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 void PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) {
84 // Copy debug information and adjust JVMState information
85 uint old_dbg_start = oldcall->tf()->domain_sig()->cnt();
86 uint new_dbg_start = newcall->tf()->domain_sig()->cnt();
87 int jvms_adj = new_dbg_start - old_dbg_start;
88 assert (new_dbg_start == newcall->req(), "argument count mismatch");
89
90 // SafePointScalarObject node could be referenced several times in debug info.
91 // Use Dict to record cloned nodes.
92 Dict* sosn_map = new Dict(cmpkey,hashkey);
93 for (uint i = old_dbg_start; i < oldcall->req(); i++) {
94 Node* old_in = oldcall->in(i);
95 // Clone old SafePointScalarObjectNodes, adjusting their field contents.
96 if (old_in != NULL && old_in->is_SafePointScalarObject()) {
97 SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject();
98 uint old_unique = C->unique();
99 Node* new_in = old_sosn->clone(sosn_map);
100 if (old_unique != C->unique()) { // New node?
101 new_in->set_req(0, C->root()); // reset control edge
102 new_in = transform_later(new_in); // Register new node.
103 }
104 old_in = new_in;
105 }
106 newcall->add_req(old_in);
107 }
108
109 // JVMS may be shared so clone it before we modify it
110 newcall->set_jvms(oldcall->jvms() != NULL ? oldcall->jvms()->clone_deep(C) : NULL);
111 for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) {
112 jvms->set_map(newcall);
113 jvms->set_locoff(jvms->locoff()+jvms_adj);
114 jvms->set_stkoff(jvms->stkoff()+jvms_adj);
115 jvms->set_monoff(jvms->monoff()+jvms_adj);
116 jvms->set_scloff(jvms->scloff()+jvms_adj);
117 jvms->set_endoff(jvms->endoff()+jvms_adj);
118 }
119 }
120
121 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
122 Node* cmp;
123 if (mask != 0) {
124 Node* and_node = transform_later(new AndXNode(word, MakeConX(mask)));
125 cmp = transform_later(new CmpXNode(and_node, MakeConX(bits)));
126 } else {
127 cmp = word;
128 }
129 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne));
130 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
131 transform_later(iff);
132
133 // Fast path taken.
134 Node *fast_taken = transform_later(new IfFalseNode(iff));
135
136 // Fast path not-taken, i.e. slow path
137 Node *slow_taken = transform_later(new IfTrueNode(iff));
138
139 if (return_fast_path) {
140 region->init_req(edge, slow_taken); // Capture slow-control
141 return fast_taken;
142 } else {
143 region->init_req(edge, fast_taken); // Capture fast-control
144 return slow_taken;
145 }
146 }
147
148 //--------------------copy_predefined_input_for_runtime_call--------------------
149 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
150 // Set fixed predefined input arguments
151 call->init_req( TypeFunc::Control, ctrl );
152 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) );
153 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
154 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
155 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
156 }
157
158 //------------------------------make_slow_call---------------------------------
159 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type,
160 address slow_call, const char* leaf_name, Node* slow_path,
161 Node* parm0, Node* parm1, Node* parm2) {
162
163 // Slow-path call
164 CallNode *call = leaf_name
165 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
166 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM );
167
168 // Slow path call has no side-effects, uses few values
169 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
170 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0);
171 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1);
172 if (parm2 != NULL) call->init_req(TypeFunc::Parms+2, parm2);
173 copy_call_debug_info(oldcall, call);
174 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
175 _igvn.replace_node(oldcall, call);
176 transform_later(call);
177
178 return call;
179 }
180
181 void PhaseMacroExpand::extract_call_projections(CallNode *call) {
182 _fallthroughproj = NULL;
183 _fallthroughcatchproj = NULL;
184 _ioproj_fallthrough = NULL;
185 _ioproj_catchall = NULL;
186 _catchallcatchproj = NULL;
187 _memproj_fallthrough = NULL;
188 _memproj_catchall = NULL;
189 _resproj = NULL;
190 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) {
191 ProjNode *pn = call->fast_out(i)->as_Proj();
192 switch (pn->_con) {
193 case TypeFunc::Control:
194 {
195 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj
196 _fallthroughproj = pn;
197 DUIterator_Fast jmax, j = pn->fast_outs(jmax);
198 const Node *cn = pn->fast_out(j);
199 if (cn->is_Catch()) {
200 ProjNode *cpn = NULL;
201 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) {
202 cpn = cn->fast_out(k)->as_Proj();
203 assert(cpn->is_CatchProj(), "must be a CatchProjNode");
204 if (cpn->_con == CatchProjNode::fall_through_index)
205 _fallthroughcatchproj = cpn;
206 else {
207 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index.");
208 _catchallcatchproj = cpn;
209 }
210 }
211 }
212 break;
213 }
214 case TypeFunc::I_O:
215 if (pn->_is_io_use)
216 _ioproj_catchall = pn;
217 else
218 _ioproj_fallthrough = pn;
219 break;
220 case TypeFunc::Memory:
221 if (pn->_is_io_use)
222 _memproj_catchall = pn;
223 else
224 _memproj_fallthrough = pn;
225 break;
226 case TypeFunc::Parms:
227 _resproj = pn;
228 break;
229 default:
230 assert(false, "unexpected projection from allocation node.");
231 }
232 }
233
234 }
235
236 void PhaseMacroExpand::eliminate_gc_barrier(Node* p2x) {
237 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
238 bs->eliminate_gc_barrier(this, p2x);
239 }
240
241 // Search for a memory operation for the specified memory slice.
242 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
243 Node *orig_mem = mem;
244 Node *alloc_mem = alloc->in(TypeFunc::Memory);
245 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
246 while (true) {
247 if (mem == alloc_mem || mem == start_mem ) {
248 return mem; // hit one of our sentinels
249 } else if (mem->is_MergeMem()) {
250 mem = mem->as_MergeMem()->memory_at(alias_idx);
251 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
252 Node *in = mem->in(0);
253 // we can safely skip over safepoints, calls, locks and membars because we
254 // already know that the object is safe to eliminate.
255 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
256 return in;
257 } else if (in->is_Call()) {
258 CallNode *call = in->as_Call();
259 if (call->may_modify(tinst, phase)) {
260 assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape");
261 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) {
262 return in;
263 }
264 }
265 mem = in->in(TypeFunc::Memory);
266 } else if (in->is_MemBar()) {
267 ArrayCopyNode* ac = NULL;
268 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) {
269 assert(ac != NULL && ac->is_clonebasic(), "Only basic clone is a non escaping clone");
270 return ac;
271 }
272 mem = in->in(TypeFunc::Memory);
273 } else {
274 assert(false, "unexpected projection");
275 }
276 } else if (mem->is_Store()) {
277 const TypePtr* atype = mem->as_Store()->adr_type();
278 int adr_idx = phase->C->get_alias_index(atype);
279 if (adr_idx == alias_idx) {
280 assert(atype->isa_oopptr(), "address type must be oopptr");
281 int adr_offset = atype->flattened_offset();
282 uint adr_iid = atype->is_oopptr()->instance_id();
283 // Array elements references have the same alias_idx
284 // but different offset and different instance_id.
285 if (adr_offset == offset && adr_iid == alloc->_idx)
286 return mem;
287 } else {
288 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
289 }
290 mem = mem->in(MemNode::Memory);
291 } else if (mem->is_ClearArray()) {
292 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
293 // Can not bypass initialization of the instance
294 // we are looking.
295 debug_only(intptr_t offset;)
296 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
297 InitializeNode* init = alloc->as_Allocate()->initialization();
298 // We are looking for stored value, return Initialize node
299 // or memory edge from Allocate node.
300 if (init != NULL)
301 return init;
302 else
303 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers).
304 }
305 // Otherwise skip it (the call updated 'mem' value).
306 } else if (mem->Opcode() == Op_SCMemProj) {
307 mem = mem->in(0);
308 Node* adr = NULL;
309 if (mem->is_LoadStore()) {
310 adr = mem->in(MemNode::Address);
311 } else {
312 assert(mem->Opcode() == Op_EncodeISOArray ||
313 mem->Opcode() == Op_StrCompressedCopy, "sanity");
314 adr = mem->in(3); // Destination array
315 }
316 const TypePtr* atype = adr->bottom_type()->is_ptr();
317 int adr_idx = phase->C->get_alias_index(atype);
318 if (adr_idx == alias_idx) {
319 DEBUG_ONLY(mem->dump();)
320 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
321 return NULL;
322 }
323 mem = mem->in(MemNode::Memory);
324 } else if (mem->Opcode() == Op_StrInflatedCopy) {
325 Node* adr = mem->in(3); // Destination array
326 const TypePtr* atype = adr->bottom_type()->is_ptr();
327 int adr_idx = phase->C->get_alias_index(atype);
328 if (adr_idx == alias_idx) {
329 DEBUG_ONLY(mem->dump();)
330 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
331 return NULL;
332 }
333 mem = mem->in(MemNode::Memory);
334 } else {
335 return mem;
336 }
337 assert(mem != orig_mem, "dead memory loop");
338 }
339 }
340
341 // Generate loads from source of the arraycopy for fields of
342 // destination needed at a deoptimization point
343 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) {
344 BasicType bt = ft;
345 const Type *type = ftype;
346 if (ft == T_NARROWOOP) {
347 bt = T_OBJECT;
348 type = ftype->make_oopptr();
349 }
350 Node* res = NULL;
351 if (ac->is_clonebasic()) {
352 assert(ac->in(ArrayCopyNode::Src) != ac->in(ArrayCopyNode::Dest), "clone source equals destination");
353 Node* base = ac->in(ArrayCopyNode::Src)->in(AddPNode::Base);
354 Node* adr = _igvn.transform(new AddPNode(base, base, MakeConX(offset)));
355 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
356 res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::UnknownControl);
357 } else {
358 if (ac->modifies(offset, offset, &_igvn, true)) {
359 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
360 uint shift = exact_log2(type2aelembytes(bt));
361 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
362 #ifdef _LP64
363 diff = _igvn.transform(new ConvI2LNode(diff));
364 #endif
365 diff = _igvn.transform(new LShiftXNode(diff, intcon(shift)));
366
367 Node* off = _igvn.transform(new AddXNode(MakeConX(offset), diff));
368 Node* base = ac->in(ArrayCopyNode::Src);
369 Node* adr = _igvn.transform(new AddPNode(base, base, off));
370 const TypePtr* adr_type = _igvn.type(base)->is_ptr();
371 if (adr_type->isa_aryptr()) {
372 // In the case of a flattened value type array, each field has its
373 // own slice so we need to extract the field being accessed from
374 // the address computation
375 adr_type = adr_type->is_aryptr()->add_field_offset_and_offset(offset);
376 } else {
377 adr_type = adr_type->add_offset(offset);
378 }
379 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
380 // Don't emit a new load from src if src == dst but try to get the value from memory instead
381 return value_from_mem(ac->in(TypeFunc::Memory), ctl, ft, ftype, adr_type->isa_oopptr(), alloc);
382 }
383 if (adr_type->isa_aryptr()) {
384 adr = _igvn.transform(new CastPPNode(adr, adr_type));
385 }
386 res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::UnknownControl);
387 }
388 }
389 if (res != NULL) {
390 res = _igvn.transform(res);
391 if (ftype->isa_narrowoop()) {
392 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
393 assert(res->isa_DecodeN(), "should be narrow oop");
394 res = _igvn.transform(new EncodePNode(res, ftype));
395 }
396 return res;
397 }
398 return NULL;
399 }
400
401 //
402 // Given a Memory Phi, compute a value Phi containing the values from stores
403 // on the input paths.
404 // Note: this function is recursive, its depth is limited by the "level" argument
405 // Returns the computed Phi, or NULL if it cannot compute it.
406 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) {
407 assert(mem->is_Phi(), "sanity");
408 int alias_idx = C->get_alias_index(adr_t);
409 int offset = adr_t->flattened_offset();
410 int instance_id = adr_t->instance_id();
411
412 // Check if an appropriate value phi already exists.
413 Node* region = mem->in(0);
414 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
415 Node* phi = region->fast_out(k);
416 if (phi->is_Phi() && phi != mem &&
417 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) {
418 return phi;
419 }
420 }
421 // Check if an appropriate new value phi already exists.
422 Node* new_phi = value_phis->find(mem->_idx);
423 if (new_phi != NULL)
424 return new_phi;
425
426 if (level <= 0) {
427 return NULL; // Give up: phi tree too deep
428 }
429 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
430 Node *alloc_mem = alloc->in(TypeFunc::Memory);
431
432 uint length = mem->req();
433 GrowableArray <Node *> values(length, length, NULL, false);
434
435 // create a new Phi for the value
436 PhiNode *phi = new PhiNode(mem->in(0), phi_type, NULL, mem->_idx, instance_id, alias_idx, offset);
437 transform_later(phi);
438 value_phis->push(phi, mem->_idx);
439
440 for (uint j = 1; j < length; j++) {
441 Node *in = mem->in(j);
442 if (in == NULL || in->is_top()) {
443 values.at_put(j, in);
444 } else {
445 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
446 if (val == start_mem || val == alloc_mem) {
447 // hit a sentinel, return appropriate 0 value
448 Node* default_value = alloc->in(AllocateNode::DefaultValue);
449 if (default_value != NULL) {
450 values.at_put(j, default_value);
451 } else {
452 assert(alloc->in(AllocateNode::RawDefaultValue) == NULL, "default value may not be null");
453 values.at_put(j, _igvn.zerocon(ft));
454 }
455 continue;
456 }
457 if (val->is_Initialize()) {
458 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
459 }
460 if (val == NULL) {
461 return NULL; // can't find a value on this path
462 }
463 if (val == mem) {
464 values.at_put(j, mem);
465 } else if (val->is_Store()) {
466 Node* n = val->in(MemNode::ValueIn);
467 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
468 n = bs->step_over_gc_barrier(n);
469 values.at_put(j, n);
470 } else if(val->is_Proj() && val->in(0) == alloc) {
471 Node* default_value = alloc->in(AllocateNode::DefaultValue);
472 if (default_value != NULL) {
473 values.at_put(j, default_value);
474 } else {
475 assert(alloc->in(AllocateNode::RawDefaultValue) == NULL, "default value may not be null");
476 values.at_put(j, _igvn.zerocon(ft));
477 }
478 } else if (val->is_Phi()) {
479 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
480 if (val == NULL) {
481 return NULL;
482 }
483 values.at_put(j, val);
484 } else if (val->Opcode() == Op_SCMemProj) {
485 assert(val->in(0)->is_LoadStore() ||
486 val->in(0)->Opcode() == Op_EncodeISOArray ||
487 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity");
488 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
489 return NULL;
490 } else if (val->is_ArrayCopy()) {
491 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc);
492 if (res == NULL) {
493 return NULL;
494 }
495 values.at_put(j, res);
496 } else {
497 #ifdef ASSERT
498 val->dump();
499 assert(false, "unknown node on this path");
500 #endif
501 return NULL; // unknown node on this path
502 }
503 }
504 }
505 // Set Phi's inputs
506 for (uint j = 1; j < length; j++) {
507 if (values.at(j) == mem) {
508 phi->init_req(j, phi);
509 } else {
510 phi->init_req(j, values.at(j));
511 }
512 }
513 return phi;
514 }
515
516 // Search the last value stored into the object's field.
