1 /*
2 * Copyright (c) 2005, 2025, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "ci/ciFlatArrayKlass.hpp"
26 #include "compiler/compileLog.hpp"
27 #include "gc/shared/collectedHeap.inline.hpp"
28 #include "gc/shared/tlab_globals.hpp"
29 #include "libadt/vectset.hpp"
30 #include "memory/universe.hpp"
31 #include "opto/addnode.hpp"
32 #include "opto/arraycopynode.hpp"
33 #include "opto/callnode.hpp"
34 #include "opto/castnode.hpp"
35 #include "opto/cfgnode.hpp"
36 #include "opto/compile.hpp"
37 #include "opto/convertnode.hpp"
38 #include "opto/graphKit.hpp"
39 #include "opto/inlinetypenode.hpp"
40 #include "opto/intrinsicnode.hpp"
41 #include "opto/locknode.hpp"
42 #include "opto/loopnode.hpp"
43 #include "opto/macro.hpp"
44 #include "opto/memnode.hpp"
45 #include "opto/narrowptrnode.hpp"
46 #include "opto/node.hpp"
47 #include "opto/opaquenode.hpp"
48 #include "opto/phaseX.hpp"
49 #include "opto/rootnode.hpp"
50 #include "opto/runtime.hpp"
51 #include "opto/subnode.hpp"
52 #include "opto/subtypenode.hpp"
53 #include "opto/type.hpp"
54 #include "prims/jvmtiExport.hpp"
55 #include "runtime/continuation.hpp"
56 #include "runtime/sharedRuntime.hpp"
57 #include "runtime/stubRoutines.hpp"
58 #include "utilities/macros.hpp"
59 #include "utilities/powerOfTwo.hpp"
60 #if INCLUDE_G1GC
61 #include "gc/g1/g1ThreadLocalData.hpp"
62 #endif // INCLUDE_G1GC
63
64
65 //
66 // Replace any references to "oldref" in inputs to "use" with "newref".
67 // Returns the number of replacements made.
68 //
69 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
70 int nreplacements = 0;
71 uint req = use->req();
72 for (uint j = 0; j < use->len(); j++) {
73 Node *uin = use->in(j);
74 if (uin == oldref) {
75 if (j < req)
76 use->set_req(j, newref);
77 else
78 use->set_prec(j, newref);
79 nreplacements++;
80 } else if (j >= req && uin == nullptr) {
81 break;
82 }
83 }
84 return nreplacements;
85 }
86
87 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
88 Node* cmp;
89 if (mask != 0) {
90 Node* and_node = transform_later(new AndXNode(word, MakeConX(mask)));
91 cmp = transform_later(new CmpXNode(and_node, MakeConX(bits)));
92 } else {
93 cmp = word;
94 }
95 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne));
96 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
97 transform_later(iff);
98
99 // Fast path taken.
100 Node *fast_taken = transform_later(new IfFalseNode(iff));
101
102 // Fast path not-taken, i.e. slow path
103 Node *slow_taken = transform_later(new IfTrueNode(iff));
104
105 if (return_fast_path) {
106 region->init_req(edge, slow_taken); // Capture slow-control
107 return fast_taken;
108 } else {
109 region->init_req(edge, fast_taken); // Capture fast-control
110 return slow_taken;
111 }
112 }
113
114 //--------------------copy_predefined_input_for_runtime_call--------------------
115 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
116 // Set fixed predefined input arguments
117 call->init_req( TypeFunc::Control, ctrl );
118 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) );
119 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
120 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
121 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
122 }
123
124 //------------------------------make_slow_call---------------------------------
125 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type,
126 address slow_call, const char* leaf_name, Node* slow_path,
127 Node* parm0, Node* parm1, Node* parm2) {
128
129 // Slow-path call
130 CallNode *call = leaf_name
131 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
132 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), TypeRawPtr::BOTTOM );
133
134 // Slow path call has no side-effects, uses few values
135 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
136 if (parm0 != nullptr) call->init_req(TypeFunc::Parms+0, parm0);
137 if (parm1 != nullptr) call->init_req(TypeFunc::Parms+1, parm1);
138 if (parm2 != nullptr) call->init_req(TypeFunc::Parms+2, parm2);
139 call->copy_call_debug_info(&_igvn, oldcall);
140 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
141 _igvn.replace_node(oldcall, call);
142 transform_later(call);
143
144 return call;
145 }
146
147 void PhaseMacroExpand::eliminate_gc_barrier(Node* p2x) {
148 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
149 bs->eliminate_gc_barrier(&_igvn, p2x);
150 #ifndef PRODUCT
151 if (PrintOptoStatistics) {
152 Atomic::inc(&PhaseMacroExpand::_GC_barriers_removed_counter);
153 }
154 #endif
155 }
156
157 // Search for a memory operation for the specified memory slice.
158 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
159 Node *orig_mem = mem;
160 Node *alloc_mem = alloc->as_Allocate()->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
161 assert(alloc_mem != nullptr, "Allocation without a memory projection.");
162 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
163 while (true) {
164 if (mem == alloc_mem || mem == start_mem ) {
165 return mem; // hit one of our sentinels
166 } else if (mem->is_MergeMem()) {
167 mem = mem->as_MergeMem()->memory_at(alias_idx);
168 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
169 Node *in = mem->in(0);
170 // we can safely skip over safepoints, calls, locks and membars because we
171 // already know that the object is safe to eliminate.
172 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
173 return in;
174 } else if (in->is_Call()) {
175 CallNode *call = in->as_Call();
176 if (call->may_modify(tinst, phase)) {
177 assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape");
178 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) {
179 return in;
180 }
181 }
182 mem = in->in(TypeFunc::Memory);
183 } else if (in->is_MemBar()) {
184 ArrayCopyNode* ac = nullptr;
185 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) {
186 if (ac != nullptr) {
187 assert(ac->is_clonebasic(), "Only basic clone is a non escaping clone");
188 return ac;
189 }
190 }
191 mem = in->in(TypeFunc::Memory);
192 } else {
193 #ifdef ASSERT
194 in->dump();
195 mem->dump();
196 assert(false, "unexpected projection");
197 #endif
198 }
199 } else if (mem->is_Store()) {
200 const TypePtr* atype = mem->as_Store()->adr_type();
201 int adr_idx = phase->C->get_alias_index(atype);
202 if (adr_idx == alias_idx) {
203 assert(atype->isa_oopptr(), "address type must be oopptr");
204 int adr_offset = atype->flat_offset();
205 uint adr_iid = atype->is_oopptr()->instance_id();
206 // Array elements references have the same alias_idx
207 // but different offset and different instance_id.
208 if (adr_offset == offset && adr_iid == alloc->_idx) {
209 return mem;
210 }
211 } else {
212 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
213 }
214 mem = mem->in(MemNode::Memory);
215 } else if (mem->is_ClearArray()) {
216 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
217 // Can not bypass initialization of the instance
218 // we are looking.
219 debug_only(intptr_t offset;)
220 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
221 InitializeNode* init = alloc->as_Allocate()->initialization();
222 // We are looking for stored value, return Initialize node
223 // or memory edge from Allocate node.
224 if (init != nullptr) {
225 return init;
226 } else {
227 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers).
228 }
229 }
230 // Otherwise skip it (the call updated 'mem' value).
231 } else if (mem->Opcode() == Op_SCMemProj) {
232 mem = mem->in(0);
233 Node* adr = nullptr;
234 if (mem->is_LoadStore()) {
235 adr = mem->in(MemNode::Address);
236 } else {
237 assert(mem->Opcode() == Op_EncodeISOArray ||
238 mem->Opcode() == Op_StrCompressedCopy, "sanity");
239 adr = mem->in(3); // Destination array
240 }
241 const TypePtr* atype = adr->bottom_type()->is_ptr();
242 int adr_idx = phase->C->get_alias_index(atype);
243 if (adr_idx == alias_idx) {
244 DEBUG_ONLY(mem->dump();)
245 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
246 return nullptr;
247 }
248 mem = mem->in(MemNode::Memory);
249 } else if (mem->Opcode() == Op_StrInflatedCopy) {
250 Node* adr = mem->in(3); // Destination array
251 const TypePtr* atype = adr->bottom_type()->is_ptr();
252 int adr_idx = phase->C->get_alias_index(atype);
253 if (adr_idx == alias_idx) {
254 DEBUG_ONLY(mem->dump();)
255 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
256 return nullptr;
257 }
258 mem = mem->in(MemNode::Memory);
259 } else {
260 return mem;
261 }
262 assert(mem != orig_mem, "dead memory loop");
263 }
264 }
265
266 // Generate loads from source of the arraycopy for fields of
267 // destination needed at a deoptimization point
268 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) {
269 BasicType bt = ft;
270 const Type *type = ftype;
271 if (ft == T_NARROWOOP) {
272 bt = T_OBJECT;
273 type = ftype->make_oopptr();
274 }
275 Node* res = nullptr;
276 if (ac->is_clonebasic()) {
277 assert(ac->in(ArrayCopyNode::Src) != ac->in(ArrayCopyNode::Dest), "clone source equals destination");
278 Node* base = ac->in(ArrayCopyNode::Src);
279 Node* adr = _igvn.transform(new AddPNode(base, base, _igvn.MakeConX(offset)));
280 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
281 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
282 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
283 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt);
284 } else {
285 if (ac->modifies(offset, offset, &_igvn, true)) {
286 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
287 uint shift = exact_log2(type2aelembytes(bt));
288 Node* src_pos = ac->in(ArrayCopyNode::SrcPos);
289 Node* dest_pos = ac->in(ArrayCopyNode::DestPos);
290 const TypeInt* src_pos_t = _igvn.type(src_pos)->is_int();
291 const TypeInt* dest_pos_t = _igvn.type(dest_pos)->is_int();
292
293 Node* adr = nullptr;
294 Node* base = ac->in(ArrayCopyNode::Src);
295 const TypeAryPtr* adr_type = _igvn.type(base)->is_aryptr();
296 if (adr_type->is_flat()) {
297 shift = adr_type->flat_log_elem_size();
298 }
299 if (src_pos_t->is_con() && dest_pos_t->is_con()) {
300 intptr_t off = ((src_pos_t->get_con() - dest_pos_t->get_con()) << shift) + offset;
301 adr = _igvn.transform(new AddPNode(base, base, _igvn.MakeConX(off)));
302 adr_type = _igvn.type(adr)->is_aryptr();
303 assert(adr_type == _igvn.type(base)->is_aryptr()->add_field_offset_and_offset(off), "incorrect address type");
304 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
305 // Don't emit a new load from src if src == dst but try to get the value from memory instead
306 return value_from_mem(ac->in(TypeFunc::Memory), ctl, ft, ftype, adr_type, alloc);
307 }
308 } else {
309 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
310 // Non constant offset in the array: we can't statically
311 // determine the value
312 return nullptr;
313 }
314 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
315 #ifdef _LP64
316 diff = _igvn.transform(new ConvI2LNode(diff));
317 #endif
318 diff = _igvn.transform(new LShiftXNode(diff, _igvn.intcon(shift)));
319
320 Node* off = _igvn.transform(new AddXNode(_igvn.MakeConX(offset), diff));
321 adr = _igvn.transform(new AddPNode(base, base, off));
322 // In the case of a flat inline type array, each field has its
323 // own slice so we need to extract the field being accessed from
324 // the address computation
325 adr_type = adr_type->add_field_offset_and_offset(offset)->add_offset(Type::OffsetBot)->is_aryptr();
326 adr = _igvn.transform(new CastPPNode(ctl, adr, adr_type));
327 }
328 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
329 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
330 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt);
331 }
332 }
333 if (res != nullptr) {
334 if (ftype->isa_narrowoop()) {
335 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
336 assert(res->isa_DecodeN(), "should be narrow oop");
337 res = _igvn.transform(new EncodePNode(res, ftype));
338 }
339 return res;
340 }
341 return nullptr;
342 }
343
344 //
345 // Given a Memory Phi, compute a value Phi containing the values from stores
346 // on the input paths.
347 // Note: this function is recursive, its depth is limited by the "level" argument
348 // Returns the computed Phi, or null if it cannot compute it.
349 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) {
350 assert(mem->is_Phi(), "sanity");
351 int alias_idx = C->get_alias_index(adr_t);
352 int offset = adr_t->flat_offset();
353 int instance_id = adr_t->instance_id();
354
355 // Check if an appropriate value phi already exists.
356 Node* region = mem->in(0);
357 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
358 Node* phi = region->fast_out(k);
359 if (phi->is_Phi() && phi != mem &&
360 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) {
361 return phi;
362 }
363 }
364 // Check if an appropriate new value phi already exists.
365 Node* new_phi = value_phis->find(mem->_idx);
366 if (new_phi != nullptr)
367 return new_phi;
368
369 if (level <= 0) {
370 return nullptr; // Give up: phi tree too deep
371 }
372 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
373 Node *alloc_mem = alloc->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
374 assert(alloc_mem != nullptr, "Allocation without a memory projection.");
375
376 uint length = mem->req();
377 GrowableArray <Node *> values(length, length, nullptr);
378
379 // create a new Phi for the value
380 PhiNode *phi = new PhiNode(mem->in(0), phi_type, nullptr, mem->_idx, instance_id, alias_idx, offset);
381 transform_later(phi);
382 value_phis->push(phi, mem->_idx);
383
384 for (uint j = 1; j < length; j++) {
385 Node *in = mem->in(j);
386 if (in == nullptr || in->is_top()) {
387 values.at_put(j, in);
388 } else {
389 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
390 if (val == start_mem || val == alloc_mem) {
391 // hit a sentinel, return appropriate value
392 Node* init_value = alloc->in(AllocateNode::InitValue);
393 if (init_value != nullptr) {
394 if (val == start_mem) {
395 // TODO 8350865 Somehow we ended up with root mem and therefore walked past the alloc. Fix this. Triggered by TestGenerated::test15
396 // Don't we need field_value_by_offset?
397 return nullptr;
398 }
399 values.at_put(j, init_value);
400 } else {
401 assert(alloc->in(AllocateNode::RawInitValue) == nullptr, "init value may not be null");
402 values.at_put(j, _igvn.zerocon(ft));
403 }
404 continue;
405 }
406 if (val->is_Initialize()) {
407 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
408 }
409 if (val == nullptr) {
410 return nullptr; // can't find a value on this path
411 }
412 if (val == mem) {
413 values.at_put(j, mem);
414 } else if (val->is_Store()) {
415 Node* n = val->in(MemNode::ValueIn);
416 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
417 n = bs->step_over_gc_barrier(n);
418 if (is_subword_type(ft)) {
419 n = Compile::narrow_value(ft, n, phi_type, &_igvn, true);
420 }
421 values.at_put(j, n);
422 } else if (val->is_Proj() && val->in(0) == alloc) {
423 Node* init_value = alloc->in(AllocateNode::InitValue);
424 if (init_value != nullptr) {
425 // TODO 8350865 Is this correct for non-all-zero init values? Don't we need field_value_by_offset?
426 values.at_put(j, init_value);
427 } else {
428 assert(alloc->in(AllocateNode::RawInitValue) == nullptr, "init value may not be null");
429 values.at_put(j, _igvn.zerocon(ft));
430 }
431 } else if (val->is_Phi()) {
432 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
433 if (val == nullptr) {
434 return nullptr;
435 }
436 values.at_put(j, val);
437 } else if (val->Opcode() == Op_SCMemProj) {
438 assert(val->in(0)->is_LoadStore() ||
439 val->in(0)->Opcode() == Op_EncodeISOArray ||
440 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity");
441 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
442 return nullptr;
443 } else if (val->is_ArrayCopy()) {
444 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc);
445 if (res == nullptr) {
446 return nullptr;
447 }
448 values.at_put(j, res);
449 } else if (val->is_top()) {
450 // This indicates that this path into the phi is dead. Top will eventually also propagate into the Region.
451 // IGVN will clean this up later.
452 values.at_put(j, val);
453 } else {
454 DEBUG_ONLY( val->dump(); )
455 assert(false, "unknown node on this path");
456 return nullptr; // unknown node on this path
457 }
458 }
459 }
460 // Set Phi's inputs
461 for (uint j = 1; j < length; j++) {
462 if (values.at(j) == mem) {
463 phi->init_req(j, phi);
464 } else {
465 phi->init_req(j, values.at(j));
466 }
467 }
468 return phi;
469 }
470
471 // Search the last value stored into the object's field.
472 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) {
473 assert(adr_t->is_known_instance_field(), "instance required");
474 int instance_id = adr_t->instance_id();
475 assert((uint)instance_id == alloc->_idx, "wrong allocation");
476
477 int alias_idx = C->get_alias_index(adr_t);
478 int offset = adr_t->flat_offset();
479 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
480 Node *alloc_mem = alloc->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
481 assert(alloc_mem != nullptr, "Allocation without a memory projection.");
482 VectorSet visited;
483
484 bool done = sfpt_mem == alloc_mem;
485 Node *mem = sfpt_mem;
486 while (!done) {
487 if (visited.test_set(mem->_idx)) {
488 return nullptr; // found a loop, give up
489 }
490 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
491 if (mem == start_mem || mem == alloc_mem) {
492 done = true; // hit a sentinel, return appropriate 0 value
493 } else if (mem->is_Initialize()) {
494 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
495 if (mem == nullptr) {
496 done = true; // Something went wrong.
