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(new AddPNode(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(new AddPNode(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(new AddPNode(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.
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;
1336 #endif
1337 yank_alloc_node(alloc);
1338 return;
1339 }
1340 }
1341
1342 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1343 Node *slow_region = nullptr;
1344 Node *toobig_false = ctrl;
1345
1346 // generate the initial test if necessary
1347 if (initial_slow_test != nullptr ) {
1348 assert (expand_fast_path, "Only need test if there is a fast path");
1349 slow_region = new RegionNode(3);
1350
1351 // Now make the initial failure test. Usually a too-big test but
1352 // might be a TRUE for finalizers.
1353 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1354 transform_later(toobig_iff);
1355 // Plug the failing-too-big test into the slow-path region
1356 Node *toobig_true = new IfTrueNode( toobig_iff );
1357 transform_later(toobig_true);
1358 slow_region ->init_req( too_big_or_final_path, toobig_true );
1359 toobig_false = new IfFalseNode( toobig_iff );
1360 transform_later(toobig_false);
1361 } else {
1362 // No initial test, just fall into next case
1363 assert(allocation_has_use || !expand_fast_path, "Should already have been handled");
1364 toobig_false = ctrl;
1365 DEBUG_ONLY(slow_region = NodeSentinel);
1366 }
1367
1368 // If we are here there are several possibilities
1369 // - expand_fast_path is false - then only a slow path is expanded. That's it.
1370 // no_initial_check means a constant allocation.
1371 // - If check always evaluates to false -> expand_fast_path is false (see above)
1372 // - If check always evaluates to true -> directly into fast path (but may bailout to slowpath)
1373 // if !allocation_has_use the fast path is empty
1374 // if !allocation_has_use && no_initial_check
1375 // - Then there are no fastpath that can fall out to slowpath -> no allocation code at all.
1376 // removed by yank_alloc_node above.
1377
1378 Node *slow_mem = mem; // save the current memory state for slow path
1379 // generate the fast allocation code unless we know that the initial test will always go slow
1380 if (expand_fast_path) {
1381 // Fast path modifies only raw memory.
1382 if (mem->is_MergeMem()) {
1383 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1384 }
1385
1386 // allocate the Region and Phi nodes for the result
1387 result_region = new RegionNode(3);
1388 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1389 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1390
1391 // Grab regular I/O before optional prefetch may change it.
1392 // Slow-path does no I/O so just set it to the original I/O.
1393 result_phi_i_o->init_req(slow_result_path, i_o);
1394
1395 // Name successful fast-path variables
1396 Node* fast_oop_ctrl;
1397 Node* fast_oop_rawmem;
1398 if (allocation_has_use) {
1399 Node* needgc_ctrl = nullptr;
1400 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1401
1402 intx prefetch_lines = length != nullptr ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1403 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1404 Node* fast_oop = bs->obj_allocate(this, mem, toobig_false, size_in_bytes, i_o, needgc_ctrl,
1405 fast_oop_ctrl, fast_oop_rawmem,
1406 prefetch_lines);
1407
1408 if (initial_slow_test != nullptr) {
1409 // This completes all paths into the slow merge point
1410 slow_region->init_req(need_gc_path, needgc_ctrl);
1411 transform_later(slow_region);
1412 } else {
1413 // No initial slow path needed!
1414 // Just fall from the need-GC path straight into the VM call.
1415 slow_region = needgc_ctrl;
1416 }
1417
1435 result_phi_i_o ->init_req(fast_result_path, i_o);
1436 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1437 } else {
1438 slow_region = ctrl;
1439 result_phi_i_o = i_o; // Rename it to use in the following code.
1440 }
1441
1442 // Generate slow-path call
1443 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1444 OptoRuntime::stub_name(slow_call_address),
1445 TypePtr::BOTTOM);
1446 call->init_req(TypeFunc::Control, slow_region);
1447 call->init_req(TypeFunc::I_O, top()); // does no i/o
1448 call->init_req(TypeFunc::Memory, slow_mem); // may gc ptrs
1449 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1450 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1451
1452 call->init_req(TypeFunc::Parms+0, klass_node);
1453 if (length != nullptr) {
1454 call->init_req(TypeFunc::Parms+1, length);
1455 }
1456
1457 // Copy debug information and adjust JVMState information, then replace
1458 // allocate node with the call
1459 call->copy_call_debug_info(&_igvn, alloc);
1460 // For array allocations, copy the valid length check to the call node so Compile::final_graph_reshaping() can verify
1461 // that the call has the expected number of CatchProj nodes (in case the allocation always fails and the fallthrough
1462 // path dies).
1463 if (valid_length_test != nullptr) {
1464 call->add_req(valid_length_test);
1465 }
1466 if (expand_fast_path) {
1467 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1468 } else {
1469 // Hook i_o projection to avoid its elimination during allocation
1470 // replacement (when only a slow call is generated).
1471 call->set_req(TypeFunc::I_O, result_phi_i_o);
1472 }
1473 _igvn.replace_node(alloc, call);
1474 transform_later(call);
1475
1476 // Identify the output projections from the allocate node and
1477 // adjust any references to them.
1478 // The control and io projections look like:
1479 //
1480 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1481 // Allocate Catch
1482 // ^---Proj(io) <-------+ ^---CatchProj(io)
1483 //
1484 // We are interested in the CatchProj nodes.
1485 //
1486 call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1487
1488 // An allocate node has separate memory projections for the uses on
1489 // the control and i_o paths. Replace the control memory projection with
1490 // result_phi_rawmem (unless we are only generating a slow call when
1491 // both memory projections are combined)
1492 if (expand_fast_path && _callprojs.fallthrough_memproj != nullptr) {
1493 migrate_outs(_callprojs.fallthrough_memproj, result_phi_rawmem);
1494 }
1495 // Now change uses of catchall_memproj to use fallthrough_memproj and delete
1496 // catchall_memproj so we end up with a call that has only 1 memory projection.
1497 if (_callprojs.catchall_memproj != nullptr ) {
1498 if (_callprojs.fallthrough_memproj == nullptr) {
1499 _callprojs.fallthrough_memproj = new ProjNode(call, TypeFunc::Memory);
1500 transform_later(_callprojs.fallthrough_memproj);
1501 }
1502 migrate_outs(_callprojs.catchall_memproj, _callprojs.fallthrough_memproj);
1503 _igvn.remove_dead_node(_callprojs.catchall_memproj);
1504 }
1505
1506 // An allocate node has separate i_o projections for the uses on the control
1507 // and i_o paths. Always replace the control i_o projection with result i_o
1508 // otherwise incoming i_o become dead when only a slow call is generated
1509 // (it is different from memory projections where both projections are
1510 // combined in such case).
1511 if (_callprojs.fallthrough_ioproj != nullptr) {
1512 migrate_outs(_callprojs.fallthrough_ioproj, result_phi_i_o);
1513 }
1514 // Now change uses of catchall_ioproj to use fallthrough_ioproj and delete
1515 // catchall_ioproj so we end up with a call that has only 1 i_o projection.
1516 if (_callprojs.catchall_ioproj != nullptr ) {
1517 if (_callprojs.fallthrough_ioproj == nullptr) {
1518 _callprojs.fallthrough_ioproj = new ProjNode(call, TypeFunc::I_O);
1519 transform_later(_callprojs.fallthrough_ioproj);
1520 }
1521 migrate_outs(_callprojs.catchall_ioproj, _callprojs.fallthrough_ioproj);
1522 _igvn.remove_dead_node(_callprojs.catchall_ioproj);
1523 }
1524
1525 // if we generated only a slow call, we are done
1526 if (!expand_fast_path) {
1527 // Now we can unhook i_o.
1528 if (result_phi_i_o->outcnt() > 1) {
1529 call->set_req(TypeFunc::I_O, top());
1530 } else {
1531 assert(result_phi_i_o->unique_ctrl_out() == call, "sanity");
1532 // Case of new array with negative size known during compilation.
1533 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1534 // following code since call to runtime will throw exception.
1535 // As result there will be no users of i_o after the call.
1536 // Leave i_o attached to this call to avoid problems in preceding graph.
1537 }
1538 return;
1539 }
1540
1541 if (_callprojs.fallthrough_catchproj != nullptr) {
1542 ctrl = _callprojs.fallthrough_catchproj->clone();
1543 transform_later(ctrl);
1544 _igvn.replace_node(_callprojs.fallthrough_catchproj, result_region);
1545 } else {
1546 ctrl = top();
1547 }
1548 Node *slow_result;
1549 if (_callprojs.resproj == nullptr) {
1550 // no uses of the allocation result
1551 slow_result = top();
1552 } else {
1553 slow_result = _callprojs.resproj->clone();
1554 transform_later(slow_result);
1555 _igvn.replace_node(_callprojs.resproj, result_phi_rawoop);
1556 }
1557
1558 // Plug slow-path into result merge point
1559 result_region->init_req( slow_result_path, ctrl);
1560 transform_later(result_region);
1561 if (allocation_has_use) {
1562 result_phi_rawoop->init_req(slow_result_path, slow_result);
1563 transform_later(result_phi_rawoop);
1564 }
1565 result_phi_rawmem->init_req(slow_result_path, _callprojs.fallthrough_memproj);
1566 transform_later(result_phi_rawmem);
1567 transform_later(result_phi_i_o);
1568 // This completes all paths into the result merge point
1569 }
1570
1571 // Remove alloc node that has no uses.
1572 void PhaseMacroExpand::yank_alloc_node(AllocateNode* alloc) {
1573 Node* ctrl = alloc->in(TypeFunc::Control);
1574 Node* mem = alloc->in(TypeFunc::Memory);
1575 Node* i_o = alloc->in(TypeFunc::I_O);
1576
1577 alloc->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1578 if (_callprojs.resproj != nullptr) {
1579 for (DUIterator_Fast imax, i = _callprojs.resproj->fast_outs(imax); i < imax; i++) {
1580 Node* use = _callprojs.resproj->fast_out(i);
1581 use->isa_MemBar()->remove(&_igvn);
1582 --imax;
1583 --i; // back up iterator
1584 }
1585 assert(_callprojs.resproj->outcnt() == 0, "all uses must be deleted");
1586 _igvn.remove_dead_node(_callprojs.resproj);
1587 }
1588 if (_callprojs.fallthrough_catchproj != nullptr) {
1589 migrate_outs(_callprojs.fallthrough_catchproj, ctrl);
1590 _igvn.remove_dead_node(_callprojs.fallthrough_catchproj);
1591 }
1592 if (_callprojs.catchall_catchproj != nullptr) {
1593 _igvn.rehash_node_delayed(_callprojs.catchall_catchproj);
1594 _callprojs.catchall_catchproj->set_req(0, top());
1595 }
1596 if (_callprojs.fallthrough_proj != nullptr) {
1597 Node* catchnode = _callprojs.fallthrough_proj->unique_ctrl_out();
1598 _igvn.remove_dead_node(catchnode);
1599 _igvn.remove_dead_node(_callprojs.fallthrough_proj);
1600 }
1601 if (_callprojs.fallthrough_memproj != nullptr) {
1602 migrate_outs(_callprojs.fallthrough_memproj, mem);
1603 _igvn.remove_dead_node(_callprojs.fallthrough_memproj);
1604 }
1605 if (_callprojs.fallthrough_ioproj != nullptr) {
1606 migrate_outs(_callprojs.fallthrough_ioproj, i_o);
1607 _igvn.remove_dead_node(_callprojs.fallthrough_ioproj);
1608 }
1609 if (_callprojs.catchall_memproj != nullptr) {
1610 _igvn.rehash_node_delayed(_callprojs.catchall_memproj);
1611 _callprojs.catchall_memproj->set_req(0, top());
1612 }
1613 if (_callprojs.catchall_ioproj != nullptr) {
1614 _igvn.rehash_node_delayed(_callprojs.catchall_ioproj);
1615 _callprojs.catchall_ioproj->set_req(0, top());
1616 }
1617 #ifndef PRODUCT
1618 if (PrintEliminateAllocations) {
1619 if (alloc->is_AllocateArray()) {
1620 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1621 } else {
1622 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1623 }
1624 }
1625 #endif
1626 _igvn.remove_dead_node(alloc);
1627 }
1628
1629 void PhaseMacroExpand::expand_initialize_membar(AllocateNode* alloc, InitializeNode* init,
1630 Node*& fast_oop_ctrl, Node*& fast_oop_rawmem) {
1631 // If initialization is performed by an array copy, any required
1632 // MemBarStoreStore was already added. If the object does not
1633 // escape no need for a MemBarStoreStore. If the object does not
1634 // escape in its initializer and memory barrier (MemBarStoreStore or
1635 // stronger) is already added at exit of initializer, also no need
1729 Node* thread = new ThreadLocalNode();
1730 transform_later(thread);
1731
1732 call->init_req(TypeFunc::Parms + 0, thread);
1733 call->init_req(TypeFunc::Parms + 1, oop);
1734 call->init_req(TypeFunc::Control, ctrl);
1735 call->init_req(TypeFunc::I_O , top()); // does no i/o
1736 call->init_req(TypeFunc::Memory , rawmem);
1737 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1738 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1739 transform_later(call);
1740 ctrl = new ProjNode(call, TypeFunc::Control);
1741 transform_later(ctrl);
1742 rawmem = new ProjNode(call, TypeFunc::Memory);
1743 transform_later(rawmem);
1744 }
1745 }
1746
1747 // Helper for PhaseMacroExpand::expand_allocate_common.
1748 // Initializes the newly-allocated storage.
1749 Node*
1750 PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1751 Node* control, Node* rawmem, Node* object,
1752 Node* klass_node, Node* length,
1753 Node* size_in_bytes) {
1754 InitializeNode* init = alloc->initialization();
1755 // Store the klass & mark bits
1756 Node* mark_node = alloc->make_ideal_mark(&_igvn, object, control, rawmem);
1757 if (!mark_node->is_Con()) {
1758 transform_later(mark_node);
1759 }
1760 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, TypeX_X->basic_type());
1761
1762 if (!UseCompactObjectHeaders) {
1763 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1764 }
1765 int header_size = alloc->minimum_header_size(); // conservatively small
1766
1767 // Array length
1768 if (length != nullptr) { // Arrays need length field
1769 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1770 // conservatively small header size:
1771 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1772 if (_igvn.type(klass_node)->isa_aryklassptr()) { // we know the exact header size in most cases:
1773 BasicType elem = _igvn.type(klass_node)->is_klassptr()->as_instance_type()->isa_aryptr()->elem()->array_element_basic_type();
1774 if (is_reference_type(elem, true)) {
1775 elem = T_OBJECT;
1776 }
1777 header_size = Klass::layout_helper_header_size(Klass::array_layout_helper(elem));
1778 }
1779 }
1780
1781 // Clear the object body, if necessary.
1782 if (init == nullptr) {
1783 // The init has somehow disappeared; be cautious and clear everything.
1784 //
1785 // This can happen if a node is allocated but an uncommon trap occurs
1786 // immediately. In this case, the Initialize gets associated with the
1787 // trap, and may be placed in a different (outer) loop, if the Allocate
1788 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1789 // there can be two Allocates to one Initialize. The answer in all these
1790 // edge cases is safety first. It is always safe to clear immediately
1791 // within an Allocate, and then (maybe or maybe not) clear some more later.
1792 if (!(UseTLAB && ZeroTLAB)) {
1793 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1794 header_size, size_in_bytes,
1795 &_igvn);
1796 }
1797 } else {
1798 if (!init->is_complete()) {
1799 // Try to win by zeroing only what the init does not store.
1800 // We can also try to do some peephole optimizations,
1801 // such as combining some adjacent subword stores.
1802 rawmem = init->complete_stores(control, rawmem, object,
1803 header_size, size_in_bytes, &_igvn);
1804 }
1805 // We have no more use for this link, since the AllocateNode goes away:
1806 init->set_req(InitializeNode::RawAddress, top());
1807 // (If we keep the link, it just confuses the register allocator,
1808 // who thinks he sees a real use of the address by the membar.)
1809 }
1810
1811 return rawmem;
1812 }
1813
1948 for ( intx i = 0; i < lines; i++ ) {
1949 prefetch_adr = new AddPNode( old_eden_top, new_eden_top,
1950 _igvn.MakeConX(distance) );
1951 transform_later(prefetch_adr);
1952 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1953 // Do not let it float too high, since if eden_top == eden_end,
1954 // both might be null.
1955 if( i == 0 ) { // Set control for first prefetch, next follows it
1956 prefetch->init_req(0, needgc_false);
1957 }
1958 transform_later(prefetch);
1959 distance += step_size;
1960 i_o = prefetch;
1961 }
1962 }
1963 return i_o;
1964 }
1965
1966
1967 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
1968 expand_allocate_common(alloc, nullptr,
1969 OptoRuntime::new_instance_Type(),
1970 OptoRuntime::new_instance_Java(), nullptr);
1971 }
1972
1973 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
1974 Node* length = alloc->in(AllocateNode::ALength);
1975 Node* valid_length_test = alloc->in(AllocateNode::ValidLengthTest);
1976 InitializeNode* init = alloc->initialization();
1977 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1978 const TypeAryKlassPtr* ary_klass_t = _igvn.type(klass_node)->isa_aryklassptr();
1979 address slow_call_address; // Address of slow call
1980 if (init != nullptr && init->is_complete_with_arraycopy() &&
1981 ary_klass_t && ary_klass_t->elem()->isa_klassptr() == nullptr) {
1982 // Don't zero type array during slow allocation in VM since
1983 // it will be initialized later by arraycopy in compiled code.
1984 slow_call_address = OptoRuntime::new_array_nozero_Java();
1985 } else {
1986 slow_call_address = OptoRuntime::new_array_Java();
1987 }
1988 expand_allocate_common(alloc, length,
1989 OptoRuntime::new_array_Type(),
1990 slow_call_address, valid_length_test);
1991 }
1992
1993 //-------------------mark_eliminated_box----------------------------------
1994 //
1995 // During EA obj may point to several objects but after few ideal graph
1996 // transformations (CCP) it may point to only one non escaping object
1997 // (but still using phi), corresponding locks and unlocks will be marked
1998 // for elimination. Later obj could be replaced with a new node (new phi)
1999 // and which does not have escape information. And later after some graph
2000 // reshape other locks and unlocks (which were not marked for elimination
2001 // before) are connected to this new obj (phi) but they still will not be
2002 // marked for elimination since new obj has no escape information.