517 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) {
518 assert(adr_t->is_known_instance_field(), "instance required");
519 int instance_id = adr_t->instance_id();
520 assert((uint)instance_id == alloc->_idx, "wrong allocation");
521
522 int alias_idx = C->get_alias_index(adr_t);
523 int offset = adr_t->flattened_offset();
524 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
525 Node *alloc_mem = alloc->in(TypeFunc::Memory);
526 Arena *a = Thread::current()->resource_area();
527 VectorSet visited(a);
528
529 bool done = sfpt_mem == alloc_mem;
530 Node *mem = sfpt_mem;
531 while (!done) {
532 if (visited.test_set(mem->_idx)) {
533 return NULL; // found a loop, give up
534 }
535 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
536 if (mem == start_mem || mem == alloc_mem) {
537 done = true; // hit a sentinel, return appropriate 0 value
538 } else if (mem->is_Initialize()) {
539 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
540 if (mem == NULL) {
541 done = true; // Something went wrong.
542 } else if (mem->is_Store()) {
543 const TypePtr* atype = mem->as_Store()->adr_type();
544 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
545 done = true;
546 }
547 } else if (mem->is_Store()) {
548 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
549 assert(atype != NULL, "address type must be oopptr");
550 assert(C->get_alias_index(atype) == alias_idx &&
551 atype->is_known_instance_field() && atype->flattened_offset() == offset &&
552 atype->instance_id() == instance_id, "store is correct memory slice");
553 done = true;
554 } else if (mem->is_Phi()) {
555 // try to find a phi's unique input
556 Node *unique_input = NULL;
557 Node *top = C->top();
558 for (uint i = 1; i < mem->req(); i++) {
559 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
560 if (n == NULL || n == top || n == mem) {
561 continue;
562 } else if (unique_input == NULL) {
563 unique_input = n;
564 } else if (unique_input != n) {
565 unique_input = top;
566 break;
567 }
568 }
569 if (unique_input != NULL && unique_input != top) {
570 mem = unique_input;
571 } else {
572 done = true;
573 }
574 } else if (mem->is_ArrayCopy()) {
575 done = true;
576 } else {
577 assert(false, "unexpected node");
578 }
579 }
580 if (mem != NULL) {
581 if (mem == start_mem || mem == alloc_mem) {
582 // hit a sentinel, return appropriate 0 value
583 Node* default_value = alloc->in(AllocateNode::DefaultValue);
584 if (default_value != NULL) {
585 return default_value;
586 }
587 assert(alloc->in(AllocateNode::RawDefaultValue) == NULL, "default value may not be null");
588 return _igvn.zerocon(ft);
589 } else if (mem->is_Store()) {
590 Node* n = mem->in(MemNode::ValueIn);
591 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
592 n = bs->step_over_gc_barrier(n);
593 return n;
594 } else if (mem->is_Phi()) {
595 // attempt to produce a Phi reflecting the values on the input paths of the Phi
596 Node_Stack value_phis(a, 8);
597 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
598 if (phi != NULL) {
599 return phi;
600 } else {
601 // Kill all new Phis
602 while(value_phis.is_nonempty()) {
603 Node* n = value_phis.node();
604 _igvn.replace_node(n, C->top());
605 value_phis.pop();
606 }
607 }
608 } else if (mem->is_ArrayCopy()) {
609 Node* ctl = mem->in(0);
610 Node* m = mem->in(TypeFunc::Memory);
611 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) {
612 // pin the loads in the uncommon trap path
613 ctl = sfpt_ctl;
614 m = sfpt_mem;
615 }
616 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc);
617 }
618 }
619 // Something went wrong.
620 return NULL;
621 }
622
623 // Search the last value stored into the value type's fields.
624 Node* PhaseMacroExpand::value_type_from_mem(Node* mem, Node* ctl, ciValueKlass* vk, const TypeAryPtr* adr_type, int offset, AllocateNode* alloc) {
625 // Subtract the offset of the first field to account for the missing oop header
626 offset -= vk->first_field_offset();
627 // Create a new ValueTypeNode and retrieve the field values from memory
628 ValueTypeNode* vt = ValueTypeNode::make_uninitialized(_igvn, vk)->as_ValueType();
629 for (int i = 0; i < vk->nof_declared_nonstatic_fields(); ++i) {
630 ciType* field_type = vt->field_type(i);
631 int field_offset = offset + vt->field_offset(i);
632 // Each value type field has its own memory slice
633 adr_type = adr_type->with_field_offset(field_offset);
634 Node* value = NULL;
635 if (vt->field_is_flattened(i)) {
636 value = value_type_from_mem(mem, ctl, field_type->as_value_klass(), adr_type, field_offset, alloc);
637 } else {
638 const Type* ft = Type::get_const_type(field_type);
639 BasicType bt = field_type->basic_type();
640 if (UseCompressedOops && !is_java_primitive(bt)) {
641 ft = ft->make_narrowoop();
642 bt = T_NARROWOOP;
643 }
644 value = value_from_mem(mem, ctl, bt, ft, adr_type, alloc);
645 if (value != NULL && ft->isa_narrowoop()) {
646 assert(UseCompressedOops, "unexpected narrow oop");
647 value = transform_later(new DecodeNNode(value, value->get_ptr_type()));
648 }
649 }
650 if (value != NULL) {
651 vt->set_field_value(i, value);
652 } else {
653 // We might have reached the TrackedInitializationLimit
654 return NULL;
655 }
656 }
657 return transform_later(vt);
658 }
659
660 // Check the possibility of scalar replacement.
661 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
662 // Scan the uses of the allocation to check for anything that would
663 // prevent us from eliminating it.
664 NOT_PRODUCT( const char* fail_eliminate = NULL; )
665 DEBUG_ONLY( Node* disq_node = NULL; )
666 bool can_eliminate = true;
667
668 Node* res = alloc->result_cast();
669 const TypeOopPtr* res_type = NULL;
670 if (res == NULL) {
671 // All users were eliminated.
672 } else if (!res->is_CheckCastPP()) {
673 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
674 can_eliminate = false;
675 } else {
676 res_type = _igvn.type(res)->isa_oopptr();
677 if (res_type == NULL) {
678 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
679 can_eliminate = false;
680 } else if (res_type->isa_aryptr()) {
681 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
682 if (length < 0) {
683 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
684 can_eliminate = false;
685 }
686 }
687 }
688
689 if (can_eliminate && res != NULL) {
690 for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
691 j < jmax && can_eliminate; j++) {
692 Node* use = res->fast_out(j);
693
694 if (use->is_AddP()) {
695 const TypePtr* addp_type = _igvn.type(use)->is_ptr();
696 int offset = addp_type->offset();
697
698 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
699 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";)
700 can_eliminate = false;
701 break;
702 }
703 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
704 k < kmax && can_eliminate; k++) {
705 Node* n = use->fast_out(k);
706 if (!n->is_Store() && n->Opcode() != Op_CastP2X &&
707 SHENANDOAHGC_ONLY((!UseShenandoahGC || !ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(n)) &&)
708 !(n->is_ArrayCopy() &&
709 n->as_ArrayCopy()->is_clonebasic() &&
710 n->in(ArrayCopyNode::Dest) == use)) {
711 DEBUG_ONLY(disq_node = n;)
712 if (n->is_Load() || n->is_LoadStore()) {
713 NOT_PRODUCT(fail_eliminate = "Field load";)
714 } else {
715 NOT_PRODUCT(fail_eliminate = "Not store field reference";)
716 }
717 can_eliminate = false;
718 }
719 }
720 } else if (use->is_ArrayCopy() &&
721 (use->as_ArrayCopy()->is_arraycopy_validated() ||
722 use->as_ArrayCopy()->is_copyof_validated() ||
723 use->as_ArrayCopy()->is_copyofrange_validated()) &&
724 use->in(ArrayCopyNode::Dest) == res) {
725 // ok to eliminate
726 } else if (use->is_SafePoint()) {
727 SafePointNode* sfpt = use->as_SafePoint();
728 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
729 // Object is passed as argument.
730 DEBUG_ONLY(disq_node = use;)
731 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
732 can_eliminate = false;
733 }
734 Node* sfptMem = sfpt->memory();
735 if (sfptMem == NULL || sfptMem->is_top()) {
736 DEBUG_ONLY(disq_node = use;)
737 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";)
738 can_eliminate = false;
739 } else {
740 safepoints.append_if_missing(sfpt);
741 }
742 } else if (use->is_ValueType() && use->isa_ValueType()->get_oop() == res) {
743 // ok to eliminate
744 } else if (use->is_Store()) {
745 // store to mark work
746 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
747 if (use->is_Phi()) {
748 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
749 NOT_PRODUCT(fail_eliminate = "Object is return value";)
750 } else {
751 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
752 }
753 DEBUG_ONLY(disq_node = use;)
754 } else {
755 if (use->Opcode() == Op_Return) {
756 NOT_PRODUCT(fail_eliminate = "Object is return value";)
757 } else {
758 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
759 }
760 DEBUG_ONLY(disq_node = use;)
761 }
762 can_eliminate = false;
763 } else {
764 assert(use->Opcode() == Op_CastP2X, "should be");
765 assert(!use->has_out_with(Op_OrL), "should have been removed because oop is never null");
766 }
767 }
768 }
769
770 #ifndef PRODUCT
771 if (PrintEliminateAllocations) {
772 if (can_eliminate) {
773 tty->print("Scalar ");
774 if (res == NULL)
775 alloc->dump();
776 else
777 res->dump();
778 } else if (alloc->_is_scalar_replaceable) {
779 tty->print("NotScalar (%s)", fail_eliminate);
780 if (res == NULL)
781 alloc->dump();
782 else
783 res->dump();
784 #ifdef ASSERT
785 if (disq_node != NULL) {
786 tty->print(" >>>> ");
787 disq_node->dump();
788 }
789 #endif /*ASSERT*/
790 }
791 }
792 #endif
793 return can_eliminate;
794 }
795
796 // Do scalar replacement.
797 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
798 GrowableArray <SafePointNode *> safepoints_done;
799
800 ciKlass* klass = NULL;
801 ciInstanceKlass* iklass = NULL;
802 int nfields = 0;
803 int array_base = 0;
804 int element_size = 0;
805 BasicType basic_elem_type = T_ILLEGAL;
806 ciType* elem_type = NULL;
807
808 Node* res = alloc->result_cast();
809 assert(res == NULL || res->is_CheckCastPP(), "unexpected AllocateNode result");
810 const TypeOopPtr* res_type = NULL;
811 if (res != NULL) { // Could be NULL when there are no users
812 res_type = _igvn.type(res)->isa_oopptr();
813 }
814
815 if (res != NULL) {
816 klass = res_type->klass();
817 if (res_type->isa_instptr()) {
818 // find the fields of the class which will be needed for safepoint debug information
819 assert(klass->is_instance_klass(), "must be an instance klass.");
820 iklass = klass->as_instance_klass();
821 nfields = iklass->nof_nonstatic_fields();
822 } else {
823 // find the array's elements which will be needed for safepoint debug information
824 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
825 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass.");
826 elem_type = klass->as_array_klass()->element_type();
827 basic_elem_type = elem_type->basic_type();
828 if (elem_type->is_valuetype() && !klass->is_value_array_klass()) {
829 assert(basic_elem_type == T_VALUETYPE, "unexpected element basic type");
830 basic_elem_type = T_OBJECT;
831 }
832 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
833 element_size = type2aelembytes(basic_elem_type);
834 if (klass->is_value_array_klass()) {
835 // Flattened value type array
836 element_size = klass->as_value_array_klass()->element_byte_size();
837 }
838 }
839 }
840 //
841 // Process the safepoint uses
842 //
843 Unique_Node_List value_worklist;
844 while (safepoints.length() > 0) {
845 SafePointNode* sfpt = safepoints.pop();
846 Node* mem = sfpt->memory();
847 Node* ctl = sfpt->control();
848 assert(sfpt->jvms() != NULL, "missed JVMS");
849 // Fields of scalar objs are referenced only at the end
850 // of regular debuginfo at the last (youngest) JVMS.
851 // Record relative start index.
852 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
853 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type,
854 #ifdef ASSERT
855 alloc,
856 #endif
857 first_ind, nfields);
858 sobj->init_req(0, C->root());
859 transform_later(sobj);
860
861 // Scan object's fields adding an input to the safepoint for each field.
862 for (int j = 0; j < nfields; j++) {
863 intptr_t offset;
864 ciField* field = NULL;
865 if (iklass != NULL) {
866 field = iklass->nonstatic_field_at(j);
867 offset = field->offset();
868 elem_type = field->type();
869 basic_elem_type = field->layout_type();
870 assert(!field->is_flattened(), "flattened value type fields should not have safepoint uses");
871 } else {
872 offset = array_base + j * (intptr_t)element_size;
873 }
874
875 const Type *field_type;
876 // The next code is taken from Parse::do_get_xxx().
