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 "compiler/compileLog.hpp"
26 #include "gc/shared/collectedHeap.inline.hpp"
27 #include "gc/shared/tlab_globals.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/intrinsicnode.hpp"
39 #include "opto/locknode.hpp"
40 #include "opto/loopnode.hpp"
41 #include "opto/macro.hpp"
42 #include "opto/memnode.hpp"
43 #include "opto/narrowptrnode.hpp"
44 #include "opto/node.hpp"
45 #include "opto/opaquenode.hpp"
46 #include "opto/phaseX.hpp"
47 #include "opto/rootnode.hpp"
48 #include "opto/runtime.hpp"
49 #include "opto/subnode.hpp"
50 #include "opto/subtypenode.hpp"
51 #include "opto/type.hpp"
52 #include "prims/jvmtiExport.hpp"
53 #include "runtime/continuation.hpp"
54 #include "runtime/sharedRuntime.hpp"
55 #include "utilities/macros.hpp"
56 #include "utilities/powerOfTwo.hpp"
57 #if INCLUDE_G1GC
58 #include "gc/g1/g1ThreadLocalData.hpp"
59 #endif // INCLUDE_G1GC
60
61
62 //
63 // Replace any references to "oldref" in inputs to "use" with "newref".
64 // Returns the number of replacements made.
65 //
66 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
67 int nreplacements = 0;
68 uint req = use->req();
69 for (uint j = 0; j < use->len(); j++) {
70 Node *uin = use->in(j);
71 if (uin == oldref) {
72 if (j < req)
73 use->set_req(j, newref);
74 else
75 use->set_prec(j, newref);
76 nreplacements++;
77 } else if (j >= req && uin == nullptr) {
78 break;
79 }
80 }
81 return nreplacements;
82 }
83
84 void PhaseMacroExpand::migrate_outs(Node *old, Node *target) {
85 assert(old != nullptr, "sanity");
86 for (DUIterator_Fast imax, i = old->fast_outs(imax); i < imax; i++) {
87 Node* use = old->fast_out(i);
88 _igvn.rehash_node_delayed(use);
89 imax -= replace_input(use, old, target);
90 // back up iterator
91 --i;
92 }
93 assert(old->outcnt() == 0, "all uses must be deleted");
94 }
95
96 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word) {
97 Node* cmp = word;
98 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne));
99 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
100 transform_later(iff);
101
102 // Fast path taken.
103 Node *fast_taken = transform_later(new IfFalseNode(iff));
104
105 // Fast path not-taken, i.e. slow path
106 Node *slow_taken = transform_later(new IfTrueNode(iff));
107
108 region->init_req(edge, fast_taken); // Capture fast-control
109 return slow_taken;
110 }
111
112 //--------------------copy_predefined_input_for_runtime_call--------------------
113 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
114 // Set fixed predefined input arguments
127 // Slow-path call
128 CallNode *call = leaf_name
129 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
130 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), TypeRawPtr::BOTTOM );
131
132 // Slow path call has no side-effects, uses few values
133 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
134 if (parm0 != nullptr) call->init_req(TypeFunc::Parms+0, parm0);
135 if (parm1 != nullptr) call->init_req(TypeFunc::Parms+1, parm1);
136 if (parm2 != nullptr) call->init_req(TypeFunc::Parms+2, parm2);
137 call->copy_call_debug_info(&_igvn, oldcall);
138 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
139 _igvn.replace_node(oldcall, call);
140 transform_later(call);
141
142 return call;
143 }
144
145 void PhaseMacroExpand::eliminate_gc_barrier(Node* p2x) {
146 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
147 bs->eliminate_gc_barrier(this, p2x);
148 #ifndef PRODUCT
149 if (PrintOptoStatistics) {
150 AtomicAccess::inc(&PhaseMacroExpand::_GC_barriers_removed_counter);
151 }
152 #endif
153 }
154
155 // Search for a memory operation for the specified memory slice.
156 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
157 Node *orig_mem = mem;
158 Node *alloc_mem = alloc->as_Allocate()->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
159 assert(alloc_mem != nullptr, "Allocation without a memory projection.");
160 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
161 while (true) {
162 if (mem == alloc_mem || mem == start_mem ) {
163 return mem; // hit one of our sentinels
164 } else if (mem->is_MergeMem()) {
165 mem = mem->as_MergeMem()->memory_at(alias_idx);
166 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
167 Node *in = mem->in(0);
170 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
171 return in;
172 } else if (in->is_Call()) {
173 CallNode *call = in->as_Call();
174 if (call->may_modify(tinst, phase)) {
175 assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape");
176 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) {
177 return in;
178 }
179 }
180 mem = in->in(TypeFunc::Memory);
181 } else if (in->is_MemBar()) {
182 ArrayCopyNode* ac = nullptr;
183 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) {
184 if (ac != nullptr) {
185 assert(ac->is_clonebasic(), "Only basic clone is a non escaping clone");
186 return ac;
187 }
188 }
189 mem = in->in(TypeFunc::Memory);
190 } else {
191 #ifdef ASSERT
192 in->dump();
193 mem->dump();
194 assert(false, "unexpected projection");
195 #endif
196 }
197 } else if (mem->is_Store()) {
198 const TypePtr* atype = mem->as_Store()->adr_type();
199 int adr_idx = phase->C->get_alias_index(atype);
200 if (adr_idx == alias_idx) {
201 assert(atype->isa_oopptr(), "address type must be oopptr");
202 int adr_offset = atype->offset();
203 uint adr_iid = atype->is_oopptr()->instance_id();
204 // Array elements references have the same alias_idx
205 // but different offset and different instance_id.
206 if (adr_offset == offset && adr_iid == alloc->_idx) {
207 return mem;
208 }
209 } else {
210 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
211 }
212 mem = mem->in(MemNode::Memory);
213 } else if (mem->is_ClearArray()) {
214 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
215 // Can not bypass initialization of the instance
216 // we are looking.
217 DEBUG_ONLY(intptr_t offset;)
218 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
219 InitializeNode* init = alloc->as_Allocate()->initialization();
220 // We are looking for stored value, return Initialize node
221 // or memory edge from Allocate node.
222 if (init != nullptr) {
227 }
228 // Otherwise skip it (the call updated 'mem' value).
229 } else if (mem->Opcode() == Op_SCMemProj) {
230 mem = mem->in(0);
231 Node* adr = nullptr;
232 if (mem->is_LoadStore()) {
233 adr = mem->in(MemNode::Address);
234 } else {
235 assert(mem->Opcode() == Op_EncodeISOArray ||
236 mem->Opcode() == Op_StrCompressedCopy, "sanity");
237 adr = mem->in(3); // Destination array
238 }
239 const TypePtr* atype = adr->bottom_type()->is_ptr();
240 int adr_idx = phase->C->get_alias_index(atype);
241 if (adr_idx == alias_idx) {
242 DEBUG_ONLY(mem->dump();)
243 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
244 return nullptr;
245 }
246 mem = mem->in(MemNode::Memory);
247 } else if (mem->Opcode() == Op_StrInflatedCopy) {
248 Node* adr = mem->in(3); // Destination array
249 const TypePtr* atype = adr->bottom_type()->is_ptr();
250 int adr_idx = phase->C->get_alias_index(atype);
251 if (adr_idx == alias_idx) {
252 DEBUG_ONLY(mem->dump();)
253 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
254 return nullptr;
255 }
256 mem = mem->in(MemNode::Memory);
257 } else {
258 return mem;
259 }
260 assert(mem != orig_mem, "dead memory loop");
261 }
262 }
263
264 // Generate loads from source of the arraycopy for fields of
265 // destination needed at a deoptimization point
266 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) {
267 BasicType bt = ft;
272 }
273 Node* res = nullptr;
274 if (ac->is_clonebasic()) {
275 assert(ac->in(ArrayCopyNode::Src) != ac->in(ArrayCopyNode::Dest), "clone source equals destination");
276 Node* base = ac->in(ArrayCopyNode::Src);
277 Node* adr = _igvn.transform(AddPNode::make_with_base(base, _igvn.MakeConX(offset)));
278 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
279 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
280 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
281 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt);
282 } else {
283 if (ac->modifies(offset, offset, &_igvn, true)) {
284 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
285 uint shift = exact_log2(type2aelembytes(bt));
286 Node* src_pos = ac->in(ArrayCopyNode::SrcPos);
287 Node* dest_pos = ac->in(ArrayCopyNode::DestPos);
288 const TypeInt* src_pos_t = _igvn.type(src_pos)->is_int();
289 const TypeInt* dest_pos_t = _igvn.type(dest_pos)->is_int();
290
291 Node* adr = nullptr;
292 const TypePtr* adr_type = nullptr;
293 if (src_pos_t->is_con() && dest_pos_t->is_con()) {
294 intptr_t off = ((src_pos_t->get_con() - dest_pos_t->get_con()) << shift) + offset;
295 Node* base = ac->in(ArrayCopyNode::Src);
296 adr = _igvn.transform(AddPNode::make_with_base(base, _igvn.MakeConX(off)));
297 adr_type = _igvn.type(base)->is_ptr()->add_offset(off);
298 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
299 // Don't emit a new load from src if src == dst but try to get the value from memory instead
300 return value_from_mem(ac->in(TypeFunc::Memory), ctl, ft, ftype, adr_type->isa_oopptr(), alloc);
301 }
302 } else {
303 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
304 #ifdef _LP64
305 diff = _igvn.transform(new ConvI2LNode(diff));
306 #endif
307 diff = _igvn.transform(new LShiftXNode(diff, _igvn.intcon(shift)));
308
309 Node* off = _igvn.transform(new AddXNode(_igvn.MakeConX(offset), diff));
310 Node* base = ac->in(ArrayCopyNode::Src);
311 adr = _igvn.transform(AddPNode::make_with_base(base, off));
312 adr_type = _igvn.type(base)->is_ptr()->add_offset(Type::OffsetBot);
313 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
314 // Non constant offset in the array: we can't statically
315 // determine the value
316 return nullptr;
317 }
318 }
319 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
320 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
321 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt);
322 }
323 }
324 if (res != nullptr) {
325 if (ftype->isa_narrowoop()) {
326 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
327 res = _igvn.transform(new EncodePNode(res, ftype));
328 }
329 return res;
330 }
331 return nullptr;
332 }
333
334 //
335 // Given a Memory Phi, compute a value Phi containing the values from stores
336 // on the input paths.
337 // Note: this function is recursive, its depth is limited by the "level" argument
338 // Returns the computed Phi, or null if it cannot compute it.
339 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) {
340 assert(mem->is_Phi(), "sanity");
341 int alias_idx = C->get_alias_index(adr_t);
342 int offset = adr_t->offset();
343 int instance_id = adr_t->instance_id();
344
345 // Check if an appropriate value phi already exists.
346 Node* region = mem->in(0);
347 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
348 Node* phi = region->fast_out(k);
349 if (phi->is_Phi() && phi != mem &&
350 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) {
351 return phi;
352 }
353 }
354 // Check if an appropriate new value phi already exists.
355 Node* new_phi = value_phis->find(mem->_idx);
356 if (new_phi != nullptr)
357 return new_phi;
358
359 if (level <= 0) {
360 return nullptr; // Give up: phi tree too deep
361 }
362 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
363 Node *alloc_mem = alloc->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
364 assert(alloc_mem != nullptr, "Allocation without a memory projection.");
365
366 uint length = mem->req();
367 GrowableArray <Node *> values(length, length, nullptr);
368
369 // create a new Phi for the value
370 PhiNode *phi = new PhiNode(mem->in(0), phi_type, nullptr, mem->_idx, instance_id, alias_idx, offset);
371 transform_later(phi);
372 value_phis->push(phi, mem->_idx);
373
374 for (uint j = 1; j < length; j++) {
375 Node *in = mem->in(j);
376 if (in == nullptr || in->is_top()) {
377 values.at_put(j, in);
378 } else {
379 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
380 if (val == start_mem || val == alloc_mem) {
381 // hit a sentinel, return appropriate 0 value
382 values.at_put(j, _igvn.zerocon(ft));
383 continue;
384 }
385 if (val->is_Initialize()) {
386 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
387 }
388 if (val == nullptr) {
389 return nullptr; // can't find a value on this path
390 }
391 if (val == mem) {
392 values.at_put(j, mem);
393 } else if (val->is_Store()) {
394 Node* n = val->in(MemNode::ValueIn);
395 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
396 n = bs->step_over_gc_barrier(n);
397 if (is_subword_type(ft)) {
398 n = Compile::narrow_value(ft, n, phi_type, &_igvn, true);
399 }
400 values.at_put(j, n);
401 } else if(val->is_Proj() && val->in(0) == alloc) {
402 values.at_put(j, _igvn.zerocon(ft));
403 } else if (val->is_Phi()) {
404 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
405 if (val == nullptr) {
406 return nullptr;
407 }
408 values.at_put(j, val);
409 } else if (val->Opcode() == Op_SCMemProj) {
410 assert(val->in(0)->is_LoadStore() ||
411 val->in(0)->Opcode() == Op_EncodeISOArray ||
412 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity");
413 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
414 return nullptr;
415 } else if (val->is_ArrayCopy()) {
416 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc);
417 if (res == nullptr) {
418 return nullptr;
419 }
420 values.at_put(j, res);
421 } else if (val->is_top()) {
422 // This indicates that this path into the phi is dead. Top will eventually also propagate into the Region.
423 // IGVN will clean this up later.
424 values.at_put(j, val);
425 } else {
426 DEBUG_ONLY( val->dump(); )
427 assert(false, "unknown node on this path");
428 return nullptr; // unknown node on this path
429 }
430 }
431 }
432 // Set Phi's inputs
433 for (uint j = 1; j < length; j++) {
434 if (values.at(j) == mem) {
435 phi->init_req(j, phi);
436 } else {
437 phi->init_req(j, values.at(j));
438 }
439 }
440 return phi;
441 }
442
443 // Search the last value stored into the object's field.
444 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) {
445 assert(adr_t->is_known_instance_field(), "instance required");
446 int instance_id = adr_t->instance_id();
447 assert((uint)instance_id == alloc->_idx, "wrong allocation");
448
449 int alias_idx = C->get_alias_index(adr_t);
450 int offset = adr_t->offset();
451 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
452 Node *alloc_ctrl = alloc->in(TypeFunc::Control);
453 Node *alloc_mem = alloc->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
454 assert(alloc_mem != nullptr, "Allocation without a memory projection.");
455 VectorSet visited;
456
457 bool done = sfpt_mem == alloc_mem;
458 Node *mem = sfpt_mem;
459 while (!done) {
460 if (visited.test_set(mem->_idx)) {
461 return nullptr; // found a loop, give up
462 }
463 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
464 if (mem == start_mem || mem == alloc_mem) {
465 done = true; // hit a sentinel, return appropriate 0 value
466 } else if (mem->is_Initialize()) {
467 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
468 if (mem == nullptr) {
469 done = true; // Something go wrong.
470 } else if (mem->is_Store()) {
471 const TypePtr* atype = mem->as_Store()->adr_type();
472 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
473 done = true;
474 }
475 } else if (mem->is_Store()) {
476 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
477 assert(atype != nullptr, "address type must be oopptr");
478 assert(C->get_alias_index(atype) == alias_idx &&
479 atype->is_known_instance_field() && atype->offset() == offset &&
480 atype->instance_id() == instance_id, "store is correct memory slice");
481 done = true;
482 } else if (mem->is_Phi()) {
483 // try to find a phi's unique input
484 Node *unique_input = nullptr;
485 Node *top = C->top();
486 for (uint i = 1; i < mem->req(); i++) {
487 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
488 if (n == nullptr || n == top || n == mem) {
489 continue;
490 } else if (unique_input == nullptr) {
491 unique_input = n;
492 } else if (unique_input != n) {
493 unique_input = top;
494 break;
495 }
496 }
497 if (unique_input != nullptr && unique_input != top) {
498 mem = unique_input;
499 } else {
500 done = true;
501 }
502 } else if (mem->is_ArrayCopy()) {
503 done = true;
504 } else if (mem->is_top()) {
505 // The slice is on a dead path. Returning nullptr would lead to elimination
506 // bailout, but we want to prevent that. Just forwarding the top is also legal,
507 // and IGVN can just clean things up, and remove whatever receives top.
508 return mem;
509 } else {
510 DEBUG_ONLY( mem->dump(); )
511 assert(false, "unexpected node");
512 }
513 }
514 if (mem != nullptr) {
515 if (mem == start_mem || mem == alloc_mem) {
516 // hit a sentinel, return appropriate 0 value
517 return _igvn.zerocon(ft);
518 } else if (mem->is_Store()) {
519 Node* n = mem->in(MemNode::ValueIn);
520 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
521 n = bs->step_over_gc_barrier(n);
522 return n;
523 } else if (mem->is_Phi()) {
524 // attempt to produce a Phi reflecting the values on the input paths of the Phi
525 Node_Stack value_phis(8);
526 Node* phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
527 if (phi != nullptr) {
528 return phi;
529 } else {
530 // Kill all new Phis
531 while(value_phis.is_nonempty()) {
532 Node* n = value_phis.node();
533 _igvn.replace_node(n, C->top());
534 value_phis.pop();
535 }
536 }
537 } else if (mem->is_ArrayCopy()) {
538 Node* ctl = mem->in(0);
539 Node* m = mem->in(TypeFunc::Memory);
540 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj()) {
541 // pin the loads in the uncommon trap path
542 ctl = sfpt_ctl;
543 m = sfpt_mem;
544 }
545 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc);
546 }
547 }
548 // Something go wrong.
549 return nullptr;
550 }
551
552 // Check the possibility of scalar replacement.
553 bool PhaseMacroExpand::can_eliminate_allocation(PhaseIterGVN* igvn, AllocateNode *alloc, GrowableArray <SafePointNode *>* safepoints) {
554 // Scan the uses of the allocation to check for anything that would
555 // prevent us from eliminating it.
556 NOT_PRODUCT( const char* fail_eliminate = nullptr; )
557 DEBUG_ONLY( Node* disq_node = nullptr; )
558 bool can_eliminate = true;
559 bool reduce_merge_precheck = (safepoints == nullptr);
560
561 Node* res = alloc->result_cast();
562 const TypeOopPtr* res_type = nullptr;
563 if (res == nullptr) {
564 // All users were eliminated.
565 } else if (!res->is_CheckCastPP()) {
566 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
567 can_eliminate = false;
568 } else {
569 res_type = igvn->type(res)->isa_oopptr();
570 if (res_type == nullptr) {
571 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
572 can_eliminate = false;
573 } else if (!res_type->klass_is_exact()) {
574 NOT_PRODUCT(fail_eliminate = "Not an exact type.";)
575 can_eliminate = false;
576 } else if (res_type->isa_aryptr()) {
577 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
578 if (length < 0) {
579 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
580 can_eliminate = false;
581 }
582 }
583 }
584
585 if (can_eliminate && res != nullptr) {
586 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
587 for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
588 j < jmax && can_eliminate; j++) {
589 Node* use = res->fast_out(j);
590
591 if (use->is_AddP()) {
592 const TypePtr* addp_type = igvn->type(use)->is_ptr();
593 int offset = addp_type->offset();
594
595 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
596 NOT_PRODUCT(fail_eliminate = "Undefined field reference";)
597 can_eliminate = false;
598 break;
599 }
600 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
601 k < kmax && can_eliminate; k++) {
602 Node* n = use->fast_out(k);
603 if (n->is_Mem() && n->as_Mem()->is_mismatched_access()) {
604 DEBUG_ONLY(disq_node = n);
605 NOT_PRODUCT(fail_eliminate = "Mismatched access");
606 can_eliminate = false;
607 }
608 if (!n->is_Store() && n->Opcode() != Op_CastP2X && !bs->is_gc_pre_barrier_node(n) && !reduce_merge_precheck) {
609 DEBUG_ONLY(disq_node = n;)
610 if (n->is_Load() || n->is_LoadStore()) {
611 NOT_PRODUCT(fail_eliminate = "Field load";)
612 } else {
613 NOT_PRODUCT(fail_eliminate = "Not store field reference";)
614 }
615 can_eliminate = false;
616 }
617 }
618 } else if (use->is_ArrayCopy() &&
619 (use->as_ArrayCopy()->is_clonebasic() ||
620 use->as_ArrayCopy()->is_arraycopy_validated() ||
621 use->as_ArrayCopy()->is_copyof_validated() ||
622 use->as_ArrayCopy()->is_copyofrange_validated()) &&
623 use->in(ArrayCopyNode::Dest) == res) {
624 // ok to eliminate
625 } else if (use->is_SafePoint()) {
626 SafePointNode* sfpt = use->as_SafePoint();
627 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
628 // Object is passed as argument.
629 DEBUG_ONLY(disq_node = use;)
630 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
631 can_eliminate = false;
632 }
633 Node* sfptMem = sfpt->memory();
634 if (sfptMem == nullptr || sfptMem->is_top()) {
635 DEBUG_ONLY(disq_node = use;)
636 NOT_PRODUCT(fail_eliminate = "null or TOP memory";)
637 can_eliminate = false;
638 } else if (!reduce_merge_precheck) {
639 safepoints->append_if_missing(sfpt);
640 }
641 } else if (reduce_merge_precheck &&
642 (use->is_Phi() || use->is_EncodeP() ||
643 use->Opcode() == Op_MemBarRelease ||
644 (UseStoreStoreForCtor && use->Opcode() == Op_MemBarStoreStore))) {
645 // Nothing to do
646 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
647 if (use->is_Phi()) {
648 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
649 NOT_PRODUCT(fail_eliminate = "Object is return value";)
650 } else {
651 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
652 }
653 DEBUG_ONLY(disq_node = use;)
654 } else {
655 if (use->Opcode() == Op_Return) {
656 NOT_PRODUCT(fail_eliminate = "Object is return value";)
657 } else {
658 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
659 }
660 DEBUG_ONLY(disq_node = use;)
661 }
662 can_eliminate = false;
663 }
664 }
665 }
666
667 #ifndef PRODUCT
668 if (PrintEliminateAllocations && safepoints != nullptr) {
669 if (can_eliminate) {
670 tty->print("Scalar ");
671 if (res == nullptr)
672 alloc->dump();
673 else
674 res->dump();
675 } else if (alloc->_is_scalar_replaceable) {
676 tty->print("NotScalar (%s)", fail_eliminate);
677 if (res == nullptr)
678 alloc->dump();
679 else
680 res->dump();
681 #ifdef ASSERT
682 if (disq_node != nullptr) {
683 tty->print(" >>>> ");
684 disq_node->dump();
685 }
686 #endif /*ASSERT*/
687 }
688 }
689
690 if (TraceReduceAllocationMerges && !can_eliminate && reduce_merge_precheck) {
691 tty->print_cr("\tCan't eliminate allocation because '%s': ", fail_eliminate != nullptr ? fail_eliminate : "");
692 DEBUG_ONLY(if (disq_node != nullptr) disq_node->dump();)
693 }
694 #endif
695 return can_eliminate;
760 // CheckCastPP result was not updated in the stack slot, and so
761 // we ended up using the CastPP. That means that the field knows
762 // that it should get an oop from an interface, but the value lost
763 // that information, and so it is not a subtype.
764 // There may be other issues, feel free to investigate further!
765 if (!is_java_primitive(value_bt)) { return; }
766
767 tty->print_cr("value not compatible for field: %s vs %s",
768 type2name(value_bt),
769 type2name(field_bt));
770 tty->print("value_type: ");
771 value_type->dump();
772 tty->cr();
773 tty->print("field_type: ");
774 field_type->dump();
775 tty->cr();
776 assert(false, "value_type does not fit field_type");
777 }
778 #endif
779
780 SafePointScalarObjectNode* PhaseMacroExpand::create_scalarized_object_description(AllocateNode *alloc, SafePointNode* sfpt) {
781 // Fields of scalar objs are referenced only at the end
782 // of regular debuginfo at the last (youngest) JVMS.
783 // Record relative start index.
