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