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 "opto/addnode.hpp"
30 #include "opto/arraycopynode.hpp"
31 #include "opto/callnode.hpp"
32 #include "opto/castnode.hpp"
33 #include "opto/cfgnode.hpp"
34 #include "opto/compile.hpp"
35 #include "opto/convertnode.hpp"
36 #include "opto/graphKit.hpp"
37 #include "opto/locknode.hpp"
38 #include "opto/loopnode.hpp"
39 #include "opto/macro.hpp"
40 #include "opto/memnode.hpp"
41 #include "opto/narrowptrnode.hpp"
42 #include "opto/node.hpp"
43 #include "opto/opaquenode.hpp"
44 #include "opto/phaseX.hpp"
45 #include "opto/rootnode.hpp"
46 #include "opto/runtime.hpp"
47 #include "opto/subnode.hpp"
48 #include "opto/type.hpp"
49 #include "opto/valuetypenode.hpp"
50 #include "runtime/sharedRuntime.hpp"
51 #if INCLUDE_G1GC
52 #include "gc/g1/g1ThreadLocalData.hpp"
53 #endif // INCLUDE_G1GC
54
55
56 //
57 // Replace any references to "oldref" in inputs to "use" with "newref".
58 // Returns the number of replacements made.
59 //
60 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
61 int nreplacements = 0;
62 uint req = use->req();
63 for (uint j = 0; j < use->len(); j++) {
64 Node *uin = use->in(j);
65 if (uin == oldref) {
66 if (j < req)
67 use->set_req(j, newref);
68 else
69 use->set_prec(j, newref);
70 nreplacements++;
71 } else if (j >= req && uin == NULL) {
72 break;
73 }
74 }
75 return nreplacements;
76 }
77
78 void PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) {
79 // Copy debug information and adjust JVMState information
80 uint old_dbg_start = oldcall->tf()->domain_sig()->cnt();
81 uint new_dbg_start = newcall->tf()->domain_sig()->cnt();
82 int jvms_adj = new_dbg_start - old_dbg_start;
83 assert (new_dbg_start == newcall->req(), "argument count mismatch");
84
85 // SafePointScalarObject node could be referenced several times in debug info.
86 // Use Dict to record cloned nodes.
87 Dict* sosn_map = new Dict(cmpkey,hashkey);
88 for (uint i = old_dbg_start; i < oldcall->req(); i++) {
89 Node* old_in = oldcall->in(i);
90 // Clone old SafePointScalarObjectNodes, adjusting their field contents.
91 if (old_in != NULL && old_in->is_SafePointScalarObject()) {
92 SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject();
93 uint old_unique = C->unique();
94 Node* new_in = old_sosn->clone(sosn_map);
95 if (old_unique != C->unique()) { // New node?
96 new_in->set_req(0, C->root()); // reset control edge
97 new_in = transform_later(new_in); // Register new node.
98 }
99 old_in = new_in;
100 }
101 newcall->add_req(old_in);
102 }
103
104 // JVMS may be shared so clone it before we modify it
105 newcall->set_jvms(oldcall->jvms() != NULL ? oldcall->jvms()->clone_deep(C) : NULL);
106 for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) {
107 jvms->set_map(newcall);
108 jvms->set_locoff(jvms->locoff()+jvms_adj);
109 jvms->set_stkoff(jvms->stkoff()+jvms_adj);
110 jvms->set_monoff(jvms->monoff()+jvms_adj);
111 jvms->set_scloff(jvms->scloff()+jvms_adj);
112 jvms->set_endoff(jvms->endoff()+jvms_adj);
113 }
114 }
115
116 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
117 Node* cmp;
118 if (mask != 0) {
119 Node* and_node = transform_later(new AndXNode(word, MakeConX(mask)));
120 cmp = transform_later(new CmpXNode(and_node, MakeConX(bits)));
121 } else {
122 cmp = word;
123 }
124 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne));
125 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
126 transform_later(iff);
127
128 // Fast path taken.
129 Node *fast_taken = transform_later(new IfFalseNode(iff));
130
131 // Fast path not-taken, i.e. slow path
132 Node *slow_taken = transform_later(new IfTrueNode(iff));
133
134 if (return_fast_path) {
135 region->init_req(edge, slow_taken); // Capture slow-control
136 return fast_taken;
137 } else {
138 region->init_req(edge, fast_taken); // Capture fast-control
139 return slow_taken;
140 }
141 }
142
143 //--------------------copy_predefined_input_for_runtime_call--------------------
144 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
145 // Set fixed predefined input arguments
146 call->init_req( TypeFunc::Control, ctrl );
147 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) );
148 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
149 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
150 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
151 }
152
153 //------------------------------make_slow_call---------------------------------
154 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type,
155 address slow_call, const char* leaf_name, Node* slow_path,
156 Node* parm0, Node* parm1, Node* parm2) {
157
158 // Slow-path call
159 CallNode *call = leaf_name
160 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
161 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM );
162
163 // Slow path call has no side-effects, uses few values
164 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
165 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0);
166 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1);
167 if (parm2 != NULL) call->init_req(TypeFunc::Parms+2, parm2);
168 copy_call_debug_info(oldcall, call);
169 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
170 _igvn.replace_node(oldcall, call);
171 transform_later(call);
172
173 return call;
174 }
175
176 void PhaseMacroExpand::extract_call_projections(CallNode *call) {
177 _fallthroughproj = NULL;
178 _fallthroughcatchproj = NULL;
179 _ioproj_fallthrough = NULL;
180 _ioproj_catchall = NULL;
181 _catchallcatchproj = NULL;
182 _memproj_fallthrough = NULL;
183 _memproj_catchall = NULL;
184 _resproj = NULL;
185 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) {
186 ProjNode *pn = call->fast_out(i)->as_Proj();
187 switch (pn->_con) {
188 case TypeFunc::Control:
189 {
190 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj
191 _fallthroughproj = pn;
192 DUIterator_Fast jmax, j = pn->fast_outs(jmax);
193 const Node *cn = pn->fast_out(j);
194 if (cn->is_Catch()) {
195 ProjNode *cpn = NULL;
196 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) {
197 cpn = cn->fast_out(k)->as_Proj();
198 assert(cpn->is_CatchProj(), "must be a CatchProjNode");
199 if (cpn->_con == CatchProjNode::fall_through_index)
200 _fallthroughcatchproj = cpn;
201 else {
202 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index.");
203 _catchallcatchproj = cpn;
204 }
205 }
206 }
207 break;
208 }
209 case TypeFunc::I_O:
210 if (pn->_is_io_use)
211 _ioproj_catchall = pn;
212 else
213 _ioproj_fallthrough = pn;
214 break;
215 case TypeFunc::Memory:
216 if (pn->_is_io_use)
217 _memproj_catchall = pn;
218 else
219 _memproj_fallthrough = pn;
220 break;
221 case TypeFunc::Parms:
222 _resproj = pn;
223 break;
224 default:
225 assert(false, "unexpected projection from allocation node.");
226 }
227 }
228
229 }
230
231 void PhaseMacroExpand::eliminate_gc_barrier(Node* p2x) {
232 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
233 bs->eliminate_gc_barrier(this, p2x);
234 }
235
236 // Search for a memory operation for the specified memory slice.
237 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
238 Node *orig_mem = mem;
239 Node *alloc_mem = alloc->in(TypeFunc::Memory);
240 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
241 while (true) {
242 if (mem == alloc_mem || mem == start_mem ) {
243 return mem; // hit one of our sentinels
244 } else if (mem->is_MergeMem()) {
245 mem = mem->as_MergeMem()->memory_at(alias_idx);
246 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
247 Node *in = mem->in(0);
248 // we can safely skip over safepoints, calls, locks and membars because we
249 // already know that the object is safe to eliminate.
250 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
251 return in;
252 } else if (in->is_Call()) {
253 CallNode *call = in->as_Call();
254 if (call->may_modify(tinst, phase)) {
255 assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape");
256 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) {
257 return in;
258 }
259 }
260 mem = in->in(TypeFunc::Memory);
261 } else if (in->is_MemBar()) {
262 ArrayCopyNode* ac = NULL;
263 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) {
264 assert(ac != NULL && ac->is_clonebasic(), "Only basic clone is a non escaping clone");
265 return ac;
266 }
267 mem = in->in(TypeFunc::Memory);
268 } else {
269 assert(false, "unexpected projection");
270 }
271 } else if (mem->is_Store()) {
272 const TypePtr* atype = mem->as_Store()->adr_type();
273 int adr_idx = phase->C->get_alias_index(atype);
274 if (adr_idx == alias_idx) {
275 assert(atype->isa_oopptr(), "address type must be oopptr");
276 int adr_offset = atype->flattened_offset();
277 uint adr_iid = atype->is_oopptr()->instance_id();
278 // Array elements references have the same alias_idx
279 // but different offset and different instance_id.
280 if (adr_offset == offset && adr_iid == alloc->_idx)
281 return mem;
282 } else {
283 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
284 }
285 mem = mem->in(MemNode::Memory);
286 } else if (mem->is_ClearArray()) {
287 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
288 // Can not bypass initialization of the instance
289 // we are looking.
290 debug_only(intptr_t offset;)
291 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
292 InitializeNode* init = alloc->as_Allocate()->initialization();
293 // We are looking for stored value, return Initialize node
294 // or memory edge from Allocate node.
295 if (init != NULL)
296 return init;
297 else
298 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers).
299 }
300 // Otherwise skip it (the call updated 'mem' value).
301 } else if (mem->Opcode() == Op_SCMemProj) {
302 mem = mem->in(0);
303 Node* adr = NULL;
304 if (mem->is_LoadStore()) {
305 adr = mem->in(MemNode::Address);
306 } else {
307 assert(mem->Opcode() == Op_EncodeISOArray ||
308 mem->Opcode() == Op_StrCompressedCopy, "sanity");
309 adr = mem->in(3); // Destination array
310 }
311 const TypePtr* atype = adr->bottom_type()->is_ptr();
312 int adr_idx = phase->C->get_alias_index(atype);
313 if (adr_idx == alias_idx) {
314 DEBUG_ONLY(mem->dump();)
315 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
316 return NULL;
317 }
318 mem = mem->in(MemNode::Memory);
319 } else if (mem->Opcode() == Op_StrInflatedCopy) {
320 Node* adr = mem->in(3); // Destination array
321 const TypePtr* atype = adr->bottom_type()->is_ptr();
322 int adr_idx = phase->C->get_alias_index(atype);
323 if (adr_idx == alias_idx) {
324 DEBUG_ONLY(mem->dump();)
325 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
326 return NULL;
327 }
328 mem = mem->in(MemNode::Memory);
329 } else {
330 return mem;
331 }
332 assert(mem != orig_mem, "dead memory loop");
333 }
334 }
335
336 // Generate loads from source of the arraycopy for fields of
337 // destination needed at a deoptimization point
338 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) {
339 BasicType bt = ft;
340 const Type *type = ftype;
341 if (ft == T_NARROWOOP) {
342 bt = T_OBJECT;
343 type = ftype->make_oopptr();
344 }
345 Node* res = NULL;
346 if (ac->is_clonebasic()) {
347 Node* base = ac->in(ArrayCopyNode::Src)->in(AddPNode::Base);
348 Node* adr = _igvn.transform(new AddPNode(base, base, MakeConX(offset)));
349 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
350 res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned);
351 } else {
352 if (ac->modifies(offset, offset, &_igvn, true)) {
353 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
354 uint shift = exact_log2(type2aelembytes(bt));
355 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
356 #ifdef _LP64
357 diff = _igvn.transform(new ConvI2LNode(diff));
358 #endif
359 diff = _igvn.transform(new LShiftXNode(diff, intcon(shift)));
360
361 Node* off = _igvn.transform(new AddXNode(MakeConX(offset), diff));
362 Node* base = ac->in(ArrayCopyNode::Src);
363 Node* adr = _igvn.transform(new AddPNode(base, base, off));
364 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
365 res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned);
366 }
367 }
368 if (res != NULL) {
369 res = _igvn.transform(res);
370 if (ftype->isa_narrowoop()) {
371 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
372 res = _igvn.transform(new EncodePNode(res, ftype));
373 }
374 return res;
375 }
376 return NULL;
377 }
378
379 //
380 // Given a Memory Phi, compute a value Phi containing the values from stores
381 // on the input paths.
382 // Note: this function is recursive, its depth is limited by the "level" argument
383 // Returns the computed Phi, or NULL if it cannot compute it.
384 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) {
385 assert(mem->is_Phi(), "sanity");
386 int alias_idx = C->get_alias_index(adr_t);
387 int offset = adr_t->flattened_offset();
388 int instance_id = adr_t->instance_id();
389
390 // Check if an appropriate value phi already exists.
391 Node* region = mem->in(0);
392 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
393 Node* phi = region->fast_out(k);
394 if (phi->is_Phi() && phi != mem &&
395 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) {
396 return phi;
397 }
398 }
399 // Check if an appropriate new value phi already exists.
400 Node* new_phi = value_phis->find(mem->_idx);
401 if (new_phi != NULL)
402 return new_phi;
403
404 if (level <= 0) {
405 return NULL; // Give up: phi tree too deep
406 }
407 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
408 Node *alloc_mem = alloc->in(TypeFunc::Memory);
409
410 uint length = mem->req();
411 GrowableArray <Node *> values(length, length, NULL, false);
412
413 // create a new Phi for the value
414 PhiNode *phi = new PhiNode(mem->in(0), phi_type, NULL, mem->_idx, instance_id, alias_idx, offset);
415 transform_later(phi);
416 value_phis->push(phi, mem->_idx);
417
418 for (uint j = 1; j < length; j++) {
419 Node *in = mem->in(j);
420 if (in == NULL || in->is_top()) {
421 values.at_put(j, in);
422 } else {
423 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
424 if (val == start_mem || val == alloc_mem) {
425 // hit a sentinel, return appropriate 0 value
426 values.at_put(j, _igvn.zerocon(ft));
427 continue;
428 }
429 if (val->is_Initialize()) {
430 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
431 }
432 if (val == NULL) {
433 return NULL; // can't find a value on this path
434 }
435 if (val == mem) {
436 values.at_put(j, mem);
437 } else if (val->is_Store()) {
438 values.at_put(j, val->in(MemNode::ValueIn));
439 } else if(val->is_Proj() && val->in(0) == alloc) {
440 values.at_put(j, _igvn.zerocon(ft));
441 } else if (val->is_Phi()) {
442 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
443 if (val == NULL) {
444 return NULL;
445 }
446 values.at_put(j, val);
447 } else if (val->Opcode() == Op_SCMemProj) {
448 assert(val->in(0)->is_LoadStore() ||
449 val->in(0)->Opcode() == Op_EncodeISOArray ||
450 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity");
451 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
452 return NULL;
453 } else if (val->is_ArrayCopy()) {
454 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc);
455 if (res == NULL) {
456 return NULL;
457 }
458 values.at_put(j, res);
459 } else {
460 #ifdef ASSERT
461 val->dump();
462 assert(false, "unknown node on this path");
463 #endif
464 return NULL; // unknown node on this path
465 }
466 }
467 }
468 // Set Phi's inputs
469 for (uint j = 1; j < length; j++) {
470 if (values.at(j) == mem) {
471 phi->init_req(j, phi);
472 } else {
473 phi->init_req(j, values.at(j));
474 }
475 }
476 return phi;
477 }
478
479 // Search the last value stored into the object's field.