784 ciInstanceKlass* iklass = nullptr;
785 BasicType basic_elem_type = T_ILLEGAL;
786 const Type* field_type = nullptr;
787 const TypeOopPtr* res_type = nullptr;
788 int nfields = 0;
789 int array_base = 0;
790 int element_size = 0;
791 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
792 Node* res = alloc->result_cast();
793
794 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
795 assert(sfpt->jvms() != nullptr, "missed JVMS");
796
797 if (res != nullptr) { // Could be null when there are no users
798 res_type = _igvn.type(res)->isa_oopptr();
799
800 if (res_type->isa_instptr()) {
801 // find the fields of the class which will be needed for safepoint debug information
802 iklass = res_type->is_instptr()->instance_klass();
803 nfields = iklass->nof_nonstatic_fields();
804 } else {
805 // find the array's elements which will be needed for safepoint debug information
806 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
807 assert(nfields >= 0, "must be an array klass.");
808 basic_elem_type = res_type->is_aryptr()->elem()->array_element_basic_type();
809 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
810 element_size = type2aelembytes(basic_elem_type);
811 field_type = res_type->is_aryptr()->elem();
812 }
813 }
814
815 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type, alloc, first_ind, sfpt->jvms()->depth(), nfields);
816 sobj->init_req(0, C->root());
817 transform_later(sobj);
818
819 // Scan object's fields adding an input to the safepoint for each field.
820 for (int j = 0; j < nfields; j++) {
821 intptr_t offset;
822 ciField* field = nullptr;
823 if (iklass != nullptr) {
824 field = iklass->nonstatic_field_at(j);
825 offset = field->offset_in_bytes();
826 ciType* elem_type = field->type();
827 basic_elem_type = field->layout_type();
828
829 // The next code is taken from Parse::do_get_xxx().
830 if (is_reference_type(basic_elem_type)) {
831 if (!elem_type->is_loaded()) {
832 field_type = TypeInstPtr::BOTTOM;
833 } else if (field != nullptr && field->is_static_constant()) {
834 ciObject* con = field->constant_value().as_object();
835 // Do not "join" in the previous type; it doesn't add value,
836 // and may yield a vacuous result if the field is of interface type.
837 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
838 assert(field_type != nullptr, "field singleton type must be consistent");
839 } else {
840 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
841 }
842 if (UseCompressedOops) {
843 field_type = field_type->make_narrowoop();
844 basic_elem_type = T_NARROWOOP;
845 }
846 } else {
847 field_type = Type::get_const_basic_type(basic_elem_type);
848 }
849 } else {
850 offset = array_base + j * (intptr_t)element_size;
851 }
852
853 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr();
854
855 Node *field_val = value_from_mem(sfpt->memory(), sfpt->control(), basic_elem_type, field_type, field_addr_type, alloc);
856
857 // We weren't able to find a value for this field,
858 // give up on eliminating this allocation.
859 if (field_val == nullptr) {
860 uint last = sfpt->req() - 1;
861 for (int k = 0; k < j; k++) {
862 sfpt->del_req(last--);
863 }
864 _igvn._worklist.push(sfpt);
865
866 #ifndef PRODUCT
867 if (PrintEliminateAllocations) {
868 if (field != nullptr) {
869 tty->print("=== At SafePoint node %d can't find value of field: ", sfpt->_idx);
870 field->print();
871 int field_idx = C->get_alias_index(field_addr_type);
872 tty->print(" (alias_idx=%d)", field_idx);
873 } else { // Array's element
874 tty->print("=== At SafePoint node %d can't find value of array element [%d]", sfpt->_idx, j);
875 }
876 tty->print(", which prevents elimination of: ");
877 if (res == nullptr)
878 alloc->dump();
879 else
880 res->dump();
881 }
882 #endif
883
884 return nullptr;
885 }
886
887 if (UseCompressedOops && field_type->isa_narrowoop()) {
888 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
889 // to be able scalar replace the allocation.
890 if (field_val->is_EncodeP()) {
891 field_val = field_val->in(1);
892 } else {
893 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
894 }
895 }
896 DEBUG_ONLY(verify_type_compatability(field_val->bottom_type(), field_type);)
897 sfpt->add_req(field_val);
898 }
899
900 sfpt->jvms()->set_endoff(sfpt->req());
901
902 return sobj;
903 }
904
905 // Do scalar replacement.
906 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
907 GrowableArray <SafePointNode *> safepoints_done;
908 Node* res = alloc->result_cast();
909 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
910
911 // Process the safepoint uses
912 while (safepoints.length() > 0) {
913 SafePointNode* sfpt = safepoints.pop();
914 SafePointScalarObjectNode* sobj = create_scalarized_object_description(alloc, sfpt);
915
916 if (sobj == nullptr) {
917 undo_previous_scalarizations(safepoints_done, alloc);
918 return false;
919 }
920
921 // Now make a pass over the debug information replacing any references
922 // to the allocated object with "sobj"
923 JVMState *jvms = sfpt->jvms();
924 sfpt->replace_edges_in_range(res, sobj, jvms->debug_start(), jvms->debug_end(), &_igvn);
925 _igvn._worklist.push(sfpt);
926
927 // keep it for rollback
928 safepoints_done.append_if_missing(sfpt);
929 }
930
931 return true;
932 }
933
934 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
935 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control);
936 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory);
937 if (ctl_proj != nullptr) {
938 igvn.replace_node(ctl_proj, n->in(0));
939 }
940 if (mem_proj != nullptr) {
941 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
942 }
943 }
944
945 // Process users of eliminated allocation.
946 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) {
947 Node* res = alloc->result_cast();
948 if (res != nullptr) {
949 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
950 Node *use = res->last_out(j);
951 uint oc1 = res->outcnt();
952
953 if (use->is_AddP()) {
954 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
955 Node *n = use->last_out(k);
956 uint oc2 = use->outcnt();
957 if (n->is_Store()) {
958 #ifdef ASSERT
959 // Verify that there is no dependent MemBarVolatile nodes,
960 // they should be removed during IGVN, see MemBarNode::Ideal().
961 for (DUIterator_Fast pmax, p = n->fast_outs(pmax);
962 p < pmax; p++) {
963 Node* mb = n->fast_out(p);
964 assert(mb->is_Initialize() || !mb->is_MemBar() ||
965 mb->req() <= MemBarNode::Precedent ||
966 mb->in(MemBarNode::Precedent) != n,
967 "MemBarVolatile should be eliminated for non-escaping object");
968 }
969 #endif
970 _igvn.replace_node(n, n->in(MemNode::Memory));
971 } else {
972 eliminate_gc_barrier(n);
973 }
974 k -= (oc2 - use->outcnt());
975 }
976 _igvn.remove_dead_node(use);
977 } else if (use->is_ArrayCopy()) {
978 // Disconnect ArrayCopy node
979 ArrayCopyNode* ac = use->as_ArrayCopy();
980 if (ac->is_clonebasic()) {
981 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
982 disconnect_projections(ac, _igvn);
983 assert(alloc->in(TypeFunc::Memory)->is_Proj() && alloc->in(TypeFunc::Memory)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
984 Node* membar_before = alloc->in(TypeFunc::Memory)->in(0);
985 disconnect_projections(membar_before->as_MemBar(), _igvn);
986 if (membar_after->is_MemBar()) {
987 disconnect_projections(membar_after->as_MemBar(), _igvn);
988 }
989 } else {
990 assert(ac->is_arraycopy_validated() ||
991 ac->is_copyof_validated() ||
992 ac->is_copyofrange_validated(), "unsupported");
993 CallProjections callprojs;
994 ac->extract_projections(&callprojs, true);
995
996 _igvn.replace_node(callprojs.fallthrough_ioproj, ac->in(TypeFunc::I_O));
997 _igvn.replace_node(callprojs.fallthrough_memproj, ac->in(TypeFunc::Memory));
998 _igvn.replace_node(callprojs.fallthrough_catchproj, ac->in(TypeFunc::Control));
999
1000 // Set control to top. IGVN will remove the remaining projections
1001 ac->set_req(0, top());
1002 ac->replace_edge(res, top(), &_igvn);
1003
1004 // Disconnect src right away: it can help find new
1005 // opportunities for allocation elimination
1006 Node* src = ac->in(ArrayCopyNode::Src);
1007 ac->replace_edge(src, top(), &_igvn);
1008 // src can be top at this point if src and dest of the
1009 // arraycopy were the same
1010 if (src->outcnt() == 0 && !src->is_top()) {
1011 _igvn.remove_dead_node(src);
1012 }
1013 }
1014 _igvn._worklist.push(ac);
1015 } else {
1016 eliminate_gc_barrier(use);
1017 }
1018 j -= (oc1 - res->outcnt());
1019 }
1020 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
1021 _igvn.remove_dead_node(res);
1022 }
1023
1024 //
1025 // Process other users of allocation's projections
1026 //
1027 if (_callprojs.resproj != nullptr && _callprojs.resproj->outcnt() != 0) {
1028 // First disconnect stores captured by Initialize node.
1029 // If Initialize node is eliminated first in the following code,
1030 // it will kill such stores and DUIterator_Last will assert.
1031 for (DUIterator_Fast jmax, j = _callprojs.resproj->fast_outs(jmax); j < jmax; j++) {
1032 Node* use = _callprojs.resproj->fast_out(j);
1033 if (use->is_AddP()) {
1034 // raw memory addresses used only by the initialization
1035 _igvn.replace_node(use, C->top());
1036 --j; --jmax;
1037 }
1038 }
1039 for (DUIterator_Last jmin, j = _callprojs.resproj->last_outs(jmin); j >= jmin; ) {
1040 Node* use = _callprojs.resproj->last_out(j);
1041 uint oc1 = _callprojs.resproj->outcnt();
1042 if (use->is_Initialize()) {
1043 // Eliminate Initialize node.
1044 InitializeNode *init = use->as_Initialize();
1045 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control);
1046 if (ctrl_proj != nullptr) {
1047 _igvn.replace_node(ctrl_proj, init->in(TypeFunc::Control));
1048 #ifdef ASSERT
1049 // If the InitializeNode has no memory out, it will die, and tmp will become null
1050 Node* tmp = init->in(TypeFunc::Control);
1051 assert(tmp == nullptr || tmp == _callprojs.fallthrough_catchproj, "allocation control projection");
1052 #endif
1053 }
1054 Node* mem = init->in(TypeFunc::Memory);
1055 #ifdef ASSERT
1056 if (init->number_of_projs(TypeFunc::Memory) > 0) {
1057 if (mem->is_MergeMem()) {
1058 assert(mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw) == _callprojs.fallthrough_memproj, "allocation memory projection");
1059 } else {
1060 assert(mem == _callprojs.fallthrough_memproj, "allocation memory projection");
1061 }
1062 }
1063 #endif
1064 init->replace_mem_projs_by(mem, &_igvn);
1065 assert(init->outcnt() == 0, "should only have had a control and some memory projections, and we removed them");
1066 } else {
1067 assert(false, "only Initialize or AddP expected");
1068 }
1069 j -= (oc1 - _callprojs.resproj->outcnt());
1070 }
1071 }
1072 if (_callprojs.fallthrough_catchproj != nullptr) {
1073 _igvn.replace_node(_callprojs.fallthrough_catchproj, alloc->in(TypeFunc::Control));
1074 }
1075 if (_callprojs.fallthrough_memproj != nullptr) {
1076 _igvn.replace_node(_callprojs.fallthrough_memproj, alloc->in(TypeFunc::Memory));
1077 }
1078 if (_callprojs.catchall_memproj != nullptr) {
1079 _igvn.replace_node(_callprojs.catchall_memproj, C->top());
1080 }
1081 if (_callprojs.fallthrough_ioproj != nullptr) {
1082 _igvn.replace_node(_callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O));
1083 }
1084 if (_callprojs.catchall_ioproj != nullptr) {
1085 _igvn.replace_node(_callprojs.catchall_ioproj, C->top());
1086 }
1087 if (_callprojs.catchall_catchproj != nullptr) {
1088 _igvn.replace_node(_callprojs.catchall_catchproj, C->top());
1089 }
1090 }
1091
1092 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1093 // If reallocation fails during deoptimization we'll pop all
1094 // interpreter frames for this compiled frame and that won't play
1095 // nice with JVMTI popframe.
1096 // We avoid this issue by eager reallocation when the popframe request
1097 // is received.
1098 if (!EliminateAllocations || !alloc->_is_non_escaping) {
1099 return false;
1100 }
1101 Node* klass = alloc->in(AllocateNode::KlassNode);
1102 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1103 Node* res = alloc->result_cast();
1104 // Eliminate boxing allocations which are not used
1105 // regardless scalar replaceable status.
1106 bool boxing_alloc = C->eliminate_boxing() &&
1107 tklass->isa_instklassptr() &&
1108 tklass->is_instklassptr()->instance_klass()->is_box_klass();
1109 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != nullptr))) {
1110 return false;
1111 }
1112
1113 alloc->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1114
1115 GrowableArray <SafePointNode *> safepoints;
1116 if (!can_eliminate_allocation(&_igvn, alloc, &safepoints)) {
1117 return false;
1118 }
1119
1120 if (!alloc->_is_scalar_replaceable) {
1121 assert(res == nullptr, "sanity");
1122 // We can only eliminate allocation if all debug info references
1123 // are already replaced with SafePointScalarObject because
1124 // we can't search for a fields value without instance_id.
1125 if (safepoints.length() > 0) {
1126 return false;
1127 }
1128 }
1129
1130 if (!scalar_replacement(alloc, safepoints)) {
1131 return false;
1132 }
1133
1134 CompileLog* log = C->log();
1135 if (log != nullptr) {
1136 log->head("eliminate_allocation type='%d'",
1137 log->identify(tklass->exact_klass()));
1138 JVMState* p = alloc->jvms();
1139 while (p != nullptr) {
1140 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1141 p = p->caller();
1142 }
1143 log->tail("eliminate_allocation");
1144 }
1145
1146 process_users_of_allocation(alloc);
1147
1148 #ifndef PRODUCT
1149 if (PrintEliminateAllocations) {
1150 if (alloc->is_AllocateArray())
1151 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1152 else
1153 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1154 }
1155 #endif
1156
1157 return true;
1158 }
1159
1160 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1161 // EA should remove all uses of non-escaping boxing node.
1162 if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != nullptr) {
1163 return false;
1164 }
1165
1166 assert(boxing->result_cast() == nullptr, "unexpected boxing node result");
1167
1168 boxing->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1169
1170 const TypeTuple* r = boxing->tf()->range();
1171 assert(r->cnt() > TypeFunc::Parms, "sanity");
1172 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1173 assert(t != nullptr, "sanity");
1174
1175 CompileLog* log = C->log();
1176 if (log != nullptr) {
1177 log->head("eliminate_boxing type='%d'",
1178 log->identify(t->instance_klass()));
1179 JVMState* p = boxing->jvms();
1180 while (p != nullptr) {
1181 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1182 p = p->caller();
1183 }
1184 log->tail("eliminate_boxing");
1185 }
1186
1187 process_users_of_allocation(boxing);
1188
1189 #ifndef PRODUCT
1190 if (PrintEliminateAllocations) {
1254 // oop flavor.
1255 //
1256 //=============================================================================
1257 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1258 // Allocations bigger than this always go the slow route.
1259 // This value must be small enough that allocation attempts that need to
1260 // trigger exceptions go the slow route. Also, it must be small enough so
1261 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1262 //=============================================================================j//
1263 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1264 // The allocator will coalesce int->oop copies away. See comment in
1265 // coalesce.cpp about how this works. It depends critically on the exact
1266 // code shape produced here, so if you are changing this code shape
1267 // make sure the GC info for the heap-top is correct in and around the
1268 // slow-path call.
1269 //
1270
1271 void PhaseMacroExpand::expand_allocate_common(
1272 AllocateNode* alloc, // allocation node to be expanded
1273 Node* length, // array length for an array allocation
1274 const TypeFunc* slow_call_type, // Type of slow call
1275 address slow_call_address, // Address of slow call
1276 Node* valid_length_test // whether length is valid or not
1277 )
1278 {
1279 Node* ctrl = alloc->in(TypeFunc::Control);
1280 Node* mem = alloc->in(TypeFunc::Memory);
1281 Node* i_o = alloc->in(TypeFunc::I_O);
1282 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1283 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1284 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1285 assert(ctrl != nullptr, "must have control");
1286
1287 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1288 // they will not be used if "always_slow" is set
1289 enum { slow_result_path = 1, fast_result_path = 2 };
1290 Node *result_region = nullptr;
1291 Node *result_phi_rawmem = nullptr;
1292 Node *result_phi_rawoop = nullptr;
1293 Node *result_phi_i_o = nullptr;
1338 #endif
1339 yank_alloc_node(alloc);
1340 return;
1341 }
1342 }
1343
1344 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1345 Node *slow_region = nullptr;
1346 Node *toobig_false = ctrl;
1347
1348 // generate the initial test if necessary
1349 if (initial_slow_test != nullptr ) {
1350 assert (expand_fast_path, "Only need test if there is a fast path");
1351 slow_region = new RegionNode(3);
1352
1353 // Now make the initial failure test. Usually a too-big test but
1354 // might be a TRUE for finalizers.
1355 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1356 transform_later(toobig_iff);
1357 // Plug the failing-too-big test into the slow-path region
1358 Node *toobig_true = new IfTrueNode( toobig_iff );
1359 transform_later(toobig_true);
1360 slow_region ->init_req( too_big_or_final_path, toobig_true );
1361 toobig_false = new IfFalseNode( toobig_iff );
1362 transform_later(toobig_false);
1363 } else {
1364 // No initial test, just fall into next case
1365 assert(allocation_has_use || !expand_fast_path, "Should already have been handled");
1366 toobig_false = ctrl;
1367 DEBUG_ONLY(slow_region = NodeSentinel);
1368 }
1369
1370 // If we are here there are several possibilities
1371 // - expand_fast_path is false - then only a slow path is expanded. That's it.
1372 // no_initial_check means a constant allocation.
1373 // - If check always evaluates to false -> expand_fast_path is false (see above)
1374 // - If check always evaluates to true -> directly into fast path (but may bailout to slowpath)
1375 // if !allocation_has_use the fast path is empty
1376 // if !allocation_has_use && no_initial_check
1377 // - Then there are no fastpath that can fall out to slowpath -> no allocation code at all.
1378 // removed by yank_alloc_node above.
1379
1380 Node *slow_mem = mem; // save the current memory state for slow path
1381 // generate the fast allocation code unless we know that the initial test will always go slow
1382 if (expand_fast_path) {
1383 // Fast path modifies only raw memory.
1384 if (mem->is_MergeMem()) {
1385 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1386 }
1387
1388 // allocate the Region and Phi nodes for the result
1389 result_region = new RegionNode(3);
1390 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1391 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1392
1393 // Grab regular I/O before optional prefetch may change it.
1394 // Slow-path does no I/O so just set it to the original I/O.
1395 result_phi_i_o->init_req(slow_result_path, i_o);
1396
1397 // Name successful fast-path variables
1398 Node* fast_oop_ctrl;
1399 Node* fast_oop_rawmem;
1400 if (allocation_has_use) {
1401 Node* needgc_ctrl = nullptr;
1402 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1403
1404 intx prefetch_lines = length != nullptr ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1405 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1406 Node* fast_oop = bs->obj_allocate(this, mem, toobig_false, size_in_bytes, i_o, needgc_ctrl,
1407 fast_oop_ctrl, fast_oop_rawmem,
1408 prefetch_lines);
1409
1410 if (initial_slow_test != nullptr) {
1411 // This completes all paths into the slow merge point
1412 slow_region->init_req(need_gc_path, needgc_ctrl);
1413 transform_later(slow_region);
1414 } else {
1415 // No initial slow path needed!
1416 // Just fall from the need-GC path straight into the VM call.
1417 slow_region = needgc_ctrl;
1418 }
1419
1437 result_phi_i_o ->init_req(fast_result_path, i_o);
1438 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1439 } else {
1440 slow_region = ctrl;
1441 result_phi_i_o = i_o; // Rename it to use in the following code.
1442 }
1443
1444 // Generate slow-path call
1445 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1446 OptoRuntime::stub_name(slow_call_address),
1447 TypePtr::BOTTOM);
1448 call->init_req(TypeFunc::Control, slow_region);
1449 call->init_req(TypeFunc::I_O, top()); // does no i/o
1450 call->init_req(TypeFunc::Memory, slow_mem); // may gc ptrs
1451 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1452 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1453
1454 call->init_req(TypeFunc::Parms+0, klass_node);
1455 if (length != nullptr) {
1456 call->init_req(TypeFunc::Parms+1, length);
1457 }
1458
1459 // Copy debug information and adjust JVMState information, then replace
1460 // allocate node with the call
1461 call->copy_call_debug_info(&_igvn, alloc);
1462 // For array allocations, copy the valid length check to the call node so Compile::final_graph_reshaping() can verify
1463 // that the call has the expected number of CatchProj nodes (in case the allocation always fails and the fallthrough
1464 // path dies).
1465 if (valid_length_test != nullptr) {
1466 call->add_req(valid_length_test);
1467 }
1468 if (expand_fast_path) {
1469 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1470 } else {
1471 // Hook i_o projection to avoid its elimination during allocation
1472 // replacement (when only a slow call is generated).
1473 call->set_req(TypeFunc::I_O, result_phi_i_o);
1474 }
1475 _igvn.replace_node(alloc, call);
1476 transform_later(call);
1477
1478 // Identify the output projections from the allocate node and
1479 // adjust any references to them.
1480 // The control and io projections look like:
1481 //
1482 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1483 // Allocate Catch
1484 // ^---Proj(io) <-------+ ^---CatchProj(io)
1485 //
1486 // We are interested in the CatchProj nodes.
1487 //
1488 call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1489
1490 // An allocate node has separate memory projections for the uses on
1491 // the control and i_o paths. Replace the control memory projection with
1492 // result_phi_rawmem (unless we are only generating a slow call when
1493 // both memory projections are combined)
1494 if (expand_fast_path && _callprojs.fallthrough_memproj != nullptr) {
1495 migrate_outs(_callprojs.fallthrough_memproj, result_phi_rawmem);
1496 }
1497 // Now change uses of catchall_memproj to use fallthrough_memproj and delete
1498 // catchall_memproj so we end up with a call that has only 1 memory projection.
1499 if (_callprojs.catchall_memproj != nullptr ) {
1500 if (_callprojs.fallthrough_memproj == nullptr) {
1501 _callprojs.fallthrough_memproj = new ProjNode(call, TypeFunc::Memory);
1502 transform_later(_callprojs.fallthrough_memproj);
1503 }
1504 migrate_outs(_callprojs.catchall_memproj, _callprojs.fallthrough_memproj);
1505 _igvn.remove_dead_node(_callprojs.catchall_memproj);
1506 }
1507
1508 // An allocate node has separate i_o projections for the uses on the control
1509 // and i_o paths. Always replace the control i_o projection with result i_o
1510 // otherwise incoming i_o become dead when only a slow call is generated
1511 // (it is different from memory projections where both projections are
1512 // combined in such case).
1513 if (_callprojs.fallthrough_ioproj != nullptr) {
1514 migrate_outs(_callprojs.fallthrough_ioproj, result_phi_i_o);
1515 }
1516 // Now change uses of catchall_ioproj to use fallthrough_ioproj and delete
1517 // catchall_ioproj so we end up with a call that has only 1 i_o projection.
1518 if (_callprojs.catchall_ioproj != nullptr ) {
1519 if (_callprojs.fallthrough_ioproj == nullptr) {
1520 _callprojs.fallthrough_ioproj = new ProjNode(call, TypeFunc::I_O);
1521 transform_later(_callprojs.fallthrough_ioproj);
1522 }
1523 migrate_outs(_callprojs.catchall_ioproj, _callprojs.fallthrough_ioproj);
1524 _igvn.remove_dead_node(_callprojs.catchall_ioproj);
1525 }
1526
1527 // if we generated only a slow call, we are done
1528 if (!expand_fast_path) {
1529 // Now we can unhook i_o.
1530 if (result_phi_i_o->outcnt() > 1) {
1531 call->set_req(TypeFunc::I_O, top());
1532 } else {
1533 assert(result_phi_i_o->unique_ctrl_out() == call, "sanity");
1534 // Case of new array with negative size known during compilation.