2003 // Mark all associated (same box and obj) lock and unlock nodes for
2004 // elimination if some of them marked already.
2005 void PhaseMacroExpand::mark_eliminated_box(Node* box, Node* obj) {
2006 BoxLockNode* oldbox = box->as_BoxLock();
2007 if (oldbox->is_eliminated()) {
2008 return; // This BoxLock node was processed already.
2009 }
2181 #ifdef ASSERT
2182 if (!alock->is_coarsened()) {
2183 // Check that new "eliminated" BoxLock node is created.
2184 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2185 assert(oldbox->is_eliminated(), "should be done already");
2186 }
2187 #endif
2188
2189 alock->log_lock_optimization(C, "eliminate_lock");
2190
2191 #ifndef PRODUCT
2192 if (PrintEliminateLocks) {
2193 tty->print_cr("++++ Eliminated: %d %s '%s'", alock->_idx, (alock->is_Lock() ? "Lock" : "Unlock"), alock->kind_as_string());
2194 }
2195 #endif
2196
2197 Node* mem = alock->in(TypeFunc::Memory);
2198 Node* ctrl = alock->in(TypeFunc::Control);
2199 guarantee(ctrl != nullptr, "missing control projection, cannot replace_node() with null");
2200
2201 alock->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
2202 // There are 2 projections from the lock. The lock node will
2203 // be deleted when its last use is subsumed below.
2204 assert(alock->outcnt() == 2 &&
2205 _callprojs.fallthrough_proj != nullptr &&
2206 _callprojs.fallthrough_memproj != nullptr,
2207 "Unexpected projections from Lock/Unlock");
2208
2209 Node* fallthroughproj = _callprojs.fallthrough_proj;
2210 Node* memproj_fallthrough = _callprojs.fallthrough_memproj;
2211
2212 // The memory projection from a lock/unlock is RawMem
2213 // The input to a Lock is merged memory, so extract its RawMem input
2214 // (unless the MergeMem has been optimized away.)
2215 if (alock->is_Lock()) {
2216 // Search for MemBarAcquireLock node and delete it also.
2217 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2218 assert(membar != nullptr && membar->Opcode() == Op_MemBarAcquireLock, "");
2219 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2220 Node* memproj = membar->proj_out(TypeFunc::Memory);
2221 _igvn.replace_node(ctrlproj, fallthroughproj);
2222 _igvn.replace_node(memproj, memproj_fallthrough);
2223
2224 // Delete FastLock node also if this Lock node is unique user
2225 // (a loop peeling may clone a Lock node).
2226 Node* flock = alock->as_Lock()->fastlock_node();
2227 if (flock->outcnt() == 1) {
2228 assert(flock->unique_out() == alock, "sanity");
2229 _igvn.replace_node(flock, top());
2230 }
2261 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2262
2263 // Make the merge point
2264 Node *region;
2265 Node *mem_phi;
2266 Node *slow_path;
2267
2268 region = new RegionNode(3);
2269 // create a Phi for the memory state
2270 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2271
2272 // Optimize test; set region slot 2
2273 slow_path = opt_bits_test(ctrl, region, 2, flock);
2274 mem_phi->init_req(2, mem);
2275
2276 // Make slow path call
2277 CallNode* call = make_slow_call(lock, OptoRuntime::complete_monitor_enter_Type(),
2278 OptoRuntime::complete_monitor_locking_Java(), nullptr, slow_path,
2279 obj, box, nullptr);
2280
2281 call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
2282
2283 // Slow path can only throw asynchronous exceptions, which are always
2284 // de-opted. So the compiler thinks the slow-call can never throw an
2285 // exception. If it DOES throw an exception we would need the debug
2286 // info removed first (since if it throws there is no monitor).
2287 assert(_callprojs.fallthrough_ioproj == nullptr && _callprojs.catchall_ioproj == nullptr &&
2288 _callprojs.catchall_memproj == nullptr && _callprojs.catchall_catchproj == nullptr, "Unexpected projection from Lock");
2289
2290 // Capture slow path
2291 // disconnect fall-through projection from call and create a new one
2292 // hook up users of fall-through projection to region
2293 Node *slow_ctrl = _callprojs.fallthrough_proj->clone();
2294 transform_later(slow_ctrl);
2295 _igvn.hash_delete(_callprojs.fallthrough_proj);
2296 _callprojs.fallthrough_proj->disconnect_inputs(C);
2297 region->init_req(1, slow_ctrl);
2298 // region inputs are now complete
2299 transform_later(region);
2300 _igvn.replace_node(_callprojs.fallthrough_proj, region);
2301
2302 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2303
2304 mem_phi->init_req(1, memproj);
2305
2306 transform_later(mem_phi);
2307
2308 _igvn.replace_node(_callprojs.fallthrough_memproj, mem_phi);
2309 }
2310
2311 //------------------------------expand_unlock_node----------------------
2312 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2313
2314 Node* ctrl = unlock->in(TypeFunc::Control);
2315 Node* mem = unlock->in(TypeFunc::Memory);
2316 Node* obj = unlock->obj_node();
2317 Node* box = unlock->box_node();
2318
2319 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2320
2321 // No need for a null check on unlock
2322
2323 // Make the merge point
2324 Node *region;
2325 Node *mem_phi;
2326
2327 region = new RegionNode(3);
2328 // create a Phi for the memory state
2329 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2330
2331 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2332 funlock = transform_later( funlock )->as_FastUnlock();
2333 // Optimize test; set region slot 2
2334 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock);
2335 Node *thread = transform_later(new ThreadLocalNode());
2336
2337 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2338 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2339 "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2340
2341 call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
2342 assert(_callprojs.fallthrough_ioproj == nullptr && _callprojs.catchall_ioproj == nullptr &&
2343 _callprojs.catchall_memproj == nullptr && _callprojs.catchall_catchproj == nullptr, "Unexpected projection from Lock");
2344
2345 // No exceptions for unlocking
2346 // Capture slow path
2347 // disconnect fall-through projection from call and create a new one
2348 // hook up users of fall-through projection to region
2349 Node *slow_ctrl = _callprojs.fallthrough_proj->clone();
2350 transform_later(slow_ctrl);
2351 _igvn.hash_delete(_callprojs.fallthrough_proj);
2352 _callprojs.fallthrough_proj->disconnect_inputs(C);
2353 region->init_req(1, slow_ctrl);
2354 // region inputs are now complete
2355 transform_later(region);
2356 _igvn.replace_node(_callprojs.fallthrough_proj, region);
2357
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
2363 _igvn.replace_node(_callprojs.fallthrough_memproj, mem_phi);
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 Node* iftrue = iff->as_If()->proj_out(1);
2383 Node* iffalse = iff->as_If()->proj_out(0);
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->is_OpaqueNotNull() ||
2504 n->is_OpaqueInitializedAssertionPredicate() ||
2505 n->Opcode() == Op_MaxL ||
2506 n->Opcode() == Op_MinL ||
2507 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n),
2508 "unknown node type in macro list");
2509 }
2510 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2511 progress = progress || success;
2512 if (success) {
2513 C->print_method(PHASE_AFTER_MACRO_ELIMINATION_STEP, 5, n);
2514 }
2515 }
2516 }
2517 #ifndef PRODUCT
2518 if (PrintOptoStatistics) {
2519 int membar_after = count_MemBar(C);
2520 AtomicAccess::add(&PhaseMacroExpand::_memory_barriers_removed_counter, membar_before - membar_after);
2521 }
2522 #endif
2523 }
2524
2525 void PhaseMacroExpand::eliminate_opaque_looplimit_macro_nodes() {
2526 if (C->macro_count() == 0) {
2527 return;
2528 }
2529 refine_strip_mined_loop_macro_nodes();
2530 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2531 bool progress = true;
2532 while (progress) {
2533 progress = false;
2534 for (int i = C->macro_count(); i > 0; i--) {
2535 Node* n = C->macro_node(i-1);
2536 bool success = false;
2537 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2538 if (n->Opcode() == Op_LoopLimit) {
2539 // Remove it from macro list and put on IGVN worklist to optimize.
2540 C->remove_macro_node(n);
2541 _igvn._worklist.push(n);
2542 success = true;
2543 } else if (n->Opcode() == Op_CallStaticJava) {
2544 // Remove it from macro list and put on IGVN worklist to optimize.
2545 C->remove_macro_node(n);
2546 _igvn._worklist.push(n);
2547 success = true;
2548 } else if (n->is_Opaque1()) {
2549 _igvn.replace_node(n, n->in(1));
2550 success = true;
2551 } else if (n->is_OpaqueNotNull()) {
2552 // Tests with OpaqueNotNull nodes are implicitly known to be true. Replace the node with true. In debug builds,
2553 // we leave the test in the graph to have an additional sanity check at runtime. If the test fails (i.e. a bug),
2554 // we will execute a Halt node.
2555 #ifdef ASSERT
2556 _igvn.replace_node(n, n->in(1));
2557 #else
2558 _igvn.replace_node(n, _igvn.intcon(1));
2559 #endif
2560 success = true;
2561 } else if (n->is_OpaqueInitializedAssertionPredicate()) {
2562 // Initialized Assertion Predicates must always evaluate to true. Therefore, we get rid of them in product
2563 // builds as they are useless. In debug builds we keep them as additional verification code. Even though
2564 // loop opts are already over, we want to keep Initialized Assertion Predicates alive as long as possible to
2565 // enable folding of dead control paths within which cast nodes become top after due to impossible types -
2566 // even after loop opts are over. Therefore, we delay the removal of these opaque nodes until now.
2567 #ifdef ASSERT
2636 // Worst case is a macro node gets expanded into about 200 nodes.
2637 // Allow 50% more for optimization.
2638 if (C->check_node_count(300, "out of nodes before macro expansion")) {
2639 return true;
2640 }
2641
2642 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2643 switch (n->class_id()) {
2644 case Node::Class_Lock:
2645 expand_lock_node(n->as_Lock());
2646 break;
2647 case Node::Class_Unlock:
2648 expand_unlock_node(n->as_Unlock());
2649 break;
2650 case Node::Class_ArrayCopy:
2651 expand_arraycopy_node(n->as_ArrayCopy());
2652 break;
2653 case Node::Class_SubTypeCheck:
2654 expand_subtypecheck_node(n->as_SubTypeCheck());
2655 break;
2656 default:
2657 switch (n->Opcode()) {
2658 case Op_ModD:
2659 case Op_ModF: {
2660 CallNode* mod_macro = n->as_Call();
2661 CallNode* call = new CallLeafPureNode(mod_macro->tf(), mod_macro->entry_point(),
2662 mod_macro->_name, TypeRawPtr::BOTTOM);
2663 call->init_req(TypeFunc::Control, mod_macro->in(TypeFunc::Control));
2664 call->init_req(TypeFunc::I_O, C->top());
2665 call->init_req(TypeFunc::Memory, C->top());
2666 call->init_req(TypeFunc::ReturnAdr, C->top());
2667 call->init_req(TypeFunc::FramePtr, C->top());
2668 for (unsigned int i = 0; i < mod_macro->tf()->domain()->cnt() - TypeFunc::Parms; i++) {
2669 call->init_req(TypeFunc::Parms + i, mod_macro->in(TypeFunc::Parms + i));
2670 }
2671 _igvn.replace_node(mod_macro, call);
2672 transform_later(call);
2673 break;
2674 }
2675 default:
2676 assert(false, "unknown node type in macro list");
2677 }
2678 }
2679 assert(C->macro_count() == (old_macro_count - 1), "expansion must have deleted one node from macro list");
2680 if (C->failing()) return true;
2681 C->print_method(PHASE_AFTER_MACRO_EXPANSION_STEP, 5, n);
2682
2683 // Clean up the graph so we're less likely to hit the maximum node
2684 // limit
2685 _igvn.set_delay_transform(false);
2686 _igvn.optimize();
2687 if (C->failing()) return true;
2688 _igvn.set_delay_transform(true);
|
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/ciInstanceKlass.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "gc/shared/collectedHeap.inline.hpp"
29 #include "gc/shared/tlab_globals.hpp"
30 #include "libadt/vectset.hpp"
31 #include "memory/universe.hpp"
32 #include "opto/addnode.hpp"
33 #include "opto/arraycopynode.hpp"
34 #include "opto/callnode.hpp"
35 #include "opto/castnode.hpp"
36 #include "opto/cfgnode.hpp"
37 #include "opto/compile.hpp"
38 #include "opto/convertnode.hpp"
39 #include "opto/graphKit.hpp"
40 #include "opto/inlinetypenode.hpp"
41 #include "opto/intrinsicnode.hpp"
42 #include "opto/locknode.hpp"
43 #include "opto/loopnode.hpp"
44 #include "opto/macro.hpp"
45 #include "opto/memnode.hpp"
46 #include "opto/narrowptrnode.hpp"
47 #include "opto/node.hpp"
48 #include "opto/opaquenode.hpp"
49 #include "opto/opcodes.hpp"
50 #include "opto/phaseX.hpp"
51 #include "opto/rootnode.hpp"
52 #include "opto/runtime.hpp"
53 #include "opto/subnode.hpp"
54 #include "opto/subtypenode.hpp"
55 #include "opto/type.hpp"
56 #include "prims/jvmtiExport.hpp"
57 #include "runtime/continuation.hpp"
58 #include "runtime/sharedRuntime.hpp"
59 #include "runtime/stubRoutines.hpp"
60 #include "utilities/globalDefinitions.hpp"
61 #include "utilities/macros.hpp"
62 #include "utilities/powerOfTwo.hpp"
63 #if INCLUDE_G1GC
64 #include "gc/g1/g1ThreadLocalData.hpp"
65 #endif // INCLUDE_G1GC
66
67
68 //
69 // Replace any references to "oldref" in inputs to "use" with "newref".
70 // Returns the number of replacements made.
71 //
72 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
73 int nreplacements = 0;
74 uint req = use->req();
75 for (uint j = 0; j < use->len(); j++) {
76 Node *uin = use->in(j);
77 if (uin == oldref) {
78 if (j < req)
79 use->set_req(j, newref);
80 else
81 use->set_prec(j, newref);
82 nreplacements++;
83 } else if (j >= req && uin == nullptr) {
84 break;
85 }
86 }
87 return nreplacements;
88 }
89
90
91 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word) {
92 Node* cmp = word;
93 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne));
94 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
95 transform_later(iff);
96
97 // Fast path taken.
98 Node *fast_taken = transform_later(new IfFalseNode(iff));
99
100 // Fast path not-taken, i.e. slow path
101 Node *slow_taken = transform_later(new IfTrueNode(iff));
102
103 region->init_req(edge, fast_taken); // Capture fast-control
104 return slow_taken;
105 }
106
107 //--------------------copy_predefined_input_for_runtime_call--------------------
108 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
109 // Set fixed predefined input arguments
122 // Slow-path call
123 CallNode *call = leaf_name
124 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
125 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), TypeRawPtr::BOTTOM );
126
127 // Slow path call has no side-effects, uses few values
128 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
129 if (parm0 != nullptr) call->init_req(TypeFunc::Parms+0, parm0);
130 if (parm1 != nullptr) call->init_req(TypeFunc::Parms+1, parm1);
131 if (parm2 != nullptr) call->init_req(TypeFunc::Parms+2, parm2);
132 call->copy_call_debug_info(&_igvn, oldcall);
133 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
134 _igvn.replace_node(oldcall, call);
135 transform_later(call);
136
137 return call;
138 }
139
140 void PhaseMacroExpand::eliminate_gc_barrier(Node* p2x) {
141 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
142 bs->eliminate_gc_barrier(&_igvn, p2x);
143 #ifndef PRODUCT
144 if (PrintOptoStatistics) {
145 AtomicAccess::inc(&PhaseMacroExpand::_GC_barriers_removed_counter);
146 }
147 #endif
148 }
149
150 // Search for a memory operation for the specified memory slice.
151 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
152 Node *orig_mem = mem;
153 Node *alloc_mem = alloc->as_Allocate()->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
154 assert(alloc_mem != nullptr, "Allocation without a memory projection.");
155 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
156 while (true) {
157 if (mem == alloc_mem || mem == start_mem ) {
158 return mem; // hit one of our sentinels
159 } else if (mem->is_MergeMem()) {
160 mem = mem->as_MergeMem()->memory_at(alias_idx);
161 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
162 Node *in = mem->in(0);
165 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
166 return in;
167 } else if (in->is_Call()) {
168 CallNode *call = in->as_Call();
169 if (call->may_modify(tinst, phase)) {
170 assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape");
171 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) {
172 return in;
173 }
174 }
175 mem = in->in(TypeFunc::Memory);
176 } else if (in->is_MemBar()) {
177 ArrayCopyNode* ac = nullptr;
178 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) {
179 if (ac != nullptr) {
180 assert(ac->is_clonebasic(), "Only basic clone is a non escaping clone");
181 return ac;
182 }
183 }
184 mem = in->in(TypeFunc::Memory);
185 } else if (in->is_LoadFlat() || in->is_StoreFlat()) {
186 mem = in->in(TypeFunc::Memory);
187 } else {
188 #ifdef ASSERT
189 in->dump();
190 mem->dump();
191 assert(false, "unexpected projection");
192 #endif
193 }
194 } else if (mem->is_Store()) {
195 const TypePtr* atype = mem->as_Store()->adr_type();
196 int adr_idx = phase->C->get_alias_index(atype);
197 if (adr_idx == alias_idx) {
198 assert(atype->isa_oopptr(), "address type must be oopptr");
199 int adr_offset = atype->flat_offset();
200 uint adr_iid = atype->is_oopptr()->instance_id();
201 // Array elements references have the same alias_idx
202 // but different offset and different instance_id.
203 if (adr_offset == offset && adr_iid == alloc->_idx) {
204 return mem;
205 }
206 } else {
207 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
208 }
209 mem = mem->in(MemNode::Memory);
210 } else if (mem->is_ClearArray()) {
211 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
212 // Can not bypass initialization of the instance
213 // we are looking.