480 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) {
481 assert(adr_t->is_known_instance_field(), "instance required");
482 int instance_id = adr_t->instance_id();
483 assert((uint)instance_id == alloc->_idx, "wrong allocation");
484
485 int alias_idx = C->get_alias_index(adr_t);
486 int offset = adr_t->flattened_offset();
487 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
488 Node *alloc_ctrl = alloc->in(TypeFunc::Control);
489 Node *alloc_mem = alloc->in(TypeFunc::Memory);
490 Arena *a = Thread::current()->resource_area();
491 VectorSet visited(a);
492
493 bool done = sfpt_mem == alloc_mem;
494 Node *mem = sfpt_mem;
495 while (!done) {
496 if (visited.test_set(mem->_idx)) {
497 return NULL; // found a loop, give up
498 }
499 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
500 if (mem == start_mem || mem == alloc_mem) {
501 done = true; // hit a sentinel, return appropriate 0 value
502 } else if (mem->is_Initialize()) {
503 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
504 if (mem == NULL) {
505 done = true; // Something went wrong.
506 } else if (mem->is_Store()) {
507 const TypePtr* atype = mem->as_Store()->adr_type();
508 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
509 done = true;
510 }
511 } else if (mem->is_Store()) {
512 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
513 assert(atype != NULL, "address type must be oopptr");
514 assert(C->get_alias_index(atype) == alias_idx &&
515 atype->is_known_instance_field() && atype->flattened_offset() == offset &&
516 atype->instance_id() == instance_id, "store is correct memory slice");
517 done = true;
518 } else if (mem->is_Phi()) {
519 // try to find a phi's unique input
520 Node *unique_input = NULL;
521 Node *top = C->top();
522 for (uint i = 1; i < mem->req(); i++) {
523 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
524 if (n == NULL || n == top || n == mem) {
525 continue;
526 } else if (unique_input == NULL) {
527 unique_input = n;
528 } else if (unique_input != n) {
529 unique_input = top;
530 break;
531 }
532 }
533 if (unique_input != NULL && unique_input != top) {
534 mem = unique_input;
535 } else {
536 done = true;
537 }
538 } else if (mem->is_ArrayCopy()) {
539 done = true;
540 } else {
541 assert(false, "unexpected node");
542 }
543 }
544 if (mem != NULL) {
545 if (mem == start_mem || mem == alloc_mem) {
546 // hit a sentinel, return appropriate 0 value
547 Node* default_value = alloc->in(AllocateNode::DefaultValue);
548 if (default_value != NULL) {
549 return default_value;
550 }
551 assert(alloc->in(AllocateNode::RawDefaultValue) == NULL, "default value may not be null");
552 return _igvn.zerocon(ft);
553 } else if (mem->is_Store()) {
554 return mem->in(MemNode::ValueIn);
555 } else if (mem->is_Phi()) {
556 // attempt to produce a Phi reflecting the values on the input paths of the Phi
557 Node_Stack value_phis(a, 8);
558 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
559 if (phi != NULL) {
560 return phi;
561 } else {
562 // Kill all new Phis
563 while(value_phis.is_nonempty()) {
564 Node* n = value_phis.node();
565 _igvn.replace_node(n, C->top());
566 value_phis.pop();
567 }
568 }
569 } else if (mem->is_ArrayCopy()) {
570 Node* ctl = mem->in(0);
571 Node* m = mem->in(TypeFunc::Memory);
572 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) {
573 // pin the loads in the uncommon trap path
574 ctl = sfpt_ctl;
575 m = sfpt_mem;
576 }
577 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc);
578 }
579 }
580 // Something went wrong.
581 return NULL;
582 }
583
584 // Search the last value stored into the value type's fields.
585 Node* PhaseMacroExpand::value_type_from_mem(Node* mem, Node* ctl, ciValueKlass* vk, const TypeAryPtr* adr_type, int offset, AllocateNode* alloc) {
586 // Subtract the offset of the first field to account for the missing oop header
587 offset -= vk->first_field_offset();
588 // Create a new ValueTypeNode and retrieve the field values from memory
589 ValueTypeNode* vt = ValueTypeNode::make_uninitialized(_igvn, vk)->as_ValueType();
590 for (int i = 0; i < vk->nof_declared_nonstatic_fields(); ++i) {
591 ciType* field_type = vt->field_type(i);
592 int field_offset = offset + vt->field_offset(i);
593 // Each value type field has its own memory slice
594 adr_type = adr_type->with_field_offset(field_offset);
595 Node* value = NULL;
596 if (vt->field_is_flattened(i)) {
597 value = value_type_from_mem(mem, ctl, field_type->as_value_klass(), adr_type, field_offset, alloc);
598 } else {
599 const Type* ft = Type::get_const_type(field_type);
600 BasicType bt = field_type->basic_type();
601 if (UseCompressedOops && !is_java_primitive(bt)) {
602 ft = ft->make_narrowoop();
603 bt = T_NARROWOOP;
604 }
605 value = value_from_mem(mem, ctl, bt, ft, adr_type, alloc);
606 if (value != NULL && ft->isa_narrowoop()) {
607 assert(UseCompressedOops, "unexpected narrow oop");
608 value = transform_later(new DecodeNNode(value, value->get_ptr_type()));
609 }
610 /*
611 // TODO attempt scalarization here?
612 if (vt->field_is_flattenable(i)) {
613 // Loading a non-flattened but flattenable value type from memory
614 if (field_type->as_value_klass()->is_scalarizable()) {
615 value = ValueTypeNode::make_from_oop(kit, value, ft->as_value_klass());
616 } else {
617 value = kit->filter_null(value, true, ft->as_value_klass());
618 }
619 }
620 */
621 }
622 if (value != NULL) {
623 vt->set_field_value(i, value);
624 } else {
625 // We might have reached the TrackedInitializationLimit
626 return NULL;
627 }
628 }
629 return vt;
630 }
631
632 // Check the possibility of scalar replacement.
633 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
634 // Scan the uses of the allocation to check for anything that would
635 // prevent us from eliminating it.
636 NOT_PRODUCT( const char* fail_eliminate = NULL; )
637 DEBUG_ONLY( Node* disq_node = NULL; )
638 bool can_eliminate = true;
639
640 Node* res = alloc->result_cast();
641 const TypeOopPtr* res_type = NULL;
642 if (res == NULL) {
643 // All users were eliminated.
644 } else if (!res->is_CheckCastPP()) {
645 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
646 can_eliminate = false;
647 } else {
648 res_type = _igvn.type(res)->isa_oopptr();
649 if (res_type == NULL) {
650 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
651 can_eliminate = false;
652 } else if (res_type->isa_aryptr()) {
653 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
654 if (length < 0) {
655 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
656 can_eliminate = false;
657 }
658 }
659 }
660
661 if (can_eliminate && res != NULL) {
662 for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
663 j < jmax && can_eliminate; j++) {
664 Node* use = res->fast_out(j);
665
666 if (use->is_AddP()) {
667 const TypePtr* addp_type = _igvn.type(use)->is_ptr();
668 int offset = addp_type->offset();
669
670 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
671 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";)
672 can_eliminate = false;
673 break;
674 }
675 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
676 k < kmax && can_eliminate; k++) {
677 Node* n = use->fast_out(k);
678 if (!n->is_Store() && n->Opcode() != Op_CastP2X &&
679 !(n->is_ArrayCopy() &&
680 n->as_ArrayCopy()->is_clonebasic() &&
681 n->in(ArrayCopyNode::Dest) == use)) {
682 DEBUG_ONLY(disq_node = n;)
683 if (n->is_Load() || n->is_LoadStore()) {
684 NOT_PRODUCT(fail_eliminate = "Field load";)
685 } else {
686 NOT_PRODUCT(fail_eliminate = "Not store field reference";)
687 }
688 can_eliminate = false;
689 }
690 }
691 } else if (use->is_ArrayCopy() &&
692 (use->as_ArrayCopy()->is_arraycopy_validated() ||
693 use->as_ArrayCopy()->is_copyof_validated() ||
694 use->as_ArrayCopy()->is_copyofrange_validated()) &&
695 use->in(ArrayCopyNode::Dest) == res) {
696 // ok to eliminate
697 } else if (use->is_SafePoint()) {
698 SafePointNode* sfpt = use->as_SafePoint();
699 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
700 // Object is passed as argument.
701 DEBUG_ONLY(disq_node = use;)
702 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
703 can_eliminate = false;
704 }
705 Node* sfptMem = sfpt->memory();
706 if (sfptMem == NULL || sfptMem->is_top()) {
707 DEBUG_ONLY(disq_node = use;)
708 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";)
709 can_eliminate = false;
710 } else {
711 safepoints.append_if_missing(sfpt);
712 }
713 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
714 if (use->is_Phi()) {
715 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
716 NOT_PRODUCT(fail_eliminate = "Object is return value";)
717 } else {
718 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
719 }
720 DEBUG_ONLY(disq_node = use;)
721 } else {
722 if (use->Opcode() == Op_Return) {
723 NOT_PRODUCT(fail_eliminate = "Object is return value";)
724 } else {
725 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
726 }
727 DEBUG_ONLY(disq_node = use;)
728 }
729 can_eliminate = false;
730 } else {
731 assert(use->Opcode() == Op_CastP2X, "should be");
732 assert(!use->has_out_with(Op_OrL), "should have been removed because oop is never null");
733 }
734 }
735 }
736
737 #ifndef PRODUCT
738 if (PrintEliminateAllocations) {
739 if (can_eliminate) {
740 tty->print("Scalar ");
741 if (res == NULL)
742 alloc->dump();
743 else
744 res->dump();
745 } else if (alloc->_is_scalar_replaceable) {
746 tty->print("NotScalar (%s)", fail_eliminate);
747 if (res == NULL)
748 alloc->dump();
749 else
750 res->dump();
751 #ifdef ASSERT
752 if (disq_node != NULL) {
753 tty->print(" >>>> ");
754 disq_node->dump();
755 }
756 #endif /*ASSERT*/
757 }
758 }
759 #endif
760 return can_eliminate;
761 }
762
763 // Do scalar replacement.
764 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
765 GrowableArray <SafePointNode *> safepoints_done;
766
767 ciKlass* klass = NULL;
768 ciInstanceKlass* iklass = NULL;
769 int nfields = 0;
770 int array_base = 0;
771 int element_size = 0;
772 BasicType basic_elem_type = T_ILLEGAL;
773 ciType* elem_type = NULL;
774
775 Node* res = alloc->result_cast();
776 assert(res == NULL || res->is_CheckCastPP(), "unexpected AllocateNode result");
777 const TypeOopPtr* res_type = NULL;
778 if (res != NULL) { // Could be NULL when there are no users
779 res_type = _igvn.type(res)->isa_oopptr();
780 }
781
782 if (res != NULL) {
783 klass = res_type->klass();
784 if (res_type->isa_instptr()) {
785 // find the fields of the class which will be needed for safepoint debug information
786 assert(klass->is_instance_klass(), "must be an instance klass.");
787 iklass = klass->as_instance_klass();
788 nfields = iklass->nof_nonstatic_fields();
789 } else {
790 // find the array's elements which will be needed for safepoint debug information
791 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
792 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass.");
793 elem_type = klass->as_array_klass()->element_type();
794 basic_elem_type = elem_type->basic_type();
795 if (elem_type->is_valuetype()) {
796 ciValueKlass* vk = elem_type->as_value_klass();
797 if (!vk->flatten_array()) {
798 assert(basic_elem_type == T_VALUETYPE, "unexpected element basic type");
799 basic_elem_type = T_OBJECT;
800 }
801 }
802 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
803 element_size = type2aelembytes(basic_elem_type);
804 if (klass->is_value_array_klass()) {
805 // Flattened value type array
806 element_size = klass->as_value_array_klass()->element_byte_size();
807 }
808 }
809 }
810 //
811 // Process the safepoint uses
812 //
813 Unique_Node_List value_worklist;
814 while (safepoints.length() > 0) {
815 SafePointNode* sfpt = safepoints.pop();
816 Node* mem = sfpt->memory();
817 Node* ctl = sfpt->control();
818 assert(sfpt->jvms() != NULL, "missed JVMS");
819 // Fields of scalar objs are referenced only at the end
820 // of regular debuginfo at the last (youngest) JVMS.
821 // Record relative start index.
822 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
823 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type,
824 #ifdef ASSERT
825 alloc,
826 #endif
827 first_ind, nfields);
828 sobj->init_req(0, C->root());
829 transform_later(sobj);
830
831 // Scan object's fields adding an input to the safepoint for each field.
832 for (int j = 0; j < nfields; j++) {
833 intptr_t offset;
834 ciField* field = NULL;
835 if (iklass != NULL) {
836 field = iklass->nonstatic_field_at(j);
837 offset = field->offset();
838 elem_type = field->type();
839 basic_elem_type = field->layout_type();
840 assert(!field->is_flattened(), "flattened value type fields should not have safepoint uses");
841 } else {
842 offset = array_base + j * (intptr_t)element_size;
843 }
844
845 const Type *field_type;
846 // The next code is taken from Parse::do_get_xxx().
