1 /*
2 * Copyright (c) 2005, 2018, 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 "precompiled.hpp"
26 #include "compiler/compileLog.hpp"
27 #include "gc/shared/collectedHeap.inline.hpp"
28 #include "libadt/vectset.hpp"
29 #include "memory/universe.hpp"
30 #include "opto/addnode.hpp"
31 #include "opto/arraycopynode.hpp"
32 #include "opto/callnode.hpp"
33 #include "opto/castnode.hpp"
34 #include "opto/cfgnode.hpp"
35 #include "opto/compile.hpp"
36 #include "opto/convertnode.hpp"
37 #include "opto/graphKit.hpp"
38 #include "opto/locknode.hpp"
39 #include "opto/loopnode.hpp"
40 #include "opto/macro.hpp"
41 #include "opto/memnode.hpp"
42 #include "opto/narrowptrnode.hpp"
43 #include "opto/node.hpp"
44 #include "opto/opaquenode.hpp"
45 #include "opto/phaseX.hpp"
46 #include "opto/rootnode.hpp"
47 #include "opto/runtime.hpp"
48 #include "opto/subnode.hpp"
49 #include "opto/type.hpp"
50 #include "opto/valuetypenode.hpp"
51 #include "runtime/sharedRuntime.hpp"
52 #include "utilities/macros.hpp"
53 #if INCLUDE_G1GC
54 #include "gc/g1/g1ThreadLocalData.hpp"
55 #endif // INCLUDE_G1GC
56 #if INCLUDE_SHENANDOAHGC
57 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
58 #endif
59
60
61 //
62 // Replace any references to "oldref" in inputs to "use" with "newref".
63 // Returns the number of replacements made.
64 //
65 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
66 int nreplacements = 0;
67 uint req = use->req();
68 for (uint j = 0; j < use->len(); j++) {
69 Node *uin = use->in(j);
70 if (uin == oldref) {
71 if (j < req)
72 use->set_req(j, newref);
73 else
74 use->set_prec(j, newref);
75 nreplacements++;
76 } else if (j >= req && uin == NULL) {
77 break;
78 }
79 }
80 return nreplacements;
81 }
82
83 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
84 Node* cmp;
85 if (mask != 0) {
86 Node* and_node = transform_later(new AndXNode(word, MakeConX(mask)));
87 cmp = transform_later(new CmpXNode(and_node, MakeConX(bits)));
88 } else {
89 cmp = word;
90 }
91 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne));
92 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
93 transform_later(iff);
94
95 // Fast path taken.
96 Node *fast_taken = transform_later(new IfFalseNode(iff));
97
98 // Fast path not-taken, i.e. slow path
99 Node *slow_taken = transform_later(new IfTrueNode(iff));
100
101 if (return_fast_path) {
102 region->init_req(edge, slow_taken); // Capture slow-control
103 return fast_taken;
104 } else {
105 region->init_req(edge, fast_taken); // Capture fast-control
106 return slow_taken;
107 }
108 }
109
110 //--------------------copy_predefined_input_for_runtime_call--------------------
111 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
112 // Set fixed predefined input arguments
113 call->init_req( TypeFunc::Control, ctrl );
114 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) );
115 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
116 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
117 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
118 }
119
120 //------------------------------make_slow_call---------------------------------
121 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type,
122 address slow_call, const char* leaf_name, Node* slow_path,
123 Node* parm0, Node* parm1, Node* parm2) {
124
125 // Slow-path call
126 CallNode *call = leaf_name
127 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
128 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM );
129
130 // Slow path call has no side-effects, uses few values
131 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
132 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0);
133 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1);
134 if (parm2 != NULL) call->init_req(TypeFunc::Parms+2, parm2);
135 call->copy_call_debug_info(&_igvn, oldcall);
136 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
137 _igvn.replace_node(oldcall, call);
138 transform_later(call);
139
140 return call;
141 }
142
143 void PhaseMacroExpand::extract_call_projections(CallNode *call) {
144 _fallthroughproj = NULL;
145 _fallthroughcatchproj = NULL;
146 _ioproj_fallthrough = NULL;
147 _ioproj_catchall = NULL;
148 _catchallcatchproj = NULL;
149 _memproj_fallthrough = NULL;
150 _memproj_catchall = NULL;
151 _resproj = NULL;
152 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) {
153 ProjNode *pn = call->fast_out(i)->as_Proj();
154 switch (pn->_con) {
155 case TypeFunc::Control:
156 {
157 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj
158 _fallthroughproj = pn;
159 DUIterator_Fast jmax, j = pn->fast_outs(jmax);
160 const Node *cn = pn->fast_out(j);
161 if (cn->is_Catch()) {
162 ProjNode *cpn = NULL;
163 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) {
164 cpn = cn->fast_out(k)->as_Proj();
165 assert(cpn->is_CatchProj(), "must be a CatchProjNode");
166 if (cpn->_con == CatchProjNode::fall_through_index)
167 _fallthroughcatchproj = cpn;
168 else {
169 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index.");
170 _catchallcatchproj = cpn;
171 }
172 }
173 }
174 break;
175 }
176 case TypeFunc::I_O:
177 if (pn->_is_io_use)
178 _ioproj_catchall = pn;
179 else
180 _ioproj_fallthrough = pn;
181 break;
182 case TypeFunc::Memory:
183 if (pn->_is_io_use)
184 _memproj_catchall = pn;
185 else
186 _memproj_fallthrough = pn;
187 break;
188 case TypeFunc::Parms:
189 _resproj = pn;
190 break;
191 default:
192 assert(false, "unexpected projection from allocation node.");
193 }
194 }
195
196 }
197
198 void PhaseMacroExpand::eliminate_gc_barrier(Node* p2x) {
199 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
200 bs->eliminate_gc_barrier(this, p2x);
201 }
202
203 // Search for a memory operation for the specified memory slice.
204 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
205 Node *orig_mem = mem;
206 Node *alloc_mem = alloc->in(TypeFunc::Memory);
207 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
208 while (true) {
209 if (mem == alloc_mem || mem == start_mem ) {
210 return mem; // hit one of our sentinels
211 } else if (mem->is_MergeMem()) {
212 mem = mem->as_MergeMem()->memory_at(alias_idx);
213 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
214 Node *in = mem->in(0);
215 // we can safely skip over safepoints, calls, locks and membars because we
216 // already know that the object is safe to eliminate.
217 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
218 return in;
219 } else if (in->is_Call()) {
220 CallNode *call = in->as_Call();
221 if (call->may_modify(tinst, phase)) {
222 assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape");
223 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) {
224 return in;
225 }
226 }
227 mem = in->in(TypeFunc::Memory);
228 } else if (in->is_MemBar()) {
229 ArrayCopyNode* ac = NULL;
230 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) {
231 assert(ac != NULL && ac->is_clonebasic(), "Only basic clone is a non escaping clone");
232 return ac;
233 }
234 mem = in->in(TypeFunc::Memory);
235 } else {
236 assert(false, "unexpected projection");
237 }
238 } else if (mem->is_Store()) {
239 const TypePtr* atype = mem->as_Store()->adr_type();
240 int adr_idx = phase->C->get_alias_index(atype);
241 if (adr_idx == alias_idx) {
242 assert(atype->isa_oopptr(), "address type must be oopptr");
243 int adr_offset = atype->flattened_offset();
244 uint adr_iid = atype->is_oopptr()->instance_id();
245 // Array elements references have the same alias_idx
246 // but different offset and different instance_id.
247 if (adr_offset == offset && adr_iid == alloc->_idx)
248 return mem;
249 } else {
250 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
251 }
252 mem = mem->in(MemNode::Memory);
253 } else if (mem->is_ClearArray()) {
254 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
255 // Can not bypass initialization of the instance
256 // we are looking.
257 debug_only(intptr_t offset;)
258 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
259 InitializeNode* init = alloc->as_Allocate()->initialization();
260 // We are looking for stored value, return Initialize node
261 // or memory edge from Allocate node.
262 if (init != NULL)
263 return init;
264 else
265 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers).
266 }
267 // Otherwise skip it (the call updated 'mem' value).
268 } else if (mem->Opcode() == Op_SCMemProj) {
269 mem = mem->in(0);
270 Node* adr = NULL;
271 if (mem->is_LoadStore()) {
272 adr = mem->in(MemNode::Address);
273 } else {
274 assert(mem->Opcode() == Op_EncodeISOArray ||
275 mem->Opcode() == Op_StrCompressedCopy, "sanity");
276 adr = mem->in(3); // Destination array
277 }
278 const TypePtr* atype = adr->bottom_type()->is_ptr();
279 int adr_idx = phase->C->get_alias_index(atype);
280 if (adr_idx == alias_idx) {
281 DEBUG_ONLY(mem->dump();)
282 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
283 return NULL;
284 }
285 mem = mem->in(MemNode::Memory);
286 } else if (mem->Opcode() == Op_StrInflatedCopy) {
287 Node* adr = mem->in(3); // Destination array
288 const TypePtr* atype = adr->bottom_type()->is_ptr();
289 int adr_idx = phase->C->get_alias_index(atype);
290 if (adr_idx == alias_idx) {
291 DEBUG_ONLY(mem->dump();)
292 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
293 return NULL;
294 }
295 mem = mem->in(MemNode::Memory);
296 } else {
297 return mem;
298 }
299 assert(mem != orig_mem, "dead memory loop");
300 }
301 }
302
303 // Generate loads from source of the arraycopy for fields of
304 // destination needed at a deoptimization point
305 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) {
306 BasicType bt = ft;
307 const Type *type = ftype;
308 if (ft == T_NARROWOOP) {
309 bt = T_OBJECT;
310 type = ftype->make_oopptr();
311 }
312 Node* res = NULL;
313 if (ac->is_clonebasic()) {
314 assert(ac->in(ArrayCopyNode::Src) != ac->in(ArrayCopyNode::Dest), "clone source equals destination");
315 Node* base = ac->in(ArrayCopyNode::Src)->in(AddPNode::Base);
316 Node* adr = _igvn.transform(new AddPNode(base, base, MakeConX(offset)));
317 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
318 res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::UnknownControl);
319 } else {
320 if (ac->modifies(offset, offset, &_igvn, true)) {
321 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
322 uint shift = exact_log2(type2aelembytes(bt));
323 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
324 #ifdef _LP64
325 diff = _igvn.transform(new ConvI2LNode(diff));
326 #endif
327 diff = _igvn.transform(new LShiftXNode(diff, intcon(shift)));
328
329 Node* off = _igvn.transform(new AddXNode(MakeConX(offset), diff));
330 Node* base = ac->in(ArrayCopyNode::Src);
331 Node* adr = _igvn.transform(new AddPNode(base, base, off));
332 const TypePtr* adr_type = _igvn.type(base)->is_ptr();
333 if (adr_type->isa_aryptr()) {
334 // In the case of a flattened value type array, each field has its
335 // own slice so we need to extract the field being accessed from
336 // the address computation
337 adr_type = adr_type->is_aryptr()->add_field_offset_and_offset(offset);
338 } else {
339 adr_type = adr_type->add_offset(offset);
340 }
341 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
342 // Don't emit a new load from src if src == dst but try to get the value from memory instead
343 return value_from_mem(ac->in(TypeFunc::Memory), ctl, ft, ftype, adr_type->isa_oopptr(), alloc);
344 }
345 if (adr_type->isa_aryptr()) {
346 adr = _igvn.transform(new CastPPNode(adr, adr_type));
347 }
348 res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::UnknownControl);
349 }
350 }
351 if (res != NULL) {
352 res = _igvn.transform(res);
353 if (ftype->isa_narrowoop()) {
354 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
355 assert(res->isa_DecodeN(), "should be narrow oop");
356 res = _igvn.transform(new EncodePNode(res, ftype));
357 }
358 return res;
359 }
360 return NULL;
361 }
362
363 //
364 // Given a Memory Phi, compute a value Phi containing the values from stores
365 // on the input paths.
366 // Note: this function is recursive, its depth is limited by the "level" argument
367 // Returns the computed Phi, or NULL if it cannot compute it.
368 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) {
369 assert(mem->is_Phi(), "sanity");
370 int alias_idx = C->get_alias_index(adr_t);
371 int offset = adr_t->flattened_offset();
372 int instance_id = adr_t->instance_id();
373
374 // Check if an appropriate value phi already exists.
375 Node* region = mem->in(0);
376 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
377 Node* phi = region->fast_out(k);
378 if (phi->is_Phi() && phi != mem &&
379 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) {
380 return phi;
381 }
382 }
383 // Check if an appropriate new value phi already exists.
384 Node* new_phi = value_phis->find(mem->_idx);
385 if (new_phi != NULL)
386 return new_phi;
387
388 if (level <= 0) {
389 return NULL; // Give up: phi tree too deep
390 }
391 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
392 Node *alloc_mem = alloc->in(TypeFunc::Memory);
393
394 uint length = mem->req();
395 GrowableArray <Node *> values(length, length, NULL, false);
396
397 // create a new Phi for the value
398 PhiNode *phi = new PhiNode(mem->in(0), phi_type, NULL, mem->_idx, instance_id, alias_idx, offset);
399 transform_later(phi);
400 value_phis->push(phi, mem->_idx);
401
402 for (uint j = 1; j < length; j++) {
403 Node *in = mem->in(j);
404 if (in == NULL || in->is_top()) {
405 values.at_put(j, in);
406 } else {
407 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
408 if (val == start_mem || val == alloc_mem) {
409 // hit a sentinel, return appropriate 0 value
410 Node* default_value = alloc->in(AllocateNode::DefaultValue);
411 if (default_value != NULL) {
412 values.at_put(j, default_value);
413 } else {
414 assert(alloc->in(AllocateNode::RawDefaultValue) == NULL, "default value may not be null");
415 values.at_put(j, _igvn.zerocon(ft));
416 }
417 continue;
418 }
419 if (val->is_Initialize()) {
420 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
421 }
422 if (val == NULL) {
423 return NULL; // can't find a value on this path
424 }
425 if (val == mem) {
426 values.at_put(j, mem);
427 } else if (val->is_Store()) {
428 Node* n = val->in(MemNode::ValueIn);
429 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
430 n = bs->step_over_gc_barrier(n);
431 values.at_put(j, n);
432 } else if(val->is_Proj() && val->in(0) == alloc) {
433 Node* default_value = alloc->in(AllocateNode::DefaultValue);
434 if (default_value != NULL) {
435 values.at_put(j, default_value);
436 } else {
437 assert(alloc->in(AllocateNode::RawDefaultValue) == NULL, "default value may not be null");
438 values.at_put(j, _igvn.zerocon(ft));
439 }
440 } else if (val->is_Phi()) {
441 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
442 if (val == NULL) {
443 return NULL;
444 }
445 values.at_put(j, val);
446 } else if (val->Opcode() == Op_SCMemProj) {
447 assert(val->in(0)->is_LoadStore() ||
448 val->in(0)->Opcode() == Op_EncodeISOArray ||
449 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity");
450 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
451 return NULL;
452 } else if (val->is_ArrayCopy()) {
453 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc);
454 if (res == NULL) {
455 return NULL;
456 }
457 values.at_put(j, res);
458 } else {
459 #ifdef ASSERT
460 val->dump();
461 assert(false, "unknown node on this path");
462 #endif
463 return NULL; // unknown node on this path
464 }
465 }
466 }
467 // Set Phi's inputs
468 for (uint j = 1; j < length; j++) {
469 if (values.at(j) == mem) {
470 phi->init_req(j, phi);
471 } else {
472 phi->init_req(j, values.at(j));
473 }
474 }
475 return phi;
476 }
477
478 // Search the last value stored into the object's field.
