2758 // worst case, that all available registers are used.
2759 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) {
2760 if (domain->field_at(i) == Type::HALF) {
2761 slow_call->init_req(i, top());
2762 handler_call->init_req(i+1, top());
2763 continue;
2764 }
2765 Node* proj = transform_later(new ProjNode(call, i));
2766 slow_call->init_req(i, proj);
2767 handler_call->init_req(i+1, proj);
2768 }
2769
2770 // We can safepoint at that new call
2771 copy_call_debug_info(call, slow_call);
2772 transform_later(slow_call);
2773 transform_later(handler_call);
2774
2775 Node* handler_ctl = transform_later(new ProjNode(handler_call, TypeFunc::Control));
2776 rawmem = transform_later(new ProjNode(handler_call, TypeFunc::Memory));
2777 Node* slowpath_false_res = transform_later(new ProjNode(handler_call, TypeFunc::Parms));
2778
2779 MergeMemNode* slowpath_false_mem = MergeMemNode::make(mem);
2780 slowpath_false_mem->set_memory_at(Compile::AliasIdxRaw, rawmem);
2781 transform_later(slowpath_false_mem);
2782
2783 Node* r = new RegionNode(4);
2784 Node* mem_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM);
2785 Node* io_phi = new PhiNode(r, Type::ABIO);
2786 Node* res_phi = new PhiNode(r, ret->bottom_type());
2787
2788 r->init_req(1, no_allocation_ctl);
2789 mem_phi->init_req(1, mem);
2790 io_phi->init_req(1, io);
2791 res_phi->init_req(1, no_allocation_res);
2792 r->init_req(2, slow_norm);
2793 mem_phi->init_req(2, slow_mem);
2794 io_phi->init_req(2, slow_io);
2795 res_phi->init_req(2, slow_res);
2796 r->init_req(3, handler_ctl);
2797 mem_phi->init_req(3, slowpath_false_mem);
2798 io_phi->init_req(3, io);
2799 res_phi->init_req(3, slowpath_false_res);
2800
2801 transform_later(r);
2802 transform_later(mem_phi);
2803 transform_later(io_phi);
2804 transform_later(res_phi);
2805
2806 assert(projs->nb_resproj == 1, "unexpected number of results");
2807 _igvn.replace_in_uses(projs->fallthrough_catchproj, r);
2808 _igvn.replace_in_uses(projs->fallthrough_memproj, mem_phi);
2809 _igvn.replace_in_uses(projs->fallthrough_ioproj, io_phi);
2810 _igvn.replace_in_uses(projs->resproj[0], res_phi);
2811 _igvn.replace_in_uses(projs->catchall_catchproj, ex_r);
2812 _igvn.replace_in_uses(projs->catchall_memproj, ex_mem_phi);
2813 _igvn.replace_in_uses(projs->catchall_ioproj, ex_io_phi);
2814
2815 _igvn.replace_node(ctl, projs->fallthrough_catchproj);
2816 _igvn.replace_node(mem, projs->fallthrough_memproj);
2817 _igvn.replace_node(io, projs->fallthrough_ioproj);
2818 _igvn.replace_node(res, projs->resproj[0]);
2819 _igvn.replace_node(ex_ctl, projs->catchall_catchproj);
2820 _igvn.replace_node(ex_mem, projs->catchall_memproj);
2821 _igvn.replace_node(ex_io, projs->catchall_ioproj);
2822 }
2823
2824 //---------------------------eliminate_macro_nodes----------------------
2825 // Eliminate scalar replaced allocations and associated locks.
2826 void PhaseMacroExpand::eliminate_macro_nodes() {
2827 if (C->macro_count() == 0)
2828 return;
2829
2830 // First, attempt to eliminate locks
2831 int cnt = C->macro_count();
2832 for (int i=0; i < cnt; i++) {
2833 Node *n = C->macro_node(i);
2834 if (n->is_AbstractLock()) { // Lock and Unlock nodes
2835 // Before elimination mark all associated (same box and obj)
2836 // lock and unlock nodes.
2837 mark_eliminated_locking_nodes(n->as_AbstractLock());
2838 }
2839 }
2840 bool progress = true;
2841 while (progress) {
2842 progress = false;
2843 for (int i = C->macro_count(); i > 0; i--) {
2844 Node * n = C->macro_node(i-1);
2845 bool success = false;
2846 debug_only(int old_macro_count = C->macro_count(););
2847 if (n->is_AbstractLock()) {
2848 success = eliminate_locking_node(n->as_AbstractLock());
2849 }
2850 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2851 progress = progress || success;
2852 }
2853 }
2854 // Next, attempt to eliminate allocations
2855 _has_locks = false;
2856 progress = true;
2857 while (progress) {
2858 progress = false;
2859 for (int i = C->macro_count(); i > 0; i--) {
2860 Node * n = C->macro_node(i-1);
2861 bool success = false;
2862 debug_only(int old_macro_count = C->macro_count(););
2863 switch (n->class_id()) {
2864 case Node::Class_Allocate:
2865 case Node::Class_AllocateArray:
2866 success = eliminate_allocate_node(n->as_Allocate());
2867 break;
2868 case Node::Class_CallStaticJava: {
2869 CallStaticJavaNode* call = n->as_CallStaticJava();
2870 if (!call->method()->is_method_handle_intrinsic()) {
2871 success = eliminate_boxing_node(n->as_CallStaticJava());
2872 }
2873 break;
2874 }
2875 case Node::Class_Lock:
2876 case Node::Class_Unlock:
2877 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2878 _has_locks = true;
2879 break;
2880 case Node::Class_ArrayCopy:
2881 break;
2882 case Node::Class_OuterStripMinedLoop:
2883 break;
2884 default:
2885 assert(n->Opcode() == Op_LoopLimit ||
2886 n->Opcode() == Op_Opaque1 ||
2887 n->Opcode() == Op_Opaque2 ||
2888 n->Opcode() == Op_Opaque3 ||
2889 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n),
2890 "unknown node type in macro list");
2891 }
2892 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2893 progress = progress || success;
2894 }
2895 }
2896 }
2897
2898 //------------------------------expand_macro_nodes----------------------
2899 // Returns true if a failure occurred.
