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 "ci/ciInstanceKlass.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "gc/shared/collectedHeap.inline.hpp"
29 #include "gc/shared/tlab_globals.hpp"
30 #include "libadt/vectset.hpp"
31 #include "memory/universe.hpp"
32 #include "opto/addnode.hpp"
33 #include "opto/arraycopynode.hpp"
34 #include "opto/callnode.hpp"
35 #include "opto/castnode.hpp"
36 #include "opto/cfgnode.hpp"
37 #include "opto/compile.hpp"
38 #include "opto/convertnode.hpp"
39 #include "opto/graphKit.hpp"
40 #include "opto/inlinetypenode.hpp"
41 #include "opto/intrinsicnode.hpp"
42 #include "opto/locknode.hpp"
43 #include "opto/loopnode.hpp"
44 #include "opto/macro.hpp"
45 #include "opto/memnode.hpp"
46 #include "opto/narrowptrnode.hpp"
47 #include "opto/node.hpp"
48 #include "opto/opaquenode.hpp"
49 #include "opto/phaseX.hpp"
50 #include "opto/rootnode.hpp"
51 #include "opto/runtime.hpp"
52 #include "opto/subnode.hpp"
53 #include "opto/subtypenode.hpp"
54 #include "opto/type.hpp"
55 #include "prims/jvmtiExport.hpp"
56 #include "runtime/continuation.hpp"
57 #include "runtime/sharedRuntime.hpp"
58 #include "runtime/stubRoutines.hpp"
59 #include "utilities/globalDefinitions.hpp"
60 #include "utilities/macros.hpp"
61 #include "utilities/powerOfTwo.hpp"
62 #if INCLUDE_G1GC
63 #include "gc/g1/g1ThreadLocalData.hpp"
64 #endif // INCLUDE_G1GC
65
66
67 //
68 // Replace any references to "oldref" in inputs to "use" with "newref".
69 // Returns the number of replacements made.
70 //
71 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
72 int nreplacements = 0;
73 uint req = use->req();
74 for (uint j = 0; j < use->len(); j++) {
75 Node *uin = use->in(j);
76 if (uin == oldref) {
77 if (j < req)
78 use->set_req(j, newref);
79 else
80 use->set_prec(j, newref);
81 nreplacements++;
82 } else if (j >= req && uin == nullptr) {
83 break;
84 }
85 }
86 return nreplacements;
87 }
88
89 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
90 Node* cmp;
91 if (mask != 0) {
92 Node* and_node = transform_later(new AndXNode(word, MakeConX(mask)));
93 cmp = transform_later(new CmpXNode(and_node, MakeConX(bits)));
94 } else {
95 cmp = word;
96 }
97 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne));
98 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
99 transform_later(iff);
100
101 // Fast path taken.
102 Node *fast_taken = transform_later(new IfFalseNode(iff));
103
104 // Fast path not-taken, i.e. slow path
105 Node *slow_taken = transform_later(new IfTrueNode(iff));
106
107 if (return_fast_path) {
108 region->init_req(edge, slow_taken); // Capture slow-control
109 return fast_taken;
110 } else {
111 region->init_req(edge, fast_taken); // Capture fast-control
112 return slow_taken;
113 }
114 }
115
116 //--------------------copy_predefined_input_for_runtime_call--------------------
117 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
118 // Set fixed predefined input arguments
119 call->init_req( TypeFunc::Control, ctrl );
120 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) );
121 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
122 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
123 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
124 }
125
126 //------------------------------make_slow_call---------------------------------
127 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type,
128 address slow_call, const char* leaf_name, Node* slow_path,
129 Node* parm0, Node* parm1, Node* parm2) {
130
131 // Slow-path call
132 CallNode *call = leaf_name
133 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
134 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), TypeRawPtr::BOTTOM );
135
136 // Slow path call has no side-effects, uses few values
137 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
138 if (parm0 != nullptr) call->init_req(TypeFunc::Parms+0, parm0);
139 if (parm1 != nullptr) call->init_req(TypeFunc::Parms+1, parm1);
140 if (parm2 != nullptr) call->init_req(TypeFunc::Parms+2, parm2);
141 call->copy_call_debug_info(&_igvn, oldcall);
142 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
143 _igvn.replace_node(oldcall, call);
144 transform_later(call);
145
146 return call;
147 }
148
149 void PhaseMacroExpand::eliminate_gc_barrier(Node* p2x) {
150 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
151 bs->eliminate_gc_barrier(&_igvn, p2x);
152 #ifndef PRODUCT
153 if (PrintOptoStatistics) {
154 Atomic::inc(&PhaseMacroExpand::_GC_barriers_removed_counter);
155 }
156 #endif
157 }
158
159 // Search for a memory operation for the specified memory slice.
160 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
161 Node *orig_mem = mem;
162 Node *alloc_mem = alloc->as_Allocate()->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
163 assert(alloc_mem != nullptr, "Allocation without a memory projection.");
164 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
165 while (true) {
166 if (mem == alloc_mem || mem == start_mem ) {
167 return mem; // hit one of our sentinels
168 } else if (mem->is_MergeMem()) {
169 mem = mem->as_MergeMem()->memory_at(alias_idx);
170 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
171 Node *in = mem->in(0);
172 // we can safely skip over safepoints, calls, locks and membars because we
173 // already know that the object is safe to eliminate.
174 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
175 return in;
176 } else if (in->is_Call()) {
177 CallNode *call = in->as_Call();
178 if (call->may_modify(tinst, phase)) {
179 assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape");
180 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) {
181 return in;
182 }
183 }
184 mem = in->in(TypeFunc::Memory);
185 } else if (in->is_MemBar()) {
186 ArrayCopyNode* ac = nullptr;
187 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) {
188 if (ac != nullptr) {
189 assert(ac->is_clonebasic(), "Only basic clone is a non escaping clone");
190 return ac;
191 }
192 }
193 mem = in->in(TypeFunc::Memory);
194 } else {
195 #ifdef ASSERT
196 in->dump();
197 mem->dump();
198 assert(false, "unexpected projection");
199 #endif
200 }
201 } else if (mem->is_Store()) {
202 const TypePtr* atype = mem->as_Store()->adr_type();
203 int adr_idx = phase->C->get_alias_index(atype);
204 if (adr_idx == alias_idx) {
205 assert(atype->isa_oopptr(), "address type must be oopptr");
206 int adr_offset = atype->flat_offset();
207 uint adr_iid = atype->is_oopptr()->instance_id();
208 // Array elements references have the same alias_idx
209 // but different offset and different instance_id.
210 if (adr_offset == offset && adr_iid == alloc->_idx) {
211 return mem;
212 }
213 } else {
214 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
215 }
216 mem = mem->in(MemNode::Memory);
217 } else if (mem->is_ClearArray()) {
218 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
219 // Can not bypass initialization of the instance
220 // we are looking.
221 debug_only(intptr_t offset;)
222 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
223 InitializeNode* init = alloc->as_Allocate()->initialization();
224 // We are looking for stored value, return Initialize node
225 // or memory edge from Allocate node.
226 if (init != nullptr) {
227 return init;
228 } else {
229 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers).
230 }
231 }
232 // Otherwise skip it (the call updated 'mem' value).
233 } else if (mem->Opcode() == Op_SCMemProj) {
234 mem = mem->in(0);
235 Node* adr = nullptr;
236 if (mem->is_LoadStore()) {
237 adr = mem->in(MemNode::Address);
238 } else {
239 assert(mem->Opcode() == Op_EncodeISOArray ||
240 mem->Opcode() == Op_StrCompressedCopy, "sanity");
241 adr = mem->in(3); // Destination array
242 }
243 const TypePtr* atype = adr->bottom_type()->is_ptr();
244 int adr_idx = phase->C->get_alias_index(atype);
245 if (adr_idx == alias_idx) {
246 DEBUG_ONLY(mem->dump();)
247 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
248 return nullptr;
249 }
250 mem = mem->in(MemNode::Memory);
251 } else if (mem->Opcode() == Op_StrInflatedCopy) {
252 Node* adr = mem->in(3); // Destination array
253 const TypePtr* atype = adr->bottom_type()->is_ptr();
254 int adr_idx = phase->C->get_alias_index(atype);
255 if (adr_idx == alias_idx) {
256 DEBUG_ONLY(mem->dump();)
257 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
258 return nullptr;
259 }
260 mem = mem->in(MemNode::Memory);
261 } else {
262 return mem;
263 }
264 assert(mem != orig_mem, "dead memory loop");
265 }
266 }
267
268 // Generate loads from source of the arraycopy for fields of
269 // destination needed at a deoptimization point
270 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) {
271 BasicType bt = ft;
272 const Type *type = ftype;
273 if (ft == T_NARROWOOP) {
274 bt = T_OBJECT;
275 type = ftype->make_oopptr();
276 }
277 Node* res = nullptr;
278 if (ac->is_clonebasic()) {
279 assert(ac->in(ArrayCopyNode::Src) != ac->in(ArrayCopyNode::Dest), "clone source equals destination");
280 Node* base = ac->in(ArrayCopyNode::Src);
281 Node* adr = _igvn.transform(new AddPNode(base, base, _igvn.MakeConX(offset)));
282 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
283 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
284 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
285 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt);
286 } else {
287 if (ac->modifies(offset, offset, &_igvn, true)) {
288 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
289 uint shift = exact_log2(type2aelembytes(bt));
290 Node* src_pos = ac->in(ArrayCopyNode::SrcPos);
291 Node* dest_pos = ac->in(ArrayCopyNode::DestPos);
292 const TypeInt* src_pos_t = _igvn.type(src_pos)->is_int();
293 const TypeInt* dest_pos_t = _igvn.type(dest_pos)->is_int();
294
295 Node* adr = nullptr;
296 Node* base = ac->in(ArrayCopyNode::Src);
297 const TypeAryPtr* adr_type = _igvn.type(base)->is_aryptr();
298 if (adr_type->is_flat()) {
299 shift = adr_type->flat_log_elem_size();
300 }
301 if (src_pos_t->is_con() && dest_pos_t->is_con()) {
302 intptr_t off = ((src_pos_t->get_con() - dest_pos_t->get_con()) << shift) + offset;
303 adr = _igvn.transform(new AddPNode(base, base, _igvn.MakeConX(off)));
304 adr_type = _igvn.type(adr)->is_aryptr();
305 assert(adr_type == _igvn.type(base)->is_aryptr()->add_field_offset_and_offset(off), "incorrect address type");
306 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
307 // Don't emit a new load from src if src == dst but try to get the value from memory instead
308 return value_from_mem(ac->in(TypeFunc::Memory), ctl, ft, ftype, adr_type, alloc);
309 }
310 } else {
311 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
312 // Non constant offset in the array: we can't statically
313 // determine the value
314 return nullptr;
315 }
316 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
317 #ifdef _LP64
318 diff = _igvn.transform(new ConvI2LNode(diff));
319 #endif
320 diff = _igvn.transform(new LShiftXNode(diff, _igvn.intcon(shift)));
321
322 Node* off = _igvn.transform(new AddXNode(_igvn.MakeConX(offset), diff));
323 adr = _igvn.transform(new AddPNode(base, base, off));
324 // In the case of a flat inline type array, each field has its
325 // own slice so we need to extract the field being accessed from
326 // the address computation
327 adr_type = adr_type->add_field_offset_and_offset(offset)->add_offset(Type::OffsetBot)->is_aryptr();
328 adr = _igvn.transform(new CastPPNode(ctl, adr, adr_type));
329 }
330 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
331 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
332 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt);
333 }
334 }
335 if (res != nullptr) {
336 if (ftype->isa_narrowoop()) {
337 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
338 assert(res->isa_DecodeN(), "should be narrow oop");
339 res = _igvn.transform(new EncodePNode(res, ftype));
340 }
341 return res;
342 }
343 return nullptr;
344 }
345
346 //
347 // Given a Memory Phi, compute a value Phi containing the values from stores
348 // on the input paths.
349 // Note: this function is recursive, its depth is limited by the "level" argument
350 // Returns the computed Phi, or null if it cannot compute it.
351 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) {
352 assert(mem->is_Phi(), "sanity");
353 int alias_idx = C->get_alias_index(adr_t);
354 int offset = adr_t->flat_offset();
355 int instance_id = adr_t->instance_id();
356
357 // Check if an appropriate value phi already exists.
358 Node* region = mem->in(0);
359 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
360 Node* phi = region->fast_out(k);
361 if (phi->is_Phi() && phi != mem &&
362 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) {
363 return phi;
364 }
365 }
366 // Check if an appropriate new value phi already exists.
367 Node* new_phi = value_phis->find(mem->_idx);
368 if (new_phi != nullptr)
369 return new_phi;
370
371 if (level <= 0) {
372 return nullptr; // Give up: phi tree too deep
373 }
374 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
375 Node *alloc_mem = alloc->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
376 assert(alloc_mem != nullptr, "Allocation without a memory projection.");
377
378 uint length = mem->req();
379 GrowableArray <Node *> values(length, length, nullptr);
380
381 // create a new Phi for the value
382 PhiNode *phi = new PhiNode(mem->in(0), phi_type, nullptr, mem->_idx, instance_id, alias_idx, offset);
383 transform_later(phi);
384 value_phis->push(phi, mem->_idx);
385
386 for (uint j = 1; j < length; j++) {
387 Node *in = mem->in(j);
388 if (in == nullptr || in->is_top()) {
389 values.at_put(j, in);
390 } else {
391 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
392 if (val == start_mem || val == alloc_mem) {
393 // hit a sentinel, return appropriate value
394 Node* init_value = alloc->in(AllocateNode::InitValue);
395 if (init_value != nullptr) {
396 if (val == start_mem) {
397 // TODO 8350865 Somehow we ended up with root mem and therefore walked past the alloc. Fix this. Triggered by TestGenerated::test15
398 // Don't we need field_value_by_offset?
399 return nullptr;
400 }
401 values.at_put(j, init_value);
402 } else {
403 assert(alloc->in(AllocateNode::RawInitValue) == nullptr, "init value may not be null");
404 values.at_put(j, _igvn.zerocon(ft));
405 }
406 continue;
407 }
408 if (val->is_Initialize()) {
409 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
410 }
411 if (val == nullptr) {
412 return nullptr; // can't find a value on this path
413 }
414 if (val == mem) {
415 values.at_put(j, mem);
416 } else if (val->is_Store()) {
417 Node* n = val->in(MemNode::ValueIn);
418 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
419 n = bs->step_over_gc_barrier(n);
420 if (is_subword_type(ft)) {
421 n = Compile::narrow_value(ft, n, phi_type, &_igvn, true);
422 }
423 values.at_put(j, n);
424 } else if (val->is_Proj() && val->in(0) == alloc) {
425 Node* init_value = alloc->in(AllocateNode::InitValue);
426 if (init_value != nullptr) {
427 // TODO 8350865 Is this correct for non-all-zero init values? Don't we need field_value_by_offset?
428 values.at_put(j, init_value);
429 } else {
430 assert(alloc->in(AllocateNode::RawInitValue) == nullptr, "init value may not be null");
431 values.at_put(j, _igvn.zerocon(ft));
432 }
433 } else if (val->is_Phi()) {
434 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
435 if (val == nullptr) {
436 return nullptr;
437 }
438 values.at_put(j, val);
439 } else if (val->Opcode() == Op_SCMemProj) {
440 assert(val->in(0)->is_LoadStore() ||
441 val->in(0)->Opcode() == Op_EncodeISOArray ||
442 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity");
443 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
444 return nullptr;
445 } else if (val->is_ArrayCopy()) {
446 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc);
447 if (res == nullptr) {
448 return nullptr;
449 }
450 values.at_put(j, res);
451 } else if (val->is_top()) {
452 // This indicates that this path into the phi is dead. Top will eventually also propagate into the Region.
453 // IGVN will clean this up later.
454 values.at_put(j, val);
455 } else {
456 DEBUG_ONLY( val->dump(); )
457 assert(false, "unknown node on this path");
458 return nullptr; // unknown node on this path
459 }
460 }
461 }
462 // Set Phi's inputs
463 for (uint j = 1; j < length; j++) {
464 if (values.at(j) == mem) {
465 phi->init_req(j, phi);
466 } else {
467 phi->init_req(j, values.at(j));
468 }
469 }
470 return phi;
471 }
472
473 // Search the last value stored into the object's field.
474 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) {
475 assert(adr_t->is_known_instance_field(), "instance required");
476 int instance_id = adr_t->instance_id();
477 assert((uint)instance_id == alloc->_idx, "wrong allocation");
478
479 int alias_idx = C->get_alias_index(adr_t);
480 int offset = adr_t->flat_offset();
481 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
482 Node *alloc_mem = alloc->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
483 assert(alloc_mem != nullptr, "Allocation without a memory projection.");
484 VectorSet visited;
485
486 bool done = sfpt_mem == alloc_mem;
487 Node *mem = sfpt_mem;
488 while (!done) {
489 if (visited.test_set(mem->_idx)) {
490 return nullptr; // found a loop, give up
491 }
492 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
493 if (mem == start_mem || mem == alloc_mem) {
494 done = true; // hit a sentinel, return appropriate 0 value
495 } else if (mem->is_Initialize()) {
496 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
497 if (mem == nullptr) {
498 done = true; // Something went wrong.