847 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) {
848 if (!elem_type->is_loaded()) {
849 field_type = TypeInstPtr::BOTTOM;
850 } else if (field != NULL && field->is_static_constant()) {
851 // This can happen if the constant oop is non-perm.
852 ciObject* con = field->constant_value().as_object();
853 // Do not "join" in the previous type; it doesn't add value,
854 // and may yield a vacuous result if the field is of interface type.
855 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
856 assert(field_type != NULL, "field singleton type must be consistent");
857 } else {
858 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
859 }
860 if (UseCompressedOops) {
861 field_type = field_type->make_narrowoop();
862 basic_elem_type = T_NARROWOOP;
863 }
864 } else {
865 field_type = Type::get_const_basic_type(basic_elem_type);
866 }
867
868 Node* field_val = NULL;
869 const TypeOopPtr* field_addr_type = res_type->add_offset(offset)->isa_oopptr();
870 if (klass->is_value_array_klass()) {
871 ciValueKlass* vk = elem_type->as_value_klass();
872 assert(vk->flatten_array(), "must be flattened");
873 field_val = value_type_from_mem(mem, ctl, vk, field_addr_type->isa_aryptr(), 0, alloc);
874 } else {
875 field_val = value_from_mem(mem, ctl, basic_elem_type, field_type, field_addr_type, alloc);
876 }
877 if (field_val == NULL) {
878 // We weren't able to find a value for this field,
879 // give up on eliminating this allocation.
880
881 // Remove any extra entries we added to the safepoint.
882 uint last = sfpt->req() - 1;
883 for (int k = 0; k < j; k++) {
884 sfpt->del_req(last--);
885 }
886 _igvn._worklist.push(sfpt);
887 // rollback processed safepoints
888 while (safepoints_done.length() > 0) {
889 SafePointNode* sfpt_done = safepoints_done.pop();
890 // remove any extra entries we added to the safepoint
891 last = sfpt_done->req() - 1;
892 for (int k = 0; k < nfields; k++) {
893 sfpt_done->del_req(last--);
894 }
895 JVMState *jvms = sfpt_done->jvms();
896 jvms->set_endoff(sfpt_done->req());
897 // Now make a pass over the debug information replacing any references
898 // to SafePointScalarObjectNode with the allocated object.
899 int start = jvms->debug_start();
900 int end = jvms->debug_end();
901 for (int i = start; i < end; i++) {
902 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
903 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
904 if (scobj->first_index(jvms) == sfpt_done->req() &&
905 scobj->n_fields() == (uint)nfields) {
906 assert(scobj->alloc() == alloc, "sanity");
907 sfpt_done->set_req(i, res);
908 }
909 }
910 }
911 _igvn._worklist.push(sfpt_done);
912 }
913 #ifndef PRODUCT
914 if (PrintEliminateAllocations) {
915 if (field != NULL) {
916 tty->print("=== At SafePoint node %d can't find value of Field: ",
917 sfpt->_idx);
918 field->print();
919 int field_idx = C->get_alias_index(field_addr_type);
920 tty->print(" (alias_idx=%d)", field_idx);
921 } else { // Array's element
922 tty->print("=== At SafePoint node %d can't find value of array element [%d]",
923 sfpt->_idx, j);
924 }
925 tty->print(", which prevents elimination of: ");
926 if (res == NULL)
927 alloc->dump();
928 else
929 res->dump();
930 }
931 #endif
932 return false;
933 }
934 if (UseCompressedOops && field_type->isa_narrowoop()) {
935 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
936 // to be able scalar replace the allocation.
937 if (field_val->is_EncodeP()) {
938 field_val = field_val->in(1);
939 } else {
940 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
941 }
942 } else if (field_val->is_ValueType()) {
943 // Keep track of value types to scalarize them later
944 value_worklist.push(field_val);
945 }
946 sfpt->add_req(field_val);
947 }
948 JVMState *jvms = sfpt->jvms();
949 jvms->set_endoff(sfpt->req());
950 // Now make a pass over the debug information replacing any references
951 // to the allocated object with "sobj"
952 int start = jvms->debug_start();
953 int end = jvms->debug_end();
954 sfpt->replace_edges_in_range(res, sobj, start, end);
955 _igvn._worklist.push(sfpt);
956 safepoints_done.append_if_missing(sfpt); // keep it for rollback
957 }
958 // Scalarize value types that were added to the safepoint
959 for (uint i = 0; i < value_worklist.size(); ++i) {
960 Node* vt = value_worklist.at(i);
961 vt->as_ValueType()->make_scalar_in_safepoints(C->root(), &_igvn);
962 }
963 return true;
964 }
965
966 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
967 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control);
968 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory);
969 if (ctl_proj != NULL) {
970 igvn.replace_node(ctl_proj, n->in(0));
971 }
972 if (mem_proj != NULL) {
973 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
974 }
975 }
976
977 // Process users of eliminated allocation.
978 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) {
979 Node* res = alloc->result_cast();
980 if (res != NULL) {
981 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
982 Node *use = res->last_out(j);
983 uint oc1 = res->outcnt();
984
985 if (use->is_AddP()) {
986 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
987 Node *n = use->last_out(k);
988 uint oc2 = use->outcnt();
989 if (n->is_Store()) {
990 #ifdef ASSERT
991 // Verify that there is no dependent MemBarVolatile nodes,
992 // they should be removed during IGVN, see MemBarNode::Ideal().
993 for (DUIterator_Fast pmax, p = n->fast_outs(pmax);
994 p < pmax; p++) {
995 Node* mb = n->fast_out(p);
996 assert(mb->is_Initialize() || !mb->is_MemBar() ||
997 mb->req() <= MemBarNode::Precedent ||
998 mb->in(MemBarNode::Precedent) != n,
999 "MemBarVolatile should be eliminated for non-escaping object");
1000 }
1001 #endif
1002 _igvn.replace_node(n, n->in(MemNode::Memory));
1003 } else if (n->is_ArrayCopy()) {
1004 // Disconnect ArrayCopy node
1005 ArrayCopyNode* ac = n->as_ArrayCopy();
1006 assert(ac->is_clonebasic(), "unexpected array copy kind");
1007 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
1008 disconnect_projections(ac, _igvn);
1009 assert(alloc->in(0)->is_Proj() && alloc->in(0)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
1010 Node* membar_before = alloc->in(0)->in(0);
1011 disconnect_projections(membar_before->as_MemBar(), _igvn);
1012 if (membar_after->is_MemBar()) {
1013 disconnect_projections(membar_after->as_MemBar(), _igvn);
1014 }
1015 } else {
1016 eliminate_gc_barrier(n);
1017 }
1018 k -= (oc2 - use->outcnt());
1019 }
1020 } else if (use->is_ArrayCopy()) {
1021 // Disconnect ArrayCopy node
1022 ArrayCopyNode* ac = use->as_ArrayCopy();
1023 assert(ac->is_arraycopy_validated() ||
1024 ac->is_copyof_validated() ||
1025 ac->is_copyofrange_validated(), "unsupported");
1026 CallProjections* callprojs = ac->extract_projections(true);
1027
1028 _igvn.replace_node(callprojs->fallthrough_ioproj, ac->in(TypeFunc::I_O));
1029 _igvn.replace_node(callprojs->fallthrough_memproj, ac->in(TypeFunc::Memory));
1030 _igvn.replace_node(callprojs->fallthrough_catchproj, ac->in(TypeFunc::Control));
1031
1032 // Set control to top. IGVN will remove the remaining projections
1033 ac->set_req(0, top());
1034 ac->replace_edge(res, top());
1035
1036 // Disconnect src right away: it can help find new
1037 // opportunities for allocation elimination
1038 Node* src = ac->in(ArrayCopyNode::Src);
1039 ac->replace_edge(src, top());
1040 // src can be top at this point if src and dest of the
1041 // arraycopy were the same
1042 if (src->outcnt() == 0 && !src->is_top()) {
1043 _igvn.remove_dead_node(src);
1044 }
1045
1046 _igvn._worklist.push(ac);
1047 } else {
1048 eliminate_gc_barrier(use);
1049 }
1050 j -= (oc1 - res->outcnt());
1051 }
1052 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
1053 _igvn.remove_dead_node(res);
1054 }
1055
1056 //
1057 // Process other users of allocation's projections
1058 //
1059 if (_resproj != NULL && _resproj->outcnt() != 0) {
1060 // First disconnect stores captured by Initialize node.
1061 // If Initialize node is eliminated first in the following code,
1062 // it will kill such stores and DUIterator_Last will assert.
1063 for (DUIterator_Fast jmax, j = _resproj->fast_outs(jmax); j < jmax; j++) {
1064 Node *use = _resproj->fast_out(j);
1065 if (use->is_AddP()) {
1066 // raw memory addresses used only by the initialization
1067 _igvn.replace_node(use, C->top());
1068 --j; --jmax;
1069 }
1070 }
1071 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) {
1072 Node *use = _resproj->last_out(j);
1073 uint oc1 = _resproj->outcnt();
1074 if (use->is_Initialize()) {
1075 // Eliminate Initialize node.
1076 InitializeNode *init = use->as_Initialize();
1077 assert(init->outcnt() <= 2, "only a control and memory projection expected");
1078 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control);
1079 if (ctrl_proj != NULL) {
1080 assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection");
1081 _igvn.replace_node(ctrl_proj, _fallthroughcatchproj);
1082 }
1083 Node *mem_proj = init->proj_out_or_null(TypeFunc::Memory);
1084 if (mem_proj != NULL) {
1085 Node *mem = init->in(TypeFunc::Memory);
1086 #ifdef ASSERT
1087 if (mem->is_MergeMem()) {
1088 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection");
1089 } else {
1090 assert(mem == _memproj_fallthrough, "allocation memory projection");
1091 }
1092 #endif
1093 _igvn.replace_node(mem_proj, mem);
1094 }
1095 } else {
1096 assert(false, "only Initialize or AddP expected");
1097 }
1098 j -= (oc1 - _resproj->outcnt());
1099 }
1100 }
1101 if (_fallthroughcatchproj != NULL) {
1102 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control));
1103 }
1104 if (_memproj_fallthrough != NULL) {
1105 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory));
1106 }
1107 if (_memproj_catchall != NULL) {
1108 _igvn.replace_node(_memproj_catchall, C->top());
1109 }
1110 if (_ioproj_fallthrough != NULL) {
1111 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O));
1112 }
1113 if (_ioproj_catchall != NULL) {
1114 _igvn.replace_node(_ioproj_catchall, C->top());
1115 }
1116 if (_catchallcatchproj != NULL) {
1117 _igvn.replace_node(_catchallcatchproj, C->top());
1118 }
1119 }
1120
1121 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1122 // Don't do scalar replacement if the frame can be popped by JVMTI:
1123 // if reallocation fails during deoptimization we'll pop all
1124 // interpreter frames for this compiled frame and that won't play
1125 // nice with JVMTI popframe.
1126 if (!EliminateAllocations || JvmtiExport::can_pop_frame() || !alloc->_is_non_escaping) {
1127 return false;
1128 }
1129 Node* klass = alloc->in(AllocateNode::KlassNode);
1130 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1131 Node* res = alloc->result_cast();
1132 // Eliminate boxing allocations which are not used
1133 // regardless scalar replacable status.
1134 bool boxing_alloc = C->eliminate_boxing() &&
1135 tklass->klass()->is_instance_klass() &&
1136 tklass->klass()->as_instance_klass()->is_box_klass();
1137 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != NULL))) {
1138 return false;
1139 }
1140
1141 extract_call_projections(alloc);
1142
1143 GrowableArray <SafePointNode *> safepoints;
1144 if (!can_eliminate_allocation(alloc, safepoints)) {
1145 return false;
1146 }
1147
1148 if (!alloc->_is_scalar_replaceable) {
1149 assert(res == NULL, "sanity");
1150 // We can only eliminate allocation if all debug info references
1151 // are already replaced with SafePointScalarObject because
1152 // we can't search for a fields value without instance_id.
1153 if (safepoints.length() > 0) {
1154 return false;
1155 }
1156 }
1157
1158 if (!scalar_replacement(alloc, safepoints)) {
1159 return false;
1160 }
1161
1162 CompileLog* log = C->log();
1163 if (log != NULL) {
1164 log->head("eliminate_allocation type='%d'",
1165 log->identify(tklass->klass()));
1166 JVMState* p = alloc->jvms();
1167 while (p != NULL) {
1168 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1169 p = p->caller();
1170 }
1171 log->tail("eliminate_allocation");
1172 }
1173
1174 process_users_of_allocation(alloc);
1175
1176 #ifndef PRODUCT
1177 if (PrintEliminateAllocations) {
1178 if (alloc->is_AllocateArray())
1179 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1180 else
1181 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1182 }
1183 #endif
1184
1185 return true;
1186 }
1187
1188 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1189 // EA should remove all uses of non-escaping boxing node.
1190 if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != NULL) {
1191 return false;
1192 }
1193
1194 assert(boxing->result_cast() == NULL, "unexpected boxing node result");
1195
1196 extract_call_projections(boxing);
1197
1198 const TypeTuple* r = boxing->tf()->range_sig();
1199 assert(r->cnt() > TypeFunc::Parms, "sanity");
1200 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1201 assert(t != NULL, "sanity");
1202
1203 CompileLog* log = C->log();
1204 if (log != NULL) {
1205 log->head("eliminate_boxing type='%d'",
1206 log->identify(t->klass()));
1207 JVMState* p = boxing->jvms();
1208 while (p != NULL) {
1209 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1210 p = p->caller();
1211 }
1212 log->tail("eliminate_boxing");
1213 }
1214
1215 process_users_of_allocation(boxing);
1216
1217 #ifndef PRODUCT
1218 if (PrintEliminateAllocations) {
1219 tty->print("++++ Eliminated: %d ", boxing->_idx);
1220 boxing->method()->print_short_name(tty);
1221 tty->cr();
1222 }
1223 #endif
1224
1225 return true;
1226 }
1227
1228 //---------------------------set_eden_pointers-------------------------
1229 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) {
1230 if (UseTLAB) { // Private allocation: load from TLS
1231 Node* thread = transform_later(new ThreadLocalNode());
1232 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset());
1233 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset());
1234 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset);
1235 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset);
1236 } else { // Shared allocation: load from globals
1237 CollectedHeap* ch = Universe::heap();
1238 address top_adr = (address)ch->top_addr();
1239 address end_adr = (address)ch->end_addr();
1240 eden_top_adr = makecon(TypeRawPtr::make(top_adr));
1241 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr);
1242 }
1243 }
1244
1245
1246 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
1247 Node* adr = basic_plus_adr(base, offset);
1248 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
1249 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered);
1250 transform_later(value);
1251 return value;
1252 }
1253
1254
1255 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
1256 Node* adr = basic_plus_adr(base, offset);
1257 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt, MemNode::unordered);
1258 transform_later(mem);
1259 return mem;
1260 }
1261
1262 //=============================================================================
1263 //
1264 // A L L O C A T I O N
1265 //
1266 // Allocation attempts to be fast in the case of frequent small objects.