479 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) {
480 assert(adr_t->is_known_instance_field(), "instance required");
481 int instance_id = adr_t->instance_id();
482 assert((uint)instance_id == alloc->_idx, "wrong allocation");
483
484 int alias_idx = C->get_alias_index(adr_t);
485 int offset = adr_t->flattened_offset();
486 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
487 Node *alloc_mem = alloc->in(TypeFunc::Memory);
488 Arena *a = Thread::current()->resource_area();
489 VectorSet visited(a);
490
491 bool done = sfpt_mem == alloc_mem;
492 Node *mem = sfpt_mem;
493 while (!done) {
494 if (visited.test_set(mem->_idx)) {
495 return NULL; // found a loop, give up
496 }
497 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
498 if (mem == start_mem || mem == alloc_mem) {
499 done = true; // hit a sentinel, return appropriate 0 value
500 } else if (mem->is_Initialize()) {
501 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
502 if (mem == NULL) {
503 done = true; // Something went wrong.
504 } else if (mem->is_Store()) {
505 const TypePtr* atype = mem->as_Store()->adr_type();
506 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
507 done = true;
508 }
509 } else if (mem->is_Store()) {
510 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
511 assert(atype != NULL, "address type must be oopptr");
512 assert(C->get_alias_index(atype) == alias_idx &&
513 atype->is_known_instance_field() && atype->flattened_offset() == offset &&
514 atype->instance_id() == instance_id, "store is correct memory slice");
515 done = true;
516 } else if (mem->is_Phi()) {
517 // try to find a phi's unique input
518 Node *unique_input = NULL;
519 Node *top = C->top();
520 for (uint i = 1; i < mem->req(); i++) {
521 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
522 if (n == NULL || n == top || n == mem) {
523 continue;
524 } else if (unique_input == NULL) {
525 unique_input = n;
526 } else if (unique_input != n) {
527 unique_input = top;
528 break;
529 }
530 }
531 if (unique_input != NULL && unique_input != top) {
532 mem = unique_input;
533 } else {
534 done = true;
535 }
536 } else if (mem->is_ArrayCopy()) {
537 done = true;
538 } else {
539 assert(false, "unexpected node");
540 }
541 }
542 if (mem != NULL) {
543 if (mem == start_mem || mem == alloc_mem) {
544 // hit a sentinel, return appropriate 0 value
545 Node* default_value = alloc->in(AllocateNode::DefaultValue);
546 if (default_value != NULL) {
547 return default_value;
548 }
549 assert(alloc->in(AllocateNode::RawDefaultValue) == NULL, "default value may not be null");
550 return _igvn.zerocon(ft);
551 } else if (mem->is_Store()) {
552 Node* n = mem->in(MemNode::ValueIn);
553 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
554 n = bs->step_over_gc_barrier(n);
555 return n;
556 } else if (mem->is_Phi()) {
557 // attempt to produce a Phi reflecting the values on the input paths of the Phi
558 Node_Stack value_phis(a, 8);
559 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
560 if (phi != NULL) {
561 return phi;
562 } else {
563 // Kill all new Phis
564 while(value_phis.is_nonempty()) {
565 Node* n = value_phis.node();
566 _igvn.replace_node(n, C->top());
567 value_phis.pop();
568 }
569 }
570 } else if (mem->is_ArrayCopy()) {
571 Node* ctl = mem->in(0);
572 Node* m = mem->in(TypeFunc::Memory);
573 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) {
574 // pin the loads in the uncommon trap path
575 ctl = sfpt_ctl;
576 m = sfpt_mem;
577 }
578 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc);
579 }
580 }
581 // Something went wrong.
582 return NULL;
583 }
584
585 // Search the last value stored into the value type's fields.
586 Node* PhaseMacroExpand::value_type_from_mem(Node* mem, Node* ctl, ciValueKlass* vk, const TypeAryPtr* adr_type, int offset, AllocateNode* alloc) {
587 // Subtract the offset of the first field to account for the missing oop header
588 offset -= vk->first_field_offset();
589 // Create a new ValueTypeNode and retrieve the field values from memory
590 ValueTypeNode* vt = ValueTypeNode::make_uninitialized(_igvn, vk)->as_ValueType();
591 for (int i = 0; i < vk->nof_declared_nonstatic_fields(); ++i) {
592 ciType* field_type = vt->field_type(i);
593 int field_offset = offset + vt->field_offset(i);
594 // Each value type field has its own memory slice
595 adr_type = adr_type->with_field_offset(field_offset);
596 Node* value = NULL;
597 if (vt->field_is_flattened(i)) {
598 value = value_type_from_mem(mem, ctl, field_type->as_value_klass(), adr_type, field_offset, alloc);
599 } else {
600 const Type* ft = Type::get_const_type(field_type);
601 BasicType bt = field_type->basic_type();
602 if (UseCompressedOops && !is_java_primitive(bt)) {
603 ft = ft->make_narrowoop();
604 bt = T_NARROWOOP;
605 }
606 value = value_from_mem(mem, ctl, bt, ft, adr_type, alloc);
607 if (value != NULL && ft->isa_narrowoop()) {
608 assert(UseCompressedOops, "unexpected narrow oop");
609 value = transform_later(new DecodeNNode(value, value->get_ptr_type()));
610 }
611 }
612 if (value != NULL) {
613 vt->set_field_value(i, value);
614 } else {
615 // We might have reached the TrackedInitializationLimit
616 return NULL;
617 }
618 }
619 return transform_later(vt);
620 }
621
622 // Check the possibility of scalar replacement.
623 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
624 // Scan the uses of the allocation to check for anything that would
625 // prevent us from eliminating it.
626 NOT_PRODUCT( const char* fail_eliminate = NULL; )
627 DEBUG_ONLY( Node* disq_node = NULL; )
628 bool can_eliminate = true;
629
630 Node* res = alloc->result_cast();
631 const TypeOopPtr* res_type = NULL;
632 if (res == NULL) {
633 // All users were eliminated.
634 } else if (!res->is_CheckCastPP()) {
635 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
636 can_eliminate = false;
637 } else {
638 res_type = _igvn.type(res)->isa_oopptr();
639 if (res_type == NULL) {
640 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
641 can_eliminate = false;
642 } else if (res_type->isa_aryptr()) {
643 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
644 if (length < 0) {
645 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
646 can_eliminate = false;
647 }
648 }
649 }
650
651 if (can_eliminate && res != NULL) {
652 for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
653 j < jmax && can_eliminate; j++) {
654 Node* use = res->fast_out(j);
655
656 if (use->is_AddP()) {
657 const TypePtr* addp_type = _igvn.type(use)->is_ptr();
658 int offset = addp_type->offset();
659
660 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
661 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";)
662 can_eliminate = false;
663 break;
664 }
665 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
666 k < kmax && can_eliminate; k++) {
667 Node* n = use->fast_out(k);
668 if (!n->is_Store() && n->Opcode() != Op_CastP2X &&
669 SHENANDOAHGC_ONLY((!UseShenandoahGC || !ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(n)) &&)
670 !(n->is_ArrayCopy() &&
671 n->as_ArrayCopy()->is_clonebasic() &&
672 n->in(ArrayCopyNode::Dest) == use)) {
673 DEBUG_ONLY(disq_node = n;)
674 if (n->is_Load() || n->is_LoadStore()) {
675 NOT_PRODUCT(fail_eliminate = "Field load";)
676 } else {
677 NOT_PRODUCT(fail_eliminate = "Not store field reference";)
678 }
679 can_eliminate = false;
680 }
681 }
682 } else if (use->is_ArrayCopy() &&
683 (use->as_ArrayCopy()->is_arraycopy_validated() ||
684 use->as_ArrayCopy()->is_copyof_validated() ||
685 use->as_ArrayCopy()->is_copyofrange_validated()) &&
686 use->in(ArrayCopyNode::Dest) == res) {
687 // ok to eliminate
688 } else if (use->is_SafePoint()) {
689 SafePointNode* sfpt = use->as_SafePoint();
690 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
691 // Object is passed as argument.
692 DEBUG_ONLY(disq_node = use;)
693 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
694 can_eliminate = false;
695 }
696 Node* sfptMem = sfpt->memory();
697 if (sfptMem == NULL || sfptMem->is_top()) {
698 DEBUG_ONLY(disq_node = use;)
699 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";)
700 can_eliminate = false;
701 } else {
702 safepoints.append_if_missing(sfpt);
703 }
704 } else if (use->is_ValueType() && use->isa_ValueType()->get_oop() == res) {
705 // ok to eliminate
706 } else if (use->is_Store()) {
707 // store to mark work
708 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
709 if (use->is_Phi()) {
710 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
711 NOT_PRODUCT(fail_eliminate = "Object is return value";)
712 } else {
713 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
714 }
715 DEBUG_ONLY(disq_node = use;)
716 } else {
717 if (use->Opcode() == Op_Return) {
718 NOT_PRODUCT(fail_eliminate = "Object is return value";)
719 } else {
720 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
721 }
722 DEBUG_ONLY(disq_node = use;)
723 }
724 can_eliminate = false;
725 } else {
726 assert(use->Opcode() == Op_CastP2X, "should be");
727 assert(!use->has_out_with(Op_OrL), "should have been removed because oop is never null");
728 }
729 }
730 }
731
732 #ifndef PRODUCT
733 if (PrintEliminateAllocations) {
734 if (can_eliminate) {
735 tty->print("Scalar ");
736 if (res == NULL)
737 alloc->dump();
738 else
739 res->dump();
740 } else if (alloc->_is_scalar_replaceable) {
741 tty->print("NotScalar (%s)", fail_eliminate);
742 if (res == NULL)
743 alloc->dump();
744 else
745 res->dump();
746 #ifdef ASSERT
747 if (disq_node != NULL) {
748 tty->print(" >>>> ");
749 disq_node->dump();
750 }
751 #endif /*ASSERT*/
752 }
753 }
754 #endif
755 return can_eliminate;
756 }
757
758 // Do scalar replacement.
759 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
760 GrowableArray <SafePointNode *> safepoints_done;
761
762 ciKlass* klass = NULL;
763 ciInstanceKlass* iklass = NULL;
764 int nfields = 0;
765 int array_base = 0;
766 int element_size = 0;
767 BasicType basic_elem_type = T_ILLEGAL;
768 ciType* elem_type = NULL;
769
770 Node* res = alloc->result_cast();
771 assert(res == NULL || res->is_CheckCastPP(), "unexpected AllocateNode result");
772 const TypeOopPtr* res_type = NULL;
773 if (res != NULL) { // Could be NULL when there are no users
774 res_type = _igvn.type(res)->isa_oopptr();
775 }
776
777 if (res != NULL) {
778 klass = res_type->klass();
779 if (res_type->isa_instptr()) {
780 // find the fields of the class which will be needed for safepoint debug information
781 assert(klass->is_instance_klass(), "must be an instance klass.");
782 iklass = klass->as_instance_klass();
783 nfields = iklass->nof_nonstatic_fields();
784 } else {
785 // find the array's elements which will be needed for safepoint debug information
786 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
787 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass.");
788 elem_type = klass->as_array_klass()->element_type();
789 basic_elem_type = elem_type->basic_type();
790 if (elem_type->is_valuetype() && !klass->is_value_array_klass()) {
791 assert(basic_elem_type == T_VALUETYPE, "unexpected element basic type");
792 basic_elem_type = T_OBJECT;
793 }
794 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
795 element_size = type2aelembytes(basic_elem_type);
796 if (klass->is_value_array_klass()) {
797 // Flattened value type array
798 element_size = klass->as_value_array_klass()->element_byte_size();
799 }
800 }
801 }
802 //
803 // Process the safepoint uses
804 //
805 Unique_Node_List value_worklist;
806 while (safepoints.length() > 0) {
807 SafePointNode* sfpt = safepoints.pop();
808 Node* mem = sfpt->memory();
809 Node* ctl = sfpt->control();
810 assert(sfpt->jvms() != NULL, "missed JVMS");
811 // Fields of scalar objs are referenced only at the end
812 // of regular debuginfo at the last (youngest) JVMS.
813 // Record relative start index.
814 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
815 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type,
816 #ifdef ASSERT
817 alloc,
818 #endif
819 first_ind, nfields);
820 sobj->init_req(0, C->root());
821 transform_later(sobj);
822
823 // Scan object's fields adding an input to the safepoint for each field.
824 for (int j = 0; j < nfields; j++) {
825 intptr_t offset;
826 ciField* field = NULL;
827 if (iklass != NULL) {
828 field = iklass->nonstatic_field_at(j);
829 offset = field->offset();
830 elem_type = field->type();
831 basic_elem_type = field->layout_type();
832 assert(!field->is_flattened(), "flattened value type fields should not have safepoint uses");
833 } else {
834 offset = array_base + j * (intptr_t)element_size;
835 }
836
837 const Type *field_type;
838 // The next code is taken from Parse::do_get_xxx().
