1 /*
2 * Copyright (c) 2005, 2019, 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 "memory/universe.hpp"
30 #include "opto/addnode.hpp"
31 #include "opto/arraycopynode.hpp"
32 #include "opto/callnode.hpp"
33 #include "opto/castnode.hpp"
34 #include "opto/cfgnode.hpp"
35 #include "opto/compile.hpp"
36 #include "opto/convertnode.hpp"
37 #include "opto/graphKit.hpp"
38 #include "opto/locknode.hpp"
39 #include "opto/loopnode.hpp"
40 #include "opto/macro.hpp"
41 #include "opto/memnode.hpp"
42 #include "opto/narrowptrnode.hpp"
43 #include "opto/node.hpp"
44 #include "opto/opaquenode.hpp"
45 #include "opto/phaseX.hpp"
46 #include "opto/rootnode.hpp"
47 #include "opto/runtime.hpp"
48 #include "opto/subnode.hpp"
49 #include "opto/type.hpp"
50 #include "opto/valuetypenode.hpp"
51 #include "runtime/sharedRuntime.hpp"
52 #include "utilities/macros.hpp"
53 #if INCLUDE_G1GC
54 #include "gc/g1/g1ThreadLocalData.hpp"
55 #endif // INCLUDE_G1GC
56 #if INCLUDE_SHENANDOAHGC
57 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
58 #endif
59
60
61 //
62 // Replace any references to "oldref" in inputs to "use" with "newref".
63 // Returns the number of replacements made.
64 //
65 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
66 int nreplacements = 0;
67 uint req = use->req();
68 for (uint j = 0; j < use->len(); j++) {
69 Node *uin = use->in(j);
70 if (uin == oldref) {
71 if (j < req)
72 use->set_req(j, newref);
73 else
74 use->set_prec(j, newref);
75 nreplacements++;
76 } else if (j >= req && uin == NULL) {
77 break;
78 }
79 }
80 return nreplacements;
81 }
82
83 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
84 Node* cmp;
85 if (mask != 0) {
86 Node* and_node = transform_later(new AndXNode(word, MakeConX(mask)));
87 cmp = transform_later(new CmpXNode(and_node, MakeConX(bits)));
88 } else {
89 cmp = word;
90 }
91 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne));
92 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
93 transform_later(iff);
94
95 // Fast path taken.
96 Node *fast_taken = transform_later(new IfFalseNode(iff));
97
98 // Fast path not-taken, i.e. slow path
99 Node *slow_taken = transform_later(new IfTrueNode(iff));
100
101 if (return_fast_path) {
102 region->init_req(edge, slow_taken); // Capture slow-control
103 return fast_taken;
104 } else {
105 region->init_req(edge, fast_taken); // Capture fast-control
106 return slow_taken;
107 }
108 }
109
110 //--------------------copy_predefined_input_for_runtime_call--------------------
111 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
112 // Set fixed predefined input arguments
113 call->init_req( TypeFunc::Control, ctrl );
114 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) );
115 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
116 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
117 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
118 }
119
120 //------------------------------make_slow_call---------------------------------
121 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type,
122 address slow_call, const char* leaf_name, Node* slow_path,
123 Node* parm0, Node* parm1, Node* parm2) {
124
125 // Slow-path call
126 CallNode *call = leaf_name
127 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
128 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM );
129
130 // Slow path call has no side-effects, uses few values
131 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
132 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0);
133 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1);
134 if (parm2 != NULL) call->init_req(TypeFunc::Parms+2, parm2);
135 call->copy_call_debug_info(&_igvn, oldcall);
136 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
137 _igvn.replace_node(oldcall, call);
138 transform_later(call);
139
140 return call;
141 }
142
143 void PhaseMacroExpand::extract_call_projections(CallNode *call) {
144 _fallthroughproj = NULL;
145 _fallthroughcatchproj = NULL;
146 _ioproj_fallthrough = NULL;
147 _ioproj_catchall = NULL;
148 _catchallcatchproj = NULL;
149 _memproj_fallthrough = NULL;
150 _memproj_catchall = NULL;
151 _resproj = NULL;
152 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) {
153 ProjNode *pn = call->fast_out(i)->as_Proj();
154 switch (pn->_con) {
155 case TypeFunc::Control:
156 {
157 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj
158 _fallthroughproj = pn;
159 DUIterator_Fast jmax, j = pn->fast_outs(jmax);
160 const Node *cn = pn->fast_out(j);
161 if (cn->is_Catch()) {
162 ProjNode *cpn = NULL;
163 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) {
164 cpn = cn->fast_out(k)->as_Proj();
165 assert(cpn->is_CatchProj(), "must be a CatchProjNode");
166 if (cpn->_con == CatchProjNode::fall_through_index)
167 _fallthroughcatchproj = cpn;
168 else {
169 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index.");
170 _catchallcatchproj = cpn;
171 }
172 }
173 }
174 break;
175 }
176 case TypeFunc::I_O:
177 if (pn->_is_io_use)
178 _ioproj_catchall = pn;
179 else
180 _ioproj_fallthrough = pn;
181 break;
182 case TypeFunc::Memory:
183 if (pn->_is_io_use)
184 _memproj_catchall = pn;
185 else
186 _memproj_fallthrough = pn;
187 break;
188 case TypeFunc::Parms:
189 _resproj = pn;
190 break;
191 default:
192 assert(false, "unexpected projection from allocation node.");
193 }
194 }
195
196 }
197
198 void PhaseMacroExpand::eliminate_gc_barrier(Node* p2x) {
199 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
200 bs->eliminate_gc_barrier(this, p2x);
201 }
202
203 // Search for a memory operation for the specified memory slice.
204 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
205 Node *orig_mem = mem;
206 Node *alloc_mem = alloc->in(TypeFunc::Memory);
207 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
208 while (true) {
209 if (mem == alloc_mem || mem == start_mem ) {
210 return mem; // hit one of our sentinels
211 } else if (mem->is_MergeMem()) {
212 mem = mem->as_MergeMem()->memory_at(alias_idx);
213 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
214 Node *in = mem->in(0);
215 // we can safely skip over safepoints, calls, locks and membars because we
216 // already know that the object is safe to eliminate.
217 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
218 return in;
219 } else if (in->is_Call()) {
220 CallNode *call = in->as_Call();
221 if (call->may_modify(tinst, phase)) {
222 assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape");
223 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) {
224 return in;
225 }
226 }
227 mem = in->in(TypeFunc::Memory);
228 } else if (in->is_MemBar()) {
229 ArrayCopyNode* ac = NULL;
230 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) {
231 assert(ac != NULL && ac->is_clonebasic(), "Only basic clone is a non escaping clone");
232 return ac;
233 }
234 mem = in->in(TypeFunc::Memory);
235 } else {
236 assert(false, "unexpected projection");
237 }
238 } else if (mem->is_Store()) {
239 const TypePtr* atype = mem->as_Store()->adr_type();
240 int adr_idx = phase->C->get_alias_index(atype);
241 if (adr_idx == alias_idx) {
242 assert(atype->isa_oopptr(), "address type must be oopptr");
243 int adr_offset = atype->flattened_offset();
244 uint adr_iid = atype->is_oopptr()->instance_id();
245 // Array elements references have the same alias_idx
246 // but different offset and different instance_id.
247 if (adr_offset == offset && adr_iid == alloc->_idx)
248 return mem;
249 } else {
250 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
251 }
252 mem = mem->in(MemNode::Memory);
253 } else if (mem->is_ClearArray()) {
254 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
255 // Can not bypass initialization of the instance
256 // we are looking.
257 debug_only(intptr_t offset;)
258 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
259 InitializeNode* init = alloc->as_Allocate()->initialization();
260 // We are looking for stored value, return Initialize node
261 // or memory edge from Allocate node.
262 if (init != NULL)
263 return init;
264 else
265 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers).
266 }
267 // Otherwise skip it (the call updated 'mem' value).
268 } else if (mem->Opcode() == Op_SCMemProj) {
269 mem = mem->in(0);
270 Node* adr = NULL;
271 if (mem->is_LoadStore()) {
272 adr = mem->in(MemNode::Address);
273 } else {
274 assert(mem->Opcode() == Op_EncodeISOArray ||
275 mem->Opcode() == Op_StrCompressedCopy, "sanity");
276 adr = mem->in(3); // Destination array
277 }
278 const TypePtr* atype = adr->bottom_type()->is_ptr();
279 int adr_idx = phase->C->get_alias_index(atype);
280 if (adr_idx == alias_idx) {
281 DEBUG_ONLY(mem->dump();)
282 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
283 return NULL;
284 }
285 mem = mem->in(MemNode::Memory);
286 } else if (mem->Opcode() == Op_StrInflatedCopy) {
287 Node* adr = mem->in(3); // Destination array
288 const TypePtr* atype = adr->bottom_type()->is_ptr();
289 int adr_idx = phase->C->get_alias_index(atype);
290 if (adr_idx == alias_idx) {
291 DEBUG_ONLY(mem->dump();)
292 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
293 return NULL;
294 }
295 mem = mem->in(MemNode::Memory);
296 } else {
297 return mem;
298 }
299 assert(mem != orig_mem, "dead memory loop");
300 }
301 }
302
303 // Generate loads from source of the arraycopy for fields of
304 // destination needed at a deoptimization point
305 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) {
306 BasicType bt = ft;
307 const Type *type = ftype;
308 if (ft == T_NARROWOOP) {
309 bt = T_OBJECT;
310 type = ftype->make_oopptr();
311 }
312 Node* res = NULL;
313 if (ac->is_clonebasic()) {
314 assert(ac->in(ArrayCopyNode::Src) != ac->in(ArrayCopyNode::Dest), "clone source equals destination");
315 Node* base = ac->in(ArrayCopyNode::Src)->in(AddPNode::Base);
316 Node* adr = _igvn.transform(new AddPNode(base, base, MakeConX(offset)));
317 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
318 res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::UnknownControl);
319 } else {
320 if (ac->modifies(offset, offset, &_igvn, true)) {
321 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
322 uint shift = exact_log2(type2aelembytes(bt));
323 Node* src_pos = ac->in(ArrayCopyNode::SrcPos);
324 Node* dest_pos = ac->in(ArrayCopyNode::DestPos);
325 const TypeInt* src_pos_t = _igvn.type(src_pos)->is_int();
326 const TypeInt* dest_pos_t = _igvn.type(dest_pos)->is_int();
327
328 Node* adr = NULL;
329 Node* base = ac->in(ArrayCopyNode::Src);
330 const TypePtr* adr_type = _igvn.type(base)->is_ptr();
331 assert(adr_type->isa_aryptr(), "only arrays here");
332 if (src_pos_t->is_con() && dest_pos_t->is_con()) {
333 intptr_t off = ((src_pos_t->get_con() - dest_pos_t->get_con()) << shift) + offset;
334 adr = _igvn.transform(new AddPNode(base, base, MakeConX(off)));
335 adr_type = _igvn.type(adr)->is_ptr();
336 assert(adr_type == _igvn.type(base)->is_aryptr()->add_field_offset_and_offset(off), "incorrect address type");
337 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
338 // Don't emit a new load from src if src == dst but try to get the value from memory instead
339 return value_from_mem(ac->in(TypeFunc::Memory), ctl, ft, ftype, adr_type->isa_oopptr(), alloc);
340 }
341 } else {
342 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
343 // Non constant offset in the array: we can't statically
344 // determine the value
345 return NULL;
346 }
347 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
348 #ifdef _LP64
349 diff = _igvn.transform(new ConvI2LNode(diff));
350 #endif
351 diff = _igvn.transform(new LShiftXNode(diff, intcon(shift)));
352
353 Node* off = _igvn.transform(new AddXNode(MakeConX(offset), diff));
354 adr = _igvn.transform(new AddPNode(base, base, off));
355 // In the case of a flattened value type array, each field has its
356 // own slice so we need to extract the field being accessed from
357 // the address computation
358 adr_type = adr_type->is_aryptr()->add_field_offset_and_offset(offset)->add_offset(Type::OffsetBot);
359 adr = _igvn.transform(new CastPPNode(adr, adr_type));
360 }
361 res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::UnknownControl);
362 }
363 }
364 if (res != NULL) {
365 res = _igvn.transform(res);
366 if (ftype->isa_narrowoop()) {
367 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
368 assert(res->isa_DecodeN(), "should be narrow oop");
369 res = _igvn.transform(new EncodePNode(res, ftype));
370 }
371 return res;
372 }
373 return NULL;
374 }
375
376 //
377 // Given a Memory Phi, compute a value Phi containing the values from stores
378 // on the input paths.
379 // Note: this function is recursive, its depth is limited by the "level" argument
380 // Returns the computed Phi, or NULL if it cannot compute it.
381 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) {
382 assert(mem->is_Phi(), "sanity");
383 int alias_idx = C->get_alias_index(adr_t);
384 int offset = adr_t->flattened_offset();
385 int instance_id = adr_t->instance_id();
386
387 // Check if an appropriate value phi already exists.
388 Node* region = mem->in(0);
389 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
390 Node* phi = region->fast_out(k);
391 if (phi->is_Phi() && phi != mem &&
392 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) {
393 return phi;
394 }
395 }
396 // Check if an appropriate new value phi already exists.