1267 // It breaks down like this:
1268 //
1269 // 1) Size in doublewords is computed. This is a constant for objects and
1270 // variable for most arrays. Doubleword units are used to avoid size
1271 // overflow of huge doubleword arrays. We need doublewords in the end for
1272 // rounding.
1273 //
1274 // 2) Size is checked for being 'too large'. Too-large allocations will go
1275 // the slow path into the VM. The slow path can throw any required
1276 // exceptions, and does all the special checks for very large arrays. The
1277 // size test can constant-fold away for objects. For objects with
1278 // finalizers it constant-folds the otherway: you always go slow with
1279 // finalizers.
1280 //
1281 // 3) If NOT using TLABs, this is the contended loop-back point.
1282 // Load-Locked the heap top. If using TLABs normal-load the heap top.
1283 //
1284 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
1285 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
1286 // "size*8" we always enter the VM, where "largish" is a constant picked small
1287 // enough that there's always space between the eden max and 4Gig (old space is
1288 // there so it's quite large) and large enough that the cost of entering the VM
1289 // is dwarfed by the cost to initialize the space.
1290 //
1291 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1292 // down. If contended, repeat at step 3. If using TLABs normal-store
1293 // adjusted heap top back down; there is no contention.
1294 //
1295 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
1296 // fields.
1297 //
1298 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1299 // oop flavor.
1300 //
1301 //=============================================================================
1302 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1303 // Allocations bigger than this always go the slow route.
1304 // This value must be small enough that allocation attempts that need to
1305 // trigger exceptions go the slow route. Also, it must be small enough so
1306 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1307 //=============================================================================j//
1308 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1309 // The allocator will coalesce int->oop copies away. See comment in
1310 // coalesce.cpp about how this works. It depends critically on the exact
1311 // code shape produced here, so if you are changing this code shape
1312 // make sure the GC info for the heap-top is correct in and around the
1313 // slow-path call.
1314 //
1315
1316 void PhaseMacroExpand::expand_allocate_common(
1317 AllocateNode* alloc, // allocation node to be expanded
1318 Node* length, // array length for an array allocation
1319 const TypeFunc* slow_call_type, // Type of slow call
1320 address slow_call_address // Address of slow call
1321 )
1322 {
1323
1324 Node* ctrl = alloc->in(TypeFunc::Control);
1325 Node* mem = alloc->in(TypeFunc::Memory);
1326 Node* i_o = alloc->in(TypeFunc::I_O);
1327 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1328 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1329 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1330
1331 assert(ctrl != NULL, "must have control");
1332 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1333 // they will not be used if "always_slow" is set
1334 enum { slow_result_path = 1, fast_result_path = 2 };
1335 Node *result_region = NULL;
1336 Node *result_phi_rawmem = NULL;
1337 Node *result_phi_rawoop = NULL;
1338 Node *result_phi_i_o = NULL;
1339
1340 // The initial slow comparison is a size check, the comparison
1341 // we want to do is a BoolTest::gt
1342 bool always_slow = false;
1343 int tv = _igvn.find_int_con(initial_slow_test, -1);
1344 if (tv >= 0) {
1345 always_slow = (tv == 1);
1346 initial_slow_test = NULL;
1347 } else {
1348 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1349 }
1350
1351 if (C->env()->dtrace_alloc_probes() ||
1352 (!UseTLAB && !Universe::heap()->supports_inline_contig_alloc())) {
1353 // Force slow-path allocation
1354 always_slow = true;
1355 initial_slow_test = NULL;
1356 }
1357
1358 Node *slow_region = NULL;
1359 Node *toobig_false = ctrl;
1360
1361 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent");
1362 // generate the initial test if necessary
1363 if (initial_slow_test != NULL ) {
1364 if (slow_region == NULL) {
1365 slow_region = new RegionNode(1);
1366 }
1367 // Now make the initial failure test. Usually a too-big test but
1368 // might be a TRUE for finalizers or a fancy class check for
1369 // newInstance0.
1370 IfNode* toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1371 transform_later(toobig_iff);
1372 // Plug the failing-too-big test into the slow-path region
1373 Node* toobig_true = new IfTrueNode(toobig_iff);
1374 transform_later(toobig_true);
1375 slow_region ->add_req(toobig_true);
1376 toobig_false = new IfFalseNode(toobig_iff);
1377 transform_later(toobig_false);
1378 } else { // No initial test, just fall into next case
1379 toobig_false = ctrl;
1380 }
1381
1382 Node *slow_mem = mem; // save the current memory state for slow path
1383 // generate the fast allocation code unless we know that the initial test will always go slow
1384 if (!always_slow) {
1385 // Fast path modifies only raw memory.
1386 if (mem->is_MergeMem()) {
1387 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1388 }
1389
1390 Node* eden_top_adr;
1391 Node* eden_end_adr;
1392
1393 set_eden_pointers(eden_top_adr, eden_end_adr);
1394
1395 // Load Eden::end. Loop invariant and hoisted.
1396 //
1397 // Note: We set the control input on "eden_end" and "old_eden_top" when using
1398 // a TLAB to work around a bug where these values were being moved across
1399 // a safepoint. These are not oops, so they cannot be include in the oop
1400 // map, but they can be changed by a GC. The proper way to fix this would
1401 // be to set the raw memory state when generating a SafepointNode. However
1402 // this will require extensive changes to the loop optimization in order to
1403 // prevent a degradation of the optimization.
1404 // See comment in memnode.hpp, around line 227 in class LoadPNode.
1405 Node *eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS);
1406
1407 // allocate the Region and Phi nodes for the result
1408 result_region = new RegionNode(3);
1409 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1410 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1411 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1412
1413 // We need a Region for the loop-back contended case.
1414 enum { fall_in_path = 1, contended_loopback_path = 2 };
1415 Node *contended_region;
1416 Node *contended_phi_rawmem;
1417 if (UseTLAB) {
1418 contended_region = toobig_false;
1419 contended_phi_rawmem = mem;
1420 } else {
1421 contended_region = new RegionNode(3);
1422 contended_phi_rawmem = new PhiNode(contended_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1423 // Now handle the passing-too-big test. We fall into the contended
1424 // loop-back merge point.
1425 contended_region ->init_req(fall_in_path, toobig_false);
1426 contended_phi_rawmem->init_req(fall_in_path, mem);
1427 transform_later(contended_region);
1428 transform_later(contended_phi_rawmem);
1429 }
1430
1431 // Load(-locked) the heap top.
1432 // See note above concerning the control input when using a TLAB
1433 Node *old_eden_top = UseTLAB
1434 ? new LoadPNode (ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered)
1435 : new LoadPLockedNode(contended_region, contended_phi_rawmem, eden_top_adr, MemNode::acquire);
1436
1437 transform_later(old_eden_top);
1438 // Add to heap top to get a new heap top
1439 Node *new_eden_top = new AddPNode(top(), old_eden_top, size_in_bytes);
1440 transform_later(new_eden_top);
1441 // Check for needing a GC; compare against heap end
1442 Node *needgc_cmp = new CmpPNode(new_eden_top, eden_end);
1443 transform_later(needgc_cmp);
1444 Node *needgc_bol = new BoolNode(needgc_cmp, BoolTest::ge);
1445 transform_later(needgc_bol);
1446 IfNode *needgc_iff = new IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN);
1447 transform_later(needgc_iff);
1448
1449 // Plug the failing-heap-space-need-gc test into the slow-path region
1450 Node *needgc_true = new IfTrueNode(needgc_iff);
1451 transform_later(needgc_true);
1452 if (slow_region != NULL) {
1453 slow_region->add_req(needgc_true);
1454 // This completes all paths into the slow merge point
1455 transform_later(slow_region);
1456 } else {
1457 // Just fall from the need-GC path straight into the VM call.
1458 slow_region = needgc_true;
1459 }
1460 // No need for a GC. Setup for the Store-Conditional
1461 Node *needgc_false = new IfFalseNode(needgc_iff);
1462 transform_later(needgc_false);
1463
1464 // Grab regular I/O before optional prefetch may change it.
1465 // Slow-path does no I/O so just set it to the original I/O.
1466 result_phi_i_o->init_req(slow_result_path, i_o);
1467
1468 i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem,
1469 old_eden_top, new_eden_top, length);
1470
1471 // Name successful fast-path variables
1472 Node* fast_oop = old_eden_top;
1473 Node* fast_oop_ctrl;
1474 Node* fast_oop_rawmem;
1475
1476 // Store (-conditional) the modified eden top back down.
1477 // StorePConditional produces flags for a test PLUS a modified raw
1478 // memory state.
1479 if (UseTLAB) {
1480 Node* store_eden_top =
1481 new StorePNode(needgc_false, contended_phi_rawmem, eden_top_adr,
1482 TypeRawPtr::BOTTOM, new_eden_top, MemNode::unordered);
1483 transform_later(store_eden_top);
1484 fast_oop_ctrl = needgc_false; // No contention, so this is the fast path
1485 fast_oop_rawmem = store_eden_top;
1486 } else {
1487 Node* store_eden_top =
1488 new StorePConditionalNode(needgc_false, contended_phi_rawmem, eden_top_adr,
1489 new_eden_top, fast_oop/*old_eden_top*/);
1490 transform_later(store_eden_top);
1491 Node *contention_check = new BoolNode(store_eden_top, BoolTest::ne);
1492 transform_later(contention_check);
1493 store_eden_top = new SCMemProjNode(store_eden_top);
1494 transform_later(store_eden_top);
1495
1496 // If not using TLABs, check to see if there was contention.
1497 IfNode *contention_iff = new IfNode (needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN);
1498 transform_later(contention_iff);
1499 Node *contention_true = new IfTrueNode(contention_iff);
1500 transform_later(contention_true);
1501 // If contention, loopback and try again.
1502 contended_region->init_req(contended_loopback_path, contention_true);
1503 contended_phi_rawmem->init_req(contended_loopback_path, store_eden_top);
1504
1505 // Fast-path succeeded with no contention!
1506 Node *contention_false = new IfFalseNode(contention_iff);
1507 transform_later(contention_false);
1508 fast_oop_ctrl = contention_false;
1509
1510 // Bump total allocated bytes for this thread
1511 Node* thread = new ThreadLocalNode();
1512 transform_later(thread);
1513 Node* alloc_bytes_adr = basic_plus_adr(top()/*not oop*/, thread,
1514 in_bytes(JavaThread::allocated_bytes_offset()));
1515 Node* alloc_bytes = make_load(fast_oop_ctrl, store_eden_top, alloc_bytes_adr,
1516 0, TypeLong::LONG, T_LONG);
1517 #ifdef _LP64
1518 Node* alloc_size = size_in_bytes;
1519 #else
1520 Node* alloc_size = new ConvI2LNode(size_in_bytes);
1521 transform_later(alloc_size);
1522 #endif
1523 Node* new_alloc_bytes = new AddLNode(alloc_bytes, alloc_size);
1524 transform_later(new_alloc_bytes);
1525 fast_oop_rawmem = make_store(fast_oop_ctrl, store_eden_top, alloc_bytes_adr,
1526 0, new_alloc_bytes, T_LONG);
1527 }
1528
1529 InitializeNode* init = alloc->initialization();
1530 fast_oop_rawmem = initialize_object(alloc,
1531 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1532 klass_node, length, size_in_bytes);
1533
1534 // If initialization is performed by an array copy, any required
1535 // MemBarStoreStore was already added. If the object does not
1536 // escape no need for a MemBarStoreStore. If the object does not
1537 // escape in its initializer and memory barrier (MemBarStoreStore or
1538 // stronger) is already added at exit of initializer, also no need
1539 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore
1540 // so that stores that initialize this object can't be reordered
1541 // with a subsequent store that makes this object accessible by
1542 // other threads.
1543 // Other threads include java threads and JVM internal threads
1544 // (for example concurrent GC threads). Current concurrent GC
1545 // implementation: CMS and G1 will not scan newly created object,
1546 // so it's safe to skip storestore barrier when allocation does
1547 // not escape.
1548 if (!alloc->does_not_escape_thread() &&
1549 !alloc->is_allocation_MemBar_redundant() &&
1550 (init == NULL || !init->is_complete_with_arraycopy())) {
1551 if (init == NULL || init->req() < InitializeNode::RawStores) {
1552 // No InitializeNode or no stores captured by zeroing
1553 // elimination. Simply add the MemBarStoreStore after object
1554 // initialization.
1555 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1556 transform_later(mb);
1557
1558 mb->init_req(TypeFunc::Memory, fast_oop_rawmem);
1559 mb->init_req(TypeFunc::Control, fast_oop_ctrl);
1560 fast_oop_ctrl = new ProjNode(mb,TypeFunc::Control);
1561 transform_later(fast_oop_ctrl);
1562 fast_oop_rawmem = new ProjNode(mb,TypeFunc::Memory);
1563 transform_later(fast_oop_rawmem);
1564 } else {
1565 // Add the MemBarStoreStore after the InitializeNode so that
1566 // all stores performing the initialization that were moved
1567 // before the InitializeNode happen before the storestore
1568 // barrier.