1535 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1536 // following code since call to runtime will throw exception.
1537 // As result there will be no users of i_o after the call.
1538 // Leave i_o attached to this call to avoid problems in preceding graph.
1539 }
1540 return;
1541 }
1542
1543 if (_callprojs.fallthrough_catchproj != nullptr) {
1544 ctrl = _callprojs.fallthrough_catchproj->clone();
1545 transform_later(ctrl);
1546 _igvn.replace_node(_callprojs.fallthrough_catchproj, result_region);
1547 } else {
1548 ctrl = top();
1549 }
1550 Node *slow_result;
1551 if (_callprojs.resproj == nullptr) {
1552 // no uses of the allocation result
1553 slow_result = top();
1554 } else {
1555 slow_result = _callprojs.resproj->clone();
1556 transform_later(slow_result);
1557 _igvn.replace_node(_callprojs.resproj, result_phi_rawoop);
1558 }
1559
1560 // Plug slow-path into result merge point
1561 result_region->init_req( slow_result_path, ctrl);
1562 transform_later(result_region);
1563 if (allocation_has_use) {
1564 result_phi_rawoop->init_req(slow_result_path, slow_result);
1565 transform_later(result_phi_rawoop);
1566 }
1567 result_phi_rawmem->init_req(slow_result_path, _callprojs.fallthrough_memproj);
1568 transform_later(result_phi_rawmem);
1569 transform_later(result_phi_i_o);
1570 // This completes all paths into the result merge point
1571 }
1572
1573 // Remove alloc node that has no uses.
1574 void PhaseMacroExpand::yank_alloc_node(AllocateNode* alloc) {
1575 Node* ctrl = alloc->in(TypeFunc::Control);
1576 Node* mem = alloc->in(TypeFunc::Memory);
1577 Node* i_o = alloc->in(TypeFunc::I_O);
1578
1579 alloc->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1580 if (_callprojs.resproj != nullptr) {
1581 for (DUIterator_Fast imax, i = _callprojs.resproj->fast_outs(imax); i < imax; i++) {
1582 Node* use = _callprojs.resproj->fast_out(i);
1583 use->isa_MemBar()->remove(&_igvn);
1584 --imax;
1585 --i; // back up iterator
1586 }
1587 assert(_callprojs.resproj->outcnt() == 0, "all uses must be deleted");
1588 _igvn.remove_dead_node(_callprojs.resproj);
1589 }
1590 if (_callprojs.fallthrough_catchproj != nullptr) {
1591 migrate_outs(_callprojs.fallthrough_catchproj, ctrl);
1592 _igvn.remove_dead_node(_callprojs.fallthrough_catchproj);
1593 }
1594 if (_callprojs.catchall_catchproj != nullptr) {
1595 _igvn.rehash_node_delayed(_callprojs.catchall_catchproj);
1596 _callprojs.catchall_catchproj->set_req(0, top());
1597 }
1598 if (_callprojs.fallthrough_proj != nullptr) {
1599 Node* catchnode = _callprojs.fallthrough_proj->unique_ctrl_out();
1600 _igvn.remove_dead_node(catchnode);
1601 _igvn.remove_dead_node(_callprojs.fallthrough_proj);
1602 }
1603 if (_callprojs.fallthrough_memproj != nullptr) {
1604 migrate_outs(_callprojs.fallthrough_memproj, mem);
1605 _igvn.remove_dead_node(_callprojs.fallthrough_memproj);
1606 }
1607 if (_callprojs.fallthrough_ioproj != nullptr) {
1608 migrate_outs(_callprojs.fallthrough_ioproj, i_o);
1609 _igvn.remove_dead_node(_callprojs.fallthrough_ioproj);
1610 }
1611 if (_callprojs.catchall_memproj != nullptr) {
1612 _igvn.rehash_node_delayed(_callprojs.catchall_memproj);
1613 _callprojs.catchall_memproj->set_req(0, top());
1614 }
1615 if (_callprojs.catchall_ioproj != nullptr) {
1616 _igvn.rehash_node_delayed(_callprojs.catchall_ioproj);
1617 _callprojs.catchall_ioproj->set_req(0, top());
1618 }
1619 #ifndef PRODUCT
1620 if (PrintEliminateAllocations) {
1621 if (alloc->is_AllocateArray()) {
1622 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1623 } else {
1624 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1625 }
1626 }
1627 #endif
1628 _igvn.remove_dead_node(alloc);
1629 }
1630
1631 void PhaseMacroExpand::expand_initialize_membar(AllocateNode* alloc, InitializeNode* init,
1632 Node*& fast_oop_ctrl, Node*& fast_oop_rawmem) {
1633 // If initialization is performed by an array copy, any required
1634 // MemBarStoreStore was already added. If the object does not
1635 // escape no need for a MemBarStoreStore. If the object does not
1636 // escape in its initializer and memory barrier (MemBarStoreStore or
1637 // stronger) is already added at exit of initializer, also no need
1731 Node* thread = new ThreadLocalNode();
1732 transform_later(thread);
1733
1734 call->init_req(TypeFunc::Parms + 0, thread);
1735 call->init_req(TypeFunc::Parms + 1, oop);
1736 call->init_req(TypeFunc::Control, ctrl);
1737 call->init_req(TypeFunc::I_O , top()); // does no i/o
1738 call->init_req(TypeFunc::Memory , rawmem);
1739 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1740 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1741 transform_later(call);
1742 ctrl = new ProjNode(call, TypeFunc::Control);
1743 transform_later(ctrl);
1744 rawmem = new ProjNode(call, TypeFunc::Memory);
1745 transform_later(rawmem);
1746 }
1747 }
1748
1749 // Helper for PhaseMacroExpand::expand_allocate_common.
1750 // Initializes the newly-allocated storage.
1751 Node*
1752 PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1753 Node* control, Node* rawmem, Node* object,
1754 Node* klass_node, Node* length,
1755 Node* size_in_bytes) {
1756 InitializeNode* init = alloc->initialization();
1757 // Store the klass & mark bits
1758 Node* mark_node = alloc->make_ideal_mark(&_igvn, control, rawmem);
1759 if (!mark_node->is_Con()) {
1760 transform_later(mark_node);
1761 }
1762 rawmem = make_store_raw(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, TypeX_X->basic_type());
1763
1764 if (!UseCompactObjectHeaders) {
1765 rawmem = make_store_raw(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1766 }
1767 int header_size = alloc->minimum_header_size(); // conservatively small
1768
1769 // Array length
1770 if (length != nullptr) { // Arrays need length field
1771 rawmem = make_store_raw(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1772 // conservatively small header size:
1773 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1774 if (_igvn.type(klass_node)->isa_aryklassptr()) { // we know the exact header size in most cases:
1775 BasicType elem = _igvn.type(klass_node)->is_klassptr()->as_instance_type()->isa_aryptr()->elem()->array_element_basic_type();
1776 if (is_reference_type(elem, true)) {
1777 elem = T_OBJECT;
1778 }
1779 header_size = Klass::layout_helper_header_size(Klass::array_layout_helper(elem));
1780 }
1781 }
1782
1783 // Clear the object body, if necessary.
1784 if (init == nullptr) {
1785 // The init has somehow disappeared; be cautious and clear everything.
1786 //
1787 // This can happen if a node is allocated but an uncommon trap occurs
1788 // immediately. In this case, the Initialize gets associated with the
1789 // trap, and may be placed in a different (outer) loop, if the Allocate
1790 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1791 // there can be two Allocates to one Initialize. The answer in all these
1792 // edge cases is safety first. It is always safe to clear immediately
1793 // within an Allocate, and then (maybe or maybe not) clear some more later.
1794 if (!(UseTLAB && ZeroTLAB)) {
1795 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1796 header_size, size_in_bytes,
1797 true,
1798 &_igvn);
1799 }
1800 } else {
1801 if (!init->is_complete()) {
1802 // Try to win by zeroing only what the init does not store.
1803 // We can also try to do some peephole optimizations,
1804 // such as combining some adjacent subword stores.
1805 rawmem = init->complete_stores(control, rawmem, object,
1806 header_size, size_in_bytes, &_igvn);
1807 }
1808 // We have no more use for this link, since the AllocateNode goes away:
1809 init->set_req(InitializeNode::RawAddress, top());
1810 // (If we keep the link, it just confuses the register allocator,
1811 // who thinks he sees a real use of the address by the membar.)
1812 }
1813
1814 return rawmem;
1815 }
1950 for (intx i = 0; i < lines; i++) {
1951 prefetch_adr = AddPNode::make_off_heap(new_eden_top,
1952 _igvn.MakeConX(distance));
1953 transform_later(prefetch_adr);
1954 prefetch = new PrefetchAllocationNode(i_o, prefetch_adr);
1955 // Do not let it float too high, since if eden_top == eden_end,
1956 // both might be null.
1957 if (i == 0) { // Set control for first prefetch, next follows it
1958 prefetch->init_req(0, needgc_false);
1959 }
1960 transform_later(prefetch);
1961 distance += step_size;
1962 i_o = prefetch;
1963 }
1964 }
1965 return i_o;
1966 }
1967
1968
1969 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
1970 expand_allocate_common(alloc, nullptr,
1971 OptoRuntime::new_instance_Type(),
1972 OptoRuntime::new_instance_Java(), nullptr);
1973 }
1974
1975 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
1976 Node* length = alloc->in(AllocateNode::ALength);
1977 Node* valid_length_test = alloc->in(AllocateNode::ValidLengthTest);
1978 InitializeNode* init = alloc->initialization();
1979 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1980 const TypeAryKlassPtr* ary_klass_t = _igvn.type(klass_node)->isa_aryklassptr();
1981 address slow_call_address; // Address of slow call
1982 if (init != nullptr && init->is_complete_with_arraycopy() &&
1983 ary_klass_t && ary_klass_t->elem()->isa_klassptr() == nullptr) {
1984 // Don't zero type array during slow allocation in VM since
1985 // it will be initialized later by arraycopy in compiled code.
1986 slow_call_address = OptoRuntime::new_array_nozero_Java();
1987 } else {
1988 slow_call_address = OptoRuntime::new_array_Java();
1989 }
1990 expand_allocate_common(alloc, length,
1991 OptoRuntime::new_array_Type(),
1992 slow_call_address, valid_length_test);
1993 }
1994
1995 //-------------------mark_eliminated_box----------------------------------
1996 //
1997 // During EA obj may point to several objects but after few ideal graph
1998 // transformations (CCP) it may point to only one non escaping object
1999 // (but still using phi), corresponding locks and unlocks will be marked
2000 // for elimination. Later obj could be replaced with a new node (new phi)
2001 // and which does not have escape information. And later after some graph
2002 // reshape other locks and unlocks (which were not marked for elimination
2003 // before) are connected to this new obj (phi) but they still will not be
2004 // marked for elimination since new obj has no escape information.
2005 // Mark all associated (same box and obj) lock and unlock nodes for
2006 // elimination if some of them marked already.
2007 void PhaseMacroExpand::mark_eliminated_box(Node* box, Node* obj) {
2008 BoxLockNode* oldbox = box->as_BoxLock();
2009 if (oldbox->is_eliminated()) {
2010 return; // This BoxLock node was processed already.
2011 }
2183 #ifdef ASSERT
2184 if (!alock->is_coarsened()) {
2185 // Check that new "eliminated" BoxLock node is created.
2186 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2187 assert(oldbox->is_eliminated(), "should be done already");
2188 }
2189 #endif
2190
2191 alock->log_lock_optimization(C, "eliminate_lock");
2192
2193 #ifndef PRODUCT
2194 if (PrintEliminateLocks) {
2195 tty->print_cr("++++ Eliminated: %d %s '%s'", alock->_idx, (alock->is_Lock() ? "Lock" : "Unlock"), alock->kind_as_string());
2196 }
2197 #endif
2198
2199 Node* mem = alock->in(TypeFunc::Memory);
2200 Node* ctrl = alock->in(TypeFunc::Control);
2201 guarantee(ctrl != nullptr, "missing control projection, cannot replace_node() with null");
2202
2203 alock->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
2204 // There are 2 projections from the lock. The lock node will
2205 // be deleted when its last use is subsumed below.
2206 assert(alock->outcnt() == 2 &&
2207 _callprojs.fallthrough_proj != nullptr &&
2208 _callprojs.fallthrough_memproj != nullptr,
2209 "Unexpected projections from Lock/Unlock");
2210
2211 Node* fallthroughproj = _callprojs.fallthrough_proj;
2212 Node* memproj_fallthrough = _callprojs.fallthrough_memproj;
2213
2214 // The memory projection from a lock/unlock is RawMem
2215 // The input to a Lock is merged memory, so extract its RawMem input
2216 // (unless the MergeMem has been optimized away.)
2217 if (alock->is_Lock()) {
2218 // Search for MemBarAcquireLock node and delete it also.
2219 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2220 assert(membar != nullptr && membar->Opcode() == Op_MemBarAcquireLock, "");
2221 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2222 Node* memproj = membar->proj_out(TypeFunc::Memory);
2223 _igvn.replace_node(ctrlproj, fallthroughproj);
2224 _igvn.replace_node(memproj, memproj_fallthrough);
2225
2226 // Delete FastLock node also if this Lock node is unique user
2227 // (a loop peeling may clone a Lock node).
2228 Node* flock = alock->as_Lock()->fastlock_node();
2229 if (flock->outcnt() == 1) {
2230 assert(flock->unique_out() == alock, "sanity");
2231 _igvn.replace_node(flock, top());
2232 }
2263 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2264
2265 // Make the merge point
2266 Node *region;
2267 Node *mem_phi;
2268 Node *slow_path;
2269
2270 region = new RegionNode(3);
2271 // create a Phi for the memory state
2272 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2273
2274 // Optimize test; set region slot 2
2275 slow_path = opt_bits_test(ctrl, region, 2, flock);
2276 mem_phi->init_req(2, mem);
2277
2278 // Make slow path call
2279 CallNode* call = make_slow_call(lock, OptoRuntime::complete_monitor_enter_Type(),
2280 OptoRuntime::complete_monitor_locking_Java(), nullptr, slow_path,
2281 obj, box, nullptr);
2282
2283 call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
2284
2285 // Slow path can only throw asynchronous exceptions, which are always
2286 // de-opted. So the compiler thinks the slow-call can never throw an
2287 // exception. If it DOES throw an exception we would need the debug
2288 // info removed first (since if it throws there is no monitor).
2289 assert(_callprojs.fallthrough_ioproj == nullptr && _callprojs.catchall_ioproj == nullptr &&
2290 _callprojs.catchall_memproj == nullptr && _callprojs.catchall_catchproj == nullptr, "Unexpected projection from Lock");
2291
2292 // Capture slow path
2293 // disconnect fall-through projection from call and create a new one
2294 // hook up users of fall-through projection to region
2295 Node *slow_ctrl = _callprojs.fallthrough_proj->clone();
2296 transform_later(slow_ctrl);
2297 _igvn.hash_delete(_callprojs.fallthrough_proj);
2298 _callprojs.fallthrough_proj->disconnect_inputs(C);
2299 region->init_req(1, slow_ctrl);
2300 // region inputs are now complete
2301 transform_later(region);
2302 _igvn.replace_node(_callprojs.fallthrough_proj, region);
2303
2304 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2305
2306 mem_phi->init_req(1, memproj);
2307
2308 transform_later(mem_phi);
2309
2310 _igvn.replace_node(_callprojs.fallthrough_memproj, mem_phi);
2311 }
2312
2313 //------------------------------expand_unlock_node----------------------
2314 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2315
2316 Node* ctrl = unlock->in(TypeFunc::Control);
2317 Node* mem = unlock->in(TypeFunc::Memory);
2318 Node* obj = unlock->obj_node();
2319 Node* box = unlock->box_node();
2320
2321 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2322
2323 // No need for a null check on unlock
2324
2325 // Make the merge point
2326 Node* region = new RegionNode(3);
2327
2328 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2329 funlock = transform_later( funlock )->as_FastUnlock();
2330 // Optimize test; set region slot 2
2331 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock);
2332 Node *thread = transform_later(new ThreadLocalNode());
2333
2334 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2335 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2336 "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2337
2338 call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
2339 assert(_callprojs.fallthrough_ioproj == nullptr && _callprojs.catchall_ioproj == nullptr &&
2340 _callprojs.catchall_memproj == nullptr && _callprojs.catchall_catchproj == nullptr, "Unexpected projection from Lock");
2341
2342 // No exceptions for unlocking
2343 // Capture slow path
2344 // disconnect fall-through projection from call and create a new one
2345 // hook up users of fall-through projection to region
2346 Node *slow_ctrl = _callprojs.fallthrough_proj->clone();
2347 transform_later(slow_ctrl);
2348 _igvn.hash_delete(_callprojs.fallthrough_proj);
2349 _callprojs.fallthrough_proj->disconnect_inputs(C);
2350 region->init_req(1, slow_ctrl);
2351 // region inputs are now complete
2352 transform_later(region);
2353 _igvn.replace_node(_callprojs.fallthrough_proj, region);
2354
2355 if (_callprojs.fallthrough_memproj != nullptr) {
2356 // create a Phi for the memory state
2357 Node* mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2358 Node* memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2359 mem_phi->init_req(1, memproj);
2360 mem_phi->init_req(2, mem);
2361 transform_later(mem_phi);
2362 _igvn.replace_node(_callprojs.fallthrough_memproj, mem_phi);
2363 }
2364 }
2365
2366 void PhaseMacroExpand::expand_subtypecheck_node(SubTypeCheckNode *check) {
2367 assert(check->in(SubTypeCheckNode::Control) == nullptr, "should be pinned");
2368 Node* bol = check->unique_out();
2369 Node* obj_or_subklass = check->in(SubTypeCheckNode::ObjOrSubKlass);
2370 Node* superklass = check->in(SubTypeCheckNode::SuperKlass);
2371 assert(bol->is_Bool() && bol->as_Bool()->_test._test == BoolTest::ne, "unexpected bool node");
2372
2373 for (DUIterator_Last imin, i = bol->last_outs(imin); i >= imin; --i) {
2374 Node* iff = bol->last_out(i);
2375 assert(iff->is_If(), "where's the if?");
2376
2377 if (iff->in(0)->is_top()) {
2378 _igvn.replace_input_of(iff, 1, C->top());
2379 continue;
2380 }
2381
2382 IfTrueNode* iftrue = iff->as_If()->true_proj();
2383 IfFalseNode* iffalse = iff->as_If()->false_proj();
2384 Node* ctrl = iff->in(0);
2385
2386 Node* subklass = nullptr;
2387 if (_igvn.type(obj_or_subklass)->isa_klassptr()) {
2388 subklass = obj_or_subklass;
2389 } else {
2390 Node* k_adr = basic_plus_adr(obj_or_subklass, oopDesc::klass_offset_in_bytes());
2391 subklass = _igvn.transform(LoadKlassNode::make(_igvn, C->immutable_memory(), k_adr, TypeInstPtr::KLASS));
2392 }
2393
2394 Node* not_subtype_ctrl = Phase::gen_subtype_check(subklass, superklass, &ctrl, nullptr, _igvn, check->method(), check->bci());
2395
2396 _igvn.replace_input_of(iff, 0, C->top());
2397 _igvn.replace_node(iftrue, not_subtype_ctrl);
2398 _igvn.replace_node(iffalse, ctrl);
2399 }
2400 _igvn.replace_node(check, C->top());
2401 }
2402
2403 // Perform refining of strip mined loop nodes in the macro nodes list.
2404 void PhaseMacroExpand::refine_strip_mined_loop_macro_nodes() {
2405 for (int i = C->macro_count(); i > 0; i--) {
2406 Node* n = C->macro_node(i - 1);
2407 if (n->is_OuterStripMinedLoop()) {
2408 n->as_OuterStripMinedLoop()->adjust_strip_mined_loop(&_igvn);
2409 }
2410 }
2411 }
2412
2413 //---------------------------eliminate_macro_nodes----------------------
2414 // Eliminate scalar replaced allocations and associated locks.
2415 void PhaseMacroExpand::eliminate_macro_nodes() {
2416 if (C->macro_count() == 0)
2417 return;
2418
2419 if (StressMacroElimination) {
2420 C->shuffle_macro_nodes();
2421 }
2422 NOT_PRODUCT(int membar_before = count_MemBar(C);)
2423
2424 // Before elimination may re-mark (change to Nested or NonEscObj)
2425 // all associated (same box and obj) lock and unlock nodes.
2426 int cnt = C->macro_count();
2427 for (int i=0; i < cnt; i++) {
2428 Node *n = C->macro_node(i);
2429 if (n->is_AbstractLock()) { // Lock and Unlock nodes
2430 mark_eliminated_locking_nodes(n->as_AbstractLock());
2431 }
2432 }
2433 // Re-marking may break consistency of Coarsened locks.
2434 if (!C->coarsened_locks_consistent()) {
2435 return; // recompile without Coarsened locks if broken
2436 } else {
2437 // After coarsened locks are eliminated locking regions
2438 // become unbalanced. We should not execute any more
2439 // locks elimination optimizations on them.
2440 C->mark_unbalanced_boxes();
2441 }
2442
2443 // First, attempt to eliminate locks
2444 bool progress = true;
2445 while (progress) {
2446 progress = false;
2447 for (int i = C->macro_count(); i > 0; i = MIN2(i - 1, C->macro_count())) { // more than 1 element can be eliminated at once
2448 Node* n = C->macro_node(i - 1);
2449 bool success = false;
2450 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2451 if (n->is_AbstractLock()) {
2452 success = eliminate_locking_node(n->as_AbstractLock());
2453 #ifndef PRODUCT
2454 if (success && PrintOptoStatistics) {
2455 AtomicAccess::inc(&PhaseMacroExpand::_monitor_objects_removed_counter);
2456 }
2457 #endif
2458 }
2459 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2460 progress = progress || success;
2461 if (success) {
2462 C->print_method(PHASE_AFTER_MACRO_ELIMINATION_STEP, 5, n);
2463 }
2464 }
2465 }
2466 // Next, attempt to eliminate allocations
2467 progress = true;
2468 while (progress) {
2469 progress = false;
2470 for (int i = C->macro_count(); i > 0; i = MIN2(i - 1, C->macro_count())) { // more than 1 element can be eliminated at once
2471 Node* n = C->macro_node(i - 1);
2472 bool success = false;
2473 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2474 switch (n->class_id()) {
2475 case Node::Class_Allocate:
2476 case Node::Class_AllocateArray:
2477 success = eliminate_allocate_node(n->as_Allocate());
2478 #ifndef PRODUCT
2479 if (success && PrintOptoStatistics) {
2480 AtomicAccess::inc(&PhaseMacroExpand::_objs_scalar_replaced_counter);
2481 }
2482 #endif
2483 break;
2484 case Node::Class_CallStaticJava:
2485 success = eliminate_boxing_node(n->as_CallStaticJava());
2486 break;
2487 case Node::Class_Lock:
2488 case Node::Class_Unlock:
2489 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2490 break;
2491 case Node::Class_ArrayCopy:
2492 break;
2493 case Node::Class_OuterStripMinedLoop:
2494 break;
2495 case Node::Class_SubTypeCheck:
2496 break;
2497 case Node::Class_Opaque1:
2498 break;
2499 default:
2500 assert(n->Opcode() == Op_LoopLimit ||
2501 n->Opcode() == Op_ModD ||
2502 n->Opcode() == Op_ModF ||
2503 n->Opcode() == Op_PowD ||
2504 n->is_OpaqueConstantBool() ||
2505 n->is_OpaqueInitializedAssertionPredicate() ||
2506 n->Opcode() == Op_MaxL ||
2507 n->Opcode() == Op_MinL ||
2508 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n),
2509 "unknown node type in macro list");
2510 }
2511 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2512 progress = progress || success;
2513 if (success) {
2514 C->print_method(PHASE_AFTER_MACRO_ELIMINATION_STEP, 5, n);
2515 }
2516 }
2517 }
2518 #ifndef PRODUCT
2519 if (PrintOptoStatistics) {
2520 int membar_after = count_MemBar(C);
2521 AtomicAccess::add(&PhaseMacroExpand::_memory_barriers_removed_counter, membar_before - membar_after);
2522 }
2523 #endif
2524 }
2525
2526 void PhaseMacroExpand::eliminate_opaque_looplimit_macro_nodes() {
2527 if (C->macro_count() == 0) {
2528 return;
2529 }
2530 refine_strip_mined_loop_macro_nodes();
2531 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2532 bool progress = true;
2533 while (progress) {
2534 progress = false;
2535 for (int i = C->macro_count(); i > 0; i--) {
2536 Node* n = C->macro_node(i-1);
2537 bool success = false;
2538 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2539 if (n->Opcode() == Op_LoopLimit) {
2540 // Remove it from macro list and put on IGVN worklist to optimize.