497 } else if (mem->is_Store()) {
498 const TypePtr* atype = mem->as_Store()->adr_type();
499 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
500 done = true;
501 }
502 } else if (mem->is_Store()) {
503 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
504 assert(atype != nullptr, "address type must be oopptr");
505 assert(C->get_alias_index(atype) == alias_idx &&
506 atype->is_known_instance_field() && atype->flat_offset() == offset &&
507 atype->instance_id() == instance_id, "store is correct memory slice");
508 done = true;
509 } else if (mem->is_Phi()) {
510 // try to find a phi's unique input
511 Node *unique_input = nullptr;
512 Node *top = C->top();
513 for (uint i = 1; i < mem->req(); i++) {
514 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
515 if (n == nullptr || n == top || n == mem) {
516 continue;
517 } else if (unique_input == nullptr) {
518 unique_input = n;
519 } else if (unique_input != n) {
520 unique_input = top;
521 break;
522 }
523 }
524 if (unique_input != nullptr && unique_input != top) {
525 mem = unique_input;
526 } else {
527 done = true;
528 }
529 } else if (mem->is_ArrayCopy()) {
530 done = true;
531 } else {
532 DEBUG_ONLY( mem->dump(); )
533 assert(false, "unexpected node");
534 }
535 }
536 if (mem != nullptr) {
537 if (mem == start_mem || mem == alloc_mem) {
538 // hit a sentinel, return appropriate value
539 Node* init_value = alloc->in(AllocateNode::InitValue);
540 if (init_value != nullptr) {
541 if (adr_t->is_flat()) {
542 if (init_value->is_EncodeP()) {
543 init_value = init_value->in(1);
544 }
545 assert(adr_t->is_aryptr()->field_offset().get() != Type::OffsetBot, "Unknown offset");
546 offset = adr_t->is_aryptr()->field_offset().get() + init_value->bottom_type()->inline_klass()->payload_offset();
547 init_value = init_value->as_InlineType()->field_value_by_offset(offset, true);
548 if (ft == T_NARROWOOP) {
549 init_value = transform_later(new EncodePNode(init_value, init_value->bottom_type()->make_ptr()));
550 }
551 }
552 return init_value;
553 }
554 assert(alloc->in(AllocateNode::RawInitValue) == nullptr, "init value may not be null");
555 return _igvn.zerocon(ft);
556 } else if (mem->is_Store()) {
557 Node* n = mem->in(MemNode::ValueIn);
558 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
559 n = bs->step_over_gc_barrier(n);
560 return n;
561 } else if (mem->is_Phi()) {
562 // attempt to produce a Phi reflecting the values on the input paths of the Phi
563 Node_Stack value_phis(8);
564 Node* phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
565 if (phi != nullptr) {
566 return phi;
567 } else {
568 // Kill all new Phis
569 while(value_phis.is_nonempty()) {
570 Node* n = value_phis.node();
571 _igvn.replace_node(n, C->top());
572 value_phis.pop();
573 }
574 }
575 } else if (mem->is_ArrayCopy()) {
576 Node* ctl = mem->in(0);
577 Node* m = mem->in(TypeFunc::Memory);
578 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj()) {
579 // pin the loads in the uncommon trap path
580 ctl = sfpt_ctl;
581 m = sfpt_mem;
582 }
583 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc);
584 }
585 }
586 // Something went wrong.
587 return nullptr;
588 }
589
590 // Search the last value stored into the inline type's fields (for flat arrays).
591 Node* PhaseMacroExpand::inline_type_from_mem(Node* mem, Node* ctl, ciInlineKlass* vk, const TypeAryPtr* adr_type, int offset, AllocateNode* alloc) {
592 // Subtract the offset of the first field to account for the missing oop header
593 offset -= vk->payload_offset();
594 // Create a new InlineTypeNode and retrieve the field values from memory
595 InlineTypeNode* vt = InlineTypeNode::make_uninitialized(_igvn, vk);
596 transform_later(vt);
597 for (int i = 0; i < vk->nof_declared_nonstatic_fields(); ++i) {
598 ciType* field_type = vt->field_type(i);
599 int field_offset = offset + vt->field_offset(i);
600 Node* value = nullptr;
601 if (vt->field_is_flat(i)) {
602 // TODO 8350865 Fix this
603 // assert(vt->field_is_null_free(i), "Unexpected nullable flat field");
604 if (!vt->field_is_null_free(i)) {
605 return nullptr;
606 }
607 value = inline_type_from_mem(mem, ctl, field_type->as_inline_klass(), adr_type, field_offset, alloc);
608 } else {
609 const Type* ft = Type::get_const_type(field_type);
610 BasicType bt = type2field[field_type->basic_type()];
611 if (UseCompressedOops && !is_java_primitive(bt)) {
612 ft = ft->make_narrowoop();
613 bt = T_NARROWOOP;
614 }
615 // Each inline type field has its own memory slice
616 adr_type = adr_type->with_field_offset(field_offset);
617 value = value_from_mem(mem, ctl, bt, ft, adr_type, alloc);
618 if (value != nullptr && ft->isa_narrowoop()) {
619 assert(UseCompressedOops, "unexpected narrow oop");
620 if (value->is_EncodeP()) {
621 value = value->in(1);
622 } else if (!value->is_InlineType()) {
623 value = transform_later(new DecodeNNode(value, value->get_ptr_type()));
624 }
625 }
626 }
627 if (value != nullptr) {
628 vt->set_field_value(i, value);
629 } else {
630 // We might have reached the TrackedInitializationLimit
631 return nullptr;
632 }
633 }
634 return vt;
635 }
636
637 // Check the possibility of scalar replacement.
638 bool PhaseMacroExpand::can_eliminate_allocation(PhaseIterGVN* igvn, AllocateNode *alloc, GrowableArray <SafePointNode *>* safepoints) {
639 // Scan the uses of the allocation to check for anything that would
640 // prevent us from eliminating it.
641 NOT_PRODUCT( const char* fail_eliminate = nullptr; )
642 DEBUG_ONLY( Node* disq_node = nullptr; )
643 bool can_eliminate = true;
644 bool reduce_merge_precheck = (safepoints == nullptr);
645
646 Unique_Node_List worklist;
647 Node* res = alloc->result_cast();
648 const TypeOopPtr* res_type = nullptr;
649 if (res == nullptr) {
650 // All users were eliminated.
651 } else if (!res->is_CheckCastPP()) {
652 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
653 can_eliminate = false;
654 } else {
655 worklist.push(res);
656 res_type = igvn->type(res)->isa_oopptr();
657 if (res_type == nullptr) {
658 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
659 can_eliminate = false;
660 } else if (!res_type->klass_is_exact()) {
661 NOT_PRODUCT(fail_eliminate = "Not an exact type.";)
662 can_eliminate = false;
663 } else if (res_type->isa_aryptr()) {
664 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
665 if (length < 0) {
666 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
667 can_eliminate = false;
668 }
669 }
670 }
671
672 while (can_eliminate && worklist.size() > 0) {
673 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
674 res = worklist.pop();
675 for (DUIterator_Fast jmax, j = res->fast_outs(jmax); j < jmax && can_eliminate; j++) {
676 Node* use = res->fast_out(j);
677
678 if (use->is_AddP()) {
679 const TypePtr* addp_type = igvn->type(use)->is_ptr();
680 int offset = addp_type->offset();
681
682 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
683 NOT_PRODUCT(fail_eliminate = "Undefined field reference";)
684 can_eliminate = false;
685 break;
686 }
687 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
688 k < kmax && can_eliminate; k++) {
689 Node* n = use->fast_out(k);
690 if (!n->is_Store() && n->Opcode() != Op_CastP2X && !bs->is_gc_pre_barrier_node(n) && !reduce_merge_precheck) {
691 DEBUG_ONLY(disq_node = n;)
692 if (n->is_Load() || n->is_LoadStore()) {
693 NOT_PRODUCT(fail_eliminate = "Field load";)
694 } else {
695 NOT_PRODUCT(fail_eliminate = "Not store field reference";)
696 }
697 can_eliminate = false;
698 }
699 }
700 } else if (use->is_ArrayCopy() &&
701 (use->as_ArrayCopy()->is_clonebasic() ||
702 use->as_ArrayCopy()->is_arraycopy_validated() ||
703 use->as_ArrayCopy()->is_copyof_validated() ||
704 use->as_ArrayCopy()->is_copyofrange_validated()) &&
705 use->in(ArrayCopyNode::Dest) == res) {
706 // ok to eliminate
707 } else if (use->is_SafePoint()) {
708 SafePointNode* sfpt = use->as_SafePoint();
709 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
710 // Object is passed as argument.
711 DEBUG_ONLY(disq_node = use;)
712 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
713 can_eliminate = false;
714 }
715 Node* sfptMem = sfpt->memory();
716 if (sfptMem == nullptr || sfptMem->is_top()) {
717 DEBUG_ONLY(disq_node = use;)
718 NOT_PRODUCT(fail_eliminate = "null or TOP memory";)
719 can_eliminate = false;
720 } else if (!reduce_merge_precheck) {
721 assert(!res->is_Phi() || !res->as_Phi()->can_be_inline_type(), "Inline type allocations should not have safepoint uses");
722 safepoints->append_if_missing(sfpt);
723 }
724 } else if (use->is_InlineType() && use->as_InlineType()->get_oop() == res) {
725 // Look at uses
726 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); k < kmax; k++) {
727 Node* u = use->fast_out(k);
728 if (u->is_InlineType()) {
729 // Use in flat field can be eliminated
730 InlineTypeNode* vt = u->as_InlineType();
731 for (uint i = 0; i < vt->field_count(); ++i) {
732 if (vt->field_value(i) == use && !vt->field_is_flat(i)) {
733 can_eliminate = false; // Use in non-flat field
734 break;
735 }
736 }
737 } else {
738 // Add other uses to the worklist to process individually
739 worklist.push(use);
740 }
741 }
742 } else if (use->Opcode() == Op_StoreX && use->in(MemNode::Address) == res) {
743 // Store to mark word of inline type larval buffer
744 assert(res_type->is_inlinetypeptr(), "Unexpected store to mark word");
745 } else if (res_type->is_inlinetypeptr() && (use->Opcode() == Op_MemBarRelease || use->Opcode() == Op_MemBarStoreStore)) {
746 // Inline type buffer allocations are followed by a membar
747 } else if (reduce_merge_precheck &&
748 (use->is_Phi() || use->is_EncodeP() ||
749 use->Opcode() == Op_MemBarRelease ||
750 (UseStoreStoreForCtor && use->Opcode() == Op_MemBarStoreStore))) {
751 // Nothing to do
752 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
753 if (use->is_Phi()) {
754 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
755 NOT_PRODUCT(fail_eliminate = "Object is return value";)
756 } else {
757 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
758 }
759 DEBUG_ONLY(disq_node = use;)
760 } else {
761 if (use->Opcode() == Op_Return) {
762 NOT_PRODUCT(fail_eliminate = "Object is return value";)
763 } else {
764 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
765 }
766 DEBUG_ONLY(disq_node = use;)
767 }
768 can_eliminate = false;
769 } else {
770 assert(use->Opcode() == Op_CastP2X, "should be");
771 assert(!use->has_out_with(Op_OrL), "should have been removed because oop is never null");
772 }
773 }
774 }
775
776 #ifndef PRODUCT
777 if (PrintEliminateAllocations && safepoints != nullptr) {
778 if (can_eliminate) {
779 tty->print("Scalar ");
780 if (res == nullptr)
781 alloc->dump();
782 else
783 res->dump();
784 } else {
785 tty->print("NotScalar (%s)", fail_eliminate);
786 if (res == nullptr)
787 alloc->dump();
788 else
789 res->dump();
790 #ifdef ASSERT
791 if (disq_node != nullptr) {
792 tty->print(" >>>> ");
793 disq_node->dump();
794 }
795 #endif /*ASSERT*/
796 }
797 }
798
799 if (TraceReduceAllocationMerges && !can_eliminate && reduce_merge_precheck) {
800 tty->print_cr("\tCan't eliminate allocation because '%s': ", fail_eliminate != nullptr ? fail_eliminate : "");
801 DEBUG_ONLY(if (disq_node != nullptr) disq_node->dump();)
802 }
803 #endif
804 return can_eliminate;
805 }
806
807 void PhaseMacroExpand::undo_previous_scalarizations(GrowableArray <SafePointNode *> safepoints_done, AllocateNode* alloc) {
808 Node* res = alloc->result_cast();
809 int nfields = 0;
810 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
811
812 if (res != nullptr) {
813 const TypeOopPtr* res_type = _igvn.type(res)->isa_oopptr();
814
815 if (res_type->isa_instptr()) {
816 // find the fields of the class which will be needed for safepoint debug information
817 ciInstanceKlass* iklass = res_type->is_instptr()->instance_klass();
818 nfields = iklass->nof_nonstatic_fields();
819 } else {
820 // find the array's elements which will be needed for safepoint debug information
821 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
822 assert(nfields >= 0, "must be an array klass.");
823 }
824 }
825
826 // rollback processed safepoints
827 while (safepoints_done.length() > 0) {
828 SafePointNode* sfpt_done = safepoints_done.pop();
829 // remove any extra entries we added to the safepoint
830 uint last = sfpt_done->req() - 1;
831 for (int k = 0; k < nfields; k++) {
832 sfpt_done->del_req(last--);
833 }
834 JVMState *jvms = sfpt_done->jvms();
835 jvms->set_endoff(sfpt_done->req());
836 // Now make a pass over the debug information replacing any references
837 // to SafePointScalarObjectNode with the allocated object.
838 int start = jvms->debug_start();
839 int end = jvms->debug_end();
840 for (int i = start; i < end; i++) {
841 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
842 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
843 if (scobj->first_index(jvms) == sfpt_done->req() &&
844 scobj->n_fields() == (uint)nfields) {
845 assert(scobj->alloc() == alloc, "sanity");
846 sfpt_done->set_req(i, res);
847 }
848 }
849 }
850 _igvn._worklist.push(sfpt_done);
851 }
852 }
853
854 SafePointScalarObjectNode* PhaseMacroExpand::create_scalarized_object_description(AllocateNode *alloc, SafePointNode* sfpt,
855 Unique_Node_List* value_worklist) {
856 // Fields of scalar objs are referenced only at the end
857 // of regular debuginfo at the last (youngest) JVMS.
858 // Record relative start index.
859 ciInstanceKlass* iklass = nullptr;
860 BasicType basic_elem_type = T_ILLEGAL;
861 const Type* field_type = nullptr;
862 const TypeOopPtr* res_type = nullptr;
863 int nfields = 0;
864 int array_base = 0;
865 int element_size = 0;
866 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
867 Node* res = alloc->result_cast();
868
869 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
870 assert(sfpt->jvms() != nullptr, "missed JVMS");
871
872 if (res != nullptr) { // Could be null when there are no users
873 res_type = _igvn.type(res)->isa_oopptr();
874
875 if (res_type->isa_instptr()) {
876 // find the fields of the class which will be needed for safepoint debug information
877 iklass = res_type->is_instptr()->instance_klass();
878 nfields = iklass->nof_nonstatic_fields();
879 } else {
880 // find the array's elements which will be needed for safepoint debug information
881 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
882 assert(nfields >= 0, "must be an array klass.");
883 basic_elem_type = res_type->is_aryptr()->elem()->array_element_basic_type();
884 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
885 element_size = type2aelembytes(basic_elem_type);
886 field_type = res_type->is_aryptr()->elem();
887 if (res_type->is_flat()) {
888 // Flat inline type array
889 element_size = res_type->is_aryptr()->flat_elem_size();
890 }
891 }
892
893 if (res->bottom_type()->is_inlinetypeptr()) {
894 // Nullable inline types have an IsInit field which is added to the safepoint when scalarizing them (see
895 // InlineTypeNode::make_scalar_in_safepoint()). When having circular inline types, we stop scalarizing at depth 1
896 // to avoid an endless recursion. Therefore, we do not have a SafePointScalarObjectNode node here, yet.
897 // We are about to create a SafePointScalarObjectNode as if this is a normal object. Add an additional int input
898 // with value 1 which sets IsInit to true to indicate that the object is always non-null. This input is checked
899 // later in PhaseOutput::filLocArray() for inline types.
900 sfpt->add_req(_igvn.intcon(1));
901 }
902 }
903
904 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type, alloc, first_ind, sfpt->jvms()->depth(), nfields);
905 sobj->init_req(0, C->root());
906 transform_later(sobj);
907
908 // Scan object's fields adding an input to the safepoint for each field.
909 for (int j = 0; j < nfields; j++) {
910 intptr_t offset;
911 ciField* field = nullptr;
912 if (iklass != nullptr) {
913 field = iklass->nonstatic_field_at(j);
914 offset = field->offset_in_bytes();
915 ciType* elem_type = field->type();
916 basic_elem_type = field->layout_type();
917 assert(!field->is_flat(), "flat inline type fields should not have safepoint uses");
918
919 ciField* flat_field = iklass->get_non_flat_field_by_offset(offset);
920 if (flat_field != nullptr && flat_field->is_flat() && !flat_field->is_null_free()) {
921 // TODO 8353432 Add support for nullable, flat fields in non-value class holders
922 // Below code only iterates over the flat representation and therefore misses to
923 // add null markers like we do in InlineTypeNode::add_fields_to_safepoint for value
924 // class holders.