877 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) {
878 if (!elem_type->is_loaded()) {
879 field_type = TypeInstPtr::BOTTOM;
880 } else if (field != NULL && field->is_static_constant()) {
881 // This can happen if the constant oop is non-perm.
882 ciObject* con = field->constant_value().as_object();
883 // Do not "join" in the previous type; it doesn't add value,
884 // and may yield a vacuous result if the field is of interface type.
885 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
886 assert(field_type != NULL, "field singleton type must be consistent");
887 } else {
888 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
889 }
890 if (UseCompressedOops) {
891 field_type = field_type->make_narrowoop();
892 basic_elem_type = T_NARROWOOP;
893 }
894 } else {
895 field_type = Type::get_const_basic_type(basic_elem_type);
896 }
897
898 Node* field_val = NULL;
899 const TypeOopPtr* field_addr_type = res_type->add_offset(offset)->isa_oopptr();
900 if (klass->is_value_array_klass()) {
901 ciValueKlass* vk = elem_type->as_value_klass();
902 assert(vk->flatten_array(), "must be flattened");
903 field_val = value_type_from_mem(mem, ctl, vk, field_addr_type->isa_aryptr(), 0, alloc);
904 } else {
905 field_val = value_from_mem(mem, ctl, basic_elem_type, field_type, field_addr_type, alloc);
906 }
907 if (field_val == NULL) {
908 // We weren't able to find a value for this field,
909 // give up on eliminating this allocation.
910
911 // Remove any extra entries we added to the safepoint.
912 uint last = sfpt->req() - 1;
913 for (int k = 0; k < j; k++) {
914 sfpt->del_req(last--);
915 }
916 _igvn._worklist.push(sfpt);
917 // rollback processed safepoints
918 while (safepoints_done.length() > 0) {
919 SafePointNode* sfpt_done = safepoints_done.pop();
920 // remove any extra entries we added to the safepoint
921 last = sfpt_done->req() - 1;
922 for (int k = 0; k < nfields; k++) {
923 sfpt_done->del_req(last--);
924 }
925 JVMState *jvms = sfpt_done->jvms();
926 jvms->set_endoff(sfpt_done->req());
927 // Now make a pass over the debug information replacing any references
928 // to SafePointScalarObjectNode with the allocated object.
929 int start = jvms->debug_start();
930 int end = jvms->debug_end();
931 for (int i = start; i < end; i++) {
932 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
933 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
934 if (scobj->first_index(jvms) == sfpt_done->req() &&
935 scobj->n_fields() == (uint)nfields) {
936 assert(scobj->alloc() == alloc, "sanity");
937 sfpt_done->set_req(i, res);
938 }
939 }
940 }
941 _igvn._worklist.push(sfpt_done);
942 }
943 #ifndef PRODUCT
944 if (PrintEliminateAllocations) {
945 if (field != NULL) {
946 tty->print("=== At SafePoint node %d can't find value of Field: ",
947 sfpt->_idx);
948 field->print();
949 int field_idx = C->get_alias_index(field_addr_type);
950 tty->print(" (alias_idx=%d)", field_idx);
951 } else { // Array's element
952 tty->print("=== At SafePoint node %d can't find value of array element [%d]",
953 sfpt->_idx, j);
954 }
955 tty->print(", which prevents elimination of: ");
956 if (res == NULL)
957 alloc->dump();
958 else
959 res->dump();
960 }
961 #endif
962 return false;
963 }
964 if (UseCompressedOops && field_type->isa_narrowoop()) {
965 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
966 // to be able scalar replace the allocation.
967 if (field_val->is_EncodeP()) {
968 field_val = field_val->in(1);
969 } else {
970 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
971 }
972 } else if (field_val->is_ValueType()) {
973 // Keep track of value types to scalarize them later
974 value_worklist.push(field_val);
975 }
976 sfpt->add_req(field_val);
977 }
978 JVMState *jvms = sfpt->jvms();
979 jvms->set_endoff(sfpt->req());
980 // Now make a pass over the debug information replacing any references
981 // to the allocated object with "sobj"
982 int start = jvms->debug_start();
983 int end = jvms->debug_end();
984 sfpt->replace_edges_in_range(res, sobj, start, end);
985 _igvn._worklist.push(sfpt);
986 safepoints_done.append_if_missing(sfpt); // keep it for rollback
987 }
988 // Scalarize value types that were added to the safepoint
989 for (uint i = 0; i < value_worklist.size(); ++i) {
990 Node* vt = value_worklist.at(i);
991 vt->as_ValueType()->make_scalar_in_safepoints(&_igvn);
992 }
993 return true;
994 }
995
996 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
997 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control);
998 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory);
999 if (ctl_proj != NULL) {
1000 igvn.replace_node(ctl_proj, n->in(0));
1001 }
1002 if (mem_proj != NULL) {
1003 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
1004 }
1005 }
1006
1007 // Process users of eliminated allocation.
1008 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) {
1009 Node* res = alloc->result_cast();
1010 if (res != NULL) {
1011 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
1012 Node *use = res->last_out(j);
1013 uint oc1 = res->outcnt();
1014
1015 if (use->is_AddP()) {
1016 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
1017 Node *n = use->last_out(k);
1018 uint oc2 = use->outcnt();
1019 if (n->is_Store()) {
1020 #ifdef ASSERT
1021 // Verify that there is no dependent MemBarVolatile nodes,
1022 // they should be removed during IGVN, see MemBarNode::Ideal().
1023 for (DUIterator_Fast pmax, p = n->fast_outs(pmax);
1024 p < pmax; p++) {
1025 Node* mb = n->fast_out(p);
1026 assert(mb->is_Initialize() || !mb->is_MemBar() ||
1027 mb->req() <= MemBarNode::Precedent ||
1028 mb->in(MemBarNode::Precedent) != n,
1029 "MemBarVolatile should be eliminated for non-escaping object");
1030 }
1031 #endif
1032 _igvn.replace_node(n, n->in(MemNode::Memory));
1033 } else if (n->is_ArrayCopy()) {
1034 // Disconnect ArrayCopy node
1035 ArrayCopyNode* ac = n->as_ArrayCopy();
1036 assert(ac->is_clonebasic(), "unexpected array copy kind");
1037 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
1038 disconnect_projections(ac, _igvn);
1039 assert(alloc->in(0)->is_Proj() && alloc->in(0)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
1040 Node* membar_before = alloc->in(0)->in(0);
1041 disconnect_projections(membar_before->as_MemBar(), _igvn);
1042 if (membar_after->is_MemBar()) {
1043 disconnect_projections(membar_after->as_MemBar(), _igvn);
1044 }
1045 } else {
1046 eliminate_gc_barrier(n);
1047 }
1048 k -= (oc2 - use->outcnt());
1049 }
1050 _igvn.remove_dead_node(use);
1051 } else if (use->is_ArrayCopy()) {
1052 // Disconnect ArrayCopy node
1053 ArrayCopyNode* ac = use->as_ArrayCopy();
1054 assert(ac->is_arraycopy_validated() ||
1055 ac->is_copyof_validated() ||
1056 ac->is_copyofrange_validated(), "unsupported");
1057 CallProjections* callprojs = ac->extract_projections(true);
1058
1059 _igvn.replace_node(callprojs->fallthrough_ioproj, ac->in(TypeFunc::I_O));
1060 _igvn.replace_node(callprojs->fallthrough_memproj, ac->in(TypeFunc::Memory));
1061 _igvn.replace_node(callprojs->fallthrough_catchproj, ac->in(TypeFunc::Control));
1062
1063 // Set control to top. IGVN will remove the remaining projections
1064 ac->set_req(0, top());
1065 ac->replace_edge(res, top());
1066
1067 // Disconnect src right away: it can help find new
1068 // opportunities for allocation elimination
1069 Node* src = ac->in(ArrayCopyNode::Src);
1070 ac->replace_edge(src, top());
1071 // src can be top at this point if src and dest of the
1072 // arraycopy were the same
1073 if (src->outcnt() == 0 && !src->is_top()) {
1074 _igvn.remove_dead_node(src);
1075 }
1076
1077 _igvn._worklist.push(ac);
1078 } else if (use->is_ValueType()) {
1079 assert(use->isa_ValueType()->get_oop() == res, "unexpected value type use");
1080 _igvn.rehash_node_delayed(use);
1081 use->isa_ValueType()->set_oop(_igvn.zerocon(T_VALUETYPE));
1082 } else if (use->is_Store()) {
1083 _igvn.replace_node(use, use->in(MemNode::Memory));
1084 } else {
1085 eliminate_gc_barrier(use);
1086 }
1087 j -= (oc1 - res->outcnt());
1088 }
1089 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
1090 _igvn.remove_dead_node(res);
1091 }
1092
1093 //
1094 // Process other users of allocation's projections
1095 //
1096 if (_resproj != NULL && _resproj->outcnt() != 0) {
1097 // First disconnect stores captured by Initialize node.
1098 // If Initialize node is eliminated first in the following code,
1099 // it will kill such stores and DUIterator_Last will assert.
1100 for (DUIterator_Fast jmax, j = _resproj->fast_outs(jmax); j < jmax; j++) {
1101 Node *use = _resproj->fast_out(j);
1102 if (use->is_AddP()) {
1103 // raw memory addresses used only by the initialization
1104 _igvn.replace_node(use, C->top());
1105 --j; --jmax;
1106 }
1107 }
1108 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) {
1109 Node *use = _resproj->last_out(j);
1110 uint oc1 = _resproj->outcnt();
1111 if (use->is_Initialize()) {
1112 // Eliminate Initialize node.
1113 InitializeNode *init = use->as_Initialize();
1114 assert(init->outcnt() <= 2, "only a control and memory projection expected");
1115 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control);
1116 if (ctrl_proj != NULL) {
1117 _igvn.replace_node(ctrl_proj, init->in(TypeFunc::Control));
1118 #ifdef ASSERT
1119 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1120 Node* tmp = init->in(TypeFunc::Control);
1121 while (bs->is_gc_barrier_node(tmp)) {
1122 Node* tmp2 = bs->step_over_gc_barrier_ctrl(tmp);
1123 assert(tmp != tmp2, "Must make progress");
1124 tmp = tmp2;
1125 }
1126 assert(tmp == _fallthroughcatchproj, "allocation control projection");
1127 #endif
1128 }
1129 Node *mem_proj = init->proj_out_or_null(TypeFunc::Memory);
1130 if (mem_proj != NULL) {
1131 Node *mem = init->in(TypeFunc::Memory);
1132 #ifdef ASSERT
1133 if (mem->is_MergeMem()) {
1134 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection");
1135 } else {
1136 assert(mem == _memproj_fallthrough, "allocation memory projection");
1137 }
1138 #endif
1139 _igvn.replace_node(mem_proj, mem);
1140 }
1141 } else {
1142 assert(false, "only Initialize or AddP expected");
1143 }
1144 j -= (oc1 - _resproj->outcnt());
1145 }
1146 }
1147 if (_fallthroughcatchproj != NULL) {
1148 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control));
1149 }
1150 if (_memproj_fallthrough != NULL) {
1151 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory));
1152 }
1153 if (_memproj_catchall != NULL) {
1154 _igvn.replace_node(_memproj_catchall, C->top());
1155 }
1156 if (_ioproj_fallthrough != NULL) {
1157 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O));
1158 }
1159 if (_ioproj_catchall != NULL) {
1160 _igvn.replace_node(_ioproj_catchall, C->top());
1161 }
1162 if (_catchallcatchproj != NULL) {
1163 _igvn.replace_node(_catchallcatchproj, C->top());
1164 }
1165 }
1166
1167 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1168 // Don't do scalar replacement if the frame can be popped by JVMTI:
1169 // if reallocation fails during deoptimization we'll pop all
1170 // interpreter frames for this compiled frame and that won't play
1171 // nice with JVMTI popframe.
1172 if (!EliminateAllocations || JvmtiExport::can_pop_frame() || !alloc->_is_non_escaping) {
1173 return false;
1174 }
1175 Node* klass = alloc->in(AllocateNode::KlassNode);
1176 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1177 Node* res = alloc->result_cast();
1178 // Eliminate boxing allocations which are not used
1179 // regardless scalar replacable status.
1180 bool boxing_alloc = C->eliminate_boxing() &&
1181 tklass->klass()->is_instance_klass() &&
1182 tklass->klass()->as_instance_klass()->is_box_klass();
1183 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != NULL))) {
1184 return false;
1185 }
1186
1187 extract_call_projections(alloc);
1188
1189 GrowableArray <SafePointNode *> safepoints;
1190 if (!can_eliminate_allocation(alloc, safepoints)) {
1191 return false;
1192 }
1193
1194 if (!alloc->_is_scalar_replaceable) {
1195 assert(res == NULL, "sanity");
1196 // We can only eliminate allocation if all debug info references
1197 // are already replaced with SafePointScalarObject because
1198 // we can't search for a fields value without instance_id.
1199 if (safepoints.length() > 0) {
1200 return false;
1201 }
1202 }
1203
1204 if (!scalar_replacement(alloc, safepoints)) {
1205 return false;
1206 }
1207
1208 CompileLog* log = C->log();
1209 if (log != NULL) {
1210 log->head("eliminate_allocation type='%d'",
1211 log->identify(tklass->klass()));
1212 JVMState* p = alloc->jvms();
1213 while (p != NULL) {
1214 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1215 p = p->caller();
1216 }
1217 log->tail("eliminate_allocation");
1218 }
1219
1220 process_users_of_allocation(alloc);
1221
1222 #ifndef PRODUCT
1223 if (PrintEliminateAllocations) {
1224 if (alloc->is_AllocateArray())
1225 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1226 else
1227 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1228 }
1229 #endif
1230
1231 return true;
1232 }
1233
1234 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1235 // EA should remove all uses of non-escaping boxing node.