1569
1570 Node* init_ctrl = init->proj_out_or_null(TypeFunc::Control);
1571 Node* init_mem = init->proj_out_or_null(TypeFunc::Memory);
1572
1573 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1574 transform_later(mb);
1575
1576 Node* ctrl = new ProjNode(init,TypeFunc::Control);
1577 transform_later(ctrl);
1578 Node* mem = new ProjNode(init,TypeFunc::Memory);
1579 transform_later(mem);
1580
1581 // The MemBarStoreStore depends on control and memory coming
1582 // from the InitializeNode
1583 mb->init_req(TypeFunc::Memory, mem);
1584 mb->init_req(TypeFunc::Control, ctrl);
1585
1586 ctrl = new ProjNode(mb,TypeFunc::Control);
1587 transform_later(ctrl);
1588 mem = new ProjNode(mb,TypeFunc::Memory);
1589 transform_later(mem);
1590
1591 // All nodes that depended on the InitializeNode for control
1592 // and memory must now depend on the MemBarNode that itself
1593 // depends on the InitializeNode
1594 if (init_ctrl != NULL) {
1595 _igvn.replace_node(init_ctrl, ctrl);
1596 }
1597 if (init_mem != NULL) {
1598 _igvn.replace_node(init_mem, mem);
1599 }
1600 }
1601 }
1602
1603 if (C->env()->dtrace_extended_probes()) {
1604 // Slow-path call
1605 int size = TypeFunc::Parms + 2;
1606 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1607 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base),
1608 "dtrace_object_alloc",
1609 TypeRawPtr::BOTTOM);
1610
1611 // Get base of thread-local storage area
1612 Node* thread = new ThreadLocalNode();
1613 transform_later(thread);
1614
1615 call->init_req(TypeFunc::Parms+0, thread);
1616 call->init_req(TypeFunc::Parms+1, fast_oop);
1617 call->init_req(TypeFunc::Control, fast_oop_ctrl);
1618 call->init_req(TypeFunc::I_O , top()); // does no i/o
1619 call->init_req(TypeFunc::Memory , fast_oop_rawmem);
1620 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1621 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1622 transform_later(call);
1623 fast_oop_ctrl = new ProjNode(call,TypeFunc::Control);
1624 transform_later(fast_oop_ctrl);
1625 fast_oop_rawmem = new ProjNode(call,TypeFunc::Memory);
1626 transform_later(fast_oop_rawmem);
1627 }
1628
1629 // Plug in the successful fast-path into the result merge point
1630 result_region ->init_req(fast_result_path, fast_oop_ctrl);
1631 result_phi_rawoop->init_req(fast_result_path, fast_oop);
1632 result_phi_i_o ->init_req(fast_result_path, i_o);
1633 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1634 } else {
1635 slow_region = ctrl;
1636 result_phi_i_o = i_o; // Rename it to use in the following code.
1637 }
1638
1639 // Generate slow-path call
1640 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1641 OptoRuntime::stub_name(slow_call_address),
1642 alloc->jvms()->bci(),
1643 TypePtr::BOTTOM);
1644 call->init_req( TypeFunc::Control, slow_region );
1645 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o
1646 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs
1647 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) );
1648 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) );
1649
1650 call->init_req(TypeFunc::Parms+0, klass_node);
1651 if (length != NULL) {
1652 call->init_req(TypeFunc::Parms+1, length);
1653 }
1654
1655 // Copy debug information and adjust JVMState information, then replace
1656 // allocate node with the call
1657 copy_call_debug_info((CallNode *) alloc, call);
1658 if (!always_slow) {
1659 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1660 } else {
1661 // Hook i_o projection to avoid its elimination during allocation
1662 // replacement (when only a slow call is generated).
1663 call->set_req(TypeFunc::I_O, result_phi_i_o);
1664 }
1665 _igvn.replace_node(alloc, call);
1666 transform_later(call);
1667
1668 // Identify the output projections from the allocate node and
1669 // adjust any references to them.
1670 // The control and io projections look like:
1671 //
1672 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1673 // Allocate Catch
1674 // ^---Proj(io) <-------+ ^---CatchProj(io)
1675 //
1676 // We are interested in the CatchProj nodes.
1677 //
1678 extract_call_projections(call);
1679
1680 // An allocate node has separate memory projections for the uses on
1681 // the control and i_o paths. Replace the control memory projection with
1682 // result_phi_rawmem (unless we are only generating a slow call when
1683 // both memory projections are combined)
1684 if (!always_slow && _memproj_fallthrough != NULL) {
1685 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) {
1686 Node *use = _memproj_fallthrough->fast_out(i);
1687 _igvn.rehash_node_delayed(use);
1688 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem);
1689 // back up iterator
1690 --i;
1691 }
1692 }
1693 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete
1694 // _memproj_catchall so we end up with a call that has only 1 memory projection.
1695 if (_memproj_catchall != NULL ) {
1696 if (_memproj_fallthrough == NULL) {
1697 _memproj_fallthrough = new ProjNode(call, TypeFunc::Memory);
1698 transform_later(_memproj_fallthrough);
1699 }
1700 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) {
1701 Node *use = _memproj_catchall->fast_out(i);
1702 _igvn.rehash_node_delayed(use);
1703 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough);
1704 // back up iterator
1705 --i;
1706 }
1707 assert(_memproj_catchall->outcnt() == 0, "all uses must be deleted");
1708 _igvn.remove_dead_node(_memproj_catchall);
1709 }
1710
1711 // An allocate node has separate i_o projections for the uses on the control
1712 // and i_o paths. Always replace the control i_o projection with result i_o
1713 // otherwise incoming i_o become dead when only a slow call is generated
1714 // (it is different from memory projections where both projections are
1715 // combined in such case).
1716 if (_ioproj_fallthrough != NULL) {
1717 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) {
1718 Node *use = _ioproj_fallthrough->fast_out(i);
1719 _igvn.rehash_node_delayed(use);
1720 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o);
1721 // back up iterator
1722 --i;
1723 }
1724 }
1725 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete
1726 // _ioproj_catchall so we end up with a call that has only 1 i_o projection.
1727 if (_ioproj_catchall != NULL ) {
1728 if (_ioproj_fallthrough == NULL) {
1729 _ioproj_fallthrough = new ProjNode(call, TypeFunc::I_O);
1730 transform_later(_ioproj_fallthrough);
1731 }
1732 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) {
1733 Node *use = _ioproj_catchall->fast_out(i);
1734 _igvn.rehash_node_delayed(use);
1735 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough);
1736 // back up iterator
1737 --i;
1738 }
1739 assert(_ioproj_catchall->outcnt() == 0, "all uses must be deleted");
1740 _igvn.remove_dead_node(_ioproj_catchall);
1741 }
1742
1743 // if we generated only a slow call, we are done
1744 if (always_slow) {
1745 // Now we can unhook i_o.
1746 if (result_phi_i_o->outcnt() > 1) {
1747 call->set_req(TypeFunc::I_O, top());
1748 } else {
1749 assert(result_phi_i_o->unique_ctrl_out() == call, "");
1750 // Case of new array with negative size known during compilation.
1751 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1752 // following code since call to runtime will throw exception.
1753 // As result there will be no users of i_o after the call.
1754 // Leave i_o attached to this call to avoid problems in preceding graph.
1755 }
1756 return;
1757 }
1758
1759
1760 if (_fallthroughcatchproj != NULL) {
1761 ctrl = _fallthroughcatchproj->clone();
1762 transform_later(ctrl);
1763 _igvn.replace_node(_fallthroughcatchproj, result_region);
1764 } else {
1765 ctrl = top();
1766 }
1767 Node *slow_result;
1768 if (_resproj == NULL) {
1769 // no uses of the allocation result
1770 slow_result = top();
1771 } else {
1772 slow_result = _resproj->clone();
1773 transform_later(slow_result);
1774 _igvn.replace_node(_resproj, result_phi_rawoop);
1775 }
1776
1777 // Plug slow-path into result merge point
1778 result_region ->init_req( slow_result_path, ctrl );
1779 result_phi_rawoop->init_req( slow_result_path, slow_result);
1780 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough );
1781 transform_later(result_region);
1782 transform_later(result_phi_rawoop);
1783 transform_later(result_phi_rawmem);
1784 transform_later(result_phi_i_o);
1785 // This completes all paths into the result merge point
1786 }
1787
1788
1789 // Helper for PhaseMacroExpand::expand_allocate_common.
1790 // Initializes the newly-allocated storage.
1791 Node* PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1792 Node* control, Node* rawmem, Node* object,
1793 Node* klass_node, Node* length,
1794 Node* size_in_bytes) {
1795 InitializeNode* init = alloc->initialization();
1796 // Store the klass & mark bits
1797 Node* mark_node = NULL;
1798 // For now only enable fast locking for non-array types
1799 if ((EnableValhalla || UseBiasedLocking) && length == NULL) {
1800 mark_node = make_load(control, rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
1801 } else {
1802 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype()));
1803 }
1804 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
1805
1806 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1807 int header_size = alloc->minimum_header_size(); // conservatively small
1808
1809 // Array length
1810 if (length != NULL) { // Arrays need length field
1811 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1812 // conservatively small header size:
1813 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1814 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1815 if (k->is_array_klass()) // we know the exact header size in most cases:
1816 header_size = Klass::layout_helper_header_size(k->layout_helper());
1817 }
1818
1819 // Clear the object body, if necessary.
1820 if (init == NULL) {
1821 // The init has somehow disappeared; be cautious and clear everything.
1822 //
1823 // This can happen if a node is allocated but an uncommon trap occurs
1824 // immediately. In this case, the Initialize gets associated with the
1825 // trap, and may be placed in a different (outer) loop, if the Allocate
1826 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1827 // there can be two Allocates to one Initialize. The answer in all these
1828 // edge cases is safety first. It is always safe to clear immediately
1829 // within an Allocate, and then (maybe or maybe not) clear some more later.
1830 if (!(UseTLAB && ZeroTLAB)) {
1831 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1832 alloc->in(AllocateNode::DefaultValue),
1833 alloc->in(AllocateNode::RawDefaultValue),
1834 header_size, size_in_bytes,
1835 &_igvn);
1836 }
1837 } else {
1838 if (!init->is_complete()) {
1839 // Try to win by zeroing only what the init does not store.
1840 // We can also try to do some peephole optimizations,
1841 // such as combining some adjacent subword stores.
1842 rawmem = init->complete_stores(control, rawmem, object,
1843 header_size, size_in_bytes, &_igvn);
1844 }
1845 // We have no more use for this link, since the AllocateNode goes away:
1846 init->set_req(InitializeNode::RawAddress, top());
1847 // (If we keep the link, it just confuses the register allocator,
1848 // who thinks he sees a real use of the address by the membar.)
1849 }
1850
1851 return rawmem;
1852 }
1853
1854 // Generate prefetch instructions for next allocations.
1855 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
1856 Node*& contended_phi_rawmem,
1857 Node* old_eden_top, Node* new_eden_top,
1858 Node* length) {
1859 enum { fall_in_path = 1, pf_path = 2 };
1860 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
1861 // Generate prefetch allocation with watermark check.
1862 // As an allocation hits the watermark, we will prefetch starting
1863 // at a "distance" away from watermark.
1864
1865 Node *pf_region = new RegionNode(3);
1866 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY,
1867 TypeRawPtr::BOTTOM );
1868 // I/O is used for Prefetch
1869 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO );
1870
1871 Node *thread = new ThreadLocalNode();
1872 transform_later(thread);
1873
1874 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread,
1875 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
1876 transform_later(eden_pf_adr);
1877
1878 Node *old_pf_wm = new LoadPNode(needgc_false,
1879 contended_phi_rawmem, eden_pf_adr,
1880 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM,
1881 MemNode::unordered);
1882 transform_later(old_pf_wm);
1883
1884 // check against new_eden_top
1885 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm );
1886 transform_later(need_pf_cmp);
1887 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge );
1888 transform_later(need_pf_bol);
1889 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol,
1890 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1891 transform_later(need_pf_iff);
1892
1893 // true node, add prefetchdistance
1894 Node *need_pf_true = new IfTrueNode( need_pf_iff );
1895 transform_later(need_pf_true);
1896
1897 Node *need_pf_false = new IfFalseNode( need_pf_iff );
1898 transform_later(need_pf_false);
1899
1900 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm,
1901 _igvn.MakeConX(AllocatePrefetchDistance) );
1902 transform_later(new_pf_wmt );
1903 new_pf_wmt->set_req(0, need_pf_true);
1904
1905 Node *store_new_wmt = new StorePNode(need_pf_true,
1906 contended_phi_rawmem, eden_pf_adr,
1907 TypeRawPtr::BOTTOM, new_pf_wmt,
1908 MemNode::unordered);
1909 transform_later(store_new_wmt);
1910
1911 // adding prefetches
1912 pf_phi_abio->init_req( fall_in_path, i_o );
1913
1914 Node *prefetch_adr;
1915 Node *prefetch;
1916 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1917 uint step_size = AllocatePrefetchStepSize;
1918 uint distance = 0;
1919
1920 for ( uint i = 0; i < lines; i++ ) {
1921 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt,
1922 _igvn.MakeConX(distance) );
1923 transform_later(prefetch_adr);
1924 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1925 transform_later(prefetch);
1926 distance += step_size;
1927 i_o = prefetch;
1928 }
1929 pf_phi_abio->set_req( pf_path, i_o );
1930
1931 pf_region->init_req( fall_in_path, need_pf_false );
1932 pf_region->init_req( pf_path, need_pf_true );
1933
1934 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
1935 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
1936
1937 transform_later(pf_region);
1938 transform_later(pf_phi_rawmem);
1939 transform_later(pf_phi_abio);
1940
1941 needgc_false = pf_region;
1942 contended_phi_rawmem = pf_phi_rawmem;
1943 i_o = pf_phi_abio;
1944 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) {
1945 // Insert a prefetch instruction for each allocation.
1946 // This code is used to generate 1 prefetch instruction per cache line.
1947
1948 // Generate several prefetch instructions.
1949 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1950 uint step_size = AllocatePrefetchStepSize;
1951 uint distance = AllocatePrefetchDistance;
1952
1953 // Next cache address.
1954 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top,
1955 _igvn.MakeConX(step_size + distance));
1956 transform_later(cache_adr);
1957 cache_adr = new CastP2XNode(needgc_false, cache_adr);
1958 transform_later(cache_adr);
1959 // Address is aligned to execute prefetch to the beginning of cache line size
1960 // (it is important when BIS instruction is used on SPARC as prefetch).
