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