499 } else if (mem->is_Store()) {
500 const TypePtr* atype = mem->as_Store()->adr_type();
501 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
502 done = true;
503 }
504 } else if (mem->is_Store()) {
505 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
506 assert(atype != nullptr, "address type must be oopptr");
507 assert(C->get_alias_index(atype) == alias_idx &&
508 atype->is_known_instance_field() && atype->flat_offset() == offset &&
509 atype->instance_id() == instance_id, "store is correct memory slice");
510 done = true;
511 } else if (mem->is_Phi()) {
512 // try to find a phi's unique input
513 Node *unique_input = nullptr;
514 Node *top = C->top();
515 for (uint i = 1; i < mem->req(); i++) {
516 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
517 if (n == nullptr || n == top || n == mem) {
518 continue;
519 } else if (unique_input == nullptr) {
520 unique_input = n;
521 } else if (unique_input != n) {
522 unique_input = top;
523 break;
524 }
525 }
526 if (unique_input != nullptr && unique_input != top) {
527 mem = unique_input;
528 } else {
529 done = true;
530 }
531 } else if (mem->is_ArrayCopy()) {
532 done = true;
533 } else {
534 DEBUG_ONLY( mem->dump(); )
535 assert(false, "unexpected node");
536 }
537 }
538 if (mem != nullptr) {
539 if (mem == start_mem || mem == alloc_mem) {
540 // hit a sentinel, return appropriate value
541 Node* init_value = alloc->in(AllocateNode::InitValue);
542 if (init_value != nullptr) {
543 if (adr_t->is_flat()) {
544 if (init_value->is_EncodeP()) {
545 init_value = init_value->in(1);
546 }
547 assert(adr_t->is_aryptr()->field_offset().get() != Type::OffsetBot, "Unknown offset");
548 offset = adr_t->is_aryptr()->field_offset().get() + init_value->bottom_type()->inline_klass()->payload_offset();
549 init_value = init_value->as_InlineType()->field_value_by_offset(offset, true);
550 if (ft == T_NARROWOOP) {
551 init_value = transform_later(new EncodePNode(init_value, init_value->bottom_type()->make_ptr()));
552 }
553 }
554 return init_value;
555 }
556 assert(alloc->in(AllocateNode::RawInitValue) == nullptr, "init value may not be null");
557 return _igvn.zerocon(ft);
558 } else if (mem->is_Store()) {
559 Node* n = mem->in(MemNode::ValueIn);
560 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
561 n = bs->step_over_gc_barrier(n);
562 return n;
563 } else if (mem->is_Phi()) {
564 // attempt to produce a Phi reflecting the values on the input paths of the Phi
565 Node_Stack value_phis(8);
566 Node* phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
567 if (phi != nullptr) {
568 return phi;
569 } else {
570 // Kill all new Phis
571 while(value_phis.is_nonempty()) {
572 Node* n = value_phis.node();
573 _igvn.replace_node(n, C->top());
574 value_phis.pop();
575 }
576 }
577 } else if (mem->is_ArrayCopy()) {
578 Node* ctl = mem->in(0);
579 Node* m = mem->in(TypeFunc::Memory);
580 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj()) {
581 // pin the loads in the uncommon trap path
582 ctl = sfpt_ctl;
583 m = sfpt_mem;
584 }
585 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc);
586 }
587 }
588 // Something went wrong.
589 return nullptr;
590 }
591
592 // Search the last value stored into the inline type's fields (for flat arrays).
593 Node* PhaseMacroExpand::inline_type_from_mem(ciInlineKlass* vk, const TypeAryPtr* elem_adr_type, int elem_idx, int offset_in_element, bool null_free, AllocateNode* alloc, SafePointNode* sfpt) {
594 auto report_failure = [&](int field_offset_in_element) {
595 #ifndef PRODUCT
596 if (PrintEliminateAllocations) {
597 ciInlineKlass* elem_klass = elem_adr_type->elem()->inline_klass();
598 int offset = field_offset_in_element + elem_klass->payload_offset();
599 ciField* flattened_field = elem_klass->get_field_by_offset(offset, false);
600 assert(flattened_field != nullptr, "must have a field of type %s at offset %d", elem_klass->name()->as_utf8(), offset);
601 tty->print("=== At SafePoint node %d can't find value of field [%s] of array element [%d]", sfpt->_idx, flattened_field->name()->as_utf8(), elem_idx);
602 tty->print(", which prevents elimination of: ");
603 alloc->dump();
604 }
605 #endif // PRODUCT
606 };
607
608 // Create a new InlineTypeNode and retrieve the field values from memory
609 InlineTypeNode* vt = InlineTypeNode::make_uninitialized(_igvn, vk, false);
610 transform_later(vt);
611 if (null_free) {
612 vt->set_is_init(_igvn);
613 } else {
614 int nm_offset_in_element = offset_in_element + vk->null_marker_offset_in_payload();
615 const TypeAryPtr* nm_adr_type = elem_adr_type->with_field_offset(nm_offset_in_element);
616 Node* nm_value = value_from_mem(sfpt->memory(), sfpt->control(), T_BOOLEAN, TypeInt::BOOL, nm_adr_type, alloc);
617 if (nm_value != nullptr) {
618 vt->set_is_init(_igvn, nm_value);
619 } else {
620 report_failure(nm_offset_in_element);
621 return nullptr;
622 }
623 }
624
625 for (int i = 0; i < vk->nof_declared_nonstatic_fields(); ++i) {
626 ciType* field_type = vt->field_type(i);
627 int field_offset_in_element = offset_in_element + vt->field_offset(i) - vk->payload_offset();
628 Node* field_value = nullptr;
629 if (vt->field_is_flat(i)) {
630 field_value = inline_type_from_mem(field_type->as_inline_klass(), elem_adr_type, elem_idx, field_offset_in_element, vt->field_is_null_free(i), alloc, sfpt);
631 } else {
632 const Type* ft = Type::get_const_type(field_type);
633 BasicType bt = type2field[field_type->basic_type()];
634 if (UseCompressedOops && !is_java_primitive(bt)) {
635 ft = ft->make_narrowoop();
636 bt = T_NARROWOOP;
637 }
638 // Each inline type field has its own memory slice
639 const TypeAryPtr* field_adr_type = elem_adr_type->with_field_offset(field_offset_in_element);
640 field_value = value_from_mem(sfpt->memory(), sfpt->control(), bt, ft, field_adr_type, alloc);
641 if (field_value == nullptr) {
642 report_failure(field_offset_in_element);
643 } else if (ft->isa_narrowoop()) {
644 assert(UseCompressedOops, "unexpected narrow oop");
645 if (field_value->is_EncodeP()) {
646 field_value = field_value->in(1);
647 } else if (!field_value->is_InlineType()) {
648 field_value = transform_later(new DecodeNNode(field_value, field_value->get_ptr_type()));
649 }
650 }
651 }
652 if (field_value != nullptr) {
653 vt->set_field_value(i, field_value);
654 } else {
655 return nullptr;
656 }
657 }
658 return vt;
659 }
660
661 // Check the possibility of scalar replacement.
662 bool PhaseMacroExpand::can_eliminate_allocation(PhaseIterGVN* igvn, AllocateNode *alloc, GrowableArray <SafePointNode *>* safepoints) {
663 // Scan the uses of the allocation to check for anything that would
664 // prevent us from eliminating it.
665 NOT_PRODUCT( const char* fail_eliminate = nullptr; )
666 DEBUG_ONLY( Node* disq_node = nullptr; )
667 bool can_eliminate = true;
668 bool reduce_merge_precheck = (safepoints == nullptr);
669
670 Unique_Node_List worklist;
671 Node* res = alloc->result_cast();
672 const TypeOopPtr* res_type = nullptr;
673 if (res == nullptr) {
674 // All users were eliminated.
675 } else if (!res->is_CheckCastPP()) {
676 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
677 can_eliminate = false;
678 } else {
679 worklist.push(res);
680 res_type = igvn->type(res)->isa_oopptr();
681 if (res_type == nullptr) {
682 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
683 can_eliminate = false;
684 } else if (!res_type->klass_is_exact()) {
685 NOT_PRODUCT(fail_eliminate = "Not an exact type.";)
686 can_eliminate = false;
687 } else if (res_type->isa_aryptr()) {
688 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
689 if (length < 0) {
690 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
691 can_eliminate = false;
692 }
693 }
694 }
695
696 while (can_eliminate && worklist.size() > 0) {
697 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
698 res = worklist.pop();
699 for (DUIterator_Fast jmax, j = res->fast_outs(jmax); j < jmax && can_eliminate; j++) {
700 Node* use = res->fast_out(j);
701
702 if (use->is_AddP()) {
703 const TypePtr* addp_type = igvn->type(use)->is_ptr();
704 int offset = addp_type->offset();
705
706 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
707 NOT_PRODUCT(fail_eliminate = "Undefined field reference";)
708 can_eliminate = false;
709 break;
710 }
711 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
712 k < kmax && can_eliminate; k++) {
713 Node* n = use->fast_out(k);
714 if (!n->is_Store() && n->Opcode() != Op_CastP2X && !bs->is_gc_pre_barrier_node(n) && !reduce_merge_precheck) {
715 DEBUG_ONLY(disq_node = n;)
716 if (n->is_Load() || n->is_LoadStore()) {
717 NOT_PRODUCT(fail_eliminate = "Field load";)
718 } else {
719 NOT_PRODUCT(fail_eliminate = "Not store field reference";)
720 }
721 can_eliminate = false;
722 }
723 }
724 } else if (use->is_ArrayCopy() &&
725 (use->as_ArrayCopy()->is_clonebasic() ||
726 use->as_ArrayCopy()->is_arraycopy_validated() ||
727 use->as_ArrayCopy()->is_copyof_validated() ||
728 use->as_ArrayCopy()->is_copyofrange_validated()) &&
729 use->in(ArrayCopyNode::Dest) == res) {
730 // ok to eliminate
731 } else if (use->is_SafePoint()) {
732 SafePointNode* sfpt = use->as_SafePoint();
733 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
734 // Object is passed as argument.
735 DEBUG_ONLY(disq_node = use;)
736 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
737 can_eliminate = false;
738 }
739 Node* sfptMem = sfpt->memory();
740 if (sfptMem == nullptr || sfptMem->is_top()) {
741 DEBUG_ONLY(disq_node = use;)
742 NOT_PRODUCT(fail_eliminate = "null or TOP memory";)
743 can_eliminate = false;
744 } else if (!reduce_merge_precheck) {
745 assert(!res->is_Phi() || !res->as_Phi()->can_be_inline_type(), "Inline type allocations should not have safepoint uses");
746 safepoints->append_if_missing(sfpt);
747 }
748 } else if (use->is_InlineType() && use->as_InlineType()->get_oop() == res) {
749 // Look at uses
750 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); k < kmax; k++) {
751 Node* u = use->fast_out(k);
752 if (u->is_InlineType()) {
753 // Use in flat field can be eliminated
754 InlineTypeNode* vt = u->as_InlineType();
755 for (uint i = 0; i < vt->field_count(); ++i) {
756 if (vt->field_value(i) == use && !vt->field_is_flat(i)) {
757 can_eliminate = false; // Use in non-flat field
758 break;
759 }
760 }
761 } else {
762 // Add other uses to the worklist to process individually
763 worklist.push(use);
764 }
765 }
766 } else if (use->Opcode() == Op_StoreX && use->in(MemNode::Address) == res) {
767 // Store to mark word of inline type larval buffer
768 assert(res_type->is_inlinetypeptr(), "Unexpected store to mark word");
769 } else if (res_type->is_inlinetypeptr() && (use->Opcode() == Op_MemBarRelease || use->Opcode() == Op_MemBarStoreStore)) {
770 // Inline type buffer allocations are followed by a membar
771 } else if (reduce_merge_precheck &&
772 (use->is_Phi() || use->is_EncodeP() ||
773 use->Opcode() == Op_MemBarRelease ||
774 (UseStoreStoreForCtor && use->Opcode() == Op_MemBarStoreStore))) {
775 // Nothing to do
776 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
777 if (use->is_Phi()) {
778 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
779 NOT_PRODUCT(fail_eliminate = "Object is return value";)
780 } else {
781 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
782 }
783 DEBUG_ONLY(disq_node = use;)
784 } else {
785 if (use->Opcode() == Op_Return) {
786 NOT_PRODUCT(fail_eliminate = "Object is return value";)
787 } else {
788 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
789 }
790 DEBUG_ONLY(disq_node = use;)
791 }
792 can_eliminate = false;
793 } else {
794 assert(use->Opcode() == Op_CastP2X, "should be");
795 assert(!use->has_out_with(Op_OrL), "should have been removed because oop is never null");
796 }
797 }
798 }
799
800 #ifndef PRODUCT
801 if (PrintEliminateAllocations && safepoints != nullptr) {
802 if (can_eliminate) {
803 tty->print("Scalar ");
804 if (res == nullptr)
805 alloc->dump();
806 else
807 res->dump();
808 } else {
809 tty->print("NotScalar (%s)", fail_eliminate);
810 if (res == nullptr)
811 alloc->dump();
812 else
813 res->dump();
814 #ifdef ASSERT
815 if (disq_node != nullptr) {
816 tty->print(" >>>> ");
817 disq_node->dump();
818 }
819 #endif /*ASSERT*/
820 }
821 }
822
823 if (TraceReduceAllocationMerges && !can_eliminate && reduce_merge_precheck) {
824 tty->print_cr("\tCan't eliminate allocation because '%s': ", fail_eliminate != nullptr ? fail_eliminate : "");
825 DEBUG_ONLY(if (disq_node != nullptr) disq_node->dump();)
826 }
827 #endif
828 return can_eliminate;
829 }
830
831 void PhaseMacroExpand::undo_previous_scalarizations(GrowableArray <SafePointNode *> safepoints_done, AllocateNode* alloc) {
832 Node* res = alloc->result_cast();
833 int nfields = 0;
834 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
835
836 if (res != nullptr) {
837 const TypeOopPtr* res_type = _igvn.type(res)->isa_oopptr();
838
839 if (res_type->isa_instptr()) {
840 // find the fields of the class which will be needed for safepoint debug information
841 ciInstanceKlass* iklass = res_type->is_instptr()->instance_klass();
842 nfields = iklass->nof_nonstatic_fields();
843 } else {
844 // find the array's elements which will be needed for safepoint debug information
845 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
846 assert(nfields >= 0, "must be an array klass.");
847 }
848 }
849
850 // rollback processed safepoints
851 while (safepoints_done.length() > 0) {
852 SafePointNode* sfpt_done = safepoints_done.pop();
853 // remove any extra entries we added to the safepoint
854 uint last = sfpt_done->req() - 1;
855 for (int k = 0; k < nfields; k++) {
856 sfpt_done->del_req(last--);
857 }
858 JVMState *jvms = sfpt_done->jvms();
859 jvms->set_endoff(sfpt_done->req());
860 // Now make a pass over the debug information replacing any references
861 // to SafePointScalarObjectNode with the allocated object.
862 int start = jvms->debug_start();
863 int end = jvms->debug_end();
864 for (int i = start; i < end; i++) {
865 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
866 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
867 if (scobj->first_index(jvms) == sfpt_done->req() &&
868 scobj->n_fields() == (uint)nfields) {
869 assert(scobj->alloc() == alloc, "sanity");
870 sfpt_done->set_req(i, res);
871 }
872 }
873 }
874 _igvn._worklist.push(sfpt_done);
875 }
876 }
877
878 void PhaseMacroExpand::process_field_value_at_safepoint(const Type* field_type, Node* field_val, SafePointNode* sfpt, Unique_Node_List* value_worklist) {
879 if (UseCompressedOops && field_type->isa_narrowoop()) {
880 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
881 // to be able scalar replace the allocation.
882 if (field_val->is_EncodeP()) {
883 field_val = field_val->in(1);
884 } else if (!field_val->is_InlineType()) {
885 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
886 }
887 }
888
889 // Keep track of inline types to scalarize them later
890 if (field_val->is_InlineType()) {
891 value_worklist->push(field_val);
892 } else if (field_val->is_Phi()) {
893 PhiNode* phi = field_val->as_Phi();
894 // Eagerly replace inline type phis now since we could be removing an inline type allocation where we must
895 // scalarize all its fields in safepoints.
896 field_val = phi->try_push_inline_types_down(&_igvn, true);
897 if (field_val->is_InlineType()) {
898 value_worklist->push(field_val);
899 }
900 }
901 sfpt->add_req(field_val);
902 }
903
904 bool PhaseMacroExpand::add_array_elems_to_safepoint(AllocateNode* alloc, const TypeAryPtr* array_type, SafePointNode* sfpt, Unique_Node_List* value_worklist) {
905 const Type* elem_type = array_type->elem();
906 BasicType basic_elem_type = elem_type->array_element_basic_type();
907
908 intptr_t elem_size;
909 if (array_type->is_flat()) {
910 elem_size = array_type->flat_elem_size();
911 } else {
912 elem_size = type2aelembytes(basic_elem_type);
913 }
914
915 int n_elems = alloc->in(AllocateNode::ALength)->get_int();
916 for (int elem_idx = 0; elem_idx < n_elems; elem_idx++) {
917 intptr_t elem_offset = arrayOopDesc::base_offset_in_bytes(basic_elem_type) + elem_idx * elem_size;
918 const TypeAryPtr* elem_adr_type = array_type->with_offset(elem_offset);
919 Node* elem_val;
920 if (array_type->is_flat()) {
921 ciInlineKlass* elem_klass = elem_type->inline_klass();
922 assert(elem_klass->maybe_flat_in_array(), "must be flat in array");
923 elem_val = inline_type_from_mem(elem_klass, elem_adr_type, elem_idx, 0, array_type->is_null_free(), alloc, sfpt);
924 } else {
925 elem_val = value_from_mem(sfpt->memory(), sfpt->control(), basic_elem_type, elem_type, elem_adr_type, alloc);
926 #ifndef PRODUCT
927 if (PrintEliminateAllocations && elem_val == nullptr) {
928 tty->print("=== At SafePoint node %d can't find value of array element [%d]", sfpt->_idx, elem_idx);
929 tty->print(", which prevents elimination of: ");
930 alloc->dump();
931 }
932 #endif // PRODUCT
933 }
934 if (elem_val == nullptr) {
935 return false;
936 }
937
938 process_field_value_at_safepoint(elem_type, elem_val, sfpt, value_worklist);
939 }
940
941 return true;
942 }
943
944 // Recursively adds all flattened fields of a type 'iklass' inside 'base' to 'sfpt'.