839 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) {
840 if (!elem_type->is_loaded()) {
841 field_type = TypeInstPtr::BOTTOM;
842 } else if (field != NULL && field->is_static_constant()) {
843 // This can happen if the constant oop is non-perm.
844 ciObject* con = field->constant_value().as_object();
845 // Do not "join" in the previous type; it doesn't add value,
846 // and may yield a vacuous result if the field is of interface type.
847 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
848 assert(field_type != NULL, "field singleton type must be consistent");
849 } else {
850 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
851 }
852 if (UseCompressedOops) {
853 field_type = field_type->make_narrowoop();
854 basic_elem_type = T_NARROWOOP;
855 }
856 } else {
857 field_type = Type::get_const_basic_type(basic_elem_type);
858 }
859
860 Node* field_val = NULL;
861 const TypeOopPtr* field_addr_type = res_type->add_offset(offset)->isa_oopptr();
862 if (klass->is_value_array_klass()) {
863 ciValueKlass* vk = elem_type->as_value_klass();
864 assert(vk->flatten_array(), "must be flattened");
865 field_val = value_type_from_mem(mem, ctl, vk, field_addr_type->isa_aryptr(), 0, alloc);
866 } else {
867 field_val = value_from_mem(mem, ctl, basic_elem_type, field_type, field_addr_type, alloc);
868 }
869 if (field_val == NULL) {
870 // We weren't able to find a value for this field,
871 // give up on eliminating this allocation.
872
873 // Remove any extra entries we added to the safepoint.
874 uint last = sfpt->req() - 1;
875 for (int k = 0; k < j; k++) {
876 sfpt->del_req(last--);
877 }
878 _igvn._worklist.push(sfpt);
879 // rollback processed safepoints
880 while (safepoints_done.length() > 0) {
881 SafePointNode* sfpt_done = safepoints_done.pop();
882 // remove any extra entries we added to the safepoint
883 last = sfpt_done->req() - 1;
884 for (int k = 0; k < nfields; k++) {
885 sfpt_done->del_req(last--);
886 }
887 JVMState *jvms = sfpt_done->jvms();
888 jvms->set_endoff(sfpt_done->req());
889 // Now make a pass over the debug information replacing any references
890 // to SafePointScalarObjectNode with the allocated object.
891 int start = jvms->debug_start();
892 int end = jvms->debug_end();
893 for (int i = start; i < end; i++) {
894 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
895 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
896 if (scobj->first_index(jvms) == sfpt_done->req() &&
897 scobj->n_fields() == (uint)nfields) {
898 assert(scobj->alloc() == alloc, "sanity");
899 sfpt_done->set_req(i, res);
900 }
901 }
902 }
903 _igvn._worklist.push(sfpt_done);
904 }
905 #ifndef PRODUCT
906 if (PrintEliminateAllocations) {
907 if (field != NULL) {
908 tty->print("=== At SafePoint node %d can't find value of Field: ",
909 sfpt->_idx);
910 field->print();
911 int field_idx = C->get_alias_index(field_addr_type);
912 tty->print(" (alias_idx=%d)", field_idx);
913 } else { // Array's element
914 tty->print("=== At SafePoint node %d can't find value of array element [%d]",
915 sfpt->_idx, j);
916 }
917 tty->print(", which prevents elimination of: ");
918 if (res == NULL)
919 alloc->dump();
920 else
921 res->dump();
922 }
923 #endif
924 return false;
925 }
926 if (UseCompressedOops && field_type->isa_narrowoop()) {
927 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
928 // to be able scalar replace the allocation.
929 if (field_val->is_EncodeP()) {
930 field_val = field_val->in(1);
931 } else {
932 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
933 }
934 } else if (field_val->is_ValueType()) {
935 // Keep track of value types to scalarize them later
936 value_worklist.push(field_val);
937 }
938 sfpt->add_req(field_val);
939 }
940 JVMState *jvms = sfpt->jvms();
941 jvms->set_endoff(sfpt->req());
942 // Now make a pass over the debug information replacing any references
943 // to the allocated object with "sobj"
944 int start = jvms->debug_start();
945 int end = jvms->debug_end();
946 sfpt->replace_edges_in_range(res, sobj, start, end);
947 _igvn._worklist.push(sfpt);
948 safepoints_done.append_if_missing(sfpt); // keep it for rollback
949 }
950 // Scalarize value types that were added to the safepoint
951 for (uint i = 0; i < value_worklist.size(); ++i) {
952 Node* vt = value_worklist.at(i);
953 vt->as_ValueType()->make_scalar_in_safepoints(&_igvn);
954 }
955 return true;
956 }
957
958 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
959 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control);
960 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory);
961 if (ctl_proj != NULL) {
962 igvn.replace_node(ctl_proj, n->in(0));
963 }
964 if (mem_proj != NULL) {
965 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
966 }
967 }
968
969 // Process users of eliminated allocation.
970 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) {
971 Node* res = alloc->result_cast();
972 if (res != NULL) {
973 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
974 Node *use = res->last_out(j);
975 uint oc1 = res->outcnt();
976
977 if (use->is_AddP()) {
978 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
979 Node *n = use->last_out(k);
980 uint oc2 = use->outcnt();
981 if (n->is_Store()) {
982 #ifdef ASSERT
983 // Verify that there is no dependent MemBarVolatile nodes,
984 // they should be removed during IGVN, see MemBarNode::Ideal().
985 for (DUIterator_Fast pmax, p = n->fast_outs(pmax);
986 p < pmax; p++) {
987 Node* mb = n->fast_out(p);
988 assert(mb->is_Initialize() || !mb->is_MemBar() ||
989 mb->req() <= MemBarNode::Precedent ||
990 mb->in(MemBarNode::Precedent) != n,
991 "MemBarVolatile should be eliminated for non-escaping object");
992 }
993 #endif
994 _igvn.replace_node(n, n->in(MemNode::Memory));
995 } else if (n->is_ArrayCopy()) {
996 // Disconnect ArrayCopy node
997 ArrayCopyNode* ac = n->as_ArrayCopy();
998 assert(ac->is_clonebasic(), "unexpected array copy kind");
999 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
1000 disconnect_projections(ac, _igvn);
1001 assert(alloc->in(0)->is_Proj() && alloc->in(0)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
1002 Node* membar_before = alloc->in(0)->in(0);
1003 disconnect_projections(membar_before->as_MemBar(), _igvn);
1004 if (membar_after->is_MemBar()) {
1005 disconnect_projections(membar_after->as_MemBar(), _igvn);
1006 }
1007 } else {
1008 eliminate_gc_barrier(n);
1009 }
1010 k -= (oc2 - use->outcnt());
1011 }
1012 _igvn.remove_dead_node(use);
1013 } else if (use->is_ArrayCopy()) {
1014 // Disconnect ArrayCopy node
1015 ArrayCopyNode* ac = use->as_ArrayCopy();
1016 assert(ac->is_arraycopy_validated() ||
1017 ac->is_copyof_validated() ||
1018 ac->is_copyofrange_validated(), "unsupported");
1019 CallProjections* callprojs = ac->extract_projections(true);
1020
1021 _igvn.replace_node(callprojs->fallthrough_ioproj, ac->in(TypeFunc::I_O));
1022 _igvn.replace_node(callprojs->fallthrough_memproj, ac->in(TypeFunc::Memory));
1023 _igvn.replace_node(callprojs->fallthrough_catchproj, ac->in(TypeFunc::Control));
1024
1025 // Set control to top. IGVN will remove the remaining projections
1026 ac->set_req(0, top());
1027 ac->replace_edge(res, top());
1028
1029 // Disconnect src right away: it can help find new
1030 // opportunities for allocation elimination
1031 Node* src = ac->in(ArrayCopyNode::Src);
1032 ac->replace_edge(src, top());
1033 // src can be top at this point if src and dest of the
1034 // arraycopy were the same
1035 if (src->outcnt() == 0 && !src->is_top()) {
1036 _igvn.remove_dead_node(src);
1037 }
1038
1039 _igvn._worklist.push(ac);
1040 } else if (use->is_ValueType()) {
1041 assert(use->isa_ValueType()->get_oop() == res, "unexpected value type use");
1042 _igvn.rehash_node_delayed(use);
1043 use->isa_ValueType()->set_oop(_igvn.zerocon(T_VALUETYPE));
1044 } else if (use->is_Store()) {
1045 _igvn.replace_node(use, use->in(MemNode::Memory));
1046 } else {
1047 eliminate_gc_barrier(use);
1048 }
1049 j -= (oc1 - res->outcnt());
1050 }
1051 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
1052 _igvn.remove_dead_node(res);
1053 }
1054
1055 //
1056 // Process other users of allocation's projections
1057 //
1058 if (_resproj != NULL && _resproj->outcnt() != 0) {
1059 // First disconnect stores captured by Initialize node.
1060 // If Initialize node is eliminated first in the following code,
1061 // it will kill such stores and DUIterator_Last will assert.
1062 for (DUIterator_Fast jmax, j = _resproj->fast_outs(jmax); j < jmax; j++) {
1063 Node *use = _resproj->fast_out(j);
1064 if (use->is_AddP()) {
1065 // raw memory addresses used only by the initialization
1066 _igvn.replace_node(use, C->top());
1067 --j; --jmax;
1068 }
1069 }
1070 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) {
1071 Node *use = _resproj->last_out(j);
1072 uint oc1 = _resproj->outcnt();
1073 if (use->is_Initialize()) {
1074 // Eliminate Initialize node.
1075 InitializeNode *init = use->as_Initialize();
1076 assert(init->outcnt() <= 2, "only a control and memory projection expected");
1077 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control);
1078 if (ctrl_proj != NULL) {
1079 _igvn.replace_node(ctrl_proj, init->in(TypeFunc::Control));
1080 #ifdef ASSERT
1081 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1082 Node* tmp = init->in(TypeFunc::Control);
1083 while (bs->is_gc_barrier_node(tmp)) {
1084 Node* tmp2 = bs->step_over_gc_barrier_ctrl(tmp);
1085 assert(tmp != tmp2, "Must make progress");
1086 tmp = tmp2;
1087 }
1088 assert(tmp == _fallthroughcatchproj, "allocation control projection");
1089 #endif
1090 }
1091 Node *mem_proj = init->proj_out_or_null(TypeFunc::Memory);
1092 if (mem_proj != NULL) {
1093 Node *mem = init->in(TypeFunc::Memory);
1094 #ifdef ASSERT
1095 if (mem->is_MergeMem()) {
1096 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection");
1097 } else {
1098 assert(mem == _memproj_fallthrough, "allocation memory projection");
1099 }
1100 #endif
1101 _igvn.replace_node(mem_proj, mem);
1102 }
1103 } else {
1104 assert(false, "only Initialize or AddP expected");
1105 }
1106 j -= (oc1 - _resproj->outcnt());
1107 }
1108 }
1109 if (_fallthroughcatchproj != NULL) {
1110 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control));
1111 }
1112 if (_memproj_fallthrough != NULL) {
1113 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory));
1114 }
1115 if (_memproj_catchall != NULL) {
1116 _igvn.replace_node(_memproj_catchall, C->top());
1117 }
1118 if (_ioproj_fallthrough != NULL) {
1119 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O));
1120 }
1121 if (_ioproj_catchall != NULL) {
1122 _igvn.replace_node(_ioproj_catchall, C->top());
1123 }
1124 if (_catchallcatchproj != NULL) {
1125 _igvn.replace_node(_catchallcatchproj, C->top());
1126 }
1127 }
1128
1129 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1130 // Don't do scalar replacement if the frame can be popped by JVMTI:
1131 // if reallocation fails during deoptimization we'll pop all
1132 // interpreter frames for this compiled frame and that won't play
1133 // nice with JVMTI popframe.
1134 if (!EliminateAllocations || JvmtiExport::can_pop_frame() || !alloc->_is_non_escaping) {
1135 return false;
1136 }
1137 Node* klass = alloc->in(AllocateNode::KlassNode);
1138 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1139 Node* res = alloc->result_cast();
1140 // Eliminate boxing allocations which are not used
1141 // regardless scalar replacable status.
1142 bool boxing_alloc = C->eliminate_boxing() &&
1143 tklass->klass()->is_instance_klass() &&
1144 tklass->klass()->as_instance_klass()->is_box_klass();
1145 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != NULL))) {
1146 return false;
1147 }
1148
1149 extract_call_projections(alloc);
1150
1151 GrowableArray <SafePointNode *> safepoints;
1152 if (!can_eliminate_allocation(alloc, safepoints)) {
1153 return false;
1154 }
1155
1156 if (!alloc->_is_scalar_replaceable) {
1157 assert(res == NULL, "sanity");
1158 // We can only eliminate allocation if all debug info references
1159 // are already replaced with SafePointScalarObject because
1160 // we can't search for a fields value without instance_id.
1161 if (safepoints.length() > 0) {
1162 return false;
1163 }
1164 }
1165
1166 if (!scalar_replacement(alloc, safepoints)) {
1167 return false;
1168 }
1169
1170 CompileLog* log = C->log();
1171 if (log != NULL) {
1172 log->head("eliminate_allocation type='%d'",
1173 log->identify(tklass->klass()));
1174 JVMState* p = alloc->jvms();
1175 while (p != NULL) {
1176 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1177 p = p->caller();
1178 }
1179 log->tail("eliminate_allocation");
1180 }
1181
1182 process_users_of_allocation(alloc);
1183
1184 #ifndef PRODUCT
1185 if (PrintEliminateAllocations) {
1186 if (alloc->is_AllocateArray())
1187 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1188 else
1189 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1190 }
1191 #endif
1192
1193 return true;
1194 }
1195
1196 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1197 // EA should remove all uses of non-escaping boxing node.