397 Node* new_phi = value_phis->find(mem->_idx);
398 if (new_phi != NULL)
399 return new_phi;
400
401 if (level <= 0) {
402 return NULL; // Give up: phi tree too deep
403 }
404 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
405 Node *alloc_mem = alloc->in(TypeFunc::Memory);
406
407 uint length = mem->req();
408 GrowableArray <Node *> values(length, length, NULL, false);
409
410 // create a new Phi for the value
411 PhiNode *phi = new PhiNode(mem->in(0), phi_type, NULL, mem->_idx, instance_id, alias_idx, offset);
412 transform_later(phi);
413 value_phis->push(phi, mem->_idx);
414
415 for (uint j = 1; j < length; j++) {
416 Node *in = mem->in(j);
417 if (in == NULL || in->is_top()) {
418 values.at_put(j, in);
419 } else {
420 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
421 if (val == start_mem || val == alloc_mem) {
422 // hit a sentinel, return appropriate 0 value
423 Node* default_value = alloc->in(AllocateNode::DefaultValue);
424 if (default_value != NULL) {
425 values.at_put(j, default_value);
426 } else {
427 assert(alloc->in(AllocateNode::RawDefaultValue) == NULL, "default value may not be null");
428 values.at_put(j, _igvn.zerocon(ft));
429 }
430 continue;
431 }
432 if (val->is_Initialize()) {
433 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
434 }
435 if (val == NULL) {
436 return NULL; // can't find a value on this path
437 }
438 if (val == mem) {
439 values.at_put(j, mem);
440 } else if (val->is_Store()) {
441 Node* n = val->in(MemNode::ValueIn);
442 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
443 n = bs->step_over_gc_barrier(n);
444 values.at_put(j, n);
445 } else if(val->is_Proj() && val->in(0) == alloc) {
446 Node* default_value = alloc->in(AllocateNode::DefaultValue);
447 if (default_value != NULL) {
448 values.at_put(j, default_value);
449 } else {
450 assert(alloc->in(AllocateNode::RawDefaultValue) == NULL, "default value may not be null");
451 values.at_put(j, _igvn.zerocon(ft));
452 }
453 } else if (val->is_Phi()) {
454 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
455 if (val == NULL) {
456 return NULL;
457 }
458 values.at_put(j, val);
459 } else if (val->Opcode() == Op_SCMemProj) {
460 assert(val->in(0)->is_LoadStore() ||
461 val->in(0)->Opcode() == Op_EncodeISOArray ||
462 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity");
463 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
464 return NULL;
465 } else if (val->is_ArrayCopy()) {
466 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc);
467 if (res == NULL) {
468 return NULL;
469 }
470 values.at_put(j, res);
471 } else {
472 #ifdef ASSERT
473 val->dump();
474 assert(false, "unknown node on this path");
475 #endif
476 return NULL; // unknown node on this path
477 }
478 }
479 }
480 // Set Phi's inputs
481 for (uint j = 1; j < length; j++) {
482 if (values.at(j) == mem) {
483 phi->init_req(j, phi);
484 } else {
485 phi->init_req(j, values.at(j));
486 }
487 }
488 return phi;
489 }
490
491 // Search the last value stored into the object's field.
492 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) {
493 assert(adr_t->is_known_instance_field(), "instance required");
494 int instance_id = adr_t->instance_id();
495 assert((uint)instance_id == alloc->_idx, "wrong allocation");
496
497 int alias_idx = C->get_alias_index(adr_t);
498 int offset = adr_t->flattened_offset();
499 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
500 Node *alloc_mem = alloc->in(TypeFunc::Memory);
501 Arena *a = Thread::current()->resource_area();
502 VectorSet visited(a);
503
504 bool done = sfpt_mem == alloc_mem;
505 Node *mem = sfpt_mem;
506 while (!done) {
507 if (visited.test_set(mem->_idx)) {
508 return NULL; // found a loop, give up
509 }
510 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
511 if (mem == start_mem || mem == alloc_mem) {
512 done = true; // hit a sentinel, return appropriate 0 value
513 } else if (mem->is_Initialize()) {
514 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
515 if (mem == NULL) {
516 done = true; // Something went wrong.
517 } else if (mem->is_Store()) {
518 const TypePtr* atype = mem->as_Store()->adr_type();
519 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
520 done = true;
521 }
522 } else if (mem->is_Store()) {
523 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
524 assert(atype != NULL, "address type must be oopptr");
525 assert(C->get_alias_index(atype) == alias_idx &&
526 atype->is_known_instance_field() && atype->flattened_offset() == offset &&
527 atype->instance_id() == instance_id, "store is correct memory slice");
528 done = true;
529 } else if (mem->is_Phi()) {
530 // try to find a phi's unique input
531 Node *unique_input = NULL;
532 Node *top = C->top();
533 for (uint i = 1; i < mem->req(); i++) {
534 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
535 if (n == NULL || n == top || n == mem) {
536 continue;
537 } else if (unique_input == NULL) {
538 unique_input = n;
539 } else if (unique_input != n) {
540 unique_input = top;
541 break;
542 }
543 }
544 if (unique_input != NULL && unique_input != top) {
545 mem = unique_input;
546 } else {
547 done = true;
548 }
549 } else if (mem->is_ArrayCopy()) {
550 done = true;
551 } else {
552 assert(false, "unexpected node");
553 }
554 }
555 if (mem != NULL) {
556 if (mem == start_mem || mem == alloc_mem) {
557 // hit a sentinel, return appropriate 0 value
558 Node* default_value = alloc->in(AllocateNode::DefaultValue);
559 if (default_value != NULL) {
560 return default_value;
561 }
562 assert(alloc->in(AllocateNode::RawDefaultValue) == NULL, "default value may not be null");
563 return _igvn.zerocon(ft);
564 } else if (mem->is_Store()) {
565 Node* n = mem->in(MemNode::ValueIn);
566 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
567 n = bs->step_over_gc_barrier(n);
568 return n;
569 } else if (mem->is_Phi()) {
570 // attempt to produce a Phi reflecting the values on the input paths of the Phi
571 Node_Stack value_phis(a, 8);
572 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
573 if (phi != NULL) {
574 return phi;
575 } else {
576 // Kill all new Phis
577 while(value_phis.is_nonempty()) {
578 Node* n = value_phis.node();
579 _igvn.replace_node(n, C->top());
580 value_phis.pop();
581 }
582 }
583 } else if (mem->is_ArrayCopy()) {
584 Node* ctl = mem->in(0);
585 Node* m = mem->in(TypeFunc::Memory);
586 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) {
587 // pin the loads in the uncommon trap path
588 ctl = sfpt_ctl;
589 m = sfpt_mem;
590 }
591 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc);
592 }
593 }
594 // Something went wrong.
595 return NULL;
596 }
597
598 // Search the last value stored into the value type's fields.
599 Node* PhaseMacroExpand::value_type_from_mem(Node* mem, Node* ctl, ciValueKlass* vk, const TypeAryPtr* adr_type, int offset, AllocateNode* alloc) {
600 // Subtract the offset of the first field to account for the missing oop header
601 offset -= vk->first_field_offset();
602 // Create a new ValueTypeNode and retrieve the field values from memory
603 ValueTypeNode* vt = ValueTypeNode::make_uninitialized(_igvn, vk)->as_ValueType();
604 for (int i = 0; i < vk->nof_declared_nonstatic_fields(); ++i) {
605 ciType* field_type = vt->field_type(i);
606 int field_offset = offset + vt->field_offset(i);
607 // Each value type field has its own memory slice
608 adr_type = adr_type->with_field_offset(field_offset);
609 Node* value = NULL;
610 if (vt->field_is_flattened(i)) {
611 value = value_type_from_mem(mem, ctl, field_type->as_value_klass(), adr_type, field_offset, alloc);
612 } else {
613 const Type* ft = Type::get_const_type(field_type);
614 BasicType bt = field_type->basic_type();
615 if (UseCompressedOops && !is_java_primitive(bt)) {
616 ft = ft->make_narrowoop();
617 bt = T_NARROWOOP;
618 }
619 value = value_from_mem(mem, ctl, bt, ft, adr_type, alloc);
620 if (value != NULL && ft->isa_narrowoop()) {
621 assert(UseCompressedOops, "unexpected narrow oop");
622 value = transform_later(new DecodeNNode(value, value->get_ptr_type()));
623 }
624 }
625 if (value != NULL) {
626 vt->set_field_value(i, value);
627 } else {
628 // We might have reached the TrackedInitializationLimit
629 return NULL;
630 }
631 }
632 return transform_later(vt);
633 }
634
635 // Check the possibility of scalar replacement.
636 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
637 // Scan the uses of the allocation to check for anything that would
638 // prevent us from eliminating it.
639 NOT_PRODUCT( const char* fail_eliminate = NULL; )
640 DEBUG_ONLY( Node* disq_node = NULL; )
641 bool can_eliminate = true;
642
643 Node* res = alloc->result_cast();
644 const TypeOopPtr* res_type = NULL;
645 if (res == NULL) {
646 // All users were eliminated.
647 } else if (!res->is_CheckCastPP()) {
648 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
649 can_eliminate = false;
650 } else {
651 res_type = _igvn.type(res)->isa_oopptr();
652 if (res_type == NULL) {
653 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
654 can_eliminate = false;
655 } else if (res_type->isa_aryptr()) {
656 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
657 if (length < 0) {
658 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
659 can_eliminate = false;
660 }
661 }
662 }
663
664 if (can_eliminate && res != NULL) {
665 for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
666 j < jmax && can_eliminate; j++) {
667 Node* use = res->fast_out(j);
668
669 if (use->is_AddP()) {
670 const TypePtr* addp_type = _igvn.type(use)->is_ptr();
671 int offset = addp_type->offset();
672
673 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
674 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";)
675 can_eliminate = false;
676 break;
677 }
678 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
679 k < kmax && can_eliminate; k++) {
680 Node* n = use->fast_out(k);
681 if (!n->is_Store() && n->Opcode() != Op_CastP2X &&
682 SHENANDOAHGC_ONLY((!UseShenandoahGC || !ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(n)) &&)
683 !(n->is_ArrayCopy() &&
684 n->as_ArrayCopy()->is_clonebasic() &&
685 n->in(ArrayCopyNode::Dest) == use)) {
686 DEBUG_ONLY(disq_node = n;)
687 if (n->is_Load() || n->is_LoadStore()) {
688 NOT_PRODUCT(fail_eliminate = "Field load";)
689 } else {
690 NOT_PRODUCT(fail_eliminate = "Not store field reference";)
691 }
692 can_eliminate = false;
693 }
694 }
695 } else if (use->is_ArrayCopy() &&
696 (use->as_ArrayCopy()->is_arraycopy_validated() ||
697 use->as_ArrayCopy()->is_copyof_validated() ||
698 use->as_ArrayCopy()->is_copyofrange_validated()) &&
699 use->in(ArrayCopyNode::Dest) == res) {
700 // ok to eliminate
701 } else if (use->is_SafePoint()) {
702 SafePointNode* sfpt = use->as_SafePoint();
703 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
704 // Object is passed as argument.
705 DEBUG_ONLY(disq_node = use;)
706 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
707 can_eliminate = false;
708 }
709 Node* sfptMem = sfpt->memory();
710 if (sfptMem == NULL || sfptMem->is_top()) {
711 DEBUG_ONLY(disq_node = use;)
712 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";)
713 can_eliminate = false;
714 } else {
715 safepoints.append_if_missing(sfpt);
716 }
717 } else if (use->is_ValueType() && use->isa_ValueType()->get_oop() == res) {
718 // ok to eliminate
719 } else if (use->Opcode() == Op_StoreX && use->in(MemNode::Address) == res) {
720 // store to mark work
721 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
722 if (use->is_Phi()) {
723 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
724 NOT_PRODUCT(fail_eliminate = "Object is return value";)
725 } else {
726 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
727 }
728 DEBUG_ONLY(disq_node = use;)
729 } else {
730 if (use->Opcode() == Op_Return) {
731 NOT_PRODUCT(fail_eliminate = "Object is return value";)
732 } else {
733 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
734 }
735 DEBUG_ONLY(disq_node = use;)
736 }
737 can_eliminate = false;
738 } else {
739 assert(use->Opcode() == Op_CastP2X, "should be");
740 assert(!use->has_out_with(Op_OrL), "should have been removed because oop is never null");
741 }
742 }
743 }
744
745 #ifndef PRODUCT
746 if (PrintEliminateAllocations) {
747 if (can_eliminate) {
748 tty->print("Scalar ");
749 if (res == NULL)
750 alloc->dump();
751 else
752 res->dump();
753 } else if (alloc->_is_scalar_replaceable) {
754 tty->print("NotScalar (%s)", fail_eliminate);
755 if (res == NULL)
756 alloc->dump();
757 else
758 res->dump();
759 #ifdef ASSERT
760 if (disq_node != NULL) {
761 tty->print(" >>>> ");
762 disq_node->dump();
763 }
764 #endif /*ASSERT*/
765 }
766 }
767 #endif
768 return can_eliminate;
769 }
770
771 // Do scalar replacement.
772 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
773 GrowableArray <SafePointNode *> safepoints_done;
774
775 ciKlass* klass = NULL;
776 ciInstanceKlass* iklass = NULL;
777 int nfields = 0;
778 int array_base = 0;
779 int element_size = 0;
780 BasicType basic_elem_type = T_ILLEGAL;
781 ciType* elem_type = NULL;
782
783 Node* res = alloc->result_cast();
784 assert(res == NULL || res->is_CheckCastPP(), "unexpected AllocateNode result");
785 const TypeOopPtr* res_type = NULL;
786 if (res != NULL) { // Could be NULL when there are no users
787 res_type = _igvn.type(res)->isa_oopptr();
788 }
789
790 if (res != NULL) {
791 klass = res_type->klass();
792 if (res_type->isa_instptr()) {
793 // find the fields of the class which will be needed for safepoint debug information
794 assert(klass->is_instance_klass(), "must be an instance klass.");
795 iklass = klass->as_instance_klass();
796 nfields = iklass->nof_nonstatic_fields();
797 } else {
798 // find the array's elements which will be needed for safepoint debug information
799 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
800 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass.");
801 elem_type = klass->as_array_klass()->element_type();
802 basic_elem_type = elem_type->basic_type();
803 if (elem_type->is_valuetype() && !klass->is_value_array_klass()) {
804 assert(basic_elem_type == T_VALUETYPE, "unexpected element basic type");
805 basic_elem_type = T_OBJECT;
806 }
807 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
808 element_size = type2aelembytes(basic_elem_type);
809 if (klass->is_value_array_klass()) {
810 // Flattened value type array
811 element_size = klass->as_value_array_klass()->element_byte_size();
812 }
813 }
814 }
815 //
816 // Process the safepoint uses
817 //
818 Unique_Node_List value_worklist;
819 while (safepoints.length() > 0) {
820 SafePointNode* sfpt = safepoints.pop();
821 Node* mem = sfpt->memory();
822 Node* ctl = sfpt->control();
823 assert(sfpt->jvms() != NULL, "missed JVMS");
824 // Fields of scalar objs are referenced only at the end
825 // of regular debuginfo at the last (youngest) JVMS.
826 // Record relative start index.
827 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
828 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type,
829 #ifdef ASSERT
830 alloc,
831 #endif
832 first_ind, nfields);
833 sobj->init_req(0, C->root());
834 transform_later(sobj);
835
836 // Scan object's fields adding an input to the safepoint for each field.