214 DEBUG_ONLY(intptr_t offset;)
215 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
216 InitializeNode* init = alloc->as_Allocate()->initialization();
217 // We are looking for stored value, return Initialize node
218 // or memory edge from Allocate node.
219 if (init != nullptr) {
224 }
225 // Otherwise skip it (the call updated 'mem' value).
226 } else if (mem->Opcode() == Op_SCMemProj) {
227 mem = mem->in(0);
228 Node* adr = nullptr;
229 if (mem->is_LoadStore()) {
230 adr = mem->in(MemNode::Address);
231 } else {
232 assert(mem->Opcode() == Op_EncodeISOArray ||
233 mem->Opcode() == Op_StrCompressedCopy, "sanity");
234 adr = mem->in(3); // Destination array
235 }
236 const TypePtr* atype = adr->bottom_type()->is_ptr();
237 int adr_idx = phase->C->get_alias_index(atype);
238 if (adr_idx == alias_idx) {
239 DEBUG_ONLY(mem->dump();)
240 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
241 return nullptr;
242 }
243 mem = mem->in(MemNode::Memory);
244 } else if (mem->Opcode() == Op_StrInflatedCopy) {
245 Node* adr = mem->in(3); // Destination array
246 const TypePtr* atype = adr->bottom_type()->is_ptr();
247 int adr_idx = phase->C->get_alias_index(atype);
248 if (adr_idx == alias_idx) {
249 DEBUG_ONLY(mem->dump();)
250 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
251 return nullptr;
252 }
253 mem = mem->in(MemNode::Memory);
254 } else {
255 return mem;
256 }
257 assert(mem != orig_mem, "dead memory loop");
258 }
259 }
260
261 // Generate loads from source of the arraycopy for fields of
262 // destination needed at a deoptimization point
263 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) {
264 BasicType bt = ft;
269 }
270 Node* res = nullptr;
271 if (ac->is_clonebasic()) {
272 assert(ac->in(ArrayCopyNode::Src) != ac->in(ArrayCopyNode::Dest), "clone source equals destination");
273 Node* base = ac->in(ArrayCopyNode::Src);
274 Node* adr = _igvn.transform(new AddPNode(base, base, _igvn.MakeConX(offset)));
275 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
276 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
277 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
278 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt);
279 } else {
280 if (ac->modifies(offset, offset, &_igvn, true)) {
281 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
282 uint shift = exact_log2(type2aelembytes(bt));
283 Node* src_pos = ac->in(ArrayCopyNode::SrcPos);
284 Node* dest_pos = ac->in(ArrayCopyNode::DestPos);
285 const TypeInt* src_pos_t = _igvn.type(src_pos)->is_int();
286 const TypeInt* dest_pos_t = _igvn.type(dest_pos)->is_int();
287
288 Node* adr = nullptr;
289 Node* base = ac->in(ArrayCopyNode::Src);
290 const TypeAryPtr* adr_type = _igvn.type(base)->is_aryptr();
291 if (adr_type->is_flat()) {
292 shift = adr_type->flat_log_elem_size();
293 }
294 if (src_pos_t->is_con() && dest_pos_t->is_con()) {
295 intptr_t off = ((src_pos_t->get_con() - dest_pos_t->get_con()) << shift) + offset;
296 adr = _igvn.transform(new AddPNode(base, base, _igvn.MakeConX(off)));
297 adr_type = _igvn.type(adr)->is_aryptr();
298 assert(adr_type == _igvn.type(base)->is_aryptr()->add_field_offset_and_offset(off), "incorrect address type");
299 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
300 // Don't emit a new load from src if src == dst but try to get the value from memory instead
301 return value_from_mem(ac->in(TypeFunc::Memory), ctl, ft, ftype, adr_type, alloc);
302 }
303 } else {
304 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
305 // Non constant offset in the array: we can't statically
306 // determine the value
307 return nullptr;
308 }
309 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
310 #ifdef _LP64
311 diff = _igvn.transform(new ConvI2LNode(diff));
312 #endif
313 diff = _igvn.transform(new LShiftXNode(diff, _igvn.intcon(shift)));
314
315 Node* off = _igvn.transform(new AddXNode(_igvn.MakeConX(offset), diff));
316 adr = _igvn.transform(new AddPNode(base, base, off));
317 // In the case of a flat inline type array, each field has its
318 // own slice so we need to extract the field being accessed from
319 // the address computation
320 adr_type = adr_type->add_field_offset_and_offset(offset)->add_offset(Type::OffsetBot)->is_aryptr();
321 adr = _igvn.transform(new CastPPNode(ctl, adr, adr_type));
322 }
323 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
324 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
325 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt);
326 }
327 }
328 if (res != nullptr) {
329 if (ftype->isa_narrowoop()) {
330 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
331 assert(res->isa_DecodeN(), "should be narrow oop");
332 res = _igvn.transform(new EncodePNode(res, ftype));
333 }
334 return res;
335 }
336 return nullptr;
337 }
338
339 //
340 // Given a Memory Phi, compute a value Phi containing the values from stores
341 // on the input paths.
342 // Note: this function is recursive, its depth is limited by the "level" argument
343 // Returns the computed Phi, or null if it cannot compute it.
344 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) {
345 assert(mem->is_Phi(), "sanity");
346 int alias_idx = C->get_alias_index(adr_t);
347 int offset = adr_t->flat_offset();
348 int instance_id = adr_t->instance_id();
349
350 // Check if an appropriate value phi already exists.
351 Node* region = mem->in(0);
352 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
353 Node* phi = region->fast_out(k);
354 if (phi->is_Phi() && phi != mem &&
355 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) {
356 return phi;
357 }
358 }
359 // Check if an appropriate new value phi already exists.
360 Node* new_phi = value_phis->find(mem->_idx);
361 if (new_phi != nullptr)
362 return new_phi;
363
364 if (level <= 0) {
365 return nullptr; // Give up: phi tree too deep
366 }
367 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
368 Node *alloc_mem = alloc->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
369 assert(alloc_mem != nullptr, "Allocation without a memory projection.");
370
371 uint length = mem->req();
372 GrowableArray <Node *> values(length, length, nullptr);
373
374 // create a new Phi for the value
375 PhiNode *phi = new PhiNode(mem->in(0), phi_type, nullptr, mem->_idx, instance_id, alias_idx, offset);
376 transform_later(phi);
377 value_phis->push(phi, mem->_idx);
378
379 for (uint j = 1; j < length; j++) {
380 Node *in = mem->in(j);
381 if (in == nullptr || in->is_top()) {
382 values.at_put(j, in);
383 } else {
384 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
385 if (val == start_mem || val == alloc_mem) {
386 // hit a sentinel, return appropriate value
387 Node* init_value = alloc->in(AllocateNode::InitValue);
388 if (init_value != nullptr) {
389 if (val == start_mem) {
390 // TODO 8350865 Scalar replacement does not work well for flat arrays.
391 // Somehow we ended up with root mem and therefore walked past the alloc. Fix this. Triggered by TestGenerated::test15
392 // Don't we need field_value_by_offset?
393 return nullptr;
394 }
395 values.at_put(j, init_value);
396 } else {
397 assert(alloc->in(AllocateNode::RawInitValue) == nullptr, "init value may not be null");
398 values.at_put(j, _igvn.zerocon(ft));
399 }
400 continue;
401 }
402 if (val->is_Initialize()) {
403 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
404 }
405 if (val == nullptr) {
406 return nullptr; // can't find a value on this path
407 }
408 if (val == mem) {
409 values.at_put(j, mem);
410 } else if (val->is_Store()) {
411 Node* n = val->in(MemNode::ValueIn);
412 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
413 n = bs->step_over_gc_barrier(n);
414 if (is_subword_type(ft)) {
415 n = Compile::narrow_value(ft, n, phi_type, &_igvn, true);
416 }
417 values.at_put(j, n);
418 } else if (val->is_Proj() && val->in(0) == alloc) {
419 Node* init_value = alloc->in(AllocateNode::InitValue);
420 if (init_value != nullptr) {
421 // TODO 8350865 Scalar replacement does not work well for flat arrays.
422 // Is this correct for non-all-zero init values? Don't we need field_value_by_offset?
423 values.at_put(j, init_value);
424 } else {
425 assert(alloc->in(AllocateNode::RawInitValue) == nullptr, "init value may not be null");
426 values.at_put(j, _igvn.zerocon(ft));
427 }
428 } else if (val->is_Phi()) {
429 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
430 if (val == nullptr) {
431 return nullptr;
432 }
433 values.at_put(j, val);
434 } else if (val->Opcode() == Op_SCMemProj) {
435 assert(val->in(0)->is_LoadStore() ||
436 val->in(0)->Opcode() == Op_EncodeISOArray ||
437 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity");
438 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
439 return nullptr;
440 } else if (val->is_ArrayCopy()) {
441 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc);
442 if (res == nullptr) {
443 return nullptr;
444 }
445 values.at_put(j, res);
446 } else if (val->is_top()) {
447 // This indicates that this path into the phi is dead. Top will eventually also propagate into the Region.
455 }
456 }
457 // Set Phi's inputs
458 for (uint j = 1; j < length; j++) {
459 if (values.at(j) == mem) {
460 phi->init_req(j, phi);
461 } else {
462 phi->init_req(j, values.at(j));
463 }
464 }
465 return phi;
466 }
467
468 // Search the last value stored into the object's field.
469 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) {
470 assert(adr_t->is_known_instance_field(), "instance required");
471 int instance_id = adr_t->instance_id();
472 assert((uint)instance_id == alloc->_idx, "wrong allocation");
473
474 int alias_idx = C->get_alias_index(adr_t);
475 int offset = adr_t->flat_offset();
476 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
477 Node *alloc_mem = alloc->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
478 assert(alloc_mem != nullptr, "Allocation without a memory projection.");
479 VectorSet visited;
480
481 bool done = sfpt_mem == alloc_mem;
482 Node *mem = sfpt_mem;
483 while (!done) {
484 if (visited.test_set(mem->_idx)) {
485 return nullptr; // found a loop, give up
486 }
487 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
488 if (mem == start_mem || mem == alloc_mem) {
489 done = true; // hit a sentinel, return appropriate 0 value
490 } else if (mem->is_Initialize()) {
491 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
492 if (mem == nullptr) {
493 done = true; // Something went wrong.
494 } else if (mem->is_Store()) {
495 const TypePtr* atype = mem->as_Store()->adr_type();
496 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
497 done = true;
498 }
499 } else if (mem->is_Store()) {
500 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
501 assert(atype != nullptr, "address type must be oopptr");
502 assert(C->get_alias_index(atype) == alias_idx &&
503 atype->is_known_instance_field() && atype->flat_offset() == offset &&
504 atype->instance_id() == instance_id, "store is correct memory slice");
505 done = true;
506 } else if (mem->is_Phi()) {
507 // try to find a phi's unique input
508 Node *unique_input = nullptr;
509 Node *top = C->top();
510 for (uint i = 1; i < mem->req(); i++) {
511 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
512 if (n == nullptr || n == top || n == mem) {
513 continue;
514 } else if (unique_input == nullptr) {
515 unique_input = n;
516 } else if (unique_input != n) {
517 unique_input = top;
518 break;
519 }
520 }
521 if (unique_input != nullptr && unique_input != top) {
522 mem = unique_input;
523 } else {
524 done = true;
525 }
526 } else if (mem->is_ArrayCopy()) {
527 done = true;
528 } else if (mem->is_top()) {
529 // The slice is on a dead path. Returning nullptr would lead to elimination
530 // bailout, but we want to prevent that. Just forwarding the top is also legal,
531 // and IGVN can just clean things up, and remove whatever receives top.
532 return mem;
533 } else {
534 DEBUG_ONLY( mem->dump(); )
535 assert(false, "unexpected node");
536 }
537 }
538 if (mem != nullptr) {
539 if (mem == start_mem || mem == alloc_mem) {
540 // hit a sentinel, return appropriate value
541 Node* init_value = alloc->in(AllocateNode::InitValue);
542 if (init_value != nullptr) {
543 if (adr_t->is_flat()) {
544 if (init_value->is_EncodeP()) {
545 init_value = init_value->in(1);
546 }
547 assert(adr_t->is_aryptr()->field_offset().get() != Type::OffsetBot, "Unknown offset");
548 offset = adr_t->is_aryptr()->field_offset().get() + init_value->bottom_type()->inline_klass()->payload_offset();
549 init_value = init_value->as_InlineType()->field_value_by_offset(offset, true);
550 if (ft == T_NARROWOOP) {
551 init_value = transform_later(new EncodePNode(init_value, init_value->bottom_type()->make_ptr()));
552 }
553 }
554 return init_value;
555 }
556 assert(alloc->in(AllocateNode::RawInitValue) == nullptr, "init value may not be null");
557 return _igvn.zerocon(ft);
558 } else if (mem->is_Store()) {
559 Node* n = mem->in(MemNode::ValueIn);
560 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
561 n = bs->step_over_gc_barrier(n);
562 return n;
563 } else if (mem->is_Phi()) {
564 // attempt to produce a Phi reflecting the values on the input paths of the Phi
565 Node_Stack value_phis(8);
566 Node* phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
567 if (phi != nullptr) {
568 return phi;
569 } else {
570 // Kill all new Phis
571 while(value_phis.is_nonempty()) {
572 Node* n = value_phis.node();
573 _igvn.replace_node(n, C->top());
574 value_phis.pop();
575 }
576 }
577 } else if (mem->is_ArrayCopy()) {
578 Node* ctl = mem->in(0);
579 Node* m = mem->in(TypeFunc::Memory);
580 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj()) {
581 // pin the loads in the uncommon trap path
582 ctl = sfpt_ctl;
583 m = sfpt_mem;
584 }
585 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc);
586 }
587 }
588 // Something went wrong.
589 return nullptr;
590 }
591
592 // Search the last value stored into the inline type's fields (for flat arrays).
593 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) {
594 auto report_failure = [&](int field_offset_in_element) {
595 #ifndef PRODUCT
596 if (PrintEliminateAllocations) {
597 ciInlineKlass* elem_klass = elem_adr_type->elem()->inline_klass();
598 int offset = field_offset_in_element + elem_klass->payload_offset();
599 ciField* flattened_field = elem_klass->get_field_by_offset(offset, false);
600 assert(flattened_field != nullptr, "must have a field of type %s at offset %d", elem_klass->name()->as_utf8(), offset);
601 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);
602 tty->print(", which prevents elimination of: ");
603 alloc->dump();
604 }
605 #endif // PRODUCT
606 };
607
608 // Create a new InlineTypeNode and retrieve the field values from memory
609 InlineTypeNode* vt = InlineTypeNode::make_uninitialized(_igvn, vk, false);
610 transform_later(vt);
611 if (null_free) {
612 vt->set_null_marker(_igvn);
613 } else {
614 int nm_offset_in_element = offset_in_element + vk->null_marker_offset_in_payload();
615 const TypeAryPtr* nm_adr_type = elem_adr_type->with_field_offset(nm_offset_in_element);
616 Node* nm_value = value_from_mem(sfpt->memory(), sfpt->control(), T_BOOLEAN, TypeInt::BOOL, nm_adr_type, alloc);
617 if (nm_value != nullptr) {
618 vt->set_null_marker(_igvn, nm_value);
619 } else {
620 report_failure(nm_offset_in_element);
621 return nullptr;
622 }
623 }
624
625 for (int i = 0; i < vk->nof_declared_nonstatic_fields(); ++i) {
626 ciType* field_type = vt->field_type(i);
627 int field_offset_in_element = offset_in_element + vt->field_offset(i) - vk->payload_offset();
628 Node* field_value = nullptr;
629 if (vt->field_is_flat(i)) {
630 field_value = inline_type_from_mem(field_type->as_inline_klass(), elem_adr_type, elem_idx, field_offset_in_element, vt->field_is_null_free(i), alloc, sfpt);
631 } else {
632 const Type* ft = Type::get_const_type(field_type);
633 BasicType bt = type2field[field_type->basic_type()];
634 if (UseCompressedOops && !is_java_primitive(bt)) {
635 ft = ft->make_narrowoop();
636 bt = T_NARROWOOP;
637 }
638 // Each inline type field has its own memory slice
639 const TypeAryPtr* field_adr_type = elem_adr_type->with_field_offset(field_offset_in_element);
640 field_value = value_from_mem(sfpt->memory(), sfpt->control(), bt, ft, field_adr_type, alloc);
641 if (field_value == nullptr) {
642 report_failure(field_offset_in_element);
643 } else if (ft->isa_narrowoop()) {
644 assert(UseCompressedOops, "unexpected narrow oop");
645 if (field_value->is_EncodeP()) {
646 field_value = field_value->in(1);
647 } else if (!field_value->is_InlineType()) {
648 field_value = transform_later(new DecodeNNode(field_value, field_value->get_ptr_type()));
649 }
650 }
651 }
652 if (field_value != nullptr) {
653 vt->set_field_value(i, field_value);
654 } else {
655 return nullptr;
656 }
657 }
658 return vt;
659 }
660
661 // Check the possibility of scalar replacement.
662 bool PhaseMacroExpand::can_eliminate_allocation(PhaseIterGVN* igvn, AllocateNode *alloc, GrowableArray <SafePointNode *>* safepoints) {
663 // Scan the uses of the allocation to check for anything that would
664 // prevent us from eliminating it.
665 NOT_PRODUCT( const char* fail_eliminate = nullptr; )
666 DEBUG_ONLY( Node* disq_node = nullptr; )
667 bool can_eliminate = true;
668 bool reduce_merge_precheck = (safepoints == nullptr);
669
670 Unique_Node_List worklist;
671 Node* res = alloc->result_cast();
672 const TypeOopPtr* res_type = nullptr;
673 if (res == nullptr) {
674 // All users were eliminated.