1198 if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != NULL) {
1199 return false;
1200 }
1201
1202 assert(boxing->result_cast() == NULL, "unexpected boxing node result");
1203
1204 extract_call_projections(boxing);
1205
1206 const TypeTuple* r = boxing->tf()->range_sig();
1207 assert(r->cnt() > TypeFunc::Parms, "sanity");
1208 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1209 assert(t != NULL, "sanity");
1210
1211 CompileLog* log = C->log();
1212 if (log != NULL) {
1213 log->head("eliminate_boxing type='%d'",
1214 log->identify(t->klass()));
1215 JVMState* p = boxing->jvms();
1216 while (p != NULL) {
1217 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1218 p = p->caller();
1219 }
1220 log->tail("eliminate_boxing");
1221 }
1222
1223 process_users_of_allocation(boxing);
1224
1225 #ifndef PRODUCT
1226 if (PrintEliminateAllocations) {
1227 tty->print("++++ Eliminated: %d ", boxing->_idx);
1228 boxing->method()->print_short_name(tty);
1229 tty->cr();
1230 }
1231 #endif
1232
1233 return true;
1234 }
1235
1236 //---------------------------set_eden_pointers-------------------------
1237 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) {
1238 if (UseTLAB) { // Private allocation: load from TLS
1239 Node* thread = transform_later(new ThreadLocalNode());
1240 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset());
1241 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset());
1242 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset);
1243 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset);
1244 } else { // Shared allocation: load from globals
1245 CollectedHeap* ch = Universe::heap();
1246 address top_adr = (address)ch->top_addr();
1247 address end_adr = (address)ch->end_addr();
1248 eden_top_adr = makecon(TypeRawPtr::make(top_adr));
1249 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr);
1250 }
1251 }
1252
1253
1254 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
1255 Node* adr = basic_plus_adr(base, offset);
1256 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
1257 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered);
1258 transform_later(value);
1259 return value;
1260 }
1261
1262
1263 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
1264 Node* adr = basic_plus_adr(base, offset);
1265 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt, MemNode::unordered);
1266 transform_later(mem);
1267 return mem;
1268 }
1269
1270 //=============================================================================
1271 //
1272 // A L L O C A T I O N
1273 //
1274 // Allocation attempts to be fast in the case of frequent small objects.
1275 // It breaks down like this:
1276 //
1277 // 1) Size in doublewords is computed. This is a constant for objects and
1278 // variable for most arrays. Doubleword units are used to avoid size
1279 // overflow of huge doubleword arrays. We need doublewords in the end for
1280 // rounding.
1281 //
1282 // 2) Size is checked for being 'too large'. Too-large allocations will go
1283 // the slow path into the VM. The slow path can throw any required
1284 // exceptions, and does all the special checks for very large arrays. The
1285 // size test can constant-fold away for objects. For objects with
1286 // finalizers it constant-folds the otherway: you always go slow with
1287 // finalizers.
1288 //
1289 // 3) If NOT using TLABs, this is the contended loop-back point.
1290 // Load-Locked the heap top. If using TLABs normal-load the heap top.
1291 //
1292 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
1293 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
1294 // "size*8" we always enter the VM, where "largish" is a constant picked small
1295 // enough that there's always space between the eden max and 4Gig (old space is
1296 // there so it's quite large) and large enough that the cost of entering the VM
1297 // is dwarfed by the cost to initialize the space.
1298 //
1299 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1300 // down. If contended, repeat at step 3. If using TLABs normal-store
1301 // adjusted heap top back down; there is no contention.
1302 //
1303 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
1304 // fields.
1305 //
1306 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1307 // oop flavor.
1308 //
1309 //=============================================================================
1310 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1311 // Allocations bigger than this always go the slow route.
1312 // This value must be small enough that allocation attempts that need to
1313 // trigger exceptions go the slow route. Also, it must be small enough so
1314 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1315 //=============================================================================j//
1316 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1317 // The allocator will coalesce int->oop copies away. See comment in
1318 // coalesce.cpp about how this works. It depends critically on the exact
1319 // code shape produced here, so if you are changing this code shape
1320 // make sure the GC info for the heap-top is correct in and around the
1321 // slow-path call.
1322 //
1323
1324 void PhaseMacroExpand::expand_allocate_common(
1325 AllocateNode* alloc, // allocation node to be expanded
1326 Node* length, // array length for an array allocation
1327 const TypeFunc* slow_call_type, // Type of slow call
1328 address slow_call_address // Address of slow call
1329 )
1330 {
1331
1332 Node* ctrl = alloc->in(TypeFunc::Control);
1333 Node* mem = alloc->in(TypeFunc::Memory);
1334 Node* i_o = alloc->in(TypeFunc::I_O);
1335 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1336 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1337 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1338
1339 assert(ctrl != NULL, "must have control");
1340 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1341 // they will not be used if "always_slow" is set
1342 enum { slow_result_path = 1, fast_result_path = 2 };
1343 Node *result_region = NULL;
1344 Node *result_phi_rawmem = NULL;
1345 Node *result_phi_rawoop = NULL;
1346 Node *result_phi_i_o = NULL;
1347
1348 // The initial slow comparison is a size check, the comparison
1349 // we want to do is a BoolTest::gt
1350 bool always_slow = false;
1351 int tv = _igvn.find_int_con(initial_slow_test, -1);
1352 if (tv >= 0) {
1353 always_slow = (tv == 1);
1354 initial_slow_test = NULL;
1355 } else {
1356 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1357 }
1358
1359 if (C->env()->dtrace_alloc_probes() ||
1360 (!UseTLAB && !Universe::heap()->supports_inline_contig_alloc())) {
1361 // Force slow-path allocation
1362 always_slow = true;
1363 initial_slow_test = NULL;
1364 }
1365
1366 Node *slow_region = NULL;
1367 Node *toobig_false = ctrl;
1368
1369 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent");
1370 // generate the initial test if necessary
1371 if (initial_slow_test != NULL ) {
1372 if (slow_region == NULL) {
1373 slow_region = new RegionNode(1);
1374 }
1375 // Now make the initial failure test. Usually a too-big test but
1376 // might be a TRUE for finalizers or a fancy class check for
1377 // newInstance0.
1378 IfNode* toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1379 transform_later(toobig_iff);
1380 // Plug the failing-too-big test into the slow-path region
1381 Node* toobig_true = new IfTrueNode(toobig_iff);
1382 transform_later(toobig_true);
1383 slow_region ->add_req(toobig_true);
1384 toobig_false = new IfFalseNode(toobig_iff);
1385 transform_later(toobig_false);
1386 } else { // No initial test, just fall into next case
1387 toobig_false = ctrl;
1388 }
1389
1390 Node *slow_mem = mem; // save the current memory state for slow path
1391 // generate the fast allocation code unless we know that the initial test will always go slow
1392 if (!always_slow) {
1393 // Fast path modifies only raw memory.
1394 if (mem->is_MergeMem()) {
1395 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1396 }
1397
1398 // allocate the Region and Phi nodes for the result
1399 result_region = new RegionNode(3);
1400 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1401 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1402 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1403
1404 // Grab regular I/O before optional prefetch may change it.
1405 // Slow-path does no I/O so just set it to the original I/O.
1406 result_phi_i_o->init_req(slow_result_path, i_o);
1407
1408 Node* needgc_ctrl = NULL;
1409 // Name successful fast-path variables
1410 Node* fast_oop_ctrl;
1411 Node* fast_oop_rawmem;
1412
1413 intx prefetch_lines = length != NULL ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1414
1415 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1416 Node* fast_oop = bs->obj_allocate(this, ctrl, mem, toobig_false, size_in_bytes, i_o, needgc_ctrl,
1417 fast_oop_ctrl, fast_oop_rawmem,
1418 prefetch_lines);
1419
1420 if (slow_region != NULL) {
1421 slow_region->add_req(needgc_ctrl);
1422 // This completes all paths into the slow merge point
1423 transform_later(slow_region);
1424 } else {
1425 // Just fall from the need-GC path straight into the VM call.
1426 slow_region = needgc_ctrl;
1427 }
1428
1429 InitializeNode* init = alloc->initialization();
1430 fast_oop_rawmem = initialize_object(alloc,
1431 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1432 klass_node, length, size_in_bytes);
1433
1434 // If initialization is performed by an array copy, any required
1435 // MemBarStoreStore was already added. If the object does not
1436 // escape no need for a MemBarStoreStore. If the object does not
1437 // escape in its initializer and memory barrier (MemBarStoreStore or
1438 // stronger) is already added at exit of initializer, also no need
1439 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore
1440 // so that stores that initialize this object can't be reordered
1441 // with a subsequent store that makes this object accessible by
1442 // other threads.
1443 // Other threads include java threads and JVM internal threads
1444 // (for example concurrent GC threads). Current concurrent GC
1445 // implementation: CMS and G1 will not scan newly created object,
1446 // so it's safe to skip storestore barrier when allocation does
1447 // not escape.
1448 if (!alloc->does_not_escape_thread() &&
1449 !alloc->is_allocation_MemBar_redundant() &&
1450 (init == NULL || !init->is_complete_with_arraycopy())) {
1451 if (init == NULL || init->req() < InitializeNode::RawStores) {
1452 // No InitializeNode or no stores captured by zeroing
1453 // elimination. Simply add the MemBarStoreStore after object
1454 // initialization.
1455 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1456 transform_later(mb);
1457
1458 mb->init_req(TypeFunc::Memory, fast_oop_rawmem);
1459 mb->init_req(TypeFunc::Control, fast_oop_ctrl);
1460 fast_oop_ctrl = new ProjNode(mb,TypeFunc::Control);
1461 transform_later(fast_oop_ctrl);
1462 fast_oop_rawmem = new ProjNode(mb,TypeFunc::Memory);
1463 transform_later(fast_oop_rawmem);
1464 } else {
1465 // Add the MemBarStoreStore after the InitializeNode so that
1466 // all stores performing the initialization that were moved
1467 // before the InitializeNode happen before the storestore
1468 // barrier.
1469
1470 Node* init_ctrl = init->proj_out_or_null(TypeFunc::Control);
1471 Node* init_mem = init->proj_out_or_null(TypeFunc::Memory);
1472
1473 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1474 transform_later(mb);
1475
1476 Node* ctrl = new ProjNode(init,TypeFunc::Control);
1477 transform_later(ctrl);
1478 Node* mem = new ProjNode(init,TypeFunc::Memory);
1479 transform_later(mem);
1480
1481 // The MemBarStoreStore depends on control and memory coming
1482 // from the InitializeNode
1483 mb->init_req(TypeFunc::Memory, mem);
1484 mb->init_req(TypeFunc::Control, ctrl);
1485
1486 ctrl = new ProjNode(mb,TypeFunc::Control);
1487 transform_later(ctrl);
1488 mem = new ProjNode(mb,TypeFunc::Memory);
1489 transform_later(mem);
1490
1491 // All nodes that depended on the InitializeNode for control
1492 // and memory must now depend on the MemBarNode that itself
1493 // depends on the InitializeNode
1494 if (init_ctrl != NULL) {
1495 _igvn.replace_node(init_ctrl, ctrl);
1496 }
1497 if (init_mem != NULL) {
1498 _igvn.replace_node(init_mem, mem);
1499 }
1500 }
1501 }
1502
1503 if (C->env()->dtrace_extended_probes()) {
1504 // Slow-path call
1505 int size = TypeFunc::Parms + 2;
1506 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1507 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base),
1508 "dtrace_object_alloc",
1509 TypeRawPtr::BOTTOM);
1510
1511 // Get base of thread-local storage area
1512 Node* thread = new ThreadLocalNode();
1513 transform_later(thread);
1514
1515 call->init_req(TypeFunc::Parms+0, thread);
1516 call->init_req(TypeFunc::Parms+1, fast_oop);
1517 call->init_req(TypeFunc::Control, fast_oop_ctrl);
1518 call->init_req(TypeFunc::I_O , top()); // does no i/o
1519 call->init_req(TypeFunc::Memory , fast_oop_rawmem);
1520 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1521 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1522 transform_later(call);
1523 fast_oop_ctrl = new ProjNode(call,TypeFunc::Control);
1524 transform_later(fast_oop_ctrl);
1525 fast_oop_rawmem = new ProjNode(call,TypeFunc::Memory);
1526 transform_later(fast_oop_rawmem);
1527 }
1528
1529 // Plug in the successful fast-path into the result merge point
1530 result_region ->init_req(fast_result_path, fast_oop_ctrl);
1531 result_phi_rawoop->init_req(fast_result_path, fast_oop);
1532 result_phi_i_o ->init_req(fast_result_path, i_o);
1533 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1534 } else {
1535 slow_region = ctrl;
1536 result_phi_i_o = i_o; // Rename it to use in the following code.
1537 }
1538
1539 // Generate slow-path call
1540 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1541 OptoRuntime::stub_name(slow_call_address),
1542 alloc->jvms()->bci(),
1543 TypePtr::BOTTOM);
1544 call->init_req( TypeFunc::Control, slow_region );
1545 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o
1546 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs
1547 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) );
1548 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) );
1549
1550 call->init_req(TypeFunc::Parms+0, klass_node);
1551 if (length != NULL) {
1552 call->init_req(TypeFunc::Parms+1, length);
1553 }
1554
1555 // Copy debug information and adjust JVMState information, then replace
1556 // allocate node with the call
1557 call->copy_call_debug_info(&_igvn, alloc);
1558 if (!always_slow) {
1559 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1560 } else {
1561 // Hook i_o projection to avoid its elimination during allocation
1562 // replacement (when only a slow call is generated).
1563 call->set_req(TypeFunc::I_O, result_phi_i_o);
1564 }
1565 _igvn.replace_node(alloc, call);
1566 transform_later(call);
1567
1568 // Identify the output projections from the allocate node and
1569 // adjust any references to them.
1570 // The control and io projections look like:
1571 //
1572 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1573 // Allocate Catch
1574 // ^---Proj(io) <-------+ ^---CatchProj(io)
1575 //
1576 // We are interested in the CatchProj nodes.
1577 //
1578 extract_call_projections(call);
1579
1580 // An allocate node has separate memory projections for the uses on
1581 // the control and i_o paths. Replace the control memory projection with
1582 // result_phi_rawmem (unless we are only generating a slow call when
1583 // both memory projections are combined)
1584 if (!always_slow && _memproj_fallthrough != NULL) {
1585 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) {
1586 Node *use = _memproj_fallthrough->fast_out(i);
1587 _igvn.rehash_node_delayed(use);
1588 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem);
1589 // back up iterator
1590 --i;
1591 }
1592 }
1593 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete
1594 // _memproj_catchall so we end up with a call that has only 1 memory projection.