925 return nullptr;
926 }
927
928 // The next code is taken from Parse::do_get_xxx().
929 if (is_reference_type(basic_elem_type)) {
930 if (!elem_type->is_loaded()) {
931 field_type = TypeInstPtr::BOTTOM;
932 } else if (field != nullptr && field->is_static_constant()) {
933 ciObject* con = field->constant_value().as_object();
934 // Do not "join" in the previous type; it doesn't add value,
935 // and may yield a vacuous result if the field is of interface type.
936 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
937 assert(field_type != nullptr, "field singleton type must be consistent");
938 } else {
939 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
940 }
941 if (UseCompressedOops) {
942 field_type = field_type->make_narrowoop();
943 basic_elem_type = T_NARROWOOP;
944 }
945 } else {
946 field_type = Type::get_const_basic_type(basic_elem_type);
947 }
948 } else {
949 offset = array_base + j * (intptr_t)element_size;
950 }
951
952 Node* field_val = nullptr;
953 const TypeOopPtr* field_addr_type = res_type->add_offset(offset)->isa_oopptr();
954 if (res_type->is_flat()) {
955 ciInlineKlass* inline_klass = res_type->is_aryptr()->elem()->inline_klass();
956 assert(inline_klass->flat_in_array(), "must be flat in array");
957 field_val = inline_type_from_mem(sfpt->memory(), sfpt->control(), inline_klass, field_addr_type->isa_aryptr(), 0, alloc);
958 } else {
959 field_val = value_from_mem(sfpt->memory(), sfpt->control(), basic_elem_type, field_type, field_addr_type, alloc);
960 }
961
962 // We weren't able to find a value for this field,
963 // give up on eliminating this allocation.
964 if (field_val == nullptr) {
965 uint last = sfpt->req() - 1;
966 for (int k = 0; k < j; k++) {
967 sfpt->del_req(last--);
968 }
969 _igvn._worklist.push(sfpt);
970
971 #ifndef PRODUCT
972 if (PrintEliminateAllocations) {
973 if (field != nullptr) {
974 tty->print("=== At SafePoint node %d can't find value of field: ", sfpt->_idx);
975 field->print();
976 int field_idx = C->get_alias_index(field_addr_type);
977 tty->print(" (alias_idx=%d)", field_idx);
978 } else { // Array's element
979 tty->print("=== At SafePoint node %d can't find value of array element [%d]", sfpt->_idx, j);
980 }
981 tty->print(", which prevents elimination of: ");
982 if (res == nullptr)
983 alloc->dump();
984 else
985 res->dump();
986 }
987 #endif
988
989 return nullptr;
990 }
991
992 if (UseCompressedOops && field_type->isa_narrowoop()) {
993 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
994 // to be able scalar replace the allocation.
995 if (field_val->is_EncodeP()) {
996 field_val = field_val->in(1);
997 } else if (!field_val->is_InlineType()) {
998 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
999 }
1000 }
1001
1002 // Keep track of inline types to scalarize them later
1003 if (field_val->is_InlineType()) {
1004 value_worklist->push(field_val);
1005 } else if (field_val->is_Phi()) {
1006 PhiNode* phi = field_val->as_Phi();
1007 // Eagerly replace inline type phis now since we could be removing an inline type allocation where we must
1008 // scalarize all its fields in safepoints.
1009 field_val = phi->try_push_inline_types_down(&_igvn, true);
1010 if (field_val->is_InlineType()) {
1011 value_worklist->push(field_val);
1012 }
1013 }
1014 sfpt->add_req(field_val);
1015 }
1016
1017 sfpt->jvms()->set_endoff(sfpt->req());
1018
1019 return sobj;
1020 }
1021
1022 // Do scalar replacement.
1023 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
1024 GrowableArray <SafePointNode *> safepoints_done;
1025 Node* res = alloc->result_cast();
1026 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
1027 const TypeOopPtr* res_type = nullptr;
1028 if (res != nullptr) { // Could be null when there are no users
1029 res_type = _igvn.type(res)->isa_oopptr();
1030 }
1031
1032 // Process the safepoint uses
1033 assert(safepoints.length() == 0 || !res_type->is_inlinetypeptr() || C->has_circular_inline_type(),
1034 "Inline type allocations should have been scalarized earlier");
1035 Unique_Node_List value_worklist;
1036 while (safepoints.length() > 0) {
1037 SafePointNode* sfpt = safepoints.pop();
1038 SafePointScalarObjectNode* sobj = create_scalarized_object_description(alloc, sfpt, &value_worklist);
1039
1040 if (sobj == nullptr) {
1041 undo_previous_scalarizations(safepoints_done, alloc);
1042 return false;
1043 }
1044
1045 // Now make a pass over the debug information replacing any references
1046 // to the allocated object with "sobj"
1047 JVMState *jvms = sfpt->jvms();
1048 sfpt->replace_edges_in_range(res, sobj, jvms->debug_start(), jvms->debug_end(), &_igvn);
1049 _igvn._worklist.push(sfpt);
1050
1051 // keep it for rollback
1052 safepoints_done.append_if_missing(sfpt);
1053 }
1054 // Scalarize inline types that were added to the safepoint.
1055 // Don't allow linking a constant oop (if available) for flat array elements
1056 // because Deoptimization::reassign_flat_array_elements needs field values.
1057 bool allow_oop = (res_type != nullptr) && !res_type->is_flat();
1058 for (uint i = 0; i < value_worklist.size(); ++i) {
1059 InlineTypeNode* vt = value_worklist.at(i)->as_InlineType();
1060 vt->make_scalar_in_safepoints(&_igvn, allow_oop);
1061 }
1062 return true;
1063 }
1064
1065 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
1066 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control);
1067 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory);
1068 if (ctl_proj != nullptr) {
1069 igvn.replace_node(ctl_proj, n->in(0));
1070 }
1071 if (mem_proj != nullptr) {
1072 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
1073 }
1074 }
1075
1076 // Process users of eliminated allocation.
1077 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc, bool inline_alloc) {
1078 Unique_Node_List worklist;
1079 Node* res = alloc->result_cast();
1080 if (res != nullptr) {
1081 worklist.push(res);
1082 }
1083 while (worklist.size() > 0) {
1084 res = worklist.pop();
1085 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
1086 Node *use = res->last_out(j);
1087 uint oc1 = res->outcnt();
1088
1089 if (use->is_AddP()) {
1090 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
1091 Node *n = use->last_out(k);
1092 uint oc2 = use->outcnt();
1093 if (n->is_Store()) {
1094 for (DUIterator_Fast pmax, p = n->fast_outs(pmax); p < pmax; p++) {
1095 MemBarNode* mb = n->fast_out(p)->isa_MemBar();
1096 if (mb != nullptr && mb->req() <= MemBarNode::Precedent && mb->in(MemBarNode::Precedent) == n) {
1097 // MemBarVolatiles should have been removed by MemBarNode::Ideal() for non-inline allocations
1098 assert(inline_alloc, "MemBarVolatile should be eliminated for non-escaping object");
1099 mb->remove(&_igvn);
1100 }
1101 }
1102 _igvn.replace_node(n, n->in(MemNode::Memory));
1103 } else {
1104 eliminate_gc_barrier(n);
1105 }
1106 k -= (oc2 - use->outcnt());
1107 }
1108 _igvn.remove_dead_node(use);
1109 } else if (use->is_ArrayCopy()) {
1110 // Disconnect ArrayCopy node
1111 ArrayCopyNode* ac = use->as_ArrayCopy();
1112 if (ac->is_clonebasic()) {
1113 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
1114 disconnect_projections(ac, _igvn);
1115 assert(alloc->in(TypeFunc::Memory)->is_Proj() && alloc->in(TypeFunc::Memory)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
1116 Node* membar_before = alloc->in(TypeFunc::Memory)->in(0);
1117 disconnect_projections(membar_before->as_MemBar(), _igvn);
1118 if (membar_after->is_MemBar()) {
1119 disconnect_projections(membar_after->as_MemBar(), _igvn);
1120 }
1121 } else {
1122 assert(ac->is_arraycopy_validated() ||
1123 ac->is_copyof_validated() ||
1124 ac->is_copyofrange_validated(), "unsupported");
1125 CallProjections* callprojs = ac->extract_projections(true);
1126
1127 _igvn.replace_node(callprojs->fallthrough_ioproj, ac->in(TypeFunc::I_O));
1128 _igvn.replace_node(callprojs->fallthrough_memproj, ac->in(TypeFunc::Memory));
1129 _igvn.replace_node(callprojs->fallthrough_catchproj, ac->in(TypeFunc::Control));
1130
1131 // Set control to top. IGVN will remove the remaining projections
1132 ac->set_req(0, top());
1133 ac->replace_edge(res, top(), &_igvn);
1134
1135 // Disconnect src right away: it can help find new
1136 // opportunities for allocation elimination
1137 Node* src = ac->in(ArrayCopyNode::Src);
1138 ac->replace_edge(src, top(), &_igvn);
1139 // src can be top at this point if src and dest of the
1140 // arraycopy were the same
1141 if (src->outcnt() == 0 && !src->is_top()) {
1142 _igvn.remove_dead_node(src);
1143 }
1144 }
1145 _igvn._worklist.push(ac);
1146 } else if (use->is_InlineType()) {
1147 assert(use->as_InlineType()->get_oop() == res, "unexpected inline type ptr use");
1148 // Cut off oop input and remove known instance id from type
1149 _igvn.rehash_node_delayed(use);
1150 use->as_InlineType()->set_oop(_igvn, _igvn.zerocon(T_OBJECT));
1151 const TypeOopPtr* toop = _igvn.type(use)->is_oopptr()->cast_to_instance_id(TypeOopPtr::InstanceBot);
1152 _igvn.set_type(use, toop);
1153 use->as_InlineType()->set_type(toop);
1154 // Process users
1155 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); k < kmax; k++) {
1156 Node* u = use->fast_out(k);
1157 if (!u->is_InlineType()) {
1158 worklist.push(u);
1159 }
1160 }
1161 } else if (use->Opcode() == Op_StoreX && use->in(MemNode::Address) == res) {
1162 // Store to mark word of inline type larval buffer
1163 assert(inline_alloc, "Unexpected store to mark word");
1164 _igvn.replace_node(use, use->in(MemNode::Memory));
1165 } else if (use->Opcode() == Op_MemBarRelease || use->Opcode() == Op_MemBarStoreStore) {
1166 // Inline type buffer allocations are followed by a membar
1167 assert(inline_alloc, "Unexpected MemBarRelease");
1168 use->as_MemBar()->remove(&_igvn);
1169 } else {
1170 eliminate_gc_barrier(use);
1171 }
1172 j -= (oc1 - res->outcnt());
1173 }
1174 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
1175 _igvn.remove_dead_node(res);
1176 }
1177
1178 //
1179 // Process other users of allocation's projections
1180 //
1181 if (_callprojs->resproj[0] != nullptr && _callprojs->resproj[0]->outcnt() != 0) {
1182 // First disconnect stores captured by Initialize node.
1183 // If Initialize node is eliminated first in the following code,
1184 // it will kill such stores and DUIterator_Last will assert.
1185 for (DUIterator_Fast jmax, j = _callprojs->resproj[0]->fast_outs(jmax); j < jmax; j++) {
1186 Node* use = _callprojs->resproj[0]->fast_out(j);
1187 if (use->is_AddP()) {
1188 // raw memory addresses used only by the initialization
1189 _igvn.replace_node(use, C->top());
1190 --j; --jmax;
1191 }
1192 }
1193 for (DUIterator_Last jmin, j = _callprojs->resproj[0]->last_outs(jmin); j >= jmin; ) {
1194 Node* use = _callprojs->resproj[0]->last_out(j);
1195 uint oc1 = _callprojs->resproj[0]->outcnt();
1196 if (use->is_Initialize()) {
1197 // Eliminate Initialize node.
1198 InitializeNode *init = use->as_Initialize();
1199 assert(init->outcnt() <= 2, "only a control and memory projection expected");
1200 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control);
1201 if (ctrl_proj != nullptr) {
1202 _igvn.replace_node(ctrl_proj, init->in(TypeFunc::Control));
1203 #ifdef ASSERT
1204 // If the InitializeNode has no memory out, it will die, and tmp will become null
1205 Node* tmp = init->in(TypeFunc::Control);
1206 assert(tmp == nullptr || tmp == _callprojs->fallthrough_catchproj, "allocation control projection");
1207 #endif
1208 }
1209 Node *mem_proj = init->proj_out_or_null(TypeFunc::Memory);
1210 if (mem_proj != nullptr) {
1211 Node *mem = init->in(TypeFunc::Memory);
1212 #ifdef ASSERT
1213 if (mem->is_MergeMem()) {
1214 assert(mem->in(TypeFunc::Memory) == _callprojs->fallthrough_memproj, "allocation memory projection");
1215 } else {
1216 assert(mem == _callprojs->fallthrough_memproj, "allocation memory projection");
1217 }
1218 #endif
1219 _igvn.replace_node(mem_proj, mem);
1220 }
1221 } else if (use->Opcode() == Op_MemBarStoreStore) {
1222 // Inline type buffer allocations are followed by a membar
1223 assert(inline_alloc, "Unexpected MemBarStoreStore");
1224 use->as_MemBar()->remove(&_igvn);
1225 } else {
1226 assert(false, "only Initialize or AddP expected");
1227 }
1228 j -= (oc1 - _callprojs->resproj[0]->outcnt());
1229 }
1230 }
1231 if (_callprojs->fallthrough_catchproj != nullptr) {
1232 _igvn.replace_node(_callprojs->fallthrough_catchproj, alloc->in(TypeFunc::Control));
1233 }
1234 if (_callprojs->fallthrough_memproj != nullptr) {
1235 _igvn.replace_node(_callprojs->fallthrough_memproj, alloc->in(TypeFunc::Memory));
1236 }
1237 if (_callprojs->catchall_memproj != nullptr) {
1238 _igvn.replace_node(_callprojs->catchall_memproj, C->top());
1239 }
1240 if (_callprojs->fallthrough_ioproj != nullptr) {
1241 _igvn.replace_node(_callprojs->fallthrough_ioproj, alloc->in(TypeFunc::I_O));
1242 }
1243 if (_callprojs->catchall_ioproj != nullptr) {
1244 _igvn.replace_node(_callprojs->catchall_ioproj, C->top());
1245 }
1246 if (_callprojs->catchall_catchproj != nullptr) {
1247 _igvn.replace_node(_callprojs->catchall_catchproj, C->top());
1248 }
1249 }
1250
1251 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1252 // If reallocation fails during deoptimization we'll pop all
1253 // interpreter frames for this compiled frame and that won't play
1254 // nice with JVMTI popframe.
1255 // We avoid this issue by eager reallocation when the popframe request
1256 // is received.
1257 if (!EliminateAllocations) {
1258 return false;
1259 }
1260 Node* klass = alloc->in(AllocateNode::KlassNode);
1261 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1262
1263 // Attempt to eliminate inline type buffer allocations
1264 // regardless of usage and escape/replaceable status.
1265 bool inline_alloc = tklass->isa_instklassptr() &&
1266 tklass->is_instklassptr()->instance_klass()->is_inlinetype();
1267 if (!alloc->_is_non_escaping && !inline_alloc) {
1268 return false;
1269 }
1270 // Eliminate boxing allocations which are not used
1271 // regardless scalar replaceable status.
1272 Node* res = alloc->result_cast();
1273 bool boxing_alloc = (res == nullptr) && C->eliminate_boxing() &&
1274 tklass->isa_instklassptr() &&
1275 tklass->is_instklassptr()->instance_klass()->is_box_klass();
1276 if (!alloc->_is_scalar_replaceable && !boxing_alloc && !inline_alloc) {
1277 return false;
1278 }
1279
1280 _callprojs = alloc->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1281
1282 GrowableArray <SafePointNode *> safepoints;
1283 if (!can_eliminate_allocation(&_igvn, alloc, &safepoints)) {
1284 return false;
1285 }
1286
1287 if (!alloc->_is_scalar_replaceable) {
1288 assert(res == nullptr || inline_alloc, "sanity");
1289 // We can only eliminate allocation if all debug info references
1290 // are already replaced with SafePointScalarObject because
1291 // we can't search for a fields value without instance_id.
1292 if (safepoints.length() > 0) {
1293 assert(!inline_alloc || C->has_circular_inline_type(),
1294 "Inline type allocations should have been scalarized earlier");
1295 return false;
1296 }
1297 }
1298
1299 if (!scalar_replacement(alloc, safepoints)) {
1300 return false;
1301 }
1302
1303 CompileLog* log = C->log();
1304 if (log != nullptr) {
1305 log->head("eliminate_allocation type='%d'",
1306 log->identify(tklass->exact_klass()));
1307 JVMState* p = alloc->jvms();
1308 while (p != nullptr) {
1309 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1310 p = p->caller();
1311 }
1312 log->tail("eliminate_allocation");
1313 }
1314
1315 process_users_of_allocation(alloc, inline_alloc);
1316
1317 #ifndef PRODUCT
1318 if (PrintEliminateAllocations) {
1319 if (alloc->is_AllocateArray())
1320 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1321 else
1322 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1323 }
1324 #endif
1325
1326 return true;
1327 }
1328
1329 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1330 // EA should remove all uses of non-escaping boxing node.