675 } else if (!res->is_CheckCastPP()) {
676 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
677 can_eliminate = false;
678 } else {
679 worklist.push(res);
680 res_type = igvn->type(res)->isa_oopptr();
681 if (res_type == nullptr) {
682 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
683 can_eliminate = false;
684 } else if (!res_type->klass_is_exact()) {
685 NOT_PRODUCT(fail_eliminate = "Not an exact type.";)
686 can_eliminate = false;
687 } else if (res_type->isa_aryptr()) {
688 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
689 if (length < 0) {
690 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
691 can_eliminate = false;
692 }
693 }
694 }
695
696 while (can_eliminate && worklist.size() > 0) {
697 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
698 res = worklist.pop();
699 for (DUIterator_Fast jmax, j = res->fast_outs(jmax); j < jmax && can_eliminate; j++) {
700 Node* use = res->fast_out(j);
701
702 if (use->is_AddP()) {
703 const TypePtr* addp_type = igvn->type(use)->is_ptr();
704 int offset = addp_type->offset();
705
706 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
707 NOT_PRODUCT(fail_eliminate = "Undefined field reference";)
708 can_eliminate = false;
709 break;
710 }
711 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
712 k < kmax && can_eliminate; k++) {
713 Node* n = use->fast_out(k);
714 if ((n->is_Mem() && n->as_Mem()->is_mismatched_access()) || n->is_LoadFlat() || n->is_StoreFlat()) {
715 DEBUG_ONLY(disq_node = n);
716 NOT_PRODUCT(fail_eliminate = "Mismatched access");
717 can_eliminate = false;
718 }
719 if (!n->is_Store() && n->Opcode() != Op_CastP2X && !bs->is_gc_pre_barrier_node(n) && !reduce_merge_precheck) {
720 DEBUG_ONLY(disq_node = n;)
721 if (n->is_Load() || n->is_LoadStore()) {
722 NOT_PRODUCT(fail_eliminate = "Field load";)
723 } else {
724 NOT_PRODUCT(fail_eliminate = "Not store field reference";)
725 }
726 can_eliminate = false;
727 }
728 }
729 } else if (use->is_ArrayCopy() &&
730 (use->as_ArrayCopy()->is_clonebasic() ||
731 use->as_ArrayCopy()->is_arraycopy_validated() ||
732 use->as_ArrayCopy()->is_copyof_validated() ||
733 use->as_ArrayCopy()->is_copyofrange_validated()) &&
734 use->in(ArrayCopyNode::Dest) == res) {
735 // ok to eliminate
736 } else if (use->is_SafePoint()) {
737 SafePointNode* sfpt = use->as_SafePoint();
738 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
739 // Object is passed as argument.
740 DEBUG_ONLY(disq_node = use;)
741 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
742 can_eliminate = false;
743 }
744 Node* sfptMem = sfpt->memory();
745 if (sfptMem == nullptr || sfptMem->is_top()) {
746 DEBUG_ONLY(disq_node = use;)
747 NOT_PRODUCT(fail_eliminate = "null or TOP memory";)
748 can_eliminate = false;
749 } else if (!reduce_merge_precheck) {
750 assert(!res->is_Phi() || !res->as_Phi()->can_be_inline_type(), "Inline type allocations should not have safepoint uses");
751 safepoints->append_if_missing(sfpt);
752 }
753 } else if (use->is_InlineType() && use->as_InlineType()->get_oop() == res) {
754 // Look at uses
755 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); k < kmax; k++) {
756 Node* u = use->fast_out(k);
757 if (u->is_InlineType()) {
758 // Use in flat field can be eliminated
759 InlineTypeNode* vt = u->as_InlineType();
760 for (uint i = 0; i < vt->field_count(); ++i) {
761 if (vt->field_value(i) == use && !vt->field_is_flat(i)) {
762 can_eliminate = false; // Use in non-flat field
763 break;
764 }
765 }
766 } else {
767 // Add other uses to the worklist to process individually
768 worklist.push(use);
769 }
770 }
771 } else if (use->Opcode() == Op_StoreX && use->in(MemNode::Address) == res) {
772 // Store to mark word of inline type larval buffer
773 assert(res_type->is_inlinetypeptr(), "Unexpected store to mark word");
774 } else if (res_type->is_inlinetypeptr() && (use->Opcode() == Op_MemBarRelease || use->Opcode() == Op_MemBarStoreStore)) {
775 // Inline type buffer allocations are followed by a membar
776 } else if (reduce_merge_precheck &&
777 (use->is_Phi() || use->is_EncodeP() ||
778 use->Opcode() == Op_MemBarRelease ||
779 (UseStoreStoreForCtor && use->Opcode() == Op_MemBarStoreStore))) {
780 // Nothing to do
781 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
782 if (use->is_Phi()) {
783 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
784 NOT_PRODUCT(fail_eliminate = "Object is return value";)
785 } else {
786 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
787 }
788 DEBUG_ONLY(disq_node = use;)
789 } else {
790 if (use->Opcode() == Op_Return) {
791 NOT_PRODUCT(fail_eliminate = "Object is return value";)
792 } else {
793 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
794 }
795 DEBUG_ONLY(disq_node = use;)
796 }
797 can_eliminate = false;
798 } else {
799 assert(use->Opcode() == Op_CastP2X, "should be");
800 assert(!use->has_out_with(Op_OrL), "should have been removed because oop is never null");
801 }
802 }
803 }
804
805 #ifndef PRODUCT
806 if (PrintEliminateAllocations && safepoints != nullptr) {
807 if (can_eliminate) {
808 tty->print("Scalar ");
809 if (res == nullptr)
810 alloc->dump();
811 else
812 res->dump();
813 } else {
814 tty->print("NotScalar (%s)", fail_eliminate);
815 if (res == nullptr)
816 alloc->dump();
817 else
818 res->dump();
819 #ifdef ASSERT
820 if (disq_node != nullptr) {
821 tty->print(" >>>> ");
822 disq_node->dump();
823 }
824 #endif /*ASSERT*/
825 }
826 }
827
828 if (TraceReduceAllocationMerges && !can_eliminate && reduce_merge_precheck) {
829 tty->print_cr("\tCan't eliminate allocation because '%s': ", fail_eliminate != nullptr ? fail_eliminate : "");
830 DEBUG_ONLY(if (disq_node != nullptr) disq_node->dump();)
831 }
832 #endif
833 return can_eliminate;
898 // CheckCastPP result was not updated in the stack slot, and so
899 // we ended up using the CastPP. That means that the field knows
900 // that it should get an oop from an interface, but the value lost
901 // that information, and so it is not a subtype.
902 // There may be other issues, feel free to investigate further!
903 if (!is_java_primitive(value_bt)) { return; }
904
905 tty->print_cr("value not compatible for field: %s vs %s",
906 type2name(value_bt),
907 type2name(field_bt));
908 tty->print("value_type: ");
909 value_type->dump();
910 tty->cr();
911 tty->print("field_type: ");
912 field_type->dump();
913 tty->cr();
914 assert(false, "value_type does not fit field_type");
915 }
916 #endif
917
918 void PhaseMacroExpand::process_field_value_at_safepoint(const Type* field_type, Node* field_val, SafePointNode* sfpt, Unique_Node_List* value_worklist) {
919 if (UseCompressedOops && field_type->isa_narrowoop()) {
920 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
921 // to be able scalar replace the allocation.
922 if (field_val->is_EncodeP()) {
923 field_val = field_val->in(1);
924 } else if (!field_val->is_InlineType()) {
925 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
926 }
927 }
928
929 // Keep track of inline types to scalarize them later
930 if (field_val->is_InlineType()) {
931 value_worklist->push(field_val);
932 } else if (field_val->is_Phi()) {
933 PhiNode* phi = field_val->as_Phi();
934 // Eagerly replace inline type phis now since we could be removing an inline type allocation where we must
935 // scalarize all its fields in safepoints.
936 field_val = phi->try_push_inline_types_down(&_igvn, true);
937 if (field_val->is_InlineType()) {
938 value_worklist->push(field_val);
939 }
940 }
941 DEBUG_ONLY(verify_type_compatability(field_val->bottom_type(), field_type);)
942 sfpt->add_req(field_val);
943 }
944
945 bool PhaseMacroExpand::add_array_elems_to_safepoint(AllocateNode* alloc, const TypeAryPtr* array_type, SafePointNode* sfpt, Unique_Node_List* value_worklist) {
946 const Type* elem_type = array_type->elem();
947 BasicType basic_elem_type = elem_type->array_element_basic_type();
948
949 intptr_t elem_size;
950 uint header_size;
951 if (array_type->is_flat()) {
952 elem_size = array_type->flat_elem_size();
953 header_size = arrayOopDesc::base_offset_in_bytes(T_FLAT_ELEMENT);
954 } else {
955 elem_size = type2aelembytes(basic_elem_type);
956 header_size = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
957 }
958
959 int n_elems = alloc->in(AllocateNode::ALength)->get_int();
960 for (int elem_idx = 0; elem_idx < n_elems; elem_idx++) {
961 intptr_t elem_offset = header_size + elem_idx * elem_size;
962 const TypeAryPtr* elem_adr_type = array_type->with_offset(elem_offset);
963 Node* elem_val;
964 if (array_type->is_flat()) {
965 ciInlineKlass* elem_klass = elem_type->inline_klass();
966 assert(elem_klass->maybe_flat_in_array(), "must be flat in array");
967 elem_val = inline_type_from_mem(elem_klass, elem_adr_type, elem_idx, 0, array_type->is_null_free(), alloc, sfpt);
968 } else {
969 elem_val = value_from_mem(sfpt->memory(), sfpt->control(), basic_elem_type, elem_type, elem_adr_type, alloc);
970 #ifndef PRODUCT
971 if (PrintEliminateAllocations && elem_val == nullptr) {
972 tty->print("=== At SafePoint node %d can't find value of array element [%d]", sfpt->_idx, elem_idx);
973 tty->print(", which prevents elimination of: ");
974 alloc->dump();
975 }
976 #endif // PRODUCT
977 }
978 if (elem_val == nullptr) {
979 return false;
980 }
981
982 process_field_value_at_safepoint(elem_type, elem_val, sfpt, value_worklist);
983 }
984
985 return true;
986 }
987
988 // Recursively adds all flattened fields of a type 'iklass' inside 'base' to 'sfpt'.
989 // 'offset_minus_header' refers to the offset of the payload of 'iklass' inside 'base' minus the
990 // payload offset of 'iklass'. If 'base' is of type 'iklass' then 'offset_minus_header' == 0.
991 bool PhaseMacroExpand::add_inst_fields_to_safepoint(ciInstanceKlass* iklass, AllocateNode* alloc, Node* base, int offset_minus_header, SafePointNode* sfpt, Unique_Node_List* value_worklist) {
992 const TypeInstPtr* base_type = _igvn.type(base)->is_instptr();
993 auto report_failure = [&](int offset) {
994 #ifndef PRODUCT
995 if (PrintEliminateAllocations) {
996 ciInstanceKlass* base_klass = base_type->instance_klass();
997 ciField* flattened_field = base_klass->get_field_by_offset(offset, false);
998 assert(flattened_field != nullptr, "must have a field of type %s at offset %d", base_klass->name()->as_utf8(), offset);
999 tty->print("=== At SafePoint node %d can't find value of field: ", sfpt->_idx);
1000 flattened_field->print();
1001 int field_idx = C->alias_type(flattened_field)->index();
1002 tty->print(" (alias_idx=%d)", field_idx);
1003 tty->print(", which prevents elimination of: ");
1004 base->dump();
1005 }
1006 #endif // PRODUCT
1007 };
1008
1009 for (int i = 0; i < iklass->nof_declared_nonstatic_fields(); i++) {
1010 ciField* field = iklass->declared_nonstatic_field_at(i);
1011 if (field->is_flat()) {
1012 ciInlineKlass* fvk = field->type()->as_inline_klass();
1013 int field_offset_minus_header = offset_minus_header + field->offset_in_bytes() - fvk->payload_offset();
1014 bool success = add_inst_fields_to_safepoint(fvk, alloc, base, field_offset_minus_header, sfpt, value_worklist);
1015 if (!success) {
1016 return false;
1017 }
1018
1019 // The null marker of a field is added right after we scalarize that field
1020 if (!field->is_null_free()) {
1021 int nm_offset = offset_minus_header + field->null_marker_offset();
1022 Node* null_marker = value_from_mem(sfpt->memory(), sfpt->control(), T_BOOLEAN, TypeInt::BOOL, base_type->with_offset(nm_offset), alloc);
1023 if (null_marker == nullptr) {
1024 report_failure(nm_offset);
1025 return false;
1026 }
1027 process_field_value_at_safepoint(TypeInt::BOOL, null_marker, sfpt, value_worklist);
1028 }
1029
1030 continue;
1031 }
1032
1033 int offset = offset_minus_header + field->offset_in_bytes();
1034 ciType* elem_type = field->type();
1035 BasicType basic_elem_type = field->layout_type();
1036
1037 const Type* field_type;
1038 if (is_reference_type(basic_elem_type)) {
1039 if (!elem_type->is_loaded()) {
1040 field_type = TypeInstPtr::BOTTOM;
1041 } else {
1042 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
1043 }
1044 if (UseCompressedOops) {
1045 field_type = field_type->make_narrowoop();
1046 basic_elem_type = T_NARROWOOP;
1047 }
1048 } else {
1049 field_type = Type::get_const_basic_type(basic_elem_type);
1050 }
1051
1052 const TypeInstPtr* field_addr_type = base_type->add_offset(offset)->isa_instptr();
1053 Node* field_val = value_from_mem(sfpt->memory(), sfpt->control(), basic_elem_type, field_type, field_addr_type, alloc);
1054 if (field_val == nullptr) {
1055 report_failure(offset);
1056 return false;
1057 }
1058 process_field_value_at_safepoint(field_type, field_val, sfpt, value_worklist);
1059 }
1060
1061 return true;
1062 }
1063
1064 SafePointScalarObjectNode* PhaseMacroExpand::create_scalarized_object_description(AllocateNode* alloc, SafePointNode* sfpt,
1065 Unique_Node_List* value_worklist) {
1066 // Fields of scalar objs are referenced only at the end
1067 // of regular debuginfo at the last (youngest) JVMS.
1068 // Record relative start index.
1069 ciInstanceKlass* iklass = nullptr;
1070 const TypeOopPtr* res_type = nullptr;
1071 int nfields = 0;
1072 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
1073 Node* res = alloc->result_cast();
1074
1075 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
1076 assert(sfpt->jvms() != nullptr, "missed JVMS");
1077 uint before_sfpt_req = sfpt->req();
1078
1079 if (res != nullptr) { // Could be null when there are no users
1080 res_type = _igvn.type(res)->isa_oopptr();
1081
1082 if (res_type->isa_instptr()) {
1083 // find the fields of the class which will be needed for safepoint debug information
1084 iklass = res_type->is_instptr()->instance_klass();
1085 nfields = iklass->nof_nonstatic_fields();
1086 } else {
1087 // find the array's elements which will be needed for safepoint debug information
1088 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
1089 assert(nfields >= 0, "must be an array klass.");
1090 }
1091
1092 if (res->bottom_type()->is_inlinetypeptr()) {
1093 // Nullable inline types have a null marker field which is added to the safepoint when scalarizing them (see
1094 // InlineTypeNode::make_scalar_in_safepoint()). When having circular inline types, we stop scalarizing at depth 1
1095 // to avoid an endless recursion. Therefore, we do not have a SafePointScalarObjectNode node here, yet.
1096 // We are about to create a SafePointScalarObjectNode as if this is a normal object. Add an additional int input
1097 // with value 1 which sets the null marker to true to indicate that the object is always non-null. This input is checked
1098 // later in PhaseOutput::filLocArray() for inline types.
1099 sfpt->add_req(_igvn.intcon(1));
1100 }
1101 }
1102
1103 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type, alloc, first_ind, sfpt->jvms()->depth(), nfields);
1104 sobj->init_req(0, C->root());
1105 transform_later(sobj);
1106
1107 if (res == nullptr) {
1108 sfpt->jvms()->set_endoff(sfpt->req());
1109 return sobj;
1110 }
1111
1112 bool success;
1113 if (iklass == nullptr) {
1114 success = add_array_elems_to_safepoint(alloc, res_type->is_aryptr(), sfpt, value_worklist);
1115 } else {
1116 success = add_inst_fields_to_safepoint(iklass, alloc, res, 0, sfpt, value_worklist);
1117 }
1118
1119 // We weren't able to find a value for this field, remove all the fields added to the safepoint
1120 if (!success) {
1121 for (uint i = sfpt->req() - 1; i >= before_sfpt_req; i--) {
1122 sfpt->del_req(i);
1123 }
1124 _igvn._worklist.push(sfpt);
1125 return nullptr;
1126 }
1127
1128 sfpt->jvms()->set_endoff(sfpt->req());
1129 return sobj;
1130 }
1131
1132 // Do scalar replacement.
1133 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
1134 GrowableArray <SafePointNode *> safepoints_done;
1135 Node* res = alloc->result_cast();
1136 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
1137 const TypeOopPtr* res_type = nullptr;
1138 if (res != nullptr) { // Could be null when there are no users
1139 res_type = _igvn.type(res)->isa_oopptr();
1140 }
1141
1142 // Process the safepoint uses
1143 Unique_Node_List value_worklist;
1144 while (safepoints.length() > 0) {
1145 SafePointNode* sfpt = safepoints.pop();
1146 SafePointScalarObjectNode* sobj = create_scalarized_object_description(alloc, sfpt, &value_worklist);
1147
1148 if (sobj == nullptr) {
1149 undo_previous_scalarizations(safepoints_done, alloc);
1150 return false;
1151 }
1152
1153 // Now make a pass over the debug information replacing any references
1154 // to the allocated object with "sobj"
1155 JVMState *jvms = sfpt->jvms();
1156 sfpt->replace_edges_in_range(res, sobj, jvms->debug_start(), jvms->debug_end(), &_igvn);
1157 _igvn._worklist.push(sfpt);
1158
1159 // keep it for rollback
1160 safepoints_done.append_if_missing(sfpt);
1161 }
1162 // Scalarize inline types that were added to the safepoint.
1163 // Don't allow linking a constant oop (if available) for flat array elements
1164 // because Deoptimization::reassign_flat_array_elements needs field values.