2541 C->remove_macro_node(n);
2542 _igvn._worklist.push(n);
2543 success = true;
2544 } else if (n->Opcode() == Op_CallStaticJava) {
2545 // Remove it from macro list and put on IGVN worklist to optimize.
2546 C->remove_macro_node(n);
2547 _igvn._worklist.push(n);
2548 success = true;
2549 } else if (n->is_Opaque1()) {
2550 _igvn.replace_node(n, n->in(1));
2551 success = true;
2552 } else if (n->is_OpaqueConstantBool()) {
2553 // Tests with OpaqueConstantBool nodes are implicitly known. Replace the node with true/false. In debug builds,
2554 // we leave the test in the graph to have an additional sanity check at runtime. If the test fails (i.e. a bug),
2555 // we will execute a Halt node.
2556 #ifdef ASSERT
2557 _igvn.replace_node(n, n->in(1));
2558 #else
2559 _igvn.replace_node(n, _igvn.intcon(n->as_OpaqueConstantBool()->constant()));
2560 #endif
2561 success = true;
2562 } else if (n->is_OpaqueInitializedAssertionPredicate()) {
2563 // Initialized Assertion Predicates must always evaluate to true. Therefore, we get rid of them in product
2564 // builds as they are useless. In debug builds we keep them as additional verification code. Even though
2565 // loop opts are already over, we want to keep Initialized Assertion Predicates alive as long as possible to
2566 // enable folding of dead control paths within which cast nodes become top after due to impossible types -
2567 // even after loop opts are over. Therefore, we delay the removal of these opaque nodes until now.
2568 #ifdef ASSERT
2637 // Worst case is a macro node gets expanded into about 200 nodes.
2638 // Allow 50% more for optimization.
2639 if (C->check_node_count(300, "out of nodes before macro expansion")) {
2640 return true;
2641 }
2642
2643 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2644 switch (n->class_id()) {
2645 case Node::Class_Lock:
2646 expand_lock_node(n->as_Lock());
2647 break;
2648 case Node::Class_Unlock:
2649 expand_unlock_node(n->as_Unlock());
2650 break;
2651 case Node::Class_ArrayCopy:
2652 expand_arraycopy_node(n->as_ArrayCopy());
2653 break;
2654 case Node::Class_SubTypeCheck:
2655 expand_subtypecheck_node(n->as_SubTypeCheck());
2656 break;
2657 default:
2658 switch (n->Opcode()) {
2659 case Op_ModD:
2660 case Op_ModF:
2661 case Op_PowD: {
2662 CallLeafPureNode* call_macro = n->as_CallLeafPure();
2663 CallLeafPureNode* call = call_macro->inline_call_leaf_pure_node();
2664 _igvn.replace_node(call_macro, call);
2665 transform_later(call);
2666 break;
2667 }
2668 default:
2669 assert(false, "unknown node type in macro list");
2670 }
2671 }
2672 assert(C->macro_count() == (old_macro_count - 1), "expansion must have deleted one node from macro list");
2673 if (C->failing()) return true;
2674 C->print_method(PHASE_AFTER_MACRO_EXPANSION_STEP, 5, n);
2675
2676 // Clean up the graph so we're less likely to hit the maximum node
|
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 "ci/ciFlatArrayKlass.hpp"
26 #include "ci/ciInlineKlass.hpp"
27 #include "ci/ciInstanceKlass.hpp"
28 #include "compiler/compileLog.hpp"
29 #include "gc/shared/collectedHeap.inline.hpp"
30 #include "gc/shared/tlab_globals.hpp"
31 #include "libadt/vectset.hpp"
32 #include "memory/universe.hpp"
33 #include "opto/addnode.hpp"
34 #include "opto/arraycopynode.hpp"
35 #include "opto/callnode.hpp"
36 #include "opto/castnode.hpp"
37 #include "opto/cfgnode.hpp"
38 #include "opto/compile.hpp"
39 #include "opto/convertnode.hpp"
40 #include "opto/graphKit.hpp"
41 #include "opto/inlinetypenode.hpp"
42 #include "opto/intrinsicnode.hpp"
43 #include "opto/locknode.hpp"
44 #include "opto/loopnode.hpp"
45 #include "opto/macro.hpp"
46 #include "opto/memnode.hpp"
47 #include "opto/narrowptrnode.hpp"
48 #include "opto/node.hpp"
49 #include "opto/opaquenode.hpp"
50 #include "opto/opcodes.hpp"
51 #include "opto/phaseX.hpp"
52 #include "opto/rootnode.hpp"
53 #include "opto/runtime.hpp"
54 #include "opto/subnode.hpp"
55 #include "opto/subtypenode.hpp"
56 #include "opto/type.hpp"
57 #include "prims/jvmtiExport.hpp"
58 #include "runtime/continuation.hpp"
59 #include "runtime/sharedRuntime.hpp"
60 #include "runtime/stubRoutines.hpp"
61 #include "utilities/globalDefinitions.hpp"
62 #include "utilities/macros.hpp"
63 #include "utilities/powerOfTwo.hpp"
64 #if INCLUDE_G1GC
65 #include "gc/g1/g1ThreadLocalData.hpp"
66 #endif // INCLUDE_G1GC
67
68
69 //
70 // Replace any references to "oldref" in inputs to "use" with "newref".
71 // Returns the number of replacements made.
72 //
73 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
74 int nreplacements = 0;
75 uint req = use->req();
76 for (uint j = 0; j < use->len(); j++) {
77 Node *uin = use->in(j);
78 if (uin == oldref) {
79 if (j < req)
80 use->set_req(j, newref);
81 else
82 use->set_prec(j, newref);
83 nreplacements++;
84 } else if (j >= req && uin == nullptr) {
85 break;
86 }
87 }
88 return nreplacements;
89 }
90
91
92 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word) {
93 Node* cmp = word;
94 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne));
95 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
96 transform_later(iff);
97
98 // Fast path taken.
99 Node *fast_taken = transform_later(new IfFalseNode(iff));
100
101 // Fast path not-taken, i.e. slow path
102 Node *slow_taken = transform_later(new IfTrueNode(iff));
103
104 region->init_req(edge, fast_taken); // Capture fast-control
105 return slow_taken;
106 }
107
108 //--------------------copy_predefined_input_for_runtime_call--------------------
109 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
110 // Set fixed predefined input arguments
123 // Slow-path call
124 CallNode *call = leaf_name
125 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
126 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), TypeRawPtr::BOTTOM );
127
128 // Slow path call has no side-effects, uses few values
129 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
130 if (parm0 != nullptr) call->init_req(TypeFunc::Parms+0, parm0);
131 if (parm1 != nullptr) call->init_req(TypeFunc::Parms+1, parm1);
132 if (parm2 != nullptr) call->init_req(TypeFunc::Parms+2, parm2);
133 call->copy_call_debug_info(&_igvn, oldcall);
134 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
135 _igvn.replace_node(oldcall, call);
136 transform_later(call);
137
138 return call;
139 }
140
141 void PhaseMacroExpand::eliminate_gc_barrier(Node* p2x) {
142 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
143 bs->eliminate_gc_barrier(&_igvn, p2x);
144 #ifndef PRODUCT
145 if (PrintOptoStatistics) {
146 AtomicAccess::inc(&PhaseMacroExpand::_GC_barriers_removed_counter);
147 }
148 #endif
149 }
150
151 // Search for a memory operation for the specified memory slice.
152 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
153 Node *orig_mem = mem;
154 Node *alloc_mem = alloc->as_Allocate()->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
155 assert(alloc_mem != nullptr, "Allocation without a memory projection.");
156 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
157 while (true) {
158 if (mem == alloc_mem || mem == start_mem ) {
159 return mem; // hit one of our sentinels
160 } else if (mem->is_MergeMem()) {
161 mem = mem->as_MergeMem()->memory_at(alias_idx);
162 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
163 Node *in = mem->in(0);
166 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
167 return in;
168 } else if (in->is_Call()) {
169 CallNode *call = in->as_Call();
170 if (call->may_modify(tinst, phase)) {
171 assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape");
172 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) {
173 return in;
174 }
175 }
176 mem = in->in(TypeFunc::Memory);
177 } else if (in->is_MemBar()) {
178 ArrayCopyNode* ac = nullptr;
179 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) {
180 if (ac != nullptr) {
181 assert(ac->is_clonebasic(), "Only basic clone is a non escaping clone");
182 return ac;
183 }
184 }
185 mem = in->in(TypeFunc::Memory);
186 } else if (in->is_LoadFlat() || in->is_StoreFlat()) {
187 mem = in->in(TypeFunc::Memory);
188 } else {
189 #ifdef ASSERT
190 in->dump();
191 mem->dump();
192 assert(false, "unexpected projection");
193 #endif
194 }
195 } else if (mem->is_Store()) {
196 const TypePtr* atype = mem->as_Store()->adr_type();
197 int adr_idx = phase->C->get_alias_index(atype);
198 if (adr_idx == alias_idx) {
199 assert(atype->isa_oopptr(), "address type must be oopptr");
200 int adr_offset = atype->flat_offset();
201 uint adr_iid = atype->is_oopptr()->instance_id();
202 // Array elements references have the same alias_idx
203 // but different offset and different instance_id.
204 if (adr_offset == offset && adr_iid == alloc->_idx) {
205 return mem;
206 }
207 } else {
208 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
209 }
210 mem = mem->in(MemNode::Memory);
211 } else if (mem->is_ClearArray()) {
212 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
213 // Can not bypass initialization of the instance
214 // we are looking.
215 DEBUG_ONLY(intptr_t offset;)
216 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
217 InitializeNode* init = alloc->as_Allocate()->initialization();
218 // We are looking for stored value, return Initialize node
219 // or memory edge from Allocate node.
220 if (init != nullptr) {
225 }
226 // Otherwise skip it (the call updated 'mem' value).
227 } else if (mem->Opcode() == Op_SCMemProj) {
228 mem = mem->in(0);
229 Node* adr = nullptr;
230 if (mem->is_LoadStore()) {
231 adr = mem->in(MemNode::Address);
232 } else {
233 assert(mem->Opcode() == Op_EncodeISOArray ||
234 mem->Opcode() == Op_StrCompressedCopy, "sanity");
235 adr = mem->in(3); // Destination array
236 }
237 const TypePtr* atype = adr->bottom_type()->is_ptr();
238 int adr_idx = phase->C->get_alias_index(atype);
239 if (adr_idx == alias_idx) {
240 DEBUG_ONLY(mem->dump();)
241 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
242 return nullptr;
243 }
244 mem = mem->in(MemNode::Memory);
245 } else if (mem->Opcode() == Op_StrInflatedCopy) {
246 Node* adr = mem->in(3); // Destination array
247 const TypePtr* atype = adr->bottom_type()->is_ptr();
248 int adr_idx = phase->C->get_alias_index(atype);
249 if (adr_idx == alias_idx) {
250 DEBUG_ONLY(mem->dump();)
251 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
252 return nullptr;
253 }
254 mem = mem->in(MemNode::Memory);
255 } else {
256 return mem;
257 }
258 assert(mem != orig_mem, "dead memory loop");
259 }
260 }
261
262 // Generate loads from source of the arraycopy for fields of
263 // destination needed at a deoptimization point
264 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) {
265 BasicType bt = ft;
270 }
271 Node* res = nullptr;
272 if (ac->is_clonebasic()) {
273 assert(ac->in(ArrayCopyNode::Src) != ac->in(ArrayCopyNode::Dest), "clone source equals destination");
274 Node* base = ac->in(ArrayCopyNode::Src);
275 Node* adr = _igvn.transform(AddPNode::make_with_base(base, _igvn.MakeConX(offset)));
276 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
277 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
278 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
279 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt);
280 } else {
281 if (ac->modifies(offset, offset, &_igvn, true)) {
282 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
283 uint shift = exact_log2(type2aelembytes(bt));
284 Node* src_pos = ac->in(ArrayCopyNode::SrcPos);
285 Node* dest_pos = ac->in(ArrayCopyNode::DestPos);
286 const TypeInt* src_pos_t = _igvn.type(src_pos)->is_int();
287 const TypeInt* dest_pos_t = _igvn.type(dest_pos)->is_int();
288
289 Node* adr = nullptr;
290 Node* base = ac->in(ArrayCopyNode::Src);
291 const TypeAryPtr* adr_type = _igvn.type(base)->is_aryptr();
292 if (adr_type->is_flat()) {
293 shift = adr_type->flat_log_elem_size();
294 }
295 if (src_pos_t->is_con() && dest_pos_t->is_con()) {
296 intptr_t off = ((src_pos_t->get_con() - dest_pos_t->get_con()) << shift) + offset;
297 adr = _igvn.transform(AddPNode::make_with_base(base, base, _igvn.MakeConX(off)));
298 adr_type = _igvn.type(adr)->is_aryptr();
299 assert(adr_type == _igvn.type(base)->is_aryptr()->add_field_offset_and_offset(off), "incorrect address type");
300 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
301 // Don't emit a new load from src if src == dst but try to get the value from memory instead
302 return value_from_mem(ac->in(TypeFunc::Memory), ctl, ft, ftype, adr_type, alloc);
303 }
304 } else {
305 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
306 // Non constant offset in the array: we can't statically
307 // determine the value
308 return nullptr;
309 }
310 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
311 #ifdef _LP64
312 diff = _igvn.transform(new ConvI2LNode(diff));
313 #endif
314 diff = _igvn.transform(new LShiftXNode(diff, _igvn.intcon(shift)));
315
316 Node* off = _igvn.transform(new AddXNode(_igvn.MakeConX(offset), diff));
317 adr = _igvn.transform(AddPNode::make_with_base(base, base, off));
318 // In the case of a flat inline type array, each field has its
319 // own slice so we need to extract the field being accessed from
320 // the address computation
321 adr_type = adr_type->add_field_offset_and_offset(offset)->add_offset(Type::OffsetBot)->is_aryptr();
322 adr = _igvn.transform(new CastPPNode(ctl, adr, adr_type));
323 }
324 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
325 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
326 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt);
327 }
328 }
329 if (res != nullptr) {
330 if (ftype->isa_narrowoop()) {
331 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
332 assert(res->isa_DecodeN(), "should be narrow oop");
333 res = _igvn.transform(new EncodePNode(res, ftype));
334 }
335 return res;
336 }
337 return nullptr;
338 }
339
340 //
341 // Given a Memory Phi, compute a value Phi containing the values from stores
342 // on the input paths.
343 // Note: this function is recursive, its depth is limited by the "level" argument
344 // Returns the computed Phi, or null if it cannot compute it.
345 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) {
346 assert(mem->is_Phi(), "sanity");
347 int alias_idx = C->get_alias_index(adr_t);
348 int offset = adr_t->flat_offset();
349 int instance_id = adr_t->instance_id();
350
351 // Check if an appropriate value phi already exists.
352 Node* region = mem->in(0);
353 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
354 Node* phi = region->fast_out(k);
355 if (phi->is_Phi() && phi != mem &&
356 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) {
357 return phi;
358 }
359 }
360 // Check if an appropriate new value phi already exists.
361 Node* new_phi = value_phis->find(mem->_idx);
362 if (new_phi != nullptr)
363 return new_phi;
364
365 if (level <= 0) {
366 return nullptr; // Give up: phi tree too deep
367 }
368 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
369 Node *alloc_mem = alloc->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
370 assert(alloc_mem != nullptr, "Allocation without a memory projection.");
371
372 uint length = mem->req();
373 GrowableArray <Node *> values(length, length, nullptr);
374
375 // create a new Phi for the value
376 PhiNode *phi = new PhiNode(mem->in(0), phi_type, nullptr, mem->_idx, instance_id, alias_idx, offset);
377 transform_later(phi);
378 value_phis->push(phi, mem->_idx);
379
380 for (uint j = 1; j < length; j++) {
381 Node *in = mem->in(j);
382 if (in == nullptr || in->is_top()) {
383 values.at_put(j, in);
384 } else {
385 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
386 if (val == start_mem || val == alloc_mem) {
387 // hit a sentinel, return appropriate value
388 Node* init_value = value_from_alloc(ft, adr_t, alloc);
389 if (init_value == nullptr) {
390 return nullptr;
391 } else {
392 values.at_put(j, init_value);
393 continue;
394 }
395 }
396 if (val->is_Initialize()) {
397 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
398 }
399 if (val == nullptr) {
400 return nullptr; // can't find a value on this path
401 }
402 if (val == mem) {
403 values.at_put(j, mem);
404 } else if (val->is_Store()) {
405 Node* n = val->in(MemNode::ValueIn);
406 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
407 n = bs->step_over_gc_barrier(n);
408 if (is_subword_type(ft)) {
409 n = Compile::narrow_value(ft, n, phi_type, &_igvn, true);
410 }
411 values.at_put(j, n);
412 } else if (val->is_Proj() && val->in(0) == alloc) {
413 Node* init_value = value_from_alloc(ft, adr_t, alloc);
414 if (init_value == nullptr) {
415 return nullptr;
416 } else {
417 values.at_put(j, init_value);
418 }
419 } else if (val->is_Phi()) {
420 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
421 if (val == nullptr) {
422 return nullptr;
423 }
424 values.at_put(j, val);
425 } else if (val->Opcode() == Op_SCMemProj) {
426 assert(val->in(0)->is_LoadStore() ||
427 val->in(0)->Opcode() == Op_EncodeISOArray ||
428 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity");
429 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
430 return nullptr;
431 } else if (val->is_ArrayCopy()) {
432 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc);
433 if (res == nullptr) {
434 return nullptr;
435 }
436 values.at_put(j, res);
437 } else if (val->is_top()) {
438 // This indicates that this path into the phi is dead. Top will eventually also propagate into the Region.
439 // IGVN will clean this up later.
440 values.at_put(j, val);
441 } else {
442 DEBUG_ONLY( val->dump(); )
443 assert(false, "unknown node on this path");
444 return nullptr; // unknown node on this path
445 }
446 }
447 }
448 // Set Phi's inputs
449 for (uint j = 1; j < length; j++) {
450 if (values.at(j) == mem) {
451 phi->init_req(j, phi);
452 } else {
453 phi->init_req(j, values.at(j));
454 }
455 }
456 return phi;
457 }
458
459 // Extract the initial value of a field in an allocation
460 Node* PhaseMacroExpand::value_from_alloc(BasicType ft, const TypeOopPtr* adr_t, AllocateNode* alloc) {
461 Node* init_value = alloc->in(AllocateNode::InitValue);
462 if (init_value == nullptr) {
463 assert(alloc->in(AllocateNode::RawInitValue) == nullptr, "conflicting InitValue and RawInitValue");
464 return _igvn.zerocon(ft);
465 }
466
467 const TypeAryPtr* ary_t = adr_t->isa_aryptr();
468 assert(ary_t != nullptr, "must be a pointer into an array");
469
470 // If this is not a flat array, then it must be an oop array with elements being init_value
471 if (ary_t->is_not_flat()) {
472 #ifdef ASSERT
473 BasicType init_bt = init_value->bottom_type()->basic_type();
474 assert(ft == init_bt ||
475 (!is_java_primitive(ft) && !is_java_primitive(init_bt) && type2aelembytes(ft, true) == type2aelembytes(init_bt, true)) ||
476 (is_subword_type(ft) && init_bt == T_INT),
477 "invalid init_value of type %s for field of type %s", type2name(init_bt), type2name(ft));
478 #endif // ASSERT
479 return init_value;
480 }
481
482 assert(ary_t->klass_is_exact() && ary_t->is_flat(), "must be an exact flat array");
483 assert(ary_t->field_offset().get() != Type::OffsetBot, "unknown offset");
484 if (init_value->is_EncodeP()) {
485 init_value = init_value->in(1);
486 }
487 // Cannot look through init_value if it is an oop
488 if (!init_value->is_InlineType()) {
489 return nullptr;
490 }
491
492 ciInlineKlass* vk = init_value->bottom_type()->inline_klass();
493 if (ary_t->field_offset().get() == vk->null_marker_offset_in_payload()) {
494 init_value = init_value->as_InlineType()->get_null_marker();
495 } else {
496 init_value = init_value->as_InlineType()->field_value_by_offset(ary_t->field_offset().get() + vk->payload_offset(), true);
497 }
498
499 if (ft == T_NARROWOOP) {
500 assert(init_value->bottom_type()->isa_ptr(), "must be a pointer");
501 init_value = transform_later(new EncodePNode(init_value, init_value->bottom_type()->make_narrowoop()));
502 }
503
504 #ifdef ASSERT
505 BasicType init_bt = init_value->bottom_type()->basic_type();
506 assert(ft == init_bt ||
507 (!is_java_primitive(ft) && !is_java_primitive(init_bt) && type2aelembytes(ft, true) == type2aelembytes(init_bt, true)) ||
508 (is_subword_type(ft) && init_bt == T_INT),
509 "invalid init_value of type %s for field of type %s", type2name(init_bt), type2name(ft));
510 #endif // ASSERT
511
512 return init_value;
513 }
514
515 // Search the last value stored into the object's field.
516 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) {
517 assert(adr_t->is_known_instance_field(), "instance required");
518 int instance_id = adr_t->instance_id();
519 assert((uint)instance_id == alloc->_idx, "wrong allocation");
520
521 int alias_idx = C->get_alias_index(adr_t);
522 int offset = adr_t->flat_offset();
523 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
524 Node *alloc_mem = alloc->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
525 assert(alloc_mem != nullptr, "Allocation without a memory projection.");
526 VectorSet visited;
527
528 bool done = sfpt_mem == alloc_mem;
529 Node *mem = sfpt_mem;
530 while (!done) {
531 if (visited.test_set(mem->_idx)) {
532 return nullptr; // found a loop, give up
533 }
534 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
535 if (mem == start_mem || mem == alloc_mem) {
536 done = true; // hit a sentinel, return appropriate 0 value
537 } else if (mem->is_Initialize()) {
538 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
539 if (mem == nullptr) {
540 done = true; // Something went wrong.
541 } else if (mem->is_Store()) {
542 const TypePtr* atype = mem->as_Store()->adr_type();
543 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
544 done = true;
545 }
546 } else if (mem->is_Store()) {
547 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
548 assert(atype != nullptr, "address type must be oopptr");
549 assert(C->get_alias_index(atype) == alias_idx &&
550 atype->is_known_instance_field() && atype->flat_offset() == offset &&
551 atype->instance_id() == instance_id, "store is correct memory slice");
552 done = true;
553 } else if (mem->is_Phi()) {
554 // try to find a phi's unique input
555 Node *unique_input = nullptr;
556 Node *top = C->top();
557 for (uint i = 1; i < mem->req(); i++) {
558 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
559 if (n == nullptr || n == top || n == mem) {
560 continue;
561 } else if (unique_input == nullptr) {
562 unique_input = n;
563 } else if (unique_input != n) {
564 unique_input = top;
565 break;
566 }
567 }
568 if (unique_input != nullptr && unique_input != top) {
569 mem = unique_input;
570 } else {
571 done = true;
572 }
573 } else if (mem->is_ArrayCopy()) {
574 done = true;
575 } else if (mem->is_top()) {
576 // The slice is on a dead path. Returning nullptr would lead to elimination
577 // bailout, but we want to prevent that. Just forwarding the top is also legal,
578 // and IGVN can just clean things up, and remove whatever receives top.