1236 if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != NULL) {
1237 return false;
1238 }
1239
1240 assert(boxing->result_cast() == NULL, "unexpected boxing node result");
1241
1242 extract_call_projections(boxing);
1243
1244 const TypeTuple* r = boxing->tf()->range_sig();
1245 assert(r->cnt() > TypeFunc::Parms, "sanity");
1246 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1247 assert(t != NULL, "sanity");
1248
1249 CompileLog* log = C->log();
1250 if (log != NULL) {
1251 log->head("eliminate_boxing type='%d'",
1252 log->identify(t->klass()));
1253 JVMState* p = boxing->jvms();
1254 while (p != NULL) {
1255 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1256 p = p->caller();
1257 }
1258 log->tail("eliminate_boxing");
1259 }
1260
1261 process_users_of_allocation(boxing);
1262
1263 #ifndef PRODUCT
1264 if (PrintEliminateAllocations) {
1265 tty->print("++++ Eliminated: %d ", boxing->_idx);
1266 boxing->method()->print_short_name(tty);
1267 tty->cr();
1268 }
1269 #endif
1270
1271 return true;
1272 }
1273
1274 //---------------------------set_eden_pointers-------------------------
1275 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) {
1276 if (UseTLAB) { // Private allocation: load from TLS
1277 Node* thread = transform_later(new ThreadLocalNode());
1278 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset());
1279 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset());
1280 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset);
1281 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset);
1282 } else { // Shared allocation: load from globals
1283 CollectedHeap* ch = Universe::heap();
1284 address top_adr = (address)ch->top_addr();
1285 address end_adr = (address)ch->end_addr();
1286 eden_top_adr = makecon(TypeRawPtr::make(top_adr));
1287 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr);
1288 }
1289 }
1290
1291
1292 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
1293 Node* adr = basic_plus_adr(base, offset);
1294 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
1295 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered);
1296 transform_later(value);
1297 return value;
1298 }
1299
1300
1301 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
1302 Node* adr = basic_plus_adr(base, offset);
1303 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt, MemNode::unordered);
1304 transform_later(mem);
1305 return mem;
1306 }
1307
1308 //=============================================================================
1309 //
1310 // A L L O C A T I O N
1311 //
1312 // Allocation attempts to be fast in the case of frequent small objects.
1313 // It breaks down like this:
1314 //
1315 // 1) Size in doublewords is computed. This is a constant for objects and
1316 // variable for most arrays. Doubleword units are used to avoid size
1317 // overflow of huge doubleword arrays. We need doublewords in the end for
1318 // rounding.
1319 //
1320 // 2) Size is checked for being 'too large'. Too-large allocations will go
1321 // the slow path into the VM. The slow path can throw any required
1322 // exceptions, and does all the special checks for very large arrays. The
1323 // size test can constant-fold away for objects. For objects with
1324 // finalizers it constant-folds the otherway: you always go slow with
1325 // finalizers.
1326 //
1327 // 3) If NOT using TLABs, this is the contended loop-back point.
1328 // Load-Locked the heap top. If using TLABs normal-load the heap top.
1329 //
1330 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
1331 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
1332 // "size*8" we always enter the VM, where "largish" is a constant picked small
1333 // enough that there's always space between the eden max and 4Gig (old space is
1334 // there so it's quite large) and large enough that the cost of entering the VM
1335 // is dwarfed by the cost to initialize the space.
1336 //
1337 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1338 // down. If contended, repeat at step 3. If using TLABs normal-store
1339 // adjusted heap top back down; there is no contention.
1340 //
1341 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
1342 // fields.
1343 //
1344 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1345 // oop flavor.
1346 //
1347 //=============================================================================
1348 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1349 // Allocations bigger than this always go the slow route.
1350 // This value must be small enough that allocation attempts that need to
1351 // trigger exceptions go the slow route. Also, it must be small enough so
1352 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1353 //=============================================================================j//
1354 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1355 // The allocator will coalesce int->oop copies away. See comment in
1356 // coalesce.cpp about how this works. It depends critically on the exact
1357 // code shape produced here, so if you are changing this code shape
1358 // make sure the GC info for the heap-top is correct in and around the
1359 // slow-path call.
1360 //
1361
1362 void PhaseMacroExpand::expand_allocate_common(
1363 AllocateNode* alloc, // allocation node to be expanded
1364 Node* length, // array length for an array allocation
1365 const TypeFunc* slow_call_type, // Type of slow call
1366 address slow_call_address // Address of slow call
1367 )
1368 {
1369
1370 Node* ctrl = alloc->in(TypeFunc::Control);
1371 Node* mem = alloc->in(TypeFunc::Memory);
1372 Node* i_o = alloc->in(TypeFunc::I_O);
1373 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1374 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1375 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1376
1377 assert(ctrl != NULL, "must have control");
1378 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1379 // they will not be used if "always_slow" is set
1380 enum { slow_result_path = 1, fast_result_path = 2 };
1381 Node *result_region = NULL;
1382 Node *result_phi_rawmem = NULL;
1383 Node *result_phi_rawoop = NULL;
1384 Node *result_phi_i_o = NULL;
1385
1386 // The initial slow comparison is a size check, the comparison
1387 // we want to do is a BoolTest::gt
1388 bool always_slow = false;
1389 int tv = _igvn.find_int_con(initial_slow_test, -1);
1390 if (tv >= 0) {
1391 always_slow = (tv == 1);
1392 initial_slow_test = NULL;
1393 } else {
1394 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1395 }
1396
1397 if (C->env()->dtrace_alloc_probes() ||
1398 (!UseTLAB && !Universe::heap()->supports_inline_contig_alloc())) {
1399 // Force slow-path allocation
1400 always_slow = true;
1401 initial_slow_test = NULL;
1402 }
1403
1404 Node *slow_region = NULL;
1405 Node *toobig_false = ctrl;
1406
1407 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent");
1408 // generate the initial test if necessary
1409 if (initial_slow_test != NULL ) {
1410 if (slow_region == NULL) {
1411 slow_region = new RegionNode(1);
1412 }
1413 // Now make the initial failure test. Usually a too-big test but
1414 // might be a TRUE for finalizers or a fancy class check for
1415 // newInstance0.
1416 IfNode* toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1417 transform_later(toobig_iff);
1418 // Plug the failing-too-big test into the slow-path region
1419 Node* toobig_true = new IfTrueNode(toobig_iff);
1420 transform_later(toobig_true);
1421 slow_region ->add_req(toobig_true);
1422 toobig_false = new IfFalseNode(toobig_iff);
1423 transform_later(toobig_false);
1424 } else { // No initial test, just fall into next case
1425 toobig_false = ctrl;
1426 }
1427
1428 Node *slow_mem = mem; // save the current memory state for slow path
1429 // generate the fast allocation code unless we know that the initial test will always go slow
1430 if (!always_slow) {
1431 // Fast path modifies only raw memory.
1432 if (mem->is_MergeMem()) {
1433 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1434 }
1435
1436 // allocate the Region and Phi nodes for the result
1437 result_region = new RegionNode(3);
1438 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1439 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1440 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1441
1442 // Grab regular I/O before optional prefetch may change it.
1443 // Slow-path does no I/O so just set it to the original I/O.
1444 result_phi_i_o->init_req(slow_result_path, i_o);
1445
1446 Node* needgc_ctrl = NULL;
1447 // Name successful fast-path variables
1448 Node* fast_oop_ctrl;
1449 Node* fast_oop_rawmem;
1450
1451 intx prefetch_lines = length != NULL ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1452
1453 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1454 Node* fast_oop = bs->obj_allocate(this, ctrl, mem, toobig_false, size_in_bytes, i_o, needgc_ctrl,
1455 fast_oop_ctrl, fast_oop_rawmem,
1456 prefetch_lines);
1457
1458 if (slow_region != NULL) {
1459 slow_region->add_req(needgc_ctrl);
1460 // This completes all paths into the slow merge point
1461 transform_later(slow_region);
1462 } else {
1463 // Just fall from the need-GC path straight into the VM call.
1464 slow_region = needgc_ctrl;
1465 }
1466
1467 InitializeNode* init = alloc->initialization();
1468 fast_oop_rawmem = initialize_object(alloc,
1469 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1470 klass_node, length, size_in_bytes);
1471
1472 // If initialization is performed by an array copy, any required
1473 // MemBarStoreStore was already added. If the object does not
1474 // escape no need for a MemBarStoreStore. If the object does not
1475 // escape in its initializer and memory barrier (MemBarStoreStore or
1476 // stronger) is already added at exit of initializer, also no need
1477 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore
1478 // so that stores that initialize this object can't be reordered
1479 // with a subsequent store that makes this object accessible by
1480 // other threads.
1481 // Other threads include java threads and JVM internal threads
1482 // (for example concurrent GC threads). Current concurrent GC
1483 // implementation: CMS and G1 will not scan newly created object,
1484 // so it's safe to skip storestore barrier when allocation does
1485 // not escape.
1486 if (!alloc->does_not_escape_thread() &&
1487 !alloc->is_allocation_MemBar_redundant() &&
1488 (init == NULL || !init->is_complete_with_arraycopy())) {
1489 if (init == NULL || init->req() < InitializeNode::RawStores) {
1490 // No InitializeNode or no stores captured by zeroing
1491 // elimination. Simply add the MemBarStoreStore after object
1492 // initialization.
1493 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1494 transform_later(mb);
1495
1496 mb->init_req(TypeFunc::Memory, fast_oop_rawmem);
1497 mb->init_req(TypeFunc::Control, fast_oop_ctrl);
1498 fast_oop_ctrl = new ProjNode(mb,TypeFunc::Control);
1499 transform_later(fast_oop_ctrl);
1500 fast_oop_rawmem = new ProjNode(mb,TypeFunc::Memory);
1501 transform_later(fast_oop_rawmem);
1502 } else {
1503 // Add the MemBarStoreStore after the InitializeNode so that
1504 // all stores performing the initialization that were moved
1505 // before the InitializeNode happen before the storestore
1506 // barrier.
1507
1508 Node* init_ctrl = init->proj_out_or_null(TypeFunc::Control);
1509 Node* init_mem = init->proj_out_or_null(TypeFunc::Memory);
1510
1511 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1512 transform_later(mb);
1513
1514 Node* ctrl = new ProjNode(init,TypeFunc::Control);
1515 transform_later(ctrl);
1516 Node* mem = new ProjNode(init,TypeFunc::Memory);
1517 transform_later(mem);
1518
1519 // The MemBarStoreStore depends on control and memory coming
1520 // from the InitializeNode
1521 mb->init_req(TypeFunc::Memory, mem);
1522 mb->init_req(TypeFunc::Control, ctrl);
1523
1524 ctrl = new ProjNode(mb,TypeFunc::Control);
1525 transform_later(ctrl);
1526 mem = new ProjNode(mb,TypeFunc::Memory);
1527 transform_later(mem);
1528
1529 // All nodes that depended on the InitializeNode for control
1530 // and memory must now depend on the MemBarNode that itself
1531 // depends on the InitializeNode
1532 if (init_ctrl != NULL) {
1533 _igvn.replace_node(init_ctrl, ctrl);
1534 }
1535 if (init_mem != NULL) {
1536 _igvn.replace_node(init_mem, mem);
1537 }
1538 }
1539 }
1540
1541 if (C->env()->dtrace_extended_probes()) {
1542 // Slow-path call
1543 int size = TypeFunc::Parms + 2;
1544 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1545 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base),
1546 "dtrace_object_alloc",
1547 TypeRawPtr::BOTTOM);
1548
1549 // Get base of thread-local storage area
1550 Node* thread = new ThreadLocalNode();
1551 transform_later(thread);
1552
1553 call->init_req(TypeFunc::Parms+0, thread);
1554 call->init_req(TypeFunc::Parms+1, fast_oop);
1555 call->init_req(TypeFunc::Control, fast_oop_ctrl);
1556 call->init_req(TypeFunc::I_O , top()); // does no i/o
1557 call->init_req(TypeFunc::Memory , fast_oop_rawmem);
1558 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1559 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1560 transform_later(call);
1561 fast_oop_ctrl = new ProjNode(call,TypeFunc::Control);
1562 transform_later(fast_oop_ctrl);
1563 fast_oop_rawmem = new ProjNode(call,TypeFunc::Memory);
1564 transform_later(fast_oop_rawmem);
1565 }
1566
1567 // Plug in the successful fast-path into the result merge point
1568 result_region ->init_req(fast_result_path, fast_oop_ctrl);
1569 result_phi_rawoop->init_req(fast_result_path, fast_oop);
1570 result_phi_i_o ->init_req(fast_result_path, i_o);
1571 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1572 } else {
1573 slow_region = ctrl;
1574 result_phi_i_o = i_o; // Rename it to use in the following code.
1575 }
1576
1577 // Generate slow-path call
1578 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1579 OptoRuntime::stub_name(slow_call_address),
1580 alloc->jvms()->bci(),
1581 TypePtr::BOTTOM);
1582 call->init_req( TypeFunc::Control, slow_region );
1583 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o
1584 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs
1585 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) );
1586 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) );
1587
1588 call->init_req(TypeFunc::Parms+0, klass_node);
1589 if (length != NULL) {
1590 call->init_req(TypeFunc::Parms+1, length);
1591 }
1592
1593 // Copy debug information and adjust JVMState information, then replace
1594 // allocate node with the call
1595 copy_call_debug_info((CallNode *) alloc, call);
1596 if (!always_slow) {
1597 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1598 } else {
1599 // Hook i_o projection to avoid its elimination during allocation
1600 // replacement (when only a slow call is generated).
1601 call->set_req(TypeFunc::I_O, result_phi_i_o);
1602 }
1603 _igvn.replace_node(alloc, call);
1604 transform_later(call);
1605
1606 // Identify the output projections from the allocate node and
1607 // adjust any references to them.
1608 // The control and io projections look like:
1609 //
1610 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1611 // Allocate Catch
1612 // ^---Proj(io) <-------+ ^---CatchProj(io)
1613 //
1614 // We are interested in the CatchProj nodes.
1615 //
1616 extract_call_projections(call);
1617
1618 // An allocate node has separate memory projections for the uses on
1619 // the control and i_o paths. Replace the control memory projection with
1620 // result_phi_rawmem (unless we are only generating a slow call when
1621 // both memory projections are combined)
1622 if (!always_slow && _memproj_fallthrough != NULL) {
1623 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) {
1624 Node *use = _memproj_fallthrough->fast_out(i);
1625 _igvn.rehash_node_delayed(use);
1626 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem);
1627 // back up iterator
1628 --i;
1629 }
1630 }
1631 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete
1632 // _memproj_catchall so we end up with a call that has only 1 memory projection.