1165 bool allow_oop = (res_type != nullptr) && !res_type->is_flat();
1166 for (uint i = 0; i < value_worklist.size(); ++i) {
1167 InlineTypeNode* vt = value_worklist.at(i)->as_InlineType();
1168 vt->make_scalar_in_safepoints(&_igvn, allow_oop);
1169 }
1170 return true;
1171 }
1172
1173 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
1174 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control);
1175 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory);
1176 if (ctl_proj != nullptr) {
1177 igvn.replace_node(ctl_proj, n->in(0));
1178 }
1179 if (mem_proj != nullptr) {
1180 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
1181 }
1182 }
1183
1184 // Process users of eliminated allocation.
1185 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc, bool inline_alloc) {
1186 Unique_Node_List worklist;
1187 Node* res = alloc->result_cast();
1188 if (res != nullptr) {
1189 worklist.push(res);
1190 }
1191 while (worklist.size() > 0) {
1192 res = worklist.pop();
1193 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
1194 Node *use = res->last_out(j);
1195 uint oc1 = res->outcnt();
1196
1197 if (use->is_AddP()) {
1198 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
1199 Node *n = use->last_out(k);
1200 uint oc2 = use->outcnt();
1201 if (n->is_Store()) {
1202 for (DUIterator_Fast pmax, p = n->fast_outs(pmax); p < pmax; p++) {
1203 MemBarNode* mb = n->fast_out(p)->isa_MemBar();
1204 if (mb != nullptr && mb->req() <= MemBarNode::Precedent && mb->in(MemBarNode::Precedent) == n) {
1205 // MemBarVolatiles should have been removed by MemBarNode::Ideal() for non-inline allocations
1206 assert(inline_alloc, "MemBarVolatile should be eliminated for non-escaping object");
1207 mb->remove(&_igvn);
1208 }
1209 }
1210 _igvn.replace_node(n, n->in(MemNode::Memory));
1211 } else {
1212 eliminate_gc_barrier(n);
1213 }
1214 k -= (oc2 - use->outcnt());
1215 }
1216 _igvn.remove_dead_node(use);
1217 } else if (use->is_ArrayCopy()) {
1218 // Disconnect ArrayCopy node
1219 ArrayCopyNode* ac = use->as_ArrayCopy();
1220 if (ac->is_clonebasic()) {
1221 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
1222 disconnect_projections(ac, _igvn);
1223 assert(alloc->in(TypeFunc::Memory)->is_Proj() && alloc->in(TypeFunc::Memory)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
1224 Node* membar_before = alloc->in(TypeFunc::Memory)->in(0);
1225 disconnect_projections(membar_before->as_MemBar(), _igvn);
1226 if (membar_after->is_MemBar()) {
1227 disconnect_projections(membar_after->as_MemBar(), _igvn);
1228 }
1229 } else {
1230 assert(ac->is_arraycopy_validated() ||
1231 ac->is_copyof_validated() ||
1232 ac->is_copyofrange_validated(), "unsupported");
1233 CallProjections* callprojs = ac->extract_projections(true);
1234
1235 _igvn.replace_node(callprojs->fallthrough_ioproj, ac->in(TypeFunc::I_O));
1236 _igvn.replace_node(callprojs->fallthrough_memproj, ac->in(TypeFunc::Memory));
1237 _igvn.replace_node(callprojs->fallthrough_catchproj, ac->in(TypeFunc::Control));
1238
1239 // Set control to top. IGVN will remove the remaining projections
1240 ac->set_req(0, top());
1241 ac->replace_edge(res, top(), &_igvn);
1242
1243 // Disconnect src right away: it can help find new
1244 // opportunities for allocation elimination
1245 Node* src = ac->in(ArrayCopyNode::Src);
1246 ac->replace_edge(src, top(), &_igvn);
1247 // src can be top at this point if src and dest of the
1248 // arraycopy were the same
1249 if (src->outcnt() == 0 && !src->is_top()) {
1250 _igvn.remove_dead_node(src);
1251 }
1252 }
1253 _igvn._worklist.push(ac);
1254 } else if (use->is_InlineType()) {
1255 assert(use->as_InlineType()->get_oop() == res, "unexpected inline type ptr use");
1256 // Cut off oop input and remove known instance id from type
1257 _igvn.rehash_node_delayed(use);
1258 use->as_InlineType()->set_oop(_igvn, _igvn.zerocon(T_OBJECT));
1259 use->as_InlineType()->set_is_buffered(_igvn, false);
1260 const TypeOopPtr* toop = _igvn.type(use)->is_oopptr()->cast_to_instance_id(TypeOopPtr::InstanceBot);
1261 _igvn.set_type(use, toop);
1262 use->as_InlineType()->set_type(toop);
1263 // Process users
1264 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); k < kmax; k++) {
1265 Node* u = use->fast_out(k);
1266 if (!u->is_InlineType() && !u->is_StoreFlat()) {
1267 worklist.push(u);
1268 }
1269 }
1270 } else if (use->Opcode() == Op_StoreX && use->in(MemNode::Address) == res) {
1271 // Store to mark word of inline type larval buffer
1272 assert(inline_alloc, "Unexpected store to mark word");
1273 _igvn.replace_node(use, use->in(MemNode::Memory));
1274 } else if (use->Opcode() == Op_MemBarRelease || use->Opcode() == Op_MemBarStoreStore) {
1275 // Inline type buffer allocations are followed by a membar
1276 assert(inline_alloc, "Unexpected MemBarRelease");
1277 use->as_MemBar()->remove(&_igvn);
1278 } else {
1279 eliminate_gc_barrier(use);
1280 }
1281 j -= (oc1 - res->outcnt());
1282 }
1283 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
1284 _igvn.remove_dead_node(res);
1285 }
1286
1287 //
1288 // Process other users of allocation's projections
1289 //
1290 if (_callprojs->resproj[0] != nullptr && _callprojs->resproj[0]->outcnt() != 0) {
1291 // First disconnect stores captured by Initialize node.
1292 // If Initialize node is eliminated first in the following code,
1293 // it will kill such stores and DUIterator_Last will assert.
1294 for (DUIterator_Fast jmax, j = _callprojs->resproj[0]->fast_outs(jmax); j < jmax; j++) {
1295 Node* use = _callprojs->resproj[0]->fast_out(j);
1296 if (use->is_AddP()) {
1297 // raw memory addresses used only by the initialization
1298 _igvn.replace_node(use, C->top());
1299 --j; --jmax;
1300 }
1301 }
1302 for (DUIterator_Last jmin, j = _callprojs->resproj[0]->last_outs(jmin); j >= jmin; ) {
1303 Node* use = _callprojs->resproj[0]->last_out(j);
1304 uint oc1 = _callprojs->resproj[0]->outcnt();
1305 if (use->is_Initialize()) {
1306 // Eliminate Initialize node.
1307 InitializeNode *init = use->as_Initialize();
1308 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control);
1309 if (ctrl_proj != nullptr) {
1310 _igvn.replace_node(ctrl_proj, init->in(TypeFunc::Control));
1311 #ifdef ASSERT
1312 // If the InitializeNode has no memory out, it will die, and tmp will become null
1313 Node* tmp = init->in(TypeFunc::Control);
1314 assert(tmp == nullptr || tmp == _callprojs->fallthrough_catchproj, "allocation control projection");
1315 #endif
1316 }
1317 Node* mem = init->in(TypeFunc::Memory);
1318 #ifdef ASSERT
1319 if (init->number_of_projs(TypeFunc::Memory) > 0) {
1320 if (mem->is_MergeMem()) {
1321 assert(mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw) == _callprojs->fallthrough_memproj, "allocation memory projection");
1322 } else {
1323 assert(mem == _callprojs->fallthrough_memproj, "allocation memory projection");
1324 }
1325 }
1326 #endif
1327 init->replace_mem_projs_by(mem, &_igvn);
1328 assert(init->outcnt() == 0, "should only have had a control and some memory projections, and we removed them");
1329 } else if (use->Opcode() == Op_MemBarStoreStore) {
1330 // Inline type buffer allocations are followed by a membar
1331 assert(inline_alloc, "Unexpected MemBarStoreStore");
1332 use->as_MemBar()->remove(&_igvn);
1333 } else {
1334 assert(false, "only Initialize or AddP expected");
1335 }
1336 j -= (oc1 - _callprojs->resproj[0]->outcnt());
1337 }
1338 }
1339 if (_callprojs->fallthrough_catchproj != nullptr) {
1340 _igvn.replace_node(_callprojs->fallthrough_catchproj, alloc->in(TypeFunc::Control));
1341 }
1342 if (_callprojs->fallthrough_memproj != nullptr) {
1343 _igvn.replace_node(_callprojs->fallthrough_memproj, alloc->in(TypeFunc::Memory));
1344 }
1345 if (_callprojs->catchall_memproj != nullptr) {
1346 _igvn.replace_node(_callprojs->catchall_memproj, C->top());
1347 }
1348 if (_callprojs->fallthrough_ioproj != nullptr) {
1349 _igvn.replace_node(_callprojs->fallthrough_ioproj, alloc->in(TypeFunc::I_O));
1350 }
1351 if (_callprojs->catchall_ioproj != nullptr) {
1352 _igvn.replace_node(_callprojs->catchall_ioproj, C->top());
1353 }
1354 if (_callprojs->catchall_catchproj != nullptr) {
1355 _igvn.replace_node(_callprojs->catchall_catchproj, C->top());
1356 }
1357 }
1358
1359 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1360 // If reallocation fails during deoptimization we'll pop all
1361 // interpreter frames for this compiled frame and that won't play
1362 // nice with JVMTI popframe.
1363 // We avoid this issue by eager reallocation when the popframe request
1364 // is received.
1365 if (!EliminateAllocations) {
1366 return false;
1367 }
1368 Node* klass = alloc->in(AllocateNode::KlassNode);
1369 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1370
1371 // Attempt to eliminate inline type buffer allocations
1372 // regardless of usage and escape/replaceable status.
1373 bool inline_alloc = tklass->isa_instklassptr() &&
1374 tklass->is_instklassptr()->instance_klass()->is_inlinetype();
1375 if (!alloc->_is_non_escaping && !inline_alloc) {
1376 return false;
1377 }
1378 // Eliminate boxing allocations which are not used
1379 // regardless scalar replaceable status.
1380 Node* res = alloc->result_cast();
1381 bool boxing_alloc = (res == nullptr) && C->eliminate_boxing() &&
1382 tklass->isa_instklassptr() &&
1383 tklass->is_instklassptr()->instance_klass()->is_box_klass();
1384 if (!alloc->_is_scalar_replaceable && !boxing_alloc && !inline_alloc) {
1385 return false;
1386 }
1387
1388 _callprojs = alloc->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1389
1390 GrowableArray <SafePointNode *> safepoints;
1391 if (!can_eliminate_allocation(&_igvn, alloc, &safepoints)) {
1392 return false;
1393 }
1394
1395 if (!alloc->_is_scalar_replaceable) {
1396 assert(res == nullptr || inline_alloc, "sanity");
1397 // We can only eliminate allocation if all debug info references
1398 // are already replaced with SafePointScalarObject because
1399 // we can't search for a fields value without instance_id.
1400 if (safepoints.length() > 0) {
1401 return false;
1402 }
1403 }
1404
1405 if (!scalar_replacement(alloc, safepoints)) {
1406 return false;
1407 }
1408
1409 CompileLog* log = C->log();
1410 if (log != nullptr) {
1411 log->head("eliminate_allocation type='%d'",
1412 log->identify(tklass->exact_klass()));
1413 JVMState* p = alloc->jvms();
1414 while (p != nullptr) {
1415 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1416 p = p->caller();
1417 }
1418 log->tail("eliminate_allocation");
1419 }
1420
1421 process_users_of_allocation(alloc, inline_alloc);
1422
1423 #ifndef PRODUCT
1424 if (PrintEliminateAllocations) {
1425 if (alloc->is_AllocateArray())
1426 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1427 else
1428 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1429 }
1430 #endif
1431
1432 return true;
1433 }
1434
1435 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1436 // EA should remove all uses of non-escaping boxing node.
1437 if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != nullptr) {
1438 return false;
1439 }
1440
1441 assert(boxing->result_cast() == nullptr, "unexpected boxing node result");
1442
1443 _callprojs = boxing->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1444
1445 const TypeTuple* r = boxing->tf()->range_sig();
1446 assert(r->cnt() > TypeFunc::Parms, "sanity");
1447 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1448 assert(t != nullptr, "sanity");
1449
1450 CompileLog* log = C->log();
1451 if (log != nullptr) {
1452 log->head("eliminate_boxing type='%d'",
1453 log->identify(t->instance_klass()));
1454 JVMState* p = boxing->jvms();
1455 while (p != nullptr) {
1456 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1457 p = p->caller();
1458 }
1459 log->tail("eliminate_boxing");
1460 }
1461
1462 process_users_of_allocation(boxing);
1463
1464 #ifndef PRODUCT
1465 if (PrintEliminateAllocations) {
1529 // oop flavor.
1530 //
1531 //=============================================================================
1532 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1533 // Allocations bigger than this always go the slow route.
1534 // This value must be small enough that allocation attempts that need to
1535 // trigger exceptions go the slow route. Also, it must be small enough so
1536 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1537 //=============================================================================j//
1538 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1539 // The allocator will coalesce int->oop copies away. See comment in
1540 // coalesce.cpp about how this works. It depends critically on the exact
1541 // code shape produced here, so if you are changing this code shape
1542 // make sure the GC info for the heap-top is correct in and around the
1543 // slow-path call.
1544 //
1545
1546 void PhaseMacroExpand::expand_allocate_common(
1547 AllocateNode* alloc, // allocation node to be expanded
1548 Node* length, // array length for an array allocation
1549 Node* init_val, // value to initialize the array with
1550 const TypeFunc* slow_call_type, // Type of slow call
1551 address slow_call_address, // Address of slow call
1552 Node* valid_length_test // whether length is valid or not
1553 )
1554 {
1555 Node* ctrl = alloc->in(TypeFunc::Control);
1556 Node* mem = alloc->in(TypeFunc::Memory);
1557 Node* i_o = alloc->in(TypeFunc::I_O);
1558 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1559 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1560 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1561 assert(ctrl != nullptr, "must have control");
1562
1563 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1564 // they will not be used if "always_slow" is set
1565 enum { slow_result_path = 1, fast_result_path = 2 };
1566 Node *result_region = nullptr;
1567 Node *result_phi_rawmem = nullptr;
1568 Node *result_phi_rawoop = nullptr;
1569 Node *result_phi_i_o = nullptr;
1612 #endif
1613 yank_alloc_node(alloc);
1614 return;
1615 }
1616 }
1617
1618 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1619 Node *slow_region = nullptr;
1620 Node *toobig_false = ctrl;
1621
1622 // generate the initial test if necessary
1623 if (initial_slow_test != nullptr ) {
1624 assert (expand_fast_path, "Only need test if there is a fast path");
1625 slow_region = new RegionNode(3);
1626
1627 // Now make the initial failure test. Usually a too-big test but
1628 // might be a TRUE for finalizers.
1629 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1630 transform_later(toobig_iff);
1631 // Plug the failing-too-big test into the slow-path region
1632 Node* toobig_true = new IfTrueNode(toobig_iff);
1633 transform_later(toobig_true);
1634 slow_region ->init_req( too_big_or_final_path, toobig_true );
1635 toobig_false = new IfFalseNode(toobig_iff);
1636 transform_later(toobig_false);
1637 } else {
1638 // No initial test, just fall into next case
1639 assert(allocation_has_use || !expand_fast_path, "Should already have been handled");
1640 toobig_false = ctrl;
1641 DEBUG_ONLY(slow_region = NodeSentinel);
1642 }
1643
1644 // If we are here there are several possibilities
1645 // - expand_fast_path is false - then only a slow path is expanded. That's it.
1646 // no_initial_check means a constant allocation.
1647 // - If check always evaluates to false -> expand_fast_path is false (see above)
1648 // - If check always evaluates to true -> directly into fast path (but may bailout to slowpath)
1649 // if !allocation_has_use the fast path is empty
1650 // if !allocation_has_use && no_initial_check
1651 // - Then there are no fastpath that can fall out to slowpath -> no allocation code at all.
1652 // removed by yank_alloc_node above.
1653
1654 Node *slow_mem = mem; // save the current memory state for slow path
1655 // generate the fast allocation code unless we know that the initial test will always go slow
1656 if (expand_fast_path) {
1657 // Fast path modifies only raw memory.
1658 if (mem->is_MergeMem()) {
1659 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1660 }
1661
1662 // allocate the Region and Phi nodes for the result
1663 result_region = new RegionNode(3);
1664 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1665 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1666
1667 // Grab regular I/O before optional prefetch may change it.
1668 // Slow-path does no I/O so just set it to the original I/O.
1669 result_phi_i_o->init_req(slow_result_path, i_o);
1670
1671 // Name successful fast-path variables
1672 Node* fast_oop_ctrl;
1673 Node* fast_oop_rawmem;
1674
1675 if (allocation_has_use) {
1676 Node* needgc_ctrl = nullptr;
1677 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1678
1679 intx prefetch_lines = length != nullptr ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1680 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1681 Node* fast_oop = bs->obj_allocate(this, mem, toobig_false, size_in_bytes, i_o, needgc_ctrl,
1682 fast_oop_ctrl, fast_oop_rawmem,
1683 prefetch_lines);
1684
1685 if (initial_slow_test != nullptr) {
1686 // This completes all paths into the slow merge point
1687 slow_region->init_req(need_gc_path, needgc_ctrl);
1688 transform_later(slow_region);
1689 } else {
1690 // No initial slow path needed!
1691 // Just fall from the need-GC path straight into the VM call.
1692 slow_region = needgc_ctrl;
1693 }
1694
1712 result_phi_i_o ->init_req(fast_result_path, i_o);
1713 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1714 } else {
1715 slow_region = ctrl;
1716 result_phi_i_o = i_o; // Rename it to use in the following code.