1331 if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != nullptr) {
1332 return false;
1333 }
1334
1335 assert(boxing->result_cast() == nullptr, "unexpected boxing node result");
1336
1337 _callprojs = boxing->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1338
1339 const TypeTuple* r = boxing->tf()->range_sig();
1340 assert(r->cnt() > TypeFunc::Parms, "sanity");
1341 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1342 assert(t != nullptr, "sanity");
1343
1344 CompileLog* log = C->log();
1345 if (log != nullptr) {
1346 log->head("eliminate_boxing type='%d'",
1347 log->identify(t->instance_klass()));
1348 JVMState* p = boxing->jvms();
1349 while (p != nullptr) {
1350 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1351 p = p->caller();
1352 }
1353 log->tail("eliminate_boxing");
1354 }
1355
1356 process_users_of_allocation(boxing);
1357
1358 #ifndef PRODUCT
1359 if (PrintEliminateAllocations) {
1360 tty->print("++++ Eliminated: %d ", boxing->_idx);
1361 boxing->method()->print_short_name(tty);
1362 tty->cr();
1363 }
1364 #endif
1365
1366 return true;
1367 }
1368
1369
1370 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
1371 Node* adr = basic_plus_adr(base, offset);
1372 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
1373 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered);
1374 transform_later(value);
1375 return value;
1376 }
1377
1378
1379 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
1380 Node* adr = basic_plus_adr(base, offset);
1381 mem = StoreNode::make(_igvn, ctl, mem, adr, nullptr, value, bt, MemNode::unordered);
1382 transform_later(mem);
1383 return mem;
1384 }
1385
1386 //=============================================================================
1387 //
1388 // A L L O C A T I O N
1389 //
1390 // Allocation attempts to be fast in the case of frequent small objects.
1391 // It breaks down like this:
1392 //
1393 // 1) Size in doublewords is computed. This is a constant for objects and
1394 // variable for most arrays. Doubleword units are used to avoid size
1395 // overflow of huge doubleword arrays. We need doublewords in the end for
1396 // rounding.
1397 //
1398 // 2) Size is checked for being 'too large'. Too-large allocations will go
1399 // the slow path into the VM. The slow path can throw any required
1400 // exceptions, and does all the special checks for very large arrays. The
1401 // size test can constant-fold away for objects. For objects with
1402 // finalizers it constant-folds the otherway: you always go slow with
1403 // finalizers.
1404 //
1405 // 3) If NOT using TLABs, this is the contended loop-back point.
1406 // Load-Locked the heap top. If using TLABs normal-load the heap top.
1407 //
1408 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
1409 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
1410 // "size*8" we always enter the VM, where "largish" is a constant picked small
1411 // enough that there's always space between the eden max and 4Gig (old space is
1412 // there so it's quite large) and large enough that the cost of entering the VM
1413 // is dwarfed by the cost to initialize the space.
1414 //
1415 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1416 // down. If contended, repeat at step 3. If using TLABs normal-store
1417 // adjusted heap top back down; there is no contention.
1418 //
1419 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
1420 // fields.
1421 //
1422 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1423 // oop flavor.
1424 //
1425 //=============================================================================
1426 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1427 // Allocations bigger than this always go the slow route.
1428 // This value must be small enough that allocation attempts that need to
1429 // trigger exceptions go the slow route. Also, it must be small enough so
1430 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1431 //=============================================================================j//
1432 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1433 // The allocator will coalesce int->oop copies away. See comment in
1434 // coalesce.cpp about how this works. It depends critically on the exact
1435 // code shape produced here, so if you are changing this code shape
1436 // make sure the GC info for the heap-top is correct in and around the
1437 // slow-path call.
1438 //
1439
1440 void PhaseMacroExpand::expand_allocate_common(
1441 AllocateNode* alloc, // allocation node to be expanded
1442 Node* length, // array length for an array allocation
1443 Node* init_val, // value to initialize the array with
1444 const TypeFunc* slow_call_type, // Type of slow call
1445 address slow_call_address, // Address of slow call
1446 Node* valid_length_test // whether length is valid or not
1447 )
1448 {
1449 Node* ctrl = alloc->in(TypeFunc::Control);
1450 Node* mem = alloc->in(TypeFunc::Memory);
1451 Node* i_o = alloc->in(TypeFunc::I_O);
1452 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1453 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1454 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1455 assert(ctrl != nullptr, "must have control");
1456
1457 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1458 // they will not be used if "always_slow" is set
1459 enum { slow_result_path = 1, fast_result_path = 2 };
1460 Node *result_region = nullptr;
1461 Node *result_phi_rawmem = nullptr;
1462 Node *result_phi_rawoop = nullptr;
1463 Node *result_phi_i_o = nullptr;
1464
1465 // The initial slow comparison is a size check, the comparison
1466 // we want to do is a BoolTest::gt
1467 bool expand_fast_path = true;
1468 int tv = _igvn.find_int_con(initial_slow_test, -1);
1469 if (tv >= 0) {
1470 // InitialTest has constant result
1471 // 0 - can fit in TLAB
1472 // 1 - always too big or negative
1473 assert(tv <= 1, "0 or 1 if a constant");
1474 expand_fast_path = (tv == 0);
1475 initial_slow_test = nullptr;
1476 } else {
1477 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1478 }
1479
1480 if (!UseTLAB) {
1481 // Force slow-path allocation
1482 expand_fast_path = false;
1483 initial_slow_test = nullptr;
1484 }
1485
1486 bool allocation_has_use = (alloc->result_cast() != nullptr);
1487 if (!allocation_has_use) {
1488 InitializeNode* init = alloc->initialization();
1489 if (init != nullptr) {
1490 init->remove(&_igvn);
1491 }
1492 if (expand_fast_path && (initial_slow_test == nullptr)) {
1493 // Remove allocation node and return.
1494 // Size is a non-negative constant -> no initial check needed -> directly to fast path.
1495 // Also, no usages -> empty fast path -> no fall out to slow path -> nothing left.
1496 #ifndef PRODUCT
1497 if (PrintEliminateAllocations) {
1498 tty->print("NotUsed ");
1499 Node* res = alloc->proj_out_or_null(TypeFunc::Parms);
1500 if (res != nullptr) {
1501 res->dump();
1502 } else {
1503 alloc->dump();
1504 }
1505 }
1506 #endif
1507 yank_alloc_node(alloc);
1508 return;
1509 }
1510 }
1511
1512 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1513 Node *slow_region = nullptr;
1514 Node *toobig_false = ctrl;
1515
1516 // generate the initial test if necessary
1517 if (initial_slow_test != nullptr ) {
1518 assert (expand_fast_path, "Only need test if there is a fast path");
1519 slow_region = new RegionNode(3);
1520
1521 // Now make the initial failure test. Usually a too-big test but
1522 // might be a TRUE for finalizers.
1523 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1524 transform_later(toobig_iff);
1525 // Plug the failing-too-big test into the slow-path region
1526 Node* toobig_true = new IfTrueNode(toobig_iff);
1527 transform_later(toobig_true);
1528 slow_region ->init_req( too_big_or_final_path, toobig_true );
1529 toobig_false = new IfFalseNode(toobig_iff);
1530 transform_later(toobig_false);
1531 } else {
1532 // No initial test, just fall into next case
1533 assert(allocation_has_use || !expand_fast_path, "Should already have been handled");
1534 toobig_false = ctrl;
1535 debug_only(slow_region = NodeSentinel);
1536 }
1537
1538 // If we are here there are several possibilities
1539 // - expand_fast_path is false - then only a slow path is expanded. That's it.
1540 // no_initial_check means a constant allocation.
1541 // - If check always evaluates to false -> expand_fast_path is false (see above)
1542 // - If check always evaluates to true -> directly into fast path (but may bailout to slowpath)
1543 // if !allocation_has_use the fast path is empty
1544 // if !allocation_has_use && no_initial_check
1545 // - Then there are no fastpath that can fall out to slowpath -> no allocation code at all.
1546 // removed by yank_alloc_node above.
1547
1548 Node *slow_mem = mem; // save the current memory state for slow path
1549 // generate the fast allocation code unless we know that the initial test will always go slow
1550 if (expand_fast_path) {
1551 // Fast path modifies only raw memory.
1552 if (mem->is_MergeMem()) {
1553 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1554 }
1555
1556 // allocate the Region and Phi nodes for the result
1557 result_region = new RegionNode(3);
1558 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1559 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1560
1561 // Grab regular I/O before optional prefetch may change it.
1562 // Slow-path does no I/O so just set it to the original I/O.
1563 result_phi_i_o->init_req(slow_result_path, i_o);
1564
1565 // Name successful fast-path variables
1566 Node* fast_oop_ctrl;
1567 Node* fast_oop_rawmem;
1568
1569 if (allocation_has_use) {
1570 Node* needgc_ctrl = nullptr;
1571 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1572
1573 intx prefetch_lines = length != nullptr ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1574 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1575 Node* fast_oop = bs->obj_allocate(this, mem, toobig_false, size_in_bytes, i_o, needgc_ctrl,
1576 fast_oop_ctrl, fast_oop_rawmem,
1577 prefetch_lines);
1578
1579 if (initial_slow_test != nullptr) {
1580 // This completes all paths into the slow merge point
1581 slow_region->init_req(need_gc_path, needgc_ctrl);
1582 transform_later(slow_region);
1583 } else {
1584 // No initial slow path needed!
1585 // Just fall from the need-GC path straight into the VM call.
1586 slow_region = needgc_ctrl;
1587 }
1588
1589 InitializeNode* init = alloc->initialization();
1590 fast_oop_rawmem = initialize_object(alloc,
1591 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1592 klass_node, length, size_in_bytes);
1593 expand_initialize_membar(alloc, init, fast_oop_ctrl, fast_oop_rawmem);
1594 expand_dtrace_alloc_probe(alloc, fast_oop, fast_oop_ctrl, fast_oop_rawmem);
1595
1596 result_phi_rawoop->init_req(fast_result_path, fast_oop);
1597 } else {
1598 assert (initial_slow_test != nullptr, "sanity");
1599 fast_oop_ctrl = toobig_false;
1600 fast_oop_rawmem = mem;
1601 transform_later(slow_region);
1602 }
1603
1604 // Plug in the successful fast-path into the result merge point
1605 result_region ->init_req(fast_result_path, fast_oop_ctrl);
1606 result_phi_i_o ->init_req(fast_result_path, i_o);
1607 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1608 } else {
1609 slow_region = ctrl;
1610 result_phi_i_o = i_o; // Rename it to use in the following code.
1611 }
1612
1613 // Generate slow-path call
1614 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1615 OptoRuntime::stub_name(slow_call_address),
1616 TypePtr::BOTTOM);
1617 call->init_req(TypeFunc::Control, slow_region);
1618 call->init_req(TypeFunc::I_O, top()); // does no i/o
1619 call->init_req(TypeFunc::Memory, slow_mem); // may gc ptrs
1620 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1621 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1622
1623 call->init_req(TypeFunc::Parms+0, klass_node);
1624 if (length != nullptr) {
1625 call->init_req(TypeFunc::Parms+1, length);
1626 if (init_val != nullptr) {
1627 call->init_req(TypeFunc::Parms+2, init_val);
1628 }
1629 } else {
1630 // Let the runtime know if this is a larval allocation
1631 call->init_req(TypeFunc::Parms+1, _igvn.intcon(alloc->_larval));
1632 }
1633
1634 // Copy debug information and adjust JVMState information, then replace
1635 // allocate node with the call
1636 call->copy_call_debug_info(&_igvn, alloc);
1637 // For array allocations, copy the valid length check to the call node so Compile::final_graph_reshaping() can verify
1638 // that the call has the expected number of CatchProj nodes (in case the allocation always fails and the fallthrough
1639 // path dies).
1640 if (valid_length_test != nullptr) {
1641 call->add_req(valid_length_test);
1642 }
1643 if (expand_fast_path) {
1644 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1645 } else {
1646 // Hook i_o projection to avoid its elimination during allocation
1647 // replacement (when only a slow call is generated).
1648 call->set_req(TypeFunc::I_O, result_phi_i_o);
1649 }
1650 _igvn.replace_node(alloc, call);
1651 transform_later(call);
1652
1653 // Identify the output projections from the allocate node and
1654 // adjust any references to them.
1655 // The control and io projections look like:
1656 //
1657 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1658 // Allocate Catch
1659 // ^---Proj(io) <-------+ ^---CatchProj(io)
1660 //
1661 // We are interested in the CatchProj nodes.
1662 //
1663 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1664
1665 // An allocate node has separate memory projections for the uses on
1666 // the control and i_o paths. Replace the control memory projection with
1667 // result_phi_rawmem (unless we are only generating a slow call when
1668 // both memory projections are combined)
1669 if (expand_fast_path && _callprojs->fallthrough_memproj != nullptr) {
1670 _igvn.replace_in_uses(_callprojs->fallthrough_memproj, result_phi_rawmem);
1671 }
1672 // Now change uses of catchall_memproj to use fallthrough_memproj and delete
1673 // catchall_memproj so we end up with a call that has only 1 memory projection.
1674 if (_callprojs->catchall_memproj != nullptr) {
1675 if (_callprojs->fallthrough_memproj == nullptr) {
1676 _callprojs->fallthrough_memproj = new ProjNode(call, TypeFunc::Memory);
1677 transform_later(_callprojs->fallthrough_memproj);
1678 }
1679 _igvn.replace_in_uses(_callprojs->catchall_memproj, _callprojs->fallthrough_memproj);
1680 _igvn.remove_dead_node(_callprojs->catchall_memproj);
1681 }
1682
1683 // An allocate node has separate i_o projections for the uses on the control
1684 // and i_o paths. Always replace the control i_o projection with result i_o
1685 // otherwise incoming i_o become dead when only a slow call is generated
1686 // (it is different from memory projections where both projections are
1687 // combined in such case).
1688 if (_callprojs->fallthrough_ioproj != nullptr) {
1689 _igvn.replace_in_uses(_callprojs->fallthrough_ioproj, result_phi_i_o);
1690 }
1691 // Now change uses of catchall_ioproj to use fallthrough_ioproj and delete
1692 // catchall_ioproj so we end up with a call that has only 1 i_o projection.
1693 if (_callprojs->catchall_ioproj != nullptr) {
1694 if (_callprojs->fallthrough_ioproj == nullptr) {
1695 _callprojs->fallthrough_ioproj = new ProjNode(call, TypeFunc::I_O);
1696 transform_later(_callprojs->fallthrough_ioproj);
1697 }
1698 _igvn.replace_in_uses(_callprojs->catchall_ioproj, _callprojs->fallthrough_ioproj);
1699 _igvn.remove_dead_node(_callprojs->catchall_ioproj);
1700 }
1701
1702 // if we generated only a slow call, we are done
1703 if (!expand_fast_path) {
1704 // Now we can unhook i_o.
1705 if (result_phi_i_o->outcnt() > 1) {
1706 call->set_req(TypeFunc::I_O, top());
1707 } else {
1708 assert(result_phi_i_o->unique_ctrl_out() == call, "sanity");
1709 // Case of new array with negative size known during compilation.
1710 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1711 // following code since call to runtime will throw exception.
1712 // As result there will be no users of i_o after the call.
1713 // Leave i_o attached to this call to avoid problems in preceding graph.
1714 }
1715 return;
1716 }
1717
1718 if (_callprojs->fallthrough_catchproj != nullptr) {
1719 ctrl = _callprojs->fallthrough_catchproj->clone();
1720 transform_later(ctrl);
1721 _igvn.replace_node(_callprojs->fallthrough_catchproj, result_region);
1722 } else {
1723 ctrl = top();
1724 }
1725 Node *slow_result;
1726 if (_callprojs->resproj[0] == nullptr) {
1727 // no uses of the allocation result
1728 slow_result = top();
1729 } else {
1730 slow_result = _callprojs->resproj[0]->clone();
1731 transform_later(slow_result);
1732 _igvn.replace_node(_callprojs->resproj[0], result_phi_rawoop);
1733 }
1734
1735 // Plug slow-path into result merge point
1736 result_region->init_req( slow_result_path, ctrl);
1737 transform_later(result_region);
1738 if (allocation_has_use) {
1739 result_phi_rawoop->init_req(slow_result_path, slow_result);
1740 transform_later(result_phi_rawoop);
1741 }
1742 result_phi_rawmem->init_req(slow_result_path, _callprojs->fallthrough_memproj);
1743 transform_later(result_phi_rawmem);
1744 transform_later(result_phi_i_o);
1745 // This completes all paths into the result merge point
1746 }
1747
1748 // Remove alloc node that has no uses.