1633 if (_memproj_catchall != NULL ) {
1634 if (_memproj_fallthrough == NULL) {
1635 _memproj_fallthrough = new ProjNode(call, TypeFunc::Memory);
1636 transform_later(_memproj_fallthrough);
1637 }
1638 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) {
1639 Node *use = _memproj_catchall->fast_out(i);
1640 _igvn.rehash_node_delayed(use);
1641 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough);
1642 // back up iterator
1643 --i;
1644 }
1645 assert(_memproj_catchall->outcnt() == 0, "all uses must be deleted");
1646 _igvn.remove_dead_node(_memproj_catchall);
1647 }
1648
1649 // An allocate node has separate i_o projections for the uses on the control
1650 // and i_o paths. Always replace the control i_o projection with result i_o
1651 // otherwise incoming i_o become dead when only a slow call is generated
1652 // (it is different from memory projections where both projections are
1653 // combined in such case).
1654 if (_ioproj_fallthrough != NULL) {
1655 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) {
1656 Node *use = _ioproj_fallthrough->fast_out(i);
1657 _igvn.rehash_node_delayed(use);
1658 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o);
1659 // back up iterator
1660 --i;
1661 }
1662 }
1663 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete
1664 // _ioproj_catchall so we end up with a call that has only 1 i_o projection.
1665 if (_ioproj_catchall != NULL ) {
1666 if (_ioproj_fallthrough == NULL) {
1667 _ioproj_fallthrough = new ProjNode(call, TypeFunc::I_O);
1668 transform_later(_ioproj_fallthrough);
1669 }
1670 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) {
1671 Node *use = _ioproj_catchall->fast_out(i);
1672 _igvn.rehash_node_delayed(use);
1673 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough);
1674 // back up iterator
1675 --i;
1676 }
1677 assert(_ioproj_catchall->outcnt() == 0, "all uses must be deleted");
1678 _igvn.remove_dead_node(_ioproj_catchall);
1679 }
1680
1681 // if we generated only a slow call, we are done
1682 if (always_slow) {
1683 // Now we can unhook i_o.
1684 if (result_phi_i_o->outcnt() > 1) {
1685 call->set_req(TypeFunc::I_O, top());
1686 } else {
1687 assert(result_phi_i_o->unique_ctrl_out() == call, "");
1688 // Case of new array with negative size known during compilation.
1689 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1690 // following code since call to runtime will throw exception.
1691 // As result there will be no users of i_o after the call.
1692 // Leave i_o attached to this call to avoid problems in preceding graph.
1693 }
1694 return;
1695 }
1696
1697 if (_fallthroughcatchproj != NULL) {
1698 ctrl = _fallthroughcatchproj->clone();
1699 transform_later(ctrl);
1700 _igvn.replace_node(_fallthroughcatchproj, result_region);
1701 } else {
1702 ctrl = top();
1703 }
1704 Node *slow_result;
1705 if (_resproj == NULL) {
1706 // no uses of the allocation result
1707 slow_result = top();
1708 } else {
1709 slow_result = _resproj->clone();
1710 transform_later(slow_result);
1711 _igvn.replace_node(_resproj, result_phi_rawoop);
1712 }
1713
1714 // Plug slow-path into result merge point
1715 result_region ->init_req( slow_result_path, ctrl );
1716 result_phi_rawoop->init_req( slow_result_path, slow_result);
1717 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough );
1718 transform_later(result_region);
1719 transform_later(result_phi_rawoop);
1720 transform_later(result_phi_rawmem);
1721 transform_later(result_phi_i_o);
1722 // This completes all paths into the result merge point
1723 }
1724
1725
1726 // Helper for PhaseMacroExpand::expand_allocate_common.
1727 // Initializes the newly-allocated storage.
1728 Node* PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1729 Node* control, Node* rawmem, Node* object,
1730 Node* klass_node, Node* length,
1731 Node* size_in_bytes) {
1732 InitializeNode* init = alloc->initialization();
1733 // Store the klass & mark bits
1734 Node* mark_node = alloc->make_ideal_mark(&_igvn, object, control, rawmem, klass_node);
1735 if (!mark_node->is_Con()) {
1736 transform_later(mark_node);
1737 }
1738 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, TypeX_X->basic_type());
1739
1740 BasicType bt = T_METADATA;
1741 Node* metadata = klass_node;
1742 Node* properties = alloc->in(AllocateNode::StorageProperties);
1743 if (properties != NULL) {
1744 // Encode array storage properties into klass pointer
1745 assert(EnableValhalla, "array storage properties not supported");
1746 if (UseCompressedClassPointers) {
1747 // Compress the klass pointer before inserting the storage properties value
1748 metadata = transform_later(new EncodePKlassNode(metadata, metadata->bottom_type()->make_narrowklass()));
1749 metadata = transform_later(new CastN2INode(metadata));
1750 metadata = transform_later(new OrINode(metadata, transform_later(new ConvL2INode(properties))));
1751 bt = T_INT;
1752 } else {
1753 metadata = transform_later(new CastP2XNode(NULL, metadata));
1754 metadata = transform_later(new OrXNode(metadata, properties));
1755 bt = T_LONG;
1756 }
1757 }
1758 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), metadata, bt);
1759
1760 int header_size = alloc->minimum_header_size(); // conservatively small
1761
1762 // Array length
1763 if (length != NULL) { // Arrays need length field
1764 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1765 // conservatively small header size:
1766 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1767 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1768 if (k->is_array_klass()) // we know the exact header size in most cases:
1769 header_size = Klass::layout_helper_header_size(k->layout_helper());
1770 }
1771
1772 // Clear the object body, if necessary.
1773 if (init == NULL) {
1774 // The init has somehow disappeared; be cautious and clear everything.
1775 //
1776 // This can happen if a node is allocated but an uncommon trap occurs
1777 // immediately. In this case, the Initialize gets associated with the
1778 // trap, and may be placed in a different (outer) loop, if the Allocate
1779 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1780 // there can be two Allocates to one Initialize. The answer in all these
1781 // edge cases is safety first. It is always safe to clear immediately
1782 // within an Allocate, and then (maybe or maybe not) clear some more later.
1783 if (!(UseTLAB && ZeroTLAB)) {
1784 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1785 alloc->in(AllocateNode::DefaultValue),
1786 alloc->in(AllocateNode::RawDefaultValue),
1787 header_size, size_in_bytes,
1788 &_igvn);
1789 }
1790 } else {
1791 if (!init->is_complete()) {
1792 // Try to win by zeroing only what the init does not store.
1793 // We can also try to do some peephole optimizations,
1794 // such as combining some adjacent subword stores.
1795 rawmem = init->complete_stores(control, rawmem, object,
1796 header_size, size_in_bytes, &_igvn);
1797 }
1798 // We have no more use for this link, since the AllocateNode goes away:
1799 init->set_req(InitializeNode::RawAddress, top());
1800 // (If we keep the link, it just confuses the register allocator,
1801 // who thinks he sees a real use of the address by the membar.)
1802 }
1803
1804 return rawmem;
1805 }
1806
1807 // Generate prefetch instructions for next allocations.
1808 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
1809 Node*& contended_phi_rawmem,
1810 Node* old_eden_top, Node* new_eden_top,
1811 intx lines) {
1812 enum { fall_in_path = 1, pf_path = 2 };
1813 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
1814 // Generate prefetch allocation with watermark check.
1815 // As an allocation hits the watermark, we will prefetch starting
1816 // at a "distance" away from watermark.
1817
1818 Node *pf_region = new RegionNode(3);
1819 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY,
1820 TypeRawPtr::BOTTOM );
1821 // I/O is used for Prefetch
1822 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO );
1823
1824 Node *thread = new ThreadLocalNode();
1825 transform_later(thread);
1826
1827 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread,
1828 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
1829 transform_later(eden_pf_adr);
1830
1831 Node *old_pf_wm = new LoadPNode(needgc_false,
1832 contended_phi_rawmem, eden_pf_adr,
1833 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM,
1834 MemNode::unordered);
1835 transform_later(old_pf_wm);
1836
1837 // check against new_eden_top
1838 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm );
1839 transform_later(need_pf_cmp);
1840 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge );
1841 transform_later(need_pf_bol);
1842 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol,
1843 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1844 transform_later(need_pf_iff);
1845
1846 // true node, add prefetchdistance
1847 Node *need_pf_true = new IfTrueNode( need_pf_iff );
1848 transform_later(need_pf_true);
1849
1850 Node *need_pf_false = new IfFalseNode( need_pf_iff );
1851 transform_later(need_pf_false);
1852
1853 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm,
1854 _igvn.MakeConX(AllocatePrefetchDistance) );
1855 transform_later(new_pf_wmt );
1856 new_pf_wmt->set_req(0, need_pf_true);
1857
1858 Node *store_new_wmt = new StorePNode(need_pf_true,
1859 contended_phi_rawmem, eden_pf_adr,
1860 TypeRawPtr::BOTTOM, new_pf_wmt,
1861 MemNode::unordered);
1862 transform_later(store_new_wmt);
1863
1864 // adding prefetches
1865 pf_phi_abio->init_req( fall_in_path, i_o );
1866
1867 Node *prefetch_adr;
1868 Node *prefetch;
1869 uint step_size = AllocatePrefetchStepSize;
1870 uint distance = 0;
1871
1872 for ( intx i = 0; i < lines; i++ ) {
1873 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt,
1874 _igvn.MakeConX(distance) );
1875 transform_later(prefetch_adr);
1876 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1877 transform_later(prefetch);
1878 distance += step_size;
1879 i_o = prefetch;
1880 }
1881 pf_phi_abio->set_req( pf_path, i_o );
1882
1883 pf_region->init_req( fall_in_path, need_pf_false );
1884 pf_region->init_req( pf_path, need_pf_true );
1885
1886 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
1887 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
1888
1889 transform_later(pf_region);
1890 transform_later(pf_phi_rawmem);
1891 transform_later(pf_phi_abio);
1892
1893 needgc_false = pf_region;
1894 contended_phi_rawmem = pf_phi_rawmem;
1895 i_o = pf_phi_abio;
1896 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) {
1897 // Insert a prefetch instruction for each allocation.
1898 // This code is used to generate 1 prefetch instruction per cache line.
1899
1900 // Generate several prefetch instructions.
1901 uint step_size = AllocatePrefetchStepSize;
1902 uint distance = AllocatePrefetchDistance;
1903
1904 // Next cache address.
1905 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top,
1906 _igvn.MakeConX(step_size + distance));
1907 transform_later(cache_adr);
1908 cache_adr = new CastP2XNode(needgc_false, cache_adr);
1909 transform_later(cache_adr);
1910 // Address is aligned to execute prefetch to the beginning of cache line size
1911 // (it is important when BIS instruction is used on SPARC as prefetch).
1912 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1));
1913 cache_adr = new AndXNode(cache_adr, mask);
1914 transform_later(cache_adr);
1915 cache_adr = new CastX2PNode(cache_adr);
1916 transform_later(cache_adr);
1917
1918 // Prefetch
1919 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr );
1920 prefetch->set_req(0, needgc_false);
1921 transform_later(prefetch);
1922 contended_phi_rawmem = prefetch;
1923 Node *prefetch_adr;
1924 distance = step_size;
1925 for ( intx i = 1; i < lines; i++ ) {
1926 prefetch_adr = new AddPNode( cache_adr, cache_adr,
1927 _igvn.MakeConX(distance) );
1928 transform_later(prefetch_adr);
1929 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr );
1930 transform_later(prefetch);
1931 distance += step_size;
1932 contended_phi_rawmem = prefetch;
1933 }
1934 } else if( AllocatePrefetchStyle > 0 ) {
1935 // Insert a prefetch for each allocation only on the fast-path
1936 Node *prefetch_adr;
1937 Node *prefetch;
1938 // Generate several prefetch instructions.
1939 uint step_size = AllocatePrefetchStepSize;
1940 uint distance = AllocatePrefetchDistance;
1941 for ( intx i = 0; i < lines; i++ ) {
1942 prefetch_adr = new AddPNode( old_eden_top, new_eden_top,
1943 _igvn.MakeConX(distance) );
1944 transform_later(prefetch_adr);
1945 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1946 // Do not let it float too high, since if eden_top == eden_end,
1947 // both might be null.
1948 if( i == 0 ) { // Set control for first prefetch, next follows it
1949 prefetch->init_req(0, needgc_false);
1950 }
1951 transform_later(prefetch);
1952 distance += step_size;
1953 i_o = prefetch;
1954 }
1955 }
1956 return i_o;
1957 }
1958
1959
1960 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
1961 expand_allocate_common(alloc, NULL,
1962 OptoRuntime::new_instance_Type(),
1963 OptoRuntime::new_instance_Java());
1964 }
1965
1966 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
1967 Node* length = alloc->in(AllocateNode::ALength);
1968 InitializeNode* init = alloc->initialization();
1969 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1970 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1971 address slow_call_address; // Address of slow call
1972 if (init != NULL && init->is_complete_with_arraycopy() &&
1973 k->is_type_array_klass()) {
1974 // Don't zero type array during slow allocation in VM since
1975 // it will be initialized later by arraycopy in compiled code.
1976 slow_call_address = OptoRuntime::new_array_nozero_Java();
1977 } else {
1978 slow_call_address = OptoRuntime::new_array_Java();
1979 }
1980 expand_allocate_common(alloc, length,
1981 OptoRuntime::new_array_Type(),
1982 slow_call_address);
1983 }
1984
1985 //-------------------mark_eliminated_box----------------------------------
1986 //
1987 // During EA obj may point to several objects but after few ideal graph
1988 // transformations (CCP) it may point to only one non escaping object
1989 // (but still using phi), corresponding locks and unlocks will be marked
1990 // for elimination. Later obj could be replaced with a new node (new phi)
1991 // and which does not have escape information. And later after some graph
1992 // reshape other locks and unlocks (which were not marked for elimination
1993 // before) are connected to this new obj (phi) but they still will not be
1994 // marked for elimination since new obj has no escape information.
1995 // Mark all associated (same box and obj) lock and unlock nodes for
1996 // elimination if some of them marked already.