945 // 'offset_minus_header' refers to the offset of the payload of 'iklass' inside 'base' minus the
946 // payload offset of 'iklass'. If 'base' is of type 'iklass' then 'offset_minus_header' == 0.
947 bool PhaseMacroExpand::add_inst_fields_to_safepoint(ciInstanceKlass* iklass, AllocateNode* alloc, Node* base, int offset_minus_header, SafePointNode* sfpt, Unique_Node_List* value_worklist) {
948 const TypeInstPtr* base_type = _igvn.type(base)->is_instptr();
949 auto report_failure = [&](int offset) {
950 #ifndef PRODUCT
951 if (PrintEliminateAllocations) {
952 ciInstanceKlass* base_klass = base_type->instance_klass();
953 ciField* flattened_field = base_klass->get_field_by_offset(offset, false);
954 assert(flattened_field != nullptr, "must have a field of type %s at offset %d", base_klass->name()->as_utf8(), offset);
955 tty->print("=== At SafePoint node %d can't find value of field: ", sfpt->_idx);
956 flattened_field->print();
957 int field_idx = C->alias_type(flattened_field)->index();
958 tty->print(" (alias_idx=%d)", field_idx);
959 tty->print(", which prevents elimination of: ");
960 base->dump();
961 }
962 #endif // PRODUCT
963 };
964
965 for (int i = 0; i < iklass->nof_declared_nonstatic_fields(); i++) {
966 ciField* field = iklass->declared_nonstatic_field_at(i);
967 if (field->is_flat()) {
968 ciInlineKlass* fvk = field->type()->as_inline_klass();
969 int field_offset_minus_header = offset_minus_header + field->offset_in_bytes() - fvk->payload_offset();
970 bool success = add_inst_fields_to_safepoint(fvk, alloc, base, field_offset_minus_header, sfpt, value_worklist);
971 if (!success) {
972 return false;
973 }
974
975 // The null marker of a field is added right after we scalarize that field
976 if (!field->is_null_free()) {
977 int nm_offset = offset_minus_header + field->null_marker_offset();
978 Node* null_marker = value_from_mem(sfpt->memory(), sfpt->control(), T_BOOLEAN, TypeInt::BOOL, base_type->with_offset(nm_offset), alloc);
979 if (null_marker == nullptr) {
980 report_failure(nm_offset);
981 return false;
982 }
983 process_field_value_at_safepoint(TypeInt::BOOL, null_marker, sfpt, value_worklist);
984 }
985
986 continue;
987 }
988
989 int offset = offset_minus_header + field->offset_in_bytes();
990 ciType* elem_type = field->type();
991 BasicType basic_elem_type = field->layout_type();
992
993 const Type* field_type;
994 if (is_reference_type(basic_elem_type)) {
995 if (!elem_type->is_loaded()) {
996 field_type = TypeInstPtr::BOTTOM;
997 } else {
998 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
999 }
1000 if (UseCompressedOops) {
1001 field_type = field_type->make_narrowoop();
1002 basic_elem_type = T_NARROWOOP;
1003 }
1004 } else {
1005 field_type = Type::get_const_basic_type(basic_elem_type);
1006 }
1007
1008 const TypeInstPtr* field_addr_type = base_type->add_offset(offset)->isa_instptr();
1009 Node* field_val = value_from_mem(sfpt->memory(), sfpt->control(), basic_elem_type, field_type, field_addr_type, alloc);
1010 if (field_val == nullptr) {
1011 report_failure(offset);
1012 return false;
1013 }
1014 process_field_value_at_safepoint(field_type, field_val, sfpt, value_worklist);
1015 }
1016
1017 return true;
1018 }
1019
1020 SafePointScalarObjectNode* PhaseMacroExpand::create_scalarized_object_description(AllocateNode* alloc, SafePointNode* sfpt,
1021 Unique_Node_List* value_worklist) {
1022 // Fields of scalar objs are referenced only at the end
1023 // of regular debuginfo at the last (youngest) JVMS.
1024 // Record relative start index.
1025 ciInstanceKlass* iklass = nullptr;
1026 const TypeOopPtr* res_type = nullptr;
1027 int nfields = 0;
1028 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
1029 Node* res = alloc->result_cast();
1030
1031 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
1032 assert(sfpt->jvms() != nullptr, "missed JVMS");
1033 uint before_sfpt_req = sfpt->req();
1034
1035 if (res != nullptr) { // Could be null when there are no users
1036 res_type = _igvn.type(res)->isa_oopptr();
1037
1038 if (res_type->isa_instptr()) {
1039 // find the fields of the class which will be needed for safepoint debug information
1040 iklass = res_type->is_instptr()->instance_klass();
1041 nfields = iklass->nof_nonstatic_fields();
1042 } else {
1043 // find the array's elements which will be needed for safepoint debug information
1044 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
1045 assert(nfields >= 0, "must be an array klass.");
1046 }
1047
1048 if (res->bottom_type()->is_inlinetypeptr()) {
1049 // Nullable inline types have an IsInit field which is added to the safepoint when scalarizing them (see
1050 // InlineTypeNode::make_scalar_in_safepoint()). When having circular inline types, we stop scalarizing at depth 1
1051 // to avoid an endless recursion. Therefore, we do not have a SafePointScalarObjectNode node here, yet.
1052 // We are about to create a SafePointScalarObjectNode as if this is a normal object. Add an additional int input
1053 // with value 1 which sets IsInit to true to indicate that the object is always non-null. This input is checked
1054 // later in PhaseOutput::filLocArray() for inline types.
1055 sfpt->add_req(_igvn.intcon(1));
1056 }
1057 }
1058
1059 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type, alloc, first_ind, sfpt->jvms()->depth(), nfields);
1060 sobj->init_req(0, C->root());
1061 transform_later(sobj);
1062
1063 if (res == nullptr) {
1064 sfpt->jvms()->set_endoff(sfpt->req());
1065 return sobj;
1066 }
1067
1068 bool success;
1069 if (iklass == nullptr) {
1070 success = add_array_elems_to_safepoint(alloc, res_type->is_aryptr(), sfpt, value_worklist);
1071 } else {
1072 success = add_inst_fields_to_safepoint(iklass, alloc, res, 0, sfpt, value_worklist);
1073 }
1074
1075 // We weren't able to find a value for this field, remove all the fields added to the safepoint
1076 if (!success) {
1077 for (uint i = sfpt->req() - 1; i >= before_sfpt_req; i--) {
1078 sfpt->del_req(i);
1079 }
1080 _igvn._worklist.push(sfpt);
1081 return nullptr;
1082 }
1083
1084 sfpt->jvms()->set_endoff(sfpt->req());
1085 return sobj;
1086 }
1087
1088 // Do scalar replacement.
1089 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
1090 GrowableArray <SafePointNode *> safepoints_done;
1091 Node* res = alloc->result_cast();
1092 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
1093 const TypeOopPtr* res_type = nullptr;
1094 if (res != nullptr) { // Could be null when there are no users
1095 res_type = _igvn.type(res)->isa_oopptr();
1096 }
1097
1098 // Process the safepoint uses
1099 Unique_Node_List value_worklist;
1100 while (safepoints.length() > 0) {
1101 SafePointNode* sfpt = safepoints.pop();
1102 SafePointScalarObjectNode* sobj = create_scalarized_object_description(alloc, sfpt, &value_worklist);
1103
1104 if (sobj == nullptr) {
1105 undo_previous_scalarizations(safepoints_done, alloc);
1106 return false;
1107 }
1108
1109 // Now make a pass over the debug information replacing any references
1110 // to the allocated object with "sobj"
1111 JVMState *jvms = sfpt->jvms();
1112 sfpt->replace_edges_in_range(res, sobj, jvms->debug_start(), jvms->debug_end(), &_igvn);
1113 _igvn._worklist.push(sfpt);
1114
1115 // keep it for rollback
1116 safepoints_done.append_if_missing(sfpt);
1117 }
1118 // Scalarize inline types that were added to the safepoint.
1119 // Don't allow linking a constant oop (if available) for flat array elements
1120 // because Deoptimization::reassign_flat_array_elements needs field values.
1121 bool allow_oop = (res_type != nullptr) && !res_type->is_flat();
1122 for (uint i = 0; i < value_worklist.size(); ++i) {
1123 InlineTypeNode* vt = value_worklist.at(i)->as_InlineType();
1124 vt->make_scalar_in_safepoints(&_igvn, allow_oop);
1125 }
1126 return true;
1127 }
1128
1129 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
1130 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control);
1131 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory);
1132 if (ctl_proj != nullptr) {
1133 igvn.replace_node(ctl_proj, n->in(0));
1134 }
1135 if (mem_proj != nullptr) {
1136 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
1137 }
1138 }
1139
1140 // Process users of eliminated allocation.
1141 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc, bool inline_alloc) {
1142 Unique_Node_List worklist;
1143 Node* res = alloc->result_cast();
1144 if (res != nullptr) {
1145 worklist.push(res);
1146 }
1147 while (worklist.size() > 0) {
1148 res = worklist.pop();
1149 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
1150 Node *use = res->last_out(j);
1151 uint oc1 = res->outcnt();
1152
1153 if (use->is_AddP()) {
1154 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
1155 Node *n = use->last_out(k);
1156 uint oc2 = use->outcnt();
1157 if (n->is_Store()) {
1158 for (DUIterator_Fast pmax, p = n->fast_outs(pmax); p < pmax; p++) {
1159 MemBarNode* mb = n->fast_out(p)->isa_MemBar();
1160 if (mb != nullptr && mb->req() <= MemBarNode::Precedent && mb->in(MemBarNode::Precedent) == n) {
1161 // MemBarVolatiles should have been removed by MemBarNode::Ideal() for non-inline allocations
1162 assert(inline_alloc, "MemBarVolatile should be eliminated for non-escaping object");
1163 mb->remove(&_igvn);
1164 }
1165 }
1166 _igvn.replace_node(n, n->in(MemNode::Memory));
1167 } else {
1168 eliminate_gc_barrier(n);
1169 }
1170 k -= (oc2 - use->outcnt());
1171 }
1172 _igvn.remove_dead_node(use);
1173 } else if (use->is_ArrayCopy()) {
1174 // Disconnect ArrayCopy node
1175 ArrayCopyNode* ac = use->as_ArrayCopy();
1176 if (ac->is_clonebasic()) {
1177 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
1178 disconnect_projections(ac, _igvn);
1179 assert(alloc->in(TypeFunc::Memory)->is_Proj() && alloc->in(TypeFunc::Memory)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
1180 Node* membar_before = alloc->in(TypeFunc::Memory)->in(0);
1181 disconnect_projections(membar_before->as_MemBar(), _igvn);
1182 if (membar_after->is_MemBar()) {
1183 disconnect_projections(membar_after->as_MemBar(), _igvn);
1184 }
1185 } else {
1186 assert(ac->is_arraycopy_validated() ||
1187 ac->is_copyof_validated() ||
1188 ac->is_copyofrange_validated(), "unsupported");
1189 CallProjections* callprojs = ac->extract_projections(true);
1190
1191 _igvn.replace_node(callprojs->fallthrough_ioproj, ac->in(TypeFunc::I_O));
1192 _igvn.replace_node(callprojs->fallthrough_memproj, ac->in(TypeFunc::Memory));
1193 _igvn.replace_node(callprojs->fallthrough_catchproj, ac->in(TypeFunc::Control));
1194
1195 // Set control to top. IGVN will remove the remaining projections
1196 ac->set_req(0, top());
1197 ac->replace_edge(res, top(), &_igvn);
1198
1199 // Disconnect src right away: it can help find new
1200 // opportunities for allocation elimination
1201 Node* src = ac->in(ArrayCopyNode::Src);
1202 ac->replace_edge(src, top(), &_igvn);
1203 // src can be top at this point if src and dest of the
1204 // arraycopy were the same
1205 if (src->outcnt() == 0 && !src->is_top()) {
1206 _igvn.remove_dead_node(src);
1207 }
1208 }
1209 _igvn._worklist.push(ac);
1210 } else if (use->is_InlineType()) {
1211 assert(use->as_InlineType()->get_oop() == res, "unexpected inline type ptr use");
1212 // Cut off oop input and remove known instance id from type
1213 _igvn.rehash_node_delayed(use);
1214 use->as_InlineType()->set_oop(_igvn, _igvn.zerocon(T_OBJECT));
1215 const TypeOopPtr* toop = _igvn.type(use)->is_oopptr()->cast_to_instance_id(TypeOopPtr::InstanceBot);
1216 _igvn.set_type(use, toop);
1217 use->as_InlineType()->set_type(toop);
1218 // Process users
1219 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); k < kmax; k++) {
1220 Node* u = use->fast_out(k);
1221 if (!u->is_InlineType()) {
1222 worklist.push(u);
1223 }
1224 }
1225 } else if (use->Opcode() == Op_StoreX && use->in(MemNode::Address) == res) {
1226 // Store to mark word of inline type larval buffer
1227 assert(inline_alloc, "Unexpected store to mark word");
1228 _igvn.replace_node(use, use->in(MemNode::Memory));
1229 } else if (use->Opcode() == Op_MemBarRelease || use->Opcode() == Op_MemBarStoreStore) {
1230 // Inline type buffer allocations are followed by a membar
1231 assert(inline_alloc, "Unexpected MemBarRelease");
1232 use->as_MemBar()->remove(&_igvn);
1233 } else {
1234 eliminate_gc_barrier(use);
1235 }
1236 j -= (oc1 - res->outcnt());
1237 }
1238 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
1239 _igvn.remove_dead_node(res);
1240 }
1241
1242 //
1243 // Process other users of allocation's projections
1244 //
1245 if (_callprojs->resproj[0] != nullptr && _callprojs->resproj[0]->outcnt() != 0) {
1246 // First disconnect stores captured by Initialize node.
1247 // If Initialize node is eliminated first in the following code,
1248 // it will kill such stores and DUIterator_Last will assert.
1249 for (DUIterator_Fast jmax, j = _callprojs->resproj[0]->fast_outs(jmax); j < jmax; j++) {
1250 Node* use = _callprojs->resproj[0]->fast_out(j);
1251 if (use->is_AddP()) {
1252 // raw memory addresses used only by the initialization
1253 _igvn.replace_node(use, C->top());
1254 --j; --jmax;
1255 }
1256 }
1257 for (DUIterator_Last jmin, j = _callprojs->resproj[0]->last_outs(jmin); j >= jmin; ) {
1258 Node* use = _callprojs->resproj[0]->last_out(j);
1259 uint oc1 = _callprojs->resproj[0]->outcnt();
1260 if (use->is_Initialize()) {
1261 // Eliminate Initialize node.
1262 InitializeNode *init = use->as_Initialize();
1263 assert(init->outcnt() <= 2, "only a control and memory projection expected");
1264 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control);
1265 if (ctrl_proj != nullptr) {
1266 _igvn.replace_node(ctrl_proj, init->in(TypeFunc::Control));
1267 #ifdef ASSERT
1268 // If the InitializeNode has no memory out, it will die, and tmp will become null
1269 Node* tmp = init->in(TypeFunc::Control);
1270 assert(tmp == nullptr || tmp == _callprojs->fallthrough_catchproj, "allocation control projection");
1271 #endif
1272 }
1273 Node *mem_proj = init->proj_out_or_null(TypeFunc::Memory);
1274 if (mem_proj != nullptr) {
1275 Node *mem = init->in(TypeFunc::Memory);
1276 #ifdef ASSERT
1277 if (mem->is_MergeMem()) {
1278 assert(mem->in(TypeFunc::Memory) == _callprojs->fallthrough_memproj, "allocation memory projection");
1279 } else {
1280 assert(mem == _callprojs->fallthrough_memproj, "allocation memory projection");
1281 }
1282 #endif
1283 _igvn.replace_node(mem_proj, mem);
1284 }
1285 } else if (use->Opcode() == Op_MemBarStoreStore) {
1286 // Inline type buffer allocations are followed by a membar
1287 assert(inline_alloc, "Unexpected MemBarStoreStore");
1288 use->as_MemBar()->remove(&_igvn);
1289 } else {
1290 assert(false, "only Initialize or AddP expected");
1291 }
1292 j -= (oc1 - _callprojs->resproj[0]->outcnt());
1293 }
1294 }
1295 if (_callprojs->fallthrough_catchproj != nullptr) {
1296 _igvn.replace_node(_callprojs->fallthrough_catchproj, alloc->in(TypeFunc::Control));
1297 }
1298 if (_callprojs->fallthrough_memproj != nullptr) {
1299 _igvn.replace_node(_callprojs->fallthrough_memproj, alloc->in(TypeFunc::Memory));
1300 }
1301 if (_callprojs->catchall_memproj != nullptr) {
1302 _igvn.replace_node(_callprojs->catchall_memproj, C->top());
1303 }
1304 if (_callprojs->fallthrough_ioproj != nullptr) {
1305 _igvn.replace_node(_callprojs->fallthrough_ioproj, alloc->in(TypeFunc::I_O));
1306 }
1307 if (_callprojs->catchall_ioproj != nullptr) {
1308 _igvn.replace_node(_callprojs->catchall_ioproj, C->top());
1309 }
1310 if (_callprojs->catchall_catchproj != nullptr) {
1311 _igvn.replace_node(_callprojs->catchall_catchproj, C->top());
1312 }
1313 }
1314
1315 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1316 // If reallocation fails during deoptimization we'll pop all
1317 // interpreter frames for this compiled frame and that won't play
1318 // nice with JVMTI popframe.