579 return mem;
580 } else {
581 DEBUG_ONLY( mem->dump(); )
582 assert(false, "unexpected node");
583 }
584 }
585 if (mem != nullptr) {
586 if (mem == start_mem || mem == alloc_mem) {
587 // hit a sentinel, return appropriate value
588 return value_from_alloc(ft, adr_t, alloc);
589 } else if (mem->is_Store()) {
590 Node* n = mem->in(MemNode::ValueIn);
591 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
592 n = bs->step_over_gc_barrier(n);
593 return n;
594 } else if (mem->is_Phi()) {
595 // attempt to produce a Phi reflecting the values on the input paths of the Phi
596 Node_Stack value_phis(8);
597 Node* phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
598 if (phi != nullptr) {
599 return phi;
600 } else {
601 // Kill all new Phis
602 while(value_phis.is_nonempty()) {
603 Node* n = value_phis.node();
604 _igvn.replace_node(n, C->top());
605 value_phis.pop();
606 }
607 }
608 } else if (mem->is_ArrayCopy()) {
609 Node* ctl = mem->in(0);
610 Node* m = mem->in(TypeFunc::Memory);
611 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj()) {
612 // pin the loads in the uncommon trap path
613 ctl = sfpt_ctl;
614 m = sfpt_mem;
615 }
616 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc);
617 }
618 }
619 // Something went wrong.
620 return nullptr;
621 }
622
623 // Search the last value stored into the inline type's fields (for flat arrays).
624 Node* PhaseMacroExpand::inline_type_from_mem(ciInlineKlass* vk, const TypeAryPtr* elem_adr_type, int elem_idx, int offset_in_element, bool null_free, AllocateNode* alloc, SafePointNode* sfpt) {
625 auto report_failure = [&](int field_offset_in_element) {
626 #ifndef PRODUCT
627 if (PrintEliminateAllocations) {
628 ciInlineKlass* elem_klass = elem_adr_type->elem()->inline_klass();
629 int offset = field_offset_in_element + elem_klass->payload_offset();
630 ciField* flattened_field = elem_klass->get_field_by_offset(offset, false);
631 assert(flattened_field != nullptr, "must have a field of type %s at offset %d", elem_klass->name()->as_utf8(), offset);
632 tty->print("=== At SafePoint node %d can't find value of field [%s] of array element [%d]", sfpt->_idx, flattened_field->name()->as_utf8(), elem_idx);
633 tty->print(", which prevents elimination of: ");
634 alloc->dump();
635 }
636 #endif // PRODUCT
637 };
638
639 // Create a new InlineTypeNode and retrieve the field values from memory
640 InlineTypeNode* vt = InlineTypeNode::make_uninitialized(_igvn, vk, false);
641 transform_later(vt);
642 if (null_free) {
643 vt->set_null_marker(_igvn);
644 } else {
645 int nm_offset_in_element = offset_in_element + vk->null_marker_offset_in_payload();
646 const TypeAryPtr* nm_adr_type = elem_adr_type->with_field_offset(nm_offset_in_element);
647 Node* nm_value = value_from_mem(sfpt->memory(), sfpt->control(), T_BOOLEAN, TypeInt::BOOL, nm_adr_type, alloc);
648 if (nm_value != nullptr) {
649 vt->set_null_marker(_igvn, nm_value);
650 } else {
651 report_failure(nm_offset_in_element);
652 return nullptr;
653 }
654 }
655
656 for (int i = 0; i < vk->nof_declared_nonstatic_fields(); ++i) {
657 ciField* field = vt->field(i);
658 ciType* field_type = field->type();
659 int field_offset_in_element = offset_in_element + field->offset_in_bytes() - vk->payload_offset();
660 Node* field_value = nullptr;
661 assert(!field->is_flat() || field->type()->is_inlinetype(), "must be an inline type");
662 if (field->is_flat()) {
663 field_value = inline_type_from_mem(field_type->as_inline_klass(), elem_adr_type, elem_idx, field_offset_in_element, field->is_null_free(), alloc, sfpt);
664 } else {
665 const Type* ft = Type::get_const_type(field_type);
666 BasicType bt = type2field[field_type->basic_type()];
667 if (UseCompressedOops && !is_java_primitive(bt)) {
668 ft = ft->make_narrowoop();
669 bt = T_NARROWOOP;
670 }
671 // Each inline type field has its own memory slice
672 const TypeAryPtr* field_adr_type = elem_adr_type->with_field_offset(field_offset_in_element);
673 field_value = value_from_mem(sfpt->memory(), sfpt->control(), bt, ft, field_adr_type, alloc);
674 if (field_value == nullptr) {
675 report_failure(field_offset_in_element);
676 } else if (ft->isa_narrowoop()) {
677 assert(UseCompressedOops, "unexpected narrow oop");
678 if (field_value->is_EncodeP()) {
679 field_value = field_value->in(1);
680 } else if (!field_value->is_InlineType()) {
681 field_value = transform_later(new DecodeNNode(field_value, field_value->get_ptr_type()));
682 }
683 }
684 }
685 if (field_value != nullptr) {
686 vt->set_field_value(i, field_value);
687 } else {
688 return nullptr;
689 }
690 }
691 return vt;
692 }
693
694 // Check the possibility of scalar replacement.
695 bool PhaseMacroExpand::can_eliminate_allocation(PhaseIterGVN* igvn, AllocateNode *alloc, GrowableArray <SafePointNode *>* safepoints) {
696 // Scan the uses of the allocation to check for anything that would
697 // prevent us from eliminating it.
698 NOT_PRODUCT( const char* fail_eliminate = nullptr; )
699 DEBUG_ONLY( Node* disq_node = nullptr; )
700 bool can_eliminate = true;
701 bool reduce_merge_precheck = (safepoints == nullptr);
702
703 Unique_Node_List worklist;
704 Node* res = alloc->result_cast();
705 const TypeOopPtr* res_type = nullptr;
706 if (res == nullptr) {
707 // All users were eliminated.
708 } else if (!res->is_CheckCastPP()) {
709 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
710 can_eliminate = false;
711 } else {
712 worklist.push(res);
713 res_type = igvn->type(res)->isa_oopptr();
714 if (res_type == nullptr) {
715 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
716 can_eliminate = false;
717 } else if (!res_type->klass_is_exact()) {
718 NOT_PRODUCT(fail_eliminate = "Not an exact type.";)
719 can_eliminate = false;
720 } else if (res_type->isa_aryptr()) {
721 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
722 if (length < 0) {
723 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
724 can_eliminate = false;
725 }
726 }
727 }
728
729 while (can_eliminate && worklist.size() > 0) {
730 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
731 res = worklist.pop();
732 for (DUIterator_Fast jmax, j = res->fast_outs(jmax); j < jmax && can_eliminate; j++) {
733 Node* use = res->fast_out(j);
734
735 if (use->is_AddP()) {
736 const TypePtr* addp_type = igvn->type(use)->is_ptr();
737 int offset = addp_type->offset();
738
739 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
740 NOT_PRODUCT(fail_eliminate = "Undefined field reference";)
741 can_eliminate = false;
742 break;
743 }
744 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
745 k < kmax && can_eliminate; k++) {
746 Node* n = use->fast_out(k);
747 if ((n->is_Mem() && n->as_Mem()->is_mismatched_access()) || n->is_LoadFlat() || n->is_StoreFlat()) {
748 DEBUG_ONLY(disq_node = n);
749 NOT_PRODUCT(fail_eliminate = "Mismatched access");
750 can_eliminate = false;
751 }
752 if (!n->is_Store() && n->Opcode() != Op_CastP2X && !bs->is_gc_pre_barrier_node(n) && !reduce_merge_precheck) {
753 DEBUG_ONLY(disq_node = n;)
754 if (n->is_Load() || n->is_LoadStore()) {
755 NOT_PRODUCT(fail_eliminate = "Field load";)
756 } else {
757 NOT_PRODUCT(fail_eliminate = "Not store field reference";)
758 }
759 can_eliminate = false;
760 }
761 }
762 } else if (use->is_ArrayCopy() &&
763 (use->as_ArrayCopy()->is_clonebasic() ||
764 use->as_ArrayCopy()->is_arraycopy_validated() ||
765 use->as_ArrayCopy()->is_copyof_validated() ||
766 use->as_ArrayCopy()->is_copyofrange_validated()) &&
767 use->in(ArrayCopyNode::Dest) == res) {
768 // ok to eliminate
769 } else if (use->is_SafePoint()) {
770 SafePointNode* sfpt = use->as_SafePoint();
771 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
772 // Object is passed as argument.
773 DEBUG_ONLY(disq_node = use;)
774 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
775 can_eliminate = false;
776 }
777 Node* sfptMem = sfpt->memory();
778 if (sfptMem == nullptr || sfptMem->is_top()) {
779 DEBUG_ONLY(disq_node = use;)
780 NOT_PRODUCT(fail_eliminate = "null or TOP memory";)
781 can_eliminate = false;
782 } else if (!reduce_merge_precheck) {
783 assert(!res->is_Phi() || !res->as_Phi()->can_be_inline_type(), "Inline type allocations should not have safepoint uses");
784 safepoints->append_if_missing(sfpt);
785 }
786 } else if (use->is_InlineType() && use->as_InlineType()->get_oop() == res) {
787 // Look at uses
788 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); k < kmax; k++) {
789 Node* u = use->fast_out(k);
790 if (u->is_InlineType()) {
791 // Use in flat field can be eliminated
792 InlineTypeNode* vt = u->as_InlineType();
793 for (uint i = 0; i < vt->field_count(); ++i) {
794 if (vt->field_value(i) == use && !vt->field(i)->is_flat()) {
795 can_eliminate = false; // Use in non-flat field
796 break;
797 }
798 }
799 } else {
800 // Add other uses to the worklist to process individually
801 worklist.push(use);
802 }
803 }
804 } else if (use->Opcode() == Op_StoreX && use->in(MemNode::Address) == res) {
805 // Store to mark word of inline type larval buffer
806 assert(res_type->is_inlinetypeptr(), "Unexpected store to mark word");
807 } else if (res_type->is_inlinetypeptr() && (use->Opcode() == Op_MemBarRelease || use->Opcode() == Op_MemBarStoreStore)) {
808 // Inline type buffer allocations are followed by a membar
809 } else if (reduce_merge_precheck &&
810 (use->is_Phi() || use->is_EncodeP() ||
811 use->Opcode() == Op_MemBarRelease ||
812 (UseStoreStoreForCtor && use->Opcode() == Op_MemBarStoreStore))) {
813 // Nothing to do
814 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
815 if (use->is_Phi()) {
816 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
817 NOT_PRODUCT(fail_eliminate = "Object is return value";)
818 } else {
819 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
820 }
821 DEBUG_ONLY(disq_node = use;)
822 } else {
823 if (use->Opcode() == Op_Return) {
824 NOT_PRODUCT(fail_eliminate = "Object is return value";)
825 } else {
826 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
827 }
828 DEBUG_ONLY(disq_node = use;)
829 }
830 can_eliminate = false;
831 } else {
832 assert(use->Opcode() == Op_CastP2X, "should be");
833 assert(!use->has_out_with(Op_OrL), "should have been removed because oop is never null");
834 }
835 }
836 }
837
838 #ifndef PRODUCT
839 if (PrintEliminateAllocations && safepoints != nullptr) {
840 if (can_eliminate) {
841 tty->print("Scalar ");
842 if (res == nullptr)
843 alloc->dump();
844 else
845 res->dump();
846 } else {
847 tty->print("NotScalar (%s)", fail_eliminate);
848 if (res == nullptr)
849 alloc->dump();
850 else
851 res->dump();
852 #ifdef ASSERT
853 if (disq_node != nullptr) {
854 tty->print(" >>>> ");
855 disq_node->dump();
856 }
857 #endif /*ASSERT*/
858 }
859 }
860
861 if (TraceReduceAllocationMerges && !can_eliminate && reduce_merge_precheck) {
862 tty->print_cr("\tCan't eliminate allocation because '%s': ", fail_eliminate != nullptr ? fail_eliminate : "");
863 DEBUG_ONLY(if (disq_node != nullptr) disq_node->dump();)
864 }
865 #endif
866 return can_eliminate;
931 // CheckCastPP result was not updated in the stack slot, and so
932 // we ended up using the CastPP. That means that the field knows
933 // that it should get an oop from an interface, but the value lost
934 // that information, and so it is not a subtype.
935 // There may be other issues, feel free to investigate further!
936 if (!is_java_primitive(value_bt)) { return; }
937
938 tty->print_cr("value not compatible for field: %s vs %s",
939 type2name(value_bt),
940 type2name(field_bt));
941 tty->print("value_type: ");
942 value_type->dump();
943 tty->cr();
944 tty->print("field_type: ");
945 field_type->dump();
946 tty->cr();
947 assert(false, "value_type does not fit field_type");
948 }
949 #endif
950
951 void PhaseMacroExpand::process_field_value_at_safepoint(const Type* field_type, Node* field_val, SafePointNode* sfpt, Unique_Node_List* value_worklist) {
952 if (UseCompressedOops && field_type->isa_narrowoop()) {
953 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
954 // to be able scalar replace the allocation.
955 if (field_val->is_EncodeP()) {
956 field_val = field_val->in(1);
957 } else if (!field_val->is_InlineType()) {
958 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
959 }
960 }
961
962 // Keep track of inline types to scalarize them later
963 if (field_val->is_InlineType()) {
964 value_worklist->push(field_val);
965 } else if (field_val->is_Phi()) {
966 PhiNode* phi = field_val->as_Phi();
967 // Eagerly replace inline type phis now since we could be removing an inline type allocation where we must
968 // scalarize all its fields in safepoints.
969 field_val = phi->try_push_inline_types_down(&_igvn, true);
970 if (field_val->is_InlineType()) {
971 value_worklist->push(field_val);
972 }
973 }
974 DEBUG_ONLY(verify_type_compatability(field_val->bottom_type(), field_type);)
975 sfpt->add_req(field_val);
976 }
977
978 bool PhaseMacroExpand::add_array_elems_to_safepoint(AllocateNode* alloc, const TypeAryPtr* array_type, SafePointNode* sfpt, Unique_Node_List* value_worklist) {
979 const Type* elem_type = array_type->elem();
980 BasicType basic_elem_type = elem_type->array_element_basic_type();
981
982 intptr_t elem_size;
983 uint header_size;
984 if (array_type->is_flat()) {
985 elem_size = array_type->flat_elem_size();
986 header_size = arrayOopDesc::base_offset_in_bytes(T_FLAT_ELEMENT);
987 } else {
988 elem_size = type2aelembytes(basic_elem_type);
989 header_size = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
990 }
991
992 int n_elems = alloc->in(AllocateNode::ALength)->get_int();
993 for (int elem_idx = 0; elem_idx < n_elems; elem_idx++) {
994 intptr_t elem_offset = header_size + elem_idx * elem_size;
995 const TypeAryPtr* elem_adr_type = array_type->with_offset(elem_offset);
996 Node* elem_val;
997 if (array_type->is_flat()) {
998 ciInlineKlass* elem_klass = elem_type->inline_klass();
999 assert(elem_klass->maybe_flat_in_array(), "must be flat in array");
1000 elem_val = inline_type_from_mem(elem_klass, elem_adr_type, elem_idx, 0, array_type->is_null_free(), alloc, sfpt);
1001 } else {
1002 elem_val = value_from_mem(sfpt->memory(), sfpt->control(), basic_elem_type, elem_type, elem_adr_type, alloc);
1003 #ifndef PRODUCT
1004 if (PrintEliminateAllocations && elem_val == nullptr) {
1005 tty->print("=== At SafePoint node %d can't find value of array element [%d]", sfpt->_idx, elem_idx);
1006 tty->print(", which prevents elimination of: ");
1007 alloc->dump();
1008 }
1009 #endif // PRODUCT
1010 }
1011 if (elem_val == nullptr) {
1012 return false;
1013 }
1014
1015 process_field_value_at_safepoint(elem_type, elem_val, sfpt, value_worklist);
1016 }
1017
1018 return true;
1019 }
1020
1021 // Recursively adds all flattened fields of a type 'iklass' inside 'base' to 'sfpt'.
1022 // 'offset_minus_header' refers to the offset of the payload of 'iklass' inside 'base' minus the
1023 // payload offset of 'iklass'. If 'base' is of type 'iklass' then 'offset_minus_header' == 0.
1024 bool PhaseMacroExpand::add_inst_fields_to_safepoint(ciInstanceKlass* iklass, AllocateNode* alloc, Node* base, int offset_minus_header, SafePointNode* sfpt, Unique_Node_List* value_worklist) {
1025 const TypeInstPtr* base_type = _igvn.type(base)->is_instptr();
1026 auto report_failure = [&](int offset) {
1027 #ifndef PRODUCT
1028 if (PrintEliminateAllocations) {
1029 ciInstanceKlass* base_klass = base_type->instance_klass();
1030 ciField* flattened_field = base_klass->get_field_by_offset(offset, false);
1031 assert(flattened_field != nullptr, "must have a field of type %s at offset %d", base_klass->name()->as_utf8(), offset);
1032 tty->print("=== At SafePoint node %d can't find value of field: ", sfpt->_idx);
1033 flattened_field->print();
1034 int field_idx = C->alias_type(flattened_field)->index();
1035 tty->print(" (alias_idx=%d)", field_idx);
1036 tty->print(", which prevents elimination of: ");
1037 base->dump();
1038 }
1039 #endif // PRODUCT
1040 };
1041
1042 for (int i = 0; i < iklass->nof_declared_nonstatic_fields(); i++) {
1043 ciField* field = iklass->declared_nonstatic_field_at(i);
1044 if (field->is_flat()) {
1045 ciInlineKlass* fvk = field->type()->as_inline_klass();
1046 int field_offset_minus_header = offset_minus_header + field->offset_in_bytes() - fvk->payload_offset();
1047 bool success = add_inst_fields_to_safepoint(fvk, alloc, base, field_offset_minus_header, sfpt, value_worklist);
1048 if (!success) {
1049 return false;
1050 }
1051
1052 // The null marker of a field is added right after we scalarize that field
1053 if (!field->is_null_free()) {
1054 int nm_offset = offset_minus_header + field->null_marker_offset();
1055 Node* null_marker = value_from_mem(sfpt->memory(), sfpt->control(), T_BOOLEAN, TypeInt::BOOL, base_type->with_offset(nm_offset), alloc);
1056 if (null_marker == nullptr) {
1057 report_failure(nm_offset);
1058 return false;
1059 }
1060 process_field_value_at_safepoint(TypeInt::BOOL, null_marker, sfpt, value_worklist);
1061 }
1062
1063 continue;
1064 }
1065
1066 int offset = offset_minus_header + field->offset_in_bytes();
1067 ciType* elem_type = field->type();
1068 BasicType basic_elem_type = field->layout_type();
1069
1070 const Type* field_type;
1071 if (is_reference_type(basic_elem_type)) {
1072 if (!elem_type->is_loaded()) {
1073 field_type = TypeInstPtr::BOTTOM;
1074 } else {
1075 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
1076 }
1077 if (UseCompressedOops) {
1078 field_type = field_type->make_narrowoop();
1079 basic_elem_type = T_NARROWOOP;
1080 }
1081 } else {
1082 field_type = Type::get_const_basic_type(basic_elem_type);
1083 }
1084
1085 const TypeInstPtr* field_addr_type = base_type->add_offset(offset)->isa_instptr();
1086 Node* field_val = value_from_mem(sfpt->memory(), sfpt->control(), basic_elem_type, field_type, field_addr_type, alloc);
1087 if (field_val == nullptr) {
1088 report_failure(offset);
1089 return false;
1090 }
1091 process_field_value_at_safepoint(field_type, field_val, sfpt, value_worklist);
1092 }
1093
1094 return true;
1095 }
1096
1097 SafePointScalarObjectNode* PhaseMacroExpand::create_scalarized_object_description(AllocateNode* alloc, SafePointNode* sfpt,
1098 Unique_Node_List* value_worklist) {
1099 // Fields of scalar objs are referenced only at the end
1100 // of regular debuginfo at the last (youngest) JVMS.
1101 // Record relative start index.
1102 ciInstanceKlass* iklass = nullptr;
1103 const TypeOopPtr* res_type = nullptr;
1104 int nfields = 0;
1105 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
1106 Node* res = alloc->result_cast();
1107
1108 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
1109 assert(sfpt->jvms() != nullptr, "missed JVMS");
1110 uint before_sfpt_req = sfpt->req();
1111
1112 if (res != nullptr) { // Could be null when there are no users
1113 res_type = _igvn.type(res)->isa_oopptr();
1114
1115 if (res_type->isa_instptr()) {
1116 // find the fields of the class which will be needed for safepoint debug information
1117 iklass = res_type->is_instptr()->instance_klass();
1118 nfields = iklass->nof_nonstatic_fields();
1119 } else {
1120 // find the array's elements which will be needed for safepoint debug information
1121 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
1122 assert(nfields >= 0, "must be an array klass.");
1123 }
1124
1125 if (res->bottom_type()->is_inlinetypeptr()) {
1126 // Nullable inline types have a null marker field which is added to the safepoint when scalarizing them (see
1127 // InlineTypeNode::make_scalar_in_safepoint()). When having circular inline types, we stop scalarizing at depth 1
1128 // to avoid an endless recursion. Therefore, we do not have a SafePointScalarObjectNode node here, yet.
1129 // We are about to create a SafePointScalarObjectNode as if this is a normal object. Add an additional int input
1130 // with value 1 which sets the null marker to true to indicate that the object is always non-null. This input is checked
1131 // later in PhaseOutput::filLocArray() for inline types.
1132 sfpt->add_req(_igvn.intcon(1));
1133 }
1134 }
1135
1136 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type, alloc, first_ind, sfpt->jvms()->depth(), nfields);
1137 sobj->init_req(0, C->root());
1138 transform_later(sobj);
1139
1140 if (res == nullptr) {
1141 sfpt->jvms()->set_endoff(sfpt->req());
1142 return sobj;
1143 }
1144
1145 bool success;
1146 if (iklass == nullptr) {
1147 success = add_array_elems_to_safepoint(alloc, res_type->is_aryptr(), sfpt, value_worklist);
1148 } else {
1149 success = add_inst_fields_to_safepoint(iklass, alloc, res, 0, sfpt, value_worklist);
1150 }
1151
1152 // We weren't able to find a value for this field, remove all the fields added to the safepoint
1153 if (!success) {
1154 for (uint i = sfpt->req() - 1; i >= before_sfpt_req; i--) {
1155 sfpt->del_req(i);
1156 }
1157 _igvn._worklist.push(sfpt);
1158 return nullptr;
1159 }
1160
1161 sfpt->jvms()->set_endoff(sfpt->req());
1162 return sobj;
1163 }
1164
1165 // Do scalar replacement.
1166 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
1167 GrowableArray <SafePointNode *> safepoints_done;
1168 Node* res = alloc->result_cast();
1169 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
1170 const TypeOopPtr* res_type = nullptr;
1171 if (res != nullptr) { // Could be null when there are no users
1172 res_type = _igvn.type(res)->isa_oopptr();
1173 }
1174
1175 // Process the safepoint uses
1176 Unique_Node_List value_worklist;
1177 while (safepoints.length() > 0) {
1178 SafePointNode* sfpt = safepoints.pop();
1179 SafePointScalarObjectNode* sobj = create_scalarized_object_description(alloc, sfpt, &value_worklist);
1180
1181 if (sobj == nullptr) {
1182 undo_previous_scalarizations(safepoints_done, alloc);
1183 return false;
1184 }
1185
1186 // Now make a pass over the debug information replacing any references
1187 // to the allocated object with "sobj"
1188 JVMState *jvms = sfpt->jvms();
1189 sfpt->replace_edges_in_range(res, sobj, jvms->debug_start(), jvms->debug_end(), &_igvn);
1190 _igvn._worklist.push(sfpt);
1191
1192 // keep it for rollback
1193 safepoints_done.append_if_missing(sfpt);
1194 }
1195 // Scalarize inline types that were added to the safepoint.
1196 // Don't allow linking a constant oop (if available) for flat array elements
1197 // because Deoptimization::reassign_flat_array_elements needs field values.