1961 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1));
1962 cache_adr = new AndXNode(cache_adr, mask);
1963 transform_later(cache_adr);
1964 cache_adr = new CastX2PNode(cache_adr);
1965 transform_later(cache_adr);
1966
1967 // Prefetch
1968 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr );
1969 prefetch->set_req(0, needgc_false);
1970 transform_later(prefetch);
1971 contended_phi_rawmem = prefetch;
1972 Node *prefetch_adr;
1973 distance = step_size;
1974 for ( uint i = 1; i < lines; i++ ) {
1975 prefetch_adr = new AddPNode( cache_adr, cache_adr,
1976 _igvn.MakeConX(distance) );
1977 transform_later(prefetch_adr);
1978 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr );
1979 transform_later(prefetch);
1980 distance += step_size;
1981 contended_phi_rawmem = prefetch;
1982 }
1983 } else if( AllocatePrefetchStyle > 0 ) {
1984 // Insert a prefetch for each allocation only on the fast-path
1985 Node *prefetch_adr;
1986 Node *prefetch;
1987 // Generate several prefetch instructions.
1988 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1989 uint step_size = AllocatePrefetchStepSize;
1990 uint distance = AllocatePrefetchDistance;
1991 for ( uint i = 0; i < lines; i++ ) {
1992 prefetch_adr = new AddPNode( old_eden_top, new_eden_top,
1993 _igvn.MakeConX(distance) );
1994 transform_later(prefetch_adr);
1995 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1996 // Do not let it float too high, since if eden_top == eden_end,
1997 // both might be null.
1998 if( i == 0 ) { // Set control for first prefetch, next follows it
1999 prefetch->init_req(0, needgc_false);
2000 }
2001 transform_later(prefetch);
2002 distance += step_size;
2003 i_o = prefetch;
2004 }
2005 }
2006 return i_o;
2007 }
2008
2009
2010 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
2011 expand_allocate_common(alloc, NULL,
2012 OptoRuntime::new_instance_Type(),
2013 OptoRuntime::new_instance_Java());
2014 }
2015
2016 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
2017 Node* length = alloc->in(AllocateNode::ALength);
2018 InitializeNode* init = alloc->initialization();
2019 Node* klass_node = alloc->in(AllocateNode::KlassNode);
2020 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
2021 address slow_call_address; // Address of slow call
2022 if (init != NULL && init->is_complete_with_arraycopy() &&
2023 k->is_type_array_klass()) {
2024 // Don't zero type array during slow allocation in VM since
2025 // it will be initialized later by arraycopy in compiled code.
2026 slow_call_address = OptoRuntime::new_array_nozero_Java();
2027 } else {
2028 slow_call_address = OptoRuntime::new_array_Java();
2029 }
2030 expand_allocate_common(alloc, length,
2031 OptoRuntime::new_array_Type(),
2032 slow_call_address);
2033 }
2034
2035 //-------------------mark_eliminated_box----------------------------------
2036 //
2037 // During EA obj may point to several objects but after few ideal graph
2038 // transformations (CCP) it may point to only one non escaping object
2039 // (but still using phi), corresponding locks and unlocks will be marked
2040 // for elimination. Later obj could be replaced with a new node (new phi)
2041 // and which does not have escape information. And later after some graph
2042 // reshape other locks and unlocks (which were not marked for elimination
2043 // before) are connected to this new obj (phi) but they still will not be
2044 // marked for elimination since new obj has no escape information.
2045 // Mark all associated (same box and obj) lock and unlock nodes for
2046 // elimination if some of them marked already.
2047 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) {
2048 if (oldbox->as_BoxLock()->is_eliminated())
2049 return; // This BoxLock node was processed already.
2050
2051 // New implementation (EliminateNestedLocks) has separate BoxLock
2052 // node for each locked region so mark all associated locks/unlocks as
2053 // eliminated even if different objects are referenced in one locked region
2054 // (for example, OSR compilation of nested loop inside locked scope).
2055 if (EliminateNestedLocks ||
2056 oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj)) {
2057 // Box is used only in one lock region. Mark this box as eliminated.
2058 _igvn.hash_delete(oldbox);
2059 oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value
2060 _igvn.hash_insert(oldbox);
2061
2062 for (uint i = 0; i < oldbox->outcnt(); i++) {
2063 Node* u = oldbox->raw_out(i);
2064 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) {
2065 AbstractLockNode* alock = u->as_AbstractLock();
2066 // Check lock's box since box could be referenced by Lock's debug info.
2067 if (alock->box_node() == oldbox) {
2068 // Mark eliminated all related locks and unlocks.
2069 #ifdef ASSERT
2070 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4");
2071 #endif
2072 alock->set_non_esc_obj();
2073 }
2074 }
2075 }
2076 return;
2077 }
2078
2079 // Create new "eliminated" BoxLock node and use it in monitor debug info
2080 // instead of oldbox for the same object.
2081 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
2082
2083 // Note: BoxLock node is marked eliminated only here and it is used
2084 // to indicate that all associated lock and unlock nodes are marked
2085 // for elimination.
2086 newbox->set_eliminated();
2087 transform_later(newbox);
2088
2089 // Replace old box node with new box for all users of the same object.
2090 for (uint i = 0; i < oldbox->outcnt();) {
2091 bool next_edge = true;
2092
2093 Node* u = oldbox->raw_out(i);
2094 if (u->is_AbstractLock()) {
2095 AbstractLockNode* alock = u->as_AbstractLock();
2096 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) {
2097 // Replace Box and mark eliminated all related locks and unlocks.
2098 #ifdef ASSERT
2099 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5");
2100 #endif
2101 alock->set_non_esc_obj();
2102 _igvn.rehash_node_delayed(alock);
2103 alock->set_box_node(newbox);
2104 next_edge = false;
2105 }
2106 }
2107 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) {
2108 FastLockNode* flock = u->as_FastLock();
2109 assert(flock->box_node() == oldbox, "sanity");
2110 _igvn.rehash_node_delayed(flock);
2111 flock->set_box_node(newbox);
2112 next_edge = false;
2113 }
2114
2115 // Replace old box in monitor debug info.
2116 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
2117 SafePointNode* sfn = u->as_SafePoint();
2118 JVMState* youngest_jvms = sfn->jvms();
2119 int max_depth = youngest_jvms->depth();
2120 for (int depth = 1; depth <= max_depth; depth++) {
2121 JVMState* jvms = youngest_jvms->of_depth(depth);
2122 int num_mon = jvms->nof_monitors();
2123 // Loop over monitors
2124 for (int idx = 0; idx < num_mon; idx++) {
2125 Node* obj_node = sfn->monitor_obj(jvms, idx);
2126 Node* box_node = sfn->monitor_box(jvms, idx);
2127 if (box_node == oldbox && obj_node->eqv_uncast(obj)) {
2128 int j = jvms->monitor_box_offset(idx);
2129 _igvn.replace_input_of(u, j, newbox);
2130 next_edge = false;
2131 }
2132 }
2133 }
2134 }
2135 if (next_edge) i++;
2136 }
2137 }
2138
2139 //-----------------------mark_eliminated_locking_nodes-----------------------
2140 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) {
2141 if (EliminateNestedLocks) {
2142 if (alock->is_nested()) {
2143 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity");
2144 return;
2145 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened
2146 // Only Lock node has JVMState needed here.
2147 // Not that preceding claim is documented anywhere else.
2148 if (alock->jvms() != NULL) {
2149 if (alock->as_Lock()->is_nested_lock_region()) {
2150 // Mark eliminated related nested locks and unlocks.
2151 Node* obj = alock->obj_node();
2152 BoxLockNode* box_node = alock->box_node()->as_BoxLock();
2153 assert(!box_node->is_eliminated(), "should not be marked yet");
2154 // Note: BoxLock node is marked eliminated only here
2155 // and it is used to indicate that all associated lock
2156 // and unlock nodes are marked for elimination.
2157 box_node->set_eliminated(); // Box's hash is always NO_HASH here
2158 for (uint i = 0; i < box_node->outcnt(); i++) {
2159 Node* u = box_node->raw_out(i);
2160 if (u->is_AbstractLock()) {
2161 alock = u->as_AbstractLock();
2162 if (alock->box_node() == box_node) {
2163 // Verify that this Box is referenced only by related locks.
2164 assert(alock->obj_node()->eqv_uncast(obj), "");
2165 // Mark all related locks and unlocks.
2166 #ifdef ASSERT
2167 alock->log_lock_optimization(C, "eliminate_lock_set_nested");
2168 #endif
2169 alock->set_nested();
2170 }
2171 }
2172 }
2173 } else {
2174 #ifdef ASSERT
2175 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region");
2176 if (C->log() != NULL)
2177 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output
2178 #endif
2179 }
2180 }
2181 return;
2182 }
2183 // Process locks for non escaping object
2184 assert(alock->is_non_esc_obj(), "");
2185 } // EliminateNestedLocks
2186
2187 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2188 // Look for all locks of this object and mark them and
2189 // corresponding BoxLock nodes as eliminated.
2190 Node* obj = alock->obj_node();
2191 for (uint j = 0; j < obj->outcnt(); j++) {
2192 Node* o = obj->raw_out(j);
2193 if (o->is_AbstractLock() &&
2194 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2195 alock = o->as_AbstractLock();
2196 Node* box = alock->box_node();
2197 // Replace old box node with new eliminated box for all users
2198 // of the same object and mark related locks as eliminated.
2199 mark_eliminated_box(box, obj);
2200 }
2201 }
2202 }
2203 }
2204
2205 // we have determined that this lock/unlock can be eliminated, we simply
2206 // eliminate the node without expanding it.
2207 //
2208 // Note: The membar's associated with the lock/unlock are currently not
2209 // eliminated. This should be investigated as a future enhancement.
2210 //
2211 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2212
2213 if (!alock->is_eliminated()) {
2214 return false;
2215 }
2216 #ifdef ASSERT
2217 if (!alock->is_coarsened()) {
2218 // Check that new "eliminated" BoxLock node is created.
2219 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2220 assert(oldbox->is_eliminated(), "should be done already");
2221 }
2222 #endif
2223
2224 alock->log_lock_optimization(C, "eliminate_lock");
2225
2226 #ifndef PRODUCT
2227 if (PrintEliminateLocks) {
2228 if (alock->is_Lock()) {
2229 tty->print_cr("++++ Eliminated: %d Lock", alock->_idx);
2230 } else {
2231 tty->print_cr("++++ Eliminated: %d Unlock", alock->_idx);
2232 }
2233 }
2234 #endif
2235
2236 Node* mem = alock->in(TypeFunc::Memory);
2237 Node* ctrl = alock->in(TypeFunc::Control);
2238 guarantee(ctrl != NULL, "missing control projection, cannot replace_node() with NULL");
2239
2240 extract_call_projections(alock);
2241 // There are 2 projections from the lock. The lock node will
2242 // be deleted when its last use is subsumed below.
2243 assert(alock->outcnt() == 2 &&
2244 _fallthroughproj != NULL &&
2245 _memproj_fallthrough != NULL,
2246 "Unexpected projections from Lock/Unlock");
2247
2248 Node* fallthroughproj = _fallthroughproj;
2249 Node* memproj_fallthrough = _memproj_fallthrough;
2250
2251 // The memory projection from a lock/unlock is RawMem
2252 // The input to a Lock is merged memory, so extract its RawMem input
2253 // (unless the MergeMem has been optimized away.)
2254 if (alock->is_Lock()) {
2255 // Seach for MemBarAcquireLock node and delete it also.
2256 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2257 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, "");
2258 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2259 Node* memproj = membar->proj_out(TypeFunc::Memory);
2260 _igvn.replace_node(ctrlproj, fallthroughproj);
2261 _igvn.replace_node(memproj, memproj_fallthrough);
2262
2263 // Delete FastLock node also if this Lock node is unique user
2264 // (a loop peeling may clone a Lock node).
2265 Node* flock = alock->as_Lock()->fastlock_node();
2266 if (flock->outcnt() == 1) {
2267 assert(flock->unique_out() == alock, "sanity");
2268 _igvn.replace_node(flock, top());
2269 }
2270 }
2271
2272 // Seach for MemBarReleaseLock node and delete it also.
2273 if (alock->is_Unlock() && ctrl->is_Proj() && ctrl->in(0)->is_MemBar()) {
2274 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2275 assert(membar->Opcode() == Op_MemBarReleaseLock &&
2276 mem->is_Proj() && membar == mem->in(0), "");
2277 _igvn.replace_node(fallthroughproj, ctrl);
2278 _igvn.replace_node(memproj_fallthrough, mem);
2279 fallthroughproj = ctrl;
2280 memproj_fallthrough = mem;
2281 ctrl = membar->in(TypeFunc::Control);
2282 mem = membar->in(TypeFunc::Memory);
2283 }
2284
2285 _igvn.replace_node(fallthroughproj, ctrl);
2286 _igvn.replace_node(memproj_fallthrough, mem);
2287 return true;
2288 }
2289
2290
2291 //------------------------------expand_lock_node----------------------
2292 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
2293
2294 Node* ctrl = lock->in(TypeFunc::Control);
2295 Node* mem = lock->in(TypeFunc::Memory);
2296 Node* obj = lock->obj_node();
2297 Node* box = lock->box_node();
2298 Node* flock = lock->fastlock_node();
2299
2300 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2301
2302 // Make the merge point
2303 Node *region;
2304 Node *mem_phi;
2305 Node *slow_path;
2306
2307 if (UseOptoBiasInlining) {
2308 /*
2309 * See the full description in MacroAssembler::biased_locking_enter().
2310 *
2311 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) {
2312 * // The object is biased.
2313 * proto_node = klass->prototype_header;
2314 * o_node = thread | proto_node;
2315 * x_node = o_node ^ mark_word;
2316 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ?
2317 * // Done.
2318 * } else {
2319 * if( (x_node & biased_lock_mask) != 0 ) {
2320 * // The klass's prototype header is no longer biased.
2321 * cas(&mark_word, mark_word, proto_node)
2322 * goto cas_lock;
2323 * } else {
2324 * // The klass's prototype header is still biased.
2325 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch?
2326 * old = mark_word;
2327 * new = o_node;
2328 * } else {
2329 * // Different thread or anonymous biased.
2330 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask);
2331 * new = thread | old;
2332 * }
2333 * // Try to rebias.
2334 * if( cas(&mark_word, old, new) == 0 ) {
2335 * // Done.
2336 * } else {
2337 * goto slow_path; // Failed.