1997 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) {
1998 if (oldbox->as_BoxLock()->is_eliminated())
1999 return; // This BoxLock node was processed already.
2000
2001 // New implementation (EliminateNestedLocks) has separate BoxLock
2002 // node for each locked region so mark all associated locks/unlocks as
2003 // eliminated even if different objects are referenced in one locked region
2004 // (for example, OSR compilation of nested loop inside locked scope).
2005 if (EliminateNestedLocks ||
2006 oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj)) {
2007 // Box is used only in one lock region. Mark this box as eliminated.
2008 _igvn.hash_delete(oldbox);
2009 oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value
2010 _igvn.hash_insert(oldbox);
2011
2012 for (uint i = 0; i < oldbox->outcnt(); i++) {
2013 Node* u = oldbox->raw_out(i);
2014 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) {
2015 AbstractLockNode* alock = u->as_AbstractLock();
2016 // Check lock's box since box could be referenced by Lock's debug info.
2017 if (alock->box_node() == oldbox) {
2018 // Mark eliminated all related locks and unlocks.
2019 #ifdef ASSERT
2020 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4");
2021 #endif
2022 alock->set_non_esc_obj();
2023 }
2024 }
2025 }
2026 return;
2027 }
2028
2029 // Create new "eliminated" BoxLock node and use it in monitor debug info
2030 // instead of oldbox for the same object.
2031 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
2032
2033 // Note: BoxLock node is marked eliminated only here and it is used
2034 // to indicate that all associated lock and unlock nodes are marked
2035 // for elimination.
2036 newbox->set_eliminated();
2037 transform_later(newbox);
2038
2039 // Replace old box node with new box for all users of the same object.
2040 for (uint i = 0; i < oldbox->outcnt();) {
2041 bool next_edge = true;
2042
2043 Node* u = oldbox->raw_out(i);
2044 if (u->is_AbstractLock()) {
2045 AbstractLockNode* alock = u->as_AbstractLock();
2046 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) {
2047 // Replace Box and mark eliminated all related locks and unlocks.
2048 #ifdef ASSERT
2049 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5");
2050 #endif
2051 alock->set_non_esc_obj();
2052 _igvn.rehash_node_delayed(alock);
2053 alock->set_box_node(newbox);
2054 next_edge = false;
2055 }
2056 }
2057 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) {
2058 FastLockNode* flock = u->as_FastLock();
2059 assert(flock->box_node() == oldbox, "sanity");
2060 _igvn.rehash_node_delayed(flock);
2061 flock->set_box_node(newbox);
2062 next_edge = false;
2063 }
2064
2065 // Replace old box in monitor debug info.
2066 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
2067 SafePointNode* sfn = u->as_SafePoint();
2068 JVMState* youngest_jvms = sfn->jvms();
2069 int max_depth = youngest_jvms->depth();
2070 for (int depth = 1; depth <= max_depth; depth++) {
2071 JVMState* jvms = youngest_jvms->of_depth(depth);
2072 int num_mon = jvms->nof_monitors();
2073 // Loop over monitors
2074 for (int idx = 0; idx < num_mon; idx++) {
2075 Node* obj_node = sfn->monitor_obj(jvms, idx);
2076 Node* box_node = sfn->monitor_box(jvms, idx);
2077 if (box_node == oldbox && obj_node->eqv_uncast(obj)) {
2078 int j = jvms->monitor_box_offset(idx);
2079 _igvn.replace_input_of(u, j, newbox);
2080 next_edge = false;
2081 }
2082 }
2083 }
2084 }
2085 if (next_edge) i++;
2086 }
2087 }
2088
2089 //-----------------------mark_eliminated_locking_nodes-----------------------
2090 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) {
2091 if (EliminateNestedLocks) {
2092 if (alock->is_nested()) {
2093 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity");
2094 return;
2095 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened
2096 // Only Lock node has JVMState needed here.
2097 // Not that preceding claim is documented anywhere else.
2098 if (alock->jvms() != NULL) {
2099 if (alock->as_Lock()->is_nested_lock_region()) {
2100 // Mark eliminated related nested locks and unlocks.
2101 Node* obj = alock->obj_node();
2102 BoxLockNode* box_node = alock->box_node()->as_BoxLock();
2103 assert(!box_node->is_eliminated(), "should not be marked yet");
2104 // Note: BoxLock node is marked eliminated only here
2105 // and it is used to indicate that all associated lock
2106 // and unlock nodes are marked for elimination.
2107 box_node->set_eliminated(); // Box's hash is always NO_HASH here
2108 for (uint i = 0; i < box_node->outcnt(); i++) {
2109 Node* u = box_node->raw_out(i);
2110 if (u->is_AbstractLock()) {
2111 alock = u->as_AbstractLock();
2112 if (alock->box_node() == box_node) {
2113 // Verify that this Box is referenced only by related locks.
2114 assert(alock->obj_node()->eqv_uncast(obj), "");
2115 // Mark all related locks and unlocks.
2116 #ifdef ASSERT
2117 alock->log_lock_optimization(C, "eliminate_lock_set_nested");
2118 #endif
2119 alock->set_nested();
2120 }
2121 }
2122 }
2123 } else {
2124 #ifdef ASSERT
2125 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region");
2126 if (C->log() != NULL)
2127 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output
2128 #endif
2129 }
2130 }
2131 return;
2132 }
2133 // Process locks for non escaping object
2134 assert(alock->is_non_esc_obj(), "");
2135 } // EliminateNestedLocks
2136
2137 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2138 // Look for all locks of this object and mark them and
2139 // corresponding BoxLock nodes as eliminated.
2140 Node* obj = alock->obj_node();
2141 for (uint j = 0; j < obj->outcnt(); j++) {
2142 Node* o = obj->raw_out(j);
2143 if (o->is_AbstractLock() &&
2144 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2145 alock = o->as_AbstractLock();
2146 Node* box = alock->box_node();
2147 // Replace old box node with new eliminated box for all users
2148 // of the same object and mark related locks as eliminated.
2149 mark_eliminated_box(box, obj);
2150 }
2151 }
2152 }
2153 }
2154
2155 // we have determined that this lock/unlock can be eliminated, we simply
2156 // eliminate the node without expanding it.
2157 //
2158 // Note: The membar's associated with the lock/unlock are currently not
2159 // eliminated. This should be investigated as a future enhancement.
2160 //
2161 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2162
2163 if (!alock->is_eliminated()) {
2164 return false;
2165 }
2166 #ifdef ASSERT
2167 const Type* obj_type = _igvn.type(alock->obj_node());
2168 assert(!obj_type->isa_valuetype() && !obj_type->is_valuetypeptr(), "Eliminating lock on value type");
2169 if (!alock->is_coarsened()) {
2170 // Check that new "eliminated" BoxLock node is created.
2171 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2172 assert(oldbox->is_eliminated(), "should be done already");
2173 }
2174 #endif
2175
2176 alock->log_lock_optimization(C, "eliminate_lock");
2177
2178 #ifndef PRODUCT
2179 if (PrintEliminateLocks) {
2180 if (alock->is_Lock()) {
2181 tty->print_cr("++++ Eliminated: %d Lock", alock->_idx);
2182 } else {
2183 tty->print_cr("++++ Eliminated: %d Unlock", alock->_idx);
2184 }
2185 }
2186 #endif
2187
2188 Node* mem = alock->in(TypeFunc::Memory);
2189 Node* ctrl = alock->in(TypeFunc::Control);
2190 guarantee(ctrl != NULL, "missing control projection, cannot replace_node() with NULL");
2191
2192 extract_call_projections(alock);
2193 // There are 2 projections from the lock. The lock node will
2194 // be deleted when its last use is subsumed below.
2195 assert(alock->outcnt() == 2 &&
2196 _fallthroughproj != NULL &&
2197 _memproj_fallthrough != NULL,
2198 "Unexpected projections from Lock/Unlock");
2199
2200 Node* fallthroughproj = _fallthroughproj;
2201 Node* memproj_fallthrough = _memproj_fallthrough;
2202
2203 // The memory projection from a lock/unlock is RawMem
2204 // The input to a Lock is merged memory, so extract its RawMem input
2205 // (unless the MergeMem has been optimized away.)
2206 if (alock->is_Lock()) {
2207 // Seach for MemBarAcquireLock node and delete it also.
2208 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2209 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, "");
2210 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2211 Node* memproj = membar->proj_out(TypeFunc::Memory);
2212 _igvn.replace_node(ctrlproj, fallthroughproj);
2213 _igvn.replace_node(memproj, memproj_fallthrough);
2214
2215 // Delete FastLock node also if this Lock node is unique user
2216 // (a loop peeling may clone a Lock node).
2217 Node* flock = alock->as_Lock()->fastlock_node();
2218 if (flock->outcnt() == 1) {
2219 assert(flock->unique_out() == alock, "sanity");
2220 _igvn.replace_node(flock, top());
2221 }
2222 }
2223
2224 // Seach for MemBarReleaseLock node and delete it also.
2225 if (alock->is_Unlock() && ctrl->is_Proj() && ctrl->in(0)->is_MemBar()) {
2226 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2227 assert(membar->Opcode() == Op_MemBarReleaseLock &&
2228 mem->is_Proj() && membar == mem->in(0), "");
2229 _igvn.replace_node(fallthroughproj, ctrl);
2230 _igvn.replace_node(memproj_fallthrough, mem);
2231 fallthroughproj = ctrl;
2232 memproj_fallthrough = mem;
2233 ctrl = membar->in(TypeFunc::Control);
2234 mem = membar->in(TypeFunc::Memory);
2235 }
2236
2237 _igvn.replace_node(fallthroughproj, ctrl);
2238 _igvn.replace_node(memproj_fallthrough, mem);
2239 return true;
2240 }
2241
2242
2243 //------------------------------expand_lock_node----------------------
2244 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
2245
2246 Node* ctrl = lock->in(TypeFunc::Control);
2247 Node* mem = lock->in(TypeFunc::Memory);
2248 Node* obj = lock->obj_node();
2249 Node* box = lock->box_node();
2250 Node* flock = lock->fastlock_node();
2251
2252 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2253
2254 // Make the merge point
2255 Node *region;
2256 Node *mem_phi;
2257 Node *slow_path;
2258
2259 if (UseOptoBiasInlining) {
2260 /*
2261 * See the full description in MacroAssembler::biased_locking_enter().
2262 *
2263 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) {
2264 * // The object is biased.
2265 * proto_node = klass->prototype_header;
2266 * o_node = thread | proto_node;
2267 * x_node = o_node ^ mark_word;
2268 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ?
2269 * // Done.
2270 * } else {
2271 * if( (x_node & biased_lock_mask) != 0 ) {
2272 * // The klass's prototype header is no longer biased.
2273 * cas(&mark_word, mark_word, proto_node)
2274 * goto cas_lock;
2275 * } else {
2276 * // The klass's prototype header is still biased.
2277 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch?
2278 * old = mark_word;
2279 * new = o_node;
2280 * } else {
2281 * // Different thread or anonymous biased.
2282 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask);
2283 * new = thread | old;
2284 * }
2285 * // Try to rebias.
2286 * if( cas(&mark_word, old, new) == 0 ) {
2287 * // Done.
2288 * } else {
2289 * goto slow_path; // Failed.
2290 * }
2291 * }
2292 * }
2293 * } else {
2294 * // The object is not biased.
2295 * cas_lock:
2296 * if( FastLock(obj) == 0 ) {
2297 * // Done.
2298 * } else {
2299 * slow_path:
2300 * OptoRuntime::complete_monitor_locking_Java(obj);
2301 * }
2302 * }
2303 */
2304
2305 region = new RegionNode(5);
2306 // create a Phi for the memory state
2307 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2308
2309 Node* fast_lock_region = new RegionNode(3);
2310 Node* fast_lock_mem_phi = new PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM);
2311
2312 // First, check mark word for the biased lock pattern.
2313 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2314
2315 // Get fast path - mark word has the biased lock pattern.
2316 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node,
2317 markWord::biased_lock_mask_in_place,
2318 markWord::biased_lock_pattern, true);
2319 // fast_lock_region->in(1) is set to slow path.
2320 fast_lock_mem_phi->init_req(1, mem);
2321
2322 // Now check that the lock is biased to the current thread and has
2323 // the same epoch and bias as Klass::_prototype_header.
2324
2325 // Special-case a fresh allocation to avoid building nodes:
2326 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn);
2327 if (klass_node == NULL) {
2328 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
2329 klass_node = transform_later(LoadKlassNode::make(_igvn, NULL, mem, k_adr, _igvn.type(k_adr)->is_ptr()));
2330 #ifdef _LP64
2331 if (UseCompressedClassPointers && klass_node->is_DecodeNKlass()) {
2332 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity");
2333 klass_node->in(1)->init_req(0, ctrl);
2334 } else
2335 #endif
2336 klass_node->init_req(0, ctrl);
2337 }
2338 Node *proto_node = make_load(ctrl, mem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeX_X, TypeX_X->basic_type());
2339
2340 Node* thread = transform_later(new ThreadLocalNode());
2341 Node* cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2342 Node* o_node = transform_later(new OrXNode(cast_thread, proto_node));
2343 Node* x_node = transform_later(new XorXNode(o_node, mark_node));
2344
2345 // Get slow path - mark word does NOT match the value.
2346 STATIC_ASSERT(markWord::age_mask_in_place <= INT_MAX);
2347 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node,
2348 (~(int)markWord::age_mask_in_place), 0);
2349 // region->in(3) is set to fast path - the object is biased to the current thread.
2350 mem_phi->init_req(3, mem);
2351
2352
2353 // Mark word does NOT match the value (thread | Klass::_prototype_header).
2354
2355
2356 // First, check biased pattern.
2357 // Get fast path - _prototype_header has the same biased lock pattern.
2358 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node,
2359 markWord::biased_lock_mask_in_place, 0, true);
2360
2361 not_biased_ctrl = fast_lock_region->in(2); // Slow path
2362 // fast_lock_region->in(2) - the prototype header is no longer biased
2363 // and we have to revoke the bias on this object.
2364 // We are going to try to reset the mark of this object to the prototype
2365 // value and fall through to the CAS-based locking scheme.