837 for (int j = 0; j < nfields; j++) {
838 intptr_t offset;
839 ciField* field = NULL;
840 if (iklass != NULL) {
841 field = iklass->nonstatic_field_at(j);
842 offset = field->offset();
843 elem_type = field->type();
844 basic_elem_type = field->layout_type();
845 assert(!field->is_flattened(), "flattened value type fields should not have safepoint uses");
846 } else {
847 offset = array_base + j * (intptr_t)element_size;
848 }
849
850 const Type *field_type;
851 // The next code is taken from Parse::do_get_xxx().
852 if (is_reference_type(basic_elem_type)) {
853 if (!elem_type->is_loaded()) {
854 field_type = TypeInstPtr::BOTTOM;
855 } else if (field != NULL && field->is_static_constant()) {
856 // This can happen if the constant oop is non-perm.
857 ciObject* con = field->constant_value().as_object();
858 // Do not "join" in the previous type; it doesn't add value,
859 // and may yield a vacuous result if the field is of interface type.
860 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
861 assert(field_type != NULL, "field singleton type must be consistent");
862 } else {
863 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
864 }
865 if (UseCompressedOops) {
866 field_type = field_type->make_narrowoop();
867 basic_elem_type = T_NARROWOOP;
868 }
869 } else {
870 field_type = Type::get_const_basic_type(basic_elem_type);
871 }
872
873 Node* field_val = NULL;
874 const TypeOopPtr* field_addr_type = res_type->add_offset(offset)->isa_oopptr();
875 if (klass->is_value_array_klass()) {
876 ciValueKlass* vk = elem_type->as_value_klass();
877 assert(vk->flatten_array(), "must be flattened");
878 field_val = value_type_from_mem(mem, ctl, vk, field_addr_type->isa_aryptr(), 0, alloc);
879 } else {
880 field_val = value_from_mem(mem, ctl, basic_elem_type, field_type, field_addr_type, alloc);
881 }
882 if (field_val == NULL) {
883 // We weren't able to find a value for this field,
884 // give up on eliminating this allocation.
885
886 // Remove any extra entries we added to the safepoint.
887 uint last = sfpt->req() - 1;
888 for (int k = 0; k < j; k++) {
889 sfpt->del_req(last--);
890 }
891 _igvn._worklist.push(sfpt);
892 // rollback processed safepoints
893 while (safepoints_done.length() > 0) {
894 SafePointNode* sfpt_done = safepoints_done.pop();
895 // remove any extra entries we added to the safepoint
896 last = sfpt_done->req() - 1;
897 for (int k = 0; k < nfields; k++) {
898 sfpt_done->del_req(last--);
899 }
900 JVMState *jvms = sfpt_done->jvms();
901 jvms->set_endoff(sfpt_done->req());
902 // Now make a pass over the debug information replacing any references
903 // to SafePointScalarObjectNode with the allocated object.
904 int start = jvms->debug_start();
905 int end = jvms->debug_end();
906 for (int i = start; i < end; i++) {
907 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
908 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
909 if (scobj->first_index(jvms) == sfpt_done->req() &&
910 scobj->n_fields() == (uint)nfields) {
911 assert(scobj->alloc() == alloc, "sanity");
912 sfpt_done->set_req(i, res);
913 }
914 }
915 }
916 _igvn._worklist.push(sfpt_done);
917 }
918 #ifndef PRODUCT
919 if (PrintEliminateAllocations) {
920 if (field != NULL) {
921 tty->print("=== At SafePoint node %d can't find value of Field: ",
922 sfpt->_idx);
923 field->print();
924 int field_idx = C->get_alias_index(field_addr_type);
925 tty->print(" (alias_idx=%d)", field_idx);
926 } else { // Array's element
927 tty->print("=== At SafePoint node %d can't find value of array element [%d]",
928 sfpt->_idx, j);
929 }
930 tty->print(", which prevents elimination of: ");
931 if (res == NULL)
932 alloc->dump();
933 else
934 res->dump();
935 }
936 #endif
937 return false;
938 }
939 if (field_val->is_ValueType()) {
940 // Keep track of value types to scalarize them later
941 value_worklist.push(field_val);
942 } else if (UseCompressedOops && field_type->isa_narrowoop()) {
943 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
944 // to be able scalar replace the allocation.
945 if (field_val->is_EncodeP()) {
946 field_val = field_val->in(1);
947 } else {
948 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
949 }
950 }
951 sfpt->add_req(field_val);
952 }
953 JVMState *jvms = sfpt->jvms();
954 jvms->set_endoff(sfpt->req());
955 // Now make a pass over the debug information replacing any references
956 // to the allocated object with "sobj"
957 int start = jvms->debug_start();
958 int end = jvms->debug_end();
959 sfpt->replace_edges_in_range(res, sobj, start, end);
960 _igvn._worklist.push(sfpt);
961 safepoints_done.append_if_missing(sfpt); // keep it for rollback
962 }
963 // Scalarize value types that were added to the safepoint
964 for (uint i = 0; i < value_worklist.size(); ++i) {
965 Node* vt = value_worklist.at(i);
966 vt->as_ValueType()->make_scalar_in_safepoints(&_igvn);
967 }
968 return true;
969 }
970
971 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
972 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control);
973 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory);
974 if (ctl_proj != NULL) {
975 igvn.replace_node(ctl_proj, n->in(0));
976 }
977 if (mem_proj != NULL) {
978 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
979 }
980 }
981
982 // Process users of eliminated allocation.
983 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) {
984 Node* res = alloc->result_cast();
985 if (res != NULL) {
986 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
987 Node *use = res->last_out(j);
988 uint oc1 = res->outcnt();
989
990 if (use->is_AddP()) {
991 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
992 Node *n = use->last_out(k);
993 uint oc2 = use->outcnt();
994 if (n->is_Store()) {
995 #ifdef ASSERT
996 // Verify that there is no dependent MemBarVolatile nodes,
997 // they should be removed during IGVN, see MemBarNode::Ideal().
998 for (DUIterator_Fast pmax, p = n->fast_outs(pmax);
999 p < pmax; p++) {
1000 Node* mb = n->fast_out(p);
1001 assert(mb->is_Initialize() || !mb->is_MemBar() ||
1002 mb->req() <= MemBarNode::Precedent ||
1003 mb->in(MemBarNode::Precedent) != n,
1004 "MemBarVolatile should be eliminated for non-escaping object");
1005 }
1006 #endif
1007 _igvn.replace_node(n, n->in(MemNode::Memory));
1008 } else if (n->is_ArrayCopy()) {
1009 // Disconnect ArrayCopy node
1010 ArrayCopyNode* ac = n->as_ArrayCopy();
1011 assert(ac->is_clonebasic(), "unexpected array copy kind");
1012 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
1013 disconnect_projections(ac, _igvn);
1014 assert(alloc->in(0)->is_Proj() && alloc->in(0)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
1015 Node* membar_before = alloc->in(0)->in(0);
1016 disconnect_projections(membar_before->as_MemBar(), _igvn);
1017 if (membar_after->is_MemBar()) {
1018 disconnect_projections(membar_after->as_MemBar(), _igvn);
1019 }
1020 } else {
1021 eliminate_gc_barrier(n);
1022 }
1023 k -= (oc2 - use->outcnt());
1024 }
1025 _igvn.remove_dead_node(use);
1026 } else if (use->is_ArrayCopy()) {
1027 // Disconnect ArrayCopy node
1028 ArrayCopyNode* ac = use->as_ArrayCopy();
1029 assert(ac->is_arraycopy_validated() ||
1030 ac->is_copyof_validated() ||
1031 ac->is_copyofrange_validated(), "unsupported");
1032 CallProjections* callprojs = ac->extract_projections(true);
1033
1034 _igvn.replace_node(callprojs->fallthrough_ioproj, ac->in(TypeFunc::I_O));
1035 _igvn.replace_node(callprojs->fallthrough_memproj, ac->in(TypeFunc::Memory));
1036 _igvn.replace_node(callprojs->fallthrough_catchproj, ac->in(TypeFunc::Control));
1037
1038 // Set control to top. IGVN will remove the remaining projections
1039 ac->set_req(0, top());
1040 ac->replace_edge(res, top());
1041
1042 // Disconnect src right away: it can help find new
1043 // opportunities for allocation elimination
1044 Node* src = ac->in(ArrayCopyNode::Src);
1045 ac->replace_edge(src, top());
1046 // src can be top at this point if src and dest of the
1047 // arraycopy were the same
1048 if (src->outcnt() == 0 && !src->is_top()) {
1049 _igvn.remove_dead_node(src);
1050 }
1051
1052 _igvn._worklist.push(ac);
1053 } else if (use->is_ValueType()) {
1054 assert(use->isa_ValueType()->get_oop() == res, "unexpected value type use");
1055 _igvn.rehash_node_delayed(use);
1056 use->isa_ValueType()->set_oop(_igvn.zerocon(T_VALUETYPE));
1057 } else if (use->is_Store()) {
1058 _igvn.replace_node(use, use->in(MemNode::Memory));
1059 } else {
1060 eliminate_gc_barrier(use);
1061 }
1062 j -= (oc1 - res->outcnt());
1063 }
1064 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
1065 _igvn.remove_dead_node(res);
1066 }
1067
1068 //
1069 // Process other users of allocation's projections
1070 //
1071 if (_resproj != NULL && _resproj->outcnt() != 0) {
1072 // First disconnect stores captured by Initialize node.
1073 // If Initialize node is eliminated first in the following code,
1074 // it will kill such stores and DUIterator_Last will assert.
1075 for (DUIterator_Fast jmax, j = _resproj->fast_outs(jmax); j < jmax; j++) {
1076 Node *use = _resproj->fast_out(j);
1077 if (use->is_AddP()) {
1078 // raw memory addresses used only by the initialization
1079 _igvn.replace_node(use, C->top());
1080 --j; --jmax;
1081 }
1082 }
1083 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) {
1084 Node *use = _resproj->last_out(j);
1085 uint oc1 = _resproj->outcnt();
1086 if (use->is_Initialize()) {
1087 // Eliminate Initialize node.
1088 InitializeNode *init = use->as_Initialize();
1089 assert(init->outcnt() <= 2, "only a control and memory projection expected");
1090 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control);
1091 if (ctrl_proj != NULL) {
1092 _igvn.replace_node(ctrl_proj, init->in(TypeFunc::Control));
1093 #ifdef ASSERT
1094 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1095 Node* tmp = init->in(TypeFunc::Control);
1096 while (bs->is_gc_barrier_node(tmp)) {
1097 Node* tmp2 = bs->step_over_gc_barrier_ctrl(tmp);
1098 assert(tmp != tmp2, "Must make progress");
1099 tmp = tmp2;
1100 }
1101 assert(tmp == _fallthroughcatchproj, "allocation control projection");
1102 #endif
1103 }
1104 Node *mem_proj = init->proj_out_or_null(TypeFunc::Memory);
1105 if (mem_proj != NULL) {
1106 Node *mem = init->in(TypeFunc::Memory);
1107 #ifdef ASSERT
1108 if (mem->is_MergeMem()) {
1109 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection");
1110 } else {
1111 assert(mem == _memproj_fallthrough, "allocation memory projection");
1112 }
1113 #endif
1114 _igvn.replace_node(mem_proj, mem);
1115 }
1116 } else {
1117 assert(false, "only Initialize or AddP expected");
1118 }
1119 j -= (oc1 - _resproj->outcnt());
1120 }
1121 }
1122 if (_fallthroughcatchproj != NULL) {
1123 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control));
1124 }
1125 if (_memproj_fallthrough != NULL) {
1126 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory));
1127 }
1128 if (_memproj_catchall != NULL) {
1129 _igvn.replace_node(_memproj_catchall, C->top());
1130 }
1131 if (_ioproj_fallthrough != NULL) {
1132 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O));
1133 }
1134 if (_ioproj_catchall != NULL) {
1135 _igvn.replace_node(_ioproj_catchall, C->top());
1136 }
1137 if (_catchallcatchproj != NULL) {
1138 _igvn.replace_node(_catchallcatchproj, C->top());
1139 }
1140 }
1141
1142 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1143 // Don't do scalar replacement if the frame can be popped by JVMTI:
1144 // if reallocation fails during deoptimization we'll pop all
1145 // interpreter frames for this compiled frame and that won't play
1146 // nice with JVMTI popframe.
1147 if (!EliminateAllocations || JvmtiExport::can_pop_frame() || !alloc->_is_non_escaping) {
1148 return false;
1149 }
1150 Node* klass = alloc->in(AllocateNode::KlassNode);
1151 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1152 Node* res = alloc->result_cast();
1153 // Eliminate boxing allocations which are not used
1154 // regardless scalar replacable status.
1155 bool boxing_alloc = C->eliminate_boxing() &&
1156 tklass->klass()->is_instance_klass() &&
1157 tklass->klass()->as_instance_klass()->is_box_klass();
1158 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != NULL))) {
1159 return false;
1160 }
1161
1162 extract_call_projections(alloc);
1163
1164 GrowableArray <SafePointNode *> safepoints;
1165 if (!can_eliminate_allocation(alloc, safepoints)) {
1166 return false;
1167 }
1168
1169 if (!alloc->_is_scalar_replaceable) {
1170 assert(res == NULL, "sanity");
1171 // We can only eliminate allocation if all debug info references
1172 // are already replaced with SafePointScalarObject because
1173 // we can't search for a fields value without instance_id.
1174 if (safepoints.length() > 0) {
1175 return false;
1176 }
1177 }
1178
1179 if (!scalar_replacement(alloc, safepoints)) {
1180 return false;
1181 }
1182
1183 CompileLog* log = C->log();
1184 if (log != NULL) {
1185 log->head("eliminate_allocation type='%d'",
1186 log->identify(tklass->klass()));
1187 JVMState* p = alloc->jvms();
1188 while (p != NULL) {
1189 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1190 p = p->caller();
1191 }
1192 log->tail("eliminate_allocation");
1193 }
1194
1195 process_users_of_allocation(alloc);
1196
1197 #ifndef PRODUCT
1198 if (PrintEliminateAllocations) {
1199 if (alloc->is_AllocateArray())
1200 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1201 else
1202 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1203 }
1204 #endif
1205
1206 return true;
1207 }
1208
1209 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1210 // EA should remove all uses of non-escaping boxing node.