2338 * }
2339 * }
2340 * }
2341 * } else {
2342 * // The object is not biased.
2343 * cas_lock:
2344 * if( FastLock(obj) == 0 ) {
2345 * // Done.
2346 * } else {
2347 * slow_path:
2348 * OptoRuntime::complete_monitor_locking_Java(obj);
2349 * }
2350 * }
2351 */
2352
2353 region = new RegionNode(5);
2354 // create a Phi for the memory state
2355 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2356
2357 Node* fast_lock_region = new RegionNode(3);
2358 Node* fast_lock_mem_phi = new PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM);
2359
2360 // First, check mark word for the biased lock pattern.
2361 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2362
2363 // Get fast path - mark word has the biased lock pattern.
2364 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node,
2365 markOopDesc::biased_lock_mask_in_place,
2366 markOopDesc::biased_lock_pattern, true);
2367 // fast_lock_region->in(1) is set to slow path.
2368 fast_lock_mem_phi->init_req(1, mem);
2369
2370 // Now check that the lock is biased to the current thread and has
2371 // the same epoch and bias as Klass::_prototype_header.
2372
2373 // Special-case a fresh allocation to avoid building nodes:
2374 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn);
2375 if (klass_node == NULL) {
2376 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
2377 klass_node = transform_later(LoadKlassNode::make(_igvn, NULL, mem, k_adr, _igvn.type(k_adr)->is_ptr()));
2378 #ifdef _LP64
2379 if (UseCompressedClassPointers && klass_node->is_DecodeNKlass()) {
2380 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity");
2381 klass_node->in(1)->init_req(0, ctrl);
2382 } else
2383 #endif
2384 klass_node->init_req(0, ctrl);
2385 }
2386 Node *proto_node = make_load(ctrl, mem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeX_X, TypeX_X->basic_type());
2387
2388 Node* thread = transform_later(new ThreadLocalNode());
2389 Node* cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2390 Node* o_node = transform_later(new OrXNode(cast_thread, proto_node));
2391 Node* x_node = transform_later(new XorXNode(o_node, mark_node));
2392
2393 // Get slow path - mark word does NOT match the value.
2394 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node,
2395 (~markOopDesc::age_mask_in_place), 0);
2396 // region->in(3) is set to fast path - the object is biased to the current thread.
2397 mem_phi->init_req(3, mem);
2398
2399
2400 // Mark word does NOT match the value (thread | Klass::_prototype_header).
2401
2402
2403 // First, check biased pattern.
2404 // Get fast path - _prototype_header has the same biased lock pattern.
2405 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node,
2406 markOopDesc::biased_lock_mask_in_place, 0, true);
2407
2408 not_biased_ctrl = fast_lock_region->in(2); // Slow path
2409 // fast_lock_region->in(2) - the prototype header is no longer biased
2410 // and we have to revoke the bias on this object.
2411 // We are going to try to reset the mark of this object to the prototype
2412 // value and fall through to the CAS-based locking scheme.
2413 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
2414 Node* cas = new StoreXConditionalNode(not_biased_ctrl, mem, adr,
2415 proto_node, mark_node);
2416 transform_later(cas);
2417 Node* proj = transform_later(new SCMemProjNode(cas));
2418 fast_lock_mem_phi->init_req(2, proj);
2419
2420
2421 // Second, check epoch bits.
2422 Node* rebiased_region = new RegionNode(3);
2423 Node* old_phi = new PhiNode( rebiased_region, TypeX_X);
2424 Node* new_phi = new PhiNode( rebiased_region, TypeX_X);
2425
2426 // Get slow path - mark word does NOT match epoch bits.
2427 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node,
2428 markOopDesc::epoch_mask_in_place, 0);
2429 // The epoch of the current bias is not valid, attempt to rebias the object
2430 // toward the current thread.
2431 rebiased_region->init_req(2, epoch_ctrl);
2432 old_phi->init_req(2, mark_node);
2433 new_phi->init_req(2, o_node);
2434
2435 // rebiased_region->in(1) is set to fast path.
2436 // The epoch of the current bias is still valid but we know
2437 // nothing about the owner; it might be set or it might be clear.
2438 Node* cmask = MakeConX(markOopDesc::biased_lock_mask_in_place |
2439 markOopDesc::age_mask_in_place |
2440 markOopDesc::epoch_mask_in_place);
2441 Node* old = transform_later(new AndXNode(mark_node, cmask));
2442 cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2443 Node* new_mark = transform_later(new OrXNode(cast_thread, old));
2444 old_phi->init_req(1, old);
2445 new_phi->init_req(1, new_mark);
2446
2447 transform_later(rebiased_region);
2448 transform_later(old_phi);
2449 transform_later(new_phi);
2450
2451 // Try to acquire the bias of the object using an atomic operation.
2452 // If this fails we will go in to the runtime to revoke the object's bias.
2453 cas = new StoreXConditionalNode(rebiased_region, mem, adr, new_phi, old_phi);
2454 transform_later(cas);
2455 proj = transform_later(new SCMemProjNode(cas));
2456
2457 // Get slow path - Failed to CAS.
2458 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0);
2459 mem_phi->init_req(4, proj);
2460 // region->in(4) is set to fast path - the object is rebiased to the current thread.
2461
2462 // Failed to CAS.
2463 slow_path = new RegionNode(3);
2464 Node *slow_mem = new PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM);
2465
2466 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control
2467 slow_mem->init_req(1, proj);
2468
2469 // Call CAS-based locking scheme (FastLock node).
2470
2471 transform_later(fast_lock_region);
2472 transform_later(fast_lock_mem_phi);
2473
2474 // Get slow path - FastLock failed to lock the object.
2475 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0);
2476 mem_phi->init_req(2, fast_lock_mem_phi);
2477 // region->in(2) is set to fast path - the object is locked to the current thread.
2478
2479 slow_path->init_req(2, ctrl); // Capture slow-control
2480 slow_mem->init_req(2, fast_lock_mem_phi);
2481
2482 transform_later(slow_path);
2483 transform_later(slow_mem);
2484 // Reset lock's memory edge.
2485 lock->set_req(TypeFunc::Memory, slow_mem);
2486
2487 } else {
2488 region = new RegionNode(3);
2489 // create a Phi for the memory state
2490 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2491
2492 // Optimize test; set region slot 2
2493 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2494 mem_phi->init_req(2, mem);
2495 }
2496
2497 // Make slow path call
2498 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(),
2499 OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path,
2500 obj, box, NULL);
2501
2502 extract_call_projections(call);
2503
2504 // Slow path can only throw asynchronous exceptions, which are always
2505 // de-opted. So the compiler thinks the slow-call can never throw an
2506 // exception. If it DOES throw an exception we would need the debug
2507 // info removed first (since if it throws there is no monitor).
2508 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2509 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2510
2511 // Capture slow path
2512 // disconnect fall-through projection from call and create a new one
2513 // hook up users of fall-through projection to region
2514 Node *slow_ctrl = _fallthroughproj->clone();
2515 transform_later(slow_ctrl);
2516 _igvn.hash_delete(_fallthroughproj);
2517 _fallthroughproj->disconnect_inputs(NULL, C);
2518 region->init_req(1, slow_ctrl);
2519 // region inputs are now complete
2520 transform_later(region);
2521 _igvn.replace_node(_fallthroughproj, region);
2522
2523 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2524 mem_phi->init_req(1, memproj );
2525 transform_later(mem_phi);
2526 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2527 }
2528
2529 //------------------------------expand_unlock_node----------------------
2530 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2531
2532 Node* ctrl = unlock->in(TypeFunc::Control);
2533 Node* mem = unlock->in(TypeFunc::Memory);
2534 Node* obj = unlock->obj_node();
2535 Node* box = unlock->box_node();
2536
2537 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2538
2539 // No need for a null check on unlock
2540
2541 // Make the merge point
2542 Node *region;
2543 Node *mem_phi;
2544
2545 if (UseOptoBiasInlining) {
2546 // Check for biased locking unlock case, which is a no-op.
2547 // See the full description in MacroAssembler::biased_locking_exit().
2548 region = new RegionNode(4);
2549 // create a Phi for the memory state
2550 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2551 mem_phi->init_req(3, mem);
2552
2553 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2554 ctrl = opt_bits_test(ctrl, region, 3, mark_node,
2555 markOopDesc::biased_lock_mask_in_place,
2556 markOopDesc::biased_lock_pattern);
2557 } else {
2558 region = new RegionNode(3);
2559 // create a Phi for the memory state
2560 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2561 }
2562
2563 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2564 funlock = transform_later( funlock )->as_FastUnlock();
2565 // Optimize test; set region slot 2
2566 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2567 Node *thread = transform_later(new ThreadLocalNode());
2568
2569 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2570 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2571 "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2572
2573 extract_call_projections(call);
2574
2575 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2576 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2577
2578 // No exceptions for unlocking
2579 // Capture slow path
2580 // disconnect fall-through projection from call and create a new one
2581 // hook up users of fall-through projection to region
2582 Node *slow_ctrl = _fallthroughproj->clone();
2583 transform_later(slow_ctrl);
2584 _igvn.hash_delete(_fallthroughproj);
2585 _fallthroughproj->disconnect_inputs(NULL, C);
2586 region->init_req(1, slow_ctrl);
2587 // region inputs are now complete
2588 transform_later(region);
2589 _igvn.replace_node(_fallthroughproj, region);
2590
2591 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2592 mem_phi->init_req(1, memproj );
2593 mem_phi->init_req(2, mem);
2594 transform_later(mem_phi);
2595 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2596 }
2597
2598 // A value type is returned from the call but we don't know its
2599 // type. Either we get a buffered value (and nothing needs to be done)
2600 // or one of the values being returned is the klass of the value type
2601 // and we need to allocate a value type instance of that type and
2602 // initialize it with other values being returned. In that case, we
2603 // first try a fast path allocation and initialize the value with the
2604 // value klass's pack handler or we fall back to a runtime call.
2605 void PhaseMacroExpand::expand_mh_intrinsic_return(CallStaticJavaNode* call) {
2606 Node* ret = call->proj_out(TypeFunc::Parms);
2607 if (ret == NULL) {
2608 return;
2609 }
2610 // TODO fix this with the calling convention changes
2611 //assert(ret->bottom_type()->is_valuetypeptr()->is__Value(), "unexpected return type from MH intrinsic");
2612 const TypeFunc* tf = call->_tf;
2613 const TypeTuple* domain = OptoRuntime::store_value_type_fields_Type()->domain_cc();
2614 const TypeFunc* new_tf = TypeFunc::make(tf->domain_sig(), tf->domain_cc(), tf->range_sig(), domain);
2615 call->_tf = new_tf;
2616 // Make sure the change of type is applied before projections are
2617 // processed by igvn
2618 _igvn.set_type(call, call->Value(&_igvn));
2619 _igvn.set_type(ret, ret->Value(&_igvn));
2620
2621 // Before any new projection is added:
2622 CallProjections* projs = call->extract_projections(true, true);
2623
2624 Node* ctl = new Node(1);
2625 Node* mem = new Node(1);
2626 Node* io = new Node(1);
2627 Node* ex_ctl = new Node(1);
2628 Node* ex_mem = new Node(1);
2629 Node* ex_io = new Node(1);
2630 Node* res = new Node(1);
2631
2632 Node* cast = transform_later(new CastP2XNode(ctl, res));
2633 Node* mask = MakeConX(0x1);
2634 Node* masked = transform_later(new AndXNode(cast, mask));
2635 Node* cmp = transform_later(new CmpXNode(masked, mask));
2636 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
2637 IfNode* allocation_iff = new IfNode(ctl, bol, PROB_MAX, COUNT_UNKNOWN);
2638 transform_later(allocation_iff);
2639 Node* allocation_ctl = transform_later(new IfTrueNode(allocation_iff));
2640 Node* no_allocation_ctl = transform_later(new IfFalseNode(allocation_iff));
2641
2642 Node* no_allocation_res = transform_later(new CheckCastPPNode(no_allocation_ctl, res, TypeInstPtr::NOTNULL));
2643
2644 Node* mask2 = MakeConX(-2);
2645 Node* masked2 = transform_later(new AndXNode(cast, mask2));
2646 Node* rawklassptr = transform_later(new CastX2PNode(masked2));
2647 Node* klass_node = transform_later(new CheckCastPPNode(allocation_ctl, rawklassptr, TypeKlassPtr::OBJECT_OR_NULL));
2648
2649 Node* slowpath_bol = NULL;
2650 Node* top_adr = NULL;
2651 Node* old_top = NULL;
2652 Node* new_top = NULL;
2653 if (UseTLAB) {
2654 Node* end_adr = NULL;
2655 set_eden_pointers(top_adr, end_adr);
2656 Node* end = make_load(ctl, mem, end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS);
2657 old_top = new LoadPNode(ctl, mem, top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered);
2658 transform_later(old_top);
2659 Node* layout_val = make_load(NULL, mem, klass_node, in_bytes(Klass::layout_helper_offset()), TypeInt::INT, T_INT);
2660 Node* size_in_bytes = ConvI2X(layout_val);
2661 new_top = new AddPNode(top(), old_top, size_in_bytes);
2662 transform_later(new_top);
2663 Node* slowpath_cmp = new CmpPNode(new_top, end);
2664 transform_later(slowpath_cmp);
2665 slowpath_bol = new BoolNode(slowpath_cmp, BoolTest::ge);
2666 transform_later(slowpath_bol);
2667 } else {
2668 slowpath_bol = intcon(1);
2669 top_adr = top();
2670 old_top = top();
2671 new_top = top();
2672 }
2673 IfNode* slowpath_iff = new IfNode(allocation_ctl, slowpath_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN);
2674 transform_later(slowpath_iff);
2675
2676 Node* slowpath_true = new IfTrueNode(slowpath_iff);
2677 transform_later(slowpath_true);
2678
2679 CallStaticJavaNode* slow_call = new CallStaticJavaNode(OptoRuntime::store_value_type_fields_Type(),
2680 StubRoutines::store_value_type_fields_to_buf(),
2681 "store_value_type_fields",
2682 call->jvms()->bci(),
2683 TypePtr::BOTTOM);
2684 slow_call->init_req(TypeFunc::Control, slowpath_true);
2685 slow_call->init_req(TypeFunc::Memory, mem);
2686 slow_call->init_req(TypeFunc::I_O, io);
2687 slow_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2688 slow_call->init_req(TypeFunc::ReturnAdr, call->in(TypeFunc::ReturnAdr));
2689 slow_call->init_req(TypeFunc::Parms, res);
2690
2691 Node* slow_ctl = transform_later(new ProjNode(slow_call, TypeFunc::Control));
2692 Node* slow_mem = transform_later(new ProjNode(slow_call, TypeFunc::Memory));
2693 Node* slow_io = transform_later(new ProjNode(slow_call, TypeFunc::I_O));
2694 Node* slow_res = transform_later(new ProjNode(slow_call, TypeFunc::Parms));
2695 Node* slow_catc = transform_later(new CatchNode(slow_ctl, slow_io, 2));
2696 Node* slow_norm = transform_later(new CatchProjNode(slow_catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci));
2697 Node* slow_excp = transform_later(new CatchProjNode(slow_catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci));
2698
2699 Node* ex_r = new RegionNode(3);
2700 Node* ex_mem_phi = new PhiNode(ex_r, Type::MEMORY, TypePtr::BOTTOM);
2701 Node* ex_io_phi = new PhiNode(ex_r, Type::ABIO);
2702 ex_r->init_req(1, slow_excp);
2703 ex_mem_phi->init_req(1, slow_mem);
2704 ex_io_phi->init_req(1, slow_io);
2705 ex_r->init_req(2, ex_ctl);
2706 ex_mem_phi->init_req(2, ex_mem);
2707 ex_io_phi->init_req(2, ex_io);
2708
2709 transform_later(ex_r);
2710 transform_later(ex_mem_phi);
2711 transform_later(ex_io_phi);
2712
2713 Node* slowpath_false = new IfFalseNode(slowpath_iff);
2714 transform_later(slowpath_false);
2715 Node* rawmem = new StorePNode(slowpath_false, mem, top_adr, TypeRawPtr::BOTTOM, new_top, MemNode::unordered);
2716 transform_later(rawmem);
2717 Node* mark_node = makecon(TypeRawPtr::make((address)markOopDesc::always_locked_prototype()));
2718 rawmem = make_store(slowpath_false, rawmem, old_top, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
2719 rawmem = make_store(slowpath_false, rawmem, old_top, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
2720 if (UseCompressedClassPointers) {
2721 rawmem = make_store(slowpath_false, rawmem, old_top, oopDesc::klass_gap_offset_in_bytes(), intcon(0), T_INT);
2722 }
2723 Node* pack_handler = make_load(slowpath_false, rawmem, klass_node, in_bytes(ValueKlass::pack_handler_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
2724
2725 CallLeafNoFPNode* handler_call = new CallLeafNoFPNode(OptoRuntime::pack_value_type_Type(),
2726 NULL,
2727 "pack handler",
2728 TypeRawPtr::BOTTOM);
2729 handler_call->init_req(TypeFunc::Control, slowpath_false);
2730 handler_call->init_req(TypeFunc::Memory, rawmem);
2731 handler_call->init_req(TypeFunc::I_O, top());
2732 handler_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2733 handler_call->init_req(TypeFunc::ReturnAdr, top());
2734 handler_call->init_req(TypeFunc::Parms, pack_handler);
2735 handler_call->init_req(TypeFunc::Parms+1, old_top);
2736
2737 // We don't know how many values are returned. This assumes the
2738 // worst case, that all available registers are used.