2366 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
2367 Node* cas = new StoreXConditionalNode(not_biased_ctrl, mem, adr,
2368 proto_node, mark_node);
2369 transform_later(cas);
2370 Node* proj = transform_later(new SCMemProjNode(cas));
2371 fast_lock_mem_phi->init_req(2, proj);
2372
2373
2374 // Second, check epoch bits.
2375 Node* rebiased_region = new RegionNode(3);
2376 Node* old_phi = new PhiNode( rebiased_region, TypeX_X);
2377 Node* new_phi = new PhiNode( rebiased_region, TypeX_X);
2378
2379 // Get slow path - mark word does NOT match epoch bits.
2380 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node,
2381 markWord::epoch_mask_in_place, 0);
2382 // The epoch of the current bias is not valid, attempt to rebias the object
2383 // toward the current thread.
2384 rebiased_region->init_req(2, epoch_ctrl);
2385 old_phi->init_req(2, mark_node);
2386 new_phi->init_req(2, o_node);
2387
2388 // rebiased_region->in(1) is set to fast path.
2389 // The epoch of the current bias is still valid but we know
2390 // nothing about the owner; it might be set or it might be clear.
2391 Node* cmask = MakeConX(markWord::biased_lock_mask_in_place |
2392 markWord::age_mask_in_place |
2393 markWord::epoch_mask_in_place);
2394 Node* old = transform_later(new AndXNode(mark_node, cmask));
2395 cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2396 Node* new_mark = transform_later(new OrXNode(cast_thread, old));
2397 old_phi->init_req(1, old);
2398 new_phi->init_req(1, new_mark);
2399
2400 transform_later(rebiased_region);
2401 transform_later(old_phi);
2402 transform_later(new_phi);
2403
2404 // Try to acquire the bias of the object using an atomic operation.
2405 // If this fails we will go in to the runtime to revoke the object's bias.
2406 cas = new StoreXConditionalNode(rebiased_region, mem, adr, new_phi, old_phi);
2407 transform_later(cas);
2408 proj = transform_later(new SCMemProjNode(cas));
2409
2410 // Get slow path - Failed to CAS.
2411 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0);
2412 mem_phi->init_req(4, proj);
2413 // region->in(4) is set to fast path - the object is rebiased to the current thread.
2414
2415 // Failed to CAS.
2416 slow_path = new RegionNode(3);
2417 Node *slow_mem = new PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM);
2418
2419 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control
2420 slow_mem->init_req(1, proj);
2421
2422 // Call CAS-based locking scheme (FastLock node).
2423
2424 transform_later(fast_lock_region);
2425 transform_later(fast_lock_mem_phi);
2426
2427 // Get slow path - FastLock failed to lock the object.
2428 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0);
2429 mem_phi->init_req(2, fast_lock_mem_phi);
2430 // region->in(2) is set to fast path - the object is locked to the current thread.
2431
2432 slow_path->init_req(2, ctrl); // Capture slow-control
2433 slow_mem->init_req(2, fast_lock_mem_phi);
2434
2435 transform_later(slow_path);
2436 transform_later(slow_mem);
2437 // Reset lock's memory edge.
2438 lock->set_req(TypeFunc::Memory, slow_mem);
2439
2440 } else {
2441 region = new RegionNode(3);
2442 // create a Phi for the memory state
2443 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2444
2445 // Optimize test; set region slot 2
2446 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2447 mem_phi->init_req(2, mem);
2448 }
2449
2450 const TypeOopPtr* objptr = _igvn.type(obj)->make_oopptr();
2451 if (objptr->can_be_value_type()) {
2452 // Deoptimize and re-execute if a value
2453 assert(EnableValhalla, "should only be used if value types are enabled");
2454 Node* mark = make_load(slow_path, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2455 Node* value_mask = _igvn.MakeConX(markWord::always_locked_pattern);
2456 Node* is_value = _igvn.transform(new AndXNode(mark, value_mask));
2457 Node* cmp = _igvn.transform(new CmpXNode(is_value, value_mask));
2458 Node* bol = _igvn.transform(new BoolNode(cmp, BoolTest::eq));
2459 Node* unc_ctrl = generate_slow_guard(&slow_path, bol, NULL);
2460
2461 int trap_request = Deoptimization::make_trap_request(Deoptimization::Reason_class_check, Deoptimization::Action_none);
2462 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
2463 const TypePtr* no_memory_effects = NULL;
2464 JVMState* jvms = lock->jvms();
2465 CallNode* unc = new CallStaticJavaNode(OptoRuntime::uncommon_trap_Type(), call_addr, "uncommon_trap",
2466 jvms->bci(), no_memory_effects);
2467
2468 unc->init_req(TypeFunc::Control, unc_ctrl);
2469 unc->init_req(TypeFunc::I_O, lock->i_o());
2470 unc->init_req(TypeFunc::Memory, mem); // may gc ptrs
2471 unc->init_req(TypeFunc::FramePtr, lock->in(TypeFunc::FramePtr));
2472 unc->init_req(TypeFunc::ReturnAdr, lock->in(TypeFunc::ReturnAdr));
2473 unc->init_req(TypeFunc::Parms+0, _igvn.intcon(trap_request));
2474 unc->set_cnt(PROB_UNLIKELY_MAG(4));
2475 copy_call_debug_info(lock, unc);
2476
2477 assert(unc->peek_monitor_box() == box, "wrong monitor");
2478 assert(unc->peek_monitor_obj() == obj, "wrong monitor");
2479
2480 // pop monitor and push obj back on stack: we trap before the monitorenter
2481 unc->pop_monitor();
2482 unc->grow_stack(unc->jvms(), 1);
2483 unc->set_stack(unc->jvms(), unc->jvms()->stk_size()-1, obj);
2484
2485 _igvn.register_new_node_with_optimizer(unc);
2486
2487 Node* ctrl = _igvn.transform(new ProjNode(unc, TypeFunc::Control));
2488 Node* halt = _igvn.transform(new HaltNode(ctrl, lock->in(TypeFunc::FramePtr)));
2489 C->root()->add_req(halt);
2490 }
2491
2492 // Make slow path call
2493 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(),
2494 OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path,
2495 obj, box, NULL);
2496
2497 extract_call_projections(call);
2498
2499 // Slow path can only throw asynchronous exceptions, which are always
2500 // de-opted. So the compiler thinks the slow-call can never throw an
2501 // exception. If it DOES throw an exception we would need the debug
2502 // info removed first (since if it throws there is no monitor).
2503 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2504 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2505
2506 // Capture slow path
2507 // disconnect fall-through projection from call and create a new one
2508 // hook up users of fall-through projection to region
2509 Node *slow_ctrl = _fallthroughproj->clone();
2510 transform_later(slow_ctrl);
2511 _igvn.hash_delete(_fallthroughproj);
2512 _fallthroughproj->disconnect_inputs(NULL, C);
2513 region->init_req(1, slow_ctrl);
2514 // region inputs are now complete
2515 transform_later(region);
2516 _igvn.replace_node(_fallthroughproj, region);
2517
2518 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2519 mem_phi->init_req(1, memproj );
2520 transform_later(mem_phi);
2521 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2522 }
2523
2524 //------------------------------expand_unlock_node----------------------
2525 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2526
2527 Node* ctrl = unlock->in(TypeFunc::Control);
2528 Node* mem = unlock->in(TypeFunc::Memory);
2529 Node* obj = unlock->obj_node();
2530 Node* box = unlock->box_node();
2531
2532 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2533
2534 // No need for a null check on unlock
2535
2536 // Make the merge point
2537 Node *region;
2538 Node *mem_phi;
2539
2540 if (UseOptoBiasInlining) {
2541 // Check for biased locking unlock case, which is a no-op.
2542 // See the full description in MacroAssembler::biased_locking_exit().
2543 region = new RegionNode(4);
2544 // create a Phi for the memory state
2545 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2546 mem_phi->init_req(3, mem);
2547
2548 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2549 ctrl = opt_bits_test(ctrl, region, 3, mark_node,
2550 markWord::biased_lock_mask_in_place,
2551 markWord::biased_lock_pattern);
2552 } else {
2553 region = new RegionNode(3);
2554 // create a Phi for the memory state
2555 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2556 }
2557
2558 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2559 funlock = transform_later( funlock )->as_FastUnlock();
2560 // Optimize test; set region slot 2
2561 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2562 Node *thread = transform_later(new ThreadLocalNode());
2563
2564 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2565 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2566 "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2567
2568 extract_call_projections(call);
2569
2570 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2571 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2572
2573 // No exceptions for unlocking
2574 // Capture slow path
2575 // disconnect fall-through projection from call and create a new one
2576 // hook up users of fall-through projection to region
2577 Node *slow_ctrl = _fallthroughproj->clone();
2578 transform_later(slow_ctrl);
2579 _igvn.hash_delete(_fallthroughproj);
2580 _fallthroughproj->disconnect_inputs(NULL, C);
2581 region->init_req(1, slow_ctrl);
2582 // region inputs are now complete
2583 transform_later(region);
2584 _igvn.replace_node(_fallthroughproj, region);
2585
2586 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2587 mem_phi->init_req(1, memproj );
2588 mem_phi->init_req(2, mem);
2589 transform_later(mem_phi);
2590 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2591 }
2592
2593 // A value type might be returned from the call but we don't know its
2594 // type. Either we get a buffered value (and nothing needs to be done)
2595 // or one of the values being returned is the klass of the value type
2596 // and we need to allocate a value type instance of that type and
2597 // initialize it with other values being returned. In that case, we
2598 // first try a fast path allocation and initialize the value with the
2599 // value klass's pack handler or we fall back to a runtime call.
2600 void PhaseMacroExpand::expand_mh_intrinsic_return(CallStaticJavaNode* call) {
2601 assert(call->method()->is_method_handle_intrinsic(), "must be a method handle intrinsic call");
2602 Node* ret = call->proj_out_or_null(TypeFunc::Parms);
2603 if (ret == NULL) {
2604 return;
2605 }
2606 const TypeFunc* tf = call->_tf;
2607 const TypeTuple* domain = OptoRuntime::store_value_type_fields_Type()->domain_cc();
2608 const TypeFunc* new_tf = TypeFunc::make(tf->domain_sig(), tf->domain_cc(), tf->range_sig(), domain);
2609 call->_tf = new_tf;
2610 // Make sure the change of type is applied before projections are processed by igvn
2611 _igvn.set_type(call, call->Value(&_igvn));
2612 _igvn.set_type(ret, ret->Value(&_igvn));
2613
2614 // Before any new projection is added:
2615 CallProjections* projs = call->extract_projections(true, true);
2616
2617 Node* ctl = new Node(1);
2618 Node* mem = new Node(1);
2619 Node* io = new Node(1);
2620 Node* ex_ctl = new Node(1);
2621 Node* ex_mem = new Node(1);
2622 Node* ex_io = new Node(1);
2623 Node* res = new Node(1);
2624
2625 Node* cast = transform_later(new CastP2XNode(ctl, res));
2626 Node* mask = MakeConX(0x1);
2627 Node* masked = transform_later(new AndXNode(cast, mask));
2628 Node* cmp = transform_later(new CmpXNode(masked, mask));
2629 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
2630 IfNode* allocation_iff = new IfNode(ctl, bol, PROB_MAX, COUNT_UNKNOWN);
2631 transform_later(allocation_iff);
2632 Node* allocation_ctl = transform_later(new IfTrueNode(allocation_iff));
2633 Node* no_allocation_ctl = transform_later(new IfFalseNode(allocation_iff));
2634
2635 Node* no_allocation_res = transform_later(new CheckCastPPNode(no_allocation_ctl, res, TypeInstPtr::BOTTOM));
2636
2637 Node* mask2 = MakeConX(-2);
2638 Node* masked2 = transform_later(new AndXNode(cast, mask2));
2639 Node* rawklassptr = transform_later(new CastX2PNode(masked2));
2640 Node* klass_node = transform_later(new CheckCastPPNode(allocation_ctl, rawklassptr, TypeKlassPtr::OBJECT_OR_NULL));
2641
2642 Node* slowpath_bol = NULL;
2643 Node* top_adr = NULL;
2644 Node* old_top = NULL;
2645 Node* new_top = NULL;
2646 if (UseTLAB) {
2647 Node* end_adr = NULL;
2648 set_eden_pointers(top_adr, end_adr);
2649 Node* end = make_load(ctl, mem, end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS);
2650 old_top = new LoadPNode(ctl, mem, top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered);
2651 transform_later(old_top);
2652 Node* layout_val = make_load(NULL, mem, klass_node, in_bytes(Klass::layout_helper_offset()), TypeInt::INT, T_INT);
2653 Node* size_in_bytes = ConvI2X(layout_val);
2654 new_top = new AddPNode(top(), old_top, size_in_bytes);
2655 transform_later(new_top);
2656 Node* slowpath_cmp = new CmpPNode(new_top, end);
2657 transform_later(slowpath_cmp);
2658 slowpath_bol = new BoolNode(slowpath_cmp, BoolTest::ge);
2659 transform_later(slowpath_bol);
2660 } else {
2661 slowpath_bol = intcon(1);
2662 top_adr = top();
2663 old_top = top();
2664 new_top = top();
2665 }
2666 IfNode* slowpath_iff = new IfNode(allocation_ctl, slowpath_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN);
2667 transform_later(slowpath_iff);
2668
2669 Node* slowpath_true = new IfTrueNode(slowpath_iff);
2670 transform_later(slowpath_true);
2671
2672 CallStaticJavaNode* slow_call = new CallStaticJavaNode(OptoRuntime::store_value_type_fields_Type(),
2673 StubRoutines::store_value_type_fields_to_buf(),
2674 "store_value_type_fields",
2675 call->jvms()->bci(),
2676 TypePtr::BOTTOM);
2677 slow_call->init_req(TypeFunc::Control, slowpath_true);
2678 slow_call->init_req(TypeFunc::Memory, mem);
2679 slow_call->init_req(TypeFunc::I_O, io);
2680 slow_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2681 slow_call->init_req(TypeFunc::ReturnAdr, call->in(TypeFunc::ReturnAdr));
2682 slow_call->init_req(TypeFunc::Parms, res);
2683
2684 Node* slow_ctl = transform_later(new ProjNode(slow_call, TypeFunc::Control));
2685 Node* slow_mem = transform_later(new ProjNode(slow_call, TypeFunc::Memory));
2686 Node* slow_io = transform_later(new ProjNode(slow_call, TypeFunc::I_O));
2687 Node* slow_res = transform_later(new ProjNode(slow_call, TypeFunc::Parms));
2688 Node* slow_catc = transform_later(new CatchNode(slow_ctl, slow_io, 2));
2689 Node* slow_norm = transform_later(new CatchProjNode(slow_catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci));
2690 Node* slow_excp = transform_later(new CatchProjNode(slow_catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci));
2691
2692 Node* ex_r = new RegionNode(3);
2693 Node* ex_mem_phi = new PhiNode(ex_r, Type::MEMORY, TypePtr::BOTTOM);
2694 Node* ex_io_phi = new PhiNode(ex_r, Type::ABIO);
2695 ex_r->init_req(1, slow_excp);
2696 ex_mem_phi->init_req(1, slow_mem);
2697 ex_io_phi->init_req(1, slow_io);
2698 ex_r->init_req(2, ex_ctl);
2699 ex_mem_phi->init_req(2, ex_mem);
2700 ex_io_phi->init_req(2, ex_io);
2701
2702 transform_later(ex_r);
2703 transform_later(ex_mem_phi);
2704 transform_later(ex_io_phi);
2705
2706 Node* slowpath_false = new IfFalseNode(slowpath_iff);
2707 transform_later(slowpath_false);
2708 Node* rawmem = new StorePNode(slowpath_false, mem, top_adr, TypeRawPtr::BOTTOM, new_top, MemNode::unordered);
2709 transform_later(rawmem);
2710 Node* mark_node = makecon(TypeRawPtr::make((address)markWord::always_locked_prototype().value()));
2711 rawmem = make_store(slowpath_false, rawmem, old_top, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
2712 rawmem = make_store(slowpath_false, rawmem, old_top, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
2713 if (UseCompressedClassPointers) {
2714 rawmem = make_store(slowpath_false, rawmem, old_top, oopDesc::klass_gap_offset_in_bytes(), intcon(0), T_INT);
2715 }
2716 Node* fixed_block = make_load(slowpath_false, rawmem, klass_node, in_bytes(InstanceKlass::adr_valueklass_fixed_block_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
2717 Node* pack_handler = make_load(slowpath_false, rawmem, fixed_block, in_bytes(ValueKlass::pack_handler_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
2718
2719 CallLeafNoFPNode* handler_call = new CallLeafNoFPNode(OptoRuntime::pack_value_type_Type(),
2720 NULL,
2721 "pack handler",
2722 TypeRawPtr::BOTTOM);
2723 handler_call->init_req(TypeFunc::Control, slowpath_false);
2724 handler_call->init_req(TypeFunc::Memory, rawmem);
2725 handler_call->init_req(TypeFunc::I_O, top());
2726 handler_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2727 handler_call->init_req(TypeFunc::ReturnAdr, top());
2728 handler_call->init_req(TypeFunc::Parms, pack_handler);
2729 handler_call->init_req(TypeFunc::Parms+1, old_top);
2730
2731 // We don't know how many values are returned. This assumes the
2732 // worst case, that all available registers are used.