1319 // We avoid this issue by eager reallocation when the popframe request
1320 // is received.
1321 if (!EliminateAllocations) {
1322 return false;
1323 }
1324 Node* klass = alloc->in(AllocateNode::KlassNode);
1325 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1326
1327 // Attempt to eliminate inline type buffer allocations
1328 // regardless of usage and escape/replaceable status.
1329 bool inline_alloc = tklass->isa_instklassptr() &&
1330 tklass->is_instklassptr()->instance_klass()->is_inlinetype();
1331 if (!alloc->_is_non_escaping && !inline_alloc) {
1332 return false;
1333 }
1334 // Eliminate boxing allocations which are not used
1335 // regardless scalar replaceable status.
1336 Node* res = alloc->result_cast();
1337 bool boxing_alloc = (res == nullptr) && C->eliminate_boxing() &&
1338 tklass->isa_instklassptr() &&
1339 tklass->is_instklassptr()->instance_klass()->is_box_klass();
1340 if (!alloc->_is_scalar_replaceable && !boxing_alloc && !inline_alloc) {
1341 return false;
1342 }
1343
1344 _callprojs = alloc->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1345
1346 GrowableArray <SafePointNode *> safepoints;
1347 if (!can_eliminate_allocation(&_igvn, alloc, &safepoints)) {
1348 return false;
1349 }
1350
1351 if (!alloc->_is_scalar_replaceable) {
1352 assert(res == nullptr || inline_alloc, "sanity");
1353 // We can only eliminate allocation if all debug info references
1354 // are already replaced with SafePointScalarObject because
1355 // we can't search for a fields value without instance_id.
1356 if (safepoints.length() > 0) {
1357 return false;
1358 }
1359 }
1360
1361 if (!scalar_replacement(alloc, safepoints)) {
1362 return false;
1363 }
1364
1365 CompileLog* log = C->log();
1366 if (log != nullptr) {
1367 log->head("eliminate_allocation type='%d'",
1368 log->identify(tklass->exact_klass()));
1369 JVMState* p = alloc->jvms();
1370 while (p != nullptr) {
1371 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1372 p = p->caller();
1373 }
1374 log->tail("eliminate_allocation");
1375 }
1376
1377 process_users_of_allocation(alloc, inline_alloc);
1378
1379 #ifndef PRODUCT
1380 if (PrintEliminateAllocations) {
1381 if (alloc->is_AllocateArray())
1382 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1383 else
1384 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1385 }
1386 #endif
1387
1388 return true;
1389 }
1390
1391 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1392 // EA should remove all uses of non-escaping boxing node.
1393 if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != nullptr) {
1394 return false;
1395 }
1396
1397 assert(boxing->result_cast() == nullptr, "unexpected boxing node result");
1398
1399 _callprojs = boxing->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1400
1401 const TypeTuple* r = boxing->tf()->range_sig();
1402 assert(r->cnt() > TypeFunc::Parms, "sanity");
1403 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1404 assert(t != nullptr, "sanity");
1405
1406 CompileLog* log = C->log();
1407 if (log != nullptr) {
1408 log->head("eliminate_boxing type='%d'",
1409 log->identify(t->instance_klass()));
1410 JVMState* p = boxing->jvms();
1411 while (p != nullptr) {
1412 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1413 p = p->caller();
1414 }
1415 log->tail("eliminate_boxing");
1416 }
1417
1418 process_users_of_allocation(boxing);
1419
1420 #ifndef PRODUCT
1421 if (PrintEliminateAllocations) {
1422 tty->print("++++ Eliminated: %d ", boxing->_idx);
1423 boxing->method()->print_short_name(tty);
1424 tty->cr();
1425 }
1426 #endif
1427
1428 return true;
1429 }
1430
1431
1432 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
1433 Node* adr = basic_plus_adr(base, offset);
1434 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
1435 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered);
1436 transform_later(value);
1437 return value;
1438 }
1439
1440
1441 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
1442 Node* adr = basic_plus_adr(base, offset);
1443 mem = StoreNode::make(_igvn, ctl, mem, adr, nullptr, value, bt, MemNode::unordered);
1444 transform_later(mem);
1445 return mem;
1446 }
1447
1448 //=============================================================================
1449 //
1450 // A L L O C A T I O N
1451 //
1452 // Allocation attempts to be fast in the case of frequent small objects.
1453 // It breaks down like this:
1454 //
1455 // 1) Size in doublewords is computed. This is a constant for objects and
1456 // variable for most arrays. Doubleword units are used to avoid size
1457 // overflow of huge doubleword arrays. We need doublewords in the end for
1458 // rounding.
1459 //
1460 // 2) Size is checked for being 'too large'. Too-large allocations will go
1461 // the slow path into the VM. The slow path can throw any required
1462 // exceptions, and does all the special checks for very large arrays. The
1463 // size test can constant-fold away for objects. For objects with
1464 // finalizers it constant-folds the otherway: you always go slow with
1465 // finalizers.
1466 //
1467 // 3) If NOT using TLABs, this is the contended loop-back point.
1468 // Load-Locked the heap top. If using TLABs normal-load the heap top.
1469 //
1470 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
1471 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
1472 // "size*8" we always enter the VM, where "largish" is a constant picked small
1473 // enough that there's always space between the eden max and 4Gig (old space is
1474 // there so it's quite large) and large enough that the cost of entering the VM
1475 // is dwarfed by the cost to initialize the space.
1476 //
1477 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1478 // down. If contended, repeat at step 3. If using TLABs normal-store
1479 // adjusted heap top back down; there is no contention.
1480 //
1481 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
1482 // fields.
1483 //
1484 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1485 // oop flavor.
1486 //
1487 //=============================================================================
1488 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1489 // Allocations bigger than this always go the slow route.
1490 // This value must be small enough that allocation attempts that need to
1491 // trigger exceptions go the slow route. Also, it must be small enough so
1492 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1493 //=============================================================================j//
1494 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1495 // The allocator will coalesce int->oop copies away. See comment in
1496 // coalesce.cpp about how this works. It depends critically on the exact
1497 // code shape produced here, so if you are changing this code shape
1498 // make sure the GC info for the heap-top is correct in and around the
1499 // slow-path call.
1500 //
1501
1502 void PhaseMacroExpand::expand_allocate_common(
1503 AllocateNode* alloc, // allocation node to be expanded
1504 Node* length, // array length for an array allocation
1505 Node* init_val, // value to initialize the array with
1506 const TypeFunc* slow_call_type, // Type of slow call
1507 address slow_call_address, // Address of slow call
1508 Node* valid_length_test // whether length is valid or not
1509 )
1510 {
1511 Node* ctrl = alloc->in(TypeFunc::Control);
1512 Node* mem = alloc->in(TypeFunc::Memory);
1513 Node* i_o = alloc->in(TypeFunc::I_O);
1514 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1515 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1516 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1517 assert(ctrl != nullptr, "must have control");
1518
1519 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1520 // they will not be used if "always_slow" is set
1521 enum { slow_result_path = 1, fast_result_path = 2 };
1522 Node *result_region = nullptr;
1523 Node *result_phi_rawmem = nullptr;
1524 Node *result_phi_rawoop = nullptr;
1525 Node *result_phi_i_o = nullptr;
1526
1527 // The initial slow comparison is a size check, the comparison
1528 // we want to do is a BoolTest::gt
1529 bool expand_fast_path = true;
1530 int tv = _igvn.find_int_con(initial_slow_test, -1);
1531 if (tv >= 0) {
1532 // InitialTest has constant result
1533 // 0 - can fit in TLAB
1534 // 1 - always too big or negative
1535 assert(tv <= 1, "0 or 1 if a constant");
1536 expand_fast_path = (tv == 0);
1537 initial_slow_test = nullptr;
1538 } else {
1539 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1540 }
1541
1542 if (!UseTLAB) {
1543 // Force slow-path allocation
1544 expand_fast_path = false;
1545 initial_slow_test = nullptr;
1546 }
1547
1548 bool allocation_has_use = (alloc->result_cast() != nullptr);
1549 if (!allocation_has_use) {
1550 InitializeNode* init = alloc->initialization();
1551 if (init != nullptr) {
1552 init->remove(&_igvn);
1553 }
1554 if (expand_fast_path && (initial_slow_test == nullptr)) {
1555 // Remove allocation node and return.
1556 // Size is a non-negative constant -> no initial check needed -> directly to fast path.
1557 // Also, no usages -> empty fast path -> no fall out to slow path -> nothing left.
1558 #ifndef PRODUCT
1559 if (PrintEliminateAllocations) {
1560 tty->print("NotUsed ");
1561 Node* res = alloc->proj_out_or_null(TypeFunc::Parms);
1562 if (res != nullptr) {
1563 res->dump();
1564 } else {
1565 alloc->dump();
1566 }
1567 }
1568 #endif
1569 yank_alloc_node(alloc);
1570 return;
1571 }
1572 }
1573
1574 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1575 Node *slow_region = nullptr;
1576 Node *toobig_false = ctrl;
1577
1578 // generate the initial test if necessary
1579 if (initial_slow_test != nullptr ) {
1580 assert (expand_fast_path, "Only need test if there is a fast path");
1581 slow_region = new RegionNode(3);
1582
1583 // Now make the initial failure test. Usually a too-big test but
1584 // might be a TRUE for finalizers.
1585 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1586 transform_later(toobig_iff);
1587 // Plug the failing-too-big test into the slow-path region
1588 Node* toobig_true = new IfTrueNode(toobig_iff);
1589 transform_later(toobig_true);
1590 slow_region ->init_req( too_big_or_final_path, toobig_true );
1591 toobig_false = new IfFalseNode(toobig_iff);
1592 transform_later(toobig_false);
1593 } else {
1594 // No initial test, just fall into next case
1595 assert(allocation_has_use || !expand_fast_path, "Should already have been handled");
1596 toobig_false = ctrl;
1597 debug_only(slow_region = NodeSentinel);
1598 }
1599
1600 // If we are here there are several possibilities
1601 // - expand_fast_path is false - then only a slow path is expanded. That's it.
1602 // no_initial_check means a constant allocation.
1603 // - If check always evaluates to false -> expand_fast_path is false (see above)
1604 // - If check always evaluates to true -> directly into fast path (but may bailout to slowpath)
1605 // if !allocation_has_use the fast path is empty
1606 // if !allocation_has_use && no_initial_check
1607 // - Then there are no fastpath that can fall out to slowpath -> no allocation code at all.
1608 // removed by yank_alloc_node above.
1609
1610 Node *slow_mem = mem; // save the current memory state for slow path
1611 // generate the fast allocation code unless we know that the initial test will always go slow
1612 if (expand_fast_path) {
1613 // Fast path modifies only raw memory.
1614 if (mem->is_MergeMem()) {
1615 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1616 }
1617
1618 // allocate the Region and Phi nodes for the result
1619 result_region = new RegionNode(3);
1620 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1621 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1622
1623 // Grab regular I/O before optional prefetch may change it.
1624 // Slow-path does no I/O so just set it to the original I/O.
1625 result_phi_i_o->init_req(slow_result_path, i_o);
1626
1627 // Name successful fast-path variables
1628 Node* fast_oop_ctrl;
1629 Node* fast_oop_rawmem;
1630
1631 if (allocation_has_use) {
1632 Node* needgc_ctrl = nullptr;
1633 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1634
1635 intx prefetch_lines = length != nullptr ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1636 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1637 Node* fast_oop = bs->obj_allocate(this, mem, toobig_false, size_in_bytes, i_o, needgc_ctrl,
1638 fast_oop_ctrl, fast_oop_rawmem,
1639 prefetch_lines);
1640
1641 if (initial_slow_test != nullptr) {
1642 // This completes all paths into the slow merge point
1643 slow_region->init_req(need_gc_path, needgc_ctrl);
1644 transform_later(slow_region);
1645 } else {
1646 // No initial slow path needed!
1647 // Just fall from the need-GC path straight into the VM call.
1648 slow_region = needgc_ctrl;
1649 }
1650
1651 InitializeNode* init = alloc->initialization();
1652 fast_oop_rawmem = initialize_object(alloc,
1653 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1654 klass_node, length, size_in_bytes);
1655 expand_initialize_membar(alloc, init, fast_oop_ctrl, fast_oop_rawmem);
1656 expand_dtrace_alloc_probe(alloc, fast_oop, fast_oop_ctrl, fast_oop_rawmem);
1657
1658 result_phi_rawoop->init_req(fast_result_path, fast_oop);
1659 } else {
1660 assert (initial_slow_test != nullptr, "sanity");
1661 fast_oop_ctrl = toobig_false;
1662 fast_oop_rawmem = mem;
1663 transform_later(slow_region);
1664 }
1665
1666 // Plug in the successful fast-path into the result merge point
1667 result_region ->init_req(fast_result_path, fast_oop_ctrl);
1668 result_phi_i_o ->init_req(fast_result_path, i_o);
1669 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1670 } else {
1671 slow_region = ctrl;
1672 result_phi_i_o = i_o; // Rename it to use in the following code.
1673 }
1674
1675 // Generate slow-path call
1676 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1677 OptoRuntime::stub_name(slow_call_address),
1678 TypePtr::BOTTOM);
1679 call->init_req(TypeFunc::Control, slow_region);
1680 call->init_req(TypeFunc::I_O, top()); // does no i/o
1681 call->init_req(TypeFunc::Memory, slow_mem); // may gc ptrs
1682 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1683 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1684
1685 call->init_req(TypeFunc::Parms+0, klass_node);
1686 if (length != nullptr) {
1687 call->init_req(TypeFunc::Parms+1, length);
1688 if (init_val != nullptr) {
1689 call->init_req(TypeFunc::Parms+2, init_val);
1690 }
1691 } else {
1692 // Let the runtime know if this is a larval allocation
1693 call->init_req(TypeFunc::Parms+1, _igvn.intcon(alloc->_larval));
1694 }
1695
1696 // Copy debug information and adjust JVMState information, then replace
1697 // allocate node with the call
1698 call->copy_call_debug_info(&_igvn, alloc);
1699 // For array allocations, copy the valid length check to the call node so Compile::final_graph_reshaping() can verify
1700 // that the call has the expected number of CatchProj nodes (in case the allocation always fails and the fallthrough
1701 // path dies).
1702 if (valid_length_test != nullptr) {
1703 call->add_req(valid_length_test);
1704 }
1705 if (expand_fast_path) {
1706 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1707 } else {
1708 // Hook i_o projection to avoid its elimination during allocation
1709 // replacement (when only a slow call is generated).
1710 call->set_req(TypeFunc::I_O, result_phi_i_o);
1711 }
1712 _igvn.replace_node(alloc, call);
1713 transform_later(call);
1714
1715 // Identify the output projections from the allocate node and
1716 // adjust any references to them.
1717 // The control and io projections look like:
1718 //
1719 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1720 // Allocate Catch
1721 // ^---Proj(io) <-------+ ^---CatchProj(io)
1722 //
1723 // We are interested in the CatchProj nodes.
1724 //
1725 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1726
1727 // An allocate node has separate memory projections for the uses on
1728 // the control and i_o paths. Replace the control memory projection with
1729 // result_phi_rawmem (unless we are only generating a slow call when
1730 // both memory projections are combined)
1731 if (expand_fast_path && _callprojs->fallthrough_memproj != nullptr) {
1732 _igvn.replace_in_uses(_callprojs->fallthrough_memproj, result_phi_rawmem);
1733 }
1734 // Now change uses of catchall_memproj to use fallthrough_memproj and delete
1735 // catchall_memproj so we end up with a call that has only 1 memory projection.
1736 if (_callprojs->catchall_memproj != nullptr) {
1737 if (_callprojs->fallthrough_memproj == nullptr) {
1738 _callprojs->fallthrough_memproj = new ProjNode(call, TypeFunc::Memory);
1739 transform_later(_callprojs->fallthrough_memproj);
1740 }
1741 _igvn.replace_in_uses(_callprojs->catchall_memproj, _callprojs->fallthrough_memproj);
1742 _igvn.remove_dead_node(_callprojs->catchall_memproj);
1743 }
1744
1745 // An allocate node has separate i_o projections for the uses on the control
1746 // and i_o paths. Always replace the control i_o projection with result i_o
1747 // otherwise incoming i_o become dead when only a slow call is generated
1748 // (it is different from memory projections where both projections are
1749 // combined in such case).
1750 if (_callprojs->fallthrough_ioproj != nullptr) {
1751 _igvn.replace_in_uses(_callprojs->fallthrough_ioproj, result_phi_i_o);
1752 }
1753 // Now change uses of catchall_ioproj to use fallthrough_ioproj and delete
1754 // catchall_ioproj so we end up with a call that has only 1 i_o projection.
1755 if (_callprojs->catchall_ioproj != nullptr) {
1756 if (_callprojs->fallthrough_ioproj == nullptr) {
1757 _callprojs->fallthrough_ioproj = new ProjNode(call, TypeFunc::I_O);
1758 transform_later(_callprojs->fallthrough_ioproj);
1759 }
1760 _igvn.replace_in_uses(_callprojs->catchall_ioproj, _callprojs->fallthrough_ioproj);
1761 _igvn.remove_dead_node(_callprojs->catchall_ioproj);
1762 }
1763
1764 // if we generated only a slow call, we are done
1765 if (!expand_fast_path) {
1766 // Now we can unhook i_o.
1767 if (result_phi_i_o->outcnt() > 1) {
1768 call->set_req(TypeFunc::I_O, top());
1769 } else {
1770 assert(result_phi_i_o->unique_ctrl_out() == call, "sanity");
1771 // Case of new array with negative size known during compilation.