1211 if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != NULL) {
1212 return false;
1213 }
1214
1215 assert(boxing->result_cast() == NULL, "unexpected boxing node result");
1216
1217 extract_call_projections(boxing);
1218
1219 const TypeTuple* r = boxing->tf()->range_sig();
1220 assert(r->cnt() > TypeFunc::Parms, "sanity");
1221 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1222 assert(t != NULL, "sanity");
1223
1224 CompileLog* log = C->log();
1225 if (log != NULL) {
1226 log->head("eliminate_boxing type='%d'",
1227 log->identify(t->klass()));
1228 JVMState* p = boxing->jvms();
1229 while (p != NULL) {
1230 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1231 p = p->caller();
1232 }
1233 log->tail("eliminate_boxing");
1234 }
1235
1236 process_users_of_allocation(boxing);
1237
1238 #ifndef PRODUCT
1239 if (PrintEliminateAllocations) {
1240 tty->print("++++ Eliminated: %d ", boxing->_idx);
1241 boxing->method()->print_short_name(tty);
1242 tty->cr();
1243 }
1244 #endif
1245
1246 return true;
1247 }
1248
1249 //---------------------------set_eden_pointers-------------------------
1250 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) {
1251 if (UseTLAB) { // Private allocation: load from TLS
1252 Node* thread = transform_later(new ThreadLocalNode());
1253 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset());
1254 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset());
1255 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset);
1256 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset);
1257 } else { // Shared allocation: load from globals
1258 CollectedHeap* ch = Universe::heap();
1259 address top_adr = (address)ch->top_addr();
1260 address end_adr = (address)ch->end_addr();
1261 eden_top_adr = makecon(TypeRawPtr::make(top_adr));
1262 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr);
1263 }
1264 }
1265
1266
1267 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
1268 Node* adr = basic_plus_adr(base, offset);
1269 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
1270 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered);
1271 transform_later(value);
1272 return value;
1273 }
1274
1275
1276 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
1277 Node* adr = basic_plus_adr(base, offset);
1278 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt, MemNode::unordered);
1279 transform_later(mem);
1280 return mem;
1281 }
1282
1283 //=============================================================================
1284 //
1285 // A L L O C A T I O N
1286 //
1287 // Allocation attempts to be fast in the case of frequent small objects.
1288 // It breaks down like this:
1289 //
1290 // 1) Size in doublewords is computed. This is a constant for objects and
1291 // variable for most arrays. Doubleword units are used to avoid size
1292 // overflow of huge doubleword arrays. We need doublewords in the end for
1293 // rounding.
1294 //
1295 // 2) Size is checked for being 'too large'. Too-large allocations will go
1296 // the slow path into the VM. The slow path can throw any required
1297 // exceptions, and does all the special checks for very large arrays. The
1298 // size test can constant-fold away for objects. For objects with
1299 // finalizers it constant-folds the otherway: you always go slow with
1300 // finalizers.
1301 //
1302 // 3) If NOT using TLABs, this is the contended loop-back point.
1303 // Load-Locked the heap top. If using TLABs normal-load the heap top.
1304 //
1305 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
1306 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
1307 // "size*8" we always enter the VM, where "largish" is a constant picked small
1308 // enough that there's always space between the eden max and 4Gig (old space is
1309 // there so it's quite large) and large enough that the cost of entering the VM
1310 // is dwarfed by the cost to initialize the space.
1311 //
1312 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1313 // down. If contended, repeat at step 3. If using TLABs normal-store
1314 // adjusted heap top back down; there is no contention.
1315 //
1316 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
1317 // fields.
1318 //
1319 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1320 // oop flavor.
1321 //
1322 //=============================================================================
1323 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1324 // Allocations bigger than this always go the slow route.
1325 // This value must be small enough that allocation attempts that need to
1326 // trigger exceptions go the slow route. Also, it must be small enough so
1327 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1328 //=============================================================================j//
1329 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1330 // The allocator will coalesce int->oop copies away. See comment in
1331 // coalesce.cpp about how this works. It depends critically on the exact
1332 // code shape produced here, so if you are changing this code shape
1333 // make sure the GC info for the heap-top is correct in and around the
1334 // slow-path call.
1335 //
1336
1337 void PhaseMacroExpand::expand_allocate_common(
1338 AllocateNode* alloc, // allocation node to be expanded
1339 Node* length, // array length for an array allocation
1340 const TypeFunc* slow_call_type, // Type of slow call
1341 address slow_call_address // Address of slow call
1342 )
1343 {
1344
1345 Node* ctrl = alloc->in(TypeFunc::Control);
1346 Node* mem = alloc->in(TypeFunc::Memory);
1347 Node* i_o = alloc->in(TypeFunc::I_O);
1348 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1349 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1350 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1351
1352 assert(ctrl != NULL, "must have control");
1353 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1354 // they will not be used if "always_slow" is set
1355 enum { slow_result_path = 1, fast_result_path = 2 };
1356 Node *result_region = NULL;
1357 Node *result_phi_rawmem = NULL;
1358 Node *result_phi_rawoop = NULL;
1359 Node *result_phi_i_o = NULL;
1360
1361 // The initial slow comparison is a size check, the comparison
1362 // we want to do is a BoolTest::gt
1363 bool always_slow = false;
1364 int tv = _igvn.find_int_con(initial_slow_test, -1);
1365 if (tv >= 0) {
1366 always_slow = (tv == 1);
1367 initial_slow_test = NULL;
1368 } else {
1369 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1370 }
1371
1372 if (C->env()->dtrace_alloc_probes() ||
1373 (!UseTLAB && !Universe::heap()->supports_inline_contig_alloc())) {
1374 // Force slow-path allocation
1375 always_slow = true;
1376 initial_slow_test = NULL;
1377 }
1378
1379 Node *slow_region = NULL;
1380 Node *toobig_false = ctrl;
1381
1382 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent");
1383 // generate the initial test if necessary
1384 if (initial_slow_test != NULL ) {
1385 if (slow_region == NULL) {
1386 slow_region = new RegionNode(1);
1387 }
1388 // Now make the initial failure test. Usually a too-big test but
1389 // might be a TRUE for finalizers or a fancy class check for
1390 // newInstance0.
1391 IfNode* toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1392 transform_later(toobig_iff);
1393 // Plug the failing-too-big test into the slow-path region
1394 Node* toobig_true = new IfTrueNode(toobig_iff);
1395 transform_later(toobig_true);
1396 slow_region ->add_req(toobig_true);
1397 toobig_false = new IfFalseNode(toobig_iff);
1398 transform_later(toobig_false);
1399 } else { // No initial test, just fall into next case
1400 toobig_false = ctrl;
1401 }
1402
1403 Node *slow_mem = mem; // save the current memory state for slow path
1404 // generate the fast allocation code unless we know that the initial test will always go slow
1405 if (!always_slow) {
1406 // Fast path modifies only raw memory.
1407 if (mem->is_MergeMem()) {
1408 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1409 }
1410
1411 // allocate the Region and Phi nodes for the result
1412 result_region = new RegionNode(3);
1413 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1414 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1415 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1416
1417 // Grab regular I/O before optional prefetch may change it.
1418 // Slow-path does no I/O so just set it to the original I/O.
1419 result_phi_i_o->init_req(slow_result_path, i_o);
1420
1421 Node* needgc_ctrl = NULL;
1422 // Name successful fast-path variables
1423 Node* fast_oop_ctrl;
1424 Node* fast_oop_rawmem;
1425
1426 intx prefetch_lines = length != NULL ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1427
1428 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1429 Node* fast_oop = bs->obj_allocate(this, ctrl, mem, toobig_false, size_in_bytes, i_o, needgc_ctrl,
1430 fast_oop_ctrl, fast_oop_rawmem,
1431 prefetch_lines);
1432
1433 if (slow_region != NULL) {
1434 slow_region->add_req(needgc_ctrl);
1435 // This completes all paths into the slow merge point
1436 transform_later(slow_region);
1437 } else {
1438 // Just fall from the need-GC path straight into the VM call.
1439 slow_region = needgc_ctrl;
1440 }
1441
1442 InitializeNode* init = alloc->initialization();
1443 fast_oop_rawmem = initialize_object(alloc,
1444 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1445 klass_node, length, size_in_bytes);
1446
1447 // If initialization is performed by an array copy, any required
1448 // MemBarStoreStore was already added. If the object does not
1449 // escape no need for a MemBarStoreStore. If the object does not
1450 // escape in its initializer and memory barrier (MemBarStoreStore or
1451 // stronger) is already added at exit of initializer, also no need
1452 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore
1453 // so that stores that initialize this object can't be reordered
1454 // with a subsequent store that makes this object accessible by
1455 // other threads.
1456 // Other threads include java threads and JVM internal threads
1457 // (for example concurrent GC threads). Current concurrent GC
1458 // implementation: G1 will not scan newly created object,
1459 // so it's safe to skip storestore barrier when allocation does
1460 // not escape.
1461 if (!alloc->does_not_escape_thread() &&
1462 !alloc->is_allocation_MemBar_redundant() &&
1463 (init == NULL || !init->is_complete_with_arraycopy())) {
1464 if (init == NULL || init->req() < InitializeNode::RawStores) {
1465 // No InitializeNode or no stores captured by zeroing
1466 // elimination. Simply add the MemBarStoreStore after object
1467 // initialization.
1468 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1469 transform_later(mb);
1470
1471 mb->init_req(TypeFunc::Memory, fast_oop_rawmem);
1472 mb->init_req(TypeFunc::Control, fast_oop_ctrl);
1473 fast_oop_ctrl = new ProjNode(mb,TypeFunc::Control);
1474 transform_later(fast_oop_ctrl);
1475 fast_oop_rawmem = new ProjNode(mb,TypeFunc::Memory);
1476 transform_later(fast_oop_rawmem);
1477 } else {
1478 // Add the MemBarStoreStore after the InitializeNode so that
1479 // all stores performing the initialization that were moved
1480 // before the InitializeNode happen before the storestore
1481 // barrier.
1482
1483 Node* init_ctrl = init->proj_out_or_null(TypeFunc::Control);
1484 Node* init_mem = init->proj_out_or_null(TypeFunc::Memory);
1485
1486 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1487 transform_later(mb);
1488
1489 Node* ctrl = new ProjNode(init,TypeFunc::Control);
1490 transform_later(ctrl);
1491 Node* mem = new ProjNode(init,TypeFunc::Memory);
1492 transform_later(mem);
1493
1494 // The MemBarStoreStore depends on control and memory coming
1495 // from the InitializeNode
1496 mb->init_req(TypeFunc::Memory, mem);
1497 mb->init_req(TypeFunc::Control, ctrl);
1498
1499 ctrl = new ProjNode(mb,TypeFunc::Control);
1500 transform_later(ctrl);
1501 mem = new ProjNode(mb,TypeFunc::Memory);
1502 transform_later(mem);
1503
1504 // All nodes that depended on the InitializeNode for control
1505 // and memory must now depend on the MemBarNode that itself
1506 // depends on the InitializeNode
1507 if (init_ctrl != NULL) {
1508 _igvn.replace_node(init_ctrl, ctrl);
1509 }
1510 if (init_mem != NULL) {
1511 _igvn.replace_node(init_mem, mem);
1512 }
1513 }
1514 }
1515
1516 if (C->env()->dtrace_extended_probes()) {
1517 // Slow-path call
1518 int size = TypeFunc::Parms + 2;
1519 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1520 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base),
1521 "dtrace_object_alloc",
1522 TypeRawPtr::BOTTOM);
1523
1524 // Get base of thread-local storage area
1525 Node* thread = new ThreadLocalNode();
1526 transform_later(thread);
1527
1528 call->init_req(TypeFunc::Parms+0, thread);
1529 call->init_req(TypeFunc::Parms+1, fast_oop);
1530 call->init_req(TypeFunc::Control, fast_oop_ctrl);
1531 call->init_req(TypeFunc::I_O , top()); // does no i/o
1532 call->init_req(TypeFunc::Memory , fast_oop_rawmem);
1533 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1534 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1535 transform_later(call);
1536 fast_oop_ctrl = new ProjNode(call,TypeFunc::Control);
1537 transform_later(fast_oop_ctrl);
1538 fast_oop_rawmem = new ProjNode(call,TypeFunc::Memory);
1539 transform_later(fast_oop_rawmem);
1540 }
1541
1542 // Plug in the successful fast-path into the result merge point
1543 result_region ->init_req(fast_result_path, fast_oop_ctrl);
1544 result_phi_rawoop->init_req(fast_result_path, fast_oop);
1545 result_phi_i_o ->init_req(fast_result_path, i_o);
1546 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1547 } else {
1548 slow_region = ctrl;
1549 result_phi_i_o = i_o; // Rename it to use in the following code.
1550 }
1551
1552 // Generate slow-path call
1553 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1554 OptoRuntime::stub_name(slow_call_address),
1555 alloc->jvms()->bci(),
1556 TypePtr::BOTTOM);
1557 call->init_req( TypeFunc::Control, slow_region );
1558 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o
1559 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs
1560 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) );
1561 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) );
1562
1563 call->init_req(TypeFunc::Parms+0, klass_node);
1564 if (length != NULL) {
1565 call->init_req(TypeFunc::Parms+1, length);
1566 } else {
1567 // Let the runtime know if this is a larval allocation
1568 call->init_req(TypeFunc::Parms+1, _igvn.intcon(alloc->_larval));
1569 }
1570
1571 // Copy debug information and adjust JVMState information, then replace
1572 // allocate node with the call
1573 call->copy_call_debug_info(&_igvn, alloc);
1574 if (!always_slow) {
1575 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1576 } else {
1577 // Hook i_o projection to avoid its elimination during allocation
1578 // replacement (when only a slow call is generated).
1579 call->set_req(TypeFunc::I_O, result_phi_i_o);
1580 }
1581 _igvn.replace_node(alloc, call);
1582 transform_later(call);
1583
1584 // Identify the output projections from the allocate node and
1585 // adjust any references to them.
1586 // The control and io projections look like:
1587 //
1588 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1589 // Allocate Catch
1590 // ^---Proj(io) <-------+ ^---CatchProj(io)
1591 //
1592 // We are interested in the CatchProj nodes.