1595 if (_memproj_catchall != NULL ) {
1596 if (_memproj_fallthrough == NULL) {
1597 _memproj_fallthrough = new ProjNode(call, TypeFunc::Memory);
1598 transform_later(_memproj_fallthrough);
1599 }
1600 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) {
1601 Node *use = _memproj_catchall->fast_out(i);
1602 _igvn.rehash_node_delayed(use);
1603 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough);
1604 // back up iterator
1605 --i;
1606 }
1607 assert(_memproj_catchall->outcnt() == 0, "all uses must be deleted");
1608 _igvn.remove_dead_node(_memproj_catchall);
1609 }
1610
1611 // An allocate node has separate i_o projections for the uses on the control
1612 // and i_o paths. Always replace the control i_o projection with result i_o
1613 // otherwise incoming i_o become dead when only a slow call is generated
1614 // (it is different from memory projections where both projections are
1615 // combined in such case).
1616 if (_ioproj_fallthrough != NULL) {
1617 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) {
1618 Node *use = _ioproj_fallthrough->fast_out(i);
1619 _igvn.rehash_node_delayed(use);
1620 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o);
1621 // back up iterator
1622 --i;
1623 }
1624 }
1625 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete
1626 // _ioproj_catchall so we end up with a call that has only 1 i_o projection.
1627 if (_ioproj_catchall != NULL ) {
1628 if (_ioproj_fallthrough == NULL) {
1629 _ioproj_fallthrough = new ProjNode(call, TypeFunc::I_O);
1630 transform_later(_ioproj_fallthrough);
1631 }
1632 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) {
1633 Node *use = _ioproj_catchall->fast_out(i);
1634 _igvn.rehash_node_delayed(use);
1635 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough);
1636 // back up iterator
1637 --i;
1638 }
1639 assert(_ioproj_catchall->outcnt() == 0, "all uses must be deleted");
1640 _igvn.remove_dead_node(_ioproj_catchall);
1641 }
1642
1643 // if we generated only a slow call, we are done
1644 if (always_slow) {
1645 // Now we can unhook i_o.
1646 if (result_phi_i_o->outcnt() > 1) {
1647 call->set_req(TypeFunc::I_O, top());
1648 } else {
1649 assert(result_phi_i_o->unique_ctrl_out() == call, "");
1650 // Case of new array with negative size known during compilation.
1651 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1652 // following code since call to runtime will throw exception.
1653 // As result there will be no users of i_o after the call.
1654 // Leave i_o attached to this call to avoid problems in preceding graph.
1655 }
1656 return;
1657 }
1658
1659 if (_fallthroughcatchproj != NULL) {
1660 ctrl = _fallthroughcatchproj->clone();
1661 transform_later(ctrl);
1662 _igvn.replace_node(_fallthroughcatchproj, result_region);
1663 } else {
1664 ctrl = top();
1665 }
1666 Node *slow_result;
1667 if (_resproj == NULL) {
1668 // no uses of the allocation result
1669 slow_result = top();
1670 } else {
1671 slow_result = _resproj->clone();
1672 transform_later(slow_result);
1673 _igvn.replace_node(_resproj, result_phi_rawoop);
1674 }
1675
1676 // Plug slow-path into result merge point
1677 result_region ->init_req( slow_result_path, ctrl );
1678 result_phi_rawoop->init_req( slow_result_path, slow_result);
1679 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough );
1680 transform_later(result_region);
1681 transform_later(result_phi_rawoop);
1682 transform_later(result_phi_rawmem);
1683 transform_later(result_phi_i_o);
1684 // This completes all paths into the result merge point
1685 }
1686
1687
1688 // Helper for PhaseMacroExpand::expand_allocate_common.
1689 // Initializes the newly-allocated storage.
1690 Node* PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1691 Node* control, Node* rawmem, Node* object,
1692 Node* klass_node, Node* length,
1693 Node* size_in_bytes) {
1694 InitializeNode* init = alloc->initialization();
1695 // Store the klass & mark bits
1696 Node* mark_node = alloc->make_ideal_mark(&_igvn, object, control, rawmem, klass_node);
1697 if (!mark_node->is_Con()) {
1698 transform_later(mark_node);
1699 }
1700 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, TypeX_X->basic_type());
1701
1702 BasicType bt = T_METADATA;
1703 Node* metadata = klass_node;
1704 Node* properties = alloc->in(AllocateNode::StorageProperties);
1705 if (properties != NULL) {
1706 // Encode array storage properties into klass pointer
1707 assert(EnableValhalla, "array storage properties not supported");
1708 if (UseCompressedClassPointers) {
1709 // Compress the klass pointer before inserting the storage properties value
1710 metadata = transform_later(new EncodePKlassNode(metadata, metadata->bottom_type()->make_narrowklass()));
1711 metadata = transform_later(new CastN2INode(metadata));
1712 metadata = transform_later(new OrINode(metadata, transform_later(new ConvL2INode(properties))));
1713 bt = T_INT;
1714 } else {
1715 metadata = transform_later(new CastP2XNode(NULL, metadata));
1716 metadata = transform_later(new OrXNode(metadata, properties));
1717 bt = T_LONG;
1718 }
1719 }
1720 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), metadata, bt);
1721
1722 int header_size = alloc->minimum_header_size(); // conservatively small
1723
1724 // Array length
1725 if (length != NULL) { // Arrays need length field
1726 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1727 // conservatively small header size:
1728 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1729 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1730 if (k->is_array_klass()) // we know the exact header size in most cases:
1731 header_size = Klass::layout_helper_header_size(k->layout_helper());
1732 }
1733
1734 // Clear the object body, if necessary.
1735 if (init == NULL) {
1736 // The init has somehow disappeared; be cautious and clear everything.
1737 //
1738 // This can happen if a node is allocated but an uncommon trap occurs
1739 // immediately. In this case, the Initialize gets associated with the
1740 // trap, and may be placed in a different (outer) loop, if the Allocate
1741 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1742 // there can be two Allocates to one Initialize. The answer in all these
1743 // edge cases is safety first. It is always safe to clear immediately
1744 // within an Allocate, and then (maybe or maybe not) clear some more later.
1745 if (!(UseTLAB && ZeroTLAB)) {
1746 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1747 alloc->in(AllocateNode::DefaultValue),
1748 alloc->in(AllocateNode::RawDefaultValue),
1749 header_size, size_in_bytes,
1750 &_igvn);
1751 }
1752 } else {
1753 if (!init->is_complete()) {
1754 // Try to win by zeroing only what the init does not store.
1755 // We can also try to do some peephole optimizations,
1756 // such as combining some adjacent subword stores.
1757 rawmem = init->complete_stores(control, rawmem, object,
1758 header_size, size_in_bytes, &_igvn);
1759 }
1760 // We have no more use for this link, since the AllocateNode goes away:
1761 init->set_req(InitializeNode::RawAddress, top());
1762 // (If we keep the link, it just confuses the register allocator,
1763 // who thinks he sees a real use of the address by the membar.)
1764 }
1765
1766 return rawmem;
1767 }
1768
1769 // Generate prefetch instructions for next allocations.
1770 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
1771 Node*& contended_phi_rawmem,
1772 Node* old_eden_top, Node* new_eden_top,
1773 intx lines) {
1774 enum { fall_in_path = 1, pf_path = 2 };
1775 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
1776 // Generate prefetch allocation with watermark check.
1777 // As an allocation hits the watermark, we will prefetch starting
1778 // at a "distance" away from watermark.
1779
1780 Node *pf_region = new RegionNode(3);
1781 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY,
1782 TypeRawPtr::BOTTOM );
1783 // I/O is used for Prefetch
1784 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO );
1785
1786 Node *thread = new ThreadLocalNode();
1787 transform_later(thread);
1788
1789 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread,
1790 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
1791 transform_later(eden_pf_adr);
1792
1793 Node *old_pf_wm = new LoadPNode(needgc_false,
1794 contended_phi_rawmem, eden_pf_adr,
1795 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM,
1796 MemNode::unordered);
1797 transform_later(old_pf_wm);
1798
1799 // check against new_eden_top
1800 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm );
1801 transform_later(need_pf_cmp);
1802 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge );
1803 transform_later(need_pf_bol);
1804 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol,
1805 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1806 transform_later(need_pf_iff);
1807
1808 // true node, add prefetchdistance
1809 Node *need_pf_true = new IfTrueNode( need_pf_iff );
1810 transform_later(need_pf_true);
1811
1812 Node *need_pf_false = new IfFalseNode( need_pf_iff );
1813 transform_later(need_pf_false);
1814
1815 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm,
1816 _igvn.MakeConX(AllocatePrefetchDistance) );
1817 transform_later(new_pf_wmt );
1818 new_pf_wmt->set_req(0, need_pf_true);
1819
1820 Node *store_new_wmt = new StorePNode(need_pf_true,
1821 contended_phi_rawmem, eden_pf_adr,
1822 TypeRawPtr::BOTTOM, new_pf_wmt,
1823 MemNode::unordered);
1824 transform_later(store_new_wmt);
1825
1826 // adding prefetches
1827 pf_phi_abio->init_req( fall_in_path, i_o );
1828
1829 Node *prefetch_adr;
1830 Node *prefetch;
1831 uint step_size = AllocatePrefetchStepSize;
1832 uint distance = 0;
1833
1834 for ( intx i = 0; i < lines; i++ ) {
1835 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt,
1836 _igvn.MakeConX(distance) );
1837 transform_later(prefetch_adr);
1838 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1839 transform_later(prefetch);
1840 distance += step_size;
1841 i_o = prefetch;
1842 }
1843 pf_phi_abio->set_req( pf_path, i_o );
1844
1845 pf_region->init_req( fall_in_path, need_pf_false );
1846 pf_region->init_req( pf_path, need_pf_true );
1847
1848 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
1849 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
1850
1851 transform_later(pf_region);
1852 transform_later(pf_phi_rawmem);
1853 transform_later(pf_phi_abio);
1854
1855 needgc_false = pf_region;
1856 contended_phi_rawmem = pf_phi_rawmem;
1857 i_o = pf_phi_abio;
1858 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) {
1859 // Insert a prefetch instruction for each allocation.
1860 // This code is used to generate 1 prefetch instruction per cache line.
1861
1862 // Generate several prefetch instructions.
1863 uint step_size = AllocatePrefetchStepSize;
1864 uint distance = AllocatePrefetchDistance;
1865
1866 // Next cache address.
1867 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top,
1868 _igvn.MakeConX(step_size + distance));
1869 transform_later(cache_adr);
1870 cache_adr = new CastP2XNode(needgc_false, cache_adr);
1871 transform_later(cache_adr);
1872 // Address is aligned to execute prefetch to the beginning of cache line size
1873 // (it is important when BIS instruction is used on SPARC as prefetch).
1874 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1));
1875 cache_adr = new AndXNode(cache_adr, mask);
1876 transform_later(cache_adr);
1877 cache_adr = new CastX2PNode(cache_adr);
1878 transform_later(cache_adr);
1879
1880 // Prefetch
1881 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr );
1882 prefetch->set_req(0, needgc_false);
1883 transform_later(prefetch);
1884 contended_phi_rawmem = prefetch;
1885 Node *prefetch_adr;
1886 distance = step_size;
1887 for ( intx i = 1; i < lines; i++ ) {
1888 prefetch_adr = new AddPNode( cache_adr, cache_adr,
1889 _igvn.MakeConX(distance) );
1890 transform_later(prefetch_adr);
1891 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr );
1892 transform_later(prefetch);
1893 distance += step_size;
1894 contended_phi_rawmem = prefetch;
1895 }
1896 } else if( AllocatePrefetchStyle > 0 ) {
1897 // Insert a prefetch for each allocation only on the fast-path
1898 Node *prefetch_adr;
1899 Node *prefetch;
1900 // Generate several prefetch instructions.
1901 uint step_size = AllocatePrefetchStepSize;
1902 uint distance = AllocatePrefetchDistance;
1903 for ( intx i = 0; i < lines; i++ ) {
1904 prefetch_adr = new AddPNode( old_eden_top, new_eden_top,
1905 _igvn.MakeConX(distance) );
1906 transform_later(prefetch_adr);
1907 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1908 // Do not let it float too high, since if eden_top == eden_end,
1909 // both might be null.
1910 if( i == 0 ) { // Set control for first prefetch, next follows it
1911 prefetch->init_req(0, needgc_false);
1912 }
1913 transform_later(prefetch);
1914 distance += step_size;
1915 i_o = prefetch;
1916 }
1917 }
1918 return i_o;
1919 }
1920
1921
1922 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
1923 expand_allocate_common(alloc, NULL,
1924 OptoRuntime::new_instance_Type(),
1925 OptoRuntime::new_instance_Java());
1926 }
1927
1928 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
1929 Node* length = alloc->in(AllocateNode::ALength);
1930 InitializeNode* init = alloc->initialization();
1931 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1932 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1933 address slow_call_address; // Address of slow call
1934 if (init != NULL && init->is_complete_with_arraycopy() &&
1935 k->is_type_array_klass()) {
1936 // Don't zero type array during slow allocation in VM since
1937 // it will be initialized later by arraycopy in compiled code.
1938 slow_call_address = OptoRuntime::new_array_nozero_Java();
1939 } else {
1940 slow_call_address = OptoRuntime::new_array_Java();
1941 }
1942 expand_allocate_common(alloc, length,
1943 OptoRuntime::new_array_Type(),
1944 slow_call_address);
1945 }
1946
1947 //-------------------mark_eliminated_box----------------------------------
1948 //
1949 // During EA obj may point to several objects but after few ideal graph
1950 // transformations (CCP) it may point to only one non escaping object
1951 // (but still using phi), corresponding locks and unlocks will be marked
1952 // for elimination. Later obj could be replaced with a new node (new phi)
1953 // and which does not have escape information. And later after some graph
1954 // reshape other locks and unlocks (which were not marked for elimination
1955 // before) are connected to this new obj (phi) but they still will not be
1956 // marked for elimination since new obj has no escape information.
1957 // Mark all associated (same box and obj) lock and unlock nodes for
1958 // elimination if some of them marked already.