2733 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) {
2734 if (domain->field_at(i) == Type::HALF) {
2735 slow_call->init_req(i, top());
2736 handler_call->init_req(i+1, top());
2737 continue;
2738 }
2739 Node* proj = transform_later(new ProjNode(call, i));
2740 slow_call->init_req(i, proj);
2741 handler_call->init_req(i+1, proj);
2742 }
2743
2744 // We can safepoint at that new call
2745 copy_call_debug_info(call, slow_call);
2746 transform_later(slow_call);
2747 transform_later(handler_call);
2748
2749 Node* handler_ctl = transform_later(new ProjNode(handler_call, TypeFunc::Control));
2750 rawmem = transform_later(new ProjNode(handler_call, TypeFunc::Memory));
2751 Node* slowpath_false_res = transform_later(new ProjNode(handler_call, TypeFunc::Parms));
2752
2753 MergeMemNode* slowpath_false_mem = MergeMemNode::make(mem);
2754 slowpath_false_mem->set_memory_at(Compile::AliasIdxRaw, rawmem);
2755 transform_later(slowpath_false_mem);
2756
2757 Node* r = new RegionNode(4);
2758 Node* mem_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM);
2759 Node* io_phi = new PhiNode(r, Type::ABIO);
2760 Node* res_phi = new PhiNode(r, TypeInstPtr::BOTTOM);
2761
2762 r->init_req(1, no_allocation_ctl);
2763 mem_phi->init_req(1, mem);
2764 io_phi->init_req(1, io);
2765 res_phi->init_req(1, no_allocation_res);
2766 r->init_req(2, slow_norm);
2767 mem_phi->init_req(2, slow_mem);
2768 io_phi->init_req(2, slow_io);
2769 res_phi->init_req(2, slow_res);
2770 r->init_req(3, handler_ctl);
2771 mem_phi->init_req(3, slowpath_false_mem);
2772 io_phi->init_req(3, io);
2773 res_phi->init_req(3, slowpath_false_res);
2774
2775 transform_later(r);
2776 transform_later(mem_phi);
2777 transform_later(io_phi);
2778 transform_later(res_phi);
2779
2780 assert(projs->nb_resproj == 1, "unexpected number of results");
2781 _igvn.replace_in_uses(projs->fallthrough_catchproj, r);
2782 _igvn.replace_in_uses(projs->fallthrough_memproj, mem_phi);
2783 _igvn.replace_in_uses(projs->fallthrough_ioproj, io_phi);
2784 _igvn.replace_in_uses(projs->resproj[0], res_phi);
2785 _igvn.replace_in_uses(projs->catchall_catchproj, ex_r);
2786 _igvn.replace_in_uses(projs->catchall_memproj, ex_mem_phi);
2787 _igvn.replace_in_uses(projs->catchall_ioproj, ex_io_phi);
2788
2789 _igvn.replace_node(ctl, projs->fallthrough_catchproj);
2790 _igvn.replace_node(mem, projs->fallthrough_memproj);
2791 _igvn.replace_node(io, projs->fallthrough_ioproj);
2792 _igvn.replace_node(res, projs->resproj[0]);
2793 _igvn.replace_node(ex_ctl, projs->catchall_catchproj);
2794 _igvn.replace_node(ex_mem, projs->catchall_memproj);
2795 _igvn.replace_node(ex_io, projs->catchall_ioproj);
2796 }
2797
2798 //---------------------------eliminate_macro_nodes----------------------
2799 // Eliminate scalar replaced allocations and associated locks.
2800 void PhaseMacroExpand::eliminate_macro_nodes() {
2801 if (C->macro_count() == 0)
2802 return;
2803
2804 // First, attempt to eliminate locks
2805 int cnt = C->macro_count();
2806 for (int i=0; i < cnt; i++) {
2807 Node *n = C->macro_node(i);
2808 if (n->is_AbstractLock()) { // Lock and Unlock nodes
2809 // Before elimination mark all associated (same box and obj)
2810 // lock and unlock nodes.
2811 mark_eliminated_locking_nodes(n->as_AbstractLock());
2812 }
2813 }
2814 bool progress = true;
2815 while (progress) {
2816 progress = false;
2817 for (int i = C->macro_count(); i > 0; i--) {
2818 Node * n = C->macro_node(i-1);
2819 bool success = false;
2820 debug_only(int old_macro_count = C->macro_count(););
2821 if (n->is_AbstractLock()) {
2822 success = eliminate_locking_node(n->as_AbstractLock());
2823 }
2824 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2825 progress = progress || success;
2826 }
2827 }
2828 // Next, attempt to eliminate allocations
2829 _has_locks = false;
2830 progress = true;
2831 while (progress) {
2832 progress = false;
2833 for (int i = C->macro_count(); i > 0; i--) {
2834 Node * n = C->macro_node(i-1);
2835 bool success = false;
2836 debug_only(int old_macro_count = C->macro_count(););
2837 switch (n->class_id()) {
2838 case Node::Class_Allocate:
2839 case Node::Class_AllocateArray:
2840 success = eliminate_allocate_node(n->as_Allocate());
2841 break;
2842 case Node::Class_CallStaticJava: {
2843 CallStaticJavaNode* call = n->as_CallStaticJava();
2844 if (!call->method()->is_method_handle_intrinsic()) {
2845 success = eliminate_boxing_node(n->as_CallStaticJava());
2846 }
2847 break;
2848 }
2849 case Node::Class_Lock:
2850 case Node::Class_Unlock:
2851 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2852 _has_locks = true;
2853 break;
2854 case Node::Class_ArrayCopy:
2855 break;
2856 case Node::Class_OuterStripMinedLoop:
2857 break;
2858 default:
2859 assert(n->Opcode() == Op_LoopLimit ||
2860 n->Opcode() == Op_Opaque1 ||
2861 n->Opcode() == Op_Opaque2 ||
2862 n->Opcode() == Op_Opaque3 ||
2863 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n),
2864 "unknown node type in macro list");
2865 }
2866 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2867 progress = progress || success;
2868 }
2869 }
2870 }
2871
2872 //------------------------------expand_macro_nodes----------------------
2873 // Returns true if a failure occurred.
2874 bool PhaseMacroExpand::expand_macro_nodes() {
2875 // Last attempt to eliminate macro nodes.
2876 eliminate_macro_nodes();
2877
2878 // Make sure expansion will not cause node limit to be exceeded.
2879 // Worst case is a macro node gets expanded into about 200 nodes.
2880 // Allow 50% more for optimization.
2881 if (C->check_node_count(C->macro_count() * 300, "out of nodes before macro expansion" ) )
2882 return true;
2883
2884 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2885 bool progress = true;
2886 while (progress) {
2887 progress = false;
2888 for (int i = C->macro_count(); i > 0; i--) {
2889 Node * n = C->macro_node(i-1);
2890 bool success = false;
2891 debug_only(int old_macro_count = C->macro_count(););
2892 if (n->Opcode() == Op_LoopLimit) {
2893 // Remove it from macro list and put on IGVN worklist to optimize.
2894 C->remove_macro_node(n);
2895 _igvn._worklist.push(n);
2896 success = true;
2897 } else if (n->Opcode() == Op_CallStaticJava) {
2898 CallStaticJavaNode* call = n->as_CallStaticJava();
2899 if (!call->method()->is_method_handle_intrinsic()) {
2900 // Remove it from macro list and put on IGVN worklist to optimize.
2901 C->remove_macro_node(n);
2902 _igvn._worklist.push(n);
2903 success = true;
2904 }
2905 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) {
2906 _igvn.replace_node(n, n->in(1));
2907 success = true;
2908 #if INCLUDE_RTM_OPT
2909 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) {
2910 assert(C->profile_rtm(), "should be used only in rtm deoptimization code");
2911 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), "");
2912 Node* cmp = n->unique_out();
2913 #ifdef ASSERT
2914 // Validate graph.
2915 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), "");
2916 BoolNode* bol = cmp->unique_out()->as_Bool();
2917 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() &&
2918 (bol->_test._test == BoolTest::ne), "");
2919 IfNode* ifn = bol->unique_out()->as_If();
2920 assert((ifn->outcnt() == 2) &&
2921 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != NULL, "");
2922 #endif
2923 Node* repl = n->in(1);
2924 if (!_has_locks) {
2925 // Remove RTM state check if there are no locks in the code.
2926 // Replace input to compare the same value.
2927 repl = (cmp->in(1) == n) ? cmp->in(2) : cmp->in(1);
2928 }
2929 _igvn.replace_node(n, repl);
2930 success = true;
2931 #endif
2932 } else if (n->Opcode() == Op_OuterStripMinedLoop) {
2933 n->as_OuterStripMinedLoop()->adjust_strip_mined_loop(&_igvn);
2934 C->remove_macro_node(n);
2935 success = true;
2936 }
2937 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2938 progress = progress || success;
2939 }
2940 }
2941
2942 // expand arraycopy "macro" nodes first
2943 // For ReduceBulkZeroing, we must first process all arraycopy nodes
2944 // before the allocate nodes are expanded.
2945 int macro_idx = C->macro_count() - 1;
2946 while (macro_idx >= 0) {
2947 Node * n = C->macro_node(macro_idx);
2948 assert(n->is_macro(), "only macro nodes expected here");
2949 if (_igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())) {
2950 // node is unreachable, so don't try to expand it
2951 C->remove_macro_node(n);
2952 } else if (n->is_ArrayCopy()){
2953 int macro_count = C->macro_count();
2954 expand_arraycopy_node(n->as_ArrayCopy());
2955 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2956 }
2957 if (C->failing()) return true;
2958 macro_idx --;
2959 }
2960
2961 // expand "macro" nodes
2962 // nodes are removed from the macro list as they are processed
2963 while (C->macro_count() > 0) {
2964 int macro_count = C->macro_count();
2965 Node * n = C->macro_node(macro_count-1);
2966 assert(n->is_macro(), "only macro nodes expected here");
2967 if (_igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())) {
2968 // node is unreachable, so don't try to expand it
2969 C->remove_macro_node(n);
2970 continue;
2971 }
2972 switch (n->class_id()) {
2973 case Node::Class_Allocate:
2974 expand_allocate(n->as_Allocate());
2975 break;
2976 case Node::Class_AllocateArray:
2977 expand_allocate_array(n->as_AllocateArray());
2978 break;
2979 case Node::Class_Lock:
2980 expand_lock_node(n->as_Lock());
2981 break;
2982 case Node::Class_Unlock:
2983 expand_unlock_node(n->as_Unlock());
2984 break;
2985 case Node::Class_CallStaticJava:
2986 expand_mh_intrinsic_return(n->as_CallStaticJava());
2987 C->remove_macro_node(n);
2988 break;
2989 default:
2990 assert(false, "unknown node type in macro list");
2991 }
2992 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2993 if (C->failing()) return true;
2994 }
2995
2996 _igvn.set_delay_transform(false);
2997 _igvn.optimize();
2998 if (C->failing()) return true;
2999 return false;
3000 }