2739 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) {
2740 if (domain->field_at(i) == Type::HALF) {
2741 slow_call->init_req(i, top());
2742 handler_call->init_req(i+1, top());
2743 continue;
2744 }
2745 Node* proj = transform_later(new ProjNode(call, i));
2746 slow_call->init_req(i, proj);
2747 handler_call->init_req(i+1, proj);
2748 }
2749
2750 // We can safepoint at that new call
2751 copy_call_debug_info(call, slow_call);
2752 transform_later(slow_call);
2753 transform_later(handler_call);
2754
2755 Node* handler_ctl = transform_later(new ProjNode(handler_call, TypeFunc::Control));
2756 rawmem = transform_later(new ProjNode(handler_call, TypeFunc::Memory));
2757 Node* slowpath_false_res = transform_later(new ProjNode(handler_call, TypeFunc::Parms));
2758
2759 MergeMemNode* slowpath_false_mem = MergeMemNode::make(mem);
2760 slowpath_false_mem->set_memory_at(Compile::AliasIdxRaw, rawmem);
2761 transform_later(slowpath_false_mem);
2762
2763 Node* r = new RegionNode(4);
2764 Node* mem_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM);
2765 Node* io_phi = new PhiNode(r, Type::ABIO);
2766 Node* res_phi = new PhiNode(r, ret->bottom_type());
2767
2768 r->init_req(1, no_allocation_ctl);
2769 mem_phi->init_req(1, mem);
2770 io_phi->init_req(1, io);
2771 res_phi->init_req(1, no_allocation_res);
2772 r->init_req(2, slow_norm);
2773 mem_phi->init_req(2, slow_mem);
2774 io_phi->init_req(2, slow_io);
2775 res_phi->init_req(2, slow_res);
2776 r->init_req(3, handler_ctl);
2777 mem_phi->init_req(3, slowpath_false_mem);
2778 io_phi->init_req(3, io);
2779 res_phi->init_req(3, slowpath_false_res);
2780
2781 transform_later(r);
2782 transform_later(mem_phi);
2783 transform_later(io_phi);
2784 transform_later(res_phi);
2785
2786 assert(projs->nb_resproj == 1, "unexpected number of results");
2787 _igvn.replace_in_uses(projs->fallthrough_catchproj, r);
2788 _igvn.replace_in_uses(projs->fallthrough_memproj, mem_phi);
2789 _igvn.replace_in_uses(projs->fallthrough_ioproj, io_phi);
2790 _igvn.replace_in_uses(projs->resproj[0], res_phi);
2791 _igvn.replace_in_uses(projs->catchall_catchproj, ex_r);
2792 _igvn.replace_in_uses(projs->catchall_memproj, ex_mem_phi);
2793 _igvn.replace_in_uses(projs->catchall_ioproj, ex_io_phi);
2794
2795 _igvn.replace_node(ctl, projs->fallthrough_catchproj);
2796 _igvn.replace_node(mem, projs->fallthrough_memproj);
2797 _igvn.replace_node(io, projs->fallthrough_ioproj);
2798 _igvn.replace_node(res, projs->resproj[0]);
2799 _igvn.replace_node(ex_ctl, projs->catchall_catchproj);
2800 _igvn.replace_node(ex_mem, projs->catchall_memproj);
2801 _igvn.replace_node(ex_io, projs->catchall_ioproj);
2802 }
2803
2804 //---------------------------eliminate_macro_nodes----------------------
2805 // Eliminate scalar replaced allocations and associated locks.
2806 void PhaseMacroExpand::eliminate_macro_nodes() {
2807 if (C->macro_count() == 0)
2808 return;
2809
2810 // First, attempt to eliminate locks
2811 int cnt = C->macro_count();
2812 for (int i=0; i < cnt; i++) {
2813 Node *n = C->macro_node(i);
2814 if (n->is_AbstractLock()) { // Lock and Unlock nodes
2815 // Before elimination mark all associated (same box and obj)
2816 // lock and unlock nodes.
2817 mark_eliminated_locking_nodes(n->as_AbstractLock());
2818 }
2819 }
2820 bool progress = true;
2821 while (progress) {
2822 progress = false;
2823 for (int i = C->macro_count(); i > 0; i--) {
2824 Node * n = C->macro_node(i-1);
2825 bool success = false;
2826 debug_only(int old_macro_count = C->macro_count(););
2827 if (n->is_AbstractLock()) {
2828 success = eliminate_locking_node(n->as_AbstractLock());
2829 }
2830 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2831 progress = progress || success;
2832 }
2833 }
2834 // Next, attempt to eliminate allocations
2835 _has_locks = false;
2836 progress = true;
2837 while (progress) {
2838 progress = false;
2839 for (int i = C->macro_count(); i > 0; i--) {
2840 Node * n = C->macro_node(i-1);
2841 bool success = false;
2842 debug_only(int old_macro_count = C->macro_count(););
2843 switch (n->class_id()) {
2844 case Node::Class_Allocate:
2845 case Node::Class_AllocateArray:
2846 success = eliminate_allocate_node(n->as_Allocate());
2847 break;
2848 case Node::Class_CallStaticJava: {
2849 CallStaticJavaNode* call = n->as_CallStaticJava();
2850 if (!call->method()->is_method_handle_intrinsic()) {
2851 success = eliminate_boxing_node(n->as_CallStaticJava());
2852 }
2853 break;
2854 }
2855 case Node::Class_Lock:
2856 case Node::Class_Unlock:
2857 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2858 _has_locks = true;
2859 break;
2860 case Node::Class_ArrayCopy:
2861 break;
2862 case Node::Class_OuterStripMinedLoop:
2863 break;
2864 default:
2865 assert(n->Opcode() == Op_LoopLimit ||
2866 n->Opcode() == Op_Opaque1 ||
2867 n->Opcode() == Op_Opaque2 ||
2868 n->Opcode() == Op_Opaque3 ||
2869 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n),
2870 "unknown node type in macro list");
2871 }
2872 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2873 progress = progress || success;
2874 }
2875 }
2876 }
2877
2878 //------------------------------expand_macro_nodes----------------------
2879 // Returns true if a failure occurred.
2880 bool PhaseMacroExpand::expand_macro_nodes() {
2881 // Last attempt to eliminate macro nodes.
2882 eliminate_macro_nodes();
2883
2884 // Make sure expansion will not cause node limit to be exceeded.
2885 // Worst case is a macro node gets expanded into about 200 nodes.
2886 // Allow 50% more for optimization.
2887 if (C->check_node_count(C->macro_count() * 300, "out of nodes before macro expansion" ) )
2888 return true;
2889
2890 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2891 bool progress = true;
2892 while (progress) {
2893 progress = false;
2894 for (int i = C->macro_count(); i > 0; i--) {
2895 Node * n = C->macro_node(i-1);
2896 bool success = false;
2897 debug_only(int old_macro_count = C->macro_count(););
2898 if (n->Opcode() == Op_LoopLimit) {
2899 // Remove it from macro list and put on IGVN worklist to optimize.
2900 C->remove_macro_node(n);
2901 _igvn._worklist.push(n);
2902 success = true;
2903 } else if (n->Opcode() == Op_CallStaticJava) {
2904 CallStaticJavaNode* call = n->as_CallStaticJava();
2905 if (!call->method()->is_method_handle_intrinsic()) {
2906 // Remove it from macro list and put on IGVN worklist to optimize.
2907 C->remove_macro_node(n);
2908 _igvn._worklist.push(n);
2909 success = true;
2910 }
2911 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) {
2912 _igvn.replace_node(n, n->in(1));
2913 success = true;
2914 #if INCLUDE_RTM_OPT
2915 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) {
2916 assert(C->profile_rtm(), "should be used only in rtm deoptimization code");
2917 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), "");
2918 Node* cmp = n->unique_out();
2919 #ifdef ASSERT
2920 // Validate graph.
2921 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), "");
2922 BoolNode* bol = cmp->unique_out()->as_Bool();
2923 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() &&
2924 (bol->_test._test == BoolTest::ne), "");
2925 IfNode* ifn = bol->unique_out()->as_If();
2926 assert((ifn->outcnt() == 2) &&
2927 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != NULL, "");
2928 #endif
2929 Node* repl = n->in(1);
2930 if (!_has_locks) {
2931 // Remove RTM state check if there are no locks in the code.
2932 // Replace input to compare the same value.
2933 repl = (cmp->in(1) == n) ? cmp->in(2) : cmp->in(1);
2934 }
2935 _igvn.replace_node(n, repl);
2936 success = true;
2937 #endif
2938 } else if (n->Opcode() == Op_OuterStripMinedLoop) {
2939 n->as_OuterStripMinedLoop()->adjust_strip_mined_loop(&_igvn);
2940 C->remove_macro_node(n);
2941 success = true;
2942 }
2943 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2944 progress = progress || success;
2945 }
2946 }
2947
2948 // expand arraycopy "macro" nodes first
2949 // For ReduceBulkZeroing, we must first process all arraycopy nodes
2950 // before the allocate nodes are expanded.
2951 int macro_idx = C->macro_count() - 1;
2952 while (macro_idx >= 0) {
2953 Node * n = C->macro_node(macro_idx);
2954 assert(n->is_macro(), "only macro nodes expected here");
2955 if (_igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())) {
2956 // node is unreachable, so don't try to expand it
2957 C->remove_macro_node(n);
2958 } else if (n->is_ArrayCopy()){
2959 int macro_count = C->macro_count();
2960 expand_arraycopy_node(n->as_ArrayCopy());
2961 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2962 }
2963 if (C->failing()) return true;
2964 macro_idx --;
2965 }
2966
2967 // expand "macro" nodes
2968 // nodes are removed from the macro list as they are processed
2969 while (C->macro_count() > 0) {
2970 int macro_count = C->macro_count();
2971 Node * n = C->macro_node(macro_count-1);
2972 assert(n->is_macro(), "only macro nodes expected here");
2973 if (_igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())) {
2974 // node is unreachable, so don't try to expand it
2975 C->remove_macro_node(n);
2976 continue;
2977 }
2978 switch (n->class_id()) {
2979 case Node::Class_Allocate:
2980 expand_allocate(n->as_Allocate());
2981 break;
2982 case Node::Class_AllocateArray:
2983 expand_allocate_array(n->as_AllocateArray());
2984 break;
2985 case Node::Class_Lock:
2986 expand_lock_node(n->as_Lock());
2987 break;
2988 case Node::Class_Unlock:
2989 expand_unlock_node(n->as_Unlock());
2990 break;
2991 case Node::Class_CallStaticJava:
2992 expand_mh_intrinsic_return(n->as_CallStaticJava());
2993 C->remove_macro_node(n);
2994 break;
2995 default:
2996 assert(false, "unknown node type in macro list");
2997 }
2998 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2999 if (C->failing()) return true;
3000 }
3001
3002 _igvn.set_delay_transform(false);
3003 _igvn.optimize();
3004 if (C->failing()) return true;
3005 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
3006 return bs->expand_macro_nodes(this);
3007 }