1772 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1773 // following code since call to runtime will throw exception.
1774 // As result there will be no users of i_o after the call.
1775 // Leave i_o attached to this call to avoid problems in preceding graph.
1776 }
1777 return;
1778 }
1779
1780 if (_callprojs->fallthrough_catchproj != nullptr) {
1781 ctrl = _callprojs->fallthrough_catchproj->clone();
1782 transform_later(ctrl);
1783 _igvn.replace_node(_callprojs->fallthrough_catchproj, result_region);
1784 } else {
1785 ctrl = top();
1786 }
1787 Node *slow_result;
1788 if (_callprojs->resproj[0] == nullptr) {
1789 // no uses of the allocation result
1790 slow_result = top();
1791 } else {
1792 slow_result = _callprojs->resproj[0]->clone();
1793 transform_later(slow_result);
1794 _igvn.replace_node(_callprojs->resproj[0], result_phi_rawoop);
1795 }
1796
1797 // Plug slow-path into result merge point
1798 result_region->init_req( slow_result_path, ctrl);
1799 transform_later(result_region);
1800 if (allocation_has_use) {
1801 result_phi_rawoop->init_req(slow_result_path, slow_result);
1802 transform_later(result_phi_rawoop);
1803 }
1804 result_phi_rawmem->init_req(slow_result_path, _callprojs->fallthrough_memproj);
1805 transform_later(result_phi_rawmem);
1806 transform_later(result_phi_i_o);
1807 // This completes all paths into the result merge point
1808 }
1809
1810 // Remove alloc node that has no uses.
1811 void PhaseMacroExpand::yank_alloc_node(AllocateNode* alloc) {
1812 Node* ctrl = alloc->in(TypeFunc::Control);
1813 Node* mem = alloc->in(TypeFunc::Memory);
1814 Node* i_o = alloc->in(TypeFunc::I_O);
1815
1816 _callprojs = alloc->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1817 if (_callprojs->resproj[0] != nullptr) {
1818 for (DUIterator_Fast imax, i = _callprojs->resproj[0]->fast_outs(imax); i < imax; i++) {
1819 Node* use = _callprojs->resproj[0]->fast_out(i);
1820 use->isa_MemBar()->remove(&_igvn);
1821 --imax;
1822 --i; // back up iterator
1823 }
1824 assert(_callprojs->resproj[0]->outcnt() == 0, "all uses must be deleted");
1825 _igvn.remove_dead_node(_callprojs->resproj[0]);
1826 }
1827 if (_callprojs->fallthrough_catchproj != nullptr) {
1828 _igvn.replace_in_uses(_callprojs->fallthrough_catchproj, ctrl);
1829 _igvn.remove_dead_node(_callprojs->fallthrough_catchproj);
1830 }
1831 if (_callprojs->catchall_catchproj != nullptr) {
1832 _igvn.rehash_node_delayed(_callprojs->catchall_catchproj);
1833 _callprojs->catchall_catchproj->set_req(0, top());
1834 }
1835 if (_callprojs->fallthrough_proj != nullptr) {
1836 Node* catchnode = _callprojs->fallthrough_proj->unique_ctrl_out();
1837 _igvn.remove_dead_node(catchnode);
1838 _igvn.remove_dead_node(_callprojs->fallthrough_proj);
1839 }
1840 if (_callprojs->fallthrough_memproj != nullptr) {
1841 _igvn.replace_in_uses(_callprojs->fallthrough_memproj, mem);
1842 _igvn.remove_dead_node(_callprojs->fallthrough_memproj);
1843 }
1844 if (_callprojs->fallthrough_ioproj != nullptr) {
1845 _igvn.replace_in_uses(_callprojs->fallthrough_ioproj, i_o);
1846 _igvn.remove_dead_node(_callprojs->fallthrough_ioproj);
1847 }
1848 if (_callprojs->catchall_memproj != nullptr) {
1849 _igvn.rehash_node_delayed(_callprojs->catchall_memproj);
1850 _callprojs->catchall_memproj->set_req(0, top());
1851 }
1852 if (_callprojs->catchall_ioproj != nullptr) {
1853 _igvn.rehash_node_delayed(_callprojs->catchall_ioproj);
1854 _callprojs->catchall_ioproj->set_req(0, top());
1855 }
1856 #ifndef PRODUCT
1857 if (PrintEliminateAllocations) {
1858 if (alloc->is_AllocateArray()) {
1859 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1860 } else {
1861 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1862 }
1863 }
1864 #endif
1865 _igvn.remove_dead_node(alloc);
1866 }
1867
1868 void PhaseMacroExpand::expand_initialize_membar(AllocateNode* alloc, InitializeNode* init,
1869 Node*& fast_oop_ctrl, Node*& fast_oop_rawmem) {
1870 // If initialization is performed by an array copy, any required
1871 // MemBarStoreStore was already added. If the object does not
1872 // escape no need for a MemBarStoreStore. If the object does not
1873 // escape in its initializer and memory barrier (MemBarStoreStore or
1874 // stronger) is already added at exit of initializer, also no need
1875 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore
1876 // so that stores that initialize this object can't be reordered
1877 // with a subsequent store that makes this object accessible by
1878 // other threads.
1879 // Other threads include java threads and JVM internal threads
1880 // (for example concurrent GC threads). Current concurrent GC
1881 // implementation: G1 will not scan newly created object,
1882 // so it's safe to skip storestore barrier when allocation does
1883 // not escape.
1884 if (!alloc->does_not_escape_thread() &&
1885 !alloc->is_allocation_MemBar_redundant() &&
1886 (init == nullptr || !init->is_complete_with_arraycopy())) {
1887 if (init == nullptr || init->req() < InitializeNode::RawStores) {
1888 // No InitializeNode or no stores captured by zeroing
1889 // elimination. Simply add the MemBarStoreStore after object
1890 // initialization.
1891 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1892 transform_later(mb);
1893
1894 mb->init_req(TypeFunc::Memory, fast_oop_rawmem);
1895 mb->init_req(TypeFunc::Control, fast_oop_ctrl);
1896 fast_oop_ctrl = new ProjNode(mb, TypeFunc::Control);
1897 transform_later(fast_oop_ctrl);
1898 fast_oop_rawmem = new ProjNode(mb, TypeFunc::Memory);
1899 transform_later(fast_oop_rawmem);
1900 } else {
1901 // Add the MemBarStoreStore after the InitializeNode so that
1902 // all stores performing the initialization that were moved
1903 // before the InitializeNode happen before the storestore
1904 // barrier.
1905
1906 Node* init_ctrl = init->proj_out_or_null(TypeFunc::Control);
1907 Node* init_mem = init->proj_out_or_null(TypeFunc::Memory);
1908
1909 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1910 transform_later(mb);
1911
1912 Node* ctrl = new ProjNode(init, TypeFunc::Control);
1913 transform_later(ctrl);
1914 Node* mem = new ProjNode(init, TypeFunc::Memory);
1915 transform_later(mem);
1916
1917 // The MemBarStoreStore depends on control and memory coming
1918 // from the InitializeNode
1919 mb->init_req(TypeFunc::Memory, mem);
1920 mb->init_req(TypeFunc::Control, ctrl);
1921
1922 ctrl = new ProjNode(mb, TypeFunc::Control);
1923 transform_later(ctrl);
1924 mem = new ProjNode(mb, TypeFunc::Memory);
1925 transform_later(mem);
1926
1927 // All nodes that depended on the InitializeNode for control
1928 // and memory must now depend on the MemBarNode that itself
1929 // depends on the InitializeNode
1930 if (init_ctrl != nullptr) {
1931 _igvn.replace_node(init_ctrl, ctrl);
1932 }
1933 if (init_mem != nullptr) {
1934 _igvn.replace_node(init_mem, mem);
1935 }
1936 }
1937 }
1938 }
1939
1940 void PhaseMacroExpand::expand_dtrace_alloc_probe(AllocateNode* alloc, Node* oop,
1941 Node*& ctrl, Node*& rawmem) {
1942 if (C->env()->dtrace_alloc_probes()) {
1943 // Slow-path call
1944 int size = TypeFunc::Parms + 2;
1945 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1946 CAST_FROM_FN_PTR(address,
1947 static_cast<int (*)(JavaThread*, oopDesc*)>(SharedRuntime::dtrace_object_alloc)),
1948 "dtrace_object_alloc",
1949 TypeRawPtr::BOTTOM);
1950
1951 // Get base of thread-local storage area
1952 Node* thread = new ThreadLocalNode();
1953 transform_later(thread);
1954
1955 call->init_req(TypeFunc::Parms + 0, thread);
1956 call->init_req(TypeFunc::Parms + 1, oop);
1957 call->init_req(TypeFunc::Control, ctrl);
1958 call->init_req(TypeFunc::I_O , top()); // does no i/o
1959 call->init_req(TypeFunc::Memory , rawmem);
1960 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1961 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1962 transform_later(call);
1963 ctrl = new ProjNode(call, TypeFunc::Control);
1964 transform_later(ctrl);
1965 rawmem = new ProjNode(call, TypeFunc::Memory);
1966 transform_later(rawmem);
1967 }
1968 }
1969
1970 // Helper for PhaseMacroExpand::expand_allocate_common.
1971 // Initializes the newly-allocated storage.
1972 Node* PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1973 Node* control, Node* rawmem, Node* object,
1974 Node* klass_node, Node* length,
1975 Node* size_in_bytes) {
1976 InitializeNode* init = alloc->initialization();
1977 // Store the klass & mark bits
1978 Node* mark_node = alloc->make_ideal_mark(&_igvn, control, rawmem);
1979 if (!mark_node->is_Con()) {
1980 transform_later(mark_node);
1981 }
1982 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, TypeX_X->basic_type());
1983
1984 if (!UseCompactObjectHeaders) {
1985 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1986 }
1987 int header_size = alloc->minimum_header_size(); // conservatively small
1988
1989 // Array length
1990 if (length != nullptr) { // Arrays need length field
1991 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1992 // conservatively small header size:
1993 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1994 if (_igvn.type(klass_node)->isa_aryklassptr()) { // we know the exact header size in most cases:
1995 BasicType elem = _igvn.type(klass_node)->is_klassptr()->as_instance_type()->isa_aryptr()->elem()->array_element_basic_type();
1996 if (is_reference_type(elem, true)) {
1997 elem = T_OBJECT;
1998 }
1999 header_size = Klass::layout_helper_header_size(Klass::array_layout_helper(elem));
2000 }
2001 }
2002
2003 // Clear the object body, if necessary.
2004 if (init == nullptr) {
2005 // The init has somehow disappeared; be cautious and clear everything.
2006 //
2007 // This can happen if a node is allocated but an uncommon trap occurs
2008 // immediately. In this case, the Initialize gets associated with the
2009 // trap, and may be placed in a different (outer) loop, if the Allocate
2010 // is in a loop. If (this is rare) the inner loop gets unrolled, then
2011 // there can be two Allocates to one Initialize. The answer in all these
2012 // edge cases is safety first. It is always safe to clear immediately
2013 // within an Allocate, and then (maybe or maybe not) clear some more later.
2014 if (!(UseTLAB && ZeroTLAB)) {
2015 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
2016 alloc->in(AllocateNode::InitValue),
2017 alloc->in(AllocateNode::RawInitValue),
2018 header_size, size_in_bytes,
2019 &_igvn);
2020 }
2021 } else {
2022 if (!init->is_complete()) {
2023 // Try to win by zeroing only what the init does not store.
2024 // We can also try to do some peephole optimizations,
2025 // such as combining some adjacent subword stores.
2026 rawmem = init->complete_stores(control, rawmem, object,
2027 header_size, size_in_bytes, &_igvn);
2028 }
2029 // We have no more use for this link, since the AllocateNode goes away:
2030 init->set_req(InitializeNode::RawAddress, top());
2031 // (If we keep the link, it just confuses the register allocator,
2032 // who thinks he sees a real use of the address by the membar.)
2033 }
2034
2035 return rawmem;
2036 }
2037
2038 // Generate prefetch instructions for next allocations.
2039 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
2040 Node*& contended_phi_rawmem,
2041 Node* old_eden_top, Node* new_eden_top,
2042 intx lines) {
2043 enum { fall_in_path = 1, pf_path = 2 };
2044 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
2045 // Generate prefetch allocation with watermark check.
2046 // As an allocation hits the watermark, we will prefetch starting
2047 // at a "distance" away from watermark.
2048
2049 Node *pf_region = new RegionNode(3);
2050 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY,
2051 TypeRawPtr::BOTTOM );
2052 // I/O is used for Prefetch
2053 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO );
2054
2055 Node *thread = new ThreadLocalNode();
2056 transform_later(thread);
2057
2058 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread,
2059 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
2060 transform_later(eden_pf_adr);
2061
2062 Node *old_pf_wm = new LoadPNode(needgc_false,
2063 contended_phi_rawmem, eden_pf_adr,
2064 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM,
2065 MemNode::unordered);
2066 transform_later(old_pf_wm);
2067
2068 // check against new_eden_top
2069 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm );
2070 transform_later(need_pf_cmp);
2071 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge );
2072 transform_later(need_pf_bol);
2073 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol,
2074 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
2075 transform_later(need_pf_iff);
2076
2077 // true node, add prefetchdistance
2078 Node *need_pf_true = new IfTrueNode( need_pf_iff );
2079 transform_later(need_pf_true);
2080
2081 Node *need_pf_false = new IfFalseNode( need_pf_iff );
2082 transform_later(need_pf_false);
2083
2084 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm,
2085 _igvn.MakeConX(AllocatePrefetchDistance) );
2086 transform_later(new_pf_wmt );
2087 new_pf_wmt->set_req(0, need_pf_true);
2088
2089 Node *store_new_wmt = new StorePNode(need_pf_true,
2090 contended_phi_rawmem, eden_pf_adr,
2091 TypeRawPtr::BOTTOM, new_pf_wmt,
2092 MemNode::unordered);
2093 transform_later(store_new_wmt);
2094
2095 // adding prefetches
2096 pf_phi_abio->init_req( fall_in_path, i_o );
2097
2098 Node *prefetch_adr;
2099 Node *prefetch;
2100 uint step_size = AllocatePrefetchStepSize;
2101 uint distance = 0;
2102
2103 for ( intx i = 0; i < lines; i++ ) {
2104 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt,
2105 _igvn.MakeConX(distance) );
2106 transform_later(prefetch_adr);
2107 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
2108 transform_later(prefetch);
2109 distance += step_size;
2110 i_o = prefetch;
2111 }
2112 pf_phi_abio->set_req( pf_path, i_o );
2113
2114 pf_region->init_req( fall_in_path, need_pf_false );
2115 pf_region->init_req( pf_path, need_pf_true );
2116
2117 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
2118 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
2119
2120 transform_later(pf_region);
2121 transform_later(pf_phi_rawmem);
2122 transform_later(pf_phi_abio);
2123
2124 needgc_false = pf_region;
2125 contended_phi_rawmem = pf_phi_rawmem;
2126 i_o = pf_phi_abio;
2127 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) {
2128 // Insert a prefetch instruction for each allocation.
2129 // This code is used to generate 1 prefetch instruction per cache line.
2130
2131 // Generate several prefetch instructions.
2132 uint step_size = AllocatePrefetchStepSize;
2133 uint distance = AllocatePrefetchDistance;
2134
2135 // Next cache address.
2136 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top,
2137 _igvn.MakeConX(step_size + distance));
2138 transform_later(cache_adr);
2139 cache_adr = new CastP2XNode(needgc_false, cache_adr);
2140 transform_later(cache_adr);
2141 // Address is aligned to execute prefetch to the beginning of cache line size
2142 // (it is important when BIS instruction is used on SPARC as prefetch).
2143 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1));
2144 cache_adr = new AndXNode(cache_adr, mask);
2145 transform_later(cache_adr);
2146 cache_adr = new CastX2PNode(cache_adr);
2147 transform_later(cache_adr);
2148
2149 // Prefetch
2150 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr );
2151 prefetch->set_req(0, needgc_false);
2152 transform_later(prefetch);
2153 contended_phi_rawmem = prefetch;
2154 Node *prefetch_adr;
2155 distance = step_size;
2156 for ( intx i = 1; i < lines; i++ ) {
2157 prefetch_adr = new AddPNode( cache_adr, cache_adr,
2158 _igvn.MakeConX(distance) );
2159 transform_later(prefetch_adr);
2160 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr );
2161 transform_later(prefetch);
2162 distance += step_size;
2163 contended_phi_rawmem = prefetch;
2164 }
2165 } else if( AllocatePrefetchStyle > 0 ) {
2166 // Insert a prefetch for each allocation only on the fast-path
2167 Node *prefetch_adr;
2168 Node *prefetch;
2169 // Generate several prefetch instructions.
2170 uint step_size = AllocatePrefetchStepSize;
2171 uint distance = AllocatePrefetchDistance;
2172 for ( intx i = 0; i < lines; i++ ) {
2173 prefetch_adr = new AddPNode( old_eden_top, new_eden_top,
2174 _igvn.MakeConX(distance) );
2175 transform_later(prefetch_adr);
2176 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
2177 // Do not let it float too high, since if eden_top == eden_end,
2178 // both might be null.