1749 void PhaseMacroExpand::yank_alloc_node(AllocateNode* alloc) {
1750 Node* ctrl = alloc->in(TypeFunc::Control);
1751 Node* mem = alloc->in(TypeFunc::Memory);
1752 Node* i_o = alloc->in(TypeFunc::I_O);
1753
1754 _callprojs = alloc->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1755 if (_callprojs->resproj[0] != nullptr) {
1756 for (DUIterator_Fast imax, i = _callprojs->resproj[0]->fast_outs(imax); i < imax; i++) {
1757 Node* use = _callprojs->resproj[0]->fast_out(i);
1758 use->isa_MemBar()->remove(&_igvn);
1759 --imax;
1760 --i; // back up iterator
1761 }
1762 assert(_callprojs->resproj[0]->outcnt() == 0, "all uses must be deleted");
1763 _igvn.remove_dead_node(_callprojs->resproj[0]);
1764 }
1765 if (_callprojs->fallthrough_catchproj != nullptr) {
1766 _igvn.replace_in_uses(_callprojs->fallthrough_catchproj, ctrl);
1767 _igvn.remove_dead_node(_callprojs->fallthrough_catchproj);
1768 }
1769 if (_callprojs->catchall_catchproj != nullptr) {
1770 _igvn.rehash_node_delayed(_callprojs->catchall_catchproj);
1771 _callprojs->catchall_catchproj->set_req(0, top());
1772 }
1773 if (_callprojs->fallthrough_proj != nullptr) {
1774 Node* catchnode = _callprojs->fallthrough_proj->unique_ctrl_out();
1775 _igvn.remove_dead_node(catchnode);
1776 _igvn.remove_dead_node(_callprojs->fallthrough_proj);
1777 }
1778 if (_callprojs->fallthrough_memproj != nullptr) {
1779 _igvn.replace_in_uses(_callprojs->fallthrough_memproj, mem);
1780 _igvn.remove_dead_node(_callprojs->fallthrough_memproj);
1781 }
1782 if (_callprojs->fallthrough_ioproj != nullptr) {
1783 _igvn.replace_in_uses(_callprojs->fallthrough_ioproj, i_o);
1784 _igvn.remove_dead_node(_callprojs->fallthrough_ioproj);
1785 }
1786 if (_callprojs->catchall_memproj != nullptr) {
1787 _igvn.rehash_node_delayed(_callprojs->catchall_memproj);
1788 _callprojs->catchall_memproj->set_req(0, top());
1789 }
1790 if (_callprojs->catchall_ioproj != nullptr) {
1791 _igvn.rehash_node_delayed(_callprojs->catchall_ioproj);
1792 _callprojs->catchall_ioproj->set_req(0, top());
1793 }
1794 #ifndef PRODUCT
1795 if (PrintEliminateAllocations) {
1796 if (alloc->is_AllocateArray()) {
1797 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1798 } else {
1799 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1800 }
1801 }
1802 #endif
1803 _igvn.remove_dead_node(alloc);
1804 }
1805
1806 void PhaseMacroExpand::expand_initialize_membar(AllocateNode* alloc, InitializeNode* init,
1807 Node*& fast_oop_ctrl, Node*& fast_oop_rawmem) {
1808 // If initialization is performed by an array copy, any required
1809 // MemBarStoreStore was already added. If the object does not
1810 // escape no need for a MemBarStoreStore. If the object does not
1811 // escape in its initializer and memory barrier (MemBarStoreStore or
1812 // stronger) is already added at exit of initializer, also no need
1813 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore
1814 // so that stores that initialize this object can't be reordered
1815 // with a subsequent store that makes this object accessible by
1816 // other threads.
1817 // Other threads include java threads and JVM internal threads
1818 // (for example concurrent GC threads). Current concurrent GC
1819 // implementation: G1 will not scan newly created object,
1820 // so it's safe to skip storestore barrier when allocation does
1821 // not escape.
1822 if (!alloc->does_not_escape_thread() &&
1823 !alloc->is_allocation_MemBar_redundant() &&
1824 (init == nullptr || !init->is_complete_with_arraycopy())) {
1825 if (init == nullptr || init->req() < InitializeNode::RawStores) {
1826 // No InitializeNode or no stores captured by zeroing
1827 // elimination. Simply add the MemBarStoreStore after object
1828 // initialization.
1829 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1830 transform_later(mb);
1831
1832 mb->init_req(TypeFunc::Memory, fast_oop_rawmem);
1833 mb->init_req(TypeFunc::Control, fast_oop_ctrl);
1834 fast_oop_ctrl = new ProjNode(mb, TypeFunc::Control);
1835 transform_later(fast_oop_ctrl);
1836 fast_oop_rawmem = new ProjNode(mb, TypeFunc::Memory);
1837 transform_later(fast_oop_rawmem);
1838 } else {
1839 // Add the MemBarStoreStore after the InitializeNode so that
1840 // all stores performing the initialization that were moved
1841 // before the InitializeNode happen before the storestore
1842 // barrier.
1843
1844 Node* init_ctrl = init->proj_out_or_null(TypeFunc::Control);
1845 Node* init_mem = init->proj_out_or_null(TypeFunc::Memory);
1846
1847 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1848 transform_later(mb);
1849
1850 Node* ctrl = new ProjNode(init, TypeFunc::Control);
1851 transform_later(ctrl);
1852 Node* mem = new ProjNode(init, TypeFunc::Memory);
1853 transform_later(mem);
1854
1855 // The MemBarStoreStore depends on control and memory coming
1856 // from the InitializeNode
1857 mb->init_req(TypeFunc::Memory, mem);
1858 mb->init_req(TypeFunc::Control, ctrl);
1859
1860 ctrl = new ProjNode(mb, TypeFunc::Control);
1861 transform_later(ctrl);
1862 mem = new ProjNode(mb, TypeFunc::Memory);
1863 transform_later(mem);
1864
1865 // All nodes that depended on the InitializeNode for control
1866 // and memory must now depend on the MemBarNode that itself
1867 // depends on the InitializeNode
1868 if (init_ctrl != nullptr) {
1869 _igvn.replace_node(init_ctrl, ctrl);
1870 }
1871 if (init_mem != nullptr) {
1872 _igvn.replace_node(init_mem, mem);
1873 }
1874 }
1875 }
1876 }
1877
1878 void PhaseMacroExpand::expand_dtrace_alloc_probe(AllocateNode* alloc, Node* oop,
1879 Node*& ctrl, Node*& rawmem) {
1880 if (C->env()->dtrace_alloc_probes()) {
1881 // Slow-path call
1882 int size = TypeFunc::Parms + 2;
1883 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1884 CAST_FROM_FN_PTR(address,
1885 static_cast<int (*)(JavaThread*, oopDesc*)>(SharedRuntime::dtrace_object_alloc)),
1886 "dtrace_object_alloc",
1887 TypeRawPtr::BOTTOM);
1888
1889 // Get base of thread-local storage area
1890 Node* thread = new ThreadLocalNode();
1891 transform_later(thread);
1892
1893 call->init_req(TypeFunc::Parms + 0, thread);
1894 call->init_req(TypeFunc::Parms + 1, oop);
1895 call->init_req(TypeFunc::Control, ctrl);
1896 call->init_req(TypeFunc::I_O , top()); // does no i/o
1897 call->init_req(TypeFunc::Memory , rawmem);
1898 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1899 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1900 transform_later(call);
1901 ctrl = new ProjNode(call, TypeFunc::Control);
1902 transform_later(ctrl);
1903 rawmem = new ProjNode(call, TypeFunc::Memory);
1904 transform_later(rawmem);
1905 }
1906 }
1907
1908 // Helper for PhaseMacroExpand::expand_allocate_common.
1909 // Initializes the newly-allocated storage.
1910 Node* PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1911 Node* control, Node* rawmem, Node* object,
1912 Node* klass_node, Node* length,
1913 Node* size_in_bytes) {
1914 InitializeNode* init = alloc->initialization();
1915 // Store the klass & mark bits
1916 Node* mark_node = alloc->make_ideal_mark(&_igvn, control, rawmem);
1917 if (!mark_node->is_Con()) {
1918 transform_later(mark_node);
1919 }
1920 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, TypeX_X->basic_type());
1921
1922 if (!UseCompactObjectHeaders) {
1923 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1924 }
1925 int header_size = alloc->minimum_header_size(); // conservatively small
1926
1927 // Array length
1928 if (length != nullptr) { // Arrays need length field
1929 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1930 // conservatively small header size:
1931 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1932 if (_igvn.type(klass_node)->isa_aryklassptr()) { // we know the exact header size in most cases:
1933 BasicType elem = _igvn.type(klass_node)->is_klassptr()->as_instance_type()->isa_aryptr()->elem()->array_element_basic_type();
1934 if (is_reference_type(elem, true)) {
1935 elem = T_OBJECT;
1936 }
1937 header_size = Klass::layout_helper_header_size(Klass::array_layout_helper(elem));
1938 }
1939 }
1940
1941 // Clear the object body, if necessary.
1942 if (init == nullptr) {
1943 // The init has somehow disappeared; be cautious and clear everything.
1944 //
1945 // This can happen if a node is allocated but an uncommon trap occurs
1946 // immediately. In this case, the Initialize gets associated with the
1947 // trap, and may be placed in a different (outer) loop, if the Allocate
1948 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1949 // there can be two Allocates to one Initialize. The answer in all these
1950 // edge cases is safety first. It is always safe to clear immediately
1951 // within an Allocate, and then (maybe or maybe not) clear some more later.
1952 if (!(UseTLAB && ZeroTLAB)) {
1953 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1954 alloc->in(AllocateNode::InitValue),
1955 alloc->in(AllocateNode::RawInitValue),
1956 header_size, size_in_bytes,
1957 &_igvn);
1958 }
1959 } else {
1960 if (!init->is_complete()) {
1961 // Try to win by zeroing only what the init does not store.
1962 // We can also try to do some peephole optimizations,
1963 // such as combining some adjacent subword stores.
1964 rawmem = init->complete_stores(control, rawmem, object,
1965 header_size, size_in_bytes, &_igvn);
1966 }
1967 // We have no more use for this link, since the AllocateNode goes away:
1968 init->set_req(InitializeNode::RawAddress, top());
1969 // (If we keep the link, it just confuses the register allocator,
1970 // who thinks he sees a real use of the address by the membar.)
1971 }
1972
1973 return rawmem;
1974 }
1975
1976 // Generate prefetch instructions for next allocations.
1977 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
1978 Node*& contended_phi_rawmem,
1979 Node* old_eden_top, Node* new_eden_top,
1980 intx lines) {
1981 enum { fall_in_path = 1, pf_path = 2 };
1982 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
1983 // Generate prefetch allocation with watermark check.
1984 // As an allocation hits the watermark, we will prefetch starting
1985 // at a "distance" away from watermark.
1986
1987 Node *pf_region = new RegionNode(3);
1988 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY,
1989 TypeRawPtr::BOTTOM );
1990 // I/O is used for Prefetch
1991 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO );
1992
1993 Node *thread = new ThreadLocalNode();
1994 transform_later(thread);
1995
1996 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread,
1997 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
1998 transform_later(eden_pf_adr);
1999
2000 Node *old_pf_wm = new LoadPNode(needgc_false,
2001 contended_phi_rawmem, eden_pf_adr,
2002 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM,
2003 MemNode::unordered);
2004 transform_later(old_pf_wm);
2005
2006 // check against new_eden_top
2007 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm );
2008 transform_later(need_pf_cmp);
2009 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge );
2010 transform_later(need_pf_bol);
2011 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol,
2012 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
2013 transform_later(need_pf_iff);
2014
2015 // true node, add prefetchdistance
2016 Node *need_pf_true = new IfTrueNode( need_pf_iff );
2017 transform_later(need_pf_true);
2018
2019 Node *need_pf_false = new IfFalseNode( need_pf_iff );
2020 transform_later(need_pf_false);
2021
2022 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm,
2023 _igvn.MakeConX(AllocatePrefetchDistance) );
2024 transform_later(new_pf_wmt );
2025 new_pf_wmt->set_req(0, need_pf_true);
2026
2027 Node *store_new_wmt = new StorePNode(need_pf_true,
2028 contended_phi_rawmem, eden_pf_adr,
2029 TypeRawPtr::BOTTOM, new_pf_wmt,
2030 MemNode::unordered);
2031 transform_later(store_new_wmt);
2032
2033 // adding prefetches
2034 pf_phi_abio->init_req( fall_in_path, i_o );
2035
2036 Node *prefetch_adr;
2037 Node *prefetch;
2038 uint step_size = AllocatePrefetchStepSize;
2039 uint distance = 0;
2040
2041 for ( intx i = 0; i < lines; i++ ) {
2042 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt,
2043 _igvn.MakeConX(distance) );
2044 transform_later(prefetch_adr);
2045 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
2046 transform_later(prefetch);
2047 distance += step_size;
2048 i_o = prefetch;
2049 }
2050 pf_phi_abio->set_req( pf_path, i_o );
2051
2052 pf_region->init_req( fall_in_path, need_pf_false );
2053 pf_region->init_req( pf_path, need_pf_true );
2054
2055 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
2056 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
2057
2058 transform_later(pf_region);
2059 transform_later(pf_phi_rawmem);
2060 transform_later(pf_phi_abio);
2061
2062 needgc_false = pf_region;
2063 contended_phi_rawmem = pf_phi_rawmem;
2064 i_o = pf_phi_abio;
2065 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) {
2066 // Insert a prefetch instruction for each allocation.
2067 // This code is used to generate 1 prefetch instruction per cache line.
2068
2069 // Generate several prefetch instructions.
2070 uint step_size = AllocatePrefetchStepSize;
2071 uint distance = AllocatePrefetchDistance;
2072
2073 // Next cache address.
2074 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top,
2075 _igvn.MakeConX(step_size + distance));
2076 transform_later(cache_adr);
2077 cache_adr = new CastP2XNode(needgc_false, cache_adr);
2078 transform_later(cache_adr);
2079 // Address is aligned to execute prefetch to the beginning of cache line size
2080 // (it is important when BIS instruction is used on SPARC as prefetch).
2081 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1));
2082 cache_adr = new AndXNode(cache_adr, mask);
2083 transform_later(cache_adr);
2084 cache_adr = new CastX2PNode(cache_adr);
2085 transform_later(cache_adr);
2086
2087 // Prefetch
2088 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr );
2089 prefetch->set_req(0, needgc_false);
2090 transform_later(prefetch);
2091 contended_phi_rawmem = prefetch;
2092 Node *prefetch_adr;
2093 distance = step_size;
2094 for ( intx i = 1; i < lines; i++ ) {
2095 prefetch_adr = new AddPNode( cache_adr, cache_adr,
2096 _igvn.MakeConX(distance) );
2097 transform_later(prefetch_adr);
2098 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr );
2099 transform_later(prefetch);
2100 distance += step_size;
2101 contended_phi_rawmem = prefetch;
2102 }
2103 } else if( AllocatePrefetchStyle > 0 ) {
2104 // Insert a prefetch for each allocation only on the fast-path
2105 Node *prefetch_adr;
2106 Node *prefetch;
2107 // Generate several prefetch instructions.
2108 uint step_size = AllocatePrefetchStepSize;
2109 uint distance = AllocatePrefetchDistance;
2110 for ( intx i = 0; i < lines; i++ ) {
2111 prefetch_adr = new AddPNode( old_eden_top, new_eden_top,
2112 _igvn.MakeConX(distance) );
2113 transform_later(prefetch_adr);
2114 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
2115 // Do not let it float too high, since if eden_top == eden_end,
2116 // both might be null.
2117 if( i == 0 ) { // Set control for first prefetch, next follows it
2118 prefetch->init_req(0, needgc_false);
2119 }
2120 transform_later(prefetch);
2121 distance += step_size;
2122 i_o = prefetch;
2123 }
2124 }
2125 return i_o;
2126 }
2127
2128
2129 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
2130 expand_allocate_common(alloc, nullptr, nullptr,
2131 OptoRuntime::new_instance_Type(),
2132 OptoRuntime::new_instance_Java(), nullptr);
2133 }
2134
2135 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
2136 Node* length = alloc->in(AllocateNode::ALength);
2137 Node* valid_length_test = alloc->in(AllocateNode::ValidLengthTest);
2138 InitializeNode* init = alloc->initialization();
2139 Node* klass_node = alloc->in(AllocateNode::KlassNode);
2140 Node* init_value = alloc->in(AllocateNode::InitValue);
2141 const TypeAryKlassPtr* ary_klass_t = _igvn.type(klass_node)->isa_aryklassptr();
2142 const TypeFunc* slow_call_type;
2143 address slow_call_address; // Address of slow call
2144 if (init != nullptr && init->is_complete_with_arraycopy() &&
2145 ary_klass_t && ary_klass_t->elem()->isa_klassptr() == nullptr) {
2146 // Don't zero type array during slow allocation in VM since
2147 // it will be initialized later by arraycopy in compiled code.