1959 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) {
1960 if (oldbox->as_BoxLock()->is_eliminated())
1961 return; // This BoxLock node was processed already.
1962
1963 // New implementation (EliminateNestedLocks) has separate BoxLock
1964 // node for each locked region so mark all associated locks/unlocks as
1965 // eliminated even if different objects are referenced in one locked region
1966 // (for example, OSR compilation of nested loop inside locked scope).
1967 if (EliminateNestedLocks ||
1968 oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj)) {
1969 // Box is used only in one lock region. Mark this box as eliminated.
1970 _igvn.hash_delete(oldbox);
1971 oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value
1972 _igvn.hash_insert(oldbox);
1973
1974 for (uint i = 0; i < oldbox->outcnt(); i++) {
1975 Node* u = oldbox->raw_out(i);
1976 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) {
1977 AbstractLockNode* alock = u->as_AbstractLock();
1978 // Check lock's box since box could be referenced by Lock's debug info.
1979 if (alock->box_node() == oldbox) {
1980 // Mark eliminated all related locks and unlocks.
1981 #ifdef ASSERT
1982 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4");
1983 #endif
1984 alock->set_non_esc_obj();
1985 }
1986 }
1987 }
1988 return;
1989 }
1990
1991 // Create new "eliminated" BoxLock node and use it in monitor debug info
1992 // instead of oldbox for the same object.
1993 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
1994
1995 // Note: BoxLock node is marked eliminated only here and it is used
1996 // to indicate that all associated lock and unlock nodes are marked
1997 // for elimination.
1998 newbox->set_eliminated();
1999 transform_later(newbox);
2000
2001 // Replace old box node with new box for all users of the same object.
2002 for (uint i = 0; i < oldbox->outcnt();) {
2003 bool next_edge = true;
2004
2005 Node* u = oldbox->raw_out(i);
2006 if (u->is_AbstractLock()) {
2007 AbstractLockNode* alock = u->as_AbstractLock();
2008 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) {
2009 // Replace Box and mark eliminated all related locks and unlocks.
2010 #ifdef ASSERT
2011 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5");
2012 #endif
2013 alock->set_non_esc_obj();
2014 _igvn.rehash_node_delayed(alock);
2015 alock->set_box_node(newbox);
2016 next_edge = false;
2017 }
2018 }
2019 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) {
2020 FastLockNode* flock = u->as_FastLock();
2021 assert(flock->box_node() == oldbox, "sanity");
2022 _igvn.rehash_node_delayed(flock);
2023 flock->set_box_node(newbox);
2024 next_edge = false;
2025 }
2026
2027 // Replace old box in monitor debug info.
2028 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
2029 SafePointNode* sfn = u->as_SafePoint();
2030 JVMState* youngest_jvms = sfn->jvms();
2031 int max_depth = youngest_jvms->depth();
2032 for (int depth = 1; depth <= max_depth; depth++) {
2033 JVMState* jvms = youngest_jvms->of_depth(depth);
2034 int num_mon = jvms->nof_monitors();
2035 // Loop over monitors
2036 for (int idx = 0; idx < num_mon; idx++) {
2037 Node* obj_node = sfn->monitor_obj(jvms, idx);
2038 Node* box_node = sfn->monitor_box(jvms, idx);
2039 if (box_node == oldbox && obj_node->eqv_uncast(obj)) {
2040 int j = jvms->monitor_box_offset(idx);
2041 _igvn.replace_input_of(u, j, newbox);
2042 next_edge = false;
2043 }
2044 }
2045 }
2046 }
2047 if (next_edge) i++;
2048 }
2049 }
2050
2051 //-----------------------mark_eliminated_locking_nodes-----------------------
2052 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) {
2053 if (EliminateNestedLocks) {
2054 if (alock->is_nested()) {
2055 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity");
2056 return;
2057 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened
2058 // Only Lock node has JVMState needed here.
2059 // Not that preceding claim is documented anywhere else.
2060 if (alock->jvms() != NULL) {
2061 if (alock->as_Lock()->is_nested_lock_region()) {
2062 // Mark eliminated related nested locks and unlocks.
2063 Node* obj = alock->obj_node();
2064 BoxLockNode* box_node = alock->box_node()->as_BoxLock();
2065 assert(!box_node->is_eliminated(), "should not be marked yet");
2066 // Note: BoxLock node is marked eliminated only here
2067 // and it is used to indicate that all associated lock
2068 // and unlock nodes are marked for elimination.
2069 box_node->set_eliminated(); // Box's hash is always NO_HASH here
2070 for (uint i = 0; i < box_node->outcnt(); i++) {
2071 Node* u = box_node->raw_out(i);
2072 if (u->is_AbstractLock()) {
2073 alock = u->as_AbstractLock();
2074 if (alock->box_node() == box_node) {
2075 // Verify that this Box is referenced only by related locks.
2076 assert(alock->obj_node()->eqv_uncast(obj), "");
2077 // Mark all related locks and unlocks.
2078 #ifdef ASSERT
2079 alock->log_lock_optimization(C, "eliminate_lock_set_nested");
2080 #endif
2081 alock->set_nested();
2082 }
2083 }
2084 }
2085 } else {
2086 #ifdef ASSERT
2087 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region");
2088 if (C->log() != NULL)
2089 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output
2090 #endif
2091 }
2092 }
2093 return;
2094 }
2095 // Process locks for non escaping object
2096 assert(alock->is_non_esc_obj(), "");
2097 } // EliminateNestedLocks
2098
2099 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2100 // Look for all locks of this object and mark them and
2101 // corresponding BoxLock nodes as eliminated.
2102 Node* obj = alock->obj_node();
2103 for (uint j = 0; j < obj->outcnt(); j++) {
2104 Node* o = obj->raw_out(j);
2105 if (o->is_AbstractLock() &&
2106 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2107 alock = o->as_AbstractLock();
2108 Node* box = alock->box_node();
2109 // Replace old box node with new eliminated box for all users
2110 // of the same object and mark related locks as eliminated.
2111 mark_eliminated_box(box, obj);
2112 }
2113 }
2114 }
2115 }
2116
2117 // we have determined that this lock/unlock can be eliminated, we simply
2118 // eliminate the node without expanding it.
2119 //
2120 // Note: The membar's associated with the lock/unlock are currently not
2121 // eliminated. This should be investigated as a future enhancement.
2122 //
2123 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2124
2125 if (!alock->is_eliminated()) {
2126 return false;
2127 }
2128 #ifdef ASSERT
2129 const Type* obj_type = _igvn.type(alock->obj_node());
2130 assert(!obj_type->isa_valuetype() && !obj_type->is_valuetypeptr(), "Eliminating lock on value type");
2131 if (!alock->is_coarsened()) {
2132 // Check that new "eliminated" BoxLock node is created.
2133 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2134 assert(oldbox->is_eliminated(), "should be done already");
2135 }
2136 #endif
2137
2138 alock->log_lock_optimization(C, "eliminate_lock");
2139
2140 #ifndef PRODUCT
2141 if (PrintEliminateLocks) {
2142 if (alock->is_Lock()) {
2143 tty->print_cr("++++ Eliminated: %d Lock", alock->_idx);
2144 } else {
2145 tty->print_cr("++++ Eliminated: %d Unlock", alock->_idx);
2146 }
2147 }
2148 #endif
2149
2150 Node* mem = alock->in(TypeFunc::Memory);
2151 Node* ctrl = alock->in(TypeFunc::Control);
2152 guarantee(ctrl != NULL, "missing control projection, cannot replace_node() with NULL");
2153
2154 extract_call_projections(alock);
2155 // There are 2 projections from the lock. The lock node will
2156 // be deleted when its last use is subsumed below.
2157 assert(alock->outcnt() == 2 &&
2158 _fallthroughproj != NULL &&
2159 _memproj_fallthrough != NULL,
2160 "Unexpected projections from Lock/Unlock");
2161
2162 Node* fallthroughproj = _fallthroughproj;
2163 Node* memproj_fallthrough = _memproj_fallthrough;
2164
2165 // The memory projection from a lock/unlock is RawMem
2166 // The input to a Lock is merged memory, so extract its RawMem input
2167 // (unless the MergeMem has been optimized away.)
2168 if (alock->is_Lock()) {
2169 // Seach for MemBarAcquireLock node and delete it also.
2170 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2171 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, "");
2172 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2173 Node* memproj = membar->proj_out(TypeFunc::Memory);
2174 _igvn.replace_node(ctrlproj, fallthroughproj);
2175 _igvn.replace_node(memproj, memproj_fallthrough);
2176
2177 // Delete FastLock node also if this Lock node is unique user
2178 // (a loop peeling may clone a Lock node).
2179 Node* flock = alock->as_Lock()->fastlock_node();
2180 if (flock->outcnt() == 1) {
2181 assert(flock->unique_out() == alock, "sanity");
2182 _igvn.replace_node(flock, top());
2183 }
2184 }
2185
2186 // Seach for MemBarReleaseLock node and delete it also.
2187 if (alock->is_Unlock() && ctrl->is_Proj() && ctrl->in(0)->is_MemBar()) {
2188 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2189 assert(membar->Opcode() == Op_MemBarReleaseLock &&
2190 mem->is_Proj() && membar == mem->in(0), "");
2191 _igvn.replace_node(fallthroughproj, ctrl);
2192 _igvn.replace_node(memproj_fallthrough, mem);
2193 fallthroughproj = ctrl;
2194 memproj_fallthrough = mem;
2195 ctrl = membar->in(TypeFunc::Control);
2196 mem = membar->in(TypeFunc::Memory);
2197 }
2198
2199 _igvn.replace_node(fallthroughproj, ctrl);
2200 _igvn.replace_node(memproj_fallthrough, mem);
2201 return true;
2202 }
2203
2204
2205 //------------------------------expand_lock_node----------------------
2206 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
2207
2208 Node* ctrl = lock->in(TypeFunc::Control);
2209 Node* mem = lock->in(TypeFunc::Memory);
2210 Node* obj = lock->obj_node();
2211 Node* box = lock->box_node();
2212 Node* flock = lock->fastlock_node();
2213
2214 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2215
2216 // Make the merge point
2217 Node *region;
2218 Node *mem_phi;
2219 Node *slow_path;
2220
2221 if (UseOptoBiasInlining) {
2222 /*
2223 * See the full description in MacroAssembler::biased_locking_enter().
2224 *
2225 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) {
2226 * // The object is biased.
2227 * proto_node = klass->prototype_header;
2228 * o_node = thread | proto_node;
2229 * x_node = o_node ^ mark_word;
2230 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ?
2231 * // Done.
2232 * } else {
2233 * if( (x_node & biased_lock_mask) != 0 ) {
2234 * // The klass's prototype header is no longer biased.
2235 * cas(&mark_word, mark_word, proto_node)
2236 * goto cas_lock;
2237 * } else {
2238 * // The klass's prototype header is still biased.
2239 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch?
2240 * old = mark_word;
2241 * new = o_node;
2242 * } else {
2243 * // Different thread or anonymous biased.
2244 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask);
2245 * new = thread | old;
2246 * }
2247 * // Try to rebias.
2248 * if( cas(&mark_word, old, new) == 0 ) {
2249 * // Done.
2250 * } else {
2251 * goto slow_path; // Failed.
2252 * }
2253 * }
2254 * }
2255 * } else {
2256 * // The object is not biased.
2257 * cas_lock:
2258 * if( FastLock(obj) == 0 ) {
2259 * // Done.
2260 * } else {
2261 * slow_path:
2262 * OptoRuntime::complete_monitor_locking_Java(obj);
2263 * }
2264 * }
2265 */
2266
2267 region = new RegionNode(5);
2268 // create a Phi for the memory state
2269 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2270
2271 Node* fast_lock_region = new RegionNode(3);
2272 Node* fast_lock_mem_phi = new PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM);
2273
2274 // First, check mark word for the biased lock pattern.
2275 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2276
2277 // Get fast path - mark word has the biased lock pattern.
2278 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node,
2279 markWord::biased_lock_mask_in_place,
2280 markWord::biased_lock_pattern, true);
2281 // fast_lock_region->in(1) is set to slow path.
2282 fast_lock_mem_phi->init_req(1, mem);
2283
2284 // Now check that the lock is biased to the current thread and has
2285 // the same epoch and bias as Klass::_prototype_header.
2286
2287 // Special-case a fresh allocation to avoid building nodes:
2288 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn);
2289 if (klass_node == NULL) {
2290 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
2291 klass_node = transform_later(LoadKlassNode::make(_igvn, NULL, mem, k_adr, _igvn.type(k_adr)->is_ptr()));
2292 #ifdef _LP64
2293 if (UseCompressedClassPointers && klass_node->is_DecodeNKlass()) {
2294 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity");
2295 klass_node->in(1)->init_req(0, ctrl);
2296 } else
2297 #endif
2298 klass_node->init_req(0, ctrl);
2299 }
2300 Node *proto_node = make_load(ctrl, mem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeX_X, TypeX_X->basic_type());
2301
2302 Node* thread = transform_later(new ThreadLocalNode());
2303 Node* cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2304 Node* o_node = transform_later(new OrXNode(cast_thread, proto_node));
2305 Node* x_node = transform_later(new XorXNode(o_node, mark_node));
2306
2307 // Get slow path - mark word does NOT match the value.
2308 STATIC_ASSERT(markWord::age_mask_in_place <= INT_MAX);
2309 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node,
2310 (~(int)markWord::age_mask_in_place), 0);
2311 // region->in(3) is set to fast path - the object is biased to the current thread.
2312 mem_phi->init_req(3, mem);
2313
2314
2315 // Mark word does NOT match the value (thread | Klass::_prototype_header).
2316
2317
2318 // First, check biased pattern.
2319 // Get fast path - _prototype_header has the same biased lock pattern.
2320 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node,
2321 markWord::biased_lock_mask_in_place, 0, true);
2322
2323 not_biased_ctrl = fast_lock_region->in(2); // Slow path
2324 // fast_lock_region->in(2) - the prototype header is no longer biased
2325 // and we have to revoke the bias on this object.
2326 // We are going to try to reset the mark of this object to the prototype
2327 // value and fall through to the CAS-based locking scheme.