2179 if( i == 0 ) { // Set control for first prefetch, next follows it
2180 prefetch->init_req(0, needgc_false);
2181 }
2182 transform_later(prefetch);
2183 distance += step_size;
2184 i_o = prefetch;
2185 }
2186 }
2187 return i_o;
2188 }
2189
2190
2191 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
2192 expand_allocate_common(alloc, nullptr, nullptr,
2193 OptoRuntime::new_instance_Type(),
2194 OptoRuntime::new_instance_Java(), nullptr);
2195 }
2196
2197 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
2198 Node* length = alloc->in(AllocateNode::ALength);
2199 Node* valid_length_test = alloc->in(AllocateNode::ValidLengthTest);
2200 InitializeNode* init = alloc->initialization();
2201 Node* klass_node = alloc->in(AllocateNode::KlassNode);
2202 Node* init_value = alloc->in(AllocateNode::InitValue);
2203 const TypeAryKlassPtr* ary_klass_t = _igvn.type(klass_node)->isa_aryklassptr();
2204 const TypeFunc* slow_call_type;
2205 address slow_call_address; // Address of slow call
2206 if (init != nullptr && init->is_complete_with_arraycopy() &&
2207 ary_klass_t && ary_klass_t->elem()->isa_klassptr() == nullptr) {
2208 // Don't zero type array during slow allocation in VM since
2209 // it will be initialized later by arraycopy in compiled code.
2210 slow_call_address = OptoRuntime::new_array_nozero_Java();
2211 slow_call_type = OptoRuntime::new_array_nozero_Type();
2212 } else {
2213 slow_call_address = OptoRuntime::new_array_Java();
2214 slow_call_type = OptoRuntime::new_array_Type();
2215
2216 if (init_value == nullptr) {
2217 init_value = _igvn.zerocon(T_OBJECT);
2218 } else if (UseCompressedOops) {
2219 init_value = transform_later(new DecodeNNode(init_value, init_value->bottom_type()->make_ptr()));
2220 }
2221 }
2222 expand_allocate_common(alloc, length, init_value,
2223 slow_call_type,
2224 slow_call_address, valid_length_test);
2225 }
2226
2227 //-------------------mark_eliminated_box----------------------------------
2228 //
2229 // During EA obj may point to several objects but after few ideal graph
2230 // transformations (CCP) it may point to only one non escaping object
2231 // (but still using phi), corresponding locks and unlocks will be marked
2232 // for elimination. Later obj could be replaced with a new node (new phi)
2233 // and which does not have escape information. And later after some graph
2234 // reshape other locks and unlocks (which were not marked for elimination
2235 // before) are connected to this new obj (phi) but they still will not be
2236 // marked for elimination since new obj has no escape information.
2237 // Mark all associated (same box and obj) lock and unlock nodes for
2238 // elimination if some of them marked already.
2239 void PhaseMacroExpand::mark_eliminated_box(Node* box, Node* obj) {
2240 BoxLockNode* oldbox = box->as_BoxLock();
2241 if (oldbox->is_eliminated()) {
2242 return; // This BoxLock node was processed already.
2243 }
2244 assert(!oldbox->is_unbalanced(), "this should not be called for unbalanced region");
2245 // New implementation (EliminateNestedLocks) has separate BoxLock
2246 // node for each locked region so mark all associated locks/unlocks as
2247 // eliminated even if different objects are referenced in one locked region
2248 // (for example, OSR compilation of nested loop inside locked scope).
2249 if (EliminateNestedLocks ||
2250 oldbox->as_BoxLock()->is_simple_lock_region(nullptr, obj, nullptr)) {
2251 // Box is used only in one lock region. Mark this box as eliminated.
2252 oldbox->set_local(); // This verifies correct state of BoxLock
2253 _igvn.hash_delete(oldbox);
2254 oldbox->set_eliminated(); // This changes box's hash value
2255 _igvn.hash_insert(oldbox);
2256
2257 for (uint i = 0; i < oldbox->outcnt(); i++) {
2258 Node* u = oldbox->raw_out(i);
2259 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) {
2260 AbstractLockNode* alock = u->as_AbstractLock();
2261 // Check lock's box since box could be referenced by Lock's debug info.
2262 if (alock->box_node() == oldbox) {
2263 // Mark eliminated all related locks and unlocks.
2264 #ifdef ASSERT
2265 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4");
2266 #endif
2267 alock->set_non_esc_obj();
2268 }
2269 }
2270 }
2271 return;
2272 }
2273
2274 // Create new "eliminated" BoxLock node and use it in monitor debug info
2275 // instead of oldbox for the same object.
2276 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
2277
2278 // Note: BoxLock node is marked eliminated only here and it is used
2279 // to indicate that all associated lock and unlock nodes are marked
2280 // for elimination.
2281 newbox->set_local(); // This verifies correct state of BoxLock
2282 newbox->set_eliminated();
2283 transform_later(newbox);
2284
2285 // Replace old box node with new box for all users of the same object.
2286 for (uint i = 0; i < oldbox->outcnt();) {
2287 bool next_edge = true;
2288
2289 Node* u = oldbox->raw_out(i);
2290 if (u->is_AbstractLock()) {
2291 AbstractLockNode* alock = u->as_AbstractLock();
2292 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) {
2293 // Replace Box and mark eliminated all related locks and unlocks.
2294 #ifdef ASSERT
2295 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5");
2296 #endif
2297 alock->set_non_esc_obj();
2298 _igvn.rehash_node_delayed(alock);
2299 alock->set_box_node(newbox);
2300 next_edge = false;
2301 }
2302 }
2303 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) {
2304 FastLockNode* flock = u->as_FastLock();
2305 assert(flock->box_node() == oldbox, "sanity");
2306 _igvn.rehash_node_delayed(flock);
2307 flock->set_box_node(newbox);
2308 next_edge = false;
2309 }
2310
2311 // Replace old box in monitor debug info.
2312 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
2313 SafePointNode* sfn = u->as_SafePoint();
2314 JVMState* youngest_jvms = sfn->jvms();
2315 int max_depth = youngest_jvms->depth();
2316 for (int depth = 1; depth <= max_depth; depth++) {
2317 JVMState* jvms = youngest_jvms->of_depth(depth);
2318 int num_mon = jvms->nof_monitors();
2319 // Loop over monitors
2320 for (int idx = 0; idx < num_mon; idx++) {
2321 Node* obj_node = sfn->monitor_obj(jvms, idx);
2322 Node* box_node = sfn->monitor_box(jvms, idx);
2323 if (box_node == oldbox && obj_node->eqv_uncast(obj)) {
2324 int j = jvms->monitor_box_offset(idx);
2325 _igvn.replace_input_of(u, j, newbox);
2326 next_edge = false;
2327 }
2328 }
2329 }
2330 }
2331 if (next_edge) i++;
2332 }
2333 }
2334
2335 //-----------------------mark_eliminated_locking_nodes-----------------------
2336 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) {
2337 if (!alock->is_balanced()) {
2338 return; // Can't do any more elimination for this locking region
2339 }
2340 if (EliminateNestedLocks) {
2341 if (alock->is_nested()) {
2342 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity");
2343 return;
2344 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened
2345 // Only Lock node has JVMState needed here.
2346 // Not that preceding claim is documented anywhere else.
2347 if (alock->jvms() != nullptr) {
2348 if (alock->as_Lock()->is_nested_lock_region()) {
2349 // Mark eliminated related nested locks and unlocks.
2350 Node* obj = alock->obj_node();
2351 BoxLockNode* box_node = alock->box_node()->as_BoxLock();
2352 assert(!box_node->is_eliminated(), "should not be marked yet");
2353 // Note: BoxLock node is marked eliminated only here
2354 // and it is used to indicate that all associated lock
2355 // and unlock nodes are marked for elimination.
2356 box_node->set_eliminated(); // Box's hash is always NO_HASH here
2357 for (uint i = 0; i < box_node->outcnt(); i++) {
2358 Node* u = box_node->raw_out(i);
2359 if (u->is_AbstractLock()) {
2360 alock = u->as_AbstractLock();
2361 if (alock->box_node() == box_node) {
2362 // Verify that this Box is referenced only by related locks.
2363 assert(alock->obj_node()->eqv_uncast(obj), "");
2364 // Mark all related locks and unlocks.
2365 #ifdef ASSERT
2366 alock->log_lock_optimization(C, "eliminate_lock_set_nested");
2367 #endif
2368 alock->set_nested();
2369 }
2370 }
2371 }
2372 } else {
2373 #ifdef ASSERT
2374 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region");
2375 if (C->log() != nullptr)
2376 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output
2377 #endif
2378 }
2379 }
2380 return;
2381 }
2382 // Process locks for non escaping object
2383 assert(alock->is_non_esc_obj(), "");
2384 } // EliminateNestedLocks
2385
2386 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2387 // Look for all locks of this object and mark them and
2388 // corresponding BoxLock nodes as eliminated.
2389 Node* obj = alock->obj_node();
2390 for (uint j = 0; j < obj->outcnt(); j++) {
2391 Node* o = obj->raw_out(j);
2392 if (o->is_AbstractLock() &&
2393 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2394 alock = o->as_AbstractLock();
2395 Node* box = alock->box_node();
2396 // Replace old box node with new eliminated box for all users
2397 // of the same object and mark related locks as eliminated.
2398 mark_eliminated_box(box, obj);
2399 }
2400 }
2401 }
2402 }
2403
2404 // we have determined that this lock/unlock can be eliminated, we simply
2405 // eliminate the node without expanding it.
2406 //
2407 // Note: The membar's associated with the lock/unlock are currently not
2408 // eliminated. This should be investigated as a future enhancement.
2409 //
2410 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2411
2412 if (!alock->is_eliminated()) {
2413 return false;
2414 }
2415 #ifdef ASSERT
2416 if (!alock->is_coarsened()) {
2417 // Check that new "eliminated" BoxLock node is created.
2418 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2419 assert(oldbox->is_eliminated(), "should be done already");
2420 }
2421 #endif
2422
2423 alock->log_lock_optimization(C, "eliminate_lock");
2424
2425 #ifndef PRODUCT
2426 if (PrintEliminateLocks) {
2427 tty->print_cr("++++ Eliminated: %d %s '%s'", alock->_idx, (alock->is_Lock() ? "Lock" : "Unlock"), alock->kind_as_string());
2428 }
2429 #endif
2430
2431 Node* mem = alock->in(TypeFunc::Memory);
2432 Node* ctrl = alock->in(TypeFunc::Control);
2433 guarantee(ctrl != nullptr, "missing control projection, cannot replace_node() with null");
2434
2435 _callprojs = alock->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2436 // There are 2 projections from the lock. The lock node will
2437 // be deleted when its last use is subsumed below.
2438 assert(alock->outcnt() == 2 &&
2439 _callprojs->fallthrough_proj != nullptr &&
2440 _callprojs->fallthrough_memproj != nullptr,
2441 "Unexpected projections from Lock/Unlock");
2442
2443 Node* fallthroughproj = _callprojs->fallthrough_proj;
2444 Node* memproj_fallthrough = _callprojs->fallthrough_memproj;
2445
2446 // The memory projection from a lock/unlock is RawMem
2447 // The input to a Lock is merged memory, so extract its RawMem input
2448 // (unless the MergeMem has been optimized away.)
2449 if (alock->is_Lock()) {
2450 // Search for MemBarAcquireLock node and delete it also.
2451 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2452 assert(membar != nullptr && membar->Opcode() == Op_MemBarAcquireLock, "");
2453 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2454 Node* memproj = membar->proj_out(TypeFunc::Memory);
2455 _igvn.replace_node(ctrlproj, fallthroughproj);
2456 _igvn.replace_node(memproj, memproj_fallthrough);
2457
2458 // Delete FastLock node also if this Lock node is unique user
2459 // (a loop peeling may clone a Lock node).
2460 Node* flock = alock->as_Lock()->fastlock_node();
2461 if (flock->outcnt() == 1) {
2462 assert(flock->unique_out() == alock, "sanity");
2463 _igvn.replace_node(flock, top());
2464 }
2465 }
2466
2467 // Search for MemBarReleaseLock node and delete it also.
2468 if (alock->is_Unlock() && ctrl->is_Proj() && ctrl->in(0)->is_MemBar()) {
2469 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2470 assert(membar->Opcode() == Op_MemBarReleaseLock &&
2471 mem->is_Proj() && membar == mem->in(0), "");
2472 _igvn.replace_node(fallthroughproj, ctrl);
2473 _igvn.replace_node(memproj_fallthrough, mem);
2474 fallthroughproj = ctrl;
2475 memproj_fallthrough = mem;
2476 ctrl = membar->in(TypeFunc::Control);
2477 mem = membar->in(TypeFunc::Memory);
2478 }
2479
2480 _igvn.replace_node(fallthroughproj, ctrl);
2481 _igvn.replace_node(memproj_fallthrough, mem);
2482 return true;
2483 }
2484
2485
2486 //------------------------------expand_lock_node----------------------
2487 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
2488
2489 Node* ctrl = lock->in(TypeFunc::Control);
2490 Node* mem = lock->in(TypeFunc::Memory);
2491 Node* obj = lock->obj_node();
2492 Node* box = lock->box_node();
2493 Node* flock = lock->fastlock_node();
2494
2495 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2496
2497 // Make the merge point
2498 Node *region;
2499 Node *mem_phi;
2500 Node *slow_path;
2501
2502 region = new RegionNode(3);
2503 // create a Phi for the memory state
2504 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2505
2506 // Optimize test; set region slot 2
2507 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2508 mem_phi->init_req(2, mem);
2509
2510 // Make slow path call
2511 CallNode* call = make_slow_call(lock, OptoRuntime::complete_monitor_enter_Type(),
2512 OptoRuntime::complete_monitor_locking_Java(), nullptr, slow_path,
2513 obj, box, nullptr);
2514
2515 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2516
2517 // Slow path can only throw asynchronous exceptions, which are always
2518 // de-opted. So the compiler thinks the slow-call can never throw an
2519 // exception. If it DOES throw an exception we would need the debug
2520 // info removed first (since if it throws there is no monitor).
2521 assert(_callprojs->fallthrough_ioproj == nullptr && _callprojs->catchall_ioproj == nullptr &&
2522 _callprojs->catchall_memproj == nullptr && _callprojs->catchall_catchproj == nullptr, "Unexpected projection from Lock");
2523
2524 // Capture slow path
2525 // disconnect fall-through projection from call and create a new one
2526 // hook up users of fall-through projection to region
2527 Node *slow_ctrl = _callprojs->fallthrough_proj->clone();
2528 transform_later(slow_ctrl);
2529 _igvn.hash_delete(_callprojs->fallthrough_proj);
2530 _callprojs->fallthrough_proj->disconnect_inputs(C);
2531 region->init_req(1, slow_ctrl);
2532 // region inputs are now complete
2533 transform_later(region);
2534 _igvn.replace_node(_callprojs->fallthrough_proj, region);
2535
2536 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2537
2538 mem_phi->init_req(1, memproj);
2539
2540 transform_later(mem_phi);
2541
2542 _igvn.replace_node(_callprojs->fallthrough_memproj, mem_phi);
2543 }
2544
2545 //------------------------------expand_unlock_node----------------------
2546 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2547
2548 Node* ctrl = unlock->in(TypeFunc::Control);
2549 Node* mem = unlock->in(TypeFunc::Memory);
2550 Node* obj = unlock->obj_node();
2551 Node* box = unlock->box_node();
2552
2553 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2554
2555 // No need for a null check on unlock
2556
2557 // Make the merge point
2558 Node *region;
2559 Node *mem_phi;
2560
2561 region = new RegionNode(3);
2562 // create a Phi for the memory state
2563 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2564
2565 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2566 funlock = transform_later( funlock )->as_FastUnlock();
2567 // Optimize test; set region slot 2
2568 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2569 Node *thread = transform_later(new ThreadLocalNode());
2570
2571 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2572 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2573 "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2574
2575 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2576 assert(_callprojs->fallthrough_ioproj == nullptr && _callprojs->catchall_ioproj == nullptr &&
2577 _callprojs->catchall_memproj == nullptr && _callprojs->catchall_catchproj == nullptr, "Unexpected projection from Lock");
2578
2579 // No exceptions for unlocking
2580 // Capture slow path
2581 // disconnect fall-through projection from call and create a new one
2582 // hook up users of fall-through projection to region
2583 Node *slow_ctrl = _callprojs->fallthrough_proj->clone();
2584 transform_later(slow_ctrl);
2585 _igvn.hash_delete(_callprojs->fallthrough_proj);
2586 _callprojs->fallthrough_proj->disconnect_inputs(C);
2587 region->init_req(1, slow_ctrl);
2588 // region inputs are now complete
2589 transform_later(region);
2590 _igvn.replace_node(_callprojs->fallthrough_proj, region);
2591
2592 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2593 mem_phi->init_req(1, memproj );
2594 mem_phi->init_req(2, mem);
2595 transform_later(mem_phi);
2596
2597 _igvn.replace_node(_callprojs->fallthrough_memproj, mem_phi);
2598 }
2599
2600 // An inline type might be returned from the call but we don't know its
2601 // type. Either we get a buffered inline type (and nothing needs to be done)
2602 // or one of the values being returned is the klass of the inline type
2603 // and we need to allocate an inline type instance of that type and
2604 // initialize it with other values being returned. In that case, we
2605 // first try a fast path allocation and initialize the value with the
2606 // inline klass's pack handler or we fall back to a runtime call.
2607 void PhaseMacroExpand::expand_mh_intrinsic_return(CallStaticJavaNode* call) {
2608 assert(call->method()->is_method_handle_intrinsic(), "must be a method handle intrinsic call");
2609 Node* ret = call->proj_out_or_null(TypeFunc::Parms);
2610 if (ret == nullptr) {
2611 return;
2612 }
2613 const TypeFunc* tf = call->_tf;
2614 const TypeTuple* domain = OptoRuntime::store_inline_type_fields_Type()->domain_cc();
2615 const TypeFunc* new_tf = TypeFunc::make(tf->domain_sig(), tf->domain_cc(), tf->range_sig(), domain);
2616 call->_tf = new_tf;
2617 // Make sure the change of type is applied before projections are processed by igvn
2618 _igvn.set_type(call, call->Value(&_igvn));
2619 _igvn.set_type(ret, ret->Value(&_igvn));
2620
2621 // Before any new projection is added:
2622 CallProjections* projs = call->extract_projections(true, true);
2623
2624 // Create temporary hook nodes that will be replaced below.