2148 slow_call_address = OptoRuntime::new_array_nozero_Java();
2149 slow_call_type = OptoRuntime::new_array_nozero_Type();
2150 } else {
2151 slow_call_address = OptoRuntime::new_array_Java();
2152 slow_call_type = OptoRuntime::new_array_Type();
2153
2154 if (init_value == nullptr) {
2155 init_value = _igvn.zerocon(T_OBJECT);
2156 } else if (UseCompressedOops) {
2157 init_value = transform_later(new DecodeNNode(init_value, init_value->bottom_type()->make_ptr()));
2158 }
2159 }
2160 expand_allocate_common(alloc, length, init_value,
2161 slow_call_type,
2162 slow_call_address, valid_length_test);
2163 }
2164
2165 //-------------------mark_eliminated_box----------------------------------
2166 //
2167 // During EA obj may point to several objects but after few ideal graph
2168 // transformations (CCP) it may point to only one non escaping object
2169 // (but still using phi), corresponding locks and unlocks will be marked
2170 // for elimination. Later obj could be replaced with a new node (new phi)
2171 // and which does not have escape information. And later after some graph
2172 // reshape other locks and unlocks (which were not marked for elimination
2173 // before) are connected to this new obj (phi) but they still will not be
2174 // marked for elimination since new obj has no escape information.
2175 // Mark all associated (same box and obj) lock and unlock nodes for
2176 // elimination if some of them marked already.
2177 void PhaseMacroExpand::mark_eliminated_box(Node* box, Node* obj) {
2178 BoxLockNode* oldbox = box->as_BoxLock();
2179 if (oldbox->is_eliminated()) {
2180 return; // This BoxLock node was processed already.
2181 }
2182 assert(!oldbox->is_unbalanced(), "this should not be called for unbalanced region");
2183 // New implementation (EliminateNestedLocks) has separate BoxLock
2184 // node for each locked region so mark all associated locks/unlocks as
2185 // eliminated even if different objects are referenced in one locked region
2186 // (for example, OSR compilation of nested loop inside locked scope).
2187 if (EliminateNestedLocks ||
2188 oldbox->as_BoxLock()->is_simple_lock_region(nullptr, obj, nullptr)) {
2189 // Box is used only in one lock region. Mark this box as eliminated.
2190 oldbox->set_local(); // This verifies correct state of BoxLock
2191 _igvn.hash_delete(oldbox);
2192 oldbox->set_eliminated(); // This changes box's hash value
2193 _igvn.hash_insert(oldbox);
2194
2195 for (uint i = 0; i < oldbox->outcnt(); i++) {
2196 Node* u = oldbox->raw_out(i);
2197 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) {
2198 AbstractLockNode* alock = u->as_AbstractLock();
2199 // Check lock's box since box could be referenced by Lock's debug info.
2200 if (alock->box_node() == oldbox) {
2201 // Mark eliminated all related locks and unlocks.
2202 #ifdef ASSERT
2203 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4");
2204 #endif
2205 alock->set_non_esc_obj();
2206 }
2207 }
2208 }
2209 return;
2210 }
2211
2212 // Create new "eliminated" BoxLock node and use it in monitor debug info
2213 // instead of oldbox for the same object.
2214 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
2215
2216 // Note: BoxLock node is marked eliminated only here and it is used
2217 // to indicate that all associated lock and unlock nodes are marked
2218 // for elimination.
2219 newbox->set_local(); // This verifies correct state of BoxLock
2220 newbox->set_eliminated();
2221 transform_later(newbox);
2222
2223 // Replace old box node with new box for all users of the same object.
2224 for (uint i = 0; i < oldbox->outcnt();) {
2225 bool next_edge = true;
2226
2227 Node* u = oldbox->raw_out(i);
2228 if (u->is_AbstractLock()) {
2229 AbstractLockNode* alock = u->as_AbstractLock();
2230 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) {
2231 // Replace Box and mark eliminated all related locks and unlocks.
2232 #ifdef ASSERT
2233 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5");
2234 #endif
2235 alock->set_non_esc_obj();
2236 _igvn.rehash_node_delayed(alock);
2237 alock->set_box_node(newbox);
2238 next_edge = false;
2239 }
2240 }
2241 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) {
2242 FastLockNode* flock = u->as_FastLock();
2243 assert(flock->box_node() == oldbox, "sanity");
2244 _igvn.rehash_node_delayed(flock);
2245 flock->set_box_node(newbox);
2246 next_edge = false;
2247 }
2248
2249 // Replace old box in monitor debug info.
2250 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
2251 SafePointNode* sfn = u->as_SafePoint();
2252 JVMState* youngest_jvms = sfn->jvms();
2253 int max_depth = youngest_jvms->depth();
2254 for (int depth = 1; depth <= max_depth; depth++) {
2255 JVMState* jvms = youngest_jvms->of_depth(depth);
2256 int num_mon = jvms->nof_monitors();
2257 // Loop over monitors
2258 for (int idx = 0; idx < num_mon; idx++) {
2259 Node* obj_node = sfn->monitor_obj(jvms, idx);
2260 Node* box_node = sfn->monitor_box(jvms, idx);
2261 if (box_node == oldbox && obj_node->eqv_uncast(obj)) {
2262 int j = jvms->monitor_box_offset(idx);
2263 _igvn.replace_input_of(u, j, newbox);
2264 next_edge = false;
2265 }
2266 }
2267 }
2268 }
2269 if (next_edge) i++;
2270 }
2271 }
2272
2273 //-----------------------mark_eliminated_locking_nodes-----------------------
2274 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) {
2275 if (!alock->is_balanced()) {
2276 return; // Can't do any more elimination for this locking region
2277 }
2278 if (EliminateNestedLocks) {
2279 if (alock->is_nested()) {
2280 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity");
2281 return;
2282 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened
2283 // Only Lock node has JVMState needed here.
2284 // Not that preceding claim is documented anywhere else.
2285 if (alock->jvms() != nullptr) {
2286 if (alock->as_Lock()->is_nested_lock_region()) {
2287 // Mark eliminated related nested locks and unlocks.
2288 Node* obj = alock->obj_node();
2289 BoxLockNode* box_node = alock->box_node()->as_BoxLock();
2290 assert(!box_node->is_eliminated(), "should not be marked yet");
2291 // Note: BoxLock node is marked eliminated only here
2292 // and it is used to indicate that all associated lock
2293 // and unlock nodes are marked for elimination.
2294 box_node->set_eliminated(); // Box's hash is always NO_HASH here
2295 for (uint i = 0; i < box_node->outcnt(); i++) {
2296 Node* u = box_node->raw_out(i);
2297 if (u->is_AbstractLock()) {
2298 alock = u->as_AbstractLock();
2299 if (alock->box_node() == box_node) {
2300 // Verify that this Box is referenced only by related locks.
2301 assert(alock->obj_node()->eqv_uncast(obj), "");
2302 // Mark all related locks and unlocks.
2303 #ifdef ASSERT
2304 alock->log_lock_optimization(C, "eliminate_lock_set_nested");
2305 #endif
2306 alock->set_nested();
2307 }
2308 }
2309 }
2310 } else {
2311 #ifdef ASSERT
2312 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region");
2313 if (C->log() != nullptr)
2314 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output
2315 #endif
2316 }
2317 }
2318 return;
2319 }
2320 // Process locks for non escaping object
2321 assert(alock->is_non_esc_obj(), "");
2322 } // EliminateNestedLocks
2323
2324 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2325 // Look for all locks of this object and mark them and
2326 // corresponding BoxLock nodes as eliminated.
2327 Node* obj = alock->obj_node();
2328 for (uint j = 0; j < obj->outcnt(); j++) {
2329 Node* o = obj->raw_out(j);
2330 if (o->is_AbstractLock() &&
2331 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2332 alock = o->as_AbstractLock();
2333 Node* box = alock->box_node();
2334 // Replace old box node with new eliminated box for all users
2335 // of the same object and mark related locks as eliminated.
2336 mark_eliminated_box(box, obj);
2337 }
2338 }
2339 }
2340 }
2341
2342 // we have determined that this lock/unlock can be eliminated, we simply
2343 // eliminate the node without expanding it.
2344 //
2345 // Note: The membar's associated with the lock/unlock are currently not
2346 // eliminated. This should be investigated as a future enhancement.
2347 //
2348 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2349
2350 if (!alock->is_eliminated()) {
2351 return false;
2352 }
2353 #ifdef ASSERT
2354 if (!alock->is_coarsened()) {
2355 // Check that new "eliminated" BoxLock node is created.
2356 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2357 assert(oldbox->is_eliminated(), "should be done already");
2358 }
2359 #endif
2360
2361 alock->log_lock_optimization(C, "eliminate_lock");
2362
2363 #ifndef PRODUCT
2364 if (PrintEliminateLocks) {
2365 tty->print_cr("++++ Eliminated: %d %s '%s'", alock->_idx, (alock->is_Lock() ? "Lock" : "Unlock"), alock->kind_as_string());
2366 }
2367 #endif
2368
2369 Node* mem = alock->in(TypeFunc::Memory);
2370 Node* ctrl = alock->in(TypeFunc::Control);
2371 guarantee(ctrl != nullptr, "missing control projection, cannot replace_node() with null");
2372
2373 _callprojs = alock->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2374 // There are 2 projections from the lock. The lock node will
2375 // be deleted when its last use is subsumed below.
2376 assert(alock->outcnt() == 2 &&
2377 _callprojs->fallthrough_proj != nullptr &&
2378 _callprojs->fallthrough_memproj != nullptr,
2379 "Unexpected projections from Lock/Unlock");
2380
2381 Node* fallthroughproj = _callprojs->fallthrough_proj;
2382 Node* memproj_fallthrough = _callprojs->fallthrough_memproj;
2383
2384 // The memory projection from a lock/unlock is RawMem
2385 // The input to a Lock is merged memory, so extract its RawMem input
2386 // (unless the MergeMem has been optimized away.)
2387 if (alock->is_Lock()) {
2388 // Search for MemBarAcquireLock node and delete it also.
2389 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2390 assert(membar != nullptr && membar->Opcode() == Op_MemBarAcquireLock, "");
2391 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2392 Node* memproj = membar->proj_out(TypeFunc::Memory);
2393 _igvn.replace_node(ctrlproj, fallthroughproj);
2394 _igvn.replace_node(memproj, memproj_fallthrough);
2395
2396 // Delete FastLock node also if this Lock node is unique user
2397 // (a loop peeling may clone a Lock node).
2398 Node* flock = alock->as_Lock()->fastlock_node();
2399 if (flock->outcnt() == 1) {
2400 assert(flock->unique_out() == alock, "sanity");
2401 _igvn.replace_node(flock, top());
2402 }
2403 }
2404
2405 // Search for MemBarReleaseLock node and delete it also.
2406 if (alock->is_Unlock() && ctrl->is_Proj() && ctrl->in(0)->is_MemBar()) {
2407 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2408 assert(membar->Opcode() == Op_MemBarReleaseLock &&
2409 mem->is_Proj() && membar == mem->in(0), "");
2410 _igvn.replace_node(fallthroughproj, ctrl);
2411 _igvn.replace_node(memproj_fallthrough, mem);
2412 fallthroughproj = ctrl;
2413 memproj_fallthrough = mem;
2414 ctrl = membar->in(TypeFunc::Control);
2415 mem = membar->in(TypeFunc::Memory);
2416 }
2417
2418 _igvn.replace_node(fallthroughproj, ctrl);
2419 _igvn.replace_node(memproj_fallthrough, mem);
2420 return true;
2421 }
2422
2423
2424 //------------------------------expand_lock_node----------------------
2425 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
2426
2427 Node* ctrl = lock->in(TypeFunc::Control);
2428 Node* mem = lock->in(TypeFunc::Memory);
2429 Node* obj = lock->obj_node();
2430 Node* box = lock->box_node();
2431 Node* flock = lock->fastlock_node();
2432
2433 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2434
2435 // Make the merge point
2436 Node *region;
2437 Node *mem_phi;
2438 Node *slow_path;
2439
2440 region = new RegionNode(3);
2441 // create a Phi for the memory state
2442 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2443
2444 // Optimize test; set region slot 2
2445 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2446 mem_phi->init_req(2, mem);
2447
2448 // Make slow path call
2449 CallNode* call = make_slow_call(lock, OptoRuntime::complete_monitor_enter_Type(),
2450 OptoRuntime::complete_monitor_locking_Java(), nullptr, slow_path,
2451 obj, box, nullptr);
2452
2453 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2454
2455 // Slow path can only throw asynchronous exceptions, which are always
2456 // de-opted. So the compiler thinks the slow-call can never throw an
2457 // exception. If it DOES throw an exception we would need the debug
2458 // info removed first (since if it throws there is no monitor).
2459 assert(_callprojs->fallthrough_ioproj == nullptr && _callprojs->catchall_ioproj == nullptr &&
2460 _callprojs->catchall_memproj == nullptr && _callprojs->catchall_catchproj == nullptr, "Unexpected projection from Lock");
2461
2462 // Capture slow path
2463 // disconnect fall-through projection from call and create a new one
2464 // hook up users of fall-through projection to region
2465 Node *slow_ctrl = _callprojs->fallthrough_proj->clone();
2466 transform_later(slow_ctrl);
2467 _igvn.hash_delete(_callprojs->fallthrough_proj);
2468 _callprojs->fallthrough_proj->disconnect_inputs(C);
2469 region->init_req(1, slow_ctrl);
2470 // region inputs are now complete
2471 transform_later(region);
2472 _igvn.replace_node(_callprojs->fallthrough_proj, region);
2473
2474 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2475
2476 mem_phi->init_req(1, memproj);
2477
2478 transform_later(mem_phi);
2479
2480 _igvn.replace_node(_callprojs->fallthrough_memproj, mem_phi);
2481 }
2482
2483 //------------------------------expand_unlock_node----------------------
2484 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2485
2486 Node* ctrl = unlock->in(TypeFunc::Control);
2487 Node* mem = unlock->in(TypeFunc::Memory);
2488 Node* obj = unlock->obj_node();
2489 Node* box = unlock->box_node();
2490
2491 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2492
2493 // No need for a null check on unlock
2494
2495 // Make the merge point
2496 Node *region;
2497 Node *mem_phi;
2498
2499 region = new RegionNode(3);
2500 // create a Phi for the memory state
2501 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2502
2503 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2504 funlock = transform_later( funlock )->as_FastUnlock();
2505 // Optimize test; set region slot 2
2506 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2507 Node *thread = transform_later(new ThreadLocalNode());
2508
2509 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2510 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2511 "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2512
2513 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2514 assert(_callprojs->fallthrough_ioproj == nullptr && _callprojs->catchall_ioproj == nullptr &&
2515 _callprojs->catchall_memproj == nullptr && _callprojs->catchall_catchproj == nullptr, "Unexpected projection from Lock");
2516
2517 // No exceptions for unlocking
2518 // Capture slow path
2519 // disconnect fall-through projection from call and create a new one
2520 // hook up users of fall-through projection to region
2521 Node *slow_ctrl = _callprojs->fallthrough_proj->clone();
2522 transform_later(slow_ctrl);
2523 _igvn.hash_delete(_callprojs->fallthrough_proj);
2524 _callprojs->fallthrough_proj->disconnect_inputs(C);
2525 region->init_req(1, slow_ctrl);
2526 // region inputs are now complete
2527 transform_later(region);
2528 _igvn.replace_node(_callprojs->fallthrough_proj, region);
2529
2530 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2531 mem_phi->init_req(1, memproj );
2532 mem_phi->init_req(2, mem);
2533 transform_later(mem_phi);
2534
2535 _igvn.replace_node(_callprojs->fallthrough_memproj, mem_phi);
2536 }
2537
2538 // An inline type might be returned from the call but we don't know its
2539 // type. Either we get a buffered inline type (and nothing needs to be done)
2540 // or one of the values being returned is the klass of the inline type
2541 // and we need to allocate an inline type instance of that type and
2542 // initialize it with other values being returned. In that case, we
2543 // first try a fast path allocation and initialize the value with the
2544 // inline klass's pack handler or we fall back to a runtime call.
2545 void PhaseMacroExpand::expand_mh_intrinsic_return(CallStaticJavaNode* call) {
2546 assert(call->method()->is_method_handle_intrinsic(), "must be a method handle intrinsic call");
2547 Node* ret = call->proj_out_or_null(TypeFunc::Parms);
2548 if (ret == nullptr) {
2549 return;
2550 }
2551 const TypeFunc* tf = call->_tf;
2552 const TypeTuple* domain = OptoRuntime::store_inline_type_fields_Type()->domain_cc();
2553 const TypeFunc* new_tf = TypeFunc::make(tf->domain_sig(), tf->domain_cc(), tf->range_sig(), domain);
2554 call->_tf = new_tf;
2555 // Make sure the change of type is applied before projections are processed by igvn
2556 _igvn.set_type(call, call->Value(&_igvn));
2557 _igvn.set_type(ret, ret->Value(&_igvn));
2558
2559 // Before any new projection is added:
2560 CallProjections* projs = call->extract_projections(true, true);
2561
2562 // Create temporary hook nodes that will be replaced below.
2563 // Add an input to prevent hook nodes from being dead.