1593 //
1594 extract_call_projections(call);
1595
1596 // An allocate node has separate memory projections for the uses on
1597 // the control and i_o paths. Replace the control memory projection with
1598 // result_phi_rawmem (unless we are only generating a slow call when
1599 // both memory projections are combined)
1600 if (!always_slow && _memproj_fallthrough != NULL) {
1601 _igvn.replace_in_uses(_memproj_fallthrough, result_phi_rawmem);
1602 }
1603 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete
1604 // _memproj_catchall so we end up with a call that has only 1 memory projection.
1605 if (_memproj_catchall != NULL) {
1606 if (_memproj_fallthrough == NULL) {
1607 _memproj_fallthrough = new ProjNode(call, TypeFunc::Memory);
1608 transform_later(_memproj_fallthrough);
1609 }
1610 _igvn.replace_in_uses(_memproj_catchall, _memproj_fallthrough);
1611 _igvn.remove_dead_node(_memproj_catchall);
1612 }
1613
1614 // An allocate node has separate i_o projections for the uses on the control
1615 // and i_o paths. Always replace the control i_o projection with result i_o
1616 // otherwise incoming i_o become dead when only a slow call is generated
1617 // (it is different from memory projections where both projections are
1618 // combined in such case).
1619 if (_ioproj_fallthrough != NULL) {
1620 _igvn.replace_in_uses(_ioproj_fallthrough, result_phi_i_o);
1621 }
1622 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete
1623 // _ioproj_catchall so we end up with a call that has only 1 i_o projection.
1624 if (_ioproj_catchall != NULL) {
1625 if (_ioproj_fallthrough == NULL) {
1626 _ioproj_fallthrough = new ProjNode(call, TypeFunc::I_O);
1627 transform_later(_ioproj_fallthrough);
1628 }
1629 _igvn.replace_in_uses(_ioproj_catchall, _ioproj_fallthrough);
1630 _igvn.remove_dead_node(_ioproj_catchall);
1631 }
1632
1633 // if we generated only a slow call, we are done
1634 if (always_slow) {
1635 // Now we can unhook i_o.
1636 if (result_phi_i_o->outcnt() > 1) {
1637 call->set_req(TypeFunc::I_O, top());
1638 } else {
1639 assert(result_phi_i_o->unique_ctrl_out() == call, "");
1640 // Case of new array with negative size known during compilation.
1641 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1642 // following code since call to runtime will throw exception.
1643 // As result there will be no users of i_o after the call.
1644 // Leave i_o attached to this call to avoid problems in preceding graph.
1645 }
1646 return;
1647 }
1648
1649 if (_fallthroughcatchproj != NULL) {
1650 ctrl = _fallthroughcatchproj->clone();
1651 transform_later(ctrl);
1652 _igvn.replace_node(_fallthroughcatchproj, result_region);
1653 } else {
1654 ctrl = top();
1655 }
1656 Node *slow_result;
1657 if (_resproj == NULL) {
1658 // no uses of the allocation result
1659 slow_result = top();
1660 } else {
1661 slow_result = _resproj->clone();
1662 transform_later(slow_result);
1663 _igvn.replace_node(_resproj, result_phi_rawoop);
1664 }
1665
1666 // Plug slow-path into result merge point
1667 result_region ->init_req( slow_result_path, ctrl );
1668 result_phi_rawoop->init_req( slow_result_path, slow_result);
1669 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough );
1670 transform_later(result_region);
1671 transform_later(result_phi_rawoop);
1672 transform_later(result_phi_rawmem);
1673 transform_later(result_phi_i_o);
1674 // This completes all paths into the result merge point
1675 }
1676
1677
1678 // Helper for PhaseMacroExpand::expand_allocate_common.
1679 // Initializes the newly-allocated storage.
1680 Node* PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1681 Node* control, Node* rawmem, Node* object,
1682 Node* klass_node, Node* length,
1683 Node* size_in_bytes) {
1684 InitializeNode* init = alloc->initialization();
1685 // Store the klass & mark bits
1686 Node* mark_node = alloc->make_ideal_mark(&_igvn, control, rawmem);
1687 if (!mark_node->is_Con()) {
1688 transform_later(mark_node);
1689 }
1690 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, TypeX_X->basic_type());
1691
1692 BasicType bt = T_METADATA;
1693 Node* metadata = klass_node;
1694 Node* properties = alloc->in(AllocateNode::StorageProperties);
1695 if (properties != NULL) {
1696 // Encode array storage properties into klass pointer
1697 assert(EnableValhalla, "array storage properties not supported");
1698 if (UseCompressedClassPointers) {
1699 // Compress the klass pointer before inserting the storage properties value
1700 metadata = transform_later(new EncodePKlassNode(metadata, metadata->bottom_type()->make_narrowklass()));
1701 metadata = transform_later(new CastN2INode(metadata));
1702 metadata = transform_later(new OrINode(metadata, transform_later(new ConvL2INode(properties))));
1703 bt = T_INT;
1704 } else {
1705 metadata = transform_later(new CastP2XNode(NULL, metadata));
1706 metadata = transform_later(new OrXNode(metadata, properties));
1707 bt = T_LONG;
1708 }
1709 }
1710 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), metadata, bt);
1711
1712 int header_size = alloc->minimum_header_size(); // conservatively small
1713
1714 // Array length
1715 if (length != NULL) { // Arrays need length field
1716 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1717 // conservatively small header size:
1718 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1719 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1720 if (k->is_array_klass()) // we know the exact header size in most cases:
1721 header_size = Klass::layout_helper_header_size(k->layout_helper());
1722 }
1723
1724 // Clear the object body, if necessary.
1725 if (init == NULL) {
1726 // The init has somehow disappeared; be cautious and clear everything.
1727 //
1728 // This can happen if a node is allocated but an uncommon trap occurs
1729 // immediately. In this case, the Initialize gets associated with the
1730 // trap, and may be placed in a different (outer) loop, if the Allocate
1731 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1732 // there can be two Allocates to one Initialize. The answer in all these
1733 // edge cases is safety first. It is always safe to clear immediately
1734 // within an Allocate, and then (maybe or maybe not) clear some more later.
1735 if (!(UseTLAB && ZeroTLAB)) {
1736 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1737 alloc->in(AllocateNode::DefaultValue),
1738 alloc->in(AllocateNode::RawDefaultValue),
1739 header_size, size_in_bytes,
1740 &_igvn);
1741 }
1742 } else {
1743 if (!init->is_complete()) {
1744 // Try to win by zeroing only what the init does not store.
1745 // We can also try to do some peephole optimizations,
1746 // such as combining some adjacent subword stores.
1747 rawmem = init->complete_stores(control, rawmem, object,
1748 header_size, size_in_bytes, &_igvn);
1749 }
1750 // We have no more use for this link, since the AllocateNode goes away:
1751 init->set_req(InitializeNode::RawAddress, top());
1752 // (If we keep the link, it just confuses the register allocator,
1753 // who thinks he sees a real use of the address by the membar.)
1754 }
1755
1756 return rawmem;
1757 }
1758
1759 // Generate prefetch instructions for next allocations.
1760 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
1761 Node*& contended_phi_rawmem,
1762 Node* old_eden_top, Node* new_eden_top,
1763 intx lines) {
1764 enum { fall_in_path = 1, pf_path = 2 };
1765 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
1766 // Generate prefetch allocation with watermark check.
1767 // As an allocation hits the watermark, we will prefetch starting
1768 // at a "distance" away from watermark.
1769
1770 Node *pf_region = new RegionNode(3);
1771 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY,
1772 TypeRawPtr::BOTTOM );
1773 // I/O is used for Prefetch
1774 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO );
1775
1776 Node *thread = new ThreadLocalNode();
1777 transform_later(thread);
1778
1779 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread,
1780 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
1781 transform_later(eden_pf_adr);
1782
1783 Node *old_pf_wm = new LoadPNode(needgc_false,
1784 contended_phi_rawmem, eden_pf_adr,
1785 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM,
1786 MemNode::unordered);
1787 transform_later(old_pf_wm);
1788
1789 // check against new_eden_top
1790 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm );
1791 transform_later(need_pf_cmp);
1792 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge );
1793 transform_later(need_pf_bol);
1794 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol,
1795 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1796 transform_later(need_pf_iff);
1797
1798 // true node, add prefetchdistance
1799 Node *need_pf_true = new IfTrueNode( need_pf_iff );
1800 transform_later(need_pf_true);
1801
1802 Node *need_pf_false = new IfFalseNode( need_pf_iff );
1803 transform_later(need_pf_false);
1804
1805 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm,
1806 _igvn.MakeConX(AllocatePrefetchDistance) );
1807 transform_later(new_pf_wmt );
1808 new_pf_wmt->set_req(0, need_pf_true);
1809
1810 Node *store_new_wmt = new StorePNode(need_pf_true,
1811 contended_phi_rawmem, eden_pf_adr,
1812 TypeRawPtr::BOTTOM, new_pf_wmt,
1813 MemNode::unordered);
1814 transform_later(store_new_wmt);
1815
1816 // adding prefetches
1817 pf_phi_abio->init_req( fall_in_path, i_o );
1818
1819 Node *prefetch_adr;
1820 Node *prefetch;
1821 uint step_size = AllocatePrefetchStepSize;
1822 uint distance = 0;
1823
1824 for ( intx i = 0; i < lines; i++ ) {
1825 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt,
1826 _igvn.MakeConX(distance) );
1827 transform_later(prefetch_adr);
1828 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1829 transform_later(prefetch);
1830 distance += step_size;
1831 i_o = prefetch;
1832 }
1833 pf_phi_abio->set_req( pf_path, i_o );
1834
1835 pf_region->init_req( fall_in_path, need_pf_false );
1836 pf_region->init_req( pf_path, need_pf_true );
1837
1838 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
1839 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
1840
1841 transform_later(pf_region);
1842 transform_later(pf_phi_rawmem);
1843 transform_later(pf_phi_abio);
1844
1845 needgc_false = pf_region;
1846 contended_phi_rawmem = pf_phi_rawmem;
1847 i_o = pf_phi_abio;
1848 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) {
1849 // Insert a prefetch instruction for each allocation.
1850 // This code is used to generate 1 prefetch instruction per cache line.
1851
1852 // Generate several prefetch instructions.
1853 uint step_size = AllocatePrefetchStepSize;
1854 uint distance = AllocatePrefetchDistance;
1855
1856 // Next cache address.
1857 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top,
1858 _igvn.MakeConX(step_size + distance));
1859 transform_later(cache_adr);
1860 cache_adr = new CastP2XNode(needgc_false, cache_adr);
1861 transform_later(cache_adr);
1862 // Address is aligned to execute prefetch to the beginning of cache line size
1863 // (it is important when BIS instruction is used on SPARC as prefetch).
1864 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1));
1865 cache_adr = new AndXNode(cache_adr, mask);
1866 transform_later(cache_adr);
1867 cache_adr = new CastX2PNode(cache_adr);
1868 transform_later(cache_adr);
1869
1870 // Prefetch
1871 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr );
1872 prefetch->set_req(0, needgc_false);
1873 transform_later(prefetch);
1874 contended_phi_rawmem = prefetch;
1875 Node *prefetch_adr;
1876 distance = step_size;
1877 for ( intx i = 1; i < lines; i++ ) {
1878 prefetch_adr = new AddPNode( cache_adr, cache_adr,
1879 _igvn.MakeConX(distance) );
1880 transform_later(prefetch_adr);
1881 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr );
1882 transform_later(prefetch);
1883 distance += step_size;
1884 contended_phi_rawmem = prefetch;
1885 }
1886 } else if( AllocatePrefetchStyle > 0 ) {
1887 // Insert a prefetch for each allocation only on the fast-path
1888 Node *prefetch_adr;
1889 Node *prefetch;
1890 // Generate several prefetch instructions.
1891 uint step_size = AllocatePrefetchStepSize;
1892 uint distance = AllocatePrefetchDistance;
1893 for ( intx i = 0; i < lines; i++ ) {
1894 prefetch_adr = new AddPNode( old_eden_top, new_eden_top,
1895 _igvn.MakeConX(distance) );
1896 transform_later(prefetch_adr);
1897 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1898 // Do not let it float too high, since if eden_top == eden_end,
1899 // both might be null.
1900 if( i == 0 ) { // Set control for first prefetch, next follows it
1901 prefetch->init_req(0, needgc_false);
1902 }
1903 transform_later(prefetch);
1904 distance += step_size;
1905 i_o = prefetch;
1906 }
1907 }
1908 return i_o;
1909 }
1910
1911
1912 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
1913 expand_allocate_common(alloc, NULL,
1914 OptoRuntime::new_instance_Type(),
1915 OptoRuntime::new_instance_Java());
1916 }
1917
1918 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
1919 Node* length = alloc->in(AllocateNode::ALength);
1920 InitializeNode* init = alloc->initialization();
1921 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1922 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1923 address slow_call_address; // Address of slow call
1924 if (init != NULL && init->is_complete_with_arraycopy() &&
1925 k->is_type_array_klass()) {
1926 // Don't zero type array during slow allocation in VM since
1927 // it will be initialized later by arraycopy in compiled code.
1928 slow_call_address = OptoRuntime::new_array_nozero_Java();
1929 } else {
1930 slow_call_address = OptoRuntime::new_array_Java();
1931 }
1932 expand_allocate_common(alloc, length,
1933 OptoRuntime::new_array_Type(),
1934 slow_call_address);
1935 }
1936
1937 //-------------------mark_eliminated_box----------------------------------
1938 //
1939 // During EA obj may point to several objects but after few ideal graph
1940 // transformations (CCP) it may point to only one non escaping object
1941 // (but still using phi), corresponding locks and unlocks will be marked
1942 // for elimination. Later obj could be replaced with a new node (new phi)
1943 // and which does not have escape information. And later after some graph
1944 // reshape other locks and unlocks (which were not marked for elimination
1945 // before) are connected to this new obj (phi) but they still will not be
1946 // marked for elimination since new obj has no escape information.
1947 // Mark all associated (same box and obj) lock and unlock nodes for
1948 // elimination if some of them marked already.
1949 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) {
1950 if (oldbox->as_BoxLock()->is_eliminated())
1951 return; // This BoxLock node was processed already.
1952
1953 // New implementation (EliminateNestedLocks) has separate BoxLock
1954 // node for each locked region so mark all associated locks/unlocks as
1955 // eliminated even if different objects are referenced in one locked region
1956 // (for example, OSR compilation of nested loop inside locked scope).