2625 // Add an input to prevent hook nodes from being dead.
2626 Node* ctl = new Node(call);
2627 Node* mem = new Node(ctl);
2628 Node* io = new Node(ctl);
2629 Node* ex_ctl = new Node(ctl);
2630 Node* ex_mem = new Node(ctl);
2631 Node* ex_io = new Node(ctl);
2632 Node* res = new Node(ctl);
2633
2634 // Allocate a new buffered inline type only if a new one is not returned
2635 Node* cast = transform_later(new CastP2XNode(ctl, res));
2636 Node* mask = MakeConX(0x1);
2637 Node* masked = transform_later(new AndXNode(cast, mask));
2638 Node* cmp = transform_later(new CmpXNode(masked, mask));
2639 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
2640 IfNode* allocation_iff = new IfNode(ctl, bol, PROB_MAX, COUNT_UNKNOWN);
2641 transform_later(allocation_iff);
2642 Node* allocation_ctl = transform_later(new IfTrueNode(allocation_iff));
2643 Node* no_allocation_ctl = transform_later(new IfFalseNode(allocation_iff));
2644 Node* no_allocation_res = transform_later(new CheckCastPPNode(no_allocation_ctl, res, TypeInstPtr::BOTTOM));
2645
2646 // Try to allocate a new buffered inline instance either from TLAB or eden space
2647 Node* needgc_ctrl = nullptr; // needgc means slowcase, i.e. allocation failed
2648 CallLeafNoFPNode* handler_call;
2649 const bool alloc_in_place = UseTLAB;
2650 if (alloc_in_place) {
2651 Node* fast_oop_ctrl = nullptr;
2652 Node* fast_oop_rawmem = nullptr;
2653 Node* mask2 = MakeConX(-2);
2654 Node* masked2 = transform_later(new AndXNode(cast, mask2));
2655 Node* rawklassptr = transform_later(new CastX2PNode(masked2));
2656 Node* klass_node = transform_later(new CheckCastPPNode(allocation_ctl, rawklassptr, TypeInstKlassPtr::OBJECT_OR_NULL));
2657 Node* layout_val = make_load(nullptr, mem, klass_node, in_bytes(Klass::layout_helper_offset()), TypeInt::INT, T_INT);
2658 Node* size_in_bytes = ConvI2X(layout_val);
2659 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2660 Node* fast_oop = bs->obj_allocate(this, mem, allocation_ctl, size_in_bytes, io, needgc_ctrl,
2661 fast_oop_ctrl, fast_oop_rawmem,
2662 AllocateInstancePrefetchLines);
2663 // Allocation succeed, initialize buffered inline instance header firstly,
2664 // and then initialize its fields with an inline class specific handler
2665 Node* mark_node = makecon(TypeRawPtr::make((address)markWord::inline_type_prototype().value()));
2666 fast_oop_rawmem = make_store(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
2667 fast_oop_rawmem = make_store(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
2668 if (UseCompressedClassPointers) {
2669 fast_oop_rawmem = make_store(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::klass_gap_offset_in_bytes(), intcon(0), T_INT);
2670 }
2671 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);
2672 Node* pack_handler = make_load(fast_oop_ctrl, fast_oop_rawmem, fixed_block, in_bytes(InlineKlass::pack_handler_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
2673 handler_call = new CallLeafNoFPNode(OptoRuntime::pack_inline_type_Type(),
2674 nullptr,
2675 "pack handler",
2676 TypeRawPtr::BOTTOM);
2677 handler_call->init_req(TypeFunc::Control, fast_oop_ctrl);
2678 handler_call->init_req(TypeFunc::Memory, fast_oop_rawmem);
2679 handler_call->init_req(TypeFunc::I_O, top());
2680 handler_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2681 handler_call->init_req(TypeFunc::ReturnAdr, top());
2682 handler_call->init_req(TypeFunc::Parms, pack_handler);
2683 handler_call->init_req(TypeFunc::Parms+1, fast_oop);
2684 } else {
2685 needgc_ctrl = allocation_ctl;
2686 }
2687
2688 // Allocation failed, fall back to a runtime call
2689 CallStaticJavaNode* slow_call = new CallStaticJavaNode(OptoRuntime::store_inline_type_fields_Type(),
2690 StubRoutines::store_inline_type_fields_to_buf(),
2691 "store_inline_type_fields",
2692 TypePtr::BOTTOM);
2693 slow_call->init_req(TypeFunc::Control, needgc_ctrl);
2694 slow_call->init_req(TypeFunc::Memory, mem);
2695 slow_call->init_req(TypeFunc::I_O, io);
2696 slow_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2697 slow_call->init_req(TypeFunc::ReturnAdr, call->in(TypeFunc::ReturnAdr));
2698 slow_call->init_req(TypeFunc::Parms, res);
2699
2700 Node* slow_ctl = transform_later(new ProjNode(slow_call, TypeFunc::Control));
2701 Node* slow_mem = transform_later(new ProjNode(slow_call, TypeFunc::Memory));
2702 Node* slow_io = transform_later(new ProjNode(slow_call, TypeFunc::I_O));
2703 Node* slow_res = transform_later(new ProjNode(slow_call, TypeFunc::Parms));
2704 Node* slow_catc = transform_later(new CatchNode(slow_ctl, slow_io, 2));
2705 Node* slow_norm = transform_later(new CatchProjNode(slow_catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci));
2706 Node* slow_excp = transform_later(new CatchProjNode(slow_catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci));
2707
2708 Node* ex_r = new RegionNode(3);
2709 Node* ex_mem_phi = new PhiNode(ex_r, Type::MEMORY, TypePtr::BOTTOM);
2710 Node* ex_io_phi = new PhiNode(ex_r, Type::ABIO);
2711 ex_r->init_req(1, slow_excp);
2712 ex_mem_phi->init_req(1, slow_mem);
2713 ex_io_phi->init_req(1, slow_io);
2714 ex_r->init_req(2, ex_ctl);
2715 ex_mem_phi->init_req(2, ex_mem);
2716 ex_io_phi->init_req(2, ex_io);
2717 transform_later(ex_r);
2718 transform_later(ex_mem_phi);
2719 transform_later(ex_io_phi);
2720
2721 // We don't know how many values are returned. This assumes the
2722 // worst case, that all available registers are used.
2723 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) {
2724 if (domain->field_at(i) == Type::HALF) {
2725 slow_call->init_req(i, top());
2726 if (alloc_in_place) {
2727 handler_call->init_req(i+1, top());
2728 }
2729 continue;
2730 }
2731 Node* proj = transform_later(new ProjNode(call, i));
2732 slow_call->init_req(i, proj);
2733 if (alloc_in_place) {
2734 handler_call->init_req(i+1, proj);
2735 }
2736 }
2737 // We can safepoint at that new call
2738 slow_call->copy_call_debug_info(&_igvn, call);
2739 transform_later(slow_call);
2740 if (alloc_in_place) {
2741 transform_later(handler_call);
2742 }
2743
2744 Node* fast_ctl = nullptr;
2745 Node* fast_res = nullptr;
2746 MergeMemNode* fast_mem = nullptr;
2747 if (alloc_in_place) {
2748 fast_ctl = transform_later(new ProjNode(handler_call, TypeFunc::Control));
2749 Node* rawmem = transform_later(new ProjNode(handler_call, TypeFunc::Memory));
2750 fast_res = transform_later(new ProjNode(handler_call, TypeFunc::Parms));
2751 fast_mem = MergeMemNode::make(mem);
2752 fast_mem->set_memory_at(Compile::AliasIdxRaw, rawmem);
2753 transform_later(fast_mem);
2754 }
2755
2756 Node* r = new RegionNode(alloc_in_place ? 4 : 3);
2757 Node* mem_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM);
2758 Node* io_phi = new PhiNode(r, Type::ABIO);
2759 Node* res_phi = new PhiNode(r, TypeInstPtr::BOTTOM);
2760 r->init_req(1, no_allocation_ctl);
2761 mem_phi->init_req(1, mem);
2762 io_phi->init_req(1, io);
2763 res_phi->init_req(1, no_allocation_res);
2764 r->init_req(2, slow_norm);
2765 mem_phi->init_req(2, slow_mem);
2766 io_phi->init_req(2, slow_io);
2767 res_phi->init_req(2, slow_res);
2768 if (alloc_in_place) {
2769 r->init_req(3, fast_ctl);
2770 mem_phi->init_req(3, fast_mem);
2771 io_phi->init_req(3, io);
2772 res_phi->init_req(3, fast_res);
2773 }
2774 transform_later(r);
2775 transform_later(mem_phi);
2776 transform_later(io_phi);
2777 transform_later(res_phi);
2778
2779 // Do not let stores that initialize this buffer be reordered with a subsequent
2780 // store that would make this buffer accessible by other threads.
2781 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
2782 transform_later(mb);
2783 mb->init_req(TypeFunc::Memory, mem_phi);
2784 mb->init_req(TypeFunc::Control, r);
2785 r = new ProjNode(mb, TypeFunc::Control);
2786 transform_later(r);
2787 mem_phi = new ProjNode(mb, TypeFunc::Memory);
2788 transform_later(mem_phi);
2789
2790 assert(projs->nb_resproj == 1, "unexpected number of results");
2791 _igvn.replace_in_uses(projs->fallthrough_catchproj, r);
2792 _igvn.replace_in_uses(projs->fallthrough_memproj, mem_phi);
2793 _igvn.replace_in_uses(projs->fallthrough_ioproj, io_phi);
2794 _igvn.replace_in_uses(projs->resproj[0], res_phi);
2795 _igvn.replace_in_uses(projs->catchall_catchproj, ex_r);
2796 _igvn.replace_in_uses(projs->catchall_memproj, ex_mem_phi);
2797 _igvn.replace_in_uses(projs->catchall_ioproj, ex_io_phi);
2798 // The CatchNode should not use the ex_io_phi. Re-connect it to the catchall_ioproj.
2799 Node* cn = projs->fallthrough_catchproj->in(0);
2800 _igvn.replace_input_of(cn, 1, projs->catchall_ioproj);
2801
2802 _igvn.replace_node(ctl, projs->fallthrough_catchproj);
2803 _igvn.replace_node(mem, projs->fallthrough_memproj);
2804 _igvn.replace_node(io, projs->fallthrough_ioproj);
2805 _igvn.replace_node(res, projs->resproj[0]);
2806 _igvn.replace_node(ex_ctl, projs->catchall_catchproj);
2807 _igvn.replace_node(ex_mem, projs->catchall_memproj);
2808 _igvn.replace_node(ex_io, projs->catchall_ioproj);
2809 }
2810
2811 void PhaseMacroExpand::expand_subtypecheck_node(SubTypeCheckNode *check) {
2812 assert(check->in(SubTypeCheckNode::Control) == nullptr, "should be pinned");
2813 Node* bol = check->unique_out();
2814 Node* obj_or_subklass = check->in(SubTypeCheckNode::ObjOrSubKlass);
2815 Node* superklass = check->in(SubTypeCheckNode::SuperKlass);
2816 assert(bol->is_Bool() && bol->as_Bool()->_test._test == BoolTest::ne, "unexpected bool node");
2817
2818 for (DUIterator_Last imin, i = bol->last_outs(imin); i >= imin; --i) {
2819 Node* iff = bol->last_out(i);
2820 assert(iff->is_If(), "where's the if?");
2821
2822 if (iff->in(0)->is_top()) {
2823 _igvn.replace_input_of(iff, 1, C->top());
2824 continue;
2825 }
2826
2827 Node* iftrue = iff->as_If()->proj_out(1);
2828 Node* iffalse = iff->as_If()->proj_out(0);
2829 Node* ctrl = iff->in(0);
2830
2831 Node* subklass = nullptr;
2832 if (_igvn.type(obj_or_subklass)->isa_klassptr()) {
2833 subklass = obj_or_subklass;
2834 } else {
2835 Node* k_adr = basic_plus_adr(obj_or_subklass, oopDesc::klass_offset_in_bytes());
2836 subklass = _igvn.transform(LoadKlassNode::make(_igvn, C->immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
2837 }
2838
2839 Node* not_subtype_ctrl = Phase::gen_subtype_check(subklass, superklass, &ctrl, nullptr, _igvn, check->method(), check->bci());
2840
2841 _igvn.replace_input_of(iff, 0, C->top());
2842 _igvn.replace_node(iftrue, not_subtype_ctrl);
2843 _igvn.replace_node(iffalse, ctrl);
2844 }
2845 _igvn.replace_node(check, C->top());
2846 }
2847
2848 // FlatArrayCheckNode (array1 array2 ...) is expanded into:
2849 //
2850 // long mark = array1.mark | array2.mark | ...;
2851 // long locked_bit = markWord::unlocked_value & array1.mark & array2.mark & ...;
2852 // if (locked_bit == 0) {
2853 // // One array is locked, load prototype header from the klass
2854 // mark = array1.klass.proto | array2.klass.proto | ...
2855 // }
2856 // if ((mark & markWord::flat_array_bit_in_place) == 0) {
2857 // ...
2858 // }
2859 void PhaseMacroExpand::expand_flatarraycheck_node(FlatArrayCheckNode* check) {
2860 bool array_inputs = _igvn.type(check->in(FlatArrayCheckNode::ArrayOrKlass))->isa_oopptr() != nullptr;
2861 if (array_inputs) {
2862 Node* mark = MakeConX(0);
2863 Node* locked_bit = MakeConX(markWord::unlocked_value);
2864 Node* mem = check->in(FlatArrayCheckNode::Memory);
2865 for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
2866 Node* ary = check->in(i);
2867 const TypeOopPtr* t = _igvn.type(ary)->isa_oopptr();
2868 assert(t != nullptr, "Mixing array and klass inputs");
2869 assert(!t->is_flat() && !t->is_not_flat(), "Should have been optimized out");
2870 Node* mark_adr = basic_plus_adr(ary, oopDesc::mark_offset_in_bytes());
2871 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));
2872 mark = _igvn.transform(new OrXNode(mark, mark_load));
2873 locked_bit = _igvn.transform(new AndXNode(locked_bit, mark_load));
2874 }
2875 assert(!mark->is_Con(), "Should have been optimized out");
2876 Node* cmp = _igvn.transform(new CmpXNode(locked_bit, MakeConX(0)));
2877 Node* is_unlocked = _igvn.transform(new BoolNode(cmp, BoolTest::ne));
2878
2879 // BoolNode might be shared, replace each if user
2880 Node* old_bol = check->unique_out();
2881 assert(old_bol->is_Bool() && old_bol->as_Bool()->_test._test == BoolTest::ne, "unexpected condition");
2882 for (DUIterator_Last imin, i = old_bol->last_outs(imin); i >= imin; --i) {
2883 IfNode* old_iff = old_bol->last_out(i)->as_If();
2884 Node* ctrl = old_iff->in(0);
2885 RegionNode* region = new RegionNode(3);
2886 Node* mark_phi = new PhiNode(region, TypeX_X);
2887
2888 // Check if array is unlocked
2889 IfNode* iff = _igvn.transform(new IfNode(ctrl, is_unlocked, PROB_MAX, COUNT_UNKNOWN))->as_If();
2890
2891 // Unlocked: Use bits from mark word
2892 region->init_req(1, _igvn.transform(new IfTrueNode(iff)));
2893 mark_phi->init_req(1, mark);
2894
2895 // Locked: Load prototype header from klass
2896 ctrl = _igvn.transform(new IfFalseNode(iff));
2897 Node* proto = MakeConX(0);
2898 for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
2899 Node* ary = check->in(i);
2900 // Make loads control dependent to make sure they are only executed if array is locked
2901 Node* klass_adr = basic_plus_adr(ary, oopDesc::klass_offset_in_bytes());
2902 Node* klass = _igvn.transform(LoadKlassNode::make(_igvn, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
2903 Node* proto_adr = basic_plus_adr(klass, in_bytes(Klass::prototype_header_offset()));
2904 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));
2905 proto = _igvn.transform(new OrXNode(proto, proto_load));
2906 }
2907 region->init_req(2, ctrl);
2908 mark_phi->init_req(2, proto);
2909
2910 // Check if flat array bits are set
2911 Node* mask = MakeConX(markWord::flat_array_bit_in_place);
2912 Node* masked = _igvn.transform(new AndXNode(_igvn.transform(mark_phi), mask));
2913 cmp = _igvn.transform(new CmpXNode(masked, MakeConX(0)));
2914 Node* is_not_flat = _igvn.transform(new BoolNode(cmp, BoolTest::eq));
2915
2916 ctrl = _igvn.transform(region);
2917 iff = _igvn.transform(new IfNode(ctrl, is_not_flat, PROB_MAX, COUNT_UNKNOWN))->as_If();
2918 _igvn.replace_node(old_iff, iff);
2919 }
2920 _igvn.replace_node(check, C->top());
2921 } else {
2922 // Fall back to layout helper check
2923 Node* lhs = intcon(0);
2924 for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
2925 Node* array_or_klass = check->in(i);
2926 Node* klass = nullptr;
2927 const TypePtr* t = _igvn.type(array_or_klass)->is_ptr();
2928 assert(!t->is_flat() && !t->is_not_flat(), "Should have been optimized out");
2929 if (t->isa_oopptr() != nullptr) {
2930 Node* klass_adr = basic_plus_adr(array_or_klass, oopDesc::klass_offset_in_bytes());
2931 klass = transform_later(LoadKlassNode::make(_igvn, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
2932 } else {
2933 assert(t->isa_klassptr(), "Unexpected input type");
2934 klass = array_or_klass;
2935 }
2936 Node* lh_addr = basic_plus_adr(klass, in_bytes(Klass::layout_helper_offset()));
2937 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));
2938 lhs = _igvn.transform(new OrINode(lhs, lh_val));
2939 }
2940 Node* masked = transform_later(new AndINode(lhs, intcon(Klass::_lh_array_tag_flat_value_bit_inplace)));
2941 Node* cmp = transform_later(new CmpINode(masked, intcon(0)));
2942 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
2943 Node* m2b = transform_later(new Conv2BNode(masked));
2944 // The matcher expects the input to If nodes to be produced by a Bool(CmpI..)