2900 bool PhaseMacroExpand::expand_macro_nodes() {
2901 // Last attempt to eliminate macro nodes.
2902 eliminate_macro_nodes();
2903
2904 // Make sure expansion will not cause node limit to be exceeded.
2905 // Worst case is a macro node gets expanded into about 200 nodes.
2906 // Allow 50% more for optimization.
2907 if (C->check_node_count(C->macro_count() * 300, "out of nodes before macro expansion" ) )
2908 return true;
2909
2910 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2911 bool progress = true;
2912 while (progress) {
2913 progress = false;
2914 for (int i = C->macro_count(); i > 0; i--) {
2915 Node * n = C->macro_node(i-1);
2916 bool success = false;
2917 debug_only(int old_macro_count = C->macro_count(););
2918 if (n->Opcode() == Op_LoopLimit) {
2919 // Remove it from macro list and put on IGVN worklist to optimize.
2920 C->remove_macro_node(n);
2921 _igvn._worklist.push(n);
2922 success = true;
2923 } else if (n->Opcode() == Op_CallStaticJava) {
2924 CallStaticJavaNode* call = n->as_CallStaticJava();
2925 if (!call->method()->is_method_handle_intrinsic()) {
2926 // Remove it from macro list and put on IGVN worklist to optimize.
2927 C->remove_macro_node(n);
2928 _igvn._worklist.push(n);
2929 success = true;
2930 }
2931 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) {
2932 _igvn.replace_node(n, n->in(1));
2933 success = true;
2934 #if INCLUDE_RTM_OPT
2935 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) {
2936 assert(C->profile_rtm(), "should be used only in rtm deoptimization code");
2937 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), "");
2938 Node* cmp = n->unique_out();
2939 #ifdef ASSERT
2940 // Validate graph.
2941 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), "");
2942 BoolNode* bol = cmp->unique_out()->as_Bool();
2943 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() &&
2944 (bol->_test._test == BoolTest::ne), "");
2945 IfNode* ifn = bol->unique_out()->as_If();
2946 assert((ifn->outcnt() == 2) &&
2947 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != NULL, "");
2948 #endif
2949 Node* repl = n->in(1);
2950 if (!_has_locks) {
2951 // Remove RTM state check if there are no locks in the code.
2952 // Replace input to compare the same value.
2953 repl = (cmp->in(1) == n) ? cmp->in(2) : cmp->in(1);
2954 }
2955 _igvn.replace_node(n, repl);
2956 success = true;
2957 #endif
2958 } else if (n->Opcode() == Op_OuterStripMinedLoop) {
2959 n->as_OuterStripMinedLoop()->adjust_strip_mined_loop(&_igvn);
2960 C->remove_macro_node(n);
2961 success = true;
2962 }
2963 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2964 progress = progress || success;
2965 }
2966 }
2967
2968 // expand arraycopy "macro" nodes first
2969 // For ReduceBulkZeroing, we must first process all arraycopy nodes
2970 // before the allocate nodes are expanded.
2971 int macro_idx = C->macro_count() - 1;
2972 while (macro_idx >= 0) {
2973 Node * n = C->macro_node(macro_idx);
2974 assert(n->is_macro(), "only macro nodes expected here");
2975 if (_igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())) {
2976 // node is unreachable, so don't try to expand it
2977 C->remove_macro_node(n);
2978 } else if (n->is_ArrayCopy()){
2979 int macro_count = C->macro_count();
2980 expand_arraycopy_node(n->as_ArrayCopy());
2981 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2982 }
2983 if (C->failing()) return true;
2984 macro_idx --;
2985 }
2986
2987 // expand "macro" nodes
2988 // nodes are removed from the macro list as they are processed
2989 while (C->macro_count() > 0) {
2990 int macro_count = C->macro_count();
2991 Node * n = C->macro_node(macro_count-1);
2992 assert(n->is_macro(), "only macro nodes expected here");
2993 if (_igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())) {
2994 // node is unreachable, so don't try to expand it
2995 C->remove_macro_node(n);
2996 continue;
2997 }
2998 switch (n->class_id()) {
2999 case Node::Class_Allocate:
3000 expand_allocate(n->as_Allocate());
3001 break;
3002 case Node::Class_AllocateArray:
3003 expand_allocate_array(n->as_AllocateArray());
3004 break;
3005 case Node::Class_Lock:
3006 expand_lock_node(n->as_Lock());
3007 break;
3008 case Node::Class_Unlock:
3009 expand_unlock_node(n->as_Unlock());
3010 break;
3011 case Node::Class_CallStaticJava:
3012 expand_mh_intrinsic_return(n->as_CallStaticJava());
3013 C->remove_macro_node(n);
3014 break;
3015 default:
3016 assert(false, "unknown node type in macro list");
3017 }
3018 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
3019 if (C->failing()) return true;
3020 }
3021
3022 _igvn.set_delay_transform(false);
3023 _igvn.optimize();
3024 if (C->failing()) return true;
3025 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
3026 return bs->expand_macro_nodes(this);
3027 }