2564 Node* ctl = new Node(call);
2565 Node* mem = new Node(ctl);
2566 Node* io = new Node(ctl);
2567 Node* ex_ctl = new Node(ctl);
2568 Node* ex_mem = new Node(ctl);
2569 Node* ex_io = new Node(ctl);
2570 Node* res = new Node(ctl);
2571
2572 // Allocate a new buffered inline type only if a new one is not returned
2573 Node* cast = transform_later(new CastP2XNode(ctl, res));
2574 Node* mask = MakeConX(0x1);
2575 Node* masked = transform_later(new AndXNode(cast, mask));
2576 Node* cmp = transform_later(new CmpXNode(masked, mask));
2577 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
2578 IfNode* allocation_iff = new IfNode(ctl, bol, PROB_MAX, COUNT_UNKNOWN);
2579 transform_later(allocation_iff);
2580 Node* allocation_ctl = transform_later(new IfTrueNode(allocation_iff));
2581 Node* no_allocation_ctl = transform_later(new IfFalseNode(allocation_iff));
2582 Node* no_allocation_res = transform_later(new CheckCastPPNode(no_allocation_ctl, res, TypeInstPtr::BOTTOM));
2583
2584 // Try to allocate a new buffered inline instance either from TLAB or eden space
2585 Node* needgc_ctrl = nullptr; // needgc means slowcase, i.e. allocation failed
2586 CallLeafNoFPNode* handler_call;
2587 const bool alloc_in_place = UseTLAB;
2588 if (alloc_in_place) {
2589 Node* fast_oop_ctrl = nullptr;
2590 Node* fast_oop_rawmem = nullptr;
2591 Node* mask2 = MakeConX(-2);
2592 Node* masked2 = transform_later(new AndXNode(cast, mask2));
2593 Node* rawklassptr = transform_later(new CastX2PNode(masked2));
2594 Node* klass_node = transform_later(new CheckCastPPNode(allocation_ctl, rawklassptr, TypeInstKlassPtr::OBJECT_OR_NULL));
2595 Node* layout_val = make_load(nullptr, mem, klass_node, in_bytes(Klass::layout_helper_offset()), TypeInt::INT, T_INT);
2596 Node* size_in_bytes = ConvI2X(layout_val);
2597 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2598 Node* fast_oop = bs->obj_allocate(this, mem, allocation_ctl, size_in_bytes, io, needgc_ctrl,
2599 fast_oop_ctrl, fast_oop_rawmem,
2600 AllocateInstancePrefetchLines);
2601 // Allocation succeed, initialize buffered inline instance header firstly,
2602 // and then initialize its fields with an inline class specific handler
2603 Node* mark_node = makecon(TypeRawPtr::make((address)markWord::inline_type_prototype().value()));
2604 fast_oop_rawmem = make_store(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
2605 fast_oop_rawmem = make_store(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
2606 if (UseCompressedClassPointers) {
2607 fast_oop_rawmem = make_store(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::klass_gap_offset_in_bytes(), intcon(0), T_INT);
2608 }
2609 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);
2610 Node* pack_handler = make_load(fast_oop_ctrl, fast_oop_rawmem, fixed_block, in_bytes(InlineKlass::pack_handler_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
2611 handler_call = new CallLeafNoFPNode(OptoRuntime::pack_inline_type_Type(),
2612 nullptr,
2613 "pack handler",
2614 TypeRawPtr::BOTTOM);
2615 handler_call->init_req(TypeFunc::Control, fast_oop_ctrl);
2616 handler_call->init_req(TypeFunc::Memory, fast_oop_rawmem);
2617 handler_call->init_req(TypeFunc::I_O, top());
2618 handler_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2619 handler_call->init_req(TypeFunc::ReturnAdr, top());
2620 handler_call->init_req(TypeFunc::Parms, pack_handler);
2621 handler_call->init_req(TypeFunc::Parms+1, fast_oop);
2622 } else {
2623 needgc_ctrl = allocation_ctl;
2624 }
2625
2626 // Allocation failed, fall back to a runtime call
2627 CallStaticJavaNode* slow_call = new CallStaticJavaNode(OptoRuntime::store_inline_type_fields_Type(),
2628 StubRoutines::store_inline_type_fields_to_buf(),
2629 "store_inline_type_fields",
2630 TypePtr::BOTTOM);
2631 slow_call->init_req(TypeFunc::Control, needgc_ctrl);
2632 slow_call->init_req(TypeFunc::Memory, mem);
2633 slow_call->init_req(TypeFunc::I_O, io);
2634 slow_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2635 slow_call->init_req(TypeFunc::ReturnAdr, call->in(TypeFunc::ReturnAdr));
2636 slow_call->init_req(TypeFunc::Parms, res);
2637
2638 Node* slow_ctl = transform_later(new ProjNode(slow_call, TypeFunc::Control));
2639 Node* slow_mem = transform_later(new ProjNode(slow_call, TypeFunc::Memory));
2640 Node* slow_io = transform_later(new ProjNode(slow_call, TypeFunc::I_O));
2641 Node* slow_res = transform_later(new ProjNode(slow_call, TypeFunc::Parms));
2642 Node* slow_catc = transform_later(new CatchNode(slow_ctl, slow_io, 2));
2643 Node* slow_norm = transform_later(new CatchProjNode(slow_catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci));
2644 Node* slow_excp = transform_later(new CatchProjNode(slow_catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci));
2645
2646 Node* ex_r = new RegionNode(3);
2647 Node* ex_mem_phi = new PhiNode(ex_r, Type::MEMORY, TypePtr::BOTTOM);
2648 Node* ex_io_phi = new PhiNode(ex_r, Type::ABIO);
2649 ex_r->init_req(1, slow_excp);
2650 ex_mem_phi->init_req(1, slow_mem);
2651 ex_io_phi->init_req(1, slow_io);
2652 ex_r->init_req(2, ex_ctl);
2653 ex_mem_phi->init_req(2, ex_mem);
2654 ex_io_phi->init_req(2, ex_io);
2655 transform_later(ex_r);
2656 transform_later(ex_mem_phi);
2657 transform_later(ex_io_phi);
2658
2659 // We don't know how many values are returned. This assumes the
2660 // worst case, that all available registers are used.
2661 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) {
2662 if (domain->field_at(i) == Type::HALF) {
2663 slow_call->init_req(i, top());
2664 if (alloc_in_place) {
2665 handler_call->init_req(i+1, top());
2666 }
2667 continue;
2668 }
2669 Node* proj = transform_later(new ProjNode(call, i));
2670 slow_call->init_req(i, proj);
2671 if (alloc_in_place) {
2672 handler_call->init_req(i+1, proj);
2673 }
2674 }
2675 // We can safepoint at that new call
2676 slow_call->copy_call_debug_info(&_igvn, call);
2677 transform_later(slow_call);
2678 if (alloc_in_place) {
2679 transform_later(handler_call);
2680 }
2681
2682 Node* fast_ctl = nullptr;
2683 Node* fast_res = nullptr;
2684 MergeMemNode* fast_mem = nullptr;
2685 if (alloc_in_place) {
2686 fast_ctl = transform_later(new ProjNode(handler_call, TypeFunc::Control));
2687 Node* rawmem = transform_later(new ProjNode(handler_call, TypeFunc::Memory));
2688 fast_res = transform_later(new ProjNode(handler_call, TypeFunc::Parms));
2689 fast_mem = MergeMemNode::make(mem);
2690 fast_mem->set_memory_at(Compile::AliasIdxRaw, rawmem);
2691 transform_later(fast_mem);
2692 }
2693
2694 Node* r = new RegionNode(alloc_in_place ? 4 : 3);
2695 Node* mem_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM);
2696 Node* io_phi = new PhiNode(r, Type::ABIO);
2697 Node* res_phi = new PhiNode(r, TypeInstPtr::BOTTOM);
2698 r->init_req(1, no_allocation_ctl);
2699 mem_phi->init_req(1, mem);
2700 io_phi->init_req(1, io);
2701 res_phi->init_req(1, no_allocation_res);
2702 r->init_req(2, slow_norm);
2703 mem_phi->init_req(2, slow_mem);
2704 io_phi->init_req(2, slow_io);
2705 res_phi->init_req(2, slow_res);
2706 if (alloc_in_place) {
2707 r->init_req(3, fast_ctl);
2708 mem_phi->init_req(3, fast_mem);
2709 io_phi->init_req(3, io);
2710 res_phi->init_req(3, fast_res);
2711 }
2712 transform_later(r);
2713 transform_later(mem_phi);
2714 transform_later(io_phi);
2715 transform_later(res_phi);
2716
2717 // Do not let stores that initialize this buffer be reordered with a subsequent
2718 // store that would make this buffer accessible by other threads.
2719 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
2720 transform_later(mb);
2721 mb->init_req(TypeFunc::Memory, mem_phi);
2722 mb->init_req(TypeFunc::Control, r);
2723 r = new ProjNode(mb, TypeFunc::Control);
2724 transform_later(r);
2725 mem_phi = new ProjNode(mb, TypeFunc::Memory);
2726 transform_later(mem_phi);
2727
2728 assert(projs->nb_resproj == 1, "unexpected number of results");
2729 _igvn.replace_in_uses(projs->fallthrough_catchproj, r);
2730 _igvn.replace_in_uses(projs->fallthrough_memproj, mem_phi);
2731 _igvn.replace_in_uses(projs->fallthrough_ioproj, io_phi);
2732 _igvn.replace_in_uses(projs->resproj[0], res_phi);
2733 _igvn.replace_in_uses(projs->catchall_catchproj, ex_r);
2734 _igvn.replace_in_uses(projs->catchall_memproj, ex_mem_phi);
2735 _igvn.replace_in_uses(projs->catchall_ioproj, ex_io_phi);
2736 // The CatchNode should not use the ex_io_phi. Re-connect it to the catchall_ioproj.
2737 Node* cn = projs->fallthrough_catchproj->in(0);
2738 _igvn.replace_input_of(cn, 1, projs->catchall_ioproj);
2739
2740 _igvn.replace_node(ctl, projs->fallthrough_catchproj);
2741 _igvn.replace_node(mem, projs->fallthrough_memproj);
2742 _igvn.replace_node(io, projs->fallthrough_ioproj);
2743 _igvn.replace_node(res, projs->resproj[0]);
2744 _igvn.replace_node(ex_ctl, projs->catchall_catchproj);
2745 _igvn.replace_node(ex_mem, projs->catchall_memproj);
2746 _igvn.replace_node(ex_io, projs->catchall_ioproj);
2747 }
2748
2749 void PhaseMacroExpand::expand_subtypecheck_node(SubTypeCheckNode *check) {
2750 assert(check->in(SubTypeCheckNode::Control) == nullptr, "should be pinned");
2751 Node* bol = check->unique_out();
2752 Node* obj_or_subklass = check->in(SubTypeCheckNode::ObjOrSubKlass);
2753 Node* superklass = check->in(SubTypeCheckNode::SuperKlass);
2754 assert(bol->is_Bool() && bol->as_Bool()->_test._test == BoolTest::ne, "unexpected bool node");
2755
2756 for (DUIterator_Last imin, i = bol->last_outs(imin); i >= imin; --i) {
2757 Node* iff = bol->last_out(i);
2758 assert(iff->is_If(), "where's the if?");
2759
2760 if (iff->in(0)->is_top()) {
2761 _igvn.replace_input_of(iff, 1, C->top());
2762 continue;
2763 }
2764
2765 Node* iftrue = iff->as_If()->proj_out(1);
2766 Node* iffalse = iff->as_If()->proj_out(0);
2767 Node* ctrl = iff->in(0);
2768
2769 Node* subklass = nullptr;
2770 if (_igvn.type(obj_or_subklass)->isa_klassptr()) {
2771 subklass = obj_or_subklass;
2772 } else {
2773 Node* k_adr = basic_plus_adr(obj_or_subklass, oopDesc::klass_offset_in_bytes());
2774 subklass = _igvn.transform(LoadKlassNode::make(_igvn, C->immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
2775 }
2776
2777 Node* not_subtype_ctrl = Phase::gen_subtype_check(subklass, superklass, &ctrl, nullptr, _igvn, check->method(), check->bci());
2778
2779 _igvn.replace_input_of(iff, 0, C->top());
2780 _igvn.replace_node(iftrue, not_subtype_ctrl);
2781 _igvn.replace_node(iffalse, ctrl);
2782 }
2783 _igvn.replace_node(check, C->top());
2784 }
2785
2786 // FlatArrayCheckNode (array1 array2 ...) is expanded into:
2787 //
2788 // long mark = array1.mark | array2.mark | ...;
2789 // long locked_bit = markWord::unlocked_value & array1.mark & array2.mark & ...;
2790 // if (locked_bit == 0) {
2791 // // One array is locked, load prototype header from the klass
2792 // mark = array1.klass.proto | array2.klass.proto | ...
2793 // }
2794 // if ((mark & markWord::flat_array_bit_in_place) == 0) {
2795 // ...
2796 // }
2797 void PhaseMacroExpand::expand_flatarraycheck_node(FlatArrayCheckNode* check) {
2798 bool array_inputs = _igvn.type(check->in(FlatArrayCheckNode::ArrayOrKlass))->isa_oopptr() != nullptr;
2799 if (array_inputs) {
2800 Node* mark = MakeConX(0);
2801 Node* locked_bit = MakeConX(markWord::unlocked_value);
2802 Node* mem = check->in(FlatArrayCheckNode::Memory);
2803 for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
2804 Node* ary = check->in(i);
2805 const TypeOopPtr* t = _igvn.type(ary)->isa_oopptr();
2806 assert(t != nullptr, "Mixing array and klass inputs");
2807 assert(!t->is_flat() && !t->is_not_flat(), "Should have been optimized out");
2808 Node* mark_adr = basic_plus_adr(ary, oopDesc::mark_offset_in_bytes());
2809 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));
2810 mark = _igvn.transform(new OrXNode(mark, mark_load));
2811 locked_bit = _igvn.transform(new AndXNode(locked_bit, mark_load));
2812 }
2813 assert(!mark->is_Con(), "Should have been optimized out");
2814 Node* cmp = _igvn.transform(new CmpXNode(locked_bit, MakeConX(0)));
2815 Node* is_unlocked = _igvn.transform(new BoolNode(cmp, BoolTest::ne));
2816
2817 // BoolNode might be shared, replace each if user
2818 Node* old_bol = check->unique_out();
2819 assert(old_bol->is_Bool() && old_bol->as_Bool()->_test._test == BoolTest::ne, "unexpected condition");
2820 for (DUIterator_Last imin, i = old_bol->last_outs(imin); i >= imin; --i) {
2821 IfNode* old_iff = old_bol->last_out(i)->as_If();
2822 Node* ctrl = old_iff->in(0);
2823 RegionNode* region = new RegionNode(3);
2824 Node* mark_phi = new PhiNode(region, TypeX_X);
2825
2826 // Check if array is unlocked
2827 IfNode* iff = _igvn.transform(new IfNode(ctrl, is_unlocked, PROB_MAX, COUNT_UNKNOWN))->as_If();
2828
2829 // Unlocked: Use bits from mark word
2830 region->init_req(1, _igvn.transform(new IfTrueNode(iff)));
2831 mark_phi->init_req(1, mark);
2832
2833 // Locked: Load prototype header from klass
2834 ctrl = _igvn.transform(new IfFalseNode(iff));
2835 Node* proto = MakeConX(0);
2836 for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
2837 Node* ary = check->in(i);
2838 // Make loads control dependent to make sure they are only executed if array is locked
2839 Node* klass_adr = basic_plus_adr(ary, oopDesc::klass_offset_in_bytes());
2840 Node* klass = _igvn.transform(LoadKlassNode::make(_igvn, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
2841 Node* proto_adr = basic_plus_adr(klass, in_bytes(Klass::prototype_header_offset()));
2842 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));
2843 proto = _igvn.transform(new OrXNode(proto, proto_load));
2844 }
2845 region->init_req(2, ctrl);
2846 mark_phi->init_req(2, proto);
2847
2848 // Check if flat array bits are set
2849 Node* mask = MakeConX(markWord::flat_array_bit_in_place);
2850 Node* masked = _igvn.transform(new AndXNode(_igvn.transform(mark_phi), mask));
2851 cmp = _igvn.transform(new CmpXNode(masked, MakeConX(0)));
2852 Node* is_not_flat = _igvn.transform(new BoolNode(cmp, BoolTest::eq));
2853
2854 ctrl = _igvn.transform(region);
2855 iff = _igvn.transform(new IfNode(ctrl, is_not_flat, PROB_MAX, COUNT_UNKNOWN))->as_If();
2856 _igvn.replace_node(old_iff, iff);
2857 }
2858 _igvn.replace_node(check, C->top());
2859 } else {
2860 // Fall back to layout helper check
2861 Node* lhs = intcon(0);
2862 for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
2863 Node* array_or_klass = check->in(i);
2864 Node* klass = nullptr;
2865 const TypePtr* t = _igvn.type(array_or_klass)->is_ptr();
2866 assert(!t->is_flat() && !t->is_not_flat(), "Should have been optimized out");
2867 if (t->isa_oopptr() != nullptr) {
2868 Node* klass_adr = basic_plus_adr(array_or_klass, oopDesc::klass_offset_in_bytes());
2869 klass = transform_later(LoadKlassNode::make(_igvn, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
2870 } else {
2871 assert(t->isa_klassptr(), "Unexpected input type");
2872 klass = array_or_klass;
2873 }
2874 Node* lh_addr = basic_plus_adr(klass, in_bytes(Klass::layout_helper_offset()));
2875 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));
2876 lhs = _igvn.transform(new OrINode(lhs, lh_val));
2877 }
2878 Node* masked = transform_later(new AndINode(lhs, intcon(Klass::_lh_array_tag_flat_value_bit_inplace)));
2879 Node* cmp = transform_later(new CmpINode(masked, intcon(0)));
2880 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
2881 Node* m2b = transform_later(new Conv2BNode(masked));
2882 // The matcher expects the input to If nodes to be produced by a Bool(CmpI..)