2945 // pattern, but the input to other potential users (e.g. Phi) to be some
2946 // other pattern (e.g. a Conv2B node, possibly idealized as a CMoveI).
2947 Node* old_bol = check->unique_out();
2948 for (DUIterator_Last imin, i = old_bol->last_outs(imin); i >= imin; --i) {
2949 Node* user = old_bol->last_out(i);
2950 for (uint j = 0; j < user->req(); j++) {
2951 Node* n = user->in(j);
2952 if (n == old_bol) {
2953 _igvn.replace_input_of(user, j, user->is_If() ? bol : m2b);
2954 }
2955 }
2956 }
2957 _igvn.replace_node(check, C->top());
2958 }
2959 }
2960
2961 //---------------------------eliminate_macro_nodes----------------------
2962 // Eliminate scalar replaced allocations and associated locks.
2963 void PhaseMacroExpand::eliminate_macro_nodes() {
2964 if (C->macro_count() == 0) {
2965 return;
2966 }
2967 NOT_PRODUCT(int membar_before = count_MemBar(C);)
2968
2969 int iteration = 0;
2970 while (C->macro_count() > 0) {
2971 if (iteration++ > 100) {
2972 assert(false, "Too slow convergence of macro elimination");
2973 break;
2974 }
2975
2976 // Before elimination may re-mark (change to Nested or NonEscObj)
2977 // all associated (same box and obj) lock and unlock nodes.
2978 int cnt = C->macro_count();
2979 for (int i=0; i < cnt; i++) {
2980 Node *n = C->macro_node(i);
2981 if (n->is_AbstractLock()) { // Lock and Unlock nodes
2982 mark_eliminated_locking_nodes(n->as_AbstractLock());
2983 }
2984 }
2985 // Re-marking may break consistency of Coarsened locks.
2986 if (!C->coarsened_locks_consistent()) {
2987 return; // recompile without Coarsened locks if broken
2988 } else {
2989 // After coarsened locks are eliminated locking regions
2990 // become unbalanced. We should not execute any more
2991 // locks elimination optimizations on them.
2992 C->mark_unbalanced_boxes();
2993 }
2994
2995 bool progress = false;
2996 for (int i = C->macro_count(); i > 0; i = MIN2(i - 1, C->macro_count())) { // more than 1 element can be eliminated at once
2997 Node* n = C->macro_node(i - 1);
2998 bool success = false;
2999 DEBUG_ONLY(int old_macro_count = C->macro_count();)
3000 switch (n->class_id()) {
3001 case Node::Class_Allocate:
3002 case Node::Class_AllocateArray:
3003 success = eliminate_allocate_node(n->as_Allocate());
3004 #ifndef PRODUCT
3005 if (success && PrintOptoStatistics) {
3006 Atomic::inc(&PhaseMacroExpand::_objs_scalar_replaced_counter);
3007 }
3008 #endif
3009 break;
3010 case Node::Class_CallStaticJava: {
3011 CallStaticJavaNode* call = n->as_CallStaticJava();
3012 if (!call->method()->is_method_handle_intrinsic()) {
3013 success = eliminate_boxing_node(n->as_CallStaticJava());
3014 }
3015 break;
3016 }
3017 case Node::Class_Lock:
3018 case Node::Class_Unlock:
3019 success = eliminate_locking_node(n->as_AbstractLock());
3020 #ifndef PRODUCT
3021 if (success && PrintOptoStatistics) {
3022 Atomic::inc(&PhaseMacroExpand::_monitor_objects_removed_counter);
3023 }
3024 #endif
3025 break;
3026 case Node::Class_ArrayCopy:
3027 break;
3028 case Node::Class_OuterStripMinedLoop:
3029 break;
3030 case Node::Class_SubTypeCheck:
3031 break;
3032 case Node::Class_Opaque1:
3033 break;
3034 case Node::Class_FlatArrayCheck:
3035 break;
3036 default:
3037 assert(n->Opcode() == Op_LoopLimit ||
3038 n->Opcode() == Op_ModD ||
3039 n->Opcode() == Op_ModF ||
3040 n->is_OpaqueNotNull() ||
3041 n->is_OpaqueInitializedAssertionPredicate() ||
3042 n->Opcode() == Op_MaxL ||
3043 n->Opcode() == Op_MinL ||
3044 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n),
3045 "unknown node type in macro list");
3046 }
3047 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
3048 progress = progress || success;
3049 }
3050
3051 // Ensure the graph after PhaseMacroExpand::eliminate_macro_nodes is canonical (no igvn
3052 // transformation is pending). If an allocation is used only in safepoints, elimination of
3053 // other macro nodes can remove all these safepoints, allowing the allocation to be removed.
3054 // Hence after igvn we retry removing macro nodes if some progress that has been made in this
3055 // iteration.
3056 _igvn.set_delay_transform(false);
3057 _igvn.optimize();
3058 if (C->failing()) {
3059 return;
3060 }
3061 _igvn.set_delay_transform(true);
3062
3063 if (!progress) {
3064 break;
3065 }
3066 }
3067 #ifndef PRODUCT
3068 if (PrintOptoStatistics) {
3069 int membar_after = count_MemBar(C);
3070 Atomic::add(&PhaseMacroExpand::_memory_barriers_removed_counter, membar_before - membar_after);
3071 }
3072 #endif
3073 }
3074
3075 //------------------------------expand_macro_nodes----------------------
3076 // Returns true if a failure occurred.
3077 bool PhaseMacroExpand::expand_macro_nodes() {
3078 // Do not allow new macro nodes once we started to expand
3079 C->reset_allow_macro_nodes();
3080 if (StressMacroExpansion) {
3081 C->shuffle_macro_nodes();
3082 }
3083
3084 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
3085 bool progress = true;
3086 while (progress) {
3087 progress = false;
3088 for (int i = C->macro_count(); i > 0; i--) {
3089 Node* n = C->macro_node(i-1);
3090 bool success = false;
3091 DEBUG_ONLY(int old_macro_count = C->macro_count();)
3092 if (n->Opcode() == Op_LoopLimit) {
3093 // Remove it from macro list and put on IGVN worklist to optimize.
3094 C->remove_macro_node(n);
3095 _igvn._worklist.push(n);
3096 success = true;
3097 } else if (n->Opcode() == Op_CallStaticJava) {
3098 CallStaticJavaNode* call = n->as_CallStaticJava();
3099 if (!call->method()->is_method_handle_intrinsic()) {
3100 // Remove it from macro list and put on IGVN worklist to optimize.
3101 C->remove_macro_node(n);
3102 _igvn._worklist.push(n);
3103 success = true;
3104 }
3105 } else if (n->is_Opaque1()) {
3106 _igvn.replace_node(n, n->in(1));
3107 success = true;
3108 } else if (n->is_OpaqueNotNull()) {
3109 // Tests with OpaqueNotNull nodes are implicitly known to be true. Replace the node with true. In debug builds,
3110 // we leave the test in the graph to have an additional sanity check at runtime. If the test fails (i.e. a bug),
3111 // we will execute a Halt node.
3112 #ifdef ASSERT
3113 _igvn.replace_node(n, n->in(1));
3114 #else
3115 _igvn.replace_node(n, _igvn.intcon(1));
3116 #endif
3117 success = true;
3118 } else if (n->is_OpaqueInitializedAssertionPredicate()) {
3119 // Initialized Assertion Predicates must always evaluate to true. Therefore, we get rid of them in product
3120 // builds as they are useless. In debug builds we keep them as additional verification code. Even though
3121 // loop opts are already over, we want to keep Initialized Assertion Predicates alive as long as possible to
3122 // enable folding of dead control paths within which cast nodes become top after due to impossible types -
3123 // even after loop opts are over. Therefore, we delay the removal of these opaque nodes until now.
3124 #ifdef ASSERT
3125 _igvn.replace_node(n, n->in(1));
3126 #else
3127 _igvn.replace_node(n, _igvn.intcon(1));
3128 #endif // ASSERT
3129 } else if (n->Opcode() == Op_OuterStripMinedLoop) {
3130 n->as_OuterStripMinedLoop()->adjust_strip_mined_loop(&_igvn);
3131 C->remove_macro_node(n);
3132 success = true;
3133 } else if (n->Opcode() == Op_MaxL) {
3134 // Since MaxL and MinL are not implemented in the backend, we expand them to
3135 // a CMoveL construct now. At least until here, the type could be computed
3136 // precisely. CMoveL is not so smart, but we can give it at least the best
3137 // type we know abouot n now.
3138 Node* repl = MaxNode::signed_max(n->in(1), n->in(2), _igvn.type(n), _igvn);
3139 _igvn.replace_node(n, repl);
3140 success = true;
3141 } else if (n->Opcode() == Op_MinL) {
3142 Node* repl = MaxNode::signed_min(n->in(1), n->in(2), _igvn.type(n), _igvn);
3143 _igvn.replace_node(n, repl);
3144 success = true;
3145 }
3146 assert(!success || (C->macro_count() == (old_macro_count - 1)), "elimination must have deleted one node from macro list");
3147 progress = progress || success;
3148 if (success) {
3149 C->print_method(PHASE_AFTER_MACRO_EXPANSION_STEP, 5, n);
3150 }
3151 }
3152 }
3153
3154 // Clean up the graph so we're less likely to hit the maximum node
3155 // limit
3156 _igvn.set_delay_transform(false);
3157 _igvn.optimize();
3158 if (C->failing()) return true;
3159 _igvn.set_delay_transform(true);
3160
3161
3162 // Because we run IGVN after each expansion, some macro nodes may go
3163 // dead and be removed from the list as we iterate over it. Move
3164 // Allocate nodes (processed in a second pass) at the beginning of
3165 // the list and then iterate from the last element of the list until
3166 // an Allocate node is seen. This is robust to random deletion in
3167 // the list due to nodes going dead.
3168 C->sort_macro_nodes();
3169
3170 // expand arraycopy "macro" nodes first
3171 // For ReduceBulkZeroing, we must first process all arraycopy nodes
3172 // before the allocate nodes are expanded.
3173 while (C->macro_count() > 0) {
3174 int macro_count = C->macro_count();
3175 Node * n = C->macro_node(macro_count-1);
3176 assert(n->is_macro(), "only macro nodes expected here");
3177 if (_igvn.type(n) == Type::TOP || (n->in(0) != nullptr && n->in(0)->is_top())) {
3178 // node is unreachable, so don't try to expand it
3179 C->remove_macro_node(n);
3180 continue;
3181 }
3182 if (n->is_Allocate()) {
3183 break;
3184 }
3185 // Make sure expansion will not cause node limit to be exceeded.
3186 // Worst case is a macro node gets expanded into about 200 nodes.
3187 // Allow 50% more for optimization.
3188 if (C->check_node_count(300, "out of nodes before macro expansion")) {
3189 return true;
3190 }
3191
3192 DEBUG_ONLY(int old_macro_count = C->macro_count();)
3193 switch (n->class_id()) {
3194 case Node::Class_Lock:
3195 expand_lock_node(n->as_Lock());
3196 break;
3197 case Node::Class_Unlock:
3198 expand_unlock_node(n->as_Unlock());
3199 break;
3200 case Node::Class_ArrayCopy:
3201 expand_arraycopy_node(n->as_ArrayCopy());
3202 break;
3203 case Node::Class_SubTypeCheck:
3204 expand_subtypecheck_node(n->as_SubTypeCheck());
3205 break;
3206 case Node::Class_CallStaticJava:
3207 expand_mh_intrinsic_return(n->as_CallStaticJava());
3208 C->remove_macro_node(n);
3209 break;
3210 case Node::Class_FlatArrayCheck:
3211 expand_flatarraycheck_node(n->as_FlatArrayCheck());
3212 break;
3213 default:
3214 switch (n->Opcode()) {
3215 case Op_ModD:
3216 case Op_ModF: {
3217 bool is_drem = n->Opcode() == Op_ModD;
3218 CallNode* mod_macro = n->as_Call();
3219 CallNode* call = new CallLeafNode(mod_macro->tf(),
3220 is_drem ? CAST_FROM_FN_PTR(address, SharedRuntime::drem)
3221 : CAST_FROM_FN_PTR(address, SharedRuntime::frem),
3222 is_drem ? "drem" : "frem", TypeRawPtr::BOTTOM);
3223 call->init_req(TypeFunc::Control, mod_macro->in(TypeFunc::Control));
3224 call->init_req(TypeFunc::I_O, mod_macro->in(TypeFunc::I_O));
3225 call->init_req(TypeFunc::Memory, mod_macro->in(TypeFunc::Memory));
3226 call->init_req(TypeFunc::ReturnAdr, mod_macro->in(TypeFunc::ReturnAdr));
3227 call->init_req(TypeFunc::FramePtr, mod_macro->in(TypeFunc::FramePtr));
3228 for (unsigned int i = 0; i < mod_macro->tf()->domain_cc()->cnt() - TypeFunc::Parms; i++) {
3229 call->init_req(TypeFunc::Parms + i, mod_macro->in(TypeFunc::Parms + i));
3230 }
3231 _igvn.replace_node(mod_macro, call);
3232 transform_later(call);
3233 break;
3234 }
3235 default:
3236 assert(false, "unknown node type in macro list");
3237 }
3238 }
3239 assert(C->macro_count() == (old_macro_count - 1), "expansion must have deleted one node from macro list");
3240 if (C->failing()) return true;
3241 C->print_method(PHASE_AFTER_MACRO_EXPANSION_STEP, 5, n);
3242
3243 // Clean up the graph so we're less likely to hit the maximum node
3244 // limit
3245 _igvn.set_delay_transform(false);
3246 _igvn.optimize();
3247 if (C->failing()) return true;
3248 _igvn.set_delay_transform(true);
3249 }
3250
3251 // All nodes except Allocate nodes are expanded now. There could be
3252 // new optimization opportunities (such as folding newly created
3253 // load from a just allocated object). Run IGVN.
3254
3255 // expand "macro" nodes
3256 // nodes are removed from the macro list as they are processed
3257 while (C->macro_count() > 0) {
3258 int macro_count = C->macro_count();
3259 Node * n = C->macro_node(macro_count-1);
3260 assert(n->is_macro(), "only macro nodes expected here");
3261 if (_igvn.type(n) == Type::TOP || (n->in(0) != nullptr && n->in(0)->is_top())) {
3262 // node is unreachable, so don't try to expand it
3263 C->remove_macro_node(n);
3264 continue;
3265 }
3266 // Make sure expansion will not cause node limit to be exceeded.
3267 // Worst case is a macro node gets expanded into about 200 nodes.
3268 // Allow 50% more for optimization.
3269 if (C->check_node_count(300, "out of nodes before macro expansion")) {
3270 return true;
3271 }
3272 switch (n->class_id()) {
3273 case Node::Class_Allocate:
3274 expand_allocate(n->as_Allocate());
3275 break;
3276 case Node::Class_AllocateArray:
3277 expand_allocate_array(n->as_AllocateArray());
3278 break;
3279 default:
3280 assert(false, "unknown node type in macro list");
3281 }
3282 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
3283 if (C->failing()) return true;
3284 C->print_method(PHASE_AFTER_MACRO_EXPANSION_STEP, 5, n);
3285
3286 // Clean up the graph so we're less likely to hit the maximum node
3287 // limit
3288 _igvn.set_delay_transform(false);
3289 _igvn.optimize();
3290 if (C->failing()) return true;
3291 _igvn.set_delay_transform(true);
3292 }
3293
3294 _igvn.set_delay_transform(false);
3295 return false;
3296 }
3297
3298 #ifndef PRODUCT
3299 int PhaseMacroExpand::_objs_scalar_replaced_counter = 0;
3300 int PhaseMacroExpand::_monitor_objects_removed_counter = 0;
3301 int PhaseMacroExpand::_GC_barriers_removed_counter = 0;
3302 int PhaseMacroExpand::_memory_barriers_removed_counter = 0;
3303
3304 void PhaseMacroExpand::print_statistics() {
3305 tty->print("Objects scalar replaced = %d, ", Atomic::load(&_objs_scalar_replaced_counter));
3306 tty->print("Monitor objects removed = %d, ", Atomic::load(&_monitor_objects_removed_counter));
3307 tty->print("GC barriers removed = %d, ", Atomic::load(&_GC_barriers_removed_counter));
3308 tty->print_cr("Memory barriers removed = %d", Atomic::load(&_memory_barriers_removed_counter));
3309 }
3310
3311 int PhaseMacroExpand::count_MemBar(Compile *C) {
3312 if (!PrintOptoStatistics) {
3313 return 0;
3314 }
3315 Unique_Node_List ideal_nodes;
3316 int total = 0;
3317 ideal_nodes.map(C->live_nodes(), nullptr);
3318 ideal_nodes.push(C->root());
3319 for (uint next = 0; next < ideal_nodes.size(); ++next) {
3320 Node* n = ideal_nodes.at(next);
3321 if (n->is_MemBar()) {
3322 total++;
3323 }
3324 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3325 Node* m = n->fast_out(i);
3326 ideal_nodes.push(m);
3327 }
3328 }
3329 return total;
3330 }
3331 #endif