1198 bool allow_oop = (res_type != nullptr) && !res_type->is_flat();
1199 for (uint i = 0; i < value_worklist.size(); ++i) {
1200 InlineTypeNode* vt = value_worklist.at(i)->as_InlineType();
1201 vt->make_scalar_in_safepoints(&_igvn, allow_oop);
1202 }
1203 return true;
1204 }
1205
1206 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
1207 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control);
1208 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory);
1209 if (ctl_proj != nullptr) {
1210 igvn.replace_node(ctl_proj, n->in(0));
1211 }
1212 if (mem_proj != nullptr) {
1213 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
1214 }
1215 }
1216
1217 // Process users of eliminated allocation.
1218 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc, bool inline_alloc) {
1219 Unique_Node_List worklist;
1220 Node* res = alloc->result_cast();
1221 if (res != nullptr) {
1222 worklist.push(res);
1223 }
1224 while (worklist.size() > 0) {
1225 res = worklist.pop();
1226 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
1227 Node *use = res->last_out(j);
1228 uint oc1 = res->outcnt();
1229
1230 if (use->is_AddP()) {
1231 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
1232 Node *n = use->last_out(k);
1233 uint oc2 = use->outcnt();
1234 if (n->is_Store()) {
1235 for (DUIterator_Fast pmax, p = n->fast_outs(pmax); p < pmax; p++) {
1236 MemBarNode* mb = n->fast_out(p)->isa_MemBar();
1237 if (mb != nullptr && mb->req() <= MemBarNode::Precedent && mb->in(MemBarNode::Precedent) == n) {
1238 // MemBarVolatiles should have been removed by MemBarNode::Ideal() for non-inline allocations
1239 assert(inline_alloc, "MemBarVolatile should be eliminated for non-escaping object");
1240 mb->remove(&_igvn);
1241 }
1242 }
1243 _igvn.replace_node(n, n->in(MemNode::Memory));
1244 } else {
1245 eliminate_gc_barrier(n);
1246 }
1247 k -= (oc2 - use->outcnt());
1248 }
1249 _igvn.remove_dead_node(use);
1250 } else if (use->is_ArrayCopy()) {
1251 // Disconnect ArrayCopy node
1252 ArrayCopyNode* ac = use->as_ArrayCopy();
1253 if (ac->is_clonebasic()) {
1254 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
1255 disconnect_projections(ac, _igvn);
1256 assert(alloc->in(TypeFunc::Memory)->is_Proj() && alloc->in(TypeFunc::Memory)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
1257 Node* membar_before = alloc->in(TypeFunc::Memory)->in(0);
1258 disconnect_projections(membar_before->as_MemBar(), _igvn);
1259 if (membar_after->is_MemBar()) {
1260 disconnect_projections(membar_after->as_MemBar(), _igvn);
1261 }
1262 } else {
1263 assert(ac->is_arraycopy_validated() ||
1264 ac->is_copyof_validated() ||
1265 ac->is_copyofrange_validated(), "unsupported");
1266 CallProjections* callprojs = ac->extract_projections(true);
1267
1268 _igvn.replace_node(callprojs->fallthrough_ioproj, ac->in(TypeFunc::I_O));
1269 _igvn.replace_node(callprojs->fallthrough_memproj, ac->in(TypeFunc::Memory));
1270 _igvn.replace_node(callprojs->fallthrough_catchproj, ac->in(TypeFunc::Control));
1271
1272 // Set control to top. IGVN will remove the remaining projections
1273 ac->set_req(0, top());
1274 ac->replace_edge(res, top(), &_igvn);
1275
1276 // Disconnect src right away: it can help find new
1277 // opportunities for allocation elimination
1278 Node* src = ac->in(ArrayCopyNode::Src);
1279 ac->replace_edge(src, top(), &_igvn);
1280 // src can be top at this point if src and dest of the
1281 // arraycopy were the same
1282 if (src->outcnt() == 0 && !src->is_top()) {
1283 _igvn.remove_dead_node(src);
1284 }
1285 }
1286 _igvn._worklist.push(ac);
1287 } else if (use->is_InlineType()) {
1288 assert(use->as_InlineType()->get_oop() == res, "unexpected inline type ptr use");
1289 // Cut off oop input and remove known instance id from type
1290 _igvn.rehash_node_delayed(use);
1291 use->as_InlineType()->set_oop(_igvn, _igvn.zerocon(T_OBJECT));
1292 use->as_InlineType()->set_is_buffered(_igvn, false);
1293 const TypeOopPtr* toop = _igvn.type(use)->is_oopptr()->cast_to_instance_id(TypeOopPtr::InstanceBot);
1294 _igvn.set_type(use, toop);
1295 use->as_InlineType()->set_type(toop);
1296 // Process users
1297 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); k < kmax; k++) {
1298 Node* u = use->fast_out(k);
1299 if (!u->is_InlineType() && !u->is_StoreFlat()) {
1300 worklist.push(u);
1301 }
1302 }
1303 } else if (use->Opcode() == Op_StoreX && use->in(MemNode::Address) == res) {
1304 // Store to mark word of inline type larval buffer
1305 assert(inline_alloc, "Unexpected store to mark word");
1306 _igvn.replace_node(use, use->in(MemNode::Memory));
1307 } else if (use->Opcode() == Op_MemBarRelease || use->Opcode() == Op_MemBarStoreStore) {
1308 // Inline type buffer allocations are followed by a membar
1309 assert(inline_alloc, "Unexpected MemBarRelease");
1310 use->as_MemBar()->remove(&_igvn);
1311 } else {
1312 eliminate_gc_barrier(use);
1313 }
1314 j -= (oc1 - res->outcnt());
1315 }
1316 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
1317 _igvn.remove_dead_node(res);
1318 }
1319
1320 //
1321 // Process other users of allocation's projections
1322 //
1323 if (_callprojs->resproj[0] != nullptr && _callprojs->resproj[0]->outcnt() != 0) {
1324 // First disconnect stores captured by Initialize node.
1325 // If Initialize node is eliminated first in the following code,
1326 // it will kill such stores and DUIterator_Last will assert.
1327 for (DUIterator_Fast jmax, j = _callprojs->resproj[0]->fast_outs(jmax); j < jmax; j++) {
1328 Node* use = _callprojs->resproj[0]->fast_out(j);
1329 if (use->is_AddP()) {
1330 // raw memory addresses used only by the initialization
1331 _igvn.replace_node(use, C->top());
1332 --j; --jmax;
1333 }
1334 }
1335 for (DUIterator_Last jmin, j = _callprojs->resproj[0]->last_outs(jmin); j >= jmin; ) {
1336 Node* use = _callprojs->resproj[0]->last_out(j);
1337 uint oc1 = _callprojs->resproj[0]->outcnt();
1338 if (use->is_Initialize()) {
1339 // Eliminate Initialize node.
1340 InitializeNode *init = use->as_Initialize();
1341 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control);
1342 if (ctrl_proj != nullptr) {
1343 _igvn.replace_node(ctrl_proj, init->in(TypeFunc::Control));
1344 #ifdef ASSERT
1345 // If the InitializeNode has no memory out, it will die, and tmp will become null
1346 Node* tmp = init->in(TypeFunc::Control);
1347 assert(tmp == nullptr || tmp == _callprojs->fallthrough_catchproj, "allocation control projection");
1348 #endif
1349 }
1350 Node* mem = init->in(TypeFunc::Memory);
1351 #ifdef ASSERT
1352 if (init->number_of_projs(TypeFunc::Memory) > 0) {
1353 if (mem->is_MergeMem()) {
1354 assert(mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw) == _callprojs->fallthrough_memproj, "allocation memory projection");
1355 } else {
1356 assert(mem == _callprojs->fallthrough_memproj, "allocation memory projection");
1357 }
1358 }
1359 #endif
1360 init->replace_mem_projs_by(mem, &_igvn);
1361 assert(init->outcnt() == 0, "should only have had a control and some memory projections, and we removed them");
1362 } else if (use->Opcode() == Op_MemBarStoreStore) {
1363 // Inline type buffer allocations are followed by a membar
1364 assert(inline_alloc, "Unexpected MemBarStoreStore");
1365 use->as_MemBar()->remove(&_igvn);
1366 } else {
1367 assert(false, "only Initialize or AddP expected");
1368 }
1369 j -= (oc1 - _callprojs->resproj[0]->outcnt());
1370 }
1371 }
1372 if (_callprojs->fallthrough_catchproj != nullptr) {
1373 _igvn.replace_node(_callprojs->fallthrough_catchproj, alloc->in(TypeFunc::Control));
1374 }
1375 if (_callprojs->fallthrough_memproj != nullptr) {
1376 _igvn.replace_node(_callprojs->fallthrough_memproj, alloc->in(TypeFunc::Memory));
1377 }
1378 if (_callprojs->catchall_memproj != nullptr) {
1379 _igvn.replace_node(_callprojs->catchall_memproj, C->top());
1380 }
1381 if (_callprojs->fallthrough_ioproj != nullptr) {
1382 _igvn.replace_node(_callprojs->fallthrough_ioproj, alloc->in(TypeFunc::I_O));
1383 }
1384 if (_callprojs->catchall_ioproj != nullptr) {
1385 _igvn.replace_node(_callprojs->catchall_ioproj, C->top());
1386 }
1387 if (_callprojs->catchall_catchproj != nullptr) {
1388 _igvn.replace_node(_callprojs->catchall_catchproj, C->top());
1389 }
1390 }
1391
1392 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1393 // If reallocation fails during deoptimization we'll pop all
1394 // interpreter frames for this compiled frame and that won't play
1395 // nice with JVMTI popframe.
1396 // We avoid this issue by eager reallocation when the popframe request
1397 // is received.
1398 if (!EliminateAllocations) {
1399 return false;
1400 }
1401 Node* klass = alloc->in(AllocateNode::KlassNode);
1402 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1403
1404 // Attempt to eliminate inline type buffer allocations
1405 // regardless of usage and escape/replaceable status.
1406 bool inline_alloc = tklass->isa_instklassptr() &&
1407 tklass->is_instklassptr()->instance_klass()->is_inlinetype();
1408 if (!alloc->_is_non_escaping && !inline_alloc) {
1409 return false;
1410 }
1411 // Eliminate boxing allocations which are not used
1412 // regardless scalar replaceable status.
1413 Node* res = alloc->result_cast();
1414 bool boxing_alloc = (res == nullptr) && C->eliminate_boxing() &&
1415 tklass->isa_instklassptr() &&
1416 tklass->is_instklassptr()->instance_klass()->is_box_klass();
1417 if (!alloc->_is_scalar_replaceable && !boxing_alloc && !inline_alloc) {
1418 return false;
1419 }
1420
1421 _callprojs = alloc->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1422
1423 GrowableArray <SafePointNode *> safepoints;
1424 if (!can_eliminate_allocation(&_igvn, alloc, &safepoints)) {
1425 return false;
1426 }
1427
1428 if (!alloc->_is_scalar_replaceable) {
1429 assert(res == nullptr || inline_alloc, "sanity");
1430 // We can only eliminate allocation if all debug info references
1431 // are already replaced with SafePointScalarObject because
1432 // we can't search for a fields value without instance_id.
1433 if (safepoints.length() > 0) {
1434 return false;
1435 }
1436 }
1437
1438 if (!scalar_replacement(alloc, safepoints)) {
1439 return false;
1440 }
1441
1442 CompileLog* log = C->log();
1443 if (log != nullptr) {
1444 log->head("eliminate_allocation type='%d'",
1445 log->identify(tklass->exact_klass()));
1446 JVMState* p = alloc->jvms();
1447 while (p != nullptr) {
1448 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1449 p = p->caller();
1450 }
1451 log->tail("eliminate_allocation");
1452 }
1453
1454 process_users_of_allocation(alloc, inline_alloc);
1455
1456 #ifndef PRODUCT
1457 if (PrintEliminateAllocations) {
1458 if (alloc->is_AllocateArray())
1459 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1460 else
1461 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1462 }
1463 #endif
1464
1465 return true;
1466 }
1467
1468 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1469 // EA should remove all uses of non-escaping boxing node.
1470 if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != nullptr) {
1471 return false;
1472 }
1473
1474 assert(boxing->result_cast() == nullptr, "unexpected boxing node result");
1475
1476 _callprojs = boxing->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1477
1478 const TypeTuple* r = boxing->tf()->range_sig();
1479 assert(r->cnt() > TypeFunc::Parms, "sanity");
1480 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1481 assert(t != nullptr, "sanity");
1482
1483 CompileLog* log = C->log();
1484 if (log != nullptr) {
1485 log->head("eliminate_boxing type='%d'",
1486 log->identify(t->instance_klass()));
1487 JVMState* p = boxing->jvms();
1488 while (p != nullptr) {
1489 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1490 p = p->caller();
1491 }
1492 log->tail("eliminate_boxing");
1493 }
1494
1495 process_users_of_allocation(boxing);
1496
1497 #ifndef PRODUCT
1498 if (PrintEliminateAllocations) {
1562 // oop flavor.
1563 //
1564 //=============================================================================
1565 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1566 // Allocations bigger than this always go the slow route.
1567 // This value must be small enough that allocation attempts that need to
1568 // trigger exceptions go the slow route. Also, it must be small enough so
1569 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1570 //=============================================================================j//
1571 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1572 // The allocator will coalesce int->oop copies away. See comment in
1573 // coalesce.cpp about how this works. It depends critically on the exact
1574 // code shape produced here, so if you are changing this code shape
1575 // make sure the GC info for the heap-top is correct in and around the
1576 // slow-path call.
1577 //
1578
1579 void PhaseMacroExpand::expand_allocate_common(
1580 AllocateNode* alloc, // allocation node to be expanded
1581 Node* length, // array length for an array allocation
1582 Node* init_val, // value to initialize the array with
1583 const TypeFunc* slow_call_type, // Type of slow call
1584 address slow_call_address, // Address of slow call
1585 Node* valid_length_test // whether length is valid or not
1586 )
1587 {
1588 Node* ctrl = alloc->in(TypeFunc::Control);
1589 Node* mem = alloc->in(TypeFunc::Memory);
1590 Node* i_o = alloc->in(TypeFunc::I_O);
1591 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1592 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1593 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1594 assert(ctrl != nullptr, "must have control");
1595
1596 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1597 // they will not be used if "always_slow" is set
1598 enum { slow_result_path = 1, fast_result_path = 2 };
1599 Node *result_region = nullptr;
1600 Node *result_phi_rawmem = nullptr;
1601 Node *result_phi_rawoop = nullptr;
1602 Node *result_phi_i_o = nullptr;
1647 #endif
1648 yank_alloc_node(alloc);
1649 return;
1650 }
1651 }
1652
1653 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1654 Node *slow_region = nullptr;
1655 Node *toobig_false = ctrl;
1656
1657 // generate the initial test if necessary
1658 if (initial_slow_test != nullptr ) {
1659 assert (expand_fast_path, "Only need test if there is a fast path");
1660 slow_region = new RegionNode(3);
1661
1662 // Now make the initial failure test. Usually a too-big test but
1663 // might be a TRUE for finalizers.
1664 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1665 transform_later(toobig_iff);
1666 // Plug the failing-too-big test into the slow-path region
1667 Node* toobig_true = new IfTrueNode(toobig_iff);
1668 transform_later(toobig_true);
1669 slow_region ->init_req( too_big_or_final_path, toobig_true );
1670 toobig_false = new IfFalseNode(toobig_iff);
1671 transform_later(toobig_false);
1672 } else {
1673 // No initial test, just fall into next case
1674 assert(allocation_has_use || !expand_fast_path, "Should already have been handled");
1675 toobig_false = ctrl;
1676 DEBUG_ONLY(slow_region = NodeSentinel);
1677 }
1678
1679 // If we are here there are several possibilities
1680 // - expand_fast_path is false - then only a slow path is expanded. That's it.
1681 // no_initial_check means a constant allocation.
1682 // - If check always evaluates to false -> expand_fast_path is false (see above)
1683 // - If check always evaluates to true -> directly into fast path (but may bailout to slowpath)
1684 // if !allocation_has_use the fast path is empty
1685 // if !allocation_has_use && no_initial_check
1686 // - Then there are no fastpath that can fall out to slowpath -> no allocation code at all.
1687 // removed by yank_alloc_node above.
1688
1689 Node *slow_mem = mem; // save the current memory state for slow path
1690 // generate the fast allocation code unless we know that the initial test will always go slow
1691 if (expand_fast_path) {
1692 // Fast path modifies only raw memory.
1693 if (mem->is_MergeMem()) {
1694 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1695 }
1696
1697 // allocate the Region and Phi nodes for the result
1698 result_region = new RegionNode(3);
1699 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1700 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1701
1702 // Grab regular I/O before optional prefetch may change it.
1703 // Slow-path does no I/O so just set it to the original I/O.
1704 result_phi_i_o->init_req(slow_result_path, i_o);
1705
1706 // Name successful fast-path variables
1707 Node* fast_oop_ctrl;
1708 Node* fast_oop_rawmem;
1709
1710 if (allocation_has_use) {
1711 Node* needgc_ctrl = nullptr;
1712 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1713
1714 intx prefetch_lines = length != nullptr ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1715 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1716 Node* fast_oop = bs->obj_allocate(this, mem, toobig_false, size_in_bytes, i_o, needgc_ctrl,
1717 fast_oop_ctrl, fast_oop_rawmem,
1718 prefetch_lines);
1719
1720 if (initial_slow_test != nullptr) {
1721 // This completes all paths into the slow merge point
1722 slow_region->init_req(need_gc_path, needgc_ctrl);
1723 transform_later(slow_region);
1724 } else {
1725 // No initial slow path needed!
1726 // Just fall from the need-GC path straight into the VM call.
1727 slow_region = needgc_ctrl;
1728 }
1729
1747 result_phi_i_o ->init_req(fast_result_path, i_o);
1748 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1749 } else {
1750 slow_region = ctrl;
1751 result_phi_i_o = i_o; // Rename it to use in the following code.
1752 }
1753
1754 // Generate slow-path call
1755 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1756 OptoRuntime::stub_name(slow_call_address),
1757 TypePtr::BOTTOM);
1758 call->init_req(TypeFunc::Control, slow_region);
1759 call->init_req(TypeFunc::I_O, top()); // does no i/o
1760 call->init_req(TypeFunc::Memory, slow_mem); // may gc ptrs
1761 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1762 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1763
1764 call->init_req(TypeFunc::Parms+0, klass_node);
1765 if (length != nullptr) {
1766 call->init_req(TypeFunc::Parms+1, length);
1767 if (init_val != nullptr) {
1768 call->init_req(TypeFunc::Parms+2, init_val);
1769 }
1770 } else {
1771 // Let the runtime know if this is a larval allocation
1772 call->init_req(TypeFunc::Parms+1, _igvn.intcon(alloc->_larval));
1773 }
1774
1775 // Copy debug information and adjust JVMState information, then replace
1776 // allocate node with the call
1777 call->copy_call_debug_info(&_igvn, alloc);
1778 // For array allocations, copy the valid length check to the call node so Compile::final_graph_reshaping() can verify
1779 // that the call has the expected number of CatchProj nodes (in case the allocation always fails and the fallthrough
1780 // path dies).
1781 if (valid_length_test != nullptr) {
1782 call->add_req(valid_length_test);
1783 }
1784 if (expand_fast_path) {
1785 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1786 } else {
1787 // Hook i_o projection to avoid its elimination during allocation
1788 // replacement (when only a slow call is generated).
1789 call->set_req(TypeFunc::I_O, result_phi_i_o);
1790 }
1791 _igvn.replace_node(alloc, call);
1792 transform_later(call);
1793
1794 // Identify the output projections from the allocate node and
1795 // adjust any references to them.
1796 // The control and io projections look like:
1797 //
1798 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1799 // Allocate Catch
1800 // ^---Proj(io) <-------+ ^---CatchProj(io)
1801 //
1802 // We are interested in the CatchProj nodes.
1803 //
1804 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1805
1806 // An allocate node has separate memory projections for the uses on
1807 // the control and i_o paths. Replace the control memory projection with
1808 // result_phi_rawmem (unless we are only generating a slow call when
1809 // both memory projections are combined)
1810 if (expand_fast_path && _callprojs->fallthrough_memproj != nullptr) {
1811 _igvn.replace_in_uses(_callprojs->fallthrough_memproj, result_phi_rawmem);
1812 }
1813 // Now change uses of catchall_memproj to use fallthrough_memproj and delete
1814 // catchall_memproj so we end up with a call that has only 1 memory projection.
1815 if (_callprojs->catchall_memproj != nullptr) {
1816 if (_callprojs->fallthrough_memproj == nullptr) {
1817 _callprojs->fallthrough_memproj = new ProjNode(call, TypeFunc::Memory);
1818 transform_later(_callprojs->fallthrough_memproj);
1819 }
1820 _igvn.replace_in_uses(_callprojs->catchall_memproj, _callprojs->fallthrough_memproj);
1821 _igvn.remove_dead_node(_callprojs->catchall_memproj);
1822 }
1823
1824 // An allocate node has separate i_o projections for the uses on the control
1825 // and i_o paths. Always replace the control i_o projection with result i_o
1826 // otherwise incoming i_o become dead when only a slow call is generated
1827 // (it is different from memory projections where both projections are
1828 // combined in such case).
1829 if (_callprojs->fallthrough_ioproj != nullptr) {
1830 _igvn.replace_in_uses(_callprojs->fallthrough_ioproj, result_phi_i_o);
1831 }
1832 // Now change uses of catchall_ioproj to use fallthrough_ioproj and delete
1833 // catchall_ioproj so we end up with a call that has only 1 i_o projection.
1834 if (_callprojs->catchall_ioproj != nullptr) {
1835 if (_callprojs->fallthrough_ioproj == nullptr) {
1836 _callprojs->fallthrough_ioproj = new ProjNode(call, TypeFunc::I_O);
1837 transform_later(_callprojs->fallthrough_ioproj);
1838 }
1839 _igvn.replace_in_uses(_callprojs->catchall_ioproj, _callprojs->fallthrough_ioproj);
1840 _igvn.remove_dead_node(_callprojs->catchall_ioproj);
1841 }
1842
1843 // if we generated only a slow call, we are done
1844 if (!expand_fast_path) {
1845 // Now we can unhook i_o.
1846 if (result_phi_i_o->outcnt() > 1) {
1847 call->set_req(TypeFunc::I_O, top());
1848 } else {
1849 assert(result_phi_i_o->unique_ctrl_out() == call, "sanity");
1850 // Case of new array with negative size known during compilation.
1851 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1852 // following code since call to runtime will throw exception.
1853 // As result there will be no users of i_o after the call.
1854 // Leave i_o attached to this call to avoid problems in preceding graph.
1855 }
1856 return;
1857 }
1858
1859 if (_callprojs->fallthrough_catchproj != nullptr) {
1860 ctrl = _callprojs->fallthrough_catchproj->clone();
1861 transform_later(ctrl);
1862 _igvn.replace_node(_callprojs->fallthrough_catchproj, result_region);
1863 } else {
1864 ctrl = top();
1865 }
1866 Node *slow_result;
1867 if (_callprojs->resproj[0] == nullptr) {
1868 // no uses of the allocation result
1869 slow_result = top();
1870 } else {
1871 slow_result = _callprojs->resproj[0]->clone();
1872 transform_later(slow_result);
1873 _igvn.replace_node(_callprojs->resproj[0], result_phi_rawoop);
1874 }
1875
1876 // Plug slow-path into result merge point
1877 result_region->init_req( slow_result_path, ctrl);
1878 transform_later(result_region);
1879 if (allocation_has_use) {
1880 result_phi_rawoop->init_req(slow_result_path, slow_result);
1881 transform_later(result_phi_rawoop);
1882 }
1883 result_phi_rawmem->init_req(slow_result_path, _callprojs->fallthrough_memproj);
1884 transform_later(result_phi_rawmem);
1885 transform_later(result_phi_i_o);
1886 // This completes all paths into the result merge point
1887 }
1888
1889 // Remove alloc node that has no uses.