2883 // pattern, but the input to other potential users (e.g. Phi) to be some
2884 // other pattern (e.g. a Conv2B node, possibly idealized as a CMoveI).
2885 Node* old_bol = check->unique_out();
2886 for (DUIterator_Last imin, i = old_bol->last_outs(imin); i >= imin; --i) {
2887 Node* user = old_bol->last_out(i);
2888 for (uint j = 0; j < user->req(); j++) {
2889 Node* n = user->in(j);
2890 if (n == old_bol) {
2891 _igvn.replace_input_of(user, j, user->is_If() ? bol : m2b);
2892 }
2893 }
2894 }
2895 _igvn.replace_node(check, C->top());
2896 }
2897 }
2898
2899 //---------------------------eliminate_macro_nodes----------------------
2900 // Eliminate scalar replaced allocations and associated locks.
2901 void PhaseMacroExpand::eliminate_macro_nodes() {
2902 if (C->macro_count() == 0)
2903 return;
2904 NOT_PRODUCT(int membar_before = count_MemBar(C);)
2905
2906 // Before elimination may re-mark (change to Nested or NonEscObj)
2907 // all associated (same box and obj) lock and unlock nodes.
2908 int cnt = C->macro_count();
2909 for (int i=0; i < cnt; i++) {
2910 Node *n = C->macro_node(i);
2911 if (n->is_AbstractLock()) { // Lock and Unlock nodes
2912 mark_eliminated_locking_nodes(n->as_AbstractLock());
2913 }
2914 }
2915 // Re-marking may break consistency of Coarsened locks.
2916 if (!C->coarsened_locks_consistent()) {
2917 return; // recompile without Coarsened locks if broken
2918 } else {
2919 // After coarsened locks are eliminated locking regions
2920 // become unbalanced. We should not execute any more
2921 // locks elimination optimizations on them.
2922 C->mark_unbalanced_boxes();
2923 }
2924
2925 // First, attempt to eliminate locks
2926 bool progress = true;
2927 while (progress) {
2928 progress = false;
2929 for (int i = C->macro_count(); i > 0; i = MIN2(i - 1, C->macro_count())) { // more than 1 element can be eliminated at once
2930 Node* n = C->macro_node(i - 1);
2931 bool success = false;
2932 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2933 if (n->is_AbstractLock()) {
2934 success = eliminate_locking_node(n->as_AbstractLock());
2935 #ifndef PRODUCT
2936 if (success && PrintOptoStatistics) {
2937 Atomic::inc(&PhaseMacroExpand::_monitor_objects_removed_counter);
2938 }
2939 #endif
2940 }
2941 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2942 progress = progress || success;
2943 }
2944 }
2945 // Next, attempt to eliminate allocations
2946 _has_locks = false;
2947 progress = true;
2948 while (progress) {
2949 progress = false;
2950 for (int i = C->macro_count(); i > 0; i = MIN2(i - 1, C->macro_count())) { // more than 1 element can be eliminated at once
2951 Node* n = C->macro_node(i - 1);
2952 bool success = false;
2953 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2954 switch (n->class_id()) {
2955 case Node::Class_Allocate:
2956 case Node::Class_AllocateArray:
2957 success = eliminate_allocate_node(n->as_Allocate());
2958 #ifndef PRODUCT
2959 if (success && PrintOptoStatistics) {
2960 Atomic::inc(&PhaseMacroExpand::_objs_scalar_replaced_counter);
2961 }
2962 #endif
2963 break;
2964 case Node::Class_CallStaticJava: {
2965 CallStaticJavaNode* call = n->as_CallStaticJava();
2966 if (!call->method()->is_method_handle_intrinsic()) {
2967 success = eliminate_boxing_node(n->as_CallStaticJava());
2968 }
2969 break;
2970 }
2971 case Node::Class_Lock:
2972 case Node::Class_Unlock:
2973 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2974 _has_locks = true;
2975 break;
2976 case Node::Class_ArrayCopy:
2977 break;
2978 case Node::Class_OuterStripMinedLoop:
2979 break;
2980 case Node::Class_SubTypeCheck:
2981 break;
2982 case Node::Class_Opaque1:
2983 break;
2984 case Node::Class_FlatArrayCheck:
2985 break;
2986 default:
2987 assert(n->Opcode() == Op_LoopLimit ||
2988 n->Opcode() == Op_ModD ||
2989 n->Opcode() == Op_ModF ||
2990 n->is_OpaqueNotNull() ||
2991 n->is_OpaqueInitializedAssertionPredicate() ||
2992 n->Opcode() == Op_MaxL ||
2993 n->Opcode() == Op_MinL ||
2994 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n),
2995 "unknown node type in macro list");
2996 }
2997 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2998 progress = progress || success;
2999 }
3000 }
3001 #ifndef PRODUCT
3002 if (PrintOptoStatistics) {
3003 int membar_after = count_MemBar(C);
3004 Atomic::add(&PhaseMacroExpand::_memory_barriers_removed_counter, membar_before - membar_after);
3005 }
3006 #endif
3007 }
3008
3009 //------------------------------expand_macro_nodes----------------------
3010 // Returns true if a failure occurred.
3011 bool PhaseMacroExpand::expand_macro_nodes() {
3012 // Do not allow new macro nodes once we started to expand
3013 C->reset_allow_macro_nodes();
3014 if (StressMacroExpansion) {
3015 C->shuffle_macro_nodes();
3016 }
3017 // Last attempt to eliminate macro nodes.
3018 eliminate_macro_nodes();
3019 if (C->failing()) return true;
3020
3021 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
3022 bool progress = true;
3023 while (progress) {
3024 progress = false;
3025 for (int i = C->macro_count(); i > 0; i--) {
3026 Node* n = C->macro_node(i-1);
3027 bool success = false;
3028 DEBUG_ONLY(int old_macro_count = C->macro_count();)
3029 if (n->Opcode() == Op_LoopLimit) {
3030 // Remove it from macro list and put on IGVN worklist to optimize.
3031 C->remove_macro_node(n);
3032 _igvn._worklist.push(n);
3033 success = true;
3034 } else if (n->Opcode() == Op_CallStaticJava) {
3035 CallStaticJavaNode* call = n->as_CallStaticJava();
3036 if (!call->method()->is_method_handle_intrinsic()) {
3037 // Remove it from macro list and put on IGVN worklist to optimize.
3038 C->remove_macro_node(n);
3039 _igvn._worklist.push(n);
3040 success = true;
3041 }
3042 } else if (n->is_Opaque1()) {
3043 _igvn.replace_node(n, n->in(1));
3044 success = true;
3045 } else if (n->is_OpaqueNotNull()) {
3046 // Tests with OpaqueNotNull nodes are implicitly known to be true. Replace the node with true. In debug builds,
3047 // we leave the test in the graph to have an additional sanity check at runtime. If the test fails (i.e. a bug),
3048 // we will execute a Halt node.
3049 #ifdef ASSERT
3050 _igvn.replace_node(n, n->in(1));
3051 #else
3052 _igvn.replace_node(n, _igvn.intcon(1));
3053 #endif
3054 success = true;
3055 } else if (n->is_OpaqueInitializedAssertionPredicate()) {
3056 // Initialized Assertion Predicates must always evaluate to true. Therefore, we get rid of them in product
3057 // builds as they are useless. In debug builds we keep them as additional verification code. Even though
3058 // loop opts are already over, we want to keep Initialized Assertion Predicates alive as long as possible to
3059 // enable folding of dead control paths within which cast nodes become top after due to impossible types -
3060 // even after loop opts are over. Therefore, we delay the removal of these opaque nodes until now.
3061 #ifdef ASSERT
3062 _igvn.replace_node(n, n->in(1));
3063 #else
3064 _igvn.replace_node(n, _igvn.intcon(1));
3065 #endif // ASSERT
3066 } else if (n->Opcode() == Op_OuterStripMinedLoop) {
3067 n->as_OuterStripMinedLoop()->adjust_strip_mined_loop(&_igvn);
3068 C->remove_macro_node(n);
3069 success = true;
3070 } else if (n->Opcode() == Op_MaxL) {
3071 // Since MaxL and MinL are not implemented in the backend, we expand them to
3072 // a CMoveL construct now. At least until here, the type could be computed
3073 // precisely. CMoveL is not so smart, but we can give it at least the best
3074 // type we know abouot n now.
3075 Node* repl = MaxNode::signed_max(n->in(1), n->in(2), _igvn.type(n), _igvn);
3076 _igvn.replace_node(n, repl);
3077 success = true;
3078 } else if (n->Opcode() == Op_MinL) {
3079 Node* repl = MaxNode::signed_min(n->in(1), n->in(2), _igvn.type(n), _igvn);
3080 _igvn.replace_node(n, repl);
3081 success = true;
3082 }
3083 assert(!success || (C->macro_count() == (old_macro_count - 1)), "elimination must have deleted one node from macro list");
3084 progress = progress || success;
3085 if (success) {
3086 C->print_method(PHASE_AFTER_MACRO_EXPANSION_STEP, 5, n);
3087 }
3088 }
3089 }
3090
3091 // Clean up the graph so we're less likely to hit the maximum node
3092 // limit
3093 _igvn.set_delay_transform(false);
3094 _igvn.optimize();
3095 if (C->failing()) return true;
3096 _igvn.set_delay_transform(true);
3097
3098
3099 // Because we run IGVN after each expansion, some macro nodes may go
3100 // dead and be removed from the list as we iterate over it. Move
3101 // Allocate nodes (processed in a second pass) at the beginning of
3102 // the list and then iterate from the last element of the list until
3103 // an Allocate node is seen. This is robust to random deletion in
3104 // the list due to nodes going dead.
3105 C->sort_macro_nodes();
3106
3107 // expand arraycopy "macro" nodes first
3108 // For ReduceBulkZeroing, we must first process all arraycopy nodes
3109 // before the allocate nodes are expanded.
3110 while (C->macro_count() > 0) {
3111 int macro_count = C->macro_count();
3112 Node * n = C->macro_node(macro_count-1);
3113 assert(n->is_macro(), "only macro nodes expected here");
3114 if (_igvn.type(n) == Type::TOP || (n->in(0) != nullptr && n->in(0)->is_top())) {
3115 // node is unreachable, so don't try to expand it
3116 C->remove_macro_node(n);
3117 continue;
3118 }
3119 if (n->is_Allocate()) {
3120 break;
3121 }
3122 // Make sure expansion will not cause node limit to be exceeded.
3123 // Worst case is a macro node gets expanded into about 200 nodes.
3124 // Allow 50% more for optimization.
3125 if (C->check_node_count(300, "out of nodes before macro expansion")) {
3126 return true;
3127 }
3128
3129 DEBUG_ONLY(int old_macro_count = C->macro_count();)
3130 switch (n->class_id()) {
3131 case Node::Class_Lock:
3132 expand_lock_node(n->as_Lock());
3133 break;
3134 case Node::Class_Unlock:
3135 expand_unlock_node(n->as_Unlock());
3136 break;
3137 case Node::Class_ArrayCopy:
3138 expand_arraycopy_node(n->as_ArrayCopy());
3139 break;
3140 case Node::Class_SubTypeCheck:
3141 expand_subtypecheck_node(n->as_SubTypeCheck());
3142 break;
3143 case Node::Class_CallStaticJava:
3144 expand_mh_intrinsic_return(n->as_CallStaticJava());
3145 C->remove_macro_node(n);
3146 break;
3147 case Node::Class_FlatArrayCheck:
3148 expand_flatarraycheck_node(n->as_FlatArrayCheck());
3149 break;
3150 default:
3151 switch (n->Opcode()) {
3152 case Op_ModD:
3153 case Op_ModF: {
3154 bool is_drem = n->Opcode() == Op_ModD;
3155 CallNode* mod_macro = n->as_Call();
3156 CallNode* call = new CallLeafNode(mod_macro->tf(),
3157 is_drem ? CAST_FROM_FN_PTR(address, SharedRuntime::drem)
3158 : CAST_FROM_FN_PTR(address, SharedRuntime::frem),
3159 is_drem ? "drem" : "frem", TypeRawPtr::BOTTOM);
3160 call->init_req(TypeFunc::Control, mod_macro->in(TypeFunc::Control));
3161 call->init_req(TypeFunc::I_O, mod_macro->in(TypeFunc::I_O));
3162 call->init_req(TypeFunc::Memory, mod_macro->in(TypeFunc::Memory));
3163 call->init_req(TypeFunc::ReturnAdr, mod_macro->in(TypeFunc::ReturnAdr));
3164 call->init_req(TypeFunc::FramePtr, mod_macro->in(TypeFunc::FramePtr));
3165 for (unsigned int i = 0; i < mod_macro->tf()->domain_cc()->cnt() - TypeFunc::Parms; i++) {
3166 call->init_req(TypeFunc::Parms + i, mod_macro->in(TypeFunc::Parms + i));
3167 }
3168 _igvn.replace_node(mod_macro, call);
3169 transform_later(call);
3170 break;
3171 }
3172 default:
3173 assert(false, "unknown node type in macro list");
3174 }
3175 }
3176 assert(C->macro_count() == (old_macro_count - 1), "expansion must have deleted one node from macro list");
3177 if (C->failing()) return true;
3178 C->print_method(PHASE_AFTER_MACRO_EXPANSION_STEP, 5, n);
3179
3180 // Clean up the graph so we're less likely to hit the maximum node
3181 // limit
3182 _igvn.set_delay_transform(false);
3183 _igvn.optimize();
3184 if (C->failing()) return true;
3185 _igvn.set_delay_transform(true);
3186 }
3187
3188 // All nodes except Allocate nodes are expanded now. There could be
3189 // new optimization opportunities (such as folding newly created
3190 // load from a just allocated object). Run IGVN.
3191
3192 // expand "macro" nodes
3193 // nodes are removed from the macro list as they are processed
3194 while (C->macro_count() > 0) {
3195 int macro_count = C->macro_count();
3196 Node * n = C->macro_node(macro_count-1);
3197 assert(n->is_macro(), "only macro nodes expected here");
3198 if (_igvn.type(n) == Type::TOP || (n->in(0) != nullptr && n->in(0)->is_top())) {
3199 // node is unreachable, so don't try to expand it
3200 C->remove_macro_node(n);
3201 continue;
3202 }
3203 // Make sure expansion will not cause node limit to be exceeded.
3204 // Worst case is a macro node gets expanded into about 200 nodes.
3205 // Allow 50% more for optimization.
3206 if (C->check_node_count(300, "out of nodes before macro expansion")) {
3207 return true;
3208 }
3209 switch (n->class_id()) {
3210 case Node::Class_Allocate:
3211 expand_allocate(n->as_Allocate());
3212 break;
3213 case Node::Class_AllocateArray:
3214 expand_allocate_array(n->as_AllocateArray());
3215 break;
3216 default:
3217 assert(false, "unknown node type in macro list");
3218 }
3219 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
3220 if (C->failing()) return true;
3221 C->print_method(PHASE_AFTER_MACRO_EXPANSION_STEP, 5, n);
3222
3223 // Clean up the graph so we're less likely to hit the maximum node
3224 // limit
3225 _igvn.set_delay_transform(false);
3226 _igvn.optimize();
3227 if (C->failing()) return true;
3228 _igvn.set_delay_transform(true);
3229 }
3230
3231 _igvn.set_delay_transform(false);
3232 return false;
3233 }
3234
3235 #ifndef PRODUCT
3236 int PhaseMacroExpand::_objs_scalar_replaced_counter = 0;
3237 int PhaseMacroExpand::_monitor_objects_removed_counter = 0;
3238 int PhaseMacroExpand::_GC_barriers_removed_counter = 0;
3239 int PhaseMacroExpand::_memory_barriers_removed_counter = 0;
3240
3241 void PhaseMacroExpand::print_statistics() {
3242 tty->print("Objects scalar replaced = %d, ", Atomic::load(&_objs_scalar_replaced_counter));
3243 tty->print("Monitor objects removed = %d, ", Atomic::load(&_monitor_objects_removed_counter));
3244 tty->print("GC barriers removed = %d, ", Atomic::load(&_GC_barriers_removed_counter));
3245 tty->print_cr("Memory barriers removed = %d", Atomic::load(&_memory_barriers_removed_counter));
3246 }
3247
3248 int PhaseMacroExpand::count_MemBar(Compile *C) {
3249 if (!PrintOptoStatistics) {
3250 return 0;
3251 }
3252 Unique_Node_List ideal_nodes;
3253 int total = 0;
3254 ideal_nodes.map(C->live_nodes(), nullptr);
3255 ideal_nodes.push(C->root());
3256 for (uint next = 0; next < ideal_nodes.size(); ++next) {
3257 Node* n = ideal_nodes.at(next);
3258 if (n->is_MemBar()) {
3259 total++;
3260 }
3261 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3262 Node* m = n->fast_out(i);
3263 ideal_nodes.push(m);
3264 }
3265 }
3266 return total;
3267 }
3268 #endif