2328 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
2329 Node* cas = new StoreXConditionalNode(not_biased_ctrl, mem, adr,
2330 proto_node, mark_node);
2331 transform_later(cas);
2332 Node* proj = transform_later(new SCMemProjNode(cas));
2333 fast_lock_mem_phi->init_req(2, proj);
2334
2335
2336 // Second, check epoch bits.
2337 Node* rebiased_region = new RegionNode(3);
2338 Node* old_phi = new PhiNode( rebiased_region, TypeX_X);
2339 Node* new_phi = new PhiNode( rebiased_region, TypeX_X);
2340
2341 // Get slow path - mark word does NOT match epoch bits.
2342 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node,
2343 markWord::epoch_mask_in_place, 0);
2344 // The epoch of the current bias is not valid, attempt to rebias the object
2345 // toward the current thread.
2346 rebiased_region->init_req(2, epoch_ctrl);
2347 old_phi->init_req(2, mark_node);
2348 new_phi->init_req(2, o_node);
2349
2350 // rebiased_region->in(1) is set to fast path.
2351 // The epoch of the current bias is still valid but we know
2352 // nothing about the owner; it might be set or it might be clear.
2353 Node* cmask = MakeConX(markWord::biased_lock_mask_in_place |
2354 markWord::age_mask_in_place |
2355 markWord::epoch_mask_in_place);
2356 Node* old = transform_later(new AndXNode(mark_node, cmask));
2357 cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2358 Node* new_mark = transform_later(new OrXNode(cast_thread, old));
2359 old_phi->init_req(1, old);
2360 new_phi->init_req(1, new_mark);
2361
2362 transform_later(rebiased_region);
2363 transform_later(old_phi);
2364 transform_later(new_phi);
2365
2366 // Try to acquire the bias of the object using an atomic operation.
2367 // If this fails we will go in to the runtime to revoke the object's bias.
2368 cas = new StoreXConditionalNode(rebiased_region, mem, adr, new_phi, old_phi);
2369 transform_later(cas);
2370 proj = transform_later(new SCMemProjNode(cas));
2371
2372 // Get slow path - Failed to CAS.
2373 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0);
2374 mem_phi->init_req(4, proj);
2375 // region->in(4) is set to fast path - the object is rebiased to the current thread.
2376
2377 // Failed to CAS.
2378 slow_path = new RegionNode(3);
2379 Node *slow_mem = new PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM);
2380
2381 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control
2382 slow_mem->init_req(1, proj);
2383
2384 // Call CAS-based locking scheme (FastLock node).
2385
2386 transform_later(fast_lock_region);
2387 transform_later(fast_lock_mem_phi);
2388
2389 // Get slow path - FastLock failed to lock the object.
2390 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0);
2391 mem_phi->init_req(2, fast_lock_mem_phi);
2392 // region->in(2) is set to fast path - the object is locked to the current thread.
2393
2394 slow_path->init_req(2, ctrl); // Capture slow-control
2395 slow_mem->init_req(2, fast_lock_mem_phi);
2396
2397 transform_later(slow_path);
2398 transform_later(slow_mem);
2399 // Reset lock's memory edge.
2400 lock->set_req(TypeFunc::Memory, slow_mem);
2401
2402 } else {
2403 region = new RegionNode(3);
2404 // create a Phi for the memory state
2405 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2406
2407 // Optimize test; set region slot 2
2408 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2409 mem_phi->init_req(2, mem);
2410 }
2411
2412 const TypeOopPtr* objptr = _igvn.type(obj)->make_oopptr();
2413 if (objptr->can_be_value_type()) {
2414 // Deoptimize and re-execute if a value
2415 assert(EnableValhalla, "should only be used if value types are enabled");
2416 Node* mark = make_load(slow_path, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2417 Node* value_mask = _igvn.MakeConX(markWord::always_locked_pattern);
2418 Node* is_value = _igvn.transform(new AndXNode(mark, value_mask));
2419 Node* cmp = _igvn.transform(new CmpXNode(is_value, value_mask));
2420 Node* bol = _igvn.transform(new BoolNode(cmp, BoolTest::eq));
2421 Node* unc_ctrl = generate_slow_guard(&slow_path, bol, NULL);
2422
2423 int trap_request = Deoptimization::make_trap_request(Deoptimization::Reason_class_check, Deoptimization::Action_none);
2424 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
2425 const TypePtr* no_memory_effects = NULL;
2426 JVMState* jvms = lock->jvms();
2427 CallNode* unc = new CallStaticJavaNode(OptoRuntime::uncommon_trap_Type(), call_addr, "uncommon_trap",
2428 jvms->bci(), no_memory_effects);
2429
2430 unc->init_req(TypeFunc::Control, unc_ctrl);
2431 unc->init_req(TypeFunc::I_O, lock->i_o());
2432 unc->init_req(TypeFunc::Memory, mem); // may gc ptrs
2433 unc->init_req(TypeFunc::FramePtr, lock->in(TypeFunc::FramePtr));
2434 unc->init_req(TypeFunc::ReturnAdr, lock->in(TypeFunc::ReturnAdr));
2435 unc->init_req(TypeFunc::Parms+0, _igvn.intcon(trap_request));
2436 unc->set_cnt(PROB_UNLIKELY_MAG(4));
2437 unc->copy_call_debug_info(&_igvn, lock);
2438
2439 assert(unc->peek_monitor_box() == box, "wrong monitor");
2440 assert(unc->peek_monitor_obj() == obj, "wrong monitor");
2441
2442 // pop monitor and push obj back on stack: we trap before the monitorenter
2443 unc->pop_monitor();
2444 unc->grow_stack(unc->jvms(), 1);
2445 unc->set_stack(unc->jvms(), unc->jvms()->stk_size()-1, obj);
2446
2447 _igvn.register_new_node_with_optimizer(unc);
2448
2449 Node* ctrl = _igvn.transform(new ProjNode(unc, TypeFunc::Control));
2450 Node* halt = _igvn.transform(new HaltNode(ctrl, lock->in(TypeFunc::FramePtr)));
2451 C->root()->add_req(halt);
2452 }
2453
2454 // Make slow path call
2455 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(),
2456 OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path,
2457 obj, box, NULL);
2458
2459 extract_call_projections(call);
2460
2461 // Slow path can only throw asynchronous exceptions, which are always
2462 // de-opted. So the compiler thinks the slow-call can never throw an
2463 // exception. If it DOES throw an exception we would need the debug
2464 // info removed first (since if it throws there is no monitor).
2465 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2466 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2467
2468 // Capture slow path
2469 // disconnect fall-through projection from call and create a new one
2470 // hook up users of fall-through projection to region
2471 Node *slow_ctrl = _fallthroughproj->clone();
2472 transform_later(slow_ctrl);
2473 _igvn.hash_delete(_fallthroughproj);
2474 _fallthroughproj->disconnect_inputs(NULL, C);
2475 region->init_req(1, slow_ctrl);
2476 // region inputs are now complete
2477 transform_later(region);
2478 _igvn.replace_node(_fallthroughproj, region);
2479
2480 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2481 mem_phi->init_req(1, memproj );
2482 transform_later(mem_phi);
2483 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2484 }
2485
2486 //------------------------------expand_unlock_node----------------------
2487 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2488
2489 Node* ctrl = unlock->in(TypeFunc::Control);
2490 Node* mem = unlock->in(TypeFunc::Memory);
2491 Node* obj = unlock->obj_node();
2492 Node* box = unlock->box_node();
2493
2494 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2495
2496 // No need for a null check on unlock
2497
2498 // Make the merge point
2499 Node *region;
2500 Node *mem_phi;
2501
2502 if (UseOptoBiasInlining) {
2503 // Check for biased locking unlock case, which is a no-op.
2504 // See the full description in MacroAssembler::biased_locking_exit().
2505 region = new RegionNode(4);
2506 // create a Phi for the memory state
2507 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2508 mem_phi->init_req(3, mem);
2509
2510 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2511 ctrl = opt_bits_test(ctrl, region, 3, mark_node,
2512 markWord::biased_lock_mask_in_place,
2513 markWord::biased_lock_pattern);
2514 } else {
2515 region = new RegionNode(3);
2516 // create a Phi for the memory state
2517 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2518 }
2519
2520 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2521 funlock = transform_later( funlock )->as_FastUnlock();
2522 // Optimize test; set region slot 2
2523 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2524 Node *thread = transform_later(new ThreadLocalNode());
2525
2526 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2527 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2528 "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2529
2530 extract_call_projections(call);
2531
2532 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2533 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2534
2535 // No exceptions for unlocking
2536 // Capture slow path
2537 // disconnect fall-through projection from call and create a new one
2538 // hook up users of fall-through projection to region
2539 Node *slow_ctrl = _fallthroughproj->clone();
2540 transform_later(slow_ctrl);
2541 _igvn.hash_delete(_fallthroughproj);
2542 _fallthroughproj->disconnect_inputs(NULL, C);
2543 region->init_req(1, slow_ctrl);
2544 // region inputs are now complete
2545 transform_later(region);
2546 _igvn.replace_node(_fallthroughproj, region);
2547
2548 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2549 mem_phi->init_req(1, memproj );
2550 mem_phi->init_req(2, mem);
2551 transform_later(mem_phi);
2552 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2553 }
2554
2555 // A value type might be returned from the call but we don't know its
2556 // type. Either we get a buffered value (and nothing needs to be done)
2557 // or one of the values being returned is the klass of the value type
2558 // and we need to allocate a value type instance of that type and
2559 // initialize it with other values being returned. In that case, we
2560 // first try a fast path allocation and initialize the value with the
2561 // value klass's pack handler or we fall back to a runtime call.
2562 void PhaseMacroExpand::expand_mh_intrinsic_return(CallStaticJavaNode* call) {
2563 assert(call->method()->is_method_handle_intrinsic(), "must be a method handle intrinsic call");
2564 Node* ret = call->proj_out_or_null(TypeFunc::Parms);
2565 if (ret == NULL) {
2566 return;
2567 }
2568 const TypeFunc* tf = call->_tf;
2569 const TypeTuple* domain = OptoRuntime::store_value_type_fields_Type()->domain_cc();
2570 const TypeFunc* new_tf = TypeFunc::make(tf->domain_sig(), tf->domain_cc(), tf->range_sig(), domain);
2571 call->_tf = new_tf;
2572 // Make sure the change of type is applied before projections are processed by igvn
2573 _igvn.set_type(call, call->Value(&_igvn));
2574 _igvn.set_type(ret, ret->Value(&_igvn));
2575
2576 // Before any new projection is added:
2577 CallProjections* projs = call->extract_projections(true, true);
2578
2579 Node* ctl = new Node(1);
2580 Node* mem = new Node(1);
2581 Node* io = new Node(1);
2582 Node* ex_ctl = new Node(1);
2583 Node* ex_mem = new Node(1);
2584 Node* ex_io = new Node(1);
2585 Node* res = new Node(1);
2586
2587 Node* cast = transform_later(new CastP2XNode(ctl, res));
2588 Node* mask = MakeConX(0x1);
2589 Node* masked = transform_later(new AndXNode(cast, mask));
2590 Node* cmp = transform_later(new CmpXNode(masked, mask));
2591 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
2592 IfNode* allocation_iff = new IfNode(ctl, bol, PROB_MAX, COUNT_UNKNOWN);
2593 transform_later(allocation_iff);
2594 Node* allocation_ctl = transform_later(new IfTrueNode(allocation_iff));
2595 Node* no_allocation_ctl = transform_later(new IfFalseNode(allocation_iff));
2596
2597 Node* no_allocation_res = transform_later(new CheckCastPPNode(no_allocation_ctl, res, TypeInstPtr::BOTTOM));
2598
2599 Node* mask2 = MakeConX(-2);
2600 Node* masked2 = transform_later(new AndXNode(cast, mask2));
2601 Node* rawklassptr = transform_later(new CastX2PNode(masked2));
2602 Node* klass_node = transform_later(new CheckCastPPNode(allocation_ctl, rawklassptr, TypeKlassPtr::OBJECT_OR_NULL));
2603
2604 Node* slowpath_bol = NULL;
2605 Node* top_adr = NULL;
2606 Node* old_top = NULL;
2607 Node* new_top = NULL;
2608 if (UseTLAB) {
2609 Node* end_adr = NULL;
2610 set_eden_pointers(top_adr, end_adr);
2611 Node* end = make_load(ctl, mem, end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS);
2612 old_top = new LoadPNode(ctl, mem, top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered);
2613 transform_later(old_top);
2614 Node* layout_val = make_load(NULL, mem, klass_node, in_bytes(Klass::layout_helper_offset()), TypeInt::INT, T_INT);
2615 Node* size_in_bytes = ConvI2X(layout_val);
2616 new_top = new AddPNode(top(), old_top, size_in_bytes);
2617 transform_later(new_top);
2618 Node* slowpath_cmp = new CmpPNode(new_top, end);
2619 transform_later(slowpath_cmp);
2620 slowpath_bol = new BoolNode(slowpath_cmp, BoolTest::ge);
2621 transform_later(slowpath_bol);
2622 } else {
2623 slowpath_bol = intcon(1);
2624 top_adr = top();
2625 old_top = top();
2626 new_top = top();
2627 }
2628 IfNode* slowpath_iff = new IfNode(allocation_ctl, slowpath_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN);
2629 transform_later(slowpath_iff);
2630
2631 Node* slowpath_true = new IfTrueNode(slowpath_iff);
2632 transform_later(slowpath_true);
2633
2634 CallStaticJavaNode* slow_call = new CallStaticJavaNode(OptoRuntime::store_value_type_fields_Type(),
2635 StubRoutines::store_value_type_fields_to_buf(),
2636 "store_value_type_fields",
2637 call->jvms()->bci(),
2638 TypePtr::BOTTOM);
2639 slow_call->init_req(TypeFunc::Control, slowpath_true);
2640 slow_call->init_req(TypeFunc::Memory, mem);
2641 slow_call->init_req(TypeFunc::I_O, io);
2642 slow_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2643 slow_call->init_req(TypeFunc::ReturnAdr, call->in(TypeFunc::ReturnAdr));
2644 slow_call->init_req(TypeFunc::Parms, res);
2645
2646 Node* slow_ctl = transform_later(new ProjNode(slow_call, TypeFunc::Control));
2647 Node* slow_mem = transform_later(new ProjNode(slow_call, TypeFunc::Memory));
2648 Node* slow_io = transform_later(new ProjNode(slow_call, TypeFunc::I_O));
2649 Node* slow_res = transform_later(new ProjNode(slow_call, TypeFunc::Parms));
2650 Node* slow_catc = transform_later(new CatchNode(slow_ctl, slow_io, 2));
2651 Node* slow_norm = transform_later(new CatchProjNode(slow_catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci));
2652 Node* slow_excp = transform_later(new CatchProjNode(slow_catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci));
2653
2654 Node* ex_r = new RegionNode(3);
2655 Node* ex_mem_phi = new PhiNode(ex_r, Type::MEMORY, TypePtr::BOTTOM);
2656 Node* ex_io_phi = new PhiNode(ex_r, Type::ABIO);
2657 ex_r->init_req(1, slow_excp);
2658 ex_mem_phi->init_req(1, slow_mem);
2659 ex_io_phi->init_req(1, slow_io);
2660 ex_r->init_req(2, ex_ctl);
2661 ex_mem_phi->init_req(2, ex_mem);
2662 ex_io_phi->init_req(2, ex_io);
2663
2664 transform_later(ex_r);
2665 transform_later(ex_mem_phi);
2666 transform_later(ex_io_phi);
2667
2668 Node* slowpath_false = new IfFalseNode(slowpath_iff);
2669 transform_later(slowpath_false);
2670 Node* rawmem = new StorePNode(slowpath_false, mem, top_adr, TypeRawPtr::BOTTOM, new_top, MemNode::unordered);
2671 transform_later(rawmem);
2672 Node* mark_node = makecon(TypeRawPtr::make((address)markWord::always_locked_prototype().value()));
2673 rawmem = make_store(slowpath_false, rawmem, old_top, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
2674 rawmem = make_store(slowpath_false, rawmem, old_top, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
2675 if (UseCompressedClassPointers) {
2676 rawmem = make_store(slowpath_false, rawmem, old_top, oopDesc::klass_gap_offset_in_bytes(), intcon(0), T_INT);
2677 }
2678 Node* fixed_block = make_load(slowpath_false, rawmem, klass_node, in_bytes(InstanceKlass::adr_valueklass_fixed_block_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
2679 Node* pack_handler = make_load(slowpath_false, rawmem, fixed_block, in_bytes(ValueKlass::pack_handler_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
2680
2681 CallLeafNoFPNode* handler_call = new CallLeafNoFPNode(OptoRuntime::pack_value_type_Type(),
2682 NULL,
2683 "pack handler",
2684 TypeRawPtr::BOTTOM);
2685 handler_call->init_req(TypeFunc::Control, slowpath_false);
2686 handler_call->init_req(TypeFunc::Memory, rawmem);
2687 handler_call->init_req(TypeFunc::I_O, top());
2688 handler_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2689 handler_call->init_req(TypeFunc::ReturnAdr, top());
2690 handler_call->init_req(TypeFunc::Parms, pack_handler);
2691 handler_call->init_req(TypeFunc::Parms+1, old_top);
2692
2693 // We don't know how many values are returned. This assumes the
2694 // worst case, that all available registers are used.