1890 void PhaseMacroExpand::yank_alloc_node(AllocateNode* alloc) {
1891 Node* ctrl = alloc->in(TypeFunc::Control);
1892 Node* mem = alloc->in(TypeFunc::Memory);
1893 Node* i_o = alloc->in(TypeFunc::I_O);
1894
1895 _callprojs = alloc->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1896 if (_callprojs->resproj[0] != nullptr) {
1897 for (DUIterator_Fast imax, i = _callprojs->resproj[0]->fast_outs(imax); i < imax; i++) {
1898 Node* use = _callprojs->resproj[0]->fast_out(i);
1899 use->isa_MemBar()->remove(&_igvn);
1900 --imax;
1901 --i; // back up iterator
1902 }
1903 assert(_callprojs->resproj[0]->outcnt() == 0, "all uses must be deleted");
1904 _igvn.remove_dead_node(_callprojs->resproj[0]);
1905 }
1906 if (_callprojs->fallthrough_catchproj != nullptr) {
1907 _igvn.replace_in_uses(_callprojs->fallthrough_catchproj, ctrl);
1908 _igvn.remove_dead_node(_callprojs->fallthrough_catchproj);
1909 }
1910 if (_callprojs->catchall_catchproj != nullptr) {
1911 _igvn.rehash_node_delayed(_callprojs->catchall_catchproj);
1912 _callprojs->catchall_catchproj->set_req(0, top());
1913 }
1914 if (_callprojs->fallthrough_proj != nullptr) {
1915 Node* catchnode = _callprojs->fallthrough_proj->unique_ctrl_out();
1916 _igvn.remove_dead_node(catchnode);
1917 _igvn.remove_dead_node(_callprojs->fallthrough_proj);
1918 }
1919 if (_callprojs->fallthrough_memproj != nullptr) {
1920 _igvn.replace_in_uses(_callprojs->fallthrough_memproj, mem);
1921 _igvn.remove_dead_node(_callprojs->fallthrough_memproj);
1922 }
1923 if (_callprojs->fallthrough_ioproj != nullptr) {
1924 _igvn.replace_in_uses(_callprojs->fallthrough_ioproj, i_o);
1925 _igvn.remove_dead_node(_callprojs->fallthrough_ioproj);
1926 }
1927 if (_callprojs->catchall_memproj != nullptr) {
1928 _igvn.rehash_node_delayed(_callprojs->catchall_memproj);
1929 _callprojs->catchall_memproj->set_req(0, top());
1930 }
1931 if (_callprojs->catchall_ioproj != nullptr) {
1932 _igvn.rehash_node_delayed(_callprojs->catchall_ioproj);
1933 _callprojs->catchall_ioproj->set_req(0, top());
1934 }
1935 #ifndef PRODUCT
1936 if (PrintEliminateAllocations) {
1937 if (alloc->is_AllocateArray()) {
1938 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1939 } else {
1940 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1941 }
1942 }
1943 #endif
1944 _igvn.remove_dead_node(alloc);
1945 }
1946
1947 void PhaseMacroExpand::expand_initialize_membar(AllocateNode* alloc, InitializeNode* init,
1948 Node*& fast_oop_ctrl, Node*& fast_oop_rawmem) {
1949 // If initialization is performed by an array copy, any required
1950 // MemBarStoreStore was already added. If the object does not
1951 // escape no need for a MemBarStoreStore. If the object does not
1952 // escape in its initializer and memory barrier (MemBarStoreStore or
1953 // stronger) is already added at exit of initializer, also no need
2047 Node* thread = new ThreadLocalNode();
2048 transform_later(thread);
2049
2050 call->init_req(TypeFunc::Parms + 0, thread);
2051 call->init_req(TypeFunc::Parms + 1, oop);
2052 call->init_req(TypeFunc::Control, ctrl);
2053 call->init_req(TypeFunc::I_O , top()); // does no i/o
2054 call->init_req(TypeFunc::Memory , rawmem);
2055 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
2056 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
2057 transform_later(call);
2058 ctrl = new ProjNode(call, TypeFunc::Control);
2059 transform_later(ctrl);
2060 rawmem = new ProjNode(call, TypeFunc::Memory);
2061 transform_later(rawmem);
2062 }
2063 }
2064
2065 // Helper for PhaseMacroExpand::expand_allocate_common.
2066 // Initializes the newly-allocated storage.
2067 Node* PhaseMacroExpand::initialize_object(AllocateNode* alloc,
2068 Node* control, Node* rawmem, Node* object,
2069 Node* klass_node, Node* length,
2070 Node* size_in_bytes) {
2071 InitializeNode* init = alloc->initialization();
2072 // Store the klass & mark bits
2073 Node* mark_node = alloc->make_ideal_mark(&_igvn, control, rawmem);
2074 if (!mark_node->is_Con()) {
2075 transform_later(mark_node);
2076 }
2077 rawmem = make_store_raw(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, TypeX_X->basic_type());
2078
2079 if (!UseCompactObjectHeaders) {
2080 rawmem = make_store_raw(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
2081 }
2082 int header_size = alloc->minimum_header_size(); // conservatively small
2083
2084 // Array length
2085 if (length != nullptr) { // Arrays need length field
2086 rawmem = make_store_raw(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
2087 // conservatively small header size:
2088 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
2089 if (_igvn.type(klass_node)->isa_aryklassptr()) { // we know the exact header size in most cases:
2090 BasicType elem = _igvn.type(klass_node)->is_klassptr()->as_instance_type()->isa_aryptr()->elem()->array_element_basic_type();
2091 if (is_reference_type(elem, true)) {
2092 elem = T_OBJECT;
2093 }
2094 header_size = Klass::layout_helper_header_size(Klass::array_layout_helper(elem));
2095 }
2096 }
2097
2098 // Clear the object body, if necessary.
2099 if (init == nullptr) {
2100 // The init has somehow disappeared; be cautious and clear everything.
2101 //
2102 // This can happen if a node is allocated but an uncommon trap occurs
2103 // immediately. In this case, the Initialize gets associated with the
2104 // trap, and may be placed in a different (outer) loop, if the Allocate
2105 // is in a loop. If (this is rare) the inner loop gets unrolled, then
2106 // there can be two Allocates to one Initialize. The answer in all these
2107 // edge cases is safety first. It is always safe to clear immediately
2108 // within an Allocate, and then (maybe or maybe not) clear some more later.
2109 if (!(UseTLAB && ZeroTLAB)) {
2110 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
2111 alloc->in(AllocateNode::InitValue),
2112 alloc->in(AllocateNode::RawInitValue),
2113 header_size, size_in_bytes,
2114 true,
2115 &_igvn);
2116 }
2117 } else {
2118 if (!init->is_complete()) {
2119 // Try to win by zeroing only what the init does not store.
2120 // We can also try to do some peephole optimizations,
2121 // such as combining some adjacent subword stores.
2122 rawmem = init->complete_stores(control, rawmem, object,
2123 header_size, size_in_bytes, &_igvn);
2124 }
2125 // We have no more use for this link, since the AllocateNode goes away:
2126 init->set_req(InitializeNode::RawAddress, top());
2127 // (If we keep the link, it just confuses the register allocator,
2128 // who thinks he sees a real use of the address by the membar.)
2129 }
2130
2131 return rawmem;
2132 }
2267 for (intx i = 0; i < lines; i++) {
2268 prefetch_adr = AddPNode::make_off_heap(new_eden_top,
2269 _igvn.MakeConX(distance));
2270 transform_later(prefetch_adr);
2271 prefetch = new PrefetchAllocationNode(i_o, prefetch_adr);
2272 // Do not let it float too high, since if eden_top == eden_end,
2273 // both might be null.
2274 if (i == 0) { // Set control for first prefetch, next follows it
2275 prefetch->init_req(0, needgc_false);
2276 }
2277 transform_later(prefetch);
2278 distance += step_size;
2279 i_o = prefetch;
2280 }
2281 }
2282 return i_o;
2283 }
2284
2285
2286 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
2287 expand_allocate_common(alloc, nullptr, nullptr,
2288 OptoRuntime::new_instance_Type(),
2289 OptoRuntime::new_instance_Java(), nullptr);
2290 }
2291
2292 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
2293 Node* length = alloc->in(AllocateNode::ALength);
2294 Node* valid_length_test = alloc->in(AllocateNode::ValidLengthTest);
2295 InitializeNode* init = alloc->initialization();
2296 Node* klass_node = alloc->in(AllocateNode::KlassNode);
2297 Node* init_value = alloc->in(AllocateNode::InitValue);
2298 const TypeAryKlassPtr* ary_klass_t = _igvn.type(klass_node)->isa_aryklassptr();
2299 assert(!ary_klass_t || !ary_klass_t->klass_is_exact() || !ary_klass_t->exact_klass()->is_obj_array_klass() ||
2300 ary_klass_t->is_refined_type(), "Must be a refined array klass");
2301 const TypeFunc* slow_call_type;
2302 address slow_call_address; // Address of slow call
2303 if (init != nullptr && init->is_complete_with_arraycopy() &&
2304 ary_klass_t && ary_klass_t->elem()->isa_klassptr() == nullptr) {
2305 // Don't zero type array during slow allocation in VM since
2306 // it will be initialized later by arraycopy in compiled code.
2307 slow_call_address = OptoRuntime::new_array_nozero_Java();
2308 slow_call_type = OptoRuntime::new_array_nozero_Type();
2309 } else {
2310 slow_call_address = OptoRuntime::new_array_Java();
2311 slow_call_type = OptoRuntime::new_array_Type();
2312
2313 if (init_value == nullptr) {
2314 init_value = _igvn.zerocon(T_OBJECT);
2315 } else if (UseCompressedOops) {
2316 init_value = transform_later(new DecodeNNode(init_value, init_value->bottom_type()->make_ptr()));
2317 }
2318 }
2319 expand_allocate_common(alloc, length, init_value,
2320 slow_call_type,
2321 slow_call_address, valid_length_test);
2322 }
2323
2324 //-------------------mark_eliminated_box----------------------------------
2325 //
2326 // During EA obj may point to several objects but after few ideal graph
2327 // transformations (CCP) it may point to only one non escaping object
2328 // (but still using phi), corresponding locks and unlocks will be marked
2329 // for elimination. Later obj could be replaced with a new node (new phi)
2330 // and which does not have escape information. And later after some graph
2331 // reshape other locks and unlocks (which were not marked for elimination
2332 // before) are connected to this new obj (phi) but they still will not be
2333 // marked for elimination since new obj has no escape information.
2334 // Mark all associated (same box and obj) lock and unlock nodes for
2335 // elimination if some of them marked already.
2336 void PhaseMacroExpand::mark_eliminated_box(Node* box, Node* obj) {
2337 BoxLockNode* oldbox = box->as_BoxLock();
2338 if (oldbox->is_eliminated()) {
2339 return; // This BoxLock node was processed already.
2340 }
2512 #ifdef ASSERT
2513 if (!alock->is_coarsened()) {
2514 // Check that new "eliminated" BoxLock node is created.
2515 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2516 assert(oldbox->is_eliminated(), "should be done already");
2517 }
2518 #endif
2519
2520 alock->log_lock_optimization(C, "eliminate_lock");
2521
2522 #ifndef PRODUCT
2523 if (PrintEliminateLocks) {
2524 tty->print_cr("++++ Eliminated: %d %s '%s'", alock->_idx, (alock->is_Lock() ? "Lock" : "Unlock"), alock->kind_as_string());
2525 }
2526 #endif
2527
2528 Node* mem = alock->in(TypeFunc::Memory);
2529 Node* ctrl = alock->in(TypeFunc::Control);
2530 guarantee(ctrl != nullptr, "missing control projection, cannot replace_node() with null");
2531
2532 _callprojs = alock->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2533 // There are 2 projections from the lock. The lock node will
2534 // be deleted when its last use is subsumed below.
2535 assert(alock->outcnt() == 2 &&
2536 _callprojs->fallthrough_proj != nullptr &&
2537 _callprojs->fallthrough_memproj != nullptr,
2538 "Unexpected projections from Lock/Unlock");
2539
2540 Node* fallthroughproj = _callprojs->fallthrough_proj;
2541 Node* memproj_fallthrough = _callprojs->fallthrough_memproj;
2542
2543 // The memory projection from a lock/unlock is RawMem
2544 // The input to a Lock is merged memory, so extract its RawMem input
2545 // (unless the MergeMem has been optimized away.)
2546 if (alock->is_Lock()) {
2547 // Search for MemBarAcquireLock node and delete it also.
2548 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2549 assert(membar != nullptr && membar->Opcode() == Op_MemBarAcquireLock, "");
2550 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2551 Node* memproj = membar->proj_out(TypeFunc::Memory);
2552 _igvn.replace_node(ctrlproj, fallthroughproj);
2553 _igvn.replace_node(memproj, memproj_fallthrough);
2554
2555 // Delete FastLock node also if this Lock node is unique user
2556 // (a loop peeling may clone a Lock node).
2557 Node* flock = alock->as_Lock()->fastlock_node();
2558 if (flock->outcnt() == 1) {
2559 assert(flock->unique_out() == alock, "sanity");
2560 _igvn.replace_node(flock, top());
2561 }
2592 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2593
2594 // Make the merge point
2595 Node *region;
2596 Node *mem_phi;
2597 Node *slow_path;
2598
2599 region = new RegionNode(3);
2600 // create a Phi for the memory state
2601 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2602
2603 // Optimize test; set region slot 2
2604 slow_path = opt_bits_test(ctrl, region, 2, flock);
2605 mem_phi->init_req(2, mem);
2606
2607 // Make slow path call
2608 CallNode* call = make_slow_call(lock, OptoRuntime::complete_monitor_enter_Type(),
2609 OptoRuntime::complete_monitor_locking_Java(), nullptr, slow_path,
2610 obj, box, nullptr);
2611
2612 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2613
2614 // Slow path can only throw asynchronous exceptions, which are always
2615 // de-opted. So the compiler thinks the slow-call can never throw an
2616 // exception. If it DOES throw an exception we would need the debug
2617 // info removed first (since if it throws there is no monitor).
2618 assert(_callprojs->fallthrough_ioproj == nullptr && _callprojs->catchall_ioproj == nullptr &&
2619 _callprojs->catchall_memproj == nullptr && _callprojs->catchall_catchproj == nullptr, "Unexpected projection from Lock");
2620
2621 // Capture slow path
2622 // disconnect fall-through projection from call and create a new one
2623 // hook up users of fall-through projection to region
2624 Node *slow_ctrl = _callprojs->fallthrough_proj->clone();
2625 transform_later(slow_ctrl);
2626 _igvn.hash_delete(_callprojs->fallthrough_proj);
2627 _callprojs->fallthrough_proj->disconnect_inputs(C);
2628 region->init_req(1, slow_ctrl);
2629 // region inputs are now complete
2630 transform_later(region);
2631 _igvn.replace_node(_callprojs->fallthrough_proj, region);
2632
2633 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2634
2635 mem_phi->init_req(1, memproj);
2636
2637 transform_later(mem_phi);
2638
2639 _igvn.replace_node(_callprojs->fallthrough_memproj, mem_phi);
2640 }
2641
2642 //------------------------------expand_unlock_node----------------------
2643 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2644
2645 Node* ctrl = unlock->in(TypeFunc::Control);
2646 Node* mem = unlock->in(TypeFunc::Memory);
2647 Node* obj = unlock->obj_node();
2648 Node* box = unlock->box_node();
2649
2650 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2651
2652 // No need for a null check on unlock
2653
2654 // Make the merge point
2655 Node* region = new RegionNode(3);
2656
2657 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2658 funlock = transform_later( funlock )->as_FastUnlock();
2659 // Optimize test; set region slot 2
2660 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock);
2661 Node *thread = transform_later(new ThreadLocalNode());
2662
2663 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2664 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2665 "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2666
2667 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2668 assert(_callprojs->fallthrough_ioproj == nullptr && _callprojs->catchall_ioproj == nullptr &&
2669 _callprojs->catchall_memproj == nullptr && _callprojs->catchall_catchproj == nullptr, "Unexpected projection from Lock");
2670
2671 // No exceptions for unlocking
2672 // Capture slow path
2673 // disconnect fall-through projection from call and create a new one
2674 // hook up users of fall-through projection to region
2675 Node *slow_ctrl = _callprojs->fallthrough_proj->clone();
2676 transform_later(slow_ctrl);
2677 _igvn.hash_delete(_callprojs->fallthrough_proj);
2678 _callprojs->fallthrough_proj->disconnect_inputs(C);
2679 region->init_req(1, slow_ctrl);
2680 // region inputs are now complete
2681 transform_later(region);
2682 _igvn.replace_node(_callprojs->fallthrough_proj, region);
2683
2684 if (_callprojs->fallthrough_memproj != nullptr) {
2685 // create a Phi for the memory state
2686 Node* mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2687 Node* memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2688 mem_phi->init_req(1, memproj);
2689 mem_phi->init_req(2, mem);
2690 transform_later(mem_phi);
2691 _igvn.replace_node(_callprojs->fallthrough_memproj, mem_phi);
2692 }
2693 }
2694
2695 // An inline type might be returned from the call but we don't know its
2696 // type. Either we get a buffered inline type (and nothing needs to be done)
2697 // or one of the values being returned is the klass of the inline type
2698 // and we need to allocate an inline type instance of that type and
2699 // initialize it with other values being returned. In that case, we
2700 // first try a fast path allocation and initialize the value with the
2701 // inline klass's pack handler or we fall back to a runtime call.
2702 void PhaseMacroExpand::expand_mh_intrinsic_return(CallStaticJavaNode* call) {
2703 assert(call->method()->is_method_handle_intrinsic(), "must be a method handle intrinsic call");
2704 Node* ret = call->proj_out_or_null(TypeFunc::Parms);
2705 if (ret == nullptr) {
2706 return;
2707 }
2708 const TypeFunc* tf = call->_tf;
2709 const TypeTuple* domain = OptoRuntime::store_inline_type_fields_Type()->domain_cc();
2710 const TypeFunc* new_tf = TypeFunc::make(tf->domain_sig(), tf->domain_cc(), tf->range_sig(), domain);
2711 call->_tf = new_tf;
2712 // Make sure the change of type is applied before projections are processed by igvn
2713 _igvn.set_type(call, call->Value(&_igvn));
2714 _igvn.set_type(ret, ret->Value(&_igvn));
2715
2716 // Before any new projection is added:
2717 CallProjections* projs = call->extract_projections(true, true);
2718
2719 // Create temporary hook nodes that will be replaced below.
2720 // Add an input to prevent hook nodes from being dead.
2721 Node* ctl = new Node(call);
2722 Node* mem = new Node(ctl);
2723 Node* io = new Node(ctl);
2724 Node* ex_ctl = new Node(ctl);
2725 Node* ex_mem = new Node(ctl);
2726 Node* ex_io = new Node(ctl);
2727 Node* res = new Node(ctl);
2728
2729 // Allocate a new buffered inline type only if a new one is not returned
2730 Node* cast = transform_later(new CastP2XNode(ctl, res));
2731 Node* mask = MakeConX(0x1);
2732 Node* masked = transform_later(new AndXNode(cast, mask));
2733 Node* cmp = transform_later(new CmpXNode(masked, mask));
2734 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
2735 IfNode* allocation_iff = new IfNode(ctl, bol, PROB_MAX, COUNT_UNKNOWN);
2736 transform_later(allocation_iff);
2737 Node* allocation_ctl = transform_later(new IfTrueNode(allocation_iff));
2738 Node* no_allocation_ctl = transform_later(new IfFalseNode(allocation_iff));
2739 Node* no_allocation_res = transform_later(new CheckCastPPNode(no_allocation_ctl, res, TypeInstPtr::BOTTOM));
2740
2741 // Try to allocate a new buffered inline instance either from TLAB or eden space
2742 Node* needgc_ctrl = nullptr; // needgc means slowcase, i.e. allocation failed
2743 CallLeafNoFPNode* handler_call;
2744 const bool alloc_in_place = UseTLAB;
2745 if (alloc_in_place) {
2746 Node* fast_oop_ctrl = nullptr;
2747 Node* fast_oop_rawmem = nullptr;
2748 Node* mask2 = MakeConX(-2);
2749 Node* masked2 = transform_later(new AndXNode(cast, mask2));
2750 Node* rawklassptr = transform_later(new CastX2PNode(masked2));
2751 Node* klass_node = transform_later(new CheckCastPPNode(allocation_ctl, rawklassptr, TypeInstKlassPtr::OBJECT_OR_NULL));
2752 Node* layout_val = make_load_raw(nullptr, mem, klass_node, in_bytes(Klass::layout_helper_offset()), TypeInt::INT, T_INT);
2753 Node* size_in_bytes = ConvI2X(layout_val);
2754 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2755 Node* fast_oop = bs->obj_allocate(this, mem, allocation_ctl, size_in_bytes, io, needgc_ctrl,
2756 fast_oop_ctrl, fast_oop_rawmem,
2757 AllocateInstancePrefetchLines);
2758 // Allocation succeed, initialize buffered inline instance header firstly,
2759 // and then initialize its fields with an inline class specific handler
2760 Node* mark_word_node;
2761 if (UseCompactObjectHeaders) {
2762 // COH: We need to load the prototype from the klass at runtime since it encodes the klass pointer already.
2763 mark_word_node = make_load_raw(fast_oop_ctrl, fast_oop_rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
2764 } else {
2765 // Otherwise, use the static prototype.
2766 mark_word_node = makecon(TypeRawPtr::make((address)markWord::inline_type_prototype().value()));
2767 }
2768
2769 fast_oop_rawmem = make_store_raw(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::mark_offset_in_bytes(), mark_word_node, T_ADDRESS);
2770 if (!UseCompactObjectHeaders) {
2771 // COH: Everything is encoded in the mark word, so nothing left to do.
2772 fast_oop_rawmem = make_store_raw(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
2773 if (UseCompressedClassPointers) {
2774 fast_oop_rawmem = make_store_raw(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::klass_gap_offset_in_bytes(), intcon(0), T_INT);
2775 }
2776 }
2777 Node* members = make_load_raw(fast_oop_ctrl, fast_oop_rawmem, klass_node, in_bytes(InlineKlass::adr_members_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
2778 Node* pack_handler = make_load_raw(fast_oop_ctrl, fast_oop_rawmem, members, in_bytes(InlineKlass::pack_handler_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
2779 handler_call = new CallLeafNoFPNode(OptoRuntime::pack_inline_type_Type(),
2780 nullptr,
2781 "pack handler",
2782 TypeRawPtr::BOTTOM);
2783 handler_call->init_req(TypeFunc::Control, fast_oop_ctrl);
2784 handler_call->init_req(TypeFunc::Memory, fast_oop_rawmem);
2785 handler_call->init_req(TypeFunc::I_O, top());
2786 handler_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2787 handler_call->init_req(TypeFunc::ReturnAdr, top());
2788 handler_call->init_req(TypeFunc::Parms, pack_handler);
2789 handler_call->init_req(TypeFunc::Parms+1, fast_oop);
2790 } else {
2791 needgc_ctrl = allocation_ctl;
2792 }
2793
2794 // Allocation failed, fall back to a runtime call
2795 CallStaticJavaNode* slow_call = new CallStaticJavaNode(OptoRuntime::store_inline_type_fields_Type(),
2796 StubRoutines::store_inline_type_fields_to_buf(),
2797 "store_inline_type_fields",
2798 TypePtr::BOTTOM);
2799 slow_call->init_req(TypeFunc::Control, needgc_ctrl);
2800 slow_call->init_req(TypeFunc::Memory, mem);
2801 slow_call->init_req(TypeFunc::I_O, io);
2802 slow_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2803 slow_call->init_req(TypeFunc::ReturnAdr, call->in(TypeFunc::ReturnAdr));
2804 slow_call->init_req(TypeFunc::Parms, res);
2805
2806 Node* slow_ctl = transform_later(new ProjNode(slow_call, TypeFunc::Control));
2807 Node* slow_mem = transform_later(new ProjNode(slow_call, TypeFunc::Memory));
2808 Node* slow_io = transform_later(new ProjNode(slow_call, TypeFunc::I_O));
2809 Node* slow_res = transform_later(new ProjNode(slow_call, TypeFunc::Parms));
2810 Node* slow_catc = transform_later(new CatchNode(slow_ctl, slow_io, 2));
2811 Node* slow_norm = transform_later(new CatchProjNode(slow_catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci));
2812 Node* slow_excp = transform_later(new CatchProjNode(slow_catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci));
2813
2814 Node* ex_r = new RegionNode(3);
2815 Node* ex_mem_phi = new PhiNode(ex_r, Type::MEMORY, TypePtr::BOTTOM);
2816 Node* ex_io_phi = new PhiNode(ex_r, Type::ABIO);
2817 ex_r->init_req(1, slow_excp);
2818 ex_mem_phi->init_req(1, slow_mem);
2819 ex_io_phi->init_req(1, slow_io);
2820 ex_r->init_req(2, ex_ctl);
2821 ex_mem_phi->init_req(2, ex_mem);
2822 ex_io_phi->init_req(2, ex_io);
2823 transform_later(ex_r);
2824 transform_later(ex_mem_phi);
2825 transform_later(ex_io_phi);
2826
2827 // We don't know how many values are returned. This assumes the
2828 // worst case, that all available registers are used.