1717 }
1718
1719 // Generate slow-path call
1720 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1721 OptoRuntime::stub_name(slow_call_address),
1722 TypePtr::BOTTOM);
1723 call->init_req(TypeFunc::Control, slow_region);
1724 call->init_req(TypeFunc::I_O, top()); // does no i/o
1725 call->init_req(TypeFunc::Memory, slow_mem); // may gc ptrs
1726 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1727 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1728
1729 call->init_req(TypeFunc::Parms+0, klass_node);
1730 if (length != nullptr) {
1731 call->init_req(TypeFunc::Parms+1, length);
1732 if (init_val != nullptr) {
1733 call->init_req(TypeFunc::Parms+2, init_val);
1734 }
1735 } else {
1736 // Let the runtime know if this is a larval allocation
1737 call->init_req(TypeFunc::Parms+1, _igvn.intcon(alloc->_larval));
1738 }
1739
1740 // Copy debug information and adjust JVMState information, then replace
1741 // allocate node with the call
1742 call->copy_call_debug_info(&_igvn, alloc);
1743 // For array allocations, copy the valid length check to the call node so Compile::final_graph_reshaping() can verify
1744 // that the call has the expected number of CatchProj nodes (in case the allocation always fails and the fallthrough
1745 // path dies).
1746 if (valid_length_test != nullptr) {
1747 call->add_req(valid_length_test);
1748 }
1749 if (expand_fast_path) {
1750 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1751 } else {
1752 // Hook i_o projection to avoid its elimination during allocation
1753 // replacement (when only a slow call is generated).
1754 call->set_req(TypeFunc::I_O, result_phi_i_o);
1755 }
1756 _igvn.replace_node(alloc, call);
1757 transform_later(call);
1758
1759 // Identify the output projections from the allocate node and
1760 // adjust any references to them.
1761 // The control and io projections look like:
1762 //
1763 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1764 // Allocate Catch
1765 // ^---Proj(io) <-------+ ^---CatchProj(io)
1766 //
1767 // We are interested in the CatchProj nodes.
1768 //
1769 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1770
1771 // An allocate node has separate memory projections for the uses on
1772 // the control and i_o paths. Replace the control memory projection with
1773 // result_phi_rawmem (unless we are only generating a slow call when
1774 // both memory projections are combined)
1775 if (expand_fast_path && _callprojs->fallthrough_memproj != nullptr) {
1776 _igvn.replace_in_uses(_callprojs->fallthrough_memproj, result_phi_rawmem);
1777 }
1778 // Now change uses of catchall_memproj to use fallthrough_memproj and delete
1779 // catchall_memproj so we end up with a call that has only 1 memory projection.
1780 if (_callprojs->catchall_memproj != nullptr) {
1781 if (_callprojs->fallthrough_memproj == nullptr) {
1782 _callprojs->fallthrough_memproj = new ProjNode(call, TypeFunc::Memory);
1783 transform_later(_callprojs->fallthrough_memproj);
1784 }
1785 _igvn.replace_in_uses(_callprojs->catchall_memproj, _callprojs->fallthrough_memproj);
1786 _igvn.remove_dead_node(_callprojs->catchall_memproj);
1787 }
1788
1789 // An allocate node has separate i_o projections for the uses on the control
1790 // and i_o paths. Always replace the control i_o projection with result i_o
1791 // otherwise incoming i_o become dead when only a slow call is generated
1792 // (it is different from memory projections where both projections are
1793 // combined in such case).
1794 if (_callprojs->fallthrough_ioproj != nullptr) {
1795 _igvn.replace_in_uses(_callprojs->fallthrough_ioproj, result_phi_i_o);
1796 }
1797 // Now change uses of catchall_ioproj to use fallthrough_ioproj and delete
1798 // catchall_ioproj so we end up with a call that has only 1 i_o projection.
1799 if (_callprojs->catchall_ioproj != nullptr) {
1800 if (_callprojs->fallthrough_ioproj == nullptr) {
1801 _callprojs->fallthrough_ioproj = new ProjNode(call, TypeFunc::I_O);
1802 transform_later(_callprojs->fallthrough_ioproj);
1803 }
1804 _igvn.replace_in_uses(_callprojs->catchall_ioproj, _callprojs->fallthrough_ioproj);
1805 _igvn.remove_dead_node(_callprojs->catchall_ioproj);
1806 }
1807
1808 // if we generated only a slow call, we are done
1809 if (!expand_fast_path) {
1810 // Now we can unhook i_o.
1811 if (result_phi_i_o->outcnt() > 1) {
1812 call->set_req(TypeFunc::I_O, top());
1813 } else {
1814 assert(result_phi_i_o->unique_ctrl_out() == call, "sanity");
1815 // Case of new array with negative size known during compilation.
1816 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1817 // following code since call to runtime will throw exception.
1818 // As result there will be no users of i_o after the call.
1819 // Leave i_o attached to this call to avoid problems in preceding graph.
1820 }
1821 return;
1822 }
1823
1824 if (_callprojs->fallthrough_catchproj != nullptr) {
1825 ctrl = _callprojs->fallthrough_catchproj->clone();
1826 transform_later(ctrl);
1827 _igvn.replace_node(_callprojs->fallthrough_catchproj, result_region);
1828 } else {
1829 ctrl = top();
1830 }
1831 Node *slow_result;
1832 if (_callprojs->resproj[0] == nullptr) {
1833 // no uses of the allocation result
1834 slow_result = top();
1835 } else {
1836 slow_result = _callprojs->resproj[0]->clone();
1837 transform_later(slow_result);
1838 _igvn.replace_node(_callprojs->resproj[0], result_phi_rawoop);
1839 }
1840
1841 // Plug slow-path into result merge point
1842 result_region->init_req( slow_result_path, ctrl);
1843 transform_later(result_region);
1844 if (allocation_has_use) {
1845 result_phi_rawoop->init_req(slow_result_path, slow_result);
1846 transform_later(result_phi_rawoop);
1847 }
1848 result_phi_rawmem->init_req(slow_result_path, _callprojs->fallthrough_memproj);
1849 transform_later(result_phi_rawmem);
1850 transform_later(result_phi_i_o);
1851 // This completes all paths into the result merge point
1852 }
1853
1854 // Remove alloc node that has no uses.
1855 void PhaseMacroExpand::yank_alloc_node(AllocateNode* alloc) {
1856 Node* ctrl = alloc->in(TypeFunc::Control);
1857 Node* mem = alloc->in(TypeFunc::Memory);
1858 Node* i_o = alloc->in(TypeFunc::I_O);
1859
1860 _callprojs = alloc->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1861 if (_callprojs->resproj[0] != nullptr) {
1862 for (DUIterator_Fast imax, i = _callprojs->resproj[0]->fast_outs(imax); i < imax; i++) {
1863 Node* use = _callprojs->resproj[0]->fast_out(i);
1864 use->isa_MemBar()->remove(&_igvn);
1865 --imax;
1866 --i; // back up iterator
1867 }
1868 assert(_callprojs->resproj[0]->outcnt() == 0, "all uses must be deleted");
1869 _igvn.remove_dead_node(_callprojs->resproj[0]);
1870 }
1871 if (_callprojs->fallthrough_catchproj != nullptr) {
1872 _igvn.replace_in_uses(_callprojs->fallthrough_catchproj, ctrl);
1873 _igvn.remove_dead_node(_callprojs->fallthrough_catchproj);
1874 }
1875 if (_callprojs->catchall_catchproj != nullptr) {
1876 _igvn.rehash_node_delayed(_callprojs->catchall_catchproj);
1877 _callprojs->catchall_catchproj->set_req(0, top());
1878 }
1879 if (_callprojs->fallthrough_proj != nullptr) {
1880 Node* catchnode = _callprojs->fallthrough_proj->unique_ctrl_out();
1881 _igvn.remove_dead_node(catchnode);
1882 _igvn.remove_dead_node(_callprojs->fallthrough_proj);
1883 }
1884 if (_callprojs->fallthrough_memproj != nullptr) {
1885 _igvn.replace_in_uses(_callprojs->fallthrough_memproj, mem);
1886 _igvn.remove_dead_node(_callprojs->fallthrough_memproj);
1887 }
1888 if (_callprojs->fallthrough_ioproj != nullptr) {
1889 _igvn.replace_in_uses(_callprojs->fallthrough_ioproj, i_o);
1890 _igvn.remove_dead_node(_callprojs->fallthrough_ioproj);
1891 }
1892 if (_callprojs->catchall_memproj != nullptr) {
1893 _igvn.rehash_node_delayed(_callprojs->catchall_memproj);
1894 _callprojs->catchall_memproj->set_req(0, top());
1895 }
1896 if (_callprojs->catchall_ioproj != nullptr) {
1897 _igvn.rehash_node_delayed(_callprojs->catchall_ioproj);
1898 _callprojs->catchall_ioproj->set_req(0, top());
1899 }
1900 #ifndef PRODUCT
1901 if (PrintEliminateAllocations) {
1902 if (alloc->is_AllocateArray()) {
1903 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1904 } else {
1905 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1906 }
1907 }
1908 #endif
1909 _igvn.remove_dead_node(alloc);
1910 }
1911
1912 void PhaseMacroExpand::expand_initialize_membar(AllocateNode* alloc, InitializeNode* init,
1913 Node*& fast_oop_ctrl, Node*& fast_oop_rawmem) {
1914 // If initialization is performed by an array copy, any required
1915 // MemBarStoreStore was already added. If the object does not
1916 // escape no need for a MemBarStoreStore. If the object does not
1917 // escape in its initializer and memory barrier (MemBarStoreStore or
1918 // stronger) is already added at exit of initializer, also no need
2012 Node* thread = new ThreadLocalNode();
2013 transform_later(thread);
2014
2015 call->init_req(TypeFunc::Parms + 0, thread);
2016 call->init_req(TypeFunc::Parms + 1, oop);
2017 call->init_req(TypeFunc::Control, ctrl);
2018 call->init_req(TypeFunc::I_O , top()); // does no i/o
2019 call->init_req(TypeFunc::Memory , rawmem);
2020 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
2021 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
2022 transform_later(call);
2023 ctrl = new ProjNode(call, TypeFunc::Control);
2024 transform_later(ctrl);
2025 rawmem = new ProjNode(call, TypeFunc::Memory);
2026 transform_later(rawmem);
2027 }
2028 }
2029
2030 // Helper for PhaseMacroExpand::expand_allocate_common.
2031 // Initializes the newly-allocated storage.
2032 Node* PhaseMacroExpand::initialize_object(AllocateNode* alloc,
2033 Node* control, Node* rawmem, Node* object,
2034 Node* klass_node, Node* length,
2035 Node* size_in_bytes) {
2036 InitializeNode* init = alloc->initialization();
2037 // Store the klass & mark bits
2038 Node* mark_node = alloc->make_ideal_mark(&_igvn, control, rawmem);
2039 if (!mark_node->is_Con()) {
2040 transform_later(mark_node);
2041 }
2042 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, TypeX_X->basic_type());
2043
2044 if (!UseCompactObjectHeaders) {
2045 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
2046 }
2047 int header_size = alloc->minimum_header_size(); // conservatively small
2048
2049 // Array length
2050 if (length != nullptr) { // Arrays need length field
2051 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
2052 // conservatively small header size:
2053 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
2054 if (_igvn.type(klass_node)->isa_aryklassptr()) { // we know the exact header size in most cases:
2055 BasicType elem = _igvn.type(klass_node)->is_klassptr()->as_instance_type()->isa_aryptr()->elem()->array_element_basic_type();
2056 if (is_reference_type(elem, true)) {
2057 elem = T_OBJECT;
2058 }
2059 header_size = Klass::layout_helper_header_size(Klass::array_layout_helper(elem));
2060 }
2061 }
2062
2063 // Clear the object body, if necessary.
2064 if (init == nullptr) {
2065 // The init has somehow disappeared; be cautious and clear everything.
2066 //
2067 // This can happen if a node is allocated but an uncommon trap occurs
2068 // immediately. In this case, the Initialize gets associated with the
2069 // trap, and may be placed in a different (outer) loop, if the Allocate
2070 // is in a loop. If (this is rare) the inner loop gets unrolled, then
2071 // there can be two Allocates to one Initialize. The answer in all these
2072 // edge cases is safety first. It is always safe to clear immediately
2073 // within an Allocate, and then (maybe or maybe not) clear some more later.
2074 if (!(UseTLAB && ZeroTLAB)) {
2075 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
2076 alloc->in(AllocateNode::InitValue),
2077 alloc->in(AllocateNode::RawInitValue),
2078 header_size, size_in_bytes,
2079 &_igvn);
2080 }
2081 } else {
2082 if (!init->is_complete()) {
2083 // Try to win by zeroing only what the init does not store.
2084 // We can also try to do some peephole optimizations,
2085 // such as combining some adjacent subword stores.
2086 rawmem = init->complete_stores(control, rawmem, object,
2087 header_size, size_in_bytes, &_igvn);
2088 }
2089 // We have no more use for this link, since the AllocateNode goes away:
2090 init->set_req(InitializeNode::RawAddress, top());
2091 // (If we keep the link, it just confuses the register allocator,
2092 // who thinks he sees a real use of the address by the membar.)
2093 }
2094
2095 return rawmem;
2096 }
2097
2232 for ( intx i = 0; i < lines; i++ ) {
2233 prefetch_adr = new AddPNode( old_eden_top, new_eden_top,
2234 _igvn.MakeConX(distance) );
2235 transform_later(prefetch_adr);
2236 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
2237 // Do not let it float too high, since if eden_top == eden_end,
2238 // both might be null.
2239 if( i == 0 ) { // Set control for first prefetch, next follows it
2240 prefetch->init_req(0, needgc_false);
2241 }
2242 transform_later(prefetch);
2243 distance += step_size;
2244 i_o = prefetch;
2245 }
2246 }
2247 return i_o;
2248 }
2249
2250
2251 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
2252 expand_allocate_common(alloc, nullptr, nullptr,
2253 OptoRuntime::new_instance_Type(),
2254 OptoRuntime::new_instance_Java(), nullptr);
2255 }
2256
2257 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
2258 Node* length = alloc->in(AllocateNode::ALength);
2259 Node* valid_length_test = alloc->in(AllocateNode::ValidLengthTest);
2260 InitializeNode* init = alloc->initialization();
2261 Node* klass_node = alloc->in(AllocateNode::KlassNode);
2262 Node* init_value = alloc->in(AllocateNode::InitValue);
2263 const TypeAryKlassPtr* ary_klass_t = _igvn.type(klass_node)->isa_aryklassptr();
2264 assert(!ary_klass_t || !ary_klass_t->klass_is_exact() || !ary_klass_t->exact_klass()->is_obj_array_klass() ||
2265 ary_klass_t->is_refined_type(), "Must be a refined array klass");
2266 const TypeFunc* slow_call_type;
2267 address slow_call_address; // Address of slow call
2268 if (init != nullptr && init->is_complete_with_arraycopy() &&
2269 ary_klass_t && ary_klass_t->elem()->isa_klassptr() == nullptr) {
2270 // Don't zero type array during slow allocation in VM since
2271 // it will be initialized later by arraycopy in compiled code.
2272 slow_call_address = OptoRuntime::new_array_nozero_Java();
2273 slow_call_type = OptoRuntime::new_array_nozero_Type();
2274 } else {
2275 slow_call_address = OptoRuntime::new_array_Java();
2276 slow_call_type = OptoRuntime::new_array_Type();
2277
2278 if (init_value == nullptr) {
2279 init_value = _igvn.zerocon(T_OBJECT);
2280 } else if (UseCompressedOops) {
2281 init_value = transform_later(new DecodeNNode(init_value, init_value->bottom_type()->make_ptr()));
2282 }
2283 }
2284 expand_allocate_common(alloc, length, init_value,
2285 slow_call_type,
2286 slow_call_address, valid_length_test);
2287 }
2288
2289 //-------------------mark_eliminated_box----------------------------------
2290 //
2291 // During EA obj may point to several objects but after few ideal graph
2292 // transformations (CCP) it may point to only one non escaping object
2293 // (but still using phi), corresponding locks and unlocks will be marked
2294 // for elimination. Later obj could be replaced with a new node (new phi)
2295 // and which does not have escape information. And later after some graph
2296 // reshape other locks and unlocks (which were not marked for elimination
2297 // before) are connected to this new obj (phi) but they still will not be
2298 // marked for elimination since new obj has no escape information.
2299 // Mark all associated (same box and obj) lock and unlock nodes for
2300 // elimination if some of them marked already.
2301 void PhaseMacroExpand::mark_eliminated_box(Node* box, Node* obj) {
2302 BoxLockNode* oldbox = box->as_BoxLock();
2303 if (oldbox->is_eliminated()) {
2304 return; // This BoxLock node was processed already.
2305 }
2477 #ifdef ASSERT
2478 if (!alock->is_coarsened()) {
2479 // Check that new "eliminated" BoxLock node is created.
2480 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2481 assert(oldbox->is_eliminated(), "should be done already");
2482 }
2483 #endif
2484
2485 alock->log_lock_optimization(C, "eliminate_lock");
2486
2487 #ifndef PRODUCT
2488 if (PrintEliminateLocks) {
2489 tty->print_cr("++++ Eliminated: %d %s '%s'", alock->_idx, (alock->is_Lock() ? "Lock" : "Unlock"), alock->kind_as_string());
2490 }
2491 #endif
2492
2493 Node* mem = alock->in(TypeFunc::Memory);
2494 Node* ctrl = alock->in(TypeFunc::Control);
2495 guarantee(ctrl != nullptr, "missing control projection, cannot replace_node() with null");
2496
2497 _callprojs = alock->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2498 // There are 2 projections from the lock. The lock node will
2499 // be deleted when its last use is subsumed below.
2500 assert(alock->outcnt() == 2 &&
2501 _callprojs->fallthrough_proj != nullptr &&
2502 _callprojs->fallthrough_memproj != nullptr,
2503 "Unexpected projections from Lock/Unlock");
2504
2505 Node* fallthroughproj = _callprojs->fallthrough_proj;
2506 Node* memproj_fallthrough = _callprojs->fallthrough_memproj;
2507
2508 // The memory projection from a lock/unlock is RawMem
2509 // The input to a Lock is merged memory, so extract its RawMem input
2510 // (unless the MergeMem has been optimized away.)
2511 if (alock->is_Lock()) {
2512 // Search for MemBarAcquireLock node and delete it also.