1957 if (EliminateNestedLocks ||
1958 oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj)) {
1959 // Box is used only in one lock region. Mark this box as eliminated.
1960 _igvn.hash_delete(oldbox);
1961 oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value
1962 _igvn.hash_insert(oldbox);
1963
1964 for (uint i = 0; i < oldbox->outcnt(); i++) {
1965 Node* u = oldbox->raw_out(i);
1966 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) {
1967 AbstractLockNode* alock = u->as_AbstractLock();
1968 // Check lock's box since box could be referenced by Lock's debug info.
1969 if (alock->box_node() == oldbox) {
1970 // Mark eliminated all related locks and unlocks.
1971 #ifdef ASSERT
1972 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4");
1973 #endif
1974 alock->set_non_esc_obj();
1975 }
1976 }
1977 }
1978 return;
1979 }
1980
1981 // Create new "eliminated" BoxLock node and use it in monitor debug info
1982 // instead of oldbox for the same object.
1983 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
1984
1985 // Note: BoxLock node is marked eliminated only here and it is used
1986 // to indicate that all associated lock and unlock nodes are marked
1987 // for elimination.
1988 newbox->set_eliminated();
1989 transform_later(newbox);
1990
1991 // Replace old box node with new box for all users of the same object.
1992 for (uint i = 0; i < oldbox->outcnt();) {
1993 bool next_edge = true;
1994
1995 Node* u = oldbox->raw_out(i);
1996 if (u->is_AbstractLock()) {
1997 AbstractLockNode* alock = u->as_AbstractLock();
1998 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) {
1999 // Replace Box and mark eliminated all related locks and unlocks.
2000 #ifdef ASSERT
2001 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5");
2002 #endif
2003 alock->set_non_esc_obj();
2004 _igvn.rehash_node_delayed(alock);
2005 alock->set_box_node(newbox);
2006 next_edge = false;
2007 }
2008 }
2009 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) {
2010 FastLockNode* flock = u->as_FastLock();
2011 assert(flock->box_node() == oldbox, "sanity");
2012 _igvn.rehash_node_delayed(flock);
2013 flock->set_box_node(newbox);
2014 next_edge = false;
2015 }
2016
2017 // Replace old box in monitor debug info.
2018 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
2019 SafePointNode* sfn = u->as_SafePoint();
2020 JVMState* youngest_jvms = sfn->jvms();
2021 int max_depth = youngest_jvms->depth();
2022 for (int depth = 1; depth <= max_depth; depth++) {
2023 JVMState* jvms = youngest_jvms->of_depth(depth);
2024 int num_mon = jvms->nof_monitors();
2025 // Loop over monitors
2026 for (int idx = 0; idx < num_mon; idx++) {
2027 Node* obj_node = sfn->monitor_obj(jvms, idx);
2028 Node* box_node = sfn->monitor_box(jvms, idx);
2029 if (box_node == oldbox && obj_node->eqv_uncast(obj)) {
2030 int j = jvms->monitor_box_offset(idx);
2031 _igvn.replace_input_of(u, j, newbox);
2032 next_edge = false;
2033 }
2034 }
2035 }
2036 }
2037 if (next_edge) i++;
2038 }
2039 }
2040
2041 //-----------------------mark_eliminated_locking_nodes-----------------------
2042 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) {
2043 if (EliminateNestedLocks) {
2044 if (alock->is_nested()) {
2045 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity");
2046 return;
2047 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened
2048 // Only Lock node has JVMState needed here.
2049 // Not that preceding claim is documented anywhere else.
2050 if (alock->jvms() != NULL) {
2051 if (alock->as_Lock()->is_nested_lock_region()) {
2052 // Mark eliminated related nested locks and unlocks.
2053 Node* obj = alock->obj_node();
2054 BoxLockNode* box_node = alock->box_node()->as_BoxLock();
2055 assert(!box_node->is_eliminated(), "should not be marked yet");
2056 // Note: BoxLock node is marked eliminated only here
2057 // and it is used to indicate that all associated lock
2058 // and unlock nodes are marked for elimination.
2059 box_node->set_eliminated(); // Box's hash is always NO_HASH here
2060 for (uint i = 0; i < box_node->outcnt(); i++) {
2061 Node* u = box_node->raw_out(i);
2062 if (u->is_AbstractLock()) {
2063 alock = u->as_AbstractLock();
2064 if (alock->box_node() == box_node) {
2065 // Verify that this Box is referenced only by related locks.
2066 assert(alock->obj_node()->eqv_uncast(obj), "");
2067 // Mark all related locks and unlocks.
2068 #ifdef ASSERT
2069 alock->log_lock_optimization(C, "eliminate_lock_set_nested");
2070 #endif
2071 alock->set_nested();
2072 }
2073 }
2074 }
2075 } else {
2076 #ifdef ASSERT
2077 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region");
2078 if (C->log() != NULL)
2079 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output
2080 #endif
2081 }
2082 }
2083 return;
2084 }
2085 // Process locks for non escaping object
2086 assert(alock->is_non_esc_obj(), "");
2087 } // EliminateNestedLocks
2088
2089 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2090 // Look for all locks of this object and mark them and
2091 // corresponding BoxLock nodes as eliminated.
2092 Node* obj = alock->obj_node();
2093 for (uint j = 0; j < obj->outcnt(); j++) {
2094 Node* o = obj->raw_out(j);
2095 if (o->is_AbstractLock() &&
2096 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2097 alock = o->as_AbstractLock();
2098 Node* box = alock->box_node();
2099 // Replace old box node with new eliminated box for all users
2100 // of the same object and mark related locks as eliminated.
2101 mark_eliminated_box(box, obj);
2102 }
2103 }
2104 }
2105 }
2106
2107 // we have determined that this lock/unlock can be eliminated, we simply
2108 // eliminate the node without expanding it.
2109 //
2110 // Note: The membar's associated with the lock/unlock are currently not
2111 // eliminated. This should be investigated as a future enhancement.
2112 //
2113 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2114
2115 if (!alock->is_eliminated()) {
2116 return false;
2117 }
2118 #ifdef ASSERT
2119 const Type* obj_type = _igvn.type(alock->obj_node());
2120 assert(!obj_type->isa_valuetype() && !obj_type->is_valuetypeptr(), "Eliminating lock on value type");
2121 if (!alock->is_coarsened()) {
2122 // Check that new "eliminated" BoxLock node is created.
2123 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2124 assert(oldbox->is_eliminated(), "should be done already");
2125 }
2126 #endif
2127
2128 alock->log_lock_optimization(C, "eliminate_lock");
2129
2130 #ifndef PRODUCT
2131 if (PrintEliminateLocks) {
2132 if (alock->is_Lock()) {
2133 tty->print_cr("++++ Eliminated: %d Lock", alock->_idx);
2134 } else {
2135 tty->print_cr("++++ Eliminated: %d Unlock", alock->_idx);
2136 }
2137 }
2138 #endif
2139
2140 Node* mem = alock->in(TypeFunc::Memory);
2141 Node* ctrl = alock->in(TypeFunc::Control);
2142 guarantee(ctrl != NULL, "missing control projection, cannot replace_node() with NULL");
2143
2144 extract_call_projections(alock);
2145 // There are 2 projections from the lock. The lock node will
2146 // be deleted when its last use is subsumed below.
2147 assert(alock->outcnt() == 2 &&
2148 _fallthroughproj != NULL &&
2149 _memproj_fallthrough != NULL,
2150 "Unexpected projections from Lock/Unlock");
2151
2152 Node* fallthroughproj = _fallthroughproj;
2153 Node* memproj_fallthrough = _memproj_fallthrough;
2154
2155 // The memory projection from a lock/unlock is RawMem
2156 // The input to a Lock is merged memory, so extract its RawMem input
2157 // (unless the MergeMem has been optimized away.)
2158 if (alock->is_Lock()) {
2159 // Seach for MemBarAcquireLock node and delete it also.
2160 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2161 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, "");
2162 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2163 Node* memproj = membar->proj_out(TypeFunc::Memory);
2164 _igvn.replace_node(ctrlproj, fallthroughproj);
2165 _igvn.replace_node(memproj, memproj_fallthrough);
2166
2167 // Delete FastLock node also if this Lock node is unique user
2168 // (a loop peeling may clone a Lock node).
2169 Node* flock = alock->as_Lock()->fastlock_node();
2170 if (flock->outcnt() == 1) {
2171 assert(flock->unique_out() == alock, "sanity");
2172 _igvn.replace_node(flock, top());
2173 }
2174 }
2175
2176 // Seach for MemBarReleaseLock node and delete it also.
2177 if (alock->is_Unlock() && ctrl->is_Proj() && ctrl->in(0)->is_MemBar()) {
2178 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2179 assert(membar->Opcode() == Op_MemBarReleaseLock &&
2180 mem->is_Proj() && membar == mem->in(0), "");
2181 _igvn.replace_node(fallthroughproj, ctrl);
2182 _igvn.replace_node(memproj_fallthrough, mem);
2183 fallthroughproj = ctrl;
2184 memproj_fallthrough = mem;
2185 ctrl = membar->in(TypeFunc::Control);
2186 mem = membar->in(TypeFunc::Memory);
2187 }
2188
2189 _igvn.replace_node(fallthroughproj, ctrl);
2190 _igvn.replace_node(memproj_fallthrough, mem);
2191 return true;
2192 }
2193
2194
2195 //------------------------------expand_lock_node----------------------
2196 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
2197
2198 Node* ctrl = lock->in(TypeFunc::Control);
2199 Node* mem = lock->in(TypeFunc::Memory);
2200 Node* obj = lock->obj_node();
2201 Node* box = lock->box_node();
2202 Node* flock = lock->fastlock_node();
2203
2204 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2205
2206 // Make the merge point
2207 Node *region;
2208 Node *mem_phi;
2209 Node *slow_path;
2210
2211 if (UseOptoBiasInlining) {
2212 /*
2213 * See the full description in MacroAssembler::biased_locking_enter().
2214 *
2215 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) {
2216 * // The object is biased.
2217 * proto_node = klass->prototype_header;
2218 * o_node = thread | proto_node;
2219 * x_node = o_node ^ mark_word;
2220 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ?
2221 * // Done.
2222 * } else {
2223 * if( (x_node & biased_lock_mask) != 0 ) {
2224 * // The klass's prototype header is no longer biased.
2225 * cas(&mark_word, mark_word, proto_node)
2226 * goto cas_lock;
2227 * } else {
2228 * // The klass's prototype header is still biased.
2229 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch?
2230 * old = mark_word;
2231 * new = o_node;
2232 * } else {
2233 * // Different thread or anonymous biased.
2234 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask);
2235 * new = thread | old;
2236 * }
2237 * // Try to rebias.
2238 * if( cas(&mark_word, old, new) == 0 ) {
2239 * // Done.
2240 * } else {
2241 * goto slow_path; // Failed.
2242 * }
2243 * }
2244 * }
2245 * } else {
2246 * // The object is not biased.
2247 * cas_lock:
2248 * if( FastLock(obj) == 0 ) {
2249 * // Done.
2250 * } else {
2251 * slow_path:
2252 * OptoRuntime::complete_monitor_locking_Java(obj);
2253 * }
2254 * }
2255 */
2256
2257 region = new RegionNode(5);
2258 // create a Phi for the memory state
2259 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2260
2261 Node* fast_lock_region = new RegionNode(3);
2262 Node* fast_lock_mem_phi = new PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM);
2263
2264 // First, check mark word for the biased lock pattern.
2265 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2266
2267 // Get fast path - mark word has the biased lock pattern.
2268 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node,
2269 markWord::biased_lock_mask_in_place,
2270 markWord::biased_lock_pattern, true);
2271 // fast_lock_region->in(1) is set to slow path.
2272 fast_lock_mem_phi->init_req(1, mem);
2273
2274 // Now check that the lock is biased to the current thread and has
2275 // the same epoch and bias as Klass::_prototype_header.
2276
2277 // Special-case a fresh allocation to avoid building nodes:
2278 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn);
2279 if (klass_node == NULL) {
2280 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
2281 klass_node = transform_later(LoadKlassNode::make(_igvn, NULL, mem, k_adr, _igvn.type(k_adr)->is_ptr()));
2282 #ifdef _LP64
2283 if (UseCompressedClassPointers && klass_node->is_DecodeNKlass()) {
2284 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity");
2285 klass_node->in(1)->init_req(0, ctrl);
2286 } else
2287 #endif
2288 klass_node->init_req(0, ctrl);
2289 }
2290 Node *proto_node = make_load(ctrl, mem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeX_X, TypeX_X->basic_type());
2291
2292 Node* thread = transform_later(new ThreadLocalNode());
2293 Node* cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2294 Node* o_node = transform_later(new OrXNode(cast_thread, proto_node));
2295 Node* x_node = transform_later(new XorXNode(o_node, mark_node));
2296
2297 // Get slow path - mark word does NOT match the value.
2298 STATIC_ASSERT(markWord::age_mask_in_place <= INT_MAX);
2299 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node,
2300 (~(int)markWord::age_mask_in_place), 0);
2301 // region->in(3) is set to fast path - the object is biased to the current thread.
2302 mem_phi->init_req(3, mem);
2303
2304
2305 // Mark word does NOT match the value (thread | Klass::_prototype_header).
2306
2307
2308 // First, check biased pattern.
2309 // Get fast path - _prototype_header has the same biased lock pattern.
2310 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node,
2311 markWord::biased_lock_mask_in_place, 0, true);
2312
2313 not_biased_ctrl = fast_lock_region->in(2); // Slow path
2314 // fast_lock_region->in(2) - the prototype header is no longer biased
2315 // and we have to revoke the bias on this object.
2316 // We are going to try to reset the mark of this object to the prototype
2317 // value and fall through to the CAS-based locking scheme.
2318 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
2319 Node* cas = new StoreXConditionalNode(not_biased_ctrl, mem, adr,
2320 proto_node, mark_node);
2321 transform_later(cas);
2322 Node* proj = transform_later(new SCMemProjNode(cas));
2323 fast_lock_mem_phi->init_req(2, proj);
2324
2325
2326 // Second, check epoch bits.
2327 Node* rebiased_region = new RegionNode(3);
2328 Node* old_phi = new PhiNode( rebiased_region, TypeX_X);
2329 Node* new_phi = new PhiNode( rebiased_region, TypeX_X);
2330
2331 // Get slow path - mark word does NOT match epoch bits.