2829 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) {
2830 if (domain->field_at(i) == Type::HALF) {
2831 slow_call->init_req(i, top());
2832 if (alloc_in_place) {
2833 handler_call->init_req(i+1, top());
2834 }
2835 continue;
2836 }
2837 Node* proj = transform_later(new ProjNode(call, i));
2838 slow_call->init_req(i, proj);
2839 if (alloc_in_place) {
2840 handler_call->init_req(i+1, proj);
2841 }
2842 }
2843 // We can safepoint at that new call
2844 slow_call->copy_call_debug_info(&_igvn, call);
2845 transform_later(slow_call);
2846 if (alloc_in_place) {
2847 transform_later(handler_call);
2848 }
2849
2850 Node* fast_ctl = nullptr;
2851 Node* fast_res = nullptr;
2852 MergeMemNode* fast_mem = nullptr;
2853 if (alloc_in_place) {
2854 fast_ctl = transform_later(new ProjNode(handler_call, TypeFunc::Control));
2855 Node* rawmem = transform_later(new ProjNode(handler_call, TypeFunc::Memory));
2856 fast_res = transform_later(new ProjNode(handler_call, TypeFunc::Parms));
2857 fast_mem = MergeMemNode::make(mem);
2858 fast_mem->set_memory_at(Compile::AliasIdxRaw, rawmem);
2859 transform_later(fast_mem);
2860 }
2861
2862 Node* r = new RegionNode(alloc_in_place ? 4 : 3);
2863 Node* mem_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM);
2864 Node* io_phi = new PhiNode(r, Type::ABIO);
2865 Node* res_phi = new PhiNode(r, TypeInstPtr::BOTTOM);
2866 r->init_req(1, no_allocation_ctl);
2867 mem_phi->init_req(1, mem);
2868 io_phi->init_req(1, io);
2869 res_phi->init_req(1, no_allocation_res);
2870 r->init_req(2, slow_norm);
2871 mem_phi->init_req(2, slow_mem);
2872 io_phi->init_req(2, slow_io);
2873 res_phi->init_req(2, slow_res);
2874 if (alloc_in_place) {
2875 r->init_req(3, fast_ctl);
2876 mem_phi->init_req(3, fast_mem);
2877 io_phi->init_req(3, io);
2878 res_phi->init_req(3, fast_res);
2879 }
2880 transform_later(r);
2881 transform_later(mem_phi);
2882 transform_later(io_phi);
2883 transform_later(res_phi);
2884
2885 // Do not let stores that initialize this buffer be reordered with a subsequent
2886 // store that would make this buffer accessible by other threads.
2887 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
2888 transform_later(mb);
2889 mb->init_req(TypeFunc::Memory, mem_phi);
2890 mb->init_req(TypeFunc::Control, r);
2891 r = new ProjNode(mb, TypeFunc::Control);
2892 transform_later(r);
2893 mem_phi = new ProjNode(mb, TypeFunc::Memory);
2894 transform_later(mem_phi);
2895
2896 assert(projs->nb_resproj == 1, "unexpected number of results");
2897 _igvn.replace_in_uses(projs->fallthrough_catchproj, r);
2898 _igvn.replace_in_uses(projs->fallthrough_memproj, mem_phi);
2899 _igvn.replace_in_uses(projs->fallthrough_ioproj, io_phi);
2900 _igvn.replace_in_uses(projs->resproj[0], res_phi);
2901 _igvn.replace_in_uses(projs->catchall_catchproj, ex_r);
2902 _igvn.replace_in_uses(projs->catchall_memproj, ex_mem_phi);
2903 _igvn.replace_in_uses(projs->catchall_ioproj, ex_io_phi);
2904 // The CatchNode should not use the ex_io_phi. Re-connect it to the catchall_ioproj.
2905 Node* cn = projs->fallthrough_catchproj->in(0);
2906 _igvn.replace_input_of(cn, 1, projs->catchall_ioproj);
2907
2908 _igvn.replace_node(ctl, projs->fallthrough_catchproj);
2909 _igvn.replace_node(mem, projs->fallthrough_memproj);
2910 _igvn.replace_node(io, projs->fallthrough_ioproj);
2911 _igvn.replace_node(res, projs->resproj[0]);
2912 _igvn.replace_node(ex_ctl, projs->catchall_catchproj);
2913 _igvn.replace_node(ex_mem, projs->catchall_memproj);
2914 _igvn.replace_node(ex_io, projs->catchall_ioproj);
2915 }
2916
2917 void PhaseMacroExpand::expand_subtypecheck_node(SubTypeCheckNode *check) {
2918 assert(check->in(SubTypeCheckNode::Control) == nullptr, "should be pinned");
2919 Node* bol = check->unique_out();
2920 Node* obj_or_subklass = check->in(SubTypeCheckNode::ObjOrSubKlass);
2921 Node* superklass = check->in(SubTypeCheckNode::SuperKlass);
2922 assert(bol->is_Bool() && bol->as_Bool()->_test._test == BoolTest::ne, "unexpected bool node");
2923
2924 for (DUIterator_Last imin, i = bol->last_outs(imin); i >= imin; --i) {
2925 Node* iff = bol->last_out(i);
2926 assert(iff->is_If(), "where's the if?");
2927
2928 if (iff->in(0)->is_top()) {
2929 _igvn.replace_input_of(iff, 1, C->top());
2930 continue;
2931 }
2932
2933 IfTrueNode* iftrue = iff->as_If()->true_proj();
2934 IfFalseNode* iffalse = iff->as_If()->false_proj();
2935 Node* ctrl = iff->in(0);
2936
2937 Node* subklass = nullptr;
2938 if (_igvn.type(obj_or_subklass)->isa_klassptr()) {
2939 subklass = obj_or_subklass;
2940 } else {
2941 Node* k_adr = basic_plus_adr(obj_or_subklass, oopDesc::klass_offset_in_bytes());
2942 subklass = _igvn.transform(LoadKlassNode::make(_igvn, C->immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
2943 }
2944
2945 Node* not_subtype_ctrl = Phase::gen_subtype_check(subklass, superklass, &ctrl, nullptr, _igvn, check->method(), check->bci());
2946
2947 _igvn.replace_input_of(iff, 0, C->top());
2948 _igvn.replace_node(iftrue, not_subtype_ctrl);
2949 _igvn.replace_node(iffalse, ctrl);
2950 }
2951 _igvn.replace_node(check, C->top());
2952 }
2953
2954 // FlatArrayCheckNode (array1 array2 ...) is expanded into:
2955 //
2956 // long mark = array1.mark | array2.mark | ...;
2957 // long locked_bit = markWord::unlocked_value & array1.mark & array2.mark & ...;
2958 // if (locked_bit == 0) {
2959 // // One array is locked, load prototype header from the klass
2960 // mark = array1.klass.proto | array2.klass.proto | ...
2961 // }
2962 // if ((mark & markWord::flat_array_bit_in_place) == 0) {
2963 // ...
2964 // }
2965 void PhaseMacroExpand::expand_flatarraycheck_node(FlatArrayCheckNode* check) {
2966 bool array_inputs = _igvn.type(check->in(FlatArrayCheckNode::ArrayOrKlass))->isa_oopptr() != nullptr;
2967 if (array_inputs) {
2968 Node* mark = MakeConX(0);
2969 Node* locked_bit = MakeConX(markWord::unlocked_value);
2970 Node* mem = check->in(FlatArrayCheckNode::Memory);
2971 for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
2972 Node* ary = check->in(i);
2973 const TypeOopPtr* t = _igvn.type(ary)->isa_oopptr();
2974 assert(t != nullptr, "Mixing array and klass inputs");
2975 assert(!t->is_flat() && !t->is_not_flat(), "Should have been optimized out");
2976 Node* mark_adr = basic_plus_adr(ary, oopDesc::mark_offset_in_bytes());
2977 Node* mark_load = _igvn.transform(LoadNode::make(_igvn, nullptr, mem, mark_adr, mark_adr->bottom_type()->is_ptr(), TypeX_X, TypeX_X->basic_type(), MemNode::unordered));
2978 mark = _igvn.transform(new OrXNode(mark, mark_load));
2979 locked_bit = _igvn.transform(new AndXNode(locked_bit, mark_load));
2980 }
2981 assert(!mark->is_Con(), "Should have been optimized out");
2982 Node* cmp = _igvn.transform(new CmpXNode(locked_bit, MakeConX(0)));
2983 Node* is_unlocked = _igvn.transform(new BoolNode(cmp, BoolTest::ne));
2984
2985 // BoolNode might be shared, replace each if user
2986 Node* old_bol = check->unique_out();
2987 assert(old_bol->is_Bool() && old_bol->as_Bool()->_test._test == BoolTest::ne, "unexpected condition");
2988 for (DUIterator_Last imin, i = old_bol->last_outs(imin); i >= imin; --i) {
2989 IfNode* old_iff = old_bol->last_out(i)->as_If();
2990 Node* ctrl = old_iff->in(0);
2991 RegionNode* region = new RegionNode(3);
2992 Node* mark_phi = new PhiNode(region, TypeX_X);
2993
2994 // Check if array is unlocked
2995 IfNode* iff = _igvn.transform(new IfNode(ctrl, is_unlocked, PROB_MAX, COUNT_UNKNOWN))->as_If();
2996
2997 // Unlocked: Use bits from mark word
2998 region->init_req(1, _igvn.transform(new IfTrueNode(iff)));
2999 mark_phi->init_req(1, mark);
3000
3001 // Locked: Load prototype header from klass
3002 ctrl = _igvn.transform(new IfFalseNode(iff));
3003 Node* proto = MakeConX(0);
3004 for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
3005 Node* ary = check->in(i);
3006 // Make loads control dependent to make sure they are only executed if array is locked
3007 Node* klass_adr = basic_plus_adr(ary, oopDesc::klass_offset_in_bytes());
3008 Node* klass = _igvn.transform(LoadKlassNode::make(_igvn, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
3009 Node* proto_adr = basic_plus_adr(top(), klass, in_bytes(Klass::prototype_header_offset()));
3010 Node* proto_load = _igvn.transform(LoadNode::make(_igvn, ctrl, C->immutable_memory(), proto_adr, proto_adr->bottom_type()->is_ptr(), TypeX_X, TypeX_X->basic_type(), MemNode::unordered));
3011 proto = _igvn.transform(new OrXNode(proto, proto_load));
3012 }
3013 region->init_req(2, ctrl);
3014 mark_phi->init_req(2, proto);
3015
3016 // Check if flat array bits are set
3017 Node* mask = MakeConX(markWord::flat_array_bit_in_place);
3018 Node* masked = _igvn.transform(new AndXNode(_igvn.transform(mark_phi), mask));
3019 cmp = _igvn.transform(new CmpXNode(masked, MakeConX(0)));
3020 Node* is_not_flat = _igvn.transform(new BoolNode(cmp, BoolTest::eq));
3021
3022 ctrl = _igvn.transform(region);
3023 iff = _igvn.transform(new IfNode(ctrl, is_not_flat, PROB_MAX, COUNT_UNKNOWN))->as_If();
3024 _igvn.replace_node(old_iff, iff);
3025 }
3026 _igvn.replace_node(check, C->top());
3027 } else {
3028 // Fall back to layout helper check
3029 Node* lhs = intcon(0);
3030 for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
3031 Node* array_or_klass = check->in(i);
3032 Node* klass = nullptr;
3033 const TypePtr* t = _igvn.type(array_or_klass)->is_ptr();
3034 assert(!t->is_flat() && !t->is_not_flat(), "Should have been optimized out");
3035 if (t->isa_oopptr() != nullptr) {
3036 Node* klass_adr = basic_plus_adr(array_or_klass, oopDesc::klass_offset_in_bytes());
3037 klass = transform_later(LoadKlassNode::make(_igvn, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
3038 } else {
3039 assert(t->isa_klassptr(), "Unexpected input type");
3040 klass = array_or_klass;
3041 }
3042 Node* lh_addr = basic_plus_adr(top(), klass, in_bytes(Klass::layout_helper_offset()));
3043 Node* lh_val = _igvn.transform(LoadNode::make(_igvn, nullptr, C->immutable_memory(), lh_addr, lh_addr->bottom_type()->is_ptr(), TypeInt::INT, T_INT, MemNode::unordered));
3044 lhs = _igvn.transform(new OrINode(lhs, lh_val));
3045 }
3046 Node* masked = transform_later(new AndINode(lhs, intcon(Klass::_lh_array_tag_flat_value_bit_inplace)));
3047 Node* cmp = transform_later(new CmpINode(masked, intcon(0)));
3048 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
3049 Node* m2b = transform_later(new Conv2BNode(masked));
3050 // The matcher expects the input to If/CMove nodes to be produced by a Bool(CmpI..)
3051 // pattern, but the input to other potential users (e.g. Phi) to be some
3052 // other pattern (e.g. a Conv2B node, possibly idealized as a CMoveI).
3053 Node* old_bol = check->unique_out();
3054 for (DUIterator_Last imin, i = old_bol->last_outs(imin); i >= imin; --i) {
3055 Node* user = old_bol->last_out(i);
3056 for (uint j = 0; j < user->req(); j++) {
3057 Node* n = user->in(j);
3058 if (n == old_bol) {
3059 _igvn.replace_input_of(user, j, (user->is_If() || user->is_CMove()) ? bol : m2b);
3060 }
3061 }
3062 }
3063 _igvn.replace_node(check, C->top());
3064 }
3065 }
3066
3067 // Perform refining of strip mined loop nodes in the macro nodes list.
3068 void PhaseMacroExpand::refine_strip_mined_loop_macro_nodes() {
3069 for (int i = C->macro_count(); i > 0; i--) {
3070 Node* n = C->macro_node(i - 1);
3071 if (n->is_OuterStripMinedLoop()) {
3072 n->as_OuterStripMinedLoop()->adjust_strip_mined_loop(&_igvn);
3073 }
3074 }
3075 }
3076
3077 //---------------------------eliminate_macro_nodes----------------------
3078 // Eliminate scalar replaced allocations and associated locks.
3079 void PhaseMacroExpand::eliminate_macro_nodes(bool eliminate_locks) {
3080 if (C->macro_count() == 0) {
3081 return;
3082 }
3083
3084 if (StressMacroElimination) {
3085 C->shuffle_macro_nodes();
3086 }
3087 NOT_PRODUCT(int membar_before = count_MemBar(C);)
3088
3089 int iteration = 0;
3090 while (C->macro_count() > 0) {
3091 if (iteration++ > 100) {
3092 assert(false, "Too slow convergence of macro elimination");
3093 break;
3094 }
3095
3096 // Postpone lock elimination to after EA when most allocations are eliminated
3097 // because they might block lock elimination if their escape state isn't
3098 // determined yet and we only got one chance at eliminating the lock.
3099 if (eliminate_locks) {
3100 // Before elimination may re-mark (change to Nested or NonEscObj)
3101 // all associated (same box and obj) lock and unlock nodes.
3102 int cnt = C->macro_count();
3103 for (int i=0; i < cnt; i++) {
3104 Node *n = C->macro_node(i);
3105 if (n->is_AbstractLock()) { // Lock and Unlock nodes
3106 mark_eliminated_locking_nodes(n->as_AbstractLock());
3107 }
3108 }
3109 // Re-marking may break consistency of Coarsened locks.
3110 if (!C->coarsened_locks_consistent()) {
3111 return; // recompile without Coarsened locks if broken
3112 } else {
3113 // After coarsened locks are eliminated locking regions
3114 // become unbalanced. We should not execute any more
3115 // locks elimination optimizations on them.
3116 C->mark_unbalanced_boxes();
3117 }
3118 }
3119
3120 bool progress = false;
3121 for (int i = C->macro_count(); i > 0; i = MIN2(i - 1, C->macro_count())) { // more than 1 element can be eliminated at once
3122 Node* n = C->macro_node(i - 1);
3123 bool success = false;
3124 DEBUG_ONLY(int old_macro_count = C->macro_count();)
3125 switch (n->class_id()) {
3126 case Node::Class_Allocate:
3127 case Node::Class_AllocateArray:
3128 success = eliminate_allocate_node(n->as_Allocate());
3129 #ifndef PRODUCT
3130 if (success && PrintOptoStatistics) {
3131 AtomicAccess::inc(&PhaseMacroExpand::_objs_scalar_replaced_counter);
3132 }
3133 #endif
3134 break;
3135 case Node::Class_CallStaticJava: {
3136 CallStaticJavaNode* call = n->as_CallStaticJava();
3137 if (!call->method()->is_method_handle_intrinsic()) {
3138 success = eliminate_boxing_node(n->as_CallStaticJava());
3139 }
3140 break;
3141 }
3142 case Node::Class_Lock:
3143 case Node::Class_Unlock:
3144 if (eliminate_locks) {
3145 success = eliminate_locking_node(n->as_AbstractLock());
3146 #ifndef PRODUCT
3147 if (success && PrintOptoStatistics) {
3148 AtomicAccess::inc(&PhaseMacroExpand::_monitor_objects_removed_counter);
3149 }
3150 #endif
3151 }
3152 break;
3153 case Node::Class_ArrayCopy:
3154 break;
3155 case Node::Class_OuterStripMinedLoop:
3156 break;
3157 case Node::Class_SubTypeCheck:
3158 break;
3159 case Node::Class_Opaque1:
3160 break;
3161 case Node::Class_FlatArrayCheck:
3162 break;
3163 default:
3164 assert(n->Opcode() == Op_LoopLimit ||
3165 n->Opcode() == Op_ModD ||
3166 n->Opcode() == Op_ModF ||
3167 n->Opcode() == Op_PowD ||
3168 n->is_OpaqueConstantBool() ||
3169 n->is_OpaqueInitializedAssertionPredicate() ||
3170 n->Opcode() == Op_MaxL ||
3171 n->Opcode() == Op_MinL ||
3172 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n),
3173 "unknown node type in macro list");
3174 }
3175 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
3176 progress = progress || success;
3177 if (success) {
3178 C->print_method(PHASE_AFTER_MACRO_ELIMINATION_STEP, 5, n);
3179 }
3180 }
3181
3182 // Ensure the graph after PhaseMacroExpand::eliminate_macro_nodes is canonical (no igvn
3183 // transformation is pending). If an allocation is used only in safepoints, elimination of
3184 // other macro nodes can remove all these safepoints, allowing the allocation to be removed.
3185 // Hence after igvn we retry removing macro nodes if some progress that has been made in this
3186 // iteration.
3187 _igvn.set_delay_transform(false);
3188 _igvn.optimize();
3189 if (C->failing()) {
3190 return;
3191 }
3192 _igvn.set_delay_transform(true);
3193
3194 if (!progress) {
3195 break;
3196 }
3197 }
3198 #ifndef PRODUCT
3199 if (PrintOptoStatistics) {
3200 int membar_after = count_MemBar(C);
3201 AtomicAccess::add(&PhaseMacroExpand::_memory_barriers_removed_counter, membar_before - membar_after);
3202 }
3203 #endif
3204 }
3205
3206 void PhaseMacroExpand::eliminate_opaque_looplimit_macro_nodes() {
3207 if (C->macro_count() == 0) {
3208 return;
3209 }
3210 refine_strip_mined_loop_macro_nodes();
3211 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
3212 bool progress = true;
3213 while (progress) {
3214 progress = false;
3215 for (int i = C->macro_count(); i > 0; i--) {
3216 Node* n = C->macro_node(i-1);
3217 bool success = false;
3218 DEBUG_ONLY(int old_macro_count = C->macro_count();)
3219 if (n->Opcode() == Op_LoopLimit) {
3220 // Remove it from macro list and put on IGVN worklist to optimize.
3221 C->remove_macro_node(n);
3222 _igvn._worklist.push(n);
3223 success = true;
3224 } else if (n->Opcode() == Op_CallStaticJava) {
3225 CallStaticJavaNode* call = n->as_CallStaticJava();
3226 if (!call->method()->is_method_handle_intrinsic()) {
3227 // Remove it from macro list and put on IGVN worklist to optimize.
3228 C->remove_macro_node(n);
3229 _igvn._worklist.push(n);
3230 success = true;
3231 }
3232 } else if (n->is_Opaque1()) {
3233 _igvn.replace_node(n, n->in(1));
3234 success = true;
3235 } else if (n->is_OpaqueConstantBool()) {
3236 // Tests with OpaqueConstantBool nodes are implicitly known. Replace the node with true/false. In debug builds,
3237 // we leave the test in the graph to have an additional sanity check at runtime. If the test fails (i.e. a bug),
3238 // we will execute a Halt node.
3239 #ifdef ASSERT
3240 _igvn.replace_node(n, n->in(1));
3241 #else
3242 _igvn.replace_node(n, _igvn.intcon(n->as_OpaqueConstantBool()->constant()));
3243 #endif
3244 success = true;
3245 } else if (n->is_OpaqueInitializedAssertionPredicate()) {
3246 // Initialized Assertion Predicates must always evaluate to true. Therefore, we get rid of them in product
3247 // builds as they are useless. In debug builds we keep them as additional verification code. Even though
3248 // loop opts are already over, we want to keep Initialized Assertion Predicates alive as long as possible to
3249 // enable folding of dead control paths within which cast nodes become top after due to impossible types -
3250 // even after loop opts are over. Therefore, we delay the removal of these opaque nodes until now.
3251 #ifdef ASSERT
3320 // Worst case is a macro node gets expanded into about 200 nodes.
3321 // Allow 50% more for optimization.
3322 if (C->check_node_count(300, "out of nodes before macro expansion")) {
3323 return true;
3324 }
3325
3326 DEBUG_ONLY(int old_macro_count = C->macro_count();)
3327 switch (n->class_id()) {
3328 case Node::Class_Lock:
3329 expand_lock_node(n->as_Lock());
3330 break;
3331 case Node::Class_Unlock:
3332 expand_unlock_node(n->as_Unlock());
3333 break;
3334 case Node::Class_ArrayCopy:
3335 expand_arraycopy_node(n->as_ArrayCopy());
3336 break;
3337 case Node::Class_SubTypeCheck:
3338 expand_subtypecheck_node(n->as_SubTypeCheck());
3339 break;
3340 case Node::Class_CallStaticJava:
3341 expand_mh_intrinsic_return(n->as_CallStaticJava());
3342 C->remove_macro_node(n);
3343 break;
3344 case Node::Class_FlatArrayCheck:
3345 expand_flatarraycheck_node(n->as_FlatArrayCheck());
3346 break;
3347 default:
3348 switch (n->Opcode()) {
3349 case Op_ModD:
3350 case Op_ModF:
3351 case Op_PowD: {
3352 CallLeafPureNode* call_macro = n->as_CallLeafPure();
3353 CallLeafPureNode* call = call_macro->inline_call_leaf_pure_node();
3354 _igvn.replace_node(call_macro, call);
3355 transform_later(call);
3356 break;
3357 }
3358 default:
3359 assert(false, "unknown node type in macro list");
3360 }
3361 }
3362 assert(C->macro_count() == (old_macro_count - 1), "expansion must have deleted one node from macro list");
3363 if (C->failing()) return true;
3364 C->print_method(PHASE_AFTER_MACRO_EXPANSION_STEP, 5, n);
3365
3366 // Clean up the graph so we're less likely to hit the maximum node
|