2513 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2514 assert(membar != nullptr && membar->Opcode() == Op_MemBarAcquireLock, "");
2515 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2516 Node* memproj = membar->proj_out(TypeFunc::Memory);
2517 _igvn.replace_node(ctrlproj, fallthroughproj);
2518 _igvn.replace_node(memproj, memproj_fallthrough);
2519
2520 // Delete FastLock node also if this Lock node is unique user
2521 // (a loop peeling may clone a Lock node).
2522 Node* flock = alock->as_Lock()->fastlock_node();
2523 if (flock->outcnt() == 1) {
2524 assert(flock->unique_out() == alock, "sanity");
2525 _igvn.replace_node(flock, top());
2526 }
2557 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2558
2559 // Make the merge point
2560 Node *region;
2561 Node *mem_phi;
2562 Node *slow_path;
2563
2564 region = new RegionNode(3);
2565 // create a Phi for the memory state
2566 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2567
2568 // Optimize test; set region slot 2
2569 slow_path = opt_bits_test(ctrl, region, 2, flock);
2570 mem_phi->init_req(2, mem);
2571
2572 // Make slow path call
2573 CallNode* call = make_slow_call(lock, OptoRuntime::complete_monitor_enter_Type(),
2574 OptoRuntime::complete_monitor_locking_Java(), nullptr, slow_path,
2575 obj, box, nullptr);
2576
2577 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2578
2579 // Slow path can only throw asynchronous exceptions, which are always
2580 // de-opted. So the compiler thinks the slow-call can never throw an
2581 // exception. If it DOES throw an exception we would need the debug
2582 // info removed first (since if it throws there is no monitor).
2583 assert(_callprojs->fallthrough_ioproj == nullptr && _callprojs->catchall_ioproj == nullptr &&
2584 _callprojs->catchall_memproj == nullptr && _callprojs->catchall_catchproj == nullptr, "Unexpected projection from Lock");
2585
2586 // Capture slow path
2587 // disconnect fall-through projection from call and create a new one
2588 // hook up users of fall-through projection to region
2589 Node *slow_ctrl = _callprojs->fallthrough_proj->clone();
2590 transform_later(slow_ctrl);
2591 _igvn.hash_delete(_callprojs->fallthrough_proj);
2592 _callprojs->fallthrough_proj->disconnect_inputs(C);
2593 region->init_req(1, slow_ctrl);
2594 // region inputs are now complete
2595 transform_later(region);
2596 _igvn.replace_node(_callprojs->fallthrough_proj, region);
2597
2598 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2599
2600 mem_phi->init_req(1, memproj);
2601
2602 transform_later(mem_phi);
2603
2604 _igvn.replace_node(_callprojs->fallthrough_memproj, mem_phi);
2605 }
2606
2607 //------------------------------expand_unlock_node----------------------
2608 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2609
2610 Node* ctrl = unlock->in(TypeFunc::Control);
2611 Node* mem = unlock->in(TypeFunc::Memory);
2612 Node* obj = unlock->obj_node();
2613 Node* box = unlock->box_node();
2614
2615 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2616
2617 // No need for a null check on unlock
2618
2619 // Make the merge point
2620 Node *region;
2621 Node *mem_phi;
2622
2623 region = new RegionNode(3);
2624 // create a Phi for the memory state
2625 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2626
2627 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2628 funlock = transform_later( funlock )->as_FastUnlock();
2629 // Optimize test; set region slot 2
2630 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock);
2631 Node *thread = transform_later(new ThreadLocalNode());
2632
2633 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2634 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2635 "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2636
2637 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2638 assert(_callprojs->fallthrough_ioproj == nullptr && _callprojs->catchall_ioproj == nullptr &&
2639 _callprojs->catchall_memproj == nullptr && _callprojs->catchall_catchproj == nullptr, "Unexpected projection from Lock");
2640
2641 // No exceptions for unlocking
2642 // Capture slow path
2643 // disconnect fall-through projection from call and create a new one
2644 // hook up users of fall-through projection to region
2645 Node *slow_ctrl = _callprojs->fallthrough_proj->clone();
2646 transform_later(slow_ctrl);
2647 _igvn.hash_delete(_callprojs->fallthrough_proj);
2648 _callprojs->fallthrough_proj->disconnect_inputs(C);
2649 region->init_req(1, slow_ctrl);
2650 // region inputs are now complete
2651 transform_later(region);
2652 _igvn.replace_node(_callprojs->fallthrough_proj, region);
2653
2654 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2655 mem_phi->init_req(1, memproj );
2656 mem_phi->init_req(2, mem);
2657 transform_later(mem_phi);
2658
2659 _igvn.replace_node(_callprojs->fallthrough_memproj, mem_phi);
2660 }
2661
2662 // An inline type might be returned from the call but we don't know its
2663 // type. Either we get a buffered inline type (and nothing needs to be done)
2664 // or one of the values being returned is the klass of the inline type
2665 // and we need to allocate an inline type instance of that type and
2666 // initialize it with other values being returned. In that case, we
2667 // first try a fast path allocation and initialize the value with the
2668 // inline klass's pack handler or we fall back to a runtime call.
2669 void PhaseMacroExpand::expand_mh_intrinsic_return(CallStaticJavaNode* call) {
2670 assert(call->method()->is_method_handle_intrinsic(), "must be a method handle intrinsic call");
2671 Node* ret = call->proj_out_or_null(TypeFunc::Parms);
2672 if (ret == nullptr) {
2673 return;
2674 }
2675 const TypeFunc* tf = call->_tf;
2676 const TypeTuple* domain = OptoRuntime::store_inline_type_fields_Type()->domain_cc();
2677 const TypeFunc* new_tf = TypeFunc::make(tf->domain_sig(), tf->domain_cc(), tf->range_sig(), domain);
2678 call->_tf = new_tf;
2679 // Make sure the change of type is applied before projections are processed by igvn
2680 _igvn.set_type(call, call->Value(&_igvn));
2681 _igvn.set_type(ret, ret->Value(&_igvn));
2682
2683 // Before any new projection is added:
2684 CallProjections* projs = call->extract_projections(true, true);
2685
2686 // Create temporary hook nodes that will be replaced below.
2687 // Add an input to prevent hook nodes from being dead.
2688 Node* ctl = new Node(call);
2689 Node* mem = new Node(ctl);
2690 Node* io = new Node(ctl);
2691 Node* ex_ctl = new Node(ctl);
2692 Node* ex_mem = new Node(ctl);
2693 Node* ex_io = new Node(ctl);
2694 Node* res = new Node(ctl);
2695
2696 // Allocate a new buffered inline type only if a new one is not returned
2697 Node* cast = transform_later(new CastP2XNode(ctl, res));
2698 Node* mask = MakeConX(0x1);
2699 Node* masked = transform_later(new AndXNode(cast, mask));
2700 Node* cmp = transform_later(new CmpXNode(masked, mask));
2701 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
2702 IfNode* allocation_iff = new IfNode(ctl, bol, PROB_MAX, COUNT_UNKNOWN);
2703 transform_later(allocation_iff);
2704 Node* allocation_ctl = transform_later(new IfTrueNode(allocation_iff));
2705 Node* no_allocation_ctl = transform_later(new IfFalseNode(allocation_iff));
2706 Node* no_allocation_res = transform_later(new CheckCastPPNode(no_allocation_ctl, res, TypeInstPtr::BOTTOM));
2707
2708 // Try to allocate a new buffered inline instance either from TLAB or eden space
2709 Node* needgc_ctrl = nullptr; // needgc means slowcase, i.e. allocation failed
2710 CallLeafNoFPNode* handler_call;
2711 const bool alloc_in_place = UseTLAB;
2712 if (alloc_in_place) {
2713 Node* fast_oop_ctrl = nullptr;
2714 Node* fast_oop_rawmem = nullptr;
2715 Node* mask2 = MakeConX(-2);
2716 Node* masked2 = transform_later(new AndXNode(cast, mask2));
2717 Node* rawklassptr = transform_later(new CastX2PNode(masked2));
2718 Node* klass_node = transform_later(new CheckCastPPNode(allocation_ctl, rawklassptr, TypeInstKlassPtr::OBJECT_OR_NULL));
2719 Node* layout_val = make_load(nullptr, mem, klass_node, in_bytes(Klass::layout_helper_offset()), TypeInt::INT, T_INT);
2720 Node* size_in_bytes = ConvI2X(layout_val);
2721 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2722 Node* fast_oop = bs->obj_allocate(this, mem, allocation_ctl, size_in_bytes, io, needgc_ctrl,
2723 fast_oop_ctrl, fast_oop_rawmem,
2724 AllocateInstancePrefetchLines);
2725 // Allocation succeed, initialize buffered inline instance header firstly,
2726 // and then initialize its fields with an inline class specific handler
2727 Node* mark_word_node;
2728 if (UseCompactObjectHeaders) {
2729 // COH: We need to load the prototype from the klass at runtime since it encodes the klass pointer already.
2730 mark_word_node = make_load(fast_oop_ctrl, fast_oop_rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
2731 } else {
2732 // Otherwise, use the static prototype.
2733 mark_word_node = makecon(TypeRawPtr::make((address)markWord::inline_type_prototype().value()));
2734 }
2735
2736 fast_oop_rawmem = make_store(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::mark_offset_in_bytes(), mark_word_node, T_ADDRESS);
2737 if (!UseCompactObjectHeaders) {
2738 // COH: Everything is encoded in the mark word, so nothing left to do.
2739 fast_oop_rawmem = make_store(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
2740 if (UseCompressedClassPointers) {
2741 fast_oop_rawmem = make_store(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::klass_gap_offset_in_bytes(), intcon(0), T_INT);
2742 }
2743 }
2744 Node* fixed_block = make_load(fast_oop_ctrl, fast_oop_rawmem, klass_node, in_bytes(InstanceKlass::adr_inlineklass_fixed_block_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
2745 Node* pack_handler = make_load(fast_oop_ctrl, fast_oop_rawmem, fixed_block, in_bytes(InlineKlass::pack_handler_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
2746 handler_call = new CallLeafNoFPNode(OptoRuntime::pack_inline_type_Type(),
2747 nullptr,
2748 "pack handler",
2749 TypeRawPtr::BOTTOM);
2750 handler_call->init_req(TypeFunc::Control, fast_oop_ctrl);
2751 handler_call->init_req(TypeFunc::Memory, fast_oop_rawmem);
2752 handler_call->init_req(TypeFunc::I_O, top());
2753 handler_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2754 handler_call->init_req(TypeFunc::ReturnAdr, top());
2755 handler_call->init_req(TypeFunc::Parms, pack_handler);
2756 handler_call->init_req(TypeFunc::Parms+1, fast_oop);
2757 } else {
2758 needgc_ctrl = allocation_ctl;
2759 }
2760
2761 // Allocation failed, fall back to a runtime call
2762 CallStaticJavaNode* slow_call = new CallStaticJavaNode(OptoRuntime::store_inline_type_fields_Type(),
2763 StubRoutines::store_inline_type_fields_to_buf(),
2764 "store_inline_type_fields",
2765 TypePtr::BOTTOM);
2766 slow_call->init_req(TypeFunc::Control, needgc_ctrl);
2767 slow_call->init_req(TypeFunc::Memory, mem);
2768 slow_call->init_req(TypeFunc::I_O, io);
2769 slow_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2770 slow_call->init_req(TypeFunc::ReturnAdr, call->in(TypeFunc::ReturnAdr));
2771 slow_call->init_req(TypeFunc::Parms, res);
2772
2773 Node* slow_ctl = transform_later(new ProjNode(slow_call, TypeFunc::Control));
2774 Node* slow_mem = transform_later(new ProjNode(slow_call, TypeFunc::Memory));
2775 Node* slow_io = transform_later(new ProjNode(slow_call, TypeFunc::I_O));
2776 Node* slow_res = transform_later(new ProjNode(slow_call, TypeFunc::Parms));
2777 Node* slow_catc = transform_later(new CatchNode(slow_ctl, slow_io, 2));
2778 Node* slow_norm = transform_later(new CatchProjNode(slow_catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci));
2779 Node* slow_excp = transform_later(new CatchProjNode(slow_catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci));
2780
2781 Node* ex_r = new RegionNode(3);
2782 Node* ex_mem_phi = new PhiNode(ex_r, Type::MEMORY, TypePtr::BOTTOM);
2783 Node* ex_io_phi = new PhiNode(ex_r, Type::ABIO);
2784 ex_r->init_req(1, slow_excp);
2785 ex_mem_phi->init_req(1, slow_mem);
2786 ex_io_phi->init_req(1, slow_io);
2787 ex_r->init_req(2, ex_ctl);
2788 ex_mem_phi->init_req(2, ex_mem);
2789 ex_io_phi->init_req(2, ex_io);
2790 transform_later(ex_r);
2791 transform_later(ex_mem_phi);
2792 transform_later(ex_io_phi);
2793
2794 // We don't know how many values are returned. This assumes the
2795 // worst case, that all available registers are used.
2796 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) {
2797 if (domain->field_at(i) == Type::HALF) {
2798 slow_call->init_req(i, top());
2799 if (alloc_in_place) {
2800 handler_call->init_req(i+1, top());
2801 }
2802 continue;
2803 }
2804 Node* proj = transform_later(new ProjNode(call, i));
2805 slow_call->init_req(i, proj);
2806 if (alloc_in_place) {
2807 handler_call->init_req(i+1, proj);
2808 }
2809 }
2810 // We can safepoint at that new call
2811 slow_call->copy_call_debug_info(&_igvn, call);
2812 transform_later(slow_call);
2813 if (alloc_in_place) {
2814 transform_later(handler_call);
2815 }
2816
2817 Node* fast_ctl = nullptr;
2818 Node* fast_res = nullptr;
2819 MergeMemNode* fast_mem = nullptr;
2820 if (alloc_in_place) {
2821 fast_ctl = transform_later(new ProjNode(handler_call, TypeFunc::Control));
2822 Node* rawmem = transform_later(new ProjNode(handler_call, TypeFunc::Memory));
2823 fast_res = transform_later(new ProjNode(handler_call, TypeFunc::Parms));
2824 fast_mem = MergeMemNode::make(mem);
2825 fast_mem->set_memory_at(Compile::AliasIdxRaw, rawmem);
2826 transform_later(fast_mem);
2827 }
2828
2829 Node* r = new RegionNode(alloc_in_place ? 4 : 3);
2830 Node* mem_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM);
2831 Node* io_phi = new PhiNode(r, Type::ABIO);
2832 Node* res_phi = new PhiNode(r, TypeInstPtr::BOTTOM);
2833 r->init_req(1, no_allocation_ctl);
2834 mem_phi->init_req(1, mem);
2835 io_phi->init_req(1, io);
2836 res_phi->init_req(1, no_allocation_res);
2837 r->init_req(2, slow_norm);
2838 mem_phi->init_req(2, slow_mem);
2839 io_phi->init_req(2, slow_io);
2840 res_phi->init_req(2, slow_res);
2841 if (alloc_in_place) {
2842 r->init_req(3, fast_ctl);
2843 mem_phi->init_req(3, fast_mem);
2844 io_phi->init_req(3, io);
2845 res_phi->init_req(3, fast_res);
2846 }
2847 transform_later(r);
2848 transform_later(mem_phi);
2849 transform_later(io_phi);
2850 transform_later(res_phi);
2851
2852 // Do not let stores that initialize this buffer be reordered with a subsequent
2853 // store that would make this buffer accessible by other threads.
2854 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
2855 transform_later(mb);
2856 mb->init_req(TypeFunc::Memory, mem_phi);
2857 mb->init_req(TypeFunc::Control, r);
2858 r = new ProjNode(mb, TypeFunc::Control);
2859 transform_later(r);
2860 mem_phi = new ProjNode(mb, TypeFunc::Memory);
2861 transform_later(mem_phi);
2862
2863 assert(projs->nb_resproj == 1, "unexpected number of results");
2864 _igvn.replace_in_uses(projs->fallthrough_catchproj, r);
2865 _igvn.replace_in_uses(projs->fallthrough_memproj, mem_phi);
2866 _igvn.replace_in_uses(projs->fallthrough_ioproj, io_phi);
2867 _igvn.replace_in_uses(projs->resproj[0], res_phi);
2868 _igvn.replace_in_uses(projs->catchall_catchproj, ex_r);
2869 _igvn.replace_in_uses(projs->catchall_memproj, ex_mem_phi);
2870 _igvn.replace_in_uses(projs->catchall_ioproj, ex_io_phi);
2871 // The CatchNode should not use the ex_io_phi. Re-connect it to the catchall_ioproj.