2695 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) {
2696 if (domain->field_at(i) == Type::HALF) {
2697 slow_call->init_req(i, top());
2698 handler_call->init_req(i+1, top());
2699 continue;
2700 }
2701 Node* proj = transform_later(new ProjNode(call, i));
2702 slow_call->init_req(i, proj);
2703 handler_call->init_req(i+1, proj);
2704 }
2705
2706 // We can safepoint at that new call
2707 slow_call->copy_call_debug_info(&_igvn, call);
2708 transform_later(slow_call);
2709 transform_later(handler_call);
2710
2711 Node* handler_ctl = transform_later(new ProjNode(handler_call, TypeFunc::Control));
2712 rawmem = transform_later(new ProjNode(handler_call, TypeFunc::Memory));
2713 Node* slowpath_false_res = transform_later(new ProjNode(handler_call, TypeFunc::Parms));
2714
2715 MergeMemNode* slowpath_false_mem = MergeMemNode::make(mem);
2716 slowpath_false_mem->set_memory_at(Compile::AliasIdxRaw, rawmem);
2717 transform_later(slowpath_false_mem);
2718
2719 Node* r = new RegionNode(4);
2720 Node* mem_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM);
2721 Node* io_phi = new PhiNode(r, Type::ABIO);
2722 Node* res_phi = new PhiNode(r, TypeInstPtr::BOTTOM);
2723
2724 r->init_req(1, no_allocation_ctl);
2725 mem_phi->init_req(1, mem);
2726 io_phi->init_req(1, io);
2727 res_phi->init_req(1, no_allocation_res);
2728 r->init_req(2, slow_norm);
2729 mem_phi->init_req(2, slow_mem);
2730 io_phi->init_req(2, slow_io);
2731 res_phi->init_req(2, slow_res);
2732 r->init_req(3, handler_ctl);
2733 mem_phi->init_req(3, slowpath_false_mem);
2734 io_phi->init_req(3, io);
2735 res_phi->init_req(3, slowpath_false_res);
2736
2737 transform_later(r);
2738 transform_later(mem_phi);
2739 transform_later(io_phi);
2740 transform_later(res_phi);
2741
2742 assert(projs->nb_resproj == 1, "unexpected number of results");
2743 _igvn.replace_in_uses(projs->fallthrough_catchproj, r);
2744 _igvn.replace_in_uses(projs->fallthrough_memproj, mem_phi);
2745 _igvn.replace_in_uses(projs->fallthrough_ioproj, io_phi);
2746 _igvn.replace_in_uses(projs->resproj[0], res_phi);
2747 _igvn.replace_in_uses(projs->catchall_catchproj, ex_r);
2748 _igvn.replace_in_uses(projs->catchall_memproj, ex_mem_phi);
2749 _igvn.replace_in_uses(projs->catchall_ioproj, ex_io_phi);
2750
2751 _igvn.replace_node(ctl, projs->fallthrough_catchproj);
2752 _igvn.replace_node(mem, projs->fallthrough_memproj);
2753 _igvn.replace_node(io, projs->fallthrough_ioproj);
2754 _igvn.replace_node(res, projs->resproj[0]);
2755 _igvn.replace_node(ex_ctl, projs->catchall_catchproj);
2756 _igvn.replace_node(ex_mem, projs->catchall_memproj);
2757 _igvn.replace_node(ex_io, projs->catchall_ioproj);
2758 }
2759
2760 //---------------------------eliminate_macro_nodes----------------------
2761 // Eliminate scalar replaced allocations and associated locks.
2762 void PhaseMacroExpand::eliminate_macro_nodes() {
2763 if (C->macro_count() == 0)
2764 return;
2765
2766 // First, attempt to eliminate locks
2767 int cnt = C->macro_count();
2768 for (int i=0; i < cnt; i++) {
2769 Node *n = C->macro_node(i);
2770 if (n->is_AbstractLock()) { // Lock and Unlock nodes
2771 // Before elimination mark all associated (same box and obj)
2772 // lock and unlock nodes.
2773 mark_eliminated_locking_nodes(n->as_AbstractLock());
2774 }
2775 }
2776 bool progress = true;
2777 while (progress) {
2778 progress = false;
2779 for (int i = C->macro_count(); i > 0; i--) {
2780 Node * n = C->macro_node(i-1);
2781 bool success = false;
2782 debug_only(int old_macro_count = C->macro_count(););
2783 if (n->is_AbstractLock()) {
2784 success = eliminate_locking_node(n->as_AbstractLock());
2785 }
2786 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2787 progress = progress || success;
2788 }
2789 }
2790 // Next, attempt to eliminate allocations
2791 _has_locks = false;
2792 progress = true;
2793 while (progress) {
2794 progress = false;
2795 for (int i = C->macro_count(); i > 0; i--) {
2796 Node * n = C->macro_node(i-1);
2797 bool success = false;
2798 debug_only(int old_macro_count = C->macro_count(););
2799 switch (n->class_id()) {
2800 case Node::Class_Allocate:
2801 case Node::Class_AllocateArray:
2802 success = eliminate_allocate_node(n->as_Allocate());
2803 break;
2804 case Node::Class_CallStaticJava: {
2805 CallStaticJavaNode* call = n->as_CallStaticJava();
2806 if (!call->method()->is_method_handle_intrinsic()) {
2807 success = eliminate_boxing_node(n->as_CallStaticJava());
2808 }
2809 break;
2810 }
2811 case Node::Class_Lock:
2812 case Node::Class_Unlock:
2813 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2814 _has_locks = true;
2815 break;
2816 case Node::Class_ArrayCopy:
2817 break;
2818 case Node::Class_OuterStripMinedLoop:
2819 break;
2820 default:
2821 assert(n->Opcode() == Op_LoopLimit ||
2822 n->Opcode() == Op_Opaque1 ||
2823 n->Opcode() == Op_Opaque2 ||
2824 n->Opcode() == Op_Opaque3 ||
2825 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n),
2826 "unknown node type in macro list");
2827 }
2828 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2829 progress = progress || success;
2830 }
2831 }
2832 }
2833
2834 //------------------------------expand_macro_nodes----------------------
2835 // Returns true if a failure occurred.
2836 bool PhaseMacroExpand::expand_macro_nodes() {
2837 // Last attempt to eliminate macro nodes.
2838 eliminate_macro_nodes();
2839
2840 // Make sure expansion will not cause node limit to be exceeded.
2841 // Worst case is a macro node gets expanded into about 200 nodes.
2842 // Allow 50% more for optimization.
2843 if (C->check_node_count(C->macro_count() * 300, "out of nodes before macro expansion" ) )
2844 return true;
2845
2846 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2847 bool progress = true;
2848 while (progress) {
2849 progress = false;
2850 for (int i = C->macro_count(); i > 0; i--) {
2851 Node * n = C->macro_node(i-1);
2852 bool success = false;
2853 debug_only(int old_macro_count = C->macro_count(););
2854 if (n->Opcode() == Op_LoopLimit) {
2855 // Remove it from macro list and put on IGVN worklist to optimize.
2856 C->remove_macro_node(n);
2857 _igvn._worklist.push(n);
2858 success = true;
2859 } else if (n->Opcode() == Op_CallStaticJava) {
2860 CallStaticJavaNode* call = n->as_CallStaticJava();
2861 if (!call->method()->is_method_handle_intrinsic()) {
2862 // Remove it from macro list and put on IGVN worklist to optimize.
2863 C->remove_macro_node(n);
2864 _igvn._worklist.push(n);
2865 success = true;
2866 }
2867 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) {
2868 _igvn.replace_node(n, n->in(1));
2869 success = true;
2870 #if INCLUDE_RTM_OPT
2871 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) {
2872 assert(C->profile_rtm(), "should be used only in rtm deoptimization code");
2873 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), "");
2874 Node* cmp = n->unique_out();
2875 #ifdef ASSERT
2876 // Validate graph.
2877 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), "");
2878 BoolNode* bol = cmp->unique_out()->as_Bool();
2879 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() &&
2880 (bol->_test._test == BoolTest::ne), "");
2881 IfNode* ifn = bol->unique_out()->as_If();
2882 assert((ifn->outcnt() == 2) &&
2883 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != NULL, "");
2884 #endif
2885 Node* repl = n->in(1);
2886 if (!_has_locks) {
2887 // Remove RTM state check if there are no locks in the code.
2888 // Replace input to compare the same value.
2889 repl = (cmp->in(1) == n) ? cmp->in(2) : cmp->in(1);
2890 }
2891 _igvn.replace_node(n, repl);
2892 success = true;
2893 #endif
2894 } else if (n->Opcode() == Op_OuterStripMinedLoop) {
2895 n->as_OuterStripMinedLoop()->adjust_strip_mined_loop(&_igvn);
2896 C->remove_macro_node(n);
2897 success = true;
2898 }
2899 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2900 progress = progress || success;
2901 }
2902 }
2903
2904 // expand arraycopy "macro" nodes first
2905 // For ReduceBulkZeroing, we must first process all arraycopy nodes
2906 // before the allocate nodes are expanded.
2907 int macro_idx = C->macro_count() - 1;
2908 while (macro_idx >= 0) {
2909 Node * n = C->macro_node(macro_idx);
2910 assert(n->is_macro(), "only macro nodes expected here");
2911 if (_igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())) {
2912 // node is unreachable, so don't try to expand it
2913 C->remove_macro_node(n);
2914 } else if (n->is_ArrayCopy()){
2915 int macro_count = C->macro_count();
2916 expand_arraycopy_node(n->as_ArrayCopy());
2917 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2918 }
2919 if (C->failing()) return true;
2920 macro_idx --;
2921 }
2922
2923 // expand "macro" nodes
2924 // nodes are removed from the macro list as they are processed
2925 while (C->macro_count() > 0) {
2926 int macro_count = C->macro_count();
2927 Node * n = C->macro_node(macro_count-1);
2928 assert(n->is_macro(), "only macro nodes expected here");
2929 if (_igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())) {
2930 // node is unreachable, so don't try to expand it
2931 C->remove_macro_node(n);
2932 continue;
2933 }
2934 switch (n->class_id()) {
2935 case Node::Class_Allocate:
2936 expand_allocate(n->as_Allocate());
2937 break;
2938 case Node::Class_AllocateArray:
2939 expand_allocate_array(n->as_AllocateArray());
2940 break;
2941 case Node::Class_Lock:
2942 expand_lock_node(n->as_Lock());
2943 break;
2944 case Node::Class_Unlock:
2945 expand_unlock_node(n->as_Unlock());
2946 break;
2947 case Node::Class_CallStaticJava:
2948 expand_mh_intrinsic_return(n->as_CallStaticJava());
2949 C->remove_macro_node(n);
2950 break;
2951 default:
2952 assert(false, "unknown node type in macro list");
2953 }
2954 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2955 if (C->failing()) return true;
2956 }
2957
2958 _igvn.set_delay_transform(false);
2959 _igvn.optimize();
2960 if (C->failing()) return true;
2961 return false;
2962 }