2332 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node,
2333 markWord::epoch_mask_in_place, 0);
2334 // The epoch of the current bias is not valid, attempt to rebias the object
2335 // toward the current thread.
2336 rebiased_region->init_req(2, epoch_ctrl);
2337 old_phi->init_req(2, mark_node);
2338 new_phi->init_req(2, o_node);
2339
2340 // rebiased_region->in(1) is set to fast path.
2341 // The epoch of the current bias is still valid but we know
2342 // nothing about the owner; it might be set or it might be clear.
2343 Node* cmask = MakeConX(markWord::biased_lock_mask_in_place |
2344 markWord::age_mask_in_place |
2345 markWord::epoch_mask_in_place);
2346 Node* old = transform_later(new AndXNode(mark_node, cmask));
2347 cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2348 Node* new_mark = transform_later(new OrXNode(cast_thread, old));
2349 old_phi->init_req(1, old);
2350 new_phi->init_req(1, new_mark);
2351
2352 transform_later(rebiased_region);
2353 transform_later(old_phi);
2354 transform_later(new_phi);
2355
2356 // Try to acquire the bias of the object using an atomic operation.
2357 // If this fails we will go in to the runtime to revoke the object's bias.
2358 cas = new StoreXConditionalNode(rebiased_region, mem, adr, new_phi, old_phi);
2359 transform_later(cas);
2360 proj = transform_later(new SCMemProjNode(cas));
2361
2362 // Get slow path - Failed to CAS.
2363 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0);
2364 mem_phi->init_req(4, proj);
2365 // region->in(4) is set to fast path - the object is rebiased to the current thread.
2366
2367 // Failed to CAS.
2368 slow_path = new RegionNode(3);
2369 Node *slow_mem = new PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM);
2370
2371 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control
2372 slow_mem->init_req(1, proj);
2373
2374 // Call CAS-based locking scheme (FastLock node).
2375
2376 transform_later(fast_lock_region);
2377 transform_later(fast_lock_mem_phi);
2378
2379 // Get slow path - FastLock failed to lock the object.
2380 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0);
2381 mem_phi->init_req(2, fast_lock_mem_phi);
2382 // region->in(2) is set to fast path - the object is locked to the current thread.
2383
2384 slow_path->init_req(2, ctrl); // Capture slow-control
2385 slow_mem->init_req(2, fast_lock_mem_phi);
2386
2387 transform_later(slow_path);
2388 transform_later(slow_mem);
2389 // Reset lock's memory edge.
2390 lock->set_req(TypeFunc::Memory, slow_mem);
2391
2392 } else {
2393 region = new RegionNode(3);
2394 // create a Phi for the memory state
2395 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2396
2397 // Optimize test; set region slot 2
2398 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2399 mem_phi->init_req(2, mem);
2400 }
2401
2402 const TypeOopPtr* objptr = _igvn.type(obj)->make_oopptr();
2403 if (objptr->can_be_value_type()) {
2404 // Deoptimize and re-execute if a value
2405 assert(EnableValhalla, "should only be used if value types are enabled");
2406 Node* mark = make_load(slow_path, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2407 Node* value_mask = _igvn.MakeConX(markWord::always_locked_pattern);
2408 Node* is_value = _igvn.transform(new AndXNode(mark, value_mask));
2409 Node* cmp = _igvn.transform(new CmpXNode(is_value, value_mask));
2410 Node* bol = _igvn.transform(new BoolNode(cmp, BoolTest::eq));
2411 Node* unc_ctrl = generate_slow_guard(&slow_path, bol, NULL);
2412
2413 int trap_request = Deoptimization::make_trap_request(Deoptimization::Reason_class_check, Deoptimization::Action_none);
2414 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
2415 const TypePtr* no_memory_effects = NULL;
2416 JVMState* jvms = lock->jvms();
2417 CallNode* unc = new CallStaticJavaNode(OptoRuntime::uncommon_trap_Type(), call_addr, "uncommon_trap",
2418 jvms->bci(), no_memory_effects);
2419
2420 unc->init_req(TypeFunc::Control, unc_ctrl);
2421 unc->init_req(TypeFunc::I_O, lock->i_o());
2422 unc->init_req(TypeFunc::Memory, mem); // may gc ptrs
2423 unc->init_req(TypeFunc::FramePtr, lock->in(TypeFunc::FramePtr));
2424 unc->init_req(TypeFunc::ReturnAdr, lock->in(TypeFunc::ReturnAdr));
2425 unc->init_req(TypeFunc::Parms+0, _igvn.intcon(trap_request));
2426 unc->set_cnt(PROB_UNLIKELY_MAG(4));
2427 unc->copy_call_debug_info(&_igvn, lock);
2428
2429 assert(unc->peek_monitor_box() == box, "wrong monitor");
2430 assert(unc->peek_monitor_obj() == obj, "wrong monitor");
2431
2432 // pop monitor and push obj back on stack: we trap before the monitorenter
2433 unc->pop_monitor();
2434 unc->grow_stack(unc->jvms(), 1);
2435 unc->set_stack(unc->jvms(), unc->jvms()->stk_size()-1, obj);
2436
2437 _igvn.register_new_node_with_optimizer(unc);
2438
2439 Node* ctrl = _igvn.transform(new ProjNode(unc, TypeFunc::Control));
2440 Node* halt = _igvn.transform(new HaltNode(ctrl, lock->in(TypeFunc::FramePtr), "monitor enter on value-type"));
2441 C->root()->add_req(halt);
2442 }
2443
2444 // Make slow path call
2445 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(),
2446 OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path,
2447 obj, box, NULL);
2448
2449 extract_call_projections(call);
2450
2451 // Slow path can only throw asynchronous exceptions, which are always
2452 // de-opted. So the compiler thinks the slow-call can never throw an
2453 // exception. If it DOES throw an exception we would need the debug
2454 // info removed first (since if it throws there is no monitor).
2455 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2456 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2457
2458 // Capture slow path
2459 // disconnect fall-through projection from call and create a new one
2460 // hook up users of fall-through projection to region
2461 Node *slow_ctrl = _fallthroughproj->clone();
2462 transform_later(slow_ctrl);
2463 _igvn.hash_delete(_fallthroughproj);
2464 _fallthroughproj->disconnect_inputs(NULL, C);
2465 region->init_req(1, slow_ctrl);
2466 // region inputs are now complete
2467 transform_later(region);
2468 _igvn.replace_node(_fallthroughproj, region);
2469
2470 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2471 mem_phi->init_req(1, memproj );
2472 transform_later(mem_phi);
2473 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2474 }
2475
2476 //------------------------------expand_unlock_node----------------------
2477 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2478
2479 Node* ctrl = unlock->in(TypeFunc::Control);
2480 Node* mem = unlock->in(TypeFunc::Memory);
2481 Node* obj = unlock->obj_node();
2482 Node* box = unlock->box_node();
2483
2484 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2485
2486 // No need for a null check on unlock
2487
2488 // Make the merge point
2489 Node *region;
2490 Node *mem_phi;
2491
2492 if (UseOptoBiasInlining) {
2493 // Check for biased locking unlock case, which is a no-op.
2494 // See the full description in MacroAssembler::biased_locking_exit().
2495 region = new RegionNode(4);
2496 // create a Phi for the memory state
2497 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2498 mem_phi->init_req(3, mem);
2499
2500 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2501 ctrl = opt_bits_test(ctrl, region, 3, mark_node,
2502 markWord::biased_lock_mask_in_place,
2503 markWord::biased_lock_pattern);
2504 } else {
2505 region = new RegionNode(3);
2506 // create a Phi for the memory state
2507 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2508 }
2509
2510 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2511 funlock = transform_later( funlock )->as_FastUnlock();
2512 // Optimize test; set region slot 2
2513 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2514 Node *thread = transform_later(new ThreadLocalNode());
2515
2516 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2517 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2518 "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2519
2520 extract_call_projections(call);
2521
2522 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2523 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2524
2525 // No exceptions for unlocking
2526 // Capture slow path
2527 // disconnect fall-through projection from call and create a new one
2528 // hook up users of fall-through projection to region
2529 Node *slow_ctrl = _fallthroughproj->clone();
2530 transform_later(slow_ctrl);
2531 _igvn.hash_delete(_fallthroughproj);
2532 _fallthroughproj->disconnect_inputs(NULL, C);
2533 region->init_req(1, slow_ctrl);
2534 // region inputs are now complete
2535 transform_later(region);
2536 _igvn.replace_node(_fallthroughproj, region);
2537
2538 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2539 mem_phi->init_req(1, memproj );
2540 mem_phi->init_req(2, mem);
2541 transform_later(mem_phi);
2542 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2543 }
2544
2545 // A value type might be returned from the call but we don't know its
2546 // type. Either we get a buffered value (and nothing needs to be done)
2547 // or one of the values being returned is the klass of the value type
2548 // and we need to allocate a value type instance of that type and
2549 // initialize it with other values being returned. In that case, we
2550 // first try a fast path allocation and initialize the value with the
2551 // value klass's pack handler or we fall back to a runtime call.
2552 void PhaseMacroExpand::expand_mh_intrinsic_return(CallStaticJavaNode* call) {
2553 assert(call->method()->is_method_handle_intrinsic(), "must be a method handle intrinsic call");
2554 Node* ret = call->proj_out_or_null(TypeFunc::Parms);
2555 if (ret == NULL) {
2556 return;
2557 }
2558 const TypeFunc* tf = call->_tf;
2559 const TypeTuple* domain = OptoRuntime::store_value_type_fields_Type()->domain_cc();
2560 const TypeFunc* new_tf = TypeFunc::make(tf->domain_sig(), tf->domain_cc(), tf->range_sig(), domain);
2561 call->_tf = new_tf;
2562 // Make sure the change of type is applied before projections are processed by igvn
2563 _igvn.set_type(call, call->Value(&_igvn));
2564 _igvn.set_type(ret, ret->Value(&_igvn));
2565
2566 // Before any new projection is added:
2567 CallProjections* projs = call->extract_projections(true, true);
2568
2569 Node* ctl = new Node(1);
2570 Node* mem = new Node(1);
2571 Node* io = new Node(1);
2572 Node* ex_ctl = new Node(1);
2573 Node* ex_mem = new Node(1);
2574 Node* ex_io = new Node(1);
2575 Node* res = new Node(1);
2576
2577 Node* cast = transform_later(new CastP2XNode(ctl, res));
2578 Node* mask = MakeConX(0x1);
2579 Node* masked = transform_later(new AndXNode(cast, mask));
2580 Node* cmp = transform_later(new CmpXNode(masked, mask));
2581 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
2582 IfNode* allocation_iff = new IfNode(ctl, bol, PROB_MAX, COUNT_UNKNOWN);
2583 transform_later(allocation_iff);
2584 Node* allocation_ctl = transform_later(new IfTrueNode(allocation_iff));
2585 Node* no_allocation_ctl = transform_later(new IfFalseNode(allocation_iff));
2586
2587 Node* no_allocation_res = transform_later(new CheckCastPPNode(no_allocation_ctl, res, TypeInstPtr::BOTTOM));
2588
2589 Node* mask2 = MakeConX(-2);
2590 Node* masked2 = transform_later(new AndXNode(cast, mask2));
2591 Node* rawklassptr = transform_later(new CastX2PNode(masked2));
2592 Node* klass_node = transform_later(new CheckCastPPNode(allocation_ctl, rawklassptr, TypeKlassPtr::OBJECT_OR_NULL));
2593
2594 Node* slowpath_bol = NULL;
2595 Node* top_adr = NULL;
2596 Node* old_top = NULL;
2597 Node* new_top = NULL;
2598 if (UseTLAB) {
2599 Node* end_adr = NULL;
2600 set_eden_pointers(top_adr, end_adr);
2601 Node* end = make_load(ctl, mem, end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS);
2602 old_top = new LoadPNode(ctl, mem, top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered);
2603 transform_later(old_top);
2604 Node* layout_val = make_load(NULL, mem, klass_node, in_bytes(Klass::layout_helper_offset()), TypeInt::INT, T_INT);
2605 Node* size_in_bytes = ConvI2X(layout_val);
2606 new_top = new AddPNode(top(), old_top, size_in_bytes);
2607 transform_later(new_top);
2608 Node* slowpath_cmp = new CmpPNode(new_top, end);
2609 transform_later(slowpath_cmp);
2610 slowpath_bol = new BoolNode(slowpath_cmp, BoolTest::ge);
2611 transform_later(slowpath_bol);
2612 } else {
2613 slowpath_bol = intcon(1);
2614 top_adr = top();
2615 old_top = top();
2616 new_top = top();
2617 }
2618 IfNode* slowpath_iff = new IfNode(allocation_ctl, slowpath_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN);
2619 transform_later(slowpath_iff);
2620
2621 Node* slowpath_true = new IfTrueNode(slowpath_iff);
2622 transform_later(slowpath_true);
2623
2624 CallStaticJavaNode* slow_call = new CallStaticJavaNode(OptoRuntime::store_value_type_fields_Type(),
2625 StubRoutines::store_value_type_fields_to_buf(),
2626 "store_value_type_fields",
2627 call->jvms()->bci(),
2628 TypePtr::BOTTOM);
2629 slow_call->init_req(TypeFunc::Control, slowpath_true);
2630 slow_call->init_req(TypeFunc::Memory, mem);
2631 slow_call->init_req(TypeFunc::I_O, io);
2632 slow_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2633 slow_call->init_req(TypeFunc::ReturnAdr, call->in(TypeFunc::ReturnAdr));
2634 slow_call->init_req(TypeFunc::Parms, res);
2635
2636 Node* slow_ctl = transform_later(new ProjNode(slow_call, TypeFunc::Control));
2637 Node* slow_mem = transform_later(new ProjNode(slow_call, TypeFunc::Memory));
2638 Node* slow_io = transform_later(new ProjNode(slow_call, TypeFunc::I_O));
2639 Node* slow_res = transform_later(new ProjNode(slow_call, TypeFunc::Parms));
2640 Node* slow_catc = transform_later(new CatchNode(slow_ctl, slow_io, 2));
2641 Node* slow_norm = transform_later(new CatchProjNode(slow_catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci));
2642 Node* slow_excp = transform_later(new CatchProjNode(slow_catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci));
2643
2644 Node* ex_r = new RegionNode(3);
2645 Node* ex_mem_phi = new PhiNode(ex_r, Type::MEMORY, TypePtr::BOTTOM);
2646 Node* ex_io_phi = new PhiNode(ex_r, Type::ABIO);
2647 ex_r->init_req(1, slow_excp);
2648 ex_mem_phi->init_req(1, slow_mem);
2649 ex_io_phi->init_req(1, slow_io);
2650 ex_r->init_req(2, ex_ctl);
2651 ex_mem_phi->init_req(2, ex_mem);
2652 ex_io_phi->init_req(2, ex_io);
2653
2654 transform_later(ex_r);
2655 transform_later(ex_mem_phi);
2656 transform_later(ex_io_phi);
2657
2658 Node* slowpath_false = new IfFalseNode(slowpath_iff);
2659 transform_later(slowpath_false);
2660 Node* rawmem = new StorePNode(slowpath_false, mem, top_adr, TypeRawPtr::BOTTOM, new_top, MemNode::unordered);
2661 transform_later(rawmem);
2662 Node* mark_node = makecon(TypeRawPtr::make((address)markWord::always_locked_prototype().value()));
2663 rawmem = make_store(slowpath_false, rawmem, old_top, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
2664 rawmem = make_store(slowpath_false, rawmem, old_top, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
2665 if (UseCompressedClassPointers) {
2666 rawmem = make_store(slowpath_false, rawmem, old_top, oopDesc::klass_gap_offset_in_bytes(), intcon(0), T_INT);
2667 }
2668 Node* fixed_block = make_load(slowpath_false, rawmem, klass_node, in_bytes(InstanceKlass::adr_valueklass_fixed_block_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
2669 Node* pack_handler = make_load(slowpath_false, rawmem, fixed_block, in_bytes(ValueKlass::pack_handler_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
2670
2671 CallLeafNoFPNode* handler_call = new CallLeafNoFPNode(OptoRuntime::pack_value_type_Type(),
2672 NULL,
2673 "pack handler",
2674 TypeRawPtr::BOTTOM);
2675 handler_call->init_req(TypeFunc::Control, slowpath_false);
2676 handler_call->init_req(TypeFunc::Memory, rawmem);
2677 handler_call->init_req(TypeFunc::I_O, top());
2678 handler_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2679 handler_call->init_req(TypeFunc::ReturnAdr, top());
2680 handler_call->init_req(TypeFunc::Parms, pack_handler);
2681 handler_call->init_req(TypeFunc::Parms+1, old_top);
2682
2683 // We don't know how many values are returned. This assumes the
2684 // worst case, that all available registers are used.