2872 Node* cn = projs->fallthrough_catchproj->in(0);
2873 _igvn.replace_input_of(cn, 1, projs->catchall_ioproj);
2874
2875 _igvn.replace_node(ctl, projs->fallthrough_catchproj);
2876 _igvn.replace_node(mem, projs->fallthrough_memproj);
2877 _igvn.replace_node(io, projs->fallthrough_ioproj);
2878 _igvn.replace_node(res, projs->resproj[0]);
2879 _igvn.replace_node(ex_ctl, projs->catchall_catchproj);
2880 _igvn.replace_node(ex_mem, projs->catchall_memproj);
2881 _igvn.replace_node(ex_io, projs->catchall_ioproj);
2882 }
2883
2884 void PhaseMacroExpand::expand_subtypecheck_node(SubTypeCheckNode *check) {
2885 assert(check->in(SubTypeCheckNode::Control) == nullptr, "should be pinned");
2886 Node* bol = check->unique_out();
2887 Node* obj_or_subklass = check->in(SubTypeCheckNode::ObjOrSubKlass);
2888 Node* superklass = check->in(SubTypeCheckNode::SuperKlass);
2889 assert(bol->is_Bool() && bol->as_Bool()->_test._test == BoolTest::ne, "unexpected bool node");
2890
2891 for (DUIterator_Last imin, i = bol->last_outs(imin); i >= imin; --i) {
2892 Node* iff = bol->last_out(i);
2893 assert(iff->is_If(), "where's the if?");
2894
2895 if (iff->in(0)->is_top()) {
2896 _igvn.replace_input_of(iff, 1, C->top());
2897 continue;
2898 }
2899
2900 Node* iftrue = iff->as_If()->proj_out(1);
2901 Node* iffalse = iff->as_If()->proj_out(0);
2902 Node* ctrl = iff->in(0);
2903
2904 Node* subklass = nullptr;
2905 if (_igvn.type(obj_or_subklass)->isa_klassptr()) {
2906 subklass = obj_or_subklass;
2907 } else {
2908 Node* k_adr = basic_plus_adr(obj_or_subklass, oopDesc::klass_offset_in_bytes());
2909 subklass = _igvn.transform(LoadKlassNode::make(_igvn, C->immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
2910 }
2911
2912 Node* not_subtype_ctrl = Phase::gen_subtype_check(subklass, superklass, &ctrl, nullptr, _igvn, check->method(), check->bci());
2913
2914 _igvn.replace_input_of(iff, 0, C->top());
2915 _igvn.replace_node(iftrue, not_subtype_ctrl);
2916 _igvn.replace_node(iffalse, ctrl);
2917 }
2918 _igvn.replace_node(check, C->top());
2919 }
2920
2921 // FlatArrayCheckNode (array1 array2 ...) is expanded into:
2922 //
2923 // long mark = array1.mark | array2.mark | ...;
2924 // long locked_bit = markWord::unlocked_value & array1.mark & array2.mark & ...;
2925 // if (locked_bit == 0) {
2926 // // One array is locked, load prototype header from the klass
2927 // mark = array1.klass.proto | array2.klass.proto | ...
2928 // }
2929 // if ((mark & markWord::flat_array_bit_in_place) == 0) {
2930 // ...
2931 // }
2932 void PhaseMacroExpand::expand_flatarraycheck_node(FlatArrayCheckNode* check) {
2933 bool array_inputs = _igvn.type(check->in(FlatArrayCheckNode::ArrayOrKlass))->isa_oopptr() != nullptr;
2934 if (array_inputs) {
2935 Node* mark = MakeConX(0);
2936 Node* locked_bit = MakeConX(markWord::unlocked_value);
2937 Node* mem = check->in(FlatArrayCheckNode::Memory);
2938 for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
2939 Node* ary = check->in(i);
2940 const TypeOopPtr* t = _igvn.type(ary)->isa_oopptr();
2941 assert(t != nullptr, "Mixing array and klass inputs");
2942 assert(!t->is_flat() && !t->is_not_flat(), "Should have been optimized out");
2943 Node* mark_adr = basic_plus_adr(ary, oopDesc::mark_offset_in_bytes());
2944 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));
2945 mark = _igvn.transform(new OrXNode(mark, mark_load));
2946 locked_bit = _igvn.transform(new AndXNode(locked_bit, mark_load));
2947 }
2948 assert(!mark->is_Con(), "Should have been optimized out");
2949 Node* cmp = _igvn.transform(new CmpXNode(locked_bit, MakeConX(0)));
2950 Node* is_unlocked = _igvn.transform(new BoolNode(cmp, BoolTest::ne));
2951
2952 // BoolNode might be shared, replace each if user
2953 Node* old_bol = check->unique_out();
2954 assert(old_bol->is_Bool() && old_bol->as_Bool()->_test._test == BoolTest::ne, "unexpected condition");
2955 for (DUIterator_Last imin, i = old_bol->last_outs(imin); i >= imin; --i) {
2956 IfNode* old_iff = old_bol->last_out(i)->as_If();
2957 Node* ctrl = old_iff->in(0);
2958 RegionNode* region = new RegionNode(3);
2959 Node* mark_phi = new PhiNode(region, TypeX_X);
2960
2961 // Check if array is unlocked
2962 IfNode* iff = _igvn.transform(new IfNode(ctrl, is_unlocked, PROB_MAX, COUNT_UNKNOWN))->as_If();
2963
2964 // Unlocked: Use bits from mark word
2965 region->init_req(1, _igvn.transform(new IfTrueNode(iff)));
2966 mark_phi->init_req(1, mark);
2967
2968 // Locked: Load prototype header from klass
2969 ctrl = _igvn.transform(new IfFalseNode(iff));
2970 Node* proto = MakeConX(0);
2971 for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
2972 Node* ary = check->in(i);
2973 // Make loads control dependent to make sure they are only executed if array is locked
2974 Node* klass_adr = basic_plus_adr(ary, oopDesc::klass_offset_in_bytes());
2975 Node* klass = _igvn.transform(LoadKlassNode::make(_igvn, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
2976 Node* proto_adr = basic_plus_adr(klass, in_bytes(Klass::prototype_header_offset()));
2977 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));
2978 proto = _igvn.transform(new OrXNode(proto, proto_load));
2979 }
2980 region->init_req(2, ctrl);
2981 mark_phi->init_req(2, proto);
2982
2983 // Check if flat array bits are set
2984 Node* mask = MakeConX(markWord::flat_array_bit_in_place);
2985 Node* masked = _igvn.transform(new AndXNode(_igvn.transform(mark_phi), mask));
2986 cmp = _igvn.transform(new CmpXNode(masked, MakeConX(0)));
2987 Node* is_not_flat = _igvn.transform(new BoolNode(cmp, BoolTest::eq));
2988
2989 ctrl = _igvn.transform(region);
2990 iff = _igvn.transform(new IfNode(ctrl, is_not_flat, PROB_MAX, COUNT_UNKNOWN))->as_If();
2991 _igvn.replace_node(old_iff, iff);
2992 }
2993 _igvn.replace_node(check, C->top());
2994 } else {
2995 // Fall back to layout helper check
2996 Node* lhs = intcon(0);
2997 for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
2998 Node* array_or_klass = check->in(i);
2999 Node* klass = nullptr;
3000 const TypePtr* t = _igvn.type(array_or_klass)->is_ptr();
3001 assert(!t->is_flat() && !t->is_not_flat(), "Should have been optimized out");
3002 if (t->isa_oopptr() != nullptr) {
3003 Node* klass_adr = basic_plus_adr(array_or_klass, oopDesc::klass_offset_in_bytes());
3004 klass = transform_later(LoadKlassNode::make(_igvn, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
3005 } else {
3006 assert(t->isa_klassptr(), "Unexpected input type");
3007 klass = array_or_klass;
3008 }
3009 Node* lh_addr = basic_plus_adr(klass, in_bytes(Klass::layout_helper_offset()));
3010 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));
3011 lhs = _igvn.transform(new OrINode(lhs, lh_val));
3012 }
3013 Node* masked = transform_later(new AndINode(lhs, intcon(Klass::_lh_array_tag_flat_value_bit_inplace)));
3014 Node* cmp = transform_later(new CmpINode(masked, intcon(0)));
3015 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
3016 Node* m2b = transform_later(new Conv2BNode(masked));
3017 // The matcher expects the input to If/CMove nodes to be produced by a Bool(CmpI..)
3018 // pattern, but the input to other potential users (e.g. Phi) to be some
3019 // other pattern (e.g. a Conv2B node, possibly idealized as a CMoveI).
3020 Node* old_bol = check->unique_out();
3021 for (DUIterator_Last imin, i = old_bol->last_outs(imin); i >= imin; --i) {
3022 Node* user = old_bol->last_out(i);
3023 for (uint j = 0; j < user->req(); j++) {
3024 Node* n = user->in(j);
3025 if (n == old_bol) {
3026 _igvn.replace_input_of(user, j, (user->is_If() || user->is_CMove()) ? bol : m2b);
3027 }
3028 }
3029 }
3030 _igvn.replace_node(check, C->top());
3031 }
3032 }
3033
3034 // Perform refining of strip mined loop nodes in the macro nodes list.
3035 void PhaseMacroExpand::refine_strip_mined_loop_macro_nodes() {
3036 for (int i = C->macro_count(); i > 0; i--) {
3037 Node* n = C->macro_node(i - 1);
3038 if (n->is_OuterStripMinedLoop()) {
3039 n->as_OuterStripMinedLoop()->adjust_strip_mined_loop(&_igvn);
3040 }
3041 }
3042 }
3043
3044 //---------------------------eliminate_macro_nodes----------------------
3045 // Eliminate scalar replaced allocations and associated locks.
3046 void PhaseMacroExpand::eliminate_macro_nodes(bool eliminate_locks) {
3047 if (C->macro_count() == 0) {
3048 return;
3049 }
3050
3051 if (StressMacroElimination) {
3052 C->shuffle_macro_nodes();
3053 }
3054 NOT_PRODUCT(int membar_before = count_MemBar(C);)
3055
3056 int iteration = 0;
3057 while (C->macro_count() > 0) {
3058 if (iteration++ > 100) {
3059 assert(false, "Too slow convergence of macro elimination");
3060 break;
3061 }
3062
3063 // Postpone lock elimination to after EA when most allocations are eliminated
3064 // because they might block lock elimination if their escape state isn't
3065 // determined yet and we only got one chance at eliminating the lock.
3066 if (eliminate_locks) {
3067 // Before elimination may re-mark (change to Nested or NonEscObj)
3068 // all associated (same box and obj) lock and unlock nodes.
3069 int cnt = C->macro_count();
3070 for (int i=0; i < cnt; i++) {
3071 Node *n = C->macro_node(i);
3072 if (n->is_AbstractLock()) { // Lock and Unlock nodes
3073 mark_eliminated_locking_nodes(n->as_AbstractLock());
3074 }
3075 }
3076 // Re-marking may break consistency of Coarsened locks.
3077 if (!C->coarsened_locks_consistent()) {
3078 return; // recompile without Coarsened locks if broken
3079 } else {
3080 // After coarsened locks are eliminated locking regions
3081 // become unbalanced. We should not execute any more
3082 // locks elimination optimizations on them.
3083 C->mark_unbalanced_boxes();
3084 }
3085 }
3086
3087 bool progress = false;
3088 for (int i = C->macro_count(); i > 0; i = MIN2(i - 1, C->macro_count())) { // more than 1 element can be eliminated at once
3089 Node* n = C->macro_node(i - 1);
3090 bool success = false;
3091 DEBUG_ONLY(int old_macro_count = C->macro_count();)
3092 switch (n->class_id()) {
3093 case Node::Class_Allocate:
3094 case Node::Class_AllocateArray:
3095 success = eliminate_allocate_node(n->as_Allocate());
3096 #ifndef PRODUCT
3097 if (success && PrintOptoStatistics) {
3098 AtomicAccess::inc(&PhaseMacroExpand::_objs_scalar_replaced_counter);
3099 }
3100 #endif
3101 break;
3102 case Node::Class_CallStaticJava: {
3103 CallStaticJavaNode* call = n->as_CallStaticJava();
3104 if (!call->method()->is_method_handle_intrinsic()) {
3105 success = eliminate_boxing_node(n->as_CallStaticJava());
3106 }
3107 break;
3108 }
3109 case Node::Class_Lock:
3110 case Node::Class_Unlock:
3111 if (eliminate_locks) {
3112 success = eliminate_locking_node(n->as_AbstractLock());
3113 #ifndef PRODUCT
3114 if (success && PrintOptoStatistics) {
3115 AtomicAccess::inc(&PhaseMacroExpand::_monitor_objects_removed_counter);
3116 }
3117 #endif
3118 }
3119 break;
3120 case Node::Class_ArrayCopy:
3121 break;
3122 case Node::Class_OuterStripMinedLoop:
3123 break;
3124 case Node::Class_SubTypeCheck:
3125 break;
3126 case Node::Class_Opaque1:
3127 break;
3128 case Node::Class_FlatArrayCheck:
3129 break;
3130 default:
3131 assert(n->Opcode() == Op_LoopLimit ||
3132 n->Opcode() == Op_ModD ||
3133 n->Opcode() == Op_ModF ||
3134 n->is_OpaqueNotNull() ||
3135 n->is_OpaqueInitializedAssertionPredicate() ||
3136 n->Opcode() == Op_MaxL ||
3137 n->Opcode() == Op_MinL ||
3138 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n),
3139 "unknown node type in macro list");
3140 }
3141 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
3142 progress = progress || success;
3143 if (success) {
3144 C->print_method(PHASE_AFTER_MACRO_ELIMINATION_STEP, 5, n);
3145 }
3146 }
3147
3148 // Ensure the graph after PhaseMacroExpand::eliminate_macro_nodes is canonical (no igvn
3149 // transformation is pending). If an allocation is used only in safepoints, elimination of
3150 // other macro nodes can remove all these safepoints, allowing the allocation to be removed.
3151 // Hence after igvn we retry removing macro nodes if some progress that has been made in this
3152 // iteration.
3153 _igvn.set_delay_transform(false);
3154 _igvn.optimize();
3155 if (C->failing()) {
3156 return;
3157 }
3158 _igvn.set_delay_transform(true);
3159
3160 if (!progress) {
3161 break;
3162 }
3163 }
3164 #ifndef PRODUCT
3165 if (PrintOptoStatistics) {
3166 int membar_after = count_MemBar(C);
3167 AtomicAccess::add(&PhaseMacroExpand::_memory_barriers_removed_counter, membar_before - membar_after);
3168 }
3169 #endif
3170 }
3171
3172 void PhaseMacroExpand::eliminate_opaque_looplimit_macro_nodes() {
3173 if (C->macro_count() == 0) {
3174 return;
3175 }
3176 refine_strip_mined_loop_macro_nodes();
3177 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
3178 bool progress = true;
3179 while (progress) {
3180 progress = false;
3181 for (int i = C->macro_count(); i > 0; i--) {
3182 Node* n = C->macro_node(i-1);
3183 bool success = false;
3184 DEBUG_ONLY(int old_macro_count = C->macro_count();)
3185 if (n->Opcode() == Op_LoopLimit) {
3186 // Remove it from macro list and put on IGVN worklist to optimize.
3187 C->remove_macro_node(n);
3188 _igvn._worklist.push(n);
3189 success = true;
3190 } else if (n->Opcode() == Op_CallStaticJava) {
3191 CallStaticJavaNode* call = n->as_CallStaticJava();
3192 if (!call->method()->is_method_handle_intrinsic()) {
3193 // Remove it from macro list and put on IGVN worklist to optimize.
3194 C->remove_macro_node(n);
3195 _igvn._worklist.push(n);
3196 success = true;
3197 }
3198 } else if (n->is_Opaque1()) {
3199 _igvn.replace_node(n, n->in(1));
3200 success = true;
3201 } else if (n->is_OpaqueNotNull()) {
3202 // Tests with OpaqueNotNull nodes are implicitly known to be true. Replace the node with true. In debug builds,
3203 // we leave the test in the graph to have an additional sanity check at runtime. If the test fails (i.e. a bug),
3204 // we will execute a Halt node.
3205 #ifdef ASSERT
3206 _igvn.replace_node(n, n->in(1));
3207 #else
3208 _igvn.replace_node(n, _igvn.intcon(1));
3209 #endif
3210 success = true;
3211 } else if (n->is_OpaqueInitializedAssertionPredicate()) {
3212 // Initialized Assertion Predicates must always evaluate to true. Therefore, we get rid of them in product
3213 // builds as they are useless. In debug builds we keep them as additional verification code. Even though
3214 // loop opts are already over, we want to keep Initialized Assertion Predicates alive as long as possible to
3215 // enable folding of dead control paths within which cast nodes become top after due to impossible types -
3216 // even after loop opts are over. Therefore, we delay the removal of these opaque nodes until now.
3217 #ifdef ASSERT
3286 // Worst case is a macro node gets expanded into about 200 nodes.
3287 // Allow 50% more for optimization.
3288 if (C->check_node_count(300, "out of nodes before macro expansion")) {
3289 return true;
3290 }
3291
3292 DEBUG_ONLY(int old_macro_count = C->macro_count();)
3293 switch (n->class_id()) {
3294 case Node::Class_Lock:
3295 expand_lock_node(n->as_Lock());
3296 break;
3297 case Node::Class_Unlock:
3298 expand_unlock_node(n->as_Unlock());
3299 break;
3300 case Node::Class_ArrayCopy:
3301 expand_arraycopy_node(n->as_ArrayCopy());
3302 break;
3303 case Node::Class_SubTypeCheck:
3304 expand_subtypecheck_node(n->as_SubTypeCheck());
3305 break;
3306 case Node::Class_CallStaticJava:
3307 expand_mh_intrinsic_return(n->as_CallStaticJava());
3308 C->remove_macro_node(n);
3309 break;
3310 case Node::Class_FlatArrayCheck:
3311 expand_flatarraycheck_node(n->as_FlatArrayCheck());
3312 break;
3313 default:
3314 switch (n->Opcode()) {
3315 case Op_ModD:
3316 case Op_ModF: {
3317 CallNode* mod_macro = n->as_Call();
3318 CallNode* call = new CallLeafPureNode(mod_macro->tf(), mod_macro->entry_point(),
3319 mod_macro->_name, TypeRawPtr::BOTTOM);
3320 call->init_req(TypeFunc::Control, mod_macro->in(TypeFunc::Control));
3321 call->init_req(TypeFunc::I_O, C->top());
3322 call->init_req(TypeFunc::Memory, C->top());
3323 call->init_req(TypeFunc::ReturnAdr, C->top());
3324 call->init_req(TypeFunc::FramePtr, C->top());
3325 for (unsigned int i = 0; i < mod_macro->tf()->domain_cc()->cnt() - TypeFunc::Parms; i++) {
3326 call->init_req(TypeFunc::Parms + i, mod_macro->in(TypeFunc::Parms + i));
3327 }
3328 _igvn.replace_node(mod_macro, call);
3329 transform_later(call);
3330 break;
3331 }
3332 default:
3333 assert(false, "unknown node type in macro list");
3334 }
3335 }
3336 assert(C->macro_count() == (old_macro_count - 1), "expansion must have deleted one node from macro list");
3337 if (C->failing()) return true;
3338 C->print_method(PHASE_AFTER_MACRO_EXPANSION_STEP, 5, n);
3339
3340 // Clean up the graph so we're less likely to hit the maximum node
3341 // limit
3342 _igvn.set_delay_transform(false);
3343 _igvn.optimize();
3344 if (C->failing()) return true;
3345 _igvn.set_delay_transform(true);
|