2685 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) {
2686 if (domain->field_at(i) == Type::HALF) {
2687 slow_call->init_req(i, top());
2688 handler_call->init_req(i+1, top());
2689 continue;
2690 }
2691 Node* proj = transform_later(new ProjNode(call, i));
2692 slow_call->init_req(i, proj);
2693 handler_call->init_req(i+1, proj);
2694 }
2695
2696 // We can safepoint at that new call
2697 slow_call->copy_call_debug_info(&_igvn, call);
2698 transform_later(slow_call);
2699 transform_later(handler_call);
2700
2701 Node* handler_ctl = transform_later(new ProjNode(handler_call, TypeFunc::Control));
2702 rawmem = transform_later(new ProjNode(handler_call, TypeFunc::Memory));
2703 Node* slowpath_false_res = transform_later(new ProjNode(handler_call, TypeFunc::Parms));
2704
2705 MergeMemNode* slowpath_false_mem = MergeMemNode::make(mem);
2706 slowpath_false_mem->set_memory_at(Compile::AliasIdxRaw, rawmem);
2707 transform_later(slowpath_false_mem);
2708
2709 Node* r = new RegionNode(4);
2710 Node* mem_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM);
2711 Node* io_phi = new PhiNode(r, Type::ABIO);
2712 Node* res_phi = new PhiNode(r, TypeInstPtr::BOTTOM);
2713
2714 r->init_req(1, no_allocation_ctl);
2715 mem_phi->init_req(1, mem);
2716 io_phi->init_req(1, io);
2717 res_phi->init_req(1, no_allocation_res);
2718 r->init_req(2, slow_norm);
2719 mem_phi->init_req(2, slow_mem);
2720 io_phi->init_req(2, slow_io);
2721 res_phi->init_req(2, slow_res);
2722 r->init_req(3, handler_ctl);
2723 mem_phi->init_req(3, slowpath_false_mem);
2724 io_phi->init_req(3, io);
2725 res_phi->init_req(3, slowpath_false_res);
2726
2727 transform_later(r);
2728 transform_later(mem_phi);
2729 transform_later(io_phi);
2730 transform_later(res_phi);
2731
2732 assert(projs->nb_resproj == 1, "unexpected number of results");
2733 _igvn.replace_in_uses(projs->fallthrough_catchproj, r);
2734 _igvn.replace_in_uses(projs->fallthrough_memproj, mem_phi);
2735 _igvn.replace_in_uses(projs->fallthrough_ioproj, io_phi);
2736 _igvn.replace_in_uses(projs->resproj[0], res_phi);
2737 _igvn.replace_in_uses(projs->catchall_catchproj, ex_r);
2738 _igvn.replace_in_uses(projs->catchall_memproj, ex_mem_phi);
2739 _igvn.replace_in_uses(projs->catchall_ioproj, ex_io_phi);
2740
2741 _igvn.replace_node(ctl, projs->fallthrough_catchproj);
2742 _igvn.replace_node(mem, projs->fallthrough_memproj);
2743 _igvn.replace_node(io, projs->fallthrough_ioproj);
2744 _igvn.replace_node(res, projs->resproj[0]);
2745 _igvn.replace_node(ex_ctl, projs->catchall_catchproj);
2746 _igvn.replace_node(ex_mem, projs->catchall_memproj);
2747 _igvn.replace_node(ex_io, projs->catchall_ioproj);
2748 }
2749
2750 //---------------------------eliminate_macro_nodes----------------------
2751 // Eliminate scalar replaced allocations and associated locks.
2752 void PhaseMacroExpand::eliminate_macro_nodes() {
2753 if (C->macro_count() == 0)
2754 return;
2755
2756 // First, attempt to eliminate locks
2757 int cnt = C->macro_count();
2758 for (int i=0; i < cnt; i++) {
2759 Node *n = C->macro_node(i);
2760 if (n->is_AbstractLock()) { // Lock and Unlock nodes
2761 // Before elimination mark all associated (same box and obj)
2762 // lock and unlock nodes.
2763 mark_eliminated_locking_nodes(n->as_AbstractLock());
2764 }
2765 }
2766 bool progress = true;
2767 while (progress) {
2768 progress = false;
2769 for (int i = C->macro_count(); i > 0; i--) {
2770 Node * n = C->macro_node(i-1);
2771 bool success = false;
2772 debug_only(int old_macro_count = C->macro_count(););
2773 if (n->is_AbstractLock()) {
2774 success = eliminate_locking_node(n->as_AbstractLock());
2775 }
2776 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2777 progress = progress || success;
2778 }
2779 }
2780 // Next, attempt to eliminate allocations
2781 _has_locks = false;
2782 progress = true;
2783 while (progress) {
2784 progress = false;
2785 for (int i = C->macro_count(); i > 0; i--) {
2786 Node * n = C->macro_node(i-1);
2787 bool success = false;
2788 debug_only(int old_macro_count = C->macro_count(););
2789 switch (n->class_id()) {
2790 case Node::Class_Allocate:
2791 case Node::Class_AllocateArray:
2792 success = eliminate_allocate_node(n->as_Allocate());
2793 break;
2794 case Node::Class_CallStaticJava: {
2795 CallStaticJavaNode* call = n->as_CallStaticJava();
2796 if (!call->method()->is_method_handle_intrinsic()) {
2797 success = eliminate_boxing_node(n->as_CallStaticJava());
2798 }
2799 break;
2800 }
2801 case Node::Class_Lock:
2802 case Node::Class_Unlock:
2803 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2804 _has_locks = true;
2805 break;
2806 case Node::Class_ArrayCopy:
2807 break;
2808 case Node::Class_OuterStripMinedLoop:
2809 break;
2810 default:
2811 assert(n->Opcode() == Op_LoopLimit ||
2812 n->Opcode() == Op_Opaque1 ||
2813 n->Opcode() == Op_Opaque2 ||
2814 n->Opcode() == Op_Opaque3 ||
2815 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n),
2816 "unknown node type in macro list");
2817 }
2818 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2819 progress = progress || success;
2820 }
2821 }
2822 }
2823
2824 //------------------------------expand_macro_nodes----------------------
2825 // Returns true if a failure occurred.
2826 bool PhaseMacroExpand::expand_macro_nodes() {
2827 // Last attempt to eliminate macro nodes.
2828 eliminate_macro_nodes();
2829
2830 // Make sure expansion will not cause node limit to be exceeded.
2831 // Worst case is a macro node gets expanded into about 200 nodes.
2832 // Allow 50% more for optimization.
2833 if (C->check_node_count(C->macro_count() * 300, "out of nodes before macro expansion" ) )
2834 return true;
2835
2836 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2837 bool progress = true;
2838 while (progress) {
2839 progress = false;
2840 for (int i = C->macro_count(); i > 0; i--) {
2841 Node* n = C->macro_node(i-1);
2842 bool success = false;
2843 debug_only(int old_macro_count = C->macro_count(););
2844 if (n->Opcode() == Op_LoopLimit) {
2845 // Remove it from macro list and put on IGVN worklist to optimize.
2846 C->remove_macro_node(n);
2847 _igvn._worklist.push(n);
2848 success = true;
2849 } else if (n->Opcode() == Op_CallStaticJava) {
2850 CallStaticJavaNode* call = n->as_CallStaticJava();
2851 if (!call->method()->is_method_handle_intrinsic()) {
2852 // Remove it from macro list and put on IGVN worklist to optimize.
2853 C->remove_macro_node(n);
2854 _igvn._worklist.push(n);
2855 success = true;
2856 }
2857 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) {
2858 _igvn.replace_node(n, n->in(1));
2859 success = true;
2860 #if INCLUDE_RTM_OPT
2861 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) {
2862 assert(C->profile_rtm(), "should be used only in rtm deoptimization code");
2863 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), "");
2864 Node* cmp = n->unique_out();
2865 #ifdef ASSERT
2866 // Validate graph.
2867 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), "");
2868 BoolNode* bol = cmp->unique_out()->as_Bool();
2869 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() &&
2870 (bol->_test._test == BoolTest::ne), "");
2871 IfNode* ifn = bol->unique_out()->as_If();
2872 assert((ifn->outcnt() == 2) &&
2873 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != NULL, "");
2874 #endif
2875 Node* repl = n->in(1);
2876 if (!_has_locks) {
2877 // Remove RTM state check if there are no locks in the code.
2878 // Replace input to compare the same value.
2879 repl = (cmp->in(1) == n) ? cmp->in(2) : cmp->in(1);
2880 }
2881 _igvn.replace_node(n, repl);
2882 success = true;
2883 #endif
2884 } else if (n->Opcode() == Op_OuterStripMinedLoop) {
2885 n->as_OuterStripMinedLoop()->adjust_strip_mined_loop(&_igvn);
2886 C->remove_macro_node(n);
2887 success = true;
2888 }
2889 assert(!success || (C->macro_count() == (old_macro_count - 1)), "elimination must have deleted one node from macro list");
2890 progress = progress || success;
2891 }
2892 }
2893
2894 // expand arraycopy "macro" nodes first
2895 // For ReduceBulkZeroing, we must first process all arraycopy nodes
2896 // before the allocate nodes are expanded.
2897 for (int i = C->macro_count(); i > 0; i--) {
2898 Node* n = C->macro_node(i-1);
2899 assert(n->is_macro(), "only macro nodes expected here");
2900 if (_igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())) {
2901 // node is unreachable, so don't try to expand it
2902 C->remove_macro_node(n);
2903 continue;
2904 }
2905 debug_only(int old_macro_count = C->macro_count(););
2906 switch (n->class_id()) {
2907 case Node::Class_Lock:
2908 expand_lock_node(n->as_Lock());
2909 assert(C->macro_count() == (old_macro_count - 1), "expansion must have deleted one node from macro list");
2910 break;
2911 case Node::Class_Unlock:
2912 expand_unlock_node(n->as_Unlock());
2913 assert(C->macro_count() == (old_macro_count - 1), "expansion must have deleted one node from macro list");
2914 break;
2915 case Node::Class_ArrayCopy:
2916 expand_arraycopy_node(n->as_ArrayCopy());
2917 assert(C->macro_count() == (old_macro_count - 1), "expansion must have deleted one node from macro list");
2918 break;
2919 }
2920 if (C->failing()) return true;
2921 }
2922
2923 // All nodes except Allocate nodes are expanded now. There could be
2924 // new optimization opportunities (such as folding newly created
2925 // load from a just allocated object). Run IGVN.
2926 _igvn.set_delay_transform(false);
2927 _igvn.optimize();
2928 if (C->failing()) return true;
2929
2930 _igvn.set_delay_transform(true);
2931
2932 // expand "macro" nodes
2933 // nodes are removed from the macro list as they are processed
2934 while (C->macro_count() > 0) {
2935 int macro_count = C->macro_count();
2936 Node * n = C->macro_node(macro_count-1);
2937 assert(n->is_macro(), "only macro nodes expected here");
2938 if (_igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())) {
2939 // node is unreachable, so don't try to expand it
2940 C->remove_macro_node(n);
2941 continue;
2942 }
2943 switch (n->class_id()) {
2944 case Node::Class_Allocate:
2945 expand_allocate(n->as_Allocate());
2946 break;
2947 case Node::Class_AllocateArray:
2948 expand_allocate_array(n->as_AllocateArray());
2949 break;
2950 case Node::Class_CallStaticJava:
2951 expand_mh_intrinsic_return(n->as_CallStaticJava());
2952 C->remove_macro_node(n);
2953 break;
2954 default:
2955 assert(false, "unknown node type in macro list");
2956 }
2957 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2958 if (C->failing()) return true;
2959 }
2960
2961 _igvn.set_delay_transform(false);
2962 _igvn.optimize();
2963 if (C->failing()) return true;
2964 return false;
2965 }